Graphtheory

Lecture Notes on

GRAPH THEORY
Tero Harju
Department of Mathematics
University of Turku
FIN-20014 Turku, Finland
e-mail: harju@utu.ﬁ

1994 – 2011

Contents

1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Graphs and their plane ﬁgures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Subgraphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Paths and cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2

Connectivity of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1 Bipartite graphs and trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3

Tours and Matchings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1 Eulerian graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Hamiltonian graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Matchings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

29
29
31
35

4

Colourings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1 Edge colourings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Ramsey Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Vertex colourings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

43
43
47
53

5

Graphs on Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1 Planar graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Colouring planar graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Genus of a graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

61
61
68
76

6

Directed Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.1 Digraphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.2 Network Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

1
Introduction
Graph theory may be said to have its beginning in 1736 when E ULER considered the (general case of the) Königsberg bridge problem:
Does there exist a walk crossing each of the
seven bridges of Königsberg exactly once? (Solutio Problematis ad geometriam situs pertinentis, Commentarii Academiae Scientiarum Imperialis Petropolitanae 8 (1736), pp. 128-140.)
It took 200 years before the first book on graph theory was written. This was “Theorie der endlichen und unendlichen Graphen” ( Teubner, Leipzig, 1936) by K ÖNIG in
1936. Since then graph theory has developed into an extensive and popular branch of
mathematics, which has been applied to many problems in mathematics, computer
science, and other scientific and not-so-scientific areas. For the history of early graph
theory, see
N.L. B IGGS , R.J. L LOYD
Press, 1986.

AND

R.J. W ILSON, “Graph Theory 1736 – 1936”, Clarendon

There are no standard notations for graph theoretical objects. This is natural, because the names one uses for the objects reflect the applications. Thus, for instance, if
we consider a communications network (say, for email) as a graph, then the computers taking part in this network, are called nodes rather than vertices or points. On the
other hand, other names are used for molecular structures in chemistry, flow charts
in programming, human relations in social sciences, and so on.
These lectures study finite graphs and majority of the topics is included in
J.A. B ONDY, U.S.R. M URTY, “Graph Theory with Applications”, Macmillan, 1978.
R. D IESTEL, “Graph Theory”, Springer-Verlag, 1997.
F. H ARARY, “Graph Theory”, Addison-Wesley, 1969.
D.B. W EST, “Introduction to Graph Theory”, Prentice Hall, 1996.
R.J. W ILSON, “Introduction to Graph Theory”, Longman, (3rd ed.) 1985.
In these lectures we study combinatorial aspects of graphs. For more algebraic topics
and methods, see
N. B IGGS, “Algebraic Graph Theory”, Cambridge University Press, (2nd ed.) 1993.
C. G ODSIL , G.F. R OYLE, “Algebraic Graph Theory”, Springer, 2001.
and for computational aspects, see
S. E VEN, “Graph Algorithms”, Computer Science Press, 1979.

3
In these lecture notes we mention several open problems that have gained respect
among the researchers. Indeed, graph theory has the advantage that it contains easily
formulated open problems that can be stated early in the theory. Finding a solution
to any one of these problems is another matter.
Sections with a star (∗) in their heading are optional.

Notations and notions
• For a finite set X, |X | denotes its size (cardinality, the number of its elements).
• Let
[1, n] = {1, 2, . . . , n},
and in general,

[i, n] = {i, i + 1, . . . , n}
for integers i ≤ n.
• For a real number x, the floor and the ceiling of x are the integers

⌊ x⌋ = max{k ∈ Z | k ≤ x} and ⌈ x⌉ = min{k ∈ Z | x ≤ k}.
• A family {X1 , X2 , . . . , Xk } of subsets Xi ⊆ X of a set X is a partition of X, if
X=

Xi

and

Xi ∩ X j = ∅ for all different i and j .

i∈[1,k]

• For two sets X and Y,
X × Y = {( x, y) | x ∈ X, y ∈ Y }
is their Cartesian product, and
X △Y = ( X Y ) ∪ (Y X )
is their symmetric difference. Here X Y = { x | x ∈ X, x ∈ Y }.
/
• Two integers n, k ∈ N (often n = |X | and k = |Y | for sets X and Y) have the same
parity, if both are even, or both are odd, that is, if n ≡ k (mod 2). Otherwise, they
have opposite parity.
Graph theory has abundant examples of NP-complete problems. Intuitively, a
problem is in P 1 if there is an efficient (practical) algorithm to find a solution to it. On
the other hand, a problem is in NP 2 , if it is first efficient to guess a solution and then
efficient to check that this solution is correct. It is conjectured (and not known) that
P = NP. This is one of the great problems in modern mathematics and theoretical
computer science. If the guessing in NP-problems can be replaced by an efficient
systematic search for a solution, then P=NP. For any one NP-complete problem, if it
is in P, then necessarily P=NP.
1
2

Solvable – by an algorithm – in polynomially many steps on the size of the problem instances.
Solvable nondeterministically in polynomially many steps on the size of the problem instances.

1.1 Graphs and their plane figures

4

Let V be a finite set, and denote by
E(V ) = {{u, v} | u, v ∈ V, u = v} .
the 2-sets of V, i.e., subsets of two distinct elements.
D EFINITION . A pair G = (V, E) with E ⊆ E(V ) is called a graph (on V). The elements
of V are the vertices of G, and those of E the edges of G. The vertex set of a graph G
is denoted by VG and its edge set by EG . Therefore G = (VG , EG ).
In literature, graphs are also called simple graphs; vertices are called nodes or points;
edges are called lines or links. The list of alternatives is long (but still finite).
A pair {u, v} is usually written simply as uv. Notice that then uv = vu. In order to
simplify notations, we also write v ∈ G and e ∈ G instead of v ∈ VG and e ∈ EG .
D EFINITION . For a graph G, we denote
νG = |VG | and ε G = | EG | .
The number νG of the vertices is called the order of G, and ε G is the size of G. For an
edge e = uv ∈ G, the vertices u and v are its ends. Vertices u and v are adjacent or
neighbours, if uv ∈ G. Two edges e1 = uv and e2 = uw having a common end, are
adjacent with each other.
A graph G can be represented as a plane figure by
drawing a line (or a curve) between the points u and
v (representing vertices) if e = uv is an edge of G.
The figure on the right is a geometric representation
of the graph G with VG = {v1 , v2 , v3 , v4 , v5 , v6 } and
E G = { v1 v2 , v1 v3 , v2 v3 , v2 v4 , v5 v6 }.

v1

v3

v6

v2

v4

v5

Often we shall omit the identities (names v) of the vertices in our figures, in which
case the vertices are drawn as anonymous circles.
Graphs can be generalized by allowing loops vv and parallel (or multiple) edges
between vertices to obtain a multigraph G = (V, E, ψ), where E = {e1 , e2 , . . . , em } is
a set (of symbols), and ψ : E → E(V ) ∪ {vv | v ∈ V } is a function that attaches an
unordered pair of vertices to each e ∈ E: ψ(e) = uv.
Note that we can have ψ(e1 ) = ψ(e2 ). This is drawn in
the figure of G by placing two (parallel) edges that connect the common ends. On the right there is (a drawing of) a multigraph G with vertices V = {a, b, c}
and edges ψ(e1 ) = aa, ψ(e2 ) = ab, ψ(e3 ) = bc, and
ψ(e4 ) = bc.

b

a

c


5

Later we concentrate on (simple) graphs.
D EFINITION . We also study directed graphs or digraphs
D = (V, E), where the edges have a direction, that is, the
edges are ordered: E ⊆ V × V. In this case, uv = vu.
The directed graphs have representations, where the edges are drawn as arrows.
A digraph can contain edges uv and vu of opposite directions.
Graphs and digraphs can also be coloured, labelled, and weighted:
D EFINITION . A function α : VG → K is a vertex colouring of G by a set K of colours.
A function α : EG → K is an edge colouring of G. Usually, K = [1, k] for some k ≥ 1.
If K ⊆ R (often K ⊆ N), then α is a weight function or a distance function.

Isomorphism of graphs
D EFINITION . Two graphs G and H are isomorphic, denoted by G ∼ H, if there exists
=
a bijection α : VG → VH such that
uv ∈ EG ⇐⇒ α(u)α(v) ∈ E H
for all u, v ∈ G.
Hence G and H are isomorphic if the vertices of H are renamings of those of G.
Two isomorphic graphs enjoy the same graph theoretical properties, and they are often
identified. In particular, all isomorphic graphs have the same plane figures (excepting
the identities of the vertices). This shows in the figures, where we tend to replace the
vertices by small circles, and talk of ‘the graph’ although there are, in fact, infinitely
many such graphs.
v2
v3
2
4
Example 1.1. The following graphs are
isomorphic. Indeed, the required isov5
1
morphism is given by v1 → 1, v2 → 3,
v3 → 4, v4 → 2, v5 → 5.
v1
v4
3
5
Isomorphism Problem. Does there exist an efficient algorithm to check whether any two
given graphs are isomorphic or not?
n

The following table lists the number 2( 2 ) of all graphs on a given set of n vertices,
and the number of all nonisomorphic graphs on n vertices. It tells that at least for
computational purposes an efficient algorithm for checking whether two graphs are
isomorphic or not would be greatly appreciated.
n

1 2 3

4

5

6

7

8

graphs
1 2 8 64 1024 32 768 2 097 152 268 435 456
156
1044
12 346
nonisomorphic 1 2 4 11 34

9
236

> 6 · 1010
274 668


6

Other representations
Plane figures catch graphs for our eyes, but if a problem on graphs is to be programmed, then these figures are, to say the least, unsuitable. Integer matrices are ideal
for computers, since every respectable programming language has array structures
for these, and computers are good in crunching numbers.
Let VG = {v1 , . . . , vn } be ordered. The adjacency matrix of G is the n × n-matrix M with entries Mij = 1
or Mij = 0 according to whether vi v j ∈ G or vi v j ∈ G.
/
For instance, the graph in Example 1.1 has an adjacency matrix on the right. Notice that the adjacency
matrix is always symmetric (with respect to its diagonal consisting of zeros).



0
1

1

0
1

1
0
0
1
1

1
0
0
1
0

0
1
1
0
0


1
1

0

0
0

A graph has usually many different adjacency matrices, one for each ordering of
its set VG of vertices. The following result is obvious from the definitions.
Theorem 1.1. Two graphs G and H are isomorphic if and only if they have a common adjacency matrix. Moreover, two isomorphic graphs have exactly the same set of adjacency matrices.
Graphs can also be represented by sets. For this, let X = {X1 , X2 , . . . , Xn } be a
family of subsets of a set X, and define the intersection graph GX as the graph with
vertices X1 , . . . , Xn , and edges Xi X j for all i and j (i = j) with Xi ∩ X j = ∅.
Theorem 1.2. Every graph is an intersection graph of some family of subsets.
Proof. Let G be a graph, and define, for all v ∈ G, a set
Xv = {{v, u} | vu ∈ G }.
Then Xu ∩ Xv = ∅ if and only if uv ∈ G.

⊔
⊓

Let s( G ) be the smallest size of a base set X such that G can be represented as an
intersection graph of a family of subsets of X, that is,
s( G ) = min{| X | | G ∼ GX for some X ⊆ 2X } .
=
How small can s( G ) be compared to the order νG (or the size ε G ) of the graph? It was
shown by K OU , S TOCKMEYER AND W ONG (1976) that it is algorithmically difficult to
determine the number s( G ) – the problem is NP-complete.
Example 1.2. As yet another example, let A ⊆ N be a finite set of natural numbers,
and let G A = ( A, E) be the graph with rs ∈ E if and only if r and s (for r = s) have a
common divisor > 1. As an exercise, we state: All graphs can be represented in the form
G A for some set A of natural numbers.

1.2 Subgraphs

7

1.2 Subgraphs
Ideally, given a nice problem the local properties of a graph determine a solution.
In these situations we deal with (small) parts of the graph (subgraphs), and a solution can be found to the problem by combining the information determined by the
parts. For instance, as we shall later see, the existence of an Euler tour is very local, it
depends only on the number of the neighbours of the vertices.

Degrees of vertices
D EFINITION . Let v ∈ G be a vertex a graph G. The neighbourhood of v is the set
NG (v) = {u ∈ G | vu ∈ G } .
The degree of v is the number of its neighbours:
dG (v) = | NG (v)| .
If dG (v) = 0, then v is said to be isolated in G, and if dG (v) = 1, then v is a leaf of the
graph. The minimum degree and the maximum degree of G are defined as
δ( G ) = min{dG (v) | v ∈ G }

and

∆( G ) = max{dG (v) | v ∈ G } .

The following lemma, due to E ULER (1736), tells that if several people shake
hands, then the number of hands shaken is even.
Lemma 1.1 (Handshaking lemma). For each graph G,

∑ d G ( v) = 2 · ε G .
v∈ G

Moreover, the number of vertices of odd degree is even.
Proof. Every edge e ∈ EG has two ends. The second claim follows immediately from
the first one.
⊔
⊓
Lemma 1.1 holds equally well for multigraphs, when dG (v) is defined as the number of edges that have v as an end, and when each loop vv is counted twice.
Note that the degrees of a graph G do not determine G. Indeed, there are graphs
G = (V, EG ) and H = (V, E H ) on the same set of vertices that are not isomorphic, but
for which dG (v) = d H (v) for all v ∈ V.

1.2 Subgraphs

8

Subgraphs
D EFINITION . A graph H is a subgraph of a graph G, denoted by H ⊆ G, if VH ⊆ VG
and E H ⊆ EG . A subgraph H ⊆ G spans G (and H is a spanning subgraph of G), if
every vertex of G is in H, i.e., VH = VG .
Also, a subgraph H ⊆ G is an induced subgraph, if E H = EG ∩ E(VH ). In this
case, H is induced by its set VH of vertices.
In an induced subgraph H ⊆ G, the set E H of edges consists of all e ∈ EG such that
e ∈ E(VH ). To each nonempty subset A ⊆ VG , there corresponds a unique induced
subgraph
G [ A] = ( A, EG ∩ E( A)) .
To each subset F ⊆ EG of edges there corresponds a unique spanning subgraph of G,
G [ F ] = (VG , F ) .

G

subgraph

spanning

induced

For a set F ⊆ EG of edges, let
G − F = G [ EG F ]
be the subgraph of G obtained by removing (only) the edges e ∈ F from G. In particular, G −e is obtained from G by removing e ∈ G.
Similarly, we write G + F, if each e ∈ F (for F ⊆ E(VG )) is added to G.
For a subset A ⊆ VG of vertices, we let G − A ⊆ G be the subgraph induced by
VG A, that is,
G − A = G [VG A] ,
and, e.g., G −v is obtained from G by removing the vertex v together with the edges
that have v as their end.
Reconstruction Problem. The famous open problem, Kelly-Ulam problem or the Reconstruction Conjecture, states that a graph of order at least 3 is determined up to isomorphism by its vertex deleted subgraphs G −v (v ∈ G): if there exists a bijection α : VG → VH
such that G −v ∼ H −α(v) for all v, then G ∼ H.
=
=

1.2 Subgraphs

9

2-switches
D EFINITION . For a graph G, a 2-switch with respect
to the edges uv, xy ∈ G with ux, vy ∈ G replaces the
/
edges uv and xy by ux and vy. Denote

v

y

v

y

u

x

u

x

2s

G− H
→
if there exists a ﬁnite sequence of 2-switches that carries G to H.
2s
2s
Note that if G − H then also H − G since we can apply the sequence of 2→
→
switches in reverse order.
Before proving Berge’s switching theorem we need the following tool.
Lemma 1.2. Let G be a graph of order n with a degree sequence d1 ≥ d2 ≥ · · · ≥ dn , where
2s

dG (vi ) = di . Then there is a graph G ′ such that G − G ′ with NG′ (v1 ) = {v2 , . . . , vd1 +1 }.
→
Proof. Let d = ∆( G ) (= d1 ). Suppose that there is a vertex vi with 2 ≤ i ≤ d + 1 such
that v1 vi ∈ G. Since dG (v1 ) = d, there exists a v j with
/
j ≥ d + 2 such that v1 v j ∈ G. Here di ≥ d j , since j > i.
vj
v1
vi
Since v1 v j ∈ G, there exists a vt (2 ≤ t ≤ n) such that
vi vt ∈ G, but v j vt ∈ G. We can now perform a 2-switch
/
with respect to the vertices v1 , v j , vi , vt . This gives a new
vt
graph H, where v1 vi ∈ H and v1 v j ∈ H, and the other
/
neighbours of v1 remain to be its neighbours.
When we repeat this process for all indices i with v1 vi ∈ G for 2 ≤ i ≤ d + 1, we
/
′ as required.
obtain a graph G
⊔
⊓
Theorem 1.3 (B ERGE (1973)). Two graphs G and H on a common vertex set V satisfy
dG (v) = d H (v) for all v ∈ V if and only if H can be obtained from G by a sequence of
2-switches.
2s

Proof. If G − H, then clearly H has the same degrees as G.
→
In converse, we use induction on the order νG . Let G and H have the same degrees.
2s

By Lemma 1.2, we have a vertex v and graphs G ′ and H ′ such that G − G ′ and
→
2s

H − H ′ with NG′ (v) = NH ′ (v). Now the graphs G ′ −v and H ′ −v have the same
→
2s

2s

degrees. By the induction hypothesis, G ′ −v − H ′ −v, and thus also G ′ − H ′ .
→
→
2s

Finally, we observe that H ′ − H by the ‘reverse 2-switches’, and this proves the
→
claim.
⊔
⊓
D EFINITION . Let d1 , d2 , . . . , dn be a descending sequence of nonnegative integers, that
is, d1 ≥ d2 ≥ · · · ≥ dn . Such a sequence is said to be graphical, if there exists a graph
G = (V, E) with V = {v1 , v2 , . . . , vn } such that di = dG (vi ) for all i.

1.2 Subgraphs

10

Using the next result recursively one can decide whether a sequence of integers is
graphical or not.
Theorem 1.4 (H AVEL (1955), H AKIMI (1962)). A sequence d1 , d2 , . . . , dn (with d1 ≥ 1 and
n ≥ 2) is graphical if and only if
d2 − 1, d3 − 1, . . . , dd1 +1 − 1, dd1 +2 , dd1 +3 , . . . , dn

(1.1)

is graphical (when put into nonincreasing order).
Proof. (⇐) Consider G of order n − 1 with vertices (and degrees)
dG (v2 ) = d2 − 1, . . . , dG (vd1 +1 ) = dd1 +1 − 1,
d G ( v d1 + 2 ) = d d1 + 2 , . . . , d G ( v n ) = d n
as in (1.1). Add a new vertex v1 and the edges v1 vi for all i ∈ [2, dd1 +1 ]. Then in the
new graph H, d H (v1 ) = d1 , and d H (vi ) = di for all i.
(⇒) Assume dG (vi ) = di . By Lemma 1.2 and Theorem 1.3, we can suppose that
NG (v1 ) = {v2 , . . . , vd1 +1 }. But now the degree sequence of G −v1 is in (1.1).
⊔
⊓
Example 1.3. Consider the sequence s = 4, 4, 4, 3, 2, 1. By Theorem 1.4,
s is graphical ⇐⇒ 3, 3, 2, 1, 1 is graphical
2, 1, 1, 0 is graphical

v2

0, 0, 0 is graphical.
The last sequence corresponds to a graph with no
edges, and hence also our original sequence s is graphical. Indeed, the graph G on the right has this degree
sequence.

v4
v6

v1

v3

v5

Special graphs
D EFINITION . A graph G = (V, E) is trivial, if it has only one vertex, i.e., νG = 1;
otherwise G is nontrivial.
The graph G = KV is the complete graph on V, if every
two vertices are adjacent: E = E(V ). All complete graphs
of order n are isomorphic with each other, and they will be
denoted by Kn .
/
The complement of G is the graph G on VG , where EG = {e ∈ E(V ) | e ∈ EG }. The
complements G = KV of the complete graphs are called discrete graphs. In a discrete
graph EG = ∅. Clearly, all discrete graphs of order n are isomorphic with each other.
A graph G is said to be regular, if every vertex of G has the same degree. If this
degree is equal to r, then G is r-regular or regular of degree r.

1.3 Paths and cycles

11

A discrete graph is 0-regular, and a complete graph Kn is (n − 1)-regular. In particular, ε Kn = n(n − 1)/2, and therefore ε G ≤ n(n − 1)/2 for all graphs G that have
order n.
Many problems concerning (induced) subgraphs are algorithmically difficult. For
instance, to find a maximal complete subgraph (a subgraph Km of maximum order)
of a graph is unlikely to be even in NP.
Example 1.4. The graph on the right is the Petersen
graph that we will meet several times (drawn differently). It is a 3-regular graph of order 10.
Example 1.5. Let k ≥ 1 be an integer, and consider the set B k of all binary strings
of length k. For instance, B3 = {000, 001, 010, 100, 011, 101, 110, 111}. Let Qk be the
graph, called the k-cube, with VQk = B k , where uv ∈ Qk if and only if the strings u
and v differ in exactly one place.
110
111

2k ,

The order of Qk is νQk =
the number of binary
strings of length k. Also, Qk is k-regular, and so, by the
handshaking lemma, ε Qk = k · 2k−1 . On the right we
have the 3-cube, or simply the cube.

101

100

010
000

011
001

Example 1.6. Let n ≥ 4 be any even number. We show by induction that there exists
a 3-regular graph G with νG = n. Notice that all 3-regular graphs have even order by
the handshaking lemma.
If n = 4, then K4 is 3-regular. Let G be a 3-regular
graph of order 2m − 2, and suppose that uv, uw ∈ EG .
Let VH = VG ∪ { x, y}, and E H = ( EG {uv, uw}) ∪
{ux, xv, uy, yw, xy}. Then H is 3-regular of order 2m.

y

x

w

v
u

The most fundamental notions in graph theory are practically oriented. Indeed, many
graph theoretical questions ask for optimal solutions to problems such as: find a
shortest path (in a complex network) from a given point to another. This kind of
problems can be difficult, or at least nontrivial, because there are usually choices what
branch to choose when leaving an intermediate point.

Walks
D EFINITION . Let ei = ui ui+1 ∈ G be edges of G for i ∈ [1, k]. The sequence W =
e1 e2 . . . ek is a walk of length k from u1 to uk+1 . Here ei and ei+1 are compatible in the
sense that ei is adjacent to ei+1 for all i ∈ [1, k − 1].


12

We write, more informally,
W : u1 − u2 − . . . − u k − u k+1
→
→
→
→

or

k

W : u1 − u k+1 .
→

⋆
Write u − v to say that there is a walk of some length from u to v. Here we under→
⋆
stand that W : u − v is always a speciﬁc walk, W = e1 e2 . . . ek , although we sometimes
→
do not care to mention the edges ei on it. The length of a walk W is denoted by |W |.

D EFINITION . Let W = e1 e2 . . . ek (ei = ui ui+1 ) be a walk.
W is closed, if u1 = uk+1 .
W is a path, if ui = u j for all i = j.
W is a cycle, if it is closed, and ui = u j for i = j except that u1 = uk+1 .
W is a trivial path, if its length is 0. A trivial path has no edges.
For a walk W : u = u1 − . . . − uk+1 = v, also
→
→
W −1 : v = u k+1 − . . . − u1 = u
→
→
is a walk in G, called the inverse walk of W.
A vertex u is an end of a path P, if P starts or ends in u.
⋆
⋆
⋆
The join of two walks W1 : u − v and W2 : v − w is the walk W1 W2 : u − w.
→
→
→
(Here the end v must be common to the walks.)
Paths P and Q are disjoint, if they have no vertices in common, and they are
independent, if they can share only their ends.
Clearly, the inverse walk P−1 of a path P is a path (the inverse path of P). The join
of two paths need not be a path.
A (sub)graph, which is a path (cycle) of length
k − 1 (k, resp.) having k vertices is denoted by
Pk (Ck , resp.). If k is even (odd), we say that the
path or cycle is even (odd). Clearly, all paths of
length k are isomorphic. The same holds for cycles of ﬁxed length.

P5

C6

⋆
⋆
Lemma 1.3. Each walk W : u − v with u = v contains a path P : u − v, that is, there is a
→
→
⋆
path P : u − v that is obtained from W by removing edges and vertices.
→

Proof. Let W : u = u1 − . . . − uk+1 = v. Let i < j be indices such that ui = u j .
→
→
⋆
If no such i and j exist, then W, itself, is a path. Otherwise, in W = W1 W2 W3 : u −
→
⋆
⋆
⋆
⋆
ui − u j − v the portion U1 = W1 W3 : u − ui = u j − v is a shorter walk. By
→
→
→
→
⋆
repeating this argument, we obtain a sequence U1 , U2 , . . . , Um of walks u − v with
→
|W | > |U1 | > · · · > |Um |. When the procedure stops, we have a path as required.
(Notice that in the above it may very well be that W1 or W3 is a trivial walk.)
⊔
⊓


13

D EFINITION . If there exists a walk (and hence a path) from u to v in G, let
k

dG (u, v) = min{k | u − v}
→
⋆
be the distance between u and v. If there are no walks u − v, let dG (u, v) = ∞ by
→
convention. A graph G is connected, if dG (u, v) < ∞ for all u, v ∈ G; otherwise, it
is disconnected. The maximal connected subgraphs of G are its connected components. Denote

c( G ) = the number of connected components of G .
If c( G ) = 1, then G is, of course, connected.
The maximality condition means that a subgraph H ⊆ G is a connected component if and only if H is connected and there are no edges leaving H, i.e., for every vertex v ∈ H, the subgraph G [VH ∪ {v}] is disconnected. Apparently, every connected
/
component is an induced subgraph, and
∗
NG (v) = {u | dG (v, u) < ∞}

is the connected component of G that contains v ∈ G. In particular, the connected
components form a partition of G.

Shortest paths
D EFINITION . Let G α be an edge weighted graph, that is, G α is a graph G together
with a weight function α : EG → R on its edges. For H ⊆ G, let
α( H ) =

∑ α( e)
e∈ H

be the (total) weight of H. In particular, if P = e1 e2 . . . ek is a path, then its weight is
α( P) = ∑k=1 α(ei ). The minimum weighted distance between two vertices is
i
⋆
dα (u, v) = min{α( P) | P : u − v} .
→
G

In extremal problems we seek for optimal subgraphs H ⊆ G satisfying speciﬁc
conditions. In practice we encounter situations where G might represent
• a distribution or transportation network (say, for mail), where the weights on
edges are distances, travel expenses, or rates of ﬂow in the network;
• a system of channels in (tele)communication or computer architecture, where the
weights present the rate of unreliability or frequency of action of the connections;
• a model of chemical bonds, where the weights measure molecular attraction.


14

In these examples we look for a subgraph with the smallest weight, and which
connects two given vertices, or all vertices (if we want to travel around). On the other
hand, if the graph represents a network of pipelines, the weights are volumes or
capacities, and then one wants to find a subgraph with the maximum weight.
We consider the minimum problem. For this, let G be a graph with an integer
weight function α : EG → N. In this case, call α(uv) the length of uv.
The shortest path problem: Given a connected graph G with a weight function α : EG →
N, find dα (u, v) for given u, v ∈ G.
G
Assume that G is a connected graph. Dijkstra’s algorithm solves the problem for
every pair u, v, where u is a fixed starting point and v ∈ G. Let us make the convention that α(uv) = ∞, if uv ∈ G.
/
Dijkstra’s algorithm:
(i) Set u0 = u, t(u0 ) = 0 and t(v) = ∞ for all v = u0 .
(ii) For i ∈ [0, νG − 1]: for each v ∈ {u1 , . . . , ui },
/
replace t(v) by min{t(v), t(ui ) + α(ui v)} .
Let ui+1 ∈ {u1 , . . . , ui } be any vertex with the least value t(ui+1 ).
/
(iii) Conclusion: dα (u, v) = t(v).
G
Example 1.7. Consider the following weighted graph G. Apply Dijkstra’s algorithm
to the vertex v0 .

• u0 = v0 , t(u0 ) = 0, others are ∞.
2
• t(v1 ) = min{∞, 2} = 2, t(v2 ) = min{∞, 3} = 3,
1
v1
v3
others are ∞. Thus u1 = v1 .
1
2
• t(v2 ) = min{3, t(u1 ) + α(u1 v2 )} = min{3, 4} = 3,
3
v0
2
2
t(v3 ) = 2 + 1 = 3, t(v4 ) = 2 + 3 = 5, t(v5 ) = 2 + 2 = 4.
3
2
Thus choose u2 = v3 .
1
v2
v4
• t(v2 ) = min{3, ∞} = 3, t(v4 ) = min{5, 3 + 2} = 5,
t(v5 ) = min{4, 3 + 1} = 4. Thus set u3 = v2 .
• t(v4 ) = min{5, 3 + 1} = 4, t(v5 ) = min{4, ∞} = 4. Thus choose u4 = v4 .
• t(v5 ) = min{4, 4 + 1} = 4. The algorithm stops.
We have obtained:
t(v1 ) = 2, t(v2 ) = 3, t(v3 ) = 3, t(v4 ) = 4, t(v5 ) = 4 .
These are the minimal weights from v0 to each vi .

v5


15

The steps of the algorithm can also be rewritten as a table:
v1 2 - v2 3 3 3
v3 ∞ 3 v4 ∞ 5 5
v5 ∞ 4 4

4
4

4

The correctness of Dijkstra’s algorithm can veriﬁed be as follows.
⋆
⋆
Let v ∈ V be any vertex, and let P : u0 − u − v be a shortest path from u0 to v,
→ →
where u is any vertex u = v on such a path, possibly u = u0 . Then, clearly, the ﬁrst
⋆
⋆
part of the path, u0 − u, is a shortest path from u0 to u, and the latter part u − v
→
→
is a shortest path from u to v. Therefore, the length of the path P equals the sum of
⋆
⋆
the weights of u0 − u and u − v. Dijkstra’s algorithm makes use of this observation
→
→
iteratively.

2
Connectivity of Graphs
2.1 Bipartite graphs and trees
In problems such as the shortest path problem we look for minimum solutions that
satisfy the given requirements. The solutions in these cases are usually subgraphs
without cycles. Such connected graphs will be called trees, and they are used, e.g., in
search algorithms for databases. For concrete applications in this respect, see
T.H. C ORMEN , C.E. L EISERSON AND R.L. R IVEST, “Introduction to Algorithms”,
MIT Press, 1993.
+
Certain structures with operations are representable
as trees. These trees are sometimes called construction
y
·
trees, decomposition trees, factorization trees or grammatical
trees. Grammatical trees occur especially in linguistics,
x
+
where syntactic structures of sentences are analyzed.
On the right there is a tree of operations for the arithy
z
metic formula x · (y + z) + y.

Bipartite graphs
D EFINITION . A graph G is called bipartite, if VG has a partition to two subsets X and
Y such that each edge uv ∈ G connects a vertex of X and a vertex of Y. In this case,
( X, Y ) is a bipartition of G, and G is ( X, Y )-bipartite.
A bipartite graph G (as in the above) is complete (m, k)bipartite, if | X | = m, |Y | = k, and uv ∈ G for all u ∈ X
and v ∈ Y.
All complete (m, k)-bipartite graphs are isomorphic. Let
Km,k denote such a graph.
A subset X ⊆ VG is stable, if G [ X ] is a discrete graph.

K2,3

The following result is clear from the deﬁnitions.
Theorem 2.1. A graph G is bipartite if and only if VG has a partition to two stable subsets.
Example 2.1. The k-cube Qk of Example 1.5 is bipartite for all k. Indeed, consider
A = {u | u has an even number of 1′ s} and B = {u | u has an odd number of 1′ s}.
Clearly, these sets partition B k , and they are stable in Qk .


17

Theorem 2.2. A graph G is bipartite if and only if G it has no odd cycles (as subgraph).
Proof. (⇒) Observe that if G is ( X, Y )-bipartite, then so are all its subgraphs. However, an odd cycle C2k+1 is not bipartite.
(⇐) Suppose that all cycles in G are even. First, we note that it suffices to show
the claim for connected graphs. Indeed, if G is disconnected, then each cycle of G
is contained in one of the connected components G1 , . . . , G p of G. If Gi is ( Xi , Yi )bipartite, then G has the bipartition ( X1 ∪ X2 ∪ · · · ∪ X p , Y1 ∪ Y2 ∪ · · · ∪ Yp ).
Assume thus that G is connected. Let v ∈ G be a chosen vertex, and define
X = { x | dG (v, x) is even}

and

Y = {y | dG (v, y) is odd} .

Since G is connected, VG = X ∪ Y. Also, by the definition of distance, X ∩ Y = ∅.
⋆
⋆
Let then u, w ∈ G be both in X or both in Y, and let P : v − u and Q : v − w
→
→
be (among the) shortest paths from v to u and w. Assume that x is the last common
⋆
⋆
vertex of P and Q: P = P1 P2 , Q = Q1 Q2 , where P2 : x − u and Q2 : x − w are
→
→
⋆
independent. Since P and Q are shortest paths, P1 and Q1 are shortest paths v − x.
→
Consequently, | P1 | = | Q1 |.
u
Thus | P2 | and | Q2 | have the same parity and hence the
P2
P1
−1
sum | P2 | + | Q2 | is even, i.e., the path P2 Q2 is even,
v
x
uw
and so uw ∈ EG by assumption. Therefore X and Y are
/
Q1
stable subsets, and G is bipartite as claimed.
⊔
⊓
Q2
w
1
Checking whether a graph is bipartite is easy. Indeed,
2
1
this can be done by using two ‘opposite’ colours, say
1 and 2. Start from any vertex v1 , and colour it by 1.
2
2
2
Then colour the neighbours of v1 by 2, and proceed by
1
1
colouring all neighbours of an already coloured vertex
2
1
by the opposite colour.
If the whole graph can be coloured without contradiction, then G is ( X, Y )-bipartite,
where X consists of those vertices with colour 1, and Y of those vertices with colour
2; otherwise, at some point one of the vertices gets both colours, and in this case, G is
not bipartite.
Example 2.2 (E RDÖS (1965)). We show that each graph G has a bipartite subgraph
H ⊆ G such that ε H ≥ 1 ε G . Indeed, let VG = X ∪ Y be a partition such that the
2
number of edges between X and Y is maximum. Denote
F = EG ∩ {uv | u ∈ X, v ∈ Y } ,
and let H = G [ F ]. Obviously H is a spanning subgraph, and it is bipartite.
By the maximum condition, d H (v) ≥ dG (v)/2, since, otherwise, v is on the wrong
side. (That is, if v ∈ X, then the pair X ′ = X {v}, Y ′ = Y ∪ {v} does better that the
pair X, Y.) Now
1
1
1
1
ε H = ∑ d H ( v) ≥ ∑ d G ( v) = ε G .
2 v∈ H
2 v∈ G 2
2


18

Bridges
D EFINITION . An edge e ∈ G is a bridge of the graph G,
if G −e has more connected components than G, that is,
if c( G −e) > c( G ). In particular, and most importantly,
an edge e in a connected G is a bridge if and only if G −e
is disconnected.
On the right (only) the two horizontal lines are bridges.
We note that, for each edge e ∈ G,
e = uv is a bridge ⇐⇒ u, v in different connected components of G −e .
Theorem 2.3. An edge e ∈ G is a bridge if and only if e is not in any cycle of G.
⋆
Proof. (⇒) If there is a cycle in G containing e, say C = PeQ, then QP : v − u is a
→
path in G −e, and so e is not a bridge.
(⇐) If e = uv is not a bridge, then u and v are in the same connected component
⋆
⋆
of G −e, and there is a path P : v − u in G −e. Now, eP : u − v − u is a cycle in G
→
→ →
containing e.
⊔
⊓

Lemma 2.1. Let e be a bridge in a connected graph G.
(i) Then c( G −e) = 2.
(ii) Let H be a connected component of G −e. If f ∈ H is a bridge of H, then f is a bridge
of G.
Proof. For (i), let e = uv. Since e is a bridge, the ends u and v are not connected in
⋆
G −e. Let w ∈ G. Since G is connected, there exists a path P : w − v in G. This is a
→
⋆
⋆
path of G −e, unless P : w − u → v contains e = uv, in which case the part w − u is
→
→
a path in G −e.
For (ii), if f ∈ H belongs to a cycle C of G, then C does not contain e (since e is in
no cycle), and therefore C is inside H, and f is not a bridge of H.
⊔
⊓

Trees
D EFINITION . A graph is called acyclic, if it has no cycles. An acyclic graph is also
called a forest. A tree is a connected acyclic graph.
By Theorem 2.3 and the deﬁnition of a tree, we have
Corollary 2.1. A connected graph is a tree if and only if all its edges are bridges.
Example 2.3. The following enumeration result for trees has many different proofs,
the ﬁrst of which was given by C AYLEY in 1889: There are nn−2 trees on a vertex set V of
n elements. We omit the proof.


19

On the other hand, there are only a few trees up to isomorphism:
n
trees

1
1

2
1

3
1

4
2

5
3

6
6

7 8
11 23

n
9 10 11 12
13
14
15
16
trees 47 106 235 551 1301 3159 7741 19 320
The nonisomorphic trees of order 6 are:

⋆
We say that a path P : u − v is maximal in a graph G, if there are no edges e ∈ G
→
for which Pe or eP is a path. Such paths exist, because νG is ﬁnite.
⋆
Lemma 2.2. Let P : u − v be a maximal path in a graph G. Then NG (v) ⊆ P. Moreover, if
→
G is acyclic, then dG (v) = 1.

Proof. If e = vw ∈ EG with w ∈ P, then also Pe is a path, which contradicts the
/
maximality assumption for P. Hence NG (v) ⊆ P. For acyclic graphs, if wv ∈ G, then
w belongs to P, and wv is necessarily the last edge of P in order to avoid cycles.
⊔
⊓
Corollary 2.2. Each tree T with νT ≥ 2 has at least two leaves.
Proof. Since T is acyclic, both ends of a maximal path have degree one.

⊔
⊓

Theorem 2.4. The following are equivalent for a graph T.
(i) T is a tree.
(ii) Any two vertices are connected in T by a unique path.
(iii) T is acyclic and ε T = νT − 1.
Proof. Let νT = n. If n = 1, then the claim is trivial. Suppose thus that n ≥ 2.
⋆
(i)⇒(ii) Let T be a tree. Assume the claim does not hold, and let P, Q : u − v
→
be two different paths between the same vertices u and v. Suppose that | P| ≥ | Q|.
Since P = Q, there exists an edge e which belongs to P but not to Q. Each edge of T
is a bridge, and therefore u and v belong to different connected components of T −e.
Hence e must also belong to Q; a contradiction.

(ii)⇒(iii) We prove the claim by induction on n. Clearly, the claim holds for n = 2,
and suppose it holds for graphs of order less than n. Let T be any graph of order n
satisfying (ii). In particular, T is connected, and it is clearly acyclic.


20

⋆
Let P : u − v be a maximal path in T. By Lemma 2.2, we have dT (v) = 1. In this
→
⋆
case, P : u − w − v, where vw is the unique edge having an end v. The subgraph
→
→
T −v is connected, and it satisﬁes the condition (ii). By induction hypothesis, ε T −v =
n − 2, and so ε T = ε T −v + 1 = n − 1, and the claim follows.

(iii)⇒(i) Assume (iii) holds for T. We need to show that T is connected. Indeed,
let the connected components of T be Ti = (Vi , Ei ), for i ∈ [1, k]. Since T is acyclic, so
are the connected graphs Ti , and hence they are trees, for which we have proved that
|Ei | = |Vi | − 1. Now, νT = ∑k=1 |Vi |, and ε T = ∑k=1 |Ei |. Therefore,
i
i
k

k

i=1

n − 1 = εT =

i=1

∑ (|Vi | − 1) = ∑ |Vi | − k = n − k ,

which gives that k = 1, that is, T is connected.

⊔
⊓

Example 2.4. Consider a cup tournament of n teams. If during a round there are k
teams left in the tournament, then these are divided into ⌊k⌋ pairs, and from each
pair only the winner continues. If k is odd, then one of the teams goes to the next
round without having to play. How many plays are needed to determine the winner?
So if there are 14 teams, after the ﬁrst round 7 teams continue, and after the second
round 4 teams continue, then 2. So 13 plays are needed in this example.
The answer to our problem is n − 1, since the cup tournament is a tree, where a
play corresponds to an edge of the tree.

Spanning trees
Theorem 2.5. Each connected graph has a spanning tree, that is, a spanning graph that is
a tree.
Proof. Let T ⊆ G be a maximum order subtree of G (i.e., subgraph that is a tree). If
VT = VG , there exists an edge uv ∈ EG such that u ∈ T and v ∈ T. But then T is not
/
/
maximal; a contradiction.
⊔
⊓
Corollary 2.3. For each connected graph G, ε G ≥ νG − 1. Moreover, a connected graph G is
a tree if and only if ε G = νG − 1.
Proof. Let T be a spanning tree of G. Then ε G ≥ ε T = νT − 1 = νG − 1. The second
claim is also clear.
⊔
⊓
Example 2.5. In Shannon’s switching game a positive player P and a negative player
N play on a graph G with two special vertices: a source s and a sink r. P and N alternate turns so that P designates an edge by +, and N by −. Each edge can be des⋆
ignated at most once. It is P’s purpose to designate a path s − r (that is, to designate
→
⋆
all edges in one such path), and N tries to block all paths s − r (that is, to designate
→
at least one edge in each such path). We say that a game ( G, s, r) is


21

• positive, if P has a winning strategy no matter who begins the game,
• negative, if N has a winning strategy no matter who begins the game,
• neutral, if the winner depends on who begins the game.
r
The game on the right is neutral.
s
L EHMAN proved in 1964 that Shannon’s switching game ( G, s, r) is positive if and only
if there exists H ⊆ G such that H contains s and r and H has two spanning trees with no
edges in common.
In the other direction the claim can be proved along the following lines. Assume
that there exists a subgraph H containing s and r and that has two spanning trees
with no edges in common. Then P plays as follows. If N marks by − an edge from
one of the two trees, then P marks by + an edge in the other tree such that this
edge reconnects the broken tree. In this way, P always has two spanning trees for the
subgraph H with only edges marked by + in common.
In converse the claim is considerably more difficult to prove.
There remains the problem to characterize those Shannon’s switching games
( G, s, r) that are neutral (negative, respectively).

The connector problem
To build a network connecting n nodes (towns, computers, chips in a computer) it
is desirable to decrease the cost of construction of the links to the minimum. This is
the connector problem. In graph theoretical terms we wish to find an optimal spanning subgraph of a weighted graph. Such an optimal subgraph is clearly a spanning
tree, for, otherwise a deletion of any nonbridge will reduce the total weight of the
subgraph.
Let then G α be a graph G together with a weight function α : EG → R + (positive reals) on the edges. Kruskal’s algorithm (also known as the greedy algorithm)
provides a solution to the connector problem.
Kruskal’s algorithm: For a connected and weighted graph G α of order n:
(i) Let e1 be an edge of smallest weight, and set E1 = {e1 }.
(ii) For each i = 2, 3, . . . , n − 1 in this order, choose an edge ei ∈ Ei−1 of smallest
/
possible weight such that ei does not produce a cycle when added to G [ Ei−1 ], and
let Ei = Ei−1 ∪ {ei }.
The final outcome is T = (VG , En−1 ).


22

By the construction, T = (VG , En−1 ) is a spanning tree of G, because it contains no
cycles, it is connected and has n − 1 edges. We now show that T has the minimum
total weight among the spanning trees of G.
Suppose T1 is any spanning tree of G. Let ek be the ﬁrst edge produced by the
algorithm that is not in T1 . If we add ek to T1 , then a cycle C containing ek is created.
Also, C must contain an edge e that is not in T. When we replace e by ek in T1 , we
still have a spanning tree, say T2 . However, by the construction, α(ek ) ≤ α(e), and
therefore α( T2 ) ≤ α( T1 ). Note that T2 has more edges in common with T than T1 .
Repeating the above procedure, we can transform T1 to T by replacing edges, one
by one, such that the total weight does not increase. We deduce that α( T ) ≤ α( T1 ).
The outcome of Kruskal’s algorithm need not be unique. Indeed, there may exist
several optimal spanning trees (with the same weight, of course) for a graph.
Example 2.6. When applied to the weighted
graph on the right, the algorithm produces the sequence: e1 = v2 v4 , e2 = v4 v5 , e3 = v3 v6 , e4 = v2 v3
and e5 = v1 v2 . The total weight of the spanning
tree is thus 9.
Also, the selection e1 = v2 v5 , e2 = v4 v5 , e3 = v5 v6 ,
e4 = v3 v6 , e5 = v1 v2 gives another optimal solution (of weight 9).

v1

3

2

v2

v3

1
1

4

2

2

2
1

v4

2

v5

v6

3

Problem. Consider trees T with weight functions α : ET → N. Each tree T of order n
has exactly (n) paths. (Why is this so?) Does there exist a weighted tree T α of order n such
2
that the (total) weights of its paths are 1, 2, . . . , (n)?
2
In such a weighted tree T α different paths have
different weights, and each i ∈ [1, ( n)] is a weight
2
of one path. Also, α must be injective.
No solutions are known for any n ≥ 7.

1

4
5

2

8

TAYLOR (1977) proved: if T of order n exists, then necessarily n = k2 or n = k2 + 2 for
some k ≥ 1.
Example 2.7. A computer network can be presented as a graph G, where the vertices
are the node computers, and the edges indicate the direct links. Each computer v has
an address a(v), a bit string (of zeros and ones). The length of an address is the number
of its bits. A message that is sent to v is preceded by the address a(v). The Hamming
distance h( a(v), a(u)) of two addresses of the same length is the number of places,
where a(v) and a(u) differ; e.g., h(00010, 01100) = 3 and h(10000, 00000) = 1.
It would be a good way to address the vertices so that the Hamming distance
of two vertices is the same as their distance in G. In particular, if two vertices were
adjacent, their addresses should differ by one symbol. This would make it easier for
a node computer to forward a message.

2.2 Connectivity

23

A graph G is said to be addressable, if it has an
addressing a such that dG (u, v) = h( a(u), a(v)).

010
000

110

111

100

We prove that every tree T is addressable. Moreover, the addresses of the vertices of T can
be chosen to be of length νT − 1.
The proof goes by induction. If νT ≤ 2, then the claim is obvious. In the case
νT = 2, the addresses of the vertices are simply 0 and 1.
Let then VT = {v1 , . . . , vk+1 }, and assume that dT (v1 ) = 1 (a leaf) and v1 v2 ∈ T. By
the induction hypothesis, we can address the tree T −v1 by addresses of length k − 1.
We change this addressing: let ai be the address of vi in T −v1 , and change it to 0ai .
Set the address of v1 to 1a2 . It is now easy to see that we have obtained an addressing
for T as required.
The triangle K3 is not addressable. In order to gain more generality, we modify
the addressing for general graphs by introducing a special symbol ∗ in addition to
0 and 1. A star address will be a sequence of these three symbols. The Hamming
distance remains as it was, that is, h(u, v) is the number of places, where u and v
have a different symbol 0 or 1. The special symbol ∗ does not affect h(u, v). So, h(10 ∗
∗01, 0 ∗ ∗101) = 1 and h(1 ∗ ∗ ∗ ∗∗, ∗00 ∗ ∗∗) = 0. We still want to have h(u, v) =
dG (u, v).
We star address this graph as follows:
a(v1 ) = 0000 ,
a(v3 ) = 1 ∗ 01 ,

a(v4 ) = ∗ ∗ 11 .

v3

a(v2 ) = 10 ∗ 0 ,

These addresses have length 4. Can you design a
star addressing with addresses of length 3?

v1

v2
v4

W INKLER proved in 1983 a rather unexpected result: The minimum star address
length of a graph G is at most νG − 1.
For the proof of this, see VAN L INT AND W ILSON, “A Course in Combinatorics”.

2.2 Connectivity
Spanning trees are often optimal solutions to problems, where cost is the criterion.
We may also wish to construct graphs that are as simple as possible, but where two
vertices are always connected by at least two independent paths. These problems occur especially in different aspects of fault tolerance and reliability of networks, where
one has to make sure that a break-down of one connection does not affect the functionality of the network. Similarly, in a reliable network we require that a break-down
of a node (computer) should not result in the inactivity of the whole network.

2.2 Connectivity

24

Separating sets
D EFINITION . A vertex v ∈ G is a cut vertex, if c( G −v) > c( G ).
A subset S ⊆ VG is a separating set, if G −S is disconnected. We
also say that S separates the vertices u and v and it is a (u, v)separating set, if u and v belong to different connected components of G −S.
If G is connected, then v is a cut vertex if and only if G −v is disconnected, that is,
{v} is a separating set. The following lemma is immediate.
⋆
Lemma 2.3. If S ⊆ VG separates u and v, then every path P : u − v visits a vertex of S.
→

Lemma 2.4. If a connected graph G has no separating sets, then it is a complete graph.
Proof. If νG ≤ 2, then the claim is clear. For νG ≥ 3, assume that G is not complete,
and let uv ∈ G. Now VG {u, v} is a separating set. The claim follows from this. ⊔
/
⊓
D EFINITION . The (vertex) connectivity number κ ( G ) of G is deﬁned as
κ ( G ) = min{k | k = |S|, G −S disconnected or trivial, S ⊆ VG } .
A graph G is k-connected, if κ ( G ) ≥ k.
In other words,
• κ ( G ) = 0, if G is disconnected,
• κ ( G ) = νG − 1, if G is a complete graph, and
• otherwise κ ( G ) equals the minimum size of a separating set of G.
Clearly, if G is connected, then it is 1-connected.
D EFINITION . An edge cut F of G consists of edges so that G − F is disconnected. Let
κ ′ ( G ) = min{k | k = | F |, G − F disconnected, F ⊆ EG } .
For trivial graphs, let κ ′ ( G ) = 0. A graph G is k-edge connected, if κ ′ ( G ) ≥ k. A
minimal edge cut F ⊆ EG is a bond (F {e} is not an edge cut for any e ∈ F).
Example 2.8. Again, if G is disconnected, then
κ ′ ( G ) = 0. On the right, κ ( G ) = 2 and κ ′ ( G ) = 2.
Notice that the minimum degree is δ( G ) = 3.
Lemma 2.5. Let G be connected. If e = uv is a bridge, then either G = K2 or one of u or v is
a cut vertex.

2.2 Connectivity

25

Proof. Assume that G = K2 and thus that νG ≥ 3, since G is connected. Let Gu =
∗
∗
NG−e (u) and Gv = NG−e (v) be the connected components of G −e containing u and
v. Now, either νGu ≥ 2 (and u is a cut vertex) or νGv ≥ 2 (and v is a cut vertex).
⊔
⊓
Lemma 2.6. If F be a bond of a connected graph G, then c( G − F ) = 2.
Proof. Since G − F is disconnected, and F is minimal, the subgraph H = G −( F {e})
is connected for given e ∈ F. Hence e is a bridge in H. By Lemma 2.1, c( H −e) = 2,
and thus c( G − F ) = 2, since H −e = G − F.
⊔
⊓
Theorem 2.6 (W HITNEY (1932)). For any graph G,
κ ( G ) ≤ κ ′ ( G ) ≤ δ( G ) .
Proof. Assume G is nontrivial. Clearly, κ ′ ( G ) ≤ δ( G ), since if we remove all edges
with an end v, we disconnect G. If κ ′ ( G ) = 0, then G is disconnected, and in this case
also κ ( G ) = 0. If κ ′ ( G ) = 1, then G is connected and contains a bridge. By Lemma 2.5,
either G = K2 or G has a cut vertex. In both of these cases, also κ ( G ) = 1.
Assume then that κ ′ ( G ) ≥ 2. Let F be an edge cut of G with | F | = κ ′ ( G ), and let
e = uv ∈ F. Then F is a bond, and G − F has two connected components.
Consider the connected subgraph
H = G −( F {e}) = ( G − F ) + e,

G .
.
.

.
.
.

H .
.
.

e

.
.
.

where e is a bridge.
F
Now for each f ∈ F {e} choose an end different from u and v. (The choices for
different edges need not be different.) Note that since f = e, either end of f is different
from u or v. Let S be the collection of these choices. Thus |S| ≤ | F | − 1 = κ ′ ( G ) − 1,
and G −S does not contain edges from F {e}.
If G −S is disconnected, then S is a separating set and so κ ( G ) ≤ |S| ≤ κ ′ ( G ) − 1
and we are done. On the other hand, if G −S is connected, then either G −S = K2 (= e),
or either u or v (or both) is a cut vertex of G −S (since H −S = G −S, and therefore
G −S ⊆ H is an induced subgraph of H). In both of these cases, there is a vertex
of G −S, whose removal results in a trivial or a disconnected graph. In conclusion,
κ ( G ) ≤ |S| + 1 ≤ κ ′ ( G ), and the claim follows.
⊔
⊓

Menger’s theorem
Theorem 2.7 (M ENGER (1927)). Let u, v ∈ G be nonadjacent vertices of a connected graph
G. Then the minimum number of vertices separating u and v is equal to the maximum number
of independent paths from u to v.
⋆
Proof. If a subset S ⊆ VG is (u, v)-separating, then every path u − v of G visits S.
→
Hence |S| is at least the number of independent paths from u to v.

2.2 Connectivity

26

Conversely, we use induction on m = νG + ε G to show that if S = {w1 , w2 , . . . , wk }
is a (u, v)-separating set of the smallest size, then G has at least (and thus exactly) k
⋆
independent paths u − v.
→
The case for k = 1 is clear, and this takes care of the small values of m, required
for the induction.
(1) Assume first that u and v have a common neighbour w ∈ NG (u) ∩ NG (v). Then
necessarily w ∈ S. In the smaller graph G −w the set S {w} is a minimum (u, v)separating set, and the induction hypothesis yields that there are k − 1 independent
⋆
paths u − v in G −w. Together with the path u − w − v, there are k independent
→
→ →
⋆
paths u − v in G as required.
→
∗
(2) Assume then that NG (u) ∩ NG (v) = ∅, and denote by Hu = NG−S (u) and
∗
Hv = NG−S (v) the connected components of G −S for u and v.

(2.1) Suppose next that S

NG (u) and S

NG (v).

Let v be a new vertex, and define Gu to be the graph
on Hu ∪ S ∪ {v } having the edges of G [ Hu ∪ S] together with vwi for all i ∈ [1, k]. The graph Gu is connected and it is smaller than G. Indeed, in order for
S to be a minimum separating set, all wi ∈ S have
to be adjacent to some vertex in Hv . This shows that
ε Gu ≤ ε G , and, moreover, the assumption (2.1) rules
out the case Hv = {v}. So | Hv | ≥ 2 and νGu < νG .

wk
...
w2

v

u
w1

If S′ is any (u, v)-separating set of Gu , then S′ will separate u from all wi ∈ S S′ in
G. This means that S′ separates u and v in G. Since k is the size of a minimum (u, v)separating set, we have |S′ | ≥ k. We noted that Gu is smaller than G, and thus by the
⋆
induction hypothesis, there are k independent paths u − v in Gu . This is possible
→
⋆
only if there exist k paths u − wi , one for each i ∈ [1, k], that have only the end u in
→
common.
By the present assumption, also u is nonadjacent to some vertex of S. A symmetric
argument applies to the graph Gv (with a new vertex u), which is defined similarly
⋆
to Gu . This yields that there are k paths wi − v that have only the end v in common.
→
⋆
When we combine these with the above paths u − wi , we obtain k independent
→
⋆
⋆
paths u − wi − v in G.
→
→
(2.2) There remains the case, where for all (u, v)-separating sets S of k elements,
either S ⊆ NG (u) or S ⊆ NG (v). (Note that then, by (2), S ∩ NG (v) = ∅ or S ∩
NG (u) = ∅.)
⋆
⋆
Let P = e f Q be a shortest path u − v in G, where e = ux, f = xy, and Q : y − v.
→
→
Notice that, by the assumption (2), | P| ≥ 3, and so y = v. In the smaller graph G − f ,
let S′ be a minimum set that separates u and v.
⋆
If |S′ | ≥ k, then, by the induction hypothesis, there are k independent paths u − v
→
in G − f . But these are paths of G, and the claim is clear in this case.

2.2 Connectivity

27

If, on the other hand, |S′ | < k, then u and v are still connected in G −S′ . Every path
u − v in G −S′ necessarily travels along the edge f = xy, and so x, y ∈ S′ .
→
/
Let
Sx = S′ ∪ { x}
and
Sy = S ′ ∪ {y} .
⋆

These sets separate u and v in G (by the above fact), and they have size k. By our
current assumption, the vertices of Sy are adjacent to v, since the path P is shortest
and so uy ∈ G (meaning that u is not adjacent to all of Sy ). The assumption (2) yields
/
that u is adjacent to all of Sx , since ux ∈ G. But now both u and v are adjacent to the
vertices of S′ , which contradicts the assumption (2).
⊔
⊓
Theorem 2.8 (M ENGER (1927)). A graph G is k-connected if and only if every two vertices
are connected by at least k independent paths.
Proof. If any two vertices are connected by k independent paths, then it is clear
that κ ( G ) ≥ k.
In converse, suppose that κ ( G ) = k, but that G has vertices u and v connected by at
most k − 1 independent paths. By Theorem 2.7, it must be that e = uv ∈ G. Consider
the graph G −e. Now u and v are connected by at most k − 2 independent paths in
G −e, and by Theorem 2.7, u and v can be separated in G −e by a set S with |S| = k − 2.
Since νG > k (because κ ( G ) = k), there exists a w ∈ G that is not in S ∪ {u, v}. The
vertex w is separated in G −e by S from u or from v; otherwise there would be a path
⋆
u − v in ( G −e)−S. Say, this vertex is u. The set S ∪ {v} has k − 1 elements, and it
→
separates u from w in G, which contradicts the assumption that κ ( G ) = k. This proves
the claim.
⊔
⊓
We state without a proof the corresponding separation property for edge connectivity.
D EFINITION . Let G be a graph. A uv-disconnecting set is a set F ⊆ EG such that
⋆
every path u − v contains an edge from F.
→
Theorem 2.9. Let u, v ∈ G with u = v in a graph G. Then the maximum number of edge⋆
disjoint paths u − v equals the minimum number k of edges in a uv-disconnecting set.
→
Corollary 2.4. A graph G is k-edge connected if and only if every two vertices are connected
by at least k edge disjoint paths.
Example 2.9. Recall the deﬁnition of the cube Qk from Example 1.5. We show that
κ ( Qk ) = k.
First of all, κ ( Qk ) ≤ δ( Qk ) = k. In converse, we show the claim by induction.
Extract from Qk the disjoint subgraphs: G0 induced by {0u | u ∈ B k−1 } and G1
induced by {1u | u ∈ B k−1 }. These are (isomorphic to) Qk−1 , and Qk is obtained from
the union of G0 and G1 by adding the 2k−1 edges (0u, 1u) for all u ∈ B k−1 .

2.2 Connectivity

28

Let S be a separating set of Qk with |S| ≤ k. If both G0 −S and G1 −S were connected, also Qk −S would be connected, since one pair (0u, 1u) necessarily remains in
Qk −S. So we can assume that G0 −S is disconnected. (The case for G1 −S is symmetric.) By the induction hypothesis, κ ( G0 ) = k − 1, and hence S contains at least k − 1
vertices of G0 (and so |S| ≥ k − 1). If there were no vertices from G1 in S, then, of
course, G1 −S is connected, and the edges (0u, 1u) of Qk would guarantee that Qk −S
is connected; a contradiction. Hence |S| ≥ k.
Example 2.10. We have κ ′ ( Qk ) = k for the k-cube. Indeed, by Whitney’s theorem,
κ ( G ) ≤ κ ′ ( G ) ≤ δ( G ). Since κ ( Qk ) = k = δ( Qk ), also κ ′ ( Qk ) = k.
Algorithmic Problem. The connectivity problems tend to be algorithmically difficult.
In the disjoint paths problem we are given a set (ui , vi ) of pairs of vertices for i =
⋆
1, 2, . . . , k, and it is asked whether there exist paths Pi : ui − vi that have no vertices in
→
common. This problem was shown to be NP-complete by K NUTH in 1975. (However,
for fixed k, the problem has a fast algorithm due to R OBERTSON and S EYMOUR (1986).)

Dirac’s fans
D EFINITION . Let v ∈ G and S ⊆ VG such that v ∈ S
/
in a graph G. A set of paths from v to a vertex in S is
called a (v, S)-fan, if they have only v in common.
Theorem 2.10 (D IRAC (1960)). A graph G is k-connected
if and only if νG > k and for every v ∈ G and S ⊆ VG with
|S| ≥ k and v ∈ S, there exists a (v, S)-fan of k paths.
/
Proof. Exercise.

...

∗

v

∗
∗

S

⊔
⊓

Theorem 2.11 (D IRAC (1960)). Let G be a k-connected graph for k ≥ 2. Then for any k
vertices, there exists a cycle of G containing them.
Proof. First of all, since κ ( G ) ≥ 2, G has no cut vertices, and thus no bridges. It
follows that every edge, and thus every vertex of G belongs to a cycle.
Let S ⊆ VG be such that |S| = k, and let C be a cycle of G that contains the
maximum number of vertices of S. Let the vertices of S ∩ VC be v1 , . . . , vr listed in
order around C so that each pair (vi , vi+1 ) (with indices modulo r) defines a path
along C (except in the special case where r = 1). Such a path is referred to as a segment
of C. If C contains all vertices of S, then we are done; otherwise, suppose v ∈ S is not
on C.
It follows from Theorem 2.10 that there is a (v, VC )-fan of at least min{k, |VC |}
⋆
⋆
paths. Therefore there are two paths P : v − u and Q : v − w in such a fan that end
→
→
⋆
⋆
in the same segment (vi , vi+1 ) of C. Then the path W : u − w (or w − u) along C
→
→
contains all vertices of S ∩ VC . But now PWQ−1 is a cycle of G that contains v and all
vi for i ∈ [1, r]. This contradicts the choice of C, and proves the claim.
⊔
⊓

3
Tours and Matchings
3.1 Eulerian graphs
The ﬁrst proper problem in graph theory was the Königsberg bridge problem. In general, this problem concerns of travelling in a graph such that one tries to avoid using
any edge twice. In practice these eulerian problems occur, for instance, in optimizing distribution networks – such as delivering mail, where in order to save time each
street should be travelled only once. The same problem occurs in mechanical graph
plotting, where one avoids lifting the pen off the paper while drawing the lines.

Euler tours
D EFINITION . A walk W = e1 e2 . . . en is a trail, if ei = e j for all i = j. An Euler trail of
a graph G is a trail that visits every edge once. A connected graph G is eulerian, if it
has a closed trail containing every edge of G. Such a trail is called an Euler tour.
Notice that if W = e1 e2 . . . en is an Euler tour (and so EG = {e1 , e2 , . . . , en }), also
ei ei+1 . . . en e1 . . . ei−1 is an Euler tour for all i ∈ [1, n]. A complete proof of the following
Euler’s Theorem was ﬁrst given by H IERHOLZER in 1873.
Theorem 3.1 (E ULER (1736), H IERHOLZER (1873)). A connected graph G is eulerian if
and only if every vertex has an even degree.
⋆
Proof. (⇒) Suppose W : u − u is an Euler tour. Let v (= u) be a vertex that occurs
→
k times in W. Every time an edge arrives at v, another edge departs from v, and
therefore dG (v) = 2k. Also, dG (u) is even, since W starts and ends at u.
(⇐) Assume G is a nontrivial connected graph such that dG (v) is even for all v ∈
G. Let
⋆
W = e1 e2 . . . en : v0 − vn with ei = vi−1 vi
→

be a longest trail in G. It follows that all e = vn w ∈ G are among the edges of W, for,
otherwise, W could be prolonged to We. In particular, v0 = vn , that is, W is a closed
trail. (Indeed, if it were vn = v0 and vn occurs k times in W, then dG (vn ) = 2(k − 1) + 1
and that would be odd.)
If W is not an Euler tour, then, since G is connected, there exists an edge f = vi u ∈
G for some i, which is not in W. However, now
e i + 1 . . . e n e1 . . . e i f
is a trail in G, and it is longer than W. This contradiction to the choice of W proves
the claim.
⊔
⊓

3.1 Eulerian graphs

30

Example 3.1. The k-cube Qk is eulerian for even integers k, because Qk is k-regular.
Theorem 3.2. A connected graph has an Euler trail if and only if it has at most two vertices
of odd degree.
⋆
Proof. If G has an Euler trail u − v, then, as in the proof of Theorem 3.1, each vertex
→
w ∈ {u, v} has an even degree.
/
Assume then that G is connected and has at most two vertices of odd degree. If G
has no vertices of odd degree then, by Theorem 3.1, G has an Euler trail. Otherwise,
by the handshaking lemma, every graph has an even number of vertices with odd
degree, and therefore G has exactly two such vertices, say u and v. Let H be a graph
obtained from G by adding a vertex w, and the edges uw and vw. In H every vertex
⋆
has an even degree, and hence it has an Euler tour, say u − v − w − u. Here the
→ → →
⋆
beginning part u − v is an Euler trail of G.
→
⊔
⊓

The Chinese postman
The following problem is due to G UAN M EIGU (1962). Consider a village, where a
postman wishes to plan his route to save the legs, but still every street has to be
walked through. This problem is akin to Euler’s problem and to the shortest path
problem.
Let G be a graph with a weight function α : EG → R + . The Chinese postman
problem is to find a minimum weighted tour in G (starting from a given vertex, the
post office).
If G is eulerian, then any Euler tour will do as a solution, because such a tour
traverses each edge exactly once and this is the best one can do. In this case the weight
of the optimal tour is the total weight of the graph G, and there is a good algorithm
for finding such a tour:
Fleury’s algorithm:
• Let v0 ∈ G be a chosen vertex, and let W0 be the trivial path on v0 .
• Repeat the following procedure for i = 1, 2, . . . as long as possible: suppose a trail
Wi = e1 e2 . . . ei has been constructed, where e j = v j−1 v j .
Choose an edge ei+1 (= e j for j ∈ [1, i ]) so that
(i) ei+1 has an end vi , and
(ii) ei+1 is not a bridge of Gi = G −{e1 , . . . , ei }, unless there is no alternative.
Notice that, as is natural, the weights α(e) play no role in the eulerian case.
Theorem 3.3. If G is eulerian, then any trail of G constructed by Fleury’s algorithm is an
Euler tour of G.
Proof. Exercise.

⊔
⊓

3.2 Hamiltonian graphs

31

If G is not eulerian, the poor postman has to walk at least one street twice. This
happens, e.g., if one of the streets is a dead end, and in general if there is a street corner
of an odd number of streets. We can attack this case by reducing it to the eulerian case
as follows. An edge e = uv will be duplicated, if it is added to G parallel to an existing
edge e′ = uv with the same weight, α(e′ ) = α(e).
4

4

3
3

2

1

4

3
3

2

2

2

1

2

2

2

3
3
1

3

2

Above we have duplicated two edges. The rightmost multigraph is eulerian.
There is a good algorithm by E DMONDS AND J OHNSON (1973) for the construction
of an optimal eulerian supergraph by duplications. Unfortunately, this algorithm is
somewhat complicated, and we shall skip it.

In the connector problem we reduced the cost of a spanning graph to its minimum.
There are different problems, where the cost is measured by an active user of the
graph. For instance, in the travelling salesman problem a person is supposed to visit
each town in his district, and this he should do in such a way that saves time and
money. Obviously, he should plan the travel so as to visit each town once, and so
that the overall flight time is as short as possible. In terms of graphs, he is looking
for a minimum weighted Hamilton cycle of a graph, the vertices of which are the
towns and the weights on the edges are the flight times. Unlike for the shortest path
and the connector problems no efficient reliable algorithm is known for the travelling
salesman problem. Indeed, it is widely believed that no practical algorithm exists for
this problem.

Hamilton cycles
D EFINITION . A path P of a graph G is a Hamilton path,
if P visits every vertex of G once. Similarly, a cycle C is
a Hamilton cycle, if it visits each vertex once. A graph is
hamiltonian, if it has a Hamilton cycle.
Note that if C : u1 → u2 → · · · → un is a Hamilton cycle, then so is ui → . . . un →
u1 → . . . ui−1 for each i ∈ [1, n], and thus we can choose where to start the cycle.
Example 3.2. It is obvious that each Kn is hamiltonian whenever n ≥ 3. Also, as is
easily seen, Kn,m is hamiltonian if and only if n = m ≥ 2. Indeed, let Kn,m have a


32

bipartition ( X, Y ), where | X | = n and |Y | = m. Now, each cycle in Kn,m has even
length as the graph is bipartite, and thus the cycle visits the sets X, Y equally many
times, since X and Y are stable subsets. But then necessarily | X | = |Y |.
Unlike for eulerian graphs (Theorem 3.1) no good characterization is known for
hamiltonian graphs. Indeed, the problem to determine if G is hamiltonian is NPcomplete. There are, however, some interesting general conditions.
Lemma 3.1. If G is hamiltonian, then for every nonempty subset S ⊆ VG ,
c( G − S ) ≤ |S | .
⋆
Proof. Let ∅ = S ⊆ VG , u ∈ S, and let C : u − u be a Hamilton cycle of G. Assume
→
G −S has k connected components, Gi , i ∈ [1, k]. The case k = 1 is trivial, and hence
suppose that k > 1. Let ui be the last vertex of C that belongs to Gi , and let vi be the
vertex that follows ui in C. Now vi ∈ S for each i by the choice of ui , and v j = vt for
all j = t, because C is a cycle and ui vi ∈ G for all i. Thus |S| ≥ k as required.
⊔
⊓

Example 3.3. Consider the graph on the right. In G,
c( G −S) = 3 > 2 = |S| for the set S of black vertices. Therefore G does not satisfy the condition of
Lemma 3.1, and hence it is not hamiltonian. Interestingly this graph is ( X, Y )-bipartite of even order with
|X | = |Y |. It is also 3-regular.
Example 3.4. Consider the Petersen graph on the right,
which appears in many places in graph theory as a
counter example for various conditions. This graph
is not hamiltonian, but it does satisfy the condition
c( G −S) ≤ |S| for all S = ∅. Therefore the conclusion
of Lemma 3.1 is not sufﬁcient to ensure that a graph is
hamiltonian.
The following theorem, due to O RE, generalizes an earlier result by D IRAC (1952).
Theorem 3.4 (O RE (1962)). Let G be a graph of order νG ≥ 3, and let u, v ∈ G be such that
dG (u) + dG (v) ≥ νG .
Then G is hamiltonian if and only if G + uv is hamiltonian.
Proof. Denote n = νG . Let u, v ∈ G be such that dG (u) + dG (v) ≥ n. If uv ∈ G, then
there is nothing to prove. Assume thus that uv ∈ G.
/
(⇒) This is trivial since if G has a Hamilton cycle C, then C is also a Hamilton
cycle of G + uv.
(⇐) Denote e = uv and suppose that G + e has a Hamilton cycle C. If C does not
use the edge e, then it is a Hamilton cycle of G. Suppose thus that e is on C. We may
⋆
then assume that C : u − v − u. Now u = v1 − v2 − . . . − vn = v is a Hamilton
→ →
→
→
→


33

path of G. There exists an i with 1 < i < n such that uvi ∈ G and vi−1 v ∈ G. For,
otherwise, dG (v) < n − dG (u) would contradict the assumption.
v1

v2

◦

◦

v i −1

vi

◦

◦

vn

⋆
⋆
But now u = v1 − vi−1 − vn − vn−1 − vi+1 − vi − v1 = u is a Hamilton cycle
→
→
→
→
→
→
in G.
⊔
⊓

Closure
D EFINITION . For a graph G, define inductively a sequence G0 , G1 , . . . , Gk of graphs
such that
G0 = G and Gi+1 = Gi + uv ,
where u and v are any vertices such that uv ∈ Gi and dGi (u) + dGi (v) ≥ νG . This
/
procedure stops when no new edges can be added to Gk for some k, that is, in Gk , for
all u, v ∈ G either uv ∈ Gk or dGk (u) + dGk (v) < νG . The result of this procedure is the
closure of G, and it is denoted by cl ( G ) (= Gk ) .
In each step of the construction of cl ( G ) there are usually alternatives which edge
uv is to be added to the graph, and therefore the above procedure is not deterministic.
However, the final result cl ( G ) is independent of the choices.
Lemma 3.2. The closure cl ( G ) is uniquely defined for all graphs G of order νG ≥ 3.
Proof. Denote n = νG . Suppose there are two ways to close G, say
H = G + {e1 , . . . , er } and H ′ = G + { f 1 , . . . , f s } ,
where the edges are added in the given orders. Let Hi = G + {e1 , . . . , ei } and Hi′ =
′
G + { f 1 , . . . , f i }. For the initial values, we have G = H0 = H0 . Let ek = uv be the
first edge such that ek = f i for all i. Then d Hk −1 (u) + d Hk −1 (v) ≥ n, since ek ∈ Hk ,
but ek ∈ Hk−1 . By the choice of ek , we have Hk−1 ⊆ H ′ , and thus also d H ′ (u) +
/
d H ′ (v) ≥ n, which means that e = uv must be in H ′ ; a contradiction. Therefore H ⊆
H ′ . Symmetrically, we deduce that H ′ ⊆ H, and hence H ′ = H.
⊔
⊓
Theorem 3.5. Let G be a graph of order νG ≥ 3.
(i) G is hamiltonian if and only if its closure cl ( G ) is hamiltonian.
(ii) If cl ( G ) is a complete graph, then G is hamiltonian.
Proof. First, G ⊆ cl ( G ) and G spans cl ( G ), and thus if G is hamiltonian, so is cl ( G ).
In the other direction, let G = G0 , G1 , . . . , Gk = cl ( G ) be a construction sequence
of the closure of G. If cl ( G ) is hamiltonian, then so are Gk−1 , . . . , G1 and G0 by Theorem 3.4.
The Claim (ii) follows from (i), since each complete graph is hamiltonian.
⊔
⊓


34

Theorem 3.6. Let G be a graph of order νG ≥ 3. Suppose that for all nonadjacent vertices u
and v, dG (u) + dG (v) ≥ νG . Then G is hamiltonian. In particular, if δ( G ) ≥ 1 νG , then G is
2
hamiltonian.
Proof. Since dG (u) + dG (v) ≥ νG for all nonadjacent vertices, we have cl ( G ) = Kn
for n = νG , and thus G is hamiltonian. The second claim is immediate, since now
dG (u) + dG (v) ≥ νG for all u, v ∈ G whether adjacent or not.
⊔
⊓

Chvátal’s condition
The hamiltonian problem of graphs has attracted much attention, at least partly because the problem has practical significance. (Indeed, the first example where DNA
computing was applied, was the hamiltonian problem.)
There are some general improvements of the previous results of this chapter, and
quite many improvements in various special cases, where the graphs are somehow
restricted. We become satisfied by two general results.
Theorem 3.7 (C HVÁTAL (1972)). Let G be a graph with VG = {v1 , v2 , . . . , vn }, for n ≥ 3,
ordered so that d1 ≤ d2 ≤ · · · ≤ dn , for di = dG (vi ). If for every i < n/2,
di ≤ i =⇒ dn−i ≥ n − i ,

(3.1)

then G is hamiltonian.
Proof. First of all, we may suppose that G is closed, G = cl ( G ), because G is hamiltonian if and only if cl ( G ) is hamiltonian, and adding edges to G does not decrease
any of its degrees, that is, if G satisfies (3.1), so does G + e for every e. We show that,
in this case, G = Kn , and thus G is hamiltonian.
Assume on the contrary that G = Kn , and let uv ∈ G with dG (u) ≤ dG (v) be
/
such that dG (u) + dG (v) is as large as possible. Because G is closed, we must have
dG (u) + dG (v) < n, and therefore dG (u) = i < n/2. Let A = {w | vw ∈ G, w = v}.
/
By our choice, dG (w) ≤ i for all w ∈ A, and, moreover,

| A | = ( n − 1) − d G ( v ) ≥ d G ( u ) = i .
Consequently, there are at least i vertices w with dG (w) ≤ i, and so di ≤ dG (u) = i.
Similarly, for each vertex from B = {w | uw ∈ G, w = u}, dG (w) ≤ dG (v) <
/
n − dG (u) = n − i, and

| B | = ( n − 1) − d G ( u ) = ( n − 1) − i .
Also dG (u) < n − i, and thus there are at least n − i vertices w with dG (w) < n −
i. Consequently, dn−i < n − i. This contradicts the obtained bound di ≤ i and the
condition (3.1).
⊔
⊓
Note that the condition (3.1) is easily checkable for any given graph.

3.3 Matchings

35

3.3 Matchings
In matching problems we are given an availability relation between the elements of
a set. The problem is then to find a pairing of the elements so that each element is
paired (matched) uniquely with an available companion.
A special case of the matching problem is the marriage problem, which is stated
as follows. Given a set X of boys and a set Y of girls, under what condition can each
boy marry a girl who cares to marry him? This problem has many variations. One of
them is the job assignment problem, where we are given n applicants and m jobs,
and we should assign each applicant to a job he is qualified. The problem is that
an applicant may be qualified for several jobs, and a job may be suited for several
applicants.

Maximum matchings
D EFINITION . For a graph G, a subset M ⊆ EG is a matching of G, if M contains no
adjacent edges. The two ends of an edge e ∈ M are matched under M. A matching
M is a maximum matching, if for no matching M ′ , | M | < | M ′ |.
The two vertical edges on the right constitute a matching M that is not a maximum matching, although you
cannot add any edges to M to form a larger matching.
This matching is not maximum because the graph has
a matching of three edges.
D EFINITION . A matching M saturates v ∈ G, if v is an
end of an edge in M. Also, M saturates A ⊆ VG , if it
saturates every v ∈ A. If M saturates VG , then M is a
perfect matching.
It is clear that every perfect matching is maximum.
On the right the horizontal edges form a perfect matching.
D EFINITION . Let M be a matching of G. An odd path
P = e1 e2 . . . e2k+1 is M-augmented, if
• P alternates between EG M and M
(that is, e2i+1 ∈ G − M and e2i ∈ M), and
• the ends of P are not saturated.
Lemma 3.3. If G is connected with ∆( G ) ≤ 2, then G is a path or a cycle.
Proof. Exercise.

⊔
⊓

We start with a result that gives a necessary and sufficient condition for a matching to be maximum. One can use the first part of the proof to construct a maximum

3.3 Matchings

36

matching in an iterative manner starting from any matching M and from any Maugmented path.
Theorem 3.8 (B ERGE (1957)). A matching M of G is a maximum matching if and only if
there are no M-augmented paths in G.
Proof. (⇒) Let a matching M have an M-augmented path P = e1 e2 . . . e2k+1 in G.
Here e2 , e4 , . . . , e2k ∈ M, e1 , e3 , . . . , e2k+1 ∈ M. Deﬁne N ⊆ EG by
/
N = ( M {e2i | i ∈ [1, k]}) ∪ {e2i+1 | i ∈ [0, k]} .
Now, N is a matching of G, and | N | = | M | + 1. Therefore M is not a maximum
matching.
(⇐) Assume N is a maximum matching, but M is not. Hence | N | > | M |. Consider
the subgraph H = G [ M △ N ] for the symmetric difference M △ N. We have d H (v) ≤
2 for each v ∈ H, because v is an end of at most one edge in M and N. By Lemma 3.3,
each connected component A of H is either a path or a cycle.
Since no v ∈ A can be an end of two edges from N or from M, each connected
component (path or a cycle) A alternates between N and M. Now, since | N | > | M |,
there is a connected component A of H, which has more edges from N than from
M. This A cannot be a cycle, because an alternating cycle has even length, and it
⋆
thus contains equally many edges from N and M. Hence A : u − v is a path (of
→
odd length), which starts and ends with an edge from N. Because A is a connected
component of H, the ends u and v are not saturated by M, and, consequently, A is an
M-augmented path. This proves the theorem.
⊔
⊓
Example 3.5. Consider the k-cube Qk for k ≥ 1. Each maximum matching of Qk has
2k−1 edges. Indeed, the matching M = {(0u, 1u) | u ∈ B k−1 }, has 2k−1 edges, and it
is clearly perfect.

Hall’s theorem
For a subset S ⊆ VG of a graph G, denote
NG (S) = {v | uv ∈ G for some u ∈ S} .
If G is ( X, Y )-bipartite, and S ⊆ X, then NG (S) ⊆ Y.
The following result, known as the
Theorem 3.9 (H ALL (1935)). Let G be a ( X, Y )-bipartite graph. Then G contains a matching
M saturating X if and only if

|S| ≤ | NG (S)|

for all S ⊆ X.

(3.2)

3.3 Matchings

37

Proof. (⇒) Let M be a matching that saturates X. If |S| > | NG (S)| for some S ⊆ X,
then not all x ∈ S can be matched with different y ∈ NG (S).
(⇐) Let G satisfy Hall’s condition (3.2). We prove the claim by induction on | X |.
If | X | = 1, then the claim is clear. Let then | X | ≥ 2, and assume (3.2) implies the
existence of a matching that saturates every proper subset of X.
If | NG (S)| ≥ |S| + 1 for every nonempty S ⊆ X with S = X, then choose an
edge uv ∈ G with u ∈ X, and consider the induced subgraph H = G −{u, v}. For all
S ⊆ X {u}, | NH (S)| ≥ | NG (S)| − 1 ≥ |S| , and hence, by the induction hypothesis,
H contains a matching M saturating X {u}. Now M ∪ {uv} is a matching saturating
X in G, as was required.
Suppose then that there exists a nonempty subset R ⊆ X with R = X such that
| NG ( R)| = | R|. The induced subgraph H1 = G [ R ∪ NG ( R)] satisfies (3.2) (since G
does), and hence, by the induction hypothesis, H1 contains a matching M1 that saturates R (with the other ends in NG ( R)).
Also, the induced subgraph H2 = G [VG A], for A = R ∪ NG ( R), satisfies (3.2).
Indeed, if there were a subset S ⊆ X R such that | NH2 (S)| < |S|, then we would
have

| NG (S ∪ R)| = | NH2 (S)| + | NH1 ( R)| < |S| + | NG ( R)| = |S| + | R| = |S ∪ R|
(since S ∩ R = ∅), which contradicts (3.2) for G. By the induction hypothesis, H2 has
a matching M2 that saturates X R (with the other ends in Y NG ( R)). Combining
the matchings for H1 and H2 , we get a matching M1 ∪ M2 saturating X in G.
⊔
⊓
Second proof. This proof of the direction (⇐) uses Menger’s theorem. Let H be the
graph obtained from G by adding two new vertices x, y such that x is adjacent to each
v ∈ X and y is adjacent to each v ∈ Y. There exists a matching saturating X if (and
⋆
only if) the number of independent paths x − y is equal to | X |. For this, by Menger’s
→
theorem, it suffices to show that every set S that separates x and y in H has at least
|X | vertices.
Let S = A ∪ B, where A ⊆ X and B ⊆ Y. Now,
vertices in X A are not adjacent to vertices of Y B,
and hence we have NG ( X A) ⊆ B, and thus that
|X A| ≤ | NG ( X A)| ≤ | B| using the condition (3.2).
We conclude that |S| = | A| + | B| ≥ | X |.

y

x
XA
A

YB
B

⊔
⊓

Corollary 3.1 (F ROBENIUS (1917)). If G is a k-regular bipartite graph with k > 0, then G
has a perfect matching.
Proof. Let G be k-regular ( X, Y )-bipartite graph. By regularity, k · | X | = ε G = k · |Y |,
and hence | X | = |Y |. Let S ⊆ X. Denote by E1 the set of the edges with an end in
S, and by E2 the set of the edges with an end in NG (S). Clearly, E1 ⊆ E2 . Therefore,
k · | NG (S)| = | E2 | ≥ | E1 | = k · |S|, and so | NG (S)| ≥ |S|. By Theorem 3.9, G has a
matching that saturates X. Since | X | = |Y |, this matching is necessarily perfect.
⊔
⊓

3.3 Matchings

38

Applications of Hall’s theorem
D EFINITION . Let S = {S1 , S2 , . . . , Sm } be a family of finite nonempty subsets of a set
S. (Si need not be distinct.) A transversal (or a system of distinct representatives) of
S is a subset T ⊆ S of m distinct elements one from each Si .
As an example, let S = [1, 6], and let S1 = S2 = {1, 2}, S3 = {2, 3} and S4 =
{1, 4, 5, 6}. For S = {S1 , S2 , S3 , S4 }, the set T = {1, 2, 3, 4} is a transversal. If we add
the set S5 = {2, 3} to S , then it is impossible to find a transversal for this new family.
The connection of transversals to the Marriage Theorem is as follows. Let S = Y
and X = [1, m]. Form an ( X, Y )-bipartite graph G such that there is an edge (i, s)
if and only if s ∈ Si . The possible transversals T of S are then obtained from the
matchings M saturating X in G by taking the ends in Y of the edges of M.
Corollary 3.2. Let S be a family of finite nonempty sets. Then S has a transversal if and only
if the union of any k of the subsets Si of S contains at least k elements.
Example 3.6. An m × n latin rectangle is an m × n integer matrix M with entries
Mij ∈ [1, n] such that the entries in the same row and in the same column are different.
Moreover, if m = n, then M is a latin square. Note that in a m × n latin rectangle M,
we always have that m ≤ n.
We show the following: Let M be an m × n latin rectangle (with m < n). Then M can
be extended to a latin square by the addition of n − m new rows.
The claim follows when we show that M can be extended to an (m + 1) × n latin
rectangle. Let Ai ⊆ [1, n] be the set of those elements that do not occur in the i-th
column of M. Clearly, | Ai | = n − m for each i, and hence ∑i∈ I | Ai | = | I |(n − m) for
all subsets I ⊆ [1, n]. Now |∪i∈ I Ai | ≥ | I |, since otherwise at least one element from
the union would be in more than n − m of the sets Ai with i ∈ I. However, each row
has all the n elements, and therefore each i is missing from exactly n − m columns. By
Marriage Theorem, the family { A1 , A2 , . . . , An } has a transversal, and this transversal
can be added as a new row to M. This proves the claim.

Tutte’s theorem
The next theorem is a classic characterization of perfect matchings.
D EFINITION . A connected component of a graph G is said to be odd (even), if it has
an odd (even) number of vertices. Denote by codd ( G ) the number of odd connected
components in G.
Denote by m( G ) be the number of edges in a maximum matching of a graph G.

3.3 Matchings

39

Theorem 3.10 (Tutte-Berge Formula). Each maximum matching of a graph G has
m( G ) = min

S ⊆VG

νG + |S| − codd ( G −S)
2

(3.3)

elements.
Note that the condition in (ii) includes the case, where S = ∅.
Proof. We prove the result for connected graphs. The result then follows for disconnected graphs by adding the formulas for the connected components.
We observe ﬁrst that ≤ holds in (3.3), since, for all S ⊆ VG ,
ν + |S| − codd ( G −S)
|VG S| − codd ( G −S)
= G
.
2
2
Indeed, each odd component of G −S must have at least one unsaturated vertex.
The proof proceeds by induction on νG . If νG = 1, then the claim is trivial. Suppose
that νG ≥ 2.
Assume ﬁrst that there exists a vertex v ∈ G such that v is saturated by all
maximum matchings. Then m( G −v) = m( G ) − 1. For a subset S′ ⊆ G −v, denote
S = S′ ∪ {v}. By the induction hypothesis, for all S′ ⊆ G −v,
m ( G ) ≤ |S | + m ( G −S ) ≤ |S | +

1
(νG − 1) + |S′ | − codd ( G −(S′ ∪ {v}))
2
1
= ((νG + |S| − codd ( G −S))) − 1.
2

m( G ) − 1 ≥

The claim follows from this.
Suppose then that for each vertex v, there is a maximum matching that does not
saturate v. We claim that m( G ) = (νG − 1)/2. Suppose to the contrary, and let M be a
maximum matching having two different unsaturated vertices u and v, and choose M
so that the distance dG (u, v) is as small as possible. Now dG (u, v) ≥ 2, since otherwise
uv ∈ G could be added to M, contradicting the maximality of M. Let w be an inter⋆
mediate vertex on a shortest path u − v. By assumption, there exists a maximum
→
matching N that does not saturate w. We can choose N such that the intersection
M ∩ N is maximal. Since dG (u, w) < dG (u, v) and dG (w, v) < dG (u, v), N saturates
both u and v. The (maximum) matchings N and M leave equally many vertices unsaturated, and hence there exists another vertex x = w saturated by M but which is
unsaturated by N. Let e = xy ∈ M. If y is also unsaturated by N, then N ∪ {e} is a
matching, contradicting maximality of N. It also follows that y = w. Therefore there
exists an edge e′ = yz in N, where z = x. But now N ′ = N ∪ {e} {e′ } is a maximum
matching that does not saturate w. However, N ∩ M ⊂ N ′ ∩ M contradicts the choice
of N. Therefore, every maximum matching leaves exactly one vertex unsaturated, i.e.,
m( G ) = (νG − 1)/2.
In this case, for S = ∅, the right hand side of (3.3) gets value (νG − 1)/2, and
hence, by the beginning of the proof, this must be the minimum of the right hand
side.
⊔
⊓

3.3 Matchings

40

For perfect matchings we have the following corollary, since for a perfect matching
we have m( G ) = (1/2)νG .
Theorem 3.11 (T UTTE (1947)). Let G be a nontrivial graph. The following are equivalent.
(i) G has a perfect matching.
(ii) For every proper subset S ⊂ VG , codd ( G −S) ≤ |S|.
Tutte’s theorem does not provide a good algorithm for constructing a perfect
matching, because the theorem requires ‘too many cases’. Its applications are mainly
in the proofs of other results that are related to matchings. There is a good algorithm
due to E DMONDS (1965), which uses ‘blossom shrinkings’, but this algorithm is somewhat involved.
Example 3.7. The simplest connected graph that has no perfect matching is the path
P3 . Here removing the middle vertex creates two odd components.
The next 3-regular graph (known as the Sylvester
graph) does not have a perfect matching, because removing the black vertex results in a graph with three
odd connected components. This graph is the smallest
regular graph with an odd degree that has no perfect
matching.
Using Theorem 3.11 we can give a short proof of P ETERSEN’s result for 3-regular
graphs (1891).
Theorem 3.12 (P ETERSEN (1891)). If G is a bridgeless 3-regular graph, then it has a perfect
matching.
Proof. Let S be a proper subset of VG , and let Gi , i ∈ [1, t], be the odd connected
components of G −S. Denote by mi the number of edges with one end in Gi and the
other in S. Since G is 3-regular,

∑

dG (v) = 3 · νGi

v ∈ Gi

and

∑ d G ( v) = 3 · |S | .
v∈S

The ﬁrst of these implies that
mi =

∑

d G ( v ) − 2 · ε Gi

v ∈ Gi

is odd. Furthermore, mi = 1, because G has no bridges, and therefore mi ≥ 3. Hence
the number of odd connected components of G −S satisﬁes
t≤

1
1 t
∑ m i ≤ 3 ∑ d G ( v) = |S | ,
3 i=1
v∈S

and so, by Theorem 3.11, G has a perfect matching.

⊔
⊓

3.3 Matchings

41

Stable Marriages
D EFINITION . Consider a bipartite graph G with a bipartition ( X, Y ) of the vertex set.
In addition, each vertex x ∈ G supplies an order of preferences of the vertices of
NG ( x). We write u < x v, if x prefers v to u. (Here u, v ∈ Y, if x ∈ X, and u, v ∈ X, if
x ∈ Y.) A matching M of G is said to be stable, if for each unmatched pair xy ∈ M
/
(with x ∈ X and y ∈ Y), it is not the case that x and y prefer each other better than
their matched companions:
xv ∈ M and y < x v, or uy ∈ M and x <y u.
We omit the proof of the next theorem.
Theorem 3.13. For bipartite graphs G, a stable matching exists for all lists of preferences.
Example 3.8. That was the good news. There is a catch, of
course. A stable matching need not saturate X and Y. For
instance, the graph on the right does have a perfect matching (of 4 edges).
Suppose the preferences are the following:

4

8

3

7

2

6

1

5

1: 5

2: 6 < 8 < 7

3: 8 < 5

4: 7 < 5

5: 4 < 1 < 3

6: 2

7: 2 < 4

8: 3 < 2

Then there is no stable matchings of four edges. A stable matching of G is the following: M = {28, 35, 47}, which leaves 1 and 6 unmatched. (You should check that there
is no stable matching containing the edges 15 and 26.)
Theorem 3.14. Let G = Kn,n be a complete bipartite graph. Then G has a perfect and stable
matching for all lists of preferences.
Proof. Let the bipartition be ( X, Y ). The algorithm by G ALE
works as follows.

AND

S HAPLEY (1962)

Procedure.
Set M0 = ∅, and P( x) = ∅ for all x ∈ X.
Then iterate the following process until all vertices are saturated:
Choose a vertex x ∈ X that is unsaturated in Mi−1 . Let y ∈ Y be the
most preferred vertex for x such that y ∈ P( x).
/
(1) Add y to P( x).
(2) If y is not saturated, then set Mi = Mi−1 ∪ { xy}.
(3) If zy ∈ Mi−1 and z <y x, then set Mi = ( Mi−1 {zy}) ∪ { xy}.
First of all, the procedure terminates, since a vertex x ∈ X takes part in the iteration at most n times (once for each y ∈ Y). The ﬁnal outcome, say M = Mt , is a
perfect matching, since the iteration continues until there are no unsaturated vertices
x ∈ X.

3.3 Matchings

42

Also, the matching M = Mt is stable. Note ﬁrst that, by (3), if xy ∈ Mi and zy ∈ M j
for some x = z and i < j, then x <y z. Assume the that xy ∈ M, but y < x z for some
z ∈ Y. Then xy is added to the matching at some step, xy ∈ Mi , which means that
z ∈ P( x) at this step (otherwise x would have ‘proposed’ z). Hence x took part in the
iteration at an earlier step Mk , k < i (where z was put to the list P( x), but xz was not
added). Thus, for some u ∈ X, uz ∈ Mk−1 and x <z u, and so in M the vertex z is
matched to some w with x <z w.
Similarly, if x <y v for some v ∈ X, then y <v z for the vertex z ∈ Y such that
vz ∈ M.
⊔
⊓

4
Colourings
4.1 Edge colourings
Colourings of edges and vertices of a graph G are useful, when one is interested in
classifying relations between objects.
There are two sides of colourings. In the general case, a graph G with a colouring
α is given, and we study the properties of this pair G α = ( G, α). This is the situation,
e.g., in transportation networks with bus and train links, where the colour (buss, train)
of an edge tells the nature of a link.
In the chromatic theory, G is first given and then we search for a colouring that the
satisfies required properties. One of the important properties of colourings is ‘properness’. In a proper colouring adjacent edges or vertices are coloured differently.

Edge chromatic number
D EFINITION . A k-edge colouring α : EG → [1, k] of a graph G is an assignment of k
colours to its edges. We write G α to indicate that G has the edge colouring α.
A vertex v ∈ G and a colour i ∈ [1, k] are incident with each other, if α(vu) = i for
some vu ∈ G. If v ∈ G is not incident with a colour i, then i is available for v.
The colouring α is proper, if no two adjacent edges obtain the same colour: α(e1 ) =
α(e2 ) for adjacent e1 and e2 .
The edge chromatic number χ′ ( G ) of G is defined as
χ′ ( G ) = min{k | there exists a proper k-edge colouring of G } .
A k-edge colouring α can be thought of as a partition {E1 , E2 , . . . , Ek } of EG , where
Ei = {e | α(e) = i }. Note that it is possible that Ei = ∅ for some i. We adopt a
simplified notation
G α [ i 1 , i 2 , . . . , i t ] = G [ Ei1 ∪ Ei2 ∪ · · · ∪ Ei t ]
for the subgraph of G consisting of those edges that have a colour i1 , i2 , . . . , or it . That
is, the edges having other colours are removed.
Lemma 4.1. Each colour set Ei in a proper k-edge colouring is a matching. Moreover, for each
graph G, ∆( G ) ≤ χ′ ( G ) ≤ ε G .
Proof. This is clear.

⊔
⊓

4.1 Edge colourings

44

Example 4.1. The three numbers in Lemma 4.1 can be equal. This happens, for instance, when G = K1,n is a star. But often the inequalities are strict.

A star, and a graph with χ′ ( G ) = 4.

Optimal colourings
We show that for bipartite graphs the lower bound is always optimal: χ′ ( G ) = ∆( G ).
Lemma 4.2. Let G be a connected graph that is not an odd cycle. Then there exists a 2-edge
colouring (that need not be proper), in which both colours are incident with each vertex v with
dG (v) ≥ 2.
Proof. Assume that G is nontrivial; otherwise, the claim is trivial.
(1) Suppose first that G is eulerian. If G is an even cycle, then a 2-edge colouring
exists as required. Otherwise, since now dG (v) is even for all v, G has a vertex v1 with
dG (v1 ) ≥ 4. Let e1 e2 . . . et be an Euler tour of G, where ei = vi vi+1 (and vt+1 = v1 ).
Define
1, if i is odd ,
α( ei ) =
2, if i is even .
Hence the ends of the edges ei for i ∈ [2, t − 1] are incident with both colours. All
vertices are among these ends. The condition dG (v1 ) ≥ 4 guarantees this for v1 . Hence
the claim holds in the eulerian case.
(2) Suppose then that G is not eulerian. We define a new graph G0 by adding a
vertex v0 to G and connecting v0 to each v ∈ G of odd degree.
In G0 every vertex has even degree including v0 (by
the handshaking lemma), and hence G0 is eulerian. Let
e0 e1 . . . et be an eulerian tour of G0 , where ei = vi vi+1 .
By the previous case, there is a required colouring α of
G0 as above. Now, α restricted to EG is a colouring of G
as required by the claim, since each vertex vi with odd
degree dG (vi ) ≥ 3 is entered and departed at least once
in the tour by an edge of the original graph G: ei−1 ei .

2
v0

D EFINITION . For a k-edge colouring α of G, let
cα (v) = |{i | v is incident with i ∈ [1, k]}| .

2

1
2

1

1

⊔
⊓

4.1 Edge colourings

45

A k-edge colouring β is an improvement of α, if

∑ c β ( v) > ∑ cα ( v) .
v∈ G

v∈ G

Also, α is optimal, if it cannot be improved.
Notice that we always have cα (v) ≤ dG (v), and if α is proper, then cα (v) = dG (v),
and in this case α is optimal. Thus an improvement of a colouring is a change towards
a proper colouring. Note also that a graph G always has an optimal k-edge colouring,
but it need not have any proper k-edge colourings.
The next lemma is obvious.
Lemma 4.3. An edge colouring α of G is proper if and only if cα (v) = dG (v) for all vertices
v ∈ G.
Lemma 4.4. Let α be an optimal k-edge colouring of G, and let v ∈ G. Suppose that the
colour i is available for v, and the colour j is incident with v at least twice. Then the connected
component H of G α [i, j] that contains v, is an odd cycle.
Proof. Suppose the connected component H is not an odd cycle. By Lemma 4.2, H
has a 2-edge colouring γ : E H → {i, j}, in which both i and j are incident with each
vertex x with d H ( x) ≥ 2. (We have renamed the colours 1 and 2 to i and j.) We obtain
a recolouring β of G as follows:
β( e) =

γ ( e ),
α ( e ),

if e ∈ H ,
if e ∈ H .
/

Since d H (v) ≥ 2 (by the assumption on the colour j) and in β both colours i and j
are now incident with v, c β (v) = cα (v) + 1. Furthermore, by the construction of β,
we have c β (u) ≥ cα (u) for all u = v. Therefore ∑u∈ G c β (u) > ∑u∈ G cα (u), which
contradicts the optimality of α. Hence H is an odd cycle.
⊔
⊓
Theorem 4.1 (K ÖNIG (1916)). If G is bipartite, then χ′ ( G ) = ∆( G ).
Proof. Let α be an optimal ∆-edge colouring of a bipartite G, where ∆ = ∆( G ). If
there were a v ∈ G with cα (v) < dG (v), then by Lemma 4.4, G would contain an odd
cycle. But a bipartite graph does not contain such cycles. Therefore, for all vertices v,
cα (v) = dG (v). By Lemma 4.3, α is a proper colouring, and ∆ = χ′ ( G ) as required. ⊔
⊓

Vizing’s theorem
In general we can have χ′ ( G ) > ∆( G ) as one of our examples did show. The following
important theorem, due to V IZING, shows that the edge chromatic number of a graph
G misses ∆( G ) by at most one colour.
Theorem 4.2 (V IZING (1964)). For any graph G, ∆( G ) ≤ χ′ ( G ) ≤ ∆( G ) + 1.
Proof. Let ∆ = ∆( G ). We need only to show that χ′ ( G ) ≤ ∆ + 1. Suppose on the
contrary that χ′ ( G ) > ∆ + 1, and let α be an optimal (∆ + 1)-edge colouring of G.

4.1 Edge colourings

46

We have (trivially) dG (u) < ∆ + 1 < χ′ ( G ) for all u ∈ G, and so
Claim 1. For each u ∈ G, there exists an available colour b(u) for u.
Moreover, by the counter hypothesis, α is not a proper colouring, and hence there
exists a v ∈ G with cα (v) < dG (v), and hence a colour i1 that is incident with v at least
twice, say
α(vu1 ) = i1 = α(vx) .
(4.1)
Claim 2. There is a sequence of vertices u1 , u2 , . . . such that
α(vu j ) = i j and i j+1 = b(u j ) .
Indeed, let u1 be as in (4.1). Assume we have already found the vertices u1 , . . . , u j ,
with j ≥ 1, such that the claim holds for these. Suppose, contrary to the claim, that v
is not incident with b(u j ) = i j+1 .
We can recolour the edges vuℓ by iℓ+1 for ℓ ∈ [1, j], and
obtain in this way an improvement of α. Here v gains a
new colour i j+1 . Also, each uℓ gains a new colour iℓ+1
(and may loose the colour iℓ ). Therefore, for each uℓ either its number of colours remains the same or it increases by one. This contradicts the optimality of α, and
proves Claim 2.

ur

u r −1
.
.
.

..
.
i r = i t +1

i r −1

v
u2

i1

Indeed, c β (v) = cα (v) and c β (u) ≥ cα (u) for all u, since
each u j (1 ≤ j ≤ r − 1) gains a new colour ji+1 although
it may loose one of its old colours.
Let then the colouring γ be obtained from β by recolouring the edges vu j by i j+1 for r ≤ j ≤ t. Now,
vut is recoloured by ir = it+1 .

ur

u r −1
.
.
.

..
.
i r = i t +1

ir

v
u2

i2
u1

ut

i1
x
ur

u r −1

..

.
.
.

i r +1

ir

v
u2

it

i3

Claim 4. γ is an optimal (∆ + 1)-edge colouring of G.
Indeed, the fact ir = it+1 ensures that ir is a new colour
incident with ut , and thus that cγ (ut ) ≥ c β (ut ). For all
other vertices, cγ (u) ≥ c β (u) follows as for β.

i1
x

Let t be the smallest index such that for some r < t,
it+1 = ir . Such an index t exists, because dG (v) is ﬁnite.

Claim 3. β is an optimal (∆ + 1)-edge colouring of G.

ut

i2
u1

Let β be a recolouring of G such that for 1 ≤ j ≤ r − 1,
β(vu j ) = i j+1 , and for all other edges e, β(e) = α(e).

it

i3
i2
u1

i1
x

ir

.

ut

4.2 Ramsey Theory

47

By Claim 1, there is a colour i0 = b(v) that is available for v. By Lemma 4.4, the
connected components H1 of G β [i0 , ir ] and H2 of G γ [i0 , ir ] containing the vertex v are
cycles, that is, H1 is a cycle (vur −1 ) P1 (ur v) and H2 is a cycle (vur −1 ) P2 (ut v), where
⋆
⋆
both P1 : ur −1 − ur and P2 : ur −1 − ut are paths. However, the edges of P1 and P2
→
→
have the same colours with respect to β and γ (either i0 or ir ). This is not possible,
since P1 ends in ur while P2 ends in a different vertex ut . This contradiction proves
the theorem.
⊔
⊓
Example 4.2. We show that χ′ ( G ) = 4 for the Petersen graph. Indeed, by Vizing’
theorem, χ′ ( G ) = 3 or 4. Suppose 3 colours suffice. Let C : v1 − . . . − v5 − v1 be
→
→
→
the outer cycle and C ′ : u1 − . . . − u5 − u1 the inner cycle of G such that vi ui ∈ EG
→
→
→
for all i.
Observe that every vertex is adjacent to all colours 1, 2, 3. Now C uses one colour
(say 1) once and the other two twice. This can be done uniquely (up to permutations):
1

2

3

2

3

v1 − v2 − v3 − v4 − v5 − v1 .
→
→
→
→
→
2

3

1

1

1

Hence v1 − u1 , v2 − u2 , v3 − u3 , v4 − u4 , v5 − u5 . However, this means that 1
→
→
→
→
→
′ . Since C ′ needs three colours, the claim follows.
cannot be a colour of any edge in C
Edge Colouring Problem. Vizing’s theorem (nor its present proof) does not offer any
characterization for the graphs, for which χ′ ( G ) = ∆( G ) + 1. In fact, it is one of the
famous open problems of graph theory to find such a characterization. The answer
is known (only) for some special classes of graphs. By H OLYER (1981), the problem
whether χ′ ( G ) is ∆( G ) or ∆( G ) + 1 is NP-complete.
The proof of Vizing’s theorem can be used to obtain a proper colouring of G with
at most ∆( G ) + 1 colours, when the word ‘optimal’ is forgotten: colour first the edges
as well as you can (if nothing better, then arbitrarily in two colours), and use the proof
iteratively to improve the colouring until no improvement is possible – then the proof
says that the result is a proper colouring.

4.2 Ramsey Theory
In general, Ramsey theory studies unavoidable patterns in combinatorics. We consider an instance of this theory mainly for edge colourings (that need not be proper).
A typical example of a Ramsey property is the following: given 6 persons each pair of
whom are either friends or enemies, there are then 3 persons who are mutual friends
or mutual enemies. In graph theoretic terms this means that each colouring of the
edges of K6 with 2 colours results in a monochromatic triangle.

4.2 Ramsey Theory

48

Turan’s theorem for complete graphs
We shall first consider the problem of finding a general condition for K p to appear
in a graph. It is clear that every graph contains K1 , and that every nondiscrete graph
contains K2 .
D EFINITION . A complete p-partite graph G consists of p discrete and disjoint induced subgraphs
G1 , G2 , . . . , G p ⊆ G, where uv ∈ G if and only if u and
v belong to different parts, Gi and Gj with i = j.
Note that a complete p-partite graph is completely determined by its discrete parts Gi , i ∈ [1, p].
Let p ≥ 3, and let H = Hn,p be the complete ( p − 1)-partite graph of order n =
t( p − 1) + r, where r ∈ [1, p − 1] and t ≥ 0, such that there are r parts H1 , . . . , Hr of
order t + 1 and p − 1 − r parts Hr +1 , . . . , H p−1 of order t (when t > 0). (Here r is the
positive residue of n modulo ( p − 1), and is thus determined by n and p.)
By its definition, K p
H. One can compute that the number ε H of edges of H is
equal to
r
p−2 2 r
n −
1−
.
(4.2)
T (n, p) =
2( p − 1)
2
p−1
The next result shows that the above bound T (n, p) is optimal.
Theorem 4.3 (T URÁN (1941)). If a graph G of order n has ε G > T (n, p) edges, then G
contains a complete subgraph K p .
Proof. Let n = ( p − 1)t + r for 1 ≤ r ≤ p − 1 and t ≥ 0. We prove the claim by
induction on t. If t = 0, then T (n, p) = n(n − 1)/2, and there is nothing to prove.
Suppose then that t ≥ 1, and let G be a graph of order n such that ε G is maximum
subject to the condition K p G.
Now G contains a complete subgraph G [ A] = K p−1, since adding any one edge to
G results in a K p , and p − 1 vertices of this K p induce a subgraph K p−1 ⊆ G.
Each v ∈ A is adjacent to at most p − 2 vertices of A; otherwise G [ A ∪ {v}] = K p .
/
Also, K p G − A, and νG− A = n − p + 1. Because n − p + 1 = (t − 1)( p − 1) + r, we
can apply the induction hypothesis to obtain ε G− A ≤ T (n − p + 1, p). Now
ε G ≤ T (n − p + 1, p) + (n − p + 1)( p − 2) +

( p − 1)( p − 2)
= T (n, p) ,
2

which proves the claim.

⊔
⊓

When Theorem 4.3 is applied to triangles K3 , we have the following interesting
case.
Corollary 4.1 ( M ANTEL (1907)). If a graph G has ε G >
triangle K3 .

1 2
4 νG

edges, then G contains a

4.2 Ramsey Theory

49

Ramsey’s theorem
D EFINITION . Let α be an edge colouring of G. A subgraph H ⊆ G is said to be (i-)
monochromatic, if all edges of H have the same colour i.
The following theorem is one of the jewels of combinatorics.
Theorem 4.4 (R AMSEY (1930)). Let p, q ≥ 2 be any integers. Then there exists a (smallest)
integer R( p, q) such that for all n ≥ R( p, q), any 2-edge colouring of Kn → [1, 2] contains a
1-monochromatic K p or a 2-monochromatic Kq .
Before proving this, we give an equivalent statement. Recall that a subset X ⊆ VG
is stable, if G [ X ] is a discrete graph.
Theorem 4.5. Let p, q ≥ 2 be any integers. Then there exists a (smallest) integer R( p, q)
such that for all n ≥ R( p, q), any graph G of order n contains a complete subgraph of order
p or a stable set of order q.
Be patient, this will follow from Theorem 4.6. The number R( p, q) is known as the
Ramsey number for p and q.
It is clear that R( p, 2) = p and R(2, q) = q.
Theorems 4.4 and 4.5 follow from the next result which shows (inductively) that
an upper bound exists for the Ramsey numbers R( p, q).
Theorem 4.6 (E RDÖS and S ZEKERES (1935)). The Ramsey number R( p, q) exists for all
p, q ≥ 2, and
R( p, q) ≤ R( p, q − 1) + R( p − 1, q) .
Proof. We use induction on p + q. It is clear that R( p, q) exists for p = 2 or q = 2, and
it is thus exists for p + q ≤ 5.
It is now sufficient to show that if G is a graph of order R( p, q − 1) + R( p − 1, q),
then it has a complete subgraph of order p or a stable subset of order q.
Let v ∈ G, and denote by A = VG ( NG (v) ∪ {v}) the set of vertices that are not
adjacent to v. Since G has R( p, q − 1) + R( p − 1, q) − 1 vertices different from v, either
| NG (v)| ≥ R( p − 1, q) or | A| ≥ R( p, q − 1) (or both).
Assume first that | NG (v)| ≥ R( p − 1, q). By the definition of Ramsey numbers,
G [ NG (v)] contains a complete subgraph B of order p − 1 or a stable subset S of order
q. In the first case, B ∪ {v} induces a complete subgraph K p in G, and in the second
case the same stable set of order q is good for G.
If | A| ≥ R( p, q − 1), then G [ A] contains a complete subgraph of order p or a stable
subset S of order q − 1. In the first case, the same complete subgraph of order p is
good for G, and in the second case, S ∪ {v} is a stable subset of G of q vertices. This
proves the claim.
⊔
⊓

4.2 Ramsey Theory

50

A concrete upper bound is given in the following result.
Theorem 4.7 ( E RDÖS and S ZEKERES (1935)). For all p, q ≥ 2,
R( p, q) ≤

p+q−2
.
p−1

Proof. For p = 2 or q = 2, the claim is clear. We use induction on p + q for the general
statement. Assume that p, q ≥ 3. By Theorem 4.6 and the induction hypothesis,
R( p, q) ≤ R( p, q − 1) + R( p − 1, q)

≤

p+q−3
+
p−1

p+q−3
p−2

=

p+q−2
,
p−1

which is what we wanted.

⊔
⊓

In the table below we give some known values and estimates for the Ramsey numbers R( p, q). As can be read from the table1 , not so much is known about these numbers.
pq 3 4
5
6
7
8
9
10
3
6 9
14
18
23
28
36
40-43
4
9 18 25 35-41 49-61 55-84 69-115 80-149
5
14 25 43-49 58-87 80-143 95-216 121-316 141-442
The ﬁrst unknown R( p, p) (where p = q) is for p = 5. It has been veriﬁed that
43 ≤ R(5, 5) ≤ 49, but to determine the exact value is an open problem.

Generalizations
Theorem 4.4 can be generalized as follows.
Theorem 4.8. Let qi ≥ 2 be integers for i ∈ [1, k] with k ≥ 2. Then there exists an integer R = R(q1 , q2 , . . . , qk ) such that for all n ≥ R, any k-edge colouring of Kn has an
i-monochromatic Kqi for some i.
Proof. The proof is by induction on k. The case k = 2 is treated in Theorem 4.4. For
k > 2, we show that R(q1 , . . . , qk ) ≤ R(q1 , . . . , qk−2 , p), where p = R(qk−1 , qk ).
Let n = R(q1 , . . . , qk−2 , p), and let α : EKn → [1, k] be an edge colouring. Let
β : EKn → [1, k − 1] be obtained from α by identifying the colours k − 1 and k:
β( e) =
1

α(e) if α(e) < k − 1 ,
k − 1 if α(e) = k − 1 or k .

S.P. R ADZISZOWSKI, Small Ramsey numbers, Electronic J. of Combin., 2000 on the Web

4.2 Ramsey Theory

51
β

By the induction hypothesis, Kn has an i-monochromatic Kqi for some 1 ≤ i ≤ k − 2
β

α
(and we are done, since this subgraph is monochromatic in Kn ) or Kn has a (k − 1)β = K . In the latter case, by Theorem 4.4, H α and thus
monochromatic subgraph H
p
α
Kn has a (k − 1)-monochromatic or a k-monochromatic subgraph, and this proves the
claim.
⊔
⊓

Since for each graph H, H ⊆ Km for m = νH , we have
Corollary 4.2. Let k ≥ 2 and H1 , H2 , . . . , Hk be arbitrary graphs. Then there exists an integer R( H1 , H2 , . . . , Hk ) such that for all complete graphs Kn with n ≥ R( H1 , H2 , . . . , Hk )
α
and for all k-edge colourings α of Kn , Kn contains an i-monochromatic subgraph Hi for some
i.
This generalization is trivial from Theorem 4.8. However, the generalized Ramsey
numbers R( H1 , H2 , . . . , Hk ) can be much smaller than their counter parts (for complete graphs) in Theorem 4.8.
Example 4.3. We leave the following statement as an exercise: If T is a tree of order
m, then
R( T, Kn ) = (m − 1)(n − 1) + 1 ,
that is, any graph G of order at least R( T, Kn ) contains a subgraph isomorphic to T,
or the complement of G contains a complete subgraph Kn .

Examples of Ramsey numbers∗
Some exact values are known in Corollary 4.2, even in more general cases, for some
dear graphs (see R ADZISZOWSKI’s survey). Below we list some of these results for
cases, where the graphs are equal. To this end, let
Rk ( G ) = R( G, G, . . . , G )

(k times G ).

The best known lower bound of R2 ( G ) for connected graphs was obtained by B URR
AND E RDÖS (1976),
R2 ( G ) ≥

4νG − 1
3

( G connected).

Here is a list of some special cases:
n
R2 ( Pn ) = n +
− 1,
2

6
if n = 3 or n = 4,

R2 (Cn ) = 2n − 1
if n ≥ 5 and n odd ,


3n/2 − 1 if n ≥ 6 and n even,
R2 (K1,n ) =

2n − 1
2n

R2 (K2,3 ) = 10,

if n is even,
if n is odd,

R2 (K3,3 ) = 18.

4.2 Ramsey Theory

52

The values R2 (K2,n ) are known for n ≤ 16, and in general, R2 (K2,n ) ≤ 4n − 2. The
value R2 (K2,17 ) is either 65 or 66.
Let Wn denote the wheel on n vertices. It is a cycle Cn−1 , where a vertex v with
degree n − 1 is attached. Note that W4 = K4 . Then R2 (W5 ) = 15 and R2 (W6 ) = 17.
For three colours, much less is known. In fact, the only nontrivial result for complete graphs is: R3 (K3 ) = 17. Also, 128 ≤ R3 (K4 ) ≤ 235, and 385 ≤ R3 (K5 ), but
no nontrivial upper bound is known for R3 (K5 ). For the square C4 , we know that
R3 (C4 ) = 11.
Needless to say that no exact values are known for Rk (Kn ) for k ≥ 4 and n ≥ 3.
It follows from Theorem 4.4 that for any complete Kn , there exists a graph G
(well, any sufficiently large complete graph) such that any 2-edge colouring of G
has a monochromatic (induced) subgraph Kn . Note, however, that in Corollary 4.2
the monochromatic subgraph Hi is not required to be induced.
The following impressive theorem improves the results we have mentioned in this
chapter and it has a difficult proof.
Theorem 4.9 (D EUBER, E RDÖS, H AJNAL, P ÓSA, and R ÖDL (around 1973)). Let H be
any graph. Then there exists a graph G such that any 2-edge colouring of G has an monochromatic induced subgraph H.
Example 4.4. As an application of Ramsey’s theorem, we shortly describe Schur’s
theorem. For this, consider the partition {1, 4, 10, 13}, {2, 3, 11, 12}, {5, 6, 7, 8, 9} of
the set N13 = [1, 13]. We observe that in no partition class there are three integers
such that x + y = z. However, if you try to partition N14 into three classes, then you
are bound to find a class, where x + y = z has a solution.
S CHUR (1916) solved this problem in a general setting. The following gives a short
proof using Ramsey’s theorem.
For each n ≥ 1, there exists an integer S(n) such that any partition S1 , . . . , Sn of N S(n) has
a class Si containing two integers x, y such that x + y ∈ Si .
Indeed, let S(n) = R(3, 3, . . . , 3), where 3 occurs n times, and let K be a complete
on N S(n) . For a partition S1 , . . . , Sn of N S(n) , define an edge colouring α of K by
α(ij) = k, if |i − j| ∈ Sk .
By Theorem 4.8, K α has a monochromatic triangle, that is, there are three vertices
i, j, t such that 1 ≤ i < j < t ≤ S(n) with t − j, j − i, t − i ∈ Sk for some k. But
(t − j) + ( j − i) = t − i proves the claim.
There are quite many interesting corollaries to Ramsey’s theorem in various parts
of mathematics including not only graph theory, but also, e.g., geometry and algebra,
see
R.L. G RAHAM , B.L. R OTHSCHILD AND J.L. S PENCER, “Ramsey Theory”, Wiley, (2nd
ed.) 1990.

4.3 Vertex colourings

53

The vertices of a graph G can also be classified using colourings. These colourings
tell that certain vertices have a common property (or that they are similar in some
respect), if they share the same colour. In this chapter, we shall concentrate on proper
vertex colourings, where adjacent vertices get different colours.

The chromatic number
D EFINITION . A k-colouring (or a k-vertex colouring) of a graph G is a mapping
α : VG → [1, k]. The colouring α is proper, if adjacent vertices obtain a different colour:
for all uv ∈ G, we have α(u) = α(v). A colour i ∈ [1, k] is said to be available for a
vertex v, if no neighbour of v is coloured by i.
A graph G is k-colourable, if there is a proper k-colouring for G. The (vertex)
chromatic number χ( G ) of G is defines as
χ( G ) = min{k | there exists a proper k-colouring of G } .
If χ( G ) = k, then G is k-chromatic.
Each proper vertex colouring α : VG → [1, k] provides a partition {V1 , V2 , . . . , Vk }
of the vertex set VG , where Vi = {v | α(v) = i }.
Example 4.5. The graph on the right, which is often called a
wheel (of order 7), is 3-chromatic.
By the definitions, a graph G is 2-colourable if and only if it
is bipartite.
Again, the ‘names’ of the colours are immaterial:
Lemma 4.5. Let α be a proper k-colouring of G, and let π be any permutation of the colours.
Then the colouring β = πα is a proper k-colouring of G.
Proof. Indeed, if α : VG → [1, k] is proper, and if π : [1, k] → [1, k] is a bijection, then
uv ∈ G implies that α(u) = α(v), and hence also that πα(u) = πα(v). It follows that
πα is a proper colouring.
⊔
⊓
Example 4.6. A graph is triangle-free, if it has no subgraphs isomorphic to K3 . We
show that there are triangle-free graphs with arbitrarily large chromatic numbers.
The following construction is due to G RÖTZEL: Let G be any triangle-free graph
with VG = {v1 , v2 , . . . , vn }. Let G t be a new graph obtained by adding n + 1 new
vertices v and u1 , u2 , . . . , un such that G t has all the edges of G plus the edges ui v and
ui x for all x ∈ N (vi ) and for all i ∈ [1, n].


54

Claim. G t is triangle-free and it is k + 1-chromatic
Indeed, let U = {u1 , . . . , un }. We show first that G t is triangle-free. Now, U is
stable, and so a triangle contains at most (and thus exactly) one vertex ui ∈ U. If
{ui , v j , vk } induces a triangle, so does {vi , v j , vk } by the definition of G t , but the latter
triangle is already in G; a contradiction.
For the chromatic number we notice first that χ( G t ) ≤ (k + 1). If α is a proper
k-colouring of G, extend it by setting α(ui ) = α(vi ) and α(v) = k + 1.
Secondly, χ( G t ) > k. Assume that α is a proper k-colouring of G t , say with α(v) =
k. Then α(ui ) = k. Recolour each vi by α(ui ). This gives a proper (k − 1)-colouring to
G; a contradiction. Therefore χ( G t ) = k + 1.
Now using inductively the above construction starting from the triangle-free
graph K2 , we obtain larger triangle -free graphs with high chromatic numbers.

Critical graphs
D EFINITION . A k-chromatic graph G is said to be k-critical, if χ( H ) < k for all H ⊆ G
with H = G.
In a critical graph an elimination of any edge and of any vertex will reduce the
chromatic number: χ( G −e) < χ( G ) and χ( G −v) < χ( G ) for e ∈ G and v ∈ G. Each
Kn is n-critical, since in Kn −(uv) the vertices u and v can gain the same colour.
Example 4.7. The graph K2 = P2 is the only 2-critical graph. The 3-critical graphs are
exactly the odd cycles C2n+1 for n ≥ 1, since a 3-chromatic G is not bipartite, and thus
must have a cycle of odd length.
Theorem 4.10. If G is k-critical for k ≥ 2, then it is connected, and δ( G ) ≥ k − 1.
Proof. Note that for any graph G with the connected components G1 , G2 , . . . , Gm ,
χ( G ) = max{χ( Gi ) | i ∈ [1, m]} . Connectivity claim follows from this observation.
Let then G be k-critical, but δ( G ) = dG (v) ≤ k − 2 for v ∈ G. Since G is critical,
there is a proper (k − 1)-colouring of G −v. Now v is adjacent to only δ( G ) < k − 1
vertices. But there are k colours, and hence there is an available colour i for v. If we
recolour v by i, then a proper (k − 1)-colouring is obtained for G; a contradiction. ⊔
⊓
The case (iii) of the next theorem is due to S ZEKERES
Theorem 4.11. Let G be any graph with k = χ( G ).
(i) G has a k-critical subgraph H.
(ii) G has at least k vertices of degree ≥ k − 1.
(iii) k ≤ 1 + max H ⊆ G δ( H ).

AND

W ILF (1968).


55

Proof. For (i), we observe that a k-critical subgraph H ⊆ G is obtained by removing
vertices and edges from G as long as the chromatic number remains k.
For (ii), let H ⊆ G be k-critical. By Theorem 4.10, d H (v) ≥ k − 1 for every v ∈ H.
Of course, also dG (v) ≥ k − 1 for every v ∈ H. The claim follows, because, clearly,
every k-critical graph H must have at least k vertices.
For (iii), let H ⊆ G be k-critical. By Theorem 4.10, χ( G ) − 1 ≤ δ( H ), which proves
this claim.
⊔
⊓
Lemma 4.6. Let v be a cut vertex of a connected graph G, and let Ai , for i ∈ [1, m], be the
connected components of G −v. Denote Gi = G [ Ai ∪ {v}]. Then χ( G ) = max{χ( Gi ) | i ∈
[1, m]}. In particular, a critical graph does not have cut vertices.
Proof. Suppose each Gi has a proper k-colouring αi . By Lemma 4.5, we may take
αi (v) = 1 for all i. These k-colourings give a k-colouring of G.
⊔
⊓

Brooks’ theorem
For edge colourings we have Vizing’s theorem, but no such strong results are known
for vertex colouring.
Lemma 4.7. For all graphs G, χ( G ) ≤ ∆( G ) + 1. In fact, there exists a proper colouring
α : VG → [1, ∆( G ) + 1] such that α(v) ≤ dG (v) + 1 for all vertices v ∈ G.
Proof. We use greedy colouring to prove the claim. Let VG = {v1 , . . . , vn } be ordered
in some way, and deﬁne α : VG → N inductively as follows: α(v1 ) = 1, and
α(vi ) = min{ j | α(vt ) = j for all t < i with vi vt ∈ G } .
Then α is proper, and α(vi ) ≤ dG (vi ) + 1 for all i. The claim follows from this.

⊔
⊓

Although, we always have χ( G ) ≤ ∆( G ) + 1, the chromatic number χ( G ) usually
takes much lower values – as seen in the bipartite case. Moreover, the maximum
value ∆( G ) + 1 is obtained only in two special cases as was shown by B ROOKS in
1941.
The next proof of Brook’s theorem is by L OVÁSZ (1975) as modiﬁed by B RYANT
(1996).
Lemma 4.8. Let G be a 2-connected graph. Then the following are equivalent:
(i) G is a complete graph or a cycle.
(ii) For all u, v ∈ G, if uv ∈ G, then {u, v} is a separating set.
/
(iii) For all u, v ∈ G, if dG (u, v) = 2, then {u, v} is a separating set.


56

Proof. It is clear that (i) implies (ii), and that (ii) implies (iii). We need only to show
that (iii) implies (i). Assume then that (iii) holds.
We shall show that either G is a complete graph or dG (v) = 2 for all v ∈ G, from
which the theorem follows.
First of all, dG (v) ≥ 2 for all v, since G is 2-connected. Let w be a vertex of maximum degree, dG (w) = ∆( G ).
If the neighbourhood NG (w) induces a complete subgraph, then G is complete.
Indeed, otherwise, since G is connected, there exists a vertex u ∈ NG (w) ∪ {w} such
/
that u is adjacent to a vertex v ∈ NG (w). But then dG (v) > dG (w), and this contradicts
the choice of w.
Assume then that there are different vertices u, v ∈ NG (w) such that uv ∈ G. This
/
means that dG (u, v) = 2 (the shortest path is u − w − v), and by (iii), {u, v} is a
→
→
separating set of G. Consequently, there is a partition VG = W ∪ {u, v} ∪ U, where
w ∈ W, and all paths from a vertex of W to a vertex of U go through either u or v.
We claim that W = {w}, and thus that ∆( G ) = 2 as required. Suppose on the
contrary that |W | ≥ 2. Since w is not a cut vertex (since G has no cut vertices), there
exists an x ∈ W with x = w such that xu ∈ G or xv ∈ G, say xu ∈ G.
Since v is not a cut vertex, there exists a y ∈ U such
that uy ∈ G. Hence dG ( x, y) = 2, and by (iii), { x, y} is a
separating set. Thus VG = W1 ∪ { x, y} ∪ U1 , where all
paths from W1 to U1 pass through x or y. Assume that
w ∈ W1 , and hence that also u, v ∈ W1 . (Since uw, vw ∈
VG −{ x, y}).

u
w

y

x
v

There exists a vertex z ∈ U1 . Note that U1 ⊆ W ∪ U. If z ∈ W (or z ∈ U, respectively), then all paths from z to u must pass through x (or y, respectively), and x (or
y, respectively) would be a cut vertex of G. This contradiction, proves the claim. ⊔
⊓
Theorem 4.12 ( B ROOKS (1941)). Let G be connected. Then χ( G ) = ∆( G ) + 1 if and only
if either G is an odd cycle or a complete graph.
Proof. (⇐=) Indeed, χ(C2k+1 ) = 3, ∆(C2k+1 ) = 2, and χ(Kn ) = n, ∆(Kn ) = n − 1.
(=⇒) Assume that k = χ( G ). We may suppose that G is k-critical. Indeed, assume
the claim holds for k-critical graphs. Let k = ∆( G ) + 1, and let H ⊂ G be a k-critical
proper subgraph. Since χ( H ) = k = ∆( G ) + 1 > ∆( H ), we must have χ( H ) =
∆( H ) + 1, and thus H is a complete graph or an odd cycle. Now G is connected, and
therefore there exists an edge uv ∈ G with u ∈ H and v ∈ H. But then dG (u) > d H (u),
/
and ∆( G ) > ∆( H ), since H = Kn or H = Cn .
Let then G be any k-critical graph for k ≥ 2. By Lemma 4.6, it is 2-connected. If G
is an even cycle, then k = 2 = ∆( G ). Suppose now that G is neither complete nor a
cycle (odd or even). We show that χ( G ) ≤ ∆( G ).


57

By Lemma 4.8, there exist v1 , v2 ∈ G with dG (v1 , v2 ) = 2, say v1 w, wv2 ∈ G with
v1 v2 ∈ G, such that H = G −{v1 , v2 } is connected. Order VH = {v3 , v4 , . . . , vn } such
/
that vn = w, and for all i ≥ 3,
d H ( vi , w ) ≥ d H ( vi+1 , w ) .
Therefore for each i ∈ [1, n − 1], we find at least one j > i such that vi v j ∈ G (possibly
v j = w). In particular, for all 1 ≤ i < n,

| NG (vi ) ∩ {v1 , . . . , vi−1 }| < dG (vi ) ≤ ∆( G ) .

(4.3)

Then colour v1 , v2 , . . . , vn in this order as follows: α(v1 ) = 1 = α(v2 ) and
α(vi ) = min{r | r = α(v j ) for all v j ∈ NG (vi ) with j < i } .
The colouring α is proper.
By (4.3), α(vi ) ≤ ∆( G ) for all i ∈ [1, n − 1]. Also, w = vn has two neighbours, v1
and v2 , of the same colour 1, and since vn has at most ∆( G ) neighbours, there is an
available colour for vn , and so α(vn ) ≤ ∆( G ). This shows that G has a proper ∆( G )colouring, and, consequently, χ( G ) ≤ ∆( G ).
⊔
⊓
Example 4.8. Suppose we have n objects V = {v1 , . . . , vn }, some of which are not
compatible (like chemicals that react with each other, or worse, graph theorists who
will fight during a conference). In the storage problem we would like to find a partition of the set V with as few classes as possible such that no class contains two incompatible elements. In graph theoretical terminology we consider the graph G = (V, E),
where vi v j ∈ E just in case vi and v j are incompatible, and we would like to colour
the vertices of G properly using as few colours as possible. This problem requires that
we find χ( G ).
Unfortunately, no good algorithms are known for determining χ( G ), and, indeed,
the chromatic number problem is NP-complete. Already the problem if χ( G ) = 3 is
NP-complete. (However, as we have seen, the problem whether χ( G ) = 2 has a fast
algorithm.)

The chromatic polynomial
A given graph G has many different proper vertex colourings α : VG → [1, k] for
sufficiently large natural numbers k. Indeed, see Lemma 4.5 to be certain on this point.
D EFINITION . The chromatic polynomial of G is the function χ G : N → N, where
χ G (k) = |{α | α : VG → [1, k] a proper colouring}| .


58

This notion was introduced by B IRKHOFF (1912), B IRKHOFF AND L EWIS (1946),
to attack the famous 4-Colour Theorem, but its applications have turned out to be
elsewhere.
If k < χ( G ), then clearly χ G (k) = 0, and, indeed,
χ( G ) = min{k | χ G (k) = 0} .
Therefore, if we can find the chromatic polynomial of G, then we easily compute
the chromatic number χ( G ) just by evaluating χ G (k) for k = 1, 2, . . . until we hit a
nonzero value. Theorem 4.13 will give the tools for constructing χ G .
Example 4.9. Consider the complete graph K4 on {v1 , v2 , v3 , v4 }. Let k ≥ χ(K4 ) = 4.
The vertex v1 can be first given any of the k colours, after which k − 1 colours are
available for v2 . Then v3 has k − 2 and finally v4 has k − 3 available colours. Therefore
there are k(k − 1)(k − 2)(k − 3) different ways to properly colour K4 with k colours,
and so
χK4 (k) = k(k − 1)(k − 2)(k − 3) .
On the other hand, in the discrete graph K 4 has no edges, and thus any k-colouring is
a proper colouring. Therefore
χ K 4 ( k ) = k4 .
Remark. The considered method for checking the number of possibilities to colour a
‘next vertex’ is exceptional, and for more nonregular graphs it should be avoided.
D EFINITION . Let G be a graph, e = uv ∈ G, and let x = x(uv) be a new contracted
vertex. The graph G ∗ e on
VG∗e = (VG {u, v}) ∪ { x}
is obtained from G by contracting the edge e, when
EG∗e = { f | f ∈ EG , f has no end u or v} ∪ {wx | wu ∈ G or wv ∈ G } .
Hence G ∗ e is obtained by introducing a new vertex x, and by replacing all edges wu and wv by wx,
and the vertices u and v are deleted.
(Of course, no loops or parallel edges are allowed
in the new graph G ∗ e.)

u
e
v

Theorem 4.13. Let G be a graph, and let e ∈ G. Then
χ G ( k ) = χ G − e ( k ) − χ G ∗ e ( k ).

x


59

Proof. Let e = uv. The proper k-colourings α : VG → [1, k] of G −e can be divided into
two disjoint cases, which together show that χ G−e (k) = χ G (k) + χ G∗e (k):
(1) If α(u) = α(v), then α corresponds to a unique proper k-colouring of G, namely
α. Hence the number of such colourings is χ G (k).
(2) If α(u) = α(v), then α corresponds to a unique proper k-colouring of G ∗ e,
namely α, when we set α( x) = α(u) for the contracted vertex x = x(uv). Hence the
number of such colourings is χ G∗e (k).
⊔
⊓
Theorem 4.14. The chromatic polynomial is a polynomial.
Proof. The proof is by induction on ε G . Indeed, χKn (k) = kn for the discrete graph,
and for two polynomials P1 and P2 , also P1 − P2 is a polynomial. The claim follows
from Theorem 4.13, since there G −e and G ∗ e have less edges than G.
⊔
⊓
The connected components of a graph can be coloured independently, and so
Lemma 4.9. Let the graph G have the connected components G1 , G2 , . . . , Gm . Then
χ G (k) = χ G1 (k)χ G2 (k) . . . χ Gm (k) .
Theorem 4.15. Let T be a tree of order n. Then χ T (k) = k(k − 1)n−1 .
Proof. We use induction on n. For n ≤ 2, the claim is obvious. Suppose that n ≥ 3,
and let e = vu ∈ T, where v is a leaf. By Theorem 4.13, χ T (k) = χ T −e (k) − χ T ∗e (k).
Here T ∗ e is a tree of order n − 1, and thus, by the induction hypothesis, χ T ∗e (k) =
k(k − 1)n−2 . The graph T −e consists of the isolated v and a tree of order n − 1.
By Lemma 4.9, and the induction hypothesis, χ T −e (k) = k · k(k − 1)n−2 . Therefore
χ T ( k ) = k ( k − 1) n − 1 .
⊔
⊓
Example 4.10. Consider the graph G of order 4 from the above. Then we have the
following reductions.

=

e

G−e

G

f

G−e

−

=

G∗e

−
G − {e, f }

( G − e) ∗ f


60

Theorem 4.13 reduces the computation of χ G to the discrete graphs. However, we
know the chromatic polynomials for trees (and complete graphs, as an exercise), and
so there is no need to prolong the reductions beyond these. In our example, we have
obtained
χ G−e (k) = χ G−{e, f } (k) − χ( G−e)∗ f (k)

= k ( k − 1) 3 − k ( k − 1) 2 = k ( k − 1) 2 ( k − 2) ,
and so
χ G (k) = χ G−e (k) − χ G∗e (k) = k(k − 1)2 (k − 2) − k(k − 1)(k − 2)

= k(k − 1)(k − 2)2 = k4 − 5k3 + 8k2 − 4k .
For instance, for 3 colours, there are 6 proper colourings of the given graph.
Chromatic Polynomial Problems. It is difficult to determine χ G of a given graph,
since the reduction method provided by Theorem 4.13 is time consuming. Also, there
is known no characterization, which would tell from any polynomial P(k) whether it
is a chromatic polynomial of some graph. For instance, the polynomial k4 − 3k3 + 3k2
is not a chromatic polynomial of any graph, but it seems to satisfy the general properties (that are known or conjectured) of these polynomials. R EED (1968) conjectured
that the coefficients of a chromatic polynomial should first increase and then decrease
in absolute value. R EED (1968) and T UTTE (1974) proved that for each G of order
νG = n:
•
•
•
•
•
•
•

The degree of χ G (k) equals n.
The coefficient of kn equals 1.
The coefficient of kn−1 equals −ε G .
The constant term is 0.
The coefficients alternate in sign.
χ G (m) ≤ m(m − 1)n − 1 for all positive integers m, when G is connected.
χ G ( x) = 0 for all real numbers 0 < x < 1.

5
Graphs on Surfaces
5.1 Planar graphs
The plane representations of graphs are by no means unique. Indeed, a graph G can
be drawn in arbitrarily many different ways. Also, the properties of a graph are not
necessarily immediate from one representation, but may be apparent from another.
There are, however, important families of graphs, the surface graphs, that rely on
the (topological or geometrical) properties of the drawings of graphs. We restrict ourselves in this chapter to the most natural of these, the planar graphs. The geometry
of the plane will be treated intuitively.
A planar graph will be a graph that can be drawn in the plane so that no two
edges intersect with each other. Such graphs are used, e.g., in the design of electrical
(or similar) circuits, where one tries to (or has to) avoid crossing the wires or laser
beams. Planar graphs come into use also in some parts of mathematics, especially in
group theory and topology.
There are fast algorithms (linear time algorithms) for testing whether a graph is
planar or not. However, the algorithms are all rather difficult to implement. Most of
them are based on an algorithm designed by A USLANDER AND PARTER (1961) see
Section 6.5 of
S. S KIENA, “Implementing Discrete Mathematics: Combinatorics and Graph Theory
with Mathematica”, Addison-Wesley, 1990.

Definition
D EFINITION . A graph G is a planar graph, if it has a
plane figure P( G ), called the plane embedding of G,
where the lines (or continuous curves) corresponding to the edges do not intersect each other except at
their ends.
The complete bipartite graph K2,4 is a planar graph.
D EFINITION . An edge e = uv ∈ G is subdivided, when it is replaced by a path
u − x − v of length two by introducing a new vertex x. A subdivision H of a graph
→ →
G is obtained from G by a sequence of subdivisions.

5.1 Planar graphs

62

The following result is clear.
Lemma 5.1. A graph is planar if and only if its subdivisions are planar.

Geometric properties
It is clear that the graph theoretical properties of G are inherited by all of its plane
embeddings. For instance, the way we draw a graph G in the plane does not change
its maximum degree or its chromatic number. More importantly, there are – as we
shall see – some nontrivial topological (or geometric) properties that are shared by
the plane embeddings.
We recall ﬁrst some elements of the plane geometry. Let F be an open set of the
plane R × R, that is, every point x ∈ F has a disk centred at x and contained in F.
Then F is a region, if any two points x, y ∈ F can be joined by a continuous curve the
points of which are all in F. The boundary ∂( F ) of a region F consists of those points
for which every neighbourhood contains points from F and its complement.
Let G be a planar graph, and P( G ) one of its plane embeddings. Regard now each
edge e = uv ∈ G as a line from u to v. The set (R × R ) EG is open, and it is divided
into a ﬁnite number of disjoint regions, called the faces of P( G ).
D EFINITION . A face of P( G ) is an interior face, if it is
bounded. The (unique) face that is unbounded is called the
exterior face of P( G ). The edges that surround a face F constitute the boundary ∂( F ) of F. The exterior boundary is the
boundary of the exterior face. The vertices (edges, resp.) on
the exterior boundary are called exterior vertices exterior
edges, resp.). Vertices (edges, resp.) that are not on the exterior boundary are interior vertices interior edges, resp.).

F2

F3

F1
F0

Embeddings P( G ) satisfy some properties that we accepts at face value.
Lemma 5.2. Let P( G ) be a plane embedding of a planar graph G.
(i) Two different faces F1 and F2 are disjoint, and their boundaries can intersect only on edges.
(ii) P( G ) has a unique exterior face.
(iii) Each edge e belongs to the boundary of at most two faces.
(iv) Each cycle of G surrounds (that is, its interior contains) at least one internal face of P( G ).
(v) A bridge of G belongs to the boundary of only one face.
(vi) An edge that is not a bridge belongs to the boundary of exactly two faces.

5.1 Planar graphs

63

If P( G ) is a plane embedding of a graph G, then so is any drawing P′ ( G ) which is
obtained from P( G ) by an injective mapping of the plane that preserves continuous
curves. This means, in particular, that every planar graph has a plane embedding inside
any geometric circle of arbitrarily small radius, or inside any geometric triangle.

Euler’s formula
Lemma 5.3. A plane embedding P( G ) of a planar graph G has no interior faces if and only if
G is acyclic, that is, if and only if the connected components of G are trees.
Proof. This is clear from Lemma 5.2.

⊔
⊓

The next general form of Euler’s formula was proved by L EGENDRE (1794).
Theorem 5.1 (Euler’s formula). Let G be a connected planar graph, and let P( G ) be any of
its plane embeddings. Then
νG − ε G + ϕ = 2 ,
where ϕ is the number of faces of P( G ).
Proof. We shall prove the claim by induction on the number of faces ϕ of a plane
embedding P( G ). First, notice that ϕ ≥ 1, since each P( G ) has an exterior face.
If ϕ = 1, then, by Lemma 5.3, there are no cycles in G, and since G is connected, it
is a tree. In this case, by Theorem 2.4, we have ε G = νG − 1, and the claim holds.
Suppose then that the claim is true for all plane embeddings with less than ϕ faces
for ϕ ≥ 2. Let P( G ) be a plane embedding of a connected planar graph such that
P( G ) has ϕ faces.
Let e ∈ G be an edge that is not a bridge. The subgraph G −e is planar with a plane
embedding P( G −e) = P( G )−e obtained by simply erasing the edge e. Now P( G −e)
has ϕ − 1 faces, since the two faces of P( G ) that are separated by e are merged into
one face of P( G −e). By the induction hypothesis, νG− e − ε G−e + ( ϕ − 1) = 2, and
hence νG − (ε G − 1) + ( ϕ − 1) = 2, and the claim follows.
⊔
⊓
In particular, we have the following invariant property of planar graphs.
Corollary 5.1. Let G be a planar graph. Then every plane embedding of G has the same
number of faces:
ϕ G = ε G − νG + 2

Maximal planar graphs
Lemma 5.4. If G is a planar graph of order νG ≥ 3, then ε G ≤ 3νG − 6. Moreover, if G has
no triangles C3 , then ε G ≤ 2νG − 4.

5.1 Planar graphs

64

Proof. If G is disconnected with connected components Gi , for i ∈ [1, k], and if the
claim holds for these smaller (necessarily planar) graphs Gi , then it holds for G, since
νG

νG

i=1

εG =

i=1

∑ ε Gi ≤ 3 ∑ νGi − 6k = 3νG − 6k ≤ 3νG − 6 .

It is thus sufﬁcient to prove the claim for connected planar graphs.
Also, the case where ε G ≤ 2 is clear. Suppose thus that ε G ≥ 3.
Each face F of an embedding P( G ) contains at least three edges on its boundary
∂( F ). Hence 3ϕ ≤ 2ε G , since each edge lies on at most two faces. The ﬁrst claim
follows from Euler’s formula.
The second claim is proved similarly except that, in this case, each face F of P( G )
contains at least four edges on its boundary (when G is connected and ε G ≥ 4).
⊔
⊓
An upper bound for δ( G ) for planar graphs was achieved by H EAWOOD.
Theorem 5.2 (H EAWOOD (1890)). If G is a planar graph, then δ( G ) ≤ 5.
Proof. If νG ≤ 2, then there is nothing to prove. Suppose νG ≥ 3. By the handshaking
lemma and the previous lemma,
δ( G ) · νG ≤

∑ dG (v) = 2ε G ≤ 6νG − 12 .
v∈ G

It follows that δ( G ) ≤ 5.

⊔
⊓

Theorem 5.3. K5 and K3,3 are not planar graphs.
Proof. By Lemma 5.4, a planar graph of order 5 has at most 9 edges, but K5 has 5
vertices and 10 edges. By the second claim of Lemma 5.4, a triangle-free planar graph
of order 6 has at most 8 edges, but K3,3 has 6 vertices and 9 edges.
⊔
⊓
D EFINITION . A planar graph G is maximal, if G + e is nonplanar for every e ∈ G.
/
Example 5.1. Clearly, if we remove one edge from K5 , the result is a maximal planar
graph. However, if an edge is removed from K3,3 , the result is not maximal!
Lemma 5.5. Let F be a face of a plane embedding P( G ) that has at least four edges on its
boundary. Then there are two nonadjacent vertices on the boundary of F.
Proof. Assume that the set of the boundary vertices of F induces a complete subgraph K. The edges of K are either on the boundary or they are not inside F (since F
is a face.) Add a new vertex x inside F, and connect the vertices of K to x. The result
is a plane embedding of a graph H with VH = VG ∪ { x} (that has G as its induced
subgraph). The induced subgraph H [K ∪ { x}] is complete, and since H is planar, we
have |K | < 4 as required.
⊔
⊓

5.1 Planar graphs

65

By the previous lemma, if a face has a boundary of at least four edges, then an
edge can be added to the graph (inside the face), and the graph remains to be planar.
Hence we have proved
Corollary 5.2. If G is a maximal planar graph with νG ≥ 3, then G is triangulated, that is,
every face of a plane embedding P( G ) has a boundary of exactly three edges.
Theorem 5.4. For a maximal planar graph G of order νG ≥ 3, ε G = 3νG − 6 .
Proof. Each face F of an embedding P( G ) is a triangle having three edges on its
boundary. Hence 3ϕ = 2ε G , since there are now no bridges. The claim follows from
Euler’s formula.
⊔
⊓

Kuratowski’s theorem
Theorem 5.5 will give a simple criterion for planarity of graphs. This theorem (due to
K URATOWSKI in 1930) is one of the jewels of graph theory. In fact, the theorem was
proven earlier by P ONTRYAGIN (1927-1928), and also independently by F RINK AND
S MITH (1930). For history of the result, see
J.W. K ENNEDY, L.V. Q UINTAS , AND M.M. S YSLO, The theorem on planar graphs.
Historia Math. 12 (1985), 356 – 368.
The graphs K5 and K3,3 are the smallest nonplanar graphs, and, by Lemma 5.1, if
G contains a subdivision of K5 or K3,3 as a subgraph, then G is not planar. We prove
the converse of this result in what follows. Therefore
Theorem 5.5 (K URATOWSKI (1930)). A graph is planar if and only if it contains no subdivision of K5 or K3,3 as a subgraph.
We prove this result along the lines of T HOMASSEN (1981) using 3-connectivity.
Example 5.2. The cube Qk is planar only for k = 1, 2, 3. Indeed, the graph Q4 contains
a subdivision of K3,3 , and thus by Theorem 5.5 it is not planar. On the other hand, each
Qk with k ≥ 4 has Q4 as a subgraph, and therefore they are nonplanar. The subgraph
of Q4 that is a subdivision of K3,3 is given below.
0000

1010
0100

1110

1001
1101

0001
0011

1000

1100

0010

5.1 Planar graphs

66

D EFINITION . A graph G is called a Kuratowski graph, if it is a subdivision of K5 or
K3,3 .
Lemma 5.6. Let E ⊆ EG be the set of the boundary edges of a face F in a plane embedding of
G. Then there exists a plane embedding P( G ), where the edges of E are exterior edges.
Proof. This is a geometric proof. Choose a circle that contains every point of the plane
embedding (including all points of the edges) such that the centre of the circle is
inside the given face. Then use geometric inversion with respect to this circle. This
will map the given face as the exterior face of the image plane embedding.
⊔
⊓
Lemma 5.7. Let G be a nonplanar graph without Kuratowski graphs of minimal total size
ε G + νG . Then G is 3-connected.
Proof. By the minimality assumption, G is connected. We show then that G is 2connected. On the contrary, assume that v is a cut vertex of G, and let A1 , . . . , Ak be
the connected components of G −v.
Since G is minimal nonplanar with respect to ε G , the
subgraphs Gi = G [ Ai ∪ {v}] have plane embeddings
A1
A2
P( Gi ), where v is an exterior vertex. We can glue these
plane embeddings together at v to obtain a plane embedding of G, and this will contradict the choice of G.
Assume then that G has a separating set S = {u, v}. Let G1 and G2 be any subgraphs of G such that EG = EG1 ∪ EG2 , S = VG1 ∩ VG2 , and both G1 and G2 contain a
connected component of G −S. Since G is 2-connected (by the above), there are paths
⋆
u − v in G1 and G2 . Indeed, both u and v are adjacent to a vertex of each connected
→
component of G −S. Let Hi = Gi + uv. (Maybe uv ∈ G.)
If both H1 and H2 are planar, then, by Lemma 5.6, they
have plane embeddings, where uv is an exterior edge.
H1
H2
It is now easy to glue H1 and H2 together on the edge
uv to obtain a plane embedding of G + uv, and thus
of G.
We conclude that H1 or H2 is nonplanar, say H1 . Now ε H1 < ε G , and so, by the
⋆
minimality of G, H1 contains a Kuratowski graph H. However, there is a path u − v
→
in H2 , since G2 ⊆ H2 . This path can be regarded as a subdivision of uv, and thus G
contains a Kuratowski graph. This contradiction shows that G is 3-connected.
⊔
⊓
Lemma 5.8. Let G be a 3-connected graph of order νG ≥ 5. Then there exists an edge e ∈ G
such that the contraction G ∗ e is 3-connected.
Proof. On the contrary suppose that for any e ∈ G, the graph G ∗ e has a separating
set S with |S| = 2. Let e = uv, and let x = x(uv) be the contracted vertex. Necessarily
x ∈ S, say S = { x, z} (for, otherwise, S would separate G already). Therefore T =

5.1 Planar graphs

67

{u, v, z} separates G. Assume that e and S are chosen such that G − T has a connected
component A with the least possible number of vertices.
There exists a vertex y ∈ A with zy ∈ G. (Otherwise
{u, v} would separate G.) The graph G ∗ (zy) is not 3connected by assumption, and hence, as in the above,
there exists a vertex w such that R = {z, y, w} separates
G. It can be that w ∈ {u, v}, but by symmetry we can
suppose that w = u.

B

u
T v
z

A
y

Since uv ∈ G, G − R has a connected component B such that u, v ∈ B. For each
/
⋆
y′ ∈ B, there exists a path P : u − y′ in G −{z, w}, since G is 3-connected, and hence
→
this P goes through y. Therefore y′ is connected to y also in G − T, that is, y′ ∈ A, and
so B ⊆ A. The inclusion is proper, since y ∈ B. Hence | B| < | A|, and this contradicts
/
the choice of A.
⊔
⊓
By the next lemma, a Kuratowski graph cannot be created by contractions.
Lemma 5.9. Let G be a graph. If for some e ∈ G the contraction G ∗ e has a Kuratowski
subgraph, then so does G.
Proof. The proof consists of several cases depending on the Kuratowski graph, and
how the subdivision is made. We do not consider the details of these cases.
Let H be a Kuratowski graph of G ∗ e, where x = x(uv) is the contracted vertex for
e = uv. If d H ( x) = 2, then the claim is obviously true. Suppose then that d H ( x) = 3
or 4. If there exists at most one edge xy ∈ H such that uy ∈ G (or vy ∈ G), then one
easily sees that G contains a Kuratowski graph.
There remains only one case, where H is a subdivision of K5 , and both u and v
have 3 neighbours in the subgraph of G corresponding to H. In this case, G contains
a subdivision of K3,3 .
⊔
⊓
v2

v4

v2

x
v1

v4
u

v3

v1

v
v3

Lemma 5.10. Every 3-connected graph G without Kuratowski subgraphs is planar.
Proof. The proof is by induction on νG . The only 3-connected graph of order 4 is the
planar graph K4 . Therefore we can assume that νG ≥ 5.
By Lemma 5.8, there exists an edge e = uv ∈ G such that G ∗ e (with a contracted
vertex x) is 3-connected. By Lemma 5.9, G ∗ e has no Kuratowski subgraphs, and
hence G ∗ e has a plane embedding P( G ∗ e) by the induction hypothesis. Consider

5.2 Colouring planar graphs

68

the part P( G ∗ e)− x, and let C be the boundary of the face of P( G ∗ e)− x containing
x (in P( G ∗ e)). Here C is a cycle of G (since G is 3-connected).
Now since G −{u, v} = ( G ∗ e)− x, P( G ∗ e)− x is a plane embedding of G −{u, v},
and NG (u) ⊆ VC ∪ {v} and NG (v) ⊆ VC ∪ {u}. Assume, by symmetry, that dG (v) ≤
⋆
dG (u). Let NG (v) {u} = {v1 , v2 , . . . , vk } in order along the cycle C. Let Pi,j : vi − v j
→
be the path along C from vi to v j . We obtain a plane embedding of G −u by drawing
(straight) edges vvi for 1 ≤ i ≤ k.
(1) If NG (u) {v} ∈ Pi,i+1 (i + 1 is taken modulo k) for some i, then, clearly, G has a
plane embedding (obtained from P( G )−u by putting u inside the triangle (v, vi , vi+1 )
and by drawing the edges with an end u inside this triangle).
y

(2) Assume there are y, z ∈ NG (u) {v} such that y ∈ Pij
and z ∈ Pij for some i and j, where y, z ∈ {vi , v j }. Now,
/
/
{u, vi , vi+1 } ∪ {v, z, y} form a subdivision of K3,3.

By (1) and (2), we can assume that NG (u) {v} ⊆ NG (v).
Therefore, NG (u) {v} = NG (v) {u} by the assumption
dG (v) ≤ dG (u). Also, by (1), dG (v) = dG (u) > 3. But now
u, v, v1 , v2 , v3 give a subdivision of K5 .
⊔
⊓

z

u
v

u
v

Proof of Theorem 5.5. By Theorem 5.3 and Lemma 5.1, we need to show that each
nonplanar graph G contains a Kuratowski subgraph. On the contrary, suppose that
G is a nonplanar graph that has a minimal size ε G such that G does not contain a
Kuratowski subgraph. Then, by Lemma 5.7, G is 3-connected, and by Lemma 5.10, it
is planar. This contradiction proves the claim.
⊔
⊓
Example 5.3. Any graph G can be drawn in the plane so that three of its edges never
intersect at the same point. The crossing number ×( G ) is the minimum number of
intersections of its edges in such plane drawings of G. Therefore G is planar if and
only if ×( G ) = 0, and, for instance, ×(K5 ) = 1.
We show that ×(K6 ) = 3. For this we need to show that ×(K6 ) ≥ 3. For the equality, one is invited to design a drawing with exactly 3 crossings.
Let X (K6 ) be a drawing of K6 using c crossings so that two edges cross at most
once. Add a new vertex at each crossing. This results in a planar graph G on c + 6
vertices and 2c + 15 edges. Now c ≥ 3, since ε G = 2c + 15 ≤ 3(c + 6) − 6 = 3νG − 6.

The most famous problem in the history of graph theory is that of the chromatic
number of planar graphs. The problem was known as the 4-Colour Conjecture for
more than 120 years, until it was solved by A PPEL AND H AKEN in 1976: if G is a


69

planar graph, then χ( G ) ≤ 4. The 4-Colour Conjecture has had a deep influence on
the theory of graphs during the last 150 years. The solution of the 4-Colour Theorem
is difficult, and it requires the assistance of a computer.

The 5-colour theorem
We prove H EAWOOD’s result (1890) that each planar graph is properly 5-colourable.
Lemma 5.11. If G is a planar graph, then χ( G ) ≤ 6.
Proof. The proof is by induction on νG . Clearly, the claim holds for νG ≤ 6. By Theorem 5.2, a planar graph G has a vertex v with dG (v) ≤ 5. By the induction hypothesis,
χ( G −v) ≤ 6. Since dG (v) ≤ 5, there is a colour i available for v in the 6-colouring of
G −v, and so χ( G ) ≤ 6.
⊔
⊓
The proof of the following theorem is partly geometric in nature.
Theorem 5.6 (H EAWOOD (1890)). If G is a planar graph, then χ( G ) ≤ 5.
Proof. Suppose the claim does not hold, and let G be a 6-critical planar graph. Recall that for k-critical graphs H, δ( H ) ≥ k − 1, and thus there exists a vertex v with
dG (v) = δ( G ) ≥ 5. By Theorem 5.2, dG (v) = 5.
Let α be a proper 5-colouring of G −v. Such a colouring
exists, because G is 6-critical. By assumption, χ( G ) > 5,
and therefore for each i ∈ [1, 5], there exists a neighbour vi ∈ NG (v) such that α(vi ) = i. Suppose these
neighbours vi of v occur in the plane in the geometric
order of the figure.

v5

v1
v

v4

v2

P13

v3

Consider the subgraph G [i, j] ⊆ G made of colours i and j. The vertices vi and
v j are in the same connected component of G [i, j] (for, otherwise we interchange the
colours i and j in the connected component containing v j to obtain a recolouring of
G, where vi and v j have the same colour i, and then recolour v with the remaining
colour j).
⋆
Let Pij : vi − v j be a path in G [i, j], and let C = (vv1 ) P13 (v3 v). By the geometric
→
assumption, exactly one of v2 , v4 lies inside the region enclosed by the cycle C. Now,
the path P24 must meet C at some vertex of C, since G is planar. This is a contradiction,
since the vertices of P24 are coloured by 2 and 4, but C contains no such colours. ⊔
⊓
The final word on the chromatic number of planar graphs was proved by A PPEL
AND H AKEN in 1976.
Theorem 5.7 (4-Colour Theorem). If G is a planar graph, then χ( G ) ≤ 4.
By the following theorem, each planar graph can be decomposed into two bipartite graphs.


70

Theorem 5.8. Let G = (V, E) be a 4-chromatic graph, χ( G ) ≤ 4. Then the edges of G can
be partitioned into two subsets E1 and E2 such that (V, E1 ) and (V, E2 ) are both bipartite.
Proof. Let Vi = α−1 (i ) be the set of vertices coloured by i in a proper 4-colouring α of
G. The define E1 as the subset of the edges of G that are between the sets V1 and V2 ;
V1 and V4 ; V3 and V4 . Let E2 be the rest of the edges, that is, they are between the sets
V1 and V3 ; V2 and V3 ; V2 and V4 . It is clear that (V, E1 ) and (V, E2 ) are bipartite, since
the sets Vi are stable.
⊔
⊓

Map colouring∗
The 4-Colour Conjecture was originally stated for maps. In the map-colouring problem we are given several countries with common borders, and we wish to colour
each country so that no neighbouring countries obtain the same colour. How many
colours are needed?
A border between two countries is assumed to have a positive length – in particular, countries that have only one point in common are not allowed in the map
colouring.
Formally, we define a map as a connected planar (embedding of a) graph with no
bridges. The edges of this graph represent the boundaries between countries. Hence
a country is a face of the map, and two neighbouring countries share a common edge
(not just a single vertex). We deny bridges, because a bridge in such a map would be
a boundary inside a country.
The map-colouring problem is restated as follows:
How many colours are needed for the faces of a plane
embedding so that no adjacent faces obtain the same
colour.
The illustrated map can be 4-coloured, and it cannot be coloured using only 3 colours, because every two faces have a common border.
Let F1 , F2 , . . . , Fn be the countries of a map M, and define a graph G with VG =
{v1 , v2 , . . . , vn } such that vi v j ∈ G if and only if the countries Fi and Fj are neighbours.
It is easy to see that G is a planar graph. Using this notion of a dual graph, we can
state the map-colouring problem in new form: What is the chromatic number of a planar
graph? By the 4-Colour Theorem it is at most four.
Map-colouring can be used in rather generic topological setting, where the maps
are defined by curves in the plane. As an example, consider finitely many simple
closed curves in the plane. These curves divide the plane into regions. The regions are
2-colourable.


71

That is, the graph where the vertices correspond to the regions, and the edges correspond
to the neighbourhood relation, is bipartite. To
see this, colour a region by 1, if the region is inside an odd number of curves, and, otherwise,
colour it by 2.

2
1

1
2
2

1

2
1
1

2

1

2

History of the 4-Colour Theorem
That four colours suffice planar maps was conjectured around 1850 by F RANCIS
G UTHRIE, a student of D E M ORGAN at University College of London. During the
following 120 years many outstanding mathematicians tried to solve the problem,
and some of them even thought that they had been successful.
In 1879 C AYLEY pointed out some difficulties that lie in the conjecture. The same
year A LFRED K EMPE published a paper, where he claimed a proof of the 4CC. The
basic idea in K EMPE’s argument (known later as Kempe chains) was the same as later
used by H EAWOOD to prove the 5-Colour Theorem, (Theorem 5.6).
For more than 10 years K EMPE’s proof was considered to be valid. For instance,
TAIT published two papers on the 4CC in the 1880’s that contained clever ideas, but
also some further errors. In 1890 H EAWOOD showed that K EMPE’s proof had serious gaps. As we shall see in the next chapter, H EAWOOD discovered the number
of colours needed for all maps on other surfaces than the plane. Also, he proved
that if the number of edges around each region is divisible by 3, then the map is
4-colourable.
One can triangulate any planar graph G (drawn in the plane), by adding edges
to divide the faces into triangles. B IRKHOFF introduced one of the basic notions (reducibility) needed in the proof of the 4CC. In a triangulation, a configuration is a part
that is contained inside a cycle. An unavoidable set is a set of configurations such that
any triangulation must contain one of the configurations in the set. A configuration
is said to be reducible, if it is not contained in a triangulation of a minimal counter
example to the 4CC.
The search for avoidable sets began in 1904 with work of W EINICKE, and in 1922
F RANKLIN showed that the 4CC holds for maps with at most 25 regions. This number
was increased to 27 by R EYNOLDS (1926), to 35 by W INN (1940), to 39 by O RE AND
S TEMPLE (1970), to 95 by M AYER (1976).
The final notion for the solution was due to H EESCH, who in 1969 introduced
discharging. This consists of assigning to a vertex v the charge 6 − dG (v). From Euler’s
formula we see that for the sum of the charges, we have

∑(6 − dG (v)) = 12.
v

Now, a given set S of configurations can be proved to be unavoidable, if for a triangulation, that does not contain a configuration from S, one can ‘redistribute’ the charges
so that no v comes up with a positive charge.


72

According to H EESCH one might be satisfied with a set of 8900 configurations to
prove the 4CC. There were difficulties with his approach that were solved in 1976
by A PPEL AND H AKEN. They based the proof on reducibility using Kempe chains,
and ended up with an unavoidable set with over 1900 configurations and some 300
discharging rules. The proof used 1200 hours of computer time. (K OCH assisted with
the computer calculations.) A simplified proof by R OBERTSON , S ANDERS , S EYMOUR
AND T HOMAS (1997) uses 633 configurations and 32 discharging rules. Because of
these simplifications also the computer time is much less than in the original proof.
The following book contains the ideas of the proof of the 4-Colour Theorem.
T.L. S AATY AND P.C. K AINEN, “The Four-Color Problem”, Dover, 1986.

List colouring
D EFINITION . Let G be a graph so that each of its vertices v is given a list (set) Λ(v) of
colours. A proper colouring α : VG → [1, m] of G is a (Λ-)list colouring, if each vertex
v gets a colour from its list, α(v) ∈ Λ(v).
The list chromatic number χℓ ( G ) is the smallest integer k such that G has a Λ-list
colouring for all lists of size k, |Λ(v)| = k}. Also, G is k-choosable, if χℓ ( G ) ≤ k.
Example 5.4. The bipartite graph K3,3 is not 2choosable. Indeed, let the bipartition of K3,3 be
( X, Y ), where X = { x1 , x2 , x3 } and Y = {y1 , y2 , y3 }.
The lists for the vertices shown in the figure show
that χℓ (K3,3 ) > 2.

{1, 2}

{1, 3}

{2, 3}

y1

y2

y3

x1

x2

x3

{1, 2}

{1, 3}

{2, 3}

Obviously χ( G ) ≤ χℓ ( G ), since proper colourings are special cases of list colourings, but equality does not hold in general. However, it was proved by V IZING (1976)
and E RDÖS , R UBIN AND TAYLOR (1979) that
χ ℓ ( G ) ≤ ∆( G ) + 1 .
For planar graphs we do not have a ‘4-list colour theorem’. Indeed, it was shown by
V OIGT (1993) that there exists a planar graph with χℓ ( G ) = 5. At the moment, the
smallest such a graph was produced by M IRZAKHANI (1996), and it is of order 63.
Theorem 5.9 (T HOMASSEN (1994)). Let G be a planar graph. Then χℓ ( G ) ≤ 5.
In fact, T HOMASSEN proved a stronger statement:
Theorem 5.10. Let G be a planar graph and let C be the cycle that is the boundary of the
exterior face. Let Λ consist of lists such that |Λ(v)| = 3 for all v ∈ C, and |Λ(v)| = 5 for all
v ∈ C. Then G has a Λ-list colouring α.
/
Proof. We can assume that the planar graph G is connected, and that it is given by
a near-triangulation; an embedding, where the interior faces are triangles. (If the


73

boundary of a face has more than 3 edges, then we can add an edge inside the face.)
This is because adding edges to a graph can only make the list colouring more difficult. Note that the exterior boundary is unchanged by a triangulation of the interior
faces.
The proof is by induction on νG under the additional constraint that one of the
vertices of C has a fixed colour. (Thus we prove a stronger statement than claimed.)
For νG ≤ 3, the claim is obvious. Suppose then that νG ≥ 4.
Let x ∈ C be a vertex, for which we fix a colour c ∈ Λ( x). Let v ∈ C be a vertex
⋆
adjacent to x, that is, C : v → x − v.
→
Let NG (v) = { x, v1 , . . . , vk , y}, where y ∈ C, and vi are
v
y
ordered such that the faces are triangles as in the figure.
x
It can be that NG (v) = { x, y}, in which case xy ∈ G.
v1

vk
v2
Consider the subgraph H = G −v. The exterior bound...
⋆
ary of H is the cycle x → v1 → · · · → vk → y − x.
→
Since |Λ(v)| = 3, there are two colours r, s ∈ Λ(v) that
differ from the colour c of x.
We define new lists for H as follows: Λ′ (vi ) ⊆ Λ(vi ) {r, s} such that |Λ′ (vi )| = 3
for each i ∈ [1, k], and otherwise Λ′ (z) = Λ(z).
Now νH = νG − 1, and by the induction hypothesis (with c still fixed for x), H has a
Λ′ -list colouring α. For the vertex v, we choose α(v) = r or s such that α(v) = α(y).
This gives a Λ′ -list colouring for G. Since Λ′ (z) ⊆ Λ(z) for all z, we have that α is a
Λ-list colouring of G.
⊔
⊓

Straight lines and kissing circles∗
We state an interesting result of WAGNER, the proof of which can be deduced from
the above proof of Kuratowski’s theorem. The result is known as Fáry’s Theorem.
Theorem 5.11 (WAGNER (1936)). A planar graph G has a plane embedding, where the edges
are straight lines.
This raises a difficult problem:
Integer Length Problem. Can all planar graphs be drawn in the plane such that the
edges are straight lines of integer lengths?
We say that two circles kiss in the plane, if they intersect in one point and their interiors do not intersect. For
a set of circles, we draw a graph by putting an edge between two midpoints of kissing circles.
The following improvement of the above theorem is
due to K OEBE (1936), and it was rediscovered independently by A NDREEV (1970) and T HURSTON (1985).

·
·

·

·

·

·
·

·

Theorem 5.12 (K OEBE (1936)). A graph is planar if and only if it is a kissing graph of circles.


74

Graphs can be represented as plane figures in many different ways. For this, consider a set S of curves of the plane (that are continuous between their end points).
The string graph of S is the graph G = (S, E), where uv ∈ E if and only if the curves
u and v intersect. At first it might seem that every graph is a string graph, but this is
not the case.
It is known that all planar graphs are string graphs (this is a trivial result).
Line Segment Problem. A graph is a line segment graph if it is a string graph for a
set L of straight line segments in the plane. Is every planar graph a line segment graph
for some set L of lines?
Note that there are also nonplanar graphs that are line segment graphs. Indeed,
all complete graphs are such graphs.
The above question remains open even in the case
when the slopes of the lines are +1, −1, 0 and ∞. +1
A positive answer to this 4-slope problem for planar graphs would prove the 4-Colour Theorem.

−1

0

∞

The Minor Theorem∗
D EFINITION . A graph H is a minor of G, denoted by H G, if H is isomorphic to a
graph obtained from a subgraph of G by successively contracting edges.
A recent result of R OBERTSON AND S EYMOUR (1983-2000) on graph minors is (one
of) the deepest results of graph theory. The proof goes beyond these lectures. Indeed,
the proof of Theorem 5.13 is around 500 pages long.

e

G

a subgraph

a contraction

Note that every subgraph H ⊆ G is a minor, H G.
The following properties of the minor relation are easily established:
(i) G
(ii) H
(iii) H

G,
G and G
L and L

H imply G ∼ H,
=
G imply H G.

The conditions (i) and (iii) ensure that the relation
is a quasi-order, that is, it is
reflexive and transitive. It turns out to be a well-quasi-order, that is, every infinite
sequence G1 , G2 , . . . of graphs has two graphs Gi and Gj with i < j such that Gi Gj .


75

Theorem 5.13 (Minor Theorem). The minor order is a well-quasi-order on graphs. In
particular, in any infinite family F of graphs, one of the graphs is a (proper) minor of another.
Each property P of graphs defines a family of graphs, namely, the family of those
graphs that satisfy this property.
D EFINITION . A family F of graphs is said to be minor closed, if every minor H of a
graph G ∈ F is also in F . A property P of graphs is said to be inherited by minors,
if all minors of a graph G satisfy P whenever G does.
The following families of graphs are minor closed: the family of (1) all graphs, (2)
planar graphs (and their generalizations to other surfaces), (3) acyclic graphs.
The acyclic graphs include all trees. However, the family of trees is not closed under
taking subgraphs, and thus it is not minor closed. More importantly, the subgraph
order of trees (T1 ⊆ T2 ) is not a well-quasi-order.
WAGNER proved a minor version of Kuratowski’s theorem:
Theorem 5.14 (WAGNER (1937)). A graph G is nonplanar if and only if K5
G.

G or K3,3

Proof. Exercise.
R OBERTSON

AND

⊔
⊓
S EYMOUR (1998) proved the Wagner’s conjecture:

Theorem 5.15 (Minor Theorem 2). Let P be a property of graphs inherited by minors.
Then there exists a finite set F of graphs such that G satisfies P if and only if G does not
have a minor from F .
One of the impressive application of Theorem 5.15 concerns embeddings of graphs
on surfaces, see the next chapters. By Theorem 5.15, one can test (with a fast algorithm) whether a graph can be embedded onto a surface.
Every graph can be drawn in the 3-dimensional space without crossing edges. An
old problem asks if there exists an algorithm that would determine whether a graph
can be drawn so that its cycles do not form (nontrivial) knots. This problem is solved
by the above results, since the property ‘knotless’ is inherited by minors: there exists
a fast algorithm to do the job. However, this algorithm is not known!
Hadwiger’s Problem. H ADWIGER conjectured in 1943 that for every graph G,
Kχ( G)

G,

that is, if χ( G ) ≥ r, then G has a complete graph Kr as its minor. The conjecture is trivial
for r = 2, and it is known to hold for all r ≤ 6. The cases for r = 5 and 6 follow from
the 4-Colour Theorem.

5.3 Genus of a graph

76

A graph is planar, if it can be drawn in the plane without crossing edges. A plane
is an important special case of a surface. In this section we study shortly drawing
graphs in other surfaces.
There are quite many interesting surfaces many of which are rather difficult to
draw. We shall study the ‘easy surfaces’ – those that are compact and orientable.
These are surfaces that have both an inside and an outside, and can be entirely characterized by the number of holes in them. This number is the genus of the surface.
There are also non-orientable compact surfaces such as the Klein bottle and the projective plane.

Background on surfaces
We shall first have a quick look at the general surfaces and their classification without
going into the details. Consider the space R3 , which has its (usual) distance function
d( x, y) ∈ R of its points.
Two figures (i.e., sets of points) A and B are topologically equivalent (or homeomorphic) if there exists a bijection f : A → B such that f and its inverse f −1 : B → A
are continuous. In particular, two figures are topologically equivalent if one can be
deformed to the other by bending, squeezing, stretching, and shrinking without tearing it apart or gluing any of its parts together. All these deformations should be such
that they can be undone.
A set of points X is a surface, if X is connected (there is a continuous line inside
X between any two given points) and every point x ∈ X has a neighbourhood that is
topologically equivalent to an open planar disk D ( a) = { x | dist ( a, x) < 1}.
We deal with surfaces of the real space, and in this case a surface X is compact, if
X is closed and bounded. Note that the plane is not compact, since it it not bounded.
A subset of a compact surface X is a triangle if it is topologically equivalent to a
triangle in the plane. A finite set of triangles Ti , i = 1, 2, . . . , m, is a triangulation of
X if X = ∪m 1 Ti and any nonempty intersection Ti ∩ Tj with i = j is either a vertex or
i=
an edge.
The following is due to R ADÓ (1925).
Theorem 5.16. Every compact surface has a triangulation.
Each triangle of a surface can be oriented by choosing an order for its vertices up
to cyclic permutations. Such a permutation induces a direction for the edges of the triangle. A triangulation is said to be oriented if the triangles are assigned orientations
such that common edges of two triangles are always oriented in reverse directions. A
surface is orientable if it admits an oriented triangulation.
Equivalently, orientability can be described as follows.
Theorem 5.17. A compact surface X is orientable if and only if it has no subsets that are
topologically equivalent to the Möbius band.


77

In the Möbius band (which itself is not a surface according
the above definition) one can travel around and return to the
starting point with left and right reversed.
A connected sum X#Y of two compact surfaces is obtained by cutting an open
disk off from both surfaces and then gluing the surfaces together along the boundary
of the disks. (Such a deformation is not allowed by topological equivalence.)
The next result is known as the classification theorem of compact surfaces.
Theorem 5.18 (D EHN

AND

H EEGAARD (1907)). Let X be a compact surface. Then

(i) if X is orientable, then it is topologically equivalent to a sphere S = S0 or a connected
sum of tori: Sn = S1 #S1 # . . . #S1 for some n ≥ 1, where S1 is a torus.
(ii) if X is nonorientable, then X is topologically equivalent to a connected sum of projective
planes: Pn = P#P# . . . #P for some n ≥ 1, where P is a projective plane.
It is often difficult to imagine how a figure (say, a graph) can be drawn in a surface. There is a helpful, and difficult to prove, result due to R ADÓ (1920), stating that
every compact surface (orientable or not) has a description by a plane model, which
consists of a polygon in the plane such that
• each edge of the polygon is labelled by a letter,
• each letter is a label of exactly two edges of the polygon, and
• each edge is given an orientation (clockwise or counter clockwise).
Given a plane model M, a compact surface is obtained by gluing together the edges
having the same label in the direction that they have.
a
a

a
b b

b
Sphere

a
b b

a
Torus

a
b a

a
Klein bottle

b

b
Projective plane

From a plane model one can easily determine if the surface is oriented or not. It
is nonoriented if and only if, for some label a, the edges labelled by a have the same
direction when read clockwise. (This corresponds to the Möbius band.)
A plane model, and thus a compact surface, can also be represented by a (circular)
word by reading the model clockwise, and concatenating the labels with the convention that a−1 is chosen if the direction of the edge is counter clockwise. Hence, the
sphere is represented by the word abb−1 a−1 , the torus by aba−1 b−1 , the Klein bottle
by aba−1 b and the projective plane by abb−1 a.


78

These surfaces, as do the other surfaces, have many
other plane models and representing words as well.
A word representing a connected sum of two surfaces,
represented by words W1 and W2 , is obtained by concatenating these words to W1 W2 . By studying the relations of the representing words, Theorem 5.18 can be
proved.

Klein bottle

Drawing a graph (or any ﬁgure) in a surface can be elaborated compared to drawing in a plane model, where a line that enters an edge of the polygon must continue
by the corresponding point of the other edge with the same label (since these points
are identiﬁed when we glue the edges together).
e4

Example 5.5. On the right we have drawn K6 in
the Klein bottle. The black dots indicate, where
the lines enter and leave the edges of the plane
model. Recall that in the plane model for the
Klein bottle the vertical edges of the square
have the same direction.

a

e5

e3

e1

b e2
e1

e2 b
e3

e4

e5
a

Sphere
D EFINITION . In general, if S is a surface, then a graph G has an S-embedding, if G
can be drawn in S without crossing edges.
Let S0 be (the surface of) a sphere. According to the next
theorem a sphere has exactly the same embeddings as do
the plane. In the one direction the claim is obvious: if G is
a planar graph, then it can be drawn in a bounded area of
the plane (without crossing edges), and this bounded area
can be ironed on the surface of a large enough sphere.

¡
¢£

Clearly, if a graph can be embedded in one sphere, then it can be embedded in any
sphere – the size of the sphere is of no importance. On the other hand, if G is embeddable in a sphere S0 , then there is a small area of the sphere, where there are no
points of the edges. We then puncture the sphere at this area, and stretch it open until
it looks like a region of the plane. In this process no crossings of edges can be created,
and hence G is planar.
Another way to see this is to use a projection of the sphere to a plane:


79

Theorem 5.19. A graph G has an S0 -embedding if and only if it is planar.
Therefore instead of planar embeddings we can equally well study embeddings of
graphs in a sphere. This is sometimes convenient, since the sphere is closed and it has
no boundaries. Most importantly, a planar graph drawn in a sphere has no exterior
face – all faces are bounded (by edges).
If a sphere is deformed by pressing or stretching, its embeddability properties will
remain the same. In topological terms the surface has been distorted by a continuous
transformation.

Torus

¡
¢£
¡ ¡ ¡
¢£ ¢£ ¢£
¡ ¡
¢£ ¢£

Consider next a surface which is obtained from the sphere
S0 by pressing a hole in it. This is a torus S1 (or an orientable surface of genus 1). The S1 -embeddable graphs
are said to have genus equal to 1.

Sometimes it is easier to consider handles than holes: a torus S1 can be deformed (by
a continuous transformation) into a sphere with a handle.

If a graph G is S1 -embeddable, then it can be drawn in any one of the above surfaces
without crossing edges.

1

1

5

Example 5.6. The smallest nonplanar graphs K5 and
K3,3 have genus 1. Also, K7 has genus 1 as can be seen
from the plane model (of the torus) on the right.

2
6
3
7
4
1

1


80

Genus
Let Sn (n ≥ 0) be a sphere with n holes in it. The drawing of an S4 can already be
quite complicated, because we do not put any restrictions on the places of the holes
(except that we must not tear the surface into disjoint parts). However, once again an
Sn can be transformed (topologically) into a sphere with n handles.

¡¢ ¡¢
£¤¥ £¤¥

D EFINITION . We define the genus g( G ) of a graph G as the smallest integer n, for
which G is Sn -embeddable.
For planar graphs, we have g( G ) = 0, and, in particular, g(K4 ) = 0. For K5 , we
have g(K5 ) = 1, since K5 is nonplanar, but is embeddable in a torus. Also, g(K3,3 ) = 1.
The next theorem states that any graph G can be embedded in some surface Sn
with n ≥ 0.
Theorem 5.20. Every graph has a genus.
This result has an easy intuitive verification. Indeed,
consider a graph G and any of its plane (or sphere)
drawing (possibly with many crossing edges) such that
no three edges cross each other in the same point (such
a drawing can be obtained). At each of these crossing
points create a handle so that one of the edges goes below the handle and the other uses the handle to cross
over the first one.
We should note that the above argument does not determine g( G ), only that G can be embedded in some Sn .
However, clearly g( G ) ≤ n, and thus the genus g( G ) of
G exists.
The same handle can be utilized by several edges.

¡¢
£¤¥

¡¢
£¤¥


81

Euler’s formula with genus∗
The drawing of a planar graph G in a sphere has the advantage that the faces of the
embedding are not divided into internal and external. The external face of G becomes
an ‘ordinary face’ after G has been drawn in S0 .
In general, a face of an embedding of G in Sn (with g( G ) = n) is a region of Sn
surrounded by edges of G. Let again ϕ G denote the number of faces of an embedding
of G in Sn . We omit the proof of the next generalization of Euler’s formula.
Theorem 5.21. If G is a connected graph, then
νG − ε G + ϕ G = 2 − 2g( G ) .
If G is a planar graph, then g( G ) = 0, and the above formula is the Euler’s formula
for planar graphs.
D EFINITION . A face of an embedding P( G ) in a surface is a 2-cell, if every simple
closed curve (that does not intersect with itself) can be continuously deformed to a
single point.
The complete graph K4 can be embedded in a torus such that it has a face that is not
a 2-cell. But this is because g(K4 ) = 0, and the genus of the torus is 1. We omit the
proof of the general condition discovered by Y OUNGS:
Theorem 5.22 (Y OUNGS (1963)). The faces of an embedding of a connected graph G in a
surface of genus g( G ) are 2-cells.
Lemma 5.12. For a connected G with νG ≥ 3 we have 3ϕ G ≤ 2ε G .
Proof. If νG = 3, then the claim is trivial. Assume thus that νG ≥ 4. In this case we
need the knowledge that ϕ G is counted in a surface that determines the genus of G
(and in no surface with a larger genus). Now every face has a border of at least three
edges, and, as before, every nonbridge is on the boundary of exactly two faces.
⊔
⊓
Theorem 5.23. For a connected G with νG ≥ 3,
1
1
ε G − (νG − 2) .
6
2
Proof. By the previous lemma, 3ϕ G ≤ 2ε G , and by the generalized Euler’s formula,
ϕ G = ε G − νG + 2 − 2g( G ). Combining these we obtain that 3ε G − 3νG + 6 − 6g( G ) ≤
2ε G , and the claim follows.
⊔
⊓
g( G ) ≥

By this theorem, we can compute lower bounds for the genus g( G ) without drawing any embeddings. As an example, let G = K8 . In this case νG = 8, ε G = 28, and so
5
g( G ) ≥ 3 . Since the genus is always an integer, g( G ) ≥ 2. We deduce that K8 cannot
be embedded in the surface S1 of the torus.
If H ⊆ G, then clearly g( H ) ≤ g( G ), since H is obtained from G by omitting
vertices and edges. In particular,


82

Lemma 5.13. For a graph G of order n, g( G ) ≤ g(Kn ).
For the complete graphs Kn a good lower bound was found early.
Theorem 5.24 (H EAWOOD (1890)). If n ≥ 3, then
g( K n ) ≥

(n − 3)(n − 4)
.
12

Proof. The number of edges in Kn is equal to ε G = 1 n(n − 1). By Theorem 5.23, we
2
obtain g(Kn ) ≥ (1/6)ε G − (1/2)(n − 2) = (1/12)(n − 3)(n − 4) .
⊔
⊓
This result was dramatically improved to obtain
Theorem 5.25 (R INGEL

AND

Y OUNGS (1968)). If n ≥ 3, then
g( K n ) =

(n − 3)(n − 4)
.
12

Therefore g(K6 ) = ⌈3 · 2/12⌉ = ⌈1/2⌉ = 1. Also, g(K7 ) = 1, but g(K8 ) = 2.
By Theorem 5.25,
Theorem 5.26. For all graphs G of order n ≥ 3,
g( G ) ≤

(n − 3)(n − 4)
.
12

Also, we know the exact genus for the complete bipartite graphs:
Theorem 5.27 ( R INGEL (1965)). For the complete bipartite graphs,
g(Km,n ) =

(m − 2)(n − 2)
.
4

Chromatic numbers∗
For the planar graphs G, the proof of the 4-Colour Theorem, χ( G ) ≤ 4, is extremely
long and difﬁcult. This in mind, it is surprising that the generalization of the 4-Colour
Theorem for genus ≥ 1 is much easier. H EAWOOD proved a hundred years ago:
Theorem 5.28 (H EAWOOD ). If g( G ) = g ≥ 1, then
χ( G ) ≤

7+

1 + 48g
.
2

Notice that for g = 0 this theorem would be the 4-colour theorem. H EAWOOD
proved it ‘only’ for g ≥ 1.
Using the result of R INGEL AND Y OUNGS and some elementary computations we
can prove that the above theorem is the best possible.


83

Theorem 5.29. For each g ≥ 1, there exists a graph G with genus g( G ) = g so that
χ( G ) =

7+

1 + 48g
2

.

If a nonplanar graph G can be embedded in a torus, then g( G ) = 1, and χ( G ) ≤
⌊(7 + 1 + 48g)/2⌋ = 7. Moreover, for G = K7 we have that χ(K7 ) = 7 and g(K7 ) =
1.

Three dimensions∗
Every graph can be drawn without crossing edges in the 3-dimensional space. Such
a drawing is called spatial embedding of the graph. Indeed, such an embedding can
be achieved by putting all vertices of G on a line, and then drawing the edges in
different planes that contain the line. Alternatively, the vertices of G can be put in a
sphere, and drawing the edges as straight lines crossing the sphere inside.
A spatial embedding of a graph G is said to have linked cycles, if two cycles of
G form a link (they cannot be separated in the space). By C ONWAY and G ORDON in
1983 every spatial embedding of K6 contains linked cycles.
It was shown by R OBERTSON , S EYMOUR AND T HOMAS (1993) that there is a set
of 7 graphs such that a graph G has a spatial embedding without linked cycles if and
only if G does not have a minor belonging to this set.
This family of forbidden graphs was originally found by S ACHS (without proof),
and it contains K6 and the Petersen graph. Every graph in the set has 15 edges, which
is curious.
For further results and proofs concerning graphs in surfaces, see
B. M OHAR

AND

C. T HOMASSEN, “Graphs on Surfaces”, Johns Hopkins, 2001.

6
Directed Graphs
6.1 Digraphs
In some problems the relation between the objects
is not symmetric. For these cases we need directed
graphs, where the edges are oriented from one vertex
to another.
As an example consider a map of a small town. Can
you make the streets one-way, and still be able to drive
from one house to another (or exit the town)?

Definitions
D EFINITION . A digraph (or a directed graph) D = (VD , ED ) consists of the vertices
VD and (directed) edges ED ⊆ VD × VD (without loops vv). We still write uv for
(u, v), but note that now uv = vu. For each pair e = uv define the inverse of e as
e−1 = vu (= (v, u)).
Note that e ∈ D does not imply e−1 ∈ D.
D EFINITION . Let D be a digraph. Then A is its
• subdigraph, if VA ⊆ VD and E A ⊆ ED ,
• induced subdigraph, A = D [ X ], if VA = X and E A = ED ∩ ( X × X ).
The underlying graph U ( D ) of a digraph D is
the graph on VD such that if e ∈ D, then the
undirected edge with the same ends is in U ( D ).
A digraph D is an orientation of a graph G, if G = U ( D ) and e ∈ D implies
∈ D. In this case, D is said to be an oriented graph.
/

e −1

⋆
D EFINITION . Let D be a digraph. A walk W = e1 e2 . . . ek : u − v of U ( D ) is a directed
→
walk, if ei ∈ D for all i ∈ [1, k]. Similarly, we define directed paths and directed cycles
as directed walks and closed directed walks without repetitions of vertices.
⋆
The digraph D is di-connected, if, for all u = v, there exist directed paths u − v
→
⋆
and v − u. The maximal induced di-connected subdigraphs are the di-components
→
of D.

6.1 Digraphs

85

Note that a graph G = U ( D ) might be connected, although the digraph D is not
di-connected.
D EFINITION . The indegree and the outdegree of a vertex are deﬁned as follows
I
dD (v) = |{e ∈ D | e = xv}|,

dO (v) = |{e ∈ D | e = vx}| .
D

We have the following handshaking lemma. (You offer and accept a handshake.)
Lemma 6.1. Let D be a digraph. Then
I
∑ dD (v) = |D| = ∑ dO (v).
D
v∈ D

v∈ D

Directed paths
The relationship between paths and directed paths is in
general rather complicated. This digraph has a path of
length ﬁve, but its directed paths are of length one.
There is a nice connection between the lengths of directed paths and the chromatic
number χ( D ) = χ(U ( D )).
Theorem 6.1 (R OY (1967),G ALLAI (1968)). A digraph D has a directed path of length
χ( D ) − 1.
Proof. Let A ⊆ ED be a minimal set of edges such that the subdigraph D − A contains
no directed cycles. Let k be the length of the longest directed path in D − A.
For each vertex v ∈ D, assign a colour α(v) = i, if a longest directed path from v
has length i − 1 in D − A. Here 1 ≤ i ≤ k + 1.
⋆
First we observe that if P = e1 e2 . . . er (r ≥ 1) is any directed path u − v in D − A,
→
⋆
then α(u) = α(v). Indeed, if α(v) = i, then there exists a directed path Q : v − w of
→
length i − 1, and PQ is a directed path, since D − A does not contain directed cycles.
⋆
Since PQ : u − w, α(u) = i = α(v). In particular, if e = uv ∈ D − A, then α(u) = α(v).
→
Consider then an edge e = vu ∈ A. By the minimality of A, ( D − A) + e contains a
⋆
⋆
directed cycle C : u − v − u, where the part u − v is a directed path in D − A, and
→ →
→
hence α(u) = α(v). This shows that α is a proper colouring of U ( D ), and therefore
χ( D ) ≤ k + 1, that is, k ≥ χ( D ) − 1.
⊔
⊓
The bound χ( D ) − 1 is the best possible in the following sense:
Theorem 6.2. Every graph G has an orientation D, where the longest directed paths have
lengths χ( G ) − 1.

6.1 Digraphs

86

Proof. Let k = χ( G ) and let α be a proper k-colouring of G. As usual the set of colours
is [1, k]. We orient each edge uv ∈ G by setting uv ∈ D, if α(u) < α(v). Clearly, the so
obtained orientation D has no directed paths of length ≥ k − 1.
⊔
⊓
D EFINITION . An orientation D of an undirected graph G is acyclic, if it has no directed cycles. Let a( G ) be the number of acyclic orientations of G.
The next result is charming, since χ G (−1) measures the number of proper colourings of G using −1 colours!
Theorem 6.3 (S TANLEY (1973)). Let G be a graph of order n. Then the number of the acyclic
orientations of G is
a( G ) = (−1)n χ G (−1) ,
where χ G is the chromatic polynomial of G.
Proof. The proof is by induction on ε G . First, if G is discrete, then χ G (k) = kn , and
a( G ) = 1 = (−1)n (−1)n = (−1)n χ G (−1) as required.
Now χ G (k) is a polynomial that satisfies the recurrence χ G (k) = χ G−e (k) −
χ G∗e (k). To prove the claim, we show that a( G ) satisfies the same recurrence.
Indeed, if
a( G ) = a( G − e) + a( G ∗ e)

(6.1)

then, by the induction hypothesis,
a( G ) = (−1)n χ G−e (−1) + (−1)n−1 χ G∗e (−1) = (−1)n χ G (−1) .
For (6.1), we observe that every acyclic orientation of G gives an acyclic orientation
of G −e. On the other hand, if D is an acyclic orientation of G −e for e = uv, it extends
to an acyclic orientation of G by putting e1 : u → v or e2 : v → u. Indeed, if D has no
⋆
⋆
directed path u − v, we choose e2 , and if D has no directed path v − u, we choose
→
→
⋆
⋆
e1 . Note that since D is acyclic, it cannot have both ways u − v and v − u.
→
→
We conclude that a( G ) = a( G − e) + b, where b is the number of acyclic orientations D of G −e that extend in both ways e1 and e2 . The acyclic orientations D that
extend in both ways are exactly those that contain
⋆
⋆
neither u − v nor v − u as a directed path.
→
→

(6.2)

Each acyclic orientation of G ∗ e corresponds in a natural way to an acyclic orientation D of G −e that satisfies (6.2). Therefore b = a( G ∗ e), and the proof is completed.
⊔
⊓

6.1 Digraphs

87

One-way traffic
Every graph can be oriented, but the result may not be di-connected. In the oneway traffic problem the resulting orientation should be di-connected, for otherwise
someone is not able to drive home. R OBBINS’ theorem solves this problem.
D EFINITION . A graph G is di-orientable, if there is a di-connected oriented graph D
such that G = U ( D ).
Theorem 6.4 (R OBBINS (1939)). A connected graph G is di-orientable if and only if G has
no bridges.
Proof. If G has a bridge e, then any orientation of G has at least two di-components
(both sides of the bridge).
Suppose then that G has no bridges. Hence G has a cycle C, and a cycle is always
di-orientable. Let then H ⊆ G be maximal such that it has a di-orientation D H . If
H = G, then we are done.
Otherwise, there exists an edge e = vu ∈ G such that
u ∈ H but v ∈ H (because G is connected). The edge e
/
is not a bridge and thus there exists a cycle
⋆
⋆
C ′ = ePQ : v − u − w − v
→ → →

u
P′

P

e

v

Q

w

in G, where w is the last vertex inside H.
⋆
In the di-orientation D H of H there is a directed path P′ : u − w. Now, we orient
→
⋆
e : v − u and the edges of Q in the direction Q : w − v to obtain a directed cycle
→
→
⋆
⋆
′ Q : v − u − w − v. In conclusion, G [V ∪ V ] has a di-orientation, which
eP
→
→
→
H
C
contradicts the maximality assumption on H. This proves the claim.
⊔
⊓
Example 6.1. Let D be a digraph. A directed Euler tour of D is a directed closed walk
that uses each edge exactly once. A directed Euler trail of D is a directed walk that
uses each edge exactly once.
The following two results are left as exercises.
(1) Let D be a digraph such that U ( D ) is connected. Then D has a directed Euler tour if and
I
only if dD (v) = dO (v) for all vertices v.
D
(2) Let D be a digraph such that U ( D ) is connected. Then D has a directed Euler trail if and
I
only if dD (v) = dO (v) for all vertices v with possibly excepting two vertices x, y for which
D
I ( v) − dO ( v)| = 1.
|d D
D
The above results hold equally well for multidigraphs, that is, for directed graphs,
where we allow parallel directed edges between the vertices.
Example 6.2. The following problem was first studied by H UTCHINSON AND W ILF
(1975) with a motivation from DNA sequencing. Consider words over an alphabet
A = {a1 , a2 , . . . , an } of n letters, that is, each word w is a sequence of letters. In the case

6.1 Digraphs

88

of DNA, the letters are A, T, C, G. In a problem instance, we are given nonnegative
integers si and rij for 1 ≤ i, j ≤ n, and the question is: does there exist a word w in
which each letter ai occurs exactly si times, and ai is followed by a j exactly rij times.
For instance, if n = 2, s1 = 3, and r11 = 1, r12 = 2, r21 = 1, r22 = 0, then the word
a1 a2 a1 a1 a2 is a solution to the problem.
Consider a multidigraph D with VD = A for which there are rij edges ai a j . It
is rather obvious that a directed Euler trail of D gives a solution to the sequencing
problem.

Tournaments
D EFINITION . A tournament T is an orientation of a complete graph.
Example 6.3. There are four tournaments of four vertices that are not isomorphic with
each other. (Isomorphism of directed graphs is deﬁned in the obvious way.)

Theorem 6.5 (R ÉDEI (1934)). Every tournament has a directed Hamilton path.
Proof. The chromatic number of Kn is χ(Kn ) = n, and hence by Theorem 6.1, a tournament T of order n has a directed path of length n − 1. This is then a directed Hamilton path visiting each vertex once.
⊔
⊓
The vertices of a tournament can be easily reached from one vertex (sometimes
called the king).
Theorem 6.6 (L AUDAU (1953)). Let v be a vertex of a tournament T of maximum outdegree.
⋆
Then for all u, there is a directed path v − u of length at most two.
→
Proof. Let T be an orientation of Kn , and let dO (v) = d be the maximum outdegree in
T
T. Suppose that there exists an x, for which the directed distance from v to x is at least
three. It follows that xv ∈ T and xu ∈ T for all u with vu ∈ T. But there are d vertices
in A = {y | vy ∈ T }, and thus d + 1 vertices in {y | xy ∈ T } = A ∪ {v}. It follows
that the outdegree of x is d + 1, which contradicts the maximality assumption made
for v.
⊔
⊓
Problem. Ádám’s conjecture states that in every digraph D with a directed cycle there
exists an edge uv the reversal of which decreases the number of directed cycles. Here the new
digraph has the edge vu instead of uv.

6.1 Digraphs

89

Example 6.4. Consider a tournament of n teams that play once against each other, and
suppose that each game has a winner. The situation can be presented as a tournament,
where the vertices correspond to the teams vi , and there is an edge vi v j , if vi won v j
in their mutual game.
D EFINITION . A team v is a winner (there may be more than one winner), if v comes
out with the most victories in the tournament.
Theorem 6.6 states that a winner v either defeated a team u or v defeated a team
that defeated u.
A ranking of a tournament is a linear ordering of the teams vi1 > vi2 > · · · >
vin that should reflect the scoring of the teams. One way of ranking a tournament
could be by a Hamilton path: the ordering can be obtained from a directed Hamilton
→
→
→
path P : vi1 − vi2 − . . . − vin . However, a tournament may have several directed
Hamilton paths, and some of these may do unjust for the ‘real’ winner.
Example 6.5. Consider a tournament of six teams
1, 2, . . . , 6, and let T be the scoring digraph as in the
figure. Here 1 − 2 − 4 − 5 − 6 − 3 is a di→
→
→
→
→
rected Hamilton path, but this extends to a directed
Hamilton cycle (by adding 3 − 1)! So for every team
→
there is a Hamilton path, where it is a winner, and in
another, it is a looser.

1

2

6

3
5

4

Let s1 ( j) = dO ( j) be the winning number of the team j (the number of teams
T
beaten by j). In the above tournament,
s1 (1) = 4, s1 (2) = 3, s1 (3) = 3, s1 (4) = 2, s1 (5) = 2, s1 (6) = 1 .
So, is team 1 the winner? If so, is 2 or 3 next? Define the second-level scoring for each
team by
s2 ( j ) = ∑ s1 ( i ) .
ji∈ T

This tells us how good teams j beat. In our example, we have
s2 (1) = 8, s2 (2) = 5, s2 (3) = 9, s2 (4) = 3, s2 (5) = 4, s2 (6) = 3 .
Now, it seems that 3 is the winner,but 4 and 6 have the same score. We continue by
defining inductively the mth-level scoring by
sm ( j) =

∑ s m −1 (i ) .
ji∈ T

It can be proved (using matrix methods) that for a di-connected tournament with at
least four teams, the level scorings will eventually stabilize in a ranking of the tournament:
there exits an m for which the mth-level scoring gives the same ordering as do the

6.2 Network Flows

90

(m + k)th-level scorings for all k ≥ 1. If T is not di-connected, then the level scoring
should be carried out with respect to the di-components.
In our example the level scoring gives 1 − 3 − 2 − 5 − 4 − 6 as the ranking
→ → → → →
of the tournament.

6.2 Network Flows
Various transportation networks or water pipelines are conveniently represented by
weighted directed graphs. These networks usually possess also some additional requirements. Goods are transported from specific places (warehouses) to final locations (marketing places) through a network of roads. In modelling a transportation
network by a digraph, we must make sure that the number of goods remains the
same at each crossing of the roads. The problem setting for such networks was proposed by T.E. Harris in the 1950s. The connection to Kirchhoff’s Current Law (1847) is
immediate. According to this law, in every electrical network the amount of current
flowing in a vertex equals the amount flowing out that vertex.

Flows
D EFINITION . A network N consists of

4

• an underlying digraph D = (V, E),
• two distinct vertices s and r, called the source and
the sink of N, and
• a capacity function α : V × V → R + (nonnegative
real numbers), for which α(e) = 0, if e ∈ E.
/

5

5
4

s

4

2
2

6

r
5

4

Denote VN = V and EN = E.
Let A ⊆ VN be a set of vertices, and f : VN × VN → R any function such that
f (e) = 0, if e ∈ N. We adopt the following notations:
/

[ A, A] = {e ∈ D | e = uv, u ∈ A, v ∈ A} ,
/
f + ( A) =

∑

f ( e)

and

f − ( A) =

f + (u) =

∑

f ( e) .

e ∈[ A,A ]

e ∈[ A,A]

In particular,

∑

f (uv)

and

v∈ N

f − (u) =

∑

f (vu) .

v∈ N

D EFINITION . A flow in a network N is a function f : VN × VN → R + such that
0 ≤ f (e) ≤ α(e) for all e,

and

f − (v) = f + (v) for all v ∈ {s, r} .
/

Example 6.6. The value f (e) can be taught of as the rate at which transportation actually happens along the channel e which has the maximum capacity α(e). The second
condition states that there should be no loss.

6.2 Network Flows

91

If N = ( D, s, r, α) is a network of water pipes, then
the value α(e) gives the capacity (x m3 /min) of the
pipe e.
The previous network has a flow that is indicated
on the right.

4
4
5

1
3

s

3

0

4

r
4

4

A flow f in N is something that the network can handle. E.g., in the above figure
the source should not try to feed the network the full capacity (11 m3 /min) of its
pipes, because the junctions cannot handle this much water.
D EFINITION . Every network N has a zero flow defined by f (e) = 0 for all e. For a
flow f and each subset A ⊆ VN , define the resultant flow from A and the value of f
as the numbers
val( f A ) = f + ( A) − f − ( A)

val( f ) = val( f s ) (= f + (s) − f − (s)) .

and

A flow f of a network N is a maximum flow, if there does not exist any flow f ′ such
that val( f ) < val( f ′ ).
The value val( f ) of a flow is the overall number of goods that are (to be) transported through the network from the source to the sink. In the above example,
val( f ) = 9.
Lemma 6.2. Let N = ( D, s, r, α) be a network with a flow f .
(i) If A ⊆ N {s, r}, then val( f A ) = 0.
(ii) val( f ) = −val( f r ).
Proof. Let A ⊆ N {s, r}. Then
0=

∑ ( f + (v) − f − (v)) = ∑
v∈ A

v∈ A

f + ( v) −

∑

f − (v) = f + ( A) − f − ( A) = val( f A ),

v∈ A

where the third equality holds since the values of the edges uv with u, v ∈ A cancel
each out.
The second claim is also clear.
⊔
⊓

Improvable flows
Let f be a flow in a network N, and let P = e1 e2 . . . en be an undirected path in N where
an edge ei is along P, if ei = vi vi+1 ∈ N, and against P, if ei = vi+1 vi ∈ N.
We define a nonnegative number ι( P) for P as follows:
ι( P) = min ι(e),
ei

where ι(e) =

α( e) − f ( e)
f ( e)

if e is along P ,
if e is against P.

6.2 Network Flows

92
4

D EFINITION . Let f be a flow in a network N. A
⋆
path P : s − r is ( f -)improvable, if ι( P) > 0.
→

4

5
3

s
4

On the right, the bold path has value ι( P) = 1,
and therefore this path is improvable.

3

1
0

r
4

4

Lemma 6.3. Let N be a network. If f is a maximum flow of N, then it has no improvable
paths.
Proof. Define


 f ( e) + ι( P )

′
f ( e) = f ( e) − ι( P )


f ( e)

if e is along P ,
if e is against P ,
if e is not in P .
4

Then f ′ is a flow, since at each intermediate vertex v ∈ {s, r}, we have ( f ′ )− (v) = ( f ′ )+ (v), and
/
the capacities of the edges are not exceeded. Now
val( f ′ ) = val( f ) + ι( P), since P has exactly one
edge sv ∈ N for the source s. Hence, if ι( P) > 0,
then we can improve the flow.

4
5

2
3

s

2

0

5

r
4

4

⊔
⊓

Max-Flow Min-Cut Theorem
D EFINITION . Let N = ( D, s, r, α) be a network. For a subset S ⊂ VN with s ∈ S and
r ∈ S, let the cut by S be
/

[S] = [S, S] (= {uv ∈ N | u ∈ S, v ∈ S}) .
/
The capacity of the cut [S] is the sum
α[S ] = α+ (S ) =

∑

α( e) .

e ∈[S ]

A cut [S] is a minimum cut, if there is no cut [ R]
with α[ R] < α[S].
Example 6.7. In our original network the capacity
of the cut for the indicated vertices is equal to 10.

4
5
5

2
4

s

4

6

Lemma 6.4. For a flow f and a cut [S] of N,
val( f ) = val( f S ) = f + (S) − f − (S) .

2

r
5

4

6.2 Network Flows

93

Proof. Let S I = S {s}. Now val(S I ) = 0 (since S I ⊆ N {s, r}), and val( f ) =
val( f s ). Hence
val( f S ) =val( f s ) −

∑

f (sv) +

+ val( f S I ) +

∑

f (vs)

v∈S I

v∈S I

∑

f (sv) −

∑

f (vs)

v∈S I

v∈S I

=val( f s ) = val( f ) .
⊔
⊓
Theorem 6.7. For a flow f and any cut [S] of N, val( f ) ≤ α[S]. Furthermore, equality holds
if and only if for each u ∈ S and v ∈ S,
/
(i) if e = uv ∈ N, then f (e) = α(e),
(ii) if e = vu ∈ N, then f (e) = 0.
Proof. By the definition of a flow,
f + (S) =

∑
e ∈[S ]

f ( e) ≤

∑

α( e) = α[S ] ,

e ∈[S ]

and f − (S) ≥ 0. By Lemma 6.4, val( f ) = val( f S ) = f + (S) − f − (S), and hence
val( f ) ≤ α[S], as required. Also, the equality val( f ) = α[S] holds if and only if
(1) f + (S) = α[S] and (2) f − (S) = 0. This holds if and only if f (e) = α(e) for all e ∈ [S]
(since f (e) ≤ α(e)), and
(2) f (e) = 0 for all e = vu with u ∈ S, v ∈ S.
/
This proves the claim.

⊔
⊓

In particular, if f is a maximum flow and [S] a minimum cut, then
val( f ) ≤ α[S] .
Corollary 6.1. If f is a flow and [S] a cut such that val( f ) = α[S], then f is a maximum
flow and [S] a minimum cut.
The following main result of network flows was proved independently by E LIAS ,
F EINSTEIN , S HANNON, by F ORD AND F ULKERSON, and by R OBACKER in 1955 – 56.
The present approach is due to Ford and Fulkerson.
Theorem 6.8. A flow f of a network N is maximum if and only if there are no f -improvable
paths in N.
Proof. By Lemma 6.3, a maximum flow cannot have improvable paths.
Conversely, assume that N contains no f -improvable paths, and let
⋆
S I = {u ∈ N | for some path P : s − u, ι( P) > 0} .
→

6.2 Network Flows

94

Set S = S I ∪ {s}.
Consider an edge e = uv ∈ N, where u ∈ S and v ∈ S. Since u ∈ S, there exists a
/
⋆
⋆
path P : s − u with ι( P) > 0. Moreover, since v ∈ S, ι( Pe) = 0 for the path Pe : s − v.
→
/
→
Therefore ι(e) = 0, and so f (e) = α(e).
By the same argument, for an edge e = vu ∈ N with v ∈ S and u ∈ S, f (e) = 0.
/
By Theorem 6.7, we have val( f ) = α[S]. Corollary 6.1 implies now that f is a
maximum flow (and [S] is a minimum cut).
⊔
⊓
Theorem 6.9. Let N be a network, where the capacity function α : V × V → N has integer
values. Then N has a maximum flow with integer values.
Proof. Let f 0 be the zero flow, f0 (e) = 0 for all e ∈ V × V. A maximum flow is
constructed using Lemma 6.3 by increasing and decreasing the values of the edges
by integers only.
⊔
⊓
The proof of Theorem 6.8 showed also
Theorem 6.10 (Max-Flow Min-Cut). In a network N, the value val( f ) of a maximum flow
equals the capacity α[S] of a minimum cut.

Applications to graphs⋆
The Max-Flow Min-Cut Theorem is a strong result, and many of our previous results
follow from it.
We mention a connection to the Marriage Theorem, Theorem 3.9. For this, let G be
a bipartite graph with a bipartition ( X, Y ), and consider a network N with vertices
{s, r} ∪ X ∪ Y. Let the edges (with their capacities) be sx ∈ N (α(sx) = 1), yr ∈ N
(α(yr) = 1) for all x ∈ X, y ∈ Y together with the edges xy ∈ N (α( xy) = | X | + 1), if
xy ∈ G for x ∈ X, y ∈ Y. Then G has a matching that saturates X if and only if N has
a maximum flow of value | X |. Now Theorem 6.10 gives Theorem 3.9.
Next we apply the theorem to unit networks, where the capacities of the edges
are equal to one (α(e) = 1 for all e ∈ N). We obtain results for (directed) graphs.
Lemma 6.5. Let N be a unit network with source s and sink r.
(i) The value val( f ) of a maximum flow equals the maximum number of edge-disjoint di⋆
rected paths s − r.
→
(ii) The capacity of a minimum cut [S] equals the minimum number of edges whose removal
⋆
destroys the directed connections s − r from s to r.
→
Proof. Exercise.

⊔
⊓

Corollary 6.2. Let u and v be two vertices of a digraph D. The maximum number of edge⋆
disjoint directed paths u − v equals the minimum number of edges, whose removal destroys
→
⋆
all the directed connections u − v from D.
→

6.2 Network Flows

95

Proof. A network N with source s and sink r is obtained by setting the capacities
equal to 1. The claim follows from Lemma 6.5 and Corollary 6.10.
⊔
⊓
Corollary 6.3. Let u and v be two vertices of a graph G. The maximum number of edge⋆
disjoint paths u − v equals the minimum number of edges, whose removal destroys all the
→
⋆
connections u − v from G.
→
Proof. Consider the digraph D that is obtained from G by replacing each (undirected)
edge uv ∈ G by two directed edges uv ∈ D and vu ∈ D. The claim follows then easily
from Corollary 6.2.
⊔
⊓
The next corollary is Menger’s Theorem for edge connectivity.
Corollary 6.4. A graph G is k-edge connected if and only if any two distinct vertices of G are
connected by at least k independent paths.
Proof. The claim follows immediately from Corollary 6.3.

⊔
⊓

Seymour’s 6-flows∗
D EFINITION . A k-flow ( H, α) of an undirected graph G is an orientation H of G together with an edge colouring α : E H → [0, k − 1] such that for all vertices v ∈ V,

∑

α( e) =

e = vu ∈ H

∑

α( f ) ,

(6.3)

f = uv∈ H

that is, the sum of the incoming values equals the sum of the outgoing values. A
k-flow is nowhere zero, if α(e) = 0 for all e ∈ H.
In the k-flows we do not have any source or sink. For convenience, let α(e−1 ) =
−α(e) for all e ∈ H in the orientation H of G so that the condition (6.3) becomes

∑

α( e) = 0 .

(6.4)

e = vu ∈ H
1

1
2

Example 6.8. A graph with a nowhere zero 4-flow.

1

3

2

2

The condition (6.4) generalizes to the subsets A ⊆ VG in a natural way,

∑

α( e) = 0 ,

e ∈[ A,A]

since the values of the edges inside A cancel out each other. In particular,

(6.5)

6.2 Network Flows

96

Lemma 6.6. If G has a nowhere zero k-flow for some k, then G has no bridges.
Tutte’s Problem. It was conjectured by T UTTE (1954) that every bridgeless graph has
a nowhere zero 5-flow. The Petersen graph has a nowhere zero 5-flow but does not
have any nowhere 4-flows, and so 5 is the best one can think of. Tutte’s conjecture
resembles the 4-Colour Theorem, and indeed, the conjecture is known to hold for the
planar graphs. The proof of this uses the 4-Colour Theorem.
In order to fully appreciate Seymour’s result, Theorem 6.11, we mention that it
was proved as late as 1976 (by J AEGER) that every bridgeless G has a nowhere zero
k-flow for some integer k.
S EYMOUR’s remarkable result reads as follows:
Theorem 6.11 (S EYMOUR’s (1981)). Every bridgeless graph has a nowhere zero 6-flow.
Proof. Omitted.

⊔
⊓

D EFINITION . The flow number f ( G ) of a bridgeless graph G is the least integer k for
which G has a nowhere zero k-flow.
Theorem 6.12. A connected graph G has a flow number f ( G ) = 2 if and only if it is eulerian.
Proof. Suppose G is eulerian, and consider an Euler tour W of G. Let D be the orientation of G corresponding to the direction of W. If an edge uv ∈ D, let α(e) = 1. Since
W arrives and leaves each vertex equally many times, the function α is a nowhere
zero 2-flow.
Conversely, let α be a nowhere zero 2-flow of an orientation D of G. Then necessarily the degrees of the vertices are even, and so G is eulerian.
⊔
⊓
Example 6.9. For each 3-regular bipartite graph G, we have f ( G ) ≤ 3. Indeed, let
G be ( X, Y )-bipartite. By Corollary 3.1, a 3-regular graph has a perfect matching M.
Orient the edges e ∈ M from X to Y, and set α(e) = 2. Orient the edges e ∈ M from
/
Y to X, and set α(e) = 1. Since each x ∈ X has exactly one neighbour y1 ∈ Y such
that xy1 ∈ M, and two neighbours y2 , y3 ∈ Y such that xy2 , xy3 ∈ M, we have that
/
f ( G ) ≤ 3.
Theorem 6.13. We have f (K4 ) = 4, and if n > 4, then
f (Kn ) =
Proof. Exercise.

2
3

if n is odd ,
if n is even .

⊔
⊓

Index

2-sets, 4
acyclic, 18
digraph, 86
addressable, 23
adjacency matrix, 6
adjacent, 4
augmented path, 35
available colour, 43, 53
bipartite, 16
bond, 24
boundary, 62
bridge, 18
capacity, 92
capacity function, 90
choosable, 72
chromatic number χ( G ), 53
chromatic polynomial χ G , 57
closed walk, 12
colouring, 53
complement G, 10
complete bipartite Km,k , 16
complete graph Kn , 10
connected, 24
connected (component), 13
connected sum, 77
connectivity number κ ( G ), 24
contracted vertex, 58
critical, 54
crossing number, 68
cube, 11
cut (in a network), 92
cut vertex, 24
cycle, 12
degree d G (v), 7
di-connected di-component, 84

di-orientable, 87
digraph, 84
directed
Euler tour, trail, 87
walk, path, cycle, 84
directed graphs (digraph), 5
disconnected, 13
disconnecting set, 27
discrete graph, 10
disjoint walks, 12
distance, 13
distance function, 5
edge, 4
edge chromatic number χ′ ( G ), 43
edge colouring, 5, 43
edge cut, 24
embedding, 78
end (of a path), 12
end (of an edge), 4
Euler trail, tour, 29
Euler’s formula, 63
eulerian, 29
even component, 38
even cycle, 12
exterior: face, boundary,vertex,edge, 62
face, 62
fan, 28
ﬂow, 90
ﬂow number, 96
forest, 18
genus, 80
graph, 4
graphical sequence, 9
Hamilton path, cycle, 31
hamiltonian, 31

Index
Hamming distance, 22
homeomorphic, 76
improvable (path), 92
improvement (colouring), 45
incident colours, 43
indegree, 85
independent paths, 12
induced subdigraph, 84
induced subgraph, 8
interior: face, vertex, edge, 62
intersection graph, 6
inverse pair, 84
inverse walk (path), 12
isolated vertex, 7
isomorphic, 5
join of walks, 12
king, 88
kiss (circles), 73
Kuratowski graph, 66
latin rectangle, 38
latin square, 38
leaf, 7
line segment graph, 74
linked cycles, 83
list chromatic number, 72
list colouring, 72
loop, 4
Möbius band, 77
map, 70
matching, 35
maximal planar graph, 64
maximum
degree ∆ ( G ), 7
ﬂow, 91
matching, 35
minimum
cut, 92
degree δ( G ), 7
weighted distance, 13
minor, 74
monochromatic, 49
multigraph, 4
near-triangulation, 72
neighbour, 4
neighbourhood NG (v), 7
network, 90
nontrivial graph, 10
NP-complete problems, 3

98
odd component, 38
odd cycle, 12
optimal colouring, 45
order νG , 4
orientable surface, 76
orientation, 84
oriented, 76
oriented graph, 84
outdegree, 85
parallel edges, 4
partition, 3
path, 12
perfect matching, 35
Petersen graph, 11, 32
planar graph, 61
plane embedding, 61
plane model, 77
proper colouring, 43, 53
Ramsey number, 49
ranking, 89
regular graph, 10
resultant ﬂow, 91
same parity, 3
saturate (matching), 35
separates, 24
separating set, 24
Shannon’s switching game, 20
sink and source, 90
size ε G , 4
spanning subgraph, 8
spanning tree, 20
spatial embedding, 83
sphere, 78
sphere with a handle, 79
stable matching, 41
stable set, 16
subdigraph, 84
subdivision, 61
subgraph, 8
surface, 76
symmetric difference, 3
topologically equivalent, 76
torus, 79
tournament, 88
trail, 29
transversal, 38
tree, 18
triangle, 76
triangle-free, 53
triangulation, 76
trivial graph, 10

Index
trivial path, 12
2-cell, 81
2-switch, 9

99
vertex colouring, 5

underlying digraph, 90
underlying graph, 84
unit networks, 94

walk, 11
weight, 13
weight function, 5
wheel, 52
winning number, 89

vertex, 4

zero ﬂow, 91

Graphtheory

More Related Content

What's hot (20)

Similar to Graphtheory (20)

More from Praveen Kumar (20)

Recently uploaded (20)

Graphtheory