THE REGULARITY LEMMA AND GRAPH THEORY

ARIEL HAFFTKA
Abstract. The Szemeredi Regularity Lemma states that any suciently large
graph G can be partitioned into a bounded (independent of the size of the
graph) number of regular, or random-looking, components. The resulting
partition can be viewed as a regularity graph R. The Key Lemma shows
that under certain conditions, the existence of a subgraph H in R implies its
existence in G. We prove the Regularity Lemma and the Key Lemma.
Contents
1. Introduction 1
2. Denitions, Examples, and Statement of the Regularity Lemma 2
3. Proof of the Regularity Lemma 5
4. The Key Lemma 14
5. Acknowledgements 20
References 20
1. Introduction
In 1975 Endre Szemeredi proved a 40 year old Ramsey Theory type conjecture
of Erdos and Turan: every set of natural numbers of positive upper density con-
tains arithmetic progressions of arbitrary length. Szemeredi began his proof with
a lemma on bipartite graphs [1]. In 1976, he generalized this result to simple
graphs [2], a result which became known as the Regularity Lemma. The Regularity
Lemma states that any graph can almost entirely be partitioned into a bounded
number of components, such that the edge density between any two of these com-
ponents is nearly random, or -regular, as will be dened in Section 2. Notably,
for a given lower bound m and , the upper bound M = M(m, ) given by the
Regularity Lemma depends only on m and and holds for arbitrarily large graphs.
One application of the Regularity Lemma is in proving that certain conditions
on a large graph G (such high enough edge density) can force the existence of a
given subgraph H. The partition resulting from the Regularity Lemma gives rise to
a regularity graph R whose vertices are the components of the partition and where
two components are connected by an edge if and only if the edge density between
the two components in G is suciently high. The Key Lemma states that under
the right conditions, the existence of a bounded degree subgraph H in a certain
generalized regularity graph (see Section 4) implies its existence in G.
Date: August 22, 2008.
1
2 ARIEL HAFFTKA
A
B
Figure 1. A complete (3, 4)-bipartite graph.
In this paper we prove the Regularity Lemma and the Key Lemma. We shall
draw main ideas, lemma statements, and some notations from [3]. The illustrations,
examples, and most of the explanations are original.
2. Definitions, Examples, and Statement of the Regularity Lemma
Let G = (V, E) be a simple graph on n vertices, i.e., a connected, undirected
graph with no loops.
Denition 2.1. (Edge density) For disjoint A, B V , let
d(A, B) =
e(A, B)
[A[[B[
,
where e(A, B) is the number of edges with a vertex in A and a vertex in B.
First observe some obvious properties of edge density. The quantity d(A, B)
is the ratio of the number of edges between A and B to the maximum possible
number of such edges. Thus if the subgraph on A and B is a complete bipartite
graph, the density is 1 (as in Figure 1). If no edges connect A and B, the density
is 0. Edge density is always a rational number in [0, 1]. Here are two slightly less
trivial examples: In Figure 2, we have d(A, B) = 15/16 and d(C, D) = 5/8.
We shall now introduce -regular pairs. Consider two disjoint vertex sets, A and
B. Intuitively, the pair of sets A, B is said to be -regular if the edges between
A and B are distributed fairly homogeneously.
Denition 2.2. (-Regular Pair) Let > 0. Finite nonempty disjoint sets A, B
V are said to be -regular if for all A

A and B

B, we have
[A

[ [A[ and [B

[ [B[ = [d(A, B) d(A

, B

)[ .
First observe that any A, B is trivially -regular for all 1, since it is
impossible for [d(A, B) d(A

, B

)[ to be greater than 1. Thus -regularity is only

meaningful for < 1.
THE REGULARITY LEMMA AND GRAPH THEORY 3
C D A B
(ii) (i)
Figure 2. (i ) The pair A, B is (1/3)-regular. (ii ) The pair
C, D is not (1/3)-regular.
The justication for the above denition is by comparison to large random
graphs. Suppose, for example, that A and B are countably innite. Let p [0, 1]
and for each a A and b B let a, b be an edge with probability p. Then for
large nite subsets A

A and B

B (taking the limit as [A

[, [B

[ ), the
expected number of edges would be p[A

[[B

[. So the expected density would be

p|A

||B

|
|A

||B

|
= p. But of course, if we were to pick out a single vertex a A and a single
vertex b B, then d(a, b) = 0 or 1. Thus we can only be reasonably certain
that the edge density will be very close to p for certain large subsets of A and B.
We use to specify an upper bound for the size of sets which do not necessarily
behave randomly.
Here are some examples of -regular pairs. The pair A, B in Figure 1 is -
regular for all , since d(A, B) = 1 and since for any subsets A

A and B

B,
we also have d(A

, B

) = 1. More generally, if the subgraph induced by A and B is

a complete bipartite graph, then A, B is -regular for all . Any pair a, b
of singletons is -regular for all , since the only nonempty subset of a singleton is
itself.
In Figure 2, the pair A, B is (1/3)-regular while the pair C, D is not. Let us
see why. We already observed that d(A, B) = 15/16 and d(C, D) = 5/8. To check
(1/3)-regularity, we need to look at subsets of size at least (1/3)4| = 2. First we
show that A, B is (1/3)-regular. Suppose not. there are subsets A

A and B

B of size at least 2 with [d(A

, B

) 15/16[ > 1/3. Thus d(A

, B

) < 29/48 < 3/4.

But since the bipartite graph between A and B is complete except for one edge and
since [A

[ 2 and [B

[ 2, we also have d(A

, B

) 3/4, a contradiction.
Now we show that C, D is not (1/3)-regular. Let C

be the two vertices on the

top left and D

be the two vertices on the bottom right, in the right side of Figure
2. Then d(C

, D

) = 1/4, so [d(C

, D

) d(C, D)[ = [1/4 5/8[ = 3/8 > 1/3, which

contradicts (1/3)-regularity.
Now we introduce the partitions we will use. In fact, usually we will not use
partitions, but rather collections of disjoint sets covering most, but not all, vertices
of a graph. The phrase collection of disjoint sets will be used with the implied
meaning of almost partition, or a partition of all of a graph except for a few
4 ARIEL HAFFTKA
A
B
C
Figure 3. A (2/3)-regular collection.
leftover vertices. We will only use the word partition when its precise meaning
is intended, that is to refer to a collection of disjoint sets which cover a set.
Given a large dense graph, we would like to group its vertices into components,
in such a way that most choices of two components form -regular pairs and such
that all components are the same size. We also would like to group most of the
vertices, although not necessarily all, into these components.
Denition 2.3. (-Regular Collection) A collection of pairwise disjoint vertex sets
/ = A
1
, . . . , A
k
is said to be -regular if
(1) (regular pairs) At most k
2
pairs A
i
, A
j
are not -regular, for 1 i <
j k.
(2) (equal size) [A
1
[ = = [A
k
[.
(3) (almost a partition) [V

AA
[ [V [.
An example of a (2/3)-regular collection is Figure 3. It is easy to check condition
(1) of (2/3)-regularity: that the pairs A, B, B, C, and A, C are (2/3)-regular.
As an example, we will prove that A, C is (2, 3)-regular. First observe that
d(A, C) = 2/9. To check -regularity, we must show that for any subsets A

A
and C

C of size at least (2/3)(3) = 2, we have [d(A

, C

) 2/9[ 2/3. Suppose

this were false. Then there exist A

and C

such that d(A

, C

) > 2/3 + 2/9 = 8/9.

But since d(A

, B

) is of the form i/9 for some i, we have d(A

, C

) = 1, which
implies the graph between A

and C

is a complete (2, 2) bipartite graph. This

contradicts the picture.
Now we check the two other conditions of (2/3)-regularity. Condition (2) is
satised because [A[ = [B[ = [C[, while condition (3) is satised because only 1
vertex is excluded from the collection, and 1 6 = (2/3)3
2
= k
2
.
Since every pair of singletons is -regular, every graph on n vertices trivially
admits an -regular collection of n disjoint singletons. Thus the importance of the
Regularity Lemma is in the claim that we can put an upper (and lower) bound on
the number of sets needed to form an -regular collection, independent of how large
a graph is.
THE REGULARITY LEMMA AND GRAPH THEORY 5
Theorem 2.4. (Regularity Lemma) For every > 0 and m N, there exists an
integer M = M(, m) such that every graph G on m or more vertices admits an
-regular collection / = A
1
, . . . , A
k
with m k M.
3. Proof of the Regularity Lemma
Our proof will follow [3]. The main idea is as follows. We dene an index of
a pair of disjoint vertex sets A and B. We extend this denition so becomes a
function from the set collections of disjoint subsets of V R. In Lemma 3.2
we show that is bounded above. Then we describe an algorithm that takes an -
irregular collection and produces a renement with index increased by the positive
constant
5
(1)
2
(assuming < 1) depending only on . Thus by the boundedness
of , the algorithm will eventually result in an -regular collection.
The algorithm involves two steps. In Lemma 3.6 we show that if a pair of disjoint
sets A, B is not -regular, then there exist partitions of / of A and B of B such
that the index of /, B is greater than the index of A, B by at least

4
|A||B|
|V |
2
. In
Lemma 3.11, we use the previous result to show that if a collection / of disjoint
vertex sets is not -regular, then a renement exists with index increased by at least
the nonzero constant
5
(1 )
2
, depending only on .
We begin by dening the index , rst for a pair of disjoint vertex sets; then
for a nite collection A
i

iI
of disjoint vertex sets; then for a pair of collections
A
i

iI
of disjoint sets in A and B
j

jJ
of disjoint sets in B; and nally we dene
an extended version

which will only be used in Lemma 3.11.

Denition 3.1. (Index )
(1) For disjoint A, B V , let
(A, B) =
e(A, B)
2
[A[[B[[V [
2
.
(2) For a collection / = A
i

iI
of pairwise disjoint vertex sets, let
(/) =

{i,j}I
(A
i
, A
j
).
(3) Let A and B be disjoint vertex sets. For collections / = A
i

iI
of pairwise
disjoint subsets of A and B = B
j

jJ
of pairwise disjoint subsets of B, let
(/, B) =

iI,jJ
(A
i
, B
j
).
(4) Furthermore, let

(/) =
_
/
_
x

x V
_
AA
A
_
_
.
We now show that the index is bounded above.
Lemma 3.2. Let / = A
i

iI
be a collection of pairwise disjoint vertex sets. Then
(/) 1.
6 ARIEL HAFFTKA
Proof.
(/) =

{i,j}I
(A
i
, A
j
)
=

{i,j}I
d(A
i
, A
j
)
2
[A
i
[[A
j
[
[V [
2

1
[V [
2

{i,j}I
[A
i
[[A
j
[

1
[V [
2
_

iI
[A
i
[
_
_
_

jJ
[A
j
[
_
_

1
[V [
2
[V [[V [
= 1.

The Cauchy-Schwarz inequality

(3.3)
n

i=1

2
i

n
i

i
)
2

n
i=1

i
.
for
i
R and
i
R
+
, will be used in Lemmas 3.4 and 3.6.
Lemma 3.4. ( does not decrease when taking renements.) Let A, B V be
disjoint, / = A
i

iI
be a partition of A, and B = B
j

jJ
be a partition of B.
Then
(/, B) (A, B).
Furthermore, if /

is a partition of A that renes /, then

(/

) (/).
Proof. (1) First we show that (/, B) (A, B).
THE REGULARITY LEMMA AND GRAPH THEORY 7
(/, B) =

iI,jJ
(A
i
, B
j
)
=

iI,jJ
e(A
i
, B
j
)
2
[A
i
[[B
j
[[V [
2

iI,jJ
e(A
i
, B
j
)
_
2
[V [
2

iI,jJ
[A
i
[[B
j
[
, (by Cauchy-Schwarz 3.3) .
=
e(A, B)
2
[V [
2
(

iI
[A
i
[)(

jJ
[B
j
[)
,
_
since

iI,jJ
e(A
i
, B
j
) counts all the (3.5)
edges from A to B
_
.
=
e(A, B)
2
[V [
2
[A[[B[
=(A, B).
(2) The second claim follows from the rst. For each X /, let
/
X
= A

[A

X,
so that /
X
is the partition of X induced by /

. Then
(/

) =

{A

,A

}A

(A

, A

{X,X

}A

A
X
,A

A
X

(A

, A

) (since /

renes /) .
=

{X,X

}A
(/
X
, /
X
)

{X,X

}A
(X, X

), (by part 1) .
=(/).

The next lemma is the rst of two crucial parts of our proof. We look at a single
pair of disjoint vertex sets A, B and show that if it is not -regular, then there
exist partitions of A and of B, each into two sets, with index increased by at least

4
|A||B|
|V |
2
. The main idea is the following: Since A, B is not -regular, there exist
subsets of A and of B that contradict the denition of -regularity. We use these
sets to construct partitions of A and B. The rest of the proof involves computation
and an application of Cauchy-Schwarz to prove the inequality.
Lemma 3.6. Let A, B V be disjoint. If the pair A, B is not -regular, there
exist partitions / of A and B of B such that [/[ 2, [B[ 2, and
(/, B) (A, B) +

4
[A[[B[
[V [
2
.
8 ARIEL HAFFTKA
Proof. 1. Constructing the partitions.
Since A, B is not -regular, there exist A
1
A and B
1
B such that [A
1
[
[A[, [B
1
[ [B[, and
(3.7) [d(A
1
, B
1
) d(A, B)[ .
Let A
2
= AA
1
and B
2
= BB
1
, so that / := A
1
, A
2
is a partition of A and
B := B
1
, B
2
is a partition of B. At least one of A
2
and B
2
is nonempty, since
otherwise (3.7) would be false. Without loss of generality, assume A
2
is nonempty.
Then we can write
/ = A
i

iI
; B = B
j

jJ
;
where I = 1, 2 and J = 1, 2 if B
2
is nonempty and J = 1 if B
2
is empty.
2. Proving the inequality.
We claim that (/, B) is suciently larger than (A, B). For brevity, let
:=d(A
1
, B
1
) d(A, B);
a :=[A[; b := [B[, e := e(A, B);
a
i
:=[A
i
[; b
j
:= [B
j
[; e
ij
= e(A
i
, B
j
);
with b
2
= 0 = e
i2
if B
2
is empty.
We are borrowing some notation from [3]. The goal is to express the dierence
between (/, B) and (A, B) in terms of , and then the claim will follow from
[[ .
Since by denition, =
e11
a1b1

e
ab
, we have e
11
= a
1
b
1
+
ea1b1
ab
. Thus we have
two simple identities.
First,
e
2
11
=
_
a
1
b
1
+
ea
1
b
1
ab
_
2
=
2
a
2
1
b
2
1
+
2ea
2
1
b
2
1
ab
+
e
2
a
2
1
b
2
1
a
2
b
2
. (3.8)
Second,
(e e
11
)
2
=
_
e
_
ab a
1
b
1
ab
_
a
1
b
1
_
2
= e
2
_
ab a
1
b
1
ab
_
2
2a
1
b
1
e
_
ab a
1
b
1
ab
_
+
2
a
2
1
b
2
1
. (3.9)
THE REGULARITY LEMMA AND GRAPH THEORY 9
Putting it all together,
[V [
2
(/, B) =[V [
2

iI,jJ
(A
i
, B
j
)
=

iI,jJ
e
2
ij
a
i
b
j
=
e
2
11
a
1
b
1
+

iI,jJ,(i,j)=(1,1)
e
2
ij
a
i
b
j

e
2
11
a
1
b
1
+
_

iI,jJ,(i,j)=(1,1)
e
ij
_
2

iI,jJ,(i,j)=(1,1)
a
i
b
j
, (by Cauchy-Schwarz 3.3) .
=
e
2
11
a
1
b
1
+
(e e
11
)
2
ab a
1
b
1
,
_
_
since a =

iI
a
i
, b =

jJ
b
j
, and e =

iI,jJ
e
ij
_
_
.
=

2
a
2
1
b
2
1
+
2ea
2
1
b
2
1
ab
+
e
2
a
2
1
b
2
1
a
2
b
2
a
1
b
1
+
e
2
_
aba1b1
ab
_
2
2a
1
b
1
e
_
aba1b1
ab
_
+
2
a
2
1
b
2
1
ab a
1
b
1
,
(by 3.8 and 3.9) .
=
_

2
a
1
b
1
+
2ea
1
b
1
ab
+
e
2
a
1
b
1
a
2
b
2
_
+
e
2
(ab a
1
b
1
)
a
2
b
2

2a
1
b
1
e
ab
+

2
a
2
1
b
2
1
ab a
1
b
1
=
2
a
1
b
1
+
e
2
ab
a
2
b
2
+

2
a
2
1
b
2
1
ab a
1
b
1

4
ab +
e
2
ab
,
_
since [[ ; a
1
a; b
1
b; and a
1
< a [since A
2
is
nonempty]
_
.
Thus (/, B) (A, B) +

4
|A||B|
|V |
2
.
The next lemma is the second crucial part of our proof. In it we show that if
a collection T = A
1
, . . . , A
k
of pairwise disjoint, equal-sized vertex sets covering
most of V is not -regular, then there exists a renement T

of T, covering almost
as much as T, with index increased by a positive constant depending only on . We
use the following broad denition of renement.
Denition 3.10. (Renement) Let T and T

be collections of pairwise disjoint

subsets of V . We say that T

renes T if each set in T

is a subset of some set in

T.
Trivially, renement is a partial ordering on the set collections of disjoint vertex
sets, with T

T if T

renes T. In the proof of the following, we will use the

term maximal with respect to this partial ordering and a given condition.
Lemma 3.11. Let 0 < < 1; T := A
1
, . . . , A
k
be a collection of pairwise disjoint
subsets of V ; [A
1
[ = = [A
k
[ := c; and [V

k
i=1
A
i
[ [V [. If T is not -regular,
then there exists a collection T

= A

1
, . . . , A

l
of pairwise disjoint equal-sized
subsets of V rening T such that
10 ARIEL HAFFTKA
(1) (The collection T

covers almost as much as T):

[V
l
_
i=1
A

i
[ [V
k
_
i=1
A
i
[ +
[V [
2
k+1
.
(2) (The collection T

has extended index

increased by a positive constant

depending only on ):

(T

(T) +
5
(1 )
2
.
(3) (The collection T

is bounded): k l k4
k
Proof. The proof consists of two parts: rst constructing the partitions and second
proving the inequalities.
1. Constructing the partitions.
The main idea of the construction is as follows: Since the initial collection T is
not -regular, there exist more than k
2
-irregular pairs. We will apply Lemma 3.6
to obtain, for each such pair A
i
, A
j
, a partition /
ij
of A
i
and a partition /
ji
of
A
j
with index increased by at least

4
|A||B|
|V |
2
.
The next step will be to use the partitions /
ij
to create the desired collection T

.
First we will take the renement / of T induced by all the /
ij
and /
ji
. Second
we will slice the sets of / into sets of equal size, to produce the desired collection
T

. See Figure 4 for a diagram of the construction.

We begin the construction. For 1 i < j k, choose partitions /
ij
of A
i
and
/
ji
of A
j
as follows.
(1) If A
i
, A
j
is -regular, let /
ij
:= A
i
and /
ji
:= A
j
.
(2) If A
i
, A
j
is not -regular, then by Lemma 3.6 there exist partitions /
ij
of A
i
and /
ji
of A
j
such that [/
ij
[, [/
ji
[ 2 and
(3.12) (/
ij
, /
ji
) (A
i
, A
j
) +
4
c
2
[V [
2
.
For each i, let /
i
be the unique partition of A
i
induced by all partitions /
ij
with j [1, k]i. In terms of topology, /
i
is the set of minimal nonempty open
sets, or atoms, in the nite topology generated on A
i
by the sets in /
ij
with
j [1, k]i. Thus each /
i
contains at most 2
k1
sets, since each starting set A
i
has been cut into two pieces at most k 1 times.
Let
(3.13) / :=
k
_
i=1
/
i
.
It follows from [/
i
[ 2
k1
that
k [/[ k2
k1
,
satisfying part (3) of the claim.
Now we further rene / to obtain a collection with all sets of the same size. Let
T

:= A

1
, . . . A

l
be a maximal collection of disjoint sets of size d :=
c
4
k
| that
renes /. A trivial example exists: Since each /
i
covers the set A
i
, [A
i
[ = c, and
/
i
has at most 2
k1
sets, it follows that at least one component of /
i
has size
at least
c
2
k1
> d. Thus we could pick a single set of d vertices from at least one
component of each /
i
. Since V is nite, a maximal collection certainly exists. We
claim T

has the desired properties.

THE REGULARITY LEMMA AND GRAPH THEORY 11
A
1
A
2
A
3
A
12
A
13
A
21
A
23
A
31
A
32
A
1
A
2
A
3
A
1
A
2
A
3
P
P
A
(i)
(ii)
(iii)
(iv)
Figure 4. The construction of T

, for k = 3 and assuming all

pairs A
i
, A
j
start -irregular. (i ) We begin with the collection
T. (ii ) Using Lemma 3.6, each set is partitioned in two dierent
ways, with respect to the other sets. For example, the partition
/
12
is a partition of A
1
with respect to A
2
. (iii ) The induced
partition /
i
is taken on each set A
i
. (iv) The sets are cut into
pieces of size d.
2. Proving the inequalities.
Since each component of T

is of size d =
c
4
k
| and is contained in at least one
set A
i
of size c, we have k l k4
k
.
12 ARIEL HAFFTKA
Since T

excludes less than d vertices from each component of /,

[V
l
_
i=1
A

i
[ [V
k
_
i=1
A
i
[ +d[/[
[V
k
_
i=1
A
i
[ +
_
c
4
k
_
k2
k1
=[V
k
_
i=1
A
i
[ +
ck
2
k+1
[V
k
_
i=1
A
i
[ +
[V [
2
k+1
, (since T is a partition of [V [ or fewer vertices into k sets of size c) .
We claim that

(T

(T) +
5
(1 )
2
. Let
/

0
=
_
x[x V
l
_
i=1
A

i
_
; /
0
=
_
x[x V
k
_
i=1
A
i
_
.
The goal is to use Lemma 3.4 to show that

(T

(T). We must use

, not
, for the following reason. Although the collection T

renes T, it may cover fewer

vertices than T. Lemma 3.4, however, only applies if T

renes T and covers the

same vertices. The extended index

avoids this problem by considering all the

vertices not covered by a partition to be singletons in the partition. By denition,

(T

) = (T

0
). Now we can make two observations: First, T

0
renes
//
0
(since T

renes / and each set in /

0
is a singleton). Second, T

0
and
/ /
0
both cover all of V . Thus T

0
and / /
0
satisfy the conditions of
Lemma 3.4.

(T

) = (T

0
)
(/ /
0
) ,
_
by Lemma 3.4, since T

0
renes / /
0
and by the above
discussion
_
.
=
_
k
_
i=0
/
i
_
, by (3.13).
THE REGULARITY LEMMA AND GRAPH THEORY 13
=

1i<jk
(/
i
, /
j
) +
k

i=0
(/
i
) +
k

i=1
(/
0
, /
i
),

1i<jk
(/
ij
, /
ji
) +(/
0
) +
k

i=1
(/
0
, A
i
),
_
by Lemma 3.4, since
in the rst term /
i
renes /
ij
and /
j
renes /
ji
; in the second term the
summands are nonnegative; and in the third term /
i
renes A
i

_
.

1i<jk
(A
i
, A
j
) +k
2

4
c
2
[V [
2
+(/
0
) +
k

i=1
(/
0
, A
i
),
_
since at least k
2
irregular pairs A
i
, A
j
satisfy (3.12)
_
.
=

(T) +
5
k
2
c
2
[V [
2
, (by denition of

) .

(T) +
5
(1 )
2
,
_
since [V [ ck = [A
0
[ [V [ and 0 < < 1 imply
_
ck
[V [
_
2
(1 )
2
_
.

The Regularity Lemma now follows by repeated application of Lemma 3.11 and
by the boundedness of

.
Proof. (Proof of the Regularity Lemma)
Let > 0 and m N. Lemma 3.2 shows that , and consequently

, is bounded
above by 1. Thus
b :=
_
1
e
5
(1 )
2
_
is an upper bound on the number of times Lemma 3.11 can be applied before a
partition is forced to be -regular. Choose k
0
N large enough that
2
k0
b/.
Let : N N be given by (x) = x4
x
. Let
M = max
_

b
(k
0
),
2k
0

_
,
where the exponent denotes iteration.
Let G be a graph on n m vertices. We aim to show that G admits an -regular
collection containing between m and M sets. If n M, G admits an -regular
collection into singletons. We can thus assume n > M.
Let / = A
1
, . . . , A
k0
be a collection of disjoint subsets of V , each of size
n
k0
|
and let /
s
= A
s
1
, . . . , A
s
ks
denote a result of iterating Lemma 3.11 s times.
We need to check that after each iteration, /
s
either is -regular or satises the
hypotheses of Lemma 3.11. The sets of /
s
are of the same size, so all there is to
14 ARIEL HAFFTKA
check is that /
s
covers all but at most n vertices. For 0 s b, we have
[V
ks
_
i=1
A
s
i
[ =[V
k0
_
i=1
A
i
[ +
sn
2
k0+1
, (by Lemma 3.11) .
k
0
+
bn
2
k0+1
,
_
since / covers all but at most k
0
vertices
and since s b
_
.
k
0
+
n
2
,
_
since 2
k0

implies
b
2
k0

_
.

n
2
+
n
2
,
_
since n M
2k
0

implies k
0

n
2
_
.
=n.
Thus Lemma 3.11 can be iterated until / becomes -regular. The nal -regular
collection will have at most
b
(k) = M components.
4. The Key Lemma
We have taken the Key Lemma, regularity graphs, and some notations from [3].
The Regularity Lemma allows one to approximate the structure of a large dense
graph G with a smaller graph R, called a regularity graph. The main idea of the
Key Lemma is that the regularity graph can be used to describe sugraphs of the
original graph.
Suppose for example that G is a large dense graph and that the Regularity
Lemma has provided an -regular collection / = A
1
, . . . , A
k
. We let R be the
graph whose vertex set is /. We let A
i
, A
j
be an edge of R if the pair is -regular
and if the edge density d(A
i
, A
j
) is greater than some xed choice of minimum
density d (0, 1]
_
we will make this precise in Denition 4.1
_
.
For each edge A
i
, A
j
of the regularity graph, we can put a lower bound on the
number of edges of G with one vertex in A
i
and one vertex in A
j
. In fact, we can
show that almost all vertices of A
i
are connected to at least a d fraction of the
vertices of A
j
(Lemma 4.4). Thus, if for example the regularity graph contains a
triangle, it would be reasonable to hope that a triangle also exists in G.
Denition 4.1. (Regularity Graph) Let G be a graph, > 0, d (0, 1], and
/ = A
1
, . . . , A
k
be an -regular collection with [A
1
[ = = [A
k
[ := l. The
regularity graph of G with respect to / and with parameters , d, and l is the
graph R with vertex set V (R) := / and edge set
E(R) :=
_
A
i
, A
j

A
i
, A
j
is an -regular pair of density d(A
i
, A
j
) d
_
.
The regularity graph can be generalized as follows.
Denition 4.2. (s-fold Regularity Graph) Let s N and let R be a regularity
graph, as above. The s-fold regularity graph R
s
is obtained by replacing each vertex
A
i
of R with an s-set A
s
i
:= v
1
i
, . . . , v
s
i
of s distinct vertices and by replacing each
edge A
i
, A
j
with a complete bipartite graph between A
s
i
and A
s
j
.
The Key Lemma asserts that under the right conditions, the existence of a graph
in R
s
implies its existence in G.
THE REGULARITY LEMMA AND GRAPH THEORY 15
R
A
1
A
2
A
3
A
1
A
2
A
3
Figure 5. An example of a regularity graph produced by an -
regular collection.
A
4
1
A
4
2
A
4
3
R
4
Figure 6. The 4-fold regularity graph produced by R from Figure 5.
Theorem 4.3. (The Key Lemma) For all d (0, 1] and any integer 0, there
exists
0
> 0 such that the following holds: if G is any graph; s is any natural
number; R
s
is any s-fold regularity graph of G with parameters
0
, d, and
l s/
0
, and H is any graph with maximum degree (H) ; then
H R
s
= H G.
Before proving the Key Lemma, we need the following trivial lemma.
Lemma 4.4. Let A, B be an -regular pair of density d; Y B with [Y [ [B[;
and
X = x A[ deg(x, Y ) < (d )[Y [.
Then [X[ < [A[.
16 ARIEL HAFFTKA
Proof.
d(X, Y ) =

xX
deg(x, Y )
[X[ [Y [
<
(d )[X[[Y [
[X[[Y [
= d .
Therefore [d(X, Y ) d(A, B)[ > . Since A, B is an -regular pair and since
[Y [ , this implies [X[ < [A[.
Lemma 4.4 provides an upper bound on the number of vertices x A with less
than the expected number of neighbors in Y , provided Y is large enough. We will
use this fact repeatedly in our proof of the Key Lemma. It will allow us to embed H
into G by picking vertices one by one and, at each stage, using Lemma 4.4 to ensure
that such a choice is possible. The proof involves keeping track, at each stage of the
embedding, of which vertices in each component of the -regular collection are still
valid choices and of which vertices can no longer be used because they contradict
previously established adjacency relations. We will provide an algorithm and then
prove by induction that the algorithm works.
While our proof borrows the main ideas and notation from [3], the algorithm is
slightly based on that of [4] and the inductive part of the proof is largely original.
We will use the following notation in our proof:
Notation 4.5. (Neighbors) Let G = (V, E) be a graph. For v V , let N(v) be
the set of vertices adjacent to v.
Proof. (Proof of the Key Lemma.) Choose
0
< d small such that
(4.6) ( + 1)
0
(d
0
)

.
Assume that H is a subgraph of R
s
and let V (H) := u
1
, . . . , u
h
. Each vertex
u
i
of H is contained in one of the s-sets A
s
j
for some j [1, k]. This denes a
function : [1, h] [1, k] satisfying
u
i
A
s
(i)
for each i [1, h].
We aim to embed H into G by associating each vertex u
i
of H with a unique
vertex v
i
in A
(i)
. The goal is to show that v
1
A
(1)
, . . . , v
h
A
(h)
can be
chosen such that
(1) (The v
i
s are unique.) We have u
i
,= u
j
= v
i
,= v
j
.
(2) (Edges are preserved.) If u
i
, u
j
E(H), then v
i
, v
j
E(G).
We will pick v
1
, . . . , v
h
inductively. Immediately after having picked v
1
, . . . , v
n
,
we denote by Y
n
i
the set of vertices in A
(i)
that are still acceptable choices for v
i
.
At the start, Y
0
1
= A
(1)
, . . . , Y
0
h
= A
(h)
, since there are no established adjacency
relations that might restrict the choices for v
1
, . . . , v
h
. At each stage, a vertex
v
n
is chosen in Y
n1
n
and the sets Y
n
n
, . . . , Y
n
h
are created from the previous sets
Y
n1
n
, . . . , Y
n1
h
based on the adjacency relations introduced by the choice of v
n
.
The sets thus satisfy
A
(i)
= Y
0
i
Y
1
i
Y
i1
i
Y
i
i
= v
i
,
THE REGULARITY LEMMA AND GRAPH THEORY 17
u
1
u
2
u
3
u
4
u
12
u
5
u
6
u
7
u
9
u
8
u
10
u
11
R
s
H
A
s
(1)
A
s
(3)
A
s
(9)
A
s
(7)
A
s
(11)
A
s
(4)
Figure 7. The graph H as a subgraph of R
s
. In this example, R
is a 6-cycle and s = 3.
for each i [1, h].
The First Stage of the Algorithm
The rst stage of the algorithm is to pick the vertex v
1
Y
0
1
= A
(1)
and to
compute the sets Y
1
1
, . . . Y
1
h
. This stage is almost identical to the succeding stages.
We explain it in detail to give the reader a picture of how the general algorithm
will work, although this entire section is not really part of the proof. So the reader
may want to skip to the section titled Algorithm.
The rst step is to pick v
1
A
(1)
. We need to pick v
1
in such a way that there
is a good chance of a successful embedding. This can be done by looking at the set
N(u
1
) of vertices in H adjacent to u
1
. There are at most such vertices. We label
these vertices N(u
1
) = u
(1)
, . . . u
(d)
, with d := [N(u
1
)[ . Our goal can be
accomplishing by picking v
1
A
(1)
such that deg(v
1
, A
(i)
) is large for each u
(i)
adjacent to u
1
. Lemma 4.4 will ensure that such a choice is possible. This will be
made precise in the proof of proposition P(n) later in the proof.
Once v
1
has been chosen, the next step is to update the sets Y
0
1
, . . . Y
0
h
so that
they accurately reect the remaining acceptable choices for the to be determined
vertices v
2
, . . . , v
h
. Since v
1
has been chosen, the set Y
0
1
collapses to a single point,
and we have Y
1
1
= v
1
.
18 ARIEL HAFFTKA
Updating Y
0
2
, . . . , Y
0
h
is slightly less trivial. In fact, we claim that
Y
1
i
:=
_
Y
0
i
N(v
1
), if u
i
, u
1
E(H)
Y
0
i
, else
for i [2, h]. To prove this, consider the two cases. First, assume that u
i
, u
1

E(H). Then for any vertex v Y
0
i
we have:
v is an acceptable choice for v
i
v
1
, v E(H),
_
since u
i
, u
1
E(H)
_
v N(v
1
).
For the second case, if u
i
, u
1
, E(H), then the choice of v
1
has created no new
restrictions on the acceptable choices for v
i
, so Y
1
i
= Y
0
i
, as claimed. The general
algorithm is as follows:
Algorithm. For n = 1, 2, . . . , h 1,
(1) Pick v
n
Y
n1
n
such that v
n
, v
1
, . . . , v
n1
and
deg(v
n
, Y
n1
i
) (d )[Y
n1
i
[
holds for all i [n + 1, h] with u
i
, u
n
E(H).
(2) Set
Y
n
i
:=
_
Y
n1
i
N(v
n
), if u
i
, u
n
E(H)
Y
n1
i
, else
for all i [n + 1, h].
To prove that the algorithm works, we shall induct on the following statement.
Proposition P(n)
(1) Vertices v
1
, . . . , v
n
have been chosen according to the algorithm.
(2) If n < h, then for each i [n + 1, h],
[Y
n
i
[ (d )

(n,i)
l,
where
(n, i) := deg
_
u
i
, u
1
, . . . , u
n

_
is the number of edges in H with one vertex equal to u
i
and another vertex in the
set u
1
, . . . , u
n
and where (0, i) := 0.
Proof of the base case.
Let n = 0. Then part (1) of P(0) is true, since 0 vertices have been chosen. Part
(2) is true because
[Y
0
i
[ = [A
(i)
[ = l = (d )
0
l = (d )

(0,i)
l.
Proof of the inductive step.
Let n [1, h] and assume P(n 1).
First we prove part (1) of P(n). The goal is to show that Y
n1
i
is not too small,
and then Lemma 4.4 will show that there are at least s acceptable choices for v
n
.
Then we use the Pigeonhole Principle to show that at least one of these s choices
is distinct from v
1
, . . . , v
n1
.
THE REGULARITY LEMMA AND GRAPH THEORY 19
First we claim that for each i [n, h], the set Y
n1
i
is not too small.
[Y
n1
i
[ (d )

(n,i)
l, ( by inductive hypothesis.)
(d )

l,
_
since
(n,i)
.
_
( + 1)
0
l, ( by assumption.)
=
0
l +
0
l
l +s, ( since
0
and l s/
0
) . (4.7)
Now we can apply Lemma 4.4. For each i [n+1, h] such that u
i
, u
n
E(H),
the sets
A := A
(n)
; B := A
(i)
;
form an -regular pair. And letting Y := Y
n1
i
B, we have
[Y [ = [Y
n1
i
[ l = [B[, by (4.7) and since 1.
Thus by Lemma 4.4,
[v A
(n)
[ deg(v, Y ) < (d )[Y [[ < l.
Applying the above result to the at most values of i for which u
i
is adjacent to u
n
,
it follows that at most l elements v A
(n)
are such that for some i [n+1, h],
both deg(v, Y
n1
i
) < (d )[Y
n1
i
[ and u
i
, u
n
E(H). Thus there are at least
[Y
n1
n
[ l l +s l,
_
by 4.7.
_
= s. (4.8)
acceptable choices for v
n
.
Now we must check that of the at least s acceptable choices for v
n
in Y
n1
n
given by (4.8), at least one such choice will be such that v
n
is distinct from all the
previously chosen vertices v
1
, . . . , v
n1
. Suppose indirectly that all of the at least
s acceptable choices for v
n
are equal to v
i
for some i [1, n 1]. Let
I =
_
i [1, n 1]

v
i
is one of the at least s acceptable choices for v
n
in Y
n1
n
_
.
Then by our assumption [I[ s. By the inductive hypothesis and since I [1, n1],
we have successfully embedded u
i
v
i
for all i I. Since v
i

iI
Y
n1
n
A
(n)
,
we have u
i

iI
A
s
(n)
. Since we also know u
n
A
s
(n)
(as given), we have
(4.9) u
i

iI
u
n
A
s
(n)
,
It is also given that u
1
, . . . , u
n
are pairwise distinct, thus
(4.10) [u
i

iI
u
n
[ = [I[ + 1 s + 1.
Lines (4.9) and (4.10) contradict the fact that [A
s
(n)
[ = s, by denition. We thus
conclude that a distinct and acceptable choice for v
n
exists in Y
n1
n
A
(n)
.
Next we prove part (2) of P(n). Since the vertex v
n
has been chosen, part (2) of
the algorithm denes the sets Y
n
i
for i [n+1, h]. We assert that [Y
n
i
[ (d)

(n,i)
for each i [n + 1, h]. Fix i. To prove the inequality, let
:=
_
1, if u
i
, u
n
E(H)
0, else.
20 ARIEL HAFFTKA
to indicate whether or not u
i
, u
n
is an edge. Then
[Y
n
i
[ (d )

[Y
n1
i
[, (by the choice of v
n
and Y
n
i
in the algorithm).
(d )

(d )

(n1,i)
, (by inductive hypothesis).
= (d )

(n1,i)
+
= (d )

(n,i)
,
_
since
(n,i)
= (n 1, i) +, trivially
_
.
This proves part (2) of P(n). P(h) follows by induction.
Finally, we observe that P(h) implies the claim. Part (1) of P(h) asserts that
vertices v
1
, . . . , v
h
have been chosen according to the algorithm. Part (2) of the
algorithm ensures that our embedding u
i
v
i
is injective and edge-preserving.
Thus v
1
, . . . , v
h
is a copy of H in G.

5. Acknowledgements
I thank Asaf Hadari and Mike Shulman for reading my work and providing many
corrections.
References
[1] E. Szemeredi On sets of integers containing no k elements in arithmetic progression. Acta
Arithmetica. XXVII., pp. 199-245. 1975.
[2] E. Szemeredi Regular Partitions of Graphs In: Problemes Combinatoires et Theorie des
Graphes, No. 260, p. 399. Proceedings July 9-13, 1976, in Orsay. Centre National de la
Recherche Scientique (CNRS). Paris, 1978.
[3] R. Diestel Graph Theory Springer-Verlag. 1997.
[4] J. Komlos, M. Simonovits Szemeredis Regularity Lemma and Its Applications in Graph The-
ory In: Combinatorics, Paul Erdos is Eighty, vol. 2 Janos Bolyai Mathematical Society. Bu-
dapest, 1996.