MATH 239: Introduction to Combinatorics

Chris Thomson

Winter 2013, University of Waterloo

Notes written from Bertrand Guenin’s lectures.

I’d love to hear your feedback. Feel free to email me at [email protected].

See cthomson.ca/notes for updates. Last modified: August 28, 2015 at 10:40 PM (80e35dc).

1
Contents
1 Introduction, Permutations, and Combinations 4
1.1 Course Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Sample Counting Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Sample Graph Theory Problems . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Permutations and Combinations . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.1 Set Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.2 Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.3 Application: Binomial Theorem . . . . . . . . . . . . . . . . . . . . . . 6

2 Simple Tools for Counting 6

2.1 Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Pascal’s Triangle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Injections, Bijections, and Onto . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Application: Binomial Coefficients . . . . . . . . . . . . . . . . . . . . . . . . 8

3 Power Series and Generating Functions 8

3.1 Power Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1.1 A General Counting Problem . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Generating Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Working With (Formal) Power Series . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.1 Finding Inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.2 Compositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.3 Cartesian Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 The Sum and Product Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4.1 Sum Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4.2 Product Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4 Binary Strings 21
4.1 Generating Functions for Strings . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Counting Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2.1 Recursive String Definitions . . . . . . . . . . . . . . . . . . . . . . . . 25

5 Coefficients from Rational Functions 28

6 Solving Recurrence Relations 30

6.1 Non-homogeneous Recurrences . . . . . . . . . . . . . . . . . . . . . . . . . . 34

7 Introduction to Graph Theory 35

7.1 Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.2 Degrees of Vertices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.3 Bipartite Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.4 Walks, Cycles, and Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.4.1 Hamiltonian Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.5 Disjoint Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
7.5.1 Menger’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.5.2 Connectedness, Subgraphs, and Components . . . . . . . . . . . . . . 48
7.5.3 Equivalence Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2
 7.5.4 Bridges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

8 Trees 52
8.1 More on Bipartite Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.2 Graph Colouring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

9 Planar Graphs 57
9.1 Platonic Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
9.2 Review of Planar Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
9.3 Planar Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
9.4 Non-Planar Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
9.5 Kuratowski’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
9.6 Graph Colouring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
9.7 Dual Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

10 Matching 75
10.1 Application: co-op . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
10.2 Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
10.2.1 Algorithm for Maximum Matching . . . . . . . . . . . . . . . . . . . . 79
10.3 Vertex Covers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
10.4 XY-Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
10.4.1 Algorithm For Finding a Maximum Matching and Minimal Vertex Cover 82
10.5 Hall’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

3
1 Introduction, Permutations, and Combinations
← January 7, 2013
1.1 Course Structure
The grading scheme is 50% 55% final, 30% midterm, 10% quizzes 5% participation marks
(clicker questions), and 10% assignments. There are ten quizzes and ten assignments. The
quizzes will not be announced in advance, and each quiz consists of a single clicker question.
Assignments are typically due on Friday mornings at 10 am. in the dropboxes outside of
MC 4066. The midterm exam is scheduled for March 7, 2013, from 4:30 pm - 6:20 pm.
There is no textbook for the course, but there are course notes available at Media.doc in
MC, and they’re highly recommended.

MATH 239 is split into two parts: counting (weeks 1-5) and graph theory (weeks 6-12).

See the course syllabus for more information – it’s available on Waterloo LEARN.

1.2 Sample Counting Problems

Problem 1.1. How many ways are there to cut a string of length 5 into parts of sizes 1
and 2?

Here are a few example cuts:

1 2 3 4 5 5 cuts: size 1, size 1, size 1, size 1, size 1.

1 1 2 2 3 3 cuts: size 2, size 2, size 1.
1 2 2 3 3 3 cuts: size 1, size 2, size 2.

This is a finite problem. You could count all of the possibilities manually in this case.
However, this problem could be made more complicated to a point where manually count-
ing all possibilities would become quite cumbersome, as is the case in the next problem.

Problem 1.2 (Cuts). How many ways are there to cut a string of size 372,694 into parts
of sizes 3, 17, 24, and 96?
Definition. A positive integer n has a composition (m1 , m2 , . . . , mk ), where m1 , . . . , mk
are positive integers and where n = m1 + m2 + . . . + mk . m1 , . . . , mk are the parts of the
composition.

Problem 1 could be rephrased as looking for the number of compositions of 5 where all parts
are 1 or 2.
Problem 1.3. How many compositions of n exist such that all parts are odd?

Problem 1.4 (Binary Strings). Let S = a1 , a2 , . . . , an where ai ∈ {0, 1}. How many strings
S exist?
For each ai , there is the choice between 0 or 1, and that choice is independent for each
character of the string (for each ai ). So, there are 2n binary strings of length n.
Problem 1.5. How many binary strings of size n exist that do not include the substring
1100?

4
For example: 10101100101 6∈ S.
Problem 1.6. How many binary strings of size n exist such that there is no odd-length
sequences of zeroes?
For example: 100100100011 6∈ S.

Problem 1.7 (Recurrences). How many times does a recursive function get called for a
particular input n?

1.3 Sample Graph Theory Problems

Problem 1.8. For any arbitrary map of regions, colour the regions such that no two touch-
ing boundaries do not have the same colour, with the least number of colours possible.
The four-colour theorem states that you can always do this with four colours. It’s also
always possible to colour these regions with five colours. It’s sometimes possible to colour
the regions with three or fewer colours, depending on the layout of the regions and their
boundaries.

1.4 Permutations and Combinations

1.4.1 Set Notation
The usual set and sequence notation is used in this course. (1, 2, 3) is a sequence (where
order matters), and {1, 2, 3} is a set (where order does not matter).

We will also be using one piece of notation you may not be familiar with: [n] := {1, 2, . . . , n}.

1.4.2 Permutations
Definition. A permutation of [n] is a rearrangement of the elements of [n]. The number
of permutations of a set of n objects is n × (n − 1) × . . . × 1 = n!.
For example: the number of permutations of 6 objects is 6×5×4×3×2×1 = 6! permutations.

Why is this the case? Simple: there are n choices for the first position, (n − 1) choices
for the second position, (n − 2) choices for the third position, and so on, until there’s 1
choice for the nth position.

Definition. A k-subset is a subset of size k.

Problem 1.9. How many k-subsets of [n] exist?
6×5×4×3
Let’s consider a more specific case: how many 4-subsets of 6 are there? 4! .

For simplicity’s sake, we

will introduce notation for this, which we will refer to as a combi-
nation, denoted as nk = k!(n−k)!
n!
where n, k ∈ Z ≥ 0.

Proposition. There are nk k-subsets of [n].

5
1.4.3 Application: Binomial Theorem
← January 9, 2013
n  
X n k
(1 + x)n = x
k
k=0

Why is this true?

1 2 3        
z }| { z }| { z }| { 3 3 3 2 3 3
(1 + x) = (1 + x) (1 + x) (1 + x) = 1 + 3x + 3x2 + x3 =
3
+ x+ x + x
0 1 2 3
Proof.
1 2 n
z }| { z }| { z }| {
n
(1 + x) = (1 + x) (1 + x) · · · (1 + x)
n


In order to get xk , we need to choose x in k of {1, . . . , n}. There are k ways of doing
this.

2 Simple Tools for Counting

2.1 Partitioning
Sets S1 , S2 partition the set S if S = S1 ∪ S2 and S1 ∩ S2 = ∅.
Example 2.1.
(
S1 = {1, 2}
S = [5] =
S2 = {3, 4, 5}
|S| = |S1 | + |S2 |
n  
X n
Proposition. 2n =
k
k=0

Proof. We will discuss two proof methods.

1. Algebraic proof. Set x = 1 in the Binomial Theorem.
2. Combinatorial proof. We will count the left-hand side and the right-hand side in
different ways to reach the same result.
Let S be the set of subsets of [n]. |S| = 2n , since for every element of [n] we have two
possibilities: include or don’t include the element in S.
Aside: suppose n = 2. Then S = {∅, {1}, {2}, {1, 2}}.
Partition S into S0 , S1 , . . . , Sn , where Sk is the set of k-subsets of [n].

|S| = |S0 | + |S1 | + · · · + |Sn |

|{z} |{z} |{z} |{z}
2n (n0 ) (n1 ) (nn)

6
      
n n−1 n−1
Proposition. = +
k k k−1

Proof. Let S be the set of k-subsets of [n]. Then |S| = nk . Partitioning S, let S1 be the

subsets of S containing n, and let S2 be the subsets of S not containing n.

It’s easy to see that |S1 | = n−1 n−1

k−1 (n
is already included in our choices) and |S2 | = k .
We now have |S| = |S1 | + |S2 | = n−1 n−1


k−1 + k .

2.2 Pascal’s Triangle

Pascal’s Triangle is a triangle where each value is determined by the sum of its two direct
parents. The uppermost value is 1.

n = 0: 1
n = 1: 1 1
n = 2: 1 2 1
n = 3: 1 3 3 1
n = 4: 1 4 6 4 1

  r  
q+r X q+i−1
Proposition. =
q i=0
q−1
       
5 2 3 4
For example: let q = 3, r = 2. Then we have: = + + .
3 2 2 2

Proof. Let S be the set of q-subsets of [q + r], so |S| = q+r

q . Partition S such that Si is
the set of q-subsets where the largest element is q + i (where i = 0, . . . , r).
q+i−1


We have: |S| = |S0 | + |S1 | + · · · + |Sr |. Note that |Si | = q−1 . That gives us:

Xr
|S| = |Si |
|{z}
q+r i=0
( q ) | {z }
(q+i−1
q−1 )

2.3 Injections, Bijections, and Onto

← January 11, 2013
Let’s consider a function f : S → T .
Definition. The function f is an injection if for all x1 , x2 ∈ S, x1 6= x2 such that f (x1 ) 6=
f (x2 ). That is, no element in the codomain T is the image of more than one element in S.

7
Definition. The function f is onto (or a surjection) if for all y ∈ T , there exists x ∈ S
such that f (x) = y. That is, all elements in the codomain T are the image of an element in
S. Multiple elements in S can map to the same element in T .
Definition. The function f is a bijection if it is both injective and onto. That is, there is
a one-to-one mapping between elements in S and T , and vice versa.

Proposition. If S and T are finite, |S| = |T | if f is bijective.

Definition. g is the inverse of f if:
1. For all x ∈ S, g(f (x)) = x.

2. For all y ∈ T, f (g(y)) = y.

In order to show that a function has a bijection, find the inverse function.

2.4 Application: Binomial Coefficients

   
n n
Proposition. =
k n−k
Proof. I will prove this proposition in a combinatorial proof using the bijection technique.
We want to show that the cardinalities are the same.

|S1 | = nk , as shown earlier. Let S2 be the set

Let S1 be the set of k-subsets of [n], so
n


of (n − k)-subsets of [n], so |S2 | = n−k . We need to show that |S1 | = |S2 |, so we need to
show that there is a bijection between S1 and S2 , f : S1 → S2 .

Aside: suppose n = 5 and k = 2. Then, the bijection could be {1, 3} → {2, 4, 5} (the
complement function).

The bijective function is f (A) = [n]\A (the complement function). Check: f is its own
inverse (let g = f ).

3 Power Series and Generating Functions

3.1 Power Series
Definition. Let (a0 , a1 , . . .) be a sequence of rational numbers. Then:
X
A(x) = ai xi
i≥0

This is called a power series.

Example 3.1. ai = 2i =⇒ A(x) = 1 + 2x + 4x2 + 8x3 + · · ·
If the sequence is finite, it is just a polynomial.

8
3.1.1 A General Counting Problem
Let S be the set of objects σ. Each object σ has a weight w(σ). We want to know how
many objects of S have some weight k.
Example 3.2. Let σ ⊆ [n], w(σ) = |σ|. The question becomes: how many k-subsets of [n]?
Example 3.3. Let σ be a set of coins (1¢, 5¢, 10¢, 25¢, $1, $2), and let w(σ) be the total
value of the coins in σ.

The question becomes: how many sets of coins have total value k? In other words, how
many ways are there to give k amount of change?

3.2 Generating Functions

Definition. Given S as the set of objects σ and a weight function w(σ):
X
φS (x) = xw(σ)
σ∈S

This is called the generating function for S, w.

Example 3.4. Let S = {σ|σ ⊆ {1, 2, 3}}, w(σ) = |σ|.

σ w(σ) xw(σ)
∅ 0 1
{1} 1 x
{2} 1 x
{3} 1 x
{1, 2} 2 x2
{1, 3} 2 x2
{2, 3} 2 x2
{1, 2, 3} 3 x3

φS (x) = 1 + 3x + 3x2 + x3 = (1 + x)3 is the generating function. Notice that the coefficients
of each term are the number of objects who have that the weight indicated by the exponent
of the x in that term.
Remember: the generating function for S with weights w is φS (x) = k≥0 ak xk , where
P
ak is the number of objects of size k in S.
← January 14, 2013
Example 3.5. Let S be the set of subsets of [n], and w(σ) = |σ|.
X n
φS (x) = xk = (1 + x)n
k
k≥0

Note that nk is included because that’s the number of k-subsets of [n]. φS (x) = (1 + x)n

by the binomial theorem.

Proposition. Let φS (x) be the generating function for finite-size S with weight w. Then:
1. φS (1) = |S|

9
 2. φ0S (1) = sum of the weight of all the objects in S.
Together, we get:

φ0S (1)
= average weight of objects in S
φS (1)

Considering the previous example again, we know that |S| = 2n and the average weight is
clearly n2 . We can verify that with this proposition.

φS (x) = (1 + x)n =⇒ φS (1) = (1 + 1)n = 2n

φ0S (x) = n(1 + x)n−1 =⇒ φ0S (1) = n2n−1
φ0S (1) n2n−1 n
average weight = = =
φS (1) 2n 2

We’ll now prove this proposition more generally.

Proof. We will prove the two parts of the proposition separately. The average weight clearly
follows from those two results.
1.
X
φS (x) = xw(σ)
σ∈S
P
Choose x = 1. Then φS (1) = σ∈S 1 = |S|.
2.
X
φ0S (x) = w(σ)xw(σ)−1
σ∈S

Choose x = 1. Then φ0S (1) =

P
σ∈S w(σ) · 1 = total weight of all objects in S.

In order to work further with generating functions, we’ll first need to learn how to manipulate
power series generally.

3.3 Working With (Formal) Power Series

For the following definitions, we will assume the following power series are defined:

A(x) = a0 + a1 x + a2 x2 + · · ·
B(x) = b0 + b1 x + b2 x2 + · · ·

Definition. We define addition of power series as follows.

X
A(x) + B(x) := (an + bn )xn
n≥0

Note that this definition is consistent with the definition of addition for polynomials.

10
Definition. We define multiplication of power series as follows.
n
!
X X
A(x)B(x) := ak bn−k xn
n≥0 k=0

Example 3.6.

(1 + x + x2 + x3 + · · · )(1 − x)
= 1 · 1 + (1 · −1 + 1 · 1)x + (1 · 1 + 1 · −1)x2 + · · ·
=1

Definition. B(x) is the inverse of A(x) if A(x)B(x) = 1 (alternatively, B(x)A(x) = 1).

1
We will use the notation B(x) = A(x) to indicate that B(x) is the inverse of A(x), and
vice versa.
Example 3.7. The inverse of (1 + x + x2 + · · · ) is (1 − x), as shown in the previous example.
Question: does every power series have an inverse? No. For example, (x + x2 ) does not
have an inverse.
Proof. Suppose (x+x2 ) has an inverse. If (x+x2 ) has an inverse, there would exist constants
bi such that

(x + x2 )(b0 + b1 x + b2 x2 + b3 x3 + · · · ) = 1

Clearly, this is impossible. In order to equal 1, b0 must multiply by some constant term,
however there is no other constant term. Therefore, our assumption was incorrect, meaning
(x + x2 ) does not have an inverse.
Remark: if a power series does not have a constant term then it has no inverse, and vice
versa.
Theorem. If the constant term of A(x) is non-zero, then A(x) has an inverse, and we can
find it.
Notation: given a power series A(x), we say [xk ]A(x) represents the coefficient of xk .
Example 3.8. Find the inverse of 1 − x + x2 − x3 + x4 − · · · .

(1 − x + x2 − x3 + x4 − · · · )(b0 + b1 x + b2 x2 + · · · ) = 1
| {z }
(?)

[x0 ](?) = 1 =⇒ 1 · b0 = 1 =⇒ b0 = 1
[x1 ](?) = 0 =⇒ 1 · b1 − 1 · b0 = 0 =⇒ b1 = b0 = 1
[x2 ](?) = 0 =⇒ 1 · b2 − 1 · b1 + 1 · b0 = 0 =⇒ b2 = b1 − b0 = 0

Similarly, b3 = b4 = · · · = 0. Thus, the inverse is (1 + x).

It’s clear that the inverse for a power series is unique. We made no choices when determining
the inverse, therefore it must be unique.

11
3.3.1 Finding Inverses
← January 16, 2013
Given:
X
A(x) = an xn where (a0 6= 0)
n≥0

We want to find:
X
B(x) = bn x n
n≥0

We’ll start by finding b0 :

1
[x0 ]A(x)B(x) = a0 b0 = 1 =⇒ b0 =
a0
Notice that if we didn’t have the restriction on a0 , we would’ve run into trouble here. Now,
suppose you found b0 , b1 , . . . , bn−1 for n ≥ 1. Find bn .
1
bn = · (−a1 bn−1 − a2 bn−2 − . . . − an b0 )
a0
Proposition. Let A(x) and P (x) be formal power series. Suppose the constant for A(x) is
not 0. Then there exists a unique B(x) such that A(x)B(x) = P (x).
Some useful formulæ:
1
= 1 + x + x2 + x3 + · · · since 1 = (1 + x + x2 + · · · )(1 − x)
1−x
1 − xk+1
= 1 + x + x2 + · · · + xk since 1 − xk+1 = (1 + x + x2 + · · · + xk )(1 − x)
1−x

3.3.2 Compositions
Definition. Let A(x) and B(x) be formal power series defined by:
X X
A(x) = an xn B(x) = bn x n
n≥0 n≥0

We define the composition as:

X
A(B(x)) := an [B(x)]n
n≥0

Example 3.9. Let A(x) = 1 + x + x2 + x3 + · · · and B(x) = 2x. Then the composition is
A(B(x)) = 1 + 2x + 4x2 + 8x3 + · · · .
Question: is the composition of two formal power series a formal power series itself? No.
Example 3.10. Suppose we pick A(x) = 1 + x + x2 + x3 + · · · and B(x) = 1 + x.

We have A(B(x)) = 1 + (1 + x) + (1 + x)2 + (1 + x)3 + · · · . This cannot be a formal

power series because a formal power series requires that all coefficients (ak ) need to be
rational numbers, and in this case a0 is infinite.

12
Example 3.11. Suppose we pick A(x) = 1 + x + x2 + x3 + · · · and B(x) = x + x2 .

We have A(B(x)) = 1 + (x + x2 )1 + (x + x2 )2 + · · · . In order to show that this is a

formal power series, we need to show that the coefficient of xk (which is ak ) is finite for all
k.

Theorem. Let A(x) and B(x) be formal power series, and [x0 ]B(x) = 0 (B(x)’s constant
is zero). Then A(B(x)) is a formal power series.
Aside:

A(B(x)) = 1 + (x + x2 )1 + (x + x2 )2 + (x + x2 )3 + · · ·
| {z } | {z } | {z }
x(1+x) x(1+x) x(1+x)

Proof. We need to show that for any fixed k, [xk ]A(B(x)) is finite.

Let R(x) be such that B(x) = xR(x). Then we have:

X
[xk ]A(B(x)) = [xk ] an [B(x)]n
n≥0
X
= [xk ] an xn R(x)
n≥0
k
X
= [xk ] an xn (R(x))n
n≥0

We made the last sum finite because we’re interested in the coefficient of xk only. Note that
this last line shows we can determine the coefficient in a finite number of steps, since k is
finite.

Example 3.12. Let y 7→ x2 (I will call x2 , y).

1 1
= = 1 + y + y2 + y3 + · · ·
1 − x2 1−y
= 1 + x2 + x4 + x6 + · · ·

This is a composition with A(x) = 1 + y + y 2 + · · · and B(x) = x2 (constant is zero). We

were only allowed to do this because [x0 ]B(x) = 0.

3.3.3 Cartesian Product

Definition. Let A and B be sets. The cartesian product is defined as:

A × B := {(a, b) a ∈ A, b ∈ B}

Note that (a, b) is an ordered pair.

Example 3.13. Let A = {1, 2, 3} and B = {x, y}. Then A×B = {(1, x), (2, x), (3, x), (1, y), (2, y), (3, y)}.
The cardinality of A × B is the product of the cardinalities of A and B: |A × B| = |A||B|.

13
3.4 The Sum and Product Lemmas
Let’s consider a bin of two red marbles – one large marble (denoted R), and one small
marble (denoted r). We define A to be the set of all selections of ≥ 1 marbles. We know:
A = {{r}, {R}, {r, R}}
w(σ) = |σ|
φA (x) = x + x + x2 = 2x + x2
Let’s now consider a bin of green marbles – two large marbles (denoted G). We define B to
be the set of all selections of ≥ 1 marbles. We know:
B = {{G}, {G, G}}
w(σ) = |σ|
φB (x) = x + x2

3.4.1 Sum Lemma

Let S = A ∪ B = {{r}, {R}, {r, R}, {G}, {G, G}}, and note that A ∩ B = ∅. What is the
generating function for S?
φS (x) = (x + x + x2 ) + (x + x2 )
| {z } | {z }
φA (x) φB (x)

This works in general.

Theorem (Sum Lemma). We have a set S of objects with weight w. Let A and B be
partitions of S. Then:
φS (x) = φA (x) + φB (x)
Proof.
X X X
φS (x) = xw(σ) = xw(σ) + xw(σ)
σ∈S σ∈A σ∈B
| {z } | {z }
φA (x) φB (x)

3.4.2 Product Lemma

← January 18, 2013
Let S = A×B = {({r}, {G}), ({R}, {G}), ({r, R}, {G}), ({r}, {G, G}), ({R}, {G, G}), ({r, R}, {G, G})}.
S is the various ways of selecting marbles from both bins. Let w(σ) be the number of mar-
bles selected in total.

Note: you can think of a generating series as the sum of all objects’ xw(σ) , or you can
think of it as the sum of all ak xk .
φS (x) = x2 + x2 + x3 + x3 + x3 + x4
= 2x2 + 3x3 + x4
= (x + x + x2 )(x + x2 )
= φA (x) · φB (x)

14
This works in general.

Each term in φA (x) corresponds to an object in A, and each term in φB (x) corresponds
to an object in B. When you find the product, it automatically does the counting for you,
due to exponentiation laws.

Theorem (Product Lemma). Let A be a set where objects have weight α, and let B be a
set where objects have weight β. Let S = A × B be a set of all objects (a, b) ∈ S, where
w[(a, b)] = α(a) + β(b). Then:

φS (x) = φA (x) · φB (x)

Proof.
X
φS (x) = xw[(a,b)]
(a,b)∈S
X
= xα(a)+β(b)
(a,b)∈S
X  X 
= xα(a) · xβ(b)
a∈A b∈B
| {z } | {z }
φA (x) φB (x)

Example 3.14. Given integers n ≥ 0 and k ≥ 0, how many k-tuples (m1 , m2 , . . . , mk )

(where mi ≥ 0, mi ∈ Z) exist where m1 + m2 + . . . + mk = n?

Suppose n = 3 and k = 2. You’ll have (0, 3), (1, 2), (2, 1), and (3, 0) – four possibilities
(notice order matters).

We have N0 = {0, 1, 2, 3, . . .}, w(a) = a (for a ∈ N0 ). Then:

X 1
φN0 (x) = an xn = 1 + x + x2 + x3 + . . . =
1−x
n≥0

N appears k times
z 0 }| {
Now, let S = N0 × N0 × . . . × N0 . The objects in S are sequences of form (m1 , m2 , . . . , mk ),
where mi ≥ 0, mi ∈ Z. We define w[(m1 , m2 , . . . , mk )] = m1 + m2 + . . . + mk .
1
φS (x) = [φN0 (x)]k = by the product lemma
(1 − x)k

Therefore, the number of k-tuples (m1 , m2 , . . . , mk ) (where mi ≥ 0, mi ∈ Z) where m1 +

1
m2 + . . . + mk = n is given by [xn ] (1−x) k . We cannot figure out this coefficient. Let’s leave

generating functions for now, and find a combinatorial proof for this.
When you use the product lemma, you need to ensure:
1. Ensure it’s used on something found by taking the cartesian product.

15
 2. The weight is equal to the sum of the weights.
Example 3.15. Let n = 5 and k = 3. We have:

(1, 2, 2): • | • • | • •
(2, 3, 0): • • | • • • |

We have n dots and k − 1 intervals, which gives us:

   
n + (k − 1) 7
= (in this case)
k−1 2

We always want two bars because k = 3. In general, we always want k − 1 bars. We want
n dots. The combination above is formed by the need to place k − 1 bars among all of the
positions.
By combining the previous two examples, we have proven the following theorem.
Theorem (Theorem 1.6.5).
 
n 1 n+k−1
[x ] =
(1 − x)k k−1

Definition. Some integer n has composition (m1 , m2 , . . . , mk ) if m1 + m2 + . . . + mk = n,

where mi ≥ 1, mi ∈ Z. Note that the only difference from earlier is that mi ≥ 1 instead of
mi ≥ 0.
Example 3.16. We have N = {1, 2, 3, . . .}, w(a) = a (for a ∈ N), which gives us:
x
φN (x) = x + x2 + x3 + . . . = x(1 + x + x2 + . . .) =
1−x
Note that φN (x) is missing the object of size zero (the constant term).
k occurrences of N here
z }| {
Let S = N × N × . . . × N . We also define the weight as w[(m1 , m2 , . . . , mk )] = m1 +
m2 + . . . + mk . Then:

xk
φS (x) = [φN (x)]k =
(1 − x)k
1
We’re interested in finding [xn ]xk (1−x) k. ← January 21, 2013

1 1
[xn ]xk k
= [xn−k ]
(1 − x) (1 − x)k
 
(n − k) + k − 1
=
k−1
 
n−1
=
k−1

16
Example 3.17. We now want to know the number of compositions of n (where n ≥ 1)
where we have an arbitrary number of parts. We have: S = N ∪ N2 ∪ N3 ∪ N4 ∪ N5 ∪ . . ..
The weight function is the same as before – the sum of the parts. That gives us:

X X  x k
φS (x) = φNk (x) =
1−x
k≥1 k≥1

We want:
X  x k
[xn ]
1−x
k≥1

Which is a composition of:

y
A(y) = y + y 2 + y 3 + . . . = y(1 + y + y 2 + . . .) =
1−y
x
B(x) =
1−x
The composition is A(B(x)), but we need to be careful before using compositions. We can
easily check that the constant of B(x) is not zero, as required:
x
B(x) = = x(1 + x + x2 + . . .) =⇒ [x0 ]B(x) = 0
1−x
The composition is well-defined. So, we have:

X  x k x
x
= 1−xx =
1−x 1 − 1−x 1 − 2x
k≥1

x
We still need to figure out the coefficient [xn ] 1−2x .

x 1
[xn ] = [xn−1 ] (1)
1 − 2x 1 − 2x
= [xn−1 ](1 + y + y 2 + . . .) (2)
n−1 2
= [x ](1 + (2x) + (2x) + . . .) (3)
n−1
=2 (4)

(1) is a common, useful trick. We eliminate any instances of x that are simply multiplied
by something, by reducing the power of the coefficient we’re looking for. (4) is true since
every term is in the form (2x)k .
We might be able to avoid generating functions in some cases by using a combinatorial proof.
Now, we’ll look at an arbitrary, convoluted example where a combinatorial proof would not
suffice. This aims to show that generating functions are solid and can apply in many more
situations than combinatorial proofs.
Example 3.18. We want to find the number of compositions of n with 2k parts, where the
first k parts are ≤ 5 and the last k parts are ≥ 3.

17
A specific instance of this problem is when n = 22 and k = 3, we have:
(|{z}
. . . , |{z}
... )
≤5 ≥3

(1, 3, 5, 3, 6, 4)
| {z } | {z }
≤5 ≥3

It’s useful to think about what you’re counting. In this case, you’re counting the k-tuples,
so this is the cartesian product of six sets (the first three being the set of positive integers
≤ 5 and the last three being the set of integers ≥ 3).
N≤5 = {1, 2, 3, 4, 5}
N≥3 = {3, 4, 5, . . .}

S = (N≤5 )k × (N≥3 )k
The weight is the sum of the six individual numbers in the tuple, since N≤5 and N≥3 ’s
weights are both defined by w(a) = a, which gives us:
φN≤5 (x) = x + x2 + x3 + x4 + x5
φN≥3 (x) = x3 + x4 + x5 + . . .
← January 23, 2013
We can expand the generating functions φN≤5 (x) and φN≥3 (x).

φN≤5 (x) = x + x2 + x3 + x4 + x5
= x(1 + x + x2 + x3 + x4 )
1 − x5
 
=x
1−x

φN≥3 (x) = x3 + x4 + x5 + . . .
= x3 (1 + x + x2 + . . .)
x3
=
1−x
Now, we have S = (N≤5 )k ×(N≥3 )k with weight function w[(a1 , . . . , a2k )] = a1 +a2 +. . .+a2k .
By the product lemma, we get:
5 k
   3 k
k 1−x x
φS (x) = x
1−x 1−x
= x4k (1 − x5 )k (1 − x)−2k
We’re still interested in finding [xn ]φS (x). This becomes tedious, but the general idea is:
[xn ] x4k
|{z} · (1 − x5 )k · (1 − x)−2k
| {z } | {z }
collapse into [xn ] to become [xn−4k ] expand with binomial theorem use formula [xn ] 1
(1−x)k

18
After much tedious work, you’ll discover that the final result is:
b n−4k
5 c   
n
X k
i n − 5i − 2k − 1
[x ]φS (x) = (−1)
i=0
i 2k − 1

This couldn’t have been found with a combinatorial proof. That’s why generating functions
are powerful – they work even in convoluted situations like this one.
Example 3.19. We’re interested in the number of compositions of n where all parts are
odd. There can be an arbitrary number of parts.

For example, with n = 5, the five solutions are:

{(1, 1, 1, 1, 1), (5), (3, 1, 1), (1, 3, 1), (1, 1, 3)}
We’ll define n = 0 to have a unique composition (), the empty tuple.

We have Nodd = {1, 3, 5, 7, . . .}, with its weight function defined to be w(a) = a for any
a ∈ Nodd . That gives us its generating function:
φNodd (x) = x + x3 + x5 + x7 + . . . (1)
2 4 6
= x(1 + x + x + x + . . .) (2)
2 3 2
= x(1 + y + y + y + . . .) where y = x (3)
x
= (4)
1−y
x
= (5)
1 − x2
(3) is an acceptable composition because x2 has a constant term of zero.
[
Now, we have that S = () Nkodd . By the sum and product lemmas, the generating
k≥1
function for S is:
X
φS (x) = 1 + φkNodd (x) (1)
k≥1
X k
x
=1+ (2)
1 − x2
k≥1
X x k
= (3)
1−x
k≥0
x
= 1 + z + z 2 + z 3 + . . . where z = (4)
1 − x2
1
= (5)
1−z
1
= x (6)
1 − 1−x 2

1 − x2
= (7)
1 − x − x2

19
 x
(4) is a valid composition because 1−x2 has a constant term of zero.
2
1−x
What is [xn ] 1−x−x 2 ? (Don’t think of assigning values to x. Just focus on canceling them

out.)

1 − x2 X
2
= an xn
1−x−x
n≥0

1 − x = (1 − x − x2 )(a0 + a1 x + a2 x2 + . . .)
2

= a0 + a1 x + a2 x2 + a3 x3 + . . . − a0 x − a1 x2 − a2 x3 − a3 x4 − . . . − a0 x2 − a1 x3 − a2 x4 − a3 x5 − . . .
= a0 + (a1 − a0 )x + (a2 − a1 − a0 )x2 + (a3 − a2 − a1 )x3 + . . .
We know a0 = 1 and (a1 − a0 ) = 0, and so on. We can express a recurrence relation for this.

a0 = 1
a1 − a0 = 0 =⇒ a1 = 1
a2 − a1 − a0 = −1 =⇒ a2 = −1 + a1 + a0 = 1
a3 − a2 − a1 = 0 =⇒ a3 = a1 + a2 = 2
a4 = a3 + a2 = 3
a5 = a4 + a3 = 5
..
.
For all n ≥ 3, we have that an = an−1 + an−2 . This defines the fibonacci numbers.
a+b
Definition. The golden ratio is a pair of integers a and b such that a = ab . It’s a ratio
that comes up in nature a lot, and is aesthetically pleasing.
an
You can approximate the golden ratio with the Fibonacci numbers. an−1 approaches the
golden ratio as n → ∞.
Example 3.20. Let Sn be the set of compositions into odd numbers of n. |Sn | = |Sn−1 | +
|Sn−2 | for n ≥ 2. Sn is partitioned by Sn−1 and Sn−2 . Is there a bijection between Sn and
Sn−1 ∪ Sn−2 ?

Let’s look at the case where n = 5. We have: ← January 25, 2013

Sn
z }| {
{(1, 1, 1, 1, 1), (3, 1, 1), (1, 3, 1), (1, 1, 3), (5)}
{(1, 1, 1, 1), (3, 1), (1, 3), (1, 1, 1), (3)}
| {z } | {z }
Sn−1 Sn−2

Notice that each element in Sn−1 and Sn−2 can be obtained from Sn by either removing the
last element entirely, or by subtracting 2 from the last element. That’s the bijection we’re
looking for, which we will now define more formally as f : Sn → Sn−1 ∪ Sn−2 :
(
(a1 , . . . , ak−1 ) ak = 1
f (a1 , . . . , ak ) =
(a1 , . . . , ak−2 ) ak ≥ 3

20
Note that f (a1 , . . . , ak ) is undefined for 1 < k < 3, but that’s okay because we’re only
interested in odd natural numbers.

We have to somehow show that f actually is a bijection. We can do that by construct-

ing its inverse. In our case, the inverse is simply providing an inverse procedure.

Therefore, |Sn | = |Sn−1 | + |Sn−2 |, which is a general property of bijections.

4 Binary Strings
Definition. A binary string is a string composed of 0s and 1s. For example, 011001110011
is a binary string.
Definition. We define the operation of concatenation of two binary strings a1 , a2 as
a1 a2 . For example, if a1 = 110 and a2 = 011, the concatenation a1 a2 = 110011.
Definition. The concatenation AB of two sets (A and B) of binary strings is de-
fined as all possible strings formed as concatenations of one string from set A followed
by one string from set B. For example, if A = {011, 11} and B = {10, 0}, then AB =
{01110, 0110, 1110, 110}.
Definition. Suppose A is a set of strings. Then Ak = AAAAA
| {z . . . A}.
k times

? 2 3
Definition. We define A = {} ∪ A ∪ A ∪ A ∪ . . ., where  represents the empty string.
?
Example 4.1. What is {0, 1} ? It’s the set of all binary strings.
5 5 5
Example 4.2. What’s {0, 1} ? {0, 1} = {0, 1}{0, 1}{0, 1}{0, 1}{0, 1}. {0, 1} is the set of
all binary strings of length 5.
Example 4.3. Describe the set S of all binary strings with no three consecutive 1s. For
example: a = 01100011101001 6∈ S, b = 011110 6∈ S, but c = 0110010011011 ∈ S. We aim
to express S using expression-? concatenation.

Let’s take a second look at c.

11 ∈ S
c = 0, 110, 0, 10, 0, 110, |{z}
| {z }
(1) (2)

?
1. A sequence of at most two ones followed by zero, repeated: {0, 10, 110}
2. A sequence of zero, one, or two ones: {, 1, 11}.
? ?
Therefore, S = {0, 10, 110} {, 1, 11}. Note that  is an element of {anything} .
Definition. A block in a binary string is equal to the maximum sequence of 0s or 1s.
Example 4.4. List all the blocks in the string 1100011110110101.

11 |{z}
|{z} 000 1111 0 |{z}
|{z } |{z} 11 |{z}
0 |{z}
1 |{z}
0 |{z}
1

21
Example 4.5. Describe the set S of strings with no blocks of length 2. For example,
a = 1011000101 6∈ S, but b = 1110111000101000 ∈ S.

Let’s take another look at b.

b = |{z} 000 ∈ S
111 , 0111, 0001, 01, |{z}
| {z }
(1) (2) (3)

1. A sequence of ones, of any length other than length 2: {, 1, 111, 1111, 11111, . . .}.
2. A sequence of zeroes (of any non-zero length other than length 2), followed by a se-
quence of ones (of any non-zero length other than length 2), repeated: [{0, 000, 0000, . . .}{1, 111, 1111, . . .}]? .
3. A sequence of zeroes, of any length other than length 2: {, 0, 000, 0000, 00000, . . .}.
?
It’s important to note that {1, 111, 1111, . . . , 0, 000, 0000, . . .} does not work because you
could pick two zeroes or two ones consecutively, which is not allowed. You have to be very
careful that you don’t generate strings that you don’t want to generate.

4.1 Generating Functions for Strings

We’d like to have something like the product lemma but for strings.

Let S be the set of binary strings. Then the generating function for S is defined as:
X
φS (x) = an xn where an is the number of strings of length n ∈ S
n≥0
← January 28, 2013
Example 4.6. Let A = {0, 01} and B = {0, 10}. Their generating functions are φA (x) =
φB (x) = x + x2 . What is the generating function for AB?

It’s easy to see that AB = {00, 010, 0110}, which gives us φAB (x) = x2 + x3 + x4 . Also note
that:

φA (x) · φB (x) = (x + x2 ) · (x + x2 )
= x2 + 2x3 + x4

Therefore, in this case, φAB (x) 6= φA (x)·φB (x). Why is this? A×B = {(0, 0), (01, 0), (0, 10), (01, 10)}.
Note that the string for (01, 0) and (0, 10) is the same (010), but it’s only going to be included
in AB once. That’s where the trouble is.
Definition. A, B is ambiguous if all three of these properties hold:
• (a1 , b1 ), (a2 , b2 ) ∈ A × B
• (a1 , b1 ) 6= (a2 , b2 )
• a1 b1 = a2 b2
In the previous example, (0, 10)(01, 0) ∈ A × B and (0, 10) 6= (10, 0), but 010 = 010.

Note that the definition of ambiguity is equivalent to saying that |A × B| =

6 |AB|.

22
Theorem (Product Lemma for Strings). Let A, B be sets of binary strings. If A, B are
unambiguous, then φAB (x) = φA (x) · φB (x).
Proof. In order to prove this theorem, we need to show that under the given assumptions,
|AB| = |A × B|, which we can show by showing there is a bijection between AB and A × B.

Let Sn be the set of strings of length n of AB, and let Tn be the set of pairs (a, b) ∈ A × B
such that length(a) + length(b) = n. Our claim is that f [(a, b)] = ab, f : Tn → Sn is a
bijection.

In order to prove that Tn → Sn is a bijection, two things must be proven. We have to

show that it is both onto and injective.
1. Onto. Every set in Sn is of the form ab where a ∈ A and b ∈ B such that ab = f [(a, b)].
2. Injective. Let (a1 , b1 ), (a2 , b2 ) ∈ Tn . Suppose:
f [(a1 , b1 )] = f [(a2 , b2 )]
| {z } | {z }
a1 b1 a2 b2

Then since A, B is unambiguous, (a1 , b1 ) = (a2 , b2 ).

Using the claim we just proved, |Sn | = |Tn |. By the product lemma (on generating functions
in general), we have:
|Sn | = |Tn | = [xn ]φA×B (x) = [xn ]φA (x) · φB (x)

Theorem (Sum Lemma for Strings). Let S be a set of binary strings. Let A, B be partitions
of S. Then φS (x) = φA (x) + φB (x).
The proof of the sum lemma for strings is the same as the proof of the sum lemma for
generating functions in general.
Definition. An expression is unambiguous if there exists a unique way of writing every
string according to the expression.
?
For example: {0, 10, 110} {, 1, 11} is the set of strings with no more than two consecutive
ones. Let’s examine the arbitrary string 0001101011001. If we split the string after every
zero, we get 0, 0, 0, 110, 10, 110, 0, 1, which is a unique way of breaking down the string.
1
Proposition. Suppose A? is unambiguous and  6∈ A. Then φA? (x) = 1−φA (x) .

Proof. We have A? = {} ∪ A ∪ A2 ∪ A3 ∪ A4 ∪ . . .. What is φAk (x)?

φAk (x) = [φA (x)]k by the product lemma for strings
= 1 + φA (x) + φA2 (x) + φA3 (x) + . . . by the sum lemma
= 1 + φA (x) + [φA (x)]2 + [φA (x)]3 + . . .
= 1 + y + y 2 + y 3 by letting y = φA (x)
1
=
1−y
1
=
1 − φA (x)

23
Note that the composition used in this proof is valid because  6∈ A, so the constant of φA (x)
is zero as required. That is, [x0 ]φA (x) = 0 makes the composition acceptable.

4.2 Counting Strings

← January 30, 2013
Example 4.7. Let S be the set of binary strings with no three consecutive ones (111). We
?
can express S as S = {0, 10, 110} {, 1, 11}. You can verify that this is an unambiguous
expression.

We have:

φ{0,10,110} (x) = x + x2 + x3
1
φ{0,10,110}? (x) = by proposition
1 − x − x2 − x3
φ{,1,11} (x) = 1 + x + x2

φS (x) = φ{0,10,110}? (x) · φ{,1,11} (x) by the product lemma for strings
1 + x + x2
=
1 − x − x2 − x3
Example 4.8. Let S be the set of binary strings where no block has length 2. Recall from
before:

S = ABC
A = {, 1, 111, 1111, . . .}
B = ({0, 000, 0000, . . .}{1, 111, 1111, . . .})?

For the sake of convenience, we’ll define f as follows.

f = x + x3 + x4 + x5 + . . .
= x + x2 + x4 + x5 + . . . − x2 by adding and removing x2
= x(1 + x + x2 + . . .) − x2
x
= − x2
1−x
Now, we can easily state the generating functions for A, B, and C:

φA (x) = 1 + f
1
φB (x) =
1 − f2
φC (x) = 1 + f

24
We can now use the product lemma on ABC to find φS (x).

φS (x) = φA (x) · φB (x) · φC (x)

(1 + f )2
=
1 − f2
(1 + f )(1 + f )
=
(1 + f )(1 − f )
1+f
=
1−f
..
.
1 − x2 + x3
=
1 − 2x + x2 − x3

4.2.1 Recursive String Definitions

There’s another way of constructing binary strings: recursive definitions.
← February 1, 2013
Example 4.9. Let S be the set of all binary strings. We can define S recursively as
S = {} ∪ {0, 1}S. Whenever you write an expression like this, you should take great care
to ensure that both sides are actually equal.

One might attempt to describe S as {, 0, 1}S, however that is ambiguous. S = {}∪{0, 1}S
is unambiguous. Since there are two terms of length 1, applying the sum & product lemmas,
we get:

φS (x) = φ{} (x) + φ{0,1} (x) · φS (x) by the sum lemma

= 1 + 2x · φS (x)
1
=
2−x
= 1 + 2x + 4x2 + 8x3 + . . . by composition

Note that [xn ]φS (x) = 2n because the number of strings with length n is 2n .
Example 4.10. Let S be the set of all binary strings with no three consecutive ones (111).

Consider the arbitrary string 110000100110100011 ∈ S. There are no three consecutive

ones in that string. We can prefix this string with 0, 10, or 110 while maintaining the con-
straint that the string cannot have three consecutive ones. With this recursive definition,
we’re building the string from right to left.

We can express S as S = {, 1, 11} ∪ S{0, 01, 011}.

In order to check this, we must verify that the two sides of the union are disjoint. That’s
easy to confirm: the elements on the left have no zeroes in them, and all elements on the

25
right have a zero. Therefore, the sum lemma applies.

φS (x) = φ{,1,11} (x) + φ{0,10,110} (x) · φS (x)

= 1 + x + x2 + (x + x2 + x3 ) · φS (x)
1 + x + x2
=
1 − x − x2 − x3
Example 4.11 (A harder problem). Let S be the set of all strings not containing 11010.
That is, 11010 is the forbidden string. Let F be the set of all strings ending in 11010, where
the entire string except the last character is an element in S. That is, F is the set of strings
where 11010 is at the end of the string but nowhere else in the string.

Idea:
1. Write two equations relating S and F . This is the trickiest part.
2. Write two equations relating φS (x) and φF (x).
3. Solve and get an expression for φS (x).
The first equation relating S and F is {} ∪ S{0, 1} = S ∪ F . We need to verify this by
showing both directions of the equality.

Showing {} ∪ S{0, 1} ⊆ S ∪ F : let a ∈ {} ∪ S{0, 1}. We need to show that a ∈ S or a ∈ F .

If a = , then a ∈ S since  ∈ S. If a = a0 0 or a = a0 1 where a0 ∈ S, then either:

• a does not have the forbidden string in it, so a ∈ S.
• a0 has the forbidden string at the end, since a0 ∈ S, so a ∈ F .
Showing {} ∪ S{0, 1} ⊇ S ∪ F : let a ∈ S ∪ F . We need to show that a ∈ {} ∪ S{0, 1}.

Suppose a ∈ S, or a = . Then clearly, a ∈ S{0, 1}. Suppose a ∈ F , then a ∈ S{0, 1}.

The second equation relating S and F is S{11010} = F . As before, we need to verify

this by showing both directions of the equality. ← February 4, 2013

Showing S{11010} ⊇ F : let a ∈ F . Then we know that the string a with its last character
truncated off is an element of S (it no longer can contain the forbidden string). Therefore,
the prefix to 11010 is also an element of S, as required.

Showing S{11010} ⊆ F : let a ∈ S{11010}. We need to show that the string ends in 11010,
and that the forbidden string does not occur elsewhere. But what if the forbidden string is
partially in the ∈ S part and partially in the 11010 part? Shifting the forbidden string to the
left by one character several times shows that it is not possible to construct the forbidden
string partially from the prefix and partially from the suffix (11010), for all shifts.

Now, let’s find the generating functions of both sides of these two equations.

The intersection of {} and S{0, 1} is empty, so the sum lemma applies. We’ll also apply the

26
product lemma to S{0, 1}. The generating function for {}∪S{0, 1} is therefore 1+φS (x)2x.
On the other side of equation 1, we get φF (x) + φS (x) by the sum lemma, since it’s a par-
tition (F only contains elements ending in the forbidden string, and S does not contain the
forbidden string at all). So, for the first equation we have 1 + φS (x)2x = φF (x) + φS (x).

By the product lemma, we get the generating function for S{11010} to be φS (x)x5 . The
other side of equation 2 has a generating function of φF (x). So, for equation 2 we have
φS (x)x5 = φF (x).

Next, we’ll substitute φF (x) from equation 2 into the generating function for equation 1:

1 + φS (x)2x = φS (x)x5 + φS (x)

..
.
1
φS (x) =
1 − 2x + x5
Example 4.12. Let S be the set of strings not containing 1010, and let F be the set of
strings that end in 1010, but if the last character of f ∈ F is removed, f ∈ S. We’ll apply
a similar method as in the previous example.

The first equation is {} ∪ S{0, 1} = S ∪ F , as before. There is nothing about this specific
forbidden string in this equation, so it applies generally.
?
The second equation we’d logically like to try is S{1010} = F . Let’s see if that works.
Let A ∈ S{1010}. If we’re unlucky, then the prefix of 1010 will be 10, which means the
forbidden string will occur with 10 in the prefix and 10 in the suffix (forming 1010). So,
S{1010} =6 F.

We could instead use S{1010} = F ∪ F {10}. Note that in the case where the forbid-
den string occurs partially in the prefix and partially in the suffix, we could say that is an
instance of a forbidden string followed by an additional 10, which is what this equation is ex-
pressing. We would also need to check the opposite direction, ∈ F ∪ F {10} =⇒ ∈ S{1010},
but we’ll omit that here.

Finding the generating functions, we get 1 + φS (x)2x = φS (x) + φF (x) (for equation 1)
and φS (x)x4 = φF (x) + φF (x)x2 (for equation 2), which gives us:

x4
φF (x) = · φS (x)
1 + x2
Substituting φF (x) into the generating function for equation 1 gives us:
−1
x4

φS (x) = 1 − 2x +
1 + x2
Example 4.13. Let S be the set of strings where there is no odd block of 0s that is imme-
diately followed by an odd block of 1s. For example, 010001011 6∈ S, but 111011000 ∈ S.

First, let A be the set of a block of zeroes followed by a block of ones. That is, strings

27
in A are composed of at least one 0 followed by at least one 1. We can express A by
? ?
A = {0}{0} {1}{1} .

Next, let B be the set of an odd block of zeroes followed by an odd block of ones. We
? ?
can express B as B = {0}{00} {1}{11} .

Next, let C be the set of a block of zeroes followed by a block of ones where both blocks
cannot be odd simultaneously (one can be odd, but not both of them). C can be expressed
as C = A \ B.
? ?
We then have that S = {1} (A \ B)? {0} .

5 Coefficients from Rational Functions

← February 6, 2013
We’re interested in finding [xn ] fg(x)
(x)
where f (x) and g(x) are polynomials.

Suppose deg(f ) ≥ deg(g), which implies f (x) = q(x)·g(x)+r(x), where q(x) is a polynomial
and r(x) is a polynomial of a lower degree than f . Dividing both sides by g(x) gives:

f (x) r(x)
= q(x) +
g(x) g(x)
f (x) r(x)
[xn ] = [xn ]q(x) + [xn ]
g(x) g(x)

Example 5.1.

6x4 − 5x3 − 5x2 1 − 5x

2
= (x2 − 1) + by long division of polynomials
1 − 5x + 6x 1 − 5x + 6x2
Now, we may assume that deg(f ) < deg(g). We will have to apply a partial fraction
expansion. For example:
1 − 5x 1 − 5x A B
2
= = +
1 − 5x + 6x (1 − 2x)(1 − 3x) 1 − 2x 1 − 3x
=⇒ 1 − 5x = A(1 − 3x) + B(1 − 2x)
1 − 5x = (A + B) + x(−3A − 2B)

We need A + B = 1 and −3A − 2B = −5. After solving this system of two equations in two
unknowns, you’ll get A = 3 and B = −2. So, we now have:
1 − 5x 1 1
= 3[xn ] −2[xn ]
1 − 5x + 6x2 1 − 2x 1 − 3x
| {z } | {z }
1+(2x)1 +(2x)2 +... 1+(3x)1 +(3x)2 +...
n n
=3·2 −2·3

We can generalize this result.

Generalize: we want [xn ] fg(x)

(x)
where deg(f ) < deg(g), where g(x) = (1 − a1 x)(1 −

28
a2 x) . . . (1 − ak x) where a1 , . . . , ak are distinct. We can only apply this partial fractions
method if all the roots are distinct (there are subtle differences if the roots aren’t distinct).
We know that:
f (x) c1 c2 ck
= + + ... +
g(x) 1 − a1 x 1 − a2 x 1 − ak x
1
However, note that 1−a1 x = 1 + (a1 x) + (a1 x)2 + . . .. So, we have:

f (x)
[xn ] = c1 an1 + c2 an2 + . . . + ck ank
g(x)

The previous example used this generalized form but with k = 2, c1 = A = 3, and c2 = B =
−2, which made a1 = 2 and a2 = 3 (from the denominators).
Example 5.2. Suppose we want to find the n-th coefficient of:

1 + 4x2
(?) =
1 − 6x + 12x2 − 8x3
1 + 4x2
=
1 + 3(−2x) + 3(−2x)2 + 1(−2x)3
1 + 4x2
= (binomial coefficients)
(1 − 2x)3

So, we’re interested in finding:

1 + 4x2
[xn ](?) = [xn ]
(1 − 2x)3
 
n A B C
= [x ] + +
1 − 2x (1 − 2x)2 (1 − 2x)3

After solving for A, B, and C, you’ll discover that A = 1, B = −2, and C = 2. Therefore,
we’re now looking for:
1 1 1
[xn ](?) = [xn ] − 2[xn ] + 2[xn ]
1 − 2x (1 − 2x)2 (1 − 2x)3
| {z } | {z } | {z }
(1) (2) (3)

1. 2n , as before.
2. Recall:
1 X n + p − 1 
= yn
(1 − y)p p−1
n≥0

Let y = 2x. Then we get:

   
n+2−1 n+1
2n = 2n = (n + 1)2n
2−1 1

29
 3. Consider y = 2x again. We get:
   
n+3−1 n+2 n
2n = 2
3−1 2

Putting (1), (2), and (3) together gives us:

   
n+2 n n+2
[xn ](?) = 2n − 2(n + 1)2n + 2 2 = 2n [1 − 2(n + 1) + ]
2 2
| {z }
degree 2

We can generalize this result.

Generalize: we’re interested in [xn ] fg(x)

(x)
where deg(f ) < deg(g), and where g(x) = (1−ax)k .
We know that:
 
n 1 n n+p−1
[x ] =a
(1 − ax)p p−1

Which gives us:

k  
n f (x) X n+p−1 n
[x ] = Cp a
g(x) p=1
p−1

(where Cp is the coefficient determined from the partial fraction expansion)

Remark:
f (x)
[xn ] = an × (polynomial of n of degree k − 1)
g(x)
← February 11, 2013

f (x) A B C
[xn ] 2
= [xn ] + [xn ] 2
+ [xn ]
(x − 3) (x + 2) (x − 3) (x − 3) x+2
= (polynomial of n of degree ≤ 1) · 3n + (constant) · (−2)n

(for f (x) of degree 2)

6 Solving Recurrence Relations

Example 6.1. We’re given:

(?) = an − 3an−1 + 2an−2 = 0 for n ≥ 2

a0 = 1
a1 = 4

Let’s take the coefficients of an , an−1 , and an−2 to guess:

(1 − 3x + 2x2 )(a0 + a1 x + a2 x2 + . . .) = c0 + c1 x

30
We want a polynomial of degree at most 1 on the right-hand side.

We know that [xn ]LHS = [xn ]RHS, so for n ≥ 2, [xn ]RHS = 0. Multiplying out, we
get (?). We can say that for n ≥ 2, (1 − 3x + 2x2 )(a0 + a1 x + . . .) = c0 + c1 x encodes (?).

Let’s continue:

(1 − 3x + 2x2 )(a0 + a1 x + a2 x2 + . . .) = c0 + c1 x from earlier

X c0 + c1
an xn =
1 − 3x + 2x2
n≥0
A B
= + by partial fraction expansion
1 − 2x 1 − x
This holds when the roots are distinct, otherwise the partial fraction expansion would work
out differently.

So, we have that an = A · 2n + B · 1n . We can solve for A and B using the given val-
ues for a0 and a1 .

a0 = 1 =⇒ A · 20 + B · 10 = 1
a1 = 4 =⇒ A · 21 + B · 11 = 4

=⇒ A = 3, B = −2

Substituting the values for A and B back into our equation from earlier, we get an = 3·2n −2.
Let’s now look at a more efficient (lazy) approach to this.

q(x) = 1 − 3x + 2x2
= (1 − 2x)(1 − 1x)

Now, we will define p(x) such that it is a polynomial with roots 2 and 1.
 
2 1
p(x) = x q
x
  
2 1 1
=x 1−2 1−
x x
= (x − 2)(x − 1)

Now, we’re just interested in the roots of p(x) instead of coefficients within the factored
form of q(x), which is much cleaner.
Definition. A homogeneous recurrence relation is an + c1 an−1 + c2 an−2 + . . . +
cm an−m = 0 with initial conditions a0 , a1 , . . . , am−1 (m initial terms).
Strategy:
1. Find the general form for an .
2. Use the initial conditions a0 , a1 , . . . , am−1 .

31
We have the characteristic polynomial:

p(x) = xm + c1 xm−1 + . . . + cm x0

Theorem. If p(x) has m distinct roots β1 , . . . , βm then the general form for an is:

A1 β1n + A2 β2n + . . . + Am βm
n

Example 6.2. We have the Fibonacci sequence an − an−1 − an−2 = 0 for n ≥ 2. The initial
conditions are a0 = 0 and a1 = 1.

We have:

p(x) = x2 − x − 1 = 0
√
1± 5
=⇒ x =
2
The general form is:
 √ n  √ n
1+ 5 1− 5
(?) = A1 + A2
2 2

Using the initial conditions a0 = 0 and a1 = 1, we get:

n = 0 =⇒ A1 + A2 = 0
 √   √ 
1+ 5 1− 5
n = 1 =⇒ A1 + A2 =1
2 2

1
=⇒ A1 = √
5
1
=⇒ A2 = − √
5
Substituting these values for A1 and A2 into (?) we get:
 √ n  √ n
1 1+ 5 1 1− 5
√ −√
5 2 5 2

=⇒ a20 = 4181

Example 6.3. We have the recurrence relation an + 3an−1 − 4an−3 = 0 for n ≥ 3. The
initial conditions are a0 = 0, a1 = 4, and a2 = 2.

We have:

p(x) = x3 + 3x2 − 4
= (x + 2)2 (x − 1)

32
Note that we cannot use the previous theorem here. Instead, let’s pick:
(1 + 3x − 4x3 )(a0 + a1 x + a2 x2 + . . .) = f (x)
(for n ≥ 3 where f (x) is of degree 2)

We know that [xn ]LHS = [xn ]RHS = 0, and we know that the LHS is an + 3an−1 − 4an−3 .
So, we get:
X f (x)
a n xn =
1 + 3x − 4x3
n≥0
f (x)
=
(1 + 2x)2 (1 − x)
We know that an = (A+Bn)·(−2)n +C·1n . Solving, we get that A = −2, B = 1, and C = 2. ← February 13, 2013

We do the same transformation as we did before.

q(x) = (1 + 2x)2 (1 − x)
 
3 1
p(x) = x q
x
 2  
1 1
= x3 1 + 2 1−
x x
= (x + 2)2 (x − 1)
Theorem. If we have:
an + c1 an−1 + c2 an−2 + . . . + cm an−m = 0 where cm 6= 0
The characteristic polynomial would be:
p(x) = xm + c1 xm−1 + c2 xm−2 + . . . + cm xm−m
= xm + c1 xm−1 + c2 xm−2 + . . . + cm
Pk
Suppose p(x) has root βi with multiplicity mi , for i = 1, . . . , k. That is, i=1 mi = m.
k
X
The general form for an is Q(mi )βin where Q is a polynomial of degree mi − 1.
i=1
This result generalizes the result from last time, except now the roots don’t need to be
distinct.
Example 6.4. We have the recurrence relation an + 3an−1 + 3an−2 + an−3 = 0 for n ≥ 3,
with initial conditions that we don’t care about. Find the general form.

First, we’ll write the characteristic polynomial, which will be of degree 3 since we have
an−3 included as one of the terms:
p(x) = x3 + 3x2 + 3x + 1 = (1 + x)3 by binomial theorem
There is one root of p(x), -1, with multiplicity 3. The general form of an is:
an = (A + Bn + Cn2 )(−1)n

33
6.1 Non-homogeneous Recurrences
Example 6.5. Suppose we have the recurrence relation (?1 ) = an + 3an−1 − 4an−3 = 6 × 2n
for n ≥ 3. The initial conditions are a0 = 5, a1 = 3, and a3 = 25.

Let’s ignore the initial conditions for now. Let’s take a guess and pick an = α2n , where α
is some constant that we’ll decide on later.

α2n + 3α2n−1 − 4α2n−3 = 6 × 2n

3 4
2n (α + α − α) = 6 × 2n
2 8
2α = 6
α=3

We still haven’t satisfied the initial conditions though. We want to find bn such that (?2 ) =
bn + 3bn−1 − 4bn−3 = 0 (which is the same relation as before, but = 0). Then:

(an + bn ) + 3(an−1 + bn−1 ) − 4(an−3 + bn−3 ) = 6 × 2n

an is a solution to (?1 ), ignoring the initial conditions. If bn is a solution to (?2 ), then an +bn
is a solution to (?1 ).

We previously showed the general form for (?2 ) to be:

bn = (A + Bn )(−2)n + C

Therefore, the general form for (?1 ) is:

an + bn = (A + Bn )(−2)n + C + 3 × 2n

Using the initial conditions, we can solve to get A = 1, B = 1, and C = 1.

Example 6.6. Suppose we have the recurrence relation an + an−1 − 2an−2 = 9, with initial
conditions a0 = 2 and a1 = 2.
?
We’ll guess that an = α (a constant). We’d then have that α + α − 2α = 9, but 0 6= 9, so
this won’t work for any value of α. It was a bad guess.

That guess didn’t work, so we’ll try a higher degree. Our second guess is an = αn for
some constant α, which would give us:

αn + α(n − 1) − 2α(n − 2) = 9
α=3

So, an = 3n will satisfy an + an−1 − 2an−2 = 9 for any n.

This concludes the counting portion of the course.

34
7 Introduction to Graph Theory
← February 15, 2013
Definition. A graph G is defined as G = (V, E), where V is a finite set of vertices and E
is a set of edges, in the form of a set of unordered pairs of V .
Example 7.1. Let’s say we have the graph G = (V, E), where V = {1, 2, 3, 4, 5} and
E = {12, 13, 14, 23, 34, 25, 45}. Note that for edges, we’re using the shorthand of xy to mean
{x, y} (for example: edge 12 = {1, 2}).

5
1 2

2 4
3

1 3
4 5

These are two different illustrations that both represent G.

Definition. The complement of the graph G, denoted G, is the graph with V (G) = V (G)
and edges uv ∈ E(G) if and only if uv 6∈ E(G).
Definition. Given the following graph:

i j

If ij ∈ E, we say i and j are endpoints of ij. We also say that i, j are adjacent, and that
edge ij is incident to i and j.

The neighbours of v ∈ V are defined as {u ∈ V |vu ∈ E}.

Example 7.2. Given the following graph:

a b
5
4 d c 6

e f g

1 3
2

The endpoints of f are 2 and 5. 1 and 2 are adjacent. c, d, and f are the edges incident to
5. The neighbours of 2 are 1, 3, and 5.
Example 7.3. If we were to represent a street map as a graph, V would be the set of
intersections and E would be the set of streets.

35
Example 7.4. If we were to represent a network of people, V would be a set of individuals
and E would be the various relationships between those people (sets of friends).
Example 7.5. Let V be the set of binary strings of length n.

uv ∈ E if u and v differ in exactly k positions. This is the definition of an n-cube. We have

n = 3 (since at most 3 differ) and k = 1. We get the following graph:

010 011

110 111

100 101

000 001

Example 7.6. If we have the following interval graph:

The overlaps are the edges. We can represent the same graph as:

1 2

4 3

36
7.1 Isomorphism
Are the following two graphs the same?

G1 , with vertices V = {1, 2, 3, 4}:

1 2

4 3

G2 , with vertices V = {a, b, c, d}:

a b

d c

Definition. G1 = (V1 , E1 ) and G2 = (V2 , E2 ) are isomorphic if there exists a bijection

φ : V1 → V2 such that uv ∈ E1 → φ(u)φ(v) ∈ E2 and uv 6∈ E1 → φ(u)φ(v) 6∈ E2 .

In other words, two graphs are isomorphic if adjacency is preserved.

Examining the two graphs above, we have:

φ(1) = a
φ(2) = b
φ(3) = c
φ(4) = d

Pairing 1 and 3, we get 13 ∈ E1 → φ(1)φ(3) = ac 6∈ E2 .

Pairing 4 and 3, we get 43 ∈ E1 → φ(4)φ(3) = dc ∈ E2 .

We must continue to consider these pairings for all edges.

Example 7.7. Are the following graphs isomorphic?

a f
1 4

d b
2 5

3 6 e c

37
Yes! There exists a bijection φ:

1→a
2→b
3→e
4→f
5→d
6→c

There is no general algorithm for finding a bijection. In this case, you could see that 1, 2,
and 3 have no edges and a, b, and e also have no edges.
Example 7.8. Are the following graphs isomorphic?

1 2 a c

5 e

4 3 b d

No. 1, 4, and 5 are pairwise adjacent but there are no corresponding vertices on the second
graph.

7.2 Degrees of Vertices

Definition. The degree of a vertex is the number of edges incident to that vertex.
Example 7.9. We have the following graph, with each vertex indicating its degree:

deg 3

deg 3

deg 3 deg 3
deg 2

Note that |E| = 7, and the sum of all degrees is 3 + 3 + 3 + 3 + 2 = 14.

X
Theorem. For any graph G = (V, E), (?) = deg(v) = 2|E|.
v∈V

Corollary (Handshake Lemma). There is an even number of vertices of odd degree.

38
Example 7.10. Given the following table:
Individuals # of friends among {a, . . . , i}
a 6
b 3
c 4
d 1
e 5
f 3
g 7
h 2
i 4
There are 5 odd numbers of friends. One of them must be lying.
← February 25, 2013

Example 7.11. Given the following graph:

b (deg 2)

d (deg 1)
a (deg 2) c (deg 3)

Let’s examine the degrees of a, b, c, and d:

deg(a) + deg(b) + deg(c) + deg(d)

= |{ab, ac}| + |{ab, bc}| + |{ac, bc, cd}| + |{cd}|

Note that each edge appears exactly twice in the above equation – once for each endpoint.
This is the general idea behind the proof for the previous theorem.
Proof. For each edge ab ∈ E, ab contributes 1 to (?) for deg(a) and ab contributes 1 to (?)
for deg(b).

Definition. A graph in which every vertex has degree k (for some fixed k) is called a
k-regular graph.
Definition. A complete graph is a graph in which all pairs of distinct vertices are adjacent
to each other. That is, each vertex is joined to every other vertex. The complete graph with
p vertices is denoted Kp (where p ≥ 1).

7.3 Bipartite Graphs

Definition. A graph G = (V, E) is bipartite if there exists a partition of V into V1 and
V2 such that every edge has one endpoint in V1 and one endpoint in V2 .
Definition. The complete bipartite graph Km,n is a bipartite graph with partitions A
and B such that all vertices in A are adjacent to the all the vertices in B, where |A| = m
and |B| = n.

39
Example 7.12. Is the following graph bipartite?
1 4

2 5

3 6

Yes, this graph is bipartite because V1 could be the vertices on the left-hand side (1, 2, and
3), and V2 could be the vertices on the right-hand side (4, 5, and 6).
Example 7.13. Is the following graph bipartite?

Yes, this graph is bipartite.

Not every graph is bipartite, however.
Example 7.14. Is the following graph bipartite?
?

No, this graph is not bipartite as shown. This is the simplest graph that is not bipartite, in
fact.
Example 7.15. Is the following graph bipartite, and is the partitioning unique?

Yes, this graph is bipartite, however its partitioning is not unique. We can flip the colours
(of both the squares of the graph) and still show the graph to be bipartite. So, in general,
the partitioning of V into V1 and V2 is not unique.

40
Definition. The n-cube is the graph G = (V, E) where V is the set of binary strings of
length n and uv is an edge if u and v differ in exactly one position.
For n = 2, we have the following n-cube:

10 11

00 01

Is the n-cube bipartite for any n? Yes!

Proof. Let V1 be the set of strings with an even number of ones, and let V2 be the set of
strings with an odd number of ones. Let uv ∈ E.

Note that we can split u and v each into two pieces, split on the single element (char-
acter) that differs between u and v. Corresponding pieces of u and v are the same.

The parity of the number of ones in u is different from the number of ones in v, since
aside from that single character difference, the parity of ones is the same between u and v.
Therefore, u and v are not both in V1 or both in V2 . So, the n-cube is bipartite.
Proposition. Let G = (V, E) be a bipartite graph with partition V = V1 , V2 . Then:
X X
deg(v) = deg(v) = |E|
v∈V1 v∈V2
| {z } | {z }
(1) (2)

Each edge has one endpoint in V1 and one endpoint in V2 , which is why this intuitively
makes sense. Let’s state this more formally.
Proof. Consider any edge ab ∈ E. Then ab contributes 1 to (1) and ab contributes 1 to
(2).

Corollary. If |V1 | = |V2 |, then the average degree of vertices in V1 is equal to the average
degree of vertices in V2 .

7.4 Walks, Cycles, and Paths

A walk is like a comet flying off track – they’re continuous lines.

start end

Definition. A walk W on graph G = (V, E) from v0 to vn is a sequence v0 e1 v1 e2 v2 e3 v3 . . . vn−1 en vn

for i = 1, . . . , n where edge ei has endpoints vi−1 and vi .

Definition. A path is a walk W in which v0 , v1 , . . . , vn are all distinct.

41
Definition. A cycle occurs if v0 , v1 , . . . , vn−1 are distinct but v0 = vn (the start and ending
vertices are the same).
Example 7.16. We have the following graph:
a 4 b

1 7

s 2 6 t

3 8

c 5 d

In this graph, we’ve visualized a path P (in orange) and a cycle C (in blue). We can express
P as P = s1a2c5d8t and we can express C similarly as C = b7t8d6b.

Proposition. Let G = (V, E) be a graph, and x, y ∈ V where x 6= y. If there exists a walk

from x to y, then there also exists a path from x to y.
To find the path from a walk, when you reach a vertex that is equal to another, you ignore
that cycle. ← February 27, 2013

Example 7.17. Let’s say you have the following graph:

x Vi Vj y

Heading from x to y, while going through the nodes in the triangle in the middle, is a walk.
However, if we were to skip visiting the Vi -th node entirely, we would have a path.
Proof. Among all walks from x to y, pick a walk W with as few edges as possible. We would
pick:

W = V0 e1 V1 e2 V2 . . . en Vn where V0 = x1 and Vn = y

We need to show that V0 , V1 , . . . , Vn are all distinct. Suppose for a contradiction that there
exists Vi = Vj for some i < j. That would mean:

W = V0 . . . ei Vi ei+1 . . . ej Vj ej+1 . . . Vn
| {z }
cut off
=⇒ W 0 = V0 . . . ei Vj ej+1 . . . Vn

W 0 is still a walk from x to y. But W 0 has fewer edges than W , which is a contradiction.
Corollary. If there exists a path from x to y and there exists a path from y to z, then there
exists a path from x to z.

42
This corollary is not as trivial to show as you might initially think. You can’t just join the
two paths, because the two paths may share some vertices.
Proof. The path from x to y followed by the path from y to z is a walk from x to z (not a
path). By the previous result, there exists a path from x to z.

Does every graph have a cycle?

This graph does clearly not have a cycle, but it’s not a terribly interesting graph. This more
complicated graph also does not have a cycle:

But what if every vertex has degree at least 2? In that case, it is not possible to create a
graph that does not have a cycle. In order for each vertex to have degree of at least 2, it
must loop back somewhere.

Proposition. Let G = (V, E). If for all v ∈ V , deg(v) ≥ 2, then G has a cycle.
Proof. Among all paths, pick a path P with as many edges as possible. We have:

P = V0 e1 V1 e2 . . . en Vn

Since deg(Vn ) ≥ 2, there exists Vn w 6= en . Since P is the longest path, w = Vi for some
i ∈ {0, . . . , n − 2}. Then we have the cycle C:

C = Vi ei Vi+1 . . . en Vn en+1 Vi

Note that it doesn’t necessarily have to loop back to the “first” vertex because you may have
a graph like this:

43
Definition. The girth of a graph G, denoted g(G), is the length of the shortest cycle in
G. If G has no cycles, then g(G) is infinite.

7.4.1 Hamiltonian Cycles

Definition. A cycle is Hamiltonian if every vertex is in the cycle.
You can think of a Hamiltonian cycle in terms of a graph of cities, where you have a person
who wants to visit all the cities once and then return to their initial city.
Example 7.18. Here’s an example of a Hamiltonian cycle:

Example 7.19. Here’s another example of a graph with a Hamiltonian cycle.

Do we always have a Hamiltonian cycle? No, not necessarily.

44
 |V |
Theorem (Dirac’s Theorem). Let G = (V, E). If for all v ∈ V , deg(v) ≥ 2 , then there
exists a Hamiltonian cycle.

7.5 Disjoint Paths

← March 1, 2013
Example 7.20. Given the following graph:
s

Is there a path from s to t? Yes, there are several.

Example 7.21. Given the following graph:
s

Is there a path from s to t? No. The two sets of vertices have no edges between them.
Definition. A pair of paths is edge disjoint if the paths do not share an edge.
Example 7.22. Given the following graph:

s t

Are there two edge disjoint paths from s to t? Yes. There is no edge that is both blue and
green, so the two paths are disjoint.
Example 7.23. Given the following graph:

45
 s

Are there two edge disjoint paths from s to t? No. For the same reason as the two examples
above, the center edge must be used in all paths because it is the only edge that connects
two otherwise disjoint sets of vertices.
Example 7.24. Given the following graph:
s

Are there three edge disjoint paths from s to t? Yes, as shown.

Example 7.25. Given the following graph:

s t

This is an example of a graph with two disjoint paths but not three. There are just two
edges connecting two otherwise disjoint sets of vertices.
Definition. Let G = (V, E), and let S ⊆ V, S 6= ∅, S 6= V (we pick a proper subset of
vertices). The cut induced by S is:

δ(S) := {uv ∈ E|u ∈ S, v 6∈ S}

We say it’s a cut because if we removed (cut) those edges, we would get two sets of edges
with no edges connecting the two sets, so the graph would fall apart.

Informally, δ(S) is the set of edges between V and V \S.

Example 7.26. Given the following graph:

46
 a

c d
s

t
b

If S = {s, a, b, c}, we get δ(S) = {cd}.

Example 7.27. Given the following graph:
b d

a f

c e

If S = {b, e}, we get δ(S) = {ab, bc, bd, ec, ed, ef }.

Definition. Let s, t ∈ V (s 6= t). If s ∈ S and t 6∈ S, then δ(S) is an st-cut.
In Example 7.25, we have a st-cut of size 2, which is why we cannot find three disjoint paths
from s to t.

7.5.1 Menger’s Theorem

Theorem. Let G = (V, E) be a graph, and s, t ∈ V (s 6= t). If there exists edge disjoint
paths P1 , P2 , . . . , Pk from s to t, then for all st-cuts δ(S), |δ(S)| ≥ k.

Equivalently, if you have an st-cut of size less than k, you cannot find k disjoint paths
from s to t.
Proof. Let δ(S) be an st-cut (s ∈ S, t 6∈ S). Let i ∈ {1, . . . , k}. Follow Pi from s to t. Let
vi be the last vertex in S of Pi , and let wi be the first vertex not in S of Pi . Then ei = vi wi
is an edge of Pi and ei ∈ δ(S). This implies {e1 , e2 , . . . , ek } ⊆ δ(S).

Since P1 , P2 , . . . , Pk are edge disjoint, e1 , e2 , . . . , ek are all different. This implies that
|δ(S)| ≥ k. Since δ(S) was an arbitrary cut, we have shown that all st-cuts have at least k
edges.
The converse is also true, which is known as Menger’s Theorem.
Theorem (Menger’s Theorem). Let G = (V, E) be a graph, and s, t ∈ V (s 6= t). If for all
st-cuts δ(S), |δ(S)| ≥ k, then there exist edge disjoint paths P1 , . . . , Pk from s to t.

47
Proof. The proof of Menger’s Theorem is by performing induction on |E| + |V |. The base
case is where you have two vertices and no edge between them, as seen here:

s t

There are then two cases:

1. There exists e1 and no st-cut that contains e1 has exactly k edges.

2. Say, k = 3, for instance. Divide the graph in two, and shrink the vertices from half the
graph into a single consolidated vertex. Then apply the same logic for the other half
of the graph. By induction, we can find three paths on both of the smaller graphs.
We want to combine these two smaller paths, which we can do by connecting the
corresponding paths of the two sub-graphs.

← March 4, 2013
Corollary. There exists a path from s to t (1) if and only if for all S ⊆ V, s ∈ S, δ(S) 6= ∅
(2).
Proof. Let us prove that (2) implies (1).

Let U = {s} ∪ {v|∃ a path from s to v}. If t ∈ U , we’re done.

Let’s assume t 6∈ U . By (2), there exists ab ∈ δ(v), say a ∈ U, b 6∈ U . That is, we

have at least one edge leaving U .

Informally, we know we have a path from s to a. We can take that path and add the
edge ab to get us a path from s to b. However, then b should be in U as well, because there
exists a path from s to b.

Formally, since a ∈ U , there exists a path Pa from s to a. Construct a path Pb by adding

edge ab at the end of Pa . This implies that b ∈ U , which is a contradiction.

7.5.2 Connectedness, Subgraphs, and Components

Definition. A graph is connected if there exists a path between any pair of vertices. If a
graph is not connected, it is disconnected.
Example 7.28. This graph is disconnected:

s t

Example 7.29. This graph (considered as a whole) is disconnected:

48
This corollary follows immediately from our characterization of st-cuts:
Corollary. G is connected (1) if and only if for all X ⊆ V, X 6= ∅, X 6= V, δ(X) 6= ∅ (2).
Definition. H = (VH , EH ) is a subgraph of G = (VG , EG ) if VH ⊆ VG and EH ⊆ EG .
Example 7.30. The highlighted portion of this graph is a subgraph:

Note that in order to keep an edge in a subgraph, we must keep both of its vertices. When
we discard a vertex, we must discard all edges that have that vertex as an endpoint.
Definition. A component of a graph is a maximal connected subgraph H of G. That is,
H is connected and no subgraph of G that properly contains H is connected.
Example 7.31. Given the following graph G (considered as a whole):

Each of these three obviously-disjoint subgraphs are components. The highlighted portion
is not itself a component (as expected), because it is not a maximal subgraph.
Corollary. G is connected if and only if G has a single component.

7.5.3 Equivalence Relations

Definition. ∼ is an equivalence relation for S (an infinite set) if all of the following hold:
1. For all a ∈ S, a ∼ a (reflexivity).
2. For all a, b ∈ S, a ∼ b =⇒ b ∼ a (symmetry).
3. For all a, b, c ∈ S, if a ∼ b and b ∼ c, then a ∼ c (transitivity).

49
Example 7.32. Let S be the set of all integers. Pick some integer m ≥ 1. A sample
equivalence relation would be a ∼ b if a − b is a multiple of m.
You can always partition your set S into equivalence classes.
Example 7.33. Let G = (V, E) be a graph. Given a, b ∈ V , a ∼ b if a = b or there exists
a path from a to b. We propose that ∼ is an equivalence relation, and we want to prove it.
Proof. We have to prove all three parts of the definition of equivalence relations.

(1) is true by definition.

(2) is true because if there exists a path from a to b, then there exists a path from b to
a.

(3) is true because if there exists a path from a to b and there exists a path from b to
c, then there exists a path from a to c (by theorem).

Therefore, we have shown that ∼ is indeed an equivalence relation.

Question: what do the equivalence classes correspond to? Any two vertices with one com-
ponent should have a path and one vertex from a component and one vertex from another
component should not have a path between them. The equivalence classes correspond to
the vertices of each of the components of G.

Exercise: look at set S (the set of all possible graphs). We propose that G ∼ H if G
is isomorphic to H. Prove that ∼ is an equivalence relation.

Remark: it’s useful to think of paths and cycles as sets of edges. Reflexivity makes a
lot more sense under this interpretation, for instance. Cycles become a set of edges of a
connected subgraph where every vertex has degree two. Paths are edges of a connected
subgraph where two vertices have degree one and every other vertex degree two.

7.5.4 Bridges
Definition. Let G = (V, E) be a graph, and e ∈ E. Then, G\e is the graph obtained from
G by deleting edge e. We say e is a bridge of G if G\e has more components than G.
Example 7.34. Given the following graph:

x y

The edge xy is a bridge, because if it was removed the number of components will increase
from one to two. If this edge was removed, x and y would be in different components. This
can be proven by transitivity.

50
 ← March 6, 2013
Proposition. If e = xy is a bridge of G, then x and y are in distinct components of G\e.
Proof. There exists ab such that a and b are in the same component of G, but a and b are
in distinct components of G\e. This implies that there exists a path P in G between a and
b that uses e.

Suppose for a contradiction that x and y are in the same component of G\e. We know:
• a and x are in the same component of G\e (because of P ).
• b and y are in the same component of G\e (because of P ).
By transitivity, a and b must be in the same component, which is a contradiction.
Proposition. Let G = (V, E) be a connected graph, and let e be a bridge of G. Then G\e
has exactly two components.
Proof. Let X be the vertices in the component of G\e that contains x. Let Y = V \X. We
need to show that any two vertices in Y are connected by a path in G\e.

Let a, b ∈ Y . Since G is connected, there exists a path Pa between a and x in G, and

there exists a path Pb between b and x in G. Moreover, Pa and Pb both use the edge e.
This implies:
• a and y are in the same component of G\e.
• b and y are in the same component of G\e.
By transitivity, a and b are in the same component of G\e.
Note that we didn’t just prove that there exists a path from a and b to y because we wanted
to ensure that the path did not use the edge e.
Proposition. Let G = (V, E) be a graph, and e ∈ E. e is not a bridge (1) if and only if e
is in a cycle (2).
In fact, this is the only reason you wouldn’t have a bridge.
Proof. Assume e = xy. We’ll start by proving that (1) implies (2).

The fact that e = xy is not a bridge implies that there exists a path P between x and
y in G\e. Then, P ∪ {e} is a cycle.

Now, we’ll prove (2) implies (1).

If e is in a cycle C, then C\e is a path between x and y. This implies that e is not a
bridge, otherwise x and y are in distinct components of G\e, as proven earlier.
Proposition. Let G = (V, E) be a graph. Let a, b ∈ V (where a 6= b). Suppose there exists
paths P1 , P2 between a and b where P1 and P2 are distinct. Then G has a cycle.
We want to show that there is a cycle that uses the edge e. So, we need to show that there
exists a path from x to y that does not use the edge e.

51
Proof. Since P1 is distinct from P2 , there exists an edge e = xy in P2 that is not in P1 .
After possibly interchanging a and b, we know:
• a and x are in the same component of G\e (because of P2 ).
• y and b are in the same component of G\e (because of P2 ).
• a and b are in the same component of G\e (because of P1 ).
Using those facts, by transitivity we know there exists a path P between x and y in G\e.
Then P ∪ {e} is a cycle.

8 Trees
Definition. A graph G is a tree if G is connected and G has no cycle.
Example 8.1. This graph is not a tree because it contains cycles:

Example 8.2. This graph (considered as a whole) is not a tree because it is disconnected:

Example 8.3. This graph is a tree because it is connected and has no cycles:

52
Example 8.4. This graph is also a tree:

Example 8.5. This is the simplest graph that is a tree:

Keep this graph (just one vertex with no edges) in mind when doing inductive proofs in-
volving trees.
← March 8, 2013
Proposition. If G is a tree, then:
1. Any two distinct vertices are joined by a unique path.
2. Every edge is a bridge.
Proof. We will prove each part of this proposition in turn.
1. Since the tree is connected, there exists at least one path. If there exists two distinct
paths between a pair of vertices, then there exists a cycle, which is a contradiction.
2. If an edge is not a bridge, then that edge is in a cycle, which is a contradiction.
Proposition. Let H = (V, E) be a tree. Let a, b ∈ V be two distinct vertices with no edge
between them. Let G be the graph obtained from H by adding the edge ab. Then G has
exactly one cycle.
Example 8.6. Let’s add the edge ab to the following tree:
a

According to the previous proposition, this graph should contain only one cycle.
Proof. C is a cycle of G if and only if C\{e} is a path of H between a and b. We know that
there exists a unique such path.
You could add an edge to create a cycle, then remove any edge from within the cycle to
get a different tree. This technique could be used by shipping companies to find the most
optimal cost-efficient route to ship a package.
Definition. Let G0 = (V 0 , E 0 ) be a subgraph of G = (V, E). G0 is a spanning subgraph
if V 0 = V .

53
Example 8.7. This is not a spanning subgraph because one vertex in V is not included in
V 0:

Example 8.8. This is a spanning subgraph because V 0 = V :

Proposition. Let G = (V, E) be a graph. G is connected (1) if and only if G contains a

spanning tree (2).
Example 8.9. This is a graph that is connected. Its spanning tree is highlighted:

Example 8.10. We have to specify that G contains a spanning tree because otherwise we
would say the following graph (considered as a whole) is connected (which it is not):

Proof. We’ll first prove that (2) implies (1). For all a, b ∈ V (a 6= b), there exists a path
between a and b in the spanning tree. But this is also a path of G.

54
Now, we’ll prove that (1) implies (2). Among all spanning subgraphs of G that are con-
nected, pick one with as few edges as possible. Call that subgraph T .

We claim that T is a tree. We need to verify that.

• T is clearly connected.

• T does not contain a cycle, otherwise any edge on that cycle could be removed while
still keeping the graph connected (edges in cycles are not bridges).

Let n be the number of vertices in a graph G. If n = 1, we must have zero edges. If n = 2,

we must have one edge. If n = 3, we must have two edges. If n = 4, we must have three
edges. If n = 5, we must have four edges. Let’s generalize this.
Proposition. In a tree, the number of edges is equal to the number of vertices minus one.
In order to prove this with induction, we want to start with the larger graph to see how we
built it up. We can’t do induction the normal way initially with graphs in general because
of how graphs can be constructed (there are multiple graphs we can form where n = 5 from
a graph where n = 4).
Proof. Let G = (V, E) be a tree. By induction on |V |:

Base case: |V | = 1, |E| = 0.

Assume |V | ≥ 2. We claim that there exists a vertex v of degree 1 (otherwise there ex-
ists a cycle). So, v is adjacent to a unique vertex w.

Let G0 = (V 0 , E 0 ) be the graph obtained from G by deleting the edge wv and the ver-
tex v. Note that G0 is a tree because the degree of v is one in G.

By induction, |E 0 | = |V 0 | − 1. We know that |E| = |E 0 | + 1 and |V | = |V 0 | + 1, which

gives us |E| = |V | − 1, as required.
If you’re ever asked to determine a result involving the number of edges in a tree, you’ll
likely use this proposition.

8.1 More on Bipartite Graphs

Proposition. If a graph has a cycle with an odd number of edges, then it is not bipartite.
← March 11, 2013
This graph is bipartite:

But this one is not:

55
 _
¨

Proof. Let C = V1 , V2 , . . . , V2n , where V1 = V2n .

Suppose there exists partition X, Y . We may assume V1 ∈ X, V2 ∈ Y, V3 ∈ X, . . .. This

implies that V2n ∈ Y since all vertices with even indices are members of Y .

Notice that V1 ∈ X and V2n ∈ Y , but V1 = V2n . This is a contradiction.

We want to show that this is the only reason why a graph would not be bipartite.
Proposition. A graph is bipartite if and only if all its components are bipartite.
The proof of this proposition is left as an exercise.
Proposition. Trees are bipartite.
For example, this tree is bipartite:

There are a couple of different proof techniques we could use to prove this proposition.
1. We could call one vertex V1 and then define even/odd vertices based on whether their
distance from V1 is even or odd (since there exists a unique path from V1 to any other
vertex in a tree).
2. More simplistically, we could build up the partition recursively. As we add an addi-
tional edge, each endpoint of the edge should be in the opposite partition.
We’ll use approach (2) to prove this proposition.
Proof. Let T = (V, E). We’re going to perform induction on |V |.

Base case: |V | = 1, i.e. V = {v}, set X = {v}, Y = ∅.

Otherwise, |V | ≥ 2. Since T is a tree, there exists vertex u of degree 1. Then, u is ad-

jacent to a unique vertex w.

Let T 0 be obtained by removing uv and u. Note that T 0 is still a tree. We can verify
that fact by seeing that we’re certainly not adding any cycles (not possible when removing

56
an edge) and since u has degree 1, the tree remains connected.

By induction, there exists a partition X 0 , Y 0 of the vertices of T 0 such that every edge
of T 0 has one end in X 0 and one end in Y 0 .

If w ∈ X 0 then X = X 0 and Y = Y 0 ∪ {u}. Otherwise (if w ∈ Y 0 ), X = X 0 ∪ {u}

and Y = Y 0 .
Theorem. G = (V, E) is bipartite (1) if and only if G has no odd cycles (2).
Proof. We proved NOT (2) implies NOT (1), which is equivalent to proving (1) implies (2).
Now, we’ll prove the other direction: (2) implies (1).

Due to the proposition from earlier, we may assume G is connected, without loss of gener-
ality. This implies that G contains a spanning tree T = (V, E 0 ).

We showed that T is bipartite. Let X, Y be the corresponding partition of V .

Claim: for all ab ∈ E, a ∈ X and b ∈ Y , or vice versa (a ∈ Y, b ∈ X). We may as-

sume ab ∈ E\E 0 .

Let P be the unique path in T between a and b. Suppose a, b ∈ X (or a, b ∈ Y ). Then P

has an even number of edges.

This implies that the cycle P ∪ {ab} is an odd cycle, which is a contradiction.

8.2 Graph Colouring

Problem 8.1. Assign colours to vertices such that every edge has endpoints of different
colours. The goal is to use as few colours as possible.
What graphs can be coloured with two colours? Easy: bipartite graphs.

What graphs can be coloured with three colours? No clue. There’s no efficient or reli-
able way of colouring an arbitrary graph with any constant (even very large) number of
colours.

9 Planar Graphs
We will allow:

• Loops:

• Parallel edges:

57
Definition. A planar embedding is a drawing of a graph where no two edges cross.

Example 9.1. This is not a planar embedding of a graph G because two edges cross:

1
5
2
4
3

This is a planar embedding for the same graph:

2 4

Observation: the same graph can have many different planar embeddings. We’ll formally
define “different” a bit later.
Example 9.2. Here are two planar embeddings of the same graph G1 :

1 5 1
5

2 4 6 2 6 4

7
3 7 3

Example 9.3. Here are two planar embeddings of the same graph G2 that look similar,
except the edges are labeled differently:

58
 5 2

1 2 1 5

4 6 3 6

3 4

Definition. A graph is planar if there exists a planar embedding.

Example 9.4. This graph is planar:

It is possible to find a planar embedding for this graph.

Example 9.5. These two graphs are not planar:

← March 13, 2013

This is a planar graph G:

59
G is a planar graph because a planar embedding exists for G. In particular, the following is
one planar embedding of G (no two edges cross):

inside

outside

Theorem (Jordan Curve Theorem). A closed, continuous, non-self-crossing curve on the

plane partitions the plane into two connected components.

F5

F4

F1 F2 F3

60
Definition. The Jordan Curve Theorem implies that a planar embedding partitions the
plane into connected components, called faces.

F3 : infinite face

F1

F2

Theorem. G has an embedding on the plane if and only if G has an embedding on the
sphere.
The truthfulness of this theorem is easy to see. To show an embedding on a plane implies
an embedding on the sphere, place your plane flat on the ground, then it’s a plane on a
sphere (Earth). For the opposite direction, just imagine a sphere so huge that the entire
graph appears on a portion of the sphere that appears like a flat plane.
Definition. For any face F , the boundary walk of F is the closed walk obtained by
following the edges bounding the face.
Suppose we have this graph G:

1 5
F4
2 F1 4 F2 7 F3 10

3 6

In G, we have the following boundary walks:

F1 : 1, 3, 2
F2 : 5, 7, 6
F3 : 8, 10, 9, 7 (in green above)
F4 : 1, 4, 5, 8, 10, 9, 6, 4, 3, 2

Note that in the boundary walk for F4 , the edge 4 appears twice. This is acceptable under
the usual definition of walks.
Definition. For a face F , deg(F ) is equal to the number of edges in the boundary walk of
F.

61
For example, we have deg(F1 ) = 3, deg(F2 ) = 3, deg(F3 ) = 4, and deg(F4 ) = 10. Note that
in deg(F4 ), we must count the edge “4” twice.

Notice:
4
X
deg(Fi ) = deg(F1 ) + deg(F2 ) + deg(F3 ) + deg(F4 )
i=1
= 20
= 2 × number of edges

This is similar to a theorem proven when we talked about degrees under the earlier definition
for degrees. That’s why we defined the degree of a face in the way we did.
X
Theorem. Consider a planar embedding of G = (V, E). Then (?) = deg(F ) = 2|E|.
F ∈faces

Why is this true? Each edge contributes to two faces, except edge “4” which is only in the
boundary walk of F4 . However, edge “4” occurs twice in that boundary walk so it’s still
counted twice. We’ll state this proof more formally now.

Proof. We want to show that every edge e is counted twice in (?). There are two cases.

Case 1: suppose e is not a bridge. Then e appears in exactly two different faces.

Case 2: suppose e is a bridge. Then e appears in one face but twice in the boundary
walk for that face.
Let p be the number of vertices, q be the number of edges, and f be the number of faces.
Here are G1 , G2 , G3 , and G4 :

G1 G2 G3

G4

In G1 , we have p = 6, q = 5, and f = 1. In G2 , we have p = 6, q = 6, and f = 2. In G3 , we

have p = 6, q = 7, and f = 3. In G4 , we have p = 6, q = 8, and f = 4.

G1 is a spanning tree, so we have p = q + 1. This implies that p − q + f = 2.

Theorem (Euler’s Formula). For any connected planar embedding of a graph G = (V, E),
we have p − q + f = 2, where p is the number of vertices, q is the number of edges, and f is
the number of faces.

62
Proof. We’ll prove this theorem by performing induction on q.

Base case: G is a tree. This implies that q = p − 1, as proven previously. Since G is a

tree, f = 1. So, we have p − q + f = 2.

Otherwise, let e ∈ E be an edge in a cycle (we can assume this because we aren’t in
our base case). Let G0 = G\e.

Note G0 has a planar embedding obtained by erasing edge e. G0 is connected because e

is not a bridge of G.

By induction, p0 − q 0 + f 0 = 2, where p0 is the number of vertices of G0 , q 0 is the num-

ber of edges of G0 , and f 0 is the number of faces in the embedding of G0 .

We know that p0 = p and q 0 = (q − 1), but we need to show that f 0 = (f − 1). f 0 = (f − 1)

follows from the Jordan Curve Theorem, because by adding another edge we’re adding an
additional boundary which creates a new face.
← March 15, 2013
Definition. A graph is simple if it has no parallel edges and it has no loops. A graph that
is not simple is called a multigraph.
Proposition. Let G be a simple planar graph. Then:
1. q ≤ 3p − 6.
2. q ≤ 2p − 4 when G has no triangle (no cycle of length 3).
Proof. We’ll prove each part of this proposition in turn.
1. The fact that G is planar implies that there exists a planar embedding with f faces.
Since the graph is simple, every face F must have deg(F ) ≥ 3. So, we know:
X
2q = deg(F )
| {z }
F ∈faces ≥3

=⇒ 2q ≥ 3f
≥ 3(2 − p + q)
≥ 6 − 3p + 3q
=⇒ 6 − 3p + q ≤ 0

Euler’s formula can be used as a general trick when we want to compare two of these
quantities.
2. Now, let’s also exclude triangles (cycles of length 3). We have:
X
2q = deg(F )
| {z }
F ∈faces ≥4

=⇒ 2q ≥ 4f
≥ 4(2 − p + q)
≥ 8 − 4p + 4q
=⇒ q ≤ 2p − 4

63
Let Kn denote the complete graph with n vertices. Is K5 planar?

Well, how many edges can a planar graph with 5 vertices have? Note that # of edges
≤ 3 × 5 − 6 = 9. But K5 has 52 = 10 edges. So, no, K5 is not planar.

Let Km,n denote the complete bipartite graph with m vertices and n vertices in each part
of the partition. Is K3,3 planar?

We have 9 ≤ 3 × 6 − 6, which does not yield a contradiction. However, K3,3 is bipartite, so

we know there’s no cycle of length 3. Applying (2), we get 9 ≤ 2 × 6 − 4 6≤ 8, which is a
contradiction. Therefore, K3,3 is not planar.

9.1 Platonic Graphs

Definition. A graph is platonic if it has a planar embedding such that every face has
degree l ≥ 3 and every vertex has degree k ≥ 3.

Theorem. There exist only five platonic graphs.

The platonic graph where k = 3, l = 3 is:

64
The platonic graph where k = 3, l = 4 is:

The platonic graph where k = 4, l = 3 is:

There are also two more graphs (k = 3, l = 5 and k = 5, l = 3), which we won’t show here
for simplicity.

9.2 Review of Planar Graphs

← March 18, 2013
• deg(f ) is the length of the boundary walk about the face f .
• Handshake Lemma: X
deg(v) = 2q
v∈V

• Dual Handshake Lemma: X

deg(f ) = 2q
f ∈faces

• Euler’s Formula:
p−q+s=1+c
(where p is the number of vertices, q is the number of edges, s is the number of faces
(earlier, we denoted this as f ), and c is the number of components.)

65
9.3 Planar Solids
There are solids that correspond to the five platonic graphs. What do they have in common?
Every vertex has the same degree and every face has the same degree.

We’ll now finally define platonic graphs formally.

Definition. A platonic graph is a graph such that there are integers k, l ≥ 3 such that
every vertex has degree k and every face has degree l.
Does the data of (k, l) determine the graph? Yes! When drawing these graphs, they are
unique because there are no ambiguous decisions to be made.

Which values of (k, l) are possible? If we try drawing the graphs for (4, 4) and (5, 4),
we’ll see that as we’re constructing them, they’d each need to be infinite in size to satisfy
the k and l constraints. Since these graphs cannot be finite, we say (4, 4) and (5, 4) are two
particular instances of (k, l) for which the constraints are not satisfied.

Idea: use the facts about planar graphs to constrain (k, l).
P
• By the handshake lemma, deg(v) = kp = 2q.
P 2q
• By the dual handshake lemma, deg(f ) = ls = 2q. This implies that p = k =s=
2q
l . This is known as the duality of graphs.

• By Euler’s formula, p − q + s = 2.

This gives us:

2q 2q
−q+ =2>0
k l
2 2 2
−1+ = >0
k l q
2 2
+ >1
k l
2l + 2k > kl
kl − 2k − 2l < 0
kl − 2k − 2l + 4 − 4 < 0
(k − 2)(l − 2) < 4

So, we know that k ≥ 3 and l ≥ 3. This gives us:

(k − 2)(l − 2) < 4
=⇒ k − 2, l − 2 ≥ 1 and k − 2, l − 2 < 4
=⇒ k, l ≤ 5

All of this gives us:

• If k = 3, then l can be 3, 4, or 5.
• If k = 4, then l can only be 3.

66
 • If k = 5, then l can only be 3.
Theorem. If G is a platonic graph, then (k, l) must be one of (3, 3), (3, 4), (3, 5), (4, 3),
or (5, 3).
Does each of these possibilities of (k, l) occur for a platonic graph? That is, we know these
are all possibilities, but are they all actually possible? For any choice of (k, l), how many
platonic graphs are there?

Claim: each of the five possibilities for (k, l) can be achieved by a platonic graph.
Proof. Just use the five platonic solids that we know.

• (3, 3): tetrahedron.

• (3, 4): cube.
• (3, 5): dodecahedron.
• (4, 3): octahedron.

• (5, 3): icosahedron.

Claim: the only platonic graphs are the five coming from platonic solids.
Proof. Just try to draw six platonic graphs. As you keep drawing, the graph will be uniquely
determined.

How many faces, edges, and vertices does each have?

 
2 2
q + +1 =2
k l
2q
p=
k
2q
s=
l

2
=⇒ q = 2 2
k −1
+ l
2kl
=
2l + 2k − kl

2q
=⇒ s =
l
4k
=
2l + 2k − kl
12
We can verify this. If (k, l) = (3, 5), then s = 10+6−15 = 12, as expected.

67
9.4 Non-Planar Graphs
← March 20, 2013

K5 K3,3

K5 and K3,3 are non-planar because they have edges that unavoidably must cross. Are
there any other planar graphs? Yes! You can just add more edges to K5 or K3,3 . We could
also add some vertices in the middle or edges, if we’d like.

The two following graphs are also non-planar, and are built by adding more edges and
vertices to K5 .

Definition. A graph H is a subdivision of a graph G if H is obtained from G by replacing

each edge by a path.
Example 9.6. Here’s a graph G and a subdivision of G.

In addition, the following is not a subdivision of G because a vertex was introduced where
there was previously a crossing:

68
Theorem. If G has a subgraph that is a subdivision of K5 or K3,3 , then G is not planar.
Example 9.7. Consider the following graph, G.

Katz claims he has a bridge in New York and some cheap land in Florida to sell you, and
he claims he has a planar embedding for G. Why must he be lying?

Essentially, we know he’s lying about having a planar embedding for G, since then he could
erase subdivisions and get a planar embedding for K3,3 . Let’s state that more formally.
Proof. Suppose G is planar. Then, take a planar embedding of G. Since H is a subgraph
of G, we get a planar embedding of H. Undo that subdivision. Then we get a planar
embedding of K3,3 or K5 , and that’s “unpossible.”

9.5 Kuratowski’s Theorem

Theorem (Kuratowski’s Theorem). A graph G is non-planar if and only if it contains a
subdivision of K3,3 or K5 .
Definition. The crossing number of a graph G is the minimum number of crossings in
drawings of G.
That is, there may be some ways to draw non-planar graphs that are smarter than other
ways. Smarter ways involve drawing the graph in such a way that the crossing number is
minimized.

Observation: a graph is planar if and only if it has crossing number zero.

It’s generally hard to compute crossing numbers. It’s an open question to d the crossing
number of Km,n , since we’d have to check all possible drawings.

9.6 Graph Colouring

Definition. Let n be a positive integer. A graph colouring is an assignment of one of
{1, 2, . . . , n} to each vertex such that no edge connects vertices with the same label. (In
other words, no two adjacent vertices are assigned the same colour.)

69
Example 9.8. This is K4 :

Since every vertex of K4 has an edge to every other vertex, every edge needs its own colour.
Note: Kn needs n colours. You must colour every vertex differently.

Note: Km,n needs two colours, because it’s a bipartite graph.

Theorem (Four Colour Theorem). Every planar graph is 4-colourable.
The four colour theorem is known to be true, but it’s not known to be true in a nice way.
The proof is basically a bunch of counterexamples. The proof also involved a computer to
generate various cases. ← March 22, 2013

We only consider simple graphs when discussing graph colouring. We can’t colour loops, for
instance.

We can expand on the four colour theorem and also introduce two very similar theorems.
Theorem (Five Colour Theorem). Every simple planar graph can be 5-coloured.
Theorem (Six Colour Theorem). Every simple planar graph can be 6-coloured.
We know that simple planar graphs cannot have too many edges. This implies that there
exists a vertex of low degree. We’ll use this information to derive a lemma.
Theorem. Let G = (V, E) be a simple graph. There exists a vertex v ∈ V of degree < k.
(k = 6)
Proof. Assume that for all v ∈ V , deg(v) ≥ k. We have:
X
2q = deg(v) ≥ pk
| {z }
v∈V ≥k

We also know that q ≤ 3p − 6, since G is a simple planar graph. That means 2q ≤ 6p − 12.
Combining these expressions, we get that pk ≤ 2q, so pk ≤ 6p − 12. We get a contradiction
for k ≥ 6.
So, we can update our lemma’s statement with k = 6.

Let’s now prove the six colour theorem.

70
Proof. Let G = (V, E) be a simple planar graph. By induction on k = |V |:

Base case: |V | = 1. You can colour this vertex any colour.

Assume for k ≥ 2 that we can 6-colour every simple planar graph with ≤ k − 1 vertices.

By the lemma we just introduced, there exists some v ∈ V such that deg(v) ≤ 5.

Let H = G\v be the graph obtained by deleting v from G. Note that H is a simple
planar graph. By induction, there exists a 6-colouring of H.

Since deg(v) ≤ 5, one colour, say c, does not appear among the vertices adjacent to v.
Pick colour c for v.
Before we can prove the five colour theorem, we need one more ingredient: contraction.
Definition. Let G = (V, E) be a graph, and e = uv ∈ E. Then G/e is the graph obtained
by identifying u and v in G and deleting e. We call this contracting the edge.
Example 9.9. Consider the following graph.

If we were to contract edge e, we would get the following graph:

When you apply contraction, you could introduce loops into the graph. Be careful.
Example 9.10. We’ll take a graph and then contract edge 5, then contract edge 6. Notice
that this introduces a loop into the graph.
1 1
5 1
2 4 2 4 6 2 4
6 3
3 3

71
Proposition. If G is planar, then G/e is planar.
Our goal is to 5-colour a planar graph. We can do this by taking a graph with five vertices,
removing a vertex, applying colours, then adding back the vertex that we removed.
Example 9.11. We’re going to 5-colour the following graph.

First, we’ll remove the vertex in the middle and colour the remaining vertices:

1 2

3 4

Then, we’ll put the middle vertex back and give it the remaining colour:

1 2

3 4

Example 9.12. Let’s now look at a slightly more complex graph – a graph with six vertices.

We’re going to contract the edges indicated in red, and then assign colours to the remaining
vertices.

72
 1

3 2

Finally, let’s uncontract and assign a colour to the re-introduced vertices.

4 1

5
4

3 2

Note that in this example, it could’ve been possible for a loop to be created, if the two
endpoints of the red path had an edge.
We know there must exist a pair of vertices that do not have an edge between them. Oth-
erwise, we’d have K5 , but we were told the graph is planar.

Let’s now prove the five colour theorem.

Proof. Let G = (V, E) be a simple planar graph. By induction on k = |V |:

Base case: |V | = k = 1. You can colour this vertex any colour.

Assume for k ≥ 2 that every simple planar graph with ≤ k − 1 vertices is 5-colourable.

Case 1 (there exists a vertex of degree ≤ 4):

Let H = G\v be the graph obtained by deleting v from G. Note that H is a simple
planar graph. By induction, there exists a 5-colouring of H.

Since deg(v) ≤ 5, one colour, say c, does not appear among the vertices adjacent to v.
Pick colour c for v.

Otherwise, for all v ∈ V , deg(v) ≥ 5. By lemma, we have Case 2.

Case 2 (there exists v ∈ V with deg(v) = 5):

We claim that there exists a, b ∈ V , adjacent to v, such that a, b are not adjacent (oth-
erwise, G contains K5 , so G would not be planar).

73
Let H = (G/av)/vb. Because of our claim, H has no loop. By induction, there exists
a 5-colouring of H.

Extend the colouring of H to G such that vertices a, b of G are assigned the same colour as
the vertex a = b = v in H.

Thus, we have coloured all vertices of G except v, and the same colour appears twice among
the neighbours of v. We can pick colour c, unused among the neighbours of v, to colour
v.

9.7 Dual Graphs

← March 25, 2013
Whenever you find a planar graph, you actually find two. You get two planar graphs for
the price of one!

Example 9.13. Here’s a planar graph, and its dual.

Here, we started with five faces, and created one new vertex per face (including the infinite
face). The faces in the original planar graph correspond to vertices in the dual (indicated
as the green graph, above). For each edge in the boundary walk of each face, draw an edge
between the two faces. We call this second graph the dual graph.
If we had a bridge in our original graph, we could create an edge with both endpoints being
the same, but the edge would have to cross the bridge visually. That is, we make an edge
that is a loop.
Definition. Let G = (V, E) be a graph with a planar embedding and faces F1 , F2 , . . . , Ff .
The dual graph G? is the graph where:
• Vertices of G? are F1 , F2 , . . . , Ff .

• If e is in the boundary of Fi , Fj in G, then e is joining Fi , Fj in G? .

• If e appears twice in the boundary walk Fj in G, then e is a loop of G? with endpoints
Fj .
We’re only going to define dual graphs for connected graphs, just so we don’t have to worry
about messy edge cases.

74
Remark: G? is planar, but G? need not be simple even if G is simple. Consider the case
when G is a triangle. Then G? is one edge connecting the face with the infinite face, as
well as two additional parallel edges, which prevents G? from being simple. Or, if G is just
two vertices joined by an edge, G? is just a single loop, which is not allowed in simple graphs.

Recall that if G has a planar embedding, then:

X
deg(f ) = 2q
f ∈faces

where q is the number of edges of G. We will now provide an alternate proof of this using
dual graphs.
Proof. Consider the graph G? . We have:
X
2q = deg(v)by the handshake lemma
v∈V (G? )
X
= deg(f )
f ∈faces(G)

Note that G?? = G, as long as the graph G is connected.

How do colourings work with dual graphs? Let’s look at the four colour theorem, stated in
a different way.

Theorem (Four Colour Theorem – Face Version). Let G be a graph with a planar embedding.
Assume G is connected and has no bridge. Then we can colour the faces of G with four
colours such that any two faces sharing an edge are assigned distinct colours.
Proof. Let G? be the dual of G. Since G has no bridge, G? has no loop. By the vertex
version of the four colour theorem, there exists a 4-colouring of the vertices of G? , which
gives a 4-colouring of the faces of G.
That’s all we’re going to discuss about planar graphs.

10 Matching
We will assume graphs are simple for the rest of the course.
Definition. Let G = (V, E) be a graph. The matching M of G is a subset of the edges
such that no two edges of M share an endpoint.
It’s trivial to see that ∅ is a matching, as is any single edge in the graph.
Example 10.1. The highlighted edges in the following graph form a matching because no
two edges share a vertex.

75
Our goal for the remainder of the course: we want to find an algorithm to produce a match-
ing with as many edges as possible. That’s a tough goal, so we will restrict ourselves to
finding a maximum matching in bipartite graphs only.

Ideally, the algorithm will be fast and easily verifiable. We need to be able to quickly
produce a result, and we need to be able to easily convince someone that it is indeed a
maximum matching.
Definition. A maximum matching is a matching with as many edges as possible in G.

Definition. A matching is said to be maximal if adding any additional edge will result in
a set that is no longer a matching.
← March 27, 2013
Example 10.2. Find a maximum matching (a matching with as many edges as possible)
for the following graph.

(The bottom-center edge is both green and red.)

The red edges represent a matching that is not maximum. The green edges represent a
maximum matching, as required.

10.1 Application: co-op

Let A be a set of applicants, and let J be a set of jobs. For all a ∈ A and j ∈ J, we know if
applicant a is qualified for job j.

Suppose A = {alice, bob, carol} and J = {1, 2, 3, 4}. We also have that:
• Alice is qualified for jobs 1 and 3.
• Bob is qualified for jobs 2 and 4.
• Carol is qualified for jobs 1, 2, and 4.

76
We want to assign applicants to jobs such that:
• No applicant is assigned two or more jobs.
• No job is assigned two or more applicants.

We want to represent this as a graph problem. Note that it’s a bipartite graph, as illustrated
here:

alice 2

bob 3

carol 4

The edges represent all cases where a particular applicant is considered to be qualified for a
job. The vertices are partitioned into applicants and jobs.

We want to assign as many applicants to jobs as possible. That’s equivalent to saying

we want to find a maximum matching.

10.2 Saturation
In order for us to move on, we need a few definitions.
Definition. Let G be a graph with matching M . Vertex v is saturated if v is the endpoint
of an edge of M . Otherwise, v is said to be unsaturated.

Example 10.3. Consider the following matching M , illustrated as green edges:

The green vertices are saturated.

Note that choosing a different matching may give you a different set of saturated vertices.
Definition. A matching is perfect is every vertex is saturated.

Example 10.4. The following matching is perfect:

77
Suppose you’re given Kn,m (a complete bipartite graph). How many perfect matchings are
there? Well, we have two cases:
1. n 6= m. Zero perfect matchings exist. One vertex will not have any corresponding
vertex in the other partition.
2. n = m. n! perfect matchings exist. For the first vertex, we have n choices, n − 1
choices for the second vertex, and so on.
Writing down every matching to determine a maximum matching is probably not a great
idea. Factorials become large very quickly.

We want to know how to make a matching bigger. Consider this graph:

s t

In this graph, the red edges are a smaller, non-perfect matching, and the green edges repre-
sent a perfect matching.

To get a larger matching, we need to remove something and add something – we can’t
just add an edge. In the red path, the first and last edges are not in the matching, and
they’re alternating. We can just shift our matching over to add that additional edge (which
is what the green matching represents). This will result in matching that is larger by one
edge.
Definition. Let G be a graph with matching M , and let P = e1 , e2 , . . . , ek be a path
between vertices s and t. P is alternating if for all i = 1, . . . , k − 1, exactly one of ei , ei+1
is in M .
Definition. P is augmenting if it is alternating and s and t are unsaturated.
Remark: let G be a graph, and let M be a matching of G. Let P be an augmenting
path. Let M 0 = M ∆P (where M ∆P = (M ∪ P )\(M ∩ P )). Then M 0 is a matching and
|M 0 | = |M | + 1.

78
10.2.1 Algorithm for Maximum Matching
← April 1, 2013
A high-level algorithm for finding a maximum matching is as follows.
M = ∅;
while (loop) do
if M is maximum then STOP;
else
find an augmenting path P ;
M := M ∆P ;
end
end
This algorithm is a nice idea, but it is not nearly precise enough – you couldn’t program it
currently. How do we know if M is maximum? We can’t simply enumerate all matchings,
that’d get out of hand quickly. Also, how do we find an augmenting path P ?

We will determine how to know if M is maximum, and how to find an augmenting path P ,
on bipartite graphs.

10.3 Vertex Covers

Informally, a vertex cover is a set of vertices where every edge in a graph has at least one
endpoint that is a vertex in that set.
Example 10.5. The vertices in green are a vertex cover of the graph:

Definition. C ⊇ V is a vertex cover if every edge has at least one endpoint in C.

Note that the easiest vertex cover is the cover where C = V . However, we’re interested in
picking a cover with as few vertices as possible. This is similar to how the empty set is a
matching, but we aren’t interested in that – we’re interested in matchings with as many
edges as possible.
Example 10.6. A minimum vertex cover is indicated in green:

79
The minimum vertex cover problem is where you aim to find a vertex cover with as
few vertices as possible.

Let’s examine the size of a graph’s maximum matching with the size of its minimum vertex
cover.
Example 10.7. Consider the following graph.

In this graph, |M | = 1 and |C| = 2.

Example 10.8. Consider the following graph.

In this graph, |M | = 4 and |C| = 4.

Example 10.9. Consider the following graph.

In this graph, |M | = 2 and |C| = 3.

80
We’ll use these results to introduce a lemma.
Theorem. Let M be a matching, and let C be a cover. Then |M | ≤ |C|.
To prove this lemma, just look at the edges in the matching. At least one vertex is needed
for every edge in the matching, and those vertices are all distinct. Let’s formalize this proof.
Proof. Let M = {e1 , . . . , ek } be a matching. Since C is a cover, for every edge ei , C contains
one endpoint, say wi , of ei .

Moreover, since M is a matching, w1 , . . . , wk are all distinct.

As C ⊇ {w1 , . . . , wk }, |C| ≥ k.
Now, how can we convince someone that a particular matching is a maximum matching?
We can note that there exists a cover of a certain size, which implies that no matching of a
greater size exists, according to the lemma we just introduced.
Theorem. Let M be a matching, and let C be a cover. Suppose |M | = |C|. Then:
1. M is a maximum matching.
2. C is a minimum cover.
Proof. For (1): let M 0 be any matching. We know that |M 0 | ≤ |C|, by the previous theorem,
and we know that |C| = |M | (by definition in this lemma). Therefore, |M 0 | ≤ |M |.

For (2): let C 0 be any cover. We know that |C 0 | ≥ |M |, by the previous theorem, and
we know that |M | = |C| (by definition in this theorem). Therefore, |C 0 | ≥ |C|.
Note that the second theorem we just introduced implies that max{|M | : M is a matching} ≤
min{|C| : C is a cover}. However, is this ≤, or is it always =? It turns out, if the graph
contains an odd cycle, then |M | =
6 |C|.
Theorem (König’s Theorem). In a bipartite graph, max{|M | : M is a matching} ≤ min{|C| :
C is a cover} holds with equality.
That is, König’s Theorem states that the size of the maximum matching of a bipartite graph
is equal to the size of the smallest cover of that graph.

Before we can discuss the proof of König’s Theorem, we’ll need to introduce some addi-
tional terminology.

10.4 XY-Structures
Let G = (V, E) be a bipartite graph with partition A, B of V . Let M be a matching of G.
Definition. A vertex v is special if:
1. v ∈ A and v is not saturated, or
2. There exists an alternating path P (v) where one endpoint is v and the other is of type
(1).

81
 ← April 3, 2013
Definition. Let X denote the special vertices in partition A, and let Y denote the special
vertices in partition B.

Let’s now prove König’s Theorem.

Proof. Let M be a maximum matching, and let C = (A\X) ∪ Y .

We aim to show:

1. C is a cover.
2. |C| = |M |.
Let’s start with (1). Suppose C is not a cover. Then there exists an edge uv with u ∈ X
and v ∈ B\Y .

Case 1: u is unsaturated. The edge uv implies that v must be special, which implies that
v ∈ Y , which is a contradiction.

Case 2: u is the endpoint of P (u). Then P (u) together with uv is an alternating path.
This implies that v is special, which means v ∈ Y , which is a contradiction.

We’ve proven (1). Let’s now prove (2).

For (2), we want to show that |C| = |M |. Intuitively, we want to show that |M | =
|Y | + |A\X|, since |Y | + |A\X| = |C|. In order to prove this, we really need to prove
these three things:

(i) Edges of M are between X and Y and between A\X and B\Y .
(ii) Vertices in A\X are saturated.
(iii) Vertices in Y are saturated.
Let’s start with (i). Suppose (i) does not hold. Then there exists an edge uv ∈ M where
u ∈ Y and v ∈ A\X. Since v is special, there exists an alternating path P (u). The alternat-
ing path obtained by adding uv to P (u) implies that v is special, which implies that v ∈ X,
which is a contradiction.

The proof for (ii) is very simple. (ii) holds because unsaturated vertices of A are in X.

Finally, let’s prove (iii). Suppose there exists u ∈ Y where u is not saturated. Then
P (u) is an alternating path. This is a contradiction since M is a maximum matching.
We’ll use a similar approach to that of this proof for the algorithm we need.

10.4.1 Algorithm For Finding a Maximum Matching and Minimal Vertex Cover
Input: a bipartite graph G = (V, E) with partition A, B.

82
Output: a matching M and a cover C where |M | = |C|.

M = ∅ (or any other matching);

while (loop) do
Construct X, Y ;
if there exists u ∈ Y where u is unsaturated then
M := M ∆P (u);
else
C = (A\X) ∪ Y is a cover;
|C| = |M |, STOP;
end
end
In this algorithm, S1 ∆S2 (for any sets S1 and S2 ) is defined as the symmetric difference of
S1 and S2 (the same way as earlier). That is, S1 ∆S2 := (S1 ∪ S2 )\(S1 ∩ S2 ).
← April 5, 2013
Example 10.10. Let’s say you’ve already run the algorithm for a few iterations, and you
have this bipartite graph:
A B

1 5

2 6

3 7

4 8

X is defined as all unsaturated vertices in A and all vertices accessible to such vertices via
an alternating path. The definition of Y is similar, but on partition B.

So, we have:
X Y X Y

2 5 1

4 8 3
7

We now have that 7 ∈ Y is unsaturated. P (7) = 48, 83, 37. We now have M 0 :
M 0 = {15, 38}∆{48, 38, 37} = {15, 48, 37}

83
We now have to repeat the process with our new matching M 0 .

A B

1 5

2 6

3 7

4 8

At this point we have:

X Y X

2 5 1

We have that X = {1, 2} and Y = {5}. Y contains only saturated vertices, so we stop. The
cover we found is:

C = (A\X) ∪ Y = {1, 2, 3, 4}\{1, 2} ∪ {5} = {3, 4, 5}

Take the time to verify that the cover is valid.

Example 10.11. Once again, suppose you have run the algorithm for a few iterations and
you currently have this:

1 6 3 8

5 2 7 4

This is a bipartite graph with partitions A = {1, 2, 3, 4} (red) and B = {5, 6, 7, 8} (green).

We have 4 ∈ A as an unsaturated vertex, so we get:

84
 X Y X Y X

8 3 6 1

7 2 5

We now have 5 ∈ Y which is an unsaturated vertex. P (5) = 47, 72, 25. We now have M 0 :

M 0 = M ∆P (5) = {16, 38, 27}\{47, 72, 25} = {16, 38, 47, 25}

We have to repeat the procedure on M 0 .

1 6 3 8

5 2 7 4

We now have that X = Y = ∅. There is no vertex in A that is unsaturated, so we should

be done. The cover we found is:

C = (A\X) ∪ Y = A = {1, 2, 3, 4}

10.5 Hall’s Theorem

Consider the following graph:

A B

Does this graph have a perfect matching? No, since there exists a cover of size 3 (indicated
in blue).

85
Let’s look at another reason we wouldn’t have a perfect matching. Look at the vertices
indicated in green. One of these vertices will not be covered by a matching.

We need at least the same number of vertices in N (S) (the neighbours of S), otherwise
a perfect matching is not possible.
Definition. Let G = (V, E) be a bipartite graph with partitions A and B. Let S ⊆ A.
Then N (S) = {v ∈ V |∃uv ∈ E, u ∈ S}. That is, N (S) denotes the neighbours of S.
Example 10.12. Consider the following graph.

A B

1 4

2 5

3 6

In this graph, if S = {1}, N (S) = {4, 5}. Similarly, if S = {2, 3}, N (S) = {4, 6}.
Proposition. Suppose the bipartite graph G = (V, E) has a perfect matching. Then:
1. |A| = |B|.
2. ∀S ⊆ A, |N (S) ≥ |S|.
((1) and (2) are necessary conditions.)
Proof. Let M be a perfect matching.
1. This is like a function, because each vertex in A should map to exactly one vertex in
B. More formally, M defines a bijection between A & B. Therefore, |A| = |B|.
2. Let S = {v1 , . . . , vk }. Since M is a perfect matching, there exist edges of M vi wi
for i = i, . . . , k. This implies that w1 , . . . , wk ∈ N (S), since wi , . . . , wk are distinct.
Therefore, |N (S)| ≥ k = |S|.
Theorem (Hall’s Theorem). A bipartite graph G = (V, E) with partitions A and B has a
perfect matching if and only if |A| = |B| and for all S ⊆ A, |N (S)| ≥ |S|.
← April 8, 2013
Proof. We proved the forwards direction of this statement in the previous proposition.

For the reverse direction, assume |A| = |B| and suppose there does not exist a perfect
matching. We need to show that there exists S ⊆ A such that |N (S)| < S (that is, we’re
aiming to prove the contrapositive).

Since there does not exist a perfect matching, we know that the size of the maximum

86
matching is less than |A|. By König’s theorem, we know there exists a cover C where
|C| < |M |.

Claim: N (A\C) ⊆ B ∩ C, since C is a cover. We have:

|N (A\C)| ≤ |B ∩ C|
≤ |B ∩ C| + |A ∩ C| −|A ∩ C|
| {z }
|C|

|C| − |A ∩ C| < |A| − |A ∩ C|

< |A\C| since |C| < |A| by König’s Theorem

Therefore, |N (A\C)| < |A\C|, as necessary.

We’ll now look at one application of this.

Corollary. Let G = (V, E) be a bipartite graph with partitions A, B. Suppose G is k-regular
(k ≥ 1). Then G has a perfect matching.
Proof. We’ll check the hypothesis of Hall’s Theorem.

1. Check that |A| = |B|. We have:

X X
k|A| = deg(v) = |E| = deg(v) = k|B|
v∈A v∈B

This implies that |A| = |B| by dividing by k (since k 6= 0).

2. Check that for every S ⊆ A, |N (S) ≥ |S|. We have:

X X
k|S| = deg(v) ≤ deg(v) = k|N (S)|
v∈S v∈N (S)

This implies that |S| ≤ |N (S)|.

Therefore, by Hall’s Theorem, G is a perfect matching.

87
Clicker Questions
← January 21, 2013
• How many bijections are there from {1, . . . , n} to {1, . . . , n}? n!, since you can’t map
two elements in the first set to the same one element in the second set.
• Let S be a set of objects with weight function w. Suppose that S = A ∪ B. Does
φS (x) = φA (x) + φB (x)? No, not always. The sum lemma only applies when A and
B are partitions of S.
• Let S = {(a, b)|a, b ≥ 1 ∈ Z}, w[(a, b)] = a − b. What is the generating function for
S? There is no generating function for S, since it is not a power series. There are an
infinite number of pairings – that is, there are an infinite number of objects with the
same weight.

• Does x2 − x4 + x6 − x8 + x1 0 − . . . have an inverse? No. There is no constant term

that isn’t equal to zero, therefore it does not have an inverse.
1
• Is a composition of A(x) = 1 + x + x2 + x3 + . . . and B(x) = 1−x a power series?
No. B(x) must have a constant term that is not zero in order for a composition to be
acceptable.
← January 30, 2013
2 3 4 5 x2
• What is the power series x + x + x + x + . . . equal to? It is equal to 1−x , since:

x2
 
1
x2 + x3 + x4 + x5 + . . . = x2 (1 + x + x2 + . . .) = x2 =
1−x 1−x

1
• Is A(B(x)) a power series for A(x) = 1 + x + x2 + . . . and B(x) = 1−x ? No! We can
0 1 2
only get a power series if [x ]B(x) = 0, but B(x) = 1−x = 1 + x + x + . . ..
1 9


• What is the coefficient of x3 for (1−x) 7 ? The coefficient is 2 . This could be found
using the formula:
 
1 n+k−1
[xn ] =
(1 − x)k k−1

• The number of ways of spending exactly $237 to buy oranges ($1 each) and/or mangos
($2 each) is. . . ? It’s [x237 ](1 + x + x2 + x3 + . . .)(1 + x2 + x4 + x6 + . . .).

You can represent this as a pair, (a, b), where a is the number of oranges and b is
the number of mangos. Define a weight function w[(a, b)] = a + 2b. We also have
S = {(a, b)}, A = {0, 1, . . .}, B = {0, 1, . . .}. Also, define weight functions for A and
B: w(a) = a, w(b) = 2b. The generating series are φA (x) = 1 + x + x2 + . . . and
φB (x) = 1 + x2 + x4 + . . .. Finally, use the product lemma on φA (x) · φB (x).
• The set of binary strings without consecutive zeroes can be described as. . . ? The set
can be described as ({, 0}{1})? {, 0}.

We have to make sure this expression does not generate anything that is not in the set.
Then we think about how to generate any arbitrary string contained in the set we’re
trying to express. There is a unique way to generate each string with the expression.

88
 ← February 27, 2013
• The general solution of the recurrence relation an − 4an−2 = 0 is given by an =
c1 2n + c2 (−2)n .

an−2 indicates that the degree of the characteristic polynomial is 2. So, we get that
p(x) = x2 − 4 = (x + 2)(x − 2), which has roots -2 and 2.
• Suppose an = (c1 n + c2 )3n . Then it is the general solution to the recurrence an −
6an−1 + 9an−2 = 0.

Note that 3 corresponds to the root, which appears with multiplicity 2. The char-
acteristic polynomial is p(x) = (x − 3)2 = x2 − 6x + 9.
• How many vertices are there in a 4-regular graph with 40 edges? 20.
P
Recall that v∈V deg(v) = 2|E| (by theorem), so we get:
X
deg(v) = 2|E|
v∈V
4|V | = 2|40|
4|V | = 80
|V | = 20

• Are the following two graphs isomorphic?

a b x y

c d z w

Yes! Here’s a bijection between their vertices:

a→x
d→y
c→z
b→w

• Consider this graph. Is W = a1c4d3a2b a walk, cycle, and/or path?

a 2 b

1 3 5

c 4 d

W is a walk but not a cycle nor a path.

89
 ← March 15, 2013
• Let T be a tree with 123 vertices. How many bridges does it have? 122, since no edge
is in a cycle in a tree, and the number of edges is equal to the number of vertices minus
one (in a tree).
• Let G be a graph with p vertices and q edges. Suppose G has 17 components and no
cycles. Then q = p + 17. Each of the 17 components is a tree, and every tree has one
more vertex than edges.
• Let G be a bipartite graph with partitions X, Y . Suppose G is k ≥ 1 regular. Is
|X| = |Y | true? Yes, always, since |X|k is equal to the number of edges, which equals
|Y |k.
• Is the following graph G planar?

Yes, G is planar. Here’s a planar embedding for G:

• Consider the following graph:

1 2

8 3

6 F 7 4

What is the boundary walk for F ? 7, 8, 8, 2, 1, 6. (Note that 8 is included twice.)

• Can 2, 2, 3, 3, 4, 4, 5, 6 be the degree sequence for the faces of a planar embedding?
No. There must be an even number of odd faces. ← March 27, 2013
• Is the following graph planar?

90
 No. There is a subdivision of K5 in this graph, which makes it non-planar.
• What is the minimum number of colours needed to colour this graph (not shown here)?
Four, because this was contained within the graph:

• Do the following embeddings have the same duals?

2 2
4
4
5 5
1 3 1 3
7
7
6 6

No. There are no boundary walks of length five in the first graph, but there are in the
second. The planar duals are not the same. Note that the dual is of the drawing, not
of the graph itself.
• How many faces does a connected planar graph with 30 vertices and 100 edges have?
By Euler’s formula, p − q + f 2, we get that f = 2 − p + q = 2 − 30 + 100 = 72.
• What is the minimum number of edges you need to remove from this graph to make
it planar?

91
 6


3. Note that the number of edges is 2 = 15. A planar graph with six vertices can
only have 3 · 6 − 6 = 12 edges.
• Do the red edges in the following form a matching?

Yes.

92