Introduction To Combinatorics, Course Notes PDF

Introduction to Combinatorics
Course Notes for Math 239
Department of Combinatorics and Optimization

University of Waterloo
©August 18, 2016

ii
Contents
1 Combinatorial Analysis 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Sums and Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Binomial Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Bijections (One-to-One Correspondence) . . . . . . . . . . . . . . . 4
1.5 Combinatorial Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Generating Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Formal Power Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.8 The Sum and Product Lemmas . . . . . . . . . . . . . . . . . . . . . 24
2 Compositions and Strings 29

2.1 Compositions of an Integer. . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Subsets with Restrictions. . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3 Binary Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Unambiguous Expressions . . . . . . . . . . . . . . . . . . . . . . . 42
2.5 Some Decomposition Rules . . . . . . . . . . . . . . . . . . . . . . . 43
2.6 Sum and Product Rules for Strings . . . . . . . . . . . . . . . . . . . 44
2.7 Decomposition Using Blocks . . . . . . . . . . . . . . . . . . . . . . 45
2.8 Recursive Decompositions of Binary Strings . . . . . . . . . . . . . 47
2.9 Bivariate Generating Series . . . . . . . . . . . . . . . . . . . . . . . 54
3 Recurrences, Binary Trees and Sorting 63

3.1 Coefficients of Rational Functions . . . . . . . . . . . . . . . . . . . 63
3.2 Solutions to Recurrence Equations . . . . . . . . . . . . . . . . . . . 65
3.3 Nonhomogeneous Recurrence Equations . . . . . . . . . . . . . . . 70
3.4 Aysmptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.5 Other Recurrence Relations . . . . . . . . . . . . . . . . . . . . . . . 73
3.6 Binary Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
iii
iv CONTENTS
3.7 The Binomial Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3.8 The Quicksort Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.9 The Mergesort Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 88
4 Introduction to Graph Theory 93

4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2 Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3 Degree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.4 Bipartite Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5 How to Specify a Graph . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.6 Paths and Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.7 Equivalence Relations . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.8 Connectedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.9 Eulerian Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.10 Bridges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.11 Certifying Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5 Trees 125
5.1 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.2 Spanning Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3 Characterizing Bipartite Graphs . . . . . . . . . . . . . . . . . . . . 130
5.4 Breadth-First Search . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.5 Applications of Breadth-First Search . . . . . . . . . . . . . . . . . . 139
5.6 Minimum Spanning Tree . . . . . . . . . . . . . . . . . . . . . . . . 140
6 Codes 145
6.1 Vector Spaces and Fundamental Cycles . . . . . . . . . . . . . . . . 145
6.2 Graphical Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7 Planar Graphs 157

7.1 Planarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2 Euler’s Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.3 Stereographic Projection . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.4 Platonic Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.5 Nonplanar Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.6 Kuratowski’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 171
7.7 Colouring and Planar Graphs . . . . . . . . . . . . . . . . . . . . . . 177
7.8 Dual Planar Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
CONTENTS v
8 Matchings 185
8.1 Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
8.2 Covers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8.3 König’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
8.4 Applications of König’s Theorem . . . . . . . . . . . . . . . . . . . . 197
8.5 Systems of Distinct Representatives . . . . . . . . . . . . . . . . . . 198
8.6 Perfect Matchings in Bipartite Graphs . . . . . . . . . . . . . . . . . 200
8.7 Edge-colouring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.8 An Application to Timetabling . . . . . . . . . . . . . . . . . . . . . 204
vi CONTENTS
Chapter 1
Combinatorial Analysis
1.1 Introduction
Our goal in the first half of this course is to learn how to solve a number of
interesting counting problems.
We offer some examples. A composition of a non-negative integer n is a
sequence (m 1 , . . . , m r ) of positive integers such that
m 1 + · · · + m r = n.
The numbers m 1 , . . . , m r are the parts of the composition. We define the weight
of a composition to be the sum of its parts. For example, the number of different
ways we can cut a piece of string of length n centimetres into pieces of length 1
and 2 cm is the number of compositions with weight n where all parts are equal
to 1 or 2. (We will often use weight where the word “size” would do, because
sometimes there is more than one way of defining size.)
Other questions we might ask about compositions include
(a) How many different compositions of n are there?
(b) How many compositions of n are there with exactly k parts?
(c) How many compositions of n are there where all parts of the composition
are odd?
A second important class of counting problems involves binary strings. A
binary string of length n is a sequence a 1 · · · a n where each a i is 0 or 1. Clearly
there are 2n binary strings of length n (so that’s solved). But there are other
questions:
1
2 CHAPTER 1. COMBINATORIAL ANALYSIS
(a) How many binary strings of length n are there that do not contain an odd
string of 0’s as a maximal substring? (So 1001 is OK, but 10001 is not.)
(b) How many binary strings of length n are there that do not contain 0101 as
a substring?
(c) How many binary strings of length n are there that contain exactly k 1’s?
1.2 Sums and Products

A natural way to determine the size of a set is to show that it has the same size
as some set of known size. Usually the “known” set will be built from smaller
known sets using unions and products.
We recall that if A and B are sets, then the union A ∪ B is defined by
A ∪ B := {x : x ∈ A or x ∈ B }.
If A and B are disjoint, that is, A ∩ B = ; then
|A ∪ B | = |A| + |B |.
Here |S| denotes the number of elements in the set S.

The Cartesian product A × B of sets A and B is the set of all ordered pairs
whose first element is an element of A and second element is an element of B ,
that is
A × B := {(a, b) : a ∈ A, b ∈ B }.
Then
|A × B | = |A| |B |.
We can similarly define the Cartesian power A k as ordered k-tuples of elements
from A.
A k := {(a 1 , a 2 , . . . , a k ) : a 1 , a 2 , . . . , a k ∈ A}.
When A is finite, |A k | = |A|k . Thus if
A = {1, 2, 3, 4, 5, 6}
and k = 3, then the elements of A 3 correspond to the 63 = 216 possible out-

comes when we toss a 6-sided die three times.
For a second example, if A = {0, 1}, then the elements of A n are the n-tuples
(a 1 , . . . , a n ), a i ∈ {0, 1}.
1.3. BINOMIAL COEFFICIENTS 3
1.3 Binomial Coefficients

We are interested in “enumeration problems"; these are problems that involve
counting various kinds of combinatorial objects. We begin with the following
easy question: How many k–element subsets are there of an n–element set?
Theorem 1.3.1. For non–negative integers n and k, the number of k-element
subsets of an n-element set is
n(n − 1) · · · (n − k + 1)
.
k!
Proof: Let L be the set of all ordered lists of k distinct numbers from the set
{1, . . . , n}. There are n choices for the first number in the list, and then there are
n − 1 remaining choices for the second number. In general, when choosing the
i –th number in our list, we have already used i −1 of the n numbers and, hence,
we have n − i + 1 numbers to choose from. Thus,
|L | = n(n − 1) · · · (n − k + 1).
Now consider a different way of generating the elements of L . We first

choose the k elements and then order the elements in all possible ways. Let
there be x ways to choose k elements from {1, . . . , n}. There are k! = k(k −1) · · · 3·
2 · 1 ways to order, or permute, each k–element set. Thus,
|L | = x(k!).
Hence
n(n − 1) · · · (n − k + 1)
x= ,
k!
as required.
We define nk , which we read as “n choose k”, to be the number of k-element
¡ ¢
subsets of {1, . . . , n}. So by Theorem 1.3.1,

Ã !
n n(n − 1) · · · (n − k + 1)
= .
k k!
Note that nk = 0 whenever n < k as it should be. It is natural to define the

¡ ¢
product of zero numbers to be one; for example, 52 = 5·5 = 25, 5¡1 = 5 and 50 = 1,
n¢
the product of zero 5’s, is 1. With this interpretation, 0! = 1 and 0 = 1.
The numerical values of the binomial coefficients form a structure known
as Pascal’s triangle. A few rows of the Pascal’s triangle are given here.
1
1 1
1 2 1
1 3 3 1
1 4 6 4 1
1 5 10 10 5 1
1 6 15 20 15 6 1
The values of each row represent the values of n0 , n1 , . . . , nn in order, where

¡ ¢¡ ¢ ¡ ¢
n = 0 in the first row, n = 1 in the second row, etc.

Notice that this structure is symmetric, meaning that for integers n and k
where 0 ≤ k ≤ n, Ã ! Ã !
n n
= .
k n −k
This can be easily explained through some arithmetic:
Ã ! Ã !
n n(n − 1) · · · (n − k + 1) n! n! n
= = = = .
k k! k!(n − k)! (n − k)!(n − (n − k))! n −k
At first this seems surprising, but it follows easily from the definitions. Indeed,
the number of ways of selecting k elements from an n–element set is the same
as the number of ways of discarding (n − k) elements from the set. We can for-
malize this relation using a bijection, which is described in the next section.
1.4 Bijections (One-to-One Correspondence)

Another way of establishing that two sets have equal size is through the use of a
bijection. We begin with a review of a few definitions on functions. Let S and T
be sets. Let f : S → T be a function (or mapping). In particular, this means that
for any x ∈ S, f (x) is an element in T . Then
• f is 1-1 or injective if for any x 1 , x 2 ∈ S, f (x 1 ) = f (x 2 ) implies x 1 = x 2 . In

other words, every element in S is being mapped to a unique element in
T.
• f is onto or surjective if for all y ∈ T , there exists x ∈ S such that f (x) =

y. In other words, every element in T is being mapped to from some
element in S.
1.4. BIJECTIONS (ONE-TO-ONE CORRESPONDENCE) 5
• f is a 1-1 correspondence or bijection if it is both 1-1 and onto.
We can visualize these definitions using the following diagram:
f f f
a
1 1 1 a
a
b
2 2 2 b
c b
3 3 3 c
d
S T S T S T
1-1 onto bijection
The functions that you have seen in high school or calculus are typically
defined for f : R → R. For example, f (x) = e x is 1-1, but not onto; f (x) = x 3 − x
is onto, but not 1-1; and f (x) = 2x + 1 is both 1-1 and onto, hence a bijection.
In combinatorics, we use mappings to compare the cardinalities of finite
sets S and T . If there exists a mapping f : S → T that is 1-1, then |S| ≤ |T |. This
is because the |S| elements of S must be mapped to distinct elements in T , so
there must be at least |S| distinct elements in T . On the other hand, if there
exists a mapping f : S → T that is onto, then |S| ≥ |T |. This is because for the |T |
elements in T , each must be mapped to by a distinct element in S. Therefore, if
there exists a bijection f : S → T , then |S| = |T |, as f is both 1-1 and onto.
In addition to showing that two sets have equal size, bijections have the nice
property that they “pair up” elements of S with elements of T exactly. This gives
a correspondence between elements of S and T .
We can prove that a function is a bijection by going through the definition
of 1-1 and onto. Alternatively, we could show that the mapping is “reversible”,
that is, there exists an inverse function. For f : S → T , the inverse (if it exists) of
f is a function f −1 : T → S such that for all x ∈ S, f −1 ( f (x)) = x, and for all y ∈ T ,
f ( f −1 (y)) = y.
Theorem 1.4.1. If a function f : S → T has an inverse, then f is a bijection.

Proof: Let f −1 : T → S be an inverse of f . We need to prove that f is 1-1 and

onto. Suppose f (x 1 ) = f (x 2 ). Then f −1 ( f (x 1 )) = f −1 ( f (x 2 )). By definition of
inverse, x 1 = x 2 , so f is 1-1. Let y ∈ T . Since f −1 is a function, f −1 (y) = x for
some x ∈ S. Then f ( f −1 (y)) = f (x), and by the definition of inverse, y = f (x).
Therefore, x is mapped to y, and f is onto.
Usually if a bijection is clear, its inverse is also clear. So in most cases, in-
stead of proving that a function is 1-1 and onto (which can be tedious), we can
simply provide the inverse function.
Problem 1.4.2. For some 0 ≤ k ≤ n, let S be the set of k-subsets of {1, . . . , n}, and
let T be the set of (n − k)-subsets of {1, . . . , n}. Find a bijection between S and T .
Solution: We can define the following mapping:
f : S → T, f (A) = {1, . . . , n} \ A ∀ A ∈ S.
First, we need to check that each element of S is being mapped to an element

in T by f . For any A ∈ S, |A| = k. So f (A) = {1, . . . , n} \ A has cardinality n − k,
meaning f (A) ∈ T .
It should be clear that in order to “reverse” the function, we simply need to
apply the same operation again. So the inverse is f −1 : T → S where for each
B ∈ T , f −1 (B ) = {1, . . . , n} \ B .
We can check that for each A ∈ S,
f −1 ( f (A)) = f −1 ({1, . . . , n} \ A) = {1, . . . , n} \ ({1, . . . , n} \ A) = A.
And for each B ∈ T ,
f ( f −1 (B )) = f ({1, . . . , n} \ B ) = {1, . . . , n} \ ({1, . . . , n} \ B ) = B.
So f −1 is indeed the inverse of f , which proves that f is a bijection.
By Theorem 1.3.1, |S| = nk¡ and

¡ ¢ ¡ n ¢
¢ ¡ n ¢ n−k . So this bijection is also a combi-
|T | =
natorial proof of the identity nk = n−k
We illustrate this bijection by matching the elements of S with elements of
T for the case where n = 5, k = 2. The function f maps in the → direction,
the inverse f −1 maps in the ← direction. For example, f ({1, 5}) = {2, 3, 4} while
f −1 ({1, 2, 5}) = {3, 4}.
1.4. BIJECTIONS (ONE-TO-ONE CORRESPONDENCE) 7
S T
{1, 2} ←→ {3, 4, 5}
{1, 3} ←→ {2, 4, 5}
{1, 4} ←→ {1, 3, 5}
{1, 5} ←→ {2, 3, 4}
{2, 3} ←→ {1, 4, 5}
{2, 4} ←→ {1, 3, 5}
{2, 5} ←→ {1, 3, 4}
{3, 4} ←→ {1, 2, 5}
{3, 5} ←→ {1, 2, 4}
{4, 5} ←→ {1, 2, 3}
We can establish a bijection between two completely different looking sets

of objects in order to find a correspondence between the two sets.
Problem 1.4.3. Let S be the set of all subsets of {1, . . . , n}, and let T be the set of
all binary strings of length n. Find a bijection between S and T .
Solution: We define f : S → T in the following way: For a subset A of {1, . . . , n},

we can create a string f (A) = a 1 a 2 · · · a n length n where
½
0 i 6∈ A
ai =
1 i∈A
It is clear that f (A) is a binary string of length n, hence in T .

This mapping is reversible: Let t = b 1 b 2 · · · b n ∈ T . Define f −1 : T → S where
f −1 (t ) = {i ∈ {1, . . . , n} | b i = 1}.
You can check that f −1 ( f (A)) = A and f ( f −1 (t )) = t , so f is a bijection.
This bijection tells us that the number of subsets of {1, . . . , n} is equal to

the number of binary strings of length n, which is 2n . In addition, we can re-
late some subset problems to binary string problems. For example, subsets of
{1, . . . , n} of size k are mapped to binary strings of length
¡n ¢ n with exactly k 1’s by
f . So the number of such binary strings is equal to k .
We illustrate this bijection for n = 3.
S T
; ←→ 000
{1} ←→ 100
{2} ←→ 010
{3} ←→ 001
{1, 2} ←→ 110
{1.3} ←→ 101
{2, 3} ←→ 011
{1, 2, 3} ←→ 111
1.5 Combinatorial Proofs

Any proof that involves some kind of counting argument is a combinatorial
proof. We have seen combinatorial proofs in Theorem 1.3.1 and the previous
section involving bijections. We give an example here which involves proving
an algebraic identity by counting combinatorial objects.
Theorem 1.5.1. For any non–negative integer n,

Ã !
n n
n
xk .
X
(1 + x) =
k=0 k
This is also known as the binomial theorem. We first give the main idea of
the proof: In the expansion of (1 + x)n , each term in the expansion consists of
a product of n items, each one being either 1 or x. So each term is a product
a 1 a 2 · · · a n where each a i ∈ {1, x}. This term is equal to x k whenever exactly k
of these a i ’s are x’s. The coefficient of x k is then the number of ways that this
happens, which ¡n ¢ is essentially the same as choosing subsets of¡ {1, . . . , n} of size
k n¢
k. There are k of these subsets, hence the coefficient of x is k . For example,
when n = 3,
(1 + x)3 = (1 + x)(1 + x)(1 + x)

= 1 · 1 · 1 + 1 · 1 · x + 1 · x · 1 + 1 · x · x + x · 1 · 1 + x · 1 · x + x · x · 1 + x · x · x.
= 1 + 3x + 3x 2 + x 3 .
We now give a formal proof of the theorem.

1.5. COMBINATORIAL PROOFS 9
Proof: We first expand (1 + x)n algebraically:
(1 + x)n = (1 + x)(1 + x) · · · (1 + x)
Ã !Ã ! Ã !
X X X
= a1 a2 · · · an
a 1 ∈{1,x} a 2 ∈{1,x} a n ∈{1,x}
X X X
= ··· a1 a2 · · · an
a 1 ∈{1,x} a 2 ∈{1,x} a n ∈{1,x}
X
= a1 a2 · · · an .
(a 1 ,a 2 ,...,a n )∈{1,x}n
Define S k to be the set of elements of {1, x}n where exactly k of the n terms are
x’s. So if (a 1 , . . . , a n ) ∈ S k , then a 1 a 2 · · · a n = x k . Also,
{1, x}n = S 0 ∪ S 1 ∪ · · · ∪ S n .
This is a disjoint union of sets, so we can break up the sum at the end of the
equation above to get
Ã !
n
n
X X
(1 + x) = a1 a2 · · · an
k=0 (a 1 ,a 2 ,...,a n )∈S k
Ã !
n
xk
X X
=
k=0 (a 1 ,a 2 ,...,a n )∈S k
n
|S k |x k .
X
=
k=0
Let Tk be the set of all subsets of {1, . . . , n} of size k. Then there is a bijection
between S k and Tk : define f : S k → Tk by f (a 1 , a 2 , . . . , a n ) = {i | a i = x}. (This¡ is¢
similar to the bijection defined in Problem 1.4.3.) This gives us |S k | = |Tk | = nk .
Therefore, Ã !
n n
(1 + x)n = xk .
X
k=0 k
Sometimes an algebraic identity can be established by giving a set of objects,

and count the size of this set in two different ways. Since we are counting the
same set of objects, the two ways of counting must result in the same number.
We give two examples here.
Problem 1.5.2. For integers 1 ≤ k ≤ n, show that

Ã ! Ã ! Ã !
n n −1 n −1
= + .
k k −1 k
Solution:
¡n ¢ Let S be the set of all k-subsets of {1, . . . , n}. By Theorem 1.3.1, |S| =
k
.
We now count |S| in a different way. Partition S into two sets S 1 and S 2 where
S 1 contains all k-subsets of {1, . . . , n} which include the element n, and S 2 con-
tains all k-subsets of {1, . . . , n} which do not include the element n. Since S 1 and
S 2 are disjoint, |S| = |S 1 | + |S 2 |.
¡n−1¢ element of S 1 consists of n and a (k − 1)-subset of {1, . . . , n¡n−1
Now each − 1},
¢
so |S 1 | = k−1 . Each element of S 2 is a k-subset of {1, . . . , n − 1}, so |S 2 | = k .
Hence nk = n−1
¡ ¢ ¡ ¢ ¡n−1¢
k−1 + k .
Problem 1.5.3. For non–negative integers n and k, show that

Ã ! Ã !
n +k X k n +i −1
= .
n i =0 n −1
Solution: Let S denote the set of all n–element subsets of {1, . . . , n + k}. By
Theorem 1.3.1,
Ã !
n +k
|S| = .
n
Note that for any set A ∈ S, the largest element of A is at least n. Now, for each
i ∈ {0, . . . , k}, let S i be the set of all sets in S whose largest element is n + i . Thus,
(S 0 , S 1 , . . . , S k ) is a partition of S, so
|S| = |S 0 | + |S 1 | + · · · + |S k |.
Now, if A ∈ S i then n+i ∈ A and A\{n+i } is an (n−1)–element subset of {1, . . . , n+

i − 1}. Then, by Theorem 1.3.1,
Ã !
n +i −1
|S i | = .
n −1
1.5. COMBINATORIAL PROOFS 11
Hence,
|S| = |S 0 | + |S 1 | + · · · + |S k |
Ã ! Ã ! Ã !
n −1 n n +k −1
= + +···+
n −1 n −1 n −1
Ã !
X k n +i −1
= .
i =0 n −1
Thus, Ã ! Ã !
n +k X k n +i −1
= ,
n i =0 n −1
as required.
Problem Set 1.5

P ¡n ¢ P ¡ n ¢
1. Consider the identity i ≥0 2i = i ≥0 2i +1 .
(a) Give a combinatorial proof of this identity. (Hint: Find a bijection be-
tween the even and odd subsets of {1, . . . , n}.)
(b) Give an algebraic proof for the identity. (Hint: Expand (1 − 1)n .)
2. Consider the identity

Ã ! Ã !Ã !
m +n X k m n
= .
k i =0 i k −i
¡m ¢ We could also use min{k, m} as the upper limit of this sum. Be-
(Remark:
cause i = 0 whenever i > m, k is also a valid upper limit of this sum.)
(a) Give a combinatorial proof of this identity. (Hint: The left side counts
the number of ways of choosing k elements from a set of (m + n) ele-
ments — determine a similar interpretation for the right side.)
(b) Give an algebraic proof for the identity. (Hint: Expand both sides of the
equation (1 + x)m+n = (1 + x)m (1 + x)n using Theorem 1.5.1.)
Pn ¡ n ¢ n−1
3. Consider the identity i =0 i i = n2 .
(a) Give a combinatorial proof of this identity. (Hint: if S is a subset of

{1, . . . , n} and S̄ is its complement then |S| + |S̄| = n. )
(b) Give an algebraic proof for the identity. (Hint: Expand (1+ x)n and take
the derivatives of both versions.)
4. Show that ni=0 ni i (i − 1) = n(n − 1)2n−2 .

P ¡ ¢
5. (a) Consider the following two sets.
Sk = {(A, B ) | A, B ⊆ {1, . . . , n}, |A| = |B | = m, |A ∩ B | = k},

Tk = {(X , Y , Z ) | X , Y , Z ⊆ {1, . . . , n}, |X | = k, |Y | = |Z | = m − k,
X ∩ Y = X ∩ Z = Y ∩ Z = ;}.
Define a bijection f : S k → Tk , explain why f (A, B ) ∈ Tk for any (A, B ) ∈

S k , and provide the inverse of your bijection.
(b) Consider the set S = {(A, B ) | A, B ⊆ {1, . . . , n}, |A| = |B | = m}. By count-
ing S in two different ways, give a combinatorial proof of the following
identity:
Ã !2 Ã !Ã !Ã !
n Xm n n −k n −m
= .
m k=0 k m − k m − k
1.6 Generating Series

We begin with a very general enumeration problem.
Problem 1.6.1. Suppose that S is a set of “configurations” and, for each config-
uration σ ∈ S, we have a non–negative integer weight w(σ). For a given integer
k, how many elements of S have weight k?
Consider, for example, the problem from the previous section of counting
subsets. Let S be the set of all subsets of {1, . . . , n}. For each subset σ ∈ S, we
¡ ¢its weight to be w(σ) = |σ|. Thus, the number of elements of S of weight
define
k is nk , the number of k element subsets of {1, . . . , n}.
Definition 1.6.2. Let S be a set of configurations with a weight function w. The

generating series for S with respect to w is defined by
X w(σ)
ΦS (x) = x .
σ∈S
1.6. GENERATING SERIES 13
By collecting like–powers of x in ΦS (x), we get

X w(σ)
ΦS (x) = x
σ∈S
Ã !
1 xk
X X
=
k≥0 σ∈S,w(σ)=k
ak x k ,
X
=
k≥0
where a k denotes the number of elements in S with weight k. That is, the coef-
ficient of x k in ΦS (x) counts the number of elements of weight k in S.
Consider, again, the set S of all subsets of {1, 2, . . . , n}. Let the weight of set
σ ∈ S be |σ|. Then
Ã ! Ã ! Ã ! Ã ! Ã !
n n n 2 n n X n n
Φs (x) = + x+ x +···+ x = x k = (1 + x)n .
0 1 2 n k=0 k
Consider the case where S is the set of all integer pairs (a, b) where a ∈
{1, 3, 5}, b ∈ {2, 4, 6} and the weight of (a, b) is defined as a + b. Thus,
ΦS (x) = x 1+2 + x 1+4 + x 1+6 + x 3+2 + x 3+4 + x 3+6 + x 5+2 + x 5+4 + x 5+6
= x 3 + 2x 5 + 3x 7 + 2x 9 + x 11 .
So, there is 1 configuration of weight 3, 2 configurations of weight 5, 3 configu-

rations of weight 7, 2 configurations of weight 9, and 1 configuration of weight
11. Also, note that ΦS (x) factors:
ΦS (x) = x 1+2 + x 1+4 + x 1+6 + x 3+2 + x 3+4 + x 3+6 + x 5+2 + x 5+4 + x 5+6
= (x 1 + x 3 + x 5 )(x 2 + x 4 + x 6 ).
Now, for any integer k ≥ 0, let N≥k be the set of all integers ≥ k; thus N≥k =
{k, k + 1, k + 2, . . .}, and the weight of a number is its value. Then
ΦN≥k (x) = x k + x k+1 + x k+2 + · · ·

= x k (1 + x + x 2 + x 3 + · · · ).
In the first example above the generating series is just a polynomial, since
S is finite. In the second example, N≥k is infinite, so the generating series is a
“formal power series”; we discuss these in greater detail in the next section.
Theorem 1.6.3. Let ΦS (x) be the generating series for a finite set S with respect
to a weight function w. Then,
(a) ΦS (1) = |S|,
(b) the sum of the weights of the elements in S is Φ0S (1), and
(c) the average weight of an element in S is Φ0S (1)/ΦS (1).

Proof: Note that
ΦS (1) = 1w(σ) = |S|,
X
σ∈S
which proves (a). Similarly,
Φ0S (x) = w(σ)x w(σ)−1 .

X
σ∈S
Thus,
Φ0S (1) =
X
w(σ),
σ∈S
which proves (b). Finally, (c) follows immediately from (a) and (b).
Consider again the example in which S is the set of all subsets of {1, . . . , n},
¡n ¢ set σ ∈ S , we define its weight to be w(σ) = |σ|. By definition,
and for each
there are k elements of S of weight k. Thus,
Ã !
n n
ΦS (x) = x k = (1 + x)n .
X
k=0 k
Now,
ΦS (1) = (1 + 1)n = 2n .
That is, there are 2n subsets of {1, . . . , n}. Now,
Φ0S (x) = n(1 + x)n−1 ,
so the sum of the sizes of the subsets of {1, . . . , n} is
Φ0S (1) = n2n−1 .
Therefore, the average size of a subset of {1, . . . , n} is
Φ0S (1)/ΦS (1) = n2n−1 /2n = n/2,
as expected.
1.6. GENERATING SERIES 15
Problem Set 1.6

1. Make a list of all the four-letter “words” that can be formed from the “al-
phabet” {a, b}. (There are 16 of them altogether). If the weight of a word is
defined to be the number of occurrences of ‘ab’ in it, determine how many
words there are of weight 0, 1 and 2. Determine the generating series.
2. Do the same for five letter words over the same alphabet, but, preferably,
without listing all the words separately.
3. Two six–sided dice, one red, one blue, are thrown. Consider all 36 combi-
nations of faces showing. Taking the weight of a throw to be the sum of the
spots showing (the usual score, in other words), make a table showing the
number of throws of each possible weight.
4. Construct a table, as in problem 3, for the case where the weight of a throw is
defined to be the absolute value of the difference between the faces showing.
5. Find the generating series ΦS (x) if:
(a) S = {1, 2, . . .} and w(i ) = i for all i ;

(b) S = {1, 2, . . .} and w(i ) = i if i is even, w(i ) = i − 1 if i is odd.
6. If M (x) is the generating series for a finite set S of objects, obtain an expres-
sion for the sum of the squares of the weights of the elements of the set S in
terms of M 0 (1) and M 00 (1).
7. (a) Let s be a subset of {1, 2, 3, 4, 5}. The weight w(s) of s is defined to be the
number of pairs of consecutive integers in s. (For example, {1, 2, 4, 5}
has 2 pairs of consecutive integers, namely {1, 2} and {4, 5}, so w({1, 2, 4, 5}) =
2). Let S be the set of all subsets of {1, 2, 3, 4, 5}. Find the generating se-
ries f (x) of S with respect to w.
(b) Calculate f (1), and justify your answer.
8. A 3 cm by 3 cm piece of cardboard is marked on one side by ruled lines

into 9 equal squares. These 1 cm by 1 cm squares are coloured, indepen-
dently, black or white. The weight of an object is defined to be the number
of black squares it possesses. We say 2 objects are the same if one can be
obtained from the other simply by rotating the first through 0◦ , 90◦ , 180◦ or
270◦ . Find the generating series for the set of different objects with respect
to this weight, up to and including the quadratic term.
1.7 Formal Power Series

When S is a finite set, ΦS (x) is a polynomial. However, we are interested in cases
where S is infinite, and there are some subtleties that we address in this section.
For example, we implicitly assume that there are only finitely many elements of
S of any given weight, so that ΦS (x) is well defined.
Let (a 0 , a 1 , a 2 , . . .) be a sequence of rational numbers; then A(x) = a 0 + a 1 x +
a 2 x 2 + · · · is called a formal power series. We say that a n is the coefficient of
x n in A(x) and write a n = [x n ]A(x). A formal power series should be regarded
as a way of encoding a sequence of numbers. Thus two formal power series are
equal if and only if they have the same sequence of coefficients.
In many respects, formal power series behave much like polynomials; for ex-
ample, addition and multiplication are well defined. If A(x) = a 0 + a 1 x + a 2 x 2 +
· · · and B (x) = b 0 + b 1 x + b 2 x 2 + · · · are formal power series then we define
(a n + b n )x n ,
X
A(x) + B (x) =
n≥0
and Ã !
n
a k b n−k x n .
X X
A(x)B (x) =
n≥0 k=0
In the case that A(x) and B (x) are polynomials, these definitions are consis-
tent with their familiar usage. Indeed, this is easy to see for addition, and for
multiplication
Ã !Ã !
i j
X X
A(x)B (x) = ai x bj x
i ≥0 j ≥0
ai b j x i + j ,
XX
= (now, let k = i and n = i + j )
i ≥0 j ≥0
Ã !
n
a k b n−k x n .
X X
=
n≥0 k=0
We need to be a bit careful when performing infinite sums and products, but
here there is no ambiguity, since there are only finitely many terms contributing
to any given coefficient. Thus, A(x) + B (x) and A(x)B (x) are well defined.
In many of the problems that we will encounter, we will need to solve linear
equations; given two formal power series P (x) and Q(x) we will want to find a
formal power series A(x) such that Q(x)A(x) = P (x). For example, suppose that
1.7. FORMAL POWER SERIES 17
A(x) = a 0 + a 1 x + a 2 x 2 + · · · where a n is the number of {0, 1}-strings of length

n that do not contain two consecutive ones. Using the techniques that we will
develop in Chapter 2, it is straightforward to show that A(x) satisfies the linear
equation (1 − x − x 2 )A(x) = 1 + x.
Problem 1.7.1. Show that the following equation has a solution and that the
solution is unique:
(1 − x − x 2 )A(x) = 1 + x. (1.7.1)
Solution: We can rewrite Equation 1.7.1 as:
a 0 + (a 1 − a 0 )x + (a 2 − a 1 − a 0 )x 2 + (a 3 − a 2 − a 1 )x 3 + · · · = 1 + x. (1.7.2)
Recall that two formal power series are equal if and only if they have the same
sequence of coefficients. Therefore, by equating coefficients on either side of
Equation 1.7.2, we have equality if and only if:
• a 0 = 1,
• a 1 − a 0 = 1, and
• for each n ≥ 2, a n − a n−1 − a n−2 = 0.
Therefore a 0 = 1, a 1 = 2, and, for each n ≥ 2,
a n = a n−1 + a n−2 .
From this recurrence relation and the two initial conditions, we can uniquely
determine the sequence {a n }n≥0 . Therefore, Equation 1.7.1 has a solution and
the solution is unique.
We can extend the method developed in Problem 1.7.1 to solve any linear
equation Q(x)A(x) = P (x) so long as the constant term of Q(x) is non-zero.
Theorem 1.7.2. Let A(x) = a 0 + a 1 x + a 2 x 2 +· · · , P (x) = p 0 + p 1 x + p 2 x 2 +· · · , and

Q(x) = 1 − q 1 x − q 2 x 2 − · · · be formal power series. Then
Q(x)A(x) = P (x)
if and only if for each n ≥ 0,
a n = p n + q 1 a n−1 + q 2 a n−2 + · · · + q n a 0 .
Proof: By definition, Q(x)A(x) = P (x) if and only if [x n ] (Q(x)A(x)) = [x n ]P (x)

for all n ≥ 0. Now [x n ]P (x) = p n and
[x n ] (Q(x)A(x)) = a n − q 1 a n−1 − q 2 a n−2 − · · · − q n a 0 .
Thefore Q(x)A(x) = P (x) if and only if, for each n ≥ 0,
a n − (q 1 a n−1 + q 2 a n−2 + · · · + q n a 0 ) = p n .
The result follows by a simple rearrangement of this equation.
Corollary 1.7.3. Let P (x) and Q(x) be formal power series. If the constant term
of Q(x) is non-zero, then there is a formal power series A(x) satisfying
Q(x)A(x) = P (x).
Moreover, the solution, A(x), is unique.
Proof: If Q(0) 6= 0, then, by dividing both P (x) and Q(x) by Q(0), we may assume
that Q(0) = 1. Then we can write Q(x) = 1 − q 1 x − q 2 x 2 − · · · and P (x) = p 0 +
p 1 x + p 2 x 2 + · · · . Now, by Theorem 1.7.2, A(x) = a 0 + a 1 x + a 2 x 2 + · · · satisfies
Q(x)A(x) = P (x) if and only if for each n ≥ 0,
a n = p n + q 1 a n−1 + q 2 a n−2 + · · · + q n a 0 .
Thus a 0 = p 0 and, for each n > 0, we can determine a n uniquely from a 0 , . . . , a n−1 .
So a solution exists and it is unique.
Definition 1.7.4. We say that B (x) is the inverse of A(x) if
A(x)B (x) = 1;
1
we denote this by B (x) = A(x)−1 or by B (x) = A(x) .
Problem 1.7.5. Show that the inverse of 1 − x is
1 + x + x2 + x3 + · · ·
Solution: Note that
(1 − x)(1 + x + x 2 + · · · ) = (1 + x + x 2 + · · · ) − (x + x 2 + x 3 + · · · ) = 1,
as required.
The formal power series 1+ x + x 2 +· · · arises frequently, since it is the gener-

ating series for the non-negative integers. We call this a geometric series. The
powers of geometric series arise often in problems that we will face later.
Problem 1.7.6. Show that
Ã !
n + k − 1
(1 − x)−k = xn
X
n≥0 k − 1
Solution: We prove by induction on k. When k = 1, n+k−1

¡ ¢ ¡n ¢
k−1 = 0 = 1. So
−1 P n P ¡n+1−1¢ n
(1 − x) = n≥0 x = n≥0 k−1 x . So the base case holds.
Assume that for some positive integer m, (1 − x)−m = n≥0 n+m−1
P ¡ ¢ n
m−1 x .
We need to prove the equation for m + 1. We see that
(1 − x)−(m+1) = (1 − x)−m (1 − x)−1 .
By induction hypothesis, [x i ](1 − x)−m = i +m−1 . Also, we know that [x i ](1 −
¡ ¢
m−1
−1
x) = 1. Using rules of multiplication of power series, we get
Ã ! Ã !
n n i +m −1 n + m
[x n ](1−x)−(m+1) = ([x i ](1−x)−m )([x n−i ](1−x)−1 ) =
X X
·1 =
i =0 i =0 m −1 m
where the final step uses Problem 1.5.3. Therefore,
Ã !
n + m
(1 − x)−(m+1) = xn.
X
n≥0 m
The result holds by induction.

This identity is sometimes known as the negative binomial series.
Another important set is S = {0, 1, . . . , k}. Its generating series is
ΦS (x) = 1 + x + · · · + x k .
Note that
1 − x k+1
= (1 − x k+1 )(1 + x + x 2 + · · · )
1−x
= (1 + x + x 2 + · · · ) − (x k+1 + x k+2 + x k+3 + · · · )
= 1 + x + x2 + · · · + xk .
Thus,
1 − x k+1
= 1 + x + x2 + · · · + xk .
1−x
Problem 1.7.7. Does x have an inverse?
Solution: If B (x) = b 0 + b 1 x + b 2 x 2 + · · · is any formal power series, then
xB (x) = b 0 x + b 1 x 2 + b 2 x 3 + · · ·
If B (x) were the inverse of x then we would have
b 0 x + b 1 x 2 + b 2 x 3 + · · · = 1 + 0x + 0x 2 + · · ·
(Here we have just expanded 1 as a formal power series 1 + 0x + 0x 2 + · · · .) This

cannot be the case since the series on the left has constant term 0 while the
series on the right has constant term 1. So x does not have an inverse.
The following theorem tells us exactly which formal power series have an
inverse.
Theorem 1.7.8. A formal power series has an inverse if and only if it has a non-
zero constant term. Moreover, if the constant term is non-zero, then the inverse
is unique.
Proof: Let Q(x) be a formal power series. Then a formal power series A(x) is the
inverse of Q(x) if and only if it satisfies the linear equation
Q(x)A(x) = 1. (1.7.3)
If Q(0) 6= 0, then, by Corollary 1.7.3, Equation 1.7.3 has a unique solution. Hence
Q(x) has an inverse and it is unique. If Q(0) = 0, then the constant term on the
right side of Equation 1.7.3 is 1 but the constant term on the left is Q(0)A(0) = 0,
so there is no solution and, hence, Q(x) does not have an inverse.
As with polynomials, we can also consider the composition of formal power
series.
Definition 1.7.9. The composition of formal power series A(x) = a 0 + a 1 x +

a 2 x 2 + · · · and B (x) is defined by
A(B (x)) = a 0 + a 1 B (x) + a 2 (B (x))2 + · · ·
However, unlike for polynomials, this composition operation is not always

well defined. Consider, for example, the case that A(x) = 1 + x + x 2 + · · · and
B (x) = (1 + x). Then
A(B (x)) = 1 + (1 + x) + (1 + x)2 + · · · .

The constant term of the right-hand side has non-zero contributions from
an infinite number of terms, so A(B (x)) is not a formal power series. The fol-
lowing result shows that A(B (x)) is well-defined so long as B (x) has its constant
term equal to zero (that is, B (0) = 0).
Theorem 1.7.10. If A(x) and B (x) are formal power series with the constant
term of B (x) equal to zero, then A(B (x)) is a formal power series.
Proof: We can write A(x) = a 0 + a 1 x + a 2 x 2 + · · · and B (x) = xC (x). Then
A(B (x)) = A(xC (x))
= a 0 + a 1 xC (x) + a 2 (xC (x))2 + · · ·
= a 0 + a 1 xC (x) + a 2 x 2 (C (x))2 + · · ·
For each k ≥ 0, note that a k x k (C (x))k is a formal power series (since it is the
product of formal power series). Moreover, for each n < k, we have [x n ](a k x k (C (x))k ) =
0. Therefore,
[x n ]A(B (x)) = [x n ](a 0 + a 1 xC (x) + a 2 x 2 (C (x))2 + · · · )
= [x n ](a 0 + a 1 xC (x) + a 2 x 2 (C (x))2 + · · · + a n x n (C (x))n ).
Now, a 0 + a 1 xC (x) + a 2 x 2 (C (x))2 + · · · + a n x n (C (x))n is a formal power series
(since it is a finite sum of formal power series), so [x n ]A(B (x)) is well-defined.
We can use the formula for the inverse of 1 − x to compute other inverses.
Problem 1.7.11. Determine the inverse of 1 − x + 2x 2 .
Solution: Let P (x) = 1 − x. Note that 1 − x + 2x 2 = 1 − (x − 2x 2 ) = P (x − 2x 2 ). In
Problem 1.7.5 we showed that
P (x)−1 = 1 + x + x 2 + · · ·
Therefore
(1 − x + 2x 2 )−1 = P (x − 2x 2 )−1
= 1 + (x − 2x 2 ) + (x − 2x 2 )2 + · · ·
(We could further simplify this expression, but we only wish to demonstrate the
method of computing inverses via composition.)
Problem Set 1.7

1. Find the following coefficients:
(a) [x 8 ](1 − x)−7

(b) [x 10 ]x 6 (1 − 2x)−5
(c) [x 8 ](x 3 + 5x 4 )(1 + 3x)6
(d) [x 9 ]{(1 − 4x)5 + (1 − 3x)−2 }
(e) [x n ](1 − 2t x)−k
(f) [x n+1 ]x k (1 − 4x)−2k
(g) [x n ]x k (1 − x 2 )−m
(h) [x n ]{(1 − x 2 )−k + (1 − 7x 3 )−k }
2. (a) Write out the first 5 terms and the nth term of the following power se-
ries:
(i) (1 − x)−1
(ii) (1 + x)−1
(iii) (1 − x)−2
(iv) (1 + x)−2
(v) (1 − x)−3
(vi) (1 + x)−3
(b) Determine the appropriate coefficient in each case.
(i) [x 8 ](1 − x)−1 ,
(ii) [x 5 ](1 + x)−1 ,
(iii) [x 6 ](1 − 3x 2 )−1 ,
(iv) [x 12 ](1 − x)−2 ,
(v) [x 9 ](1 + 4x 3 )−3 ,
(vi) [x 8 ](1 − 2x 3 )−2 .
(c) Write each of the following power series in closed form.
(i) 1 − 2x + 3x 2 − 4x 3 + 5x 4 − 6x 5 + . . .
(ii) 1 + 3x + 6x 2 + 10x 3 + 15x 4 + 21x 5 + . . .
(iii) 1 − x 3 + x 6 − x 9 + x 12 − x 15 + . . .
(iv) 1 + 2x 2 + 4x 4 + 8x 6 + 16x 8 + 32x 10 + . . .
(v) 1 − 4x 2 + 12x 4 − 32x 6 + 80x 8 − 192x 10 + . . .

(vi) 1 + 6x + 24x 2 + 80x 3 + 240x 4 + 672x 5 + . . .
3. Show that
¡n+2k−1¢
(a) [x n ](1 − 2x + x 2 )−k = n
¡i +k−1¢¡ n−i
2 +k−1
(b) [x n ](1 − x − x 2 + x 3 )−k =
P ¢
i ≥0 n−i
n−i ≡0 (mod 2)
i
2
1−x 2 1
4. (a) Prove that 1+x 3
= x 2 .
1+ 1−x
(b) By expanding each side of the identity in (a) as a power series, and con-
sidering the coefficient of x N , prove that
¯ Ã !¯ ½
¯X N − k − 1 ¯
0 if N ≡ 1 (mod 3)
¯ (−1)k ¯=
¯ ¯
¯k≥0 N − 2k ¯ 1 otherwise.
5. Calculate [x n ](1 + x)−2 (1 − 2x)−2 , and give the simplest expression you can
find.
r n+r −1 m
P ¡ ¢¡ ¢ ¡m−n ¢
6. Prove that r +s=t (−1) r s = t .
7. Prove that nm=k m

P ¡ ¢ ¡n+1¢ −1 −(k+1)
k = k+1 . (Hint: Use the identity (1 − x) (1 − x) =
−(k+2)
(1 − x) .)
8. Let S be a set of configurations with weight function w. Show that for any
non–negative integer n,
ΦS (x)
[x n ]
1−x
counts the number of configurations in S with weight at most n.
9. Let A(x) and B (x) be formal power series.
(a) Show that if A(x)B (x) = 0, then A(x) = 0 or B (x) = 0. (Hint: If A(x) 6= 0
then we can write A(x) = x k Â(x) where Â(x) has a non–zero constant
term.)
(b) Show that if A(x)2 = B (x)2 , then A(x) = ±B (x).
1.8 The Sum and Product Lemmas

In this section we show that addition and multiplication of generating series
correspond to natural combinatorial constructions.
Theorem 1.8.1 (The Sum Lemma). Let (A, B ) be a partition of a set S. (That is,
A and B are disjoint sets whose union is S.) Then,
ΦS (x) = Φ A (x) + ΦB (x).
Proof:
ΦS (x) = x w(σ)
X
σ∈S
x w(σ) + x w(σ)
X X
= (since (A, B ) partitions S)
σ∈A σ∈B
= Φ A (x) + ΦB (x),
as required.
More generally, if A and B are sets where A ∩ B need not be empty, then
Φ A∪B (x) = Φ A (x) + ΦB (x) − Φ A∩B (x).
Recall that, given sets A and B , the cartesian product of A and B is the set
A × B = {(a, b) : a ∈ A, b ∈ B }.
Theorem 1.8.2 (The Product Lemma). Let A and B be sets of configurations

with weight functions α and β respectively. If w(σ) = α(a) + β(b) for each σ =
(a, b) ∈ A × B , then
Φ A×B (x) = Φ A (x)ΦB (x).
Proof:
Φ A×B (x) = x w(σ)

X
σ∈A×B
x α(a)+β(b) (since w(σ) = α(a) + β(b))
X
=
(a,b)∈A×B
Ã !Ã !
x α(a) x β(b)
X X
=
a∈A b∈B
= Φ A (x)ΦB (x),
as required.
1.8. THE SUM AND PRODUCT LEMMAS 25
More generally, if A 1 , . . . , A k are sets then A 1 × · · · × A k denotes the set of

all k–tuples (a 1 , a 2 , . . . , a k ) where a i ∈ A i for all i . Now, suppose that αi is a
weight function for A i and that w is a weight function for A 1 ×· · ·× A k . If w(σ) =
α1 (a 1 ) + · · · + αk (a k ) for each k-tuple σ = (a 1 , . . . , a k ), then
Φ A 1 ×···×A k (x) = Φ A 1 (x) · · · Φ A k (x).
We conclude this section with some applications of the Product Lemma.
Problem 1.8.3. Let k, n be fixed non-negative integers. How many solutions are
there to t 1 + · · · + t n = k, where t 1 , . . . , t n = 0 or 1?
Solution: The solutions are n-tuples (t 1 , . . . , t n ) where t i ∈ {0, 1} for each i =

1, . . . , n. So we are dealing with the set {0, 1}n . The generating series for the
solutions is
¢n
Φ{0,1}n (x) = Φ{0,1} (x)
¡
= (1 + x)n
and the number of solutions is

Ã !
k n
n
[x ](1 + x) = .
k
There is a direct combinatorial way to obtain the solution to Problem 1.8.3.

A solution (t 1 , . . . , t n ) is a sequence of 0s and 1s of length n containing exactly
k ones. The number of ways of choosing which of the k entries is 1, is nk , as
¡ ¢
expected.
Problem 1.8.4. Let S be the set of all k-tuples (a 1 , . . . , a k ) where each a i is a non-
negative integer. The weight of a k-tuple σ = (a 1 , . . . , a k ) is defined as w(σ) =
a 1 + · · · + a k . Show that ΦS (x) = (1 − x)−k .
Solution: Let N0 = {0, 1, 2, 3, . . .}. Now S = N0 × · · · × N0 = Nk0 . Then, by the

Product Lemma, ¢k
ΦS (x) = ΦN0 (x) .
¡
Now, N0 = {0, 1, 2, 3, . . .}, so

1
ΦN0 (x) = 1 + x + x 2 + · · · = .
1−x
Therefore,
µ ¶k
1
ΦS (x) = .
1−x
We have previously proved a formula involving (1 − x)−k in Problem 1.7.6.

We now give another proof by counting the number of k-tuples in S of weight
n.
Theorem 1.8.5. For any positive integer k,

Ã !
n + k − 1
(1 − x)−k = xn.
X
n≥0 k − 1
Proof: Let S n be the set of all k-tuples (a 1 , . . . , a k ) ∈ Nk0 with a 1 + · · · + a k = n. By

Problem 1.8.4,
|S n | = [x n ](1 − x)−k .
Let Tn denote the set of all binary strings of length n + k − 1 with k − 1 1’s.
Then we know that Ã !
n +k −1
|Tn | = .
k −1
We will establish a bijection between S n and Tn , and, hence, prove that
Ã !
n + k − 1
[x n ](1 − x)−k = |S n | = |Tn | = ,
k −1
as required.
We define f : S n → Tn as follows: For a k-tuple A = (a 1 , . . . , a k ) in S n ,
f (A) = 0a1 10a2 1 · · · 10ak
where 0ai represents a i 0’s in a row. Notice that f (A) is a string with exactly k −1
1’s, and the length of the string is a 1 +· · ·+a k +(k −1) = n +k −1 (note that A ∈ S n ,
hence a 1 + · · · + a k = n). Therefore f (A) ∈ Tn .
We can easily find the inverse of f : Any binary string in Tn has the form
0 10a2 1 · · · 10ak where a 1 +· · ·+a k = n. So the inverse of f is f −1 : Tn → S n where
a1
f −1 (0a1 10a2 1 · · · 10ak ) = (a 1 , . . . , a k ). This shows that f is a bijection between S n

to Tn .
1.8. THE SUM AND PRODUCT LEMMAS 27
Problem 1.8.6. Let n ∈ N0 . Suppose we have an unlimited supply of Canadian

nickels, dimes, quarters, loonies, and toonies. How many ways can we make n
cents using these coins?
Solution: Let C k = {0, k, 2k, 3k, 4k, . . .}. Each collection of coins can be repre-
sented by a 5-tuple (n, d , q, l , t ) ∈ C 5 × C 10 × C 25 × C 100 × C 200 . Let S = C 5 × C 10 ×
C 25 ×C 100 ×C 200 . We define the weight of any 5-tuple in S to be w(n, d , q, l , t ) =
n + d + q + l + t , which represents the value of this collection of coins in cents.
Using the weight function αk (a) = a for the set C k , we see that the generat-
ing series for C k is
1
ΦC k (x) = 1 + x k + x 2k + x 3k + · · · = .
1 − xk
Notice that for our set S, the weight function w(n, d , q, l , t ) = α5 (n) + α10 (d ) +
α25 (q) + α100 (l ) + α200 (t ), so we can apply the Product Lemma. The generating
series for S is then
ΦS (x) = ΦC 5 (x) · ΦC 10 (x) · ΦC 25 (x) · ΦC 100 (x) · ΦC 200 (x)

1
= .
(1 − x 5 )(1 − x 10 )(1 − x 25 )(1 − x 100 )(1 − x 200 )
The number of collections that are valued at n cents is then [x n ]ΦS (x).
Problem Set 1.8

1. Let A and B be sets of configurations with weight functions α and β respec-
tively. Now define a weight function w on A×B such that w(σ) = α(a)+2β(b)
for each σ = (a, b) ∈ A × B . Then,
Φ A×B (x) = Φ A (x)ΦB (x 2 ).
2. Extend the sum lemma to any finite number of sets. That is, if A 1 , A 2 , . . . , A k
are disjoint sets whose union is S, prove that
ΦS (x) = Φ A 1 (x) + Φ A 2 (x) + · · · + Φ A k (x) .
3. Extend the product lemma to any finite number of sets. That is, if S = B 1 ×
B 2 × · · · × B k with weight functions w 1 , . . . , w k respectively, the elements of S
being k-tuples σ = (b 1 , b 2 , . . . , b k ) with b i ∈ B i for each i , and if the weight of

σ is
w(σ) = w 1 (b 1 ) + w 2 (b 2 ) + · · · + w k (b k ),
prove that
ΦS (x) = ΦB 1 (x)ΦB 2 (x) · · · ΦB k (x).
4. (a) Prove that there are n! permutations of n distinct objects (n! = n(n −
1) · · · 2 · 1), by establishing a 1-1 correspondence between these permu-
tations and Nn × Nn−1 × · · · × N1 , where Ni = {1, 2, . . . , i }, i ≥ 1.
(b) Prove that there are nk = k!(n−k)! n!
¡ ¢
subsets of size k in a set of n dis-
tinct objects, by establishing a 1-1 correspondence between P n and
P k × P n−k × B n,k , where P i is the set of all permutations of i distinct
objects, i ≥ 1, and B n,k is the set of all k-subsets of a set of n distinct
objects.
5. List the set of all formal products of (1 + x 2 + x 4 )2 (1 + x + x 2 )2 in which the

exponents sum to 4.
6. Determine a generating series (as a product of polynomials or power series)

for the number of 5-combinations (i.e. collections of 5 elements, where 2 or
more may be alike) of the letters M, A, T, H in which M and A can appear any
number of times but T and H can appear at most once. Which coefficient in
this generating series do we want?
7. (a) Determine a generating series (as a product of polynomials or power

series) for the number of different election outcomes in an election for
class president if 27 students cast a vote among four candidates. Which
coefficient do we want?
(b) Suppose each student who is a candidate votes for herself or himself.
Now what is the generating series and the required coefficient?
(c) Repeat part (a) under the assumption that no candidate receives a ma-
jority of the votes.
8. Find a generating series (again, as a product of polynomials or power series)

for the number of integers between 0 and 999,999 which have the sum of
their digits equal to r .
Chapter 2
Compositions and Strings
2.1 Compositions of an Integer.

Definition 2.1.1. For nonnegative integers n and k, a composition of n with k
parts is an ordered list (c 1 , . . . , c k ) of positive integers c 1 , . . . , c k , such that c 1 +· · ·+
c k = n. The positive integers c 1 , . . . , c k are called the parts of the composition.
There is one composition of 0, the empty composition, which is a composition
with 0 parts.
The compositions of 4 are (4), with 1 part; (3, 1), (2, 2), (1, 3), with 2 parts;
(2, 1, 1), (1, 2, 1), (1, 1, 2), with 3 parts; and (1, 1, 1, 1), with 4 parts. Thus these are
eight compositions of 4, one with 1 part, three with 2 parts, three with 3 parts,
and one with 4 parts.
Let N = {1, 2, 3, 4, . . .} be the set of all positive integers. Then each part of a
composition is in N.
Problem 2.1.2. How many compositions of n are there with k parts, for n ≥ k ≥
1?
Solution: We let S = Nk . Then a composition (c 1 , . . . , c k ) of n with k parts is

an element of S with the restriction that c 1 + · · · + c k = n. Now let w(c 1 , . . . , c k ) =
c 1 + · · · + c k , and w(c i ) = c i for c i ∈ N, i = 1, . . . , k. Then the compositions of
n with k parts are the elements of S of weight n, and we have w(c 1 , . . . , c k ) =
w(c 1 ) + · · · + w(c k ). Thus the generating series for the number of compositions
29
30 CHAPTER 2. COMPOSITIONS AND STRINGS
of n with k parts is, from Theorem 1.6,
ΦS (x) = ΦNk (x)

= (ΦN (x))k , from Product Lemma
Ã !k
X i
= x
i ≥1
¢k
= x(1 − x)−1
¡
, from geometric series
Therefore, the number of compositions of n with k parts is
[x n ]x k (1 − x)−k = [x n−k ](1 − x)−k

Ã !
n −k +k −1
=
k −1
Ã !
n −1
= , for n ≥ k ≥ 1.
k −1
¡n−1¢
Thus there are k−1
compositions of n with k parts, for n ≥ k ≥ 1.
The solution method for Problem 2.1.2 adapts easily to many other prob-
lems involving compositions. In general, we will be asking questions in the
form of “how many compositions of n has such and such properties?” And the
way we are going to answer them is to follow these steps:
1. Find the set S of all compositions that satisfy these properties (without
regard for what n is);
2. Find the generating series for S (the weight of a composition is the sum
of its parts, so we can apply the Sum and Product Lemmas); and
3. The answer to our question is the coefficient of x n in ΦS (x). Find an ex-

plicit formula for this coefficient if possible.
Problem 2.1.3. How many k-part compositions of n are there in which each
part is an odd number?
Solution: In this case let S = Nkod d , where Nod d = {1, 3, 5, 7, . . .}. Then the re-
quired compositions are the elements of S of weight n, so the required generat-
2.1. COMPOSITIONS OF AN INTEGER. 31
ing series is
ΦS (x) = ΦNk (x)

od d
¢k
= ΦNod d (x) , by Product Lemma
¡
Ã !k
X 2i +1
= x
i ≥0
= (x(1 − x 2 )−1 )k , by Geometric Series
It follows that the number of compositions of the required form is
[x n ](x(1 − x 2 )−1 )k = [x n ]x k (1 − x 2 )−k

= [x n−k ](1 − x 2 )−k
Ã !
n−k
X i + k − 1 2i
= [x ] x .
i ≥0 k −1
The coefficient is zero if n−k is odd. If n−k is even then the required coefficient
occurs for i = n−k
2
, and is
Ã ! Ã !
n−k n+k−2
2
+k −1 2
= .
k −1 k −1
Thus this is the number of k-part compositions of n in which each part is odd.
Problem 2.1.4. How many compositions of n are there with k parts, each part
at most 5?
Solution: In this case let S = {1, 2, 3, 4, 5}k , and the required compositions are
the elements of S of weight n. Thus the required generating series is
ΦS (x) = Φ{1,2,3,4,5}k (x)

¢k
= Φ{1,2,3,4,5} (x) , by Product Lemma
¡
= (x 1 + x 2 + x 3 + x 4 + x 5 )k
¢k
= x(1 − x 5 )(1 − x)−1 , by finite Geometric Sum
¡
= x k (1 − x 5 )k (1 − x)−k .
Therefore, the number of compositions is
Ã ! Ã !
k j + k − 1
[x n ]x k (1 − x 5 )k (1 − x)−k = [x n−k ] (−x 5 )i xj,
X X
i ≥0 i j ≥0 k − 1
Ã ! Ã !
k j + k − 1
= [x n−k ] (−1)i x 5i + j
X X
i ≥0 j ≥0 i k − 1
Ã ! Ã !
X k i j +k −1
= (−1)
n
i ≥0, j ≥0 i k −1
5i + j =n−k
b n−k
5 c
Ã ! Ã !
X k i n − k − 5i + k − 1
= (−1) ,
i =0 i k −1
j k
n−k n−k
where 5
is the “integer part" of 5
(we use j = n − k − 5i , so j ≥ 0 gives
n−k
i≤ 5 ).
We also have instances when the number of parts in a composition is not

specified.
Problem 2.1.5. How many compositions of n are there, n ≥ 0?
Solution: Here we let S = ∪k≥0 Nk , where the term k = 0 is the empty compo-
sition (of 0). Then the compositions of n are elements of S of weight n, so the
generating series is
ΦS (x) = Φ∪k≥0 Nk (x)

ΦNk (x), by Sum Lemma
X
=
k≥0
(ΦN (x))k , by Product Lemma
X
=
k≥0
X ³X ´k
= xi
k≥0 i ≥1
¢k
x(1 − x)−1
X¡
= , by Geometric Series
k≥0
1
= , by Geometric Series
1 − x(1 − x)−1
1−x
= , multiplying by 1 − x on top and bottom
1−x −x
1 − 2x + x
=
³ 1 − 2xx ´
= 1+ .
1 − 2x
Hence the number of compositions is
Ã !
³ x ´
[x n ] 1 + = [x n ] 1 + x (2x)i , by Geometric Series
X
1 − 2x i ≥0
Ã !
= [x n ] 1 + 2i x i +1
X
i ≥0
(
1, n = 0;
=
2n−1 , n ≥ 1
(We use i + 1 = n, so i = n − 1, for n ≥ 1.) Thus the number of compositions of n

is 1 for n = 0, and 2n−1 for n ≥ 1.
Sometimes the generating series might not give us an explicit formula for
the set of compositions that we are counting. But they may give us additional
information regarding the structure of these compositions.
Problem 2.1.6. How many compositions of n are there where each part is odd?
(The number of parts is not restricted.)
Solution: Let Nod d = {1, 3, 5, 7, . . .}. The set of all compositions where each part
is odd is [ k
S= Nod d .
k≥0
The generating series of Nod d is

x
ΦNod d (x) = x + x 3 + x 5 + · · · = .
1 − x2
By the product lemma, the generating series for compositions with k odd parts
is then ³ x ´k
ΦNk (x) = .
od d 1 − x2
By the sum lemma, the generating series for our problem is
ΦS (x) = Φ Nk
X
od d
k≥0
X³ x ´k
= 2
k≥0 1 − x
1
= x by Geometric Series
1 − 1−x 2
1 − x2
= .
1 − x − x2
2
1−x
The answer to our question is [x n ] 1−x−x 2.
Here we could not use existing tools to figure out an explicit formula for
1−x 2
the coefficient of 1−x−x 2 . But let’s derive a recurrence based on the method
described in Problem 1.7.1. If we let ΦS (x) = n≥0 a n x n , then we get
P
1 − x 2 = (1 − x − x 2 ) a n x n = a 0 + (a 1 − a 0 )x + (a n − a n−1 − a n−2 )x n .
X X
n≥0 n≥2
By comparing coefficients on both sides, we get
a 0 = 1, a 1 = 1, a 2 = 1, a n = a n−1 + a n−2 for n ≥ 3.
Other than a 0 , the rest are terms in the Fibonacci sequence. So for n ≥ 1, the
number of compositions of n where each part is odd is the n-th term in the
Fibonacci sequence.
We will now give a combinatorial proof of the recurrence. Let S n be the set
of all compositions of n where each part is odd. The recurrence a n = a n−1 +a n−2
for n ≥ 3 means that |S n | = |S n−1 | + |S n−2 |. So we can prove the recurrence by
finding a bijection between S n and S n−1 ∪ S n−2 . Here is one possible bijection.
Let f : S n → S n−1 ∪ S n−2 where for each composition (a 1 , . . . , a k ) ∈ S n ,
½
(a 1 , . . . , a k−1 ) ak = 1
f (a 1 , . . . , a k ) =
(a 1 , . . . , a k−1 , a k − 2) a k ≥ 3
This means that if the last part is 1, we remove that part to get a composition of
n − 1; if the last part is 3 or more, we subtract 2 from it to get a composition of
n − 2. Either composition still consist of parts that are odd, so they are in S n−1
and S n−2 respectively.
This bijection is reversible. For the inverse f −1 : S n−1 ∪ S n−2 → S n ,

½
−1 (b 1 , . . . , b l , 1) b1 + · · · + bl = n − 1
f (b 1 , . . . , b l ) =
(b 1 , . . . , b l −1 , b l + 2) b 1 + · · · + b l = n − 2
So f is a bijection, and this proves that |S n | = |S n−1 ∪ S n−2 | = |S n−1 | + |S n−2 |.

One nice consequence of this bijection is that it gives a method to recur-
sively build the set of all possible compositions of n with parts that are odd.
Suppose we have built up all such compositions up to n − 1, we can get all such
compositions for n by adding a new part 1 to every composition of n − 1, and
adding 2 to the last part of every composition of n −2. Here are some examples:
S 3 = {(3), (1, 1, 1)}

S 4 = {(3, 1), (1, 3), (1, 1, 1, 1)}
S 5 = {(5), (1, 1, 3), (3, 1, 1), (1, 3, 1), (1, 1, 1, 1, 1)}
S 6 = {(3, 3), (1, 5), (1, 1, 1, 3),
(5, 1), (1, 1, 3, 1), (3, 1, 1, 1), (1, 3, 1, 1), (1, 1, 1, 1, 1, 1)}
Problem Set 2.1

1. How many compositions of n are there where each part is at most m? (The
number of parts is not restricted.)
2. Let k be a fixed integer. How many compositions of n with k parts are there
where each part is congruent to 1 modulo 5? Determine an explicit formula.
3. Let n be a non-negative integer. How many compositions of n are there

where the i -th part has the same parity as i ? For example, compositions
of 7 that satisfy this condition are
(7), (5, 2), (3, 4), (1, 6), (1, 2, 1, 2, 1).
4. How many k-tuples (a 1 , . . . , a k ) of positive integers satisfy the inequality a 1 +

· · · + a k < n?
5. Let a n denote the number of compositions of n. From Problem 2.1.2, we see

that a n = 2n−1 for n ≥ 1. This tells us that for n ≥ 2, a n satisfies the recurrence
a n = 2a n−1 . Give a combinatorial proof of this recurrence.
6. Let n be a non-negative integer, and let S n be the set of all compositions of

n where each part is greater than 1. (The number of parts is not restricted.)
Let a n = |S n |.
(a) Prove that

1−x
a n = [x n ] .
1 − x − x2
(b) Derive a recurrence that a n satisfies together with sufficient initial con-
ditions
(c) Give a combinatorial proof of the recurrence from part (b).
7. Let n ∈ N. Consider the problem of finding the number of compositions of

n with exactly 3 parts (a 1 , a 2 , a 3 ) such that 1 ≤ a 1 < a 2 < a 3 . For example,
when n = 9, there are three such compositions: (1, 2, 6), (1, 3, 5), (2, 3, 4). Let
S be the set of all such compositions, i.e. S = {(a 1 , a 2 , a 3 ) | 1 ≤ a 1 < a 2 < a 3 }.
We will determine the generating series of S with the help of another set
T = N × N × N, which is the set of all compositions with exactly 3 parts.
(a) Define a bijection f : S → T , and write down its inverse f −1 . Illustrate

your bijection by determining f (2, 3, 9) and f −1 (3, 1, 4).
(b) Let w be the weight function on S where w(a 1 , a 2 , a 3 ) = a 1 +a 2 +a 3 . For
each (b 1 , b 2 , b 3 ) ∈ T , define a weight w ∗ (b 1 , b 2 , b 3 ) such that w ∗ ( f (a 1 , a 2 , a 3 )) =
w(a 1 , a 2 , a 3 ) for all (a 1 , a 2 , a 3 ) ∈ S.
(c) Determine the generating series of T with respect to w ∗ , and explain
why this is the same as the generating series of S with respect to w.
2.2 Subsets with Restrictions.

There are many combinatorial problems that can be solved using the product
lemma, but frequently one has to recast the problem first in order for it to meet
the requirements of the lemma. The next problem is an example of this.
We have already noted ¡ ¢ that the number of k-subsets (i.e. subsets of size k)
of Nn = {1, 2, . . . , n} is nk . Now consider the related problem:
Problem 2.2.1. How many k-subsets of Nn are there which contain no two con-
secutive integers?
2.2. SUBSETS WITH RESTRICTIONS. 37
(For example, when n = 6 and k¡=¢ 3, these subsets are {1,3,5}, {1,3,6}, {1,4,6}
and {2,4,6} - considerably less than 63 = 20.)
To specify an example of such a subset we can list its elements in increasing
order, as above, say as a vector α = (α1 , α2 , . . . , αk ) where α1 < α2 < · · · < αk ; but
this is not enough. We must also indicate the value of n. (The subsets listed
above, for example, could just as well be subsets of N7 or N519 ). Accordingly we
define a set C k consisting of elements (α, n) where n is a positive integer and
α represents a subset of Nn in the manner just described. For example, when
k = 3, then ({1, 3, 5}, 6), ({1, 4, 6}, 6) and ({1,4,6},13) are all in C 3 (though ({1,3,5},4)
and ({2,5,6},10) are not).
Suppose that we let the weight of σ = (α, n) be w(σ) = n. Then Problem
2.2.1 can be rephrased as
Restatement of Problem 2.2.1: How many configurations of weight n are there
in C k ?
Thus the required number is [x n ]ΦC k (x), and we show how the product
lemma can be used to obtain ΦC k (x).
For an arbitrary element σ = ({α1 , . . . , αk }, n) in C k consider the vector of
differences d (σ) = (α1 , α2 − α1 , . . . , αk − αk−1 , n − αk ). Thus for σ = ({1, 3, 6}, 6),
we have d (σ) = (1, 3 − 1, 6 − 3, 6 − 6) = (1, 2, 3, 0). Note that σ can be uniquely
recovered from d (σ). In the present example, σ = ({1, 1 + 2, 1 + 2 + 3}, 1 + 2 +
3 + 0) = ({1, 3, 6}, 6), and in general, if d (σ) = (d 1 , . . . , d k+1 ), then σ = ({d 1 , d 1 +
d 2 , . . . , d 1 + · · · + d k }, d 1 + · · · + d k+1 ). Note that d 1 is any positive integer, d k+1
is any non-negative integer, and d 2 , . . . , d k are positive integers greater than 1.
(Since we don’t allow any pairs of consecutive integers). Thus there is a 1-1 cor-
respondence, or bijection, between elements of C k of weight n and solutions to
t 1 +· · ·+t k+1 = n, with t 1 ≥ 1, t k+1 ≥ 0, t 2 , . . . , t k ≥ 2. Thus the required generating
series is
ΦC k (x) = ΦN (x){ΦN≥2 (x)}k−1 ΦN≥0 (x)
= (x + x 2 + . . .) (x 2 + x 3 + . . .)k−1 (1 + x + x 2 + . . .)
↑ ↑ ↑
Power of x gives Power of x gives Power of x gives
smallest element of difference between difference between (2.2.1)
subset successive pairs in n and largest
subset element in subset
of Nn
= x(1 − x)−1 x 2(k−1) (1 − x)−(k−1) (1 − x)−1 , by the Geometric series.

Hence the required number is
[x n ]x 2k−1 (1 − x)−(k+1) = [x n−2k+1 ](1 − x)−(k+1)

Ã !
n − 2k + 1 + (k + 1) − 1
=
n − 2k + 1
Ã !
n −k +1
=
n − 2k + 1
Ã !
n −k +1
= .
k
To check a special case, we found 4 subsets when n = 6 and k = 3 and

Ã ! Ã ! Ã !
n −k +1 6−3+1 4
= = = 4,
k 3 3
which is the number of subsets predicted by our formula.

This seems like a complicated way to obtain such a simple answer (Q: How
does one obtain this formula by a “direct” argument?), but an examination of
(2.2.1) allows us to solve any question involving differences of successive ele-
ments in subsets. In the following problems let α = {α1 , . . . , αk } be a subset of
Nn with α1 < . . . < αk .
Problem 2.2.2. (The Terquem Problem) How many subsets α of Nn are there,
which have size k, and in which αi ≡ i (mod 2) for i = 1, . . . , k? (i.e. the odd
positions in the subset contain odd numbers, and the even positions contain
even numbers.)
Solution: The first element in such a subset must be odd, and the difference
between successive pairs must also be odd. The difference between n and the
largest element in the subset can be any non-negative integer. Thus we modify
(2.2.1) to obtain the required number as
[x n ](x + x 3 + x 5 + . . .)(x + x 3 + x 5 + . . .)k−1 (1 + x + x 2 + . . .)

= [x n ]x k (1 + x 2 + x 4 + . . .)k (1 + x + x 2 + . . .)
= [x n−k ](1 − x 2 )−k (1 − x)−1
= [x n−k ](1 − x 2 )−k (1 − x 2 )−1 (1 + x)
= [x n−k ](1 − x 2 )−(k+1) (1 + x)
2.2. SUBSETS WITH RESTRICTIONS. 39
Ã !
n−k
X k + i 2i
= [x ](1 + x) x
i ≥0 i
Ã !
n−k
X k +i
= [x ] (x 2i + x 2i +1 )
i ≥0 i
Ã !
k +i
= , if n − k = 2i or 2i + 1 .
i
The two cases can be unified by using the notation b n−k2 c, the “integer part”
of 2 , which is defined to be the greatest integer which is not larger than n−k
n−k
2
.
n−k
Thus i = b 2 c both when n − k = 2i and n − k = 2i + 1. The answer to problem
¡b n−k
2 c+k
¢
2.2.2 is therefore n−k . When n = 7, k = 3, the subsets are {1,2,3}, {1,2,5},
b 2 c
{1,2,7}, {1,4,5},{1,4,7}, {1,6,7}, {3,4,5},{3,4,7},{3,6,7},
¡2+3 ¢ ¡5¢ {5,6,7}, which verifies the for-
mula, since b 7−32
c = 2, and 2
= 2
= 10.
Problem 2.2.3. How many subsets α of Nn are there, which have size k, and in
which the difference between successive pairs of elements is at least c, but less
than d , where c < d ? (i.e. c ≤ αi +1 − αi < d for i = 1, . . . , k − 1).
Solution: The required generating series, from modifying (2.2.1) is:
(x + x 2 + . . .)(x c + x c+1 + · · · + x d −1 )k−1 (1 + x + x 2 + . . .)
= x(1 − x)−1 {(x c − x d )(1 − x)−1 }k−1 (1 − x)−1 ,
by summing a partial geometric series. Hence the required number is
[x n ]x c(k−1)+1 (1 − x)−(k+1) (1 − x d −c )k−1

¡k+i ¢ j k−1 (d −c) j
= [x n−c(k−1)−1 ] xi
P P ¡ ¢
i ≥0 i j ≥0 (−1) j
x
j k+i k−1
P ¡ ¢¡ ¢
= i +(d −c) j =n−c(k−1)−1 (−1) i j
.
This answer is not as compact as the two previous ones but is still a finite
summation with single index j (since i can be determined from j by the re-
striction relating i and j ), and is easier to calculate than actually counting the
subsets.
Problem Set 2.2

1. How many ¡n ¢subsets of Nn are there with size k? (Note: We know that the
answer is k , so this problem is included to provide a check on our method).
2. How many subsets of Nn are there which have size k, and in which the dif-
ference between successive pairs of elements is at least c?
3. How many subsets of Nn are there which have size k, contain no consecutive
integers, and in which the i t h smallest element is congruent to i ( mod 5), for
i = 1, . . . , k?
4. Show that the number of subsets α of Nn , of size k, with αi +1 − αi ≥ i + 1 for

¡ n−(k2) ¢
i = 1, . . . , k − 1, is .
n−(k+1
2 )
5. Show that the number of subsets of Nn , which have size k, and in which the
difference between successive pairs of elements is never equal to 2, is
Ã !Ã !
k−1
X k −1 n −k −i +1
.
i =0 i n − 2i − k
6. Show that the number of subsets of N2m of size k, in which the difference
between any pair of elements is never equal to 2, is
Ã !Ã !
X k m −i +1 m −k +i +1
.
i =0 i k −i
(Hint: Consider the odd elements and even elements separately.)
2.3 Binary Strings

Definition 2.3.1. A binary string or {0,1}-string is a string of 0’s and 1’s, its
length is the number of occurrences of 0 and 1 in the string.
For example, 01101 is a binary string of length five. We will use `(α) to de-
note the length of a string α. There is a single string, ε, of length 0, called the
empty string. Now there are clearly 2n binary strings of length n, since there
are 2 choices, 0 or 1, for each of the n symbols in a string of length n. In this
section, we count the number of binary strings of length n subject to various
2.3. BINARY STRINGS 41
restrictions. For example, we might consider strings that have no three consec-
utive ones. Since we will always be interested in the number of strings of a given
length, the generating series for strings in this section will all use the length of
a string as the weight function. Thus, for a set A of binary strings, we have
X `(a)
Φ A (x) = x .
a∈A
We will build up longer binary strings by piecing together short binary strings.
If a 1 and a 2 are binary strings, then a 1 a 2 is the string formed by concatenating
a 1 and a 2 . For example, if a 1 = 0100 and a 2 = 001, then a 1 a 2 = 0100001. Further,
if A and B are sets of binary strings, we define
AB = {ab : a ∈ A, b ∈ B }.
We also define
A ∗ = {ε} ∪ A ∪ A A ∪ A A A ∪ · · ·
= {ε} ∪ A ∪ A 2 ∪ A 3 ∪ · · · .
So A ∗ is the set of strings formed by concatenating any number of strings in A.
For example, in this notation, the set of all binary strings is given by {0, 1}∗ .
For {0, 1}-strings a, b, we say that b is a substring of a if a = cbd for some
{0, 1}-strings c, d . Our problems will often be expressed in terms of the blocks
of a binary string, which are maximal nonempty substrings consisting of only
0’s or only 1’s. For example, the string 00111001 has four blocks: (00, 111, 00, 1).
Problem 2.3.2. Let a n be the number of {0, 1}-strings of length n which contain
no three consecutive 1’s (or, which contain no substring “111”). Prove that
1 + x + x2
a n = [x n ] , for n ≥ 0.
1 − x − x2 − x3
Solution: Let S be the set of {0, 1}-strings with no three consecutive 1’s. We will
decompose each string by decomposing it after every occurrence of 0. For ex-
ample, the string σ1 = 00011010100 decomposes into (0, 0, 0, 110, 10, 10, 0) and
the string σ2 = 10111001 decomposes into (10, 1110, 0, 1). This decomposition
has two important properties. First, each string is decomposed uniquely. Sec-
ond, the decomposition does not separate consecutive ones, so a string is in S
if and only if no piece in the decomposition contains three consecutive ones.
So, just by examining the pieces, we see that S contains σ1 but not σ2 . When
we decompose a string in S, each piece will be in {0, 10, 110}, except possibly for
the last peice which can be 1 or 11. So we can write S as
S = {0, 10, 110}∗ {ε, 1, 11}.
Since the length is additive under composition, we can use this expression for
S together with the Sum and Product Lemmas to compute ΦS (x). Note that
Φ{0,10,110} (x) = x + x 2 + x 3 , and

Φ{ε,1,11} (x) = 1 + x + x 2 .
Moreover,
{0, 10, 110}∗ = {ε} ∪ {0, 10, 110} ∪ {0, 10, 110}2 ∪ · · · .
So
¢2
Φ{0,10,110}∗ (x) = 1 + Φ{0,10,110} (x) + Φ{0,10,110} (x) + · · ·
¡
1
= .
1 − Φ{0,10,110} (x)
Therefore
1
ΦS (x) = Φ{ε,1,11} (x)
1 − Φ{0,10,110} (x)
1 + x + x2
= ,
1 − x − x2 − x3
as required.
2.4 Unambiguous Expressions

The solution to Problem 2.3.2 glosses over a subtle but important point. We
used the Product Lemma to expand Φ AB (x), but the Product Lemma concerns
the expansion of Φ A×B (x). The sets A × B and AB are closely related:
AB = {ab : (a, b) ∈ A × B }.
So concatenation defines a map from A × B to AB . The problem lies in the

fact that this map need not be a bijection; that is, there may be distinct pairs
(a 1 , b 1 ) and (a 2 , b 2 ) in A ×B with a 1 b 1 = a 2 b 2 . Moreover, when this happens, we
2.5. SOME DECOMPOSITION RULES 43
will not have Φ AB (x) = Φ A (x)ΦB (x). For example, if A = {0, 01} and B = {0, 10},
then A × B = {(0, 0), (0, 10), (01, 0), (01, 10)} and AB = {00, 010, 0110}. So Φ AB (x) =
x 2 + x 3 + x 4 whereas Φ A (x)ΦB (x) = Φ A×B (x) = x 2 +2x 3 + x 4 . The problem here is
that the string 010 was created by concatenating two distinct pairs, (0, 10) and
(01, 0), in A × B .
We say that the expression AB is ambiguous if there exist distinct pairs
(a 1 , b 1 ) and (a 2 , b 2 ) in A × B with a 1 b 1 = a 2 b 2 —otherwise we say that AB is
an unambiguous expression. In other words, AB is unambiguous if each
string uniquely decomposes into a string in A concatenated with a string in
B . Note that if A and B are finite sets, then AB is unambiguous if and only if
|AB | = |A × B |.
We will use similar terminology for other constructions of binary strings.
For example, the expression A ∪ B is unambiguous when A ∩ B = ;. More gen-
erally, we will compose expressions for sets of strings using concatenation and
union of smaller sets; such an expression is unambiguous when each concate-
nation operation and each union operation is unambiguous. Thus A ∗ is unam-
biguous if the sets {ε}, A, A 2 , . . . are disjoint and, for each i ∈ N≥0 , A i is unam-
biguous.
It is important to note that it is the description of the set of strings using
sum, concatenation and the ∗-operation which might be ambiguous. A set of
strings itself is never ambiguous or unambiguous.
2.5 Some Decomposition Rules

Let’s return briefly to the solution of Problem 2.3.2. We expressed the set S as
{0, 10, 110}∗ {ε, 1, 11}.
It is not difficult to show directly that this expression is unambiguous, however,

the reason that it is unambiguous is implicit in our solution. We considered
the decomposition of a binary string into a sequence of substrings, obtained by
breaking the string after each occurrence of 0. The expression above is gener-
ating the pieces of that decomposition. Since the decomposition rule is unam-
biguous, the expression is unambiguous.
Here are some of the common decomposition rules for the set S of all binary
strings.
• Decompose a string after each 0 or 1. Each piece in the decomposition

will be either 0 or 1. This gives rise to the expression
S = {0, 1}∗ .
• Decompose a string after each occurrence of 0. Each piece in the decom-

position will be from {0, 10, 110, . . .} = {1}∗ {0} except possibly for the last
piece which may consist only of 1s. This gives rise to the expression
S = ({1}∗ {0})∗ {1}∗ .
• Decompose a string after each block of 0s. Each piece in the decomposi-
tion, except possibly the first and last pieces, will consist of a block of 1s
followed by a block of 0s. The first piece may consist only of a block of 0s
and the last piece may consist only of a block of 1s. This gives rise to the
expression
S = {0}∗ ({1}{1}∗ {0}{0}∗ )∗ {1}∗ .
Moreover, since each of these decomposition rules is unambiguous, the corre-

sponding expressions are unambiguous.
2.6 Sum and Product Rules for Strings

If A and B are sets of strings and A ∩ B = ;, then
Φ A∪B (x) = Φ A (x) + ΦB (x),
as we would hope. We also have the following.
Theorem 2.6.1. Let A, B be sets of {0, 1}-strings.
(a) If the expression AB is unambiguous, then
Φ AB (x) = Φ A (x)ΦB (x).
(b) If the expression A ∗ is unambiguous, then
Φ A ∗ (x) = (1 − Φ A (x))−1 .
2.7. DECOMPOSITION USING BLOCKS 45
Proof:
(a)
Φ AB (x) = x `(s)
X
s∈AB
x `(ab)
X X
=
a∈A b∈B
since s = ab for unique a ∈ A, b ∈ B . But length (ab) = length (a)+ length

(b) for all {0, 1}-strings a, b, so
X X `(a)+`(b)
Φ AB (x) = x
a∈A b∈B
X `(a)
x `(b)
X
= x
a∈A b∈B
= Φ A (x)ΦB (x).
(b)
Φ A ∗ (x) = Φ A k (x),
X
k≥0
by the Sum Lemma, since A ∗ is unambiguous. Note that ε 6∈ A, so the
constant term of Φ A (x) is 0. Now Φ A k (x) = (Φ A (x))k from part (a), so
Φ A ∗ (x) = (Φ A (x))k = (1 − Φ A (x))−1

X
k≥0
by Geometric Series.
2.7 Decomposition Using Blocks

The following problems involve considering strings subject to a condition on
their blocks.
Problem 2.7.1. Let b n be the number of {0, 1}-strings of length n in which no

block has length exactly two. Prove that
1 − x2 + x3
b n = [x n ] , n ≥ 0.
1 − 2x + x 2 − x 3
Solution: Let B be the set of all {0, 1}-strings in which no block has length ex-
actly two. Consider decomposing a string after each block of 1s. Except possibly
for the first and last piece, each piece has the form
{0, 000, 0000, · · · }{1, 111, 1111, · · · }.
The first piece may consist only of 1s and the last piece may consist only of 0s.
This gives the expression
B = {ε, 1, 111, 1111, · · · }({0, 000, 0000, · · · }{1, 111, 1111, · · · })∗ {ε, 0, 000, 0000, · · · }.
Moreover, since the decomposition rule is unambiguous, the expression is un-

ambiguous. So
(1 + x + x 3 + x 4 + · · · )2
ΦB (x) =
1 − (x + x 3 + x 4 + · · · )2
3
x
(1 + x + 1−x )2
= 3
, by geometric series
x
1 − (x + 1−x )2
3
x
1 + x + 1−x (1 + A)2 1 + A
= , since =
x
1 − (x + 1−x
3
) 1 − A2 1− A
1 − x2 + x3
= , multiplying top and bottom by 1 − x.
1 − 2x + x 2 − x 3
Now b n = [x n ]ΦB (x), and the result follows.
Problem 2.7.2. Find the generating series with respect to length for {0, 1}-strings
in which an odd block of 0’s is never followed by an odd block of 1’s.
Solution: Let S be the set of all {0, 1}-strings in which an odd block of "0"s is
never followed by an odd block of "1"s. We decompose a binary string after each
block of 1s, so a string has the required property if and only if each of the pieces
has. Decomposing all {0, 1}-strings in this way gives rise to the unambiguous
expression
{0, 1}∗ = {1}∗ ({0}{0}∗ {1}{1}∗ )∗ {0}∗ .
We are excluding pieces of the form {0}{00}∗ {1}{11}∗ . So
S = {1}∗ M ∗ {0}∗
2.8. RECURSIVE DECOMPOSITIONS OF BINARY STRINGS 47
where
M = {0}{0}∗ {1}{1}∗ \ {0}{00}∗ {1}{11}∗ .
Thus
x x x x
ΦM (x) = −
1 − x 1 − x 1 − x2 1 − x2
x 3 (2 + x)
=
(1 − x 2 )2
and the required generating series is
¶−1
x 3 (2 + x)
µ
1 1
Φ 1∗ M ∗ 0∗ = 1− 2 2
1−x (1 − x ) 1−x
2
(1 + x)
= .
1 − 2x 2 (1 + x)
2.8 Recursive Decompositions of Binary Strings

Another type of decomposition that is often useful is a recursive decomposi-
tion, in which a set is decomposed in terms of itself. For example, let S be the
set of all {0, 1}-strings. We can define S in the following way.
(i) The empty string is in S.
(ii) Any other element of S consists of a symbol (either 0 or 1) followed by an

element of S.
Such a definition, in which the set S is defined in terms of itself, is called a re-
cursive definition. This leads immediately to the recursive decomposition
S = {ε} ∪ {0, 1}S.
Here the right side gives an unambiguous expression for S and so from this we
obtain
ΦS (x) = 1 + Φ{0,1} (x)ΦS (x).
Solving for ΦS (x) gives
1 − Φ{0,1} (x) ΦS (x) = 1,
¡ ¢
so ¢−1 X n n
ΦS (x) = 1 − Φ{0,1} (x) = (1 − 2x)−1 =
¡
2 x ,
n≥0
n
as expected (since there are 2 binary strings of length n).
Below we give a different solution to Problem 2.3.2 based on a recursive de-
composition; we restate the problem here for convenience.
Problem 2.8.1. Let a n be the number of {0, 1}-strings of length n which contain
no three consecutive 1’s. Prove that
1 + x + x2
a n = [x n ] , for n ≥ 0.
1 − x − x2 − x3
Solution: Let S be the set of all {0, 1}-strings that do not contain three con-
secutive 1s. Consider decomposing a string after the first occurrence of 0. The
strings {ε, 1, 11} are indecomposable. Every other string σ ∈ S can be written as
σ = σ1 σ2 where σ1 ∈ {0, 10, 110} and σ2 ∈ S. This gives rise to the expression
S = {², 1, 11} ∪ {0, 10, 110}S.
Moreover the right hand side is an unambiguous expression, so
ΦS (x) = (1 + x + x 2 ) + (x + x 2 + x 3 )ΦS (x).
Solving this gives

1 + x + x2
ΦS (x) = .
1 − x − x2 − x3
2
1+x+x
Hence a n = 1−x−x 2 −x 3 , as required.
In the previous problem we forbade the substring 111. The technique used
does not adapt well to more complicated forbidden substrings, but such prob-
lems can be handled using the method outlined in the following two problems.
Problem 2.8.2. Let L denote the set of binary strings which do not contain 11010
as a substring. Find the generating series for L.
Solution: Let M be the set of binary strings that contain exactly one copy of
11010, at their right end or (in other words) as a suffix. We have two equations:
{ε} ∪ L{0, 1} = L ∪ M ,
L{11010} = M .
For the first equation, if we append a bit to ε or a string in L, the resulting string
is either in L or M , which shows that {ε} ∪ L{0, 1} ⊆ L ∪ M . To show the other
inclusion, let α ∈ L ∪ M . If α ∈ L, then α is either the empty string, or we can
remove the last bit to obtain another string in L, so α ∈ L{0, 1}. If α ∈ M , then
by removing the last bit from α, it destroys the only copy of 11010 in α. The
remaining string is in L, so α ∈ L{0, 1} as well. This shows that {ε}∪L{0, 1} ⊇ L∪M ,
hence equality holds.
For the second equation, it should be clear that if α ∈ M , then α ∈ L{11010},
which shows that L{11010} ⊇ M . On the other hand, if α ∈ L, then α11010 cer-
tainly contains 11010 as a substring, but it takes a little effort to verify that it
does not contain a second occurrence of 11010. Since α does not contain 11010,
a second copy of 11010 would have to use some digits from the end of α and
some digits from the start of the appended sequence 11010. We can consider
cases depending upon the size of the overlap of the two copies of 11010.
µ ¶
∗ ∗ ∗ ∗ 1 1 0 1 0
• Overlap in 1 digit: · · · . This cannot hap-
1 1 0 1 0 ∗ ∗ ∗ ∗
pen, since 1 6= 0.
µ ¶
∗ ∗ ∗ 1 1 0 1 0
• Overlap in 2 digits: · · · . This cannot happen,
1 1 0 1 0 ∗ ∗ ∗
since 11 6= 10.
µ ¶
∗ ∗ 1 1 0 1 0
• Overlap in 3 digits: · · · . This cannot happen,
1 1 0 1 0 ∗ ∗
since 110 6= 010.
µ ¶
∗ 1 1 0 1 0
• Overlap in 4 digits: · · · . This cannot happen, since
1 1 0 1 0 ∗
1101 6= 1010.
This shows that L{11010} ⊆ M , hence the second equation holds.

If we convert our two equations to generating series, we get
1 + ΦL (x)Φ{0,1} (x) = ΦL (x) + ΦM (x)

ΦL (x)Φ{11001} (x) = ΦM (x).
Consequently
1 + 2xΦL (x) = ΦL (x) + ΦM (x)

x 5 ΦL (x) = ΦM (x).
Substituting the second equation into the first equation gives
1 + 2xΦL (x) = ΦL (x) + x 5 ΦL (x).
Therefore
1
ΦL (x) = .
1 − 2x + x 5
Problem 2.8.3. Let L denote the set of binary strings which do not contain 1010
as a substring. Find the generating series for L.
Solution: Let M be the set of binary strings that contain exactly one copy of
1010, at their right end. We have two equations:
{ε} ∪ L{0, 1} = L ∪ M ,
L{1010} = M ∪ M {10}.
The argument to establish the first equation is similar to the case in Problem
2.8.2. For the second equation, it should be clear that L{1010} ⊇ M ∪ M {10}. On
the other hand, if α ∈ L, then α1010 certainly contains 1010 as a substring, but
it can contain two copies. For example, 0110010 ∈ L and 01100101010 contains
a second copy of 1010 as shown: 01100(1010)10. More generally, if α ∈ L and
α1010 contains two copies of 0101, then α1010 ends with 101010 and if we drop
the last 10 we get a string in M .
If we convert our two equations to generating series, we get
1 + ΦL (x)Φ{0,1} (x) = ΦL (x) + ΦM (x)

ΦL (x)Φ{1001} (x) = ΦM (x) + ΦM (x)Φ{10} .
Consequently
1 + 2xΦL (x) = ΦL (x) + ΦM (x)

x 4 ΦL (x) = ΦM (x) + x 2 ΦM (x).
From the second equation here
x4
ΦM (x) = ΦL (x)
1 + x2
and if we substitute this into the first equation, we have
1 = ΦL (x) − 2xΦL (x) + ΦM (x)

x4
µ ¶
= 1 − 2x + ΦL (x)
1 + x2
and therefore
1
ΦL (x) = 4
.
x
1 − 2x + 1+x 2
Problem Set 2.8

1. Let A = {10, 101} and B = {001, 100, 1001}.
(a) Is AB an unambiguous expression? Find the generating series for AB

with respect to length.
(b) Is B A an unambiguous expression? Find the generating series for B A
with respect to length.
2. Let A = {00, 101, 11} and B = {00, 001, 10, 110}. Prove that A ∗ is unambiguous,
and that the B ∗ is ambiguous. Find the generating series for A ∗ with respect
to length.
3. For each of the following sets, write an unambiguous expression generating

the set of strings.
(a) The {0, 1}-strings that have no substring of 0s with length 3, and no sub-
string of 1s of length 2.
(b) The {0, 1}-strings that have no block of 0s of size 3, and no block of 1s of
size 2.
(c) The set of {0, 1}-strings in which the substring 011 does not occur.
(d) The set of {0, 1}-strings in which the substring 0110 does not occur.
(e) The set of {0, 1}-strings where the 0-blocks have even lengths and the
1-blocks have odd lengths.
(f) The set of all {0, 1}-strings where each odd-length block of 0s is followed
by a non-empty even block of 1s and each even-length block of 0s is
followed by an odd-length block of 1s.
4. Let S = {0}∗ ({1}{11}∗ {00}{00}∗ ∪ {11}{11}∗ {0}{00}∗ )∗ .
(a) Describe in words the strings that belong to S.

(b) Find the generating series for S.
(c) Let b n equal the number of strings in S with length n. Establish a recur-
rence relation with sufficient initial conditions to uniquely determine
the values b n .
5. (a) Show that the generating series by length for binary strings in which
every block of 0’s has length at least 2 and every block of 1’s has length
at least 3, is
(1 − x + x 3 )(1 − x + x 2 )
.
1 − 2x + x 2 − x 5
(b) Obtain a recurrence relation for the coefficients in this generating se-
ries.
6. (a) Let a n be the number of {0,1}-strings of length n such that each even
block of 0’s is followed by exactly one 1 and each odd block of 0’s is
followed by exactly two 1’s. Show that
1+x
a n = [x n ] .
1 − x 2 − 2x 3
(b) Prove that a n = a n−2 + 2a n−3 , n ≥ 3.

(c) Find the initial values a 0 , a 1 , a 2 .
7. (a) Let b n be the number of {0,1}-strings of length n in which any block of

1’s which immediately follows a block of 0’s must have length at least
as great as the length of the block of 0’s. Find n≥0 b n x n .
P
(b) Prove that b n = b n−1 + 2b n−2 − b n−3 , n ≥ 3.

(c) Find the initial values b 0 , b 1 , b 2 .
8. Let a n be the number of {0, 1}-strings of length n in which the substring

“0110” does not occur.
(a) Prove that

1 + x3
a n = [x n ] , n ≥ 0.
1 − 2x + x 3 − x 4
(b) Deduce from (a) a linear recurrence equation for the sequence {a n }n≥0 ,
and give enough initial conditions so that {a n }n≥0 is uniquely specified.
9. Find the generating series, with respect to length, for {0,1}-strings in which
no block has length greater than m.
10. Find the generating series, with respect to length, for {0,1}-strings in which
no block of 0’s has length greater than m, and no block of 1’s has length
greater than k.
11. Show that the generating series by length for the number of binary strings in
which the substring 01110 does not occur is
1 + x4
.
1 − 2x + x 4 − x 5
12. (a) Let s n be the number of {0, 1}-strings of length n in which every block of
0’s is followed by a block of 1’s of the same parity in length (i.e. a block
with an even number of 0’s is followed by a block with an even number
of 1’s, etc.). Determine the generating series Φ(x) = n≥0 s n x n .
P
n
(b) Show that if n ≥ 2 then s n = 2 · 3b 2 c−1 .
13. Let A i be the set of {0,1}-strings with no occurrences of substring “00” whose
first symbol is an i , for i = 0, 1.
(a) Show, without determining A 0 or A 1 , that
˙
A 0 = {0}∪{0}A 1
˙
A 1 = {1}∪{1}(A ˙ 0 ).
1 ∪A
(b) Determine Φ A 0 (x) and Φ A 1 (x) using part (a).

(c) Deduce from (b) that the number of {0,1}-strings with no occurrences
of substring “00” is
1+x
[x n ] .
1 − x − x2
14. Let b n be the number of {0, 1}-strings of length n in which every block of 0’s is
followed by a block of 1’s of the same length. (For example, 111010001110011
and 100000111110011 are strings of length 15 of this kind.)
(a) Prove that

1+x
bn x n =
X
.
n≥0 1 − 2x 2
(b) Evaluate b n , n ≥ 0.
2.9 Bivariate Generating Series

In the above problems the weight function is the length of a string, or number
of symbols in a string. Thus the appropriate generating series is obtained by
writing down (generating) all strings in the required set and marking the occur-
rence of each symbol (0 or 1) by a single power of x. There is no reason why we
can’t record the number of 0’s and 1’s separately, say marking each 0 by x, and
each 1 by y. Using the decomposition S = {0, 1}∗ for the set of all binary strings,
the number of strings with m occurrences of 1 and n occurrences of 0 is:
1
[x m y n ] .
1 − (x + y)
What we have here is a “bivariate” weight function and a bivariate generating

series, that is, a series of two variables. We can solve other problems by marking
various particular substrings, not just a 0 or 1, and our generating series can
have any number of variables. Fortunately the sum and product lemmas extend
to this more general setting.
Problem 2.9.1. Let B (x, u) denote the multivariate generating series where [x n u k ]B (x, u)
counts the number of binary strings with length n and k occurrences of the sub-
string 001. Show that
1
B (x, u) = .
1 − 2x − (u − 1)x 3
Solution: Let S be the set of all {0, 1}-strings. We use the decomposition
S = {1, 01, 001, . . .}∗ {0}∗ .

2.9. BIVARIATE GENERATING SERIES 55
This gives
1
B (x, u) = (1 + x + x 2 + . . .)
1 − (x + x 2 + ux 3 + ux 4 + · · · )
1 1
= 3 3
x−x
1−x 1−
1−x
−u x 1−x
1
= ,
1 − 2x − (u − 1)x 3
as required.
Problem 2.9.2. Let A(x, y, u) denote the multivariate generating series where
[x m y n u k ]A(x, y, u) counts the number of binary strings with m ones, n zeros,
and k blocks. Show that
(1 + (u − 1)x)(1 + (u − 1)y)
A(x, y, u) = .
1 − x − y − (u 2 − 1)x y
Solution: Let S be the set of all {0, 1}-strings. We use the decomposition
S = {ε, 1, 11, . . .}({0, 00, . . .}{1, 11, . . .})∗ {ε, 0, 00, . . .}.
This gives
(1 + ux + ux 2 + . . .)(1 + u y + u y 2 + . . .)
A(x, y, u) =
1 − (u y + u y 2 + · · · )(ux + ux 2 + · · · )
¡ ux
¢³ uy
´
1 + 1−x 1 + 1−y
= ³ ´
u y ¡ ux ¢
1 − 1−y 1−x
((1 − x) + ux)((1 − y) + u y)
=
(1 − x)(1 − y) − u 2 x y
(1 + (u − 1)x)(1 + (u − 1)y)
= ,
1 − x − y − (u 2 − 1)x y
as required.
By making appropriate variable substituitions we can extract other interest-
ing generating series. For example, A(x, y, 1) gives the generating series where
x marks occurrences of 1 and y marks occurrences of 0, and A(x, x, u) gives the
generating series where x the length of a string and u marks each block. By
Problem 2.9.2,
1
A(x, y, 1) =
1 − (x + y)
and
(1 + (u − 1)x)2
A(x, x, u) =
1 − 2x − (u 2 − 1)x 2
(1 + (u − 1)x)2
=
(1 − (u − 1)x)(1 + (u − 1)x)
1 + (u − 1)x
= .
1 − (u − 1)x
Bivariate generating series are useful for finding averages, using the follow-
ing result, where “Property P " would be specified in any particular example. We
will use F u (x, u) as an abbreviation for
∂
F (x, u).
∂u
Theorem 2.9.3. Let R be a set of {0, 1}-strings, and c n,k be the number of strings
in R of length n with k occurrences of property P , n, k ≥ 0. Let µn be the aver-
age number of occurrences of property P among the {0, 1}-strings in R of length
n, and let
c n,k x n u k .
X X
F (x, u) =
n≥0 k≥0
Then
[x n ]F u (x, 1)
µn = , n ≥ 0.
[x n ]F (x, 1)
Proof: The total number of strings in R of length n is k≥0 c n,k . The total num-
P
P
ber of occurrences of property P is strings of length n is k≥0 kc n,k . Therefore,
N
µn = ,
D
P P
where D = k≥0 c n,k and N = k≥0 kc n,k . Now
c n,k x n ,
X X
F (x, 1) =
n≥0 k≥0
so D = [x n ]F (x, 1). Also

∂
c n,k x n ku k−1 ,
X X
F (x, u) =
∂u n≥0 k≥0
so
kc n,k x n .
X X
F u (x, 1) =
n≥0 k≥0
³ ´
∂
Since N = [x n ] ∂u
F (x, u) |u=1 , the result follows.
Problem 2.9.4. (a) Let c n,k be the number of {0, 1}-strings of length n with k
blocks. Prove that
1 − x + ux
c n,k = [x n u k ] , n, k ≥ 0.
1 − x − ux
(b) Let µn be the average number of blocks among {0, 1}-strings of length n.
Determine µn , n ≥ 0.
Solution: We use the block decomposition
{0, 1}∗ = {0}∗ ({1}{1}∗ {0}{0}∗ )∗ {1}∗
and mark each block by a u. Then
c n,k x n u k
X X
F (x, u) =
n≥0 k≥0
is the generating series for {0, 1}∗ with x marking length and u marking blocks,
so
³ ux ´ ³ ux ux ´−1 ³ ux ´
F (x, u) = 1 + 1− 1+
1−x 1−x 1−x 1−x
¶−1
ux ux
µ
³ ´2 ³ ´ 2
= 1+ 1−
1−x 1−x
³ ux ´2 ³³ ux ´ ³ ux ´´−1
= 1+ 1− 1+
1−x 1−x 1−x
ux
1 + 1−x
= ux
1 − 1−x
1 − x + ux
= ,
1 − x − ux
and so (a) follows.

From Theorem 2.9.3, we have R = {0, 1}∗ , and property P is the occurrence
of a block, so
[x n ]F u (x, 1)
µn = n , n ≥ 0.
[x ]F (x, 1)
Now in this case
1−x +x 1
= (2x)i ,
X
F (x, 1) = =
1 − x − x 1 − 2x i ≥0
so [x n ]F (x, 1) = 2n , n ≥ 0. We also have
∂ 1 − x + ux
µ ¶
F u (x, u) =
∂u 1 − x − ux
x(1 − x − ux) − (−x)(1 − x + ux)
= ,
(1 − x − ux)2
so
x(1 − 2x) + x
F u (x, 1) =
(1 − 2x)2
x x
= +
1 − 2x (1 − 2x)2
Ã !
i
X j +1
(2x) j
X
= x (2x) + x
i ≥0 j ≥0 j
X i i +1 X
= 2x + ( j + 1)2 j x j +1 .
i ≥0 j ≥0
Thus if n = 0 then [x n ]F u (x, 1) = 0 and if n ≥ 1, then
[x n ]F u (x, 1) = 2n−1 + n2n−1 = (n + 1)2n−1
and so (
0, n = 0;
µn = n+1
2 , n ≥ 1.
This settles (b).
Problem 2.9.5. (a) Let a n,k be the number of {0, 1}-strings of length n with k
occurrences of 0011 as a substring. Prove that
a n,k = [x n u k ]{1 − 2x − (u − 1)x 4 }−1 .

(b) Let µn be the average number of occurrences of 0011 as a substring among

{0, 1}-strings of length n. Determine µn , n ≥ 0.
Solution: Let
a n,k x n u k ,
X X
F (x, u) =
n≥0 k≥0
∗
the generating series for {0, 1} with x marking length and u marking occur-
rences of 0011 as a substring. We use the block decomposition
{0, 1}∗ = {1}∗ ({0}{0}∗ {1}{1}∗ )∗ {0}∗ ,
and note that 0011 occurs as a substring whenever a block of 2 or more 0’s is
immediately followed by a block of 2 or more 1’s. Thus
¶¾−1
x x x2 x2 x2 x2
½ µ
1 1
A(x, u) = 1− − +u
1−x 1−x 1−x 1−x 1−x 1−x 1−x 1−x
2 2 4 4 −1
= {(1 − x) − (x − x + ux )} ,
= {1 − 2x − (u − 1)x 4 }−1 .
This proves (a)

From Theorem 2.9.3 we have
[x n ]F u (x, 1)
µn = .
[x n ]F (x, 1)
In this case we have
1
= (2x)i ,
X
F (x, 1) =
1 − 2x i ≥0
so
[x n ]F (x, 1) = 2n , n ≥ 0.
We also have
−(−x 4 )
F u (x, u) = ,
(1 − 2x − (u − 1)x 4 )2
so
x4
F u (x, 1) =
(1 − 2x)2
= x 4 ( j + 1)(2x) j
X
j ≥0
( j + 1)2 j x j +4 .
X
=
j ≥0
Thus (
0, n = 0, 1, 2, 3;
[x n ]F u (x, 1) = n−4
(n − 3)2 , n ≥ 4.
so (
0, n = 0, 1, 2, 3,
µn = n−3
16 , n ≥ 4.
Note that, in general,

F u (x, 1)
µ ¶
n
µn 6= [x ] ,
F (x, 1)
so be careful to avoid this mistake!
Problem Set 2.9

1. Let S = N≥0 3
. Consider the two weight functions ω1 (t 1 , t 2 , t 3 ) = t 1 + t 2 + t 3
and ω2 (t 1 , t 2 , t 3 ) = number of odd numbers among t 1 , t 2 , t 3 . For example,
ω1 (3, 5, 6) = 14 and ω2 (3, 5, 6)=2 (since 3 and 5 are odd).
Find the generating series for S with respect to the weight functions ω1 and
ω2 where we use the variable x to record the ω1 weight and y to record the
ω2 weight.
2. (a) Let a n,k be the number of binary strings of length n containing exactly
k singleton 1’s, i.e. blocks consisting of a single 1. Show that
(1 − x)−1 + x(t − 1)
a n,k x n t k =
X
.
n≥0 1 − x − x 2 (1 − x)−1 − x 2 (t − 1)
k≥0
(b) Hence find the average number of singleton 1’s in strings of length n.
3. (a) Show that the number of {0,1}-strings of length n with m occurrences

of 01 (for example 01011 has two occurrences of 01) is
Ã !
n +1
n − 2m
(b) Show that the average number of occurrences of 01 is (n − 1)/4, n ≥ 1.

4. Find the average number of blocks of 1’s in a {0,1}-string of length n.
5. Find the generating series, by numbers of 0’s and 1’s, for {0,1}-strings which
satisfy both of the following conditions:
(a) all blocks have odd length, and

(b) each block of 0’s has a different length from the block of 1’s which fol-
lows it.
6. Find the generating series, with respect to number of 0’s and number of 1’s,
for {0,1}-strings in which each block of 0’s immediately following a block of
1’s of length k has length at least k and at most 2k − 1 for all k ≥ 1.
7. (a) Find the generating series for {0,1}-strings with respect to length and
with respect to numbers of occurrences of “011” as a substring.
(b) Find the average number of occurrences of “011” as a substring in {0,1}-
strings of length n.
(c) A gambler offers you the following proposition: First, you give him $50.
Then, you flip a fair coin 100 times. Each time you get a tail immedi-
ately followed by two heads, he gives you back $4. On average, how
much money do you win or lose in this game?
8. Let S be the set of all binary strings that do not contain the substring 000.
(a) Determine a recursive decomposition of the set S.

(b) Use the recursive decomposition to determine a generating series for
S.
9. (a) Prove that the generating series for the number c(n, k) of {0,1}-strings
of length n with k occurrences of a block of 3 0’s immediately followed
by a block of two 1’s is
1
c(n, k)x n u k =
X
.
k,n≥0 1 − 2x + (1 − u)x 5 (1 − x)2
(b) Prove that the average number of occurrences of a block of three 0’s
immediately followed by a block of 2 1’s in {0,1}-strings of length n is
1
128 (n − 2) for n > 5.
10. Let α be a {0,1}-string; α is said to be simple if there are no {0,1}-strings u, v, w

such that uw = α = w v, where w is non-empty, and u, v are not both empty.
(a) Show that 10100 is simple, but that 10101 is not.

(b) Let S α be the set of {0,1}-strings not containing α as a substring. If α is
simple, prove that
{0, 1}∗ = S α {αS α }∗ .
(c) If α is simple, of length p, find the number of strings in S α of length n.
11. The set of all {0,1,2}-strings can be decomposed as
{0, 1, 2}∗ = {0}∗ (({1, 2}∗ − ε){0}{0}∗ )∗ {1, 2}∗
(a) Using a similar decomposition, find the generating series by length for
{0,1,2}-strings with no “22” substring. (You may assume that the gener-
ating series by length for {0, 1}∗ is (1 − 2x)−1 .)
(b) Let a n be the number of strings from (a) with length n. Find a recur-
rence relation for a n .
(c) Find a formula for a n by the method of partial fractions.
12. (a) A Dyck word in {0, 1}∗ is a string ε or σ1 · · · σ2n such that (i) σ1 = 0; (ii)
σ1 + · · · + σi ≤ i /2 for i = 1, . . . , 2n − 1; (iii) σ1 + · · · + σ2n = n. Let D be the
set of all Dyck words. Prove that
D − {ε} = 0D1D
where ε is the empty word.

(b) Find the number of Dyck words of length 2n.
Chapter 3
Recurrences, Binary Trees and

Sorting
3.1 Coefficients of Rational Functions

Suppose that we have computed that the generating series for the sequence
{a n }n≥0 is
1
Φ A (x) = an x n =
X
.
n≥0 (1 − 2x)(1 − 3x)
We show one way we can use this to produce an explicit formula for the terms
of the sequence. The first step is to notice that
1 −2 3
= + .
(1 − 2x)(1 − 3x) 1 − 2x 1 − 3x
(You would not be expected to guess this, but certainly you can easily check
that it is correct.) It follows that
−2 3
a n = [x n ]Φ A (x) = [x n ] + [x n ]
1 − 2x 1 − 3x
1 1
= (−2)[x n ] + 3[x n ]
1 − 2x 1 − 3x
= −2n+1 + 3n+1 .
We will show how we can compute coefficients whenever the generating

series is a rational function whose denominator has constant term 1 and factors
63
64 CHAPTER 3. RECURRENCES, BINARY TREES AND SORTING
into linear terms. Suppose that we have

f (x)
Φ A (x) = .
g (x)
This Φ A (x) is a rational function. If this rational function is not proper, that is,
if deg( f ) ≥ deg(x) then
f (x) = q(x)g (x) + r (x)
where deg(r ) < deg(g ) and
r (x)
Φ A (x) = q(x) + .
g (x)
This means that we can compute the coefficients of rational function provided
we can do it for proper rational functions. We can do this in two steps.
Lemma 3.1.1. If f (x) is a polynomial of degree less than r , then there is a poly-
nomial P (x) with degree less than r such that
f (x)
[x n ] = P (n)θ n .
(1 − θx)r
Proof: This is an easy exercise using the binomial theorem.
Note that if r = 1, then the polynomial P (x) will be a constant polynomial.
Lemma 3.1.2. Suppose f and g are polynomials with deg( f ) < deg(g ). If g (x) =
g 1 (x)g 2 (x) where g 1 and g 2 are coprime, there are polynomials f 1 and f 2 such
that deg( f i ) < deg(g i ) for i = 1, 2 and
f (x) f 1 (x) f 2 (x)
= + .
g (x) g 1 (x) g 2 (x)
Proof: Recall that, since g 1 and g 2 are coprime, we can use the extended Eu-
clidean algorithm to find polynomials a 1 and a 2 such that deg(a 1 ) < deg(g 2 ),
deg(a 2 ) < deg(g 1 ) and
a 1 (x)g 1 (x) + a 2 (x)g 2 (x) = 1.
Suppose now that deg( f ) < deg(g ), let the polynomial b 1 (x) be the remainder
when we divide f (x)a 1 (x) by g 2 (x) and let the polynomial b 2 (x) be the remain-
der when we divide f (x)a 2 (x) by g 1 (x). Then modulo g (x) = g 1 (x)g 2 (x) we have
b 1 (x)g 1 (x) + b 2 (x)g 2 (x) ≡ f (x);
since both sides have degree less than the degree of g , they are equal.
3.2. SOLUTIONS TO RECURRENCE EQUATIONS 65
The previous two lemmas yield the following theorem. Note that the num-
bers θi are the reciprocals of the zeros of g (x) and the integer m i is the multi-
plicity of θi−1 as a zero of g (x).
Theorem 3.1.3. Suppose f and g are polynomials such that deg( f ) < deg(g ). If
for i = 1, . . . , k there are complex numbers θi and positive integers m i such that
g (x) = (1 − θi x)mi
Y
i
then there are polynomials P i such that deg(P i ) < m i and
f (x) X k
[x n ] = P i (n)θin .
g (x) i =1
3.2 Solutions to Recurrence Equations

In earlier sections we have met linear recurrences for sequences, when the gen-
erating series for the sequence is a rational function.
Theorem 3.2.1. Let C (x) = n≥0 c n x n where the coefficients c n satisfy the recur-
P
rence
c n + q 1 c n−1 + · · · + q k c n−k = 0, (n ≥ k). (3.2.1)
If
g (x) := 1 + q 1 x + · · · + q k x k ,
there is a polynomial f (x) with degree less than k such that
f (x)
C (x) = .
g (x)
Proof: From the rule for the coefficients of a product of power series, if n ≥ k
then
c n + q 1 c n−1 + · · · + q k c n−k = [x n ](1 + q 1 x + · · · + q k x k )C (x).
So [x n ]g (x)C (x) = 0 if n ≥ k and therefore g (x)C (x) = f (x) for some polynomial
f (x) of degree less than k.
The recurrence in Equation 3.2.1 is called a homogeneous equation be-
cause the right-hand side of the equation is zero. (The term ‘homogeneous’
was used in the same way in your linear algebra courses.) The polynomial
h(x) = x k g (x −1 ) in this theorem is called the characteristic polynomial of the
recurrence. Using Theorem 3.1.3 we have at once:
Theorem 3.2.2. Suppose (c n )n≥0 satisfies the recurrence equation (3.2.1). If the
characteristic polynomial of this recurrence has root βi with multiplicity m i ,
for i = 1, . . . , j , then the general solution to (3.2.1) is
c n = P 1 (n)βn1 + · · · + P j (n)βnj (3.2.2)
where P i (n) is a polynomial in n with degree less than m i , and these polynomi-
als are determined by the c 0 , . . . , c k−1 .
Recall that to find roots of a polynomial, we use the Factor Theorem:: if F (x)
is a polynomial and F (a) = 0 then x − a is a factor of F (x).
Problem 3.2.3. Find c n explicitly, where
c n − 4c n−1 + 5c n−2 − 2c n−3 = 0, (n ≥ 3)
with initial conditions c 0 = 1, c 1 = 1, c 2 = 2.
Solution: The characteristic polynomial is
x 3 − 4x 2 + 5x − 2 = (x − 1)2 (x − 2)
has root β1 = 1 with multiplicity m 1 = 2, and root β2 = 2 with multiplicity m 2 = 1.

Thus
c n = (A + B n) · (1)n +C · 2n , (n ≥ 0)
where A, B and C are constants. From the initial conditions, we have
1 = A + C
1 = A + B + 2C
2 = A + 2B + 4C
and solving this system gives (check this)
A = 0, B = −1, C = 1.
Thus, the solution to the recurrence is
c n = 2n − n, (n ≥ 0).
Problem 3.2.4. Find c n explicitly, where
c n + 4c n−1 − 3c n−2 − 18c n−3 = 0, (n ≥ 3)
with initial conditions c 0 = 0, c 1 = 2, c 2 = 13.
Solution: The characteristic polynomial
x 3 + 4x 2 − 3x − 18 = (x − 2)(x + 3)2
has root β1 = 2 with multiplicity m 1 = 1, and root β2 = −3 with multiplicity

m 2 = 2. Thus
c n = A · 2n + (B +C n) · (−3)n , (n ≥ 0),
where A, B,C are constants. From the initial conditions, we have
0 = A + B
2 = 2A − 3B − 3C
13 = 4A + 9B + 18C
and solving this system gives (check this)
A = 1, B = −1,C = 1.
Thus the solution to the recurrence is
c n = 2n + (n − 1)(−3)n , (n ≥ 0).
Problem 3.2.5. Let a n = 3n − 2n , b n = 3n + 2n , n ≥ 0. Find linear homogeneous

recurrence equations with initial conditions for
(a) c n = a n2 .
(b) d n = a n2 + b n2 .
(c) e n = a n b n .
(d) f n = b n + 5.
Solution: (a) We have c n = 9n − 2 · 6n + 4n , so c n satisfies the linear recurrence

equation with characteristic polynomial
(x − 9)(x − 6)(x − 4) = x 3 − 19x 2 + 114x − 216,
and thus a linear recurrence equation for c n is
c n − 19c n−1 + 114c n−2 − 216c n−3 = 0, n ≥ 3.
Initial conditions are obtained from the formula for c n above, which gives c 0 =
0, c 1 = 1, c 2 = 25.
(b) We have d n = 2·9n +2·4n , so d n satisfies the linear recurrence equation with
characteristic polynomial
(x − 9)(x − 4) = x 2 − 13x + 36,
and thus a linear recurrence equation for d n is
d n − 13d n−1 + 36d n−2 = 0, n ≥ 2.
Initial conditions are obtained from the formula for d n above, which gives d 0 =
4, d 1 = 26.
(c) We have e n = 9n −4n , so e n satisfies the linear recurrence equation with char-
acteristic polynomial
(x − 9)(x − 4) = x 2 − 13x + 36,
and thus a linear recurrence equation for e n is
e n − 13e n−1 + 36e n−2 = 0, n ≥ 2.
Initial conditions are obtained from the formula for e n above, which gives e 0 =
0, e 1 = 5.
(d) We have f n = 3n +2n +5 = 3n +2n +5·1n , so f n satisfies the linear recurrence
equation with characteristic polynomial
(x − 3)(x − 2)(x − 1) = x 3 − 6x 2 + 11x − 6,
and thus a linear recurrence equation for f n is
f n − 6 f n−1 + 11 f n−2 − 6 f n−3 = 0, n ≥ 3.
Initial conditions are obtained from the formula for f n above, which gives f 0 = 7,
f 1 = 10, f 2 = 18.
Problem Set 3.2

1. Consider the recurrence equation c n = 5c n−1 − 3c n−2 − 9c n−3 = 0, for n ≥ 3
with c 0 = 1, c 1 = 1, c 2 = 29.
(a) Determine c n explicitly.

(b) What is the asymptotic form for c n ?
2. Find, explicitly, b n where b n − 5b n−1 + 8b n−2 − 4b n−3 = 0 for n ≥ 3 with initial

conditions b 0 = 2, b 1 = 2, b 2 = 0.
3. Solve the recurrence relation c n+4 = 5c n+3 −9c n+2 +7c n+1 −2c n = 0 with initial
values c 0 = 1, c 1 = 0, c 2 = 1, c 3 = 0.
4. Consider the recurrence c n = 5c n−1 −7c n−2 +3c n−3 with the initial conditions
c 0 = 3, c 1 = 8, c 2 = 21.
(a) Give a formula for c n .

(b) Determine the asymptotic behaviour of c n .
5. Consider the recurrence equation c n = 4c n−1 − c n−2 − 6c n−3 , for n ≥ 3, with

c 0 = 5, c 1 = 0,c 2 = 18.
(a) Determine c n explicitly.

(b) What is the asymptotic form for c n ?
6. Consider the recurrence c n+5 = 5c n+4 − 10c n+3 + 10c n+2 − 5c n+1 + c n , n ≥ 0,
with c 0 = 1, c 1 = 5, c 2 = 15, c 3 = 35, c 4 = 70. Determine c n explicitly, n ≥ 0.
7. A sequence {c n } of integers is defined by the recurrence c n = c n−1 +2c n−2 , for

n ≥ 2, together with the initial conditions c 0 = c 1 = 1.
1
(a) Show that c n = [x n ] 1−x−2x 2.
(b) Deduce that

µ ¶
2n 1 1 1
c 2n = [x ] +
2 1 − x − 2x 2 1 + x − 2x 2
and that
1 − 2x
c 2n x n =
X
n≥0 1 − 5x + 4x 2
and deduce, or prove otherwise, that c 2n = 5c 2n−2 − 4c 2n−4 for n ≥ 2.
(c) Find a similar equation connecting c 2n , c 2n−4 and c 2n−8 for n ≥ 2.

(d) Obtain an explicit expression for c n from (a).
3.3 Nonhomogeneous Recurrence Equations

We now wish to determine the sequence b 0 , b 1 , . . . satisfying the recurrence
b n + q 1 b n−1 + . . . + q k b n−k = f (n), (n ≥ k), (3.3.1)
where q 1 , . . . , q k are constants, and f (n) is a specified function of n. To uniquely

determine the sequence b 0 , b 1 , . . ., we would specify the values of b 0 , . . . , b k−1 as
initial conditions.
Theorem 3.3.1. Suppose that a 0 , a 1 , . . . is a solution to (3.3.1) ( any solution;

without checking the initial conditions). Then the general solution to (3.3.1)
is given by
bn = an + cn , n ≥ 0,
where c n is given by Theorem 3.2.2, and the k constants b 11 , . . . , b j m j in c n can

be chosen to fit the initial conditions for b n .
Proof: We are given that
a n + q 1 a n−1 + · · · + q k a n−k = f (n), n ≥ k,
and subtracting this equation from (3.3.1) above, we obtain
(b n − a n ) + q 1 (b n−1 − a n−1 ) + · · · + q k (b n−k − a n−k ) = 0, n ≥ k.
But this means that b n − a n is a solution to the homogeneous recurrence

(3.2.1), and the result follows from Theorem 3.2.2.
Problem 3.3.2. Solve
b n − 4 b n−1 + 5 b n−2 − 2 b n−3 = 24 (−1)n , n ≥ 3, (3.3.2)
with initial conditions b 0 = −1, b 1 = −3, b 2 = 2.

3.3. NONHOMOGENEOUS RECURRENCE EQUATIONS 71
Solution: If we suppose that some solution to (3.3.2) has the form b n = α(−1)n ,
for some choice of constant α, then substituting b n = α(−1)n on the left side of
(3.3.2), we obtain
α(−1)n −4α(−1)n−1 +5α(−1)n−2 −2α(−1)n−3 = (1+4+5+2)α(−1)n = 12α(−1)n ,
and from the RHS of (3.3.2), we deduce that 12 α = 24, so α = 2, and thus b n =
2(−1)n is a solution to (3.3.2). Now, from the solution to Problem (3.2.4) and the
result above, we find that the general solution to (3.3.2)is
b n = 2 (−1)n + A + B n +C · 2n , n ≥ 0,
where A, B,C are constants, whose values can be determined from the initial
conditions:
−1 = 2 + A + 0 · B +C
−3 = −2 + A + B + 2C
2 = 2 + A + 2B + 4C .
Solving this system gives (check this) A = −2, B = 3, C = −1, so the solution to
the recurrence is
b n = 2 (−1)n − 2 + 3n − 2n , (n ≥ 0).
It is difficult to specify exactly how one might obtain some solution a n to

a nonhomogeneous recurrence. However, in general, it is useful to consider
a family of solutions with unknown constants that could be solved for, like the
constant α in Problem 3.3.2. In the following example, the right side is a specific
polynomial in n of degree 1, and the family we consider is a general polynomial
in n of degree 1. When the right side is a polynomial in n, you should always
begin by trying an arbitrary polynomial, starting with one of the same degree,
and if that fails, again with a polynomial of higher degree.
Problem 3.3.3. Solve
b n + 4 b n−1 − 3 b n−2 − 18 b n−3 = 2n + 1, n ≥ 3, (3.3.3)
with initial conditions b 0 = 0, b 1 = 27/8, b 2 = −67/4.

Solution: If we suppose that some solution to (3.3.3) has the form b n = αn + γ,

for some choice of constants α, γ, then substituting b n = αn + γ, on the LHS
of (3.3.3), we obtain
(αn + γ) + 4(α(n − 1) + γ) − 3((α(n − 2) + γ) − 18((α(n − 3) + γ)

= −16αn + 56α − 16γ,
and from the RHS of (3.3.3), we deduce that −16α = 2 and 56α − 16γ = 1, so
α = −1/8, γ = −1/2, and thus b n = −(n + 4)/8 is a solution to (3.3.3). Now, from
the solution to Problem 3.2.4 and the result above, we have the general solution
to (3.3.3) as
n +4
bn = − + A · 2n + (B +C n)(−3)n , n ≥ 0,
8
where A, B,C are constants, whose values can be determined from the initial
conditions:
1
0 = − + A +B
2
27 5
= − + 2A − 3B − 3C
8 8
67 3
− = − + 4A + 9B + 18C ,
4 4
and solving this system gives (check this) A = 1/2, B = 0, C = −1, so the solu-
tion to the recurrence is
n +4
bn = − + 2n−1 − n(−3)n , (n ≥ 0).
8
3.4 Aysmptotics
We say that c n is asymptotic to g (n) as n → ∞, written c n ∼ g (n), if
cn
lim = 1.
n→∞ g (n)
3.5. OTHER RECURRENCE RELATIONS 73
For example, suppose that the roots β1 , . . . , β j in (3.2.2) are all real, and that
β1 has the largest absolute value of β1 , . . . , β j . Then
c n = A 1 βn1 + A 2 βn2 + · · · + A j βnj

A 2 β2 n A j βj n
½ µ ¶ µ ¶ ¾
n
= A 1 β1 1 + +···+
A 1 β1 A 1 β1
where A 1 , A 2 , . . . A j are polynomials in n. Thus
cn A 2 β2 n A j βj n
µ ¶ µ ¶
= 1+ +···+
A 1 βn1 A 1 β1 A 1 β1
and
cn
lim = 1+0+···+0
n→∞ A 1 βn
1
β βj
since β21 , . . . , β1 all have absolute value less than 1. Thus we have c n ∼ A 1 βn1 . For
a more compact result, here we have
A 1 = b 11 + b 12 n + · · · + b 1m1 n m1 −1 .
So, if b 1m1 6= 0,
cn
lim = 1,
n→∞ b 1m n m 1 −1 βn
1 1
and we also write
c n ∼ b 1m1 n m1 −1 βn1 .
This answers the question, “Roughly how fast does c n grow as n increases?”
Example 3.4.1. In Problem 3.2.3, with c n = 2n − n, we have c n ∼ 2n . In Problem
3.2.4, with c n = 2n + (n − 1)(−3)n , we have c n ∼ (n − 1)(−3)n and c n ∼ n(−3)n .
3.5 Other Recurrence Relations

We have seen that if a generating function C (x) = n≥0 c n x n is a rational func-
P
tion, i.e. the ratio of two polynomials then we can easily deduce a linear ho-
mogeneous recurrence equation of fixed order, with constant coefficients, by
simply multiplying on both sides of
P (x)
C (x) =
Q(x)
by Q(x). Also we can obtain an exact form for c n for such generating functions
by considering the partial fraction expansion or by repeated use of the binomial
theorem. Another method of deducing a recurrence equation is by looking at
the derivatives of a generating function.
By way of example, consider F (x) = m≥0 f m x m = (1 − x)−k . Suppose we
P
have forgotten the binomial theorem and wish to find a recurrence relation for
f n . Now
F 0 (x) = m f m x m−1 = k(1 − x)−k−1
X
n≥0
so F (x) = k(1 − x) F (x), or (1 − x)F 0 (x) = kF (x).
0 −1
Thus (1 − x) m≥0 m f m x m−1 = k m≥0 f m x m and comparing coefficients of

P P
x n−1 gives, when n ≥ 1

n f n − (n − 1) f n−1 = k f n−1 ,
n f n − (n − 1 + k) f n−1 = 0.
The initial condition is f 0 = 1, since f 0 = F (0).
Conversely, there are situations where instead of being given a generating
function and asked to find a recurrence equation for the coefficients, we start
with a recurrence equation and are asked to find an explicit form, or generating
function, for the coefficients. Such a situation arises naturally in the analysis
of algorithms, which are often specified recursively, and thus naturally lead to
recurrence equations. Examples are given in Sections 3.8 and 3.9.
Problem Set 3.5

1. (a) For the non-homogeneous recurrence relation
c n − 4c n−1 + 4c n−2 = 3
c 0 = 4, c 1 = 11,
P n
determine n≥0 c n x .
(b) Derive a homogeneous recurrence relation for c n . Solve to obtain a
formula for c n .
(c) What is the asymptotic behaviour of c n ?
2. Solve the nonhomogeneous linear recurrence relation

b n − b n−1 − 16b n−2 − 20b n−3 = 36n + 15
having initial conditions b 0 = 0, b 1 = 1, b 2 = 12.
3.6. BINARY TREES 75
3. Solve the nonhomogeneous linear recurrence relation
b n − 3b n−1 + 4b n−3 = 3 · 2n
having initial conditions b 0 = 3, b 1 = 1, b 2 = 27.
3.6 Binary Trees

In this section we count binary trees. A binary tree is either
(a) the empty tree, or
(b) a tree with a fixed root vertex such that each vertex has a left branch and
a right branch (either of which may be empty).
Below, for example, we list all binary trees with three vertices.
Figure 3.1: Binary trees with three vertices
We let ² denote the empty tree and we let • denote the tree with one vertex.
Let T be the set of all binary trees. For a tree T ∈ T , we let n(T ) be the number
of vertices in T . We let T (x) denote the generating function for T with respect
to the weight function n. Thus [x n ]T (x) is the number of binary trees with n
vertices. The following problem detemines T (x) implicitly.
Problem 3.6.1. Show that
(2xT (x) − 1)2 = 1 − 4x.
Solution: For any binary tree T 6= ², we can decompose T into (•, T1 , T2 ), where
T1 and T2 are, respectively, the left and right branches of the root vertex of T .
This decomposition provides a bijection from T − {²} to {•} × T × T . So, with a

small abuse of notation, we can write
T = {²} ∪ {•} × T × T .
Moreover, since n(T ) = 1 + n(T1 ) + n(T2 ), we can apply the Product Lemma to
obtain:
T (x) = 1 + xT (x)2 .
Now
0 = 4x xT (x)2 − T (x) + 1
¡ ¢
= 4x 2 T (x)2 − 4xT (x) + 4x

= (2xT (x) − 1)2 + 4x − 1.
Therefore
(2xT (x) − 1)2 = 1 − 4x,
as required.
Up to this point in the course, our generating series have been determined
by linear equations, now we have a quadratic equation to solve. To determine
T (x) explicitly, we need to use a result, Lemma 3.7.2, from the next section giv-
1
ing the expansion of (1 − 4x) 2 .
Theorem 3.6.2. The number of binary trees with n ≥ 0 vertices is
Ã !
1 2n
.
n +1 n
Proof: By Problem 3.6.1 and Lemma 3.7.2,
1
2xT (x) − 1 = (1 − 4x) 2
Ã Ã Ã ! ! !
X 1 2n n+1
= ± 2 x −1 .
n≥0 n + 1 n
Considering the constant term reveals the sign, so

Ã !
X 1 2n n
T (x) = x .
n≥0 n + 1 n
Hence the number of binary trees with n ≥ 0 vertices is

Ã !
1 2n
,
n +1 n
as required.
3.7. THE BINOMIAL SERIES 77
1 2n
¡ ¢
The numbers n+1 n
are called the Catalan numbers; these arise in many
other counting problems.
Problem Set 3.6

1. Let t n denote the number of binary trees with n edges. Show that, for each
n ≥ 1, we have
t n = t 0 t n−1 + t 1 t n−2 + · · · + t n−1 t 0 .
2. Place 2n points equally spaced on a circle. A chord is a straight line joining

two of these points. Let d n be the number of ways we can place n chords
so that no two chords have a point in common (so they do not cross, and
they pair up the 2n points on the circle). Prove that d n+1 is the n-th Catalan
number.
3. Show that the number of ways we can fills the entries of a 2 × n matrix with
the integers 1, . . . , 2n such that entries increase as we move from right to left
along a row, and as we move down a column, is the n-th Catalan number.
4. If C n is the n-th Catalan number, prove that

Cn
lim → 0.
n→∞ 4n
5. For which values of n is C n odd?

p
6. Prove that 1 − 4x is not a rational function, i.e., there are no polynomials
p(x) and q(x) such that
p p(x)
1 − 4x = .
q(x)
Deduce that the Catalan numbers cannot satisfy a linear recurrence of finite
order.
3.7 The Binomial Series

By Theorems 1.5.1 and 1.8.5, we have expansions for (1 + x)n for any integer n.
1
What happens when n is not an integer? Indeed, what do we mean by (1 + x) 2 ?
If B (x) is a formal power series such that B (x)2 = 1 + x then we write B (x) = (1 +
1
x) 2 . This is much the same as our interpretation of an inverse; however, unlike
inverses, square–roots are not uniquely determined since (−B (x))2 = B (x)2 . We
3
can extend this interpretation to other rational powers. For example, (1 + x) 2 =
1 3
³ ´ 3 3
(1 + x) 2 , and (1 + x)− 2 is the inverse of (1 + x) 2 . Similarly, for any rational
number a, we can interpret (1 + x)a as a formal power series.
For any real number a and non–negative integer k we define
Ã !
a a(a − 1) · · · (a − k + 1)
= ;
k k!
this is called a binomial coefficient

¡a ¢ . It is convenient to extend this definition
to all real values of k by defining k = 0 when k 6∈ {0, 1, 2, . . .}.
Theorem 3.7.1 (The Binomial Theorem). For any rational number a,
Ã !
a
X a k
(1 + x) = x .
k≥0 k
We will not prove the Binomial Theorem; though, the reader should recog-
nize this, from Calculus, as the MacLaurin series expansion of (1+x)a . In Calcu-
lus, we place additional conditions on x since we view power series as functions
of x. Here, on the other hand, we do not evaluate formal power series; instead,
we are concerned with the sequence of coefficients. Therefore, we need not
address questions of convergence.
We show below that the Binomial Theorem is consistent with Theorems 1.5.1
¢ consider the case that a = n for some non–negative integer n.
and 1.8.5. ¡ First
Note that nk = 0 when n < k. Thus, by the Binomial Theorem,
Ã !
n n
n
xk ,
X
(1 + x) =
k=0 k
as expected.
Now consider the case that a = −n for some positive integer n. Thus,
Ã !
−n (−n)(−n − 1) · · · (−n − k + 1)
=
k k!
(n + k − 1)(n + k − 2) · · · n
= (−1)k
Ã ! k!
n +k −1
= (−1)k
k
Ã !
n + k − 1
= (−1)k .
n −1
Then, by the Binomial Theorem,
Ã !
X −n
(1 − x)−n = (−x)k
k≥0 k
Ã !
X n +k −1 k
= x ,
k≥0 n −1
as expected.
1
We can use the Binomial Theorem to compute (1 − 4x) 2 , which we needed
to enumerate binary trees; see Theorem 3.6.2.
Lemma 3.7.2.
Ã !
1 X 1 2n n+1
(1 − 4x) = 1 − 2
2 x .
n≥0 n + 1 n
Proof: Note that, for k ≥ 1,

1 −1 1
Ã !
1 − k + 1)
2 2 2 ···(2
=
k k!
k−1
(−1) (1 · 3 · · · (2k − 3))
=
2k k!
k−1
(−1) (1 · 2 · · · (2k − 3) · (2k − 2))
=
2k k! (2 · 4 · · · (2k − 2))
(−1)k−1 (2k − 2)!
=
2k k!2k−1 (1 · 2 · · · (k − 1))
(−1)k−1 (2k − 2)!
=
22k−1 k!(k − 1)!
Ã !
(−1)k−1 2k − 2
= .
22k−1 k k − 1
Hence, by the Binomial Theorem,

Ã !
1
1
xk
X
2
(1 + x) 2 =
k≥0 k
Ã ! Ã !
1 1
1
xk
X
2 2
(1 + x) 2 = +
0 k≥1 k
Ã !
X (−1)k−1 2k − 2 k
= 1+ 2k−1 k
x ,
k≥1 2 k −1
as required.
Problem Set 3.7

1. Calculate the binomial coefficients:
¡− 1 ¢ ¡1¢
(a) 73 (b) −2
¡¢ ¡ ¢
4
(c) 3
2 (d) 53
Answers:
5 22
(a) 35 (b) 5 (c) − 16 (d) 729
2. Find the following coefficients:
(a) [x 8 ](1 − x)−7 Answer: 3003

(b) [x 10 ]x 6 (1 − 2x)−5 Answer: 1120

(c) [x 8 ](x 3 + 5x 4 )(1 + 3x)6 Answer: 7533
(d) [x 9 ]{(1 − 4x)5 + (1 − 3x)−2 } Answer: 196830
(e) [x n ](1 − 2t x)−k Answer: n+k−1 (2t )n
¡ ¢
n
¡ n+k ¢ n+1−k
(f) [x n+1 ]x k (1 − 4x)−2k Answer: n+1−k 4
¡ n−k
2 +m−1
(g) [x n ]x k (1 − x 2 )−m
¢
Answer: n−k
2
¡ n +k−1¢ ¡ n +k−1¢ n
(h) [x n ]{(1 − x 2 )−k + (1 − 7x 3 )−k } Answer: 2
n + 3 n 73
2 3
3. (a) Write out the first 5 terms and the nth term of the following power se-
ries:
(i) (1 − x)−1
(ii) (1 + x)−1
(iii) (1 − x)−2
(iv) (1 + x)−2
(v) (1 − x)−3
(vi) (1 + x)−3
(b) Determine the appropriate coefficient in each case.
(i) [x 8 ](1 − x)−1 , Answer: 1
5 −1
(ii) [x ](1 + x) , Answer: −1
(iii) [x 6 ](1 − 3x 2 )−1 , Answer: 27
12 −2
(iv) [x ](1 − x) , Answer: 13
9 3 −3
(v) [x ](1 + 4x ) , Answer: −640
8 3 −2
(vi) [x ](1 − 2x ) . Answer: 0
(c) Write each of the following power series in closed form.
(i) 1 − 2x + 3x 2 − 4x 3 + 5x 4 − 6x 5 + . . .
(ii) 1 + 3x + 6x 2 + 10x 3 + 15x 4 + 21x 5 + . . .
(iii) 1 − x 3 + x 6 − x 9 + x 12 − x 15 + . . .
(iv) 1 + 2x 2 + 4x 4 + 8x 6 + 16x 8 + 32x 10 + . . .
(v) 1 − 4x 2 + 12x 4 − 32x 6 + 80x 8 − 192x 10 + . . .
(vi) 1 + 6x + 24x 2 + 80x 3 + 240x 4 + 672x 5 + . . .
4. The coefficient notation can be extended to series of more than one variable.
Thus [x n y k ] denotes the coefficient of x n y k in the expansion of the series
that follows.
Find [x n y k ] 1x y . [Hint: Treat y as a constant, and find the coefficient of x n ;
1− 1−y
then find the coefficient of y k in the result.]
5. Show that
¡n+2k−1¢
(a) [x n ](1 − 2x + x 2 )−k = n
¡i +k−1¢¡ n−i
2 +k−1
(b) [x n ](1 − x − x 2 + x 3 )−k =
P ¢
i ≥0 n−i
n−i ≡0 (mod 2)
i
2
1−x 2 1
6. (a) Prove that 1+x 3
= x 2 .
1+ 1−x
(b) By expanding each side of the identity in (a) as a power series, and con-
sidering the coefficient of x N , prove that
¯ Ã !¯ ½
k N −k −1 ¯ 0 if N ≡ 1 (mod 3)
¯X ¯
¯ (−1) ¯=
¯
¯k≥0 N − 2k ¯ 1 otherwise.
7. Calculate [x n ](1 + x)−2 (1 − 2x)−2 , and give the simplest expression you can
find.
8. Prove that r +s=t (−1)r n+rr −1 ms = m−n

P ¡ ¢¡ ¢ ¡ ¢
t
.
9. Prove that nm=k m

P ¡ ¢ ¡n+1¢ −1 −(k+1)
k = k+1 . (Hint: Use the identity (1 − x) (1 − x) =
−(k+2)
(1 − x) .)
10. Using the Binomial Theorem, show that

Ã !
− 32
X (k + 1) 2k + 1 k
(1 − x) = 2k
x .
k≥0 2 k
R1 1
11. By using the fact that 0 x i d x = i +1 for i ≥ 0, show that
1 n 1
(a) ni=0 i +1 n+1
P ¡ ¢
i = n+1 {2 − 1}.
Pn (−1)i ¡n ¢ Pn 1
(b) i =1 i i = − k=1 k
i +1 ¡ ¢
(c) ni=0 2i +1 ni = n+11
{3n+1 − 1}.
P
3.8. THE QUICKSORT ALGORITHM 83
(a) Let F (x) = (i 1 ,...,i k )∈S g (i 1 , . . . , i k )x h(i 1 ,...,i k ) , where h(i 1 , . . . , i k ) is an integer-
P
12.
valued function of i 1 , . . . , i k and g (i 1 , . . . , i k ) is independent of x. Show
that
[x n ]F (x) =
X
g (i 1 , . . . , i k ).
n
(i 1 ,...,i k )∈S
h(i 1 ,...,i k )=n
(b) Prove that

Ã !Ã !
3i j a + j − 1 2a
[x n ]e 3x (1 − 5x 2 )−a (1 + x 3 )2a =
X
5
i , j ,k≥0 i! j k
i +2 j +3k=n
b n3 c b n−3k
2 c
Ã !Ã !
3n−2 j −3k a + j − 1 2a
5j
X X
= .
k=0 j =0 (n − 2 j − 3k)! j k
3.8 The Quicksort Algorithm

Suppose that we are given a list of n distinct numbers x 1 , . . . , x n , and are asked
to sort them (put them in increasing order, say). Consider the following proce-
dure:
1. Compare x 1 with each of x 2 . . . , x n , thus forming two lists, one consisting

of those elements of x 2 , . . . , x n which are bigger than x 1 , and those which
are smaller.
2. Repeat (1) for each of these lists unless their length is 0 or 1. This is called
“Quicksort”, and is implemented as follows.
Example 3.8.1. Sort the list 4 2 8 9 3 5.
STEP 1.
23 895
smaller bigger
STEP 2.
2 8
3 5 9
and we’re finished. We have constructed a binary tree, which will be discussed
later.
Problem 3.8.2. Assuming the original list is in random order (i.e. all n! possi-
ble orderings are equiprobable), what is the average number of comparisons
needed by this algorithm to completely sort a list of n distinct numbers?
Solution: Let this average value be a n . Then from the algorithm,
1X n
an = n − 1 + (a i −1 + a n−i ). (3.8.1)
n i =1
The n − 1 arises from comparing a 1 to a 2 , . . . , a n . Further, since the list is in ran-

dom order, it is equally likely (probability = n1 ) that a 1 is the i t h ordered element
in the list, for i = 1, . . . , n. But if a 1 is the i t h ordered element, we must then sort
the lists of smaller and larger elements, of lengths i − 1 and n − i respectively.
For initial conditions, we have a 0 = 0.
Multiply both sides of (3.8.1) by n, to give
n−1
X
na n = n(n − 1) + 2 aj, n ≥ 1.
j =0
Now multiply both sides by x n−1 and sum for n ≥ 1. Then
n−1
na n x n−1 = n(n − 1)x n−1 + 2 x n−1
X X X X
aj
n≥1 n≥1 n≥1 j =0
∞
na n x n−1 = x n(n − 1)x n−2 + 2 x n−1 ,
X X X X
aj
n≥1 n≥2 j ≥0 n= j +1
since
X n−1
X X X X X
= = .
n≥1 j =0 n≥ j j ≥0 j ≥0 n> j
Now let
an x n
X
A(x) =
n≥0
so that
A 0 (x) = na n x n−1 .
X
n≥1
Then the above equation becomes
xj
A 0 (x) = x n(n − 1)x n−2 + 2
X X
aj
n≥2 j ≥0 1−x
which gives
A 0 (x) = 2x(1 − x)−3 + 2(1 − x)−1 A(x).
Hence we must solve the differential equation
A 0 (x) − 2(1 − x)−1 A(x) = 2x(1 − x)−3 . (3.8.2)
This type of differential equation is called a linear differential equation of

the first order. The general form is
dy
+ P (x)y = Q(x)
dx
where P and Q are functions of x alone. There is a standard method of solution
for such equations.
The equation is essentially the same if we multiply both sides by any func-
tion of x, say µ(x). We get
dy
µ(x) + µ(x)P (x)y = µ(x)Q(x) (3.8.3)
dx
Now the left hand side of (3.8.3) starts off like the derivative of the product of
two functions µ(x) and y. We would have
d dy
[µ(x)y] = µ(x) + µ0 (x)y.
dx dx
In fact, if we choose µ(x) so that µ0 (x) = µ(x)P (x) then the left hand side of
(3.8.3) would be the derivative of µ(x)y. So let us take
µ0 (x)
= P (x)
µ(x)
or, on integrating, log µ(x) = P (x)d x. Then provided we can integrate P (x) we
R
have the required µ(x). The equation now becomes

d
[µ(x)y] = µ(x)Q(x)
dx
and another integration gives
Z
µ(x)y = µ(x)Q(x)d x +C
which gives us the solution, y.

The function µ(x) is called an integrating factor. In our example we have
y = A(x), P (x) = −2(1 − x)−1 and Q(x) = 2x(1 − x)−3 .
R R 1
P (x)d x = −2 1−x dx
= 2 log(1 − x)
making µ(x) = (1 − x)2 . Multiplying by this we get
(1 − x)2 A 0 (x) − 2(1 − x)A(x) = 2x(1 − x)−1
or
d 2x
dx
[(1 − x)2 A(x)] = 1−x
2
= 1−x −2
Integrating, we have
(1 − x)2 A(x) = −2 log(1 − x) − 2x + c .
Hence
A(x) = {−2 log(1 − x) − 2x + c}(1 − x)−2 .
We know that A(0) = 0, so c = 0. To find a n , we expand A(x) in powers of x.
A(x) = {2 log(1 − x)−1 − 2x}(1 − x)−2
xk
+ 1)x i .
P P
=2 k≥2 k i ≥0 (i
Hence
i +1
a n = [x n ]A(x) = 2
P P
k≥2 i ≥0 k
i +k=n
Pn n−k+1 1
© Pn ª
=2 k=2 k
= 2(n + 1) k=2 k
− 2(n − 1) .
Thus ( )
Xn 1
a n = 2(n + 1) − 4n ,
k=1 k
is the average number of comparisons needed. This might not give you much
feel for how big a n is, but a result from advanced calculus (which we won’t go
into) tells us that nk=1 k1 is asymptotically log n+γ where γ is a constant, known
P
as Euler’s constant. Its value is 0.5772... . It follows from this that
a n ∼ 2n log n.
Exercise: Although the result just quoted is outside the scope of this course, it
is fairly easy to show that
1 1
log(n + 1) < 1 + + · · · + < log n + 1
2 n
by considering the area under the curve y = x1 . (Use the integral test for series
convergence in first year calculus).
Problem Set 3.8

1. Suppose that
X xn
F (x) = fn
n≥0 n!
0
satisfies the differential equation F (x) = (1 + 2x)F (x), with F (0) = 3. Find a
recurrence equation for f n , together with initial conditions.
2. Suppose that
X xn
A(x) = an
n≥0 n!
satisfies the differential equation
(1 − x)A 0 (x) = 5(1 + x)A(x),

with A(0) = 2. Find the recurrence relation for a n , n ≥ 0 together with initial
conditions.
3. If
X xn
F (x) = fn
n≥0 n!
F 00 (x) + (1 − x)F 0 (x) − x 2 F (x) = 0,
with F (0) = 1, F 0 (0) = 2, find a recurrence for f n , together with initial condi-
tions.
4. Suppose
∞
X xk
f (x) = fk
k=0 k!
(1 + 2x)2 f 0 (x) = 4x f (x)
with the boundary condition f (0) = 1.
(a) Find a recurrence equation for f n .

(b) Solve the differential equation to find f (x).
(c) Find an expression for f n in terms of n, for n ≥ 2.
3.9 The Mergesort Algorithm

A list of n distinct numbers can also be sorted by the following procedure:
1. Divide the n numbers into two groups, which have¥ sizes

¦ as close to equal
n
as possible, (i.e., one group consists
§n ¨ of the first 2 numbers, the sec-
ond group consists of the last 2 numbers, where dxe = smallest integer
which is not less than x).
2. Sort these two groups individually by applying stage (1) to them and re-
peat, until each group has size 1. When these groups are sorted, merge
them to create an ordered list.
3.9. THE MERGESORT ALGORITHM 89
4 9
2 8 3 5
Figure 3.2: Some graphs
This is called “Mergesort”. We might implement it to sort 4 2 8 9 3 5 by first

creating a binary tree with subtrees of an equal size as possible, e.g.
We now merge successive pairs of subtrees, as follows:
STEP 1.
4 28 9 35
STEP 2.
248 359
STEP 3. 2 3 4 5 8 9.
We wish to find a n , the number of comparisons needed to sort such a list. Then,
from the algorithm,
a n = a b n c + a d n e + (number of comparisons needed to merge.)

2 2
But the number of comparisons needed to merge lists of sizes n2 and n2 is at

§ ¨ ¥ ¦
most n − 1, so
an ≤ n − 1 + ab n c + ad n e (3.9.1)
2 2
Thus if we solve the recurrence
t n = n − 1 + t b n c + t d n e , n ≥ 2,
2 2
with initial condition t 1 = 0, we will know that a n ≤ t n , and we have found an

upper bound for the number of comparisons needed. The determination of the
average number of comparisons is difficult, but the upper bound turns out to
be quite close to the actual average. To solve our recurrence for t n , multiply by
x n and sum for n ≥ 2, giving
tn x n = (n − 1)x n + (t b n c + t d n e )x n ,
X X X
2 2
n≥2 n≥2 n≥2
P n
and letting T (x) = n≥1 t n x , t 1 = 0, gives
i −1
T (x) = x 2 2m
+ t k+1 )x 2k+1
P P P
i ≥0 i x + m≥1 2t m x + k≥1 (t k
i −1
= x2 + 2T (x 2 ) + xT (x 2 ) + x −1 T (x 2 ), since t 1 = 0
P
i ≥0 i x
= x 2 (1 − x)−2 + (2 + x + x −1 )T (x 2 )
Hence
T (x) = x 2 (1 − x)−2 + x −1 (1 + x)2 T (x 2 ),
and we must solve this functional equation. First let S(x) = x −2 (1 − x)2 (T (x), so
that S(x 2 ) = x −4 (1 − x 2 )2 T (x 2 ). The equation for T (x) becomes
S(x) = 1 + xS(x 2 ). (3.9.2)
Applying this iteratively gives
S(x) = 1 + x(1 + x 2 S(x 4 ))
= 1 + x(1 + x 2 (1 + x 4 (. . .)))
= 1 + x + x3 + x7 + · · ·
P 2i −1
= i ≥0 x ,
which easily checks to satisfy (3.9.2) above. (Exercise).

3.9. THE MERGESORT ALGORITHM 91
Thus we want t n = [x n ]T (x), where
2i −1
T (x) = x 2 (1 − x)−2
P
i ≥0 x
2i −1
= x2 + 1)x j
P P
j ≥0 ( j i ≥0 x .
Therefore t n = i ( j +1), where the summation is for i ≥ 0, j ≥ 0, and 2i +1+ j =

P
n. Let m be the largest possible value for i , given by 2m ≤ n −1, or m = blog2 (n −

1)c. Then
m 2m+1 − 1
(n − 2i ) = (m + 1)n − = (m + 1)n + 1 − 2m+1
X
tn =
i =0 2−1
where m = blog2 (n − 1)c. More usefully, we have
t n ∼ n log2 n,
since n ≤ 2m+1 ≤ 2(n − 1). Finally, recall that this gives an upper bound on the
average number of comparisons needed by MERGESORT.
If you take a Computer Science course on the analysis of algorithms, you
will spend a lot of time developing asymptotic results of the above types (though
by quite different methods).
Chapter 4
Introduction to Graph Theory
4.1 Definitions
Graph theory is the study of mathematical objects known as “graphs" — a word
to which graph theorists have given a rather special meaning. So we must start
by defining exactly what a graph is.
Definition 4.1.1. A graph G is a finite nonempty set, V (G), of objects, called

vertices, together with a set, E (G), of unordered pairs of distinct vertices. The
elements of E (G) are called edges.
For example, we might have
V (G) = {1, 2, 3, 4, 5}
and
E (G) = {{1, 2}, {1, 3}, {1, 4}, {2, 3}, {2, 5}, {3, 4}, {3, 5}, {4, 5}}.
For the sorts of results with which we are concerned, it is most convenient to
consider the following geometric representation or diagram or drawing of a
graph. On the page we draw a small circle to correspond to each vertex. For
each edge we then draw a line between the corresponding pair of vertices. The
only restriction on such a line is that it does not intersect the circle correspond-
ing to any other vertex. For example, the above graph is represented in Figure
4.1(i), (ii) and (iii) in three ways.
If e = {u, v} then we say that u and v are adjacent vertices, and that edge e
is incident with vertices u and v. We can also say that the edge e joins u and v.
93
94 CHAPTER 4. INTRODUCTION TO GRAPH THEORY
2
2 2
1 1
1 3 5
3 3
5 5
4
4 4
(i) (ii) (iii)
Figure 4.1: Three drawings of the same graph
Vertices adjacent to a vertex u are called neighbours of u. The set of neighbours

of u is denoted N (u). A graph is completely specified by the pairs of vertices
that are adjacent, and the only function of a line in the drawing, representing
an edge, is to indicate that two vertices are adjacent. In (i) of Figure 4.1, no edge
crosses another; in (ii) of Figure 4.1, the edge {1, 4} crosses edges {2, 5} and {3, 5}.
A graph which can be represented with no edges crossing is said to be planar,
so our graph G is planar by Figure 4.1(i).
We give a few examples to illustrate the variety of settings in which graphs
arise.
Example 4.1.2. The word graph Wn is the graph having V (Wn ) equal to the set
of all English words having exactly n letters. Two words are adjacent if one can
be obtained from the other by replacing exactly one letter by another (in the
same position). For example, seat is adjacent to sent in W4 . In Figure 4.2 we
show a drawing of part of W3 .
Example 4.1.3. Given the street map of a city, one can define a street map graph
as follows. There is a vertex for each street intersection, and an edge for each
part of a street joining two intersections and traversing no other intersections.
An example is given in Figure 4.3, where intersections are numbered to make
the correspondence clear. Such graphs are useful in solving certain kinds of
routing and scheduling problems, such as garbage pickup, or delivery of news-
papers to carriers.
4.1. DEFINITIONS 95
rat
ran
cat eat
run
ear
sun
can car
Figure 4.2: Part of W3
1 2 Euler Way 3 1 2 3
Queen St
King St
Jack St
4 5 Main St 6 4 5 6
7 Erb St 8 7 8
Figure 4.3: Map of Smalltown and its Graph

Example 4.1.4. Another way to obtain a graph from a map, is to begin with a
political map, such as the map of the countries of a continent. There is a vertex
for each country and two countries are adjacent if they share a boundary. One
of the most famous problems in graph theory arose from the question of how
many colours are needed to colour such maps so that adjacent countries are
not assigned the same colour. Figure 4.4 shows the graph obtained from South
America.
Gu
Su
Ve FG
Co
Br Ur
Ec
Pa
Pe Bo Ar
Ch
Figure 4.4: Graph of South American Countries
Example 4.1.5. Graphs are often defined from other mathematical objects. For
example, we can define the graph S n,k to have V (S n,k ) equal to the set of k-
element subsets of {1, 2, . . . , n}. Two such subsets are adjacent if they have ex-
actly k − 1 elements in common. Figure 4.5 shows a drawing of the graph S 4,2 .
Some important points arise from our definition of a graph.
(1) Edges are unordered pairs of vertices. Thus the edge {v 1 , v 2 } is not from v 1
to v 2 or vice versa; it is simply “between" v 1 and v 2 . If we change Definition
4.1.1 to read “ordered pairs" we obtain the definition of a different kind of
graph, a directed graph or digraph.
(2) E (G), being a set, either contains a pair {v 1 , v 2 } or it does not. Thus we
do not allow the possibility of “multiple edges” such as exist between the
vertices a and b in Figure 4.6.
4.1. DEFINITIONS 97
{1, 2}
{2, 3} {2, 4}
{3, 4}
{1, 3} {1, 4}
Figure 4.5: A drawing of S 4,2
d a
c e
f b
Figure 4.6: A multigraph

(3) The edges are pairs of distinct vertices. Hence we cannot have a loop, i.e.,
an edge joining a vertex to itself as shown in Figure 4.6 at vertex c. Neverthe-
less in some circumstances it can be convenient to consider loops and/or
multiple edges. If we wish to allow loops and multiple edges we will use the
term multigraph instead of graph. (In some texts “graph” is used to mean
“multigraph”, and if loops and multiple edges are not allowed the term ”sim-
ple graph” is used.)
(4) Note that V (G), and hence E (G), is a finite set. If we remove this condition
we find ourselves in the realm of infinite graphs—and that is a whole new
ballgame!
4.2 Isomorphism
Figure 4.7(i) is the diagram of the graph G, where
V (G) = {p, q, r, s}, E (G) = {{p, q}, {p, r }, {q, r }, {q, s}}.
Figure 4.7(ii) is the diagram of the graph H , where
V (H ) = {a, b, c, d }, E (H ) = {{a, b}, {a, c}, {a, d }, {c, d }}.
The graphs G and H are not the same—G has vertices p, q, r, s and H has ver-
tices a, b, c, d —but for almost all purposes they are indistinguishable. We make
this idea precise.
p c
q s a b
r d
(i) (ii)
Figure 4.7: Isomorphic graphs

4.2. ISOMORPHISM 99
Definition 4.2.1. Two graphs G 1 and G 2 are isomorphic if there exists a bijec-
tion f : V (G 1 ) → V (G 2 ) such that vertices f (u) and f (v) are adjacent in G 2 if and
only if u and v are adjacent in G 1 . (We might say that f preserves adjacency.)
The bijection f in this definition that preserves adjacency is called an iso-

morphism. For example, an isomorphism from G to H is g , defined by
g (p) = c, g (q) = a, g (r ) = d , g (s) = b.
Another isomorphism is h, defined by
h(p) = d , h(q) = a, h(r ) = c, h(s) = b.
Figure 4.8 shows two other graphs, G and H , that are isomorphic. One iso-
morphism is the mapping f : V (G) → V (H ) given by
f (1) = a, f (2) = b, f (3) = c, f (4) = h,

f (5) = i , f (6) = j , f (7) = d , f (8) = e,
f (9) = f , f (10) = g .
(This graph is called the Petersen graph.)
1 a
b i
6
5 2
8 9 c h
j
10 7
d g
4 3
e f
G H
Figure 4.8: Two drawings of the Petersen graph
The collection of graphs that are isomorphic to G forms the isomorphism

class of G. In almost all cases, a graph has some property if and only if all graphs
in its isomorphism class have the property. Thus we generally regard isomor-
phic graphs as ‘the same’ even if formally they might not be equal. Even if G
has only one vertex, there are infinitely many graphs in its isomorphism class.
Fortunately though, the number of isomorphism classes of graphs with a given
finite set of vertices is finite. For example, there are exactly 11 isomorphism
classes of graphs on 4 vertices, pictured in Figure 4.9. Note that in this figure,
the vertices of the graphs are not given explicitly, because however we assign
vertices to the drawing, we will still get a graph in the same isomorphism class.
The identity map on V (G) is an isomorphism from the graph G to itself. An
isomorphism from G to itself is called an automorphism of G.
Figure 4.9: The graphs on 4 vertices, up to isomorphism
4.3 Degree
The number of edges incident with a vertex v is called the degree of v, and is
denoted by deg(v). For example in G of Figure 4.7 we have deg(p) = deg(r ) =
2, deg(q) = 3, deg(s) = 1; in G of Figure 4.8, all vertices have degree 3. In what
follows we generally use ‘p’ for the number of vertices and ‘q’ for the number
of edges.
4.4. BIPARTITE GRAPHS 101
Theorem 4.3.1. For any graph G we have

X
deg(v) = 2|E (G)|.
v∈V (G)
Proof: Each edge has two ends, and when we sum the degrees of the vertices,
we are counting the edges twice, once for each end.
This is known as the Handshaking Lemma or the Degree-Sum Formula.
Corollary 4.3.2. The number of vertices of odd degree in a graph is even.
Proof: The sum of all vertex degrees is 2q, an even number. The sum of the
vertices of even degrees is even. Hence the sum of the vertices of odd degrees
is also an even number. This implies that there must be an even number of
vertices with odd degree.
Corollary 4.3.3. The average degree of a vertex in the graph G is

2|E (G)|
.
|V (G)|
A graph in which every vertex has degree k, for some fixed k, is called a k-
regular graph (or just a regular graph). We note one important class of regular
graphs.
Definition 4.3.4. A complete graph is one in which all pairs of distinct vertices
are adjacent. (Thus each vertex is joined to every other vertex). The complete
graph with p vertices is denoted by K p , p ≥ 1.
In K n each vertex is adjacent to the n − 1 vertices distinct from¡it,¢ thus K n is
regular with degree n − 1. The number of edges in K n is therefore n2 , following
Corollary 4.3.2.
Figure 4.10 shows K 4 , the complete graph on 4 vertices, and therefore 6
edges.
4.4 Bipartite Graphs

A graph in which the vertices can be partitioned into two sets A and B , so that
all edges join a vertex in A to a vertex in B , is called a bipartite graph, with bipar-
tition (A, B ). The complete bipartite graph K m,n has all vertices in A adjacent
to all vertices in B , with |A| = m, |B | = n. For example, Figure 4.11 is a drawing
of K 2,3 .
Figure 4.10: K 4
Figure 4.11: The complete bipartite graph K 2,3
Definition 4.4.1. For n ≥ 0, the n-cube is the graph whose vertices are the {0, 1}-
strings of length n, and two strings are adjacent if and only if they differ in ex-
actly one position.
For example, Figure 4.12 shows the 3-cube.
Problem 4.4.2. Determine the numbers of vertices and edges in the n-cube, for
n ≥ 0.
Solution: The number of {0, 1}-strings of length n is 2n , for n ≥ 0, so the n-cube

has p = 2n vertices. Also, every vertex has degree n, since a string of length n is
adjacent to the n strings that can be obtained by switching each single element
in the string in turn. Thus the n-cube is an n-regular graph, and Theorem 4.3.1
gives
X2n
n = 2q
i =1
n2n = 2q,
so the n-cube has q = n2n−1 edges, for n ≥ 0.
000 001
010 011
100 101
110 111
Figure 4.12: The 3-cube
Problem 4.4.3. Show that the n–cube is bipartite, for n ≥ 0.
Solution: Let V be the set of all {0, 1}–strings of length n; V is the vertex set of
the n-cube. Partition V into the set A of strings containing an even number of
ones and the set B of strings containing an odd number of ones. If vertices x
and y are adjacent then the strings x and y differ in exactly one position. Thus,
exactly one of x and y contain an even number of ones. Therefore (A, B ) is a
bipartition and hence the n-cube is bipartite.
Problem Set 4.4

1. For the graphs G 1 ,G 2 ,G 3 and H in Figure 4.13, prove that no two of G 1 ,G 2 or
G 3 are isomorphic. Prove that one of them (which?) is isomorphic to H by
giving a suitable bijection.
2. A cubic graph is one in which every vertex has degree three. Find all the
nonisomorphic cubic graphs with 4, 6 and 8 vertices.
3. For the subset graph S n,k defined in Example 4.1.5, find the number of ver-
tices and the number of edges.
4. The odd graph O n is the graph whose vertices are the n-subsets of a (2n +1)-
set, two such subsets being adjacent if and only if they are disjoint.
G1 G2 G3
Figure 4.13: Isomorphism exercise
(a) Draw O 1 and O 2 .

(b) Prove that O 2 is isomorphic to the Petersen graph (see Figure 4.8).
(c) How many vertices and edges does O n have?
5. The line-graph L(G) of a graph G is the graph whose vertex set is E (G) and
in which two vertices are adjacent if and only if the corresponding edges of
G are incident with a common vertex.
(a) Find a graph G such that L(G) is isomorphic to G.

0 0
(b) Find nonisomorphic graphs G,G such that L(G) is isomorphic to L(G ).
(c) If G is the graph
find L(G), L(L(G)) and L(L(L(G))).

6. For integer n ≥ 0, define the graph G n as follows: V (G n ) is the set of all binary
strings of length n having at most one block of 1’s. Two vertices are adjacent
if they differ in exactly one position.
(a) Find |V (G n )|
(b) Make drawings of G 3 and G 4 .
(c) Find |E (G n )|.
7. (a) Draw K m,n for all m, n such that 1 ≤ m ≤ n ≤ 3.

(b) How many vertices and edges does K m,n have?
(c) Let K be a complete bipartite graph on p vertices. Prove that K has at
most bp 2 /4c edges.
(d) Let G be a bipartite graph on p vertices. Prove that G has at most bp 2 /4c
edges.
(e) Let G be a k-regular bipartite graph with bipartition (X , Y ). Prove that
|X | = |Y | if k > 0. Is this still valid when k = 0?
8. The complement of the graph G, denoted Ḡ is the graph with V (Ḡ) = V (G)
and the edge {u, v} ∈ E (Ḡ) if and only if {u, v} 6∈ E (G).
(a) Let G have vertices 1, 2, 3, 4 and edges {1, 2}, {2, 3}, {3, 4}, {1, 4}. Draw Ḡ.
(b) Find a 5-vertex graph that is isomorphic to its complement.
(c) Prove that no 6-vertex graph is isomorphic to its complement.
(d) Let G 1 and G 2 be two graphs. Prove that G 1 is isomorphic to G 2 if and
only if Ḡ 1 is isomorphic to G¯2 .
(e) Find all 2-regular non-isomorphic graphs on 6 vertices (prove that these
are the only ones).
(f) Prove that there are only two 3-regular non-isomorphic graphs on 6
vertices.
9. Make drawings of the 15 nonisomorphic graphs having six vertices and six
edges, such that every vertex has degree at least one.
10. Are the graphs in Figure 4.14 isomorphic? Justify your answer.
1 2 a b
8 3 h c
7 4 g d
6 5 f e
11. For n a positive integer, define the prime graph B n to be the graph with
vertex set {1, 2, . . . , n}, where {u, v} is an edge if and only if u + v is a prime
number. Prove that B n is bipartite.
12. (a) Are the two graphs in Figure 4.15 isomorphic? Prove your claim.
(b) Are the two graphs in Figure 4.16 isomorphic? Prove your claim.
1 2 A B
8 3 H C
7 4 G D
6 5 F E
4.5 How to Specify a Graph

One way of specifying a particular graph is to display a drawing of it; but this is
not always convenient. Another method is by means of adjacency or incidence
matrices.
4.5. HOW TO SPECIFY A GRAPH 107
1 2 A B
7 3 G C
6 4 F D
5 E
Definition 4.5.1. The adjacency matrix of a graph G having vertices v 1 , v 2 , . . . , v p

is the p × p matrix A = [a i j ] where
(
1, if v i and v j are adjacent;
ai j =
0, otherwise.
Clearly A is a symmetric matrix and, since we do not allow loops, its diago-
nal elements are all zero.
To define an incidence matrix we must name the edges of G; we shall call
them e 1 , e 2 , . . . , e q .
Definition 4.5.2. The incidence matrix of a graph G with vertices v 1 . . . v p and

edges e 1 . . . e q is the p × q matrix B = [b i j ] where
(
1, if v i is incident with e j ;
bi j =
0, otherwise.
Clearly, each column of B contains exactly two 1’s.

Consider the product B B t . Its (i , j )-element is
q
X
bi k b j k .
k=1
For i 6= j this sum is the number of edges incident with both v i and v j ; for i = j
it is the number of edges incident with v i , which is deg(v i ). Thus
B B t = A + diag(deg(v i ), . . . , deg(v p )).

Another way of specifying a graph is to give, for each vertex, a list of the
vertices adjacent to it. For example:
vertex adjacent vertices

1 3457
2 357
3 12456
4 1356
5 1234
6 347
7 126
This is called an adjacency list. In practice this would be stored as a dictio-

nary or hash.
Finally we could also specify a graph by giving its vertex set together with
the list of its edges. This is likely to be most useful when the graph is sparse, i.e.,
does not have many edges.
It is important to note that in many cases the vertices of the graph may not
arrive as non-negative integers. There were examples of this in Problem Set 4.4.
Problem Set 4.5

1. (a) Find the adjacency matrix A and the incidence matrix B for the graph
in Figure 4.17.
(b) Give an interpretation of
(i) the diagonal elements of the matrix A 2 ; and
(ii) the off-diagonal elements of the matrix A 2 .
2. (a) Order the vertices· of the graph

¸ in Figure 4.18 so that the adjacency ma-
0 M
trix has the form for some matrix M.
Mt 0
(b) Prove that any bipartite graph has a vertex ordering which gives an ad-
jacency matrix of the above form for some M.
4.5. HOW TO SPECIFY A GRAPH 109
e1 e4
e2 e3
5 6 7 2
e5 e6 e7
e8 e9 e11 e12
4 3
e10
Figure 4.17: Graph for problem 1
Figure 4.18: Graph for problem 2

4.6 Paths and Cycles

A subgraph of a graph G is a part of G (or possibly the whole of G). More rigor-
ously we have
Definition 4.6.1. A subgraph of a graph G is a graph whose vertex set is a subset

U of V (G) and whose edge set is a subset of those edges of G that have both
vertices in U .
Thus if H is a subgraph of G, then H has some (perhaps all) the vertices of

G and some (perhaps all) the edges that, in G, join vertices of H . (Clearly we
cannot have any edges in H that involve vertices not in H ). If V (H ) = V (G), that
is H has all vertices of G, we say H is a spanning subgraph of G.
If H is a subgraph of G and H is not equal to G, we say it is a proper sub-
graph of G. If D is a subgraph of G and C is a subgraph of D, then C is a sub-
graph of G.
A walk in a graph G from v 0 to v n , n ≥ 0, is an alternating sequence of ver-
tices and edges of G
v 0 e 1 v 1 e 2 v 2 . . . v n−1 e n v n
which begins with vertex v 0 , ends with vertex v n and, for 1 ≤ i ≤ n, edge e i =
{v i −1 , v i }. Such a walk can also be called a v 0 , v n -walk. Note that the length of a
walk is the number of edges in it (in this case, n). Also, because we can reverse a
walk to get a walk from v n to v 0 , we refer, where convenient, to a walk between
a pair of vertices, instead of from one to the other. A walk is said to be closed if
v0 = vn .
A path is a walk in which all the vertices are distinct. A path that starts at
v 0 and ends at v n is called a v 0 , v n -path. Observe that since all the vertices in a
path are distinct, so are all the edges.
Figure 4.19 shows (by heavy lines) a path in a graph, from vertex 1 to vertex
2.
Since graphs have no multiple edges, consecutive vertices v i −1 and v i de-
termine the edge e i of a walk. Hence, in describing a walk we often omit the
edges.
Theorem 4.6.2. If there is a walk from vertex x to vertex y in G, then there is a

path from x to y in G.
Proof: We use a form of proof by contradiction.

4.6. PATHS AND CYCLES 111
5 6
1 4
2 3
8
7 9
10
Figure 4.19: A path
Suppose that v 0 , . . . , v n is a walk in G from v 0 = x to v n = y with minimal

length (among all walks from x to y). If the walk is not a path, then v i = v j
for some i < j , and then v 0 . . . v i v j +1 . . . v n would be a walk from x to y whose
length is less than the minimum possible. But this contradicts our initial choice,
and so we conclude that the vertices in our original walk are all distinct, and
therefore it is a path.
Corollary 4.6.3. Let x, y, z be vertices of G. If there is a path from x to y in G and

a path from y to z in G, then there is a path from x to z in G.
Proof: The path from x to y followed by the path from y to z is a walk from x to
z in G, so by Theorem 4.6.2 there is a path from x to z in G.
A cycle in a graph G is a subgraph with n distinct vertices v 0 , v 1 . . . v n−1 , n ≥
1, and n distinct edges {v 0 , v 1 }, {v 1 , v 2 }, . . . {v n−2 , v n−1 }, {v n−1 , v 0 }. Equivalently,
a cycle is a connected graph that is regular of degree two. (The definition of
connectedness is given in Section 4.8.)
The subgraph we get from a cycle by deleting one edge is called a path. This
is inconsistent with the original definition of path, because we defined a path
to be a type of walk, and a walk is a sequence of vertices and edges and so is not
a graph. But it is very convenient to use the word path in both senses, and we
will.
A cycle with n edges is called an n-cycle or a cycle of length n. A cycle of
length 1 has one vertex, v 0 , and one edge {v 0 , v 0 } which is a loop. Therefore,
since graphs have no loops, a graph has no cycle of length one. A cycle of length
2 has two vertices, v 0 and v 1 , and two distinct edges, called multiple edges, join-
ing v 0 and v 1 . Therefore, a graph has no cycle of length two. Thus the shortest
possible cycle in a graph is a 3-cycle, often called a triangle.
Note that for a cycle with n ≥ 3 vertices, v 0 , . . . , v n−1 , then v i v i +1 . . . v n−1 v 0 . . . v i
and v i v i −1 . . . v 0 v n−1 . . . v i are both closed walks for each i = 0, . . . , n − 1, and in
this way there are 2n closed walks of length n associated with a given n-cycle.
Figure 4.20 shows (by heavy lines) an 8-cycle in a graph.
5 6
1 4
2 3
8
7 9
10
Figure 4.20: A cycle
In a cycle, every vertex has degree exactly 2. There’s one condition that guar-
antees that a graph contains a cycle.
Theorem 4.6.4. If every vertex in G has degree at least 2, then G contains a cycle.
Proof: Let v 0 , v 1 , . . . , v k be a longest path in G. The vertex v 0 has v 1 as one neigh-

bour. Since v 0 has degree at least 2, v 0 has another neighbour x. If x is not on
the path, then x, v 0 , v 1 , . . . , v k is a path longer than our longest path, which is a
contradiction. Hence x must be on the path, and x = v i for some i ≥ 2. Then
v 0 , v 1 , . . . , v i , v 0 is a cycle in G.
The girth of a graph G is the length of the shortest cycle in G, and is denoted
by g (G). If G has no cycles, then g (G) is infinite (but you may choose to ignore
this fact).
4.6. PATHS AND CYCLES 113
A spanning cycle in a graph is known as a Hamilton cycle (so the cycle in

Figure 4.20 is not a Hamilton cycle). Although it is easy to decide if a given cycle
is a Hamilton cycle, it can be surprisingly difficult to find one in a graph or to
certify that it has no Hamilton cycle.
Problem Set 4.6

1. Let G be a graph with minimum degree k, where k ≥ 2. Prove that
(a) G contains a path of length at least k;

(b) G contains a cycle of length at least k + 1.
2. Let A be the adjacency matrix of a graph G.
(a) Show that the (i , j )t h entry of A k is the number of walks of length k

from i to j .
(b) Assume A satisfies the matrix equation A 2 + A = (k − 1)I + J where I is
the identity matrix and J is the matrix of all 1’s. Explain in graph theory
terms the properties G possesses given by the matrix equation.
3. (a) If A is the adjacency matrix of a graph G, show that for i 6= j , the (i , j )t h

element of A 2 is the number of paths of length 2 in G between vertices
v i and v j .
(b) What are the diagonal elements of A 2 ?
4. Let G be the graph whose set of vertices is the set of all “lower 48” states of
the United States, plus Washington, DC, with two vertices being adjacent if
they share a boundary. (For example, California is adjacent to Arizona.) Let
H be the subgraph of G whose vertices are those of G whose first letter is one
of W, O, M, A, N, and whose edges are the edges of G whose ends have this
property. (For example, California is not a vertex of H , but Arizona and New
Mexico are, and they are adjacent in H .) Find a path in H from Washington
to Washington, DC.
5. Consider the word graph Wn defined in Example 4.1.2.
(a) Find a cycle through math in W4 .

(b) Find a path from pink to blue in W4 .
6. For n ≥ 2, prove that the n-cube contains a Hamilton cycle.
7. Prove that the complete bipartite graph K m,n has a Hamilton cycle if and
only if m = n and m > 1.
8. Show that if there is a closed walk of odd length in the graph G, then G con-
tains an odd cycle (that is, G has a subgraph which is a cycle on an odd num-
ber of vertices).
9. A diagonal of a cycle in a graph is an edge that joins vertices that are not
consecutive in the cycle.
(a) Prove that a shortest cycle (if one exists) has no diagonal.
(b) Prove that a shortest odd cycle (if one exists) has no diagonal.
(c) Give an example of a graph in which a shortest even cycle has a diago-
nal.
10. (a) Prove that a k-regular graph of girth 4 has at least 2k vertices (k ≥ 2).
(b) For k = 2, 3, find a k-regular graph of girth 4 with precisely 2k vertices.
Generalize these examples, i.e. find one for each k ≥ 2.
(c) Prove that a k-regular graph of girth 5 has at least k 2 + 1 vertices (k ≥ 2).
Remark: The only values of k for which such a graph with exactly k 2 +1
vertices can exist are k = 2, 3, 7, 57. This surprising result can be proved
using elementary matrix theory (i.e., what you study in MATH 235). Ex-
amples are known for k = 2, 3, 7, but no example has yet been found for
k = 57. Such a graph would have 572 + 1 = 3250 vertices.
t
(d) Prove that a k-regular graph of girth 2t , where t ≥ 2, has at least 2(k−1)
k−2
−2
vertices.
(e) Prove that a k-regular graph of girth 2t + 1, where t ≥ 2, has at least
k(k−1)t −2
k−2 vertices.
(f) For k = 2, 3, give an example of a k-regular graph of girth five with ex-
actly k 2 + 1 vertices.
4.7. EQUIVALENCE RELATIONS 115
4.7 Equivalence Relations
You have met equivalence relations in your first algebra course, but these are
important in nearly all areas of mathematics, including graph theory.
Formally, if S and T are sets, then a relation R between S and T is a subset
of S × T . The idea is that if a ∈ S and b ∈ T , then a and b are related if and only
if (a, b) belongs to the subset. If a and b are related we may say that they are
incident. Thus if G is a graph, then “is contained in” is a relation on V (G)×E (G).
We will be most concerned with the case where S = T . In this case we usu-
ally refer to a relation on S, and do not mention S × S. By way of example, “is
adjacent to” is a relation on the vertices of a graph. A relation on S is reflexive if
each element of S is related to itself. So “is adjacent to” is not a reflexive relation
on the vertex set V (G) of a graph G. However this relation is symmetric, that is,
if a is related to b then b is related to a. On the integers, the relation “divides” is
reflexive but not symmetric.
There is a third important property a relation may have. Suppose that we
are given a relation on a set V and, if a, b ∈ V , we write a ≈ b to denote that a
and b are related. We say the relation is transitive if whenever a ≈ b and b ≈ c,
then a ≈ c. The relation “divides” on the integers is transitive, as is the relation
≤ on R. The relation “is a subgraph of” on the subgraphs of G is reflexive and
transitive.
We say a relation is an equivalence relation if it is reflexive, symmetric and
transitive.
As an example, the relation “is joined by a walk to” on the vertices of a graph
G is an equivalence relation—each of the three properties is very easy to verify.
The canonical example is equality. Another example you have met is “con-
gruent modulo m” on the integers. We generally use equivalence relations to
partition things. Thus the equivalence relation “congruent mod 5” splits the in-
tegers into five classes. The key is that if ≈ is an equivalence relation on a set V
and C (a) is the set
{v ∈ V : v ≈ a}
then any two elements of C (a) are equivalent, and any element of v that is
equivalent to something in C (a) is itself an element of C (a). Hence if b ∈ V
then either C (a) = C (b) or C (a) ∩C (b) = ;.
4.8 Connectedness
Definition 4.8.1. A graph G is connected if, for each two vertices x and y, there
is a path from x to y.
Theorem 4.8.2. Let G be a graph and let v be a vertex in G. If for each vertex w
in G there is a path from v to w in G, then G is connected.
Proof: For any vertices x and y in G, there is a path from v to x and a path from
v to y. If we reverse the path from v to x we obtain a path from x to v, and now
Corollary 4.6.3 implies that there is a path from x to y in G, so G is connected
by Definition 4.8.1.
If G is a graph on n vertices, then to certify that G is connected
¡n ¢ according
to the definition of connected, we must provide a total of 2 paths. The above
theorem reduces the workload: only n paths are needed.
Problem 4.8.3. Prove that the n-cube is connected for each n ≥ 0.
Solution: We use Theorem 4.8.2, and prove that there is a path from vertex
v 0 = 0 . . . 0 (with n 0’s) to x for all other vertices x in the n-cube. Now x is a {0, 1}-
string of length n, and suppose that x has k 1’s in positions i 1 , . . . , i k , where
1 ≤ i 1 < . . . < i k ≤ n, with 1 ≤ k ≤ n. Now let v j be the {0, 1}-string with 1’s in
positions i 1 , . . . , i j , and 0’s elsewhere for j = 1, . . . , k. Then v 0 v 1 . . . v k is a path
from 0 . . . 0 to x, so the n-cube is connected, by Theorem 4.8.2.
Definition 4.8.4. A component of G is subgraph C of G such that
(a) C is connected.
(b) No subgraph of G that properly contains C is connected.
These two conditions may sometimes be stated in the form “a component

of G is a subgraph which is maximal, subject to being connected”. Here when
we say a subgraph C of G is maximal, subject to having some property, then this
means that any subgraph that properly contains C does not have this property.
Note that we could have defined a component of G to be the equivalence
class of a vertex, relative to the relation “is joined by a walk to”. Since this is
indeed an equivalence relation, it follows immediately that the components of
G partition its vertex set.
4.8. CONNECTEDNESS 117
Figure 4.21: A disconnected graph
Figure 4.21 shows a graph having three components. Note that there are
paths between every pair of vertices in the same component, but not between
pairs of vertices in different components.
While it is easy to convince someone of the existence of a path between two
vertices (show them the path), it is less clear how you might convince them that
a path does not exist. We introduce a convenient way of doing this.
If we are given a partition (X , Y ) of V (G) such that there are no edges having
an end in X and an end in Y , then there is no path from any vertex in X to any
vertex in Y . So, if X and Y are both nonempty, G is not connected. Given a
subset X of the vertices of G, the cut induced by X is the set of edges that have
exactly one end in X .
Theorem 4.8.5. A graph G is not connected if and only if there exists a proper
nonempty subset X of V (G) such that the cut induced by X is empty.
Proof: Let G be a connected graph, and let X be a proper nonempty subset

of V (G). Choose vertices u ∈ X and v ∈ V (G) \ X . Since G is connected, there
exists a path x 0 x 1 . . . x n from u to v. Choose k as large as possible such that
x k ∈ X . Since x n = v 6∈ X , k < n, the edge x k x k+1 is in the cut induced by X and
hence this cut is not empty.
Conversely, suppose that G is not connected and let C be a component of
G. Consider the partition (X , Y ) of V (G) where X = V (C ) and Y = V (G) \ V (C ).
Since C is connected and G is not, X is a proper non-empty subset of V (G).
Since C is a component, if the vertex y is adjacent to a vertex in C , then y ∈ V (C ).
Hence the cut induced by X is empty.
4.9 Eulerian Circuits

One of the earliest problems in graph theory is the Seven Bridges of Königs-
berg problem. In the 18th century in the town of Königsberg (now Kaliningrad,
Russia), there were 7 bridges that crossed the river Pregel, which cuts through
the city and there were 2 islands in the middle of the river. The layout can be
roughly represented in the following diagram:
The question is that can a resident of the city leave home, cross every bridge
exactly once, and then return home? We can formulate the layout of the city as
a graph: Create 4 vertices representing the land areas (two shores and two is-
lands), and create an edge for each bridge, joining the vertices representing the
two land areas on either side of the bridge. In this case, we obtain the following
graph (with multiple edges):
In terms of graph theory, the question becomes “is there a closed walk that
uses every edge exactly once?” We use a definition for this type of walk.
4.9. EULERIAN CIRCUITS 119
Definition 4.9.1. An Eulerian circuit of a graph G is a closed walk that contains

every edge of G exactly once.
In 1736, Swiss mathematician Leonhard Euler noted that there cannot be an
Eulerian circuit for the 7 bridges of Königsberg problem. The reason being that
if an Eulerian circuit exists, each time we visit a vertex, we must use 2 distinct
edges incident with that vertex: use one edge to go to the vertex, use one edge
to leave the vertex. So every vertex must have even degree. However, in this
graph, every vertex has odd degree, so such a walk is not possible.
On the other hand, suppose we have a connected graph and every vertex
has even degree. Does it have an Eulerian circuit? Euler noted that this is indeed
true, and we give a proof here.
Theorem 4.9.2. Let G be a connected graph. Then G has an Eulerian circuit if
and only if every vertex has even degree.
Proof: (⇒) A closed walk contributes 2 to the degree of a vertex for each visit.
Since an Eulerian circuit uses each edge of the graph exactly once, each vertex
in the graph must have even degree.
(⇐) We will prove by strong induction on the number of edges m in G.
Base case: If G has 0 edges, then G consists of exactly one isolated vertex. It
has a trivial closed walk as an Eulerian circuit.
Induction hypothesis: Assume that for some m ≥ 1, any connected graph
with less than m edges where every vertex has even degree has an Eulerian cir-
cuit.
Induction step: Let G be a connected graph with m edges where every vertex
has even degree. This implies that every vertex has degree at least 2. Therefore,
by Theorem 4.6.4, there exists a cycle C in G, say the vertices on the cycle in
order are v 1 , v 2 , . . . , v k , v 1 .
v2
v1
vk
Remove edges of C from G and remove isolated vertices to obtain G 0 . Since

every vertex is incident with either 0 or 2 edges in C , every vertex in G 0 still has
even degree. Now G 0 consists of components C 1 , . . . ,C l , each containing less
than m edges.
C2
C1
C3
By induction hypothesis, each component C i has an Eulerian circuit Wi .

Moreover, each component must have a common vertex with C , for otherwise
G is disconnected. Let v ai be one vertex of C i . Rearrange the components so
that a 1 < a 2 < · · · < a l , and let Wi start and end at v ai .
va 2
va 1
W2
W1
W3
va 3
Then we can construct an Eulerian circuit for G by walking along C and

making detours Wi as we hit v ai :
v 1 , . . . , v a1 −1 ,W1 , v a1 +1 , . . . , v a2 −1 ,W2 , v a2 +1 , . . . . . . . . . , v al −1 ,Wl , v al +1 , . . . , v k , v 1 .

4.10. BRIDGES 121
va 2
va 1
W2
W1
v2
W3
va 3
v1
vk
4.10 Bridges
If e ∈ E (G), we denote by G −e (or by G\e) the graph whose vertex set is V (G) and
whose edge set is E (G)\{e}. (So G − e is the graph obtained from G by deleting
the edge e.)
Definition 4.10.1. An edge e of G is a bridge if G −e has more components than

G.
Thus if G is connected, a bridge is an edge such that G − e is not connected.

Figure 4.22 shows a connected graph with a bridge e. Some texts, and some
instructors, use cut-edge as a synonym for bridge.
Figure 4.22: A bridge

Lemma 4.10.2. If e = {x, y} is a bridge of a connected graph G, then G − e has

precisely two components; furthermore, x and y are in different components.
Proof: Let e = {x, y}. If e is a bridge, then G − e has at least two components. Let
Vx be the set of vertices in the same component of G −e as x. Let z be any vertex
of G − e not in Vx . Because there exists a path from x to z in G but not in G − e,
every path from x to z in G contains edge e and so must be of the form
xe ye 2 v 2 e 3 v 3 · · · v n−1 e n z.
But ye 2 v 2 · · · e n z is a path from y to z in G − e. Thus z is in the same component

of G − e as y, for every vertex z not in Vx . Hence, G − e has 2 components, one
containing vertex x and the other containing vertex y.
Theorem 4.10.3. An edge e is a bridge of a graph G if and only if it is not con-

tained in any cycle of G.
Proof: We begin by proving the following implication: if edge e = {x, y} is an

edge of some cycle of G, then e is not a bridge of G.
By hypothesis, there is a cycle
x, e 1 , v 1 , e 2 , v 2 , . . . , v n−1 , e n , y, e, x
in G. Then
x, e 1 , v 1 , e 2 , v 2 , . . . , v n−1 , e n , y
is a path from x to y in G −e. Hence e is not a bridge of G by Lemma 4.10.2. This
establishes the implication.
To complete the proof, we must establish the converse – if edge e is not a
bridge of graph G, then e is an edge of some cycle. Suppose e = uv is not a
bridge. Then u and v must lie in the same component of G − e, and so there is
a path P that joins them. Together with e this path forms a cycle that contains
e.
Corollary 4.10.4. If there are two distinct paths from vertex u to vertex v in G,
then G contains a cycle.
Proof: Let P 1 and P 2 be distinct paths from u to v. Suppose that P 1 is given by

x 0 x 1 . . . x n and P 2 is given by y 0 y 1 . . . y m . Thus u = x 0 = y 0 and v = x n = y m . Let
i be the first index such that x i +1 6= y i +1 . Then e = {x i , x i +1 } is an edge in P 1 but
not in P 2 .
4.10. BRIDGES 123
Consider the walk
x i , x i −1 , . . . , x 0 = u = y 0 , y 1 , . . . , y m = v = x n , x n−1 , . . . , x i +1 .
It is a walk from x i to x i +1 that does not use the edge e. So it is also a walk in
G − e, hence x i , x i +1 are in the same component in G − e. By Lemma 4.10.2, e is
not a bridge. By Theorem 4.10.3, e must be part of a cycle. Hence G contains a
cycle.
We will often use the contrapositive form of this result: If graph G has no
cycles, then each pair of vertices is joined by at most one path.
Problem Set 4.10

1. Prove that the prime graph B n defined in Problem Set 4.4 is connected for
every n. You may use without proof the following fact: For every integer
k ≥ 2 there is a prime number r such that k < r < 2k.
2. Prove that, if G is connected, any two longest paths have a vertex in com-
mon.
3. Which graphs, with at least one edge, have the property that every edge is a
bridge?
4. If every vertex of a graph H with p vertices has degree at least p/5, prove that
H cannot have more than 4 components.
5. If edge e is not a bridge of a connected graph G, prove that e is an edge of

some cycle.
6. Prove that a 4-regular graph has no bridge.
7. Let A n be the graph whose vertices are the {0, 1}-strings of length n, and
edges are between strings that differ in exactly two positions, n ≥ 2.
(a) How many edges does A n have?

(b) Is A n bipartite for any n ≥ 2?
(c) How many components does A n have?
8. Let B n be the graph whose vertices are the {0, 1}-strings of length n, and
edges are between strings that differ in exactly two consecutive positions,
n ≥ 2.
(a) How many edges does B n have?

(b) How many components does B n have?
9. Let G be a graph in which exactly two of the vertices u, v have odd degree.
Prove that G contains a path from u to v.
4.11 Certifying Properties

In graph theory we usually find that each time we meet a new property, two
questions arise. First, how do we certify that a graph has the property. Second,
how do we certify that a graph does not have that property.
So we can certify that a graph is connected by providing, for each pair of
distinct vertices, a path in the graph the joins the two vertices. We can certify
that a graph G is not connected by producing a cut—two non-empty sets of
vertices A and B that partition V (G), such that no edge of G joins a vertex in A
to a vertex in B .
As another example, you can certify that an edge e = uv in a connected
graph G is a bridge by producing a cut (A, B ) for G − e where u ∈ A and v ∈ B .
You can certify that e is not a bridge by producing a cycle that contains e. (You
could also certify this by showing that G − e is connected, but this would be
more work as a rule.)
Certificates are required to be easy to verify; they need not be easy to find.
So you might have to work hard to find a certificate, but it must still be easy for
the marker to verify. (The precise rule is that it must be possible to check the
correctness of your certificate in polynomial time.)
It is an experimental fact that if there is a good certificate for verifying that
a graph has some property, and a good certificate for verifying that it does not,
then there is an efficient algorithm for testing if a graph has the property.
Chapter 5
Trees
5.1 Trees
A very special and important kind of graph is a tree.
Definition 5.1.1. A tree is a connected graph with no cycles.
Figure 5.1 shows a typical tree.
Figure 5.1: A tree
Lemma 5.1.2. There is a unique path between every pair of vertices u and v in
a tree T .
125
126 CHAPTER 5. TREES
Proof: For any 2 vertices u and v in T , there is at least 1 path joining them since
T is connected. Since T has no cycles, there is at most one path by Corollary
4.10.4. This establishes the result.
Lemma 5.1.3. Every edge of a tree T is a bridge.

Proof: An edge e of T is not in a cycle, so, by Theorem 4.10.3, e is a bridge.
Theorem 5.1.4. A tree with at least two vertices has at least two vertices of de-
gree one.
Proof: Find a longest path w 0 w 1 w 2 . . . w n in the tree T , say between vertices
u = w 0 and v = w n . This path has length at least one, since any edge in the
non-trivial tree is a path of length one, so u 6= v.
Thus one vertex adjacent to v is in the path, namely w n−1 . If deg(v) > 1
there must be another vertex, w, adjacent to v. Vertex w cannot be in the path,
since this would imply a cycle in T , whereas T has no cycles. Since w is not
in the path, we can extend the path by adding edge {v, w} to it, and we get a
longer path w 0 w 1 . . . w n w. This is a contradiction. Hence d eg (v) = 1. Similarly
d eg (u) = 1, which proves the theorem.
This proof works if, instead of choosing a longest path, we choose a path
which is not a subgraph of a path in T with more edges. (We might say that
our path is “maximal under inclusion”.) The advantage of this choice is that it
is easier to decide if a path is maximal under inclusion than to decide if it is a
longest path—for to do the latter we must consider all paths in the tree.
One main property of trees is that any tree with p vertices have the same
number of edges: p − 1.
Theorem 5.1.5. If T is a tree, then |E (T )| = |V (T )| − 1.
Proof: The proof is by mathematical induction on p, the number of vertices.
When p = 1 there is just one tree. It has one vertex and no edges. It is called the
“trivial" tree, and the theorem holds for it.
Suppose that the theorem is true for trees on fewer than p vertices, and let T
be any tree with p vertices, for p ≥ 2. By the previous theorem, there is a vertex
u in T with degree one. Let v be the neighbor of u in T and let e be the edge
{u, v}. Since e is a bridge, T \ e is not connected and, by Lemma 4.10.2, it has
exactly two components, one of which is the vertex u. The other component is
a connected graph on p − 1 vertices with no cycles. Therefore it is a tree and, by
induction, it has exactly p − 2 edges. This implies that |E (T )| = p − 1.
5.1. TREES 127
In Theorem 5.1.4, we proved that every tree has at least two vertices of de-
gree one. The following alternative proof gives more detailed information about
how many vertices of degree one a tree can have given the degrees of other ver-
tices.
Alternate proof of Theorem 5.1.4: Let T be a tree and let n r denote the number
of vertices of degree r in T . Set p = |V (T )| and assume p ≥ 2. By Theorem 4.3.1
we have X
2p − 2 = deg(v)
v∈V (T )
and therefore
X n−1
X
−2 = (deg(v) − 2) = n r (r − 2).
v r =0
In the last sum, n 0 = 0 (because trees are connected) and so we find that
X
−2 = −n 1 + (r − 2)n r .
r ≥3
Therefore X
n1 = 2 + (r − 2)n r .
r ≥3
Since (r − 2)n r ≥ 0 when r ≥ 3, it follows that n 1 ≥ 2.

The vertices of degree 1 in a tree are often called “end-vertices” or some-
times “leaves”. The above proof implies that if T contains a vertex of degree r ,
where r ≥ 3, then n 1 ≥ 2 + (r − 2) = r . As an exercise, use a version of the first
proof to show that a tree that contains a vertex of degree r has at least r vertices
of degree one.
Problem Set 5.1

1. (a) Draw one tree from each isomorphism class of trees on six or fewer
vertices.
(b) For each tree in (a), exhibit a bipartition (X , Y ) by coloring the vertices
in X with one colour and the vertices in Y with another.
2. Prove that every tree is bipartite.
3. What is the smallest number of vertices of degree 1 in a tree with 3 vertices of

degree 4 and 2 vertices of degree 5? Justify your answer by proving that this
is the smallest possible number, and by giving a tree which has this many
vertices of degree 1.
4. Find the smallest number r of vertices in a tree having two vertices of degree
3, one vertex of degree 4, and two vertices of degree 6. Justify your answer by
proving that any such tree has at least r vertices, and by giving an example
of such a tree with exactly r vertices.
5. A cubic tree is a tree whose vertices have degree either 3 or 1. Prove that a
cubic tree with exactly k vertices of degree 1 has 2(k − 1) vertices.
6. A forest is a graph with no cycles. Prove that a forest with p vertices and q
edges has p − q components.
7. Let p ≥ 2. Show that a sequence (d 1 , d 2 , . . . , d p ) of positive integers is the

Pp
degree sequence of a tree on p vertices if and only if i =1 d i = 2p − 2. (Hint:
use induction on p.)
5.2 Spanning Trees

A spanning subgraph which is also a tree is called a spanning tree. The reason
that spanning trees are important is that, of all the spanning subgraphs, they
have the fewest edges while remaining connected. Figure 5.2 shows a graph
with a spanning tree indicated by heavy lines.
Theorem 5.2.1. A graph G is connected if and only if it has a spanning tree.
Proof: (“if” part.) We are given that G has a spanning tree T . Then Lemma 5.1.2
implies that there is a path in T between every pair of vertices of T . But each
of these paths is also contained in G, and G has the same vertices as T , so from
Definition 4.8.1 we conclude that G is connected.
(“only if” part.) We are given that G is connected. If G has no cycles, then
G itself is a spanning tree of G. Otherwise G has a cycle. Remove any edge e of
some cycle. Then G − e is connected, by Theorem 4.10.3, and has fewer cycles
than G.
Repeat this process, removing an edge on a cycle at each stage, until we have
a connected, spanning subgraph with no cycles. This subgraph is a spanning
tree of G.
5.2. SPANNING TREES 129
5 6
1 4
2 3
8
7 9
10
Figure 5.2: A spanning tree
To show that a graph is connected, using the definition, you need to give
a path between any pair of vertices. However, Theorem 5.2.1 provides a much
more succinct method: give a spanning tree.
Corollary 5.2.2. If G is connected, with p vertices and q = p −1 edges, then G is

a tree.
Proof: Let G be a connected graph with p vertices and q = p − 1 edges. By

Theorem 5.2.1, G has a spanning tree T . Now T is a tree with p vertices, so, by
Theorem 5.1.5, T has p − 1 edges. However, as G has only p − 1 edges, it must
be the case that G = T . Therefore, G is a tree.
There are a couple of ways to get different spanning trees by exchanging two
edges.
Theorem 5.2.3. If T is a spanning tree of G and e is an edge not in T , then T + e

contains exactly one cycle C . Moreover, if e 0 is any edge on C , then T + e − e 0 is
also a spanning tree of G.
Proof: Let e = {u, v}. Any cycle in T +e must use e, since T has no cycles. Such a
cycle consists of e along with a u, v-path in T . By Lemma 5.1.2, there is a unique
u, v-path in T , hence there is exactly one cycle C in T + e.
If e 0 is any edge in C , then e 0 is not a bridge (Theorem 4.10.3). So T + e − e 0 is
still connected. Since it has n − 1 edges, by Corollary 5.2.2, it is a tree.
Theorem 5.2.4. If T is a spanning tree of G and e is an edge in T , then T − e

has 2 components. If e 0 is in the cut induced by one of the components, then
T − e + e 0 is also a spanning tree of G.
Proof: The first statement is a direct consequence of Lemma 4.10.2. Let C 1 and
C 2 be the two components of T − e. Suppose e 0 = {u, v} where u ∈ V (C 1 ) and
v ∈ V (C 2 ).
We wish to show that T − e + e 0 is connected using Theorem 4.8.2. Let x ∈
V (C 1 ). For any y ∈ V (C 1 ), there exists an x, y-path since C 1 is connected. Sup-
pose y ∈ V (C 2 ). Since C 1 and C 2 are connected, there exist an x, u-path P 1 and
a v, y-path P 2 . Then P 1 , e, P 2 form an x, y-path. Since there exists an x, y-path
for any vertex y, T − e + e 0 is connected.
Since T − e + e 0 has n − 1 edges, by Corollary 5.2.2, it is a tree.
5.3 Characterizing Bipartite Graphs

Using Theorem 5.2.1 we will prove an important characterization of bipartite
graphs. Note that, we can convince someone that a graph is bipartite by giv-
ing them a bipartition. However, it is not so clear how you would convince
them that a graph does not have a bipartition. (Checking every partition of
V (G) would be tedious.)
One idea is to note that any subgraph of a bipartite graph is bipartite, and so
we could try to find a subgraph that is “obviously” not bipartite. An odd cycle is
a cycle on an odd number of vertices.
Lemma 5.3.1. An odd cycle is not bipartite.
Proof: Suppose that C is a cycle with vertex set
{−k, −k + 1, . . . , k}
where i ∼ i + 1 if −k ≤ i < k and k ∼ −k. Suppose C is bipartite with bipartition

(A, B ). Without loss of generality, 0 ∈ A. Then 1, −1 ∈ B and it follows easily
that for j = 1, . . . , k, the vertices j and − j must be in the same partition. But k
and −k are adjacent and they are in the same partition. This is a contradiction,
hence C is not bipartite.
So you could certify a graph is not bipartite by producing a subgraph that is
an odd cycle. It is quite surprising, but this is all you need to do.
The converse is also true, as we see in the following result.
5.3. CHARACTERIZING BIPARTITE GRAPHS 131
Theorem 5.3.2. A graph is bipartite if and only if it has no odd cycles.
Proof: Given Lemma 5.3, it suffices to prove that if G is not bipartite, then it
contains an odd cycle.
Since G is not bipartite, at least one component H of G is not bipartite. (If
all components are bipartite, then we could find a bipartition of G by combin-
ing the bipartitions of the individual components.) Since H is connected, by
Theorem 5.2.1, there exists a spanning tree T in H .
Trees are bipartite (see Problem Set 5.1, Problem 2), so let (A, B ) be a bipar-
tition of T . Since H is not bipartite, (A, B ) is not a bipartition of H and therefore
there exists an edge {u, v} of H such that both u and v are in A, or both are in B .
By swapping A and B if needed, we may assume that u, v ∈ A.
Since T is connected, there exists a u, v-path P in T , with vertices x 0 x 1 . . . x n
where u = x 0 and v = x n . Since x 0 = u ∈ A and T is bipartite, the vertices along P
must alternate between A and B . So x 0 , x 2 , x 4 , . . . ∈ A and x 1 , x 3 , x 5 , . . . ∈ B . Since
x n ∈ A, n must be even, hence P has even length. However, x 0 x n = uv ∈ E (H ),
so P + {u, v} is an odd cycle in H , which is in G. Hence G contains an odd cycle,
as claimed.
Problem Set 5.3

1. Let r be a fixed vertex of a tree T . For each vertex v of T , let d (v) be the
length of the path from v to r in T . Prove that
(a) for each edge uv of T , d (u) 6= d (v), and

(b) for each vertex x of T other than r , there exists a unique vertex y such
that y is adjacent to x and d (y) < d (x).
2. Let r be a fixed vertex in a graph G. Suppose that, for each vertex v of G we

have an integer d (v) such that
(i) for each edge uv of G, d (u) 6= d (v), and

(ii) for each vertex x of G other than r , there exists a unique vertex y such
that y is adjacent to x and d (y) < d (x).
Prove that G is a tree.

5.4 Breadth-First Search

We now consider an algorithm for finding a spanning tree in a graph G, if one
exists. Note that, by Theorem 5.2.1, this is also a very good way of deciding
whether a graph is connected. To describe the algorithm properly we must
decide how our graph will be presented to us, and also in what form we will
present its output. We assume that the graph is given as a list of edges.
We could present the spanning tree (if it exists) as a list of edges but, given a
list of edges it is not obvious if it is a tree, and so we take a more sophisticated
approach. We will represent a tree by a function.
Suppose T is a tree and u is a vertex in T . For each vertex x in T distinct
from u there is a unique path of length at least one from x to u. Define pr(x) to
be the neighbour of x in this path. We define pr(u) to be ;. The pr is a function
from V (T ) to V (T ). We might call pr(x) the parent or predecessor of x.
Clearly if we are given a tree it is easy to write down its predecessor function
and, conversely, given the predecessor function we can recover the tree.
So suppose we are given a graph G and we want to find a spanning tree in G.
Choose a vertex u in G and set D = {u}. Define pr(u) = ;.
Now suppose we are given a subset D and a partially completed predecessor
function. If D 6= V (G), look for an edge that joins a vertex in D to a vertex not
in D. If there is none, then D determines an empty cut and we deduce that G
is not connected. If there is an edge v w where v ∈ D and w ∉ D, then add w to
D and define pr(w) = v. If D = V (G), then our predecessor function describes a
spanning tree in G.
Algorithm 5.4.1. To find a spanning tree of a graph G: Select any vertex r of G

as the initial subgraph D, with pr(r ) = ;. At each stage, find an edge in G that
joins a vertex u of D to a vertex v not in D. Add vertex v and edge {u, v} to D,
with pr(v) = u. Stop when there is no such edge.
Claim:
If |V (D)| = |V (G)| when the algorithm terminates, then D is a spanning tree
of G. If |(D)| < |V (G)| when the algorithm terminates, then G is not connected
and so, from Theorem 5.2.1, G has no spanning tree.
Proof:
We begin by using mathematical induction on the number of iterations to
show that the subgraphs D produced by the algorithm are subtrees of graph G.
Basis Case: Initially, D is a tree with 1 vertex.
5.4. BREADTH-FIRST SEARCH 133
Induction Hypothesis: For k ≥ 0, assume that the subgraph D produced in

the k-th iteration is a tree with k + 1 vertices and k edges.
Inductive Step: If the algorithm terminates in the (k + 1)st iteration, then
there is nothing to prove. Otherwise, the (k +1)st iteration produces a subgraph
E by adding vertex v and edge {u, v} to D. Therefore E has k +2 vertices and k +1
edges. Since D is tree, there is a path from u to every vertex in D. Since u, v is
a path from u to v, there is a path from u to every vertex in E . Therefore, E is
connected by Theorem 4.8.2. By Corollary 5.2.2, E is a tree. The iteration ends
by redefining D to be the subtree E . By Mathematical Induction, D is a subtree
of G in every iteration.
If D has p vertices when the algorithm terminates, then D is a spanning tree
because it includes all the vertices of G, as required.
If D has fewer that p vertices when the algorithm terminates, then V (D) is a
nonempty, proper subset of V (G). Since the algorithm has terminated, no edge
joins a vertex u in D to a vertex v not in D. Therefore, the cut induced by V (D)
is empty. This implies that G is not connected by Theorem 4.8.5, as required.
Spanning trees have the important property of being connected. If u and v
are vertices of a connected graph, then in any spanning tree there is a unique
path joining u and v. Hence spanning trees provide a nice “small" structure
by which one can search all the vertices in a graph using only edges of the tree.
This is the reason for the parent function pr(v) = u, in which vertex v is referred
to as a child of u. In a diagram we display this information by placing an arrow
on the edge between u and v, pointing from the child v to the parent u, as in
Figure 5.3. The initial vertex r is called the root vertex, and has no parent. One
can now easily recover paths in the tree. For vertex v 6= r , there is a unique
positive integer k such that prk (v) = r , and the path from v to r is
v, pr(v), pr2 (v), . . . , prk (v) = r
A path between two vertices u, v can be recovered from Corollary 4.6.3, using
the paths from u and v to the root r . Alternatively, we examine parents from
u and v until we find a common “ancestor". A spanning tree with this extra
structure, provided by the parent functions (the pointers, or the arrows on the
diagram) is often called a search tree. If prk (v) = r we say that the level of v is
k, and write level (v) = k. We define level (r ) to be 0.
In Figure 5.3, r is at level 0, u and y are at level 1, while v, w and x are at level
two.
In Algorithm 5.4.1, we looked for edges incident with a vertex in the tree and
a vertex not in the tree. Thus it is convenient at each stage to say that a vertex in
r
pr(r) = φ
pr(u) = r
pr(y) = r
u y
pr(v) = u
pr(w) = u
pr(x) = u
v w x
Figure 5.3: A search tree
the tree is exhausted if it is not adjacent to a vertex outside the tree. Of course,
if a vertex is exhausted at any stage, then it will remain exhausted at all later
stages. In Algorithm 5.4.1, we can ignore edges incident with exhausted vertices,
since the only possible edges that will allow the tree to increase in size must be
incident with an unexhausted vertex in the tree. Now we consider a refinement
of Algorithm 5.4.1 called breadth-first search, in which the unexhausted vertex
u at each stage is chosen in a special way.
Algorithm 5.4.2. Breadth-first search. Follow Algorithm 5.4.1 with the follow-
ing refinement: At each stage consider the unexhausted vertex u that joined
the tree earliest among all unexhausted vertices (called the active vertex), and
choose an edge incident with this vertex and a vertex v not in the tree.
A breadth-first search tree is any spanning tree that is created by applying

breadth-first search to a connected graph.
One way to implement breadth-first search is to use a “queue", that is, a
first-in, first-out list for vertices in the tree, where vertices are placed at the end
of the queue when they are first added to the tree. At each stage the vertex at the
head of the queue is examined. If it is exhausted, it is removed from the queue.
Otherwise, it is the active vertex.
In Figure 5.5, we illustrate an example of breadth-first search, applied to
the connected graph G. We begin the breadth-first search tree B by (arbitrarily)
selecting vertex 9 as the root, with pr(9) = ;, and thus the queue then consists
of vertex 9 alone. Vertex 9 becomes active, and we add vertices 3, 8, 10 to B , at
level 1, in that (arbitrary) order, with pr(3) = pr(8) = pr(10) = 9, and the queue
becomes 9, 3, 8, 10 (ordered with the head of the queue on the left, and each
1 2 3 4
5 6 G
7 8 9 10
Figure 5.4: Graph G
3 8 10
B
4 2 7 5 6
Figure 5.5: A re-drawing of G, with the edges of a breadth-first search tree B

indicated with arrows
new vertex joining the queue on the right). Now vertex 9 is exhausted, so it is
removed from the queue, and vertex 3 becomes active. Next, vertices 4 and 2
are added to B , with pr(4) = pr(2) = 3, and so vertices 4 and 2 are at level 2, and
the queue becomes 3, 8, 10, 4, 2. Vertex 3 is now exhausted, so it is removed
from the queue, vertex 8 becomes active, and we add vertices 7, 5, 6 to B in
that order, so the queue becomes 8, 10, 4, 2, 7, 5, 6. Now vertices 8, 10, 4 are
all exhausted, so the queue becomes 2, 7, 5, 6, and vertex 1 is added to B , at
level 3. Finally, all vertices of G are now in B , so all vertices in the queue are
exhausted, and we stop. In Figure 5.5, we have placed the vertices at each level
from left to right, in the order that they joined the breadth-first search tree B .
As a summary, note that the order in which the vertices of G joined B is 9, 3, 8,
10, 4, 2, 7, 5, 6, 1.
In Figure 5.5, the edges of B have an arrow to specify the parent. The edges
of G that are not contained in the breadth-first search tree B (called non-tree
edges) have been added to the drawing of B , without arrows. This gives a re-
drawing of G that will illustrate the primary property of breadth-first search,
given below. First we need a preliminary lemma.
Lemma 5.4.3. The vertices enter a breadth-first search tree in non-decreasing

order of level.
Proof: We prove this by induction on the number of vertices in the tree at each
stage of the algorithm. The first vertex in the tree is the root vertex, with level 0,
and the result is true for this first stage.
Now we make the induction hypothesis, that the result is true for the first
m vertices in the tree, m ≥ 1, and consider the next vertex v, that joins the tree
at stage m + 1. Now pr(v) = u, where u is active when v joins the tree, and
level(v) = level(u) + 1. Consider any other non-root vertex x in the tree at stage
m + 1. Then pr(x) = y, and level(x) = level(y) + 1. But either y = u or y is ac-
tive before u. In the latter case, y joined the tree before u, so by the induction
hypothesis, in either case we have level(y) ≤ level(u). Thus we have
level(v) = level(u) + 1 ≥ level(y) + 1 = level(x),
so level(v) ≥ level(x) for all other vertices x in the tree at stage m + 1, so the
result is true at stage m + 1.
Hence, the result is true by mathematical induction.
For example, it is easy to check that Lemma 5.4.3 holds for the breadth-first
search tree B given in Figure 5.5.
This result allows us to establish the following important fact about breadth-
first search.
Theorem 5.4.4. (The primary property of breadth-first search.)

In a connected graph with a breadth-first search tree, each non-tree edge in the
graph joins vertices that are at most one level apart in the search tree (of course
each tree edge joins vertices that are exactly one level apart).
Proof: Suppose that vertices u and v are joined by an edge, and without loss of
generality, that u joins the tree before v. Thus u is active before v, and there are
two cases:
Case 1. v is in the tree when u first is active. Then u and v are joined by a non-
tree edge, and pr(v) = w, where w joined the tree before u. Thus level(w) ≤
level(u), by Lemma 5.4.3, so we have
level(v) = level(w) + 1 ≤ level(u) + 1.
But also level(v) ≥ level(u), by Lemma 5.4.3, so we conclude in this case that u
and v are joined by a non-tree edge with
level(u) ≤ level(v) ≤ level(u) + 1,
and the result is true in this case.

Case 2. v is not in the tree when u first is active. Then v will be added to the
tree, with pr(v) = u and level (v) = level (u) + 1, and the edge between u and v
is a tree edge.
For example, it is easy to check that Theorem 5.4.4 holds for graph G and
breadth-first search tree B in Figure 5.5. Here, non-tree edges {3, 10}, {5, 6} join
vertices at the same level, and {4, 8}, {1, 7} join vertices one level apart.
Problem Set 5.4

1. Prove that if Algorithm 5.4.1 terminates with subgraph D containing fewer
than p vertices, then D is a spanning tree for the component of G containing
the initially chosen vertex.
2. Consider a graph G with V (G) = {3, 4, 5, . . . , 25} and with {p, q} ∈ E (G) if and
only if either p|q or q|p. (Definition: p|q if and only if q = pr for some
integer r .) Give a spanning forest of G with the largest number of edges and
hence determine the number of components of G.
3. Construct a breadth-first search tree for the graph below, taking vertex la-
belled 1 as root. When considering the vertices adjacent to the vertex being
examined, take them in increasing order of their labels.
1 2 5
13 14 6 4
11 10 7 3
12 9 8
4. What graphs have the property that, for a suitable choice of root, the breadth-
first search algorithm yields a tree in which all vertices (except the root) are
at level 1?
5. Explain why the search tree, rooted at d and indicated by dotted lines in the
graph below is not a breadth-first search tree.
a b c
d e f g
h i j
6. Prove that if a non-tree edge joins vertices u and v in adjacent levels, say
level (v) = level (u)+1, then the parent of v is at the same level as u, and was
active before u.
5.5. APPLICATIONS OF BREADTH-FIRST SEARCH 139
5.5 Applications of Breadth-First Search

Theorem 5.3.2 provides a nice characterization of bipartite graphs. Moreover,
the ideas in Section 5.3 do provide an algorithm to find a bipartition or to find
an odd cycle. However, breadth-first search provides a compact and efficient
algorithm, as given in the proof of the following result. We refer to this proof
as constructive because it allows us to construct directly either an odd cycle or
a bipartition. Of course, for a graph that is not connected, we would carry this
out by applying the construction for each component (and an odd cycle in any
component would demonstrate that the graph is not bipartite).
Theorem 5.5.1. A connected graph G with breadth-first search tree T has an

odd cycle if and only if it has a non-tree edge joining vertices at the same level
in T .
Proof: (i) Suppose G has a non-tree edge joining distinct vertices u, v at the
same level in T . Then the paths from u and v to the root vertex first meet at a
vertex (possibly the root vertex) which is m levels less than u and v, for some
m ≥ 1. Then the path in T between u and v has length 2m, and together with
the non-tree edge {u, v} this gives a cycle of length 2m + 1 in G, so G has an odd
cycle. (ii) If G has no non-tree edge joining vertices at the same level, then G is
bipartite, with bipartition into sets of vertices A = {v ∈ V (G) : level(v) is odd },
B = {v ∈ V (G) : level(v) is even}. All edges in G are incident with vertices whose
levels differ by one; hence one level must be even and one must be odd, so
every edge of G is incident with one vertex in A and one vertex in B . But, from
Lemma 5.3, this means that G has no odd cycles.
To find an odd cycle in a connected graph G, we need not construct the
entire breadth-first search tree and then look for an edge between vertices at
the same level. Instead, we can look for these edges when the tree is being built.
At each stage, when we examine the vertex at the head of the queue to see if it is
exhausted, we check if there is an edge joining it to a vertex already in the tree,
at the same level.
Theorem 5.5.1 illustrates an important application of breadth-first search
dealing with parity of cycle lengths. We now consider an application of breadth-
first search dealing with shortest paths.
Theorem 5.5.2. The length of a shortest path from u to v in a connected graph

G is equal to the level of v in any breadth-first search tree of G with u as the
root.
Proof: From the primary property of breadth-first search trees we know that no
edge of G joins vertices that are more than one level apart. If vertex v is at level
k in a breadth-first search tree rooted at u, then there is a path (in the tree) of
length k from u to v. There can be no shorter path from u to v in G, since such
a path would have to contain an edge joining vertices more than one level apart.
The length of the shortest path between two vertices is often called the dis-
tance between the vertices. Theorem 5.5.2 implies that we can determine the
distance between vertex u and any other vertex in a connected graph G by find-
ing a breadth-first search tree of G rooted at u.
Problem Set 5.5

1. Describe a general procedure to determine a maximal bipartite subgraph of
any connected graph G. (A proof that your procedure works is not neces-
sary.)
2. The diameter of a graph is the largest of the distances between the pairs of
vertices in the graph. Let G be a connected graph of diameter three with
exactly 20 vertices at distance three from a given vertex v. Prove that G has
some spanning tree T with exactly 20 vertices of degree one at level three.
3. (a) Describe an algorithm to determine the diameter of a graph.

(b) Prove that this algorithm works.
4. Let m, n be integers with m, n ≥ 1. Let G be a graph (connected) with m

vertices at distance n from a given vertex v. Prove that G has a spanning tree
with at least m vertices of degree 1.
5. Suppose that a connected graph G has a breadth-first search tree T for which
every non-tree edge joins vertices at equal levels. Prove that every cycle of G
contains an even number of tree edges.
5.6 Minimum Spanning Tree

In the minimum spanning tree (MST) problem, we are given a connected graph
G and a weight function on the edges w : E (G) → R. The goal is to find a span-
ning tree in G whose total edge weight is minimized.
5.6. MINIMUM SPANNING TREE 141
For example, in the graph to the left, we have a connected graph with the
edges being labelled with their weights. In the graph to the right, we have a
spanning tree of total weight 31. But is there another spanning tree of smaller
weight?
a 9 a 9
b b
4 8 4 8
5 8 7 5 8 7
c d c d
8 3 8 3
e f e f
7 7
Edge-weighted connected graph G A spanning tree of weight 31.
This problem is useful in the design of various networks, such as computer

networks, road networks, and power grids. The edge weights would represent
the costs of building between the two locations. A minimum spanning tree
would represent the least amount of cost required to completely connect every
location.
There are several efficient algorithms that can solve the MST problem. We
present Prim’s algorithm here. The idea of Prim’s algorithm is that we start with
a vertex, and iteratively grow the tree one edge at a time. Each time we grow the
tree, we increase the total weight as small as possible.
Prim’s algorithm:
1. Let v be an arbitrary vertex in G, let T be the tree consists of just v.
2. While T is not a tree...
(a) Look at all the edges in the cut induced by V (T ).

(b) Let e = uv be an edge with the smallest weight in the cut (where
u ∈ V (T ), v 6∈ V (T )).
(c) Add v to V (T ) and add e to E (T ).
Here is an illustration of the algorithm using the example above. We start

with the vertex a. In the first iteration, we look at the edges in the cut induced
by {a}, which are {ab, ac, ae}. Since ac has the smallest weight, we add it to the
tree along with vertex c.
In the second iteration, we look at the edges in the cut induced by {a, c},
which are {ab, cd , ce, ae}. Since ae has the smallest weight, we add it to the tree
along with the vertex e.
We repeat this process until we have a spanning tree. We claim that the tree
we produce is a minimum spanning tree.
a 9 a 9
b b
4 8 4 8
5 8 7 5 8 7
c d c d
8 3 8 3
e f e f
7 7
a 9 a 9
b b
4 8 4 8
5 8 7 5 8 7
c d c d
8 3 8 3
e f e f
7 7
a 9 a 9
b b
4 8 4 8
5 8 7 5 8 7
c d c d
8 3 8 3
e f e f
7 7
Prim’s algorithm is a greedy algorithm, meaning at each step in the process,

we pick the edge that is “best” for our problem. But how do we know that it will
always produce a tree of minimum weight? This requires the following proof.
P
For any graph H , we will use the notation w(H ) = e∈E (H ) w(e).
5.6. MINIMUM SPANNING TREE 143
Theorem 5.6.1. Prim’s algorithm produces a minimum spanning tree for G.
Proof: Let T1 , T2 , . . . , Tn be the trees produced by the algorithm at each step,

where the order of selection of the edges is e 1 , e 2 , . . . , e n−1 (i.e. you get Ti +1 by
adding e i to Ti ). We will prove by induction on k that there exists a MST con-
taining Tk as a subgraph.
Base case: For k = 1, it is just a vertex, so every MST contains T1 as a sub-
graph.
Induction hypothesis: Assume that there exists a MST containing Tk as a
subgraph.
Induction step: We need to prove that there is a MST containing Tk+1 as a
subgraph.
Let T ∗ be a MST that contains Tk as a subgraph, which is assumed by the in-
duction hypothesis. If T ∗ also contains e k , then T ∗ contains Tk+1 as a subgraph,
and we are done. If not, then e k 6∈ E (T ∗ ). This means that T ∗ + e k contains a
unique cycle C . Now C − e k is a path between the two endpoints of e k , one of
which is in V (Tk ) and the other is not. Therefore, there is at least one edge e 0 in
C − e k in the cut induced by V (Tk ). By Theorem 5.2.3, T 0 = T ∗ + e k − e 0 is also a
spanning tree.
In Prim’s algorithm, when we picked e k , it is an edge in the cut induced by
V (Tk ) of minimum weight. So w(e 0 ) ≥ w(e k ). If w(e 0 ) > w(e k ), then w(T 0 ) =
w(T ∗ ) + w(e k ) − w(e 0 ) < w(T ∗ ), which is not possible since T ∗ is a minimum
spanning tree. Therefore, w(e 0 ) = w(e k ), which means that w(T 0 ) = w(T ∗ ). So
T 0 is also a minimum spanning tree, which contains all edges in Tk+1 . Therefore,
T 0 is the tree we are looking for.
This induction tells us that there is a MST that contains Tn , meaning Tn
must equal to the MST. Hence the algorithm produces a MST.
Chapter 6
Codes
6.1 Vector Spaces and Fundamental Cycles

Let S(G) be the set of all spanning subgraphs of the graph G. Let the element of
S(G) with no edges be denoted by Z .
The field GF (2) of integers modulo 2 contains two elements, 0 and 1, whose
elements are added and multiplied modulo 2 (e.g. 0⊕0 = 1⊕1 = 0, 1⊕0 = 0⊕1 =
1, 0 · 0 = 0 · 1 = 1 · 0 = 0, 1 · 1 = 1). For any H ∈ S(G), we define multiplication of
H by an element of GF (2) as follows:
0· H = Z , 1· H = H.
Addition of elements of S(G) is also defined.
Definition 6.1.1. For H1 , H2 ∈ S(G) the modulo 2 sum of H1 and H2 , denoted

by H1 ⊕ H2 , is the element of S(G) whose edge set consists of all the edges of G
that are in H1 or H2 but not in both (the “symmetric difference" of E (H1 ) and
E (H2 )).
The set of all spanning subgraphs of G forms a vector space under the oper-
ation of mod 2 sum.
Theorem 6.1.2. The set S(G) is a vector space over GF (2).
Proof: Look up the definition of a vector space in your linear algebra text. You
will find that the scalar multiplication and vector addition defined above must
satisfy a list of axioms. These are all easy to verify.
145
146 CHAPTER 6. CODES
For example, for any H ∈ S(G),
Z ⊕ H = H ⊕ Z = H,
so Z is the “zero" element in S(G). Also
H ⊕H = Z,
so each element of S(G) has an additive inverse, namely itself. We have defined
0·H = Z,
and have, for instance,
1 · H ⊕ 1 · H = H ⊕ H = Z = 0 · H = (1 ⊕ 1) · H .
Proving that all the axioms of a vector space are satisfied by S(G) is left as an
exercise.
Suppose that the edges of G are e 1 , e 2 , . . . , e q , and let A i be the element of
S(G) that contains edge e i , and no other edges, for i = 1, 2, . . . , q.
Theorem 6.1.3. {A 1 , A 2 , . . . A q } forms a basis for S(G).
Proof: The graphs A 1 , A 2 , . . . A q are all contained in S(G). To prove that they
form a basis for S(G), it is sufficient to prove that they (i) span S(G) and (ii) form
a linearly independent set.
(i) The elements of S(G) are uniquely specified by their edges. For H ∈ S(G),
if H contains edges e i 1 , e i 2 , . . . , e i k , then H can be written as the linear com-
bination
H = α1 A 1 ⊕ · · · ⊕ αq A q
where αi 1 = αi 2 = · · · = αi k = 1, and all other α’s are equal to 0. Thus
{A 1 , A 2 , . . . , A q } spans S(G).
(ii) Suppose β1 , . . . , βq are such that
β1 A 1 ⊕ · · · ⊕ β q A q = Z .
If β j = 1 for some j , then the graph on the LHS contains edge e j . But the
graph on the RHS contains no edges, so for equality to hold, we must have
β j = 0 for all j = 1, . . . , q. Thus {A 1 , A 2 , . . . , A q } is a linearly independent
set.
6.1. VECTOR SPACES AND FUNDAMENTAL CYCLES 147
This means that S(G) is a vector space of dimension q and contains 2q ele-
ments. Now we look at a subspace of S(G).
Definition 6.1.4. A graph in which all degrees are even non-negative integers is
called an even graph.
Let C (G) be the set of even spanning subgraphs of a graph G, so C (G) is a
subset of S(G).
Theorem 6.1.5. The set C (G) forms a vector space over GF (2). (This is a sub-
space of S(G).)
Proof: Clearly Z ∈ C (G), so C (G) is not empty. Thus, to prove that C (G) is a
subspace of S(G), it is sufficient to prove that (i) C (G) is closed under scalar
multiplication and (ii) C (G) is closed under addition.
(i) Suppose H ∈ C (G). Then 1 · H = H ∈ C (G), and 0 · H = Z ∈ C (G), since Z
has all vertex degrees equal to zero, which is an even non-negative inte-
ger.
(ii) Suppose H1 , H2 ∈ C (G), and let d 1 (v) and d 2 (v) denote the degree of v in
H1 and H2 , respectively, for all v ∈ V (G). Let m(v) be the number of edges
incident with v that are contained in both H1 and H2 . Then the degree of
v in H1 ⊕ H2 is
(d 1 (v) − m(v)) + (d 2 (v) − m(v)) = d 1 (v) + d 2 (v) − 2m(v),
which is even, since both d 1 (v) and d 2 (v) are even, and non-negative,
since d 1 (v) ≥ m(v) and d 2 (v) ≥ m(v). Thus H1 ⊕ H2 ∈ C (G).
We have proved that C (G) is a subspace of S(G), so C (G) is a vector space

itself.
In order to find a basis for C (G), we consider a result about trees.
Theorem 6.1.6. Let T be a spanning tree of a connected graph G. If e is an edge
of G that is not in T , then T + e contains a unique cycle. (This cycle contains
edge e.)
Proof: Suppose e = {u, v}. Since T contains no cycles, any cycle in T + e must
contain edge e, so it must consist of e together with a path from u to v in T .
But from the Lemma 5.1.2 there is a unique such path, and so T + e contains a
unique cycle, which passes through e.
Edges e as described in the above result are called non-tree edges. If G has
q edges and p vertices, then T has p − 1 edges, so G has q − (p − 1) = q − p + 1
non-tree edges. Suppose that the non-tree edges are e 1 , e 2 , . . . , e q−p+1 . Let C i
be the spanning subgraph of G whose edges are the edges of the unique cycle
in T + e i , for i = 1, . . . , q − p + 1. Then C i is called a fundamental cycle, and
{C 1 ,C 2 , . . . ,C q−p+1 } is the set of fundamental cycles of G determined by T .
Lemma 6.1.7. For a fixed spanning tree T of a connected graph G, no two ele-
ments of C (G) contain exactly the same set of non-tree edges.
Proof: Consider H ∈ C (G) with no non-tree edges. Then H is a spanning sub-

graph of T , and is thus a spanning forest of G. If any component of H has more
than one vertex, then there are at least two vertices of degree 1 in that compo-
nent, by Theorem 5.1.4. But H is an even graph, so it can have no vertices of
degree 1. Thus all components of H are isolated vertices, so H must be equal to
Z . Hence Z is the unique elment in C (G) with no non-tree edges.
Suppose that H1 , H2 ∈ C (G), where H1 and H2 contain exactly the same set
of non-tree edges. Then H1 ⊕ H2 has no non-tree edges. But H1 ⊕ H2 ∈ C (G),
so H1 ⊕ H2 = Z , since Z is the unique element of C (G) with no non-tree edges.
This gives H1 = H2 and the result follows.
Theorem 6.1.8. {C 1 ,C 2 , . . . ,C q−p+1 } forms a basis for C (G), where G is connected.
Proof: We have C i ∈ C (G), for i = 1, . . . q −p +1, since the vertex degrees in C i are
either 2 (for vertices on the cycle in T +e i ) or 0 (for the remaining vertices). Thus,
to prove that {C 1 , . . . ,C q−p+1 } forms a basis for C (G), it is sufficient to prove that
(i) it spans C (G) and (ii) it is a linearly independent set.
(i) From Lemma 6.1.7, the elements of C (G) are uniquely specified by their
non-tree edges. For H ∈ C (G), where H contains non-tree edges e i 1 , e i 2 , . . . , e i m ,
then H can be written as the linear combination
H = α1C 1 ⊕ · · · ⊕ αq−p+1C q−p+1
where αi 1 = αi 2 = · · · = a i m = 1 and all other α’s are equal to 0. Thus

{C 1 , . . . ,C q−p+1 } spans C (G).
(ii) Suppose β1 , . . . , βq−p+1 such that
β1C 1 ⊕ · · · ⊕ βq−p+1C q−p+1 = Z .

If β j = 1 for some j , then the graph on the LHS contains non-tree edge e j ,
since C j is the only fundamental cycle containing non-tree edge e j . But
the graph on the RHS contains no edges, so for equality to hold, we must
have β j = 0 for all j = 1, . . . , q − p + 1. Thus {C 1 , . . . ,C q−p+1 } is a linearly
independent set.
This means that C (G) has dimension q −p +1 and contains 2q−p+1 elements.
The cycles of G (together with some isolated vertices) are elements of C (G) and
thus can be uniquely expressed as a mod 2 sum of fundamental cycles. We call
C (G) the cycle space of G. The number q − p + 1 is known as the cyclomatic
number of G.
Problem Set 6.1

e1 e2
e3 e4 e5 e6
e7 e8
e9 e10 e11 e12
e13 e14
1. (a) Find the fundamental cycles corresponding to the spanning tree T with
edges
e 1 , e 2 , e 5 , e 7 , e 10 , e 11 , e 14
in the above graph. (List their edges.)
(b) Express the spanning even subgraphs with the following edge sets as
modulo 2 sums of fundamental cycles from (a):
(i) E (H1 ) = {e 1 , e 2 , e 4 , e 5 , e 7 , e 8 , e 10 , e 11 , e 13 , e 14 }
(ii) E (H2 ) = {e 3 , e 7 , e 9 , e 11 , e 12 , e 14 }
2. (a) Let T be the spanning tree given by dotted lines in the graph below.
Write down the edges in each of the fundamental cycles determined by
T.
1
e1 e7 e2
6
5 e6 e12 e13 2
8 9 e8
e11 e14
e5 e15 e3
10 7
e10
e9
4 3
e4
(b) Express the cycle whose edges are e 1 , e 7 , e 13 , e 14 , e 15 , e 10 and e 5 , as a

modulo 2 sum of fundamental cycles from (a).
3. In the graph shown below, let T be the spanning tree with edges
e 1 , e 5 , e 6 , e 8 , e 9 , e 10 , e 13 , e 14 , e 15 .
Write down the fundamental cycles determined by T . Also, express the cycle
C , whose edges are
e 2 , e 7 , e 12 , e 11 , e 15 , e 14 , e 8
as a linear combination of fundamental cycles.

e1 e7 e2
e6 e12 e13 e8
e11 e14
e5 e15 e3
e10
e9
e4
4. (a) In the graph below, a spanning tree has been constructed. Find the fun-
damental cycles C a ,C b and C c that are determined by the edges a, b
and c.
13 16
12 17
14 15
10 11 18 19
a
1 9
b 8 7 6
2 3 4 5
(b) Find also the subgraph C a ⊕C b ⊕C c .
5. G is a connected graph with a spanning tree T . For each statement below,

either give a proof or find a counterexample.
(a) If G is bipartite, then every fundamental cycle of T must have even

length.
(b) If every fundamental cycle of T has even length, then every cycle of G
must have even length.
(c) If every fundamental cycle of T has length divisible by 3, then every
cycle of G must have length divisible by 3.
6. Prove that a graph which is connected, with 13 vertices and 18 edges, must
have at least 6 distinct cycles. Can you find such a graph with 6 edge-disjoint
cycles (edge-disjoint means that no edge may belong to more than one of
the cycles)?
7. Prove that an edge belongs to H1 ⊕ H2 ⊕ · · · ⊕ Hk if, and only if, it belongs to

an odd number of the Hi ’s where H1 , H2 , . . . , Hk ∈ S(G).
8. What is the dimension of C (G) if G has m components, for arbitrary m ≥ 1?
9. Let F(G) be the set of spanning subgraphs with an even number of edges of
a graph G. (The spanning subgraph containing no edges belongs to F(G).
Note: F(G) is not the set of even spanning subgraphs of G.)
(a) If F 1 , F 2 ∈ F(G), prove that F 1 ⊕ F 2 ∈ F(G), where F 1 ⊕ F 2 is the modulo 2

sum of F 1 and F 2 .
(b) Part (a) implies that F(G) is a vector space over GF (2). Find a basis for
F(G), and prove that it is a basis. What is the dimension of F(G), in
terms of p and q, the number of vertices and edges in G?
6.2 Graphical Codes

We now explore a nice application of the cycle space of a graph.
Let G = (V, E ), E = {e 1 , . . . , e q } be a fixed connected graph. The characteris-
tic vector (x 1 , . . . , x q ) of subgraph (V, F ) has x i = 1 if e i ∈ F and x i = 0 otherwise
for 1 ≤ i ≤ q. Suppose that x and y are characteristic vectors of two even sub-
graphs of G. How many nonzero elements does x ⊕ y contain? It may contain
none, if x = y. But if x 6= y, then x ⊕ y is an even subgraph which necessarily
contains a cycle; then x ⊕ y must have at least three nonzero entries.
Definition 6.2.1. The Hamming distance between two binary vectors x and y
of the same length is the number of 1’s in x ⊕ y.
6.2. GRAPHICAL CODES 153
We can equally well speak about the Hamming distance between subgraphs,
using their characteristic vectors.
We have essentially proved the following:
Lemma 6.2.2. In the cycle space of a graph G, any two distinct vectors have
Hamming distance at least three.
Now let us see how to use this observation. Suppose that two people want
to communicate over a “channel” (such as a telephone line), capable of send-
ing 0, 1-sequences, but this channel on rare occasions introduces an error by
transmitting a 0 when a 1 was intended, or vice versa.
The two people, say Alice and Bob, can agree on a graph G = (V, E ) and
a labelling E = {e 1 , . . . , e q } of its edges, and further agree only to transmit as
messages the characteristic vectors of even subgraphs of G. Suppose that Alice
sends a message m to Bob and that Bob receives a message m 0 ; both m and m 0
are binary vectors of length q. Bob can check whether m 0 is an even subgraph
easily, but can he be sure that the message m 0 is in fact the message m? If m 0 is
not an even subgraph, Bob is absolutely sure that m 0 6= m (naturally, assuming
that Alice is playing by the rules and transmitting an even subgraph). But if m 0
is an even subgraph, Bob can only be sure that either
- m 0 is indeed m; or
- the channel introduced at least three errors (since m 0 and m must be at

Hamming distance at least three by Lemma 6.2.2).
If errors on the channel are indeed rare, then Bob can reasonably suppose that,
if m 0 is even, it is indeed the message sent by Alice.
Bob can actually do better than this. If only one error was introduced (i.e.
m ⊕m 0 has only one ‘1’ in it), then there is a unique even subgraph closest to m.
For suppose that m 1 ⊕ m 0 and m 2 ⊕ m 0 both have only one 1; then (m 1 ⊕ m 0 ) ⊕
(m 2 ⊕ m 0 ) = m 1 ⊕ m 2 has at most two 1’s, and by Lemma 6.2.2, we must have
m 1 = m 2 . So, in particular, m is the unique closest vector to m 0 if only one error
was made. Of course, this does not tell us how to correct the single error, just
that a single error has a unique correction; more on this later.
Let us formalize this. Let C be a set of binary vectors of length q forming a
vector space of dimension d ; then C has 2d vectors in it. The distance t of C is
the minimum Hamming distance between any two distinct vectors of C . Such
a set C is called a binary (q, d , t )-error correcting code, or just a (q, d , t )-code.
Vectors in C are called codewords. The distance of a code is of fundamental
importance in its use:
Lemma 6.2.3. If fewer than t errors are made in the transmission of a code-
word, then the received message is either the original codeword, or it is not a
codeword at all.
Lemma 6.2.4. If at most b t −1

2
c errors are made in the transmission of a code-
word m, the Hamming distance between the received message m 0 and a code-
word is minimum for the unique codeword m.
Prove these two lemmas for yourself.
Lemma 6.2.5. The cycle space of a connected graph G on p vertices and q edges
is a
(q, q − p + 1, t )-code for some t ≥ 3.
This code is called the even graphical code of G. What is the distance of
such a code?
If G has a cycle of length `, this cycle has Hamming distance ` from the void
graph Z . So evidently the girth g of G is an upper bound on the distance. Can
two even subgraphs S 1 and S 2 have Hamming distance less than g ? Consider
S 1 ⊕ S 2 . If S 1 6= S 2 , then S 1 ⊕ S 2 is an even subgraph, and hence it contains a
cycle. But all cycles have at least g edges, and so
Lemma 6.2.6. The distance of the even graphical code of G equals the girth of
G.
As an aside, let’s observe that we can extend our definition of cycle space to
multigraphs in which repeated edges are permitted. If we choose the two vertex
multigraph with q edges between the vertices, the cycle space is just the set of
all binary vectors of length q with an even number of 1’s. This is the standard
even parity code.
We are left with a significant question: correcting errors. Suppose that an
even subgraph M is transmitted, but that a subgraph S is received which is not
even. Now S = M ⊕ E , where E marks the positions in which a transmission
error was made; equivalently, M = S ⊕ E . Both E and M are unknown to us;
however, we do know all of the valid codewords, and when G has girth g , we
g −1
assume that at most b 2 c errors were introduced. Under this assumption, E
cannot contain a cycle (why?). To find E , let’s observe that a vertex v has odd
degree in S if and only if it has odd degree in E (for it surely has even degree
in M ). So E is a subgraph of G with a specified set of odd degree vertices and
as few edges as possible. There is only one candidate for E , since if S ⊕ E 1 and
6.2. GRAPHICAL CODES 155
S ⊕ E 2 are codewords and E 1 6= E 2 , E 1 ⊕ E 2 is a nonzero codeword — so we have

g −1
a contradiction unless more than b 2 c errors were made. When the number of
errors is “small", E can easily be found by brute-force techniques; when larger,
algorithms for weighted matching in graphs can be used (this material is cov-
ered in C&O 450).
(Aside: Essentially what we do in the general case is to identify the set W of
odd degree vertices of S. There must be an even number 2s of them by the hand-
shake theorem. We choose a labelling of the vertices in W as x 1 , . . . , x s , y 1 , . . . , y s
so that the sum of the distances between x i and y i for i = 1, . . . , s is minimum
over all labellings of the vertices. Then find paths P 1 , . . . , P s where P i is a short-
est path between x i and y i . Determine E = P 1 ⊕ P 2 ⊕ · · · ⊕ P s . Finally compute
M = S ⊕ E to correct the errors, if any.)
One topic of interest is to determine which graphs give “good" codes (large
distance, large dimension and small length). These amount to requiring “large"
girth, “large" cyclomatic number, and “small" number of edges. The Petersen
graph gives a (15, 6, 5)-code, for example. Finding the best codes (and even the
best graphical codes) is a challenging problem, discussed in much more detail
in CO 331.
Problem Set 6.2

1. Produce all codewords in the even graphical code of K 4 . Determine the
length, distance and dimension.
2. Produce all codewords in the even graphical code of K 3,3 . Determine the
length, distance and dimension.
3. Prove that every connected graph on an even number of vertices has a span-
ning subgraph in which all vertices have odd degree. (Hint: consider modulo
2 sums of paths.)
4. If G is a connected graph on an even number of vertices, Problem 3 ensures

that G has an odd subgraph S. Consider the vector space C (G) of even sub-
graphs and let C + (G) = C (G) ∪ {H ⊕ S : H ∈ C (G)}. Is C + (G) a vector space
under modulo 2 sum? Compute the length, distance and dimension of the
corresponding code.
5. Using Problem 4, extend the (15, 6, 5)-code from the Petersen graph to a
(15, 7, 5)-code. Describe all codewords.
Chapter 7
Planar Graphs
7.1 Planarity
Definition 7.1.1. A graph G is planar if it has a drawing in the plane so that its
edges intersect only at their ends, and so that no two vertices coincide. The
actual drawing is called a planar embedding of G, or a planar map.
For example, the 3-cube, which we previously considered in Figure 4.12, is a
planar graph, with a planar embedding given in Figure 7.1. A planar graph may
have a number of essentially different embeddings.
000 001
010 011
110 111
100 101
Figure 7.1: A planar embedding of the 3-cube
It is clear that a graph is planar if and only if each of its components is planar.
So it is often sufficient to consider only connected planar graphs and connected
157
158 CHAPTER 7. PLANAR GRAPHS
planar embeddings.
A planar embedding partitions the plane into connected regions called faces;
one of these regions, called the outer face, is unbounded. For example, the pla-
nar embedding given in Figure 7.2 has 4 faces, identified as f 1 , f 2 , f 3 , f 4 in the
diagram. In this case, the outer face is f 4 .
f4
f1 f3
f2
Figure 7.2: A planar embedding with 4 faces
Consider a planar embedding of a connected graph G. The subgraph formed

by the vertices and edges in a face is called the boundary of the face. We say that
two faces are adjacent if they are incident with a common edge. Assume, for the
moment, that G is connected. As one moves around the entire perimeter of a
face f , one encounters the vertices and edges in a fixed order, say
W f = (v 0 , e 1 , v 1 , e 2 , v 2 , . . . , v n−1 , e n , v n )
where v n = v 0 . This sequence is a closed walk of the graph G, and we call it

the boundary walk of face f . (The boundary walk can start at any vertex, and
can proceed around the perimeter in either a clockwise or counterclockwise
direction.) The number of edges in the boundary walk W f is called the degree
of the face f . For example, in Figure 7.2,
deg( f 1 ) = 6, deg( f 2 ) = 3, deg( f 3 ) = 5, deg( f 4 ) = 14.
In Figure 7.1, all faces have degree 4. Note that a bridge of a planar embedding
is incident with just one face, and is contained in the boundary walk of that
face twice, once for each side. Thus a bridge contributes 2 to the degree of
the face with which it is incident. On the other hand, if e is an edge of a cycle
7.2. EULER’S FORMULA 159
of an embedding, e is incident with exactly two faces, and is contained in the

boundary walk of each face precisely once.
Every edge in a tree is a bridge, so a planar embedding of a tree T has a
single face of degree 2|E (T )| = 2|V (T )| − 2.
In what follows we may use ‘s’ for the number of faces in a planar embed-
ding.
Theorem 7.1.2. If we have a planar embedding of a connected graph G with

faces f 1 , . . . , f s , then
s
X
deg( f i ) = 2|E (G)|.
i =1
Proof: Each edge has two sides, and when we sum the degrees of the faces we
are counting the edges twice, once for each side.
Note the similarity between Theorem 7.1.2 and Theorem 4.3.1. This the-
orem is colloquially known as the Faceshaking Lemma or the Handshaking
Lemma for Faces. We shall make a direct link between these results later when
we consider the dual of a planar embedding.
Corollary 7.1.3. If the connected graph G has a panar embedding with f faces,
the average degree of a face in the embedding is 2|Ef(G)| .
7.2 Euler’s Formula

There are often a number of completely different planar embeddings of a planar
graph. However, every planar embedding of a given connected planar graph
has the same number of faces, a fact that we can deduce from the following
result, called Euler’s Formula.
Theorem 7.2.1. (Euler’s Formula) Let G be a connected graph with p vertices

and q edges. If G has a planar embedding with f faces, then
p − q + f = 2.
Proof: For each positive integer p, we prove this result by induction on q. Since
G is a connected, it has a spanning tree and so q ≥ p − 1.
As a tree has no cycles, any planar embedding of a tree has just one face,
and the theorem holds.
So assume q > p − 1, and assume inductively that Euler’s formula is true for
any connected graph on p vertices with fewer than q edges. Suppose that we
have a planar embedding of G with f faces. Since q ≥ p we see that G is not
a tree and therefore it has an edge e = {u, v} that is not a bridge. Then we also
have a planar embedding of G \ e (the graph we get from G by deleting the edge
e). Since G \ e has p vertices and q − 1 edges and is connected, it follows by
induction that if it has f 1 faces, then
p − (q − 1) + f 1 = 2
and therefore f 1 = q + 1 − p. If we put e back into our drawing, it divides a face

into two. So the embedding of G has one more face than that of G \ e. Hence
the number of faces in the embedding of G is q + 2 − p and then
|V (G)| − |E (G)| + q + 2 − p = p − q + q + 2 − p = 2.
As an example of Euler’s Formula, consider the connected planar embed-

ding in Figure 7.2. In this case there are 12 vertices, 14 edges and 4 faces and
12 − 14 + 4 = 2, as expected.
7.3 Stereographic Projection

There is a lack of symmetry among the faces of a planar map—one of the faces
is unbounded. Symmetry can be achieved, making all the faces bounded, by
considering embeddings on the surface of a sphere rather than on the plane.
The main result in this section is the following:
Theorem 7.3.1. A graph is planar if and only if it can be drawn on the surface of
a sphere.
Any drawing on the plane can be converted to a drawing on the sphere via
stereographic projection. Let the sphere be tangent to the plane at point A,
and let B be antipodal to A on the sphere. In stereographic projection, the im-
age of each point x on the plane is the unique point x 0 on the surface of the
sphere that lies on the line between x and B . This is illustrated in Figure 7.3.
If we apply stereographic projection to a planar embedding, we obtain a
drawing of the graph on the surface of the sphere (with no edges crossing), in
which all faces are bounded, and B is in the interior of a face.
7.3. STEREOGRAPHIC PROJECTION 161
B
x0
A
Figure 7.3: Stereographic projection
On the other hand, given any face f of an embedding G on the sphere, stere-
ographic projection provides a way to obtain a planar embedding H in which
the outer (unbounded) face corresponds to f —turn the sphere so that point B
lies in face f , and then project the embedding G to the plane to get H . If we
redraw an embedding so that a different face becomes the outside face, we con-
sider this to be the same as the original embedding. Roughly speaking, we have
the same graph, and the same faces, and the same faces are incident with the
the same edges, so they are essentially the same embedding. In particular the
number of faces of degree i in two embeddings related in this way will be equal.
A graph may have a number of essentially different planar embeddings. Fig-
ure 7.4 exhibits two embeddings of a planar graph. In the first embedding there
are two faces of degree three and two of degree five; in the second, there are two
faces of degree three, one of degree four and one of degree six. It is reassuring
to note that in both embeddings there are four faces (so Euler is happy) and the
sum of the faces degrees is 16 in both embeddings, as it should be by Theorem
7.1.2.
Problem Set 7.3

1. Prove that every planar embedding has either a vertex of degree at most 3 or
a face of degree 3.
2. Prove that each of the graphs shown in Figure 7.5 is planar, by exhibiting a
planar embedding.
5 5
1 2 1 2
4 3 4 3
Figure 7.4: Embeddings of a Planar Graph
a f
a b
b g
f g c
c h
i h
d i
e d
e
(a) (b)
Figure 7.5: Planarity exercises

7.4. PLATONIC SOLIDS 163
3. Let n ≥ 3 be an integer. Suppose that a convex n-gon is drawn in the plane,

and then each pair of nonadjacent corner points is joined by a straight line
through the interior. Suppose that no 3 of these lines through the interior
meet at a common point in the interior. (Figure 7.6 shows such a drawing
with n = 6.)
Figure 7.6: A convex 6-gon with diagonals
Let f n be the number of regions into which the interior of the n-gon is di-
vided by this process. (So f 3 = 1, f 4 = 4, f 5 = 11.) Use Euler’s Formula to find
f n . (Hint: any set of 4 corner points of the n–gon uniquely determines a pair
of intersecting lines in the interior.)
7.4 Platonic Solids

Consider the two geometric solids in Figure 7.7; the cube and the tetrahedron.
These polyhedra exhibit a great deal of symmetry. In particular, the faces have
the same degree and the vertices have the same degree. We call all such poly-
hedra platonic solids. Surprisingly, there are just five platonic solids: the tetra-
hedron, the cube, the octahedron, the dodecahedron, and the icosahedron. In
this section we will outline a proof of this remarkable fact.
From each platonic solid, we can obtain a planar embedding in which all
vertices have the same degree d ≥ 3 and all faces have the same degree d ∗ ≥ 3;
Figure 7.7: The tetrahedron and the cube
see Figure 7.8. We call a graph platonic if it admits a planar embedding in which
each vertex has the same degree d ≥ 3 and each face has the same degree d ∗ ≥ 3.
We will show that the only platonic graphs are those given in Figure 7.8, from
which it is easy to deduce that there are just five platonic solids.
Theorem 7.4.1. There are exactly five platonic graphs.
We require the following lemma.
Lemma 7.4.2. Let G be a planar embedding with p vertices, q edges and s faces,
in which each vertex has degree d ≥ 3 and each face has degree d ∗ ≥ 3. Then
(d , d ∗ ) is one of the five pairs
{(3, 3), (3, 4), (4, 3), (3, 5) (5, 3)}.
Proof: Assume p = |V (G)|, q = |E (G)| and that there are exactly f faces in the
embedding. From Euler,
p − q + f = 2.
Since G is regular with degree d , we have 2q = d p by Theorem 4.3.1. Since
each face of the embedding has degree d ∗ , we have 2q = d ∗ f by Theorem 7.1.2.
Hence p = 2q/d and f = 2q/d ∗ . So we can write Euler’s equation as
µ ¶
2q 2q 2 2
2= −q + ∗ = q −1+ ∗ .
d d d d
We rewrite this in turn as

2 2 2
+ ∗ = 1+ . (7.4.1)
d d q
The basic idea now is to note that the right side of this equality is greater than
1, while the left is struggling to reach 1.
7.4. PLATONIC SOLIDS 165
(a) (b) (c)
(d) (e)
Figure 7.8: (a) the tetrahedron; (b) the octahedron; (c) the cube;(d) the icosahe-
dron; (e) the dodecahedron
If d = 3 and d ∗ ≥ 6, then
2 2 2 2
+ ∗ ≤ + = 1.
d d 3 6
2 2 2 2
+ ∗ ≤ + = 1.
d d 4 4
2 2 2 2 18
+ ∗≤ + = < 1.
d d 5 4 20
Finally, if d ≥ 6 and d ∗ ≥ 3, then
2 2 2 2
+ ∗ ≤ + = 1.
d d 6 3
It follows that (d , d ∗ ) must be one of the five pairs in the statement of lemma.
Lemma 7.4.3. If G is a platonic graph with p vertices, q edges and f faces, where
each vertex has degree d and each face degree d ∗ , then
2d d ∗
q=
2d + 2d ∗ − d d ∗
and p = 2q/d and f = 2q/d ∗ .
Proof: In proving the previous lemma we saw that

µ ¶
2 2
2=q −1+ ∗ .
d d
and the formula for q follows at once from this. We also saw that 2q = pd = f d ∗ ,
which yields the other two claims.
To prove Theorem 7.4.1, we look at the five possible values for (d , d ∗ ) given
in the previous lemma, and show that in each case there is a unique planar
embedding with the required parameters. We will consider the first two cases;
the remaining three cases are left as exercises.
Case 1: d = 3, d ∗ = 3.
2·3·3
Thus q = 2·3+2·3−3·3 = 6, p = 2·6 2·6
3 = 4, and s = 3 = 4. Note that K 4 is the only
graph having 4 vertices and 6 edges.
7.5. NONPLANAR GRAPHS 167
Case 2: d = 3, d ∗ = 4.
2·3·4
Thus q = 2·3+2·4−3·4 = 12, p = 2·12 2·12
3 = 8, and s = 4 = 6. Consider a planar
embedding G with these parameters. Since each face has degree 4, no vertex
can be repeated in a boundary walk, so each face boundary is a 4–cycle. Let
C = (v 1 , v 2 , v 3 , v 4 , v 1 ) be the boundary of one of the faces of G. If v 1 is adjacent to
v 3 then (v 2 , v 3 , v 1 , v 2 ) is part of some boundary walk. This contradicts that each
face is bounded by a 4–cycle. Therefore, v 1 is not adjacent to v 3 , and, similarly,
v 2 is not adjacent to v 4 . Since each vertex has degree 3, v i has a unique neigh-
bour u i not in C , for i = 1, 2, 3, 4. Note that u 1 , v 1 , v 2 , u 2 is part of the boundary
walk of some face, so u 1 6= u 2 and u 1 u 2 is an edge. Similarly u 2 u 3 , u 3 u 4 , and
u 4 u 1 are also edges. If u 1 = u 3 then u 4 is only incident with two faces, namely
(u 4 , v 4 , v 1 , u 1 , u 4 ) and (u 4 , v 4 , v 3 , u 1 , u 4 ). This contradicts that u 4 has degree 3,
and, hence, u 1 6= u 3 . By symmetry, u 2 6= u 4 . Therefore G is the 3–cube.
Problem Set 7.4

1. Show that there is a unique planar embedding in which each vertex has de-
gree 4 and each face has degree 3.
7.5 Nonplanar Graphs

In order to prove that a graph is planar, we can find a planar embedding. How
would we prove that a graph is nonplanar? In this section we will see that in
some case we can use Euler’s formula to do this.
We need one technical result before we can start.
Lemma 7.5.1. If G is connected and not a tree, then in a planar embedding of

G, the boundary of each face contains a cycle.
Proof: Since G has a cycle, it has more than one face.

Therefore, every face f is adjacent to at least one other face, say g .
Let e 1 = {v 0 , v 1 } be an edge that is incident with both f and g . Let H be the

component in the boundary graph of face f containing the edge e 1 . Let
W f = (v 0 , e 1 , v 1 , e 2 , v 2 , . . . , v n−1 , e n , v 0 )
be the boundary walk of H . Since the edge e 1 is incident with both f and g , it is
contained in W f precisely once.
The edge e 1 is not a bridge of H because
(v 1 , e 2 , v 2 , . . . , v n−1 , e n , v 0 )
is a walk from v 1 to v 0 in H − e 1 . Therefore, by Theorem 4.10.3, H contains a

cycle. This establishes the result.
Lemma 7.5.2. Let G be a connected planar embedding with p vertices and q

edges. If each face of G has degree at least d ∗ , then (d ∗ − 2)q ≤ d ∗ (p − 2).
Proof: Let f 1 , f 2 , . . . , f s be the faces of G. Thus, applying Theorem 7.1.2, we have

s s
d ∗ = sd ∗ .
X X
2q = deg( f i ) ≥
i =1 i =1
Now, by Euler’s Formula, s ≥ q + 2 − p. So,
2q ≥ d ∗ s ≥ d ∗ (q + 2 − p) = d ∗ q + 2d ∗ − d ∗ p.
Therefore, (d ∗ − 2)q ≤ d ∗ (p − 2), as required.

Note that this lemma also applies to planar graphs that are not connected.
This is because we could add edges on the outer face to connect the compo-
nents together, resulting in a connected planar graphs. After applying the lemma,
we then remove the edges we have added, and we still apply the same bound.
Lemma 7.5.2 relies on a given planar embedding, since it involves the face
degrees. We would like similar inequalities for planar graphs. The following
lemma allows us to relate face degrees to the lengths of cycles, which are inde-
pendent of the embedding.
Note that a graph with p vertices can have as many as p(p − 1)/2 edges, a
quadratic function of p. However, this is not the case for planar graphs. The
following theorem shows that the number of edges in a planar graph is bounded
by a linear function in the number of vertices p. (Note that if we allow multiple
edges or loops, fixing p does not bound the number of edges; for example, the
graph in Figure 7.9 has p = 2 vertices and q edges for any positive integer q.)
7.5. NONPLANAR GRAPHS 169
e1
e2
u . v
.
.
eq
Figure 7.9: Multiple edges in a planar multigraph
Theorem 7.5.3. In a connected planar graph G with p ≥ 3 vertices and q edges,

we have
q ≤ 3p − 6.
Proof: If G is a tree, then q = p − 1 ≤ 3p − 6 whenever p ≥ 3.

If G is not a tree, then by Lemma 7.5.1, every face contains a cycle. Since
each cycle in G has length at least three, each face has degree at least 3. By
Lemma 7.5.2,
q ≤ 3(p − 2)/(3 − 2) = 3p − 6.
Theorem 7.5.3 gives an inequality that holds for all planar graphs. Therefore,
if a graph does not satisfy the inequality, then it is not planar. We illustrate this
idea below on K 5 .
Corollary 7.5.4. K 5 is a not planar.
Proof: We have Ã !
5
|E (K 5 )| = = 10.
2
But 3p − 6 = 15 − 6 = 9, so for K 5 we have
q = 10 > 9 = 3p − 6,
and the inequality q ≤ 3p − 6 does not hold. We conclude from Theorem 7.5.3
that K 5 is not a planar graph.
Notice that Theorem 7.5.3 is only a necessary condition for a graph to be

planar. If a graph satisfies |E (G)| ≤ 3|V (G)|−6, it does not follow that it is planar.
As an example, consider the graph of Figure 7.10. Since it has K 5 as a subgraph,
it cannot be planar, but |E (G)| = 11 < 12 = 3|V (G)| − 6.
Figure 7.10: A nonplanar graph with q ≤ 3p − 6
Corollary 7.5.5. A planar graph has a vertex of degree at most five.
Proof: This is true if |V (G)| ≤ 2. If p = |V (G)| ≥ 3 and q = |E (G)| then by Theo-

rem 7.5.3,
q ≤ 3p − 6
and therefore
2q 12
≤ 6− .
p p
This shows that the average degree of a vertex in G is less than six, and therefore
G contains a vertex of degree at most five.
After a few attempts to make a planar embedding of K 3,3 , you will conclude
that it also is not planar. However, neither Theorem 7.5.3 nor Corollary 7.5.5
implies that K 3,3 is not planar. (You should check this.) Nevertheless, one can
repeat the proof of Theorem 7.5.3 with an additional observation to obtain a
result strong enough to prove that K 3,3 is not planar. Recall that the girth of a
graph G having a cycle, is the length of a shortest cycle of G.
7.6. KURATOWSKI’S THEOREM 171
Theorem 7.5.6. In a connected bipartite planar graph G with p ≥ 3 vertices and

q edges, we have
q ≤ 2p − 4.
Proof: If G is a tree, then q = p − 1 ≤ 2p − 4 whenever p ≥ 3.

If G is not a tree, then by Lemma 7.5.1, every face contains a cycle. Since G
is bipartite, it does not have cycles of length 3. Hence each cycle in G has length
at least four, and each face has degree at least 4. By Lemma 7.5.2,
q ≤ 4(p − 2)/(4 − 2) = 2p − 4.
Lemma 7.5.7. K 3,3 is not planar.
Proof: We have
|E (K 3,3 )| = 3 · 3 = 9.
But 2p − 4 = 2 · 6 − 4 = 8, so for K 3,3 we have
q = 9 > 8 = 2p − 4
and the inequality q ≤ 2p − 4 does not hold. We conclude from Theorem 7.5.6
that K 3,3 is not a planar graph.
The proof of this lemma shows that if G is a planar graph with girth at least
four, then
|E (G)| ≤ 2|V (G)| − 4
and thus if p = |V (G)|, then the average degree of a vertex is
8
4− .
p
You might prove that if this bound holds with equality, then all faces have de-
gree four. (The 3-cube provides an example.)
7.6 Kuratowski’s Theorem

We have been able to prove by counting arguments that two graphs, K 5 and K 3,3 ,
are not planar. We are going to see how we can use graphs based on K 5 and K 3,3
to certify that a graph is not planar.
First we need some language to properly state our result.

An edge subdivision of a graph G is obtained by applying the following oper-
ation, independently, to each edge of G: replace the edge by a path of length 1 or
more; if the path has length m > 1, then there are m − 1 new vertices and m − 1
new edges created; if the path has length m = 1, then the edge is unchanged.
For example, Figure 7.11 shows a graph H , and an edge subdivision G of H .
Note that the operation of edge subdivision does not change planarity: if G
is a planar graph, then all edge subdivisions of G are planar; if G is nonplanar,
then all edge subdivisions of G are nonplanar. Similarly, note that if a graph G
has a nonplanar subgraph, then G is nonplanar.
From these observations, we can immediately conclude that if a graph G
has a subgraph isomorphic to an edge subdivision of K 3,3 or K 5 , then G is non-
planar. One of the most famous results of graph theory, known as Kuratowski’s
Theorem, establishes that the converse of this result is also true.
Theorem 7.6.1. A graph is not planar if and only if it has a subgraph that is an
edge subdivision of K 5 or K 3,3 .
(We omit the proof of this result as it is beyond the scope of our current
study.)
Note that Kuratowski’s Theorem has the following surprising consequence.
Suppose we begin with a nonplanar graph G and do the following operation
as long as it is possible. Delete a vertex v or an edge e whose deletion leaves
0
a nonplanar graph. When this process ends, what will the final graph G be?
0
According to Kuratowski’s Theorem, G must be an edge subdivision of K 5 or
K 3,3 !
For example, in Figure 7.11, the graph M 5 has subgraph G, and G is an edge
subdivision of H . Also, H is actually K 3,3 , where the vertices in the two vertex
classes are {1, 5, 8} and {3, 6, 10}. Thus we conclude from Kuratowski’s Theorem
that M 5 is nonplanar. Figure 7.11 illustrates one convenient strategy when look-
ing for an edge subdivision of K 3,3 or K 5 : first find a long cycle in the graph – in
this case the cycle has 10 vertices; then find edges or paths (subdivided edges)
across the cycle – in this case edges {1, 10} and {5, 6}, together with any one of
edges {2, 7}, {3, 8}, {4, 9}, would create an edge subdivision of K 3,3 .
Note that M 5 has 10 vertices, 15 edges and girth 4, so our counting results
are not strong enough to prove that M 5 is not planar.
1 2 3 4 5
M5
6 7 8 9 10
1
6 2
7 3
G
8 4
9 5
10
6 3
8 5
10
Figure 7.11: Edge subdivisions and an example of Kuratowski’s Theorem

Problem Set 7.6

1. For each of the graphs in Figure 7.12, determine if it is planar. Prove your
conclusion in each case.
2. Let G be a connected planar graph with p vertices and q edges and girth k.
Show that
k(p − 2)
q≤ .
k −2
Show also that if equality holds, all faces of G have degree k.
3. Prove that the Petersen graph is nonplanar, without using any form of Kura-
towski’s theorem.
4. (a) Prove that the Petersen graph is nonplanar by Kuratowski’s Theorem,

finding a subgraph that is an edge subdivision of K 3,3 .
(b) Show that there exist two edges of the Petersen graph whose deletion
leaves a planar graph.
5. Prove that the n-cube is not planar when n ≥ 4, without using any form of
Kuratowski’s theorem.
6. (a) Prove that the 4-cube is nonplanar by Kuratowski’s Theorem, finding a

subgraph that is an edge subdivision of K 3,3 .
(b) Show that there exist four edges of the 4-cube whose deletion leaves a
planar graph.
(c) Show that no matter which 3 edges are deleted from a 4-cube, the re-
sulting graph is not planar.
7. Consider the graph G n with V (G n ) = {3, 4, . . . , n} and {u, v} ∈ E (G n ) if and only

if u|v or v|u.
(a) Find the largest value k such that G k is planar and give a planar embed-
ding of G k .
(b) Prove that G k+1 is not planar using Kuratowski’s Theorem.
(c) Find n < 50 so that G n has K 5 as a subgraph.
(d) Find n < 40 so that G n has K 3,3 as a subgraph.
8. Consider the prime graph B n defined in Problem 11 of Problem Set 4.4.

1
1 8
11 2
2 9 10 3
3 10 9 4
4 11
8 5
7 6
5 12
(b)
6 13
a f
7 14
b g
(a)
a
j c h
b
i c
d i
h d
e j
g e
f
(c) (d)
Figure 7.12:
(a) Prove that B 8 is planar.

(b) Using Kuratowski’s Theorem, prove that B 9 is not planar.
9. Prove that every planar bipartite graph G has a vertex of degree at most three.
10. Let G denote the graph below. (You may assume, without proof, that G has
girth six.)
(a) Let H be any graph obtained from G by deleting two edges. Prove that
H is not planar.
(b) Prove that there exist three edges that can be deleted from G so that the
resulting graph is planar.
a b
n c
m d
l e
k f
j g
i h
Figure 7.13: Planarity exercise.
11. Prove that every planar graph having girth at least six has a vertex of degree
at most two. Prove that this is false if the girth is five.
12. Prove that if G is a planar graph in which every vertex has degree at least five,
then |V (G)| ≥ 12. Find such a graph with |V (G)| = 12.
13. (a) Prove that the complement of the 3-cube is nonplanar.

7.7. COLOURING AND PLANAR GRAPHS 177
(b) Prove that if G has p ≥ 11 vertices, then at least one of G and Ḡ is non-
planar.
7.7 Colouring and Planar Graphs

Definition 7.7.1. A k-colouring of a graph G is a function from V (G) to a set
of size k (whose elements are called colours), so that adjacent vertices always
have different colours. A graph with a k-colouring is called a k-colourable
graph.
For example, Figure 7.14 gives a 4-colouring of a graph. The colours in this
case are the integers 1, 2, 3, 4, and the colour assigned (by the function) to each
vertex is written beside the vertex in the diagram.
1 2 3
4 3 4
1 2
Figure 7.14: A 4-colouring of a graph
Theorem 7.7.2. A graph is 2-colourable if and only if it is bipartite.
Proof: If a graph has bipartition (A, B ), then we obtain a 2-colouring by assign-

ing one colour to all vertices in A, and the second colour to all vertices in B ;
no edge joins two vertices of the same colour since (A, B ) is a bipartition (so all
edges join a vertex in A to a vertex in B ).
Conversely, if a graph has a 2-colouring, then let the vertices of one colour
be set A, and the vertices of the second colour be set B . Then (A, B ) is a biparti-
tion since all edges join vertices of different colours.
It follows that we have a description of all the 2-colourable graphs — they

are precisely the graphs in which all cycles have even length. However, 3-colourable
graphs are not well understood. There is no simple description known, and
graph theorists believe that none exists.
Similarly, whereas there is an efficient algorithm to determine whether a
graph is 2-colourable (Theorem 5.5.1), graph theorists believe that no efficient
algorithm exists to determine whether a graph is 3-colourable.
Theorem 7.7.3. K n is n-colourable, and not k-colourable for any k < n.
Proof: Any graph on n vertices is n-colourable; assign different colours to the
vertices, so no two vertices have the same colour. K n is not k-colourable for
any k < n, since such a colouring would assign the same colour to some pair of
vertices. But all pairs of vertices are adjacent in K n , so we would have assigned
the same colour to adjacent vertices. Thus K n has no k-colouring for k < n.
Theorem 7.7.4. Every planar graph is 6-colourable.

Proof: The proof is by induction on the number p of vertices. First note that all
graphs on one vertex are 6-colourable, so the result is true for p = 1.
For the induction hypothesis, assume the result is true for all planar graphs
on p ≤ k vertices, where k ≥ 1.
Now consider a planar graph G on p = k + 1 vertices. From Corollary 7.5.5,
G has a vertex v with deg(v) ≤ 5. Suppose we remove vertex v, and all edges
incident to v, from G, and call the resulting graph G 0 . Then G 0 has k vertices,
and is a planar graph (all subgraphs of a planar graph are planar). Thus we
can apply the induction hypothesis to G 0 , so G 0 is 6-colourable. Now find a 6-
colouring of G 0 . There are at most 5 vertices in G 0 that are adjacent to v in G,
so these vertices are assigned at most 5 different colours in the 6-colouring of
G 0 . Thus there is at least one of the 6 colours remaining. Assign one of these
remaining colours to v, so that v has a different colour from all of its adjacent
vertices in G. Thus we have a 6-colouring of G, and the result is true for p = k +1.
We have now proved that the result is true by mathematical induction.
Let us proceed to the Five-Colour Theorem. The proof relies on the notion
of edge-contraction which we now define.
Definition 7.7.5. Let G be a graph and let e = {x, y} be an edge of G. The graph
G/e obtained from G by contracting the edge e is the graph with vertex set
V (G) \ {x, y} ∪ {z}, where z is a new vertex, and edge set
{{u, v} ∈ E (G) : {u, v}∩{x, y} = ;}∪{{u, z} : u ∉ {x, y}, {u, w} ∈ E (G) for some w ∈ {x, y}}.
7.7. COLOURING AND PLANAR GRAPHS 179
Intuitively, we can think of the operation of contracting e as allowing the

“length” of e to decrease to 0, so that the vertices x and y are identified into a
new vertex z. Any other vertex that was adjacent to one (or both) of x and y is
adjacent to z in the new graph G/e. See Figure 7.15 for an example.
If G has p vertices and q edges then G/e has p − 1 vertices and at most q − 1
edges.
5 2 5 2
13
e
−→
4 3 4
G G/e
Figure 7.15: Contraction
Remark: G/e is planar whenever G is.

The converse of this remark is not true; G/e may be planar when G is non-
planar. See Figure 7.7 for an example.
Now we are ready to prove the Five-Colour Theorem.
Proof: The proof is by mathematical induction on the number of vertices p of

a planar graph.
For any graph G having one vertex, G is 5-colourable.
Induction Hypothesis: Assume that every planar graph on p ≤ k vertices is 5-
colourable, where k ≥ 1.
Let G be any planar graph on p = k + 1 vertices.
Case 1: Graph G has a vertex of degree 4 or less.
Proceeding as in the 6-colour theorem above, it can be shown that G is 5-colourable.
−→
e
K3,3 K3,3/e
Figure 7.16: Contraction
Case 2: Graph G has no vertex of degree 4 or less.

Then, by Corollary 7.5.5, G has a vertex, say v, of degree 5.
If the 5 vertices joined to v are mutually adjacent, then G has a subgraph
isomorphic to K 5 . This is impossible since G is planar by hypothesis. Hence, at
least 2 neighbours of v are not adjacent, say a and b.
Contract edge e = {a, v} to get graph H = G/e, and label the new vertex v.
Contract edge f = {b, v} of H to get graph K = H / f , and label the new vertex v.
Since G is planar, so are H and K ; furthermore, K has k − 1 vertices. Therefore
K is 5-colourable by the induction hypothesis.
Use the 5-colouring of K to colour all the vertices of G except for a, b, and
v. Since a and b are not adjacent in G, they can both accept the same colour.
Colour both vertices a and b with the colour assigned to vertex v in K . Hence,
at most four distinct colours appear on the five neighbours of v. Colour v with
one of the absent colours. Then we have a valid 5-colouring of G.
In both cases, G has a 5-colouring.
Therefore, by mathematical induction it follows that every planar graph has
a 5-colouring.
The most famous result of graph theory is the “Four colour theorem". The
known proofs are by mathematical induction, but involve many hundreds of
cases, and use computer verification for cases. Its statement is as follows.
It is important to note that the converse of this result does not hold. For ex-
ample, from Theorem 7.7.3 we know that K 5 is not 4-colourable, so it is correct
to deduce from the Four Colour Theorem that K 5 is not planar. However, from
7.8. DUAL PLANAR MAPS 181
Theorem 7.7.2 we know that K 3,3 is 2-colourable (and thus 4-colourable), but
we know already from Corollary 7.5.7 that K 3,3 is not planar.
7.8 Dual Planar Maps

Note that our colourings are colourings of vertices, with the restriction that
vertices joined by an edge have different colours. However, most descriptions
of the Four Colour Theorem refer to colouring of regions, say countries in a
map, with the restriction that regions with a common boundary have different
colours.
In fact, these are completely equivalent results, as we show now. Given a
connected planar embedding G, the dual G ∗ is a planar embedding constructed
as follows: G ∗ has one vertex for each face of G. Two vertices of G ∗ are joined by
an edge whenever the corresponding faces of G have an edge in common (one
side for each face), and the edge in G ∗ is drawn to cross this common boundary
edge in G. For example, Figure 7.17 illustrates the construction of G ∗ from a
planar embedding G. Note that the faces of G ∗ now correspond to the vertices
of G.
There are a number of things to note about the relationship between G and
G . First, a face of degree k in G becomes a vertex of degree k in G ∗ , and a
∗
vertex of degree j in G becomes a face of degree j in G ∗ . Thus, Theorems 4.3.1

and 7.1.2 are the same result for planar embeddings, since Theorem 4.3.1 for G
becomes Theorem 7.1.2 for G ∗ and vice versa. This connection between G and
G ∗ is even stronger, since (G ∗ )∗ and G are the same graph.
Note that a bridge in G gives an edge in G ∗ between a vertex and itself (such
an edge is called a loop), and more than one edge between two faces in G gives
more than one edge between a pair of vertices (these are called, together, a mul-
tiple edge). Thus G ∗ may be a multigraph rather than a graph.
These complications aside, it is now clear that the Four Colour Theorem for
colouring vertices in planar graphs is equivalent to the Four Colour Theorem
for colouring faces in planar embeddings, via duality.
Problem Set 7.8

1. (a) Prove that every planar graph without a triangle (that is, a cycle of length
three) has a vertex of degree three or less.
Figure 7.17: The dual of a planar embedding

7.8. DUAL PLANAR MAPS 183
(b) Without using Theorem 7.7.7, prove that every planar graph without a
triangle is 4-colourable.
2. Prove that if G is a planar graph with girth at least six, then G is 3-colourable.
(You may use the result of Problem 11 in Problem Set 7.6.)
3. Let G be a connected planar embedding with p vertices and q edges, and

suppose that the dual graph G ∗ is isomorphic to G.
(a) Prove that q = 2p − 2.

(b) Give an example of such a graph with six vertices.
4. Let G be a connected planar embedding in which every face has even degree.
Prove that G is a bipartite graph.
5. Find a planar bipartite 3-regular graph with 14 vertices.
6. Show that a planar graph with p > 2 vertices and 2p−3 edges is not 2-colourable.
7. Show that a graph with 2m vertices and m 2 + 1 edges is not 2-colourable.
8. Show that K 5 can be obtained by contracting five edges of the Petersen graph.
Hence deduce from the nonplanarity of K 5 and the remark following the def-
inition of edge-contraction, that the Petersen graph is nonplanar.
Chapter 8
Matchings
8.1 Matching
A matching in a graph G is a set M of edges of G such that no two edges in
M have a common end. (Another way to express the condition is that in the
spanning subgraph of G with edge set M , every vertex has degree at most 1.) So
M matches certain pairs of adjacent vertices—hence the name. The thick edges
in Figure 8.1 form a matching. We say that a vertex v of G is saturated by M , or
that M saturates v, if v is incident with an edge in M . Of course, every graph has
a matching; for example the empty set ; is always a matching. The question we
will be most interested in is to find a largest matching in G, called a maximum
matching of G. In Figure 8.1 the matching M indicated there has size 3, and
therefore is not a maximum matching, since it is easy to find a matching of size
4. A special kind of maximum matching is one having size p/2, that is, one that
saturates every vertex, called a perfect matching. Of course, not every graph
has a perfect matching.
We will be concentrating on matching problems for bipartite graphs. Here
is a way to restate the problem in case G is bipartite.
Job assignment problem. We are given a set A of workers and a set B of jobs,
and for each job, the set of workers capable of doing the job. We want to as-
sign as many jobs as possible to workers able to do them, but each worker is to
be assigned to at most one job, and each job is to be assigned to at most one
worker.
For example, suppose that A = {a, b, c, d , e, f } and B = {g , h, i , j , k, l }, and the
lists of workers that can do the jobs are:
185
186 CHAPTER 8. MATCHINGS
a b c d
i h g e
Figure 8.1: G with matching M
For g : c, e
For h : a, c
For i : a, b, c, d , f
For j : c, e
For k : c, e
For l : b, d , f
Why is the job assignment problem equivalent to the maximum matching prob-
lem of bipartite graphs? From the data of the job assignment problem we can
construct a bipartite graph G with vertex set A ∪ B and with u ∈ A adjacent to
v ∈ B if and only if worker u can do job v. (The graph corresponding to the sam-
ple data above is shown in Figure 8.2.) Conversely, given a bipartite graph with
bipartition A, B we make a worker for each element of A and a job for each ele-
ment of B and declare worker u to be able to do job v if and only if {u, v} ∈ E (G).
The condition that jobs be assigned to workers that can do them, means that an
assignment is a set of edges. The condition that each worker be assigned to at
most one job and that each job be assigned to at most one worker, corresponds
to the condition that the assigned edges form a matching.
If we have a matching M of G, certain kinds of paths are useful for obtaining
a larger matching. We say that a path v 0 v 1 v 2 . . . v n is an alternating path with
respect to M if one of the following is true:
{v i , v i +1 } ∈ M if i is even and {v i , v i +1 } ∉ M if i is odd
{v i , v i +1 } ∉ M if i is even and {v i , v i +1 } ∈ M if i is odd.
8.2. COVERS 187
A a b c d e f
B g h i j k l
Figure 8.2: A bipartite example
That is, edges of the path are alternately in and not in M . In the graph of Figure
8.1, with respect to the matching indicated there, the following are examples
of alternating paths: (i) ahbg , (ii) i ahg b, (iii) i ahbg f . An augmenting path
with respect to M is an alternating path joining two distinct vertices neither of
which is saturated by M . The path (iii) above is an augmenting path in Figure
8.1. Note that augmenting paths have odd length because they begin and end
with nonmatching edges.
Lemma 8.1.1. If M has an augmenting path, it is not a maximum matching.
Proof: Let P be an augmenting path v 0 v 1 v 2 . . . v n . Then {v i , v i +1 } ∈ M if i is odd,

and {v i , v i +1 } ∉ M if i is even. Moreover, n must be odd. So there are fewer edges
of P in M than not in M . If we replace the edges of M that are in P by the other
edges of P , we get a matching M 0 that is larger than M .
8.2 Covers
A cover of a graph G is a set C of vertices such that every edge of G has at least
one end in C . In Figure 8.1 {a, h, g , c, e} is a cover. It is easy to find large covers,
just as it is easy to find small matchings. For example, in any graph G, V (G) is a
cover. Also, if G is bipartite with bipartition A, B , then A is a cover, and so is B .
A very useful observation about matchings and covers is the following.
Lemma 8.2.1. If M is a matching of G and C is a cover of G, then |M | ≤ |C |.

Proof: For each edge {u, v} of M , u or v is in C . Moreover, for two different edges
of M , any vertices of C they saturate must be different, since M is a matching.
Therefore, |M | ≤ |C |.
Sometimes we can use a cover to prove that a matching is maximum.
Lemma 8.2.2. If M is matching and C is a cover and |M | = |C |, then M is a max-

imum matching and C is a minimum cover.
Proof: Let M 0 be any matching. Then by Lemma 8.2.1
|M 0 | ≤ |C | = |M |.
It follows that M is a maximum matching. Now let C 0 be any cover. Then by

Lemma 8.2.1
|C 0 | ≥ |M | = |C |.
It follows that C is a minimum cover.
We will show that if G is bipartite, then it is always possible to find such M
and C , that is, that the maximum size of a matching is the minimum size of a
cover. This is König’s Theorem, the subject of the next section.
Problem Set 8.2

1. Show that a tree has at most one perfect matching.
2. How many perfect matchings are there in K n ? How many in K m,n ?
3. How many perfect matchings has the graph L n of Figure 8.3? (There are n
vertical edges.) (Hint: where a n denotes the number of perfect matchings of
L n , find a recurrence relation for a n .)
4. Show that for n ≥ 1, the n-cube has a perfect matching.
5. Show that the 64 squares of a chessboard can be covered with 32 dominoes,

each of which covers two adjacent squares.
6. Show that if two opposite corner squares of a chessboard are removed, then
the resulting board cannot be covered with 31 dominoes.
7. Let G be a graph with even number of vertices. Prove that if G has a Hamilton
cycle, then G has a perfect matching.
8.2. COVERS 189
···
Figure 8.3: L n
8. In the previous problem, suppose in addition that G is bipartite, with bipar-

tition A, B . Let u ∈ A, v ∈ B , and let H denote the graph obtained from G by
deleting u and v and their incident edges. Prove that H has a perfect match-
ing.
9. Show that if two squares of the chessboard having opposite colours are re-
moved, then the resulting board can be covered by 31 dominoes. Hint: Use
the previous exercise.
10. Consider the prime graph B n introduced in Problem 11 of Problem Set 4.4.
Use induction on n to show that, if n is even, B n has a perfect matching. You
may use without proof the fact that there is a prime number between k and
2k for k ≥ 2.
11. Prove that C is a cover of G if and only if V (G) \ C is a set of pairwise nonad-
jacent vertices.
12. Show that it is not always true that there exist a matching M and a cover C
of the same size.
13. Let N be a matrix. We want to find a largest set of non-zero entries of N

such that no two are in the same row or in the same column. Formulate this
problem as one of finding a maximum matching in a bipartite graph.
14. In the previous problem interpret the meaning of a cover of the bipartite
graph in terms of the matrix N .
15. Find a bipartite graph G with bipartition A, B where |A| = |B | = 5, and having
the following properties. Every vertex has degree at least 2, the total number
of edges is 16, and G has no perfect matching. Why does your graph not have
a perfect matching?
16. Suppose that for some n ≥ 1, graph G with p vertices satisfies p = 2n and
d eg (v) ≥ n for every vertex v. Prove that G has a perfect matching. (Hint:
Prove that if M is a matching that is not perfect, then there exists an aug-
menting path of length 1 or 3.)
17. Suppose that M is a matching of G that is not contained in any larger match-
ing, and that M 0 is a maximum matching of G. Prove that |M 0 | ≤ 2|M |.
8.3 König’s Theorem

The main result about matching in bipartite graphs is the following theorem of
König.
Theorem 8.3.1. (König’s Theorem) In a bipartite graph the maximum size of a
matching is the minimum size of a cover.
Though minimum covers have the same size as maximum matchings in bi-
¥ p ¦ a minimum cover in K p
partite graphs, this is not true in general. For example,
has size p − 1 and a maximum matching has size 2 .
Let A, B be a bipartition of G, and let M be a matching of G. In the following
XY-construction we use alternating paths to define sets X , Y , and show that
they have certain properties. This will allow us to prove König’s Theorem, and
also to give an efficient algorithm to find a maximum matching.
Let X 0 be the set of vertices in A not saturated by M and let Z denote the
set of vertices in G that are joined by to a vertex in X 0 by an alternating path. If
v ∈ Z we use P (v) to denote an alternating path that joins v to X 0 . Now define:
(a) X = A ∩ Z .
(b) Y = B ∩ Z .
For example, in the graph of Figure 8.2 we have
X 0 = {b, d }, X = {a, b, c, d , e, f }, Y = {g , h, i , j , k, l }.
As a second example, consider the same graph but with a different matching,
shown in Figure 8.4. Then we have
X 0 = {d }, X = {d , b, f }, Y = {i , l }.
8.3. KÖNIG’S THEOREM 191
A a b c d e f
B g h i j k l
Figure 8.4: A second bipartite example
Notice that any alternating path P (v) has even length if v ∈ X and odd
length if v ∈ Y . Since the first edge of any alternating path beginning at a ver-
tex in X 0 is not a matching edge, and every second edge is a matching edge, it
follows that
• If v ∈ X , then the last edge of P (v) is in M (this is true vacuously if v ∈ X 0 ).
• If v ∈ Y , then the last edge of P (v) is not in M .
One more easy observation: If w is a vertex of an alternating path P (v) from X 0

to v ∈ Z , then w ∈ Z .
Lemma 8.3.2. Let M be a matching of bipartite graph G with bipartition A, B,

and let X and Y be as defined above. Then:
(a) There is no edge of G from X to B \ Y ;
(b) C = Y ∪ (A \ X ) is a cover of G;
(c) There is no edge of M from Y to A \ X ;
(d) |M | = |C | − |U | where U is the set of unsaturated vertices in Y;
(e) There is an augmenting path to each vertex in U .
Proof: Suppose that (a) is false, and let u ∈ X , v ∈ B \ Y , {u, v} ∈ E (G). Then
adding v to the even-length alternating path P (u) from X 0 to u, gives us an
odd-length alternating path to v, which implies that v ∈ Y , a contradiction.
For (b), the only edges of G that are not incident with an element of C are
those from X to B \ Y . However, no such edges exist by part (a), so C is a cover.
Now suppose that (c) is false, and let u ∈ Y , v ∈ A\X , {u, v} ∈ M . Then adding
v to the odd-length alternating path P (u) from X 0 to u, gives us an even-length
alternating path to v, which implies that v ∈ X , a contradiction.
For (d ), from part (c), every edge of M joins a vertex in Y to a vertex in X , or
joins a vertex in A \ X to a vertex in B \ Y . The number of edges of the first type
is |Y | − |U |. Also, by the fact that X 0 ⊆ X , every vertex in A \ X is saturated by M ,
so the number of edges of the second type is |A \ X |. It follows that
|M | = |Y | − |U | + |A \ X | = |C | − |U |.
Finally, (e) is easy—if such a vertex v exists, then P (v) is an augmenting path.
We can check that the conclusions of Lemma 8.3.2 are satisfied in the two
examples above. Notice in particular, that in Figure 8.2 we have U = {g , j }, and,
for example, augmenting paths to g and j are given by bi c g and bi c j . If we use
the latter path to get a larger matching, then we get the matching of Figure 8.4.
In Figure 8.4, there are no unsaturated vertices in Y , so U = ;.
Proof of Theorem 8.3.1: Let M be a maximum matching of G. Then from
Lemma 8.1.1 and part (e) of Lemma 8.3.2, U must be the empty set, so |U | = 0.
Therefore, from parts (b) and (d ) of Lemma 8.3.2, C = Y ∪ (A \ X ) is a cover of G
with |C | = |M |, and the result follows immediately from Lemma 8.2.2.
Notice in the example of Figure 8.4 that there is no unsaturated vertex in
Y . The construction of the proof of König’s Theorem then gives the cover C =
Y ∪ (A \ X ) = {a, c, e, i , l }. It has size 5, and therefore shows that the matching
shown there is maximum.
Problem 8.3.3. Let G be a bipartite graph with bipartition A, B , where |A| =
|B | = n. Prove that if G has q edges, then G has a matching of size at least q/n.
Solution: Notice that what is needed, is to prove that the maximum size of
a matching is at least q/n. By König’s Theorem, it is enough to show that the
minimum size of a cover is at least q/n. Suppose that C is a cover of G. There
can be at most n edges incident to any element of C , so there can be at most
n|C | edges incident with one or more elements of C . But C is a cover, so every
edge must be incident with one or more elements of C . Therefore, n|C | ≥ q,
or |C | ≥ q/n. Since every cover contains at least q/n vertices, therefore, the
minimum size of a cover is at least q/n, and we are done.
An algorithm for maximum matching in bipartite graphs

The X Y -construction and Lemma 8.3.2 essentially provides an algorithm to
find a maximum matching:
Step 1. Begin with any matching M .
Step 2. Construct X and Y .
Step 3. If there is an unsaturated vertex v in Y , find an augmenting path P (v)
ending at v, use it to construct a larger matching M 0 , and replace M by M 0 .
Then go to Step 2.
Step 4. If every vertex in Y is saturated, stop. M is a maximum matching, and
C = Y ∪ (A \ X ) is a cover of minimum size.
Now we give a more explicit way to construct X and Y and, at the same time,
the alternating paths P (v) that define them. It is very much like breadth-first
search—we find the elements of X , Y in levels. We find first the vertices that
are reachable by alternating paths beginning in X 0 and having 0 edges (this is
just X 0 , of course), then those that are reachable by alternating paths having 1
edge, then those reachable by alternating paths having 2 edges, et cetera. The
vertices v in the next level are exactly those (whose level has not yet been as-
signed) that are joined by an edge to a vertex u in the previous level. The only
additional rule is that, when we are creating an even level, that edge {u, v} must
be a matching edge. Just as in breadth-first search, when a vertex v is recog-
nized as an element of X or of Y , we assign it a parent pr(v), which is the vertex
u in the previous level from which we reached v. When we find an unsaturated
vertex v in Y , then we can trace the path vpr(v)pr2 (v) . . . w, where w ∈ X 0 , and
this is an augmenting path. There is no need to construct the rest of X and Y in
this case—we can immediately use the augmenting path to get a bigger match-
ing. Here is a statement of the resulting algorithm. In it, we use X̂ to represent
the set of elements of X that we have found so far, and similarly for Ŷ . Initially,
X̂ = X 0 and Ŷ = ;.
Bipartite matching algorithm
Step 0. Let M be any matching of G.
Step 1. Set X̂ = {v ∈ A : v is unsaturated }, set Ŷ = ;, and set pr(v) to be unde-
fined for all v ∈ V (G).
Step 2. For each vertex v ∈ B \ Ŷ such that there is an edge {u, v} with u ∈ X̂ , add
v to Ŷ and set pr(v) = u.
Step 3. If Step 2 added no vertex to Ŷ , return the maximum matching M and
the minimum cover C = Ŷ ∪ (A \ X̂ ), and stop.
Step 4. If Step 2 added an unsaturated vertex v to Ŷ , use pr values to trace an

augmenting path from v to an unsaturated element of X̂ , use the path to pro-
duce a larger matching M 0 , replace M by M 0 , and go to Step 1.
Step 5. For each vertex v ∈ A \ X̂ such that there is an edge {u, v} ∈ M with u ∈ Ŷ ,
add v to X̂ and set pr(v) = u. Go to Step 2.
Here is an example of the application of the algorithm. Consider the graph
of Figure 8.5 with the matching M indicated there. Take A = {1, 2, 3, 4}.
1 2 3 4
5 6 7 8
Figure 8.5: Example of the algorithm
Initially, X̂ = {1, 4} and Ŷ = ;. Then 7 is added to Ŷ with pr(7) = 1. Then 3

is added to X̂ with pr(3) = 7. Then 5 and 8 are added to Ŷ with pr(5) =pr(8) = 3.
Since 5 (for example) is unsaturated we have the augmenting path 5371. This
gives the new matching M = {{2, 6}, {3, 5}, {1, 7}}.
Beginning again (with no parent values), we have X̂ = {4}, Ŷ = ;. Then 7 is
added to Ŷ with pr(7) = 4. Then 1 is added to X̂ with pr(1) = 7. Now nothing
can be added to Ŷ and so the algorithm terminates with the current matching
M and the cover C = (A \ X̂ ) ∪ Ŷ = {2, 3, 7}.
Problem Set 8.3

1. Let G be a bipartite graph, and let ∆ be the largest degree of any vertex of G.
Prove that G has a matching of size at least q/∆. Also, show that this is false
in general if G is not bipartite.
2. Let n be a positive integer. Construct a bipartite graph with bipartition A, B ,

where |A| = |B | = n, for which the size of a maximum matching is less than
(q + 1)/n. (In other words, show that the value q/n is the best possible in
Problem 8.3.3.)
3. (Difficult) Let G be a bipartite graph with bipartition A, B , where |A| = |B | =

n, and suppose that every vertex of G has degree at least δ < n. Prove that G
has a matching of size at least the minimum of n and (q − δ2 )/(n − δ).
4. Construct a bipartite graph with bipartition A, B , where |A| = |B | = 8, and

having minimum degree 2, for which the size of a maximum matching is
less than (q −3)/6. (This shows that the value (q −δ2 )/(n −δ) in the previous
exercise, cannot be increased.)
5. Find a maximum matching and a minimum cover in the graph of Figure 8.6,
by applying the algorithm, beginning with the matching indicated.
A 1 2 3 4 5
B a b c d e
Figure 8.6: Matching exercise
by applying the algorithm, beginning with the matching indicated.
by applying the algorithm, beginning with the matching of size 18 consisting
of all the edges oriented from northwest to southeast.
8. Let G be bipartite with bipartition A, B . Suppose that C and C 0 are both cov-
ers of G. Prove that Ĉ = (A ∩C ∩C 0 ) ∪ (B ∩ (C ∪C 0 )) is also a cover of G.
9. In the previous exercise, prove that if C and C 0 are minimum covers, then so
is Ĉ .

8.4. APPLICATIONS OF KÖNIG’S THEOREM 197
8.4 Applications of König’s Theorem

If A, B is a bipartition of G, no matching of G can have size bigger than |A|. It
is interesting to determine whether there is one of exactly this size. (A perfect
matching has this property, but if |A| < |B |, there will be no perfect matchings.)
This problem was raised in a different context by Hall, and the resulting theo-
rem (although it follows from König’s Theorem) is called Hall’s Theorem.
Consider the subset D of A. If its elements are to be saturated by a match-
ing, there must be at least |D| distinct elements of B that are adjacent in G to at
least one element in D, since the matching edges incident to the elements of D
must have distinct ends. For example, in Figure 8.4, the subset D = {a, b, d , f } of
A is joined by edges of the graph only to the three vertices h, i , l of B . It follows
that there can be no matching saturating all of D, and therefore there can be
no matching saturating all of A. Let us define, for any subset D of vertices of a
graph G the neighbour set N (D) of D to be {v ∈ V (G) : there exists u ∈ D with {u, v} ∈
E (G)}. Then if there is a matching saturating A, we must have at least |D| ele-
ments in N (D). Hall proved that, if every subset D satisfies this condition, then
there will exist such a matching. As we indicated above, we can prove this using
König’s Theorem.
Theorem 8.4.1. (Hall’s Theorem) A bipartite graph G with bipartition A, B has a

matching saturating every vertex in A, if and only if every subset D of A satisfies
|N (D)| ≥ |D|.
Proof: First, suppose that G has a matching M saturating every vertex in A.

Then for any subset D of A, N (D) contains the other end of the edge of M inci-
dent with v for every v ∈ D, and these vertices must all be distinct, so |N (D)| ≥
|D|.
We prove the contrapositive of the “if” part of Hall’s Theorem; namely, if
there is no matching that saturates every vertex of A, then |N (D)| < |D| for some
subset D ⊆ A.
By hypothesis, there is no matching that saturates every vertex of A. Then,
by König’s Theorem, there exist a maximum matching M and a minimum cov-
ering C such that |C | = |M | < |A|. The sets A, B, C partition the vertices of G
into 4 subsets A ∩C , A \C , B ∩C , B \C as shown in Figure 8.9. (Remark: some of
these subsets may be empty; for example, if C = B , then B \C = A∩C = ;.) Since
C is a cover, no edge joins a vertex of B \C to a vertex of A \C , so N (A \C ) ⊆ B ∩C .
Let D = A \ C . Then
|N (D)| ≤ |B ∩C | = |C | − |A ∩C | < |A| − |A ∩C | = |A \C | = |D|.
This completes the proof of the “if” part of Hall’s Theorem
A A\C C ∩A
B C ∩B B\C
Figure 8.9: Proof of Hall’s Theorem
Note that, if there is no matching saturating all vertices in A, the maximum

matching algorithm will find a set D ⊆ A such that |N (D)| < |D|. Namely, at
termination of the algorithm, we take D = A\C , where C is the cover Y ∪ (A\X ).
Therefore D = X and N (D) = Y .
8.5 Systems of Distinct Representatives

The problem that led Hall to prove Theorem 8.4.1 can be described as follows.
Suppose there are several interest groups in some population, for example, bridge
players, socialists, football fans, and so on. We want to find a representative for
each of these groups, perhaps to serve on some decision-making body. There
are two rules, first, that a representative must belong to the group she repre-
sents, and second, that no one can represent two different groups. More for-
mally, given a collection Q 1 ,Q 2 , . . . ,Q n of subsets of a finite set Q, a system
of distinct representatives (SDR) for the collection is a sequence (or n-tuple)
(q 1 , q 2 , . . . , q n ) of n distinct elements of Q such that
q i ∈ Q i for i = 1, 2, . . . , n.
8.5. SYSTEMS OF DISTINCT REPRESENTATIVES 199
For example, suppose that Q 1 = {b, d }, Q 2 = {a, b, c, d }, Q 3 = {b, c}, Q 4 = {b, d }.

Then we can choose q 1 = b, q 2 = a, q 3 = c, q 4 = d . As a second example, sup-
pose that Q 1 = {b, d }, Q 2 = {a, b, c, d }, Q 3 = {b}, Q 4 = {b, d }. Then it is not hard
to convince ourselves that there is no SDR. One way to do so, is to observe that
the three subsets Q 1 ,Q 3 ,Q 4 have only 2 elements (b and d ) among them, and
yet an SDR would have to have 3 different elements from those 3 subsets. More
generally, if we have a subcollection of the given collection of subsets, that have
among them fewer elements than the number of subsets of the subcollection,
then there cannot exist an SDR. Hall proved (see Corollary 8.5.1 below) that this
is the only thing that can go wrong, that is, that if no such nasty subcollection
exists, then there does exist an SDR.
What do SDR’s have to do with matching? Given the collection Q 1 ,Q 2 , . . . ,Q n
of subsets of Q, we can construct the bipartite graph with bipartition A = {1, 2, . . . , n},
B = Q having the following adjacencies: vertices k ∈ A and b ∈ B are adjacent
if and only if b ∈ Q k . (For example, the graph associated with the first example
above is shown in Figure 8.10.) Now, corresponding to each SDR (q 1 , q 2 , . . . , q n ),
the edges {1, q 1 }, {2, q 2 }, . . . , {n, q n } in G make up a matching saturating A. Con-
versely, for each matching of G saturating A, say M = {{1, b 1 }, {2, b 2 }, . . . , {n, b n }},
the sequence (b 1 , b 2 , . . . , b n ) is an SDR of Q 1 ,Q 2 , . . . ,Q n . In summary, the collec-
tion Q 1 ,Q 2 , . . . ,Q n has an SDR if and only if graph G has a matching saturating
every vertex in A.
1 2 3 4
a b c d
Figure 8.10: Graph constructed from first SDR example
Moreover, the condition described above for an SDR to exist, corresponds

exactly to the requirement that in the graph, every subset D of A have |N (D)| ≥
|D|. Thus the following theorem is nothing but a restatement of Theorem 8.4.1
in the language of SDR’s.
Corollary 8.5.1. (Hall’s SDR Theorem) The collection Q 1 ,Q 2 , . . . ,Q n of subsets

of the finite set Q has an SDR if and only if, for every subset J of {1, 2, . . . , n}, we
have ¯[ ¯
¯ Q i ¯ ≥ |J |.
¯ ¯
i ∈J
8.6 Perfect Matchings in Bipartite Graphs

We can use Hall’s Theorem to obtain a condition for a bipartite graph to have a
perfect matching.
Corollary 8.6.1. A bipartite graph G with bipartition A, B has a perfect matching

if and only if |A| = |B | and every subset D of A satisfies
|N (D)| ≥ |D|.
Proof: Clearly, if |A| 6= |B |, then G has no perfect matching. On the other hand,
if |A| = |B |, then G has a perfect matching if and only if it has a matching satu-
rating every vertex in A, and then the result follows from Hall’s Theorem.
Another application is to regular bipartite graphs. We can show that these
always have perfect matchings, since it is easy to show that they always satisfy
the condition of Corollary 8.6.1. This result will be used in the next section,
when we discuss edge-colouring.
Theorem 8.6.2. If G is a k-regular bipartite graph with k ≥ 1, then G has a per-

fect matching.
Proof: Let A, B be a bipartition of G. Then since every edge has one end in A
P P
and the other in B , we have v∈A d eg (v) = v∈B d eg (v). It follows that k|A| =
k|B |, and therefore, since k > 0, that |A| = |B |. Now let D ⊆ A. Then since every
edge incident with a vertex in D has its other end in N (D), we have
X X
d eg (v) ≤ d eg (v).
v∈D v∈N (D)
It follows that k|D| ≤ k|N (D)|, and therefore (again, since k > 0) that |N (D)| ≥
|D|. Now by Corollary 8.6.1, G has a perfect matching.
Note: Theorem 8.6.2 works even if G contains multiple edges.
8.6. PERFECT MATCHINGS IN BIPARTITE GRAPHS 201
Problem Set 8.6

1. For a bipartition A, B of the graph of Problem Set 8.3, #6, find a set D ⊆ A
such that |N (D)| < |D|.
2. In the graph of Problem Set 8.3, #7, find a set D of vertices such that
|N (D)| < |D|.
3. Let G be a bipartite graph with bipartition A, B , let M be a matching of G,

and let D ⊆ A. Prove that |M | ≤ |A| − |D| + |N (D)|.
4. Let G be a bipartite graph with bipartition A, B . Prove that the maximum

size of a matching of G is the minimum, over subsets D of A, of |A|−|D|+
|N (D)|.
5. Let G be a bipartite graph with bipartition A, B such that |A| = |B |, and for
every proper nonempty subset D of A, we have |N (D)| > |D|. Prove that
for every edge e ∈ E (G) there is a perfect matching containing e.
6. Check that the proof of Theorem 8.6.2 works even for bipartite multi-
graphs. (Note that a bipartite multigraph can have multiple edges, but
is not allowed to have loops.)
7. A deck of playing cards is arranged in a rectangular array of four rows

and thirteen columns. Prove that there exist thirteen cards, no two in the
same column and no two of the same value.
8. Find a 3-regular graph having no perfect matching. (Such a graph must

be nonbipartite.)
9. Show how to find in a bipartite graph, a largest set of mutually nonadja-

cent vertices.
10. Let N be a matrix. Let us define the size of a submatrix N 0 of N to be the

number of rows of N 0 plus the number of columns of N 0 . How could one
find a maximum size submatrix of N with the property that each of its
entries is 0?
11. Prove Theorem 8.6.2 directly from König’s Theorem.
12. Let G be a bipartite graph with bipartition A, B where |A| = |B | = 2n. Sup-
pose that, for every subset X ⊆ A and every subset X ⊆ B such that |X | ≤ n,
|N (X )| ≥ |X |. Prove that G has a perfect matching.
8.7 Edge-colouring
An edge k-colouring of a graph G is an assignment of one of a set of k colours
to each edge of G so that two edges incident with the same vertex are assigned
different colours. Consider the set of edges M that are assigned a particular
colour. Then for any vertex v, there can be at most one edge of M incident
with v. That is, M is a matching, and so an edge k-colouring of G amounts to
a partitioning of the edges of G into k matchings. Figure 8.11 shows an edge
3-colouring of a graph, where the colours are indicated by different ways of
drawing the edges.
Figure 8.11: A graph with an edge 3-colouring
Obviously, we need at least d eg (v) different colours at each vertex v. There-

fore, for G to have an edge k-colouring, we must have k ≥ ∆, where ∆ denotes
the maximum degree of vertices of G. The main result of this section is that for
bipartite graphs this bound can be achieved.
Theorem 8.7.1. A bipartite graph with maximum degree ∆ has an edge ∆-colouring.
Lemma 8.7.2. Let G be a bipartite graph having at least one edge. Then G has a
matching saturating each vertex of maximum degree.
(Note that, in the special case of regular bipartite graphs, all of the vertices
have maximum degree, and the lemma says that there must be a perfect match-
ing. In fact, we already have seen this special case—it is Theorem 8.6.2.)
8.7. EDGE-COLOURING 203
Proof: Let A, B be a bipartition of G and K = {v ∈ V : d eg (v) = ∆}. Let M be a

maximum matching that leaves as few elements of K unsaturated as possible.
If M saturates all elements of K , we are finished. Otherwise, we may assume,
by interchanging A with B if necessary, that there is a vertex in A ∩ K that is not
saturated by M . Apply the X Y -construction, except that we define X 0 to con-
sist only of the unsaturated vertices in A ∩ K . Then Y contains no unsaturated
vertex, since M is a maximum matching.
Suppose there is a vertex w ∈ X having degree less than ∆. Consider the
alternating path P (w). It has even length. If we replace the edges of M that are
in P (w) by those that are not, we get another maximum matching, but one that
leaves fewer elements of K unsaturated, a contradiction to the choice of M . So
every vertex in X has degree ∆.
By the construction and the fact that M is maximum, for every u ∈ Y there
is an edge of M joining u to some vertex in X . Since X contains at least one
unsaturated vertex, |X | > |Y |. Moreover, again by the construction, there is no
edge from X to B \ Y , so N (X ) ⊆ Y . Therefore,
∆|Y | < ∆|X | = d eg (v) ≤ ∆|Y |,

X X
d eg (v) ≤
v∈X v∈Y
a contradiction. Therefore, there are no vertices in K not saturated by M , and

we are done.
Proof of Theorem 8.7.1: The result can be proved by induction on ∆. First, if

∆ = 0, then G has no edges, and has an edge 0-colouring. Now suppose m ≥ 1
and assume that every bipartite graph having ∆ < m has an edge ∆-colouring.
Let G be a bipartite graph with ∆ = m. By Lemma 8.7.2, G has a matching M
saturating all the vertices of degree m. Delete M from G to obtain the graph H .
Then H is bipartite and has ∆ = m−1, and so by the induction hypothesis, it has
an edge (m −1)-colouring. If we use this colouring on G, and use one additional
colour for the edges of M , then we have an edge m-colouring of G. The result is
proved by induction.
Note that the proof of Lemma 8.7.2 suggests a modification of the matching
algorithm which will efficiently find a matching saturating all vertices of max-
imum degree. Therefore, since the proof of Theorem 8.7.1 just requires us to
find and delete such a matching and repeat, there is an efficient algorithm for
finding an edge ∆-colouring of a given bipartite graph.
8.8 An Application to Timetabling

We have sets of instructors and courses, and for each instructor, the list of courses
she teaches. We need to schedule the courses into as few timetable slots as pos-
sible, so that no instructor is required to be in two places at the same time. We
have the additional requirement that two sections of the same course cannot be
given in the same slot. (This may be because we want courses to be available at
several different times.)
Let us construct the bipartite graph G having a vertex for each instructor
and a vertex for each course, with an edge joining a course vertex to an instructor-
vertex if that instructor teaches that course. (If an instructor teaches more than
one section of the same course, this would be a multigraph.) The set of edges
of the graph corresponds to all of the classes that must be scheduled. Now sup-
pose that we have arranged a schedule, and consider the set of classes taught
in a particular slot. These cannot involve any instructor more than once, nor
can they involve the same course more than once. That is, they correspond to a
matching of G. If we take the slots to correspond to colours, then our problem
is one of colouring the edges of G with the minimum number of colours.
Since G is bipartite, we know from Theorem 8.7.1 that the minimum num-
ber of slots needed is the maximum degree of the graph G constructed from
the data. This is the largest number of sections of a single course, or the largest
number of classes taught by a single instructor, whichever is larger. We also
know how to solve the problem algorithmically, and actually find a schedule
that achieves this minimum.
But suppose now that there is also a classroom limitation, so that at most
m classes can be assigned to the same slot. Then of course we cannot find a
schedule having fewer than q/m slots, where q is the total number of classes to
be scheduled (and is the number of edges of the graph G above). Our problem
now becomes:
Bounded edge-colouring problem: What is the smallest number of colours

needed to edge-colour a bipartite graph G, given that no colour can be assigned
to more than m edges?
The answer turns out to be the smallest integer that is at least q/m and also
is at least ∆. This follows from the next result.
Theorem 8.8.1. Let G be a graph with q edges, and suppose k, m are positive
integers such that
8.8. AN APPLICATION TO TIMETABLING 205
(a) G has an edge k-colouring;
(b) q ≤ km.
Then G has an edge k-colouring in which every colour is used at most m times.
Proof: Suppose that the colouring does not already have the desired property,
so that some colour, say red, is used at least m + 1 times. If every other colour is
used at least m times, then
q ≥ m + 1 + m(k − 1) > km,
a contradiction. So there exists a second colour, say blue, that is used at most
m − 1 times.
Now the red edges and the blue edges form disjoint matchings M and M 0 of
G. Consider the spanning subgraph H of G having edgeset M ∪ M 0 . Then every
vertex of this graph has degree 0 or 1 or 2. Therefore, each component of H
consists of a path or a cycle with edges alternately in M and M 0 . Moreover, any
cycle is even. Since |M | > |M 0 |, there must exist a component of H containing
more edges of M than of M 0 . Such a component must consist of an augmenting
path for M 0 . If we use that path to make M 0 larger by one and to make M smaller
by one, then we have a new edge k-colouring of G such that red is used fewer
times, and blue is still used at most m times. We can repeat this argument until
we find a colouring with the desired property.
Now we can solve the bounded edge-colouring problem for bipartite graphs by
combining Theorems 8.7.1 and 8.8.1.
Corollary 8.8.2. In a bipartite graph G, there is an edge k-colouring in which

each colour is used at most m times if and only if
(a) ∆ ≤ k, and
(b) q ≤ km.
Problem Set 8.8

1. Show that the number of colours required to edge-colour a nonbipartite
graph G can exceed ∆.
2. Prove that every subgraph of an edge k-colourable graph is edge k-colourable.

3. Prove that a k-regular bipartite graph has an edge k-colouring, using Theo-
rem 8.6.2.
4. Show that the Petersen graph cannot be edge 3-coloured.
5. Show that Theorem 8.8.1 does not work for vertex-colouring. That is, show
that if a graph can be k-coloured and mk ≥ p, it is not necessarily true that
there is a k-colouring in which each colour is used at most m times.
6. An n by n permutation matrix is a matrix having one 1 and n − 1 0’s in every

row and in every column. Let N be an n by n matrix such that every row
and every column contains k 1’s and n − k 0’s. Prove that N is the sum of k
permutation matrices.
7. State an algorithm for finding in a bipartite graph a matching saturating all

vertices of maximum degree.
Index
k-regular, 101 codewords, 153

n-cube, 102 coefficient, 16
colourable, 177
active vertex, 134 colouring, 177
adjacency list, 108 complement, 105
adjacency matrix, 107 complete, 101
adjacent, 93 complete graph, 101
algorithm for maximum matching, 193 component, 116
alternating path, 186 composition, 20, 29
ancestor, 133 concatenating, 41
asymptotic, 72 connected, 116
augmenting path, 187 constructive, 139
automorphism, 100 cover, 187
average number of occurrences, 56 cut, 117
cut-edge, 121
binary, 153 cycle, 111
binomial coefficient, 78 cycle space, 149
bipartite matching algorithm, 193 cyclomatic, 149
bivariate generating series, 54
bivariate weight function, 54 degree, 100, 158
blocks of a string, 41 diagonal, 114
boundary, 158 diameter, 140
boundary walk, 158 digraph, 96
bounded edge-colouring problem, 204 directed graph, 96
breadth-first search, 134 distance, 140, 153
bridge, 121 drawing, 93
dual, 181
cartesian product, 24
child, 133 edge k-colouring, 202
closed, 110 edge subdivision, 172
code, 153 Edge-colouring, 202
207
208 INDEX
edge-contraction, 178 level, 133

edges, 93 line-graph, 104
empty string, 40
end-vertices, 127 matching, 185
error correcting code, 153 maximum matching, 185
Euler’s Formula, 159 modulo 2 sum, 145
Eulerian circuit, 119 multipleedge, 98
even graphical code, 154 neighbour, 94
even parity, 154 neighbour set, 197
exhausted vertices, 134 non-tree, 148
Nonhomogeneous Recurrence Equations,
faces, 158
70
Factor Theorem, 66
formal power series, 16 odd graph, 103
Four colour theorem, 180
parent, 132
generating series, 12 path, 110, 111
girth, 112, 170 perfect matching, 185
graph, 93 Petersen graph, 99
planar, 94
Hall’s SDR Theorem, 200
planar embedding, 157
Hall’s Theorem, 197
planar graph, 157
Hamilton cycle, 113
planar map, 157
Handshaking lemma, 101
predecessor, 132
homogeneous equation, 65
recursive decomposition, 47
incidence matrix, 107
recursive definition, 47
incident, 93
regular, 101
inverse, 18
isomorphism class, 99 saturated, 185
saturates, 185
Job assignment problem, 185
SDR, 198
join, 93
search tree, 133
König’s Theorem, 190 spanning subgraph, 110
Kuratowski’s Theorem, 172 spanning tree, 128
stereographic projection, 160
leaves, 127 subgraph, 110
length of a string, 40 substring, 41
length of a walk, 110 symmetric difference, 145
INDEX 209
system of distinct representatives, 198
Terquem Problem, 38
tree, 125
Unambiguous expression, 43
unique, 153
vertices, 93
walk, 110
XY-construction, 190

Introduction To Combinatorics, Course Notes PDF

Uploaded by

Introduction To Combinatorics, Course Notes PDF

Uploaded by

Introduction to Combinatorics

Course Notes for Math 239

Department of Combinatorics and Optimization

©August 18, 2016

2 Compositions and Strings 29

3 Recurrences, Binary Trees and Sorting 63

3.7 The Binomial Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

4 Introduction to Graph Theory 93

7 Planar Graphs 157

(b) How many compositions of n are there with exactly k parts?

1.2 Sums and Products

If A and B are disjoint, that is, A ∩ B = ; then

Here |S| denotes the number of elements in the set S.

and k = 3, then the elements of A 3 correspond to the 63 = 216 possible out-

1.3 Binomial Coefficients

Now consider a different way of generating the elements of L . We first

subsets of {1, . . . , n}. So by Theorem 1.3.1,

Note that nk = 0 whenever n < k as it should be. It is natural to define the

The values of each row represent the values of n0 , n1 , . . . , nn in order, where

n = 0 in the first row, n = 1 in the second row, etc.

1.4 Bijections (One-to-One Correspondence)

• f is 1-1 or injective if for any x 1 , x 2 ∈ S, f (x 1 ) = f (x 2 ) implies x 1 = x 2 . In

• f is onto or surjective if for all y ∈ T , there exists x ∈ S such that f (x) =

• f is a 1-1 correspondence or bijection if it is both 1-1 and onto.

We can visualize these definitions using the following diagram:

1-1 onto bijection

Theorem 1.4.1. If a function f : S → T has an inverse, then f is a bijection.

Proof: Let f −1 : T → S be an inverse of f . We need to prove that f is 1-1 and

Solution: We can define the following mapping:

First, we need to check that each element of S is being mapped to an element

f −1 ( f (A)) = f −1 ({1, . . . , n} \ A) = {1, . . . , n} \ ({1, . . . , n} \ A) = A.

And for each B ∈ T ,

f ( f −1 (B )) = f ({1, . . . , n} \ B ) = {1, . . . , n} \ ({1, . . . , n} \ B ) = B.

So f −1 is indeed the inverse of f , which proves that f is a bijection.

By Theorem 1.3.1, |S| = nk¡ and

We can establish a bijection between two completely different looking sets

Solution: We define f : S → T in the following way: For a subset A of {1, . . . , n},

It is clear that f (A) is a binary string of length n, hence in T .

You can check that f −1 ( f (A)) = A and f ( f −1 (t )) = t , so f is a bijection.

This bijection tells us that the number of subsets of {1, . . . , n} is equal to

We illustrate this bijection for n = 3.

1.5 Combinatorial Proofs

Theorem 1.5.1. For any non–negative integer n,

(1 + x)3 = (1 + x)(1 + x)(1 + x)

We now give a formal proof of the theorem.

Proof: We first expand (1 + x)n algebraically:

Sometimes an algebraic identity can be established by giving a set of objects,

Problem 1.5.2. For integers 1 ≤ k ≤ n, show that

Problem 1.5.3. For non–negative integers n and k, show that

Now, if A ∈ S i then n+i ∈ A and A\{n+i } is an (n−1)–element subset of {1, . . . , n+

Problem Set 1.5

2. Consider the identity

(a) Give a combinatorial proof of this identity. (Hint: if S is a subset of

4. Show that ni=0 ni i (i − 1) = n(n − 1)2n−2 .

5. (a) Consider the following two sets.

Sk = {(A, B ) | A, B ⊆ {1, . . . , n}, |A| = |B | = m, |A ∩ B | = k},

Define a bijection f : S k → Tk , explain why f (A, B ) ∈ Tk for any (A, B ) ∈

1.6 Generating Series

Definition 1.6.2. Let S be a set of configurations with a weight function w. The

By collecting like–powers of x in ΦS (x), we get

So, there is 1 configuration of weight 3, 2 configurations of weight 5, 3 configu-

ΦN≥k (x) = x k + x k+1 + x k+2 + · · ·

(c) the average weight of an element in S is Φ0S (1)/ΦS (1).

Φ0S (x) = w(σ)x w(σ)−1 .

Φ0S (x) = n(1 + x)n−1 ,

so the sum of the sizes of the subsets of {1, . . . , n} is

Φ0S (1) = n2n−1 .

Therefore, the average size of a subset of {1, . . . , n} is

Φ0S (1)/ΦS (1) = n2n−1 /2n = n/2,

Problem Set 1.6

5. Find the generating series ΦS (x) if:

(a) S = {1, 2, . . .} and w(i ) = i for all i ;

8. A 3 cm by 3 cm piece of cardboard is marked on one side by ruled lines

1.7 Formal Power Series

A(x) = a 0 + a 1 x + a 2 x 2 + · · · where a n is the number of {0, 1}-strings of length

Solution: We can rewrite Equation 1.7.1 as: