Graph Theory

Lectured by I. B. Leader, Michaelmas Term 2007

Chapter 1 Introduction 1

Chapter 2 Connectivity and Matchings 9

Chapter 3 Extremal Problems 15

Chapter 4 Colourings 21

Chapter 5 Ramsey Theory 29

Chapter 6 Random Graphs 34

Chapter 7 Algebraic Methods 40

Examples Sheets

Last updated: Tue 21st Aug, 2012

Please let me know of any corrections: [email protected]
 Course schedule

GRAPH THEORY (D) 24 lectures, Michaelmas term

No specific prerequisites.

Introduction
Basic definitions. Trees and spanning trees. Bipartite graphs. Euler circuits. Elementary
properties of planar graphs. Statement of Kuratowski’s theorem. [3]

Connectivity and matchings

Matchings in bipartite graphs; Hall’s theorem and its variants. Connectivity and Menger’s
theorem. [3]

Extremal graph theory

Long paths, long cycles and Hamilton cycles. Complete subgraphs and Turan’s theorem.
Bipartite subgraphs and the problem of Zarankiewicz. The Erdős-Stone theorem; *sketch of
proof*. [5]

Eigenvalue methods
The adjacency matrix and the Laplacian. Strongly regular graphs. [2]

Graph colouring
Vertex and edge colourings; simple bounds. The chromatic polynomial. The theorems of
Brooks and Vizing. Equivalent forms of the four colour theorem; the five colour theorem.
Heawood’s theorem for surfaces; the torus and the Klein bottle. [5]

Ramsey theory
Ramsey’s theorem (finite and infinite forms). Upper bounds for Ramsey numbers. [3]

Probabilistic methods
Basic notions; lower bounds for Ramsey numbers. The model G(n, p); graphs of large girth
and large chromatic number. The clique number. [3]

Appropriate books

B.Bollobás Modern Graph Theory. Springer 1998 (£26.00 paperback).

J.A.Bondy and U.S.R.Murty Graph Theory with Applications. Elsevier 1976

(available online via http://www.ecp6.jussieu.fr/pageperso/bondy/bondy.html).

R.Diestel Graph Theory. Springer 2000 (£38.50 paperback).

D.West Introduction to Graph Theory. Prentice Hall 1999 (£45.99 hardback).

 Chapter 1 : Introduction

A graph is a pair (V, E), where V is a set and E is a subset of V (2) = {{x, y} : x, y ∈ V, x 6= y},
the set of unordered pairs from V .
x1s sx2
@ s x5
E.g., @  has V = {x1 , x2 , x3 , x4 , x5 }, E = {x1 x2 , x1 x4 , x2 x3 , x3 x4 , x4 x5 }.
s @s
x3 x4

Notes. 1. No loops: s


2. No multiple edges: s s

3. No directed edges: s > s

Unless otherwise stated, V is finite.

We call V = V (G) the vertex set of G, and E = E(G) the edge set of G.

The order of G is |G| = |V (G)|, and the size is e(G) = |E(G)|.

Often write, say, x ∈ G to mean x ∈ V (G).

Examples.

1. The empty graph En .

x1 s s x2
V = {x1 , . . ., xn }, E = ∅. s s
xn . . . x3
So |G| = n, e(G) = 0.

2. The complete graph Kn .

s x2
(2) x1  sa
AH
@! Hs x3
V = {x1 , . . ., xn }, E = V .
@A! a!

A@
A
s !
A@
Aa @As
a
 
! a
n xnA
A @
A
x4
So |G| = n, e(G) = . ... ...
2

3. The path Pn of length n

sx2 sx4 sxn+1
V = {x1 , . . ., xn+1 }, @ ..
s @s s
.
E = {x1 x2 , . . ., xn xn+1 } = {xi , xi+1 : 1 6 i 6 n}. x1 x3 xn

So |G| = n + 1, e(G) = n.

4. The cycle Cn of length n. s x2

x1 sHHs x3

A
V = {x1 , . . ., xn }, E = {xi , xi+1 : 1 6 i 6 n − 1} ∪ {xn x1 }. s
As
xnA
x4
So |G| = n, e(G) = n. ... ...

1
Graphs G = (V, E) and H = (V ′ , E ′ ) are isomorphic if there is a bijection f : V → V ′ such
that xy ∈ E ⇐⇒ f (x)f (y) ∈ E ′ .
s s
s  BB s
#c
s
# cs 
E.g., B  and ZZ  B  are isomorphic.
Bs s s ZBs

Say that H = (V ′ , E ′ ) is a subgraph of G = (V, E) if V ′ ⊂ V and E ′ ⊂ E.

E.g., Cn is a subgraph of Kn .

For xy ∈ E, write G − xy for the graph (V, E \ {xy}).

For xy ∈
/ E, write G + xy for the graph (V, E ∪ {xy}).

If xy ∈ E, say x, y are adjacent or neighbours.

The neighbourhood of x is Γ(x) = {y ∈ V : xy ∈ E}, and the degree of x is d(x) = |Γ(x)|.

x1s sx2
E.g., in @ s x5 we have Γ(x4 ) = {x1 , x3 , x5 }, so d(x4 ) = 3.
@ 
s @s
x3 x4

If V (G) = {x1 , . . ., xn }, the degree sequence of G is d(x1 ), . . ., d(xn ).

The maximum degree of G is ∆(G) = max d(xi ).

16i6n

The minimum degree of G is δ(G) = min d(xi ).

16i6n

E.g., the above example has degree sequence: 2, 2, 2, 3, 1. So δ(G) = 1 and ∆(G) = 3.

Say G is regular of degree k, or k-regular, if d(x) = k for all x ∈ V (G).

E.g., Cn is regular of degree 2, and Kn is (n − 1)-regular.

In a graph G, an x-y path is a sequence x1 , . . ., xk (k > 1) of distinct vertices of G with

x1 = x, xk = y and xi xi+1 ∈ E for all 1 6 i 6 k − 1. It has length k − 1.

Say G is connected if for all x, y ∈ V , there is an x-y path in G.

s s s s s s

A
A
s s
s As s s s
As
E.g., is connected, but is not connected.

Write x ∼ y if there is an x-y path. Note, ∼ is an equivalence relation.

s x3
If x1 . . .xk and xk . . .xl are paths, then x1 , . . ., xl need not be a path, e.g., s s s
x1 x2 , x4 x5

– but it contains a path from x1 to xl . (E.g., choose minimal 1 6 i 6 l for which there is
some k 6 j 6 l with xi = xj , and then x1 , . . ., xi , xj+1 , . . ., xl is a path.)

2
The equivalence classes of ∼ are the components of G.

So the component of x is {y : ∃ x-y path in G}.

s s s

A has two components.
s s s
As
E.g.,

A walk is a sequence x1 , . . ., xk such that xi xi+1 ∈ E(G) for all 1 6 i 6 k − 1.

Thus G has an x-y walk ⇐⇒ G has an x-y path.

Trees
A graph is acyclic if it has no cycle.

A tree is a connected acyclic graph.

In a tree T , a vertex x with d(x) = 1 is called a leaf or endvertex.

Proposition 1. Let G be a graph. The following are equivalent.

(a) G is a tree.

(b) G is minimal connected (i.e., G connected, G − xy disconnected for all xy ∈ E).

(c) G is maximal acyclic (i.e., G acyclic, G + xy has a cycle for all xy ∈

/ E).

Proof. (a) ⇒ (b). If G − xy is connected, then it has an x-y path P , but then P yx is a
cycle in G. /\/\

(b) ⇒ (a). If G has a cycle C, choose xy ∈ E(C). Then G − xy is connected, because

for all a, b ∈ V , we have an a-b path in G, and if it used edge xy then replace xy with
C − xy to obtain an a-b walk in G − xy. /\/\

(a) ⇒ (c). We have x-y path P in G, so G + xy contains the cycle P yx.

(c) ⇒ (a). If G is not connected, choose x, y in different components of G, then G + xy

is acyclic. /\/\ 

Proposition 2. Let T be a tree (|T | > 2). Then T has a leaf.

Proof. Let P = x1 . . .xk be a longest path in T (which exists as we have a finite graph).
Then Γ(xk ) ⊂ P (by maximality of P ), and Γ(xk ) ∩ P = {xk−1 }, since T is acyclic.
So d(xk ) = 1. 

Remark. The proof actually shows there are > 2 leaves, as the same proof works for x1 .

Alternatively, if T has no leaf, “go for a walk”. Choose some x1 , x2 ∈ E and then choose
x3 , x4 , x5 , . . . as follows. Having chosen x1 , . . ., xk , choose xk+1 ∈ Γ(xk ) \ {xk−1 }.
This must repeat (as T finite), giving a cycle. /\/\ .

3
Proposition 3. Let T be a tree on n vertices (n > 1). Then e(T ) = n − 1.

For G a graph, W ⊂ V , write G[W ] for the subgraph spanned by W . That is, G[W ] has
vertex set W , edge set E ∩ W (2) . For x ∈ V , write G − x for G[V \ {x}].

Proof of 3. Induction on n. Done if n = 1.

Given T on n vertices (n > 2), choose a leaf x. Then G−x is a tree, with |G−x| = n−1,
so e(G − x) = n − 2, by induction. Thus e(G) = e(G − x) + 1 = n − 1. 

For G a connected graph, a spanning tree of G is a subgraph T of G, with V (T ) = V (G),

that is a tree.
s s s s s s s s
s s s s s s s s
E.g., has as a spanning tree.

Clearly every connected G does have a spanning tree: just remove edges until we get a
minimal connected graph.

For a non-inductive proof of Proposition 3 we’ll show that any connected G has a spanning
tree T on n − 1 edges (then done as if G is a tree, then the only spanning tree of G is G
itself, e.g., by minimal connectedness).

For x, y ∈ G, the distance from x to y, d(x, y) is the shortest length of any x-y path.

We construct our spanning tree in G, as follows.

Fix x0 ∈ G, then for each x ∈ V \ {x0 } choose a shortest x-x0 path, say xx′ . . .x0 .
(So d(x′ , x0 ) = d(x, x0 ) − 1.)

Let T consist of all the xx′ for x ∈ V \ {x0 }. Then e(T ) = n − 1.

T is connected, as for all x, xx′ x′′ . . . forms a path to x0 .

T is acyclic. Suppose T has a cycle C. On C choose x at maximum distance from

x0 , say d(x, x0 ) = k. Then both neighbours of x are at distance 6 k from x0 , /\/\ of
construction. 

Notes. 1. A forest is an acyclic graph.

Thus G is a forest ⇐⇒ every component of G is a tree.

2. For G a connected graph, xy ∈ E, say xy is a bridge if G − xy is disconnected.

s sx sy

A
s s
s As
Thus G is a tree =⇒ every edge is a bridge. And, e.g., in xy is a bridge.

3. For G connected, x ∈ G, say x is a cutvertex if G − x is disconnected.

Clearly, if G has a bridge then it has a cutvertex (for |G| > 2).
s s
s
Q
s Qs
Q
The converse is false, e.g.,

4
Bipartite Graphs
A graph G is bipartite on vertex classes V1 and V2 if V1 , V2 partition V (i.e., V1 ∪ V2 = V ,
V1 ∩ V2 = ∅) and E(G) ⊂ {xy : x ∈ V1 , y ∈ V2 } (i.e., no edges inside V1 , V2 ).
 
s s
V1 s s V
s s
E.g., a path: 2
 

The complete bipartite graph Kn,m has |V1 | = n, |V2 | = m, and E = {xy : x ∈ V1 , y ∈ V2 }.

So e(Kn,m ) = nm.

(s
s h
( ( (


(
hs
h
H h(h
s h
E.g., so K2,3 is: 
H(
( (

( H
Hs
h

h H
h
h


Proposition 4. G bipartite ⇐⇒ G has no odd cycle.

A circuit in a graph G is a closed walk (i.e., a walk of the form x1 . . .xk where x1 = xk ).

Note that if G has an odd circuit, then it has an odd cycle. Indeed, if x1 , . . .xk x1 is an odd
circuit and xi = xj (some 1 6 i 6 j 6 k), then one of xi xi+1 . . .xj and xj xj+1 . . .xk x1 . . .xi is
an odd circuit. Then done by induction on k.

Proof of 4. (⇒) The vertices in a cycle must alternate between V1 and V2 .

(⇐) Wlog G is connected (as if each component of G is bipartite, then so is G).

Fix x1 ∈ G and put V1 = {x ∈ G : d(x, x1 ) even}, V2 = {x ∈ G : d(x, x1 ) odd}.
(The only possible choices.) If we had x, y ∈ V1 or x, y ∈ V2 with xy ∈ E, then xy
together with shortest paths from x to x1 and y to x1 gives an odd circuit. /\/\ 

Planar Graphs

A graph G is planar if it can be drawn in the plane without crossing edges. A plane graph
is such a drawing.
s s s s
@ @
E.g., @ is not a plane graph, but @ is. So K4 is planar.
s @s s @s

Examples.
s s s
s @s @s
@ @
1. Any path is planar:
s
s H Hs
2. Any cycle is planar: s s
H Hs
s s
s
3. The empty graph is planar: s s
4. K4 is planar (as shown above).

5
Which graphs are planar? How do we check if a given graph is not planar?

Given a graph G, R2 − G (i.e., the plane with G removed) splits up into connected regions
called faces. The boundary of a face consists of the vertices and edges that touch it.
F4
s s F2
?Z
E.g.,
@
F5 - @s Zs F1
Here, the boundaries of F2 , F3 , F4 , F5 , F6 consist of
s 6 s

@@
I
@
 3-cycles, and of F1 consists of a 5-cycle.
F3
F6
s s
@s s
Warning. 1. The boundary of a face need not be a cycle: s
s F s
s s
s s
2. The boundary of a face need not even be connected: s
s F s
s s
@
s @s
3. The two faces on either side of an edge may be the same:
F

Formal bit: 
is officially  Q
Q

Let x, y ∈ R2 , x 6= y. A polygonal arc from x to y is a finite union of (closed) straight

line segments, x1 x2 ∪ x2 x3 ∪ . . . ∪ xn−1 xn , with x = x1 , y = xn , that are disjoint except for
xi−1 xi ∩ xi xi+1 = {xi }.

For a graph G with vertex set {v1 , . . ., vn }, a drawing of G consists of distinct points
x1 , . . ., xn ∈ R2 together with, for each vi vj ∈ E, a polygonal arc pij from xi to xj such
that pij ∩ pkl = ∅ if i, j, k, l distinct, and pij ∩ pjk = {xj }.

For x, y ∈ R2 − G, write x ∼ y if there exists a polygonal arc in R2 − G from x to y. The

components (equivalence classes) of ∼ are called the faces. The boundary of a face consists
of G intersect its closure.

We’ll assume various facts about R2 , like “a cycle has two faces”, or “a boundary of a face
consists of vertices and (whole) edges”. (Proved by induction on the total number of straight
line segments in the drawing.)

End of formal bit!

Remarks. 1. Every tree is planar, with exactly one face. (Proof by induction, via removing
a leaf.)

2. A planar graph may have genuinely different drawings.

s s s s
E.g., s JJs
has a face with a 6-cycle as boundary, but s JJs
doesn’t.
JJs s JJs s

6
Is the number of faces fixed? Yes:

Theorem 5 (Euler’s Formula). Let G be a connected plane graph with n vertices,

m edges and f faces. Then n − m + f = 2.

Note. We need G to be connected. E.g., En has n vertices, 0 edges, 1 face.

Proof of 5. If G has no cycles, then G is a tree (acyclic and connected), so m = n − 1,

f = 1, so n − m + f = 2.

If G has a cycle, choose an edge e that is on a cycle. Then G − e is connected (as e

was on a cycle), and has n vertices, m − 1 edges, and f − 1 faces (as e was on a cycle).
But by induction, n − (m − 1) + (f − 1) = 2, so n − m + f = 2. 

Theorem 6. Let G be a plane graph on n vertices (n > 3) with m edges. Then m 6 3n − 6.

Notes. 
1. This
 is2a linear bound on m – whereas in general a graph could have anything up
n n −n
to = edges.
2 2
s `
`
P
H
@HPP```
s @sH
PsP
Ps . . . (line of n − 2 vertices)
``
2. This bound is best possible, e.g.: H 



 
s 

Here, the number of edges is (n − 3) + 2(n − 2) + 1 = 3n − 6.

Proof of 6. Wlog, G connected. (If not, add edges to make it so.) Thus n − m + f = 2.

If we sum, for each face f , the number of edges in the boundary of f , we obtain > 3f ,
because each face has > 3 edges in its boundary (for n > 3 – theorem trivial for n 6 2).
But we also obtain 6 2m (since each edge counted 6 2).

2m 2m m
Thus 3f 6 2m, i.e. f 6 . So n − m + > 2, so 6 n − 2. 
3 3 3
Corollary 7. K5 is not planar.

Proof. n = 5, m = 10, so 10 6 15 − 6. /\/\ 

Hence any graph containing K5 is not planar. E.g., Kn , n > 5.

A subdivision of a graph G is obtained by replacing edges of G with disjoint paths.

s s
#Bc #Bcs s
E.g., K5 : sZ
#  B cs
B 
s#  B cs
Subdivided K5 : ZBs Z B
B Z   s
Bs ZBs Bs ZBs

So any subdivision of K5 is not planar.

7
Proposition 8. K3,3 is not planar.

Remarks. 1. We cannot use Theorem 6, as n = 6, m = 9 holds.

2. So we cannot, e.g., connect up 3 houses to 3 utilities without pipes crossing.

Proof of 8. K3,3 is triangle-free, so if drawn in the plane, we must have > 4 edges on the
boundary of each face. So 4f 6 2m, i.e. f 6 m/2.
m
So n − m + > 2, i.e. m 6 2(n − 2). But n = 6, m = 9. /\/\ 
2

Remark. The girth of a graph is the length of a shortest cycle (with girth ∞ if the graph
has no cycle).
 
g
The proof above yields: if G is planar, girth > g, then m 6 max (n − 2), n − 1 .
g−2
↑
silly bit

Corollary 9. If G contains a subdivision of K5 or K3,3 , then G is not planar. 

Kuratowski’s Theorem says that these are the only obstruction to being planar:

G planar ⇐⇒ G 6⊃ subdivided K5 or K3,3 .

So, to show G planar: draw it.

To show G not planar: find a subdivided K5 or K3,3 .

8
 Chapter 2 : Connectivity and Matchings

Let G be a bipartite graph, with vertex classes X and Y . A matching from X to Y is a set
{xx′ : x ∈ X} of edges of G, such that x 7→ x′ is injective. In other words, it consists of |X|
independent edges (i.e., no vertices in common).

When does G have a matching?

“Matchmaker” terminology: let X = {boys}, Y = {girls}, with x joined to y if x knows y.

Can we pair up each boy with a girl he knows?

Clearly impossible if d(x) = 0 for some x ∈ X, or if there are (distinct) x1 , x2 ∈ X with

Γ(x1 ) = Γ(x2 ) = {y}, some y ∈ Y .
[
For A ⊂ X, write Γ(A) = Γ(x), then we clearly need |Γ(A)| > |A| for all A ⊂ X.
x∈A

Are there any other obstructions to a matching?

Theorem 1 (Hall’s “Marriage” Theorem). Let G be a bipartite graph, with vertex

classes X, Y . Then G has a matching from X to Y ⇐⇒ |Γ(A)| > |A| for all A ⊂ X.
(This is “Hall’s condition”.)

Proof (1). (⇒) Trivial.

(⇐) Induction on |X|. |X| = 1 is trivial.

We have G, vertex classes X, Y , with |X| > 1, such that |Γ(A)| > |A| for all A ⊂ X.

Question: do we have |Γ(A)| > |A| for all A ⊂ X (A 6= ∅, X)?

If yes, choose any x ∈ X and any y ∈ Γ(x). Let G′ = G − x − y.

Claim. G′ has a matching from X \ {x} to Y \ {y}.

Proof of claim. We need |ΓG′ (A)| > |A| for all A ⊂ X \ {x} (A 6= ∅).
But |ΓG (A)| > |A| + 1, so |ΓG′ (A)| > |A| + 1 − 1.

If no, then there is some A ⊂ X with |Γ(A)| = |A|. Let G′ = G[A ∪ Γ(A)] and
G′′ = G[(X \ A) ∪ (Y \ Γ(A))].

Claim 1. G′ has a matching from A to Γ(A).

Proof of claim. For B ⊂ A, have |ΓG′ (B)| > |B|, because ΓG′ (B) = ΓG (B).

Claim 2. G′′ has a matching from X \ A to Y \ Γ(A).

Proof of claim. For B ⊂ X \ A, consider A ∪ B. Have |ΓG (A ∪ B)| > |A| + |B|,
so |ΓG (A ∪ B) \ ΓG (A)| > |B|, so |ΓG′′ (B)| > |B|. 

9
Proof (2). Form a directed network as follows: add a source s joined to each x ∈ X by an
edge of capacity 1, and a sink t joined to each y ∈ Y by an edge of capacity 1, and also
direct each xy ∈ E(G) from x to y, capacity ∞ (i.e., some huge number).
∞ s 1
1s - X zX
s 
s 
X
:
- s -s X
- Xs
X
zXXs 
:  : t
s 

Then an integer-valued flow of size |X| is precisely a matching from X to Y . So, by inte-
grality theorem of max-flow min-cut, we just need to show that every cut has size > |X|.
So given a cut {s} ∪ A ∪ B (A ⊂ X, B ⊂ Y ), wlog Γ(A) ⊂ B (else capacity ∞),
so capacity = |X| − |A| + |B| > |X| as |B| > |A|, since B ⊃ Γ(A). 

A matching of deficiency d in a bipartite graph, vertex classes X and Y , consists of |X|−d

independent edges.
 
sh
(s
sh
hhh
(h
( ( ( h
(s
E.g., – a matching of deficiency 1.
s( (( 
h hh
(


Corollary 2 (Defect Hall). Let G be a bipartite graph, vertex classes X, Y . Then G has
a matching of deficiency d from X to Y ⇐⇒ |Γ(A)| > |A| − d for all A ⊂ X.

Proof. (⇒) Trivial.

(⇐) Form G′ by adding d new points to Y , each joined to all of X. Then |ΓG′ (A)| > |A|
for all A ⊂ X, so by Hall there is a matching in G′ , giving a matching of deficiency d
in G. 

(In terms of boys and girls: add d imaginary girls to Y , known to all boys. Hall gives a
matching, and at most d boys are paired with imaginary girls, so at least |X| − d are paired
with real girls.)

Let S1 , . . ., Sn be sets. A transversal for S1 , . . ., Sn consists of distinct points x1 , . . ., xn

with xi ∈ Si for all i.

E.g., {1, 2, 3}, {3, 4}, {4, 5}, {5} has a transversal 1, 3, 4, 5.

When is there a transversal?

[
Corollary 3. Sets S1 , . . ., Sn have a transversal ⇐⇒ Si > |A| for all A ⊂ {1, . . ., n}.
i∈A

Proof. (⇒) Trivial.

(⇐) Wlog all Si are finite. Form a bipartite graph as follows: X = {1, . . ., n}, Y =
S1 ∪ . . . ∪ Sn , with i ∈ X joined to j ∈ Y if j ∈S
Si . Thus a transversal is precisely a
matching from X to Y . But for A ⊂ X, have i∈A Si > |A|, i.e. |Γ(A)| > |A|, so
done by Hall. 

10
Remarks. 1. Actually, corollary 3 is equivalent to Hall, since a matching in a bipartite
graph G, vertex classes {x1 , . . ., xn } and Y is exactly a transversal for Γ(x1 ), . . ., Γ(xn ).

2.S There is also a defect form: there exists a transversal for all but d of S1 , . . ., Sn ⇐⇒
i∈A Si > |A| − d for all A ⊂ {1, . . ., n}.

A typical application of Hall

Let G be a finite group, H a subgroup of G. Have the left cosets L1 , . . ., Lk (k = |G|/|H|),

say g1 H, . . ., gk H, and have the right cosets R1 , . . ., Rk , say Hg1′ , . . ., Hgk′ .

Can we choose representatives of the left cosets that are also representatives of the right cosets
– that is, g1 , . . ., gk such that g1 H, . . ., gk H are the left cosets and Hg1 , . . ., Hgk are the right
cosets? In other words, we seek a permutation π of {1, . . ., k} such that Li ∩ Rπ(i) 6= ∅ ∀ i.
(“Pair up each left coset with a right coset that it meets.”)

Thus we seek a matching from X to Y in G, where X = {L1 , . . ., Lk }, Y = {R1 , . . ., Rk }, and

joined to Rj if Li ∩ RjS6= ∅. Thus, by Hall, need to check that |Γ(A)| > |A| ∀ A ⊂ X.
Li is S
But i∈A Li = |A||H|, so i∈A Li must meet at least |A| of the Rj , as |Rj | = |H| for all j.
So done. 

Connectivity

Idea. How “connected” is a graph?

A tree is connected, but we can remove a point to disconnect it.

A cycle G is connected, and G − x is connected for all x ∈ G (although we can remove two
points to disconnect it).
s s
s s
s s — G − x − y still connected ∀ x, y ∈ G.
s s

For G connected, |G| > 1, the connectivity of G, κ(G), is the smallest |S| such that
S ⊂ V (G) and G − S is disconnected, or a single point (because of complete graphs).

Say G is k-connected if κ(G) > k.

Thus G is k-connected ⇐⇒ no set of size < k disconnects G (or makes it a single point).

Equivalently, G is k-connected ⇐⇒ |G| > k and no set of size < k disconnects G.

Thus: G 1-connected ⇐⇒ G is connected (|G| > 1)

G 2-connected ⇐⇒ G has no cutvertex (|G| > 2)
G 3-connected ⇐⇒ G cannot be disconnected by removing 2 vertices (|G| > 3)

11
Examples.

1. Any tree T is not 2-connected (|T | > 1)

2. A cycle Cn is 2-connected, but not 3-connected.

s s
3. s s s s is 3-connected.
s s

4. Kn is (n − 1)-connected.
s s
s
H s x
H
sH s
Warning. We can have κ(G − x) > κ(G), e.g.
Hs

Remark. Always have κ(G) 6 δ(G) = minimum degree in G: choose x with d(x) = δ(x)
and remove S = Γ(x). Then |S| = δ(x) and G is disconnected (or a single point).

We know: G connected =⇒ ∀ a, b there is an a-b path in G.

It would be nice if: G k-connected =⇒ ∀ a, b there are k independent (i.e., disjoint vertices
apart from a, b) a-b paths.

For G connected, a, b distinct vertices of G, say S ⊂ V (G) \ {a, b} separates a from b, or is

an a-b separator, if a, b are in different components of G − S (i.e., every a-b path meets S).
s s s  s
E.g., @
s @s s  s
a S b

Theorem 4 (Menger’s Theorem). Let G be a connected graph, and a, b distinct non-

adjacent vertices of G. If all a-b separators have size > k then there exists a family of
k independent a-b paths.

Remarks. 1. Converse trivial: any separator contains > 1 point from each of the k paths.

2. Equivalent form: minimum size of an a-b separator = maximum number of indepen-

dent a-b paths.

3. Need “non-adjacent”, else no separators!

4. Menger generalises Hall. Given bipartite G on X, Y , with |Γ(A)| > |A| ∀ A ⊂ X,

form G′ by adding s (joined to all of X) and t (joined to all of Y ). Then a matching is
precisely a family of X independent s-t paths, so by Menger, it’s enough to check that
every separator has size > |X|. So let S = A ∪ B be a separator, where A ⊂ X, B ⊂ Y .
Then Γ(X \ A) ⊂ B, so |A| + |B| > |Γ(X \ A)| + |A| > |A| + |X \ A| = |X|. 

5. Cannot prove Menger by choosing one point on each path in a maximum-sized

“family of independent a-b paths”.
s s
E.g., a s
 HHs
HH@ s @s
b

12
Proof (1). We may assume that k > 2, as the theorem is trivial for k = 1.

Let k be the minimum size of any a-b separator. We need k independent paths from
a to b. If not possible, take a minimal counterexample (say minimal k, then minimal
e(G) for k). Let S be an a-b separator, |S| = k.

Suppose first that S 6⊂ Γ(a), S 6⊂ Γ(b).

Form G′ from G by replacing the component of G − S containing a by a single point a′

joined to all of S. Then e(G′ ) < e(G), as S 6⊂ Γ(a). In G′ , there is no a′ -b separator of
size < k (else the same set separates a and b in G), so by minimality of e(G) we have
k independent a′ -b paths in G′ . I.e., we have k paths B1 , . . ., Bk from b to S, disjoint
except at b.

Similarly, we have k paths A1 , . . ., Ak from a to S, disjoint except at a. No Ai can meet

a Bj except on S (else S is not a separator), so put the Ai and Bj together to form k
independent a-b paths. /\/\

Now suppose that every a-b separator S of size k has S ⊂ Γ(a) or S ⊂ Γ(b).

We cannot have any x ∈ Γ(a)∩Γ(b), because if so then consider G−x. All a-b separators
in G − x have size > k − 1 (as with x, they separate a from b in G), so by minimality of
k there exist k − 1 independent paths in G − x. Now add axb, and obtain k independent
paths in G. /\/\

Choose a shortest a-b path, say ax1 . . .xr b (r > 2) and consider G − x1 x2 . In G − x1 x2
have a separator S of size k − 1 (by minimality), so S ∪{x1 } and S ∪{x2 } are separators
in G. Since x1 ∈ / Γ(b), we must have S ∪ {x1 } ⊂ Γ(a). And since x2 ∈ / Γ(a), we must
have S ∪ {x2 } ⊂ Γ(b). So S ⊂ Γ(a) ∩ Γ(b), contradicting Γ(a) ∩ Γ(b) = ∅. (If S = ∅
then k = 1.) /\/\ 

Proof (2). We’ll apply vertex capacity form of max-flow min-cut. Form a directed network
by replacing each edge xy by directed edges xy ~ and yx
~ and give each capacity 1. Then
an integer-valued flow of size k is exactly a family of k independent a-b paths. So, by
integrality form of max-flow min-cut, just need that every vertex cut set has size > k,
i.e. every a-b separator has size > k. 

Corollary 5. Let G be a graph, |G| > 1. Then G is k-connected ⇐⇒ for all a, b ∈ G, there
exist k independent a-b paths.

Remark. Sometimes also called “Menger’s Theorem”.

Proof. (⇐) Certainly G is connected, with |G| > k. Also, no set of size < k can disconnect
a from b (else choose a, b in different components).

(⇒) If a, b are non-adjacent, we are done by Menger (as G is k-connected, so no set of

size < k can separate a, b).

If a, b are adjacent, let G′ = G− ab. Then G′ is (k − 1)-connected (as G is k-connected),

so by Menger there exist k − 1 independent a-b paths in G. Now add edge ab as a k th
path. 

13
For G connected, |G| > 1, the edge connectivity λ(G) of G is the smallest size of a set
W ⊂ E(G) such that G − W is disconnected.

Say G is k-edge-connected if λ(G) > k.

Thus: G 1-edge-connected ⇐⇒ G connected (|G| > 1)

G 2-edge-connected ⇐⇒ G connected and has no bridge (|G| > 1)
(Note, different from 2-connected: while bridge ⇒ cutvertex, the converse is false.)

s s s
sH Hs
s @ s @s has λ(G) = 2, κ(G) = 1.
@
E.g., s s has λ(G) = 2, and @
HHs @ @s @s

We always have λ(G) 6 δ(G).

Theorem 6 (Edge version of Menger). Let G be a connected graph, and a, b distinct

vertices of G. Then the minimum size of W ⊂ E(G) separating a from b equals the
maximum number of edge-disjoint a-b paths.

Idea. “View edges of G as vertices, and apply vertex Menger.”

The line graph L(G) of a graph G has vertex set E(G), with e joined to f if they meet
(share a vertex) in G.

Proof of 6. Form G′ from L(G) by adding new vertices a′ , b′ to L(G) with edges a′ e for
each e ∈ E(G) = V (L(G)) with a ∈ e, and edges b′ e for each e ∈ E(G) with b ∈ e.

Then ∃ a-b path in G ⇐⇒ ∃ a′ -b′ path in G′ , and deleting edges in G corresponds to

deleting vertices (not a′ , b′ ) in G′ . So done by vertex Menger applied to G′ . 

s s s
s es s s s
E.g., if G is a s @s @sb
@ e @
g g
s s
then L(G) is
@@s f f

s s s s s
sH
 A

a′ Hs s A
sb′
and G′ is

Corollary 7. Let G be a graph, |G| > 1. Then G is k-edge-connected ⇐⇒ for all distinct
a, b ∈ G there exist k edge-disjoint a-b paths.

Proof. (⇐) Obvious.

(⇒) Edge form of Menger. 

Remark. Or prove Theorem 6 and Corollary 7 by max-flow min-cut (usual edge capacity
version).

14
 Chapter 3 : Extremal Problems

An Euler circuit, or Eulerian circuit, in a graph G is a circuit passing through each edge
exactly once. I.e., x1 , . . ., xk (xk = x1 ) such that if xy ∈ E then there is a unique 1 6 i 6 k−1
with xy = xi xi+1 .

Say G is Eulerian if it has an Euler circuit.

s s
E.g., s A
s
s is Eulerian. But any graph with a bridge is not Eulerian.
AAs
AAs

Which graphs are Eulerian?

Proposition 1. Let G be a connected graph. Then G is Eulerian ⇐⇒ d(x) is even for all
x ∈ G.

Remark. Hence G is Eulerian ⇐⇒ all degrees are even and at most one component contains
an edge.

Proof of 1. (⇒) If an Eulerian circuit passes through x k times, then d(x) = 2k.

(⇐) Use induction on e(G). Done if e(G) = 0.

Given G connected, with e(G) > 0 and d(x) even for all x ∈ G, suppose that G is not
Eulerian and let C be a largest circuit in G with no edges repeated. Note that e(C) > 0
as G has a cycle (since d(x) > 2 for all x ∈ G, so G is not a tree).

Let H be a component of G − E(C) with e(H) > 0. Then H is connected and dH (x)
is even for all x ∈ H (as dG (x) and dC (x) are even for all x). So, by the induction
hypothesis, H has an Euler circuit C ′ . But now C and C ′ are edge disjoint circuits that
share a vertex (we do have V (H) ∩ V (C) 6= ∅ as G is connected), so we can combine
them to form a circuit longer than C. /\/\ 

Let G be a graph of order n. Say G is Hamiltonian if it has a cycle of length n (i.e., a cycle
through all vertices). Such a cycle is called a Hamilton cycle.

There is no nice “⇔” condition known for Hamiltonicity. We can’t have a parity type of
condition, as G Hamiltonian ⇒ G + xy Hamiltonian.
s s
E.g., s
A
s
AAs is Hamiltonian (use the perimeter).


AAs
AAs

But any graph with a cutvertex is not Hamiltonian.

How “large” does a graph on n vertices have to be to ensure it is Hamiltonian?

Silly question. How many edges do we need to ensure that G is Hamiltonian?

This is silly because any x with d(x) = 1 stops G from being Hamiltonian, so take G to
be the complement
(i.e., G = (V, V (2) \ E), for G = (V, E)) of {x1 x2 , x1 x3 , . . ., x1 xn−1 }.
Then e(G) = n2 − (n − 2), but G is not Hamiltonian.

15
Better question. What δ(G) ensures that G is Hamiltonian?
n
E.g., n even, take two disjoint copies of Kn/2 . Then δ(G) = − 1, but G is not Hamiltonian.
2
n−1
E.g., n odd, take two copies of K(n+1)/2 , meeting at a single point. Then δ(G) = , but
2
there is no Hamilton cycle.

n
Theorem 2. Let G be a graph of order n (n > 3) Then δ(G) > =⇒ G is Hamiltonian.
2

Proof. G is connected, since if x, y are non-adjacent vertices then Γ(x), Γ(y) ⊂ V \ {x, y},
so Γ(x) ∩ Γ(y) 6= ∅, as |Γ(x)|, |Γ(y)| > n/2 but |V \ {x, y}| < n.

Let x1 x2 . . .xl be a longest path in G. (Note l > 3 since G is connected and n > 3.)

Wlog G has no cycle of length l, because: if l = n, we’re done; and if l < n then as G
is connected, there exists x ∈
/ cycle adjacent to some y ∈ cycle, which yields a path on
l + 1 points.

Thus xl x1 ∈
/ E. Moreover we cannot have 2 6 i 6 l with x1 xi , xi−1 xl ∈ E, or else we
 xi . . . xl
s- s s  s
have a cycle:
x1 . . . xi−1 -
Now, Γ(x1 ) ⊂ {x2 , . . ., xl } and Γ(xl ) ⊂ {x1 , . . ., xl−1 } (by maximality of the path),
and, by above, Γ(x1 ) is disjoint from Γ+ (xl ) = {xi : 2 6 i 6 l : xi−1 ∈ Γ(xl )}. But
|Γ(x1 )|, |Γ+ (xl )| > n/2 and Γ(x1 ), Γ+ (xl ) ⊂ {x2 , . . ., xl }. /\/\ 

Remark. We didn’t use the full strength of δ(G) > n/2. We used only that x, y non-adjacent
⇒ d(x) + d(y) > n.

Similarly,

k
Proposition 3. Let G be a graph of order n (n > 3). Then G connected and δ(G) > ,
2
where k < n =⇒ G has a path of length k.

Note. We do need k < n: e.g., G = Kk . And we do need G to be connected: e.g., two

disjoint copies of Kk .

Proof of 3. Let x1 . . .xl be a longest path in G. (l > 3 since G is connected and n > 3.)
Suppose that l < k. Then, as in the proof of Theorem 2: wlog G has no l-cycle, and
thus Γ(x1 ) and Γ+ (xl ) are disjoint subsets of {x2 , . . ., xl }, each of size > k/2. /\/\ 

n(k − 1)
Theorem 4. Let G be a graph of order n with e(G) > . Then G contains a path
2
of length k.

n(k − 1)
Remarks. 1. Equivalently, G 6⊃ Pk ⇒ e(G) 6 .
2

2. This cannot be improved, if k divides n: e.g., n/k disjoint Kk .

16
Proof of 4. Use induction on n. Done if n 6 2.

Given G, with |G| = n > 3 and G 6⊃ Pk , we want e(G) 6 12 n(k − 1).

Wlog, G is connected – if not, then components P G1 , . . ., Gr of orders n1 , . . ., nr have

e(Gi ) 6 12 ni (k − 1) by induction, whence e(G) 6 1 1
2 ni (k − 1) = 2 n(k − 1).

Thus G has a vertex x of degree 6 21 (k − 1), by Proposition 3.

(Note: we may assume that k < n, as if n 6 k then e(G) 6 21 n(k − 1) trivially).

But then G − x is on n − 1 vertices and has no Pk , so e(G − x) 6 12 (n − 1)(k − 1) by

induction, whence e(G) 6 21 (n − 1)(k − 1) + d(x) 6 12 n(k − 1). 

Remark. Both Theorem 2 and Theorem 4 are extremal results. They answer “how large
can a graph be with a certain property?” Often this property is non-containment of a given
graph. E.g., how big can e(G) be if G is triangle-free?

Turán’s Theorem

How many edges guarantee a Kr ?

Equivalently, how many edges can a graph on n vertices have if it does not have Kr as a
subgraph?

For r = 3, we would try G = Ka,b where a + b = n. Take a = b = n/2 if n is even, and

a = (n + 1)/2, b = (n − 1)/2 if n is odd.

Say G is k-partite on classes V1 , . . ., Vk if V1 , . . ., Vk partition V and G[Vi ] = ∅ for all i.

So G (r − 1)-partite ⇒ G 6⊃ Kr (else > 2 vertices of Kr in some Vi ).

If in addition e(G) = {xy : x ∈ Vi , y ∈ Vj , some i 6= j}, then say G is complete k-partite.

The Turán graph Tk (n) is the complete k-partite graph on n vertices, with vertex classes
V1 , . . ., Vk , where |V1 |, . . ., |Vk | are as equal as posisble.

(Integers are “as equal as possible” if |ai − aj | 6 1 ∀ i, j.)

s`` s `s` s
s @ JJ ``` D s s @ JJ `` `D s
Q
DD Q
@  @ D
 Q Q @  #D 
E.g. T4 (8): @Q J@ D  and T4 (7):
C@
C @Q J@##D 
D @ Q
J@ Q
Q D C @ #J @
 Q D
s`
D ` @J @ Q Ds CsX
#X@J @
 Q Ds
Ds ``@Js
` XX
@Js

Certainly Tr−1 (n) 6⊃ Kr (as Tr−1 (n) is (r − 1)-partite), and Tr−1 (n) is maximal Kr -free: if
we add any edge then we make a Kr .

If k divides n then all classes have size n/k, so d(x) = n − n/k = n(1 − 1/k) for all x. In
general, the classes have size ⌈n/k⌉ or ⌊n/k⌋, so all degrees are n − ⌈n/k⌉ or n − ⌊n/k⌋.

17
Notes. 1. To obtain Tk (n − 1) from Tk (n), remove a point from a largest class, i.e. a point
of minimum degree.

2. And to obtain Tk (n) from Tk (n − 1), add a point to a smallest class.

Theorem 5 (Turán’s Theorem). Let G be a graph on n vertices. Then e(G) > e(Tr−1 (n))
=⇒ G ⊃ Kr .

Remarks. 1. This is best possible, as Tr−1 (n) 6⊃ Kr .

2. If we knew G was (r − 1)-partite, done by some kind of AM-GM inequality. But

there’s no reason why G should be (r − 1)-partite: e.g. C5 is K3 -free, but not bipartite.

3. It looks like the proof has to be fiddly, as e(Tr−1 (n)) is a complicated function of n
and r.

Proof. Key idea: we’ll strengthen the theorem to make it easier to prove. We’ll prove:
|G| = n, e(G) = e(Tr−1 (n)), G 6⊃ Kr =⇒ G ∼ = Tr−1 (n). (This certainly implies
Theorem 5, by maximality of Tr−1 (n).)

Use induction on n: n 6 r − 1 is trivial.

We have G with |G| = n, e(G) = e(Tr−1 (n)), G 6⊃ Kr .

Claim. δ(G) 6 δ(Tr−1 (n)).

X X
Proof of claim. We have e(G) = e(Tr−1 (n)) and so dG (y) = dTr−1 (n) (y).
y∈G y∈Tr−1 (n)

But the dTr−1 (n) (y) are as equal as possible, so δ(G) 6 δ(Tr−1 (n)), as claimed.

Choose x ∈ G with d(x) = δ(G) and let G′ = G − x. Then |G′ | = n − 1, G′ 6⊃ Kr , and

e(G′ ) = e(G) − δ(G) > e(Tr−1 (n)) − δ(Tr−1 (n)) = e(Tr−1 (n − 1)) (by note 1 above).
Thus δ(G) = δ(Tr−1 (n)) and G′ ∼= Tr−1 (n − 1) by the induction hypothesis.

Let G′ have vertex classes V1 , . . ., Vr−1 . We cannot have x joined to a point in every Vi ,
else G ⊃ KSr , so Γ(x) ∩ Vi = ∅ for some i. But d(x) = n − 1 − min |Vi | (by note 2 above).
So Γ(x) = j6=i Vj , for some i with |Vi | minimal. Thus G is complete (r − 1)-partite,
with vertex classes Vj (j 6= i) and Vi ∪ {x}. Thus G ∼ = Tr−1 (n). 

There are many other proofs of Turán.

The Problem of Zarankiewicz

“Bipartite version of Turán’s Theorem” : how many edges can a bipartite graph G (with n
vertices in each class) have if G 6⊃ Kt,t ?

Write Z(n, t) for this maximum. How large is Z(n, t)?

Theorem 6. Let t > 2. Then Z(n, t) 6 n2−1/t t1/t + nt. In particular, Z(n, t) 6 2n2−1/t for
n sufficiently large (i.e., for all n > n0 (t)).

18
Proof. Let G be bipartite, vertex classes X, Y , where |X| = |Y | = n, with G 6⊃ Kt,t . Let
degrees in X be d1 , . . ., dn . We’ll see that the average degree d is 6 n1−1/t t1/t + t.

Wlog each di > t − 1. (If di 6 t − 2, add an edge.)

For each t-set A in Y , how many x ∈ X have A ⊂ Γ(x)? At most t − 1, as G 6⊃ Kt,t .

 
n
Thus, the number of (x, A) with x ∈ X, A ⊂ Γ(x), |A| = t is 6 (t − 1) .
t

X di  X di   
n
But this number also equals , so 6 (t − 1) .
t t t
 
x x(x − 1). . .(x − t + 1)
Now, the function = is a convex function of x for x > t − 1.
t   t!
x (y + t − 1). . .y
(E.g., put y = x − t + 1, then = , which is a non-negative linear
t t!
combination of powers of y.)

X di   
d
 
d
 
n n(d − t + 1)t (t − 1)nt
Thus >n , so n 6 (t − 1) , and so 6 .
t t t t t! t!

Thus d − t + 1 6 n1−1/t (t − 1)1/t , whence d 6 n1−1/t t1/t + t. 

Is this the right value? Does Z(n, t) actually grow as n2−1/t (t fixed)?

t = 2. G bipartite, G 6⊃ K2,2 = C4 . Can e(G) be as large as cn3/2 ? Linear e(G) is easy,

e.g. a 2n-cycle. But n1.01 ? In fact, there are examples of G with e(G) = cn3/2 (coming from
algebra – projective planes).

t = 3. Here, cn5/3 is correct (but harder).

t = 4 is unknown!

** Non-examinable section **

The Erdős-Stone Theorem

For a fixed graph H, write Ex(n, H) = max{e(G) : |G| = n, G 6⊃ H}.
  
1 n
E.g., Turán’s Theorem says: Ex(n, Kr ) ∼ 1− .
r−1 2

Ex(n, Kr ) 1
Or, more precisely, Turán says: n

→1− as r → ∞.
2
r−1

e(G)
Note. n

is called the density of G.
2

19
 n(k − 1)
And Theorem 4 says: Ex(x, Pk ) ∼ .
2

How does Ex(n, H) behave for general H (as n → ∞)?

e(G) 1 e(G) 1
Turán says: n

>1− =⇒ G ⊃ Kr . But what if n

>1− + 0.001?
2
r−1 2
r−1

Write Kr (t) for Tr (rt). (“Kr blown up by t”.)

e(G) 1
Remarkably, n

>1− + 0.001 =⇒ G ⊃ Kr (t) for any t (for n large enough).
2
r−1

In general, we have the. . .

e(G) 1
Erdős-Stone Theorem. For all r, ǫ, t, n

> 1− +ǫ =⇒ G ⊃ Kr (t) for n > n0 (r, ǫ, t).
2
r − 1
 
1
Sketch proof. We have G, average degree > 1 − + ǫ n.
r−1
 
1
1. Get H ⊂ G (H large) with δ(H) > 1− + ǫ n′ (n′ = |H|).
r−1

(Similar to: average degree d ⇒ can get δ > d/2.)

2. By induction on r, H ⊃ Kr−1 (t′ ), some t′ large. Write K for the Kr−1 (t′ ).

3. Have lots of points in H − K, each joined to > t of each class of K (by δ(H)).

4. Get > t of these points joined to the same t-set in each class of K (by pigeonhole
principle). 

For given H, choose least r with H r-partite. (E.g., H = Petersen graph: not bipartite as it
has a 5-cycle, but it is 3-partite.) So then H ⊂ Kr (t), some t.

Ex(n, H) 1
Then Tr−1 (n) 6⊃ H (as Tr−1 (n) is (r − 1)-partite), so n

>1− .
2
r−1

Ex(n, H) 1
But Erdős-Stone says n

>1− + ǫ =⇒ G ⊃ any Kr (t), so G ⊃ H.
2
r−1

Ex(n, H) 1
Conclusion: n

→1− (where r is least such that H r-partite).
2
r−1

Ex(n, H)
Remark. If H bipartite, this says n

→ 0. But what is the growth speed of Ex(n, H)?
2

Unknown for most H. (E.g., H = C2n , n > 6).

** End of non-examinable section **

20
 Chapter 4 : Colourings

An r-colouring of a graph G is a function c : V (G) → [r] = {1, . . ., r}, such that c(x) 6= c(y)
whenever x, y are adjacent.

The chromatic number χ(G) of G is the smallest r such that G has an r-colouring.

Examples.
s2 s2 s2
s1 @@s 1 @@s 1
1. χ(Pn ) = 2 :

s1 s3

2 if n even 2s H Hs2 s
# #ccs1
1 sH s1
2. χ(Cn ) = : 2
B 
Hs 1 Bs s2
3 if n odd
2
3. χ(Kn ) = n (as all vertices need a different colour).

4. χ(En ) = 1.

5. T a tree ⇒ χ(T ) = 2. (E.g., induction: remove a leaf.)

6. χ(Km,n ) = 2 (one colour for each class).

Clearly any bipartite graph is 2-colourable, and conversely if c is a 2-colouring of G, then

X = {x : c(x) = 1}, Y = {y : c(y) = 2} show that G is bipartite. Similarly, G r-colourable
⇐⇒ G r-partite.
 
Ex(n, H) 1
So Erdős-Stone corollary says: n

→1− as r → ∞.
2
χ(H) − 1

If |G| = n then trivially χ(G) 6 n. We can improve this.

Proposition 1. Let G be a graph with maximum degree ∆. Then χ(G) 6 ∆ + 1.

Note. This is best possible: e.g., G a complete graph or odd cycle.

Proof. Order V (G) as x1 , . . ., xn and colour x1 , . . ., xn in turn. When we come to colour xi ,

we find 6 ∆ colours used already on neighbours of xi , so there is > 1 colour we can
use for xi . 

Remarks. 1. We can have χ(G) much less than ∆.

s
s s
H s
s Hs
s
E.g., K1,n is 2-colourable:  (also called a star)

2. Could view proof of Proposition 1 as an application of the greedy algorithm:

for a given ordering x1 , . . ., xn , colour each in turn, always using the smallest colour
available.

21
Warning. 1. Greedy algorithm may use more than χ(G) colours.

s x3 s x2
E.g., in sx @ @s x greedy gives 3 colours.
1 4

2. G ⊃ Kr ⇒ χ(G) > r, but G 6⊃ Kr 6⇒ χ(G) < r. E.g., C5 6⊃ C3 , but not 2-colourable.

In fact, there is no simple formula for χ(G).

Colouring Planar Graphs

How many colours do we need to colour a planar graph?

3 isn’t enough, e.g. K4 .

Proposition 2 (Six Colour Theorem). G planar =⇒ G is 6-colourable.

Proof. Use induction on |G|. Done if |G| 6 6.

We have planar G, |G| > 6.

Claim. δ(G) 6 5.
P
Proof of claim. We have e(G) 6 3n − 6 (as G is planar), so x∈G d(x) 6 6n − 12, so
some x has d(x) 6 5.

Choose x such that d(x) 6 5. Then G−x is planar, so by induction it has a 6-colouring.
Then Γ(x) receives 6 5 colours, so we can colour x with the remaining colour. 

Theorem 3 (Five Colour Theorem). G planar =⇒ G is 5-colourable.

Proof. Use induction on |G|. Done if |G| 6 5.

Given planar G, |G| > 5, we have δ(G) 6 5 (as before), so choose x ∈ G with d(x) 6 5.
By the induction hypothesis, we have a 5-colouring of G − x, so done unless d(x) = 5
and all 5 colours appear in Γ(x).

Say Γ(x) is x1 , . . ., x5 (clockwise) with xi having colour i for each i.

Question: is there a 1-3 path from x1 to x3 ? (A 1-3-path is a path on which colours

are alternately 1 and 3.)

If no, let H be the 1-3-component of x1 (i.e., all vertices we can reach from x1 by
a 1-3-path). So x3 ∈
/ H. Swap colours 1 and 3 on H. This is still a colouring of
G − x, but we now have x1 of colour 3, so we can use colour 1 for x.

If yes, we have a 1-3-path P from x1 to x3 , but then there is no 2-4-path from x2 to

x4 (as it would have to meet P , as G is planar). So finish as above: swap 2, 4 on
the 2-4-component of x2 and then use colour 2 for x. 

Remark. The i-j-paths in the proof are called Kempe chains.

22
Suppose we wanted to colour the faces of a plane graph such that distinct faces sharing an
edge get different colours – i.e., “colouring a plane map”.

Given a plane graph G, form the dual graph G′ by taking a point for each face, and joining
two points if they share an edge. So G′ planar and a colouring G′ corresponds exactly to
colouring the faces of G. Thus Theorem 3 tells us that every planar map is 5-colourable.

** Non-examinable section **

Theorem (Four Colour Theorem). G planar =⇒ G is 4-colourable.

“Proof ”. Use induction on |G|. Done if |G| 6 4.

Given planar G, |G| > 4, we have δ(G) 6 5 (as usual), so choose x ∈ G with d(x) 6 5.
We can 4-colour G−x by induction, so done unless either d(x) = 4 or 5, with all colours
appearing in Γ(x).

If d(x) = 4, proceed as in the Five Colour Theorem. If there is no 1-3-path from x1 to

x3 , swap the colours 1 and 3 on the 1-3-component of x1 and use colour 1 for x. And
if there is a 1-3-path from x1 to x3 , then there is no 2-4-path from x2 to x4 , so swap
colours 2 and 4 on the 2-4-component of x2 and use colour 2 for x.
s1′ s2
1Xs X s 2s 1Xs Xs s1′
If d(x) = 5, we could have or

s As

A s
As

A
4 3 4 3

First picture. If there is no 2-4-path from x2 to x4 then done (swap colours 2 and 4 on
the 2-4-component of x2 ).

If there is a 2-4-path from x2 to x4 then there is no 1-3-path from x3 to x1 or x1′ , so

done (swap colours 1 and 3 on the 1-3-component of x3 ).

Second picture. If there is no 2-4-path from x2 to x4 then done, as before.

So wlog there is a 2-4-path from x2 to x4 (so there is no 1-3-path from x1 to x3 ).

Similarly, wlog there is a 2-3-path from x2 to x3 (so there is no 1-4-path from x1′ to
x4 ). So swap colours 1 and 3 on the 1-3-component of x1 , and colours 1′ and 4 on the
1′ -4-component of x1′ . This leaves colour 1 for use at x. 

The above “proof” was given by Kempe in 1879.

In 1890, Heawood spotted a mistake – but where is the mistake?

The theorem was eventually proved in 1976 by Appel and Haken. In the proof of Theorem
3, we used the fact that “a vertex of degree 1, 2, 3, 4 or 5” forms an unavoidable (every
planar graph contains > one) set of reducible (cannot be present in a minimal counterex-
ample) configurations. Appel and Haken found an unavoidable set of about 1900 reducible
configurations for the Four Colour Theorem (using computers a lot).

** End of non-examinable section **

23
We have χ(G) 6 ∆ + 1 for all graphs G, and we can have equality – for example Kn and
Codd . We will show that in fact χ(G) 6 ∆ for all connected G except Kn and Codd .

Remark. If G is connected and not regular, we can certainly colour with ∆ colours.

Indeed, choose xn with d(xn ) < ∆. Then choose xn−1 adjacent to xn (as G connected).
Then choose xn−2 ∈ G − {xn , xn−1 } adjacent to {xn , xn−1 } (as G connected). And so
on. Then run the greedy algorithm on the order x1 , . . ., xn .

Each xi has a forward edge, i.e. xi xj ∈ E, some j > i (for all i 6= n), so uses 6 ∆
colours. (The presence of a forward edge means that when choosing the colour of a
vertex, we have already chosen colours for at most ∆ − 1 of its neighbours.) 

Proposition 4 (Brooks’ Theorem). Let G be connected, but not a complete graph or

odd cycle. Then χ(G) 6 ∆.

Proof. Wlog G is regular (by the remark above), and ∆ > 3 (as ∆ = 1 is trivial, and ∆ = 2
⇒ G is a cycle).

Let G be a minimal counterexample (i.e., |G| minimal). Wlog G is 2-connected: as if x

is a cutvertex, let G1 , . . ., Gk be the components of G − x together with x; then each Gi
is ∆-colourable (by remark above, as dGi (x) < ∆ for all i), so G itself is ∆-colourable.

Case 1. G is 3-connected.

We want an ordering such that every vertex has a forward edge, and xn has two
neighbours of the same colour.

Choose any xn . There must be some x1 , x2 ∈ Γ(xn ) with x1 x2 ∈ / E (or else

xn ∪ Γ(xn ) forms a K∆+1 , whence we must have a point of degree > ∆, as G is
connected but not itself a complete graph /\/\).

Now, G−{x1 , x2 } is connected (as G is 3-connected) so order its vertices x3 , . . ., xn

(as in the remark above) such that for all 3 6 i 6 n − 1, there is j > i for which
xi xj ∈ E(G). Then run the greedy algorithm on the ordering x1 , . . ., xn . This
uses 6 ∆ colours. 

Case 2. G is not 3-connected.

Choose a separator {x, y} (i.e., G − {x, y} is disconnected), and let G1 , . . ., Gk

be the components of G − {x, y} together with x and y. Then each Gi has a
∆-colouring (by the remark above, as dGi (x) 6 ∆ − 1 for all i).

If xy ∈ E, then x and y have different colours in the colouring of Gi for each i, so

we can recolour and combine to form a ∆-colouring of G.

So suppose xy ∈/ E. If each Gi has at least one of dGi (x), dGi (y) being 6 ∆ − 2,
we can recolour to ensure x and y have different colours in Gi .

So done, unless some Gi has dGi (x) = dGi (y) = ∆ − 1, say i = 1. Then we must
have k = 2, with dG2 (x) = dG2 (y) = 1 (as d(x), d(y) 6 ∆). Let ΓG2 (x) = {u},
ΓG2 (y) = {v}. Then {x, v} is a separator, not of this form. 

24
The Chromatic Polynomial
We are carrying more information on G than the number χ(G). For any graph G and
t = 1, 2, 3, . . ., let PG (t) be the number of t-colourings of G. (So χ(G) equals the least t such
that PG (t) > 0.)

Examples.

PKn (t) = t(t − 1)(t − 2). . .(t − n + 1).

PEn (t) = tn .
st−1 s . . . s
PPn (t) = t(t − 1)n : st @@st−1
@ @s . . .

And in general PTn (t) = t(t − 1)n−1 for any tree T on n vertices (by induction)
st − 1
PCn (t) = ..? t s
# #ccst − 1
B 
? Bs s. . .

Is PG (t) always a polynomial?

For a graph G and e = xy ∈ E(G), the contraction of G by e, written G/e, is formed from
G by replacing x and y by a new vertex e, joined to all points that were joined to x or y.
s s
E.g., s
x e y
@@ s - se
s s s @@s
G G/e
Lemma 5. Let G be a graph, and e ∈ E. Then PG = PG−e − PG/e .

Remark. Called the deletion-contraction relation or cut-fuse relation.

Proof. The t-colourings of G − e in which x and y get different colours correspond exactly
with the t-colourings of G.

And the t-colourings of G − e in which x and y get the same colour correspond exactly
with the t-colourings of G/e.

Thus PG−e (t) = PG (t) + PG/e (t). 

Note. We could not use Lemma 5 (and a base case of G = En ) to define PG , as it’s not
clear that it gives a unique function PG for all G.

Proposition 6. PG (t) is a polynomial in t.

Proof. Use induction on e(G). PEn (t) = tn , so e(G) = 0 is done.

Given G, e(G) > 0, choose e ∈ E. Then PG−e and PG/e are polynomials, by induction.
So PG = PG−e − PG/e is also a polynomial. 

25
For a tree T , we had PT (t) = tn − (n − 1)tn−1 + . . .

Proposition 7. Let G be a graph on n vertices with m edges. Then the leading terms of
PG (t) are tn − mtn−1 + . . .

Proof. Use induction on e(G). PEn (t) = tn , so e(G) = 0 is done.

Given G, e(G) > 0, choose e ∈ E. Then, by induction, PG−e = tn − (m − 1)tn−1 + . . .

and PG/e = tn−1 + . . .. So PG (t) = tn − mtn−1 + . . .. 

Remarks. 1. We can get other information about G from
its chromatic polynomial.

n−2 For
example, it turns out that PG (t) = tn − mtn−1 + m
2 − #triangles of G t + . . ..

2. Since PG is a polynomial, we can talk about PG (t) for any real or complex t.

3. Four Colour Theorem says: planar G has PG (4) > 0. I.e., PG has no root at 4. No
√ that PG (4) 6= 0 for all planar G. It is known
polynomial-style direct proof is known
that PG (ϕ + 2) 6= 0 where ϕ = (1 + 5)/2.

Edge Colouring
A k-edge-colouring of G is a map c : E(G) → {1, . . ., n} such that c(e) 6= c(f ) whenever
e, f share a vertex.

The smallest such k is called the edge-chromatic number or chromatic index of G,

written χ′ (G). (Thus χ′ (G) = χ(L(G)).)

2 if n even
E.g., χ′ (Cn ) =
3 if n odd

But also, χ′ (G) can be far from χ(G). E.g., K1,n (a star) has χ(G) = 2, χ′ (G) = n.

Clearly χ′ (G) > ∆(G) for all G. We can have χ′ (G) > ∆(G), for example Codd .

But very surprisingly:

Theorem 8 (Vizing’s Theorem). For any G, we have χ′ (G) = ∆(G) or ∆(G) + 1.

Proof. Let ∆ = ∆(G). We must show that χ′ (G) 6 ∆ + 1, i.e. that G can be (∆ + 1)-edge-
coloured. Use induction on e(G). Done if e(G) = 0.

Given G, e(G) > 0, choose e ∈ E(G). We have a (∆ + 1)-edge-colouring of G − e (by

induction). Let e be xy1 . At each vertex, some colour is not being used (as ∆+1 > d(y)
for all y ∈ G).

Choose a maximal sequence of distinct vertices s

x
y1 , . . ., yk as follows. Given edge xyi , choose a uncoloured HH
 c1 c2 HHck−1
colour ci missing at yi . If there is a new edge s
 s s Hs
... H
from x with colour ci , let xyi+1 be this edge. y1 y2 y3 yk

This must stop, as G is finite. And it must stop because either ck is not used at x, or
ck = cj for some j < k.

26
 If ck is not used at x, then recolour by giving xyi colour ci for all 1 6 i 6 k. This is a
(∆ + 1)-edge-colouring of G.

If ck = cj for some j < k, wlog j = 1, because we can recolour xyi with colour ci for
1 6 i 6 j − 1, leaving xyj as the uncoloured edge. Let c be a colour not used at x.

If there is no c − c1 path from y1 to x, then swap c and c1 on all edges of the c − c1

component of y1 . This leaves c missing at y1 (since c1 was missing at y1 previously)
and at x, so we can give xy1 colour c, so done.

Similarly, if there is no c − c1 path from yk to x (recall ck = c1 ), then swap c and c1

on the c − c1 component of yk . We can now use colour c for xyk and colour ci for xyi
(1 6 i 6 k − 1), so done.

Otherwise, the c − c1 component of x (call it H) contains y1 and yk , but H is connected

and has ∆(H) 6 2 (as H is 2-edge-coloured) and so is a path or cycle, but dH (x) =
dH (y1 ) = dH (yk ) = 1. /\/\ 

Graphs on Surfaces
We know that G drawn on the plane or a sphere has χ(G) 6 5 (well, actually 6 4). What
about G drawn on other surfaces?

1 sPP s1

s PJ 3


s
J J
P
PJP
s 2J 5P P

JJ
E.g., we can draw K7 on a torus: J P J
P s (with edges identified).
PJ s
4 J 7
JPP

6JPP s

1 s
J
J PJP JJs
P 1

In general?

The surface of genus g (or the compact orientable surface of genus g) consists of a
sphere with g handles attached.

For plane/sphere, n − m + f = 2 for G connected, and n − m + f > 2 for any planar G (add
edges to make G connected).

G on a torus?

n−m+f =2 n−m+f =1 n−m+f =0

Fact. For G on a surface of genus g, we have n − m + f > 2 − 2g. (This is the Euler
characteristic, E = 2 − 2g.)

So n − m + f > E, and 3f 6 2m (for m > 3) as usual.

2m
Thus n − m + > E, and so m 6 3(n − E).
3

27
Theorem 9 (Heawood’s Theorem). Let G be a graph drawn on a surface of Euler char-
acteristic E 6 0.
 √ 
7+ 49 − 24E
Then χ(G) 6 H(E) = .
2

Note. Tantalizingly, H(2) = 4, so “almost” have the Four Colour Theorem (but need E = 2).

Proof. Let G be a graph drawn on the surface, with χ(G) = k. We want k 6 H(E). Choose
minimal G with χ(G) = k. Then δ(G) > k − 1 (by minimality of G) and n > k.

We know m 6 3(n − E), so the sum of all degrees = 2m 6 6(n − E), and so δ(G) 6
6(n − E)/n = 6 − 6E/n.

Thus k − 1 6 δ(G) 6 6 − 6E/n 6 6 − 6E/k (as n > k and E 6 0, which is why this
fails for the Four Colour Theorem).

So k 2 − k 6 6k − 6E, i.e. k 2 − 7k + 6E 6 0. Whence result. 

Remark. Equality holds. We can draw KH(E) on a surface of characteristic E – this took
75 years to prove!

28
 Chapter 5 : Ramsey Theory

Can we find order in enough disorder?

Suppose we 2-edge-color K6 , e.g., c : E(K6 ) → {1, 2}. Can we always find a monochromatic
K3 (i.e., a K3 on which c is constant)?

Choose x ∈ V (K6 ). We have d(x) = 5, so there are at least 3 edges from x with the same
colour – say xy1 , xy2 , xy3 have colour 1. If any edge yi yj (i 6= j) also has colour 1, then xyi yj
is a colour-1 triangle. But if all edges yi yj are colour 2, then y1 y2 y3 is a colour-2 triangle.

What about K4 ? Is there an n such that Kn 2-edge-coloured ⇒ ∃ a monochromatic K4 ?

In general, write R(s) for the smallest n (if it exists) such that whenever Kn is 2-coloured
(we now mean “edge-coloured”), there exists a monochromatic Ks (i.e., a Ks on which
c : E(Kn ) → {1, 2} is constant).

Aim. To show that R(s) exists for all s (and find out roughly how fast R(s) grows).
s
#c
E.g., the above shows that R(3) 6 6. In fact, R(3) = 6, because of s# cs
B 
Bs s

Idea. “To go from monochromatic K3 to monochromatic K4 , first try to get a red K3 or

blue K4 .”

For s, t > 2, write R(s, t) for the smallest n (if it exists) such that whenever Kn is 2-coloured,
we have either a red Ks or a blue Kt .

So: R(s, s) = R(s),

R(s, t) = R(t, s),
R(s, 2) = s.

Equivalently, R(s, t) is the smallest n (if it exists) such that every graph G on n vertices has
either Ks ⊂ G or Kt ⊂ G.

Theorem 1 (Ramsey’s Theorem). R(s, t) exists for all s, t. And moreover, we have
R(s, t) 6 R(s − 1, t) + R(s, t − 1), for s, t > 3.

Proof. It is enough to show that if both R(s − 1, t) and R(s, t − 1) exist, then we have
R(s, t) 6 R(s − 1, t) + R(s, t − 1), as then R(s, t) exists for all s, t by induction on s + t.

So let a = R(s − 1, t), b = R(s, t − 1). Choose a 2-colouring of Ka+b , and choose
x ∈ Ka+b . Then d(x) = a + b − 1, so we have either > a red edges or > b blue edges
incident with x.

If > a are red, then consider the Ka spanned by the endpoints of the edges from x. By
the definition of a, this Ka contains either a red Ks−1 or a blue Kt .

If > b are blue, similarly. 

29
Remarks. 1. So any graph on n vertices has Ks ⊂ G or Ks ⊂ G for n large enough.

2. Very few of the Ramsey numbers R(s, t) are known exactly. (See later.)
 
s+t−2
Corollary 2. Let s, t > 2. Then R(s, t) 6 . In particular, R(s) 6 22s .
s−1

Proof. Induction on s + t. Done if s = 2 or t = 2. Given s, t > 3,

     
s+t−3 s+t−3 s+t−2
R(s, t) 6 R(s − 1, t) + R(s, t − 1) 6 + = . 
s−2 s−1 s−1

What about more colours?

Write Rk (s1 , . . ., sk ) for the smallest n (if it exists) such that whenever Kn is k-coloured,
there exists a Ksi of colour i, for some 1 6 i 6 k.

Corollary 3. Rk (s1 , . . ., sk ) exists for all k > 1 and s1 , . . ., sk > 2.

Proof (“Turquoise Spectacles”). Induction on k, the number of colours. Done if k = 1.

(Or if k = 2, by Ramsey’s Theorem.)

Given k > 2 and s1 , . . ., sk , let n = R(s1 , Rk−1 (s2 , . . ., sk )). Then any k-colouring of
Kn may be viewed as a 2-colouring, with colours 1 and “2 or 3 or . . . or k”.

So, by the choice of n, we have either a Ks1 coloured 1 (so done), or a KRk−1 (s2 ,...,sk )
coloured with colours 2, 3, . . . , k (i.e., (k − 1)-coloured), so done by the definition of
Rk−1 (s2 , . . ., sn ). 

What about r-sets?

E.g., r = 3. Colour each triangle red/blue – do we get a 4-set all of whose triangles are the
same colour? (We are asking for a much denser monochromatic structure.)

For X a set and r = 1, 2, . . ., write X (r) = {A ⊂ X : |A| = r}. Unless otherwise stated,
X = [n] = {1, . . ., n}.

Write R(r) (s, t) for the smallest n (if it exists) such that whenever X (r) is 2-coloured (i.e.,
we have c : X (r) → {1, 2}), there exists a red s-set (i.e., A ⊂ X, |A| = s, with c(B) = 1 for
all B ∈ A(r) ) or a blue t-set.

So: R(2) (s, t) = R(s, t),

R(1) (s, t) = s + t − 1 = (s − 1) + (t − 1) + 1 (pigeonhole principle),
R(r) (s, t) = R(r) (t, s),
R(r) (s, r) = R(r) (r, s) = s.

Theorem 4 (Ramsey for r-sets). Let r > 1, s, t > r. Then R(r) (s, t) exists.

Idea. In the proof of Theorem 1 (r = 2), we used the case r = 1.

30
Proof. Induction on r. Done if r = 1 (pigeonhole principle) or r = 2 (Theorem 1).

Given r > 1, use induction on s + t. Done if s = r or t = r. So, suppose s, t > r.

Claim. R(r) (s, t) 6 R(r−1) R(r) (s − 1, t), R(r) (s, t − 1) + 1.

Proof of claim. Let a = R(r) (s − 1, t), b = R(r) (s, t − 1), and n = R(r−1) (a, b) + 1.

Given a 2-colouring c of X (r) , choose x ∈ X and let Y = X \ {x}. Then c induces

a 2-colouring c′ of Y (r−1) by c′ (A) = c(A ∪ {x}), for each A ∈ Y (r−1) . So by the
definition of R(r−1) (a, b), we have either a red a-set for c′ or a blue b-set for c′ .

By symmetry, wlog we have a red a-set Z for c′ , i.e. A ∪ {x} is red for all
A ∈ Z (r−1) . But by the definition of a, Z contains either a red (s − 1)-set for c or
a blue t-set for c.

If a blue t-set, then we are done.

If a red (s − 1)-set, then add x and obtain a red s-set. 

Remarks. 1. Similarly for k colours – e.g., by “turquoise spectacles”.

2. The bounds we get on R(s,t) are quite large: “to get R(r) , iterate R(r−1) about s + t
times”.

Define f1 , f2 , f3 , . . . : N → N as follows:

f1 (x) = 2x, and for n > 2, fn (x) = fn−1 (fn−1 (fn−1 . . .(fn−1 (x)). . .)) .
| {z }
x times

} x times.
2
..
x 2.
So f2 (x) = 2 , f3 (x) = 2 And f4 (x)..?

Well, f4 (1) = 2, f4 (2) = 22 = 4, f3 (2) = 22

22 ..
= 65536, f4 (4) = 22
.2
} 65536 times
Then our bound on R(r) (s, t) is of the form fr (s + t).

(These sorts of bounds are often a feature of such double induction proofs.)

Infinite Ramsey Theory

Given a 2-colouring c of N(2) , can we always find an infinite monochromatic subset? (I.e.,
M ⊂ N, M infinite, with c constant on M (2) ?)

Examples. 1. Colour ij red if i + j is even, and blue if odd. Take, e.g., M = {n : n even}.

2. Colour ij red if max{n : 2n divides i + j} is even, and blue otherwise. Take, e.g.,
M = {22 , 24 , 26 , 28 , . . .}.

3. Colour ij red if i + j has an even number of (distinct) prime factors, and blue if odd.

31
Note. Asking for an infinite red set is more than asking for arbitrarily large finite red
sets. For example, consider the colouring for which all edges within the sets {1, 2}, {3, 4, 5},
{6, 7, 8, 9}, {10, 11, 12, 13, 14}, {15, 16, 17, 18, 19, 20},.. . are coloured red, and all other edges
are coloured blue. Then there are arbitrarily large finite red sets, but no infinite red set.

Theorem 5 (Infinite Ramsey). Let N(2) be 2-coloured. Then there exists an infinite
monochromatic M ⊂ N.

 
Proof. Choose a1 ∈ N. Then there exists infinite B1 ⊂ N \ {a1 }
! 

s  s !
!
((  
(!

such that all edges from a1 to B1 have the same colour, 
(
!
h
P hh a!
(
h
P h
aa a B2

P
1 aP P
aP
a
a
 
say c1 . Choose a2 ∈ B1 . Then there exists infinite 2
B2 ∈ B1 \ {a2 } such that all edges from a2 to B2 are the B1
same colour, say c2 . Continue inductively.

We obtain points a1 , a2 , . . . in N and colours c1 , c2 , . . . such that ai aj has colour ci (for

i < j). We must have infinitely many of the ci the same colour, say ci1 = ci2 = . . ..
Then M = {ai1 , ai2 , . . .} is infinite monochromatic. 

Remarks. 1. Similarly for k colours. (E.g., by “turquoise spectacles”.)

2. This is called a 2-pass proof.

3. In the third example above (prime factors), no explicit example is known.

Example. Any sequence in R (or any totally ordered set) has a monotonic subsequence.
Indeed, given sequence x1 , x2 , . . ., 2-colour N(2) by giving ij the colour “up” if xi 6 xj and
“down” if xi > xj . Now apply infinite Ramsey.

What about r-sets?

E.g., r = 3. 2-colour N(3) by giving ijk (i < j < k) the colour red if i divides j + k, and blue
if not. We can take M = {1, 2, 4, 8, 16, . . .}.

Theorem 6 (Infinite Ramsey for r-sets). Let r > 1, and let N(r) be 2-coloured. Then
there exists an infinite monochromatic subset M ⊂ N.

Proof. Induction on r. Done if r = 1 (or if r = 2, by Theorem 5).

Given r > 1, and given a 2-colouring c of N(r) , choose a1 ∈ N. This induces a 2-

colouring c′ of (N \ {a1 })(r−1) by c′ (F ) = c(F ∪ {a1 }). So there exists an infinite
monochromatic B1 ⊂ N \ {a1 } for c′ , by the induction hypothesis. I.e., we have some
colour c1 such that every r-set of the form {a1 } ∪ F , for F ⊂ B1 , has colour c1 .

Choose a2 ∈ B1 . This induces a 2-colouring of (B1 \ {a2 })(r−1) , so we get an infinite

B2 ⊂ B1 \ {a2 } and a colour c2 such that every r-set of the form {a2 } ∪ F , for F ⊂ B2 ,
has colour c2 . Continue inductively.

We obtain points a1 , a2 , . . . in N and colours c1 , c2 , . . . such that any r-set ai1 . . .air (for
i1 < . . . < ir ) has colour ci . But infinitely many of the ci agree, say ci1 = ci2 = . . ..
Then M = {ai1 , ai2 , . . .} is monochromatic. 

32
Example. We saw that given points (1, x1 ), (2, x2 ), . . . in R2 , we can find a subsequence
such that the induced function (i.e., piecewise linear, through these dots) is monotonic. In
fact, we can guarantee that the induced function is convex or concave. Colour N(3) by giving
ijk the colour “convex” if `r ` r r and “concave” if  r `r and apply Theorem 6.
r `
xi x xk xi xj xk
j

Exact Ramsey Numbers

Very few non-trivial (s, t, > 3) of the R(s, t) are known exactly:

R(3, 3) = 6, R(3, 4) = 9, R(3, 5) = 14, R(3, 6) = 18, R(3, 7) = 23, R(3, 8) = 28, R(3, 9) = 36.

R(4, 4) = 18, R(4, 5) = 25. Known that 43 6 R(5, 5) 6 49

(2)
For example, to show R(4, 4) > 17, 2-colour Z17 by giving ij colour red if i − j is a square
mod 17, and blue if not. (Have to check that there is no monochromatic K4 .)

For more than two colours, the only number is R3 (3, 3, 3) = 17.

For r-sets, the only known number is R(3) (4, 4) = 13.

This is hard because was are asking “exactly how much disorder” guarantees a given amount
of order – hard to analyse.

“Put it on a computer?”


5-sets, in each of 2( 2 ) colourings.
43
43
To show, for example, R(5, 5) > 43, need to examine 5
But 2( 2 ) > 2800 > 10250 , so no chance.
43

33
 Chapter 6 : Random Graphs

How fast does R(s) grow?

We know R(s) 6 4s . What about a lower bound?

It’s easy to see that R(s) > (s − 1)2 : take s − 1 copies of Ks−1 , then colour the edges within
each Ks−1 red, and those between different copies blue.

It was believed (in the 1940s) that perhaps R(s) ∼ cs2 . However. . .

Theorem 1 (Erdős, 1947). Let s > 3. Then R(s) > 2s/2 .

Proof. Choose a colouring of Kn at random, by taking each edge to be red or blue with prob-
s
ability 1/2 each (independently). Then P(a fixed s-set is monochromatic) = 2( 21 )(2) .
   
n n 1−(2s)
The number of s-sets is , so P(∃ a monochromatic s-set) 6 2 .
s s
   
n 1−(2s) n s
Thus we must have R(s) > n if 2 < 1, i.e. if < 2(2)−1 .
s s
 
n ns s
But 6 and s! > 2 2 +1 (for s > 3, by induction on s).
s s!
2
So done if ns 6 2s /2
, i.e. if n 6 2s/2 

Remarks. 1. The above is a random graphs argument.

n
2. We could rewrite it as: there are 2( 2 ) colourings, and a given s-set is monochromatic
n s
n s n
in 2 × 2( s )−(2) of them, so done if ns × 21+( 2 )−(2) < 2( 2 ) .

(But this is a bad viewpoint for later, when we won’t have all graphs equally likely.)

3. The proof gives no hint as to how to construct such a colouring.

4. No construction giving an exponential lower bound on R(s) is known!

So, to find an actual graph G on 106 points with no K40 in G or G, the best thing to do is
just toss a coin for each edge.
√ s √ √
We have now 2 6 R(s) 6 4s . No better bounds √ s ( s2 → 2 + ǫ or 4 → 4 − ǫ) are known.
(There exists a heuristic argument for each of 2 , 2 , 4s .)

The probability space G(n, p) is defined on the set of all graphs on {1, . . ., n} as follows.
We choose each edge to be present with probability p, and absent with probability 1 − p.

So, e.g., in the proof of Theorem 1, we worked inside G(n, 21 ).

34
  s2 
Example. In G(5, p), P 1 s
# #ccs3 = p6 (1 − p)4 .
B 
5Bs s4

It can be useful to look at p other than 1/2.

Recall Zarankiewicz. We had Z(n, t) 6 2n2−1/t (for t fixed, n large).

What about a lower bound? (Hopefully better than the trivial lower bound.)

Could: choose a random bipartite graph G, vertex classes X, Y (with |X| = |Y | = n), choos-
ing each edge independently with probability p, and choosing p to make the expected
number of Kt,t less than 1.
 2
n 2
Now, the number of Kt,t is , and P(∃ fixed Kt,t ⊂ G) = pt .
t
 2
n 2 1 2
So the expected number of Kt,t in G is pt < n2t pt .
t 4

Take p = n−2/t to get E(#Kt,t in G) < 1/4. Whence P(G has no Kt,t ) > 3/4.

Also, E(#edges) = pn2 , so we would get P(e(G) > 12 pn2 ) > 1/2, so there exists G, with
no Kt,t , with e(G) > 12 pn2 = 21 n2−2/t . 

We can do better, however.

1 2−2/(t+1)
Theorem 2. Z(n, t) > n .
2
Idea. If a graph G has m edges and r copies of Kt,t , remove an edge from each copy of Kt,t
to obtain a graph with > m − r edges and no Kt,t , showing Z(n, t) > m − r.

Proof. Choose a random bipartite graph G, vertex classes X, Y (|X| = |Y | = n), by taking
each edge independently with probability p. Let M = e(G), and R = the number of
Kt,t in G.
 2
n 2
2
As before, E(M ) = pn and E(R) = pt .
t
 2
n 2 1 2
So, using linearity of expectation, E(M − R) = pn2 − pt > pn2 − n2t pt .
t 2

1 2 1
Taking p = n−2/(t+1) , we have E(M − R) > n2−2/(t+1) − n2t−2t /(t+1) = n2−2/(t+1) .
2 2
1 2−2/(t+1) 1
Thus there exists a graph with m − r > n , so Z(n, t) > n2−2/(t+1) . 
2 2

Remark. The above proof is called modifying a random graph.

35
Graphs with Large χ(G)
To make χ(G) > k – can ensure this by G ⊃ Kk .

To have χ(G) > k – need not have G ⊃ Kk . E.g., G = C5 .

In fact, we can have χ(G) much larger than the clique number = max{k : G ⊃ Kk },
because:

(a) the graph G in Theorem 1 (e.g., the red edges) is on n = 2s/2 vertices, with no K ⊂ G
and no independent set (i.e., a set with no edges) of size s. But in any colouring, each
s/2
colour class is an independent set, so χ(G) > 2s−1 – much more than the clique number,
which is 6 s − 1.

(b) We can construct a graph G to be triangle-free (i.e., K3 6⊂ G) but with χ(G) large
(although this is not easy).

Could we have large girth (length of shortest cycle) and still have χ(G) large, for example,
with girth > 10, χ(G) > 100? This seems unlikely, however. . .

Theorem 3. For all k, g, there exists a graph G with χ(G) > k and girth(G) > g.

Idea. Try to find G on n vertices such that the number of short cycles (length 6 g) is 6 n/2,
and every independent set has size 6 n/2k. Then done by removing a vertex in each short
cycle to obtain a graph H with girth(H) > g and χ(H) > (n/2)/(n/2k) = k.

Proof. Choose a random G ∈ G(n, p), where p = n−1+1/g .

Let xi be the number of i-cycles in G, 3 6 i 6 g − 1, and let x be the number of cycles

of length < g, so x = x3 + . . . + xg−1 .

Then E(xi ) = (# possible i-cycles) × P(given i-cycle is present) 6 ni pi .

g−1 g−1
X X g n
So E(x) 6 (np)i = ni/g 6 gn(g−1)/g = n < for n large (as g/n1/g → 0).
i=3 i=3
n1/g 4

Thus P(x 6 n/2) > 1/2 (else P(x > n/2) > 1/2, whence E(x) > n/4 /\/\).

Let t = n/2k (for n a multiple of 2k) and let y be the number of t-sets that are
independent. Then
 
n t
E(y) = (#t-sets) × P(given t-set independent) = (1 − p)(2)
t
t
6 nt e−p(2) (because 1 − x 6 e−x ∀ x)
 
n n2
6 exp log n − 2 n−1+1/g
2k 8k

→ 0 as n → ∞ (because n log n − nn1/g → −∞).

So E(y) < 1/2 (for n large). So P(y = 0) > 1/2.

Thus there exists G ∈ G(n, p) with x 6 n/2, y = 0, so done. 

36
The Structure of a Random Graph
What does a “typical” random graph G ∈ G(n, p) look like?

How do the properties vary as p varies?

For example, does G have no isolated vertices?

We would expect a gradual increase of P(no isolated vertices) as we increase p. But in fact
we get a threshold effect:

16
P(no isolated
vertices)
- p
1

This jump or threshold happens below p = constant, since if p = constant then P(a given
vertex is isolated) is exponentially small, whence P(∃ isolated vertex) is also small.

So, where does that jump happen?

Probability digression/reminder

Let X be a random variable, taking values in {0, 1, 2, . . .}.

To show P(X = 0) is large, it is enough to show that the mean µ = E(X) is small. Indeed,
for any t, we have µ > P(X > t)t whence P(X > t) 6 µ/t (known as Markov’s inequality),
so P(X > 1) 6 µ, so P(X = 0) > 1 − µ.

To show P(X = 0) is small, it is not enough to show that µ is large. E.g., take X = 0 with
probability 999/1000, and X = 1010 with probability 1/1000.

So instead look at the variance V = Var(X) = E((X − µ)2 ) = E(X 2 ) − E2 (X). For any t,
P(|X − µ| > t) = P(|X − µ|2 > t2 ) 6 V /t2 , by Markov (known as Chebyshev’s inequality).

Thus P(|X − µ| > µ) 6 V /µ2 . So in particular P(X = 0) 6 V /µ2 .

Conclusion: to show P(X = 0) small, show V /µ2 small.

P
Suppose X is the number of some events A that occur. Then µ = E(X) = n P(A).
P P P
Variance? E2 (X) = A,B P(A)P(B), and E(X 2 ) = A,B P(A ∩ B) = A,B P(A)P(B|A).
P P P
(Using, e.g., X = A IA , so X 2 = A,B IA IB = A,B IA∩B .)
P
So variance V = Var(X) = A,B P(A)(P(B|A) − P(B)).

(Note: no contribution to this from independent A, B.)

The phrase almost surely means “with probability → 1 as n → ∞”.

37
 log n
Theorem 4. Let λ be constant, and let G ∈ G(n, p) where p = λ .
n
If λ < 1 then almost surely G has an isolated vertex. While if λ > 1, then almost surely
G has no isolated vertex.

Remark. Theorem 4 tells us that p = log n/n is a threshold for the property of having an
isolated vertex.

Proof of 4. Let X be the number of isolated vertices of G.

n
Then µ = E(X) = n(1 − p)n−1 = (1 − p)n .
1−p

n −pn n −λ log n n1−λ

For λ > 1, we have µ 6 e = e = → 0 as n → ∞,
1−p 1−p 1−p
so certainly P(X = 0) → 1.

For λ < 1, we have 1 − p > e−(1+δ)p , for any δ (p small enough).

n −(1+δ)pn n −(1+δ)λ log n n1−(1+δ)λ

So µ > e = e = .
1−p 1−p 1−p

Choosing δ such that (1 + δ)λ < 1, we have µ → ∞ as n → ∞.

Var(X) = n(1 − p)n−1 (1 − (1 − p)n−1 ) + n(n − 1)(1 − p)n−1 ((1 − p)n−2 − (1 − p)n−1 )
| {z } | {z }
n terms in which A = B n(n − 1) terms in which A 6= B

6 µ + n(n − 1)(1 − p)n−1 p(1 − p)n−2

p p
6µ+ n2 (1 − p)2n−2 = µ + µ2 .
1−p 1−p

V 1 p
Thus 2
6 + → 0 as n → ∞. So P(X = 0) → 0 as n → ∞. 
µ µ 1−p

A different kind of threshold effect is the clique number of a random graph. Let 0 < p < 1
be fixed. What is the distribution of the clique number of G ∈ G(n, p)?

We’d guess: 6 but in fact: 6

1 1
P(clique P(clique
number) number)
- -
n a a+1 n
clique number clique number

It turns out that, almost surely, the clique number is a or a + 1 (some a).

38
  
n (d2)
Theorem 5. Let 0 < p < 1 be fixed, and let d be a real number with p = 1.
d

Then, almost surely, G ∈ G(n, p) has clique number equal to ⌈d⌉, ⌊d⌋, or ⌊d⌋ − 1.

Remark. With more work, we could check that only two values occur.

Sketch proof of 5. Fix an integer k, and let X be the number of Kk in G.

 
n (k2 )
So µ = E(X) = p .
k

Task: if k 6 d − 1, then almost surely X 6= 0, while if k > d + 1 then almost surely

X = 0.

If k > d + 1, then µ → 0 (check), so P(X = 0) → 1.

If k 6 d − 1, then µ → ∞ (check).
  k   
n (k
)
X k n − k  (k2)−(s2) k

Now, V = p 2 p − p( 2 ) .
k s=2 |
s k−s
{z }
#B meeting A in s points

k   
V 1 X k n−k k s
So, 2 6 p(2)−(2) .
µ µ s=2 s k−s

Check that the first and last terms dominate – i.e., sum 6 (first + last)× constant.
   
V k n − k (k2 )−1
So, 2 6 constant× p + 1 , whence V /µ2 → 0.
µ 2 k−2

Thus P(X = 0) → 0. 

39
 Chapter 7 : Algebraic Methods

The diameter of a graph G is max{d(x, y) : x, y, ∈ G}.

So, e.g., G has diameter 1 ⇐⇒ G is complete.

What about diameter 2? How many vertices can G have if G has diameter 2 and maximum
degree ∆?

Expanding from a point x, we see that V (G) = {x} ∪ Γ(x) ∪ Γ(Γ(x)). And |Γ(x)| 6 ∆, so
|G| 6 1 + ∆ + ∆(∆ − 1) = 1 + ∆2 .

If |G| = 1 + ∆2 , then G must be regular.

A k-regular graph of diameter 2 on n = 1 + k 2 vertices is called a Moore graph, or Moore

graph of diameter 2.

Equivalently, a k-regular graph is a Moore graph ⇐⇒ for all x 6= y there exists a unique path
of length 6 2 from x to y (for k 6= 1).

Examples.
s
#c
s# cs
1. k = 2. C5 : B 
Bs s
s
#c
s cs
# c
s# s
PP
Petersen Graph : B Z B s

B s ZBs 
2. k = 3.

Bs
 S Ss
s
s s s
s s
Try s s s s s and add further required edges. But this fails!
s s
3. k = 4.
s s s
s
4. k > 4. Ought to keep failing. Why?

Let G be a graph on vertex set [n] = {1, . . ., n}. 

1 if ij ∈ E(G)
The adjacency matrix of G is the n × n matrix A with Aij = .
0 otherwise

s5
 
0 1 0 1 0
#c
s cs3
1 0 1 0 1
#
2 B
 
E.g., for  0
we find A =  1 0 1 1
s s

1 B  4
1 0 1 0 0
0 1 1 0 0

Note that A is always real and symmetric.

The matrix A contains all of the information of G.

40
 P
E.g., look at A2 . (A2 )ij = k Aik Akj = the number of walks of length 2 from i to j.

Similarly, (A3 )ij = the number of walks of length 3 from i to j.

P
We have a linear map x 7→ Ax from Rn to Rn . Thus (Ax)i = j Aij xj .

s0 s3 + 7
#c #c
E.g., so 3 s#
B
cs 7

- 6+7+0 # s
B
cs 0 + 3 + (−5)

6 B s s

−5 3 + (−5) Bs s

6+7

I.e., add the values at neighbours. So if x = (6, 3, 7, −5, 0) then Ax = (−2, 13, −2, 13, 10).

E.g., if A is k-regular, then (1, 1, . . ., 1) 7→ (k, k, . . ., k).

Since A is real symmetric, it is diagonalisable: it has a basis of eigenvectors, say e1 , e2 , . . ., en

with corresponding eigenvalues λ1 > λ2 > . . . > λn . (We can take the e1 , . . ., en to be an
orthonormal basis.)
P
Often write λmax = λ1 and λmin = λn . Note that λi = 0, as trace(A) = 0. So λmax > 0
and λmin < 0 (unless G = En ).

To find eigenvalues, we often don’t need lots of calculation.

 
0 1 0 1
1 0 1 0
Example. C4 has A = 
0
.
1 0 1
1 0 1 0

So rank(A) = 2, so we have 0 as an eigenvalue twice. The other two?

1s s1 2s s2
A- so 2 is an eigenvalue.
1s s1 2s s2

1s s−1 −2 s s2
A- so −2 is an eigenvalue (or use
P
= 0)
−1 s s1 2s s−2

Thus the eigenvalues are 2, 0, 0, −2.

n
P
Let e1 , . . ., en P
be an orthonormal basis
2
P of eigenvectors. TakeP x ∈2 R , say x = ci ei , with
||x|| = 1, i.e., ci = 1. Then Ax = ci λi ei , so hx, Axi = λi ci .

So, 
min hx, Axi = λmin (take cn = 1, all other ci = 0) 

||x||=1 
(∗)

max hx, Axi = λmax (take c1 = 1, all other ci = 0) 

||x||=1

41
Proposition 1. For any graph G,

(i) λ an eigenvalue =⇒ |λ| 6 ∆

(ii) For G connected: ∆ is an eigenvalue ⇐⇒ G is regular

(iii) For G connected: −∆ is an eigenvalue ⇐⇒ G is regular and bipartite

(iv) λmax > δ.

P x for λ, and choose i with |xi | maximal. Wlog, xi = 1.

Proof. (i) Choose an eigenvector
But then |(Ax)i | = | j∈Γ(i) xj | 6 ∆ × 1. So |λ| 6 ∆.

(ii) (⇐) Let x = (1, . . ., 1), then Ax = (∆, . . ., ∆).

(⇒) From (i) we must have d(i) = ∆ and that all j ∈ Γ(i) also have xj = 1. So
we can repeat for each j ∈ Γ(i), then for each k ∈ Γ(j), etc. We obtain d(k) = ∆
for all k (as G is connected).

(iii) (⇐) Choose x = (1, 1, . . ., −1, −1, . . .).

| {z } | {z }
on X on Y

(⇒) From (i) we must have d(j) = ∆ and xj = −1 for all j ∈ Γ(i). Repeat for
each k ∈ Γ(j), etc. We obtain d(j) = ∆ for all j ∈ G. and for all jk ∈ E we have
xj = 1, xk = −1 or xj = −1, xk = 1. Thus G is regular and has no odd cycle.

(iv) Let x = (1, . . ., 1). Then (Ax)i > δ for all i, so hAx, xi > δhx, xi = δn. Hence
λmax > δ, by (∗) (immediately before Proposition). 

Remark. In (ii), if ∆ is an eigenvalue, it has multiplicity 1 (as it has the unique eigenvector
(1, . . ., 1)).

Eigenvalues can link in to other graph parameters. We know that χ(G) 6 ∆ + 1. We can
strengthen this to

Proposition 2. For any graph G, χ(G) 6 λmax + 1.

Proof. Induction on |G|. Done if |G| = 1.

Given G with |G| > 1, choose v ∈ G with d(v) = δ(G).

Claim. λmax (G − v) 6 λmax .

Then done: colour G − v by induction, and we can colour v as d(v) 6 λmax .

Proof of claim. Let B be A with row and column v removed – wlog, the last row and
column.

For any x = (x1 , . . ., xn−1 ), let y = (x1 , . . ., xn−1 , 0). Then (Ay).y = (Bx).x, and
so λmax (G − v) 6 λmax , by (∗). 

42
Towards Moore Graphs
A graph G is strongly regular with parameters (k, a, b) if G is k-regular, with any two
adjacent points having exactly a common neighbours, and any two non-adjacent points having
exactly b common neighbours.

Examples.
s s
s s
C4 : has (2, 0, 2).

s
#c
s
# cs
C5 : B  has (2, 0, 1)
Bs s

– in general a Moore graph of degree k is strongly regular with (k, 0, 1).

s

A
s
As
s A

s
A has (4, 1, 2).

A A

A
s
A
s A
s As

A huge constraint on such G is given by. . .

Theorem 3 (Rationality condition for strongly regular graph). Let G be a graph on

n vertices, strongly regular with parameters (k, a, b), with b > 1.
!
1 (n − 1)(b − a) − 2k
Then the the numbers n−1± p are integers.
2 (a − b)2 + 4(k − b)

Proof. G is connected, as b > 1. So k is an eigenvalue with multiplicity 1. What are the

other eigenvalues?

 k if i = j
We have (A2 )ij = a 6 j, ij ∈ E .
if i =

b if i =6 j, ij ∈
/E

Thus A2 = kI + aA + b(J − I − A), where J has all entries equal to 1.

So A2 + (b − a)A + (b − k)I = bJ.

For λ an eigenvalue (λ 6= k) with eigenvector x, we have (1, . . ., 1).x = 0, since the

eigenvectors are orthogonal. I.e., Jx = 0.

Thus (λ2 + (b − a)λ + (b − k))x = 0, so λ2 + (b − a)λ + (b − k) = 0.

 p 
So eigenvalues not equal to k are λ = 21 a − b ± (b − a)2 + 4(k − b) – say λ, µ, with
multiplicities r, s. So r + s = n − 1, and λr + µs = −k, as the eigenvalues sum to 0
(e.g., consider the trace). Solving for r, s, we get the numbers in the theorem. 

43
Finally,

Corollary 4. If there exists a Moore graph of degree k, then k ∈ {2, 3, 7, 57}.

Remark. k = 2 : C5

k = 3 : Petersen graph

k = 7 : exists

k = 57 : unknown

Proof of 4. By Theorem 3, with n = k 2 + 1, a = 0, b = 1, we must have

!
1 2 k 2 − 2k
k ±p ∈ Z.
2 1 + 4(k − 1)

So either k 2 − 2k = 0 or 1 + 4(k − 1) = 4k − 3 is a perfect square.

If k 2 − 2k = 0, then k = 2. Done.
 2  
t2 + 3 t2 + 3
If 4k − 3 = t2 , then t divides k 2 − 2k = −2 .
4 4

So t divides (t2 + 3)2 − 8(t2 + 3) = t4 − 2t2 + 15.

Thus t divides 15, so t = 1, 3, 5, 15, giving k = 1 (not possible), 3, 7, 57. 

44
 Mich. 2007 GRAPH THEORY – EXAMPLES 1 IBL

1. Construct a 3-regular graph on 8 vertices. Is there a 3-regular graph on 9 vertices?

2. How many spanning trees does K4 have?

3. Prove that every connected graph has a vertex that is not a cutvertex.

4. Let G be a graph on n vertices, G 6= Kn . Show that G is a tree if and only if the

addition of any edge to G produces exactly 1 new cycle.

5. Let n ≥ 2, and let d1 ≤ d2 . . . ≤ dn be a sequence of integers.

P Show that there is a tree
with degree sequence d1 , . . . , dn if and only if d1 ≥ 1 and di = 2n − 2.

6. Let T1 , . . . , Tk be subtrees of a tree T , any two of which have at least one vertex in
common. Prove that there is a vertex in all the Ti .

7. Show that every graph of average degree d contains a subgraph of minimum degree at
least d/2.

8. The clique number of a graph G is the maximum order of a complete subgraph of G.

Show that the possible clique numbers for a regular graph on n vertices are 1, 2, . . . , ⌊n/2⌋
and n.

9. Let G be a graph on vertex set V . Show that there is a partition V1 ∪ V2 of V such that
in each of G[V1 ] and G[V2 ] all vertices have even degree.

10. For which n and m is the complete bipartite graph Kn,m planar?

11. Prove that the Petersen graph (shown) is not planar.

12. The square of a graph G has vertex set that of G and edge set {xy : dG (x, y) ≤ 2}.
For which n is the square of the n-cycle planar?

13. Prove that every planar graph has a drawing in the plane in which every edge is a
straight-line segment.
+
14. The group of all isomorphisms from a graph G to itself is called the automorphism
group of G. Show that every finite group is the automorphism group of some graph. Is
every group the automorphism group of some (possibly infinite) graph?
 Mich. 2007 GRAPH THEORY – EXAMPLES 2 IBL

1. For which n and m is the complete bipartite graph Kn,m Hamiltonian? Is the Petersen
graph Hamiltonian?
2. Let G be a graph of order n with e(G) > n2 − (n − 2). Prove that G is Hamiltonian.

3. Let G be a bipartite graph with vertex classes X, Y . Show that if G has a matching
from X to Y then there exists x ∈ X such that every edge incident with x extends to a
matching from X to Y .
4. Let G be a connected bipartite graph with vertex classes X, Y . Show that every edge
of G extends to a matching from X to Y if and only if |Γ(A)| > |A| for every A ⊂ X,
A 6= ∅, X.
5. Let A be a matrix with each entry 0 or 1. Prove that the minimum number of rows
and columns containing all the 1s of A equals the the maximum number of 1s that can be
found with no two in the same row or column.
6. An n × n Latin square (resp. r × n Latin rectangle) is an n × n (resp. r × n) matrix, with
each entry from {1, . . . , n}, such that no two entries in the same row or column are the
same. Prove that every r × n Latin rectangle may be extended to an n × n Latin square.
+
7. Let G be a (possibly infinite) bipartite graph, with vertex classes X, Y , such that
|Γ(A)| ≥ |A| for every A ⊂ X. Give an example to show that G need not contain a
matching from X to Y . Show however that if G is countable and d(x) < ∞ for every
x ∈ X then G does contain a matching from X to Y . Does this remain true if G is
uncountable?
8. Show that we always have κ(G) ≤ λ(G). For any positive integers k ≤ l, construct a
graph G with κ(G) = k and λ(G) = l.
9. For a set B ⊂ V (G) and a vertex a not in B, an a-B fan is a family of |B| paths from
a to B, any two meeting only at a. Show that a graph G (with |G| > k) is k-connected if
and only if there is an a-B fan for every B ⊂ V (G) with |B| = k and every vertex a not
in B.
10. Let G be a k-connected graph (k ≥ 2), and let x1 , . . . , xk be vertices of G. Show that
there is a cycle in G containing all the xi .
11. For each r ≥ 3, construct a graph G such that G does not contain Kr but G is not
(r − 1)-partite.
12. Let G be a graph of order n that does not contain an even cycle. Prove that each
vertex x of G with d(x) ≥ 3 is a cutvertex, and deduce that G has at most ⌊3(n − 1)/2⌋
edges. Give (for each n) a graph for which equality holds. How does this bound compare
with the maximum number of edges of a graph of order n containing no odd cycles?
13. A deleted Kr consists of a Kr from which an edge has been removed. Show that if G
is a graph of order n (n ≥ r + 1) with e(G) > e(Tr−1 (n)) then G contains a deleted Kr+1 .
14. A bowtie consists of two triangles meeting in one vertex. Show that if G is a graph of
order n (n ≥ 5) with e(G) > ⌊n2 /4⌋ + 1 then G contains a bowtie.
+
15. Let G be an r-regular graph on 2r + 1 vertices. Prove that G is Hamiltonian.
 Mich. 2007 GRAPH THEORY – EXAMPLES 3 IBL

1. What is the chromatic number of the Petersen graph? What is its edge-chromatic
number?

2. What is χ′ (Kn,n )? What is χ′ (Kn )?

k


3. Let G be a graph with chromatic number k. Show that e(G) ≥ 2
.

4. Show that, for any graph G, there is an ordering of the vertices of G for which the
greedy algorithm uses only χ(G) colours.

5. For each k ≥ 3, find a bipartite graph G, with an ordering v1 , v2 , . . . , vn of its vertices,

for which the greedy algorithm uses k colours. Give an example with n = 2k − 2. Is there
an example with n = 2k − 3?

6. Let G be a bipartite graph of maximum degree ∆. Must we have χ′ (G) = ∆?

7. Find the chromatic polynomial of the n-cycle.

8. Let G be a graph on n vertices, with pG (t) = tn + an−1 tn−1 + an−2 tn−2 + . . . + a1 t + a0 .

Show that the ai alternate in sign (in other words, ai ≤ 0 if n − i is odd and

ai ≥ 0 if n − i
m
is even). Show also that if G has m edges and c triangles then an−2 = 2 − c.

9. An acyclic orientation of a graph G is an assignment of a direction to each edge of G in

such a way that there is no directed cycle. Show that the number of acyclic orientations
of G is precisely |pG (−1)|.

10. Let G be a plane graph in which every face is a triangle. Show that the faces of G
may be 3-coloured, unless G = K4 .

11. Can K4,4 be drawn on the torus? What about K5,5 ?

12. A minor of a graph G is any graph that may be obtained from a subgraph of G by
successively contracting edges – equivalently, a graph H on vertex-set {v1 , . . . , vr } is a
minor of G if we can find disjoint connected subgraphs S1 , . . . , Sr of G such that whenever
vi vj ∈ E(H) there is an edge from Si to Sj . Show that for any k there is an n such that
every graph G with χ(G) ≥ n has a Kk minor. Writing c(k) for the least such n, show
that c(k + 1) ≤ 2c(k). [Hint: choose x ∈ G, and look at the sets {y ∈ G : d(x, y) = t}.]
Show that c(k) = k for 1 ≤ k ≤ 4, and explain why c(5) = 5 would imply the 4-Colour
Theorem.

13. Let G be a countable graph in which every finite subgraph can be k-coloured. Show
that G can be k-coloured.
+
14. Construct a triangle-free graph of chromatic number 1526.
 Mich. 2007 GRAPH THEORY – EXAMPLES 4 IBL

1. Show that R(3, 4) ≤ 9. By considering the graph on Z8 (the integers modulo 8) in

which x is joined to y if x − y = ±1 or ± 2, show that R(3, 4) = 9.

2. By considering the graph on Z17 in which x is joined to y if x − y is a square modulo

17, show that R(4, 4) = 18.

3. Show that R3 (3, 3, 3) ≤ 17.

4. Let A be a set of R(4) (n, 5) points in the plane, with no three points of A collinear.
Prove that A contains n points forming a convex n-gon.

5. Let A be an infinite set of points in the plane, with no three points of A collinear. Prove
that A contains an infinite set B such that no point of B is a convex combination of other
points of B.

6. Show that every graph G has a partition of its vertex-set as X ∪Y such that the number
of edges from X to Y is at least 21 e(G). Give three proofs: by induction, by choosing an
optimal partition, and by choosing a random partition.

7. In a tournament on n players, each pair play a game, with one or other player winning
(there are no draws). Prove that, for any k, there is a tournament in which, for any k
players, there is a player who beats all of them. [Hint: consider a random tournament.]
Exhibit such a tournament for k = 2.

8. Let X denote the number of copies of K4 in a random graph G chosen from G(n, p).
Find the mean and the variance of X. Deduce that p = n−2/3 is a threshold for the
existence of a K4 , in the sense that if pn2/3 → 0 then almost surely G does not contain a
K4 , while if pn2/3 → ∞ then almost surely G does contain a K4 .

9. Find the eigenvalues of Kn . Find the eigenvalues of Kn,m .

10. Prove that the matrix J (all of whose entries are 1) is a polynomial in the adjacency
matrix of a graph G if and only if G is regular and connected.

11. Let G be a graph in which every edge is in a unique triangle and every non-edge is a
diagonal of a unique 4-cycle. Show that G is k-regular, for some k, and that the number
of vertices of G is 1 + k 2 /2. Show also that k must belong to the set {2, 4, 14, 22, 112, 994}.

12. Let the infinite subsets of N be 2-coloured. Must there exist an infinite set M ⊂ N all
of whose infinite subsets have the same colour?
+
13. Let A be an uncountable set, and let A(2) be 2-coloured. Must there exist an
uncountable monochromatic set in A?