Ton Kloks and Yue-Li Wang

Advances in Graph Algorithms

October 10, 2013

Z Y
<
y
<
<

X z
Preface

We prepared this book as a course textbook for our students in Taiwan. Our
aim was to write a book about some currently popular topics such as expo-
nential algorithms, fixed-parameter algorithms and algorithms using decom-
position trees of graphs. Especially for this last topic we found it necessary
to include a chapter on graph classes. The chapter on decomposition trees
includes some basics of the graph minor theory and such topics as tree de-
compositions and rank decompositions. To explain these concepts we found
it beneficial to include a chapter which explains the classes of chordal graphs
and distance-hereditary graphs.

In the chapter on exponential algorithms we start with some introduc-

tory examples on independent sets, chromatic number, domatic partitions, set
cover, etc. The highlight of this chapter is the method based on the inclusion
- exclusion principle. We exemplify this method by algorithms for chromatic
number and triangle partitions.

In the chapter on graph classes we put an emphasis on perfect graphs.

We explain the perfect graph theorem and the strong perfect graph theorem.
We discuss the special classes of chordal graphs and interval graphs, compa-
rability graphs, cographs and distance-hereditary graphs. We put an empha-
sis on the clique trees of chordal graphs, cotrees for cographs and the split-
decomposition trees for distance-hereditary graphs. We briefly explain the use
of these data structures by some elementary examples.

In the chapter on fixed-parameter algorithms we explain the basic notions

of the search tree technique and the concept of a kernelization. We explain
the search tree technique by some basic examples such as vertex cover and
minimum fill-in. As an example of a kernelization we show how a matching
can be used to find a kernel for vertex cover. We explain the iterative com-
pression techniques by the fixed-parameter algorithm for graph bipartization,
feedback vertex set and homogeneous colorings of perfect graphs.
VI Preface

In the chapter on decomposition trees we start with an explanation of the

graph minor theory. As a basic example we show that this implies that feed-
back vertex set is fixed-parameter tractable. Next, we introduce treewidth
as a parametrization of chordal graphs. We show that the class of graphs of
bounded treewidth is closed under minors and, as a consequence, that the
class can be recognized in O(n2 ) time. We give an easy linear-time algorithm
for the recognition of graphs with treewidth two. For general k we explain
the historic O(nk+2 ) algorithm of Arnborg, Corneil and Proskurowski, for
the recognition of partial k-trees. We introduce tree decompositions as the
clique trees of chordal embeddings of a graph. Likewise, we introduce rank
decompositions as a parametrization of the decomposition trees for distance-
hereditary graphs. We close the chapter with a brief discussion of monadic
second-order logic.

After each chapter we included some basic exercises. For experienced stu-
dents these exercises are probably too easy. We made the decision to concen-
trate on elementary exercises in order not to distract the student too much
from the main topics. The exercises are primarily meant as a check for the
students that they understand the material of the chapter.

One of us, Ton Kloks, taught this course during the spring semester of 2012
at the Department of Computer Science of National Tsing Hua University. We
are indebted to our students for their enthusiasm. Special thanks go to Ching-
Hao Liu, who operated as a teaching assistant. He spotted numerous typos in
the original manuscript. Thanks to him this text has greatly improved.

Ton Kloks also wishes to express his gratitude towards the Department of
Computer Science of National Tsing Hua University for their hospitality and
for giving him the opportunity to teach this course. He is indebted to the
National Science Council of Taiwan for their financial support.

Ton Kloks
Department of Computer Science
National Tsing Hua University
Hsinchu, Taiwan.
Yue-Li Wang
Department of Information Management
National Taiwan University of Science and Technology
No. 43, Sec. 4, Keelung Rd., Taipei, 106, Taiwan
[email protected]
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Exponential Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Independent set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Chromatic number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Three-coloring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Domatic partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 The traveling salesman problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Set cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6 Dominating set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.1 Dominating set in bipartite graphs . . . . . . . . . . . . . . . . . . . 21
2.7 Inclusion – exclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7.1 Triangle partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7.2 An improved algorithm for triangle partition . . . . . . . . . . 27
2.8 Subset sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Graph Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1 Perfect graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Comparability graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.1 Recognition of comparability graphs . . . . . . . . . . . . . . . . . 47
3.3 Cographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.1 Cotrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.2 Finding cliques in cographs . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4 Distance-hereditary graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.1 Decomposition trees for DH-graphs . . . . . . . . . . . . . . . . . . 60
3.4.2 Feedback vertex set in DH-graphs . . . . . . . . . . . . . . . . . . . . 62
3.5 Chordal graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.5.1 Clique trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.2 Algorithms for independent set, clique and vertex
coloring in chordal graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 71
VIII Contents

3.6 Interval graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.7 Permutation graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.7.1 Interval containment graphs . . . . . . . . . . . . . . . . . . . . . . . . 82
3.7.2 Cliques and independent sets in permutation graphs . . . 83
3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

4 Fixed-parameter Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.1 Vertex cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.2 A kernel for vertex cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3 A better search-tree algorithm for vertex cover . . . . . . . . . . . . . . . 98
4.4 Minimum fill-in . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5 Odd cycle transversals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.5.1 Feedback vertex set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.6 Homogeneous coloring of perfect graphs . . . . . . . . . . . . . . . . . . . . 114
4.7 Color coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5 Decomposition Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

5.1 Graph minors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.2 Parameterized feedback vertex set . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.3 Treewidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3.1 An algorithm for treewidth two . . . . . . . . . . . . . . . . . . . . . . 133
5.3.2 k-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.3.3 An O(nk+2 ) algorithm for treewidth . . . . . . . . . . . . . . . . . 137
5.3.4 Maximum clique in graphs of bounded treewidth . . . . . . 139
5.3.5 Chromatic number for graphs of bounded treewidth . . . . 141
5.3.6 Disjoint cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.4 Pathwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.5 Rankwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.6 Monadic second-order logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
1
Introduction

This book is about graphs with an emphasis on algorithms. The choice of

our topics is personal. The topics that appear as chapters in this book are
a selection of the currently ‘hot’ research areas in graph theory and graph
algorithms.

We expect that our students are familiar with the basic notions of graphs
and algorithms. In this chapter we review some of our notations.

A graph G is an ordered pair of sets, G = (V, E). The elements of the set V
are called the vertices of G. Sometimes we call the vertices ‘points.’ We require
that the set V is nonempty. The elements of the set E are called the edges of
G. Sometimes we call an edge a ‘line.’ An edge is a set of two vertices. When
the graph G is not clear from the context we use V(G) and E(G) to denote its
set of vertices and its set of edges. The two vertices of an edge are called the
endpoints of the edge. Two vertices x and y are adjacent in G if {x, y} ∈ E;
they are nonadjacent if {x, y} ∈ / E. When e = {x, y} ∈ E we say that e is an
edge which runs between x and y or, simpler, we say that e is an edge between
x and y.

The neighborhood of a vertex x is the set of vertices that are adjacent to x.

We use the notation N(x) to denote the neighborhood of x. When the graph
G is not clear from the context we use the notation NG (x). When two vertices
are adjacent we also call them neighbors of each other. When two vertices
are not adjacent we call them nonneighbors. We use N[x] = {x} ∪ N(x) as a
notation for the ‘closed neighborhood’ of a vertex x. The degree of a vertex x
is the number of its neighbors, that is, |N(x)|. For a subset C ⊆ V of vertices
we write
N(C) = { y | y ∈ V \ C and N(y) ∩ C 6= ∅ }.
Thus, for C ⊆ V, N(C) is the set of vertices that are not in C and that have a
neighbor in C. We also write N[C] = C ∪ N(C) for the closed neighborhood of
a subset C.
2 1 Introduction

All our graphs will be finite, that means that V is a finite set. A graph
G = (V, E) is called empty if E = ∅.

Let G = (V, E) be a graph. A graph G0 = (V 0 , E0 ) is a subgraph of G if

V ⊆ V and E0 ⊆ E. The complement of a graph G = (V, E) is the graph Ḡ
0

with

V(Ḡ) = V and E(Ḡ) = { {x, y} | x ∈ V and y ∈ V and {x, y} ∈

/ E }.

Let G = (V, E) be a graph and let S ⊆ V be some nonempty subset of vertices.

The graph G[S] induced by S has the set S as its set of vertices. As its set of
edges it has those elements of E of which both endpoints are in S. If G = (V, E)
is a graph and if S ⊂ V is a set of vertices such that V \ S 6= ∅ then we write
G − S for the subgraph of G induced by V \ S. If S = {x} then we write G − x
instead of G − {x}. If F ⊆ E is a set of edges then we write G − F = (V, E \ F).
If F = {e} then we write G − e instead of G − {e}.

Let G = (V, E) be a graph. The linegraph L(G) of G is the graph which

has the edges of G as its vertices. Two vertices of L(G) are adjacent if the
corresponding edges in G have a vertex in common.

Let G = (V, E) be a graph. A path is an ordered sequence of distinct ver-

tices. A path has at least one vertex. We use the notation

P = [x1 , . . . , xt ] (1.1)

to denote a path P. The set of vertices of P is {x1 , . . . , xt } ⊆ V. There is an

edge between any two vertices in P if and only if the vertices are consecutive
in the ordering. In some cases we allow edges between vertices that are not
consecutive in the ordering. In that case, those edges are called chords. When
we want to emphasize that a path has no chords, we will call it ‘chordless.’

The vertices x1 and xt of a path P as in (1.1) are called the endvertices, or

endpoints of P. The length of a path is the number of edges in the path. When
the endvertices of a path P are a and b then we refer to P as an a, b-path.

A cycle in G = (V, E) is an ordered sequence of distinct vertices. We use

the notation
C = [c1 , . . . , ct ] (1.2)
to denote a cycle C. A cycle has at least three vertices. There is an edge be-
tween consecutive vertices and between the first and last vertex of C; these
edges are called the edges of C. A cycle is even or odd if the number of edges
of C is even or odd. If there are no other edges between the vertices of C,
we say that C is chordless. If there are other edges between vertices of C then
these edges are called chords. The length of a cycle is the number of vertices
 1 Introduction 3

in it. If the cycle is chordless then the length is also the number of edges in it.
When the set of vertices of a chordless path or a chordless cycle is the set of
vertices V of G, we say that G is a path or a cycle. Notice that we call a graph
a path or cycle only when it is chordless.

A graph G = (V, E) is connected if there is a path between any two ver-

tices. Being connected by a path defines an equivalence relation on the ver-
tices of G. The equivalence classes are called the components of G. If a com-
ponent consists of only one vertex then this vertex is called an isolated vertex
of G.

A tree is a connected graph without any cycles. When T = (V, E) is a tree

then |E| = |V| − 1. Especially when T is a tree we call its vertices points and its
edges lines. When G is a graph in which all components induce trees in G we
call G a forest. When T = (V, E) is a tree then a vertex of degree one is called
a leaf of T .

A clique in a graph G = (V, E) is a nonempty set of vertices Ω such that

every pair of vertices in Ω is adjacent. When Ω is a clique with a maximal
number of elements among all cliques in G we say that Ω is a maximum clique
in G. A maximal clique in G is a clique Ω such that every vertex z ∈ V \ Ω has
at least one nonneighbor in Ω.

An independent set in a graph G = (V, E) is a nonempty set of vertices A

such that there is no edge between any pair of vertices in A. In other words, A
is an independent set in G if A induces a clique in Ḡ. An independent set A in
a graph G is maximum if it has the largest cardinality among all independent
sets in G. Of course, there could be more than one maximum independent
set. An independent set A in G is maximal if every vertex of V \ A has at least
one neighbor in A.

A coloring of a graph G = (V, E) is a partition of its vertices such that

each part of the partition is an independent set. The parts of the partition are
called the color classes. The minimum number of color classes in a coloring
of G is the chromatic number of G. We denote the chromatic number of G by
χ(G). A graph is bipartite if its chromatic number is at most two. A graph is
bipartite if and only if it has no cycles with an odd length.

An edge coloring of a graph G = (V, E) is a coloring of its linegraph L(G).

Each color class is a matching in G, that is, it’s a set of edges in G such that
no two edges in the set have a vertex in common. The chromatic index of G
is the chromatic number of L(G).

A clique cover of a graph G = (V, E) is a coloring of its complement Ḡ.

Thus each color class induces a clique in G. The minimal number of color
4 1 Introduction

classes in a clique cover of G is called the clique cover number; we denote

it by κ(G). Sometimes we allow the cliques that form the color classes of
a clique cover to overlap. It is always easy to change a clique cover with
overlapping cliques into one, with at most as many cliques, where the cliques
form a partition.

Usually we denote the number of vertices and edges of a graph G = (V, E)

as n and m. We express the worst-case runtime of an algorithm on G in the
number of vertices of G. An algorithm on G is polynomial if it can be im-
plemented to run in O(p(n)) time, for some polynomial p(n). Usually we
express the runtime only in the number of vertices. We make an exception for
the linear-time algorithms. Such algorithms run in time O(n+m). A quadratic
algorithm is an algorithm that runs in O(n2 ) time. A cubic time algorithm runs
in O(n3 ) time, etc.

In our book we won’t go into complexity theory. We assume that our reader
is familiar with the basic notion of NP-completeness. Primarily, we will be con-
cerned with the question whether there exists a polynomial-time algorithm for
a problem or whether the problem is NP-complete.
We won’t go into many details about data structures. Usually it will be
sufficient to assume that the graph is represented by adjacency lists, or an
adjacency matrix, or a mix of these. The adjacency matrix of a graph G =
(V, E) with n vertices is an n × n matrix. The rows and columns are indexed
by the vertices. A matrix entry with row x and column y is 1 if {x, y} ∈ E
and otherwise the entry is 0. The adjacency matrix needs O(n2 ) space. The
advantage of the adjacency matrix is that one can test in O(1) time if two
vertices are adjacent or not. Adjacency lists require only linear space, that
is O(n + m) space. A clever mix of the two data structures, which uses the
adjacency matrix without initializing it, makes it possible to test if two vertices
are adjacent in O(1) time while using only linear space.

Many other concepts will pass the revue. We will explain them when we
meet them, as clearly as we can. Let’s just get started. We hope the reader will
enjoy the choice of our topics and the way we present them in this book.
2
Exponential Algorithms

Let’s face it: lots of interesting problems on graphs are NP-complete. In this
chapter we have a look at exponential algorithms.

Let’s look at an example. Suppose we want to solve the maximum indepen-

dent set problem on some graph G = (V, E). An independent set is a subset
M ⊆ V of vertices such that no two vertices in M are adjacent. The maximum
independent set problem asks for an independent set M in G such that |M| is
maximal. An answer to the problem is called a ‘maximum’ independent set.
We denote the cardinality of a maximum independent set in G with α(G).

An easy way to solve the problem is as follows. First, make a list of all sub-
sets of vertices. Next, check which subsets are independent sets. Then count
the number of vertices in each independent set and take the largest one.

What is the time-complexity of this algorithm? Let n be the number of

vertices in the graph G. Obviously, there are 2n subsets of vertices. Let M
be a subset. We need to check if M is an independent set. We assume that
the graph G is represented by an adjacency matrix A. Then we can check if
two vertices are adjacent in constant time. Since M has O(n2 ) pairs we can
check if M is an independent set in O(n2 ) time. Thus our algorithm runs in
O(n2 · 2n ) time. In the next section we show that we can do better.

When we are dealing with exponential algorithm we don’t care so much

about the polynomial factors in the time-bound. The O∗ -notation neglects the
polynomial factors. So instead of O(n2 · 2n ) we write O∗ (2n ).

The O∗ -notation is pretty useful. For example, suppose our graph G is

represented by adjacency lists. That is, for each vertex x in G there is a linked
list L(x) of its neighbors. Now, to check if two vertices x and y are adjacent
we need to check if y appears in the list L(x) or not. In the worst case L(x)
contains n − 1 vertices, so checking if y is in this list takes more than constant
6 2 Exponential Algorithms

time. (Of course, we can construct the adjacency matrix A in O(n2 ) time and
then proceed as before.)

If we use the O∗ -notation then we don’t need to worry about such details.
Let p(n) be some polynomial, for example p(n) = 10 · n5 . Suppose we can
check if any subset M is an independent set in at most p(n) time. Then the
algorithm described above runs in O(p(n) · 2n ) = O∗ (2n ) time.

2.1 Independent set

An independent set M in a graph G = (V, E) is maximal if every vertex in
V \ M has at least one neighbor in M. Moon and Moser1 have shown that
any graph with n vertices has at most 3n/3 maximal independent sets. Notice
that a graph which is the union of n3 triangles achieves this bound.

There are algorithms that list all the maximal independent sets with poly-
nomial delay. That means that there exists some polynomial p(n) such that
the algorithm spends at most p(n) time before it generates the next (or the
first) independent set. For example, the algorithm of Tsukiyama, et al., takes
O(nm) time per maximal independent set, where n and m are the num-
ber of vertices and edges in the graph.2 These two results yield the following
theorem.

Theorem 2.1. There exists an O∗ (1.4422n ) algorithm that solves the maximum
independent set problem on a graph G, where n is the number of vertices in G.

Proof. We use an algorithm which lists all maximal independent sets in G in

O(p(n) · 3n/3 ) time for some polynomial p(n). Notice that

O(p(n) · 3n/3 ) = O∗ (3n/3 ) = O∗ (1.4422n ).

This proves the theorem. t

u

Of course, this algorithm is much better than the O∗ (2n ) algorithm that
we started with. In the rest of this section we show that we can still do a little
bit better.

1
J. W. Moon and L. Moser, On cliques in graphs, Israel Journal of Mathematics 3
(1965), pp. 23–28.
2
S. Tsukiyama, M. Ide, H. Ariyoshi and I. Shirakawa, A new algorithm for generating
all the maximal independent sets, SIAM Journal on Computing 6 (1977), pp. 505–
517.
 2.1 Independent set 7

Let x be a vertex of G. As usual we use the notation N(x) for the set of
neighbors of a vertex x. The degree of x is |N(x)|. We also use the notation
N[x] for the closed neighborhood of x, which is

N[x] = N(x) ∪ {x}.

Let x be a vertex in G. There are two types of independent sets, namely

those that contain x and those that do not contain x. Consider a maximum
independent set M. The next two lemmas show how to reduce the graph in
each of the two cases.

Lemma 2.2. Let M be a maximum independent set in G and let x ∈

/ M. Then
M is a maximum independent set in G − x.

Proof. The graph G − x is the subgraph of G induced by V \ {x}. Let M be a

maximum independent set in G and let x ∈/ M. Then M is an independent set
in G − x.
Of course, any independent set in G − x is also an independent set in G.
Thus G − x cannot have an independent set M0 with |M0 | > |M| since this
contradicts the assumption that M is a maximum independent set in G.
This proves the lemma. t
u

Lemma 2.3. Let M be a maximum independent set in G and let x ∈ M. Then

M = {x} or M \ {x} is a maximum independent set in G − N[x].

Proof. The graph H = G − N[x] is the subgraph of G induced by V \ N[x].

Let M be a maximum independent set in G and assume that x ∈ M. Notice
that, unless M = {x}, M \ {x} is an independent set in G − N[x].
Suppose that H has an independent set M0 which is larger than M \ {x}. Since
M0 is an independent set in G − N[x], M0 contains no neighbors of x. Thus
M0 ∪{x} is an independent set in G which is larger than M and this contradicts
the assumption.
This proves the lemma. t
u

In other words, Lemmas 2.2 and 2.3 show that, for any vertex x, a maxi-
mum independent set M can be derived from a maximimum independent set
in G − x or from a maximum independent set in G − N[x].

The algorithm builds a rooted binary tree T as follows. The root of T cor-
responds with the graph G. If the graph has only one vertex, then T consists
of the root only. Otherwise, choose a vertex x in G. The root has two children.
8 2 Exponential Algorithms
G

G-x G-N[x]

Fig. 2.1. This figure illustrates the search tree for maximum independent set.

The left child is the root of a binary tree which corresponds with the graph
G − x. The right child is the root of a binary tree which corresponds with the
graph G − N[x].

Our algorithm computes a maximum independent set in G as follows. If

G has only one vertex then α(G) = 1. Otherwise, the algorithm recursively
computes the maximum independent set in the left subtree and in the right
subtree. By the two lemmas above,
α(G) = max { α(G − x), 1 + α(G − N[x]) }. (2.1)

We need to make a remark here. If x is the only vertex in G then G − x is

not a graph since, by definition, a graph has at least one vertex. (The ‘empty
graph’ is the graph without any edge.) In Formula (2.1), if V = {x} then we
define α(G − x) = 0. A similar situation occurs when x is adjacent to all other
vertices, i.e., when N[x] = V. In that case we define α(G − N[x]) = 0.

For example, assume that the graph G is a clique. A clique is a subset C

of vertices such that every pair of vertices in C is adjacent. In other words, a
clique in the graph is an independent set in the complement Ḡ of the graph G
and vice versa.
Since G is a clique, any maximal independent set in G has only one vertex.
Thus G has exactly n maximal independent sets. In this case, every pair of
vertices in G is adjacent, thus N[x] = V for any vertex x. If x is the only vertex
in G then α(G − x) = 0 and otherwise α(G − x) = 1.

In order to obtain a good timebound we like to reduce the graph in at least

one of the two branches as much as possible. One branch only removes the
vertex x and this reduces the graph by one vertex. The other branch removes
|N[x]| vertices from the graph. To make this graph as small as possible we
choose the vertex x such that it has the largest degree.

An isolated vertex in G is a vertex without neighbors. A pendant vertex is a

vertex with exactly one neighbor. The next lemma deals with graphs in which
every vertex has degree at most two.
 2.1 Independent set 9

Lemma 2.4. Let G be a graph and assume that every vertex of G has degree at
most 2. Then a maximum independent set in G can be computed in linear time.
Proof. Since every vertex has degree at most 2, the graph is the union of a
collection of paths and cycles.
To compute the maximum independent set of G we can compute the maxi-
mum independent set in each of the (connected) components of G and add
them up. We leave it as an exercise to check the following claims, for example
by using (2.1).
The length of a path or cycle is the number of edges in the path or the cycle.

(1) If C is a cycle of length 2k then α(C) = k.

(2) If C is a cycle of length 2k + 1 then α(C) = k.
(3) If P is a path of length 2` > 0 then α(P) = ` + 1.
(4) If P is a path of length 2` + 1 > 0 then α(P) = ` + 1.
Thus, in order to compute α(G) it suffices to compute the lengths of the com-
ponents of G, and this can be done in linear time. t
u

We now change the algorithm a little bit, as follows. As long as there exists
a vertex x in the graph with degree at least three, then the algorithm chooses
such a vertex to grow the tree T . When, at some point, a reduced graph H
has no more vertices of degree more than two then the algorithm uses the
linear-time algorithm described in Lemma 2.4 to compute α(H).

Let T (n) be the worst-case timebound that the algorithm needs to com-
pute α(G) for a graph G with n vertices. Then, by the Formula (2.1) and by
Lemma 2.4 we have the following recurrence relation for T (n).

T (n) 6 T (n − 1) + T (n − 4) + O(n + m). (2.2)

Perhaps you wonder why we write T (n − 4). If x is a vertex of degree i then

the Formula (2.1) says T (n − i − 1) instead of T (n − 4). However, notice that
T (n) is a non-decreasing function (see Exercise 2.3), and since the degree of
x is at least three

i>3 ⇒ T (n − i − 1) 6 T (n − 4).

It is an easy exercise to check that T (n) 6 ωn · p(n), where p(n) is some

polynomial and ω is the largest real root of the equation

ω4 = ω3 + 1.

Some calculations (or a calculator) yield ω ≈ 1.3803. This proves the follow-
ing theorem.
10 2 Exponential Algorithms

Theorem 2.5. There exists an O∗ (1.3803n ) algorithm which solves the maxi-
mum independent set problem on a graph G, where n is the number of vertices
in G.

Remark 2.6. The technique that we used is sometimes called a ‘pruned search
tree technique.’ Pruning means that the search tree is cut short. In our case
the leaves of the search tree are graphs in which every vertex has degree at
most two. There are many ways to prune the search tree further (see, e.g.,
Exercise 2.2). It is one of the well-known techniques that are used in the
design of exponential algorithms.

Remark 2.7. Very often the timebound for an exponential algorithm follows
from a recurrence relation like Formula (2.2). If you want to study exact al-
gorithms it’s a good idea to obtain some basic knowledge about solving recur-
rence relations.

Remark 2.8. The highly-praised Bron-Kerbosch algorithm generates all max-

imal independent sets in O(3n/3 ) time. The Bron-Kerbosch algorithm is not
output-sensitive.3 Of course, for Theorem 2.1 this does not matter.

Remark 2.9. The best time-bound for an algorithm that solves the maximum
independent set problem seems to be Robson’s algorithm.4 This algorithm
uses exponential space. Robson claims that he improved his algorithm such
that, nowadays, it runs in O∗ (1.1844n ). This algorithm uses a computer-
generated case analysis with thousands of different cases. Jian designed an
algorithm for maximum independent set which runs in O(1.2346n ).5 This
algorithm uses only polynomial space.

2.2 Chromatic number

Definition 2.10. The chromatic number of a graph G is the minimal number
of colors needed to color the vertices of G such that no two adjacent vertices in G
have the same color.

Usually, one denotes the chromatic number of a graph G by χ(G).

The chromatic number problem is one of the most studied problems in

graph algorithms. The problem is NP-complete for graphs in general. Even
3
C. Bron and J. Kerbosch, Algorithm 457: finding all cliques of an undirected graph,
Communications of the ACM 16 (1973), pp. 575–577.
4
J. M. Robson, Algorithms for maximum independent sets, Journal of Algorithms 7
(1986), pp. 425–440.
5
T. Jian, An O(20.304n ) algorithm for solving maximum independent set problem,
IEEE Transactions on Computers 35 (1986), pp. 847–851.
 2.2 Chromatic number 11

the problem to color a planar graph with three colors is NP-complete. (By
the 4-color theorem every planar graph can be colored with four colors, but
checking if a planar graph can be colored with three colors is NP-complete.) In
this section we look at an exact algorithm for the chromatic number problem.

Consider a coloring of a graph G = (V, E) with k colors. Consider a set T of

vertices that all have the same color. Then, by definition, T is an independent
set. Notice also that we may assume that T is maximal. This can be seen as
follows. Suppose T is contained in a maximal independent set T 0 . Then we
can re-color the vertices of T 0 \ T such that all the vertices of T 0 have the same
color. Thus we obtain a coloring of G with k colors and now T 0 is maximal.
This proves that the chromatic number of G is determined by the following
formula.

χ(G) = min { 1 + χ(G − T ) | T is a maximal independent set in G }. (2.3)

When G is an independent set then V − T = ∅ and then G − T is not a graph.

In that case we define χ(G − T ) = 0.

Recall the result of Moon and Moser: every graph with n vertices has at
most 3n/3 maximal independent sets and these can be listed in O(p(n) · 3n/3 )
time, where p(n) = nc is some polynomial (in Tsukiyama’s algorithm c 6 3).

Our algorithm considers all possible subsets S of vertices in G and it colors

G[S]. Here, G[S] is the graph induced by S. The subsets are processed in order
of increasing cardinality. Thus, when S is considered by the algorithm, all
the subsets of S have already been colored. When S has ` vertices, then the
algorithm lists all the maximal independent sets in G[S] in O(p(`) · 3`/3 ) time.

By Formula (2.3) the algorithm computes χ(G[S]) in O(p(`) · 3`/3 ) time,

since all the values of
G[S] − T = G[S \ T ]
were determined earlier (because S\T ⊂ S), for all maximal independent sets
T in G[S].

When G has n vertices, then there are n` subsets S with ` vertices. This

shows that our algorithm has a running time proportional to

X
n  
n X
n  
n `/3
p(`) 3 `/3
6 p(n) 3 = p(n)(1 + 31/3 )n .
` `
`=0 `=0

An easy calculation shows that 1 + 31/3 ≈ 2.4422 and this proves the
following theorem.
12 2 Exponential Algorithms

Theorem 2.11. There exists an O∗ (2.4422n ) algorithm to compute the chro-

matic number of a graph G, where n is the number of vertices in G.

Remark 2.12. Notice that this algorithm uses exponential space.

Remark 2.13. Recently, some improvements were obtained by Björklund, et

al.6 They show that the chromatic number problem can be solved in O∗ (2n ).
We look at their method in Section 2.7.

2.2.1 Three-coloring

In this section we consider the problem whether χ(G) 6 3 for a graph G.

Lawler describes the following algorithm, which is a pruning of the search
tree.7

Consider a 3-coloring of G. Notice that the graph induced by vertices of

any two colors is bipartite. A graph is bipartite when it has a 2-coloring.

Lawler’s algorithm is very simple. Generate all maximal independent sets

in G and check if G − S is bipartite for some maximal independent set S. It is
easy to see that one can check if a graph is bipartite in linear time. Thus the
time needed to check if a graph G can be colored with three colors is simply
the time needed to list all the maximal independent sets, i.e., O∗ (1.4422n ).

Theorem 2.14. There exists an O∗ (1.4422n ) algorithm which checks if a graph

G with n vertices can be colored with three colors.

Remark 2.15. Until now, the best algorithm for solving the 3-coloring problem
is an algorithm by Eppstein.8 This algorithm runs in O∗ (1.3289n ).

6
A. Björklund, T. Husfeldt and M. Koivisto, Set partitioning via inclusion-exclusion,
SIAM Journal on Computing 39 (2009), pp. 546–563.
7
E. L. Lawler, A note on the complexity of the chromatic number problem, Informa-
tion Processing Letters 5 (1976), pp. 66–67.
8
D. Eppstein, Improved algorithms for 3-coloring, 3-edge-coloring, and constraint
satisfaction, Proceedings of the 12th ACM-SIAM Symposium on Discrete Algorithms
SIAM, 2001.
 2.3 Domatic partition 13

2.3 Domatic partition

Definition 2.16. A dominating set in a graph G = (V, E) is a set D ⊆ V of
vertices such that every vertex x ∈ V \ D has at least one neighbor in D.

In other words, a set D ⊆ V is a dominating set in G = (V, E) if and only if

N[x] ∩ D 6= ∅ for every vertex x ∈ V,

where N[x] is the closed neighborhood of x.

By definition, if G = (V, E) is a graph, then V 6= ∅. It follows that the

empty set is not a dominating set in G. On the other hand, for any graph
G = (V, E), the set V is a dominating set. Also, any maximal independent set
in G is a dominating set.

The dominating set problem asks for a dominating set of minimal cardi-
nality. Usually, one denotes the minimal cardinality of a dominating set in G
by γ(G).

Remark 2.17. It is easy to see that the dominating set problem can be solved
in O∗ (2n ) time, by simply testing every subset of V. In Section 2.6 we show
that the problem can be solved in O∗ (1.7088n ). In their recent book, Fomin
and Kratsch improve this. They show that the dominating set problem can be
solved in O∗ (1.5259n ) time.9

The dominating set problem is NP-complete. Notice that it is easy to check

whether a set D is a dominating set in G or not, namely, simply check if each
vertex x ∈ V \ D has a neighbor in D. Since there are 2n subsets of V, there
is an easy O∗ (2n ) algorithm which solves the dominating set problem.

Definition 2.18. A domatic partition of a graph G = (V, E) is a partition

{ D1 , . . . , Dk }

of V such that each Di is a dominating set in G.

Remark 2.19. A set {D1 , . . . , Dk } is a partition of a set V if

(i) for each 1 6 i 6 k, Di 6= ∅, and
(ii) for all 1 6 i < j 6 k, Di ∩ Dj = ∅, and
(iii) ∪ki=1 Di = V.
14 2 Exponential Algorithms

a a

f e f e

b d c b d c
Fig. 2.2. This figure shows two domatic partitions; on the left {{a, b, c}, {d, e, f}} and on
the right {{a, d}, {b, e}, {c, f}}. The one on the right is a solution for the domatic partition
problem.

The domatic partition problem asks for a domatic partition of the graph with
a maximal number of dominating sets.

The domatic partition problem is NP-complete. In this section we show

that the domatic partition problem can be solved in O∗ (3n ) time on graphs
with n vertices.

Let G = (V, E) be a graph and let X ⊆ V be some subset of vertices. An

X-partition is a partition
{ D1 , . . . , D` }
of X such that each Di is a dominating set in G. Notice that there can be an
X-partition only if X is a dominating set in G!

If X is a dominating set in G then let f(X) denote the maximal ` for which
there is an X-partition into ` dominating sets. If X is not a dominating set then
let f(X) = 0. Notice that f(V) solves the domatic partition problem.

Our algorithm computes the maximal ` for which there is an X-partition

into ` dominating sets for each X ⊆ V.

Consider subsets X of V in order of increasing cardinality. Let X be a subset

of cardinality k. We assume that for all subsets X0 with |X0 | < |X| the algorithm
has determined f(X0 ). By definition, if X is not a dominating set then f(X) = 0.
Otherwise,

f(X) = max { 1 + f(X \ Y) | Y ⊆ X and Y is a dominating set in G }. (2.4)

9
Corollary 6.11 in: F. Fomin and D. Kratsch, Exact exponential algorithms, Springer,
2010.
 2.4 The traveling salesman problem 15

This simple observation gives us the following theorem.

Theorem 2.20. There exists an O∗ (3n ) algorithm which solves the domatic par-
tition problem for graphs with n vertices.
Proof. The timebound is upperbounded by
Xn  X
n
k  
k Xn  
n k
= 2 = 3n .
k ` k
k=0 `=0 k=0

t
u

Remark 2.21. Notice that the algorithm described above uses exponential
space. The result can be improved to O∗ (2.8718n ) time and polynomial
space.10

2.4 The traveling salesman problem

Suppose there are n cities, numbered 1, . . . , n. A traveling salesman has to
visit the cities 1, . . . , n. He starts in city 1 and travels through the remaining
cities in arbitrary order and at the end he returns to city 1. The distance
between cities i and j is denoted by d(i, j). The problem is to minimize the
total travel length.

The traveling salesman problem is NP-hard.

For each subset S ⊆ {2, . . . , n} and for each i ∈ S let f(S, i) denote the
length of a shortest tour that starts in city 1, then visits the cities in S \ {i} in
some order and then stops in city i.

We now have

d(1, i) if S = {i} and
f(S, i) = (2.5)
min { f(S \ {i}, j) + d(j, i) | j ∈ S \ {i} } otherwise.

The following theorem was proved by Held and Karp, and, independently,
by Bellman.11 12

10
Proposition 8 in: A. Björklund, T. Husfeldt and M. Koivisto, Set partitioning via
inclusion-exclusion, SIAM Journal on Computing 39 (2009), pp. 546–563.
11
R. Bellman, Dynamic programming treatment of the traveling salesman problem,
Journal of the Association for Computing Machinery 9 (1962), pp. 61–63.
12
M. Held and R. Karp, A dynamic programming approach to sequencing problems,
Journal of the Society for Industrial and Applied Mathematics 10 (1962), pp. 196–
210.
16 2 Exponential Algorithms

Theorem 2.22. There exists an O∗ (2n ) algorithm which computes the mini-
mum length of a traveling salesman tour.

Proof. The algorithm processes the subsets S ⊆ {2, . . . , n} in order of increas-

ing cardinality.
The values f(S, i) can be computed by Formula (2.5) in O(n) time. Notice that
the solution to the traveling salesman problem is given by the value of

min { f({2, . . . , n}, j) + d(j, 1) | 2 6 j 6 n }. (2.6)

There are O(n · 2n ) pairs (S, i) where S is a subset of {2, . . . , n} and i ∈ S. The
computation of each f(S, i) takes O(n) time and so the overall complexity is
bounded by O(n2 · 2n ) = O∗ (2n ) time. t
u

Remark 2.23. As far as we know, there is no algorithm that solves the traveling
salesman problem in time O∗ (cn ) for c < 2.

2.5 Set cover

Definition 2.24. Let U be a finite, nonempty set and let n = |U|. Let

S = { S1 , . . . , Sm }

be a collection of nonempty subsets of U. A subset S0 ⊆ S is a cover for U if

[
S = U.
S∈S0

The set cover problem asks for a cover S0 such that |S0 | is minimal.

Of course, a cover can exist only if every element of U is an element of at

least one subset Si .

Consider the sets

Si = {S1 , . . . , Si } for i = 1, . . . , m.

Let U0 ⊆ U and let f(U0 , i) for i > 1 be the minimal cardinality of a cover
for U0 with subsets from Si . Let f(∅, i) = 0 and
[
if U0 6= ∅ and U0 6⊆ S then define f(U0 , i) = ∞.
S∈Si
 2.6 Dominating set 17

The values f(U0 , i), for ∅ 6= U0 ⊆ U and i = 1, . . . , m, can be computed

via the following formulas. First consider the case i = 1. Then

1 if U0 ⊆ S1
f(U0 , 1) = (2.7)
∞ otherwise.

The value of f(U0 , i + 1) follows from the following formula.


0 ∞ if U0 6⊆ ∪S∈Si+1 S
f(U , i + 1) = (2.8)
min { f(U0 , i), 1 + f(U0 \ Si+1 , i) } otherwise.

Here is the proof that Formula (2.8) is correct. Either the set Si+1 is used
to cover U0 and then U0 \ Si+1 has to be covered with sets from Si , or the set
Si+1 is not used in the cover, and then U0 has to be covered with elements
from the set Si .

Theorem 2.25. There exists an O∗ (m · 2n ) algorithm for the set cover problem.

Proof. The algorithm processes the subsets U0 ⊆ U in order of increasing car-

dinality. For each U0 the algorithm computes the values f(U0 , i) for increasing
i = 1, . . . , m via the Formulas (2.7) and (2.8).
There are O(m · 2n ) pairs (U0 , i). First consider Formula (2.7). It can be
checked in O(n) time if U0 is covered by S1 or not. Therefore, the values
f(U0 , 1) can be determined in O(n · 2n ) time.
Now consider the computation of f(U0 , i + 1) via Formula (2.8). In this case
the algorithm needs to compute U0 \ Si+1 , which takes O(n) time. Thus the
total time complexity is bounded by O(nm · 2n ) = O∗ (m · 2n ).
This proves the theorem. t
u

2.6 Dominating set

Recall Definition 2.16. Let G = (V, E) be a graph. A dominating set for G is a

set D ⊆ V such that every vertex x ∈ V \ D has at least one neighbor in D.

The dominating set problem asks for a dominating set of minimal cardi-
nality. Recall that we use γ(G) to denote the cardinality of a minimum domi-
nating set. Thus,

γ(G) = min { |D| | D ⊆ V(G) and ∀x∈V(G) N[x] ∩ D 6= ∅ }.

18 2 Exponential Algorithms

The dominating set problem can be reduced to the set cover problem as
follows. Let U = V and let

S = { N[x] | x ∈ V }.

A solution for this set cover problem corresponds to a solution for the domi-
nating set problem. This can be seen as follows.

Assume that
S0 = { N[x] | x ∈ V 0 },
is a solution for the set cover problem, for some V 0 ⊆ V. Then V 0 is a domi-
nating set in G because every vertex y ∈ V \ V 0 is covered by some set in S0
and so y ∈ N[z] for some z ∈ V 0 .

To see the converse, let D be a dominating set in G. Then let

S0 = { N[x] | x ∈ D }.

Since every vertex y ∈ V \ D has a neighbor in D, there exists some z ∈ D

such that y ∈ N[z]. Thus S0 covers V.

Let I be a maximal independent set and let W = V \ I. Our algorithm

finds, for every subset D0 ⊆ W, an extension E(D0 ) ⊆ I such that D0 ∪ E(D0 ) is
a dominating set and such that |E(D0 )| is minimal.

Notice that, when we have an extension for every D0 ⊆ W, then this solves
the domination problem, namely by taking the set D0 ⊆ W which minimizes

|D0 ∪ E(D0 )| = |D0 | + |E(D0 )|. (2.9)

Let D0 ⊆ W and define

I(D0 ) = { s ∈ I | s has no neighbor in D0 }. (2.10)

Then any extension of D0 contains I(D0 ).

The vertices that have no neighbors in D0 ∪ I(D0 ) are the vertices of

U0 = { x ∈ W | N[x] ∩ (D0 ∪ I(D0 )) = ∅ }. (2.11)

To obtain a minimum extension of D0 we need to add a minimum set S of

vertices from I \ I(D0 ) such that every vertex of U0 has a neighbor in S.

We now show how to compute minimum extensions.

 2.6 Dominating set 19

Order the vertices of I, say

I = { s1 , . . . , st } and define, for ` ∈ {1, . . . , t}, I` = { s1 , . . . , s` }.
(2.12)

For each U0 ⊆ W and ` ∈ {1, . . . , t}, define

f(U0 , `) ⊆ I`
as a subset of I` such that
(i) every vertex of U0 has a neighbor in f(U0 , `) and
(ii) |f(U0 , `)| is minimal.
We obtain formulas similar to (2.7) and (2.8). When U0 = ∅ then define
f(U0 , `) = ∅ for all ` ∈ {1, . . . , t}.

Otherwise, when U0 6= ∅, first consider the case ` = 1.


0 {x1 } if ∅ 6= U0 ⊆ N(s1 ), and
f(U , 1) = (2.13)
I otherwise.

Assume that U0 6= ∅. Then we obtain the following recurrence relation for

f(U0 , ` + 1).


0 f(U0 , `) if |f(U0 , `)| < 1 + |f(U0 \ N(s`+1 ), `)|
f(U , `+1) =
f(U0 \ N(s`+1 ), `) ∪ {s`+1 } otherwise.
(2.14)

According to Theorem 2.25 the sets f(U0 , `) for U0 ⊆ W and ` = 1, . . . , t

can be computed in O∗ (2|W| ) time.

Now let D0 ⊆ W. Then the cardinality of an extension E(D0 ) of D0 is

|E(D0 )| = |I(D0 ) ∪ f(U0 , t)| = |I(D0 )| + |f(U0 , t)|, (2.15)
where t = |I| and U0 is defined by Formula (2.11).

It follows that the time for solving the dominating set problem is bounded
by some polynomial factor times
|W| |W| 
X X X |W|

|W| |W|
2 + 1=2 + = 2|W|+1 . (2.16)
k
k=0 D0 ⊆W k=0
|D |=k
0

We now easily obtain the following theorem.

20 2 Exponential Algorithms

Theorem 2.26. There exists an O∗ (1.7088n ) algorithm which solves the domi-
nating set problem on graphs with n vertices.
Proof. Let β = 0.2271. We consider two cases.
In the first case, assume that |I| > β · n. Then, according to Formula (2.16), a
minimum dominating set can be computed in
O∗ (2(1−β)n ) = O∗ (1.7088n ). (2.17)

Now assume that |I| 6 βn. Then γ(G) 6 |I| 6 βn. In that case we test every
subset of V with cardinality at most βn.
By Stirling’s formula one can obtain that, since 0 < β < 21 ,
X
bβnc 
n

6 2h(β)n where h(β) = −β log2 β − (1 − β) log2 (1 − β). (2.18)
i
i=0

The function h(β) is the binary entropy function. (See Figure 2.3.) Some cal-
culations show that 2h(β)n 6 1.7088n .

h(x)= -xlog2x-(1-x)log2(1-x)

0.2271

Fig. 2.3. This figure shows the binary entropy function; the value β ≈ 0.2270922 is
where h(β) = 1 − β.

This proves the theorem. t

u

Remark 2.27. Notice that we chose β such that h(β) = 1 − β to obtain the
best bound (see Figure 2.3). If the maximal independent set I is small then
the domination number γ(G) 6 |I| is small as well. In that case it is better to
try all subsets of cardinality at most |I|. This gives us the bound in terms of
the entropy function (2.18). When |I| is large, then the ‘universe’ W = V \ I is
small. Then it is better to use the set cover method. By the choice of β we get
the same timebound in both cases.
 2.7 Inclusion – exclusion 21

2.6.1 Dominating set in bipartite graphs

Theorem 2.26 produces a nice timebound for the dominating set problem on
bipartite graphs.13

First recall the definition of a bipartite graph.

Definition 2.28. A graph G is bipartite if χ(G) 6 2.

Theorem 2.29. There exists an O∗ (2n/2 ) algorithm that computes a minimum

dominating set on bipartite graphs with n vertices.
Proof. Let G = (V, E) be a bipartite graphs with n vertices. We claim that
α(G) > n2 , where α(G) is the maximal cardinality of an independent set
in G. To see this, consider a coloring with at most two colors. A color class
is a subset of the vertices that have the same color. The coloring partitions
the vertices into at most two color classes, and so at least one color class
has at least n2 vertices. In Exercise 2.4 we ask you to design a linear-time
algorithm to recognize bipartite graphs. This recognition algorithm produces
an independent set of size at least n2 vertices in linear time. (Notice that it does
not compute α(G) in linear time! The best way to compute α(G) in bipartite
graphs is via Edmonds’ matching algorithm.14 )
Our algorithm now runs the algorithm as described in the first case in the
proof of Theorem 2.26. We can take β = 12 in this case because the indepen-
dent set has cardinality at least n2 . According to Formula (2.16) on Page 19, a
minimum dominating set can be computed in
O∗ (2(1−β)n ) = O∗ (2n/2 ) = O∗ (1.4143n ).

This proves the theorem. t

u

2.7 Inclusion – exclusion

Good timebounds for many set covering problems can be obtained via the
method of inclusion-exclusion. The method was developed by Björklund and
Husfeldt and, independently, by Koivisto.15 We illustrate their method by
graph coloring.

We start with the inclusion-exclusion formula.

13
M. Liedloff, Finding a dominating set on bipartite graphs, Information Processing
Letters 107 (2008), pp. 154–157.
14
J. Edmonds, Paths, trees and flowers, Canadian Journal of Mathematics 17 (1965),
pp. 449–467.
15
A. Björklund, T. Husfeldt and M. Koivisto, Set partitioning via inclusion-exclusion,
SIAM Journal on Computing 39 (2009), pp. 546–563.
22 2 Exponential Algorithms

Lemma 2.30. Let

S = { A1 , A2 , . . . , At } (2.19)
be a collection of subsets of some finite set U. Then
t
[ X
t X
| Ai | = (−1)k−1 |Ai1 ∩ · · · ∩ Aik |. (2.20)
i=1 k=1 16i1 <i2 <···<ik 6t

Proof. Let A = ∪ti=1 Ai .

Define the characteristic functions f : U → {0, 1} and fi : U → {0, 1}, for
i ∈ {1, . . . , t}, as follows.
 
1 if x ∈ A, 1 if x ∈ Ai , and
f(x) = and ∀i∈{1,...,t} fi (x) =
0 if x ∈ U \ A, 0 if x ∈ U \ Ai .
(2.21)
Then we have the following identity.

Y
t
∀x∈U (f(x) − fi (x)) = 0. (2.22)
i=1

To see this, first assume that x ∈ A. Then x ∈ Ai for some i ∈ {1, . . . , t}. Then
the ith term in the product f(x) − fi (x) = 1 − 1 = 0 and so the whole product
is zero.
Now assume that x ∈ / Ai for any i ∈ {1, . . . , t}, and so every term
/ A. Then x ∈
in the product f(x) − fi (x) = 0.
Write this product as follows.

X
t X
(−1)k · f(x)t−k · fi1 (x) · fi2 (x) · · · fik (x) = 0. (2.23)
k=0 16i1 <i2 <···<ik 6t

This implies

X
t X
t
f(x) = (−1)k−1 fi1 (x) · · · fik (x). (2.24)
k=1 16i1 <···<ik 6t

Sum Equation (2.24) over all elements x ∈ U. This gives

X X
t X
f(x)t = |A| = (−1)k−1 |Ai1 ∩ · · · ∩ Aik |. (2.25)
x∈U k=1 16i1 <···<ik 6t

This proves the lemma. t

u
 2.7 Inclusion – exclusion 23

A similar formula can be obtained for the number of elements that lie in
none of the Ai . We give a different proof.

Lemma 2.31. Let S = {A1 , . . . , At } be a collection of subsets of a finite set U.

Use the convention that
∩i∈∅ Ai = U. (2.26)

Then
t X
(−1)|I| | ∩i∈I Ai |.
\
| Ai | = (2.27)
i=1 I⊆{1,...,t}

Proof. Give each element x ∈ U a weight, say w(x), and sum the weights
over the sets on the left and right hand side of (2.27). Consider x ∈ U. If
x lies in none of the Ai then it contributes w(x) on the left hand side and
it also contributes w(x) on the right hand side, namely for I = ∅ (by the
convention).
Now assume that x lies in Ai for i ∈ I, and assume that I 6= ∅. Then it is
counted zero times on the left. On the right hand side it is counted for all
I0 ⊆ I. Its total contribution is zero because, by the binomial theorem,
|I|  
X X |I|
|I0 |
(−1) w(x) = w(x) · (−1)k = 0 because I 6= ∅. (2.28)
k
I0 ⊆I k=0

t
u

Definition 2.32. Let G = (V, E) be a graph and let S be some collection of

subsets of V. A k-cover is a k-tuple over S

each Si ∈ S, and
(S1 , . . . , Sk ) such that (2.29)
∪ki=1 Si = V.

Notice that some of the sets in a k-tuple may be equal.

Denote the number of k-covers by ck .

Observe the following. When S is the set of all independent sets in G then

ck > 0 if and only if χ(G) 6 k.

24 2 Exponential Algorithms

Definition 2.33. For X ⊆ V, define a(X) as the number of sets in S that avoid
X, that is,
a(X) = | { S ∈ S | S ∩ X = ∅ } |. (2.30)

Lemma 2.34. Let G = (V, E) be a graph and let S be a collection of subsets of

V. Let ck be the number of k-covers. For X ⊆ V, let a(X) be the number of sets
in S that avoid X. Then
X
ck = (−1)|X| a(X)k . (2.31)
X⊆V

Proof. Let V = {1, . . . , n}. We use Lemma 2.31.

For the universe U we take the set of all k-tuples,

U = { (S1 , . . . , Sk ) | Si ∈ S }. (2.32)

For the subset Ai ⊆ U, i ∈ V, we take the set of k-tuples that avoid {i}, that
is,
k
[
Ai = { (S1 , . . . , Sk ) | i ∈
/ S` }. (2.33)
`=1

The number of k-covers is exactly the number of k-tuples that lie in none of
the Ai . Thus ck is the left hand side of (2.27). For X ⊆ V, ∩i∈X Ai contains
those k-tuples that avoid all elements of X. This number is of course a(X)k .
Thus \
| Ai | = a(X)k . (2.34)
i∈X

This proves the lemma. t

u

In the following theorem we illustrate the inclusion – exclusion method.

Theorem 2.35. Let G = (V, E) be a graph with n vertices. There exists an

O∗ (2n ) algorithm that computes χ(G).

Proof. Let G = (V, E) be a graph with n vertices. Let V = {1, . . . , n}. Let
S be the collection of all nonempty, independent sets in G, not necessarily
maximal.
By Formula (2.31) it is sufficient to compute a(X), for all subsets X ⊆ V.
Of course, we have that a(V) = 0. Assume that X 6= V. Fix some v ∈ V \ X.
We claim that
 2.7 Inclusion – exclusion 25

a(X) = a(X ∪ {v}) + a(X ∪ N[v]) + 1. (2.35)

To prove (2.35) we partition the sets S ∈ S that avoid X into two types, the
sets S that contain v and the sets S that do not contain v.
First consider sets S that avoid X and that do not contain v. Those sets S are
counted in a(X ∪ {v}).
Consider sets S that avoid X and that contain v. Since S is an independent set
which contains v, S contains no neighbors of v. Thus

S \ {v} ∩ (X ∪ N[v]) = ∅. (2.36)

Either S = {v}, which accounts for the +1-term, or S \ {v} 6= ∅. By (2.36), the
number of sets S with S \ {v} 6= ∅ is exactly a(X ∪ N[v]). This proves (2.35).
The recurrence relation (2.35) shows that the numbers a(X) can be computed
in O(n2n ) = O∗ (2n ) time. This proves the theorem. t
u

Remark 2.36. The algorithm of Theorem 2.35 uses exponential space. The pa-
per of Björklund, et al., shows that the chromatic number of a graph can be
computed in O∗ (2.2461n ) time and polynomial space.

2.7.1 Triangle partition

As an example of the inclusion-exclusion method, let’s have a look at the tri-

angle partition problem. Let G = (V, E) be a graph with n vertices. A triangle
in G is a clique with three vertices. A triangle partition of G is a collection
of vertex-disjoint triangles that partition V. Obviously, G can have a triangle
partition only if
n
is integer.
3

The triangle partition problem is NP-complete (even for graphs with max-
imal degree four).16

Theorem 2.37. There exists an O∗ (2n ) time and polynomial space algorithm
that checks if G has a triangle partition.

16
J. van Rooij, M. van Kooten Niekerk and H. Bodlaender, Partitioning sparse graphs
into triangles – relations to exact satisfiability and very fast exponential time algo-
rithms. Technical Report UU-CS-2010-005, Utrecht University, 2010.
26 2 Exponential Algorithms

Proof. We use the inclusion-exclusion method.

Let G be a graph with n vertices. Let S be the set of all triangles in G. Then,
obviously,
|S| = O(n3 ). (2.37)

A k-cover is a k-tuple of S:

each Si is a triangle in G, and
(S1 , . . . , Sk ) such that (2.38)
∪ki=1 Si = V.

Let ck be the number of k-covers.

We claim that G has a triangle partition if and only if
n
q= is integer, and cq > 0. (2.39)
3
To see this, notice that G has a q-cover if and only if G has a triangle partition.
By Lemma 2.34 on page 24
X
ck = (−1)|X| a(X)k , (2.40)
X⊆V

where a(X) is the number of triangles that avoid X, that is, the number of
triangles contained in V \ X.
Notice that we can easily count the number of triangles that are contained in
V \ X in polynomial time. Thus the numbers a(X) can be computed in O∗ (2n )
time and polynomial space.
To compute a(X)q we need O(log q) multiplications and we can represent
this number in O(q log n) bits. This shows that we can compute the For-
mula (2.40) in O∗ (2n ) time and polynomial space. t
u

Remark 2.38. There is an O∗ (1.7693n ) algorithm that solves the triangle par-
tition problem.17

17
M. Koivisto, Partitioning into sets of bounded cardinality, Proceedings of the 4th
International Workshop on Parameterized and Exact Computations, IWPEC 2009,
Springer-Verlag, LNCS 5917 (2009), pp. 258–263.
 2.7 Inclusion – exclusion 27

2.7.2 An improved algorithm for triangle partition

Consider the problem of partitioning the vertices of a graph G = (V, E) into

triangles. Koivisto’s algorithm, mentioned in the previous section, reaches a
timebound of O∗ (1.7693n ) for partitioning V into subsets of cardinality at
most three. In this section we illustrate Koivisto’s method for the triangle par-
tition problem. The improvement follows from the idea to consider triangle
partitions of which the elements are in lexicographic order.

Write q = n3 and let V = {1, . . . , n}. Order the triangles of G into lexico-
graphic order. Then the first triangle in any partition must contain the first
vertex, which is 1. In general, we have the following lemma.
Lemma 2.39. Consider the lexicograph ordering of the triangles in a triangle
partition P of G. Then, for j ∈ {1, . . . , q}, the first j triangles in P must contain
the vertices of {1, . . . , j}.
Proof. Since the triangles of P partition V each element of V is in some trian-
gle of P. Since the triangles are lexicographically ordered, the first j triangles
of P must contain all vertices of {1, . . . , j}. t
u

Let F be the set of all triangles in G. Define

R1 = { T | T ∈ F and 1 ∈ T },
and, recursively for k > 1,
Rk = { Y ∪ X | Y ∈ Rk−1 , X ∈ F, Y ∩ X = ∅ and min { i | i ∈ V \ Y } ∈ X }.

Lemma 2.40. Let P = (T1 , . . . , Tq ) be a triangle partition of G, where the tri-

angles of P are in lexicographic order. Then, for j ∈ {1, . . . , q},
j
[
Tk ∈ R j .
k=1

Proof. The first triangle of P must contain the vertex 1. Thus T1 ∈ R1 . We

proceed by induction. Let
j−1
[
Y= T` .
`=1

Then the j triangle must contain the smallest element of V \ Y. Thus

th

j
[
T` ∈ R j .
`=1

t
u
28 2 Exponential Algorithms

Lemma 2.41.
X
q
|Rk | = O∗ (1.7549n ).
k=1

Proof. By Lemma 2.39, the first j triangles must contain {1, . . . , j}. The rest of
the vertices of these j triangles are in {j + 1, . . . , n}. Thus,
Xq Xq 
n−k

|Rk | 6 .
2k
k=1 k=1

The binomial coefficients can be bounded as follows. Write

2α
k = αn and β = . (2.41)
1−α
First consider the tails. When β 6 13 then α 6 17 and we have
 
n−k
6 22n/7 6 1.22n .
2k
2
When β > 3 then α > 14 and we find
     
n−k n−k n
= 6 6 2n/4 6 1.2n .
2k n − 3k n − 3k

Consider the case where 13 < β < 23 , that is, 17 < α < 41 . Write β = 1+ 2 .
Then − 13 <  < 13 . With Stirling’s formula we obtain (neglecting polynomial
factors)
   
n−k 2n
∼ exp ( 2 ln(2) − (1 − )ln(1 − ) − (1 + ) ln(1 + ) ) .
2k 5+
(2.42)
Now write γ = 1− 1 2
2 . Then 3 < γ < 3 and (2.42) becomes
   
n−k 2n 1 − 3α k
∼ exp f(γ) where γ = and α = (2.43)
2k 3−γ 1−α n
and where f(γ) = −γ ln(γ) − (1 − γ) ln(1 − γ) is the entropy function. (See
Figure 2.4.)
Define
1
g(γ) = · f(γ). (2.44)
3−γ
The function g(γ) has its maximum at γ 6 0.56985 and with this value (2.43)
becomes  
n−k
= O∗ (1.7549n ). (2.45)
2k
t
u
 2.7 Inclusion – exclusion 29

2
Fig. 2.4. This figure shows the function 3−x
f(x) where f(x) is the entropy function.

We show that there is a dynamic programming algorithm which runs in

time proportional to
Xq
|Rk |.
k=1

Theorem 2.42. There exists an O∗ (1.7549n ) algorithm that checks if a graph

G with n vertices has a triangle partition.

Proof. Define, for S ⊆ V,


1 if S ∈ R1 , and
f1 (S) =
0 otherwise.

For j > 1, define

X
fj (S) = fj−1 (Y) · [ S \ Y ∈ F ] · [ min { x | x ∈ V \ Y } ∈ S \ Y ]
Y⊆S

where we write 
1 if P is true, and
[P]=
0 otherwise.
Thus fk (S) counts the number of ordered partitions of S into k triangles. No-
tice that
fk (S) = 0 if S ∈/ Rk .
30 2 Exponential Algorithms

By Lemma 2.41, the time to compute fk (S) is therefore proportional to

X
q
|Rk | · n3 = O∗ (1.7549n ).
k=1

t
u

Remark 2.43. The O∗ (2n ) algorithm of section 2.7.1 on page 25 uses polyno-
mial space. The algorithm of Theorem 2.42 uses exponential space.

Remark 2.44. For the partitioning into general (constant) subsets of size r this
method gives
   
n−k n(r − 1) 1 − rk/n
∼ exp · f(x) where x = (2.46)
(r − 1)k r−x 1 − k/n

and where f(x) = −xln(x) − (1 − x)ln(1 − x). To maximize one would need
to solve
(1 − x)r−1 = xr .
1+
When we write x = 2 we get a first approximation

1 3r − 1
≈ and x≈ .
3r − 2 2(3r − 2)

For r = 3 this gives x ≈ 4/7 ≈ 0.571 while the actual value, which we found
in the proof of Lemma 2.41, is x ≈ 0.5699.

For large r, x ≈ 1/2, f(x) ≈ ln(2) and

n2
 
n−k 1 n ln(2)
∼ 22n(r−1)/(2r−1) = 2n(1− 2r ) (1 − 2
+ O( 3 )). (2.47)
(r − 1)k 4r r

2.8 Subset sum

Let’s do an easy one to finish this chapter off.

Let a1 , . . . , an be n natural numbers, thus each ai is a positive integer.

Let also K be a natural number.
P The subset sum problem asks if there is a set
S ⊆ {1, . . . , n} such that i∈S ai = K.

The subset sum problem is NP-complete.

 2.9 Problems 31

Theorem 2.45. There exists an O(K · n) algorithm which solves the subset sum
problem.

Proof. Define
Si = { 1, . . . , i } for i = 1, . . . , n.

For i = 1, . . . , n, our algorithm computes a set

X
Ωi = { 0 6 t 6 K | there exists some I ⊆ Si such that ai = t }.
i∈I

Thus Ωi contains the sums that can be made with numbers from {a1 , . . . , ai }
and which are at most K.
First consider i = 1. Then,

{ 0, a1 } if a1 6 K, and
Ω1 = (2.48)
{0} if a1 > K.

Next, the set Ωi+1 can be computed via the following formula.

Ωi+1 = Ωi ∪ { 0 6 t 6 K | there exists a t0 ∈ Ωi such that t0 + ai+1 = t }.

(2.49)
This can be seen as follows. Let t0 be a sum which can be made with numbers
from {a1 , . . . , ai }. Then t0 and t0 + ai+1 can be made with numbers from
{a1 , . . . , ai+1 }.
To see the converse, let t be a sum that can be made with numbers from
{a1 , . . . , ai+1 }. If ai+1 is used to obtain the sum t, then t0 = t − ai+1 can be
made with numbers from {a1 , . . . , ai }. If ai+1 is not used to obtain the sum t
then t is a sum with numbers from {a1 , . . . , ai }.
It is easy to see that Ωn can be computed via Formulas (2.48) and (2.49) in
O(K · n) time. The answer to the subset problem is YES if K ∈ Ωn , and NO
otherwise. t
u

2.9 Problems
2.1. In the maximum independent set algorithm we used a linear-time algo-
rithm for the case where every vertex in G has degree at most two. Suppose
that, instead, we keep branching until no remaining vertex has any neigh-
bor, that is, until the graph is an independent set. Show that this changes
Formula (2.2) on Page 9 into

T (n) 6 T (n − 1) + T (n − 2) + O(n + m). (2.50)

32 2 Exponential Algorithms

Show that the solution of Formula (2.50) is the nth Fibonacci number times
some polynomial. Use the exact formula for the Fibonacci numbers to show
that this gives
T (n) = O∗ (1.6181n ).

Now assume that we branch until every vertex has degree at most one. This
changes Formula (2.50) into

T (n) 6 T (n − 1) + T (n − 3) + O(n + m). (2.51)

Show that Formula (2.51) gives a characteristic equation

ω3 = ω2 + 1. (2.52)

This gives T (n) = O∗ (1.4656n ).

2.2. Let x and y be two vertices of a graph G and assume that

N[x] ⊆ N[y].

Notice that this implies that x and y are adjacent. Show that

α(G) = α(G − y).

This is one of the prunings that is used in the algorithm of Fomin and Kratsch.
2.3. Let T (n) be the worst-case time-bound for any algorithm which solves the
maximum independent set problem on graphs with n vertices. In this exercise
we show that
T (n − 1) 6 T (n) for all n > 1.

Suppose that T (n − 1) > T (n) for some n > 1. We obtain a contradiction

as follows.
(a) Let G be a graph with n − 1 vertices. Let G0 be the graph obtained from G
by adding an isolated vertex to G. Prove that

α(G0 ) = α(G) + 1.

(b) Show that this proves that the maximum independent set problem for
G can be determined in T (n) time. This contradicts the assumption that
T (n − 1) > T (n).
2.4. Design a linear-time algorithm which checks if a graph is bipartite.
2.5. Design an exact algorithm for 4-coloring.
Hint: The 4-coloring problem can be solved in O∗ (1.7504n ).18
18
J. M. Byskov, Enumerating maximal independent sets with applications to graph
colouring, Operations Research Letters 32 (2004), pp. 547–556.
 2.9 Problems 33

2.6. Check Formula (2.4) on Page 14.

2.7. Check Formula (2.5) on Page 15.

2.8. Let G = (V, E) be a graph.

(a) A Hamiltonian cycle is a cycle in G which contains all vertices. Show that
there is an O∗ (2n ) algorithm which solves the Hamiltonian cycle problem.
Hint: Reduce the Hamiltonian cycle problem to the traveling salesman
problem. For any two vertices i and j in G define

1 if {i, j} ∈ E, and
d(i, j) = (2.53)
∞ if {i, j} ∈
/ E.

(b) A Hamiltonian path is a path in G which contains all vertices. The differ-
ence with the Hamiltonian cycle problem is that the two endpoints of the
path are not necessarily adjacent. Design an O∗ (2n ) algorithm that solves
the Hamiltonian path problem on graphs with n vertices.

2.9. Check the Formulas (2.13) and (2.14) and show that the sets f(X, `), as
defined by these formulas, can be computed in O∗ (2|W| ).

2.10. Design an O∗ (2n ) algorithm that solves the domatic partition problem.
Hint: Use the inclusion-exclusion method described in Section 2.7. You need
to derive a recurrence relation similar to (2.35) on Page 25.

2.11. Design an O∗ (2n ) algorithm that solves the triangle packing problem.
The triangle packing problem asks for the maximal number of vertex-disjoint
triangles in a graph.
Hint: Use the inclusion-exclusion method as in Section 2.7.1. For the set S
take the set of all triangles and all subsets that contain one vertex. Show that
there is a triangle packing with k triangles if and only if cn−2k > 0.

2.12. The subset sum problem described in Section 2.8 is NP-complete. The-
orem 2.45 on Page 31 shows that it can be solved in O(K · n) time. Does this
prove that P = NP?
3
Graph Classes

In this chapter we have a close look at a few important graph classes and we
look at some NP-complete problems that become polynomial when the graphs
are restricted to these.

A graph class, or a class of graphs, is simply a set of graphs. For example

G = { G | G is a planar graph } (3.1)

is the class of all planar graphs. A class of graphs may be finite or infinite. The
class above, of all planar graphs, is of course infinite (it contains an infinite
number of elements).

Obviously, many NP-complete problems can become polynomial when one

restricts the graphs to some special graph class. For example, the four-coloring
problem is NP-complete but it can be solved trivially when one restricts the
graphs to the class of planar graphs.

For algorithmic problems one considers usually only infinite classes of

graphs. The reason is that for finite classes of graphs most problems can be
solved in constant time by exhaustive search.

Usually, one restricts the research on infinite classes of graphs to classes

that are hereditary. A class G of graphs is hereditary if G ∈ G implies that every
induced subgraph of G is also in G. For example, the class of planar graphs is
hereditary, since, if G is planar then so is every induced subgraph of G. All the
classes that we study in this chapter are hereditary.

When one studies some graph class G then the membership of graphs in
G is an important issue. One refers to this as the recognition problem for the
class G:
Input: A graph G.
Question: Is G ∈ G?
36 3 Graph Classes

For example, the planar graphs are recognizable in linear time,1 but there
are many classes of graphs for which the recognition problem is not clear. For
example, consider the class

H = { G | G is a planar graph and χ(G) 6 3 }.

The recognition problem for H is NP-complete, since the 3-coloring problem

is NP-complete for planar graphs.

A lot of research is done on subclasses of perfect graphs. All classes that

we study in this chapter are perfect. We introduce the class of perfect graphs
in the next section.

3.1 Perfect graphs

Definition 3.1. A graph G is perfect if for every induced subgraph H of G

χ(H) = ω(H), (3.2)

where χ(H) is the chromatic number of H and ω(H) is the clique number of H.

If one wants to color a graph G such that adjacent vertices have different
colors, then all vertices of a clique in G must receive different colors. Thus for
all graphs we have

χ(G) > ω(G). (3.3)

For perfect graphs equality holds, not only for the graph itself but also for
all induced subgraphs of it. Notice that the class of perfect graphs is hereditary,
simply by definition.

Perhaps we should emphasize this. Assume that for some graph G,

χ(G) > ω(G).

For example, if G is an odd cycle of length more than 3 then

χ(G) = 3 and ω(G) = 2.

It is easy to construct a graph G0 such that χ(G0 ) = ω(G0 ) by adding a clique

of size at least χ(G) to G. The graph G0 is of course not perfect, since G is an
induced subgraph of G0 and equality in (3.3) does not hold for G.
1
S. Williamson, Depth-first search and Kuratowski subgraphs, Journal of the Associa-
tion for Computing Machinery 31 (1984), pp. 681–693.
 3.1 Perfect graphs 37

Since the structure of G0 is not essentially different from the structure of

G it cannot be expected that there are many problems that are easier to solve
for G0 than for G. For example, a maximum independent set for G0 consists
of an independent set in G plus one vertex in the clique that is added to G.
Thus the independent set problem for G0 is just as hard as it is for G. As we
will see, we get a whole new ballgame when we require that
χ=ω for every induced subgraph.

The complement of a graph G is the graph Ḡ with the same set of vertices
as G, and with two vertices in Ḡ adjacent if and only if they are not adjacent
in G. Concerning perfect graphs one of the first and most important theorems
was proved by Lovász in 1972.23

Theorem 3.2 (The perfect graph theorem). The complement of a perfect

graph is perfect.

Thus, if G is perfect then for every induced subgraph H of G we have that

α(H) = κ(H), (3.4)

where κ(H) = χ(H̄) is the smallest number of cliques that partition the set of
vertices and α(H) = ω(H̄) is the cardinality of a largest independent set in H.

In Exercise 3.3 we ask you to prove that bipartite graphs are perfect. By
Theorem 3.2 also the complements of bipartite graphs are perfect. (Can you
prove this without using Theorem 3.2?)

Another important example of a class of perfect graphs is the set of line-

graphs of bipartite graphs. The linegraph L(G) of a graph G has as its vertices
the edges of G and as its edges those pairs of edges in G that share an end-
point.

As an introductory example, let us prove that linegraphs of bipartite

graphs are perfect.
Coloring a linegraph L(G) of a graph G is equivalent to coloring the edges
of G such that any two edges that share an endpoint get different colors. By
Vizing’s theorem, for every graph G, χ(L(G)) is either ∆ or ∆ + 1, where ∆ is
the largest degree in G.4 Kőnig proved that bipartite graphs fall into the first
2
L. Lovász, Normal hypergraphs and the perfect graph conjecture, Discrete Mathe-
matics 2 (1972), pp. 253–267.
3
G. Gasparian, Minimal imperfect graphs: a simple approach, Combinatorica 16
(1996), pp. 209–216.
4
V. Vizing, On an estimate of the chromatic class of a p-graph, Diskret. Analiz 3
(1964), pp. 25–30.
38 3 Graph Classes

class, that is5

when G is bipartite: χ(L(G)) = ∆(G) = ω(L(G)).

In the following lemma we prove (3.4) for linegraphs of bipartite graphs.

Lemma 3.3. Let G be bipartite. Then L(G) is perfect.

Proof. First of all, it is sufficient to prove Equation (3.4) for L(G), since re-
moving an edge from a bipartite graph leaves it bipartite. (That is, the class
of linegraphs of bipartite graphs is hereditary.)

Notice that an independent set in L(G) is a set of edges in G of which no two

share an endpoint, i.e.,
α(L(G)) = ν(G),
where ν(G) is the cardinality of a maximum matching in G.
One of the earliest results in graph theory is the Theorem of König-Egerváry:6

If G is bipartite then ν(G) = τ(G).

Here, τ(G) is the cardinality of a smallest vertex cover in G. A vertex cover is

a set C of vertices such that every edge in G has at least one endpoint in C.
Let C = {x1 , . . . , xs } be a vertex cover of G. The set Li of edges that have the
vertex xi in common forms a clique in L(G). Since C is a vertex cover, every
edge in G is in some Li . Thus {L1 , . . . , Ls } is a clique cover in L(G). Possibly
some pairs Li and Lj are not disjoint, but it is easy to change the clique cover
into a partition of the vertices of L(G) into s cliques.
Thus, if C is a minimum vertex cover then

α(L(G)) = ν(G) = τ(G) = |C| > κ(L(G)) > α(L(G)),

since, by (3.3), α(H) 6 κ(H) for any graph H.

Thus α(L(G)) = κ(L(G)) and by Theorem 3.2 this proves the lemma. t
u

Obviously, if a graph G is perfect then it cannot have an induced odd

cycle of length at least five. By Theorem 3.2, when G is perfect it cannot
have the complement of an odd cycle of length at least five as an induced
subgraph. Claude Berge conjectured in 1961 that this characterizes perfect
5
D. Kőnig, Über Graphen und ihre Anwendung auf Determinantentheorie und Men-
genlehre, Mathematische Annalen 77 (1916), pp. 453–465.
6
D. Kőnig, Graphen und Matrizen, Matematikai Lapok 38 (1931), pp. 116–119.
 3.1 Perfect graphs 39

graphs.7 The proof of the conjecture sent a shock wave through the graph
theory community.8

Definition 3.4. Let G be a graph. A hole in G is an induced cycle of length at

least five. An antihole in G is an induced subgraph of G which is isomorphic to
the complement of a cycle of length at least five. An odd hole in G is a hole of
odd length. An odd antihole is an antihole of odd cardinality.

Fig. 3.1. The figure shows some holes and antiholes. Notice that the complement of
C5 is C5 . The complement of C6 is planar but the complement of C7 is not (and also
not the complements of longer cycles).

Theorem 3.5 (The strong perfect graph theorem). A graph is perfect if and
only if it has no odd hole and no odd antihole.

Theorem 3.5 was proved using a certain decomposition of the graph into
four basic classes of perfect graphs, namely,
(1) bipartite graphs,
(2) complements of bipartite graphs,
(3) linegraphs of bipartite graphs, and
(4) complements of linegraphs of bipartite graphs.
7
C. Berge, Färbung von Graphen, deren sämtliche bzw. deren ungerade Kreise starr
sind, Wiss. Zeitschr. Martin-Luther Univ. Halle-Wittenberg 10 (1961), pp. 114–115.
8
M. Chudnovsky, N. Robertson, P. Seymour and R. Thomas, The strong perfect graph
theorem, Annals of Mathematics 164 (2006), pp. 51–229.
40 3 Graph Classes

This decomposition theorem for perfect graphs also led to a polynomial

algorithm for recognizing perfect graphs.910

Theorem 3.6. There exists an O(n9 ) algorithm which tests if a graph with n
vertices is perfect.

One of the reasons for the popularity of perfect graphs is the following
theorem.11

Theorem 3.7. There exists a polynomial-time algorithm that computes ω(G)

and χ(G) for graphs G which satisfy ω(G) = χ(G).

The Shannon capacity of the complement of a graph is a graph parameter

which is sandwiched between the clique number and chromatic number. In
turn, the Lovász number is sandwiched between the Shannon capacity and
the clique cover number. Thus, if we write Θ(G) for the Shannon capacity
and ϑ(G) for the Lovász number, then

ω(Ḡ) = α(G) 6 Θ(G) 6 ϑ(G) 6 κ(G) = χ(Ḡ). (3.5)

The theorem above was proved by showing that the Lovász number ϑ(G) of a
graph G is computable in polynomial time.1213

3.2 Comparability graphs

Perhaps the best known class of perfect graphs is the class of comparability
graphs.

Definition 3.8. Let G = (V, E) be a graph. An orientation of G gives each edge

{x, y} a direction, either from x to y or from y to x.

Thus, in an orientation of G each edge {x, y} is directed in exactly one way,

either as −
→ or as ←
xy −
xy.

9
M. Chudnovsky, G. Cornuéjols, X. Liu, P. Seymour and K. Vušković, Recognizing
Berge graphs, Combinatorica 25 (2005), pp. 143–186.
10
P. Seymour, How the proof of the strong perfect graph conjecture was found, Gazette
des Mathématiciens 109 (2006), pp. 69–83.
11
M. Grötschel, L. Lovász and A. Schrijver, Polynomial algorithms for perfect graphs.
In (Berge, Chvátal eds.): Topics on perfect graphs, North-Holland Mathematics Stud-
ies 88 (1984), pp. 325–356.
12
L. Lovász, On the Shannon capacity of a graph, IEEE Transactions on Information
Theory 25 (1979), pp. 1–7.
13
M. Grötschel, L. Lovász and A. Schrijver, The ellipsoid method and its consequences
in combinatorial optimization, Combinatorica 1 (1981), pp. 169–197.
 3.2 Comparability graphs 41

Definition 3.9. An orientation of a graph G is transitive if for every three ver-

tices x, y and z, if there are directed edges −
→ and −
xy → then the pair {x, z} is an
yz
→
−
edge in G and in the orientation it is directed as xz.

Remark 3.10. A graph G = (V, E) equipped with a transitive orientation is

closely related to a partially ordered set, or a poset. A partially ordered set is
a pair (V, ) where V is a set and  is a binary relation. The binary relation
 is reflexive, antisymmetric and transitive. That means
(i) x  x for any element x ∈ V. This is reflexivity.
(ii) When x  y and y  x then x = y. This is antisymmetry.
(iii) When x  y and y  z then x  z. This is transitivity.
(So, the only ‘difference’ with the directed graph is the reflexivity.) When
every pair of elements of V is ordered by  then (V, ) is a total order (or
linear order). An example of a total order is the set of natural numbers, or
the real numbers, ordered by 6. In the case where (V, ) is a total order and
|V| is finite there is exactly one minimal element and one maximal element
in V. When (V, ) is a total order and if |V| is finite, then the corresponding
comparability graph is a clique (every pair of elements is comparable; some
people like to add a ‘loop’ at every vertex to have also the reflexivity).
Of course, when G = (V, E) is equipped with a transitive orientation, then
the directed graph is acyclic. Vertices with zero indegree are sometimes called
‘sources’ and vertices with zero outdegree, ‘sinks.’

Obviously, not every graph can be oriented transitively. For example, con-
sider an odd cycle C of length at least five. In any orientation of C there will
be two edges {x, y} and {y, z} that are oriented in the same direction, that is,
either −
→ and −
xy → or ←
yz − and ←
xy − If the orientation were transitive then in both
yz.
cases we would have an edge {x, z}, but since C has length at least 5, {x, z} is
not an edge of C.

Definition 3.11. A graph is a comparability graph if it has a transitive orienta-

tion.

Remark 3.12. Computing a transitive orientation can be done in linear time.14

However, in order to check that the orientation, which is obtained by this
algorithm, is indeed transitive one needs to do a matrix multiplication. Thus
the total time for comparability graph recognition takes O(n2.3727... ), using
14
M. Tedder, D. Corneil, M. Habib and C. Paul, Simple, linear-time modular decom-
position. ArXiv: 0710.3901v2, 2008.
42 3 Graph Classes

Williams’ algorithm for fast matrix multiplication.15 Alternatively, Golumbic

designed an easy O(n3 ) algorithm.16

Let G be a graph and consider a transitive orientation of G.

Definition 3.13. A chain is a directed path [x1 , . . . , xk ] such that

−
x− −−→ is a directed edge for all i ∈ {1, . . . , k − 1}.
i xi+1

Notice that, when [x1 , . . . , xk ] is a chain in a transitive orientation of a

graph G, then {x1 , . . . , xk } is a clique in G. For example, −
x−→ −−→
1 x2 and x2 x3 imply
that {x1 , x3 } is an edge in G. By induction it easily follows that x1 is adjacent
to all other vertices in the chain, and the same holds for all other vertices.

The converse holds as well.

Lemma 3.14. Let G be a graph and assume that G has a transitive orientation.
Then every clique in G corresponds with a chain in any transitive orientation of
the graph G.

Proof. Let Ω = {x1 , . . . , xk } be a clique in G. Every edge {xi , xj } has a direction,

either −
x−→ ←−−
i xj or xi xj . By transitivity there can be no directed cycles, not even
triangles.
Consider a vertex xi in Ω with a minimal number of incoming edges. Assume
that it has an incoming edge −
x−→
j xi . Then we obtain a contradiction as follows.
Each incoming edge xk xj yields an incoming edge −
−−→ x−→
k xi and so the number
of incoming edges into xi is at least one more than the number of incoming
edges into xj .
So there exists a vertex in Ω, say x1 , without any incoming edge. Notice that
x1 is unique, since any other vertex xj ∈ Ω now has at least one incoming
edge, namely −x− →
1 xj .

By induction Ω \ {x1 } forms a chain in the orientation of G; say [x2 , . . . , xk ].

Then [x1 , . . . , xk ] is also a chain in that orientation of G. t
u

Thus, if G is a graph with a transitive orientation then ω(G) is the number

of vertices in the longest chain. We want to prove that any comparability
graph is perfect. In order to do that, we need the concept of an antichain.

Definition 3.15. Let G be a graph with a transitive orientation. An antichain is

a set of vertices which is an independent set in G.
15
V. Williams, Breaking the Coppersmith-Winograd Barrier. Manuscript 2011.
16
M. Golumbic, The complexity of comparability graph recognition and coloring,
Computing 18 (1977), pp. 199–208.
 3.2 Comparability graphs 43

Theorem 3.16. Comparability graphs are perfect.

Proof. When G is a comparability graph then every induced subgraph of G is

also a comparability graph. Therefore, it is sufficient to prove that

ω(G) = χ(G).

Consider a transitive orientation of G. For each vertex x let ∆(x) be the maxi-
mal number of vertices in a chain that ends in x. Now define

Ni = { x | ∆(x) = i }.

Notice that N1 6= ∅, since there is a vertex without any incoming edge. This
follows, in the same manner as in the proof of Lemma 3.14, from the fact that
there are no directed cycles.
Also notice that, for any i > 1,

Ni+1 6= ∅ ⇒ Ni 6= ∅.

Let ω ∈ N be the maximal number such that Nω 6= ∅. Then, by Lemma 3.14,

ω = ω(G). The sets N1 , . . . , Nω form a partition of G into independent sets.
Thus χ(G) 6 ω(G).
This proves the theorem because, of course, χ(G) > ω(G) for any graph (this
is (3.3) on Page 36). t
u

As a corollary of Theorem 3.16 we obtain Dilworth’s theorem.17 This the-

orem is usually formulated for posets. The theorem can be proved in many
ways. For example, it can be seen that it is equivalent to the König-Egerváry
theorem. We obtain an easy proof via the perfect graph theorem.

Theorem 3.17 (Dilworth’s theorem). Let G = (V, E) be a graph with a tran-

sitive orientation. There exists an antichain A and a partition C of V into κ
chains
C = { C1 , . . . , Cκ } such that κ = |A|.

Proof. By the perfect graph theorem, the complement of G is perfect, and so

α(G) = κ(G).

A clique cover corresponds with a partition of V into a collection of chains.

t
u

17
R. Dilworth, A decomposition theorem for partially ordered sets, Annals of Mathe-
matics 51 (1950), pp. 161–166.
44 3 Graph Classes

Comparability graphs can be characterized by a set of forbidden induced

subgraphs. That is, there exists a collection of graphs F such that a graph G is
a comparability graph if and only if no element of F is an induced subgraph of
G. The set F is not finite, for example it contains all the odd cycles of length
at least five.18

In the following figures, Figure 3.2 (Table I) and Figure 3.3 on the facing
page (Table II), we show Gallai’s list of forbidden induced subgraphs for com-
parability graphs. Some care is needed, since the complements of the graphs
in Figure 3.3 are forbidden. By the way, the circled vertices in this figure form
asteroidal triples. We’ll talk about those in Section 3.6.
Theorem 3.18 (Gallai). A graph is a comparability graph if and only if it has
none of the graphs in Figure 3.2 and none of the complements of the graphs in
Figure 3.3 as an induced subgraph.

a0 a0
a1 a2n a1 a2n

a2 a2 a2n-1

_
n>2 _
n>2
a0 a0
a2n a1
a1 a2n

a2 a2n-1 a2 a2n-1

_
n>2 _
n>2
Fig. 3.2. Table I.

By the perfect graph theorem, the complements of comparability graphs

are perfect. Let’s have a closer look at those.
Definition 3.19. A graph G is a cocomparability graph if Ḡ is a comparability
graph.
18
T. Gallai, Transitiv orientierbare Graphen, Acta Mathematica Academiae Scientiarum
Hungaricae 18 (1967), pp. 25–66.
 3.2 Comparability graphs 45

c1 c2 cγ
_
γ>3

_
n-cycle, with n>6

c1 c2 cγ c1 c2 cγ
_
γ>1 _
γ>1

Fig. 3.3. Table II.

Cocomparability graphs have a nice intersection model.19

Theorem 3.20. A graph G is a cocomparability graph if and only if for every
vertex x ∈ V there exists a continuous function

fx : [0, 1] → R
19
M. Golumbic, D. Rotem and J. Urrutia, Comparability graphs and intersection
graphs, Discrete Mathematics 43 (1983), pp. 37–46.
46 3 Graph Classes

such that two vertices x and y are adjacent if and only if the functions fx and fy
intersect each other. By that we mean that there exists a point 0 6 r 6 1 such
that
fx (r) = fy (r).

Proof. Let P = (V, ) be a poset. A realizer for P is a collection

{ 61 , . . . , 6k }
of total orders of V such that  is the intersection of 6i , i ∈ {1, . . . , k}. By that
we mean that, for any two elements x and y in V,
xy if and only if x 61 y and ... and x 6k y.

Dushnik and Miller show that every partial order has a realizer. A finite poset
has a finite realizer. The dimension of the poset is the minimal cardinality of
a realizer.20
Consider a realizer {61 , . . . , 6k } of a poset P = (V, ). Let V = {1, . . . , n}.
Draw k vertical lines, L1 , . . . , Lk , next to each other. On each line Li label n
points {1, . . . , n} in the order of 6i . Now connect any point on line Li by a
straight line segment with the point that has the same label on line Li+1 , for
i ∈ {1, . . . , k − 1}. See Figure 3.4 on the next page.
Now we have that, for any two elements x and y in V, that x  y if and
only if x 6i y for all i ∈ {1, . . . , k}. That is, x  y if and only if the functions
fx and fy , defined as the concatenation of the straight line segments, do not
intersect.
The converse is easy. If a graph G has an intersection model by continuous
functions, as above, then the ordering (x, y), defined by “fx lies below fy ,” is
a transitive orientation of Ḡ.
This proves the theorem. t
u

Remark 3.21. In Section 3.7 we have a look at the posets of dimension two.
The corresponding comparability graphs are called permutation graphs.

Remark 3.22. It is NP-complete to determine whether the dimension of a poset

is at most k for any k > 3.21
20
B. Dushnik and E. Miller, Partially ordered sets, American Journal of Mathematics 63
(1941), pp. 600–610.
21
M. Yannakakis, The complexity of the partial order dimension problem, SIAM Jour-
nal on Algebraic and Discrete Methods 3 (1982), pp. 351–358.
 3.2 Comparability graphs 47

1
6 6 4

>
>
3 1
6 2 2 4 2

>
> 1 5 3
5

4 2 6
2
5 3
5 1 1

> >
4 6 4 3 3 5

Fig. 3.4. This figure shows the complement of C6 . Notice that C6 is a comparability
graph. The figure shows the intersection model. The dimension of the poset is three.
The picture on the right shows the three total orders of the realizer.

3.2.1 Recognition of comparability graphs

In this section we explain Golumbic’s recognition algorithm for comparability

graphs.2223 This algorithm can be implemented such that it runs in O(n3 )
time.

Let G = (V, E) be a graph. For a set E ⊆ E write

→
−
E ={−
→ −
xy, → | {x, y} ∈ E }.
yx (3.6)
→
−
We call the elements of E the directed edges of G.

Let F be a set of directed edges of a graph G. Define the sets F−1 , F2 and F̂
as follows.
1.
F−1 = { ←−|−
xy →∈F}
xy (3.7)
2.
F2 = { →
−
xz | ∃y∈V −
→ ∈ F and −
xy →∈F}
yz (3.8)
3.
F̂ = F ∪ F−1 . (3.9)

22
M. Golumbic, Comparability graphs and a new matroid, Journal of Combinatorial
Theory, Series B 22, (1977) pp. 68–90.
23
M. Golumbic, The complexity of comparability graph recognition and coloring,
Computing 18 (1977), pp. 199–208.
48 3 Graph Classes
→
−
Definition 3.23. Let G = (V, E) be a graph and let E ⊆ E. A set F ⊆ E of
directed edges is an orientation of E if
→
−
F ∪ F−1 = E and F ∩ F−1 = ∅. (3.10)

If E = E then F is an orientation of G.

Definition 3.24. Let F be an orientation of G. Then F is transitive if F2 ⊆ F. A

graph G is a comparability graph if it has a transitive orientation.

→
−
Definition 3.25. Define a binary relation Γ on E of a graph G = (V, E) as
follows.

→ −−
− → a = a0 and {b, b0 } ∈/ E, or
ab Γ a0 b0 ⇔ (3.11)
b = b0 and {a, a0 } ∈/ E.

Notice that the relation Γ is reflexive and symmetric. Its transitive closure
→
−
Γ is an equivalence relation on E . We call the equivalence classes of Γ ∗ the
∗
→
−
implication classes of E .

→
−
The Γ relation on E captures the fact that, for any transitive orientation F
of G,
→ −−→
− −→ −−→
ab Γ a0 b0 and ab ∈ F ⇒ a0 b0 ∈ F. (3.12)

Golumbic gave a simple algorithm to test whether an undirected graph

G is a comparability graph and to give it a transitive orientation if it is. The
→
−
central part of Golumbic’s algorithm is to compute a G-decomposition of E
which is defined as follows.
→
−
Definition 3.26. Let G = (V, E) be a graph. A partition of E

{ B̂1 , B̂2 , · · · , B̂k } (3.13)

is a G-decomposition if, for i ∈ {1, . . . , k}, Bi is an implication class of

k
[
B̂` . (3.14)
`=i

Golumbic’s algorithm follows directly from the following theorem.

 3.2 Comparability graphs 49

Theorem 3.27. Let G = (V, E) be a graph with a G-decomposition

k
→
− [
E = B̂i . (3.15)
i=1

The following statements are equivalent.

(i) G is a comparability graph.
(ii) A ∩ A−1 = ∅ for all implication classes A of G.
i = ∅ for i ∈ {1, . . . , k}.
(iii) Bi ∩ B−1
Furthermore, when these conditions hold, then
k
[
B` (3.16)
`=1

is a transitive orientation of E.
By Theorem 3.27, we can test whether a graph G is a comparability graph,
and give G a transitive orientation if it is a comparability graph, by computing
→
−
a G-decomposition of E . Golumbic describes an O(n3 ) algorithm to compute
a G-decomposition.

To prove Theorem 3.27 we need a lemma.

Lemma 3.28 (The Triangle Lemma). Let X, Y and Z be three implication
classes of a graph G = (V, E). Assume that

X 6= Y and X 6= Z−1 . (3.17)

Let
−
→∈Z
xy and →
−
xz ∈ Y and −
→ ∈ X.
yz (3.18)
Then −−0→0 −→0 −→0
yz ∈X ⇒ xy ∈ Z and xz ∈ Y. (3.19)
Proof. By definition, there exists a chain
−
→=−y−→ −−→ −−→ −−−→ with yk = y0 and zk = z0 .
yz 0 z 0 Γ y1 z 0 Γ y1 z 1 Γ . . . Γ yk z k (3.20)

We use induction on i. Assume that −→ ∈ Y and −

xz i
→ ∈ Z. Then we have that
xyi
(since X 6= Y and X 6= Z−1 )

Y3−→ 6Γ −
xz i y−−−→
i+1 zi ∈ X ⇒ {x, yi+1 } ∈ E
{yi+1 , yi } ∈
/E ⇒ −−−→ Γ −
xy →∈Z
xy
i+1 i
Z−1 −−−→ −−− −− −
→
3 yi+1 x 6 Γ yi+1 zi+1 ∈ X ⇒ {x, zi+1 } ∈ E
{zi , zi+1 } ∈
/E ⇒ −−−→ Γ −
xz → ∈ Y.
xz
i+1 i

Therefore, −−→ ∈ Z and −

xyk xz→
k ∈ Y. t
u
50 3 Graph Classes

Z Y
<
y

<
<
X z
Fig. 3.5. The figure illustrates the Triangle Lemma.

We are now ready to prove Theorem 3.27.

Theorem 3.29. Let G = (V, E) be a graph with a G-decomposition

E = { B̂1 , · · · , B̂k }. (3.21)

The following statements are equivalent.

(i) G is a comparability graph.
→
−
(ii) A ∩ A−1 = ∅ for all implication classes A of E .
i = ∅ for i ∈ {1, . . . , k}.
(iii) Bi ∩ B−1
Furthermore, when these conditions hold, then
k
[
B` (3.22)
`=1

is a transitive orientation of E.
Proof. Assume that G is a comparability graph. We show that (i) implies (ii).
Let A be an implication class. We show that A ∩ A−1 = ∅.
Let F be a transitive orientation. By applying (3.12) inductively, either

A∩F=∅ or A ⊆ F.

First assume that A ∩ F = ∅. Since A ⊆ F ∪ F−1 we have that A ⊆ F−1 . Since

F ∩ F−1 = ∅ we have that A ∩ A−1 = ∅. Now assume that A ⊆ F. Since F−1 is
also a transitive orientation of G, it follows similarly that A ∩ A−1 = ∅.
We now show that (ii) implies (iii). We use induction on the number of im-
plication classes in the G-decomposition. By assumption, B1 is an implication
class of G. When k = 1 we are done. Assume that the claim holds true for all
graphs G0 with less than k implication classes in some G0 -decomposition.
Consider the graph induced by E \ B̂1 . Let D be an implication class of E \ B̂1 .
We claim that either D is an implication class of E or D = C1 ∪ C2 , where C1
and C2 are two implication class of E.
 3.2 Comparability graphs 51

The removal of B̂1 causes some implication classes to merge into one. Assume
that D is the union of t implication classes and assume that t > 2. Then there
exists a triangle {x, y, z} with −
→ ∈ B̂ and either
yz 1

(→
−
xz ∈ C1 and −
→∈C )
xy 2 or (←
−∈C
xz 1 and ←− ∈ C ).
xy 2 (3.23)

Without loss of generality we may assume that →

−
xz ∈ C1 and −
→ ∈ C since the
xy 2
alternative follows by symmetry by considering D−1 .
Assume that C = C−1 . Then −
1 2
→ ∈ C and →
yx 1
−
xz ∈ C . But −→ ∈
yz / C and so
1 1
C1 = Ĉ1 = C−1 1 which implies C1 = C2 . This is a contradiction. Therefore,
Ĉ1 ∩ Ĉ2 = ∅.
Any Γ -chain in E \ B̂1 with edges in Ĉ1 ∪ Ĉ2 cannot contain edges from any
other implication class since any triangle in E with one edge in B̂1 and the
second in Ĉ1 must have its third edge in Ĉ2 . Thus t = 2 and D = C1 ∪ C2 .
If D is an implication class of E \ B̂1 then D ∩ D−1 = ∅. If D = C1 ∪ C2 , for
two implication classes C1 and C2 with Ĉ1 ∩ Ĉ2 = ∅, then

D ∩ D−1 = (C1 ∪ C2 ) ∩ (C−1 −1

1 ∪ C2 )
= (C1 ∩ C−1 −1
1 ) ∪ (C2 ∩ C2 ) = ∅.

i = ∅ for i ∈ {2, . . . , k}.

Therefore, by induction, Bi ∩ B−1
We prove that (iii) implies (i). We first show that B1 is transitive. Let −
→∈B
xy 1
−
→
and yz ∈ B1 . If {x, z} ∈
/ E then
−
→Γ −
xy →
zy ⇒ −
→∈B
zy ⇒ −
→ ∈ B−1
yz (3.24)
1 1

which is a contradiction since B1 ∩ B−11 = ∅. Thus {x, z} ∈ E. Let W be the

implication class containing →
−
xz and assume that W 6= B1 . Since B1 6= B−1 1
and −
→ ∈ B , by the triangle lemma, −
xy 1
→ ∈ W which is a contradiction. Thus
xy
→
−
xz ∈ B1 and this proves that B1 is transitive.
When k = 1 we are done. Assume that the implication holds for all graphs G0
that have a G0 -decomposition with less than k implication classes. Thus
k
[
F= Bi (3.25)
`=2

is transitive.
Let −→ and −
xy → be two elements of B ∪ F. If both are in B or if both are in F
yz 1 1
then we are done. Assume that −→ ∈ B and −
xy 1
→ ∈ F. Then
yz
−
→ 6 Γ∗ −
xy →
zy ⇒ {x, z} ∈ E. (3.26)

Assume that →
− / B1 ∪ F. Then →
xz ∈ −
zx ∈ B1 ∪ F. Now
52 3 Graph Classes

(a) →
−
zx ∈ B1 and −
→ ∈ B imply −
xy → ∈ B which is a contradiction, and
zy
1 1
(b) →
− −
→ −
→
zx ∈ F and yz ∈ F imply yx ∈ F, which is also a contradiction.
Thus →
−
xz ∈ B1 ∪ F. Similarly, −
→ ∈ F and −
xy → ∈ B imply →
yz 1
−
xz ∈ B1 ∪ F.
This proves that ∪ki=1 Bi is a transitive orientation of G.
This proves the theorem. t
u

A G-decomposition for a graph G = (V, E) can be found as follows.

1. Let i = 1 and let E1 = E.

2. Choose ei ∈ Ei .
3. Enumerate the implication class Bi of Ei which contains ei .
4. Define Ei+1 = Ei \ B̂i .
5. If Ei+1 = ∅ then k = i and STOP. Otherwise, increase i by one and continue
with step 2.

Theorem 3.30. There exists an O(n3 ) algorithm to recognize comparability

graphs.

Proof. An edge is put into an implication class at most one time. Whenever an
edge e is put in an implication class, all the neighbors of the two endpoints
are examined. For all the neighbors that are adjacent to exactly one endpoint
of e, the edges are put in the same implication class as e.
A flag on each edge is used to detect when an arc − → is detected for which
xy
←−
xy is in the same implication class. When this happens, the graph is not a
comparability graph and otherwise, by Theorem 3.27 it is.
Each edge is put into an implication class only once. When an edge e is put
in an implication class O(n) vertices are examined to see which edges are
Γ -related to e. This shows that the algorithm can be implemented to run in
O(nm) time. t
u

Remark 3.31. By Theorem 3.16 on page 43, when a transitive orientation is a

part of the input, then a maximum clique and an optimal coloring in a com-
parability graph can be computed in O(n2 ) time by dynamic programming.
Therefore, the clique and the chromatic number problem can be solved in
O(n3 ) time on comparability graphs.24

24
M. Golumbic, The complexity of comparability graph recognition and coloring,
Computing 18 (1977), pp. 199–208.
 3.3 Cographs 53

3.3 Cographs
One of the most elegant classes of graphs is the class of cographs.

Definition 3.32. A graph G is a cograph if it has no induced P4 , which is a path

with four vertices.

s s s s

Fig. 3.6. A P4 is a path with four vertices. Notice that P¯4 = P4 .

By definition, the class of cographs is hereditary, namely, if G has no in-

duced P4 then no induced subgraph of G has an induced P4 .

Lemma 3.33. If G is a cograph then G is perfect.

Proof. Let G be a cograph. We show that neither G nor Ḡ has a hole (an
induced cycle of length at least 5).
The graph G has no hole, otherwise it has an induced P4 . Notice that P¯4 is
isomorphic to P4 . Thus if G is a cograph then Ḡ is also a cograph. Thus G has
no antihole.
Since G has no hole it has no odd hole and the same holds true for Ḡ. The
claim now follows from Theorem 3.5. t
u

Theorem 3.34. A graph G = (V, E) is a cograph if and only if for every induced
subgraph H of G one of the following properties holds.
(a) H has only one vertex, or
(b) H is disconnected, or
(c) H̄ is disconnected.
Proof. Notice that P¯4 is isomorphic to P4 . This implies that if G is a cograph
then Ḡ is a cograph.
A graph with less than four vertices is a cograph. Therefore, since the class
of cographs is hereditary, it is sufficient to prove that G or Ḡ is disconnected
when G is a cograph with at least two vertices.
It is easy to check that the claim holds true when G has at most three vertices.
Assume that G has at least four vertices. Assume also that G is connected.
Let x be a vertex of G. By induction G − x or the complement of G − x is
disconnected.
54 3 Graph Classes

Assume that G − x is disconnected and let C1 , . . . , Ct be the components of

G − x. Assume that x has a neighbor and a nonneighbor in C1 . Then there
exist vertices a and b in C1 such that [a, b, x] is an induced P3 in G. Since
G is connected, x has a neighbor c in C2 . Now [a, b, x, c] is an induced P4
which is a contradiction. Thus x is adjacent to all other vertices in G. Then Ḡ
is disconnected, since {x} is a component of Ḡ.
Assume that the complement of G − x is disconnected. (We now ‘copy’ the
argument above.) Let C01 , . . . , C0s be the components of the complement of
G − x. Assume that x has a neighbor a0 and a nonneighbor b0 in C01 . Since
the complement of G[C01 ] is connected, we may assume that a0 and b0 are not
adjacent in G.
If x is adjacent to all vertices in C02 then Ḡ is disconnected with a component
C02 . Thus x has a nonneighbor c0 in C2 . Now [x, a0 , c0 , b0 ] is an induced P4 ,
which is a contradiction.
Thus we may assume that x is not adjacent to any other vertex in G, and so G
is disconnected.
This proves the theorem. t
u

Remark 3.35. There are many alternative characterizations. For example, a

graph is a cograph if and only if in every induced subgraph every maximal
clique intersects every maximal independent set (in one vertex).

Remark 3.36. A very interesting subclass of cographs is the class of trivially

perfect graphs. These are the graphs without induced P4 and C4 .
An intersection model for the trivially perfect graphs can be obtained as
follows. If G = (V, E) is a connected trivially perfect graph then there exists a
rooted tree with vertex set V such that any two vertices x and y are adjacent if
and only if one of them lies on the path to the root of the other one. It follows
that G is trivially perfect if and only if every connected, induced subgraph has
a universal vertex, that is a vertex which is adjacent to all other vertices in
that subgraph.
The reason why this class is called trivially perfect is the following. A graph
is trivially perfect if and only if in every induced subgraph H, the number of
maximal cliques is equal to the independence number α(H). (Notice that the
leaves of the tree described above form a maximum independent set and that
the paths from the leaves to the root form all maximal cliques.)
This property implies (trivially) that the graphs are perfect, since

α(H) 6 κ(H)
 3.3 Cographs 55

for any graph H and κ(H) is of course at most the number of maximal cliques,
which is α(H) by the property.25 26

3.3.1 Cotrees
Let G be a cograph. Theorem 3.34 shows us how we can build a decompo-
sition tree for the graph G. For cographs, this decomposition tree is called a
cotree.27

The decomposition tree is a pair (T , f) where T is a binary tree and f

is a bijection from the leaves of T to the vertices of G. Each internal node,
including the root, is labeled with a ⊗ - or a ⊕-operator. Consider an internal
vertex t. Let V1 and V2 be the two sets of vertices in G that are mapped to the
leaves of the left and right subtree. If t is labeled with ⊗ then every vertex of
V1 is adjacent to every vertex of V2 . If t is labeled by ⊕ then no vertex of V1
is adjacent to any vertex of V2 .

Let G be a cograph. We can build a decomposition tree as follows. If G has

only one vertex, the tree consists of a single leaf, which is mapped by f to the
vertex of G.
Otherwise, by Theorem 3.34, either G or Ḡ is disconnected.

Assume that G is disconnected and let C1 , . . . , Ct be the components of

G. Group the components into two nonempty sets, say A and B. Recursively,
build decomposition trees (T1 , f1 ) and (T2 , f2 ) for G[A] and G[B]. Create a new
root, and make it adjacent to the root of T1 and to the root of T2 . Label the
new root by ⊕.

Assume that Ḡ is disconnected. In that case, build a decomposition tree

(T̄ , f̄) for Ḡ as described above. Change all labels from ⊕ to ⊗ and vice versa.

A decomposition tree for cographs, as described above, is called a (binary)

cotree. One obtains a somewhat different cotree by recursively creating a child
for every component of the subgraph or its complement (see Figure 3.7). In
that case, the labels on every path from a leaf to the root alternate between
⊕ and ⊗.

Notice that a graph G is a cograph if and only if it has a cotree (see Exer-
cise 3.9). Corneil, Perl and Stewart proved the following theorem.28
25
E. Wolk, A note on the comparability graph of a tree, Proceedings of the American
Mathematical Society 16 (1965), pp. 17–20.
26
M. Golumbic, Trivially perfect graphs, Discrete Mathematics 24 (1978), pp. 105–
107.
27
D. Corneil, H. Lerchs and L. Stewart-Burlingham, Complement reducible graphs,
Discrete Applied Mathematics 3 (1981), pp. 163–174.
28
D. Corneil, Y. Perl and L. Stewart, A linear recognition algorithm for cographs, SIAM
Journal on Computing 14 (1985), pp. 926–934.
56 3 Graph Classes

2 7
×
+ +
3 6
1 × × 6 7

4 5 2 3 4 5
(a) a cograph (b) a cotree

Fig. 3.7. The figure shows a cograph and a cotree. Notice the difference with the binary
cotree described above. When a cotree is built by creating a child for each component
of the subgraph or its complement, the labels alternate on every path from a leaf to
the root. For algorithmic purposes one usually prefers the binary cotree.

Theorem 3.37. There exists a linear-time algorithm that recognizes cographs.

When G is a cograph then this algorithm builds a cotree for G.

3.3.2 Finding cliques in cographs

One appreciates cotrees when one looks at some examples. Let’s use it for the
computation of the clique number.

Theorem 3.38. There exists a linear-time algorithm that computes the clique
number ω(G) of a cograph G.

Proof. Let G = (V, E) be a cograph. First, the algorithm builds a cotree (T , f)

for G. By Theorem 3.37 this step takes linear time.
First assume that G has only one vertex. Then ω(G) = 1.
Now assume that |V| > 2. For an internal node p of T let V1 and V2 be the two
sets of vertices that are mapped to the leaves in the left and right subtree. Let
Gi = G[Vi ] for i ∈ {1, 2} and let

Gp = G[V1 ∪ V2 ].

First assume that the label of p in T is ⊗. Then Gp is the join of G1 and G2 ,

that is, every vertex of G1 is adjacent to every vertex of G2 . Notice that

ω(Gp ) = ω(G1 ) + ω(G2 ). (3.27)

 3.4 Distance-hereditary graphs 57

Now assume that the label of p is ⊕. Then Gp is the union of G1 and G2 , that
is, no vertex of G1 is adjacent to any vertex of G2 . Now

ω(Gp ) = max { ω(G1 ), ω(G2 ) } (3.28)

Assume that p is the root of T . The algorithm recursively computes ω(G1 ) and
ω(G2 ). Since p is the root, Gp = G and ω(G) follows from Formulas (3.27)
and (3.28).
The amount of work in each internal node takes O(1) time. So, in total the
algorithm runs in time O(n), where n = |V|, since the depth of the cotree is at
most n. In other words, when the cotree is a part of the input this algorithm
runs in O(n) time.
This proves the theorem. t
u

Remark 3.39. Obviously, there exists also a linear-time algorithm to compute

a maximum independent set via formulas similar to (3.27) and (3.28).
Recall Definition 2.16 on page 13. A dominating set in a graph G = (V, E)
is a set D of vertices such that every vertex x ∈ V \D has at least one neighbor
in D. We denote the minimal cardinality of a dominating set in G by γ(G).
The formulas to compute α(G) and γ(G) for each of the nodes in a cotree
are almost the same, except that the max-operator in Equation (3.28) for the
⊗-nodes is replaced by

min { γ(G1 ), γ(G2 ), 2 }.

It follows that also γ(G) can be computed in linear time for cographs.

3.4 Distance-hereditary graphs

Edward Howorka introduced distance-hereditary graphs (abbr. DH-graphs).29

Definition 3.40. A graph G is distance hereditary if for every pair of nonadja-

cent vertices x and y and for every connected, induced subgraph H of G which
contains x and y, the distance between x and y in H is the same as the distance
between x and y in G.

In other words, a graph G is distance hereditary if for every nonadjacent

pair x and y of vertices, all chordless paths between x and y in G have the
same length. (A path P is chordless if G[P] is a path; that is, P has no short-
cuts.)
29
E. Howorka, A characterization of distance-hereditary graphs, The Quarterly Journal
of Mathematics 28 (1977), pp. 417–420.
58 3 Graph Classes

Notice that, by definition the class of distance-hereditary graphs is hered-

itary.

There are various characterizations of distance-hereditary graphs. One of

them states that a graph is distance hereditary if and only if it has no induced
house, hole, domino or gem. See Figure 3.8.

s s s s s s
s @@s B @
s s s s s B  @s
@ bb B  " "
s s @s s s s s
b "
bB"
Fig. 3.8. A graph is distance hereditary if it has no induced house, hole, domino or
gem.

Distance-hereditary graphs are also characterized by the property that ev-

ery induced subgraph has either an isolated vertex, or a pendant vertex, or a
twin. An isolated vertex is a vertex with no neighbors. A pendant vertex is a
vertex with exactly one neighbor.

Definition 3.41. A twin is a pair of vertices x and y such that either

N(x) = N(y) or N[x] = N[y].

Let G be distance hereditary and let G0 be the graph obtained from G by

adding an isolated vertex or a pendant vertex, or a twin x of some vertex y in
G. In Exercise 3.14 we ask you to check that G0 is also distance hereditary.

Theorem 3.42. Let G = (V, E) be a graph. The following statements are equiv-
alent.
(1) The graph G is distance hereditary.
(2) Every induced subgraph of G has an isolated vertex, or a pendant vertex or
a twin.
(3) The graph G has no induced house, hole, domino or gem.

Proof. We prove

(1) ⇒ (3) ⇒ (2) ⇒ (1).

 3.4 Distance-hereditary graphs 59

It is easy to check that, when G is distance hereditary, then it has no induced

house, hole, domino or gem. We show that this implies that every induced
subgraph H of G has an isolated vertex, or a pendant vertex or a twin.
First, since the class of distance-hereditary graphs is hereditary, it is sufficient
to prove that G has an isolated vertex, or a pendant vertex or a twin. We may
assume that G is connected, since a graph is distance hereditary if and only if
every component of G is distance hereditary.
Let x be a vertex for which the largest component C of G − N[x] has a largest
number of vertices. We may assume that C 6= ∅, since otherwise G is a clique
and then we are done. Let

∆ = N(C) = { y | y ∈ V \ C and N(y) ∩ C 6= ∅ }.

Thus ∆ ⊆ N(x) and every vertex of ∆ has a neighbor in C. Notice that ∆ 6= ∅

since we assume that G is connected.
Notice that, by the choice of x,
every vertex of V \ (C ∪ ∆) is adjacent to every vertex of ∆.
(If this were not true, then there would be another vertex x0 with a larger
component than C in G − N[x0 ], namely, one component of G − N[x0 ] would
include C and some vertex of ∆.)
Now, notice that all vertices of ∆ have the same neighbors in C, since other-
wise there would be a house, hole, domino or gem. (This takes some effort to
figure out, but it is not difficult and we leave it as an exercise.)
Let
Ω = V \ (C ∪ ∆).
Thus x ∈ Ω and every vertex of Ω is adjacent to every vertex of ∆. Notice that
Ω has no induced P4 since otherwise, with some vertex δ ∈ ∆ there would be
a gem and a gem is not distance hereditary.
Thus G[Ω] is a cograph, and so, by Exercise 3.10, either it has only one vertex
or it has a twin. Any twin in G[Ω] is a twin in G, since all vertices of Ω have
the same neighbors in G − Ω, namely the vertices of ∆.
Thus, now we may assume that Ω = {x}, since otherwise there is a twin. No-
tice that G[∆] has no induced P4 , since otherwise there would be an induced
gem in G (with the vertex x), and a gem is clearly not distance hereditary.
Thus, if G[∆] has at least two vertices then it contains a twin, since G[∆] is a
cograph. (Here we use Exercise 3.10 again.)
A twin in G[∆] is a twin in G, since all vertices of ∆ have the same neighbors
in V \ ∆. Thus, finally, we may assume that |∆| = 1. But then x is a pendant
vertex.
60 3 Graph Classes

Thus, if G is distance hereditary, then every induced subgraph has an iso-

lated vertex, or a pendant vertex or a twin. We leave it as an easy exercise
(Exercise 3.14) to prove the converse.
This proves the theorem. t
u

Theorem 3.43. Distance-hereditary graphs are perfect.

Proof. Let G = (V, E) be distance hereditary. Then G has no hole. It remains

to show that G has no odd antihole.
Let x ∈ V be a vertex which is either isolated, or a pendant vertex adjacent to
some vertex y, or a twin of some vertex y. Consider G − x. By induction we
may assume that G − x has no odd hole or odd antihole.
Assume that G has an odd antihole H. Let V 0 be the vertex set of H. Then
x ∈ V 0 . Since H is connected, x is not isolated. Since H is biconnected, x is
not a pendant vertex. Thus x is a twin of some vertex y. Notice that y is not a
vertex of H, since H has no twins. Let

V 00 = (V 0 \ {x}) ∪ {y}.

Then V 00 induces an odd antihole in G − x which is a contradiction.

This proves the theorem. t
u

3.4.1 Decomposition trees for DH-graphs

A decomposition tree for a graph G = (V, E) is a pair (T , f) consisting of a

rooted binary tree T and a bijection f from V to the leaves of T .

When G is distance hereditary it has a decomposition tree (T , f) with the

following three properties.30 31

Consider an edge e = {p, c} in T where p is the parent of c. Let We ⊂ V

be the set of vertices of G that are mapped by f to the leaves in the subtree
rooted at c. Let Qe ⊆ We be the set of vertices in We that have neighbors in
G − We . The set Qe is called the twinset of e. The first property is that the
subgraph of G induced by Qe is a cograph for every edge e in T .

Consider an internal vertex p in T . Let c1 and c2 be the two children of p.

Let e1 = {p, c1 } and let e2 = {p, c2 }. Let Q1 and Q2 be the twinsets of e1 and
30
P. Hammer and F. Maffray, Completely separable graphs, Discrete Applied Mathe-
matics 27 (1990), pp. 85–99.
31
H. Bandelt and H. Mulder, Distance-hereditary graphs, Journal of Combinatorial
Theory, Series B 41 (1986), pp. 182–208.
 3.4 Distance-hereditary graphs 61

e2 . The second property is that there is a join- or a union-operation between

Q1 and Q2 . Thus either all vertices of Q1 are adjacent to all vertices of Q2 , or
they are not adjacent to any vertex of Q2 . As in cotrees, there is a label ⊕ or
⊗ at the vertex p in T that indicates which operation is performed on Q1 and
Q2 .

Remark 3.44. Notice the difference with the labels in cotrees. The ⊕- or ⊗-
operator in the decomposition tree for distance-hereditary graphs works on
the twinsets, and not, as in the cotrees, on all the vertices of We1 and We2 .

Let p be an internal vertex of T which is not the root. Let e be the line
that connects p with its parent. Let Qe be the twinset of e. Let c1 and c2 be
the two children of p in T . Let e1 = {p, c1 } and let e2 = {p, c2 }. Let Qi be the
twinset of ei , for i ∈ {1, 2}. The third, and final, property is that

Qe = ∅ or Qe = Q1 or Qe = Q2 or Qe = Q1 ∪ Q2 .

The vertex p in T has an extra label that indicates which of these four opera-
tions that define Qe occur.

Notice that the first property is a consequence of the other two. As an

example, notice that cographs are distance hereditary. A cotree is a decom-
position tree for a cograph with the three properties mentioned above. See
Figure 3.9 for another example.

1 {1, 2}
× {3, 5}
2

3
× ×
5 {1} {2} {3, 5} {4}

4 1 2
+ ×
{3} {5} {4} {6}

6 3 5 4 6

Fig. 3.9. An example of a distance-hereditary graph and a decomposition tree for it.
The labels at the edges are the twinsets. The ⊕- and ⊗-labels at the nodes represent
the union- and join-operation on the twinsets of the two children.
62 3 Graph Classes

Theorem 3.45. Let G be a graph. Then G is distance hereditary if and only if it

has a decomposition tree as described above.

Proof. Let G = (V, E) be distance hereditary. We use the property that G has
an isolated vertex, a pendant vertex or a twin.
Let x be an isolated vertex, a pendant vertex with neighbor y, or a twin of a
vertex y. Let G0 = G − x. By induction G0 has a decomposition tree (T 0 , f0 ) as
described above. We now show how to construct a decomposition tree (T , f)
for G.
First assume that x is a twin of a vertex y. The vertex y is mapped by f0 to
some leaf ` of T 0 . Create two leaves `1 and `2 and let ` be the parent of `1
and `2 in T . Let f map x to `1 and y to `2 , and let f be the same as f0 for
all other vertices z ∈ V \ {x, y}. If x and y are adjacent, then the new internal
node ` receives an ⊗-operator and otherwise it receives an ⊕-operator. Finally,
update the twinsets by adding x to all twinsets that contain y. The edges {`, `1 }
and {`, `2 } have twinsets {x} and {y}.
Assume that x is a pendant vertex with a neighbor y. The tree T and the map
f are obtained in the same manner as described above. The internal node `
receives an ⊗-operator since x and y are adjacent. Finally, x appears in only
one twinset, namely in the twinset of the new edge {`, `1 } in T .
Assume that x is isolated in G. Choose an arbitrary vertex y in V \ {x}. Create
the tree T and the map f as above. In this case, the vertex x appears in no
twinset.
Assume that a graph G has a decomposition tree as described above. You can
use the technique of Exercise 3.10 to show that every induced subgraph has
an isolated vertex, a pendant vertex or a twin. We leave it as an exercise. t
u

When G is distance hereditary then a tree-decomposition for G with the

three properties described above can be obtained in linear time.32

3.4.2 Feedback vertex set in DH-graphs

To appreciate decomposition trees for DH-graphs you need to look at some

examples. In this section we show how to use it for the computation of a
feedback vertex set.

Definition 3.46. Let G = (V, E) be a graph. A set F ⊆ V is a feedback vertex set

if G − F has no cycles, that is, G − F is a forest.
32
E. Gioan and C. Paul, Split decomposition and graph labelled trees: characteriza-
tions and fully dynamic algorithms for totally decomposable graphs. Manuscript on
arXiv: 0810.1823v2, 2012.
 3.4 Distance-hereditary graphs 63

Let G be a graph. The feedback vertex set problem asks for a feedback
vertex set in G of minimal cardinality. The feedback vertex set problem is
NP-complete.

Recall (from Section 3.3.2) that a graph G is a join of two graphs G1 and
G2 , if G is obtained from G1 and G2 by making every vertex of G1 adjacent to
every vertex of G2 .

A graph G is the union of two graphs G1 and G2 if

1. the vertex set of G is the union of the vertex sets of G1 and G2 , and
2. the edge set of G is the union of the edge sets of G1 and G2 .

When G is disconnected then a feedback vertex set for G is of course

very easy to obtain from minimum feedback vertex sets of its components. In
the following two lemmas we summarize two basic observations for the case
where G is the join of two graphs.

Lemma 3.47. Assume that G is the join of two graphs G1 and G2 . For i ∈ {1, 2},
let Vi be the vertex set of Gi . Let F be a feedback vertex set of G. Then

|V1 \ F| 6 1 or |V2 \ F| 6 1 or both.

Proof. Otherwise G − F contains a 4-cycle, that is, an induced 4-cycle or a

triangle. t
u

Lemma 3.48. Let G be a cograph which is the join of cographs G1 and G2 . Let
V1 and V2 be the vertex sets of G1 and G2 . Let F be a feedback vertex set of G.
Assume that |V1 \ F| = 1. Then G2 − F is an independent set.

Proof. Otherwise G − F contains a triangle. t

u

We are ready to prove our theorem.

Theorem 3.49. There exists a linear-time algorithm that solves the feedback
vertex set problem on distance-hereditary graphs.

Proof. Let G = (V, E) be distance hereditary. First construct a decomposition

tree (T , f) for G. This takes linear time.
Extend the tree with a new root r0 and make the parent of the old root r this
new root r0 . Define the twinset of the edge {r, r0 } as ∅. For ease of description,
call this new decomposition tree again (T , f).
64 3 Graph Classes

Let p be an internal vertex of T which is not the root r0 and let c1 and c2 be the
two children of p. Let W1 and W2 be the two sets of vertices that are mapped
to the leaves in the subtrees at c1 and c2 respectively and let W = W1 ∪ W2 .
Let Q1 ⊆ W1 and Q2 ⊆ W2 be the two twinsets of {p, c1 } and {p, c2 }. The ⊗
or ⊕ label at p indicates whether there is a join or a union of the two twinsets
Q1 and Q2 .
Let e be the edge in T that connects p with its parent p0 (possibly p0 = r0 ) and
let ei = {p, ci } for i ∈ {1, 2}. The twinset Qe is either one of Q1 or Q2 , or it is
Q1 ∪ Q2 , or it is the empty set.
In our dynamic programming algorithm we maintain the following four values
for each edge e in T with twinset Q:
(1) the minimal cardinality of a feedback vertex set of G[W];
(2) the minimal cardinality of a feedback vertex set F of G[W] such that no
two vertices of Q − F are in one component of G[W] − F;
(3) the minimal cardinality of a feedback vertex set F of G[W] such that

|Q − F| = 1 and

(4) the minimal cardinality of a feedback vertex set F of G[W] with

Q ⊆ F.

It is easy to see that these four values for an edge e = {p, p0 } in T can be
obtained from the values at the edges {p, c1 } and {p, c2 } (see Exercise 3.18).
The minimal cardinality of a feedback vertex set for G can be read from
the first value i.e., Item (1) above, at the root-edge {r, r0 } of the binary tree-
decomposition.
This completes the proof. t
u

One of the oldest classes of graphs that have been studied in great detail is
the class of chordal graphs.33 We’ll give you a crash course in chordal graphs
in the next section. Brace yourselves!

3.5 Chordal graphs

Definition 3.50. A graph is chordal if it has no induced cycle of length more
than three.
33
J. Blair, B. Peyton, An introduction to chordal graphs and clique trees. In:
(A. George, J. Gilbert, J. Liu eds.) Graph theory and sparse matrix computation,
Springer-Verlag, 1993, pp. 1–29. Notice that Theorem 2.1 of this paper (the ‘proof’
is M. Golumbic’s erroneous proof) is not correct. Dirac gives a counterexample in
his paper.
 3.5 Chordal graphs 65

Notice that the class of chordal graphs contains, for example, all trees. It
may be helpful to keep this in mind; chordal graphs look a lot like ordinary
trees.

By definition, the class of chordal graphs is hereditary. Let’s first prove that
chordal graphs are perfect.

Theorem 3.51. Chordal graphs are perfect.

Proof. Let G = (V, E) be chordal. Then, by definition, G has no holes. We

show that G has no antiholes. Since C¯5 = C5 any antihole must have at least
six vertices.
Notice that any cycle of length at least six has an induced 2K2 , which is the
complement of a 4-cycle. Thus G has no antihole. t
u

Our first characterization of chordal graphs is in terms of minimal separa-

tors.34

Definition 3.52. Let G = (V, E) be a graph and let x and y be nonadjacent

vertices. A set
S ⊆ V \ {x, y}
is an x, y-separator if x and y are in different components of G − S.

Definition 3.53. An x, y-separator S is a minimal x, y-separator if no proper

subset of S is an x, y-separator.

Definition 3.54. A set S is a minimal separator if there exist nonadjacent ver-

tices x and y such that S is a minimal x, y-separator.

Remark 3.55. Notice that one minimal separator may properly contain an-
other minimal separator. For example, consider a 4-cycle [a, b, c, d]. Add a
pendant vertex e adjacent to c. Then {a, c} is a minimal b, d-separator and
{c} is a minimal a, e-separator. So, both {a, c} and {c} are minimal separators
(separating different pairs of vertices) and

{c} ⊂ {a, c}.

(See Figure 3.10.) By the way, {b, d} is another minimal a, e-separator (dis-
joint from {c}, that is, {c} ∩ {b, d} = ∅). Minimal separators for the same pair
do not have to be disjoint, but one cannot properly contain another one.
66 3 Graph Classes

a b

d c

Fig. 3.10. This graph has a minimal separator contained in another one.

Theorem 3.56. A graph is chordal if and only if each minimal separator is a

clique.

Proof. Assume that G = (V, E) is chordal. Let S be a minimal x, y-separator

for nonadjacent vertices x and y in G. Let Cx and Cy be the components of
G − S that contain x and y.
Notice that every vertex of S has at least one neighbor in Cx and at least one
neighbor in Cy . To see this, assume some z ∈ S has no neighbors in Cx . Then
S \ {z} is also a minimal x, y-separator. This contradicts the minimality of S.
Now assume that S is not a clique. Then S contains two vertices a and b that
are not adjacent. Consider two chordless paths Px and Py from a to b; one
with internal vertices in Cx and the other with internal vertices in Cy .
A chordless path is a path without a chord, that is, the path is induced.
(In Exercise 3.20 we ask you to prove that Px and Py exist.) The two paths
together form an induced cycle of length at least four.
This proves the theorem. t
u

Our second characterization of chordal graphs is in terms of simplicial

vertices.35 To get the idea, you may think of simplicial vertices as the leaves of
a tree.

Definition 3.57. Let G = (V, E) be a graph. A vertex x in G is simplicial if N(x)

induces a clique in G.

34
Theorem 1 in: G. Dirac, On rigid circuit graphs, Abhandlungen aus dem Mathema-
tischen Seminar der Universität Hamburg 25 (1961), pp. 71–76.
35
D. Rose, R. Tarjan and G. Lueker, Algorithmic aspects of vertex elimination of
graphs, SIAM Journal on Computing 5 (1976), pp. 266–283.
 3.5 Chordal graphs 67

Theorem 3.58. A graph is chordal if and only if every induced subgraph has a
simplicial vertex.
Proof. Let G = (V, E) be a graph. First assume that every induced subgraph
of G has a simplicial vertex. Let Ω be a subset of vertices such that G[Ω] is a
cycle of length at least four. Then G[Ω] is an induced subgraph of G without
simplicial vertex. This is a contradiction.
Now assume that G is chordal. Let x be a vertex such that the largest compo-
nent C of G − N[x] is as large as possible. Let S ⊆ N(x) be the set of neighbors
of x that have a neighbor in C. Then S is a minimal x, y-separator for any
y ∈ C. Thus S is a clique.
We claim that every vertex of
V \ (C ∪ S)
is adjacent to all vertices in S.
Clearly, x is adjacent to every vertex in S since S ⊆ N(x). Assume that there
exists a vertex
z ∈ V \ (S ∪ C ∪ {x})
which is not adjacent to some vertex s ∈ S. Then C ∪ {s} is contained in a
component of G − N[z] since G[C] is connected and s has a neighbor in C.
Thus G − N[z] has a component which is larger than C. This contradicts the
choice of x.
Let
G0 = G − (S ∪ C).
By induction we may assume that G0 has a simplicial vertex z. Let N0 (z) be
the neighborhood of z in G0 . Then N0 (z) is a clique. Notice that
N(z) = N0 (z) ∪ S.

Now N(z) induces a clique in G, since

(i) S is a clique, and
(ii) N0 (z) is a clique, and
(iii) every vertex of N0 (z) is adjacent to every vertex of S.
This proves that z is a simplicial vertex in G. t
u

Definition 3.59. Let G = (V, E) be a chordal graph. A perfect elimination or-

dering for G is an ordering of the vertices
[x1 , . . . , xn ]
such that for i = 1, . . . , n the vertex xi is simplicial in the subgraph of G induced
by {xi , . . . , xn }.
68 3 Graph Classes

Theorem 3.60. A graph is chordal if and only if it has a perfect elimination

ordering.

Proof. This is an immediate consequence of Theorem 3.58. t

u

In Exercise 3.24 we ask you to prove that a perfect elimination ordering

in a chordal graph can be obtained in linear time. One of the oldest and eas-
iest algorithms to do this is an algorithm of Tarjan and Yannakakis.36 Their
algorithm computes an ordering [x1 , . . . , xn ] of the vertices in any graph. This
ordering is a perfect elimination ordering if and only if the graph is chordal.
Their paper describes first a linear-time algorithm that computes an order-
ing and next it describes a linear-time test to see if the ordering is a perfect
elimination ordering.

Their algorithm computes an ordering as follows. It labels the vertices one

by one. In each step, the unlabeled vertex that has the most labeled neighbors
is labeled next. Ties are broken arbitrarily (so, the first vertex to get a label
is arbitrary). This produces the perfect elimination ordering backwards, i.e.,
the last vertex that gets a label is the first vertex in the perfect elimination
ordering.

Lemma 3.61. Every chordal graph has at most n maximal cliques, where n is
the number of vertices in the graph.

Proof. Let G = (V, E) be a chordal graph. Let x be a simplicial in G. The only

maximal clique in G that contains x is N[x]. Thus all other maximal cliques in
G are maximal cliques in G − x. The claim follows by induction. t
u

3.5.1 Clique trees

Chordal graphs have a special decomposition tree, which is called a clique

tree. These clique trees can be defined in two ways. We describe both.

Definition 3.62. Let G = (V, E) be a graph. Let (T , S) be a pair where T is a

tree and S is a collection of subsets of V which are in 1-1 correspondence with
the points of T . For a point i in T , let Si ∈ S be the subset that is assigned to the
point i.
The pair (T , S) is a clique tree for G if the following conditions are true.
(a) S is the set of maximal cliques in G, and
36
R. Tarjan and M. Yannakakis, Simple linear-time algorithms to test chordality of
graphs, test acyclicity of hypergraphs, and selectively reduce acyclic hypergraphs,
SIAM Journal on Computing 13 (1984), pp. 566–579.
 3.5 Chordal graphs 69

(b) for every vertex x ∈ V, if x is in two subsets Si and Sj then x is in every S`

for which ` lies on the path in T from i to j.

In other words, a clique tree for G is a tree of which the vertices represent
the maximal cliques in G such that, for every vertex x in G, the maximal
cliques that contain x form a subtree of T .

Theorem 3.63. A graph is chordal if and only if it has a clique tree.

Proof. First assume that the graph G = (V, E) has a clique tree (T , S). Let ` be
a leaf in T and let S` ∈ S be the maximal clique in G which is assigned to `.
We claim that S` contains a vertex which is simplicial in G.
Let p be the neighbor of ` in T . The cliques S` and Sp are maximal cliques in
G, so there must be a vertex
z ∈ S` \ Sp .
The maximal cliques that contain z form a subtree of T , by definition of the
clique tree. Since z ∈
/ Sp , the vertex z is contained only in one maximal clique
S` in G. Thus z is simplicial.
Assume that G = (V, E) is a chordal graph. We may assume that G is not a
clique, otherwise we are done. Let S be a minimal separator in G such that |S|
is as small as possible.
Let C1 , . . . , Ct be the components of G − S. Since S is a minimal separator,
there are at least two components. We claim that every vertex in S has a
neighbor in each Ci . Assume that s ∈ S has no neighbors in C1 .
Let
S0 = S \ {s}.
Then C1 is a component of G − S0 . Thus S0 is an x, y-separator for a pair

x ∈ C1 and y ∈ ∪ti=2 Ci .

This contradicts the choice of S.

Consider clique trees Ti for the subgraphs Gi of G induced by

Ci ∪ S for i ∈ {1, . . . , t}.

Since S is a clique, it is contained in a maximal clique Mi in Gi . We leave it

as an exercise to check that Mi 6= S, since G is chordal.
Construct a clique tree T for G as follows. For i = 2, . . . , t make the vertex in
Ti that represents Mi adjacent to the vertex in T1 that represents M1 .
We prove that this is a clique tree for G. Let x ∈ Ci . The maximal cliques that
contain x are all contained in Gi . Thus these form a subtree of Ti .
70 3 Graph Classes

Now let x ∈ S. The cliques in Gi that contain x form a subtree of Ti and Mi is

one of them. By the construction, all the maximal cliques in G that contain x
form a subtree of the clique tree T for G.
This proves the theorem. t
u

1 2 6 1 2 6
2
3 7
2 2
3 7 3 7
5 5
3 7
5 5
4 8
4 5 8

Fig. 3.11. This figure shows a chordal graph and a clique tree.

The other way to describe clique trees is as follows.

Theorem 3.64. A graph G = (V, E) is a chordal graph if and only if it is the

intersection graph of a collection of subtrees of a tree. By that we mean that there
exists a tree T and a collection of subtrees

{ Tx | x ∈ V }

such that two vertices x and y of G are adjacent if and only if Tx and Ty have at
least one vertex of T in common.
Proof. First assume that G is chordal. Consider a clique tree (T , S) for G. Let
x be a vertex of G. By definition of the clique tree, the maximal cliques in G
that contain x form a subtree Tx of T .
Assume that two vertices x and y are adjacent in G. The edge {x, y} is con-
tained in some maximal clique in G. So Tx ∩ Ty 6= ∅.
Assume that x and y are two vertices and assume that p ∈ Tx ∩ Ty . The
maximal clique Sp which is assigned to the vertex p in T contains x and y.
Thus x and y are adjacent.
Now assume that G is the intersection graph of a collection of subtrees of a
tree T . Consider a leaf ` of T . If ` is not in any subtree Tx , then we may remove
the leaf from T .
If every subtree that contains ` contains also the neighbor of ` in T , then we
may remove ` from T .
 3.5 Chordal graphs 71

Finally, assume that there is a tree Tx which consists of the single vertex `. If
y and z are two neighbors of x, then Ty and Tz both contain `, and so y and
z are adjacent. That means that x is a simplicial vertex. Now remove x from
the graph and Tx from the collection of subtrees. It follows by induction that
G has a perfect elimination ordering. By Theorem 3.60, G is chordal.
This proves the theorem. t
u

3.5.2 Algorithms for independent set, clique and vertex coloring in

chordal graphs

Theorem 3.65. There exists a linear-time algorithm to compute a maximum

independent set in chordal graphs.

Proof. Let G = (V, E) be a chordal graph. Let s be a simplicial vertex in G.

We first prove that there exists a maximum independent set I in G which
contains s.
To see this, let I be a maximum independent set and assume that s ∈
/ I. If

N(s) ∩ I = ∅,

then I ∪ {s} is also an independent set, which is a contradiction.

Since s is a simplicial vertex in G, N(s) induces a clique in G. Thus

|N(s) ∩ I| = 1.

Let u ∈ N(s) ∩ I. Then consider

I0 = (I \ {u}) ∪ {s}.

Then I0 is also a maximum independent set and s ∈ I0 .

The algorithm computes a maximum independent set in G as follows. Take
any simplicial vertex s and start with I = {s}. Now let

G0 = G − N[s].

Then G0 is chordal, and by induction there exists a linear time algorithm that
computes α(G0 ). Then
α(G) = 1 + α(G0 ).
This proves the theorem. t
u

Theorem 3.66. There exists a linear-time algorithm that computes ω(G) for
chordal graphs G.
72 3 Graph Classes

Proof. Let G = (V, E) be a chordal graph. Then, by Theorem 3.60 on Page 68,
G has a perfect elimination ordering

π = [x1 , . . . , xn ].

Let
Ni = |N[xi ] ∩ {xi , . . . , xn }|.
Then
ω(G) = max { Ni | i ∈ {1, . . . , n} }.
This proves the theorem. t
u

Let G = (V, E) be a chordal graph. Then G is perfect, and χ(G) = ω(G).

By Theorem 3.66 there exists a linear-time algorithm that computes χ(G). An
actual coloring is also very easy to obtain, as we show next.
Theorem 3.67. There exists a linear-time algorithm that computes a vertex col-
oring for G with χ(G) colors for chordal graphs G.
Proof. Let G = (V, E) be a chordal graph. Let [x1 , . . . , xn ] be a perfect elim-
ination ordering for G. We color the vertices of G greedily with ω(G) colors
from the color set
Ω = { 1, . . . , ω(G) }
as follows.
For i = n down to 1, color the vertex xi with an arbitrary color from Ω that
is not used by vertices in

N(xi ) ∩ { xi+1 , . . . , xn }. (3.29)

To see that this is possible, notice that the set in Formula (3.29) contains at
most ω(G) − 1 vertices. Thus there is a color in Ω available to color xi .
The claim follows by induction. t
u

3.6 Interval graphs

As far as practical applications are concerned, the class of graphs that steals
the show is the class of interval graphs. Indeed, practical applications range
from archeology, sociology, scheduling classes at school, DNA-sequencing
problems in biology, time-schedules for airliners, et cetera, etc, &tc.

In this section we have a short look at the definition and some of the most
important properties of interval graphs.
 3.6 Interval graphs 73

1 4

2 7

3 5

2 4
3 7
1 5 6

Fig. 3.12. The figure shows an interval graph.

Definition 3.68. A graph G is an interval graph if it is the intersection graph

of a collection of intervals on the real line. By that we mean that there is an
interval Ix for every vertex x in G such that two vertices x and y are adjacent in
G if and only if Ix ∩ Iy 6= ∅.

Remark 3.69. We only look at finite graphs. Then it is easy to see that we
can have intervals such that no two endpoints coincide. Thus the question
whether the intervals are closed or open is not an issue. Also, we may assume
that all intervals are finite, that is, they have finite length.

Theorem 3.70. Interval graphs are chordal.

Proof. It is not difficult to see that one cannot construct any cycle of length
more than three as the intersection graph of intervals. If one interval Ix is
completely contained in some other interval Iy then, for the corresponding
vertices x and y in G we have
N[x] ⊆ N[y].

Consider a cycle [x1 , . . . , xk ]. Suppose it has an interval representation. By the

previous observation we may assume that
`1 < `2 < r1 < `3 < r2 < `4 < r3 < . . . ,
where `i and ri are the left and right endpoint of the interval Ii that represents
xi . Now, Ik intersects I1 and Ik−1 , but then xk is adjacent to all vertices in
{x1 , . . . , xk−1 }. Thus it is a chordless cycle (that is an induced cycle) only for
k = 3. t
u
Corollary 3.71. Interval graphs are perfect.

By Theorem 3.70 every interval graph is chordal and by Theorem 3.63 on

Page 69 every chordal graph has a clique tree. We show next that interval
graphs have a clique tree which is a path.37
37
R. Halin, Some remarks on interval graphs, Combinatorica 2 (1982), pp. 297–304.
74 3 Graph Classes

The clique tree for interval graphs is usually called

a consecutive clique arrangement.

Theorem 3.72. An interval graph G has a consecutive clique arrangement, that

is, there is a linear ordering
[M1 , . . . , Mt ]
of the maximal cliques in G such that for every vertex x, the maximal cliques
that contain x form a subsequence.

Proof. Consider an interval model for interval graph G = (V, E). Scan the
real line from left to right. At each point w on the real line, consider all the
intervals that contain w. Let

M(w) = { x ∈ V | w ∈ Ix }.

Then M(w) is a clique in G (possibly empty).

We need to prove the ‘Helly property for intervals,’ that is, if M is a clique in
G then there exists a w ∈ R such that each interval that represents a vertex in
M contains w.
If |M| = 3 then this is easy to check (see Exercise 3.28).
Assume that |M| > 3. Let x and y be two elements of M. Replace the intervals
Ix and Iy by Ix ∩ Iy . Each pair of intervals of this new collection intersects,
by the previous observation. By induction, there exists a w ∈ R which is
contained in every interval of this collection. Then w is also contained in
every interval that represents a vertex of M.
We now have that every maximal clique is in the set

{ M(w) | w ∈ R }.

This defines the linear order of the maximal cliques in G: For each maximal
clique M fix a real number w such that w is in each interval which represents
a vertex of M. By the finiteness of the system, we can choose the real numbers
such that no two maximal cliques are represented by the same real number. If
M(w1 ) and M(w2 ) are two maximal cliques, then M(w1 ) precedes M(w2 ) if
and only if w1 < w2 . Obviously, since Ix is an interval, for every vertex x the
maximal cliques that contain x are consecutive in this ordering. t
u

Definition 3.73. An asteroidal triple {x, y, z}, AT for short, in a graph G is a set
of three pairwise nonadjacent vertices in G such that for every pair of them there
is a path connecting them that avoids the neighborhood of the third.
 3.6 Interval graphs 75

Remark 3.74. Notice that, if some x, y-path contains no neighbor of z then it

also does not contain z, since z is not adjacent to x nor y.

Remark 3.75. For example, consider a 6-cycle. Let x, y and z be three ver-
tices that are pairwise not adjacent. Then {x, y, z} is an asteroidal triple. More
examples can be found in Figure 3.3 on page 45.
Actually, Gallai found those graphs that contain asteroidal triples (and that
are minimal with respect to the induced subgraph relation) and which do not
contain C5 . Ekkehard Köhler completed the list.38 The remaining graphs are
depicted in Figure 3.13.

Fig. 3.13. This figure shows the remaining minimal graphs that contain asteroidal
triples. These remaining ones are those that contain a C5 . The set of all minimal graphs
that contain asteroidal triples consists of the graphs in Figure 3.3 on page 45 plus the
ones in this figure.

Lemma 3.76. Interval graphs are AT-free, that is, they have no asteroidal triple.

Proof. Let G = (V, E) be an interval graph. Consider an interval model for G.

take three intervals Ix , Iy and Iz which pairwise do not intersect. Then one of
the intervals is between the other two. Say Ix is left of Iy and Iy is left of Iz .
Let P = [x = x1 , . . . , xk = z] be an x, z-path. Assume that y ∈/ P. Then an in-
terval Ii of some vertex xi must intersect Iy . Thus {x, y, z} is not an asteroidal
triple. t
u

38
Ekkehard Köhler, Graphs without asteroidal triples, PhD Thesis, TU-Berlin, 1999.
76 3 Graph Classes

Lekkerkerker and Boland characterize interval graphs as those chordal

graph that have no asteroidal triple.39 Halin gave an elegant proof of the
theorem. We present Halin’s proof. 40

Theorem 3.77 (Lekkerkerker and Boland, Halin). A graph G is an interval

graph if and only if G is chordal and has no asteroidal triple.

Proof. Assume that G is chordal and that it has no asteroidal triple. We prove
that G has a consecutive clique arrangement.
By Theorem 3.63 on page 69, G has a clique tree. We prove that this tree can
be turned into a path. We use induction on the number of maximal cliques.
When G is a clique then we are done. Assume that G has k+2 maximal cliques
and assume that for all AT-free chordal graphs with less maximal cliques the
claim is true.
Consider a clique tree for G. Let C be a maximal clique which is mapped to
a leaf. Remove that leaf from the clique tree. The remaining clique tree is
the clique tree of a chordal graph G0 . The vertices of G0 are the vertices of
G minus the simplicial vertices that appear only in C. Repeating the process,
each time removing a leaf from the clique tree, gives a sequence of cliques, in
reversed order:
C0 , C1 , . . . , Ck , Ck+1 = C. (3.30)
Let H be the graph induced by
k
[
Ci . (3.31)
i=0

Then H is chordal and it contains no asteroidal triple. By induction, H is an

interval graph and we may assume that the cliques can be arranged consecu-
tively, say
[ C0 , . . . , Ck ]. (3.32)

Define
Si = (C0 ∪ · · · ∪ Ci−1 ) ∩ Ci , for i ∈ { 1, . . . , k }. (3.33)
We write
S = (C0 ∪ . . . , Ck ) ∩ C.

If S ⊂ C0 or S ⊂ Ck then we are done. In that case we can place C before C0

or after Ck . This gives a consecutive clique arrangement for G.

39
C. Lekkerkerker and J. Boland, Representation of a finite graph by a set of intervals
on the real line, Fundamenta Mathematicae 51 (1962), pp. 45–64.
40
R. Halin, Some remarks on interval graphs, Combinatorica 2 (1982), pp. 297–304.
 3.6 Interval graphs 77

Notice that S is contained in at least one clique of H. Assume that S ⊂ Cm for

0 < m < k. We now consider two cases.

In the first case, assume that Si ⊆ S for at least one i ∈ {1, . . . , k}. Define
S0 = Sk+1 = ∅ and
p = max { i | Si ⊆ S and 06i6m}
q = min {i | Si ⊆ S and m + 1 6 i 6 k + 1 }.

Since we assume that S 6⊆ C0 and S 6⊆ Ck we have that

p>1 or q 6 k. (3.34)

Let
Z = Cp ∪ . . . ∪ Cm ∪ . . . ∪ Cq−1 . (3.35)
Then Z \ S is the component of H − S which contains Cm \ S. Notice that,
Sp ⊆ Cj for j ∈ {p, . . . , m}, and Sq ⊆ Cj for j ∈ {m, . . . , q − 1}. (3.36)
This follows because Sp ⊆ S ⊆ Cm and, by definition also Sp ⊆ Cp . Similarly,
Sq ⊆ S ⊆ Sm and Sq ⊆ Cq , which yields the second series of relations.
Notice that the maximal cliques of Z ∪ C are Cp , . . . , Cq−1 and C, and so
Z ∪ C has less maximal cliques than G. Thus we may assume that Z ∪ C has a
consecutive clique arrangement. Let this consecutive clique arrangement be
[ Z0 , . . . , Zt ]. (3.37)

Since S does not separate Z ∪ C, we have that

C = Z0 or C = Zt . (3.38)
Without loss of generality, assume that C = Z0 .
Assume that Zt = Cσ for some σ ∈ {p, . . . , q−1}. If σ 6 m then Sp ⊆ Cσ = Zt
and we have a consecutive clique arrangement
[ C0 , . . . , Cp−1 , Zt , . . . , Z1 , C, Cq , . . . , Ck ]. (3.39)
When σ > m + 1 then Sq ⊆ Cσ = Zt , and we have a consecutive clique
arrangement
[ C0 , . . . , Cp−1 , C, Z1 , . . . , Zt , Cq , . . . , Ck ]. (3.40)

In the second case, assume that

Si 6⊆ S for i ∈ {1, . . . , k}. (3.41)
Assume that
78 3 Graph Classes

(i) S 6⊆ Sk and Si 6⊆ Sk for i ∈ {1, . . . , k − 1}, and

6 S1 and Si 6⊆ S1 for i ∈ {2, . . . , k}.
(ii) S ⊆
Then there exist vertices

x ∈ C0 \ S1 y ∈ Ck \ Sk z ∈ C \ S. (3.42)

Then {x, y, z} is an asteroidal triple.

If Item (i) does not hold then the first case applies, with Ck and Sk instead of
C and S. If Item (ii) does not hold then the first case applies with C0 and S1
instead of C and S.
This proves the theorem. t
u

Lekkerkerker and Boland obtained a list of forbidden induced subgraphs

for the class of interval graphs.41

Theorem 3.78 (Lekkerkerker and Boland). A graph is an interval graph if

and only if it does not contain any graph of Figure 3.14 as an induced subgraph.

I II IIIn(n points; n>4)

_

_
IVn(n+4 points; n>2) _
Vn(n+5 points; n>1)

Fig. 3.14. The forbidden induced subgraphs for interval graphs. Notice that all of
them, except C4 and C5 are contained in Gallai’s table Figure 3.3 on page 45. The C5
appears in Figure 3.2 on page 44. The C4 is a comparability graph and so it does not
appear in Gallai’s list.

41
C. Lekkerkerker and J. Boland, Representation of a finite graph by a set of intervals
on the real line, Fundamenta Mathematicae 51 (1962), pp. 45–64.
 3.6 Interval graphs 79

Let G = (V, E) be a graph. Recall Section 3.2. An orientation directs every

edge {x, y} from x to y or from y to x. The orientation is transitive if the
following holds true for every three vertices x, y and z. If there is a directed
edge (x, y) from x to y and a directed edge (y, z) from y to z, then x and z are
adjacent in G and the edge {x, z} is directed from x to z.

Definition 3.79. A graph is a comparability graph if there exists a transitive

orientation of its edges.

Lemma 3.80. If G is an interval graph then its complement Ḡ is a comparability

graph.

Proof. Let G = (V, E) be an interval graph. Consider an interval model for

G. Let x and y be two nonadjacent vertices in G. Then the interval Ix lies
completely to the left, or completely to the right of the interval Iy . Say we
direct the edge {x, y} of Ḡ from x to y if Ix lies to the left of Iy . We claim that
this is a transitive orientation of Ḡ.
We need to check that, if (x, y) and (y, z) are arcs pointing from x to y and
from y to z, then there is an edge {x, z} in Ḡ and it is oriented from x to z.
I’m sure you agree that this is obvious: If Ix lies to the left of Iy and Iy lies to
the left of Iz then Ix lies to the left of Iz . t
u

Actually, Lemma 3.80 ‘almost’ characterizes interval graphs. Gilmore and

Hoffman proved the following characterization.42

Theorem 3.81 (Gilmore and Hoffman). A graph G is an interval graph if and

only if G has no induced 4-cycle and Ḡ is a comparability graph.

Proof. By Lemma 3.80 we only need to prove one direction. Assume that G is
a graph without induced C4 and assume that Ḡ has a transitive orientation.
We first prove that G is chordal.
Assume that G has a hole H, that is, H is an induced cycle in G of length at
least five. First notice that a transitive orientation of Ḡ induces a transitive
orientation of H.
Number the vertices of H consecutively [x1 , . . . , xt ]. First consider the neigh-
bors of x1 in Ḡ, that is
{x3 , . . . , xt−1 }.
Without loss of generality, assume that the edge {x1 , x3 } is directed from x1 to
x3 . Then, by transitivity, the edge {x1 , x4 } is directed from x1 to x4 , otherwise
42
P. Gilmore and A. Hoffman, A characterization of comparability graphs and of in-
terval graphs, Canadian Journal of Mathematics 16 (1964), pp. 539–548.
80 3 Graph Classes

there would have to be an edge {x3 , x4 } in Ḡ. Repeating the argument we find
that all edge {x1 , xi } for i ∈ {3, . . . , t − 1} are directed from x1 to xi .
Consider the neighbors of x2 in Ḡ, that is the set {x4 , . . . , xt }. Since {x1 , xt−1 }
is directed from x1 to xt−1 and since {x1 , x2 } is not an edge in G, by transitivity
the edge {x2 , xt−1 } in Ḡ is directed from x2 to xt−1 . (Notice that here we use
the fact that t > 4.) As above, we find that all edges {x2 , xi } for i ∈ {4, . . . , t}
are directed from x2 to xi .
We can repeat the argument, and we find that for all vertices xi of H, all the
edges {xi , xj } of Ḡ are directed from xi to xj , which is of course a contradic-
tion.
We conclude that G is chordal. From Theorem 3.20 on page 45 it follows
easily that G has no asteroidal triple. (See Exercise 3.36.) By Theorem 3.77
we conclude that G is an interval graph. t
u

We summarize some results of Halin, Lekkerkerker and Boland, and

Gilmore and Hoffman as follows.
Theorem 3.82. Let G be a graph. The following statements are equivalent.
(a) G is an interval graph.
(b) G is AT-free and chordal.
(c) G has no induced C4 and Ḡ is a comparability graph.
(d) For every three maximal cliques M1 , M2 and M3 in G there is one that
separates the others. Here we say that M1 separates M2 and M3 if M2 \ M1
and M3 \ M1 are contained in different components of G − M1 .

3.7 Permutation graphs

A permutation diagram is obtained as follows. Let L1 and L2 be two horizontal
lines in the plane, one above the other. Label n points on the topline and on
the bottom line by 1, 2, . . . , n. Connect each point on the topline by a straight
linesegment to the point with the identical label on the bottom line. A graph
G is a permutation graph if it is the intersection graph of the linesegments of
a permutation diagram.43 By that we mean that the vertices of the graph G
are the linesegments of the permutation diagram and two vertices in G are
adjacent if the two linesegments cross each other.

If G is a permutation graph then its complement Ḡ is also a permutation

graph. This is easy to see; simply reverse the ordering of the points on one of
the two horizontal lines.

43
A. Pnueli, A. Lempel and S. Even, Transitive orientation of graphs and identification
of permutation graphs, Canadian Journal of Mathematics 23 (1971), pp. 160–175.
 3.7 Permutation graphs 81

1 1 2 3 4 5

5 3

4 2 3 5 1 4 2
Fig. 3.15. A permutation graph and its permutation diagram

Lemma 3.83. Permutation graphs are perfect.

Proof. It is easy to check that one cannot construct a permutation diagram for
a graph which is a cycle of length more than four. Thus permutation graphs
have no holes.
A permutation graph G can also not have an antihole, since Ḡ is also a per-
mutation graph. t
u

Recall Definition 3.79. A comparability graph is a graph which has a tran-

sitive orientation; so we can direct any edge {x, y} either from x to y or from
y to x such that the directed graph is transitive.

In the proof of Theorem 3.20 on page 45 we introduced the dimension

of a poset, and in Remark 3.21 we promised that we would have a closer
look at posets of dimension two. By presenting the next theorem we keep that
promise.44

Theorem 3.84. A graph G is a permutation graph if and only if G and Ḡ are

comparability graphs.
Proof. Let G be a permutation graph. Consider a permutation diagram for
G. Notice this: for any independent set in G the corresponding linesegments
are noncrossing. So they can be ordered left to right, which is of course a
transitive ordering. By that we mean that, if a linesegment x is to the left of a
linesegment y, and the linesegment of y is to the left of a linesegment z, then
the linesegment of x is to the left of the linesegment z.
This shows that Ḡ is a comparability graph. Then G is also a comparability
graph since the class of permutation graphs is closed under complementa-
tions.
44
K. Baker, P. Fishburn and F. Roberts, Partial orders of dimension 2, Networks 2
(1971), pp. 11–28.
82 3 Graph Classes

Let G be a graph such that G and Ḡ are both comparability graphs. We show
that G is a permutation graph by constructing a permutation diagram.
Let F1 and F2 be transitive orientations of G and Ḡ. We claim that F1 ∪ F2 is
an acyclic orientation of the complete graph. Otherwise there is a directed
triangle, and so two edges in the triangle are directed according to one of F1
and F2 and the third is directed according to the other one of F1 and F2 . But
this contradicts the transitivity of F1 or the transitivity of F2 .
Likewise, F−1
1 ∪ F2 is acyclic. Order the vertices on the topline according to
F1 ∪ F2 and the vertices on the bottom line according to F−1
1 ∪ F2 . It is easy to
check that this yields the permutation diagram. t
u

3.7.1 Interval containment graphs

The following characterization of permutation graphs illustrates the relation

of this class of graphs to the class of interval graphs. Consider a collection
of intervals on the real line. Construct a graph of which the vertices are the
intervals and make two vertices adjacent if one of the two intervals contains
the other. Such a graph is called an interval containment graph.

Dushnik and Miller proved that classes of interval containment graphs and
permutation graphs are the same.45

Theorem 3.85. A graph is a permutation graph if and only if it is an interval

containment graph.
Proof. Consider a diagram of a permutation graph. When one moves the bot-
tom line to the right of the topline then the linesegments in the diagram
transform into intervals. It is easy to check that two linesegments intersect if
and only if one of the intervals is contained in the other one. This proves that
permutation graphs are interval containment graphs.
Now consider an interval containment graph. Construct a permutation dia-
gram as follows. Put the left endpoints of the intervals in order on the topline
and the right endpoints in order on the bottom line of the diagram. Then one
interval is contained in another interval if and only if the two linesegments
intersect. This proves that every interval containment graph is a permutation
graph. t
u

Tedder, et al. proved the following theorem.46

45
B. Dushnik and E. Miller, Partially ordered sets, American Journal of Mathematics 63
(1941), pp. 600–610.
46
M. Tedder, D. Corneil, M. Habib and C. Paul, Simpler linear-time modular de-
composition via recursive factorizing permutations, Proceedings ICALP’08, Springer,
LNCS 5125 (2008), pp. 634–645.
 3.7 Permutation graphs 83

Theorem 3.86. Permutation graphs can be recognized in linear time. If G is a

permutation graph then this algorithm can be used to construct a permutation
diagram in linear time.

Remark 3.87. Consider a collection S of subsets of some universal set U. One

can construct a graph such that each element of S is represented by a vertex
and in which two vertices are adjacent if one of the two subsets contains
the other one. It is easy to see that the containment graphs are exactly the
comparability graphs.47 48

3.7.2 Cliques and independent sets in permutation graphs

To end this chapter smoothly, we’ll have a quick and easy look at the clique
and independence number of permutation graphs. Since permutation graphs
are perfect, we get the chromatic number and the clique cover number for
free.

Let G be a permutation graph. Consider a permutation diagram for G. Let

the vertices on the topline be numbered from left to right as 1, . . . , n. Let
π = [π1 , π2 , . . . , πn ]
be the sequence of numbers as they appear in left to right order on the bottom
line.

Two vertices i and j with i < j are adjacent if i appears to the right of j on
the bottom line. Thus a clique in G corresponds with a decreasing sequence
in π.

This shows that finding ω(G) is equivalent to finding the longest decreas-
ing subsequence in π. Given the sequence π this can be done very efficiently.49

Let G = (V, E) be a permutation graph. Then Ḡ is also a permutation

graph and α(G) = ω(Ḡ).
Theorem 3.88. There exist O(n log log n) algorithms to compute α(G) and
ω(G) for permutation graphs G. Here we assume that the permutation π is
a part of the input.
47
W. Trotter, Jr., Combinatorics and partially ordered sets – Dimension theory, The
Johns Hopkins University Press, Studies in the Mathematical Sciences, Baltimore,
Maryland, 1992.
48
J. Urrutia, Partial orders and Euclidean geometry. In: (I. Rival, ed.) Algorithms and
order, Kluwer Academic Publishers, Dordrecht, The Netherlands, 1987, pp. 327–
436.
49
J. Hunt and T. Szymanski, A fast algorithm for computing longest common subse-
quences, Communications of the ACM 20 (1977), pp. 350–353.
84 3 Graph Classes

3.8 Problems
3.1. Let G be a 5-cycle. Is G perfect?

3.2. Let B be the class of bipartite graphs. Is B hereditary?

3.3. A graph G is bipartite if χ(G) 6 2. Show that G is perfect.

3.4. Prove that a graph G is bipartite if and only if all cycles in G are even. Is
it also true that G is bipartite if and only if all induced cycles are even?

3.5. Let G = (V, E) be a graph. A vertex cover in G is a set C of vertices such

that every edge in G has at least one endpoint in C. Let τ(G) be the cardinality
of a smallest vertex cover in G. Prove that

α(G) + τ(G) = n where n = |V|. (3.43)

3.6. A P3 is a path with three vertices. Let G be a graph without induced P3 .

Show that G is a union of cliques.

3.7. Let G be a comparability graph. Design a linear-time algorithm to com-

pute ω(G).
Hint: When G is a comparability graph then a transitive orientation can be
obtained in linear time. You may assume that the graph has a transitive ori-
entation.50

3.8. Is a cograph a comparability graph?

3.9. Show that a graph G is a cograph if and only if it has a cotree.

3.10. Let G = (V, E) be a graph. Two vertices x and y are twins if either

N[x] = N[y] or N(x) = N(y).

(a) Show that every cograph with at least two vertices has a twin.
(b) Show also the converse: If every induced subgraph of a graph G has either
only one vertex, or else has a twin, then G is a cograph.
Hint: Consider a cotree (T , f) of G. Let x and y be two vertices in G that are
mapped by f to two leaves in T that have the same parent in T . Prove that x
and y are twins in G.

3.11. Show that a cograph with n vertices can have 3n/3 maximal cliques.
Hint: Show that the complement of the union of n3 disjoint triangles is a
cograph.

50
M. Golumbic, Algorithmic graph theory and perfect graphs, Elsevier, Annals of Dis-
crete Mathematics 57, Amsterdam, 2004.
 3.8 Problems 85

3.12. In Section 3.3.2 on Page 56 we showed how to compute the clique

number in cographs in linear time. Design a linear-time algorithm to compute
α(G), χ(G) and κ(G).
Hint: Use the fact that G is perfect and that Ḡ is a cograph.

3.13. Check that the house, hole, domino and gem depicted in Figure 3.8 on
Page 58 are not distance hereditary.

3.14. Let G = (V, E) be distance hereditary and let G0 be the graph obtained
from G by adding one vertex x to G which is either isolated, or a pendant
vertex with a neighbor y ∈ V, or a twin of some vertex z ∈ V. Show that G0 is
distance hereditary.

3.15. Design an algorithm that constructs a DH-decomposition tree for a tree.

3.16. Is every distance-hereditary graph a comparability graphs?

3.17. Let G be distance hereditary. Prove that, for every vertex x and every
integer k, the set of vertices at distance k from x induces a cograph.

3.18. Finish the proof of Theorem 3.49 on Page 63 by showing how the four
values, that are mentioned in the proof, are computed for an edge e in the
decomposition tree T .

3.19. Which of the graphs in Figure 3.8 on Page 58 is chordal?

3.20. Finish the proof of Theorem 3.56 by showing that the paths Px and Py
exist.
Hint: There is an a, b-path Px with internal vertices in Cx since a and b both
have a neighbor in Cx and G[Cx ] is connected. You need to show how to get
rid of all the shortcuts.

3.21. Let G be a chordal graph and let S be a minimal x, y-separator for non-
adjacent vertices x and y. Let Cx be the component of G − S that contains the
vertex x. Then Cx contains a vertex x0 such that

S ⊆ N(x0 ).

3.22. Let G = (V, E) be a chordal graph and let S ⊂ V be a separator in G.

Show that S is a minimal separator if and only if there exist two components
C1 and C2 in G − S such that every vertex of S has at least one neighbor in C1
and in C2 , that is,

∀s∈S N(s) ∩ C1 6= ∅ and N(s) ∩ C2 6= ∅.

3.23. Consider a tree T . If T is not a clique then T has two simplicial vertices
which are not adjacent. Prove the analogue for chordal graph: If G is a chordal
graph and if G is not a clique, then G has two nonadjacent simplicial vertices.
86 3 Graph Classes

3.24. Design an algorithm to compute a perfect elimination ordering of a

chordal graph in linear time.
Hint:Number the vertices from n to 1 in decreasing order. For the next ver-
tex to number, choose one that has the largest number of neighbors that are
already numbered, braking ties arbitrarily.
This exercise is not easy; perhaps you like to have a look at Tarjan and Yan-
nakakis’ paper.

3.25. The proof of Theorem 3.66 is perhaps a bit sketchy. Rub the sleep out
of your eyes and make sure that you agree with all the details.

3.26. Show that the coloring algorithm of Theorem 3.67 on Page 72 can be
implemented to run in linear time.

3.27. Design a polynomial-time algorithm which checks if a graph has an

asteroidal triple.
Hint: Try all triples {x, y, z} of pairwise nonadjacent vertices in a graph G.
Notice that there is an x, z-path that avoids N(y) if and only if x and z are
both contained in a component of G − N(y). What is the timebound of your
algorithm?

3.28. Consider three intervals I1 , I2 and I3 on the real line. If any two have a
nonempty intersection then

I1 ∩ I2 ∩ I3 6= ∅.

3.29. Prove that the Helly property also holds for a collection of subtrees of a
tree. That is the following. Let T be a tree, and let S be a collection of subtrees
of T such that

∀T1 ∈S ∀T2 ∈S T1 and T2 have at least one vertex of T in common.

Then there is a vertex of T which is a vertex of every subtree of S.

3.30. Design an efficient algorithm to compute a feedback vertex set on inter-

val graphs. What is the timebound for your algorithm? Extend the algorithm
such that it works on chordal graphs.
Hint: Use the clique tree. How many vertices can a clique have that are not
in the feedback vertex set?

3.31. Let G = (V, E) be a chordal graph. By Theorem 3.64 there exist a tree
T and a collection of subtrees

T = { Tx | x ∈ V }

such that any two vertices x and y in G are adjacent if and only if Tx ∩ Ty 6= ∅.
We say that G is the intersection graph of T.
Make a subtree Te for every edge e = {a, b} ∈ E in G by defining
 3.8 Problems 87

T e = Ta ∪ Tb .

Consider the intersection graph G∗ of the subtrees

E = { Te | e ∈ E }.

Prove that G∗ is the square of the linegraph L(G) of G. By that we mean that
the vertices of G∗ are the edges of G, and two vertices e1 and e2 of G∗ are
adjacent if and only if

∃e∈E e1 ∩ e 6= ∅ and e2 ∩ e 6= ∅.

This proves that, if G is chordal then G∗ is also chordal.

3.32. Prove that every cograph G is a permutation graph by showing that

there exists a permutation diagram for G.

3.33. Check that the circled vertices in Figure 3.3 on page 45 form asteroidal
triples.

3.34. Point out an asteroidal triple in each graph in Figure 3.13 on page 75.

3.35. Are permutation graphs AT-free?

3.36. Are cocomparability graphs AT-free?

3.37. Consider a circle C drawn in the plane. A chord of C is a straight line-

segment that connects two points on C. A circle graph G is the intersection
graph of a collection of chords in a circle. By that we mean that every vertex
of G is represented by a chord and that two vertices are adjacent in G if and
only if the two chords intersect. The diagram that represents G is called a
circle diagram. (See Figure 3.16 on the next page.)
(i) Show that every permutation graph is a circle graph.
(ii) Prove that every distance-hereditary graph is a circle graph.
Hint: Use the fact that a graph is distance-hereditary if and only if every
induced subgraph has an isolated vertex, or a pendant vertex, or a twin.
Prove by induction that a distance-hereditary graph has a circle diagram.
(iii) Prove that circle graphs are closed under local complementations. By
that we mean the following. Let G be a circle graph and let x be a vertex
of G. Consider the graph G0 obtained from G by replacing the graph
induced by N(x) by its complement.
Hint: Consider reversing the order of the endpoints of chords on one of
the two parts of the circle bounded by the endpoints of the chord that
represents x.
88 3 Graph Classes

4 1 2
1 4 7 3
6
4
1
7 6 2 3
5

5 3
5 7 6
2

Fig. 3.16. A circle graph and its circle diagram.

4
Fixed-parameter Algorithms

For most NP-complete problems one can attach a parameter to the problem
and consider the parameterized version of the problem.

For example, consider the clique number ω(G) of a graph G. The clique
number problem asks to find the maximal value ω for which the graph G has
a clique with ω vertices.

Suppose that, instead, we ask the following.

Given a graph G and some number k.
Is ω(G) > k?

Any algorithm, exponential or not, that solves this parameterized clique

number problem, can be used to solve the clique number problem: Just check
for k = 1, . . . , n whether G has a clique of k vertices, and determine the
largest k for which it has one.

Just coming up with some parameter like that doesn’t help much, of
course. Well, maybe it does, a little bit. Suppose that ω(G) is at most two.
For example, any bipartite graph has clique number at most two, so there are
lots of them.

It is easy to check if ω(G) > 3: just run the following algorithm.

1. If G is an independent set then ω(G) = 1. Of course, it takes only linear
time to check if G has an edge or not.
2. When G has at least one edge, then the next step is to check if G has a
triangle {x, y, z}. If so, then ω(G) is at least three. If not, then ω(G) < 3.
To check if G has a triangle, we can check for every three vertices x, y and
z if {x, y, z} is a triangle or not. There are O(n3 ) triples to check. If we use
the adjacency matrix to represent G we can check if a given triple {x, y, z}
90 4 Fixed-parameter Algorithms

is a triangle or not in constant time, since we only need to check if the

three pairs {x, y}, {x, z} and {y, z} are three edges or not. So we can check
in O(n3 ) time if ω(G) > 3 or not.

Of course, we can generalize this. We can check in O(k2 · nk ) time if

ω(G) > k or not. Namely, to check if G has a clique with k vertices we can
try all subsets with k vertices and see if one of them is a clique. In a graph G
with n vertices there are O(nk ) subsets with k vertices. Let Ω = {x1 , . . . , xk }
be such a subset. To check if Ω is a clique, check if every pair in Ω is adja-
cent. There are O(k2 ) pairs, so, when we use the adjacency matrix to check
adjacencies, this algorithm can be implemented to run in O(k2 · nk ) time.

The above algorithm shows that for each CONSTANT k, the algorithm
to check if ω(G) > k runs in polynomial time. So we have a polynomial
algorithm to check if a graph G has a clique with three vertices, or a clique
with 10 vertices, or even if G has a clique with one million vertices. In fact,
for each CONSTANT k, the algorithm to check if ω(G) > k runs in polynomial
time.

Of course, we can say that the algorithm is polynomial only for constant
√ k.
An algorithm that runs in O(nn ), or an algorithm that runs in O(n n ) is not
polynomial. Even an algorithm that runs in O(nlog log(n) ) is not polynomial.
The inverse Ackermann function α(n) is one of the slowest growing, un-
bounded functions that exist. For example, when n is the number of atoms in
the universe, then α(n) < 5 (at least, that’s what people think). No computer
can ever be built that contains more components than the number of atoms
in the universe. An algorithm that runs in O(nα(n) ) is not polynomial, since
α(n) is an unbounded, growing function of n. For any practical situation this
algorithm runs in O(n5 ), but it is not polynomial. That’s where the ‘theoret-
ical’ in ‘theoretical computer science’ kicks in; it takes us where no man has
gone before.

When we really want to run that algorithm to check if ω(G) > 106 then
6
we run into trouble. I mean, a polynomial-time algorithm that runs in O(n10 )
is not very appetizing! So, “Big Deal!” you say, and you’re right, so far it is not
a big deal.

Here’s a definition.

Definition 4.1. A parameterized problem (P, k) is fixed-parameter tractable if

there exist
1. some function f : N → R, and
2. some constant c, and
3. some algorithm that solves (P, k) and that runs in O(f(k) · nc ) time.
 4 Fixed-parameter Algorithms 91

For example, when

f(k) = 2k and c = 2

a fixed-parameter algorithm to solve the parameterized problem (P, k) runs

in O(2k · n2 ).

Notice the difference with an algorithm that runs in O(k2 · nk ). In a fixed-

parameter algorithm the parameter k does not appear in the exponent of
n. An algorithm that runs in O(k2 · nk ) is not a fixed-parameter algorithm
because the k appears in the exponent of n.

Which of the two timebounds, O(k2 · nk ) or O(2k · n2 ) would you prefer?

For constant k, say k = 106 , both algorithms run in polynomial time. Also,
in both
√ cases, the algorithms are practical only for small values of k. When
k = n, the algorithms are not polynomial.

To compare them, let’s fix k = 10. The fixed parameter algorithm runs in
time O(210 · n2 ), so actually it is a quadratic algorithm. The other algorithm
runs in O(100 · n10 ). I think everybody agrees, the fixed-parameter algorithm
is better, unless we are fooled, for example by some crazy constant that is
hidden in the big O.

An example of a parameterized problem (P, k) is (ω(G), k) and it asks if

ω(G) > k. Or, another example of a parameterized problem is whether G
has a dominating set with at most k vertices, that is, (γ(G), k) is the problem
that asks if γ(G) 6 k. One more example is the problem (χ, k) which asks if
χ(G) 6 k.

Notice that you have to be a bit careful with the sign; sometimes you
want to maximize the cardinality of some subset and sometimes you want to
minimize it. The question whether there is a vertex coloring of G with at least
k colors makes little sense. If a graph has n vertices and if n > k then the
graph has a coloring with at least k colors.

In this chapter we look at some problems for which there are fixed-
parameter algorithms. Not every problem is like that. For example, it is
unlikely that the parameterized clique number problem (ω(G), k) is fixed-
parameter tractable. It can be shown that,

unless P = NP, there is no fixed-parameter algorithm that solves

the parameterized clique number problem (ω(G), k) on graphs.

Unless P=NP, any algorithm that solves (ω(G), k) will have a running time
where the k appears in some way in the exponent of n.
92 4 Fixed-parameter Algorithms

The book by Downey and Fellows examines which problems have a fixed-
parameter algorithm.1 The book contains a list of problems that are fixed-
parameter tractable, and a list of problems that are not fixed-parameter
tractable (unless P=NP).

It is not always easy to tell. As in the usual NP-completeness theory one can
show that some parameterized problem (P, k) is not fixed-parameter tractable
by reducing some other parameterized problem (Q, k0 ), which is hard for
fixed-parameter tractability, to (P, k).

Suppose it is known that some parameterized problem (Q, k0 ) is hard for

fixed-parameter tractability. To reduce (Q, k0 ) to (P, k) one reduces Q to P in
polynomial time. The extra condition that is put on this reduction is, that there
has to be some function that connects the parameters k and k0 , say k = g(k0 ).
Just as in ordinary NP-completeness proofs, we now need that (P, g(k0 )) has
a solution if and only if (Q, k0 ) has a solution.

Now suppose we have an algorithm that solves (P, k) in time O(f(k) · nc )

for some function f and some constant c. Then a parameterized reduction as
above solves (Q, k0 ) in time O(f(g(k0 )) · h(n)c ) where h is the polynomial that
reduces Q to P. Since we know that (Q, k0 ) has no fixed-parameter solution
unless P=NP, any fixed-parameter solution for (P, k) implies that P=NP.

Notice that the difference with the ordinary NP-completeness reductions

is the function that relates the parameters. People who are familiar with NP-
completeness proofs will have little trouble doing similar proofs for parame-
terized problems.

We show a very easy example of a parameterized reduction at the start of

the next section.

4.1 Vertex cover

Definition 4.2. Let G = (V, E) be a graph. A vertex cover for G is a set S of
vertices such that every edge in G has at least one endvertex in S.

The vertex cover problem asks for a vertex cover of smallest cardinality.

The vertex cover problem is of course equivalent to the independent set

problem.

Lemma 4.3. Let G = (V, E) be a graph. Then S is a vertex cover if and only if
V \ S is an independent set.
1
R. Downey and M. Fellows, Parameterized complexity, Springer-Verlag, 1999.
 4.1 Vertex cover 93

Proof. Let S be a vertex cover. By definition, there is no edge with both ends
in V \ S. So V \ S is an independent set.
Suppose I is an independent set. Then every edge must have at least one end
in V \ I. So V \ I is a vertex cover. t
u

We know that the independent set problem is NP-complete, so Lemma 4.3

shows that the vertex cover problem is NP-complete as well.

As we mentioned in the introduction, there is probably no fixed-parameter

algorithm that solves the parameterized clique number problem (ω(G), k).

There is a trivial parameterized reduction from (ω(G), k0 ) to (α(G), k).

The reduction takes the complement Ḡ of the graph G. The function that
relates k and k0 is just the identity function g(k) = k. Notice that G has a
clique with at least k vertices if and only if Ḡ has an independent set with k
vertices.

Since the reduction from G to Ḡ is polynomial, this shows that (α(G), k)

is not fixed-parameter tractable, unless something funny happens.

Let’s introduce some fancy notation for the vertex cover number of a
graph. Denote the smallest cardinality of a vertex cover in a graph G by ζ(G).
(That greek letter is “zeta,” or “z,” the last letter in the English alphabet. The
greek α is “a,” the first letter in the English alphabet.) Then, by Lemma 4.3,

ζ(G) = n − α(G), where n is the number of vertices in G.

(At least when n = 27 it looks OK in the English alphabet.)

Here comes the surprise!

Theorem 4.4. The vertex cover problem is fixed-parameter tractable. The prob-
lem (ζ(G), k) can be solved in O(2k · n2 ) time.

Proof. Let G = (V, E) be a graph. Let e = {x, y} be an edge in G. If Z is a

vertex cover then
x ∈ Z or y ∈ Z.
Consider the following algorithm.
Start with Z = ∅. Build a binary tree as follows. The root consists of the pair
(G, ∅). If G − Z has no edge, then we are done; Z is a vertex cover.
Now assume that {x, y} is an edge with both ends in G − Z. Then construct
two children in the tree. One for the pair

( G − (Z ∪ {x}), Z ∪ {x} )
94 4 Fixed-parameter Algorithms

and one for the pair

( G − (Z ∪ {y}), Z ∪ {y} ).
Thus in one case we put x in the vertex cover and in the other case we put y
in the vertex cover. In both cases we remove the vertex from the graph that
we put in the vertex cover.
Of course, we can keep growing the binary tree until each leaf is some pair
(G − Z∗ , Z∗ ), where Z∗ is a minimal vertex cover. When we scan all the leaves
and find the smallest cardinality of the sets Z∗ then we find ζ(G).
Here’s the trick. We want to find a vertex cover Z of size at most k. Thus, if
some node in the binary tree has a pair (G − Z0 , Z0 ) with

|Z0 | = k

then we don’t need to grow it any further. We let the algorithm check if G − Z0
is an independent set or not. If so, then Z0 is a vertex cover with k vertices.
Otherwise, we let the algorithm backtrack to its parent, and continue the
search from there.
In this way, the binary tree has a depth at most k, since each branching adds
one vertex to the set Z. Any full binary tree of depth k has 2k leaves and 2k −1
internal nodes.
In each vertex (G − Z∗ , Z∗ ) of the binary tree we have to check if there is an
edge {x, y} with both ends in G − Z∗ . This we can do in O(n2 ) time. The total
time complexity is therefore O(2k · n2 ). t
u

Remark 4.5. We have seen that (α(G), k) is not fixed-parameter tractable and
that (ζ(G), k0 ) is. Furthermore, we have that

ζ(G) = n − α(G)

for any graph G.

Obviously, we cannot have a parameterized reduction from (α(G), k) to
(ζ(G), k0 ). What is the problem?
The problem is that we need to relate the parameters in the two prob-
lems if we want to have a parameterized reduction. The parameters should
relate like k0 = n − k. There is no function doing that; there would be an “n”
somewhere in that function.

Remark 4.6. The technique used in Theorem 4.4 is called a bounded search-
tree technique. In this technique you build a search-tree of a size which is
some function of the parameter k. The bounded search-tree technique is one
of the most successful techniques that one can use to prove that some problem
is fixed-parameter tractable.
 4.2 A kernel for vertex cover 95

4.2 A kernel for vertex cover

In this section we describe another technique which is very often useful to ob-
tain fixed-parameter algorithms. We illustrate it for the vertex cover problem.

Lemma 4.7. Let G be a graph and assume that ζ(G) 6 k. Let Z be a vertex
cover for G with |Z| 6 k. If a vertex x in G has at least k + 1 neighbors then
x ∈ Z.

Proof. Suppose that x ∈

/ Z. Consider k + 1 edges

{x, y1 }, . . . , {x, yk+1 }, where {y1 , . . . , yk+1 } ⊆ N(x).

Every edge must have one endvertex in Z, thus if x ∈ / Z then yi ∈ Z, for all
i ∈ {1, . . . , k + 1}. But then |Z| > k, which is a contradiction. t
u

Lemma 4.8. Let G be a graph and assume that every vertex in G has degree at
most k. Furthermore, assume that G has no isolated vertices. Assume that G has
a vertex cover Z with at most k vertices. Let n and m be the number of vertices
and edges in G. Then

n 6 k + k2 and m 6 2k2 .

Proof. Let Z be a vertex cover in G. Since G has no isolated vertices, and since
Z is a vertex cover, every vertex of V \ Z has at least one neighbor in Z.
Since the degree of any vertex in G is at most k, any vertex in Z has at most
k neighbors in V \ Z. This proves that the number of vertices is at most

n 6 |Z| + |Z| · k = |Z|(k + 1) 6 k(k + 1)

Any edge has either two ends in Z or exactly one end in Z. Thus the number
of edges in G satisfies

m 6 |Z|2 + |Z| · k 6 2k2 .

This proves the lemma. t

u

Lemmas 4.7 and 4.8 can be used to reduce the graph G in polynomial time
to a graph H of which the number of vertices is bounded by some function
of the parameter k. Such a graph H is called a kernel for the parameterized
problem. Notice that the exponential part of the algorithm runs only on the
kernel. (You could say that the kernel is ‘the heart of the problem.’)
96 4 Fixed-parameter Algorithms

Theorem 4.9. Let G be a graph with n vertices. The parameterized vertex cover
problem (ζ(G), k) is fixed-parameter tractable. There exists an algorithm for
(ζ(G), k) which runs in O(n2 + 2k · k4 ) time.

Proof. Reduce G to a kernel H by removing vertices that have at least k + 1

neighbors in G. Let Z0 be this set of vertices. Then Z0 ⊆ Z for any vertex cover
Z with at most k vertices.
If some vertices of the remaining graph G − Z0 are isolated, then remove
them. The subgraph H induced by the remaining vertices of G − Z0 is the
kernel. The part of the algorithm described above is called the reduction to
the kernel, and it runs in O(n2 ) time.
By Lemma 4.8 we may assume that the number of vertices in H is at most
k2 + k. We need to find a vertex cover in H with at most k − |Z0 | vertices. By
Theorem 4.4 there is an O(2k · (k2 + k)2 ) = O(2k · k4 ) algorithm that solves
the parameterized vertex cover problem in the kernel.
In total, our algorithm runs in O(n2 + 2k · k4 ) time. To see that this is a fixed-
parameter algorithm, we need to show that it has the form f(k)nc for some
function f and constant c. This is obvious; namely, since k 6 n we have that

n2 + 2k · k4 6 (2k + 1) · n4 .

This proves the theorem. t

u

The following theorem is pretty useless in practice, but it may answer a

question that you had in mind.

Theorem 4.10. Any parameterized problem that is fixed-parameter tractable

can be reduced to a kernel in polynomial time.

Proof. Assume that there exists an algorithm that runs in O(f(k)·nc ) for some
function f and some constant c. We use this algorithm to reduce the input to
a kernel as follows.
1. When f(k) 6 n then the algorithm above runs in time proportional to
f(k)nc 6 nc+1 , i.e., it is polynomial. This part of the algorithm is the
reduction to the kernel.
2. Otherwise, when f(k) > n, then we have a kernel of size n < f(k).
This proves the theorem. t
u

An alternative algorithm for finding a kernel for the parameterized vertex

cover problem uses a maximum matching.

Definition 4.11. A matching in a graph G = (V, E) is a subset M of edges such

that no two edges in M have an endvertex in common.
 4.2 A kernel for vertex cover 97

The maximum matching problem asks for a matching in a graph G of max-

imal cardinality. The maximum matching problem can be solved in O(n5/2 )
on graphs G with n vertices.2

Lemma 4.12. Let G = (V, E) be a graph. Assume that ζ(G) 6 k. Then

ν(G) 6 k,

where ν(G) is the cardinality of a maximum matching in G.

Proof. If Z is a vertex cover in G then Z contains at least one endvertex of

each edge in a maximum matching M. t
u

Theorem 4.13. Let G = (V, E) be a graph and assume that G has no isolated
vertices. There exists a kernel with at most 3k vertices for (ζ(G), k).

Proof. The algorithm first computes a maximum matching M. By Lemma 4.12

if |M| > k then any vertex cover has at least k + 1 vertices.
Now assume that |M| 6 k. Let V(M) be the set of endvertices of edges in M
and let I = V \ V(M). Then I is an independent set.
Consider the graph B with vertex set V and edge set the edges with one
endpoint in V(M) and the other endpoint in I. Then B is bipartite. By the
König-Egerváry theorem (Page 39)

ζ(B) = ν(B).

Via the maximum matching algorithm we can find in polynomial time, a max-
imum matching M0 in B and a minimum vertex cover Z0 for B with endvertices
in V(M0 ).
We now consider two cases.
First assume that Z0 has at least one vertex in V(M). Define

U = V(M) ∩ Z0 .

Define I0 as the set of other endvertices of edges in M0 that have one endvertex
in U. Notice that, if U 6= ∅, then U separates I0 and V \ (U ∪ I0 ).
We claim that there exists a minimum vertex cover Z in G which contains U.
To see this, notice that each edge e ∈ M0 must have an endvertex in Z. If this
endvertex is in I0 , then we may replace it with the other endvertex of e in U.
Then the new set of vertices is also a vertex cover.
2
J. Edmonds, Paths, trees, and flowers, Canadian Journal of Mathematics 17 (1965),
pp. 449–467.
98 4 Fixed-parameter Algorithms

So, in this case we can reduce the graph by removing the vertices of
U ∪ I0
and put the vertices of U in Z. The algorithm proceeds to look for a vertex
cover with at most k − |U| vertices in
G − (U ∪ I0 ).

In the second case we assume that Z0 has no vertices in V(M). Then the
minimum vertex cover Z0 has only vertices in I. Since there are no isolated
vertices,
Z0 = I.
Then the number of vertices is at most
|I| + |V(M)| 6 k + 2k = 3k.
In other words, this set is a kernel. t
u

Remark 4.14. For every constant c there exists a polynomial-time algorithm

that reduces an instance (ζ(G), k) to a kernel with at most 2k − c log k ver-
tices.3 This seems pretty close to optimal. Unless P = NP there is no kernel-
ization which reduces the parameterized vertex cover problem to a kernel of
size (2 − )k which is a subgraph of the original graph.4

4.3 A better search-tree algorithm for vertex cover

In this section we develop a better search-tree algorithm for the parameter-
ized vertex cover problem. The improvement follows from two simple obser-
vations.

Lemma 4.15. Let G = (V, E) be a graph. Let x be a vertex of degree at most

one. Then there exists a minimum vertex cover Z for G with x ∈
/ Z.
Proof. Assume that x is isolated. Then x is not in any minimum vertex cover.
Assume that x has one neighbor, say y. If Z is a minimum vertex cover, and
x ∈ Z, then
Z0 = (Z \ {x}) ∪ {y}
is also a minimum vertex cover.
This proves the lemma. t
u
3
M. Lampis, A kernel of order 2k − c log k for vertex cover, Information Processing
Letters 111 (2011), pp. 1089–1091.
4
J. Chen, H. Fernau, I. Kanj and G. Xia, Parametric duality and kernelization: Lower-
bounds and upperbounds on kernel size, SIAM Journal on Computing 37 (2008),
pp. 1077–1106.
 4.3 A better search-tree algorithm for vertex cover 99

Lemma 4.16. Let G be a graph and let x be a vertex with two neighbors, say y
and z. Then there exists a minimum vertex cover Z for G such that either
(i) {y, z} ⊆ Z and x ∈
/ Z, or
(ii) x ∈ Z and
(N(y) ∪ N(z)) \ N(x) ⊆ Z.

Proof. Let G = (V, E) be a graph and let Z be a minimum vertex cover for G.
Let x be a vertex with two neighbors, say y and z.
Obviously, when y ∈ Z and z ∈ Z then x ∈
/ Z, since Z is minimum.
Assume that x ∈ Z, y ∈ Z and z ∈
/ Z. Then

Z0 = (Z \ {x}) ∪ {z}

is a minimum vertex cover and |Z0 | = |Z|.

Thus, we may assume that either both y and z are in Z or neither of them is
in Z. If neither y nor z is in Z, then x ∈ Z and

(N(y) ∪ N(z)) \ N(x) ⊆ Z.

This proves the lemma. t

u

Theorem 4.17. There exists an O(1.39k · n2 ) algorithm that solves the param-
eterized vertex cover problem (ζ(G), k).

Proof. The algorithm builds a search-tree as follows. The root of the search-
tree is the pair (G, ∅).
At each vertex (G0 , Z0 ) in the search tree the algorithm proceeds as follows.
(1) If there is an isolated vertex x in G0 , then remove it from G0 .
(2) If G0 has a vertex x with one neighbor y, then put y in Z0 and remove x
and y from the graph G0 .
(3) If G0 has a vertex x with at least three neighbors, then the algorithm
branches; the vertex in the search tree gets two children. In one child the
vertex x is put in Z0 and it is removed from G0 . In the other child, all the
neighbors of x are put in Z0 and all vertices of N[x] are removed from the
graph G0 .
(4) Assume that G0 has a vertex x with two neighbors, y and z. Assume that
neither y nor z has a neighbor which is not a neighbor of x. If y and z are
not adjacent then put x in Z0 and remove x, y and z from G0 . If y and z
are adjacent then put two of x, y and z in Z0 and remove x, y and z from
G0 .
Otherwise, when at least one of y and z has a neighbor which is not a
neighbor of x, the algorithm branches. In one child both y and z are put
100 4 Fixed-parameter Algorithms

in Z0 and the three vertices x, y and z are removed from G0 . In the other
child, x and all vertices of
(N(y) ∪ N(z)) \ N(x)
are put in Z0 . In this case all vertices of
N[y] ∪ N[z]
are removed from G0 . By the assumption that y or z has at least one
neighbor which is not a neighbor of x, this removes at least four vertices.
The correctness of this step follows from Lemma 4.16.

Notice that each vertex in the search tree is a leaf or it has two children.
Let T (n) be the amount of work done by the algorithm. In the first and second
case, the algorithm is greedy. We may assume that these cases do not occur.
In the third case, the branching gives the recurrence
T (n) 6 T (n − 1) + T (n − 4) + O(n2 ). (4.1)
In the fourth case, we have, apart from two greedy cases,
T (n) 6 T (n − 3) + T (n − 4) + O(n2 ). (4.2)
This is so because in one branch x, y and z are removed, and in the other
branch the vertices of N[y] ∪ N[z] are removed (and this set contains at least
four vertices).
The search tree has depth at most k. The solution for the Recurrence (4.1)
is T (k) = O(1.39k · n2 ). The solution for the Recurrence (4.2) is T (k) =
O(1.23k · n2 ). (In Exercise 4.5 we ask you to check that.) Clearly, the worst
case occurs when the algorithm branches all the time as in the third case. tu

Remark 4.18. The best algorithm that we know of for the parameterized ver-
tex cover problem (ζ(G), k) runs in O∗ (1.28k ) and polynomial space. Here we
suppressed the polynomial in n.5 This algorithm does a more extensive case
analysis. No better timebound is known, even when one allows exponential
space.6
Remark 4.19. The Exponential Time Hypothesis assumes that k-SAT cannot
be solved in O(2o(k)n ) time. It can be shown that, if (ζ(G), k) can be solved
in O(2o(k) · nO(1) ) time then the Exponential Time Hypothesis fails.7 8
5
J. Chen, I. Kanj, and G. Xia, Improved upperbounds for vertex cover, Theoretical
Computer Science 411 (2010), pp. 3736–3756.
6
L. Chandran, F. Grandoni, Refined memorisation for vertex cover, Information Pro-
cessing Letters 93 (2005), pp. 125–131.
7
R. Impagliazzo and R. Paturi, On the complexity of k-SAT, Journal of Computer and
System Sciences 62 (2001), pp. 367–375.
8
D. Marx, Fixed parameter algorithms. Open lectures for PhD students in computer
science, 2010, Warsaw, Poland.
 4.4 Minimum fill-in 101

4.4 Minimum fill-in

Recall Definition 3.50 on Page 64 of a chordal graph. A graph is chordal if it
has no induced cycle of length more than three.

Of course, when a graph G = (V, E) is not chordal then we can add some
edges to the graph such that it becomes chordal. For example, if we add all
edges between any pair of vertices x and y that are not adjacent in G, then
the new graph is a clique, and a clique is obviously chordal.

In the minimum fill-in problem one wants to add as few edges as possible
to make it chordal.

Definition 4.20. Let G = (V, E) be a graph. A graph H = (V, E0 ) is a chordal

embedding of G if
1. G and H have the same set of vertices, and
2. H is chordal, and
3. E ⊆ E0 .

A chordal embedding H of a graph G is minimal if the deletion of any edge

that is added to G, makes H non-chordal. If G is chordal then the only minimal
embedding of G is G itself.

Definition 4.21. The minimum fill-in problem asks for a chordal embedding H
of a graph G such that the number of edges in H is minimal.

The minimum fill-in problem is NP-complete.

Let f(G) be the minimal number of edges that needs to be added to G

in order to make G chordal. In this section we look at the parameterized
(f(G), k) problem.

For example, when G is a 4-cycle we need to add one edge. Furthermore,

there are two choices to add the edge. If G is a 5-cycle, we need two edges
and there are 5 choices to make G chordal by adding two edges.

If H is a t-cycle, then adding t − 3 edges from one vertex x to all vertices

of H − N[x] makes the graph chordal. In fact, any minimal chordal embedding
of H adds t − 3 edges to H, although not all chordal embeddings are like the
one above. (See Figure 4.1 on the following page.)
102 4 Fixed-parameter Algorithms

Fig. 4.1. The figure shows the different triangulations of C6 . The number of these is
the Catalan number C4 = 14.

Lemma 4.22. Let C be a cycle of length t. Then f(C) = t − 3. There are

 
1 2(t − 2)
6 4t−3
t−1 t−2

different embeddings of C into a chordal graph with t + f(C) edges.

Proof. We leave most of the proof as an exercise.

The Catalan number Cn is defined as
 
1 2n
Cn = .
n+1 n

(They are named after the Belgian mathematician E. Catalan (1814-1894).

The Chinese mathematician Antu Ming discovered these numbers around
1730, but he didn’t call them ‘Catalan numbers!’)
The Catalan numbers satisfy the recurrence relation

2(2n + 1)
Cn+1 = Cn and C1 = 1.
n+2

Notice that the bound Cn 6 4n−1 follows easily by induction from the recur-
rence relation:
2(2n + 1) 4n + 2 n−1 4(n + 2) n−1
Cn+1 = Cn 6 4 6 4 = 4n .
n+2 n+2 n+2
t
u
 4.4 Minimum fill-in 103

Theorem 4.23. The minimum fill-in problem is fixed-parameter tractable. The

parameterized problem (f(G), k) can be solved in O(4k · n5 ) time.
Proof. We use the bounded search-tree technique.
First we show how to find an induced cycle of length at least four. Let x be
a vertex and let y and z be two neighbors of x such that y and z are not
adjacent. Then {x, y, z} is contained in a chordless cycle of length at least four
if and only if y and z are contained in a component of the graph induced by

(V \ N[x]) ∪ {y, z}.

So, to look for an induced cycle of length at least four, we try all possible
triples {x, y, z} as above. For each such triple the algorithm finds a path from
y to z in
G − (N[x] \ {y, z}).
Adding the vertex x yields the chordless cycle of length at least four.
Notice that the number of these triples is bounded by
X X
d(x)(d(x) − 1) 6 n d(x) = 2nm.
x∈V x∈V

At each vertex in the search-tree T , the algorithm finds an induced cycle C

of length at least four. Let t be the length of C. By Lemma 4.22 C cannot be
made chordal when t − 3 > k. Otherwise, the search-tree branches over the
1 2t−4


t−1 t−2 possible minimal embeddings of C. The parameter k reduces by
t − 3 since there are t − 3 edges added to C.
Let Lk be the number of leaves in the search-tree T . We claim that

Lk 6 4k .

We prove this by induction. If k = 0 then the algorithm does not branch. It

checks in linear-time if the graph is chordal. If so, we are done, and if not,
there is no suitable chordal embedding of G since we cannot add any edge.
By Lemma 4.22 and by induction, for k > 1,

Lk 6 4t−3 · Lk−(t−3) 6 4k .

We claim that the search-tree T has at most 2 · 4k + 1 vertices. To see this, first
note that it has at most one vertex of degree two, namely the root. All other
internal vertices have one parent and at least two children, since every cycle
of length at least four has at least two possible minimal embeddings.
Let n0 and m0 be the number of vertices and edges in T . We write L instead
of Lk for the number of leaves in T . For a vertex x in T we write dT (x) for the
number of neighbors of x in T . Then we have
104 4 Fixed-parameter Algorithms
X X X
2m0 = 2(n0 − 1) = 1+ 2+ 3 + ...
x,dT (x)=1 x,dT (x)=2 x,dT (x)=3
X X
> L+ 3+ 4 + ...
x,dT (x)=3 x,dT (x)=4
X X
> L + 3( 1+ 1 + . . .)
x,dT (x)=3 x,dT (x)=4

> L + 3(n0 − L − 1) = 3n0 − 2L − 3.

This implies that

n0 6 2L + 1 6 2 · 4k + 1.

At each vertex in the search-tree we spend at most O(n3 ) time to look for a
suitable triple {x, y, z} and O(n2 ) time to find the y, z-path. Thus, a cycle of
length at least four is found in O(n5 ) time, if it exists. The total time that is
used by this algorithm is therefore bounded by O((2·4k +1)·n5 ) = O(4k ·n5 ).
This proves the theorem. t
u

4.5 Odd cycle transversals

Many problems can be solved in polynomial time for bipartite graphs. There-
fore, it is of some interest to remove a small number of vertices from the graph
such that the remaining graph is bipartite.

The problem to remove a minimal number of vertices from the graph such
that the remaining graph is bipartite is of course equivalent to finding a set
of vertices of minimal cardinality such that each odd cycle in the graph has
at least one vertex in this set. The problem is called the odd cycle transversal
problem.

Definition 4.24. Let G = (V, E) be a graph. An odd cycle transversal in G is a

set F ⊆ V of vertices such that G − F is bipartite.

The odd cycle transversal problem asks to find a minimum set F of vertices
in G such that G − F is bipartite. The odd cycle transversal problem is NP-
complete. In this section we show that it is fixed-parameter tractable.9

Given an integer k we show that there exists an O(4k k · mn) algorithm

that either finds an odd cycle transversal with at most k vertices or decides
that there is no odd cycle transversal with at most k vertices. Here, m and n
are the numbers of edges and vertices in the graph.

9
B. Reed, K. Smith and A. Vetta, Finding odd cycle transversals, Operations Research
Letters 32 (2004), pp. 299–301.
 4.5 Odd cycle transversals 105

The algorithm is recursive. It chooses an arbitrary vertex x and recursively

calls the procedure to find an odd cycle transversal with at most k vertices
in G − x. If G − x has no odd cycle transversal with at most k vertices, then
neither has G. In that case we are done.

Assume that G − x has an odd cycle transversal F with at most k vertices.

Let
F0 = F ∪ {x}.

Lemma 4.25. Let G be a graph and let x be a vertex of G. Let F be an odd cycle
transversal in G − x. Then F0 = F ∪ {x} is an odd cycle transversal in G.

Proof. This is obvious. Any odd cycle in G which contains no vertex of F must
contain x, otherwise it would be an odd cycle in (G − {x}) − F. Thus F0 is an
odd cycle transversal in G. t
u

If |F0 | 6 k, and we are done. Otherwise

|F0 | = k + 1.

In that case, the algorithm proceeds as follows. It calls a subroutine which,

given a graph G and an odd cycle transversal T with k+1 vertices, determines
if G has an odd cycle transversal with at most k vertices.

Assume that T is an odd cycle transversal of G = (V, E). Let

{B1 , B2 }

be a partition of the vertices of G − T into two color classes. Thus B1 and B2

induce two independent sets in G − T .

Define a bipartite graph

B = (V 0 , E0 ) (4.3)
as follows. For each vertex t ∈ T introduce two vertices t1 and t2 and put one
in each color class of B. That is,

V 0 = (V \ T ) ∪ { t1 , t2 | t ∈ T }. (4.4)

For each edge e in G we put the following edges in B, depending on the

endpoints of e.
(i) If e is an edge in G − T then e is an edge of B with the same endpoints
as e.
(ii) If e connects a vertex s ∈ Bi with a vertex f ∈ T then the edge e connects
s in Bi with the vertex f3−i in B3−i in B. Here i ∈ {1, 2}.
106 4 Fixed-parameter Algorithms

(iii) Assume e = {f, g} with {f, g} ⊆ T . The graph B contains one edge for
each edge e of this type. Either, arbitrarily, {f1 , g2 } or {f2 , g1 }.
This completes the description of the bipartite graph B. Thus the vertices of B
are partitioned into two color classes

B01 = B1 ∪ { t1 | t ∈ T } and B02 = B2 ∪ { t2 | t ∈ T }. (4.5)

Definition 4.26. Let Y ⊆ T . Write

Y 0 = { y1 , y2 | y ∈ Y }, (4.6)

where yi is the copy of y in B0i , i ∈ {1, 2}. A partition {Yα , Yβ } of Y 0 is valid if

|{ y1 , y2 } ∩ Yα | = 1 for all y ∈ Y. (4.7)

The following theorem shows when there is an odd cycle transversal in G

of cardinality smaller than |T |.

Theorem 4.27. An odd cycle transversal T is minimum if and only if for every
valid partition
{ Yα , Yβ } (4.8)
of Y 0 , for any Y ⊆ T , there are |Y| vertex-disjoint paths from Yα to Yβ in

B0 = B[Yα ∪ Yβ ∪ B1 ∪ B2 ] (4.9)

Proof. First assume that T is minimum. Let Y ⊆ T , let Y 0 be defined as in (4.6),

let {Yα , Yβ } be a valid partition of Y 0 , and assume that there are less than |Y|
vertex-disjoint paths from Yα to Yβ in

B0 = B[Yα ∪ Yβ ∪ B1 ∪ B2 ].

By the max flow – min cut theorem, or Menger’s theorem, there exists a
separator S in B0 which separates Yα from Yβ with cardinality less than |Y|.
Define a set of vertices Ω ⊆ V as follows. A vertex w ∈ V is in Ω if, either

w ∈ S or w ∈ T and at least one of w1 and w2 is in S. (4.10)

Then we obtain a smaller odd cycle transversal in G, namely,

T ∗ = Ω ∪ (T \ Y). (4.11)

To see that this is an odd cycle transversal, assume that G − T ∗ has an odd
cycle O. Then O must have a vertex y ∈ Y. The set Ω separates Yα from Yβ in
B0 , where B0 is defined as in (4.9). One copy of y is in Yα and the other copy
 4.5 Odd cycle transversals 107

is in Yβ . Since S separates y1 and y2 in B0 , the y1 , y2 -path in B0 induced by O

must have a vertex in S. Then either this is a vertex of G or it is a vertex in T .
In both cases O contains a vertex in Ω, by definition.

For the converse, assume that G has an odd cycle transversal U which is
smaller than T . That is, we assume |U| < |T |. Let { B̃1 , B̃2 } be a partition of
the vertices of G − U into two color classes.
Let Y = T \ U, and let the valid partition {Yα , Yβ } be defined by

Yα = { y1 | y ∈ Y ∩ B̃1 } ∪ {y2 | y2 ∈ Y ∩ B̃2 } (4.12)

Yβ = { y2 | y ∈ Y ∩ B̃1 } ∪ { y1 | y ∈ Y ∩ B̃2 }. (4.13)

We claim that there are not |Y| vertex-disjoint paths from Yα to Yβ in

B[Yα ∪ Yβ ∪ B1 ∪ B2 ]. (4.14)

To show that, we prove that U \ T is a separator separating Yα from Yβ . Since

|U \ T | < |T \ U| = |Y| (4.15)

this will prove the claim.

Consider a chordless path P from Yα to Yβ with vertices in

Y 0 = (Yα ∪ Yβ ∪ B1 ∪ B2 ) \ (U \ T ). (4.16)

Let u and v be the endpoints of P. We may assume that P contains no other

vertices of Y 0 .
By symmetry, we may assume that either both u and v are in Y ∩ B̃1 or that
u ∈ Y ∩ B̃1 and v ∈ Y ∩ B̃2 .
Consider the first case. Then, by (4.12) and (4.13) u = y1 for some y ∈ Y and
v = y02 for some y0 ∈ Y. Thus P has odd length. By definition, P contains no
copy of a vertex in U. Thus P must have even length, since y1 and y02 are both
in B̃1 .
Consider the second case, that is u ∈ Y ∩ B̃1 and v ∈ Y ∩ B̃2 . Then u = y1 for
some y ∈ Y and v = y01 for some y0 ∈ Y. Then P has even length. However,
u ∈ B̃1 and v ∈ B̃2 , and so P must have odd length.
This proves the theorem. t
u

We now easily obtain the result.

Theorem 4.28. For any k ∈ N there exists an O(4k k · mn) algorithm which
determines if a graph G can be made bipartite by deleting at most k vertices.
108 4 Fixed-parameter Algorithms

Proof. For a given odd cycle transversal F0 of size k + 1 we can check if there
exists a smaller one by solving a maximum flow problem for each choice of
Y ⊆ F0 and each valid partition {Yα , Yβ } of Y 0 as defined in (4.6). There are
O(2k ) choices for Y and O(2k ) choices for a valid partition.
The Ford-Fulkerson algorithm takes time O(km), since the value of the flow
is O(k). Since the subroutine is called for at most n vertices x, this proves the
theorem. t
u

Remark 4.29. The method illustrated in this section is sometimes called itera-
tive compression. Given a solution of small size k + 1 one tries to ‘compress’
it to a solution of size k or decide that no such solution exists.

4.5.1 Feedback vertex set

As another example of the iterative compression technique we discuss the

feedback vertex set problem.

Recall Definition 3.46 on page 62. Let G be a graph. A set F ⊆ V is a

feedback vertex set if G − F is a forest. Denote by f(G) the cardinality of a
minimum feedback vertex set in G.

The minimum feedback vertex set problem is NP-complete. In this section

we design a fixed-parameter algorithm for feedback vertex set.10

Let V = {v1 , . . . , vn } and define, for i ∈ {1, . . . , n},

Vi = { v1 , . . . , vi } and Gi = G[Vi ]. (4.17)

The iterative compression technique tries to find a feedback vertex set Fi

in Gi of cardinality at most k as follows.
1. Let i = 1 and F1 = ∅.
2. For i ∈ {2, . . . , n}, let
Fi = Fi−1 ∪ {vi }. (4.18)
3. Assume that |Fi−1 | 6 k. If |Fi | 6 k, then Gi has a feedback vertex set with
at most k vertices. Otherwise, |Fi | = k + 1. In that case, using iterative
compression, find a new minimum feedback vertex set Fi in Gi and check
if |Fi | 6 k.

10
J. Guo, J. Gramm, F. Hüffner, R. Niedermeier and S. Wernicke, Compression-based
fixed-parameter algorithms for feedback vertex set and edge bipartization, Journal
of Computer and System Sciences 72 (2006), pp. 1386–1396.
 4.5 Odd cycle transversals 109

We concentrate on the third item. For convenience we write G instead of

Gi , V instead of Vi and F instead of Fi . We assume that |F| 6 k + 1.

First observe that it is sufficient to describe an algorithm which solves the

following problem.

Problem 1.
Instance: A graph G = (V, E) and a feedback vertex set F in G of cardi-
nality at most k + 1.
Task: Find a minimum feedback vertex set F0 such that F0 ∩ F = ∅.

To see that we may restrict ourselves to solving this problem, let F̂ be a

minimum feedback vertex set in G. Let

X = F̂ ∩ F and Y = F̂ \ F. (4.19)

Then F \ X is a feedback vertex set in G − X and Y is a minimum feedback

vertex set in G − X which is disjoint from F \ X.

In our algorithm we try all subsets X ⊆ F. Remove X from the graph and
find a minimum feedback vertex set Y in G − X disjoint from F \ X. Since there
are 2k+1 possible subsets X ⊆ F, it is sufficient to solve the problem stated
above.

We first get rid of some vertices that are unimportant for solving the prob-
lem. We present four reduction rules which reduce an instance of Problem 1
to an instance in which all vertices of V \ F have degree at least three in G.

Lemma 4.30. Let {G, F} be a problem instance of Problem 1 and let x be a vertex
in G of degree at most one. Then a solution to Problem 1 with instance

{G − x, F \ {x}}

is a solution to Problem 1 with instance {G, F}.

Proof. If x has degree at most one then it is not contained in any cycle. Thus
any feedback vertex set in G − x is a feedback vertex set in G. Thus, a solution
of Problem 1 for instance {G−x, F\{x}} is a solution to Problem 1 with instance
{G, F}. t
u

Lemma 4.31. Let {G, F} be a problem instance of Problem 1. Let x be a vertex

of degree two and let y and z be the two neighbors of x. Assume that x ∈ F and
assume that y and z are not adjacent. Let G0 be the graph obtained from G by
removing x and adding the edge {y, z}. Let Y be a solution to Problem 1 with
instance
110 4 Fixed-parameter Algorithms

{G0 , F \ {x}}.
Then Y is a solution to Problem 1 with instance {G, F}. If there is no solution to
Problem 1 with instance {G0 , F \ {x}} then there is no solution to Problem 1 with
instance {G, F}.

Proof. Let Y be a solution to Problem 1 with instance {G0 , F\{x}}. We first show
that Y is a feedback vertex set in G, disjoint from F. Assume that G − Y has a
cycle C. If x ∈ C then C contains y and z since x has two neighbors in C and
y and z are the only two neighbors of x in G. Since y and z are not adjacent
C must contain at least one vertex of V \ {x, y, z}. But then C \ {x} is a cycle in
G0 − Y, which is a contradiction.
Let Ỹ be a feedback vertex set which solves Problem 1 with instance {G, F}.
Since x ∈ F and Ỹ ∩ F = ∅, x ∈ / Ỹ. We show that Ỹ is a feedback vertex set
in G0 . If C were a cycle in G0 − Ỹ then C had to contain the edge {y, z}, since
otherwise C would also be a cycle in G − Ỹ. But then C ∪ {x} would be a cycle
in G − Ỹ, which is a contradiction.
Now assume that there is no solution to problem 1 with instance {G0 , F \ {x}}.
Then there is a cycle C in G0 with all vertices in F. If C is a cycle in G then
{G, F} has no solution. Otherwise, C contains the edge {y, z} in G0 . Then C∪{x}
is a cycle in G with all vertices in F. Thus {G, F} has no solution. t
u

Lemma 4.32. Let {G, F} be a problem instance of Problem 1. Assume x has a

vertex of degree two and let y and z be its two neighbors. Assume that y and z
are adjacent. Assume that y ∈/ F. Let Ỹ be a solution to Problem 1 with instance

{G − x, F \ {x}}. (4.20)

(1) If y ∈ Ỹ or z ∈ Ỹ then Ỹ is a solution to Problem 1 with instance {G, F}.

(2) If y ∈
/ Ỹ and z ∈/ Ỹ then Y = Ỹ ∪ {y} is a solution to Problem 1 with instance
{G, F}.
(3) If there is no solution to Problem 1 with instance 4.20 then there is no solu-
tion to problem 1 with instance {G, F}.

Proof. Let Ỹ be a minimum feedback vertex set in G − x disjoint from F \ {x}.

Assume that y ∈ Ỹ or z ∈ Ỹ. We first show that Ỹ is a feedback vertex set in G,
disjoint from F. Any cycle C in G which is not a cycle in G − x must contain x.
Furthermore, C contains y and z since these are the only two neighbors of x
in G. Since {y, z} ∩ Ỹ 6= ∅, Ỹ has a vertex of C.
Let Y be a solution to Problem 1 with instance {G, F}. If x ∈ Y then we may
replace x in Y with y and obtain an alternative solution which does contain y
and which does not contain x. Then Y is a solution to Problem 1 with instance
{G − x, F} with y ∈ Y and then we are done. Thus we may assume that x ∈ / Y.
 4.5 Odd cycle transversals 111

Since {x, y, z} is a cycle in G, Y must contain at least one of y and z. Then Y is

a solution to Problem 1 with instance {G − x, F \ {x}}, with {y, z} ∩ Y 6= ∅. So
we are done.
Now assume that y ∈ / Ỹ and z ∈ / Ỹ. We claim that Ỹ ∪ {y} is a solution to
Problem 1 with instance {G, F}. First observe that Ỹ ∪ {y} is a feedback vertex
set in G disjoint from F. This is so because any cycle in G which is not a cycle
in G − x contains x, y and z and so it has a vertex in Ỹ ∪ {y}.
Let Y ∗ be a minimum feedback vertex set in G, disjoint from F. If y ∈ Y ∗ or
z ∈ Y ∗ then we are done, since Y ∗ \ {x} is a minimum feedback vertex set in
G − x and it contains y or z. Assume that y ∈ / Y ∗ and that z ∈ / Y ∗ . Then x ∈ Y ∗
since {x, y, z} is a cycle in G. Now (Y \ {x}) ∪ {y} is also a solution to Problem 1
∗

with instance {G, F} and it contains y.

It is easy to check that, when (4.20) has no solution then {G, F} has no solution
either. This proves the lemma. t
u

Lemma 4.33. Let {G, F} be a problem instance of Problem 1. Assume G has a

vertex x with degree two and let y and z be its neighbors. Assume that y and z
are adjacent. Assume that y ∈ F and z ∈ F. If x ∈ F then there is no solution to
Problem 1. If x ∈/ F then let Ỹ be a solution to Problem 1 with instance {G − x, F}.
Then Ỹ ∪ {x} is a solution to Problem 1 with instance {G, F}.

Proof. Since y ∈ F and z ∈ F there can only be a solution when x ∈

/ F.
Furthermore, any solution contains x. t
u

Henceforth, we may assume that all vertices in V \ F have degree at least

three in G = (V, E). We now show that if G has a feedback vertex set F0
disjoint from F and strictly smaller than F then the number of vertices in G is
at most 15|F|.

Lemma 4.34. Let G = (V, E) be a graph and let F be a feedback vertex set in
G. Assume that every vertex of V \ F has degree at least three in G. If G has a
feedback vertex set F0 with

F0 ∩ F = ∅ and |F0 | < |F| (4.21)

then G has at most 15 · |F| vertices.

Proof. Let V 0 = V \ F. Partition V 0 into the following three sets.

A = {x | x ∈ V 0 and |N(x) ∩ F| > 2 }

B = {x | x ∈ V \ A
0
and |N(x) ∩ V 0 | > 3 }
C = V \ (A ∪ B).
0
112 4 Fixed-parameter Algorithms

We bound the cardinality of A, B and C as follows.

We first show that |A| < 2|F|. Consider the bipartite graph GA with color
classes F and A and edges

EA = { {a, f} | a ∈ A and f ∈ F and {a, s} ∈ E }. (4.22)

If GA is a forest then |EA | 6 |F| + |A| − 1. Each vertex of A has at least two
neighbors in F, thus
|EA | > 2|A|. (4.23)
Hence, if GA is a forest then |A| 6 |F| − 1. If |A| > 2|F| then it is clearly
impossible to remove at most |F| vertices from A such that GA is acyclic.
We now show that |B| 6 2|F|. Since F is a feedback vertex set in G, G[V 0 ] is a
forest. All vertices of V 0 that have degree at most one in G[V 0 ] are in A, since
every vertex of G − F has degree at least three. All vertices of B have degree
at least three in G[V 0 ] and since G[V 0 ] is a forest

|B| 6 |A| < 2|F|. (4.24)

Each vertex of C has degree two in G[V 0 ]. Thus it has at least one neighbor
in F. Since C ∩ A = ∅, every vertex of C has exactly one neighbor in F. The
graph G[C] consists of paths and isolated vertices. We first bound the number
of isolated vertices. Each isolated vertex connects two components of G[A∪B]
and no two components of G[A ∪ B] are connected by more than one isolated
vertex in C. Since G[V 0 ] is a forest, the number of isolated vertices in C is at
most
|A ∪ B| − 1 < 4|S|. (4.25)

We now bound the number of vertices in C that are in paths of length at least
one. We claim that the number of vertices in G[C] that are not isolated is
at most 6|F|. First notice that each edge of G[C] creates a path between two
vertices in F. Thus, if there are at least |F| edges in G[C] then there is a cycle
in G[C ∪ S].
Assume that G[C] has more than 3|F| edges. Each vertex of G[C] is incident
with at most two edges of G[C]. If we remove at most |F| vertices from G[C]
then at least |F| edges remain and then there is a cycle by the argument above.
Thus G[C] can have at most 3|F| edges, that is, at most 6|F| vertices of G[C] are
in paths of length at least one.
In total, we find that |V 0 | is bounded by

|V 0 | = |A| + |B| + |C| < 2|F| + 2|F| + (4 + 6)|F| = 14|F|. (4.26)

This proves the lemma. t

u
 4.5 Odd cycle transversals 113

Lemma 4.35. Let G be a graph and let F be a feedback vertex set in G of cardi-
nality k + 1. There exists an O(ck · n2 ) algorithm, for some constant c, which
decides if there exists a feedback vertex set F0 with cardinality at most k.

Proof. The algorithm tries all 2k+1 subsets X ⊆ F and tries to find a feedback
vertex set in G − X which is disjoint from F \ X.
Let G∗ = G − X and reduce G∗ by repeated application of the reduction rules
implied by Lemmas 4.30, 4.31, 4.32 and 4.33. This takes at most linear time.
By Lemma 4.34, the remaining graph G0 has at most 14 · |F| vertices in V 0 \ F,
where V 0 is the set of vertices of G0 and where F is a feedback vertex set in G0
with at most k + 1 vertices. Hence, the algorithm needs to try at most

X
k 
14(k + 1)

< 214·(k+1) (4.27)
i
i=0

subsets of V 0 \ F.
In total, the time complexity is bounded by O(215·k · n2 ). t
u

By lemma 4.35 the compression step takes O(ck · n2 ) time. Since this step
is performed at most n times, namely for the graphs Gi where i = 1, . . . , n,
the overall time complexity is bounded by O(ck · n3 ) time. This proves the
following theorem.

Theorem 4.36. The parameterized feedback vertex set problem (f(G), k) can be
solved in O(ck · n3 ) time, for some constant c.

Remark 4.37. By a careful analysis the constant c can be lowered to c ≈

10.6.11

Remark 4.38. Thomassé proved that (f(G), k) can be reduced to a kernel with
at most 4k2 vertices.12

11
F. Dehne, M. Fellows, M. Langston, F. Rosamund and K. Stevens, An O(2O(k) · n3 )
FPT algorithm for the undirected feedback vertex set problem, Theory of Computing
Systems 41 (2007), pp. 479–492.
12
S. Thomassé, A 4k2 kernel for feedback vertex set, ACM Transactions on Algorithms
6 (2010), pp. 32:1–8.
114 4 Fixed-parameter Algorithms

4.6 Homogeneous coloring of perfect graphs

Definition 4.39. A homogeneous coloring of a graph is a partition of its vertices
into cliques and independent sets.

We refer to the sets of the partition as color classes. Each color class is
either a clique or an independent set. The homogeneous coloring problem
asks for a homogeneous coloring with a minimum number of color classes.

A (k, `)-coloring of a graph is a partition of its vertices into k cliques and l

independent sets, of which some may be empty.

Recall that a graph G is perfect when ω(H) = χ(H) for every induced
subgraph H of G. By the perfect graph theorem, when G is perfect, also Ḡ is
perfect. Recall also Theorem 3.7 on Page 40 which shows that the coloring
and clique number problem can be solved in polynomial time for G, and also
for Ḡ, when G is a perfect graph.

This is no longer true for homogeneous colorings. Wagner showed that

finding a homogeneous coloring of a permutation graph with a minimal num-
ber of colors is NP-complete. However, for every k and ` there exists a finite
set of graphs F(k, `) such that a perfect graph has a (k, `)-coloring if and only
if it has no element of F(k, `) as an induced subgraph.13 It follows that there
is a polynomial-time algorithm to check if a perfect graph can be colored with
k cliques and ` independent sets. Although this algorithm is polynomial for
each fixed k and `, this is not a fixed-parameter algorithm. Furthermore, the
theorem only shows the existence of a polynomial-time algorithm.

Recently, it was shown that classes of graphs that do not contain all com-
plete multipartite graphs nor their complements have a (k, k)-coloring for
some fixed k. Chudnovsky and Seymour proved the following result.14

Theorem 4.40. Let H be a complete multipartite graph and let J be a disjoint

union of cliques. There exists a k such that any graph G that does not contain H
nor J as an induced subgraph has a (k, k)-coloring.

13
K. Wagner, Monotonic coverings of finite sets, Elektronische Informationsverar-
beitung und Kybernetik 20 (1984), pp. 633–639.
T. Feder and P. Hell, Matrix partitions of perfect graphs, Discrete Mathematics 306
(2006), pp. 2450–2460.
A. Kézdy, H. Snevily and C. Wang, Partitioning permutations into increasing and
decreasing subsequences, Journal of Combinatorial Theory, Series A 73 (1996),
pp. 353–359.
14
M. Chudnovsky and P. Seymour, Extending the Gyárfás-Sumner conjecture.
Manuscript 2012.
 4.6 Homogeneous coloring of perfect graphs 115

In the following theorem we show that the (k, `)-coloring problem is fixed-
parameter tractable on perfect graphs.15

Theorem 4.41. There exists an O(f(k, `)nc ) algorithm which solves the (k, `)-
coloring problem on perfect graphs. Here

f(k, `) = (k + `)2k·`+` and c is a constant.

Proof. Let x1 , . . . , xn be an arbitrary ordering of the vertices and define

Gi = G[Vi ] where Vi = {x1 , . . . , xi }.

If there exists a partition of G into k cliques and l independent sets, of which

some may be empty, then such a partition exists for each of the graphs Gi .
In each step the algorithm checks if the current graph Gi has a partition of the
vertices with k cliques and ` independent sets. Let P be a partition of Gi−1
into k cliques and ` independent sets. Then, obviously, the graph Gi has a
partition into k + 1 cliques and ` independent sets. To see this, just consider
the partition
P ∪ {{xi }}.

Consider a partition P of Vi into cliques and independent sets. Consider a

bitvector L of length i where the tth bit is one if the vertex xt is in a clique of
P and zero if the vertex xt is in an independent set of P.
Assume that we are given the bitvector L. Then we can check in polynomial
time if it is valid by checking if
(a) the vertices with a 1 in L induce a perfect graph that has a clique cover
with at most k cliques, and
(b) the vertices with a 0 in L induce a perfect graph that has chromatic num-
ber at most `.

Let P and Q be two homogeneous colorings of Gi . Assume that P has k cliques

and ` independent sets and that Q has k0 cliques and `0 independent sets. Let
LP and LQ be the bitvectors of P and Q. The Hamming distance H(LP , LQ )
of LP and LQ is the number of bits that are different in LP and LQ . We claim
that
H(LP , LQ ) 6 k · `0 + k0 · `. (4.28)
To see this, simply observe that any independent set in P and clique in Q can
intersect in at most one vertex.

15
P. Heggernes, D. Kratsch, D. Lokshtanov, V. Raman and S. Saurabh, Fixed-parameter
algorithms for cochromatic number and disjoint rectangle stabbing, Proceedings
SWAT’10 , Springer, LNCS 6139 (2010), pp. 334–345.
116 4 Fixed-parameter Algorithms

The algorithm runs as follows. Let P be a partition of Gi−1 . Add the vertex
xi as a separate clique to P and let L be the bitvector of this partition. Let
X be the set of vertices that have a 1 in L and let Y be the set of vertices
with a 0 in L. First check if Gi [X] has a clique cover with at most k cliques
and if Gi [Y] has chromatic number at most `. If that is the case, then we are
done; the algorithm outputs a clique cover and a coloring with k cliques and
` independent sets for the graph Gi .
Now assume that the clique cover number of Gi [X] is at least k + 1. Since
Gi is perfect, it has an independent set S with k + 1 vertices. Since Gi is a
perfect graph it can find a maximum independent set S in polynomial time
and reduce this, if necessary, to k + 1 vertices.
If there is a partition Q of Gi with k cliques and ` independent sets, then, by
Equation (4.28),
H(L, LQ ) 6 µ = 2k · ` + `. (4.29)
If there exists an independent set S with k + 1 vertices in Gi [X], then there is a
vertex z ∈ S of which the corresponding bit is 0 in LQ . For each bit in L which
corresponds with a vertex z of S the algorithm does the following. It switches
the bit of z in L to 0. Let L0 be this bitvector. Then the algorithm recurses and
tries to find a partition Q which is at distance at most µ − 1 from L0 .
The case where Gi [Y] has chromatic number at least ` + 1 is similar.
To analyze the time complexity, observe that we may assume that k > 1 and
` > 1, since otherwise we can just check if there is a clique cover or coloring
with k or ` sets. The recursion tree has depth at most 2k · ` + ` since each time
the algorithm is recursively called the Hamming distance µ is decreased by
one. Each node in the recursion tree corresponds with a clique of cardinality
` + 1 or an independent set of cardinality k + 1. Since

k + 1 6 k + ` and ` + 1 6 k + ` (4.30)

every node in the recursion tree has at most k + ` children. Thus the total
number of recursions is bounded by

(k + `)2k·`+` . (4.31)

The graph searches for a partition P in Gi for i = 1, . . . , n. In each transition

from i to i + 1 the graph spends O(nc ) time in each node of a recursion tree
with (k + `)2k`+` nodes. This proves the time bound and the theorem. t
u

4.7 Color coding

Alon, et al., introduced the color coding technique to obtain good randomized
algorithms for various problems where one searches for specific subgraphs of
 4.7 Color coding 117

small cardinality.16 The technique has become a popular technique for obtain-
ing good fixed-parameter algorithms.

We illustrate the color-coding technique for finding a path of length k in

a graph G between two given vertices s and t. Obviously, this problem is NP-
complete, since by putting k = n − 1 the problem reduces to the Hamiltonian
Path problem (by trying all pairs s and t).

The trick to solve this problem is very easy. First, randomly color the ver-
tices of the graph with k colors. Say the colors are {1, . . . , k}. Now search for
a path
P = [p1 , . . . , pk ] (4.32)
where s = p1 and t = pk such that pi has color i for all i ∈ {1, . . . , k}.

When G is equipped with a k-coloring then the search algorithm for a path
P as above is very easy. Namely, first delete all edges from G except those that
connect vertices with colors i and i + 1, for i ∈ {1, . . . , k − 1}. Direct the
remaining edges, {x, y} as − → if x has color i and y has color i + 1. This makes
xy
G into a directed, acyclic graph. Next, the algorithm searches for an s, t-path
in this acyclic digraph. This takes linear time using dynamic programming;
simply maintain the subset of vertices with color i that can be reached from s
by a colored path [p1 , . . . , pi ].

Lemma 4.42. Assume that G has an s, t-path. Consider the algorithm which
tries all permutations of the k colors and which runs the algorithm described
above for each permutation. This algorithm finds an s, t-path with probability
at least e−k .

Proof. Consider an s, t-path P0 in G. The probability that all vertices of P0 are

colored with different colors is
k! (k/e)k
k
> = e−k . (4.33)
k kk
This proves the lemma. t
u

Theorem 4.43. Assume that G has an s, t-path. When the algorithm described
in Lemma 4.42 is ran at least T · ek times, then the probability that it does not
find an s, t-path is at most e−T .

16
N. Alon, R. Yuster and U. Zwick, Color-coding, Journal of the ACM 42 (1992),
pp. 844–856.
118 4 Fixed-parameter Algorithms

Proof. The probability that the algorithm of Lemma 4.42 does not find an
s, t-path is at most
1 − e−k . (4.34)
Thus, after T · ek trials, the probability that the algorithm fails to find the
s, t-path is at most

(1 − p)T/p 6 e−T where p = e−k . (4.35)

t
u

Remark 4.44. using derandomization techniques like hashing schemes this al-
gorithm can be made deterministic.17

4.8 Problems
4.1. Let’s start with a problem from elementary school. Which function grows
faster.
klog(k) or (log(k))k .
How about
2 log(k)+log(k2 )
2k or 22 .
A tricky one:

the inverse Ackermann function α(k) or log(log( · · · log(k)) · · · ).

4.2. Show that the parameterized chromatic number problem (χ(G), k) is not
fixed-parameter tractable.
Hint: Recall that 3-coloring is NP-complete.

4.3. Let S be a set and let

S = { S1 , . . . , Sn }

be a collection of n subsets of S such that |Si | = 3 for all i ∈ {1, . . . , n}. The
3-hitting set problem asks for a subset S0 ⊆ S of minimal cardinality such
that every triple Si of S contains at least one element of S0 . Use the bounded
search-tree technique to show that the parameterized 3-hitting set problem
can be solved in O(3k · n2 ) time.
Hint: Build a search-tree in which every vertex which is not a leaf has degree
at most three.
17
M. Fredman, J. Komlós and E. Szemerédi, Storing a sparse table with O(1) worst
case access time, Journal of the ACM 31 (1984), pp. 538–544.
 4.8 Problems 119

4.4. Find a kernel for the 3-hitting set that we defined in Exercise 4.3.
Hint: Consider an element x ∈ S that appears in at least k2 + 1 subsets Si .
Design a proof, similar to the proof of Lemma 4.8, which shows that x is in
any 3-hitting set with at most k elements.

4.5. Solve the recurrences in (4.1) and (4.2).

4.6. Check that there are 14 minimal chordal embeddings of a labeled 6-cycle.

4.7. Prove Lemma 4.22 on Page 102.

Hint: Prove that the number of minimal chordal embeddings Cn of an (n+2)-
cycle with vertices numbered 1, . . . , n + 2 satisfies the Catalan recurrence
relation
X n
Cn+1 = Ci Cn−i .
i=0

4.8. Let G be a graph and let k ∈ N, k > 3. Design an algorithm that checks
if G has a chordless cycle of length at least k and that out-puts one if it has. Is
your algorithm a fixed-parameter algorithm?

4.9. Can we use the linear-time recognition algorithm for chordal graph of
Tarjan and Yannakakis to find an induced cycle of length at least four?

4.10. Let H be a graph with c vertices. Design an algorithm that checks if

a graph G has an induced subgraph which is isomorphic to H. What is the
timebound for your algorithm?
Hint: This problem is not fixed-parameter tractable. That is easy to see. When
H is an independent set with α vertices, the induced subgraph problem asks
for an independent set in G with α vertices.

4.11. Check the details in the proof of Theorem 4.41. Especially

(a) check Formula 4.28,
(b) check Formula 4.29, and
(c) check Formula 4.31.

4.12. Design a fixed-parameter algorithm that checks if a graph G = (V, E)

has a set F of at most k edges such that H = (V, E \ F) is bipartite.
Hint: The following paper describes an O(2k · m2 ) algorithm.
J. Guo, J. Gramm, F. Hüffner, R. Niedermeier and S. Wernicke, Improved
fixed-parameter algorithms for two feedback set problems, Proceedings WADS
2005, Springer, LNCS 3608 (2005), pp. 158–168.

4.13. Given a graph G = (V, E). The k-maximum cut problem asks if there is a
partition {A, B} of V such that at least k edges have one endpoint in A and the
other endpoint in B. Design a fixed-parameter algorithm for the k-maximum
cut problem.
Hint: This is a difficult exercise. The following paper describes a reduction to
120 4 Fixed-parameter Algorithms

a kernel with 2k edges.

E. Prieto, The method of extremal structure on the k-maximum Cut prob-
lem, Proceedings CATS 05, In (M. Atkinson and E. Dehne, eds.) Conferences
in Research and Practice in Information Technology 41, Australian Computer
Society, 2005.
Another method is described in this paper:
M. Mahajan and V. Raman, Parameterizing above guaranteed values: MaxSat
and MaxCut, Journal of Algorithms 31 (1999), pp. 335–354.
5
Decomposition Trees

We have seen a few examples of decomposition trees of graphs already.

1. When G is a chordal graph it has a clique tree. We explained that concept
in Section 3.5.1.
2. When G is a cograph then it has a cotree. We explained that in Sec-
tion 3.3.1.
3. In Section 3.4.1 we explained the notion of a decomposition tree for
distance-hereditary graphs.
4. Interval graphs are chordal graphs. Interval graphs have a special clique
tree, which is a path. We explained that in Theorem 3.72 on Page 74.
In Section 3.1 we hinted at a decomposition tree for perfect graphs that de-
composes the graph into four basic classes of graphs. Furthermore, in Chap-
ter 3 we have seen a few examples that show that a decomposition tree for a
graph can be very useful for solving hard problems on that graph.

In this chapter we look at some parameterized decomposition trees. The

advantage of the parametrization is that we no longer are restricted to some
special class of graphs. Any graph can be decomposed using the tree decom-
positions by a suitable choice of the parameter.

We look at two kinds of decomposition trees. The first one is based on the
clique trees for chordal graphs and the second one is based on the decompo-
sition trees for distance-hereditary graphs.

Research on decomposition trees really took off with the graph minor the-
ory. We review some of that theory in the next section.

5.1 Graph minors

Let’s start with a basic lemma on natural numbers.
122 5 Decomposition Trees

Lemma 5.1. Consider a sequence of natural numbers

n1 , n2 , n3 , . . . (5.1)

There exists an infinite subsequence which is nondecreasing.

Proof. Define
I = { i ∈ N | ∀j>i nj < ni }. (5.2)
The subsequence of (5.1) of the numbers ni with i ∈ I is strictly decreasing.
Since it is bounded from below by 1, it is finite.
Now let 
0 if I = ∅, and
k= (5.3)
max { i | i ∈ I } if I 6= ∅.
Since I is finite this maximum exists. Consider the sequence

nk+1 , nk+2 , . . .

For each element n` with ` > k, there exists some element nm with m > ` and
nm > n` , since ` ∈
/ I. Then it is easy to construct an infinite nondecreasing
subsequence. t
u

Consider an infinite sequence of graphs

G1 , G2 , . . . (5.4)

Let ni be the number of vertices of the graph Gi , for all i. By lemma 5.1,
there exists an infinite subsequence of (5.4) in which the number of vertices
is nondecreasing. A similar proof shows that there is also an infinite subse-
quence of graphs of which the number of vertices and edges is nondecreasing.
So we may assume that the sequence of graphs is ‘nondecreasing,’ if we order
them by their numbers of vertices and edges.

Can we take a different ordering? Suppose that we order the set of all
graphs by the induced subgraph relation. For two graphs G and H define
G  H if the graph G is an induced subgraph of H. Notice that this relation is
transitive, that is,

if F  G and G  H then F  H.

Not every pair of graphs is related by . We call  a quasi-order. If we call

two graphs G and H ‘the same’ if G  H and H  G, then the order is called a
partial order. (In that case, we call two graphs that are isomorphic ‘the same.’)
 5.1 Graph minors 123

Question:
Is it true that there always exist some integers i and j with i < j such
that Gi  Gj ?

Notice that, by Lemma 5.1 this is true for a sequence (5.1) of natural
numbers; just take two elements of a (infinite) nondecreasing subsequence.
Thus, if  is the ordering of graphs by numbers of vertices and edges, then
the answer to the question is yes.

No; for the sequence of graphs this is not true if  is the ordering by
induced subgraphs. An easy counterexample is

C3 , C4 , C5 , . . . ,

where Ci is a cycle of length i. No cycle Ci is an induced subgraph of another

cycle Cj .
For the sequence of paths

P1 , P2 , P3 , . . . ,

where Pi is the path with i vertices, the statement is true of course, since
any path Pi is an induced subgraph of a path Pj with j > i. But for general
sequences of graphs we cannot have that.

To make the statement above true for general sequences of graphs, we

need to relax the condition that some graph Gi is an induced subgraph of an-
other graph Gj . If we take the subgraph-relation, that is, if we only insist that
Gi is a subgraph of Gj , then this is still not true. The same counterexample
applies.

The way to make the statement true is to look at graph minors. To define
graph minors we need the concept of an edge contraction.
Definition 5.2. Let G = (V, E) be a graph. Let e = {x, y} be an edge in G. Con-
tracting the edge e in G is the following operation which transforms G into a
graph G0 . Replace the two vertices x and y by one vertex, say xy. The neighbor-
hood of xy in G0 is the set

N(x) ∪ N(y) \ {x, y}.

Thus, contracting the edge e = {x, y} squeezes the edge down to a single
vertex xy. When z is a vertex that is adjacent to both x and y, then squeezing
the edge {x, y} to a single vertex xy, creates two edges {xy, z} in G0 . The double
edge {xy, z} is then replaced in G0 by a single edge.

For example, let C be a cycle with n vertices. If we contract an edge in C

we get a cycle with n − 1 vertices.
124 5 Decomposition Trees

We now have the following. If (5.4) is an infinite sequence of graphs Gi

such that each Gi is a cycle, then there exist i < j such that Gi is obtained
from Gj by a sequence of edge contractions. Just take two cycles Gi and Gj
with i < j such that Gj has at least as many vertices as Gi . We can obtain Gi
from Gj by contracting sufficiently many edges in Gj (zero when Gi = Gj ).
In other words, we now have the following. Define G  H if G is obtained
from H by a sequence of edge contractions. If (5.4) is a sequence of cycles,
then there exist i < j such that Gi  Gj .

The minor ordering is a little bit more general.

Definition 5.3. A graph G is a minor of a graph H if G can be obtained from H

by a sequence of operations, where each operation is one of the following.
(i) Deletion of a vertex.
(ii) Deletion of an edge.
(iii) Contraction of an edge.

Alternatively, we can define graph minors as follows. Let G = (V, E) be a

graph and let
V = { x1 , . . . , xn }.
Then G = (V, E) is a minor of a graph H = (V 0 , E0 ) if there exist n subsets of
vertices
Xi ⊆ V 0 , for i = 1, . . . , n
such that
(a) Xi ∩ Xj = ∅, when i 6= j, and
(b) H[Xi ] is connected for all i, and
(c) Some vertex in Xi is adjacent to some vertex in Xj when {xi , xj } ∈ E.
See Figure 5.1 on the facing page for an illustration.

Theorem 5.4 (The Graph Minor Theorem). Let G m H if G is a minor of

H. Then in any infinite sequence of graphs

G1 , G2 , . . .

there exist two elements i < j with Gi m Gj .

It took Robertson and Seymour more than ten years to prove this theorem.
They finished the proof in 2004, and they wrote it down in a sequence of 20
papers. The total length of the proof is more than 500 pages; so we skip it.
 5.1 Graph minors 125

x1 x2 x3
G

H X1

X2
X3

Fig. 5.1. The graph G is a minor of the graph H. Each induced subgraph H[Xi ] con-
tracts to the vertex xi . If we only contract the subsets Xi , without removing other
edges, then we end up with a ‘paw,’ that is a triangle {x1 , x2 , x3 } with one pendant
vertex adjacent to x1 . The graph G is a subgraph of the paw.

The first step in the proof is Kruskal’s theorem. Kruskal proved the theorem
in 1960 for the case where G1 , G2 , . . . is a sequence of trees.1 Of course, for
trees it is sufficient to consider only edge contractions; if T1 and T2 are trees
then T1 is a minor of T2 if and only if T1 can be obtained from T2 by a sequence
of edge contractions.
An important consequence of the graph minor theorem is
The Finite Basis Theorem.

Theorem 5.5. Let G be a class of graphs which is closed under taking minors.
That is, if G ∈ G and if H is a minor of G, then H ∈ G. There exists a finite set of
graphs Ω such that G ∈ G if and only if no element of Ω is a minor of G.
Proof. Let Ω be the set of graphs that are not in G but for which every proper
minor is in G. So, if G ∈ Ω then G ∈/ G, but if we delete a vertex or an edge,
or if we contract an edge in G, then the new graph is in G.
Suppose that Ω is not finite. Then we can construct an infinite sequence
G1 , G2 , . . . ,
of graphs in Ω. Furthermore, we may assume that no two graphs Gi and Gj
are the same (isomorphic). By Theorem 5.4, there exist i < j such that Gi is
a minor of Gj . Since Gi and Gj are not the same, Gi is a proper minor of Gj .
But this is a contradiction; every proper minor of Gj is in G, so Gi ∈
/ Ω. t
u
1
J. Kruskal, Well-quasi-ordering, the tree theorem, and Vazsonyi’s conjecture, Trans-
actions of the American Mathematical Society 95 (1960), pp. 210–225.
126 5 Decomposition Trees

The set Ω is called the obstruction set for the class G.

Let’s look at an example. Let G be the class of all planar graphs. Let G ∈ G.
If we delete a vertex x from G, then the remaining graph G − x is also planar.
It is also easy to see that, if we delete an edge from G, then the remaining
graph is planar. Finally, if we contract an edge in G, then it is fairly easy to
see that the new graph G0 is still planar. Thus the class G of planar graphs is
closed under taking minors. By Theorem 5.5 there is a finite obstruction set
Ω. For the class of planar graphs this set is

Ω = { K5 , K3,3 }.

Thus a graph is planar if and only if it has no K5 or K3,3 as a minor. Probably

you know this theorem as Kuratowski’s theorem, which was originally formu-
lated a little bit different (without using edge contractions). Wagner showed
that Kuratowski’s theorem is equivalent to the formulation above.2

In the next section we illustrate the power of the graph minor theorem
by another example. We end this section with another important result of
Robertson and Seymour.

Theorem 5.6 (The Minor Test Theorem). Let H be a graph. There exists an
O(n3 ) algorithm that tests if H m G for graphs G, where n is the number of
vertices of G.

In other words, the parameterized graph minor problem (m (G), H),
which asks if the fixed graph H is a minor of a graph G, is fixed-parameter
tractable. Notice that, when H is not fixed the problem is NP-complete. For
example, when we take H a cycle with n vertices and G is a graph on n ver-
tices, then H m G if and only if G has a Hamiltonian cycle. To test if G is
Hamiltonian is NP-complete.

Notice that Theorem 5.6 provides an O(n3 ) algorithm to test if a graph G is

planar. The theorem says we can test if G has a K5 -minor in O(n3 ) time. Also,
we can test if G has a K3,3 -minor in O(n3 ) time. So, because the obstruction
set is finite (it has only two elements), the total time to test if a graph is planar
takes 2 · O(n3 ) = O(n3 ).
Of course, we know that there is a linear-time algorithm to test planarity,
but the result is much more general; it says that the recognition problem (see
Page 35) for graph classes that are closed under taking minors can be solved
in polynomial time.

2
K. Wagner, Über eine Erweiterung eines Satzes von Kuratowski, Deutsche Mathe-
matik 2 (1937), pp. 280–285.
 5.2 Parameterized feedback vertex set 127

Theorem 5.7. Let G be a class of graphs which is closed under taking minors.
There exists an O(n3 ) algorithm to test if a graph G ∈ G.

Proof. By Theorem 5.5 the class G has a finite obstruction set. By Theorem 5.6
we can test for each element in Ω whether it is a minor of a graph G or not.
Since Ω is finite we only need to perform a constant number of these tests.
This proves the theorem. t
u

Remark 5.8. Kawarabayashi and Wollan are working on a shorter proof of

Theorem 5.5. Recently, they published a shorter proof of Theorem 5.6. They
claim that it shortens the proof of the graph minor algorithm by at least 200
pages.3

Remark 5.9. Kawarabayashi, Kobayashi and Reed recently announced an O(n2 )

algorithm for the minor test, that is, Theorem 5.6.4

5.2 Parameterized feedback vertex set

Recall Definition 3.46 on Page 62. A set F of vertices in a graph G = (V, E) is
a feedback vertex set in G if every cycle in G has at least one vertex in F. In
other words, F is a feedback vertex set in G if G − F is a forest.

For a graph G let f(G) denote the minimal cardinality of a feedback ver-
tex set in G. In this section we show that the parameterized feedback vertex
set problem (f(G), k) is fixed-parameter tractable. We already proved that in
Section 4.5.1 on page 108 but in this section we give a much easier argument.

Lemma 5.10. Let k ∈ N ∪ {0}. Let G(k) be the class of graphs G with f(G) 6 k.
The class G(k) is closed under taking minors.

Proof. Let G ∈ G(k). We prove that every minor of G is in G(k).

Let F be a feedback vertex set in G = (V, E) with |F| 6 k.
Let x ∈ V and let G0 = G − x. If x ∈ F then F0 = F \ {x} is a feedback vertex set
in G0 , since G0 − F0 is a forest. When x ∈
/ F, then G0 − F is a forest, since the
class of forests is minor closed.

3
K. Kawarabayashi and P. Wollan, A shorter proof of the graph minor algorithm –
The unique linkage theorem, proceedings STOC’10, ACM (2010), pp. 687–694.
4
K. Kawarabayashi, Y. Kobayashi and B. Reed, The disjoint paths problem in
quadratic time, Journal of Combinatorial Theory, Series B 102 (2012), pp. 424–435.
128 5 Decomposition Trees

Let e = {x, y} ∈ E. Let G0 be the graph obtained from G by deleting the edge
from E. Then F is a feedback vertex set in G0 , since G0 − F is a forest.
Let e = {x, y} ∈ E and let G0 be the graph obtained from G by contracting the
edge {x, y} to a single vertex xy. First assume that x ∈ / F and that y ∈/ F. Then
F is a feedback vertex set in G0 since the class of forests is closed under taking
minors.
Assume that x ∈ F and that y ∈ / F. Define

F0 = (F \ {x}) ∪ {xy}.

Then F0 is a feedback vertex set for G0 since G0 − F0 = G − F − {y} is a forest.

Assume that x ∈ F and that y ∈ F. Define

F0 = (F \ {x, y}) ∪ {xy}.

Then F0 is a feedback vertex set in G0 since G0 − F0 = G − F is a forest.

This proves the lemma, since |F0 | 6 k in all cases. t
u

We’re done! The following theorem states the result.

Theorem 5.11. The parameterized feedback vertex set problem (f(G), k) is

fixed-parameter tractable.

Proof. Let k ∈ N ∪ {0}. Let G(k) be the class of graphs that have a feedback
vertex set with at most k vertices. By Lemma 5.10 the class G(k) is minor
closed. By Theorem 5.5 there is a finite obstruction set Ω(k) such that

G ∈ G(k) if and only if ∀H∈Ω(k) H 6m G.

By Theorem 5.6 we can test H m G in O(n3 ) time, for each H ∈ Ω(k).

Since |Ω(k)| is constant, the total time needed to test if G ∈ G(k) takes O(n3 )
time. t
u

5.3 Treewidth
Recall Definition 4.20 on Page 101. Let G = (V, E) be a graph. A chordal
embedding of G is a chordal graph H = (V, E0 ) with E ⊆ E0 .

Definition 5.12. Let G = (V, E) be a graph. The treewidth of G is

tw(G) = min { ω(H) − 1 | H is a chordal embedding of G }. (5.5)

 5.3 Treewidth 129

Robertson and Seymour came up with the graph parameter treewidth dur-
ing their research on graph minors.5 It turns out that, if G is a class of graphs
which is closed under taking minors, and if G does not contain all planar
graphs then there exists a k ∈ N ∪ {0} such that all graphs in G have treewidth
at most k.

For example, let G(`) be the class of graphs that have a feedback vertex
set with at most ` vertices. By Lemma 5.10 this class is closed under taking
minors. Furthermore, G(`) does not contain all planar graphs. For example, if
we take a sequence of larger and larger grids then it is easy to see that these
have an increasing number of vertices in a minimum feedback vertex set. The
result of Robertson and Seymour mentioned above now says that there exists
a k (which is a function of `) such that all graphs in G(`) have treewidth at
most k.

In this section we take a close look at this important graph parameter.

For example, let T be a tree. Then T has no cycles. Thus T is chordal, since
it has no induced cycles of length more than three. Any chordal embedding
of T has a clique number at least equal to the clique number of T . Therefore,

0 if T has only one vertex
tw(T ) = (5.6)
1 if T has at least two vertices.

Thus any nontrivial tree has treewidth one. This explains why we sub-
tract one from ω(H) in Definition 5.12. Namely, in this way we have that any
nontrivial tree has treewidth one, which is nice. In the following example we
show the converse, that is, if the treewidth of a graph G is at most one then
G is a forest.

Let’s look at another example. Consider a cycle C. We claim that the

treewidth of C is at least two. Assume that it is one. Let H be a chordal em-
bedding of C and assume that ω(H) = 2.

Recall Theorem 3.60 on page 68. Since H is chordal it has a perfect elimi-
nation ordering, say
[x1 , . . . , xn ].
By Theorem 3.66 on page 71, every maximal clique in H is one of the sets

Ωi = { xj | j > i and xj ∈ N[xi ] }.

Since ω(H) = 2, we have that

5
N. Robertson and P. Seymour, Graph minors. II. Algorithmic aspects of tree-width,
Journal of Algorithms 7 (1986), pp. 309–322.
130 5 Decomposition Trees

|Ωi | 6 2 for all i ∈ { 1, . . . , n }.

Thus the perfect elimination ordering removes vertices one by one, and at
each step the vertex has at most one neighbor in the remaining graph. But
this shows that H is a forest! Since C is a subgraph of H, also C is a forest,
which is a contradiction because C is a cycle. Thus the treewidth of C is at
least two.

We now show that the treewidth of C is exactly two. Let x be a vertex of

C. Add edges from x to all other vertices in C. The new graph is chordal and
has clique number three. This proves that tw(C) = 2.

Thus the treewidth of a graph G is one if and only if G is a forest.

In Section 5.3.1 we look at a greedy algorithm that checks if the treewidth

of a graph G is at most two.

In the following theorem we prove that the class of graphs with treewidth
at most k is closed under taking minors. For k = 1 this is true, since the class
of graphs with treewidth at most one is the class of all forests.

To prove the general case we start with three easy lemmas.

Lemma 5.13. Let G be a graph and let G0 be a subgraph of G. Then

tw(G0 ) 6 tw(G).

Proof. Let G = (V, E) be a graph. Write k = tw(G). Let G0 = (V 0 , E0 ) be a

subgraph of G. Let H be a chordal embedding of G with ω(H) = k + 1. Then
H[V 0 ] is a chordal graph, since the class of chordal graphs is hereditary. The
graph H[V 0 ] is a chordal embedding of G0 since every edge in E0 is also an
edge in E, and so it is an edge in H[V 0 ] since H is a chordal embedding of G.
Obviously,
ω(H[V 0 ]) 6 ω(H) = k + 1,
and this implies that tw(G0 ) 6 k.
This proves the lemma. t
u

Lemma 5.13 shows that the class of graphs with treewidth at most k is
closed under taking subgraphs. To prove that this class is also closed under
edge contractions, we prove this first for the class of chordal graphs.

By the way, of course the class of chordal graphs is not closed under mi-
nors, since it is not even closed under taking subgraphs. Every graph is a
subgraph (and thus also a minor) of a large enough clique!
 5.3 Treewidth 131

Lemma 5.14. The class of chordal graphs is closed under edge contractions.

Proof. Let G = (V, E) be a chordal graph. Let e = {x, y} be an edge in G and

let G0 be the graph obtained from G by contracting the edge e to a single
vertex xy. We prove that G0 is chordal.
Recall Theorem 3.64 on Page 70. Since G is chordal there exist a tree T and a
collection of subtrees of T
{ Tx | x ∈ V } (5.7)
such that any two vertices a and b are adjacent if and only if Ta ∩ Tb 6= ∅.
Consider the subtrees Tx and Ty . Since x and y are adjacent Tx ∩ Ty 6= ∅. For
the new vertex xy define the subtree

Txy = Tx ∪ Ty . (5.8)

Then Txy is a subtree of T . Any other vertex z is adjacent to xy in G0 if and

only if z is adjacent to x or to y in G. Any subtree Tz intersects Tx or Ty if and
only if Tz intersects Txy .
Thus the graph G0 is the intersection graph of a set of subtrees of a tree, and
so, G0 is chordal. t
u

Lemma 5.15. Let G = (V, E) be a chordal graph and let e = {x, y} be an edge
in G. Let G0 be the graph obtained from G by contracting the edge e to a single
vertex xy. Then
ω(G0 ) 6 ω(G). (5.9)

Proof. Let Ω0 be a maximal clique in G0 . If xy ∈

/ Ω0 then Ω0 is a clique in G
and so
|Ω0 | 6 ω(G). (5.10)

Now assume that xy ∈ Ω0 . If x is adjacent to all other vertices z ∈ Ω0 \ {xy}

then
Ω = (Ω0 \ {xy}) ∪ {x}
is a clique in G and so (5.10) holds.
Of course, by symmetry, (5.10) also holds when y is adjacent to all other
vertices of Ω0 \ {xy}.
Assume that there exist vertices x0 and y0 in Ω0 \ {xy} such that

x0 ∈ N(x) \ N(y) and y0 ∈ N(y) \ N(x). (5.11)

Then [x, x0 , y0 , y] is a 4-cycle in G which is a contradiction. t

u

The proof of the next theorem is now a piece of cake.

132 5 Decomposition Trees

Theorem 5.16. Let k ∈ N ∪ {0}. Let T(k) be the class of graphs G with
tw(G) 6 k.
Then T(k) is minor closed.
Proof. Let G = (V, E) be a graph and assume that tw(G) 6 k. Let G0 be a
minor of G. We prove that tw(G0 ) 6 k.
When G0 is a subgraph of G then the claim follows from Lemma 5.13.
Let e = {x, y} be an edge in G. Assume that G0 is obtained from G by contract-
ing the edge e to a single vertex xy.
By definition, there exists a chordal embedding H of G with ω(H) = k + 1.
Since H is a chordal embedding of G, e is also an edge in H.
By Lemma 5.14 the graph H0 , obtained from H by contracting the edge e in H
to a single vertex xy, is chordal. Then H0 is a chordal embedding of G0 .
By Lemma 5.15 ω(H0 ) 6 ω(H).
This proves the theorem. t
u

Via the theory on graph minors we immediately obtain the following the-
orem.
Theorem 5.17. The parameterized treewidth problem (tw(G), k), which asks
if a graph G has treewidth at most k, is fixed-parameter tractable. For each
k ∈ N ∪ {0} there exists an O(n3 ) algorithm which checks if the treewidth of a
graph G is at most k.
Proof. By Theorem 5.16 the class T(k) of graphs with treewidth at most k is
minor closed. By Theorem 5.5 there exists a finite obstruction set Ω(k) and,
by Theorem 5.6 we can test for each element H ∈ Ω(k) in O(n3 ) time if it is
a minor of G. Now tw(G) 6 k if and only if none of the graphs in Ω(k) is a
minor of G. t
u

Remark 5.18. The theorem above only shows that there exists an O(n3 ) al-
gorithm which checks if a graph has treewidth at most k. The graph minor
theory does not provide the algorithm since the obstruction set is unknown.
However, for each k ∈ N ∪ {0} there is an explicit linear-time algorithm which
checks if a graph has treewidth at most k.6 7 In Section 5.3.3 we show that
there exists an O(nk+2 ) algorithm which checks if the treewidth of a graph
is at most k. Although this is not a fixed-parameter algorithm, it is useful for
small values of k.
6
T. Kloks, Treewidth - Computations and Approximations, Springer-Verlag, Lecture
Notes in Computer Science 842, 1994.
7
L. Perković and B. Reed, An improved algorithm for finding tree decompositions of
small width, International Journal of Foundations of Computer Science 11 (2000),
pp. 365–371.
 5.3 Treewidth 133

Remark 5.19. The obstruction set for the class of graphs with treewidth one
has only one element, namely the triangle. This is so because a graph is a
forest if and only if it has no cycle. Therefore, a graph is a forest if and only if
it has no triangle as a minor. The obstruction set for graphs of treewidth two
also has only one element, namely K4 . The following Figure 5.2 shows the
obstruction set for the class of graphs with treewidth three.8 9 10

Fig. 5.2. The figure shows the obstruction set for treewidth three.

5.3.1 An algorithm for treewidth two

To get in the mood, let’s do something easy, first.

In this section we show that there exists a linear-time algorithm to check

if the treewidth of a graph is at most two.

We present the algorithm first. We prove that it is correct in Theorem 5.21.

8
H. Bodlaender, A partial k-arboretum of graphs with bounded treewidth, Theoretical
Computer Science 209 (1998), pp. 1–45.
9
S. Arnborg, A. Proskurowski and D. Corneil, Forbidden minors characterization of
partial 3-trees, Discrete Mathematics 80 (1990), pp. 1–19.
10
A. Satyanarayana and L. Tung, A characterization of partial 3-trees, Networks 20
(1990), pp. 299–322.
134 5 Decomposition Trees

Let G = (V, E) be a graph. The algorithm to check if the treewidth of G is

at most two runs as follows.
1. If there exists a vertex x with at most one neighbor then delete it from the
graph. Let G0 = G − x. Check if the treewidth of G0 is at most two. The
treewidth of G is at most two if and only if the treewidth of G0 is at most
two.
2. Let x be a vertex with exactly two neighbors, y and z. Assume that y and
z are adjacent. Then let G0 = G − x. The treewidth of G is at most two if
and only if the treewidth of G0 is at most two.
3. Let x be a vertex with exactly two neighbors, y and z. Assume that y and
z are not adjacent. Then add the edge {y, z} to the graph and remove the
vertex x. Let G0 be this graph. The treewidth of G is at most two if and
only if the treewidth of G0 is at most two.
4. Assume that every vertex in G has at least three neighbors. Then the
treewidth of G is more than two.

As you see, in the algorithm we use the fact that every graph with
treewidth at most two has a vertex with at most two neighbors. We prove
that in the following lemma. In Exercise 5.18 we ask you to prove a similar
lemma for the graphs with treewidth at most k.

Lemma 5.20. Let G = (V, E) be a graph and assume that tw(G) 6 2. Then
every subgraph of G has a vertex with at most two neighbors.
Proof. By Lemma 5.13 the class of graphs with treewidth at most two is closed
under taking subgraphs. Therefore, it is sufficient to prove the claim for G.
Let H be a chordal embedding of G with ω(H) 6 3. To avoid confusion,
we use the notation NH (x) to denote the neighborhood of a vertex x in the
chordal graph H.
By Theorem 3.60 on Page 68 there exists a perfect elimination ordering for H,
say
σ = [x1 , . . . , xn ].
The vertex x1 is a simplicial vertex in H and so its neighborhood is a clique.
Since ω(H) 6 3 we have that |NH (x1 )| 6 2, and so x1 has at most two
neighbors in H. Because H is a chordal embedding of G we have that

NG (x1 ) ⊆ NH (x1 ),

and so the vertex x1 has at most two neighbors in G.

This proves the lemma. t
u

We now show that the algorithm above is correct.

 5.3 Treewidth 135

Theorem 5.21. There exists an O(n) algorithm which checks if the treewidth of
a graph G is at most two.
Proof. By Lemma 5.20, if G has n vertices and more than 2n edges then
tw(G) > 2. Assume that G has at most 2n edges. Then we can compute
the degree of every vertex x in O(n) time. This shows that the algorithm that
we described can be implemented to run in O(n) time.
Let x be a vertex with at most two neighbors.
First assume that x is isolated. Let G0 = G − x. By Lemma 5.13, if tw(G0 ) > 2
then also tw(G) > 2. Assume that tw(G0 ) 6 2. Let H0 be a chordal embedding
for G0 with ω(H0 ) 6 3. Add the vertex x as an isolated vertex to H0 and let H
be this graph. Then H is a chordal embedding of G and ω(H) 6 2.
Assume that x has one neighbor, say y. Let G0 = G − x. If tw(G0 ) > 2, then
also tw(G) > 2. Assume tw(G0 ) 6 2 and let H0 be a chordal embedding of G0
with ω(H0 ) 6 3. Add the vertex x to H0 and make it adjacent to y. Let H be
this graph. Then H is a chordal embedding of G and ω(H) 6 3.
Assume that x has two neighbors, y and z, and assume that y and z are
adjacent. Let G0 = G−x. If tw(G0 ) > 2 then tw(G) > 2 and so we may assume
that tw(G0 ) 6 2. Let H0 be a chordal embedding of G0 with ω(H0 ) 6 3. Let
σ0 = [x2 , . . . , xn ]
be a perfect elimination ordering for H0 . Let H be the graph obtained from H0
by adding the vertex x to H0 and by making x adjacent to y and z. Consider
σ = [x, x2 , . . . , xn ].
Since x is adjacent to y and z and since {y, z} is an edge in H, the vertex x is a
simplicial vertex in H. This proves that σ is a perfect elimination ordering for
H. Thus, H is a chordal embedding for G and ω(H) 6 2.
Finally, assume that x has exactly two neighbors, say y and z, and assume that
y and z are not adjacent. Let G0 be the graph obtained from G by adding the
edge {y, z} and by removing the vertex x. We claim that G0 is a minor of G. To
see that, observe that contracting the edge {x, y} in G produces the graph G0 .
By Theorem 5.16, if tw(G0 ) > 2 then tw(G) > 2. Assume that tw(G0 ) 6 2
and let H0 be a chordal embedding of G0 with ω(H0 ) 6 3. Notice that {y, z} is
an edge in H0 . Add the vertex x as a simplicial to H0 , by making it adjacent to
y and z. Then H is a chordal embedding of G and ω(H) 6 3.
This proves the theorem. t
u

5.3.2 k-Trees
When a graph G has treewidth k then it has an embedding into a chordal
graph H with clique number k + 1. In this section we show that there is a
special chordal graph embedding for G.
136 5 Decomposition Trees

Definition 5.22. Let k ∈ N ∪ {0}. A k-tree is a chordal graph in which

(i) every maximal clique has cardinality k + 1, and
(ii) every minimal separator is a k-clique.

Theorem 5.23. Let G be a graph and let k = tw(G). Then there exists an em-
bedding of G into a k-tree.

Proof. Let H be a chordal graph embedding of G with ω(H) = k + 1. Write

NH (x) for the neighbors of x in the graph H.
We prove that there exists an embedding H0 of H such that H0 is a k-tree.
By Theorem 3.60 there exists a perfect elimination ordering for H, say

σ = [x1 , . . . , xn ].

For i = n down to 1 add neighbors to NH (xi ) as follows. If i > n − k then

make xi adjacent to all vertices of

{ xi+1 , . . . , xn }.

This step makes a (k + 1)-clique of {xn−k , . . . , xn }.

For i = n − k − 1 down to 1 consider

Ni = NH (xi ) ∩ { xi+1 , . . . , xn }.

Then |Ni | 6 k since it is a clique in H of cardinality at most k.

By induction, every maximal clique in

H0i = H0 [{ xi+1 , . . . , xn }]

has cardinality k + 1. Since Ni is a clique it is contained in a maximal clique

Ωi in H0i . Furthermore, |Ni | 6 k.
Choose a subset N0i ⊆ Ωi which contains Ni and which has cardinality k.
Make xi adjacent to all vertices N0i .
The graph H0 is a chordal graph embedding of G. Every maximal clique in H0
has cardinality k + 1 and every minimal separator is a k-clique.
This proves the theorem. t
u
 5.3 Treewidth 137

5.3.3 An O(nk+2 ) algorithm for treewidth

Computing the treewidth of a graph is NP-complete. Arnborg, et al., designed
the first polynomial-time algorithm which checks if a graph has treewidth at
most k.11 In this section we explain their algorithm.

Recall Theorem 3.56 on Page 66 which says that a graph is chordal if and
only if every minimal separator is a clique.

Lemma 5.24. Let G = (V, E) be a chordal graph and let S be a minimal sepa-
rator in G. Let C be a component of G − S. Then C contains a vertex x such that
x is simplicial in G.
Proof. Consider the graph G[S∪C]. Since the class of chordal graphs is heredi-
tary, this induced subgraph is chordal. If G[S ∪ C] is a clique then we are done.
In that case every vertex x ∈ C is simplicial since N(x) ⊆ S ∪ C and so N(x) is
a clique.
Otherwise, there are two nonadjacent vertices a and b in G[S ∪ C]. Let S0 be
a minimal a, b-separator in G[S ∪ C]. We claim that S0 is also a minimal a, b-
separator in G. To see this, assume that there exists an a, b-path P in G − S0 .
Then P must contain at least one vertex which is not in S ∪ C. But then P
contains two vertices from S, and so P has a shortcut. This proves the claim.
Consider the components of G − S0 . Let Ca and Cb be the components of
G − S0 which contain a and b, respectively. Since S is a clique, the set S \ S0
can have vertices in at most one of Ca and Cb . Assume that Ca contains no
vertices from S. Then Ca ⊆ C.
Assume that S0 ⊆ S. Then Ca = C, b ∈ S and |S0 | < |S|. In that case we can
remove vertices of S that have no neighbors in C. By induction there exists a
vertex in Ca which is simplicial in G.
Otherwise, when S0 has at least one vertex of C, then |Ca | < |C|. By induction
Ca contains a vertex x which is simplicial in G. t
u

Let G be a chordal graph and let S be a minimal separator in G. Let C

be a component of G − S. By Lemma 5.24 there exists a perfect elimination
ordering for G which eliminates all vertices of C first.

The algorithm of Arnborg, et al., is based on the following observation.

Let H be a k-tree embedding of a graph G. Let S be a minimal separator

in H. Then |S| = k. Furthermore, S is a separator in G. Let C1 , . . . , Ct be the
components of G − S.

11
S. Arnborg, D. Corneil and A. Proskurowski, Complexity of finding embeddings in a
k-tree, SIAM Journal on Algebraic and Discrete Methods 8 (1987), pp. 277–284.
138 5 Decomposition Trees

Lemma 5.25. There exists a k-tree embedding H0 of G such that

1. S is a separator in H0 , and
2. each component Ci of G − S is a component of H0 − S, and
3. each H0 [S ∪ Ci ] is a k-tree.

Proof. The proof is easy.

Let Gi be the subgraph of G induced by S ∪ Ci . The graphs H[S ∪ Ci ] are
chordal embeddings of G[S ∪ Ci ]. Since the treewidth of each H[S ∪ Ci ] is at
most k, there exist k-tree embeddings Hi for H[S ∪ Ci ]. Define the k-tree H0
as the union of the graphs Hi . t
u

Theorem 5.26. Let G be a graph and let k ∈ N ∪ {0}. There exists an O(nk+2 )
algorithm which checks if the treewidth of G is at most k.

Proof. When G has at most k + 1 vertices then tw(G) 6 k. In that case the
algorithm makes a clique of G and it reports YES. Otherwise, if G has more
than k + 1 vertices, by Theorem 5.23 the graph has tw(G) 6 k if and only if G
has a k-tree embedding. The algorithm that we describe below finds a k-tree
embedding H if tw(G) 6 k.
The algorithm first makes a list of all pairs (S, C) where
(i) S is a separator in G and |S| = k, and
(ii) C is a component of G − S.
Notice that G has at most nk separators S of cardinality k. The number of
components in G − S is at most n for each separator S, so the list contains at
most nk+1 pairs (S, C).
For a pair (S, C) let G∗ (S, C) be the graph obtained from G[S ∪ C] by making
a clique of S.
The algorithm checks if there exists a separator S with |S| = k, such that for
each component C of G − S, G∗ (S, C) has a k-tree embedding. If that is the
case, then the treewidth of G is at most k, and otherwise the treewidth of G
is more than k.
First, the algorithm sorts the pairs (S, C) according to nondecreasing cardi-
nalities |C|. It does that by bucket sort in time O(nk+1 ). It processes the pairs
(S, C) in that order as follows.
When there is a k-tree embedding of G∗ (S, C) then, by Lemma 5.24, there
is a perfect elimination ordering that eliminates the vertices of C first. Let
c ∈ C be the last vertex in this elimination ordering. Then c is adjacent to all
vertices of S in the k-tree embedding. The algorithm tries all vertices c ∈ C as
candidates.
 5.3 Treewidth 139

Consider a pair (S, C). Let c ∈ C and define

S(c) = S ∪ {c}.

Let C1 (c), . . . , Ct (c) be the components of

G[C \ S(c)].

For each i ∈ {1, . . . , t} check if there exists a separator Si (c) such that
(i) { y | y has a neighbor in Ci (c) } ⊆ Si (c), and
(ii) |Si (c)| = k, and
(iii) (Si (c), Ci (c)) is a YES instance, that is, there exists a k-tree embedding
of G∗ (Si (c), Ci (c)).

We claim that the pair (S, C) can be processed in O(|C|2 ) time. Notice that a
k-tree has O(kn) edges, which is O(n) since k is fixed. Thus for each choice
of c ∈ C the components Ci (c) can be determined in O(|C|) time (see Exer-
cise 5.18). Since C is connected, the vertex c has a neighbor in each compo-
nent Ci (c). There are |C| choices for c and O(k) candidates for the separators
Si (c) because
c ∈ Si (c) and |S \ Si (c)| = 1.
For each component Ci (c) it can be checked if there is a suitable separator
Si (c) (which reports YES) in constant time. This proves the claim.
Notice that this shows that the overall time complexity of the algorithm is
bounded by O(nk+2 ), since there are O(nk ) separators S of cardinality k and
the summation of |C|2 over all components C of G − S is bounded by O(n2 ).
If there exist a vertex c ∈ C and a collection of separators Si (c) such that
the answer is YES for all G∗ (Si (c), Ci (c)) then the algorithm answers YES for
G∗ (S, C). Otherwise, it answers NO for G∗ (S, C).
If there exists a separator S with |S| = k such that for all components C of
G − S the algorithm above reports a YES for the pair (S, C), then the treewidth
of G is at most k. Otherwise the treewidth is more than k. t
u

5.3.4 Maximum clique in graphs of bounded treewidth

As an example, we show that, for every k ∈ N ∪ {0} there exists a linear-time

algorithm to compute the clique number for graphs of treewidth at most k.

The usual strategy to solve NP-complete problems for graphs of bounded

treewidth is dynamic programming on the clique tree of the chordal embed-
ding (see Definition 3.62 on page 68). This clique tree of the chordal embed-
ding of the graph G is called the tree decomposition for G.
140 5 Decomposition Trees

For the clique number problem there is an easier algorithm, that we de-
scribe below.

Let k ∈ N ∪ {0} and let G be a graph with treewidth at most k. Then there
exists a chordal embedding H of G with ω(H) 6 k + 1. We mentioned that
there exists a linear-time algorithm to construct the graph H, although we did
not give you the details of this algorithm.

Theorem 5.27. Let k ∈ N ∪ {0}. There exists an O(n) algorithm to compute

ω(G) for graph G ∈ T(k).

Proof. Let G = (V, E) be a graph and assume that tw(G) 6 k. We assume that
we have a chordal embedding H of G with ω(H) 6 k + 1.
Obviously, if M is a maximal clique in G then M is a clique in H. Let M be the
set of all maximal cliques in H. Then

ω(G) = max { |W| | ∃M∈M W ⊆ M and G[W] is a clique }. (5.12)

By Lemma 3.61 on Page 68, H has at most n maximal cliques, where n is the
number of vertices of H.
Let
σ = [x1 , . . . , xn ] (5.13)
be a perfect elimination ordering for H. This can be obtained in linear time by
the algorithm of Tarjan and Yannakakis. Notice that H has at most n · k edges,
and so this algorithm runs in O(n · k) = O(n) time, since k is a constant.
Define
Ni = { xj | j > i and xj ∈ N[xi ] }. (5.14)

Then Formula (5.12) becomes

ω(G) = max { |W| | W ⊆ Ni , i ∈ {1, . . . , n} and G[W] is a clique }. (5.15)

Each Ni has at most k + 1 vertices, so the number of subsets of Ni is bounded

by 2k+1 . Using a suitable data structure we can check in constant time if
two vertices are adjacent in G. So, for each subset W ⊆ Ni we can check in
O(k2 ) time if G[W] is a clique or not. Thus the overall time complexity of this
algorithm is O(2k · k2 · n) = O(n), since k is a constant. t
u
 5.3 Treewidth 141

5.3.5 Chromatic number for graphs of bounded treewidth

In this section we illustrate the standard technique, namely dynamic program-

ming on the decomposition tree. As an example, we show how to compute the
chromatic number for graphs of bounded treewidth.

Theorem 5.28. Let k ∈ N ∪ {0}. There exists a linear-time algorithm that com-
putes χ(G) for graphs G ∈ T(k).

Proof. Let k ∈ N ∪ {0} and let G = (V, E) be a graph in T(k). There exists
a linear-time algorithm which computes a chordal embedding H of G with
ω(H) 6 k + 1. Since H is perfect χ(H) 6 k + 1 and so, since G is a subgraph
of H also χ(G) 6 k + 1.
Consider a clique tree (T , S) for H. For a vertex p in T let Sp ∈ S be the
maximal clique of H that is assigned to p.
Root T at some arbitrary vertex r. For a vertex p in T let Tp be the subtree of
T which is rooted at p. Thus Tr = T .
For a vertex p in T let

Vp = { x ∈ V | ∃i i is a vertex in Tp and x ∈ Si }. (5.16)

Consider all vertex colorings of Sp . A coloring of Sp with ` colors is a partition

P = { Q1 , . . . , Q` }

of Sp into ` color classes. Some of these color classes may be empty.

Since |Sp | 6 k + 1, there are at most `k+1 different partitions of Sp , namely,
each vertex in Sp has one of ` different colors.
A partition P of Sp is valid when no two vertices in the same color class of P
are adjacent in G.
Let 1 6 ` 6 k + 1. Let P be a valid partition of Sp with ` color classes.
Let bp (P, `) be a boolean variable which is TRUE if and only if there exists a
coloring of G[Vp ] with ` colors such that the vertices of Sp are colored as in
P. For each vertex p the algorithm determines all values bp (P, `) as follows.
First assume that p is a leaf. Then Vp = Sp . The algorithm determines all
values bp (P, `) by trying all partitions of Sp . If P has ` color classes (possibly
some empty), and if each color class is an independent set in G[Sp ], then
bp (P, `) is TRUE. Otherwise it is FALSE.
Let p be an internal vertex of T and let c1 , . . . , ct be the children of p in T . For
each i ∈ {1, . . . , t} let Si ∈ S be the maximal clique of H which is assigned to
the vertex ci in T .
142 5 Decomposition Trees

The important observation is that no two vertices

x ∈ Si \ Sp and y ∈ Sj \ Sp

are adjacent when i 6= j. (This follows from Definition 3.62 on Page 68 of the
clique tree.)
Let P = {Q1 , . . . , Q` } be a valid partition of Sp . Then bp (P, `) is TRUE if and
only if for each child ci of p there exists a valid partition P0 = {Q01 , . . . , Q0` }
such that
(i) bci (P0 , `) is TRUE, and
(ii) if Q0j ∈ P0 then
Q0j ∩ Sp ⊆ Qj .

The graph G has a coloring with ` colors if and only if there exists a valid
partition P with ` color classes at the root r, such that br (P, `) is TRUE.
This proves the theorem. t
u

5.3.6 Disjoint cycles

As an example of what’s in the magic box, openened by the Finite Basis The-
orem 5.5 and the Minor Test Theorem 5.6, we look at the problem to find k
vertex-disjoint cycles.

The problem to find a maximum collection of vertex-disjoint cycles in a

graph is NP-complete. Even for chordal graphs, where the problem reduces to
finding a maximum collection of vertex-disjoint triangles, the problem is NP-
complete. Also for planar graphs and for bipartite graphs the disjoint cycles
problem is NP-complete.

In the parameterized version of the disjoint cycles problem we want to find

k vertex-disjoint cycles. Let cp(G) be the maximal number of vertex-disjoint
cycles in a graph G (‘cp’ stands for ‘cycle packing’).

Theorem 5.29. The parameterized problem

(cp(G), k)

is fixed-parameter tractable. For every natural number k there exists an O(n3 )

algorithm to test if a graph G has k vertex-disjoint cycles.

Proof. Let H be the graph which is the union of k triangles. We claim that a
graph G has k vertex-disjoint cycles if and only if H is a minor of G.
First assume that G has k vertex-disjoint cycles, C1 , . . . , Ck . Remove all ver-
tices and edges from G except those vertices that appear in the cycles and
except those edges that connect consecutive vertices in one of the cycles.
 5.3 Treewidth 143

Now contract edges in the cycles such that each cycle becomes a triangle.
This proves that H is a minor of G.
Now assume that H is a minor of G. Let {h1 , . . . , h3k } be the set of vertices of
H. By the alternative definition of a minor, (see Figure 5.1 on page 125) there
exist a collection of disjoint sets of vertices V1 , . . . , V3k , one for each vertex
in H, such that each G[Vi ] is connected, and such that there is an edge in G
between some vertex in Vi and some vertex in Vj whenever hi is adjacent to
hj in H.
Consider a triangle {h1 , h2 , h3 } in H. Let x1 ∈ V1 and x2 ∈ V1 be two vertices
in V1 which are adjacent to some vertex in V2 and V3 , respectively. Since
G[V1 ] is connected, there is a path P1 in G[V1 ] that connects x1 and x2 . In the
same manner we find paths P2 and P3 in G[V2 ] and G[V3 ] such that the union
P1 ∪ P2 ∪ P3 contains a cycle. Since all sets Vi are vertex-disjoint, this gives a
collection of k vertex-disjoint cycles in G.
Thus, an algorithm to check if G has k vertex-disjoint cycles only needs to test
if H is a minor of G. By Theorem 5.6 on page 126 this can be done on O(n3 )
time. t
u

In the rest of this section we show that the time complexity for solving the
parameterized problem (cp(G), k) can be reduced to linear time.

A brick is a 6-cycle. An elementary h-wall, for h ∈ N and h > 2, consists

of h levels and each level contains h bricks. An elementary wall of height 8 is
depicted in Figure 5.3 on the next page. An elementary h-wall, for h > 2 is
defined similarly. A wall is obtained from an elementary wall by subdividing
some of the edges, that is, one may replace each edge in the elementary wall
by an induced path of length at least one.

Theorem 5.30 (Robertson and Seymour12 ). There exists a function

f:N→N

such that if a graph G has treewidth at least f(k) then G contains a wall of
height k as a subgraph.

5
Remark 5.31. The best-known upperbound for f(k) in Theorem 5.30 is 202k .

Notice that a wall of height 2k has at least k2 disjoint cycles.

12
N. Robertson and P. Seymour, Graph minors. V. Excluding a planar graph, Journal
of Combinatorial Theory, Series B 41 (1986), pp. 92–114.
144 5 Decomposition Trees

r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r r
r r r r r r r r r r r r r r r r r

Fig. 5.3. This figure shows an elementary 8-wall.

Theorem 5.32. For each constant k there exists a linear-time algorithm which
checks if a graph has k vertex-disjoint cycles.
Proof (Sketch). Let k ∈ N. We can check in linear time whether the treewidth
of G is at most f(2k) (or, whether it is at most the known upperbound for
f(2k)). If the treewidth is more than f(2k) then, by Theorem 5.30, G con-
tains a wall of height 2k. In that case G has at least k2 disjoint cycles. (See
Remark 5.33 below to see how these cycles can be found in linear time.)
Otherwise, when the treewidth of G is at most f(2k) then one can find the
maximum number of vertex-disjoint cycles in G by standard dynamic pro-
gramming techniques on the tree decomposition of G (that is, the clique tree
of the chordal embedding of G). t
u

Remark 5.33. Assume that the treewidth of a graph G is more than k for some
k ∈ N. Perković and Reed show that there is a linear-time algorithm which
computes a subgraph G0 with treewidth more than k but at most 2k.13
When the treewidth of G is more than f(2k), one can construct the tree
decomposition of a subgraph G0 of G with treewidth more than f(2k) but at
most 2 · f(2k). Then G0 has a wall of height 2k. Since the treewidth of G0 is
bounded by the constant 2 · f(2k) one can construct this wall in linear time
by dynamic programming on the tree decomposition of G0 . This observation
shows that the collection of k vertex disjoint cycles in G can be found in linear
time (if it exists).
13
L. Perković and B. Reed, An improved algorithm for finding tree decompositions of
small width, International Journal of Foundations of Computer Science 11 (2000),
pp. 365–371.
 5.4 Pathwidth 145

5.4 Pathwidth
Pathwidth is a graph parameter which is closely related to treewidth. One can
see that right away from the definition.

An interval embedding of a graph G = (V, E) is a graph H = (V, E0 ) which

is an interval graph with E ⊆ E0 .
Definition 5.34. Let G = (V, E) be a graph. The pathwidth of G is
pw(G) = min { ω(H) − 1 | H is an interval embedding of G }. (5.17)

Lemma 5.35. For any graph G,

tw(G) > pw(G). (5.18)
Proof. Any interval graph embedding of G is a chordal embedding. t
u

Let’s start our investigations with the easy case; trees.

Definition 5.36. A caterpillar is a tree T which has a path P such that every
other vertex of T is adjacent to a vertex in P.

Fig. 5.4. The figure shows a caterpillar.

Lemma 5.37. Let T be a caterpillar. Then


0 if T has one vertex
pw(T ) = (5.19)
1 if T has at least two vertices.

Proof. If T has only one vertex then, by definition, pw(T ) = 0 since ω(T ) = 1.
Assume that T is a caterpillar with at least two vertices. Then pw(T ) > 1 since
T is a chordal graph and
pw(T ) > tw(T ) = 1. (5.20)
Notice that a caterpillar has no asteroidal triple. Therefore, by Theorem 3.77
on page 76 T is an interval graph. This proves that pw(T ) = 1. t
u
146 5 Decomposition Trees

Theorem 5.38. A graph G has pathwidth one if and only if every component of
G is a caterpillar and at least one component has at least two vertices.

Proof. When G has pathwidth one then its treewidth is also one. Therefore,
G is a tree with at least two vertices. If G is not a caterpillar then it has
an asteroidal triple. We claim that any interval embedding of G has clique
number at least three.
We may assume that G is connected and that G has at least two vertices. Let
H be an interval embedding of G and assume that ω(H) = 2. Then H = G,
since adding any edge to G creates a cycle and the treewidth of a cycle is two.
Since H is an interval graph it has no asteroidal triple. This implies that H is
a caterpillar (see Exercise 5.20). t
u

We know that every cycle C has treewidth two. However, not every chordal
embedding of a cycle is an interval graph. For example, a C6 has a chordal
embedding which has an asteroidal triple. However, consider the following
chordal embedding of C. Take any vertex of C and make it adjacent to all
other vertices. Then this is a chordal embedding without any asteroidal triple,
and so it is an interval embedding with clique number three. This shows that
the pathwidth of cycles is two (see Figure 5.5).

u u u u
HH @
@
H
HH@
@
u @u
H u
@u
@  @
 
 
@  @

@u
  u @u


u

Fig. 5.5. This figure shows a 3-sun, which is a triangulation of C6 which has an as-
teroidal triple. The graph on the right is a triangulation of C6 which is an interval
graph.

We want to prove that, for any integer k, the class of graphs with path-
width at most k is closed under taking minors. For k = 1 this is true, since
any minor of a caterpillar is a caterpillar. To prove the general result, we fol-
low the same trajectory as for treewidth. Let’s start with the following easy
lemma.
 5.4 Pathwidth 147

Lemma 5.39. Let G0 be a subgraph of a graph G. Then

pw(G0 ) 6 pw(G). (5.21)

Proof. Write G = (V, E) and G0 = (V 0 , E0 ). By assumption

V0 ⊆ V and E0 ⊆ E. (5.22)

Let H be an interval graph embedding of G and assume that

pw(G) = ω(H) − 1.

Then H[V 0 ] is an interval graph, since the class of interval graphs is closed
under taking induced subgraph. We claim that H[V 0 ] is an interval embedding
of G0 . Let e = {x, y} be an edge of G0 . Then x and y are both vertices of V 0 and
{x, y} ∈ E. Therefore, {x, y} is an edge of H[V 0 ].
Obviously,
ω(H[V 0 ]) 6 ω(H) = pw(G) + 1. (5.23)
This proves the lemma. t
u

Lemma 5.40. The class of interval graphs is closed under edge contractions.

Proof. Let G be an interval graph. Consider a collection of intervals on the

real line such that G is the intersection graph of these intervals.
Let e = {x, y} be an edge of G. Let Ix and Iy be the two intervals that represent
x and y. Then
Ix ∩ Iy 6= ∅ (5.24)
since x and y are adjacent. Replace the two intervals Ix and Iy by one interval
Ixy , which is their union. The new collection of intervals represent an interval
graph, say G0 . We claim that G0 is the graph obtained from G by contracting
the edge {x, y}.
/ {x, y} which is adjacent to x or y is represented by an interval
Every vertex z ∈
Iz which intersects Ix or Iy . Then Iz intersects Ixy in the new interval model.
Furthermore, if z is not adjacent to x or y then Iz does not intersect Ix nor Iy .
Then it has to be to the left or to the right of both Ix and Iy , since Ix and Iy
intersect. Therefore, Iz does not intersect the interval Ixy in the model of G0 .
This proves the claim and the lemma. t
u

Following the same path as we did in Section 5.3 we have one more step
to go.
148 5 Decomposition Trees

Lemma 5.41. Let G be an interval graph. Let e = {x, y} be an edge in G and let
G0 be the interval graph obtained from G by contracting the edge {x, y}. Then

ω(G0 ) 6 ω(G). (5.25)

Proof. The proof of this lemma is the same as the proof of Lemma 5.15 on
page 131. In that lemma we showed that the clique number of a chordal
graph does not increase under edge contractions. Since interval graphs are
chordal, we are done. t
u

In the following theorem we show that the class of graphs with pathwidth
at most k is closed under taking minors.

Theorem 5.42. Let k ∈ N ∪ {0}. Let

P(k) = { G | G is a graph and pw(G) 6 k }. (5.26)

Then P(k) is minor closed.

Proof. Let G be a graph with pw(G) 6 k. Let G0 be a minor of G. When G0 is

a subgraph of G then the claim follows from Lemma 5.39.
Assume that G0 is obtained from G by contracting an edge {x, y} in G to a
single vertex xy. Let H be an interval embedding of G with ω(H) 6 k + 1. By
lemma 5.40 the graph H0 obtained from H by contracting the edge {x, y} is an
interval graph. By Lemma 5.41 ω(H0 ) 6 ω(H).
We claim that the graph H0 is an interval embedding of G0 . This follows from
the observation that any edge {a, b} in G0 is also an edge in H0 .
This proves the theorem. t
u

Remark 5.43. By Theorem 5.5 on page 125, for any k ∈ N ∪ {0}, the class
of graphs with pathwidth at most k is characterized by a finite collection
of forbidden minors. For k = 1 this set consists of two graphs, the triangle
and the smallest tree which has an asteroidal triple (see Figure 5.6). The
obstruction set for graphs with pathwidth two is also known. It contains 110
graphs.14

14
N. Kinnersley and M. Langston, Obstruction set isolation for the gate matrix layout
problem, Discrete Applied Mathematics 54 (1994), pp. 169–213.
 5.4 Pathwidth 149

u
u
J u
J
u Ju u
@
u @u
@
u @u

Fig. 5.6. The obstruction set for pathwidth one.

Remark 5.44. By Theorem 5.6 on page 126 the parameterized problem

(pw(G), k)

which asks if the pathwidth of a graph G is at most k, is fixed-parameter

tractable. For each k ∈ N ∪ {0} there exists an O(n3 ) algorithm which checks
if the pathwidth of a graph G is at most k. This algorithm is not constructive,
since the graph minor theory does not provide the obstruction set. However,
for each k ∈ N ∪ {0} there is an explicit linear-time algorithm which checks if
a graph G has pathwidth at most k.15

Remark 5.45. The pathwidth problem asks for an interval embedding of a

graph G with minimal clique number. The pathwidth of a tree with n ver-
tices can be as large as log n. Scheffler showed that there is a linear-time
algorithm which computes the pathwidth of a tree.16 The pathwidth problem
is also polynomial for outerplanar graphs and permutation graphs (which
includes cographs). The pathwidth problem remains NP-complete for graph
classes such as chordal graphs, planar graphs, bipartite distance-hereditary
graphs and cocomparability graphs.

We end this section on pathwidth with an interesting observation. Recall

that a chordal embedding of a graph G adds some edges to G such that the re-
sult is a chordal graph. When the deletion of any new edge creates a chordless
cycle of length at least four, then the embedding is minimal.
15
T. Kloks, Treewidth - Computations and Approximations, Springer-Verlag, Lecture
Notes in Computer Science 842, 1994.
16
P. Scheffler, A linear algorithm for the pathwidth of trees. In (R. Bodendieck and
R. Henn eds.) Topics in Combinatorics and Graph Theory, Springer Physica-Verlag
(1990), pp. 613–620.
150 5 Decomposition Trees

Definition 5.46. A chordal embedding H = (V, E0 ) of a graph G = (V, E) is

minimal if for for any edge e ∈ E0 \ E the graph H0 = (V, E0 \ {e}) is not chordal.

Theorem 5.47 (Kloks, Möhring). Let G be an AT-free graph. Then every mini-
mal chordal embedding of G is an interval graph. Consequently,

pw(G) = tw(G). (5.27)

Remark 5.48. There exists an elegant O(n3 ) algorithm which computes the
treewidth of circle graphs.17 When the graph is a permutation graph it is
AT-free and then this algorithm computes the pathwidth. However, for per-
mutation graphs there is also a linear-time algorithm which computes the
pathwidth.

5.5 Rankwidth

Recall the definition of a decomposition tree (T , f) of a distance-hereditary

graph G = (V, E) as described in Section 3.4.1 on page 60. Thus T is a rooted
binary tree and f is a bijection from the vertices of G to the leaves of T .

For each edge {p, c} in T where c is the child of p, let We be the set of
vertices of G that are mapped to leaves in the subtree of c. The twinset Qe is
the subset of We that have neighbors in V \ We .

By definition of the decomposition tree, all vertices of Qe have the same

neighbors in V \ We .

Consider the adjacency matrix A of G. This is an n×n matrix with diagonal

elements Ai,i = 0 for all i. For i 6= j, Ai,j = 1 if the vertices xi and xj are
adjacent in G and Ai,j = 0 if xi and xj are not adjacent in G.

Let e = {p, c} be an edge in T .

The cutmatrix Ce is the submatrix of A of which the rows are

the vertices of We and the columns are V \ We .

17
T. Kloks, Treewidth of circle graphs, International Journal of Foundations of Com-
puter Science 7 (1996), pp. 111–120.
 5.5 Rankwidth 151

Since all vertices of Qe have the same neighbors in V \ We , the cutmatrix

Ce has a shape  
J0
Ce = (5.28)
00
where J is a matrix with all ones. The first set of rows of Ce correspond with
the vertices of Qe and the first set of columns correspond with the neighbors
of vertices in Qe in V \ We .

When Qe = ∅, this matrix becomes the zero matrix. If Qe = We then the

shape of the matrix becomes

J0 .
When the vertices of Qe are adjacent to all vertices of V \ We then the matrix
becomes  
J
.
0
So in general, the cutmatrix is a submatrix of the matrix in (5.28).

The binary field GF[2] has two elements, 0 and 1. In this field we can add
elements and multiply elements. The addition in the field GF[2] is defined by
the following rules.

0+0=0 and 0+1=1+0=1 and 1 + 1 = 0.

The multiplication in GF[2] is defined by

0·0=0 and 0·1=1·0=0 and 1 · 1 = 1.

Column vectors with entries in GF[2] can be added by using the addition
rules of GF[2] entrywise. The multiplication of a vector with a scalar α ∈ {0, 1}
is done entrywise, with the rules for multiplication of GF[2].

A set of column vectors a1 , . . . , at is linearly independent if

α1 · a1 + · · · + αt · at = 0 (for some α1 , . . . , αt ∈ {0, 1}) ⇒

α1 = · · · = αt = 0 (5.29)

where 0 is the all-zero vector.

The rank of a matrix over GF[2] is the maximal number of linearly indepen-
dent columns.
152 5 Decomposition Trees

The rank of a matrix is computed using the well-known Gauss elimination

method.

Definition 5.49. A graph G has rankwidth k if there exists a decomposition tree

(T , f) such that every cutmatrix has rank over GF[2] at most k.

Here, a decomposition tree (T , f) is defined as above, so T is a rooted binary

tree and f is a bijection from the vertices of G to the leaves of T .

We denote the class of graphs with rankwidth k by R(k).

Since we are only interested in matrices that have small rank, the follow-
ing observation is useful. It shows that, we can avoid the Gaussian elimination
and instead just look at the number of different rows or columns.

Lemma 5.50. Let A be a matrix and let k be its rank over GF[2]. Then A has at
most 2k different columns.

Proof. Since the rank over GF[2] of A is k there is a basis {a1 , . . . , ak }. Then
every column c of A can be written as a linear combination,

c = α1 · a1 + · · · + αk · ak .

The scalars αi ∈ {0, 1}, for all i ∈ {1, . . . , k}. There are at most 2k different
linear combinations and so there are at most 2k different columns in A. t
u

Lemma 5.51. Let G be a graph with rankwidth k. Then the rankwidth of its
complement Ḡ is at most k + 1.

Proof. Let (T , f) be a decomposition tree for G such that the rank over GF[2]
of every edge in T is at most k.
For Ḡ we use the same decomposition tree. We write J for the all-one matrix.
Each cutmatrix Ce changes to J + Ce , which switches the zeroes and ones in
Ce into ones and zeroes.
Since the rank over GF[2] of Ce is k, there is a basis {a1 , . . . , ak } for the
column space of Ce . Then every column of J + Ce can be written as

j + c = j + α1 · a1 + · · · + αk · ak ,

where j is the all-one vector. This shows that the dimension of the columns
space of J + Ce is at most k + 1. Here we use the linear algebra property
that {j, a1 , . . . , ak } contains a basis for the columns of J + Ce , with a minimal
number of elements. t
u
 5.5 Rankwidth 153

Let’s look at an example.

Lemma 5.52. A graph is a cograph if and only if it has a decomposition tree

such that every cutmatrix or its transpose has a shape


J0

or a submatrix of that. Here J is the all-ones matrix. Consequently, cographs

have rankwidth one.

Proof. By Theorem 3.34, a graph G = (V, E) is a cograph if and only if it has a

cotree (see Section 3.3.1). Any binary tree with at least two vertices has two
leaves that have the same parent. Let x and y be two vertices of G that are
mapped to sibling leaves. We claim that x and y are twins.
To see that, notice that x is adjacent to z 6= x if and only if their common
ancestor is labeled with an ⊗-operator. Let z ∈ V \ {x, y}. Then the common
ancestor of x and z is the same as the common ancestor of y and z. This proves
the claim.
Actually, a graph is a cograph if and only if every induced subgraph with at
least two vertices has a twin. Namely, if H is a cograph, then the operation
which creates a twin of some vertex x in H does not create an induced P4 .
Thus the class of cographs is closed under creating twins.
We prove the lemma by induction on the number of vertices. Consider a cotree
(T , f) for the cograph G. Let x and y be twins, mapped to sibling leaves of T .
Remove the vertex x and let G0 = G − x.
A decomposition tree (T 0 , f0 ) for G0 is obtained from the decomposition tree
(T , F) for G by removing the leaf that is mapped to x, and by contracting the
edge in T that connects y to its parent. The ⊗ or ⊕ label of the parent of x
and y in T disappears.
By induction, the cutmatrix of every edge in T 0 has the shape as claimed in
the lemma. Now consider the edges of T . For every edge e in T which is not
incident with x or y, the cutmatrix Ce is obtained from the cutmatrix C0e in T 0
by making a copy of the row or column that represents the vertex y since x
and y have the same neighbors.
Consider an edge e of T which is incident with a leaf. Let a be the vertex that
is mapped to that leaf. Then the cutmatrix of e is


1···1 0···0 ,

where the single row represents a. The first set of columns, those with a 1,
are the neighbors of a and the final set of columns, those with a 0, are the
nonneighbors of a. Thus in any decomposition tree, the cutmatrix of an edge
which is incident with a leaf, has the desired shape.
This proves the lemma. t
u
154 5 Decomposition Trees

Remark 5.53. Notice that the class R(k) is closed under creating twins.

In the same manner as in the proof of Lemma 5.52 one can show that
every cutmatrix in the decomposition tree of a distance-hereditary graph, has
a shape  
J0
.
00
We ask you to prove that in Exercise 5.21.

Many NP-complete problems can be solved in polynomial time for graphs

of R(k). The reason is this. Consider a decomposition tree (T , f) for a graph
in R(k). Let e = {p, c} be an edge in T , where p is the parent of c. Let We
be the set of vertices that are mapped to leaves in the subtree rooted at c. By
Lemma 5.50 the vertices of We have at most 2k different neighborhoods in
V \ We . In the next section we discuss a class of problems that can be solved
in polynomial time for graph in R(k).

Algorithms for graphs in R(k) use the dynamic programming strategy on

the decomposition tree. Luckily, this decomposition tree can be obtained in
O(n3 ) time.18 We omit the description of this algorithm.

We have seen in Theorem 5.16 that the class T(k) of graphs with treewidth
at most k is closed under taking minors. As a consequence, this class of graphs
is characterized by a finite obstruction set.

Unfortunately, this is not true for the class R(k) of graphs with rankwidth
at most k. For example, consider a clique with 5 vertices. This graph has
rankwidth one (it is a cograph). Now remove edges such that a 5-cycle re-
mains. The 5-cycle is not distance-hereditary; its rankwidth is two. Thus the
5-cycle is a minor of the 5-clique but the rankwidth of the 5-cycle is bigger
than the rankwidth of the 5-clique. Actually, this example shows that the class
R(k) is not even closed under taking subgraphs.

For the class R(k) we define a different ordering, based on

Local Complementation.

Definition 5.54. Let G = (V, E) be a graph. Let x ∈ V. The local complementa-

tion at x is the operation which replaces all edges in N(x) by nonedges and all
nonedges in N(x) by edges.
18
P. Hliněný and S. Oum, Finding branch-decompositions and rank-decompositions,
SIAM Journal on Computing 38 (2008), pp. 1012–1032.
 5.6 Monadic second-order logic 155

Definition 5.55. A graph H is a vertex-minor of a graph G if H can be obtained

by a sequence of operations, each of which is either
(a) a deletion of a vertex, or
(b) a local complementation.

In Exercise 5.22 we ask you to prove that R(k) is closed under taking
vertex-minors.

Oum and Seymour proved the following theorem.

Theorem 5.56. Let k ∈ N ∪ {0} and let

G1 , G2 , . . . (5.30)

be an infinite sequence of graphs in R(k). There exist indices i < j such that Gi
is a vertex-minor of Gj .

As a consequence of this theorem the graphs in R(k) are characterized by a

finite collection of forbidden vertex-minors. To prove this, Oum and Seymour
showed that, for k > 1, each graph in the obstruction set has at most

6k+1 − 1
5
vertices.

This finite obstruction set for R(k) leads to a polynomial recognition algo-
rithm. However, this is non-constructive since the obstruction set is unknown.
Furthermore, the timebound for this algorithm is much worse than the con-
structive O(n3 ) algorithm by Hliněný and Oum, mentioned above. When H
is a fixed graph, then checking if a graph G contains H as a vertex-minor is
quite complicated.

5.6 Monadic second-order logic

The most natural way to express and classify graph-theoretic problems is by
means of logic.

In monadic second order logic a finite sentence is a formula that uses

quantifiers ∀ and ∃. The quantification is over vertices, edges, and subsets of
vertices and edges. Relational symbols are ¬, ∈, =, and, or, ⊆, ∪, ∩, and ⇒.
Some of these are superfluous.

Although the minimization or maximization of the cardinality of a subset

is not part of the logic, one usually includes them.
156 5 Decomposition Trees

As an example we show how the feedback vertex set can be formulated

in monadic second-order logic. Obviously, we can formulate that V \ F is a
forest by a sentence which expresses that every subset W of V \ F has a vertex
of degree at most one but the following method to formulate that a graph is
a forest is more informative.

First we show that the property that a graph is connected can be formu-
lated in this logic. A graph is disconnected if the vertex set V has a partition
{V1 , V2 } such that there is no edge between a vertex in V1 and a vertex in V2 .
Note that this property can be formulated in monadic second-order logic.

Next we show that we can formulate the property that a graph has no
induced cycle of length more than three in this logic. A graph has an induced
cycle of length at least four if there exist three vertices w, w1 , and w2 such
that w is adjacent to w1 and w2 , and w1 and w2 are not adjacent, and such
that the following holds. Let

V 0 = (V \ N[w]) ∪ {w1 , w2 }.

Obviously, there exists an induced cycle containing {w, w1 , w2 } if and only if

w1 and w2 are contained in a component of G[V 0 ].

Checking for a triangle is easy and so, one can formulate the property that
a graph is a forest in monadic second-order logic. Finally, the fact that a subset
F is a feedback vertex set can be formulated by stating that V \ F induces a
forest.

Another easy example is this. Fix some graph H. Then one can formulate in
monadic second-order logic if a graph G contains H as an induced subgraph.

One more example. It is only a little bit more complicated to show that for
every graph H there is a monadic second-order formula that expresses that a
graph G contains H as a minor. We leave it as an exercise.19

Let k ∈ N ∪ {0} and let Ω(k) be the finite obstruction set for T(k). Then
one can formulate the question if the treewidth of a graph G is at most k
by a finite monadic second-order formula. Namely, write down the monadic
second order formula which checks if some H ∈ Ω(k) is a minor of G.

Courcelle popularized monadic second-order logic by proving that any

graph-theoretic problem that can be formulated in monadic second-order
logic can be solved in linear time for graphs of bounded treewidth.20
19
Hint: Use the alternative formulation of a minor, which appears after Defini-
tion 5.3.
20
B. Courcelle, The monadic second-order logic or graphs. I. Recognizable sets of
finite graphs, Information and Computation 85 (1990), pp. 12–75.
 5.6 Monadic second-order logic 157

Theorem 5.57. Let k ∈ N∪{0}. Any problem that can be formulated in monadic
second-order logic can be solved in linear time for graphs in T(k).

A restricted form of this logic is where one does not allow quantification
over subsets of edges.

The C2 MS-logic is a restricted monadic second-order logic, where one

does not allow quantification over subsets of edges but where one can use a
test whether the cardinality of a subset of vertices is even or odd. Using this
logic one can formulate, for example, whether a fixed graph H is a vertex-
minor of a graph G.21

Remark 5.58. Notice the difference. One can formulate that a graph is Hamil-
tonian by expressing this as the existence of a suitable subset of the edges.
However, this formulation is not valid in C2 MS-logic.

The class of graphs that have rankwidth at most k is much larger than the
class of graphs with treewidth at most k. (There exists a function

f:N→N

such that if a graph has treewidth at most k then its rankwidth is at most
f(k). The converse is of course not true; any clique has rankwidth one, but its
treewidth is the number of vertices minus one.) The set of problems that can
be solved in polynomial time for graphs of bounded rankwidth is consequently
a bit smaller. For example, the Hamiltonian cycle problem can be formulated
in monadic second-order logic but not in C2 MS-logic.

Theorem 5.59. Let k ∈ N ∪ {0}. Any problem that can be formulated in C2 MS-
logic can be solved in O(n3 ) time for graphs in R(k). When a decomposition tree
for the graph is a part of the input, then these algorithms run in linear time.

Remark 5.60. The Hamiltonian cycle problem can be solved in polynomial

time for graphs of bounded rankwidth. However, this algorithm does not run
in linear time (not even when a rank-decomposition tree is a part of the in-
put).22 23
21
B. Courcelle and S. Oum, Vertex-minors, monadic second-order logic, and a conjec-
ture by Seese, Journal of Combinatorial Theory, Series B 97 (2007), pp. 91–126.
22
E. Wanke, k-NLC graphs and polynomial algorithms, Discrete Applied Mathematics
54 (1994), pp. 251–266.
23
F. Fomin, P. Golovach, D. Lokshtanov and S. Saurabh, Intractability of clique-width
parameterizations, SIAM Journal on Computing 39 (2010), pp. 1941–1956.
158 5 Decomposition Trees

5.7 Problems
5.1. Consider an infinite sequence of trees

T1 , T 2 , . . . (5.31)

Let  denote the induced subgraph relation. Is it true that for any sequence
of trees, as in (5.31) there always exist integers i < j such that Ti  Tj ?
Hint: Consider a sequence of paths P1 , P2 , . . . In each path Pi , say with ends
ai and bi , add two leaves, a∗i and b∗i . Make a∗i adjacent to N(ai ) and make
b∗i adjacent to N(bi ) (in other words; create a twin of each end).

5.2. Prove the alternative definition of a minor, which appears after Defini-
tion 5.3.

5.3. Prove that, for two trees T1 and T2 , if T1 m T2 then there exists a se-
quence of edge contractions in T2 that produces T1 .

5.4. Check that the class of all planar graphs is closed under taking minors.

5.5. Let T be the class of all forests.

(a) Prove that G ∈ T if and only if G has no triangle as a minor.
(b) Prove that you can test in O(n3 ) time if a graph G is a forest.
(c) Can you do better? (This is a joke.)

5.6. A graph is outerplanar if it has a plane embedding such that all vertices
are on the outerface. Let O be the class of outerplanar graphs.
(i) Prove that O is closed under taking minors.
(ii) Prove that a graph is outerplanar if and only if it has no K4 nor K2,3 as a
minor. Thus the obstruction set for O is

Ω = { K4 , K2,3 }.

5.7. Consider graphs of treewidth two.

(1) Prove that a graph has treewidth at most two if and only if it has no K4 as
a minor.
(2) Use Problem 5.6 to show that every outerplanar graph has treewidth at
most two.
(3) Give an example of a graph which has treewidth two but which is not
outerplanar.

5.8. Consider the class of graphs that you can draw on the surface of a torus
(that is a donut) without crossing lines. The graphs in this class are called
toroidal. Show that this class of graphs is minor closed.
The obstruction set is finite, but it seems that it contains at least 16000 ele-
ments. Can you think of one element?
 5.7 Problems 159

Hint: It is not difficult to show that you can draw K5 on a torus. Also, K6 and
K7 are toroidal. In general, if you can draw a graph G in the plane with at
most one crossing then G is toroidal. Toroidal graphs have chromatic number
at most 7; thus K8 is not toroidal.
5.9. Show that K3,3 is a minor of the Petersen graph.
5.10. Let k ∈ N ∪ {0}. Let C(k) be the class of graphs G that have at most k
vertex-disjoint cycles.
(i) Show that C(k) is closed under taking minors.
(ii) What is the obstruction set Ω(k)?
Hint: Consider the graph Q which is the union of k + 1 triangles. Prove
that Ω(k) = {Q}.
(iii) Let G(k) be the class of graphs that have a feedback vertex set with at
most k vertices. Show that G(k) ⊆ C(k). Can you think of a graph that is
in C(k) but not in G(k)?
Hint: Consider the 5-wheel, that is, a 5-cycle plus one vertex adjacent
to all vertices in the cycle. How many vertex-disjoint cycles are there in
the 5-wheel? What is the minimal cardinality of a feedback vertex set?
5.11. Let k ∈ N ∪ {0}. Let K(k) be the class of graphs that have a vertex cover
with at most k vertices.
(a) Show that K(k) is closed under taking minors.
(b) Consider the graph H which is the union of k + 1 edges. Then H ∈/ K(k).
(c) Can you think of another graph than H which is in the obstruction set of
K(k)?
5.12. Let k ∈ N ∪ {0}. Let D(k) be the class of graphs G = (V, E) for which
there is a subset D of vertices with |D| 6 k such that every vertex of G − D
has degree at most two in G − D.
(1) Show that D(k) is minor closed.
(2) Show that there exists an O(n3 ) algorithm to check if G ∈ D(k).
5.13. Let T be a tree and let

{ Tx | x ∈ V } (5.32)

be a collection of subtrees of T . Define a graph G = (V, E) by

{x, y} ∈ E if and only if |Tx ∩ Ty | > 2. (5.33)

I. What can you say about G?

Hint: Show that any graph is the edge-intersection graph of a collection
of subtrees of a star (a tree of diameter two).
II. Let T be a star and fix the degree of T by some k ∈ N. Let G(k) be the
class of graphs that are edge-intersection graphs of T . What can you say
about G(k)?
160 5 Decomposition Trees

III. Is the class G(k) minor closed?24

5.14. Let G be a graph. Prove that

ω(G) 6 tw(G) + 1 and χ(G) 6 tw(G) + 1.

5.15. Let G be a k-tree and consider a vertex coloring of G with k + 1 colors

such that no two adjacent vertices have the same color. Let 1 6 p 6 k + 1 and
let Cp be any subset of p colors. Let Gp be the subgraph of G induced by the
vertices that have colors in Cp . Prove that Gp is a (p − 1)-tree.
Hint: Use a perfect elimination ordering for G and prove the claim by induc-
tion.

5.16. Let G be a chordal graph. Prove that if G is not a clique then it has two
simplicial vertices that are not adjacent.
Hint: Use Lemma 5.24.

5.17. Let G be a k-tree. Prove that every minimal separator in G is the inter-
section of two maximal cliques of cardinality k + 1.

5.18. Let G be a graph and let k = tw(G). Prove that every induced subgraph
of G has a vertex with at most k neighbors.

5.19. Prove that the number of edges in a k-tree is

   
k+1 k+1
+ (n − k − 1)k = nk − .
2 2

5.20. Let T be a tree. prove that T is an interval graph if and only if it is a

caterpillar.

5.21. Let (T , f) be a decomposition tree for a distance-hereditary graph. Prove

that the cutmatrix of every edge in T has a shape
 
J0
.
00

Hint: A distance-hereditary graph has an isolated vertex, or a pendant vertex,

or a twin. Use this fact to prove that for every edge e in T the vertices of We
have only two different neighborhoods in V \ Ee , namely, all vertices of the
twinset Qe have the same neighbors in V \ We and all vertices of We \ Qe
have the same neighbors (zero) in V \ We .

5.22. Let k > 1. Prove that the class R(k) of graphs of rankwidth at most k is
closed under taking vertex-minors.
24
J. Gramm, J. Guo, F. Hüffner and R. Niedermeier, Data reduction, exact, and heuris-
tic algorithms for clique cover, proceedings 8th ALENEX06, SIAM (2006), pp. 86–94.
 5.7 Problems 161

5.23. Let G = (V, E) be a graph. Let S ⊆ V be a subset of vertices. A switch

of G with respect to S is the following operation. Change all edges with one
endvertex in S and the other in V \ S into nonedges, and change all nonedges
with one endvertex in S and the other in V \ S into edges.
Let G be the class of graphs that can be switched to a cograph. In other
words, G ∈ G if there exists a set of vertices in G such that the switch of G
with respect to S changes G into a cograph.
Prove that graphs in G have rankwidth at most two.

5.24. Consider the following class G of graphs G for which there is a color-
ing of the vertices with colors black and white, such that for every induced
subgraph H of G, one of the following holds.
(i) H has one vertex (either black or white).
(ii) There exists a partition {V1 , V2 } of the vertices of H such that either
(a) every vertex x of V1 is adjacent to all vertices of V2 that have the
same color as x, or
(b) every vertex x of V1 is adjacent to all vertices of V2 that have the
opposite color of x.
(1) Prove that the graphs in G have rankwidth at most two.
(2) Prove that there exists a polynomial-time algorithm to check if a graph G
is in G.

5.25. Let k ∈ N ∪ {0}. Prove that the chromatic number problem can be solved
in linear time on graphs with treewidth at most k by showing that the problem
can be formulated in monadic second-order logic.
Hint: Use the result of Exercise 5.14, that is,

tw(G) 6 k implies that χ(G) 6 k + 1.

5.26. Prove that the domatic partition problem can be solved in linear time
on graphs of bounded treewidth.
Hint: First prove that the domatic number of a graph G ∈ T(k) is bounded
by k + 1.
References

1. R. Bellman, Dynamic programming, Princeton University Press, Landmarks in

Mathematics, 2012.
2. R. Niedermeier, Invitation to fixed-parameter algorithms, Oxford Lecture Series in
Mathematics 31, 2006.
3. R. Diestel, Graph theory, Springer-Verlag, Graduate Text in Mathematics 173,
2010.
4. M. Golumbic, Algorithmic graph theory and perfect graphs, Elsevier, Annals of Dis-
crete Mathematics 57, 2004.
5. T. Kloks, Treewidth – computations and approximations, Springer-Verlag, Lecture
Notes in Computer Science 842, 1994.
6. J. Edmonds, A glimpse of heaven. In (J. Lenstra, A. Rinnooy Kan and A. Schrijver,
eds.) History of Mathematical Programming – A collection of personal reminis-
cences, CWI Amsterdam and North-Holland, Amsterdam 1991, pp. 32–54.
7. P. Seymour, How the proof of the strong perfect graph conjecture was found,
Gazette des Mathématiciens 109 (2006), pp. 69–83.
8. F. Fomin and D. Kratsch, Exact exponential algorithms, Springer-Verlag, 2010.
Index

(k, `)-coloring, 114 γ(G), 13, 18, 57

2K2 , 65 κ(G), 4, 37
3-coloring, 12 N, 139
3-edge-coloring, 12 D(k), 159
C2 MS-logic, 155 K(k), 159
E(G), 1, 2 R(k), 152
G-decomposition, 48 T(k), 131, 154
G − F, 2 ν(G), 38
G − N[x], 7 ω(G), 8, 37, 56, 57
G − S, 2, 7 ⊕-operator, 55, 61
G − e, 2 ⊗-operator, 55, 61
G − x, 2, 7 m -ordering, 124
G = (V, E), 5 τ(G), 84
GF[2], 151 ϑ(G), 40
G[S], 2, 11 ζ(G), 93
K5 , 159 a, b-path, 2, 137
K3,3 , 159 cp(G), 142
L(G), 2, 37 h-wall, 143
N(C), 1, 59 k-cover, 23
N(x), 2, 7 k-tree, 135
N[C], 1 k-tree embedding, 136, 138
N[x], 2, 7 k-tuple, 23
O∗ -notation, 5 t-cycle, 101
P3 , 54 tw(G), 129
P3 -free graph, 84 x, y-separator, 65
P4 , 53 1-1 correspondence, 68
V(G), 1, 2 3-hitting set, 118
X-partition, 14 3-hitting set problem, 118
Ḡ, 2, 8 4-cycle, 63
∆, 38 5-wheel, 159
Θ(G), 40
α(G), 5, 21, 37, 57 Ackermann, W., 118
χ(G), 10, 37 acyclic orientation, 82
166 Index

addition rule, 151 Chandran, L., 100

adjacency, 1 charateristic equation, 32
adjacency matrix, 5 Chen, J., 98, 100
adjacent, 1, 32 child, 61
algorithm, 8 chord, 2, 3
all-ones matrix, 151 chord in a circle, 87
all-zero vector, 151 chordal embedding, 101, 140
Alon, N., 117 chordal graph, 64, 101
antichain, 43 chordal graphs, 64
antihole, 39 chordless cycle, 3, 73
antisymmetric, 41 chordless path, 2, 57, 66
arc, 79 chromatic index, 3
Ariyoshi, H., 6 chromatic number, 3, 10, 12
Arnborg, S., 137 Chudnovsky, M., 40, 114
asteroidal triple, 74 circle diagram, 87
AT-free, 75 circle graph, 87
AT-free graph, 75, 150 class of forests, 158
class of graphs, 35
backtrack, 94 class of planar graphs, 35
Baker, K., 81 class one graph, 38
Bandelt, H., 60 clique, 3, 8, 101
basic classes of perfect graphs, 39 clique cover, 4, 37
Bellman, R., 15, 163 with overlapping cliques, 4
Berge, C., 38 clique cover number, 37, 40
big deal, 90 clique cover of edges, 159
bijection, 55 clique cover problem, 37
bijective map, 150 clique number, 56
binary entropy function, 20 clique problem, 3, 37, 91
binary field, 151 clique separator, 66
binary rank, 152 clique tree, 68
binary tree, 7 closed neighborhood, 1, 2, 7
bipartite graph, 3, 12, 21 closed neighborhood of set, 1
bitvector, 115 cograph, 53, 55
Björklund, A., 15, 21 color class, 21, 48, 114
Blair, J., 64 color coding, 116
Bodlaender, H., 25, 133 coloring of a bipartite graph, 21
Boland, J., 76 coloring of a graph, 3
boolean variable, 141 coloring problem, 3, 10, 12, 35, 37, 91
bottom line, 82 column basis, 152
bounded search tree technique, 94 comparability, 41
branch, 8 comparability graph, 40, 79, 81, 83
branch-decomposition, 154 comparable, 41
Bron, C., 10 complement, 8, 37
bucket sort, 138 complement of a graph, 2
Byskov, J. M., 32 complete bipartite graph, 115
complete graph, 82
Catalan number, 102 complete multipartite graph, 115
Catalan, E. C., 102 component, 9
caterpillar, 145 component of a graph, 3
 Index 167

connected, 53 dominating set in cographs, 57

connected component, 9 dominating set on bipartite graphs, 21
connected graph, 3, 53 dominating set problem, 13, 17, 91
consecutive clique arrangement, 74 domination problem, 13, 17
constant time, 6 domino, 58
constraint satisfaction, 12 donut, 158
containment graph, 83 Downey, R., 92
contraction of edge, 124 Dushnik, B., 46, 82
Corneil, D., 41, 55, 133, 137 dynamic programming, 15, 139
Cornuéjols, G., 40
cotree, 55 edge, 1, 2
Courcelle, B., 156, 157 edge bipartization, 119
crash course, 64 edge coloring problem, 3
crossing linesegments, 80 edge contraction, 123, 124
cubic algorithm, 4 edge set, 2
cutmatrix, 151 edge-intersection graph, 159
cycle, 3, 9 Edmonds, J., 21, 97, 163
chord of a cycle, 3 elementary school, 118
edge of a cycle, 3 elementary wall, 143
cycle packing, 142 elimination process, 66
empty graph, 1, 2, 8
data structure, 140 endpoints (of a path), 2
decomposition tree, 55, 121 endpoints (of an edge), 2
decomposition tree for DH-graph, 60 entropy function, 20, 28
degree, 1, 9 Eppstein, D., 12
degree of a vertex, 2 equivalence class, 48
Dehne, F., 113 Euclidean geometry, 83
deletion of edge, 124 even cycle, 2
deletion of vertex, 124 Even, S., 80
depth-first search, 36 exhaustive search, 35
DH-graph, 60 exponential algorithm, 5, 10
Diestel, R., 163 exponential space, 10
Dilworth’s theorem, 44 exponential time hypothesis (ETH), 100
Dilworth, R., 43 extension, 18
dimension of a poset, 46, 81
Dirac, G., 64 fast matrix multiplication, 42
directed edge, 79 Feder, T., 114
directed graph, 81 feedback vertex set, 62, 108, 127
directed triangle, 82 feedback vertex set in DH-graphs, 62
disconnected, 55, 156 feedback vertex set problem, 63, 108
disjoint cycles, 142 Fellows, M., 92, 113
disjoint separators, 65 Fernau, H., 98
disjoint sets, 65 Fibonacci number, 32
distance, 57 Fibonacci, L., 32
distance-hereditary graph, 57 finite basis theorem, 125
domatic number, 161 finite class, 35
domatic partition, 13 finite graph, 1
domatic partition problem, 13 Fishburn, P., 81
dominating set, 13, 17, 57 fixed-parameter algorithm, 89
168 Index

fixed-parameter tractable, 91 hole, 39, 53, 58

flower, 97 homogeneous coloring, 114
Fomin, F., 13, 157, 163 house, 58
Ford, L., 108 Howorka, E., 57
Ford–Fulkerson algorithm, 108 Hunt, J. W., 83
forest, 3, 62 Husfeldt, T., 15, 21
four color theorem, 11
Fredman, M., 118 Ide, M., 6
Fulkerson, D., 108 Impagliazzo, R., 100
full binary tree, 94 implication class, 48
function intersection-model, 46 inclusion - exclusion, 15, 21
inclusion – exclusion formula, 22
Gallai, T., 44 incomparable, 41
Gasparian, G., 37 independence number, 40
Gauss elimination method, 152 independent set, 3, 5
Gauss, J. C. F., 152 independent set in bipartite graph, 21
gem, 58 independent set in linegraph, 38
Gilmore, P., 79 independent set problem, 3, 5, 6, 92
Gioan, E., 62 induced cycle, 64
Golovach, P., 157 induced path, 66
Golumbic, M., 42, 45, 47, 55, 84, 163 induced subgraph, 2, 7
Grötschel, M., 40 induced subgraph relation, 122
Gramm, J., 108, 119, 160 infinite class, 35
Grandoni, F., 100 infinite sequence, 122
graph, 1, 5, 8 infinite sequence of cycles, 123
graph class, 35 infinite sequence of graphs, 122
graph coloring, 3, 86 infinite sequence of paths, 123
graph minor, 121, 124 infinite sequence of trees, 158
graph minor theorem, 124 infinite subsequence, 122
graph minor theory, 121 internal node, 56
Guo, J., 108, 119, 160 internal vertex, 55, 66
intersection graph, 70
Hüffner, F., 108, 119, 160 interval, 73
Habib, M., 41, 82 interval containment graph, 82
Halin, R., 73, 76 interval embedding, 145
Hamilton, W. R., 33 interval graph, 72
Hamiltonian cycle, 33 inverse Ackermann function, 118
Hamiltonian path, 33 isolated vertex, 3, 8, 58, 85, 87
Hamiltonian path problem, 117 isomorphic, 122
Hammer, P., 60 iterative compression, 108
Hamming distance, 115
Hamming, R., 115 Jian, T., 10
Heggernes, P., 115 join, 56, 63
Held, M., 15 join-operation, 61
Hell, P., 114 joke, 158
Helly property, 74
hereditary class, 35 Köhler, E., 75
Hliněný, P., 154 König-Egerváry theorem, 38, 97
Hoffman, A., 79 Kézdy, S., 114
 Index 169

Kőnig, D., 37, 38 Lueker, G., 66

Kanj, I., 98, 100
Karp, R., 15 Möhring, R., 150
Kawarabayashi, K., 127 Maffray, F., 60
Kerbosch, J., 10 Mahajan, M., 120
kernel, 95, 96 Marx, D., 100
kernelization, 95 matching, 3, 21, 38, 96
Kobayashi, Y., 127 matrix partition, 114
Koivisto, M., 15, 21 max flow – min cut theorem, 106
Komlós, J., 118 maximal clique, 3, 69
Kooten Niekerk, M. van, 25 maximal independent set, 3, 6
Kratsch, D., 13, 115, 163 of minimal cardinality, 57
Kruskal, J., 124 maximum clique, 3, 71
Kuratowski theorem, 126 maximum clique in cographs, 57
Kuratowski, K., 126 maximum cut, 119
maximum cut problem, 119
label, 55 maximum flow problem, 108
Lampis, M., 98 maximum independent set, 3, 5, 71
Langston, M., 113 maximum independent set in cographs,
Lawler, E. L., 12 57
leaf, 3, 55 maximum independent set problem, 10,
left child, 8 37
left endpoint, 82 maximum matching, 21, 38, 96
left subtree, 56 maximum matching problem, 97
Lekkerkerker, C., 76 membership, 35
Lempel, A., 80 Menger’s theorem, 106
length of cycle, 9 Menger, K., 106
length of path, 9 Miller, E., 46, 82
Lenstra, J., 163 Ming, A., 102
Lerchs, H., 55 minimal chordal embedding, 150
Liedloff, M., 21 minimal embedding, 103
line, 1, 2 minimal separator, 65, 137
line of a tree, 3, 68 minimum extension, 18
linear algorithm, 4 minimum fill-in, 101
linear order, 41 minimum fill-in problem, 101
linear-time algorithm, 9 minor, 121, 124
linearly independent columns, 151 minor ordering, 124
linegraph, 2, 37 minor test, 126
linegraph of bipartite graph, 37 minor, alternative definition, 124
linesegment, 80 minor-closed class of graphs, 126
linked list, 5 modular decomposition, 82
listing algorithm, 6 monadic second-order logic, 155, 156
Liu, X., 40 Moon, J. W., 6
local complement, 154 Moser, L., 6
logic, 155 Mulder, H., 60
Lokshtanov, D., 115, 157 multiplication rule, 151
loop, 41
Lovász number, 40 natural number, 30, 122
Lovász, L., 37, 40 neighbor, 1, 2, 5
170 Index

neighborhood, 1, 2, 7 perfect graph, 36, 114

neighborhood of set, 1 perfect graph theorem, 37
new ballgame, 37 Perković, L., 132
Niedermeier, R., 108, 119, 160 Perl, Y., 55
Niermeier, R., 163 permutation diagram, 80
non-chordal, 101 permutation graph, 80, 114
nondecreasing, 122 Petersen graph, 159
nonneighbor, 2, 54 Peyton, B., 64
NP-complete, 13 piece of cake, 131
NP-completeness, 4, 5 planar graph, 11
Pnueli, A., 80
obstruction set, 126 point, 1, 2
obstruction set of vertex-minors, 155 point of a tree, 3, 68
odd antihole, 39 polynomial algorithm, 4
odd cycle, 2, 38 polynomial delay, 6
odd cycle transversal, 104 polynomial factor, 19
odd cycle transversal problem, 104 polynomial space, 10
odd hole, 39 poset, 41, 46, 81
ordering by induced subgraph relation, poset dimension, 46
122 Prieto, E., 120
orientation, 40, 48, 79 problem kernel, 95
orientation of a graph, 48 proportional, 11
Oum, S., 154, 157 Proskurowski, A., 133, 137
outerface, 158 pruned search tree, 10
outerplanar graph, 158
quadratic algorithm, 4
parameterized clique problem, 89
quantification, 155
parameterized complexity, 92
quantifier, 155
parameterized decomposition tree, 121
quasi-order, 122
parameterized feedback vertex set
quick-and-dirty, 83
problem, 127
parameterized graph minor problem,
126 Raman, V., 115, 120
parameterized problem, 89 rank of a matrix, 152
parameterized reduction, 94 rank-decomposition, 154
parent, 61 rankwidth, 150, 152
partial order, 41, 83, 122 real line, 73
dimension two, 81 realizer, 46
partially ordered set, 41, 46, 82, 83 recognition algorithm, 21, 47
partition of a set, 13 recognition of planar graph, 36
path, 2, 9 recognition problem, 35
chord of a path, 2 recurrence relation, 9
edge of a path, 2 reduced graph, 9
pathwidth, 145 reduction rule, 109
Paturi, R., 100 Reed, B., 104, 127, 132
Paul, C., 41, 62, 82 reflexive, 41
paw, 124 reflexive graph, 41
pendant vertex, 8, 58, 85, 87 relation, 122
perfect elimination ordering, 67, 134 relational symbol, 155
 Index 171

restricted monadic second-order logic, Stirling, J., 20

156 Strassen’s algorithm, 42
right child, 8 Strassen, V., 42
right endpoint, 82 strong perfect graph theorem, 39
right subtree, 56 subdivision of edge, 143
rigid circuit graph, 64 subgraph, 2, 124
Rinnooy Kan, A., 163 subgraph relation, 123
Roberts, F., 81 subset sum problem, 30
Robertson, N., 39, 124, 143 subtree, 56
Robson, J. M., 10 switch, 161
Rooij, J. van, 25 symmetric closure, 48
rooted binary tree, 7 Szemerédi, E., 118
Rosamund, F., 113 Szymanski, T. G., 83
Rose, D., 66
Rotem, D., 45 Tarjan, R., 66, 68
Tedder, M., 41, 82
sandwich, 40 Thomas, R., 39
Satyanarayana, A., 133 Thomassé, S., 113
Saurabh, S., 115, 157 three-coloring, 12
Scheffler, P., 149 time-complexity, 5
Schrijver, A., 40, 163 timebound, 8
search tree, 10 topline, 82
search-tree for vertex cover, 99 toroidal graph, 158
Seese, D., 156 torus, 158
Seidel switch, 161 total order, 41
Seidel, J., 161 transitive, 41
separator, 65 transitive closure, 48
set cover, 16 transitive orientation, 41, 48, 79, 81
set cover problem, 16 transitive relation, 122
set partitioning, 15 traveling salesman problem, 15
Seymour, P., 39, 40, 114, 124, 143, 163 traveling salesman tour, 16
Shannon capacity, 40 tree, 3, 68
Shannon, C., 40 tree decomposition, 121, 139
shape of matrix, 151 treewidth, 128, 129
Shirakawa, I., 6 triangle, 6
sibling, 153 triangle packing, 33
sibling leaves, 153 triangle packing problem, 33
simplicial, 66, 134 triangle partition, 25
simplicial vertex, 66 triangle partition problem, 25
Smith, K., 104 triangulated graph, 66
Snevily, H., 114 triple, 103
square of a graph, 87 trivially perfect graph, 54
square of linegraph, 87 Trotter, W. Jr., 83
squeeze, 123 Tsukiyama S., 6
star, 159 Tung, L., 133
star of the show, 72 twin, 58, 84, 85, 87, 153
Stevens, K., 113 twinset, 60, 150
Stewart, L., 55
Stirling formula, 20 union, 57, 63
172 Index

union of cliques, 84 Wagner, K., 114, 126

union-operation, 61 wall, 143
Urrutia, J., 45, 83 Wang, C., 114
Wanke, E., 157
valid bitvector, 115 Wernicke, S., 108, 119
valid partition, 106 wheel, 159
Vazsonyi’s conjecture, 125 Williams, V., 42
vertex, 1, 2, 7 Williamson, S., 36
vertex coloring, 10 Wolk, E., 55
vertex cover, 38, 84, 92, 159 Wollan, P., 127
vertex cover problem, 38, 92 worst-case timebound, 9
vertex elimination, 66
vertex ordering, 115 Xia, G., 98, 100
vertex set, 2
vertex-disjoint cycles, 159 Yannakakis, M., 46, 68
vertex-minor, 154 Yuster, R., 117
Vetta, A., 104
Vizing, V., 37 zero matrix, 151
Vušković, K., 40 Zwick, U., 117