1.

Introduction
Graph theory owes its evolution to the study of some physical problems involving sets
of objects and binary relations among them. It is difficult to pinpoint its formulation to a
single source; in fact, graph theory can be said to have been discovered many times, each
discovery being independent of the other. The earliest known studies appear in the works
of Euler, Kirchoff, Cayley and Hamilton. The twentieth century witnessed considerable activity in this area, with new discoveries and proofs being proposed as solutions to classical
problems including the celebrated four colour problem.

1.1 Basic Concepts

Let V be a nonempty set. The cartesian product of V with itself, denoted by V V , is the
set of all unordered pairs of elements of V . That is, V V = {(u, v) : u, v V }. We denote
by V(2) the set of unordered pairs of distinct elements of V, by V[2] the set of unordered pairs
of elements of V , not necessarily distinct, and by V (2) the set of ordered pairs of distinct
elements of V .
Definition: A simple graph (or briefly, a graph) G is a finite nonempty set V together with
a symmetric, irreflexive relation R on V . The elements of the set V are called the vertices of
the graph and the relation R is called the adjacency relation. If u is related to v by R, then
u is said to be adjacent to v and we write uRv. An example of a simple graph is given in
Figure 1.1(a).
Since R is a symmetric relation, it defines a subset E of V(2) . The elements of the set E
are called the edges of the graph.

(a) Simple graph

(b) Multigraph
Fig. 1.1

(c) Infinite graph

Introduction

From the above definition, we observe that a graph G is a pair (V, E ), where V is a
nonempty set whose elements are called the vertices of G and E is a subset of V(2) whose
elements are called the edges of G.
We observe that there is an incidence relation I between the vertex set V and the edge
set E of a graph. If the element e E , there is a pair of distinct vertices u and v such that
e = {u, v}. The vertices u and v are called end vertices of e, and u and v are said to be
incident with e (uIe and vIe). Also, e is said to be incident with u and v, and in that case we
write eI 0 u and eI 0 v, where I 0 is the relation converse to I .
A graph with a finite number of vertices and finite number of edges is called a finite
graph, otherwise it is an infinite graph (Figure 1(c)).
We represent a graph G with vertex set V and edge set E by (V (G), E(G)). Since we are
only going to deal with finite graphs, we write V (G) = {v1 , v2 , . . ., vn }, E(G) = {e1 , e2 , . . ., em }.
We define |V | = n to be the order of G and |E| = m to be the size of G. Such a graph is
called an (n, m) graph. If there is an edge e between the vertices u and v, we briefly write
e = uv and say edge e joins the vertices u and v. A vertex is said to be isolated if it is not
adjacent to any other vertex.
Multigraph: A multigraph is a pair (V , E ), where V is a nonempty set of vertices and E is
a multiset of edges, being a multi-subset of V[2] . The number of times an edge e = uv occurs
in E is called the multiplicity of e and edges with multiplicity greater than one are called
multiple edges. An example of a multigraph is given in Figure 1(b).
General graph: A general graph is a pair (V , E ), where V is a nonempty set of Moment
and E is a multiset of edges, being a multi-subset of V(3) . An edge of the form e = uu, u V
is called a loop. An edge which is not a loop is called a proper edge or link. The number of
times edge e occurs is called its multiplicity, and proper edges with multiplicity greater than
one are called multiple edges. Loops with multiplicity greater than one are called multiple
loops (Fig. 1.2).
If u, v V in a general graph or multigraph G, then the multiplicity of the edge uv is
denoted by qG [u, v]. If uv is not an edge, then qG [u, v] = 0. Similarly, if u is a vertex of a
general graph G, then the number of loops at u in G is denoted by qG [u].
The graph obtained by replacing all multiple edges by single edges in a multigraph G is
called the underlying graph of G. Similarly, if G is a general graph, the graph H obtained
by removing all its loops and replacing all multiple edges by single edges is called the
underlying graph of G.

Fig. 1.2 General graph

Graph Theory

1.2 Degrees
In a graph G, the degree of a vertex v is the number of edges of G which are incident to v
and is denoted by d(v) or d(v|G). We have d(v) = |{e E: e = uv, for u V }|. The minimum
degree and the maximum degree of a graph G are denoted by (G) and (G) respectively.
In the graph of Figure 1.3(a), d(v1 ) = d(v3 ) = d(v6 ) = 3, d(v2 ) = 1, d(v4 ) = 0 and d(v5 ) = 2.
A graph is said to be regular if all its vertices are of same degree and k-regular if all
its vertices are of degree k. A 3-regular graph is also called a cubic graph. A vertex with
degree zero is an isolated vertex, a vertex with degree one is a pendant vertex and the
unique edge incident to a pendant vertex is a pendant edge. A vertex of odd degree is an
odd vertex and a vertex of even degree is an even vertex. In Figure 1.3(a), v4 is an isolated
vertex and v2 is a pendant vertex. The graph shown in Figure 1.3(b) is 2-regular and that
given in Figure 1.3(c) is 3-regular, i.e., cubic.
In a general graph G, a loop incident to a vertex v is counted as two edges incident to
v. Therefore d(v) is the number of non-loop edges incident to v plus twice the number of
loops at v.

(a)

(b)

(c)

Fig. 1.3 Graphs (b) and (c) are regular

The following result is due to Euler [76].

Theorem 1.1
edges.

The sum of the degrees of a graph is even, being twice the number of

Proof Let m be the number of edges in a graph G = (V, E). Since each edge contributes
two to the degrees, one at the beginning vertex and one at the end vertex of the edge, the
sum of the degrees is even and equal to twice the number of edges. Hence,

d(v) = 2m.

v V

Theorem 1.2

Proof

In any graph there is an even number of vertices of odd degree.

Let G = (V, E) be a graph and d(v) be the degree of the vertex v V . Let |E| = m.

Then d(v) = 2m and therefore,

v V

Introduction
d(v j )

odd degree
vertices

d(vk ) = 2m.
even degree
vertices

(1.2.1)

II

Since the right hand side of (1.2.1) is even, and (II) in (1.2.1) is also even, therefore (I)
in (1.2.1) is even. Hence,
d(v j ) = even.
odd degree
vertices

This is only possible when the number of vertices with odd degree is even.

1.3 Isomorphism
Let G and H be general graphs. Let f be a oneone mapping of V (G) onto V (H), and g
be a oneone mapping of E(G) onto E(H). Let denote the ordered pair ( f , g). Then
is an isomorphism of G onto H , when the vertex x is incident with the edge e in G if and
only if the vertex f x is incident with the edge ge in H (Fig. 1.4). If such an isomorphism
exists, the graphs G and H are said to be isomorphic and is denoted by G
= H . We have,
|V (G) | = |V (H) | and |E(G)| = |E(H)|.

ge

x
fx
G

Fig. 1.4

We can think to be an operation transforming G into H and write G = H . Also, we

write v = f v and e = ge for each vertex v and each edge e of G. Clearly, G and H can
be represented by the same diagram. The representative of an edge or vertex x of G can be
reinterpreted as the representative of x in H .
An isomorphism of a graph G onto itself is called an automorphism of G. Any graph G
has the identical or trivial automorphism I such that Ix = x for each edge or each vertex x
of G.
Clearly, two graphs G and G0 are isomorphic to each other if there is a oneone correspondence between their vertices, and between their edges such that the incidence relationship is preserved. For example, the graphs shown in Figure 1.5(a), Figure 1.5(b) and
Figure 1.5(c) are isomorphic.

Graph Theory

(a) Isomorphic graphs

(b) Isomorphic graphs

(c) Isomorphic graphs

Fig. 1.5

It follows from the definition of isomorphism that two isomorphic graphs have
i. the same number of vertices,
ii. the same number of edges, and
iii. an equal number of vertices with a given degree.
However, these conditions are not sufficient. To see this, consider the two graphs given in
Figure 1.6.

Fig. 1.6 Non-isomorphic graphs

These graphs satisfy all the three conditions, but they are not isomorphic because the
vertex x in (a) corresponds to vertex y in (b) as there are no other vertices of degree three,
and in (b) there is only one pendant vertex w adjacent to y, while in (a) there are two
pendant vertices u and v adjacent to x.
The graphs in Figure. 1.7 also satisfy conditions (i), (ii) and (iii) but are not isomorphic.

Introduction

Fig. 1.7 Non-isomorphic graphs

We have the following observation on isomorphism of graphs.

Theorem 1.3

The relation isomorphism in graphs is an equivalence relation.

Proof The relation of isomorphism between graphs is reflexive because of the trivial
automorphisms.
Let = ( f , g) be an isomorphism of a graph G onto a graph H , so G
= H . Then there is
an inverse isomorphism 1 = ( f 1 , g1 ) of H onto G. So H
= G. Therefore
= is symmetric.
Now, let = ( f , g) be an isomorphism of G onto H , and = ( f 1 f , g1 g) of G onto K . Here
f 1 f is a mapping obtained by applying first f and then f 1 . Similarly, is the isomorphism
obtained applying first and then . Thus
= is transitive.
Hence the relation isomorphism is an equivalence relation.
q
Remark

The multiplication of the isomorphism defined above is associative.

Since the relation isomorphism is an equivalence relation, it partitions the class of all
graphs into disjoint nonempty subclasses called isomorphism classes, such that two graphs
belong to the same isomorphism class if and only if they are isomorphic.
Theorem 1.4 Let G and H be graphs and let f be a oneone mapping V (G) onto V (H)
such that two distinct vertices x and y of G are adjacent if and only if the corresponding
vertices f x and f y of H are adjacent in H . Then there is a uniquely determined oneone
mapping g of E(G) onto E(H) such that ( f , g) is an isomorphism of G onto H .

Proof Let e be any edge of G, having distinct ends x and y. By hypothesis, there is a
uniquely determined edge e0 of H whose ends are f x and f y. We define a oneone mapping
g by the rule ge = e0 , for each edge e of G. It is then clear that ( f , g) is an isomorphism of G
onto H .
Conversely, let g be a mapping such that ( f , g) is an isomorphism of G onto H . Then for
each edge e of G, there is an edge e0 of H such that ge = e0 .
q
An isomorphism of a graph G onto a graph H is defined as a oneone mapping of V (G)
onto V (H) that preserves adjacency. This specialisation can be regarded as an application
of Theorem 1.4.
Definition: Two graphs G(V, E) and H(U, F) are label-isomorphic if and only if V = U ,
and for any pair u, v in V , uv E if and only if uv F . The graphs of Figure 1.8 are
isomorphic, but not label-isomorphic.

Graph Theory

Fig. 1.8 Isomorphic but not label-isomorphic graphs

Definition: A graph invariant is a function f from the set of all graphs to any range of
values (numerical, vectorial or any other) such that f takes the same value on isomorphic
graphs. When the range of values is numerical (real, rational or integral) the invariant is
called a parameter. The order and size of a graph are graph parameters.

1.4 Types of graphs

Simple directed graph or simple digraph: A simple digraph (or simply digraph) D is
a pair (V , A), where V is a nonempty set of vertices and A is a subset of V (2) whose elements
are called arcs of D.
Multidigraph: A multidigraph D is a pair (V , A), where V is a nonempty set of vertices,
and A is a multiset of arcs of V (2) . The number of times an arc occurs in D is called its
multiplicity and arcs with multiplicity greater than one are called multiple arcs of D.
General digraph: A general digraph D is a pair (V , A), where V is a nonempty set of
vertices, and A is a multiset of arcs, being a multisubset V V . An arc of the form uu is
called a loop of D and arcs which are not loops are called proper arcs of D. The number
of times an arc occurs is called its multiplicity. A loop with multiplicity greater than one is
called a multiple loop.
An arc (u, v) A of a digraph is denoted by uv, implying that it is directed from u to v, u
being the initial vertex and v the terminal vertex. Clearly, a digraph is an irreflexive binary
relation on V .
Various types of digraphs are shown in Figure 1.9.

Fig. 1.9

If D(V, A) is a digraph, the graph G(V, E), where uv E whenever uv or vu or both are in
A, is called the underlying graph of D (also called the covering graph C(D) of D).

Introduction

If D(V, A) is a general digraph, the digraph D1 (V, A1 ) obtained from D by removing all
loops, and by replacing all multiple arcs by single arcs is the digraph underlying D. The
underlying graph of D1 is the underlying graph of D.
Mixed graph: A mixed graph G{V, A E} consists of a nonempty set V of vertices, a set
A of arcs (A V (2) ), and a set E of edges (E V(2) ), such that if uv E then neither uv, nor
vu is in A.
We represent a general, multi and simple graph by g-graph, m-graph and s-graph respectively.
Subgraphs: A subgraph of a graph G(V, E) is a graph H(U, F) with U V and F E .
We denote it by H < G (G is also called the super graph of H .) If U = V then H is called
the spanning subgraph of G, and is denoted by H G. Here G is called the spanning super
graph of H and is denoted by G H .
If F consists of all those edges of G joining pairs of vertices of U , then H is called the
vertex induced subgraph of G and is denoted by H =< U >. If F E , and U is the set of
end vertices of the edges of F , then H(U, F) is called an edge induced subgraph of G and
is denoted by H =< F >. These definitions are illustrated in Figure 1.10.

Fig. 1.10

If S and T are two disjoint subsets of the vertex set V of a graph G(V, E), we define [S, T ] =
{uv E : u S, v T } and q[S, T ] = |[S, T ]|.
If D = (V, A) is a digraph and S and T are disjoint subsets of V , we denote (S, T ) = {uv
A : u S and v T } and q(S, T ) = |(S, T )|.

Graph Theory

9


Complete graph: A graph of order n with all possible edges m =

n(n 1)
2

is called a

complete graph of order n and is denoted by Kn . A graph of order n with no edges is called
an empty graph and is denoted by Kn . Each graph of order n is clearly a spanning subgraph
of Kn . Some examples of complete graphs are shown in Figure 1.11.

Fig. 1.11 Complete graphs

r-partite graph: A graph G(V, E) is said to be r-partite (where r is a positive integer)

if its vertex set can be partitioned into disjoint sets V1 , V2 , . . ., Vr with V = V1 V2 . . . Vr
such that uv is an edge of G if u is in some Vi and v in some V j , i 6= j . That is, every
one of the induced subgraphs < Vi > is an empty graph. We denote r-partite graph by
G(V1 , V2 , . . ., Vr , E).
If an r-partite graph has all possible edges, that is, uv E , for every u Vi and every
v V j , for all i, j , i 6= j , then it is called a complete r-partite graph. If |Vi | = ni , we denote it
by Kn1 , n2 , ..., nr .
Bipartite graph: A graph G(V, E) is said to be bipartite, or 2-partite, if its vertex set can
be partitioned into two different sets V1 and V2 with V = V1 V2 , such that uv E if u V1 and
v V2 . The bipartite graph is said to be complete if uv E , for every u V1 and every v V2 .
When |V1 | = n1 , |V2 | = n2 , we denote the complete bipartite graph by Kn1 , n2 . For example,
K2, 2 and place K2, 3 are shown in Figure 1.12(a) and (b).
The complete bipartite graph K1, n is called an n-star or n-claw. For example, 3-star can
be seen in Figure 1.12(e).
Complement of a graph: The complement G(V, E) of a graph G(V, E) is the graph
having the same vertex set as G, and its edge set E is the complement of E in V(2) , that
is, uv is an edge of G if and only if uv is not an edge of G. In Figure 1.12(c) and (d) the
complements of K4 and K2, 3 respectively are shown.
A graph G is said to be self-complementary if G
= G. The complement Kn of the complete
graph of order n is clearly the empty graph of order n.

10

Introduction

Fig. 1.12 Complete bipartite graphs

Removal of edges: Let G(V, E) be a graph and let F E . The graph H(V, E F) with
vertex set V and edge set E F is said to be obtained from G by removing the edges in F . It
is denoted by G F . If F consists of a single edge e of G, the graph obtained by removing
e is denoted by G e.
Now, GF may contain isolated vertices which are not isolated vertices of G. The graph
obtained by removing these newly created isolated vertices from G F is denoted by G \ F .
Similarly, the graph obtained by removing isolated vertices from G e is denoted by G \ e.
Figure 1.13 illustrates this operation.

Fig. 1.13

Removal of vertices: Let G(V, E) be a graph and let v V . Let Ev be the set of all edges
of G incident with v. The graph H(V {v}, E Ev ) is said to be obtained from G by the
removal of the vertex v and is denoted by G v.
If U is a subset of V , the graph obtained by removing the vertices of G which are in U is
denoted by G U . If H is a subgraph of G, we denote G V (H) by G H , and G E(H) by
H(G). Here H(G) is called the relative complement of H in G (Fig. 1.14).

Fig. 1.14

Graph Theory

11

Addition of edges: Let G(V, E) be a graph and let f be an edge of G . The graph H(V, E
{ f }) is said to be obtained from G by the addition of the edge f and is denoted by G + f . If
F is a subset of edges of G , the graph obtained from G by adding the edges of F is denoted
by G + F (Fig. 1.15(a) and (b)).

Fig. 1.15

Addition of vertices: Let G(V, E) be a graph and let v

/ V . The graph H with vertex set
V {v} and edge set E {uv, for all u V } obtained from G by adding a vertex v, is denoted
by G + v. Thus G + v is obtained from G by adding a new vertex and joining it to all vertices
of G. For illustration, refer to Figure 1.15(c).
Join of graphs: Let G(V, E) and H(U, F) be two graphs with disjoint vertex sets (V U =
). The join of G and H denoted by GV H is the graph with vertex set V U and edge set
Et F t[V, t]. So the join is obtained from G and H by joining every vertex of G to each
vertex of H by an edge. Clearly, G + v = GV K1 = GV K1 . This operation is illustrated in
Figure 1.16.

Fig. 1.16

1.5 Graph properties

Parametric property: A property P is called a parametric property of a graph if for a
graph G having property P, every graph isomorphic to G also has property P. A graph with
property P is denoted as P-graph. A subgraph H of a graph G is said to be maximal with
respect to a property P (or P-maximal) if H is a P-graph, and there is no subgraph K of G
having property P which properly contains H . So H is P-maximal if H is a P-graph, and

12

Introduction

for every e E(G) E(H), H + e is not a P-graph, that is, the addition of any edge to H
destroys the property P of H . A subgraph H of a graph G is said be minimal with respect
to a property P (or P-minimal) if H is a P-graph and H has no proper subgraph K which is
also a P-graph. So H is minimal if and only if H is a P-graph, and for every e E(H), H e
is not a P-graph, that is, if and only if the removal of any edge destroys the property P.
P-critical: A graph G is said to be P-critical if G is a P-graph and for every v V , G v is
not a P-graph.
Hereditary property: A property P is said to be a hereditary property of a graph G if a
graph G has the property P, then every subgraph of G also has the property P. It is called
an induced-hereditary property if every induced subgraph of G also has the property.
Monotone property: A property P is said to be a monotone property of a graph G, if G
has the property P, then for every e E(G), G + e also has the property P.
A vertex subset U of a graph G is said to be an independent set of G if the induced
subgraph < U > is an empty graph. An independent set of G with maximum number of
vertices is called a maximum independent set (MIS) of G. The number of vertices in a MIS
of G is called the independent number of G and is denoted by 0 (G).
A maximal complete subgraph of G is a clique of G. The order of a maximum clique of
G is the clique number of G and is denoted by w(G). Consider the graph given in Figure
1.17. Here {1, 4} is independent set, {1, 3, 5} is an MIS, 0 (G) = 3, {1, 2, 6} is a clique,
{2, 3, 4, 6} is a maximum clique and w(G) = 4.

Fig. 1.17

The property of being a complete subgraph is an induced hereditary property. The property
of being a planar graph is hereditary and the property of being a nonplanar graph is a
monotone property.

1.6 Paths, Cycles and Components

The incidence relation I between the elements of V and the elements of E induced by the
adjacency relation R further induces an adjacency relation among the edges, namely, two
edges e and f are adjacent if and only if they have an end vertex in common. This relation

Graph Theory

13

is denoted by L. Two edges e and f of a graph G(V, E) are adjacent if and only if e = uv and
f = vw, for some three vertices u, v, w. If e and f are adjacent, it is denoted by eL f .
Walks: An alternating sequence of vertices and edges, beginning and ending with vertices
such that no edge is traversed or covered more than once is called a walk. A vertex may
appear more than once in a walk. v1 e1 v2 e2 v3 . . .vk ek vk+1 is called a v1 vk+1 walk W , v1 and
vk+1 are called the initial and terminal vertices of the walk W , and k, the number of edges
in W is called length of W . The walk is said to be open if v1 and vk+1 are distinct, and closed
if v1 = vk+1 . If vr and vs are two vertices in W , s > r, then the walk vr er vr+1 er+1 . . .es1 vs
is called a subwalk of W , or vr vs section of W , and is denoted by W vr vs . The walk
vk+1ek vk . . .e1 v1 is called the reverse walk of W and is denoted by W 1 .
Paths: A path is an open walk in which no vertex (and therefore no edge) is repeated. A
closed walk in which no vertex (and edge) is repeated is called a cycle. A path of length
n is called an n-path and is denoted by Pn . A cycle of length n is called an n-cycle and is
denoted by Cn . A loop is 1-cycle and a pair of edges joining two vertices form a 2-cycle.
An n-cycle is proper only if n 3.
As the edges and vertices in a path or cycle are not repeated, these are denoted by the
sequence of vertices only. For example u1 u2 . . .uk , where ui V is a k 1 path.
Two distinct vertices u and v of a graph G are said to be connected or joined if there
is a u v walk in G. By convention, a vertex is connected to itself. A graph is said to be
connected if every two of its vertices are connected, otherwise it is disconnected. The relation of connectedness is an equivalence relation on the vertex set V of a graph. The graphs
induced on the equivalence classes of this relation are called the components of the graph.
Component of a graph: A maximal connected subgraph of a graph G is called a component of G. A component which is K1 is called a trivial component. The number of components of a graph G is denoted by k(G). A component of G with an odd (even) number of
vertices is called an odd (even) component of G. The number of odd components of G is denoted by k0 (G). Consider the graph shown in Figure 1.18. Here v1 e1 v2 e2 v3 e3 v4 e4 v2 e5 v5 e6 v6
is a walk, v1 e1 v2 e2 v3 e3 v4 is a path and v2 e2 v3 e3 v4 e4 v2 is a 3-cycle.

Fig. 1.18

Union of graphs: Let G(V, E) and H(U, F) be two graphs with V U = . The union of
G and H , denoted by G H , is the graph with vertex set V U and edge set E F . Clearly,

14

Introduction

when G and H are connected graphs, G H is a disconnected graph whose components are
G and H (Fig. 1.19).

Fig. 1.19

A graph G which has k components, all isomorphic to H , is H H . . . H , or G = kH . We

write G = k1 G1 k2 G2 . . . kr Gr , if G has ki components isomorphic to Gi , 1 i r.
Theorem 1.5 A graph G is disconnected if and only if its vertex set V can be partitioned
into two nonempty, disjoint subsets V1 and V2 , such that there exists no edge in G whose
one end vertex is in V1 and the other in V2 .

Proof Let G be a graph whose vertex set can be partitioned into two nonempty disjoint
subsets V1 and V2 , so that no edge of G has one end in V1 and the other in V2 . Let v1 and v2 be
any two vertices of G such that v1 V1 and v2 V2 . Then there is no path between vertices
v1 and v2 , since there is no edge joining them. This shows that G is disconnected.
Conversely, let G be a disconnected graph. Consider a vertex v in G. Let V1 be the set of
all vertices that are joined by paths to v. Since G is disconnected, V1 does not contain all
vertices of G. Let V2 be the set of the remaining vertices. Clearly, no vertex in V1 is joined
to any vertex in V2 by an edge, proving the converse.
q
Theorem 1.6
by a path.

If a graph has exactly two vertices of odd degree, they must be connected

Proof Let G be a graph with all its vertices of even degree, except for v1 and v2 which are
of odd degree. Consider the component C to which v1 belongs. Then C has an even number
of vertices of odd degree. Therefore C must contain v2 , the only other vertex of odd degree.
Thus v1 and v2 are in the same component, and since a component is connected, there is a
path between v1 and v2 .
q
Lemma 1.1
k

i=1

Proof

n2i

For any set of positive integers n1 , n2 , . . ., nk

k

ni
i=1

!2

(k 1) 2 ni k .
i=1

i=1

i=1

We have, (ni 1) = ni k.

Graph Theory

15

Squaring both sides, we get

"

#2

(ni 1)

i=1

"

ni k

i=1

#2

,
k

ni

or [(n1 1) + (n2 1) + . . .+ (nk 1)] =

i=1
2

or (n1 1) + (n2 1) + ... + (nk 1) +

i=1

Therefore,
i=1

or

i=1
k

or

i=1

n2i
n2i

n2i 2

i=1

2 ni k +
i=1

ni

i=1

2

ni
i=1

(k 1)

i=1

(ni 1)(n j 1)=

j=1
i 6= j

(ni 1)(n j 1) =
j=1
i 6= j

i=1

2k ni + k2 ,

ni + k +

Theorem 1.7
1
2

!2

2

ni
i=1

2

ni
i=1

2

2k ni + k2 .
i=1

2k ni + k2 ,
i=1

2k ni + k2 ,
i=1

2 ni k
i=1

A graph with n vertices and k components cannot have more than

(n k)(n k + 1) edges.

Proof The number of components is k and let the number of vertices in ith component
be ni , 1 i k.
So,

n1 + n2 + . . .+ nk = ni = n.
i=1

Now, a component with ni vertices will have the maximum possible number of edges
when it is complete, and in that case it has 21 ni (ni 1) edges.
Thus the maximum number of edges in G =
k

1
1
n2i 2 ni 2
2 i=1
i=1

ni
i=1

!2

1
2

ni (ni 1)
i=1
k

(k 1) 2 ni k
i=1

1 k
ni
2 i=1

1
 1
1 2
1
n (k 1) (2n k) n = n2 2nk + k2 + n k = (n k)(n k + 1).
2
2
2
2

The following result is due to Turan [247].

16

Introduction

Theorem
number of edges among all n vertex graphs with no trianh maximum
i
h 2 i1.8 The
2
n2
n
gles is 4 , where 4 is the greatest integer not exceeding the number n4 .

Proof The result is proved by induction, taking the cases of n odd and n even separately.
Let n be even. The result is obvious for small values of n. Assume the result to be true
for all even n 2p. We then prove it for n = 2p + 2. So assume G is a graph with n = 2p + 2
vertices and no triangles. Since G is not totally disconnected, there are adjacent vertices u
and v. The subgraph G0 = G{u,
h 2 i v} has 2 p vertices and no triangles, so that by the induction
hypothesis, G0 has at most 4p4 = p2 edges.
There can be no vertex w such that u and v are both adjacent to w, for then u, v and w
form a triangle of G. Therefore, if u is adjacent to k vertices of G0 , v is adjacent to at most
2n k vertices. Then G has at most
p2 + k + (2p k) + 1 = p2 + 2p + 1 = (p + 1)2 =

 2
n2
n
=
edges.
4
4

It can be shown that for all even p, there exists a ( p,

graph is formed as follows. Take two sets V1 and V2 of
of V1 with each vertex of V2 .
Example

p2
4 ) graph with no triangles. Such a
p
2 vertices each, and join each vertex

When n = 6, the graph formed is shown in Figure 1.20.

Fig. 1.20

The following result due to Konig [136] characterises bipartite graphs.

Theorem 1.9

A graph is bipartite if and only if it contains no odd cycles.

Proof Let G be a bipartite graph. Then its vertex set V can be partitioned into two sets
V1 and V2 , so that every edge of G joins a vertex of V1 with a vertex of V2 . Thus every cycle
v1 v2 . . .vn v1 in G has its oddly subscripted vertices in V1 , say, and the others in V2 , so that its
length is even.
Conversely, we assume, without loss of generality, that G is connected, for otherwise
we can consider the components of G separately. Take any vertex v1 V , and let V1 consist
of v1 and all vertices of G whose shortest paths from v1 (in terms of the number of edges)
contain an even number of edges, while V2 = VV1 . Since G has no odd cycles, every edge
of G joins a vertex of V1 with a vertex of V2 (Fig. 1.21).

Graph Theory

17

For suppose there is an edge uv joining two vertices of V1 . The length of v1 to u is even,
the length of v1 to v is even and the union of edges from v1 to u and from v1 to v together
with the edge uv contains an odd cycle, which is a contradiction. Hence G is bipartite. q

Fig. 1.21

Theorem 1.10

A graph and its complement are not both disconnected.

Proof Let G be a disconnected graph and u, v be any two vertices of G, the complement
of G. If u and v belong to different components of G, then u and v are adjacent in G, by
definition. If u and v belong to the same component, say Gi of G, then let w be a vertex of
some other component, say G j of G. By definition both u and w, and v and w are adjacent
in G. In either case u is connected to v by a path in G. Thus G is connected.
q
Theorem 1.11

For any graph G with six vertices, G or G contains a triangle.

Proof Let v be a vertex of a graph G with six vertices. Since v is adjacent either in G or
G to the other five vertices of G, we assume without loss of generality, that there are three
vertices u1 , u2 and u3 adjacent to v in G. If any two of three vertices are adjacent, then they
are two vertices of a triangle whose third vertex is v. If no two of them are adjacent in G,
then u1 , u2 and u3 are the vertices of a triangle in G (Fig. 1.22).
q

Fig. 1.22

Theorem 1.12

For any graph G of order n, (G) n 1.

18

Introduction

Proof Since a vertex in a graph can be joined to at most n1 other vertices, the maximum
degree a vertex can have is n 1. So, (G) n 1.
q
Theorem 1.13 Every graph G contains a bipartite spanning subgraph whose size is at
least half the size of G.

Proof Let B be a bipartite spanning subgraph of G with B being of maximum size and
the bipartition of its vertex set be V = V1 V2 . Clearly, all the edges of G with one end in V1
and the other in V2 must be in B.
Now, if any vertex u V1 is joined to p vertices of V1 , then u is joined to at least p vertices
of V2 . That is, the number of vertices of V2 joined to u is greater or equal to the number of
vertices of V1 joined to u (Fig. 1.23).

Fig. 1.23

For if u V1 is joined to more vertices of V1 than of V2 , then V1 {u}, V2 {u} is a bipartition

of V giving a bipartite subgraph of G with larger size, which is a contradiction to our
assumption (Fig. 1.24).

Fig. 1.24

Therefore, d(u|B)

1
2 d(u|B),

for each u V .

Summing over all u V , we get m(B) m(G).

Theorem 1.14 If G is a graph with at least n0 vertices and at least n0 .n(G) 2n0 +1 + 1
edges, then G contains a subgraph H with (H) n0 + 1, where n(G) is the number of
vertices in G and n(G) n0 .

Graph Theory

Proof

19

Let Gn0 be the set of all such graphs G with





n0 + 1
Gn0 = G : n(G) > n0 , m(G) n0 .n(G)
+1 .
2

If G Gn0 , then n(G) > no, because if n(G) = no, then






n0 + 1
n0 + 1
+ 1 gives m(G) n0 .n0
+ 1.
2
2


n0 (n0 1)
(n0 + 1)n0
2
2
2
Therefore, m(G) n0
+ 1 = n0 n0
+1
2
2


(n0 1)n0
n0
+1 =
+ 1.
=
2
2


n0
Thus, m(G)
+ 1, which is a contradiction, as the number of edges can be at
2


n0
most
.
2
m(G) n0 .n(G)

Now, if (G) n0 , let u be a vertex with d(u) n0 . Then it is easily verified that G u
also belongs to Gn0 . By repeating this process of removing vertices with degree smaller
than n0 + 1 (i.e., n0 ), we arrive at a graph H Gn0 with (H) = n0 + 1.
q
Example

Consider the graph G of Figure 1.25. Here n(G) = 6. Let n0 = 3. Clearly, m(G)
4
36
+ 1 = 13. In G u, we have n(G u) = 5 and again let n0 = 3. Then m(G u)
35

2

4
+ 1 = 10, and thus satisfies given conditions. Hence G u Gn0 .
2

Fig. 1.25

Theorem 1.15 Any graph G has a regular supergraph H of degree (G) such that G is an
induced subgraph of H .

Proof We have to prove that for a graph G, there exists a graph H which is regular and the
degree of every vertex of H is (G), and H is the supergraph of G, (G) being the maximum
degree in G. If G is regular, there is nothing to prove.
Now let G be not regular and = (G) be the maximum degree in G. Let di be the degree
n
of the vertex vi in G. Let d = ( di ).
i=1

20

Introduction

Case 1 If d is even, take d2 disjoint copies of the graph K+1 e (where e is an edge). For
1 i n, join vi to di different vertices from these copies having degree 1. We get a
graph H which is -regular. Clearly, G is an induced subgraph of H .

Example Consider the graph of Figure 1.26. Here = 2, d = 2 1 + 2 1 + 2 2 = 2

(even). We take d2 = 22 = 1 copy of K2+1 e = K3 e. Join vi to 2 di different vertices from
this copy having degree 2 1 = 1. Join v1 to 2 1 = 1 vertex of this copy having degree 1.
Join v2 to 2 1 = 1 vertex.

Fig. 1.26
(d1)

Case 2 If d is odd, take 2 disjoint copies of the graph K+1 e. For 1 i n, join vi to
di different vertices of degree 1 from these copies. This can be done for all but one
vertex of G, say v j . Then d(v j ) = 1 in the graph H1 so constructed. Take another copy
0
0
0
of H1 say H1 with d(v j ) = 1. Then H is obtained by joining v j and v j by an edge. Hence
H is regular and G is an induced subgraph of H .
q

Example

Consider the graph G of Figure 1.27. Here = 3, d = 3 1 + 3 1 + 3 1 + 3

(d1)
(71)
2 = 2 + 2 + 2 + 1 = 7, 2 = 2 = 3. We take 3 copies of K3+1 e = K4 e.

Fig. 1.27

Graph Theory

21

Theorem 1.16 If G(X, Y, E) is a bipartite graph with maximum degree , then G has a
-regular bipartite supergraph H(U, V, F) with X U , Y V , E F .

Proof Let be the maximum degree in G(X, Y, E). Take copies of G (each isomorphic
to G) say Gi (Xi , Yi , Ei ), 1 i . For each x X1 with d(x) < , take r = d(x) vertices
say x1 , x2 , . . ., xr . Let Yo be the set of all such vertices as x ranges over X1 and let V =
Yo Y1 . . . Y. Join each vertex xi to one vertex x(d(x) < ) from each of the copies X1 ,
X2 , . . ., X .
Similarly, for each y Y1 with d(y) , take s = d(y) vertices say y1 , y2 , . . ., ys . Let
Xo be that set of all such vertices as y ranges over Y1 and let U = X0 X1 . . . X . Join
each vertex y j to one vertex y(d(y) < ) from each of the copies Y1 , Y2 , . . ., Y . The resulting
bipartite graph H has clearly induced bipartite subgraphs Gi each isomorphic to G (Fig.
1.28).
q

Fig. 1.28

Example 1 Consider the graph G of Figure 1.29. Here = 3, d(u) = 2 < 3, d(x) = 2 < 3,
d(y) = 2 < 3, d(z) = 1 < 3. Now, r = d(u) = 3 2 = 1 and s = d(y)+d(y)+d(z) =
3 2 + 3 2 + 3 1 = 4.

Fig. 1.29

Example 2 In the graph G of Figure 1.30, = 2 and d(v) = 0 < 2, d(x) = 1 < 2, d(y) = 1 < 2
so that r = d(v) = 2 0 = 2 and s = 2 1 + 2 1 = 2.

22

Introduction

Fig. 1.30

1.7 Operations on Graphs

Let G1 (V1 , E1 ) and G2 (V2 , E2 ) be two graphs. The union of G1 and G2 , denoted by G1 G2 ,
is the graph whose vertex set is V1 V2 and edge set is E1 E2 . The intersection of G1 and
G2 , denoted by G1 G2 , is the graph consisting only of those vertices and edges that are
both in G1 and G2 . Clearly, G1 G2 = (V1 V2 , E1 E2 ).
The ring sum of G1 and G2 , denoted by G1 G2 , is a graph whose vertex set is V1 V2 and
whose edges are that of either G1 or G2 , but not of both. Examples of union, intersection
and ring sum are given in Figure 1.31.

Fig. 1.31

We observe that these three operations are commutative, that is,

G1 G2 = G2 G1 , G1 G2 = G2 G1 , G1 G2 = G2 G1 .

If G1 and G2 are edge disjoint, then G1 G2 is a null graph and G1 G2 = G1 G2 . If G1

and G2 are vertex disjoint, then G1 G2 is empty. For any graph G, G G = G G = G and
G G = null graph.
If H is a subgraph of G, then G H is by definition, that subgraph of G which remains
after all the edges in H have been removed from G. We write, G H = G H , whenever
H G. G H = G H is also called complement of H in G. Figure 1.32 illustrates this
operation.

Graph Theory

23

Fig. 1.32

Decomposition: A graph G is said to be decomposed into two subgraphs G1 and G2

if G1 G2 = G and G1 G2 = a null graph. In other words, every edge of G occurs either
in G1 or in G2 , but not in both, while as some of the vertices can occur in both G1 and
G2 . In decomposition, isolated vertices are disregarded. In Figure 1.33, graph G has been
decomposed into subgraph G1 and G2 .

Fig. 1.33

Deletion: Let G be a graph and v be any vertex in G. Then G v denotes the subgraph
of G by deleting vertex v, and all the edges of G which are incident with v. An example of
deletion of a vertex is given in Figure 1.34.

Fig. 1.34

If e is any edge of G, then Ge is a subgraph of G obtained by deleting e from G (Fig. 1.34).

Deletion of an edge does not imply deletion of its end vertices. Therefore G e = G e .
Fusion: A pair of vertices u and v in a graph are said to be fused (merged or identified) if u
and v are replaced by a single new vertex such that every edge incident on u or v is incident
on this new vertex. Therefore, fusion of vertices does not alter the number of edges, but

24

Introduction

reduces the number of vertices by one. In Figure 1.35, vertices u and v are fused to a single
vertex w.

Fig. 1.35

Let G1 (V1 , E1 ) and G2 (V2 , E2 ) be two graphs with V1 V2 = and E1 E2 = . Let G be

the compound graph whose vertex set is V1 V2 and edge set is E1 E2 E3 , where E3 is a
subset of the set of edges {vi v j : vi V1 , v j V2 }. We represent E3 by the symmetric binary
relation V1 V2 , and the compound graph G by G1 G2 . If = , then we get the union
of G1 and G2 , denoted by G1 G2 . If = V1 V2 (that is all edges between V1 and V2 ), we
get the join of G1 and G2 , denoted by G1 G2 . If is a function from V1 to V2 , we have
the function graph G1 f G2 . If defines a homomorphism from G1 to G2 , we have the
homomorphism graph G1 G2 . If G
= G2 = G (say) and is a bijection of V1 to V2 , we
have a permutation graph G G. This is also denoted by P (G). If G1
= G2 = G (say) and
is an automorphism of G, then G G is an automorphism graph. These operations are
illustrated in Figure 1.36.

Fig. 1.36

Let G1 (V1 , E1 ) and G2 (V2 , E2 ) be two graphs, and let the vertex sets V1 and V2 be labelled
by the same set of labels. We note here that each of V1 and V2 need not have all the labels.

Graph Theory

25

The sum of G1 and G2 , denoted by G1 +G2 , is the graph G = (V, E), where V = V1 V2 and
an edge vi v j E if and only if vi v j is an edge of G1 or G2 or both. The direct sum denoted
by G1 (+)G2 is the graph G(V, E) with V = V1 V2 and an edge vi v j E if and only if vi v j is
an edge of G1 or G2 , but not of both. The superposition graph G = G1 s G2 has vertex set
V = V1 V2 and the edge set E contains all the edges of G1 and G2 with the identity of edges
of G1 and G2 in G being preserved by assigning two different labels to these edges. So, if
vi v j is an edge in both G1 and G2 , then there are two edges vi v j in G with different labels.
For illustration of these operations, see Figure 1.37. We can extend these operations to a
finite number of graphs.
We form the compound graphs G from G1 and G2 with vertex sets V1 V2 as follows.

Fig. 1.37

The cartesian product of G1 and G2 , denoted by G = G1 G2 , is the graph whose vertex

set is V = V1 V2 and for any two vertices w1 = (u1 , v1 ) and w2 = (u2 , v2 ) in V , u1 , u2 V1
and v1 , v2 V2 , there is an edge w1 w2 E(G) if and only if either (a) u1 = u2 and v1 v2 E2 ,
or (b) v1 = v2 and u1 u2 E1 .
The tensor product (conjunction), denoted by G = G1 G2 , is the graph with vertex set
V = V1 V2 and for any two vertices w1 = (u1 , v1 ) and w2 = (u2 , v2 ) in V ; u1 , u2 V1 and
v1 , v2 V2 , there is an edge w1 w2 E(G) if and only if u1 u2 E1 and v1 v2 E2 .
The normal product (strong product), denoted by G = G1 G2 , is the graph with vertex
set V = V1 V2 and for any two vertices w1 = (u1 , v1 ) and w2 = (u2 , v2 ) in V ; u1 , u2 V1 and
v1 , v2 V2 , there is an edge w1 w2 E(G) if and only if one of the following holds:
(a) u1 = u2 and v1 v2 E2 (b) v1 = v2 and u1 u2 E1 (c) u1 u2 E1 and v1 v2 E2 .
It is easy to see that G1 G2 and G1 G2 are spanning subgraphs of G1 G2 and also
G1 G2 = (G1 G2 ) (G1 G2 ). Figure 1.38 illustrates these operations.

Fig. 1.38

26

Introduction

The composition of G1 and G2 , denoted by G = G1 [G2 ], is the graph with vertex set
V = V1 V2 , and for any two vertices w1 = (u1 , v1 ) and w2 = (u2 , v2 ) in V ; u1 , u2 V1 and
v1 , v2 V2 , there is an edge w1 w2 E(G) if and only if either (a) u1 u2 E1 or (b) u1 = u2 and
v1 v2 E2 (Fig. 1.39).

Fig. 1.39

1.8 Topological Operations

Definition: A subdivision of the edge e = uv of a graph G is the replacement of the edge e
by a new vertex w and two new edges uw and wv. This operation is also called an elementary
subdivision of G (Fig. 1.40).

Fig. 1.40

A graph H obtained by a sequence of elementary subdivisions from a graph G is said

to be a subdivision graph of G, or to be homeomorphic (or a homeomorph) of G. Two
graphs H1 and H2 which are homeomorphs of the same graph G are said to be homeomorphic to each other (or homeomorphically reducible to G). A graph G is homeomorphically
irreducible if whenever a graph H is homeomorphic to G, then H is homeomorphic from
G. In Figure 1.41, H1 and H2 are homeomorphs of K4 and K4 itself is homeomorphically
irreducible.

Graph Theory

27

Fig. 1.41

The relation being homeomorphic to each other is an equivalence relation and each
equivalence class contains a unique homeomorphically irreducible graph which is taken as
the representative of the class.
Identification of the vertices: Let G(V, E) be a general graph and let U be a subset of
the vertex set V . Let H be the general graph obtained from G by the following operations.
i. Replace the set of vertices U by a single new vertex u.
ii. Replace the edge e = ab with a U and b V U by a corresponding edge e0 = ub.
iii. Replace each edge e = ab with a, b U (b possibly being the same as a) by a loop
at u.
Let K be the multigraph obtained from H by dropping all loops and L be the graph obtained
from K by replacing each multiple edge by a single edge. Then H, K, L are respectively
said to be obtained from G by a general, multiple, simple identification of the vertices of
U . We write H = G U, K = G : U, L = G.U . This operation is illustrated in Figure 1.42.
Remarks

1. The operation in (ii) may result in the generation of multiple edges and that in (iii),
in the generation of loops even when G is a simple graph.


2. For b 6= u in H , qH [a, b] = qG [a, b] : a U .

Fig. 1.42

28

Introduction

Coalescence of G: Let G(V, E) be a general graph and V = U1 U2 . . .Ur be a partition

of the vertex set V with Ui U j 6= , i 6= j . Let H(U, E 0 ) be the general graph defined by
replacing
i. each Ui by a single new vertex ui ,
ii. each edge e = vi v j with vi Ui and v j U j (i 6= j) by a corresponding edge e0 = ui u j
and
iii. each edge e = ab with a, b Ui by a loop at ui .
Let K be the multigraph obtained from H by dropping the loops and L be the graph
obtained from K by replacing the multiple edges by single edges. Then H, K, L are respectively called the general/multiple/simple coalescence of G and we write H = G , K = G :
and L = G.. Coalescence of a graph is illustrated in Figure 1.43.
Remarks

1. Clearly H is obtained from G by sequentially identifying the vertices in Ui , 1 i r

in any order.
2. To coalescence a partition 1 of a proper subset V1 of V , we may adopt the above
definition by augmenting 1 to a partition of V by adjoining the singleton sets corresponding to the vertices in V V1 .

Fig. 1.43

Let G(V, E) be a general graph and f : V W be a map of V onto W = {w1, w2 , . . ., wr }.

Let Ui = f 1 (w) and V =

r
S

Ui be the partition of V induced by f . Then the general/multiple/

i=1

simple coalescences of G are denoted by f g (G), f m (G), f s(G) or G f , G : f , G. f , respectively,

and are referred to as a general/multiple/simple homomorph or homomorphic image of G.
We will use f (G) as a common symbol for these and may often abbreviate this to f G.
Let M(W 0 , E 0 ) be a general/multiple/simple graph with W W 0 such that f (G) is a general/multiple/simple subgraph of M . Then f is said to be a homomorphism of G into M . If
W 0 = W , then f is said to be an onto (surjective) homomorphism. If F(G) =< f (V ) >, then
f is called a full homomorphism. In case f (G) = M , then f is called a full onto homomorphism.

Graph Theory

29

Remarks

1. An injective homomorphism (i.e., f is injective) is a monomorphism. A full onto

monomorphism is an isomorphism.
2. If every Ui is an independent set in G, then f is a discrete homomorphism. If U1 =
{u, v} and other Ui s are singleton sets of V , then f is called an elementary homomorphism, and if uv / E , then f is a discrete elementary homomorphism.
3. If every < Ui > is a connected subgraph of G, then f is called connected homomorphism. An elementary homomorphism with uv E is called a connected elementary
homomorphism.
4. A mapping f : V W 0 , where G(V, E) and M(W 0 , E 0 ) are simple graphs, is a homomorphism of G into M if and only if uv E implies that f (u) f (v) E 0 .
The following result is due to Ore [178].
Theorem 1.17 Any homomorphism is the product of a connected and a discrete homomorphism.

Proof Let f be a homomorphism of G(V, E) into M(W 0 , E 0 ) with f (V ) = W (i.e., f (Vi ) =

Wi ). So, V = f 1 (W ) and Vi = f 1 (Wi ). Then each < f 1 (Wi ) > can consist of a number
of components (which are, in fact, connected) Ci1 , Ci2 , . . ., Ci j , . . ., Cini with vertex sets
Vi1 , Vi2 , . . ., Vi j , . . ., Vini .
The coalescence of Vi = f 1 (Wi ) can be done in two stages. First, we perform a multiple
coalescence of Vi = Vi1 Vi2 . . . Vi j . . . Vini sending vertex set Vi j to a single vertex Wi j .
This, when done for all Vi = f 1 (Wi )s, corresponds to a connected homomorphism of G.
Next, identify the vertices Wi j , 1 j ni of (G) into a single vertex Wi for 1 i |W |. This
corresponds to a discrete homomorphism of (G) into M . Clearly, f (G) = ((G)). q
Contractions: Let G(V, E) be a general graph and F be a subset of E such that the
edge-induced subgraph < F > is connected on the vertex subset U = V (< F >). Then the
general/multiple/simple graphs obtained by the identification of the vertex set U 0 in the
general graph G F are called the general/multiple/simple contractions of F in G and
are denoted by G| F , G||F and G|F respectively. Consider the graph of Figure 1.44. The
vertices of U = {v1 , v2 , v3 , v4 } in G F have been identified.

30

Introduction

Fig. 1.44

Let G(V, E) be a general graph and let F be a subset of E such that the edge-induced
subgraph < F > has components < Fi > on vertex sets V (< Fi >) = Ui , 1 i r. Let be the
partition of V induced by the Ui , vertices in Ui being taken as singleton sets of the partition.
The general/multiple/simple coalescences in the general graph G F are called the general/multiple/simple contractions of F in G and are denoted by G| F , G||F and G|F respectively.
Let G(V, E) be a general graph and let W = {w1, w2 , . . ., wr }. Suppose f : V W is a map
such that < Ui > is a connected general subgraph of G, where Ui = f 1 (wi ), 1 i r( f (Ui ) =
wi ). Let Fi E(Ui ) be such that Fi induces a spanning connected subgraph of < Ui > and
let F =

r
S

Fi . Then G| F , G||F and G|F are respectively called general/ multiple/simple

i=1

contraction induced by f and F and are denoted by f F (G). Here f is called a contraction
mapping of G.
If f F (G) = H , then G is said to be g/m/s- contractible to H and H is called a g/m/scontraction of G, according as H is a g/m/s-graph. We call simple contractible and simple
contraction as contractible and contraction.
If a subgraph of G is contractible to H (equivalently, if H is a subgraph of a contraction
of G), then H is called a subcontraction of G, and we also say that G is subcontractible to
H . We denote this by G}H .
Remarks

1. It can be observed that corresponding to every general contraction there exists a

unique contraction mapping f but given a mapping f from V (G) to V (H), with G
and H being general graphs, there can be various contractions of G into H defined
by suitable edge subsets of

k
S

< f 1 (wi ) >. Any two such contractions of the same

i=1

graph G corresponding to the same function f differ only in the number of loops at
wi W = V (H). Thus all such contractions correspond to a unique simple contraction
induced by f .
2. We note the distinction between a contraction and a connected homomorphism. A
connected homomorphism f is a particular case of a contraction induced by f when
the edge subsets Fi are the null subsets of E < f 1 (wi ) >. This is called the null con-

Graph Theory

31

traction corresponding to f and is denoted by f 0 (G). Also, the connected homomorphism induced by f differs from any proper contraction induced by f only in the number of loops at wi W . The contraction induced by f in which each Fi = E < f 1 (wi ) >
is called the full contraction induced by f .
3. If e = uv E and F = {e}, then G| e is called an elementary contraction of G and
differs from the corresponding connected elementary homomorphism G {v, u} by
the absence of a loop at the common vertex w = {u, v} in H .
The following result is due to Behzad and Chartrand [15].
Theorem 1.18

If a graph H is homeomorphic from a graph G, then G is a contraction of H .

Proof If G = H , then obviously G is the null contraction of H corresponding to the identity mapping f : V (H) V (G).
Let G 6= H . Then H is obtained from G by a sequence of elementary subdivisions, say
1 , 2 , . . ., k , through the intermediate graphs Gi =i (Gi1 ), G0 = G, Gk = H . Let Gi be
obtained from Gi1 by subdividing the edge uv, introducing a new vertex w. Then Gi1 is
obtainable from Gi by the elementary contraction with f = {uw}. That is, Gi1 = Gi |uw =
(Gi ), say. Then, clearly we have G = 1 (2 . . .(k (H) . . .). Therefore G can be obtained
from H by a sequence of elementary contractions and is thus a contraction of H .
q
The converse of the above theorem is not true in general. To see this, consider the graphs
in Fig. 1.45. Here the graph H is a contraction of the graph G by using the mapping f (a) =
f (b) = f (c) = a0 , f (g) = g0 , f (h) = h0 , f (d) = f (e) = f ( f ) = d 0 . But G is not homeomorphic
from H .

Fig. 1.45

Corollary 1.1 If a graph H contains a subgraph K homeomorphic from a non-trivial

connected graph G, then G has a sub contraction of H .
The next result due to Halin can be found in Ore [177].
Theorem 1.19 If a graph G is contractible to a graph H and (H) 3, then G has a
subgraph homeomorphic from H .

32

Introduction

Proof Let G be a graph contractible to the graph H . Hence there is a mapping f : V (G)
V (H) inducing the contraction H . Therefore, Gi =< f 1 (vi ) > for vi V (H) are connected
induced subgraphs of G which have been contracted to the vertices vi of H . Let Ui = f 1 (vi ).
As (H) 3, for given U1 there are at most three Ui , say U2 , U3 , U4 such that there is
an edge from U1 to Ui . Let the end vertices in U1 of three such edges from U1 to U2 , U3 ,
U4 , respectively be u2 , u3 , u4 . Since G1 is connected, there is u2 u3 path P and a u4 u5
path Q in G1 , where u5 is the first vertex that Q has in common with P. (If there are only
two vertices, we consider only the path P.) For each edge vi v j in H , we choose one edge
between Ui and U j in G and then the set of subgraphs PUQ in Gi . The graph H1 so formed is
a subgraph of G and clearly from the construction it is homeomorphic from H (Fig. 1.46).
q

Fig. 1.46

Corollary 1.2 If the graph G has a sub contraction H with (H) 3, then G has a subgraph homeomorphic from H .
Theorem 1.20 If G is subcontractible to K5 , then either G has a subgraph homeomorphic
from K5 , or G has a subcontraction to the graph L.

Proof We assume without loss of generality, that the given graph G is contractible to K5 .
Let V (K5 ) = {v1 , v2 , v3 , v4 , v5 } and let Ui = f 1 (vi ), 1 i 5, where f : V (G) V (K5 ) is the
contraction mapping. Then Gi =< Ui > are connected subgraphs of G and there is an edge
in G between Ui and U j for each pair (i, j ). Let u2 , u3 , u4 , u5 be the end vertices in U1 , of
four such edges selected from U1 to the other Ui (2 i 5). Since G1 is connected, there is
a u2 u5 path P in G, and there are u4 u04 , u5 u05 paths Q and R, where u04 and u05 are the
first vertices that these paths have in common with P.
If u04 6= u05 , we can contract each Ui , 2 i 5 to a vertex to get a K4 and this together with
the selected edges from U1 to the Ui , and the path P, Q, R is a graph homeomorphic to L, so
that G has a subcontraction to L.
If u04 = u05 , for any choice of P and for every set Ui , G has a subgraph homeomorphic from
K5 . The same is the case when in each Ui , the four end vertices u j coincide (Fig. 1.47). q

Graph Theory

33

Fig. 1.47

1.9 Distance and eccentricity

Definition: Distance between two vertices u and v of a connected graph G is the length of
the shortest path between u and v. It is denoted by d(u, v). Clearly, by convection, d(u, u) =
0. The shortest path is called a u v geodesic, or a u v distance path.
Theorem 1.21
vertex set.

The distance function for a connected graph is a metric defined on its

Proof Let G be a connected graph and d(u, v) be the distance function, with u and v being
vertices of G.
1. Clearly, d(u, v) 0, with d(u, u) = 0.
2. Also, d(u, v) = d(v, u), because the length of the shortest path from u to v is same as
the length of the shortest path from v to u.
3. For any vertices u, v, w in G, we have d(u, v) d(u, w) + d(w, v)
Thus d is a metric.

Note The distance between two vertices u and v in different components of a disconnected graph is infinite and this distance function does not induce a metric on the vertex set.

Neighbourhood: Let v be any vertex of a connected graph G. The ith neighbourhood

of v is Ni (v) = {u V : d(v, u) = i}. We set N0 (v) = {v} and denote N1 (v) simply by N(v), and
call N(v) as the neighbourhood of v, or the neighbours of v.
The s-ball at v is Bs (v) =

s
S

N j (v).

j=0

Eccentricity of a vertex: Let G be a connected graph. The eccentricity of a vertex v

in G is the distance of the vertex u farthest from v. It is denoted by e(v). That is, e(v) =
max{d(u, v) : u V }.
The minimum eccentricity is called the radius of G and the maximum eccentricity is
called the diameter of G. The radius is denoted by r and diameter by d .
Therefore, r = min{e(v) : v V } and d = max{e(v) : v V }.

34

Introduction

When the graph is to be mentioned, we use the notations, e(v|G), r(G) and d(G) for
eccentricity, radius and diameter respectively.
The vertices of minimum eccentricity in G are called the centres of G and the vertices of
maximum eccentricity are called the periphery of G. That is, the centre of G is C(G) = {v
V : e(v) = r} and periphery of G is P(G) = {v V ; e(v) = d}.
A vertex in C(G) is called a central vertex and a vertex in P(G) is called a peripheral
vertex. A graph having C(G) = V (G) is called a self-centered graph.
A longest geodesic (maximum among the shortest paths between any two vertices) of a
graph is called a diametral path of G. Note that this is not same as a longest path in G. The
diameter is the length of any diametral path. A graph G is said to be geodetic if any two of
its vertices are joined by a unique geodesic.
Theorem 1.22

The radius and diameter of a graph are related as r d 2r.

Proof Clearly, r d follows from the definition of r and d . Let u and v be the ends of a
diametral path and w be a central vertex. Then,
d = d(u, v) d(u, w) + d(w, v) r + r = 2r,
by using the triangle inequality of the metric and definition of r.
q
Example Consider the graph shown in Figure 1.48. The eccentricities are 5, 4, 3, 3, 3,
4, 5, 4, 5; r(G) = 3, d(G) = 5, C(G) = {v3 , v4 , v5 }, d(G) = {v1 , v7 , v9 }. Diametral paths are
v1 v2 v4 v5 v6 v7 , v1 v2 v3 v5 v6 v7 , v1 v2 v4 v5 v8 v9 and v1 v2 v3 v5 v8 v9 .

Fig. 1.48

Detours: A detour path between the vertices u and v in a graph G is the path of maximum
length between u and v. The length of detour path is called detour distance and is denoted
by (u, v).
The detour eccentricity of a vertex v is defined by (v) = max{ (v, u) : u V }.
The detour diameter, or detour number of G is defined as (G) = max{ (v) : v V } and
the detour radius is defined by r (G) = min{ (v) : v V }.
A longest path in G is called a detour path in G and its length is (G). The detour center
of G is C (G) = {v V : (v) = r (G)}.

Graph Theory

Theorem 1.23

35

If a connected graph G of order n has minimum degree , then

(G) = min(n 1, 2 ).

(1.23.1)

Proof

Clearly, (G) n 1. If (G) = n 1, then (1.23.1) is obviously true, as min (n

1, 2 ) = 2, or n 1 according as = 1, or = n 1. Now, let (G) < n 1. Assume the

contrary,
(G) < min(n 1, 2 ).

(1.23.2)

Let P = v0 v1 . . .v be a path of length = (G) and H be the subgraph induced by the

vertex set of P. Then H does not contain a spanning cycle, because otherwise some vertex
v V (G) V (H) will be adjacent to some vertex vi of V (H) giving a path of length + 1.
Thus v0 v
/ E , and for similar reasons v0 and v are adjacent only to vertices in H .
Let S = {vi V (H) : v0 vi E} and T = {vi V (H) : vi1 v E}.
Since H does not contain a spanning cycle, S T = . Also, v0
/ SUT , so that |S T| .
Also, |S| and |T | , therefore |S T| = |S| + |T| + = 2 . That is,
2 .

(1.23.3)

Therefore, 2 < n 1, so that 2 < n 1.

Thus (1.23.2) gives < min(n 1, 2 ) implying < 2 (because min (n 1, 2 ) = 2 as
2 < n 1), which contradicts (1.23.3) and so our assumption (1.23.2) is wrong.
Hence, (G) min(n 1, 2 ).
q

1.9 Exercises
1. If G is a graph with n vertices and m edges and if k is the smallest positive integer
such that k 2m/n, then prove G has a vertex of degree at least k.
2. If G is a graph with n vertices and if r vertices out of n have degree k and the others
have degree k + 1, then prove r = (k + 1)n 2m, where m is the number of edges in G.
3. If G is a k-regular graph (k being odd), prove that the number of edges in G is a
multiple of k.
4. If G is a graph with n vertices and exactly n 1 edges, then prove that G has either a
vertex of degree 1 or a vertex of degree zero.
5. If in a graph G,

n1
2 ,

then show that G is connected.

6. If a graph G is not connected, then show that G is connected.

7. Show that if a self-complementary graph contains pendent vertex, then it must have
at least another pendent vertex.

36

Introduction

8. Prove that any two connected graphs with n vertices, all of degree 2, are isomorphic.
9. Prove that if a connected graph G is decomposed into two subgraphs G1 and G2 , there
must be at least one vertex common between G1 and G2 .
10. Prove that a connected graph G remains connected after removing an edge e from G
if and only if e is in some cycle of G.
11. If the intersection of two paths is a disconnected graph, show that the union of the
two paths has at least one cycle.
12. Show that the order of a self-complementary graph is of the form 4n or 4n + 1, where
n is a positive integer.
13. Draw all the non-isomorphic self complementary graphs on four vertices.
14. Show that the bipartition of a connected graph is unique. Prove that a graph with
n(n > 1) vertices has at least (n 1)([n/2]) cycles.
15. Prove that a graph with n vertices must be connected if it has more than (n 1)(n
2)/2 edges.
16. Prove that a graph with n vertices (n > 2) cannot be bipartite if it has more than n2 /4
edges.
17. Prove that every graph with n vertices is isomorphic to a subgraph of Kn .
18. Prove that if a graph has more edges than vertices then it must possesses at least one
cycle.

2. Degree Sequences
The concept of degrees in graphs has provided a framework for the study of various structural properties of graphs and has therefore attracted the attention of many graph theorists.
Here we deliberate on the various criteria for a non-decreasing sequence of non-negative
integers to be a degree sequence of some graph.

2.1 Degree Sequences

Let di , 1 i n, be the degrees of the vertices vi of a graph in any order. The sequence [di]n1
is called the degree sequence of the graph. The non-negative sequence [di ]n1 is called the
degree sequence of the graph if it is the degree sequence of some graph, and the graph is
said to realise the sequence.
The set of distinct non-negative integers occurring in a degree sequence of a graph is
called its degree set. A set of non-negative integers is called a degree set if it is the degree
set of some graph, and the graph is said to realise the degree set.
Two graphs with the same degree sequence are said to be degree equivalent. In the graph
of Figure 2.1(a), the degree sequence is D = [1, 2, 3, 3, 3, 4] or D = [1 2 33 4] and its degree
set is {1, 2, 3, 4}, while the degree sequence of the graph in Figure 2.1(b) is [1, 1, 2, 3, 3]
and its degree set is {1, 2, 3}.

Fig. 2.1

If the degree sequence is arranged as the non-decreasing positive sequence d1n1 , d2n2 , . . .
(d1 < d2 < . . . < dk ), the sequence n1 , n2 , . . ., nk is called the frequency sequence of the
graph.
n
dk k ,

38

Degree Sequences

The two necessary conditions implied by Theorem 1.1 and Theorem 1.12 are not sufficient to ensure that a non-negative sequence is a degree sequence of a graph. To see this,
consider the sequence [1, 2, 3, 4, . . ., 4, n 1, n 1]. The sum of the degrees is clearly even
and = n 1. However, this is not a degree sequence, since there are two vertices with
degree n 1, and this requires that each of the two vertices is joined to all the other vertices,
and therefore 2. But the minimum number in the sequence is 1.
A degree sequence is perfect if no two of its elements are equal, that is, if the frequency
sequence is 1, 1, . . ., 1. A degree sequence is quasi-perfect if exactly two of its elements
are same.
Definition: Let D = [di ]n1 be a non-negative sequence and k be any integer 1 k n.
Let D0 = [di0 ]n1 be the sequence obtained from D by setting dk = 0 and di0 = di 1 for the
dk largest elements of D other than dk . Let Hk be the graph obtained on the vertex set
V = {v1 , v2 , . . ., vn } by joining vk to the dk vertices corresponding to the dk elements used
to obtain D0 . This operation of getting D0 and Hk is called laying off dk and D0 is called the
residual sequence, and Hk the subgraph obtained by laying off dk .
Example Let D = [2, 2, 3, 3, 4, 4]. Take d3 = 0. Then D0 = [2, 2, 0, 2, 3, 3]. The subgraph
Hk in this case is shown in Figure 2.2.

Fig. 2.2

2.2 Criteria for Degree Sequences

Havel [112] and Hakimi [99] independently obtained recursive necessary and sufficient
conditions for a degree sequence, in terms of laying off a largest integer in the sequence.
Wang and Kleitman [261] proved the necessary and sufficient conditions for arbitrary layoffs.
Theorem 2.1 A non-negative sequence is a degree sequence if and only if the residual
sequence obtained by laying off any non-zero element of the sequence is a degree sequence.

Proof
Sufficiency Let the non-negative sequence be [di ]n1 . Suppose dk is the non-zero element
laid off and the residual sequence [di0 ]n1 is a degree sequence. Then there exists a graph G0

Graph Theory

39

realising [di0 ]n1 in which vk has degree zero and some dk vertices, say vi j , 1 j dk have
degree di j 1. Now, by joining vk to these vertices we get a graph G with degree sequence
[di]n1 . (Observe that the subgraph obtained by such joining is precisely the subgraph Hk
obtained by laying off dk ).
Necessity We are given that there is a graph realising D = [di ]n1 . Let dk be the element to
be laid off. First, we claim there is a graph realising D in which vk is adjacent to all the
vertices in the set S of dk largest elements of D {dk }. If not, let G be a graph realising
D such that vk is adjacent to the maximum possible number of vertices in S. Then there
is a vertex vi in S to which vk is not adjacent and hence a vertex v j outside S to which vk
is adjacent (since d(vk ) = |S|). By definition of S, d j di . Therefore there is a vertex vh in
V {vk } adjacent to vi , but not adjacent to v j . Note that vh may be in S (Fig. 2.3).

Fig. 2.3

Construct a graph H from G by deleting the edges v j vk and vh vi and adding the edges v j vh
and vi vk . This operation does not change the degree sequence. Thus H is a graph realising
the given sequence, in which one more vertex, namely vi of S is adjacent to vk , than in G.
This contradicts the choice of G and establishes the claim.
To complete the proof, if G is a graph realising the given sequence and in which vk is
adjacent to all vertices of S, let G0 = G vk . Then G0 has the residual degree sequence
obtained by laying off dk .
q
Definition: Let the subgraph H on the vertices vi , v j , vr , vs of a multigraph G contain the
edges vi v j and vr vs . The operation of deleting these edges and introducing a pair of new
edges vi vs and v j vr , or vi vr and v j vs is called an elementary degree preserving transformation
(EDT), or simple exchange, or 2-switching, or elementary degree-invariant transformation.
Remarks

1. The result of an EDT is clearly a degree equivalent multigraph.

2. If an EDT is applied to a graph, the result will be a graph only if the latter pair of
edges (vi vs and v j vr ), or (vi vr and v j vs ) does not exist in G.

40

Degree Sequences

Theorem 2.2 (Havel, Hakimi) The non-negative integer sequence D = [di ]n1 is graphic
if and only if D0 is graphic, where D0 is the sequence (having n 1 elements) obtained from
D by deleting its largest element and subtracting 1 from its next largest elements.

Proof
Sufficiency Let D = [di ]n1 be the non-negative sequence with d1 d2 . . . dn . Let G0 be
the graph realising the sequence D0 . We add a new vertex adjacent to vertices in G0 having
degrees d2 1, . . ., d+1 1. Those di are the largest elements of D after itself. (But the
numbers d2 1, . . ., d+1 1 need not be the largest elements in D0 ).
Necessity Let G be a graph realising D = [di ]n1 , d1 d2 . . . dn . We produce a graph G0
realising D0 , where D0 is the sequence obtained from D by deleting the largest entry d1 and
subtracting 1 from d1 next largest entries.
Let w be a vertex of degree d1 in G and N(w) be the set of vertices which are adjacent to
w. Let S be the set of d1 number of vertices in G having the desired degrees d2 , . . ., dd1 +1 .
If N(w) = S, we can delete w to obtain G0 . Otherwise, some vertex of S is missing from
N(w). In this case, we modify G to increase |N(w) S| without changing the degree of any
vertex. Since |N(w) S| can increase at most d1 times, repeating this procedure converts an
arbitrary graph G that realises D, into a graph G that realises D, and has N(w) = S. From
G , we then delete w to obtain the desired graph G0 realising D0 .
If N(w) 6= S, let x S and z / S, so that wz is an edge and wx is not an edge, since
d(w) = d1 = |S|. By this choice of S, d(x) d(z) (Fig. 2.4).

Fig. 2.4

We would like to add wx and delete wz without changing their respective degrees. It
suffices to find a vertex y outside T = {x, z, w} such that yx is an edge, while yz is not. If
such a y exists, then we also delete xy and add zy. Let q be the number of copies of the
edge xz (0 or 1). Now x has d(x) q neighbours outside T , and z has d(z) 1 q neighbours
outside T . Since d(x) d(z), the desired y outside T exists and we can perform the EDT
(elementary degree preserving transformation or 2-switch).
q
Algorithm: The above recursive conditions give an algorithm to check whether a nonnegative sequence is a degree sequence and if so to construct a graph realising it.

Graph Theory

41

The algorithm starts with an empty graph on vertex set V = {v1 , v2 , . . ., vn } and at the kth
iteration generates a subgraph Hk of G by deleting (laying off) a vertex of maximum degree
in the residual sequence at that stage. If the given sequence is a degree sequence, we end
up with a null degree sequence (i.e., for each i, di = 0) and the graph realising the original
sequence is simply the sum of the subgraphs H j . If not, at some stage, one of the elements
of the residual sequence becomes negative, and the algorithm reports non-realisability of
the sequence.
An obvious modification of the algorithm, obtained by choosing an arbitrary vertex of
positive degree, gives the Wang-Kleitman algorithm for generating a graph with a given
degree sequence.
Remarks

1. There can be many non-isomorphic graphs with the same degree sequence. The
smallest example is the pair shown in Figure 2.5 on five vertices with the degree
sequence [2, 2, 2, 1, 1].

Fig. 2.5

The problem of generating all non-isomorphic graphs of given order and size involves the problem of graph isomorphism for which a good algorithm is not yet
known. So also is the problem of generating all non-isomorphic graphs with given
degree sequence. In fact, even the problem of finding the number of non-isomorphic
graphs with given order and size, or with given degree sequence (and several other
problems of similar nature) has not been satisfactorily solved.
2. The Wang-Kleitman algorithm is certainly more general than the Havel-Hakimi algorithm, as it can generate more number of non-isomorphic graphs with a given degree
sequence, because of the arbitrariness of the laid-off vertex. For example, not all the
five non-isomorphic graphs with the degree sequence [3, 3, 2, 2, 1, 1] can be generated by the Havel-Hakimi algorithm unlike the Wang-Kleitman algorithm.
3. Even the Wang-Kleithman algorithm cannot always generate all graphs with a given
degree sequence. For example, the graph G with degree sequence [3, 3, 3, 3, 2, 2, 2, 2,
1, 1, 1, 1] shown in Figure 2.6, cannot be generated by this algorithm. For
a. if we lay off a 3, it has to be laid off against the other 3s and will generate a
graph in which a vertex with degree 3 is adjacent to three other vertices with
degree 3,
b. if we lay off a 2 it will generate a graph with a vertex of degree 2 adjacent to
two vertices of degree 3,

42

Degree Sequences

c. if we lay off a one it will generate a graph in which a vertex of degree one is
adjacent to a vertex of degree 3. None of these cases is realised in the given
graph G.

Fig. 2.6

However, there are other methods of generating all graphs realising a degree sequence D from any one graph realising D based on a theorem by Hakimi [98]. But
those will also be inefficient unless some efficient isomorphism testing is developed.
4. The graphs in Figure 2.5 show that the same degree sequence may be realised by a
connected as well as a disconnected graph. Such degree sequences are called potentially connected, where as a degree sequence D such that every graph realising D is
connected is called a forcibly connected degree sequence.
Definition: If P is a graph property, and D = [di ]n1 is a degree sequence, then D is said to
be potentially-P, if at least one graph realising D is a P-graph, and it is said to be forcibly-P
if every graph realising it is a P-graph.
Theorem 2.3 (Hakimi) If G1 and G2 are degree equivalent graphs, then one can be
obtained from the other by a finite sequence of EDTs.

Proof Superimpose G1 and G2 such that each vertex of G2 coincides with a vertex of G1
with the same degree. Imagine the edges of G1 are coloured blue and the edges of G2 are
coloured red. Then in the superimposed multigraph H , the number of blue edges incident
equals the number of red edges incident at every vertex. We refer to this as blue-red parity.
If there is a blue edge vi v j and a red edge vi v j in H , we call it a blue-red parallel pair.
Let K be the graph obtained from H by deleting all such parallel pairs. Then K is the null
graph if and only if G1 and G2 are label-isomorphic in H and hence originally isomorphic.
If this is not the case, we show that we can create more parallel pairs by a sequence of
EDTs and delete them till the final resultant graph is null. This will prove the theorem.
Let B and R denote the sets of blue and red edges in K . If vi v j B, we show that we
can produce a parallel pair at vi v j , so that the pair can be deleted. This would establish the
claim made above.
Now, by construction, there is a blue-red degree parity at every vertex of K . So there are
red edges vi vk , v j vr in K . If vk 6= vr (Fig. 2.7(a)) an EDT in G2 switching the red edges to
vi v j , vk vr produces a blue-red parallel at vi v j .

Graph Theory

43

Fig. 2.7

If vk = vr , again by degree parity, at vk there are at least two blue edges. Let vk vs be one
such blue edge. Then vs is distinct from both vi and v j , for otherwise, there is a blue-red
parallel pair vi vk or v j vr . Then there is another red edge vs vt , vt distinct from vi or v j .
Let vt 6= vi . The two subcases vt = v j and vt 6= v j are shown in Figure 2.7(b) and (c). In the
case of (b), one EDT of G2 switching vi vk and vs vt to positions vi v j and vs vk produces a bluered pair at vi v j and vk vs . In the case of (c), one EDT of G2 switching vi vk and vt vs to positions
vs vk and vt vi produces a blue-red parallel pair at vk vs (which can be deleted). Another EDT
of G2 switching the blue-red pair vt vi and v j vk to positions vi v j and vs vk produces a blue-red
pair vi v j .
Since in both cases we get a blue-red pair at vi v j position, our claim is established and
the proof of the theorem is complete.
q
In the related context of a (0, 1) matrix A (that is, a matrix A whose elements
are 0s or 1s), Ryser [227] defined an
as a transformation of the elements
of
 interchange



Remarks

A that changes a minor of type A1 =

1 0
0 1

into a minor of the type A1 =

0 1
, or vice
1 0

versa and proved an interchange theorem which can be interpreted as EDT theorem for
bipartite graphs and digraphs.

The next result is a combinatorial characterisation of degree sequences, due to Erdos

and Gallai [73]. Several proofs of the criterion exist; the first proof given here is due to
Choudam [58] and the second one is due to Tripathi et al [246].
Theorem 2.4 (Erdos-Gallai) A non-increasing sequence [di ]n1 of non-negative integers
is a degree sequence if and only if D = [di]n1 is even and the inequality

44

Degree Sequences
k

di k(k 1) +
i=1

(2.4.1)

min(di , k)

i=k+1

is satisfied for each integer k, 1 k n.

Proof
Necessity

Evidently di is even. Let U denote the subset of vertices with the k highest
i=1
k

degrees in D. Then the sum s = di can be split as s1 + s2 , where s1 is the contribution

i=1

to s from edges joining vertices in U , each edge contributing 2 to the sum, and s2 is the
contribution to s from the edges between vertices in U and U (where U = V U ), each edge
contributing 1 to the sum (Fig. 2.8).
s1 is clearly bounded above by the degree sum of a complete graph on k-vertices, i.e.,
k(k 1). Also, each vertex vi of U can be joined to at most min (di , k) vertices of U , so that
n
s2 is bounded above by min(di , k). Together, we get (2.4.1).
i=k+1

Fig. 2.8

Sufficiency

We induct on the sum s = di and use the obvious inequality

i=1

min(a, b) 1 min(a 1, b),

(2.4.2)

for positive integers a and b.

For s = 2, clearly K2 (n 2)K1 realises the only sequence [1, 1, 0, 0, . . . 0] or [120n2 ]
satisfying the conditions (2.4.1).
As induction hypothesis, let all non-increasing sequences of non-negative integers with
even sum at most s 2 and satisfying (2.4.1) be degree sequences.
Let D = [di ]n1 be a sequence with sum s and satisfying (2.4.1). We produce a new nonincreasing sequence D0 of non-negative integers by subtracting one each from two positive
terms of D and verify that D0 satisfies the hypothesis of the theorem. Since the trailing

Graph Theory

45

zeros in the non-increasing sequences of non-negative integers do not essentially affect the
argument, there is no loss of generality in assuming that dn > 0, and we assume this to
simplify the expression.
To define D0 , let t be the smallest integer ( 1) such that dt > dt+1 . That is, let D be
d1 = d2 = . . . = dt > dt+1 dt+2 . . . dn > 0.
If D is regular (that is, di = d > 0, for all i) then let t be n 1.

di ,
dt 1,
Then di0 =

dn 1,

f or 1 i t 1 and t + 1 i n 1 ,
f or i = t ,
f or i = n .
n

Clearly, D0 is a non-increasing sequence of non-negative integers and di0 = s 2 is

i=1
even.
We verify that D0 satisfies (2.4.1) by considering several cases depending on the relative
position of k and the magnitudes of dk and dn .
Case I

i=1

i=1

Let k = n. Therefore, di0 = di 2 n(n 1) 2 < n(n 1) = RHS of (2.4.1)

for D0 .
Case II Let t k n 1.
k

i=1

i=1

i=k+1

Then di0 = di 1 k(k 1) + min(di , k) 1 (since D satisfies (2.4.1))

n1

= k(k 1) + min(di0 , k) + min(dn , k) 1

i=k+1
n1

k(k 1) + min(di0 , k) + min(dn 1, k)

by (2.4.2)

i=k+1
n1

= k(k 1) + min(di0 , k) + min(dn0 , k)

i=k+1

i=1

i=k+1

Therefore, di0 k(k 1) + min(di0 , k).

Case III Let k t 1.
Subcase III.1

Assume dk k 1.

i=1

i=k+1

Then di0 = kdk k(k 1) k(k 1) + min(di0 , k),

since the second term is non-negative.
Subcase III.2

Every d j = k, 1 j k. We first observe that dk+2 + . . . + dn 2.

46

Degree Sequences

This is obvious if k + 2 n 1, because dn > 0 gives dn 1 and dn1 1. When k + 2 = n,

we have k = n 2. As k t 1, t k + 1 = n 2 + 1 = n 1. Since t > n 1 is not possible,
t = n 1.
The sequence D is [n 2, n 2, . . ., n 2, dn ], or [(n 2)n1 dn ]. Then s = (n 1) (n 2)+dn .
Since s is even, dn is even and hence dn 2. Thus, dk+2 + . . .+ dn 2.
Therefore, dk+2 + . . . + dn 2 0.
Now,
k

di0

i=1

di = k.k = k2 = k2 k + k
i=1

k2 k + dk+1 , (because k t 1, and d1 = . . . = dt1 = dt ,

so if dt1 = k, then dt = k, and if dk = k, dk+1 = k).

k

i=1

i=k+1

Thus, di0 k2 k + dk+1 + (dk+2 + . . . + dn 2) = k(k 1) + min(di , k) 2,

(because min (dk+1 , k) = dk+1, min (dk+2 , k) = k = dk+2, . . ., min (dt , k) = k = dt , . . ., min
(dt+1 , k) = dt+1 (as dt+1 < dt = k), . . ., min (dn , k) = dn (as dn < dt = k)).
k

i=1

i=k+1
i6=t, n

Hence, di0 k(k 1) + min(di , k) + min(dt , k) + min(dn , k) 2

n

= k(k 1) + min(di0 , k) + min(dt0 + 1, k) + min(dn0 + 1, k) 2

i=k+1
i6=t, n
n

k(k 1) + min(di0 , k) + min(dt0 , k) + 1 + min(dn0 , k) + 1 2

i=k+1
i6=t, n
n

= k(k 1) + min(di0 , k).

i=k+1

Subcase III.3 Let dk k + 1.

i. Let dn k + 1.
k

i=1

i=1

i=k+1

Then di0 = di k(k 1) + min(di , k)

n

(since D satisfies (2.4.1))

= k(k 1) + min(di , k) + min(dt , k) + min(dn , k)

k+1
i6=t, n

Graph Theory

47
n

= k(k 1) + min(di0 , k) + min(dt 1, k) + min(dn 1, k),

k+1
i6=t, n

(because min(dt , k) = min(dt 1, k) = k, min(dn , k) = min(dn 1, k) = k, as dt k +

1, dn k + 1 implies that dt 1 k, dn 1 k).
k

i=1

k+1
i6=t, n

So, di0 k(k 1) + min(di0 , k) + min(dt0 , k) + min(dn0 , k)

n

= k(k 1) + min(di0 , k).

i=k+1

ii. Let dn k and let r be the smallest integer such that dt+r+1 k. We verify that in
(2.4.1), D can not attain equality for such a choice of k. For, with equality, we have
k

t+r

i=1

k+1

t+r+1

di = kdk = k(k 1) + min(di , k) + min(di , k)

n

= k(k 1) + (t + r k)k + di ,
t+r+1

because min (di , k) =

k,

f or i = k + 1, ..., t + r as di k + 1,

di ,

f or i = t + r + 1, ..., n as di k .

So, kdk = k(t + r 1) + di .

t+r+1

)
1 n
Then di = (k + 1)dk = (k + 1) (t + r 1) + di , (using dk from above)
k t+r+1
i=1
(

k+1

= (k + 1)(t + r 1) +

n
k+1 n
di > (k + 1)(t + r 1) + di

k t+r+1
t+r+1
n

= (k + 1)k (k + 1)k + (k + 1)(t + r 1) +

di

t+r+1
n

= (k + 1)k + (k + 1)(t + r k 1) +

t+r+1

= (k + 1)k +

t+r+1

min (di , k + 1),

t+r

di = (k + 1)k +

(k + 1) +
k+1

t+r+1

di

48

Degree Sequences

because

min (di , k + 1) = k + 1 for i = k + 1, . . ., t + r, and

min (di , k + 1) = di ,
k+1

So,

for i = t + r + 1, . . ., n.

di > k(k + 1) + min (di , k + 1).

i=1

k+1
k+1

di > k(k + 1) + (k + 1) + min (di, k + 1),

Therefore,

i=1

k+2

which is a contradiction to (2.4.1), for D for k + 1. Hence D has strict inequality for k.
Therefore,

i=1

i=1

k+1

di0 = di < k(k 1) + min (di, k).

i=1

i=1

k+1

Thus, di0 = di k(k 1) + min (di , k) 1

n1

= k(k 1) +

min (di , k) + min (dt , k) + min (dn , k) 1

i=k+1
i6=t
n1

k(k 1) +

min (di0 , k) + min (dt 1, k) + min (dn 1, k),

i=k+1
i6=t

as min (dn , k) 1 min (dn 1, k), min(dt , k) = k (since dt k + 1), min(dt 1, k) = k

(since dt 1 k).
Therefore,

i=1

k+1

di0 k(k 1) + min (di0 , k).

Hence in all cases D0 satisfies (2.4.1).

Therefore by induction hypothesis, there is a graph G0 realising D0 . If vt vn / E(G0 ), then
G + vt vn gives a realisation G of D. If vt vn E(G0 ), since d(vt |G0 ) = dt 1 n 2, there is
a vertex vr such that vr vt
/ E(G0 ). Also, since d(vr |G0 ) > d(vn |G0 ), there is a vertex vs such
0
that vs vn / E(G ). Making an EDT exchanging the edge pair vt vn , vr vs for the edge pair vt vr ,
vs vn , we get a realisation G00 of D0 with vt vn
/ E(G00 ). Then G00 + vt vn realises D.
0

Second Proof of Sufficiency (Tripathi et al.) Let a subrealisation of a non-increasing

sequence [d1, d1 , . . ., dn] be a graph with vertices v1 , v1 , . . ., vn such that d(vi ) = di for 1 i n,
where d(vi ) denotes the degree of vi . Given a sequence [d1, d1 , . . ., dn] with an even sum that
satisfies (2.4.1), we construct a realisation through successive subrealisations. The initial
subrealisation has n vertices and no edges.

Graph Theory

49

In a subrealisation, the critical index r is the largest index such that d(vi ) = di for 1
i < r. Initially, r = 1 unless the sequence is all 0, in which case the process is complete.
While r n, we obtain a new subrealisation with smaller deficiency dr d(vr ) at vertex vr
while not changing the degree of any vertex vi with i < r (the degree sequence increases
lexicograpically). The process can only stop when the subrealisation of d .
Let S = {vr+1 , . . ., vn }. We maintain the condition that S is an independent set, which certainly holds initially. Write ui v j when vi v j E(G); otherwise, vi 6 v j
Case 0 vr 6 vi for some vertex vi such that d(vi ) < di . Add the edge ur vi .
Case 1 vr 6 vi for some i with i < r. Since d(vi ) = di dr > d(vr ), there exists u N(ui )
(N(vr ) {vr }), where N(z) = {y : z y}. If dr d(vr ) 2, then replace uvi with {uvr , vi vr }.
If dr d(vr ) = 1, then since di d(vi ) is even there is an index k with k > r such that
d(vk ) < dk . Case 0 applies unless vr vk ; replace {vr vk , uvi } with {uvr , vi ur }.
Case 2 v1 , . . ., vr1 N(vr ) and d(vk ) 6= min{r, dk } for some k with k > r. In a subrealisation,
d(vk ) dk . Since S is independent, d(vk ) r. Hence d(vk ) < min{r, dk }, and case 0 applies
unless uk vr . Since d(vk ) < r, there exists i with i < r such that uk 6 vi . Since d(vi ) > d(vr ),
there exists u N(vi ) (N(vr ) {ur}). Replace uvi with {uvr , vi vk }.
Case 3 v1 , . . ., vr1 6 N(vr ) and vi vi for some i and j with i < j < r. Case 1 applies
unless vi , v j N(vr ). Since d(vi ) d(vi ) > d(vr ), there exists u N(vi ) (N(vr ) {vr }) and
w N(v j )(N(vr ) {vr }) (possibly u = w). Since u, w 6 N(vr ), Case 1 applies unless u, w S.
Replace {vi v j , uvr } with {uvr , v, vr}.

If none of these case apply, then v1 , . . ., vr are pairwise adjacent, and d(vk ) = min{r, dk }
for k > r. Since S is independent, ri=1 d(vi ) = r(r 1) + nk=r+1 min{r, dk}. By (2.4.1), ri=1 d1
is bounded by the right side. Hence we have already eliminated the deficiency at vertex r.
Increase r by 1 and continue.
q
Tripathi and Vijay [245] have shown that the Erdos-Gallai condition characterising graphical degree sequences of length n needs to be checked only for as many k as there are distinct
terms in the sequence and not for all k, 1 k n.

2.3 Degree Set of a Graph

The set of distinct non-negative integers occurring in a degree sequence of a graph is called
its degree set. For example, let the degree sequence be D = [2, 2, 3, 3, 4, 4], then degree set
is {2, 3, 4}. A set of distinct non-negative integers is called a degree set if it is the degree
set of some graph and the graph is said to realise the degree set.
Let S = {d1 , d2 , . . ., dk } be the set of distinct non-negative integers. Clearly, S is the
degree set as the graph
G = Kd1 +1 Kd2 +1 . . . Kdk +1 ,

50

Degree Sequences

realises S. This graph has d1 + d2 + . . .+ dk + k vertices.

Example Let S = {1, 3, 4}. Then G = K2 K4 K5 (Fig. 2.9).

Fig. 2.9

The following result is due to Kapoor, Polimeni and Wall [126].

Theorem 2.5 Any set S of distinct positive integers is the degree set of a connected
graph and the minimum order of such a graph is M + 1, where M is the maximum integer
in the set S.

Proof Let S be a degree set and n0 (S) denote the minimum order of a graph G realising
S. As M is the maximum integer in S, therefore in G there is a vertex adjacent to M other
vertices, i.e., n0 (S) M + 1. Now, if there exists a graph of order M + 1 with S as degree
set, then n0 (S) = M + 1. The existence of such a graph is established by induction on the
number of elements p of S.
Let S = {a1 , a2 , . . ., a p } with a1 < a2 < . . . < a p .
For p = 1, the complete graph Ka1 +1 realises {a1 } as degree set.
For p = 2, we have S = {a1 , a2 }. Let G = Ka1 V K a2 a1 +1 (join of two graphs). Here every
vertex of Ka1 has degree a2 and every other vertex has degree a1 and therefore G realises
{a1, a2 } (Fig. 2.10(a)).
For p = 3, we have S = {a1 , a2 , a3 }. Then G = Ka1 V (K a3 a2 H), where H is the graph realising the degree set {a2 a1 } with a2 a1 + 1 vertices, realises {a1 , a2 , a3 } (Fig. 2.10 (b)).
(Note that d(u) = a1 1 + a3 a2 + a2 a1 + 1 = a3 , d(v) = a1 , d(w) = a2 a2 + a1 = a2 ).

Graph Theory

51

Fig. 2.10

Let every set with h positive integers, 1 h k, be the degree set. Let S1 = {b1 , b2 , . . ., bk+1}
be a (k + 1) set of positive integers arranged in increasing order. By induction hypothesis,
there is a graph H realising the degree set {b2 b1 , b3 b1 , . . ., bk b1 } with order bk b1 +1.
The graph G = Kb1 V (K bk+1 bk H), with order bk+1 + 1 realises S1 (Fig. 2.10 (c)). Clearly
by construction, all these graphs are connected.
Hence the result follows by induction.
q
Note that d(ui ) = b1 1 +bk+1 bk +bk b1 +1 = b
k +1, d(vi ) = b1 , d(wi ) = bi+1 b1 +b1 =
bi+1, that is d(w1 ) = b2 , d(w2 ) = b3 , . . ., d wbk b1 +1 = bk b1 + b1 = bk . Some results on

degree sets in bipartite and tripartite graphs can be seen in [262].

2.4 New Criterion

We have the following notations. Let D = [di ]n1 be a non-decreasing sequence of nonnegative integers with 0 di n 1 for all i. Let n p1 be the greatest integer, n p1 p2 ,
k

the second greatest integer and n pr , the kth greatest integer in D, 1 pr n (r 1).
r=1

Let the number of times the kth greatest integer appears in D be denoted by ak . Also, we
take


k
k
tk = n n pr = pr , 1 pr n (r 1) and jk = 1, 2, . . ., pk+1.
r=1

r=1

The following result due to Pirzada and YinJian [208] is another criterion for a nonnegative sequence of integers in non-decreasing order to be the degree sequence of some
graph.
n

Theorem 2.6 A non-decreasing sequence [di ]n1 of non-negative integers, where di is

i=1

even and 0 di n 1 for all i, is a degree sequence of a graph if and only if

tk + jk 1

i=1

di { jk + (k m)} am
m=1

(2.6.1)

52

Degree Sequences
k

for all tk + jk 1 + am n.
m=1

In the above criterion, the inequalities (2.6.1) are to be checked only for tk + jk

Note
k

1 + am n (but not for greater than n).

m=1

We now illustrate the theorem with the help of the following examples.
Example 1

Let D = [1, 2, 2, 4, 6, 6, 6, 7, 8, 8].

Here, n = 10, a1 = 2, a2 = 1, a3 = 3, a4 = 1, p1 = 2, p2 = 1, p3 = 1, p4 = 2, so t1 = 2, t2 = 3,
t3 = 4, t4 = 6.
Also, j1 = 1, j2 = 1, j3 = 1, 2.
Now, for j1 = 1,

t1 + j11

2+11

i=1

i=1

di = di = di = 1 + 2 = 3,

i=1
k

m=1

m=1

and [ jk + (k m)] am = [ j1 + (1 m)]am = j1 a1 = 2.

So inequalities (2.6.1) hold.
For j2 = 1,

t2 + j21

i=1

3+11

i=1

i=1

di = di = di = 5

m=1

m=1

and [ jk + (k m)] am = [ j2 + (2 m)]am = 2a1 + a2 = 4 + 1 = 5.

So inequalities (2.6.1) hold.
For j3 = 1,

t3 + j31

i=1

4+11

i=1

i=1

di = di = di = 9

m=1

m=1

and [ j3 + (3 m)] am = [1 + (3 m)] am = 3a1 + 2a2 + a3 = 6 + 2 + 3 = 11.

Since the inequalities (2.6.1) do not hold (as 9 > 11 is not true), D is not the degree
sequence.
Example 2

Let D = [1, 2, 3, 4, 5, 6, 6, 7, 8, 8].

Here, n = 10, a1 = 2, a2 = 1, a3 = 2, a4 = 1, p1 = 2, p2 = 1, p3 = 1, p4 = 1, p5 = 1. So
t1 = 2, t2 = 3, t3 = 4, t4 = 5.
Also, j1 = 1, j2 = 1, j3 = 1, j4 = 1.

Graph Theory

53

For j1 = 1,

t1 + j11

i=1

2+11

i=1

i=1

di = di = di = 3,

and [ j1 + (1 m)] am = a1 = 2.
m=1

Obviously the inequalities (2.6.1) hold.

t2 + j21

For j2 = 1,

i=1

3+11

i=1

i=1

di = di = di = 6

m=1

m=1

and [ j2 + (2 m)] am = [1 + (2 m)] am = 2a1 + a2 = 4 + 1 = 5 .

Here again the inequalities (2.6.1) hold.
For j3 = 1,

t3 + j31

i=1

4+11

i=1

i=1

di = di = di = 10

m=1

m=1

and [ j3 + (3 m)] am = [1 + (3 m)] am = 3a1 + 2a2 + a3 = 6 + 2 + 2 = 10 .

Therefore the inequalities (2.6.1) hold.
For j4 = 1, t4 + j4 1 = 5 + 1 1 = 5 and a1 + a2 + a3 + a4 = 2 + 1 + 2 + 1 = 6, therefore
4

t4 + j4 1 + am = 5 + 6 = 11 > 10 and no further verification of the inequalities is to be

m=1

done.
Hence D is the degree sequence.

2.5 Equivalence of Seven Criteria

We list the seven criteria for integer sequences to be graphic.
A. The Ryser Criterion (Bondy and Murty [36] and Ryser [227]) A sequence [a1 , . . .,
a p ; b1 , . . ., bn ] is called bipartite-graphic if and only if there is a simple bipartite
graph such that one component has degree sequence [a1, . . ., a p ] and the other one
has [b1 , . . ., bn ]. Define f = max{i : di i} and d1 = di + 1 if i h f i (= {1, . . ., f }) and
d1 = di otherwise. The criterion can be stated as follows.

The integer sequence [d1 , ..., dn; d1 , ..., dn]is bipartite-graphic. (A)
B. The Berge Criterion (Berge [23]) Define [d1 , . . ., dn ] as follows: For i hni, di is the
ith column sum of the (0, 1) matrix, which has for each k and dk leading terms in row

54

Degree Sequences
k equal to 1 except for the (k, k)th term that is 0 and also the remaining entries are 0.
If d1 = 3, d2 = 2, d3 = 2, d4 = 2, d5 = 1, then d1 = 4, d2 = 3, d3 = 2, d4 = 1, d5 = 0, and
the (0, 1) matrix becomes

0
1
1
1
1

1
0
1
1
0

1
1
0
0
0

1
0
0
0
0

0
0
0
0
0

The criterion is
k

i=1

i=1

di di for each k hni.

(B)

C. The Erdos-Gallai Criterion. (Bondy and Murty [36])

k

i=1

j=k+1

di (k)(k 1) + min{k, d j } for each k hni.

(C)

D. The Fulkerson-Hoffman-McAndrew Criterion (Fulkerson[83] and Grunbaum [92)

k

di (k)(n m 1) +
i=1

di for each k hni, m 0 and k + m n.

(D)

i=nm+1

E. The Bollobas Criterion (Bollabas[29]))

k

i=1

j=k+1

i=1

di di + min{d j , k 1} for each k hni.

(E)

F. The Grunbaum Criterion (Grunbaum [92]).

k

i=1

i=k+1

max{k 1, di } (k)(k 1) + di for each k hni.

(F)

G. The Hasselbarth Criterion (Hasselbarth [111]) Define [di0 , . . ., dn0 ] as follows. For
i hni, di0 is the ith column sum of the (0, 1)-matrix in which the di leading terms in
row i are 1s and the remaining entries are 0s. The criterion is
k

i=1

i=1

di (di 1) for each k h f i,

with f = max{i : di i}.

(G)

Graph Theory

55

The following result due to Sierksma and Hoogeveen [235] gives the equivalence among
the above seven criteria.
Theorem 2.7 (Sierksma and Hoogeveen [235]) Let [d1 , . . ., dn ] be a positive integer sequence with even sum. Then each of the criteria (A) (G) is equivalent to the statement
that [d1 , . . ., dn ] is graphic.

Proof

Refer to Ryser [227].

2.6 Signed Graphs

A signed graph is a graph in which every edge is labelled with a+ or a. An edge uv
labelled with a+ is called a positive edge, and is denoted by uv+ . An edge uv labelled
with a is called a negative edge, and is denoted by uv . In a signed graph G(V, E), the
positive degree of a vertex u is deg+ (u) = |{uv : uv+ E}|, the negative degree of a vertex u
is deg (u) = |{uv : uv E}|, the signed degree of u is sdeg(u) = deg+ (u) deg (u) and the
degree of u is deg(u) = deg+ (u) + deg (u). An edge uv labelled with a+ is called a positive
edge, and is denoted by uv+ . An edge uv labelled with a is called a negative edge, and
is denoted by uv .
An integral sequence [di ]n1 is the signed degree sequence of a signed graph G = (V, E)
with V = {v1 , v2 , . . ., vn } if s deg(vi ) = di , for 1 i n.
Chartrand et al. [50] have given the characterisation of signed degree sequences of
signed paths, signed stars, signed double stars and complete signed graphs. An integral
sequence is s-graphical if it is the signed degree sequence of a signed graph. An integral
sequence [di ]n1 is standard if n 1 d1 d2 . . . dn and d1 |dn |.
The following lemma shows that a signed degree sequence can be modified and rearranged into an equivalent standard form.
Lemma 2.1 If [di ]n1 is the signed degree sequence of a signed graph G, then [di ]n1 is the
signed degree sequence of the signed graph G0 obtained from G by interchanging positive
edges with negative edges.
The following necessary and sufficient condition under which an integral sequence is
s-graphical is due to Chartrand et al. [50].

A standard integral sequence [di ]n1 is s-graphical if and only if the sequence
[d2 1, dd1 +s+1 1, dd1 +s+2 , . . ., dns , dns+1 + 1, . . ., dn + 1] is s-graphical for some 0 s
(n 1 d1)/2.

Theorem 2.8

Remark We note that Hakimis theorem for degree sequences is a case of Theorem 2.8
by taking s = 0. This leads to an efficient algorithm for recognising the degree sequences
of a graph. But the wide degree of latitude for choosing s in Theorem 2.8 makes it harder
to devise an efficient algorithm implementation.

56

Degree Sequences

The following result due to Yan et al. [271] provides a good choice for parameter s
in Theorem 2.8. It leads to a polynomial time algorithm for recognising signed degree
sequences.
A standard sequence D = [di ]n1 is s-graphical if and only if Dm = [d2
1, dd1 +m+1 1, . . ., dd1 +m+2 , . . ., dnm , dnm+1 + 1 . . ., dn + 1] is s-graphical, where m is the
maximum non-negative integer such that dd1 +m+1 > dnm+1 .

Theorem 2.9

Proof Let D be the signed degree sequence of a signed graph G = (V, E) with V =
{v1 , v2 , . . ., vn } and sdeg(vi ) = di , for 1 i n. For each s, 0 s (n 1 d1 )/2, consider
the sequence
Ds = [d2 1, . . ., dd1 +s+1 1, dd1+s+2 , . . ., dns , dns+1 + 1, . . ., dn + 1].

By Theorem 2.8, Ds is s-graphical for some s. We may choose s such that |s m| is

minimum. Suppose G0 = (V 0 , E 0 ) is a signed graph with V 0 = {v2 , v3 , . . ., vn } whose signed
degree sequence is Ds .
If s < m, then da > db by the choice of m, where a = d1 + s + 2 and b = n s. Since da > db ,
there exists some vertex vk of G0 different from va and vb and satisfies one of the following
conditions.

i. va v+
k is a positive edge and vb vk is a negative edge.

ii. va v+
k is a positive edge and vb is not adjacent to vk
iii. va is not adjacent to vk and vb v
k is a negative edge
+

0
0
For (i), remove va v+
k and vb vk to G , and for (ii), remove va vk from G and add a new
+

0
0
positive edge vb vk to G and for (iii), remove vb vk from G and a new negative edge va v
k to
G0 . These modifications result in a signed graph G00 whose signed degree sequence Ds+1 .
This contradicts the minimality of |s m| .
If s > m, then dd1 +s+1 = dns+1 , and therefore, dd1 +s+1 1 < dns+1 1. An argument
similar to the above leads to a contradiction in the choice of s. Therefore, s = m and Dm is
s-graphical.
Conversely, suppose Dm is the signed degree sequence of a signed graph G0 = (V 0 , E 0 )
in which V 0 = {v2 , v3 , . . ., vn }. If G is the signed graph obtained from G0 by adding a new

vertex v1 and new positive edges v1 v+

i for 2 i d1 + m + 1 and new negative edges v1 v j
for n m + 1 j n, then D is the signed degree sequence of G.
q

In a signed graph G = (V, E) with |V | = n, |E| = m, we denote by m+ and m respectively,

the numbers of positive edges and negative edges of G. Further, n+ , n0 and n denote
respectively, the numbers of vertices with positive, zero and negative signed degrees.
The following result is due to Chartrand et al. [50].
Lemma 2.2

If G = (V, E) is a signed graph with |V | = n, |E| = m, then k = s deg(v)

vV

2m(mod4), m+ = 41 (2m + k) and m = 14 (2m k).

Graph Theory

57

The next result is due to Yan et al [271].

Lemma 2.3

For any signed graph G = (V, E) without isolated vertices, |sdeg(v)| +

vV

2n0 2m.

Proof First, each |sdeg(v)| = | deg+ (v) deg (v)| deg+ (v)+ deg (v). Since G has no
isolated vertices, 2 deg+ (v)+ deg (v) when sdeg(v) = 0. Thus,
|s deg (v)| + 2n0 (deg+ (v)+deg (v)| = 2m+ + 2m = 2m.
vV

vV

Lemma 2.4

For any connected signed graph G = (V, E), |s deg (v)|+2

vV

|sdeg(v)|

sdeg(v)<0

6m + 4 4 4n+ 4n0 , where = 1 if n+n > 0 and = 0 otherwise.

Proof Consider the subgraph G0 = (V 0 , E 0 ) of G induced by those edges incident to vertices with non-negative signed degrees. We have,
|sdeg (v)| 2 (number of positive edges in G0 )

sdeg(v)>0

(number of negative edges in G0 ) 3m+ |E 0 |.

Since G is connected, each component of G0 contains at least one vertex of negative
signed degree except for the case of G0 = G.
Therefore, n+ + n0 1 + |E 0 |. Thus,

|s deg(v)| + n+ + n0 1 + 3m = 3

sdeg(v).0

Hence, |s deg(v)| + 2
vV

!
1
1
m + s deg(v) .
2
4 vV

|s deg(v)| 6m + 4 4 4n+ 4n0 .

sdeg(v)<0

For any integer k, k copies of vi v j means k copies of positive edges vi v+j if k > 0, no edges
if k = 0 and k copies of negative edges vi vj if k < 0. The next result for signed graphs with
loops or multiple edges is due to Yan et al. [271].
Theorem 2.10
n

An integral sequence [di ]n1 is the signed degree sequence of a signed if

and only if di is even.

i=1

Proof

The necessity follows from Lemma 2.2.

Sufficiency

Let di be even. Then the number of odd terms is even, say di = 2ei + 1 for
i=1

1 i 2k and di = 2ei for 2k + 1 i p. Then [d1 , d2 , . . ., dn ] is the signed degree sequence

58

Degree Sequences

of the signed graph with vertex set {v1 , v2 , . . ., vn } and edge set {d3 =
v1 v2 } {d2 + d3

1
2

di copies of v2 v3 } {d1 + d3
i=1

1
2

1
2

di copies of
i=1

di copies of v1 v3 } {di copies of

i=1

v3 vi : 4 i n}.

Various results on signed degrees in signed graphs can be found in [259], [263], [264]
and [266].

2.7 Exercises
1. Verify whether or not the following sequences are degree sequences.
a. [ 1, 1, 1, 2, 3, 4, 5, 6, 7 ],
b. [ 1, 1, 1, 2, 2, 2 ],
c. [ 4, 4, 4, 4, 4, 4 ],
d. [ 2, 2, 2, 2, 4, 4 ].
2. Show that there is no perfect degree sequence.
3. What conditions on n and k will ensure that kn is a degree sequence?
4. Give an example of a graph that can not be generated by the Wang-Kleitman algorithm.
5. Draw the five non isomorphic graphs with degree sequence [3, 3, 2, 2, 1, 1].
6. Show that a graph and its complement have the same frequency sequence.
7. Construct a graph with a degree sequence [3, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1] by using
Havel-Hakimi algorithm.

3. Eulerian and Hamiltonian Graphs

There are many games and puzzles which can be analysed by graph theoretic concepts.
In fact, the two early discoveries which led to the existence of graphs arose from puzzles, namely, the Konigsberg Bridge Problem and Hamiltonian Game, and these puzzles
also resulted in the special types of graphs, now called Eulerian graphs and Hamiltonian
graphs. Due to the rich structure of these graphs, they find wide use both in research and
application.

3.1 Euler Graphs

A closed walk in a graph G containing all the edges of G is called an Euler line in G. A
graph containing an Euler line is called an Euler graph.
We know that a walk is always connected. Since the Euler line (which is a walk) contains
all the edges of the graph, an Euler graph is connected except for any isolated vertices the
graph may contain. As isolated vertices do not contribute anything to the understanding
of an Euler graph, it is assumed now onwards that Euler graphs do not have any isolated
vertices and are thus connected.
Example Consider the graph shown in Figure 3.1. Clearly, v1 e1 v2 e2 v3 e3 v4 e4 v5 e5 v3 v6
e7 v1 in (a) is an Euler line, whereas the graph shown in (b) is non-Eulerian.

Fig. 3.1

60

Eulerian and Hamiltonian Graphs

The following theorem due to Euler [74] characterises Eulerian graphs. Euler proved
the necessity part and the sufficiency part was proved by Hierholzer [115].
Theorem 3.1 (Euler) A connected graph G is an Euler graph if and only if all vertices
of G are of even degree.

Proof
Necessity Let G(V, E) be an Euler graph. Thus G contains an Euler line Z , which is a
closed walk. Let this walk start and end at the vertex u V . Since each visit of Z to an
intermediate vertex v of Z contributes two to the degree of v and since Z traverses each edge
exactly once, d(v) is even for every such vertex. Each intermediate visit to u contributes
two to the degree of u, and also the initial and final edges of Z contribute one each to the
degree of u. So the degree d(u) of u is also even.
Sufficiency Let G be a connected graph and let degree of each vertex of G be even.
Assume G is not Eulerian and let G contain least number of edges. Since 2, G has a
cycle. Let Z be a closed walk in G of maximum length. Clearly, G E(Z) is an even degree
graph. Let C1 be one of the components of G E(Z). As C1 has less number of edges than
G, it is Eulerian and has a vertex v in common with Z . Let Z 0 be an Euler line in C1 . Then
Z 0 Z is closed in G, starting and ending at v. Since it is longer than Z , the choice of Z is
contradicted. Hence G is Eulerian.
Second proof for sufficiency Assume that all vertices of G are of even degree. We construct a walk starting at an arbitrary vertex v and going through the edges of G such that
no edge of G is traced more than once. The tracing is continued as far as possible. Since
every vertex is of even degree, we exit from the vertex we enter and the tracing clearly
cannot stop at any vertex but v. As v is also of even degree, we reach v when the tracing
comes to an end. If this closed walk Z we just traced includes all the edges of G, then G is
an Euler graph. If not, we remove from G all the edges in Z and obtain a subgraph Z 0 of G
formed by the remaining edges. Since both G and Z have all their vertices of even degree,
the degrees of the vertices of Z 0 are also even. Also, Z 0 touches Z at least at one vertex say
u, because G is connected. Starting from u, we again construct a new walk in Z 0 . As all the
vertices of Z 0 are of even degree, therefore this walk in Z 0 terminates at vertex u. This walk
in Z 0 combined with Z forms a new walk, which starts and ends at the vertex v and has more
edges than Z . This process is repeated till we obtain a closed walk that traces all the edges
of G. Hence G is an Euler graph (Fig. 3.2)
q

Fig. 3.2

Graph Theory

61

3.2 Konigsberg Bridge Problem

Two islands A and B formed by the Pregal river (now Pregolya) in Konigsberg (then the
capital of east Prussia, but now renamed Kaliningrad and in west Soviet Russia) were
connected to each other and to the banks C and D with seven bridges. The problem is to
start at any of the four land areas, A, B, C, or D, walk over each of the seven bridges exactly
once and return to the starting point.
Euler modeled the problem representing the four land areas by four vertices, and the
seven bridges by seven edges joining these vertices. This is illustrated in Figure 3.3.
C
C
A

D
D
Konigsberg Bridges

Underlying Graph

Fig. 3.3

We see from the graph G of the Konigsberg bridges that not all its vertices are of even
degree. Thus G is not an Euler graph, and implies that there is no closed walk in G containing all the edges of G. Hence it is not possible to walk over each of the seven bridges
exactly once and return to the starting point.
Note Two additional bridges have been built since Eulers day. The first has been built
between land areas C and D and the second between the land areas A and B. Now in the
graph of Konigsberg bridge problem with nine bridges, every vertex is of even degree and
the graph is thus Eulerian. Hence it is now possible to walk over each of the nine bridges
exactly once and return to the starting point (Fig 3.4).

Fig. 3.4

62

Eulerian and Hamiltonian Graphs

The following characterisation of Eulerian graphs is due to Veblen [254].

Theorem 3.2 A connected graph G is Eulerian if and only if its edge set can be decomposed into cycles.

Proof Let G(V, E) be a connected graph and let G be decomposed into cycles. If k of
these cycles are incident at a particular vertex v, then d(v) = 2k. Therefore the degree of
every vertex of G is even and hence G is Eulerian.
Conversely, let G be Eulerian. We show G can be decomposed into cycles. To prove
this, we use induction on the number of edges.
Since d(v) 2 for each v V, G has a cycle C. Then G E(C) is possibly a disconnected
graph, each of whose components C1 , C2 , . . ., Ck is an even degree graph and hence Eulerian. By the induction hypothesis, each Ci is a disjoint union of cycles. These together with
C provide a partition of E(G) into cycles.
q
The following result is due to Toida [244].
Theorem 3.3
of u v paths.

If W is a walk from vertex u to vertex v, then W contains an odd number

Proof Let W be a walk which we consider as a graph in itself, and not as a subgraph of
some other graph. Let u and v be initial and final vertices of the walk W . Clearly, d(u|W )
and d(v|W ) are odd, and d(w|W ) is even, for every w V (W ) {u, v}. We count the number
of distinct u v walks in W . These walks are the subgraphs of W .
When we take a u v walk by successively selecting the edges e1 , e2 , . . ., es , initial
vertex of e1 being u and terminal vertex of es being v, for each edge there are an odd number
of choices. The total number of such edges is the product of these odd numbers and is
therefore odd. Now from these walks, we find the u v paths. If a u v walk W1 is not a
path, then it contains one or more cycles. The traversal of these cycles in the two possible
alternative directions (clockwise and anticlockwise) produces in all an even number of
walks, all with the same edge set as W1 . Omitting these even number of walks which are
not paths from the total odd collection of u v walks, gives an odd number of u v paths.
q

Toida [244] proved the necessity part and McKee [157] the sufficiency part of the next
characterisation. The second proof of this result can be found in Fleischner [79], [80].
Theorem 3.4 A connected graph is Eulerian if and only if each of its edges lies on an
odd number of cycles.

Proof
Necessity Let G be a connected Eulerian graph and let e = uv be any edge of G. Then
G e is a u v walk W , and so G e = W contains an odd number of u v paths. Thus each
of the odd number of u v paths in W together with e gives a cycle in G containing e and
these are the only such cycles. Therefore there are an odd number of cycles in G containing
e.

Graph Theory

63

Sufficiency Let G be a connected graph so that each of its edges lies on an odd number of
cycles. Let v be any vertex of G and Ev = {e1 , . . ., ed } be the set of edges of G incident on
v, then |Ev | = d(v) = d . For each i, 1 i d , let ki be the number of cycles of G containing
ei . By hypothesis, each ki is odd. Let c(v) be the number of cycles of G containing v. Then
1
2

clearly c(v) =

i=1

i=1

ki implying that 2c(v) = ki . Since 2c(v) is even and each ki is odd, d is

even. Hence G is Eulerian.

Corollary 3.1
graph is even.

The number of edgedisjoint paths between any two vertices of an Euler

A consequence of Theorem 3.4 is the result of Bondy and Halberstam [37], which gives
yet another characterisation of Eulerian graphs.
Corollary 3.2
positions.

A graph is Eulerian if and only if it has an odd number of cycle decom-

Proof In one direction, the proof is trivial. If G has an odd number of cycle decompositions, then it has at least one, and hence G is Eulerian.
Conversely, assume that G is Eulerian. Let e E(G) and let C1 , . . ., Cr be the cycles
containing e. By Theorem 3.4, r is odd. We proceed by induction on m = |E(G)|, with G
being Eulerian.
If G is just a cycle, then the result is true. Now assume that G is not a cycle. This means
that for each i, 1 i r, by the induction assumption, Gi = G E(Ci ) has an odd number,
say si , of cycle decompositions. (If Gi is disconnected, apply the induction assumption to
each of the nontrivial components of Gi ). The union of each of these cycle decompositions
of Gi and Ci yields a cycle decomposition of G. Hence the number of cycle decompositions
of G containing Ci is si , 1 i r. Let s(G) denote the number of cycle decompositions of
G. Then
r

s(G) si r(mod 2)

(since si 1(mod 2))

i=1

1(mod 2).

Two examples of Euler graphs are shown in Figure 3.5.

Fig. 3.5

64

Eulerian and Hamiltonian Graphs

3.3 Unicursal Graphs

An open walk that includes (or traces) all edges of a graph without retracing any edge is
called a unicursal line or open Euler line. A connected graph that has a unicursal line is
called a unicursal graph. Figure 3.6 shows a unicursal graph.

Fig. 3.6

Clearly by adding an edge between the initial and final vertices of a unicursal line, we
get an Euler line.
The following characterisation of unicursal graphs can be easily derived from Theorem 3.1.
Theorem 3.5
odd degree.

A connected graph is unicursal if and only if it has exactly two vertices of

Proof Let G be a connected graph and let G be unicursal. Then G has a unicursal line,
say from u to v, where u and v are vertices of G. Join u and v to a new vertex w of G to get a
graph H . Then H has an Euler line and therefore each vertex of H is of even degree. Now,
by deleting the vertex w, the degree of vertices u and v each get reduced by one, so that u
and v are of odd degree.
Conversely, let u and v be the only vertices of G with odd degree. Join u and v to a new
vertex w to get the graph H . So every vertex of H is of even degree and thus H is Eulerian.
Therefore, G = H w has a u v unicursal line so that G is unicursal.
q
The following result is the generalisation of Theorem 3.5.
Theorem 3.6 In a connected graph G with exactly 2k odd vertices, there exists k edge
disjoint subgraphs such that they together contain all edges of G and that each is a unicursal
graph.

Proof Let G be a connected graph with exactly 2k odd vertices. Let these odd vertices
be named v1 , v2 , . . ., vk ; w1 , w2 , . . ., wk in any arbitrary order. Add k edges to G between
the vertex pairs (v1 , w1 ), (v2 , w2 ), . . ., (vk , wk ) to form a new graph H , so that every vertex
of H is of even degree. Therefore H contains an Euler line Z .
Now, if we remove from Z the k edges we just added (no two of these edges are incident
on the same vertex), then Z is divided into k walks, each of which is a unicursal line.
The first removal gives a single unicursal line, the second removal divides that into two
unicursal lines, and each successive removal divides a unicursal line into two unicursal
lines, until there are k of them. Hence the result.
q

Graph Theory

65

3.4 Arbitrarily Traceable Graphs

An Eulerian graph G is said to be arbitrarily traceable (or randomly Eulerian) from a vertex
v if every walk with initial vertex v can be extended to an Euler line of G. A graph is said
to be arbitrarily traceable if it is arbitrarily traceable from every vertex (Fig. 3.7).

Fig. 3.7

The following characterisation of arbitrarily traceable graphs is due to Ore [174]. Such
graphs were also characterised by Chartrand and White [56] .
Theorem 3.7 An Eulerian graph G is arbitrarily traceable from a vertex v if and only if
every cycle of G passes through v.

Proof
Necessity Let the Eulerian graph G be arbitrarily traceable from a vertex v. Assume there
is a cycle C not passing through v. Let H = G E(C). Then every vertex of H has an even
degree and the component of H containing v is Eulerian. This component of H can be
traversed as an Euler line Z , starting and ending with v and contains all those edges of G
which are incident at v. Clearly, this v v walk cannot be extended to contain the edges of
C also, contradicting that G contains v. Thus every cycle in G contains v.
Sufficiency Let every cycle of the Eulerian graph G pass through the vertex v of G. We
show that G is arbitrarily traceable from v. Assume, on the contrary, that G is not arbitrarily
traceable from v. Then there is a v v closed walk W of G containing all the edges of G
incident with v and yet not containing all the edges of G. Let one such edge be incident at
a vertex u on W . So every vertex of H = G E(W ) is of even degree and v is an isolated
vertex of H and u is not. The component of H containing u is therefore Eulerian subgraph
of G not passing through v, contradicting the assumption. Hence the result follows.
q
Corollary 3.3

Cycles are the only arbitrarily traceable graphs.

66

Eulerian and Hamiltonian Graphs

3.5 Sub-Eulerian Graphs

A graph G is said to be sub-Eulerian if it is a spanning subgraph of some Eulerian graph.
The following characterisation of sub-Eulerian graphs is due to Boesch, Suffel and Tindell [28].
Theorem 3.8 A connected graph G is sub-Eulerian if and only if G is not spanned by a
complete bipartite graph.

Proof
Necessity We prove that no spanning supergraph H of an odd complete bipartite graph G
is Eulerian. Let V1 V2 be the bipartition of the vertex set of G. Since degree of each vertex
of G is odd, and G is complete bipartite, therefore |V1 | and |V2 | are odd. If H1 is the induced
subgraph of H on V1 , then at least one vertex, say v, of V1 has even degree in H1 , since |V1 |
is odd. But then d(v|H) = d(v|H) + |V2|, which is odd. Therefore H is not Eulerian.
Sufficiency

Refer Boesch et. al., [28].

Super-Eulerian graphs
A non-Eulerian graph G is said to be super-Eulerian if it has a spanning Eulerian subgraph.
The following sufficient conditions for super-Eulerian graphs are due to Lesniak-Foster
and Williams [148].
Theorem 3.9 If a graph G is such that n 6, 2 and d(u) + d(v) n 1, for every pair
of non-adjacent vertices u and v, then G is super-Eulerian.
The following result is due to Balakrishnan and Paulraja [12].
Theorem 3.10 If G is any connected graph and if each edge of G belongs to a triangle
in G, then G has a spanning Eulerian subgraph.

Proof Since G has a triangle, G has a closed walk. Let W be the longest closed walk in
G. Then W must be a spanning Eulerian subgraph of G. If not, there exists a vertex v
/W
and v is adjacent to a vertex u of W . By hypothesis, uv belongs to a triangle, say uvw. If
none of the edges of this triangle is in W , then W {uv, vw, wu} yields a closed walk longer
than W (Fig. 3.8). If uw W , then (W uw) {uv, vw} would be a closed walk longer than
W . This contradiction proves that W is a spanning closed walk in G.
q

Graph Theory

67

Fig. 3.8

3.6 Hamiltonian Graphs

A cycle passing through all the vertices of a graph is called a Hamiltonian cycle. A graph
containing a Hamiltonian cycle is called a Hamiltonian graph. A path passing through all
the vertices of a graph is called a Hamiltonian path and a graph containing a Hamiltonian
path is said to be traceable. Examples of Hamiltonian graphs are given in Figure 3.9.

Fig. 3.9

If the last edge of a Hamiltonian cycle is dropped, we get a Hamiltonian path. However,
a non-Hamiltonian graph can have a Hamiltonian path, that is, Hamiltonian paths cannot
always be used to form Hamiltonian cycles. For example, in Figure 3.10, G1 has no Hamiltonian path, and so no Hamiltonian cycle; G2 has the Hamiltonian path v1 v2 v3 v4 , but has no
Hamiltonian cycle, while G3 has the Hamiltonian cycle v1 v2 v3 v4 v1 .

Fig. 3.10

68

Eulerian and Hamiltonian Graphs

Hamiltonian graphs are named after Sir William Hamilton, an Irish Mathematician
(18051865), who invented a puzzle, called the Icosian game, which he sold for 25 guineas
to a game manufacturer in Dublin. The puzzle involved a dodecahedron on which each of
the 20 vertices was labelled by the name of some capital city in the world. The aim of the
game was to construct, using the edges of the dodecahedron a closed walk of all the cities
which traversed each city exactly once, beginning and ending at the same city. In other
words, one had essentially to form a Hamiltonian cycle in the graph corresponding to the
dodecahedron. Figure 3.11 shows such a cycle.

Fig. 3.11

Clearly, the n-cycle Cn with n distinct vertices (and n edges) is Hamiltonian. Now, given
any Hamiltonian graph G, the supergraph G0 (obtained by adding in new edges between
non-adjacent vertices of G) is also Hamiltonian. This is because any Hamiltonian cycle in
G is also a Hamiltonian cycle of G0 . For instance, Kn is a supergraph of an n-cycle and so
Kn is Hamiltonian.
A multigraph or general graph is Hamiltonian if and only if its underlying graph is
Hamiltonian, because if G is Hamiltonian, then any Hamiltonian cycle in G remains a
Hamiltonian cycle in the underlying graph of G. Conversely, if the underlying graph of a
graph G is Hamiltonian, then G is also Hamiltonian.
Let G be a graph with n vertices. Clearly, G is a subgraph of the complete graph Kn .
From G, we construct step by step supergraphs of G to get Kn , by adding an edge at each
step between two vertices that are not already adjacent (Fig. 3.12).

Fig. 3.12

Now, let us start with a graph G which is not Hamiltonian. Since the final outcome of
the procedure is the Hamiltonian graph Kn , we change from a non-Hamiltonian graph to
a Hamiltonian graph at some stage of the procedure. For example, the non-Hamiltonian

Graph Theory

69

graph G1 above is followed by the Hamiltonian graph G2 . Since supergraphs of Hamiltonian graphs are Hamiltonian, once a Hamiltonian graph is reached in the procedure, all the
subsequent supergraphs are Hamiltonian.
Definition: A simple graph G is called maximal non-Hamiltonian if it is not Hamiltonian
and the addition of an edge between any two non-adjacent vertices of it forms a Hamiltonian graph. For example, G1 above is maximal non-Hamiltonian. Figure 3.13 shows a
maximal non-Hamiltonian graph.

Fig. 3.13

It follows from the above procedure that any non-Hamiltonian graph with n-vertices is
a subgraph of a maximal non-Hamiltonian graph with n vertices.
The above procedure is used to prove the following sufficient conditions due to Dirac
[68].
Theorem 3.11 (Dirac) If G is a graph with n vertices, where n 3 and d(v) n/2, for
every vertex v of G, then G is Hamiltonian.

Proof Assume that the result is not true. Then for some value n 3, there is a nonHamiltonian graph H in which d(v) n/2, for every vertex of H . In any spanning super
graph K (i.e., with the same vertex set) of H, d(v) n/2 for every vertex of K , since any
proper supergraph of this form is obtained by adding more edges. Thus there is a maximal
non-Hamiltonian graph G with n vertices and d(v) n/2 for every v in G. Using this G, we
obtain a contradiction.
Clearly, G 6= Kn , as Kn is Hamiltonian. Therefore there are non-adjacent vertices u and
v in G. Let G + uv be the supergraph of G by adding an edge between u and v. Since G
is maximal non-Hamiltonian, G + uv is Hamiltonian. Also, if C is a Hamiltonian cycle of
G + uv, then C contains the edge uv, since otherwise C is a Hamiltonian cycle of G, which
is not possible. Let this Hamiltonian cycle C be u = v1 , v2 , . . ., vn = v, u.
Now, let S = {vi C : there is an edge from u to vi+1 in G} and T = {v j C : there is an
edge from v to v j in G}.
Then vn / T , since otherwise there is an edge from v to vn = v, that is a loop, which is
impossible.

70

Eulerian and Hamiltonian Graphs

Also vn / S, (taking vn+1 as v1 ), since otherwise we again get a loop from u to v1 = u.

Therefore, vn S T (Fig. 3.14).

Fig. 3.14

Let |S|, |T | and |S T | be the number of elements in S, T and S T respectively. So

|S T | < n. Also, for every edge incident with u, there corresponds one vertex vi in S.
Therefore, |S| = d(u). Similarly, |T | = d(v).
Now, if vk is a vertex belonging to both S and T , there is an edge e joining u to vk+1 and
an edge f joining v to vk . This implies that C0 = v1 , vk+1, vk+2 , . . ., vn , vk , vk1 , . . ., v2 , v1 is
a Hamiltonian cycle in G, which is a contradiction as G is non-Hamiltonian. This shows
that there is no vertex vk in S T , so that S T = .
Thus |S T | = |S| + |T| |S T | gives |S| + |T| = |S T|, so that d(u) + d(v) < n. This is
a contradiction, because d(u) n/2 for all u in G, and so d(u) + d(v) n/2 + n/2 giving
d(u) + d(v) n. Hence the theorem follows.
q
The following result is due to Ore [176].
Theorem 3.12 (Ore) Let G be a graph with n vertices and let u and v be non-adjacent
vertices in G such that d(u) + d(v) n. Let G + uv denote the super graph of G obtained by
joining u and v by an edge. Then G is Hamiltonian if and only if G + uv is Hamiltonian.

Proof Let G be a graph with n vertices and suppose u and v are non-adjacent vertices
in G such that d(u) + d(v) n. Let G + uv be the super graph of G obtained by adding the
edge uv. Let G be Hamiltonian. Then obviously G + uv is Hamiltonian. Conversely, let
G + uv be Hamiltonian. We have to show that G is Hamiltonian. Then, as in Theorem 3.11,
we get d(u) + d(v) < n, which contradicts the hypothesis that d(u) + d(v) n. Hence G is
Hamiltonian.
q
The following is the proof of Bondy [35] of Theorem 3.12, and this proof bears a close
resemblance to the proof of Diracs theorem given by Newman [170], but is more direct.

Proof (Bondy [35]) Consider the complete graph K on the vertex set of G in which the
edges of G are coloured blue and the remaining edges of K are coloured red. Let C be a

Graph Theory

71

Hamiltonian cycle of K with as many blue edges as possible. We show that every edge of
C is blue, in other words, that C is Hamiltonian cycle of G.
Suppose to the contrary, C has a red edge uu (where u is the successor of u on C). Consider the set S of vertices joined to u by blue edges (that is, the set of neighbours of u in G).
The successor u of u on C must be joined by a blue edge to some vertex v of S , because
if u is adjacent in C only to vertices V (SU{u}), dG (u) + dG (u ) = |NG (u)| + |NG(u )|
|S| + (|V | |S | 1) = |V (G)| 1, contradicting the hypothesis that dG (u) + dG (u ) |V (G)|,
u and u being non-adjacent in G. But now the cycle C obtained from C by exchanging the
edges uu and vv has more blue edges than C, which is a contradiction.
q
Definition: Let G be a graph with n vertices. If there are two non-adjacent vertices u1
and v1 in G such that d(u1 ) + d(v1 ) n, join u1 and v1 by an edge to form the super graph
G1 . Now, if there are two non-adjacent vertices u2 and v2 in G1 such that d(u2 ) + d(v2 ) n,
join u2 and v2 by an edge to form supergraph G2 . Continue in this way, recursively joining
pairs of non-adjacent vertices whose degree sum is at least n until no such pair remains.
The final supergraph thus obtained is called the closure of G and is denoted by c(G).
The example in Figure 3.15 illustrates the closure operation.

Fig. 3.15

We observe in this example that there are different choices of pairs of non-adjacent
vertices u and v with d(u) + d(v) n. Therefore the closure procedure can be carried out in
several different ways and each different way gives the same result.
In the graph shown in Figure 3.16, n = 7 and d(u) + d(v) < 7, for any pair u, v of adjacent
vertices. Therefore, c(G) = G.
G

Fig. 3.16

72

Eulerian and Hamiltonian Graphs

The importance of c(G) is given in the following result due to Bondy and Chvatal [36].
Theorem 3.13

A graph G is Hamiltonian if and only if its closure c(G) is Hamiltonian.

Proof Let c(G) be the closure of the graph G. Since c(G) is a supergraph of G, therefore,
if G is Hamiltonian, then c(G) is also Hamiltonian.
Conversely, let c(G) be Hamiltonian. Let G, G1 , G2 , . . ., Gk1 , Gk = c(G) be the sequence
of graphs obtained by performing the closure procedure on G. Since c(G) = Gk is obtained
from Gk1 by setting Gk = Gk1 + uv, where u, v is a pair of non adjacent vertices in Gk1
with d(u) + d(v) n, therefore it follows that Gk1 is Hamiltonian. Similarly Gk2 , so Gk3,
. . ., G1 and thus G is Hamiltonian.
q
Corollary 3.4
Hamiltonian.

Let G be a graph with n vertices with n 3. If c(G) is complete, then G is

There can be more than one Hamiltonian cycle in a given graph, but the interest lies in the
edge-disjoint Hamiltonian cycles. The following result gives the number of edge-disjoint
Hamiltonian cycles in a complete graph with odd number of vertices.
Theorem 3.14 In a complete graph with n vertices there are (n 1)/2 edge-disjoint
Hamiltonian cycles, if n is an odd number, n 3.

Proof A complete graph G of n vertices has n(n 1)/2 edges and a Hamiltonian cycle in
G contains n edges. Therefore the number of edge-disjoint Hamiltonian cycles in G cannot
exceed (n 1)/2. When n is odd, we show there are (n 1)/2 edge-disjoint Hamiltonian
cycles.
The subgraph of a complete graph with n vertices shown in Figure 3.17 is a Hamiltonian
cycle. Keeping the vertices fixed on a circle, rotate the polygonal pattern clockwise by
360
360
n3 360
n1 , 2. n1 , . . ., 2 . n1 degrees. We see that each rotation produces a Hamiltonian cycle
that has no edge in common with any of the previous ones. Therefore, there are (n 3)/2
new Hamiltonian cycles, all disjoint from the one in Figure 3.17, and also edge-disjoint
among themselves. Thus there are (n 1)/2 edge disjoint Hamiltonian cycles.
q

Fig. 3.17

Graph Theory

73

The next result involving degrees give the sufficient conditions for a graph to be Hamiltonian.
Theorem 3.15 Let D = [di ]n1 be a degree sequence of a graph G = (V, E), d1 d2 . . . dn .
Each of the following gives the sufficient conditions for G to be Hamiltonian.
A. 1 k n dk

n
(Dirac [68])
2

B. uv
/ E d(u) + d(v) n (Ore [176])
C. 1 k n2 dk > k (Posa [210]).
D. j < k, d j j and dk k 1 d j + dk n (Bondy [33])
E. dk k < n2 dnk n k (Chvatal [59])
F. For every i and j with 1 i n, 1 j n, i + j n, vi v j / E , d(vi ) i and d(v j )
j 1 d(vi ) + d(v j ) n (Las Vergnas [256].
G. c(G) is complete (Bondy and Chvatal [36]).

Proof

We first prove that

(i)

(ii)

(iii)

(iv)

(v)

(vi)

A B C D E F G.

i. This can be easily established.

ii. Assume that (C) is not true, so that there exists a k with 1 k < n2 and dk k. Then
the induced subgraph on the vertices v1 , v2 , . . ., vk is a complete graph. For, if there
are vertices i and j with 1 i < j k and vi v j / E , then di + d j 2dk < n, contradicting
(B). Since dk k, each vi , 1 i k, is adjacent to at most one v j , k + 1 j n. Also,
n k > k, because k < 2n . Therefore there is a vertex v j , k + 1 j n not adjacent to
any of the vertices v1 , v2 , . . ., vk . For this v j , we have d j n k 1. But then d j + dk
(n k 1) + k = n 1. Thus there is a v j vk
/ E with d j + dk n 1, contradicting (B).
Hence proving (ii).
iii. Assume that (D) is not true, so that there exist j and k with j < k, d j j, dk k 1 and
d j + dk < n. This gives i = d j < n2 . But then d j j gives dd j d j , since the sequence
is non-decreasing. Therefore, di d j = i. Thus there is an i, 1 i 2n with di i,
contradicting (C). This proves (iii).
iv. If (E) is not true, there is a k with dk k < n2 and dnk n k 1. Then dk +dnk n 1.
Setting n k = j , we have k < j, dk k, d j j 1 and d j + dk n 1. This contradicts
(D) and so (iv) is proved.
v. Assume that (F) is not true, so that there is a pair of vertices vi and v j , i < j with
vi v j
/ E and violating (F). Choose i to be the least such possible integer. Then by
minimality of i, di1 > i 1. Thus di di1 i and since di i, we obtain di = i. If

74

Eulerian and Hamiltonian Graphs

i 2n , we get di + d j 2di n, contradicting the violation of (F). Therefore, i < 2n .
Thus there is an i, 1 i < 2n with di = i. Now, if (E) is satisfied, we have dni n i
and since j n i, we obtain d j dni n i. By minimality of i, di = i and we have
d j + di (n i) + i = n, again contradicting the violation of (F). Thus negation of (F)

implies negation of (E) and (V) is established.

vi. Assume that c(G) = H is not complete. Let vi and v j be non-adjacent vertices in H
such that (a) j is as large as possible and (b) i is as large as possible subject to (a).
Then i < j , and since H is the closure of G, therefore
d(vi |H) + d(v j |H) n 1,

(3.15.1)

d(vi |G) + d(v j |G) n 1.

By the choice of j, vi is adjacent in H to all vk with k > j , so that

d(vi |H) n j .

(3.15.2)

Again, by the choice of i, v j is adjacent in H to all vk with k > i, k 6= j , so that

d(v j |H) n i 1.

(3.15.3)

From (3.15.1) and (3.15.2), we have

d(v j |G) d(v j |H) (n 1) (n j) = j 1.

From (3.15.1) and (3.15.3), we have

d(v j |G) d(vi |H) (n 1) (n i 1) = i.

From (3.15.2) and (3.15.3), we have

i + j (2n 1) d(vi |H) d(v j |H) n.

(using (1))

Therefore i and j contradict the given conditions. Thus H = c(G) is complete. This
proves (vi).
By Theorem 3.13 it follows that if (G) holds, then G is Hamiltonian.

The next result is due to Nash-Williams [168].

Theorem 3.16 (Nash-Williams)
nian.

Every k-regular graph on 2k + 1 vertices is Hamilto-

Proof Let G be a k-regular graph on 2k + 1 vertices. Add a new vertex w and join it by
an edge to each vertex of G. The resulting graph H on 2k + 2 vertices has = k + 1. Thus
by Theorem 3.15 (A), H is Hamiltonian. Removing w from H , we get a Hamiltonian path,
say v0 v1 . . .v2k .

Graph Theory

75

Assume that G is not Hamiltonian, so that (a) if v0 vi E , then vi1 v2k

/ E , (b) if v0 vi
/ E,
then vi1 v2k E , since d(v0 ) = d(v2k ) = k.
The following cases arise.
Case (i) v0 is adjacent to v1 , v2 , . . ., vk , and v2k is adjacent to vk , vk+1 , . . ., v2k1 . Then
there is an i with 1 i k such that vi is not adjacent to some v j for 0 j k( j 6= i). But
d(vi ) = k. So vi is adjacent to v j for some j with k + 1 j 2k 1. Then the cycle C given
by vi vi1 . . .v0 vi+1 . . .v j1 v2k v2k+1 . . .v j is a Hamiltonian cycle of G (Fig 3.18).

Fig. 3.18
Case (ii) There is an i with 1 i 2k 1 such that vi+1 v0 E , but vi v0
/ E . Then by
(b), vi1 v2k E . Thus G contains the 2k-cycle vi1 vi2 . . .v0 vi+1 . Renaming the 2k-cycle
C as u1 u2 . . .u2k and let u0 be the vertex of G not on C . Then u0 cannot be adjacent to
two consecutive vertices on C and hence u0 is adjacent to every second vertex on C, say
u1 , u3 , . . ., u2k1 . Replacing u2i by u0 , we obtain another maximum cycle C 0 of G and hence
u2i must be adjacent to u1 , u3 , . . ., u2k1 . But then u1 is adjacent to u0 , u2 , . . ., u2k , implying
d(u1 ) k + 1. This is a contradiction and hence G is Hamiltonian.
q

3.7 Pancyclic Graphs

Definition: A graph G of order n( 3) is pancyclic if G contains all cycles of lengths
from 3 to n. G is called vertex-pancyclic if each vertex v of G belongs to a cycle of every
length `, 3 ` n.

76

Eulerian and Hamiltonian Graphs

Example Clearly, a vertex-pancyclic graph is pancyclic. However, the converse is not

true. Figure 3.19 displays a pancyclic graph that is not vertex-pancyclic.

Fig. 3.19

The result of pancyclic graphs was initiated by Bondy [34], who showed that Ores
sufficient condition for a graph G to be Hamiltonian (Theorem 6.2.5) actually implies much
2
more. Note that if 2n , then m n2 . The proof of the following result due to Thomassen
can be found in Bollobas [29].
h 2i

Theorem 3.17 Let G be a simple Hamiltonian graph on n vertices with at least n2

edges. Then G is either pancyclic or else is the complete bipartite graph K 2n , n2 . In particular,
2
if G is Hamiltonian and m > n4 , then G is pancyclic.

Proof The result can easily be verified for n = 3. We may therefore assume that n 4. We
apply induction on n. Suppose the result is true for all graphs of order at most n 1(n 4),
and let G be a graph of order n.
First, assume that G has a cycle C = v0 v1 . . .vn2 v0 of length n 1. Let v be the (unique)
vertex of G not belonging to C. If d(v) n2 , v is adjacent to two consecutive vertices on
C and hence G has a cycle of length 3. Suppose for some r, 2 r n1
2 , C has no pair of
vertices u and w on C adjacent to v in G with dC (u, w) = r. Then if vi1 , vi2 , . . .vid(v) are the
vertices of C that are adjacent to v in G (recall that C contains all the vertices of G except v),
then vi1 +r , vi2 +r , . . ., vid(v) +r are nonadjacent to v in G, where the suffixes are taken modulo
(n 1). Thus, 2d(v) n 1, a contradiction. Hence, for each r, 2 r n1
2 , C has a pair of
vertices u and w on C adjacent to v in G with dC (u, w) = r. Thus for each r, 2 r n1
2 , G
has a cycle of length r + 2 as well as a cycle of length n 1 r + 2 = n r + 1 (Fig. 3.20).
Thus G is pancyclic.

Fig. 3.20

Graph Theory

77
2

n
n1
If d(v) n1
2 , then G[V (C)], the subgraph of G induced by V (C) has at least 4 2 >
2
(n1)
edges. So, by the induction assumption, G[V (C)] is pancyclic and hence G is pan4
cyclic. (By hypothesis, G is Hamiltonian).
Next, assume that G has no cycle of length n 1. Then G is not pancyclic. In this case,
we show that G is K 2n , n2 .
Let C = v0 v1 v2 . . .vn1 v0 be a Hamilton cycle of G. We claim that of the two pairs vi vk and
vi+1 vk+2 (where suffixes are taken modulo n), at most only one of them can be an edge of
G. Otherwise, vk vk1 vk2 . . .vi+1 vk+2 vk+3 vk+4 . . .vi vk is an (n 1)-cycle in G, a contradiction.
Hence, if d(vi ) = r, then there are r vertices adjacent to vi in G and hence at least r vertices
(including vi+1 since vi vi1 E(G)) that are nonadjacent to vi+1 . Thus, d(vi+1 ) n r and
d(vi ) + d(vi+1 ) n.
Summing the last inequality over i from 0 to n 1, we get 4m n2 . But by hypothesis,
2
4m n2 . Hence, m = n4 and so n must be even.
This gives d(vi ) + d(vi+1 ) = n for each i, and thus for each i and k, exactly one of vi vk and
vi+1 vk+2 is an edge of G.
(3.17.1)
Thus, if G 6= K 2n , n2 , then certainly there exist i and j such that vi v j E and i j (mod 2).
Hence for some j , there exists an even positive integer s such that v j+1 v j+1+s E . Choose s
to be the least even positive integer with the above property. Then v j v j+1+s
/ E . Hence, s 4
(as s = 2 would mean that v j v j+1 / E ). Again, by (3.17.1), v j1 v j+s3 = v j1 v j1+(s2) E(G)
contradicting the choice of s. Thus, G = K 2n , n2 . The last part follows from the fact that
2
q
|E(K 2n , n2 , )| = n4 .

Theorem 3.18 Let G 6= K n2 , n2 , be a simple graph with n 3 vertices and let d(u)+d(v) n
for every pair of non-adjacent vertices of G. Then G is pancyclic.

Proof By Ores Theorem (Theorem 3.12), G is Hamiltonian. We show that G is pan2

cyclic by first proving that m n4 and then invoking Theorem 3.17. This is true if n2
n

(as 2m = di n n2 /2). So assume that < 2n .

i=1

Let S be the set of vertices of degree in G. For every pair (u, v) of vertices of degree
, d(u) + d(v) < 2n + n2 = n. Hence by hypothesis, S induces a clique of G and |S| + 1. If
|S| = + 1, then G is disconnected with G[S] as a component, which is impossible (as G is
Hamiltonian). Thus, |S| . Further, if v S, v is nonadjacent to n 1 vertices of G. If
u is such a vertex, d(v) + d(u) n implies that d(u) n . Further, v is adjacent to at least
one vertex w / S and d(w) + 1, by the choice of S. These facts give that
n

2m = di (n 1)(n ) + 2 + ( + 1),
i=1

where the last ( + 1) comes out of the degree of w. Thus,

2m n2 n(2 + 1) + 2 2 + 2 + 1,

78

Eulerian and Hamiltonian Graphs

which implies that

4m 2n2 2n(2 + 1) + 4 2 + 4 + 2
= (n (2 + 1)2 + n2 + 1
n2 + 1, since n > 2 .

Consequently, m >

n2
, and by Theorem 3.17, G is pancyclic.
4

3.8 Exercises
1. Prove that the wheel Wn is Hamiltonian for every n 2, and n-cube Qn is Hamiltonian
for each n 2.
2. If G is a k-regular graph with 2k 1 vertices, then prove that G is Hamiltonian.
3. Show that if a cubic graph G has a spanning closed walk, then G is Hamiltonian.
4. If G = G(X, Y ) is a bipartite Hamiltonian graph, then show that |X| = |Y |.
5. Prove that for each n 1, the complete tripartite graph Kn, 2n, 3n is Hamiltonian, but
Kn, 2n, 3n+1 is not Hamiltonian.
6. How many spanning cycles are there in the complete bipartite graphs K3, 3 and K4, 3 ?
7. Prove that a graph G with n 3 vertices is arbitrarily traceable if and only if it is one
of the graphs Cn , Kn or Kn, n with n = 2p.
8. Prove that a graph G with n 3 vertices is randomly traceable if and only if it is
randomly Hamiltonian.
9. Find the closure of the graph given in Figure 3.2. Is it Hamiltonian?
10. Does there exist an Eulerian graph with
i. an even number of vertices and an odd number of edges,
ii. and odd number of vertices and an even number of edges.
Draw such a graph if it exists.
11. Characterise graphs which are both Eulerian and Hamiltonian.
12. Characterise graphs which possess Hamiltonian paths but not Hamiltonian cycles.
13. Characterise graphs which are unicursal but not Eulerian.
14. Give an example of a graph which is neither pancyclic nor bipartite, but whose nclosure is complete.

4. Trees
One of the important classes of graphs is the trees. The importance of trees is evident from
their applications in various areas, especially theoretical computer science and molecular
evolution.

4.1 Basics
Definition: A graph having no cycles is said to be acyclic. A forest is an acyclic graph.
Definition: A tree is a connected graph without any cycles, or a tree is a connected
acyclic graph. The edges of a tree are called branches. It follows immediately from the
definition that a tree has to be a simple graph (because self-loops and parallel edges both
form cycles). Figure 4.1(a) displays all trees with fewer than six vertices.

Fig. 4.1(a)

The following result characterises trees.

Theorem 4.1
of its vertices.

A graph is a tree if and only if there is exactly one path between every pair

80

Trees

Proof Let G be a graph and let there be exactly one path between every pair of vertices
in G. So G is connected. Now G has no cycles, because if G contains a cycle, say between
vertices u and v, then there are two distinct paths between u and v, which is a contradiction.
Thus G is connected and is without cycles, therefore it is a tree.
Conversely, let G be a tree. Since G is connected, there is at least one path between
every pair of vertices in G. Let there be two distinct paths between two vertices u and v of
G. The union of these two paths contains a cycle which contradicts the fact that G is a tree.
Hence there is exactly one path between every pair of vertices of a tree.
q
The next two results give alternative methods for defining trees.
Theorem 4.2

A tree with n vertices has n 1 edges.

Proof We prove the result by using induction on n, the number of vertices. The result
is obviously true for n = 1, 2 and 3. Let the result be true for all trees with fewer than n
vertices. Let T be a tree with n vertices and let e be an edge with end vertices u and v. So
the only path between u and v is e. Therefore deletion of e from T disconnects T . Now,
T e consists of exactly two components T1 and T2 say, and as there were no cycles to
begin with, each component is a tree. Let n1 and n2 be the number of vertices in T1 and T2
respectively, so that n1 + n2 = n. Also, n1 < n and n2 < n. Thus, by induction hypothesis,
number of edges in T1 and T2 are respectively n1 1 and n2 1. Hence the number of edges
in T = n1 1 + n2 1 + 1 = n1 + n2 1 = n 1.
q
Theorem 4.3

Any connected graph with n vertices and n 1 edges is a tree.

Proof Let G be a connected graph with n vertices and n 1 edges. We show that G
contains no cycles. Assume to the contrary that G contains cycles.
Remove an edge from a cycle so that the resulting graph is again connected. Continue
this process of removing one edge from one cycle at a time till the resulting graph H is a
tree. As H has n vertices, so number of edges in H is n 1. Now, the number of edges in G
is greater than the number of edges in H . So n 1 > n 1, which is not possible. Hence, G
has no cycles and therefore is a tree.
q
Definition: A graph is said to be minimally connected if removal of any one edge from it
disconnects the graph. Clearly, a minimally connected graph has no cycles.
Here is the next characterisation of trees.
Theorem 4.4

A graph is a tree if and only if it is minimally connected.

Proof Let the graph G be minimally connected. Then G has no cycles and therefore is a
tree.
Conversely, let G be a tree. Then G contains no cycles and deletion of any edge from G
disconnects the graph. Hence G is minimally connected.
q

Graph Theory

81

The following results give some more properties of trees.

Theorem 4.5

A graph G with n vertices, n 1 edges and no cycles is connected.

Proof Let G be a graph without cycles with n vertices and n 1 edges. We have to prove
that G is connected. Assume that G is disconnected. So G consists of two or more components and each component is also without cycles. We assume without loss of generality that
G has two components, say G1 and G2 (Fig. 4.1(b)). Add an edge e between a vertex u in
G1 and a vertex v in G2 . Since there is no path between u and v in G, adding e did not create
a cycle. Thus G e is a connected graph (tree) of n vertices, having n edges and no cycles.
This contradicts the fact that a tree with n vertices has n 1 edges. Hence G is connected.

Fig. 4.1(b)

Theorem 4.6

Any tree with at least two vertices has at least two pendant vertices.

Proof Let the number of vertices in a given tree T be n(n > 1). So the number of edges in
T is n 1. Therefore the degree sum of the tree is 2(n 1). This degree sum is to be divided
among the n vertices. Since a tree is connected it cannot have a vertex of 0 degree. Each
vertex contributes at least 1 to the above sum. Thus there must be at least two vertices of
degree exactly 1.
Second proof We use induction on n. The result is obviously true for all trees having
fewer than n vertices. We know that T has n 1 edges, and if every edge of T is incident
with a pendant vertex, then T has at least two pendant vertices, and the proof is complete.
So let there be some edge of T that is not incident with a pendant vertex and let this edge
be e = uv (Fig. 4.2). Removing the edge e, we see that the graph T e consists of a pair
of trees say T1 and T2 with each having fewer than n-vertices. Let u V (T1 ), v V (T2 ), and
|V (T1 )|= n1 , |V (T2 )|= n2 . Applying induction hypothesis on both T1 and T2 , we observe that
each of T1 and T2 has two pendant vertices. This shows that each of T1 and T2 has at least
one pendant vertex that is not incident with the edge e. Thus the graph T e + e = T has at
least two pendant vertices.

Fig. 4.2

82

Trees

Third proof Let T be a tree with n(n > 1) vertices. The number of edges in T is n 1 and
the sum of degrees in T is 2(n 1), that is, di = 2(n 1). Assume T has exactly one vertex
v1 of degree one, while all the other n 1 vertices have degree 2. Then sum of degrees
is d(v1 ) + d(v2 ) + . . . + d(vn ) 1 + 2 + 2 + . . .+ 2 = 1 + 2(n 1). So, 2(n 1) 1 + 2(n 1),
implying 0 1, which is absurd. Hence T has at least two vertices of degree one.
q
The following result characterises tree degree sequences.
Theorem 4.7
only if

The sequence [di ]n1 of positive integers is a degree sequence of a tree if and
n

(i) di 1 for all i, 1 i n and (ii) di = 2n 2.

i=1

Proof
Necessity

Since a tree has no isolated vertex, therefore di 1 for all i. Also, di =

i=1

2(n 1), as a tree with n vertices has n 1 edges.

Sufficiency We use induction on n. For n = 2, the sequence is [1, 1] and is obviously the
degree sequence of K2 . Suppose the claim is true for all positive sequences of length less
than n.
Let [di ]n1 be the non-decreasing positive sequence of n terms, satisfying conditions (i)
and (ii). Then d1 = 1 and dn > 1 (by Theorem 4.5).
Now, consider the sequence D0 = [d2, d3 , . . ., dn1 , dn 1], which is a sequence of length
n 1. Obviously in D0 , di 1 and di = d2 + d3 + . . . + dn1 + dn 1 = d1 + d2 + d3 + . . . +
dn1 + dn 1 1 = 2n 2 2 = 2(n 1) 2 (because d1 = 1). So D0 satisfies conditions (i)
and (ii), and by induction hypothesis there is a tree T 0 realising D0 . In T 0 , add a new vertex
and join it to the vertex having degree dn 1 to get a tree T . Therefore the degree sequence
of T is [d1, d2 , . . ., dn ].
Theorem 4.8

A forest of k trees which have a total of n vertices has n k edges.

Proof

Let G be a forest and T1 , T2 , . . ., Tk be the k trees of G. Let G have n vertices and

T1 , T2 , . . ., Tk have respectively n1 , n2 , . . ., nk vertices. Then n1 + n2 + . . . + nk = n. Also, the
number of edges in T1 , T2 , . . ., Tk are respectively n1 1, n2 1, . . ., nk 1. Thus number of
edges in G = n1 1 + n2 1 + . . .+ nk 1 = n1 + n2 + . . . + nk k = n k.
q
The following result characterises trees as subgraphs of a graph.

Theorem 4.9 Let T be a tree with k edges. If G is a graph whose minimum degree
satisfies (G) k, then G contains T as a subgraph. Alternatively, G contains every tree of
order atmost (G) + 1 as a subgraph.

Proof We use induction on k. If k = 0, then T = K1 and it is clear that K1 is a subgraph of

any graph. Further, if k = 1, then T = K2 and K2 is a subgraph of any graph whose minimum

Graph Theory

83

degree is one. Assume the result is true for all trees with k 1 edges (k 2) and consider
a tree T with exactly k edges. We know that T contains at least two pendant vertices. Let
v be one of them and let w be the vertex that is adjacent to v. Consider the graph T v.
Since T v has k 1 edges, the induction hypothesis applies, so T v is a subgraph of G.
We can think of T v as actually sitting inside G (meaning w is a vertex of G, too). Since
G contains at least k + 1 vertices, and T v contains k vertices, there exist vertices of G that
are not a part of the subgraph T v. Further, since the degree of w in G is at least k, there
must be a vertex u not in T v that is adjacent to w. The subgraph T v together with u
forms the tree T as a subgraph of G (Fig. 4.3).
q

Fig. 4.3

4.2 Rooted and Binary Trees

A tree in which one vertex (called the root) is distinguished from all the others is called a
rooted tree.
A binary tree is defined as a tree in which there is exactly one vertex of degree two and
each of the remaining vertices is of degree one or three. Obviously, a binary tree has three
or more vertices. Since the vertex of degree two is distinct from all other vertices, it serves
as a root, and so every binary tree is a rooted tree.
q
Below are given some properties of binary trees.
Theorem 4.10

Every binary tree has an odd number of vertices.

Proof Apart from the root, every vertex in a binary tree is of odd degree. We know that
there are even number of such odd vertices. Therefore when the root (which is of even
degree) is added to this number, the total number of vertices is odd.
Corollary 4.1

There are 21 (n + 1) pendant vertices in any binary tree with n vertices.

Proof Let T be a binary tree with n vertices. Let q be the number of pendant vertices in
T . Therefore there are n q internal vertices in T and so n q 1 vertices of degree 3. Thus
the number of edges in T = 21 [3(n q 1) + 2 + q]. But the number of edges in T is n 1.

84

Trees

Hence, 12 [3(n q 1) + 2 + q] = n 1, so that q = 12 (n + 1).

The following result is due to Jordan [122].

Theorem 4.11 (Jordan)

Every tree has either one or two centers.

Proof The maximum distance, max d(v, vi ) from a given vertex v to any other vertex
occurs only when vi is a pendant vertex. With this observation, let T be a tree having
more than two vertices. Tree T has two or more pendant vertices. Deleting all the pendant
vertices from T , the resulting graph T 0 is again a tree. The removal of all pendant vertices
from T uniformly reduces the eccentricities of the remaining vertices (vertices in T 0 ) by
one. Therefore the centers of T are also the centers of T 0 . From T 0 we remove all pendant
vertices and get another tree T 00 . Continuing this process, we either get a vertex, which is a
center of T , or an edge whose end vertices are the two centers of T .
Definition: Trees with center K1 are called unicentral and trees with center K2 are called
bicentral trees.

Spanning trees
A tree is said to be a spanning tree of a connected graph G, if T is a subgraph of G and T
contains all vertices of G.
Example

Consider the graph of Fig. 4.4, where the bold lines represent a spanning tree.

Fig. 4.4
q

The following result shows the existence of spanning trees in connected graphs.
Theorem 4.12

Every connected graph has at least one spanning tree.

Proof Let G be a connected graph. If G has no cycles, then it is its own spanning tree.
If G has cycles, then on deleting one edge from each of the cycles, the graph remains
connected and cycle free containing all the vertices of G.
Definition: An edge in a spanning tree T is called a branch of T . An edge of G that is
not in a given spanning tree T is called a chord. It may be noted that branches and chords

Graph Theory

85

are defined only with respect to a given spanning tree. An edge that is a branch of one
spanning tree T1 (in a graph G) may be a chord with respect to another spanning tree T2 . In
Figure 4.5, u1 u2 u3 u4 u5 u6 is a spanning tree, u2 u4 and u4 u6 are chords.

Fig. 4.5

A connected graph G can be considered as a union of two subgraphs T and T , that is

G = T T , where T is a spanning tree, T is the complement of T in G. T being the set of
chords is called the co tree, or chord set.
The following result provides the number of chords in any graph with a spanning tree.
Theorem 4.13 With respect to any of its spanning trees, a connected graph of n vertices
and m edges has n 1 tree branches and m n + 1 chords.

Proof Let G be a connected graph with n vertices and m edges. Let T be the spanning
tree. Since T contains all n vertices of G, T has n 1 edges and thus the number of chords
in G is equal to m (n 1) = m n + 1.
q
Definition: Let G be a graph with n vertices, m edges and k components. The rank r and
nullity of G are defined as r = n k and = m n + k.
Clearly, the rank of a connected graph is n 1 and the nullity is m n + 1.
It can be seen that rank of G = number of branches in any spanning tree (or forest) of G.
Also, nullity of G = number of chords in G. So, rank + nullity = number of edges in G.
The nullity of a graph is also called its cyclomatic number, or first Betti number.
Theorem 4.14 If T is a tree with 2k 0 vertices of odd degree, then E(T ) is the union of
k pair-wise edge-disjoint paths.

Proof We prove the result for every forest G, using induction on k. If k = 0, then G has
no pendant vertex and therefore no edge. Let k > 0 and let each forest with 2k 2 vertices
of odd degree has decomposition into k 1 paths. Since k > 0, some component of G is a
tree with at least two vertices. This component has at least two pendant vertices. Let P be
the path connecting two pendant vertices. Deleting E(P) changes the parity of the vertex
degree only for the end vertices of P and it makes them even. Thus G E(P) is a forest
with 2k 2 vertices of odd degree. So by the induction hypothesis, G E(P) is the union of

86

Trees

k 1 pair wise edge-disjoint paths. These k 1 edge-disjoint paths together with P partition
E(G) into k pair wise edge-disjoint paths (Fig. 4.6).

Fig. 4.6

Theorem 4.15 Let T be a non-trivial tree with the vertex set S and |S|= 2k, k 1. Then
there exists a set of k pairwise edge-disjoint paths whose end vertices are all the vertices
of S.

Proof Obviously, there exists a set of k paths in T whose end vertices are all the vertices
of S. Let P = {P1 , P2 , . . ., Pk } be such a set of k paths and let the sum of their lengths be the
minimum.
We show that the paths of P are pairwise edge-disjoint. Assume to the contrary, and let
Pi and Pj , i 6= j , be paths having an edge in common. Then Pi and Pj have path Pi j of length
1 in common. Therefore, Pi Pj the symmetric difference of Pi and Pj is a disjoint union of
two paths, say Qi and Q j , with their end vertices being disjoint pairs of vertices belonging
to S (Fig. 4.7).
If Pi and Pj are replaced by Qi and Q j in P, then the resulting set of paths has the property
that their end vertices are all the vertices of S and that the sum of their lengths is less than
the sum of the lengths of the paths in P. This is a contradiction to the choice of P.

Fig. 4.7

Graph Theory

Theorem 4.16

87

If u is a vertex of an n-vertex tree T , then

d(u, v)
vV (T )

n
2

Proof Let T (V, E) be a tree with |V |= n. Let u be any vertex of T . We use induction on
n. If n = 2, the result is trivial. Let n > 2. The graph T u is a forest and let the components
of T u be T1 , T2 , . . ., Tk , where k 1. Since T is connected, u has a neighbour in each Ti .
Also, since T has no cycles, u has exactly one neighbour vi in each Ti . For any v V (Ti ), the
unique u v path in T passes through vi and we have dT (u, v) = 1 + dTi (vi , v). Let ni = n(Ti )
(Fig. 4.8). Then we have

dT (u, v) = ni + dTi (vi , v).

vV (Ti )

(4.16.1)

vV (Ti )

Fig. 4.8

By the induction hypothesis, we have

dTi (vi , v)

vV (Ti )

n 
i

We now sum the formula (4.16.1) for distances from u over all the components of T u
and we get

dT (u, v) (n 1) +
i

vV (Ti )

n 
i

Now, we have ni = n 1. Clearly,

i
Kn1 ,

ni

2

ni
2

, because the right side counts the

and the left side counts the edges in a subgraph of K ni , the subgraph
edges in K ni or
being union of disjoint cliques Kn1 , Kn1 , . . ., Knk .
Thus,

dT (u, v) (n 1) +

vV (T )

n1
2

n
2

88

Trees

Corollary 4.2

The sum of the distances from a pendant vertex of the path Pn to all other

n1

i=

vertices is

i=0

Corollary 4.3

n
2

If H is a subgraph of a graph G, then dG (u, v) dH (u, v).

Proof Every u v path in H appears also in G, and G may have additional u v paths that
are shorter than any u v path in H .
Corollary 4.4

If u is a vertex of a connected graph G, then

d(u, v)
vV (G)

Proof


n(G)
.
2

Let T be a spanning tree of G. Then dG (u, v) dT (u, v), so that

dG (u, v)

vV (G)

dT (u, v)

vV (G)


n(G)
.
2

The sum of the distances over all pairs of distinct vertices in a graph G is the Wiener
index W (G) = d(u, v). On assigning vertices for the atoms and edges for the atomic
u, vV (G)

bonds, we can use graphs to study molecules. Wiener [268] originally used this to study
the boiling point of paraffin.
Theorem 4.17 Let v be any vertex of a connected graph G. Then G has a spanning tree
preserving the distances from v.

Proof

Let G be a connected graph. We find a spanning tree T of G such that for each
u V = V (G) = V (T ), dG (v, u) = dT (v, u).
Consider the neighbourhoods of v,
Ni (v) = {u V : dG (v, u) = i}, 1 i e, where e = e(v).
Let H be the graph obtained from G by removing all edges in each < Ni (v) >. Clearly, H
is connected. Let < Bi (v) >H denote the induced subgraph of H , induced by the ball Bi (v).
Clearly, < B1 (v) >H does not contain any cycle. If < B2 (v) >H contains cycles, remove
edges from [ N1(v), N2 (v) ] sequentially, one edge from each cycle, till it becomes acyclic.
Proceeding successively by removing edges from [ Ni(v), Ni+1(v) ] to make < Bi+1 (v) >H
acyclic for 1 i e 1, we get a spanning tree of H and hence of G.
Since in this procedure one distance path from v to each of the other vertices remains
intact, we have dG (v, u) = dT (v, u) for each u V .
q

Graph Theory

89

Remarks The above result implies that for any vertex v of a connected graph G, there
exists an image v(G) which is a spanning tree of G preserving distances from v. This is
called an isometric tree of G at v. If there is only one such tree (upto isomorphism) at v, we
say that G has a unique isometric tree at v. If G has the same unique isometric tree at each
vertex v, then G is said to have a unique isometric tree (or unique distance tree). K2, 2 and
the Peterson graph are examples of graphs having unique isometric trees, while K3, 3 does
not have a unique isometric tree at any vertex. Every tree has a unique isometric tree.

The next result due to Chartrand and Stewart [52] gives the necessary condition for a
graph to have a unique isometric tree.
Theorem 4.18 Let G be a connected graph with d = 2r, which has a unique isometric
tree. Then the end vertices of every diametral path of G has degree 1.

Proof Let G be a connected graph with d = 2r and let P be a diametral path with end
vertices u and v. If possible let d(u|G) > 1. Let Tu be the isometric tree at u. It is easy to see
that Tu can be chosen to contain P.
Since u has degree at least 2 in G, there is a vertex ui adjacent to u and not lying in P.
Clearly, dTu (ui , v) = 1 + d .
Let c be a central vertex of G. Then for any two vertices w1 and w2 of G, we have
dG (w1 , c) r = 21 d and dG (w2 , c) r = 21 d .

Therefore, dG (w1 , w2 ) d(w1 , c) + d(c, w2 ) d .

Since Tc is isometric with G at c, we also have dT c(w1 , w2 ) d .
Thus no path Tc has length greater than d , whereas there is a path in Tu of length 1 +
d . Therefore Tc 6= Tu and G does not have a unique isometric tree. This contradicts the
hypothesis. Hence the result follows.
q
Remark The above condition is necessary but not sufficient. To see this, consider the
graph given in Figure 4.9.

Fig. 4.9

Graph without a unique isometric tree

Chartrand and Schuster [54], and Kundu [142] have given some more results on the graphs
with unique isometric trees.
Definition: The complexity (G) of a graph G is the number of different spanning trees
of G.

90

Trees

The following result gives a recursive formula for (G).

Theorem 4.19

For any cyclic edge e of a graph G, (G) = (G e) + (G|pe).

Proof Let S be the set of spanning trees of G and let S be partitioned as S1 S2 , where S1
is the set of spanning trees of G not containing e and S2 is the set of the spanning trees of G
containing e.
Since e is a cyclic edge, G e is connected and there is a one-one correspondence between the elements of S1 and the spanning trees of G e. Also, there is a one-one correspondence between the spanning trees of G|e and the elements of S2 .
Thus, (G) =|S1 |+|S2|= (G e) + (G|e).
q
Remarks

1. The above recurrence relation is valid even if e is a cut edge. This is because (G
e) = 0 and every spanning tree of G contains every cut edge.
2. The recurrence relation is valid even if G is a general graph and e is a multiple edge,
but not when e is a loop.
3. The complexity of any graph G is computed by repeatedly applying the above recurrence. We observe that on applying the elementary contraction to a multiple edge,
the resulting graph can have a loop and by remark (2) the procedure can be still continued. At each stage of the algorithm, only an edge belonging to the proper cycle is
chosen. The algorithm starts with a given graph and produces two graphs (possibly
general) at the end of the first stage. At each subsequent stage one proper cyclic edge
from each graph is chosen (if it exists) for applying the recurrence. On termination
of the algorithm, we get a set of graphs (or general graphs) none of which have a
proper cycle. Then (G) is the sum of the number of these graphs. If H is any of
these graphs, then (H) is the product of its edges, ignoring the loops.
Example

Consider the graph G given in Figure 4.10.

Fig. 4.10

Graph Theory

91

Label the edges of G arbitrarily. Choose e1 as the first cyclic edge. Then (G) is the sum
of the complexities of the graphs given in Figure 4.10(b) and (c). Now, choose e4 in both
G1 and G2 as the next cyclic edge. Then (G) is the sum of the complexities of the graphs
in Figure 4.10(d) and (e). Since there are no more cyclic edges, the algorithm terminates,
and we have (G) = 1 + 2 + 2 + 3 = 8.

4.3 Number of Labelled Trees

Let us consider the problem of constructing all simple graphs with n vertices and m edges.
There are n(n 1)/2 unordered pairs of vertices. If the vertices are distinguishable from
each other (i.e., labelled
graphs), then the number of ways of selecting m edges to form the
!
n(n1)
2

graph is

Thus the number of simple labelled graphs with n vertices and m edges is
n(n1)
2

(A)

Clearly, many of these graphs can be isomorphic (that is they are same except for the
labels of their vertices). Thus the number of simple, unlabelled graphs of n vertices and m
edges is much smaller than that given by (A) above.
Theorem 4.20

The number of simple, labelled graphs of n vertices is 2

n(n1)
2

Proof The number of simple graphs of n vertices and 0, 1, 2, . . ., n(n 1)/2 edges are
obtained by substituting 0, 1, 2, . . ., n(n 1)/2 for m in (A). The sum of all such numbers
is the number of all simple graphs with n vertices.
Therefore the total number of simple, labelled graphs of n vertices is
n(n1)
2

n(n1)
2

by using the identity

n(n1)
2

+ . . .+

n(n1)
2
n(n1)
2

=2

n(n1)
2

     
 
k
k
k
k
+
+
+ . . .+
= 2k .
0
1
2
k

The following result was proved independently by Tutte [252] and Nash-Williams [167].
We prove the necessity and for sufficiency the reader is referred to the original papers of
Tutte and Nash-Williams.
Theorem 4.21 A simple connected graph G contains k pairwise edge-disjoint spanning
trees if and only if, for each partition of V (G) into p parts, the number m ( ) of edges of
G joining distinct parts is at least k(p 1).

92

Trees

Proof
Necessity Let G has k pairwise edge-disjoint spanning trees. If T is one of them, and if
= {V1 , V2 , . . ., Vp } is a partition of V (G) into p parts, then identification of each part Vi
into a single vertex vi , 1 i p, results in a connected graph G0 (possibly with multiple
edges) on {V1 , V2 , . . ., Vp }. Clearly, G0 contains a spanning tree with p 1 edges, and each
such edge belongs to T , and joins distinct partite sets of . Since this is true for each of the
k edge disjoint spanning trees of G, the number of edges joining distinct parts of is at
least k(p 1).
q
Cayley [46] in 1889 discovered the formula (Kn ) = nn2 . Clearly, the number of spanning trees of Kn is same as the number of non-label-isomorphic trees on n vertices. Several
proofs of this result have appeared since Cayleys discovery. Moon [164] has outlined ten
such proofs, and a complete presentation of some of these can also be found in Lovasz
[152]. Here we give two proofs, and the first is due to Prufer [212].
Theorem 4.22 (Cayley)

There are nn2 labelled trees with n vertices, n 2.

Proof Let T be a tree with n vertices and let the vertices be labelled 1, 2, . . ., n. Remove
the pendant vertex (and the edge incident to it) having the smallest label, say u1 . Let v1 be
the vertex adjacent to u1 . From the remaining n 1 vertices, let u2 be the pendant vertex
with the smallest label and let v2 be the vertex adjacent to u2 . We remove u2 and the edge
incident on it. We repeat this operation on the remaining n 2 vertices, then on n 3
vertices, and so on. This process completes after n 2 steps, when only two vertices are
left.
Let the vertices after each removal have labels v1 , v2 , . . ., vn2 . Clearly, the tree T
uniquely defines the sequence
(v1 , v2 , . . ., vn2 ).

(4.22.1)

Conversely, given a sequence of n 2 labels, an n-vertex tree is constructed uniquely as

follows. Determine the first number in the sequence
1, 2, 3, . . ., n,

(4.22.2)

that does not appear in (4.22.1). Let this number be u1 . Thus the edge (u1 , v1 ) is defined.
Remove v1 from sequence (4.22.1) and u1 from (4.22.2). In the remaining sequence of
(4.22.2), find the first number which does not appear in the remaining sequence of (4.22.1).
Let this be u2 and thus the edge (u2 , v2 ) is defined. The construction is continued till the
sequence (4.22.1) has no element left. Finally, the last two vertices remaining in (4.22.2)
are joined.
For each of the n 2 elements in sequence (4.22.1), we choose any one of the n numbers,
thus forming nn2 (n 2)-tuples, each defining a distinct labelled tree of n vertices. Since
each tree defines one of these sequences uniquely, there is a oneone correspondence
between the trees and the nn2 sequences.
q

Graph Theory

Example

93

Consider the tree shown in Figure 4.11. Pendant vertex with smallest label is

u1 . Remove u1 . Let v1 be adjacent to u1 (label of v1 is 1). Pendant vertex with smallest label
is 4. Remove 4. Here 4 is adjacent to 1. Pendant vertex with smallest label is 1. Remove 1.
Here 1 is adjacent to 3. Remove 3. Then 3 is adjacent to 5. Remove 6. So 6 is adjacent to 5.
Remove 5. Remove 7. 7 is adjacent to 5. So 5 is adjacent to 9. Sequence (v1 , v2 , . . ., vn2 )
is (1, 1, 3, 5, 5, 5, 9).

Fig. 4.11

Theorem 4.23 If D = [ di]n1 is the degree sequence of a tree, then the number of labelled
trees with this degree sequence is
(n 2)!
.
(d1 1)!(d2 1)! . . .(dn 1)!

Proof

We first observe that, when asking for all possible trees with the vertex label set

V = {v1 , v2 , . . ., vn } with degree sequence D = [ di]n1 , it is not necessary that di = d (vi ) and it

is not necessary that the sequence be monotonic non-decreasing.

Therefore we assume that D = [ di]n1 is an integer sequence satisfying the conditions di =
2(n 1) and di 1. We use induction on n. The result is obvious for n = 1, 2. For n = 2,
the sequence is [ d1, d2 ] and the only degree sequence in this case is [ 1, 1 ]. Clearly, there is
only one labelled tree with this degree sequence.
Also,

(n 2)!
(2 2)!
=
= 1.
(d1 1)! . . .(dn 1) (1 1)!(1 1)!

Now, assume that the result is true for all sequences of length n 1. Let D = [ di]n1 be an n
length sequence. By assumption there is a di = 1 and let it be dn = 1. Let Tn be a tree realising
D = [ di ]n1 . Now, removing vn , we get a tree Tn1 on the vertex set {v1 , v2 , . . ., vn1 } with
degrees d1 , . . ., d j1, d j 1, d j+1, . . ., dn1 , where v j is the vertex to which vn is adjacent in

94

Trees

Tn . Clearly, the converse is also true. Therefore, by induction hypothesis, the number of
trees Tn1 is
(n 3)!
(d1 1)!...(d j1 1)!(d j 1 1)!(d j+1 1)!...(dn1 1)!
=

(n 3)!(d j 1)
(d1 1)!...(d j1 1)! [(d j 1) (d j 2)!] (d j+1 1)!...(dn1 1)!

(n 3)!(d j 1)
(d1 1)!...(d j 1)!...(dn1 1)!(dn 1)!
(n 3)!(d j 1)

(d j 1)!
j=1

Since v j is any one of the vertices v1 , . . ., vn1 , the number of trees Tn is

n1 (n3)!(d 1)
j

j=1

(d j 1)!

j=1

(n3)!

(d j 1), as dn = 1 and dn 1 = 1 1 = 0

(d j 1)! j=1

j=1

(n3)!
n

(d j 1)!

(n 2), since (d j 1) = 2(n 1) n = n 2

j=1

j=1

(n2)!
n

(d j 1)!

j=1

Now, we use Theorem 4.22 to obtain (Kn ) = nn2 , which forms the second proof of
Cayleys Theorem.

Second Proof of Theorem 4.22

degree sequence [di ]n1 is
(n 2)!
n

We know the number of labelled trees with a given

(d j 1)!
j=1

The total number of labelled trees with n vertices is obtained by adding the number of
labelled trees with all possible degree sequences.

Graph Theory

95

Therefore, (Kn ) =
di 1
n
di =2n2
i=1

(n2)!
n

(d j 1)!

j=1

Let di 1 = ki . So di 1 gives di 1 0, or ki 0.
n

i=1

i=1

i=1

Also, ki = (di 1) = di n = 2n 2 n = n 2.
Thus, (Kn ) =

(n 2)!
(n 2)! k 1 k2
=
1 1 . . .1kn
k1 !k2 ! . . .kn ! k 0 k1 !k2 !kn !

ki 0
n
ki =n2
1

i
n
ki =n2
1

= (1 + 1 + . . .+ 1)n2 , by multinomial theorem.

Hence, (Kn ) == nn2 .

Note

The multinomial distribution is given by

n
n!
x
x x
p11 p22 . . . pkk = (p1 + p2 + . . . + pk )n , where xi = n.
x1 !x2 ! . . .xk !
i=1

4.4 The Fundamental Cycles

Definition: Let T be a spanning tree of a connected graph G. Let T be the spanning
subgraph of G containing only the edges of G which are not in T (i.e., T is the relative
complement of T in G). Then T is called the co tree of T in G. The edges of T are called
branches and the edges of T are called chords of G relative to the spanning tree T .
Theorem 4.24 If T is a spanning tree of a connected graph G and f is a chord of G
relative to T , then T + f contains a unique cycle of G.

Proof Let f = uv. Then there is a unique u v path P in T . Clearly, P + f is a cycle of

G, since T is acyclic, any cycle C of T + e should contain e, and C e is a u v path in T .
Since there is a unique path in T , T + e contains a unique cycle of G.
q
Remarks

1. If f 1 and f 2 are two distinct chords of the connected graph G relative to a spanning tree
T , then there are two unique distinct cycles C1 and C2 of G containing respectively f 1
and f 2 .
2. If e E(G) and T is a spanning tree of G, then T + e contains a unique cycle of Kn .

96

Trees

Definition: Let G be a connected graph with n vertices and m edges. The number of
chords of G relative to a spanning tree T of G is m n + 1 = . The distinct cycles of a
connected graph G corresponding to the distinct chords of G relative to a spanning tree T
of G are said to form a set of fundamental cycles of G.
If G is a disconnected graph with k components G1 , G2 , . . ., Gk and Ti , 1 i k, are a set
of k spanning trees of Gi , then the union of the set of fundamental cycles of Gi with respect
to Ti is a set of fundamental cycles for G. It is to be noted that different spanning trees give
different sets of fundamental cycles.
The following result characterises cycles in terms of the set of all spanning trees.
Theorem 4.25 Any cycle of a connected graph G contains at least one chord of every
spanning tree of G.

Proof Let C be a cycle and assume the result is not true. So there exists a spanning
tree T of G such that C is contained in the edge set E(G) E(T ), where T is the cotree
of G corresponding to T . This means that the tree T contains the cycle C, which is a
contradiction.
q
Theorem 4.26 A set of edges C of a connected graph G is a cycle of G if and only if it
is a minimal set of edges containing at least one chord of every spanning tree of G.

Proof Let C be a cycle of G. Then it contains at least one chord of every spanning tree
of G. If C0 is any proper subset of C, then C0 does not contain a cycle and is a forest.
A spanning tree T of G can therefore be constructed containing C0 . Clearly, C0 does not
contain any chord of T . Thus no proper subset of C has the stated property, proving that C
is minimal with respect to the property.
To prove sufficiency, let C be minimal set with the stated property. Then C is not acyclic.
Therefore C contains at least a cycle C0 . But by the necessary part, C0 is minimal with
respect to the property and hence C0 = C, that is, C is a cycle.
q

4.5 Generation of Trees

Definition: Let T1 and T2 be two spanning trees of a connected graph G and let there
be edges e1 T1 and e2 T2 such that T1 e1 + e2 = T2 (and hence T2 e2 + e1 = T1 ). The
transformation T1 T2 is called an elementary tree transformation (ETT), or a fundamental
exchange. If e1 and e2 are adjacent in G, then the ETT is called a neighbour transformation
(NT). If e is a pendant edge of T1 (and hence e2 is a pendant edge of T2 ) the ETT is called
a pendant-edge transformation (PET) or an end-line transformation.
Definition: Let I be the collection of all spanning trees of a connected graph G. Let
Tr(G) be the graph whose vertices ti correspond to the elements Ti of I , and in which ti
and t j are adjacent if and only if there is an ETT between Ti and T j , that is, if and only
if E(Ti )E(T j ) = {ei , e j }. Then Tr(G) is called the tree graph of G. The distance d(Ti , T j )

Graph Theory

97

between the spanning trees Ti and T j of G is defined to be the distance between ti and t j in
Tr(G).
Theorem 4.27

The tree graph Tr(G) of a connected graph is connected.

Proof Let G be a connected graph with n vertices and let Tr(G) be its tree graph. To
prove that Tr(G) is connected, it is enough to show that any two spanning trees of G can be
obtained from each other by a finite sequence of ETTs.
Let T and T 0 be two distinct spanning trees of G. Then there is a set S = {e1 , e2 , . . ., ek }
of some k edges of T which are not in T 0 . Since a spanning tree has n 1 edges, there is
a corresponding set S0 = {e01 , e02 , . . ., e0k } of edges of T 0 which are not in T . Thus T + e01
contains a unique fundamental cycle T e01 . As T 0 is a tree, at least one edge of T e01 (which
is a branch of T ) will not be in T 0 and thus is a member of S. Without loss of generality,
let this edge be e1 . Define T1 = T e1 + e01 . Then T1 can be obtained from T by an ETT and
therefore T1 and T 0 have one more edge in common.
Repeating this process k 1 more times, we get a sequence of spanning trees T0 =
T, T1 , T2 , . . ., Tk1 , Tk = T 0 such that there is an ETT Ti Ti+1 , 0 i k 1.
q
Theorem 4.28 An elementary tree transformation can be obtained by a sequence of
neighbour transformations.

Proof Let T and T 0 = T x + y be spanning trees of the graph G, where x and y are
non-adjacent edges of G. Then we can choose a set of edges e1 , e2 , . . ., ek such that
x, e1 , e2 , . . ., ek , y is a path in T + y. Define T1 = T x + e1 and Ti = Ti1 + ei1 + ei , 2 i k
and Tk+1 = Tk ek + y. Then Tk+1 = T 0 , and is obtained from T by a sequence of k + 1
neighbour transformations through the intermediate trees Ti , 1 i k.
q
Definition: A spanning tree of a graph G corresponding to a central vertex of the tree
Tr(G) is called a central tree.

The set of diameters of the spanning trees of a connected graph G is the tree diameter
set of G. A set of positive integers is a feasible tree diameter set if it is the tree diameter set
of some graph. For example, the graph in Figure 4.12 has one spanning tree of diameter
seven and all others of diameter five.

Fig. 4.12

The girth g(G) of a graph G is the length of a smallest cycle of G. A cycle of smallest
length is called a girdle of G. The circumference c(G) of a graph G is the length of the
longest cycle of G. A cycle of maximum length is called a hem of G.

98

Trees

Let n ( , g) denote the minimum order (minimum vertices) of a graph with minimum
degree at least ( 3) and girth at least g ( 2). Let n (, g) denote the maximum order of
a graph with degree at most and girth at most g.
The following upper bound for n ( , g) can be found in Bollobas [29].
Theorem 4.29 (Bollobas)

n ( , g) (2 )g .

Proof Clearly, n ( , g) denotes the minimum order of a graph with minimum degree at
least ( 3) and girth at least g ( 2). Therefore we construct a graph with atmost (2 )g
vertices with these properties. Let n = (2 )g .
Consider all graphs
with vertex set V = {1, 2, . . ., n} and having exactly n edges.
Since there are n2 possible positions to accommodate these n edges, the number of
such graphs
=

n

2
.

Among the n available vertices, the number of ways an h-cycle can be formed is
=

1 n
(h 1)!
2 h

Obviously,

1 n
1
(h 1)! < nh .
2 h
2h

The number of graphs in the set which contain a given h-cycle is

=


n

2 h .

nh

Hence the average number of cycles of length at most g 1 in these graphs

g1

1
< nh
2h
h=3

 n

n

2 h
2

nh

g1

< (2 )h < (2 )g = n.
h=3

Since the average is less than n, there is an element in the set with value less than or
equal to n 1. Thus there is a graph G on n vertices with n edges and at most n 1
cycles of length at most g 1. Removing one edge from each of these cycles, we get
a graph G0 with girth at least g. The number of edges removed is atmost n 1, so that
m(Go ) n (n 1) n ( 1) + 1 and n (G0 ) = n. Thus G0 G 1 , and hence G0 contains

Graph Theory

99

a subgraph H with (H) . By construction, g(H) g and n(H) n = (2 )g . Thus we

have constructed a graph H with the desired properties.
q
If G is a graph with at least n0 vertices and at least n0 n(G)

Note

then G contains a subgraph H with (H) n0 + 1.



We denote by Gno = G : n(G) > n0 , m(G) n0 .n(G)


n0 + 1
+ 1 edges,
2



n0 + 1
+1 .
2

The following lower bound for n ( , g) is due to Tutte [248].

Theorem 4.30 (Tutte)

g1

( 1) 2 ,
g2
n( , g)

2( 1) 2 1

,
2

if g is odd,
if g is even.

Proof

i. Let g be odd, say g = 2d + 1. Then clearly the diameter of G is at least d . Let v be a

vertex with eccentricity at least d . Consider the neighbourhoods
Ni = Ni (V ), 1 i d = (g 1)/2.

Obviously, no vertex of Ni is adjacent to more than one vertex of Ni1 , because

otherwise, there will be a cycle of length 1 2i < g. Similarly, there is no edge in
< Ni >.
Therefore, for every u Ni , we have
|N(u) Ni1|= 1, |N(u) Ni+1|= d(u) 1 and
T

(4.30.1)

|Ni+1 | = {d(u) 1} ( 1) |Ni | .

uNi

As V {v}

d
SS

Ni (v), therefore

i=1

n 1 + |Ni | 1 + + ( 1) + ... + ( 1)d1

i=1

= 1+

( 1)d 1 =
2
n

( 1)

g1
2

100

Trees

ii. Let g be even, say g = 2d . Then again the diameter is at least d . Let xy be an edge of
G and let
Si = {v V : d(x, v) = i, or d(y, v) = i}, for 1 i d 1 and S0 = {x, y}.

The girth requirement forces that there are no edges in < Si >, for 1 i d 2,
and that each vertex of Si be adjacent to at most one vertex of Si1, for 1 i d 1.
Thus, for each u Si , we have |N(u) Si1| = 1, |N(u) Si+1| = d(u) 1 and
|Si+1 | = (d(u) 1) ( 1) |Si |.

(4.30.2)

u Si

Since V {x, y}

S d1
S

Si ,

i=1

d1

d1

i=0

i=0

n=

|Si| 2 ( 1)i =

i
g
2 h
( 1) 2 1 .
2

By using arguments as in Theorem 4.30 and by replacing by , we obtain the following

result.
Theorem 4.31

g1

( 1) 2 2

h g2 i
n(,
g)
2 ( 1) 2 1

if g is odd,
if g is even.

Definition: A k-regular graph with girth g and with minimum order n(k, g) is called a (k,
g)-cage.

g1

k(k 1) 2 2

h k 2g
i
The integer n0 =

2 (k 1) 2 1

,
k2

i f g is odd ,
i f g is even ,

is called the Moore bound for a k-regular graph with g.

4.6 Helly Property

Definition: ATfamily {Ai : i I} of subsets of T
a set A is said to satisfy the Helly property
if J I , and Ai A j 6= , for every i, j J , then
A j 6= .
jJ

Graph Theory

101

The following result is reported by Balakrishnan and Ranganathan [13].

Theorem 4.32

A family of subtrees of a tree satisfies the Helly property.

Proof

Let = {Ti : i I} be a T
family of subtrees of a tree T . Suppose for all i, j J
I, Ti T j 6= . We have to prove
T j 6= . If some tree Ti , i J , is a single vertex tree

{v} (that is, K1 ), then clearly,

T jJ

T j = {v}. So assume that each tree Ti T with i J has at

jJ

least two vertices.

We induct on the number of vertices of T . Suppose the result is true for all trees with at
most n vertices and let T be a tree with (n + 1) vertices. Let v0 be an end vertex of T and u0
its unique neighbour in T . Let Ti0 = Ti v0 , i J and T 0 = T v0 . By induction hypothesis,
the result is true for the tree T 0 . Also, Ti0 T j0 6= , for any i, j J . In fact, if Ti and T j have
a vertex u (6= v0 ) in common then Ti0 and T j0 also have u in common, whereas if Ti and T j
have v0 in common, then
T and T j have u0 also in common, and so do Ti0 and T j0 . Hence by
T
T i0
q
induction hypothesis, T j 6= and therefore T j 6= .
jJ

jJ

4.7 Signed Trees

The following result by Yan et al. [271] characterises signed degree sequences in signed
trees.
Theorem 4.33 Let D = [di ]n1 be an integral sequence of n 2 terms and let D has n+
positive terms, n0 zero and n negative terms. Let = 1 if n+ n > 0, and = 0, otherwise.
Then D is the signed degree sequence of a signed tree if and only if (i) to (iv) hold.
n

i.

di 2n 2(mod4).
i=1
n

ii.

|di| 2n 2 2n0.
i=1
n

iii.

|di| + 2 |di| 2n 2 4 + 4p .
i=1

di >0

iv.

|di| + 2 |di| 2n 2 4 + 4p+ .

i=1

di >0

Proof Note that condition (iv) for D is same as condition (iii) for D. The necessity of
the theorem follows from the fact that m = n 1 and Lemmas 2.2, 2.3 and 2.4.
We prove the sufficiency by induction on n. For n = 2, by (i) and (iii), d1 = d2 = 1 or
1. Therefore D is the signed degree sequence of K2 with positive edge or a negative edge.
Assume that the theorem is true for n 1. Let n 3.

102

Trees

By (ii), D has at least two terms in which |di |= 1. After rearranging the terms in D or
taking D, we may assume without loss of generality that dn = 1 and one of the following
holds.
1. |di |= 1, for 1 i n, d1 0 and d1 = 0, if n0 > 0.
2. d1 2.
3. di 1 but di 6= 1 for 1 i n and d1 = 0 and = 1.
4. di = 1 or di 2, for 1 i n and d1 = = 1.
0

For any of the above, consider the sequence D0 = [di0 ]ni , where n0 = n 1 and di0 = d1 1
and di0 = di , for 2 i n 1!.
Note that

n0

i=1

i=1

di0 = di

2 (2n 2) 2 2n0 2(mod4), that is, (i) holds for D0 . We

check conditions (ii) to (iv) for D0 according to the four cases above.
Case 1

In this case, |di0 | 1, for 1 i n 1, we have

n0

|d 0 | = n0+ , |di0 | = n0 .
|di0| = n0+ + n0 ,
0
0

i=1

di >0

di <0

Thus (ii) to (iv) holds for D0 as n0+ + n0 2.

Case 2

In this case, since d1 2 and dn = 1, we have

n0 = n 1, n0+ = n+ 1, n00 = n0 , n0 = n , 0 = ,
n0

i=1

i=1

|di0 | = |di0 | 2, |di0 | = |di0 |.

|di0| = |di0| 2,
0
0
0
0
di >0

di >0

di <0

di <0

Therefore (ii) to (iv) holding for D imply that (ii) to (iv) hold for D0 .
Case 3

In this case, since d1 = 0 and dn = 1, we have

n0 = n 1, n0+ = n+ 1, n00 = n0 1, n0 = n + 1, 0 ,
n0

i=1

i=1

|di0 | = |di0 | + 1.
|di0| = |di|, |di0| = |di0| 1,
0
0
di >0

di >0

di <0

di <0

So (ii) and (iii) holding for D imply that (ii) and (iii) hold for D0 . Since di0 1 for
n0

1 i n 1, |di0 | = n0+. By (iii) for D and the fact that di 2 when di < 0,
di0

n0

n+ + 6n |di | + 2 |di| 2n 2 4 + 4n = 2n+ + 2n0 + 6n 6,

i=1

di <0

Graph Theory

103

and so 6 n+ + 2n0 . Therefore, 3 n0+ + 2n00 and then 4 2n0+ + 2n00 .

This together with (ii) for D0 and |di0 | = n0+ implies (iv) for D0 .
di0 >0

Case 4

In this case, since d1 = dn = 1, therefore

n0 + n 1, n0+ = n+ 2, n00 = n0 + 1 = 1, n0 = n, 0 ,
n0

i=1

i=1

|di0 | = |di0 | 2, |di0 | = |di0 |.

|di0| = |di| 2,
0
0
0
di >0

di >0

di <0

di <0

(iii) for D implies that (iii) holds for D0 . As in the argument for Case 3, we have |di | = n0+
di0 >0

and 6 n+ +2n0 . Therefore, 4 n0+. Adding 2 |di0 | = 2n0+ to the equality in (iii) for D0 and
di0 >0

dividing the resulting equality by 3, we get (ii) for D0 as 2n00 2n0+. Adding 2 |di0 | = 2n0+
D0 ,

D0

4 0

2n00 + 2n0+.

di0 >0

to the equality in (ii) for

we get (iv) for as
From the above discussion, D0 satisfies (i) to (iv). By the induction hypothesis, there
exists a signed tree T 0 with the vertex set {v1 , v2 , . . ., vn1 } and signed degree T 0 (vi ) = di0 ,
for 1 i n 1. Suppose T is the signed tree obtained from T 0 by adding a new vertex vn
and a new positive edge v1 v+
q
n , then T has a signed degree sequence D.
Corollary Let D = [di]n1 be an integral sequence of n 3 terms. Let D has at least two
terms in which |di |= 1, |dn |= 1 and one of the following condition holds.
1. |di | 1, for 1 i n, di 0, and d1 = 0 if no > 0.
2. d1 2.
3. di 1 but di 6= 1 for 1 i n, and d1 = 0 and = 1
4. di = 1 or di 2 for 1 i n, and d1 = = 1.
Then D is the signed degree sequence of a signed tree if and only if D0 = [d1 1, d2 , . . ., dn1 ]
is the signed degree sequence of a signed tree.

4.8 Exercises
1. Draw all unlabelled trees with seven and eight vertices.
2. Draw a tree which has radius five and diameter ten.
3. If a tree has an even number of edges, then show that it contains at least one vertex
of even degree.

104

Trees

4. If the maximum degree of a vertex in a tree is , then show that it has pendant
vertices.
5. If T is a tree such that every vertex adjacent to a pendant vertex has degree at least
three, then prove that some pair of pendant vertices in T has a common neighbour.
6. Show that a path is its own spanning tree.
7. Prove that every tree is a bipartite graph.
8. If for a simple graph G, m(G) n(G), prove that G contains a cycle.
9. Show that for a unicentral tree, d = 2r, and for a bicentral tree, d = 2r 1.
10. Prove that if Kr, s is a tree, then it must be a star.
11. How many spanning trees does K4 have?
12. Prove that each spanning tree of a connected graph G contains all the pendant edges
of G.
13. Prove that each edge of a connected graph G belongs to at least one spanning tree of
G.

5. Connectivity
In a connected graph there is at least one path between every pair of its vertices. If in a
graph, it happens that by deleting a vertex, or by removing an edge, or performing both,
the graph becomes disconnected, we can say that such vertices or edges hold the whole
graph, or in other words have the property of destroying the connectedness of a graph. For
example, consider a communication network which is modelled as the graph G shown in
Figure 5.1, where vertices correspond to communication centers and the edges represent
communication channels. Clearly, deletion of vertex v results in the breakdown of the communication. This implies that in the above communication network, the center represented
by vertex v has the property of destroying the communication system and thus communication network depends on the connectivity. We start this chapter with the following
definitions.

Fig. 5.1

5.1 Basic Concepts

Cut vertex: Let G be a graph with k(G) components. A vertex v of G is called a cut
vertex of G if k (G v) > k (G). For example, in the graph of Figure 5.2, the vertices u and
v are cut vertices.
Cut edge: An edge e of a graph G is said to be a cut edge if k (G e) > k(G). In the graph
of Figure 5.2, e and f are cut edges.

106

Connectivity

Fig. 5.2

The following observations are the immediate consequences of the definitions introduced above.
1. Removal of a vertex may increase the number of components in a graph by at least
one, while removal of an edge may increase the number of components by at most
one.
2. The end vertices of a cut edge are cut vertices if their degree is more than one.
3. Every non-pendant vertex of a tree is a cut vertex.
We now give the first result which characterises cut vertices.
Theorem 5.1 If G(V, E) is a connected graph, then v is a cut vertex if there exist vertices
u, w V {v} such that every u w path in G passes through v.

Proof Let G(V, E) be a connected graph and let v be the cut vertex of G. Then G v
is disconnected. Let G1 , G2 , . . ., Gk be the components of G v. Let U = V (G1 ) and W =
k
S

V (Gi ). Also, let u U and w W , and to be definite, let w V (Gi ), i 6= 1. If there is a

i=2

u w path P in G not passing through v, then P connects u and w in G v also. Therefore

G1 Gi is a single component in G v, contradicting our assumption. Thus every u w path
in G passes through v.
Conversely, let there be vertices u, w V {v} such that every u w path in G passes
through v. Then there is no u w path in G v. Therefore u and w belong to different
components of G v. Thus G v is disconnected and v is a cut vertex of G.
q

The following result characterises cut edges.

Theorem 5.2

For a connected graph G, the following statements are equivalent.

i. e is a cut edge of G.
ii. If e = ab, there is a partition of the edge subset E {e} as E1 E2 with a V (< E1 >)
and b V (< E2 >) such that for any u V (< E1 >) and any w V (< E2 >), every u w
path contains e.

Graph Theory

107

iii. There exist vertices u and w such that every u w path in G contains e.
iv. e is not a cycle edge of G.

Proof (i) (ii). Let e be a cut edge of G. So G e is disconnected. Let G1 and G2 be

two components of G e and E1 = E(G1 ) and E2 = E(G2 ). If u V (G1 ) and w v(G2 ) exist
such that there is a u w path P in G which does not contain e, then u and w are connected
in G e by the path P. This implies that G1 G2 , that is, G e is connected, contradicting
the hypothesis. This proves (ii).
(ii) (iii). Obvious.
(iii) (iv). Suppose e lies on a cycle C. Then C e gives an a b path Q not containing e.
With vertices u and w following the condition given in (iii), let P be any u w path. Without
loss of generality, assume that a and b occur in that order in P. Let u0 and w0 be the first
and last vertices that P has in common with C (the possibility of these coinciding with a, b,
u or w is not ruled out). Then Pu, uo Qu0 , w0 Pw0 , w is a u w path P0 of G which does not
contain e, contradicting (iii). (See Figure 5.3(a), where broken curves represent path Q and
thick curves represent path P.)

Fig. 5.3(a)

(iv) (i). Let e be not a cyclic edge of G. We have to prove that e is a cut edge of G, that
is, G e is disconnected. Assume G e is connected. Then there is an a b path P in G e.
But then P e is a cycle containing e, which contradicts (iv).
q
Block: A block is a connected graph which does not have any cut edge. We observe that
a block does not have any cut edge. The graph K2 = ({a, b}, e) does not have a cut vertex
and hence is a block. However, e is a cut edge in this case. We call K2 a trivial block. All
other blocks are non-trivial.
Separable graph: A connected graph with at least one cut vertex is called a separable
graph. A block of a graph G is a maximal graph fH of G such that H is a block. That is, H
has no cut vertex, but for any v V (G) V (H), hV (H) {v}i is either a disconnected graph
or a separable graph.
The next result characterises blocks.
Theorem 5.3

For a connected graph G, the following are equivalent.

i. G is a non-trivial block.

108

Connectivity

ii. Any two vertices of G lie on a cycle.

iii. Given any vertex u and any edge vw, there is a cycle of G containing both.
iv. Given any pair of edges e = uv and e0 = u0 v0 , there is a cycle of G containing both.
v. Given any pair of vertices u and u0 and any edge e = vw, there is a u u0 path of G
containing e.

Proof
a. (i) (ii). The proof is by induction on the distance between the vertices. If d(u, v) =
1, then uv is an edge, and since G is a block, uv is not a cut edge. Hence uv is a cyclic
edge, and so u and v lie on a cycle. Now, for the induction hypothesis, we assume
that if u is any vertex, then any vertex v0 at a distance at most k 1 from u lies on a
cycle with u.
Let v be a vertex at a distance k from u. We prove that u and v lie on a cycle. Let P
be a shortest u v path and v0 the nearest vertex on P from u. By induction hypothesis
there is a cycle C containing u and v0 . Since v0 is not a cut vertex of G, there is a u v
path Q not passing through v0 . Let z be the last vertex from u that Q has in common
with C. Then Cuv0 {v0 v} Qvz Czu is a cycle of G containing u and v. (Here Cuv0 is
the uv0 segment of C not containing z, and Czu is the zu segment of C not containing v0 )
(Fig. 5.3(b)).

Fig. 5.3(b)

b. (ii) (i). Let any two vertices of G lie on a cycle. We prove that G is a non-trivial
block, that is, G has no cut vertex. Assume to the contrary that G has a cut vertex
u. Then there are vertices v and w such that every v w path passes through u. But
then there is no cycle containing v and w, which is a contradiction. Thus G has no cut
vertex.
c. (ii) (iii). Let any two vertices of G lie on a cycle. Let vertex u and edge vw be
given. So by (b), G is a block and therefore vw is not a cut edge.
Let C be a cycle containing vw. If C contains u, the proof is complete. If not, by
(ii), there is a cycle Z containing u and v. Taking any orientation of Z , let x and y be
the first and last vertices from u that Z has in common with C. Then the u x segment

Graph Theory

109

of Z , the x y segment of C containing vw and the yu segment of Z constitute a cycle

of G containing u and vw.
d. (iii) (ii). Let u and v be any two vertices. Since v cannot be an isolated vertex,
there is an edge vw. By (iii) there is a cycle containing u and vw (Fig. 5.4).

Fig. 5.4

e. (iii) (iv). Let uv and u0 v0 be the given edges. By (iii) there is a cycle C through
u containing u0 v0 . If it passes through v, then the u v segment of C containing u0 v0
and the edge vu constitute a cycle as required. If not, then as the earlier implications
show that G is a block, u is not a cut vertex, and hence there is a v v0 path P in G not
passing through u (Fig. 5.5).

Fig. 5.5

Let w be the first vertex from v that P has in common with C. Then v w segment of
P, the w u segment of C containing u0 v0 and the edge uv constitute a cycle of G as
desired (Fig. 5.6).

Fig. 5.6

f. (iv) (iii). This can be proved as in (d).

g. (iv) (v). Let u and u0 be the vertices and vw the given edge. If uv0 is also an edge,
then there is nothing to prove. If not, since u is not an isolated vertex, there is an edge

110

Connectivity
uv1 and by (iv) there is a cycle C containing uv1 and vw. By previous implications, G
is a block and hence u is not a cut vertex. Therefore there is a u0 w path P not passing
through u. Let x be the first vertex from u0 that P has in common with C. Then the
u0 x segment of P and the xu segment of C containing vw constitute a path as desired

(Fig. 5.7).

Fig. 5.7

h. (v) (iv). Obvious.

Remark Property (iv) of the above theorem can be used to define an equivalence relation
on the edge set E of a graph G by e f if and only if e and f lie on a common cycle in G.
The equivalence classes are simply the blocks of G and the edge set E is partitioned into
blocks. These blocks are joined at cut vertices, two blocks having at most one vertex in
common.

5.2 BlockCut Vertex Tree

Let B be the set of blocks and C be the set of cut vertices of a separable graph G. Construct a
graph H with vertex set B C in which adjacencies are defined as follows: ci C is adjacent
to b j B if and only if the block b j of G contains the cut vertex ci of G. The bipartite graph
H constructed above is called the block-cut vertex tree of G.
Example Consider the graph in Figure 5.8. The blocks are b1 = < 1, 2 >, b2 = < 2, 3, 4 >,
b3 = < 2, 5, 6, 7 >, b4 = < 7, 8, 9, 10, 11 >, b5 = < 8, 12, 13, 14, 15 >, b6 = < 10, 16 >,
b7 = < 10, 17, 18 > and cut vertices are c1 = 2, c2 = 7, c3 = 8, c4 = 10.

Fig. 5.8

Graph Theory

111

End block:
of G.

A block of a graph G containing only one cut vertex is called an end block

We have the following result on blocks.

Theorem 5.4 Every separable graph has at least one cut vertex and therefore has at least
two end blocks.

Proof A separable graph G has at least one cut vertex and therefore has at least two
blocks. Thus its block-cut vertex tree T has at least three vertices. Now, for any separable
graph the end blocks correspond to the pendant vertices of its block-cut vertex tree. Also,
any tree with at least two vertices has two vertices of degree one. Thus the block-cut vertex
tree T has at least two pendant vertices. Hence G has at least two end-blocks.
q
The next result is due to Harary and Norman [109].
Theorem 5.5

The center of any connected graph G lies on a block of G.

Proof If not, let B1 , B2 be blocks of G containing central vertices. If b1 , b2 are the vertices
of the block-cut vertex tree T of G corresponding to B1 and B2 , then there is at least one
vertex c in the unique b1 b2 path of T , corresponding to a cut-vertex c of G. So there are
two components G1 and G2 of G c such that B1 c G1 and B2 c G2 . Let c be an
eccentric vertex of c in G and P be a c c path of G having length e(c). Then at least one
of the components G1 and G2 , say G2 , contains no vertex of P. Let s be a central vertex in
G2 and Q be a shortest s c path in G. Then Q P is clearly an s c path in G and thus
d(s, c)
= d(s, c) + e(c). Therefore e(s) > e(c), contradicting the fact that s is a central vertex.
Thus the center of G lies in a single block. Hence the center of any connected graph G lies
on a block of G.
q
Definition: If G is a separable graph and c a cut-vertex of G, then a maximal connected
subgraph of G containing c in which c is not a cut-vertex is called a branch of G at c. The
induced subgraph hCi on the central vertices of G is called the central graph of G. If G has
a unique central vertex c, then G is said to be a unicentric graph. The unique block B of G
to which the center c of G belongs is called the central block of G. This is unambiguously
defined except when G is unicentric and the unique central vertex is a cut-vertex of G.
When B = hCi = G, then G is called a self-centered graph. If the unique central vertex c of
G is a cut-vertex of G, the unique block of any of the branches of G at c in which c has an
eccentric vertex c may be taken as the chosen central block of G.
We note that the central graph of a tree is either K1 or K2 . Buckley, Miller and Slater [53]
have studied graphs with specified central graphs. The following result is atributed to
Hedetniemi and is reported in Parthasarathy [180].
Theorem 5.6 For any graph H there exists a graph G with 4 more vertices such that H is
the central graph of G.

112

Connectivity

Proof Take two new vertices v and w, and join each to every vertex of H . Take two other
vertices x and y, join x to v, and y to w. Then in the resulting graph G, e(x) = e(y) = 4, e(v) =
e(w) = 3 and e(u) = 2, for every vertex u V (H). Thus H is an induced subgraph of G and
the central graph of G.
q

5.3 Connectivity Parameters

Assume that a graph does not get disconnected by deleting a single vertex, or by removing
a single edge. A natural question then arises: what is the minimum number of vertices or
edges required to disconnect a graph? This and other related questions are answered in this
section. Before proceeding, we have the following definitions.
Definition: Let G = (V, E) be a graph. A subset S of V E is called a disconnecting set
of the graph G if k(G S) > k(G), or G S is the trivial graph.
If a disconnecting set S is a subset of V , it is called a vertex cut of G, and if it is a subset
of E it is called an edge cut of G. If a disconnecting set S contains vertices and edges it is
called a mixed cut.
Example For the graph shown in Figure 5.9, S = {3, e3 } is a mixed cut, S = {3} is a
vertex cut and S = {e1 , e3 } is an edge cut.

Fig. 5.9

A mixed cut/vertex cut/edge cut S is minimal if no proper subset of S has the same
property as S. A mixed cut/vertex cut/edge cut S is minimum if it has least cardinality
among all such minimal sets. A minimal vertex cut is called a knot and a minimum vertex
cut is called a clot. The cardinality of a clot is called the vertex-connectivity number, or
clot number of the graph G and is denoted by (G).
A minimal edge cut is called a bond and a minimum edge cut is called a band. The
cardinality of a band is called the edge-connectivity number, or band number of the graph
G, and is denoted by (G).
The minimum cardinality of a mixed set is denoted by (G).
Let S be a disconnected set of the graph G(V, E). Let vertices s and t be in the same
component of G, but in different components of G S. Then S is called an s t separating
set in G. Minimal s t separating vertex cut is called an s t knot, and the minimum s t
separating vertex cut is called an s t clot. Minimal s t separating edge cut is called an
st bond, and the minimum st separating edge cut is called an st band. The cardinality

Graph Theory

113

of an s t clot is called the s t clot number and is denoted by (s, t), and the cardinality
of an s t band, called the s t band number, is denoted by (s, t). The cardinality of a
minimum s t separating mixed cut is denoted by (s, t).
The following result gives vertex connectivity of complete graphs and an upper bound
for non-complete graphs.
Theorem 5.7

(Kn ) = n 1. If G is incomplete, then (G) n 2.

Proof
i. Clearly, Kn is a connected graph with n vertices. Deletion of a vertex v1 keeps the
graph G v1 connected. Clearly, G v1 has n 1 vertices. Deleting one more vertex,
say v2 from G v1 , gives a graph G {v1 , v2 }, which is again connected. Continuing
this process, we observe that deleting any number of ertices i, 1 i n 1 does not
disconnect the graph, but deleting exactly n 1 vertices gives a trivial graph with one
vertex. Thus, (Kn ) = n 1.
ii. Let G be an incomplete graph with n vertices. Then there are at least two vertices,
say vi and v j which are not adjacent. If there is exactly one edge vi v j missing, then
deleting the n 2 vertices other than vi and v j disconnects the graph. So in this case
(G) = n 2. If there are more edges missing, then clearly (G) < n 2 (Fig. 5.10).
Hence, (G) n 2.
q

Fig. 5.10

The following result is obvious.

Theorem 5.8

(G) min (s, t), (G) = min (s, t), (G) = min (s, t).
s, tv

s, tv

s, tv

st
/E

Cut of a graph: Let G(V, E) be a graph and let A be any non-empty sub-set of the vertex
set V . Let A = V A. The set of all edges with one end in A and the other end in A , denoted
is called a cut of G. The concept of a cut of a graph is intermediate between that
by [A, A]
of an edge cut and a bond.
We note that every cut is an edge cut, but the converse is not true. Consider the graph in
Figure 5.11. Here F = {e1 , e2 , e3 , e4 , e5 } is an edge cut, but F is not a cut.

114

Connectivity

Fig. 5.11

Also, every bond is a cut, but the converse is not true. This is illustrated by the graph in
= {e1 , e2 , e3 , e4 }. Here
Figure 5.12. Let A = {1, 2, 3, 4}. Then A = {5, 6, 7, 8}. So [A, A]

[A, A] is a cut but not a bond.

Fig. 5.12

Theorem 5.9

Every minimal cut is a bond and every bond is a minimal cut.

be a cut of G. Assume C to be a minimal

Proof Let G(V, E) be a graph and let C = [A, A]
cut. Then no subset
of the edges of C is a cut and this implies that G C has only two
components hAi and A . Therefore C is a bond.
Conversely, let F be a bond. Then G F has only two components, say C1 and C2 . Then
F = [V1 , V2 ], with V2 = V1 . Thus F is a cut and hence a minimal cut.
q

Theorem 5.10

Every cut is a disjoint union of minimal cuts.

Proof Let G(V, E) be a graph

 and let C = [A, A] be a cut of G. Let C be not a minimal cut.

Then at least one of hAi, or A has more than one component.
Assume
C1 , C2 , . . ., Cr to be the components of hAi and C10 , C20 , . . ., Cs0 be the components


of A . (Clearly, at least one of r and s is greater than one.) Let Ci be coalesced to vertices
ci , 1 i r and Ci0 be coalesced to vertices c0i , 1 i s, and let H be the simple coalescence
thus obtained. Obviously in H , there are no edges of the form ci c j and c0i c0j , i 6= j . Thus H is
a bipartite graph (because there are edges ci c0j in H ) (Fig. 5.13).

Graph Theory

115

Fig. 5.13

If we can partition the edge set of H into a disjoint union of bonds of H , the edges of G
corresponding to these bonds will be disjoint bonds of G whose union is C. To achieve
such a partition of E(H), we first take the cut edges of H as members of the partition and
let F be the set of such cut edges. For the remaining members of the partition, we take the
stars at the remaining (non-isolated) vertices ci (or c0i ). This gives the required partition and
hence the result follows.
q
Illustration Consider the graph of Figure 5.14. Partition of E(H) is {e1 } {e2 } {e3 , e4 }
{e5 , e6 }. e1 is a cut edge, e2 is a cut edge, {e3 , e4 } form the star K1, 2 and {e5 , e6 } form the
star K1, 2 .

Fig. 5.14
Remark Though the equivalence of minimal edge cuts and minimal cuts is brought out
by Theorem 5.10, there is an essential difference between edge cuts and cuts as already

116

Connectivity

mentioned. To emphasise this, we observe that Theorem 5.10 cannot be generalised to

state that every edge cut is a disjoint union of bonds. The example in Figure 5.11 illustrates
this point.
Since it is enough to consider connected graphs for discussing connectivity concepts, in
what follows, we shall assume that graphs are connected, unless stated otherwise.
The following results are reported by Harary and Frisck [105].
Theorem 5.11 In a connected graph G(V, E), if st / E , then (s, t) (s, t).

Proof

Let G(V, E) be a connected graph, and s, t be vertices in V such that st / E . Let

(s, t) be the cardinality of the minimum s t separating vertex cut (s t clot). Let (s, t)
be the cardinality of a minimum s t separating mixed cut. We prove that from any mixed
s t separating set, we can get an s t separating vertex cut with no more elements.
Let S be a minimum mixed s t separating set. If i j is an edge in S, then both i and j
cannot coincide with s and t , since st / E .
If i = s, add i to S, and remove from S all edges with i as an end vertex. If i 6= s, add j to
S and remove from S all edges with j as an end vertex. The resulting, possibly mixed set is
clearly an s t separating set with no more elements than S.
We repeat this process and remove all edges from S and obtain a vertex cut S0 with atmost
|S| elements.
Since (s, t) |S0 | |S| = (s, t), we have (s, t) (s, t).
q
Corollary

In a connected graph G(V, E), if st / E , then (s, t) (s, t).

Note If st E , then (s, t) is not defined.

Theorem 5.12 For any graph G, (G) = (G).

Proof
Case (i) When G = Kn , then (G) = n 1 and (G) = n 1.
Let S be a minimum mixed disconnecting set of G and let S = T F , where T V , F E ,
and |T | = n1 , |F| = m1 . Then G T is Knn1 . Therefore, |F| (Knn1 1 ) = n n1 1. Thus,
m1 n n1 1. So, = |S| = m1 + n1 n n1 1 + n1 = n 1. Therefore, n 1 = .
Also, . Hence, = .
Case (ii) When G is incomplete, then clearly . We have to prove that = when
G is complete. If possible, let there be a minimum s t separating mixed set S = M {st}
with = |S| < . Now, M can be replaced by a set of vertices T (a subset of the vertex set
of the induced subgraph hMi) to provide a vertex cut of G = G st with cardinality at most
|M|.
Let C1 and C2 be the components of G S to which s and t respectively belong. Let
there be another component C3 of G S and let v be a vertex of C3 . Then T {s} is a v t
separating vertex cut of G. But then |T {s}| |S| < , a contradiction (Fig. 5.15).

Graph Theory

117

Fig. 5.15

Thus C1 and C2 are the only components of G S. Also, if u V (C1 ), and u 6= s, then T {s}
is a u t separating vertex cut of G, again leading to a contradiction. Thus C1 = {s}, and
similarly C2 = {t}. So G has |V (M)| + 2 vertices, and is incomplete. Therefore (G)
n 2 = |V (M)| + 2 2 = |V (M)| < implying (G) < , a contradiction. Thus 6< . Hence
= .
q
The following inequalities are due to Whitney [265].
Theorem 5.13 (Whitney) For any graph G, (G) (G) (G).

Proof We first prove (G) (G).

If G has no edges, then = 0 and = 0. If G has edges, then we get a disconnected
graph, when all edges incident with a vertex of minimum degree are removed. Thus, in
either case, (G) (G).
We now prove (G) (G). For this, we consider the various cases. If G = Kn , then
(G) = (G) = n 1. Now let G be an incomplete graph. In case G is disconnected or
trivial, then obviously = = 0.
If G is disconnected and has a cut edge (bridge) x, then = 1. In this case, = 1, since
either G has a cut vertex incident with x, or G is K2 .
Finally, let G have 2 edges whose removal disconnects it. Clearly, the removal of
1 of these edges produces a graph with a cut edge (bridge) x = uv. For each of these 1
edges, select an incident vertex different from u or v. The removal of these vertices also
removes the 1 edges and quite possibly more. If the resulting graph is disconnected,
then < . If not, x is a cut edge (bridge) and hence the removal of u or v will result in
either a disconnected or a trivial graph, so that in every case.
q
Illustration
= 4.

We illustrate this by the graph shown in Figure 5.16. Here = 2, = 3 and

118

Connectivity

Fig. 5.16

Theorem 5.14

For any v V and any e E of a graph G(V, E), (G) 1 (G v) and

(G) 1 < (G e) (G).

Proof We observe that the removal of a vertex or an edge from a graph can bring down
or by at most one, and that while may be increased by the removal of a vertex,
cannot be increased by the removal of an edge.
q
Theorem 5.15 For any three integers r, s, t with 0 < r s t , there is a graph G with
= r, = s and = t .

Proof Take two disjoint copies of Kt+1 . Let A be a set of r vertices in one of them and
B be a set of s vertices in the other. Join the vertices of A and B by s edges utilising all the
vertices of B and all the vertices of A. Since A is a vertex cut and the set of these s edges is
an edge cut of the resulting graph G, it is clear that (G) = r and (G) = s. Also, there is at
least one vertex which is not in A B, and it has degree t , so that (G) = t .
q
Illustration Let r = 1, s = 2, t = 3. Take two copies of K4 . Here (G) = 1, (G) = 2,
(G) = 3 (Fig. 5.17).

Fig. 5.17

Theorem 5.16

For a graph,

n
ensures = .
2
n

Proof Let G be a graph with . Let < . Let F be a set of edges disconnecting
2
G. Let C1 and C2 be the components of G F , and A1 and A2 be the end vertices of F in C1
and C2 , respectively.
Suppose |A1 | = r, |A2 | = s and also V (C1 ) = A1 . Then each vertex of C1 is adjacent with
at least one edge of F . So the number m1 of edges in C1 satisfies the inequality

Graph Theory

119

1
1
m1 (r ) > (r ), since < by assumption.
2
2
1
2

1
2

Therefore, m1 > (r 1) > (r 1)r, since r |F| = < .

1

But a graph on n vertices cannot have more than r(r 1) edges. Thus, |V (C1 )| > |A1 |.
2
Similarly, V (C2 ) > |A2 |. Thus each of C1 and C2 contains at least + 1 vertices.
n
Therefore, n = |V (G)| 2( + 1) 2( + 1) = n + 2 or n n + 2, which is a contradiction.
2
Hence < is not possible. So = .
q

5.4 Mengers Theorem

Harary [104] listed eighteen variations of Mengers theorem including those for digraphs.
Clearly, all these are equivalent and one can be obtained from the other. Several proofs of
the various forms of Mengers theorems have appeared, for example, in Dirac [67], Ford
and Fulkerson [81], Lovasz [150], McCuaig [156], Menger [158], Nash-Williams and Tutte
[169], ONeil [173], Pym [213] and Wilson [269].
Let u and v be two distinct vertices of a connected graph G. Two paths joining u and v
are called disjoint (vertex disjoint) if they have no vertices other than u and v (and hence
no edges) in common. The maximum number of such paths between u and v is denoted by
p(u, v). If the graph G is to be specified, it is denoted by p(u, v|G).
The following is the vertex form of Mengers theorem. The proof is due to NashWilliams [9] and Tutte [169].
Theorem 5.17 (Menger-vertex form) The minimum number of vertices separating
two non-adjacent vertices s and t is equal to the maximum number of disjoint s t paths,
that is, for any pair of non-adjacent vertices s and t , the clot number equals the maximum
number of disjoint s t paths. That is, (s, t) = p(s, t), for every pair s, t V with st / E .

Proof Let G(V, E) be a graph with |E| = m. We use induction on m, the number of edges.
The result is obvious for a graph with m = 1 or m = 2. Assume that the result is true for
all graphs with less than m edges. Let the result be not true for the graph G with m edges.
Then we have
p(s, t|G) < (s, t|G) = q (say),

(5.17.1)

as for any graph, we obviously have p(s, t) (s, t).

Let e = uv be an edge of G. The deletion graph G1 = G e, and the contraction graph
G2 = G|e have less number of edges than G. Therefore, by induction hypothesis, we have
p(s, t|G1) = (s, t|G1 ) and p(s, t|G2) = (s, t|G2).
Let I be an (s, t) clot in G1 and J 0 be an (s, t) clot in G2 . Then we have

(5.17.2)

120

Connectivity
|I| = (s, t|G1) = p(s, t|G1) p(s, t|G) < q and

|J 0 | = (s, t|G2) = p(s, t|G2 ) p(s, t|G) < q, by using (5.17.2) and (5.17.1).

So |J 0 | < q and therefore |J 0 | q 1.

Now to J 0 there corresponds an (s t) vertex cut J of G such that |J| |J 0 | + 1, since, by
elementary contraction, (s, t) can be decreased by at most one, and this decrease actually
occurs when e E(hJi).
Thus, |J| |J 0 | + 1 q 1 + 1 = q,
that is, |J| q.

(5.17.3)

Since J is an (s, t) vertex cut in G, (s, t) |J|, q |J|.

Thus, q |J| q, so that |J| = q.
Therefore, |I| < q and |J| = q and u, v J by (5.17.3).

(5.17.4)

Let
Hs = {w I J : there exists an s w path in G, vertex-disjoint from I J {w}} and
Ht = {w I J : there exists a t w path in G, vertex-disjoint from I J {w}}.
Clearly, Hs and Ht are (s t) separating vertex cuts in G. Therefore,
|Hs | q and |Ht | q.

(5.17.5)

Obviously, Hs Ht I J .
We claim that Hs Ht I J . For this, let w Hs Ht . Then there exists an s w path
P1 and w t path P2 in G vertex disjoint from I J {w}. So P1 P2 contains a path, say P.
If e P then we have u, v V (P) J {w}, which is impossible. Therefore e / P and so
P G e. Since I is an (s, t ) separator in G e and J is an separator in G, P has a vertex
common with I and also with J . So w I J . Thus, Hs Ht I J .
Combining (5.17.4) and (5.17.5), and the above observation, we have
q + q |Hs | + |Ht | = |Hs Ht | + |Hs Ht | |I J| + |I J|
= |I| + |J| < q + q,

which is a contradiction.
Thus (5.17.1) is not true, and therefore, we have
(s, t|G) = p(s, t|G).

Graph Theory

121

Definition: Two paths joining u and v are said to be edge-disjoint if they have no edges
in common. The maximum number of edge-disjoint paths between u and v is denoted by
l(u, v).
The following is the edge form of Mengers theorem and the proof is adopted from
Wilson [196].
Theorem 5.18 (Menger-edge form) For any pair of vertices s and t of a graph G, the
minimum number of edges separating s and t equals the maximum number of edge-disjoint
paths joining s and t , that is, (s, t) = l(s, t) for every pair s, t V .

Proof Let G(V, E) be a graph and let |E| = m. We use induction on the number of edges
m of G. For m = 1, 2, the result is obvious. Assume the result to be true for all graphs with
fewer than m edges. Let (s, t) = k. We have two cases to consider.
Suppose G has an (s t) band F such that not all edges of F are incident with s,
nor all edges of F are incident with t . Then G F consists of two non-trivial components
C1 and C2 with s C1 and t C2 . Let G1 be the graph obtained from G by contracting the
edges of C1 , and G2 be a graph obtained from G by contracting the edges of C2 . Therefore,
Case (i)

G1 = G||E(C1 ) and G2 = G||E(C2).

Since G1 and G2 have less edges than G, the induction hypothesis applies to them. Also,
the edges corresponding to F provide an (s t) band in G1 and G2 , so that (s, t|G1) = k
and (s, t|G2) = k. Thus, by induction hypothesis, there are k edge-disjoint paths joining s
and t in G1 , and there are k edge-disjoint paths joining s and t in G2 . Thus l(s, t|G1 ) = k and
l(s, t|G2) = k.
The section of the path of the k edge-disjoint paths joining s and t in G2 which are in
C1 and the section of the paths of the k edge-disjoint paths joining s and t in G1 which are
in C2 can now be combined to get k edge disjoint paths between s and t in G. Hence
l(s, t|G) = k.
Case (ii) Every (s t) band of G is such that either all its edges are incident with s, or all
its edges are incident with t .
If G has an edge e which is not in any (s t) band of G, then (s, t|G e) = (s, t|G) =
k. Since the induction hypothesis is applicable to G e, there are k edge-disjoint paths
between s and t in G e and thus in G. Hence l(s, t|G) = k.
Now, assume that every edge of G is in at least one (s t) band of G. Then every s t
path P of G is either a single edge or a pair of edges. Any such path P can therefore contain
at most one edge of any (s t) band. Then G E(P) = G1 is a graph with (s, t|G1) = 1.
Appling induction hypothesis, we have l(s, t|G1 ) = 1. Together with P, we get
l(s, t|G) = k.
q

Definition: A graph G is said to be n-(vertex) connected if (G) n and n-(edge) connected if (G) n. Thus a separable graph ( = 1) is 1-connected and not 2-connected. A
separable graph without cut edges is only 1-edge connected.

122

Connectivity

5.5 Some Properties of a Bond

We give some properties of a bond (bond is also called a cut-set). The first property follows.
Theorem 5.19 Every bond in a connected graph G connects at least one branch of every
spanning tree of G.

Proof Let G be a connected graph and T be a spanning tree of G. Let S be an arbitrary

bond in G. Clearly, there are edges which are common in S and T . For, if there is no edge
of S which is also in T , then removal of the bond S from G will not disconnect the graph, as
G S contains T and is therefore connected. Thus S and T have at least one common edge.
q

Theorem 5.20 In a connected graph G, any minimal set of edges containing at least one
branch of every spanning tree of G is a bond.

Proof Let G be a connected graph and let Q be a minimal set of edges containing at least
one branch of every spanning tree of G.
Consider G Q, the subgraph that remains after removing the edges in Q from G. Since
G Q contains no spanning tree of G, therefore G Q is disconnected (one component of
which may just consist of an isolated vertex). Also, since Q is a minimal set of edges with
this property, therefore any edge e from Q returned to G Q creates at least one spanning
tree. Thus the subgraph G Q + e is a connected graph. Therefore Q is a minimal set of
edges whose removal from G disconnects G. This, by definition, is a bond.
q
Theorem 5.21

Every cycle has an even number of edges in common with any bond.

Proof Let G be a graph and let S be a bond of G. Let the removal of S partition the
vertices of G into two mutually disjoint subsets V1 and V2 . Consider a cycle C in G (Fig.
5.18).

Fig. 5.18

Graph Theory

123

If all the vertices in C are entirely within vertex set V1 (or V2 ), then the number of edges
common to S and C is zero, which is an even number. If, on the other hand, some vertices
in C are in V1 and some in V2 , we traverse back and forth between the sets V1 and V2 as we
traverse the cycle. Because of the closed nature of a cycle, the number of edges between V1
and V2 must be even. And, since every edge in S has one end in V1 and other in V2 , and no
other edge in G has the property of separating sets V1 and V2 , the number of edges common
to S and C is even.
q

5.6 Fundamental Bonds

Consider a spanning tree T of a connected graph G. Take any branch b in T . Since {b} is a
bond in T , therefore {b} partitions all vertices of T into two disjoint sets, one at each end
of b. Consider the same partition of vertices in G and the bond S in G that corresponds to
this partition. Bond S will contain only one branch b of T and the rest (if any) of the edges
in S are chords with respect to T . Such a bond S containing exactly one branch of a tree T
is called a fundamental bond with respect to T .
Theorem 5.22 The ring sum of any two bonds is either a third bond, or an edge-disjoint
union of bonds.

Proof Let G be a connected graph, and S1 and S2 be two bonds. Let V1 and V2 be the
unique and disjoint partitioning of the vertex set V of G corresponding to S1 . Let V3 and V4
be the partitioning corresponding to S2 .
Clearly, V1 V2 = V, V1 V2 = , V3 V4 = V and V3 V4 = (Fig. 5.19(a) and (b)).
Now, let (V1 V4 ) (V2 V3 ) = V5 and (V1 V3 ) (V2 V4 ) = V6 .
Clearly, V5 = V1 V3 and V6 = V2 V3 (Fig. 5.19(c)).
The ring sum of two bonds S1 S2 consists only of edges that join vertices in V5 to those
in V6 . Also, there are no edges outside S1 S2 that joins vertices in V5 to those in V6 . Thus
the set of edges S1 S2 produces a partitioning of V into V5 and V6 such that V5 V6 = V
and V5 V6 = . Hence S1 S2 is a bond if the subgraphs containing V5 and V6 each remain
connected after S1 S2 is removed from G. Otherwise, S1 S2 is an edge disjoint union of
bonds.
q

124

Connectivity

Fig. 5.19

Example Consider the graph in Figure 5.20. Here {d, e, f } { f , g, h} = {d, e, g, h}

is a bond, {a, b} {b, c, e, f } = {a, c, e, f } is another bond and {d, e, g, h} { f , g, k} =
{d, e, f , h, k} = {d, e, f } {h, k} an edge disjoint union of bonds.

Fig. 5.20

Theorem 5.23 With respect to a given spanning tree T , a chord ci that determines a
fundamental cycle C occurs in every fundamental bond associated with the branches in C
and in no other.

Proof Let G be a connected graph and T be a spanning tree of G. Let ci be a chord with
respect to T and let the fundamental cycle made by ci be called C, consisting of k branches
b1 , b2 , . . ., bk in addition to the chord ci . So C = {ci , b1 , b2 , . . ., ck } is a fundamental cycle
with respect to T .
Now every branch of any spanning tree has a fundamental bond associated with it. So
let S1 be the fundamental bond associated with b1 , consisting of q chords in addition to the
branch b1 . Thus, S1 = {b1 , c1 , c2 , . . ., cq } is a fundamental bond with respect to T .
We know there are even number of edges common to C and S1 . Clearly, b1 is in both C
and S1 . So there is exactly one more edge which is in both C and S1 . Obviously, the edge ci
in C can possibly be in S1 . Thus ci is one of the chords c1 , c2 , . . ., cq .
The same argument holds for fundamental bonds associated with b2 , b3 , . . ., bk . Thus
the chord ci is contained in every fundamental bond associated with branches in C.

Graph Theory

125

Now we show that the chord ci is not in any other fundamental bond S0 with respect to
T , besides those associated with b1 , b2 , . . ., bk . Let this be possible. Then since none of
the branches in C are in S0 , there is only one edge ci common to S0 and C, which gives a
contradiction to the fact that there are even number of edges common to a fundamental
bond and a cycle.
q
Example In the graph of Figure 5.20, consider the spanning tree {b, c, e, h, k}. The
fundamental cycle made by the chord f is C = { f , e, h, k}. The three fundamental bonds
determined by the three branches e, h and k are as follows: (i) determined by e is {d, e, f },
(ii) determined by h is { f , g, h} and (iii) determined by k is { f , g, k}. Clearly, chord f
occurs in each of these three fundamental bonds and there is no other fundamental bond
that contains f .
Theorem 5.24 With respect to a given spanning tree T , a branch bi that determines a
fundamental bond S is contained in every fundamental cycle associated with the chords in
S, and in no others.

Proof Let G be a connected graph and T be a spanning tree in G. Let the fundamental
bond determined by a branch bi be S = {bi , c1 , c2 , . . ., c p }.
Let C1 be the fundamental cycle determined by chord c1 , so that
C1 = {c1 , b1 , b2 , . . ., bq }.

We know that S and C1 have even number of edges in common. One common edge is
obviously c1 . Thus the second common edge should be bi , so that bi is also in C1 . Therefore
bi is one of the branches b1 , b2 , . . ., bq .
The same is true for the fundamental cycles made by the chords c2 , c3 , . . ., c p .
Now assume that bi occurs in a fundamental cycle Cp+1 made by a chord other than
c1 , c2 , . . ., c p . Since none of the chords c1 , c2 , . . ., c p is in C p+1 , there is only one edge bi
common to a cycle Cp+1 and the bond S, which is not possible. Hence the result follows.
q

Example Consider the graph of Figure 5.20. Consider the branch e of spanning tree
{b, c, e, h, k}. The fundamental bond determined by e is {e, d, f }. The two fundamental
cycles determined by chords d and f are respectively {d, c, e} and { f , e, h, k}. Clearly,
branch e is contained in both these fundamental cycles and none of the remaining three
fundamental cycles contains branch e.
Theorem 5.25 Let A, B be two disjoint vertex subsets of a graph G and let any vertex
subset of G which meets every A B path in G have at least k vertices. Then there are k
vertex disjoint A B paths in G.

Proof Let G be a graph and let A and B be two disjoint vertex subsets of G. Let S be any
vertex subset of G which meets every A B path in G and let |S| k.

126

Connectivity

Take two new vertices s and t , and join s by an edge to each vertex of A, and join t by an
edge to each vertex of B. Let G0 be the resulting graph, and in G0 we have (s, t) k.
Hence by Mengers theorem, there are k vertex disjoint paths between s and t in G0 .
Omitting the edges incident with s and t in these paths, we get k vertex-disjoint A B paths
in G.
q
Definition: A graph G is k-connected if (G) = k, and G is k-edge connected if (G) =
k. A k-connected (k-edge connected) graph is r-connected (r-edge 1-connected) for each
r, 0 r k 1. Clearly, a separable graph ( = 1) is connected and not 2-connected. A
separable graph without cut edge is 2-edge connected. A separable graph with cut edges is
only 1-connected.
The following result is due to Whitney [265].
Theorem 5.26 A graph G with at least three vertices is 2-connected if and only if any
two vertices of G are connected by at least two internally disjoint paths.

Proof Let G be 2-connected so that G contains no cut vertex. Let u and v be two distinct
vertices of G. To prove the result, we induct on d(u, v).
If d(u, v) = 1, let e = uv. Since G is 2-connected and n(G) 3, therefore e cannot be a
cut edge of G. For, if e is a cut edge, then at least one of u and v is a cut vertex. Now, by
Theorem 5.2, e belongs to a cycle C in G. Then C e is a u v path in G, internally disjoint
from the path uv.
Assume any two vertices x and y of G, such that d(x, y) = t 1, t 2, are joined by two
internally disjoint x y paths in G. Let d(u, v) = t and let P be a u v path of length t , and w
be the vertex before v on P. Then d(u, w) = t 1. Therefore, by induction hypothesis, there
are two internally disjoint u w paths, say P1 and P2 , in G. Since G has no cut vertex, G w
is connected and therefore there exists a u v path Q in G w. Clearly, Q is a u v path in
G not containing w. Suppose x is the vertex of Q such that x v section of Q contains only
the vertex x in common with P1 P2 (Fig. 5.21). Assume x belongs to P1 . Then the union
of the u x section of P1 and x v section of Q together with P2 {wv} are two internally
disjoint u v paths in G.
Conversely, assume any two distinct vertices of G are connected by at least two internally
disjoint paths. Then G is connected. Also, G has no cut vertex. For, if v is a cut vertex of
G, then there exist vertices u and w such that every u w path contains v, contradicting the
hypothesis. Thus G is 2-connected.
q
The following property of 3-connected graphs is given in Thomassen [241] and is attributed to Barnette and Grunbaum [14] and Titov [243].
Theorem 5.27 If G is a 3-connected graph with at least five vertices, then G has an edge
e such that G e is a subdivision of a 3-connected graph.

Proof
of K4 .

Since G is 3-connected, 3, and so by Mengers theorem, G has a subdivision

Graph Theory

127

Fig. 5.21

Let H be a proper subgraph of G which is a subdivision of a 3-connected graph, and let

H have maximum possible number of edges. If H is a 3-connected spanning subgraph of
G, then by the maximality of H , G has an edge e such that H = G e is a subgraph of G
with the desired property.
Now, let H be 3-connected but not spanning (Fig. 5.22). Then there is a vertex v
V (G) V (H), so that there are three v V (H) paths, say P1 , P2 and P3 which have only
vertex v in common. Let the other end vertices of these paths in V (H) be v1 , v2 and v3 . If
v1 v2 = e E(H), then H + P1 + P2 e is a subdivision of a 3-connected graph. Otherwise,
H + P1 + P2 is a subdivision of a 3-connected graph. In both cases the maximality of H
is contradicted.

Fig. 5.22

Let H be not 3-connected. Then H has a suspended path P of length at least 2. Let u
and v be the end vertices of P. Since G is 3-connected, G {u, v} has a path P0 joining
an internal vertex w of P to a vertex in V (H) V (P). But then H P0 is a subdivision of a
3-connected graph. By the choice of H, H P0 = G and P0 consists of a single edge e0 . q

128

Connectivity

The following property of 3-connected graphs is attributed to Thomassan [241].

Theorem 5.28 If G is a 3-connected graph with at least five vertices, then G has an edge
e such that G|e is 3-connected.

Proof Let G be a 3-connected graph with at least five vertices. Let e = uv be an edge of
G such that G|e is not 3-connected. Then G|e is 2-connected.
Let {x, y} be a vertex cut of G|e and let z be the vertex into which x and y have been
coalesced. Assume both x and y are different from z. Then G|e {x, y} is a graph obtained
by contracting an edge of the connected graph G {x, y}. This implies that G|e {x, y} is
connected, which is a contradiction. Thus one of x and y coincides with z. Renaming the
other as w, we see that G has a vertex cut {u, v, w}.
Let G1 be the smallest component of G {u, v, w}. Since G is 3-connected, G1 is joined
to w by an edge e1 = wx1 . If G|e1 is not 3-connected, by a similar argument, there is a vertex
y1 such that G {w, x1 , y1 } = G2 is disconnected. But then the smallest component of G2 is
a proper subgraph of G1 .
Continuing in this way, we reach a stage when the smallest sub-graph is a single vertex
and the edge f joining it to the previous vertex cut is such that G| f is a 3-connected graph.
q

Illustration We illustrate this in Figure 5.23, where graph G is 3-connected having vertex cut {u, v, w}. G|e is 2-connected with vertex cut {z, w} and G|e1 is 3-connected. In
G {u, v, w}, we observe that the smallest subgraph is a single vertex.

Fig. 5.23

5.7 Block Graphs and Cut Vertex Graphs

The block graph B(G) of a graph G is a graph whose vertices are the blocks of G and two
of these vertices are adjacent whenever the corresponding blocks contain a common cut
vertex of G. The cut vertex graph C(G) of a graph G has vertices as cut vertices of G and
two such vertices are adjacent if the cut vertices of G to which they belong lie on a common
block (Fig. 5.24).

Graph Theory

129

Fig. 5.24

The following characterisation of block graphs is due to Harary [103].

Theorem 5.29 A graph H is the block graph of some graph if and only if every block
of H is complete.

Proof Let H = B(G) and assume there is a block Hi of H which is not complete. Then
there are two vertices in Hi which are non-adjacent and lie on a shortest common cycle Z
of length at least 4. But the union of the blocks of G corresponding to the vertices of Hi
which lie on Z is then connected and has no cut vertex, so it itself is contained in a block,
contradicting the maximality property of a block of a graph.
Conversely, let H be a given graph in which every block is complete. Form B(H), and
then form a new graph G by adding to each vertex Hi of B(H) a number of end edges equal
to the number of vertices of the block Hi which are not cut vertices of H . Then it is easy to
see that B(G) is isomorphic to H .
q
1-isomorphism A separable graph consists of two or more non-separable subgraphs,
and each of the largest non-separable subgraph is a block. The graph in Figure 5.25 has
five blocks and three cut vertices u, v and w. We note that a non-separable connected graph
consists of just one block.

Fig. 5.25

130

Connectivity

Now, compare the disconnected graphs of Figure 5.26 with the graph of Figure 5.25.
Clearly, these two graphs are not isomorphic, as they do not have the same number of vertices. Evidently, the blocks of the graph of Figure 5.25 are isomorphic to the components
of the graph of Figure 5.26. We call such graphs 1-isomorphic.

Fig. 5.26

These observations lead to the following definition.

Definition: Two graphs G1 and G2 are said to be 1-isomorphic if they become isomorphic to each other under repeated application of the following operation.
Split a cut vertex into two vertices to produce two disjoint subgraphs.
This definition implies that two non-separable graphs are 1-isomorphic if and only if
they are isomorphic.
Operation 1

Two 1-isomorphic graphs have the following property.

Theorem 5.30 If G1 and G2 are 1-isomorphic graphs, then rank G1 = rank G2 and nullity
G1 = nullity G2 .

Proof Under operation 1, whenever a cut vertex in a graph G is split into two vertices,
the number of components in G increases by one. Therefore, rank G = number of vertices
in G-number of components in G remains invariant under operation 1.
Since no edges are destroyed or new edges created by operation 1, two 1-isomorphic
graphs have the same number of edges. Two graphs with same rank, and same number of
edges have the same nullity, since nullity = number of edgesrank.
q
Suppose the two vertices x and y belonging to different components of the graph in
Figure 5.26 are superimposed, then the graph obtained is shown in Figure 5.27. Clearly,
the graph in Figure 5.27 is 1-isomorphic to the graph in Figure 5.26. Also, since the blocks
of the graph in Figure 5.27 are isomorphic to the blocks of the graph in Figure 5.25, these
two graphs are 1-isomorphic. Hence the three graphs in Figures 5.25, 5.26 and 5.27 are
1-isomorphic.

Graph Theory

131

Fig. 5.27

We have seen that a graph G1 is 1-isomorphic to a graph G2 if the blocks of G1 are

isomorphic to the blocks of G2 . Since a non-separable graph is a block, 1-isomorphism
for non-separable graphs is same as isomorphism. For separable graphs, obviously 1isomorphism is different from isomorphism. In fact, graphs that are isomorphic are also
1-isomorphic, but the converse need not be true.
2-isomorphism In a 2-connected graph G, let x and y be a pair of vertices whose removal from G, leaves the remaining graph disconnected. That is, G consists of a subgraph
H and its complement H such that H and H have exactly two vertices x and y, in common.
Now, we perform the following operation on G.
Operation 2 Split the vertex x into x1 and x2 , and the vertex y into y1 and y2 such that G
is split into H and H . Let vertices x1 and y1 go with H , and x2 and y2 with H . Now, rejoin
the graphs H and H by merging x1 with y2 and x2 with y1 . Clearly, edges whose end vertices
are x and y in G can go with H or H , without affecting the final graph.
Two graphs are said to be 2-isomorphic if they become isomorphic after undergoing
operation 1, or operation 2, or both any number of times. For example, Figure 5.28 shows
how the two graphs in (a) and (d) are 2-isomorphic.

Fig. 5.28

132

Connectivity

It follows from the definition that isomorphic graphs are always 1-isomorphic and 1isomorphic graphs are always 2-isomorphic. But 2-isomorphic graphs are not necessarily 1-isomorphic and 1-isomorphic graphs are not necessarily isomorphic. However, for
graphs with three or more connectivity, isomorphism, 1-isomorphism and 2-isomorphism
are same.
Clearly, no edges or vertices are created or destroyed under operation 2. So the rank and
nullity of a graph remain unchanged under operation 2. Therefore the 2-isomorphic graphs
are equal in rank and equal in nullity.
Cycle correspondence: Two graphs G1 and G2 are said to have a cycle correspondence
if there is a one-one correspondence between the edges of G1 and G2 , and a one-one correspondence between the cycles of G1 and G2 , such that a cycle in G1 formed by certain
edges of G1 has a corresponding cycle in G2 formed by the corresponding edges of G2 , and
vice versa. Clearly, isomorphic graphs have cycle correspondence. Since in a separable
graph G, every cycle is confined to a particular block, every cycle in G retains its edges as
G undergoes operation 1. Thus 1-isomorphic graphs have cycle correspondence.
The following result for 2-isomorphic graphs is due to Whitney [266].
Theorem 5.31
dence.

Two graphs are 2-isomorphic if and only if they have cycle correspon-

5.8 Exercises
1. Prove that a vertex v of a tree is a cut vertex if and only if d(v) > 1.
2. Prove that a unicentric graph need not be separable.
3. Prove that a graph H is the block-cut vertex graph of some graph G if and only if it is
a tree in which the distance between any two end vertices is even.
4. Prove that a unicentric graph need not have d = 2r.
5. Prove that a non-separable graph with at least two edges has nullity greater than zero.
6. Prove that a non-separable graph of nullity one is a cycle and its converse.
7. If v is a cut vertex of a simple connected graph G, prove that v is not a cut vertex of G .
8. Prove that a connected k-regular bipartite graph is 2-connected.
9. Show that a simple connected graph with at least three vertices is a path if and only
if it has exactly two vertices that are not cut vertices.
10. If b(v) denotes the number of blocks of a simple connected graph G containing vertex
v, prove that the number of blocks b(G) of G is given by
b(G) = 1 + (b(v) 1).
vV (G)

Graph Theory

11. Prove that if a graph G is k-connected or k-edge-connected, then m

133
nk
.
2

12. Prove that a connected graph with at least two vertices contains at least two vertices
that are not cut vertices.
13. Prove that a 3-regular connected graph has a cut vertex if and only if it has a cut edge.
14. Prove that the connectivity and edge connectivity of a cubic graph are equal.
15. Prove that a graph with at least three vertices is 2-connected if and only if any two
vertices of G lie on a common cycle.
16. Prove that a graph is 2-connected if and only if for every pair of disjoint connected
subgraphs G1 and G2 , there exist two internally disjoint paths P1 and P2 of G between
G1 and G2 .
17. In a 2-connected graph G, prove that any two longest cycles have at least two vertices
in common.
18. Prove that a connected graph G is 3-connected if and only if every edge of G is the
exact intersection of the edge sets of two cycles of G.
19. Prove that a connected graph is Eulerian if and only if each of its blocks is Eulerian.
20. Prove that a connected graph is Eulerian if and only if each of its edge cuts has an
even number of edges.

6. Planarity
Let G(V, E) be a graph with V = {v1 , v2 , . . ., vn } and E = {e1 , e2 , . . ., em}. Let S be any
surface (like the plane, sphere) and P = {p1 , p2 , . . ., pn } be a set of n distinct points of S,
pi corresponding to vi , 1 i n. If ei = v j vk , draw a Jordan arc Ji on S from p j to pk such
that Ji does not pass through any other pi . Then P {J1 , J2 , . . ., Jm } is called a drawing of
G on S, or a diagram representing G on S. The pi are called the points of the diagram and
Ji , the lines of the diagram.
An embedding of a graph G on a surface S is a diagram of G drawn on the surface such
that the Jordan arcs representing any two edges of G do not intersect except at a point
representing a vertex of G.
A graph is planar if it has an embedding on the plane. A graph which has no embedding
on the plane is nonplanar. That is, a graph G is said to be planar if there exists some
geometric representation of G which can be drawn on a plane such that no two of its edges
intersect and a graph that cannot be drawn on a plane without a crossover between its edges
is called nonplanar.
In order that a graph G is nonplanar, we have to show that of all possible geometric
representations of G, none can be embedded in a plane. Equivalently, a geometric graph G
is planar if there exists a graph isomorphic to G that is embedded in a plane.
An embedding of a planar graph G on a plane is called a plane representation of G.
Figure 6.1 shows three diagrams of the same graph which is planar. The two graphs in
Figure 6.2 represent the same planar graph.

Fig. 6.1

Fig. 6.2

Graph Theory

135

6.1 Kuratowskis Two Graphs

The complete graph K5 and the complete bipartite graph K3,3 are called Kuratowskis
graphs, after the polish mathematician Kasimir Kurtatowski, who found that K5 and K3,3
are nonplanar.
Theorem 6.1

The complete graph K5 with five vertices is nonplanar.

Proof Let the five vertices in the complete graph be named v1 , v2 , v3 , v4 , v5 . Since in a
complete graph every vertex is joined to every other vertex by means of an edge, there is a
cycle v1 v2 v3 v4 v5 v1 that is a pentagon. This pentagon divides the plane of the paper into two
regions, one inside and the other outside, Figure 6.3(a).
Since vertex v1 is to be connected to v3 by means of an edge, this edge may be drawn
inside or outside the pentagon (without intersecting the five edges drawn previously). Suppose we choose to draw the line from v1 to v3 inside the pentagon, Figure 6.3(b). In case we
choose outside, we end with the same argument. Now we have to draw an edge from v2 to
v4 and another from v2 to v5 . Since neither of these edges can be drawn inside the pentagon
without crossing over the edge already drawn, we draw both these edges outside the pentagon, Figure 6.3(c). The edge connecting v3 and v5 cannot be drawn outside the pentagon
without crossing the edge between v2 and v4 . Therefore v3 and v5 have to be connected with
an edge inside the pentagon, Figure 6.3(d). Now, we have to draw an edge between v1 and
v4 and this cannot be placed inside or outside the pentagon without a crossover. Thus the
graph cannot be embedded in a plane.
q

Fig. 6.3

136

Theorem 6.2

Planarity

The complete bipartite graph K3,3 is nonplanar.

Proof The complete bipartite graph has six vertices and nine edges. Let the vertices be
u1 , u2 , u3 , v1 , v2 , v3 . We have edges from every ui to each vi , 1 i 3. First we take the
edges from u1 to each v1 , v2 and v3 . Then we take the edges between u2 to each v1 , v2 and
v3 , Figure 6.4(a). Thus we get three regions namely I, II and III. Finally we have to draw
the edges between u3 to each v1 , v2 and v3 . We can draw the edge between u3 and v3 inside
the region II without any crossover, Figure 6.4(b). But the edges between u3 and v1 , and
u3 and v2 drawn in any region have a crossover with the previous edges. Thus the graph
cannot be embedded in a plane. Hence K3,3 is nonplanar.
q

Fig. 6.4

We observe that the two graphs K5 and K3,3 have the following common properties.
1. Both are regular.
2. Both are nonplanar.
3. Removal of one edge or a vertex makes each a planar graph.
4. K5 is a nonplanar graph with the smallest number of vertices, and K3,3 is the nonplanar
graph with smallest number of edges.
Thus both are the simplest nonplanar graphs.
The following result given independently by Fary [77] and Wagner [260] implies that
there is no need to bend edges in drawing a planar graph to avoid edge intersections.
Theorem 6.3 (Fary) Every triangulated planar graph has a straight line representation.

Proof The proof is by induction on the number of vertices. The result is obvious for
n = 4. So, let n 5 and assume that the result is true for all planar graphs with fewer than
n vertices. Let G be a plane graph with n vertices.
First, we show that G has an edge e belonging to just two triangles. For this, let x be any
vertex in the interior of a triangle T and choose x and T such that the number of regions
inside T is minimal.

Graph Theory

137

Let y be a neighbour of x, and the edge xy lies inside T , and let xy belong to three triangles
xyz1 , xyz2 and xyz3 . Then one of these triangles lies completely inside another. Assume that
z3 lies inside xyz1 . Then z3 and xyz1 contradict the choice of x and T (Fig. 6.5).

Fig. 6.5

Thus there is an edge e = xy lying in just two triangles xyz1 and xyz2 . Contracting xy to
a vertex u, we get a new graph G0 with a pair of double edges between u and z1 , and u and
z2 . Remove one each of this pair of double edges to get a graph G00 which is a triangulated
graph with n 1 vertices. By the induction hypothesis, it has a straight line representation
H 00 . The edges of G00 correspond to uz1 , uz2 in H 00 . Divide the angle around u into two parts
in one of which the pre-images of the edges adjacent to x in G lie, and in the other, the
pre-images of the edges adjacent to y in G. Thus u can be pulled apart to x and y, and the
edge xy is restored by a straight line to get a straight line representation of G.
q

6.2 Region
A plane representation of a graph divides the plane into regions (also called windows,
faces or meshes). A region is characterised by the set of edges (or the set of vertices)
forming its boundary. We note that a region is not defined in a nonplanar graph, or even in

138

Planarity

a planar graph not embedded in a plane. Thus a region is a property of the specific plane
representation of a graph and not of an abstract graph.

Fig. 6.6

Infinite region: The portion of the plane lying outside a graph embedded in a plane such
as region 4 in the graph given in Figure 6.6, is infinite in its extent. Such a region is called
an infinite, unbounded, outer or exterior region for that particular plane representation.
Like other regions, the infinite region is also characterised by a set of edges (or vertices).
Clearly, by changing the embedding of a given planar graph, we can change the infinite
region. Consider the graphs of Figure 6.7. In the two embeddings of the same graph, finite
region v1 v3 v5 in (a) is infinite in (b).

Fig. 6.7

Embedding on a sphere: A graph is spherical if it can be embedded on the surface of

a sphere.
The following result implies that a planar graph can often be embedded on the surface
of a sphere, so that there is no distinction between finite and infinite region.
Theorem 6.4
in a plane.

A graph can be embedded on the surface of a sphere if it can be embedded

Proof Consider the stereographic projection of a sphere on the plane. Put the sphere on
the plane and call the point of contact as SP (south-pole). At point SP, draw a straight line
perpendicular to the plane, and let the point where this line intersects the surface of the
sphere be called NP (north-pole).

Graph Theory

139

Fig. 6.8

Now, corresponding to any point p on the plane, there exists a unique point p0 on the
sphere, and vice versa, where p0 is the point where the straight line from point p to point
NP intersects the surface of the sphere. Thus there is a one-one correspondence between
the points of the sphere and the finite points on the plane, and points at infinity in the plane
corresponding to the point NP on the sphere.
Therefore from this construction, it is clear that any graph that can be embedded in a
plane can also be embedded on the surface of the sphere, and vice versa.
q
We now have the following result.
Theorem 6.5 A planar graph can be embedded in a plane such that any specified region,
specified by the edges forming it, can be made the infinite region.

Proof A planar graph embedded on the surface of the sphere divides the surface of the
sphere into different regions. Each region of the sphere is finite, the infinite region on the
plane having been mapped onto the region containing the point NP. Clearly, by suitably
rotating the sphere, we can make any specified region map onto the infinite region on the
plane. Hence the result.
q
Remark Thinking in terms of the regions on the sphere, we observe that there is no real
difference between the infinite region and the finite regions on the plane. Therefore, when
we discuss the regions in a plane representation of a graph, we include the infinite region.
Also, since there is no essential difference between an embedding of a planar graph on a
plane, or on a sphere (a plane can be regarded as the surface of the sphere of infinitely large
radius), the term plane representation of a graph is often used to include spherical as well
as plane embedding.

6.3 Eulers Theorem

The following important result due to Euler gives a relation between the number of vertices,
edges, regions and the components of a planar graph.
Theorem 6.6
then

If G is a planar graph with n vertices, m edges, f regions and k components,

n m + f = k + 1.

(6.6.1)

140

Planarity

Proof We construct the graph G by the addition of successive edges starting from the
null graph Kn . For this starting graph, k = n, m = 0, f = 1, so that (6.6.1) is true.
Let Gi1 be the graph at the start of ith stage and Gi be the graph obtained from Gi1 by
addition of the ith edge e. If e connects two components of Gi1, then f is not altered, m is
increased by 1 and k is reduced by 1, so that (6.6.1) holds for Gi as it holds for Gi1. If e
joins two vertices of the same components of Gi1, k is unaltered, m is increased by 1 and
f is increased by 1, so that again (6.6.1) holds for Gi .
q
The following relation between the number of vertices, edges and regions is the discovery of Euler [75] and is also called as Eulers formula for planar graphs.
Theorem 6.7 (Euler) In a connected planar graph with n vertices, m edges and f regions
(faces), n m + f = 2.
(The proof can also be deduced from Theorem 6.6 by taking k = 1, as the connected
graph has one component.)

Proof Without loss of generality, assume that the planar graph is simple. Since any
simple planar graph can have a plane representation such that each edge is a straight line,
any planar graph can be drawn such that each region is a polygon (a polygon net). Let
the polygon net representing the given graph consist of f regions. Let k p be the number of
p-sided regions. Since each edge is on the boundary of exactly two regions,
3k3 + 4k4 + 5k5 + . . . + rkr = 2m,

(6.7.1)

where kr is the number of polygons with r edges.

Also, k3 + k4 + k5 + . . . + kr = f .

(6.7.2)

The sum of all angles subtended at each vertex in the polygon net is 2 n.
(6.7.3.)
Now, the sum of all interior angles of a p-sided polygon is (p 2) and the sum of the
exterior angles is (p +2). The expression in (6.7.3) is the total sum of all interior angles of
f 1 finite regions plus the sum of the exterior angles of the polygon defining the infinite
region. This sum is
(3 2)k3 + (4 2)k4 + . . .+ (r 2)kr + 4
= [3k3 + 4k4 + . . .+ rkr + 2(k3 + k4 + . . .+ kr )] + 4
= (2m 2 f ) + 4 = 2 (m f + 2).

(6.7.4)

Equating (6.7.3) and (6.7.4) we get

2 (m f + 2) = 2n , so that f = m n 2.

Definition: Let be a region of a planar graph G. We define the degree of , denoted

by d( ), as the number of edges on the boundary of .

Graph Theory

141

Second proof of Theorem 6.7 We use induction on m, the number of edges. If m = 0,

then G is K1 , a graph with 1 vertex and 1 region. So, n m + f = 1 0 + 1 = 2, and the result
is true. If m = 1, then the number of vertices in G is either one or two, the first possibility
occurring when the edge is a loop. These two possibilities give rise respectively to two
regions and one region.

Fig. 6.9

Therefore, n m + f =

11+2,
21+1,

in the loop case ,

in the non loop case .

Thus, n m + f = 2, and again the result is true.

Assume that the result is true for all connected planar graphs with fewer than m edges.
Let G have m edges.
Suppose G is a tree. Then m = n 1 and f = 1, because a planar representation of
a tree has only one region. Thus, n m + f = n (n 1) + 1 = 2, and the result holds.

Case I

Suppose G is not a tree. Then G has cycles. Let C be a cycle in G. Let e be an

edge of C. The graph G e has one edge less than the graph G. Also the number of vertices
in G e and G are same. Since removing e coalesces two regions in G into one in G e,
therefore G e has one region less than that in G. Thus by induction hypothesis in G e,
we have n (m 1) + ( f 1) = 2, so that n m + f = 2. Hence the result.
q

Case II

The next result gives an upper bound for the edges of a simple planar graph.
Theorem 6.8 Let G be a simple planar graph with n vertices and m edges, where n 3.
Then m 3n 6.

Proof First assume that the planar graph G is connected. If n = 3, then since G is simple,
therefore G has at most three edges. Thus m 3, that is, m 3 3 6, and the result is true.
Now, let n = 4. First let G be a tree so that m = n 1. Since n 4, obviously we have
n 1 3n 6.
Now, let G be not a tree. Then G has cycles. Clearly, there is a cycle in G, all of whose
edges lie on the boundary of the exterior region of G. Then, since G is simple, we have
d( ) 3, for each region of G.
Let b = d ( ), where F denotes the set of all regions of G.
F

142

Planarity

Since each region has at least three edges on its boundary, therefore we have b 3 f ,
where f is the number of regions of G. However, when we sum to get b, each edge of G
is counted either once or twice (twice when it occurs as a boundary edge for two regions).
Therefore, b 2m so that 3 f b 2m. In particular, we have 3 f 2m.
Using Eulers formula n m + f = 2, we get m 3n 6.
Let G be not connected and let G1 , G2 , . . ., Gt be its connected components. For each
i, 1 i t , let ni and mi denote the number of vertices and edges in Gi . Since each Gi is a
planar simple graph, by the above argument, we have mi 3ni 6, for each i, 1 i t .
t

i=1

i=1

Also, n = ni and m = mi .
t

i=1

i=1

i=1

Hence, m = mi (3ni 6) = 3 ni 6t 3n 6.

The following result gives the existence of a vertex of degree less than six in a simple
planar graph.
Theorem 6.9

If G is a simple planar graph, then G has a vertex v of degree less than 6.

Proof If G has only one vertex, then this vertex has degree zero. If G has only two
vertices, then both vertices have degree at most one.
Let n 3. Assume degree of every vertex in G is at least six.
Then,

d (v) 6n.
vV (G)

We know

d (v) 2m.
vV (G)

Thus, 2m 6n so that m 3n.

This is not possible because, by Theorem 6.8, we have m 3n 6. Thus we get a
contradiction. Hence G has at least one vertex of degree less than 6.
q
Corollary 6.1

K5 is nonplanar.

Proof Here n = 5 and m = 10. So, 3n 6 = 15 6 = 9. Thus m > 3n 6. Therefore K5 is

nonplanar.
Corollary 6.2

K3, 3 is nonplanar.

Proof Since K3, 3 is bipartite, it contains no odd cycles, and so no cycle of length three.
It follows that every region of a plane drawing of K3, 3 if it exists, has at least four boundary
edges. We have d( ) 4, for each region of G.

Graph Theory

143

Let b = d ( ), where F denotes the set of all regions of G. Since each region has at
F

least four edges on its boundary, we have b 4 f , where f is the number of regions of G.
Now, when we sum up to get b, each edge of G is counted either once or twice, and so
b 2m. Thus, 4 f b 2m, so that 2f m.
9
For K3, 3 , we have m = 9, and so 2 f 9 giving f . But by Eulers formula, f =
2
m n + 2 = 9 6 + 2 = 5, a contradiction. Hence K3, 3 is nonplanar.
q
Eulers formula gives a necessary condition for connected planar graphs.
Theorem 6.10

For a connected simple planar graph with n vertices and m edges, and

girth g, we have m

g (n 2)
.
g2

Proof Let G be a connected simple planar graph with n vertices, m edges and girth g.
Let F be the set of regions and d( ) be the degree of the region . Then d ( ) = 2m. Let
F

f be the number of regions, that is |F| = f . As g is the length of the smallest cycle in G,
therefore g is the smallest degree of a region. So, d ( ) g f . Thus, 2m g f . Now, using
F

Eulers formula, f = m n + 2, we have 2m g(m n + 2), so that m

Corollary 6.3

Proof

g (n 2)
.
g2

The Peterson graph P is nonplanar.

The girth
 of the
 Perterson graph P is 5, n(P) = 10, and m(P) = 15. Thus, if P were

planar, 15 = 5

10 2
, which is not true. Hence P is nonplanar.
52

Corollary 6.4 Any graph containing a homeomorph of a K5 or K3, 3 is nonplanar. This

follows from the fact that any homeomorph of a nonplanar graph is nonplanar.
The following result characterises nonplanar 3-connected graphs with minimum edges.
Theorem 6.11 A nonplanar connected graph G with minimum number of edges that
contains no subdivision of K5 or K3, 3 is simple and 3-connected.

Proof Since G has minimum number of edges, G is minimal with respect to these properties. The minimality of G ensures that it is simple. Also, a graph is planar if and only if
each of its blocks is planar. This implies that a minimal nonplanar graph is a block.
Suppose such a graph has a two vertex cut S = {u, v}. Then G S is disconnected. Let
G1 be one of its components and G2 be the union of all of its other components.
Let H1 = < V (G1 ) S > + e and H2 = < V (G2 ) S > + e, where e = uv.
Clearly, these are nonempty graphs with fewer edges than G. Since G is nonplanar, at
least one of H1 and H2 is nonplanar. If both are planar, let H 1 be a plane embedding of H1
and let be a region of H 1 containing the edge e.

144

Planarity

A plane embedding H 2 of H2 inside can then be obtained with the edge e of both H 1 and
H 2 coinciding. Then H 1 H 2 e is a plane embedding of G, contradicting our hypothesis.
Suppose H1 is nonplanar. Since m(H1 ) < m(G), therefore H1 contains a subdivision K of
either K5 or K3, 3 according to the hypothesis. Clearly e E , since otherwise K G, contradicting the assumption. Now, replacing e of K by a u v path in H2 , we get a homeomorph
of K contained in G. This is again a contradiction and shows that G has no 2-vertex cut.
Hence G is 3-connected.
q
The following result is used in proving Kuratowskis theorem.
Theorem 6.12
a. If G|e contains a subdivision of K5 , then G contains a subdivision of K5 or K3, 3 .
b. If G|e contains a subdivision of K3, 3 , then G contains a subdivision of K3, 3 .

Proof Let G0 = G|e be a graph obtained by contracting the edge e = xy of G. Let w be the
vertex of G0 obtained by contracting e = xy.
a. Let G|e contains a subdivision of K5 , say H . If w is not a branch vertex of H , then G
also contains a subdivision of K5 , obtained by expanding w back into the edge xy, if
necessary (Fig. 6.10).

Fig. 6.10

Assume w is a branch vertex of H and each of x, y is incident in G to two of the four

edges incident to w in H . Let u1 and u2 be the branch vertices of H that are at the
other ends of the paths leaving w on edges incident to x in G. Let v1 , v2 be the branch
vertices of H that are at the other ends of the paths leaving w on edges incident to y
in G (Fig. 6.11).
By deleting the u1 u2 path and v1 v2 path from H , we obtain a subdivision of K3, 3
in G, in which y, u1 , u2 are branch vertices for one partite set, and x, v1 , v2 are branch
vertices of the other.

Graph Theory

145

Fig. 6.11

b. Let G|e contain a subdivision of K3, 3 , say H . If w is not a branch vertex of H , then G
also contains a subdivision of K3, 3 , obtained by expanding w back into the edge xy, if
necessary (Fig. 6.12).

Fig. 6.12

Now, assume that w is a branch vertex in H and at most one of the edges incident to
w in H is incident to x in G. Then w can be expanded into xy to lengthen that path and
y becomes the corresponding branch vertex of K3, 3 in G (Fig. 6.13).
q

Fig. 6.13

6.4 Kuratowskis Theorem

This theorem was independently given by Kuratowski [144] and Frink and Smith. In 1954,
Dirac and Schuster [69] found a proof that was slightly shorter than the original proof. The
proof given here is due to Thomassen [241].

146

Theorem 6.13 (Kuratowski)

subdivision of K5 or K3, 3 .

Planarity

A graph is planar if and only if it does not have any

Proof
Necessity Let G be a planar graph. Then any of its subgraphs is neither K5 nor K3, 3 nor
does it contain any subdivision of K5 or K3, 3 .
Sufficiency It is enough to prove sufficiency for 3-connected graphs. Let G be a 3connected graph with n vertices. We prove that the 3-connected graph G either contains
a subdivision of K5 or K3, 3 or has a convex plane representation. This we prove by using
induction on n. Since G is 3-connected, therefore n 4. For n = 4, G = K4 and clearly has
a plane representation.
Now, let n 5. Assume the result to be true for all 3-connected graphs with fewer than n
vertices. Since G is 3-connected, G has an edge e such that G|e is 3-connected. Let e = xy.
If G|e contains a subdivision of K5 or K3, 3 , then G also contains a subdivision of K5 or K3, 3 .
Therefore, let G|e = H have a convex plane representation. Let z be the vertex obtained by
contraction of e = xy. The plane graph obtained by deleting the edges incident to z has a
region containing z (this may be the exterior region). Let C be the cycle of H z bounding
this region.
Since we started with a convex plane representation of H , we have straight segments
from z to all its neighbours. Let x1 , x2 , . . ., xk be the neighbours of x in that order on C.
If all the neighbours of y belong to a single segment from xi to xi+1 on C, then we obtain
a convex plane representation of G by putting x at z in H , and putting y at a point close to z
in the wedge formed by x xi and x xi+1 .
If all the neighbours of y do not belong to any single segment xi xi+1 on C (1 i k,
xk+1 = x1 ), then we have the following cases (Fig. 6.14).
a. y shares three neighbours with x. In this case C together with these six edges involving
x and y form a subdivision of K5 .
b. y has two u, v in C that are in different components of the subgraph of C obtained by
deleting xi and xi+1 , for some i. In this case, C together with the paths uyv, xi xxi+1 and
xy form a subdivision of K3, 3 .
q

Fig. 6.14

Graph Theory

147

6.5 Geometric Dual

Let G be a plane graph. The dual of G is defined to be the graph G constructed as follows.
To each region f of G there is a corresponding vertex f of G and to each edge e of G there
is corresponding edge e in G such that if the edge e occurs on the boundary of the two
regions f and g, then the edge e joins the corresponding vertices f and g in G . If the
edge e is a bridge, i.e., the edge e lies entirely in one region f , then the corresponding edge
e is a loop incident with the vertex f in G . For example, consider the graph shown in
Figure 6.15.

Fig. 6.15

Theorem 6.14

The dual G of a plane graph is planar.

Proof Let G be a plane graph and let G be the dual of G. The following construction of
G shows that G is planar.
Place each vertex f k of G inside its corresponding region f i . If the edge ei lies on the
boundary of two regions f j and f k of G, join the two vertices f j and f k by the edge ei ,
drawing so that it crosses the edge e exactly once and crosses no other edge of G (Fig.
6.16).
q

Fig. 6.16
Remarks Clearly, there is one-one correspondence between the edges of plane graph G
and its dual G with one edge of G intersecting one edge of G.

148

Planarity

1. An edge forming a self-loop in G gives a pendant edge in G (An edge incident on a

pendant vertex is called a pendant edge).
2. A pendant edge in G gives a self loop in G .
3. Edges that are in series in G produce parallel edges in G .
4. Parallel edges in G produce edges in series in G .
5. The number of edges forming the boundary of a region f i in G is equal to the degree
of the corresponding vertex f i in G .
6. Considering the process of drawing a dual G from G, it is evident that G is a dual
of G . Therefore, instead of calling G a dual of G, we usually say that G and G are
dual graphs.
7. Let n, m, f , r and denote the number of vertices, edges, regions, rank and nullity of
a connected plane graph G and let n , m , f , r and be the corresponding numbers
in G . Then n = f , m = m, f = n. We have r = n 1, = m n + 1, r = n 1,
= m n + 1. So, r = f 1, = m f + 1, r = n 1, = m n + 1.
Using Eulers formula, n m + f = 2 or f = m n + 2, we have r = m n 2 1 =
m n + 1 = and = m f + 1 = n + f 2 f + 1 = n 1 = r.
We now have the following result.
Theorem 6.15

The edge e is a loop in G if and only if e is a bridge in G .

Proof Let the edge e be a loop in a plane graph G. Then it is the edge on the common
boundary of two regions on which, say f , lies within the area of the plane surrounded by e
with the other, say g, lying outside the area. Thus, from definition of G , e is the only path
from f to g in G . Thus e is a bridge in G .
Conversely, let e be a bridge in G , joining vertices f and g . Thus e is the only path
in G from f to g . This implies, again from the definition of G , that the edge e in G
completely encloses one of the regions f and g. So e is a loop in G.
q
Remark The occurrence of parallel edges in G is easily described. Given two regions f
and g of G, there are k parallel edges between f and g if and only if f and g have k edges

on their common boundary.

Note We have defined the dual of a plane graph instead of a planar graph. The reason
for this is that different plane drawings G1 and G2 of the same planar graph G may result in
non-isomorphic duals G1 and G2 .

2-isomorphism Two graphs G(V, E)and H(W, F) are said to be 2-isomorphic if there
exists a bijection : E F such that both and 1 preserve cycles.

Graph Theory

149

Two graphs G and H are said to be 2-isomorphic if there exists a one-one correspondence
between their edge sets such that the edges of a cycle in G correspond to the edges of a
cycle in G2 and vice versa.
Example

Figure 6.17(a) shows 2-isomorphic graphs.

1-isomorphism Graphs which become isomorphic after the splitting of all of their cut
vertices are said to be 1-isomorphic.
Example

Figure 6.17(b) shows 1-isomorphic graphs.

Fig. 6.17

Consider the two non-isomorphic graphs G1 and G2 shown in Fig. 6.17(a) and (b). They
both have a single cut vertex, namely v. If this cut vertex is split into two vertices in each of
G1 and G2 , we obtain the edge disjoint graphs G01 and G02 as shown in (c) and (d). Clearly,
G01 and G02 are isomorphic. Thus G1 and G2 are 1-isomorphic.
The next result gives a correspondence between the edges of a graph with the edges of
its dual.
Theorem 6.16 Edges in a plane graph G form a cycle in G if and only if the corresponding dual edges form a bond in G .

150

Planarity

Proof Let D E(G). It is sufficient to prove that D contains a cycle if and only if the
set D of dual edges contains a bond of G . Now, let D contain a cycle C. Then by the
Jordan Curve theorem, some region of G lies inside C and some region lies outside C.
These regions correspond to vertices v , w in G , one drawn inside C and one outside C.
A v w path in G crosses C and hence uses an edge of G , that is dual to an edge of C.
Thus D disconnects v from w and hence D contains a bond (Fig. 6.18(a)).
Conversely, let D contain a bond. We show that D contains a cycle. Assume that D does
not contain a cycle. Then D encloses no region. It remains possible to reach each region of
G from every other without crossing D. Hence G D is connected, and so D contains no
bond. This is a contradiction. Hence D contains a cycle.
q

Fig. 6.18(a)

Theorem 6.17 All duals of a planar graph G are 2-isomorphic and any graph H
2-isomorphic to a dual G of G is itself a dual of G.

Proof Let G (V , E ) be a dual of G. Then there is an edge bijection : E E such

that cycles in G correspond to bonds in G. Let H(W, F) be any other dual of G. Then there
is an edge bijection : E E such that bonds in G correspond to cycles in H . Therefore
: E F is an edge bijection which preserves cycles between G and H . Hence G and
H are 2-isomorphic.
Conversely, let H be a graph which is 2-isomorphic to a dual G of G. Then there is an
edge bijection : E F which preserves cycles between G and H . Also, there is an edge
bijection : E E such that cycles in G correspond to bonds in G. Thus we have an edge
bijection : F E such that cycles in F correspond to bonds in G. Hence H is a dual of
G.
q
The following result provides the condition for two given graphs to be duals of each other.
Theorem 6.18 A necessary and sufficient condition for two planar graphs G1 and G2 to
be the duals of each other is that there should be a one-one correspondence of the edges in
G1 with the edges in G2 such that the subset of edges in G1 forms a cycle if and only if the
corresponding set of edges in G2 forms a bond.

Proof
Necessity Consider a plane representation of a planar graph G1 and let G1 be the dual of
G1 . Let C be an arbitrary cycle in G1 .

Graph Theory

151

Clearly, C will form some closed simple curve in the plane representation of G1 , dividing
the plane into two areas. Thus the vertices of G1 are partitioned into two non-empty,
mutually exclusive subsets, one inside C and the other outside. In other words, the set of
edges C in G1 corresponding to the set C in G1 is a bond in G1 .
Similarly, corresponding to a bond S in G1 , there is a unique cycle consisting of the
corresponding edge set S in G1 such that S is a cycle.
Sufficiency Let G be a planar graph, and let G0 be a graph for which there is a one-one
correspondence between the bonds of G and the cycles of G0 and vice versa. Let G be the
dual graph of G. Therefore there is a one-one correspondence between the bonds of G and
cycles of G . So there is a one-one correspondence between the cycles of G0 and G . Thus
G0 and G are 2-isomorphic. Hence G0 is the dual of G.
q
Let G be a plane graph and G be the dual of G. So to every vertex of G we have a
region of G . Let the region of the plane graph G corresponding to the vertex v of G has
e1 , e2 , . . ., ek as its boundary edges. Then each of these edges ei crosses the corresponding
edge ei of G, and these edges are all incident at the vertex v. It follows that contains the
vertex v. As G is a plane graph, we can construct the dual of G , called the double dual of
G and is denoted by G .
We now have the following observation.
Theorem 6.19
dual G .

Let G be a plane connected graph. Then G is isomorphic to its double

Proof Let G be a plane connected graph and G be the dual of G. Now, any region of
dual G contains exactly one vertex of G, namely its corresponding vertex v of G. This is
because the number of the regions of G is same as the number of vertices of G.
Thus in the construction of G , we take the vertex v to be the vertex in G corresponding
to the region of G . This choice gives the required isomorphism.
q
Definition: A connected graph G is called self-dual if it is isomorphic to its dual G . For
example, K4 is self-dual.
The next result is a characterisation of bipartite graphs in terms of duality.
Theorem 6.20
Eulerian.

A connected plane graph G is bipartite if and only if its dual graph G is

Proof Let G be a bipartite graph. Then G does not contain odd cycles. Since G is a plane,
all the regions of G are of even length. So degree of every region in G is even. Let G be
the dual graph of G. Since the vertex degrees of G are same as the corresponding degrees
of the regions of G, degree of every vertex in G is even. Hence G is Eulerian.
Conversely, assume G is Eulerian. Then degree of every vertex in G is even. Since
the degree of a vertex in G corresponds to the degree of the region in G, degree of every
region in G is even. This shows that G contains only even cycles. Hence G is bipartite. q

152

Planarity

Dual of a subgraph: Let G be a plane graph and let G be its dual. Let e be an edge in
G and let e be the corresponding edge in G . We delete the edge e and find dual of G e.
If edge e is on the boundary of two regions, then removal of e merges these two regions
into one.
Thus the dual (G e) is obtained from G by deleting the corresponding edge e and
then fusing the two end vertices of e in G e . On the other hand, if edge e is not on the
boundary, then e forms a self-loop. In that case, G e is the same as (G e) .
Thus, if a graph G has a dual G , then the dual of any subgraph of G can be obtained by
successive application of this procedure.
Dual of a homeomorphic graph: Let G be a plane graph and G be its dual. Let e be
an edge in G and let e be the corresponding edge in G . We create an additional vertex
in G by introducing a vertex of degree two in edge e (e now becomes two edges in series).
This simply adds an edge parallel to e in G . Likewise, the reverse process of merging two
edges in series will simply eliminate one of the corresponding parallel edges in G . Thus,
by this procedure, if the graph G has dual G , the dual of any graph homeomorphic to G
can be obtained from G . These are illustrated in Figure 6.18(b).
The following important result is due to Whitney [267].
Theorem 6.21 (Whitney) A graph has a dual if and only if it is planar.

Proof We have only to prove that a nonplanar graph does not have a dual. Let G be
a nonplanar graph. Then according to Kuratowskis theorem, G contains K5 or K3, 3 , or a
sub-division of K5 or K3, 3 .
We know that a graph G has a dual only if every subgraph H of G and every subdivision
of G has a dual. Thus to prove the result, we show that neither K5 nor K3, 3 has a dual.

Fig. 6.18(b)

Graph Theory

153

a. Suppose that K3, 3 has a dual D. Then the bonds in K3, 3 correspond to cycles in D and
vice-versa. Since K3, 3 has no bond consisting of two edges, D has no cycle consisting
of two edges. That is, D contains no pair of parallel edges. Since every cycle in K3, 3
is of length four or six, D has no bond with less than four edges. Therefore the degree
of every vertex in D is at least four. As D has no parallel edges and the degree of every
vertex is at least four, D has at least five vertices each of degree four or more ( 4).
That is, D has at least 54
2 = 10 edges. This is a contradiction to the fact that K3, 3 has
nine edges. Hence K3, 3 has no dual.
b. Suppose K5 has a dual H . We note that K5 has
i. 10 edges,
ii. no pair of parallel edges,
iii. no bond with two edges, and
iv. bonds with only four or six edges.
Therefore graph H has
i. 10 edges,
ii. no vertex with degree less than three,
iii. no bond with two edges, and
iv. cycles of length four or six only.
Now graph H contains a hexagon (a cycle of length six) and no more than three edges
can be added to a hexagon without creating a cycle of length three, or a pair of parallel
edges. Since both of these are not present in H , and H has 10 edges, there must be at least
seven vertices in H . The degree of each of these vertices is at least three. This implies that
H has at least 11 edges, which is a contradiction. Hence K5 has no dual.
q
Definition: Two plane graphs G and G are said to be duals (or combinatorial duals) of
each other if there is a one-one correspondence between the edges of G and G such that if
H is any subgraph of G, and H is the corresponding subgraphs of G , then rank (G H ) =
rank G nullity H .

6.6 Polyhedron
A polyhedron is a solid bounded by surfaces, called faces, each of which is a plane. That
is, a solid bounded by plane surfaces is called a polyhedron. For example, brick, window
frame, tetrahedron.
A polyhedron is said to be convex if any two of its interior points can be joined by
a straight line lying entirely within the interior. For example, a brick is convex, but a
window frame is not convex.

154

Planarity

The vertices and edges of a polyhedron, which form a skeleton of the solid, give a simple
graph in three dimensional space. It can be shown that for a convex polyhedron this graph
is planar.
To see this, imagine the faces of the convex polyhedron to be made of rubber, with
one face, say at the base, missing. Then taking hold of the edges at the missing face,
we are able to stretch out the rubber to form one plane sheet. The vertices and edges
in this transformation now form a plane graph. Moreover, each face of this plane graph
corresponds to a face of a solid, with the exterior face of the graph corresponding to the
missing face we used in stretching process.
Clearly, the plane graph is also connected, the degree of every vertex is at least 3 and
the degree of every face is also at least 3. Also, the graph is simple.
A simple connected plane graph G is called polyhedral if d(v) 3, for each vertex v of
G, and d( ) 3, for every face of G.
The following is a simple but useful property of polyhedral graphs.
Theorem 6.22 Let P be a convex polyhedron and G be its polyhedral graph. For each
n 3, let ni denote the number of vertices of G of degree i and let f i denote the number of
faces of G of degree i. Then
a. ini = i f i = 2m , where m is the number of edges of G.
i3

i3

b. The polyhedron P, and so the graph G, has at least one face bounded by a cycle of
length n for either n = 3, 4 or 5.

Proof
a. The expression ini is simply
i3

We know,

d(v), since d(v) 3 for each vertex v.

vV (G)

d(v) = 2m, so that ini = 2m .

i3

vV (G)

For each n 3, the expression i f i is obtained by going round the boundary of each
face having a cycle of length i as its boundary, counting up the edges as we go. If we
do this over all possible i, then we count each edge twice, and so i f i = 2m .
i3

b. Assume to the contrary that P has no faces bounded by cycles of length 3, 4 or 5.

Then f 3 = f 4 = f 5 = 0.
So by (a), 2m = i f i 6 f i = 6 f i = 6 f , where f is the total number of faces
i6

i6

i6

of P. Thus, f 31 m.
Also by (a), 2m = ini 3ni = 3 ni = 3n, where n is the total number of
i3

i3

i3

vertices of P. Thus, n 23 m .
Now, by Eulers formula, m = n + f 2 and we have

Graph Theory

155

2
1
m m + m 2 = m 2 , and this implies that 2 0,
3
3

which is impossible. Hence P must have a face of length 3, 4 or 5.

Regular polyhedra: A polyhedron is called regular if it is convex and its faces are congruent regular polygons (so that the polyhedral angles are all equal). The regular polyhedra
are also called Platonic bodies or Platonic solids.
A fact that has been known for at least 2000 years is that there are only five regular
polyhedra, which is proved next by using graph theoretic argument.
Theorem 6.23 The only regular polyhedra are the tetrahedron, the cube, the octahedron,
the dodecahedron and the icosahedron.

Proof Let P be a regular polyhedron and let G be its corresponding polyhedral graph.
Let n, m and f denote the number of vertices, edges and faces of G. Since the faces of
G are congruent to each other, each one is bounded by the same number of edges. Thus,
d( ) = k, k 3 for each face of G. Also, we know that there is at least one cycle of length
3, 4 or 5, therefore, 3 k 5. Similarly, since the polyhedral angles are all equal to each
other, the graph G is regular, say r-regular, where r 3. Thus we have
rn = 2m and k f = 2m, so that
2m = rn = k f .

(6.23.1)

Using this in Eulers formula n m + f = 2, we have

8 = 4n 4m + 4 f = 4n 2m + 4 f 2m = 4n rn + 4 f k f .

Therefore, 8 = (4 r)n + (4 k) f .

(6.23.2)

Since n and f are both positive and also 3 k 5 and r = 3, there are only five possible
cases (Fig. 6.19).
Case 1

Let r = 3 and k = 3. Then 8 = (4 3)n + (4 3) f , so that 8 = n + f .

From (6.23.1.), n = f . So, n + f = 8 gives n = 4, f = 4.

Thus, n = 4, m = 6, f = 4, which is a tetrahedron.
Case 2

Let r = 3, k = 4. Then from (6.23.2), 8 = (4 3)n + (4 4) f .

So, n = 8. Then, (6.23.1) gives 3n = 4 f , so that f = 6. Also, m = 12.

So, n = 8, m = 12, f = 6, and thus is a cube.

156

Planarity

Fig. 6.19

Graph Theory
Case 3

157

Let r = 3, k = 5. Then (6.23.2) gives 8 = (4 3)n + (4 5) f , so that 8 = n f .

From (6.23.1), 3n = 5 f . Solving 8 = n f and 3n = 5 f , we get n = 20, f = 12.

Thus, n = 20, m = 30, f = 12, which is a dodecahedron.
Case 4

Let r = 4, k = 3. Then we get n = 6, f = 8.

So, n = 6, m = 12, f = 8, which is an octahedron.

Case 5

Let r = 5, k = 3. Then we have n = 12, f = 20.

Thus, n = 12, m = 30, f = 20, which is an icosahedron.

The following is an elegant necessary condition for a plane graph to be Hamiltonian and
is due to Grinberg [86].
Theorem 6.24

If G is a loopless plane graph having a Hamiltonian cycle C, then (i

i=2

2)(i0 i00 ) = 0, where i0 and i00 are the number of regions of G of degree i contained in
interior C and exterior C respectively.

Proof Let E 0 and E 00 denote the sets of edges of G contained in int Cand ext C, respectively, and let |E 0 | = m0 and |E 00 | = m00 . Then int C contains exactly m0 + 1 regions (Fig. 6.20)
and so
n

i0 = m0 + 1,

(6.24.1)

i=2

Since G is loopless, 10 = 100 = 0.

Fig. 6.20

158

Planarity

Also, each edge in int C is on the boundary of exactly two regions in int C and each edge
of C is on the boundary of exactly one region in int C. Thus, counting the edges of all the
regions in int C, we obtain
n

ii00 = 2m0 + n.

(6.24.2)

i=2

Eliminating m0 between (6.24.1) and (6.24.2), we get

n

(i 2)i00 = n 2.

(6.24.3)

i=2

Similarly, (i 2)i00 = n 2.

(6.24.4)

i=2

The required result follows from equation (6.24.3) and (6.24.4).

Example Consider the Herschel graph G shown in Figure 6.21. Clearly, G has 9 regions
and all the regions are of degree 4. Thus, if G were Hamiltonian, then 2(40 400 ) = 0. This
implies 40 = 400 , which is impossible, since 40 + 400 = number of regions of degree 4 in
G = 9 is odd. Hence G is non-Hamiltonian.

Fig. 6.21

Konig observed that every graph is embeddable on some orientable surface. This is seen
by drawing an arbitrary graph G on the plane, possibly with edges that cross over each
other and then attaching a handle to the plane at each crossing and allowing one edge to
go over the handle and the other under it. Konig further showed that any embedding of
a graph on an orientable surface with a minimum number of handles has all its regions
simply connected.
Obviously, planar graphs can be embedded on a sphere. A toroidal graph is a graph
embedded on a torus, for instance, K5 , K3, 3 , K7 and K4, 4 .

Graph Theory

159

Definition: The genus (G)of a graph G is the minimum number of handles which are
added to the sphere so that G can be embedded on the resulting surface. Obviously (G) = 0
if and only if G is planar. The homeomorphic graphs have the same genus.
The genus of a polyhedron is the number of handles required on the sphere for a surface
to contain the polyhedron.
Definition: The least number of planar subgraphs whose union is a given graph G is
called the thickness of G. Clearly the thickness of any planar graph is 1. Also, K5 and K3, 3
have thickness 2, while the thickness of K9 is 3.
The crossing number of a graph G is the minimum number of pairwise intersections of
its edges when G is drawn in the plane. Clearly, crossing number of a graph is zero if and
only if G is planar.
The following result is due Euler and the proof can be found in Courant and Robbins [61].
Theorem 6.25

For a polyhedron of genus with n vertices, m edges and f regions

n m + f = 2 2 .

We have the following observations.

If G is a connected graph of genus , then

m=

3(n 2 + 2 ),

2(n 2 + 2 ),

i f every region in G is a triangle,

i f every region in G is a quadrilateral.

From this, we have the following.

If G is a connected graph of genus , then 61 m 12 (n 2), and if G has no triangles,
then 41 m 12 (n 2).
The proof of the following result is due to Ringel and Youngs [223] and some details
can be found in Harary [104].
Theorem 6.26 For n 3, the genus of the complete graph is
(Kn ) =

(n 3)(n 4)
.
12

6.7 Decomposition of Some Planar Graphs

Schnyder [232] proved that each triangulated planar graph G can be decomposed into three
edge disjoint trees. If instead triangles all faces (or regions) of G are quadrilaterals, and G
is without loops and multiple edges, Petrovic [183] proved that two trees are sufficient.

160

Planarity

The following two lemmas are used in proving Theorem 6.27.

Lemma 6.1 If G is a planar graph on n(n 4) vertices whose all regions are quadrilaterals, then |E(G)| = 2n 4.
Lemma 6.2 If G is a planar graph on n(n 4) vertices whose all regions are quadrilaterals, then G contains at least three vertices of degree 3.
The following result is due to Petrovic [183].
Theorem 6.27 (Petrovic) Let G be a planar graph on n(n 4) vertices whose all faces
are quadrilaterals, and vr and vb any two non-adjacent vertices of G (we assume vr is
coloured red and vb blue). Then the edges of G can be partitioned into red and blue ones so
that red ones form a spanning tree Tr of G vb , and blue ones a spanning tree Tb of G vr .

Proof We call Tr the red, and Tb the blue tree. A vertex of G is red (respectively blue)
if it belongs to Tr (respectively Tb ). A vertex is red-blue if it belongs to both Tr and Tb .
Thus vr is the only red, and vb is only blue vertex in G. We proceed by induction on n. For
n = 4, the assertion is obvious. Assume that it holds for each graph on less than n(n > 4)
vertices, and consider a graph G on n vertices. Let u (u 6= vr , vb ) be a vertex of G of degree
3 existing by Lemma 6.2.
a. d(u) = 2. Let uv1 , uv2 be the edges, and uv1 xv2 u, uv1 yv2 u the faces incident with u.
Since G 6= C4 , each of vertices v1 and v2 has degree 3. It implies that all regions of
G u are quadrilaterals. By induction hypothesis G u can be partitioned into trees
Tr0 and Tb0 . Now, independent of the kind of vertices v1 and v2 , (red, blue or red-blue)
we can always colour the edges uv1 and uv2 so that the obtained trees Tr and Tb are as
desired.
b. d(u) = 3. Let uv1 , uv2 , uv3 be the edges and uv1 xv2 u, uv2 yv3 u, uv3 zv1 u be the faces
incident with u.
First assume that all vertices x, y, z are distinct. Since no triangle can be partitioned
into disjoint quadrilaterals whose vertices are vertices of the triangle and some its inner
points, xy, yz, zx / E(G). Suppose one of vertices x, y, z, is blue and there is no edge
connecting it with the opposite vertex of the hexagon v1 xv2 yv3 z. Without loss of generality,
we may assume it is x. Then xv3
/ E(G). Denote by G00 the graph obtained from G by
deleting the edges xv1 and xv2 and identifying vertices u and x. Since each face of G00 is
a quadrilateral it can be partitioned into two trees Tr00 and Tb00 , where V (Tr00 )= V (G00 vb )
and V (Tb00 ) = V (G00 vr ). If edges uv1 and uv2 are coloured differently in G00 0 , say uv1 red
and uv2 blue, we colour xv1 red and xv2 blue obtaining the required trees Tr = Tr0 + xv1 and
Tb = Tb00 + xv2 in G. Therefore assume that uv1 and uv2 are monocoloured, say red. Then one
of vertices v1 and v2 , say v1 , is redblue. As u is also a red-blue vertex, there is the unique
blue u v1 path P in G00 . If P contains uv3 , we colour xv1 and xv2 blue and red, respectively.
If not, we colour both edges xv1 and xv2 red, and recolour uv1 blue. It is routine to check
that obtained trees Tr = Tr00 + xv1 , Tb = Tb00 + xv2 in the first case, and Tr = Tr00 uv1 + xv1 + xv2 ,
Tb = Tb00 + uv1 in the second, are as required.

Graph Theory

161

Now, assume none of the vertices x, y, z satisfy the condition above. Since at most one of
edges xv3 , yv1 , zv2 can exist, we may assume without loss of generality that x is red-blue, y
is the red, z is the blue vertex, xv3 E(G) and yv1 / E(G). Then each region of the graph G000
obtained from G by deleting edges yv2 and yv3 and identifying y and u is a quadrilateral. By
induction ypothesis, G000 can be composed into trees Tr000 and Tb000 , where V (T 000 ) = V (G000 z)
and V (Tb000 ) = V (G000 y). Colouring all edges incident with y red, and recolouring the edge
uv3 blue, we obtain the decomposition of G into trees Tr and Tb , where V (Tr ) = V (G z) and
V (Tb ) = V (G y).
Suppose two of vertices x, y, z, say y and z, coincide. Then d(v3 ) = 2. If v3 is redblue, we finish the proof as in case (a). If x is red-blue we proceed as above. So we
may assume that {v3 , x} = {vr , vb ). Let v3 = vr , x = vb . By induction hypothesis the graph
GIV = G v3 can be decomposed into trees TrIV and TbIV , where V (TrIV ) = V (GIV x) and
V (TbIV ) = V (GIV u). Then edges uv1 , uv2 are coloured red and edges xv1 , xv2 , blue. It
implies that edges yv1 , yv2 are coloured differently, yv1 , red and yv2 , blue. Now, colouring
v3 u and v3 y red and recolouring uv1 blue, we get from TrIV and TbIV the desired trees Tr and
Tb , where V (Tr ) = V (Tr ) = V (G x) and V (Tb ) = V (G v3 ).
If x = y = z, then v(G) = {u, x, v1, v2 , v3 }, E(G) = {uv1 , uv2 , uv3, xv1 , xv2 , xv3 } and the statement holds trivially.
q

6.8 Exercises
1. If G is a connected planar graph of order less than 12, prove that (G) = 4.
2. If G is a planar graph of order 24 and is regular of degree 3, then what is the number
of regions in a planar representation of G?
3. Prove that Eulers formula fails for disconnected graphs.
4. Show that every graph with at most three cycles is planar.
5. Find a simple graph G with degree sequence [4, 4, 3, 3, 3, 3] such that
a. G is planar,
b. G is nonplanar.
6. Show that every simple bipartite cubic planar graph contains a C4 .
7. Prove that a simple planar graph has at least 4 vertices of degree 5 at most.
8. Let G be a planar, triangle free graph of order n. Prove that G has no more than 2n 4
edges.
9. Let G be of order n = 11. Show that at least one of G or G is non planar.
10. Show that the average degree of a planar graph is less than 6.
11. Use Kuratowskis theorem to prove that the Peterson graph is nonplanar.

162

Planarity

12. Show that the edges forming a spanning tree in a planar graph G correspond to the
edges forming a set of chords in the dual G .
13. Prove that the complete graph K4 is self-dual.

7. Colourings
Colouring is one of the important branches of graph theory and has attracted the attention
of almost all graph theorists, mainly because of the four colour theorem, the details of
which can be seen in Chapter 12.

7.1 Vertex colouring

A vertex colouring (or simply colouring) of a graph G is a labelling f : V (G) {1, 2, . . .};
the labels called colours, such that no two adjacent vertices get the same colour and each
vertex gets one colour. A k-colouring of a graph G consists of k different colours and G is
then called k-colourable. A 2-colourable and a 3-colourable graph are shown in Figure 7.1.
It follows from this definition that the k-colouring of a graph G(V, E) partitions the vertex
set V into k independent sets V1 , V2 , . . ., Vk such that V = V1 V2 . . . Vk . The independent
sets V1 , V2 , . . ., Vk are called the colour classes and the function f : V (G) {1, 2, . . ., k}
such that f (v) = i for v Vi , 1 i k, is called the colour function.

Fig. 7.1

164

Colourings

The minimum number k for which there is a k-colouring of the graph is called the chromatic
number (chromatic index) of G and is denoted by (G). If (G) = k, the graph G is said to
be k-chromatic.
We observe that colouring any one of the components in a disconnected graph does
not affect the colouring of its other components. Also, parallel edges can be replaced by
single edges, since it does not affect the adjacencies of the vertices. Thus, for colouring
considerations, we opt only for simple connected graphs.
The following observations are the immediate consequences of the definitions introduced above.
1. A graph is 1-chromatic if and only if it is totally disconnected.
2. A graph having at least one edge is at least 2-chromatic (bichromatic).
3. A graph G having n vertices has (G) n.
4. If H is subgraph of a graph G, then (H) (G).
5. A complete graph with n vertices is n-chromatic, because all its vertices are adjacent.
So, (Kn ) = n and (K n ) = 1. Therefore we see that a graph containing a complete
graph of r vertices is at least r-chromatic. For example, every graph containing a
triangle is at least 3-chromatic.
6. A cycle of length n 3 is 2-chromatic if n is even and 3-chromatic if n is odd. To see
this, let the vertices of the cycle be labelled 1, 2, . . ., n, and assign one colour to odd
vertices and another to even. If n is even, no adjacent vertices get the same colour,
if n is odd, the nth vertex and the first vertex are adjacent and have the same colour,
therefore need the third colour for colouring.
7. If G1 , G2, . . ., Gr are the components of a disconnected graph G, then
(G) = max (Gi ).
1ir

We note that trees with greater or equal to two vertices are bichromatic as is seen in the
following result.
Theorem 7.1

Every tree with n 2 vertices is 2-chromatic.

Proof Let T be a tree with n 2 vertices. Consider any vertex v of T and assume T to
be rooted at vertex v (Fig. 7.2). Assign colour 1 to v. Then assign colour 2 to all vertices
which are adjacent to v. Let v1 , v2 , . . ., vr be the vertices which have been assigned colour
2. Now assign colour 1 to all the vertices which are adjacent to v1 , v2 , . . ., vr . Continue this
process till every vertex in T has been assigned the colour. We observe that in T all vertices
at odd distances from v have colour 2, and v and vertices at even distances from v have
colour 1. Therefore along any path in T , the vertices are of alternating colours. Since there
is one and only one path between any two vertices in a tree, no two adjacent vertices have
the same colour. Thus T is coloured with two colours. Hence T is 2-chromatic.
q

Graph Theory

165

Fig. 7.2

The converse of the above theorem is not true, i. e., every 2-chromatic graph need not
be a tree. To see this, consider the graph shown in Figure 7.3. Clearly, G is 2-chromatic,
but is not a tree.

Fig. 7.3

The next result due to Konig [134] characterises 2-chromatic graphs.

Theorem 7.2 (Konig)
cycles.

A graph is bicolourable (2-chromatic) if and only if it has no odd

Proof Let G be a connected graph with cycles of only even length and let T be a spanning
tree in G. Then, by Theorem 7.1, T can be coloured with two colours. Now add the chords
to T one by one. As G contains cycles of even length only, the end vertices of every chord
get different colours of T . Thus G is coloured with two colours and hence is 2-chromatic.
Conversely, let G be bicolourable, that is, 2-chromatic. We prove G has even cycles only.
Assume to the contrary that G has an odd cycle. Then by observation (6), G is 3-chromatic,
a contradiction. Hence G has no odd cycles.
q
Corollary 7.1

For a graph G, (G) 3 if and only if G has an odd cycle.

The following result is yet another characterisation of 2-chromatic graphs.

Theorem 7.3

A nonempty graph G is bicolourable if and only if G is bipartite.

166

Colourings

Proof Let G be a bipartite graph. Then its vertex set V can be partitioned into two
nonempty disjoint sets V1 and V2 such that V = V1 V2 . Now assigning colour 1 to all vertices in V1 and colour 2 to all vertices in V2 gives a 2-colouring of G. Since G is nonempty,
(G) = 2.
Conversely, let G be bicolourable, that is, G has a 2-colouring. Denote by V1 the set of
all those vertices coloured 1 and by V2 the set of all those vertices coloured 2. Then no two
vertices in V1 are adjacent and no two vertices in V2 are adjacent. Thus any edge in G joins
a vertex in V1 and a vertex in V2 . Hence G is bipartite with bipartition V = V1 V2 .
q

7.2 Critical Graphs

If G is a k-chromatic graph and (G v) = k 1 for every vertex v in G, then G is called a kcritical graph. A 4-critical graph is shown in Figure 7.4. If G is k-chromatic, but (G e) =
k 1 for each edge e of G, then G is called k-edge-critical graph, or k-minimal. A graph G
is said to be contraction critical or con-critical if (H) < (G) for every proper contraction
H of G. A graph G is said to be critical if (H) < (G) for every proper subgraph H of G.

Fig. 7.4

We have the following observations.

1. Every critical or minimal graph is connected.
2. Every connected k-chromatic graph contains a critical or minimal k-chromatic graph.
3. (G) = max { (B) : B is a block of G}.
4. The only 1-critical or 1-minimal graph is K1 , the only 2-critical or 2-minimal graph
is K2 and the only 3-critical or 3-minimal graphs are C2n+1 , n 1, that is, odd cycles.
The following result due to Dirac [66] describes some of the important properties of a
k-critical graph.

Theorem 7.4

If G is a k-critical graph, then

a. G is connected,

Graph Theory

167

b. (G) k 1,
c. G has no pair of subgraphs G1 and G2 for which G = G1 G2 , and G1 G2 is a complete
graph,
d. G v is connected for every vertex v of G, provided k > 1.

Proof
a. Assume G is not connected. Since (G) = k, by observation 7, there is a component
G1 of G such that (G1 ) = k. If v is any vertex of G which is not in G1 , then G1 is a
component of the subgraph G v. Therefore, (G v) = (G1 ) = k. This contradicts
the fact that G is k-critical. Hence G is connected.
b. Let v be a vertex of G so that d(v) < k 1. Since G is k-critical, the subgraph G v
has a (k 1)-colouring. As v has at most k 2 neighbours, these neighbours use at
most k 2 colours in this (k 1)-colouring of G v . Now, colour v with the unused
colour and this gives a (k 1)-colouring of G. This contradicts the given assumption
that (G) = k. Hence every vertex v has degree at least k 1.
c. Let G = G1 G2 , where G1 and G2 are subgraphs with G1 G2 = Kt . Since G is kcritical, therefore G1 and G2 both have chromatic number at most k 1. Consider
a (k 1)-colouring of G1 and a (k 1)-colouring of G2 . As G1 G2 is complete, in
the overlap, every vertex in G1 G2 has a different colour (in each of the (k 1)colourings). This implies that colours in the (k 1)-colouring of G2 can be rearranged
such that it assigns the same colour to each vertex in G1 G2 , as is given by the
colouring of G1 . Combining the two colourings then produces a (k 1)-colouring of
all of G. This is impossible, since (G) = k. Thus no subgraphs of the type G1 and G2
exist.
d. Assume G v is disconnected, for some vertex v of G. Then G v has a subgraph H1
and H2 with H1 H2 = G v and H1 H2 = . Let G1 and G2 be the subgraphs of G,
where G1 is induced by H1 and v, while G2 is induced by H2 and v. Then G = G1 G2
and G1 G2 = K1 (with K1 as a single vertex). This contradicts (c) and thus G v is
connected.
q
Let S = {u, v} be a 2-vertex cut of a critical k-chromatic graph G. Since no separating
set of a critical graph is a complete graph, therefore uv is not an edge of G. Let Gi be the
S-component of G. Gi is said to be of type 1 if every (k 1) colouring of Gi assigns the same
colour to u and v, Gi is of type 2 if every (k 1)-colouring of Gi assigns different colours to
u and v, and Gi is of type 3 if some (k 1)-colouring of Gi assigns same colour to u and v,
while some other (k 1)-colouring assigns different colours to u and v.
The following characterisation of k-critical graphs with a 2-vertex cut is due to Dirac.
Theorem 7.5 If G is a minimal k-chromatic graph with a 2-vertex cut S = {u, v}, then (i)
G = G1 G2 , where Gi is the S-component of type i, i = 1, 2 and (ii) both G + uv and G : uv
are k-minimal.

168

Colourings

Proof Let G be a minimal k-chromatic graph with a 2-vertex cut S = {u, v}. So the
S-components of G are (k 1)-colourable. If there is no set of (k 1)-colourings of the Scomponents all of which agree on S, then G is (k 1)-colourable. Therefore there is a type
1 S-component G1 and a type 2 S-component G2 . Then G1 G2 is not (k 1)-colourable.
Since G is k-critical, there is no third S-component G3 . Hence, G = G1 G2 .
Let H1 = G1 + uv and H2 = G2 : uv. We prove that H2 is k-minimal. Since G2 is of type 2,
therefore every (k 1)-colouring of G2 assigns different colours to u and v. As u and v are
identified to a single vertex, say w in H2 , so a k-colouring is necessary to colour H2 , that
is, H2 is k-chromatic. We further prove that (H2 e) = k 1 for any edge of H2 . Any such
edge e can be considered to belong to G and in the (k 1)-colouring of G e, u and v get
the same colour, since they can be considered to belong to G1 which is a subgraph of G e.
The restriction of such a colouring of G e to H2 e (with u and v identified as w with the
common colour of u and v) is a (k 1)-colouring of H2 e. This proves the result.
That H1 is minimal, can be proved in a similar manner.
q
Theorem 7.6
graph.

Every k-chromatic graph can be contracted into a con-critical chromatic

Proof Let G be a k-chromatic graph and let the edge e of G be contracted. Then a colouring of G can be used to give a colouring of G|e except that, possibly the vertex formed
by the contraction may be assigned an extra colour. Thus, (G|e) (G) + 1. On the other
hand, a colouring of G|e can be used to get a colouring for G by using an extra colour
for one of the end vertices of e. Therefore, (G) (G|e) + 1. Thus the contraction of an
edge changes the chromatic number by at most one. Sometimes contraction of an edge
may increase the chromatic number, but by repeated contractions, the number of edges and
therefore the chromatic number gets reduced. Clearly, the connected graph can be contracted to a single vertex whose chromatic number is one. In between, a stage arises where
the chromatic number of the graph is the same as the original, but the contraction of any
edge reduces the chromatic number by one.
q
As noted earlier, every connected k chromatic graph contains a critical or minimal kchromatic graph. To see this, we observe that if G is not k-critical, then (G v) = k, for
some vertex v of G. If G v is kcritical, then this is the required subgraph. If not, then
G {v, w} = (G v) w has chromatic number k, for some vertex w in G v. If this new
subgraph is k-critical, then again this is the required subgraph. If not, we continue this
vertex deletion procedure, and we will clearly get a k-critical subgraph.
We have the following immediate observation.
Theorem 7.7

Any k-chromatic graph has at least k vertices of degree at least k 1 each.

Proof Let G be a k-chromatic graph and let H be a k-critical subgraph of G. Then, by

Theorem 7.4 (b), every vertex of H has degree at least k 1 in H and hence in G. Since H
is k-chromatic, H has at least k vertices. This completes the proof.
q

Graph Theory

169

We note that there is no easy characterisation of graphs with chromatic number greater
or equal to three. The graph vertex - colouring problem is a standard NP-complete problem
and no good algorithm for finding (G) has been discovered for the class of all graphs,
though for some special classes of graphs polynomial time algorithms have been found.
There are various results which give upper bounds for the chromatic number of an arbitrary
graph G, provided the degrees of all the vertices of G are known. The first of these is due
to Szekeres and Wilf [237].
Theorem 7.8
k 1.

Let G be a graph and k = max{ (G0 ) : G0 is a subgraph of G}. Then (G) =

Proof

Let H be a k-minimal subgraph of G. Then H is a subgraph of G and therefore

(H) k. Using Theorem 7.4, we have, (H) (H) 1 = (G) 1. Thus, (G) (H) +
1 = k + 1.
q
The next result is due to Welsh and Powell [262] and its proof is due to Bondy [32].
Theorem 7.9 Let G b a graph with degree sequence [di ]n1 such that d1 d2 . . . dn .
Then, (G) max{min{i, di + 1}}.

Proof Let G be k-chromatic. Then, by Theorem 7.8, G has at least k vertices of degree at
least k 1. Therefore, dk k 1 and max {min {i, di + 1}} min {k, dk + 1} = k = (G).

We have the following upper bounds for chromatic number.

Theorem 7.10

For any graph G, (G) 4(G) + 1.

Proof Let G be any graph with n vertices. To prove the result, we induct on n. For n =
1, G = K1 and (G) = 1 and 4(G) = 0. Therefore the result is true for n = 1.
Assume that the result is true for all graphs with n 1 vertices and therefore by induction
hypothesis, (G) 4(Gv)+1. This shows that Gv can be coloured by using 4(Gv)+1
colours. Since 4(G) is the maximum degree of a vertex in G, vertex v has at most 4(G)
neighbours in G. Thus these neighbours use up at most 4(G) colours in the colouring of
G v.
If 4(G) = 4(G v), then there is at least one colour not used by vs neighbours and that
can be used to colour v giving a 4(G) + 1 colouring for G.
In case 4(G) 6= 4(G v), then 4(G v) < 4(G). Therefore, using a new colour for v,
we have a 4(G v) + 2 colouring of G and clearly, 4(G v) + 2 4(G) + 1. Hence in both
cases, it follows that (G) 4(G) + 1.
q
Remarks

1. Clearly, Theorem 7.10 is a simple consequence of Theorem 7.7. This is because if G

is k chromatic, then Theorem 7.4 gives 4 k 1, that is, 4 + 1.
2. The equality in Theorem 7.10 holds if G = C2n+1 , n 1 and if G = Km .

170

Colourings

7.3 Brooks Theorem

Greedy colouring algorithm: The greedy colouring with respect to a vertex ordering
v1 , v2 , . . ., vn of V (G) is obtained by colouring vertices in the order v1 , v2 , . . ., vn assigning
to vi the smallest indexed colour not already used on its lower indexed neighbours.
This is reported in West [263].
The following recolouring technique as noted in Clark and Holton [60] is due to Kempe
[128].
Kempe Chain argument: Let G be a graph with a colouring using at least two different
colours represented by i and j . Let H(i, j) denote the subgraph of G induced by all the
vertices of G coloured either i or j and let K be a connected component of the subgraph
H(i, j). If we interchange the colours i and j on the vertices of K and keep the colours
of all other vertices of G unchanged, then we get a new colouring of G, which uses the
same colours with which we started. This subgraph K is called a Kempe chain and the
recolouring technique is called the Kempe chain argument (Fig. 7.5).

Fig. 7.5

The following result due to Brooks [39] is an improvement of the bounds obtained in
Theorem 7.10. We give two proofs of Brooks theorem, the first given by Lovasz [150] uses
greedy colouring, and the second proof uses Kempe chain argument.
Theorem 7.11 (Brooks)
cycle, then (G) 4(G).

If G is a connected graph which is neither complete nor an odd

Proof Let G be a connected graph with vertex set V = {v1 , v2 , . . ., vn } which is neither a
complete graph, nor an odd cycle and let 4 = k. Since G is a complete graph for k = 1 and
G is an odd cycle or a bipartite graph for k = 2, let k 3.

Graph Theory

171

Assume G is not k-regular. Then there exists a vertex say v = vn such that d(v) < k. Since
G is connected, we form a spanning tree of G starting from vn and whose vertices are
arranged in the order vn , vn1 , . . ., v1 (Fig. 7.6).

Fig. 7.6

Clearly, each vertex vi other than vn in the resulting order vn , vn1 , . . ., v1 has a higher
indexed neighbour along the path to vn in the tree. Therefore each vertex vi has atmost k 1
lower indexed neighbours and the greedy colouring needs at most k colours (Fig. 7.7).

Fig. 7.7

Now, let G be k-regular. Assume G has a cut vertex say x and let G0 be a subgraph
containing a component of G x together with the edges of G x to x. Clearly, d(x|G0 ) < k.
Therefore, by using the above argument, we have a k-colouring of G0 . By making use of
the permutations of the colours, it can be seen that this is true for all such subgraphs. Thus
G is k-colourable.
Now, let G be 2-connected. We claim that G has an induced 3-vertex path, with vertices
say v1 , vn , v2 in order, such that G {v1 , v2 } is connected.
To prove the claim, let x be any vertex of G. If k(G x) 2, let v1 be x and let v2 be a
vertex with distance two from x, which clearly exists, as G is regular and not a complete
graph. If k(G x) = 1, then x has a neighbour in every end block of G x, since G has no cut
vertex. Let v1 and v2 be the neighbours of x in two such blocks. Clearly v1 and v2 are non
adjacent. Also, since blocks have no cut vertices, G {x, v1 , v2 } is connected. As k 3, so
G {v1 , v2 } is connected and we let x = vn , proving the claim.
Now arrange the vertices of a spanning tree of G {v1 , v2 } as v3 , v4 , . . ., vn . As before,
each vertex before n has atmost k 1 lower indexed neighbours. The greedy colouring uses
at most k 1 colours on neighbours of vn , since v1 and v2 get the same colour.

Second Proof (Using Kempe chain argument) Let G be a connected graph with n
vertices which is neither complete nor an odd cycle. Let 4(G) = k. For k = 1, G is complete
and for k = 2, G is an odd cycle or a bipartite graph. Therefore, assume k 3.

172

Colourings

We induct on n. Since k 3, the induction starts from n = 4. As G is not complete, then

for n = 4, G is one of the graphs given in Figure 7.8.

Fig. 7.8

Clearly, for each such graph, the chromatic index is at most three.
Now, let n 5, and assume the result to be true for all graphs with fewer than n vertices.
If G has a vertex v of degree less than k, then it follows from Theorem 7.10 that G can
be coloured by k colours, since the neighbours of v use up atmost k 1 colours. Therefore
the result is true in this case.
Now assume that degree of every vertex of G is k, that is, G is k-regular. We show that G
has a k-colouring.
Let v be any vertex of G. Then by induction hypothesis, the subgraph G v has a kcolouring. If the neighbours of v in G do not use all the k colours in the k-colouring of
G v, then any unused colour is assigned to v giving a k-colouring of G. Assume that the k
neighbours of v are assigned all the k colours in the k-colouring of Gv. Let the neighbours
of v be v1 , v2 , . . ., vk which are coloured by the colours 1, 2, . . ., k respectively.
Let the Kempe chains Hvi (i, j) and Hv j (i, j) containing the neighbours vi and v j be different. That is, vi and v j are in different components of the subgraph H(i, j) induced by
the colours i and j . Therefore, using Kempe chain argument, the colours in Hvi (i, j) are
interchanged to give a k-colouring of G v, where now vi has been assigned the colour j .
This implies that the neighbours of v use less than k colours and the unused colour assigned
to v gives a k-colouring of G.
Now assume that for each i and j , the neighbours vi and v j are in the same Kempe
chain, which is briefly denoted by H . If the degree of vi in H is greater than one, then vi
is adjacent to at least two vertices coloured j . Therefore there is a third colour, say `, not
used in colouring the neighbours of vi . Recolour vi by ` and then colour v by i, giving the
k-colouring of G. Assume that vi and v j both are of degree one in H . Let P be a path from
vi to v j in H and let there be a vertex in P with degree at least three in H (Fig. 7.9). Let u
be the first such vertex and coloured i. (If u is coloured j , the same argument is used as in
case of i). Then at least three neighbours of u are coloured j and therefore there is a colour,
say `, not used by these neighbours. Recolour u by ` and interchange colours i and j on the
vertices of P from vi upto u, excluding u. So we get a colouring of G v, where vi and v j
are now both coloured j . This allows v to be coloured by i.

Fig. 7.9

Graph Theory

173

Let all the vertices on a path vi to v j , excluding the end vertices vi and v j , be of degree
two in H . Clearly H contains a single path from vi to v j .
Let all the Kempe chains be paths. Let H and K be such chains corresponding to vi , v j
and vi , v` respectively, with j 6= `. Let w 6= vi be a vertex present in both the chains (Fig.
7.10). Then w is coloured i, has two neighbours coloured j and two neighbours coloured
`. Therefore there is a fourth colour, say s, not used by the neighbours of w. Now colour w
by s, and interchange colours ` and i on the vertices of K beyond w upto and including v` ,
we get a colouring of G v, where vi and v` are now both coloured i. This allows v to be
coloured by `.

Fig. 7.10

Thus assume that two such Kempe chains meet only at their common end vertex vi .
Let vi and v j be two neighbours of v which are nonadjacent and let x be the vertex
coloured j , adjacent to vi on the Kempe chain H from vi to v j . With ` 6= j , let K denote the
Kempe chain from vi to v j . Then by the Kempe chain argument, we interchange the colours
in K , without changing the colours of the other vertices. This results in vi coloured `, and v`
coloured i. Since x is adjacent to vi , it is in the Kempe chain for colours ` and j . However,it
is also the Kempe chain for colours i and j . This contradicts the assumption that Kempe
chains have at most one vertex in common, the end vertex. This contradiction implies that
any two vi and v j are adjacent. In other words, all neighbours of v are also neighbours of
each other. This shows that G is the complete graph Kk , a contradiction to the hypothesis
of G.
q
Definition: In the depth first search tree (DFS), a search tree T is used to represent the
edge examination process. In DFS a new adjacent vertex is selected, which is incident
with the first edge incident with v. In other words, in DFS we leavev as quickly as possible,
examining only one of its incident edges and replacing v by a new vertex, which is adjacent
to v.
The following result is due to Chartrand and Kronk [51].
Theorem 7.12 Let G be a connected graph every depth-first search tree of which is a
Hamiltonian path. Then G is a cycle, a complete graph, or a complete bipartite graph Kn, n .

174

Colourings

Proof Let P be a Hamiltonian path of G, with origin u. Because the path P u extends to
a Hamiltonian path of G, the path P extends to a Hamiltonian cycle C of G.
When C has no chord, G = C is a cycle. So let uv be a chord of C. Then u v is one
too, because uCvuC1v is a Hamiltonian path of G, likewise, u v is a chord of C (where
u denotes the successor of u on C and u is the successor of u ). And if the length of
uCv is at least four, uv and u v are also chords of C , in view of the Hamiltonian path
uCv uC 1 v u uv and the fact that u v = (u ) v .
When C has a chord uw of length two, let v = u(= w). Then vw E . Moreover, if

vw E , then vw(1) E in view of the Hamiltonian path w(1)CuwCwv. It follows that

v is adjacent to every vertex of G. But then G is complete, because uw is a chord of length
two for all i. If C has no chord of length two, every chord of C is odd, moreover, every odd
chord must be present. Thus, G = Kn, n , where |V (G)| = 2n.
q
The following is the third proof of Brooks theorem which is due to Bondy [35].

Bondys Proof Suppose first that G is not regular. Let u be a vertex of degree and let T
be a search tree of G rooted at u. Colour the vertices with the colours 1, 2, . . ., 4 according
to the greedy heuristic, selecting at each step a pendent vertex of the subtree of T induced
by the vertices not yet coloured, assigning to it the smallest available colour and ending
with the root u of T . When each vertex v different from x is coloured, it is adjacent (in T )
to at least one uncoloured vertex and so is adjacent to at most d(v) 1 < 4 1 coloured
vertices. It is therefore assigning one of the colours 1, 2, . . ., 4, because d(u) = 4 1.
The greedy heuristic therefore produces a 4-colouring of G.
Now, let G be regular. If G has a cut vertex u, then G = G1 G2 , where G1 and G2 are
connected and G1 G2 = {u}. Because the degree of u in G is less than 4(G), neither
subgraph of G is regular, so (G) 4(Gi ) = 4(G), i = 1, 2 and (G) = max{ (G1 ), (G2 )}
4(G). Therefore we assume that G is 2-connected.
If every depth- first search tree of G is a Hamiltonian path, then G is a cycle, a complete
graph, or a complete bipartite graph Kn, n , by Theorem 7.12. Since by hypothesis, G is
neither an odd cycle nor a complete graph, (G) = 2 4(G). Suppose then, that T is a
depth-first search tree of G, but not a path. Let u be a vertex of T with at least two children,
v and w. Because G is 2-connected, both G v and G w are connected. Thus there are
proper descendants of v and w, each of which is joined to an ancestor of u, and it follows
that G {v, w} is connected. Consider a search tree T with root u in G. By colouring v and
w with colour 1 and then the vertices of T by the greedy heuristic as above, ending with the
root u, we obtain a 4-colouring of G.
q
Brooks theorem and the observation that in a graph G containing Kn as a subgraph,
(G) n, provide estimates for the chromatic number. For instance, in Figure 7.11(a) for
the graph G1 , 4(G1 ) = 8 and G1 has K4 as a subgraph. Therefore, 4 (G1 ) 8. It can be
easily seen that (G1 ) = 4. Similarly for G2 in Fig. 7.11(b) known as the Birkhoff diamond,
4(G2 ) = 5 and G2 has K3 as a subgraph. So, 3 (G2 ) = 5. In fact, (G2 ) = 4.
Independent set: A set of vertices in a graph G is independent if no two of them
are adjacent. The largest number of vertices in such a set is called the vertex indepen-

Graph Theory

175

dence numberindependence number of G and is denoted by 0 (G) or 0 . Analogously, an

independent set of edges of G has no two of its edges adjacent and the maximum cardinality of such a set
h n iis the edge independence number 1 (G) or 1 . For the complete graph
Kn , 0 = 1, 1 =
. In the graph of Figure 7.12, 0 (G) = 2 and 1 (G) = 3.
2

Fig. 7.11

Fig. 7.12

A lower bound, noted in Berge [22] and Ore [178] and an upper bound by Harary and
Hedetniemi [106] involve the vertex independent number 0 of a graph.
Theorem 7.13

For any graph G,

n
n

(G) n o + 1.

(G)
o

Proof If (G) = k, then V can be partitioned into k colour classes V1 , V2 , . . ., Vk , each of

which is an independent set of vertices.
If |Vi | = ni , then every ni o , so that n = ni k o . This proves the middle inequality.
Now, let S be a maximal independent set containing o vertices. Clearly, (G S)
(G) 1.
As G S has n o vertices, (G S) n o , and (G) (G S) + 1 n o + 1,
proving the last inequality.
As G has a complete subgraph of order o (G),
(G) o (G), or

n
n

, proving the first inequality.

(G) o (G)

176

Colourings

The following result due to Nordhaus and Gaddum [172] gives the bounds on the sum
and product of the chromatic numbers of a graph and its complement.
Theorem 7.14

For any graph of order n,

2 n + n + 1,

and n

n+1
2

(7.14.1)
2

(7.14.2)

where = (G) and = (G).

Proof

Evidently from Theorem 7.13, we have

n.

(7.14.3)

Since the arithmetic mean is greater than or equal to the geometric mean,
+ p
.
2

(7.14.4)

Combining (7.14.3) and (7.14.4), we get

+
n.
2

Therefore the left inequalities of (7.14.1) and (7.14.2) are proved.

Now, let d1 d2 . . . dn be the degree sequence of G. Then d 1 d 2 . . . d n , where
di = n 1 dn+1i , is the degree sequence of G. Then by using Theorem 7.9, we have
(G) + (G) max min{di + 1, i} + max min{n dn+1i , i}
i

= max min{di + 1, i} + (n + 1) min max{dn+1i + 1, n + 1 i}

i

= max min{di + 1, i} + (n + 1) min max{d j + 1, j}.

i

Thus, (G) + (G) n + 1.

p
(G) + (G) p
+
n+1
. Therefore,

. Thus,
Also,
2
2
2

n+1
2

2

Second Proof Let G be k-chromatic and let v1 , v2 , . . ., vk be the colour classes of G,

where |Vi | = ni Then ni = n and max ni n/k. Since each Vi induces a complete subgraph
of G, max ni n/k, so that n. As the geometric mean of twopositive numbers is
always less or equal to their arithmetic mean, it follows that + 2 n.

Graph Theory

177

To prove + n + 1, we induct on n. Clearly, the equality holds for n = 1. We assume that (G) + (G) n for all graphs G having fewer than n vertices. Let H and H
be complementary graphs with n vertices and let v be a vertex of H . Then G = H v and
G = H v are complementary graphs with n 1 vertices. Let the degree of v in H be d ,
so that degree of v in H is n d 1. Clearly, (H) (G) + 1 and (H) (G) + 1. If either (H) < (G) + 1 or (H) < (G) + 1, then (H) + (H) n + 1. If (H) = (G) + 1
and (H) = (G) + 1 , then the removal of v from H , producing G, decreases the chromatic
number, so that d (G). Similarly, n d 1 (G). Thus, (G) + (G) n 1. Therefore
we always have (H) + (H) n + 1. Applying now the inequality 4 ( + )2 , we get

n+1
2

2

We now have the following result.

Theorem 7.15 If a connected k-chromatic graph has exactly one vertex of degree exceeding k 1, then it is minimal.

Proof Let G be a connected k-chromatic graph having exactly one vertex of degree exceeding k 1. Let e be any edge of G. Then (G e) k 2 (otherwise, G will have at least
two vertices of degree exceeding k 1).
For every induced subgraph H of G e, we have (H) k 2. Thus, by Theorem 7.7,
(G e) k 1 and hence (G e) k 1. Since e is arbitrary, therefore G is minimal. q
We observe from Theorem 7.4(b) that the number of edges m of a k-critical graph is at
least n(k 1)/2. Dirac extended this to the inequality 2m n(k + 1) 2 for a (k + 1)-critical
graph, the proof of which can be found in Bollobas [29].

7.4 Edge colouring

An edge colouring of a nonempty graph G is a labelling f : E(G) {1, 2, . . .}; the labels are
called colours, such that adjacent edges are assigned different colours. A k-edge colouring
of G is a colouring of G which consists of k different colours and in this case G is said to
be k-edge colourable.
The definition implies that the k-edge colouring of a graph G(V, E) partitions the edge
set E into k independent sets E1 , E2 , . . ., Ek such that E = E1 E2 . . . Ek . The independent
sets Ei , 1 i k are called the colour classes and the function f : E(G) {1, 2, . . ., k} such
that f (e) = i, for each e Ei , 1 i k, is called the colour function. The minimum number k
for which there is a k-colouring of G is called the edge chromatic umber (or edge chromatic
index) and is denoted by 0 (G).
We have the following observations.
1. If H is a subgraph of a graph G, then 0 (H) 0 (G).
2. For any graph G, 0(G) 4(G).

178

Colourings

If v is any vertex of G with d(v) = 4(G), then the 4(G) edges incident with v have a
different colour in any edge colouring of G.
3. 0 (Cn ) =

2,
3,

if n is even ,
if n is odd .

The following is a recolouring technique for edge colouring, called Kempe chain argument.
Let G be a graph with an edge colouring using at least two different colours say i and j .
Let H(i, j) represent the subgraph of G induced by all the edges coloured either i or j . Let
K be a connected component of the subgraph H(i, j). It can be easily verified that K is
a path whose edges are alternately coloured by i and j . If the colours on these edges are
interchanged and the colours on all other edges of G are kept unchanged, the result is a new
colouring of G, using the same initial colours. The component K is called Kempe chain and
this recolouring method is called the Kempe chain argument.
Definition: Let i be a colour used in the edge colouring of a graph G. If there is an edge
coloured i incident at the vertex v of G, we say i is present at v, and if there is no edge
coloured i at v, we say i is absent from v.
The following result is due to Konig [136].
Theorem 7.16 (Konig)

For a nonempty bipartite graph G, 0 (G) = 4(G).

Proof The proof is by induction on the number of edges of G. If G has only one edge,
the result is trivial.
Let G have more than one edge and assume that the result is true for all nonempty
bipartite graphs having fewer edges than G. Since 4(G) 0 (G), it is enough to prove that
G has a 4(G)-edge colouring. We let 4(G) = k. Let e = uvbe an edge of G. Then G e is
bipartite with less edges than G. Therefore, by induction hypothesis, G e has a 4(G e)edge colouring. Since 4(G e) 4(G) = k, G e has a k-colouring. We show that the same
k colours are used to colour G.
As d(u) k in G and the edge e is uncoloured, there is at least one of the k colours absent
from u. Similarly, at least one of these colours is absent from v.
If one of the colours absent at u and v is same, we use this to colour e and we get a k-edge
colouring of G.
Now take the case of a colour i present at u, but absent from v and a colour j present at
v, but absent from u.
Let K be the Kempe chain containing u in the subgraph H(i, j) induced by the edges
coloured i or j . We claim that v does not belong to the Kempe chain K .
For if v belongs to K , then there is a path P in K from u to v. Since u and v are adjacent,
they do not belong to the same bipartition subset of the bipartite graph G and therefore the
length of the path P is odd. As the colour i is present at u, the first edge of P is coloured
i. Since the edges of P are alternately coloured i and j , and P is of odd length, therefore
the last edge of P, which is incident at v, is also coloured i. This is a contradiction, as i is
absent from v, proving our claim.

Graph Theory

179

Using Kempe chain argument on K , the interchanging of colours now makes i absent
from u and does not affect the colours of the edges incident at v. Therefore i is absent from
both u and v in this new edge colouring and colouring edge e by i gives a k-edge colouring
of G.
q
The next result gives edge chromatic number of complete graphs.
Theorem 7.17
0 (G) =

If G = Kn is a complete graph with n vertices, n 2, then

n 1,
n,

i f n is even ,
i f n is odd .

Proof Let G = Kn be a complete graph with n vertices.

Assume n is odd. Draw G so that its vertices form a regular polygon. Clearly, there are n
edges of equal length on the boundary of the polygon. Colour the edges along the boundary
using a different colour for each edge. Now, each of the remaining internal edges of G is
parallel to exactly one edge on the boundary. Each such edge is coloured with the same
colour as the boundary edge. So two edges have the same colour if they are parallel and
therefore we have the edge colouring of G. Since it uses n colours, we have shown that
0 (G) n.
Let G have an (n 1)-colouring. From the definition of an edge colouring, the edges
of one particular colour form a matching in G (set of independent edges). Since n is odd,
therefore the maximum possible number of these is (n 1)/2. This implies that there are
atmost (n 1)(n 1)/2 edges in G. This is a contradiction, as Kn has n(n 1)/2 edges. Thus
G does not have an (n 1) colouring. Hence, 0 (G) = n.
Now, let n be even, and let v be any vertex of G. Clearly G v is complete with n 1
vertices. Since n 1 is odd, G v has an (n 1)-colouring. With this colouring, there is a
colour absent from each vertex and different vertices having different absentees. Reform G
from G v by joining each vertex w of G v to v by an edge and colour each such edge by
the colour absent from w. This gives an (n 1)-colouring of G and therefore 0 (G) = n 1.
q

The above result is illustrated by taking K5 and K6 in Figure 7.13.

Fig. 7.13

180

Colourings

Since 4(G) = n 1 in a complete graph, Theorem 7.17 shows that 0 (Kn ) is either 4(G)
or 4(G) + 1. The next result obtained by Vizing [258] and independently by Gupta [94]
gives the tight bounds for edge chromatic number of a simple graph.
Theorem 7.18 (Vizing)

For any graph G, 4(G) 0 (G) 4(G) + 1.

Proof Let G be a simple graph, we always have 4(G) = 0 (G).

To prove 0 (G) 4(G)+1, we use induction on the number of edges of G. Let 4(G) = k.
If G has only one edge, then k = 1 = 0 (G). Therefore assume that G has more than one edge
and that the result is true for all graphs having fewer edges than G.
Let e = v1 v2 be an edge of G. Then by induction hypothesis the subgraph G e has
(k + 1)-edge colouring and let the colours used be 1, 2, . . ., k + 1.
Since d(v1 ) k and d(v2 ) k, out of these k + 1 colours at least one colour is absent from
v1 and at least one colour is absent from v2 . If there is a common colour absent from both
v1 and v2 , then we use this to colour e and get a (k + 1)-colouring of G. Therefore in this
case, 0 (G) k + 1.
We now assume that there is a colour, say 1, absent from v1 but present at v2 and there
is a colour, say 2, absent from v2 but present at v1 . We start from v1 and v2 and construct a
sequence of distinct vertices v1 , v2 , . . ., v j , where each vi for i 2 is adjacent to v1 . Let v1 v3
be coloured 2. This v3 exists, because 2 is present at v1 . We observe that not all the k + 1
colours are present at v3 and assume that the colour 3 is absent from v3 . But the colour 3 is
present at v1 and choose the vertex v4 so that v1 v4 is coloured 3. Continuing in this way, we
choose a new colour i absent from vi but present at v1 , so that v1 vi+1 is the edge coloured i.
In this way, we get a sequence of vertices v1 , v2 , v3 , . . ., v j1, v j such that
a. vi is adjacent to v1 for each i > 1,
b. the colour i is absent from each i = 1, 2, . . ., j 1 and
c. the edge v1 vi+1 is coloured i for each i = 1, 2, . . ., j 1.
This is illustrated in Figure 7.14.

Fig. 7.14

Graph Theory

181

As d(v1 ) k, (a) implies that such a sequence has at most k + 1 terms, that is, j k + 1.
Assume that v1 , v2 , . . ., v j is a longest such sequence, that is, the sequence for which it is
not possible to find a new colour j , absent from v j , together with a new neighbour v j+1 of
v1 such that v1 v j+1 is coloured j .
We first assume that for some colour j absent from v j there is no edge of that colour
present at v1 . We colour the edge e = v1 v2 by colour 2 and then recolour the edges v1 v j by
colour i, for i = 3, . . ., j 1. Since i was absent from vi , for each i = 2, . . ., j 1, this gives a
(k + 1)-colouring of the subgraph G v1 v j . Now as the colour j is absent from both v j and
v1 , recolour v1 v j by the colour j . This gives a (k + 1)-colouring of G (Fig. 7.15).

Fig. 7.15

Now assume that whenever j is absent from v j , j is present at v1 . If v j+1 is a new

neighbour of v1 so that v1 v j+1 is coloured by j , then we have extended our sequence to
v1 , v2 , . . ., v j , v j+1 which is a contradiction to the assumption that v1 , v2 , . . ., v j is the
longest sequence. Thus one of the edges v1 v3 , . . ., v1 v j1 is to be coloured by j , say v1 v` ,
with 3 ` j 1. Now colour e = v1 v2 by 2, and for i = 3, . . ., ` 1, recolour each of the
edges v1 vi by i while unaltering the colours of the edges v1 vi , for i = ` + 1, . . ., j . Removing
the colour j from v1 v` , we have a (k + 1)-colouring of the edge deleted subgraph G v1 v` .
Let H(1, j) represent the subgraph of G induced by the edges coloured 1 or j in this
partial colouring of G. Since the degree of every vertex in H(1, j) is either 1 or 2, each
component of H(1, j) is either a path or a cycle. As 1 is absent from v1 and j is absent from
both v j and v` , it follows that all these three vertices do not belong to the same connected
component of H(1, j). Therefore, if K and L represent the corresponding Kempe chains
containing v j and v` respectively, then either v1
/ K or v1
/ L. Let v1
/ L. Then interchanging
the colours of L, the Kempe chain argument gives a (k + 1)-colouring of G v1 v` in which
1 is missing from both v1 and v` . Colouring v1 v` by 1 gives a (k + 1)-colouring of G.
Now, let v1 / K . Colour the edge v1 v` by `, recolour the edges v1 vi by i, for i = `, . . .,
j 1, and remove the colour j 1 from v1 v j . Then, from the definition of the sequence
v1 , v2 , . . ., v j , we get a (k +1)-colouring of Gv1 v j without affecting two coloured subgraph
H(1, j). Using Kempe chain argument to interchange the colours of K , we obtain a (k + 1)colouring of G v1 v j in which 1 is absent from both v1 and v j . Therefore, again colouring
v1 v j by 1 gives (k + 1)-colouring of G.
q

182

Colourings

The following result is due to Vizing [259] and Alavi and Behzad [2].
Theorem 7.19

Let G be a graph of order n and let G be the complement of G. Then

a. n 1 0 (G) + 0 (G) 2(n 1),

0 0 (G) 0 (G) (n 1)2 , for even n,

b. n 0 (G) + 0 (G) 2n 3,
0 0 (G) 0 (G) (n 1)(n 2), for odd n.

Further, the bounds are the best possible for every positive integer n(n 6= 2).

Proof

Let G be a graph of order n and let G be the complement of G. Clearly,

4(G) n 1 4(G), so that 4(G) + 4(G) n 1.

Therefore, combining with 0 (G) 4(G), we get

0 (G) + 0 (G) n 1.

b. If n is odd, we have 0 (G) + 0 (G) n, since 0 (G) + 0 (G) < n implies 0 (Kn ) < n,
which is a contradiction.
Obviously, 0 (G) 0 (G) 0. It can be seen that the lower bounds are attained in
complete graphs Kn .
We now prove that 0 (G) + 0 (G) 2n 3 and 0 (G) 0 (G) (n 1)(n 2).
Clearly, for n = 1, 3, the inequalities 0 (G)+ 0 (G) 2n 3 are true. So, let n 5.
If 4(G) + 4(G) 2n 5, then by Vizings theorem, we get 0 (G) + 0 (G) 2n 3.
Otherwise, we have the following cases.
i. 4(G) = n1 and 4(G) = n2. So G has a pendant vertex v. Then 4(Gv) n2
and 0 (G v) n 2. But 0 (G v) 0 (Kn1 ) = n 2. Therefore, 0 (G v) =
n 2. Thus, 0 (G) = n 1.
As G is the disjoint union of an isolated vertex and a subgraph of Kn1 ,
0 (G) = n 2. Hence, 0 (G) + 0 (G) = 2n 3.

ii. 4(G) = n 2 and 4(G) = n 2. Again, G has a pendant vertex v and as before,
(G v) n 2 and (G) = n 2. Similarly, (G) = n 2. Thus, 0 (G) + 0 (G) =
2n 4 < 2n 3.

iii. 4(G) = n 1 and 4(G) = n 3. In this case, G has a vertex v of degree two and
so 0 (G v) = n 2.

Graph Theory

183

Let vu and vw be the edges incident with v in G, with d(u) = n 1. In (n 2)edge colouring of Gv, change the colour i of an edge uu0 (u0 6= w) to a new colour
n 1, and now colour vu by i and vw by n 1. This gives an (n 1)-colouring of
G. Now, by Vizings theorem, 0 (G) n 2.
Together, we get 0 (G) + 0 (G) 2n 3. Since 0(G) + 0 (G) 2n 3, clearly
we have 0 (G) 0 (G) (n 1)(n 2). We observe that in graph K1,n1 , the upper
bounds are attained.
a. Let n be even. Then 0 (G) + 0 (G) n 1. Also, 0 (G) 0 (G) 0.
The lower bounds are attained for complete graphs.
To get the upper bounds, since G and G are subgraphs of Kn and 0 (Kn ) = n 1 for
all even n, 0 (G) + (G) 2(n 1) and 0 (G) 0 (G) (n 1)2 . These upper bounds are
attained in the complete bipartite graphs K1,n1 , n 6= 2.
q

7.5 Region Colouring (Map Colouring)

A region colouring of a planar graph is a labeling of its regions f : R(G) {1, 2, . . .}; the
labels called colours, such that no two adjacent regions get the same colour. A k-region
colouring of a planar graph G consists of k different colours and G is then called k-region
colourable. From the definition, it follows that the k-region colouring of a planar graph
G partitions the region set R into k independent sets R1 , R2 , . . ., Rk , so that R = R1 R2
. . . Rk . The independent sets are called the colour classes, and the function f : R(G)
{1, 2, . . ., k} such that f (r) = i, for each r Ri , 1 i k, is called the colour function.
The minimum number k for which there is a k-region colouring of the planar graph G
is called the region-chromatic number of G, and is denoted by 00 (G). The colouring of
regions is also called map colouring, because of the fact that in an atlas different countries
are coloured such that countries with common boundaries are shown in different colours.
The four colour problem: Any map on a plane surface (or a sphere) can be coloured
with at most four colours so that no two adjacent regions have the same colour.
Now coming to the origin of the four colour problem, there have been reports that Mobius was familiar with the problem in 1840. But the problem was introduced in 1852 by
Francis Guthrie, student of Augustus DeMorgan and the problem first appeared in a letter
(October 23, 1852) from DeMorgan to Sir William Hamilton. DeMorgan continued the discussion of the problem with other mathematicians and in the years that followed attempts
were made to prove or disprove the problem by top mathematical minds of the world. In
1878, Cayley announced the problem to the London Mathematical Society, and in 1879,
Alfred Kempe announced that he had found a proof. An error in Kempes proof was discovered by P. J. Heawood in 1890. Kempes idea was based on the alternating paths and
Heawood used this idea to prove that five colours are sufficient. Kempes argument did not
prove the four colour problem, but did contain several ideas which formed the foundation
for many later attempts at the proof, including the successful attempts by Appel and Haken.
In 1976, K. Appel and W. Haken [5, 6, 7] with the help of J. Koch established what is now
called four colour theorem. Their proof made use of large scale computers (using over

184

Colourings

1000 hours of computer time) and this is the first time in the history of mathematics that
a mathematical proof depended upon the external factor of the availability of a large scale
computing facility. Though the Appel-Haken proof is accepted as valid, mathematicians
are still in search of alternative proof. Robertson, Sanders, Seymour and Thomas [225]
have given a short and clever proof, but their proof still requires a number of computer
calculations. Saaty [230] presents thirteen colourful variations of four colour problem.
In the year 2000, Ashay Dharwadkar [64] has given a new proof of the four colour
theorem, which will be discussed in details in Chapter 14.
The following observations are immediate from the definitions introduced above.
1. A planar graph is k-vertex colourable or k-region colourable if and only if its components have this property.
2. A planar graph is k-vertex colourable or k-region colourable if and only if its blocks
have this property.
These observations imply that for studying vertex colourings or region colourings, it
suffices to consider the graph to be a block.
Theorem 7.20
a. A planar graph G is k-region colourable if and only if its dual G is k-vertex colourable.
b. If G is a plane connected graph without loops, then G has a k-vertex colouring if and
only if its dual G has a k-region colouring.

Proof
a. Let the regions and edges of G be respectively denoted by r1 , . . ., rt and e1 , . . ., em . Let
the vertices of G be r1 , . . ., rt and edges be e1 , . . ., em . Then the vertices and edges of
G are in one-to-one correspondence with the regions and edges of G, and two vertices
r and s in G are joined by an edge e if and only if the corresponding regions r and
s in G have the corresponding edge e as a common edge on their boundary.
Let G be k-region colourable. We colour the vertices in G such that each vertex in
gets the same colour as assigned to the region r in G. Since the vertices r and s
are only adjacent in G if the corresponding regions r and s are adjacent in G, G is
k-vertex colourable.
G

Conversely, let G be k-vertex colourable. Now colour the regions of G such that
the region r in G gets the samecolour as the vertex r in G . This gives a k-region
colouring of G, since the regions r and s are adjacent in G only if the corresponding
vertices r and s are adjacent in G .
b. Since G has no loops its dual G has no bridges, and therefore G is planar. Thus by
(a), G is k-region colourable if and only if the double dual G is k-vertex colourable.
Since G is connected, G is isomorphic to G and hence the result follows.
q

Graph Theory

185

A graph has a dual if and only if it is planar and this implies that colouring
the regions of a planar graph G is equivalent to colouring the vertices of its dual G and
vice versa. It also follows from Theorem 7.17 that if G is the dual of the planar graph G,
then (G) = 00 (G ) and (G ) = 00 (G). These observations give the dual form of the four
colour problem which states that, every planar graph is 4-vertex colourable.
Since loops and multiple edges are not allowed in vertex colourings, it may be assumed
that no two regions have more than one boundary edge in common, for region colouring of
a planar graph
As every triangulation is a planar graph (in fact, a maximal planar graph) and every
planar graph is a subgraph of a triangulation, the four colour problem is true if and only if
every triangulation is 4-colourable.
Remarks

The following result shows that a planar graph is 6-colourable.

Theorem 7.21

Every planar graph is 6-colourable.

Proof Let G be a planar graph and H be the dual of G. Then it is sufficient to prove that
H has a vertex colouring of at most 6 colours. More generally, we prove that any graph H
is 6-colourable.
To prove the result, we use induction on n, the order of H . The result is trivial if H has
at most six vertices. So assume n 7.
Let all planar graphs with fewer than n vertices be 6-colourable. Obviously, H has a
vertex, say v, so that d(v) 5. Therefore v has at most five neighbours in H and these
neighbours evidently need at most five colours for colouring. The vertex deleted subgraph
H v is planar with n 1 vertices and therefore by induction hypothesis is 6-colourable.
Since at most five colours are used for colouring the neighbours of v, therefore assigning v
the sixth colour not used by the its neighbours gives the 6-colouring of H .
q
The following result is a consequence of Theorem 7.20.
Theorem 7.22

A planar graph G is 2-colourable if and only if it is an Euler graph.

Proof Let G be a planar graph which is 2-colourable. Then, if G is the geometric dual
of G, we have (G ) = 2. Therefore G is bipartite and thus G (the dual of G ) is an Euler
graph. Since G and G are isomorphic, therefore G is an Euler graph.
Conversely, let G be an Euler graph. Then its double dual G is an Euler graph and thus

G is bipartite. Therefore, (G ) = 2 and hence the planar graph G is 2-colourable.

q
The next result due to Heawood [113] is called five colour theorem and Heawood used
the Kempe chain argument in proving it.
Theorem 7.23 (Heawood)

Every planar graph is 5-colourable.

Proof Let G be a planar graph with n vertices. We use induction on n, the order of G. The
result is obvious for n 5. So, let n 6. Assume the result to be true for all planar graphs
with fewer than n vertices.

186

Colourings

Let G0 be the graph obtained from G by deleting the vertex v and removing all the edges
incident with v. The graph G0 with order n 1 is clearly planar and by induction hypothesis
is 5-colourable. Let the colours used to colour G0 be c1 , c2 , c3 , c4 , c5 .
We know for a planar graph with n 6 vertices, there exists a vertex, say v, such that
d(v) 5. Thus v has atmost five neighbours in G and all of these neighbours are the already
coloured vertices in G0 .
If in G0 less than five colours are used to colour these neighbours, then the 5-colouring
of G is obtained by using the colouring for G0 on all vertices, and by colouring v with the
colour not used to colour the neighbours of v (Fig. 7.16).
Let all the five colours be used in G0 , to colour the neighbours of v. This implies that
there are exactly five neighbours of v, say u1 , u2 , u3 , u4 and u5 . Assume without loss of
generality that ui is coloured with ci , for each i, 1 i 5.

Fig. 7.16

Consider all the vertices of G0 that are coloured with colour c1 and c3 . If u1 and u3 are in
different components of the Kempe subgraph H(c1 , c3 ) induced by those vertices coloured
c1 and c3 , then using the Kempe chain argument to interchange the colours c1 and c3 in the
component containing u1 leaves c1 unused on the set {u1 , u2 , u3 , u4 , u5 }. We colour v by c1
to get a 5-colouring of G.
Finally, if u1 and u3 are in the same component of H(c1 , c3 ), and u2 and u4 are in the
same component of H(c2 , c4 ), then the c1 c3 Kempe chain from u1 to u3 crosses the c2
c4 Kempe chain from u2 to u4 . This is impossible as the triangulation is a plane graph
(Fig. 7.17).
q

Fig. 7.17

Graph Theory

187

The following result is due to Grunbaum [91].

Theorem 7.24

Every planar graph with fewer than four triangles is 3-colourable.

The next result due to Grotzsch [90] immediately follows from Theorem 7.24.
Theorem 7.25

Every planar graph without triangles is 3-colourable.

While making various attempts to solve the four colour problem, the problem got translated into several equivalent conjectures and sometimes conjectures were made which implied the four colour problem. The details of several such conjectures can be found in Saaty
[229] and Saaty and Kainen [230]. Finch and Sachs [78] proved that every planar graph
with at most 21 triangles is 4-colourable. Ore and Stemple [179] showed that all planar
graphs with upto 39 regions are 4-colourable.
The following result is one such equivalence.
Theorem 7.26 Every planar graph is four colourable if and only if every cubic bridgeless
planar graph is 4-colourable.

Proof We observe that every planar graph is four colourable if and only if every bridgeless planar graph (without cut edges) is four colourable. This is because if G has a bridge
e and G0 is obtained from G by contracting e, then 00 (G0 ) = 00 (G) (as the elementary contraction of identifying the end vertices of a bridge affects neither the number of regions in
the planar graph nor the adjacency of any of the regions).
We now prove that every bridgeless planar graph is 4-colourable if and only if every cubic bridgeless planar graph is 4-colourable. If every bridgeless planar graph is 4-colourable,
then evidently every cubic bridgeless planar graph is 4-colourable.
Conversely, let all cubic bridgeless planar graphs be 4-colourable. Let G be a bridgeless
planar graph. We obtain G0 from G by performing the following operations. In case G has
a vertex v of degree two, let xv and yv be the edges incident with v. Subdivide xv at u and
yv at w. Take two new vertices a and b and add the edges au, aw, bu, bw and ab. Now remove
v (and uv and uw). In doing so, we have replaced v by a K4 e and we see that each new
vertex has degree three (Fig. 7.18).

Fig. 7.18

If G has a vertex v so that d(v) 4, let vx1 , vx2 , . . ., vxd be the edges incident with v.
Subdivide each vxi producing a new vertex vi , for 1 i d . We then remove v and add the

188

Colourings

new edges v1 v2 , v2 v3 , . . ., vd1 vd , vd v1 . Here we have replaced v by Cd and again the degree
of each new vertex is three (Fig. 7.19).
In both cases the graph G0 formed is a bridgeless cubic graph. If these K4 e and Cd
introduced are contracted, we get the original graph G. Hence G is 4-colourable, since we
have assumed that G0 is 4-colourable.
q

Fig. 7.19

The next result is due to Tait [238] and the details can also be found in Bondy and
Chvatal [36] and Bollobas [29].
Theorem 7.27 (Tait) Every planar graph G is 4-colourable if and only if 0 (G) = 3, for
every bridgeless cubic planar graph G.
Proof As seen in Theorem 7.26, the statement that every planar graph is 4-colourable is
equivalent to the statement that every cubic bridgeless planar graph is 4-colourable. Therefore, to prove the result, we prove that a cubic bridgeless planar graph G is 4-colourable if
and only if 0 (G) = 3.
Now assume that G is a bridgeless cubic planar graph which is 4-colourable. For the
set of colours we choose the elements of the Klein four group Q = {c0 , c1 , c2 , c3 }, where
addition in Q is defined by ci + ci = c0 and c1 + c2 = c3 , with c0 being the identity element.

Graph Theory

189

Now, define the colour of an edge to be the sum of the colours of two distinct regions
which are incident with that edge. We see that the edges are coloured with elements of the
set {c1 , c2 , c3 } and that no two adjacent edges get the same colour. Hence, 0 (G) = 3.
Conversely, let G be a bridgeless cubic planar graph with 0 (G) = 3 and colour its edges
with the three non-zero elements of Q. Consider a region, say R0 , and give the colour c0 to
it. Let R be any other region of G and let C be any curve in the plane joining the interior of
R0 with the interior of R, so that C does not pass through a vertex of G. Then the colour of
R is defined to be the sum of the colours of those edges which intersect C .
That the colours of the regions are well defined follows from the fact that the sum of the
colours of the edges which intersect any simple closed curve not passing through a vertex
of G is c0 . Suppose S is such a curve and assume q1 , q2 , . . ., qn to be the colours of the
edges which intersect S. Also assume r1 , r2 , . . ., rm be the colours of those edges interior
to S. If c(v) denotes the sum of the colours of the three edges incident with v, then we see
that c(v) = c0 . Thus, for all vertices v interior to S, we have c(v) = c0 . While on the other
hand, we have c(v) = q1 + q2 + . . . + qn + 2(r1 + r2 + . . . + rm ) = q1 + q2 + . . . + qn as every
element of Q is self-inverse. Thus, q1 + q2 + . . .. + qn = co . Hence we get the 4-colouring of
the regions of G and the colours used are c0 , c1 , c2 , c3 .
q
Remarks Because of Theorem 7.27, a three colouring of the edges of a cubic graph is
called a Tait colouring.
In an attempt to solve the four colour problem, Tait considered edge colourings of
bridgeless cubic planar graphs and proved that every such graph is 3-edge colourable.
In 1880, Tait tried to give a proof of the four colour problem by using Theorem 7.27 and
based on the wrong assumption that any bridgeless cubic planar graph is Hamiltonian. A
counter example called the Tuttle graph to Taits assumption was given by Tutte in 1946
and is shown in Figure 7.20.

Fig. 7.20

190

Colourings

7.6 Heawood Map-Colouring Theorem

Let S p be the orientable surface of genus p, so that S p is topologically equivalent to a sphere
with p handles. The chromatic number of S p , denoted by (S p ), is the maximum chromatic
number among all graphs which can be embedded on S p . The surface So is clearly the
sphere. The four colour problem states that (S0 ) = 4.
Definition: The genus (G) of a graph is the minimum number of handles which must
be added to a sphere so that G can be imbedded on the resulting surface. Clearly, (G) = 0
if and only if G is planar. Further, for a polyhedron, n m + f = 2 2 .
For n 3, the genus of the complete graph Kn is
(Kn ) =

(n 3)(n 4)
.
12

The following inequality is due to Heawood [113].

Theorem 7.28 The chromatic number of the orientable surface of positive genus p has
the upper bound

7 + 1 + 48p
(S p )
, p > 0.
2

Proof Let G be an (n, m) graph embedded on S p . Since any graph can be augmented to
a triangulation of the same genus by adding edges without reducing , we assume G to
be a triangulation. Let d 0 be the average degree of the vertices of G. Then n, m and f (the
number of regions) are related by the equations
d 0 n = 2m = 3 f .

Solving for m and f in terms of n and using Eulers equation n m + f = 2 2 , we get

d0 =

12(p 1)
+ 6.
n

As d 0 n 1, this gives n 1

(7.28.1)
12(p 1)
+ 6.
n

Solving for n and taking the positive square root, we obtain n

7 + 1 + 48p
.
2

7 + 1 + 48p
Let H(p) =
. Then we show that H(p) colours are sufficient to colour the
2
vertices of G. If n = H(p), obviously we have sufficient colours. In case n > H(p), we
substitute H(p) for n in (7.28.1) and obtain

Graph Theory
d0 <

191

12(p 1)
+ 6 = H(p) 1.
H(p)

Therefore, when n > H(p), there is a vertex v of degree at most H(p) 2. Identify v
and any adjacent vertex by an elementary contraction to obtain a new graph G0 . If n0 =
n 1 = H(p), then G0 can be coloured in H(p) colours. If n0 > H(p), repeat the argument
and evidently we get an H(p) colourable graph. It is then easy to see that the colouring of
this graph induces a colouring of the preceding one in H(p) colours, and so forth, so that G
itself is H(p)colourable. Hence the result follows.
q
The next result is called Heawood map colouring theorem, the proof of which is due to
Ringel and Youngs [223].
Theorem 7.29 (Heawood map-colouring theorem) For every positive integer p, the
chromatic number of the orientable surface of genus p is given by
(S p )=

Proof

7+

1 + 48p
, p > 0.
2

Let G be an (n, m) graph embedded on S p . Then,

7 + 1 + 48p
(S p )
, p > 0.
2

Now, if the complete graph Kn is embedded in S p , then

p (Kn ) =

(n 3)(n 4)
.
12

(7.29.1)

Setting n to be the largest integer satisfying (7.29.1), we have



(n 2)(n 3)
(n 2)(n 3)
(n 3)(n 4)
p
1 <
.
12
12
12

Solving for n, we get

5+

7+
1 + 48p
<n
2

7 + 1 + 48p
Thus, n =
.
2

1 + 48p
2

Since (Kn ) = n, we have found a graph with genus p and chromatic number equal to
H(p). This shows that H(p) is a lower bound for (Sp), completing the proof.
q

192

Colourings

7.7 Uniquely Colourable Graph

A graph G(V, E) is said to be uniquely k-vertex-colourable (or uniquely k-colourable) if
there is a unique k-part partition of the vertex set V into independent subsets. That is, in
the uniquely k-colouring of a graph G, every k-colouring of G induces the same partition
of V . The graph of Figure 7.21(a) is uniquely 3-colourable, since every 3-colouring of G
has the partition {v1 }, {v2 , v4 }, {v3 , v5 }. The graph of the Figure 7.21(b) is not uniquely
3-colourable. A pentagon is not uniquely 3-colourable, as five different partitions of its
vertex set are possible.

Fig. 7.21

We observe that the empty graphs Kn are the only uniquely 1-colourable graphs and the
connected bipartite graphs are the only uniquely 2-colourable graphs. We note that unique
k-colourability is not defined for k > n. Further, we see that Kn is uniquely n-colourable.
These observations imply to assume 3 k n, for studying uniquely k-colourable graphs.
We have the following observation.
Theorem 7.30

If G is uniquely k-colourable, then (G) k 1.

Proof Let G be a uniquely k-colourable graph. Then every vertex v of G is adjacent to at

least one vertex of every colour different from that assigned to v. For otherwise, a different
k-colouring of G is obtained by recolouring v. This implies that d(v) k 1, for every v. q
Corollary 7.2
2m n(k 1).

If G is a uniquely k-colourable graph with n vertices and m edges, then

The next result, due to Cartwright and Harary [45], gives a necessary condition for a
graph to be uniquely colourable.
Theorem 7.31 The subgraph induced by the union of any two colour classes in a kcolouring of a uniquely k-colourable graph is connected.

Graph Theory

193

Proof Let G be a uniquely k-colourable graph. Let C1 and C2 be two classes in the kcolouring of the graph G. Assume the subgraph S of G induced by C1 C2 to be disconnected, and let S1 and S2 be components of S. Then each of S1 and S2 contain vertices of
both C1 and C2 . Now interchanging the colour of the vertices in C1 S1 by the colour of the
vertices in C2 S1 gives a different k-colouring of G. This contradicts the hypothesis that G
is uniquely k-colourable. Hence S is connected.
q
The converse of Theorem 7.31 is not true. To see this, consider the 3-chromatic graph G
of Fig. 7.21(b). The graph G has the property that in any 3-colouring, the subgraph induced
by the union of any 2 colour classes is connected. But G is not uniquely colourable. It
follows from Theorem 7.30 that every uniquely k-colourable graph, k 2, is connected.
The following stronger result is due to Chartrand and Geller [49].
Theorem 7.32

Any uniquely k-colourable graph is (k 1)-connected.

Proof Let G be a uniquely k-colourable graph. In case G is Kk , then it is (k1)-connected.

Now assume G to be an incomplete graph which is not (k 1)connected. Therefore there
exists a set U of k 2 vertices whose removal disconnects G. Then in any k-colouring of
G, there are at least two distinct colours, say c1 and c2 , not assigned to any vertex of U .
By Theorem 7.30, a vertex coloured c1 is connected to any vertex coloured c2 by a path
all of whose vertices are coloured c1 or c2 . Therefore the set of vertices of G coloured
c1 or c2 lies within the same component of G U , say G1 . Another k-colouring of G can
thus be obtained by taking any vertex of G U which is not in G1 and recolouring it with
either c1 or c2 . This contradicts the hypothesis that G is uniquely k-colourable. Hence G is
(k 1)-connected.
q
Corollary 7.3 In any k-colouring of uniquely k-colouring graph, the subgraph induced
by the union of any h colour classes, 2 h k, is (h 1)-connected.
The following result due to Bollobas [30] gives a sufficient lower bound for uniquely
colourable graphs.
Theorem 7.33 If G is a k-colourable graph of order n(k 2) with (G) > n(3k 5)/(3k
2), then G is uniquely k-colourable.
The following result is due to Harary, Hedetniemi and Robinson [107].
Theorem 7.34 For all k 3, there is a uniquely k-colourable graph which contains no
subgraph isomorphic to Kk .
Clearly, a graph is uniquely 1-colourable if and only if it is 1-colourable, that is, totally
disconnected. Also, a graph is uniquely 2-colourable if and only if G is 2-chromatic and
connected.
The converse of Theorem 7.36 is not true. This is because a uniquely 3-colourable planar
graph may have more than one region which is not a triangle, as shown in Figure 7.22.

194

Colourings

Fig. 7.22 Uniquely 3-colourable planar graph

Theorem 7.35 If G is a uniquely 3-colourable planar graph with at least four vertices,
then G contains at least two triangles.
Theorem 7.36

Every uniquely 4-colourable planar graph is maximal planar.

Proof Let a 4-colouring be given to uniquely 4-colourable planar graph G with the colour
classes denoted by Vi , 1 i 4, where |Vi | = ni . Since the subgraph induced by Vi V j , i 6= j ,
is connected, G has at least (ni + n j 1) edges, 1 i < j < 4. Clearly, (ni + n j 1) =
n1 + n2 1 + n1 + n3 1 + n1 + n4 1 + n2 + n3 1 + n2 + n4 1 + n3 + n4 1 = 3(n1 + n2 + n3 +
n4 ) 6 = 3n 6. Therefore, m 3n 6. Hence G is maximal planar.
q

The next result for uniquely 5-colourable graphs is due to Hedetniemi [114].
Theorem 7.37

No planar graph is uniquely 5-colourable.

Theorem 7.38 A necessary and sufficient condition that a connected planar graph is 4colourable is that G be the sum of three subgraphs G1 , G2 and G3 such that for each vertex
v, the number of edges of each Gi incident with v are all even or odd.
The following equivalence is due to Whitney [264].
Theorem 7.39 The four colour problem holds if and only if every Hamiltonian planar
graph is 4-colourable.

7.8 Hajos Conjecture

Hajos [97] made the following conjecture.
If a graph is k-chromatic, then it contains a subdivision of Kk .
When k = 1, or 2, the conjecture trivially holds. As for k = 3, every chromatic graph contains
an odd cycle which is a subdivision of K3 , therefore proving the validity of the conjecture.
The validity of the conjecture for k = 4, as noted in Parthasarthy [180] is due to Dirac [66].
Theorem 7.40

Hajos conjecture is true for k = 4.

Graph Theory

195

Proof Assume without loss of generality that G is a 4-minimal graph. So G is a block

and = 3. In case n = 4, G is K4 and the result is obvious. Therefore assume n 5. We
induct on n.
Let G have a 2-vertex cut S = {u, v}. Then by Theorem 7.5, G = G1 G2 , where G1 is
of type 1, and G2 is of type 2 and G1 + uv is 4-minimal. By induction hypothesis, G1 + uv
contains a subdivision of K4 . Here we replace uv by a u v path P in G2 and so G1 P
contains a subdivision of K4 . Thus G also contains a subdivision of K4 . Now, let G be 3connected. Since (G) 3, therefore it has a cycle C of length at least 4. If u and v are any
two non-consecutive vertices on C, then G {u, v} is still connected and therefore there
exist vertices x, y on C, and an x y path P in G {u, v}. Similarly, there exists a u v path
Q in G {x, y}. If P and Q have no vertices in common, then C P Q is a subdivision of
K4 in G. Otherwise, let w be the first vertex of P on Q, then C Pxw Q is a subdivision of
K4 in G.
Hence in all cases G contains a subdivision of K4 .
q
For k 5, Hajos conjecture implies four colour problem. This is because if G is a planar
graph which is not colourable by 4 colours, its chromatic number is at least 5 and thus
contains a subdivision of K5 , and so cannot be planar, a contradiction. The four colour
theorem implies that a 5-chromatic graph contains a homeomorph of K5 or K3, 3 . For k = 5,
Hajos conjecture makes the stronger assertion that it contains a homeomorph of K5 . For
k = 5, or 6, the conjecture has not been settled, but for k = 7, it is disproved by Catlin. The
counter example of Catlin is the graph H = L(3C5 ) {v1 , v2 }, where 3C5 is the multigraph
obtained from C5 by replacing each edge by three edges, L represents the edge graph, and
v1 and v2 are any two non-adjacent vertices of L(3C5 ) (Fig. 7.23).

Fig. 7.23

196

Colourings

The largest integer for which a given graph G contains a T Kn is called the subdivision
number (or topological clique number or Hajos number) of G and is denoted by tw(G).
With this, Hajos conjecture is equivalent to tw(G) (G). In the above example, we see
that H contains a K6 and from any vertex outside this K6 there are no six internally disjoint
paths to the vertices of the K6 . Therefore,
tw(H) = 6. A maximum independent set of H has
 
cardinality two, so that (H)

13
= 7 and a 7-colouring of H is shown in the Fig. 7.23.
2

Thus, (H) = 7 and hence H is a counter example for Hajos conjecture. We now observe
that if G is a counter example for Hajos conjecture for k, then G + v is a counter example
for Hajos conjecture for k + 1. This can be seen from the fact that tw(G + v) = tw(G) + 1 and
(G + v) = (G) + 1. Hence Hajos conjecture is false for all k = 7. Erdos and Fajtlowicz
[72] proved that almost every graph is a counter example
to Hajos conjecture. Bollobas

and Catlin [31] proved that tw(G) is approximately 2 n for n-vertex graphs.
The following conjecture involving contractions is due to Hadwiger [95].
Hadwigers conjecture [95]

Every k-chromatic graph contains Kk as a subcontraction.

Hadwigers conjecture is trivially true for k = 1. Since 2-chromatic graphs are the bipartite graphs and 3-chromatic graphs contain an odd cycle, contractible to K3 , the conjecture
is true for k = 2 and 3. Dirac [66] proved the conjecture for k = 4. For k = 5, this conjecture
states that every 5-chromatic graph is contractible to K5 and therefore every such graph
is non-planar. Thus Hadwigers conjecture for k = 5 implies the four colour problem. The
converse of this is given by Wagner.

7.9 Exercises
1. Prove that (G) = 4(G) + 1 if and only if G is either a complete graph or a cycle of
odd length.
2. Show that if G contains exactly one odd cycle, then (G) = 3.
3. If G is a graph in which any pair of odd cycles have a common vertex, then prove that
(G) 5.
4. Find the chromatic number of the Peterson graph and the Birkhoff Diamond.
5. Determine the chromatic number of the graphs in Figure 7.24.

Graph Theory

197

Fig. 7.24

6. If G is k-regular, prove that (G)

n
.
nk

7. If G is connected and m n, show that (G) 3.

8. Prove that the 3-critical graphs are the odd cycles C2n+1 .
9. Prove that the wheel W2n1 is a 4-critical graph for each n 2.
10. Prove that the wheel W2n is a 4-critical graph for each n 2.
11. Prove that the Peterson graph has edge chromatic number 4.
12. If m(G) is the number of edges in a longest path of G, prove that (G) 1 + m(G).
13. Show that if G is 3-regular Hamiltonian graph, then 0 (G) = 3.
14. Show that a triangulation with a vertex of degree 2 or 3 can be coloured with five
colours.
15. Prove that for every k 1 there is a k-chromatic graph Mk with no triangle subgraphs.
(Mycielski, 1955).
16. If G is a graph in which no set of four vertices induces P4 as a subgraph, then prove
that (G) = cl(G) (Seinsche, 1974).
17. Obtain proofs of Theorems 2.24 and 2.25.
18. Show that a uniquely 3-colourable graph contains three Hamiltonian cycles.

8. Matchings and Factors

Consider the formation of an executive council by the parliament committee. Each committee needs to designate one of its members as an official representative to sit on the council,
and council policy states that no senator can be the official representative for more than
one committee. For example, let there be five committees 1, 2, 3, 4 and 5. Let A, C, D, E
be members of committee 1; B, C, D be members of committee 2; A, B, E, F be members
of committee 3; and A, F be members of committee 5. Here the executive council can be
formed by members A, C, B, F and E , representing committees 1, 2, 3, 4 and 5 respectively. The problem of formation of the executive council and many related problems can
be solved using the concept of matchings in graphs.

8.1 Matchings
Definition: A matching in a graph is a set of independent edges. That is, a subset M
of the edge set E of a graph G(V, E) is a matching if no two edges of M have a common
vertex. A matching M is said to be maximal if there is no matching M 0 strictly containing
M , that is, M is maximal if it cannot be enlarged. A matching M is said to be maximum if it
has the largest possible cardinality. That is, M is maximum if there is no matching M 0 such
that |M 0 | > |M|.
Consider the graph G shown in Figure 8.1. The examples of matchings in G are M1 =
{v1 v5 , v2 v6 , v3 v4 , v7 v8 }, M2 = {v1 v5 , v2 v6 , v3 v4 , v7 v9 }, M3 = {v1 v5 , v2 v6 , v3 v4 , v7 v10 }, M4 =
{v1 v5 , v2 v6 , v3 v4 } and M5 = {v4 v7 , v1 v6 , v2 v3 }. Clearly, M1 is a maximum matching, whereas
M5 is maximal but not maximum.

Fig. 8.1

Graph Theory

199

The cardinality of a maximum matching is denoted by 1 (G) and is called the matching
number of G (or the edge-independence number of G).
Definition: Let M be a matching in a graph G. A vertex v in G is said to be M-saturated
(or saturated by M ) if there is an edge e M incident with v. A vertex which is not incident
with any edge of M is said to be M-unsaturated. In other words, given a matching M in
a graph G, the vertices belonging to the edges of M are M -saturated and the vertices not
belonging to the edges of M are M -unsaturated. Consider the graph shown in Figure 8.2.
Clearly, M = {v1 v2 , v3 v7 , v4 v5 } is a matching and the vertices v1 , v2 , v3 , v4 , v5 , v7 are M saturated but v6 is M -unsaturated.

Fig. 8.2

Definition: A matching M in a graph G is said to be a perfect matching if M saturates

every vertex of G. In Figure 8.3(a), G1 has a perfect matching M = {v1 v2 , v3 v6 , v4 v5 }, but
G2 has no perfect matching.

Fig. 8.3(a)

Definition: If G is a bipartite graph with bipartition V = V1 V2 and |V1 | = n1 , |V2 | = n2 ,

then a matching M of G saturating all the vertices in V1 is called a complete matching
(or V1 matched into V2 ).
Definition: Let H be a subgraph of a graph G. An H -alternating path (cycle) in G is a
path (cycle) whose edges are alternately in E(G) E(H) and E(H).
If M is matching in a graph G, then an M -alternating path (cycle) in G is a path (cycle)
whose edges are alternately in E(G) M and M . That is, in an M-alternating path, the edges

200

Matchings and Factors

alternate between M -edges and non-M -edges. An M -alternating path whose end vertices
are M -unsaturated is said to be an M -augmenting path.

Fig. 8.3(b)

Consider the graph G shown in Figure 8.3(b). An example of a matching in G is M =

{v2 v10 , v3 v9 , v4 v8 , v5 v7 }. Clearly v10 , v2 , v3 , v9 , v8 , v4 is an M -alternating path and v1 , v10 , v2 , v9 , v3 , v11
is an M -augmenting path.
The following result due to Berge [21] characterises a maximum matching in a graph.
This forms the foundation of an efficient algorithm for obtaining a maximum matching.
Theorem 8.1 (Berge) A matching M of a graph G is maximum if and only if G contains
no M -augmenting paths.

Proof Let M be a maximum matching in a graph G. Assume P = v1 v2 . . .vk is an M augmenting path in G. Due to the alternating nature of M -augmenting path, we observe that
k is even and the edges v2 v3 , v4 v5 , . . ., vk2 vk1 belong to M . Also the edges v1 v2 , v3 v4 , . . .,
vk1vk do not belong to M (Figure 8.4).

Fig. 8.4

Now, let M1 be the set of edges given by

M1 = [M {v2 v3 , . . ., vk2 vk1 }] {v1v2 , . . ., vk1vk }.

Then M1 is a matching and clearly M1 contains one more edge than M . This contradicts
the assumption that M is maximum. Thus G contains no M-augmenting paths.
Conversely, assume that G has no M -augmenting paths. Let M 0 be a matching such that
|M 0| > |M|. Let H be a subgraph of G with V (H) = V (G) and E(H) be the set of edges of G
that appear in exactly one of M and M 0 . Since every vertex of G lies on at most one edge
from M and at most one edge from M 0 , therefore degree (in H ) of each vertex of H is at most
2. This implies that each connected component of H is either a single vertex, or a path, or
a cycle (Fig. 8.5). If a component is a cycle, then it is an even cycle, because the edges
alternate between M -edges and M 0-edges. Since |M 0| > |M|, there is at least one component
in H that is a path which begins and ends with edges from M 0. Clearly, this path is an M -

Graph Theory

201

augmenting path, contradicting the assumption. Thus no such matching M 0 can exist and
hence M is maximum.
q

Fig. 8.5

Definition: The neighbourhood of a set of vertices S, denoted by N(S), is the union of

the neighbourhood of the vertices of S.
The following classic result due to P. Hall [101] characterises a complete matching in a
bipartite graph.
Theorem 8.2 (Hall) If G(V1 , V2 , E) is a bipartite graph with |V1 | |V2 |, then G has a
matching saturating every vertex of V1 , if and only if |N(S)| |S|, for every subset S V1 .

Proof Let G have a complete matching M that saturates all the vertices of V1 and let S be
any subset of V1 . Then every vertex in S is matched by M into a different vertex in N(S), so
that |S| |N(S)| (Fig. 8.6 (a)).

Fig. 8.6

202

Matchings and Factors

Conversely, let |N(S)| |S| for all subsets S of V1 and let M be a maximum matching.
Assume G has no complete matching. Then there exists a vertex, say v V1 , which is
M -unsaturated (Fig. 8.6(b)). Let Z be the set of vertices of G that can be joined to v by
M -alternating paths.

Fig. 8.7

Since M is a maximum matching, by Theorem 8.1, G has no M -augmenting path. This

implies that v is the only vertex of Z that is M -unsaturated. Let S = Z V1 and T = Z V2 (Fig.
8.7). Then every vertex of T is matched under M to some vertex of S {v} and conversely
and N(S) = T . Since |T| = |S| 1, therefore |N(S)| = |T | = |S| 1 < |S|, contradicting the
given assumption. Hence G has a complete matching.
q
The aim of the above theorem was basically to obtain the conditions for the existence of
a system of distinct representatives (SDR) for a collection of subsets of a given set. Before
giving such conditions, we have the following definition.
Definition: If A = {Ai : i N} is a family of sets, then a system of distinct representatives
(SDR) of A is a set of elements {ai : i N} such that ai Ai for every i N and ai 6= a j whenever i 6= j . For example, if A1 = {2, 8}, A2 = {8}, A3 = {5, 7}, A4 = {2, 4, 8} and A5 = {2, 4},
then the family {A1 , A2 , A3 , A4 } has an SDR {2, 8, 7, 4}, but the family {A1 , A2 , A4 , A5 } has
no SDR.
Theorem 8.3 (Halls SDR Theorem) A family of finite nonempty sets A = {Ai : 1 i
r} has an SDR if and only if for every k, 1 k r, the union of any k of these sets contains
at least k elements.

Proof

Clearly Ai is finite, since each of the sets A1 , A2 , . . ., Ar is finite. Let Ai =

i=1
i=1
{a1, a2 , . . ., an }.

Graph Theory

203

Construct a bipartite graph G with partite sets X = {A1 , . . ., Ar } and Y = {a1 , . . ., an }, as

shown in Figure 8.8. We take an edge between Ai and a j if and only if a j Ai .

Fig. 8.8

Clearly, A has an SDR if and only if G has a matching that saturates all the vertices of
X . Now, Theorem 8.2 implies that G has such a matching if and only if |S| |N(S)| for all
subsets S of X , that is, if and only if |S| | Ai | proving the result.
q
ai S

Remarks

1. Halls theorem is often referred to as Halls marriage theorem.

2. Since Halls theorem, there has been remarkable progress in the theory of SDR, and
besides other references the reader can refer to the book of Mirsky [161].
The following result due to M. Hall [100] counts the number of complete matchings.
Theorem 8.4 If G(V1 , V2 , E) is a bipartite graph with |V1 | = n1 n2 = |V2 | and satisfying
Halls condition |S| |N(S)| for all S V1 , then G has at least
r( , n1 ) =

( + 1 i)

1imin( , n1 )

complete matchings, where = min{d(u) : u V1 }.

Proof Induct on n1 . The result is trivial for n1 = 1. Assume the result to be true for all
values of n1 m 1 and let G be a bipartite graph satisfying the given conditions and
|V1 | = m.
For each nonempty proper subset S of V1 , |S| < |N(S)|. Now, for a given u V1 ,
choose N(u). This v can be chosen in at least ways, since = min{d(u) : u V1 }. Obviously, G {u, v} is a bipartite graph with |V1 | = m 1, and min{d(x) : x V1 } 1. Therefore by induction hypothesis, G {u, v} has at least r( 1, m 1) =
( i)

Case 1

1imin( 1,m1)

complete matchings. These together with uv, give at least

[r( 1, m 1)] =

1imin( 1, m1)

complete matchings in G.

( i) =

1imin( , m)

( + 1 i) = r( , m)

204

Matchings and Factors

Case 2 There is a non empty proper subset S of V1 with |S| = |N(S)|. Let |S| = s, and
G1 =< S N(S) > and G2 = G < S N(S) >. Then the values of n1 and in G1 are s and ,
and in G2 are m s and 2 = max{ s, 1}. Therefore G1 and G2 are disjoint graphs with at
least r( , s) and r(2 , m s) complete matchings, by induction hypothesis. Thus G has at
least r( , s) r(2 , m s) = r( , m) complete matchings.
q

The following result is due to Ore [175].

Theorem 8.5 If a bipartite graph G(V1 , V2 , E) satisfies |S| d + |N(S)| for all S V1 ,
where d is a given positive integer, then G contains a matching M with at most d vertices
of V1 being M -unsaturated.
Let H(V1 , U2 , E 0 ) be the bipartite graph with U2 = V2 W2, |W2 | = d and E 0 = E
[V1 , W2 ]. Then the given condition of G implies Halls condition for H and so H has a
complete matching M . Obviously, at most d of the edges of M can be in E 0 E , proving the
result.
q

Proof

The following observation is an immediate consequence of Theorem 8.5.

Corollary 8.1 For a bipartite graph G(V1 , V2 , E), with |V1 | = n1 , 1 (G) = n1 max{|S|
|N(S)| : S V1 } = min{|V1 S| + |N(S)| : S V1 }.
The following result is a consequence of Halls theorem.
Theorem 8.6

A k( 1) regular bipartite graph has a perfect matching.

Proof Let G be a k regular bipartite graph with partite sets V1 and V2 . Then E(G) is
equal to the set of edges incident to the vertices of V1 and also E(G) is equal to the set of
edges incident to the vertices of V2 . Therefore, k|V1 | = k|V2 | = E(G) and thus |V1 | = | V2 |. If
S V1 , then N(S) V2 and thus N(N(S)) contains S. Let E1 and E2 be the sets of edges of
G incident respectively to S and N(S). Then E1 E2 , |E1 | = k|S| and |E2 | k|N(S)|. Since
|E2 | |E1 |,|N(S)| |S|. Thus by Theorem 8.2, G has a matching that saturates all the vertices
of V1 , that is G has a complete matching M . Now deleting the edges of M from G gives
a (k 1)-regular bipartite graph G0 , with V (G) = V (G0 ). By the same argument, G0 has a
complete matching M 0 . Deletion of the edges of M 0 from G0 results in a (k 2)-regular
bipartite graph. Repeating this process k 1 times, we arrive at a 1-regular bipartite graph
G such that |V (G )| = |V (G)|. Clearly G has a perfect matching.
q
Definition: A set C of vertices in a graph G is said to cover the edges of G if every
edge of G is incident with at least one vertex of C. Such a set C is called a covering of G.
Consider the graphs G1 and G2 shown in Figure 8.9. Clearly, {v1 , v2 , v4 , v5 } and {v1 , v5 , v6 }
are coverings of G1 . Also, we observe that there is no covering of G1 with fewer than three
vertices. In G2 , each of {v1 , v2 , v6 } and {v1 , v2 , v3 , v4 , v5 , v6 } is a covering.

Graph Theory

205

Fig. 8.9

The number of vertices in a minimum covering of G is called the covering number of G

and is denoted by (G).
Definition: An edge covering of a graph G is a subset L of E such that every vertex of
G is incident to some edge of L. Clearly, an edge covering of G exists if and only if > 0.
The cardinality of a minimum edge covering of G is denoted by 1 (G). For example, in the
wheel W5 of Figure 8.10, the set {v1 v5 , v2 v3 , v4 v6 } is a minimum edge covering.

Fig. 8.10

The following results are immediate.

Lemma 8.1

A subset S of V is independent if and only if V S is a covering of G.

Proof Clearly, S is independent if and only if no two vertices in S are adjacent in G.

Thus every edge of G is incident to a vertex of V S. This is possible if and only if V S is
a covering of G.
q
Lemma 8.2

For any graph G, + = n.

Proof If S is a maximum independent set of G, then by Lemma 8.1, V S is a covering

of G. Therefore, |V S| = n . Similarly, if C is a minimum covering of G, then V C

206

Matchings and Factors

is independent and therefore, |V S| = n . Combining the two inequalities, we obtain

+ = n.
q

Remark Consider the graph of Figure. 8.11. Clearly the set E1 = {e3 , e4 } is independent
and E E1 = {e1 , e2 , e5 } is not an edge covering of G. Also, E2 = {e1 , e3 , e4 } is an edge
covering of G and E E2 is not independent in G. These observations imply that the edge
analogue of Lemma 8.1 is not true in general.

Fig. 8.11

The following result gives the relation between 1 and 1 of a graph G.

Theorem 8.7(a)

In a graph G with n vertices and > 0, 1 + 1 = n.

Proof Let M be a maximum matching in G, so that |M| = 1 . Let X be the set of M unsaturated vertices in G. Since M is maximum, X is an independent set of vertices with
|X| = n 21 . As > 0, choose one edge for each vertex in X incident with it and let F
be the set of these edges chosen. Then M F is an edge covering of G. Thus, |M F| =
|M| + |F| = 1 + n 2 1 1 , and so
n 1 + 1 .

(8.7.1)

Now assume that L is a minimum edge covering of G, so that |L| = 1 . Let H = G(L), the
edge subgraph of G defined by L and let MH be a maximum matching in H . Let the set of
MH -unsaturated vertices in H be denoted by X . Since L is an edge covering of G, therefore,
H is a spanning subgraph of G. Thus, |L| |MH| = |L MH | |X| = n 2|MH |, and therefore
|L| + |MH | n. Since MH is a matching in G, so |MH | 1 , and hence
n 1 + 1 .

Combining (8.7.1) and (8.7.2), we obtain 1 + 1 = n.

(8.7.2)
q

Let M be any matching of a graph G and C be any vertex covering of G. Then in order
to cover each edge of M , at least one vertex of C is to be chosen. Thus, |M| |C|. In case,
M 0 is a maximum matching and C 0 is a minimum covering of G, then |M 0| |C 0 |.
We have the following obvious result.

Graph Theory

207

Lemma 8.3 If C is any covering and M any matching of a graph G with |C| = |M|, then C
is a minimum covering and M is a maximum matching.

Proof Let M 0 be a maximum matching and C0 be a minimum covering of G. Then |M|

|M 0| and |C| |C 0 |. Therefore, |M| |M 0 | |C 0 | |C|. As |M| = |C| we have |M| = |M 0| =
|C 0| = |C|, and the proof is complete.
q
The next result is due to Konig [135].
Theorem 8.7(b) (Konig) In a loopless bipartite graph G, the maximum number of
edges in a matching of G is equal to the minimum number of vertices in an edge cover
of G, that is, 1 = .

Proof Let G be a bipartite graph with partite sets V1 and V2 and let M be a matching of
G. Let X be the set of all M -unsaturated vertices of V1 , so that |M| = |V1 | |X| (Fig. 8.12).

Fig. 8.12

Let A be the set of those vertices of G which are connected to some vertex of X by an
M -alternating path. Further, let S = A V1 and T = A V2 . Obviously, N(S) = T and S X is
matched to T so that |X| = |S| |T |. Clearly, C = (V1 S) T is a covering of G, because if
there is an edge e not incident to any vertex in C, then one of the end vertices of e is in S
and the other in V2 T , which contradicts the fact that N(S) = T . Now, |C| = |V1 | |S| + |T| =
|V1 | |X| = |M|. Then by Lemma 8.3, M is a maximum matching and C is a minimum
covering of G.
q
The following is a new proof of Konigs theorem due to Rizzi [224].
Second proof of Theorem 8.7 (Rizzi [224]) Let G be a minimal counter-example.
Then G is connected but is not a cycle, nor is a path. Therefore G has a vertex of degree
at least three. Let u be such a vertex and v one of its neighbours. If 1 (G v) < 1 (G),
then by minimality, G v has a cover W 0 with |W 0 | < 1 (G). Thus W 0 {v} is a cover of G
with cardinality at most 1 (G). Therefore assume there exists a maximum matching M of
G having no edge incident at v. Let f be an edge of G M incident at u but not at v. Let W 0

208

Matchings and Factors

be a cover of G f with |W 0 | = 1 (G). Since no edge of M is incident at v, it follows that W 0

does not contain v. Therefore W 0 contains u and is a cover of G.
q

8.2 Factors
We start this section with the following definitions.
Definition: A factor F of a graph G is a spanning subgraph of G. A factor F is said
to be an Ffactor if the components of F are some graphs in a given collection F =
{H1 , H2 , . . ., Hk } of subgraphs of G. If F contains a single graph H , then F is called an
H-factor. Consider the graph G shown in Figure 8.13. Clearly G1 is a spanning subgraph
whose components belong to the set F = {H1 , H2 , H3 } and therefore is an F -factor. Each of
the components of the spanning subgraph G2 is a graph K3 and so G2 is a K3 -factor.

Fig. 8.13

We observe that a perfect matching is a K2 -factor, since each of the components in a

perfect matching is a graph K2 .
A vertex function of a graph G(V, E) is a function f : V No (No being the set of nonnegative integers).
Definition: A spanning subgraph F of a graph G(V, E) is called an f-factor of G if for a
vertex function f on V, d(v|F) = f (v) for all v V . For example, consider the graph G shown
in Figure 8.14. Define f : V N0 by f (vi ) = 1 or 2, for all v V in G. Then the f -factor of
G is given by graph H .

Graph Theory

209

Figure 8.14

Further, F is said to be a partial ffactor (or f -matching) if d(v|F) f (v), for all v V
and F is said to be a covering f -factor of G if d(v|F) f (v), for all v V . Some of the
details of this can be seen in Graver and Jurkat [88]. We note that an f -factor exists only if
f (v) is even.
vV

Definition: A k-factor of a graph G is a factor of G that is k-regular. Clearly, a 1-factor is

a perfect matching and exists only for graphs with an even number of vertices. A 2-factor
of G is a factor of G that is a disjoint union of cycles of G and connected 2-factor is a
Hamiltonian cycle.
Definition: If the edge set of a graph G is partitioned by the edge sets of a set of factors
F1 , F2 , . . ., Fq of G, then these factors are said to constitute a factorisation or decomposition
of G and we write G = F1 F2 . . . Fq . If each factor Fi in the factorisation is isomorphic to the same spanning subgraph F of G, then the factorisation is called an isomorphic
factorisation or an F -decomposition of G.
The factorisation is called an f-factorisation or k-factorisation according as each factor
f i of a factorisation is an f -factor or k-factor. A graph G is f -factorable or k-factorable if G
has respectively f -factorisation or k-factorisation.

Example In Figure 8.15, the three distinct 1-factors of G1 are < e1 , e6 >, < e2 , e4 > and
< e3 , e5 >. The two distinct 2-factors of G2 are < e1 , e2 , e3 , e4 , e5 > and < e6 , e7 , e8 , e9 , e10 >.

Fig. 8.15

210

Matchings and Factors

Definition: A component of a graph is odd or even according to whether it has an odd or

even number of vertices. The number of odd components of a graph G is denoted by k0 (G).
The following result of Tutte [249] characterises the graphs with a 1-factor. Several proofs
of this result exist in the literature and the proof given here is due to Lovasz [150].
Theorem 8.8 (Tutte)

A graph G has a 1-factor if and only if

k0 (G S) |S|, for all S V .

(8.8.1)

Proof Let G be a graph having 1-factor M . Let S be an arbitrary subset of V and let O1 ,
O2 , . . ., Ok be the odd components of G S. For each i, the vertices in Oi can be adjacent
only to other vertices in Oi and to vertices in S. Since G has a 1-factor (perfect matching), at
least one vertex from each Oi is to be matched with a different vertex in S. Thus the number
of vertices in S is at least k, the number of odd components. Hence |S| k, so that k0 (G S)
|S| (Fig. 8.16).

Fig. 8.16

Conversely, assume that the condition (8.8.1) holds. Let G have no 1-factor. Join pairs of
non-adjacent vertices of G and continue this process till we get a maximal super graph G
of G having no 1-factor. Since joining two odd components by an edge results in an even
component, therefore
k0 (G S) k0 (G S).

(8.8.2)

Thus for G also, we have k0 (G S) |S|.

When S = in (8.8.1), we observe that k0 (G) = 0 so that k0 (G ) = 0. Therefore, |V (G )| =
|V (G)| = n is even. Also, for every pair of non-adjacent vertices u and v of G , G + uv has
a 1-factor and any such 1-factor definitely contains the edge uv.
Now, let K be the set of those vertices of G which are having degree n 1. Clearly,
K 6= V , since otherwise G = Kn has a perfect matching.
Claim that each component of G K is complete. Assume, on the contrary, that in

G K , there exists a component G1 which is not complete. Then in G1 , there are vertices

Graph Theory

211

x, y and z such that xy, yz E(G ) and xy

/ E(G ). Also, y V (G1 ) so that d(y|G ) < n 1

and thus there exists a vertex w of G with yw / E(G ). Evidently, w / K (Fig. 8.17).

Fig. 8.17

Now, by the choice of G , both G + xz and G + yw have 1-factors, say M1 and M2

respectively. Clearly, xz M1 and yw M2 . Let H be the subgraph of G + {xz, yw} induced
by the edges that are in M1 and M2 but not in both (that is, in the symmetric difference of
M1 and M2 ). Since M1 and M2 are 1-factors, each vertex of G is saturated by both M1 and
M2 , and H is a disjoint union of even cycles in which the edges alternate between M1 and
M2 . We have two cases to consider.
Case 1 Let xz and yw belong to different components of H , as shown in Figure 8.18 (a).
If yw belongs to the cycle C, then the edge of M1 in C together with the edges of M2 not
belonging to C form a 1-factor of G , contradicting the choice of G .

Fig. 8.18
Case 2 Let xz and yw belong to the same component C of H . Since each component of
H is a cycle, C is a cycle (Fig. 8.18 (b)). By symmetry of x and z, it can be assumed that
the vertices x, y, w and z appear in that order on C. Then the edges of M1 belonging to the

212

Matchings and Factors

yw . . .z section of C , together with the edge yz, and the edges of M2 not in the yw . . . z section
of C form a 1-factor of G , again contradicting the choice of G .
Thus each component of G K is complete.
By conditions (8.8.2), k0 (G K) |K|. Therefore one vertex of each of the odd components of G K is matched to a vertex of K . This is because each vertex of K is adjacent to
every other vertex of G . Also the remaining vertices in each of the odd and even components of G K can be matched among themselves (Fig. 8.19). The total number of vertices
thus matched is even. As |V (G )| is even, the remaining vertices of K can be matched among
themselves, thus giving a 1-factor of G , which is a contradiction.
q

Fig. 8.19
Remark Among the various proofs of Theorem 8.8, one proof due to Anderson [4] uses
Halls theorem.
Tuttes characterisation of graphs admitting 1-factors has been used to obtain several
results including those for regular graphs and many of these are reported in the review
article by Akiyama and Kano [1]. We present some of these results. The first is due to
Peterson [181].

Theorem 8.9 (Peterson)

Every cubic graph without cut edges has a 1-factor.

Proof Let G be a cubic graph without cut edges and let S V . Let O1 , O2 , . . ., Ok be the
odd components of G S and let mi be the number of edges of G having one end in V (Oi )
and the other end in S. As G is cubic,
therefore,

d(v) = 3 |V (Oi ) |

(8.9.1)

vV (Oi )

and

d(v) = 3 | S |.

(8.9.2)

vS

Now, E(Oi ) = [V (Oi ), V (Oi ) S] [V (Oi ), S], where [A, B] denotes the set of edges having
one end in A and the other end in B, A V, B V . Thus, mi = | [V (Oi ), S] | = d(v)
vV (Oi )

2 |E(Oi )|. Since d(v) = 3 for each v and V (Oi ) is odd component, mi is odd for each i. Also,

Graph Theory

213

since G has no cut edges, mi 3. Therefore, k0 (G S) = k

by Theorem 8.8, G has a 1-factor.

1
3

d(v) < 31 3 |S | = | S |. Hence

vS

Remark A cubic graph with cut edges may not have a 1-factor. To see this, consider the
graph G1 of Figure 8.20 (a). If we take S = {v}, then k0 (G S) = 3 > 1 = |S| and hence G1
has no 1-factor.
Also, a cubic graph with a 1-factor may have cut edges. Consider the graph G2 of Figure
8.20 (b). Clearly, < e1 , e2 , e3 , e4 , e5 > is a 1-factor and e3 is a cut edge of G2 .

Fig. 8.20

The next result is due to Cunningham [62].

Corollary 8.2 The edge set of a simple 2-edge-connected cubic graph G can be partitioned into paths of length three.

Proof By Theorem 8.9, G is a union of a 1-factor and 2-factor. Orient the edges of each
cycle of the 2-factor in any manner so that each cycle becomes a directed cycle. Then, if
e is any edge of the 1-factor and f 1 , f 2 are the two arcs of G having their tails at the end
vertices of e, then {e, f 1 , f 2 } form typical 3-path of the edge partition of G (Fig. 8.21). q

Fig. 8.21

214

Matchings and Factors

Berge [23] obtained the following estimate for 1 (G) of a graph G, the proof of which is
given by Bollobas [29].
Theorem 8.10 For any simple graph G, 1 (G) = (n d)/2, where d = max{k0 (G S) |S| :
S V }.

Proof Clearly at most n d vertices can be saturated in any matching of G. In order to

show that this bound is attained, we use the following construction. Let H = GV Kd . Then
H satisfies Tuttes condition. For if S0 is any subset of V (H), then k0 (H S0 ) = 0 or 1 if
T
S0 6 V (Kd ). Also, if S0 V (Kd ), then k0 (H S0 ) = k0 (G S), where S = S0 V (G), so that
k0 (H S0 ) |S0 | = k0 (G S) |S| d = 0, by definition of d . Therefore, by Tuttes theorem,
H has a 1-factor M . The restriction of M of G is a matching in G which does not cover at
most d vertices of G.
q
Remark Berge actually used the theory of alternating paths to prove Theorem 8.10 and
then deduced the Tuttes theorem.

The restatement of Theorem 8.10 is as follows.

Corollary 8.3 If G is a graph with n vertices and 1 (G) = a, then there is a set S of s
vertices such that k0 (G S) = n + s 2a.
The following result is due to Babler [10].
Theorem 8.11
factor.

Every k-regular, (k 1)-edge-connected graph with even order has a 1-

Proof Let G be a k-regular, (k 1)-edge-connected graph with even order and let S be
any subset of V . Let O1 , O2 , . . ., Or be the odd components of G S with order ni and size
mi , and let ki edges join Oi to S, for 1 i r. Then ki k 1,
(8.11.1)
d(v) = kni = 2mi + ki

(8.11.2)

vV (Oi )

and

d(v) = k|S| ki + 2|E(S)|.

(8.11.3)

vS

Now, (8.11.2) gives ki = kni 2mi , so that if k is even, so is ki , and thus ki > k 1 implies
ki k. Similarly, if k is odd, so is ki , and again ki > k 1 implies ki k. Summing this over
r odd components, we obtain by using (8.11.3)
r

rk ki ki + 2|E(S)| k|S| .
i=1

i=1

Therefore, k0 (G S) = r |S| and by Tuttes theorem, G has a 1-factor.

Graph Theory

215

8.3 Antifactor Sets

Definition: A subset S V of a graph G(V, E) such that k0 (G S) > |S| is called an
antifactor set. An antifactor set S such that no proper subset of S is an antifactor set is
called a minimal antifactor set. For example, in the graph of Figure 8.20(a), {v} is an
antifactor set.
Definition: A vertex v V is a claw centre if and only if at least three vertices of G
adjacent to v form an independent set. Clearly, vertex v is a claw centre in the graph of
Figure 8.20(a).
Now assume that G is a graph of even order n and let S be an antifactor set of G. Let
k0 (G S) = k and O1 , O2 , . . ., Ok be the odd components of G S. Since n is even, |S| and k
have the same parity. Choose a vertex ui V (Oi ), 1 i k. Then |S {u1, u2 , . . ., uk }| is even,
that is, k + |S| 0 (mod 2). This implies that k |S| (mod 2) and therefore k0 (G S) |S|
(mod 2). Thus we have the following result.
Lemma 8.4

If S is an antifactor set of a graph G of even order, then k0 (G S) |S| + 2.

Tuttes theorem implies that a graph G has either a 1-factor or an antifactor set. It is
necessary to mention here that the computational complexity is high in the verification of
Tuttes condition for every subset S of V . In order to reduce the computations, the antifactor
sets are of great significance. The following result on antifactors is due to Sumner [236].
Lemma 8.5 If S is a minimal antifactor set of a connected graph G of even order and
having no 1-factor, and if k0 (G S) = r and |S| = s, then
i. r s + 2,
ii. each vertex of S is adjacent to vertices in at least r s + 1 distinct odd components of
G S,
iii. every vertex of S is a claw centre.

Proof
i. Already established (Lemma 8.4).
ii. If v S is adjacent to h distinct odd components say O1 , O2 , . . ., Oh with h r s,
then G {S {v}} has the odd components Oh+1 , Oh+2 , . . ., Or . Then, if h is odd,
k0 (G S0 ) = r h s > s 1 = |S0 |, and if h is even, then {O1 O2 . . . Oh } {v} is
also an odd component of G S0 so that k0 (G S0 ) = r h + 1 s + 1 > s 1 = |S0 |,
where S0 = S {v}. That is, S0 is an antifactor set, contradicting the minimality of S.
Thus, h r s + 1.
iii. Since any v S is adjacent to at least r s + 1 3 (by (i)) distinct odd components, v
is a claw centre, by definition.
q

216

Matchings and Factors

Remark The above observations imply that Tuttes theorem is equivalent to the following.
A graph G has a 1-factor if and only if G does not have a set S of claw centres such that
k0 (G S) > |S|. Obviously, this form of Tuttes theorem decreases the number of subsets S
of V to be verified in Tuttes condition.

An improvement of Theorem 8.11 is due to Plesnik [209].

Theorem 8.12 If G is a k-regular, (k 1)-edge-connected graph of even order and G1 is
obtained from G by removing any k 1 edges, then G1 has a 1-factor.

Proof Assume G1 has no 1-factor. Then G1 has an antifactor set S with k0 (G1 S) > |S|.
Therefore by Lemma 8.4,
k0 (G1 S) |S| + 2.

(8.12.1)

Let O1 , O2 , . . ., Or be the odd components and Or+1 , Or+2, . . ., Or+s be the even components of G1 S. For a component Oi , let ai and mi be the number of edges respectively
of E(G) E(G1 ) and E(G1 ) joining it to S and bi be the number of edges of E(G) E(G1 )
joining it to the other O j s. Then the total number of edges going out of Oi is ai + bi + mi .
Since G is (k 1)-edge-connected, ai + bi + mi k 1. By the argument used in Theorem
8.11, this implies that ai + bi + mi k, for an odd component Oi . Summing over the r odd
components, we obtain
r

i=1

i=1

i=1

ai + bi + mi rk.

(8.12.2)

The number of edges between the Oi s, and between Oi s and S, removed from G is
t

1
2

ai +
i=1

bi , where t = r + s. As this is at most k 1, we have

i=1

i=1

i=1

2 ai + bi 2(k 1).
t

i=1

i=1

Also, ai + mi d(v) = k|S|.

(8.12.3)
(8.12.4)

vS

Adding (8.12.3) and (8.12.4), we get

t

i=1

i=1

i=1

3 ai + bi + mi k(|S| + 2) 2.

(8.12.5)

As the sum on the left of (8.12.2) is less than the sum on the left of (8.12.5), we have
rk k(|S| + 2) 2 < k(|S| + 2), giving r < |S| + 2, which contradicts (8.12.1).

Graph Theory

217

We now have the following observation.

Corollary 8.4

A (k 1)-regular simple graph on 2k vertices has a 1-factor.

Proof Assume G is a (k 1)-regular simple graph with 2k vertices having no 1-factor.

Then G has an antifactor set S and by Lemma 8.4, k0 (G S) |S| + 2. Therefore, |S| + (|S| +
2) 2k and so |S| k 1 . Let |S| = k r. Then r 6= 1. For if r = 1, then |S| = k 1 and
therefore k0 (G S) = k + 1. Thus each odd component of G S is a singleton and therefore
each such vertex is adjacent to all the k 1 vertices of S, as G is (k 1) regular. This implies
that every vertex of S is of degree at least k + 1, a contradiction. Thus, |S| = k r, 2 r
k 1. If G1 is any component of G S and v V (G1 ), then v can be adjacent to at most
|S| vertices of S. Since G is (k 1)-regular, v is adjacent to at least (k 1) (k r) = r 1
vertices of G1 . So, |V (G1 )| r. Counting the vertices of S, we obtain (|S| + 2)r + |S| = 2k, or
(k r + 2)r + (k r) 2k. This gives (r 1)(r k) 0, violating the condition on r.
q
The next result is due to Sumner [236].
Theorem 8.13 If G is a connected graph of even order n and claw-free (that is, contains
no K1, 3 as an induced subgraph), then G has a 1-factor.

Proof If G has no 1-factor, then G contains a minimal antifactor set S of G. Also, there
is an edge between S and each odd component of G S. If v S and vx, vy and vz are edges
of G with x, y and z belonging to distinct odd components of G S, then a K1, 3 is induced
in G. This is not possible, by hypothesis.
Since k0 (G S) > |S|, there exists a vertex v of S and edges vu and vw of G, with u
and w in distinct odd components of G S. Assume Ou and Ow to be the odd components
containing u and w respectively. Then < Ou Ow {v} > is an odd component of G S1 ,
where S1 = S {v}. Also, k0 (G S1 ) = k0 (G S) 1 > |S| 1 = |S1| and therefore S1 is an
antifactor set of G with |S1 | = |S| 1, a contradiction to the choice of S. So G has a 1-factor.
(Clearly, by Lemma 8.4, the case |S| = 1 and k0 (G S) = 2 does not exist).
q
The following results are immediate, as each type of graph in these cases does not contain K1, 3 as an induced subgraph.
Corollary 8.5

Every connected even order edge graph has a 1-factor.

Corollary 8.6

If G is a connected graph of even order, then G2 has a 1-factor.

Corollary 8.7

Every connected total graph of even order has a 1-factor.

Corollary 8.8
a 1-factor.

Every connected cubic graph in which every vertex lies on a triangle has

Corollary 8.9

If a connected even order graph G has less than (G) claw centres, then

G has a 1-factor.

218

Matchings and Factors

The following result is the generalisation of Theorem 8.13.

Theorem 8.14
a 1-factor.

If G is n-connected even order graph with no induced K1, n+1 , then G has

Proof Assume G has no 1-factor. Then G has a minimal antifactor set S. Let |S| = s and
let the odd components of G S be O1 , O2 , . . ., Ok . If S0i = {v S : v is adjacent to some
vertex to Oi }, then S0i is a vertex cut of G and therefore |S0i | n. Thus there are at least n
edges from Oi incident with n distinct vertices of S. The number of such edges as i ranges
from 1 to k is at least nk and this is greater or equal to n(s + 2), by Lemma 8.4. Thus at least
one vertex v of S is incident with at least n + 1 of these edges. But then, v is the centre of an
induced K1, n+1 , contradicting the assumption.
q
It can be observed that a graph can have more than one 1-factors and different 1-factors
can either have no edge in common or have some edges in common. Depending on the
nature of edges, we have the following definition.
Definition: In a given graph, two 1-factors are said to be disjoint if they have no edge in
common and two 1-factors are said to be distinct if they have at least one different edge. For
example, in the graph of Figure 8.22, the 1-factors {e1 , e2 , e3 } and {e4 , e5 , e6 } are disjoint,
while as the 1-factors {e1 , e2 , e3 } and {e1 , e7 , e5 } are distinct.

Fig. 8.22

Consider maximal sets S which satisfy Tuttes condition in graphs admitting a 1-factor.
If such a graph is connected, then definitely there exist subsets S such that k0 (G S) = |S|.
For such a graph will have a vertex v which is not a cut vertex, and G v being connected
and of odd order, k0 (G v) = 1 = |v|. The maximal subsets S for which k0 (G S) = |S| holds
is called a maximal factor set.
A result on counting distinct 1-factors for bipartite graphs is given by Theorem 8.4. The
next result as reported in Bollobas [29] gives the bounds for distinct 1-factors in a graph..
Theorem 8.15 Let G be a graph admitting a 1-factor and let S0 be a maximal factor set
of G with |S0| = m and O1 , O2 , . . ., Om be the odd components of G S0 . Further, let d = min
{|N(Oi) So | : 1 i m} and d 0 be the minimum number of edges joining Oi to S0 , 1 i m.
Then the number F(G) of 1-factors of G satisfies

Graph Theory

219
min(d,m)

i. F(G) r(d, m) =

(d + 1 i) and

i=1

ii. F(G) r (d 0 , m), where r (d 0 , m) is the minimum number of 1-factors in a bipartite

multigraph with m vertices in each class, in which d 0 is the minimum degree in the
first vertex class.

Proof Let S0 be a maximal factor set of G so that k0 (G S0 ) = |S0 |.

Using Halls theorem for a complete matching in a bipartite graph, we can show that |S0 |
odd components of G S0 can be paired with the vertices of S0 . The sufficient condition is
that any k such odd components should be joined in G to at least k vertices in S0 . Clearly,
this is satisfied here. For if some set of k odd components are joined in G only to a subset
T of S0 with h vertices, h < k, then k0 (G T ) k > h = |T|, contradicting the hypothesis.
Therefore one vertex each of |S0 | odd components can be matched to one vertex of S0 . Let
O be an odd component of which a vertex a has been matched with a corresponding vertex
s of S0 . Clearly, O0 = O a is a graph on an even number of vertices. Now we verify Tuttes
condition on O0 . If there is a T V 0 = V (O0 ) such that k0 (O0 T ) > |T |, then k0 (O0 T )
> |T | + 2. Let S = S0 T {a}. Then k0 (G S) = k0 (O0 T ) + k0 (G S0 ) 1 = k0 (O0 T )+
|S0| 1 > |T| +|S0| +1 = |S|, contradicting the choice of S0 . Thus Tuttes condition is satisfied
for O0 and O0 has a 1-factor.
If G S0 has an even component C and a V (C), then G (S0 {a}) has at least one
more odd component than G S0 , contradicting the choice of S0 or the Tutte condition.
Thus G S0 has no even components.
Therefore we have seen that for each oi Oi such that there is an si S0 with oi si E ,
each Oi oi has a 1-factor forming part of a 1-factor of G. The number of such choices of
m independent edges oi si equals the number of 1-factors in a bipartite graph H(V1 V2 , E),
where V1 has m vertices, one corresponding to each odd component Oi of G S0 and V2
has the m vertices of S0 . If we take the edge set E to be such that whenever s j N(Oi ) S0 ,
we take exactly one oi si E , then we get an ordinary bipartite graph for H in which the
minimum degree in V1 is d = min{|N(Oi) S0 |} and then (i) follows from Theorem 8.4.
However, if we take E to be such that for each edge joining Oi to s j we take an edge oi s j ,
then we get a bipartite multigraph for H in which the least degree of a vertex in Vi is d 0 as
defined and then (ii) follows.
q
The following observations can be seen immediately.
Corollary 8.10

For a k-edge-connected graph with a unique 1-factor, F(G) k.

Corollary 8.11 If G is a connected graph with a unique 1-factor, then G has a cut edge
which belongs to a 1-factor.
The details of the following results can be found in Bollobas [29].
Theorem 8.16 If (n) is the minimum degree of a graph of order 2n with exactly one
1-factor, then (n) [log2(n + 1)].

220

Matchings and Factors

Theorem 8.17 If e(n) is the maximum size of a graph of order 2n with exactly one
1-factor, then e(n) = n2 .

8.4 The f-factor Theorem

The f -factor Theorem gives a set of necessary and sufficient conditions for the existence
of an f-factor in a general graph. This result was proved by Tutte [250] for loopless graphs,
using the theory of alternating paths, and later again by Tutte [251], using the 1-factor
Theorem.
First of all, we have the following definitions.
Definition: A vertex function for a general graph G(V, E) is a function f : V N0 such
that 0 f (v) d(v), for every v V .
For any function f : V N0 and for any U V , we set f (U) = f (v).
vU

Definition: Three pair-wise disjoint (possibly empty) subsets S, T and U of the vertex
set V of a general graph G, such that V = S T U , is called a decomposition D = (S, T, U)
of G. The components, say C1 , C2 , . . ., Ck , of the vertex-induced subgraph < U > of G are
called the components of U . For each Ci , let hi = f (Ci ) + qG [T, Ci ], where f is a given vertex
function of G, and qG [T, Ci ] denotes the number of edges of G, whose one end is in T and
the other end is in Ci . Then Ci is said to be an odd or even component of U , depending upon
hi being odd or even, and the number of odd components of U is denoted by k00 (D, f ).
The f-deficiency of D for the given G is defined as
(D, f ) = k00 (D, f ) f (S) + f (T ) d(T ) + q[S, T ].

The decomposition D is defined to be an f-barrier of G if (D, f ) > 0.

The term decomposition is used by Graver and Jurkat [88], and Tutte calls it a G-triple.
These decompositions play the same role as the antifactor sets in the theory of 1-factors.
Before going into the details, we first state the f -factor Theorem.
Theorem 8.18 If f is a vertex function for a graph G, then either G has an f -factor or an
f -barrier but not both.
It can be seen that both the statement and the proof of the f -factor Theorem as in Tutte
[253] definitely have similarities to those of the 1-factor Theorem. Since there are few
complications in the constructions involved which make the proof given by Tutte [253]
some what lengthy, we follow here the earlier proof by Tutte [251].
Tutte [250] defines an ordinary graph G0 corresponding to a given graph G such that G
has an f -factor if and only if G0 has a 1-factor and shows that the 1-factor condition corresponds to the f -factor condition, and derives the f -factor Theorem as a consequence of
the 1-factor Theorem. Berge [23] uses the same graph construction to derive the maximum

Graph Theory

221

f -matching Theorem. Since we have general graphs under consideration, we construct G0

with suitable changes as reported by Parthasarthy [180], and then find that the proof still
holds.
Let G be a general graph with the vertex function f . This defines for each vertex vi , its
degree di = d(vi |G), f i = f (vi ) and ki = di f i . For each vertex vi of G, take two subsets Xi
and Yi of vertices of G0 , where |Xi | = di and |Yi | = ki . Let the vertices of Yi be labelled yij ,
1 j ki , arbitrarily.Now, label the vertices of Xi in the following manner.
Label the edges of G arbitrarily as 1, 2, . . ., m. Since |Xi | = number of link edges li
incident with vi plus twice the number of loops `0i incident with vi , a vertex xij Xi is
taken corresponding to each link edge j incident with vi and two vertices xih , xih0 are taken
corresponding to each loop h incident at vi . As for edges, if link j joins vi and vk , then xij xkj is
made an edge. Corresponding to loop h incident with xih the edge xih xih0 is taken. In addition,
all the edges of the complete bipartite subgraph with vertex partition {Xi , Yi } are taken for
each i. These constitute the graph G0 . The induced subgraph < Xi Yi > of G is called the
star graph at i and is denoted by St(i). For multi graph G, it is a complete bipartite graph,
but for general graphs it will contain edges in < Xi > which we call loop edges. The other
edges of St(i) are called star edges and the remaining edges of G0 are called as link edges.
This construction is illustrated in Figure 8.23. In the general graph G, d(v1 ) = 5, d(v2 ) = 3
and d(v3 ) = 2. Let f = (3, 2, 1) be the vertex function, so that f (v1 ) = 3, f (v2 ) = 2 and
f (v3 ) = 1. Therefore, k1 = 2, k2 = 1 and k3 = 1. Here |X1 | = 5, |X2 | = 3 and |X3 | = 2, and
|Y1| = 2, |Y2 | = 1 and |Y3 | = 1.

Fig. 8.23

222

Matchings and Factors

Label the edges of G as 1, 2, 3, 4 and 5 as shown. Then Y1 = {y11 , y12 }, Y2 = {y21 } and Y3 = {y31 },
and we have X1 = {x11 , x110 , x12 , x13 , x14 }, X2 = {x22 , x23 , x25 } and X3 = {x34 , x35 }. The vertex set of
G0 is Xi Yi , 1 i 3. The edges of G0 are x11 x110 , x12 x22 , x13 x23 , x14 x34 , x25 x35 (shown by dotted
lines) and the star edges (shown by bold lines). The dotted lines in G form an f -factor with
f = (3, 2, 1).
The following result gives a relation between f -factors of G and the 1-factors of G0 .
Theorem 8.18

A general graph G has an f -factor if and only if G0 has a 1-factor.

Proof First, assume G has an f -factor F with link edges L and loop edges L0 . Let F have
gi link edges and f io loop edges at the vertex vi . Then f i = gi + 2 f i0 |Xi | = di = li + 2li0 and
|Yi | = ki = di f i = (li + 2li0 ) (gi + 2 f i0) = (li gi ) + 2(2li0 f i0 ), where li is the number of link
edges incident with vi and li0 is the number of loops incident with vi . Let L0 and ( L0 )0 be
the set of corresponding set of link and loop edges in G0 . The matching consisting of these
edges, keeps unsaturated all the vertices of Yi and a set of li gi + 2(lio f io ) = ki vertices of
Xi , for 1 i n. Now match the vertices of Yi with these unsaturated vertices of Xi using
the star edges St(i) and let S0 be this set of edges. Clearly L0U(L0 )0 S0 is a 1-factor in G0 .
Conversely, assume G0 has a 1-factor H . As the vertices in Yi are matched only to vertices
in Xi , there are exactly di (di f i ) = f i = gi + 2 f i0 vertices of Xi which are not matched by
H to vertices in Yi . Clearly, 2 f i0 vertices from these unmatched vertices of Xi correspond to
loop edges of < Xi > and thus are matched in pairs amongst themselves. The remaining
gi vertices are matched by the link edges of H to X j vertices of other star graphs St(i) of
G0 . If F is the set of edges of G corresponding to the loop edges and link edges, then d
(vi | < F >) = gi + 2 f i0 . Thus f is an f -factor of G.
q
The following definition will be required for the graph G0 constructed as above.
Definition: A subset W of the vertex set V 0 of the graph G0 is said to be simple if the
following conditions get satisfied.
i. (Xi W ) 6= implies Xi W ,
ii. (Yi W ) 6= implies Xi W and
iii. At most one of Xi and Yi is a subset of W , implies that if Yi = , then Yi W , that is
Xi W = .
Now assume W be a simple set of V 0 , S be the set of vertices vi of G, for each Xi
W and T be the set of vertices v j of G for each Y j W . We observe that S T = , and
S T (V S T ) is a decomposition D of V .
Denoting by d(G, S) = k0 (G S) |S|, for any subset S of the vertex set V of a graph,
Theorem 8.8 can be restated in the following form.
Theorem 8.19
d(G, S) > 0.

A simple graph G has a 1-factor if and only if it has no subset S V with

Graph Theory

223

Theorem 8.20 If W is a simple subset of V 0 and D is the corresponding decomposition

of V , then d (G0 , W ) = (D, f ).

Proof

Since W is a simple subset of V 0 and D is the corresponding decomposition of V ,

|W | = |X j | + |Y j | = d j + (d j f j ) = d(S) + d(T ) f (T ).
v j S

v j T

v j S

(8.20.1)

v j T

Thus, d (G0 , W ) = k0 (G0 W ) d(S) d(T ) + f (T ).

We proceed to find k0 (G0 W ) and in doing so we consider the components K of G0 W .
We have the following possibilities.
i. The component K contains a single vertex of the form xir . This is possible when Yi
W and the edge r is a link edge whose other end is in Y j W . Thus, in this case, K
is an odd component, and the number of such odd components is equal to the number
of such link edges and equals q[S, T ].
ii. The component K contains a single vertex of the form yi , and this is possible when
Xi W . This again is an odd component and the number of such odd components is
equal to

|Yi| = (di fi ) = d(S) f (S).

iS

iS

iii. K contains a single edge of the form xi x corresponding to a loop of G at . This is

possible when Yi W . Clearly, this is an even component of G0 W .
iv. K contains a single edge of the form xir xrj , where r is the link edge of G joining vertices vi and v j . Clearly, this happens when Yi W and Y j W . This also is an even
component.
v. K contains more than one edge and we call such components large components.
Clearly, they have one star graph St(i) as a subgraph. If K contains more than one
such star graph as a proper subgraph, it will also contain all edges of G0 between
such star graphs. The other edges of K are of the form xir xrj , where r is an edge of G
joining vi and v j , and St(i) K and Y j T . Let vi1 , vi2 , . . . , vik be the vertices of G
corresponding to the star graphs St(i j ), 1 j k, which are subgraphs of K . Let C be
the induced subgraph < vi1 , vi2 ,. . . , vik > of G. Then C is a component of G S T and
therefore the number of vertices of the large component K is given by
n(K) = |V (K)| = |V (St(i))| + q[C, T ]
vi C

= (di + di f i ) + q[C, T ]
vi C

224

Matchings and Factors

= f i + q[C, T ] (mod 2)
vi C

= f (C) + q[C, T ] (mod 2).

Therefore K is an odd component of G0 W if and only if C is an odd component of

G S T . So the number of large odd components of G0 W is k00 (D, f ).
Combining the above results, we observe that the number of odd components of G0 W
is given by
k0 (G0 W ) = q[S, T ] + d(S) f (S) + k00 (D, f ).

(8.20.2)

Now, (D, f ) = k00 (D, f ) f (S) + f (T ) d(T ) + q[S, T ]

= k0 (G0 W ) d(S) + f (T ) d(T ),

(by using (8.20.2))

= d(G0 , W ).

(by using (8.20.1))

Now, we have the following main result.

Theorem 8.21

A general graph G has an f -factor if and only if it has no decomposition

D with (D, f ) > 0.

Proof
Necessity Assume there are decompositions D for which (D, f ) > 0 and choose one
such decomposition for which |S| is least. Claim that for each vi S, f (vi ) < d(vi ). If not,
there is a v S such that f (v) = d(v). Consider the decomposition D0 = ( S0 , T 0 , U 0 ), where
S0 = S {v}, T 0 = T {v} and U 0 = U . Then the components of < V S0 T 0 > are the
same as those of < V S T >. Also, for such a component C, f (C) is unchanged, but
q[C, T 0 ] = q[C, T ] + q[C, v]. Therefore at most q[v, U] components can change from odd to
even. That is, k00 (G S0 T 0 ) k00 (G S T ) q[v, U].
Thus, (D0 , f ) = k00 (G S0 T 0 ) + f (T 0 ) f (S0 ) d(T 0 ) + q[S0 , T 0 ]
= k00 (G S0 T 0 ) + f (T ) + f (v) f (S) + f (v) d(T ) d(v)+
q[S, T ] q[v, T ] + q[v, S]
= k00 (G S0 T 0 ) + f (T ) f (S) d(T ) + q[S, T ] + 2 f (v) d(v)
q[v, T ] + q[v, S]
k00 (G S T ) q[v, U] + f (T) f (S) d(T ) + q[S, T ] + d(v)

Graph Theory

225
q[v, T ] + q[v, S]

(as d(v) = f (v))

= (D, f ) + d(v) q[v, U] q[v, T ] + q[v, S]

= (D, f ) + 2q[v, S] > (D, f ) > 0.

This contradicts the choice of D and thus the claim is established.

Now, suppose W = {Xi : vi S} {Y j : v j T }.
Then W is simple, as S and T are disjoint. Therefore by Theorem 8.20, d (G0 , W ) =
(D, f ) > 0 and G0 has no 1-factor. Thus by Theorem 8.18, G has no f -factor.
Sufficiency Assume that G has no f -factor, so that G0 has no 1-factor. Therefore, by
Theorem 8.19, there is a subset W of V 0 with d(G0 , W ) > 0. We choose such a W with
least cardinality and prove that W is simple.
Y j W 6= and Y j W .
If this is not true, then there is a Y j such that Y j W 6= , and Y j (V 0 W ) 6= . Define
Q = W (Y j W ) (Fig. 8.24).
Claim 1

Fig. 8.24

If X j 6 W , then G0 W and G0 Q differ in one component only, the component star

edges.
If X j W , then each component of G0 W is also a component of G0 Q.
In either case, k0 (G0 Q) k0 (G0 W ) 1 and |Q| |W | 1. Therefore, d(G0 , Q) > 0,
contradicting the choice of W .
Xi W 6= implies Yi W .
If this is not true, then there is an i such that Xi W 6= and Yi W . Let a Xi W and
define Q = W {a}. Then G0 W can have at most one component with a vertex belonging
to Yi which is adjacent to a in G. Thus,
Claim 2

k0 (G0 W ) k0 (G0 Q) + 1 and |Q| = |W | 1.

Therefore, d(G0 , Q) d(G0 , W ), contradicting the choice of W .

226

Matchings and Factors

Now claim (1) and claim (2) together imply that both Xi and Yi are not subsets of W .
Xi W 6= implies Xi W .
If not, there is an i such that Xi W 6= and Xi (V 0 W ) 6= . Let a Xi W and define
Q = W {a}. By claims (2) and (1), Yi G0 W , so that all vertices of Yi belong to a single
component of G0 W . Also there is at most one component of G0 Q which has a vertex
not belonging to Yi which is adjacent to an a in G0 (Fig. 8.25). Therefore G0 W and G0 Q
differ in at most two components.

Claim 3

Fig. 8.25

Thus, k0 (G0 Q) k0 (G0 W ) 2

and d(G0 , Q) d(G0 , W ) 1.

(8.21.1)

But k0 (G0 W ) + |W | |V 0 | and k0 (G0 Q) + |Q| |V 0 |.

So, k0 (G0 W ) |W | |V 0 | k0 (G0 Q) |Q| and d(G0 , W ) d(G0 , Q).
Therefore, (8.21.1) becomes d(G0 , Q) d(G0 , W ) > 0, contradicting the choice of W .
Hence we have proved that W is a simple set and by Theorem 8.20, for the corresponding
decomposition D of G, (D, f ) > 0.
q
Remark Theorem 8.21 shows that the 1-factor Theorem implies the f -factor Theorem.
The converse is also true as can be seen in Theorem 8.22. Before that we have the following
definition.

Definition: Suppose D = (S, T, U) is a decomposition of V and x S, y T . The decomposition D1 = (S {x}, T, U {x}) is said to be obtained by an x-transfer from S and the
decomposition D2 = (S, T {y}, U {y} is said to be obtained by y-transfer from T .
The following result is an immediate consequence of the above definition and can be
easily established.

Graph Theory

227

Lemma 8.6 An x-transfer from S does not reduce the deficiency of D if f (x) = d(x) or
d(x) 1 and a y-transfer does not reduce the deficiency of D if f (y) = 0 or 1.
Theorem 8.22

The f -factor Theorem implies the 1-factor Theorem.

Proof Clearly, by the f -factor Theorem with f (v) = 1, for all v V , G has a 1-factor if
and only if it has a 1-barrier D = (S, T, U). But by Lemma 8.6, D1 = (S, , V S) has no
less deficiency than D (by y-transfer from T ). Therefore G has no 1-factor if and only if
there is a 1-barrier of the form D1 . But (D1 , f ) = k0 (G S) f (S) d(T ) + f (T ) + q[S, T ] =
k0 (G S) |S| and D1 is a 1-barrier if and only if (D1 , f ) > 0, that is k0 (G S) > |S|. Hence
the 1-factor Theorem.
q
Corollary 8.12 The Erdos-Gallai Theorem 2.4 on degree sequences follows from the
f -factor Theorem.

Proof Theorem 2.4 gives a necessary and sufficient condition for a non-increasing sequence d = [ di ]n1 of non-negative integers to be the degree sequence of a simple graph. In
order to get this condition, we observe that d is graphic if and only if the complete graph
n
Kn has a d -factor, where d(vi ) = di , given di = 2m.
i=1

By the f -factor Theorem, d fails to be graphic if and only if Kn has a d -barrier D =

(S, T, U). Since for Kn , U can have at most one component, k0 (D, d) = 0 or 1, further
(D, d) > 0 is even, as d(Kn ) = di = 2m is even. Therefore the condition (D, d) > 0
becomes
qkn = [S, T ] d(S) (n 1) |T | + d(T ) > 0.

(8.12.1)

Now transfer of elements t from T to U , or elements s from S to U does not decrease

the deficiency (D, d). So elements t T with d(t) < |T | + |U| and elements s S with d(s)
> |T | in S can be transferred to U without decreasing the deficiency. Also it can be seen
that elements u U with d(u) |T | + |U| can be transferred to S without decreasing the
deficiency. Therefore the f -barrier can be made such that d(t) > |T| + |U| for all t T , |T |
< d(u) < |T | + |U| for all u U and d(s) |T | for all s S, and d is non-graphical if and
only if Kn has such a d -barrier D. But for such a D, q[S, T ] = r(n r |U|), where |T | = r,
r

d(T ) = di , and thus condition (8.12.1) becomes

1

r(n r |U|) d(S) r(n 1) + d(T) > 0.

r

That is, di > r(n 1) + d(S) r(n r |U|)

1

= r(r 1) + r(n r) + d(S) r(n r |U|)

= r(r 1) + r|U| + d(S)

228

Matchings and Factors

n

> r(r 1) + min{r, di },

i=r+1

which is the Erdos-Gallai condition.

8.5 Degree Factors

Let d = [di ]n1 and r = [di ki ]n1 be graphical sequences. The problem we consider here is
the existence of a graph G realising d with a spanning subgraph (factor) realising k = [ki ]n1 .
This problem was asked by Rao and Rao [215] as k-factor conjecture, when k ki k + 1
for each i, and solved by Kundu [143]. An improvement of this was due to Kleitman and
Wang [129] and further improvement by Kundu [141]. For ki = 1, 1 i n, the problem was
earlier posed by Grunbaum [93] and proved by Lovasz [149], and this proof also applies
when ki = k, 1 i n.
Theorem 8.23 (k-factor Theorem) If [di ]n1 is a graphical sequence such that [di k]n1 is
also graphical, then there is a realisation of [di ]n1 with a k-factor.

Proof

Let G1 and G2 be two graphs on V = {v1 , v2 , . . ., vn } such that d(vi |G1 ) = di and
d(vi |G2 ) = di k. Assume that G1 and G2 are chosen such that G2 has maximum edges
common with G1 . We show that all edges of G2 are in G1 . Let E1 = E(G1 ) and E2 = E(G2 ).
Claim

If vi v j E2 E1 and vi vk , v j vr E1 E2 (i, j, k, r being distinct), then vk vr E1 E2 .

If vk vr E2 , then an EDT in G2 switching the pair vi v j , vk vr to the positions vi vk , v j vr gives a
graph degree-equivalent to G2 with more edges common with G1 , contradicting the choice
of G1 and G2 . If vk vr
/ E1 , then an EDT in G1 switching the pair vi vk , v j vr to the positions
vi v j , vk vr gives a graph degree-equivalent to G1 , such that G2 has more common edges with
this graph, again a contradiction. Thus, the claim is established.
Let G2 has some edges which are not in G1 and let v1 be a vertex with a maximum
number t of edges of E2 E1 incident with it. By the degree stipulation on G1 and G2 , there
are t + k vertices v2 , v3 , . . ., vt+k+1 which are adjacent to v1 in G1 , but not in G2 . Assume
v1 vk E2 E1 and vk v E1 E2 , and consider the following two cases.
Case 1 Let v 6= vi for i, 2 i t + k + 1. Then by the above claim, vvi E1 E2 for 2 i
t +k +1, so that there are t +k +1 edges in E1 E2 adjacent with v. By the degree-stipulation
on G1 and G2 , there are t + 1 edges in E2 E1 incident with v, contradicting the choice of v.

Let v = vi for some i, 2 i t + k + 1, and without loss of generality, take v = v2 .

Then again, by the above claim, vvi E1 E2 for 3 i t + k + 1. Also, vv1 , vvk E1 E2 .
Therefore there are t + k + 1 edges in E1 E2 incident with v, again a contradiction.

Case 2

Graph Theory

229

Hence G2 has no edges which are not in G1 . But then < E(G1 ) E(G2 ) > is a k-factor
of G1 .
q
Lemma 8.7

If [ki ]n1 is a sequence of non-negative integers such that k ki k + 1 for

n

1 i n and for some k, 0 k n 2, and ki is even, then [ki ]n1 is a graphical sequence.
1

Proof

Induct on S = ki . Clearly, the result is trivial for S = 0. Assume it is true for all
1
n

values of ki < S 1 and let [ki ]n1 be a sequence with ki = S (even). Then by Theorem 2.1,
[ki ]n1 is graphical if and only if [k2 1, k3 1, . . . , kk1 +1 1, kk1 +2 , . . ., kn ] is graphical, where
k1 k2 . . . kn . But the sum of the terms in this sequence is S 2k1 , which is even, and
by induction hypothesis is graphical. Hence by Theorem 2.1, [ki ]n1 is graphical.
q

The proof given here of the following k-factor Theorem of Kundu, is due to Chen [57].
Theorem 8.24 If [di ]n1 and [di ki ]n1 are two graphical sequences satisfying k ki k + 1,
for 1 i n and some k > 0, then there is a graph realising [di ]n1 and having a k-factor
k = [ki ]n1 .

Proof Since [di ]n1 and [di ki ]n1 are both graphical, [ki ]n1 is also graphical by Lemma 8.7.
Clearly, there exists a graph with degree sequence [di]n1 containing a k-factor k = [ki ]n1
if and only if there exists a graph with degree sequence [n 1 ki ]n1 containing a k0 -factor
k0 = [n 1 di ]n1 . This can be easily established by taking the complementary graphs. Thus,
the conditions can be shifted from ki s to di s, so that it is enough to prove that if [di ]n1 and
[di ki ]n1 are graphical with k di k + 1 for 1 i n and the ki s unrestricted, then there is
a graph with degree sequence [di ]n1 and a k-factor k = [ki ]n1 .
Let G1 and G2 be graphs with degree sequences [ki ]n1 and [di ki ]n1 respectively, such that
the multigraph G, obtained by superimposing G1 and G2 has a minimum number of multiple
edges. Obviously, in the superimposed graph, there are at most two edges between any pair
of vertices, one of G1 and one of G2 . If there are no such multiple edges, then G1 is a
factor of G.
Assume there are multiple edges in G and let uv be one such multiple edge. Since
d(u|G) = d(u|G1 ) + d(u|G2) n 1 and uv is a multiple edge, there is a vertex w not adjacent
to u in G. Let V be the common vertex set of G1 , G2 and G. If for every x V u, q[v, x]
q[w, x], then d(v|G)2 d(w|G), contradicting the choice that each di is k or k +1. Therefore
there is a vertex x V u such that q[v, x] < q[w, x] (Fig. 8.26).

Fig. 8.26

230

Matchings and Factors

The following two cases arise:

1. q[v, x] = 1 and q[w, x] = 2 ,
2. q[v, x] = 0 and q[w, x] = 0.
In case 1, we have either 1(a) vx G1 , or 1(b) vx G2 , and in case 2, either 2(a) wx G1 ,
or 2(b) wx G2 . For all these cases, make an EDT switching the pair of edges vu, wx to the
positions vx, uw. In 1(a) and 2(b), this is done in G2 to obtain a graph G02 degree-equivalent
to G2 and with less number of multiple edges in G1 G02 . In 1(b) and 2(a), this is done in
G1 to get a graph G01 degree-equivalent to G1 and with less multiple edges in G01 G2 . This
contradicts the choice of G1 and G2 .
q
Remark If di s and ki s being allowed to be arbitrary with [di ]n1 , [di ki ]n1 are graphical,
and so [ki ]n1 is graphical, then [di ]n1 need not be realised with a k-factor. To see this, consider

the following example of Lovasz [149].

Let d = [8, 8, 3, 2, 2, 2, 2, 2, 3] and k = [4, 4, 0, 0, 0, 1, 2, 2, 3] so that d k = [4, 4, 3, 2, 2, 1,
0, 0, 0]. Then d and d k are graphical, realised by graphs G1 and G2 of Figure 8.27, and
evidently k is same as d k with a different labeling. Now, it is easy to see that any realisation of k = d k needs an edge joining the two vertices of degree 4 in it. But if this uv is
an edge in G2 , it will not be an edge in G1 G2 , so that there is no realisation of d with a
k-factor.

Fig. 8.27

Graph Theory

231

8.6 (g, f) and [a, b]-factors

In this section, we give a brief description of (g, f ) and [a, b] factors in a graph.
Definition: If f and g are two vertex functions of a graph G such that g(v) f (v) for all
v V and F is a factor of G such that g(v) d(v|F) f (v) for all v V , then F is called a (g,
f)-factor of G.
If H is a (g, f )-factor of G, its f-deficiency is defined as H (G, f ) = f (V ) d(H), where
d(H) = d(v|H). Clearly, H (G, f ) f (V )(mod2).
vV

If D = (S, T, U) is a decomposition of G, then k00 (D, f , g) denotes the number of odd

components of U with respect to D and f for which f (c) = g(c), for every vertex c of the
component.
We now state the (g, f )-factor Theorem of Lovasz [148].
Theorem 8.25 If f and g are vertex functions of a graph G (V , E ) such that 0 g(v)
f (v) d(v|G) for all v V , then either G has a (g, f )-factor or it has a decomposition D = (S,
T , U ) satisfying k00 (D, f , g) > f (S) + d(T ) g(T ) q[S, T ], but not both, where k00 (D, f , g)
is the number of odd components of < U >G with f (v) = g(v) for every vertex v of the
component.
Next, we state the (g, f )-factor Theorem due to Las Vergnas [255].
Theorem 8.26 Let g, f be vertex functions for a graph G(V, E) such that 0 g(v) 1
f (v) d(v|G) for all v V . For S V , let k0 (G S) be the number of odd components C of
G S which are such that either C = {v} and g(v) = 1 or < V (C) > is odd and at least 3, with
g(v) = f (v) = 1 for all v V (C). Then G has a (g, f )-factor if and only if f (S) k0 (G S) for
all S V .
The following variation of the (g, f )-factor Theorem is due to Kano and Saito [124].
Theorem 8.27 If g and f are vertex functions of a graph G(V, E) and a real number
such that 0 1, and if g(v) < f (v), g(v) d(v|G) f (v) for all v V , then G has a (g,
f )-factor.
Definition: If a and b are non-negative integers with a b and F is a factor of G such
that a d(v|F) b for all v V , then F is called an [a, b]-factor of G. We note that an [a,
b]-factor is a (g, f )-factor for constant vertex functions g(v) = a for all v V and f (v) = b
for all v V .
The following [a, b]-factor Theorem can easily be derived from the (g, f )-factor Theorem.
Theorem 8.28 If G(V, E) is a graph, and a and b are distinct integers with a < b, then
either G has an [a, b]-factor H or a decomposition D = (S, T , U ) such that q[S, T ]> d(T ) +
b|S| a|T|, but not both.

232

Matchings and Factors

The following result is due to Las Vergnas [255].

Theorem 8.29 If G is a graph and k( 2) an integer, then G has a [1, k]-factor if and only
if k|N(S)| |S|, for all independent sets S of G.
The next result can be found in Kano and Saito [125].
Theorem 8.30 If k, r, s and t are integers such that 0 k r, 0 s, 1 t and ks rt , then
every [r, r + s]-graph has a [k, k + 1]-factor.
The following result on the existence of [a, b]-factors is given by Kouider and Long
[139].
Theorem 8.31

If

Let b a + 1 and let G be a graph with minimum degree 8.

4b( a + 1)

(a + 1)2
a(G)

4b( a + 1)

,
a(a + 2)

for a odd
for a odd

then G has an [a, b]-factor.

8.7 Exercises
1. Prove that a tree can have at most one perfect matching.
2. Give an example of a cubic graph having no 1-factor.
3. Show that Kn, n and K2n are 1-factorable, and further show that the number of
1-factors of Kn, n and K2n are respectively n! and

(2n)!
.
2n n!

4. Prove that the Peterson graph is not 1-factorable.

5. Let G = (V1 , V2 , E) be a bipartite graph with |V1 | = n1 and |V2 | = n2 , and the vertices
xi V1 and y j V2 indexed such that d(x1 ) d(x2 ) . . . d(xn ) and d(y1 ) d(y2 )
. . . d(yn ). Then show that a sufficient condition for V1 to be matched into V2 is that
k

k1

i=1

i=1

n1 n2 , d(x1 ) > 0 and d(xi ) > d(y j ), 2 k n1 .

6. Prove that a bipartite graph G = (V1 , V2 , E) has a matching saturating all the vertices
of maximum degree.
7. Show that the following statements are equivalent for a bipartite graph G = (V1 , V2 , E).
a. G is connected and each edge of G is contained in a 1-factor.
b. |V1 | = |V2 | and for each non-empty subset S of V1 , |S| < |N(S)|.

Graph Theory

233

c. For each u V1 and v V2 , G u v has a 1-factor.

8. Show that a tree T has a perfect matching if and only if k0 (T v) = 1 for every vertex
v of T .
9. If G is a cubic graph without cut edges or with all cut edges on a single suspended
path, then prove that G has a 1-factor.
10. Prove directly that if G is a connected graph with no induced subgraph isomorphic to
K1, 3 , then G has an edge uv such that G {u, v} is connected.
11. If M and N are matchings in a graph G and |M| > |N|, then prove that there exist
matchings M 0 and N 0 in G such that |M 0| = |M| 1, |N 0| = |N| +1 and M 0 and N 0 have
the same union and intersection (as edge sets) as M and N .
12. Prove that deleting any perfect matching from the Peterson graph leaves the subgraph
C5 +C5 .
13. Determine the minimum size of a maximal matching in the cycle Cn .
14. If G = (V1 , V2 , E) isa bipartite graph that has a matching of V1 into V2 , then prove that
G has at most |V21 | edges belonging to no matching of V1 into V2 .
15. Describe complete tripartite graphs of the form Kn, n, n that have perfect matching.
16. For what values of n 4 does the wheel Wn have a perfect matching?
17. Prove that a graph has an f -factor if and only if it has an f 0 -factor, where f 0 (v) =
d(v) f (v) for every v V .
18. Show that a k-regular, (k 1)-edge connected graph of even order has a 1-factor.
19. If G is a connected graph of even order, then show that G2 has a 1-factor.
20. If G is a connected cubic graph without a 1-factor and S is a minimal antifactor set,
then prove that S is an independent set of claw centers and G S has exactly |S| + 2
odd components and no even components.

9. Edge Graphs and Eccentricity Sequences

Many authors discovered edge graphs independently and gave it a different name, for example, interchange graph by Ore [177], derivative by Sabidussi [231], derived graphs by
Beineke [18], edge-to-vertex dual by Seshu and Reed [233], covering graph by Kasteleyn
[127] and adjoint by Menon [159].

9.1 Edge Graphs

Definition: Let G(V, E) be a graph with V = {v1 , v2 , . . ., vn } and E = {e1 , e2 , . . ., em }.
The edge graph L(G) of G has the vertex set E and two vertices ei and e j are adjacent in
L(G) if and only if the corresponding edges ei and e j of G are adjacent in G. For example,
in Figure 9.1, L(G) is the edge graph of G. A graph G is an edge graph if it is isomorphic
to the edge graph L(H) of some graph H .

Fig. 9.1

Graph Theory

235

Since isolated vertices do not contribute anything to the study of edge graphs, we assume
that the graphs contain no isolated vertices. Besides this, the graphs under consideration
will be without loops. We have the following observations about edge graphs.
1. A graph G is connected if and only if L(G) is connected.
2. If H is a subgraph of G, then L(H) is a subgraph of L(G).
3. The edges incident at a vertex of G form a maximal complete subgraph of L(G).
4. In G, if e = uv is an edge, then the degree of e in L(G) is the number of edges of G
adjacent to e in G. Clearly, dL(G) (e) = dG (u) + dG (v) 2.
5. For n > 1, Ln (G) = L(Ln1 (G)) and L0 (G) = G.
The following result determines the number of edges in an edge graph.
The number of edges m0 in L(G) when G has degree sequence [di ]n1 is given

Theorem 9.1
by
n

1
m =
2
0

di2

i=1

m.

Proof Let [di ]n1 be the degree sequence of the graph G and let L(G) the edge graph of G,
have m0 edges.
As the degree of the vertex vi in G is di , there are di edges incident on vi . From these di
edges,
 any two are adjacent at vi in G. Hence the number of edges contributed by vi to L(G)

is

di
.
2

Thus,

 
n
di
di (di 1) 1 n 2
m =
=
= (di di )
2
2 i=1
i=1 2
i=1
0

1 n 2 1 n
di 2 di
2 i=1
i=1

1
=
2

i=1

di2

m.

The following observation is immediate.

Theorem 9.2

The edge graph of a graph G is a path if and only if G is a path.

Proof Let G be a graph with n vertices. Assume G is a path Pn . Then L(G) is the path Pn1
with n 1 vertices.

236

Edge Graphs and Eccentricity Sequences

Conversely, let L(G) be a path. Then no vertex of G has degree greater than two. For, if G
has a vertex v of degree greater than two, the edges incident to v form a complete subgraph
of L(G) with at least three vertices. Therefore G is either a cycle or a path. But G cannot be
a cycle, since the edge graph of a cycle is a cycle.
q
We now have the following stronger result.
Theorem 9.3
cycle.

A connected graph is isomorphic to its edge graph if and only if it is a

Proof Let G be a connected graph with n vertices, m edges and with degree sequence
[di]n1 . Let L(G) be the edge graph of G. The number of vertices in L(G) is m. The number of
edges m0 in L(G) is given by
1
m =
2
0

di2

i=1

m.

Clearly, L(G) is connected and L(Cn ) = Cn .

Conversely, let G
= L(G).
Then G and L(G) have the same number of vertices and edges.
1
So, n = m and m =
2

di2

i=1

m.

Therefore, n = m and

di2 = 4m.
i=1

Thus, variance
1 n
{[di ]} = di2
n i=1

1 n
di
n i=1

!2

Because Var = 1 f i x2i

N i

1
f i xi
N
i

!2

and we have f i = 1

1
1
4m 4m2
= 4m 2 (2m)2 =
2 = 4 4 = 0.
n
n
m
m

Therefore the di s are equal and G is regular of degree d , say.

Graph Theory

237

So nd = 2m implies that d =

2m 2m
=
= 2.
n
m

Thus G is a 2-regular connected graph, that is, Cn .

The next result is about the isomorphism of edge graphs.

Theorem 9.4
isomorphic.

If the graphs G1 and G2 are isomorphic, then L(G1 ) and L(G2 ) are also

Proof

Assume ( , ) to be an isomorphism of G1 onto G2 . Then is a bijection of

E(G1 ) onto E(G2 ). We show that is an isomorphism of L(G1 ) to L(G2 ) by showing that
preserves adjacency. Let ei and e j be two adjacent vertices of L(G1 ). So there exists a vertex
v of G1 incident to both ei and e j , and therefore (v) is a vertex incident to both (ei ) and
(e j ). Thus (ei ) and (e j ) are adjacent vertices in L(G2 ).
Let (ei ) and (e j ) be adjacent vertices in L(G2 ). Then they are adjacent edges in G2 and
therefore there exists a vertex v0 of G2 incident to both (ei ) and (e j ). Then 1 (v0 ) is a
vertex of G1 incident to both ei and e j , and thus ei and e j are adjacent vertices of L(G1 ).
Therefore ei and e j are adjacent vertices of L(G1 ) if and only if (ei ) and (e j ) are
adjacent vertices of L(G2 ). Hence is an isomorphism of L(G1 ) onto L(G2 ).
q
The converse of Theorem 9.4 is not true. To see this, consider the graphs K1, 3 and K3
whose edge graphs are K3 . But K1, 3 is not isomorphic to K3 , since there is a vertex of degree
three in K1, 3 while there is no such vertex in K3 . However, it was shown by Whitney [265]
that the converse holds unless one is K1, 3 and the other is K3 . The proof of this result is due
to Jung [123].
Theorem 9.5 Let G and G0 be connected graphs with isomorphic edge graphs. Then G
and G0 are isomorphic unless one is K3 and the other is K1, 3 .

Proof First suppose that n(G) and n(G0 ) are less than or equal to 4. A necessary condition for L(G) and L(G0 ) to be isomorphic is that m(G) = m(G0 ). The only nonisomorphic
connected graphs on at most four vertices are those shown in Figure 9.2.

Fig. 9.2

238

Edge Graphs and Eccentricity Sequences

In Figure 9.2, graphs G4 , G5 and G6 are the three graphs having three edges each. We
know that G4 and G6 have isomorphic edge graphs, namely K3 . The edge graph of G5 is a
path of length 2 and hence L(G5 ) cannot be isomorphic to L(G4 ) or L(G6 ). Further, G7 and
G8 are the only two graphs in the list having four edges each.
Clearly, L(G8 )
= G8 and L(G7 ) is isomorphic to G9 . Thus the edge graphs of G7 and G8
are not isomorphic. No two of the remaining graphs have the same number of edges. Hence
the only non-isomorphic graphs with at most four vertices having isomorphic edge graphs
are G4 and G6 .
Now suppose that either G or G0 , say G, has at least five vertices and that L(G) and L(G0 )
are isomorphic under an isomorphism 1 . So 1 is a bijection from the edge set of G onto
the edge set of G0 .
We now prove that 1 transforms an induced K1, 3 subgraph of G onto a K1, 3 subgraph of
G0 . Let e1 = uv1 , e2 = uv2 and e3 = uv3 be the edges of an induced K1, 3 subgraph of G. As G
has at least five vertices and is connected, there exists an edge e adjacent to only one or all
three of edges e1 , e2 and e3 , as illustrated in Figure 9.3.

Fig. 9.3

Clearly, 1 (e1 ), 1 (e2 ) and 1 (e3 ) form either a K1, 3 subgraph or a triangle in G0 . If 1 (e1 ),
1 (e2 ) and 1 (e3 ) form a triangle in G0 , 1 (e) can be adjacent to precisely two of 1 (e1 ),
1 (e2 ) and 1 (e3 ) (since L(G0 ) is simple), whereas 1 (e) must be adjacent to only one or all
the three. This contradiction shows that {1 (e1 ), 1 (e2 ), 1 (e3 )} is not a triangle in G0 and
therefore forms a K1, 3 in G0 .
It is clear that a similar result holds as well for 11 , since it is an isomorphism on L(G0 )
onto L(G).
Let S(u) denote the star subgraph of G formed by the edges of G incident at a vertex u of
G. We show that 1 maps S(u) onto the unique star subgraph S(u0 ) of G0 .
i. First suppose the degree of u is at least 2. Let f 1 and f 2 be any two edges incident at u.
The edges 1 ( f 1 ) and 1 ( f 2 ) of G0 have an end vertex u0 in common. If f is any other
edge of G incident with u, then 1 ( f ) is incident with u0 , and conversely, for every
edge f 0 of G0 incident with u0 , 11( f 0 ) is incident with u. Thus S(u) in G is mapped to
S(u0 ) in G0 .
ii. Let the degree of u be 1 and e = uv be the unique edge incident with u. As G is
connected and n(G) 5, degree of v must be at least 2 in G, and therefore, by (i), S(v)
is mapped to a star S(v0 ) in G0 . Also 1 (uv) = u0 v0 for some u0 V (G0 ). If the degree of
u0 is greater than 1, by (i), the star at u0 in G0 is transformed by 11 either to the star
at u in G or to the star at v in G. But as the star at v in G is mapped to the star at v0 in

Graph Theory

239

G0 by 1 , 11 should map the star at u0 to the star at u only. As 11 is 1 1, this means

that deg u 2, a contradiction. Therefore, deg u0 = 1 and so S(u) in G is mapped to
S(u0 ) in G0 .

Define : V (G) V (G0 ) by setting (u) = u0 if 1 (S(u)) = S(u0 ). Since S(u) = S(v) only
when u = v (G 6= K2 , G0 6= K2 ), is 1 1. is also onto since, for v0 in G0 , 11 (S(v0 )) = S(v) for
some v V (G), and by the definition of , (v) = v0 . Finally, if uv is an edge of G, then 1 (uv)
belongs to both S(u0 ) and S(v0 ), where 1 (S(u)) = S(u0 ) and 1 (S(v)) = S(v0 ). This means
u0 v0 is an edge of G0 . But u0 = (u) and v0 = (v). Consequently, (u) (v) is an edge of G0 .
If u and v are nonadjacent in G, (u) (v) must be nonadjacent in G0 . Otherwise, (u) (v)
belongs to both S( (u)) and S( (v)) and hence 11 ( (u) (v)) = uv E(G), a contradiction.
Thus G and G0 are isomorphic under .
q
The following result shows that K1, 3 is not an edge graph and thus K1, 3 is of great
significance in studying edge graphs as will be seen in further discussions.
Lemma 9.1

The star K1, 3 is not an edge graph.

Proof Assume that K1, 3 is an edge graph. Then K1, 3 = L(H) for some graph H . Since
K1, 3 has four vertices, therefore H has four edges. Also H is connected. All the connected
graphs with four edges are given in Figure 9.4.

Fig. 9.4
H is neither graph (a) nor (b), because L(C4 ) = C4 and L(K1,3 + x) = K4 x. These are
shown in Figure 9.5.

Fig. 9.5

Thus H is one of the three trees as given in (c), (d) and (e). But the edge graphs of these
trees are the path P4 , the graphs K3 . K2 and K4 , given in Figure 9.6.

240

Edge Graphs and Eccentricity Sequences

Fig. 9.6

This shows that H is none of the trees (c), (d) or (e). Hence it follows that K1, 3 is not an
edge graph.
q
Now, we proceed to give a characterisation of edge graphs which is due to Krausz [140].
Theorem 9.6 (Krausz) A graph G is the edge graph of some graph if and only if the
edges of G can be partitioned into cliques such that no vertex appears in more than two
cliques.

Proof
Necessity Let G be an edge graph of a graph H . Assume without loss of generality that H
has no isolated vertices. Then the edges in the star K1, 3 at each vertex of H induce a clique
of G and every edge of H belongs to the stars of exactly two vertices of H , therefore no
vertex of G is in more than two of these cliques.
Sufficiency Let the edges of G be partitioned into the cliques S1 , S2 , . . ., Sk such that no
vertex of G belongs to more than two of these cliques. We construct H such that L(H) = G.
As isolated vertices of G become isolated edges of H , therefore assume (G) 1. Let v1 ,
v2 , . . ., v` be the vertices of G (if any) that appear in exactly one of Si . The vertices of H
correspond to the set S = {S1 , S2 , . . ., Sk , {v1 }, {v2 }, . . ., {v` }}, with one vertex for each
member of S. Any two of these vertices are adjacent whenever their corresponding sets
intersect. Each vertex of G appears in exactly two sets in S and no two vertices appear in
the same pair of sets. Thus H is a simple graph with one edge for each vertex of G. If
vertices are adjacent in G, they appear together in some Si and the corresponding edges of
H share the vertex corresponding to Si . Hence, G = L(H).
q
Krausz characterisation of edge graphs is close to the definition. Since it characterises
edge graphs by the existence of a special edge partition, it does not directly give an efficient test. This has been improved by Van Rooij and Wilf [227] by describing the structural
criterion for a graph to be an edge graph. Before we take the Van Rooij and Wilf characterisation, we have the following definition.
Definition: An induced subgraph is a subgraph which is maximal on its vertex set. A
triangle T of a graph G is said to be odd, if there is a vertex of G adjacent to an odd number
of vertices of T , otherwise T is said to be even. That is, T is odd if |V (T ) N(v)| is odd, for
some v V (G), and T is even if |V (T ) N(v)| is even, for every v V (G). An induced copy
of K4 e is a double triangle and clearly has two triangles with a common edge.

Graph Theory

241

The following is the Van Rooij and Wilf characterisation.

Theorem 9.7 A graph G is the edge graph of some graph if and only if G does not
contain an induced subgraph K1, 3 and no double triangle of G has two odd triangles.

Proof
Necessity Let G = L(H). Clearly, G does not contain an induced subgraph K1, 3 , since
K1, 3 itself is not an edge graph. Now we observe that the vertices of a double triangle in G
correspond to the edges of a K1, 3 + e in H . In particular, one of these double triangles in G
is generated by a triangle in H . Obviously a triangle in G generated by a triangle in H is
even, since an edge incident to a triangle in H intersects exactly two edges of the triangle
in H (Fig. 9.7).

Fig. 9.7

Sufficiency Let G not contain an induced subgraph K1, 3 and let no double triangle of G
have two odd triangles. Assume G is connected, for otherwise, we apply the construction
to each component. In case G is K1, 3 free and has a double triangle with both triangles
even, then G is one of the graphs given in Figure 9.8.

Fig. 9.8

Thus we consider the case when every double triangle of G has exactly one odd triangle.
To prove the result, it suffices by Theorem 9.6, to partition E(G) into cliques that cover
each vertex at most twice. Now, let S1 , S2 , . . ., Sk be the maximal cliques of G that are not
even triangles and let T1 , T2 , . . ., T` be the edges that belong to one even triangle and no
odd triangle. We claim that B = {Si } {T j } partitions E(G) into cliques using each vertex at
most twice.
Now, every edge appears in a maximal clique, but every triangle in a clique with more
than three vertices is odd. Therefore T j is not in any clique Si . Also Si and Si0 have no
common edge, because G has no double triangles with both triangles odd. Thus the cliques

242

Edge Graphs and Eccentricity Sequences

in B are edge-disjoint. If e E(G), then e belongs to some Si unless the only maximal clique
containing e is an even triangle. In this case e is a T j , since the double triangles do not have
both triangles even (Fig. 9.9).

Fig. 9.9

We now show that each v G appears at most twice in B. Assume v belongs to B1 , B2 ,

B3 B. Edge-disjointness implies that v has neighbours x, y, z with each belonging to only
one of B1 , B2 , B3 . Since G has no induced K1,3, assume that xy is an edge. Now by edgedisjointness, the triangle vxy does not belong to a member of B. Therefore vxy is an even
triangle. Thus z has exactly one other edge to vxy, say zx, while zy is not an edge. But now
the same argument shows that zvx is an even triangle and we have a double triangle with
both triangles even. This contradicts our supposition and hence each v G appears at most
twice in B.
q
The next characterisation due to Beineke [149] displays those subgraphs which are
not present in edge graphs. These subgraphs other than K1, 3 are vertex-minimal K1, 3 -free
graphs containing a double triangle with both triangles odd. Each such graph has a double
triangle and one or two additional vertices that make the triangles odd by having one or
three neighbours in the triangle.
Theorem 9.8 A graph G is an edge graph of some graph if and only if G does not contain
an induced subgraph of any one of the graphs in Figure 9.10.

Fig. 9.10

Graph Theory

243

Proof
Necessity Let G be the edge graph of some graph H so that G = L(H). Then by Theorem
9.6, the edges of G can be partitioned into cliques such that every vertex appears in at most
two cliques. We observe that none of these nine graphs have such a partition. Since every
induced subgraph of an edge graph is itself an edge graph, G does not contain an induced
subgraph of any one of the nine graphs, in Figure 9.10.
Sufficiency Let G not contain an induced subgraph of any one of these nine graphs. We
prove that no double triangle of G has two odd triangles. Assume to the contrary that G has
a double triangle both of which are odd. Let these triangles be abc and abd with c and d non
adjacent. We discuss two cases, one in which there is a vertex v adjacent to an odd number
of vertices of both odd triangles and second when there is no such vertex.
Case 1 Assume there is a vertex v which is adjacent to an odd number of vertices in each
of the triangles abc and abd . Now two possibilities arise; either v is adjacent to exactly one
vertex of each of these triangles, or it is adjacent to more than one vertex of one of them.
If v is adjacent to exactly one vertex of each of these triangles, then either v is adjacent to a
or b giving G1 , or to both c and d giving G2 . If v is adjacent to more than one vertex of one
of the triangles, then v is adjacent to all four vertices of the two triangles, giving G3 as an
induced subgraph of G (Fig. 9.11).

Fig. 9.11
Case 2 Now, let there be no vertex adjacent to an odd number of vertices of both triangles.
Assume that the vertex u is adjacent to an odd number of vertices of triangle abc and
the vertex v is adjacent to an odd number of vertices of triangle abd . We consider three
subcases.
Case 2.1

u is adjacent to exactly one vertex of abc and v is adjacent to exactly one vertex

of abd .
Case 2.2 One of u or v is adjacent to all three vertices of its triangle and the other to only
one vertex of its triangle.
Case 2.3
abd .

u is adjacent to all three vertices of abc and v is adjacent to all three vertices of

We observe that if u or v is adjacent to a or b, then it is also adjacent to c or to d , since

otherwise G1 is an induced subgraph. Also, neither u nor v is adjacent to both c and d , since
otherwise G2 or G3 is induced.

244

Edge Graphs and Eccentricity Sequences

Let uc, vd G. Then for uv G, G4 is induced, and for uv / G, G7 is induced.

Now, let ub, vd G. Then it follows from the above observations that ud G while vc
/ G.
Therefore for uv / G, the vertices a, d , u, v induce G1 and for uv G, the vertices a, b, c, d ,
u, v induce G8 . Next, let ub, va G, then clearly ud , vc G. So, when uv
/ G, G8 is induced,
and when uv G, G2 is induced. Finally, let ub, vb G, then again ud , vc G. Therefore,
when uv G, G9 is induced, and when uv / G, G1 is induced (Fig. 9.12, Case 2.1).
Case 2.1

Case 2.2 Let ua, ub, uc G. If ud G, then G3 is induced. Take ud

/ G. Then either vd G
or vb G. If vd G, then for uv G, G2 is induced, and for uv / G, G5 is induced. If vb G,
then G3 or G1 is induced depending on whether or not v is adjacent to both c and u (Fig.

9.12, Case 2.2).

If ud , vc or uv G, then G3 is induced. The only other possibility gives G6 (Fig.
9.12, Case 2.3).
q

Case 2.3

(Case 2.1)

(Case 2.2)

Graph Theory

245

(Case 2.3)
Fig. 9.12

The following result due to Chartrand [53] characterises the edge graphs of a tree.
Theorem 9.9 (Chartrand [53]) A graph is the edge graph of a tree if and only if it is a
connected block graph in which each cut vertex is on exactly two blocks.

Proof
Necessity Let T be any tree and let G = L(T ). Then G is also B(T ) since the edges and
blocks of a tree coincide. Each cut vertex w of G corresponds to a bridge uv of T and is on
exactly those two blocks of G which correspond to the stars at u and v.
Sufficiency Let G be a block graph in which each cut vertex is on exactly two blocks.
Since each block of a block graph is complete, there exists a graph H such that L(H) = G,
by Theorem 9.6. If G = K3 , we can take H = K1, 3 . If G is any other block graph, then we
show that H is a tree. Assume H is not a tree, so that it contains a cycle. If H is itself a cycle,
then by Theorem 9.3, L(H) = H , but the only cycle which is a block graph is K3 , a case not
under consideration. Thus H properly contains a cycle, implying that H has a cycle Z and
an edge e adjacent to two edges of Z , but not adjacent to some edge f in Z . The vertices e
and f of L(H) lie on a cycle of L(H) and they are not adjacent. This contradicts the fact that
L(H) is a block graph. Hence H is a tree.
q
Consider the block graph G of Figure 9.13(a) in which each cut vertex lies on just two
blocks. Figure 9.13(b) shows the tree T of which G is the edge graph, is constructed by first
forming the block graph B(G) and then adding new vertices for the non-cut vertices of G,
and the edges joining each block with its non-cut vertices.
The edge graphs of complete graphs were independently characterised by Chang [47]
and Hoffman [116, 117], while the edge graphs of complete bipartite graphs were characterised by Moon [163] and Hoffman [118].

Fig. 9.13

246

Edge Graphs and Eccentricity Sequences

9.2 Edge Graphs and Traversability

In this section, we study Eulerian and Hamiltonian property in edge graphs. We start with
the following result.
Theorem 9.10

If G is Eulerian, then L(G) is both Eulerian and Hamiltonian.

Proof Let G be Eulerian and let {e1 , e2 , . . ., em } be the edge sequence of an Euler line in
G. Let the edge ei in G be represented by the vertex vi in L(G), 1 i m. Then v1 v2 . . .vm v1
is a Hamiltonian cycle of L(G). Now, if e = ui u j E(G) and the vertex v in L(G) represents
the edge e, then dL(G) (v) = dG (ui ) + dG (u j ) 2, which is obviously even and greater than or
equal to two, since both dG (ui ) and dG (u j ) are even (and 2). Thus in L(G) every vertex is
of even degree ( 2). Hence L(G) is Eulerian.
q
The converse of Theorem 9.11 is not true. To see this, consider the graph G shown in
Figure 9.14. Clearly L(G) is both Eulerian and Hamiltonian, but G is not Eulerian.

Fig. 9.14

Definition: A dominating walk of a graph G is a closed walk W in G (which can be just

a single vertex) such that every edge of G not in W is incident with W . For example, the
walk v1 v2 v3 v4 in the graph of Figure 9.15 is a dominating walk.

Fig. 9.15

The following characterisation of graphs that contain Hamiltonian edge graphs is due to
Harary and Nash-Williams [108].

Graph Theory

247

Theorem 9.11 The edge graph of a graph G with at least three edges is Hamiltonian if
and only if G has a dominating walk.

Proof Let W be a dominating walk of G which is represented by the edge sequence

{e1 , e2 , . . ., ek }. Let e1 and e2 be incident at v1 . Replace the subsequence {e1 , e2 } by the
sequence {e1 , e11 , e12 , . . ., e1r1 , e2 }, where e11 , e12 , . . ., e1r1 are the edges other than e1
and e2 incident at v1 . Continuing this process for all subsequences {ei , ei+1 }, 1 i k with
ek+1 = e1 , we obtain a sequence of edges e1 e11 e12 . . .e1r1 e2 e21 e22 . . .e2r2 e3 . . .ek ek1 ek2 . . .ekrk e1 .
This clearly gives the Hamiltonian cycle u1 u11 u12 . . .u1r1 u2 u21 u22 . . .u2r2 u3 . . .uk uk1 uk2 . . .ukrk
u1 in L(G), with u1 being the vertex of L(G) that corresponds to the edge e1 of G, and so on.
Conversely, let L(G) contain a Hamiltonian cycle C = u1 u2 . . .um u1 and let ei be the edge
of G corresponding to the vertex ui of L(G). Let W0 be the edge sequence e1 e2 . . .em e1 . We
delete edges from W0 in succession in the following way. If ei e j ek are the first three distinct
consecutive edges of W0 that have a common vertex, then delete e j , and let W00 = W0 e j =
e1 e2 . . .ei ek . . .em e1 . Now starting with W00 , apply the same process as is applied in W, to get
W0. Continue in this way, till no such three consecutive edges exist. Clearly, the resulting
subsequence of W0 is a dominating walk or a pair of adjacent edges incident at a vertex,
say v0 . In the later case, all the edges of G are incident at v0 and hence v0 is the dominating
walk of G.
q
The following results are simple consequences of Theorem 9.11.
Corollary 9.1

The edge graph of a Hamiltonian graph is Hamiltonian.

Proof Let G be a Hamiltonian graph with Hamiltonian cycle C. Then C is a dominating

walk of G, and hence, L(G) is Hamiltonian.
q
We note that the converse of Corollary 9.1 is not true in general. To see this, consider
the graph G as shown in Figure 9.16. Clearly L(G) is Hamiltonian but G is not.

Fig. 9.16

Corollary 9.2 If G is a connected graph and each edge of G belongs to a triangle, then
L(G) is Hamiltonian.

Proof

This follows from Theorem 9.11.

248

Edge Graphs and Eccentricity Sequences

The following result is due to Chartrand and Wall [55].

Theorem 9.12 If G is connected and (G) 3, then
L2 (G) is Hamiltonian.

Proof Since (G) 3, each vertex of L(G) belongs to a clique of size at least three
and hence each edge of L(G) belongs to a triangle. Then the result follows by applying
Corollary 9.2.
q
The next result is due to Nebesky [170].
Theorem 9.13
tonian.

If G is a connected graph with at least three vertices, then L(G2 ) is Hamil-

Proof Since G is a connected graph with at least three vertices, every edge of G2 belongs
to a triangle. Hence by Corollary 9.2, L(G2 ) is Hamiltonian.
q
Theorem 9.14 Let G be a connected graph in which every edge belongs to a triangle.
If e1 and e2 are edges of G such that G {e1 , e2 } is connected, then there exists a spanning
walk of G with e1 and e2 as its initial and terminal edges.

Proof Consider the longest walk W of G with e1 and e2 as its initial and terminal edges.
Then proceed as in Theorem 9.11.
q
The following result is due to Jaeger [121].
Theorem 9.15

The edge graph of a 4-connected graph is Hamiltonian.

Proof Let G be a 4-edge connected graph. By Theorem 9.11, it suffices to show that G
contains a spanning Eulerian subgraph.
Now, G contains two edge-disjoint spanning trees T1 and T2 . Let S be the set of vertices
of odd degree in T1 . Then |S| is even. Let |S| = 2k, k 1. By Theorem 9.12, there exists a set
of k pairwise edge-disjoint paths {P1 , P2 , . . ., Pk } in T2 with the property stated in Theorem
9.12. Then G0 = T1U (P1 P2 . . . Pk ) is a connected spanning subgraph of G in which each
vertex is of even degree. Hence G0 is a spanning Eulerian subgraph of G.
q
Let every edge in a graph G be subdivided and let S(G) be the subdivision graph. If the
graph obtained from G by inserting n new vertices of degree two into every edge of G be
denoted by Sn (G) and taking S(G) = S1 (G), we define Ln (G) = L(Sn1 (G)). We see that in
general Ln (G)  Ln (G) (Fig. 9.17).

Graph Theory

249

Fig. 9.17

An improvement of Theorem 9.13 is seen in the following result of Harary and NashWilliams [108].
Theorem 9.16 A sufficient condition for L2 (G) to be Hamiltonian is that G be Hamiltonian and a necessary condition is that L(G) be Hamiltonian.
Now, we have the following consequence.
Corollary 9.3

A graph G is Eulerian if and only if L3 (G) is Hamiltonian.

The following result is due to Chartrand [48].

Theorem 9.17 If G is a non-trivial connected graph with n vertices which is not a path,
then Lk (G) is Hamiltonian for all k n 3.

9.3 Total Graphs

Let G(V, E) be a graph. The total graph T (G) of G has vertex set V E and two vertices of
T (G) are adjacent if and only if one of the following is true.
i. the vertices are vi , v j V and vi v j is an edge in E .
ii. one vertex is v V and the other e E and the edge e of G is incident with the vertex
v of G.
iii. the edges are ei , e j E and the edges ei and e j have a vertex in common in G.
Example

The total graph of a graph G is shown in Figure 9.18.

Fig. 9.18

250

Edge Graphs and Eccentricity Sequences

It can be seen that G and L(G) are induced subgraphs of T (G) and that the remaining
edges of T (G) form a graph homeomorphic to G.

9.4 Eccentricity Sequences and Sets

The concept of eccentricity has been introduced in chapter 1 and has been further discussed
in chapter 4. Now, we study eccentricity sequences in graphs.
Definition: A positive sequence [ei ]n1 is called an eccentricity sequence if it is an eccentricity sequence of some graph. The graph is said to realise the sequence. A set of positive
integers is called an eccentricity set if it is an eccentricity set of some graph. The graph is
said to realise the set. (The set of distinct eccentricities in a graph is called the eccentricity
set of that graph.)
When eccentricity set is written in the increasing order {e1 , e2 , . . ., ek } with e1 < e2 <
n
n
n
. . . < ek , the eccentricity sequence is then expressed as [e11 , e22 , . . ., ek k ], where n1 , n2 , . . .,
nk are respectively the number of occurrences of e1 , e2 , . . ., ek , the sequence n1 , n2 , . . ., nk is
called the eccentricity frequency sequence of the graph.
Here it can be noted that e1 = r (radius) and ek = d (diameter) of the graph.
Therefore, r d 2r gives e1 ek 2e1 , which is a necessary condition for a positive
sequence to be an eccentricity sequence.
Now, we have the following observation.
Theorem 9.18

If uv is an edge of a connected graph G, then |e(u) e(v)| 1.

Proof Let w be an eccentric vertex of u (i.e., w is the farthest vertex from u). Then by the
triangle inequality for the metric d (distance), we have
d(u, w) d(u, v) + d(v, w)

so that e(u) d(u, v) + d(v, w).

(9.18.1)

But u and v are adjacent, therefore d(u, v) = 1.

Also, e(v) d(v, w) so that d(v, w) e(v).
Thus, from (9.18.1) we have
e(u) 1 + d(v, w) so that e(u) 1 + e(v).

Therefore,

e(u) e(v) 1.

Similarly, by considering an eccentric vertex of v, we have

(9.18.2)

Graph Theory

251

e(v) e(u) 1.

(9.18.3)

From (9.18.2) and (9.18.3) it follows that

|e(u) e(v)| 1.

The above result shows that the eccentricities of two adjacent vertices are either
equal or differ by 1 as |e(u) e(v)| 1 gives |e(u) e(v)| = 0 or |e(u) e(v)| = 1.
Note

An important consequence of Theorem 9.18 is as follows.

Corollary 9.4 If u0 u1 u2 . . .um is a path in a connected graph and e(u0 ) < e(um ) and k is
any integer such that e(u0 ) < k < e(um ), then there exists an integer j (0 j m) such that
e(u j ) = k.

Proof We know the difference of eccentricities of any two adjacent vertices along the
path u0 u1 u2 . . .um is always less or equal to 1. Therefore every integer between e(u0 ) and
e(um ) occurs as the eccentricity of some vertex in this path. This can also be seen in the
following way.
Assume, e(u0 ) < e(u1 ) < . . . < e(u j1),
and let j = 1 + max{i : e(ui ) < k}, that is, j 1 = max{i : e(ui ) < k}.
Thus, e(u j1) < k.
Therefore, |e(u j ) e(u j1)| 1 gives
e(u j ) e(u j1) + 1 < k + 1,

so that e(u j ) k.

(9.4.1)

But by the choice of j , we have e(u j ) k.

Hence from (9.4.1) and (9.4.2), we get e(u j ) = k.

(9.4.2)
q

The following necessary condition for a positive sequence to be an eccentricity sequence

is due to Lesniak [146].
Theorem 9.19 (Lesniak)
eccentric sequence then

If a non decreasing sequence [ei ]n1 of positive integers is an

i. 2e1 n,
ii. en min{n 1, 2e1} and
iii. for every integer k such that e1 < k en , there exists an integer i (2 i n 1) such
that ei = ei+1 = k.

252

Edge Graphs and Eccentricity Sequences

Proof
i. Let the vertices of G be labelled as v1 , v2 , . . ., vn such that e(vi ) = ei . Then G has a
spanning tree T which preserves the distance from v1 . This gives
eG (v1 ) = eT (v1 ) and eG (vi ) eT (vi ),

for 2 i n (since removal of edges cannot reduce distances)

Thus, if [a1 , a2 , . . ., an ] is the eccentricity sequence of T , we have a1 = e1 . So it is
enough to prove that 2a1 n. We prove this for any tree T .
Now let T be any tree with eccentricity sequence [a1 , a2 , . . ., an ].
If n = 2, then a1 = a2 = 1, and the result is true. So assume n 3.
Let u be a central vertex of T . Then e(u) = a1 . Also u is a cut vertex of T .
Suppose a1 = e(u)

n+1
.
2

[2a1 n + 1]

Since an eccentric vertex u of u should lie in a component of T u, there is at least

n+1

.
one component C of T u with |V (C)|
2
Now, let v be the vertex adjacent to u in C. Then for any vertex w in C, we have
d(v, w) = d(u, w) 1. So, d(v, w) < e(u), because d(u, w) e(u) and so
d(u, w) 1 < e(u).

For every vertex w in V (T ) V (C), we have d(v, w) d(u, w) = 1,

so that d(v, w) = d(u, w) + 1.
Total vertices in T is n, |V (C)|


n+1
n
2

n1
.
2

That is, |V (T ) V (C)|

Thus, d(v, w)

n+1
, therefore number of vertices in V (T ) V (C)
2

n1
n1
n3
. Therefore, d(u, w)
1 =
.
2
2
2

n3
n1
+1 =
. So, d(v, w) < e(u).
2
2

Hence for all vertices w, we have d(v, w) < e(u), and thus e(v) < e(u), so that e(v) <
a1 , which is a contradiction as a1 is the least eccentricity of a vertex of T . Thus,

Graph Theory
a1

253

n+1
n
is wrong and so, a1 .
2
2

ii. The maximum distance possible in an nvertex graph is n 1. So, en n 1. Also,

en 2e1 , [r d 2r, and here d = en , r = e1 ]. Hence, en min{(n 1), 2e1}.
iii. We have to prove that each integer between e1 and en (en inclusive) occurs at least
twice in the sequence. Let u1 be the central vertex and uk be the peripheral vertex of
G. Then e(u1 ) = e1 and e(uk ) = en . Since G is connected, there exists a u1 uk path.
Now by Corollary 9.4, if k is any integer between e1 and ek , there exists a vertex u j in
this path with e(u j ) = k. This gives the existence of a vertex whose eccentricity is k.
If e(w) > e1 , we show there is a vertex u other than w such that e(u) = e(w). Let w be
an eccentric vertex of w, that is, d(w, w) = e(w) = k, say. As we have assumed that u1
is the central vertex of G, let P = u1 . . .um (um = w) be a u1 w distance path in G. Since
e(u1 ) = e1 < e(w) = d(w, w) e(w), applying Corollary 9.4, there is a vertex u j in this
path such that e(u j ) = k. But d(w, u j ) m 1 = d(u1 , w) e(u1 ) = e1 < e(w) = d(w, w).
Therefore, d(w, u j ) < d(w w). Thus, u j 6= w and the result is proved.
q
The following characterisation of eccentricity sequences of trees is again due to Lesniak
[146].
Theorem 9.20 (Lesniak) A non-decreasing sequence [ei ]n1 of positive integers is the
eccentric sequence of a tree if and only if
i. For each integer k with e1 < k en , we have
ei = ei+1 = k, for some i, 2 i n 1,

ii. Either e1 =

en
e +1
and e1 6= e2 , or e1 = n
, e1 = e2 and e2 6= e3 .
2
2

Proof
Necessity Let the nondecreasing sequence [ei ]n1 of positive integers be the eccentric sequence of a tree. Then (i) follows from condition (iii) of the previous result.
Let r be the radius and d the diameter of the tree, so that e1 = r and en = d . Since a tree
is either unicentral or bicentral, we have d = 2r for unicentral,
and d = 2r 1 for bicentral.
In case the tree is unicentral, then the eccentricity of the center v1 is e1 and e1 6= e2 .
e
Thus, en = 2e1 which implies that e1 = n and e1 6= e2 .
2
In case the tree is bicentral, then e1 = e2 and d = 2r 1 gives en = 2e1 1, so that
e1 =

en + 1
with e2 6= e3 .
2

Sufficiency Let the nondecreasing sequence [ei ]n1 of positive integers satisfy conditions
(i) and (ii). We construct a tree with eccentric sequence [ei ]n1 in the following way.

254

Edge Graphs and Eccentricity Sequences

Let P be a path of length en = d . Then eccentric sequence of P is


2 
2


d d
d
d d 2
S1 = ,
+1 ,
+ 2 , . . .,
+
,
2 2
2
2 2

or

S2 =

d+1
2

if d is even,

2 
2

2
d +1
d +1 d +1
,
+ 1 , . . .,
+
1 ,
2
2
2

if d is odd,

d
2

that is, S1 = r, (r + 1)2 , (r + 2)2, . . ., (r + r)2 , where r = ,

or S2 = r2 , (r + 1)2 , (r + 2)2 , . . ., (r + r 1)2, where r =

d +1
,
2

where powers denote repetition of eccentricity.

Let the given sequence [ei ]n1 be written in power notation = ri1 (r + 1)i2 . . .d ik , where
i1 = 1 or 2, according as d is even or odd. If i j > 2 for any j , 1 < j k, we attach i j 2 vertices
to any vertex with eccentricity r + j 2 in the path P. This does not alter the eccentricities
of the vertices of P and the resulting tree T has eccentric sequence [ei ]n1 .
q
Example Construct a tree with eccentric sequence [42 , 54 , 63 , 74 ].
First draw a path P say u0 u1 u2 u3 u4 u5 u6 u7 of length 7. Then the eccentricities of these
vertices are 7, 6, 5, 4, 4, 5, 6, 7.

Fig. 9.19

To get two more vertices of eccentricities 5, attach a new vertex each to u3 and u4 . Let
these new vertices be u8 and u9 . (Here i2 = 4 > 2, and i2 2 = 4 2 = 2). So 2 vertices one
each are attached to the vertices of eccentricities r + j 2 = r + 2 2 = 4 + 2 2 = 4, i.e., u3
and u4 ). Now i3 = 3 > 2 and i3 2 = 3 2 = 1. So one vertex is to be attached to the vertex of
eccentricity r + j 2 = 4 + 3 2 = 5, i.e., the vertex u2 or u8 or u9 or u5 . Let this new vertex
u10 be attached to u9 say. Now i4 = 4 > 2 and i4 2 = 4 2 = 2. So two new vertices are to
be attached, one each among the vertices with eccentricities r + j 2 = 4 + 4 2 = 6, i.e., to
the vertices u1 , u6 , u10 . Let these new vertices be u11 attached to u10 and u12 attached to u1 .
The resulting tree is shown in Figure 9.19.
Remark

Clearly, there are many trees realising this eccentric sequence.

Graph Theory

255

Neighbourhood Let v be any vertex of a connected graph G. The ith neighbourhood of

v denoted by Ni (v) is the set of all those vertices in V whose distance from v is i.

i.e., Ni (v) = {u V : d(v, u) = i}.

We denote N1 (v) by N(v) and call it the neighbourhood of v.
Example

Consider the graph in Figure 9.20.

We have N1 (v) = N(v) = {u1 , u2 , u3 , u4 , u5 , u6 }.

Fig. 9.20

We have the following observations.

Lemma 9.2
eccentricity.

Any two vertices with the same neighbourhood in a graph have the same

Proof Let u and v be two vertices with the same neighbourhood. So N(u) = N(v). Therefore the path lengths from u and v to the other vertices of the graph are equal. Clearly, u
and v are not adjacent.
q
Lemma 9.3 If u and v are adjacent in G and N(u) {v} = N(v) {u}, then u and v have
the same eccentricity in G.

Proof Let G be a graph in which uv = e and N(u) {v} = N(v) {u}.

Let H = G e. Then u and v are not adjacent in H so that u and v have the same neighbourhood in H .
Therefore, e(u|H) = e(v|H) where (e(u|H) means eccentricity of vertex u in graph H ). If
e(u|G) = 1, then e(v|G) = 1 also. If not, then e(u|G) = e(u|H) = e(v|H) = e(v|G).
q
Definition: Let v be a vertex of a graph G and let H be a graph obtained from G v by
adding edge to each vertex of a new graph Kp (or K p ) to every vertex of G v to which v
was adjacent in G. Then H is said to be obtained from G by replacing v by Kp (or K p ). This
operation is illustrated in Figure 9.21.

256

Edge Graphs and Eccentricity Sequences

Fig. 9.21

The operation of replacing v in G by K p is called multiplication of the vertex in G. The

operation of replacing v in G by Kp is called the linked multiplication of v in G. Lesniak
observed that multiplication or linked multiplication of one or more vertices of a graph is
an operation preserving the eccentricity set of the graph.
Lemma 9.4 If H is the graph obtained by replacing a vertex v of a graph G with diameter
greater than one, by a Kp or K p (for any positive integer p), then G and H have the same
eccentricity sets.

Proof

Since v and any vertex of K p have the same neighbourhood in H ,

e(u|H) = e(v|G), for every u K p .

For replacement by Kp , we have for any two adjacent vertices ui and u j in Kp ,

N(ui ) {u j } = N(u j ) {ui }.

Therefore, e(ui |H) = e(u j |H). So, e(v|H) = e(u|G), for every u Kp .
Thus, e(u|H) = e(v|G), for every u Kp (K p ) and obviously
e(w|H) = e(w|G), for all other vertices.

The above result is illustrated in Figure 9.22, where we choose K2 and K 2

Fig. 9.22

The next gives the necessary and sufficient conditions for an eccentricity sequence of a
graph, due to Lesniak [146].

Graph Theory

257

Theorem 9.21 (Lesniak) A nondecreasing sequence of positive integers [er11 , er22 , ..., erkk ]
is an eccentricity sequence if and only if some of its subsequence [es11 , es22 , . . ., eskk ] with
si ri and no si = 0 for 1 i k, is an eccentricity sequence.

Proof
Necessity

Since every sequence is its own subsequence, necessity follows.

Sufficiency Let [es11 , es22 , ..., eskk ] be the eccentricity sequence of a graph G. Let vi be a vertex
of G with e(vi ) = ei , 1 i k. Let n1 = r1 s1 + 1 so that r1 = n1 + s1 1.
Now let G1 be obtained from G by replacing v1 by Kn1 or K n1 . Then every one of the
vertices of this Kn1 or K n1 has eccentricity e(v1 |G) and the eccentricities of the vertices of G
are unaltered in G1 . Thus G1 has eccentricity sequence [er11 , es22 , . . ., eskk ].
Let G2 be obtained from G1 by replacing v2 by Kn2 or K n2 . Then by similar argument, G2
has eccentricity sequence [er11 , er22 , . . ., eskk ].
Proceeding in this way, by successively replacing v3 , v4 , . . ., vk by Kn3 (K n3 ), Kn4 (K n4 ),
r
r
r
. . ., Knk (K nk ), we get a graph Gk with eccentricity sequence [e11 , e22 , . . ., ekk ].
q
Remarks

1. It is assumed that ek > 1.

2. The construction of Gk is not unique.
3. This result keeps unsolved the problem of characterising minimal eccentricity sequences, that is, those eccentricity sequences which have no proper eccentric subsequences.
The next result characterises eccentricity sets and is due to Behzad and Simpson [17].
Theorem 9.22 A non-empty set S = {e1 , e2 , . . ., ek } of positive integers arranged in increasing order is an eccentricity set if and only if k e1 + 1 and ei+1 = ei + 1 for 1 i k 1.

Proof
Necessity Let S be an eccentricity set. Then by (iii) of Theorem 4.24, ei+1 = ei + 1 for
each i, 1 i k 1. This gives ek = e1 + k 1. Since ek 2e1 , we get k e1 + 1.
Sufficiency If e1 = 1, then k = 1 or 2 and S = {1} or {1, 2}. In this case, K2 and K1,n realise
the sets.
For e1 > 1, let G be the graph obtained by identifying a vertex of a cycle C2e1 with an
end vertex of a path Pk . Let e1 k + 1 = d . Then d 0, and the eccentricity sequence of G is
easily verified to be [e2d+1
, (e1 + 1)3 , (e1 + 2)3 , . . ., (e1 + k 2)3 , (e1 + k 1)3 ]. Hence S is the
1
eccentricity set.
q

258

Edge Graphs and Eccentricity Sequences

9.5 Distance Degree Regular and Distance Regular Graphs

Let G be a connected graph and let v be any vertex of G. Let e be the eccentricity
of

vertex v and N j (v) be the j th neighbourhood of v. Assume, n j (v) = N j (v), for 0 j
e. The sequence D(G, v) = [n0 (v), n1 (v), . . ., ne (v)] is called the distance degree sequence
(DDS) of v in G. If all the vertices of G have the same distance degree sequence D(G) =
[n0, n1 , . . ., nd ], then G is said to be distance degree regular.
If G is not distance degree regular, the n vectors D(G, vi ), 1 i n, arranged lexicographically in an array with variable row sizes is called the distance degree array of G
(DDA(G)).
The distance of vertex v is defined by
e(v)

D(v) =

n j (v).
j=1

Let Di be the distance of the vertex vi . The sequence DS(G) = [D1 , D2 , . . ., Dn ] in nondecreasing order is called the distance sequence of the graph.
A vertex v with minimum distance D(v) is called a median of G and the subgraph induced
by the set M of median vertices of G is called the median subgraph of G.
Clearly, if G is distance degree regular, then all vertices in G have the same eccentricity,
and G is a self-centered graph. Also, n0 = 1 and n1 = |N1 (v)| for every v G, so that G is
n1 -regular.
Randic [214] conjectured that two trees are isomorphic if and only if they have same
DDA and Slater [187] disproved this giving the counter example as shown in Figure 9.23.

Fig. 9.23

Graph Theory

259

Definition: Let G be a connected graph with diameter d and let k = b0 , b1 , . . ., bd1 ;

l = c1 , c2 , . . ., cd be 2d non-negative integers. Then G is said to be distance regular (DR) if
for every pair of vertices u, v in G with d(u, v) = j , we have, (i) the number of vertices in
N j1(v) adjacent to u is c j , 1 j d and (ii) the number of vertices in N j+1(v) adjacent to u
is b j , 0 j d 1.
The sequence [b0 , b1 , . . ., bd1 , c1 , c2 , . . ., cd ] is called the intersection array of G.
Clearly, DR graphs are k-regular and self-centered. The examples of distance regular
graphs are Kn , Kn, n and the cubes Qn .
A graph G is said to be strongly regular (SR) with parameters (n, k, , ) if it is a kregular graph of order n in which every pair of adjacent vertices are mutually adjacent to
vertices and every pair of non-adjacent vertices are mutually adjacent to vertices.
The Petersen graph is strongly regular with parameters (10, 3, 0, 1), that is n = 10, k = 3,
= 0, = 1 (Fig. 9.24).

Fig. 9.24

9.6 Isometry
The concept of isometry as in Chartrand and Stewart [52] is as follows.
Let G1 and G2 be connected graphs with vertex sets V1 and V2 respectively. Then G2 is
said to be isometric from G1 if for each v V1 , there is a one-one map v : V1 V2 such that
v preserves distances from v, that is dG2 (u, v) = dG1 (v (v), v (u)) for every u V1 .
Two graphs G1 and G2 are said to be isometric if they are isometric from each other.
Example

Consider the graphs shown in Figure 9.25, we have

1 = 4 = 5 (1 a, 2 b, 3 c, 4 d, 5 e),
2 = (2 e, 1 a, 3 d, 4 c, 5 b), 3 = (3 e, 4 a, 2 d, 1 b, 5 c).

Here G2 is isometric from G1 .

260

Edge Graphs and Eccentricity Sequences

Fig. 9.25

Isometry between graphs as defined above does not imply isomorphism (Fig.
9.26). A pair of isometric graphs may even have same degree sequence and yet be nonisomorphic (Fig. 9.27).
Remarks

Fig. 9.26

Fig. 9.27

We now have the following results.

Theorem 9.23 If G1 and G2 are k-regular graphs of order n, where k n 1/2, then G1
and G2 are isometric.

Proof Since G1 is a k-regular graph with k n 1/2, d(G1 ) 2.

Let u V (G1 ) and v V (G2 ) be any two vertices and define
u : V (G1 ) V (G2 ) by u (u) = v.

For i = 1, 2, . . ., k, let ui N1 (u) and vi N1 (v) and define u (ui ) = vi .

Graph Theory

261

For i = k + 1, . . ., n 1, let ui N2 (u) and vi N2 (v) and again let u (ui ) = vi .

The neighbourhoods are in the appropriate graphs. Then u is an isometry of G2 from
G1 at u. Since u and v are arbitrary, it is easily seen that G2 is isometric from G1 , and G1 is
isometric from G2 .
q
Theorem 9.24 A necessary condition for two graphs to be isometric is that they have
the same degree set and the same eccentricity set.

Proof Let G1 = (V1 , E1 ) and G2 = (V2 , E2 ) be isometric graphs. As G2 is isometric from

G1 , let v be the one-one mapping from V (G1 ) to V (G2 ). Therefore, d(v|G1 ) = d(v (v)|G2 ).
Also v has the property of preserving distance, therefore e(v|G1 ) = e(v (v)|G2 ). So the
eccentricity set of G1 is included in the eccentricity set of G2 .
Again, as G1 is isometric from G2 , therefore, the degree set and eccentricity set of G2 are
included respectively in the degree set and eccentricity set of G1 .
Hence the degree sets are equal in G1 and G2 and the eccentricity sets are equal in G1
and G2 .
q

9.7 Exercises
1. Show that the edge graph of K1, n is Kn .
2. Show that the edge graph of K5 is the complement of the Petersen graph.
3. Show that if L(G) is connected and regular, then either G is regular or G is a bipartite
graph in which vertices of the same partite set have the same degree.
4. If G is k-edge-connected, then prove that L(G) is k-connected and (2k 2)-edgeconnected.
5. Show that the graph L2 (G) is Hamiltonian if and only if G has a closed spanning walk.
6. Show that the graph L2 (G) is Hamiltonian if and only if there is a closed walk in G
which includes at least one vertex incident with each edge of G.
7. Prove that T (Kn )
= L(Kn+1 ).
8. If G is Hamiltonian, then prove that T (G) is Hamiltonian.
9. If G is Eulerian, then prove that T (G) is both Eulerian and Hamiltonian.
10. Prove that T (G) of every nontrivial connected graph G contains a spanning Eulerian
subgraph.
11. Show that the edge graph of a graph G has a Hamiltonian path if and only if G has a
walk W such that every edge of G not in W is incident with W .

262

Edge Graphs and Eccentricity Sequences

12. If G is any connected graph with (G) 4, then prove that L2 (G) is Hamiltonianconnected.
13. Construct graphs with eccentricity sequence
[2, 33 , 43 ].

14. If G is a connected graph with diameter 3 and e(u|G) = 3, then show that e(u|G) = 2.
15. If [e1 , e2 , . . ., en ] with en < 2e1 1 is an eccentricity sequence, then show that each
central vertex lies on a cycle.
16. If [ei ]n1 is the eccentricity sequence of an (n, m) graph, show that
1
m
2

n ei .
i=1

17. For a distance regular graph, prove the following

1
, then bi ci .
2d
1
b. If 1 i
, then b1 ci .
d 1
c. c2 k 2b1 .

a. If 1 i

10. Graph Matrices

Since a graph is completely determined by specifying either its adjacency structure or its
incidence structure, these specifications provide far more efficient ways of representing a
large or complicated graph than a pictorial representation. As computers are more adept at
manipulating numbers than at recognising pictures, it is standard practice to communicate
the specification of a graph to a computer in matrix form. In this chapter, we study various
types of matrices associated with a graph, and our study is based on Narsing Deo [63],
Foulds [82], Harary [104] and Parthasarathy [180].

10.1 Incidence Matrix

Let G be a graph with n vertices, m edges and without self-loops. The incidence matrix A of
G is an n m matrix A = [ai j ] whose n rows correspond to the n vertices and the m columns
correspond to m edges such that
ai j =

1,
0,

i f jth edge m j is incident on the ith vertex

otherwise.

It is also called vertex-edge incidence matrix and is denoted by A(G).

Example Consider the graphs given in Figure 10.1. The incidence matrix of G1 is
e1 e2 e3 e4 e5 e6 e7 e8
v1
v2
v
A(G1 ) = 3
v4
v5
v6

0
0
0
1
0
1

0
0
0
1
0
1

0
0
0
1
1
0

1
0
0
0
1
0

The incidence matrix of G2 is

0
1
0
1
0
0

1
1
0
0
0
0

0
1
0
0
1
0

0
1
1
0
0
0

264

Graph Matrices
e1 e2 e3 e4 e5

v1
1 1
v2
1 0
A(G2 ) =
v3 0 1
v4
0 0

0
0
1
1

0
1
1
0

0
1
.
1
0

The incidence matrix of G3 is

e1 e2 e3 e4 e5

v1
1 1
v2
1 1

A(G3 ) =
v3 0 0
v4
0 0

0
1
0
1

0
0
1
1

1
0
.
0
1

Fig. 10.1

The incidence matrix contains only two types of elements, 0 and 1. This clearly is a
binary matrix or a (0, 1)-matrix.
We have the following observations about the incidence matrix A.
1. Since every edge is incident on exactly two vertices, each column of A has exactly
two ones.
2. The number of ones in each row equals the degree of the corresponding vertex.

Graph Theory

265

3. A row with all zeros represents an isolated vertex.

4. Parallel edges in a graph produce identical columns in its incidence matrix.
5. If a graph is disconnected and consists of two components G1 and G2 , the incidence
matrix A(G) of graph G can be written in a block diagonal form as
A(G) =

A(G1 )
0
0
A(G2 )

where A(G1 ) and A(G2 ) are the incidence matrices of components G1 and G2 . This
observation results from the fact that no edge in G1 is incident on vertices of G2 and
vice versa. Obviously, this is also true for a disconnected graph with any number of
components.
6. Permutation of any two rows or columns in an incidence matrix simply corresponds
to relabeling the vertices and edges of the same graph.
Note The matrix A has been defined over a field, Galois field modulo 2 or GF(2), that
is, the set {0, 1} with operation addition modulo 2 written as + such that 0 + 0 = 0, 1 + 0 =
1, 1 +1 = 0 and multiplication modulo 2 written as. such that 0.0 = 0, 1.0 = 0 = 0.1, 1.1 = 1.

The following result is an immediate consequence of the above observations.

Theorem 10.1 Two graphs G1 and G2 are isomorphic if and only if their incidence matrices A(G1 ) and A(G2 ) differ only by permutation of rows and columns.

Proof Let the graphs G1 and G2 be isomorphic. Then there is a one-one correspondence
between the vertices and edges in G1 and G2 such that the incidence relation is preserved.
Thus A(G1 ) and A(G2 ) are either same or differ only by permutation of rows and columns.
The converse follows, since permutation of any two rows or columns in an incidence
matrix simply corresponds to relabeling the vertices and edges of the same graph.
q

Rank of the incidence matrix

Let G be a graph and let A(G) be its incidence matrix. Now each row in A(G) is a vector
over GF(2) in the vector space of graph G. Let the row vectors be denoted by A1 , A2 , . . .,
An . Then,

A(G) =

A1
A2
.
.
.
An

266

Graph Matrices

Since there are exactly two ones in every column of A, the sum of all these vectors is 0
(this being a modulo 2 sum of the corresponding entries).
Thus vectors A1 , A2 , . . ., An are linearly dependent. Therefore, rank A < n.
Hence, rank A n 1.
From the above observations, we have the following result.
Theorem 10.2 If A(G) is an incidence matrix of a connected graph G with n vertices,
then rank of A(G) is n 1.

Proof Let G be a connected graph with n vertices and let the number of edges in G be m.
Let A(G) be the incidence matrix and let A1 , A2 , . . ., An be the row vector of A(G).

Then, A(G) =

A1
A2
.
.
.
An

(10.2.1)

Clearly, rank A(G) n 1.

(10.2.2)

Consider the sum of any m of these row vectors, m n 1. Since G is connected, A(G)
cannot be partitioned in the form
A(G) =

A(G1 )
0

0
A(G2 )

such that A(G1 ) has m rows and A(G2 ) has n m rows.

Thus there exists no m m submatrix of A(G) for m n 1, such that the modulo 2 sum
of these m rows is equal to zero.
As there are only two elements 0 and 1 in this field, the additions of all vectors taken m
at a time for m = 1, 2, . . ., n 1 gives all possible linear combinations of n 1 row vectors.
Thus no linear combinations of m row vectors of A, for m n 1, is zero.
Therefore, rank A(G) n 1.
Combining (10.2.2) and (10.2.3), it follows that rank A(G) = n 1.

(10.2.3)
q

Remark If G is a disconnected graph with k components, then it follows from the above
theorem that rank of A(G) is n k.
Let G be a connected graph with n vertices and m edges. Then the order of the incidence
matrix A(G) is n m. Now, if we remove any one row from A(G), the remaining (n 1) by
m submatrix is of rank (n 1). Thus the remaining (n 1) row vectors are linearly independent. This shows that only (n 1) rows of an incidence matrix are required to specify the

Graph Theory

267

corresponding graph completely, because (n 1) rows contain the same information as the
entire matrix. This follows from the fact that given (n 1) rows, we can construct the nth
row, as each column in the matrix has exactly two ones. Such an (n 1) m matrix of A
is called a reduced incidence matrix and is denoted by A f . The vertex corresponding to the
deleted row in A f is called the reference vertex. Obviously, any vertex of a connected graph
can be treated as the reference vertex.
The following result gives the nature of the incidence matrix of a tree.
Theorem 10.3

The reduced incidence matrix of a tree is non-singular.

Proof A tree with n vertices has n 1 edges and also a tree is connected. Therefore, the
reduced incidence matrix is a square matrix of order n 1, with rank n 1. Thus the result
follows.
Now a graph G with n vertices and n 1 edges which is not a tree is obviously disconnected. Therefore the rank of the incidence matrix of G is less than (n 1). Hence the
(n 1) (n 1) reduced incidence matrix of a graph is non-singular if and only if the graph
is a tree.
q

10.2 Submatrices of A(G)

Let H be a subgraph of a graph G, and let A(H) and A(G) be the incidence matrices of H
and G respectively. Clearly, A(H) is a submatrix of A(G), possibly with rows or columns
permuted. We observe that there is a one-one correspondence between each n k submatrix
of A(G) and a subgraph of G with k edges, k being a positive integer, k < m and n being the
number of vertices in G.
The following is a property of the submatrices of A(G).
Theorem 10.4 Let A(G) be the incidence matrix of a connected graph G with n vertices. An (n 1) (n 1) submatrix of A(G) is non-singular if and only if the n 1 edges
corresponding to the n 1 columns of this matrix constitutes a spanning tree in G.

Proof Let G be a connected graph with n vertices and m edges. So, m n 1.

Let A(G) be the incidence matrix of G, so that A(G) has n rows and m columns (m n 1).
We know every square submatrix of order (n 1) (n 1) in A(G) is the reduced incidence matrix of some subgraph H in G with n 1 edges, and vice versa. We also know that
a square submatrix of A(G) is non-singular if and only if the corresponding subgraph is a
tree.
Obviously, the tree is a spanning tree because it contains n 1 edges of the n-vertex
graph.
Hence (n 1) (n 1) submatrix of A(G) is non-singular if and only if n 1 edges
corresponding to n 1 columns of this matrix forms a spanning tree.
q

268

Graph Matrices

The following is another form of incidence matrix.

Definition: The matrix F = [ f i j ] of the graph G = (V, E) with V = {v1 , v2 , . . ., vn } and
E = {e1 , e2 , . . ., em }, is the n m matrix associated with a chosen orientation of the edges
of G in which for each e = (vi , v j ), one of vi or v j is taken as positive end and the other as
negative end, and is defined by

1,
1 ,
fi j =

0,

i f vi is the positive end o f e j ,

i f vi is the negative end o f e j ,
i f vi is not incident with e j .

This matrix F can also be obtained from the incidence matrix A by changing either of
the two 1s to 1 in each column.
The above arguments amount to arbitrarily orienting the edges of G, and F is then the
incidence matrix of the oriented graph.
The matrix F is then the modified definition of the incidence matrix A.
Example

Consider the graph G shown in Figure 10.2, with V = {v1 , v2 , v3 , v4 } and E =

{e1 , e2 , e3 , e4 , e5 }.

The incidence matrix is given by

e1 e2 e3
v1
1 0 0
v2
1 1 0
A=
v3 0 1 1
v4
0 0 1

e4
1
0
0
1

e5
0
1
.
0
1

Therefore,
e1
v1
1
v2
1

F=
v3 0
v4
0

e2 e3
0
0
1
0
1 1
0 1

e4
e5
1
0
0
1
0
0
1 1

Fig. 10.2

Graph Theory

Theorem 10.5

269

If G is a connected graph with n vertices, then rank F = n 1.

Proof Let G be a connected graph with V = {v1 , v2 , . . ., vn } and E = {e1 , e2 , . . ., em }.

Then the matrix F = [ f i j ]nm is given by

1,
1 ,
fi j =

0,

i f vi is the positive end o f e j ,

i f vi is the negative end o f e j ,
i f vi is not incident with e j .

Let R j be the j th row of F . Since each column of F has only one +1 and one 1, as
non-zero entries, rank F < n. Thus, rank F n 1.
n
Now, let c j R j = 0 be any other linear dependence relation of R1 , R2 , . . ., Rn with at least
1
one c j non-zero.
If cr 6= 0, then the row Rr has non-zero entries in those columns which correspond to
edges incident with vr . For each such column there is just one row, say Rs , at which there is
a non-zero entry (with opposite sign to the non-zero entry in Rr ). The dependence relation
thus requires cs = cr , for all s corresponding to vertices adjacent to vr . Since G isconnected,

n

we have c j = c, for all j = 1, 2, . . ., n. Therefore the dependence relation is c R j = 0,

1

which is same as the first one. Hence, rank F = n 1.

Alternative Proof
Then

R1
R2

F =
M .
Rn

(10.5.1)

Since each column of F has only one +1 and one 1 as non-zero entries, R j = 0.
j=1

Thus, rank F n 1.

(10.5.2)

Consider the sum of any m of these row vectors, m n 1. As G is connected, F cannot

be partitioned in the form F =

F1
0

0
F2

, such that F1 has m rows and F2 as n m rows.

Therefore there exists no m m submatrix of F for m n 1, such that the sum of these m
rows is equal to zero. Therefore,
m

R j 6= 0.
j=1

(10.5.3)

270

Graph Matrices

Also, there is no linear combination of m(m n 1) vectors of F , which is zero. For, if

c j R j = 0 is a linear combination, with at least one c j non-zero, say cr 6= 0, then the row
m

j=1

Rr has non-zero entries in those columns which correspond to edges incident with vr . So
for each such column there is just one row, say Rs at which there is a non-zero entry with
opposite sign to the non-zero entry in Rr . The linear combination thus requires cs = cr for
all s corresponding to vertices adjacent to vr . As G is connected, we have c j = c, for all
m

j=1

j=1

j = 1, . . ., m. Therefore the linear combination becomes c ( R j ) = 0, or R j = 0, which

contradicts (10.5.3).
Hence, rank F = n 1.
Theorem 10.6

Proof

If G is a disconnected graph with k components, then rank F = n k.

Since G has k components, F can be partitioned as

F1
0
0

F =
..
.
0

0
F2
0

0
0
F3

Fk

0
0
0

where Fi is the matrix of the ith component Gi of G. We have proved that rank Fi = ni 1,
where ni is the number of vertices in Gi .
Thus, rank F = n1 1 + n2 1 + . . .+ nk 1 = n1 + n2 + . . . + nk k = n k,
as the number of vertices in G is n1 + n2 + . . . + nk = n.

Corollary 10.1 A basis for the row space of F is obtained by taking for each i, 1 i k,
any ni 1 rows of Fi .
Theorem 10.7 The determinant of any square submatrix of the matrix F of a graph G
has value 1, 1, or zero.

Proof Let N be the square submatrix of F such that N has both non-zero entries +1 and
1 in each column. Then row sum of N is zero and hence |N| = 0. Clearly if N has no
non-zero entries, then |N| = 0.
Now let some column of N have only one non-zero entry. Then expanding |N| with the
help of this column, we get |N| = |N 0 |, where N 0 is a matrix obtained by omitting a row
and column of N . Continuing in this way, we either get a matrix whose determinant is zero,
or end up with a single non-zero entry of N , in which case |N| = 1.
q

Graph Theory

271

Theorem 10.8 Let X be any set of n 1 edges of the connected graph G = (V, E) and
Fx the (n 1) (n 1) submatrix of the matrix F of G, determined by any n 1 rows and
those columns which correspond to the edges of X . Then Fx is non-singular if and only if
the edge induced subgraph < X > of G is a spanning tree of G.

Proof Let F 0 be the matrix corresponding to < X >. If < X > is a spanning tree of G, then
Fx consists of n 1 rows of F 0 . Since < X > is connected, therefore rank Fx = n 1. Hence
Fx is non-singular.
Conversely, let Fx be non-singular. Then F 0 contains an (n 1) (n 1) non singular
submatrix. Therefore, rank F 0 = n 1. Since rank + nullity = m, for any graph G, and
m(< X >) = n 1 and rank (< X >) = n 1, therefore, nullity (< X >) = 0. Thus < X > is
acyclic and connected and so is a spanning tree of G.
q

10.3 Cycle Matrix

Let the graph G have m edges and let q be the number of different cycles in G. The cycle
matrix B = [bi j]qm of G is a (0, 1) matrix of order q m, with bi j = 1, if the ith cycle
includes j th edge and bi j = 0, otherwise. The cycle matrix B of a graph G is denoted by
B(G).
Example

Consider the graph G1 given in Figure 10.3.

Fig. 10.3

The graph G1 has four different cycles Z1 = {e1 , e2 }, Z2 = {e3 , e5 , e7 }, Z3 = {e4 , e6 , e7 }

and Z4 = {e3 , e4 , e6 , e5 }.
The cycle matrix is

272

Graph Matrices
e1 e2 e3 e4 e5 e6 e7 e8

z1
1 1 0
z2
0 0 1
B(G1 ) =
z3 0 0 0
z4
0 0 1

0
0
1
1

0
1
0
1

0
0
1
1

0
1
1
0

0
0
.
0
0

The graph G2 of Figure 10.3 has seven different cycles, namely, Z1 = {e1 , e2 },
Z2 = {e2 , e7 , e8 }, Z3 = {e1 , e7 , e8 }, Z4 = {e4 , e5 , e6 , e7 }, Z5 = {e2 , e4 , e5 , e6 , e8 },
Z6 = {e1 , e4 , e5 , e6 , e8 } and Z7 = {e9 }. The cycle matrix is given by
e1 e2 e3 e4 e5 e6 e7 e8 e9
z1
z2
z3
B(G2 ) = z4
z5
z6
z7

1
0
1
0
0
1
0

1
1
0
0
1
0
0

0
0
0
0
0
0
0

0
0
0
1
1
1
0

0
0
0
1
1
1
0

0
0
0
1
1
1
0

0
1
1
1
0
0
0

0
1
1
0
1
1
0

0
0
0
0
0
0
1

We have the following observations regarding the cycle matrix B(G) of a graph G.
1. A column of all zeros corresponds to a non cycle edge, that is, an edge which does
not belong to any cycle.
2. Each row of B(G) is a cycle vector.
3. A cycle matrix has the property of representing a self-loop and the corresponding
row has a single one.
4. The number of ones in a row is equal to the number of edges in the corresponding
cycle.
5. If the graph G is separable (or disconnected) and consists of two blocks (or components) H1 and H2 , then the cycle matrix B(G) can be written in a block-diagonal form
as
B(G) =

B(H1 )
0
0
B(H2 )

where B(H1 ) and B(H2 ) are the cycle matrices of H1 and H2 . This follows from the
fact that cycles in H1 have no edges belonging to H2 and vice versa.
6. Permutation of any two rows or columns in a cycle matrix corresponds to relabeling
the cycles and the edges.

Graph Theory

273

7. We know two graphs G1 and G2 are 2-isomorphic if and only if they have cycle
correspondence. Thus two graphs G1 and G2 have the same cycle matrix if and only
if G1 and G2 are 2-isomorphic. This implies that the cycle matrix does not specify a
graph completely, but only specifies the graph within 2-isomorphism.
For example, the two graphs given in Figure 10.4 have the same cycle matrix. They are
2-isomorphic, but are not isomorphic.

Fig. 10.4

The following result relates the incidence and cycle matrix of a graph without self-loops.
Theorem 10.9 If G is a graph without self-loops, with incidence matrix A and cycle
matrix B whose columns are arranged using the same order of edges, then every row of B
is orthogonal to every row of A, that is ABT = BAT 0 (mod2), where AT and BT are the
transposes of A and B respectively.

Proof Let G be a graph without self-loops, and let A and B, respectively, be the incidence
and cycle matrix of G.
We know that in G for any vertex vi and for any cycle Z j , either vi Z j or vi
/ Z j.
In case vi
/ Z j , then there is no edge of Z j which is incident on vi and if vi Z j , then there
are exactly two edges of Z j which are incident on vi .
Now, consider the ith row of A and the j th row of B (which is the j th column of BT ).
Since the edges are arranged in the same order, the rth entries in these two rows are both
non-zero if and only if the edge er is incident on the ith vertex vi and is also in the j th cycle
Z j.
We have [ABT ]i j = [A]ir [BT ]r j = [A]ir [B] jr = air b jr .
For each er of G, we have one of the following cases.
i. er is incident on vi and er
/ Z j . Here air = 1, b jr = 0.
ii. er is not incident on vi and er Z j . In this case, air = 0, b jr = 1.
iii. er is not incident on vi and er / Z j , so that air = 0, b jr = 0.
All these cases imply that the ith vertex vi is not in the j th cycle Z j and we have
[ABT ]i j = 0 0 (mod2).
iv. er is incident on vi and er Z j .

274

Graph Matrices

Here we have exactly two edges, say er and et incident on vi so that air = 1, ait = 1,
b jr = 1, b jt = 1. Therefore, [ABT ]i j = air b jr = 1 + 1 0 (mod2).
q
We illustrate the above theorem with the following example (Fig. 10.5).

Fig. 10.5

Clearly,

0
0
0
1
0
1

0
0
0
1
0
1

0
0
0
1
1
0

1
0
0
0
1
0

0
0
0
2
0
2

0
2
0
2
2
0

2
2
0
0
2
0

2
2
0
2
2
0

T
AB =

0
1
0
1
0
0

1
1
0
0
0
0

0
1
0
0
1
0

0
1
1
0
0
0

1
1
0
0
0
0
0
0

0
0
1
0
1
0
1
0

0
0
0
1
0
1
1
0

0
0
1
1
1
1
0
0

0(mod2).

We know that a set of fundamental cycles (or basic cycles) with respect to any spanning
tree in a connected graph are the only independent cycles in a graph. The remaining cycles
can be obtained as ring sums (i.e., linear combinations) of these cycles. Thus, in a cycle
matrix, if we take only those rows that correspond to a set of fundamental cycles and
remove all other rows, we do not lose any information. The removed rows can be formed
from the rows corresponding to the set of fundamental cycles. For example, in the cycle
matrix of the graph given in Figure 10.6, the fourth row is simply the mod 2 sum of the
second and the third rows. Fundamental cycles are

Graph Theory

275

Z1 = {e1 , e2 , e4 , e7 }
Z2 = {e3 , e4 , e7 }
Z3 = {e5 , e6 , e7 }

Fig. 10.6
e2 e3 e6
Z1
Z2
Z3

e1 e4 e5 e7

.
1 0 0 .. 1 1 0 1

.
.
0 1 0 .. 0 1 0 1

.
0 0 1 .. 0 0 1 1

A submatrix of a cycle matrix in which all rows correspond to a set of fundamental

cycles is called a fundamental cycle matrix B f .
The permutation of rows and/or columns do not affect B f . If n is the number of vertices,
m the number of edges in a connected graph G, then B f is an (m n + 1) m matrix because
the number of fundamental cycles is m n + 1, each fundamental cycle being produced by
one chord.
Now, arranging the columns in B f such that all the m n + 1 chords correspond to the
first m n + 1 columns and rearranging the rows such that the first row corresponds to the
fundamental cycle made by the chord in the first column, the second row to the fundamental
cycle made by the second, and so on. This arrangement is done for the above fundamental
cycle matrix.
A matrix B f thus arranged has the form
B f = [I : Bt ],

where I is an identity matrix of order = m n + 1 and Bt is the remaining (n 1)

submatrix, corresponding to the branches of the spanning tree.
From equation B f = [I : Bt ], we have rank B f = = m n + 1.
Since B f is a submatrix of the cycle matrix B, therefore, rank B rank B f and thus,
rank B m n + 1.
The following result gives the rank of the cycle matrix.

276

Graph Matrices

Theorem 10.10 If B is a cycle matrix of a connected graph G with n vertices and m

edges, then rank B = m n + 1.

Proof

Let A be the incidence matrix of the connected graph G.

Then ABT 0(mod2).

Using Sylvesters theorem (Theorem 10.13), we have rank A+ rank BT m so that
rank A+ rank B m.
Therefore, rank B m rank A.
As rank A = n 1, we get rank B m (n 1) = m n + 1.
But, rank B m n + 1.
Combining, we get rank B = m n + 1.

Theorem 10.10 can be generalised in the following form.

Theorem 10.11 If B is a cycle matrix of a disconnected graph G with n vertices, m edges
and k components, then rank B = m n + k.

Proof Let B be the cycle matrix of the disconnected graph G with n vertices, m edges
and k components. Let the k components be G1 , G2 , ..., Gk with n1 , n2, ..., nk vertices and m1 ,
m2 , . . ., mk edges respectively.
Then n1 + n2 + . . .+ nk = n and m1 + m2 + ... + mk = m.
Let B1 , B2 , . . ., Bk be the cycle matrices of G1 , G2 , . . ., Gk .

Then B(G) =

B1 (G1 )
0
0

0
B2 (G2 )
0

0
0
B3 (G3 )

0
0
0

..
.
0

Bk (Gk )

We know rank Bi = mi ni + 1, for 1 i k.

Therefore, rank B = rank B1 + . . .+ rank Bk
= (m1 n1 + 1) + . . .+ (mk nk + 1)
= (m1 + . . . + mk ) (n1 + . . . + nk ) + k = m n + k.

Graph Theory

277

Definition: Let A be a matrix of order k m, with k < m. The major determinant of A is

the determinant of the largest square submatrix of A, formed by taking any k columns of A.
That is, the determinant of any k k square submatrix is called the major determinant of A.
Let A and B be matrices of orders k m and m k respectively (k < m). If columns
i1 , i2 , . . ., ik of B are chosen for a particular major of B, then the corresponding major in A
consists of the rows i1 , i2 , . . ., ik in A.
If A is a square matrix of order n, then AX = 0 has a non trivial solution X 6= 0 if and only
if A is singular, that is |A| = 0. The set of all vectors X that satisfy AX = 0 forms a vector
space called the null space of matrix A. The rank of the null space is called the nullity of A.
Further more,
rank A + nullityA = n.
These definitions and the above equation also hold when A is a matrix of order k n, k < n.
We now give Binet-Cauchy and Sylvester theorems which will be used in the further
discussions.
Theorem 10.12 (BinetCauchy) If A and B are two matrices of the order k m and m
k respectively (k < m), then |AB| = sum of the products of corresponding major determinants
of A and B.

Proof


We multiply two (m + k) (m + k) partitioned matrices to get

Ik A
O Im



A
Im

O
B

O
Im

AB
B

where Im and Ik are identity matrices of order m and k respectively.

Therefore, det

A
Im


Thus, det(AB) =det

O
B
A
Im

= det
O
B

O
Im

AB
B

.
(10.12.1)

Now apply Cauchys expansion method to the right side of (10.12.1) and observe that the
only non-zero minors of any order in Im are its principal minors of that order. Therefore,
we see that the Cauchy expansion consists of these minors of order m k multiplied by
their cofactors of order k in A and B together.
q
Theorem 10.13 (Sylvester) If A and B are matrices of order k m and n p respectively,
then nullity AB nullity A+ nullity B.

Proof

Since every vector X satisfying BX = 0 also satisfies ABX = 0, therefore we have

nullity AB nullity B 0.

(10.13.1)

278

Graph Matrices

Let nullity B = s. So there exists a set of s linearly independent vectors {x1 , x2 , . . ., xs }

forming a basis of the null space of B. Therefore,
BXi = 0, for i = 1, 2, . . ., s.

(10.13.2)

Now let nullity AB = s + t . Thus there exists a set of t linearly independent vectors
[Xs+1 , Xs+2 , . . ., Xs+t ] such that the set {X1 , X2 , . . ., Xs , Xs+1 , . . ., Xs+t } forms a basis for
the null space of AB. Therefore,
ABXi = 0, for i = 1, 2, . . ., s, s + 1, . . ., s + t .

(10.13.3)

This implies that out of the s + t vectors Xi forming a basis of the null space of AB, the
first s vectors are made zero by B and the remaining non-zero BXi s, i = s + 1, . . ., s + t are
made zero by A.
Clearly, the vectors BXs+1 , . . ., BXs+t are linearly independent. For if
b1 BXs+1 + b2 BXs+2 + . . .+ bt BXs+t = 0,

i.e., if B(b1 Xs+1 + b2 Xs+2 + . . . + bt Xs+t ) = 0,

then the vector b1 Xs+1 + b2 Xs+2 + . . . + bt Xs+t is the null space of B, which is possible only if
b1 = b2 = . . . = bt = 0.
Thus we have seen that there are at least t linearly independent vectors which are made
zero by A. So, nullity A t .
Since t = (s + t) s, therefore t = nullity AB nullity B
Therefore, nullity AB nullity B nullity A, and so
nullity AB nullity A + nullity B.
Corollary 10.2

(10.13.4)

We know, rank A + nullity A = n, and using this in (10.13.4), we get

n rank AB n rank A + n rank B.

Therefore, rank AB rank A+ rank B n.

If in above, AB = 0, then rank A + rank B n.

10.4 Cut-Set Matrix

Let G be a graph with m edges and q cutsets. The cut-set matrix C = [ci j ]qm of G is a (0,
1)-matrix with
ci j =

1,

0,

i f ith cutset contains jth edge ,

otherwise .

Graph Theory

Example

279

Consider the graphs shown in Figure 10. 7.

Fig. 10.7(a)

Fig. 10.7(b)

In the graph G1 , E = {e1 , e2 , e3 , e4 , e5 , e6 , e7 , e8 }.

The cut-sets are c1 = {e8 }, c2 = {e1 , e2 }, c3 = {e3 , e5 }, c4 = {e5 , e6 , e7 }, c5 = {e3 , e6 , e7 }, c6 =
{e4 , e6 }, c7 = {e3 , e4 , e7 } and c8 = {e4 , e5 , e7 }.
The cut-sets for the graph G2 are c1 = {e1 , e2 }, c2 = {e3 , e4 }, c3 = {e4 , e5 }, c4 = {e1 , e6 }, c5
= {e2 , e6 }, c6 = {e3 , e5 }, c7 = {e1 , e4 , c7 }, c8 = {e2 , e3 , e7 } and c9 = {e5 , e6 , e7 }.
Thus the cut-set matrices are given by
e1 e2 e3 e4 e5 e6 e7 e8
c1
c2
c3
c
C(G1 ) = 4
c5
c6
c7
c8

0
1
0
0
0
0
0
0

0
1
0
0
0
0
0
0

0
0
1
0
1
0
1
0

0
0
0
0
0
1
1
1

0
0
1
1
0
0
0
1

0
0
0
1
1
1
0
0

0
0
0
1
1
0
1
1

1
0
0
0
0
0
0
0

, and

280

Graph Matrices
e1 e2 e3 e4 e5 e6 e7
c1
c2
c3
c
C(G2 ) = 4
c5
c6
c7
c8
c9

1
0
0
1
0
0
1
0
0

1
0
0
0
1
0
0
1
0

0
1
0
0
0
1
0
1
0

0
1
1
0
0
0
1
0
0

0
0
1
0
0
1
0
0
1

0
0
0
1
1
0
0
0
1

0
0
0
0
0
0
1
1
1

We have the following observations about the cut-set matrix C(G) of a graph G.
1. The permutation of rows or columns in a cut-set matrix corresponds simply to renaming of the cut-sets and edges respectively.
2. Each row in C(G) is a cut-set vector.
3. A column with all zeros corresponds to an edge forming a self-loop.
4. Parallel edges form identical columns in the cut-set matrix.
5. In a non-separable graph, since every set of edges incident on a vertex is a cut-set,
therefore every row of incidence matrix A(G) is included as a row in the cut-set matrix
C(G). That is, for a non-separable graph G, C(G) contains A(G). For a separable graph,
the incidence matrix of each block is contained in the cut-set matrix. For example, in
the graph G1 of Figure 10.7, the incidence matrix of the block {e3 , e4 , e5 , e6 , e7 } is
the 4 5 submatrix of C, left after deleting rows c1 , c2 , c5 , c8 and columns e1 , e2 , e8 .
6. It follows from observation 5, that rank C(G) rank A(G). Therefore, for a connected
graph with n vertices, rank C(G) n 1.
The following result for connected graphs shows that cutset matrix, incidence matrix
and the corresponding graph matrix have the same rank.
Theorem 10.14 If G is a connected graph, then the rank of a cut-set matrix C(G) is equal
to the rank of incidence matrix A(G), which equals the rank of graph G.

Proof Let A(G), B(G) and C(G) be the incidence, cycle and cut-set matrix of the connected graph G. Then we have
rank C(G) n 1.

(10.14.1)

Since the number of edges common to a cut-set and a cycle is always even, every row
in C is orthogonal to every row in B, provided the edges in both B and C are arranged in the
same order.

Graph Theory

281

Thus, BCT = CBT 0 (mod 2).

(10.14.2)

Now, applying Sylvesters theorem to equation (10.14.2), we have

rank B+ rank C m.
For a connected graph, we have rank B = m n + 1.
Therefore, rank C m rank B = m (m n + 1) = n 1.
So, rank C n 1.

(10.14.3)

It follows from (10.14.1) and (10.14.3) that rank C = n 1.

10.5 Fundamental Cut-Set Matrix

Let G be a connected graph with n vertices and m edges. The fundamental cut-set matrix C f
of G is an (n 1)m submatrix of C such that the rows correspond to the set of fundamental
cut-sets with respect to some spanning tree. Clearly, a fundamental cut-set matrix C f can
be partitioned into two submatrices, one of which is an identity matrix In1 of order n 1.
We have
C f = [Cc : In1 ],

where the last n 1 columns forming the identity matrix correspond to the n 1 branches
of the spanning tree and the first m n + 1 columns forming Cc correspond to the chords.
Example Consider the connected graphs G1 and G2 given in Figure 10.8. The spanning
tree is shown with bold lines. The fundamental cut-sets of G1 are c1 , c2 , c3 , c6 and c7 while
the fundamental cut-sets of G2 are c1 , c2 , c3 , c4 and c7 .

Fig. 10.8

The fundamental cut-set matrix of G1 and G2 , respectively are given by

282

Graph Matrices
e2
1

Cf = 0

e3 e4
0 0
1 0
0 1
1 1

0 0 0

e1 e5 e6 e7 e8
1 0 0 0 0

0 1 0 0 0

and
0 0 1 0 0

0 0 0 1 0

..
.
..
.
..
.
..
.
..
.

e1 e4
c1
c2
C f = c3
c4
c7

1 0

0 1

0 1

1 0

1 1

0 0 0 1

e2 e3 e5 e6 e7

..
.
..
.
..
.
..
.
..
.

1 0
0 1
0 0
0 0
0 0

0 0 0

0 0 0

1 0 0

0 1 0

0 0 1

10.6 Relations between A f , B f and C f

Let G be a connected graph and A f , B f and C f be respectively the reduced incidence matrix,
the fundamental cycle matrix, and the fundamental cut-set matrix of G.
We have shown that


.
B f = I .. Bt


..
and C f = Cc . In1 ,

(10.6.i)

(10.6.ii)

where Bt denotes the submatrix corresponding to the branches of a spanning tree and Cc
denotes the submatrix corresponding to the chords.
Let the spanning tree T in Equations (10.6.i) and (10.6.ii) be the same and let the order
of the edges in both equations be same. Also, in the reduced incidence matrix A f of size
(n 1) m, let the edges (i.e., the columns) be arranged in the same order as in B f and C f .
Partition A f into two submatrices given by
.
A f = Ac .. At ,


(10.6.iii)

where At consists of n 1 columns corresponding to the branches of the spanning tree T

and Ac is the spanning submatrix corresponding to the m n + 1 chords.

Graph Theory

283

Since the columns in A f and B f are arranged in the same order, the equation ABT =
= 0(mod 2) gives

BAT

A f BTf 0(mod 2),

or

or

Ac .. At ...


BtT

0(mod 2),

Ac + At BTf 0(mod 2).

(10.6.iv)

Since At is non singular, At1 exists. Now, premultiplying both sides of equation (10.6.iv)
by At1 , we have
At1 Ac + At1 At BtT 0(mod 2),

or

At1 Ac + BtT 0(mod 2).

Therefore, At1 Ac = BtT .

Since in mod 2 arithmetic 1 = 1,
BtT = At1 Ac .

(10.6.v)

Now as the columns in B f and C f are arranged in the same order, therefore (in mod 2
arithmetic) C f . BTf 0(mod 2) in mod 2 arithmetic gives C f .BTf = 0.



.
Therefore, Cc .. In1

..
.
BtT

= 0, so that Cc + BtT = 0, that is, Cc = BtT .

Thus, Cc = BtT (as 1 = 1 in mod 2 arithmetic).

Hence, Cc = At1 Ac from (10.6.v).
Remarks

We make the following observations from the above relations.

1. If A or A f is given, we can construct B f and C f starting from an arbitrary spanning

tree and its submatrix At in A f .
2. If either B f or C f is given, we can construct the other. Therefore, since B f determines
a graph within 2-isomorphism, so does C f .
3. If either B f and C f is given, then A f in general cannot be determined completely.

284

Graph Matrices

Example

Consider the graph G of Figure 10.9.

Fig. 10.9

Let {e1 , e5 , e6 , e7 , e8 } be the spanning tree.

e1 e2 e3 e4 e5 e6 e7 e8

We have, A =

0
0
0
1
0
1

0
0
0
1
0
1

0
0
0
1
1
0

1
0
0
0
1
0

0
1
0
1
0
0

1
1
0
0
0
0

0
1
0
0
1
0

0
1
1
0
0
0

Dropping the sixth row in A, we get

e2 e3 e4

Af =

0
0
0
1
0

0
0
0
1
1

1
0
0
0
1

e2 e3 e4

1
Bf = 0
0

e1 e5 e6 e7 e8
:
:
:
:
:

0
0
0
1
0

0
1
0
1
0

1
1
0
0
0

0
1
0
0
1

0
1
1
0
0

= [Ac : At ].

e1 e5 e6 e7 e8

0 0 : 1 0
1 0 : 0 1
0 1 : 0 0

0 0 0
0 1 0 = [I3 : Bt ] and
1 1 0

Graph Theory

285
e2 e3 e4 e1 e5 e6 e7 e8

1 0 0 1 0 0 0 0
0 1 0 0 1 0 0 0

0 0 1 0 0 1 0 0
= [Cc : I5 ] .
0 1 1 0 0 0 1 0
0 0 0 0 0 0 0 1

Cf =

Clearly, BtT = Cc.

We verify At1 Ac = BtT .
Now,

At =

0
0
0
1
0

Therefore,

0
1
0
1
0

1
1
0
0
0

At1 Ac

0
1
0
0
1

0
1
1
0
0
1
0
0
0
0

1 0 0

, Bt = 0 1 0

0 0 1
0
1
0
1
0

0
0
1
1
0

0 0
1 0
1 0

. Hence, At1 Ac = BtT .

10.7 Path Matrix

Let G be a graph with m edges, and u and v be any two vertices in G. The path matrix
for vertices u and v denoted by P(u, v) = [pi j ]qm , where q is the number of different paths
between u and v, is defined as
pi j =

1,

0,

i f jth edge lies in the ith path ,

otherwise .

Clearly, a path matrix is defined for a particular pair of vertices, the rows in P(u, v)
correspond to different paths between u and v, and the columns correspond to different
edges in G. For example, consider the graph in Figure 10.10.

286

Graph Matrices

Fig. 10.10

The different paths between the vertices v3 and v4 are

p1 = {e8 , e5 }, p2 = {e8 , e7 , e3 } and p3 = {e8 , e6 , e4 , e3 }.

The path matrix for v3 , v4 is given by

e1 e2 e3 e4 e5 e6 e7 e8

0 0 0
P(v3 , v4 ) = 0 0 1
0 0 1

0 1 0
0 0 0
1 0 1

0 1
1 1 .
0 1

We have the following observations about the path matrix.

1. A column of all zeros corresponds to an edge that does not lie in any path between u
and v.
2. A column of all ones corresponds to an edge that lies in every path between u and v.
3. There is no row with all zeros.
4. The ring sum of any two rows in P(u, v) corresponds to a cycle or an edge-disjoint
union of cycles.
The next result gives a relation between incidence and path matrix of a graph.
Theorem 10.15 If the columns of the incidence matrix A and the path matrix P(u, v)
of a connected graph are arranged in the same order, then under the product (mod 2).
APT (u, v) = M ,

where M is a matrix having ones in two rows u and v, and the zeros in the remaining n 2
rows.

Proof Let G be a connected graph and let vk = u and vt = v be any two vertices of G. Let
A be the incidence matrix and P(u, v) be the path matrix of (u, v) in G.

Graph Theory

287

Now for any vertex vi in G and for any u v path p j in G, either vi p j or vi / p j .

If vi / p j , then there is no edge of p j which is incident on vi .
If vi p j , then either vi is an intermediate vertex of p j , or vi = vk or vt . In case vi is an
intermediate vertex of p j , then there are exactly two edges of p j which are incident on vi
and in case vi = vk or vt , there is exactly one edge of p j which is incident on vi .
Now consider the ith row of A and the j th row of P (which is the j th column of PT (u, v)).
As the edges are arranged in the same order, the rth entries in these two rows are both
non zero if and only if the edge er is incident on the ith vertex vi and is also on the j th path
p j . Let APT (u, v) = M = [mi j ].
We have, APT


ij

= [A]ir [PT ]r j .
r=1
m

Therefore, mi j = air p jr.

r=1

For each edge er of G, we have one of the following cases.

i. er is incident on vi and er
/ p j . Here air = 1, b jr = 0.
ii. er is not incident on vi and er p j . Here air = 0, b jr = 1.
iii. er is not incident on vi and er
/ p j . Here air = 0, b jr = 0.
All these cases imply that the ith vertex vi is not in j th path p j and we have Mi j = 0
0(mod 2). (Fig. 10.11(a)).
iv. er is incident on vi and er p j (Fig. 10.11(b)).
If vi is an intermediate vertex of p j , then there are exactly two edges say er and et incident
on vi so that air = 1, ait = 1, p jr = 1, p jt = 1.
Therefore, mi j = 1 + 1 = 0(mod 2).
If vi = vk or vt then the edge er is incident on either vk or vt . So, akr = 1, p jr = 1, or
atr = 1, p jr = 1.
Thus, mk j = air p jr = 1.1 1(mod 2), and
mt j = air p jr = 1.1 1(mod 2).

Hence M = [mi j ] is a matrix, such that under modulo 2,

mi j =

1,

0,

f or i = k , t ,
q

otherwise .

288

Graph Matrices

Fig. 10.11

Example

In the graph of Figure 10.10, we have

T
AP (v3 , v4 ) =

0
0
0
1
0
1

0
0
0
1
0
1

0
0
0
1
1
0

1
0
0
0
1
0

0
1
0
1
0
0

1
1
0
0
0
0

0
1
0
0
1
0

0
1
1
0
0
0

0
0
0
0
1
0
0
1

0
0
1
0
0
0
1
1

0
0
1
1
0
1
0
1

v1
v2
v3
v4
v5
v6

0
0
1
1
0
0

0
0
1
1
0
0

0
0
1
1
0
0

(mod2).

10.8 Adjacency Matrix

Let V = (V, E) be a graph with V = {v1 , v2 , . . ., vn }, E = {e1 , e2 , . . ., em } and without parallel
edges. The adjacency matrix of G is an n n symmetric binary matrix X = [xi j ] defined over
the ring of integers such that
xi j =

Example

1,

0,

i f vi v j E ,
otherwise .

Consider the graph G given in Figure 10.12.

Fig. 10.12

Graph Theory

289

The adjacency matrix of G is given by

v1 v2 v3 v4 v5 v6
v1
v2
v
X= 3
v4
v5
v6

0
1
0
0
1
1

1
0
0
1
1
0

0
0
0
1
0
0

0
1
1
0
1
1

1
1
0
1
0
0

1
0
0
0
0
0

We have the following observations about the adjacency matrix X of a graph G.

1. The entries along the principal diagonal of X are all zeros if and only if the graph has
no self-loops. However, a self-loop at the ith vertex corresponds to xii = 1.
2. If the graph has no self-loops, the degree of a vertex equals the number of ones in the
corresponding row or column of X .
3. Permutation of rows and the corresponding columns imply reordering the vertices.
We note that the rows and columns are arranged in the same order. Therefore, when
two rows are interchanged in X , the corresponding columns are also interchanged.
Thus two graphs G1 and G2 without parallel edges are isomorphic if and only if their
adjacency matrices X(G1 ) and X(G2 ) are related by
X(G2 ) = R1 X(G1 )R,

where R is a permutation matrix.

4. A graph G is disconnected having components G1 and G2 if and only if the adjacency
matrix X(G) is partitioned as

X(G1 )
X(G) = . .
O

:
O
:
. . ,
: X(G2 )

where X(G1 ) and X(G2 ) are respectively the adjacency matrices of the components
G1 and G2 . Obviously, the above partitioning implies that there are no edges between
vertices in G1 and vertices in G2 .
5. If any square, symmetric and binary matrix Q of order n is given, then there exists a
graph G with n vertices and without parallel edges whose adjacency matrix is Q.
Definition: An edge sequence is a sequence of edges in which each edge, except the
first and the last, has one vertex in common with the edge preceding it and one vertex

290

Graph Matrices

in common with the edge following it. A walk and a path are the examples of an edge
sequence. An edge can appear more than once in an edge sequence. In the graph of Figure
10.13, v1 e1 v2 e2 v3 e3 v4 e4 v2 e2 v3 e5 v5 , or e1 e2 e3 e4 e2 e5 is an edge sequence.

Fig. 10.13

We now have the following result.

Theorem 10.16 If X = [xi j ] is the adjacency matrix of a simple graph G, then [X k ]i j is
the number of different edge sequences of length k between vertices vi and v j .

Proof We prove the result by using induction on k. The result is trivial for k = 0 and 1.
Since X 2 = X.X, X 2 is a symmetric matrix, as product of symmetric matrices is also
symmetric.
For k = 2, i 6= j , we have
[X 2 ]i j = number of ones in the product of ith row and j th column (or j th row) of X
= number of positions in which both ith and j th rows of X have ones
= number of vertices that are adjacent to both ith and j th vertices
= number of different paths of length two between ith and j th vertices

Also, [X 2 ]ii = number of ones in the ith row (or column) of X

= degree of the corresponding vertex.

This shows that [X 2 ]i j is the number of different paths and therefore different edge sequences of length 2 between the vertices vi and v j . Thus the result is true for k = 2.
Assume the result to be true for k, so that
[X k ]i j = number of different edge sequences of length k between vi and v j .
n

r=1

r=1

We have, [X k+1]i j = [X k X]i j = [X k ]ir [X]r j = [X k ]ir xr j .

Graph Theory

291

Now, every vi v j edge sequence of length k + 1 consists of a vi vr edge sequence of

length k, followed by an edge vt v j . Since there are [X k ]ir such edge sequences of length
k and xr j such edges for each vertex vr , the total number of all vi v j edge sequences of
n

length k + 1 is [X k ]ir xr j . This proves the result for k + 1 also.

r=1

We have the following observation about connectedness and adjacency matrix.

Theorem 10.17 Let G be a graph with V = {v1 , v2 , . . ., vn } and let X be the adjacency
matrix of G. Let Y = [yi j ] be the matrix Y = X + X 2 + . . .+ X n1 .
Then G is connected if and only if for all distinct i, j, yi j 6= 0. That is, if and only if Y has
no zero entries off the main diagonal.

Proof

We have, yi j = [Y ]i j = [X]i j + [X 2 ]i j + . . .+ [X n1 ]i j .

Since [X k ]i j denotes the number of distinct edge-sequences of length k from vi to v j ,

yi j = number of different vi v j edge sequence of length 1
+ number of different vi v j edge sequences of length 2 + . . .
+ number of different vi v j edge sequences of length n 1.

Therefore, yi j = number of different vi v j edge sequence of length less than n.

Now let G be connected. Then for every pair of distinct i, j there is a path from vi to
v j . Since G has n vertices, this path passes through atmost n vertices and so has length less
than n. Thus, yi j 6= 0 for each i, j with i 6= j .
Conversely, for each distinct pair i, j we have yi j 6= 0. Then from above, there is at least
one edge sequence of length less than n from vi to v j . This implies that vi is connected to
v j . Since the distinct pair i, j is chosen arbitrarily, G is connected.
q
The next result is useful in determining the distances between different pairs of vertices.
Theorem 10.18 In a connected graph, the distance between two vertices vi and v j is k if
and only if k is the smallest integer for which [X k ]i j 6= 0.

Proof Let G be a connected graph and let X = [xi j ] be the adjacency matrix of G. Let vi
and v j be vertices in G such that
d(vi , v j ) = k.

Then the length of the shortest path between vi and v j is k.

This implies that there are no paths of length 1, 2, . . ., k 1 and so no edge sequences of
length 1, 2, . . ., k 1 between vi and v j .

292

Graph Matrices

Therefore, [X]i j = 0, [X 2 ]i j = 0, . . ., [X k1]i j = 0.

Hence k is the smallest integer such that [X k ]i j 6= 0.
Conversely, suppose that k is the smallest integer such that [X k ]i j 6= 0.
Therefore, there are no edge sequences of length 1, 2, . . ., k 1 and in fact no paths of
length 1, 2, . . ., k 1 between vertices vi and v j .
Thus the shortest path between vi and v j is of length k, so that d(vi , v j ) = k.

Definition: Let G be a graph and let di be the degree of the vertex vi in G. The degree
matrix H = [hi j ] of G is defined by
hi j =

0,

f or i 6= j ,

di ,

f or i = j .

The following result gives a relation between the matrices F, X and H .

Theorem 10.19 Let F be the modified incidence matrix, X the adjacency matrix and H
the degree matrix of a graph G. Then
FF T = H X .

We have (i, j )th element of FF T ,

Proof

[FF T ]i j =

r=1

r=1

[F]ir [F T ]r j = [F]ir[F] jr.

Now, [F]ir and [F]r j are non-zero if and only if the edge er = vi v j . Then for i 6= j ,
m

[F]ir [F] jr =
r=1

1 ,

i f er = vi v j is an edge ,

0,

i f er = vi v j is not an edge .

For i = j, [F]ir [F] jr = 1 whenever [F]ik = 1, and this occurs di times corresponding to
the number of edges incident on vi . Thus,
m

[F]ir [F] jr = di, for i = j .

r=1

Therefore,
[FF T ]i j =

1 or 0 ,

di ,

according to whether f or i 6= j , vi v j is an edge or not ,

f or i = j .

Graph Theory

293

Also, [H X]i j = [H]i j [X]i j

=

di 0 ,

di ,

f or i = j ,

0 (1 or 0) , according as f or i 6= j , vi v j is an edge, or vi v j is not an edge .

f or i = j ,

1 or 0 , according as f or i 6= j , vi v j is an edge, or vi v j is not an edge .

Hence FF T = H X .

Corollary 10.3 The matrix Q = FF T is independent of the orientation used for the edges
of G in getting F .
Theorem 10.20 Let X , F and H be the adjacency, modified incidence and degree matrices of the graph G, and Q = FF T = H X . Then the matrix of cofactors of Q denoted by adj
Q is a multiple of the all ones n n matrix J .

Proof If G is disconnected, then rank Q =rank FF T < n 1 and so every cofactor of Q is

zero. Therefore, adj Q = O = O.J , where J = [Ji j ]nn with Ji j = 1 for all i, j .
Now let G be connected. Then rank Q = n 1 and therefore |Q| = 0 This implies that
every column of adj Q belongs to the kernel (null space) of Q.
But nullity Q = 1 (as rank Q+ nullity Q = n). So, if u is the n-vector of ones, then
(H X)u = 0.

and therefore u is in the null space of Q.

Thus every other vector in the null space of Q and in particular every column of adj Q is
a multiple of u.
Since Q and so adj Q are symmetric, the multiplying factor for all columns of adj Q are
same.
Hence, adj Q = cJ , where c is a constant.
q
The next result called Matrix-tree theorem can be used in finding the complexity of a
connected graph.
Theorem 10.21 (Matrix-tree theorem) If X, F and H are the adjacency, modified incidence and degree matrices of the connected graph G, and Q = FF T , J is the n n matrix
of ones and (G) is the complexity of G, then adj Q = (G).J.

Proof Let X, F and H be the adjacency, modified incidence and degree matrix of a connected graph G. We have Q = FF T and adj Q = matrix of the cofactors of Q. Also (G)J is
a matrix whose every entry is (G) as J is a matrix whose every entry is unity.

294

Graph Matrices

Therefore, to prove adj Q = (G)J , it is enough to prove that (G) = any one cofactor of
Q.

Let F0 be the matrix obtained by dropping the last row from F . Then clearly, |Fo FoT | is a
cofactor of Q.
Using Binet-Cauchy theorem of matrix theory, we have


 
 F0 F T  = | Fx | F T  ,
x
0

(10.21.1)

X E

where Fx is the square submatrix of Fo whose n 1 columns correspond to n 1 edges in

the subset X of E , the summation running over all possible such subsets.
We know |Fx | =
6 0 if and only if < X > is a spanning tree of G and |Fx| = 1.
But, |FxT | = |Fx| .
Therefore each X E such that < X > is a spanning tree of G contributes one to the sum
on the right of (10.21.1) and all other contributions are zero.
q
Hence |Fo FoT | = (G), proving the theorem.
Corollary 10.4

Proof

Prove (Kn ) = nn2 .

Here, Q = H X = (n 1)I (J I) = nI J . Therefore,

Q=

n
0
0
:
0

0 0
n 0
0 n
0 0

.. 0

.. 0

. . 0

.. n

n 1 1
1
1 n 1 1
1
1 n 1
:
1
1
1

..
..
..

1
1
1
:
1

1 1
1 1
1 1
1 1
1
1
1

. . n1

.. 1
.. 1

.. 1

.. 1

The cofactor of q11 is the (n 1) (n 1) determinant given by


 n 1 1 . . 1

 1 n 1 . . 1
cofactor of q11 = 
 :
 1
1 . . n 1

Subtracting the first row from each of the others and then adding the last n 2 columns
to the first, we get

Graph Theory





cofactor of q11 = 




295
1 1
0 n
0 0
:
0 0


1 . . 1 
0 . . 0 
n . . 0 .


0 .. n 

Expanding with the help of the first column, we have cofactor of q11 = nn2 . Thus,
(kn ) = nn2 .

10.9 Exercises
1. Characterise A f , B f , C f and X of the complete graph of n vertices.
2. Characterise simple, self-dual graphs in terms of their cycle and cut-set matrices.
3. Show that each diagonal entry in X 3 equals twice the number of triangles passing
through the corresponding vertex.
4. Characterise the adjacency matrix of a bipartite graph.
5. Prove that a graph is bipartite if and only if for all odd k, every diagonal entry of Ak
is zero.
6. Similar to the cycle or cut-set matrix, define a spanning tree matrix for a connected
graph, and observe some of its properties.
7. If X is the adjacency matrix of a graph G and L is the adjacency matrix of its edge
graph L(G) and A and H are the incidence and degree matrices, show that X = AAT H
and L = AT A 2I .
8. Use the matrix tree theorem to calculate (K4 e).
9. Prove that (G) =

1
det(J + Q).
n2

11. Digraphs
The concept of digraphs (or directed graphs) is one of the richest theories in graph theory,
mainly because of their applications to physical problems. For example, flow networks
with valves in the pipes and electrical networks are represented by digraphs. They are
applied in abstract representations of computer programs and are an invaluable tools in the
study of sequential machines. They are also used for systems analysis in control theory.
Most of the concepts and terminology of undirected graphs are also applicable to digraphs,
and hence in this chapter more emphasis will be given to those properties of digraphs that
are not found in undirected graphs.
There are two different treatments of digraphsone can be found in the book by Harray, Norman and Cartwright [191] and the other in the book by Berge [18]. The former
discusses the application of digraphs to sociological problems, and the latter gives a comprehensive mathematical treatment. One can also refer to the books by N. Deo [63], Harary
[104], Behzad, Chartrand and Lesniak-Foster [16], and Buckley and Harary [42].

11.1 Basic Definitions

Digraphs (Directed graphs): A digraph D is a pair (V, A), where V is a nonempty set
whose elements are called the vertices and A is the subset of the set of ordered pairs of
distinct elements of V . The elements of A are called the arcs of D (Fig. 11.1(a)).
Multidigraphs: A multidigraph D is a pair (V, A), where V is a nonempty set of vertices
and A is a multiset of arcs, which is a multisubset of the set of ordered pairs of distinct
elements of V . The number of times an arc occurs in D is called its multiplicity and arcs
with multiplicity greater than one are called multiple arcs of D (Fig. 11.1(b)).

General digraphs:

A general digraph D is a pair (V, A), where V is a nonempty set

of vertices, and A is a multiset of arcs, which is a multisubset of the cartesian product of
V with itself. An arc of the form uu is called a loop of D and arcs which are not loops are
called proper arcs of D. The number of times an arc occurs is called its multiplicity. A
loop with multiplicity greater than one is called a multiple loop (Fig. 11.1(c)).

Graph Theory

297

Oriented graph: A digraph containing no symmetric pair of arcs is called an oriented

graph (Fig. 11.1(d)).

Fig. 11.1

For u, v V , an arc a = (u, v) A is denoted by uv and implies that a is directed from u

to v. Here, u is the initial vertex (tail) and v is the terminal vertex (head). Also we say that
a joins u to v; a is incident with u and v; a is incident from u and a is incident to v; and u
is adjacent to v and v is adjacent from u. In case both uv and vu belong to A, then uv and
vu are called a symmetric pair of arcs. For any v V , the number of arcs incident to v is
the indegree of v and is denoted by d (v). The number of arcs incident from v is called the
outdegree of v and is denoted by d + (v). Berge calls indegree and outdegree as inner and
outer-demi degrees. The total degree (or simply degree) of v is d(v) = d (v) + d + (v).
We define N +(v) and N (v) by
N +(v) = {u V : vu A} and N (v) = {u V : uv A}.

If d(v) = k for every v V , then D is said to be a k-regular digraph. If for every v V, d (v) =
d +(v), the digraph is said to be an isograph or a balanced digraph. We note that an isograph
is an even degree digraph, but not necessarily regular. Also, every symmetric digraph is an
isograph. Edmonds and Johnson [70] call isographs as asymmetric digraphs and Berge [23]
calls them pseudo-symmetric graphs. Kotzig [137, 138] calls an anti-symmetric isograph
as oriented-in-equilibrium or a -graph.
A vertex v for which d + (v) = d (v) = 0 is called an isolate. A vertex v is called a transmitter or a receiver according as d + (v) > 0, d (v) = 0 or d + (v) = 0, d (v) > 0. A vertex v is
called a carrier if d + (v) = d (v) = 1.
Underlying graph of a digraph: Let D = (V, A) be a digraph. The graph G = (V, E),
where uv E if and only if uv or vu or both are in A, is called the underlying graph of D.
This is also called the covering graph C(D) of D. Here we denote C(D) by G(D) or simply
by G.
In case G = (V, E ) is a graph, the digraph with vertex set V and a symmetric uv whenever
uv E , is called the digraph corresponding to G, and is denoted by D(G), or D. Clearly,
D(G) is a symmetric digraph. An oriented graph obtained from the graph G = (V, E) by
replacing each edge uv E by an arc uv or vu, but not both is called an orientation of G and
is denoted by O(G) or O.

298

Digraphs

Complete symmetric digraph:

A digraph D = (V, A) is said to be complete if both
uv and vu A, for all u, v V . Obviously this corresponds to Kn , where |V | = n, and is
denoted by Kn . A complete antisymmetric digraph, or a complete oriented graph is called
a tournament. Clearly, a tournament is an orientation of Kn (Fig. 11.2).

Fig. 11.2

We note that the number of arcs in Kn is n (n 1) and the number of arcs in a tournament
is n(n1)
2 .

11.2 Digraphs and Binary Relations

Let A and B be nonempty sets. A (binary) relation R from A to B is a subset of A B. If
R A B and (a, b) R, where a A, b B, we say a is related to b by R, and we write
aRb. If a is not related to b by R, we write a /
R b. A relation R defined on a set X is a subset
of X X . For example, less than, greater than and equality are the relations in the set of
real numbers. The property is congruent to defines a relation in the set of all triangles in
a plane. Also, parallelism defines a relation in the set of all lines in a plane.
Let R define a relation on a nonempty set X . If R relates every element of X to itself, the
relation R is said to be reflexive. A relation R is said to be symmetric if for all xi x j X, xi R x j
implies x j R xi . A relation R is said to be transitive if for any three elements xi , x j and xk in
X , xi Rx j and x j R xk imply xi R xk . A binary relation is called an equivalence relation if it is
reflexive, symmetric and transitive.
A binary relation R on a set X can always be represented by a digraph. In such a representation, each xi X is represented by a vertex xi and whenever there is a relation R from
xi to x j , an arc is drawn from xi to x j , for every pair (xi , x j ). The digraph in Figure 11.3
represents the relation is less than, on a set consisting of four numbers 2, 3, 4, 6.

Fig. 11.3

Graph Theory

299

We note that every binary relation on a finite set can be represented by a digraph without
parallel edges and vice versa.
Clearly, the digraph of a reflexive relation contains a loop at every vertex (Fig. 11.4). A
digraph representing a reflexive binary relation is called a reflexive digraph.

Fig 11.4

The digraph of a symmetric relation is a symmetric digraph because for every arc from
xi to x j , there is an arc from x j to xi . Figure 11.5 shows the digraph of an irreflexive and

symmetric relation on a set of three elements.

Fig. 11.5

A digraph representing a transitive relation on its vertex set is called a transitive digraph.
Figure 11.6 shows the digraph of a transitive, which is neither reflexive, nor symmetric.

Fig. 11.6

A binary relation R on a set M can also be represented by a matrix, called a relation

matrix. This is a (0, 1), n n matrix MR = [mi j ], where n is the number of elements in M ,
and is defined by
mi j =

1
0,

i f xi R x j is true,
otherwise.

300

Digraphs

Isomorphic digraphs:
Two digraphs are said to be isomorphic if their underlying
graphs are isomorphic and the direction of the corresponding arcs are same. Two nonisomorphic digraphs are shown in Figure 11.7.

Fig. 11.7

Subdigraph: Let D = (V, A) be a digraph. A digraph H = (U, B) is the subdigraph of D

whenever U V and B A. If U = V , the subdigraph is said to be spanning.
Complement of a digraph: The complement D = (V, A) of the digraph D = (V, A) has
vertex set V and a A if and only if a
/ A. That is, D is the relative complement of D in Kn ,
where |V | = n.
Converse digraph: The converse D0 = (V, A0 ) of the digraph D = (V, A) has vertex set
V and a = uv A0 if and only if a0 = vu A. That is, A0 is obtained by reversing the direction
of each arc of D. Clearly, (D0 )0 = D00 = D.
A digraph D is self-complementary if D
= D and D is said to be self-converse if D
= D0 .
0
A digraph D is said to be self-dual if D
=D
=D.

11.3 Directed Paths and Connectedness

Directed walks: A (directed) walk in a digraph D = (V A) is a sequence v0 a1 v1 a2 . . . ak vk ,
where vi V and ai A are such that ai = vi1 vi for 1 i k, no arc being repeated. As there
is only one arc of the form vi v j , the walk can also be represented by the vertex sequence
v0 v1 . . .vk . A vertex may appear more than once in a walk. Clearly, the length of the walk
is k. If v0 6= vk , the walk is open, and if v0 = vk , the walk is closed. A walk is spanning if
V = {v0 , . . ., vk }.
A (directed) path is an open walk in which no vertex is repeated. A (directed) cycle is a
closed walk in which no vertex is repeated. A digraph is acyclic if it has no cycles.
A semiwalk is a sequence v0 a1 v1 a2 . . .ak vk with vi V and ai A such that either ai = vi1 vi
or ai vi vi1 and no arc is repeated. The length of the semiwalk is k. If v0 6= vk , the semiwalk
is open, and if vo = vk , the semiwalk is closed. If no vertex is repeated in an open (closed)
semiwalk, it is called a semi path (semicycle).
A spanning path of a digraph is called a Hamiltonian path and a spanning cycle is called
a Hamiltonian cycle. A digraph with a Hamiltonian cycle is said to be Hamiltonian.

Graph Theory

301

Fig. 11.8

In Figure 11.8, v1 a1 v2 a2 v3 a4 v4 is an open walk, v1 a1 v2 a3 v3 a4 v4 is semiwalk, v1 a1 v2 a8 v4

a5 v5 is a path and v1 a1 v2 a8 v4 a5 v5 a6 v1 is a cycle.
In a digraph D = (V, A), a vertex u is said to be joined to a vertex v, if there is a semipath
from u to v. We note that the relation is joined to is reflexive, symmetric and transitive,
and therefore is an equivalence relation on V . A vertex u is said to be reachable from a
vertex v, if there is a path from v to u. The relation is reachable from is reflexive and
transitive, but not symmetric, since there may or may not be a path from u to v. A vertex v
is called a source of D if every vertex of D is reachable from v, and v is called a sink of D,
if v is reachable from every other vertex.

Principle of duality for digraphs

While changing a digraph D to its converse D0 , we observe that the properties about D
get changed to the corresponding properties about D0 . When D0 is changed to D00 = D, the
original properties of D are obtained. Such type of a pair of properties are called dual
properties, (transmitter, receiver), (source, sink), (indegree, outdegree), (isolate, isolate)
and (carrier, carrier). The dual of a statement P about a digraph is the statement P0 obtained
from P by changing every concept in P to its dual. For any statements P and Q for digraphs,
P Q in D is true if and only if P0 Q0 in D0 and for any digraph D there is a converse D0 .
Therefore for every result in digraphs we get a dual result by changing every property to
its dual. This is called the principle of duality for digraphs.
Definition: A digraph is said to be strongly connected or strong, if every two of its
distinct vertices u and v are such that u is reachable from v and v is reachable from u.
A digraph is unilaterally connected or unilateral, if either u is reachable from v or v is
reachable from u and is weakly connected or weak, if u and v are joined by a semipath.

302

Digraphs

In Figure 11.9, (a) shows a strong digraph, (b) a unilateral digraph and (c) a weak digraph.

Fig. 11.9

A digraph is said to be disconnected if it is not even weak. A digraph is said to be strictly

weak if it is weak, but not unilateral. It is strictly unilateral, if it is unilateral but not strong.
Two vertices of a digraph D are said to be
i. 0-connected if there is no semipath joining them,
ii. 1-connected if there is a semipath joining them, but there is no u v path or vu path,
iii. 2-connected if there is a u v or a v u path, but not both,
iv. 3-connected if there is u v path and a v u path.
Definition: An arc sequence in a digraph D is an alternating sequence of vertices and
arcs of D.
The following results characterise various types of connectivity in digraphs.
Theorem 11.1

A digraph is strong if and only if it has a spanning closed arc sequence.

Proof
Necessity Let D = (V, A) be a strong digraph with V = {v1 , v2 , . . ., vn }. Then there is an
arc sequence from each vertex in V to every other vertex in V . Therefore, there exists in D,
arc sequences Q1 , Q2 , . . ., Qn1 such that the first vertex of Qi is vi and the last vertex of Qi
is vi+1 , for i = 1, 2, . . ., n 1. Also, there exists an arc sequence, say Qn , with first vertex
vn and the last vertex v1 . Then the arc sequence obtained by traversing the arc sequences
Q1 , Q2 , . . ., Qn in succession is a spanning closed arc sequence of D.
Sufficiency Let u and v be two distinct vertices of V . If v follows u in any spanning
closed arc sequence, say Q of D, then there exists a sequence of the arcs of Q forming an
arc sequence from u to v. If u follows v in Q, then there is an arc sequence from u to the
last vertex of Q and an arc sequence from that vertex to v. An arc sequence from u to v is
then obtained by traversing these two arc sequences in succession.
q
Theorem 11.2

A digraph D is unilateral if and only if it has a spanning arc sequence.

Graph Theory

303

Proof
Necessity Assume D is unilateral. Let Q be an arc sequence in D which contains maximum number of vertices and let Q begin at the vertex v1 of D and end at the vertex v2 of D.
If Q is a spanning arc sequence, there is nothing to prove. Assume that Q is not a spanning
arc sequence. Then there exists a vertex, say u of D that is not in Q. Also in D, there is
neither an arc sequence from u to v1 , nor from v2 to u. Since D is unilateral and does not
contain an arc sequence from u to v1 , D contains an arc sequence from v1 to u.
Let w(6= v2 ) be the last vertex of Q from which an arc sequence from w to u exists in
D. Let Q1 be an arc sequence from w to u in D. Let z be the vertex in D which is the
immediate successor of the last appearance of w in Q. Clearly, D does not contain an arc
sequence from z to u. Since D is unilateral, there is an arc sequence, say Q2 from u to z in
D. Traversing Q from v1 to the last appearance of w, then traversing u, then traversing Q2
to vertex z and finally traversing Q to v2 , we obtain an arc sequence from v1 to v2 which has
more distinct vertices than Q. This is a contradiction and thus Q is a spanning arc sequence
in D (Fig. 11.10).
The sufficiency follows from the definition.
q

Fig. 11.10

Theorem 11.3

A digraph is weak if and only if it has a spanning semi arc sequence.

Proof Let D = (V, A) be a weak digraph with V = {v1 , v2 , . . ., vn }. Since D is weak, there
is a semi arc sequence, say Qi from vi to vi+1 in D for i = 1, 2, . . ., n 1. The semi arc
sequence obtained by traversing the semi arc sequences Q1 , Q2 , . . ., Qn in succession is a
spanning semi arc sequence of D.
Conversely, let D be a digraph containing a spanning semi arc sequence, say Q. Let v1
and v2 be two distinct vertices of D. Clearly, v1 and v2 are in Q, since Q is spanning. The
part of Q which begins at any appearance of v1 (v2 ) and ends at any appearance of v2 (v1 )
represents a semi arc sequence from v1 to v2 (from v2 to v1 ) in D. Thus there is either a semi
arc sequence from v1 to v2 , or from v2 to v1 in D. Hence D is weak.
Definition: In a digraph D, a strong component is a maximal strong subdigraph of D.
A unilateral component is a maximal unilateral subdigraph of D and a weak component is
a maximal weak subdigraph of D. In Figure 11.1(b), the digraph has a strong component
induced by the vertex set {v2 , v3 , v4 }. In Figure 11.1(c), the digraph has a unilateral component induced by the vertex set {v1 , v4 , v2 } and a weak component which is the digraph
itself.

304

Digraphs

Theorem 11.4
i. Every vertex and every arc of a digraph D belongs to a unique weak component.
ii. Every vertex and every arc of a digraph of D belongs to at least one unilateral component.
iii. Every vertex of a digraph belongs to a unique strong component. Every arc is contained in at most one strong component and it is in a strong component if and only if
it is in a cycle.

Proof
i. If a vertex v lies in two weak components W1 and W2 , let v1 and v2 be any two vertices
of W1 W2 and W2 W1 . Then there is a vv1 and vv2 semi-path. Also there is a v1 v2
semipath, so that W1 and W2 are in the same weak component. Similar argument holds
for an arc.
ii. Since each vertex and each arc is a unilateral subdigraph, the result follows.
iii. Let v be any vertex and S be the set of vertices mutually reachable with v including v
itself. Then h S i is a strong component containing v. The uniqueness of this follows
as in (i). Also, a similar argument holds for an arc.
If an arc uv is in a cycle, then all vertices on the cycle are pairwise reachable and belong
to a strong component containing the arcs of the cycle, in particular uv. Conversely, if uv
is in a strong component, u and v are mutually reachable and hence there is vu path, which
together with uv gives a cycle.
q
We note that the vertex sets of the weak components of a digraph give a partition w
of its vertex set V and is called a weak partition. Similarly, the vertex sets of the strong
components of a digraph give a partition s of V , which is a refinement of the partition w .
This s is called the strong partition.
Definition: Let V = S1 S2 . . . Sk be a partition of the vertex set V of the digraph D
= (V, A). Consider a digraph D with vertex set V = {S1 , S2 , . . ., Sk } which has an arc Si S j
if and only if in D there is at least one arc from a vertex of Si to a vertex of S j . Then D is
called the contraction of D with respect to the partition .
The contraction of a digraph D with respect to its strong partition S is called the condensation D of D.
Definition: The symmetrisation DS of a digraph D is obtained from D by adding uv to A
whenever uv
/ A, but vu A. Equivalently, DS = D(G(D)), which is the digraph corresponding to the underlying graph of D.

Graph Theory

305

Figure 11.11 illustrates these operations.

Fig. 11.11

We note from the definition of the condensation that a digraph D and its condensation
have the same kind of connectedness.
Theorem 11.5 If S1 and S2 are two strong components of digraph D, and v1 S1 and
v2 S2 , then there is a v1 v2 path in D if and only if there is an S1 S2 path in D .

Proof Let there be a v1 v2 path P in D. If length of P is one, then there is an S1 S2

arc in D . We induct on the length of P. Let the result hold for any path of length n 1 in
D. Assume that P = v1 u1 u2 . . .un1 v2 is of length n. Let S(un1 ) be the strong component
containing un1 . Then by induction hypothesis, there is a path P from S1 to S(un1 ) in D .
If un1 S2 , this path serves as an S1 S2 path. If not, since un1 v2 A, there is an arc from
S(un1 ) to S2 in D , and thus an S1 S2 path.
Conversely, let S1 S3 S4 . . .S2 be an S1 S2 path in D . Then there are arcs u1 u3 , u03 u4 , u04
u5 , ..., u0n u2 in D, ui , u0i Si . Since the Si s are strong components, there are v1 u1 , u3
u03 , . . ., un u0n , u2 v2 paths in D. These along with the arcs given above form a v1 v2 path
in D.
q
The following result can be easily established.
Theorem 11.6 If v1 , v2 are vertices in different strong components S1 , S2 of a digraph
D, then there is a strict semipath but no path joining v1 to v2 if and only if there is a strict
semipath and no path from S1 to S2 in D .
Now we have the following result.
Theorem 11.7

The condensation D of any digraph is acyclic.

Proof If D contains a cycle, let S1 S2 be an arc in this cycle. By Theorem 11.1, S1 S2

lies on a strong component of D . Therefore S1 and S2 are mutually reachable in D , and
by Theorem 11.5, there are vertices v1 S1 and v2 S2 which are mutually reachable in D.
Thus v1 and v2 belong to the same strong component S of D. So, S1 = S2 = S, and S1 S2 is
not an arc in D , contradicting the assumption.
q

306

Digraphs

We note that a digraph D is strong if and only if D consists of a single vertex.

Definition: A cut-set in a digraph D = (V, A) is a set of arcs of A, which constitute a cutset in the multigraph G = (V, E ), obtained from D by removing the orientation from each
arc of A.

11.4 Euler Digraphs

A digraph D is said to be Eulerian if it contains a closed walk which traverses every arc of
D exactly once. Such a walk is called an Euler walk. A digraph D is said to be unicursal if
it contains an open Euler walk.
The following result characterises Eulerian digraphs.
Theorem 11.8 A digraph D = (V, A) is Eulerian if and only if D is connected and for
each of its vertices v, d (v) = d + (v).

Proof
Necessity Let D be an Eulerian digraph. Therefore, it contains an Eulerian walk, say W .
In traversing W , every time a vertex v is encountered we pass along an arc incident towards
v and then an arc incident away from v. This is true for all the vertices of W , including the
initial vertex of W , say v, because we began W by traversing an arc incident away from v
and ended W by traversing an arc incident towards v.
Sufficiency Let for every vertex v in D, d (v) = d + (v). For any arbitrary vertex v in D, we
identify a walk, starting at v and traversing the arcs of D at most once each. This traversing
is continued till it is impossible to traverse further. Since every vertex has the same number
of arcs incident towards it as away from it, we can leave any vertex that we enter along the
walk and the traversal then stops at v. Let the walk traversed so far be denoted by W . If W
includes all arcs of A, then the result follows. If not, we remove from D all the arcs of W
and consider the remainder of A. By assumption, each vertex in the remaining digraph, say
D1 , is such that the number of arcs directed towards it equals the number of arcs directed
away from it. Further, W and D1 have a vertex, say u in common, since D is connected.
Starting at u, we repeat the process of tracing a walk in D1 . If this walk does not contain
all the arcs of D1 , the process is repeated until a closed walk that traverses each of the arcs
of D exactly once is obtained. Hence D is Eulerian.
q
Theorem 11.9 A weakly connected digraph D = (V, A) is unicursal if and only if D contains vertices u and v such that d + (u) = d (u) + 1, d (v) = d + (v) + 1 and d +(w) = d (w), for
all w V , where w 6= u, v. In this case, the open Euler walk begins at u and ends at v.

Proof Let D be unicursal. Then D has an Euler walk W that begins at u and ends at v.
Therefore, as in Theorem 11.8, for every vertex w different from both u and v, we have

Graph Theory

307

d +(w) = d (w). Also, the first arc of W contributes one to the outdegree of u while every
other occurrence of u in W contributes one each to the outdegree of u. Thus, d + (u) =
d (u) + 1.

Similarly, d (v) = d + (v) + 1.

Conversely, let D be a weakly connected digraph containing vertices u and v such that
d +(u) = d (u) + 1, d (v) = d +(v) + 1 and for each w 6= u, v, d + (w) = d (w). In D, add a new
arc a joining u and v. Now, we get a digraph D1 in which the new outdegree of v is one
more than its old outdegree, so that d (v) = d +(v). Similarly, d (u) = d + (u), and for every
other vertex w, d (w) = d + (w). Also, D1 is weakly connected, and since d (z) = d + (z) for
every vertex z in D1 , it follows from Theorem 11.8 that D1 is Eulerian. Let W be an Euler
walk in D1 . Clearly, D1 contains all the arcs of D together with the added arc a. Deleting
the arc a produces an open Euler walk in D. Hence D is unicursal.
q
Theorem 11.10 A non-trivial weak digraph is an isograph if and only if it is the union
of arc-disjoint cycles.

Proof If the weak digraph D is a union of arc-disjoint cycles, each cycle contributes one
to the indegree and one to the outdegree of each vertex on it. Thus, d + (v) = d (v), for all
v V.
Conversely, let D be a non-trivial weak isograph. Then each vertex has positive outdegree and therefore D has a cycle, say Z . Removing the edges of Z from D, we get a digraph
D1 whose weak components are isographs. By using an induction argument, each such
non-trivial weak component is a union of arc-disjoint cycles. These cycles together with Z
provide a decomposition of the arc set of D into cycles.
q
Corollary 11.1

Every weak isograph is strong.

Proof If u and v are any two vertices of the weak isograph D, there is a semi path P
joining u and v, and each arc of this lies on some cycle of D. The union of these cycles
provides a closed walk containing u and v. Thus u and v are mutually reachable.
q

11.5 Hamiltonian digraphs

Definition: A spanning path of a digraph is called a Hamiltonian path and a spanning cycle a Hamiltonian cycle. A digraph containing a Hamiltonian cycle is said to be
Hamiltonian.
Regarding the results on Hamiltonian digraphs, there are surveys by Bermond and
Thomassen [24], and by Jackson. There are a number of results that are analogous to
those proved for Hamiltonian graphs in Chapter 3. Two early results on sufficient conditions are due to Ghouila-Houri [87] and Woodall [270]. But now we have a more general
result due to Meyniel [160]. The proof of Meyniels result given here is due to Bondy and
Thomassen [38].

308

Digraphs

First, we have the following observations.

Lemma 11.1 Let P = v1 v2 . . .vk be a path in the digraph D and v be any vertex in V
V (P). If there is no v1 vk path with vertex set V (P) {v}, then |{v, V (P)}| k + 1, where
{v, V (P)} is the set of all arcs in D with one end in v or V (P) and the other end in V (P) or v,
respectively.

Proof By the assumption on P, for any v V V (P), there is no path vi v vi+1 in D.

Therefore for each i, 1 i k 1, |(vi , v)| + |(v, vi+1 )| 1.
k1

Hence, |{v, V (P)}| = (|vi , v)| + |v, vi+1 | + |(v, v1 )| + |(vk , v)|
i=1

(k 1)1 + 2 = k + 1.

Let in a digraph D = (V, A), S be a proper subset of V . A u v path of length at least two
with only u and v in S is called an S-path.
Theorem 11.11 Let D be a strong non-Hamiltonian digraph and C = v1 v2 . . .vk v1 be a
cycle of D such that there is no cycle of D whose vertex set properly contains V (C) = C,
say. Then there exists a v V C, and integers p and q(1 p, q k) such that (i) v p v A,
(ii) vv p+i / A for i, 1 i q and (iii) d(v) + d(v p+q ) 2n 1 q.

Proof
Case 1 D has no C -path. Since D is strong and C is a proper subset of V , there is a vertex in
C joined to some vertex of V C by a path and joined from some vertex of V C by a path.
By the assumption on C, there is a cycle C0 having only one vertex, say v p , common with C.
Let v be the successor of v p on C0 (Fig. 11.12(a)). If v is adjacent to or from any vertex of C
other than v p , D has a C-path. Therefore, |{v,C}| 2. Clearly, |{v p+1 ,C}| 2(k 1). For any
u V C , if vuv p+1 or v p+1 uv is a 2-path, then these are C-paths in D. Thus, such adjacencies
are not possible. Therefore, |{u, {v, v p+1 }}| 2 for such u. Putting these together, we have
d(v) + d(v p+1 ) 2 + 2(k 1) + 2(n k 1) = 2n 2,

and this verifies the statements with q = 1.

Case 2 D has a C -path, say P = v p u1 u2 . . .us v p , where ui V C . Choose P such that r is
least. By the assumption on C, r > 1. Let v = u1 .

i. By the assumption on C and the minimality of r, v is not adjacent to any v p+i , 1 i

r 1. Therefore the path Q = v p+r v p+r+1 . . . vk . . . v1 . . . v p and v satisfy the hypothesis
of Lemma 11.1 and we have
|{v, V (Q)}| = |{v, C}| k r + 2.

(11.11.1)

Graph Theory

309

ii. By the minimality of r, we observe that for any u V C, vuv p+i or v p+i uv are not
possible paths in D, for 1 i r 1 (Fig. 11.12(b)). Thus, for any u V C,
|{u, {v, vp + i}}| = 2, for any such i.
(11.11.2)
iii. Since D is strong, there are v p v p+r and v p+r v p paths in D. Clearly, v p+r v p+r+1 . . .
vk v1 v2 . . .v p is such a v p+r v p path. It is possible that there are other v p+r v p
paths containing all these vertices and some more, say v p+1 v p+2 . . .. v p+i1 . Let q
be the largest integer i, 1 i r such that there is a v p+r v p path with vertex set
S = {v p+r , v p+r+1 , . . . , v1 , v2 , . . ., v p1 , v p+1 , . . ., v p+q2 , v p+q1 , v p } and let P0 be such a
path. By the assumption on C, S cannot contain all the vertices of C, since P P0 is a
cycle. Therefore, q < r. It is possible that q = 1 (Fig. 11.12 (c)). Now the lemma is
applicable for v p+q and P0 . Therefore,
|{Vp+q , V (P0 )}| k r + q + 1.

(11.11.3)

Using (11.11.2) with i = q, we have

|{u, {v, v p+q}}| 2, for each u V C .

(11.11.4)

Also, v p+q can be joined to and from any of the other r q 1 vertices among
{v p+q+1 , . . ., v p+r1 }.

(11.11.5)

310

Digraphs

Fig. 11.12

Combining all these, we get

d(v) + d(v p+q ) = |{v,C}| + |{v p+q,V(P0 )}| + |{v p+q , {v p+q+1 , . . ., v p+r1 }}|
+ |{u, {v, v p+q}}|
uV C

(k r + 2) + (k r + q + 1) + 2(r q 1) + 2(n k 1)
= 2n q 1.

Theorem 11.12 (Meyneil) If D is a strong digraph of order n such that for any pair of
non-adjacent vertices u and v, d(u) + d(v) 2n 1, then D is Hamiltonian.

Proof If such a D is non-Hamiltonian, by Theorem 11.11, there exists a pair of nonadjacent vertices v and v p+q such that d(v) + d(v p+q ) < 2n 1, contradicting the hypothesis.
q

Corollary 11.2 If D is a strong digraph such that for any vertex v,

n, then D is Hamiltonian.

d(v) = d + (v) + d (v)

The direct proof of this result can be found in Berge [18].

11.6 Trees with Directed Edges

We know that a tree in undirected graphs is a connected graph without cycles. But in case
of digraphs, a structure similar to that of a tree needs absence of cycles as well as absence
of semi cycles. We have the following definition.

Graph Theory

311

A (directed) tree is a connected digraph without cycles, neither directed cycles nor semi
cycles. We observe that a tree with n vertices has n 1 directed edges, and has properties
analogous to those of trees with undirected edges. A digraph whose weak components are
trees is called a forest. For example, Figure 11.13 shows a tree.

Fig. 11.13

Arborescence: A (directed) tree is said to be an arborescence if it contains exactly one

vertex, called the root, with no arcs directed towards it and if all the arcs on any semipath
are directed away from the root. For example, the tree in Figure 11.14 is an arborescence.
That is, every vertex other than the root has indegree exactly one. Arborescence is also
called an out-tree. If the direction of every arc in an arborescence is reversed, we get a tree
called an in-tree.

Fig. 11.14

Theorem 11.13 In an arborescence, there is a directed path from the root v to every
other vertex. Conversely, a digraph D without cycles is an arborescence if there is a vertex
v in D such that every other vertex is reachable from v and v is not reachable from any other
vertex.

Proof In an arborescence, consider a directed path P starting from the root v and continuing as far as possible. Clearly, P can end only at a pendant vertex, since otherwise, we get
a vertex whose indegree is two or more, which is a contradiction. Since an arborescence is
connected, every vertex lies on some directed path from the root v to each of the pendant
vertices.
Conversely, since every vertex in D is reachable from v and D has no cycle, D is a tree.
Further, since v is not reachable from any other vertex, d (v) = 0. Every other vertex
is reachable from v and therefore indegree of each of these vertices is at least one. The

312

Digraphs

indegree is not greater than one, because there are only n 1 arcs in D, n being the number
of vertices of D.
q
Ordered trees:
A tree in which the relative order of subtrees meeting at each vertex
is preserved is called an ordered tree or a planar tree (because the tree can be visualised
as rigidly embedded in the plane of the paper). In computer science, the term tree usually
means an ordered tree and by convention, a tree is drawn hanging down with the root at
the top.
Spanning trees:
A spanning tree is an n-vertex connected digraph analogous to a
spanning tree in an undirected graph and consists of n 1 directed arcs. A spanning arborescence in a connected digraph is a spanning tree that is an arborescence. For example,
{a, b, c, g} is a spanning arborescence in Figure 11.15.

Fig. 11.15

Theorem 11.14 In a connected isograph D of n vertices and m arcs, let W = (a1 , a2 , . . ., am )

be an Euler line, which starts and ends at a vertex v (that is, v is the initial vertex of a1 and
the terminal vertex of am ). Among the m arcs in W there are n 1 arcs that enter each of
n 1 vertices, other than v, for the first time. The subdigraph D1 of these n 1 arcs together
with the n vertices is a spanning arborescence of D, rooted at vertex v.

Proof In the subdigraph D1 , vertex v is of indegree zero, and every other vertex is of
indegree one, for D1 includes exactly one arc going to each of the n 1 vertices and no arc
going to v. Further, the way D1 is defined in W , implies that D1 is connected and contains
n 1 arcs. Therefore D1 is a spanning arborescence in D and is rooted at v.
q
Illustration: In Figure 11.16, W = (b d c e f g h a) is an Euler line, starting and ending
at vertex 2. The subdigraph {b, d, f } is a spanning arborescence rooted at vertex 2.

Fig. 11.16

Graph Theory

313

The following result is due to Van Aardenne-Ehrenfest and N.G. de Bruijn [257].
Theorem 11.15 Let D be an Euler digraph and T be a spanning in-tree in D, rooted
at a vertex v. Let a1 be an arc in D incident out of the vertex v. Then a directed walk
W = (a1 , a2 , . . ., am ) is a directed Euler line, if it is constructed as follows.
i. No arc is included in W more than once.
ii. In exiting a vertex the one arc belonging to T is not used until all other outgoing arcs
have been traversed.
iii. The walk is terminated only when a vertex is reached from which there is no arc left
on which to exit.

Proof The walk W terminates at v, since all vertices have been entered as often as they
have been left (because D is an isograph). Now assume there is an arc a in D that has not
been included in W . Let u be the terminal vertex of a. Since D is an isograph, u is also the
initial vertex of some arc b not included in W . Arc b going out of vertex u is in T , according
to (i). This omitted arc leads to another omitted arc c in T , and so on. Finally we arrive at
v and find an outgoing arc not included in W . This contradicts (iii).
q
Theorem 11.14 provides a method of obtaining a spanning arborescence rooted at any
specified vertex, provided the digraph is Eulerian. Conversely, given a spanning arborescence in an Euler digraph, an Euler line can be constructed using Theorem 11.15.
The number of distinct Euler lines formed from a given in-tree T and starting with arc
a1 at v, can be computed by considering all the choices available at each vertex, after
starting with a1 . Since there is exactly one outgoing arc in T at each vertex and this arc
is to be selected last ((ii), Theorem 11.15), the remaining d + (vi ) 1 arcs at vertex vi can
be chosen in (d + (vi ) 1)! ways. Since these are independent choices, we have altogether
n
(d + (vi ) 1)! different Euler lines that meet (i), (ii) and (iii) of Theorem 11.15.

i=1

Illustration: Consider Figure 11.17. We apply (i), (ii) and (iii) of Theorem 11.15 to obtain different Euler lines from the in-tree {a2 , a3 , a7 , a10 , a11 }, starting with arc a1 . The two
Euler lines obtained are (a1 a12 a5 a6 a7 a8 a9 a10 a11 a2 a4 a3 ) and (a1 a12 a8 a9 a10 a11 a5 a6 a7
a2 a4 a3 ).

Fig. 11.17

314

Digraphs
n

Here, (d + (vi ) 1)! = 2.

i=1

Note that these are not all the Euler lines in the digraph, but only those that are generated
by the specific in-tree in accordance with (i), (ii) and (iii) of Theorem 11.15.
Fundamental cycles in digraphs:
The arcs of a connected digraph not included in a specified spanning tree T are called
the chords with respect to T . As in undirected graphs, every chord ci added to the spanning
tree T produces a fundamental cycle, which is a directed cycle or a semi cycle.
A cut-set in a connected digraph D induces a partitioning of the vertices of D into two
disjoint subsets V1 and V2 such that the cut-set consists of all those arcs that have one end
vertex in V1 and the other in V2 . All arcs in the cut-set can be directed from V1 to V2 , or from
V2 to V1 , or some arcs can be directed from V1 to V2 and others from V2 to V1 . A cut-set in
which all arcs are oriented in the same direction is called a directed cut-set.
Consider the digraph of Figure 11.18. A spanning tree T = {a, d, f , h, k} is shown by
bold lines. Here, rank = 5, nullity = 4. The chord set with respect to T is {b, c, e, g}.
Fundamental cycles with respect to T are d f e (semi cycle), dkhc (semi cycle), khg (semi
cycle) and adkhb (directed cycle). The fundamental cut-sets with respect to T are ab, bcde,
e f , bcgk and bcgh.

Fig. 11.18

11.7 Matrices A, B and C of Digraphs

The matrices associated with a digraph are almost similar to those discussed for an undirected graph, with the difference that in matrices of digraphs consist of 1, 0, 1 instead of
only 0 and 1 for undirected graphs. The numbers 1, 0, 1 are real numbers and their addition and multiplication are interpreted as in ordinary arithmetic, not modulo 2 arithmetic
as in undirected graphs. Thus the vectors and vector spaces associated with a digraph and
its subdigraphs are over the field of all real numbers, but not modulo 2.
Incidence matrix:
The incidence matrix of a digraph with n vertices, m arcs and no self-loops is an n m
matrix A = [ai j ], whose rows correspond to vertices and columns correspond to arcs, such
that

Graph Theory

315

1,
ai j =
1,

0,

i f jth arc is incident out o f ith vertex ,

i f jth arc is incident into ith vertex ,
i f jth arc is not incident on ith vertex .

For example, consider the digraph of Figure 11.19.

Fig. 11.19

The incidence matrix is given by

a

A=

v1
v2
v3
v4
v5

1
1
0
0
0

1
0
1 1
0
1
0
0
0
0

1
1
0
0
1 0
0 1
0
0

0
0
0
0
1 1
1 1
0
0

0
0
0
1
1

Now, since the sum of each column is zero, the rank of the incidence matrix of a digraph
of n vertices is less than n. The proof of the following result is almost similar to the result
in undirected graphs.
Theorem 11.16 If A(D) is the incidence matrix of a connected digraph of n vertices,
then rank of A(D) = n 1.
We further note that after deleting any row from A, we get A f , the (n 1) m reduced
incidence matrix. The vertex corresponding to the deleted row is called the reference
vertex.
We now have the following result.
Theorem 11.17 The determinant of every square submatrix of A, which is the incidence
matrix of a digraph, is 1, or 1, or 0.

Proof Consider a k k submatrix M of A. If M has any column or row consisting of

all zeros, then clearly det M = 0. Also, det M = 0, if every column of M contains the two
non-zero entries, 1 and 1.

316

Digraphs

Now let det M 6= 0. Then the sum of entries in each column of M is not zero. Therefore
M has a column in which there is a single non-zero element that is either 1 or 1. Let this
single element be in the (i, j)th position in M . Thus,
det M = 1 det Mi j ,
where Mi j is the submatrix of M with its ith row and j th column deleted. The (k 1)
(k 1) submatrix Mi j is non-singular (because M is non-singular), therefore Mi j also has at
least one column with a single non-zero entry, say in the ( p, q)th position. Expanding det
Mi j about this element in the ( p, q)th position, we obtain
det Mi j = [det of non-singular (k 2) (k 2) submatrix of M ].
Repeated application of this procedure gives
det M = 1.

Unimodular matrix: A matrix is said to be unimodular if the determinant of its every

square submatrix is 1, 1, or 0.
Cycle matrix of a digraph: Let D be a digraph with m arcs and q cycles (directed cycles
or semi cycles). An arbitrary orientation (clockwise or counter clockwise) is assigned to
each of the q cycles. Then a cycle matrix B = [bi j ] of the digraph G is a q m matrix defined
by

1,
i f the ith cycle includes the jth arc , and the orientations o f the arc

and
cycle coincide ,

bi j =
1 , i f the ith cycle includes the jth arc , and the orientations o f the two

are opposite ,

0,
i f the ith cycle does not include the jth arc .

Example

Consider the digraph D given in Figure 11.20. The cycle matrix of D is

a b
0 0
0 0
B=
0 0
1 1

c
d
0 1
1 0
1 1
0 0

e
f
0
1
1 0
1 1
0
0

g
1
1
0
0

0
0
.
0
0

Graph Theory

317

Fig. 11.20

We note that the orientation assigned to each of the four cycles is arbitrary. The cycle
in the first row is assigned clockwise orientation, in the second row counter-clockwise, in
the third counter-clockwise, and in the fourth clockwise. Changing the orientation of any
cycle will simply change the sign of every non-zero entry in the corresponding row. Also
we observe that if first row is subtracted from second, the third is obtained. Thus the rows
are not all linearly independent (in the real field).
The next result gives the relation between incidence matrix and cycle matrix.
Theorem 11.18 Let B and A be respectively, the cycle matrix and the incidence matrix
of a digraph (without loops) such that the columns are arranged using the same order of
arcs. Then,
ABT = BAT = 0,

where T denotes the transposed matrix.

Proof Consider the pth row of A and the rth row of B. The rth cycle, say Zr , either (a)
does not, or (b) does possess an arc incident with vertex, say v p , represented by the pth row
of A. If (a), the product of the two rows is zero. If (b), there are exactly two arcs, say ai and
a j , of the rth cycle incident with v p .
We have the following four possibilities.
i. ai and a j are both incident towards v p ,
ii. ai and a j are both incident away from v p ,
iii. the directions of both ai and a j coincide with the orientation of Zr and
iv. the directions of both ai and a j do not coincide with the orientation of Zr .
It can be easily verified that in all these four cases, the product of the pth row of A and
the rth row of B is zero.
q

318

Digraphs

Figure 11.21 illustrates these four cases.

Fig. 11.21

Now, using Sylvesters theorem and Theorem 11.18, we can show that
rank B+ rank A = m.
If the digraph is connected, then rank A = n 1.
Therefore, rank B = m n + 1.
The following two results can be easily established.
Theorem 11.19 The non-singular submatrices of order n 1 of A are in one-one correspondence with spanning trees of a connected digraph of n vertices.
Theorem 11.20 The non-singular submatrices of B of order = m n + 1 are in oneone correspondence with the chord set (complement of the spanning tree) of the connected
digraph of n vertices and m edges.
Sign of a spanning tree: For a digraph, the determinant of the non-singular submatrix
of A corresponding to a spanning tree T has a value either 1 or 1. This is referred to as
the sign of T .
We note that the sign of a spanned tree is defined only for a particular ordering of vertices
and arcs in A, because interchanging two rows or columns in a matrix changes the sign of
its determinant. Thus the sign of a spanning tree is relative. Once the sign of one spanning
tree is arbitrarily chosen, the sign of every other spanning tree is determined as positive or
negative with respect to this spanning tree.
Number of spanning trees: The following result determines the number of spanning
trees in a connected digraph.
Theorem 11.21 If A f is the reduced incidence matrix of a connected digraph, then the
number of spanning trees in the graph is equal to det (A f . ATf ).

Proof

According to Binet-Cauchy theorem,

Graph Theory

319

det (A f . ATf ) = sum of the products of all corresponding majors of A f and ATf .
Every major of A f or ATf is zero unless it corresponds to a spanning tree, in which case
its value is 1 or 1. Since both majors of A f and ATf have the same value 1 or 1, the
product is 1 for each spanning tree.
q
Fundamental cycle matrix: The fundamental cycles each formed by a chord with
respect to some specified spanning tree, define a fundamental cycle matrix B f for a digraph.
The orientation assigned to each of the fundamental cycles is chosen to coincide with that
of the chord. Therefore B f , a m matrix can be expressed exactly in the same form as in
the case of an undirected graph,
B f = [I : Bt ],

where I is the identity matrix of order and the columns of Bt correspond to the arcs in a
spanning tree. This is illustrated in Figure 11.22.

Fig. 11.22
b

1 0 0
Here B f = 0 1 0
0 0 1

1 0
0 1
0
1

0 0 0
1 1 0 = [I : Bt ].
1 0 0

Cut-set matrix: Let D = (V, A) be a connected digraph with q cut-sets. The cut-set
matrix C = [ci j ] of D is a q m matrix in which the rows correspond to the cut-sets of D and
the columns to the arcs of D. Each cut-set is given an arbitrary orientation. Let Ri be the
ith cut-set of D and let Ri partition V into nonempty vertex sets Vi0 and Vi00 . The orientation
can be defined to be either from Vi0 to Vi00 or from Vi00 to Vi0 . Suppose that the orientation
is chosen to be from Vi0 to Vi00 . Then the orientation of an arc a j of cut-set Ri is said to be
the same as that of Ri if a j is of the form va vb , where va Vi0 and vb Vi00 and opposite,
otherwise. Then,

i f arc a j o f cut set Ri has the same orientation as Ri ,

1,
ci j =
1 , i f arc a j has the opposite orientation to Ri ,

0,
otherwise .

320

Digraphs

We have the following observations.

1. A permutation of the rows or columns corresponds to a relabelling of the cut-sets and
arcs of D respectively.
2. Rank C rank A.
3. Rank C n 1, by observation 2.
4. If the arcs of D are arranged in the same column order in B and C, then BCT = CBT = 0.
5. Rank B+ rank C m.
6. If D is weak, rank B = m n + 1 and rank C n 1.
7. Rank C = n 1, because of (3) and (4).
We observe that the removal of an arc, say a = vs vt (also called a branch) of a spanning
directed tree of D, partitions the vertices of a digraph D into two disjoint sets, say V1 and
V2 .
The cut-set created by the removal of a is said to be either (i) directed away from V1 and
towards V2 if vs V1 and vt V2 , or (ii) directed away from V1 and towards V2 if vs V2 and
vt V1 .
This type of cut-set is called fundamental cut-set. Clearly, not all the chords in Ri necessarily have the same orientation as vs vt . If vs vt is directed away from a vertex in V1 , there
may exist a chord in Ri which is directed towards a vertex in V1 . The orientation of a cutset on the basis of the direction of the branch giving rise to it constitutes a natural way of
orienting cut-sets. If all the chords of Ri are oriented as is vs vt , then Ri is said to be directed.
Consider the graph shown in Figure 11.23 with T shown by bold lines. The fundamental
cut-sets with respect to T are
branch
a3
a5
a6
a7

cut-set
{a3 , a2 }
{a5 , a4 }
{a6 , a1 , a4 }
{a7 , a1 , a2 }

Fig. 11.23

A fundamental cut-set is created from C, the cut-set matrix of a connected digraph with
the given directed spanning tree T , by deleting from C, all rows which do not correspond

Graph Theory

321

to fundamental cut-sets with respect to T . Therefore C f is an (n 1) m submatrix of C

such that each row represents a unique fundamental cut-set with respect to T .
The rows of any fundamental cut-set C f can be permuted to create a matrix of the form
C f = [Cc : In1 ], where Cc is an (n 1) (m n + 1) matrix whose columns correspond to the
chords of T and In1 is the identity matrix of order n 1 whose columns correspond to the
branches of T .
Relation between B f , C f and Ar Ar is the reduced incidence matrix in which an arbitrary row has been removed in order to make its rows linearly independent.
We have B f = [I : Bt ]

(11.7.i)

and C f = [Cc : In1 ],

(11.7.ii)

where t corresponds to the branches of T and c to the chords of T . Let the arcs be arranged
in the same order in (11.7.i) and (11.7.ii) and in Ar . Partition Ar as
Ar = [Ac : At ],

where Ac is an (n 1) (m n + 1) submatrix whose columns correspond to the chords of

T and At is an (n 1) (n 1) submatrix whose columns correspond to the branches of T .
Since ABT = 0, therefore Ar BTf = 0.

I
Thus, [Ac : At ] .. = 0, so that Ac + At BtT = 0.
BtT

Since At is non-singular, we have At1 [Ac + At BtT ] = 0.

Therefore, At1 Ac + BtT = 0 and so At1 Ac = BtT . Also, C f BTf = 0.

I
Therefore, [Cc : In1 ] .. = 0, and so Cc + BtT = 0.
BtT

Thus, Cc = BtT and so Cc = At1 Ac .

Example
lines.

Consider the digraph given in Figure 11.24, with spanning tree shown in bold

Fig. 11.24

322

Digraphs

We have

a1

a4

a2

a5

..
1 0 0 . 1

..

Ar = [Ac : At ] = 1 0 0 . 0

0 0 1 ... 0

a6

a1 a4 a2

a5

a7

0
1

0
1

a6

a7

a3

1
1

a3

..
. 0 1 1 0

..
. 1 1
0 0

.
0 0 1 .. 0 0
1 1

1 0 0

B f = [I3 : Bt ] = 0 1 0

a1

a4

a2

a5 a6 a7 a3

..
0 1 0 .

.
1 1 0 ..
C f = [Cc : I4 ] =

1 0 1 ...

.
0 0 1 ..

1
0
0
0

0 0 0

1 0 0

0 1 0

0 0 1

Note that the last row of A, corresponding to vertex v5 , has been removed to form A f . We
form linear combinations of the rows of C f to create 10 rows of C, representing all of the
cut-sets of the digraph in Figure 11.24.

C=

a1

a4

0
1
1
0
1
0
0
0
0
1

1
1
0
0
0
1
0
0
1
0

a2

a5

0
1
0
0
1 0
1 0
0
1
0
0
1 1
0
1
1
0
0
0

a6

a7

0
0
1
0
0
1
0
0
1 0
1 1
1 1
1
1
1 1
0
1

a3
0
0
0
1
0
1
0
1
0
1

c1

c2

c3

c4

c
1 c2

2 c3 + c4

c
2 c1 c3

c1 c2 + c3 c4

c2 c3
c3 c4

The above facts lead to the following observations.

1. Given Ar , we can construct B f and C f .
2. Given Br , we can construct C f .

Graph Theory

323

3. Given C f , we can construct B f .

Semipath matrix: The semipath matrix P(u, v) = [pi j ], of a digraph D = (V, A), where
u, v V , is the matrix with each row representing a distinct semipath from u to v and the
columns representing the arcs of D, in which pi j = 1, if the ith semipath contains the j th
arc, pi j = 1 if the ith semipath contains the converse of the j th arc, and pi j = 0 otherwise.
The matrix P(v3 , v5 ) for the digraph of Figure 11.25 is
a1 a2 a3 a4 a5

0
P(v3 , v5 ) = 0
1

0 1 0 0
1 0 0 1
0 0 1 0

a6

0
0
1

Fig. 11.25

We have the following observations about P.

1. If P(u, v) contains a column of all zeros, then the vertex that it represents does not
belong to any of the semipaths between u and v.
2. If P(u, v) contains a column of all unit entries, then the vertex that it represents belongs to every semipath between u and v.
3. The number of non-zero entries in any row of P(u, v) equals the number of arcs in the
semipath represented by the row.
Adjacency matrix of a digraph:
Let G be a digraph with n vertices and with no
parallel arcs. The adjacency matrix X = [xi j ] of the digraph G is an n n (0, 1)-matrix
defined by
xi j =

Example

1,
0,

i f there is an arc directed f rom ith vertex to jth vertex ,

otherwise .

Consider the digraph D of Figure 11.26.

324

Digraphs
n2

n1

n4

e
g
n5

n3
f

Fig. 11.26

The adjacency matrix of D is

v1
v2
X = v3
v4
v5

v1
0
1
0
0
0

v2
0
1
0
0
1

v3
1
0
0
0
1

v4
1
0
0
0
0

v5
0
0
1
0
0

We have the following observations about the adjacency matrix X of a digraph D.

1. X is a symmetric matrix if and only if D is a symmetric digraph.
2. Every non-zero element on the main diagonal element represents a loop at the corresponding vertex.
3. The parallel arcs cannot be represented by X and therefore X is defined only for a
digraph without parallel arcs.
4. The sum of each row equals the outdegree of the corresponding vertex and the sum
of each column equals the indegree of the corresponding vertex. The number of
non-zero entries of X equals the number of arcs in D.
5. Permutation of any rows together with a permutation of the corresponding columns
does not alter the digraph and thus the permutation corresponds to a reordering of
the vertices. Therefore two digraphs are isomorphic if and only if their adjacency
matrices differ only by such permutations.
6. If X is the adjacency matrix of a digraph D, then the transposed matrix X T is the
adjacency matrix of a digraph D obtained by reversing the direction of every arc in
D.
7. For any (0 1)-matrix Q of order n, there exists a unique digraph D of n vertices such
that Q is the adjacency matrix of D.

Graph Theory

325

Connectedness and adjacency matrix: A digraph is disconnected if and only if its

vertices can be ordered in such a way that its adjacency matrix X can be expressed as the
direct sum of two square submatrices X1 and X2 as

X1
X = ..
O

: O
: . . .
: X2

(11.7.iii)

This partitioning is possible if and only if the vertices in the submatrix X1 have no arc going
to or coming from the vertex of X2 .
Similarly, a digraph is weakly connected if and only if its vertices can be ordered in such
a way that its adjacency matrix can be expressed as

or

X1 :
X = .. :
X21 :

X1 :
X = .. :
O :

O
..
X2

X12
. . ,
X2

(11.7.iv)

(11.7.v)

where X1 and X2 are square submatrices.

Form (11.7.iv) represents the case when there is no arc going from the subdigraph corresponding to X1 to the one corresponding to X2 . Form (11.7.v) represents the case when there
is no arc going from the subdigraph corresponding to X2 to the subdigraph corresponding
to X1 .
Since a strongly connected digraph is neither disconnected nor weakly connected, a
digraph is strongly connected if and only if the vertices of D cannot be ordered such that
its adjacency matrix X is expressible in the form (11.7.iii), or (11.7.iv), or (11.7.v).
Theorem 11.22 [X k ]i j is the number of different arc sequences of k arcs from the ith vertex
to the j th vertex.

Proof

Induct on k. The result is trivially true for k = 1. Assume the result holds for

[X k1]i j . Now,
n

r=1

r=1

[X k ]i j = [X k1X]i j = [X k1]ir [X]r j = [X k1]ir xr j

n

= (number of all directed arc sequences of length k 1 from

r=1

vertex i to r ) xr j ,
by induction hypothesis.

(11.22.1)

326

Digraphs

In (11.22.1), xr j = 1 or 0, according as there is an arc from r to j . Therefore a term in the

sum (11.22.1) is non zero if and only if there is an arc sequence of length k from i to j ,
whose last arc is from r to j . If the term is non-zero, its value equals the number of such
arc sequences from i to j through r. This holds for every vertex r, 1 r n. Thus (11.22.1)
is equal to the number of all possible arc sequences from i to j .
q
It is to be noted that [X k ]i j gives the number of all arc sequences from vertex i to j and
these arc sequences can be of the following types.
1. Directed paths from i to j , that is, those arc sequences in which no vertex is traversed
more than once.
2. Directed walks from i to j , that is, those directed arc sequences in which a vertex may
be traversed more than once, but no arc is traversed more than once.
3. Those arc sequences in which an arc may also be traversed more than once.

11.8 Number of Arborescences

We now give a formula for counting the number of spanning arborescence in a labeled,
connected digraph (which of course is simple). First, we have the following definition.
Kirchoff matrix: For a digraph (simple) D of n vertices, the Kirchoff matrix is an n n
matrix K(D) = [ki j ] defined by
ki j =

d (vi ) , i = j, in degree o f the ith vertex ,

xi j , i 6= j (i, j)th entry in the ad jacency matrix , with a negative sign .

Example

Consider the digraph D given in Figure 11.27.

d

n1

n3
c

e
g

b
n2

f
n4
D

Fig. 11.27

The Kirchoff matrix of D is

Graph Theory

v1
v2
K(D) =
v3
v4

327
v1 v2
v3 v4

1
0 1
0
1
2
0 1

0 1
2 1
0 1 1
2

Clearly, the sum of the entries in each column in K is equal to zero, so that the n rows
are linearly independent. Thus, det K = 0.
Theorem 11.23 A digraph (simple) D of n vertices and n 1 arcs is an arborescence
rooted at v1 if and only if the (1, 1) cofactor of K(D) is equal to 1.

Proof
a. Let D be an arborescence with n vertices and rooted at vertex v1 . Relabel the vertices
as v1 , v2 , . . ., vn such that vertices along every path from the root v1 have increasing
indices. Permute the rows and columns of K(D) to conform with this relabelling.
Since the indegree of v1 equals zero, the first column contains only zeros. Other
entries in K(D) are

0,
ki j =
xi j ,

1,

i > j,
i < j,
i = j, i > 1.

Then the K matrix of an arborescence rooted at v1 is of the form

K(D) =

0 x12
0
1
0
0
0
0
:
0
0

x13
x23
1
0

x14
x24
x34
1

..
..
..
..
..
..

x1n
x2n
x3n

Clearly, the cofactor of the (1, 1) entry is 1, that is, det K11 = 1.
b. Conversely, let D be a digraph of n vertices and n 1 arcs, and let (1, 1) cofactor of
its K matrix be equal to 1, that is, det K11 = 1.
Since det K11 6= 0, every column in K11 has at least one non-zero entry. Therefore,
d (vi ) 1, for i = 2, 3, . . ., n.

There are only n 1 arcs to go around, therefore

328

Digraphs
d (vi ) = 1, for i = 2, 3, . . ., n, and d (v1 ) = 0.

Since no vertex in D has an indegree of more than one, if D can have any cycle at all, it
has to be a directed cycle. Suppose that such a directed cycle exists, which passes through
vertices vi1 , vi2 , ..., vir . Then the sum of the columns i1 , i2 , . . ., ir in K11 is zero. This is because
each of these columns contains exactly two non-zero entries, as 1 on the main diagonal
and 1 for the incoming arc from the vertex preceding it in the directed cycle. Thus the r
columns in K11 are linearly dependent. So, det K11 = 0, a contradiction. Therefore, D has
no cycles.
If D has n 1 arcs and no cycles, it must be a tree. Since in this tree d (v1 ) = 0 and

d (vi ) = 1, for i = 2, 3, . . ., n, D is an arborescence rooted at vertex v1 .

The arguments in (a) and (b) are valid for an arborescence rooted at any vertex vq . Any
reordering of the vertices in D corresponds to identical permutations of rows and columns
in K(D). Such permutations do not alter the value or sign of the determinant.
q
Theorem 11.24 If K(D) is the Kirchoff matrix of a (simple) digraph D, then the value of
the (q, q) cofactor of K(D) is equal to the number of arborescences in D rooted at the vertex
vq .

Proof The proof depends on the result of Theorem 11.23 and on the fact that the determinant of a square matrix is a linear function of its columns. In particular, if P is a square
matrix consisting of n column vectors, each of dimension n, that is
P = [p1 , p2 , . . ., (pi + p0i ), . . ., pn ],

then det P = det [p1 , p2, . . ., pi, . . ., pn ]+ det [p1 , p2 , . . ., p0i , . . ., pn ].

(11.24.1)

In digraph D, suppose that vertex v j has indegree d j . The j th column of K(D) can be
regarded as the sum of d j different columns, each corresponding to a digraph in which v j
has indegree one. And then (11.24.1) can be repeatedly applied. After this, splitting of
columns can be carried out for each j , j 6= q and det Kqq (D) can be expressed as a sum of
determinants of subdigraphs, that is
det Kqq (D) = det Kqq (D0 ),
D0

(11.24.2)

where D0 is a subdigraph of D with the following properties.

1. Every vertex in D0 has an indegree of exactly one except vq .
2. D0 has n 1 vertices and hence n 1 arcs.
From Theorem 11.23,
det Kqq (D0 ) =

1,
0,

i f and only i f D0 is an arborescence rooted at q ,

otherwise .

Thus the summation in (11.24.2) carried over all D0 s equals the number of arborescences
rooted at vq .
q

Graph Theory

329

11.9 Tournaments
A tournament is an orientation of a complete graph. Therefore in a tournament each pair of
distinct vertices vi and v j is joined by one and only one of the oriented arcs (vi , v j ) or (v j , vi ).
If the arc (vi , v j ) is in T , then we say vi dominates v j and is denoted by vi v j . The relation
of dominance thus defined is a complete, irreflexive, antisymmetric binary relation. Figure
11.28 displays all tournaments on three and four vertices.

Fig. 11.28

Definition: A triple in a tournament T is the subdigraph induced by any three vertices.

A triple (u, v, w) in T is said to be transitive if whenever (u, v) A(T ) and (v, w) A(T ),
then (u, w) A(T ). That is, whenever u v and v w, then u w.
Definition: A bipartite tournament is an orientation of a complete bipartite graph. A
k-partite tournament is an orientation of a complete k-partite graph. Figure 11.29 displays
a bipartite and a tripartite tournament.

Fig. 11.29

Theorem 11.25 If v is a vertex having maximum outdegree in the tournament T , then

for every vertex w of T there is a directed path from v to w of length at most 2.

330

Digraphs

Proof Let T be a tournament with n vertices and let v be a vertex of maximum outdegree
in T . Let d + (v) = m and let v1 , v2 , . . ., vm be the vertices in T such that there are arcs from v
to vi , 1 i m. Since T is a tournament, there are arcs from the remaining n m1 vertices,
say u1 , u2 ,. . ., unm1 to v. That is, there are arcs from u j to v, 1 j n m 1 (Fig. 11.30).

Fig. 11.30

Then for each i, 1 i m, the arc from v to vi gives a directed path of length 1 from
v to vi . We now show that there is a directed path of length 2 from v to u j for each j ,
1 j n m 1.
Given such a vertex u j , if there is an arc from vi to u j for some i, then vvi u j is a directed
path of length 2 from v to u j . Now, let there be a vertex uk , 1 k n m 1, such that no
vertex vi , 1 i m, has an arc from vi to uk . Since T is tournament, there is an arc from uk
to each of the m vertices vi . Also, there is an arc from uk to v and therefore d + (uk ) m + 1.
This contradicts the fact that v has maximum outdegree with d + (v) = m. Thus each u j must
have an arc joining it from some vi and the proof is complete by using the directed path
vvi u j .
q
Let T be a tournament with n vertices and let v be any vertex of T . Then T v is
the digraph obtained from T by removing v and all arcs incident with v. Clearly, any two
vertices of T v are joined by exactly one arc, since these two vertices are joined by exactly
one arc in T . Thus T v is again a tournament.
Definition: A directed Hamiltonian path of a digraph D is the directed path in D that
includes every vertex of D exactly once.
The following result, due to Redei [216], shows that a tournament contains a direct Hamiltonian path.
Theorem 11.26 (Redei)

Every tournament T has a directed Hamiltonian path.

Proof Let T be a tournament with n vertices. We induct on n. When n = 1, 2, or 3, the

result is trivially true (Fig. 11.31).

Graph Theory

331

Fig. 11.31

Let n 4. Assume that the result is true for all tournaments with fewer than n vertices.
Let v be any vertex of T . Then T v is a tournament with n 1 vertices and by induction
hypothesis has a directed Hamiltonian path, say P = v1 v2 . . . vn1 .
In case there is an arc from v to v1 , then P1 = vv1 v2 . . .vn1 is a directed Hamiltonian
path in T . Similarly if there is an arc from vn1 to v, then P2 = v1 v2 . . . vn1 v is a directed
Hamiltonian path in T .
Now, assume there is no arc from v to v1 and no arc from vn1 to v. Then there is at least
one vertex w on the path P with the property that there is an arc from w to v and w is not
vn1 . Let vi be the last vertex on P having this property, so that the next vertex vi+1 does not
have this property. Then there is an arc from vi to v and an arc from v to vi+1 , as shown in
Figure 11.32. Thus Q = v1 v2 . . .vi v vi+1 vi+2 . . . vn1 is a directed Hamiltonian path in T . q
n1

n2

ni

ni + 1

nn 1

Fig. 11.32

Definition: A directed Hamiltonian cycle in a digraph D is a directed cycle which includes every vertex of D. If D contains such a cycle, then D is called Hamiltonian.
The next two results are due to Camion [44].
Theorem 11.27 (Camion)
cycles of length 3, 4,. . ., n.

A strongly connected tournament T on n vertices contains

Proof First we show that T contains a cycle of length three. Let v be any vertex of T . Let
W denote the set of all vertices w of T for which there is an arc from v to w. Let Z denote
T
the set of all vertices z of T for which there is an arc from z to v. We note that W Z = ,
since T is a tournament.
Since T is strongly connected, W and Z are both nonempty. For, if W is empty, then
there is no arc going out of v, which is impossible because T is strongly connected and the
same argument can be used for Z . Again, because T is strongly connected, there is an arc
in T going from some w in W to some z in Z . This gives the directed cycle v w z v of length
3 (Fig. 11.33).

332

Digraphs

W
w

z
Z

Fig. 11.33

Now induct on n. Assume T has a cycle C of length k, where k < n and k 3 and let this
cycle be v1 v2 . . . vk v1 . We show that T has a cycle of length k + 1.
Let there be a vertex v not on the cycle C, with the property that there is an arc from v
to vi and an arc from v j to v for some vi , v j on C. Then there is a vertex vi on C with an arc
from vi1 to v and an arc from v to vi . Therefore, C1 = v1 v2 . . . vi1 v vi vi+1 . . . vk v1 is a cycle
of length k + 1 (Fig. 11.34).

Fig. 11.34

If no vertex exists with the above property, then the set of vertices not contained in the
cycle can be divided into two distinct sets W and Z , where W is the set of vertices w such
that for each i, 1 i k, there is an arc from vi to w and Z is the set of vertices z such that
for each i, 1 i k, there is an arc from z to vi . If W is empty then the vertices of C, and the
vertices of Z together make up all the vertices in T . But, by definition of Z , there is no arc
from a vertex on C to a vertex in Z , a contradiction, because T is strongly connected. Thus
W is nonempty. A similar argument shows that Z is nonempty. Again, since T is strongly
connected, there is an arc from some w in W to some z in Z . Then C1 = v1 w z v3 v4 . . . vk v1 is
a cycle of length k + 1 (Fig. 11.35). This completes the proof.
q

Graph Theory

333

Fig. 11.35

Theorem 11.28 (Camion)

connected.

A tournament T is Hamiltonian if and only if it is strongly

Proof Let T have n vertices. If T is strongly connected, then by Theorem 11.27, T has a
cycle of length n. Such a cycle is a Hamiltonian cycle, since it includes every vertex of T .
Hence T is Hamiltonian.
Conversely, let T be Hamiltonian with Hamiltonian cycle C = v1 v2 . . .vn v1 . Then given
any vi , v j with i j , in the vertex set of T , v j v j+1 . . . vi is a path P1 from v j to vi while
vi vi+1 . . . vn1 vn v1 . . . v j1 v j is a path P2 from vi to v j (Fig. 11.36). Thus each vertex is
reachable from any other vertex and so T is strongly connected.
q
n1
nn
ni 1

n2
P2
nj 1
P1
nj
nj + 1

Fig. 11.36

11.10 Exercises
1. Prove that the converse of a strong digraph is also strong.
2. Show that D , the condensation of any digraph D, is cyclic.
3. Prove that the converse of a unilateral digraph is unilateral.
4. Show that the transmitters, receivers and isolates of a digraph D retain their properties
in D .
5. Prove that the only acyclic digraph is K1
= K1 .
6. Prove that an acyclic digraph without isolates has a transmitter and a receiver.

334

Digraphs

7. Show that every vertex v of a non-trivial digraph D has total even degree if and only
if D is the union of arc-disjoint cycles.
8. Prove that every arc in a digraph belongs either to a directed cycle or a directed cutset.
9. Prove that the digraph D = (V, A) with dv+ > 0 for all v V , has a cycle.
10. Prove that the digraph distance satisfies the triangle inequality.
11. Prove that every Eulerian digraph is strong. Is the converse true?
12. If E |G| is the number of Euler lines in an n-vertex Euler digraph D, show that
2n1 .E |G| is the number of Euler lines in L(D).
13. If D is a digraph with an odd number of vertices and if each vertex of D has an odd
outdegree, prove that there is an odd number of vertices of D with odd indegree.
14. Prove by induction on n that for each n 1, there is a simple digraph D with n vertices
v1 , v2 , . . ., vn such that dv+i = i 1 and dvi = n i for each i = 1, 2,. . ., n.
15. Prove that no strictly weak digraph contains a vertex whose removal results in a
strong digraph.
16. There exists a digraph with outdegree sequence [s1 , s2 , . . ., sn ], where n 1 s1 s2
. . . sn and indegree sequences [t1, t2 , . . ., tn ] where every t j n 1 if and only if
si = ti and for each integer k < n,
k

i=1

i=1

i=k+1

si min {k 1, ti } + min {k, ti } .

17. If X is the adjacency matrix of the edge digraph of a complete symmetric digraph,
then X 2 + X has all entries 1.
18. Let T be any tournament. Prove that the converse of T and the complement of T are
isomorphic.
19. Prove that a tournament is transitive if and only if it has a unique Hamiltonian path.
20. Prove that if a simple digraph D has a cycle of length three, then it is not transitive.
21. Prove that a tournament is transitive if and only if it has no directed cycles.

12. The Four Colour Theorem

Map of Madhya Pradesh and adjoining states in India, circa 2000

Introduction
The famous four colour theorem seems to have been first proposed by Mbius in 1840,
later by DeMorgan and the Guthrie brothers in 1852 and again by Cayley in 1878. The
problem of proving this theorem has a distinguished history, details of which abound in
the literature. The statement of the theorem may be introduced as follows. In colouring a
geographical map it is customary to give different colours to any two countries that have
a segment of their boundaries in common. It has been found empirically that any map,
no matter how many countries it contains nor how they are situated, can be so coloured
by using only four different colours. The map of India requires four colours in the states
bordering Madhya Pradesh. The fact that no map was ever found whose colouring requires
more than four colours suggests the mathematical theorem.
The four colour theorem: For any subdivision of the plane into non-overlapping regions, it is always possible to mark each of the regions with one of the numbers 0, 1, 2, 3,
in such a way that no two adjacent regions receive the same number.

336

The Four Colour Theorem

Steps of the proof We shall outline the strategy of the new proof given in this chapter. In
section 12.1 on Map Colouring, we define maps on the sphere and their proper colouring.
For purposes of proper colouring it is equivalent to consider maps on the plane and furthermore, only maps which have exactly three edges meeting at each vertex. Lemma 12.1
proves the six colour theorem using Eulers formula, showing that any map on the plane
may be properly coloured by using at most six colours. We may then make the following
basic definitions.
Define N to be the minimal number of colours required to properly colour any map

from the class of all maps on the plane.

Based on the definition of N , select a specific map m(N ) on the plane, which requires
no fewer than N colours to be properly coloured.
Based on the definition of the map m(N), select a proper colouring of the regions of
the map m(N) using the N colours 0, 1, . . ., N 1.

The whole proof works with the fixed number N , the fixed map m(N) and the fixed proper
colouring of the regions of the map m(N). In Section 12.2, we define Steiner Systems and
prove Tits inequality and its consequence that if a Steiner system S(N + 1, 2N, 6N) exists,
then N cannot exceed 4. Now the goal is to demonstrate the existence of such a Steiner
system. In Section 12.3, we define Eilenberg Modules. The regions of the map m(N) are
partitioned into disjoint, nonempty equivalence classes 0, 1, . . ., N 1 according to the
colour they receive. This set is given the structure of the cyclic group ZN = {0, 1, . . ., N
1} under addition modulo N . We regard ZN as an Eilenberg module for the symmetric
group S3 on three letters and consider the split extension ZN ]S3 corresponding to the trivial
representation of S3 . By Section 12.4 on Hall Matchings, we choose a common system of
coset representatives for the left and right cosets of S3 in the full symmetric group on |ZN ]S3 |
letters. In Section 12.5, for each such common representative and for each ordered pair of
elements of S3 , on Riemann surfaces, we establish a certain action of the two-element
cyclic group on twelve copies of the partitioned map m(N) by using the twenty-fourth root
function of the sheets of the complex plane. In Section 12.6, we give the details of the Main
Construction using this action. The 6N elements of ZN ]S3 are regarded as the set of points
and Lemma 12.23 builds the blocks of 2N points with every set of N + 1 points contained
in a unique block. This constructs a Steiner system S(N + 1, 2N, 6N) which implies by Tits
inequality that N cannot exceed 4, completing the proof.

12.1 Map Colouring

A map on the sphere is a subdivision of the surface into finitely many regions. A map is
regarded as properly coloured if each region receives a colour and no two regions having a
whole segment of their boundaries in common receive the same colour. Since deformations
of the regions and their boundary lines do not affect the proper colouring of a map, we
shall confine ourselves to maps whose regions are bounded by simple closed polygons.
For purposes of proper colouring, this is equivalent to considering maps drawn on the

Graph Theory

337

plane. Any map on the sphere may be represented on the plane by boring a small hole
through the interior of one of the regions and deforming the resulting surface until it is flat.
Conversely, by a reversal of this process, any map on the plane may be represented on the
sphere. Furthermore, it suffices to consider 3-regular maps, i.e., maps with exactly three
edges meeting at each vertex, by the following argument. Replace each vertex at which
more than three edges meet by a small circle and join the interior of each such circle to
one of the regions meeting at the vertex. A new map is obtained which is 3-regular. If
this new map can be properly coloured by using at most n colours, then by shrinking the
circles down to points, the desired colouring of the original map using at most n colours is
obtained. Example of a map that requires four colours to be properly coloured is shown in
Figure 12.1.

Fig. 12.1

Lemma 12.1
colours.

A map that requires four colours to be properly coloured

Any map on the sphere can be properly coloured by using at most six

Proof Assume that the given map is 3-regular. First show that there must be at least one
region whose boundary is a polygon with fewer than six sides as follows. Let E be the
number of edges, V the number of vertices, F the number of regions and Fn the number of
regions whose boundary is a polygon with n sides in the given map. Then
F = F2 + F3 + F4 + . . .
2E = 3V = 2F2 + 3F3 + 4F4 + . . .

since a region bounded by n edges has n vertices and each vertex belongs to three regions.
By Eulers formula V E + F = 2

338

The Four Colour Theorem

6V 6E + 6F
4E 6E + 6F
6F 2E
6(F2 + F3 + F4 + . . .) (2F2 + 3F3 + 4F4 + . . .)
4F2 + 3F3 + 2F4 + F5 + 0+ negative terms

=
=
=
=
=

12
12
12
12
12.

Hence at least one of F2 , F3 , F4 , F5 must be positive. Now, if a region R with fewer than six
sides is removed from the map and the resulting map coloured with six colours inductively,
there is always a colour left for R.
q
By Lemma 12.1 the minimal number of colours required to properly colour any map
from the class of all maps on the sphere is a well-defined natural number. We may now
make the following basic definition.
Definition
Define N to be the minimal number of colours required to properly colour any map

from the class of all maps on the sphere. That is, given any map on the sphere, no
more than N colours are required to properly colour it and there exists a map on the
sphere which requires no fewer than N colours to be properly coloured.

Based on the definition of N , select a specific map m(N) on the sphere, which requires
no fewer than N colours to be properly coloured.
Based on the definition of the map m(N), select a proper colouring of the regions of
the map m(N) using the N colours 0, 1, . . ., N 1.

The natural number N , the map m(N) and the proper colouring of the regions of m(N)
is fixed for all future reference. By the example shown in Figure 12.1 and Lemma 12.1,
4 N 6. The goal is to show that N 4.

12.2 Steiner Systems

A Steiner system S(t, k, v) is a set P of points together with a set B of blocks such that
There are v points.
Each block consists of k points.
Every set of t points is contained in a unique block.

Note that by definition t, k, v are non-negative integers with t k v. Steiner systems

with v = k (only one block that contains all the points) or k = t (every k-element subset of
points is a block) are called trivial. An example of a nontrivial Steiner system is S(5, 8, 24)
due to Witt, whose blocks are known as Golay codewords of weight eight. The group of
automorphisms of S(5, 8, 24) (permutations of points which permute blocks) is the largest
of the Mathieu groups, M24 .

Graph Theory

339

Lemma 12.2 (J. Tits)

If there exists a nontrivial Steiner system S(t, k, v) then

v (t + 1)(k t + 1).

Proof First show that there exists a set X0 of t +1 points that is not contained in any block
as follows. Suppose that for every set X of t + 1 points there is a block BX that contains it.
Then this block BX must be the unique block containing X , since X has more than t points.
Let b denote the total number of blocks. Count in two ways the number of pairs (X, BX )
where X is a set of t + 1 points and B X is the unique block containing it. One finds that


v
t +1

=b 

k
t +1

Count in two ways the number of pairs (Y, BY ) where Y is a set of t points and BY is the
unique block containing it, by definition of a Steiner system. One finds
 v =b  k 
t

Hence
v

t+1
 
k
t+1

k
t+1

=   =b
k
t

and it follows that b = 1 and k = v, contradicting the hypothesis that the Steiner system is
nontrivial. Now choose a fixed set X0 of t + 1 points that is not contained in any block. For
each set Z of t points contained in X0 there is a unique block BZ containing Z . Each such
BZ has k t points not in X0 and any point not in X0 is contained in atmost one such BZ
since two such blocks already have t 1 points of X0 in common. The union of the blocks
BZ contains (t + 1) + (t + 1)(k t) points and this number cannot exceed the total number of
points v.
q
Recall the definition from section 12.1 that N is the minimal number of colours required
to properly colour any map from the class of all maps on the sphere and m(N) is a specific
map which requires all of the N colours to properly colour it. The regions of the map m(N)
have been properly coloured using the N colours 0, 1, . . ., N 1. From the map m(N) and
its fixed proper colouring, we shall construct a Steiner system S(N + 1, 2N, 6N) by defining
the points and blocks in a certain way. The next lemma shows that this construction would
force N 4.

340

The Four Colour Theorem

Lemma 12.3 Referring to the definition of N in section 12.1, if there exists a Steiner
system S(N + 1, 2N, 6N), then N 4.

Proof Since 4 N 6 by definition, the Steiner system is nontrivial if it exists. By

Lemma 12.2, 6N (N + 1 + 1)(2N N 1 + 1) = (N + 2)N . Hence 6 N + 2 and it follows
that 4 N .
q
Now the goal is to demonstrate the existence of the Steiner system S(N + 1, 2N, 6N)
based upon the definition of the map m(N).

12.3 Eilenberg Modules

Let G be a group with identity element e and let Z denote the integers. The integral group
algebra (ZG, +, ) is a ring whose elements are formal sums

ng g
gG

with g in G and ng in Z such that ng = 0 for all but a finite number of g. Addition and
multiplication in ZG are defined by

ngg + mg g = (ng + mg)g

gG

gG

gG

ngg mg g = h G(ngh 1mh)g

gG

gG

gG nG

The element n of Z is identified with the element n e of ZG and the element g of G is

identified with the element 1 g of ZG, so that Z and G are to be regarded as subsets of ZG.
The underlying additive abelian group (ZG, +) is the direct sum of copies of the integers Z
indexed by elements of G. If Q is a subgroup of G then ZQ is a subring of ZG in a natural
way. For each element g of G, the right multiplication R(g) : G G; x xg and the left
multiplication L(g) : G G; x gx are permutations of the set G. Denote the group of all
permutations of the set G by Sym(G). Then
R : G Sym(G); g R(g)
L1 : G Sym(G); g L1 (g) = L(g1 )

are embeddings of the group G in Sym(G). The images R(G), L1 (G) are called the Cayley
right and left regular representations of G, respectively. The subgroup of Sym(G) generated by the set R(G) L1 (G) = {R(g), L(g1 )|g G} is called the combinatorial multiplication group Mlt(G) of G. There is an exact sequence of groups

Graph Theory

341

T
1 C(G) G G Mlt(G) 1

where T (x, y) = R(x)L(y1 ) and c = (c, c) for an element c of the center C(G) of G. If Q
is a subgroup of G then the relative combinatorial multiplication group MltG(Q) of Q in G
is the subgroup of Mlt(G) generated by the set R(Q) L1 (Q) = {R(q), L1 (q)|q Q}. The
orbits of the action of MltG(Q) on G are the double cosets QgQ of the subgroup Q in G.
The stabilizer of the identity element e is the subgroup of MltG(Q) generated by the set
{T (q) = R(q)L1 (q)|q Q}. A representation of the group Q is usually defined as a module, i.e., an abelian group (M, +), for which there is a homomorphism T : Q Aut(M, +)
showing how Q acts as a group of automorphisms of the module. Another approach due to
Eilenberg views a module M for the group Q as follows. The set M Q equipped with the
multiplication
(m1 , q1 ) (m2 , q2 ) = (m1 + m2 T (q1 ), q1 q2 )

becomes a group M]Q known as the split extension of M by Q. There is an exact sequence
of groups

1 M M]Q Q 1

With : M M]Q; m (m, e) and : M]Q Q; (m, q) q split by 0 : Q M]Q; q (0, q).
The group action T is recovered from the split extension M]Q by mT (q) = m R((0, q))L1 ((0, q))
for m in M and q in Q. In this context we shall call M an Eilenberg module for the
group Q. For example, the trivial representation for the group Q is obtained by defining
T : Q Aut(M, +); q 1M , the identity automorphism of (M, +) and the corresponding
split extension is the group direct product M Q. The Cayley right regular representation
for the group Q is obtained by defining
T : Q Aut(ZQ, +); q ( ng g ng gR(q)).
gQ

gQ

Here, T (q) = R(q)L1 (q) with L1 (q) is acting trivially on the module elements and R(q)
is acting as the usual right multiplication. The split extension ZQ]Q has multiplication given
by
(m1 , q1 ) (m2 , q2 ) = (m1 + m2 R(q1 ), q1 q2 )

for m1 , m2 in ZQ and q1 , q2 in Q.
Referring to the definition in section 12.1, N is the minimal number of colours required
to properly colour any map from the class of all maps on the sphere and m(N) is a specific
map that requires all of N colours to be properly coloured. Note that m(N) has been properly
coloured by using the N colours 0, 1, . . ., N 1 and this proper colouring is fixed. The set of
regions of m(N) is then partitioned into subsets 0, 1, . . ., N 1 where the subset m consists

342

The Four Colour Theorem

of all the regions which receive the colour m. Note that the subsets 0, 1, . . ., N 1 are
each nonempty (since m(N) requires all of the N colours to be properly coloured) and form
a partition of the set of regions of m(N) (by virtue of proper colouring). Identify the set
{0, 1, . . ., N 1} with the underlying set of the N -element cyclic group ZN under addition
modulo N . Let S3 denote the symmetric group on three letters, identified with the dihedral
group of order six generated by , where | | = 3 and | | = 2.
Lemma 12.4
phism

(ZN , +) is an Eilenberg module for the group S3 with the trivial homomor-

T1 : S3 Aut (ZN , +); a 1 ZN

where 1 ZN denotes the identity automorphism of ZN . The corresponding split extension

ZN ]S3 has multiplication given by
(m1 , 1 ) (m2 , 2 ) = (m1 + m2 , 1 2 )

and is a group isomorphic to the direct product ZN S3 .

Proof

follows from definition.

Referring to section 12.2, the goal is to construct a Steiner system S(N + 1, 2N, 6N). We
shall take the point set of the Steiner system to be the underlying set of the split extension
ZN ]S3 . The following lemma is used in section 12.5.
Lemma 12.5 Let (Z(ZN ]S3 ), +) and (ZS3 , +) denote the underlying additive groups of the
integral group algebras Z(ZN ]S3) and ZS3 , respectively. Then (Z(ZN ]S3 ), +) is an Eilenberg
module for the group (ZS3 , +) with the trivial homomorphism
T2 : (ZS3 , +) Aut (Z(ZN ]S3), +); n 1 Z(ZN ]S3 )
S3

where 1 Z(ZN ]S3) denotes the identity automorphism of (Z(ZN ]S3 ), +). The corresponding
split extension Z(ZN ]S3)]ZS3 has multiplication given by
(

(m, )ZN ]S3

=(

n(m, )(m, ), n )(

(m, )ZN ]S3

S3

(m, )ZN ]S3

n0(m, ) (m, ), n0 )
S3

(nm, ) + n0(m ) )(m, ), (n + n0 ) )

S3

and is a group isomorphic to the direct product Z(ZN ]S3 ) ZS3 , +).

Proof

follows from definition.

Graph Theory

343

12.4 Hall Matchings

Let be a bipartite graph with vertex set V = X Y and edge set E (every edge has one end
in X and the other end in Y ). A matching from X to Y in is a subset M of E such that no
vertex is incident with more than one edge in M . A matching M from X to Y in is called
complete if every vertex in X is incident with an edge in M . If A is a subset of V then let
ad j (A) denote the set of all vertices adjacent to a vertex in A.
Lemma 12.6 (P. Hall) If |ad j(A)| |A| for every subset A of X , then there exists a complete matching from X to Y in .

Proof A matching from X to Y in G with |M| = 1 always exists by choosing a single edge
in E . Let M be a matching from X to Y in G with m edges, m < |X|. Let x0 X such that x0
is not incident with any edge in M . Since |ad j({x0 })| 1, there is a vertex y1 adjacent to
x0 by an edge in E/M . If y1 is not incident with an edge in M , then stop. Otherwise let x1
be the other end of such an edge. If x0 , x1 , . . ., xk and y1 , . . ., yk have been chosen, since
|ad j({x0, x1 , . . ., xk })| k + 1, there is a vertex yk+1, distinct from y1 , . . ., yk , that is adjacent
to at least one vertex in {x0 , x1 , . . ., xk }. If yk + 1 is not incident with an edge in M then stop.
Otherwise let xk + 1 be the other end of such an edge. This process must terminate with
some vertex, say yk +1. Now build a simple path from yk +1 to x0 as follows. Start with yk +1
and the edge in E/M joining it to, say xi1 , with i1 < k + 1. Then add the edge in M from xi1 to
yi1 . By construction yi1 is joined by an edge in E/M to some xi2 with i2 < i1 . Continue adding
edges in this way until x0 is reached. One obtains a path yk+1 , xi1 , yi1 , xi2 , yi2 , . . ., xir , yir , x0
of odd length 2r + 1 with the r + 1 edges {yk+1 , xi1 }, {yi1 , xi2 }, . . ., {yir , x0 } in E/M and the
r edges {xi1 , yi1 }, . . ., {xir , yir } in M . Define
M 0 = (M/{{xi1 , yi1 }, . . ., {xir , yir }}) {{yk+1, xi1 }, {yi1 , xi2 }, . . ., {yir , x0 }}.

Then M 0 is a matching from X to Y in , with |M 0 | = |M| r + r + 1 = |M| + 1. Repeating

this process a finite number of times must yield a complete matching from X to Y in .
Lemma 12.7 Referring to section 12.3, let Sym(ZN ]S3 ) denote the group of all permutations of the underlying set of the split extension ZN ]S3 of Lemma 12.4. Then S3 embeds in
Sym((ZN ]S3 ) via the Cayley right regular representation.

Proof Note that S3 = {(0, )| S3 } is a subgroup of ZN ]S3 . Since S3 embeds in Sym(S3 )

via the Cayley right regular representation R( ) and Sym(S3 ) is a subgroup of Sym(ZN ]S3 ),
the lemma follows.
Lemma 12.8 By Lemma 12.7, regard S3 as a subgroup of Sym(ZN ]S3 ). There exists a
common system of coset representatives 1 , . . ., k such that {1 S3 , . . ., k S3 } is the family
of left cosets of S3 in Sym(ZN ]S3 ) and {S3 1 , . . ., S3 k } is the family of right cosets of S3 in
Sym(ZN ]S3 ).

344

The Four Colour Theorem

Proof By Lagranges theorem the left cosets of S3 partition Sym(ZN ]S3 ) into k = [Sym(ZN ]S3 ) :
S3 ] disjoint nonempty equivalence classes of size |S3 | = 6. The same is true of the right
cosets. Define a bipartite graph with vertices X Y where X = {1 S3 , . . ., k S3 } is the family of left cosets of S3 in Sym(ZN ]S3 ) and Y = {S3 10 , . . ., S3 k0 } is the family of right cosets of
S3 in Sym(ZN ]S3) with an edge {i S3 , S3 0j } if and only if i S3 and S3 0j have nonempty intersection. For any subset A = { i1 S3 , . . ., ir S3 } of X , one has i1 i1 S3 , . . ., ir ir S3
and there exist distinct j1 , . . ., jr such that i1 S3 0 j1 , . . ., ir S3 0 jr . Hence in the graph
, |ad j(A)| |A|. Halls hypothesis of Lemma 12.6 is satisfied and there exists a complete
matching from X to Y in . This is precisely the statement that a common system of coset
representatives 1 , . . ., k exists.

12.5 Riemann Surfaces

Let C denote the complex plane. Consider the function C C ; z w = zn , where n 2.
There is a one-to-one correspondence between each sector
{z|(k 1)2 /n < arg z < k2 /n}(k = 1, . . ., n)

and the whole w-plane except for the positive real axis. The image of each sector is obtained
by performing a cut along the positive real axis; this cut has an upper and a lower edge.
Corresponding to the n sectors in the z-plane, take n identical copies of the w-plane with
the cut. These will be the sheets of the Riemann surface and are distinguished by a label k,
which serves to identify the corresponding sector. For k = 1, . . ., n 1 attach the lower edge
of the sheet labelled k with the upper edge of the sheet labelled k +1. To complete the cycle,
attach the lower edge of the sheet labelled n to the upper edge of the sheet labelled 1. In a
physical sense, this is not possible without self-intersection but the idealised model shall be
free of this discrepancy. The result of the construction is a Riemann surface whose points
are in one-to-one correspondence with the points of the z-plane. This correspondence is
continuous in the following sense. When z moves in its plane the corresponding point w is
free to move on the Riemann surface. The point w = 0 connects all the sheets and is called
the branch point. A curve must wind n times around the branch point before it closes. Now
consider the n-valued relation
z=

n w.

To each w 6= 0, there correspond n values of z. If the w-plane is replaced by the Riemann

surface just constructed, then each complex number w 6= 0 is represented by n points of the
Riemann surface at superposed positions. Let the point on the uppermost sheet represent

the principal value and the other n 1 points represent the other values. Then z = n w
becomes a single-valued, continuous, one-to-one correspondence of the points of the Riemann surface with the points of the z-plane. Now recall the definition of the map m(N)
from section 12.1. The map m(N) is on the sphere. Pick a region and deform the sphere so
that both 0 and are two distinct points inside this region when the sphere is regarded as

Graph Theory

345

the extended complex plane. Using the stereographic projection one obtains the map m(N)
on the complex plane C with the region containing 0 and forming a sea surrounding
the other regions which form an island. Put this copy of C on each sheet of the Riemann
surface
corresponding to w = zn . The branch point lies in the sea. The inverse function

n
z=
w results in n copies of the map m(N) on the z-plane in the sectors
{z|(k 1)2 /n < arg z < k2 /n}(k = 1, . . ., n).

The origin of the z-plane lies in the n seas. An example with n = 4 is given in Figure
12.2.

Fig. 12.2 An example with n = 4

Referring to section 12.3, the full symmetric group Sym(ZN ]S3 ) acts faithfully on the set
ZN ]S3 . The action of an element of Sym(ZN ]S3) on an element (m, ) of ZN ]S3 will be
written as (m, ) . This action extends to the integral group algebra Z(ZN ]S3) by linearity
(

(m )ZN ]S3

n(m, ) (m, )) =

(m, )ZN ]S3

n(m, ) ((m, ) ).

Referring to Lemma 12.8, fix a common coset representative i of S3 in Sym(ZN ]S3 ) and
fix a pair ( , ) S3 S3 = Mlt(S3). There are two cases depending on whether = or
whether 6= .
Case 1

Suppose 6= . Consider the composition of the functions

C C; z t = z2 and C C; t w = t 12 .

The composite is given by the assignment

z t = z2 w = t 12 = Z 24 .

There are twenty-four superposed copies of the map m(N) on the w-Riemann surface
corresponding to the sectors

346

The Four Colour Theorem

{z|(k 1)2 /24 < arg z < k2 /24}(k = 1, . . ., 24)

on the z-plane. These are divided into two sets. The first set consists of twelve superposed
copies of the map m(N) corresponding to the sectors
{z(k 1)2 /24 < arg z < k2 /24}(k = 1, . . ., 12)

of the upper half of the z-plane which comprise the upper sheet of the t -Riemann surface.
The second set consists of twelve superposed copies of the map m(N) corresponding to the
sectors
{z|(k 1)2 /24 < arg z < k2 /24}(k = 13, . . ., 24)

of the lower half of the z-plane which comprise the lower sheet of the t -Riemann surface.
Figure 12.3 shows sheets of the t -Riemann surface.

Fig. 12.3 Sheets of the t-Riemann surface

Label the twelve sectors of the upper sheet of the t -Riemann surface by elements of
Z(ZN ]S3 )]ZS3 as shown in Figure 12.4.

Fig. 12.4 Upper sheet of the t-Riemann surface

Graph Theory

347

Label the twelve sectors of the lower sheet of the t -Riemann surface by elements of
Z(ZN ]S3 )]ZS3 as shown in Figure 12.5.

Fig. 12.5 Lower sheet of the t-Riemann surface

Referring to section 12.2, the regions of the map m(N) have been partitioned into disjoint, nonempty equivalence classes 0, 1, . . ., N 1 and this set of equivalence classes forms
the underlying set of the cyclic group ZN . Hence there are twelve copies of ZN on the upper
sheet and twelve copies of ZN on the lower sheet of the t -Riemann surface. The copies of
ZN are indexed by the elements of Z(ZN ]S3)] ZS3 which label the sectors on a particular
sheet. The branch point of the t -Riemann surface is labelled by the element (0, + ) of
Z(ZN ]S3 )]ZS3 where 0 denotes the zero element of Z(ZN ]S3 ).
Lemma 12.9 Referring to Lemma 12.8, fix a common representative i of the left and
right cosets of S3 in Sym(ZN ]S3 ). Fix a pair ( , ) S3 S3 with 6= . Referring to Lemma
12.5, define a subset T( , ) of Z(ZN ]S3 )] ZS3 as follows.
T( , ) = {((m, ), + )|(m, ) ZN ]S3 }

{(0, + )}

{((m, ), + )|(m, ) ZN ]S3}.

Referring to the preceding discussion, consider the composite function

z t = z2 w = t 12 = z24

348

The Four Colour Theorem

of the complex z-plane to the w-Riemann surface. There is a copy of the set T( , ) on
the upper sheet and a copy of the set T( , ) on the lower sheet of the t -Riemann surface
according to the labels of the sectors in Figures 12.4 and 12.5 with the branch point labelled
by the element (0, + ) of both copies. The rotation of the z-plane by p radians induces a
permutation
p : T( , ) T( , )

given by
((m, )R( )i , + )p = ((m, )i R( ), + )
(0, + )p = (0, + )
((m, )i R( ), + )p = ((m, )R( )i , + )

for all (m, ) ZN ]S3 , such that each point of the copy of T( , ) on the upper sheet moves
continuously along a circular curve that winds exactly once around the branch point, to the
point superposed directly below it on the copy of T( , ) on the lower sheet of the t -Riemann
surface.

Proof T( , ) is seen to be a well-defined subset of Z(ZN ]S3)]ZS3 by setting the appropriate coefficients to zero in a typical element as described in Lemma 12.5. Each of R( )i ,
i R( ), i R( ) and R( )i are permutations of the set ZN ]S3 and the rotation of the z-plane
by p radians clearly induces a permutation p of the set T( , ) as described.
q
Lemma 12.10 Referring to Lemma 12.9, let Sym(T( , )) denote the full permutation
group of the set T( , ). Let < p > denote the cyclic subgroup of Sym(T( , )) generated by p.
Then < p > is nontrivial and acts faithfully on the set T( , ).

Proof If p = 1, then ((0, 1)R( )i , + ) = ((0, 1)R( )i , + )p = ((0, 1)i R( ), + )

which implies that (0, 1)R( )i = (0, 1)i R( ) in Z(ZN ]S3). This is impossible since 1 6= 1
in Z. Hence p 6= 1. Since the full permutation group Sym(T( , )) acts faithfully on T( , ) , so
does its subgroup < p >.
q
Lemma 12.11 Referring to Lemma 12.9 and Lemma 12.10, let 1 : C C; z z denote
the identity and : C C; z z denote the rotation through an angle of p radians of the
z-plane. Then the two-element cyclic group {1, } acts faithfully on the set < p > as follows:
pn 1 = pn and pn = p1n , for all n in Z .

Proof The set {1, } forms a two-element cyclic group < > under function composition. To show that {1, } acts on < p > as defined, observe that (pn ) = (p1n ) =
p1(1n) = p11+n = pn = pn 1 = pn ( ), for all n in Z. To show that the action is faithful,
let {1, }. If belongs to the kernel of the action, then pn = pn for all n in Z, so that
p = p which implies that = 1, since p 6= 1 by Lemma 12.10.
q

Graph Theory

349

Lemma 12.12 Putting together Lemma 12.9, Lemma 12.10 and Lemma 12.11, there is
a well-defined action of the two-element cyclic group {1, } on the set T( , ) given by
((m, )i R( ), + ) 1 = ((m, ) i R( ), + )
(0, + ) 1 = (0, + )
((m, )R( )i , + ) 1 = ((m, )R( )i , + )

and
((m, )i R( ), + ) = ((m, )R( )i , + )
(0, + ) = (0, + )
((m, )R( )i , + ) = ((m, )i R( ), + )

for all (m, ) in ZN ]S3. This action is faithful.

Proof For each x T( , ), let Orb(x) = {xpn |n Z} denote the orbit of x under < p >. The
collection {Orb(x)|x T( , )} forms a partition of the set T( , ) as follows. Let x, y T( , ).
If z Orb(x) Orb(y) then z = xpn = ypm for some m, n Z. This implies xpnm = y y
Orb(x) Orb(y) Orb(x) and ypmn = x x Orb(y) Orb(x) Orb(y). Hence Orb(x) =
Orb(y). Also, each x T( , ) belongs to an orbit, namely x Orb(x). Hence the orbits are
disjoint, nonempty and their union is all of the set T( , ). For each fixed x T( , ), define
: Orb(x) Orb(x); (xpn ) = xp1n

Then, is well-defined since xpn = xpm xpnm = x xpm = xpn xp1m = xp1n
(xpn ) = (xpm ) . Also, is a permutation of Orb(x) with 2 = 1 since for each xpn Orb(x)
we have ((xpn ) ) = (xp1n ) = xp1(1n) = xpn . Now define
: T( , ) T( , ); (xpn ) = xp1n

orbit by orbit. Then, since {Orb(x)|x T( , )} forms a partition of T( , ), is a well-defined

permutation of T( , ) with 2 = 1, the identity permutation of T( , ). Hence, using the definition of p in Lemma 12.9, we obtain an action of the two-element cyclic group {1, } on
T( , ) as follows. For all (m, ) in ZN ]S3, define
(((m, )R( )i , + )p) 1 = ((m, )R( )i , + )p
((0, + )p) 1 = (0, + )p
(((m, )i R( ), + )p) 1 = ((m, )i R( ), + )p

350

The Four Colour Theorem

and
(((m, )R( )i , + )p) = ((m, )R( )i , + )
((0, + )p) = (0, + )
(((m, )i R( ), + )p) = ((m, )i R( ), + )

Now, using the definition of p in Lemma 12.9, the action of {1, } on T( , ) may be
rewritten as
((m, )i R( ), + ) 1 = ((m, )i R( ), + )
(0, + ) 1 = (0, + )
((m, )R( )i , + ) 1 = ((m, )R( )i , + )

and
((m, )i R( ), + ) = ((m, )R( )i , + )
(0, + ) = (0, + )
((m, )R( )i , + ) = ((m, )i R( ), + )

for all (m, ) in ZN ]S3, as in the statement of this lemma. To verify that the action of {1, }
on T( , ) is faithful, note that
1 : T( , ) T( , ); x x

: T( , ) T( , ); x x

are permutations of the set T( , ). If {1, } and belongs to the kernel of the action then
x = x for all x T( , ). Then = 1, since moves ((0, 1)i R( ), + ) to ((0, 1)R( )i , +
) which are distinct elements of Z(ZN ]S3 )]ZS3 .
Case 2 Suppose = . Note that in the labelling of the sectors of the sheets of the t Riemann surface in Figures. 12.4 and 12.5, R( ) = R( ) and + = 2 in the group algebra
ZS3 .
q

Lemma 12.13 Referring to Lemma 12.8, fix a common representative i of the left and
right cosets of S3 in Sym(ZN ]S3). Fix a pair ( , ) S3 S3 . Referring to Lemma 12.5, define
a subset T( , ) of Z(ZN ]S3 )ZS3 as follows.

Graph Theory

351

T( , ) = {((m, ), 2 )|(m, ) ZN ]S3}

{(0, 2 )}

{((m, ), 2 )|(m, ) ZN ]S3 }.

Referring to the preceding discussion, consider the composite function

z t = z2 w = t 12 = z24

of the complex z-plane to the w-Riemann surface. There is a copy of the set T( , ) on
the upper sheet and a copy of the set T( , ) on the lower sheet of the t -Riemann surface
according to the labels of the sectors in Figures 12.4 and 12.5. Note that in this case R( ) =
R( ) and + = 2 with the branch point labelled by the element (0, 2 ) of both copies.
The rotation of the z-plane by p radians induces a permutation
p : T( , ) T( , )

given by
((m, )R( )i , 2 )p = ((m, )i R( ), 2 )
(0, 2 )p = (0, 2 )
((m, )i R( ), 2 )p = ((m, )R( )i , 2 )

for all (m, ) ZN ]S3 , such that each point of the copy of T( , ) on the upper sheet moves
continuously along a circular curve that winds exactly once around the branch point, to the
point superposed directly below it on the copy of T( , ) on the lower sheet of the t -Riemann
surface. Then, p = p1 so that < p > = {1, p} is a two-element cyclic subgroup of the full
permutation group Sym(T( , ) ) and < p > acts faithfully on the set T( , ).

Proof As in the proof of Lemma 12.9, T( , ) is seen to be a well-defined subset of

Z(ZN ]S3 )]ZS3 by setting the appropriate coefficients to zero in a typical element as described
in Lemma 12.5. Both i R( ), R( )i are permutations of the set ZN ]S3 and the rotation of
the z-plane by p radians clearly induces a permutation p of the set T( , ) as described. Furthermore, it is clear from the definition that p = p 1 by chasing elements of T( , ) . Then,
< p > = {1, p} is a subgroup of Sym(T( , ) ) and < p > acts faithfully on the set T( , ).
q

352

The Four Colour Theorem

Lemma 12.14

Referring to Lemma 12.13, let 1 : C C; z z denote the identity and

: C C; z z denote the rotation through an angle of p radians of the z-plane. Then
there is a well-defined action of the two-element cyclic group {1, } on the set T( , ) given
by

((m, )i R( ), 2 ) 1 = ((m, )i R( ), 2 )
(0, 2 ) 1 = (0, 2 )
((m, )R( )i , 2 ) 1 = ((m, )R( )i , 2 )

and
((m, )i R( ), 2 ) = ((m, )R( )i , 2 )
(0, 2 ) = (0, 2 )
((m, )R( )i , 2 ) = ((m, )i R( ), 2 )

for all (m, ) in ZN ]S3. This action is faithful.

Proof The isomorphism 1 1, p of the two-element cyclic groups {1, p} and {1, }
establishes the lemma.
q

12.6 Main Construction

Let us review the final goal with a brief rsum. In section 12.1, we have defined N to
be the minimal number of colours required to properly colour any map from the class of
all maps on the sphere. We know that 4 N 6. We have chosen a specific map m(N)
on the sphere that requires all of the N colours 0, 1, . . ., N 1 to properly colour it. The
map m(N) has been properly coloured and the regions of m(N) partitioned into disjoint,
nonempty equivalence classes 0, 1, . . ., N 1 according to the colour they receive. The
set {0, 1, . . ., N 1} is endowed with the structure of the cyclic group ZN under addition
modulo N . In section 12.3, we have built the split extension ZN ]S3. The underlying set ZN ]S3
of cardinality 6N is taken to be the point set of a Steiner system S(N + 1, 2N, 6N) which will
be constructed in this section. We are required to define the blocks of size 2N and show that
every set of N + 1 points is contained in a unique block. Once this goal is achieved, Lemma
12.3 shows that N 4.
Lemma 12.15 Let ZN ]S3 denote the split extension defined in Lemma 12.4 and Sym(ZN ]S3 )
denote the full permutation group on the set ZN ]S3 .

Graph Theory

353

Define
: Sym(ZN ]S3 ) Sym(ZN ]S3 )

by
= R( )i i ( ) = .

Then, is a bijection of the set Sym(ZN ]S3 ) with itself.

Proof Referring to Lemma 12.8, is well-defined since each Sym(ZN ]S3 ) may be
written uniquely as = R( )i for some S3 and some i . Then, is a surjection because
for any Sym(ZN ]S3 ) one may also write = i R( ) uniquely for some S3 and some
i , whence R( )i i R( ) = . Since Sym(ZN ]S3 ) is a finite set, must be a bijection by
counting.
q
Lemma 12.16
G=

(

Define the set G as follows.



) (
R( )i

i R( )

Sym(ZN ]S3) =

S3
i=1, ..., k

Define multiplication in G as follows:



(1 )

R(1 ) i1
i1R(1 )




 
2
1 2
=

(2 )
(1 2 )

i.e.,


 

R(2 ) i2
R(3 ) i3
=
,
i2R(2 )
i3 R(3 )

where R(3 )3 is the unique expression for R(1 ) i1 R(2 ) i2 according to the right coset
decomposition of S3 in Sym(ZN S3 ). Then, G is a group.

Proof Referring to Lemma 12.8 and Lemma 12.15, the set G is well-defined by the
decomposition of Sym(ZN ]S3 ) into the left and right cosets of S3 by the 1 . Define
0

: Sym(ZN ]S3 ) G;
.

Then, 0 is a well-defined bijection of the set Sym(ZN ]S3 ) with G since m is a bijection by Lemma 12.15. The definition of multiplication in G mirrors the multiplication in
Sym(ZN ]S3 ) via 0 and is designed to make G a group and 0 an isomorphism.
q

354

The Four Colour Theorem

Lemma 12.17


Consider the set ZN ]S3 and let

 


R( )i
=
G.
i R( )

Define


: ZN ]s3 ZN ]S3 by

(m, ) (m, )





R( )i )
=
(m,
= (m, )R( )i .

i R( )

Define

: ZN ]S3 by





R( )i
(m, ) (m, )
= (m, )
= (m, )i R( ).
i R( )

Then






(m, )
= (m, )
for all (m, ) ZN ]S3 .

Both and are well-defined, faithful and |ZN ]S3 |-transitive right actions of the group G
on the set ZN ]S3.

Proof Referring to Lemma 12.12 and Lemma 12.12, put = 1. Working in the set T( 1, ),
for each (m, ) ZN ]S3 , we have


((M, )
, 1 + )

= ((m, )


R( )i
, 1 + )
i R( )

= ( (m, )i R( ), 1 + )
= ( (m, )R( )i , 1 + )
= ( (m, )i , 1 + )
= ( (m, )R(1)i , 1 + )

Graph Theory

355

= ( (m, )i R(1), 1 + )
= ( (m, )i , 1 + )
= ( (m, )R( )i , 1 + )
= ( (m, )


R( )i ,
, 1 + )
i R( )

= ( (m, )
, 1 + )

using the action of the two-element cyclic group {1, } on the set T(1, )n according to
Lemma 12.12 and Lemma 12.12. Hence






(m, )
= (m, al pha)
for all (m, ) ZN ]S3 .

Since the action is the usual action of Sym(ZN ]S3 ) on the set ZN ]S3 , it is faithful and
|ZN ]S3| -transitive. By the last equality, so is the action.
q
Lemma 12.18 Let (m1 , 1 ), , (mr , r ) be any r distinct elements of ZN ]S3 and let
(n1 , 1 ), . . ., (ns , s ) be any s distinct elements of ZN ]S3. Let

 

Hr, s =
G

(mi , i )
and
(n j , i)

then Hr, s is a subgroup of G.

= (n j , i ) for j = 1, . . ., s

Note that if = R( )i = 1 then i = R( )1 , so that

Proof

= i R( ) = R( )1 R( ) = 1.

Then


= (mi , i ) for i = 1, . . ., r


1
Hr,s
1

since



1
(mi , i ) = (mi , i )1 = (mi , i )
1

356

The Four Colour Theorem

for i = 1, . . ., r and
(n j , j )


1
= (n j , j )1 = (n j , j )1 = (n j , j )
1

for j = 1, . . ., s. If


and


2
Hr,s ,
2

then
(mi , i )







1
= (mi , i )




2
= (mi , i )
= (mi , i ) for i = 1, . . ., r
2

and
(n j , j )







= (n j , j )

 

1
2

1
2


2
= (n j , j )
= (n j , j ) for j = 1, . . ., s.
2

Hence





2
Hr, s
2

Since G is finite, Hr, s is a subgroup of G.

Note that ZN is embedded as the subgroup {(m, 1)|m |ZN } in ZN ]S3 and S3 is embedded
as the subgroup {(0, )| S3 } in ZN ]S3 . Since ZN ]S3 = ZN S3 is the direct product of
groups by Lemma 4, both ZN and S3 are normal subgroups. Recall the notation
S3 =< , >= {1, , 2 , , , 2 }.

Lemma 12.19

Define

 

H=
G

(m, 1)
and
(0, )

= (m, 1) for all m ZN

= (0, )

Graph Theory

357

Then given


H,

either


= (m, ) for all (m, ) ZN ]S3

(m, )

or


= (m, ) for all (m, ) ZN ]S3 .

(m, )

Proof H is a well-defined subgroup of G according to Lemma 12.18. Let



 


R( )i
=
H

i R( )

and (m, ) ZN ]S3 be given. Referring to Lemmas 12.12 and 12.14, put = 1 . Working in the set T( , ), we have
((m, )

, + )

= ((m, )


R( )i
, + )
i R( )

= ((m, )i R( ), + )
= ((m, )R( ), i + )
= ((m, )i , + )
= ((m, )i , + )
= ((m, )R( )i , + )
= ((m, )i R( ), + )
= ((m, 1)R( )i R( ), + )
= ((m, 1)R( ), + )

358

The Four Colour Theorem

= ((m, ), + )

using the definition of H and the action of the two-element cyclic group {1, } on the set

T( , ) . Hence

(m, )
= (m, ) = (m, 1 ) = (m, ).

Now since
(0, ) = (0, )

= (0, )

by hypothesis, we have = . Hence = so that either = 1 or = .

Lemma 12.20 Let H be the subgroup of G defined in Lemma 12.19. Then, H is a nontrivial group of involutions of the set ZN ]S3 . In particular, every nontrivial element of H is
of order 2.

Proof

Define

:ZN ]S3 ZN ]S3 ; (m, ) (m, ).

Then
(m, 1)

= (m, 1) = (m, 1 ) = (m, 1) for all m ZN

and






(0, )
= (0, )
= (0, ) = (0, ) = (0, ).

Now 1 6= 1, so


 

1

6
=
H,
1

hence H is nontrivial. To show that each nontrivial element of H is of order 2, let



 


R( )i
=
H

i R( )

Then, by the proof of Lemma 12.19, = 1 or = . In particular, 2 = 1. Hence, for any

(m, ) ZN ]S3

Graph Theory

359

R( ) i
(m, )
i R( )


= (m, )

2

 

R( ) i
R( ) i

i R( )
i R( )

R( ) i
= (m, )
i R( )


= (m, ( ) ) = (m, ).

Since the action of G on the set ZN ]S3 is faithful,



R( )i
i R( )

2


1
,
1

the identity element of G.

Lemma 12.21

Denote the right cosets of ZN in ZN ]S3 by

ZN , ZN , Z2N ,; ZN , ZN , Z2N .

Define Fix(H ) =










(m, ) ZN ]S3 (m )
= (m ) for all
H .

Then Fix (H ) = {(m, ) ZN ]S3 |a = 1 or = . The action of a nontrivial element

of H transposes the coset ZN with the coset ZN 2 and transposes the coset ZN with the
coset ZN 2 .

Proof


By Lemmas 12.19 and 12.20, the elements



are of two kinds:

(i)

(m, )

= (m, ) for all (m, ) ZN ]S3

in which case


  
1

360

The Four Colour Theorem

the identity element of H, and

(ii)

(m, )

= (m, ) for all (m, ) ZN ]S3

in which case


 

1

6
=
1

is an element of order 2 in H . In the second case, compute according to the cosets of ZN in

ZN ]S3 :


= (m, 1 ) = (m, 1) for all m ZN




(m, )
= (m, ) = (m, ) for all m ZN




(m, )
= (m, ) = (m, 2 ) for all m ZN




2
(m, )
= (m, ( 2 ) ) = (m, ) for all m Z N




(m, )
= (m, ( ) ) = (m, 2 ) for all m Z N




(m, 2 )
= (m, ( 2 ) ) = (m, ) for all m Z N

(m, 1)

and the lemma follows.

Lemma 12.22 Let NormG(H) denote the normaliser of H in G. The action of G on ZN ]S3
restricts to an action of NormG(H) on Fix (H) which is (|ZN | + 1)-transitive.

Proof


Let


G.

First show that

 1 




Fix
H
= Fix (H)

Graph Theory

361

as follows.
1  



H
(m, ) Fix

1  






(m, )
=
(m,

)
for
all
H

1 
1


 



(m, )
= (m, )
for all
H

1
1  






(m, )

=
(m,
for
all
H






Fix (H) (m, ) Fix (H)

.
(m, )

Now let




1 
 





NormG (H) =
G H
=H .

Then

Fix (H) = Fix

1


 




H
= Fix (H)

showing that the action restricts to an action of NormG(H) on Fix (H). Now to show that
the action of NormG (H) on Fix (H) =ZN ZN is (|ZN | + 1)-transitive, label the elements
)
of ZN as (m1 , 1 ), . . ., (mN , N ) and label (0, ) = (mN+1 , N+1 ). Let (m1 , 1 ), . . ., (mN+1 , N+1
be any |ZN | + 1 distinct points of Fix (H). It is enough to show that there exists


NormG(H)

such that
(mi , ai )

= (mi , i ) for i = 1, . . ., N + 1.

Now there exists

362

The Four Colour Theorem



such that
(mi , i )

= (mi , i ) for i = 1, . . ., N + 1.

Hence
1 


for I = 1, . . ., N + 1.
(mi , i ) = (mi , i )

Note that for every



and for i = 1, . . ., N + 1:
 1 



(mi , i )

 1  
 


= (mi , i )


 


= (mi , i )




= (mi , i )
= (mi , i ).

Hence


1










H
for
all

NormG (H).

Lemma 12.23 There exists a Steiner system S(N + 1, 2N, 6N), where the points are the
elements of the set ZN ]S3 and the set of blocks is


 



Fix (H)

G
.


Graph Theory

Proof


363

There are 6N = | ZN ]S3 | points. Each block, for a fixed



contains 2N = | ZN ZN | = | Fix (H)| =






Fix (H) 



points. Label the elements of ZN as (m1 , 1 ), . . ., (mN , N ) and label (0, ) = (mN+1 , N+1 ).
) be any N +1 distinct points of Z ]S . Then there exists
Let (m1 , 1 ), . . ., (mN+1 , N+1
N 3


such that
(mi , i )

= (mi , i ) for i = 1, . . ., N + 1.

Hence there is at least one block, namely Fix (H), that contains the points (m1 , 1 ), . . .,

). It remains to show that this is the unique block that contains the points
(mN+1 , N+1

). Suppose (m , ), . . ., (m

(m1 , 1 ), . . ., (mN+1 , N+1

1
1
N+1 , N+1 ) are contained in
Fix (H)





for
some
G.
( )
( )

Then there exist points (m

1 , 1 ), . . ., (mN+1 , N+1 ) in Fix (H) such that

(mi , i ) = (m
i , i )

for i = 1, . . ., N + 1.

By Lemma 12.22, there exists



NormG (H)

such that
(m
i ,

i ) =

(mi , i )
for i = 1, . . ., N + 1.

Hence for i = 1, . . ., N + 1

364

The Four Colour Theorem



= (m
i , i )

(mi , i )

= (mi , i )

= (mi , i )

 


(mi , i ) = (mi , i )



1




for i = 1, . . ., N + 1.

Then, by Lemma 12.17

(mi , i ) = (mi , i )



1




for i = 1, . . ., N ,

and

(mN+1 , N+1 ) = (mN+1 , N+1 )



1




Hence



1


H.

Now, H is a subgroup of NormG(H)






1



NormG (H)

1
1

 


NormG (H) = NormG (H)

Graph Theory



365



1
11
 



H
=H

1
1

 



H
=H

1
1
 
 
 


H
=
H
.

Now, using the first fact in the proof of Lemma 12.22

Fix (H)

1
= Fix

= Fix





 


1
 


= Fix (H)

This establishes the uniqueness of the block.

Theorem 12.1.
colours.

Any map on the sphere can be properly coloured by using at most four

Proof Referring to section 12.1, we have defined N to be the minimal number of colours
required to properly colour any map from the class of all maps on the sphere. Based on the
definition of N , we have selected a specific map m(N) on the sphere that requires no fewer
than N colours to be properly coloured. Based on the definition of the map m(N) we have
selected a proper colouring of its regions using the N colours 0, 1, . . . , N 1 . Working with
the fixed number N , the fixed map m(N), and the fixed proper colouring of the regions of
the map m(N), Lemma 12.23 has explicitly constructed a Steiner system S(N + 1, 2N, 6N).
Now Lemma 12.3 implies that N cannot exceed four.

13. Graph Algorithms

An algorithm is a problem-solving method suitable for implementation as a computer program. While designing algorithms we are typically faced with a number of different approaches. For small problems, it hardly matters which approach we use, as long as it
is one that solves the problem correctly. However, there are many problems, including
some problems in graph theory, for which the known algorithms take so long to compute
the solution that they are practically useless. A polynomial-time algorithm is one whose
number of computational steps is always bounded by a polynomial function of the size
of the input. The class of all such problems that have polynomial-time algorithms is denoted by P. For some problems, there are no known polynomial-time algorithms but these
problems do have nondeterministic polynomial-time algorithms: try all candidates for solutions simultaneously and for each given candidate, verify whether it is a correct solution
in polynomial-time. The class of all such problems is denoted by NP. Clearly P NP.
On the other hand, there are problems that are known to be in NP and are such that any
polynomial-time algorithm for them can be transformed into a polynomial-time algorithm
for every problem in NP. Such problems are called NP-complete. Thus, if anybody ever
finds a polynomial-time algorithm for an NP-complete problem, he or she would have
proved that P= NP. One of the greatest unresolved problems in mathematics and computer
science today is whether P = NP or P 6= NP [62].
In this chapter we present algorithms for five well-known graph theory problems. First,
we discuss Dijkstras algorithm for computing shortest paths in graphs. This algorithm
is polynomial-time and hence the problem is in P. Dijkstras algorithm is widely used
in practice, for example, as part of the TCP/IP protocol suite for routing internet traffic
over shortest paths. The second is Prims algorithm for computing minimal spanning trees
in graphs. This algorithm is obtained by a trivial modification of Dijkstras algorithm,
is also polynomial-time and hence the problem again is in P. Prims algorithm is also
widely used in practice, for example, in distribution problems and broadcasting data over
computer networks. Third, we present Fleurys algorithm for finding Eulerian circuits
(cycles) in graphs. The critical point here is to decide whether a certain pair of subgraphs
partition the graph into two nontrivial components and the trick is to compute a tree using
Prims algorithm and check whether the tree spans the graph. Fleurys algorithm is also
polynomial-time and hence the problem of finding Eulerian circuits is also in P. Fourth,
we show how to construct the De Bruijn graphs and sequences using Eulerian circuits in
polynomial-time. These graphs have recently found important applications in multihop and
fault tolerant computer networks. The fifth and final algorithm is for finding Hamiltonian

Graph Theory

367

circuits (cycles) in graphs. The algorithm is an example of a nondeterministic polynomialtime algorithm. The problem of finding a Hamiltonian circuit in a graph is an example of
a NP-complete problem.
All algorithms are implemented in C++ and tested using Microsoft Visual C++ [158].

13.1 Dijkstras Algorithm for Shortest Paths

In 1956, Edsger W. Dijkstra [65, 153], was the main programmer at the Burroughs Corporation in Amsterdam, where the construction of one of the earliest computers (the ARMAC) was on the verge of completion. In order to celebrate its inauguration they needed
a nice demonstration. It should solve a problem that could be easily stated to a predominantly lay audience. For the purpose of the demonstration, Dijkstra drew a slightly simplified map of the Dutch railroad system; someone in the audience could ask for the shortest
connection between, say, Harlingen and Maastricht, and the ARMAC would print out the
shortest route town by town. The demonstration was a great success; Dijkstra reminisces
that he could show that the inversion of source and destination could influence the computation time required. The speed of the ARMAC and the size of the map were such that
one-minute computations always sufficed.
The general problem is to find shortest paths from one specified vertex to all other vertices in a weighted graph, where the edge weights correspond to distances between towns.
Together, these paths will form a (not necessarily minimal) spanning tree. The main idea
of Dijkstras algorithm is the following; if P is a shortest path from u to z and P contains v,
then the portion of the path P from u to v must be a shortest path from u to v. This suggests
that we should determine optimal routes from u to every other vertex z in increasing order
of the distance d(u, z) from u to z. We maintain a current tentative distance and use this to
update the remaining tentative distances in tabular form. The algorithm works equally well
for directed graphs. A formal description of Dijkstras algorithm is the following.
13.1.1

Dijkstras algorithm to compute distances from u

Input A weighted graph (or digraph) and a starting vertex u. The weight of edge xy is
w(xy); let w(xy) = if xy is not an edge.
Idea Maintain the set S of vertices to which the shortest route from u is known, enlarging
S to include all the vertices. To do this maintain also a tentative distance t(z) from u to each
z not in S; this is the length of the shortest path found yet from u to z.
Initialisation

Let S = {u}; d(u, u) = 0; t(z) = w(uz) for all z 6= u.

Iteration Select a vertex v outside S such that t(v) is minimum in the set {t(z)|z
/ S}. Add
v to S. For each edge vz with z
/ S, update t(z) to min{t(z), d(u, v) + w(vz)}.

368

Graph Algorithms

Continue the iteration until S = V (G) or until t(z) = for every z / S. In

the first case, all shortest paths from u have been found. In the latter case, the remaining
vertices are unreachable from u.
We give an implementation of Dijkstras algorithm in C++ below.

Termination

dijkstra.cpp

#include
#include
#include
#include
#include

<iostream>
<fstream>
<string>
<vector>
<set>

using namespace std;

ifstream infile(graph.txt);
ofstream outfile(shortest_paths.txt);
int main()
cout<<Dijkstras Algorithm.<<endl;
int m,n,i,j,k;
//Read adjacency matrix of weights from graph.txt
infile>>n;
vector< vector< float> > weight;
float val;
for(i=0; i<n; i++)
{
vector< float > row;
for(j=0; j<n; j++)
{
infile>>val;
row.push_back(val);
}
weight.push_back(row);
}
//Initialize Table
const float infinity=1000000;

Graph Theory
vector<bool> known;
for(i=0; i<n; i++) known.push_back(false);
vector<float> d;
d.push_back(0);
for(i=1; i<n; i++) d.push_back(infinity);
vector<int> p;
for(i=0; i<n; i++) p.push_back(-1);
//Print Initial Table
outfile<<endl<<INITIAL TABLE:<<endl;
outfile<<endl<<Vertex : t;
for(i=0; i<n; i++) outfile<<i<<\t;
outfile<<endl<<Known :\t;
for(i=0; i<n; i++) outfile<<known[i]<<\t;
outfile<<endl<<Distance:\t;
for(i=0; i<n; i++) outfile<<d[i]<<\t;
outfile<<endl<<Path
:\t;
for(i=0; i<n; i++) outfile<<p[i]<<\t;
outfile<<endl;
//Iteration
for(k=0; k<n; k++)
{
//Find min of d for unknown vertices
int min=0;
while(known[min]==true)min++;
for(i=0; i<n; i++)
if(known[i]==false && d[i]<d[min])min=i;
//Update Table
known[min]=true;
for (int j=0; j<n; j++)
{
if(weight[min][j]!=0 &&
d[j]>d[min]+weight[min][j] &&
known[j]==false)
{
d[j]=d[min]+weight[min][j];
p[j]=min;
}

369

370

Graph Algorithms
}
//Print Table
outfile<<endl<<TABLE No.<<k<<":<<endl;
outfile<<endl<<Vertex :\t;
for(i=0; i<n; i++) outfile<<i<<\t;
outfile<<endl<<Known :\t;
for(i=0; i<n; i++) outfile<<known[i]<<\t;
outfile<<endl<<Distance:\t;
for(i=0; i<n; i++) outfile<<d[i]<<\t;
outfile<<endl<<Path
:\t;
for(i=0; i<n; i++) outfile<<p[i]<<\t;
outfile<<endl;

}
//Find shortest paths and spanning tree
outfile<<endl<<endl<<SHORTEST PATHS:<<endl;
set< vector<int> > span;
for(i=0; i<n; i++)
{
outfile<<Path=;
vector<int> temp;
m=i;
while(m!=-1)
{
temp.push_back(m);
m=p[m];
}
outfile<<temp[temp.size()-1]<<" ";
for(j=temp.size()-2; j>=0; j--)
{
outfile<<temp[j]<<" ";
vector<int> edge;
edge.push_back(temp[j+1]);
edge.push_back(temp[j]);
span.insert(edge);
}
outfile<<Distance=<<d[i]<<endl;
}

Graph Theory

371

//Print spanning tree

outfile<<endl<<SPANNING TREE:<<endl;
set< vector<int>>::iterator it;
for(it=span.begin(); it!=span.end(); it++)
outfile<<(<<(*it)[0]<<,<<(*it)[1]<<")"<<" ";
cout<<See shortest_paths.txt for results. <<endl;system(PAUSE);
return 0;}

Example 13.1.2
13.1.

Consider the Petersen graph with vertices labelled as shown in Figure

Fig. 13.1 The Petersen graph with labelled vertices

Make a text file graph.txt as shown below and save it in the same directory as the program
dijkstra.cpp. The first line of the file graph.txt is the number of vertices and the rest is
the adjacency matrix of the labelled Petersen graph shown in Figure 13.1.

372

Graph Algorithms
graph.txt

10
0
1
0
0
1
0
1
0
0
0

1
0
1
0
0
0
0
1
0
0

0
1
0
1
0
0
0
0
1
0

0
0
1
0
1
0
0
0
0
1

1
0
0
1
0
1
0
0
0
0

0
0
0
0
1
0
0
1
1
0

1
0
0
0
0
0
0
0
1
1

0
1
0
0
0
1
0
0
0
1

0
0
1
0
0
1
1
0
0
0

0
0
0
1
0
0
1
1
0
0

Now compile and run the program dijkstra.cpp. The following output file shortest paths.txt
is produced in the programs directory. The output shows shortest paths from the vertex
labelled 0 to all other vertices:
shortest paths.txt

INITIAL TABLE
Vertex
known
Distance
Path

:
:
:
:

0
0
0
1

1
0

2
0

3
0

4
0

5
0

6
0

7
0

8
0

9
0

1
0
1
0

2
0

3
0

7
0

8
0

9
0

6
0
1
0

1
1
1
0

2
0
2
1

3
0

TABLE No.0:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

4
0
1
0

5
0

TABLE No.1:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

4
0
1
0

5
0

6
0
1
0

7
0
2
1

8
0

9
0

Graph Theory

373

TABLE No.2:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
1
1
0

2
0
2
1

3
0
2
4

4
1
1
0

5
0
2
4

6
0
1
0

7
0
2
1

8
0

9
0

1
1
1
0

2
0
2
1

3
0
2
4

4
1
1
0

5
0
2
4

6
1
1
0

7
0
2
1

8
0
2
6

9
0
2
6

1
1
1
0

2
1
2
1

3
0
2
4

4
1
1
0

5
0
2
4

6
1
1
0

7
0
2
1

8
0
2
6

9
0
2
6

1
1
1
0

2
1
2
1

3
1
2
4

4
1
1
0

5
0
2
4

6
1
1
0

7
0
2
1

8
0
2
6

9
0
2
6

1
1
1
0

2
1
2
1

3
1
2
4

4
1
1
0

5
1
2
4

6
1
1
0

7
0
2
1

8
0
2
6

9
0
2
6

TABLE No.3:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

TABLE No.4:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

TABLE No.5:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

TABLE No.6:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

374

Graph Algorithms

TABLE No.7:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
1
1
0

2
1
2
1

3
1
2
4

4
1
1
0

5
1
2
4

6
1
1
0

7
1
2
1

8
0
2
6

9
0
2
6

1
1
1
0

2
1
2
1

3
1
2
4

4
1
1
0

5
1
2
4

6
1
1
0

7
1
2
1

8
1
2
6

9
0
2
6

1
1
1
0

2
1
2
1

3
1
2
4

4
1
1
0

5
1
2
4

6
1
1
0

7
1
2
1

8
1
2
6

9
1
2
6

TABLE No.8:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

TABLE No.9:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

Shortest paths

Path=0 Distance=0
Path=0 1 Distance=1
Path=0 1 2 Distance=2
Path=0 4 3 Distance=2
Path=0 4 Distance=1
Path=0 4 5 Distance=2
Path=0 6 Distance=1
Path=0 1 7 Distance=2
Path=0 6 8 Distance=2
Path=0 6 9 Distance=2
Spanning tree

(0, 1) (0, 4) (0, 6) (1, 2) (1, 7) (4, 3) (4, 5) (6, 8) (6, 9)

Graph Theory

375

Fig 13.2 Shortest paths from vertex 0 and a spanning tree

The output of Dijkstras algorithm for the Petersen graph with starting vertex 0 is shown in
Figure 13.2. In this case, the spanning tree is minimal as the reader may verify. This is not
always the case as the next example shows.
Example 13.1.3

Consider the weighted and directed graph shown in Figure 13.3.

Fig. 13.3 A weighted and directed graph

Make a text file graph.txt as shown below and save it in the same directory as the program
dijkstra.cpp. The first line of the file graph.txt is the number of vertices and the rest is
the matrix of weights of the labelled directed graph shown in Figure 13.3 above.

376

Graph Algorithms
graph.txt

7
0
0
0
0
0
1
7

7
0
0
0
0
0
0

0
2
0
0
0
0
0

0
0
2
0
0
0
0

0
0
0
5
0
0
1

0
0
0
0
6
0
0

0
3
4
1
0
4
0

Now run the program dijkstra.cpp. The following output file shortest paths.txt is produced in the programs directory. The output shows shortest paths from the vertex labelled
0 to all other vertices:
shortest paths.txt

INITIAL TABLE:
Vertex
Known
Distance
Path

0
0
0

:
:
:
:

1
0

2
0

3
0

4
0

5
0

6
0

TABLE No.0:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
0
7
0

2
0

3
0

4
0

5
0

6
0

1
1
7
0

2
0
9
1

3
0

4
0

5
0

TABLE No.1:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

6
0
10
1

Graph Theory

377

TABLE No.2:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
1
7
0

2
1
9
1

3
0
11
2

4
0

5
0

6
0
10
1

1
1
7
0

2
1
9
1

3
0
11
2

4
0
11
6

5
0

1
1
7
0

2
1
9
1

3
1
11
2

4
0
11
6

6
1
10
1

1
1
7
0

2
1
9
1

3
1
11
2

4
1
11
6

5
0
17
4

6
1
10
1

1
1
7
0

2
1
9
1

3
1
11
2

4
1
11
6

5
1
17
4

6
1
10
1

TABLE No.3:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

6
1
10
1

TABLE No.4:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

5
0

TABLE No.5:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

TABLE No.6:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

Shortest paths

Path=0 Distance=0
Path=0 1 Distance=7
Path=0 1 2 Distance=9

378

Graph Algorithms

Path=0 1 2 3 Distance=11
Path=0 1 6 4 Distance=11
Path=0 1 6 4 5 Distance=17
Path=0 1 6 Distance=10
Spanning tree

(0, 1) (1, 2) (1, 6) (2, 3) (4, 5) (6, 4)

Fig. 13.4 Shortest paths from vertex 0 and a spanning tree

The output of Dijkstras algorithm for the graph of Figure 13.3 with starting vertex 0 is
shown in Figure 13.4. In this case, the spanning tree has total weight 21 and is not minimal.
Indeed, it is possible to find a spanning tree of total weight 19 in this graph, as the next
section will show.

13.2 Prims Algorithm for Minimal Spanning Tree

Almost simultaneously in 1957, while Dijkstras algorithm was just becoming widely
known, R.C. Prim [211, 249] of Bell Labs discovered an algorithm for finding a minimal
spanning tree in a weighted graph. Surprisingly, a minor modification of Dijkstras algorithm allows us to find a minimal spanning tree. A formal description of Prims algorithm
is the following.
13.2.1

Prims algorithm to compute a minimal spanning tree from u

Input A weighted graph and a starting vertex u. The weight of edge xy is w(xy); let
w(xy) = if xy is not an edge.

Graph Theory

379

Idea Maintain the set S of vertices to which the shortest path from u is known, enlarging
S to include all the vertices. To do this maintain also a tentative distance t(z) from u to each
z not in S; this is the length of the shortest path found yet from u to z.
Initialisation

Let S = {u}; d(u, u) = 0; t(z) = w(uz) for all z 6= u.

Iteration Select a vertex v outside S such that t(v) is minimum in the set {t(z)|z
/ S}.
Add v to S. For each edge vz with z
/ S, update t(z) to min{t(z), w(vz)}. ( Note this step is

different from the corresponding step in Dijkstras algorithm).

Continue the iteration until S = V (G) or until t(z) = for every z / S. In
the first case, all shortest paths from u have been found. Together, they yield a minimal
spanning tree. In the latter case, the remaining vertices are unreachable from u and the
shortest paths together will not span the graph.
We give an implementation of Prims algorithm in C++ below.
Termination

prim.cpp

#include
#include
#include
#include
#include

<iostream>
<fstream>
<string>
<vector>
<set>

using namespace std;

ifstream infile(graph.txt);
ofstream outfile(minimal_spanning_tree.txt);
int main()
{
cout<<Prims Algorithm.<<endl;
int m,n,i,j;
//Read adjacency matrix of weights from graph.txt
infile>>n;
vector< vector<float> > weight; {float }val;
for(i=0; i<n; i++)
{
vector<{float}> row;

380

Graph Algorithms
for(j=0; j<n; j++)
{
infile>>val;
row.push_back(val);
}
weight.push_back(row);

}
//Initialize Table
const float infinity=1000000;
vector<bool> known;
for(i=0; i<n; i++) known.push_back(false);
vector<float> d;
d.push_back(0);
for(i=1; i<n; i++) d.push_back(infinity);
vector<int> p;
for(i=0; i<n; i++) p.push_back(-1);
//Print Table
outfile<<endl<<INITIAL TABLE:<<endl;
outfile<<endl<<Vertex :\t;
for(i=0; i<n; i++) outfile<<i<<\t;
outfile<<endl<<Known :\t;
for(i=0; i<n; i++) outfile<<known[i]<<\t;
outfile<<endl<<Distance:\t;
for(i=0; i<n; i++) outfile<<d[i]<<\t;
outfile<<endl<<Path
:\t;
for(i=0; i<n; i++) outfile<<p[i]<<\t;
outfile<<endl;
//Iteration
for(m=0; m<n; m++)
{
//Find min of d for unknown vertices
int min=0;
while(known[min]==true)min++;
for(i=0; i<n; i++)
if(known[i]==false && d[i]<d[min])min=i;
//Update Table
known[min]=true;

Graph Theory
for(j=0; j<n; j++)
{
if(weight[min][j]!=0 &&
d[j]>weight[min][j] &&
known[j]==false)
{
d[j]=weight[min][j];
p[j]=min;
}
}
//Print Table
outfile<<endl<<endl<<TABLE No.<<m<<:<<endl;
outfile<<endl<<Vertex :\t;
for(i=0; i<n; i++) outfile<<i<<\t;
outfile<<endl<<Known :\t;
for(i=0; i<n; i++) outfile<<known[i]<<\t;
outfile<<endl<<Distance:\t;
for(i=0; i<n; i++) outfile<<d[i]<<\t;
outfile<<endl<<Path
:\t;
for(i=0; i<n; i++) outfile<<p[i]<<\t;
outfile<<endl;
}
//Print minimal spanning tree
outfile<<endl<<MINIMAL SPANNING TREE:<<endl;
for(i=1; i<n; i++)
outfile<<(<<i<<,<<p[i]<<) ;
cout<<See minimal spanning tree.txt.<<endl;
system(PAUSE);
return 0;
}

381

382

Graph Algorithms

Example 13.2.1

Consider the weighted graph shown in Figure 13.5.

Fig. 13.5 A weighted graph

Make a text file graph.txt as shown below and save it in the same directory as the program
prim.cpp. The first line of the file graph.txt is the number of vertices and the rest is the
matrix of weights of the labelled graph shown in Figure 13.5.
graph.txt

8
0
7
0
5
0
0
0
3

7
0
4
0
0
0
5
0

0
4
0
6
0
5
0
0

5
0
6
0
3
0
0
0

0
0
0
3
0
3
0
2

0
0
5
0
3
0
2
0

0
5
0
0
0
2
0
1

3
0
0
0
2
0
1
0

Now compile and run the program prim.cpp. The following output file minimal spanning tree.txt
is produced in the programs directory. The output shows a minimal spanning tree for the
graph.

Graph Theory

383
minimal spanning tree.txt

INITIAL TABLE:
Vertex
Known
Distance
Path

:
:
:
:

0
0

1
0

2
0

3
0

4
0

5
0

6
0

7
0

1
0
7
0

2
0

3
0
5
0

4
0

5
0

6
0

3
0
5
0

4
0
2
7

5
0

3
0
5
0

4
0
2
7

5
0
2
6

6
1
1
7

7
1
3
0

3
0
3
4

4
1
2
7

5
0
2
6

6
1
1
7

7
1
3
0

TABLE No.0:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

TABLE No.1:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
0
7
0

2
0

6
0
1
7

7
1
3
0

TABLE No.2:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
0
5
6

2
0

TABLE No.3:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
0
5
6

2
0

1
0
5
6

2
0
5
5

TABLE No.4:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

3
0
3
4

4
1
2
7

5
1
2
6

6
1
1
7

7
1
3
0

7
0
3
0

384

Graph Algorithms

TABLE No.5:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
0
5
6

2
0
5
5

3
1
3
4

4
1
2
7

5
1
2
6

6
1
1
7

7
1
3
0

:
:
:
:

0
1
0
1

1
1
5
6

2
0
4
1

3
1
3
4

4
1
2
7

5
1
2
6

6
1
1
7

7
1
3
0

:
:
:
:

0
1
0
1

1
1
5
6

2
1
4
1

3
1
3
4

4
1
2
7

5
1
2
6

6
1
1
7

7
1
3
0

TABLE No.6:
Vertex
Known
Distance
Path
TABLE No.7:
Vertex
Known
Distance
Path

Minimal spanning tree

[(1, 6) (2, 1) (3, 4) (4, 7) (5, 6) (6, 7) (7, 0)

Fig. 13.6 A minimal spanning tree for the weighted graph

The output of Prims algorithm for the graph of Figure 13.5 with starting vertex 0 is
shown in Figure 13.6. The spanning tree has total weight 20 and is of minimal total weight
compared to all other spanning trees.

Graph Theory

385

Example 13.2.1 It makes sense to try to compute a minimal spanning tree for a directed
graph. Consider the weighted directed graph of Figure 13.3. We claimed that the spanning
tree found by Dijkstras algorithm of total weight 21 is not minimal. Let us run the program
for Prims algorithm on the file graph.txt of Example 13.1.3. We obtain the following
output.
minimal spanning tree.txt

INITIAL TABLE:
Vertex
Known
Distance
Path

:
:
:
:

0
0
0
1

1
0

2
0

3
0

4
0

5
0

6
0

:
:
:
:

0
1
0
1

1
0
7
0

2
0

3
0

4
0

5
0

6
0

:
:
:
:

0
1
0
1

1
1
7
0

2
0
2
1

3
0

4
0

5
0

:
:
:
:

0
1
0
1

1
1
7
0

2
1
2
1

3
0
2
2

4
0

5
0

:
:
:
:

0
1
0
1

1
1
7
0

2
1
2
1

3
1
2
2

4
0
5
3

5
0

TABLE No.0:
Vertex
Known
Distance
Path
TABLE No.1:
Vertex
Known
Distance
Path

6
0
3
1

TABLE No.2:
Vertex
Known
Distance
Path

6
0
3
1

TABLE No.3:
Vertex
Known
Distance
Path

6
0
1
3

386

Graph Algorithms

TABLE No.4:
Vertex
Known
Distance
Path

:
:
:
:

0
1
0
1

1
1
7
0

2
1
2
1

3
1
2
2

4
0
1
6

5
0

6
1
1
3

:
:
:
:

0
1
0
1

1
1
7
0

2
1
2
1

3
1
2
2

4
1
1
6

5
0
6
4

6
1
1
3

:
:
:
:

0
1
0
1

1
1
7
0

2
1
2
1

3
1
2
2

4
1
1
6

5
1
6
4

6
1
1
3

TABLE No.5:
Vertex
Known
Distance
Path
TABLE No.6:
Vertex
Known
Distance
Path

Minimal spanning tree

(1, 0) (2, 1) (3, 2) (4, 6) (5, 4) (6, 3)

Fig. 13.7 A minimal spanning tree for the weighted directed graph

The output of Prims algorithm for the directed weighted graph of Figure 13.3 with
starting vertex 0 is shown in Figure 13.7. The spanning tree has total weight 19 and is of
minimal total weight compared to all other spanning trees, in particular, the spanning tree
of total weight 21 found by Dijkstras algorithm.

Graph Theory

387

13.3 Fleurys Algorithm for Eulerian Circuit

The problem of finding an Eulerian circuit in a graph (possibly with multiple edges) has
been studied since Leonhard Eulers solution [74] to the problem of the seven bridges of
Knigsberg in 1736 (see Chapter 3). Lewis Carroll [263], we are told in a biography by
his nephew, was fond of asking little children to draw, in one stroke, without lifting the pen
off the paper, Figure 13.8.

Fig. 13.8 Lewis Carrolls three-square graph

If we define all intersections of line segments as vertices then we obtain a planar graph
with 18 vertices, known as Lewis Carrolls three-square graph. Drawing the figure in one
stroke is equivalent to finding an Eulerian circuit in the graph. Since all vertices have even
degree 2 or 4, we know that there must be such an Eulerian circuit. In Example 13.3.1,
we show how to find this Eulerian circuit in a systematic way. Lucas [153] describes an
algorithm for finding an Eulerian trail (or circuit, if one exists) due to Fleury.
13.3.1

Fleurys algorithm

Input

A graph G with one nontrivial component and at most two odd vertices.

Initialisation

If G has an odd vertex, start at an odd vertex. Else start at any vertex.

Iteration From the current vertex, traverse any remaining edge whose deletion from the
remaining graph does not leave a graph with two nontrivial components. To check this, run
Prims algorithm and make sure that the resulting tree spans the graph.

Stop when there are no more edges left to traverse.

We give an implementation of Fleurys algorithm in C++ below. Note that this implementation is for simple graphs, but we can easily modify it to handle graphs with multiple
edges.
Termination

388

Graph Algorithms
fleury.cpp

#include
#include
#include
#include

<iostream>
<fstream>
<string>
<vector>

using namespace std;

ifstream infile(graph.txt);
ofstream outfile(eulerian_circuit.txt);
bool euler(vector<vector<int> > edge);
bool fleury(vector<vector<int> > edge, vector<int> del);
vector<vector<int> >erase(vector<vector<int> > edge, vector<int> del );
bool empty(vector<vector<int> > edge);
int main()
{
cout<<Fleurys Algorithm.<<endl;
int n, i, j;
//Read adjacency matrix of edges from graph.txt
infile>>n;
vector<vector<int> > edge; int val;
for(i=0; i<n; i++)
{
vector<int> row;
for(j=0; j<n; j++)
{
infile>>val; row.push_back(val);
}
edge.push_back(row);
}
cout<<Read graph from file graph.txt...<<endl;
if(euler(edge))
{
cout<<Finding Eulerian circuit...<<endl;
vector<int> circuit; int current=0;

Graph Theory
circuit.push_back(current); cout<<current<< ";
while(!empty(edge))
{
for(i=0; i<n; i++)
{
int previous=current;
if(edge[current][i]==1)
{
vector<int> del;
del.push_back(current);
del.push_back(i);
if(fleury(edge, del))
{
edge=erase(edge,del); current=i;
circuit.push_back(current);
cout<<current<<" ";
break;
}
}
}
}
for(i=0; i<circuit.size(); i++)
outfile<<circuit[i]<<" ";
cout<<endl<<See circuit.txt for results.<<endl;
}
else
cout<<No Eulerian circuit.<<endl;
system(PAUSE); return 0;
}
bool euler(vector<vector<int> > edge)
{
for(int i=0; i<edge.size(); i++)
{
int deg=0;
for(int j=0; j<edge[0].size(); j++)
deg+=edge[i][j];

389

390

Graph Algorithms
if(deg%2!=0) return false;

}
return true;
}
bool fleury(vector<vector<int> > edge, vector<int> del )
{
int n, i, j, k;
if(del[0]==del[1]) return false;
vector<vector<int> > edged=edge;
edged[del[0]][del[1]]=0; edged[del[1]][del[0]]=0;
n= edged[0].size();
//Initialize Table
const int infinity=1000000;
vector<bool> known;
for(i=0; i<n; i++) known.push_back(false);
vector<int> d; d.push_back(0);
for(i=1; i<n; i++) d.push_back(infinity);
vector<int> p;
for(i=0; i<n; i++) p.push_back(-1);
//Iteration
for(k=0; k<n; k++)
{
//Find min of d for unknown vertices
int min=0;
while(known[min]==true)min++;
for(i=0; i<n; i++)
if(known[i]==false && d[i]<d[min])min=i;
//Update Table
known[min]=true;
for(j=0; j<n; j++)
{
if(edged[min][j]!=0 && d[j]>edged[min][j] &&
known[j]==false)
{
d[j]=edged[min][j]; p[j]=min;
}

Graph Theory

391

}
}
bool ok=true;
//Find if resulting graph has two nontrivial //components
it for(i=1; i<n; i++)
{
if(p[i]==-1)
for (int j=0; j<n; j++)
if(edged[i][j]!=0)
{ok=false; break;}
}
return ok;
}
vector<vector<int> > erase(vector< vector<int> > edge, vector<int> del )
{
vector< vector<int> > edged=edge;
edged[del[0]][del[1]]=0; edged[del[1]][del[0]]=0;
return edged;
}
bool empty(vector<vector<int> > edge)
{
for(int i=0; i<edge.size(); i++)
for(int j=0; j<edge[0].size(); j++)
if(edge[i][j]==1)
return false; return true;
}

Example 13.3.1 Consider the following labelled graph given in Figure 13.9 corresponding to Lewis Carrolls three-square puzzle.

392

Graph Algorithms

Fig. 13.9 Labelled graph corresponding to Lewis Carrolls puzzle

Make a text file graph.txt corresponding to Figure 13.9.

graph.txt

18
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0

1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0

0
1
1
0
1
0
1
0
0
0
0
0
0
0
0
0
0
0

0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0

0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
0
0

0
0
0
1
0
1
0
1
0
0
0
0
1
0
0
0
0
0

0
0
0
0
1
0
1
0
1
0
0
0
0
0
0
1
0
0

0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1

1
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0

0
0
1
0
0
0
0
0
0
1
0
1
0
1
0
0
0
0

0
0
0
0
0
1
0
0
0
0
1
0
1
0
1
0
0
0

0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0

0
0
0
0
0
0
0
0
0
0
0
1
0
1
0
1
1
0

0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0

0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1

0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0

Graph Theory

393

Now compile and run the program fleury.cpp. The following output file eulerian circuit.txt
is produced in the programs directory.
0 1 3 2 10 11 5 6 3 4 7 6 12 11 14 15 7 8 17 16 14 13 10 9 0 The output shows the final
Eulerian circuit in the given graph. If we trace the actual computation of this Eulerian
circuit step by step, we obtain a solution to Carrolls puzzle, as shown in Figure 13.10.

Fig. 13.10 Solution to Lewis Carrolls three-square puzzle

13.4 De Bruijn Graphs

To see a good application of Eulerian circuits, we briefly consider the rotating drum problem, described as follows. Suppose the head of a rotating drum is divided into 24 = 16
sectors, with each sector labelled with either a zero or a one, as shown in Figure 13.11.

Fig. 13.11 Labelled rotating drum

394

Graph Algorithms

Can the sectors be labelled in such a way that the labels of any four consecutive sectors
uniquely determine the position of the drum? This means that the 16 possible quadruples of
consecutive binary labels on the drum should be the binary representations of the integers
0 to 15. This question was studied by N.G. de Bruijn [41] in 1946 and thus the resulting
binary circular sequences and their corresponding graphs given below are called De Bruijn
sequences and De Bruijn graphs, respectively.
We define a directed graph called G4 (the De Bruijn graph of order 4) as follows. The
vertices are all 3-bit binary strings x1 x2 x3 (each xi is either zero or one). Thus there are 8
vertices. Each vertex x1 x2 x3 has directed edges to the vertices x2 x3 0 and x2 x3 1. The directed
edge (x1 x2 x3 , x2 x3 x4 ) is labelled e j , where j = x1 23 + x2 22 + x3 21 + x4 20 is the unique binary
representation of the integer j . Thus, there are 16 directed edges. The graph is shown in
the following Figure 13.12.

Fig. 13.12 G4 , the De Bruijn graph of order 4

Now, G4 is Eulerian since the in-degree and out-degree of every vertex is 2. Using
Fleurys algorithm we find the following Eulerian circuit in G4 : 000, 000, 001, 011, 111,
111, 110, 100, 001, 010, 101, 011, 110, 101, 010, 100, 000. This Eulerian circuit corresponds to the De Bruijn sequence 0000111100101101, to be read circularly, with the
required property. This is exactly the labeling on the rotating drum shown in Figure 13.11.

Graph Theory

395

In general, we may build the De Bruijn graph of order n, denoted by Gn , and find circular De Bruijn sequences of length n corresponding to Eulerian circuits in Gn . Among
other things, this technique has recently been used to design large scale multihop and fault
tolerant computer networks [239].

13.5 Hamiltonian Circuits

A concept that is similar to Eulerian circuits but in reality quite different is that of Hamiltonian circuit. A hamiltonian circuit in a graph is a simple closed path that passes through
each vertex exactly once. In the mid 19th century, Sir William Rowan Hamilton [102]
tried to popularise the exercise of finding such a circuit in the graph of a dodecahedron.
While we have seen a polynomial-time algorithm for Eulerian circuits, no such algorithm
is known for Hamiltonian circuits. We give an implementation of a nondeterministic algorithm in C++ below.
hamilton.cpp

#include
#include
#include
#include
#include

<iostream>
<fstream>
<string>
<vector>
<algorithm>

using namespace std;

ifstream infile(graph.txt);
ofstream outfile(hamiltonian_circuits.txt);
int main()
{
cout<<Algorithm for Hamiltonian circuits.<<endl;
//Read adjacency matrix of from graph.txt
int i, j, k, l, m, n;
infile>>n;
vector< vector<bool> > graph;
bool edge;
for(i=0; i<n; i++)
{
vector<bool> row;

396

Graph Algorithms
for(j=0; j<n; j++)
{
infile>>edge;
row.push_back(edge);
}
graph.push_back(row);

}
int count=0;
vector<int> vertex;
for(k=0; k<n; k++)
vertex.push_back(k);
cout<<\nStarting search...\n;
bool found=false, circuit=true;
while(next_permutation(vertex.begin(),vertex.end())
)
{
if(vertex[0]!=0) break;
for(l=0; l<n; l++)
circuit=
circuit*graph[vertex[l%n]][vertex[(l+1)%n]];
switch(circuit)
{
case true:
found=true;
count++;
cout<<endl
<<count
<< Hamiltonian Circuits found:<<endl;
outfile<<endl
<<count
<< Hamiltonian Circuits found:
<<endl;
for(m=0; m<n; m++)
{
cout<<vertex[m]<< ;
outfile<<vertex[m]<< ;
}

Graph Theory

397

cout<<endl; outfile<<endl;
break;
default:
break;
}
circuit=true;
}
if(!found)
{
cout<<\nNo Hamiltonian Circuits found.\n;
outfile<<\nNo Hamiltonian Circuits found.\n;
}
cout<<\nSee hamilton_circuits.txt.
<<endl;
system(PAUSE);
return 0;
}

Example 13.5.1 We consider the graphs of the five platonic solids, all of which are
known to be Hamiltonian. First, consider the simplest of the platonic graphs, the tetrahedron:

Fig 13.13 Labelled graph of the tetrahedron

Make a text file graph.txt as shown below and save it in the same directory as the program
hamilton.cpp. The first line of the file graph.txt is the number of vertices and the rest
is the adjacency matrix of the labelled tetrahedron graph shown in Figure 13.13.

398

Graph Algorithms
graph.txt

4
0
1
1
1

1
0
1
1

1
1
0
1

1
1
1
0

Now compile and run the program. We find five Hamiltonian circuits: 0 1 3 2, 0 2 1 3, 0 2
3 1, 0 3 1 2, 0 3 2 1. Next, consider the graph of the octahedron:

Fig. 13.14 Labelled graph of the octahedron

The file graph.txt for the octahedron graph is shown below:

graph.txt

6
0
1
1
0
1
1

1
0
1
1
0
1

1
1
0
1
1
0

0
1
1
0
1
1

1
0
1
1
0
1

1
1
0
1
1
0

Now run the program. We find 31 Hamiltonian circuits: 0 1 2 3 5 4, 0 1 2 4 3 5, 0 1 3 2 4

5, 0 1 3 5 4 2, 0 1 5 3 2 4, 0 1 5 3 4 2, 0 1 5 4 3 2, 0 2 1 3 4 5, 0 2 1 3 5 4, 0 2 1 5 3 4, 0 2 3
1 5 4, 0 2 3 4 5 1, 0 2 4 3 1 5, 0 2 4 3 5 1, 0 2 4 5 3 1, 0 4 2 1 3 5, 0 4 2 3 1 5, 0 4 2 3 5 1, 0
4 3 2 1 5, 0 4 3 5 1 2, 0 4 5 1 3 2, 0 4 5 3 1 2, 0 4 5 3 2 1, 0 5 1 2 3 4, 0 5 1 3 2 4, 0 5 1 3 4
2, 0 5 3 1 2 4, 0 5 3 4 2 1, 0 5 4 2 3 1, 0 5 4 3 1 2, 0 5 4 3 2 1.
Next, consider the graph of the cube.

Graph Theory

399

Fig. 13.15 Labelled graph of the cube

graph.txt

8
0
1
0
1
0
1
0
0

1
0
1
0
0
0
1
0

0
1
0
1
0
0
0
1

1
0
1
0
1
0
0
0

0
0
0
1
0
1
0
1

1
0
0
0
1
0
1
0

0
1
0
0
0
1
0
1

0
0
1
0
1
0
1
0

Now run the program. We find 12 Hamiltonian circuits: 0 1 2 3 4 7 6 5, 0 1 2 7 6 5 4 3, 0

1 6 5 4 7 2 3, 0 1 6 7 2 3 4 5, 0 3 2 1 6 7 4 5, 0 3 2 7 4 5 6 1, 0 3 4 5 6 7 2 1, 0 3 4 7 2 1 6
5, 0 5 4 3 2 7 6 1, 0 5 4 7 6 1 2 3, 0 5 6 1 2 7 4 3, 0 5 6 7 4 3 2 1. Next, consider the graph
of the icosahedrons, given in Figure 13.16.

400

Graph Algorithms

Fig. 13.16 Labelled graph of the icosahedron

graph.txt

12
0
1
1
0
0
1
1
1
0
0
0
0

1
0
1
1
1
1
0
0
0
0
0
0

1
1
0
1
0
0
0
1
1
0
0
0

0
1
1
0
1
0
0
0
1
1
0
0

0
1
0
1
0
1
0
0
0
1
1
0

1
1
0
0
1
0
1
0
0
0
1
0

1
0
0
0
0
1
0
1
0
0
1
1

1
0
1
0
0
0
1
0
1
0
0
1

0
0
1
1
0
0
0
1
0
1
0
1

0
0
0
1
1
0
0
0
1
0
1
1

0
0
0
0
1
1
1
0
0
1
0
1

0
0
0
0
0
0
1
1
1
1
1
0

Now, it gets interesting. The program runs for quite a few minutes! We find 2560 Hamiltonian circuits: 0 1 2 3 4 5 6 10 9 8 11 7, . . . , 0 7 11 10 9 8 3 2 1 4 5 6. Finally, consider
the graph of the dodecahedron, the original inspiration for Hamilton [212], given in Figure
13.17.

Graph Theory

401

Fig. 13.17 Labelled graph of the dodecahedron

The file graph.txt for the dodecahedron graph is shown below:

graph.txt

20
0
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0

1
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0

0
1
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0

0
0
1
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0

1
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0

1
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0

0
1
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0

0
0
1
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0

0
0
0
1
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0

0
0
0
0
1
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0

0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
1

0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
1
0
0
0
0

0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
1
0
0
0

0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
1
0
0

0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
1
0

0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
1
0
0
1

0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
1
0
0

0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
1
0

0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
1

0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
1
0
0
1
0

402

Graph Algorithms

Depending on the speed of your computer, the program will now run for quite a few hours,
and perhaps even days. One of the Hamiltonian circuits we found was: 0 1 2 3 4 5 6 7 8 9
10 11 12 19 18 17 16 15 14 13.

14. Score Structure in Digraphs

Landau [145] associated with each tournament an ordered sequence of non-negative integers, its score structure, formed by listing the vertex outdegrees in non-decreasing order.
Since then the concept of score structure has been extended to various other classes of digraphs, namely oriented graphs and semicomplete graphs.The score structure property has
been used in the study of some structural properties of digraphs.

14.1 Score Sequences in Tournaments

Definition: In a tournament T , the score s(vi ), or simply si of a vertex vi is the number
of arcs directed away from vi and the score sequence S(T ) is formed by listing the vertex
scores in non-decreasing order. Clearly, 0 si n(n1)
2 . Further, no two scores can be zero
and no two scores can be n 1 . Tournament score sequences have also been called score
structures [145], score vectors [165] and score lists [29].
One interpretation of a tournament is as a competition where n participants play each
other once in a match that cannot end in a tie and score one point for each win. Player v
is represented in the tournament by vertex v and an arc from u to v means that u defeats
v. Then player v obtains a total of dv+ points in the competition and the vertex scores can
be ordered to obtain the score sequence of the tournament. We use u v to denote the
both, an arc from u to v and the fact that u dominates v. A result of Ryser [227] states
that an n-tournament can be obtained from any other having the same score sequence by a
sequence of arc reversals of 3-cycles.
Now, we give the characterisation of score sequences of tournaments which is due to
Landau [145].This result has attracted quite a bit of attention as nearly a dozen different proofs appear in the literature.Early proofs tested the readers patience with special
choices of subscripts,but eventually such gymnastics were replaced by more elegant arguments.Many of the existing proofs are discussed in a survey by Reid [221] and the proof
we give here is due to Thomassen [242]. Further, two new proofs can be found in [89].
Theorem 14.1 (Landau) A sequence of non-negative integers S = [si ]n1 in non-decreasing
order is a score sequence of a tournament if and only if for each subset I [n] = {1, 2, . . ., n},

404

Score Structure in Digraphs

si
iI

|I|
2

(14.1.1)

with equality when |I| = n.

Because of the monotonicity assumption s1 s2 . . . sn , the inequalities (14.1.1),
known as the Landau inequalities, are equivalent to
k

si
i=1

k
2

for 1 k n, with equality for k = n.

Proof
Necessity If a sequence of non-negative integers [si ]n1 in the non-decreasing order is the
score sequence of an n-tournament T , then the sum of the first k scores in the sequence
counts exactly one each arc in the subtournamnent W induced by {v1 , v2 , . . ., vk } plus each
arc from W to T W . Therefore the sum is at least k(k1)
2 , the number of arcs in W. Also,
since the sum of the scores of the vertices counts each arc of the tournament exactly once,
the sum of the scores is the total number of arcs, that is, n(n1)
2 .
Sufficiency (Thomassen) Let n be the smallest integer for which there is a non-decreasing
sequence S of non-negative integers satisfying Landaus conditions (14.1.2), but for which
there is no n-tournament with score sequence S. Among all such S, pick one for which s1
is as small as possible.
First consider the case where for some k < n,
k

si =
i=1

 
k
.
2

By the minimality of n, the sequence S1 = [s1 , s2 , . . ., sk ] is the score sequence of some

tournament T1 . Further,
m

m+k

(sk+i k) = si
i=1

i=1

k
2

mk

m+k
2

k
2

mk =

m
2

for each m, 1 m n k, with the equality when m = n k. Therefore, by the minimality of

n, the sequence S2 = [sk+1 k, sk+2 k, . . ., sn k] is the score sequence of some tournament
T2 . By forming the disjoint union of T1 and T2 , and adding all arcs from T2 to T1 , we obtain
a tournament with score sequence S.
Now, consider the case where each inequality in 14.1.2 is strict when k < n (in particular
s1 > 0). Then the sequence S3 = [s1 1, s2 , . . ., sn1 , sn + 1] satisfies (14.1.2) and by the

Graph Theory

405

minimality of s1 , S3 is the score sequence of some tournament T3 . Let u and v be the

vertices with scores sn + 1 and s1 1 respectively. Since the score of u is larger than that
of v, T3 has a path P from u to v of length 2. By reversing the arcs of P, we obtain a
tournament with score sequence S, a contradiction.
q
Landaus theorem is the tournament analog of the Erdos-Gallai theorem for graphical
sequences. A tournament analog of the Havel-Hakimi theorem for graphical sequences is
the following result, the proof of which can be found in Reid and Beineke [218].
Theorem 14.2 A non-decreasing sequence [si ]n1 of non-negative integers, n 2, is the
score sequence of an n-tournament if and only if the new sequence
[s1 , s2 , . . ., sm , sm+1 1, . . ., sn1 1]

where m = sn , when arranged in non-decreasing order, is the score sequence of some (n 1)tournament.
Definition: A tournament is strongly connected or strong if for every two vertices u and
v there is a path from u to v and a path from v to u. A strong component of a tournament is
a maximal strong subtournament.
The following extension of Theorem 14.1, characterises strong components. The proof
is straightforward and consequently omitted.
Theorem 14.3 A non-decreasing sequence [si ]n1 of non-negative integers is the score
sequence of a strong n-tournament if and only if
k

si >
i=1

k
2

, 1 k n, and

si =
i=1

n
2

We have the following observation from Theorem 14.3. Let S = [si ]n1 be a score sequence


p

of an n-tournament T with vertex set V = {1, 2, . . ., n}. Let si =

i=1

and si >
i=1

k
2

p
2

, si =
i=1

q
2

, for p + 1 k q 1, where 0 p < q n.

Then the subtournament induced by {p +1, . . ., q} is a strong component of T with score

sequence [s p+1 p, s p+2 p, . . ., sq p].
We say that S is strong if T is strong and the strong components of S are the score
sequences of the strong components of T . Theorem 14.3 shows that the strong components
of S are determined by the successive values of k for which
k

si =
i=1

k
2

that is, the successive values of k for which equality holds in condition (14.1.2).

406

Score Structure in Digraphs

For example, consider the score sequence

S = [1, 1, 1, 4, 4, 5, 5, 7, 9, 9, 10, 11, 11],
k

si =
i=1

k
2

for k = 3, 7, 8, and 13.

Therefore the strong components of S are, in ascending order,

[1, 1, 1], [1, 1, 2, 2], [0], and [1, 1, 2, 3, 3].
The next result due to Brualdi and Shen [40] shows that the score sequence of an ntournament satisfies inequalities (14.4.1) below, which are individually stronger than the
inequalities (14.1.1), although collectively the two sets of inequalities are equivalent.
Theorem 14.4 (Brualdi and Shen) A sequence S = [si ]n1 of non-negative integers in
non-decreasing order is a score sequence of a tournament if and only if for every subset
I [n],

si
iI

1
1
(i 1) +
2
2
iI

|I|
2

(14.4.1)

with equality when |I| = n.

Proof

The sufficiency (14.4.1) imply the inequalities (14.1.1).

We prove that the score sequence S of a tournament satisfies (14.4.1). For any subset I [n],
define
1
1
f (I) = si (i 1)
2 iI
2
iI

|I|
2

Choose I firstly to have f (I) minimum and secondly to have |I| minimum.
Claim that I = {i : 1 i |I|}.
Otherwise, there exists i / I and j I such that j = i + 1. Then si s j . Since
1
s j ( j + |I| 2) = f (I) f (I { j}) <0
2
1
2

and si (i + |I| 1) = f (I {i}) f (I) 0,

Graph Theory

407

1
1
1
(i + |I| 1) si s j < ( j + |I| 2) = (i + |I| 1),
2
2
2

which is a contradiction. This proves the claim.

|I|

|I|

Thus, f (I) = si
i=1

1
1
(i 1)

2 i=1
2

|I|
2

|I|

= si
i=1

|I|
2

0,

where the inequality follows from 14.1. By the choice of the subset I, Theorem 14.4
follows.
q
Remark Clearly, the equality can occur often in (14.4.1). For
equality
holds
 example,

n1
for regular tournaments of odd order n (with score sequence n1
,
.
.
.,
)
,
whenever
2
2
|I| = k, and I = {n k + 1, . . ., n}.
Further, Theorem 14.4 is best possible in the sense that, for any real > 0, the inequality

si
iI

1
+
2

(i 1) +
iI



|I|
2

fails for some I and some tournaments (for example, regular tournaments).
The following set of upper bounds for si is equivalent to the set of lower bounds for
iI

si in Theorem 14.4.
iI

Corollary 14.1 A sequence S = [si ]n1 of non-negative integers in non-decreasing order is a

score sequence of a tournament if and only if for any subset I [n],
1

si 2 (i 1) + 4 |I| (2n |I| 1),

iI

iI

with equality when |I| = n.

Proof

Let J = [n] I . Then,

1

si = 2 (i 1) + 4 |I| (2n |I| 1),

iI

iI

with equality when |I| = n, if and only if

si =
iJ

n
2

si
iI

1
1
(i 1) +

2 iJ
2


|J|
,
2

with equality when |J| = n. Therefore Corollary 14.1 follows from Theorem 14.4.

408

Score Structure in Digraphs

Corollary 14.2
n+i2
2 .

If S = [si ]n1 is a score sequence of a tournament, then for each i, i1

2 si

Proof Choose I = {i}. Then the result follows immediately from Theorem 14.4 and
Corollary 14.1.
q
The next result by Brualdi and Shen [40] shows that when equality occurs in the inequalities (14.4.1), there are implications concerning the strong connectedness and regularity of
every tournament with score sequence S. For any integers r and s with r s, [r, s] denotes
the set of all integers between r and s.
If S = [si ]n1 is a tournament score sequence and if

Theorem 14.5

si =
iI

1
1
(i 1) +
2
2
iI

|I|
2

for some I [n], then one of the following holds.

|I|

i. I = [1, |I|] and

si =
i=1


|I|
.
2

ii. I = [t, t + |I| 1] for some t, 2 t n |I| + 1,

t+|I|1

si =

i=1

t + |I| 1
2

and si = (t + |I| 2)/2, for all i t + |I| 1.

iii. I = [1, r] [r + t, t + |I| 1] , for some r and t such that 1 r |I| 1 and 2 t
n |I| + 1,
t+|I|1

i=1

si =

t + |I| 1
2

and si = (r + t + |I| 2)/2 for all i, r + 1 i t + |I| 1.

Remark Conditions (i), (ii) and (iii) of Theorem 14.5 are equivalent to the assertion that
every tournament with the score sequence S has one of the three structures shown in Figure
14.1. The notation Tr is used to denote a subtournament on r vertices and the double arrows
mean that all the arcs between the two parts go in that direction.

Graph Theory

409

Fig. 14.1

The next result due to Bjelica [26] gives a criterion for score segments and subsequences
with arbitrary positions of scores.
Theorem 14.6 If [ti ]m1 is a sequence of non-negative integers in non-decreasing order and
[si ]n1 is a score sequence of an n-tournament T with m < n, then the following properties are
equivalent.
j

ti

i.

i=1
t j = s j,

ii.
iii.
iv.

j
2

1 j m,
1 j m, for some T ,
1 j m, for some T and k,
1 j m, for some T and k1 < k2 < . . . < km .

t j = sk+ j ,
t j = sk j ,

The following result due to Bjelica and Lakic [27] gives the conditions for a set of integers
to be the subset ofscores
 with prescribed positions in some score sequence.
Denote b(x) =

x
2

, X(k) = xi .
i=1

Theorem 14.7 Let 0 t1 t2 . . . tm and 0 < k1 < k2 < . . . < km be two sequences of
integers. Then there exists an n-tournament T with score sequence s1 s2 . . . sn such
that t j = sk j , 1 j m, if and only if
j

(ki ki1)ti
i=1

kj
2

, 1 j m, k0 = 0.

(14.7.1)

The size of the tournament can be km if and only if in (14.7.1) the equality holds for j = m.

Proof
Necessity If, for some tournament, we have t j = sk j , where 1 j m, then monotonicity
of the score sequence and the Landau theorem give

410

Score Structure in Digraphs

j

kj

(ki ki1)ti (ki ki1)Ski Si

i=1

i=1

i=1

kj
2

, 1 j m.

Sufficiency Let some sequences t and k satisfy (14.7.1). Define the sequence u1 u2
. . . ukm which includes the sequence t as subsequence uk = t j , k j1 < k < k j , 1 j m, and
we prove that it satisfies property (i) from Theorem 14.6. In the following minorisation,
we apply piecewise linearity of U , inequalities (14.7.1) and convexity of binomial function
bU(k) = U(k j1 ) + (k k j1 )tk j = U(k j1) + (k k j1 )
=

kj k
k k j1
U(k j1 ) +
U(k j )
k j k j1
k j k j1

kj k
k k j1
b(k j1 ) +
b(k j )
k j k j1
k j k j1

kj k
k k j1
k j1 +
kj
k j k j1
k j k j1

U(k j )U(k j1 )
k j k j1

= b(k).

By property (ii) from Theorem 14.6, there exists an n-tournament T with beginning score
segment u. Hence scores from the sequence t appear on the prescribed positions k.
q

14.2 Frequency Sets in Tournaments

Definition: The number of times that a particular score occurs in a score sequence
of a tournament is called the frequency of that score. A set of distinct positive integers
F = { f 1 , f 2 , . . ., f k } is a frequency set if there exists a tournament T such that the set of
frequencies of the scores in T is exactly F . Note that in such a case T has order at least
f 1 + f 2 + . . . + f k . For example, the reversal of the orientation of three vertex disjoint arcs in
a regular 7-tournament results in a 7-tournament with score sequence [2, 2 , 2, 3, 4, 4, 4]
and frequency set F = {1, 3}.
Define N(F) to be the smallest m such that there exists a tournament on m vertices with
n
frequency set F . Clearly, N( f 1 , f 2 , . . ., f n ) f i . An almost regular tournament is an even
i=1

order tournament in which the scores of the vertices are all as nearly equal as possible.
We have the following observations due to Alspach and Reid [3].
Lemma 14.1

If f 1< f 2 <. . . < f n , n 2, are positive integers, f k is odd, and N( f 1 , f 2 , . . .,

n
f k1, f k+1, . . ., f n) = f i f k , then
i=1

N( f 1 , f 2 , . . ., f n) =

fi
i=1

Graph Theory

Proof

411

Let R be a tournament on


n
f i f k vertices with frequency set { f 1 , f 2 , . . .,

i=1

f k1, f k+1, . . ., f n }. Let Q be a regular tournament on f k vertices. Let T be the tournament

obtained from disjoint copies of 
Q andR and in which every vertex of Q dominates every

vertex of R. It is clear that T has

fi

vertices and frequency set { f 1 , f 2 , . . ., f n}.

i=1

Lemma 14.2 Let R be a regular tournament of order r and let Q1 , Q2 , . . ., Qk be almost

regular tournaments of orders q1 < q2 < . . . < qk . Then there exists a tournament T of
k

order r + qi containing disjoint copies of R, Q1 , Q2 , . . ., Qk as subtournaments such that

i=1

< R Qi > is regular of order r + qi and each vertex of Qi dominates each vertex of Q j when
i > j.
The following constructive criterion due to Alspach and Reid [3] shows that every set F

of positive integers is the frequency set of some tournament and determines the least order
N(F) of such a tournament. The proof is omitted as it is lengthy and can be found in [3].
Theorem 14.8
Then

Let f 1 < f 2 < . . . < f n be positive integers, at least one of which is odd.
n

N( f 1 , f 2, . . ., f n ) = f i

(14.8.1)

i=1

unless either
i. n = 2, f 1 6 f 2 ( mod 2 ), gcd { f 1 , f 2} = 1, and f 2 6= 2, in which case N( f 1 , f 2) = 2 f 1 + f 2
(14.8.2)
or
ii. n = 2, f 1 = 1, f 2 = 2, in which case
N (1, 2) = 5.

(14.8.3)

Definition: Given the set of even integers F = { f 1 , f 2 , . . ., f n}, e(F) denotes the largest
power of 2 that divides every f i , i.e., e(F ) = 2m if 2m / f i , i = 1, 2, . . ., n, but 2m+1 f j for
some j .
Lemma 14.3: If T is a tournament with r vertices and frequency set { f 1 , f 2 , . . ., f n }, then
for each k 1 there is a tournament with kr vertices and frequency set {k f 1 , k f 2 , . . ., k f n }.

Proof For each i, 1 i k, let Ti be a copy of T with vertices ui1 , ui2 , . . ., uir . Orient the arcs between Ti and T j , i < j , so that uik exactly dominates the r/2 vertices u j,k+1,
u j, k+2, . . ., u j, (k+r)/2 where the second subscripts are interpreted modulo r. Every vertex
has had its scores increased by r(k 1)/2 and since we started with k copies of T , the
resulting tournament has kr vertices and frequency set {k f 1 , k f 2 , . . ., k f n }.
q

412

Score Structure in Digraphs

The following result is also due to Alspach and Reid [3].

Theorem 14.9 Let f 1 < f 2 < . . . < f n be even positive integers, let e( f 1 , f 2 , . . ., f n) = 2m
and let f i = 2m ki , 1 i n. Then
i. if an even number of the ki0 s is odd, then
n

N( f 1 , f 2 , . . ., f n) = f i and
i=1

ii. if an odd number of the ki0 s is odd and j is the smallest index for which ki is odd, then
n

N( f 1 , f 2 , . . ., f n) = f j + f i .
i=1

Proof Since 2m is the highest power of 2 which is a factor of all the f i0 s, at least one of
the ki0 s is odd. Hence, in case (i) there are at least two odd ki0 s and n 2. By Theorem 14.8,
n

there is a tournament T on ki vertices with frequency set {k1 , k2 , . . ., kn }. By Lemma

i=1

14.3 with k

= 2m ,

there is a tournament on f i vertices and frequency set { f 1 , f 2 , . . ., f n }.

i=1

This proves (i).

n
Now, consider Case (ii) with (k1 , k2 ) 6= (1, 2). First, we show that N( f 1 , f 2 , . . ., f n ) = f i
i=1

is impossible. Let i be the score occurring with frequency f i . Then

n

i fi =
i=1

f 1 + f 2 + ... + f n
2

( f 1 + f 2 + ... + f n)( f 1 + f 2 + ... + f n 1)

.
2

Clearly, the left hand side is divisible by 2m , while the right hand side is not. Thus,
n
n
N( f 1, f 2 , . . ., f n) > f i . Also, N( f 1 , f 2, . . ., f n) > f k + f i , for k < j , because of the same
i=1

problem regarding divisibility by 2m . Thus,

i=1

N( f 1 , f 2 , . . ., f n) f j + f i .
i=1
n

If n 3, by Theorem 14.8, there is a tournament on ki vertices with frequency set K =

i=1

{k1 , k2 , . . ., kn }. As k j is odd, by Lemma 14.3 there is a tournament on k j + ki vertices

i=1

with frequency set K . If n = 2 and k1 is odd, there is a tournament on k1 + k1 + k2 vertices

with frequency set K (use Lemma 14.1 if gcd{k1 , k2 } > 1). If k1 is even and k2 is odd,
again there is a tournament with k1 + k2 + k2 vertices and frequency set K by application of
Theorem 14.8 and Lemma 14.1 if gcd {k1 , k2 } > 1, or by Lemma 14.2 if gcd {k1 , k2 } = 1.
In case n = 1 and k1 is odd, then an almost regular tournament of order 2k1 has frequency
set F . In all of the above cases employ Lemma 14.3 with k = 2m to obtain the result in (ii).

Graph Theory

413

Finally, let F = {2m, 2m+1 }, m 1. As above, N(2m, 2m+1 ) 2m+2 . Any tournament
with 2m scores of 2m+1 2, 2m scores of 2m+1 + 2 and 2m+1 scores of 2m+1 shows that
N(2m, 2m+1 ) = 2m+2 as required in (ii).
q

14.3 Score Sets in Tournaments

Definition: The set of distinct scores in a tournament T is called the score set of T .
It is easy to see that every singleton set {k}, k 0, is a score set. Reid [217] showed that
a nonempty set S of non-negative integers is a score set whenever |S| = 1, 2, 3 or whenever
S is either an arithmetic or geometric progression. Reid conjectured that any nonempty set
S of non-negative integers is a score set. Hager [96] proved the conjecture for |S| = 4, 5.
This problem proved to be more resistant than that of frequency set discussed above.
If the set {x1 , x2 , . . ., xk } is the score set of some n-tournament T , then there are multiplicities m1 , m2, . . ., mk (positive integers) such that xi occurs as a score exactly mi times in
T . These mi are not necessarily distinct, so they are not the frequencies discussed above.
Therefore, by Landaus Theorem, {x1 , x2 , . . ., xk } is the score set of some n-tournament T
if and only if there exist positive integers m1 , m2 , . . ., mk such that

mixi i=1 i , for 1 j k, with equality for j = k.

i=1
2
j

Consequently, the connection to tournaments is removed, and Reids conjecture becomes strictly an arithmetical supposition. Yan [271] proved the conjecture by pure arithmetical analysis.
Pirzada and Naikoo [200] proved by construction that if s1 , s2 , . . ., s p are p nonnegative
s1 < s2 < . . . < s p , then there exists a tournament with score set
 integers with 
2

S=

s1 , si , . . ., si . More results on score sets in tournaments can be found in [197,

i=1

i=1

198, 200]. Also, the reconstruction of complete tournament can be seen in [119, 120]. The
concept of scores in hyper tournament can be found in Zhou et.at [272]. The literature on
kings in tournaments can be found in [130, 131, 132, 133, 155, 182, 184, 219, 220].

14.4 Lexicographic Enumeration and Tournament

Construction
Definition: Let [si ]n1 be any sequence of integers. The (transitive) deviation sequence of
[si ]n1 is defined to be the sequence [d(i)] = [si i + 1] and d(i) is called the deviation of si .
It is easy to see that [si ]n1 is non-decreasing if 
and only
 if d(i) d(i + 1) 1 for each i < n.
k

i=1

i=1

Also, for each k = 1, 2, . . ., n, d(i) = si

k
2

. From this, it follows that a sequence

414

Score Structure in Digraphs

[si ]n1 of non-negative integers in non-decreasing order is a score sequence if and only if its
k

deviation sequence [d(i)]n1 satisfies d(i) 0 for k = 1, 2, . . ., n with equality for k = n.

i=1

Let [si ]n1 and [s0i ]n1 be score sequences of length n. Then, say that [si ]n1 precedes [s0i ]n1 if there
exists a positive integer k n such that si = s0i for each 1 i k and sk < s0k . They are equal
if equality holds for all i. In symbols, [si ] [s0i ] means [si ] proceeds [s0i ]. Further, [s0i ] is
the successor of [si ] if they are distinct; [si ] [s0i ] and [s0i ] [s00i ] whenever [si ] [s00i ]. An
enumeration of all score sequences of a given length with the property that the successor of
any score sequence follows it immediately in the list is called a exicographic enumeration.
Clearly, [0, 1, 2, . . ., n 1] is not the successor of any score sequence of length n and thus it
is always the first in the lexicographic enumeration. Also, [si ] has no successor if and only
if sn s1 1.
If we know the first sequence in a lexicographic enumeration, then we can complete the
work provided we know how to get the successor of any given sequence. The following
algorithm due to Gervacio [85] gives the successor [s0i ] of [si ], if it exists.

Algorithm
1. Determine the maximum k such that sn sk 2.
2. Let s0i = si for all i < k.
3. Let s0k = sk + 1.
j

4. Let s0j = sk + 1 until s0i <

i=1

 
j
2

.
j

5. Let t be the minimum j such that s0i <

i=1

6. Let s0i = i 1 for all i, t < i n.

 
j
2

, set st0 =

t
2

t1

s0i .
i=1

Constructing a tournament: One method of constructing a tournament with a given

score sequence can be found in [16]. The method we give here is due to Gervacio [85].
First, we have the following observations.
Lemma 14.4 Let [si ]n1 be a score sequence with deviation sequence [d(i)].
a. If max {d(i)} = M > 0, then for each 1 k M , there exists a vertex u such that
d(u) = k.
b. If min {d(i)} = m < 0, then for each 1 k m, there exists a vertex u such that
d(u) = k.

Graph Theory

415

Proof
a. Clearly, the result is true for k = M . Now, we show that it is true for all 1 k M by
induction on k. Assume the result to be true for k + 1, i. e., there exists a vertex u such
that d(u) = k + 1. Since d(n) 0, there exists t > u such that d(u) = d(t) > d(t + 1).
Since d(t) d(t + 1) + 1, d(t + 1) = d(u) 1 = k. Hence (a) holds.
b. This can be proved by using the argument as in (a).
Lemma 14.5 Let [si ]n1 be a score sequence with deviation sequence [d(i)]. If c is the
number of negative terms, then c max{d(i)}.

Proof Let p = max {d(i)}. If p = 0, then c = 0 and the result holds. For p > 0, we have
the following cases.
Case 1 There exists a vertex k such that d(k) < 0 and |d(k)| p. By Lemma 14.4, c
|d(k)| p.
Case 2

For each non-negative deviation d(k), |d(k)| < p.

Let q = max{|d(i)| : d(i) < 0} and let c < p. Then using Lemma 14.4,

|d(i)| 1 + 2 + . . .+ q

d(i)<0

+(p q)q. But

|d(i)| =

d(i)<0

Hence,

|d(i)|, and by Lemma 14.4,

d(i)>0

|d(i)| 1 + 2 + . . .+ p.

d(i)>0

q(q + 1)
p(p + 1)
>
.
2 + (p q)q
2

This gives the quadratic inequality p2 (2q 1)+q(q 1) < 0 and this implies that q 1 <
p < q, which is absurd, since p and q are integers. Hence, c p.
q
Now, we describe and validate the above algorithm.
Construction algorithm
Let [si ]n1 be a score sequence with deviation sequence [d(i)]n1. First take n vertices arranged
horizontally and labelled 1, 2, . . ., n from left to right.
Subdivide [d(i)] into maximal non-increasing segments and denote by p the number of segments in the subdivision. Let ni be the number of negative deviations in the ith
segment, counting from left to right.

Step 1

Let j be the last integer such that d( j) > 0. If no such j exists, go to step 6. Else,
determine the least integer q such that ni d( j). For each i in the segments to the left of

Step 2

iq

the qth segment such that d(i) < 0, let d 0 (i) = d(i) + 1 and draw the arc ji.
Step 3

Let = d( j) ni and choose a smallest (negatively largest) deviations d(i) in

i<q

the qth segment. For each such d(i), let d 0 (i) = d(i) + 1 and draw the arc ji. Let d 0 ( j) = 0.

416

Score Structure in Digraphs

Step 4

For all other deviations d(i) not changed in the preceding steps, let d 0 (i) = d(i).

Step 5

If [d 0 (i)] 6= [0], go to step 1 using [d 0 (i)] in place of [d(i)].

Step 6

Whenever u < v and there is no arc between u and v, draw the arc vu.

Step 7 The resulting digraph is a tournament with score sequence [si ]n1 .

Now, we analyse the algorithm to verify its validity. Clearly, step 1 can always be carried
out. Step 2 can be done in view of Lemma 14.5. Step 3 can be implemented because of
step 2. Obviously, step 4 can be done, and after this step, [d 0 (i)] satisfies d 0 (i) d 0 (i + 1) 1
for all 1 i n. Let D be the digraph formed when [d 0 (i)] = [0]. Then for each vertex i in
D,

and

di+ (D) =

d(i),
0,

i f d(i) 0
i f d(i) < 0

di (D) =

d(i),
d(i),

i f d(i) 0
i f d(i) < 0

Let T be the tournament formed after step 6 and let i be any vertex of T . If d(i) 0, then
si = di+ = di+ (D) + i 1 = si . If d(i) < 0, then di = di (D) + n i = d(i) + n i , and thus
si = di+ = (n 1) di = d(i) + i 1 = si .
Example Let [si ] = [1, 1, 2, 2]. Then [d(i)] = [1, 0, 0, 1].
The resulting digraph after using above algorithm upto [d 0 (i)] = [0] is shown in Figure
14.2. To get the tournament, add all arcs i j (i > j ).

Fig. 14.2

Graph Theory

417

14.5 Simple Score Sequences in Tournaments

Definition: A score sequence is simple (uniquely realisable) if it belongs to exactly one
tournament. Every score sequence of tournaments with fewer than five vertices is simple,
but the score sequence [1, 2, 3, 3, 3] is not simple, since the tournaments in Figure 14.3 are
not isomorphic.

Fig. 14.3

We have the following observations.

Lemma 14.6
simple.

A score sequence S is simple if and only if every strong component of S is

The following result due to Avery [8] gives a condition for determining simple score
sequence in tournaments.
Theorem 14.10 (Avery) A strong score sequence is simple if it is one of [0], [1, 1, 1],
[1, 1, 2, 2] or [2, 2, 2, 2, 2].

Fig. 14.4

Corollary 14.3 The score sequence S is simple if and only if every strong component of
S is one of [0], [1, 1, 1] , [1, 1, 2, 2] or [2, 2, 2, 2, 2].
Hence it is possible to decide whether a given score sequence S is simple by using
Theorem 14.10 to determine the strong components of S and then applying Corollary 14.3.
Let s(n) denote the number of simple score sequences of order n. It is easy to show that
s(n) satisfies the following recurrence relation, which can be used to evaluate s(n).

418

Score Structure in Digraphs

Theorem 14.11
s(0) = 1.

s(n) = s(n 1) + s(n 3) + s(n 4) + s(n 5), where s(k) = 0 if k < 0, and

14.6 Score Sequences of Self-Converse Tournaments

We know, the converse of an n-tournament Tn is the tournament Tn0 obtained by reversing
the orientation of all the arcs in Tn . A tournament is called self-converse if Tn
= Tn0 . The
transitive tournaments are examples of self-converse tournaments.
The following characterisation of score sequences of self-converse tournaments is due
to Eplett [71].
Theorem 14.12 (Eplett)
tournament if and only if

A score sequence [si ]n1 is the score sequence of a self-converse

(14.12.1)

si + sn+1i = n 1,

for 1 i n .

14.7 Score Sequences of Bipartite Tournaments

A bipartite tournament T is an orientation of a complete bipartite graph. Clearly, the vertex
set of T is the union of two disjoint nonempty sets X and Y , and arc set of T comprises
exactly one of the pairs (x, y) or (y, x) for each x X and each y Y . If the orders of X and
Y are m and n respectively, T is said to be an m n bipartite tournament.
A bipartite tournament may be used to represent competition between two teams and
each player competes against everyone on the opposing team. The score sv of vertex v is
the number of vertices it dominates and for a bipartite tournament there is a pair of score
sequences, one sequence for each set. For example, the bipartite tournament in Figure 14.5
has sequence [4, 3, 2, 0] and [2, 2, 2, 1].

Fig. 14.5

Graph Theory

419

Definition A bipartite tournament is reducible if there is a nonempty proper subset of

its vertex set to which there are no arcs from the other vertices, otherwise irreducible. A
component is a maximal irreducible sub-bipartite tournament. A non-trival component
contains at least two vertices one from each partite set.
A bipartite tournament is consistent if it contains no directed cycles. It can be easily
seen that a bipartite tournament is consistent if and only if, for v and w in the same partite
set, v dominates every vertex which w dominates if its score is at least that of w.
The converse of a bipartite tournament is obtained by reversing the direction of all its
arcs and a bipartite tournament is self-converse if it is isomorphic to its converse.
Now, assume that the partite sets of bipartite tournaments have a fixed ordering with X
first and Y second. Then a given bipartite tournament T has associated with it two bipartite
graphs (on the same sets of vertices) in a natural way, one graph containing those edges
corresponding to the arcs directed from X to Y , the other from Y to X . For example, two
graphs of the bipartite tournament T are shown in Figure 14.6. The two graphs are relative
complements as bipartite graphs.

Fig. 14.6

Clearly, the pairs of score sequences of bipartite tournaments and pairs of degree sequences
of bipartite graphs are equivalent.
Lemma 14.7

Let A = [a1, . . ., am ] and B = [b1 , . . ., bn ] be sequences of integers and let

A = [n a1 , . . ., n am] and B = [m b1 , . . ., m bn]. Then the following are equivalent.
1. A and B are the score sequences of a bipartite tournament.
2. A and B are the score sequences of a bipartite tournament.
3. A and B are the score sequences of a bipartite graph.
4. A and B are the score sequences of a bipartite graph.
The following observation can be easily established.
Lemma 14.8 If v and v0 are vertices in the same partite set of a bipartite tournament
T , if sv s0v , and if there is a vertex w which is dominated by v and which dominates v0 ,
then there is another vertex w0 which is dominated by v0 and which dominates v, that is,
v w v0 w0 v is a 4-cycle.
The following result is due to Gale [84].

420

Score Structure in Digraphs

Theorem 14.13 (Gale) If A = [a1 , . . ., am ] and B = [b1, . . ., bn ] are sequences of nonnegative integers in nondecreasing order, then A and B are the score sequences of some bipartite tournament if and only if the sequences A0 = [a1 , . . ., am1 ] and B0 = [b1 , . . ., bam , bam+1
1, . . ., bn 1] are.

Proof First assume that A0 and B0 are the score sequences of a bipartite tournament T 0 .
To the first partite set of T 0 , add a new vertex v with arcs directed from it to vertices (in
the second set) with scores b1 , . . ., bam , and to it from the others. The result is a bipartite
tournament with score sequences A and B.
For the converse, it is sufficient to show that if A and B are the score sequences of a
bipartite tournament, then in one realisation, a vertex (in the first set) of score am dominates
vertices of scores b1 , . . ., bam . Among the bipartite tournament realisations of A and B, let
T be the one in which a vertex x of score am is such that the sum S of the scores of the
am
vertices it dominates is as small as possible. Let S > b j . Then there exist vertices y and
j=1

y0 such that x y0 , y x and sy < s0y . By Lemma 14.8, T has a 4-cycle x y0 x0 y x,

and if its arcs are reversed, the result is a bipartite tournament with the same sequences,
but in which score sum of the vertices dominated by x is less than before. Since the sum
was assumed to be minimised, the result follows.
q
Theorem 14.3 gives a natural construction for a canonical tournament T (A, B) from a
given pair of score sequences A and B. The only point which needs clarification is getting B0
into non-decreasing order, i.e., we must specify dominance when a vertex vi must dominate
some but not all vertices y with a particular score. This is done by forming B0 as follows. Let
h and k denote the smallest and largest integers j for which b j = bam . Let A0 = [a1, . . ., am1],
and B0 = [b01 , . . ., b0n] with
b0j =

b j,
f or 1 j < h and h + k am j k,
b j 1 , otherwise

This reduction and the resulting construction is illustrated by the following example,
starting with sequences A = [1, 1, 3, 5, 5] and B = [1, 1, 2, 3, 4, 4] .
A = [1, 1, 3, 5, 5]
A1 = [1, 1, 3, 5]

B = [1, 1, 2, 3, 4, 4]
B1 = [1, 1, 2, 3, 3, 4]

(x5 dominates y1 , y2 , y3 , y4 , y6 )
A2 = [1, 1, 3]

B2 = [1, 1, 2, 3, 3, 3]

(x4 dominates y1 , y2 , y3 , y4 , y5 )
A3 = [1, 1]

B3 = [1, 1, 2, 2, 2, 2]

(x3 dominates y1 , y2 , y3 )

Graph Theory

421

A4 = [1]

B4 = [0, 1, 1, 1, 1, 1]

(x2 dominates y2 )
B5 = [0, 0, 0, 0, 0, 0]

A5 =

(x1 dominates y1 )
Figure 14.7 shows the X to Y arcs resulting from this construction.

Fig. 14.7

A canonical tournament T (A, B) has the following special property. In the subtournament Tr, n induced by {x1 , . . ., xr } and Y , if xr yi and yj xr , then syi syj , that is, bi b j .
The next result is due to Beineke and Moon [20].
Theorem 14.14 If two bipartite tournaments have the same score sequences, then each
can be transformed into the other by successively reversing the arcs of 4-cycles.
It can be noted that Theorem 14.14 does not imply that all bipartite tournaments with
given score sequences have the same number of 4-cycles (they need not), although the
corresponding statement does hold for 3-cycles in tournaments.
The next result first established by Moon [162] and then in the present form by Beineke
and Moon [20] gives a simple criterion for determining whether a pair of sequences are
realisable as scores.
Theorem 14.15 (Moon) A pair of sequences A and B of non-negative integers in nondecreasing order are the score sequences of some bipartite tournament if and only if
k

i=1

j=1

ai + b j kl

for 1 k m and 1 l n, with equality when k = m and l = n.

Further more, the bipartite tournament is irreducible if and only if the inequality is strict
except when k = m, and l = n.

Proof In any bipartite tournament T , the combined scores of any collection of k vertices
from the first set and l from the second must be at least kl, so that the inequalities certainly
hold. Further, if T irreducible, the inequality is strict unless k = m and l = n.

422

Score Structure in Digraphs

Sufficiency If A and B satisfy the inequalities, we show that A0 and B0 satisfy the inequalities reordered as in construction of 14.13. It is easily seen that A0 and B0 are then in
non-decreasing order, and further, their combined sum is
m1

a0i + b0j = mn am (n am ) = (m 1)n.

For a fixed value of k(1 k m 1), assume there is a value of l for which the inequality
does not hold. And let h denote the least such that
k

a0i + b0j < kh ..

It follows from the minimality of h that b0h < k, whence bh k. Now, let p and q be
the least and greatest values of j for which b j = bam and set r = max (h, q). Since the
first p 1 values of b j were unchanged, we have h p and thus bh = . . . = br . Finally, let s
denote the number of j h such that b0j = b j1. If h q, then s q am , and if h > q, then
s = (h q) + (q am) = h am . In either case, am + s r. Therefore,
k+1

h+1

ai + b j = a0i + b0j + b j + ak+1 + s < kh + (r h)bh + am + s

1

kh + (r h)k + r < (k + 1)r,

which is a contradiction. Therefore A0 and B0 satisfy the inequalities, as required.

It is easily seen that if the strict inequalities hold for A and B, no realisation can be
reducible, completing the proof.
q
The next criterion derived by Ryser [227] in the context of (0, 1)-matrices with prescribed row and column sums is equivalent to prescribed degrees in bipartite graphs.
Theorem 14.16 If A and B are sequences of non-negative integers with A in non-increasing
order, then A and B are the score sequences of some bipartite tournament if and only if
k

ai min(k, m b j )

for 1 k m, with equality when k = m. Further, the bipartite tournament is irreducible if

and only if the inequality, is strict for all k < m and 0 < b j < m for all j n.
Remarks In general, one need not check the inequalities for all values of h and k, but
only for those for which the next value in the sequence is different. Thus in order to show
that [5, 5, 5, 3, 3, 2, 2] and [6, 5, 4, 1, 1, 0] belong to a bipartite tournament, we need to
check the inequalities in Theorem 14.16 for k = 3, 5, and 7 only.

Graph Theory

423

Theorem 14.17 If A = [ai ]m1 and B = [b j ]n1 are non-decreasing integer sequences with
0 ai n and 0 b j m, and are such that
a. Am + Bn = mn and
b. Ar + Bs rs, whenever ar < ar+1 and bs < bs+1 .
Then (A, B) is realisable.

(Note that for a sequence L = [xi ]1p , Lq = xi for 1 q p and L0 = 0).

i=1

Proof

We show that the inequality

Ak + Bx kx

(4.17.1)

holds for all 1 k m and 1 x n. If this is not the case for some k and x, let q and s
be the smallest, and r and t be the largest indices such that aq+1 = ak = ar and bs+1 = bx =
bt (q, s = 0). Now, Ar + Bx < kx. Claim that at least one of Ak + Bs < ks and Ak + Bt < kt holds.
For otherwise, (x s)bx < k(x s) and (t x)bx > k(t x) which is impossible. Thus assume
(i) Ak + Bs < ks.
Now, by hypothesis, (ii) Aq + Bs qs and (iii) Ar + Bs rs (observe that if r = m, then
Am + Bn = mn and 0 b j m together imply (iii). Then (i) and (ii) give (k q)ak < (k q)s,
while (i) and (iii) give (r k)ak > (r k)s. These again lead to a contradiction. The case
Ak + Bt < kt can similarly be treated.
q
The following result can be obtained from Theorem 14.16.
Corollary 14.4 If C = [c1 , . . ., cm ] and D = [d1, . . ., dn] be two non-increasing sequences
having equal sum, then the following are equivalent.
k

i. ci min(k, d j ), for k = 1, . . ., m.
ii. d j min(x, ci ), for x = 1, . . ., n.
Now, different bipartite tournaments can have the same score sequences and they can differ
only within irreducible components. That is, they must have the same numbers of components the same numbers of vertices within components, the same scores within components, and the same dominance between components. The general procedure here is
to find a dominating component (a component is called dominating if it has no incoming
arcs), delete its vertices and repeat. While an ordinary tournament has precisely one dominating component, the situation in the bipartite case is slightly different. It is described in
the following result [19] and is a direct consequence of Theorem 14.15.
Theorem 14.18 Let A = [a1 , . . ., am ] and B = [b1 , . . ., bn ] be score sequences (in nondecreasing order) of a reducible m n bipartite tournament.

424

Score Structure in Digraphs

i. If am = n or bn = m, then there is a corresponding trivial dominating component,

consisting of one vertex which dominates all the vertices in the other partite set.
k

ii. Otherwise, if k and x are the largest indices with k < m and x < n such that ai +
x

i=1

b j < kx, then the non-trivial dominating component consists of all the vertices in

j=1

the two partite sets with scores exceeding ak and bk respectively.

The following results can be found in [11].
Theorem 14.19 If A = [ai]m1 and B = [b j ]n1 are realisable pair of score sequences with
0 < ai < n and 0 < b j < m and if |ai ak | 1 for any i, k = 1, 2, . . ., m, then any bipartite
tournament with score sequences A and B is irreducible.

Proof Assume T is a reducible bipartite tournament on partite sets X and Y with score
sequences A and B respectively. Since 0 < ai < n and 0 < b j < m, T has at least two non
trivial components, say C and C0 , with C being the dominating one. If xi C X and
xk C 0 X , then xi dominates all the vertices in Y dominated by xk . Also, there exists
y j C Y and yl C 0 Y such that xi y j xk and xi yl xk . Thus, ai = score (xi )
score (xk ) + 2 = ak + 2. This contradicts the hypothesis, and the result follows.
q
Theorem 14.20 Let A = [ai ]m1 (in non-decreasing order) and B = [bn] be sequences such
that (A, B) is realisable. Let ak , ax be two entries in A with ak > 0 and ax < n. Define a new
sequence A0 = [a0i ]m1 as follows.
a0k = ak 1, a0x = ax + 1

and a0i = ai , for i 6= k, x. Then (A0 , B) is realisable.

Proof: It follows immediately from Theorem 14.17.

Theorem 14.21 If (A, B) is irreducible, i. e., (A, B) is realisable and all its realisation are
irreducible, and if A0 is obtained from A by adding 1 to some entry and B0 is obtained from
B by subtracting 1 from some entry, then (A0 , B0 ) is realisable.

14.8 Uniquely Realisable (Simple) Pairs of Score

Sequences
If (A, B) is realisable, let T denote a realisation on partite sets X and Y . If a pair (A, B) is
realisable and all its realisations are isomorphic, then (A, B) is called uniquely realisable.
The following observations can be found in [11].

Graph Theory

425

Lemma 14.9 Let (A, B) be uniquely realisable. For any entry ai in A and b j in B, let Xi
and Y j be the subsets of X and Y consisting of vertices of scores ai and b j respectively. Then
any cycle in T contains the same number of arcs from Xi to Y j as from Y j to Xi .

Proof If this were not the case for some cycle Z , then reversal of the arcs of Z would
produce an (A, B) realisation non-isomorphic to T .
Lemma 14.10

If (A, B) is irreducible and uniquely realisable, then A or B is constant.

Proof If neither A nor B is constant, let X1 be the set of vertices of minimum score in X and
let X2 = X X1 , and similarly define Y1 and Y2 . Since T is irreducible, every arc is contained
in a cycle. Thus by Lemma 14.9, none of the four subtournaments T (Xi , Yi ), i, j = 1, 2 is
unanimous, i. e., has all its arcs directed from one partite set to the other.
Choose a vertex x1 in X1 of minimum score in T (X1 ,Y1). Then x1 dominates some y2 Y2 .
Consider two cases depending on whether or not y2 dominates some vertex in X2 .
Case (i) Every vertex in X2 dominates y2 . Let uv be an arc from Y1 to X2 . Since score
(y2 ) > score(u) in T , there exists an x X1 such that y2 x u (Fig. 14.8).

Fig. 14.8

But then x u v y2 x is a cycle which violates Lemma 14.9.

Case (ii) Some vertex x2 X2 is dominated by y2 . By the choice of x1 , there exists y1 Y1
such that y1 x1 . If x2 y1 , we again get a 4-cycle which has one arc from Y1 to X1 , but
no arcs in the other direction. So assume y1 x2 (Fig. 14.9). Since score (y2 ) > score
(y1 ) in T , there exists an x X1 with y2 x y1 . If x X2 , we again get a forbidden cycle
x1 y2 x y1 x1 . Thus x X1 . Likewise, there exists y Y1 with x2 y x. But then
x1 y2 x2 y x y1 x1 is a 6-cycle which violates Lemma 14.9.
Since all the possibilities have been exhausted, the result follows.
q

426

Score Structure in Digraphs

Fig. 14.9

We note that if in a realisable pair (A, B), one of the sequences has all entries
as 1s (the sequence is constantly 1), then (A, B) is uniquely realisable. This is illustrated
in Figure 14.10, where A = [1,1, . . ., 1] and B = [b1 , b2 , . . .., bn] and only X to Y arcs are
shown.

Remarks

Fig. 14.10

Now onwards assume that none of the sequences A, B, A and B is constantly 1. Hence the
following observations [232] can be easily proved.
Lemma 14.11

If (A, B) is irreducible and uniquely realisable, then A or B is non-constant.

Lemma 14.12

With A and B as above, the sequence B has precisely two distinct values.

Lemma 14.13 If (A, B) is irreducible, uniquely realisable and if A = [am] , 2 a n 2,

and B = [br , cs ], 1 b < c m 1, and r, s > 0, r + s = n, then r = 1 or s = 1.
It has been shown that if (A, B) is irreducible, uniquely realisable, then one of
the sequences (or its dual) consists entirely of 1s, or one of the sequences is constant and
the other has exactly two distinct values, one of which appears precisely once.
The following result due to Bagga and Beineke [11] gives necessary and sufficient conditions for unique realisability in the irreducible case.

Remarks

Theorem 14.22 (Bagga and Beineke) An irreducible pair (A, B) of score sequences is
uniquely realisable if and only if one of the following holds.

Graph Theory

427

I. (without loss of generality) A = [1m] and B is arbitrary,

I0 . the dual of (I ), that is, A = [(n 1)m ] and B is arbitrary,
II. (without loss of generality) A = [1m1, a] and B = [bn ],
II0 . the dual of II,
III. (without loss of generality) A = [1, am1 ] and B = [2n ],
III0 . the dual of III.

Proof The sufficiency of (I ) has already been noted in the remarks before Lemma 14.11
where Figure 14.10 shows the unique realisation. Now, let T be a realisation of A = [1m1 , a]
and B = [bn ] on partite sets X = {x1 , x2 , . . ., xm } and Y = {y1 , y2 , . . ., yn } respectively. If
(say) xm has score a and it dominates (say) y1 , y2 , . . ., yn , then T xm has score sequences
A1 = [1m1 ] and B1 = [(b 1)na , ba ]. Thus by (I ), (A1 , B1 ) is uniquely realisable. The
unique realisability of (A, B) follows. This proves the sufficiency of (II). The proof of (III)
is similar and the dual cases follow by the remarks before Lemma 14.11.
For proving necessity, induct on m + n. Since (A, B) is irreducible, so m, n 2. If (say) m
= 2, then B = [1n ], and the result follows. Now, assume the result holds for all irreducible
and uniquely realisable pairs of score sequences with combined length less than m + n, and
consider such a pair (A, B) with |A| = m and |B| = n(m, n 3).
Assume A and B are not of the type (I ) or (I 0 ). Then by the remarks after Lemma 14.13,
we have, without loss of generality, A = [am ] and B = [bn1 , c], with 1 < a < n 1, 1 b, c
m 1 and b 6= c.
If y is the vertex of score c in a realisation T of (A, B), then T y has score sequences
A1 = [(a 1)mc , ac] and B1 = [bn1 ]. Now, the unique realisability of (A, B) implies that of
(A1 , B1). Also, by Theorem 14.19, (A1 , B1 ) is irreducible. Thus, by the induction hypothesis,
A1 and B1 belong to one of the six given types. Consider these cases one by one.
i. If B1 = [1n1 ], then B = [1n1 , c], so that A and B are of type (II).
ii. If B1 = [(m 1)n1 ], then A and B belong to (II 0 ).
iii. If B1 = [bn1 ] and A1 = [1m1, a], then c = 1 and a = 2, so that A = [2m ] and B = [1, bn1].
This is of type (III).
iv. If B1 = [bn1 ] and A1 = [d, (n 2)m1], we get A and B of type (III 0 ).
v. If B1 = [2n1] and A1 = [1, am1 ], then b = 2, a = 2 and c = m 1. Thus, A = [2m ] and
B = [2n1 , m 1]. Using Moons theorem, we get 2m + 2(n 1) + (m 1) = mn, so that
m = 2n3
n3 . It follows that n = 6 and m = 3. But then A and B are both constant, a
contradiction. Therefore this case is not possible.
vi. The possibility of B1 = [(m 2)n1 ] and A1 = [am1 , n 2] follows by duality.

428

Score Structure in Digraphs

This exhausts all the possibilities and hence by induction, the result is completely proved.
q

Now, assume that (A, B) is a realisable pair and Q1 , Q2 , . . ., Q p are the irreducible components of a realisation T of (A, B). Also, let Qk has score sequences Ak and Bk , 1 k p.
Then (A, B) is uniquely realisable if and only if (Ak , Bk ) is uniquely realisable for all k.

14.9 Score Sequences of Oriented Graphs

An oriented graph is a digraph with no symmetric pairs of directed arcs and with no loops.
Avery [233] extended the concept of score structure to all oriented graphs.
Definition: Let D be an oriented graph with vertex set V = {1, 2, . . ., n}, and let d + (v)
and d (v) be the outdegree and indegree respectively of vertex v. Then the score of vertex
v denoted by av is defined as av = n 1 + d + (v) d (v) with 0 av 2n 2. The sequence
of scores is called the score list, and A = [a1, a2 , . . ., an ] arranged in non-decreasing order is
called the score sequence of D.
Any oriented graph can be interpreted as the result of a round robin competition in which
ties (draws) are allowed, that is, the participants play each other once, with an arc from u
to v if and only if u defeats v. A player receives two points for each win and one point for
each tie, as is frequently the case in sports such as soccer, ice hockey and cricket. With this
scoring system , player v obtains a total of av points. An arc from u to v denoted by u v
is written as u(1 0)v, and u(0 0)v means that neither u v nor v u.
Definition: A triple in an oriented graph is an induced subdigraph with three vertices.
A cyclic triple is an intransitive triple of the form u v w u. Any triple can be of the
form u(x1 x2 )v(y1 y2 )w(z1 z2 )u, where 0 xi , yi , zi 1 with 0 xi , yi , zi 1.
The following result [9] extends a result of Ryser [234] which showed that if two tournaments have the same score structure, then each can be transformed to the other by successively reversing the arcs of appropriate cyclic triples.
Theorem 14.23 Let D and D0 be two oriented graphs with the same score sequence.
Then D can be transformed to D0 by successively transforming appropriate triples in one of
the following ways.
Either (a) by changing a cyclic triple u(1 0)v(1 0)w(1 0)u to a transitive triple u(0
0)v(0 0)w(0 0)u, which has the same score sequence, or vice versa.
or (b) by changing an intransitive triple u(1 0)v(1 0)w(0 0)u to a transitive triple
u(0 0)v(0 0)w(0 1)u, which has the same score sequence, or vice versa.
The following result due to Avery [9] gives a constructive condition for a non-negative
sequence in non-decreasing order to be a score sequence of some oriented graph. A short
proof of this result is due to Pirzada et. al. [199].

Graph Theory

429

Theorem 14.24 (Avery) A sequence of non-negative integers in non-decreasing order

is the score sequence of an oriented graph if and only if
k

ai k(k 1)
i=1

for 1 k n, with equality for k = n .

The following result can be found in [9].
Theorem 14.25 Among all oriented graphs with a given score sequence, those with the
fewest arcs are transitive.

Proof Let A be a score sequence and let D be a realisation of A that is not transitive.
Then D has an intransitive triple. There are two types of intransitive triples, a cyclic triple,
which can be transformed by operation (a) of Theorem 14.23 to a triple with the same
score sequence and three arcs fewer, and a triple u(1 0)v(1 0)w(0 0)u, which can be
transformed by operation (b) of Theorem 14.23 to a triple with the same score sequence
and one arc fewer. So in either case, we obtain a realisation of A with fewer arcs.
q
The next result [9] provides a useful recursive test of whether a given sequence of nonnegative integers is the score list of an oriented graph. We note that a transmitter is a vertex
with indegree zero.
Theorem 14.26 Let A be a sequence of n integers between 0 and 2n 2 inclusive and let
A0 be obtained from A by deleting the greatest entry 2n 2 r say, and reducing each of the
greatest r remaining entries in A by one. Then A is a score list if and only if A0 is a score
list.

Proof Clearly, in a transitive oriented graph, any vertex of greatest score is a transmitter.
Let A0 be a score list of some oriented graph D0 . Then an oriented graph D with score list
A can be obtained by adding a transmitter that is adjacent to just those vertices of whose
scores are not reduced in going from A to A0 .
For the converse, we show that there is an oriented graph with score list A in which
a transmitter v with score 2n 2 r is adjacent to the (other) n 1 r vertices with least
scores. By Theorem 14.25, there is a transitive oriented graph D with score list A, in which
a vertex v with greatest score 2n 2 r is a transmitter. Let U be the set of r vertices, apart
from v, with the greatest scores in A, and let W be the set V {v U}.
Let v be adjacent in D to vertices u1 , u2 , . . ., uk of U . Then there are exactly k vertices,
say w1 , w2 , . . ., wk of W not adjacent from v. Now, ui cannot be adjacent to wi , since D is
transitive. Neither can wi be adjacent to ui , since taken together with the transitivity of
D this implies that the score of wi is greater than the score of ui , which is contrary to the
assumption. Thus wi (0 0)ui for all i.
Now, transforming all triples v(1 0)ui (0 0)wi (0 0)v to triples v(0 0)ui (0 1)wi (0
1)v, the vertex scores remain unchanged. This forms an (not necessarily transitive) oriented

430

Score Structure in Digraphs

graph D1 with score list A in which the transmitter v is adjacent to all vertices of W and none
of U , as required.
q
Theorem 14.26 provides an algorithm for determining whether a given non-decreasing
sequence A of non-negative integers is a score sequence, and for constructing a corresponding oriented graph. At each stage, we form A0 according to Theorem 14.26, such that
scores of A0 are also non-decreasing. If an = 2n 2 r, this means deleting an and reducing
the r greatest remaining entries by one each to form A0 = |a01 , a02 , . . ., a0n1 | while ensuring
that this is also non-decreasing. Arcs of an oriented graph are defined by n v if and only
if a0v = av . If this procedure is applied recursively, then first it tests whether A is a score
sequence and if A is a score sequence, an oriented graph (A) with score sequence A is
constructed.
Example
Stage

1
2
3
4

Let n = 5, A = [2, 4, 4, 4, 6].

A
[2, 4, 4, 4, 6]
[2, 3, 3, 4]
[2, 2, 2]
[1, 1]

B
[2, 3, 3, 4]
[2, 2, 2]
[1, 1]
[0]

Arcs of (A)
5 4, 5 1
41

Thus A is a score sequence.

We have the following observations [9] about (A).
Theorem 14.27

The oriented graph (A) is transitive for any score sequence A.

Theorem 14.28
than (A).

There is no oriented graph with score sequence A which has fewer arcs

One more method of constructing oriented graph with a given score sequence can be
found in Pirzada [188].
The following is an equivalent statement of Theorem 14.24. A sequence of non-negative
integers A = [ai ]n1 in non-decreasing order is a score sequence of an oriented graph if and
only if for each subset I [n] = {1, 2, . . ., n}
ai 2
iI

 
|I|
2

with equality for |I| = n.

The following inequalities for scores in oriented graphs can be found in [235].
Theorem 14.29 A sequence A = [ai ]n1 of non-negative integers in non-decreasing order
is a score sequence of an oriented graph if and only if for every subset I = [n],

Graph Theory

431

ai (i 1) +
iI

iI

|I|
2

with equality when I = [n]

Theorem 14.30 A sequence A = [ai ]n1 of non-negative integers in non-decreasing order
is a score sequence if and only if for any subset I [n],
1

ai (i 1) + 2 |I|(2n |I| 1),

iI

iI

with equality for I = [n].

Theorem 14.31

If A = [ai ]n1 is a score sequence of an oriented graph, then for each i,

i 1 ai n + i 2.

A necessary condition for a score sequence in oriented graphs to be self-converse, can

be found in [192].

14.10 Score Sets in Oriented Graphs

Definition: The set A of distinct scores of vertices in an oriented graph D is called the
score set of D.
Definition: A digraph D is diregular if dv+ = dv = k holds for each vertex v in D. In case
of an oriented graph D with n vertices, av = n 1 + dv+ dv, for each vertex v in D, and
when dv+ = dv = k (say), then av = n 1 + k k = n 1 for each v in oriented graph D. Thus
an oriented graph D with n vertices is diregular if av = n 1, for all v in D.
Now, we have the following result, the proof of which is obvious.
Lemma 14.14 The number of vertices in an oriented graph with at least two distinct
scores does not exceed its largest score.
The following result is given by Pirzada and Naikoo [196].
Theorem 14.32 (a) (Pirzada and Naikoo) Let A = {a, ad, ad 2 , . . ., ad n }, where a and
d are positive integers with a > 0 and d > 1. Then there exists an oriented graph D with
score set A, except for a = 1, d = 2, n > 0 and for a = 1, d = 3, n > 0.

Proof We induct on n. For n = 0, there is a positive integer a > 0, so that a + 1 > 0. Let
D be a diregular oriented graph having a + 1 vertices. Then av = a + 1 1 = a, for all v D.
Therefore score set of D is A = {a}. This proves the result for n = 0. If n = 1, then there are
positive integers a and d with a > 0 and d > 1, and for a = 1, d 6= 2, 3.

432

Score Structure in Digraphs

Now, three cases arise: (I) a > 1, d > 2, (II) a > 1, d = 2 and (III) a = 1, d > 3.
(I) Let a > 1, d > 2. Therefore, a + 1 > 0. Let D1 be a diregular oriented graph having
a + 1 vertices. Then av = a + 1 1 = a, for all v D1 .
Now, ad 2|V (D1 )| + 1 = ad2(a + 1) + 1 = ad 2a 1 3a 2a 1 = a 1 > 0, as d 3
and a > 1. That is, ad 2|V (D1 )| + 1 > 0. Let D2 be a diregular oriented graph having
ad 2|V (D1 )| + 1 vertices. Then au = ad 2|V (D1 )| + 1 1 = ad 2|V (D1 )|, for all
u D2 .
Let there be an arc from every vertex of D2 to each vertex of D1 , so that we get an
oriented graph D (which includes D1 to D2 together with all the new arcs from D2 and
D1 ) having |V (D1 )| + |V(D2 )| = a + 1 + ad 2|V(D1 )| + 1 = a + 1 + ad 2(a + 1) + 1 =
ad a vertices with av = a, for all v D1 , and au = ad 2|V (D1 )| + 2|V (D1 )| = ad, for
all u D2 . Therefore score set of D is A = {a, ad}.
(II) Assume a > 1, d = 2. First take a = 2, d = 2. Then ad = 4 > 0. Let D be an oriented
graph having ad = 4 vertices, say, v1 , v2 , v3 , and v4 in which v1 v3 and v2 v4 , so
that av1 = av2 = 2 + 4 2 = 4 = ad, and av3 = a44 2 = 2 = a. Therefore D is an oriented
graph having ad vertices with score set A = {a, ad}.
Now, take a > 2, d = 2. Let D1 be a diregular oriented graph having 2 vertices, say v1
and v2 . Then avi = 2 1 = 1 for all vi D1 , where 1 i 2.
Again, a > 2 or a 2 > 0. Let D2 be a diregular oriented graph having a 2 vertices,
say v3 , v4 , . . ., va . Then av j = a 2 1 = a 3, for all v j D2 , where 3 j a.
Let there be an arc from every vertex of D2 to each vertex of D1 , so that we get an
oriented graph D3 (which includes D1 and D2 together with all the new arcs from D2
to D1 ) having 2 + a 2 = a vertices with avi = 1, for all vi D1 , where 1 i 2, and
av j = a 3 + 2(2) = a + 1, for all v j D2 , where 3 j a.
Again, a > 2 > 0. Let D4 be a diregular oriented graph having a vertices, say w1 , w2 , . . .,
wa . Then awk = a 1, for all wk D4 , where 1 k a.
Let there be a arcs from a distinct vertices of D4 to a distinct vertices of D3 (wq vq ,
for all q = 1, 2, . . ., a), so that we get an oriented graph D (which includes D3 and D4
together with all the new arcs from D4 to D3 ) having a + a = 2a = ad vertices with
avi = 1 + a 1 = a, for all vi D3 , where 1 i 2, = a + 1 + a 1 = 2a, for all v j D3 ,
where 3 j a, and awk = a 1 + 2(1) + a 1 = 2a, for all wk D4 , where 1 k a.
Therefore score set of D is A = {a, 2a} = {a, ad}.
(III) Finally, let a = 1, d > 3. Therefore, a + 1 > 0. Let D1 be a diregular oriented graph
having a + 1 vertices. Then av = a + 1 1 = a, for all v D1 .
Now, ad 2|V (D1 )| + 1 = ad 2(a + 1) + 1 = ad 2a 1 4a 2a 1 = 2a 1 > 0, as
d 4 and a = 1, i.e., ad 2|V (D1 )| + 1 > 0. Then as in (I ), we have an oriented graph
D having ad a vertices with score set A = {a, ad}.
Hence in all these cases, we get an oriented graph D with score set A = {a, ad}. This
shows that the result is also true for n = 1.

Graph Theory

433

Assume the result to be true for all p 1. We show that the result is true for p + 1.
Let a and d be positive integers with a > 0 and d > 1, and for a = 1, d 6= 2, 3. Therefore
by induction hypothesis, there exists an oriented graph D1 having |V (D1 )| vertices
with score set {a, ad, ad 2 , . . ., ad p }.
Once again, we have either (I) a > 1, d > 2, or (II) a > 1, d = 2, or (III) a = 1, d > 3.
Obviously, for d > 1, in all the above possibilities, ad p+1 2ad p , and the score set of
D1 , namely, {a, ad, ad 2 , . . ., ad p } has at least two distinct scores for p 1. Therefore
by Lemma 14.14, |V (D1 )| ad p . Hence, ad p+1 2|V(D1 )|, or ad p+1 2|V (D1 )| +1 > 0.
Let D2 be a diregular oriented graph having ad p+1 2|V (D1 )| + 1 vertices. Then av =
ad p+1 2|V (D1 )| + 1 1 = ad p+1 2|V (D1 )|, for all v D2 .
Let there be an arc from every vertex of D2 to each vertex of D1 , so that we get an
oriented graph D (which includes D1 and D2 together with all the new arcs from D2
to D1 ) having |V (D1 )| + |V(D2 )| vertices with a, ad, ad 2, . . ., ad p as the scores of the
vertices of D1 , and av = ad p+1 2|V (D1 )| + 2|V (D1 )| = ad p+1 , for all v D2 . Therefore
score set of D is A = {a, ad, ad 2 , . . ., ad p , ad p+1 }, proving the result for p + 1. Hence
the result follows.
q
That no oriented graph exists when either a = 1, d = 2, n > 0 or a = 1, d = 3, n > 0, is
proved in the following theorem.
Theorem 14.32 (b) There exists no oriented graph with score set A = {a, ad, ad 2 , . . ., ad n },
n > 0, when either (i) a = 1, d = 2, or (ii) a = 1, d = 3.
We now have the following result [201].
Theorem 14.32 (c) If a1 , a2, . . ., an are n non-negative integers with a1 < a2 < . . . < an .
Then there exists an oriented graph D with score set A = {a01 , a02 , . . ., a0n }, where
a0i =

ai1 + ai + 1,
ai ,

f or i > 1,
f or i = 1.

Remarks

1. From Theorem 14.32 ( c ), it follows that every singleton set of non-negative integers
is a score set of some oriented graph.
2. As we have shown in Theorem 14.32(b), i. e., the sets {1, 2, 22 , . . ., 2n } and {1, 3, 32 , . . .,
3n } cannot be the score sets of any oriented graph for n > 0. It follows, therefore, that
the above results cannot be generalised to conclude that any set of non-negative integers forms the score set of some oriented graph. However, there can be other special
classes of non-negative integers which can form the score set of an oriented graph,
and the problem needs further investigations.
Pirzada and Naikoo [195] have obtained some results on degree frequencies in oriented
graphs. More results on scores, score sets and kings in oriented graphs can be seen in [199,
201, 203, 207].

434

Score Structure in Digraphs

14.11 Uniquely Realisable (Simple) Score

Sequences in Oriented Graphs
An oriented graph D is reducible if it is possible to partition its vertices into two nonempty
sets V1 and V2 in such a way that there is an arc from every vertex of V2 to each vertex of
V1 . Let D1 and D2 be induced digraphs having vertex sets V1 and V2 respectively. Then D
consists of D1 and D2 and arcs from every vertex of D2 to each vertex of D1 . We write D =
[D1, D2 ]. If this is not possible, then the oriented graph D is irreducible. Let D1 , D2 , . . ., Dk
be irreducible oriented graphs with disjoint vertex sets. Then D = [D1 , D2 , . . ., Dk] denotes
the oriented graph having all arcs of Di , 1 i k and there are arcs from every vertex of D j
to each vertex of Di , 1 i < j k. Here D1 , D2, . . ., Dk are called irreducible components of
D. Such a decomposition is called as irreducible component decomposition of D, which is
unique.
Definition: A score sequence A is said to be irreducible if all the oriented graphs D with
score sequence A are irreducible.
In case of ordinary tournaments, the score sequence used to decide whether a tournament
T having score sequence S, is strong or not. This is not true in case of oriented graphs. For
example, the oriented graphs D1 and D2 in Figure 14.11, both have score sequence [2, 2, 2]
but D1 is strong and D2 is not.

Fig. 14.11

The following result due to Pirzada [186] characterises irreducible oriented graphs.
Theorem 14.33 Let D be an oriented graph having score sequence. Then D is irreducible
if and only if, for k = 1, 2, . . ., n 1
k

ai k(k 1)

(14.33.1)

i=1

and

ai = n(n 1)

(14.33.2)

i=1

Proof Suppose D is an irreducible oriented graph having score sequence [ai ]n1 . Condition
(14.33.2) holds since Theorem 14.24 has already established it for any oriented graph. To
verify inequalities (14.33.1) we observe that for any integer k < n, the subdigraph induced
by any set of k vertices has a sum of scores k(k 1). Since D is irreducible, there must be

Graph Theory

435

an arc from at least one of these vertices to one of the other n k vertices, or there is no arc
from these k vertices to other n k vertices. Thus, for 1 k n 1
k

ai > k(k 1).

i=1

For the converse, suppose conditions (14.33.1) and (14.33.2) hold, we know by Theorem
14.24 that there exists an oriented graph D with these scores. Assume that such an oriented
graph D is irreducible. Let D = [D1 , D2, . . ., Dk] be the irreducible component decomposition
of D. If m is the number of vertices in D1 , then m < n, and the following equation holds,
k

ai = m(m 1),
i=1

which is a contradiction. This proves the converse part.

The following result due to Pirzada [186] can be proved easily.

Theorem 14.34 Let D be an oriented graph with score sequence A = [ai ]n1 . Suppose that
k

i=1

i=1

ai = p(p 1), ai = q(q 1) and for p + 1 k q 1, where 0 p < q n. Then the

subdigraph induced by the vertices v p+1 , v p+2 , . . ., vq is an irreducible component of D with

score sequence [a p+1 2 p , . . ., aq 2 p ].
Now, A is irreducible if D is irreducible and the irreducible components of A are the score
sequences of the irreducible components of D. Theorem 14.34 shows that the irreducible
components of A are determined by the successive values of k for which
k

ai = k(k 1), 1 k n.

(14.33.3)

i=1

We illustrate it with the following example.

Let A = [1, 2, 3, 8, 8, 8, 13, 13]. Equation (14.33.3) is satisfied for k = 3, 6, 8. Thus
irreducible components of S are [1, 2, 3], [2, 2, 2] and [1, 1] in ascending order.
Definition: A score sequence is simple if it belongs to exactly one oriented graph. We
characterise simple score sequences of oriented graphs. First we have the following observation.
Lemma 14.15 The score sequence A of an oriented graph is simple if and only if every
irreducible component of A is simple.
The following result due to Pirzada [186] determines which irreducible score sequences
are simple.

436

Score Structure in Digraphs

Theorem 14.35 (S. Pirzada) Let A be an irreducible score sequence. Then A is simple
if and only if it is one of [0], or [1, 1].

Proof Suppose A is an irreducible score sequence and let D be an oriented graph having
score sequence A. We have three cases to consider. (1) D has n 3 vertices, (2) D has two
vertices, (3) D has one vertex.
Case (1) D has n 3 vertices. Since A is irreducible, there exist vertices u, v and w such
that D has a cyclic triple u(1 0)v(1 0)w(1 0)u; or an intransitive triple u(1 0)v(1
0)w(0 0)u; or a transitive triple u(0 0)v(0 0)w(0 1)u; or a transitive triple u(0 0)v(0
0)w(0 0).
Now, if D contains the cyclic triple u(1 0)v(1 0)w(1 0)u, it can be changed to the
transitive triple u(0 0)v(0 0)w(0 0)u to form an oriented graph with the same score
sequence, or vice versa. So the number of arcs in D and D0 is different. If D contains
the intransitive triple u(1 0)v(1 0)w(0 0)u, we can transform it to the transitive triple
u(0 0)v(0 0)w(0 0)u, to form an oriented graph having the same score sequence, or
vice versa. Here also the number of arcs in D and D0 is different. Since in every case the
number of arcs in D and is not same, D is D0 not isomorphic to D0 . Thus A is not simple.
Case (2) D has two vertices. Then A = [1, 1] is the only irreducible score sequence and it
belongs to exactly one oriented graph, namely u(0 0)v.
D has just one vertex. Then A = [0] which is obviously simple.
Hence [0] and [1, 1] are the only irreducible score sequences that are simple.

Case (3)

Corollary 14.5 The score sequence A is simple if and only if every irreducible component of A is one of [0], or [1, 1].

14.12 Score Sequences in Oriented Bipartite Graphs

An oriented bipartite graph is the result of assigning a direction to each edge of a simple
bipartite graph. Let X = {x1 , x2 , . . ., xm } and Y = {y1 , y2 , . . ., yn } be the partite sets of an
oriented bipartite graph D. For any vertex u in D, let d + (u) and d (u) be the outdegree and
indegree respectively. Define ax = n +d + (x)d (x) and by = m+d + (y)d (y) as the scores
of x in X and y in Y respectively. Clearly, 0 ax 2n and 0 by 2m. Then the sequences
n
A = [ai ]m
1 and B = [b j ]1 in non-decreasing order are called a pair of score sequences of D.
An arc from x to y, that is, x y is denoted by x(1 0), and x(0 0) means neither x y
nor y x.
Definition: A tetra in an oriented bipartite graph is an induced subdigraph with two
vertices from each partite set. Define tetras of the form x(1 0)y(1 0)x0 (1 0)y0 (1 0)x
and x(1 0)y(1 0)x0 (1 0)y0 (0 0)x to be of -type, and all other tetras to be of -type.
An oriented bipartite graph is said to be of -type or -type according as all of its tetras
are of -type or -type respectively (Fig. 14.12).

Graph Theory

437

Fig. 14.12

We have the following simple observation.

Theorem 14.36 Among all oriented bipartite graphs with a given pair of score sequences, those with the fewest arcs are of -type.

Proof Let (A, B) be a given pair of score lists and let D be a realisation of (A, B) that is
not -type. Then D has a tetra of -type: x(1 0)y(1 0)x0 (1 0)y0 (1 0)x or x(1 0)y(1
0)x0 (1 0)y0 (0 0)x. Since x(1 0)y(1 0)x0 (1 0)y0 (1 0)x can be changed to x(0 0)y(0
0)x0 (0 0)y0 (0 0)x with the same score sequences and four arcs fewer , and x(1 0)y(1
0)x0 (1 0)y0 (0 0)x can be changed to x(0 0)y(0 0)x0 (0 0)y0 (0 1)x with the same score
sequences and two arcs fewer, so in either case we can obtain a realisation of (A, B) with
fewer arcs.
q
A transmitter is a vertex with indegree zero. In a -type oriented bipartite graph with
score sequences A = [a1 , a2 , . . ., am] and B = [b1 , b2, . . ., bn ] either the vertex with score am ,
or the vertex with score bn , or both may act as transmitter.
The next result due to Pirzada, Merajudin and Yin Jianhua [194] provides a useful recursive test to find whether a pair of lists is realisable.
Theorem 14.37 Suppose A = [a1 , a2 , . . ., am ] and B = [b1 , b2 , . . ., bn ] be two sequences
of non-negative integers in nondecreasing order. Let A0 be obtained from A by deleting
one entry am and B0 be obtained from B by reducing 2n am greatest entries of B by 1 each
provided am n and bn 2m 1. Then A and B are the score sequences of some oriented
bipartite graph if and only if A0 and B0 are also score sequences of some oriented bipartite
graph.
Theorem 14.37 provides an algorithm for determining whether a given pair of sequences
(A, B) of non-negative integers in nondecreasing order is a pair of score sequences and for
constructing a corresponding oriented bipartite graph. Suppose A = [a1 , a2 , . . ., am ] and
B = [b1, b2 , . . ., bn ] be a pair of score sequences of an oriented bipartite graph with parts
X = {x1 , x2 , . . ., xm } and Y = {y1 , y2 , . . ., yn }, where am n, bn 2m 1. Deleting am and
reducing 2n am greatest entries of B by 1 each to form B0 = [b01 , b02 , . . ., b0n ].. Then arcs
are defined by xm y j for which b j = b0j . Now, if at least one of the conditions am n or
bn 2m 1 does not hold, then we delete bn (obviously bn m, am 2n 1) and reduce
2m bn greatest entries of A by 1 each to form A = [a01, a02 , . . ., a0m ]. In this case arcs are

438

Score Structure in Digraphs

defined by yn xi for which ai = a0i . If this method is applied successively, then it tests
whether (A, B) is a pair of score sequences and an oriented bipartite graph (A, B) with
score sequences (A, B) is constructed.
We can interpret this algorithm in the following way. Let A = [a1 , a2 , . . ., am ] and B =
[b1, b2 , . . ., bn ] be a pair of score sequences of an oriented bipartite graph, where am
n, bn 2m 1. Let p and q denote the smallest and largest integers j for which b j = bam n .
Let A0 [a1, a2 , . . ., am1 ] as before and let then B0 = [b01, b02 , . . ., b0n ], then
b0j

b j,
b j 1,

f or 1 j p 1 and p + q (am n) j q,
otherwise.

We illustrate this reduction and the resulting construction with the following example,
beginning with lists A1 and B1 . The oriented bipartite graph constructed is shown in Figure
14.13.
A1
A2
A3
A4
A5

= [4, 4, 5]
= [4, 4]
= [4, 4]
= [4]
= [4]

B1
B2
B3
B4
B5

= [1, 1, 4, 5]
= [0, 1, 3, 4]
= [0, 1, 3]
= [0, 0, 2]
= [0, 0]

x3 y2
y4 x1 , x2
x2 y1
y3 x1
x1 y1 , y2

Fig. 14.13

Let Di be the oriented bipartite graphs with disjoint parts Xi and Yi for 1 i t . Let X =
t X and Y = U t Y . Clearly, D = [D , D , . . ., D ] denotes the oriented bipartite graph with
Ui=1
i
1
2
t
i=1 i
parts X and Y , obtained from Di for 1 i t such that the arcs of D are the arcs of Di and
each vertex of Y j is adjacent to every vertex of Xi for j < i and each vertex of Xi is adjacent
to every vertex of Y j for i < j .
The next result [194] gives a criterion for determining whether a pair of sequences are
realisable as scores.
Theorem 14.38 Let A = [a1 , a2 , . . ., am ] and B = [b1, b2 , . . ., bn ] be a pair of non-negative
integers in non-decreasing order. Then A and B are scores sequences of some oriented
bipartite graph if and only if
k

ai + b j 2kl
i=1

j=1

(14.3.1)

Graph Theory

439

for 1 k m and 1 l n, with equality when k = m and l = n.

The characterisation of scores of oriented tripartite graphs can be found in [190] and
scores (marks) in other types of digraphs can be found in [188, 196, 191].

14.13 Score Sequences of Semi Complete Digraphs

Definition: A semi complete digraph is a digraph with no directed loops and at least
one arc between every pair of distinct vertices. Clearly, a tournament is a semi complete
digraph in which there is exactly one arc between every pair of distinct vertices. Therefore
every semi complete digraph contains at least one tournament on the same vertex set and is
contained in the complete symmetric digraph on the same vertex set. The score of a vertex
v in a semi complete digraph D is the outdegree of v.
The following result is due to Reid and Zhang [222].
Theorem 14.39 A sequence of non-negative integers S = [si ]n1 in non-decreasing order is
a score sequence of some semi complete digraph of order n if and only if
k

i=1

 
k
and sk n 1,
2

(14.39.1)

for all k, 1 k n.

Proof
Necessity If S is a score sequence of some semi complete digraph D of order n, then
any k vertices of D induce a semi complete digraph of order k which, in turn, contains a
tournament W of order k. Therefore the sum of the scores in D of these k vertices is at least
the sum of their scores in W which is the total number of arcs in W, k2 . Also, a vertex
of D can dominate at most all of the other vertices, so no score in S can exceed n 1. Thus
the conditions (14.39.1) are necessary.
We require the following result for proving sufficiency.
Lemma 14.16 If S = [si ]n1 , n 1, is a sequence of integers in non-decreasing order satisfying (14.39.1), then there exists a tournament T with score sequence s0 = [s0i ]n1 , such that
s0i si for 1 i n.

Proof Define an order  on all non-decreasing sequences of integers satisfying (14.39.1)

(thus including all sequences satisfying conditions (14.1.2)) as follows. If B = [b1 , b2 , . . ., bn ]
and m is the smallest index for which bm = bn ( = max {bi : 1 i n}), then B (strictly) covers the sequence A = [a1, a2 , . . ., an ], where A and B are identical such that am = bm 1. Note
that if m > 1, then bm1 < bm = bm+1 = . . . = bn and if m = 1, then b1 = b2 = . . . = bn (n 1)/2.

440

Score Structure in Digraphs

n

Also, if B covers A, then ai =

i=1


n
bi 1 .

i=1

This implies, by Landaus theorem, that if S satisfies (14.39.1), then S is the score sequence for some tournament if and only if S covers no sequence satisfying (14.39.1). And,
if B is not the score sequence for any tournament, then B covers exactly one sequence satisfying (14.39.1). For two non-decreasing sequences of integers X and Y satisfying (14.39.1),
define X  Y if either X = Y, or there is a sequence X0 = X, X1 , X2 , . . ., X j1, X j = Y of nondecreasing sequences of integers each satisfying conditions (14.39.1) such that Xi covers
Xi1 , 1 i j.
Now, let S = [si ]n1 be a sequence of integers

in non-decreasing order satisfying conditions
n


(14.39.1). Induct on the integer e(S) si n2 . If e(S) = 0, then by Landaus theorem,
i=1

S itself is a score sequence for some tournament T . If e(S) > 0, then by the remarks above,
S covers exactly one sequence
Z = [z1 , z2 , . . ., zn] satisfying (14.39.1), such that zi si ,

n
n




for 1 i n, and e(Z) = Zi n2 = si n2 1 = e(s) 1. By the induction
i=1

i=1

hypothesis applied to Z , there is a score sequence S0 = [s0i ]n1 for some tournament T such that
s0i zi , for 1 i n. By the transitivity of , we have s0i si , for 1 i n, so and T suffice
for S, as required.
q

Sufficiency of Theorem 14.39 Let S = [si ]n1 , n 1, be a sequence of integers in nondecreasing order, satisfying conditions (14.39.1). By Lemma 14.16, there is a tournament
T of order n with score sequence S0 , where S0  S In T denote the vertex with score s0i by
vi , 1 i n. Since vi has indegree n 1 s0i n 1 si , arcs can be added from vi to any
n 1 si vertices in the inset of vi in T so as to produce a semi complete digraph D with
score sequence S.

14.15 Exercises
1. Prove that any n-tournament can be obtained from any other having the same scores
by a sequence of arc reversals of 3-cycles.
2. If an n-tournament has every score si satisfying, 41 (n 1) si 43 (n 1) then show that
it is irreducible.
3. Construct a proof for Theorem 14.2, and 14.6.
4. If S = {a, a + d, a + d + e}, where a, d, e are non-negative integers and de > 0, and if
(d, e) = g, d = a and e a + d d/2g + (1/2), then prove S is a score set of some
tournament.
5. Prove that every set of three non-negative integers is a score set of some tournament.
6. If a, b, c, d are four non-negative integers with bcd > 0, prove that there exists a tournament T with score set S = {a, a + b, a + b + c, a + b + c + d}.
7. Construct a proof of Theorem 14.11.

Graph Theory

441

8. Prove Lemma 14.7 and Lemma 14.8.

9. Construct a proof of Theorem 14.14.
10. Construct a proof of Theorem 14.16.
11. Construct proofs of Theorem 14.18 and 14.21.
12. If T is a bipartite tournament with score sequences and satisfying A = [ai ]m1 and B =
[b j ]n1 satisfying n/4 < ai < 3n/4 for 1 i m, and m/4 < b j < 3m/4 for 1 j n, then
prove that T is irreducible.
13. Construct proofs of Theorem 14.37 and Theorem 14.38.

References

1. Akiyama, J. and Kano, M., Factors and factorization of graphs, J. Graph Theory 9
(1985) 142.
2. Alavi, A. and Behzad, M., Complementary graphs and edge-chromatic numbers,
SIAM J. Apl. Math. 20 (1971) 161163.
3. Alspach, B. and Reid, K. B., Degree frequencies in digraphs and tournaments, J.
Graph Theory 2 (1978) 241249.
4. Anderson, I., Perfect matching of a graph, J. Combin. Theory Ser. B 10 (1971)
183186.
5. Appel, K. and Haken, W., Every planar map is four colorable, Bull. Amer. Math.
Soc. 82 (1976) 711712.
6. Appel, K. and Haken, W., Every planar map is four colorable, Part I, Discharging,
Illinois J. Math. 21 (1977) 429490.
7. Appel, K. and Haken, W., Every planar map is four colorable, Part II, Reducibility,
Illinois J. Math. 21 (1977) 491567.
8. Avery, P., Condition for a tournament score sequence to be simple, J. Graph Theory
4 (1980) 157164.
9. Avery, P., Score sequences in oriented graphs, J. Graph Theory 15, 3 (1991) 251257.
10. Babler, F., Uber die Zerlegung regularer streckenkomplexe ungerader Ordnung, Comment. Math. Helve. 10 (1938) 275287.
11. Bagga, K. S. and Beineke, L.W., Uniquely realizable score lists in bipartite tournaments, Czech. Math. Journal 37, 112 (1987) 323333.
12. Balakrishnan, R. and Paulraja, P., Chordal graphs and some of their derived graphs,
Congressus Numerantium 53 (1986) 3335.
13. Balakrishnan, R. and Ranganathan, K., A Textbook of Graph Theory, Springer-Verlag,
New York (2000).

Graph Theory

443

14. Barnette, D. W. and Grunbaum, B., On Steinitzs theorem concerning convex 3polytopes and on some properties of planar graphs, Lecture Notes in Math. 110
(1969) 2740.
15. Behzad, M. and Chartrand, G., Introduction to the Theory of Graphs, Allyn and
Bacon, Inc., Boston (1971).
16. Behzad, M. and Simpson, J. E., Eccentric sequences and eccentric sets in graphs,
Discrete Math. 16 (1976) 187193.
17. Behzad, M., Chartrand, M. G. and Lesniak, L., Foster, Graphs and Digraphs, Pindle,
Weber and Schmidt Int. Series, Boston Mass. (1979).
18. Beineke, L. W., Derived graphs and digraphs, in H. Sachs, H. Voss and H. Walther
(eds), Beitrage zur Graphentheorie Teubner, Leipzig (1968) 1733.
19. Beineke, L. W., A tour through tournaments, or bipartite and ordinary tournaments,
A comparative survey, Combinatorics, London Math. Soc. Lecture Note Series 52,
Cambridge University Press (1981) 4155.
20. Beineke, L. W. and Moon, J. W., On bipartite tournaments and scores, in G. Chartrand
(ed.), The Theory of Applications of Graphs (1981) 5571.
21. Berge, C., Two theorems in graph theory, Proc. Nat. Acad. Sci. U. S. A. 43 (1957)
842844.
22. Berge, C., The Theory of Graphs and its Applications, Methuen, London (1962).
23. Berge, C., Graphs and Hypergraphs, North-Holland, Amsterdam (1973).
24. Bermond, J. C. and Thomassen, C., Cycles in digraphsa survey, J. Graph Theory
5, 1 (1981) 143.
25. Biggs, N. L., Algebraic Graph Theory, Cambridge University Press (1974).
26. Bjelica, M. and Lakic, S., Criteria for scores sequences, Novi Sad J. Math. 30, 2
(2000) 1114.
27. Bjelica, M. and Lakic, S., Criteria for sets of scores with prescribed positions, IMC
Filomat (2001) 2630.
28. Boesch, F. T., Suffel, C. and Tindell, R., The spanning subgraphs of Eulerian graphs,
J. Graph Theory 1 (1977) 7984.
29. Bollabas, B., External Graph Theory, Acad. Press, New York (1978).
30. Bollobas, B., Uniquely colorable graphs, J. Combin. Theory Ser. B 24 (1978).
31. Bollobas, B. and Catlin, P. A., Topological cliques of random graphs, J. Comb. Theory Ser. B 30 (1981) 224227.

444

References

32. Bondy, J. A., Bounds on the chromatic number of a graph, J. Combin. Theory 7
(1969) 9698.
33. Bondy, J. A., Properties of graphs with constraints on degrees, Studia Sci. Math.
Hungar. 4 (1969) 473475.
34. Bondy, J. A., Pancyclic graphs, J. Combin. Theory Ser. B11 (1971) 8084.
35. Bondy, J. A., Short proofs of classical theorems, J. Graph Theory, 44 (2003) 159165.
36. Bondy, J. A. and Chvatal, V., A method in graph theory, Discrete Math. 15 (1976)
111135.
37. Bondy, J. A. and Halberstam, F. Y., Parity theorems for paths and cycles in graphs, J.
Graph Theory 10 (1986) 107115.
38. Bondy, J. A. and Thomassen, C., A short proof of Meyniels theorem, Discrete Math.
19 (1977) 195197.
39. Brooks, R. L., On coloring the nodes of a network, Proc. Cambridge Phil. Soc. 37
(1941) 194197.
40. Brualdi, R. A. and Shen, J., Landaus inequalities for tournament scores and a short
proof of a theorem on transitive sub-tournaments, J. Graph Theory 38 (2001) 244254.
41. de Bruijn, N.G., A combinatorial problem, Proc. Kon. Ned. Akad. V. Wetensch. 49
(1946) 758764.
42. Buckley, F. and Harary, F., Distance in Graphs, Addison-Wesley (1990).
43. Buckley, F., Miller, Z. and Slater, P. J., On graphs containing a given graph as center,
J. Graph Theory 5 (1981) 427434.
44. Camion, P., Chemins et circuits Hamiltoniens des graphs complets, C. R. Acad. Sci.
Paris 249 (1959) 21512152.
45. Cartwright, D. and Harary, F., On colorings of signed graphs, Elem. Math. 23 (1968)
8589.
46. Cayley, A., A theorem on trees, Quart. J. Math. 23 (1989) 376378; collected papers,
Cambridge 13 (1877) 2628.
47. Chang, L. C., The uniqueness and non uniqueness of the triangular association scheme,
Sci. Record 3 (1959) 604613.
48. Chartrand, G., On Hamiltonian line graphs, Trans. Amer. Math. Soc. 134 (1968)
559566.
49. Chartrand, G. and Geller, D., Uniquely colorable graphs, J. Combin. Theory 6 (1969)
271278.

Graph Theory

445

50. Chartrand, G., Gavlas, H., Harary F. and Schultz, M., On signed degrees in signed
graphs, Czech. Math. Journal 44 (1994) 111117.
51. Chartrand, G. and Kronck, H. V., Randomly traceable graphs, SIAM J. Appl. Math.
16 (1968) 696700.
52. Chartrand, G. and Stewart, M. J., Isometric graphs, in Capobianco et. al., Recent
Trends in Graph Theory, Springer-Verlag, Berlin, Lecture Notes in Math. 186 (1971)
6367.
53. Chartrand, G. and Stewart, M. J., The connectivity of line graphs, Math. Ann. 182
(1969) 170174.
54. Chartrand, G. and Schuster, S., Which graphs have unique distance trees, Amer.
Math. Monthly 81 (1974) 5356.
55. Chartrand, G. and Wall, C. E., On the Hamiltonian index of a graph, Studia Sci.
Math. Hungar. 8 (1973) 4348.
56. Chartrand, G. and White, A. T., Randomly traversable graphs, Elem. Math. 25 (1970)
101107.
57. Chen, Y. L., A short proof of Kundus f-factor theorem, Discrete Math. 71 (1988)
177179.
58. Choudam, S. A., A simple proof of the Erdos-Gallai theorem on graphic sequences,
Bull. Austral. Math. Soc. 36 (1986) 6770.
59. Chvatal, V., On Hamiltons ideals, J. Combin. Theory Ser. B 12 (1972) 163168.
60. Clark, J. and Holton, D. A., A First Look at Graph Theory, World Scientific (1991).
61. Courant, R. and Robbins, H. E., What is Mathematics? Oxford Univ. Press, London
(1941).
62. Cunningham, W. H., Matchings, matroids and extensions, Math. Programming B91
(2002) 515542.
63. Deo, N., Graph Theory with Applications to Engineering and Computer Science,
PHI, New Delhi (1974).
64. Dharwadkar, A., A new proof of the four color theorem, http: //www.geocities.com/
(2002).
65. Dijkstra, E. W., A note on two problems in connexion with graphs, Numer. Math. 1
(1959) 26971.
66. Dirac, G. A., A property of 4-chromatic graphs with some remarks on critical graphs,
J. London Math. Soc 27 (1952) 8592.
67. Dirac, G. A., Short proof of Mengers graph theorem, Mathematika 13 (1966) 4244.

446

References

68. Dirac, G.A., Some theorems on abstract graphs, Proc. London Math. Soc. 2 (1976)
111135.
69. Dirac, G. A. and Schuster, S., A theorem of Kuratowski, Nederl. Akad. Wetensch.
Proc. Ser. A 57 (1954) 343348.
70. Edmonds, J. and Johnson, E. L., Matching, Euler tours and the Chinese postman,
Math. Programming 5 (1973) 88124.
71. Eplett, W. J. R., Self-converse tournaments, Canad. Math. Bull. 22 (1979) 2327.
72. Erdos, P. and Fajtlowicz, S., On the conjecture of Hajos, Combinatorica 1 (1981)
141143.
73. Erdos, P. and Gallai, T. G., Graphs with prescribed degrees of vertices, Mat. Lapok
11 (1960) 264274.
74. Euler, L., Solutio problematics ad geometriam situs pertinents, Comment. Academiae
Sci. I Petropolitanae 8 (1736) 128140.
75. Euler, L., Demonstratio Nonnullarum Insignium Proprietatum QuibusSolida Hedris
Planis Inclusa Sunt Praedita, Novi Comm. Acad. Sci. Imp. Petropol 4 (1758)
140160.
76. Euler, L., The Konigsberg bridges, Sci. Amer. 189 (1853) 6670.
77. Fary, I., On straight line representation of planar graphs, Acta Sci. Math. 11 (1948)
229233.
78. Finck, H. J. and Sachs, H., Uber eine von H. S. Wilf angegebene Schranke fur die
chromatische Zahl endlicher Graphe, Math. Nachr. 39 (1969) 373386.
79. Fleischner, H., Elementary proofs of (relatively) recent characterizations of Eulerian
graphs, Discrete Applied Math. 24 (1989) 115119.
80. Fleischner, H., Eulerian graphs and related topics, Annals of Discrete Math. 45,
North-Holland, New York (1990).
81. Ford, L. R. and Fulkerson, D. R., Flows in Networks, Princeton University Press,
Princeton (1962).
82. Foulds, L. R., Graph Theory Applications, Springer-Verlag, New York (1992).
83. Fulkerson, D. R., Hofman, A. J. and Mc Andrew, M. H., Some properties of graphs
with multiple edges, Can. J. Math. 17 (1965) 166177.
84. Gale, D., A theorem on flows in networks, Pacific J. Math. 7 (1957) 10731082.
85. Gervacio, S. V., Score sequences: lexicographic enumeration and tournament construction, Discrete Math. 72 (1988) 151155.

Graph Theory

447

86. Grinberg, E. Ja., Plane homogenous graphs of degree three without Hamiltonian circuits (Russian), Latvian Math. Yearbook 4 (1968) 5158.
87. Ghouila-Houri, A., Une condition sufficiante d, existence d un circuit Hamiltonian,
C. R. Acad. Sci. Paris 25 (1960) 495497.
88. Graver, J. E. and Jurkat, W. R., f -factors and related decompositions of graphs, J.
Combin. Theory Ser. B 28 (1980) 6684.
89. Griggs, J. R. and Reid, K.B., Landaus Theorem revisited, Australasian J. Combinatorics 20 (1999) 1924.
90. Grotzsch, H., Ein Oreifarbensatz fur dreikreisfreie Netze auf der Kugel, Wiss. Z.
Martin-Luther Univ. Halle-Wittenberg Math. Naturwiss. Reihe 8 (1958) 109120.
91. Grunbaum, B., Grotzschs Theorem on 3-colorings, Michigan Math. J. 10 (1963)
303310.
92. Grunbaum, B., Graphs and Complex, Report of the University of Washington, Seattle, Math., 572B, 1969.
93. Grunbaum, B., Problem 2 in combinatorial structures and their applications, Proc.
Calgory International Conference, (1969), Gordon and Breach, New York.
94. Gupta, R. P., Bounds on the chromatic and achromatic numbers of complementary
graphs, in W. T. Tutte (ed.), Recent Progress in Combinatorics, Acad. Press, New
York 1969 (229235).
95. Hadwiger, H., Uber eine Klassification der Streckenkomplexe, Vierteljschr Naturforsch Ges Zurich 88 (1943) 133142.
96. Hager, M., On score sets for tournaments, Discrete Math. 58 (1986) 2534.
97. Hajos, G., Uber eine Konstruktion nich n-farbarrer Graphen, Wiss. Z. Martin-Cuther
Univ. Halle-wittenberg Math. Nat. Reihe 10 (1961) 116117.
98. Hakimi, S. L., On realizability of a set of integers as degree of the vertices of a linear
graph II. Uniqueness, J. Soc. Indust. Apple. Math.11 (1963) 135147.
99. Hakimi, S., On the realizability of a set of integers as the degrees of the vertices of a
graph, SIAM J.Appl.Math.10 (1962) 496506.
100. Hall, M., Distinct representatives of subsets, Bull. Amer. Math. Soc. 3 (1952)
584587.
101. Hall, P., On representation of subsets, J. London Math. Soc. 10 (1935) 2030.
102. Hamilton, W.R., Memorandum respecting a new system of roots of unity, Philosophical Magazine, 12, 4 (1856), p.446. http://www.maths.tcd.ie/pub/Hist Math/People/
Hamilton/Icosian/

448

References

103. Harary, F., A characterization of block graphs, Canad. Math. Bull.6 (1963) 16.
104. Harary, F., Graph theory, Addison-Wesley, Reading, MA. (1969).
105. Harary, F. and Frisch, I. T., Communication, Transmission and Transportation Networks, Addison-Wesley, Reading MA. (1971).
106. Harary, F. and Hedetniemi, S., The achromatic number of a graph, J. Combin. Theory
B (1970) 154161.
107. Harary, F., Hedetniemi, S. and Robinson, R. W., Uniquely colorable graphs, J. Combin. Theory 6 (1969) 264270.
108. Harary, F. and Nash-Williams, C. St. J. A., On Eulerian and Hamiltonian graphs and
line graphs, Canad. Math. Bull. 8 (1965) 701710.
109. Harary, F. and Norman, R. Z., The dissimilarity characteristic of Husimi trees, Ann.
of Math. 58 (1953) 131141.
110. Harary, F., Norman, R. Z. and Cartwright, D., Structural Models: An Introduction to
the Theory of Directed Graphs, Wiley, New York (1965).
111. Hasselbarth, W., Die verzweightheit von graphen, Match 16 (1984) 317.
112. Havel, M., A remark on the existence of finite graphs, Casopis Pest. Mat. 80 (1955)
477480.
113. Heawood, P. J., Map color theorem, Qurat. J. Pure Appl. Math. 24 (1890) 332338.
114. Hedetniemi, S., On hereditary properties of graphs, Studia Sci. Math. Hungarica.
115. Heirholzer, C., Ueber die Moglickkeit, Einen Linienzug ohne Wiederholung and
ohne Unterrechung zu umfahren, Math. Ann. 6 (1893) 3032.
116. Hoffman, A. J., On the exceptional case in a characterization of the arcs of complete
graphs, IBM J. Res. Develop. 4 (1960) 487496.
117. Hoffman, A. J., On the uniqueness of the triangular association scheme, Ann. Math.
Statist. 31 (1960) 492497.
118. Hoffman, A. J., On the line graph of the complete bipartite graph, Ann. Math. Statist.
35 (1964) 883885.
119. Ivanyi, A., Reconstruction of complete interval tournaments, Acta Univ. Sapientiae,
Informatica, 1, 1 (2009) 7188.
120. Ivanyi, A., Reconstruction of complete interval tournaments II, Acta Univ. Sapientiae, Mathematica, 2, 1 (2010) 4771.
121. Jaeger, F., A note on sub-Eulerian graphs, J. Graph Theory 3 (1979) 9193.
122. Jordan, C., Sur les assemblages de lignes, J. Reine Angew Math. 70 (1969) 185190.

Graph Theory

449

123. Jung, H. A., Zu einem isomorphiesatz von Whitney fur Graphen, Math. Ann. 164
(1966) 270271.
124. Kano, M. and Saito, A., [a, b]-factors of graphs, Discrete math. 47 (1983) 113116.
125. Kano, M. and Saito, A., [a, b]-factorization of a graph, J. Graph Theory 9 (1985)
129146.
126. Kapoor, S. F., Polimeni, A. O. and Wall, C. E., Degree sets for graphs, Fund. Math.
95 (1977) 189194.
127. Kasteleyn, P. W., Graph theory and crystal physics, in F. Harary (ed.) Chapter 2 in
Graph Theory and Theoretical Physics, Acad. Press (1967) 44110.
128. Kempe, A. B., On the geographical problem of the four colors, Amer. J. Math. 2
(1879) 193200.
129. Kleitman, D. J. and Wang, D. L., Algorithms for constructing graphs and digraphs
with given valencies and factors, Discrete Math. 6 (1973) 7888.
130. Koh, K. M. and Tan, B. P., The set and number of kings in a multipartite tournament,
Bull. Inst. Combin. Appl. 13 (1995) 1522.
131. Koh, K. M. and Tan, B. P., Kings in multipartite tournaments, Discrete Mathematics
147 (1995) 171183.
132. Koh, K. M. and Tan, B. P., Number of 4-kings in bipartite tournaments with no 3kings, Discrete Math. 154 (1996) 281287.
133. Koh, K. M. and Tan, B.P., The number of kings in multipartite tournaments, Discrete
Math., 167 (1997) 411418.
134. Kotzig, A., Cycles in a complete graph oriented in equilibrium, Mat. Fyz. Casopis
SAV 16 (1966) 175182.
135. Kotzig, A., The decomposition of a directed graph into quadratic factors consisting
of cycles, Acta F. R. N. Univ. Comen. Math. 22 (1969) 2729.
136. Konig, D., Graphen and Matrizen, Math. Fiz. Lapok 38 (1931) 116119.
137. Konig, D., Theorie der endlichen and unendlichen Graphen, Leipzig.
138. Konig, D., Uber Graphen und ihre Anwendung auf determinantentheorie und Mengenlehre, Math. Ann. 77 (1916) 453465.
139. Kouider, M. and Zbigniew Long, Stability number and [a, b]-factors in Graphs, J.
Graph Theory 46 (2004) 254264.
140. Krausz, J., Demonstration nouvelle d un theoreme de Whitney sur les reseaux, Mat.
Fiz. Lapok 50 (1943) 7589.

450

References

141. Kundu, S., Generalizations of the k-factor theorem, Discrete Math. 9 (1974) 173179.
142. Kundu, S., The Chartrand-Schuster Conjecture: Graphs with unique distance trees
are regular, J. Combin, Theory Ser. B 22 (1977) 233245.
143. Kundu, S., The k-factor conjecture is true, Discrete Math. 6 (1973) 367376.
144. Kuratowski, C., Sur le probleme des courbes gauches en topologie, Fund. Math. 15
(1930) 271283.
145. Landau, H. G., On dominance relations and the structure of animal socities: III. The
condition for a score structure, Bull. Math. Biophys. 15 (1953) 143148.
146. Lesniak, L., Eccentric sequences in graphs, Period Math. Hungar. 6 (1975) 287293.
147. Lesniak-Foster and Williams, J. E., On spanning and dominating circuits in graphs,
Canad. Math. Bull.20 (1977).
148. Lovasz, L., Subgraphs with prescribed valencies, J. Combin. Theory 8 (1970) 391416.
149. Lovasz, L., Valencies of graphs with 1-factors, Period. Math. Hungar. 5 (1974)
149151.
150. Lovasz, L., A remark on Mengers theorem, Acta Math. Acad. Sci. Hungar. 6 (1975)
287293.
151. Lovasz, L., Three short proofs in graph theroy, J. Combin Theory Ser. B 19 (1975)
111113.
152. Lovasz, L., Combinatorial Problems and Exercises, North-Holland, Amsterdam (1979).
153. Lucas, E., Rcrations Mathmatiques, Gauthier-Villares Paris (1891).
154. Mathematisch Centrum, This Weeks Citation Classic, CC/Number 7, February 14,
1983, http://www.garfield.library.upenn.edu/.
155. Maurer, S. B., The king chicken theorems, Math. Mag. 53 (1980) 6780.
156. McCuaig, W., A simple proof of Mengers theorem, J. Graph Theory 8 (1984) 427429.
157. Mckee, T. A., Recharacterizing Eulerian: Intimations of new duality, Discrete Math.
51 (1984) 237242.
158. Menger, K., Zur allgemeinen Kurventheoril, Fund. Math. 10 (1927) 96115.
159. Menon, V., On repeated interchange graphs, Amer. Math. Monthly 13 (1966) 986989.
160. Meyniel, M., Une condition sufficiante dexistance dun circuit Hamiltonian dans un
graphe oriente, J. Combin. Theory Ser. B 14 (1973) 137147.
161. Mirsky, L., Transversal Theroy, Acad. Press, New York (1971).

Graph Theory

451

162. Moon, J. W., On some combinatorial and probabilistic aspects of bipartite graphs,
Ph.D Thesis, University of Alberta, Edmonton (1962).
163. Moon, J. W., On the line graph of the complete bigraph, Ann. Math. Statist. 34
(1963) 664667.
164. Moon, J. W., Various proofs of Cayleys formula for counting trees, in F. Harary and
L. Beineke (eds), A seminar on Graph Theory, Holt, Rinehart and Winston (1967)
7078.
165. Moon, J. W., Topics on Tournaments, Holt, Rinehart and Winsten, New York (1968).
166. Muhammad A., Khan, K. K. Kayibi and S. Pirzada, Some results on tournaments,
Annales Computatorica, To appear.
167. Nash-Williams, C. St. J. A., Edge-disjoint spanning trees of finite graphs, J. London
Math. Soc. 36 (1961) 445450.
168. Nash-Williams, C. St. J. A., Edge-disjoint Hamiltonian circuits in graphs with vertices of large valency, in Studies in Pure Mathematics, Acad. Press, London (1971).
169. Nash-Williams, C. St. J. A. and Tutte, W. T., More proofs of Mengers theorem, J.
Graph Theory I (1977) 1317.
170. Nebesky, L., On the line graph of the square and the square of the line graph of a
connected graph, Casopis. Pset. Mat. 98 (1973) 285287.
171. Newman, D. J., A Problem in Graph Theory, Amer. Math. Monthly, 65 (1958) 611.
172. Nordhaus, E. A. and Gaddam, J. W., On complementary graphs, Amer. Math. Monthly
63 (1956) 175177.
173. O Neil, P. V., A new proof of Mengers theorem, J. Graph Theory 2 (1978) 257259.
174. Ore, O., A problem regarding the tracing of graphs, Elem. Math. 6 (1951) 4953.
175. Ore, O., Graphs and matching theorems, Duke Math. J. 22 (1955) 625639.
176. Ore, O., Note on Hamilton circuits, Amer. Math. Monthly 67 (1960) 55.
177. Ore, O., The Four Color Problem, Acad. Press, New York (1967).
178. Ore, O., Theory of Graphs, Amer. Math. Soc. Colloq. Publ. 38, Providence (1962).
179. Ore, O. and Stemple, G. J., Numerical methods in the four color problem, in W. T.
Tutte (ed.) Recent Progress in Combinatorics, Acad. Press, New York (1969).
180. Parthasarathy, K. R., Basic Graph Theory, Tata McGraw-Hill, New Delhi (1994).
181. Peterson, J., Die Theories der regularen graphen, Acta Math. 15 (1891) 193220.

452

References

182. Petrovic, V. and Thomassen, C., Kings in k-partite tournaments, Discrete Mathematics 98 (1991) 237238.
183. Petrovic, V., Decomposition of some planar graphs into trees, Discrete Mathematics
150 (1996) 449451.
184. Petrovic, V., Kings in bipartite tournaments, Discrete Mathematics 173 (1997) 187196.
185. Pirzada, S., Self converse score lists in oriented graphs, J. China Univ. of Science
and Tech. 26, 3 (1996) 392394.
186. Pirzada, S., Simple score sequences in oriented graphs, Novi Sad J. Math. 33, 1
(2003) 2529.
187. Pirzada, S., Construction of oriented graphs, in Balakrishnan, Wilson and Sethuraman (eds) Proc. National Conf. Graph Theory and Applications, Narosa, (2004)
146149.
188. Pirzada, S., Mark sequences in multidigraphs, Discrete Mathematics and Applications 17, 1 (2007) 7176.
189. Pirzada, S., Imbalances in digraphs, Kragujevic J. Mathematics 31 (2008) 14314
190. Pirzada, S. and Merajuddin, Score lists in oriented tripartite graphs, Novi Sad J. Math
26, 2 (1996) 19.
191. Pirzada, S. and Samee, U., Seminare Loth. de Comb. 55 Art B.(2006).
192. Pirzada, S., Merajuddin and Samee, U., On oriented graph scores, Matematicki Vesnik 60 (2008) 187191.
193. Pirzada, S., Merajuddin and Samee, U., Inequalities in oriented graph scores, Bull.
Allahabad Mathematical Society 23, 2 (2008) 389395.
194. Pirzada, S., Merajuddin and Yin Jianhua, On the scores of oriented bipartite graphs,
J. Math. Study 33, 4 (2000) 245359.
195. Pirzada, S. and Naikoo, T. A., Degree frequencies in oriented graphs, Thai J. Maths.
2 (2004) 8596.
196. Pirzada, S. and Naikoo, T. A., Inequalities on marks in digraphs, J. Mathematical
Inequalities and Applications 9, 2 (2006) 189198.
197. Pirzada, S. and Naikoo, T. A., Score sets in tournaments, Vietnam J. Mathematics 34,
2 (2006) 157161.
198. Pirzada, S. and Naikoo, T. A., Score sets in k-partite tournaments, J. Applied Mathematics and Computing 22, 1-2 (2006) 237245.
199. Pirzada, S. and Naikoo, T. A., Score sets in oriented k-partite graphs, AKCE International J. Graphs and Combinatorics 3, 2 (2006).

Graph Theory

453

200. Pirzada, S. and Naikoo, T. A., Score sets for oriented graphs, Applicable Analysis
and Discrete Mathematics 2, No. 1 (2008) 107113.
201. Pirzada, S., Naikoo, T. A. and Chishti, T. A., Score sets in oriented bipartite graphs,
Novi Sad J. Mathematics 36, 1 (2006) 3545.
202. Pirzada, S., Naikoo, T. A. and Dar, F. A., Signed degree sets in signed graphs, Czech.
Mathematical Journal 57, 3 (2007) 843848.
203. Pirzada, S., Naikoo, T. A. and Shah, N. A., Score sequences in oriented graphs, J.
Applied Mathematics and Computing 22, 12 (2007) 257268.
204. Pirzada, S., Naikoo, T. A. and Dar, F. A., Degree sets in bipartite and 3-partite graphs,
Oriental Journal of Mathematical Sciences 1, 1 (2007) 4753.
205. Pirzada, S., Naikoo, T. A. and Dar, F. A., Signed degree sequences in signed bipartite
graphs, AKCE International J. Graphs and Combinatorics 4, 2 (2007) 112.
206. Pirzada, S. Naikoo, T. A. and Dar, F. A., A note on signed degree sets in signed
bipartite graphs, Applicable Analysis and Discrete Mathematics 2, 1 (2008) 114117.
207. Pirzada, S. and Shah, N. A., Kings and serfs in oriented graphs, Mathematica Slovaca
58, 3 (2008) 277288.
208. Pirzada, S. and Yin Jian Hua, Degree sequences in graphs, J. Math. Study 39, No. 1
(2006) 2531.
209. Plesnik, J., Connectivity of regular graphs and the existence of 1-factors, Math. Casopis 22 (1972) 310318.
210. Posa, L., A theorem concerning Hamiltonian lines, Magyar Tud. Akad. Math. Kutato
Int. Kozl 7 (1962) 225226.
211. Prim, R. C., Shortest connection networks and some generalizations, Bell System
Techn. J. 36 (1957) 13891401.
212. Prufer, H., Beweis eines Satzes uber Permutationen, Arch-Math Physics 27 (1918)
742744.
213. Pym, J. S., A proof of Mengers theorem, Monatsh. Math. 73 (1969) 8183.
214. Randic, M., Characterization of atoms, molecules and classes of molecules based on
path enumerations, Proc. Bremen Konferenz zur Chemie, Univ. Bremen (1978) Part
II, Match No. 7 (1979) 564.
215. Rao, A. R. and Rao, S. B., On factorable degree sequences, J. Combin, Ser. B 13
(1972) 185191.
216. Redei, L., Ein Kombinatorischer Satz, Acta. Litt. Szeged, 7 (1934) 3943.
217. Reid, K. B., Score sets for tournaments, Congressus Num. 21 (1978) 607618.

454

References

218. Reid, K. B. and Beineke, L.W., Tournaments, in L. W. Beineke and R. J. Wilson

(eds), Chapter 7 in Selected Topics in Graph Theory, Acad. Press, London (1979)
169204.
219. Reid, K. B., Tournaments with prescribed number of kings and serfs, Congressus
Numerantium 29 (1980) 809826.
220. Reid, K. B., Every vertex a king, Discrete Mathematics 38 (1982) 9398.
221. Reid, K. B., Tournaments: Scores, kings, generalizations and special topics-a survey,
Congressus Num. 115 (1996) 171211.
222. Reid, K. B. and Zhang, C. Q., Score sequences of semi complete digraphs, Bulletin
of the ICA 24 (1998) 2732.
223. Ringel, G. and Youngs, J. W. T., Solution of the Heawood map coloring problem,
Proc. Nat. Acad. Sci. USA 60 (1968) 438445.
224. Rizzi, R., A short proof of Konigs matching theorem, J. Graph Theory 33 (2000)
138139.
225. Robertson, N., Sanders, D. P., Seymour, P. D. and Thomas, R., The four color theorem, J. Combin. Theory Ser. B. 70 (1997) 244.
226. Rooij, V. A. and Wilf, H., The interchange graphs of a finite graph, Acta. Math.
Acad. Sci. Hungar. 16 (1965) 263269.
227. Ryser, H. J., Combinatorial properties of matrices of zeros and ones, Canad. J. Math.
9 (1957) 371377.
228. Ryser, H. J., Matrices of zeros and ones in combinatorial mathematics, in Recent
Advances in Matrix Theory, Univ. Wisconsin Press (1964) 103124.
229. Saaty, T. L., Thirteen colorful variations on Guthries four color conjecture, Amer.
Math. Monthly 79 (1972) 243.
230. Saaty, T. L. and Kainen, P. C., The Four Color Problem, McGraw Hill, New York
(1977).
231. Sabidussi, G., Graph derivatives, Math. 2. 76 (1961) 385401.
232. Schnyder, W., Planar graphs and poset dimension, Order 5 (1989) a. 323343.
233. Seshu, S. and Reed, M., Linear Graphs and Electrical Networks, F. Harary(ed.),
Addison-Wesley, Reading (1961) Acad Press, London (1967) 44110.
234. Slater, P. J., A counter-example to Randics conjecture on distance degree-sequences
for trees, J. Graph Theory 6 (1982) 8992.
235. Sierksma, G. and Hoogeveen, H., Seven criteria for integer sequences being graphic,
J. Graph Theory 15 (1991) 223231.

Graph Theory

455

236. Sumner, D. P., 1-factors and antifactor sets, J. London Math. Soc. 2 (13) (1976)
351359.
237. Szekeres, G. and Wilf, H. S., An inequality for the chromatic number of a graph, J.
Combin. Theory 4 (1968) 13.
238. Tait, P. G., Remarks on the coloring of maps, Proc. Roy. Soc. Edinburgh Sect. A 10
(1880) 501503, 729.
239. Tanenbaum, A. S., Computer Networks, Prentice Hall Inc., Upper Saddle River, NJ,
1988.
240. The Mathematics Center at Bell Labs, History of Mathematics at Bell Labs,
http://cm.bell-labs.com/cm/ms/center/history.html
241. Thomassen, C., Kuratowskis theorem, J. Graph Theory 5 (1981) 225241.
242. Thomassen, C., Landaus characterization of tournament score sequences, in The Theory and Applications of Graphs, John Wiley (1981) 589591.
243. Titov, V. K., A constructive description of some classes of graphs, Ph.D. Thesis,
Moscow (1975).
244. Toida, S., Properties of an Euler graph, J. Franklin Inst. 295 (1973) 343345.
245. Tripathi, A. and Vijay, S., A note on a theorem of Erdos and Galai, Discrete Mathematics 265 (13) (2003) 417420.
246. Tripathi, A., Venugopalan, S. and West, D. B., A short proof of the Erdos-Gallai
characterization of graphic lists, Discrete Math. 310 (2010) 843844.
247. Turan, P., Eine Extremalaufgabe aus der Grapheneorie, Mat. Fiz. Lapok 48 (1941)
436452 (1936), Reprinted Chelsea, New York (1950).
248. Tutte, W. T., A family of cubical graphs, Proc. Cambridge Phil. Soc. 21 (1946)
98101.
249. Tutte, W. T., The factorization of linear graphs, J. London Math. Soc. 22 (1947)
107111.
250. Tutte, W. T., The factors of graphs, Canad. J. Math. 4 (1952) 314328.
251. Tutte, W. T., A short proof of the factor theorem for finite graphs, Canad. J. Math. 6
(1954) 347352.
252. Tutte, W. T., On the problem of decomposing a graph into n connected factors, J.
London Math. Soc. 36 (1961) 221230.
253. Tutte, W. T., Graph factors, Combinatorica 1 (1981) 7997.

456

References

254. Veblen, O., An application of modular equations in analysis situs, Ann. of Math. 14
(191213) 8694.
255. Vergnas, M. L., An extension of Tuttes 1-factor theorem, Discrete Math. 23 (1978)
241255.
256. Vergnas, M. L., M. Thesis, University of Paris, v1 (1972).
257. van Aardenne-Ehrenfest, T. and de Bruijn, N. G., Circuits and trees in oriented
graphs, Simon Stevin 28 (1951) 203217.
258. Vizing, V. G., On an estimate of the chromatic class of a p-graph (Russian) Diskret.
Analiz. 3 (1964) 2530.
259. Vizing, V. G., The chromatic class of a multigraph, Cybernetics 3 (1965) 3241.
260. Wagner, K., Bemerkungen zum Vierfarbenproblem, Jahresber Deutsch. Math. Verein
46 (1936) 2122.
261. Wang, D. L. and Kleitman, D. J., On the existence of n-connected graphs with prescribed degrees (n > 2), Networks 3 (1973) 225239.
262. Welsh, D. J. A. and Powell, M. B., An upper bound for the chromatic number of a
graph and its application to time-tabling problems, Comput. J. 10 (1967) 8586.
263. West, D. B., Introduction to Graph Theory, Prentice Hall Inc. (1996).
264. Whitney, H., A theorem on graphs, Annals Math. 32 (1931) 378390.
265. Whitney, H., Congruent graphs and the connectivity of graphs, Amer. J. Math. 54
(1932) 150168.
266. Whitney, H., 2-isomorphic graphs, Amer. J. Math. 55, (1933) 245254.
267. Whitney, H., Planar graphs, Fund. Math. 21 (1933) 7384.
268. Wiener, N., A simplification of logic of relations, Proc. Cambridge Phil. Soc. 17
(191214) 387390.
269. Wilson, R. J., Introduction to Graph Theory, Oliver and Boyd, Edindurgh (1972).
270. Woodall, D. R., Sufficient conditions for circuits in graphs, Proc. London Math.Soc.
24 (1972) 739755.
271. Yan, J., Lih, K., Kuo, D. and Chang, G. J., Signed degree sequences in signed graphs,
J. Graph Theory 26 (1997) 111117.
272. Zhou, G., Yao, T. and Zhang, K, On score sequences of k-hypertournaments, European J. Combin. 21 (8) (2000) 9931000.