Introduction to Graph Theory

Notes for MATH 4171

R.A. Litherland

January 12, 2009

Contents
0 Preliminaries 2

1 Graphs 5

2 Degree sequences 11

3 Connectedness 15

4 Graphs and matrices 22

5 Blocks 24

6 Connectivity 29

7 Trees 34

8 Counting trees 36

9 Permutation groups 39

10 Automorphism groups of graphs 42

11 Eulerian graphs 45

12 Hamiltonian graphs 48

13 Elements of plane topology 51

1
14 Planar graphs 52

15 Kuratowski’s Theorem 56

16 Crossing number 63

17 Vertex colorings 68

18 The Four-color Theorem 72

19 Edge colorings 75

20 Strongly regular graphs 78

21 Moore graphs 80

22 Kneser graphs 86

A More topology 92

0 Preliminaries
We assume that the reader is familiar with the rudiments of set theory,
including the notations x ∈ X to indicate that x is an element of the set
X and x ∈ / X to indicate that it is not, and the use of expressions such as
{x ∈ R | x > 1}, which denotes the set of all real numbers greater than
1. We recall here some (but not all) of the ideas the reader should have
encountered. Two sets are equal iff they have exactly the same elements.
There is a unique set with no elements, the empty set ∅. We use the notation
X ⊆ Y to indicate that X is a subset of Y ; that is, that x ∈ X implies x ∈ Y .
We have X = Y iff X ⊆ Y and Y ⊆ X. We use X ⊂ Y to indicate that X
is a proper subset of Y ; that is, X ⊆ Y and X 6= Y . The intersection, union
and difference (or relative complement) of sets X and Y are defined by

X ∩ Y = {x | x ∈ X and x ∈ Y },
X ∪ Y = {x | x ∈ X or x ∈ Y },
and X − Y = {x | x ∈ X and x ∈
/ Y },

respectively. (We follow the standard mathematical convention that “or” is

used inclusively, so that X ∩ Y ⊆ X ∪ Y .) The Cartesian product X × Y is
the set of all ordered pairs (x, y) with x ∈ X and y ∈ Y . We deal mostly

2
with finite sets. The number of elements, or cardinality of a finite set X will
be denoted by |X|.
We also assume familiarity with the idea of a function, and a few related
notions. The notation f : X → Y means that f is a function from the set X
to the set Y ; that is, to each element x of X is associated a unique element
f (x) of Y . Functions f : X → Y and f ′ : X ′ → Y ′ are equal iff X = X ′ ,
Y = Y ′ and f (x) = f ′ (x) for all x ∈ X. For any set X, the identity function
on X is the function idX (or just id) from X to X defined by idX (x) = x
for all x ∈ X. If f : X → Y and A ⊆ X, the image of A under f is the
subset f (A) = {f (a) | a ∈ A} of Y . The function f is injective (or an
injection) if f (x1 ) = f (x2 ) implies x1 = x2 for x1 , x2 ∈ X. It is surjective
(or a surjection) if f (X) = Y , and it is bijective (or a bijection) if it is both
injective and surjective. If f is a bijection from X to Y , there is an inverse
function f −1 : Y → X defined by f −1 (y) = x iff y = f (x) for x ∈ X and
y ∈Y.
If f : X → Y and g : Y → Z, the composite g ◦ f : X → Z is defined by
(g ◦ f )(x) = g(f (x)). We also use gf to denote the composite. If f : X → Y
is a bijection we have f −1 ◦ f = idX and f ◦ f −1 = idY .
The rest of this section discusses equivalence relations and partitions,
which the reader may not have met before. A binary relation on a set X
is a subset R of X × X. If (x, y) ∈ R, we say that x is related to y by R,
and write x R y. Typically, specific binary relations are denoted by symbols
such as =, ≤, > and ∼ rather than letters. Thus, formally, the relation >
on the set {1, 2, 3} is the set {(2, 1), (3, 1), (3, 2)} of ordered pairs. Let R be
a binary relation on X.

(1) R is reflexive if x R x for all x ∈ X.

(2) R is symmetric if x R y implies y R x for x, y ∈ X.

(3) R is transitive if x R y and y R z imply x R z for x, y and z ∈ X.

An equivalence relation is a relation that is reflexive, symmetric and transi-

tive. If R is an equivalence relation on X and x ∈ X, the equivalence class
of x is the set [x] of all y ∈ X such that x R y.

Example 0.1. The equality relation on any set is an equivalence relation,

and [x] = {x}.

Example 0.2. On the set R of real numbers, the relation ≤ is reflexive and
transitive, but not symmetric.

3
Example 0.3. On the set R, the relation < is transitive, but neither reflex-
ive nor symmetric.

Example 0.4. If n is a positive integer, the relation of congruence modulo

n on the set Z of all integers is defined by a ≡ b (mod n) if n divides a − b.
It is an equivalence relation, and [a] = {qn + a | q ∈ Z}. The distinct
equivalence classes are [0], [1], . . . , [n − 1], which we shall sometimes denote
just by 0, 1, . . . , n − 1. The set of equivalence classes will be denoted by Zn
(a notation that annoys some algebraists of my acquaintance, who prefer
Z/n or Z/nZ). We define operations of addition and multiplication on Zn
by [i] + [j] = [i + j] and [i][j] = [ij]. It is left to the reader to verify that
these are well-defined operations. They satisfy many of the usual laws of
arithmetic (commutativity, associativity, distributivity, . . . ); those who have
met the concept will know that they give Zn the structure of a ring. At one
point we shall need the operation of addition on Z2 × Z2 , defined in the
obvious way: (a, b) + (c, d) = (a + c, b + d).

Proposition 0.5. Let R be an equivalence relation on a set X, and x,

y ∈ X. Then x ∈ [x], and [x] = [y] iff x R y.

Proof. First, x ∈ [x] because R is reflexive. Second, if [x] = [y] then y ∈ [x],
so x R y. Finally, suppose x R y. If z ∈ [y] then y R z, so x R z by transitivity.
This gives z ∈ [x], and so [y] ⊆ [x]. By symmetry of R, y R x, so also
[x] ⊆ [y]; that is, [x] = [y].

A partition of a set X is a collection of disjoint, non-empty subsets of X

whose union is X.

Proposition 0.6. Let X be a set.

(1) The collection of equivalence classes under any equivalence relation

on X is a partition of X.

(2) Every partition of X is the collection of equivalence classes for a

unique equivalence relation on X.

Proof. (1) Let R be an equivalence relation on X, and P the set of equiv-

alence classes under R. The relation x ∈ [x] of Proposition 0.5 shows that
the elements of P are non-empty and have union X. Suppose [x] and [y] are
elements of P that are not disjoint. Let z ∈ [x] ∩ [y]. Then x R z and y R z,
so by Propsition 0.5, [x] = [z] = [y]. Thus P is a partition of X.
(2) Let P be a partition of X, and for x ∈ X let hxi be the unique
element of P containing X. By Proposition 0.5, if P is the set of equivalence

4
 Figure 1: The Petersen graph

classes for an equivalence relation R then x R y iff hxi = hyi, which proves
the uniqueness part. To prove existence, define a relation R by x R y if
hxi = hyi. It is trivial to check that R is an equivalence relation. We must
show that the equivalence class [x] is equal to hxi. Suppose y ∈ [x]. Then
hxi = hyi, and since y ∈ hyi by definition, y ∈ hxi. Conversely, if y ∈ hxi,
then since also y ∈ hyi, hxi and hyi are not disjoint. Hence hxi = hyi, so
x R y and y ∈ [x].

1 Graphs
Definition 1.1. A graph G is a pair (V, E), where V is a non-empty finite
set, and E is a set of unordered pairs of elements of V . The elements of V
are called the vertices of G, and the elements of E are the edges of G. We
always assume that V ∩ E = ∅.

We can represent a graph pictorially as in Figure 1. The dots represent

the vertices, and each edge {u, v} is represented by an arc connecting the
dots corresponding to the vertices u and v. We will usually write uv instead
of {u, v}; of course, uv = vu. The set of vertices of a graph G may be
written V (G); the number |V (G)| of vertices is the order of G, denoted by
n(G) or just n. The set of edges may be written E(G); the number |E(G)|
of edges is the size of G, denoted by m(G) or just m. By definition, n > 0,
but m = 0 is
allowed. If n = 1 (and so m = 0), G is trivial. In general,
0 ≤ m
≤ n2 = n(n − 1)/2. If m = 0 we have an empty graph, while if
m = n2 we have a complete graph; see Figure 2 for an example. Vertices u
and v of a graph G are adjacent if uv ∈ E(G); an edge e = uv is incident
with the vertices u and v, and distinct edges e and f are adjacent if there
is a vertex incident with both of them. Vertices u and v are independent if

5
 Figure 2: The complete graph on 5 vertices

u 6= v and uv ∈
/ E(G). If u and v are adjacent we shall also say that u and
v are neighbors.

Definition 1.2. Let G1 and G2 be graphs. An isomorphism from G1 to G2

is a bijection φ : V (G1 ) → V (G2 ) such that uv ∈ E(G1 ) iff φ(uv) ∈ E(G2 ).
(Here we are using the standard notation for the image of a set under a
function: φ(uv) = φ({u, v}) = {φ(u), φ(v)} = φ(u)φ(v).) If there exists an
isomorphism G1 → G2 , G1 is isomorphic to G2 , written G1 ∼ = G2 .

Note that any graph is isomorphic to itself; if G1 ∼ = G2 then G2 ∼ = G1 ;

∼ ∼ ∼
and if G1 = G2 and G2 = G3 then G1 = G3 . (One is tempted to say
that isomorphism is an equivalence relation on the set of graphs. However,
there is no such thing as the set of all graphs. Isomorphism is an equivalence
relation on any set of graphs.) Isomorphic graphs are “essentially the same”;
they differ only in the names of their vertices. Some authors use G1 = G2 to
mean that G1 and G2 are isomorphic, but I prefer to reserve the equal sign
for, well, equality. Nevertheless, I do follow the universal practice of referring
to the Petersen graph, or the complete graph on n vertices, meaning some
one of a collection of isomorphic graphs. The complete graph on n vertices
will be denoted by Kn .
Sometimes we wish to determine the number of graphs in a given collec-
tion, and the question arises as to whether we are counting the graphs up to
identity or up to isomorphism. We adopt a convention that should be clear
from the following example. If X is a given set of n elements, “the number
of graphs with vertex set X” (or “on X”) means precisely what it says: the
n
number of collections of unordered pairs of elements of X, which is 2( 2 ) .
However, “the number of graphs of order n” (or “on n vertices”) means the
maximum number in a collection of pairwise non-isomorphic graphs of order
n, or equivalently the number of isomorphism classes in the set of all graphs
on a fixed set of size n. This number has no known formula; the reader may
care to verify that the numbers of graphs of orders 1, 2, 3, 4 and 5 are 1, 2,

6
 G1 G2 G3

Ḡ1 Ḡ2 Ḡ3

Figure 3: The graphs of order 4 and size 3

4, 11 and 34, respectively. (For order 5, particularly, you may want to use
the following definition to reduce the tedium.)

Definition 1.3. The complement of a graph G is the graph Ḡ with V (Ḡ) =

V (G) in which, for distinct vertices u and v, uv ∈ E(Ḡ) iff uv ∈/ E(G).
Put another way, if K is the complete graph with vertex set V (G), E(Ḡ) =
E(K) − E(G).

For example, the complement K̄n of the complete graph on n vertices

is the empty graph on n vertices. If G has order n and size m then Ḡ has
n ∼


size 2 − m. A
graphn
is self-complementary if G = Ḡ. If this is so we must
n
have 2m = 2 , so 2 is even, which happens iff n ≡ 0 or 1 (mod 4). In
fact, for any such n there exist self-complementary graphs of order n. For
n = 4, the size must be 3. In Figure 3, the top row shows all graphs of order
4 and size 3, while the bottom row shows their complements. We see that
G1 and G3 are (up to isomorphism) complements of each other, while G2 is
self-complementary.
The degree of a vertex v of a graph G is the number of vertices adjacent
to v. It is denoted by degG v or just deg v. With this definition we can
finally state a theorem.
P
Theorem 1.4. Let G be a graph of size m. Then v∈V (G) deg v = 2m.

Proof. We can count the number p of pairs (v, e) where v is a vertex of the
edge e in two ways. First, there are m edges, each with two vertices,
P so p =
2m. Second, the vertex v is incident to deg v edges, so p = v∈V (G) deg v.

A vertex is called even or odd according as its degree is even or odd.

Corollary 1.5. Any graph contains an even number of odd vertices.

7
 The minmum degree of the vertices of a graph is denoted by δ(G), and
the maximum by ∆(G). Clearly 0 ≤ δ(G) ≤ ∆(G) ≤ n(G) − 1. A vertex
of degree 0 is an isolated vertex, while one of degree 1 is an end-vertex. If
δ(G) = ∆(G) then G is regular ; if the common value is d, G is d-regular
or regular of degree d. A 3-regular graph is also called a cubic graph. The
complete graph Kn is (n − 1)-regular, and the Petersen graph of Figure 1
is cubic. For a regular graph, any two of the order n, the size m and the
degree d determine the third by the equation 2m = nd.
A subgraph of a graph G is a graph H with V (H) ⊆ V (G) and E(H) ⊆
E(G). We write H ⊆ G. If U is any proper subset of V (G), G − U is the
subgraph obtained by deleting all vertices of U and all edges with at least
one vertex in U . When U = {u}, we abbreviate G − {u} to G − u. Similarly,
if F is a subset of E(G), G−F is the subgraph obtained by deleting the edges
of F , and G−{e} is written as G−e. (This notation is logically indefensible,
since an edge is a set of vertices. Nevertheless, for an edge e = uv = {u, v},
we adopt the convention that G − e and G − uv denote G with just an edge
deleted, and G − {u, v} denotes G with two vertices and all incident edges
deleted.) Note that V (G − F ) = V (G), and F = E(G) is allowed, with
G − E(G) being the empty graph on V (G). A spanning subgraph of G is
a subgraph of the same order (i.e., containing all the vertices of G). These
are precisely the subgraphs of the form G − F for F ⊆ E(G). If u and v are
independent vertices of G and f = uv, the graph G + f has vertex set V (G)
and edge set E(G) ∪ {f }.
If U is a non-empty set of vertices of the graph G, the subgraph hU i
induced by U is the subgraph with vertex set U and edge set consisting of
all edges of G with both vertices in U . A vertex-induced subgraph (or just
an induced subgraph) is a subgraph of this form. If F is a non-empty set
of edges of G, the subgraph hF i induced by F is the subgraph with vertex
set all vertices of edges of F and edge set F . An edge-induced subgraph is a
subgraph of this form.

Theorem 1.6 (König [22]). Let G be a graph and d an integer with d ≥

∆(G). Then there is a d-regular graph containing G as an induced subgraph.

Proof. The proof is by induction on d − δ(G). If this is 0, G is d-regular

and there is nothing to do. If d > δ(G), let G′ be another copy of G and
form H by adding edges joining each vertex of G of degree δ(G) to the
corresponding vertex of G′ . Then H contains G as an induced subgraph,
and has δ(H) = δ(G) + 1 and ∆(H) ≤ d. The result follows.

The graph constructed in this proof is generally not of minimal order;

8
 Figure 4: The complete bipartite graph K3,3

G1 G2 G1 ∗ G2

Figure 5: The join of two graphs

see Exercise 1.7.

A graph G is bipartite if there is a partition of V (G) into two sets V1
and V2 , called partite sets, such that every edge has one vertex in V1 and
one in V2 . It is a complete bipartite graph if uv is an edge for all u ∈ V1
and v ∈ V2 . The complete bipartite graph with r vertices in one partite set
and s in the other is denoted by Kr,s ; see Figure 4 for an example. The
complete bipartite graph K1,n is called a star. More generally, G is k-partite
if there is a partition of V (G) into k subsets (k ≥ 2) such that the vertices
of any edge are in different sets of the partition, and complete k-partite if
every pair consisting of two vertices from different sets is an edge. The
complete k-partite graph with partite sets of sizes n1 , n2 , . . . , nk is denoted
by Kn1 ,n2 ,...,nk . A graph is (complete) multipartite if it is (complete) k-partite
for some k ≥ 2. (We exclude k = 1 because the only 1-partite graphs would
be empty, and we would have a conflict in the use of Kn . At the other
extreme, any graph of order n is n-partite, with the partite sets being the
one-vertex sets, and K1,1,...,1 is a superfluous notation for Kn .)

G1 × G2

Figure 6: The product of two graphs

9
 Let G1 and G2 be graphs. Their union G1 ∪ G2 is the graph with vertex
set V (G1 )∪V (G2 ) and edge set E(G1 )∪E(G2 ). Their disjoint union G1 ⊔G2
is any graph of the form G′1 ∪ G2 where G′1 ∼ = G1 and V (G′1 ) is disjoint from
V (G2 ); it is well-defined up to isomorphism. The disjoint union of k copies
of a graph G is denoted kG. The join of G1 and G2 is the graph G1 ∗ G2
obtained from G1 ⊔ G2 by adding an edge v1 v2 for each vertex v1 of G1 and
v2 of G2 ; see Figure 5. Note that Ḡ1 ∗ Ḡ2 ∼ = G1 ⊔ G2 . The complete bipartite
graph Kr,s can be described as either K̄r ∗ K̄s or Kr ⊔ Ks . The Cartesian
product G1 × G2 has vertex set V (G1 ) × V (G2 ), and there is an edge with
vertices (u1 , u2 ) and (v1 , v2 ) iff either u1 v1 ∈ E(G1 ) and u2 = v2 , or u1 = v1
and u2 v2 ∈ E(G2 ). Figure 6 shows the product of the graphs G1 and G2 of
Figure 5.
Two classes of graphs we shall meet frequently are the paths and cycles.
The path Pn (n ≥ 1) of order n has vertices v1 , . . . , vn and edges vi vi+1
for 1 ≤ i < n. Of course, P1 ∼ = K1 and P2 ∼ = K2 . For n ≥ 3, the cycle
Cn is obtained from Pn by adding the edge vn v1 . It can be represented
as a regular n-gon. For small values of n we refer to Cn as a triangle,
quadrilateral, pentagon, . . . . Another special class of graphs consists of the
cubes. The n-cube Qn (n ≥ 0) is defined recursively by setting Q0 = K1
and Qn+1 = Qn × K2 . It may also be defined as the graph with vertices all
sequences of zeroes and ones of length n, with two sequences being adjacent
iff they differ in exactly one place.

Remark. The definition of a graph is mutable. A variant is that a graph

G consists of disjoint sets V (G) of vertices and edges E(G), together with
a relation of incidence between vertices and edges such that every edge is
incident with either one or two vertices. (Some authors also allow edges
incident with no vertices; well, I have, at times.) Edges incident with only
one vertex are loops, while distinct edges incident with the same vertices are
parallel. We shall refer to such objects as multigraphs, and hardly ever men-
tion them again. Obviously every graph determines a multigraph. Further,
every multigraph may be converted to a graph by subdividing its edges (see
§15) while preserving most of its interesting properties.

Exercises for §1

1.1. Determine all non-isomorphic graphs of order at most 5.

1.2. Show that in any non-trivial graph there are two distinct vertices with
the same degree.

10
1.3. Let G be a self-complementary graph, and let G′ be obtained from
G ⊔ P4 by adding edges from every vertex of G to the two end-vertices
of P4 . Show that G′ is self-complementary, and deduce that there are
self complementary graphs of order n for every positive integer n with
n ≡ 0 or 1 (mod 4).

1.4. Let G be a self-complementary graph of order n ≡ 1 (mod 4). Prove

that G has an odd number of vertices of degree (n − 1)/2.

1.5. Give an example of two non-isomorphic regular graphs of order 6 and

degree 3. [Hint: examine the figures in this section.]

1.6. For any integer k ≥ 2, show that there are k non-isomorphic regular
graphs, all of the same order and degree. [Hint: the disjoint union of
regular graphs of the same degree is regular.]

1.7. Show that the minimum order of a 3-regular graph containing the graph
G below as an induced subgraph is 6. Also draw the 3-regular graph
constructed from G as in the proof of Theorem 1.6.

2 Degree sequences
Let G be a graph with vertices v1 , . . . , vn , and set di = deg vi . The sequence
(d1 , . . . , dn ) is called a degree sequence of G. A sequence (d1 , . . . , dn ) of
integers is graphical if it is a degree sequence of some graph. Necessary
conditions for the sequence to be graphical are that 0 ≤ di ≤ n − 1 for all i
and (by Theorem 1.4) that d1 + · · · + dn is even. These are not sufficient; it
is easy to see that (3, 3, 3, 1) is not graphical.

Theorem 2.1 (Havel [17], Hakimi [15]). Let d = (d1 , . . . , dn ) be a sequence

of integers with n ≥ 2, d1 ≥ · · · ≥ dn ≥ 0 and d1 > 0. Then d is graphical
iff d1 ≤ n − 1 and the sequence

d′ = (d2 − 1, . . . , dd1 +1 − 1, dd1 +2 , . . . , dn )

is graphical.

11
 Evidently this theorem gives an algorithm for determining whether a
sequence d of integers is graphical. The only graphical sequence of length
1 is (0), any sequence consisting entirely of zeroes is graphical, and any
sequence with negative terms is not graphical. For any other sequence,
arrange it in descending order. If d1 ≥ n, it is not graphical. Otherwise,
form the sequence d′ as above, and repeat the process until a decision is
reached. For d = (3, 3, 3, 1), we get d′ = (2, 2, 0) and then d′′ = (1, −1), so
indeed d is not graphical.

Proof. Sufficiency is easy. Suppose that d1 ≤ n − 1 and d′ is graphical.

Let G′ be a graph with vertices v2 , . . . , vn such that degG′ vi = di − 1 for
2 ≤ i ≤ d1 + 1 and degG′ vi = di for d1 + 2 ≤ i ≤ n. Form G by adding to
G′ a new vertex v1 and edges v1 vi for 2 ≤ i ≤ d1 + 1. Then degG vi = di for
1 ≤ i ≤ n.
Conversely, suppose that d is graphical. Certainly d1 ≤ n − 1. Consider
all graphs G of order n and labellings of their vertices as v1 , . . . , vn for which
deg vi = di , 1 ≤ i ≤ n. Amongst all such, choose one for which the sum
of the degrees of the vertices adjacent to v1 is a maximum. We claim that
in this G, the degrees of the vertices adjacent to v1 are d2 , . . . , dd1 +1 , from
which it will follow that G − v1 has d′ as a degree sequence, completing the
proof.
Suppose the claim is false. Then there exist integers r and s with 2 ≤
r < s ≤ n, dr > ds , and vr not adjacent to v1 and vs adjacent to v1 . Since
deg vr > deg vs , there is a vertex vt adjacent to vr and independent of vs ;
vt is automatically distinct from v1 and vr . Deleting the edges v1 vs and
vr vt and adding edges v1 vr and vs vt gives a graph G′ with degG′ vi = di for
1 ≤ i ≤ n, but the sum of the degrees of the vertices adjacent to v1 in G′ is
greater than the sum for G by dr − ds , a contradiction.

Theorem 2.2 (Erdös-Gallai [11]). Let d = (d1 , . . . , dn ) be a sequence

P of
integers with n ≥ 2 and d1 ≥ · · · ≥ dn ≥ 0. Then d is graphical iff ni=1 di
is even and, for 1 ≤ k ≤ n − 1,
k
X n
X
(EG) di ≤ k(k − 1) + min{di , k}.
i=1 i=k+1

Proof. We first show the necessity of the conditions. That ni=1 di must be
P
even has already been remarked. Let G be a graph of order n with vertices
v1 , . . . , vn for which deg vi = di , 1 ≤ i ≤ n, and let 1 ≤ k ≤ n − 1. Consider
all ordered pairs (i, j) with 1 ≤ i ≤ k and 1 ≤ j ≤ n for which vi vj ∈ E(G).

12
For 1 ≤ i ≤ k there are di such pairs, so the total number of pairs is ki=1 di .
P
The number of such pairs with j ≤ k is at most k(k − 1), and for each j > k
the number of such pairs is at most min{dj , k}, so (EG) holds.
For the converse, we prove the stronger result that if ni=1 di is even
P
and (EG) holds for all k with 1 ≤ k ≤ dk then G is graphical. (It is easy
to see that these cases of (EG) imply the remainder, but we do not need
this.) If n = 2, the only such sequences are (0, 0) and (1, 1), which are
graphical. Suppose then that n > 2 and the result holds for sequences of
length n − 1. If d1 = 0 then d is the degree sequencePof an empty graph.
Suppose that d1 > 0. The case k = 1 of (EG) is d1 ≤ ni=2 min{di , 1}, and
the right-hand side of this inequality is the number of i ≥ 2 with di 6= 0.
Thus dˆ = dd1 +1 is positive. Let a and c be the least and greatest elements
ˆ We have a ≤ d1 + 1 ≤ c, so
of the set of those i ∈ {2, . . . , n} with di = d.
setting b = a + c − d1 − 1 we have a ≤ b ≤ c. Partition the set {2, . . . , n}
into R = {i | 2 ≤ i < a or b ≤ i ≤ c} and S = {i | a ≤ i < b or c < i ≤ n},
and note that |R| = d1 . For 2 ≤ i ≤ n, let d′i = di − 1 if i ∈ R and
d ′ = d if i ∈ S. The sequence d′ = (d′ , . . . , d′ ) has d′ ≥ · · · ≥ d′ ≥ 0 and
i i 2 n 2 n
P n ′
Pn
i=2 i =
d i=1 di −2d1 . (It is just the sequence of Theorem 2.1 arranged in
descending order.) We show that it satisfies the inequalities corresponding
to (EG), namely
k
X n
X
d′i ≤ (k − 1)(k − 2) + min{d′i , k − 1},
i=2 i=k+1

for 1 ≤ k −1 ≤ d′k . Once this has been done, it will follow from the inductive
hypothesis that d′ is graphical, and then that d is graphical as in the earlier
proof. For i ≥ c we have
d′i ≤ d′c = dˆ − 1 ≤ d1 − 1 ≤ c − 2 ≤ i − 2,
so k − 1 ≤ d′k implies k < c. Suppose first that k < a. If k ≤ d, ˆ then
′
min{di , k − 1} = k − 1 = min{di , k − 1} for i ∈ R, while clearly min{d′i , k −
1} = min{di , k − 1} for i ∈ S. Hence
k
X k
X k−1
X
d′i = di − k + 1 ≤ di − k + 1
i=2 i=2 i=1
Xn
≤ (k − 1)(k − 2) + min{di , k − 1} − k + 1
i=k
Xn
= (k − 1)(k − 2) + min{d′i , k − 1}.
i=k+1

13
If k ≥ dˆ + 1, then min{d′i , k − 1} = di = min{di , k} for i ∈ S, and of course
min{d′i , k − 1} = min{di , k} − 1 for i ∈ R. Therefore
k
X k
X
d′i = di − d1 − k + 1
i=2 i=1
n
X
≤ k(k − 1) + min{di , k} − d1 − k + 1
i=k+1
Xn
= k(k − 1) + min{d′i , k − 1} + (d1 − k + 1) − d1 − k + 1
i=k+1
n
X
= (k − 1)(k − 2) + min{d′i , k − 1}.
i=k+1

For the rest of the proof, we assume that a ≤ k. We distinguish three

cases.
ˆ
Case 1. a ≤ k < b and k ≤ d.

Case 2. a ≤ k < b and k = dˆ + 1.

ˆ
Case 3. b ≤ k ≤ d.

We have
k
X k
X
d′i ≤ (a − 2)(d1 − 1) + d′i
i=2 i=a
= (a − 2)(d1 − 1) + (k − a + 1)dˆ − xk ,

where xk is 0 in cases 1 and 2, and k − b + 1 in case 3. Also,

n
X
(k − 1)(k − 2) + min{d′i , k − 1}
i=k+1
c
X
≥ (k − 1)(k − 2) + min{d′i , k − 1}
i=k+1
= (k − 1)(k − 2) + (c − k)(k − 1) − yk
= (k − 1)(c − 2) − yk ,

14
where yk is 0 in cases 1 and 3, and c − b + 1 in case 2. It is therefore enough
to prove that zk ≥ 0, where

zk = (k − 1)(c − 2) − yk − ((a − 2)(d1 − 1) + (k − a + 1)dˆ − xk )

= (a − 2)(c − d1 − 1) + (k − a + 1)(c − dˆ − 2) + xk − yk .

Case 1. Here zk = (a − 2)(c − d1 − 1) + (k − a + 1)(c − dˆ − 2), and since

a < b we have dˆ ≤ d1 ≤ c − 2, so zk ≥ 0.
Case 2. Here

zk = (a − 2)(c − d1 − 1) + (k − a + 1)(c − dˆ − 2) − (c − b + 1)
= (a − 2)(c − d1 − 1) + (k − a)(c − dˆ − 2) + b − dˆ − 3.

As in the previous case, dˆ ≤ d1 ≤ c − 2, and also b ≥ k + 1 = dˆ + 2, so

zk ≥ 0 except perhaps if a = 2 and b = dˆ +P2. In this case c = dˆ + d1 + 1
and zk = (dˆ − 1)(d1 − 1) − 1. If dˆ = 1 then ni=1 di = d1 + c − 1 = 2d1 + 1,
which is impossible. Hence 1 < dˆ ≤ d1 and zk ≥ 0.
Case 3. Here

zk = (a − 2)(c − d1 − 1) + (k − a + 1)(c − dˆ − 2) + k − b + 1,

and since dˆ ≤ d1 ≤ c − 1, zk > 0 unless dˆ = d1 = c − 1, and in that case

b = a so zk = 0.

Exercises for §2
2.1. Let n and d be integers with 0 ≤ d < n and dn even. Prove that there
is a d-regular graph of order n.

2.2. Come up with 5 different sequences (d1 , . . . , d7 ) 0f integers with d1 ≥

· · · ≥ d7 > 0, d1 + · · · + d7 even and d1 ≤ 6, of which 2 are graphical
and 3 are not. Prove that those that aren’t graphical aren’t, and give
drawings showing that the other two are.

3 Connectedness
Definition 3.1. A walk in a graph G is a sequence W = (u0 , u1 , . . . , ul )
(l ≥ 0) of vertices of G such that ui−1 and ui are adjacent for 1 ≤ i ≤ l.
The length of W is l (the number of edges on W rather than the number
of vertices). We say that W is a u0 -ul walk, or a walk from u0 to ul . If

15
u0 = ul , W is a closed walk; otherwise it is open. A trail is a walk with
no repeated edges (ui ui+1 6= uj uj+1 for i 6= j). A path is a walk with no
repeated vertices (ui 6= uj for i 6= j). Every path is a trail, and for any
vertex u the trivial u-u walk (u) is a path. A circuit is a non-trivial closed
trail, and a cycle is a closed walk (u0 , u1 , . . . , ul ) of length l ≥ 3 such that
ui 6= uj for 1 ≤ i < j ≤ l. Every cycle is a circuit, and every circuit has
length at least 3.

Note that the subgraph induced by the edges of a path in G of length l

is Pl+1 , and that induced by the edges of a cycle of length l is Cl . We shall
often confuse a walk (trail, path, circuit, cycle) with the subgraph induced
by its edges. If W is a u-v walk, we refer to u and v as the endpoints of W ,
and when W is a path any other vertices on W are called internal vertices
of W . We say that a walk (u0 , u1 , . . . , ul ) contains a walk (v0 , v1 , . . . , vk ) if
there are integers 0 ≤ i0 < i1 < · · · < ik ≤ l such that vj = uij for 0 ≤ j ≤ k.

Theorem 3.2. Any u-v walk in a graph G contains a u-v path.

Proof. The proof is by induction on the length. Suppose

W = (u = u0 , u1 , . . . , ul = v)

is a u-v walk, and assume inductively that every u-v walk (if any) of length
less than l contains a u-v path. If W is a path, there is nothing to do.
Otherwise, there exist i and j with 0 ≤ i < j ≤ l and ui = uj . Now W
contains the u-v walk (u0 , u1 , . . . , ui , uj+1 , . . . , ul ) of length less than l, which
in turn contains a u-v path.

Definition 3.3. We say that vertices u and v of a graph G are connected

if there is a u-v path in G, and in this case we write u ∼ v. The graph G is
connected if any two of its vertices are connected.

Note that by Theorem 3.2, u and v are connected iff there is a u-v
walk. If W = (u0 , u1 , . . . , ul ) is a walk, we denote by W r the reverse walk
(ul , ul−1 , . . . , u0 ). If W ′ = (v0 , v1 , . . . , vk ) is another walk with v0 = ul we
denote by W · W ′ the walk (u0 , u1 , . . . , ul , v1 , . . . , vk ).

Lemma 3.4. Let G be a graph. The relation of Definition 3.3 is an equiv-

alence relation on V (G).

Proof. Let u, v and w be vertices of G. The trivial u-u path shows that
u ∼ u. If u ∼ v, let P be a u-v path. Then P r is a v-u path, so v ∼ u. If
also v ∼ w, let Q be a v-w path. Then P · Q is a u-w walk, so u ∼ w.

16
 The subgraphs of G induced by the equivalence classes for the relation
∼ are called the components of G. Each component is connected, and every
connected subgraph of G is contained in a component. The number of
components is denoted by k(G). Thus k(G) = 1 iff G is connected.

Definition 3.5. Let u and v be vertices of a connected graph G. The

distance from u to v, d(u, v) = dG (u, v), is the minimum of the lengths of
all paths from u to v in G. A u-v path of length d(u, v) is a u-v geodesic.

By Theorem 3.2, d(u, v) is also the minimum of the lengths of all walks
from u to v. In a disconnected graph, we may define d(u, v) as above for
vertices in the same component, and set it to ∞ for vertices in different
components.

Theorem 3.6. Let u, v and w be vertices of a connected graph G.

(1) d(u, u) = 0 and d(u, v) > 0 if u 6= v.

(2) d(u, v) = d(v, u).

(3) d(u, w) ≤ d(u, v) + d(v, w).

Proof. (1) is obvious. The other parts follow by examining the lengths of
the walks used to show symmetry and transitivity in the proof of Lemma
3.4.

A cycle in a graph is called odd or even according as its length is odd or

even.

Theorem 3.7. A non-trivial graph G is bipartite iff it has no odd cycles.

Proof. We may assume that G is connected, since a non-trivial graph is

bipartite iff all its non-trivial components are bipartite. Suppose first that
G is bipartite, with partite sets V1 and V2 , and let (u0 , u1 , . . . , uk ) be a cycle
in G. Without loss of generality, u0 ∈ V1 . It follows by an easy induction
that ui ∈ V1 if i is even and ui ∈ V2 if i is odd. Since uk = u0 ∈ V1 , k is
even.
Suppose conversely that G has no odd cycles, and pick a vertex w. Let
V1 be the set of those u ∈ V (G) with d(u, w) even, and V2 the set of those
with d(u, w) odd. We have w ∈ V1 , and since G is connected and non-
trivial, there is a vertex adjacent to w, and therefore in V2 . Thus {V1 , V2 }
is a partition of V (G). We show that G is bipartitite with respect to this
partition. Suppose for a contradiction that u and v are adjacent vertices in

17
the same set of the partition; that is, with d(u, w) ≡ d(v, w) (mod 2). Let
(w = u0 , u1 , . . . , uk = u) and (w = v0 , v1 , . . . , vl = v) be geodesics, so that
k ≡ l (mod 2). For 0 ≤ i ≤ j ≤ k, d(ui , uj ) = j − i, since otherwise we
could replace the subpath (ui , . . . , uj ) by a shorter path to obtain a shorter
w-u walk. Similarly, d(vi , vj ) = j − i for 1 ≤ i ≤ j ≤ l. Hence if ui = vj then
i = d(w, ui ) = d(w, vj ) = j. Let i be the greatest integer for which ui = vi .
Since u 6= v, (k −i)+(l −i) ≥ 1, and since k ≡ l (mod 2), (k −i)+(l −i) ≥ 2.
Now (ui , . . . , uk−1 , u, v, vl−1 , . . . , vi ) is a cycle of odd length k +l −2i+1.

Definition 3.8. Let G be a connected graph. The eccentricity eG (u) = e(u)

of a vertex u of G is defined by

e(u) = max{d(u, v) | v ∈ V (G)}.

The radius rad G of G is the minimum of e(u) over all vertices u, while the
diameter diam G is the maximum of e(u). We also have

diam G = max{d(u, v) | u, v ∈ V (G)}.

A vertex u is central if e(u) = rad G and peripheral if e(u) = diam G, and

the center Cen G and periphery Per G of G are the subgraphs induced by
the central and peripheral vertices, respectively.

Example 3.9. rad Kn = diam Kn = 1 and Cen Kn = Per Kn = Kn for

n ≥ 2.

Example 3.10. rad Cn = diam Cn = ⌊n/2⌋ and Cen Cn = Per Cn = Cn .

Example 3.11. Let Pn have vertices v1 , . . . , vn in that order, where n ≥ 3.

We have diam Pn = n − 1 and Per Pn = hv1 , vn i ∼ = K̄2 . If n = 2k + 1 then
rad Pn = k and Cen Pn = hvk+1 i ∼
= K1 , while if n = 2k then rad Pn = k and
Cen Pn = hvk , vk+1 i ∼
= K2 .

Theorem 3.12. For any connected graph G, rad G ≤ diam G ≤ 2 rad G.

Proof. That rad G ≤ diam G is immediate from the definitions. For the
other inequality we must show that d(u, v) ≤ 2 rad G for any vertices u and
v. If w is a central vertex we have

d(u, v) ≤ d(u, w) + d(w, v) ≤ 2e(w) = 2 rad G.

Theorem 3.13 (Hedetniemi; see [5]). Any graph G is the center of some
connected graph.

18
Proof. Form a graph H by adding to G four new vertices v1 , v2 , w1 and w2 ,
and edges uvi and vi wi for u ∈ V (G) and i = 1 or 2. Then H is connected,
eH (u) = 2 for u ∈ V (G), eH (v1 ) = eH (v2 ) = 3 and eH (w1 ) = eH (w2 ) = 4,
so G = Cen H.

Theorem 3.14 (Bielak and Syslo [3]). A graph G is the periphery of some
connected graph iff G is complete or no vertex of G has eccentricity 1.

If G is not connected, the eccentricity of any vertex is infinite, so the

theorem asserts that G is the periphery of some connected graph.

Proof. A complete graph is its own periphery, so suppose G is not complete.

Suppose first that there is a connected graph H with G = Per H. Then H is
also not complete, so diam H ≥ 2. For u ∈ V (G), eG (u) ≥ eH (u) = diam H,
so no vertex of G has eccentricity 1. Conversely, if no vertex of G has
eccentricity 1, set H = K1 ∗G, with new vertex v. For u ∈ V (G), eH (u) = 2,
while eH (v) = 1, so G = Per H.

Definition 3.15. If u is a vertex of theP

connected graph G, the total distance
td(u) = tdG (u) is defined by td(u) = v∈V (G) d(u, v). A median vertex of
G is a vertex with minimum total distance, and the median of G is the
subgraph Med G induced by the median vertices.

Theorem 3.16 (Hendry [18]). For any graphs G1 and G2 , there is a con-
nected graph H with Cen H ∼
= G1 and Med H ∼
= G2 .

Proof. We may assume that V (G1 ) and V (G2 ) are disjoint. For i = 1 or
2, set ni = n(Gi ) and δi = δ(Gi ). Let a and b be positive integers with
a > (n2 − n1 − δ2 − 3)/2 and b > 2a + n1 + n2 − δ2 . Form H by adding to
G1 ∪ G2 vertices xi for 1 ≤ i ≤ a + 2, yi for 1 ≤ i ≤ a and zi for 1 ≤ i ≤ b,
and edges

xi xi+1 for 1 ≤ i < a + 2;

yi yi+1 for 1 ≤ i < a;
zi zj for 1 ≤ i < j ≤ b;
u1 x1 and u1 y1 for u1 ∈ V (G1 ); and
u2 ya and u2 zi for u2 ∈ V (G2 ) and 1 ≤ i ≤ b − degG2 u2 + δ2 .

In what follows, d, e and td denote distance, eccentricity and total distance

in H. For distinct vertices ui and u′i of Gi (i = 1 or 2) we have d(ui , u′i ) = 1
if ui u′i ∈ E(Gi ) and d(ui , u′i ) = 2 otherwise. Let u2 be a vertex of G2 .

19
Since b − degG2 u2 + δ2 > 0, u2 is adjacent to z1 , and so d(u2 , zi ) ≤ 2 for
1 ≤ i ≤ b. Since b − degG2 u2 + δ2 ≤ b, the number of zi with d(u2 , zi ) = 1 is
b − degG2 u2 + δ2 . The remaining distances in H are given below; for those
involving a vertex zi , note that zi is adjacent to at least one vertex of G2 ,
namely a vertex of minimum degree in G2 . We have

d(xi , xj ) = j − i for 1 ≤ i < j ≤ a + 2;

d(xi , u1 ) = i for 1 ≤ i ≤ a + 2 and u1 ∈ V (G1 );
d(xi , yj ) = i + j for 1 ≤ i ≤ a + 2 and 1 ≤ j ≤ a;
d(xi , u2 ) = i + a + 1 for 1 ≤ i ≤ a + 2 and u2 ∈ V (G2 );
d(xi , zj ) = i + a + 2 for 1 ≤ i ≤ a + 2 and 1 ≤ j ≤ b;
d(u1 , yi ) = i for u1 ∈ V (G1 ) and 1 ≤ i ≤ a;
d(u1 , u2 ) = a + 1 for u1 ∈ V (G1 ) and u2 ∈ V (G2 );
d(u1 , zi ) = a + 2 for u1 ∈ V (G1 ) and 1 ≤ i ≤ b;
d(yi , yj ) = j − i for 1 ≤ i < j ≤ a;
d(yi , u2 ) = a − i + 1 for 1 ≤ i ≤ a and u2 ∈ V (G2 );
d(yi , zj ) = a − i + 2 for 1 ≤ i ≤ a and 1 ≤ j ≤ b; and
d(zi , zj ) = 1 for 1 ≤ i < j ≤ b.

Hence

e(xi ) = i + a + 2 for 1 ≤ i ≤ a + 2;
e(u1 ) = a + 2 for u1 ∈ V (G1 );
e(yi ) = i + a + 2 for 1 ≤ i ≤ a;
e(u2 ) = 2a + 3 for u2 ∈ V (G2 ); and
e(zi ) = 2a + 4 for 1 ≤ i ≤ b,

so Cen H = G1 . Further,
i−1
X a−i+2
X i+a
X
td(xi ) = j+ j + in1 + j + (i + a + 1)n2 + (i + a + 2)b
j=1 j=1 j=i+1

for 1 ≤ i ≤ a + 2;
a+2
X a
X
td(u1 ) ≥ j + n1 − 1 + j + (a + 1)n2 + (a + 2)b
j=1 j=1

20
for u1 ∈ V (G1 );
i+a+2
X i−1
X a−i
X
td(yi ) = j + in1 + j+ j + (a − i + 1)n2 + (a − i + 2)b
j=i+1 j=1 j=1

for 1 ≤ i ≤ a;
2a+3
X a
X
td(u2 ) = j + (a + 1)n1 + j + 2n2 + b − δ2 − 2
j=a+2 j=1

for u2 ∈ V (G2 ); and

2a+4
X a+1
X
td(zi ) ≥ j + (a + 2)n1 + j + n2 + b − 1
j=a+3 j=2

for 1 ≤ i ≤ b. For 1 ≤ i < a + 2

td(xi+1 ) − td(xi ) = 2i + n1 + n2 + b − 2 > 0,
so the minimum value of td(xi ) is
a+1
X a+1
X
td(x1 ) = j + n1 + j + (a + 2)n2 + (a + 3)b,
j=1 j=2

while for 1 ≤ i < a

td(yi ) − td(yi+1 ) = −2i − n1 + n2 + b − 2 ≥ −2a − n1 + n2 + b > 0,
so the maximum and minimum values of td(yi ) are
a+3
X a−1
X
td(y1 ) = j + n1 + j + an2 + (a + 1)b
j=2 j=1
2a+2
X a−1
X
and td(ya ) = j + an1 + j + n2 + 2b,
j=a+1 j=1

respectively. We have, for u1 ∈ V (G1 ), u2 ∈ V (G2 ) and 1 ≤ i ≤ b,

td(x1 ) − td(y1 ) = 2n2 + 2b − 4 > 0;
td(u1 ) − td(y1 ) ≥ n2 + b − 3 > 0;
td(ya ) − td(u2 ) = −2a − n1 − n2 + b + δ2 > 0; and
td(zi ) − td(u2 ) ≥ 2a + n1 − n2 + δ2 + 3 > 0,
so Med H = G2 .

21
Exercises for §3
3.1. Prove that any graph G contains a path of length δ(G).

4 Graphs and matrices

Definition 4.1. Let G be a graph of order n and size m, with vertices
v1 , . . . , vn and edges e1 , . . . , em . The adjacency matrix of G is then n × n
matrix A = A(G) with entries
(
1 if vi vj ∈ E(G);
aij =
0 if vi vj ∈
/ E(G).

The degree matrix of G is the n×n diagonal matrix D = D(G) with diagonal
entries dii = deg vi , and the Lagrangian matrix is L = L(G) = D − A.
Finally, if m > 0, the incidence matrix of G is the m × n matrix B = B(G)
with entries (
1 if ei is incident to vj ;
bij =
0 if not.
Evidently an n × n 0-1 matrix (a matrix with all entries 0 or 1) is the
adjacency matrix of some graph iff it is symmetric and has all diagonal
entries 0, and an m × n 0-1 matrix is the incidence matrix of some graph iff
every row has exactly two non-zero entries and different rows are unequal.
Either the adjacency or the incidence matrix determines the graph up to
isomorphism.
Theorem 4.2. Let G be a graph of order n with vertices v1 , . . . , vn and
(l) (l)
adjacency matrix A = [aij ]. For any integer l ≥ 0, let Al = [aij ]. Then aij
is the number of vi -vj walks of length l.
Proof. This is by induction on l, the case l = 0 being obvious. Suppose the
result holds for some l. Then
n
(l+1) (l) (l)
X X
aij = aik akj = aik ,
k=1 k:vk vj ∈E(G)

which is the number of walks of length l from vi to a vertex adjacent to vj .

This is clearly the number of walks of length l + 1 from vi to vj .

The adjacency matrix A of a graph G of order n, being a real symmetric

matrix, is diagonalizable over R; that is, there is a basis of Rn consisting of
eigenvectors of A.

22
Theorem 4.3. Let G be a graph of order n with vertices v1 , . . . , vn and adja-
cency matrix A. If G is d-regular, A has x1 = (1, 1, . . . , 1) as an eigenvector
of eigenvalue d, and for any eigenvalue λ, |λ| ≤ d. If also G is connected,
the multiplicity of the eigenvalue d is 1.

Proof. For any graph G, Ax1 = (d1 , d2 , . . . , dn ) where di = deg vi , giving

the first assertion. Suppose that G is d-regular, and let x = (x1 , . . . , xn ) be
an eigenvector of eigenvalue λ. Then, for 1 ≤ i ≤ n,

xj(i,1) + xj(i,2) + · · · + xj(i,d) = λxi ,

where vj(i,1) , vj(i,2) , . . . , vj(i,d) are the vertices adjacent to vi . Take i so that
|xi | is a maximum. Then |λxi | ≤ dk=1 |xj(i,k) | ≤ d|xi |, so |λ| ≤ d. If λ = d,
P
we must have
xj(i,1) = xj(i,2) = · · · = xj(i,d) = xi .
It follows that xj = xk if vj and vk are in the same component, completing
the proof.

Theorem 4.4. Let A be the adjacency matrix of a connected graph G. If A

has at most two eigenvalues then G is complete.

Proof. Let G have order n. If n = 1 then G = ∼ K1 , so suppose n ≥ 2. A

diagonalizable matrix with only one eigenvalue is a scalar multiple of the
identity, so A has at least two eigenvalues. Suppose it has exactly two. Then
its minimal polynomial has degree 2, so A2 = aA + bI for some a and b, and
every diagonal entry of A2 is b. The diagonal entry of A2 corresponding to
a vertex v is the number of v-v walks of length 2, which is deg v. Therefore
G is b-regular, so one eigenvalue is b and the other is −1. The minimal
poynomial is thus (x − b)(x + 1), so A(A + I) = b(A + I). That is, every
column of A + I is an eigenvector of A with eigenvalue b, and since every
diagonal entry of A + I is 1, it follows from Theorem 4.3 that A + I = J, the
n × n matrix with all entries 1. That is, A = J − I, the adjacency matrix of
Kn .

Exercises for §4

4.1. Let G be a connected d-regular graph with adjacency matrix A. Show

that −d is an eigenvalue of A iff G is bipartite, in which case it has
multiplicity 1.

23
5 Blocks
Definition 5.1. A vertex v of a graph G is a cut-vertex if G is non-trivial
and k(G − v) > k(G). A graph is non-separable if it is connected and has
no cut-vertex; otherwise it is separable.

Note that an end-vertex is never a cut-vertex.

Theorem 5.2. A vertex u of a non-trivial graph G is a cut-vertex iff there

exist vertices v and w, different from u, that belong to the same component
of G, and for which every v-w path contains u.

Proof. Since every component of G − u is contained in a component of G, u

is a cut-vertex iff some component of G contains more than one component
of G − u. Since vertices v and w of G − u lie in different components of G − u
iff every v-w path in G contains u, the result follows.

Theorem 5.3. Let G be a non-trivial connected graph. No peripheral vertex

of G is a cut-vertex, and in particular G has at least two vertices that are
not cut-vertices.

Proof. Let u be a peripheral vertex of G, and v a vertex with d(u, v) =

diam G. Suppose that u is a cut-vertex, and let w be a vertex of a component
of G − u not containing v. Then every v-w path, and in particular every v-w
geodesic, contains u, so d(v, w) > d(u, v) = diam G, a contradiction.

Definition 5.4. An edge e of a graph G is a bridge if k(G − e) > k(G).

The proof of the following theorem is similar to that of Theorem 5.2,

and is left as an exercise.

Theorem 5.5. An edge e of a graph G is a bridge iff there exist vertices

v and w that belong to the same component of G, and for which every v-w
path contains e.

It follows that if e is a bridge then k(G − e) = k(G) + 1 and the vertices

of e belong to different components of G − e. A vertex of a bridge is either a
cut-vertex or an end-vertex, and so the only connected graph with a bridge
but no cut-vertices is K2 .

Theorem 5.6. An edge e of a graph G is a bridge iff it lies on no cycle in

G.

24
Proof. Let e = uv. Suppose that e lies on a cycle C. For any two vertices
w and x in a single component of G, take a w-x path in G. If it contains
the edge e, replace e by the other u-v or v-u path in C to get a w-x path in
G − e. Hence e is not a bridge.
Now suppose e is not a bridge. Then there is a u-v path P in G − e, and
P · (v, u) is a cycle in G containing e.

Definition 5.7. A graph is acyclic if it contains no cycles. A connected

acyclic graph is a tree, and an acyclic graph (whose components are trees)
is also called a forest.

In a forest, every edge is a bridge, so every vertex is either a cut-vertex or

an end-vertex. A non-separable graph of order at least three has no bridges,
so it is not acyclic. The following result is an immediate consequence of
Theorem 5.3.

Theorem 5.8. A non-trivial tree has at least two end-vertices.

Theorem 5.9. Let G be a graph of order n ≥ 3. Then the following are

equivalent.

(1) G has no isolated vertices and any two edges of G lie on a common
cycle.

(2) G is non-empty and any vertex and edge of G lie on a common cycle.

(3) Any two vertics of G lie on a common cycle.

(4) G is non-separable.

Proof. The implications (1) ⇒ (2) ⇒ (3) are trivial. Suppose that (3) holds.
Certainly G is connected. Let u be any vertex of G, and v and w vertices
distinct from u. There is a cycle C containing v and w, and at least one
of the two v-w paths in C does not contain u, so u is not a cut-vertex by
Theorem 5.2. Thus G is non-separable.
Suppose that G is non-separable, and let e1 and e2 be edges of G. Let
k be the minimum distance from a vertex of e1 to a vertex of e2 . We prove
by induction on k that there is a cycle containing e1 and e2 . Let e1 = u1 v1
and e2 = u2 v2 , where d(u1 , u2 ) = k. We show first that there is a cycle C
containing e1 and u2 . If k = 0, this is because G has no bridges, so there
is a cycle containing e1 . If k > 0, there is a vertex w adjacent to u2 with
d(u1 , w) = k − 1, and by inductive hypothesis there is a cycle containing e1
and the edge u2 w.

25
 If e2 is on C, we are done, so suppose it is not. Since u2 is not a cut-
vertex, there is a path P in G − u2 from v2 to a vertex x of C. Taking P
to have minimum length, x is the only vertex of P on C. Let Q be the x-u2
path in C containing e1 . Then P · Q · (u2 , v2 ) is the desired cycle.

Definition 5.10. Paths P and Q in a graph G are internally disjoint if no

internal vertex of either one is on the other.

The next result is just a reformulation of the equivalence (3) ⇔ (4) of

the previous one, and is left as an exercise.

Corollary 5.11. A graph G of order n ≥ 3 is non-separable iff, for u 6= v ∈

V (G), there are at least two internally disjoint u-v paths.

Definition 5.12. A block of a graph G is a maximal non-separable subgraph

of G. (That is, it is a non-separable subgraph H such that H ⊆ H ′ ⊆ G
and H ′ non-separable imply that H = H ′ .)

Every block of G is contained in a component of G, and every edge is

contained in a block, so the only trivial blocks of G are isolated vertices.
The subgraph induced by a bridge is a block of G, called an acyclic block of
G. If G is non-separable, it is its own unique block.

Lemma 5.13. Let G be a connected graph.

(1) Two distinct blocks of G meet in at most a single vertex, which must
be a cut-vertex of G.

(2) Every cut-vertex of G is in at least two blocks.

(3) If v and w are vertices such that no block of G contains both v and w,
there is a cut-vertex u such that v and w are in different components
of G − u.

Proof. The proof is by induction on the order of G. If G has no cut-vertices,

(1)–(3) are vacuously true, so suppose G has a cut-vertex u. Let the com-
ponents of G − u be H1 , . . . , Hk (k ≥ 2), and set Gi = hV (Hi ) ∪ {u}i for
1 ≤ i ≤ k. Each Gi is a connected graph of smaller order than G, so by
induction the result holds for Gi . If v and w belong to different components
of G − u then any connected subgraph of G containing v and w contains
u as a cut-vertex, so there is no block of G containing v and w. That is,
every block of G is contained in some Gi , and so the blocks of G are just
the blocks of all the Gi ’s. Also, a vertex of Gi is a cut-vertex of Gi iff it is a

26
cut-vertex of G different from u. Consider two distinct blocks of G. If they
lie in the same Gi , they meet in at most a single vertex which is a cut-vertex
of Gi , and if not they meet in at most u. Every cut-vertex of G other than
u is in at least two blocks of some Gi , and u is in k ≥ 2 blocks of G, one
from each Gi . Finally, suppose v and w are vertices contained in no block
of G. If they lie in a single Gi , there is a cut-vertex u′ of Gi such that v and
w lie in different components of Gi − u′ , and hence of G − u′ . Otherwise, v
and w lie in different components of G − u.

Theorem 5.14 (Harary and Norman [16]). Let G be a connected graph.

Then Cen G is contained in a single block of G.

Proof. Let u and v be vertices such that no block of G contains both u and
v, and let w be the cut-vertex provided by Lemma 5.13. Let x be a vertex
with d(w, x) = e(w). At least one of u and v is in a different component of
G − w from x; say u is. Then e(u) ≥ d(u, x) = d(u, w) + d(w, x) > e(w), so
u is not a central vertex. The result follows.

Corollary 5.15. Let T be a tree. Then Cen T is isomorphic to K1 or K2 .

Proof. If T is non-trivial, every block of T is isomorphic to K2 , and the only

induced subgraphs of K2 are K1 and K2 .

If a connected graph G has a cut-vertex, every block of G contains a

cut-vertex. A block of G containing exactly one cut-vertex is called an
end-block.

Definition 5.16. The block graph Blk G of a graph G is a bipartite graph

with partite sets V1 and V2 , where V1 is the set of all blocks of G, and V2
is the set of all cut-vertices. A block B and a cut-vertex v are adjacent in
Blk G iff B contains v.

The block graph of G is the disjoint union of the block graphs of its
components, so it is enough to consider Blk G for G connected. By (2) of
Lemma 5.13, the end-vertices of Blk G are precisely the end-blocks of G.

Theorem 5.17. The block graph of a connected graph G is a tree.

Proof. If G is non-separable then Blk G is trivial, so suppose G has a cut-

vertex. Since every block contains a cut-vertex, to show that Blk G is con-
nected it is enough to show that any cut-vertices u and v of G are connected
in Blk G. Let (u0 , u1 , . . . , ur ) be a u-v path in G, and let the cut-vertices
on this path be ui0 , ui1 , . . . , uis where 0 = i0 < i1 < · · · < is = r. Since any

27
edge or non-cut-vertex of G is contained in a unique block, each sub-path
from uij−1 to uij is contained in a block Bj . Now
(u, B1 , ui1 , B2 , . . . , uis−1 , Bs , v)
is a u-v path in Blk G.
It remains to prove that Blk G is acyclic. If not, it contains a cycle
(B0 , u1 , B1 , . . . , uk , Bk = B0 )
where each Bi is a block and each ui is a cut-vertex. Let H = B1 ∪ · · · ∪ Bk ,
which is clearly connected. Let v be a vertex of H, and let Bi′ be Bi − v if
v is in Bi , and Bi otherwise. Each Bi′ is connected, and if v 6= ui then Bi−1 ′
′
and Bi have a common vertex, and hence lie in a single component of H − v.
Since v is equal to at most one ui , H − v = B1′ ∪ · · · ∪ Bk′ is connected, so H
is non-separable. This contradicts the maximality of the blocks Bi .

The next result is immediate from the previous one and Theorem 5.8.
Theorem 5.18. A connected graph with a cut-vertex contains at least two
end-blocks.
Theorem 5.19. Let G be a connected graph with a cut-vertex. Then there
is a cut-vertex u such that at most one block of G containing u is not an
end-block.
Proof. Let T be the result of deleting from Blk G all its end-vertices. Then T
is also a tree. If T is trivial, then G has a unique cut-vertex u and all blocks
of G are end-blocks, so u has the desired property. Otherwise, T has an
end-vertex, which must be a cut-vertex u of G since blocks are not adjacent
in Blk G. Now every block of G containing u is an end-block, except for the
unique vertex of T adjacent to u.

The next lemma will be used in §17.

Lemma 5.20. Let G be a non-trivial connected graph and U a set of non-
cut-vertices of G containing at most one vertex from each block. Then G−U
is connected.
Proof. This is by induction on the number of blocks, the case of a single
block being trivial. Suppose G has more than one block, let G1 be an end-
block, and let G2 be the union of the other blocks. Then G1 and G2 meet in
a single cut-vertex u of G. For i = 1 or 2, let Ui = U ∩ V (Gi ). By induction,
G1 − U1 and G2 − U2 are connected, and since they have u in common, so
is their union G − U .

28
Exercises for §5

5.1. Prove Theorem 5.5.

5.2. Prove Corollary 5.11.

5.3. Prove that a graph with only even vertices has no bridge.

5.4. Let T be a tree of order n ≥ 3, and let T ′ be the tree obtained by

deleting all end-vertices of T . Prove that for every vertex u of T ′ ,
eT ′ (u) = eT (u) − 1. Use this to give an alternative proof of Theorem
5.14.

6 Connectivity
Definition 6.1. A vertex-cut of a graph G is a proper subset U of V (G)
such that G − U is disconnected.

If G is not complete, let u and v be independent vertices. Then V (G) −

{u, v} is a vertex-cut. The connectivity of G, κ(G), is defined to be the
minimum number of vertices in a vertex-cut. If G has order n, we have 0 ≤
κ(G) ≤ n−2. A complete graph has no vertex-cut; we define κ(Kn ) = n−1.
In all cases, κ(G) is the minimum cardinality of a set U of vertices such that
G − U is disconnected or trivial. We have κ(G) = 0 iff G is disconnected or
trivial. Also κ(G) = 1 iff G is connected and non-trivial, and either G ∼= K2
or G has a cut-vertex. Thus κ(G) ≥ 2 iff G is non-separable of order at least
3.

Definition 6.2. An edge-cut of a graph G is a subset X of E(G) such that

G − X is disconnected.

If G is non-trivial, G has an edge-cut (for instance, E(G)), and the

edge-connectivity κ1 (G) is the minimum cardinality of an edge-cut. We
also set κ1 (K1 ) = 0. We have κ1 (G) = 0 iff G is disconnected or trivial and
κ1 (G) = 1 iff G is connected, non-trivial, and has a bridge. If X is a minimal
edge cut of a connected graph G then G − X has exactly two components,
and each edge of X has one vertex in each.

Theorem 6.3 (Whitney [31]). For any graph G, κ(G) ≤ κ1 (G) ≤ δ(G).

Proof. We may assume that G is connected and non-trivial. Suppose that X

is an edge-cut with |X| = κ1 (G), and let the components of G−X be G1 and
G2 . Suppose first that every vertex of G1 is adjacent to every vertex of G2 .

29
Let the orders of G, G1 and G2 be n, n1 and n2 . Then (n1 − 1)(n2 − 1) ≥ 0,
so |X| = n1 n2 ≥ n1 + n2 − 1 = n − 1 ≥ κ(G). Otherwise, let u ∈ V (G1 ) and
v ∈ V (G2 ) be non-adjacent. For every edge e of X, pick a vertex f (e) of e
different from u and v, and set U = {f (e) | e ∈ X}, so |U | ≤ |X|. Then U
is a vertex-cut, because u and v are vertices of G − U , and every u-v path
in G must contain an edge e of X, and therefore the vertex f (e) of U . We
have proved that κ(G) ≤ κ1 (G) in all cases. The set of edges incident with
any vertex is an edge-cut, which gives κ1 (G) ≤ δ(G).

Theorem 6.4. Let G be a graph of diameter 2. Then κ1 (G) = δ(G).

Proof. Let k = κ1 (G), and let X be an edge-cut with |X| = k. The graph
G − X has two components H1 and H2 . We show first that either every
vertex of H1 is adjacent to a vertex of H2 , or every vertex of H2 is adjacent
to a vertex of H1 . Suppose there is a vertex u of H1 adjacent to no vertex
of H2 . If v is any vertex of H2 , d(u, v) = 2 and there is a path (u, w, v).
Since w is adjacent to u, w is in H1 . Thus every vertex of H2 is adjacent to
a vertex of H1 .
Let n1 and n2 be the orders of H1 and H2 , and choose the numbering so
that n1 ≤ n2 . In the first case above k ≥ n1 and in the second k ≥ n2 , so in
any case k ≥ n1 . For a vertex u of H1 , let d1 (u) and d2 (u) be the numbers
of vertices of H1 and H2 adjacent to u. Then

n1 ≤ k ≤ δ(G) ≤ deg u = d1 (u) + d2 (u) ≤ n1 − 1 + d2 (u).

Hence d2 (u) ≥ 1; that is, every vertex of H1 is adjacent to a vertex of H2 .

For any vertex u of H1 ,
X
k= d2 (v) ≥ d2 (u) + n1 − 1 ≥ δ(G) ≥ k,
v∈V (G1 )

so k = δ(G).

Let u and v be vertices of a graph G. We say that a set W of vertices

of G separates u and v if u and v belong to different components of G − W ;
such a set is of course a vertex-cut.

Theorem 6.5 (Menger [24]). Let u and v be independent vertices of a graph

G. Then the minimum number of vertices that separate u and v is equal to
the maximum number of internally disjoint u-v paths.

30
Proof. For independent vertices u and v of a graph G let sG (u, v) be the
minimum number of vertices that separate u and v (which exists because the
set of all other vertices separates), and let pG (u, v) be the maximum number
of internally disjoint u-v paths. If P is any such set of paths and W is a set
of vertices separating u and v, W must contain at least one internal vertex of
every path in P , so |P | ≤ |W | and hence pG (u, v) ≤ sG (u, v). If pG (u, v) = 0,
then u and v are in different components of G, and so sG (u, v) = 0. It
follows that if sG (u, v) = 1 then pG (u, v) = 1, so if pG (u, v) < sG (u, v) then
sG (u, v) ≥ 2. Suppose there exist independent vertices u and v of a graph G
with pG (u, v) < sG (u, v). Let t be the smallest integer for which there exist
such u, v and G with sG (u, v) = t, and amongst such u, v and G let G have
minimum size. We make the following claims about G.

(1) There is no vertex adjacent to both u and v.

(2) If w1 and w2 are adjacent vertices different from u and v, there is a set
W of t − 1 vertices such that W ∪ {w1 } and W ∪ {w2 } both separate
u and v.

(3) If W is a set of t vertices separating u and v, either every vertex of

W is adjacent to u, or every vertex of W is adjacent to v.

For (1), suppose that w is adjacent to u and v, and let H = G − w. A set

W of vertices separates u and v in H iff W ∪ {w} separates them in G, so
sH (u, v) = t or t − 1. If sH (u, v) = t, then since H has smaller size than G,
pH (u, v) = t. But t internally disjoint u-v paths in H are also paths in G, a
contradiction. If sH (u, v) = t − 1, then pH (u, v) = t − 1. However, adding to
t − 1 internally disjoint u-v paths in H the path (u, w, v) gives t such paths
in G, again a contradiction.
As for (2), let H be obtained by deleting the edge w1 w2 of G. If W is
a set of vertices separating u and v in H, then W ∪ {wi } separates them
in G for i = 1 or 2, so sH (u, v) ≥ t − 1. On the other hand, any set of
vertices separating u and v in G also separates in H, so sH (u, v) ≤ t. If
sH (u, v) = t, then since H has smaller size than G, pH (u, v) = t. As before,
t internally disjoint u-v paths in H are also paths in G, a contradiction.
Hence sH (u, v) = t − 1, and a set of t − 1 vertices separating u and v in H
is the set required by the claim.
Finally, suppose, contrary to (3), that there is a set W of t vertices
separating u and v in G that contains a vertex not adjacent to u, and one
not adjacent to v. Let Ku be the component of G − W containing u, and
Hu the subgraph of G induced by all edges with at least one vertex in Ku .

31
By minimality of W , every vertex of W is in Ku , and since W has a vertex
not adjacent to u, Hu has at least t + 1 edges. Define Kv and Hv similarly.
Let Gu be the graph obtained from Hu by adding the vertex v and edges
vw for all w in W . Since Hv has at least t + 1 edges, Gu has smaller size
than G. Clearly W separates u from v in Gu ; let W ′ be any set of vertices
separating u from v in Gu . Every u-v path P in G gives a u-v path P ′
in Gu , namely the initial subpath in P from u to some w ∈ W , followed
by (w, v). Since P ′ contains a vertex of W ′ , so does P . It follows that
sGu (u, v) = t, and so pGu (u, v) = t. Now t internally disjoint u-v paths in
Gu give internally disjoint paths in Hu from u to each vertex of W . Similarly,
there are internally disjoint paths in Hv from each vertex of W to v. Putting
these together with the paths in Hu gives t internally disjoint u-v paths in
G, a contradiction.
At this point, (1)–(3) have been established, and we now derive a final
contradiction. Let (w0 , w1 , . . . , wk ) be a u-v geodesic in G. By (1), k ≥ 3.
By (2), there is a set W of t − 1 vertices such that W ∪ {w1 } and W ∪ {w2 }
both separate u and v in G. Since w1 is not adjacent to v, (3) implies
that every vertex of W is adjacent to v, and since w2 is not adjacent to u, it
implies that every vertex of W is adjacent to v. Since t ≥ 2, W is non-empty
and we have a contradiction to (1).

Definition 6.6. Let k be a positive integer. A graph G is k-connected if

κ(G) ≥ k, and edge k-connected if κ1 (G) ≥ k.
G is 2-connected iff G is non-separable of order at least 3, so Corollary
5.11 can be restated as saying that a non-trivial graph is 2-connected iff
every two vertices are connected by two internally disjoint paths. This is
generalized in Theorem 6.8 below.
Lemma 6.7. Let e be an edge of a k-connected graph G, where k ≥ 2. Then
G − e is (k − 1)-connected.
Proof. Let e = uv, and let W be a set of vertices such that G − e − W is
disconnected or trivial. If u or v is in W , G − e − W = G − W , so |W | ≥ k.
Otherwise, u and v are vertices of G−e−W , which is therefore disconnected.
If G − u − W is disconnected, we have |W | ≥ k − 1. Otherwise, G − e − W
has u as an isolated vertex, and just one other component. Hence G − v − W
is trivial or disconnected, so again |W | ≥ k − 1.

Theorem 6.8 (Whitney [31]). Let G be a non-trivial graph. Then G is

k-connected (k ≥ 1) iff for all u 6= v ∈ V (G) there are k internally disjoint
u-v paths.

32
Proof. Suppose first that such paths exist. Take distinct vertices of G. Of
a set of k internally disjoint paths connecting them, at most one has no
internal vertices, so G has order at least k + 1. Thus it suffices to show that
any vertex-cut U has at least k elements. Let P be a set of k internally
disjoint paths joining vertices of different components of G − U . Then U
contains an internal vertex of every path in P , so indeed |U | ≥ k.
Suppose conversely that G is k-connected, and let u 6= v ∈ V (G). If
k = 1 the result is trivial, so suppose k ≥ 2. If u and v are not adjacent,
any set of vertices separating u and v has at least k elements, so Menger’s
Theorem gives at least k internally disjoint u-v paths. Otherwise, G − uv is
(k − 1)-connected by Lemma 6.7, so Menger’s Theorem gives at least k − 1
internally disjoint u-v paths in G − uv, and adding the path (u, v) gives at
least k internally disjoint u-v paths in G.

Lemma 6.9. Let u1 , . . . , uk be distinct vertices of a k-connected graph G,

and form H by adding a new vertex v and edges ui v for 1 ≤ i ≤ k. Then H
is k-connected.

Proof. Let U be a set of fewer than k vertices of H. If v ∈/ U then G − U

is connected and contains at least one ui , so H − U is connected and non-
trivial. If v ∈ U then H − U = G − (U − {v}), which is connected and
non-trivial.

Theorem 6.10. Let u, v1 , . . . , vk be distinct vertices of a k-connected graph

G. Then there are internally disjoint u-vi paths for 1 ≤ i ≤ k.

Proof. Form H by adding a new vertex w and edges vi w for 1 ≤ i ≤ k. By

Lemma 6.9, H is k-connected, so there are k internally disjoint u-w paths
in H by Theorem 6.8. Deleting their last edges gives the required paths in
G.

Theorem 6.11. Let G be k-connected, where k ≥ 2. Then any k vertices

of G lie on a common cycle.

Proof. Let U be a set of k vertices of G, and let C be a cycle containing

the maximum number, say l, of vertices of U . Certainly l > 0. (In fact,
Theorem 5.9 implies that l ≥ 2, but we do not need this.) We suppose,
for a contradiction, that l < k. Let u be a vertex of U not on C. Suppose
first that C has length l. By Theorem 6.10, there are internally disjoint
paths from u to every vertex of C. Replacing an edge of C by two of these
paths gives a cycle containing at least l + 1 vertices of U , a contradiction.
Otherwise, let V be a set of l + 1 vertices of C. Since l + 1 ≤ k, Theorem

33
6.10 gives internally disjoint paths from u to every vertex of V . Replacing
these paths by their shortest initial segments ending on C, and V by the
set of terminal points of the new paths, we may assume that each path in
the set meets C only in its terminal point. Now there are vertices v1 and
v2 of V such that one of the v1 -v2 paths in C has no element of V or U as
an internal vertex. Replacing this path by a path through u gives a cycle
containing at least l + 1 vertices of U , completing the proof.

Theorem 6.12. Let G be a graph of order n ≥ 2, and let k be an integer

with 1 ≤ k ≤ n − 1. If δ(G) ≥ 12 (n + k − 2) then G is k-connected.

Proof. Suppose that G is not k-connected. Then G has a vertex-cut U with

|U | = l ≤ k − 1. Let G1 be a component of G − U of minimum order n1 .
Then n1 ≤ (n − l)/2. If v is a vertex of G1 then v is adjacent only to vertices
of U and other vertices of G1 , so

δ(G) ≤ deg v ≤ l + n1 − 1 ≤ (n + l − 2)/2 ≤ (n + k − 3)/2.

Exercises for §6

6.1. (a) Let G1 and G2 be disjoint copies of Kn , n ≥ 2, and form a graph

G by adding to G1 ∪ G2 k edges, 0 < k < n, which must have one
vertex in G1 and one in G2 . Suppose that n1 vertices of G1 and n2
of G2 are incident with one or more of these k edges. Determine
κ(G), κ1 (G) and δ(G).
(b) Show that for any positive integers a, b and c with a ≤ b ≤ c, there
is a graph G with κ(G) = a, κ1 (G) = b and δ(G) = c.

6.2. Determine the connectivity and edge-connectivity of any complete mul-

tipartite graph.

7 Trees
Recall that a spanning subgraph of a graph G is a subgraph containing all
the vertices of G.

Definition 7.1. A spanning tree of a graph is a spanning subgraph that is

a tree, and a spanning forest is a spanning subgraph that is a forest.

We shall be concerned with spanning trees more than spanning forests.

Evidently a graph with a spanning tree is connected.

34
Theorem 7.2. Any connected graph G has a spanning tree.

Proof. The proof is by induction on the size m of G. If G is itself a tree,

there is nothing to do. Otherwise, G has an edge e that is not a bridge, and
then G − e is a connected spanning subgraph of G of size m − 1, which has
a spanning tree T by induction. Now T is also a spanning tree of G.

We have seen that any non-trivial tree T has end-vertices. Deleting an

end-vertex from T leaves a tree. Conversely, adding to a tree a single vertex
v and an edge with v as a vertex yields a tree (with v as an end-vertex).

Theorem 7.3. For a graph G of order n and size m, any two of the following
conditions imply the third.

(1) G is connected.

(2) G is acyclic.

(3) m = n − 1.

Thus to show that a graph is a tree, it is enough to verify any two of

these conditions.

Proof. (a) We show that (1) and (2) imply (3) by induction on n. If n = 1
then m = 0 and (3) holds. If n > 1 then as remarked above we may delete an
end-vertex, leaving a tree of order n − 1 and size m − 1, giving the inductive
step.
(b) Suppose that (2) and (3) hold and G has k components. Applying
part (a) to each component and adding, we have m = n − k, so k = 1, as
required.
(c) Suppose that (1) and (3) hold. Then G has a spanning tree T . By
part (a), T has size m, so T = G, giving (2).

Corollary 7.4. Let G be an acyclic graph of order n and size m with k

components. Then m = n − k.

Lemma 7.5. Let u and v be vertices of a graph G, and suppose there are
distinct u-v paths P and Q of lengths k and l. Then there is a cycle of length
at most k + l in G, and if P · Qr is not a cycle, there is one of length less
than k + l.

Proof. Let P = (u0 , u1 , . . . , uk ). There is some i, 0 ≤ i < k, for which ui is

on Q and the edge ui ui+1 is not. Let j be the smallest integer greater than
i for which uj is on Q. Let P1 and Q1 be the ui -uj subpaths of P and Q or

35
Qr , respectively. Then P1 · Qr1 is a cycle in G of length at most k + l, with
equality only if P = P1 and Q = Q1 .

Theorem 7.6. A graph G is a tree iff there is a unique u-v path for all u,
v ∈ V (G).
Proof. Suppose that any two vertices of G are connected by a unique path.
Certainly G is connected. If C is a cycle in G and u and v are any two
distinct vertices of C, there are two distinct u-v paths in C, a contradiction.
Thus G is a tree. The converse is immediate from the preceding lemma.

Theorem 7.7. Let d = (d1 , . . . , dn ) be a sequence Pof positive integers, where

n
n ≥ 2. Then d is a degree sequence of a tree iff i=1 di = 2n − 2.
Proof. If d is a degree sequence of a tree of size m then ni=1 di = 2m =
P
2n − 2. For the converse, if n = 2 then d = (1, 1), the degree sequence of
the tree K2 . Suppose then that n > 2 and the result holds for sequences of
length
Pn n − 1. We may assume d1 ≥ · · · ≥ dn . Then d1 P > 1, since otherwise
n
d
i=1 i = n < 2n − 2, and d n = 1, since otherwise i=1 di ≥ 2n. The
sequence (d1 − 1, d2 , . . . , dn−1 ) of positive integers has sum 2n − 4, so it is a
degree sequence of a tree T of order n − 1. Adding to T a vertex u and an
edge uv where degT v = d1 − 1 gives a tree with degree sequence d.

Theorem 7.8. Let T be a tree of order n. If G is a graph with δ(G) ≥ n − 1

then G has a subgraph isomorphic to T .
Proof. This is by induction on n, the case n = 1 being trivial. Suppose that
n > 1, and let u be an end-vertex of T and v the vertex of T adjacent to
u. By inductive hypothesis, there is an isomorphism φ from T ′ = T − u
to a subgraph H of G. Since degT v ≤ n − 1, degT ′ v ≤ n − 2. Since
degG φ(v) ≥ n − 1, there is a vertex u′ of G adjacent to φ(v) and not in H.
Then φ may be extended to an isomorphism from T to a subgraph of G by
sending u to u′ .

8 Counting trees
There is no known formula for the number of trees of order n (up to iso-
morphism). For small values of n one can enumerate the trees of order n by
considering all ways of adding a vertex and edge to a tree of order n − 1.
The unique trees of orders 1, 2 and 3 are K1 , K2 and P3 . For n = 4, there
are two trees, the path P4 and the star K1,3 . Of course, any path or star
is a tree. Other special classes of trees are the double stars, which have

36
 Figure 7: The trees of order 5

exactly two (necessarily adjacent) vertices of degree greater than 1, and the
caterpillars, the trees for which deleting the end-vertices leaves a path (the
spine of the caterpillar). There are three trees of order 5, shown in Figure 7,
a path, a star, and a double star. There are 6 of order 6, whose enumeration
is left as an exercise.
For the number of trees on a given set of size n (up to equality), some-
times known as the number of labelled trees of order n, there is the following
theorem of Cayley [6]. The proof below is due to Prüfer [26].

Theorem 8.1. Let X = {1, . . . , n}. The number of trees with vertex set X
is nn−2 .

Proof. The cases n = 1 and 2 are obvious, so assume n ≥ 3. Let T be a tree

with V (T ) = X. We define trees Ti for 0 ≤ i ≤ n − 2 and integers si ∈ X
for 1 ≤ i ≤ n − 2 such that V (Ti ) ⊆ X and |V (Ti )| = n − i as follows. We
set T0 = T . Suppose that 1 ≤ i ≤ n − 2 and Ti−1 has been constructed. Let
vi be the least end-vertex of Ti−1 , let si be the vertex of Ti−1 adjacent to
vi and set Ti = Ti−1 − vi . The sequence (s1 , . . . , sn−2 ) so constructed is the
Prüfer sequence of T . For later use, we show that, for 0 ≤ i ≤ n − 2,

a vertex of Ti is an end-vertex
(8.1)
⇔ it is not in the set {si+1 , . . . , sn−2 }.

For i = n − 2 this is clear: Tn−2 has two vertices, which are end-vertices,
and the set in question is empty. Suppose that (8.1) holds for some i with
1 ≤ i ≤ n − 2. Ti−1 has one more vertex than Ti , namely vi , which is an
end-vertex of Ti−1 and is not in the set {si , . . . , sn−2 }. Further, a vertex
of Ti is an end-vertex of Ti−1 iff it is an end vertex of Ti and not equal to
si , which is the case iff it is not in {si , . . . , sn−2 }. This establishes (8.1) for
i − 1.
Now let (s1 , . . . , sn−2 ) be any sequence of elements of X. We construct
graphs Gi and sets Xi ⊆ X for 0 ≤ i ≤ n − 2 such that V (Gi ) = X,
|E(Gi )| = i, |Xi | = n − i, {si+1 , . . . , sn−2 } ⊆ Xi , and each component of

37
Gi has exactly one vertex in Xi . We take G0 to be the empty graph on
X and X0 = X. Suppose that 1 ≤ i ≤ n − 2 and Gi−1 and Xi−1 have
been constructed. Then at least two elements of Xi−1 are not in the set
{si , . . . , sn−2 }; let vi be the least such element. Then vi and si are in distinct
components of Gi−1 . Setting Gi = Gi−1 + vi si and Xi = Xi−1 − {vi }, Gi
and Xi have the desired properties. The set Xn−2 has two elements vn−1
and vn , and the graph Gn−2 has n − 2 edges and two components, one
containing vn−1 and the other containing vn . Hence T = Gn−2 + vn−1 vn
is a tree. For 0 ≤ i ≤ n − 2, let Ti be the subgraph of T induced by the
vertices Xi . We show that Ti is a tree whose end vertices are the elements
of Xi − {si+1 , . . . , sn−2 }. For i = n − 2, Xn−2 = {vn−1 , vn } and vn−1 vn is
an edge of T , so the result holds in this case. Suppose it holds for some i,
1 ≤ i ≤ n−2. Then Xi−1 = Xi ∪{vi }, si ∈ Xi , and vi si is an edge of Gi , and
hence of T . Thus Ti−1 = Ti + vi si , which is a tree, and its set of end-vertices
is obtained from that of Ti by deleting si (if present) and adding vi . This
proves the result for i − 1. Note also that vi is the least end vertex of Ti−1 .
It follows that the trees Ti are also the trees used in constructing the Prüfer
sequence of T , and that this sequence is equal to (s1 , . . . , sn−2 ).
Now, starting with a tree T , construct the trees T0 , . . . , Tn−2 and the
Prüfer sequence (s1 , . . . , sn−2 ) as in the first paragraph, and then the graphs
G0 , . . . , Gn−2 and the sets X0 , . . . , Xn−2 from this sequence. We show that,
for 0 ≤ i ≤ n − 2,

(8.2) Xi = V (Ti ) and E(Gi ) = E(T ) − E(Ti ),

the case i = 0 being obvious. Suppose that 1 ≤ i ≤ n − 2, and (8.2)

holds for i − 1. Let vi be the least end-vertex of Ti−1 , so that si is the
vertex of Ti−1 adjacent to vi and Ti = Ti−1 − vi . By (8.1), vi is also the
least element of V (Ti−1 ) = Xi−1 not in {si , . . . , sn−2 }, so Gi = Gi−1 + vi si
and Xi = Xi−1 − {vi }, proving (8.2) for i. If Xn−2 = {vn−1 , vn }, the case
i = n − 2 of (8.2) shows that Gn−2 + vn−1 vn = T . Thus sending a tree on X
to its Prüfer sequence gives a bijection from the set of such trees to X n−2 ,
proving the theorem.

The following result of Kirchoff [21] is known as the Matrix-Tree Theo-

rem.
Theorem 8.2. Let G be a graph. The number of spanning trees of G is
equal to any cofactor of the Lagrangian matrix L(G).
The number of trees on a given set is the same as the number of span-
ning trees in the complete graph on that set, so consistency with Cayley’s

38
Theorem requires that any cofactor of L(Kn ) is equal to nn−2 ; this is left as
an exercise.
Proof. The sum of each row or column of L(G) is zero, from which it follows
easily that all cofactors of L(G) are equal. For any non-empty proper subset
U of V = V (G), let LU (G) be the matrix obtained from L(G) by deleting
the rows and columns corresponding to elements of U . Define a U -forest in
G to be a spanning forest in which every component contains exactly one
element of U . When U has a single element, a U -forest is just a spanning
tree, so the result will follow if we show that the number of U -forests is
equal to det LU (G). This we do by induction on |V − U |. Suppose this is
1, and let v be the element of V − U . Then det LU (G) = deg v, while the
U -forests are the spanning subgraphs with just one edge, incident to v. The
result holds in this case. Now suppose that |V − U | > 1, and the result holds
for larger subsets of V . Here we use induction on the number k of edges
with one vertex in U and one in V − U . If k = 0, there are no U -forests,
while LU (G) is just the Lagrangian of the subgraph induced by V − U , and
so has determinant zero. Suppose then that k > 0, and let e = uv be an
edge with u ∈ U and v ∈ V − U . Set H = G − e. The matrix LU (G)
is obtained from LU (H) by adding 1 to the diagonal entry corresponding
to v, so det LU (G) = det LU (H) + det LU ′ (H), where U ′ = U ∪ {v}. By
the induction on k, det LU (H) is the number of U -forests in H, and by the
induction on |V − U |, det LU ′ (H) is the number of U ′ -forests in H. The
U -forests in H are precisely the U -forests in G that do not contain e. Also,
there is a one-to-one correspondence between the U ′ -forests in H and the
U -forests in G that do contain e, given by adding e to a U ′ -forest in H. The
result follows.
Exercises for §8
8.1. Draw all trees of order 6.
8.2. Let A be a square matrix in which every row and column sums to zero.
Show that all cofactors of A are equal.
8.3. Show that any cofactor of L(Kn ) is equal to nn−2 .

9 Permutation groups
In what follows, we need some knowledge of groups. However, the general
theory is not really needed, and a few easy facts about permutation groups
will suffice.

39
Definition 9.1. Let X be a finite set. A permutation of X is a bijection
π : X → X. The set of all permutations of X will be denoted by S(X).
When X = {1, . . . , n}, we write Sn for S(X), and we will often identify
S(X) with Sn when X has n elements.

For any set X we have the identity permutation id = idX ∈ S(X), and
for any π ∈ S(X) the inverse π −1 is in S(X). Finally, for any π and σ in
S(X), the composite πσ is in S(X). We have:

(1) id π = π = π id for all π ∈ S(X);

(2) ππ −1 = id = π −1 π for all π ∈ S(X);

(3) π(στ ) = (πσ)τ for all π, σ, τ ∈ S(X).

It follows easily that πσ = πτ implies σ = τ , as does σπ = τ π. The set

S(X) together with the operation of composition is called the symmetric
group of X. Clearly the number of elements (or order ) of Sn is n!.

Definition 9.2. Let x1 , . . . , xk be distinct elements of X. The permutation

k-cycle (x1 . . . xk ) is the permutation π of X defined by π(xi ) = xi+1 for
1 ≤ i < k, π(xk ) = x1 , and π(x) = x if x ∈ / {x1 , . . . , xk }. Permutation cycles
(x1 . . . xk ) and (y1 . . . yl ) are disjoint if xi 6= yj for all i and j.

Of course, the same cycle may be written in different ways: (1 2 3) =

(2 3 1) = (3 1 2). Any 1-cycle (x) is equal to the identity id. If π = (x1 . . . xk )
and σ = (y1 . . . yl ) are disjoint then π and σ commute (πσ = σπ). It is easy
to see that any permutation can be written as a product of disjoint cycles,
and that this representation is unique except for the order of the factors and
the presence or absence of 1-cycles.

Example 9.3. S3 = {id, (1 2), (1 3), (2 3), (1 2 3), (3 2 1)}.

Definition 9.4. A permutation group on the finite set X is a subset A of

S(X) such that:

(1) id ∈ A;

(2) if α ∈ A then α−1 ∈ A;

(3) if α ∈ A and β ∈ A then αβ ∈ A.

We also call A a subgroup of S(X).

40
Example 9.5. Let G be a graph. An automorphism of G is an isomorphism
α : G → G. The set Aut(G) of all automorphisms of G is a permutation
group on V (G).

Example 9.6. It is well-known that any permutation π ∈ S(X) can be

written as a product of (not necessarily disjoint) 2-cycles, or transpositions,
and that if this is done in two ways, the numbers of factors are either both
even or both odd; π is called an even or odd permutation accordingly. The
sign sgn(π) of π is 1 if π is even and −1 if π is odd. Clearly sgn(id) = 1,
sgn(π −1 ) = sgn(π) and sgn(πσ) = sgn(π) sgn(σ), and it follows that the set
Alt(X) of even permutations is a subgroup of S(X), called the alternating
group on X. When X = {1, . . . , n}, we write An for Alt(X).

Proposition 9.7. A non-empty subset A of S(X) is a subgroup iff, for any

α and β ∈ A, αβ −1 ∈ A.

Proof. That any subgroup has this property follows from (2) and (3) in
Definition 9.4. Suppose, conversely, that A is a non-empty subset of S(X)
with the given property. Take some α ∈ A; then id = αα−1 ∈ A. Hence,
for α ∈ A, α−1 = id α−1 ∈ A. Finally, for α and β ∈ A, αβ = α(β −1 )−1 ∈
A.

Theorem 9.8. Let A be a subgroup of Sn . Then the order of A divides n!.

Proof. We define a relation ∼ on S(X) by π ∼ σ if πσ −1 ∈ A, and show that

this is an equivalence relation. For π ∈ S(X), π ∼ π since ππ −1 = id ∈ A. If
π ∼ σ then πσ −1 ∈ A, so σπ −1 = (πσ −1 )−1 ∈ A, and σ ∼ π. Finally, if π ∼ σ
and σ ∼ τ then πσ −1 ∈ A and στ −1 ∈ A, so στ −1 = (πσ −1 )(στ −1 ) ∈ A,
and π ∼ τ .
Next we show that σ is in the equivalence class of π iff σ = απ for some
α ∈ A. If σ = απ then σπ −1 = α ∈ A, so σ ∼ π, and conversely if σ ∼ π
then defining α = σπ −1 we have α ∈ A and σ = απ.
Since απ = βπ implies α = β, it follows that every equivalence class
has size |A|, and hence that the size n! of Sn is the product of |A| and the
number of equivalence classes.

Example 9.9. Consider the subgroup An of Sn , and suppose n > 1 so

that there exist odd permutations and An 6= Sn . Clearly πσ −1 is even if
and only if π and σ have the same parity, so there are just two equivalence
classes under the relation in the above proof, namely An and the set of odd
permutations. Hence An has order n!/2.

41
 Figure 8: Two pictures of M4

Definition 9.10. Let A and B be permutation groups (possibly on sets of

different sizes). An isomorphism from A to B is a bijection f : A → B such
that f (αβ) = f (α)f (β) for all α, β ∈ A. (It follows that f (id) = id and
f (α−1 ) = f (α)−1 .) If there exists an isomorphism from A to B, A and B
are isomorphic, written A ∼ = B. It is easy to see that isomorphism is an
equivalence relation on any set of permutation groups.

Example 9.11. If |X| = n then S(X) ∼

= Sn .

10 Automorphism groups of graphs

Example 10.1. For any graph G, Aut(Ḡ) = Aut(G).

Example 10.2. Aut Kn = Aut K̄n ∼ = Sn , of order n!. For any other graph
G of order n, | Aut G| is less than n!, and divides n! by Theorem 9.8.

Example 10.3. Aut Pn has order 2 for any n > 1.

Example 10.4. Aut Cn has order 2n for any n ≥ 3. If Cn is drawn as

a regular n-gon, the automorphisms are n rotations (counting the identity
as a rotation), and n reflections. Aut Cn is known as the dihedral group of
order 2n.

Example 10.5. The n-rung Möbius ladder Mn (n ≥ 2) is the graph ob-

tained from a cycle C2n (called the rim) by adding edges (called the rungs)
joining each pair of opposite vertices (i.e., vertices at distance n in C2n ).
The case n = 4 is shown in Figure 8. Clearly Aut Mn contains the dihedral
group Aut C2n of order 4n. For n = 2, M2 = K4 , so Aut M2 ∼ = S4 , of order
24. For n = 3, M3 = K3,3 , so Aut M3 has order 72. Let n ≥ 4. We shall
show that Aut Mn = Aut C2n . In Mn , there is a unique cycle containing no

42
 v1

1
3 4
v6
v5 5 2 v2
1
v10 v7
5 2

4 3
2 5
v9 v8
4 3
1
v4 v3

Figure 9: Labelling the Petersen graph

rungs, namely the rim, of length 2n ≥ 8, and any cycle containing a single
rung has length n + 1 ≥ 5. Therefore a 4-cycle must contain at least two
rungs, and hence consists of a pair of opposite edges of the rim together with
the two adjacent rungs. This shows that each edge of the rim lies on just
one 4-cycle, while each rung lies on two. It follows that any automorphism
of Mn takes rim edges to rim edges (and rungs to rungs), and therefore lies
in Aut C2n .

Theorem 10.6. Let G be the Petersen graph. Then Aut(G) ∼

= S5 .
Proof. Label the edges of G with the numbers 1, . . . , 5 as in Figure 9. You
can check that two distinct edges e and f get different labels iff they are
adjacent, or there is a third edge adjacent to both. Hence, if α ∈ Aut(G) and
e and f are edges of G, then e and f have the same label iff α(e) and α(f )
have the same label, so α induces an element F (α) of S5 which takes i to the
label of α(e), where e is any edge labelled i. Clearly F (αβ) = F (α)F (β).
Now label each vertex v of G with the pair {i, j} where i and j are the
numbers that do not occur as labels of the edges at v. (For instance, v1
gets the label {2, 5}.) You can check that each two-element subset {i, j}
of {1, . . . , 5} occurs exactly once, and that vertices u and v are adjacent
iff their labels are disjoint (in which case the edge uv is labelled with the
number not occurring in either vertex label). It follows that for each π ∈ S5
there is a unique α ∈ Aut(G) with F (α) = π, sending a vertex labelled {i, j}
to the vertex labelled {π(i), π(j)}.

43
 v3

v1

v2

Figure 10: The case a = 1, b = 2, c = 3 of Theorem 10.8

Example 10.7. With the notation of Theorem 10.6,

α = (v1 v4 v7 v8 )(v2 v3 )(v5 v9 v10 v6 )

is an automorphism of the Petersen graph with F (α) = (1 2 3 4) ∈ S5 .

Vertices u and v of a graph G are similar if there is an automorphism

α of G with α(u) = v. This is an equivalence relation on V (G), and the
equivalence classes are the orbits of G. A graph with a single orbit is vertex
transitive. A vertex transitive graph is regular, but not conversely (Exercise
10.2). At the other extreme, a graph is asymmetric if its only automorphism
is the identity; equivalently, its orbits are all singletons.

Theorem 10.8. There is an asymmetric graph of order n > 1 iff n ≥ 6.

Proof. Let a, b and c be distinct positive integers. Take the disjoint union
H of Pa , Pb and Pc , and let v1 , v2 and v3 be endpoints of the three compo-
nents. Form G by adding edges v1 v2 , v1 v3 and v2 v3 (see Figure 10). Any
automorphism α of G must take the unique cycle induced by {v1 , v2 , v3 } to
itself, and so induce an automorphism of H. Since the components of H
have different orders, each must be taken to itself, and it follows that α = id.
Thus G is an asymmetric graph of order a + b + c, which can be any integer
n ≥ 6.
It remains to show that there is no asymmetric graph of order n for
2 ≤ n ≤ 5. A reader who has done Exercise 1.1 can simply examine all
the graphs of these orders. To avoid examining all of them, we can note
the following.
Since Aut(G) = Aut(Ḡ) it is enough to consider graphs of
size m ≤ 21 n2 . Any disconnected graph is either empty or has a non-trivial
component of smaller order, so if the result has already been verified for
smaller values of n it is enough to consider connected graphs. There is no
graph satisfying these conditions of order 2 or 3, and for order 4 such a
graph must be a tree, and therefore either P4 or K1,3 , neither of which is
asymmetric. For order 5, such a graph is either a tree or has size 5. The

44
three trees of order 5 are shown in Figure 7, and are not asymmetric. A
connected graph of order 5 and size 5 contains a unique cycle, of length 3, 4
or 5, so there are five such graphs; it is left to the reader to verify that they
are not asymmetric.

A reader who has done Exercise 8.1 can verify that there are no asym-
metric trees of order 6.
Exercises for §10

10.1. Find an automorphism α of the Petersen graph such that, in the

notation of Theorem 10.6, F (α) = (1 2 3)(4 5) ∈ S5 .

10.2. Find a connected regular graph that is not vertex transitive.

10.3. Show that there is an asymmetric tree of any order n ≥ 7.

10.4. Show that for n ≥ 2 there is a graph of order n with vertices of n − 1

different degrees.

10.5. Show that for integers n and k with 1 ≤ k ≤ n − 1 there is a graph

of order n with k orbits.

11 Eulerian graphs
The first paper in graph theory was Euler’s 1736 work on the problem of the
Königsberg bridges, in which he studied what are now known as Eulerian
circuits and trails. An English translation is in Biggs, Lloyd and Wilson
[4], which contains much historical detail, as well as all or part of several
important papers, and references to others.

Definition 11.1. A Eulerian circuit in a graph G is a circuit containing

every edge and vertex of G, and G is Eulerian if it contains a Eulerian
circuit. An Eulerian trail is an open trail containing every edge and vertex
of G.

The essential part of the definition of an Eulerian circuit or trail is that it

contains every edge; that it contains every vertex is automatic unless G has
isolated vertices. Recall that a trail has no repeated edges, so an Eulerian
trail or circuit contains every edge exactly once. A graph with an Eulerian
trail or circuit is connected and non-trivial.

45
Remark 11.2. We make repeated use of the following obvious facts. If C
is a circuit in a graph G, every vertex of G is incident with an even number
of edges of C, while if T is a u-v trail with u 6= v then u and v are incident
with an odd number of edges of T , and every other vertex is incident with
an even number.

Theorem 11.3 (Euler). A non-trivial connected graph G is Eulerian iff

every vertex of G is even.

Proof. The “only if” part follows from the preceding remark. Suppose that
every vertex of G is even, and let T be a trail of maximum length. Suppose
T is a u-v trail. If u 6= v, v is incident with an odd number of edges of T ,
and since deg v is even there is an edge vw not on T . But then T · (v, w) is a
longer trail, which is impossible. Thus T is a circuit. If T is not an Eulerian
circuit, there is an edge xy that is not on T but has at least the vertex x on
T . There is an x-x trail T ′ with the same edges as T , and now T ′ · (x, y) is
a longer trail than T . Thus T is indeed an Eulerian circuit.

Theorem 11.4 (Euler). A connected graph G has an Eulerian trail iff there
are exactly two odd vertices in G.

Proof. The “only if” part again follows from Remark 11.2, so suppose G has
just two odd vertices u and v. Form a graph G′ by adding to G a vertex w
and edges uw and vw. By the previous theorem, G′ is Eulerian. We may
take an Eulerian circuit in G′ of the form T · (v, w, u), and then T is an
Eulerian trail in G.

When there are more than two odd vertices, something can still be said.

Theorem 11.5. Let G be a connected graph with 2k odd vertices, k ≥ 2.

There are k trails in G whose edge-sets partition E(G), and of which at most
one is of odd length.

Proof. Let the odd vertices of G be u1 , v1 , . . . , uk , vk . Form a graph G′ by

adding to G k new vertices w1 , . . . , wk and edges ui wi and vi wi for 1 ≤ i ≤ k.
Then G′ is Eulerian by Theorem 11.3. Pick an Eulerian circuit C of G′ .
There are k sub-trails of C that start at some wi , end at some wj , but
do not otherwise pass through any wk , and their edge-sets partition E(G′ ).
Deleting their first and last edges gives k trails in G whose edge-sets partition
E(G).
We shall call a partition P of E(G) into k subsets, each the edge set of
a trail, a trail partition. For any trail partition P let ν(P) be the number

46
of elements X of P with |X| odd, and suppose for a contradiction that
ν(P) > 1 for any trail partition. We associate a graph H(P) to each trail
partition. The vertices are the elements of P, and X 6= Y ∈ P are adjacent
iff they contain adjacent edges. We show that H(P) is connected. Let X
and Y ∈ P, and let u and v be vertices of edges of X and Y respectively.
Let u = u0 , u1 , . . . , ur = v be a path in G. For 1 ≤ i ≤ r, the edge ui−1 ui
of G belongs to some Zi ∈ P. In the sequence X, Z1 , . . . , Zr , Y of vertices
of H(P), any two consecutive terms are equal or adjacent, so X and Y are
connected in H(P). Hence we may define µ(P) to be the minimum distance
in H(P) between distinct elements X and Y of P with |X| and |Y | odd.
Let P be a trail partition such that for every other trail partition P′ ,
either ν(P) < ν(P′ ), or ν(P) = ν(P′ ) and µ(P) ≤ µ(P′ ). We consider two
cases, depending on whether or not µ(P) = 1. If µ(P) = 1, there are adjacent
vertices X and Y of H(P) with |X| and |Y | odd. These are the edge sets
of trails u0 , u1 , . . . , ur and v0 , v1 , . . . vs in G, where r and s are odd, and
ui = vj for some i and j. By reversing one trail if necessary, we may assume
that i 6≡ j (mod 2). We may form a new trail partition P′ by replacing X
and Y by the edge sets X ′ and Y ′ of the trails u0 , u1 , . . . , ui , vj+1 , . . . , vs and
v0 , v1 , . . . , vj , ui+1 , . . . , ur . Since |X ′ | and |Y ′ | are even, ν(P′ ) = ν(P) − 2, a
contradiction.
Now suppose that µ(P) > 1. Then there are vertices X, Y , Z and W
of H(P) with |X| and |W | odd, |Y | even, Y adjacent to both X and Z,
and dH(P) (Z, W ) = µ(P) − 2. Then X and Y are the edge sets of trails
u0 , u1 , . . . , ur and v0 , v1 , . . . vs in G, where r is odd, s is even, ui = vj for
some i and j, and some vk is a vertex of an edge of Z. Note that j 6= k, since
otherwise X and Z would be equal or adjacent. By reversing the trails if
necessary, we may assume that j < k, and then that i 6≡ j (mod 2). We may
form a new trail partition P′ by replacing X and Y by the edge sets X ′ and Y ′
of the trails u0 , u1 , . . . , ui , vj+1 , . . . , vs and v0 , v1 , . . . , vj , ui+1 , . . . , ur . Since
|X ′ | is odd and |Y ′ | is even, ν(P′ ) = ν(P). Since X ′ and Z have adjacent
edges, µ(P′ ) ≤ dH(P′ ) (X ′ , W ) ≤ µ(P) − 1. This contradiction completes the
proof.

Theorem 11.6. A non-trivial connected graph G is Eulerian iff every edge

of G is on an odd number of cycles.

In the statement of this theorem, “cycle” means a subgraph isomorphic

to some Ck rather than a walk inducing such a subgraph.

Proof. Suppose first that G is Eulerian, so that every vertex is even, and let
e = uv be any edge. Let T be the set of all trails in G − e that start at u

47
and either do not contain v, or contain it only as the terminal vertex. For
k ≥ 0, let Tk be the set of trails in T of length at most k that either have v
as terminal vertex or are of length exactly k. For k = 0, T0 contains only the
trivial u-u path, while for sufficiently large k, Tk is the set T ′ of all trails in T
that end at v. Suppose that T ∈ Tk is a u-x trail where x 6= v. If x 6= u then
x has even degree in G − e and is incident with an odd number of edges of T ,
while if x = u then x has odd degree in G − e and is incident with an even
number of edges of T . In either case there are an odd number of edges xy of
G − e that are not on T , and for each one T · (x, y) is in Tk+1 . Every element
of Tk+1 is either of this form, or is an element of Tk ending at v. It follows
by induction that |Tk | is odd for all k, so |T ′ | is odd. Now the set P of all u-v
paths in G − e is a subset of T ′ , and is in one-to-one correspondence with
the set of all cycles in G containing e. Suppose T = (u = u0 , u1 , . . . , uk = v)
is an element of T ′ − P. Let j be the smallest integer for which there exists
i < j with ui = uj . Then
α(T ) = (u0 , u1 , . . . , ui , uj−1 , . . . , ui+1 , uj , uj+1 , . . . , uk )
is a different element of T ′ − P, and α2 (T ) = T , so |T ′ − P| is even and |P|
is odd, as required.
Now suppose that every edge of G lies on an odd number of cycles. Let
u be a vertex of G, and consider the set P of all pairs (e, C), where e is an
edge incident with u and C is a cycle containing e. Because there are an odd
number of such pairs for each edge incident with u, |P | ≡ deg u (mod 2).
Because there are two such pairs for each cycle containing u, |P | is even.
Hence deg u is even, and G is Eulerian by Theorem 11.3.

12 Hamiltonian graphs
Definition 12.1. A Hamiltonian cycle in a graph G is a cycle containing
every vertex of G, and G is Hamiltonian if it contains a Hamiltonian cycle.
A Hamiltonian graph has order at least 3 and is connected, and indeed,
by Theorem 5.9, 2-connected. Not every 2-connected graph is Hamiltonian,
however; the Petersen graph, for instance, is not Hamiltonian. The graph
of the dodecahedron, shown in Figure 11, is Hamiltonian; the association of
Hamilton’s name with this idea comes from his invention of a game based
on this fact; see [4].
Theorem 12.2. Let G be a graph of order n and u and v independent
vertices of G with deg u + deg v ≥ n. Then G is Hamiltonian iff G + uv is
Hamiltonian.

48
 Figure 11: The graph of the dodecahedron

Proof. The “only if” part is trivial. Suppose that G′ = G + uv is Hamilto-

nian. A Hamiltonian cycle in G′ not containing the edge uv is a Hamiltonian
cycle in G, so suppose that G′ has a Hamiltonian cycle

(u = u1 , u2 , . . . , un = v, u).

The set of integers i with 2 ≤ i ≤ n such that ui is adjacent to u has

cardinality deg u, and the set of those such that ui−1 is adjacent to v has
cardinality deg v, so there is some such i such that ui is adjacent to u and
ui−1 is adjacent to v. Now

(u1 , ui , ui+1 . . . , un , ui−1 , ui−2 , . . . , u1 )

is a Hamiltonian cycle in G.

Given a graph G of order n, we may form a sequence G = G0 , G1 , . . . , Gr

of graphs where, for 0 < i ≤ r, Gi = Gi−1 + ui vi for independent vertices
ui and vi of Gi−1 with degGi−1 ui + degGi−1 vi ≥ n, such that there are no
independent vertices u and v of Gr with degGr u + degGr v ≥ n. We call Gr
a closure of G.
Lemma 12.3. Any graph has a unique closure.
Proof. Let G be a graph of order n, and let G = G0 , G1 , . . . , Gr , u1 , . . . ur
and v1 , . . . , vr be as above. Let G′ be another closure of G. We show, by
induction on i, that each edge ui vi added to G in forming Gr is an edge of
G′ . Suppose then that, for some i, 1 ≤ i ≤ r, uj vj ∈ E(G′ ) for 1 ≤ j < i.
Then
degG′ ui + degG′ vi ≥ degGi−1 ui + degGi−1 vi ≥ n,
so ui and vi are not independent in G′ , and ui vi ∈ E(G′ ), proving the
inductive step. By symmetry, every edge added to G in forming G′ is an
edge of Gr , so Gr = G′ .

49
 We may therefore speak of the closure of G, denoted C(G). The next
three results are immediate consequences of Theorem 12.2.

Corollary 12.4. Let G be a graph of order n ≥ 3. If C(G) ∼

= Kn then G is
Hamiltonian.

Corollary 12.5. Let G be a graph of order n ≥ 3. If deg u + deg v ≥ n for

any independent vertices u and v of G then G is Hamiltonian.

Corollary 12.6. Let G be a graph of order n ≥ 3. If δ(G) ≥ n/2 then G is

Hamiltonian.

The following consequence of Theorem 12.2 takes a bit more work.

Theorem 12.7. Let G be a graph of order n ≥ 3 and (d1 , . . . , dn ) the degree

sequence of G with d1 ≤ · · · ≤ dn . If there is no k with 1 ≤ k < n/2 such
that dk ≤ k and dn−k ≤ n − k − 1 then G is Hamiltonian.

Proof. Let H = C(G). Note that G, and hence H, has no isolated vertices.
We claim that H ∼ = Kn . Suppose not, and let u and v be independent
vertices of H such that degH u + degH v is a maximum. By the definition of
closure, degH u+degH v ≤ n−1. Suppose that degH u ≥ degH v and set k =
degH v. Then 1 ≤ k ≤ (n−1)/2 < n/2. There are n−1−degH u ≥ k vertices
of H independent of u, and for each such vertex w, degG w ≤ degH w ≤
degH v = k. That is, dk ≤ k. There are n − 1 − degH v = n − 1 − k vertices
independent of v in H, and for each such vertex w, degG w ≤ degH w ≤
degH u ≤ n − 1 − k. Since also degG v ≤ k ≤ n − 1 − k, there are at least
n − k vertices of degree at most n − k − 1 in G, and dn−k ≤ n − k − 1.

Definition 12.8. A set U of vertices of a graph G is independent if any

two distinct elements of U are independent. The independence number of
G, β(G), is the maximum number of vertices in an independent set.

For example, β(G) = 1 iff G is complete.

Theorem 12.9. Let G be a graph of order n ≥ 3. If κ(G) ≥ β(G) then G

is Hamiltonian.

Proof. Let k = κ(G). Then k ≥ 2, since otherwise β(G) = 1 and G ∼ = Kn ,

contradicting k = 1. Hence G contains a cycle. Let C be a cycle of maximum
length l. By Theorem 6.11, l ≥ k. Suppose, for a contradiction, that there
is a vertex u not on C. By Theorem 6.10, we may find k internally disjoint
paths Q1 , . . . , Qk from u to distinct vertices of C. Let vi be the first vertex

50
of Qi that is on C and Pi the u-vi subpath of Qi . If C contains an edge
vi vj , we may replace it by Pir · Pj to get a longer cycle, which is impossible.
Let wi be the vertex following vi in one cyclic ordering of C. If some wi is
adjacent to u, we may replace the edge vi wi by Pir · (u, wi ) to get a longer
cycle, which is again impossible. Now |{u, w1 , . . . , wk }| = k + 1 > β(G), so
this set is not independent and G contains an edge wi wj (which is of course
not in C). Replacing the edges vi wi and vj wj by Pir · Pj and (wi , wj ) gives
a longer cycle, and this contradiction completes the proof.

Exercises for §12

12.1. Prove that if G is Hamiltonian of order n then β(G) ≤ n/2.

12.2. Let G be a graph of order n ≥ 3 such that deg u + deg v ≥ n for any
independent vertices u and v of G. Let (d1 , . . . , dn ) be the degree
sequence of G with d1 ≤ · · · ≤ dn . Prove that there is no k with
1 ≤ k < n/2 such that dk ≤ k and dn−k ≤ n − k − 1.

12.3. Show that the complete k-partite graph Kr1 ,...,rk , where r1 ≤ · · · ≤ rk ,
Pk−1
is Hamiltonian iff rk ≤ i=1 ri .

13 Elements of plane topology

We review briefly some definitions from MATH 2057pand a few related ideas.
The length or norm of x = (x1 , x2 ) ∈ R2 is kxk = x21 + x22 , the Euclidean
distance between x and y ∈ R2 is kx−yk, and the (open) ball of radius r > 0
centered at x ∈ R2 is B(x, r) = {y ∈ R2 | kx − yk < r}. A subset A of R2 is
open if, for every x ∈ A, there is some ǫ > 0 such that B(x, ǫ) ⊆ A, and A
is closed if R2 − A is open. The closure of A is the set Ā of all x ∈ R2 such
that B(x, ǫ) meets A for all ǫ > 0; it is the smallest closed set containing
A. The interior of A is the set A◦ of all x ∈ R2 such that there is some
ǫ > 0 such that B(x, ǫ) ⊆ A; it is the largest open set contained in A. The
boundary of A is Ā − A◦ ; alternatively, it is the set of all x ∈ R2 such that
B(x, ǫ) meets both A and R2 − A for all ǫ > 0.
A curve in R2 is a continuous function f : [0, 1] → R2 . An arc in R2 is
the image of an injective curve, and its endpoints are the images of 0 and
1. It can be shown that this is well-defined; that is, if f and g are injective
curves with the same image then f ({0, 1}) = g({0, 1}). (This can be proved
using two theorems usually stated but not proved in a first calculus course,
the Intermediate and Extreme Value Theorems.) A subset A of R2 is path-
connected if, for any x, y ∈ A, there is a curve f in A with f (0) = x

51
 (a) (b)

Figure 12: The planar graph K2,3

and f (1) = y. The path-components of A are the maximal path-connected

subsets of A; they form a partition of A and the path-components of an
open set are open. A simple closed curve is the image of a curve f such
that, for 0 ≤ s < t ≤ 1, f (s) = f (t) iff s = 0 and t = 1. If C is a simple
closed curve in R2 , R2 − C has exactly two path-components, one bounded
and one unbounded, each with boundary C. This is the celebrated Jordan
Curve Theorem, and will be taken on trust.
Suppose that A, B and C are arcs in R2 such that any two meet only
in their endpoints x and y. If X is one of A, B or C, then the union of the
other two is a simple closed curve, whose complement has path-components
UX disjoint from X, and VX containing X − {x, y}. The path-components
of R2 − (A ∪ B ∪ C) are UA , UB and UC .

14 Planar graphs
Definition 14.1. A plane embedding of a graph G consists of the following.
First, a bijection α from V (G) to a subset of R2 . Second, a bijection β
from E(G) to a set of arcs in R2 . These are required to satisfy the following
conditions. First, if e is the edge uv, then the endpoints of β(e) are α(u)
and α(v). Second, if e and f are distinct edges, then β(e) and β(f ) intersect
only if e and f are adjacent, and then only in α(u) where u is the common
vertex of e and f .
Less formally, a plane embedding of G is a picture of G in which no
edge crosses itself or any other edge. A graph is planar if it has a plane
embedding. A plane graph is a graph G together with a plane embedding
of G. Given a plane graph we usually identify each vertex v with the point
α(v) in R2 , each edge e with the arc β(e), and the graph G with the union
of these subsets of R2 . Figure 12 shows two pictures of K2,3 . That in (a)
is not a plane embedding, while the one in (b) is, and shows that K2,3 is
planar.

52
 If G is a plane graph, the path-components of the open set R2 − G
are called the regions of G. There is one unbounded region, the exterior
region. The boundary of a region R is a subgraph of G denoted by ∂R;
we say that a region is adjacent to the vertices and edges in its boundary.
A plane graph G may be re-embedded so that any given vertex or edge is
adjacent to the exterior region as follows. Regard R2 as the plane of the
first two coordinates in R3 , let S 2 be the unit sphere centered at the origin
in R3 , and let N = (0, 0, 1), the “north pole” of S 2 . There is a bijection
φ : S 2 − {N } → R2 , called stereographic projection. For x ∈ S 2 − {N }, φ(x)
is the point in which the line through N and x meets R2 . Let x be a point
of some region adjacent to the given vertex or edge, and take the image of
G under the composite of φ−1 , followed by a rotation of S 2 taking φ−1 (x)
to N , followed by φ.
Clearly a graph is planar iff each component is planar. Let G be a non-
trivial, connected plane graph. A arrow of a graph is an ordered pair (u, v)
of adjacent vertices (i.e., it is a path of length 1). For each arrow (u, v) of G,
there is a region R+ to its left and a region R− to its right. If uv is a bridge
then R+ = R− ; otherwise R+ 6= R− . A side of a region R is an arrow
having R to its left. The number of sides of R is thus twice the number
of bridges plus the number of other edges of ∂R. We may assume that at
each vertex v there is a circle centered at v meeting each edge incident to
v just once. Traversing this circle either clockwise or counterclockwise gives
a cyclic ordering of the vertices adjacent to v. For any region R of G we
construct a closed walk in its boundary as follows. Start with a side (u0 , u1 )
of R. Suppose we have constructed (u0 , u1 , . . . , uk ) so that (ui , ui+1 ) is a
side of R for 0 ≤ i < k. Let v be the vertex adjacent to uk that follows
uk−1 in the clockwise ordering, or uk−1 if uk has degree 1. Then (uk , v) is
a side of R. If uk = u0 and v = u1 , stop; otherwise extend the walk by
taking uk+1 = v. The result is a walk containing each side of R just once.
If there are no bridges in ∂R the walk is a circuit, and if there are also no
cut-vertices it is a cycle. In particular, if G is 2-connected, the boundary of
any region is a cycle.
Theorem 14.2 (Euler’s formula). Let G be a connected plane graph of order
n and size m having r regions. Then n − m + r = 2.
Proof. The proof is by induction on m. If G is a tree then m = n − 1, and
since every edge is a bridge, r = 1, so the result holds. Otherwise, let e be
an edge that lies on a cycle. Then G − e is a connected plane graph of order
n and size m − 1 having r − 1 regions (the two regions of G adjacent to e
lying in a single region of G − e). The result follows.

53
Corollary 14.3. Let G be a plane graph of order n and size m having r
regions and k components. Then n − m + r = k + 1.

Theorem 14.4. If G is a planar graph of order n ≥ 3 and size m then

m ≤ 3n − 6.

Proof. We may assume that G is connected, since otherwise we may add

edges to get a connected planar graph of order n and size greater than m.
Embed G in the plane and let the number of regions be r. The number
of arrows in G is 2m. Now each arrow is a side of a unique region, and
each region has at least three sides, so 2m ≥ 3r, and by Euler’s formula
6 = 3n − 3m + 3r ≤ 3n − m.

Example 14.5. For K5 we have m = 10 and 3n − 6 = 9, so K5 is not

planar.

Corollary 14.6. Every planar graph G has δ(G) ≤ 5.

Proof. Let G have order n and size m.

P If n < 3 the result is trivial, and
otherwise m ≤ 3n − 6. Now 2m = v∈V (G) deg v ≥ δ(G)n, so δ(G)n ≤
6n − 12.

A maximal planar graph is a planar graph G of order n ≥ 3 such that,

for any independent vertices u and v, G + uv is non-planar.

Theorem 14.7. Let G be a maximal planar graph of order n and size m.

Then m = 3n − 6, and in any plane embedding of G, the boundary of any
region is a triangle.

Proof. The second statement implies the first, because it gives 2m = 3r,
where r is the number of regions, and together with Euler’s formula this
gives m = 3n − 6. Clearly G is connected; in fact, it is 2-connected, because
if u were a cut vertex, there would be vertices v and w adjacent to u, with
w following v in the counterclockwise order of the neighbors of u, and in
different components of G − u. But then we could embed G + vw by putting
the edge vw in the region to the left of (u, v). Thus the boundary of every
region is a cycle. Suppose that the boundary of some region R has length
greater than 3, and let u1 , u2 , u3 and u4 be consecutive vertices around ∂R.
Not both of u1 u3 and u2 u4 can be edges of G, and for one that is not we can
embed a corresponding arc in R to contradict maximal planarity of G.

Theorem 14.8. A graph is planar iff each of its blocks is planar.

54
Proof. One direction is trivial: every subgraph of a planar graph is planar.
Suppose that G is a graph all of whose blocks are planar. We may assume
that G is connected. The proof is by induction on the number k of blocks,
the case k = 1 being obvious. Suppose that k > 1. By Theorem 5.18, G has
an end-block B. Let v be the cut-vertex in B, and H the union of all other
blocks of G, so that G = B ∪ H and B and H have only v in common. By
induction, H is planar. We may embed both B and H in R2 so that v is
adjacent to the exterior region. Further, we may choose these embeddings so
that v = (0, 0) in both cases, every other point (x1 , x2 ) of B has x1 < 0, and
every other point (x1 , x2 ) of H has x1 > 0. This gives a plane embedding of
G.
Definition 14.9. The girth of a graph that contains a cycle is the minimum
length of a cycle.
Theorem 14.10. Let G be a 2-connected planar graph of order n, size m
g
and girth g. Then m ≤ g−2 (n − 2).
Proof. Embed G in the plane and let the number of regions be r. Since G is
2-connected, the boundary of any region is a cycle, so it has at least g sides.
Thus 2m ≥ gr, so 2g = gn − gm + gr ≤ gn − (g − 2)m.
g
Example 14.11. For K3,3 , m = 9, n = 6, g = 4 and g−2 (n − 2) = 8, so
K3,3 is non-planar.
Example 14.12. For the Petersen graph, m = 15, n = 10, g = 5 and
g 40
g−2 (n − 2) = 3 , so it is non-planar.
Theorem 14.13. Let G be a connected plane graph with δ(G) ≥ 3. Then
there is a region with at most 5 sides.
Proof. Let the order, size and number of regions be n, m and r. We have
2m ≥ 3n and if every region has at least 6 sides 2m ≥ 6r, or m ≥ 3r. Hence
6 = 3n − 3m + 3r ≤ 0, a contradiction.
Theorem 14.14. Let G be a connected plane graph with δ(G) ≥ 3. Let nk
be the number of vertices of degree k and rk the number of regions with k
sides. Then n3 + r3 ≥ 8.
Proof. Let the order, size and number of regions
P be n, m and
P r. Note that
nk = rk = 0 for k <
P P3. We have n = k≥3 nk , r = k≥3 rk , 2m =
k≥3 kn k , and 2m = k≥3 krk . Hence
X
8 = (4n − 2m) + (4r − 2m) = (4 − k)(nk + rk ) ≤ n3 + r3 .
k≥3

55
 The following theorem is due independently to Fáry [13] and Wagner
[30]. The proof below is based on Wood [32].
Theorem 14.15. Any planar graph has a plane embedding in which each
edge is a straight line segment.
Proof. It is enough to prove this for a maximal planar graph G, which we
do by induction on the order n. In fact we prove that for any embedding of
G there is a line-segment embedding in which the counterclockwise order on
the neighbors of any vertex is unchanged. When n = 3, G is isomorphic to
K3 and the result is clear, so suppose n ≥ 4. Recall that all regions of the
embedding are bounded by triangles. If u and v are adjacent vertices, the
edge uv is in the boundary of two regions, whose other vertices are common
neighbors of u and v. We show that u and v may be chosen so that these are
their only common neighbors. If there are three common neighbors, uv is
in three triangles of G, at least one of which has a vertex in its interior and
one in its exterior. Amongst all triangles of G which have a vertex in both
their interior and their exterior, choose an innermost one, choose a vertex
u in its interior, and let v be any neighbor of u. Then indeed u and v have
just two common neighbors p and q.
With a suitable choice of notation, the neighbors of u in counterclockwise
order are v, p, x1 , . . . , xr , q, and those of v are u, q, y1 , . . . , ys , p. Form a graph
G′ by deleting the vertex u and adding edges vxi for 1 ≤ i ≤ r. (This oper-
ation is called contracting the edge uv, and will appear in the next section.)
We may extend the embedding of G − u to an embedding of G′ , in which
the neighbbors of v in counterclockwise order are p, x1 , . . . , xr , q, y1 , . . . , ys .
By inductive hypothesis we may re-embed G′ so the edges are line seg-
ments, preserving orderings of neighbors. Suppose the angle θ between
vp and vq, measured counterclockwise, is at most π. Consider the cycle
C = (v, p, x1 , . . . , xr , q, v). Its interior contains the edges vxi , except for
their endpoints. Delete these edges, take u to be in the interior of C, and
add the line segments from u to v, p, q and each xi . If u is taken sufficiently
close to v, these segments are in the interior of C, except for their endpoints
on C, and we have the desired embedding of G.
If θ > π, rename v as u, and similarly situate v in the interior of
(u, q, y1 , . . . , ys , p, u).

15 Kuratowski’s Theorem
Definition 15.1. An elementary subdivision of a graph G is the operation
of deleting an edge uv and adding a new vertex w and edges uw and vw.

56
A subdivision of G is a graph isomorphic to one obtained by zero or more
elementary subdivisions.
Example 15.2. The Petersen graph has a subgraph that is a subdivision
of K3,3 , obtained by deleting two edges with the same label in Figure 9.
Clearly a sudivision of G is planar iff G is planar, and as already re-
marked, a subgraph of a planar graph is planar. Together with Examples
14.5 and 14.11, these observations prove the “only if” part of the following
theorem.
Theorem 15.3 (Kuratowski [23]). A graph G is planar iff no subgraph of
G is a subdivision of K5 or K3,3 .
Lemma 15.4. Let G be a 2-connected graph of order n ≥ 4. Then G has
either a vertex v such that G − v is 2-connected, or a vertex of degree 2.
Proof. Suppose that G has no vertex v such that G − v is 2-connected.
Then, for evey vertex x of G there is a vertex y 6= x such that G − {x, y}
is disconnected. Let u and v be such a pair of vertices for which some
component of G − {u, v}, say G1 , has minimum order. Note that each of u
and v is adjacent to some vertex in every component of G − {u, v}. Let G2
be the union of the components of G − {u, v} other than G1 , and for i = 1
or 2, let Hi = hV (Gi ) ∪ {u, v}i. Let w1 be a vertex of G1 . Then there is a
vertex w2 6= w1 such that G − {w1 , w2 } is disconnected. If w2 is in G1 , then
because H2 is connected, some component of G − {w1 , w2 } is contained in
G1 − {w1 , w2 }, a contradiction. If w2 = u, then because hV (G2 ) ∪ {v}i is
connected, some component of G − {w1 , w2 } is contained in G1 − w1 , again
a contradiction. A similar contradiction arises if w2 = v, so w2 is in G2 .
For i = 1 or 2, every vertex of Hi − wi is connected to either u or v in
Hi − wi , because it is connected to, say, u in G − wi and a path can only
leave Hi at u or v. It follows that Hi − wi has exactly two components, one
containing u and the other containing v. Let the components of H1 − w1 be
Ku and Kv . If Ku is non-trivial, then Ku − u is a union of components of
G − {w1 , u}, contradicting the minimality of the order of G1 . Thus Ku is
trivial, and similarly Kv is trivial. This means that G1 has only the single
vertex w1 , which is therefore adjacent to only u and v, and of degree 2.

Lemma 15.5. Let G be a 2-connected graph of order n ≥ 4. Then G has

either an edge e such that G − e is 2-connected, or a vertex of degree 2.
Proof. By the previous lemma, we may assume that G has a vertex v such
that G − v is 2-connected. Let e be an edge incident with v. If G − e is

57
2-connected, we are done. Otherwise there is a vertex u such that G − e − u
is disconnected; u is not a vertex of e since otherwise G − e − u = G − u.
Now e is a bridge of G − u but its vertex v is not a cut-vertex of G − u.
Hence v has degree 1 in G − u, and therefore degree 2 in G.

Let G be a graph and Γ a cycle of G. We assume that an orientation

of Γ (that is, a cyclic ordering of its vertices) has been chosen. For distinct
vertices u and v of Γ, we let [u, v] be the u-v path in Γ determined by the
orientation. (The other one is the reverse of [v, u].) We let (u, v], [u, v) and
(u, v) be the sets obtained from the set of vertices of [u, v] by removing u,
v, or both. (The “open interval” (u, v) may be empty.) We also allow the
degenerate closed interval [u, u], the trivial u-u path. By a chord of Γ we
mean a path C in G having only its end vertices in common with Γ. We
define pieces of G relative to Γ of two kinds. For each component K of
G − V (Γ) there is a piece, which is the subgraph of G induced by all edges
with at least one vertex in K. Also, the subgraph induced by an edge not
on Γ, but with both vertices on Γ, is a piece. The edge sets of Γ and all
pieces partition E(G). If G is connected, each piece has at least one vertex
on Γ, and if G is 2-connected, at least two. If u and v are distinct vertices
of some piece P that lie on Γ, P contains a u-v chord of Γ. We say that two
subgraphs H1 and H2 of G, each of which is either a chord or a piece, cross
if there are vertices ui and vi of Hi on Γ (i = 1 or 2), such that u1 , u2 , v1
and v2 are distinct and appear in that order around Γ.
Suppose that G has been embedded in the plane. For definiteness, we
shall assume that Γ has the counterclockwise orientation. The plane is
divided by Γ into two regions, the interior and the exterior. Each vertex
and edge not on Γ, as well as each chord or piece, and each region of G,
may be classified as interior or exterior. Note that if two subgraphs (each a
chord or piece) cross, one is interior and the other exterior.

Lemma 15.6. Let G be a 2-connected plane graph, and Γ a cycle of G.

Suppose that u and v are vertices of Γ that are not adjacent to any single
interior region of G. Then there is an interior chord of Γ with one endpoint
in (u, v) and the other in (v, u).

Remark. The assumption that G is 2-connected is not essential, but sim-

plifies the proof, and is satisfied in our applications.

Proof. The sets (u, v) and (v, u) are non-empty; let w ∈ (u, v) and w′ ∈ (v, u)
be adjacent to u. We first define a w-w′ walk with all edges (and hence
vertices) on or interior to Γ. Set w0 = w. Let R1 be the interior region

58
of G adjacent to uw0 , and take the path going counterclockwise around
the boundary of R1 from w0 to its other vertex w1 adjacent to u. If uw1
is interior, let R2 be the other (necessarily interior) region of G adjacent
to uw1 , and take the path going counterclockwise around the boundary of
R2 from w1 to its other vertex w2 adjacent to u. Continue in this fashion,
stopping when uwk is on Γ, and therefore wk = w′ . Stringing these paths
together gives the desired walk W . Note that neither u nor v is on W .
Let W be w = x0 , x1 , . . . , xr = w′ . There exist i and j with 0 ≤ i < j ≤ r
such that xi is on (u, v), xj is on (v, u), and xk is interior to Γ for i < k < j.
An xi -xj path contained in the walk xi , xi+1 , . . . , xj is the required chord.

If a graph G has a subgraph that is a subdivision of K5 or K3,3 , we shall

say that G has a forbidden subgraph. The next lemma is the heart of the
proof of Kuratowski’s Theorem.

Lemma 15.7. Suppose that G is a non-planar graph that has an edge e such
that G − e is 2-connected and planar. Then G has a forbidden subgraph.

Proof. Let e = u0 u2 . (The reason for the peculiar indexing will emerge
later.) Set H = G − e. Since H is 2-connected, it has a cycle Γ containing
u0 and u2 . The pieces of G relative to Γ are the pieces of H together with
the chord C0 induced by e. For any plane embedding of H, there must be
an interior piece of H crossing C0 , since otherwise Lemma 15.6 would give
an interior region adjacent to u0 and u2 , and hence a plane embedding of
G.
Now fix some embedding of H, and choose Γ to have the maximum
number of interior regions amongst all cycles of H containing u0 and u2 .
Let P be an exterior piece of H. Since H is 2-connected, P has at least two
vertices on Γ. Suppose it has two such v and w for which (v, w) contains
neither u0 nor u2 . Then the union of the path [w, v] and a v-w chord in P is
a cycle containing u0 and u2 with more interior regions than Γ, contrary to
hypothesis. It follows that P has exactly two vertices on Γ, with one, vP , on
(u0 , u2 ), and the other, wP , on (u2 , u0 ). Amongst the regions of the subgraph
Γ ∪ P of H, there are three whose boundaries contain edges of Γ. One is the
interior of Γ. Of the others, one, RP− , has boundary meeting Γ in [wP , vP ],
and the otherRP+ , has boundary meeting Γ in [vP , wP ]. The boundary of
any other region has at most the vertices vP and wP in common with Γ, and
cannot have both. (Suppose it does. Then its boundary is the union of two
vP -wP chords of Γ, both of which must contain exterior vertices. However,
any path in Γ ∪ P from an exterior vertex of one chord to an exterior vertex
of the other must pass through vP or wP , contradicting the fact that the

59
 w1 = w2 R1

R0 u0 u2 R2

v1 v2

Figure 13: Exterior pieces and the regions Ri

exterior vertices of P belong to a single component of H − V (Γ).) If P ′

is any other exterior piece of H, its edges and exterior vertices lie in some
region of Γ ∪ P , and since it has vertices in both (u0 , u2 ) and (u2 , u0 ), this
must be RP− or RP+ . In the first case we say that P follows P ′ , and in the
second that P precedes P ′ . Then P precedes P ′ iff P ′ follows P , and if P
precedes P ′ and P ′ precedes P ′′ then P precedes P ′′ . Hence we may number
the exterior pieces of H as P1 , . . . , Pk so that Pi precedes Pi+1 for 1 ≤ i < k.
We set vi = vPi and wi = wPi . For 1 ≤ i < k, the intersection of RP+i and
RP−i+1 is a region Ri of H whose boundary meets Γ in the paths [vi , vi+1 ]
and [wi+1 , wi ]. We also set R0 = RP−1 and Rk = RP+k . (See Figure 13, in
which the interior pieces of H are not shown.)
Now consider an interior piece Q of H. If Q does not cross some Pi , all
its vertices on Γ lie on the boundary of some Ri . Thus we may re-embed
Q in Ri . We may do this simultaneously for all such interior pieces, since
they do not cross one another. In this way, we obtain a new embedding
of H in which every interior piece crosses some Pi . As remarked above, in
this embedding there must be some interior piece Q crossing C0 . Suppose
that Q also crosses Pj . Set u1 = vj and u3 = wj , so that u0 , u1 , u2 and
u3 appear in that order around Γ. We shall take the subscripts on these
vertices modulo 4. Choose a u1 -u3 chord C1 in Pj , and note that the union
of Γ, C0 and C1 is a subdivision of K4 . The piece Q has a vertex in each
of the four open intervals (ui , ui+2 ). There are two cases, (each with two
subcases) depending on whether or not it has a vertex in some (ui , ui+1 ).
Suppose first that Q has a vertex x1 in (ui , ui+1 ). If it also has a vertex
x2 in (ui+2 , ui+3 ), the union of Γ, C0 , C1 and an x1 -x2 chord in Q is a
subdivision of K3,3 . (See Figure 14(a).) Otherwise, Q has vertices x2 in
(ui+1 , ui+2 ] and x3 in [ui+3 , ui ). Take an x2 -x3 chord D1 in Q. Then D1 has
interior vertices, and each one may be joined to x1 by a path in Q whose
only vertex on Γ is x1 . Taking a shortest such path, we obtain an x1 -x4

60
 u0 u3 u0 u3
x3

x1 x2 x1
x4

x2
u1 u2 u1 u2
(a) (b)

u0 u3 u0 u3

x1

x1
x2

u1 u2 u1 u2
(c) (d)

Figure 14: Forbidden subgraphs

path D2 in Q whose only vertex on Γ is x1 and whose only vertex on D1

is the interior vertex x4 . The union of [ui+3 , ui+2 ], C0 , C1 , D1 and D2 is a
subdivision of K3,3 . (See Figure 14(b).) The proof is thus complete in this
case.
Now suppose that Q has no vertex in any (ui , ui+1 ). Then its only
possible vertex in (ui−1 , ui+1 ) is ui , so every ui is a vertex of Q. Take a
u0 -u2 chord D1 in Q. As above, we may find a u1 -x1 path D2 and a u3 -x2
path D3 in Q where the only vertex of D2 on Γ is u1 , the only vertex of D2
on D1 is the interior vertex x1 , the only vertex of D3 on Γ is u3 , and the
only vertex of D3 on D1 is the interior vertex x2 . Of the two regions into
which D1 divides the interior of Γ, the internal vertices of D2 lie in one, and
those of D3 lie in the other, so D2 and D3 are internally disjoint. If x1 = x2 ,
the union of Γ, C0 , C1 , D1 , D2 and D3 is a subdivision of K5 . (See Figure
14(c).) Otherwise, one of x1 and x2 precedes the other as we traverse D1
from u0 to u2 . Suppose it is x1 . Then the union of [u1 , u2 ], [u3 , u0 ], C0 , C1 ,

61
 u

v w
*

v w G
j
H
u
v w

Figure 15: The moves for the proof of Theorem 15.3

D1 D2 and D3 is a subdivision of K3,3 . (See Figure 14(d).) The remaining

case is similar, and we are done.

Proof of Theorem 15.3. What remains to be proved is that every non-planar

graph has a forbidden subgraph. Note that if H has a forbidden subgraph
and either H is a subgraph of G or G is a subdivision of H, then G has a
forbidden subgraph. Let us say that a graph G1 is smaller than a graph G2
if either n(G1 ) < n(G2 ) , or n(G1 ) = n(G2 ) and m(G1 ) < m(G2 ). It suffices
to prove that if G is a non-planar graph such that every smaller non-planar
graph has a forbidden subgraph, then G has a forbidden subgraph. If G is
not 2-connected, then every block of G is smaller than G, and by Theorem
14.8, G has a non-planar block, so the result follows in this case. Suppose
then that G is 2-connected. If δ(G) < 3 then G has a vertex u of degree
2. Let the two vertices adjacent to u be v and w. Form H by deleting the
vertex u and, if v and w are not adjacent in G, adding an edge vw (in which
case G is a subdivision of H). Then H is smaller than G and is non-planar,
since a plane embedding of H would give rise to one of G by one of the two
moves in Figure 15. Therefore H, and hence G, has a forbidden subgraph.
Suppose further that δ(G) ≥ 3. By Lemma 15.5, G has an edge e such that
G − e is 2-connected. If G − e is non-planar then, being smaller than G, it
has a forbidden subgraph, and hence so does G. Finally, if G − e is planar
then G has a forbidden subgraph by Lemma 15.7.

Let e = uv be an edge of a graph G. The operation of contracting e is

defined as follows. To G − {u, v} add a new vertex w together with edges
wx for all x adjacent to either u or v.

62
Definition 15.8. A minor of a graph G is a graph isomorphic to one ob-
tained from G by zero or more applications of the operations of deleting
or contracting an edge, or deleting a vertex. A minor obtained using only
contractions is a contraction of G.

Example 15.9. The Petersen graph has K5 as a contraction.

Clearly any subgraph of G is a minor of G, and if H is a minor of G and

K is a minor of H then K is a minor of G.

Lemma 15.10. If G is a subdivision of H then H is a minor of G.

Proof. It suffices to consider the case where G is obtained from H by an

elementary subdivision. If the edge of H involved is uv and the new vertex
in G is w, contracting the edge vw of G gives a graph isomorphic to H.

Theorem 15.11. A graph G is planar iff neither K5 nor K3,3 is a minor

of G.

Proof. It is obvious that if G is planar, so is any graph obtained by contract-

ing an edge, and hence any minor of G, which proves the “only if” part. The
converse is given by Kuratowski’s Theorem and the preceding lemma.

Exercises for §15

15.1. Show that any minor of G is a contraction of a subgraph of G.

15.2. Show that H is a contraction of G iff there is a surjection φ : V (G) →

V (H) such that, for vertices u and v of H, the subgraph of G induced
by φ−1 (u) is connected, and u and v are adjacent iff there is an edge
of G with one vertex in φ−1 (u) and the other in φ−1 (v).

15.3. Show, without using Kuratowski’s Theorem, that if G has K5 or K3,3

as a minor then it has a subgraph that is a subdivision of K5 or K3,3 .

16 Crossing number
In this section we need to be more precise about what we mean by a drawing
of a graph.

Definition 16.1. A (plane) drawing of a graph G consists of the following.

First, a bijection α from V (G) to a subset of R2 . Second, a bijection β
from E(G) to a set of arcs in R2 . These are required to satisfy the following

63
conditions. First, if e is the edge uv, then the endpoints of β(e) are α(u)
and α(v). Second, if e and f are adjacent edges, β(e) and β(f ) intersect
only in α(u) where u is the common vertex of e and f . Thirdly, if e and f
are distinct non-adjacent edges then β(e) and β(f ) intersect in at most one
point, at which β(e) and β(f ) cross, and which lies on the image of no other
edge. The crossing number of G is the minimum number ν(G) of crossings
in any drawing of G.

Thus ν(G) = 0 iff G is planar. As with embeddings, we identify a vertex

with its image under α and an edge with its image under β. The preceding
definition could be relaxed by allowing an edge to cross itself, or adjacent
edges to cross, or edges to cross more than once. However, it is easy to see
that such crossings could be eliminated while decreasing the total number
of crossings, so this would not change the crossing number. If we allow more
than two edges to cross at a point, we would have to replace “number of
crossings” in the definition of ν(G) by “number of unordered pairs of edges
that cross”. (For instance, if three edges have a common point at which they
mutually cross, we can perturb one slightly to produce three crossings on
two edges each.) It will sometimes be more convenient to describe drawings
in this generalized sense.

Definition 16.2. The rectilinear crossing number ν̄(G) of a graph G is

the minimum number of crossings in a drawing of G in which all edges are
straight line segments.

Clearly ν(G) ≤ ν̄(G). By the Fáry-Wagner Theorem (14.15), if ν(G) = 0

then ν̄(G) = 0, but equality does not hold in general. It is true that allowing
more than two edges to cross at a single point does not change ν̄(G) provided
one counts crossings properly; in this case, we need to perturb the vertices.
Computing crossing numbers seems to be hard; mostly what is known is
upper bounds, conjectured to be exact.

Theorem 16.3. The complete graph Kn has

    
1 n n−1 n−2 n−3
ν(Kn ) ≤ .
4 2 2 2 2

Proof. If n < 5, ν(Kn ) = 0 and the result holds, so suppose n ≥ 5. Suppose

first that n = 2k, k ≥ 3. In the strip {(r, s) | 1 ≤ r ≤ 2} of R2 , consider the
set of all straight line segments from (1, i) to (2, i+a) for i, a ∈ Z and 0 ≤ a <
k. Under the map f : R2 → R2 given by f (r, s) = (r cos 2πs 2πs
k , r sin k ), the
endpoints of these segments map to 2k points ui = f (1, i) and vi = f (2, i)

64
for 0 ≤ i < k, and the segments to k 2 arcs which are the edges of a drawing
of Kk,k . To obtain a drawing of Kn we first add edges between distinct ui
and uj that are straight line segments. These do not cross any edges of the
first type, and for any four of the ui there is just one crossing on the
edges
they induce. Thus the number of crossings on edges of this type is k4 . We
may similarly add edges between distinct vi and vj lying in the set of points
at distance at least 2 from the origin, having a further k4 crossings. We

now count, for a vertex w, the number of crossings on edges of the first type
incident to w. This is independent of w, so we take w = u0 . For given
a and b with 0 ≤ a, b < k, we determine the number of crossings of u0 va
with edges ui vi+b . If a = b, there are none. If a 6= b the corresponding line
segments cross if either i < 0 and i + b > a, or i > 0 and i + b < a; that is,
if a − b < i < 0 or 0 < i < a − b. At most one of these ranges is non-empty,
and different values of i in the range give different edges, so the number of
crossing edges is |a − b| − 1. Hence the number of crossings on edges of the
first type incident with w is
k X
X a−2 k
X
a−1 k



2 (a − b − 1) = 2 2 =2 3 .
a=3 b=1 a=3

There are 2k vertices, and each crossing is on edges incident with four
vertices,
so the
total number of crossings on edges of the first type is
2k · 2 k3 /4 = k k3 , and the number of all crossings in the diagram is
k

k

k(k−1)2 (k−2) 1
 n   n−1   n−2   n−3 
k 3 +2 4 = 4 = 4 2 2 2 2 .

Now suppose n = 2k − 1, k ≥ 3. In the drawing of K2k
above,
the

number of crossings on the edges incident with any vertex is 2 k3 + k−1
3 =
(k − 1)2 (k − 2)/2, so deleting a vertex gives a drawing of Kn with
k(k−1)2 (k−2) (k−1)2 (k−2) (k−1)2 (k−2)2
4 − 2 =
 4 n−1   n−2   n−3 
1 n
= 4 2 2 2 2

crossings.

The bound in this theorem is known to be exact for n ≤ 10. It gives 0

for n ≤ 4 and 1 for n = 5, which are clearly exact. Let us verify that the
bound ν(K6 ) ≤ 3 is also exact. Suppose we have a drawing of K6 with c
crossings. Replacing each crossing by a vertex gives a plane graph of order
c + 6 and size 2c + 15, so by Theorem 14.4, 2c + 15 ≤ 3(c + 6) − 6, or c ≥ 3.
The known case n = 8 gives ν(K8 ) = 18; it is also known that ν̄(K8 ) = 19.

65
 Figure 16: C4 × C5 with 10 crossings

Theorem 16.4. For the complete bipartite graph Kr,s ,

    
r r−1 s s−1
ν(Kr,s ) ≤ ν̄(Kr,s ) ≤ .
2 2 2 2
r
Proof. As the vertices of one  r partite set we take r+ = 2 points on the
positive x-axis and r− = 2 points on the negative x-axis, and  s for the
s
other we take s+ = 2 points on the positive y-axis and s− = 2 points
on the negative y-axis. The edges are represented by line segments. Each
crossing determines two points on one half of the x-axis and two on one half
of the y-axis, and conversely any such points determine one crossing. Hence
the number of crossings is
r + s+ r + s− r − s+ r − s− r+ r−











s+
s−



2 2 + 2 2 + 2 2 + 2 2 =  2  + 2   2 + 2
= 2r r−1 2
s
2
s−1
2 .

It is conjectured that both bounds in this Theorem are exact, and this
is known for r ≤ 6 and any s, and r = 7 and s ≤ 10.
Figure 16 shows a straight-line drawing of C4 × C5 with (4 − 2) · 5 = 10
crossings, from which the general case of Cr × Cs given in the next result
should be clear.
Theorem 16.5. ν(Cr × Cs ) ≤ ν̄(Cr × Cs ) ≤ (r − 2)s.
It is conjectured that, for r ≤ s, both these bounds are exact, and this
is known for r = 3 and 4. We shall prove it for r = 3.
Lemma 16.6. ν(C3 × C3 ) = 3.

66
Proof. What remains to be proved is that any drawing of G = C3 × C3 has
at least three crossings. Consider a drawing with c crossings. Certainly G
is not planar (see below), so c > 0. Pick an edge containing a crossing, and
then a 4-cycle containing this edge. Deleting the edges of this cycle leaves
a subdivision of K5 , so there is at least one crossing among the remaining
edges. If there are two or more then c ≥ 3, so suppose there is just one.
This means that our original drawing can be obtained from a one-crossing
drawing of K5 by picking points subdividing the edges of a 4-cycle in K5 ,
and adding an edge between the subdivision points on each adjacent pair of
edges. There is an essentially unique one-crossing diagram of K5 ; replacing
the crossing by a vertex gives the plane embedding of the graph of the
octahedron. We must consider all ways of choosing a 4-cycle in the drawing
of K5 , and for any edge on the cycle through the crossing, choosing which
side of the crossing to put the subdivision point.
We consider the drawing as lying on the unit sphere in R3 , with vertices
u± = (±1, 0, 0), v± = (0, ±1, 0) and w = (0, 0, 1), and with edges lying in
the coordinate planes. The edges u+ u− and v+ v− are semicircles crossing
at (0, 0, −1), and the other edges are quarter-circles. Rotation through a
multiple of π2 about the z-axis induces an automorphism of the K5 , as does
reflection in any of the vertical planes x = 0, y = 0, x = y and x = −y.
This means we may assume that our 4-cycle is one of the four below. In
each case, we denote the subdivision points on the edges by a1 , a2 , a3 and
a4 in the order in which the edges appear in the cycle as written.
(a) (u+ , v+ , u− , v− , u+ ). In this case every added edge must have a cross-
ing, so c ≥ 5.
(b) (u+ , v+ , v− , u− , u+ ). In this case we may further assume that a4 has
positive x-coordinate and that a2 has positive y-coordinate. Then each edge
added at a3 must have a crossing, so c ≥ 3.
(c) (u+ , v+ , u− , w, u+ ). Here each of a1 a2 and a3 a4 must have a crossing,
so c ≥ 3.
(d) (u+ , u− , v+ , w, u+ ). The edge a1 a4 must have one crossing if a1 has
positive x-coordinate, and two if it has negative x-coordinate, and if a1
has positive x-coordinate the edge a1 a2 must also have a crossing, so again
c ≥ 3.

Theorem 16.7. For any s ≥ 3, ν(C3 × Cs ) = s.

Proof. This is by induction on s, the case s = 3 being the previous lemma.

Suppose then that s > 3 and the result is true for s − 1. We let the vertices
of G = C3 × Cs be ui , vi and wi for i ∈ Zs , where the triangles are Ti =

67
hui , vi , wi i and the s-cycles are induced by the sets U = {ui | i ∈ Zs },
V = {vi | i ∈ Zs } and W = {wi | i ∈ Zs }, with edges ui ui+1 , vi vi+1 and
wi wi+1 . We must prove that, given a drawing of G with c crossings, c ≥ s.
If any edge of some Ti contains a crossing, deleting the edges of Ti gives a
drawing of a subdivision of C3 × Cs−1 with at most c − 1 crossings, so by
inductive hypothesis c − 1 ≥ s − 1.
Suppose then that there are no crossings on the edges P of any Ti . Let ce
be the number of crossings on the edge e and note that e∈E(G) ce = 2e.
P i ∈ Zs , let Xi be the set of edges {ui ui+1 , vi vi+1 , wi wi+1 }. We show that
For
e∈Xi ce ≥ 2, which will complete the proof. If two edges of Xi cross each
other this is clear, so suppose they do not. Then our drawing contains a
plane embedding of the subgraph Hi induced by the vertices of the elements
of Xi , which is isomorphic to C3 × K2 . It has five regions, two bounded by
triangles and three by 4-cycles. The triangle Ti+2 must lie in one of these
regions, say R. If R were bounded by Ti , the cycle induced by U would meet
Ti in a point other than ui , which would be a crossing on an edge of Ti . Thus
R is not bounded by Ti , and similarly not by Ti+1 . Without loss of generality,
we may assume that R is bounded by the 4-cycle (ui , vi , vi+1 , ui+1 , ui ). Now
the cycle induced by W must meet this 4-cycle in two points, which are
crossings on edges of Xi .

17 Vertex colorings
A vertex coloring, or just a coloring, of a graph G from a set C of colors is a
function γ : V (G) → C such that for adjacent vertices u and v, γ(u) 6= γ(v).
It is a k-coloring if the image of γ has k elements, and the chromatic index
χ(G) is the minimum k for which G has a k-coloring. G is k-colorable
if χ(G) ≤ G. It is clear that χ(G) is the maximum of χ(H) over the
components H of G.
Lemma 17.1. For any graph G, χ(G) is the maximum of χ(B) over all
blocks B of G.
Proof. We may assume that G is connected. Let k be the maximum over
blocks; clearly k ≤ χ(G). We prove the reverse inequality by induction on
the number of blocks, the case of a single block being trivial. Suppose that
G has more than one block. Let B be an end-block, and let G′ be the union
of the other blocks. Then B and G′ meet in a single cut-vertex u of G.
By induction, B and G′ are k-colorable. We may color them both from the
same set of k colors, so that u receives the same color from B and G′ , and
this gives a k-coloring of G.

68
 A graph has chromatic index 1 iff it is empty, and chromatic index 2 iff
it is non-empty and bipartite. More generally, a k-coloring of G partitions
V (G) into k color classes, the sets of vertices of each color, and G is k-
partite with respect to these sets. Conversely, if G is k-partite with partite
sets U1 , . . . , Uk , we may color G from {1, . . . , k} by setting γ(u) = i for
u in Ui . Some classes of graphs have easily determined chromatic indices:
χ(Kn ) = n, χ(Pn ) = 2 for n > 1, and χ(Cn ) = 2 if n is even and 3 if n is
odd. In general, though, determining χ(G) is hard.
A graph G is k-chromatic if χ(G) = k, and, for k ≥ 2, critically k-
chromatic if in addition χ(G −u) < k (and hence χ(G −u) = k −1) for every
vertex u. By Lemma 17.1, a critically k-chromatic graph is non-separable.
The only critically 2-chromatic graph is K2 . An odd cycle is critically 3-
chromatic. Conversely, if G is critically 3-chromatic, it is not bipartite, and
so contains an odd cycle. Let C be a shortest odd cycle in G. If vertices
of C are not adjacent in C, they are not adjacent in G, since otherwise we
could find two shorter cycles, one of them odd. If G had a vertex not on C,
then G − u, containing an odd cycle, would be 3-chromatic, a contradiction.
Thus G is an odd cycle.

Theorem 17.2. Let G be a critically k-chromatic graph, k ≥ 2. Then G is

edge (k − 1)-connected.

Proof. Certainly G is connected. Let X be an edge-cut of cardinality κ1 (G),

and suppose, for a contradiction that |X| < k − 1. Then G − X has two
components G1 and G2 , which are both (k − 1)-colorable. Choose colorings
of G1 and G2 from some set C of k − 1 colors. Let the color classes of
G1 that contain a vertex adjacent to some vertex of G2 be U1 , . . . , Ur , and
let the number of edges with one vertex in Ui and one in G2 be ki . Then
k1 + · · · + kr ≤ |X|. We re-color G1 from C, keeping the same color classes,
so as to obtain a coloring of G from C, which is the desired contradiction.
The vertices of U1 are adjacent to vertices of G2 having at most k1 different
colors, and since k1 < k − 1 we may choose some other color for the vertices
of U1 . Suppose we have assigned colors to the vertices of Uj for j < i so
that all adjacent vertices so far colored have different colors. The number
of different colors appearing at vertices of all Uj for j < i, or at vertices of
G2 adjacent to a vertex of Ui , is at most i − 1 + ki ≤ k1 + · · · + ki < k − 1,
so we may choose a different color for vertices of Ui .
In this way, all vertices of U1 ∪ · · · ∪ Ur are colored, and we may assign
colors to the remaining color classes of G1 subject only to the requirement
that all classes of G1 get different colors.

69
 From the inequality κ1 (G) ≤ δ(G) of Theorem 6.3 we have the following
corollary.

Corollary 17.3. Let G be a critically k-chromatic graph, k ≥ 2. Then

δ(G) ≥ k − 1.

Theorem 17.4. Let G be a connected graph that is neither complete nor an

odd cycle. Then χ(G) ≤ ∆(G).

Of course, if G is complete or an odd cycle then χ(G) = ∆(G) + 1.

Proof. Let k = χ(G). Then G contains a critically k-chromatic subgraph

H. If H is complete or an odd cycle then H 6= G. Since G is connected and
H is regular, ∆(H) < ∆(G), so k = ∆(H) + 1 ≤ ∆(G). Suppose then that
H is neither complete nor an odd cycle, which implies that k ≥ 4, and so,
by the preceding corollary, δ(H) ≥ 3. Let n be the order of H. We shall
show below that H contains vertices u and v with dH (u, v) = 2 such that
H − {u, v} is connected. Given this, we may number the remaining vertices
as w1 , . . . , wn−2 , where w1 is adjacent to u and v, and, for 1 < i ≤ n − 2,
wi is adjacent to some wj with j < i. We may now color H with the colors
{1, . . . , ∆(H)}, showing k ≤ ∆(H) ≤ ∆(G), as follows. We first give u and
v the color 1. Then, for i = n − 2, . . . , 2 in turn we give wi a color different
from any color already assigned to some vertex adjacent to wi , which is
possible since wi is adjacent to at most ∆(H) vertices, of which at least one
has not yet been colored. Finally, since w1 is adjacent to at most ∆(H)
vertices, two of which have the same color, it may be assigned a color to
complete the coloring.
Since H is not complete (and is connected), it has vertices u and v with
dH (u, v) = 2, and if H − {u, v} is connected we are done. Otherwise, H − u
is connected (since H is 2-connected) and has v as a cut-vertex. Let B1
and B2 be distinct end-blocks of H − u containing cut-vertices w1 and w2
respectively. For i = 1 or 2, since H − wi is connected, there is a vertex xi of
Bi −wi adjacent to u. Then dH (x1 , x2 ) = 2. By Lemma 5.20, H −u−{x1 , x2 }
is connected, and since degH u ≥ 3, so is H − {x1 , x2 }.

Theorem 17.5. For any graph G, χ(G) ≤ max δ(G′ ) + 1, where the maxi-
mum is over all subgraphs G′ of G.

Proof. Let k = χ(G). If k = 1 the result is clear, and otherwise G has a

subgraph G′ that is critically k-chromatic, and k ≤ δ(G′ ) + 1 by Corollary
17.3.

70
 Figure 17: The Grötzsch graph

Theorem 17.6. For any graph G, let l(G) be the maximum length of a path
in G. Then χ(G) ≤ l(G) + 1.
Proof. By the previous theorem there is a subgraph G′ of G with χ(G) ≤
δ(G′ ) + 1, and by Exercise 3.1 or Theorem 7.8, l(G) ≥ l(G′ ) ≥ δ(G′ ).

A graph is triangle-free if, well, it contains no triangles; that is, it is

either acyclic or of girth at least 4.
Theorem 17.7. For any positive integer k there is a triangle free, k-chromatic
graph.
Much more is true; there are graphs of any chromatic index of arbitrarily
high girth, but we do not prove this.

Proof. If G is a graph of order n with vertices u1 , . . . , un , form a graph G′ of

order 2n+1 by adding new vertices v1 , . . . vn and w, and edges vi uj whenever
ui and uj are adjacent, and vi w for 1 ≤ i ≤ n. If G is triangle-free, so is G′ .
Given a k-coloring of G, we get a (k + 1)-coloring of G′ by giving each vi
the same color as ui , and w a (k + 1)’st color. On the other hand, given a
(k + 1)-coloring of G′ , let c be the color of w. Color G by giving ui its color
in G′ if this is not c, and the color of vi if it is. This produces a coloring of
G with at most k colors. Thus χ(G′ ) = χ(G) + 1, and the result follows by
induction on k.

If the construction of the above proof is applied to K1 (a triangle-free

1-chromatic graph), the result is K1 ⊔ K2 , which is indeed triangle-free and
2-chromatic, but so is K2 . If the construction is applied to K2 the result is
C5 , which is clearly the unique triangle-free 3-chromatic graph of minimum
order. If the construction is applied to C5 , the result is the Grötzsch graph of
order 11, shown in Figure 17. We verify that this is the unique triangle-free
4-chromatic graph of minimum order.

71
Proposition 17.8. Let G be a triangle-free graph of order n ≤ 11 with
χ(G) ≥ 4. Then G is isomorphic to the Grötzsch graph.

Proof. We assume that n is minimal, and that amongst triangle free graphs
of order n with chromatic index at least 4, G is maximal. If we prove that
G is isomorphic to the Grötzsch graph, the result will follow since deleting
any edge of the Grötzsch graph leaves a 3-chromatic graph. Since adding
edges does not decrease the chromatic index, the maximality requirement
means that G has diameter 2.
By Theorem 17.4, ∆(G) ≥ 4. Let w be a vertex of maximum degree,
V the set of its neighbors, and U the set of all other vertices of G. Then
|U | ≤ 6. The subgraph induced by U is not bipartite, since otherwise we
could 2-color it, give the vertices of V a third color, and w one of the original
2 colors to obtain a 3-coloring of G. Hence there is a 5-cycle C with vertices
in U , and at most one vertex x of U not on C. Any vertex v ∈ V is a
adjacent to at most two vertices of C. Any vertex u of C not adjacent to v
has a common neighbor with v, which must be one of its neighbors in C, or
x. It follows that v is adjacent to two vertices of C. Moreover, distinct v1
and v2 in V cannot be adjacent to the same two vertices of C, since then the
set of neigbors of one of them, say v1 , would be a subset of the neighbors of
the other. A 3-coloring of G − v1 could then be extended to a 3-coloring of
G by giving v1 the same color as v2 .
If ∆(G) = 5, it follows that G is isomorphic to the Grötzsch graph.
Otherwise, let G′ be G − x if x exists, and G if not. Then G′ is isomorphic
to the Grötzsch graph with some vi (in the notation of the proof of Theorem
17.7) deleted. Hence there is a 3-coloring of G′ in which some color appears
only at two vertices of degree 4 in G′ . Since x, if it exists, cannot be
adjacent to these vertices, it can be given the same color to 3-color G, a
contradiction.

18 The Four-color Theorem

Theorem 18.1. Every planar graph is 4-colorable.

The original statement of this theorem was in terms of coloring regions

of a plane graph. A region coloring of a plane graph G is a function γ from
the set of regions of G to a set C of colors such that, for each arrow (u, v)
of G, the regions to the left and right of (u, v) get different colors. (This
form of the definition means that if G has a bridge then it has no region
colorings.) It is a region k-coloring if the image of γ has k elements, and the

72
region chromatic index χ∗ (G) is the minimum k for which G has a region
k-coloring. G is region k-colorable if χ∗ (G) ≤ k.

Proposition 18.2. Let k be a positive integer. The following are equivalent.

(1) Every bridgeless cubic plane graph is region k-colorable.

(2) Every planar graph is k-colorable.

(3) Every bridgeless plane graph is region k-colorable.

The basic idea here is that a coloring of the vertices of a plane graph
corresponds to a coloring of the regions of its dual. The dual of a plane graph
G is obtained by taking a vertex in every region and taking, for each edge
e of G, an edge in the dual adjacent to the vertices in the adjoining regions
and crossing e at one point. However, the dual is generally a multigraph,
and we are avoiding multigraphs in our formal treatment. (Loops in the
dual correspond to bridges in G, and parallel edges arise when regions of G
have more than one edge in common.) Therefore our proof will avoid taking
the dual of a general plane graph.

Proof. Suppose first that every bridgeless cubic planar graph is region k-
colorable. It is enough to prove that every maximal planar graph G is
k-colorable. Take a plane embedding of G. Then the dual G∗ of G is a
graph, is bridgeless and cubic, and has dual G, so a region coloring of G∗
gives a vertex coloring of G using the same colors.
Suppose now that every planar graph is k-colorable, and let G be a
bridgeless plane graph, which we may assume is connected. Form a plane
graph H as follows. Take a vertex in every region of G. For any regions
of G with an edge in common, pick one such, e, and take an edge in H
whose vertices are those in the regions and which crosses e at one point.
A coloring of the vertices of H gives a region-coloring of G using the same
colors. (Here the dual of G has no loops since G is bridgeless, and H is
obtained by dropping all but one of each set of parallel edges.)

For the early history of the four-color theorem, see [4]. A fallacious
proof was published by Kempe in 1879. The error was only pointed out by
Heawood in 1890. The four-color theorem was then reduced to the status
of a conjecture, only becoming a theorem again in 1976 (Appel, Haken and
Koch [1]). Even the simplified proof of Robertson, Sanders, Seymour and
Thomas [27] relies heavily on the use of a computer, and no more will be
said about it here. We give Kempe’s incorrect proof, Heawood’s observation

73
 y b

b r r y
g g

r y b g y b r g
r y
g y g b

b r

Figure 18: Heawood’s counterexample

of the flaw, and his modification of Kempe’s method to prove a weaker

theorem. We shall use the vertex-coloring version, rather than the dual
region-colorings of the originals.

Failed proof of Theorem 18.1. We attempt to use induction on the order n

of a planar graph G, the cases n ≤ 4 being obvious. Suppose then that n ≥ 5
and the result is true for graphs of order less than n. We may assume that
G is maximal planar, and we take a plane embedding of G. By Corollary
14.6, G has a vertex u of degree at most 5. The planar graph G − u may be
colored using (at most) four colors, which we take to be red, green, blue and
yellow (r, g, b and y). If some color does not appear at any neighbor of u,
we get a 4-coloring of G by giving u this color. One possibility is that u has
degree 4, with neigbors v1 , v2 , v3 and v4 in counterclockwise order, having
colors r, g, b and y respectively. For any two colors, we may consider the
subgraph of G − u induced by the vertices of those colors. A component of
such a graph is a Kempe chain, and an rg-chain if the colors r and g were
used. We may switch the two colors at the vertices of any Kempe chain
to produce another coloring of G − u. If the rb-chain containing v1 does
not contain v3 , then switching red and blue on this chain gives a coloring
of G − u in which red does not appear at a neighbor of u, and hence a 4-
coloring of G. If this chain does contain v3 , it contains a v1 -v3 path, which
together with (v3 , u, v1 ) gives a cycle in G separating v2 and v4 . Hence the
gy-chain containing v2 does not contain v4 , and we may use this chain to
get a 4-coloring of G.

74
 It remains to consider the case where u has degree 5, with neighbors
v1 , . . . , v5 in counterclockwise order. By maximality, (v1 , . . . , v5 , v1 ) is a
cycle, and so without loss of generality the colors of v1 , . . . , v5 are b, r, y, g,
r respectively. If the by-chain containining v1 does not contain v3 , we are
done as before, so we assume it does. Hence the rg-chain containing v2 does
not contain v4 . Similarly we may assume that the bg-chain containing v1
contains v4 , and so the ry-chain containing v5 does not contain v3 . Kempe
thought that these two chains could be used to get a coloring of G − u
in which red does not appear at a neighbor of u, completing the proof.
Heawood pointed out that switching colors on one chain can alter the other,
so that the new chain contains two neighbors of u, and gave the example
shown in Figure 18. Here G − u and its coloring are shown; u is in the
unbounded region. It will be seen that the relevant by- and bg-chains have
the properties stated, and changing colors on the rg- and ry-chains, in either
order, leaves a coloring with all four colors at neighbors of u.

Theorem 18.3 (Heawood). Every planar graph is 5-colorable.

Proof. We use induction on the order n of a planar graph G, the cases n ≤ 5

being obvious. Suppose then that n ≥ 6 and the result is true for graphs of
order less than n. Take a plane embedding of G. By Corollary 14.6, G has a
vertex u of degree at most 5. The planar graph G − u may be colored using
(at most) five colors, which we take to be red, green, blue, yellow and taupe
(r, g, b, y and t). If some color does not appear at any neighbor of u, we
get a 5-coloring of G by giving u this color. Otherwise u has degree 5, with
neigbors v1 , v2 , v3 , v4 and v5 in counterclockwise order, having colors, say,
r, g, b, y and t respectively. Either the rb-chain containining v1 does not
contain v3 , or the by-chain containing v2 does not contain v4 , and in either
case we get a coloring of G − u with a missing color at the neighbors of u,
and hence a coloring of G.

19 Edge colorings
An edge coloring of a graph G from a set C of colors is a function γ : E(G) →
C such that for adjacent edges e and f , γ(e) 6= γ(f ). It is an edge k-coloring
if the image of γ has k elements, and the edge chromatic index χ1 (G) is
the minimum k for which G has an edge k-coloring. G is edge k-colorable if
χ1 (G) ≤ G. Clearly ∆(G) ≤ χ1 (G).

Theorem 19.1 (Vizing). For any graph G, χ1 (G) ≤ ∆(G) + 1.

75
Proof. This is by induction on the size, the case of an empty graph being
obvious. Suppose that G is non-empty and the result holds for any graph
with fewer edges. Let C be some fixed set of ∆(G) + 1 colors. In this proof,
an edge coloring from C will be called just a coloring. Let e = uv0 be an
edge of G, and set H = G − e. Since ∆(H) ≤ ∆(G), H has a coloring. For
any coloring γ of H and vertex x of G, we say that a color c ∈ C appears at x
if there is an edge f of H incident with x such that γ(f ) = c. At any vertex,
there is at least one color that does not appear. Suppose that, for some
coloring γ of H and integer k ≥ 0, there are distinct vertices v0 , v1 , . . . , vk
adjacent to u such that
(1) for 1 ≤ i ≤ k, the color ai = γ(uvi ) does not appear at vi−1 ;
(2) there is a color b that appears at neither u nor vk .
Then we obtain a coloring of G by giving uvi−1 the color ai for 1 ≤ i ≤ k,
uvk the color b, and every other edge its color under γ.
Now let γ be an arbitrary coloring of H, and suppose we cannot find
vertices as above. Let k be the maximum integer for which there exist
distinct vertices v0 , v1 , . . . , vk adjacent to u satisfying (1). (There is such
an integer, because (1) is vacuously true if k = 0.) There is some color not
appearing at vk , which must appear at u, and by maximality of k it must
be aj for some j with 0 < j < k. Let b be a color not appearing at u, which
must appear at every vi for 0 ≤ i ≤ k. Consider the subgraph K of H
induced by the edges colored aj or b. Every component is either a cycle or
a path, and the endpoints of the paths are the vertices at which just one of
aj and b appears. One such vertex is vk ; let the component of K containing
it be P , and let w be the other endpoint of P . Define a new coloring γ ′ of
H by switching aj and b on the edges of P . Note that at every vertex other
than vk and w the same colors appear in γ and γ ′ . We claim that there is
some k ′ , 0 ≤ k ′ ≤ k, such that, with respect to γ ′ ,
(1′ ) for 1 ≤ i ≤ k ′ , the color a′i = γ ′ (uvi ) does not appear at vi−1 ;
(2′ ) there is a color b′ that appears at neither u nor vk′ ,
which will complete the proof. Note that a′i = ai except perhaps for i = j.
Suppose first that w = u, so the vertex adjacent to w in P is vj . Then we
may take k ′ = j − 1 and b′ = aj . If w 6= u then we also have a′j = aj .
Suppose that w = vl , where 0 ≤ l < k and l 6= j. Since b appears at vl in
γ, it does not in γ ′ and we may take k ′ = l and b′ = b. The only remaining
possibility is that w is different from u and all vi , 0 ≤ i ≤ k, and then we
may take k ′ = k and b′ = b.

76
 A graph G is of class 1 if κ1 (G) = ∆(G), and of class 2 if κ1 (G) =
∆(G) + 1.
Theorem 19.2. A regular graph of class 1 has even order.
Proof. Suppose G is d-regular of class 1. In any edge d-coloring of G, every
color appears at every vertex. Thus the edges with any given color span G.
Since they are pairwise non-adjacent, the result follows.

The converse is false; see Exercise 19.1. The existence of edge 3-colorings
of cubic graphs is related to the four-color theorem by the following result.
Theorem 19.3 (Tait). Let G be a connected, bridgeless, cubic plane graph.
Then G is edge 3-colorable iff it is region 4-colorable.
Proof. We color regions using Z2 ×Z2 and edges using the non-zero elements.
Given a region coloring, we color each edge with the sum of the colors of the
adjacent regions, which is non-zero because adjacent regions have different
colors. Since the three regions around a vertex have different colors, so do
the three edges.
Conversely, given an edge coloring, give one region the color (0, 0). Then
the color of every other region is determined by the rule that the sum of
the colors of adjacent regions is the color of their common edge. This is
well-defined because at every vertex the colors of the edges are the three
non-zero elements of Z2 × Z2 , whose sum is zero.

By Proposition 18.2, it follows that the four-color theorem is equivalent

to the edge 3-colorability of all bridgeless, cubic, planar graphs. Martin
Gardner [14] gave the name snark to a connected, bridgeless cubic graph
that has no edge 3-coloring (i.e., is of class 2), and called a planar snark
a boojum. (The names are taken from Lewis Carroll’s “The Hunting of
the Snark”.) In these terms, the four-color theorem is equivalent to the
non-existence of a boojum. Actually, snarks are required to satisfy extra
conditions, such as being edge 4-connected and of girth at least 5, to rule
out examples that are easily constructed from simpler ones. The first snark
discovered was the Petersen graph (Exercise 19.1). Another is shown in
Figure 19.
Exercises for §19
19.1. Show that the Petersen graph is a snark.

19.2. Let n be odd. Show that in any edge n-coloring of Kn , each color
appears (n − 1)/2 times.

77
 Figure 19: The flower snark

19.3. Show that Kn is of class 1 if n is even.

19.4. Show that a Hamiltonian cubic graph is not a snark.

20 Strongly regular graphs

Definition 20.1. A graph G is strongly regular with parameters (n, d, λ, µ)
if it is connected of order n, is d-regular where d < n − 1, any two adjacent
vertices have λ common neighbors, and any two independent vertices have
µ common neighbors.

The condition that d < n − 1 just rules out complete graphs, and the
requirement that G be connected rules out graphs whose components are all
complete of the same order. We have 0 ≤ λ < d, 0 < µ ≤ d and diam G = 2.

Theorem 20.2. Let G be strongly regular with parameters (n, d, λ, µ). Then
d(d − λ − 1) = (n − d − 1)µ.

Proof. For any vertex v, the number of walks of length 2 starting at v is d2 .

Of these, d end at v, and dλ end at a vertex adjacent to v, so d(d − λ − 1)
end at a vertex independent of v. On the other hand, there are n − d − 1
vertices independent of v, and µ walks of length 2 from v to each of them.
The result follows.

Lemma 20.3. Let G be strongly regular with parameters (n, d, λ, µ) and

adjacency matrix A. Then A2 + (µ − λ)A − (d − µ)I = µJ, where I is the
n × n identity matrix, and J is the n × n matrix with all entries 1.

Proof. By Theorem 4.2, A2 counts walks of length 2 in G. The number of

u-v walks of length 2 is d if u = v, λ if u and v are adjacent, and µ otherwise,
so A2 = dI + λA + µ(J − I − A), as required.

78
Theorem 20.4. Let G be strongly regular with parameters (n, d, λ, µ). Then
!
1 (n − 1)(µ − λ) − 2d
2 n−1±
p
(µ − λ)2 + 4(d − µ)
are positive integers.
Proof. By Lemma 20.3, every eigenvector of A is an eigenvector of J. Now
J has a 1-dimensional eigenspace with eigenvalue n, generated by x1 =
(1, 1, . . . , 1), and an (n − 1)-dimensional eigenspace with eigenvalue 0. The
vector x1 is an eigenvector of A, with eigenvalue d. The eigenvalue of any
independent eigenvector of A must be a root of x2 + (µ − λ)x − (d − µ). The
roots of this quadratic are
 p 
r± = 21 −µ + λ ± (µ − λ)2 + 4(d − µ) .

Let the multiplicity of r± as an eigenvalue of A be a± . By Theorem 4.4,

A must have three distinct eigenvalues, so a+ and a− are positive integers.
Then a+ + a− = n − 1 and, since A has trace 0, d + r+ a+ + r− a− = 0. These
equations give
!
(n − 1)(µ − λ) − 2d
a± = 21 n − 1 ± p .
(µ − λ)2 + 4(d − µ)
Theorem 20.5. Let G be strongly regular with parameters (n, d, λ, µ). Then
either (µ − λ)2 + 4(d − µ) is a perfect square, or λ = µ − 1, d = 2µ and
n = 4µ + 1.
Proof. By Theorem 20.4, if (µ − λ)2 + 4(d − µ) is not a perfect square then
(n − 1)(µ − λ) = 2d. Since 0 < d < n − 1, µ − λ = 1 and n = 2d + 1.
Combining these equations with Theorem 20.2 gives the result.

Theorem 20.6. Let G be a connected d-regular graph of order n whose

adjacency matrix A has exactly three distinct eigenvalues. Then G is strongly
regular.
Proof. By Theorem 4.3, A has d as an eigenvalue of multiplicity 1. Let the
sum and product of the other two eigenvalues be a and b. Then the minimal
polynomial of A is (x−d)q(x), where q(x) = x2 −ax+b. Now Aq(A) = dq(A),
so every column of q(A) is an eigenvector of A with eigenvalue d, and all
its entries are equal. Since each diagonal entry of A2 , A or I is d, 0 or 1,
respectively, every diagonal entry of q(A) is d + b, and so q(A) = (d + b)J.
That is, A2 = aA − bI + (d + b)J, so G is strongly regular with parameters
(n, d, d + a + b, d + b).

79
Exercises for §20

20.1. Let G be strongly regular with parameters (n, d, λ, µ) where µ < d.

Show that Ḡ is strongly regular with parameters
¯ d¯ − d + µ − 1, d¯ − d + λ + 1),
(n, d,

where d¯ = n−d−1. (The strongly regular graphs with µ = d are com-

plete multipartite graphs of the form Kr,r,...,r , and have disconnected
complements.)

20.2. Find a strongly regular graph with parameters (9, 4, 1, 2).

21 Moore graphs
Theorem 21.1. Let G be a connected graph of order n, diameter k and
Pk−1
maximum degree ∆. Then n ≤ 1 + ∆ i=0 (∆ − 1)i .

Proof. Fix a vertex v0 , and for 0 ≤ r ≤ k let Sr be the set of vertices v

with d(v0 , v) = r, so that V (G) = S0 ∪ S1 ∪ · · · ∪ Sk . We have S0 = {v0 }
and |S1 | ≤ ∆. For 0 < r < k, every vertex of Sr is adjacent to a vertex
of Sr−1 and is therefore adjacent to at most ∆ − 1 vertices of Sr+1 , so
|Sr+1 | ≤ (∆ − 1)|Sr |. Hence |Sr | ≤ ∆(∆ − 1)r−1 for 0 < r ≤ k, and the
result follows.

Definition 21.2. A Moore graph is a graph for which equality holds in the
inequality of Theorem 21.1.

This is the original definition of a Moore graph. The term has since
been extended to a larger class of graphs, as will be discussed later. If G
is a Moore graph, every inequality in the above proof must be an equality.
This implies that G is regular, rad G = diam G, and that for every vertex
v0 the subgraph induced by the edges of all geodesics starting at v0 is a tree
T , and the vertices of an edge not in T are end-vertices of T . The graphs of
diameter less than 2 are the complete graphs, and these are Moore graphs.
The 2-regular Moore graphs are the odd cycles. Recall that the girth of a
graph that is not a forest is the length of a shortest cycle. Clearly the girth
of a Moore graph of diameter k is 2k + 1.

Theorem 21.3. Let G be a connected graph of order n, girth 2k + 1 and

Pk−1
minimum degree δ. Then n ≥ 1 + δ i=0 (δ − 1)i .

80
Proof. As before, fix a vertex v0 and let Sr be the set of vertices v with
d(v0 , v) = r. Then S0 = {v0 } and |S1 | ≥ δ. For 0 < r ≤ k, every vertex of
Sr is adjacent to only one vertex of Sr−1 , since otherwise there would be a
cycle of length at most 2r (by Lemma 7.5). If there are adjacent vertices of
Sr then there is a cycle of length at most 2r + 1, so r ≥ k. It follows that
|Sr | ≥ δ(δ − 1)r−1 for 0 < r ≤ k.

Theorem 21.4. Let G be aPconnected graph of order n, girth 2k and mini-

k−1
mum degree δ. Then n ≥ 2 i=0 (δ − 1)i .
Proof. This time, fix adjacent vertices v0 and v1 , and let Sr be the set of
vertices v for which the minimum of d(v0 , v) and d(v1 , v) is r, so S0 =
{v0 , v1 }. For 0 < r < k, every vertex of Sr is adjacent to only one vertex
of Sr−1 , since otherwise there would be a cycle of length at most 2r + 1.
If there are adjacent vertices of Sr then there is a cycle of length at most
2r + 2, so r ≥ k − 1. It follows that |Sr | ≥ 2(δ − 1)r for 0 ≤ r < k, giving
the result.

If equality holds in Theorem 21.3, then G is regular of diameter k, and

therefore a Moore graph. If equality holds in Theorem 21.4 then G is also
regular of diameter k. Some authors use the term Moore graph to mean
a graph for which equality holds in Theorem 21.3 or 21.4, as appropriate.
We shall keep the original meaning, and use generalized Moore graph for the
broader class. The generalized Moore graphs of girth 4 are the complete
bipartitite graphs Kr,r , and the 2-regular generalized Moore graphs of even
girth are the even cycles. It can be shown that, for any integers d ≥ 2 and
g ≥ 3, there is a d-regular graph of girth g. Such a graph of minimum order is
called a (d, g)-cage. Thus every generalized Moore graph is a cage; however,
there are many pairs (d, g) for which a (d, g)-cage is not a generalized Moore
graph.
If G is d-regular
Pk−1 of diameter k and girth
Pk−12k + 1, ithen Theorems 21.1 and
i
21.3 give 1+d i=0 (d−1) ≤ n ≤ 1+d i=0 (d−1) , so G is a Moore graph.
The hypothesis of regularity is superfluous, as was shown by Singleton [28].
Theorem 21.5. Suppose that G is a graph of diameter k and girth 2k + 1.
Then G is regular.
Proof. By Lemma 7.5, for vertices u and v there is a unique u-v path of
length d(u, v), and any other u-v path has length at least 2k + 1 − d(u, v),
with equality only if the two paths give a cycle of length 2k + 1. We say that
u and v are antipodes if d(u, v) = k. Suppose this is so, and let P be the
u-v path of length k. Any neighbor w of u not on P must also have distance

81
k from v, and P , a v-w path of length k and the edge wu give a (2k + 1)-
cycle. Hence the number of (2k + 1)-cycles through u and v is deg u − 1, and
by symmetry in u and v, deg u = deg v. Suppose C = (v0 , v1 , . . . , v2k , v0 )
is a (2k + 1)-cycle, and take the subscripts modulo 2k + 1. The sequence
(v0 , vk , v2k , . . . , v(2k)k ) contains all vertices of C, and consecutive terms are
antipodes, so all vertices of C have the same degree.
Now fix a (2k + 1)-cycle C. For any vertex v, we may construct a path
of length k starting at v by taking a shortest path from v to a vertex of C,
followed by a path in C. Hence v has an antipode. It follows that no vertex
has degree 1, since its neighbor would not have an antipode. Hence any
path of length l < 2k can be extended to a path of length l + 1. This implies
that, for any vertices u and v there is some path of length k containing u
and v, and this is contained in a (2k + 1)-cycle. Thus deg u = deg v.

Theorem 21.6 (Hoffman-Singleton [19]). If there exists a d-regular Moore

graph of diameter 2, then d = 2, 3, 7 or 57.

As for the existence of Moore graphs of diameter 2, the following is

known. Obviously the only one of degree 2 is the pentagon. The Petersen
graph is an example of degree 3, and it is not hard to see that it is unique.
For degree 7, there is also a unique example, as we show below; this was
also proved in [19]. Whether there is an example of degree 57 is unknown.

Proof. If G is such a graph, it is strongly regular with parameters (d2 +

1, d, 0, 1). By Theorem√20.5, either 4d − 3 is a perfect square or d = 2.
Suppose then that s = 4d − 3 is an integer. By Theorem 20.4, s divides
1
d(d−2) = 16 (s2 +3)(s2 −5). That is, for some integer a, s4 −2s2 −16as−15 =
0, so s is a divisor of 15. If s = 1 then d = 1, which is impossible, while if
s = 3, 5 or 15 then d = 3, 7 or 57.

Example 21.7 (The Hoffman-Singleton graph). This is a 7-regular Moore

graph G of diameter 2, and hence order 50. The vertices are xi,j and yi,j
for i, j ∈ Z5 . The edges are xi,j xi,j+1 , yi,j yi,j+2 and xi,k yj,ij+k for i, j and
k ∈ Z5 . It is easy to check that G is 7-regular of diameter 2, as required.
This graph is shown Figure 20. (This drawing can be constructed as follows.
The vertices lie on a circle centered at the origin. To determine their angular
coordinates, let a = (0, 2, 24, 28, 6) and b = (1, 33, 45, 49, 37). The angle for
xi,j is π(ai + 20j)/25, and that for yi,j is π(bi + 20j)/25.)

Theorem 21.8. Let G be a Moore graph of degree 7 and diameter 2. Then

G is isomorphic to the graph of Example 21.7.

82
 Figure 20: The Hoffman-Singleton graph

The following proof is taken from Elkies [10].

Proof. In this proof we shall distinguish between a 5-cycle, which is a se-

quence of vertices, and a pentagon, the underlying graph of a 5-cycle. It is
clear that in any graph the number of 5-cycles is 10 times the number of
pentagons. All subscripts will be in Z5 .
As remarked in the previous proof, G is strongly regular with parameters
(50, 7, 0, 1). If P = (u0 , u1 , u2 , u3 ) is a path of length 3 then u0 and u3
are non-adjacent, so they have a unique common neighbor u4 , and there
is a unique 5-cycle (u0 , u1 , u2 , u3 , u4 , u0 ) starting with P . There are thus
50 · 7 · 6 · 6 = 12600 5-cycles, and hence 1260 pentagons, in G.
Let C = (u0 , u1 , u2 , u3 , u4 , u0 ) be a 5-cycle, and let Ai be the set of
vertices adjacent to ui other than ui±1 . If i 6= j ∈ Z5 then either ui and
uj have no common neighbors or they have a common neighbor in C, so
Ai and Aj are disjoint. Setting A = A0 ∪ · · · ∪ A4 and B = V (G) − A,
|A| = |B| = 25. A vertex of Ai can be adjacent to a vertex of Aj only if
i = j ± 2, and any vertex not on C can be adjacent to at most one vertex in
each Ai . If v ∈ Ai and j = i ± 2, v and uj have a common neighbor, which
must be in Aj . It follows that each vertex of A is adjacent to 2 vertices of
A and 5 of B. There are 125 edges from a vertex of A to a vertex of B, and
since each vertex v of B is adjacent to at most 5 vertices of A (those of Ai
if v = ui and at most one from each Ai if v is not on C), v is adjacent to 5

83
vertices of A and 2 of B.
Now consider a 5-cycle (v0 , v1 , v2 , v3 , v4 , v0 ). There are 25 · 5 · 4 · 4 = 2000
such cycles with v0 , v2 ∈ A and v1 , v3 ∈ B, and for each one there is
just one other cycle of the same form giving the same pentagon, namely
(v4 , v3 , v2 , v1 , v0 , v4 ) if v4 ∈ A and (v2 , v1 , v0 , v4 , v3 , v2 ) if v4 ∈ B, so these
cycles account for 1000 of the pentagons. There are 25 · 2 · 5 · 2 = 500 cycles
with v0 , v1 ∈ A and v2 , v3 ∈ B, and again for each one there is one other cycle
of the same form giving the same pentagon, namely (v0 , v4 , v3 , v2 , v1 , v0 ) if
v4 ∈ A and (v1 , v0 , v4 , v3 , v3 , v1 ) if v4 ∈ B. These cycles give a further 250
pentagons, so there are 10 pentagons not yet accounted for. These are given
by cycles in which v0 , v1 , v2 and v3 belong to the same one of A and B, and
there are 50 · 2 · 1 · 1 = 100 of these. If v4 is not in the same set, there is only
one other cycle of this form giving the same pentagon, (v3 , v2 , v1 , v0 , v4 , v3 ),
which gives too many pentagons. Hence A is the disjoint union of pentagons
P0 , . . . , P4 , and B of pentagons Q0 , . . . , Q4 .
Each vertex of Pi is adjacent to one vertex of every Qj , and vice-versa.
We shall label the vertices of Pi as xi,j and those of Qi as yi,j (i, j ∈ Z5 ),
in such a way that the edges are xi,j xi,j+1 and yi,j yi,j+2 . However this is
done, there are, for i, j ∈ Z5 , elements a(i, j) and b(i, j) = ±1 of Z5 such
that xi,k is adjacent to yj,a(i,j)+b(i,j)k . We may choose the labelling of the
x0,k arbitrarily, then that of the yj,l so that a(0, j) = 0 and b(0, j) = 1,
then that of the remaining xi,k so that a(i, 0) = 0 and b(i, 0) = 1. Suppose
that there exist i and j with b(i, j) = −1. Then for any k ∈ Z5 there is a
trail (xi,k , yj,a(i,j)−k , x0,a(i,j)−k , y0,a(i,j)−k , xi,a(i,j)−k) ) of length 4. However,
if k = 3a(i, j) this is a cycle, which is impossible. Thus xi,k is adjacent to
yj,a(i,j)+k . Since there are no 4-cycles, for i1 6= i2 and j1 6= j2 we have

(21.1) a(i1 , j1 ) − a(i2 , j1 ) + a(i2 , j2 ) − a(i1 , j2 ) 6= 0.

In particular,

a(i, j1 ) 6= a(i, j2 ) for i 6= 0 and j1 6= j2 ;

(21.2)
a(i1 , j) 6= a(i2 , j) for i1 6= i2 and j 6= 0.

We may re-number the Pi and Qj so that, in addition to a(i, 0) = a(0, j) = 0,

we have a(i, 1) = i and a(1, j) = j. All values of a(i, j) are now determined.
We have a(2, 3) 6= 0, 2 or 3 by (21.2), and taking i1 = j1 = 1, i2 = 2 and
j2 = 3 in (21.1) gives a(2, 3) 6= 4, so a(2, 3) = 1. Similarly a(3, 2) = 1, and
now (21.2) determines the remaining values. We find a(i, j) = ij, so G is
isomorphic to the Hoffman-Singleton graph.

84
 Following [19], we shall show that the only Moore graph of diameter 3
is the heptagon C7 . (It was subsequently proved by Bannai and Ito [2] and
Damerell [8] that all Moore graphs of diameter greater than 2 are odd cycles,
so we have expended a lot of effort determining the properties of the graphs
in a small, and, with the exceptions of the Petersen and Hoffman-Singleton
graphs, boring class. Such is life.)
For integers d ≥ 2 and r ≥ 0, define polynomials Fd,r ∈ Z[x] by

1
 if r = 0;
Fd,r = x+1 if r = 1;

xFd,r−1 − (d − 1)Fd,r−2 if r ≥ 2.


Fd,r is monic of degree r.

Lemma 21.9. If G is a d-regular Moore graph of diameter k with adjacency

matrix A then Fd,k (A) = J.

Proof. The discussion after Definition 21.2 shows that if u and v are vertices
with d(u, v) = r > 0 then u has one neighbor at distance r − 1 from v, and if
r < k its other neighbors are at distance r + 1 from v. We show by induction
that, for 0 ≤ r ≤ k, the entry of Fd,r (A) corresponding to vertices u and v
is 1 if d(u, v) ≤ r and 0 otherwise, the cases r = 0 and 1 being obvious, and
the case r = k giving the lemma. Suppose that r ≥ 2 and the result is true
for smaller values. The entry of AFd,r−1 (A) corresponding to u and v is the
number of neighbors of u whose distance from v is at most r − 1, which is d
if d(u, v) ≤ r − 2; 1 if d(u, v) = r − 1 or r; and 0 if d(u, v) > r. The result
follows.

Theorem 21.10. For integers d ≥ 3 and k ≥ 2, if the polynomial Fd,k

is irreducible (over Z, or equivalently over Q), there is no d-regular Moore
graph of diameter k.

Proof. Suppose there is such a Moore graph G with adjacency matrix A.

Then A has d as an eigenvalue of multiplicity 1, and every other eigenvalue
is a root of Fd,k . The minimal polynomial of A has the form (x − d)f , where
f ∈ Z[x] divides Fd,k . If Fd,k is irreducible, f = Fd,k and the characteristic
a for some positive integer a. The coefficient of
polynomial of A is (x − d)Fd,k
xk−1 in Fd,k is 1, so the sum of the roots of Fd,k is −1. Since A has trace zero
the sum of its eigenvalues (with multiplicities)
Pk−1 is zero, so a = d. Therefore
1 + kd is the order of G, which is 1 + d i=0 (d − 1)i , and this is impossible
for d ≥ 3 and k ≥ 2.

85
Corollary 21.11. There is no d-regular Moore graph of diameter 3 for
d ≥ 3.
Proof. We have Fd,3 = x3 + x2 − 2(d − 1)x − (d − 1). If this polynomial
2 (r+1)
is reducible, it has an integral root r. But then d − 1 = r 2r+1 , and since
2r + 1 is coprime to r and r + 1, this implies 2r + 1 = ±1 and d = 1, a
contradiction.

Exercises for §21

21.1. Show that the automorphism group of the Hoffman-Singleton graph
has order 252000.

22 Kneser graphs
Let n and r be integers with 0 < r < n and let X be the
set {1, . . . , n}.
We denote the collection of all r-element subsets of X by Xr . The Kneser
graph Kn(n, r) has vertex set Xr , and two vertices A and B are adjacent

iff they are disjoint. In the proof of Theorem 10.6 we showed that the
Petersen graph is isomorphic to Kn(5, 2). If s is an integer with
0 ≤ s < r,
the generalized Kneser graph Kn(n, r, s) also has vertex set Xr , and two
vertices A and B are adjacent iff |A ∩ B| ≤ s. Thus Kn(n, r) = Kn(n, r, 0).
If n < 2r − s then Kn(n, r, s) is empty, while if s = 0 and n = 2r then every
component is a K2 , so for the rest of this section we exclude these cases.
We shall always use t to denote n − 2(r − s); writing t as (n − (2r − s))
+ s
makes it clear that our restrictions imply that t > 0. Sending A ∈ Xr to
X − A gives an isomorphism from Kn(n, r, s) to Kn(n, n − r, n − 2r + s), so
we could restrict attention to the cases n ≥ 2r, but this does not seem to
simplify any arguments significantly.
Lemma 22.1. Let A, B ∈ Xr . Then A and B have a common neighbor

in Kn(n, r, s) iff |A ∩ B| ≥ r − t.
Proof. Let k = |A ∩ B|. Suppose first that there is a common neighbor C
of A and B. Then |(A ∪ B) ∩ C| ≤ 2s, so

n − 2r + k = |X − (A ∪ B)|
≥ |C − (A ∪ B)|
≥ r − 2s,

or k ≥ 3r − n − 2s = r − t. Now suppose that k ≥ r − t. It suffices to prove

that there is a set Y ⊆ X with |Y | ≥ r, |A ∩ Y | ≤ s and |B ∩ Y | ≤ s. If

86
s ≤ r −k, we may take sets A′ ⊆ A−B and B ′ ⊆ B −A with |A′ | = |B ′ | = s,
and set Y = A′ ∪ B ′ ∪ (X − (A ∪ B)). We have |A ∩ Y | = |B ∩ Y | = s and
|Y | = 2s + n − 2r + k ≥ r. If s > r − k, take Z ⊆ A ∩ B with |Z| = r − s and
set Y = X − Z. Then |A ∩ Y | = |B ∩ Y | = s and |Y | = n − (r − s) ≥ r.

Lemma 22.2. Let A, B ∈ Xr and let p ≥ 0 be an integer. There is an

A-B walk of length 2p in Kn(n, r, s) iff |A ∩ B| ≥ r − pt.

Proof. The case p = 0 is trivial. For p ≥ 1 we proceed by induction on p,

the case p = 1 being Lemma 22.1. Suppose there is an A-B walk of length
2(p + 1). Then there is a vertex C and A-C and B-C walks of lengths 2p
and 2, respectively. By induction, |A ∩ C| ≥ r − pt and |B ∩ C| ≥ r − t. Now

|A ∩ B| ≥ |A ∩ B ∩ C|
= |A ∩ C| + |B ∩ C| − |(A ∩ C) ∪ (B ∩ C)|
≥ (r − pt) + (r − t) − r
= r − (p + 1)t.

Suppose conversely that |A ∩ B| ≥ r − (p + 1)t. We show there is a vertex

C with |A ∩ C| ≥ r − pt and |B ∩ C| ≥ r − t; it will follow by induction
that there are A-C and B-C walks of lengths 2p and 2, and hence an A-
B walk of length 2(p + 1). If |A ∩ B| ≥ r − pt we may take C = B, so
suppose |A ∩ B| < r − pt. It suffices to show that there is a set Y ⊆ X
with |Y | ≤ r, |A ∩ Y | ≥ r − pt and |B ∩ Y | ≥ r − t. Let A′ ⊆ A − B and
B ′ ⊆ B − A have |A′ | = r − pt − |A ∩ B| and |B ′ | = r − t − |A ∩ B|, and
set Y = A′ ∪ B ′ ∪ (A ∩ B). Then |A ∩ Y | = r − pt, |B ∩ Y | = r − t, and
|Y | = 2r − (p + 1)t − |A ∩ B| ≤ r.

Lemma 22.3. Let A, B ∈ Xr and let p ≥ 0 be an integer. There is an

A-B walk of length 2p + 1 in Kn(n, r, s) iff |A ∩ B| ≤ pt + s.

Proof. By

Lemma 22.2, there is an A-B walk of length 2p + 1 iff there exists
C ∈ Xr with |A ∩ C| ≥ r − pt and |B ∩ C| ≤ s. Suppose such a C exists.
Then |A ∩ B| ≤ |A − C| + |B ∩ C| ≤ pt + s. Suppose conversely that
|A ∩ B| ≤ pt + s. If |A ∩ B| ≤ s we may take C = A. Otherwise, we may
take C1 ⊆ A ∩ B with |C1 | = s and, since |A − B| < r − s ≤ n − r = |X − B|,
C2 with A − B ⊆ C2 ⊆ X − B and |C2 | = r − s. Setting C = C1 ∪ C2 , we
have |C| = r, |A ∩ C| = |A − B| + s ≥ r − pt, and |B ∩ C| = s.

Definition 22.4. The odd girth of a graph G is the minimum length of an

odd cycle (assuming there is one; that is, G is not bipartite).

87
 The odd girth of Kn(n, r) was determined by Poljak and Tuza [25], and
this was extended to Kn(n, r, s) under mild restrictions on n, r and s by
Denley [9].

Theorem 22.5. The odd girth of Kn(n, r, s) is 2 r−s

 
t + 1.

Proof. For any graph, the odd girth is the minimal length of an odd closed
walk. By Lemma 22.3, there is a closed walk in Kn(n, r, s) of length 2p + 1
iff r ≤ pt + s, and the result follows.

Proposition 22.6. Let A, B ∈ Xr and set k = |A ∩ B|. The distance

d(A, B) from A to B in Kn(n, r, s) is given by

     
r−k k−s
d(A, B) = min 2 ,2 +1 .
t t

Proof. By Lemma 22.2, there is an A-B walk of length 2p iff p ≥ r−k

 
t . By
 k−s 
Lemma 22.3, there is an A-B walk of length 2p + 1 iff p ≥ t , provided
that k − s > −t. Of course, if k < s then d(A, B) = 1, but for completeness
we verify that in fact k − s > −t and our formula is valid in all cases. Now
k − s ≤ −t iff n − 2r + s ≤ −k, which can happen only if n = 2r − s and
k = 0. But we assumed that if n =
2r − s then s > 0, and in this case there
do not exist disjoint elements of Xr .

The diameter of Kn(n, r) was determined by Valencia-Pabon and Vera

[29].

Theorem 22.7. The diameter of Kn(n, r, s) is r−s−1

 
t + 1.

Proof. Set a = r−s−1

 
t + 1, so that r − s − 1 ≤ t(a − 1) ≤ r− s +  t − a2.
The
a−1
first of these inequalities may be written as r − s − 1 ≤ t 2 + 2 , or
r − t 2 ≤ s + t 2 + 1. Hence, for any integer k, either k ≥ r − t a2
a  a−1 

or k ≤ s + t a−1 . In the first case, 2 r−k

   
2 t ≤ a, and in the second
2 k−s
 
t +1 ≤ a, and it follows by Proposition 22.6 that diam Kn(n, r, s) ≤ a.
 inequality t(a −1) ≤ r − s + t − 2 may be similarly rewritten as
The
s + t a2 − t + 1 ≤ r − t a−1 2 − 1. If k is an integer with
 
jak a−1
(22.1) s+t −t+1≤k ≤r−t −1
2 2

then 2 r−k ≥ 2 a−1 + 1 ≥ a and 2 k−s +1 ≥ 2 a2 +1 ≥ a. Therefore

   
   
t 2 t
if we can find A and B in Xr with k = |A ∩ B| satisfying (22.1) we will

88
have diam Kn(n, r, s) ≥ a, completing the proof. Certainly there exist A
and B with |A ∩ B|  = k provided as ≤ k ≤ r, so it is enough to prove
that s ≤ r − t a−12  − 1 and s + t 2 − t +  1≤  r. The first inequality is
a−1 a a−1
equivalent to t 2  ≤ r − s − 1 and (since 2 ≤ 2 + 1) implies the
second. Now t a−1 2 ≤ 21 (r − s + t − 2), which is at most r− s − 1 provided
r−s−1
t ≤ r − s. On the other hand, if t > r − s then a = t + 1 ≤ 2, and so
a−1
t 2 = 0 ≤ r − s − 1.

In the rest of this section we determine the automorphism group of

Kn(n, r, s). For Kn(n, r), this can be determined using the Erdös-Ko-Rado
Theorem [12, 20], but the argument below does
work in this case. If s = r−1
n
then Kn(n, r, s) is the complete graph on r vertices, so from now on we
further assume s ≤ r
− 2, and hence 2 ≤ r ≤ n − 2. The symmetric group Sn
acts naturally on Xr as a group of automorphisms of Kn(n, r, s). Further,
if n = 2r there is an automorphism χ of Kn(n, r, s) given by χ(A) = X − A,
which is not induced by an element of Sn , and commutes with those that
are. We show that
these automorphisms generate the whole group.
For A ∈ Xr we let N (A) be the set of neighbors of A in Kn(n, r, s),
and for distinct vertices A and B we let N (A, B) = N (A) ∩ N (B). Clearly
the cardinality of N (A, B) depends only on that of A ∩ B. We let f (k) =
|N (A, B)| if there exist A and B with |A ∩ B| = k, and 0 otherwise.

Lemma 22.8. For 0 < k < r, f (k) ≥ f (k − 1), and f (r − 1) > f (r − 2).

Proof. Suppose there exist A, B ∈ Xr with |A ∩ B| = k − 1. (Otherwise

there is nothing to do.) Choose b ∈ B − A and
c ∈ A − B and set C =

B − {b} ∪ {c}, so that |A ∩ C| = k. For D ∈ Xr , we have D ∈ N (B) − N (C)
iff |B ∩ C ∩ D| = s, b ∈
/ D and c ∈ D, and similarly with the rôles of B and
C reversed. Hence there is a bijection φ : N (B) − N (C) → N (C) − N (B)
given by φ(D) = D ∪ {b} − {c}. Since |A ∩ φ(D)| = |A ∩ D| − 1, we have
f (k) ≥ f (k −1), and the inequality is strict if there exists E ∈ N (C)−N (B)
with |A ∩ E| = s. Suppose then that k = r − 1, and let x and y be the
unique elements of A − C and C − A. If s > 0 we may take E1 ⊆ A ∩ B
with |E1 | = s − 1 and E2 ⊆ X − (A ∪ C) with b ∈ E2 and |E2 | = r − s − 1,
and set E = E1 ∪ E2 ∪ {x, y}; while if n > 2r − s we may take E1 ⊆ A ∩ B
with |E1 | = s and E2 ⊆ X − (A ∪ C) with b ∈ E2 and |E2 | = r − s and
set E = E1 ∪ E2 . At least one of these constructions is possible, and we are
done.

Remark. We need not have f (k) > f (k −1), even if f (k) > 0. For example,
with n = 8, r = 4 and s = 2 we have f (1) = f (0) = 36.

89
 We define a new graph with vertex set Xr , the maximal overlap graph

M (n, r), in which A and B are adjacent iff |A ∩ B| = r − 1. (It is the

complement of Kn(n, r, r − 2).) An immediate consequence of the previous
lemma is:
Corollary 22.9. If s ≤ r − 2, every automorphism of Kn(n, r, s) is an
automorphism of M (n, r).
Let T = {A, B, C} be a triangle in M (n, r). Then |A ∩ B ∩ C| is equal to
r − 1 or r − 2, and we say that T is of the first or second kind accordingly.
Since 2 ≤ r ≤ n − 2, triangles of both kinds exist. If T is of the first kind,
there are n − r − 2 vertices adjacent to A, B and C in M (n, r) (the sets
(A ∩ B ∩ C) ∪ {x} for x ∈/ A ∪ B ∪ C), while if T is of the second kind, there
are r − 2 (the sets (A ∪ B ∪ C) − {x} for x ∈ A ∩ B ∩ C). Hence if n 6= 2r,
any automorphism of M (n, r) takes each triangle to one of the same kind.
When n = 2r this is not the case; the complementation automorphism χ
interchanges the two kinds of triangles, so the following result is non-trivial
in this case.
Lemma 22.10. Let T and T ′ be triangles in M (n, r), and α an automor-
phism of M (n, r). Then T and T ′ are of the same kind iff α(T ) and α(T ′ )
are of the same kind.
Proof. Let T be the graph whose vertices are the triangles in M (n, r), with
two triangles being adjacent in T if they share an edge. If T = {A, B, C}
and T ′ = {A, B, C ′ } are adjacent triangles, then C and C ′ are adjacent
in M (n, r) iff T and T ′ are of the same kind, so the result holds in this
case. It therefore suffices to prove that T is connected. Any triangle of the
second kind is adjacent to one of the first, so it is enough to prove that any
two triangles T1 = {A1 , B1 , C1 } and T2 = {A2 , B2 , C2 } of the first kind are
connected. Let Yi = Ai ∩ Bi ∩ Ci , so that Ai = Yi ∪ {ai }, Bi = Yi ∪ {bi }
and Ci = Yi ∪ {ci } for distinct ai , bi and ci ∈ X − Yi . The proof is by
induction on k = |Y1 − Y2 |, the case k = 0 being obvious. Suppose that
k > 0. We may assume that a1 ∈ Y2 , using the case k = 0 if necessary. Let
y ∈ Y1 − Y2 , and set Y ′ = Y1 − {y} ∪ {a1 }, A′ = A1 , B ′ = B1 − {y} ∪ {a1 }
and C ′ = C1 − {y} ∪ {a1 }. Then T ′ = {A′ , B ′ , C ′ } is a triangle of the first
kind with A′ ∩ B ′ ∩ C ′ = Y ′ , and since Y ′ − Y2 = Y1 − Y2 − {y}, T ′ is
connected to T2 by induction. Since {A1 , B1 , B ′ } is a triangle (of the second
kind) adjacent to both T1 and T ′ , T1 and T2 are connected.

Lemma 22.11. Let α be an automorphism of M (n, r) that takes each tri-

angle to one of the same kind. Then α is induced by an element of Sn .

90
 X


Proof. Let A, B and C be elements of r with A adjacent to both B and
C in M (n, r). We claim that:

(1) if A − B = A − C then α(A) − α(B) = α(A) − α(C);

(2) if B − A = C − A then α(B) − α(A) = α(C) − α(A).

If B = C this is trivial. Otherwise, A − B = A − C iff {A, B, C} is a triangle

of the first kind, and B − A = C − A iff it is a triangle of the second kind,
and the claim follows.
We now show that if A1 B1 and A2 B2 are edges of M (n, r) with A1 −B1 =
A2 − B2 , then α(A1 ) − α(B1 ) = α(A2 ) − α(B2 ). The proof is by induction
on k = |A1 − A2 |, the case k = 0 being (1). Suppose that k > 0 and pick
a1 ∈ A1 − A2 and a2 ∈ A2 − A1 . Note that A1 − B1 = A2 − B2 implies that
a1 ∈ B1 . Let B1 − A1 = {b1 }, and set

A′ = A1 − {a1 } ∪ {a2 };
B ′ = B1 − {b1 } ∪ {a2 }.

(It may be that B ′ = B1 .) We have A1 ∩ B ′ = A1 ∩ B1 , so A1 and B ′ are

adjacent and A1 − B ′ = A1 − B1 . Thus α(A1 ) − α(B ′ ) = α(A1 ) − α(B1 ) by
(1). Also A′ − B ′ = A1 − B ′ (since a1 and a2 are both in B ′ ), which implies
that A′ and B ′ are adjacent and α(A′ ) − α(B ′ ) = α(A1 ) − α(B ′ ) by (2). Now
|A′ − A2 | = k − 1, so α(A′ ) − α(B ′ ) = α(A2 ) − α(B2 ) by induction, and we
are done.
It follows that there is a well-defined function π : X → X such that if
AB is an edge of M (n, r) and A − B = {a} then α(A) − α(B) = {π(a)}. If
a 6= b ∈ X, we can pick an edge AB with A − B = {a} and B − A = {b}.

6= π(b), and so π ∈ Sn . We must show that π(A) = α(A) for
Hence π(a)
A ∈ Xr . If a ∈ A then π(a) ∈ α(A) by the definition of π. That is,
π(A) ⊆ α(A), and so π(A) = α(A).

Theorem 22.12. If s ≤ r − 2, the automorphism group of Kn(n, r, s) is

isomorphic to Sn if n 6= 2r, and to Sn × Z2 if n = 2r.

Proof. Let α be an automorphism of Kn(n, r, s). We must show that either

α is induced by an element of Sn , or n = 2r and χα is so induced. By
Corollary 22.9, α is an automorphism of M (n, r). By the discussion before
Lemma 22.10, if n 6= 2r then α takes each triangle of M (n, r) to one of the
same kind, while by that lemma, if n = 2r either α or χα has this property.
Lemma 22.11 completes the proof.

91
A More topology
Here I fill in a few details from section 13 that my inner pedant insists must
be covered. I have tried, perhaps quixotically, to limit the prerequisites to
things that should have at least been stated in 1550 – 2057. The first lemma
is left as an exercise.
Lemma A.1. Let A and B be subsets of R2 . If A and B are open (resp.
closed), then A ∪ B and A ∩ B are open (resp. closed).
Lemma A.2. Let f : [0, 1] → [0, 1] be a continuous bijection. Then f ({0, 1}) =
{0, 1}.
Proof. Suppose, for a contradiction, that 0 < f (0) < 1. Since f is surjective,
there exist a, b ∈ [0, 1] with f (a) = 0 and f (b) = 1. By the Intermediate
Value Theorem, there is some c between a and b with f (c) = f (0), contra-
dicting the injectivity of f . A similar contradiction arises if 0 < f (1) < 1,
completing the proof.
Lemma A.3. Let f : [a, b] → R2 be continuous. The image of f is a closed
and bounded subset of R2 .
Proof. Let A = f ([a, b]). We show first that A is closed. Let x ∈ R2 −A, and
define g : [a, b] → R by g(s) = kf (s) − xk. By the Extreme Value Theorem,
g has a minimum value ǫ, and ǫ > 0. If y ∈ B(x, ǫ) then y ∈ / A, so R2 − A
is open and A is closed.
Now define h : [a, b] → R by h(s) = kf (s) − 0k. If M is the maximum
value of h then A ⊆ B(0, M ).
Lemma A.4. Let f and g be injective paths in R2 with the same image.
Then the composite h = g −1 ◦ f : [0, 1] → [0, 1] is continuous.
Proof. Let x0 ∈ [0, 1], y0 = h(x0 ). and ǫ > 0. We must show that there
exists δ > 0 such that x ∈ [0, 1] and |x − x0 | < δ imply that |h(x) − y0 | < ǫ.
Let Y = {y ∈ [0, 1] | |y − y0 | ≥ ǫ}. Replacing ǫ by a smaller number, if
necessary, we may assume that Y is a closed interval or a disjoint union of
two closed intervals. By Lemmas A.1 and A.3, g(Y ) is a closed subset of
R2 . Hence there exists γ > 0 such that B(f (x0 ), γ) ⊆ R2 − g(Y ), and then
δ > 0 such that x ∈ [0, 1] and |x − x0 | < δ imply f (x) ∈ B(f (x0 ), γ), and
therefore h(x) ∈
/ Y and |h(x) − y0 | < ǫ, as required.
The next result follows immediately from Lemmas A.2 and A.4.
Theorem A.5. Let f and g be injective paths in R2 with the same image.
Then f ({0, 1}) = g({0, 1}).

92
References
[1] K. Appel, W. Haken and J. Koch, Every planar map is four-colorable,
Illinois J. Math. 21 (1977) 429–567.

[2] Elichi Bannai and Tatsuro Ito, On Moore graphs, J. Fac. Sci. Univ.
Tokyo Sect IA 20 (1973) 191–208.

[3] H. Bielak and M.M. Syslo, Peripheral vertices in graphs, Studia Sci.
Math. Hungar. 18 (1983) 269–75.

[4] Norman L. Biggs, E. Keith Lloyd and Robin J. Wilson, Graph Theory
1736–1936, Clarendon Press, Oxford 1976.

[5] F. Buckley, Z. Miller and P.J. Slater, On graphs containing a given

graph as center, J. Graph Theory 5 (1981) 427–34.

[6] A. Cayley, A theorem on trees, Quart. J. Math 23 (1889) 259–61; Col-

lected papers, Cambridge 13 (1892) 26–8.

[7] G. Chartrand and L. Lesniak, Graphs and Digraphs (4th edition), Chap-
man and Hall/CRC 2004.

[8] R.M. Damerell, On Moore graphs, Proc. Cambridge Philos. Soc. 74

(1973) 227–36.

[9] Tristan Denley, The odd girth of the generalised Kneser graph, Euro-
pean J. Combin. 18 (1997) 607–11.

[10] Noam Elkies, Existence and uniqueness of the Hoffman-Singleton graph

(outline),
http://www.math.harvard.edu/~elkies/Misc/hsgraph.pdf.

[11] P. Erdös and T. Gallai, Graphs with prescribed degrees of vertices (Hun-
garian), Mat. Lapok 11 (1960) 264–74.

[12] P. Erdös, C. Ko and R. Rado, Intersection theorems for systems of finite

sets, Q. J. Math. 12 (1961) 313–20.

[13] István Fáry, On straight line representation of planar graphs, Acta

Univ. Szeged. Sect. Sci. Math. 11 (1948) 229–33.

[14] Martin Gardner, Mathematical Games, Sci. Amer. 234 (1976) 126–30.

93
[15] S.L. Hakimi, On the realizability of a set of integers as degrees of the
vertices of a graph, J. SIAM Appl. Math. 10 (1962) 496–506.

[16] F. Harary and R.Z. Norman, The dissimilarity characteristic of Husimi

trees, Ann. of Math. 58 (1953) 134–41.

[17] V. Havel, A remark on the existence of finite graphs (Czech), Časopis

Pěst. Mat. 80 (1955) 477–80.

[18] G.R.T. Hendry, On graphs with prescribed median I, J. Graph Theory

9 (1985) 477–81.

[19] A.J. Hoffman and Robert R.Singleton, On Moore graphs of diameter

two and three, IBM J. Res. Develop. 4 (1960) 497–504.

[20] G.O.H. Katona, A simple proof of the Erds-Chao Ko-Rado theorem, J.

Combin. Theory Ser. B 13 (1972) 183–4.

[21] G. Kirchoff, Über die Auflösung der Gleichungen, auf welche man bei
der Untersuchung der linearen Verteilung galvanischer Ströme geführt
wird, Ann. Phys. Chem. 72 (1847) 497–508; Gesammelte Abhandlun-
gen, Leipzig (1882) 22–3.

[22] D. König, Über Graphen und ihre Anwendung auf Determinanttheorie

und Mengenlehre, Math. Ann. 77 (1916) 453–65.

[23] K. Kuratowski, Sur le probléme des courbes gauches en topologie, Fund.

Math. 15 (1930) 271–83.

[24] K. Menger, Zur allgemeinen Kurventheorie, Fund. Math. 10 (1927)

95–115.

[25] S. Poljak and Zs. Tuza, Minimum bipartite subgraphs of Kneser graphs,
Graphs Combin. 3 (1987) 191–9.

[26] H. Prüfer, Neuer Beweis eines Satzes über Permutationen, Arch. Math.
Phys. 27 (1918) 142–4.

[27] N. Robertson, D.P. Sanders, P.D. Seymour and R. Thomas, The four-
color theorem, J. Combin. Theory Ser. B 70 (1997) 2–44.

[28] Robert R.Singleton, There is no irregular Moore graph, Amer. Mat.

Monthly 75 (1968) 42–3.

94
[29] Mario Valencia-Pabon and Juan-Carlos Vera, On the diameter of
Kneser graphs, Discrete Math. 305 (2005) 383–5.

[30] Klaus Wagner, Bemerkung zum Vierfarbenproblem, Jber. Deutsch

Math.-Verein. 46 (1936) 26–32.

[31] H. Whitney, Congruent graphs and the connectivity of graphs, Amer. J.

Math. 54 (1932) 150–68.

[32] David R. Wood, A simple proof of the Fáry-Wagner Theorem,

http://www-ma2.upc.es/wood/papers/Wood-FaryWagner.pdf.

95
 List of errata

Problem 6.1 in the notes cannot be answered as stated. See the corrected
version in the separate collection of exercises.
Page Line reads should read
15 -9 0f of
16 13 (u0 , u1 , . . . , ul ) W1 = (u0 , u1 , . . . , ul )
16 13 (v0 , v1 , . . . , vk ) W2 = (v0 , v1 , . . . , vk )
16 14 0≤j≤k 0 ≤ j ≤ k and every edge
of W2 is an edge of W1
29 7 – 8 Theorem 5.14 Corollary 5.15
77 3 regular graph non-empty regular graph