MATH2069

Discrete Mathematics and

Graph Theory

Anthony Henderson
c School of Mathematics and Statistics,

The University of Sydney, 2008–2020
 Part I

Topics in Discrete Mathematics

Anthony Henderson
Acknowledgements

These lecture notes were written in 2008 (and revised in 2009) for the units
MATH2069 Discrete Mathematics and Graph Theory and MATH2969 Dis-
crete Mathematics and Graph Theory (Advanced), given at the University
of Sydney. I am extremely grateful to David Easdown for allowing me to
make use of his lecture notes on these topics, and for his careful correction
of, and comments on, these notes. Any errors which remain are entirely my
own fault.1

I also acknowledge a debt to the following textbooks, which students should

consult for more examples and exercises.

• Introduction to Discrete Mathematics by K.-G. Choo and D. E. Taylor.

• Combinatorics: An Introduction by K. H. Wehrhahn.

• The Art of Computer Programming Vol. 1 by D. E. Knuth.

• Enumerative Combinatorics Vol. 1 by R. P. Stanley.

Anthony Henderson

1
A few minor changes were made by Alexander Molev.

ii
Contents

0 Introduction 1

1 Counting Problems 7

1.1 Fundamental principles . . . . . . . . . . . . . . . . . . . . . . 7

1.2 Ordered selection . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.3 Binomial coefficients . . . . . . . . . . . . . . . . . . . . . . . 23

1.4 Inclusion/Exclusion . . . . . . . . . . . . . . . . . . . . . . . . 34

1.5 Stirling numbers . . . . . . . . . . . . . . . . . . . . . . . . . 40

2 Recursion and Induction 45

2.1 Examples of recursive sequences . . . . . . . . . . . . . . . . . 45

2.2 Proof by induction . . . . . . . . . . . . . . . . . . . . . . . . 51

2.3 Homogeneous linear recurrence relations . . . . . . . . . . . . 60

2.4 Non-homogeneous linear recurrence relations . . . . . . . . . . 68

iii
3 Generating Functions 75

3.1 Formal power series . . . . . . . . . . . . . . . . . . . . . . . . 76

3.2 Manipulating formal power series . . . . . . . . . . . . . . . . 81

3.3 Generating functions and recursion . . . . . . . . . . . . . . . 93

iv
Chapter 0

Introduction

What is discrete mathematics and why is it useful? It is best to begin by

considering an example of a problem in discrete mathematics.

The ‘tower of Hanoi’ is a popular mathematical toy (apparently invented

in France, despite being named after a Vietnamese building). It consists of
three pegs and a certain number of circular discs of differing sizes which sit
around the pegs. You can move one disc at a time from one peg to another,
subject to the rule that you can never place a larger disc on top of a smaller
one. The discs start off all being on peg 1, in increasing order of size from
top to bottom, and the aim is to move them all to peg 2. The mathematical
question is: what is the smallest number of moves needed to achieve this, as
a function of the number of discs?

We let n be the number of discs, and write hn for the smallest number of
moves required. If there are no discs, then there is nothing to do, so h0 = 0.
If there is one disc, then clearly one move suffices, so h1 = 1. If there are two
discs, then the rule that the larger disc cannot sit on top of the smaller one
forces you to put the smaller one temporarily on peg 3, so h2 = 3.

This illustrates a general feature: the largest disc has to move from peg 1 to
peg 2 at some stage, and while that happens all the other discs have to be
on peg 3. Clearly it takes at least hn−1 moves to get all the other discs to

1
2 Topics in Discrete Mathematics

peg 3, and then after moving the largest disc, it takes at least hn−1 moves to
get them back to peg 2. This suggests a recursive procedure for solving the
puzzle, which uses the smallest possible number of moves:

(1) Move all but the largest disc from peg 1 to peg 3 in the smallest possible
number of moves.
(2) Move the largest disc from peg 1 to peg 2.
(3) Move all the other discs from peg 3 to peg 2 in the smallest possible
number of moves.

This procedure is recursive because in steps (1) and (3) you need to go
through the whole three-step procedure, but for n − 1 discs rather than n
(the change of peg numbers is immaterial); and when you then examine each
of those, their own steps (1) and (3) call for the whole three-step procedure
for n − 2 discs, and so on. Returning to the number of moves required, we
have derived a recurrence relation
hn = 2hn−1 + 1,
which determines the answer for n in terms of the answer for n − 1. This
recurrence relation, together with the inital condition h0 = 0, determines hn
for all n. For example, continuing from where we left off,
h3 = 2h2 + 1 = 2 × 3 + 1 = 7,
h4 = 2h3 + 1 = 2 × 7 + 1 = 15,
h5 = 2h4 + 1 = 2 × 15 + 1 = 31.
You may notice that these values of hn are all one less than a power of 2. In
fact, the evidence we have so far supports the conjecture that hn = 2n − 1
always. To prove such a formula for a sequence defined recursively, the most
natural method is mathematical induction: we know that h0 = 20 − 1 is true,
and if we assume that hn−1 = 2n−1 − 1 is true, then we can deduce
hn = 2hn−1 + 1 = 2(2n−1 − 1) + 1 = 2n − 2 + 1 = 2n − 1.
By induction, hn = 2n −1 for all nonnegative integers n. So the answer to the
original question is that the smallest number of moves needed to complete
the n-disc tower of Hanoi puzzle is 2n − 1.
CHAPTER 0. INTRODUCTION 3

“Discrete” is the opposite of “continuous”, and discrete mathematics deals

with things which are composed of separate parts rather than flowing to-
gether. What made the tower of Hanoi problem discrete was that we were
trying to find a formula for something which was a function of a nonnegative
integer rather than a real variable. It would make no sense at all to ask about
the smallest number of moves needed to solve the puzzle for 4.673 discs; the
number of discs has to be a nonnegative integer.

So in this course we will almost never be dealing with functions f (x) of a

2
real variable x, like those in calculus (e.g. sin x, ex ). We will almost always
be discussing functions whose domain is the set of nonnegative integers:

N = {0, 1, 2, 3, 4, · · · }.

(In mathematics it is usually best to start counting at 0 if possible, even if

it means making some slightly contrived interpretations, such as declaring
that a tower of Hanoi with 0 discs is already solved, or that an Introduction
is Chapter 0.) It is perfectly acceptable to write such functions in the same
way as functions of a real variable – for instance, we could have written h(n)
instead of hn above – but we will tend to put the integer variable in the
subscript, as say an . This notation goes along with the idea of writing the
values of the function in a sequence:

a0 , a 1 , a 2 , a 3 , · · · .

There is in fact no logical difference between a function whose domain is N

and a sequence like this; just a slight psychological shift of viewpoint. Calling
the function a sequence is particularly natural in cases which are recursive
like the tower of Hanoi problem: every term of the sequence (past a certain
point, maybe) is determined by earlier terms, so you can imagine writing out
the sequence in succession, with what you’ve already written determining
what to write next.

In the tower of Hanoi example, the terms of the sequence were nonnegative
integers. This feature is not essential for a problem to be considered discrete.
For maximum flexibility, we will usually allow the terms of our sequences, i.e.
the values of our functions with domain N, to be complex numbers (though
examples with non-real values will be rare).
4 Topics in Discrete Mathematics

The usefulness of discrete mathematics comes from the fact that, even though
physics sometimes gives the impression that we live in a continuous universe,
there are countless phenomena in nature and society which are best under-
stood in terms of sequences (functions of a nonnegative integer variable)
rather than functions of a real variable.

Obvious examples are the regular additions to or subtractions from bank

accounts: if you want to calculate how much more interest you will earn if
it is calculated monthly rather than yearly, or what the monthly repayments
will be on a loan where interest is regularly calculated on the balance, you
are automatically dealing with quantities which depend on an integer number
of months (or years, or whatever) rather than anything continuous. And of
course, the world of computers is discrete by construction: if you want to work
out how long a program for constructing students’ university timetables will
take to run, all the variables are nonnegative integers – number of students,
number of classes, number of bytes in the memory, etc.

It is also important to see the analogies between discrete and continuous

problems, and these will be pointed out in the following chapters wherever
possible. For example, solving a recurrence relation (that is, finding a closed
formula for the terms of a sequence which is defined recursively) is strongly
analogous to solving a differential equation. Some authors speak of “differ-
ence equations” instead of “recurrence relations”, on the principle that the
difference an+1 − an is the discrete analogue of the derivative of a function in
calculus; but we will not explore this point of view.

There are some natural phenomena which could be modelled either discretely
or continuously. For example, if you wanted to predict the temperature in
Sydney tomorrow, you could think of temperature as a real-valued function
of three real variables (longitude, latitude, and time), and try to understand
what differential equations it satisfied. Or you could break up NSW into a
finite number of square kilometre blocks, and time into a discrete sequence
of days, and try to understand what recurrence relations hold: i.e. express
the temperature on a given day in a given block in terms of the temperatures
on the previous day in adjacent blocks. The discrete approach might well be
easier, in the sense of being better suited to the computer technology you
would need to use.
CHAPTER 0. INTRODUCTION 5

But it is certainly not the case that discrete mathematics is always easier
than continuous mathematics; we will encounter many challenging problems
which have no continuous analogues. For a first example, imagine a new,
improved tower of Hanoi puzzle which has four pegs rather than three. The
rules are the same as before; what is the smallest number of moves required
to move the discs from peg 1 to peg 2, if there are n discs? Nobody knows
the answer (at least, there is no known closed formula like 2n − 1).

Comments on the text

In these notes, the main results are all called “Theorem”, irrespective of their
difficulty or significance. Usually, the statement of the Theorem is followed
(perhaps after an intervening Example or Remark) by a rigorous Proof; as is
traditional, the end of each proof is marked by an open box.

To enable cross-references, the Theorems, Definitions, Examples, and Re-

marks are all numbered in sequence (the number always starts with the rel-
evant chapter number). Various equations and formulas are also numbered
(on their right-hand edge) in a separate sequence.

In definitions (either formally numbered or just in the main text), a word

or phrase being defined is underlined. The symbol “:=” is used to mean “is
defined to be” or “which is defined to be”.

The text in the Examples and Remarks is slanted, to make them stand out
on the page. Many students will understand the Theorems more easily by
studying the Examples which illustrate them than by studying their proofs.
Some of the Examples contain blank boxes for you to fill in. The Remarks
are often side-comments, and are usually less important to understand.

The more difficult Theorems, Proofs, Examples, and Remarks are marked at
the beginning with either * or **. Those marked * are at the level which
MATH2069 students will have to understand in order to be sure of getting
a Credit, or to have a chance of a Distinction or High Distinction. Those
marked ** are really intended only for the students enrolled in the Advanced
unit MATH2969, and can safely be ignored by those enrolled in MATH2069.
6 Topics in Discrete Mathematics
Chapter 1

Counting Problems

As we saw in the Introduction, the feature that characterizes discrete math-

ematics is the role played by the set of nonnegative integers:
N = {0, 1, 2, 3, 4, · · · }.
To state the obvious, the reason that N arises is that for every finite set X,
the number of elements of X is a nonnegative integer. For instance, in the
tower of Hanoi problem we had a finite set of discs, and n was the number
of them; this is why the function we were computing had domain N.

So the foundation of the whole subject is the idea of counting the number
of elements in finite sets. This can be much harder than it sounds, and in
its higher manifestations goes by the name of enumerative combinatorics.
In this chapter we will review and extend some of the basic principles of
counting.

1.1 Fundamental principles

If X is any finite set, we will write |X| for its size (also known as its
cardinality), i.e. the number of elements of X. For instance, if X = {0, 2, 4},
then |X| = 3. The most basic principle of all is:

7
8 Topics in Discrete Mathematics

Theorem 1.1 (Bijection Principle). If two sets X and Y are in bijection,

then |X| = |Y |.

A bijection, also known as a one-to-one correspondence, between X and Y

is a way of associating to each element of X a single element of Y , such that
every element of Y is associated to exactly one element of X. (“Exactly one”
means not zero, and not more than one.)
Example 1.2. If we assume that every person in the room has a single
head, and every head in the room is part of exactly one person, then the set
of people in the room is in bijection with the set of heads in the room. So
the number of people in the room equals the number of heads in the room.
Example 1.3. How many numbers between 1 and 999 are divisible by 3
(in the sense that 3 divides the number exactly, with zero remainder)? This
question is asking for the size of the set
X = {n ∈ N | 1 ≤ n ≤ 999, n is divisible by 3}
= {3, 6, 9, 12, · · · , 999}.
The answer is easy. The elements of X are exactly 3 × 1, 3 × 2, and so on
up to 3 × 333, so |X| = 333. But even this simple argument is really an
application of the Bijection Principle: we are implicitly observing that X
is in bijection with the set Y = {1, 2, · · · , 333}, so |X| = |Y | = 333. This
illustrates how the Bijection Principle is used. When it’s not obvious how to
count the elements of some set X, we try to translate the problem by finding
another set Y , whose size we do know, which is in bijection with X.
Definition 1.4. The formal definition of a bijection between X and Y is a
pair of functions, one from X to Y and one from Y to X, which are inverse
to each other. Recall that a function f : X → Y is said to be injective
(‘one-to-one’) if no function value is taken more than once, i.e. we never have
f (x1 ) = f (x2 ) for different elements x1 , x2 ∈ X; equivalently,
for every y ∈ Y there is at most one x ∈ X such that f (x) = y.
(The opposite of “injective” is “not injective”: to say that f : X → Y is not
injective is to say that there are some elements x1 , x2 ∈ X, with x1 6= x2 ,
such that f (x1 ) = f (x2 ).) A function f : X → Y is said to be surjective
(‘onto’) if every element of Y is one of the values taken by the function, i.e.
for every y ∈ Y there is at least one x ∈ X such that f (x) = y.
CHAPTER 1. COUNTING PROBLEMS 9

(The opposite of “surjective” is “not surjective”, which means something

completely different from “injective”.) A function f : X → Y is bijective if
it is both injective and surjective, i.e.
for every y ∈ Y there is exactly one x ∈ X such that f (x) = y.
It is only in this last situation that you can define an inverse function f −1 :
Y → X (which you do by sending y to the unique x such that f (x) = y),
and thus put X and Y in bijection.

The next most basic principle, again not requiring proof (since it is bound
up with such elementary matters as the definition of addition), is:
Theorem 1.5 (Sum Principle). If a finite set X is the disjoint union of two
subsets A and B, i.e.
X = A ∪ B and A ∩ B = ∅,
which means that every element of X is either in A or in B but not both,
then |X| = |A| + |B|. More generally, if X is the disjoint union of subsets
A1 , A2 , · · · , An , i.e.
X = A1 ∪ A2 ∪ · · · ∪ An and Ai ∩ Aj for all i 6= j,
then |X| = |A1 | + |A2 | + · · · + |An |.
Example 1.6. If every person in the room is either left-handed or right-
handed, and no-one claims to be both, then the total number of people
equals the number of left-handed people plus the number of right-handed
people. Of course this relies on the disjointness of the two subsets: if anyone
was ambidextrous and was counted in both the left-handed total and the
right-handed total, it would throw out the calculation (see the section on the
Inclusion/Exclusion Principle later in this chapter).
Example 1.7. How many numbers from 1 up to 999 are palindromic, in
the sense that they read the same backwards as forwards? As soon as you
start to think about this question, you have to break it up into the cases of
one-digit numbers, two-digit numbers, and three-digit numbers. So, whether
you choose to make it explicit or not, you are using the Sum Principle: in
order to count the set
X = {n ∈ N | 1 ≤ n ≤ 999, n is palindromic},
10 Topics in Discrete Mathematics

you are using the fact that it is the disjoint union of subsets A1 , A2 , A3 , where
A1 = {n ∈ N | 1 ≤ n ≤ 9, n is palindromic},
A2 = {n ∈ N | 10 ≤ n ≤ 99, n is palindromic},
A3 = {n ∈ N | 100 ≤ n ≤ 999, n is palindromic},

and then calculating |X| as the sum |A1 | + |A2 | + |A3 |. The calculation is
left for you to complete:

|A1 | = because

|A2 | = because

|A3 | = because

|X| =

A variant form of the Sum Principle involves the complement of a subset: if

A is a subset of X, its complement X \ A is defined to be {x ∈ X | x ∈
/ A},
the subset of X consisting of all elements which are not in A.
Theorem 1.8 (Difference Principle). For any subset A of a finite set X,

|X \ A| = |X| − |A|.

Proof. This is just the Sum Principle rearranged, because X is the disjoint
union of A and X \ A.
Example 1.9. How many three-digit numbers are divisible by 3? One way
to answer this is to see that the set of three-digit numbers divisible by 3 can
be written as X \ A, where

X = {n ∈ N | 1 ≤ n ≤ 999, n is divisible by 3},

A = {n ∈ N | 1 ≤ n ≤ 99, n is divisible by 3}.

As seen in Example 1.3, |X| = 333, and similarly |A| = 33. By the Difference
Principle, the answer to the question is |X \ A| = |X| − |A| = 300.
CHAPTER 1. COUNTING PROBLEMS 11

One situation in which a disjoint union naturally arises is if you have a

function f : X → Y . For any y ∈ Y , the preimage f −1 (y) is defined by

f −1 (y) := {x ∈ X | f (x) = y}, a subset of X.

Note that the condition for f to be injective is that |f −1 (y)| ≤ 1 for all y ∈ Y ,
whereas the condition for f to be surjective is that |f −1 (y)| ≥ 1 for all y ∈ Y .
Remark 1.10. If f is bijective, and therefore has an actual inverse function
f −1 : Y → X, we have a slight notational clash. If you interpret f −1 (y)
as the image of y ∈ Y under f −1 , then it is the unique x ∈ X such that
f (x) = y. If you interpret f −1 (y) as the preimage, then it is the set whose
single element is this x. In practice, the context always determines which of
the two is meant.

It is clear that X is the disjoint union of the subsets f −1 (y) as y runs over
Y , because for every x ∈ X there is a unique y ∈ Y such that f (x) = y (that
is what it means to have a well-defined function). So by the Sum Principle,
X
|X| = |f −1 (y)|. (1.1)
y∈Y

Example 1.11. Suppose X is the set of students at this university, Y is

the set of days in the year (including February 29), and f : X → Y is the
function which associates to every student their birthday. Then (1.1) is just
the obvious statement that the total number of students is the sum of the
number whose birthday is January 1, the number whose birthday is January
2, and so on. It is actually possible to deduce something from this: there
must be some day of the year which is the birthday of at least 126 students.
The reason is that the size of X is 46054 (at least it was last year), and the
size of Y is 366. If |f −1 (y)| ≤ 125 for all y ∈ Y , then the right-hand side
of (1.1) would be at most 366 × 125 = 45750, a contradiction. The magic
number 126 was obtained by dividing 46054 by 366 and then rounding up.
Definition 1.12. For any real number x, define
bxc : = x rounded down to the nearest integer
= largest integer m such that m ≤ x,
dxe : = x rounded up to the nearest integer
= smallest integer m such that m ≥ x.
12 Topics in Discrete Mathematics

Because we are dealing mainly with functions which only make sense for
integer variables, we will quite often have to carry out such roundings. We
will use without comment the obvious inequalities

x − 1 < bxc ≤ x, x ≤ dxe < x + 1. (1.2)

Theorem 1.13 (Pigeonhole Principle). If we have a function f : X → Y

between two finite (nonempty) sets, then there must be some y ∈ Y such
that  
−1 |X|
|f (y)| ≥ .
|Y |
In particular, if |X| > |Y | then there must be some x1 6= x2 in X such that
f (x1 ) = f (x2 ), i.e. f is not injective.
Example 1.14. The reason for the name is the imagined situation of pigeons
flying into pigeonholes: if there are 60 pigeons and only 25 holes, there must
be some pigeonhole that ends up with at least three pigeons in it. The “In
particular” part of the statement just says that if the number of pigeons is
more than the number of holes, there must be two pigeons which fly into
the same hole. (Some authors call this the Pigeonhole Principle and regard
Theorem 1.13 as a generalization.)

Proof. As in Example 1.11, the proof is by contradiction. We assume the

negation of what we are trying to prove, i.e. that
 
−1 |X|
|f (y)| < , for all y ∈ Y .
|Y |
|X|
Since |f −1 (y)| is an integer, this inequality implies that |f −1 (y)| < |Y |
. So
X X |X|
|f −1 (y)| < = |X|,
y∈Y y∈Y
|Y |

which contradicts (1.1). Therefore our assumptionl was m wrong, which means
−1 |X|
that there is some y ∈ Y for which |f (y)| ≥ |Y | . To deduce the “In
l m
particular” sentence, note that |X| > |Y | means that |Y | > 1, so |X|
|X|
|Y |
≥ 2.
So we have shown that there is some y ∈ Y for which there at least two
elements of X which are taken to y by f ; this gives the stated result.
CHAPTER 1. COUNTING PROBLEMS 13

Remark 1.15. Note that the Pigeonhole Principle asserts that something
is true for some y ∈ Y , but not for all y ∈ Y , and not for a particular y
specified in advance. For instance, in the example of students’ birthdays,
we cannot pick a particular date like June 16 and be sure that at least 126
students have that birthday. Indeed, it is possible that no students have that
birthday (though admittedly that’s unlikely).

The Pigeonhole Principle is surprisingly useful even in its basic form of prov-
ing non-injectivity.
Example 1.16. If there are 30 students in a tutorial, there must be two
whose surnames begin with the same letter, because there are only 26 possible
letters. This is an example of the Pigeonhole Principle: we are saying that
the function from the set of students to the set of letters given by taking
the first letter of the surname cannot be injective. Of course, we can’t pick
a particular letter like A and be sure that there are two students whose
surname begins with A: there might be none.
Example 1.17*. If m is any positive integer, the decimal expansion of 1/m
eventually gets into a repeating cycle: for instance,

1/5 = 0.20000000 · · · , 1/65 = 0.0153846153846153846 · · · .

We can prove this using the Pigeonhole Principle, as follows. Consider the
set X = {100 , 101 , · · · , 10m }. For each element of X, we can divide it by
m and find the remainder, which must be in the set Y = {0, 1, · · · , m − 1}.
This defines a function from X to Y . Since |X| = m + 1 and |Y | = m, this
function can’t be injective, so there are some nonnegative integers k1 < k2
such that 10k1 and 10k2 have the same remainder after division by m. This
k k
means that 10m1 and 10m2 are the same after the decimal point (i.e. their
k
difference is an integer). But the decimal expansion of 10m1 is obtained from
that of m1 by shifting every digit k1 places to the left, and similarly for k2 .
So for every k > k1 , the kth digit after the decimal point in m1 equals the
k
(k − k1 )th digit after the decimal point in 10m1 , which equals the (k − k1 )th
k
digit after the decimal point in 10m2 , which equals the (k + k2 − k1 )th digit
after the decimal point in m1 . So once you are sufficiently far to the right
of the decimal point, the digits in m1 repeat themselves every k2 − k1 digits
(maybe more often than that, but at least that often).
14 Topics in Discrete Mathematics

There is a related way of proving non-surjectivity:

Theorem 1.18 (Reverse Pigeonhole Principle). If |X| < |Y |, then any func-
tion f : X → Y is not surjective. That is, there must always be some y ∈ Y
such that f −1 (y) is empty.

Proof. The range of f can clearly have at most |X| elements, so it cannot
be the whole of Y .

There is an old joke that a combinatorialist is someone who, to find out how
many sheep there are in a flock, counts their legs and divides by 4. That
may not be a practical method in the case of sheep, but it is a surprisingly
useful trick in other sorts of counting problems.

Theorem 1.19 (Overcounting Principle). Let X and Y be finite sets, and m

a positive integer. Suppose there is a function f : X → Y with the property
that for all y ∈ Y , |f −1 (y)| = m (i.e. there are always m elements x ∈ X
such that f (x) = y). Then |Y | = |X|
m
.

P
Proof. By (1.1), we have |X| = y∈Y |f −1 (y)| = m|Y |.
Example 1.20. In the spurious sheep example, X is the set of legs, Y is the
set of sheep, m = 4, and f is the function which associates to each leg the
sheep it belongs to. The assumption (which we certainly do need, to apply
the method!) amounts to saying that every sheep has four legs.

Obviously, the cases where it is useful to apply this principle are those where
the set X is for some reason easier to count than the set Y (despite being
bigger). It is vital that all the preimages have the same size, so that when
we count X, we are overcounting Y by a known factor.
Example 1.21. How many edges does a cube have? If you don’t have pen
and paper handy, you can just observe that a cube has 6 faces, and every
face has 4 edges. Is the answer 6 × 4 = 24? No: this counts every edge twice,
since every edge is on the border of two faces. So the correct answer is 12,
and this argument is a case of the Overcounting Principle. To interpret it
in the above terms, Y is the set of edges, X is the set of pairs (face, edge)
where the edge belongs to the face, and f is the function which takes such a
CHAPTER 1. COUNTING PROBLEMS 15

pair and forgets the face. Here is a similar problem: the outside of a soccer
ball is made from 32 pieces of cloth, 12 of which are pentagonal (having five
edges) and 20 of which are hexagonal (having six edges). These are stitched
together along their edges, and at every vertex where an edge ends, three
faces meet. Using the Overcounting Principle, you can deduce that:

the number of edges is ,

the number of vertices is .

Our final fundamental principle relates to the most important way of con-
structing new sets from old.

Definition 1.22. If X and Y are sets, the Cartesian product X ×Y is defined

to be the set of all pairs (x, y) where the first entry is an element of X and
the second entry is an element of Y . In symbols:

X × Y := {(x, y) | x ∈ X, y ∈ Y }.

More generally, if we have n sets X1 , X2 , · · · , Xn , then their product is the

set of n-tuples where the ith entry is an element of Xi :

X1 × X2 × · · · × Xn := {(x1 , x2 , · · · , xn ) | x1 ∈ X1 , x2 ∈ X2 , · · · , xn ∈ Xn }.

If X1 = X2 = · · · = Xn = X, then X1 × X2 × · · · × Xn is simply written X n .

Example 1.23. Let X = {A, B, C, · · · , Z} be the set of letters in the alpha-

bet, and let Y = {0, 1, 2, · · · , 9}. Then X × X × X × Y × Y × Y = X 3 × Y 3
is the set of 6-tuples where the first three entries are letters and the last
three entries are single-digit numbers, such as (L, B, M, 5, 2, 3). Ignoring the
brackets and commas, this is the set of all strings of three letters followed by
three digits, such as LBM523: exactly the set of possible car numberplates,
under the old style. Asking how many possible numberplates there were
when this was the uniform style is the same as asking for the size of X 3 × Y 3 .
The answer is that there are 26 choices for the first letter, 26 choices for the
second letter, 26 choices for the third letter, and 10 choices for each of the
three digits, so a total of 263 × 103 = 17576000 possibilities.
16 Topics in Discrete Mathematics

Remark 1.24. Before we extract the last part of this argument as a general
principle, a pedantic comment. According to Definition 1.22, X 3 × Y 3 ought
to be a set of pairs where each entry of the pair is a triple: its elements look
like ((L, B, M), (5, 2, 3)). So this is not really the same set as X × X × X ×
Y × Y × Y ; however, the bijection between them is so obvious (just a matter
of deleting or inserting some brackets) that it does no harm to identify them.
A similar comment applies to the identification of 6-tuples with ‘strings’.

The following is the reason why Cartesian products are called products.

Theorem 1.25 (Product Principle). If X1 , X2 , · · · , Xn are finite sets, then

|X1 × X2 × · · · × Xn | = |X1 | × |X2 | × · · · × |Xn |.

In particular, |X n | = |X|n .

This principle encapsulates the idea of ‘independent choices’: in the definition

of the Cartesian product X1 × X2 × · · · × Xn , there are no restrictions which
relate the different entries of the n-tuple, so no matter which of the |X1 |
choices you make for the first entry, there are still |X2 | choices for the second
entry and so on. This is why the total number of n-tuples is the product of
all the individual sizes.
Remark 1.26. Just as it is useful to define x1 = x, one would always
identify X 1 with X (without worrying about whether a ‘1-tuple’ ought to
have brackets around it – see Remark 1.24). Similarly, on the same principle
which motivates setting x0 = 1, it is useful to stipulate that X 0 always has a
single element: some sort of ‘empty tuple’ which you could write as (). Then
the rule |X n | = |X|n holds for all n ∈ N.
Example 1.27. How many five-digit numbers are there? Probably the sim-
plest approach is that the five-digit numbers are those n ∈ N such that
n > 9999 and n ≤ 99999, so the answer is 99999 − 9999 = 90000; this is
an application of the Difference Principle, because we are implicitly view-
ing the set of five-digit numbers as the complement of {1, 2, · · · , 9999} in
{1, 2, · · · , 99999}. An alternative approach is that there are 9 choices for the
first digit (since it can’t be zero) and 10 choices for every other digit, so the
answer is 9 × 10 × 10 × 10 × 10 = 90000; this is an application of the Product
Principle, because we are implicitly identifying the set of five-digit numbers
CHAPTER 1. COUNTING PROBLEMS 17

with the set of 5-tuples where the first entry is an element of {1, · · · , 9}
and all the other entries are elements of {0, · · · , 9}. The second approach
is necessary for some variants of the question. For instance, the number of
five-digit numbers which are palindromes (i.e. the digits have the form abcba)
is 9 × 10 × 10 = 900, because the choice of the first three digits determines
the last two. (The Bijection Principle is also involved here, because we are
effectively using the fact that the five-digit palindromes are in bijection with
the strings formed by deleting their last two digits.)

For some situations, the above statement of the Product Principle is too
restrictive. Instead of counting the ways of making independent selections
from pre-determined sets X1 , · · · , Xn , you might have a situation where the
set X2 depends on what element you selected from X1 , and then the set X3
depends on what elements you selected from X1 and X2 , and so on. All you
really need to be able to apply the Product Principle is that the sizes |Xi |
don’t depend on the choices you’ve already made.
Example 1.28. How many two-digit numbers have two different digits?
Since there are 90 two-digit numbers in total, and the digits are the same
in 9 of them (11, 22, · · · , 99), the answer is 81 (by the Difference Principle).
Another approach is that there are 9 choices for the first digit (anything from
1 to 9), and then having chosen the first digit, there are 9 choices for the
second digit (anything except the digit you’ve already used), so the answer is
9×9 = 81. This is an application of the looser form of the Product Principle,
since the set of possibilities for the second digit varies depending on what
first digit you chose; what matters is that it always has 9 elements. What
about trying to choose the digits in the other order? Since the second digit
is allowed to be zero, there are 10 choices for it, and then there should surely
be 8 choices for the first digit (anything except zero and the chosen second
digit). But the answer isn’t 80; what’s gone wrong is that if you do choose
zero as the second digit, the number of choices for the first digit is not 8 but
9. Although this example is pretty trivial, it does illustrate the need for care
in applying the Product Principle.
18 Topics in Discrete Mathematics

1.2 Ordered selection

A classic situation where the Product Principle arises is in counting the

number of ways to carry out some ‘ordered selection’.
Example 1.29. Suppose there are three students A, B, and C, each of whom
has to be allocated to one of 4 tutorial times, with no restrictions on how
many are put at each time (in particular, we are allowed to ‘repeat a selec-
tion’, i.e. choose the same tutorial time for more than one student). If all
the students can attend all 4 times, then the number of ways of doing the
allocation is 43 = 64. If student A can only attend 2 of the times, student
B can only attend 3 of the times, and student C can only attend 1 of the
times, then the number of ways is 2 × 3 × 1 = 6. Here we are implicitly using
the Product Principle to count the elements of the set X1 × X2 × X3 , where
X1 is the set of times that student A can attend, X2 is the set of times that
student B can attend, and X3 is the set of times that student C can attend.

Since it is not always easy to think of such problems in terms of a literal

selection, it is helpful to have a more purely mathematical framework in which
to discuss them. The use of the word “allocate” in the previous example is
a hint that it is essentially about counting some functions from one set to
another (namely, from the set of students to the set of tutorial times – the
allocation is a function because every student is allocated a single tutorial
time). We have the following general result:
Theorem 1.30. If X and Y are sets with |X| = k and |Y | = n, then the
number of functions f : X → Y is nk .

Proof. We can name the elements of X as x1 , x2 , · · · , xk . (Another way

to say this is that we can choose a bijection between X and {1, 2, · · · , k}.)
Then to specify a function f : X → Y is the same as choosing the elements
f (x1 ), f (x2 ), · · · , f (xk ), all of which must lie in Y . There are n choices for
f (x1 ), n choices for f (x2 ), and so on up to f (xk ), so the answer is n×n×· · ·×n
(k factors), which is nk . Implicitly, we are using the Bijection Principle and
the Product Principle here: the set of functions f : X → Y is in bijection
with the set Y k via the association of any function f : X → Y with the k-
tuple (f (x1 ), f (x2 ), · · · , f (xk )), and the size |Y k | equals |Y |k by the Product
Principle.
CHAPTER 1. COUNTING PROBLEMS 19

Remark 1.31*. To make Theorem 1.30 true when X is empty, we need the
number of functions f : ∅ → Y to be |Y |0 = 1. Indeed, the convention is
that there is a unique function from the empty set to every other set.
Example 1.32. If there are 5 different presents under a Christmas tree, and
3 greedy children grabbing at them until they are all claimed, and no present
can be shared between more than one child, how many possible results are
there? The answer is given by Theorem 1.30; the main mental difficulty is
deciding which way round the functions go, from the set of children to the set
of presents or vice versa. Here we have a giveaway phrase: “no present can
be shared between more than one child”. That is, every present is associated
with a single child, so the “possible results” are really the functions from the
set of presents to the set of children, and the answer is 35 = 243. To phrase a
problem which amounted to counting the functions from the set of children
to the set of presents, you would have to specify that every child gets a single
present and remove the ban on sharing presents and the requirement that
every present is claimed. Then the answer would be 53 = 125.

An important consequence of Theorem 1.30 is:

Theorem 1.33. The number of subsets of a finite set X (including the empty
set and the set itself) is 2|X| .
Example 1.34. The set {1, 2, 3} has the following 8 = 23 subsets: ∅, {1},
{2}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}.

Proof. The idea of the proof is that to specify a subset is the same as
specifying, for every element of the set, whether it is in or out. We can
formalize this idea as saying that there is a bijection between the set of
subsets of X and the set of functions f : X → {0, 1}. To define this bijection,
start with any subset A of X, and associate to this the function
(
1, if x ∈ A,
fA : X → {0, 1} defined by fA (x) =
0, if x ∈ / A.

This is a bijection because every function f : X → {0, 1} arises in this way

for a unique subset A, namely the subset {x ∈ X | f (x) = 1}. Applying the
Bijection Principle, the number of subsets of X is the same as the number
of functions X → {0, 1}, which is 2|X| by Theorem 1.30.
20 Topics in Discrete Mathematics

Many problems amount to counting the number of functions f : X → Y

which satisfy the extra condition of being injective, i.e. we never have f (x1 ) =
f (x2 ) when x1 6= x2 : different inputs give different outputs. These are the
problems which involve ‘ordered selection with repetition not allowed’. We
already saw one such problem in Example 1.28; here is another.
Example 1.35. Suppose that car numberplates consisted of six letters of
the alphabet. If there was no restriction on repeating letters, the number of
possible numberplates would be

If repetition of letters was not allowed, the number of possible numberplates

would be

Theorem 1.36. If X and Y are sets with |X| = k and |Y | = n, then the
number of injective functions f : X → Y is n(n − 1)(n − 2) · · · (n − k + 1).

Proof. First suppose that n < k, i.e. Y is a smaller set than X; then there
cannot be any injective function from X to Y (by the Pigeonhole Principle).
The formula does indeed give the answer 0 in this case, because one of the
factors in the product is n − n.

Now suppose that n ≥ k. Name the elements of X as x1 , x2 , · · · , xk . To

specify an injective function f : X → Y , we can first choose f (x1 ) in n ways.
Then f (x1 ) is ruled out as a possible value for f (x2 ), so there are n − 1
options for f (x2 ); and so on, until when we come to choose f (xk ), our k − 1
previous choices are ruled out, so there are n − k + 1 options left. By the
looser form of the Product Principle, the number of injective functions is the
product of n, n − 1, · · · , n − k + 1, as claimed.

Notice that there are k factors in the product in Theorem 1.36, which is
called a falling factorial (“falling” because the factors decrease by 1 at each
step). The usual notation for these is:
n(k) := n(n − 1)(n − 2) · · · (n − k + 1). (1.3)
This makes sense when n is any complex number, which will be useful later.
CHAPTER 1. COUNTING PROBLEMS 21

Remark 1.37. A note on conventions regarding products. The definition

(1.3) of n(k) shows the first three factors of a product, a “ · · · ”, and then the
final factor. This may seem to imply that there should always be at least
five factors. But in interpreting this for small values of k, one has to follow
the pattern that there are always k factors, rather than the literal sense.
For instance, n(1) is defined to be n; what should be the last factor, namely
n − 1 + 1, is the same as the first, so we stop there already. Similarly, n(2) is
defined to be n(n − 1). As for the case k = 0, we make sense of that by the
convention that a product of 0 things means 1, so n(0) = 1 for any n.

An important special case of Theorem 1.36 is when Y = X. An injective

function from X to itself must also be surjective, because the values of the
function are all different and hence exhaust the elements of X. Thus it must
be a bijection from X to itself, which we call a permutation of X. So in this
special case, Theorem 1.36 says that the number of permutations of X is

k! := k(k) = k(k − 1)(k − 2) · · · 3 × 2 × 1, (1.4)

which is called k factorial. According to our conventions, 0! = 1 (there is a

single permutation of the empty set). The sequence of factorial continues:

1! = 1, 2! = 2, 3! = 6, 4! = 24, 5! = 120, 6! = 720, 7! = 5040, · · ·

Example 1.38. The permutations of {1, 2, 3} can be written as strings

where the first digit is f (1), the next f (2), and the third f (3):

123, 132, 213, 231, 312, 321.

There are indeed 3! = 6 of these.

In general, k! is the number of ways of ordering k objects, where “ordering”

may manifest itself differently in various examples.
Example 1.39. Here are some of the things that are counted by 4! = 24:

• the ways that four people can line up in a queue;

• the four-digit numbers whose digits are 3, 5, 7, and 9 (in some order);

• the ways of putting four different letters in four different envelopes.

22 Topics in Discrete Mathematics

We can express n(k) in terms of plain factorials, whenever n is another non-

negative integer and n ≥ k:
n(n − 1) · · · 1 n!
n(k) = n(n − 1) · · · (n − k + 1) = = . (1.5)
(n − k)(n − k − 1) · · · 1 (n − k)!
n!
But beware that (n−k)!
makes no sense if n is not an integer, or if n < k.

Counting the surjective functions from X to Y is a more subtle problem

which we will return to in the section after next. To conclude this section,
here is another example along similar lines to Theorem 1.36.
Example 1.40. In how many ways can a group of 2n people be split into
n pairs for games of table tennis? Imagine them lining up in single file. The
opponent of the first person can be any one of the other people, so there are
2n − 1 choices. Having made that choice, the first pair can both leave the
line, leaving 2n − 2 people behind; there are 2n − 3 choices for the opponent
of the first person, then they can both leave the line, leaving 2n − 4 people
behind; and so on. So the answer is
(2n − 1)!! := (2n − 1)(2n − 3) · · · 5 × 3 × 1, (1.6)
the product of all the odd numbers up to 2n − 1. (The double ! is meant to
indicate that the factors decrease by 2 at each step.) This can be expressed
in terms of ordinary factorials:
(2n − 1)!! = (2n − 1)(2n − 3) · · · 5 × 3 × 1
(2n)(2n − 1)(2n − 2)(2n − 3)(2n − 4) · · · 5 × 4 × 3 × 2 × 1
=
(2n)(2n − 2)(2n − 4) · · · 4 × 2
(2n)(2n − 1)(2n − 2)(2n − 3)(2n − 4) · · · 5 × 4 × 3 × 2 × 1
=
2n n(n − 1)(n − 2) · · · 2 × 1
(2n)!
= n .
2 n!
For instance, if we have 6 people then the answer is 15 = 5 × 3 × 1 = 236!3! .
Incidentally, it will be convenient later to allow n = 0 in the definition of
(2n − 1)!!. Applying to (1.6) the principle that a product of 0 things means
1, we get the somewhat unpleasant definition (−1)!! := 1.
Remark 1.41*. There is an alternative approach to Example 1.40. We
know that (2n)! counts the number of ways of ordering the 2n people in a
CHAPTER 1. COUNTING PROBLEMS 23

line. Any such ordering gives rise to a way of pairing them off: simply pair
the first with the second, the third with the fourth, and so on. So we have a
function from the set of orderings to the set of pairings. If we start with a
given pairing, the number of orderings which give rise to it is 2n n!, because
we can order the n pairs in n! ways, and then decide which person comes
first in each pair in 2n ways. So every one of the preimages for our function
has size 2n n!, and by the Overcounting Principle, there are (2n)!
2n n!
pairings.

1.3 Binomial coefficients

Another application of the Overcounting Principle is in counting subsets of

a prescribed size. Recall that the total number of subsets of an n-element
set is 2n .
Theorem 1.42. For n, k ∈ N, the number of k-element subsets of an n-
element set X is
n(k) n(n − 1)(n − 2) · · · (n − k + 1)
= .
k! k!
Example 1.43. The number of 2-element subsets of {1, 2, 3, 4} is six:

{1, 2}, {1, 3}, {1, 4}, {2, 3}, {2, 4}, {3, 4}.
4×3
This agrees with the stated formula: 2×1
= 6.

Proof. By Theorem 1.36, n(k) counts the injective functions {1, 2, · · · , k} →

X. That is, it is the number of ordered selections of k different elements
x1 , x2 , · · · , xk of X. What we want is the number of unordered selections
{x1 , x2 , · · · , xk } of k different elements of X. The relationship between the
set of ordered selections and the set of unordered selections is exactly that of
the Overcounting Principle: every ordered selection gives rise to an unordered
selection by forgetting the order, and for every unordered selection, there are
exactly k! orderings of it. So the number of unordered selections is the
number of ordered selections divided by k!, which is the claim.
Remark 1.44. Notice that if n < k, there are no k-element subsets of
X; the formula does indeed give 0 in that case. Another notable aspect of
24 Topics in Discrete Mathematics

Theorem 1.42 is that it proves that k! always divides n(k) (i.e. their quotient
is an integer), which is not particularly easy to see otherwise. For instance,
n(n − 1)(n − 2)(n − 3) is a multiple of 24 for all n ∈ N.

We will use the notation

 
n n(k) n(n − 1)(n − 2) · · · (n − k + 1)
:= = , (1.7)
k k! k!

read “n choose k”. This definition makes sense when n is any complex
number, which will be useful in Chapter 3. Note that
       
n n n n
= = 1, = = n. (1.8)
0 n 1 n−1

If n is an integer and n ≥ k, then the alternative formula (1.5) for n(k) gives
 
n n!
= . (1.9)
k k!(n − k)!

The formula (1.9) is symmetric in k and n − k, so

   
n n
= , for all n, k ∈ N, n ≥ k. (1.10)
k n−k

Remark 1.45. Equation (1.10) can also be proved using the Bijection Prin-
ciple: there is a bijection between the k-element subsets of an n-element set
X and the (n − k)-element subsets, which associates to every k-element sub-
set A its complement X \A. This is a bijection because every (n−k)-element
subset B of X arises in this way for exactly one k-element subset, namely
X \ B.
Example 1.46. There are 52 playing cards in a standard deck, 13 of each
suit (spades, hearts, diamonds, and clubs). If you are dealt 5 of them at
random, how many possible hands can you end up with? Since the order
in which you are dealt the cards doesn’t matter, this is just the number of
5-element subsets of the set of cards, which is
 
52 52 × 51 × 50 × 49 × 48
= = 2598960.
5 5×4×3×2×1
CHAPTER 1. COUNTING PROBLEMS 25

In card games such as poker, you need to know the probability of getting
various kinds of hands, which means you need to count how many of the
5-element subsets satisfy certain properties. For instance, how many hands:

contain no spades?

contain the king of spades?

contain exactly one spade?

Example 1.47.
The
number of ways of forming a string of seven a’s and
five b’s is 12
5
= 12
7
12!
= 7!5! = 792. This is because the string must be 12
letters long, and to specify it is the same as choosing which seven of the 12
positions are occupied by an a (or equivalently, which five of the 12 positions
are occupied by a b).

n


The numbers k
are known as binomial coefficients, from their role in:

Theorem 1.48 (Binomial Theorem). If a, b are any numbers and n ∈ N,

n  
X
n n
(a + b) = an−k bk
k
  k=0   
n n n n−1 n n−2 2
= a + a b+ a b + ···
0 1 2
     
n 2 n−2 n n−1 n n
+ ab + ab + b
n−2 n−1 n
n(n − 1) n−2 2 n(n − 1) 2 n−2
= an + nan−1 b + a b + ··· + ab + nabn−1 + bn .
2 2

Proof. Imagine expanding out (a + b)n = (a + b)(a + b) · · · (a + b). There

will be 2n terms in the expansion, each of which is (initially) a string of a’s
and b’s, of total length n; this is because you get one term for every choice
of either a or b from the first factor, a
or b from the second factor, and so on.
For every k from 0 up to n, exactly nk of these strings have a in n − k places
and b in k places; this is because there are k ways of specifying which factors
26 Topics in Discrete Mathematics

to choose b from, and then the other factors must all be ones you choose a
from (see Example 1.47). So when you collect
terms together using the fact
that ab = ba, the coefficient of an−k bk is nk .
Remark 1.49*. As the proof shows, the a and b in the Binomial Theorem
don’t have to be numbers. They could be elements of any ring (a sort of
generalized number system where addition and multiplication are defined
and satisfy the distributive law), provided that ab = ba. So, for instance, the
result is still true if applied to two square matrices A and B which commute
(but would be false if A and B did not commute).

Some special cases of the Binomial Theorem are worth pointing out. When
a = b = 1, it says that
Xn          
n n n n n n
2 = = + + ··· + + . (1.11)
k=0
k 0 1 n−1 n

This also follows from the Sum Principle, because the left-hand side counts
all subsets of a set of size n, and the terms of the right-hand side count the
number of subsets of a given size. When a = 1 and b = −1, we get
n   (
X n 1, if n = 0,
(−1)k = 0n = (1.12)
k=0
k 0, otherwise.



If n is odd this can be seen directly, because nk and n−k n
occur on the
left-hand side with opposite signs and hence cancel each other out.

There is a famous diagrammatic representation of the binomial coefficients

called Pascal’s triangle:

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
1 7 21 35 35 21 7 1
.. .. .. .. ..
. . . . .
CHAPTER 1. COUNTING PROBLEMS 27

Here the rows correspond to n = 0, n = 1, n = 2, and so on,
and the

columns
n n
correspond to k = 0, k = 1, k = 2, and so on. (Since n−k = k , each row
of Pascal’s triangle reads the same backwards as forwards. Because of this
symmetry, it is more usual to align the rows so that they are all centred on
a vertical line.) Notice that each entry in the above triangle is the sum of
the one above and to the left, and the one directly above; this suggests the
following result.
Theorem 1.50. The binomial coefficients satisfy the recurrence relation
     
n n−1 n−1
= + , for all n ≥ 1.
k k−1 k

Proof. One can prove this from the Binomial Theorem, by expressing (a+b)n
as (a + b)(a + b)n−1 . But here is a more satisfying argument using the Sum
Principle. The left-hand side counts the number of k-element subsets A of
{1, 2, · · · , n}. Now a subset A either contains the element n or it doesn’t.
(More formally, the set of k-element subsets is the disjoint union of the set
of k-element subsets which contain n and the set of k-element subsets which
don’t contain n.) Choosing a k-element subset A which contains n just
amounts
to choosing a (k − 1)-element subset of {1, 2, · · · , n − 1}, so there
are n−1 subsets which contain n (notice the implicit appeal to the Bijection
k−1

Principle in the phrase “just amounts to”). Similarly, there are n−1 k
subsets
which don’t contain n. The recurrence relation follows.

Here is a slight generalization of binomial coefficients.

Theorem 1.51. Let X be a set with n elements, and
n

let n1 , n2 , · · · , nm ∈ N
be such that n1 + n2 + · · · + nm = n. Let n1 ,n2 ,··· ,nm be the number of ways
to choose subsets A1 , A2 , · · · , Am of X such that:

(1) |Ai | = ni for i = 1, 2, · · · , m, and

(2) X is the disjoint union of the Ai ’s.

Then  
n n!
= .
n1 , n2 , · · · , nm n1 !n2 ! · · · nm !
28 Topics in Discrete Mathematics

Proof. We can choose A1 in nn1 ways. Then to maintain disjointness, A2


must be a subset of X \ A1 , so it can be chosen in n−n 1
ways. Then A3
n2
n−n1 −n2


must be a subset of X \ (A1 ∪ A2 ), so it can be chosen in n3
ways, and
so forth. So we have
      
n n n − n1 n − n1 − n2 − · · · − nm−1
= ···
n1 , n2 , · · · , nm n1 n2 nm
n! (n − n1 )! (n − n1 − n2 − · · · − nm−1 )!
= ···
n1 !(n − n1 )! n2 !(n − n1 − n2 )! nm !(n − n1 − n2 − · · · − nm )!
n!
= ,
n1 !n2 ! · · · nm !
because everything else cancels out. A slightly more elegant proof uses
the Overcounting Principle: instead of choosing unordered subsets we could
choose the elements of A1 in order, then the elements of A2 in order, and
so on. The number of ways of doing this would be n!, since it amounts
to just ordering all the elements of X; and this overcounts by a factor of
n1 !n2 ! · · · nm !, because we have ordered each of the subsets.


The numbers n1 ,n2n,··· ,nm are called multinomial coefficients because of their
occurrence in the generalization of the Binomial Theorem:
Theorem 1.52 (Multinomial Theorem). If a1 , a2 , · · · , am are any numbers
and n ∈ N, then
X  
n
n
(a1 + a2 + · · · + am ) = an1 1 an2 2 · · · anmm .
n ,n ,··· ,n ∈N
n1 , n2 , · · · , nm
1 2 m
n1 +n2 +···+nm =n

Proof. This is similar to the proof of the Binomial Theorem. When you
expand (a1 + a2 + · · · + am )n , you get a sum of terms. A term corresponds to
a set of choices of which of a1 , a2 , · · · , am to select from each of the n factors.
To get the coefficient of an1 1 an2 2 · · · anmm , you need to count how many ways
there are to select a1 from n1 factors, a2 from n2 factors, and so on. This
amounts to choosing disjoint subsets of the factors, one of size n1 , one of size
n2 , etc., which is what the multinomial coefficient counts.
Remark 1.53. A nice consequence of Theorem 1.51 is that n1 !n2 ! · · · nm !
always divides (n1 + n2 +
· · · + nm )!. Beware of a slight notational
conflict:
the binomial coefficient nk equals the multinomial coefficient k,n−k n
.
CHAPTER 1. COUNTING PROBLEMS 29

Multinomial coefficients arise in a range of counting problems.

Example 1.54. If there are 5 different presents under a Christmas tree and
3 virtuous children who wait to be given what they deserve, how many ways
are there to distribute the presents so that child A gets two, child B gets two,
and child C gets one? This amounts to choosing a disjoint union of the
set of
5
presents into three subsets of sizes 2, 2, and 1, so the answer is 2,2,1 = 30.
Example 1.55. How many ways are there to rearrange the letters of WOOL-
LOOMOOLOO? (The answer is not 13!, because of the repeated letters.) In
other words, how many different strings are there which consist of one W,
one M, three L’s, and eight O’s? As in the proof of the Multinomial Theo-
rem, to specify such a string we have to express the 13 positions as a disjoint
union of a subset of size 1 (the position of W), another subset of size 1 (the
position of M), a subset of size 3 (the positions of the L’s),

and a subset of
13 13!
size 8 (the positions of the O’s). So the answer is 1,1,3,8 = 1!1!3!8! = 25740.
Example 1.56*. How many ways are there to split kn people into n groups
of k? The answer depends on whether you consider the groups to be ordered,
in the sense of caring which group is which. If the groups are ordered (say,
because you want to assign each group to a different task), then this is a
kn

(kn)!
special case of Theorem 1.51, and the answer is k,k,··· ,k
= (k!)n . If there is no
ordering of the groups (say, because you want all the groups to work amongst
themselves on the same task), then we can use the Overcounting Principle:
every unordered grouping corresponds to exactly n! ordered groupings, so
(kn)!
the answer is (k!) n n! . Note that in the case k = 2 this becomes the same

answer that we gave in Example 1.40.

Example 1.57*. Suppose you have 16 tennis players in a tournament, 4 of
whom are seeds. How many ways are there to construct a standard four-
round knock-out draw so that no two seeds can possibly play each other
before the semi-finals? The trick here is to notice that in any such draw,
the players are divided into four groups of 4, with one seed in each group;
each group plays a mini-tournament amongst itself in the first two rounds,
with the winner progressing to the semi-finals. If you imagine each group as
named after the seed it contains, then assigning the other twelve players to
the groups amounts

to splitting them into 4 ordered groups of 3, so it can be
12
done in 3,3,3,3 ways. Then within each group, there are three ways to form
the two-round mini-tournament; and there are also three
ways to allocate
12
the group winners to semi-finals. So the answer is 35 3,3,3,3 = 89812800.
30 Topics in Discrete Mathematics

One can think of nk as the number of ways to make an unordered selection
of k things from a total of n with repetition not allowed. It is then natural
to consider what happens if you allow repetition.
Example 1.58. If you have 5 different marbles in a bag – say red, blue,
white,

yellow and green – and stick your hand in to pull out 3, there are
5
3
= 10 possible outcomes. (Remember that if you draw out the marbles
one by one, and you consider that the order in which you draw them out
matters, there are 5(3) = 60 possible outcomes.) To allow repetition, suppose
that, instead of just one marble of each colour, you have an unlimited supply
of each colour, indistinguishable amongst themselves. Then you may get
more than one marble of a particular colour when you stick your hand in
the bag and pull out 3. To specify an outcome, you just need to specify the
number of red marbles, the number of blue marbles, and so on; together,
these numbers constitute a 5-tuple (k1 , k2 , k3 , k4 , k5 ) of nonnegative integers
such that k1 +k2 +k3 +k4 +k5 = 3. One way to count the possible outcomes is
by dividing them into three disjoint subsets according to the following cases:

(1) all three marbles have the same colour (i.e. ks = 3 for some s, and all
the other ki ’s are zero); or

(2) two marbles have the same colour, and the third is a different colour
(i.e. ks = 2 and kt = 1 for some s 6= t, and the other ki ’s are zero); or

(3) all three marbles have different colours (i.e. three of the ki ’s are 1, and
the other two ki ’s are zero).

You can calculate that

the number of outcomes of type (1) is

the number of outcomes of type (2) is

the number of outcomes of type (3) is

the total number of outcomes is

CHAPTER 1. COUNTING PROBLEMS 31

Remark 1.59*. It may seem that we could count unordered selections of k

from n with repetition allowed by taking the count of ordered selections of
k from n with repetition allowed – which as we have seen is just nk – and
dividing by k!, in an application of the Overcounting Principle. But this gives
k
the wrong answer in Example 1.58, and indeed nk! is not usually an integer.
The reason the Overcounting Principle doesn’t apply is that the number of
orderings of one of our unordered selections is not always the same. For
instance, in Example 1.58, a selection of three marbles of different colours
corresponds to 3! ordered selections, whereas a selection of three marbles of
the same colour corresponds to only one ordered selection.

Here is the answer to the problem in general.

Theorem 1.60. Let n, k ∈ N. When making an unordered selection of k
things from n possibilities, with repetition allowed, the number of different
outcomes is
 
n+k−1 n(n + 1)(n + 2) · · · (n + k − 1)
= .
k k!
Remark 1.61. Notice the similarity between the formulas in Theorems 1.42
and 1.60: the difference is that in Theorem 1.60, the factors in the numerator


rise from n in steps of 1, instead of falling from n as they do in nk . Some



authors use the notation ( nk ) for this, but since it equals n+k−1 k
there
doesn’t seem much need.

Proof*. We will first give the proof in the case of Example 1.58, where
we are making an unordered selection of 3 marbles from 5 possible kinds.
Having made a selection, we can lay the marbles out in a row, with the red
marbles (if any) first, followed by the blue, the white, the yellow and the
green. Imagine putting down dividing lines to mark off the different colours:
one line between red and blue, one line between blue and white, one line
between white and yellow, and one line between yellow and green. Then you
can forget the colours, and what remains is an arrangement of 3 marbles and
4 dividing lines. For instance, the arrangement
| ◦ ◦ | ◦ ||
means that we have no red marbles, two blue marbles, one white marble,
and no yellow or green marbles (the 5-tuple is (0, 2, 1, 0, 0)). What we have
32 Topics in Discrete Mathematics

shown is that the possible outcomes are in bijection with the different strings
of three ◦’s and four |’s. By the Bijection Principle, the number
of outcomes
equals the number of such strings, which by Example 1.47 is 73 = 35. The
general case is no harder: we can imagine that we are selecting k marbles
from n different kinds, which amounts to considering

strings

of k ◦’s and
n − 1 |’s. The number of such strings is k+n−1k
= k+n−1
n−1
, as claimed.

If talk of “marbles” and “dividing lines” strikes you as too informal, the
real effect of this argument is to construct a bijection between n-tuples
(k1 , k2 , · · · , kn ) of nonnegative integers which satisfy k1 + k2 + · · · + kn = k
(these are what specify the possible outcomes of the selection, with ki record-
ing how many times the ith possibility was selected) and (n − 1)-element
subsets of the set {1, 2, · · · , n + k − 1} (these specify the positions of the
dividing lines in the string). Starting from an n-tuple (k1 , · · · , kn ), the sub-
set we associate to it is {k1 + 1, k1 + k2 + 2, · · · , k1 + · · · + kn−1 + n − 1};
it is easy to see that the elements here are listed in increasing order, so
they are indeed n − 1 different elements of {1, 2, · · · , n + k − 1}. For the
inverse function, we start with a subset {i1 , i2 , · · · , in−1 }, which we can
assume is listed in increasing order, and then associate to it the n-tuple
(i1 − 1, i2 − i1 − 1, · · · , in−1 − in−2 − 1, n + k − 1 − in−1 ), whose elements are
easily seen to be nonnegative with sum k. The verification that these maps
are inverse to each other is straightforward.
Example 1.62. If there are 5 identical presents under a Christmas tree, and
3 greedy children grabbing at them until they are all claimed, and no present
can be shared between more than one child, how many possible results are
there? Notice that the only difference between this question and Example
1.32 is that the presents are now indistinguishable. Is this a question about
ordered or unordered selection? Is repetition allowed or not allowed? Once
again, we should think that we are selecting, for each present, the child it
ends up with: thus it is an unordered selection of 5 things (children) from
3 possibilities, with repetition allowed (because more than one present
is
3+5−1
allowed to end up with the same child), and the answer is 5
= 21.
If you find yourself confused about what is being selected, it may help to
think in terms of n-tuples: in the present case, all that matters is how many
presents child A gets, how many presents child B gets, and how many presents
child C gets (because the presents are all identical). So this is the problem of
counting 3-tuples of nonnegative integers which add up to 5, not the problem
CHAPTER 1. COUNTING PROBLEMS 33

of counting 5-tuples of nonnegative integers which add up to 3 (which we

solved in Example 1.58). Finally, what happens if we add the reasonable
requirement that every child gets at least one present, so that we are counting
3-tuples of positive integers which add up to 5? We can solve this simply by
imagining that we give every child one present to start with; then there are
2 presents left, and we
are back in the original situation with 2 in place of 5,
so the answer is 3+2−1
2
= 6. Explicitly, the six 3-tuples are

(3, 1, 1), (1, 3, 1), (1, 1, 3), (2, 2, 1), (2, 1, 2), (1, 2, 2).

Remark 1.63**. In these problems about unordered selection, we are not

counting the functions between two sets like we were in the previous section:
every outcome of the allocation of presents to children corresponds to many
different functions from the set of presents to the set of children, and we
are blurring the distinction between these functions by refusing to distin-
guish between the presents. What we are really counting is the number of
equivalence classes of functions f : X → Y , where two functions f and f 0 are
thought of as equivalent if there is some permutation σ of the set X such that
f 0 (x) = f (σ(x)) for all x ∈ X (in other words, f 0 takes the same values as f
but for different inputs). Theorem 1.60 says that if |X| = k and |Y
| = n, the
number of equivalence classes of functions f : X → Y is n+k−1 k
(compare
Theorem 1.30, which counts the functions without imposing the concept of
equivalence). We also know that
the number of equivalence classes of in-
jective functions f : X → Y is nk (because we just need to specify the k
values of the function, which must all be different – this is how we counted
k-element subsets of an n-element set in the first place).

As for the number of equivalence classes of surjective functions, we can find

that by the reasoning at the end of Example 1.62. If k < n, there are no
surjective functions f : X → Y , by the Reverse Pigeonhole Principle. If
k ≥ n, then we can let A be any n-element subset of X (it doesn’t matter
which, because we are only working up to equivalence), define f on A so
that it gives a bijection between A and Y , and then define f on X \ A
in an arbitrary way. Thus the number of equivalence classes of surjective
functions f : X → Y equals the number

of
equivalence classes of functions
n+(k−n)−1 k−1
f : X \ A → Y , which is k−n
= k−n . Remember that we haven’t yet
seen how to compute the actual number of surjective functions f : X → Y
(without equivalence).
34 Topics in Discrete Mathematics

1.4 Inclusion/Exclusion

Definition 1.64. Recall that if A and B are subsets of a set X, then their
union, their intersection, and the difference of A from B are defined by:

A ∪ B := {x ∈ X | x ∈ A or x ∈ B},
A ∩ B := {x ∈ X | x ∈ A and x ∈ B},
A \ B := {x ∈ X | x ∈ A and x ∈/ B} = A ∩ (X \ B).

These operations satisfy a number of basic laws:

Theorem 1.65. If A, B, and C are subsets of X, then:

(1) A ∪ A = A ∩ A = A.

(2) A ∪ ∅ = A, A ∩ ∅ = ∅.

(3) A ∪ B = B ∪ A, A ∩ B = B ∩ A.

(4) A ∪ (B ∪ C) = (A ∪ B) ∪ C, so we just call it A ∪ B ∪ C.

(5) A ∩ (B ∩ C) = (A ∩ B) ∩ C, so we just call it A ∩ B ∩ C.

(6) A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C).

(7) A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C).

(8) A \ (B ∪ C) = (A \ B) ∩ (A \ C).

(9) A \ (B ∩ C) = (A \ B) ∪ (A \ C).

(10) If B ⊆ A, then A \ (A \ B) = B.

Proof. These are all equivalent to logical statements which are more or less
part of the axioms of mathematics. For example, we ‘prove’ (8) as follows.
To say that two subsets A1 and A2 are equal is to say that the condition for
x to lie in A1 is equivalent to the condition for x to lie in A2 (that is, for
every choice of x in the universal set X, they are either both true or both
CHAPTER 1. COUNTING PROBLEMS 35

false). In the present case, the condition for x to lie in A \ (B ∪ C) is that

x ∈ A and x ∈ / (B ∪ C). But saying that x ∈ / (B ∪ C) is equivalent to saying
that x ∈/ B and x ∈ / C (if it is false that either x ∈ B or x ∈ C, then it is
true that x ∈/ B and x ∈ / C, and conversely). So x ∈ (A \ (B ∪ C)) if and
only if
x ∈ A and x ∈ / B and x ∈/ C. (1.13)
On the other hand, the condition for x to lie in (A \ B) ∩ (A \ C) is that
x ∈ (A \ B) and x ∈ (A \ C); that is, x ∈ A and x ∈ / B and x ∈ A and
x∈ / C. Since the second “x ∈ A” is redundant, this is again equivalent to
(1.13), and (8) is proved.

Now (assuming we are dealing with finite sets) how do these operations relate
to counting? As noted in our discussion of the Sum Principle, |A ∪ B| =
|A| + |B| is true only if A and B are disjoint, i.e. A ∩ B = ∅. If there are
elements in the intersection A ∩ B, they are counted twice when we take the
sum |A| + |B|. To get the right formula for |A ∪ B|, we need to correct this:

|A ∪ B| = |A| + |B| − |A ∩ B|. (1.14)

Example 1.66. Suppose that by a show of hands it is determined that of

the 80 students in a maths class, 30 are studying physics, 40 are studying
chemistry, and 10 are studying both physics and chemistry. How many are
not studying either physics or chemistry? If we let X be the set of students
in the class, A the subset of those doing physics, and B the subset of those
doing chemistry, then the question asks for |X \ (A ∪ B)|. We have

|A ∪ B| =

|X \ (A ∪ B)| =

Similarly, if we want a formula for |A ∪ B ∪ C|, and start with |A| + |B| + |C|
as our first attempt, then we have to correct for the overcounting of the
intersections. If |A| + |B| + |C| − |A ∩ B| − |A ∩ C| − |B ∩ C| is our second
attempt, we find that it is still not quite right, because elements in the triple
intersection A ∩ B ∩ C have been added three times in |A| + |B| + |C| and
36 Topics in Discrete Mathematics

subtracted three times in −|A ∩ B| − |A ∩ C| − |B ∩ C|. So we need to add

the triple intersection back in:

|A∪B ∪C| = |A|+|B|+|C|−|A∩B|−|A∩C|−|B ∩C|+|A∩B ∩C|. (1.15)

Example 1.67. How many numbers between 1 and 1000 are divisible by at
least one of 2, 3, or 5? If we let X be {1, 2, 3, · · · , 1000}, A the subset of
numbers divisible by 2 (i.e. even), B the subset of numbers divisible by 3, and
C the subset of numbers divisible by 5, then we want to find |A ∪ B ∪ C|. We
can use (1.15), because the sizes of these subsets and their intersections can
be easily determined. The key is the observation
n that in the set {1, 2, · · · , n},
the number of multiples of m is exactly m (every mth number is divisible
by m – we round down because all the numbers after the last multiple of m
in the set are wasted). Moreover, a number is divisible by both 3 and 5 (say)
if and only if it is divisible by 15 (this relies on the fact that 3 and 5 have no
common divisors, so their least common multiple is just their product). So
 
1000
|A| = = 500,
2
 
1000
|B| = = 333,
3
 
1000
|C| = = 200,
5
 
1000
|A ∩ B| = = 166,
6
 
1000
|A ∩ C| = = 100,
10
 
1000
|B ∩ C| = = 66,
15
 
1000
|A ∩ B ∩ C| = = 33,
30

so |A ∪ B ∪ C| = 500 + 333 + 200 − 166 − 100 − 66 + 33 = 734. Calculations

like this are used to determine how many numbers up to a certain point are
prime (have no divisors except themselves and 1).

Equations (1.14) and (1.15) are the n = 2 and n = 3 cases of:

CHAPTER 1. COUNTING PROBLEMS 37

Theorem 1.68 (Inclusion/Exclusion Principle)*. If A1 , A2 , · · · , An are sub-

sets of a finite set X, then
n
X X
|A1 ∪ A2 ∪ · · · ∪ An | = (−1)k−1 |Ai1 ∩ Ai2 ∩ · · · ∩ Aik |.
k=1 1≤i1 <i2 <···<ik ≤n

Proof*. We will prove the same property we saw in the special cases: namely,
that every element x of A1 ∪ A2 ∪ · · · ∪ An , no matter which of the Ai ’s it’s
in and which it’s not in, ends up being counted exactly once after all the
additions and subtractions in the right-hand side. Let I be the (nonempty)
subset of {1, 2, 3, · · · , n} consisting of those i such that x ∈ Ai . Then Ai1 ∩
Ai2 ∩ · · · ∩ Aik contains x if and
only if i1 , i2 , · · · , ik all belong to I. So for
|I|
a fixed k, x is counted in k terms in the inner sum on the right-hand
side.
P|I| Thusk−1 the total

number of timesP|I|x is|I|counted

on the right-hand side is
|I| k−1
k=1 (−1) k
. But by (1.12), k=0 k (−1) = 0, so

|I|    
X |I| |I|
k−1
(−1) =− (−1)0−1 = 1.
k=1
k 0

Hence the right-hand side equals the left-hand side.

One important application of the Inclusion/Exclusion Principle is in counting

the number of derangements of a set S. A derangement is a permutation
f : S → S such that f (s) 6= s for all s, i.e. no element of the set is left
unchanged.
Example 1.69*. If you have four different letters and four already addressed
envelopes, how many ways are there to put the letters in the envelopes so
that no letter gets sent to the right address? This is equivalent to counting
the derangements of a set with four elements. Let X be the set of all ways
of putting the letters in the envelopes; we have seen that |X| = 4! = 24. Let
A1 be the subset of ways in which the first letter is in the right envelope,
A2 the same for the second letter, and similarly define A3 and A4 . We
want to calculate |X \ (A1 ∪ A2 ∪ A3 ∪ A4 )|, and we will do it using the
Inclusion/Exclusion Principle. For this we need to know the sizes of these
sets and their intersections. But |Ai | = 3! = 6 for all i, because if the ith
letter is in the right envelope we are just counting the number of ways of
38 Topics in Discrete Mathematics

arranging the others. Similarly, |Ai ∩ Aj | = 2! = 2 for any i < j, because if

the ith and jth letters are right we are just counting the number of ways of
arranging the other two; and |Ai ∩ Aj ∩ Ak | = 1! = 1 for any i < j < k, and
|A1 ∩ A2 ∩ A3 ∩ A4 | = 0! = 1 (one way to get all the letters right). So
|A1 ∪ A2 ∪ A3 ∪ A4 | = |A1 | + |A2 | + |A3 | + |A4 |
− |A1 ∩ A2 | − · · · − |A3 ∩ A4 |
+ |A1 ∩ A2 ∩ A3 | + · · · + |A2 ∩ A3 ∩ A4 |
− |A1 ∩ A2 ∩ A3 ∩ A4 |
= 4 × 3! − 6 × 2! + 4 × 1! − 0!
= 24 − 12 + 4 − 1 = 15.
Hence the number we want is 24 − 15 = 9.
Theorem 1.70*. The number of derangements of an n-element set is
Xn
n!
(−1)k .
k=0
k!

Proof*. We may as well assume that the set is {1, 2, · · · , n}. As in the
preceding example, let X be the set of all permutations of {1, 2, · · · , n}, and
let Ai be the subset consisting of those permutations which leave i unchanged.
Then for any 1 ≤ i1 < i2 < · · · < ik ≤ n,

|Ai1 ∩ Ai2 ∩ · · · ∩ Aik | = (n − k)!,

because once you fix k elements you are just
dealing with a permutation of
the remaining n − k. Moreover, there are nk such sequences 1 ≤ i1 < i2 <
· · · < ik ≤ n. So the Inclusion/Exclusion Principle says that
Xn   Xn
k−1 n n!
|A1 ∪ A2 ∪ · · · ∪ An | = (−1) (n − k)! = (−1)k−1 .
k=1
k k=1
k!

Therefore the number of derangements is

n
X n
X
k−1 n! n!
|X \ (A1 ∪ A2 ∪ · · · ∪ An )| = n! − (−1) = (−1)k ,
k=1
k! k=0
k!

as claimed.
CHAPTER 1. COUNTING PROBLEMS 39

Remark 1.71*. As a consequence of Theorem 1.70, the probability that a

P k
randomly chosen permutation of a set of size n is a derangement is nk=0 (−1)
k!
.
−1 1
As n tends to infinity, this probability converges to e = e (about 0.37).

Another application of the Inclusion/Exclusion Principle is in counting sur-

jective functions.
Example 1.72. Suppose we have 5 students who need to be assigned to 3
tutors, so that every tutor gets at least one student. How many ways can this
be done? Numbering the students and tutors, we see that this is equivalent
to finding the number of surjective functions f : {1, 2, 3, 4, 5} → {1, 2, 3}.
Let X be the set of all functions f : {1, 2, 3, 4, 5} → {1, 2, 3}. Recall that
|X| = 35 = 243. To count the surjective functions, we need to take away
those whose range is a proper subset. So let A be the set of functions whose
range is contained in {1, 2} (i.e. the allocations in which the third tutor
doesn’t get any students), B the set of functions whose range is contained in
{1, 3}, and C the set of functions whose range is contained in {2, 3}. Clearly
|A| = |B| = |C| = 25 = 32,
|A ∩ B| = |A ∩ C| = |B ∩ C| = 1, |A ∩ B ∩ C| = 0.
So the Inclusion/Exclusion Principle says that
|A ∪ B ∪ C| = 3 × 32 − 3 × 1 + 0 = 93,
and the number of surjective functions is |X \ (A ∪ B ∪ C)| = 243 − 93 = 150.
Theorem 1.73*. If X and Y are sets with |X| = m and |Y | = k, then the
number of surjective functions f : X → Y is
Xk        
j k m m k m k m k k
(−1) (k−j) = k − (k−1) + (k−2) −· · ·+(−1) 0m .
j=0
j 1 2 k

Proof*. As in the example, let Ay be the set of functions f : X → Y for

which y ∈/ f (X), for y ∈ Y . Then any intersection Ay1 ∩ · · · ∩ Ayj consists
of functions whose range lies in Y \ {y1 , · · · , yj }, and there are (k − j)m of
these. The Inclusion/Exclusion Principle says that
   
[  X k
  j−1 k
 Ay  = (−1) (k − j)m ,
  j
y∈Y j=1
40 Topics in Discrete Mathematics

which tells us the number of non-surjective functions. So the number of

surjective functions f : X → Y is
X k   Xk  
m j−1 k m j k
k − (−1) (k − j) = (−1) (k − j)m ,
j=1
j j=0
j

as claimed.
 
k m
k
Note that the j = k term of the sum, (−1) 0 , vanishes unless m = 0.
k
Remark 1.74*. This is a much less straightforward formula than the count
of injective functions given by Theorem 1.36. For small values of k (and
assuming m ≥ 1, so the j = k term vanishes), the formula is:
k =1: 1
k =2: 2m − 2
k =3: 3m − 3 × 2m + 3
k =4: 4m − 4 × 3m + 6 × 2m − 4
Note the pattern of alternating signs, decreasing bases to the power m, and
binomial coefficients. There are two unsatisfactory aspects. Firstly, the
number of terms is k, so if we have a problem where k is growing, the formula
gets less and less ‘closed’. Secondly, the fact that there are minus signs in the
sum means that the terms can potentially grow much bigger than the final
answer. For example, when m < k we know that the final answer must be zero
by the Reverse Pigeonhole Principle, so all the terms just cancel away! To
be sure of not making bigger computations than necessary, combinatorialists
prefer ‘positive’ formulas, which do not involve minus signs.

1.5 Stirling numbers

If you experiment with the formula given in Theorem 1.73, you will notice
that it evaluates to numbers which are more divisible than you would have
a right to expect. We will prove in this section that, just as the number of
injective functions f : X → Y (i.e. |Y |(|X|) ) is always divisible by |X|!, the
number of surjective functions f : X → Y is always divisible by |Y |!.
CHAPTER 1. COUNTING PROBLEMS 41

Definition 1.75. For n, k ∈ N, the Stirling number S(n, k) is the number of

ways of writing an n-element set as a disjoint union of k nonempty subsets. A
way of writing a set as a disjoint union of nonempty subsets is usually called
a partition of the set, and the subsets are called the blocks of the partition.
Notice that, unlike in Theorem 1.51, there is no order on these blocks, and
their sizes are not specified.

Example 1.76. The partitions of {1, 2, 3} with 2 blocks can be represented

as 1|23, 2|13, and 3|12, where 2|13 means that you write {1, 2, 3} as the
disjoint union of {2} and {1, 3}, and so on. So S(3, 2) = 3.
Example 1.77. The partitions of {1, 2, 3, 4} with 2 blocks are as follows:

So S(4, 2) = .

The partitions of {1, 2, 3, 4} with 3 blocks are as follows:

So S(4, 3) = .

Example 1.78. It is obvious that you can’t partition an n-element set into
more than n blocks, so S(n, k) = 0 when k > n.
Example 1.79. You can’t partition a nonempty set into 0 blocks, so we
have S(n, 0) = 0 when n > 0. When n = 0, the convention is to declare
that there is a single partition of the empty set into 0 blocks, so S(0, 0) = 1.
(This is the right value for subsequent formulas.)
Example 1.80. If n > 0, then there is only one way to partition an n-
element set into 1 block (everything has to be lumped in together), and only
one way to partition it into n blocks (everything has to be alone in its own
block). So S(n, 1) = S(n, n) = 1.



This is
all a bit reminiscent of the binomial coefficient facts: n0 = nn = 1,
and nk = 0 if k > n. So it seems like a good idea to display the Stirling
numbers S(n, k) in a triangle like Pascal’s triangle, with rows corresponding
to n = 1, n = 2, and so on, and columns corresponding to k = 1, k = 2, and
42 Topics in Discrete Mathematics

so on. The start of the triangle would be:

1
1 1
1 3 1
1 7 6 1
We can see already from the fourth row that we are not going to have the
same symmetry property as in Pascal’s triangle (so there’s not much point
aligning the centres of the rows). But there is still a rule, like Pascal’s
recurrence relation, for generating the nth row from the (n − 1)th:
Theorem 1.81. For n, k ≥ 1,
S(n, k) = S(n − 1, k − 1) + k S(n − 1, k).

Proof. The proof is similar to the bijective proof we gave of Theorem 1.50.
By definition, S(n, k) counts the number of partitions of the set {1, 2, · · · , n}
into k blocks. Now the element n is either in a block by itself or it’s not.
Choosing a partition in which n is in a block by itself amounts to choosing a
partition of {1, 2, · · · , n − 1} into k − 1 blocks, so there are S(n − 1, k − 1)
such partitions. Choosing a partition in which n is not on its own amounts
to choosing a partition of {1, 2, · · · , n − 1} into k blocks and then choosing
which of these k blocks to put n in. So there are k S(n − 1, k) such partitions,
and the desired recurrence relation follows.

The coefficient k is the only change from Theorem 1.50 to this, but that is
enough to make a big difference. In terms of the triangle, Theorem 1.81 tells
us that each entry is the sum of the one above and to the left and the product
of the one directly above by the column number. So we can build more of
the Stirling triangle:
1
1 1
1 3 1
1 7 6 1
1 15 25 10 1
1 31 90 65 15 1
1 63 301 350 140 21 1
.. .. .. .. ..
. . . . .
CHAPTER 1. COUNTING PROBLEMS 43

With more data, we can spot some patterns in the Stirling numbers.
Example 1.82. The second column consists of the tower of Hanoi numbers,
i.e. those which are one less than a power of 2. The reason for this is that ev-
ery partition of {1, 2, · · · , n} into two blocks gives you two nonempty proper
subsets of {1, 2, · · · , n}, which are complements of each other. (“Proper”
means “not equal to the whole set”.) Moreover, every such subset A features
in exactly one partition with two blocks, namely the partition into A and
{1, 2, · · · , n} \ A. So by the Overcounting Principle, S(n, 2) is half the num-
ber of nonempty proper subsets of {1, 2, · · · , n}. Since the total number of
subsets of {1, 2, · · · , n} is 2n , there are 2n − 2 nonempty proper subsets, so
n
S(n, 2) = 2 2−2 = 2n−1 − 1 for all n ≥ 1.
Example 1.83. The diagonal just inside the right-hand edge coincides with
one of the diagonals of Pascal’s triangle. The reason for this is that in any
partition of {1, 2, · · · , n} into n − 1 blocks, there must be two elements in
one block, and every

other element is in its
own block. The two elements can
be chosen in 2 ways, so S(n, n − 1) = n2 for all n ≥ 1.
n

Although we do not have a closed formula for the Stirling numbers, Theorem
1.81 gives us a good handle on them, which we will be able to use more
effectively in later chapters. To the extent that we understand the Stirling
numbers, the following result is a better formula for the number of surjective
functions than Theorem 1.73:
Theorem 1.84. If X and Y are sets with |X| = m and |Y | = k, then the
number of surjective functions f : X → Y is k! S(m, k).

Proof. As observed before, every function f : X → Y gives rise to a way of

writing X as a disjoint union, namely as the disjoint union of the preimages
f −1 (y) for y ∈ Y (that is, you group together elements of X on which f takes
the same value). Saying that f is surjective is the same as saying that all
these preimages are nonempty, so in that case we get a partition of X into
k blocks. For a fixed partition of X into k blocks, there are k! surjective
functions f : X → Y which give rise to it, because once you have grouped
together the elements of X which get sent to the same element of Y by f ,
you just need to decide which block gets sent to which element of Y . So the
number of surjective functions is k! times the number of partitions into k
blocks.
44 Topics in Discrete Mathematics

Example 1.85. Recall from Example 1.77 the partitions of {1, 2, 3, 4} into
three blocks. Each of these gives rise to 3! = 6 surjective functions f :
{1, 2, 3, 4} → {1, 2, 3}: for instance, the partition 23|1|4 corresponds to the
six functions for which f (2) = f (3), but f (1) and f (4) are different from this
and from each other. There are six of these because this is the number of
ways of deciding which of 1, 2, 3 is to be f (2) = f (3), which is to be f (1),
and which is to be f (4).

Generalizing the idea of Theorem 1.84, we have a relationship between four

of the key sorts of numbers we have encountered in this chapter: integer
powers, factorials, binomial coefficients, and Stirling numbers.
Theorem 1.86*. For any n, m ∈ N,
m
X  
m n
n = k! S(m, k) .
k=0
k

Proof*. By Theorem 1.30, nm counts the functions f : {1, 2, · · · , m} →

{1, 2, · · · , n}. We can count these another way using the Sum Principle,
dividing the functions according to the size of the range f ({1, 2, · · · , m}).
If we stipulate that |f ({1, 2, · · · , m})| = k, then we can choose

what k-
n
element subset J of {1, 2, · · · , n} this range actually is in k ways. Having
chosen J, the choice of f amounts to the choice of a surjective function from
{1, 2, · · · , m} to J. By Theorem 1.73, there are k! S(m, k) of these. Putting
this all together, we get the formula in the statement.
Remark 1.87*. The k = 0 term of the sum in Theorem 1.86 is usually
zero, because S(m, 0) = 0 for m ≥ 1: this is put in purely for the case when
m = 0, when the theorem says that 1 = 1 × 1 × 1.
Chapter 2

Recursion and Induction

Recursion is a phenomenon that is widespread in discrete mathematics. It

occurs when the answer to some problem which depends on a nonnegative
integer n makes use of the answers to the same problem for smaller values
of n. In other words, if the answer to the problem for n is an , then the
sequence a0 , a1 , a2 , · · · satisfies a recurrence relation, which expresses an (for
sufficiently large values of n) as a function of the earlier terms in the sequence,
i.e. a0 , a1 , a2 , · · · , an−1 . In this chapter we will examine many such sequences,
and show how to prove facts about them by induction. What we regard as the
goal, usually, is to find a closed formula for an , one which does not involve the
earlier terms (and avoids anything equally inconclusive, like an unevaluated
sum of n things). This is what is meant by ‘solving’ a recurrence relation.

2.1 Examples of recursive sequences

A sequence a0 , a1 , a2 , a3 , · · · which satisfies a recurrence relation is said to be

recursive; every term to which the recurrence relation applies is determined
by the earlier ones. To get the whole thing rolling, there must be at least
one term at the beginning defined separately as an initial condition.

45
46 Topics in Discrete Mathematics

The simplest type is when each term depends only on the previous one, i.e. an
(for n ≥ 1) is a function of an−1 , and only a0 needs to be given as an initial
condition; this was the case in the tower of Hanoi problem, for example.
But there are also important cases where earlier terms of the sequence are
involved.
Example 2.1. One of the most famous of all recursive sequences was men-
tioned first in 1202 by the Italian mathematician Fibonacci. He imagined a
rabbit farm which on Day 1 contains only a single newborn breeding pair of
rabbits. On Day 2 these rabbits mature, so that on Day 3 the female gives
birth to a new pair of rabbits. On Day 4 the new pair is still maturing, but
the older female gives birth to a third pair. On Day 5 the newest pair is
still maturing, but the other two females give birth to new pairs, so there are
now 5 pairs, and so on. If we let Fn denote the number of pairs after n days,
Fibonacci’s assumptions imply that Fn −Fn−1 (the number of new pairs born
on Day n) equals Fn−2 (the number of pairs existing on Day n−2). Formally,
we define the Fibonacci sequence by

F0 = 0, F1 = 1, Fn = Fn−1 + Fn−2 for n ≥ 2. (2.1)

(Starting the sequence with a 0 makes various formulas nicer.) Notice that
because the recurrence relation doesn’t kick in until we come to F2 , we have
to define F0 and F1 as initial conditions. The Fibonacci sequence begins:

0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, · · ·

and we will prove many interesting facts about it as we go, including a

surprising closed formula for Fn in Example 2.24.
Example 2.2. Another famous sequence is the sequence of Catalan numbers
cn . These arise in an amazingly wide range of contexts: in the book ‘Enu-
merative Combinatorics’ Vol. 2 by R. P. Stanley, there is a list of 66 counting
problems, to all of which the answer is the Catalan numbers. The one which
is most often used as the definition is that cn is the number of balanced
strings of n left brackets and n right brackets, where “balanced” means that
as you read from left to right, the number of right brackets never exceeds the
number of left brackets. For instance, c3 = 5 because the balanced strings of
3 left brackets and 3 right brackets are as follows:

()()(), ()(()), (())(), (()()), and ((())).

CHAPTER 2. RECURSION AND INDUCTION 47

What’s more important for us, however, is the recursive definition:

c0 = 1, cn = c0 cn−1 + c1 cn−2 + · · · + cn−2 c1 + cn−1 c0 for n ≥ 1. (2.2)

As usual, the right-hand side of this recurrence relation is to be interpreted

according to the pattern rather than the literal sense for small values of n: it
always has n terms, so when n = 1 it is just c0 c0 , when n = 2 it is c0 c1 + c1 c0 ,
etc. It would probably be better to avoid this problem of interpretation by
writing the recurrence relation using sigma notation:
n−1
X
cn = cm cn−1−m .
m=0

The reason that the recurrence relation holds, if you define the Catalan
numbers using balanced strings of brackets, is that every such string must
have the form

(first balanced substring) second balanced substring,

i.e. there is a unique right bracket which corresponds to the initial left bracket,
and it must occur in the (2m + 2)th position for some m between 0 and n − 1;
the substring enclosed by these brackets, and the substring which follows
them, must both themselves be balanced. For any fixed m, the number of
ways of choosing the first balanced substring is cm , and the number of ways
of choosing the second balanced substring is cn−1−m ; the recurrence relation
follows. The Catalan sequence begins 1, 1, 2, 5, 14, 42, 132, 429, · · · .

Even some familiar sequences which we wouldn’t normally think of as re-

cursive really are. The tell-tale sign is a “ · · · ” in a formula, or a verbal
instruction to carry out some operation a variable number of times.
Example 2.3. If asked for the definition of 2n , most people would say:

2n = 2 × 2 × 2 × · · · × 2 (n twos).

In other words, 2n is the result of starting with 1 and doubling n times. Since,
in this procedure, you necessarily reach 2n−1 in the step before reaching 2n ,
the definition may as well be expressed using a recurrence relation:

20 = 1, 2n = 2 × 2n−1 for n ≥ 1. (2.3)

48 Topics in Discrete Mathematics

So you could argue that our solution of the tower of Hanoi problem was not
all it was cracked up to be: we just expressed one recursive sequence, hn , in
terms of another, 2n . But, of course, the latter is something we understand
so well that expressing hn in terms of it counts as a solution.
Example 2.4. Similarly, the “ · · · ” in the definition n! = n(n − 1) · · · 2 × 1
is a sign that we can define factorials recursively:
0! = 1, n! = n × (n − 1)! for n ≥ 1. (2.4)
Although the sequence of factorials is slightly more complicated than the
sequence of powers of 2, we still regard n! as something known, and we are
quite happy if we can express some other sequence in terms of factorials. For
instance, in Example 3.45 we will prove a famous formula for the Catalan
(2n)!
numbers: cn = (n+1)! n!
.
Example 2.5. In many problems we have some function f of a nonnegative
integer variable, and we then have to consider the function of n defined by
n
X
an := f (0) + f (1) + f (2) + · · · + f (n) = f (m).
m=0

The sequence a0 , a1 , a2 , · · · can obviously be defined recursively:

a0 = f (0), an = an−1 + f (n) for n ≥ 1. (2.5)
Conversely, any recurrence relation which is of this sum type, where each
term of the sequence an equals the previous term plus
Psome fixed function of
n
n, gives rise to the corresponding sequence of sums m=0 f (m), because we
can just unravel the recurrence:
an = an−1 + f (n)
= an−2 + f (n − 1) + f (n)
= an−3 + f (n − 2) + f (n − 1) + f (n)
..
.
= a0 + f (1) + f (2) + · · · + f (n − 1) + f (n)
= f (0) + f (1) + f (2) + · · · + f (n − 1) + f (n).

Such an unravelling procedure is often possible: the idea is that you sub-
stitute in the recursive formula for an the recursive formula for an−1 , and
CHAPTER 2. RECURSION AND INDUCTION 49

keep substituting until a pattern is clear, whereupon you jump to the point
at which the terms of the sequence disappear, leaving a large formula with
an inevitable “ · · · ”. This is not as helpful as you might think, because it is
often just rephrasing the recurrence relation in a more cumbersome way.
Example 2.6. Consider the sequence defined by
p
a0 = 1, an = 1 + an−1 for n ≥ 1.
If you unravel the recurrence relation, you get
p
an = 1 + an−1
q p
= 1 + 1 + an−2
r q p
= 1 + 1 + 1 + an−3
..
.
s r q
√ √
= 1+ 1 + 1 + ··· 1 + 1 (with n signs).

However, this way of writing an tells us nothing more than the recurrence
relation did; it certainly doesn’t count as a solution of the recurrence relation.

However, there are cases where the unravelling procedure helps to reveal that
the problem is of a kind you already know how to solve.
Example 2.7. The triangular number tn is the number of dots in a triangu-
lar array with n dots on each side. Since removing one of the sides produces
a smaller triangular array, this has a recursive definition of sum type:
t0 = 0, tn = tn−1 + n for n ≥ 1.
Unravelling this recurrence relation gives the non-closed formula

tn =
Summing the arithmetic progression gives the solution of the recurrence:

tn =
This sequence begins 0, 1, 3, 6, 10, 15, 21, 28, · · · .
50 Topics in Discrete Mathematics

Example 2.8. Consider the sequence defined by

a0 = 1, an = 3an−1 + 1 for n ≥ 1.

The unravelling procedure gives:

an = 3an−1 + 1 = 3(3an−2 + 1) + 1
= 32 an−2 + 3 + 1 = 32 (3an−3 + 1) + 3 + 1
= 33 an−3 + 32 + 3 + 1
..
.
= 3 a0 + 3n−1 + 3n−2 + · · · + 3 + 1
n

= 3n + 3n−1 + 3n−2 + · · · + 3 + 1.
This would be no real help, except that we recognize this as the sum of a
geometric progression, and hence obtain the closed formula
3n+1 − 1 3n+1 − 1
an = = ,
3−1 2
which counts as a solution of the recurrence.
Remark 2.9*. It is wise to be cautious about arguments, such as that in
Example 2.8, in which a chain of equalities goes on for a variable number
of steps. After all, if we are supposed to be writing a logical proof in which
each line is deducible from the previous one, are we really allowed to say to
the reader, in effect, “the next lot of lines all follow the same pattern, but
I can’t actually tell you how many there are”? Just as horizontal dots are
a sign that some definition is really recursive, vertical dots are a sign that
some proof is really a proof by induction (see Example 2.13).
Example 2.10. In Chapter 1 we encountered two important collections of
numbers which were more naturally displayed as a triangle than as a se-
quence, because they depended

on not one but two nonnegative integers: the
binomial coefficients nk and the Stirling numbers S(n, k). We can natu-
rally extract sequences from these numbers by fixing k, which corresponds
to looking at a single column of the Pascal or Stirling triangle (for instance,
the k = 2 column of Pascal’s triangle gives us the sequence of triangular
numbers). In both cases we have a recurrence relation (Theorems 1.50 and
1.81), which expresses an element of the kth column in terms of the previous
element of the kth column and also an element of the (k − 1)th column. So
CHAPTER 2. RECURSION AND INDUCTION 51

we can’t really consider the kth column on its own as a recursive sequence,
until we have a closed formula for the elements of the (k − 1)th column; this
suggests that we should analyse the columns from left to right. This idea is
rather pointless in
the case of Pascal’s triangle, because we already know the
closed formula nk = k!(n−k)!
n!
. But it will contribute to our understanding of
the Stirling numbers in the next chapter.

2.2 Proof by induction

The principle of mathematical induction, or induction for short, is one of the

most common and most powerful methods of proof in discrete mathematics.
The context in which it is used is that one wants to prove for all n ∈ N some
statement S(n) which depends on n. The principle of induction, in its most
general form, says that if we prove the following two statements:

(1) (base case) S(0) is true;

(2) (inductive step) for any n ≥ 1, the truth of S(n) follows from the truth
of S(0), S(1), S(2), · · · , S(n − 1),

then we have proved that S(n) is true for all n ∈ N.

The reason this works is that S(0) implies S(1), and then S(0) and S(1) to-
gether imply S(2), and then S(0), S(1), and S(2) together imply S(3), and so
on. So as long as the base case can be proved (which is often a simple check),
we can assume any or all of S(0), S(1), S(2), · · · , S(n − 1) in our attempts to
prove S(n); the assumption we make is called the induction hypothesis. This
is especially useful when S(n) involves the terms of a recursive sequence.
Example 2.11. Consider the sequence defined by

a0 = 2, an = a2n−1 + 1 for n ≥ 1.

The sequence begins 2, 5, 26, 677, · · · . To give some idea of how fast it grows,
n
we want to prove that an ≥ 22 for all n ∈ N. We do this by induction, making
n
S(n) the statement an ≥ 22 . The base case S(0) asserts that 2 ≥ 21 , which
52 Topics in Discrete Mathematics

is clearly true. Then in proving S(n) for n ≥ 1, we are allowed to assume

that S(0), S(1), · · · , S(n−1) are true. Actually we only need to use the truth
of S(n − 1) in the proof:
n−1 n n
an = a2n−1 + 1 ≥ (22 )2 + 1 = 22 + 1 ≥ 22 .
n
This completes the inductive step, so an ≥ 22 is proved for all n ∈ N.
Example 2.12. Suppose that S(n) is the inequality Fn < 2n . To start the
induction, we check that S(0) is true: S(0) asserts that 0 < 1, so this is
obvious. Now we want to carry out the induction step, in which we prove
S(n) having made the assumption that S(0), S(1), · · · , S(n−1) are true. The
natural starting point is to apply the recurrence relation Fn = Fn−1 + Fn−2 .
Conveniently, our inductive hypothesis includes the statements S(n − 1) and
S(n − 2), which say that Fn−1 < 2n−1 and Fn−2 < 2n−2 . We conclude that
1 1 3
Fn < 2n−1 + 2n−2 = × 2n + × 2n = × 2n < 2n .
2 4 4
This appears to completes the inductive step. However, there is a gap in the
proof: to use the Fibonacci recurrence, we need to assume that n ≥ 2, so
we haven’t yet covered the n = 1 case. This is a dangerous gap, because in
an induction proof, all the statements S(n) for n ≥ 2 rely on the truth of
S(1). Fortunately, S(1) just says that 1 < 2, so it is obviously true and the
proof is finished. If you like, you can regard S(1) as an extra base case which
we should have checked along with S(0); alternatively, the induction step is
divided into two cases depending on whether n = 1 or n ≥ 2.

Perhaps because assuming smaller versions of what you are trying to prove
feels a bit like cheating, there is an aesthetic (rather than logical) feeling
that the induction hypothesis should be restricted as much as possible. As
the preceding examples suggest, if you are proving a statement about a
recursively-defined sequence an where the recurrence relation expresses an
in terms of an−1 , an−2 , · · · , an−k for some k, you probably only need to ap-
peal to S(n − 1), S(n − 2), · · · , S(n − k) in proving S(n). In particular, if an
is a function of an−1 only, then assuming S(n − 1) is usually enough to prove
S(n), as it was in Example 2.11; this form of the argument is sometimes dis-
tinguished by the special term “weak induction”. However, since the proof
of S(n − 1) relies on S(n − 2), whose proof relies on S(n − 3), and so on, all
the earlier statements are still implicitly involved.
CHAPTER 2. RECURSION AND INDUCTION 53

Example 2.13*. After defining a recurrence relation of sum type in (2.5),

we showed using an unravelling
Pn argument that it gave the same sequence
as the definition an = m=0 f (m). If you share the doubts expressed in
Remark 2.9 about the unravelling, the true justification is the following proof
by induction. The base case is the statement that a0 = f (0), which is given
as the initial condition.PIn proving the statement for n ≥ 1, we are allowed
to assume that an−1 = n−1 m=0 f (m) is true, and then the recurrence gives

n−1
! n
X X
an = f (m) + f (n) = f (m),
m=0 m=0

completing the inductive step and finishing the proof. The ease of such
induction proofs means that we usually don’t worry about these matters.

Of course, the same principles apply to statements which are to be proved

for all n ≥ 1, or all n ≥ 2, etc. You just have to start with the appropriate
base case. (Actually, you could put these variants in the n ∈ N framework,
by defining S 0 (n) to be S(n + 1) or S(n + 2), etc., but that’s pedantic.)
Example 2.14. We prove by induction that 2n < n! for all n ≥ 4. The base
case is that 24 < 4!, i.e. 16 < 24. Now we can assume that n ≥ 5, and that
2n−1 < (n − 1)! is true; then

2n = 2 × 2n−1 < 2 × (n − 1)! < n × (n − 1)! = n!,

completing the induction step and finishing the proof.

Example 2.15*. A famous example of induction is the proof that every
integer n ≥ 2 has a divisor which is prime. (Recall that a positive integer
n is said to be prime if it is bigger than 1 and has no divisors other than
1 and n.) For the base case we need to show that 2 has a divisor which
is prime: but clearly 2 itself is prime, and a number is always a divisor of
itself. Now we can assume that n ≥ 3, and (as our induction hypothesis)
that every number from 2 up to n − 1 has a prime divisor. If n itself is prime,
the statement is automatically true. If n is not prime, then it must have
some divisor d which is neither 1 nor n. The induction hypothesis includes
the fact that d has a prime divisor, say e. Since e is a divisor of d, and d is
a divisor of n, e is a divisor of n, so the statement is true in this case also.
This completes the proof. You may think that this fact can be easily proved
without induction, as follows: if n is prime we are finished; otherwise, we can
54 Topics in Discrete Mathematics

write n as the product of two smaller numbers which divide it; if either of
those is prime we are finished; otherwise, we write them in turn as products
of smaller numbers; and keep going in this way until we break everything
into a product of primes. There is nothing wrong with this, except that the
phrase “keep going in this way until” falls foul of the same sort of objection as
in Remark 2.9. Really it requires an induction argument to make it rigorous.

The availability of proof by induction is very useful in solving recurrence

relations. One great advantage that recurrence relations have over the some-
what similar differential equations is that you can directly work out as many
terms of the sequence as you have the patience to do. There may be a pattern
which suggests a closed formula for the general term, and you can then try
to prove this formula by induction.
Example 2.16. (For this example, forget that you know how to sum an
arithmetic progression.) Consider the sequence defined by

a0 = 0, an = an−1 + (2n − 1) for n ≥ 1.

This is a recurrence of the sum type, in the terminology introduced in the

previous section. After unravelling, it says that an is the sum of the first n
odd numbers:
an = 1 + 3 + 5 + · · · + (2n − 1).
Direct calculation shows that the sequence starts 0, 1, 4, 9, 16, 25, · · · , from
which it is not hard to guess the formula an = n2 . So we make this our
statement S(n), and aim to prove it by induction. The base case S(0) is
given to us as the initial condition, so all that remains is to prove S(n) for
n ≥ 1, assuming S(n − 1). This is easy:

an = an−1 + (2n − 1) = (n − 1)2 + 2n − 1 = n2 − 2n + 1 + 2n − 1 = n2 .

Notice that what the proof boils down to is the fact that the sequence of
squares satisfies the same initial condition and recurrence relation as an , and
therefore must be the same sequence.

The last comment in Example 2.16 is a general principle, almost trivial but
worth making explicit: if two sequences satisfy the same initial conditions
and the same recurrence relation, they must be the same. Formally:
CHAPTER 2. RECURSION AND INDUCTION 55

Theorem 2.17. Suppose that a0 , a1 , a2 , · · · and b0 , b1 , b2 , · · · are two se-

quences, the first of which satisfies some initial conditions:

a0 = I0 , a1 = I1 , a2 = I2 , · · · , ak−1 = Ik−1 ,

and some recurrence relation:

an = R(n) (a0 , a1 , · · · , an−1 ), for all n ≥ k,

where k ≥ 1 is fixed but the function R(n) possibly varies with n. If the
second sequence also satisfies the same initial conditions:

b0 = I0 , b1 = I1 , b2 = I2 , · · · , bk−1 = Ik−1 ,

and the same recurrence relation:

bn = R(n) (b0 , b1 , · · · , bn−1 ), for all n ≥ k,

then an = bn for all n ∈ N.

Proof. The reason is the principle of induction: we are given that an = bn

for all 0 ≤ n ≤ k − 1, which gives us all the base cases we could possibly
need. Then in proving an = bn , we can assume that n ≥ k and that a0 = b0 ,
a1 = b1 , and so on up to an−1 = bn−1 . We then have

an = R(n) (a0 , a1 , · · · , an−1 ) = R(n) (b0 , b1 , · · · , bn−1 ) = bn ,

completing the induction step.

So if we have guessed a formula for an , say f (n), and we want to prove that
it really is true that an = f (n) for all n ∈ N, all we need to do is to check
that f (n) satisfies the same initial conditions and recurrence relation that
an does. Although induction need not be explicitly mentioned, this will boil
down to an induction proof, since Theorem 2.17 is proved by induction.
Example 2.18. Consider the sequence defined by

1
a0 = 0, an = an−1 + for n ≥ 1.
n(n + 1)
56 Topics in Discrete Mathematics

This recurrence relation is of sum type, so we could alternatively write

1 1 1 1
an = + + ··· + + . (2.6)
1×2 2×3 (n − 1)n n(n + 1)

Working out the first few terms of the sequence, we find that
1 2 3 4
a1 = , a 2 = , a 3 = , a 4 = ,
2 3 4 5
from which we naturally guess that an = n+1 n
for all n ∈ N. To prove this, we
n
just need to show that the candidate formula n+1 satisfies the initial condition
and the recurrence relation. The initial condition is obvious: 01 = 0. For the
n
recurrence relation we need to show that n+1 = n−1n
1
+ n(n+1) , which is true
by a trivial calculation. So the claim is proved. In fact, (2.6) is an example
1
of a “telescoping sum”: since k(k+1) = k1 − k+1
1
, the right-hand side becomes

1 1 1 1 1 1 1 1
− + − + ··· + − + − ,
1 2 2 3 n−1 n n n+1
and every term cancels out except for the first and last, giving us an =
1 n
1 − n+1 = n+1 . Although such telescoping arguments are aesthetically more
pleasing, they really boil down to the same thing as the induction method
described above: the cancellation of an unspecified number of terms needs
induction to justify it properly.
Example 2.19. The formulas for the sum of an arithmetic progression and
the sum of a geometric progression can both be proved by this induction
method. In the geometric case, we have to prove that for all r 6= 1,

rn+1 − 1
1 + r + r2 + r3 + · · · + rn−1 + rn = , for n ∈ N. (2.7)
r−1
The left-hand side can be viewed as the an term of a sequence defined by
a recurrence relation of ‘sum type’, namely an = an−1 + rn , with initial
condition a0 = 1. The right-hand side is our candidate formula. That it
satisfies the initial condition is easy to see, and then all that remains is to
prove that, for n ≥ 1,

rn+1 − 1 rn − 1
= + rn .
r−1 r−1
CHAPTER 2. RECURSION AND INDUCTION 57

This is trivial, after multiplying both sides by r − 1. Of course, most people

would prove (2.7) by multiplying both sides by r − 1 in the first place; but
this results in a telescoping sum on the left-hand side, and so the principle
of induction would be implicitly involved.
Example 2.20. Consider the sequence defined by
n−1
a0 = 1, an = an−1 (22 + 1) for n ≥ 1.

Work out the next few terms:

a1 =

a2 =

a3 =

Can you guess a closed formula that fits all this data?

an =

To prove your formula is correct, all that remains is to check that it satisfies
the recurrence relation. Write this out to make sure:

for n ≥ 1.

Example 2.21. One formula which is surprisingly easy to guess is for the
sum of the first n positive integer cubes:

an = 13 + 23 + 33 + · · · + n3 .

Rephrasing the definition recursively:

a0 = 0, an = an−1 + n3 for n ≥ 1.

Working out the first few terms, we find that

a1 = 1, a2 = 9, a3 = 36, a4 = 100, a5 = 225,

58 Topics in Discrete Mathematics

and notice that all the values are squares. Moreover, the square roots seem
to be the triangular numbers, so we conjecture that
 2
n(n + 1) n2 (n + 1)2
an = = .
2 4
This formula obviously satisfies the initial condition, so all that remains to
prove is that
n2 (n + 1)2 (n − 1)2 n2
= + n3 ,
4 4
which follows from some straightforward manipulation. To sum up, in this
example we have shown that
n
X n2 (n + 1)2
m3 = , for all n ∈ N. (2.8)
m=0
4

Example 2.22. To show that guessing a formula is not always straightfor-

ward, consider the sequence defined by summing the first n positive squares,
whose recursive definition is
a0 = 0, an = an−1 + n2 for n ≥ 1.
Working out the first few terms, we find that
a1 = 1, a2 = 5, a3 = 14, a4 = 30, a5 = 55, a6 = 91,
and no very obvious pattern emerges (although if you look at the divisors of
these numbers, and spot that an seems to be often divisible by 2n + 1, you
may get on the right track). We will solve this recurrence in Example 3.47.
Example 2.23*. Consider the sequence defined by
1
a0 = 1, an = (an−1 + 2an−2 + 3an−3 + · · · + na0 ) for n ≥ 1.
n
Working out the next few terms, we find that
3 13 73 501
a1 = 1, a2 = , a3 = , a4 = , a5 = .
2 6 24 120
You have probably spotted that the denominators here are the factorials;
this suggests the general conjecture that n!an is a positive integer, which is
easy to prove by induction. But the numerators remain a bit mysterious for
now. We will return to this sequence in Example 3.51.
CHAPTER 2. RECURSION AND INDUCTION 59

Example 2.24*. Here is a formula for the Fibonacci numbers that seems
at first sight impossible to guess:
" √ !n √ !n #
1 1+ 5 1− 5
Fn = √ − . (2.9)
5 2 2
√
With all these fractions and 5’s, it’s even surprising that the right-hand
side works out to be an integer. Nevertheless, following the principle of
Theorem 2.17, we can prove this is a correct formula just by checking that
it satisfies the initial conditions and recurrence relation which define the
Fibonacci sequence. For the initial conditions:
 
√ !0 √ !0
1 1+ 5 1− 5  1
n=0: √  − = √ [1 − 1] = 0,
5 2 2 5
 
√ !1 √ !1 √
1  1+ 5 1− 5  5
n=1: √ − = √ = 1.
5 2 2 5

For the recurrence relation, in which n ≥ 2:

" √ !n √ !n #
1 1+ 5 1− 5
√ −
5 2 2
 
√ !2 √ !n−2 √ !2 √ !n−2
1 1+ 5 1+ 5 1− 5 1− 5
=√  − 
5 2 2 2 2
 
√ ! √ !n−2 √ ! √ !n−2
1 6+2 5 1+ 5 6−2 5 1− 5
=√  − 
5 4 2 4 2
 ! 
√ √ !n−2 √ ! √ !n−2
1 1+ 5 1+ 5 1− 5 1− 5
=√  +1 − +1 
5 2 2 2 2
 
√ !n−1 √ !n−1
1 1+ 5 1− 5
=√  − 
5 2 2
 
√ !n−2 √ !n−2
1 1+ 5 1− 5
+√  − .
5 2 2
60 Topics in Discrete Mathematics

2.3 Homogeneous linear recurrence relations

There is a general method for solving (constant-coefficient) linear recurrence

relations which is strongly reminiscent of the solution of linear ordinary dif-
ferential equations. As in that theory, there are two kinds, homogeneous and
non-homogeneous.

Definition 2.25. If k is a positive integer, a kth-order homogeneous linear

recurrence relation for a sequence an is one of the form

an = r1 an−1 + r2 an−2 + · · · + rk an−k , for all n ≥ k,

where r1 , · · · , rk are fixed numbers (independent of n), and rk 6= 0.

Remember that such a recurrence relation on its own is not enough to specify
the sequence: one also needs to give the first k terms a0 , a1 , · · · , ak−1 as initial
conditions. But the crucial idea is to consider all the sequences satisfying
the same recurrence relation together.
Example 2.26. A first-order homogeneous linear recurrence relation is of
the form an = ran−1 for some nonzero r. To specify the sequence, one
needs to give the value of a0 as an initial condition. If a0 = C, it is clear
that an = Crn for all n ∈ N. (Strictly speaking, we need an application
of Theorem 2.17: the sequence Crn satisfies the right initial condition and
recurrence relation, therefore it equals an .) So we have solved the first-order
homogeneous case.
Example 2.27. A second-order homogeneous linear recurrence relation is
of the form an = ran−1 + san−2 for some numbers r, s with s 6= 0 (because
having s = 0 would make it actually just first-order). For instance, the
Fibonacci sequence satisfies such a recurrence relation, with r = s = 1. The
philosophy now is that we should consider the Fibonacci sequence together
with all other solutions of an = an−1 + an−2 for which the initial conditions
are different. An example of another such is the Lucas sequence defined by

L0 = 2, L1 = 1, Ln = Ln−1 + Ln−2 for n ≥ 2. (2.10)

This sequence begins 2, 1, 3, 4, 7, 11, 18, 29, · · · .

CHAPTER 2. RECURSION AND INDUCTION 61

Example 2.28. Many of the recurrence relations we encountered in previous

sections are not covered by the above definition. For instance, the Catalan
recurrence relation cn = c0 cn−1 + · · · + cn−1 c0 is ruled out, because it involves
products of two terms of the sequence; the recurrence relation in Example
2.23, an = n1 (an−1 + 2an−2 + 3an−3 + · · · + na0 ), is ruled out because the
coefficients, and the number of terms, depend on n.

The main reason that linear recurrence relations are nicer is that if you have
two solutions, a linear combination of them is also a solution:
Theorem 2.29. If a0 , a1 , a2 , · · · and b0 , b1 , b2 , · · · are two sequences satisfying
the same kth-order homogeneous linear recurrence relation:
an = r1 an−1 + r2 an−2 + · · · + rk an−k , for all n ≥ k, and
bn = r1 bn−1 + r2 bn−2 + · · · + rk bn−k , for all n ≥ k,
then for any constants C1 and C2 , the sequence
C1 a0 + C2 b0 , C1 a1 + C2 b1 , C1 a2 + C2 b2 , · · ·
also satisfies this recurrence relation.

Proof. This is trivial: taking C1 times the first equation plus C2 times the
second equation immediately gives us
C1 an + C2 bn = r1 (C1 an−1 + C2 bn−1 ) + r2 (C1 an−2 + C2 bn−2 )
+ · · · + rk (C1 an−k + C2 bn−k ), for all n ≥ k,
which is the desired recurrence relation.

Exactly the same principle applies to a linear combination of three or more

solutions.
Remark 2.30*. In the terminology of linear algebra, what we have actually
shown here is that the set of all sequences which satisfy a given homogeneous
linear recurrence relation forms a vector space (in fact, a vector subspace of
the vector space of all sequences).

Thanks to Theorem 2.29, we can hope to build up a general solution by

finding special solutions and taking linear combinations of them. Inspired
62 Topics in Discrete Mathematics

by the first-order case, we look for special solutions of the form an = λn and
slight modifications thereof, and the next result tells us which λ to take.
Definition 2.31. The characteristic polynomial of the recurrence relation
an = r1 an−1 + r2 an−2 + · · · + rk an−k is
xk − r1 xk−1 − r2 xk−2 − · · · − rk−1 x − rk .
This is a polynomial of degree k in the indeterminate x.
Theorem 2.32. Let λ be a (possibly complex) root of the characteristic
polynomial.

(1) The sequence an = λn satisfies the recurrence relation

an = r1 an−1 + r2 an−2 + · · · + rk an−k , for all n ≥ k.

(2) If λ is a repeated root of the characteristic polynomial of multiplicity m,

then an = ns λn also satisfies this recurrence relation for 1 ≤ s ≤ m − 1.

Proof*. The fact that λ is a root of the characteristic polynomial means

that
λk − r1 λk−1 − r2 λk−2 − · · · − rk−1 λ − rk = 0. (2.11)
Moving the terms with the − sign to the other side and multiplying both
sides by λn−k , we find that
λn = r1 λn−1 + r2 λn−2 + · · · + rk λn−k , for all n ≥ k, (2.12)
which is exactly what part (1) claims. The proof of part (2) is slightly harder.
We need to show that λ is a root of the polynomial
ns xn − r1 (n − 1)s xn−1 − r2 (n − 2)s xn−2 − · · · − rk (n − k)s xn−k , (2.13)
for all n ≥ k and 1 ≤ s ≤ m − 1. What we know is that (x − λ)m divides the
characteristic polynomial, and hence also divides
xn − r1 xn−1 − r2 xn−2 − · · · − rk xn−k . (2.14)
But the polynomial (2.13) can be obtained from (2.14) by applying, s times,
the operation of differentiating and then multiplying by x. Each time we
apply this operation can only reduce by 1 the power of x − λ which divides
the polynomial, so at the end it remains divisible by (x−λ)m−s , and therefore
λ is a root of (2.13).
CHAPTER 2. RECURSION AND INDUCTION 63

Example 2.33. We can use Theorems 2.29 and 2.32 to solve the following
second-order recurrence relation:

a0 = 2, a1 = 5, an = 7an−1 − 12an−2 , for n ≥ 2.

The characteristic polynomial is x2 −7x+12, which factorizes as (x−3)(x−4);

so its roots are 3 and 4. By Theorem 2.32, both the sequence 3n and the
sequence 4n satisfy our recurrence relation (although, clearly, neither of them
satisfies our initial conditions). So by Theorem 2.29, any sequence with
formula C1 3n + C2 4n , where C1 and C2 are constants (i.e. independent of
n), also satisfies this recurrence relation. It so happens that we can find
constants C1 and C2 such that the initial conditions hold also. For this we
need to have the following:

C1 30 + C2 40 = 2, i.e. C1 + C2 = 2, and
C1 31 + C2 41 = 5, i.e. 3C1 + 4C2 = 5.

This is a system of two linear equations in the two unknowns C1 and C2 ,

and it is straightforward to find the unique solution: C1 = 3, C2 = −1.
(For example, the second equation minus three times the first equation tells
us that C2 = −1, and then C1 = 3 follows by substituting this in the first
equation.) Since 3 × 3n + (−1) × 4n satisfies the same recurrence relation and
initial conditions as an , we must have

an = 3 × 3n + (−1) × 4n = 3n+1 − 4n .

So we have solved the original recurrence relation.

Example 2.34. Consider the sequence defined by

a0 = 1, a1 = 4, an = 4an−1 − 4an−2 for n ≥ 2.

The characteristic polynomial is which has

repeated root . By Theorems 2.32 and 2.29, any sequence an

which the form an = for constants C1 and C2

must satisfy an = 4an−1 − 4an−2 . To find constants which make the initial
64 Topics in Discrete Mathematics

conditions satisfied also, we rewrite a0 = 1 and a1 = 4 as linear equations in

C1 and C2 :

The unique solution of this system of linear equations is C1 = ,

C2 = . Hence the solution of the original recurrence relation is

an =

Example 2.35*. Consider the sequence defined by

a0 = 2, a1 = 2, an = 2an−1 − 2an−2 for n ≥ 2.
The characteristic polynomial is x2 − 2x + 2, whose roots are complex: 1 ± i.
Thus both (1 + i)n and (1 − i)n satisfy the recurrence relation, and hence so
does C1 (1 + i)n + C2 (1 − i)n for any constants C1 and C2 . The latter satisfies
the initial conditions if and only if
C1 + C2 = 2,
(1 + i)C1 + (1 − i)C2 = 2,
which has unique solution C1 = C2 = 1. So we have an = (1 + i)n + (1 − i)n .
We can rewrite this in a slightly better way, which removes
√ ±i π the need for
complex numbers, if we recall the polar forms 1 ± i = 2 e 4 :
√ n nπ nπ √ n nπ
an = (1 + i)n + (1 − i)n = 2 (ei 4 + e−i 4 ) = 2 2 cos( ).
4
Thus the sequence an is almost periodic with period 8; whenever the argu-
√ 8
ment of cos recurs, the value gets multiplied by 2 = 16. If we had worked
out the first few terms of the sequence, the pattern would have emerged:
2, 2, 0, −4, −8, −8, 0, 16, 32, 32, 0, −64, −128, −128, · · · .

In all of these examples we were able to find a linear combination of the

solutions provided by Theorem 2.32 which satisfied the right initial conditions
to be the sequence we were looking for. This was no coincidence, because it
always happens:
CHAPTER 2. RECURSION AND INDUCTION 65

Theorem 2.36. The general solution of the kth-order homogeneous linear

recurrence relation

an = r1 an−1 + r2 an−2 + · · · + rk an−k , for all n ≥ k,

is a linear combination of the k solutions arising from the roots of the char-
acteristic polynomials via parts (1) and (2) of Theorem 2.32. In particular,
the general solution of the second order recurrence relation

an = ran−1 + san−2 , for all n ≥ 2,

is as follows:

(1) if x2 − rx − s = (x − λ1 )(x − λ2 ) where λ1 6= λ2 (distinct roots),

an = C1 λn1 + C2 λn2 for some constants C1 and C2 ;

(2) if x2 − rx − s = (x − λ)2 (repeated root),

an = C1 λn + C2 nλn for some constants C1 and C2 .

Proof**. First note that we do indeed always get k solutions to the re-
currence relation out of parts (1) and (2) of Theorem 2.32: this is because
every root λ gives rise to the m solutions ns λn , 0 ≤ s ≤ m − 1, where m
is the multiplicity (which is 1 if λ is not a repeated root), and the sum of
the multiplicities of all the roots equals the degree of the polynomial, which
is k. (This relies on the fact that we are allowing the roots to be complex
numbers, so the polynomial factorizes completely.)

What remains to be proved is that no matter what initial conditions a0 =

I0 , a1 = I1 , · · · , ak−1 = Ik−1 we impose, there is always a linear combination
of these k solutions which satisfies them. As in the examples, this amounts
to showing that there is a solution to a system of k linear equations in the
k unknown coefficients. By a basic theorem in linear algebra, such a system
has a solution if and only if a certain determinant is nonzero. Proving this in
general requires some knowledge of Vandermonde determinants, which would
take us too far off the main topic, so we will just go through the proof in the
case when k = 2.
66 Topics in Discrete Mathematics

In this case the initial conditions are a0 = I0 and a1 = I1 , the recurrence

relation is an = ran−1 + san−2 , and the characteristic polynomial is the
quadratic x2 − rx − s. First suppose that this has distinct roots λ1 and λ2 ;
then our linear combination is C1 λn1 + C2 λn2 , and we are trying to show that
there is always a choice of C1 and C2 which satisfies the initial conditions.
The system of linear equations this amounts to is:

C1 + C2 = I0 ,
λ1 C1 + λ2 C2 = I1 .
 
1 1
Because the determinant of the matrix is λ2 − λ1 , which we are
λ1 λ2
assuming is nonzero, there is a unique solution. In fact, that solution is:

λ2 I0 − I1 I1 − λ1 I0
C1 = , C2 = . (2.15)
λ2 − λ1 λ2 − λ1

(It’s probably not worth memorizing these formulas, given how easy it is to
solve any particular 2×2 linear system.) Now suppose that the characteristic
polynomial has a repeated root, i.e. x2 − rx − s = (x − λ)2 . Notice that we
have r = 2λ, s = λ2 . Our linear combination is C1 λn + C2 nλn , and the
equations we need to solve are

C1 = I0 ,
λC1 + λC2 = I1 .
 
1 0
Because the determinant of the matrix is λ, and λ must be nonzero
λ λ
because s 6= 0 (or else we wouldn’t be dealing with a second-order recurrence
relation at all), there is a unique solution, namely:

I1 − λI0
C1 = I0 , C2 = . (2.16)
λ
So in either case the sequence an must be expressible in the right form.
Remark 2.37. Notice that the first-order case explained in Example 2.26 is
also covered by Theorem 2.36: in this case the characteristic polynomial is
the degree-1 polynomial x − r, whose sole root is r.
CHAPTER 2. RECURSION AND INDUCTION 67

Example 2.38. We can now explain the strange formula for the Fibonacci
numbers that we proved in Example 2.24. The characteristic polynomial
√ for
the√ Fibonacci recurrence is x2 − x − 1, which has distinct roots 1+2 5 and
1− 5
2
. By Theorem 2.36, this is enough to prove that
√ !n √ !n
1+ 5 1− 5
Fn = C1 + C2 , for all n ∈ N,
2 2

for some constants C1 and C2 . The initial conditions F0 = 0 and F1 = 1 give

C1 + C2 = 0,
√ ! √ !
1+ 5 1− 5
C1 + C2 = 1,
2 2

a system whose solution is C1 = √1 , C2 = − √15 . Hence we derive (2.9).

5

Example 2.39. Recall from (2.10) that the Lucas numbers satisfy the same
recurrence relation as the Fibonacci numbers, but different initial conditions.
So they must also be of the same form:
√ !n √ !n
1+ 5 1− 5
Ln = C1 + C2 , for all n ∈ N,
2 2

for some constants C1 and C2 . The initial condition L0 = 2 and L1 = 1 give

C1 + C2 = 2,
√ ! √ !
1+ 5 1− 5
C1 + C2 = 1,
2 2

a system whose solution is C1 = C2 = 1. Hence we get a formula for the

Lucas numbers: √ !n √ !n
1+ 5 1− 5
Ln = + . (2.17)
2 2
As with the Fibonacci formula, if you just saw the right-hand side, it would
be quite surprising that it always worked out to be a positive integer.
Example 2.40. Consider the ‘stupid’ second-order recurrence relation

a0 = I0 , a1 = I1 , an = an−2 for n ≥ 2,
68 Topics in Discrete Mathematics

where I0 and I1 are some initial values. Obviously the solution is that an
is either I0 or I1 depending on whether n is even or odd; how does this fit
the framework of Theorem 2.36? The characteristic polynomial is x2 − 1 =
(x−1)(x+1), so the general solution is an = C1 1n +C2 (−1)n = C1 +C2 (−1)n .
To fit the initial conditions we must have C1 = I0 +I
2
1
and C2 = I0 −I
2
1
, so
(
I0 + I1 I0 − I1 I0 , if n is even,
an = + (−1)n =
2 2 I1 , if n is odd,

as expected.

2.4 Non-homogeneous linear recurrence rela-

tions

Definition 2.41. A kth-order non-homogeneous linear recurrence relation

for a sequence an is one of the form

an = r1 an−1 + r2 an−2 + · · · + rk an−k + f (n), for all n ≥ k,

where r1 , · · · , rk are fixed numbers (independent of n), rk 6= 0, and f (n) is

some nonzero function of n.

What makes this “non-homogeneous” is the extra term f (n), which is not of
the same kind as all the others.
Example 2.42. Recall the recurrence relation from the tower of Hanoi prob-
lem in the Introduction:

hn = 2hn−1 + 1, for all n ≥ 1.

This can be viewed as a first-order non-homogeneous recurrence relation,

with the 1 being the extra term.
Example 2.43. Suppose we want to find a closed formula for an , defined
recursively by

a0 = 3, a1 = 5, an = 2an−1 + 3an−2 + 8n − 4 for n ≥ 2.

CHAPTER 2. RECURSION AND INDUCTION 69

The recurrence relation here is non-homogeneous because of the extra term

8n − 4. So we can’t directly apply the characteristic-polynomial method.
Example 2.44*. A general first-order non-homogeneous recurrence relation
has the form
an = ran−1 + f (n), for all n ≥ 1.
Unravelling this, we obtain
an = ran−1 + f (n)
= r2 an−2 + rf (n − 1) + f (n)
= r3 an−3 + r2 f (n − 2) + rf (n − 1) + f (n)
..
.
= r a0 + rn−1 f (1) + rn−2 f (2) + · · · + rf (n − 1) + f (n),
n

so solvingPthe recurrence relation amounts to finding a closed formula for

n
the sum m=1 rn−m f (m). Currently we only know how to do this in special
cases such as when f (n) is a constant function, where we can use the formula
(2.7) for the sum of a geometric progression.

The key to tackling the non-homogeneous case is a very simple observation:

Theorem 2.45. If a0 , a1 , a2 , · · · and p0 , p1 , p2 , · · · are two sequences satisfy-
ing the same kth-order non-homogeneous linear recurrence relation:
an = r1 an−1 + r2 an−2 + · · · + rk an−k + f (n), for all n ≥ k, and
pn = r1 pn−1 + r2 pn−2 + · · · + rk pn−k + f (n), for all n ≥ k,
then the sequence an −pn satisfies the corresponding homogeneous recurrence
relation (with the extra term omitted). So the general solution of the non-
homogeneous relation has the form

an = b n + p n ,

where bn is an arbitrary solution of the homogeneous recurrence relation and

pn is one particular solution of the non-homogeneous recurrence relation.

Proof. Subtracting the two equations, we get

an − pn = r1 (an−1 − pn−1 ) + r2 (an−2 − pn−2 ) + · · · + rk (an−k − pn−k ),

70 Topics in Discrete Mathematics

because the f (n) term cancels out. This shows that an − pn is a solution of
the homogeneous recurrence relation; letting it be bn , we have an = bn + pn
as claimed. Conversely, given any solution bn of the homogeneous recurrence
relation, the sequence an defined by an = bn + pn does satisfy the non-
homogeneous recurrence relation, because pn does and the subtraction of the
two equations is true.
Remark 2.46*. This theorem implies that the set of all solutions of the
non-homogeneous recurrence relation forms an affine subspace of the vector
space of all sequences (a vector subspace translated so that it doesn’t pass
through the origin).

In the parallel theory of differential equations, bn is called the ‘complementary

function’. We know the general form of bn from Theorem 2.36, so we can
substitute that in the equation an = bn + pn to get the general form of
a solution of the non-homogeneous recurrence relation. Then we can once
again use the initial conditions to determine the undetermined constants.

The problem, of course, is that we still have to find one particular solution
pn . Indeed, we may seem not to have made any progress: in trying to find
the solution an of the non-homogeneous recurrence relation, we are reduced
to finding a solution pn of the non-homogeneous recurrence relation. But the
point is that pn doesn’t have to satisfy the initial conditions that an does,
which gives us more scope. Indeed, pn can often be found by something
approaching guesswork. If the extra term f (n) in the recurrence is a polyno-
mial in n of degree d, it is a good idea to try for a solution pn which is also a
polynomial in n of degree d; let the coefficients be unknowns, and see what
equations they need to satisfy.
Example 2.47. Return to the sequence from Example 2.43:

a0 = 3, a1 = 5, an = 2an−1 + 3an−2 + 8n − 4 for n ≥ 2.

To find a closed formula for an , we first have to find the general solution of
the recurrence relation an = 2an−1 + 3an−2 + 8n − 4. So we forget about the
initial conditions a0 = 3 and a1 = 5 for now.

According to Theorem 2.45, an = bn + pn where bn is a solution of the

homogeneous recurrence relation bn = 2bn−1 + 3bn−2 and pn is a particular
CHAPTER 2. RECURSION AND INDUCTION 71

solution of the non-homogeneous recurrence relation. Since the characteristic

polynomial x2 − 2x − 3 = (x − 3)(x + 1), the general form of bn is C1 3n +
C2 (−1)n , where C1 and C2 are constants.

Since the extra term 8n − 4 is a degree-1 polynomial in n, we try for a

particular solution of the form pn = C3 n + C4 , where C3 and C4 are two
more constants. For this to satisfy the non-homogeneous recurrence relation,
we need to have the following equation for all n ≥ 2:
C3 n + C4 = 2(C3 (n − 1) + C4 ) + 3(C3 (n − 2) + C4 ) + 8n − 4
= (2C3 + 3C3 + 8)n + (−2C3 + 2C4 − 6C3 + 3C4 − 4)
= (5C3 + 8)n + (−8C3 + 5C4 − 4).
For this to hold for all n ≥ 2, it has to be true that the coefficients of n1 = n
and n0 = 1 on both sides are the same. So we want C3 and C4 to satisfy
C3 = 5C3 + 8 and C4 = −8C3 + 5C4 − 4. The unique solution is C3 = −2
and C4 = −3, so we have found our particular solution: pn = −2n − 3.

Putting all this together, we have found the general solution of the non-
homogeneous recurrence relation an = 2an−1 + 3an−2 + 8n − 4:
an = C1 3n + C2 (−1)n − 2n − 3, for some constants C1 , C2 . (2.18)
Only now, at the end, do we remember the initial conditions we want, a0 = 3
and a1 = 5. These are equivalent to the following equations for C1 and C2 :
C1 + C2 − 0 − 3 = 3, i.e. C1 + C2 = 6, and
3C1 − C2 − 2 − 3 = 5, i.e. 3C1 − C2 = 10.
Again this is a system of linear equations with a unique solution, namely
C1 = 4 and C2 = 2. (In general, the left-hand sides are the same as if we
were solving the homogeneous recurrence relation, so there will always be a
unique solution, for the same reason as in Theorem 2.36.) We have finally
found our closed formula for the original sequence:
an = 4 × 3n + 2 × (−1)n − 2n − 3. (2.19)
Example 2.48. We can now find the general solution of the tower of Hanoi
recurrence relation an = 2an−1 + 1. The corresponding homogeneous recur-

rence relation bn = 2bn−1 has general solution bn = , where

72 Topics in Discrete Mathematics

C is a constant. Since the extra term in the non-homogenous recurrence is

also a constant (if you like, a polynomial in n of degree 0), we look for a
solution pn = C 0 where C 0 is a constant. The equation we need C 0 to sat-

isfy is which means that C 0 = . So by

Theorem 2.45, the general solution of an = 2an−1 + 1 is

an = for some constant C.

The constant C is determined by the initial condition, i.e. the value of a0 . In

the tower of Hanoi example we had h0 = 0, so hn = 2n − 1 as seen before.
Example 2.49*. Recall from Example 1.82 that we have a closed formula for
the elements of the k = 2 column of the Stirling triangle: S(n, 2) = 2n−1 − 1
for all n ≥ 1. So Theorem 1.81 gives us a recurrence relation for the elements
of the k = 3 column:

S(1, 3) = 0, S(n, 3) = 3S(n − 1, 3) + 2n−2 − 1 for n ≥ 2.

This is a first-order non-homogeneous recurrence. According to Theorem

2.45, the general solution of an = 3an−1 + 2n−2 − 1 has the form an = bn + pn ,
where bn satisfies the first-order homogeneous relation bn = 3bn−1 , and pn is
a particular solution of the recurrence relation pn = 3pn−1 + 2n−2 − 1. We
have bn = C3n for some constant C, so we just need to find pn . This time
the extra term is not a polynomial in n, but applying the same principle of
trying things of the same form as the extra term, we look for a solution of
the form pn = C1 2n + C2 where C1 and C2 are further constants. We need
to have the following equation for all n ≥ 2:

C1 2n + C2 = 3(C1 2n−1 + C2 ) + 2n−2 − 1

 
3C1 1 n
= + 2 + (3C2 − 1).
2 4

For this to hold for all n ≥ 2, it has to be true that C1 = 3C2 1 + 14 and
C2 = 3C2 − 1. Hence C1 = − 12 and C2 = 12 , and our particular solution is
−2n +1
2
, which gives us the general solution:

−2n + 1
an = C3n + , for some constant C. (2.20)
2
CHAPTER 2. RECURSION AND INDUCTION 73

To achieve the initial condition S(1, 3) = 0 we need to have 3C − 21 = 0, so

C = 61 . This gives our desired formula for the k = 3 Stirling numbers:
3n−1 − 2n + 1
S(n, 3) = , for all n ≥ 1. (2.21)
2
Unfortunately, this is not really anything new: it is the same as what you
get by combining Theorem 1.84 with the Inclusion/Exclusion formula for the
number of surjective functions, Theorem 1.73.
Example 2.50**. To show that finding a particular solution can be a tricky
business, consider the sequence defined by
a0 = a1 = 1, an = 4an−1 − 4an−2 + 2n for n ≥ 2.
We saw in Example 2.34 that the general solution of the homogeneous recur-
rence has the form bn = C1 2n + C2 n2n , but how to find a particular solution
of pn = 4pn−1 − 4pn−2 + 2n ? On the pattern of the previous examples, it
would be natural to look for one of the form pn = C3 2n , but then we would
have to have the following equation for all n ≥ 2:
C3 2n = 4C3 2n−1 − 4C3 2n−2 + 2n = (C3 + 1)2n ,
which is clearly impossible. The problem is that a constant multiple of 2n
is already part of bn , i.e. it satisfies the homogeneous relation, so it can’t
satisfy the non-homogeneous relation as well. The same objection applies to
a constant multiple of n2n . How about if we try pn = C3 n2 2n ? Then we need
to have, for all n ≥ 2:
C3 n2 2n = 4C3 (n − 1)2 2n−1 − 4C3 (n − 2)2 2n−2 + 2n ,
which (dividing both sides by 2n ) is equivalent to
C3 n2 = 2C3 (n − 1)2 − C3 (n − 2)2 + 1 = C3 n2 − 2C3 + 1.
This is indeed satisfied when C3 = 21 , so we get
1
an = C1 2n + C2 n2n + n2 2n , for some constants C1 , C2 . (2.22)
2
The initial condition a0 = 1 forces C1 = 1, and the initial condition a1 = 1
forces C2 = −1. So the solution of the original recurrence is
1
an = (1 − n + n2 )2n . (2.23)
2
We will re-derive this solution with less guesswork in Example 3.42.
74 Topics in Discrete Mathematics

We have the following general rule to find particular solutions in the case
where the sequence f (n) is quasi-polynomial; that is, f (n) = q(n)µn for some
polynomial q(n) and some constant µ. This constant is called the exponent of
the quasi-polynomial. Note that if µ = 1 then f (n) is a genuine polynomial.
Consider the kth-order non-homogeneous linear recurrence relation

an = r1 an−1 + r2 an−2 + · · · + rk an−k + q(n)µn , for all n ≥ k.

Theorem 2.51. Suppose that q(n) is a polynomial of degree d and suppose

that µ is a root of the characteristic polynomial of multiplicity l. Then a
particular solution of the non-homogeneous recurrence relation can be found
in the form pn = Q(n)nl µn , where Q(n) is a polynomial of degree d.

Note that “µ is a root of multiplicity 0” just means that µ is not a root of

the characteristic polynomial.

By the same argument as for the proof of Theorem 2.45 one verifies that
if f (n) is a sum of quasi-polynomials then a particular solution of the non-
homogeneous linear recurrence relation can be found as the corresponding
sum of particular solutions provided by Theorem 2.51.

Returning to Example 2.50, note that f (n) = 2n can be regarded as a quasi-

polynomial with q(n) = 1 and the exponent µ = 2. The exponent is a
root of the characteristic polynomial x2 − 4x + 4 of multiplicity 2 and so,
by Theorem 2.51, a particular solution can indeed be found in the form
pn = Cn2 2n as we saw in the example.
Chapter 3

Generating Functions

Generating functions are a powerful device for capturing a whole sequence

in a single entity. Manipulations of generating functions can be used to solve
many types of recurrence relations, including the linear recurrence relations
we examined in the previous chapter.
Definition 3.1. The generating function of the sequence a0 , a1 , a2 , a3 , · · · is
the formal power series
a0 + a1 z + a2 z 2 + a3 z 3 + · · ·
in the indeterminate z. This is often denoted A(z), where in general A should
be replaced by (the capital form of) whatever letter denotes the terms of the
sequence.
Example 3.2. The sequence 1, 2, 4, 8, · · · has generating function
1 + 2z + 4z 2 + 8z 3 + · · · .
The sequence 0, 1, 2, 3, · · · has generating function
z + 2z 2 + 3z 3 + · · ·
(notice that we leave off the 0+ at the beginning, and write simply z instead of
1z). The Fibonacci sequence and Catalan sequence have generating functions
F (z) = z + z 2 + 2z 3 + 3z 4 + 5z 5 + 8z 6 + · · · ,
C(z) = 1 + z + 2z 2 + 5z 3 + 14z 4 + · · · .

75
76 Topics in Discrete Mathematics

3.1 Formal power series

The first question you should ask about the definition of generating func-
tion is: what is a “formal power series”? This is a very useful idea which
generalizes the concept of a polynomial.

You are familiar with the terminology that something like 2+z +4z 2 is called
a polynomial in the indeterminate (or variable) z. This means that it is a
sum of terms, each of which consists of a different power of z multiplied by a
coefficient which is just a number. (The constant term is the term in which
the power of z is z 0 , and we usually don’t write the z 0 , just the coefficient.)
It is an important part of the concept that there are only finitely many terms
– although you could imagine, if you like, that all the other powers of z are
present but not visible because their coefficient is 0. As you know, there
are natural ways to define addition, subtraction, multiplication, and even
long division of polynomials. The other important thing you can do with a
polynomial in the indeterminate z is to substitute a number for z: e.g. if you
substitute z = 3 in 2 + z + 4z 2 , you get the result 2 + 3 + 4 × 32 = 41. This
means that you can think of polynomials as just a special kind of function.

A formal power series is a ‘polynomial with infinitely many terms’. That is,
it is something like 1 + 2z + 4z 2 + 8z 3 + · · · , in which there is a term for every
power of z. It is allowed for some of the coefficients to be 0, however; indeed,
a polynomial like 2 + z + 4z 2 can be viewed as a formal power series in which
the coefficients of z 3 , z 4 , · · · happen to be 0, and even a number like 2 can be
viewed as a formal power series in which the coefficients of all the positive
powers of z happen to be 0. But having, in general, infinitely many terms
makes the crucial difference that you can’t just substitute any number for z
any more. For example, if you tried to substitute z = 3 in 1+2z +4z 2 +8z 3 +
· · · , you would get the series (infinite sum) 1 + 2 × 3 + 4 × 32 + 8 × 33 + · · · ,
which makes no sense.
Remark 3.3. Of course, this statement is too strong. Power series such as
this arise in many parts of mathematics – you would have seen Taylor series in
calculus, for example – and one of the main aims of analysis is to make sense
of them. But even in the context of analysis, 1 + 2 × 3 + 4 × 32 + 8 × 33 + · · ·
is a divergent series, which cannot be given a numerical value. (You could
CHAPTER 3. GENERATING FUNCTIONS 77

say it equalled ∞, but that is really just a shorthand way of expressing

the divergence, not a true evaluation: whatever ∞ is, it’s not a number.)
MATH2962 includes some fascinating results about when such a power series
is convergent and when it is divergent. These results are most interesting
and powerful in the context of the complex numbers, and that is actually the
reason we are using the letter z for the indeterminate, rather than, say, x.
Somewhere at the back of our minds is the idea that it would be interesting
to substitute a complex number for z, although in this course we will not go
down that path.

The fact that we can’t (or at least won’t) substitute a number for z means
that we can’t think of a formal power series as a function, as we could a
polynomial (despite the term “generating function”). This is what is meant
by the adjective “formal”: a0 + a1 z + a2 z 2 + a3 z 3 + · · · is not a function of
z, it’s just an object in its own right, in which the letter z and its ‘powers’
play the role of keeping the terms of the sequence in their correct order. In
fact, you could think of the generating function a0 + a1 z + a2 z 2 + a3 z 3 + · · ·
as just a different notation for the sequence a0 , a1 , a2 , a3 , · · · , using + signs
and powers of z instead of commas. To get back to the sequence, you just
have to extract the coefficients of the various powers of z.

Of course, this would be a bizarre choice of notation if there were nothing

more to it. The point is that we can add and multiply formal power series, by
rules which look completely natural in this notation, if you imagine collecting
terms and multiplying powers of z in the obvious way.
Definition 3.4. If we have two formal power series,
A(z) = a0 + a1 z + a2 z 2 + a3 z 3 + · · · and B(z) = b0 + b1 z + b2 z 2 + b3 z 3 + · · · ,
then
A(z) + B(z) := (a0 + b0 ) + (a1 + b1 )z + (a2 + b2 )z 2 + (a3 + b3 )z 3 + · · · ,
A(z)B(z) := a0 b0 + (a0 b1 + a1 b0 )z + (a0 b2 + a1 b1 + a2 b0 )z 2
+ (a0 b3 + a1 b2 + a2 b1 + a3 b0 )z 3 + · · · .

The justification for this multiplication rule is that to get a z n term in the
product, you need to multiply the am z m term of A(z) P with the bn−m z n−m
n
term of B(z), for some m. So the coefficient of z n is m=0 am bn−m .
78 Topics in Discrete Mathematics

Example 3.5. According to these rules, the sum of

F (z) = z + z 2 + 2z 3 + 3z 4 + · · · and C(z) = 1 + z + 2z 2 + 5z 3 + 14z 4 + · · ·
is just obtained by adding the coefficients term-by-term:
F (z) + C(z) = 1 + 2z + 3z 2 + 7z 3 + 17z 4 + · · · .
We obtain each coefficient of the product by adding appropriate products of
one coefficient from F (z) and one coefficient from C(z):

coefficient of z 0 in F (z)C(z) =

coefficient of z 1 in F (z)C(z) =

coefficient of z 2 in F (z)C(z) =

coefficient of z 3 in F (z)C(z) =

coefficient of z 4 in F (z)C(z) =

so F (z)C(z) =

It is often useful to write formal power series in sigma notation as follows.

Definition 3.6. The symbol
X∞
an z n means a0 + a1 z + a2 z 2 + a3 z 3 + · · · .
n=0
The
P∞ lowern limit of the3 sum 4can be5 varied in the obvious way: for instance,
n=3 bn z means b3 z + b4 z + b5 z + · · · . The summation variable n can of
course be replaced with any other letter.

In this notation the rules for addition and multiplication become:

∞
! ∞
! ∞
X X X
n n
an z + bn z = (an + bn ) z n ,
n=0 n=0 n=0
∞
! ∞
! ∞ n
! (3.1)
X X X X
n n n
an z bn z = am bn−m z .
n=0 n=0 n=0 m=0
CHAPTER 3. GENERATING FUNCTIONS 79

A special case of the multiplication rule says that multiplying a0 +a1 z+a2 z 2 +
· · · by a number b0 just multiplies every coefficient by that number. So if
A(z) is the generating function of the sequence a0 , a1 , a2 , · · · , then 3A(z), for
instance, is the generating function of the sequence 3a0 , 3a1 , 3a2 , · · · .

Another important special case of the multiplication rule is when B(z) = z m

for some m. When you multiply a0 + a1 z + a2 z 2 + · · · by z m , you get
a0 z m + a1 z m−1 + a2 z m+2 + · · · , as you would expect. Notice that the co-
efficients of z 0 , z 1 , · · · , z m−1 in the result are all zero, and the previous coef-
ficients a0 , a1 , a2 , · · · have all been ‘shifted m places to the right’, so that the
coefficient of z n is an−m for all n ≥ m. Thus in sigma notation, the rule for
multiplying by a power of z is
∞
X ∞
X
m n
z ( an z ) = an−m z n . (3.2)
n=0 n=m

In terms of sequences, this means that if A(z) is the generating function of

a0 , a1 , a2 , · · · , then

zA(z) is the generating function of 0, a0 , a1 , a2 , a3 , · · · ,

z 2 A(z) is the generating function of 0, 0, a0 , a1 , a2 , · · · ,
z 3 A(z) is the generating function of 0, 0, 0, a0 , a1 , · · · ,
and so on.
P
Example 3.7. Let G(z) = ∞ n 2 3
n=0 z = 1 + z + z + z + · · · be the geometric
series, the generating function of the sequence 1, 1, 1, 1, · · · . We have
∞
X
zG(z) = z n = z + z 2 + z 3 + z 4 + · · · = G(z) − 1,
n=1

1
so (1−z)G(z) = 1, which is commonly rewritten as G(z) = 1−z
. In summary:

1
1 + z + z2 + z3 + · · · = . (3.3)
1−z
As you may know, if you substitute for z a complex number α inside the unit
circle, the left-hand side of (3.3) becomes a series which converges to the
1
number 1−α . On the other hand, if you substitute for z a complex number α
outside the unit circle, the left-hand side becomes a divergent series which has
80 Topics in Discrete Mathematics

1
nothing to do with 1−α . The equation (3.3) is not meant to assert anything
about what happens when you substitute an actual number for z. It means
solely that 1+z +z 2 +z 3 +· · · is a formal power series which has the property
that when you multiply it by 1 − z, you get 1.

To make sense of fractions of formal power series, we need the following

result. (Here 0 means the power series with all coefficients 0.)
Theorem 3.8. (1) If A(z)B(z) = 0, then either A(z) or B(z) must be 0.

(2) If A(z)F (z) = A(z)G(z), and A(z) 6= 0, then F (z) = G(z).

Proof*. We will prove (1) in the ‘contrapositive’ form, which says that if
both A(z) and B(z) are nonzero, then so is A(z)B(z). Since they are nonzero
formal power series, they each have a ‘starting term’; in other words,

A(z) = ap z p + ap+1 z p+1 + ap+2 z p+2 + · · · ,

B(z) = bq z q + aq+1 z q+1 + aq+2 z q+2 + · · · ,

where ap is nonzero and all the earlier coefficients ap0 for p0 < p are zero, and
similarly bq is nonzero and all the earlier coefficients bq0 for q 0 < q are zero.
From the multiplication rule we see that the coefficient of z p+q in A(z)B(z)
is ap bq , because every other term you would expect to be in the coefficient
involves either some ap0 for p0 < p or some bq0 for q 0 < q. Since ap bq 6= 0,
this shows that A(z)B(z) has at least one nonzero coefficient, so it is not
the zero power series. Part (2) follows by applying part (1) to the equation
A(z)(F (z) − G(z)) = 0.
Remark 3.9. Part (2) of Theorem 3.8 implies that there cannot be two
different power series F (z) and G(z) which satisfy (1−z)F (z) = (1−z)G(z) =
1
1, so our use of the notation 1−z for the geometric series is unambiguous.
Similarly, from now on, whenever we state that F (z) = B(z) A(z)
(a quotient of
formal power series), it just means that A(z)F (z) = B(z); this notation is
possible because there can be at most one formal power series F (z) which
satisfies this equation. These fractions can be manipulated in the usual way.
Remark 3.10*. It is not always the case that such a quotient formal power
series exists: there is no formal power series F (z) which satisfies zF (z) =
1 + z, for instance. But if A(z) has nonzero constant term, then one can find
CHAPTER 3. GENERATING FUNCTIONS 81

1
an inverse A(z) , i.e. a formal power series F (z) such that A(z)F (z) = 1; the
coefficients of the powers of z n in F (z) can be determined recursively. So in
this case, B(z)
A(z)
= B(z) A(z) 1
always makes sense.

3.2 Manipulating formal power series

In order to use generating functions effectively, we will need more operations

on formal power series than just addition and multiplication. Some of the
basic results involve an operation of apparently dubious validity, namely
differentiating with respect to z.
Example 3.11. Let us ‘differentiate both sides of (3.3) with respect to z’:
1
1 + 2z + 3z 2 + 4z 3 + · · · = .
(1 − z)2
This purports to be a formula for the generating function of the sequence
1, 2, 3, 4, · · · . As explained in Remark 3.9, the interpretation of this is that
1 + 2z + 3z + 4z 3 + · · · is a formal power series with the property that when
you multiply it by (1 − z)2 , you get 1. We can easily prove this rigorously,
because from the multiplication rule we get

(1 + z + z 2 + z 3 + · · · )2 = 1 + 2z + 3z 2 + 4z 3 + · · · , (3.4)

from which it follows that

(1 − z)2 (1 + 2z + 3x2 + 4z 3 + · · · ) = ((1 − z)(1 + z + z 2 + z 3 + · · · ))2 = 12 = 1.

So we do indeed have
∞
X 1
(n + 1) z n = . (3.5)
n=0
(1 − z)2

Incidentally, applying the shifting principle (3.2) to (3.5), we find

∞
X z
n z n = z + 2z 2 + 3z 3 + · · · = , (3.6)
n=1
(1 − z)2

which is a formula for the generating function of the sequence 0, 1, 2, 3, · · · .

82 Topics in Discrete Mathematics

Clearly this dubious differentiation is worth defining properly.

P
Definition 3.12. If A(z) = ∞ n 2 3
n=0 an z = a0 + a1 z + a2 z + a3 z + · · · is a
formal power series, we define its derivative to be the formal power series
∞
X
A0 (z) := (n + 1)an+1 z n = a1 + 2a2 z + 3a3 z 2 + 4a4 z 3 + · · · .
n=0
P
Remark 3.13. It would also be correct to write A0 (z) =P ∞ n=1 nan z
n−1
,
∞ n
which is what you would expect to get from differentiating n=0 an z term-
by-term (killing the constant term a0 ). In Definition 3.12 we have shifted
the summation variable so that it equals the power of z, which gives a more
convenient form.

We can also find ‘integrals’ of formal power series:

Theorem 3.14. For any formal power series A(z), there exists a formal
power series B(z) such that B 0 (z) = A(z). Any other solution of this equation
differs from B(z) only in the constant term; there is a unique solution which
has zero constant term.

P∞ n
P∞ n
Proof. If A(z) = n=0 an z and B(z) = n=0 bn z , then the equation
B 0 (z) = A(z) is equivalent to the statement that an = (n + 1)bn+1 for all
n ≥ 0. So bn must be an−1 n
for all n ≥ 1, and the value of b0 is irrelevant.
The result follows.

We can easily prove some familiar-looking differentiation rules.

P P∞
Theorem 3.15. If A(z) = ∞ n
n=0 an z and B(z) =
n
n=0 bn z , then:

(1) (Sum rule) (A(z) + B(z))0 = A0 (z) + B 0 (z).

(2) (Product rule) (A(z)B(z))0 = A0 (z)B(z) + A(z)B 0 (z).
(3) (Power rule) (A(z)k )0 = kA(z)k−1 A0 (z), for all k ≥ 1.
(4) (Quotient rule) If B(z)F (z) = A(z) for some formal power series F (z),
then B(z)2 F 0 (z) = A0 (z)B(z) − A(z)B 0 (z). In short,
 0
A(z) A0 (z)B(z) − A(z)B 0 (z)
= .
B(z) B(z)2
CHAPTER 3. GENERATING FUNCTIONS 83

Proof*. To prove an equality of formal power series as in (1), it is enough

to prove that the coefficient of z n is the same on both sides, for all n ≥ 0.
By Definition 3.12, the coefficient of z n in (A(z) + B(z))0 is (n + 1) times the
coefficient of z n+1 in A(z) + B(z), i.e. (n + 1)(an+1 + bn+1 ). The coefficient
of z n in A0 (z) + B 0 (z) is (n + 1)an+1 + (n + 1)bn+1 , so the equality is obvious.
The proof of the product P rule is slightly harder. The coefficient of z n in
n+1
(A(z)B(z))0 is (n + 1) m=0 am bn+1−m , while in A0 (z)B(z) + A(z)B 0 (z) it is
n
X n
X
(m + 1)am+1 bn−m + am (n − m + 1)bn−m+1
m=0 m=0
n+1
X n
X
= mam bn+1−m + (n − m + 1)am bn+1−m
m=1 m=0
n+1
X n+1
X
= mam bn+1−m + (n − m + 1)am bn+1−m
m=0 m=0
n+1
X
= (n + 1)am bn+1−m ,
m=0

as required. Here the second-last step used the common trick of adding an
extra term, which works out to be zero, at the top or bottom of a sum. We
prove the power rule by induction on k. The k = 1 case is obvious, so assume
that k ≥ 2 and that the result is known for k − 1. Then using the product
rule, we have
(A(z)k )0 = (A(z)k−1 A(z))0
= (A(z)k−1 )0 A(z) + A(z)k−1 A0 (z)
= (k − 1)A(z)k−2 A0 (z)A(z) + A(z)k−1 A0 (z)
= kA(z)k−1 A0 (z),
completing the induction step. To prove the quotient rule, we differentiate
both sides of B(z)F (z) = A(z) and use the product rule on the left-hand
side (the same way the quotient rule is proved in calculus, incidentally).
This gives
B 0 (z)F (z) + B(z)F 0 (z) = A0 (z),
and multiplying both sides by B(z), using B(z)F (z) = A(z), and rearranging
gives the desired equation B(z)2 F 0 (z) = A0 (z)B(z) − A(z)B 0 (z).
84 Topics in Discrete Mathematics

Example 3.16. Our initial proof of (3.5) can now be justified: we obtain it
from (3.3) by differentiating both sides, and using the quotient rule on the
right-hand side. If we differentiate both sides of (3.5) again, we obtain:
X∞
−2(1 − z)(−1) 2
(n + 1)(n + 2) z n = 4
= 3
.
n=0
(1 − z) (1 − z)

Dividing both sides by 2, this says that

X∞  
n+2 n 1
z = 3
, (3.7)
n=0
2 (1 − z)

which is a formula for the generating function of the k =
2 column of Pascal’s

2
triangle (i.e. the triangular numbers), starting from 2 = 1.

In general, the generating function of each column of Pascal’s triangle has a

very simple formula:
Theorem 3.17. For any k ≥ 0,
X∞  
n+k n 1
z = .
n=0
k (1 − z)k+1

Proof*. We proceed by induction on k. The k = 0 base case is (3.3). Assume

that k ≥ 1, and that we know the result for k − 1, i.e.
X∞  
n+k−1 n 1
z = k
.
n=0
k − 1 (1 − z)

Taking derivative of both sides (and using the quotient and power rules on
the right-hand side), we get
X∞  
n+k n 0 − k(1 − z)k−1 (−1) k
(n + 1) z = 2k
= .
n=0
k−1 (1 − z) (1 − z)k+1

Since
   
n+1 n+k (n + 1)(n + k)(n + k − 1) · · · (n + 2) n+k
= = ,
k k−1 k (k − 1)! k
we obtain the result.
CHAPTER 3. GENERATING FUNCTIONS 85

Remark 3.18*. Theorem 3.17 is equivalent to saying that for all s ≥ 1,

X∞  
2 3 s n+s−1 n
(1 + z + z + z + · · · ) = z . (3.8)
n=0
n
To get a z n term on the left-hand side of (3.8), one must select from the
various factors the z n1 , z n2 , · · · , z ns terms where n1 , n2 , · · · , ns are nonneg-
ative integers such that n1 + n2 + · · · + ns = n. So (3.8) is equivalent to
the statement that the number
of s-tuples (n1 , n2 , · · · , ns ) ∈ Ns such that
n1 + n2 + · · · + ns = n is n+s−1n
. Thus Theorem 3.17 is just a disguised form
of Theorem 1.60.

Multiplying Theorem 3.17 on both sides by z k gives a generalization of (3.6):

X∞  
n n zk
z = . (3.9)
n=0
k (1 − z)k+1
(You would normally change the summation so
that it started from n = k,
but in this case it makes no difference because nk = 0 for n = 0, 1, · · · , k−1.)

Since we can express nm as a linear combination of binomial coefficients nk
for various k by Theorem 1.86, we can use (3.9) to get a closed formula for
the generating function of the sequence 0m , 1m , 2m , · · · for any m, and hence
of any sequence where the nth term is a polynomial function of n.
Example 3.19. To find a closed formula for 02 + 12 z + 22 z 2 + 32 z 3 + · · · , we
first express n2 as an integer linear combination of binomial coefficients:

n

n


n2 = 1
+ 2
.
Hence
∞
X ∞
X ∞
X
2 n n

n n


n z = 1
z + 2
zn
n=0 n=0 n=0

Another seemingly dubious operation which is often useful is to take a formal

power series A(z) and substitute for z another formal power series F (z),
expanding out each F (z)n and collecting terms; the result is denoted A(F (z)).
86 Topics in Discrete Mathematics

Example 3.20. The simplest and most commonly encountered example is

when F (z) = cz for some complex number c; then we are just changing
A(z) = a0 + a1 z + a2 z 2 + · · · to A(cz) = a0 + a1 cz + a2 c2 z 2 + · · · , i.e.
multiplying the coefficient of z n by the number cn , for all n.
Example 3.21. A more complicated example is when F (z) = z + 21 z 2 ; then
the result of the substitution is
1 1 1
a0 + a1 (z + z 2 ) + a2 (z + z 2 )2 + a3 (z + z 2 )3 + · · ·
2 2 2
1 2 1 3 3 1
= a0 + a1 (z + z ) + a2 (z 2 + z 3 + z 4 ) + a3 (z 3 + z 4 + z 5 + z 6 ) + · · ·
2 4 2 4 8
1
= a0 + a1 z + ( a1 + a2 )z 2 + (a2 + a3 )z 3 + · · · .
2
It is important to notice in this last step that there cannot be any other
contributions to the coefficients of 1, z, z 2 , z 3 from the “ · · · ” on the previous
line, because that “ · · · ” involves fourth and higher powers of (z + 21 z 2 ),
which will expand to fourth and higher powers of z, which we are sweeping
into the “ · · · ” on the last line. In general, to work out the coefficient of z n in
A(z + 21 z 2 ), we only need to know the coefficients of 1, z, z 2 , · · · , z n in A(z).

This finiteness principle continues to work, and therefore the substitution

continues to make sense, if F (z) is any formal power series with zero constant
term; but not if F (z) has nonzero constant term.
Example 3.22*. If you try to substitute 1 + z for z in A(z), you get the
expression a0 + a1 (1 + z) + a2 (1 + 2z + z 2 ) + a3 (1 + 3z + 3z 2 + z 3 ) + · · · . This
is not a valid formal power series, because each power of z gets contributions
from infinitely many terms in the sum.
Theorem 3.23. If A(z), B(z), and F (z) are formal power series, and F (z)
has zero constant term, then:

(1) If S(z) = A(z) + B(z), then S(F (z)) = A(F (z)) + B(F (z)).

(2) If P (z) = A(z)B(z), then P (F (z)) = A(F (z))B(F (z)).

A(z) A(F (z))
(3) If G(z) = B(z)
, then G(F (z)) = B(F (z))
.

(4) (Chain rule) A(F (z))0 = A0 (F (z))F 0 (z).

CHAPTER 3. GENERATING FUNCTIONS 87

Proof*. We will only give the proof in the simple case that F (z) = cz, since
this is enough for many of our later purposes. As seen above, substituting
cz for z in a formal power series just multiplies the coefficient of z n by cn ,
for all n ≥ 0. Part (1) is therefore easy: the coefficient of z n in S(F (z)) is
cn (an + bn ), whereas that in the right-hand sideP is cn an + cn bn . Similarly, for
part (2), the coefficient Pof z n in P (F (z)) is cn nm=0 am bn−m , whereas that
in the right-hand side is nm=0 (cm am )(cn−m bn−m ), which is clearly the same.
Part (3) follows immediately from part (2), if you rearrange the equations so
that they don’t involve fractions. Part (4) in the special case F (z) = cz says
that A(cz)0 = cA0 (cz); the coefficient of z n in A(cz)0 is (n + 1)cn+1 an+1 , and
that in A0 (cz) is cn (n + 1)an+1 , so this is true.
Example 3.24. Applying part (3) in the case F (z) = cz to (3.3), we find
∞
X 1
cn z n = 1 + cz + c2 z 2 + c3 z 3 + · · · = . (3.10)
n=0
1 − cz

1
For instance, the generating function of the sequence 1, 2, 4, 8, · · · is 1−2z
.
Similarly, (3.5) implies that
∞
X 1
(n + 1) cn z n = . (3.11)
n=0
(1 − cz)2

More generally, Theorem 3.17 implies that for all k ≥ 0,

∞ 
X 
n+k 1
cn z n = . (3.12)
n=0
k (1 − cz)k+1

At this point we know enough to extract the coefficient of z n from any for-
F (z)
mal power series of the form G(z) where F (z) and G(z) are polynomials.
We may as well assume that the degree of F (z) is less than that of G(z),
since long division of polynomials reduces the general case to this. Then
F (z)
the method is to factorize G(z) into powers (1 − cz)m , and rewrite G(z) in
b
partial-fractions form, i.e. as a sum of terms of the form (1−cz)m where b and
c are constants (i.e. just numbers, not involving z). The coefficient of z n in
each term can then be read off from (3.10), (3.11), or in general (3.12).
88 Topics in Discrete Mathematics

2−9z
Example 3.25. To find the coefficient of z n in 1−7z+12z 2 , we need to write

it in partial-fractions form. The denominator factorizes as (1 − 3z)(1 − 4z),

so this amounts to finding constants C1 and C2 such that
2 − 9z C1 C2
= + .
(1 − 3z)(1 − 4z) 1 − 3z 1 − 4z
Multiplying both sides by (1 − 3z)(1 − 4z), this equation becomes

Equating coefficients, we get two linear equations for the two unknowns of

which the unique solution is C1 = and C2 = . So the

desired partial-fractions expression is:

2 − 9z
=
1 − 7z + 12z 2
We can use (3.10) to read off that

2 − 9z
the coefficient of z n in is
1 − 7z + 12z 2
5−9z
Example 3.26. Let us find the coefficient of z n in 1−6z+9z 2 . The denomi-
2
nator factorizes as (1 − 3z) ; when we have a repeated factor like this, the
partial-fractions form should include a term for every power of the factor up
to the multiplicity, which in this case means
5 − 9z C1 C2
2
= + .
(1 − 3z) 1 − 3z (1 − 3z)2
Clearing the denominators, this becomes
5 − 9z = C1 (1 − 3z) + C2 = (C1 + C2 ) − 3C1 z,
which has the unique solution C1 = 3, C2 = 2. So
5 − 9z 3 2
2
= + .
1 − 6z + 9z 1 − 3z (1 − 3z)2
By (3.10) and (3.11), the coefficient of z n is 3 × 3n + 2(n + 1)3n = (2n + 5)3n .
CHAPTER 3. GENERATING FUNCTIONS 89

2z
Example 3.27*. Let us find the coefficient of z n in (1−z)(1−2z)2 . The partial-

fractions form is
2z C1 C2 C3
2
= + + .
(1 − z)(1 − 2z) 1 − z 1 − 2z (1 − 2z)2
Clearing the denominators, this becomes
2z = C1 (1 − 2z)2 + C2 (1 − z)(1 − 2z) + C3 (1 − z)
= (C1 + C2 + C3 ) − (4C1 + 3C2 + C3 )z + (4C1 + 2C2 )z 2 .
So we must solve the following system of three linear equations:
C1 + C2 + C3 = 0,
4C1 + 3C2 + C3 = −2,
4C1 + 2C2 = 0.
The first equation minus the second plus the third gives C1 = 2, and then
one deduces C2 = −4 and C3 = 2. So
2z 2 4 2
2
= − + ,
(1 − z)(1 − 2z) 1 − z 1 − 2z (1 − 2z)2
and the coefficient of z n is 2 − 4 × 2n + 2(n + 1)2n = (n − 1)2n+1 + 2.

It is sometimes possible to find square roots, cube roots, etc. of formal power
series. By the Binomial Theorem, for any nonnegative integer a we have the
following equation of polynomials in z:
Xa  
a a n
(1 + z) = z . (3.13)
n=0
n

Definition 3.28. For any complex number α, we define the formal power
series (1 + z)α by:
X∞  
α α n
(1 + z) := z .
n=0
n


(Recall that αn = α(α−1)···(α−n+1)
n!
is defined for any α ∈ C.)

In the special case when α is a nonnegative integer a, this definition does

indeed give the polynomial (1 + z)a that we
expect: the reason is that every
term where n > a vanishes, because then na = 0.
90 Topics in Discrete Mathematics

Theorem 3.29. (1 + z)α (1 + z)β = (1 + z)α+β , for any α, β ∈ C.

Proof**. We just need to show that the coefficient of z n is the same on both
sides, for all n ≥ 0. That is, we must prove that
Xn
α(α − 1) · · · (α − m + 1) β(β − 1) · · · (β − (n − m) + 1)
m! (n − m)!
m=0 (3.14)
(α + β)(α + β − 1) · · · (α + β − n + 1)
= .
n!
One thing we certainly know is that (1 + z)a (1 + z)b = (1 + z)a+b for every
nonnegative integers a and b, so (3.14) must hold whenever α, β ∈ N. There
is a clever way to deduce the general case from this, using the fact that a
nonzero polynomial can have only finitely many roots (this is because every
root λ of a polynomial corresponds to a factor x − λ in its factorization).
First fix α to be some nonnegative integer a. Then both sides of (3.14)
are polynomial functions of β, so their difference is a polynomial function
of β, which has infinitely many roots because it vanishes whenever β is a
nonnegative integer. The only way this is possible is if their difference is the
zero polynomial. So with α fixed to equal a, (3.14) holds for all values of β.
Now we can use the same argument turned around: for any fixed value of β,
the two sides of (3.14) are polynomial functions of α which agree whenever
α is a nonnegative integer, so their difference must be zero.
Example 3.30. Directly from Theorem 3.29 we see that (1+z)1/2 is a square
root of 1 + z, in the sense that (1 + z)1/2 (1 + z)1/2 = (1 + z)1 = 1 + z. Writing
out the definition in a slightly nicer way:
X∞ 1 X∞ 1 1
1/2 2 n ( − 1)( 12 − 2) · · · ( 21 − n + 1) n
2 2
(1 + z) = z = z
n=0
n n=0
n!
∞
X 1(−1)(−3) · · · (3 − 2n)
=1+ zn
n=1
2n n!
X∞
(−1)n−1 (2n − 3)!! n
=1+ z .
n=1
2n n!

Here (2n − 3)!! is defined as in Example 1.40 (including the convention that
(−1)!! = 1). Similarly, (1 + z)1/3 is a cube root of 1 + z, and so forth.
CHAPTER 3. GENERATING FUNCTIONS 91

Example 3.31*. More generally, if F (z) is any formal power series with zero
constant term, then we can find a square root of 1+F (z) by substituting F (z)
into (1 + z)1/2 ; however, it is not often that you can get any nice formula
for the coefficients of the result. We will only use the case of the simple
substitution of cz for z:
∞
X
1/2 (−1)n−1 (2n − 3)!! cn
(1 + cz) =1+ zn. (3.15)
n=1
2n n!

You may like to ponder which formal power series have a square root: it
is pretty clear that z does not, for instance. All we will need is the easy
observation that a square root, if it does exist, is unique up to sign:

Theorem 3.32. If F (z) and G(z) are both square roots of A(z), then either
F (z) = G(z) or F (z) = −G(z).

Proof. The equation F (z)2 = G(z)2 can be rewritten as

(F (z) − G(z))(F (z) + G(z)) = 0,

and part (1) of Theorem 3.8 forces one of the two factors to be zero.
Example 3.33*. Notice that if α = −a is a negative integer, then (1 + z)−a
1 a 0
must be (1+z)a , since when multiplied by (1 + z) it gives (1 + z) = 1. So

X∞  
1 −a n
= z
(1 + z)a n=0
n
∞
X (−a)(−a − 1)(−a − 2) · · · (−a − n + 1)
= zn
n=0
n!
X∞
(−1)n a(a + 1)(a + 2) · · · (a + n − 1) n
= z ,
n=0
n!

which is a rearranged form of the c = −1 special case of (3.12).

Finally, we can even solve some ‘differential equations’ involving formal power
series. In the usual theory of (linear ordinary) differential equations, the
92 Topics in Discrete Mathematics

exponential function plays a crucial role, being the unique function ex which
d x x
satisfies the differential equation dx e =
Pe∞ subject to e0 = 1. So let us try
n
to find a formal power series exp(z) = n=0 en z such that

exp0 (z) = exp(z), subject to the condition that e0 = 1. (3.16)

Taking the coefficient of z n on either side of exp0 (z) = exp(z) gives

1
(n + 1)en+1 = en , or equivalently en+1 = en .
n+1
It is easy to see that the solution of this recurrence relation, given the initial
condition e0 = 1, is en = n!1 , so

X∞
1 n 1 1 1
exp(z) = z = 1 + z + z2 + z3 + z4 + · · · (3.17)
n=0
n! 2 6 24

is the unique solution of our ‘differential equation’ (3.16). Of course, this is

not remotely surprising, because this is exactly the Taylor series of the ex-
ponential function (and, as a power series of a complex variable, it converges
to the complex exponential function everywhere).

We can now use this exponential series to solve ‘differential equations’ of a

simple type:
Theorem 3.34*. Let F (z) be a formal power series, and let G(z) be the
unique formal power series with zero constant term such that G0 (z) = F (z).
Then A0 (z) = F (z)A(z) if and only if A(z) = a0 exp(G(z)), where a0 is the
constant term of A(z).
Remark 3.35*. This is the analogue of the fact that the solution Rx of the
d
ordinary differential equation dx a(x) = f (x)a(x) is a(x) = a(0) e 0 f (t)dt .

Proof**. The “if” direction, which just says that A(z) = a0 exp(G(z)) is
indeed a solution of the equation A0 (z) = F (z)A(z), can be proved using the
chain rule:

(a0 exp(G(z)))0 = a0 exp(G(z))G0 (z) = a0 exp(G(z))F (z).

Since the constant term of exp(G(z)) is 1, the constant term of a0 exp(G(z))

is indeed a0 . The “only if” direction says that a0 exp(G(z)) is the unique
CHAPTER 3. GENERATING FUNCTIONS 93

solution of the equation with constant term a0 . So it is P enough to prove

that, given any number a0 and any power series F (z) = ∞ n
n=0 bn z , there
is
P∞ a unique way to choose a1 , a2 , a3 , · · · so that the power series A(z) =
n 0 n
n=0 an z satisfies the equation A (z) = F (z)A(z). Extracting the z term
from both sides of this equation, we see that it is equivalent to
n
X
(n + 1)an+1 = bm an−m , for all n ≥ 0,
m=0

which is a recurrence relation determining an+1 in terms of a0 , a1 , · · · , an and

the fixed numbers bm . So it is indeed true that a1 , a2 , a3 , · · · are uniquely
determined.
Remark 3.36*. Although Theorem 3.34 enables us to express the solution
of the ‘differential equation’ neatly as a0 exp(G(z)), it may not be easy to
extract the coefficients of the powers of z after substituting G(z) in exp(z).

3.3 Generating functions and recursion

Now that we are familiar with the basic operations involving formal power
series, we explore how generating functions can be used to say something
about various kinds of recursive sequences a0 , a1 , a2 , · · · . The main idea is
that a recurrence relation expressing an in terms of earlier terms in the se-
quence should give rise to an equation expressing A(z) in terms of itself (or its
derivatives, powers, etc.). If this equation can be solved, we have a formula
for the generating function A(z). Ideally, we would then be able to extract
the coefficient of each z n , giving a closed formula for an (and thus solving
the recurrence relation). Even when this is not possible, a formula for the
generating function is a concise way of summarizing the sequence, and there
are standard methods for deducing qualitative information from it.

In our first class of examples, everything works perfectly: generating func-

tions provide an alternative way of solving the linear recurrence relations
discussed in the previous chapter. In fact, this is the way that de Moivre
originally discovered the characteristic-polynomial rule, Theorem 2.36.
94 Topics in Discrete Mathematics

Example 3.37. Recall from Example 2.33 the sequence defined by

a0 = 2, a1 = 5, an = 7an−1 − 12an−2 for n ≥ 2.
Let A(z) denote the generating function of this sequence. We use the recur-
rence relation to express A(z) in terms of itself:
A(z) = a0 + a1 z + a2 z 2 + a3 z 3 + a4 z 4 + · · ·
= 2 + 5z + (7a1 − 12a0 )z 2 + (7a2 − 12a1 )z 3 + (7a3 − 12a2 )z 4 + · · ·
= 2 + 5z + 7(a1 z 2 + a2 z 3 + a3 z 4 + · · · ) − 12(a0 z 2 + a1 z 3 + a2 z 4 + · · · )
= 2 + 5z + 7z(A(z) − 2) − 12z 2 A(z).
2−9z
Rearranging this gives (1 − 7z + 12z 2 )A(z) = 2 − 9z, i.e. A(z) = 1−7z+12z 2.
n
We found the coefficient of z in the latter expression in Example 3.25, so
we can conclude that an = 3n+1 − 4n , recovering the result of Example 2.33.

Assuming sufficient facility with the shift operation (3.2), one can write such
calculations more concisely using sigma notation rather than spelling out the
series with a “ · · · ”.
Example 3.38. Recall from Example 2.34 the sequence defined by
a0 = 1, a1 = 4, an = 4an−1 − 4an−2 for n ≥ 2.
P∞
Let A(z) = n=0 an z n be the generating function. Then
∞
X
A(z) = 1 + 4z + (4an−1 − 4an−2 ) z n
n=2
X∞ ∞
X
= 1 + 4z + 4 an−1 z n − 4 an−2 z n
n=2 n=2
2
= 1 + 4z + 4z(A(z) − 1) − 4z A(z),
1 1
which after rearrangement becomes A(z) = 1−4z+4z 2 = (1−2z)2 . Using (3.11),

we can write down straight away that an = (n + 1) 2n , as in Example 2.34.

Remark 3.39**. Following the pattern of these examples, one can give a
generating-function proof of Theorem 2.36.

Slightly more useful than solving homogeneous linear relations, which we

know how to do anyway, is tackling the non-homogeneous case, where gen-
erating functions can avoid the problem of finding a particular solution.
CHAPTER 3. GENERATING FUNCTIONS 95

Example 3.40. Recall the tower of Hanoi sequence satisfying h0 = 0 and

the recurrence relation hn = 2hn−1 + 1 for all n ≥ 1. If H(z) denotes the
generating function, we have the following equation for H(z) in terms of itself
and the geometric series:
H(z) = h0 + h1 z + h2 z 2 + h3 z 3 + · · ·
= 0 + (2h0 + 1)z + (2h1 + 1)z 2 + (2h2 + 1)z 3 + · · ·

which shows that

H(z) =

From this we deduce the formula hn = 2n − 1 immediately.

Example 3.41*. Recall from Example 2.47 the sequence defined by
a0 = 3, a1 = 5, an = 2an−1 + 3an−2 + 8n − 4 for n ≥ 2.
P
Let A(z) = ∞ n
n=0 an z be the generating function. Then
∞
X
A(z) = a0 + a1 z + an z n
n=2
∞
X
= 3 + 5z + (2an−1 + 3an−2 + 8n − 4) z n
n=2
X∞ ∞
X ∞
X ∞
X
n n n
= 3 + 5z + 2 an−1 z + 3 an−2 z + 8 nz − 4 zn
n=2 n=2 n=2 n=2
X∞ ∞
X ∞
X ∞
X
= 3 + 5z + 2z an z n + 3z 2
an z n + 8 n zn − 4 zn
n=1 n=0 n=2 n=2
2
= 3 + 5z + 2z(A(z) − 3) + 3z A(z)
   
z 1
+8 −z −4 −1−z ,
(1 − z)2 1−z
where the last line uses (3.3) and (3.6). Tidying this up gives

2 3 − 7z + 17z 2 − 5z 3
(1 − 2z − 3z )A(z) = ,
(1 − z)2
96 Topics in Discrete Mathematics

so
3 − 7z + 17z 2 − 5z 3
A(z) = . (3.18)
(1 − z)2 (1 − 3z)(1 + z)
We now want to rewrite the right-hand side using partial fractions:
3 − 7z + 17z 2 − 5z 3 C1 C2 C3 C4
= + + + ,
(1 − z)2 (1 − 3z)(1 + z) 1 − 3z 1 + z 1 − z (1 − z)2
where C1 , C2 , C3 , C4 are some constants. Clearing the denominator, we get
3 − 7z + 17z 2 − 5z 3 = C1 (1 − z)2 (1 + z) + C2 (1 − z)2 (1 − 3z)
+ C3 (1 − z)(1 − 3z)(1 + z) + C4 (1 − 3z)(1 + z)
= C1 (1 − z − z 2 + z 3 ) + C2 (1 − 5z + 7z 2 − 3z 3 )
+ C3 (1 − 3z − z 2 + 3z 3 ) + C4 (1 − 2z − 3z 2 )
= (C1 + C2 + C3 + C4 ) − (C1 + 5C2 + 3C3 + 2C4 )z
+ (−C1 + 7C2 − C3 − 3C4 )z 2 − (−C1 + 3C2 − 3C3 )z 3 ,
and equating coefficients gives us a system of four linear equations for the
four unknowns:
C1 + C2 + C3 + C4 =3
C1 + 5C2 + 3C3 + 2C4 =7
−C1 + 7C2 − C3 − 3C4 = 17
−C1 + 3C2 − 3C3 =5
Using linear algebra methods, we can solve this to obtain
C1 = 4, C2 = 2, C3 = −1, C4 = −2,
and thus we get the partial-fractions formula for the generating function:
4 2 1 2
A(z) = + − − . (3.19)
1 − 3z 1 + z 1 − z (1 − z)2
From this we can read off the answer:
an = 4 × 3n + 2 × (−1)n − 1 − 2(n + 1),
which is of course the same as in Example 2.47. For this example it is
questionable whether the generating-function approach is any better: it is
more motivated and direct, but the system of linear equations involved is
larger than anything we had to deal with in the other method.
CHAPTER 3. GENERATING FUNCTIONS 97

Example 3.42*. Recall from Example 2.50 the sequence defined by

a0 = a1 = 1, an = 4an−1 − 4an−2 + 2n for n ≥ 2.
P
Let A(z) = ∞ n
n=0 an z be the generating function. Then
∞
X
A(z) = a0 + a1 z + an z n
n=2
∞
X
=1+z+ (4an−1 − 4an−2 + 2n ) z n
n=2
∞
X ∞
X  
n 2 n 1
= 1 + z + 4z an z − 4z an z + − 1 − 2z
n=1 n=0
1 − 2z
1
= −z + 4z(A(z) − 1) − 4z 2 A(z) + .
1 − 2z
Rearranging gives
1 − 5z + 10z 2
(1 − 4z + 4z 2 )A(z) = ,
1 − 2z
so
1 − 5z + 10z 2
A(z) = . (3.20)
(1 − 2z)3
We now need to rewrite the right-hand side using partial fractions:
1 − 5z + 10z 2 C1 C2 C3
3
= + 2
+ ,
(1 − 2z) 1 − 2z (1 − 2z) (1 − 2z)3
where C1 , C2 , C3 are some numbers. Clearing the denominator, we get
1 − 5z + 10z 2 = C1 (1 − 2z)2 + C2 (1 − 2z) + C3
= (C1 + C2 + C3 ) − (4C1 + 2C2 )z + 4C1 z 2 ,
and after equating coefficients we find C1 = 5/2, C2 = −5/2, C3 = 1. So
5/2 5/2 1
A(z) = − 2
+ . (3.21)
1 − 2z (1 − 2z) (1 − 2z)3
Recalling (3.12), we can extract the coefficient of z n from the right-hand side:
 
5 n 5 n n+2 n
an = 2 − (n + 1)2 + 2 ,
2 2 2
which simplifies to the answer found in Example 2.50.
98 Topics in Discrete Mathematics

Along the same lines, we can find a formula for the generating function of
the kth column of the Stirling triangle (compare this with Theorem 3.17):
Theorem 3.43*. For any k ≥ 1,
∞
X 1
S(n + k, k) z n = .
n=0
(1 − z)(1 − 2z) · · · (1 − kz)

Proof*. We prove this by induction. The k = 1 base case is (3.3). As-

P∞ that k ≥ 2, and
sume
n
that the result is known for k − 1. Define Gk (z) :=
n=0 S(n + k, k) z . We must convert the recurrence relation for Stirling
numbers, Theorem 1.81, into an equation relating Gk (z) and Gk−1 (z):
∞
X
Gk (z) = (S(n + k − 1, k − 1) + k S(n + k − 1, k)) z n
n=0
X∞ ∞
X
= S(n + k − 1, k − 1) z n + k S(n + k − 1, k)z n
n=0 n=1
= Gk−1 (z) + kzGk (z).
Gk−1 (z)
Rearranging, this becomes Gk (z) = 1−kz
, and the result follows from the
induction hypothesis.
Remark 3.44**. To deduce a formula for an individual Stirling number
1
S(n + k, k), we need to find the coefficient of z n in (1−z)(1−2z)···(1−kz) , which
is the product

(1 + z + z 2 + · · · )(1 + 2z + 22 z 2 + · · · ) · · · (1 + kz + k 2 z 2 + · · · ).

To get a z n term from the expansion of the product, one must select the z n1
term from the first factor, the z n2 term from the second factor, etc., up to the
z nk term from the kth factor, where n1 + n2 + · · · + nk = n. The coefficient
obtained from this selection is 1n1 2n2 · · · k nk , so we find
X
S(n + k, k) = 1n1 2n2 · · · k nk . (3.22)
n1 ,n2 ,··· ,nk ∈N
n1 +n2 +···+nk =n

This is
a positive formula, but with the major drawback that the sum has
n+k−1
k−1
terms. So finding a formula for the generating function has not
CHAPTER 3. GENERATING FUNCTIONS 99

actually allowed us to solve the recurrence relation. Incidentally, (3.22) can

be proved directly from the definition of S(n + k, k): the term indexed by
n1 , n2 , · · · , nk counts those partitions of {1, 2, · · · , n + k} in which one block
has largest element nk + 1, another has largest element nk−1 + nk + 2, and
so on until the last block has largest element n1 + · · · + nk + k = n + k.

Perhaps more importantly, generating functions can be used to solve other

types of recurrence relations which are not linear.
Example 3.45*. The classic example is the Catalan sequence c0 , c1 , c2 , · · · ,
whose generating function we are calling C(z). Recall that the Catalan
numbers satisfy:
c0 = 1, cn = c0 cn−1 + c1 cn−2 + · · · + cn−2 c1 + cn−1 c0 for n ≥ 1.
The form of the recurrence relation suggests generating functions straight
Pn−1
away, because the right-hand side m=0 cm cn−1−m is exactly the coefficient
of z n−1 in C(z)2 , by the multiplication rule. Thus we have
∞
X
C(z) = 1 + cn z n = 1 + zC(z)2 .
n=1

So C(z) is a solution of the quadratic equation zC(z)2 − C(z) + 1 = 0, to

which it is tempting to apply the quadratic formula. If that seems a leap
too far, we can certainly do some completion of the square (which is how the
quadratic formula is proved anyway):
0 = 4z − 4zC(z) + 4z 2 C(z)2 = (1 − 2zC(z))2 − 1 + 4z,
which shows that 1 − 2zC(z) is a square root of 1 − 4z. By Theorem 3.32, the
only square roots of 1 − 4z are (1 − 4z)1/2 and its negative; since 1 − 2zC(z)
has constant term 1, it must equal (1 − 4z)1/2 . Setting c = −4 in (3.15), we
deduce that
X∞
(−1)n−1 (2n − 3)!! (−4)n n
1 − 2zC(z) = 1 + n n!
z
n=1
2
∞
X 2n (2n − 3)!!
=1− zn
n=1
n!
∞
X 2n (2n − 1)!!
= 1 − 2z zn,
n=0
(n + 1)!
100 Topics in Discrete Mathematics

which implies that

∞
X 2n (2n − 1)!!
C(z) = zn. (3.23)
n=0
(n + 1)!
From this we can read off our desired formula for the Catalan numbers:
2n (2n − 1)!! (2n)!
cn = = , (3.24)
(n + 1)! (n + 1)!n!
where the second equality uses (2n − 1)!! = (2n)!
2n n!
(see Example

1.40). Note


1 2n 1 2n+1
that cn is not quite a binomial coefficient, but cn = n+1 n = 2n+1 n
.

Generating functions can be very helpful in finding closed formulas for sums
f (0) + f (1) + · · · + f (n). The reason is that if A(z) is the generating function
of a0 , a1 , a2 , · · · , then by (3.3) and the multiplication rule,
X ∞
A(z)
= (a0 + a1 + · · · + an ) z n , (3.25)
1−z n=0
A(z)
so 1−z
is the generating function of the sequence of partial sums of a0 , a1 , · · · .
Example 3.46. To find a closed formula for 0×20 +1×21 +· · ·+n×2n , we first
need a formula for the generating function of the sequence 0 × 20 , 1 × 21 , · · · :
by (3.11), this is
∞
X
n2n z n =
n=0

Now divide both sides by 1 − z, and use (3.25):

0×20 +1×21 +· · ·+n×2n = coefficient of z n in

We already found this coefficient in Example 3.27: it is (n − 1)2n+1 + 2.

Example 3.47. In Example 2.22 we were unable to guess a closed formula
for the sum of squares 12 +22 +· · ·+n2 . To solve this problem using generating
functions, we first need a formula for the generating function of the sequence
of squares itself, i.e. 02 + 12 z + 22 z 2 + · · · . We found this in Example 3.19:
∞
X
2 n z 2z 2
nz = + .
n=0
(1 − z)2 (1 − z)3
CHAPTER 3. GENERATING FUNCTIONS 101

(It is convenient not to combine the two terms.) Thus (3.25) tells us that
∞
X
z 2z 2
02 + 12 + · · · + n2 z n = +
n=0
(1 − z)3 (1 − z)4
X∞   X∞  
n+1 n n+1 n
= z +2 z .
n=0
2 n=0
3

(As in (3.9), we can let the summations start from n = 0, because the early
terms are zero anyway.) From this we read off the formula
   
2 2 2 n+1 n+1 (n + 1)n(2n + 1)
1 + 2 + ··· + n = +2 = . (3.26)
2 3 6

The generalization of Example 3.47 is the following:

Theorem 3.48*. For any positive integer a,
a
X  
a a a n+1
1 + 2 + ··· + n = k! S(a, k) .
k=1
k+1

Remark 3.49*. Expressing one sum as another sum may not seem like
progress, but the sum on the right-hand side has only a terms, and we are
imagining a as being fixed while n varies.

Proof*. Using Theorem 1.86 and (3.9), we get a formula for the generating
function of the sequence of ath powers:
X∞ Xa X∞   a
a n n n X zk
n z = k! S(a, k) z = k! S(a, k) .
n=0 k=1 n=0
k k=1
(1 − z)k+1

Then (3.25) tells us that

∞ n
! a
X X X zk
a n
m z = k! S(a, k) .
n=0 m=0 k=1
(1 − z)k+2
zk n+1


Since the coefficient of z n in (1−z)k+2
is k+1
, the result follows.

To conclude, here are a couple of examples which use our (rudimentary)

knowledge of differential equations of formal power series.
102 Topics in Discrete Mathematics

Example 3.50**. Consider the sequence defined by

1
a0 = a1 = 1, an = (an−1 + an−2 ) for n ≥ 2.
n
P∞ n
Let A(z) = n=0 an z be the generating function. The fact that nan is
involved in the recurrence relation prompts us to look at the derivative:
∞
X
0
A (z) = (n + 1)an+1 z n
n=0
∞
X
=1+ (an + an−1 ) z n
n=1
∞
X ∞
X
n
=1+ an z + an−1 z n
n=1 n=1
= 1 + (A(z) − 1) + zA(z)
= (1 + z)A(z).

Theorem 3.34 says that the solution of this (subject to a0 = 1) is

 
1 2
A(z) = exp z + z .
2

So the generating function is very concisely expressed. To get a formula for

an itself, we need a bit more work:
X∞  m
1 1 2
A(z) = z+ z
m=0
m! 2
∞
X 1  m
m 1
= z 1+ z
m=0
m! 2
∞ m

X 1 mX k k
m
= z z ,
m=0
m! k=0
2k

and extracting the coefficient of z n gives

n m

n
X n−m
X 1
an = n−m
= n−m
.
m=0
2 m! n
2 (n − m)!(2m − n)!
m=d 2 e
CHAPTER 3. GENERATING FUNCTIONS 103

However, this is not a closed formula, since the number of terms in the sum
grows as n grows. With sequences such as this, we often have to be satisfied
with giving a formula for the generating function.
Example 3.51**. Recall from Example 2.23 the sequence defined by
1
a0 = 1, an = (an−1 + 2an−2 + 3an−3 + · · · + na0 ) for n ≥ 1.
n
P∞
Letting A(z) = n=0 an z n be the generating function, we have
∞
X
0
A (z) = (n + 1)an+1 z n
n=0
∞
X
= (an + 2an−1 + 3an−2 + · · · + (n + 1)a0 ) z n
n=0
= (1 + 2z + 3z 2 + 4z 3 + · · · )A(z)
1
= A(z).
(1 − z)2

Again, this is a differential equation of the type we know how to solve: the
1 z 2 3
required ‘integral’ of (1−z)2 is 1−z = z + z + z + · · · , since we need it to
z


have zero constant term. So A(z) = exp 1−z . As in the previous example,
this does not lead to a closed formula for an .
 Part I

Introduction to Graph Theory

Anthony Henderson
2

Acknowledgements

These lecture notes were written in 2009 for the units MATH2069 Discrete
Mathematics and Graph Theory and MATH2969 Discrete Mathematics and
Graph Theory (Advanced), given at the University of Sydney. I am extremely
grateful to the previous lecturer Bill Palmer for allowing me to make use of
his lecture notes on graph theory.

I also acknowledge a debt to the following textbooks, which students should

consult for more examples and exercises.

• A First Look at Graph Theory by J. Clark and D. A. Holton

• Introduction to Graph Theory by R. J. Wilson

• Graph Theory by R. Diestel

Anthony Henderson
Contents

0 Introduction 1

1 First Properties of Graphs 5

1.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2 Connectedness and subgraphs . . . . . . . . . . . . . . . . . . 13

1.3 Degrees of vertices . . . . . . . . . . . . . . . . . . . . . . . . 20

2 Special Walks in Graphs 27

2.1 Eulerian graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.2 Hamiltonian graphs . . . . . . . . . . . . . . . . . . . . . . . . 32

2.3 Minimal walks in weighted graphs . . . . . . . . . . . . . . . . 37

3 Trees 47

3.1 Trees and Cayley’s Formula . . . . . . . . . . . . . . . . . . . 47

3.2 Spanning trees . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3
4 CONTENTS

3.3 Kirchhoff’s Matrix–Tree Theorem . . . . . . . . . . . . . . . . 62

4 Colourings of Graphs 71

4.1 Vertex colourings and chromatic number . . . . . . . . . . . . 71

4.2 The chromatic polynomial . . . . . . . . . . . . . . . . . . . . 80

4.3 Edge colourings . . . . . . . . . . . . . . . . . . . . . . . . . . 87

Chapter 0

Introduction

Graph theory is one of the most important and useful areas of discrete math-
ematics. To get a preliminary idea of what it involves, let us consider some
of the real-world problems which can be translated in graph-theoretic terms.

In the modern age of telecommunications, networks and their properties are

hot topics of study. Every time you download data from the internet, the
request passes from your computer to the computer where the data is actually
stored via a whole chain of intermediaries, each connected to the next. The
speed and efficiency of the process depends vitally on how well connected the
whole network is. One way to measure this connectedness is to determine
how many computers would need to crash before the other computers would
become unable to communicate with each other. For a simple example,
imagine six computers connected as follows.

If either of the middle two computers crashes, the others would be split into
two groups which could no longer communicate. To fix this problem, suppose

1
2 Introduction to Graph Theory

you add one more connection to the network.

Then the loss of any single computer would not be fatal: the remaining
computers would still be linked to each other. However, if you examine the
new network, you will be able to find two computers such that if they both
crashed simultaneously, the remaining four computers would be split into
two groups; if this represented an unacceptable risk, you would have to add
another connection. This sort of thinking about the nodes of a network and
their connections is quintessentially graph-theoretic.

Another widespread use of graph theory is in what are called scheduling

problems. Suppose there are various university courses which need to be
timetabled so that no students have a clash. To work out how many timetable
slots are required, you need to know which pairs of courses have a student in
common (because they then can’t be put on simultaneously). For example,
suppose there are five courses named A, B, C, D, E, and the pairs which have
a student in common are joined in the following picture.
A

E B

D C
Then clearly A, B, and C need to be allocated three different time-slots, but
D could be scheduled concurrently with B and E with C. Again the problem
amounts to considering abstract properties of the picture, and which courses
are connected to which. A similar problem confronts map-makers when they
have to choose a colour for each region of the map so that no two adjacent
regions have the same colour; again, the only information which is relevant is
the pattern of adjacencies between regions, which could be displayed in the
same sort of picture as above.

One of the most famous problems in graph theory is the Travelling Salesman
Problem. The original context was that a salesman wanted to visit a certain
CHAPTER 0. INTRODUCTION 3

number of towns, returning to his starting point, and having as short a total
distance to travel as possible. The information necessary to solve the problem
is not only which towns are connected to which, but how far apart they are
(considering only the shortest way of travelling between each pair of towns).
For example, there might be four towns A, B, C, D, where A is the salesman’s
home town, with distances in kilometres given in the following picture.
B

9 7
5
A C
8

9 6

One way to solve the problem would be to list all the possible routes and
calculate their distances: for instance, the route A–B–C–D–A has distance
9+7+6+9 = 31, whereas the route A–B–D–C–A has distance 9+5+6+8 =
28. The disadvantage of this approach is that if there are n towns, there are
(n−1)! possible routes, and as n increases, the growth rate of (n−1)! is faster
than exponential. There are many smarter algorithms which cut down the
amount of calculation involved, but it is a major open question whether there
is any algorithm for solving the Travelling Salesman Problem whose running
time has polynomial growth rate. The ramifications of such an algorithm
would be huge, because there are many problems in all sorts of areas which
can be rephrased in terms of finding the shortest paths through a diagram.

Comments on the text

In these notes, the main results are all called “Theorem”, irrespective of their
difficulty or significance. Usually, the statement of the Theorem is followed
(perhaps after an intervening Example or Remark) by a rigorous Proof; as is
traditional, the end of each proof is marked by an open box.

To enable cross-references, the Theorems, Definitions, Examples, and Re-

marks are all numbered in sequence (the number always starts with the rel-
evant chapter number). Various equations and formulas are also numbered
(on their right-hand edge) in a separate sequence.
4 Introduction to Graph Theory

In definitions (either formally numbered or just in the main text), a word

or phrase being defined is underlined. The symbol “:=” is used to mean “is
defined to be” or “which is defined to be”.

The text in the Examples and Remarks is slanted, to make them stand out
on the page. Many students will understand the Theorems more easily by
studying the Examples which illustrate them than by studying their proofs.
Some of the Examples contain blank boxes for you to fill in. The Remarks
are often side-comments, and are usually less important to understand.

The more difficult Theorems, Proofs, Examples, and Remarks are marked at
the beginning with either * or **. Those marked * are at the level which
MATH2069 students will have to understand in order to be sure of getting
a Credit, or to have a chance of a Distinction or High Distinction. Those
marked ** are really intended only for the students enrolled in the Advanced
unit MATH2969, and can safely be ignored by those enrolled in MATH2069.
Chapter 1

First Properties of Graphs

What the problems in the introduction have in common is that they are all
concerned with a collection of objects (computers, courses, towns) some of
which are linked together (in a literal physical sense, or by some property
such as two courses having a student in common). The abstract definition
of a graph is meant to capture this basic idea. (Note that the word “graph”
is used here in a different sense from the graphs of functions in calculus.)

1.1 Basic definitions

Here is the definition of graph we will use in this course.

Definition 1.1. A graph G is a pair (V, E) where:

• V is a finite set whose elements are called the vertices of G,

• E is a set of unordered pairs {v, w} where v, w ∈ V and v 6= w. The
elements of E are called the edges of G.

We say that v, w ∈ V are adjacent in G if {v, w} is one of the edges of G. If

e = {v, w} ∈ E, we say that v and w are the ends of the edge e.

5
6 Introduction to Graph Theory

In applications, the vertices of the graphs may be computers and the edges
may be connections between them, or the vertices may be towns and the
edges may be roads, and so forth. But for the purpose of developing the
theory, we will usually use positive integers or letters of the alphabet as the
vertices. Wherever possible, n will denote the number of vertices.

In drawing pictures of graphs, we will represent the vertices as little dots,

labelled by the appropriate number or symbol, and the edges as lines or
curves running from one end to the other. The positions of the dots have
no significance; nor do the shapes of the lines or the places where they may
happen to cross in the picture (other than at the dots). All that matters is
which vertices are adjacent to which.
Example 1.2. The following picture represents a graph with vertex set
{1, 2, 3, 4, 5, 6}.
1

2 4
3
5 6
The edges in this graph are as follows:

{1, 2}, {1, 4}, {2, 3}, {2, 4}, {3, 4}, {3, 5}, {3, 6}, {5, 6}.

So 1 is adjacent to 2 and to 4 but not to 3, 5, or 6; 3 is adjacent to all

the other vertices except for 1; and so forth. Any of the following pictures
represents the same graph.
1
1 1
2 4 2 4 2 4
5 2 4 6
3 3 3 1
6 5 5 6 5 6 3
Remark 1.3. Many authors allow more freedom in the definition of graph.
Specifically, their graphs can have multiple edges (more than one edge join-
ing the same two vertices) or loops (edges which have the same vertex for
both ends) – both of these are ruled out in our definition, where an edge is
just a set of 2 different vertices (those other authors would call our graphs
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 7

simple graphs). On the rare occasions where we need to discuss graphs with
multiple edges or loops, we will call them multigraphs. There is also an im-
portant concept of directed graph, where the edges are ordered pairs rather
than unordered pairs, and can hence be thought of as having a definite di-
rection from one end to the other. We will ignore all these variants for now.

Our first result about graphs is an illustration of basic counting principles.

Theorem 1.4. If |V | = n, then the number of graphs with vertex set V is

n

2( 2 ) . The number of edges in such a graph could be anything from 0 to n2 .
n

For fixed k, the number of graphs with vertex set V and k edges is ( 2 ) . k

Proof. If you fix the vertex set V , then the graph is determined by specifying
the edge set E. By definition, E is a subset of the set
X of two-element
subsets of V . We know that if |V | = n,
then |X| = n2 . So the size of a
subset of X can be anything from 0 to n2 ; the number of subsets of X of size
n

k is ( 2 ) ; and the total number of subsets of X is 2( 2 ) . (Recall the reason
n

k
for this last part: specifying a subset of X is the same as deciding,
for each
element of X, whether it is in or out. In our situation, we have n2 pairs of
vertices which are ‘potential edges’, and we have to decide for each pair of
vertices whether to join them or not.)


Example 1.5. Since 32 = 3, there are 8 = 23 graphs with vertex set
{1, 2, 3}:
1 1 1 1

2 3 2 3 2 3 2 3
1 1 1 1

2 3 2 3 2 3 2 3




Of these graphs,
30 = 1 has no edges, 31 = 3 have one edge, 3
2
= 3 have
two edges, and 33 = 1 has three edges.
8 Introduction to Graph Theory

There is a natural sense in which there are ‘really’ only four graphs involved
in Example 1.5: all the graphs with one edge have ‘the same form’, if you
forget about the labels of the vertices, and so have all the graphs with two
edges. This looser kind of sameness is called isomorphism (from the Greek
for “same form”). It is formalized in the following definition.
Definition 1.6. A graph G = (V, E) is said to be isomorphic to a graph
G′ = (V ′ , E ′ ) if there is a bijection between their vertex sets under which
their edge sets correspond; that is, a bijective function f : V → V ′ such that
E ′ = {{f (v), f (w)} | {v, w} ∈ E},
i.e. f (v), f (w) are adjacent in G′ if and only if v, w are adjacent in G.
Example 1.7. Consider the two graphs represented by these pictures.

b u
a c
v w x y
d
e f z
The first graph is (V, E) where V = {a, b, c, d, e, f } and
E = {{a, b}, {a, c}, {a, d}, {b, c}, {c, d}, {d, e}, {d, f }, {e, f }},
and the second graph is (V ′ , E ′ ) where V ′ = {u, v, w, x, y, z} and
E ′ = {{u, w}, {u, x}, {v, y}, {v, z}, {w, x}, {w, z}, {x, z}, {y, z}}.
Although their pictures look superficially different, (V, E) is isomorphic to
(V ′ , E ′ ) via the following bijection:
a ↔ w, b ↔ u, c ↔ x, d ↔ z, e ↔ v, f ↔ y.
You can check from the above listings that the edge sets correspond; a more
visual way to see the isomorphism is to re-draw the second graph so that it
obviously has the same form as the first graph.
b u
a c w x

d z
e f v y
Note that the bijection given above is not the only one we could have used:
it would be just as good to let e correspond to y and f to v, for instance.
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 9

Remark 1.8. To avoid confusion, the two graphs in Example 1.7 had differ-
ent vertex sets. But it is also possible for two graphs with the same vertex set
to be isomorphic. In the special case that V ′ = V , Definition 1.6 says that
(V, E) is isomorphic to (V, E ′ ) if and only if there is some permutation of the
vertices which makes the edge sets correspond; in other words, a picture of
(V, E ′ ) can be obtained from a picture of (V, E) by permuting the labels of
the vertices while leaving the edges where they are.

It is important to prove that isomorphism of graphs is an equivalence relation,

which means the following.
Theorem 1.9. (1) Any graph G is isomorphic to itself.

(2) If G is isomorphic to G′ , then G′ is isomorphic to G.

(3) If G is isomorphic to G′ and G′ to G′′ , then G is isomorphic to G′′ .

Proof. Let G = (V, E), G′ = (V ′ , E ′ ), G′′ = (V ′′ , E ′′ ). To prove (1), we can

use the most obvious bijective function from V to itself, namely the identity
map which leaves every vertex unchanged. To prove (2), if f : V → V ′ is a
bijection under which the edge sets of the two graphs correspond, then its
inverse f −1 : V ′ → V is another such bijection. To prove (3), let f : V → V ′
and g : V ′ → V ′′ be bijections such that

v, w adjacent in G ⇐⇒ f (v), f (w) adjacent in G′ and

v ′ , w′ adjacent in G′ ⇐⇒ g(v ′ ), g(w′ ) adjacent in G′′ .

Then the composition g ◦ f : V → V ′′ is also a bijection, and from the above

two equivalences we deduce that

v, w adjacent in G ⇐⇒ g(f (v)), g(f (w)) adjacent in G′′ .

So G is isomorphic to G′′ as required.

This means that it makes sense to classify graphs into isomorphism classes,
where two graphs are in the same isomorphism class if and only if they are iso-
morphic to each other; in visual terms, if and only if they can be represented
by the same picture with (possibly) different labellings of the vertices. We
represent isomorphism classes of graphs by pictures with unlabelled vertices.
10 Introduction to Graph Theory

Remark 1.10. Here are some elementary classification principles. To show

that two graphs are isomorphic, one needs to find a bijection between their
vertex sets under which the edges correspond, as in Example 1.7; in principle,
this requires a laborious search, although there could well be short-cuts. But
to show that two graphs are not isomorphic, one just needs to find one point
of ‘essential difference’ between them, one discrepancy which could not occur
if they were isomorphic. As Example 1.7 showed, the actual names of the
vertices may well differ from one isomorphic graph to another (that is an
‘inessential difference’); but the number of vertices clearly cannot differ, nor
can the number of edges. So the classification immediately breaks down
into a lot of sub-problems, such as “classify graphs with 7 vertices and 15
edges”. We will see many more properties of graphs, and they will all be of
this intrinsic kind which is unchanged under isomorphism. So every time we
define a new property of graphs, we have a new tool for showing that two
graphs are not isomorphic, and we can refine the classification further.

I should confess straight away that we will never actually complete any gen-
eral classification of graphs; in fact, our focus will shift away from this prob-
lem in later chapters. But it is useful to examine some small examples of
such classification, to get some familiarity with basic concepts.
Example 1.11. There is only one graph with 0 vertices (the vertex set and
the edge set both have to be the empty set). By contrast, there are arguably
infinitely many graphs with 1 vertex, because that vertex could be anything
you like (the edge set has to be empty). However, what really matters is that
there is only one isomorphism class of graphs with 1 vertex: they all have
the ‘same form’, since they just consist of a single vertex and no edges.
Example 1.12. If a graph has two vertices, it can either have 0 edges or
1 edge. Clearly all the graphs with two vertices and 0 edges are isomorphic
to each other, as are all the graphs with two vertices and 1 edge. Hence
there are two isomorphism classes of graphs with two vertices, which can be
represented by the following pictures.

Remember that there is no point labelling the vertices, because the names
of the vertices are irrelevant to the isomorphism class of a graph.
Example 1.13. As foreshadowed after Example 1.5, there are four isomor-
phism classes of graphs with three vertices, one for each of the possible num-
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 11

bers of edges (from 0 to 3):

For instance, we are claiming here that every graph with three vertices and
two edges must have a vertex where the two edges meet; this is clear, because
if the two edges had no ends in common there would have to be at least
2 × 2 = 4 vertices.
Example 1.14. There is only one isomorphism class of graphs with four
vertices and no edges, and similarly for four vertices and one edge. If a graph
has four vertices and two edges, the only essential information remaining is
whether the edges meet at a vertex or not. This gives rise to two isomorphism
classes:

Classifying graphs with four vertices and three edges is a little more inter-
esting. One way to approach it is to fix the specific vertex set {1, 2, 3, 4},
and draw pictures of all the graphs with this vertex set and three edges; by
4

Theorem 1.4, there are (32) = 20 of these. Examining the pictures, we can
easily classify them into the following isomorphism classes:

We will develop more systematic ways of doing this classification later. If you
start considering the possible configurations of four edges, you will quickly
realize that it is easier to classify these graphs according to the two ‘non-
edges’, which must either meet at a vertex or not; this gives two isomorphism
12 Introduction to Graph Theory

classes:

By the same reasoning, there is only one isomorphism class of graphs with
four vertices and five edges (and one non-edge), and clearly there is only one
isomorphism class of graphs with four vertices and six edges (where every
possible edge is included).

We can formalize the idea of considering the ‘non-edges’ of a graph as follows.

Definition 1.15. The complement of the graph G = (V, E) is the graph
G = (V, E) where E is the complement of E in the set of 2-element subsets
of V . In other words, G has the same vertex set V as G, and for v 6= w ∈ V ,
v, w are adjacent in G if and only if v, w are not adjacent in G.
Example 1.16. The complement of
1 2 1 2
is
3 4 3 4
′
Theorem 1.17. Two graphs G and G are isomorphic if and only if their
complements G and G′ are isomorphic.

Proof. Let G = (V, E), G′ = (V ′ , E ′ ). The condition for G and G′ to be

isomorphic is that there is a bijective function f : V → V ′ such that for
v 6= w ∈ V ,
v, w adjacent in G ⇐⇒ f (v), f (w) adjacent in G′ .
But by basic logic, this condition is unchanged if “adjacent” is changed to
“not adjacent” on both sides, in which case it becomes the condition for G
and G′ to be isomorphic.

n


Note that if G has n vertices and k edges, then G has

n vertices and 2
−k
edges. So classifying graphs with n vertices and n2 − k edges is essentially
the same as classifying graphs with n vertices and k edges.
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 13

1.2 Connectedness and subgraphs

One glaring difference between some of the graphs we classified in the previ-
ous section was that sometimes there was a way to get from every vertex to
every other vertex along the edges, and sometimes there wasn’t: if the edges
represent lines of communication, then sometimes all the vertices could com-
municate with each other and sometimes they split into more than one group.
Definition 1.18. Let G = (V, E) be a graph. If v, w ∈ V , a walk from v to
w in the graph G is a sequence of vertices
v0 , v1 , v2 , · · · , vℓ ∈ V, with v0 = v and vℓ = w,
such that vi and vi+1 are adjacent for all i = 0, 1, · · · , ℓ − 1, or in other words
the following are all in E:
{v0 , v1 }, {v1 , v2 }, · · · , {vℓ−1 , vℓ }.
The length of such a walk is the number of steps, i.e. ℓ. We say that v is
linked to w in G if there exists a walk from v to w in G.
Theorem 1.19. In any graph G = (V, E), the relation of being linked is an
equivalence relation on the vertex set V . In other words:

(1) Every vertex v ∈ V is linked to itself.

(2) If v is linked to w, then w is linked to v.
(3) If v is linked to w and w is linked to x, then v is linked to x.

Proof. For (1), we can use the walk of length 0 which consists solely of v.
For (2), if v0 , v1 , · · · , vℓ is a walk from v to w, then its reversal vℓ , · · · , v0
is a walk from w to v. The proof of part (3) uses the fact that we can
concatenate walks: if v0 , v1 , · · · , vℓ is a walk from v to w and w0 , w1 , · · · , wk
is a walk from w to x, then v0 , v1 , · · · , vℓ , w1 , · · · , wk is a walk from v to x,
since vℓ = w = w0 .

As a consequence, the vertex set V is the disjoint union of the equivalence

classes for the equivalence relation of being linked.
14 Introduction to Graph Theory

Definition 1.20. For any vertex v of a graph G, the connected component

of G containing v is the graph whose vertex set consists of all vertices of G
which are linked to v, where two vertices are adjacent if and only if they are
adjacent in G. We say that G is connected if it has one connected component,
i.e. the vertex set is nonempty and every vertex is linked to every other vertex.
Example 1.21. Of the two graphs in Example 1.16, the first breaks into two
connected components: the vertices 1, 3, 4 are all linked to each other, while
the vertex 2 is not linked to any of these. The second graph is connected.

At the extremes of connectedness we have the following cases.

Definition 1.22. For a fixed vertex set V , the null graph is the graph (V, ∅)
in which no two vertices are adjacent; the complete graph is the graph (V, X)
where X is the set of all 2-element subsets of V , or in other words any two
vertices are adjacent (this is the complement of the null graph). We write Nn
and Kn for the null graph and complete graph with vertex set {1, 2, · · · , n}.

Note that Nn has 0 edges whereas Kn has n2 , and Nn has n connected com-
ponents (every vertex is a component on its own) whereas Kn is connected.
Example 1.23. Here are the pictures of Kn for n ≤ 4:
1 1 2

1 1 2
2 3 3 4
K1 K2 K3 K4
The following result means that the classification of graphs up to isomorphism
reduces to the case of connected graphs.
Theorem 1.24*. Let G and G′ be graphs.

(1) Suppose that G and G′ are isomorphic. Then G is connected if and

only if G′ is connected. More generally, the connected components of
G must be isomorphic to corresponding connected components of G′ :
that is, if the connected components of G are G1 , · · · , Gs , it must be
possible to number the connected components of G′ as G′1 , · · · , G′s in
such a way that Gi is isomorphic to G′i for all i.
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 15

(2) Conversely, if the connected components of G are G1 , · · · , Gs and those

of G′ are G′1 , · · · , G′s , and Gi is isomorphic to G′i for all i, then G is
isomorphic to G′ .

Proof*. Let G = (V, E), G′ = (V ′ , E ′ ). In part (1) we are assuming the

existence of a bijection f : V → V ′ which respects adjacency; it follows
easily that v, w are linked in G if and only if f (v), f (w) are linked in G′ .
Hence the connected components of G′ are just obtained from the connected
components G1 , · · · , Gs of G by applying f to each vertex set, and (1) follows.
In part (2), it is clear that we can combine the individual bijections between
the vertices of Gi and those of G′i for each i into an overall bijection f : V →
V ′ . By definition there are no edges in G or G′ other than those within a
connected component, so f respects adjacency as required.
Remark 1.25. Theorem 1.24 is fiddly to state and to prove (and indeed we
cheated slightly in the proof by omitting details which would have required
too much notation). Yet it really expresses nothing more than the obvious
idea that connected components are an intrinsic property of a graph and
don’t depend on the particular labelling of the vertices. There are several
basic results in graph theory which are obvious once one gets the idea, yet
tricky to prove rigorously. From now on we will often omit the proofs of such
results.

Connected components of a graph are a special case of the following idea.

Definition 1.26. If G = (V, E) is a graph, a subgraph of G is a graph

H = (W, F ) where W is a subset of V and F is a subset of E. The second
of these statements means that for v, w ∈ W ,

v, w are adjacent in H =⇒ v, w are adjacent in G.

Note that the converse is not required to hold. A subgraph with the same
vertex set as the whole graph is called a spanning subgraph.

A less formal way to express this definition is to say that a subgraph is

something obtained from the graph by deleting various vertices and edges,
and a spanning subgraph is obtained by deleting edges only.
16 Introduction to Graph Theory

Example 1.27. If e is an edge of G, you can form a spanning subgraph of G

by removing e from the set of edges and leaving everything else unchanged:
this subgraph will be called G − e. Conversely, if e is an edge of the comple-
ment G (i.e. it is a pair of non-adjacent vertices of G), then G + e denotes
the graph obtained by adding e to the set of edges of G; so G is a spanning
subgraph of G + e.
Example 1.28. Any graph G with vertex set {1, 2, · · · , n} is a spanning
subgraph of Kn ; you can start with Kn and delete all the edges in the com-
plement G to obtain G.
Example 1.29. If v is a vertex of G, the subgraph G − v is obtained from
G by removing v from the set of vertices and also (as is then necessary)
removing all edges which ended at v. For instance,
1 2 1 2
if G is then G − 4 is
3 4 3

It is clear that if G is isomorphic to G′ , then every subgraph of G is isomor-

phic to a corresponding subgraph of G′ . So a natural way to approach the
classification of graphs is to consider the subgraphs which belong to various
isomorphism classes. For example, if G has a subgraph isomorphic to K4 (i.e.
it contains four vertices which are all adjacent to each other) and G′ doesn’t,
then G and G′ are definitely not isomorphic. Here are two particularly useful
kinds of subgraph to consider.
Definition 1.30. The path graph Pn has vertex set {1, 2, · · · , n} and edges
{1, 2}, {2, 3}, · · · , {n − 1, n}:
1 1 2 1 2 3
···
P1 P2 P3
For n ≥ 3, the cycle graph Cn is obtained from Pn by adding the edge {n, 1}:
1 1 2

···

3 2 4 3
C3 C4
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 17

Definition 1.31. For any graph G, a path in G is a subgraph of G which

is isomorphic to Pn for some n: that is, a collection of distinct vertices
{v1 , v2 , · · · , vn } such that vi is adjacent to vi+1 for i = 1, · · · , n − 1, together
with the edges {vi , vi+1 }. The length of such a path is n − 1; its end-vertices
are v1 and vn . A cycle in G is a subgraph of G which is isomorphic to Cn for
some n ≥ 3: that is, a collection of distinct vertices {v1 , v2 , · · · , vn } such that
vi is adjacent to vi+1 for i = 1, · · · , n − 1 and vn is adjacent to v1 , together
with the edges {vi , vi+1 } and {vn , v1 }. The length of such a cycle is n, and a
cycle of length n is called an n-cycle for short.
Remark 1.32. Notice the differences between the definition of a path or a
cycle and that of a walk. Firstly, paths and cycles in a graph contain edges
as well as vertices, whereas a walk as defined in Definition 1.18 consists of a
sequence of vertices only (although the edges between them are involved in
the definition). Secondly, the vertices of a path or cycle must all be different,
whereas a walk is allowed to use a vertex more than once. Thirdly, there is
no specific ordering on the vertices of a path or cycle beyond what is implied
by the edges. Thus the path with vertices v1 , · · · , vn referred to in Definition
1.31 is the same as the path with vertices vn , · · · , v1 , although there are two
different walks of length n − 1 ‘along’ this path, one from v1 to vn and one
from vn to v1 . Similarly, the cycle with vertices v1 , · · · , vn referred to in
Definition 1.31 is the same as the cycle with vertices v2 , · · · , vn , v1 , and the
same as the cycle with vertices v3 , · · · , vn , v1 , v2 , and so forth. For any one
of these vertices vi , there are two different walks of length n ‘along’ the cycle
from vi back to itself, namely vi , vi+1 , · · · , vn , v1 , · · · , vi−1 , vi and its reversal.
Example 1.33. In the complete graph K4 , the number of:

walks of length 3 from vertex 1 to vertex 4 is

paths of length 3 with end-vertices 1 and 4 is

walks of length 3 from vertex 1 to itself is

cycles of length 3 containing vertex 1 is

The numbers of paths or cycles of various lengths in a graph are obviously

isomorphism-invariant, so potentially useful for classification.
18 Introduction to Graph Theory

There is a relationship between paths, cycles and connectedness.

Theorem 1.34. Let G be a graph, and v, w distinct vertices of G.

(1) v and w are linked in G if and only if there is a path in G with end-
vertices v and w.

(2) If e = {v, w} is an edge of G, then v and w are linked in G − e if and

only if there is a cycle in G containing e.

Proof*. For part (1), recall that v and w are said to be linked in G if there
is a walk in G from v to w. So the “if” direction is clear: if there is a path
with end-vertices v and w, then there is a walk along that path from v to
w. For the “only if” direction, we assume that there is a walk from v to
w in G, say v = v0 , v1 , v2 , · · · , vℓ = w. We need to produce a path with
end-vertices v and w; the problem is that there may be repeated vertices in
the walk. But if vi = vj for some i < j, then the part of the walk between
vi and vj is a needless detour; we can cut it out to obtain the shorter walk
v0 , · · · , vi , vj+1 , · · · , vℓ . Continuing in this way, we must eventually obtain a
walk from v to w with no repeated vertices, which (together with the edges
between consecutive vertices) constitutes a path as required.

In part (2), if there is a cycle containing e, we may as well number its vertices
v = v1 , v2 , · · · , vn = w, and then the walk v1 , v2 , · · · , vn shows that v and
w are linked without using the edge e. Conversely, suppose that v and w
are linked in G − e. By part (1), there is a path in G − e with vertices
v = v1 , v2 , · · · , vn = w, and we can add the edge e to this path to form a
cycle. (We cannot have n = 2, because then the path would contain e.)

Definition 1.35. An edge e = {v, w} in a graph G is said to be a bridge

if v and w are not linked in G − e, i.e. e is not contained in any cycle. In
particular, if G is connected, then e is a bridge if and only if G − e is not
connected, i.e. removing e disconnects the graph.

The reason for the name is that such an edge e is the only way to get from
v to w, as if they were separated by water and e was the only bridge.
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 19

Example 1.36. In the graph with the following picture

a b c w x

d z
e f v y

the bridges are {a, d}, {b, c}, {c, d}, and {d, z}.

We can now prove a result bounding the number of edges in a graph.

Theorem 1.37. Let G be a graph with n vertices and k edges.

n


(1) If G is connected, then n − 1 ≤ k ≤ 2
.


n−s+1
(2) If G has s connected components, then n − s ≤ k ≤ 2
.

Proof*. In part (1), we already know the upper bound k ≤ n2 , so the only
new thing to prove is that connectedness requires at least n − 1 edges. We
can prove this by induction on n; the n = 1 base case is obvious. Assume
that n ≥ 2 and that we know the result for graphs with fewer than n vertices.
Suppose that G contains a bridge e = {v, w}. Since G is connected, every
vertex in G − e must be linked to either v or w; that is, G − e has two
connected components, the component G1 containing v and the component
G2 containing w. Suppose that Gi has ni vertices and ki edges, for i = 1, 2.
Then n1 + n2 = n, so n1 , n2 < n; thus G1 and G2 are graphs to which
the induction hypothesis applies, and we conclude that k1 ≥ n1 − 1 and
k2 ≥ n2 − 1. Hence

k = k1 + k2 + 1 ≥ (n1 − 1) + (n2 − 1) + 1 = n1 + n2 − 1 = n − 1.

On the other hand, if G does not contain a bridge, then if we delete any
edge it remains connected; if we continue to delete edges, we must eventually
reach a spanning subgraph which does contain a bridge. Since the number
of edges in this subgraph is at least n − 1, the number of edges in G is even
more. So in either case, k ≥ n − 1 and the inductive step is complete.
20 Introduction to Graph Theory

In part (2), let G1 , · · · , Gs be the connected components of G, and suppose

that Gi has ni vertices and ki edges, so that n = n1 + · · · + ns and k =
k1 + · · · + ks . By part (1) we have ki ≥ ni − 1 for all i, so

k ≥ (n1 − 1) + · · · + (ns − 1) = (n1 + · · · + ns ) − s = n − s,

proving the required lower bound. We also have ki ≤ n2i for all i, so
   
n1 ns
k≤ + ··· +
2 2
1
= (n1 (n1 − 1) + · · · + ns (ns − 1))
2
1
≤ (n − s + 1)((n1 − 1) + · · · + (ns − 1)) (since ni ≤ n − s + 1 for all i)
2
1
= (n − s + 1)(n − s)
2 
n−s+1
= ,
2

proving the required upper bound.

Remark 1.38. The upper bound in part (1) is attained exactly when G
is a complete graph, isomorphic to Kn . If you examine the proof of (2),
you will see that the upper bound there is attained exactly when one of the
connected components of G is a complete graph and the others consist of
single points. We will explore the cases when the lower bounds are attained
in a later chapter.

1.3 Degrees of vertices

One way to analyse a graph is to narrow one’s focus to what is happening in

the neighbourhood of each vertex. This requires some definitions.

Definition 1.39. Let G be a graph. For any vertex v of G, its degree

degG (v) or just deg(v) is the number of edges of G which have v as an
end; equivalently, the number of vertices of G which are adjacent to v. The
degree sequence of G consists of the degrees of the vertices of G arranged in
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 21

weakly increasing order: if the vertices of G are numbered v1 , v2 , · · · , vn in

such a way that deg(v1 ) ≤ deg(v2 ) ≤ · · · ≤ deg(vn ), then the degree sequence
of G is (deg(v1 ), deg(v2 ), · · · , deg(vn )). We use some special notation for the
first and last entries of this sequence:

δ(G) = the smallest degree of any vertex of G,

∆(G) = the largest degree of any vertex of G.

If δ(G) = ∆(G), i.e. every vertex has the same degree d, we say that G is
regular of degree d.

It is obvious that if G has n vertices, then

0 ≤ δ(G) ≤ ∆(G) ≤ n − 1. (1.1)

Example 1.40. If G is the graph with the following picture:

a

e b

d c
then the degrees of the vertices are

deg(a) = 4, deg(b) = 2, deg(c) = 3, deg(d) = 3, deg(e) = 2.

So the degree sequence of G is (2, 2, 3, 3, 4) (note the conventional increasing

order), and δ(G) = 2, ∆(G) = 4.
Example 1.41. Some of our families of examples are regular:

the null graph Nn is regular of degree

the complete graph Kn is regular of degree

the cycle graph Cn is regular of degree

22 Introduction to Graph Theory

Example 1.42. There is a famous graph called the Petersen graph which
has 10 vertices and is regular of degree 3:

This is really an isomorphism class of graphs, because there is no particularly

privileged way to label the vertices. In fact, the vertices play completely sym-
metrical roles. In the picture, the inner 5 vertices are joined in a pentagram
pattern, but that is just another way of drawing a 5-cycle. It is crucial that
the vertices in the inner 5-cycle are joined to the vertices in the outer 5-cycle
in a way which is “out of step” with the cycles. In other words, the Petersen
graph is not isomorphic to the graph(s) with the following picture:

In the latter graph there are only two 5-cycles, the inner ring and the outer
ring; but in the Petersen graph, there are many ways other than the most
visible way to separate the vertices into two 5-cycles.

It is obvious that isomorphic graphs have the same degree sequence, so con-
sidering the degree sequence can be a useful way to distinguish between
non-isomorphic graphs. (Unfortunately, the preceding example shows that
two non-isomorphic graphs can have the same degree sequence, so it is not a
panacea.) The degree sequence contains some of the pieces of information we
have previously used for classification: the number of vertices is the number
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 23

of terms in the degree sequence, and the number of edges can be derived
from the sum of the terms, by the following result.

Theorem 1.43 (Hand-shaking Lemma). For any graph, the number of edges
is half the sum of the degrees of all the vertices.

Proof. This is a simple application of the Overcounting Principle: since

every edge has two ends, when you add up the numbers of edges at each
vertex you count every edge exactly twice.
Example 1.44. The name comes from imagining the context of a party
where various people shake hands upon meeting. If you ask every one at the
end of the night how many hands they shook, add up those numbers and
divide by two, you get the total number of hand-shakes. (Here the vertices
are the party guests, two of whom are adjacent if they shook hands.)
Example 1.45. Purely from the degree sequence (2, 2, 3, 3, 4) of the graph
in Example 1.40, we can see that the number of edges is 2+2+3+3+4
2
= 7.
10×3
Example 1.46. The number of edges in the Petersen graph is 2
= 15.

To use degree sequences in the classification of graphs, we would want a

way of deciding, given a sequence (d1 , d2 , · · · , dn ) of nonnegative integers in
weakly increasing order, whether there exists a graph with this as its degree
sequence; for brevity, we say that the sequence is graphic if there does exist
such a graph. There are some obvious requirements that (d1 , d2 , · · · , dn )
must satisfy in order to be graphic. For instance, the Hand-shaking Lemma
shows that d1 + · · · + dn must be even, which is equivalent to saying that
the number of odd di ’s is even. Another obvious requirement is that if dn
is nonzero, there must be at least dn other di ’s which are nonzero; because
if a vertex is to have degree dn , there must be dn other vertices which are
adjacent to it. There are other such constraints.
Example 1.47. Is the sequence (1, 1, 3, 3) graphic? That is, does there exist
a graph G with vertices v1 , v2 , v3 , v4 such that degG (v1 ) = 1, degG (v2 ) = 1,
degG (v3 ) = 3, degG (v4 ) = 3? The answer is
24 Introduction to Graph Theory

Example 1.48. Is the sequence (1, 2, 2, 3, 3, 3) graphic? Let’s try to actually

construct a graph G with vertices v1 , v2 , · · · , v6 such that degG (v1 ) = 1,
degG (v2 ) = 2, degG (v3 ) = 2, degG (v4 ) = 3, degG (v5 ) = 3, degG (v6 ) = 3.
We need to make v6 adjacent to three other vertices. Crossing our fingers,
we arbitrarily specify that these should be v3 , v4 , and v5 (because it seems
that we will give ourselves the best chance by using vertices whose degrees
are as large as possible). With that assumption, consider the graph G − v6
obtained by removing the vertex v6 and all the edges ending at it. We
have degG−v6 (v1 ) = 1, degG−v6 (v2 ) = 2, degG−v6 (v3 ) = 1, degG−v6 (v4 ) = 2,
degG−v6 (v5 ) = 2. So the degree sequence of the hypothetical graph G − v6
is (1, 1, 2, 2, 2). But the latter sequence is definitely graphic: it is the degree
sequence of the path graph P5 . So we can join the extra vertex v6 to the
appropriate vertices of P5 , and we have succeeded in constructing G:
v6

v1 v2 v4 v5 v3
So the answer to the original question is that (1, 2, 2, 3, 3, 3) is graphic.

To make such arguments into a hard-and-fast rule, we need the following

result, which shows that the arbitrary stipulation made in Example 1.48 was
not so arbitrary after all.
Theorem 1.49 (Havel-Hakimi Theorem). Let (d1 , d2 , · · · , dn ) be a weakly
increasing sequence of nonnegative integers. If (d1 , d2 , · · · , dn ) is graphic,
then there is a graph G with vertices v1 , v2 , · · · , vn such that deg(vi ) = di for
i = 1, 2, · · · , n and the vertices adjacent to vn are vn−1 , vn−2 , · · · , vn−dn .

Proof*. By the definition of “graphic”, there is a graph G with vertices

v1 , v2 , · · · , vn such that deg(vi ) = di for i = 1, 2, · · · , n; all we don’t know is
which of the vertices v1 , · · · , vn−1 are adjacent to vn . Let W be the set of
vertices adjacent to vn ; thus |W | = dn . Let d be the sum of the degrees (in
the graph G) of all the vertices in W . Since the sequence (d1 , d2 , · · · , dn ) is
weakly increasing,
d ≤ dn−1 + dn−2 + · · · + dn−dn . (1.2)
If equality holds in (1.2), then the vertices in W have the same collection of
degrees as vn−1 , vn−2 , · · · , vn−dn ; so after possibly renumbering some vertices
CHAPTER 1. FIRST PROPERTIES OF GRAPHS 25

with the same degree, we can arrange that W = {vn−1 , vn−2 , · · · , vn−dn } as
required. On the other hand, if the inequality in (1.2) is strict, there must
be some vertex w in W whose degree is strictly less than that of some vertex
x ∈ {v1 , · · · , vn−1 } \ W . The inequality deg(w) < deg(x) implies that there
is some vertex y 6= w which is adjacent to x but not to w. Thus vn , w, x, y
are four distinct vertices of G such that {w, vn } and {x, y} are edges of G,
but {x, vn } and {w, y} are not. Now we modify G by deleting the edges
{w, vn } and {x, y}, and adding the edges {x, vn } and {w, y}. Since each of
vn , w, x, y is involved in one of the deleted edges and one of the added edges,
this modification has not changed any of the degrees; however, since w has
been replaced by x in the set W , the quantity d has strictly increased. After
repeating this step a finite number of times, we must reach a graph for which
equality holds in (1.2), and then the result follows as seen before.

We can deduce the following recursive rule for determining whether a se-
quence is graphic.
Theorem 1.50. Let (d1 , d2 , · · · , dn ) be a weakly increasing sequence of non-
negative integers. We have three cases.

(1) If all the di ’s are zero, the sequence is graphic.

(2) If the number of nonzero di ’s is between 1 and dn , the sequence is not
graphic.
(3) If the number of nonzero di ’s is at least dn + 1 (i.e. dn−dn ≥ 1), let
(e1 , · · · , en−1 ) be the sequence obtained from (d1 , d2 , · · · , dn ) by carry-
ing out the following steps:
(a) remove the last term of the sequence, dn ;
(b) subtract 1 from each of the next dn terms from the end, i.e.
dn−1 , · · · , dn−dn ;
(c) if necessary, rearrange in weakly increasing order.
Then (d1 , d2 , · · · , dn ) is graphic if and only if (e1 , · · · , en−1 ) is graphic.

Proof*. Case (1) is obvious, because the null graph Nn has degree sequence
(0, 0, · · · , 0). We have already seen the reason for case (2): a vertex of
26 Introduction to Graph Theory

degree dn requires dn other vertices of nonzero degree to be adjacent to it.

So the main content of the result is the “if and only if” statement in case
(3). We prove the “only if” direction first, supposing that (d1 , d2 , · · · , dn ) is
graphic. By the Havel–Hakimi Theorem, there is a graph G with vertices
v1 , v2 , · · · , vn such that deg(vi ) = di and the vertices adjacent to vn are
vn−1 , vn−2 , · · · , vn−dn . By contruction, (e1 , · · · , en−1 ) is the degree sequence
of G − vn , and it is hence graphic. To prove the converse direction, we
can reverse the construction: starting from a graph H with degree sequence
(e1 , · · · , en−1 ), we construct a graph G with degree sequence (d1 , · · · , dn ) by
adding a new vertex which is adjacent to the appropriate vertices of H.

In case (3), Theorem 1.50 does not immediately decide whether (d1 , · · · , dn )
is graphic or not, but it reduces the question to the same question for the
shorter sequence (e1 , · · · , en−1 ); we can then apply Theorem 1.50 again to
this shorter sequence and so on, and since the sequence can’t go on getting
shorter indefinitely, we must eventually land in either case (1) or case (2).
(Of course, we can stop before that point if the sequence becomes obviously
graphic or not graphic by some other reasoning.)
Example 1.51. The sequence (1, 1, 1, 1, 4, 4) falls into case (3). To obtain
the shorter sequence, we apply the three steps prescribed in Theorem 1.50:

(1, 1, 1, 1, 4, 4) −→ (1, 1, 1, 1, 4) −→ (1, 0, 0, 0, 3) −→ (0, 0, 0, 1, 3).

The sequence (0, 0, 0, 1, 3) falls into case (2), so it is not graphic; hence the
original sequence (1, 1, 1, 1, 4, 4) is not graphic.
Remark 1.52. Since Havel and Hakimi provided the solution to the basic
question of whether a sequence is the degree sequence of a graph, there
has been much research on more refined questions: for instance, replacing
“graph” by “connected graph”. Unfortunately we will have to omit these
later developments.
Chapter 2

Special Walks in Graphs

The origin of graph theory was a question posed to the great 18th-century
mathematician Euler by the citizens of Königsberg. Their city was built
around two islands in the river Pregel, and included seven bridges joining
the islands to the banks and to each other; they wanted to know whether
it was possible to walk through the city crossing every bridge exactly once.
To decide this, Euler was led to introduce the concept of a graph. Many
subsequent applications of graph theory also involve special kinds of walks.

2.1 Eulerian graphs

Definition 2.1. An Eulerian circuit in a connected graph G is a walk in G,

say v0 , v1 , v2 , · · · , vℓ , which:

(1) uses every edge exactly once, i.e. for every {v, w} in V there is exactly
one i such that {vi , vi+1 } = {v, w};

(2) returns to its starting point, i.e. vℓ = v0 .

A connected graph G is said to be Eulerian if it has an Eulerian circuit.

27
28 Introduction to Graph Theory

It is clear that if an Eulerian circuit does exist, you can make it start (and
therefore also finish) at any vertex you want.
Example 2.2. Consider the following two graphs.
a a

e b e b

d c d c
The first is clearly Eulerian: an example of an Eulerian circuit is the walk
a, b, c, a, d, e, a. But if you try to find an Eulerian circuit in the second graph,
you will get stuck: the extra edge {c, d} can’t be fitted into the walk with-
out repeating some other edge. The best you can do is a walk such as
d, e, a, d, c, a, b, c, which uses every edge exactly once but doesn’t return to
its starting point. So the second graph is not Eulerian.
Example 2.3. The cycle graph Cn is obviously Eulerian: a walk around the
cycle is an Eulerian circuit.
Remark 2.4. We restricted to connected graphs in Definition 2.1, because
if a graph has more than one connected component which contains an edge,
there clearly cannot be a walk which uses all the edges of the graph.
Remark 2.5*. We may seem to be avoiding some interesting examples of
this problem by not allowing multiple edges in our graphs. Indeed, the
bridges of Königsberg formed a multigraph and not a graph, because some
of the landmasses were joined by more than one bridge. But given any multi-
graph, the question of whether it has an Eulerian circuit can be rephrased in
terms of a graph, namely the one where you sub-divide every multiple edge
into two edges by putting a new vertex in the middle of it.

The reason that the first graph in Example 2.2 is so obviously Eulerian is that
its edges can be split up into two 3-cycles. Euler realized that this property
is crucial to the existence of an Eulerian circuit, and that it depends on the
evenness of the vertex degrees; what is wrong with the vertices c and d in
the second graph in Example 2.2 is that their degrees are odd.

Theorem 2.6. A connected graph G is Eulerian if and only if deg(v) is even

for every vertex v of G.
CHAPTER 2. SPECIAL WALKS IN GRAPHS 29

Proof. We first prove the “only if” direction, that in an Eulerian graph all
the degrees are even. Let v0 , v1 , · · · , vℓ be an Eulerian circuit. A vertex v
may appear more than once in this circuit, but every time it does appear,
two of the edges ending at v are used: namely, {vi−1 , vi } and {vi , vi+1 } if
v = vi , or {v0 , v1 } and {vℓ−1 , vℓ } if v = v0 = vℓ . Since every edge is used
exactly once, the total number of edges ending at v must be even.

To prove the “if” direction, we use induction on the number of edges. The
base case of this induction is the case where there are no edges at all, i.e. the
graph consists of a single vertex; this graph is trivially Eulerian (the Eulerian
circuit has length 0). So we assume that G does have some edges, and that
all the vertex degrees are even. Our induction hypothesis says that every
connected graph with fewer edges than G and all degrees even is Eulerian,
and we want to prove that G is also. We start by showing that G contains
a cycle. Pick any vertex v0 ; since G is connected, there must be some other
vertex v1 which is adjacent to v0 ; since deg(v1 ) can’t be 1, there must be some
vertex v2 6= v0 which is adjacent to v1 ; since deg(v2 ) can’t be 1, there must
be some vertex v3 6= v1 which is adjacent to v2 ; and we can continue in this
way indefinitely. Because there are only finitely many vertices, there must
be some k < ℓ such that vk = vℓ , and we can assume that vk , vk+1 , · · · , vℓ−1
are all distinct. The subgraph C which consists of these distinct vertices and
the edges between them (including the edge {vℓ−1 , vk }) is a cycle.

Now let H be the subgraph of G obtained by deleting all the edges in C (but
keeping all the vertices). The vertex degrees in H are the same as those in
G, except that 2 is subtracted from the degree of every vertex in C; so the
vertex degrees in H are all even. H need not be connected, but if we let
H1 , H2 , · · · , Hs be the connected components of H, then each Hi is Eulerian
by the induction hypothesis. Since G is connected, every Hi must contain
at least one of the vertices vk , · · · , vℓ−1 of C; let ki be such that vki lies in
Hi , and renumber if necessary so that k1 < k2 < · · · < ks . Then we can
construct an Eulerian circuit in G as follows:

vk , · · ·, vk1 , ·c
· ·, vk1 , · · ·, vk2 , ·c
· ·, vk2 , · · · , vks , ·c
· ·, vks , · · ·, vℓ−1 , vk ,

where · · · indicates a walk around the appropriate part of the cycle C, and
·c
· · indicates an Eulerian circuit in the appropriate Hi starting and finishing
at vki . This completes the induction step.
30 Introduction to Graph Theory

Theorem 2.6 not only gives us a simple criterion for the existence of an
Eulerian circuit; its proof supplies a recursive way of finding one.
Example 2.7. Let G be the following graph, in which all degrees are even.
b w
a x
c z
d y
e v
Theorem 2.6 guarantees that an Eulerian circuit exists. To use the construc-
tion in the proof, we need to choose a cycle C, say the 4-cycle with vertices
b, w, v, e. After removing the edges of C, we obtain the graph H:
b w
a x
c z
d y
e v
We now need to find an Eulerian circuit in H; again, the first step is to
remove a cycle, say the 4-cycle with vertices a, b, z, c.
b w
a x
c z
d y
e v
This leaves four connected components, and Eulerian circuits in these are
obvious: the length-0 walks a and b in those components, the walk c, d, e, c,
and the walk z, v, y, z, w, x, z. We now patch these into the appropriate places
in the cycle a, b, z, c to form an Eulerian circuit in H:

a, b, z, v, y, z, w, x, z, c, d, e, c, a.

Finally, we make this circuit start at b instead, and patch it into the original
cycle b, w, v, e to form an Eulerian circuit in G:

b, z, v, y, z, w, x, z, c, d, e, c, a, b, w, v, e, b.

Of course, we made many choices along the way, so the result is not unique.
CHAPTER 2. SPECIAL WALKS IN GRAPHS 31

There is an easy variant of Theorem 2.6 for the case of an Eulerian trail,
which is a walk that uses every edge exactly once but finishes at a different
vertex from where it started. (We saw such a walk in Example 2.6.)

Theorem 2.8. A connected graph G has an Eulerian trail if and only if it

has exactly two vertices of odd degree. Moreover, if this the case, the trail
must start at one of these vertices of odd degree and finish at the other.

Proof. The “only if” direction, that a graph with an Eulerian trail must
have exactly two vertices of odd degree, works in the same way as in the
proof of Theorem 2.6. The difference is that the first vertex and last vertex
of the trail are no longer the same, so their appearances at the beginning and
end of the trail use up only 1 of their edges, and their total number of edges
ends up odd. Notice that this proves the second sentence in the statement:
the first and last vertex of the trail must be the two odd-degree vertices.

For the “if” direction, suppose that G is connected and has exactly two
vertices of odd degree, say v and w. If v and w are not adjacent, then we
can add the edge {v, w} to form the graph G + {v, w}; since this increases
the degrees of v and w by 1, all vertices in G + {v, w} have even degree. By
Theorem 2.6, there is an Eulerian circuit in G + {v, w}. Since this circuit
includes the edge {v, w} at some point, we may as well suppose that it starts
v, w, · · · , v; we then obtain an Eulerian trail w, · · · , v in G by deleting {v, w}
from the circuit. If v and w are adjacent in G, the argument is similar, but
instead we add a new vertex x as well as the edges {v, x} and {w, x}. (This
is to avoid having multiple edges.) Since all the degrees in the new graph are
even, it has an Eulerian circuit. Since this circuit visits x only once, we may
as well suppose that it starts v, x, w, · · · , v; again, deleting the superfluous
edges gives an Eulerian trail in G.
Example 2.9. What is the condition on n for the graph Kn to have:

an Eulerian circuit?

an Eulerian trail?

neither?
32 Introduction to Graph Theory

2.2 Hamiltonian graphs

In many real-world applications of graph theory, the vertices of the graph

represent tasks which have to be carried out, and the edges represent the
possible transitions between them. The Eulerian circuit concept is then not
so relevant, because carrying out all the tasks is more important than making
all the possible transitions. Here is a concept which is more relevant to such
interpretations.

Definition 2.10. A connected graph G with ≥ 3 vertices is Hamiltonian if

it has a walk which visits every vertex exactly once and then returns to its
starting point.

Since such a walk never repeats a vertex except for returning to where it
began, it can’t possibly repeat an edge (here we need the assumption that
there are more than 2 vertices). So the vertices and edges involved in the
walk form a cycle. Thus we can rephrase the definition more simply: G is
Hamiltonian if and only if it contains a spanning cycle (recall that “spanning”
just means “using all the vertices of G”). In other words, G is Hamiltonian
if and only if you can delete some of the edges of G (without deleting any of
the vertices) and obtain a graph isomorphic to Cn , where n is the number of
vertices of G. More precisely, “some of the edges” could read “all but n of
the edges”, since Cn has n edges. Despite the fact that this condition is as
easy to state as the existence of an Eulerian circuit, it is significantly more
difficult to give a criterion for when it holds.

One obvious principle is that if you add more edges to a Hamiltonian graph,
it remains Hamiltonian. Heuristically speaking, the more edges G has, the
more likely it is that n of them form a cycle.
Example 2.11. For any n
≥ 3, the complete graph Kn is certainly Hamil-
tonian. In fact, since all n2 possible edges are present, you can write down
the vertices in any order you like, and walk from the first to the second to
the third and so on to the last and back to the first again. So there are n!
walks which satisfy the Hamiltonian condition. (The number of spanning
n!
cycles is 2n = (n−1)!
2
, because every spanning cycle gives rise to 2n different
Hamiltonian walks, as explained in Remark 1.32.)
CHAPTER 2. SPECIAL WALKS IN GRAPHS 33

Example 2.12. The following graph is not Hamiltonian:

b e
c

a d
Indeed, the only two cycles in this graph are the visible 3-cycles, so there
is no cycle containing all the vertices. (There is an obvious Eulerian circuit
a, b, c, d, e, c, a, but that is not a cycle because it repeats the vertex c.)
Example 2.13. The graph of vertices and edges of a cube is Hamiltonian:

contains the spanning cycle

The same holds for the other regular polyhedra: tetrahedron, octahedron,
dodecahedron, and icosahedron. The origin of the term “Hamiltonian” was
a puzzle invented by the Irish mathematician Hamilton, in which you had to
find a spanning cycle in the vertices and edges of a dodecahedron.
Example 2.14*. Generalizing the ordinary 3-dimensional cube, there is an
m-dimensional cube for any positive integer m, which sits in Rm . Its vertices
are the m-tuples (x1 , x2 , · · · , xm ) where every xi is either 0 or 1 (hence there
are 2m vertices). Two such m-tuples are adjacent if and only if they differ in
exactly one coordinate. Forgetting about the m-dimensional object itself, the
vertices and edges form a graph called the cube graph Qm . These cube graphs
have a natural recursive structure: the vertices where the last coordinate is
0 form a subgraph isomorphic to Qm−1 , as do the vertices where the last
coordinate is 1. So we can construct Qm by taking two copies of Qm−1 and
joining corresponding vertices. When m = 3, this is the standard way to draw
a perspective picture of a cube: draw two squares and join the corresponding
vertices. Here is a picture of Q4 constructed in the same way:

We can prove by induction that Qm is Hamiltonian for all m ≥ 2. The base

case is clear, because Q2 (the graph of vertices and edges of a square) is
34 Introduction to Graph Theory

isomorphic to C4 . Assume that m ≥ 3 and that Qm−1 is Hamiltonian. Then

in Qm , there is a cycle containing all the vertices with last coordinate 0, and
a corresponding cycle containing all the vertices with last coordinate 1. We
then build a spanning cycle by removing an edge from the first cycle and the
corresponding edge from the second cycle, and joining the cycles together
along the edges between the corresponding affected vertices. For example, in
the above picture of Q4 we can find the following spanning cycle.

Example 2.15. The complete bipartite graph Kp,q , for positive integers p
and q, is the graph with vertex set {1, 2, · · · , p + q} where any vertex m ≤ p
is adjacent to any vertex m′ > q, but no two vertices ≤ p are adjacent, nor
are any two vertices > p. Thus the vertices fall into two parts (hence the
name “bipartite”), {1, 2, · · · , p} and {p + 1, p + 2, · · · , p + q}, where there are
no edges between vertices in the same part but every possible edge between
vertices in different parts. For instance, here is a picture of K5,3 .
1 2 3 4 5

6 7 8
In every cycle in Kp,q , the vertices must alternate between one part and the
other. So it is impossible for there to be a spanning cycle unless there are the
same number of vertices in both parts, i.e. p = q. This shows that when p 6= q,
Kp,q and all its spanning subgraphs are not Hamiltonian. Up to isomorphism,
a spanning subgraph of Kp,q is any graph where the vertices can be divided
into two parts of sizes p and q in such a way that no two vertices in the same
part are adjacent; such graphs are called bipartite. So any bipartite graph
where the two parts have different numbers of vertices is not Hamiltonian.
(This doesn’t mean, by the way, that a bipartite graph where the two parts
have the same number of vertices is Hamiltonian; for p ≥ 2, the complete
bipartite graph Kp,p is obviously Hamiltonian, but subgraphs of it need not
be.)
CHAPTER 2. SPECIAL WALKS IN GRAPHS 35

Here is a necessary condition for a graph to be Hamiltonian.

Theorem 2.16. Let G be a Hamiltonian graph.

(1) For any vertex v of G, the graph G − v is connected.

(2) More generally, for any nonempty subset S of the set of vertices of G,
the graph G − S obtained by deleting all the vertices in S and their
edges has at most |S| connected components.

Proof. Since G is Hamiltonian, it contains a spanning cycle C. For any

vertex v of G, v is also a vertex of C, and the graph C − v is a path and
hence connected. But then G − v, which has all the edges of C − v and more,
is also connected; this proves part (1).

The proof of part (2) is similar: instead of the fact that C − v is connected,
we use the fact that C − S has at most |S| connected components, which is
clear from a picture of a cycle; this implies that G − S must have at most
|S| connected components also.

Unfortunately, part (2) is not a sufficient condition for a graph to be Hamil-

tonian. So this result can be used to prove that a graph is not Hamiltonian
(if it fails to satisfy (2)), but not to prove that a graph is Hamiltonian.
Example 2.17. Let G be the following graph.
b w
a x
c z
d y
e v

A nonempty set S of vertices such that G − S has more than |S| connected

components is . Hence G is not Hamiltonian.

Example 2.18*. The Petersen graph of Example 1.42 does satisfy the con-
dition that if you delete any m vertices, the result has at most m connected
components. This is clear from the way the vertices are divided into two
36 Introduction to Graph Theory

linked 5-cycles: if you delete only vertices in the outer 5-cycle or only ver-
tices in the inner 5-cycle the graph stays connected; and if you delete m1 ≥ 1
outer vertices and m2 ≥ 1 inner vertices, then the outer cycle breaks into
at most m1 components and the inner cycle breaks into at most m2 compo-
nents, so the graph as a whole has at most m1 + m2 components. But the
Petersen graph is not Hamiltonian, as may be seen as follows. Suppose there
is a cycle C in the Petersen graph using all 10 vertices, and hence having
10 edges. At each vertex, C must use two of the three edges: in particular,
at each outer vertex, C must use at least one of the two edges in the outer
cycle. So C must contain either 4 of the edges of the outer cycle, or 3 of the
edges (two meeting at a vertex and the other disjoint from these). The choice
of which edges in the outer cycle belong to C determines which outer-inner
edges belong to C, and hence which edges in the inner cycle belong to C; it
is easy to check that both cases result in contradictions.

In the other direction, here is a sufficient, but not necessary, condition for a
graph to be Hamiltonian.
Theorem 2.19 (Ore’s Theorem). Let G be a connected graph with n vertices
where n ≥ 3. If every non-adjacent pair of vertices {v, w} satisfies deg(v) +
deg(w) ≥ n, then G is Hamiltonian.

Proof*. The proof is by contradiction, so assume that G is not Hamiltonian.

Then G certainly can’t be a complete graph, so there is some non-edge e
which we could add to form the graph G + e. If this too is not Hamiltonian,
there must be some other non-edge we can add, and so forth; there must
be some point at which adding another edge changes the graph from non-
Hamiltonian to Hamiltonian, because the complete graph is Hamiltonian.
Since adding edges can only increase the degrees, any graph obtained from
G by adding edges still has the property that every non-adjacent pair of
vertices {v, w} satisfies deg(v) + deg(w) ≥ n. So we can rename our graphs if
necessary so that G itself has the property that when you add some non-edge
it becomes Hamiltonian.

Our situation now is that G is not Hamiltonian, but for some non-adjacent
pair of vertices {v, w}, G + {v, w} is Hamiltonian. Let C be a spanning
cycle of G + {v, w}; since C is not contained in G, it must use the edge
{v, w}. So C − {v, w} is a spanning path in G with end-vertices v and w.
CHAPTER 2. SPECIAL WALKS IN GRAPHS 37

Let v1 = v, v2 , v3 , · · · , vn = w be the vertices of G in the order in which they

occur along this path; thus vi is adjacent to vi+1 for i = 1, 2, · · · , n − 1. We
also know that v1 is not adjacent to vn , and that deg(v1 ) + deg(vn ) ≥ n. We
want to use this information to show that there is a spanning cycle in G,
which contradicts the fact that G is not Hamiltonian.

The trick is to consider the sets

A = {i | 2 ≤ i ≤ n − 2, v1 is adjacent to vi+1 },
B = {i | 2 ≤ i ≤ n − 2, vn is adjacent to vi }.

Since all the vertices adjacent to v1 except v2 are of the form vi+1 for some
i ∈ A, |A| = deg(v1 ) − 1. Similarly all the vertices adjacent to vn except
vn−1 are of the form vi for some i ∈ B, so |B| = deg(vn ) − 1. Moreover,
A ∪ B ⊆ {2, 3, · · · , n − 2}, so |A ∪ B| ≤ n − 3. Hence

|A ∩ B| = deg(v1 ) + deg(vn ) − 2 − |A ∪ B| ≥ n − 2 − (n − 3) = 1.

That is, there is some i such that i ∈ A and i ∈ B, i.e. 2 ≤ i ≤ n − 2, v1 is

adjacent to vi+1 , and vn is adjacent to vi . But then the walk

v1 , vi+1 , vi+2 , · · · , vn , vi , vi−1 , · · · , v1

traverses a spanning cycle in G, as required.

Example 2.20. The condition in Ore’s Theorem is certainly satisfied if
δ(G) ≥ n2 (that is, every vertex has degree at least n2 ). The statement that
graphs with this property are Hamiltonian is known as Dirac’s Theorem (it
was proved earlier than the more general result of Ore).
Example 2.21. To see that the condition in Ore’s Theorem is not necessary
for a graph to be Hamiltonian, consider C5 : it is clearly Hamiltonian, but
the sum of the degrees of any two non-adjacent vertices is 4.

2.3 Minimal walks in weighted graphs

So far we have been treating all the edges of a graph on the same footing, but
in many applications you need to weight the edges to take account of their
38 Introduction to Graph Theory

cost, or distance, or whatever factor is important in the problem at hand.

Then you would want to find special walks in the graph which had as small
a total weight as possible.
Definition 2.22. A weighted graph is a triple (V, E, f ) where (V, E) is a
graph and f is a function from E to the positive real numbers. For every
e ∈ E, f (e) is called the weight of e. Then the weight of a subgraph H of G is
the sum of the weights of the edges of H; the weight of a walk v0 , v1 , · · · , vℓ in
G is the sum of the weights of the edges {vi , vi+1 }. For any vertices v, w ∈ V ,
the distance d(v, w) from v to w is defined to be the minimum weight of a
walk in G from v to w (or ∞ if v and w are not linked). A walk which
achieves this minimum weight is said to be a minimal walk from v to w.

Note that d(v, v) = 0 (the length-0 walk from v to itself has weight 0). It is
clear that any minimal walk from v to w must be a walk along a path, since
visiting a vertex twice would increase the weight needlessly. So d(v, w) could
also be defined as the minimal weight of a path with end-vertices v and w.
Remark 2.23. An ordinary un-weighted graph can be viewed as a weighted
graph by saying that all edges have weight 1. Then the weight of a walk is
just its length, and the distance d(v, w) is just the smallest number of edges
you have to traverse to get from v to w. The term “distance” should arguably
be restricted to this un-weighted context; its use in the weighted context is
influenced by those applications where the weights of the edges are literally
distances (between cities, and so forth).

We will represent weighted graphs by labelling each edge with its weight
(with no attempt to make the visual length of the edge proportional to it).
Example 2.24. Consider the following weighted graph.
B 2 C
1 5
A 4 1 Z
8 4
D
To find d(A, D), we need to consider walks from A to D (and we need not
consider walks which repeat a vertex). The walk along the edge {A, D} itself
CHAPTER 2. SPECIAL WALKS IN GRAPHS 39

has weight 8, which is bettered by the walk A, B, D with weight 1 + 4 = 5

and the walk A, B, C, D with weight 1 + 2 + 1 = 4. It is easy to see that
no other walk from A to D has weight ≤ 4, so d(A, D) = 4. Similarly,
d(A, C) = 1 + 2 = 3, with the walk A, B, C being the unique minimal walk
from A to C. We can use this information to deduce d(A, Z) as follows: every
walk from A to Z has to have either C or D as its second-last vertex, since
those are the only vertices adjacent to Z. Thus

the minimum weight of a walk A, · · · , C, Z is

the minimum weight of a walk A, · · · , D, Z is

the distance d(A, Z) is

the number of minimal walks from A to Z is

Remark 2.25. Notice that in Example 2.24, since the edge {A, D} does
not give a minimal walk from A to D, it can never occur in any minimal
walk: it is always preferable to get from A to D via B and C. So for the
purposes of minimal walks, the edge {A, D} may as well be deleted. From
the opposite point of view, you could imagine that we are always dealing
with complete graphs, where some of the edges have such large weights that
they are guaranteed never to be used.

When dealing with a large graph (presumably stored in computer memory

rather than pictured), it is impractical to search all possible walks from one
vertex to another in order to find a minimal walk: if the graph is complete
with n vertices, there are more than n! possible walks from one vertex to
another, and n! grows super-exponentially as n increases. A better approach
is to do what we did in Example 2.24 to find minimal walks from A to Z:
first find minimal walks from A to vertices ‘closer’ to A than Z is. Here is
an algorithm which uses this idea to find d(A, v) for every vertex v.

DIJKSTRA’S ALGORITHM. Given a weighted graph (V, E, f ) and a vertex

A, the algorithm assigns temporary labels to every vertex, which are pos-
sibly changed in the course of the algorithm until they are declared to be
permanent. The steps are:
40 Introduction to Graph Theory

(1) Give A the permanent label 0.

(2) If every v ∈ V has a permanent label, stop.

(3) Let v be the vertex which has most recently acquired a permanent
label, and let ℓ be that label.

(4) For every vertex w which is adjacent to v, do the following:

(a) if w has a permanent label, leave it unchanged;

(b) if w has no label, give it the temporary label ℓ + f ({v, w});
(c) if w has a temporary label k, compare k with ℓ + f ({v, w}), and
change the label to ℓ + f ({v, w}) if that is smaller than k.

(5) Of all the vertices which have temporary labels, choose one whose label
is smallest, and make that the permanent label of that vertex.

(6) Return to Step (2).

At the end of this algorithm, the permanent label of every vertex v is guar-
anteed to be d(A, v). We will not give a rigorous proof, but here is the idea.
At any point in the algorithm, the label of a vertex w is the minimum weight
of all walks from A to w which you have ‘considered so far’ (only implic-
itly considered, because the point of having the algorithm is that you don’t
actually need to consider all these walks). The label becomes permanent
at a point where you are certain that there are no other walks with smaller
weights. In Step (4), you are ‘considering’ walking from A to w by first doing
a minimal walk from A to v and then using the edge {v, w}. This walk has
weight ℓ + f ({v, w}), so you have to decide whether that is smaller than the
previously known minimum k; if it is, it becomes the new minimum.
Example 2.26. Let us apply the algorithm to the graph in Example 2.24.
(Of course, there is no need for the algorithm in such a small example, but
it will illustrate the general procedure.) We will put temporary labels in
parentheses and permanent labels in square brackets. The first time we
come to Step (4), v is A itself which has the label [0]; neither of the adjacent
CHAPTER 2. SPECIAL WALKS IN GRAPHS 41

vertices has a label yet, so they both get temporary labels under case (b).
(1)
B 2 C
1 5
[0] A 4 1 Z
8 4
D
(8)
Step (5) then makes the label of B permanent, since it is the smallest tempo-
rary label. Thus the next time we come to Step (4), v is B. Of the vertices
adjacent to B, vertex A is left alone under case (a); vertex C now gets a
temporary label under case (b); and vertex D gets a new temporary label
under case (c), because 1 + 4 is smaller than 8.
[1] (3)
B 2 C
1 5
[0] A 4 1 Z
8 4
D
(5)
Step (5) then makes the label of C permanent, so it is the new v; the next
pass through Step (4) gives Z a label, and changes the label of D again.
[1] [3]
B 2 C
1 5
[0] A 4 1 Z (8)
8 4
D
(4)
Step (5) then makes the label of D permanent; the next pass through Step (4)
leaves the label of Z unchanged, and the final Step (5) makes that permanent
also. The labels now display the distances from A to each vertex.
42 Introduction to Graph Theory

A good feature of Dijkstra’s algorithm is that it not only tells you the min-
imum weight of a walk from A to Z, it allows you to find all walks which
attain this minimum. The reason is that every minimal walk from A to Z
must be of the form A, · · · , Y, Z where A, · · · , Y is a minimal walk from A to
Y . Moreover, the only possible vertices Y which can occur in this way are the
vertices which are adjacent to Z and satisfy d(A, Y ) + f ({Y, Z}) = d(A, Z).
We can then use the same principle to find the minimal walks from A to Y ;
there is no chance of getting into a loop in this recursion, because d(A, Y ) is
less than d(A, Z).
Example 2.27. Consider the following weighted graph, to which Dijkstra’s
algorithm has already been applied.
[1] [4]
B 5 C
1 3
[0] A 3 1 Z [7]
3 4
D
[3]
To find all the minimal walks from A to Z, we note that the second-last vertex
Y could be either C or D, because both of these satisfy d(A, Y )+f ({Y, Z}) =
d(A, Z). In the minimal walks from A to C, the second-last vertex can only
be D, because
d(A, B) + f ({B, C}) = 1 + 4 > 4, d(A, Z) + f ({Z, C}) = 7 + 3 = 10 > 4.
Similarly, in the minimal walks from A to D the second-last vertex can only
be A itself. Thus
the unique minimal walk from A to D is A, D;
the unique minimal walk from A to C is A, D, C;
so the two minimal walks from A to Z are A, D, Z and A, D, C, Z.

Note that this method finds the walks ‘backwards’. It is tempting to try to
find minimal walks ‘forwards’ by starting at A and following the smallest-
weight edges; but Example 2.27 shows that this wouldn’t always work. (To
make it workable we would need to know the distances d(B, Z) for all vertices
B, i.e. we would need to have run Dijkstra’s algorithm for Z instead.)
CHAPTER 2. SPECIAL WALKS IN GRAPHS 43

Remark 2.28. We will not go into the details of running times for these
algorithms, but it is easy enough to see that if n is the number of vertices,
both Dijkstra’s algorithm and the above method of finding minimal walks
have running times which grow no faster than n2 . The reason is that they
both essentially consist of a loop which is executed no more than n times,
and in each pass through the loop, any vertex is considered at most once.

Here are two much-studied problems which combine the minimal-weight

property with the Eulerian and Hamiltonian properties.

Chinese Postman Problem. Given a connected weighted graph, find a

walk which uses every edge, returns to its starting point, and has minimum
weight subject to these two conditions.

Travelling Salesman Problem. Given a connected weighted graph, find a

walk which visits every vertex, returns to its starting point, and has minimum
weight subject to these two conditions.

The names come from the imagined situations of a postman doing the rounds
of various streets, and a salesman visiting various potential clients. (The
postman is Chinese in honour of the graph theorist Mei-Ko Kuan.) As with
the Eulerian and Hamiltonian problems, the starting point can be arbitrary.

Notice that in the Chinese Postman Problem, the walk has to use every edge
of the graph, but is allowed to use an edge more than once. It is obvious that
the weight of such a walk is at least the sum of the weights of all the edges.
So if the graph is Eulerian, the solutions of the Chinese Postman Problem
are exactly the Eulerian circuits (which we already have a way of finding). If
the graph is not Eulerian, i.e. it has some vertices of odd degree, then there
will inevitably be some back-tracking in the walk. The idea of the general
solution is to find the distances between the odd-degree vertices by Dijkstra’s
algorithm, and use that information to decide what is the most economical
back-tracking to do. Here is a precise result in the case of graphs which have
an Eulerian trail (see Theorem 2.8).

Theorem 2.29. Let G be a connected graph with exactly two vertices of

odd degree, say v and w. Then an Eulerian trail from v to w, followed by a
minimal walk from w to v, is a solution of the Chinese Postman Problem.
44 Introduction to Graph Theory

Proof**.
P Let E be the set of edges of G; then the proposed walk has weight
e∈E f (e) + d(v, w). We need to prove that any walk in G which returns
to its starting point and uses every edge has weight at least as large as this.
The proof is easier if we allow ourselves to talk about multigraphs. Imagine a
duplicate copy of the set of vertices in G, on which we construct a multigraph
G′ by following the walk in G and drawing an edge in G′ as the corresponding
edge in G is used – thus edges in G which are used more than once correspond
to multiple edges in G′ . We also give each edge in G′ the same weight as the
corresponding edge in G. Our walk in G then corresponds to an Eulerian
circuit in the multigraph G′ , and its weight is the weight of G′ . For the
same reason as for ordinary graphs, the degrees of an Eulerian multigraph
are all even, so every degree in G′ is even (we continue to define the degree
of u as the number of edges ending at u). Now we form a new (possibly
not connected) multigraph G′′ by deleting from G′ one copy of every edge
′′
of G. The weightP of G expresses the amount by which the weight ′′of our
walk exceeds e∈E f (e), so we need to prove that the weight of G is at
least d(v, w). It clearly suffices to prove that there is a walk from v to w in
G′′ . Since degG′′ (u) = degG′ (u) − degG (u) for every vertex u, we know that
degG′′ (v) and degG′′ (w) are odd while degG′′ (u) is even for every other vertex
u. So we construct our new walk in G′′ as follows. Starting from v, we use
any edge from v to an adjacent vertex v1 in G′′ (there must be such an edge
since degG′′ (v) is odd); if the vertex v1 is not w, then we can imagine deleting
the edge we used, which makes degG′′ (v) even and degG′′ (v1 ) odd, and then
we can continue the walk with any edge from v1 to an adjacent vertex v2 ,
and so on; we must eventually reach w, because we can’t go on deleting edges
indefinitely.
Example 2.30. Return to the graph from Example 2.27, in which B and C
are the odd-degree vertices.

An Eulerian trail from B to C is

A minimal walk from C to B is

A Chinese Postman solution is

Its weight is
CHAPTER 2. SPECIAL WALKS IN GRAPHS 45

Remark 2.31*. If a graph G with n vertices satisfied the condition of The-

orem 2.29, how long would it take to actually find a solution of the Chinese
Postman Problem, for large n? The time needed to determine which vertices
have odd degree, if that is included,

grows no faster than n2 , since the maxi-
n
mum number of edges is 2 . As noted in Remark 2.28, a minimal walk from
w to v can be found in a time which also grows no faster than n2 . The time
needed to find an Eulerian trail from v to w is probably the most significant
part: if you do it by decomposing the graph G + {v, w} into cycles, then the
time needed conceivably has cubic growth, since the time taken to find each
cycle could have the growth rate of n1 , and the number of cycles could have
the growth rate of n2 . So the growth rate of the total time is at worst cubic.

The Travelling Salesman Problem is considerably harder, but we can make

a few worthwhile comments. Firstly, we can reduce to a problem about
complete graphs, as follows. If G is the original connected weighted graph,
construct a complete graph K with the same vertex set as G, where every
edge {v, w} has weight d(v, w) (the distance from v to w in G, which then
equals the distance from v to w in K). Think of every edge of K as rep-
resenting a minimal walk between those vertices in G, which we know how
to find. Then a solution to the Travelling Salesman Problem in K can be
translated into a solution to the Travelling Salesman Problem in G, with the
same weight.
Example 2.32*. If G is the weighted graph of Example 2.24, then K is:
B 2 C
1 5
3 8 7
A Z
3 1
4 4
D
A walk in K that starts and finishes at A and visits every other vertex must
have weight at least 16, because the part of the walk from A to Z must
have weight at least 8, and the part of the walk from Z back to A must
have weight at least 8 also. Therefore the walk A, B, D, Z, C, A, which does
have weight 16, is a solution of the Travelling Salesman Problem for K. To
translate this into a solution of the Travelling Salesman Problem for G, we
replace every edge in K with a minimal walk between those vertices in G, to
obtain A, B, C, D, Z, C, B, A.
46 Introduction to Graph Theory

In the complete graph K there is no reason to visit a vertex more than once,
so the Travelling Salesman Problem in K amounts to finding a minimum-
weight spanning cycle. The brute-force approach would be to consider every
spanning cycle. Unfortunately, as seen in Example 2.11, the number of span-
ning cycles in Kn is (n−1)!
2
, which grows super-exponentially as the number of
vertices n increases. We are thus a long way from the situation of the Chinese
Postman Problem, where there is a solution with cubic growth (see Remark
2.31). In fact, it is widely believed that there is no algorithm for solving the
Travelling Salesman Problem whose running time has fixed-power growth: if
one was discovered, it would revolutionize mathematics and computing. We
will say a little more about this problem in the next chapter.
Chapter 3

Trees

One of the earliest uses of graph theory outside mathematics was determining
the possible molecular structures of various hydrocarbons; one of the most
recent has been determining the possible evolutionary relationships between
species based on their genetic closeness. In both cases, graphs without cycles
are particularly important.

3.1 Trees and Cayley’s Formula

Definition 3.1. A tree is a connected graph containing no cycles. A forest

is a general graph containing no cycles (so every connected component of a
forest is a tree). A leaf of a tree or forest is a vertex of degree 1.

Example 3.2. The graph with the following picture is a tree:

1 2
3 4

5 6

Its leaves are the vertices 1, 2, and 6.

47
48 Introduction to Graph Theory

Example 3.3. Of the graphs we have seen previously, the path graph Pn is
a tree for all n. The complete bipartite graph K1,q is also a tree for all q:
1 1 1

···
2 2 3 2 3 4
K1,1 = P2 K1,2 K1,3
The cycle Cn and the complete graph Kn for n ≥ 3 are obviously not trees.
Example 3.4. The list in Example 1.14 shows that any tree with 4 vertices
either is a path (hence is isomorphic to P4 ) or has a vertex which is adjacent
to all the others (hence is isomorphic to K1,3 ).

One reason that trees are so nice to work with is that if you remove a leaf
from a tree (and its single edge), what you have left is still a tree; this means
that trees are ideally suited for proofs by induction on the number of vertices.
Before we can use this, however, we need to ensure that leaves always exist.

Theorem 3.5. Every tree T with ≥ 2 vertices has at least two leaves.

Proof. Among all the paths in the graph T , there must be one which is of
maximal length; so if the vertices of the path are labelled v1 , v2 , · · · , vm in
order, with v1 and vm being the end-vertices, there is no path with more than
m vertices. Since T is not just a single vertex, it is clear that the maximal
m is at least 2. Now we claim that v1 is a leaf. Since we know that v1 is
adjacent to v2 , this amounts to saying that v1 is not adjacent to any other
vertex of T . But v1 can’t be adjacent to any vk with k ≥ 3, because then
we would have a cycle with vertices v1 , v2 , · · · , vk , and T contains no cycles.
Also v1 can’t be adjacent to any vertex v which is not one of the vi ’s, because
then we would have a path with vertices v, v1 , v2 , · · · , vm , contradicting the
maximality of m. So v1 is a leaf, as claimed; the same argument shows that
vm is a leaf, so T has at least two leaves.

We can now prove that trees and forests are exactly the graphs which attain
the minimum numbers of edges found in Theorem 1.37.
CHAPTER 3. TREES 49

Theorem 3.6. Let G be a graph with n vertices, k edges and s connected

components. Then G is a forest if and only if k = n − s. In particular, if G
is connected, then G is a tree if and only if k = n − 1.

Proof. We first prove that if G is a tree, it has n − 1 edges. As foreshadowed

above, we use induction on n. The n = 1 case is when G has a single vertex;
it does indeed then have 0 edges. Assume that n ≥ 2 and that the result is
known for trees with fewer vertices. By Theorem 3.5, G has a leaf v. The
graph G − v obtained by deleting v and its sole edge is a tree with n − 1
vertices, which has n − 2 edges by the induction hypothesis. So G itself has
n − 1 edges, and the induction is complete.

We can deduce the “only if” direction of the general result. If G is a forest,
then its connected components G1 , G2 , · · · , Gs are all trees. So if Gi has
ni vertices, it must have ni − 1 edges. Thus the number of edges in G is
(n1 − 1) + · · · + (ns − 1) = (n1 + · · · + ns ) − s = n − s.

Now we prove the “if” direction. Suppose that k = n − s, and let {v, w} be
any edge of G (if G has no edges, it is clearly a forest). If we delete this edge,
the graph G − {v, w} has n − s − 1 edges. According to Theorem 1.37, this is
less than the minimum number of edges a graph with s connected components
can have, so G − {v, w} must have more connected components than G. The
only way this is possible is if v and w are not linked in G − {v, w}, i.e. {v, w}
is a bridge. So we have shown that every edge of G is a bridge. We saw in
Theorem 1.34 that this is equivalent to saying that no edge of G is contained
in a cycle, so G contains no cycles and is a forest.

Theorem 3.6 means that classifying trees with n vertices is the same as
classifying connected graphs with n vertices and n − 1 edges.
Example 3.7. One of the major problems in 19th century chemistry was to
find the structure of various molecules, knowing only the molecular formula,
i.e. how many atoms of various elements the molecule contained. Molecules
were thought of as connected graphs, where the atoms were the vertices and
the bonds between atoms were the edges. Each element had a known valency,
which in graph-theoretic terms meant that the degree of all atoms of that
element was the same. For instance, in hydrocarbons every vertex is either
a carbon atom (C) of degree 4 or a hydrogen atom (H) of degree 1. Suppose
50 Introduction to Graph Theory

we know that a molecule has formula Ck H2k+2 , for some positive integer k;
this means that the graph has k vertices of degree 4 and 2k + 2 vertices of
degree 1. We can then work out the number of edges by the Hand-shaking
Lemma: it is 4k+2k+2
2
= 3k + 1. Since this is exactly 1 less than the number
of vertices, Theorem 3.6 implies that the graph is a tree. (This rules out
the possibility of cycles of carbon atoms, which give rise to special chemical
properties.) The hydrogen atoms are the leaves of the tree, so if you delete
them all and consider just the carbon atoms you still have a tree. In fact,
if you know the tree formed by the carbon atoms, you can reconstruct the
whole molecule by joining hydrogen atoms where necessary to make the C
degrees equal 4. For instance, there are only two possible structures for a
molecule with formula C4 H10 :
H

H C H

H H H H H H

H C C C C H H C C C H

H H H H H H H
Both stuctures do exist: the molecules are called butane and isobutane.

Motivated by such applications, the British mathematician Arthur Cayley

started a systematic enumeration of trees. He found that for large n, it was
easier to count all the trees with a particular vertex set, such as {1, 2, · · · , n},
rather than to count the isomorphism classes of trees with n vertices (in other
words, to count labelled rather than unlabelled trees).


Example 3.8. We have seen that there are 63 = 20 graphs with vertex
set {1, 2, 3, 4} which have 3 edges. Of these, the only ones which are not
trees are the ones with two connected components, an isolated vertex and a
3-cycle; there are 4 of these. So there are 16 trees with vertex set {1, 2, 3, 4}
(of which, incidentally, 12 are isomorphic to P4 and 4 are isomorphic to K1,3 ).

The general result, striking in its simplicity, is:

Theorem 3.9 (Cayley’s Formula). For any n ≥ 2, the number of trees with
vertex set {1, 2, · · · , n} is nn−2 .
CHAPTER 3. TREES 51

We will prove Cayley’s Formula by the Bijection Principle, finding a bi-

jection between the set of trees with vertex set {1, 2, · · · , n} and a set of
sequences which has size nn−2 . The sequence we attach to a tree is called its
Prüfer sequence, which is defined by the following recursive algorithm.

PRÜFER SEQUENCE. The input is a tree T with n ≥ 2 vertices, where

the vertex set is a subset of the positive integers. The output is a sequence
(p1 , p2 , · · · , pn−2 ) with n − 2 terms, all in the vertex set. The steps are:

(1) If n = 2, return the empty sequence () and stop.

(2) (To arrive here, we must have n ≥ 3.) Let ℓ be the smallest (in nu-
merical order) of all the leaves of T . Define the first term p1 to be the
unique vertex of T which is adjacent to ℓ.

(3) Recursively call PRÜFER SEQUENCE to find the Prüfer sequence of

the smaller tree T − ℓ, and let (p2 , · · · , pn−2 ) be this sequence.

In words, the algorithm progressively deletes the smallest leaf of whatever

is remaining, and records in the sequence not the leaf itself but the vertex
which was adjacent to it.
Example 3.10. Let T be the tree in Example 3.2. The smallest leaf is 1
which is adjacent to 5. So the first term of the Prüfer sequence is 5, and the
rest is obtained by considering the tree T − 1:

2
3 4

5 6

The smallest leaf in T − 1 is 2 which is adjacent to 4, so the next term of

the Prüfer sequence is 4, and the rest is obtained by considering (T − 1) − 2.
Continuing in this way, you can find the full Prüfer sequence of T (which
should have 6 − 2 = 4 terms):
52 Introduction to Graph Theory

Here is a useful property of the Prüfer sequence:

Theorem 3.11. If T is a tree with n ≥ 2 vertices, then each vertex v occurs
deg(v) − 1 times in the Prüfer sequence of T . In particular, the leaves of T
are exactly the vertices which don’t occur in the sequence.

Proof. Since the definition of Prüfer sequence is recursive, the obvious

method of proof is induction on n. The n = 2 base case is clear: both vertices
are leaves, and the sequence is accordingly empty. Assume that n ≥ 3 and
that the result is known for smaller trees. Let ℓ be the smallest leaf of T . The
Prüfer sequence of T is (p1 , p2 , · · · , pn−2 ), where p1 is the vertex adjacent to
ℓ and (p2 , · · · , pn−2 ) is the Prüfer sequence of T − ℓ. The claim is certainly
correct when v = ℓ: ℓ does not occur in the Prüfer sequence. The claim is
also correct when v = p1 : we have degT −ℓ (p1 ) = degT (p1 ) − 1, so the induc-
tion hypothesis tells us that p1 occurs degT (p1 ) − 2 times in (p2 , · · · , pn−2 ),
and hence it occurs degT (p1 ) − 1 times in (p1 , · · · , pn−2 ). If v is any other
vertex of T , then its degree is the same in T − ℓ as in T . So the induction
hypothesis says that v occurs deg(v) − 1 times in (p2 , · · · , pn−2 ) and hence
also in (p1 , · · · , pn−2 ). The induction step is complete.

The crucial point is that we can reconstruct a tree uniquely from its Prüfer
sequence. The easiest way to see this is to write down a recursive algorithm
going in the opposite direction to the above algorithm.

REVERSE PRÜFER. The input is a subset V of the positive integers with

n ≥ 2 elements, and a sequence (p1 , p2 , · · · , pn−2 ) whose terms belong to V .
The output is a tree T with vertex set V . The steps are:

(1) If n = 2, output the tree with vertex set V and a single edge joining
the two vertices.
(2) (To arrive here, we must have n ≥ 3.) Let ℓ be the smallest element of
V which does not occur in the sequence.
(3) Recursively call REVERSE PRÜFER with the smaller set V \ {ℓ} and
the shorter sequence (p2 , · · · , pn−2 ); let T ′ be the resulting tree.
(4) Form a tree T with vertex set V by adding to T ′ the new vertex ℓ and
a single new edge joining ℓ to p1 .
CHAPTER 3. TREES 53

Example 3.12. Suppose that V = {1, 2, 3, 4, 5} and the sequence is (4, 4, 2).
The algorithm tells us that our tree T is obtained by attaching the leaf 1 to
the vertex 4 of the tree T ′ , where T ′ is the output of the algorithm applied
to the set {2, 3, 4, 5} and the sequence (4, 2). In turn, T ′ is obtained by
attaching the leaf 3 to the vertex 4 of the tree T ′′ , where T ′′ is the output of
the algorithm applied to the set {2, 4, 5} and the sequence (2). Finally, T ′′
is obtained by attaching the leaf 4 to the vertex 2 of the unique tree with
vertex set {2, 5}. So T is:

3
2 4

5 1

We can now prove Cayley’s Formula.

Proof. Let X be the set of all trees with vertex set {1, 2, · · · , n}, and let
Y be the set of all sequences (p1 , p2 , · · · , pn−2 ) where each pi belongs to
{1, 2, · · · , n}. The PRÜFER SEQUENCE algorithm gives us a function f :
X → Y , taking each tree to its Prüfer sequence. The REVERSE PRÜFER
algorithm gives us a function g : Y → X, taking a sequence (and the set
{1, 2, · · · , n}) and constructing a tree. From the definitions of the algorithms,
it is clear that these functions are inverse to each other, i.e. if you apply
REVERSE PRÜFER to the output of PRÜFER SEQUENCE you do recover
the original tree, and if you apply PRÜFER SEQUENCE to the output of
REVERSE PRÜFER you do recover the original sequence. (You could prove
this formally by induction.) So X and Y are in bijection, and |X| = |Y |. But
clearly |Y | = nn−2 , because there are n choices for each of the n − 2 terms.
Cayley’s Formula follows.

3.2 Spanning trees

If the edges of a connected graph represent lines of communication in some

sense, it may be important to consider the ways in which you could maintain
as few of them as possible and still have all the vertices linked to each other.
54 Introduction to Graph Theory

As we have seen, a graph with the smallest possible number of edges subject
to being connected is a tree. So for a general connected graph G, we consider
the spanning trees of G, i.e. the subgraphs which use all the vertices of G,
are themselves connected, and contain no cycles.
Theorem 3.13. Every connected graph G contains a spanning tree.

Proof. If every edge of G is a bridge, then G is itself a tree. Otherwise, we

can delete some edge belonging to a cycle, and the graph remains connected.
If we keep repeating this procedure, we will eventually reach a tree (at the
point where the number of edges remaining is 1 less than the number of
vertices); this is a spanning tree of the original graph.

Note that you can’t find a spanning tree by simply deleting all edges of G
which belong to cycles in G: after all, it is perfectly possible that every edge
belongs to a cycle. What the above proof does is different: it deletes an edge
of a cycle, and then re-considers the cycles in that smaller graph (some of
the cycles in G will have been removed by the deletion), and so on. At any
stage, the choices for which edge you can delete depend on which edges you
have deleted so far, so there can be many different spanning trees.
Example 3.14. Suppose that G is the graph
b w
a x
c z
d y
e v
Two of the many spanning trees of G are:
b w b w
a x a x
c z c z
d y d y
e v e v

Most of the important algorithms for finding spanning trees take the reverse
approach: they start with a single vertex and add edges. As long as every
added edge joins a vertex already in the tree with a new vertex, it is obvious
that no cycles are created.
CHAPTER 3. TREES 55

BREADTH-FIRST SEARCH (BFS). The input is a connected graph G with

n vertices, and a chosen vertex v1 . The output is an ordering v1 , v2 , · · · , vn
of the vertices of G, and a spanning tree T of G. The steps are:

(1) Let T be the graph with the single vertex v1 .

(2) If vn has been defined, then stop.
(3) Let m be the largest number for which vm has been defined. Let ℓ ≤ m
be the smallest number such that there are vertices adjacent to vℓ which
are not in T . If there are s such vertices, name them vm+1 , · · · , vm+s
in an arbitrary order.
(4) Add to T the vertices vm+1 , · · · , vm+s , and the edges joining these ver-
tices to vℓ .
(5) Return to Step (2).

The reason for the name “search” is the idea that this algorithm searches for
the vertices of G: first finding all the vertices adjacent to v1 , then finding all
the unfound vertices adjacent to those, and so on until all vertices have been
found. Each time Step (4) is executed, the new vertices are added to T along
with the edges which record ‘through which earlier vertex they were found’.
Note that because of the choice in ordering the new vertices, the resulting
spanning tree is not unique.
Example 3.15. Let us apply this algorithm to the graph G in Example
3.14, with v1 = a. The first pass through Step (3) defines v2 = c and v3 = b,
say. The order of these two vertices is arbitrarily chosen, but it affects what
happens next, because on the next pass through Step (3) it is the vertices
adjacent to v2 which get added, say v4 = d, v5 = e, v6 = z. The earliest vi
which is adjacent to an unused vertex is now v5 , so v7 = v, v8 = w. Then v6
is responsible for naming the last two vertices v9 = x and v10 = y. Thus the
order of the vertices is a, c, b, d, e, z, v, w, x, y, and the spanning tree is:
b w
a x
c z
d y
e v
56 Introduction to Graph Theory

One application of breadth-first search is to find the shortest possible path

from v1 to another vertex. (Since we are currently dealing with unweighted
graphs, “shortest” simply means “with the smallest number of edges”.) First
note the following general result:
Theorem 3.16. If v and w are vertices of a tree T , there is a unique path
in T between v and w.

Proof. We already know by Theorem 1.34 that there is a path between v and
w, so the content of this result is the uniqueness. The idea is simple: if there
were two different paths, you could patch appropriate pieces of them together
to form a cycle, contradicting the fact that the graph is a tree. However, the
notation involved in writing this out is a bit messy, so we will disguise the
idea by giving an induction proof. Let n be the number of vertices; if n = 1
then v and w are the same and there is nothing to prove. If n ≥ 2 and the
result is known for smaller trees, then let ℓ be any leaf of T . If v = w = ℓ
there is obviously a unique path between v and w, namely the path with no
edges and single vertex ℓ. If neither v nor w equals ℓ, then we know by the
induction hypothesis that there is a unique path in T − ℓ with end-vertices v
and w. But clearly no path with end-vertices v and w can include ℓ as one of
its other vertices, because deg(ℓ) = 1; so there is a unique path in T between
v and w. The remaining cases are v = ℓ, w 6= ℓ and v 6= ℓ, w = ℓ; these
are symmetric, so we can assume v = ℓ, w 6= ℓ. If u is the unique vertex
adjacent to v, then any path whose first vertex is v must have u as its second
vertex. By the induction hypothesis, there is a unique path in T − ℓ with
end-vertices u and w, so there is a unique path in T between v and w. The
induction step is complete.

In particular, in the spanning tree produced by the BFS algorithm there is

a unique path between v1 and v for any vertex v; this path records ‘how the
algorithm found v’. Note that in the progress of the algorithm, these paths
from v1 in T are built evenly, in the sense that no paths are ever present
while there are shorter paths not yet present. This allows us to prove:
Theorem 3.17. Let T be a spanning tree of the graph G obtained by ap-
plying the BFS algorithm with the initial vertex v1 . Then for any vertex v,
the unique path in T between v1 and v has the shortest possible length of
any path in G between these vertices.
CHAPTER 3. TREES 57

Proof*. Let dT (v1 , v) denote the length (i.e. number of edges) of the unique
path between v1 and v in T , and let dG (v1 , v) denote the length of the shortest
possible path between v1 and v in G. It is obvious that dG (v1 , v) ≤ dT (v1 , v)
for all v. We suppose for a contradiction that dG (v1 , v) < dT (v1 , v) for some
v; we can choose v so that dG (v1 , v) is minimal, subject to this inequality
holding. Since dT (v1 , v) ≥ 1, v is not v1 . Let w be the vertex ‘through which’
v was added to the tree T ; then w is adjacent to v on the path in T between
v1 and v, so dT (v1 , w) = dT (v1 , v) − 1. Let u be the vertex adjacent to v on
some shortest-possible path in G between v1 and v. Then the part of the path
from v1 to u must also be as short as possible, so dG (v1 , u) = dG (v1 , v) − 1.
By our minimality assumption, we have

dT (v1 , u) = dG (v1 , u) = dG (v1 , v) − 1 < dT (v1 , v) − 1 = dT (v1 , w).

This implies that in the progress of the BFS algorithm, u was added to the
tree T before w; but then it is impossible for the vertex v to have been
added through w, because it would have already been added through u if not
through some other vertex.

Breadth-first search is so called by comparison with depth-first search, where

you follow one path as long as you can before retracing your steps to take a
new one (in a somewhat similar way to the line of succession to the throne
in a European monarchy).

DEPTH-FIRST SEARCH (DFS). The input is a connected graph G with n

vertices, and a chosen vertex v1 . The output is an ordering v1 , v2 , · · · , vn of
the vertices of G, and a spanning tree T of G. The steps are:

(1) Let T be the graph with the single vertex v1 .

(2) If vn has been defined, then stop.

(3) Let m be the largest number for which vm has been defined. Let ℓ ≤ m
be the largest number such that there are vertices adjacent to vℓ which
are not in T . Choose one of these vertices and call it vm+1 .

(4) Add to T the vertex vm+1 , and the edge joining this vertex to vℓ .

(5) Return to Step (2).

58 Introduction to Graph Theory

Example 3.18. If you carry out a depth-first search on the graph G in

Example 3.14 with v1 = a, and whenever you have a choice of vertices choose
the one which is latest in the alphabet, then the order of the vertices is

and the spanning tree T is

One useful feature of a DFS spanning tree is that it contains some information
about all the edges of the graph.
Theorem 3.19. Let T be a spanning tree of the graph G obtained by ap-
plying the DFS algorithm, and let v1 , v2 , · · · , vn be the resulting ordering of
the vertices. If vi is adjacent to vj in G, with i < j, then vi lies on the path
between v1 and vj in T .

Proof*. Let vi1 , vi2 , · · · , vis be the successive vertices of the path from v1 to
vi , with i1 = 1 and is = i; from the construction it is clear that i1 < i2 <
· · · < is . Similarly, let vj1 , vj2 , · · · , vjt be the successive vertices of the path
from v1 to vj , with j1 = 1, jt = j, and j1 < j2 < · · · < jt . Let k be maximal
so that ik = jk ; the vertex vik is the ‘latest common ancestor’ of vi and vj ,
if you think of T as a family tree of descendants of v1 . We want to prove
that, given the adjacency of vi and vj in G, this latest common ancestor is vi
itself, i.e. k = s. Suppose for a contradiction that k < s. It is impossible for
k to be t, because that would imply that j < i, so the two paths diverge and
continue through the different vertices vik+1 and vjk+1 . From the fact that
the descendant vi of vik+1 was added to the tree before the descendant vj of
vjk+1 , we deduce that ik+1 ≤ i < jk+1 . But then consider the pass through
Step (3) of the algorithm which gave vjk+1 its name. At this point of the
algorithm, the value of m must have been jk+1 − 1, so vi was already in T ;
CHAPTER 3. TREES 59

and the value of ℓ must have been ik , because vik was the vertex to which the
new vertex vjk+1 became attached in T . But that contradicts the maximality
in the choice of ℓ, because i > ik and vi was also adjacent to a vertex not yet
in T , namely vj .

In the context of weighted graphs, one would prefer a minimal spanning tree,
i.e. one whose weight (the sum of the weights of the edges) is as small as
possible. The naive approach is to build the tree by starting with a single
vertex and adding one new edge at a time, where the edge has the smallest
weight of all those you can possibly add to the tree.
Example 3.20. Consider the weighted graph
B 5 C
1 5
A 4 6 Z
3 4
D
If we start with vertex A and want to add another vertex to our tree in a way
which minimizes the total weight of the edges, we should obviously choose
B rather than D. Having done that, the possible edges we can add in the
next step are {A, D}, {B, C}, and {B, D}; since {A, D} has the smallest
weight, we choose that. We cannot now add the edge {B, D}, because that
would form a cycle: to ensure that we still have a tree, we have to add a new
vertex along with each new edge. The possible edges are {B, C}, {C, D},
and {D, Z}; the last of these has the smallest weight, so that is the one we
add. Finally, we choose between the three edges ending at C: since {B, C}
and {C, Z} have the same weight, we arbitrarily choose {B, C} to reach a
spanning tree of weight 13.
B 5 C
1
A Z
3 4
D
It is easy to see that there is no spanning tree of weight less than 13, so we
have found a minimal spanning tree for our weighted graph.
60 Introduction to Graph Theory

Somewhat remarkably, this naive method always works, and is worth describ-
ing formally.

PRIM’S ALGORITHM. The input is a connected weighted graph G with n

vertices, and a chosen vertex v1 . The output is an ordering v1 , v2 , · · · , vn of
the vertices of G, and a spanning tree T of G. The steps are:

(1) Let T be the graph with the single vertex v1 .

(2) If vn has been defined, then stop.

(3) Let m be the largest number for which vm has been defined. Consider
all the edges {vi , w} in G where i ≤ m and w is not in T ; choose one
of minimal weight, and let vm+1 be w.

(4) Add to T the vertex vm+1 , and the edge joining this vertex to vi .

(5) Return to Step (2).

Theorem 3.21. A spanning tree produced by Prim’s Algorithm is minimal.

Proof**. Let T be a spanning tree of the weighted graph G produced by

Prim’s Algorithm, and let v1 , v2 , · · · , vn be the vertices in the order in which
they were added. Thus the edges of T can be listed as {va2 , v2 }, · · · , {van , vn }
for some a2 , · · · , an ∈ {1, · · · , n} such that aj < j for j = 2, · · · , n.

Let S be a minimal spanning tree of G. If S = T we are finished, so we can

assume there is some j such that the edge {vaj , vj } of T does not belong to
S; in fact, we can define j to be minimal with this property. Since S is a tree,
there is a unique path between vaj and vj in S: let its successive vertices be
vi1 , vi2 , · · · , vis where i1 = aj and is = j. We know that s ≥ 3, because the
edge {vaj , vj } is not in S. Now let t ≤ s be minimal such that it ≥ j; then
t ≥ 2 and it−1 < j. Since the edge {vit−1 , vit } was passed over for inclusion in
T at the stage of the algorithm where {vaj , vj } was added, its weight cannot
be less than that of {vaj , vj }. So if we modify the spanning tree S by adding
the edge {vaj , vj } (temporarily forming a cycle with the existing path) and
then deleting the edge {vit−1 , vit } to return to n − 1 edges, we obtain another
minimal spanning tree S ′ .
CHAPTER 3. TREES 61

If S ′ = T we are finished. Otherwise, since the edge we deleted from S was

not any of {va2 , v2 }, · · · , {vaj−1 , vj−1 }, the quantity j ′ defined for S ′ in the
same way as j was for S is strictly larger. So if we repeat this procedure,
modifying S ′ to obtain a new minimal spanning tree S ′′ and so on, we must
eventually obtain T .

With Prim’s Algorithm to help us find minimal spanning trees, we can say a
little more about the Travelling Salesman Problem. Remember that we had
reduced the problem to finding minimal spanning cycles in a Hamiltonian
weighted graph (in fact, we saw that we could assume the graph is complete,
but it doesn’t help much).

Theorem 3.22. Let H be a Hamiltonian weighted graph, and v a vertex of

H. Then the weight of a minimal spanning cycle of H is at least equal to:

the weight of a minimal spanning tree of H − v

+ the sum of the two smallest weights of edges of H at v

Proof. Let C be any spanning cycle in H. Then the weight of C equals the
weight of C − v plus the sum of the weights of the two edges of C at v. Since
C − v is a path and contains all the vertices of H − v, it is a spanning tree
of H − v, so its weight is at least equal to the weight of a minimal spanning
tree of H − v. Trivially, the sum of the weights of the two edges of C at v
is at least equal to the sum of the two smallest weights of edges of H at v.
The result follows.

Theorem 3.22 gives us a lower bound for the weight of a minimal spanning
cycle, or rather one lower bound for each vertex, which can be computed
efficiently. However, there is no guarantee that equality holds.
Example 3.23. If K is the complete weighted graph of Example 2.32, then
the weight of a minimal spanning tree in K − A is 7, and the two smallest
weights of edges at A are 1 and 4, so the lower bound provided by Theorem
3.22 is 12. Similarly, the lower bounds obtained by considering the vertices
B, C, D, and Z are 11, 11, 12, and 13 respectively; all well short of the
actual weight of a minimal spanning cycle of K, which we saw is 16.
62 Introduction to Graph Theory

3.3 Kirchhoff ’s Matrix–Tree Theorem

In the previous section we saw several different ways to construct a spanning

tree of a connected graph. This raises the interesting combinatorial question
of how many spanning trees a given connected graph has.
Example 3.24. Let G be the graph
2 5
3

1 4
Since G has five vertices, any spanning tree must have four edges, so it must
be obtained from G by deleting two edges. Since a spanning tree cannot
contain cycles, we need to delete one edge from each of the two 3-cycles.
Clearly we can choose any edge from the left-hand 3-cycle and any edge
from the right-hand 3-cycle, so G has 3 × 3 = 9 spanning trees. (These fall
into three isomorphism classes, but we are not concerned with isomorphism
classes at the moment.)
Example 3.25. Let G be the graph
1 2

3 4


To find a spanning tree, we need to delete two edges; of the 52 = 10 possible
pairs of edges, the only cases which would leave a cycle remaining are if
we deleted {1, 2} and {2, 4}, or {1, 3} and {3, 4}. So there are 10 − 2 = 8
spanning trees.
Example 3.26. If G is a tree, then of course the unique spanning tree of G
is G itself.
Example 3.27. If G = Cn , then we obtain a spanning tree (specifically, a
path) by deleting any edge, so there are n spanning trees.
Example 3.28. Another general class for which we have already answered
the question is that of the complete graphs Kn , for n ≥ 2. A spanning tree
of Kn is just the same thing as a tree with vertex set {1, 2, · · · , n}, and by
Cayley’s Formula there are nn−2 of these.
CHAPTER 3. TREES 63

The physicist Kirchhoff (who was interested in graphs through his study of
electrical circuits) found a general formula for the number of spanning trees,
which reveals an unexpected connection with the determinants of matrices.
Definition 3.29. Let G be a graph with vertex set {1, 2, · · · , n}. The
Laplacian matrix of G is the n × n matrix M whose (i, j)-entry is defined by


deg(i) if i = j,
mij = −1 if i =6 j and {i, j} is an edge of G,


0 if i 6= j and {i, j} is not an edge of G.

That is, the diagonal entries give the degrees of the vertices, and the off-
diagonal entries are −1 in positions corresponding to an edge and 0 elsewhere.
Example 3.30. The Laplacian matrix of the graph in Example 3.24 is
 
2 −1 −1 0 0
−1 2 −1 0 0
 
−1 −1 4 −1 −1 .
 
0 0 −1 2 −1
0 0 −1 −1 2

Here are some properties of the Laplacian matrix.

Theorem 3.31. Let M be the Laplacian matrix of a graph G as above.

(1) M is symmetric (its (j, i)-entry equals its (i, j)-entry for all i, j).

(2) The sum of the entries in any row or column of M is zero.

(3) The determinant det(M ) is zero.

(4) For any k, ℓ ∈ {1, 2, · · · , n}, let M k̂,ℓ̂ denote the (n−1)×(n−1) matrix
obtained from M by deleting the kth row and the ℓth column. Then
(−1)k+ℓ det(M k̂,ℓ̂ ) is the same for all values of k, ℓ.

Proof*. Part (1) is obvious from the definition. Because M is symmetric, its
rows are the same as its columns, so we only need to prove part (2) for a row,
64 Introduction to Graph Theory

say the ith row. By definition, the nonzero entries in the ith row are a single
entry of deg(i), and an entry of −1 for every vertex adjacent to i; since there
are deg(i) adjacent vertices, these entries cancel out and give zero, proving
part (2). Now suppose we multiply the matrix M by the n × 1 column vector
v whose entries are all 1. By the definition of matrix multiplication, M v
is another column vector whose ith entry is the sum of the entries in the
ith row of M ; we have just seen that this sum is always zero, so M v is the
zero vector. In other words, v belongs to the null space of the matrix M . A
standard linear algebra result says that a square matrix has a nontrivial null
space if and only if its determinant is zero, so part (3) follows.

To prove part (4), it is enough to show that

′ ˆ′
(−1)k+ℓ det(M k̂,ℓ̂ ) = (−1)k +ℓ det(M k ,ℓ̂ ) for all k, k ′ , ℓ, (3.1)
since then the symmetry of M will imply that we also have
′ ˆ′
(−1)k+ℓ det(M k̂,ℓ̂ ) = (−1)k+ℓ det(M k̂,ℓ ) for all k, ℓ, ℓ′ , (3.2)
and the result clearly follows from (3.1) and (3.2). Moreover, it suffices to
prove the case of (3.1) where k ′ = k + 1. In that case, the only difference
ˆ′
between the matrices M k̂,ℓ̂ and M k ,ℓ̂ is in their kth rows: the kth row of
ˆ′
M k ,ℓ̂ is the original kth row of M without its ℓth entry, which is the very
row that was deleted from M to form M k̂,ℓ̂ . Since the sum of the entries
in each column of M is zero, the sum of all the rows of M is the zero row
vector, so the kth row of M is the negative of the sum of all the other rows.
ˆ′
Thus M k ,ℓ̂ is exactly what you get when you apply to the matrix M k̂,ℓ̂ the
row operation Rk ← −R1 − R2 − · · · − Rn−1 . This row operation changes
ˆ′
the sign of the determinant, so det(M k ,ℓ̂ ) = − det(M k̂,ℓ̂ ), which gives the
required case of (3.1).

The quantity (−1)k+ℓ det(M k̂,ℓ̂ ) is called the (k, ℓ)-cofactor of the matrix M .
Remark 3.32. These cofactors are what you multiply the matrix entries by
when calculating the determinant by expanding along a row or column. For
instance, the rule for calculating det(M ) by expanding along the first row is
det(M ) = m11 det(M 1̂,1̂ ) − m12 det(M 1̂,2̂ ) + · · · + m1n (−1)1+n det(M 1̂,n̂ ).
In our case there is no need to use this rule, since part (3) of Theorem 3.31
tells us that det(M ) is zero.
CHAPTER 3. TREES 65

Theorem 3.33 (Kirchhoff’s Matrix–Tree Theorem). Let M be the Laplacian

matrix of a graph G with vertex set {1, 2, · · · , n}, where n ≥ 2. Then for all
k, ℓ ∈ {1, 2, · · · , n}, the (k, ℓ)-cofactor (−1)k+ℓ det(M k̂,ℓ̂ ) equals the number
of spanning trees of G. (In particular, if G is not connected, all the cofactors
of M are zero.)

Before we give the proof, here are some examples of how to use the Matrix–
Tree Theorem to count spanning trees.
Example 3.34. Let us apply the Matrix–Tree Theorem to find the number
of spanning trees in the graph G of Example 3.24, whose Laplacian matrix
M was found in Example 3.30. Since all the cofactors of M are guaranteed
to be the same, we may as well choose to delete the row and column with
the most nonzero entries; the (3, 3)-cofactor is
 
2 −1 0 0
−1 2 0 0
(−1)3+3 det(M 3̂,3̂ ) = det 
0
.
0 2 −1
0 0 −1 2

When a matrix has block-diagonal form like this, its determinant is the prod-
uct of the determinants of the blocks. Since
 
2 −1
det = 2 × 2 − (−1) × (−1) = 3,
−1 2

the (3, 3)-cofactor of M is 3 × 3 = 9. This of course agrees with the number

of spanning trees found in Example 3.24.
Example 3.35. If G is as in Example 3.25, the Laplacian matrix M is
 
 
 .
 

Its (1, 1)-cofactor is

 

det  = .
66 Introduction to Graph Theory

Example 3.36*. We can use the Matrix–Tree Theorem to give another

proof of Cayley’s Formula for the number of spanning trees of Kn . The
Laplacian matrix M of Kn is the n × n matrix where all the diagonal entries
are n − 1 and all the off-diagonal entries are −1. The (1, 1)-cofactor is the
determinant of an (n − 1) × (n − 1) matrix with the same rule for the entries:
 
n − 1 −1 −1 · · · −1
 −1 n − 1 −1 · · · −1 
 
 −1 −1 n − 1 · · · −1 
 .
 .. .. .. ... .. 
 . . . . 
−1 −1 −1 · · · n − 1

To put this matrix in a form where the determinant is easier to compute, we

apply the row operation R1 ← R1 + R2 + · · · + Rn−1 (which does not change
the determinant):
 
1 1 1 ··· 1
−1 n − 1 −1 · · · −1 
 
−1 −1 n − 1 · · · −1 
 ,
 .. .. .. . . .. 
 . . . . . 
−1 −1 −1 · · · n − 1

and then the row operations Ri ← R1 + Ri for i = 2, · · · , n − 1:

 
1 1 1 ··· 1
0 n 0 · · · 0
 
0 0 n · · · 0
 .
 .. .. .. . . .. 
. . . . .
0 0 0 ··· n

For an upper-triangular matrix like this, the determinant is the product of

the diagonal entries, which are a 1 and n − 2 copies of n. So we recover the
answer nn−2 , re-proving Cayley’s Formula.

The proof of the Matrix–Tree Theorem uses a fact about determinants which
you may not have seen. Recall that determinants are only defined for square
matrices, and that they are multiplicative in the sense that

det(AB) = det(A) det(B), (3.3)

CHAPTER 3. TREES 67

whenever A and B are square matrices of the same size. There is a general-
ization of this multiplicativity property for possibly non-square matrices:
Theorem 3.37 (Cauchy–Binet Formula)*. Let A be an n × m matrix and
B an m × n matrix. Then
X
det(AB) = det(A−,J ) det(B J,− ),
J⊆{1,2,··· ,m}
|J|=n

where A−,J is the n × n matrix obtained from A by deleting all the columns
except those whose number lies in J, and B J,− is the n × n matrix obtained
from B by deleting all the rows except those whose number lies in J. (The
− in the superscripts is to indicate that there is no deletion of any rows in
forming A−,J or columns in forming B J,− .)

We will omit the proof. Note that if m = n, the only choice for J is all of
{1, · · · , n}, and for this J we have A−,J = A and B
J,− = B, which brings the
formula in line with (3.3). In general, there are m n
terms on the right-hand
side of the Cauchy–Binet Formula (in particular, if n > m the statement is
that det(AB) = 0).

The reason we are interested in non-square matrices is the following.

Definition 3.38. Let G be a graph with vertex set {1, 2, · · · , n} and edges
numbered e1 , e2 , · · · , ek in some fixed order (we assume that G 6= Nn ). The
incidence matrix F of G is the n × k matrix whose (i, j)-entry is defined by


1 if ej = {i, i′ } where i < i′ ,
fij = −1 if ej = {i, i′ } where i > i′ ,


0 if i is not an end of ej .
So the jth column contains a 1 in the row corresponding to the smaller end
of ej , a −1 in the row corresponding to the larger end, and zeroes elsewhere.
Example 3.39. If G is the following graph with edges numbered as shown:
2 e4 4
e1 e7
1 e3 e5 5
e2 e6
3
68 Introduction to Graph Theory

then the incidence matrix F is the following 5 × 7 matrix:

 
1 1 0 0 0 0 0
−1 0 1 1 0 0 0
 
 0 −1 −1 0 1 1 0
 .
0 0 0 −1 −1 0 1
0 0 0 0 0 −1 −1

Theorem 3.40*. Let F be the incidence matrix of a graph G as above.

(1) The sum of the entries in any column of F is zero.

(2) Let I be a subset of {1, 2, · · · , n} such that |I| = k. Let F I,− be the
k × k matrix obtained from F by deleting all the rows except those
whose number lies in I. If I contains all the vertices of some connected
component of G, then det(F I,− ) = 0.

(3) Suppose that G is a tree; hence k = n − 1. For any i ∈ {1, · · · , n},

let F î,− be the k × k matrix obtained from F by deleting the ith row.
Then det(F î,− ) = ±1.

Proof**. Part (1) is obvious, because the entries of each column of F are
a 1, a −1, and n − 2 zeroes. For part (2), let I ′ be a subset of I consisting
of all the vertices of some connected component of G. For convenience, we
continue to number the rows of the matrix F I,− by their row numbers from
F (that is, the elements of I) rather than changing their numbers to 1, · · · , k.
Let v be a row vector whose columns are indexed by the elements of I, in
which the ith entry is 1 if i ∈ I ′ and 0 otherwise. Then vF I,− is a 1 × k
row vector whose jth entry is the sum, as i runs over I ′ , of the ith entry
of the jth column of F I,− . If the edge ej does not belong to the connected
component with vertex set I ′ , all these entries are zero; if the edge ej does
belong to this connected component, the entries consist of a 1, a −1, and the
rest zeroes. In either case the sum is zero, so vF I,− is the zero row vector.
Thus the left null space of F I,− is nontrivial, proving that det(F I,− ) = 0 as
required.

We prove part (3) by induction on k. The k = 1 base case is clear, because

1
then the incidence matrix F is ( −1 ). Assume that k ≥ 2, and that the result
CHAPTER 3. TREES 69

is known for trees with fewer edges. Now by the same argument as in the
proof of Theorem 3.31, the fact that every column of F sums to zero implies
that all the determinants det(F î,− ) are the same up to sign, so it is enough to
show that just one of them equals ±1. We can therefore assume that vertex i
is a leaf of the tree G. Let i′ be the unique vertex which is adjacent to i, and
suppose that the edge {i, i′ } is ej . Then the jth column of F î,− has a single
nonzero entry, namely a ±1 in row i′ . By expanding along this column, we
conclude that
c′
det(F î,− ) = ± det(F i,i ,ĵ ),
where the meaning of the superscripts on the right-hand side is hopefully
clear. But F î,ĵ is the incidence matrix of the tree G − i, so by the induction
hypothesis, the right-hand side is ±1 as required.

The connection between the Laplacian and incidence matrices is:

Theorem 3.41. Let G be a graph with vertex set {1, 2, · · · , n} and edges
numbered e1 , e2 , · · · , ek in some fixed order. Let M be the Laplacian matrix
of G, let F be the incidence matrix of G, and let F t be the transpose of F .
Then M = F F t .

Proof*. By definition
P of matrix multiplication and transpose, the (i, i′ )-
entry of F F t is kj=1 fij fi′ j . Since fij is zero unless i is an end of ej , the only
nonzero terms in this sum are those where i and i′ are ends of ej . If i 6= i′ and
{i, i′ } is not an edge, there are no such nonzero terms, so the entry is zero. If
i 6= i′ and {i, i′ } is an edge, then the unique nonzero term is 1 × (−1) = −1,
so the entry is −1. If i = i′ , there is a nonzero term of fij2 = 1 for every j
such that i is an end of ej , so the entry is deg(i). Thus in every case the
(i, i′ )-entry of F F t agrees with the (i, i′ )-entry of M .

We can now prove the Matrix–Tree Theorem.

Proof**. Number the edges of G as e1 , · · · , ek and form the incidence matrix

F . Since all cofactors of M are equal by part (4) of Theorem 3.31, it is enough
to show that the (n, n)-cofactor det(M n̂,n̂ ) equals the number of spanning
trees of G. It follows from Theorem 3.41 that M n̂,n̂ = (F n̂,− )(F n̂,− )t , where
F n̂,− is the (n − 1) × k matrix obtained from F by deleting the nth row. Note
that since G is connected, k ≥ n − 1. Applying the Cauchy–Binet Formula,
70 Introduction to Graph Theory

we obtain
det(M n̂,n̂ ) = det((F n̂,− )(F n̂,− )t )
X
= det(F n̂,J ) det((F n̂,J )t )
J⊆{1,2,··· ,k}
|J|=n−1
X
= det(F n̂,J )2 ,
J⊆{1,2,··· ,k}
|J|=n−1

where the meaning of F n̂,J should by now be clear, and the last step uses the
fact that the transpose of a square matrix has the same determinant. Now
F −,J is the incidence matrix of a spanning subgraph GJ of G, namely the
one with edges ej for j ∈ J. Since GJ has n vertices and n − 1 edges, it is
either a tree or has more than one connected component. If GJ has more
than one connected component, then there is a connected component which
does not include the vertex n, and part (2) of Theorem 3.40 implies that
det(F n̂,J ) = 0. If GJ is a tree, then part (3) of Theorem 3.40 implies that
det(F n̂,J ) = ±1. So the nonzero terms in the above sum are all 1, and the
number of them is the number of spanning trees of G, as required.
Chapter 4

Colourings of Graphs

As mentioned in the introduction, one application of graph theory is to

scheduling problems. The common feature of such problems is that there
are a number of tasks which all have to be carried out in as short a time as
possible, with the constraint that some of the tasks cannot be done simul-
taneously: computer processes which need to access the same database, or
exams which have a candidate in common, or flights which need to land on
the same runway, or whatever. These problems can be translated into the
abstract framework of colourings of graphs.

4.1 Vertex colourings and chromatic number

Definition 4.1. Let G be a graph with vertex set V . A vertex colouring of

G is a function c : V → C, where C is a finite set whose elements are called
colours, with the property that no two adjacent vertices are assigned the
same colour: in other words, c(v) 6= c(w) for every edge {v, w} of G. We say
that G is t-colourable if there exists a vertex colouring of G where |C| = t.
The minimum number of colours required in such a vertex colouring, i.e. the
smallest t such that G is t-colourable, is called the chromatic number of G
and written χ(G).

71
72 Introduction to Graph Theory

In applications to scheduling, the vertices represent the different tasks, the

colours represent the available times for them, and the edges record which
pairs of tasks are not allowed to be assigned the same time. From a math-
ematical point of view, the colours could always just be numbers, but it is
convenient to stick with the metaphor and use white, black, red, blue, etc.
We will sometimes represent vertex colourings of a graph by labelling the
vertices with abbreviations for these colour words, in addition to the labels
which give the actual names of the vertices – although the latter may be
omitted, since we often care only about the isomorphism class of the graph.
Example 4.2. Let G be the graph in Example 3.25. It is obvious that G
is 4-colourable, because if we have four colours we can give every vertex a
different colour, and then the condition that no two adjacent vertices have
the same colour is trivially satisfied. Almost as obvious is that G is not 2-
colourable, because the vertices 1, 2, and 4 are all adjacent to each other, so
there would be a contradiction if we had only 2 colours. Suppose we want
a vertex colouring with 3 colours: red, white, and blue. By the previous
remark, 1, 2, and 4 have to get one colour each; the remaining constraint is
that 3 cannot have the same colour as 1 or 4, so it must get the same colour
as 2. Thus a possible vertex colouring of G with three colours is:
R 1 2 W

W 3 4 B
We have shown that the chromatic number χ(G) is 3. Since the colours
red, white, and blue could be shuffled arbitrarily, there are actually 3! = 6
different vertex colourings with these colours.

Example 4.3. The chromatic number of the Petersen graph is .

CHAPTER 4. COLOURINGS OF GRAPHS 73

Here are some fairly obvious principles concerning colourability.

Theorem 4.4. Let G be a graph with n vertices.

(1) G is 1-colourable if and only if it has no edges. So if G has edges,

χ(G) ≥ 2.

(2) If G is a complete graph, then χ(G) = n. If G is not a complete graph,

then χ(G) ≤ n − 1.

(3) If H is a subgraph of G and G is t-colourable, then H is t-colourable.

Hence χ(H) ≤ χ(G).

(4) If G has connected components G1 , G2 , · · · , Gs , then G is t-colourable

if and only if every Gi is t-colourable. Hence χ(G) = max{χ(Gi )}.

Proof. If G has a vertex colouring with 1 colour, then no two vertices can
be adjacent; conversely, if no two vertices are adjacent we can obviously
colour all vertices the same colour, which proves part (1). Now it is clear
that if we have n colours we can give every vertex a different colour. (In
scheduling terms, this corresponds to the easy but potentially uneconomic
option of having no simultaneous processes.) If G is a complete graph, i.e.
any two vertices are adjacent, then we are in fact forced to give every vertex
a different colour, so we cannot have a vertex colouring with fewer than n
colours. If G is not a complete graph, there must be two vertices v and w
which are not adjacent, so we can construct a vertex colouring with n − 1
colours by giving v and w the same colour and using the other n − 2 colours
for the other n − 2 vertices, one each. This proves part (2).

For part (3), any vertex colouring of G clearly restricts to give a vertex
colouring of the subgraph H, which implies the statement. The reason that
we have an inequality χ(H) ≤ χ(G) is that there could be vertex colourings
of H which cannot be extended to a vertex colouring of G: for instance,
because vertices which are not adjacent in H may be adjacent in the larger
graph G. However, in the special case where H is a connected component of
G, H and the rest of the graph are completely independent of each other; if
you have a vertex colouring of every connected component of G, then taken
together they form a vertex colouring of G. Part (4) follows.
74 Introduction to Graph Theory

As a consequence of parts (2) and (3) of Theorem 4.4, if G has a subgraph

isomorphic to the complete graph Km , then χ(G) ≥ m. In other words,
if G contains m vertices which are all adjacent to each other, then G has
no vertex colouring with fewer than m colours. (We used this principle in
Example 4.2 to show that G was not 2-colourable.) However, it is not the
case that G always contains χ(G) vertices which are all adjacent to each
other: the Petersen graph illustrates this.
Example 4.5. Consider the cycle graph Cn , for n ≥ 3. If we are to have
a vertex colouring with two colours, say white and black, then the colours
have to alternate as you move around the cycle. Clearly this is possible if n
is even. But if n is odd, you would have a contradiction when you arrived
back at the first vertex; thus you need to colour one of the vertices with a
third colour, say red, as in the following picture of C5 .
R

B W

W B
Thus χ(Cn ) = 2 if n is even, and χ(Cn ) = 3 if n is odd.

In general, a graph is 2-colourable if and only if it is bipartite in the sense

mentioned in Example 2.15: the vertices can be divided into two parts –
white vertices and black vertices, to use the colour terminology – such that
there are no edges between vertices of the same colour. Example 4.5 showed
that even cycles (that is, cycles with an even number of vertices) are bipartite
and odd cycles are not. There is a general theoretical criterion:
Theorem 4.6. A graph G is bipartite if and only if it contains no odd cycles.

Proof. If G is bipartite (i.e. 2-colourable), then every subgraph of G must

also be bipartite; since odd cycles are not bipartite, G cannot contain any
odd cycles. We must now prove the converse: if G contains no odd cycles,
then G is bipartite. Using part (4) of Theorem 4.4, we reduce to the case
that G is connected. Choose any vertex v. To colour the vertices, we divide
CHAPTER 4. COLOURINGS OF GRAPHS 75

the vertex set V into disjoint subsets V0 , V1 , V2 , · · · as follows. Firstly, V0

consists of v alone, which we colour white. Then V1 consists of those vertices
adjacent to v, which we are forced to colour black. Then V2 consists of those
vertices which are adjacent to some vertex in V1 but which are not in V0 or
V1 ; we are forced to colour these vertices white. Similarly, V3 consists of those
vertices which are adjacent to some vertex in V2 but which are not in V0 , V1 ,
or V2 ; we are forced to colour the vertices in V3 black. Continuing in this
way, we eventually reach all the vertices, because the graph is connected. In
fact, Vk consists of those vertices w such that the shortest walk from v to w
has length k. Note that we have coloured w white if w ∈ Vk for k even, and
we have coloured w black if w ∈ Vk for k odd.

The only question is whether this is a proper vertex colouring: we need to

prove that we never have w ∈ Vk being adjacent to w′ ∈ Vℓ , where k and ℓ are
both odd or both even. Suppose for a contradiction that such w, w′ do exist.
Let v = v0 , v1 , · · · , vk = w and v = v0′ , v1′ , · · · , vℓ′ = w′ be minimal-length
walks from v to w and to w′ respectively. Since vi , vi′ ∈ Vi for all i, any vertices
in common between the two walks must have the same position in both walks.
′
Let m be maximal such that vm = vm (i.e. this is the last vertex common to
′ ′ ′
both walks). Then the vertices vℓ , vℓ−1 , · · · , vm+1 , vm , vm+1 , · · · , vk , together
with the edges between successive vertices and between first and last, form
a cycle in G with (ℓ − m) + (k − m) + 1 vertices. Since k and ℓ are both odd
or both even, this number is odd, and we have our contradiction.

In particular, every tree is bipartite, because a tree contains no cycles at all.

The colouring procedure in the proof of Theorem 4.6 gives a practical way
to either find a 2-colouring or prove that none exists.
Example 4.7. The graph on the left is a tree; starting by colouring the
top-left vertex white, we construct the rest of the 2-colouring as shown. If
we try to do the same with the graph on the right, we end up wanting to
colour both vertices marked “?” white, a contradiction. So the graph on the
left is bipartite and the graph on the right is not.

W B W B W B ? ?

B W B B
W B W B
B W B W
76 Introduction to Graph Theory

If a graph is not bipartite, there is no particularly good algorithm for finding

its chromatic number or constructing a vertex colouring with that many
colours. There are various algorithms for constructing vertex colourings
where the number of colours used is not necessarily minimal; here is one.

WELSH–POWELL ALGORITHM. The input is a graph G with n vertices

and a set of colours. (We don’t know exactly how many colours will be
needed, but no more than n.) The output is a vertex colouring of G.

(1) Order the vertices v1 , v2 , · · · , vn in order of decreasing degree (for ver-

tices with the same degree, order them arbitrarily).
(2) If all the vertices have been coloured, stop.
(3) Let m1 be minimal such that vm1 has not been coloured. Let m2 > m1
be minimal such that vm2 has not been coloured and vm2 is not adjacent
to vm1 . Let m3 > m2 be minimal such that vm3 has not been coloured
and vm3 is not adjacent to vm1 or vm2 . Continue in this way to define
m1 , · · · , ms .
(4) Colour vm1 , vm2 , · · · , vms with a colour which has not yet been used.
(5) Return to Step (2).

Note that in this algorithm you colour all the red vertices (say) at once and
then never use the colour red again: to determine which vertices to colour
red, you choose the first uncoloured vertex and then whatever later vertices
you can allowably give the same colour to. The ordering by decreasing degree
is not necessary, but tends to reduce the number of colours which are needed.
Example 4.8. The vertices in the following graph have been ordered by de-
creasing degree. Use the Welsh–Powell Algorithm to give a vertex colouring.
v6 v7 v4
v9
v2 v1
v8 v10
v3 v5

Hence the chromatic number of the graph is .

CHAPTER 4. COLOURINGS OF GRAPHS 77

There is an upper bound for the chromatic number of G in terms of ∆(G), the
maximum of the vertex degrees. (We assume that the vertex set is nonempty,
so ∆(G) make sense.) We start with a weak form of the result.

Theorem 4.9. Any graph G is (∆(G) + 1)-colourable.

Proof. The proof is by induction on n, the number of vertices of G. In

the case where G is a single vertex we have ∆(G) = 0, and G is indeed
1-colourable. So we can assume that n ≥ 2 and that the result is true for
graphs with fewer than n vertices. Let v be any vertex of G. It is clear
that ∆(G − v) ≤ ∆(G), so the induction hypothesis implies that G − v is
(∆(G)+1)-colourable; choose a vertex colouring of G−v where the colour set
has size ∆(G) + 1. Now the number of vertices of G adjacent to v is deg(v),
which is at most ∆(G) by definition. So at most ∆(G) different colours are
used among the vertices adjacent to v. Therefore there is a colour which is
not used among these vertices, and we can colour v with this colour to obtain
a vertex colouring of G. The induction step is complete.

Another way to state this result is that χ(G) ≤ ∆(G) + 1. Note that equality
holds in the cases where G = Kn (since χ(Kn ) = n and ∆(Kn ) = n − 1) or
G = Cn for n ≥ 3 odd (since χ(Cn ) = 3 and ∆(Cn ) = 2). It turns out that
these are the only connected graphs for which equality holds.

Theorem 4.10 (Brooks). Let G be a connected graph which is neither

complete nor an odd cycle. Then G is ∆(G)-colourable, so χ(G) ≤ ∆(G).

Proof**. The proof is again by induction on the number of vertices of G.

The smallest case allowed by the hypotheses is when G has three vertices
but is not complete (hence G ∼ = P3 ); in this case χ(G) = ∆(G) = 2. So we
can assume that G has at least four vertices, and that the result is known for
graphs with fewer vertices. Let v be a vertex of G of minimal degree δ(G),
and let d denote ∆(G). Let G1 , · · · , Gs be the connected components of G−v
(possibly s = 1); we claim that each Gi is d-colourable. Since ∆(Gi ) ≤ d,
this is part of the induction hypothesis unless Gi is complete or an odd cycle;
but in these two cases we have ∆(Gi ) < d (because Gi is regular and some
vertex of Gi is adjacent to v), so the claim follows from Theorem 4.9. So
G − v is d-colourable; choose a vertex colouring of G − v where the colour
78 Introduction to Graph Theory

set is {1, 2, · · · , d}. We want to extend this vertex colouring to the whole of
G (possibly after some modification). If G is not regular of degree d, then
deg(v) = δ(G) < d, so there is a colour not used among the vertices adjacent
to v, and we can colour v with this colour to obtain a vertex colouring of G
as required. So henceforth we may assume that G is regular of degree d. If
d = 2 this forces G to be a cycle (an even cycle, by assumption), in which
case we know the result. So we may also assume that d ≥ 3.

We now have various cases for the vertex colouring of G − v. To save words,
we will omit to say in the description of each case that the previous cases do
not hold, but that should always be understood.

Case 1: the vertices adjacent to v do not all have distinct colours.

In this case there is a colour not used among these vertices, so as before we
can extend the vertex colouring to G.

In the remaining cases, every colour is used exactly once among the vertices
adjacent to v; we let vi denote the unique vertex adjacent to v with colour i.

Case 2: for some i, the vertices in G − v adjacent to vi do not all

have distinct colours. Since there are d − 1 such vertices, this means that
there is some colour other than i which they do not use. We can then change
the colour of vi to this other colour, putting us back in Case 1.

In the remaining cases, we let vij denote the unique vertex in G − v which is
adjacent to vi and has colour j, for all i 6= j.

Case 3: for some i 6= j, there is no walk in G − v from vi to vj

which uses only vertices coloured i and j. Let H be the subgraph
of G − v formed by the vertices with colours i and j, and whatever edges
of G join such vertices. Our assumption means that vi and vj belong to
different connected components of H. We can then swap the colours i and j
throughout the component containing vj , and still have a vertex colouring.
After this change, vj and vi both have colour i, putting us back in Case 1.

In the remaining cases, we let Hij denote the unique connected component
of the subgraph of G − v formed by vertices coloured i and j which contains
the particular vertices vi and vj .
CHAPTER 4. COLOURINGS OF GRAPHS 79

Case 4: for some i 6= j, Hij is not a path between vi and vj . Hence

as you walk in Hij from vi (without back-tracking), there is some vertex w
where you first have a choice about which vertex to walk to next. Note that
w is not vi itself, because the first step of the walk must be to vij (similarly,
w 6= vj ). Thus the degree of w in Hij is at least 3, so there are ≤ d − 3
vertices adjacent to w with colours other than i and j, and there must be
some third colour not used among these vertices. If we change w to this third
colour it has the effect of disconnecting Hij , putting us back in Case 3.

Case 5: for some distinct i, j, k, Hij and Hik have a vertex in common
other than vi . Let this vertex be u; its colour must clearly be i. Since u
has two neighbours coloured j on the path Hij and two neighbours coloured
k on the path Hik , we must have d ≥ 4; moreover, there must be a fourth
colour not used among the vertices adjacent to u. If we change u to this
fourth colour we disconnect Hij and Hik , putting us back in Case 3.

Case 6: for some i 6= j, vi is not adjacent to vj . In particular, this

implies that vij 6= vj . Let k be any third colour. We can swap the colours i
and k throughout the path Hik , and still have a vertex colouring of G − v.
We claim that this new vertex colouring falls into one of the previous cases;
assume for a contradiction that it does not. One of the changes made in the
new colouring is that the old vi is the new vk and vice versa. So the old vij
(whose colour is still j, and which is still not equal to vj ) is the new vkj . On
the other hand, the old vij was joined to vj by the path Hij − vi , and none of
the vertices on that path had their colours changed (because none of them
belonged to Hik , by virtue of us not having been in Case 5). So the new
vkj is joined to vj by a path whose vertices have colours i and j, and it thus
belongs to both the new Hij and the new Hkj , contradicting the assumption
that our new colouring did not fall into Case 5.

In the only case remaining, we know that all the vi ’s are adjacent to each
other. So the d neighbours of vi in G are exactly v and the other vj ’s; this
shows that G consists of just the vertices v, v1 , · · · , vd , and is a complete
graph. This contradicts our assumption, so the last case vanishes and the
induction step (remember that this was all inside the induction step!) is
finished.
80 Introduction to Graph Theory

If you think the proof of Brooks’ Theorem was long, you’ll be relieved that
we are only going to mention the most famous result about vertex colourings,
whose proof runs to hundreds of pages.

Theorem 4.11 (Four-Colour Theorem). Every planar graph is 4-colourable.

A graph is said to be planar if you can draw a picture of it (on an ordinary

piece of paper) in such a way that the edges don’t cross at non-vertex points.
(For instance, if you try to draw such a picture of K5 , you will find it can’t
be done; which is just as well for the sake of the Four-Colour Theorem,
because we know that K5 is not 4-colourable.) The notion of planarity is
substantially different from any property of graphs we have considered in
this course, because it brings in ideas of topology: it focuses on the spatial
properties of a picture of a graph, rather than the abstract adjacencies which
the picture represents. These topological properties of graphs are explored
further in MATH3061 Geometry and Topology.

4.2 The chromatic polynomial

In studying colourability of a graph, it is natural to ask how many different

vertex colourings a graph has with a given list of colours.

Definition 4.12. If G is a graph and t ∈ N, the chromatic polynomial PG (t)

is the number of vertex colourings of G with a fixed colour set of size t.

Thus G is t-colourable if and only if PG (t) 6= 0. The name discloses one of

the unexpected facts about PG (t), namely that it is a polynomial function of
t; we will prove this later.
Example 4.13. Let G = K3 (which is the same as C3 ). We know that K3
has no vertex colourings with fewer than 3 colours, so PK3 (0) = PK3 (1) =
PK3 (2) = 0. If our colour set has three elements – say red, white, and blue
– then to get a vertex colouring, we must assign a different colour to each
vertex; the number of ways of doing this is 3! = 6, so PK3 (3) = 6. In fact,
for a colour set with t elements, the number of vertex colourings of K3 is
CHAPTER 4. COLOURINGS OF GRAPHS 81

just the number of ways of making an ordered selection of 3 colours from t

possibilities with repetition not allowed, i.e. PK3 (t) = t(3) = t(t − 1)(t − 2).
Recall the reason for this formula: there are t choices for the colour of vertex
1, then t − 1 choices for the colour of vertex 2 (because it can’t be the same
as that of vertex 2), then t − 2 choices for the colour of vertex 3 (because it
can’t be the same as either of the two colours already used). Notice that the
formula t(t − 1)(t − 2) gives the right answer 0 for t = 0, 1, 2 also; so PK3 (t)
is indeed a polynomial in t.
Example 4.14. Let G be the path graph P3 . Given a colour set with t
colours, the number of vertex colourings of P3 is t(t − 1)(t − 1) = t(t − 1)2 ,
by similar reasoning: there are t choices for the colour of vertex 1, then t − 1
choices for the colour of vertex 2, then t − 1 choices for the colour of vertex
3 (which can be anything except that colour of vertex 2; notice that there is
no reason you can’t re-use the colour of vertex 1 for vertex 3). The fact that
0 and 1 are roots of the polynomial t(t − 1)2 , while 2 is not, is another way
of saying that χ(P3 ) = 2.

These examples can be generalized as follows.

Theorem 4.15. Let G be a graph with n ≥ 1 vertices, and t ∈ N.

(1) If G has connected components G1 , · · · , Gs , PG (t) = PG1 (t) · · · PGs (t).

(2) If G is a complete graph, PG (t) = t(t − 1)(t − 2) · · · (t − n + 1).

(3) If G is a tree, PG (t) = t(t − 1)n−1 .

Proof. Part (1) follows from the fact that choosing a vertex colouring of G
is the same as independently choosing vertex colourings of all the Gi ’s. If
G is complete, then any vertex colouring must use different colours for all
the vertices. So the number of vertex colourings is the number of ordered
selections of n colours from t possibilities with repetition not allowed, which
is given by t(n) as stated in part (2).

We prove part (3) by induction on n. If n = 1 (i.e. G consists of a single

vertex), then obviously PG (t) = t as claimed. Assume that n ≥ 2 and that
the result is known for trees with fewer vertices. Let v be a leaf of G, and w
82 Introduction to Graph Theory

the unique vertex it is adjacent to. Any vertex colouring of G is obtained by

making a vertex colouring of G − v and then choosing a colour for v which
is not the same as that of w. By the induction hypothesis, the number of
vertex colourings of G − v is t(t − 1)n−1 , and no matter what vertex colouring
of G − v is chosen, the number of ways of choosing the colour for v is t − 1.
So the number of vertex colourings of G is t(t − 1)n−2 (t − 1) = t(t − 1)n−1 ,
completing the induction step.
Remark 4.16. For obvious reasons, we are not vitally concerned with the
graph whose vertex set is empty. But according to the standard conventions,
this graph has a single vertex colouring for any colour set, so its chromatic
polynomial is 1 = t0 (and its chromatic number is 0).

Chromatic polynomials in general can sometimes be computed simply by

using the Product Principle, as in Theorem 4.15.
Example 4.17. Again let G be the graph in Example 3.25. We have a
subgraph isomorphic to K3 formed by the vertices 1, 2, and 4; the number of
vertex colourings of this subgraph is t(t − 1)(t − 2), as seen above. Given any
such vertex colouring, the number of ways to extend it to a vertex colouring
of G is t − 2, because the colour of vertex 3 can be anything except the
colours of vertices 1 and 4 (which are definitely different from each other).
So PG (t) = t(t − 1)(t − 2)2 .

The generalization is:

Theorem 4.18. Let G be a graph and t a nonnegative integer. If v is a
vertex of G such that the vertices adjacent to v are all adjacent to each
other, then PG (t) = PG−v (t)(t − deg(v)).

Proof. In any vertex colouring of G − v, the deg(v) vertices adjacent to v

must all have different colours, since by assumption they are all adjacent to
each other. So if t < deg(v), there are no vertex colourings of G − v, and
therefore none of G; so both sides are zero. If t ≥ deg(v), then for any vertex
colouring of G − v there are t − deg(v) ways to choose a colour for the vertex
v, so as to extend it to a vertex colouring of G. The result follows.

However, there need not be any vertex which satisfies the condition in The-
orem 4.18. In such cases we may need the Sum Principle as well.
CHAPTER 4. COLOURINGS OF GRAPHS 83

Example 4.19. Consider the cycle graph C4 :

1 2

4 3
If we try to use the same sort of Product Principle arguments to compute
PC4 (t), we run into a slight problem. If we first choose the colours of vertices
1, 2, and 3 (in t, t − 1, and t − 1 ways respectively), then when we come to
choose the colour of vertex 4 there could be either t − 1 or t − 2 possibilities,
depending on whether the colours we have chosen for vertices 1 and 3 are
the same or not. We need to consider each case separately:

no. of ways to colour 1, 2, 3

=
so that 1 and 3 have the same colour

no. of ways to colour 1, 2, 3

=
so that 1 and 3 have different colours

So PC4 (t) = .

The logic of Example 4.19 can be applied more generally using the following
definition.

Definition 4.20. Let G be a graph, and let v, w be distinct vertices of G.

The graph obtained from G by fusing v and w, written G[v, w], is the graph
whose vertices are the same as those of G except that v and w are replaced
by a single vertex v ∗ w. The edges between vertices other than v ∗ w are the
same as in G, and v ∗ w is adjacent to another vertex u if and only if either
{u, v} or {u, w} is an edge of G (or both are).

Example 4.21. If you start with the graph C4 , you obtain:

1∗2 2

if you fuse 1 and 2, and 1∗3 if you fuse 1 and 3.

4 3 4
84 Introduction to Graph Theory

Theorem 4.22. Let G be a graph and t a nonnegative integer.

(1) If v and w are non-adjacent vertices of G, then

PG (t) = PG+{v,w} (t) + PG[v,w] (t).

(2) If v and w are adjacent vertices of G, then

PG (t) = PG−{v,w} (t) − PG[v,w] (t).

Proof. If v and w are not adjacent, then in a vertex colouring of G there is

no constraint on whether the colours of v and w are different or not. Thus
the set of all vertex colourings of G (with a fixed colour set of size t) is the
disjoint union of the set of vertex colourings of G in which the colours of v
and w are different, and the set of vertex colourings of G in which the colours
of v and w are the same. The vertex colourings of G in which the colours of
v and w are different are obviously in bijection with the vertex colourings of
G + {v, w}, because the extra edge {v, w} imposes exactly the condition that
the colours of v and w are different. The vertex colourings of G in which the
colours of v and w are the same are obviously in bijection with the vertex
colourings of G[v, w], because if their colours have to be the same, the two
vertices may as well be fused into one. Part (1) follows.

For part (2), we apply part (1) to G − {v, w}, obtaining

PG−{v,w} (t) = PG (t) + PG[v,w] (t),

because it makes no difference to the fused graph G[v, w] whether v and w

are adjacent in G or not. Rearranging this equation gives the result.
Example 4.23. Consider the graph C5 .
1

5 2

4 3
CHAPTER 4. COLOURINGS OF GRAPHS 85

Applying part (1) of Theorem 4.22 with the vertices 2 and 5, we see that
PC5 (t) is the sum of the chromatic polynomials of the following graphs:
1 1

5 2 2∗5

4 3 4 3
In both of these graphs, Theorem 4.18 applies to vertex 1. Hence their
chromatic polynomials are t(t − 1)(t2 − 3t + 3)(t − 2) and t(t − 1)(t − 2)(t − 1)
respectively, and we thus find that

PC5 (t) = t(t − 1)(t − 2)[(t2 − 3t + 3) + (t − 1)] = t(t − 1)(t − 2)(t2 − 2t + 2).

An alternative method is to apply part (2) of Theorem 4.22 to the vertices

1 and 2. This tells us that PC5 (t) is the difference between the chromatic
polynomials of the following graphs:
1

5 2
5 1∗2

4 3 4 3
The first of these being a tree, and the second isomorphic to C4 , their chro-
matic polynomials are t(t − 1)4 and t(t − 1)(t2 − 3t + 3) respectively, so

PC5 (t) = t(t − 1)[(t3 − 3t2 + 3t − 1) − (t2 − 3t + 3)] = t(t − 1)(t3 − 4t2 + 6t − 4),

which is the same polynomial as found above.

If necessary, we can apply Theorem 4.22 recursively to the graphs appearing

on the right-hand side. Using part (2) repeatedly, we must eventually reduce
to graphs whose chromatic polynomials we know, such as trees or forests,
because the number of edges keeps decreasing. Alternatively, using part (1)
86 Introduction to Graph Theory

repeatedly, we must eventually reach complete graphs, because the number

of non-edges keeps decreasing. As a theoretical consequence of this, we can
finally justify the terminology “chromatic polynomial”.
Theorem 4.24*. Let G be a graph with n ≥ 1 vertices. There are unique
nonnegative integers a1 , · · · , an−1 , depending on G, such that for any t ∈ N,

PG (t) = tn − an−1 tn−1 + an−2 tn−2 − · · · + (−1)n−1 a1 t.

Moreover, an−1 is the number of edges of G.

Proof*. The proof is by induction on the number of edges of G. If G has

no edges, then G has n connected components, all single vertices, and its
chromatic polynomial is clearly tn ; the result holds with all ai = 0. So we
can assume that G has at least one edge, and that the result is known for
graphs with fewer edges. If {v, w} is an edge of G, then G − {v, w} and
G[v, w] both have fewer edges than G, so by the induction hypothesis,

PG−{v,w} (t) = tn − bn−1 tn−1 + bn−2 tn−2 − · · · + (−1)n−1 b1 t, and

PG[v,w] (t) = tn−1 − cn−2 tn−2 + · · · + (−1)n−2 c1 t,

for some nonnegative integers b1 , · · · , bn−1 and c1 , · · · , cn−2 (note that G[v, w]
has only n − 1 vertices). By part (2) of Theorem 4.22, we get

PG (t) = tn − (bn−1 + 1)tn−1 + (bn−2 + cn−2 )tn−2 − · · · + (−1)n−1 (b1 + c1 )t.

So the desired equation holds with ai = bi + ci for i = 1, · · · , n − 2 and

an−1 = bn−1 + 1. Moreover, the induction hypothesis includes the fact that
bn−1 is the number of edges of G−{v, w}, so an−1 is the number of edges of G.
Finally, the ai ’s must be unique, because two polynomials in the variable t
which agree for infinitely many values of t (namely, all nonnegative integers)
must have the same coefficients.
Remark 4.25. Remember that PG (t) = 0 for all nonnegative integers t <
χ(G), so the polynomial in Theorem 4.24 has χ(G) known roots. In partic-
ular, if G has any edges then 1 must be a root of the polynomial, so

1 − an−1 + an−2 − · · · + (−1)n−1 a1 = 0.

It is not known in general which sequences of nonnegative integers arise as

the sequence (a1 , · · · , an−1 ) for some graph G.
CHAPTER 4. COLOURINGS OF GRAPHS 87

4.3 Edge colourings

An important class of scheduling problems involves ‘appointments’. Imagine

a number of students who need to have one-on-one meetings with various
lecturers, where each lecturer may be required for several meetings. If we
are given the list of which lecturers each student needs to meet with, how
can we schedule the meetings so that the whole thing takes as little time
as possible? If we adopted our previous approach, we would construct a
graph which has a vertex for each meeting that has to take place (i.e. each
pair of a student and a lecturer who need to meet), and make two vertices
adjacent (thus ensuring that they are not scheduled simultaneously) if they
have either a student in common or a lecturer in common. We then need to
find the chromatic number of this graph, which may not be easy.

A better alternative is to consider the graph which is naturally suggested by

the problem: namely, where there are vertices corresponding to the students,
other vertices corresponding to the lecturers, and edges joining each student
with every lecturer whom they need to meet. Note that this graph is bipar-
tite: there are no edges joining a student to another student, or a lecturer to
another lecturer. Since it is now the edges, not the vertices, which represent
the meetings that need to be scheduled, our question amounts to finding the
edge chromatic number of the graph, in the following sense.
Definition 4.26. Let G be a graph with edge set E. An edge colouring of
G is a function c : E → C, where C is a finite set whose elements are called
colours, with the property that no two edges with the same colour have an
end in common. We say that G is t-edge-colourable if there exists an edge
colouring of G where |C| = t. The smallest t such that G is t-edge-colourable
is called the edge chromatic number of G and written χ′ (G).
Example 4.27. Suppose that students 1, 2, 3, 4, 5 need to have meetings
with lecturers a, b, c according to the edges of the following graph:
1 2 3 4 5

a b c
88 Introduction to Graph Theory

Since there are three students who need to meet lecturer b, there will have
to be at least three meeting timeslots. Can all the meetings be scheduled in
only three timeslots? Try to find an edge colouring of the graph with three
colours (say red, white, and blue). The conclusion is that the edge chromatic

number of this graph is .

Here are some obvious principles about edge colourings.

Theorem 4.28. Let G be a graph and t a nonnegative integer.

(1) If t < ∆(G), then G is not t-edge-colourable. Hence χ′ (G) ≥ ∆(G).

(2) If H is a subgraph of G and G is t-edge-colourable, then H is t-edge-

colourable. Hence χ′ (H) ≤ χ′ (G).

(3) If G has connected components G1 , · · · , Gs , then G is t-edge-colourable

if and only if each Gi is t-edge-colourable. Hence χ′ (G) = max{χ′ (Gi )}.

Proof. By definition of ∆(G), there is some vertex where ∆(G) edges all
end; these edges must have different colours in any edge colouring, which
implies part (1). Parts (2) and (3) follow from the same reasoning as in the
case of vertex colourings (see parts (3) and (4) of Theorem 4.4).

A graph G has edge chromatic number 0 if and only if it has no edges, i.e.
∆(G) = 0. If ∆(G) = 1, i.e. the graph does have edges but no two edges ever
end at the same vertex, then obviously the edges can all be given the same
colour, so the edge chromatic number is 1. In these cases, the inequality in
part (1) of Theorem 4.28 is actually equality. But there are examples where
equality fails, i.e. graphs G which are not ∆(G)-edge-colourable.
Example 4.29. Consider the cycle graph Cn for n ≥ 3. An edge colouring
of Cn requires at least ∆(Cn ) = 2 colours. In any edge colouring with 2
colours, the colours must alernate as you go around the cycle; if n is odd,
this results in a contradiction, exactly as with vertex colourings of Cn . So
the edge chromatic number is actually the same as the (vertex) chromatic
number of Cn , namely 2 if n is even and 3 if n is odd.
CHAPTER 4. COLOURINGS OF GRAPHS 89

We can find the edge chromatic numbers of complete graphs as follows.

Theorem 4.30. For any n ≥ 2,
(
n if n is odd,
χ′ (Kn ) =
n−1 if n is even.

Proof*. First consider the case of odd n ≥ 3. We have ∆(Kn ) = n − 1.

However, we can prove that Kn is not (n − 1)-edge-colourable as follows.
Suppose for a contradiction
that we had an edge colouring of Kn with n − 1
colours. Since there are n2 = n(n−1)
2
edges, the Pigeonhole Principle tells us
that there must be some colour for which there are ⌈ n2 ⌉ = n+1
2
edges with
n+1
that colour. But then the n + 1 ends of those 2 edges must all be different,
which is a contradiction since Kn has n vertices.

We can prove that Kn is n-edge-colourable by a direct construction. We

let the colour set be {1, 2, · · · , n}, the same as the vertex set. Picture the
vertices as the vertices of a regular n-gon, numbered clockwise from 1 to n.
We extend the numbering of vertices cyclically, so that vertex 0 is vertex n,
vertex −1 is vertex n−1, vertex n+1 is vertex 1, and so forth. Then the edges
which we colour with colour i are those where the ends are equidistant from i:
namely, {i−1, i+1}, {i−2, i+2}, and so on until we reach {i− n−1 2
, i+ n−1
2
},
which is the edge of the n-gon directly opposite i. These edges clearly have
no ends in common (in fact, in the picture they are all parallel). Moreover,
each edge of Kn gets a unique colour in this way, because for any two vertices
there is a unique third vertex from which they are equidistant (note that this
would not hold if n was even). So Kn is n-edge-colourable, and we have
shown that χ′ (Kn ) = n for n odd.

Now consider the case of even n. It is clear that χ′ (K2 ) = 1, so we can assume
that n ≥ 4. Since ∆(Kn ) = n − 1, we just need to prove that Kn is (n − 1)-
edge-colourable. But if we delete the vertex n, the result is the complete
graph Kn−1 , and we have just seen how to construct an edge colouring of
Kn−1 with colours 1, · · · , n − 1. It is clear from our construction that none
of the edges coloured with colour i ends at the vertex i. So when we add the
vertex n, we can colour the edge {i, n} with colour i for i = 1, · · · , n − 1, and
the result is an edge colouring of Kn with colours 1, · · · , n − 1. The proof is
finished.
90 Introduction to Graph Theory

Remark 4.31*. As hinted at the beginning of this section, the problem of

finding edge colourings of G can be rephrased in terms of vertex colourings.
Let Gb denote a graph whose vertices correspond to the edges of G, where two
vertices are adjacent if and only if the corresponding edges of G have an end in
common. Then the edge colourings of G correspond to the vertex colourings
b and χ′ (G) = χ(G).
of G, b For instance, C cn happens to be isomorphic to Cn ,
which is why the computation of χ′ (Cn ) in Example 4.29 was essentially the
same as the computation of χ(Cn ). The reason for not converting everything
into vertex colourings is just that our general results about (vertex) chromatic


numbers are a bit weak. For instance, K cn for n ≥ 4 is a graph with n
2
vertices, which is regular of degree 2n − 4 (every edge of Kn has one of
its ends in common with n − 2 other edges, and the other end in common
with n − 2 further edges). The best general result for such a graph, Brooks’
Theorem, would tell us only that χ(K cn ) ≤ 2n − 4, whereas Theorem 4.30
′
determines χ (Kn ) exactly.

The main result about edge colourings is surprisingly conclusive.

Theorem 4.32 (Vizing). Any graph G is (∆(G) + 1)-edge-colourable. So

χ′ (G) is always either ∆(G) or ∆(G) + 1.

Thus the whole universe of graphs is divide into two types, those for which
χ′ (G) = ∆(G) and those for which χ′ (G) = ∆(G) + 1. For instance, Exam-
ple 4.29 says that even cycles are the first type whereas odd cycles are the
second type; Theorem 4.30 says that complete graphs with an even number
of vertices are the first type, whereas complete graphs with an odd number
of vertices (at least 3) are the second type.

We will not give the general proof of Vizing’s Theorem, but some idea of the
argument is provided by the following result, which shows that all bipartite
graphs (including all trees and forests) are of the first type.

Theorem 4.33. Any bipartite graph G is ∆(G)-edge-colourable; hence its

edge chromatic number is ∆(G).

Proof*. The proof is by induction on the number of edges of G; we know

the result if there are no edges, so assume that G has some edges and that
CHAPTER 4. COLOURINGS OF GRAPHS 91

the result holds for bipartite graphs with fewer edges than G. Let {v, w} be
an edge of G. By the induction hypothesis, the graph G − {v, w} (which is
clearly still bipartite) has an edge colouring with ∆(G − {v, w}) colours, so it
has an edge colouring with ∆(G) colours. We choose such an edge colouring
of G − {v, w} where the colour set is {1, 2, · · · , ∆(G)}; we want to extend
this edge colouring to the whole of G (possibly after some modification).

Since the degree of v in G − {v, w} is at most ∆(G) − 1, there must be at

least one colour not used among the edges ending at v; similarly, there must
be at least one colour not used among the edges ending at w. If there is any
colour which is not used either among the edges ending at v or among the
edges ending at w, we can simply colour the edge {v, w} with that colour,
and we are finished. So we can assume that there is a colour i which is used
among the edges ending at v but not among the edges ending at w, and
another colour j which is used among the edges ending at w but not among
the edges ending at v.

Now starting from the vertex v = v0 , we can build a path in G−{v, w} where
the edges alternate between the colours i and j: the first edge {v0 , v1 } is the
unique edge ending at v of colour i, the next edge {v1 , v2 } is the unique edge
ending at v1 of colour j (if such an edge exists), the next edge {v2 , v3 } is the
unique edge ending at v2 of colour i (if such an edge exists), and so on until
the next required edge does not exist. This is indeed a path, because if any
vertex was ever repeated, we would have three edges ending at a vertex with
only two colours between them, contradicting the edge colouring property
(or if v0 was the repeated vertex, we would have two edges ending there with
the colour i, because we assumed the colour j is not used by any edge ending
there). If w belonged to this path, it would have to be the other end-vertex,
coming after an edge of colour j (there being no edge of colour i with which
to continue the path). But this would mean that w = vm for some even
integer m, and then we could add the edge {v, w} to form a cycle in G with
an odd number of vertices (namely, m + 1); since G is bipartite, it contains
no odd cycles, so this cannot be the case. So w does not belong to the path.
We can now exchange the colours i and j throughout the path; this clearly
still gives an edge colouring of G − {v, w}, in which the colour i is not used
among the edges ending at v or the edges ending at w. We can then colour
{v, w} with colour i, and we are finished.
92 Introduction to Graph Theory

Note that this proof gives a concrete recursive procedure for finding an edge
colouring of G with ∆(G) colours.

Returning to the situation of students and lecturers imagined at the start

of the section, we can now conclude that the minimum number of timeslots
required for all the meetings is the obvious lower bound, namely the largest
number of meetings that any single student or lecturer is involved in (as seen
in Example 4.27).
Example 4.34. A special case of Theorem 4.33 is that the complete bipartite
graph Kp,p has an edge colouring with p colours. We can display such an edge
colouring in a p × p array, where the entry in row i and column j is the colour
of the edge between vertices i and p + j. The fact that no two edges ending
at vertex i have the same colour means that no two entries in row i are
the same; the fact that no two edges ending at vertex p + j have the same
colour means that no two entries in column j are the same. So every row
of the array contains each of the p colours exactly once, and so does every
column; such an array is called a ‘Latin square’, and we have proved that
Latin squares of all sizes exist. The p = 9 case is familiar to anyone who has
ever done a Sudoku puzzle!

Since it is traditional for books to end with a wedding, we will use Theorem
4.33 to prove a special case of what is called the Marriage Theorem.
Example 4.35. Suppose there are n men and n women, and every woman
has a list of which of the men she would be happy to marry. Every woman’s
list has d men on it, and every man occurs on d lists, where d ≥ 1. Is there
a way to marry the men off to the women so that every woman gets one of
the men on her list? If we construct a graph G where the vertices are the
people and there is an edge between a man and a woman if the man is on
the woman’s list, then G is bipartite and regular of degree d. By Theorem
4.33, there is an edge colouring of G with d colours. Let i be any colour;
since the edges coloured i cannot have any men or women in common, there
are at most n of them. But the total number of edges is nd, so in fact there
must be exactly n edges of colour i for every i. Thus for every i, the edges
of colour i join n men to n women, and thus provide a way to marry them
off. So the answer is that there is not just one way; you can actually find d
different ways which have no weddings in common.