Western Michigan University

ScholarWorks at WMU

Honors Theses Lee Honors College

4-19-2017

The Regularity Lemma and its Applications

Elizabeth Sprangel
Western Michigan University, [email protected]

Follow this and additional works at: https://scholarworks.wmich.edu/honors_theses

Part of the Mathematics Commons

Recommended Citation
Sprangel, Elizabeth, "The Regularity Lemma and its Applications" (2017). Honors Theses. 2802.
https://scholarworks.wmich.edu/honors_theses/2802

This Honors Thesis-Open Access is brought to you for

free and open access by the Lee Honors College at
ScholarWorks at WMU. It has been accepted for inclusion
in Honors Theses by an authorized administrator of
ScholarWorks at WMU. For more information, please
contact [email protected].
 WESTERN MICHIGAN UNIVERSITY
LEE HONORS COLLEGE

The Regularity Lemma and

Its Applications

by
Elizabeth Sprangel

Committee:
Dr. Andrzej Dudek, Chair
Dr. Patrick Bennett, Member
Dr. Allen Schwenk, Member

May 2017
 Abstract
The regularity lemma (also known as Szemerédi’s Regularity Lemma) is one of the most
powerful tools used in extremal graph theory. In general, the lemma states that every
graph has some structure. That is, every graph can be partitioned into a finite number
of classes in a way such that the number of edges between any two parts is “regular.”
This thesis is an introduction to the regularity lemma through its proof and applications.
We demonstrate its applications to extremal graph theory, Ramsey theory, and number
theory.
Contents

Abstract i

1 Introduction 1

2 Szemerédi’s Regularity Lemma 3

2.1 Definitions, Notations, and Examples . . . . . . . . . . . . . . . . . . . . . 3
2.2 Szemerédi’s Regularity Lemma . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Other Forms of the Regularity Lemma 14

3.1 Degree Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Many Colors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Applications 16
4.1 Triangle Free Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Roth’s Theorem for 3-Term Arithmetic Progressions . . . . . . . . . . . . 18
4.3 Erdős and Stone’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4 Chvatál, Rödl, Szemerédi, and Trotter’s Theorem . . . . . . . . . . . . . . 21
4.5 Ramsey-Turán for K4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

Bibliography 26

ii
Chapter 1

Introduction

Combinatorics is a branch of mathematics that can be loosely defined as the study of

counting discrete structures. Many simple questions can be posed in combinatorics, in-
cluding enumerating how many ways to arrange a set of objects, or counting the number
of ways to choose a subset of a structure. Here are some introductory combinatorial
problems:

1. Given a team of 12 basketball players, how many ways can we assign these players
to 5 positions on the court?

2. How many ways can you put 6 balls in 14 boxes, where each box can have at most
one ball?

3. If there are 30 colors of paint, what is the total number of ways one can distribute
these colors to 9 rooms, ensuring that each room gets painted?

Although “counting” may sound easy, the level of difficulty rises quickly as the problems
add more variables and conditions. It is often found that the ideas and techniques of
combinatorics are being used not only in the traditional areas of mathematical appli-
cation, namely computer science and physical sciences, but also in the social sciences,
information theory and the biological sciences [4]. For example, determining the pre-
cise order of nucleotides within a DNA molecule, called DNA sequencing, is a typical
problem where combinatorial methods are successfully applied.

A graph is a set of points and a set of lines, where a line connects two points. It
is standard to call points vertices and lines edges. Extremal graph theory is a part
of combinatorics. It was first considered a topic to be studied as its own subject by
Paul Turán in 1940. Paul Erdős pioneered the subject through the problems, papers,
and lectures he produced [1]. Extremal graph theory problems can take two forms.
1
Contents 2

First, given a certain quality of a graph, such as the number of vertices or edges, what
properties does the graph have? Otherwise, given a property, how many edges or vertices
can a graph have while still forbidding the property [6]? A classic example of the first
form, proved by Mantel in 1907, is the following:
n2
Mantel’s Theorem (1907). Every graph with n vertices and at least 4 edges contains
a triangle.

This problem falls under the first form because given a certain number of vertices and
edges, a graph is guaranteed to have the property that it contains a triangle. Man-
tel’s theorem was generalized by Turán in 1941, bringing extremal graph theory to the
forefront of mathematical research.

Turán’s Theorem (1941). If G is a graph with n vertices that does not contain Kr ,
n2 (r−2)
then G has at most 2(r−1) edges.

Notice that this theorem was a solution to the second form of extremal graph problems.
Many problems in extremal graph theory are concerned with finding the maximum
number of edges a graph with n vertices can have while forbidding a certain type of
subgraph. One of the most successful techniques used in solving these kind of problems
is the so-called regularity method.
Chapter 2

Szemerédi’s Regularity Lemma

The regularity method, developed by Szemerédi, is a vital tool used in extremal graph
theory. This lemma first appeared as a part of a larger result by Szemerédi [11]. The
basic idea of the regularity lemma is that every graph has some structure, which can
be approximated as some collection of random graphs with uniformly distributed edges.
More precisely, all of the vertices of a graph G can be put into a finite number of classes.
Most of these classes are “regular,” meaning the number of edges between two subsets
of the classes is about the same as the number of edges between the whole classes [2].
The regularity method is of great importance when solving extremal problems, since it
provides structure to graphs that have seemingly little structure.

2.1 Definitions, Notations, and Examples

Before we can state and prove Szemerédi’s Regularity Lemma, we must first define a
few terms. First, we will define a few basic definitions in graph theory, and then what
it means for a graph to have “regular” parts.

A graph is a collection two sets G = (V, E) with V := V (G) denoting a set of vertices
and E := E(G) denoting a set of edges. Vertices can be thought of as nodes or points
and edges can be thought of as lines connecting the vertices. The elements of E are
2-element subsets of V , so E ⊆ [V ]2 . An edge e is typically denoted xy (or yx), where
{x, y} ∈ e. The following are a few examples of graphs:

3
Contents 4

2 3 2 4
b g
1 4
c d 1 3
a f 6 5
G H J

A vertex v is incident with an edge e if v ∈ e, or the edge has an endpoint v. Two

vertices x, y are adjacent if xy ∈ E(G). The vertex x is called a neighbor of y if the
two vertices are adjacent. For example, in graph H above, all vertices are adjacent to all
other vertices, so every vertex has every other vertex as a neighbor. When all vertices
are adjacent to one another, the graph is called complete and is denoted Kr , where
r = |V (G)|. So, H is K6 .

The degree of a vertex v, denoted d(v), is the number of neighbors v has. The mini-
mum degree of a graph G is δ(G) := min{d(v) : v ∈ V }, and the maximum degree
of G is ∆(G) := max{d(v) : v ∈ V }.

Now, we will define what it means for a graph to be “regular.” Let G = (V, E) be a graph
and X, Y be disjoint subsets of V . Let e(X, Y ) denote the number of edges between X
and Y . Then,
e(X, Y )
d(X, Y ) :=
|X| · |Y |
is the density of the pair (X, Y ). Notice that the density must always be a real number
between 0 and 1, since 0 ≤ e(X, Y ) ≤ |X| · |Y |.

A pair (A, B) of disjoint subsets of V is an ε-regular pair for some ε > 0 if for all X ⊆ A
with |X| ≥ ε|A| and Y ⊆ B with |Y | ≥ ε|B|, it follows that |d(X, Y ) − d(A, B)| < ε.

d(A, B)

d(X, Y )

X Y

A B
Contents 5

Notice that |X| ≥ ε|A| and |Y | ≥ ε|B|. This is because if we considered sets too small,
say |X| = 1 = |Y |, then the density of the pair (X, Y ) would be either 1 or 0, and thus
|d(X, Y ) − d(A, B)| = d(A, B), which might not be less than ε.

Let V = V0 ∪ V1 ∪ ... ∪ Vk be a partition of the vertices of graph G. This parition is an

ε-regular partition of G if the following are true:

(i) |V0 | ≤ ε|V |

|V |−ε|V |
(ii) |V1 | = |V2 | = ... = |Vk | = k

(iii) all but at most εk 2 of the pairs (Vi , Vj ) with 1 ≤ i < j ≤ k are ε-regular.

The set V0 is called the exceptional set. This set exists so that the other partition sets
can have the same number of vertices. These vertices are not considered when looking
at the density of G, since they are not ε-regular and there are so few.

Example. We will show that any ε-regular pair in a graph is also ε-regular in the
complement of the graph.

Consider the ε-regular pair (X, Y ) in a graph G and let A ⊆ X and B ⊆ Y with
|A| ≥ ε|X| and |B| ≥ ε|Y |. Let d(X, Y ) be the density between X and Y in G, and let
d(A, B) be the density between A and B in G. Also, let dc (X, Y ) be the density between
X and Y in Gc , and let dc (A, B) be the density between A and B in Gc .

Notice, dc (X, Y ) = 1 − d(X, Y ) and dc (A, B) = 1 − d(A, B). Now,

|dc (X, Y ) − dc (A, B)| = |1 − d(X, Y ) − 1 + d(A, B)| = |d(X, Y ) − d(A, B)| < ε.

So, (X, Y ) is an ε-regular pair in Gc .

Now that we have an idea of what regular means in the context of graphs, we can prove
the following proposition about ε-regular pairs. This proposition states that if a pair
(A, B) is a ε-regular pair, then most of the vertices in A are neighbors with almost the
density times the size of B number of vertices in B. We will use this proposition later
in the first application we consider.

Proposition 2.1. Let (A, B) be an ε-regular pair, of density d say, and let Y ⊆ B have
size |Y | ≥ ε|B|. Then all but fewer than ε|A| of the vertices in A have (each) at least
(d − ε)|Y | neighbors in Y .
Contents 6

Proof. Let X be the set of vertices in A with at most (d − ε)|Y | neighbors in Y . So,
e(X, Y ) < |X|(d − ε)|Y |. Now,

|X|(d − ε)|Y |
d(X, Y ) < = d − ε = d(A, B) − ε.
|X||Y |

Solving for ε yields

ε < d(A, B) − d(X, Y ) ≤ |d(A, B) − d(X, Y )|.

Since (A, B) is an ε-regular pair and |Y | ≥ ε|B|, it must follow that |X| < ε|A|. Other-
wise, (A, B) would not satisfy the definition of ε-regular. So, all but fewer than ε|A| of
the vertices in A have at least (d − ε)|Y | neighbors in Y .

2.2 Szemerédi’s Regularity Lemma

Szemerédi’s Regularity Lemma. For all ε > 0 and every integer m ≥ 1, there exists
an integer M such that for any graph G with |V (G)| ≥ m, there exists an ε-regular
partition P = {C0 , C1 , . . . , Ck } with m ≤ k ≤ M .

This lemma states that for any ε > 0, every graph has a partition into a bounded number
of ε-regular sets. If the number of sets was not bounded above by M , then any graph
could be trivially partitioned into singletons. This lemma is powerful for larger graphs.
We will now begin to prove the regularity lemma, using several preliminary lemmas.

Let G be a graph with |V (G)| = n. For disjoint sets A, B ⊆ V we define

|A||B| 2
q(A, B) := d (A, B).
n2

For partitions A of A and B of B we let

X
q(A, B) := q(A0 , B 0 ).
A0 ∈A,B 0 ∈B

For a partition P = {C1 , ..., Ck } of V (G) we let

X
q(P) := q(Ci , Cj ).
i<j
Contents 7

Notice, q(P) is bounded above by 1. This is because of the following:

X
q(P) = q(Ci , Cj )
i<j
X |Ci ||Cj |
= d2 (Ci , Cj )
n2
i<j
X |Ci ||Cj | e(Ci , Cj )2
=
n2 (|Ci ||Cj |)2
i<j
X 1 (|Ci ||Cj |)2
≤
n2 |Ci ||Cj |
i<j
1 X
= 2 |Ci ||Cj |
n
i<j

≤ 1.

Since every refinement of q increases by a constant and is bounded above by 1, there

must be a finite number of refinements for the partition to be ε-regular.

If P has εk 2 or more irregular pairs (Ci , Cj ) we may take the pairs (X, Y ) of subsets of
(Ci , Cj ) which violate the regularity and make X and Y into partition sets of their own.
Now, Lemma 2.2 will show that this refines P into a partition P 0 , where q(P 0 ) is greater
than q(P). To prove this, we will use the Cauchy-Schwarz Inequality.

Cauchy-Schwarz Inequality.
X X X 2
a2i b2i ≥ ai bi .

Lemma 2.2. (i) Let C, D ⊆ V be disjoint. If C is a partition of C and D is a partition

of D, then q(C, D) ≥ q(C, D).

(ii) If P, P 0 are partitions of V and P 0 refines P, then q(P 0 ) ≥ q(P).

Proof. (i) Let C = {C1 , . . . , Ck } and D = {D1 , . . . , D` }.

C1 D1
C2 D2

Ck D`

C D
Contents 8

By the definition of q(C, D) and q(Ci , Dj ),

X
q(C, D) = q(Ci , Dj )
i,j
1 X e(Ci , Dj )2
= .
n2 |Ci ||Dj |
i,j

p p
Letting ai = e(Ci , Dj )/ |Ci ||Dj | and bi = |Ci ||Dj |, the Cauchy-Schwartz In-
equality yields

!2  2
X q 2
X e(Ci , Dj ) X
p |Ci ||Dj | ≥  e(Ci , Dj )
i,j
|Ci ||Dj | i,j i,j
P 2
X e(Ci , Dj ) i,j e(C i , D j )
≥ P P .
|Ci ||Dj | i |Ci | j |Dj |
i,j

Notice that the sum of the edges between Ci and Dj for all i, j is the number of
edges between C and D. Also, the sum over i, j of |Ci | and |Dj | is |C| and |D|
respectively. Knowing this and the inequality we get from Cauchy-Schwartz, we
have
P 2
1 X e(Ci , Dj ) 2
1 i,j e(C i , D j )
≥ 2· P
n2
P
|Ci ||Dj | n i |Ci | j |Dj |
i,j
1 e(C, D)2
= ·
n2
P P
i |Ci | j |Dj |
e(C, D)2
=
n2 |C||D|
= q(C, D).

So, q(C, D) ≥ q(C, D), where C is a partition of C and D is a partition of D.

(ii) Let P := {C1 , . . . , Ck } and let Ci be the partition of Ci induced by P 0 , for i =

1, . . . , k.

C1
C2

Ci

Ck

V
Contents 9

Using the definition of q(P) and the inequality from part (i), we have
X
q(P) = q(Ci , Cj )
i<j
X
≤ q(Ci , Cj )
i<j

≤ q(P 0 )

Notice that the last inequality holds because q(P 0 ) is the sum over i, j of the
values of q within the part Ci and between parts Ci and Cj , so q(P 0 ) = i q(Ci ) +
P

0
P
i,j q(Ci , Cj ). Therefore, q(P ) ≥ q(P).

Next, we will prove that the index q(C, D), where C and D are refinements of C and D,
increases q(C, D) by at most ε4 |C||D|
n2
.

Lemma 2.3. Let ε > 0 and let C, D be disjoint subsets of V . If (C, D) is not ε-regular,
then there exists partitions C = {C1 , C2 } of C and D = {D1 , D2 } of D so that

|C||D|
q(C, D) ≥ q(C, D) + ε4 .
n2

Proof. Suppose (C, D) is not an ε-regular pair. So, there exists sets C1 ⊆ C and D1 ⊆ D
with |C1 | > ε|C| and |D1 | > ε|D| so that η =: |d(C1 , D1 )−d(C, D)| ≥ ε. Let C2 =: C\C1
and D2 =: D\D1 , and let C =: {C1 , C2 } and D =: {D1 , D2 }.

Now, define the following notation:

ci := |Ci | dj := |Dj | eij := e(Ci , Dj )

c := |C| d := |D| e := e(C, D).

We will now show that q(C, D) ≥ q(C, D) + ε4 |C||D|

n2
.

First, using the definition of q(C, D),

X
q(C, D) = q(Ci , Dj )
i,j

1 X e2ij
=
n2 cd
i,j i j
 
2 e 2
1 e X ij 
= 2  11 +
n c1 d1 ci dj
i+j>2
Contents 10

p p
Now, using the Cauchy-Schwartz Inequality with a = eij / ci dj and b = ci dj , we have
the following:
  !
2 2
P
1  e211
X eij
≥ 1 e211 i+j>2 eij
2
+ +P
n c1 d1 ci dj n2 c1 d1 i+j>2 ci dj
i+j>2
 2
(e − e11 )2

1 e11
= 2 + .
n c1 d1 cd − c1 d1

Now, we will solve η = d(C1 , D1 ) − d(C, D) for e11 :

η = d(C1 , D1 ) − d(C, D)
e11 e
η= −
c d cd
1 1 e 
e11 = η + c1 d1
cd
ec1 d1
e11 = ηc1 d1 + .
cd

Substituting this for e11 , he have

" #
e211 (e − e11 )2 (ηc1 d1 + (ec1 d1 )/(cd))2 (e − ηc1 d1 − (ec1 d1 )/(cd))2
 
1 1
+ = 2 +
n2 c1 d1 cd − c1 d1 n c1 d1 cd − c1 d1
1 η 2 c11 d21 + (2ηec21 d21 )/(cd) + (e2 c21 d21 )/(c2 d2 )

= 2
n c1 d1
#
(e(cd − c1 d1 )/(cd) − ηc1 d1 )2
+
cd − c1 d1
2ηec1 d1 e2 c1 d1

1
= 2 η 2 c1 d1 + + 2 2
n cd c d
e2 (cd−c1 d1 )2

2eηc1 d1 (cd−c1 d1 )
(c2 d2 )
− (cd) + η 2 c21 d21
+ 
cd − c1 d1

2ηec1 d1 e2 c1 d1 e2 (cd − c1 d1 )

1
= 2 η 2 c1 d1 + + 2 2 +
n cd c d c2 d2
η 2 c21 d21

2eηc1 d1
− +
cd cd − c1 d1
2 2η 2 c21 d21
 
1 e 2
= 2 0+ + η c1 d1 +
n cd cd − c1 d1
 2 
1 e
≥ 2 + η 2 c1 d1 .
n cd
Contents 11

Since |C1 | > ε|C|, |D1 | > ε|D|, and |η| > ε, we have

e2 e2
   
2 1
+ η c1 d1 ≥ 2 + η 2 · ε|C| · ε|D|
cd n cd
 2 
1 e 4
≥ 2 + ε cd
n cd
cd
= q(C, D) + ε4 2 .
n

So, we have the desired conclusion to the lemma.

Now, in the Key Lemma, we will show that if P 0 is a refinement of P, then the increase
of the index is bounded below by some constant which only depends on ε.

Key Lemma. Let 0 < ε ≤ 1/4 and let P = {C0 , C1 , . . . , Ck } be a partition of V

with exceptional set C0 with |C0 | ≤ εn and |C1 | = |C2 | = · · · = |Ck | := c. If P is
not ε-regular, then there exists P 0 = {C00 , C10 , C20 , . . . , C`0 } where k ≤ ` ≤ k4k , so that
|C00 | ≤ |C0 | + n/2k , |C10 | = |C20 | = · · · = |C`0 |, and q(P 0 ) ≥ q(P) + ε5 /2.

Proof. Assume P is not ε-regular. Then, at least εk 2 pairs (Ci , Cj ) are not ε-regular, so
we refine these pairs as follows. For each pair (Ci , Cj ) which is not ε-regular, Lemma 2
states that there exists partitions Cij of Ci and Cji of Cj so that q(Cij , Cji ) ≥ q(Ci , Cj )+
ε4 c2 /n2 . So, for each Ci , there exists a minimum parition, call it Ci , with |Ci | ≤ 2k−1 .
This is because for some Ci , each Cj such that (Ci , Cj ) is not ε-regular, Cij could parition
Ci into at most twice as many parts.

Let C0 =: {{v} | v ∈ C0 }, and let C := ∪ki=0 Ci . Notice,

|C0 | + |C1 | + · · · + |Ck | ≤ |C0 | + k2k−1 ≤ k2k .

Also,
X X X
q(C) = q(Ci , Cj ) + q(C0 , Ci ) + q(Ci ).
1≤i<j 1≤i 0≤i

Note that
X X
q(Ci ) = q(C0 ) + q(Ci )
0≤i 1≤i

≥ q(C0 ).
Contents 12

Since Ci is a partition of Cij , Cj is a partition of Cji , and Ci is a partition for {Ci },

Lemma 2(i) gives that
X X X X X
q(Ci , Cj ) + q(C0 , Ci ) + q(Ci ) ≥ q(Cij , Cji ) + q(C0 , {Ci }) + q(C0 )
1≤i<j 1≤i 0≤i 1≤i<j 1≤i
X  ε4 c2

≥ q(Ci , Cj ) +
n2
1≤i<j
X
+ q(C0 , {Ci }) + q(C0 )
1≤i
X
ε4 c2
≥ q(Ci , Cj ) + εk 2
n2
1≤i<j
X
+ q(C0 , {Ci }) + q(C0 )
1≤i
ε5 c2 k 2
= q(P) +
n2
ε 5
≥ q(P) + .
2

Now, we need to ensure that our refinement will have partitions of equal size. Let
C10 , C20 , . . . , C`0 be a collection of disjoint sets of size d := c/4k so that every Ci0 is
 

contained in some C ∈ C\{C0 } and C00 := V \ Ci0 . So, P 0 = {C00 , C10 , . . . , C`0 } is a
S

partition of V . Also, P 0 refines C, so Lemma 2.2(ii) yields q(P 0 ) ≥ q(C) ≥ q(P) + ε5 /2.

Since each Ci0 is also contained in one of C1 , . . . , Ck and no more than 4k sets Ci0 can be
contained in any one of the same Cj from the definition of d, it follows that k ≤ ` ≤ k4k .
So C10 , C20 , . . . , C`0 use all but at most d vertices for each set Ci 6= C0 in C. Thus,

|C00 | ≤ |C0 | + d|C|

c
= |C0 | + k |C|
4
c
≤ |C0 | + k k2k
4
ck
= |C0 | + k
2
n
≤ |C0 | + k
2

The last inequality holds since ck is all of the vertices except those in the exceptional
set. Therefore, P 0 satisfies the conditions in the lemma.

We are now set up to choose a value for M , and then apply the Key Lemma to any
graph until we obtain an ε-regular partition. The following is the proof of Szemerédi’s
Regularity Lemma.
Contents 13

Proof. Let G be a graph of orfer n with vertex set V . Let 0 < ε ≤ 1/4 and let m be a
positive integer. Let s := 2/ε5 be an upper bound on the number of applications of the
Key Lemma, and recall that q(P) ≤ 1.

Now, we must choose M . Consider the function f : x → x4x . Take M to be the

maximum of f 5 (k) and 2k/ε. If n ≤ M , then the partition of the graph into single
vertices satisfies the conditions, so let n > M . Let C0 ⊆ V be minimal such that k divides
|V \C0 |, and consider any partition {C1 , . . . , Ck } of the vertices where |Ci | = |V \C0 |/k.

We must now show |C0 | ≤ εn. Notice, |C0 | < k and the exceptional set grows by at
most n/2k each iteration of the Key Lemma. So, to obtain k + sn/2k ≤ ε/n, choose k
large enough so 2k−1 ≥ s/ε. Then,

ε s
≥ k
2 2
s ε
ε≥ k +
2 2
sn εn
εn ≥ k + .
2 2
2k
Recall that n > ε , since M cannot exceed this by definition and M is strictly larger
nε sn
than n. This implies that k < 2 , so it must follow that εn > 2k
+ k. This means that
|C0 | does not exceed εn. Now, we are set up to apply the Key Lemma until the partition
is ε-regular.
Chapter 3

Other Forms of the Regularity

Lemma

Since the regularity lemma is such a powerful tool used in combinatorics, it has been
adapted to many different forms to be better suited for different situtions. In this section,
we will introduce two alternate forms of the lemma.

3.1 Degree Form

The degree form of the regularity lemma is frequently used in problems for which having
more edges can only be beneficial. This form considers a subgraph G0 of the graph G
which we want to regularize. In G0 , there are no edges within any parts, all pairs are
ε-regular, and all pairs have either a density of 0 or a sufficiently large density. Typically,
the graph G0 is easier to work with, which is why degree form can be preferred when it
is applicable.

Degree Form of the Regularity Lemma. For every ε > 0, there exists M dependent
only on ε such that if G is any graph and d ∈ [0, 1], then there exists a partition
V = V0 ∪ V1 ∪ · · · ∪ Vk of the vertices of G and a subgraph G0 ⊂ G with the following
properties:

(i) k ≤ M ,

(ii) |V0 | ≤ ε|V |,

(iii) all parts Vi for i ≥ 1 are of the same size m ≤ dε|V |e,

(iv) degG0 (v) > degG (v) − (d + ε)|V | for all v ∈ V (G),

14
Contents 15

(v) e(G0 (Vi )) = 0 for all i ≥ 1,

(vi) all pairs G0 (Vi , Vj ) for 1 ≤ i < j ≤ k are ε-regular with each having density either
0 or greater than d.

This form of the regularity lemma can be derived from Szemerédi’s Regularity Lemma
by using a “cleaning procedure.”

3.2 Many Colors

In graph theory, it is common to consider coloring problems. In such problems, we will

colors the vertices or edges of a graph and study what properties such graphs can have.
Since it is common to work with colored graphs, there is a version of the regularity
lemma which accounts for graphs with colored edges.

The many colors regularity lemma states that any graph colored with r colors can be
partitioned in a way such that all but at most εk 2 pairs are epsilon regular for all r
colors. That is, if we only consider the red edges, then all but at most εk 2 pairs are
ε-regular with respect to the red edges. The following is a formal statement of this form
of the lemma.

Many Colors Regularity Lemma. For all ε ≥ 0 and integers r, m, there exists an
M such that if the edges of a graph G with n vertices are r-colored, then the vertex set
n
V (G) can be partitioned into sets V0 , V1 , ..., Vk , for some m ≤ k ≤ M , so that |Vi | = k
for every i ≥ 1, and for all but at most εk 2 pairs (Vi , Vj ), X ⊂ Vi and Y ⊂ Vj of size
|X|, |Y | > εn we have |dv (X, Y ) − dv (Vi , Vj )| < ε for v = (1, . . . , r).

The proof for the Many Colors Regularity Lemma is similar to the proof of the lemma
without colors. To account for the different colors, we use a different index which sums
over all the colors. If P = {V0 , V1 , . . . , Vk } is a partition of G and v ∈ (1, . . . r), then

k k
1 XX X 2
q(P) = dv (Vi , Vj )
k v
i=1 j=i+1

is the index used to prove the Many Colors Regularity Lemma.

Chapter 4

Applications

In this chapter, we will show how the regularity lemma can be applied to problems in
extremal combinatorics, Ramsey theory, and number theory. In the following applica-
tions, we use all three forms of the regularity lemma we have discussed thus far. First,
we will consider the problem of triangle free graphs, which is a straight-forward applica-
tion of the regularity lemma. Then, we illustrate how the regularity lemma can be used
to prove Roth’s theorem from number theory. Finally, we will look at a few results in
Ramsey theory and the study of Turán numbers.

4.1 Triangle Free Graphs

The first application of the regularity method we are going to consider is about triangle
free graphs. The theorem roughly states that if a graph contains at most some constant
times n3 triangles, then there exists at most a constant times n2 edges which can be
removed from the graph to make it triangle free.
Triangle Free Graphs Theorem. For every constant c > 0 there exists a constant
a > 0 with the following property. If G is any graph with n vertices that contains at
most an3 triangles, then it is possible to remove at most cn2 edges from G to make it
triangle-free.

Proof. Apply the Regularity Lemma with ε = c/4 and m = c obtaining a partition
V = {V0 , V1 , ..., Vk } of the vertices of G with all but at most cn/4 pairs (Vi , Vj ), where
0 < i < j ≤ k, are c/4-regular. Now, we remove all edges between pairs which fail to be
c/4-regular, which is at most
2
cn2

2c n − |V0 |
k ≤
4 k 4
16
Contents 17

edges. Now, we remove all of the edges between pairs which have density less than c/2,
which is at most
k c n − |V0 | 2 cn2
   
≤
2 2 k 4
edges. Finally, we remove all edges within each part. This is at most
 2
 n−|V0 |  n−|V0 |
k k
k ≤k·
2 2
kn2
≤
2k 2
n2
≤
2m
cn2
=
2

edges. So, in total we removed at most cn2 edges. Call this new subgraph with edges
removed G0 .

Now, we must show that G0 is triangle free. Assume, to the contrary, that G0 has a
triangle. In particular, G0 has a triangle with each vertex in a different part, say Vi ,
Vj , and V` . If there exists such a triangle, then (Vi , Vj ), (Vj , V` ), and (Vi , V` ) must be
c/4-regular and have density at least c/2. Since (Vi , Vj ) is a c/4-regular pair, all but
at most (c/4)|Vi | vertices in Vi have at least (c/2 − c/4)|Vj | neighbors in Vj , this set of
vertices in Vj be denoted by Vj0 . Similarly, all but at most (c/4)|Vi | vertices in Vi have
at least (c/2 − c/4)|V` | neighbors in V` , and call this set V`0 . Now, since c/2 − c/4 = c/4,
all but at most (c/4)|Vi0 | vertices in Vi have at least (c/2 − c/4)|V`0 | neighbors in V`0 .

So, Vi ∪ Vj ∪ Vk has at least the following number of triangles:

c 3 n − |V0 | 3 c 3 n − cn/4 3
       
2c  cc 2c  cc
1− 1− − ≥ 1− 1− −
4 4 2 4 k 4 4 2 4 k
c (1 − c/2)3 3
 
2c  c   c  3
= 1− 1− − n
4 4 2 4 k3

Now, since c and k do not depend on n, if

c 3 (1 − c/2)3
 
2c  cc
a= 1− 1− − −δ
4 4 2 4 k3

for some small δ, then G would contain less triangles than Vi ∪ Vj ∪ Vk , which is a
contradiction. Note that a is positive, since c < 1. If c ≥ 1, then it is trivial to remove
at most n2 edges to make G triangle free. So, G0 must be triangle free and thus it is
possible to remove at most cn2 edges from G to make it triangle free.
Contents 18

The converse of this statement is known as the Triangle Removal Lemma, since it states
that if it is necessary to remove at least cn2 edges from G to make it triangle-free, then
G contains at least an3 triangles. We will use the Triangle Removal Lemma in the
following application.

4.2 Roth’s Theorem for 3-Term Arithmetic Progressions

A k-term arithmetic progression, or an APk , is a sequence of k numbers such that

the difference of any successive numbers is constant. A subset A of the natural numbers
has positive upper density if
|A|
lim > 0.
n→∞ n

In 1936, Erdős and Turán conjectured that every subset of the natural numbers with
positive upper density contains arbitrarily long arithmetic progressions [8]. This conjec-
ture was proven by Szemerédi in 1975 [11]. His proof was one of the first applications
of the regularity lemma. The special case for AP3 of this theorem was first proven by
Klaus Roth in 1953 [9]. Roth’s proof for this theorem utilized analytic number theory.
In the following statement of the theorem, [n] will be used to denote the set {1, 2, . . . , n}.

Roth’s Theorem for 3-Term Arithmetic Progressions (1953). For all ε > 0 there
exists N (ε) such that for all n ≥ N any set S ⊆ [n] with |S| ≥ εn contains an AP3 .

Proof. Let S ⊆ [n] such that |S| ≥ εn, and assume that S contains no AP3 . We will
begin by constructing a 3-partite graph G = A∪B ∪C from S. Let V (A) = {a : a ∈ [n]},
V (B) = {b : b ∈ [2n]}, and V (C) = {c : c ∈ [3n]}. Now, define the edges in the following
way, where a ∈ V (A), b ∈ V (B), and c ∈ V (C):

E(G) ={ab : b − a ∈ S} ∪ {bc : c − b ∈ S} ∪ {ac : c − a ∈ 2S}

Now, the vertices a, b, and c form a triangle in G if and only if a + k = b, a + k + ` = c,

and k + ` ∈ 2S. For every vertex a and k ∈ S, there exists a triangle with vertices a,
a + k, and a + 2k. Notice that these triangles are pairwise edge-disjoint. So, at least
c ∈ V (C) vertices in G must be removed to make the graph triangle free. Since |S| ≥ εn,
|S|n ≥ εn2 . Hence, by the Triangle Removal Lemma, G contains at least γn3 triangles,
for some constant γ. Since there exists at least γn3 triangles, there must be a nontrivial
solution to a + b = c. This is a contradiction, so S must contain an AP3 .
Contents 19

4.3 Erdős and Stone’s Theorem

The following theorem published by Paul Erdős and Arthur Stone in 1946 [7] generalizes
Turán’s graph theorem to the extremal number of graphs containing no Ksr , which is
a complete r-partite graph where each part has s vertices. The Embedding Lemma,
Turán’s Theorem [12], and Lemma 4.1 will be used to prove the Erdős-Stone theorem.
Embedding Lemma. For all d ∈ (0, 1] and ∆ ≥ 1, there exists ε0 > 0 with the property
that if G is any graph, H is a graph with ∆(H) ≤ ∆, s ∈ N, and R is any regularity
2s
graph of G with parameters ε ≤ ε0 , ` ≥ d∆
, and d, then H ⊆ Rs implies H ⊆ G.
Turán’s Theorem (1941). For all integers r, n with r > 1, every graph G which is
r−2 n2
K r free with n vertices has at most tr−1 (n) = r−1 · 2 edges.

Lemma 4.1.  −1

n r−2
lim tr−1 (n) = .
n→∞ 2 r−1

In the following proof, we will use the regularity lemma to create a regularity graph.
Then, using Turán’s theorem, we will show that the regularized graph contains a com-
plete graph on r vertices as a subgraph. If such a subgraph exists in the regularized
graph, then the embedding lemma will give us that Ksr is contained in the original graph,
as desired.
Erdős-Stone Theorem (1946). For every ε > 0 and integers r ≥ 2 and s ≥ 1, there
exists integer n0 so that every graph G with n ≥ n0 vertices and at least tr−1 (n) + εn2
edges contains Ksr as a subgraph.

Proof. Let r ≥ 2 be given. If s = 1, then Turán’s Theorem yeilds that K r ⊆ G. So let

s ≥ 2 and set γ > 0. Let G be a graph of order n. Notice that if |E(G)| ≥ tr−1 (n) + γn2
then γ < 1 because if γ ≥ 1, then |E(G)| > n2 .

1
Let d := γ and ∆ := ∆(Ksr ). The Embedding Lemma gives an ε0 > 0. Now, let m > γ
γ
and ε > 0 where ε ≤ ε0 , ε < 2 < 1, and

δ := 2γ − ε2 − 4ε − d − 1/m > 0.

1
This is possible because m > γ implies that

1
2γ − d − > 2γ − γ − γ = 0.
m

Given ε and m, Szemerdi’s Regularity Lemma outputs an integer M . Let

2M s 2M s
n> = ∆ .
d∆ (1− ε) γ (1 − ε)
Contents 20

Notice that this is at least m since δ, 1−ε < 1. So Szemerdi’s Regularity Lemma gives an
ε-regular partition {V0 , V1 , ..., Vk } of G where m ≤ k ≤ M and |V1 | = |V2 | = ... = |Vk | = `
for some `. Notice that n ≥ k`, since k is the number of ε-regular partitions and ` is the
number of verticies in each set Vi with i ∈ (1, ..., k). Also, notice that

n − |V0 |
`=
k
n − εn
≥
M 
1−ε
=n
M
 
2M s 1−ε
≥ ∆
d (1 − ε) M
2s
= ∆.
d

Let R be the regularity graph of G with parameters ε, `, and d. Given these parameters,
the Embedding Lemma yields that if K r ⊆ R, then Ksr ⊆ G. To show that K r ⊆ R, we
will first show that R has more than tr−1 (k) edges. Then, Turn’s Theorem will give us
K r ⊆ R.

We will count the number of edges in G to show that there is a sufficient number. First,
within V0 there are at most |V20 | ≤ 12 (εn)2 edges. Between V0 and Vi , for i ∈ (1, ..., k),

there are at most |V0 |k` ≤ εnk` edges. There are at most εk 2 pairs (Vi , Vj ), for i, j ∈
(1, ..., k), which are not ε-regular. Between such pairs there is at most `2 edges, so
between all pairs which are not ε-regular, there are at most εk 2 `2 edges. The ε-regular
pairs with density less than d have at most d`2 edges between them, so altogether there
are at most 21 k 2 d`2 edges between pairs with insufficient density in G. Within a set Vi
there are trivially at most 2` ≤ 12 `2 k edges. The rest of the edges are between ε-regular

pairs with density at least d, which are also edges in Rs . Every edge in R corresponds
to at most `2 edges in G, so there are at most |E(R)|`2 edges of this kind. By adding it
all, we get

1 1 1
|E(G)| ≤ (εn)2 + εnk` + εk 2 `2 + k 2 d`2 + `2 k + |E(R)|`2 .
2 2 2
Contents 21

Solving for |E(R)| gives

|E(G)| − 12 (εn)2 + εnk` + εk 2 `2 + 12 k 2 d`2 + 21 `2 k

|E(R)| ≥
`2 !
1 2 |E(G)| − 21 (εn)2 − εnk` − εk 2 `2 − 12 k 2 d`2 − 21 `2 k
= k 1 2 2
2 2k `
!
1 2 tr−1 (n) + γn2 − 12 (εn)2 − εnk` 1
≥ k 1 2 − 2ε − d −
2 2n
k
!
1 tr−1 (n) 1
≥ k2 1 2 + 2γ − ε2 − 4ε − d −
2 2 n m
 −1   !
1 2 n 1
= k tr−1 (n) 1− +δ .
2 2 n

Lemma 1 can be applied to obtain

 −1   !  
1 2 n 1 1 2 r−2
k tr−1 (n) 1− +δ > k
2 2 n 2 r−1

≥ tr−1 (k).

By Turán’s Theorem, K r ⊆ R, and so Ksr ⊆ Rs . By the Embedding Lemma, Ksr ⊆ G.

4.4 Chvatál, Rödl, Szemerédi, and Trotter’s Theorem

The following theorem from Chvatál, Rödl, Szemerédi, and Trotter from 1983 [5] is
a significant result in Ramsey theory. It states that for every graph with bounded
maximum degree has a diagonal Ramsey number which is linear in the order of the
graph. The diagonal Ramsey number Rr (H) is the maximum number n such that there
exists an r-edge-coloring of the complete graph Kn which contains no monochromatic
copy of H as a subgraph. To prove this theorem, we will utilize Turán’s Theorem and
the Many Colors Form of the Regularity Lemma.

Chvatál, Rödl, Szemerédi, and Trotter (1983). For every positive integer ∆ and
r ≥ 2, there is a c(∆, r) such that Rr (H) ≤ c|H| for all graphs H with ∆(H) ≤ ∆.

1
Proof. Let ∆ ≥ 1 and r ≥ 2 be given, and define m := Rr (K∆+1 ). For d := r and ∆,
1 1
the Embedding Lemma gives ε0 . Choose ε ≤ ε0 be small enough so that 2ε < m−1 −m .
Contents 22

Notice that ε < 1. Now, let M be given by the Many Colors Regularity Lemma with
inputs ε and m. Everything defined thus far depended only on ∆ and r, so choose
2M r∆
c := 1−ε .

Let H be a graph with ∆(H) ≤ ∆ and s := |H|, and let G be a graph with order
n ≥ c|H|. The Many Colors Regularity Lemma gives that G has an ε-regular partition
{V0 , V1 , ..., Vk } with exceptional set V0 and |V1 | = ... = |Vk | = l where m ≤ k ≤ M and
|dv (X, Y ) − dv (Vi , Vj )| < ε for v = 1, ..., r. Notice,

n − |V0 | n(1 − ε) 1−ε 2s 2s

l= ≥ ≥ cs = 1 ∆ = ∆.
k M M (r) d

Let R be the regularity graph of G with parameters ε, l, and 0. By definition, R has k

vertices. Observe that
 
k
|E(R)| ≥ − εk 2
2
 
1 2 1
= k 1 − − 2ε
2 k
 
1 2 1 1 1
> k 1− − +
2 k 1−m m
 
1 1 1 1
> k2 1 − − +
2 k 1−m k
 
1 2 m−2
= k
2 m−1
≥ t(k, m − 1).

By Turan’s Theorem, R has Km as a subgraph.

Now, color the edges of R by coloring an edge corresponding to a pair with color v ∈
1
{1, ..., r} where dv (Vi , Vj ) ≥ r. That is, the edge density in the v-th color is greater
than or equal to density d. For each color, let Rv denote the subgraph of R induced
by the v-th color (with parameters ε, l, and 1r ). Since m := Rr (K∆+1 ), Km contains a
monochromatic copy of K∆+1 , say the v-th color. Since χ(H) ≤ ∆ + 1, H ⊂ (K∆+1 )s ,
and thus H ⊂ (Rv )s .

2s
Since ε ≤ ε0 , l ≥ d∆
, and d ∈ (0, 1], H ⊂ G by the Embedding Lemma.

4.5 Ramsey-Turán for K4

The study of Ramsey-Turán numbers is motivated by Ramsey theory and the study of
Turán numbers. Given integers n and s, the Ramsey-Turán number RT (n, H, s) is the
Contents 23

maximum number of edges in a graph G with n vertices and independence number at

most s so that G does not contain H as a subgraph. The following theorem was proven
by Szemerédi in 1972 [10]. The basic idea of this theorem is if G is a graph with no K4
as a subgraph and at most o(n) independent vertices, then G has less than 1/8n2 + o(n2 )
edges. In the proof, we will use the term cluster graph. Let G be a graph with partition
V = V0 ∪ V1 ∪ · · · ∪ Vk , and let d be a number between 0 and 1. The cluster graph R
of G has k vertices corresponding to the k parts of V , and there exists the edge ij if the
corresponding parts Vi and Vj have density at least d.

Ramsey-Turán for K4 , Szemerédi (1972). Let G be a graph with n vertices with

no K4 as a subgraph and only o(n) independent vertices. Then, |E(G)| < 18 n2 + o(n2 ).

Proof. Let G be a graph with n vertices. Assume, to the contrary, that |E(G)| >
1 2 ε2 M (ε)
8n + 4εn2 . Also, let α(G) ≤ M (ε) n − 1 and n ≥ ε , for ε > 0 and M (ε) obtained
from applying Degree Form to G with the parameter d = 2ε. Let G0 denote the graph
G in Degree Form, and let G00 = G0 − V0 be the usual pure graph. Also, let R be the
cluster graph of G00 . Notice,

n2
E(G00 ) > E(G) − (d + 3ε) .
2

Notice, by the assumption of the size of E(G),

1 n2
E(G00 ) > n2 + 4εn2 − (d + 3ε)
8 2
1 2 n2
= n + 4εn2 − (2ε + 3ε)
8 2
1 2 3 2
= n + εn
8 2
1 2
> n + εn2 .
8

Now, use the fact that k ≤ M (ε) to simplify α(G):

 
n2
α(G) < ε −1
M (ε)
hn i
≤ ε2 −1
k 
2 n
<ε
k
2
≤ε m

Proceed with two cases:

k2
Case 1. Suppose there are more than 4 edges in R. By Turán’s Theorem, R contains a
triangle. Let the vertices of said triangle in R correspond to the clusters Vi0 , Vj0 , and V`0
Contents 24

in G00 . Let H be the graph consisting of the clusters A, B, and C in G, corresponding

to Vi0 , Vj0 , and V`0 respectively with the edges within the clusters put back.

By Proposition 2.1, all but at most ε|A| vertices of A have at least ε|B| neighbors in
|B| and all but at most ε|A| vertices of A have at least ε|C| neighbors in |C|. So, all
but at most 2ε|A| vertices in A have at least ε|B| neighbors in |B| and ε|C| neighbors
in |C|. Choose a to be one of these vertices in A. Let the neighborhood of a in B
and C be called B 0 and C 0 accordingly. Since (B, C) is an ε-regular pair, (B 0 , C 0 ) is an
ε-regular pair. So, all but at most ε|B 0 | = ε2 |B| vertices in B 0 have at least ε|C 0 | = ε2 |C|
neighbors in C 0 . So, let b be such a vertex in B 0 . Let the neighborhood of b in C 0 be
called C 00 . Since α(G) < ε2 m = ε2 |C|, there must be two adjacent vertices in C 00 , call
them c1 and c2 . Now, there must exist a K4 , consiting of vertices a, b, c1 , and c2 .
k2
Case 2. Suppose there are at most 4 edges in R. Notice,

X E(G00 )
d(Vi , Vj ) =
m2
1≤i<j≤k
k2
≥ E(G00 ) 2
 n
1
> + ε k2 .
8

Note that the last inequality follows from the observation that E(G) > ( 81 + ε)n2 . So,
the average of the densities which are not zero is at least:

(1/8 + ε)k 2 1
d= = + 4ε.
k 2 /4 2

Since the average is at least d, there is at least one pair (Vi , Vj ) which has density
greater than d. Let H be the graph consisting of this regular pair with the edges inside
the two clusters put back, call this pair (A, B). Note that |A| = |B| = m. Notice that
1
α(H) ≤ α(G) < εm < 6εm. Let β = d − ε = 2 + 3ε, so α(H) < (2β − 1)m. Notice
that since α(H) < εm, there are at most εm independent vertices in A,so there exists at
1
least m − εm adjacent vertices in A. Also notice that since d ≥ 2 + 4ε, all but at most
1
εm vertices in A are adjacent to at least 2 + 4ε vertices in B. Choose two such vertices
in A, and call them a1 and a2 . Now, these vertices must have at least 4εm of the same
neighbors in B, call this set B 0 . Now, since α(H) < εm, at most εm vertices in B 0 are
non-adjacent. So, there are at least two vertices in B 0 which are adjacent, call them b1
and b2 . Now, there exists a K4 as a subgraph of G, with vertices a1 , a2 , b1 , and b2 .

Since in both cases there exists a K4 as a subgraph of G, there must be less than
1 2
8n + o(n2 ) edges in G.
Contents 25

This result was proven to be optimal by Bollobás and Erdős in 1976 [3]. To prove this,
they constructed a graph with o(n) independent vertices and 81 n2 − o(n2 ) edges.
Bibliography

[1] B. Bollobás, Extremal Graph Theory, Academic Press, London, 1978.

[2] , Modern Graph Theory, Springer, New York, 1998.

[3] B. Bollobás and P. Erdős, On a Ramsey-Turán type problem, J. Combin. Theory

Ser. B 21 (1976), no. 2, 166–168.

[4] B. Chen, What is combinatorics?, available online: http://www.math.ust.hk/

~mabfchen/Math232/Introduction.pdf.

[5] C. Chvatál, V. Rödl, E. Szemerédi, and W.T. Trotter, The ramsey number of a
graph with bounded maximum degree, Journal of Combinatorial Theory 34 (1983),
239–243.

[6] R. Diestel, Graph Theory, third ed., Springer, New York, 2006.

[7] P. Erdős and A.H. Stone, On the structure of linear graphs., Bulletin of the Amer-
ican Mathematical Society 52 (1946), 1087–1091.

[8] P. Erdős and P. Turán, On some sequences of integers, Journal of the London
Mathematical Society 11 (1936), 261–264.

[9] K. Roth, On certain sets of integers, Journal of the London Mathematical Soci-
ety 28 (1953), 101–109.

[10] E. Szemerédi, On graphs containing no complete subgraph with 4 vertices, Mat.

Lapok 23 (1972), 113–116 (1973).

[11] , On sets of integers containing no k elements in arithmetic progression,

Acta Arithmetica 27 (1975), 199–245, Collection of articles in memory of Juriı̆
Vladimirovič Linnik.

[12] P. Turán, An extremal problem in graph theory, Matematikai és Fizikai Lapok 48
(1941), 436–452 (in Hungarian).

26