Graph Theory and Additive Combinatorics

Lecturer: Prof. Yufei Zhao

2
Forbidding subgraphs

2.1 Mantel’s theorem: forbidding a triangle

We begin our discussion of extremal graph theory with the following

basic question.

Question 2.1. What is the maximum number of edges in an n-vertex

graph that does not contain a triangle?

Bipartite graphs are always triangle-free. A complete bipartite

graph, where the vertex set is split equally into two parts (or differing
by one vertex, in case n is odd), has n2 /4 edges. Mantel’s theorem
 

states that we cannot obtain a better bound:

Theorem 2.2 (Mantel). Every triangle-free graph on n vertices has at W. Mantel, "Problem 28 (Solution by H.
Gouwentak, W. Mantel, J. Teixeira de
most bn2 /4c edges.
Mattes, F. Schuh and W. A. Wythoff).
Wiskundige Opgaven 10, 60 —61, 1907.
We will give two proofs of Theorem 2.2.

Proof 1. Let G = (V, E) a triangle-free graph with n vertices and m x

edges. Observe that for distinct x, y ∈ V such that xy ∈ E, x and y N (x)

must not share neighbors by triangle-freeness.

Therefore, d( x ) + d(y) ≤ n, which implies that

∑ d ( x )2 = ∑ (d( x ) + d(y)) ≤ mn.

N (y)
x ∈V xy∈ E y

On the other hand, by the handshake lemma, ∑ x∈V d( x ) = 2m. Now

by the Cauchy–Schwarz inequality and the equation above, Adjacent vertices have disjoint neigh-
borhoods in a triangle-free graph.
!2 !
2
4m = ∑ d( x ) ≤n ∑ d( x ) 2
≤ mn2 ;
x ∈V x ∈V

hence m ≤ n2 /4. Since m is an integer, this gives m ≤ bn2 /4c.

y
Proof 2. Let G = (V, E) be as before. Since G is triangle-free, the
x
neighborhood N ( x ) of every vertex x ∈ V is an independent set.
z
An edge within N ( x ) creates a triangle
24 turán’s theorem: forbidding a clique

Let A ⊆ V be a maximum independent set. Then d( x ) ≤ | A| for

all x ∈ V. Let B = V \ A. Since A contains no edges, every edge of G
intersects B. Therefore,

e( G ) ≤ ∑ d(x) ≤ | A|| B|
x∈B
$ 2 %
| A| + | B| n2
 
≤ = .
AM-GM 2 4

Remark 2.3. For equality to occur in Mantel’s theorem, in the above

proof, we must have

• e( G ) = ∑ x∈ B d( x ), which implies that no edges are strictly in B.

• ∑ x∈ B d( x ) = | A|| B|, which implies that every vertex in B is com-

plete in A.

• The equality case in AM-GM must hold (or almost hold, when n is
odd), hence || A| − | B|| ≤ 1.

Thus a triangle-free graph on n vertices has exactly n2 /4 edges if

 

and only if it is the complete bipartite graph Kbn/2c,dn/2e .

2.2 Turán’s theorem: forbidding a clique

Motivated by Theorem 2.2, we turn to the following more general

question.

Question 2.4. What is the maximum number of edges in a Kr+1 -free

graph on n vertices?

Extending the bipartite construction earlier, we see that an r-partite

graph does not contain any copy of Kr+1 .

Definition 2.5. The Turán graph Tn,r is defined to be the complete,

n-vertex, r-partite graph, with part sizes either nr or nr .
   

In this section, we prove that Tn,r does, in fact, maximize the num-
ber of edges in a Kr -free graph:

Theorem 2.6 (Turán). If G is an n-vertex Kr+1 -free graph, then e( G ) ≤

e( Tn,r ). The Turán graph T10,3
P. Turán, On an extremal problem in
When r = 2, this is simply Theorem 2.2. graph theory. Math. Fiz. Lapok 48, 436
We now give three proofs of Theorem 2.6. The first two are in the —452, 1941.

same spirit as the proofs of Theorem 2.2.

 forbidding subgraphs 25

Proof 1. Fix r. We proceed by induction on n. Observe that the state-

ment is trivial if n ≤ r, as Kn is Kr+1 -free. Now, assume that n > r
and that Turán’s theorem holds for all graphs on fewer than n ver-
tices. Let G be an n-vertex, Kr+1 -free graph with the maximum pos-
sible number of edges. Note that G must contain Kr as a subgraph,
or else we could add an edge in G and still be Kr+1 -free. Let A be the
vertex set of an r-clique in G, and let B := V \ A. Since G is Kr+1 -free,
every v ∈ B has at most r − 1 neighbors in A. Therefore
 
r
e( G ) ≤ + (r − 1)| B| + e( B)
2
 
r
≤ + (r − 1)(n − r ) + e( Tn−r,r )
2
= e( Tn,r ).

The first inequality follows from counting the edges in A, B, and

everything in between. The second inequality follows from the in-
ductive hypothesis. The last equality follows by noting removing
one vertex from each of the r parts in Tn,r would remove a total of
(2r ) + (r − 1)(n − r ) edges.

Proof 2 (Zykov symmetrization). As before, let G be an n-vertex, Kr+1 -

free graph with the maximum possible number of edges.
We claim that the non-edges of G form an equivalence relation;
that is, if xy, yz ∈
/ E, then xz ∈
/ E. Symmetry and reflexivity are easy
to check. To check transitivity, Assume for purpose of contradiction
that there exists x, y, z ∈ V for which xy, yz ∈
/ E but xz ∈ E.
If d(y) < d( x ), we may replace y with a “clone” of x. That is, we x

delete y and add a new vertex x 0 whose neighbors are precisely the
as the neighbors of x (and no edge between x and x 0 ). (See figure on
the right.)
Then, the resulting graph G 0 is also Kr+1 -free since x was not in
any Kr+1 . On the other hand, G 0 has more edges than G, contradict- x0
ing maximality. x and its clone x 0
Therefore we have that d(y) ≥ d( x ) for all xy ∈ / E. Similarly,
d(y) ≥ d(z). Now, replace both x and z by “clones” of y. The new
graph G 0 is Kr+1 -free since y was not in any Kr+1 , and

e( G 0 ) = e( G ) − (d( x ) + d(z) − 1) + 2d(y) > e( G ),

contradicting maximality of e( G ). Therefore such a triple ( x, y, z)

cannot exist in G, and transitivity holds.
The equivalence relation shows that the complement of G is a
union of cliques. Therefore G is a complete multipartite graph with
at most r parts. One checks that increasing the number of parts in-
creases the number of edges in G. Similarly, one checks that if the
26 hypergraph turán problem

number of vertices in two parts differ by more than 1, moving one

vertex from the larger part to the smaller part increases the number
of edges in G. It follows that the graph that achieves the maximum
number of edges is Tn,r .

Our third and final proof uses a technique called the probabilistic
method. In this method, one introduces randomness to a determinis-
tic problem in a clever way to obtain deterministic results.

Proof 3. Let G = (V, E) be an n-vertex, Kr+1 -free graph. Consider a

uniform random ordering σ of the vertices. Let

X = {v ∈ V : v is adjacent to all earlier vertices in σ}.

Observe that the set of vertices in X form a clique. Since the permuta-
tion was chosen uniformly at random, we have

1
P(v ∈ X ) = P(v appears before all non-neighbors) = .
n − d(v)

Therefore,

1 convexity n
r ≥ E| X | = ∑ P( v ∈ X ) = ∑ n − d(v)
≥
n − 2m/n
.
v ∈V v ∈V
  2
Rearranging gives m ≤ 1 − 1r n2 (a bound that is already good for
most purposes). Note that if n is divisible by r, then the bound imme-
diately gives a proof of Turán’s theorem. When n is not divisible by
r, one needs to a bit more work and use convexity to argue that the
d(v) should be as close as possible. We omit the details.

2.3 Hypergraph Turán problem

The short proofs given in the previous sections make problems in

extremal graph theory seem deceptively simple. In reality, many
generalizations of what we just discussed remain wide open.
Here we discuss one notorous open problem that is a hypergraph
generalization of Mantel/Turán.
An r-uniform hypergraph consists of a vertex set V and an edge
set, where every edge is now an r-element subset of V. Graphs corre-
spond to r = 2.

Question 2.7. What is the maximum number of triples in an n vertex

3-uniform hypergraph without a tetrahedron?

Turán proposed the following construction, which is conjectured to

be optimal.
 forbidding subgraphs 27

Example 2.8 (Turán). Let V be a set of n vertices. Partition V into

3 (roughly) equal sets V1 , V2 , V3 . Add a triple { x, y, z} to e( G ) if it
satisfies one of the four following conditions:

• x, y, z are in different partitions

• x, y ∈ V1 and z ∈ V2

• x, y ∈ V2 and z ∈ V3

• x, y ∈ V3 and z ∈ V1

where we consider x, y, z up to permutation (See Example 2.8). One

checks that the 3-uniform hypergraph constructed is tetrahedron-free,
and that it has edge density 5/9.

On the other hand, the best known upper bound is approximately

0.562 , obtained recently using the technique of flag algebras.
Turán’s construction of a tetrahedron-
free 3-uniform hypergraph
2.4 Erdős–Stone–Simonovits theorem (statement): forbidding a
Keevash (2011)
general subgraph Baber and Talbot (2011)
Razborov (2010)
One might also wonder what happens if Kr+1 in Theorem 2.6 were
replaced with an arbitrary graph H:

Question 2.9. Fix some graph H. If G is an n vertex graph in which

Notice that we only require H to be a
H does not appear as a subgraph, what is the maximum possible subgraph, not necessarily an induced
number of edges in G? subgraph. An induced subgraph H 0
of G must contain all edges present
between the vertices of H 0 , while there
Definition 2.10. For a graph H and n ∈ N, define ex(n, H ) to be the is no such restriction for arbitrary
subgraphs.
maximum number of edges in an n-vertex H-free graph.

For example, Theorem 2.6 tells us that for any given r,

  
1 n
ex(n, Kr+1 ) = e( Tn,r ) = 1 − + o (1)
r 2

where o (1) represents some quantity that goes to zero as n → ∞.

At a first glance, one might not expect a clean answer to Ques-
tion 2.9. Indeed, the solution would seem to depend on various char-
acteristics of H (for example, its diameter or maximum degree). Sur-
prisingly, it turns out that a single parameter, the chromatic number
of H, governs the growth of ex(n, H ).

Definition 2.11. The chromatic number of a graph G, denoted χ( G ),

is the minimal number of colors needed to color the vertices of G
such that no two adjacent vertices have the same color.

Example 2.12. χ(Kr+1 ) = r + 1 and χ( Tn,r ) = r.

28 k ővári–sós–turán theorem: forbidding a complete bipartite graph

Observe that if H ⊆ G, then χ( H ) ≤ χ( G ). Indeed, any proper

coloring of G restricts to a proper coloring of H. From this, we gather
that if χ( H ) = r + 1, then Tn,r is H-free. Therefore,
  
1 n
ex(n, H ) ≥ e( Tn,r ) = 1 − + o (1) .
r 2
Is this the best we can do? The answer turns out to be affirmative.
Theorem 2.13 (Erdős–Stone–Simonovits). For all graphs H, we have Erdős and Stone (1946)
Erdős and Simonovits (1966)
ex(n, H ) 1
lim = 1− .
n→∞ (n2 ) χ( H ) − 1
We’ll skip the proof for now.
Remark 2.14. Later in the book we will show how to deduce Theo-
rem 2.13 from Theorem 2.6 using the Szemerédi regularity lemma.
Example 2.15. When H = K3 , Theorem 2.13 tells us that
ex(n, H ) 1
lim = ,
n→∞ (n2 ) 2
in agreement with Theorem 2.6.
When H = K4 , we get
ex(n, H ) 2
lim n = ,
n→∞ (2) 3
also in agreement with Theorem 2.6.
When H is the Peterson graph, Theorem 2.13 tells us that
ex(n, H ) 1 2
lim n = ,
n→∞ (2) 2 1
1 1
which is the same answer as for H = K3 ! This is surprising since the 3 2

Peterson graph seems much more complicated than the triangle.

3 2

2 3
2.5 Kővári–Sós–Turán theorem: forbidding a complete bipartite The Peterson graph with a proper
graph 3-coloring.

The Erdős–Stone–Simonovits Theorem (Theorem 2.13) gives a first-

order approximation of ex(n, H ) when χ( H ) > 2. Unfortunately,
Theorem 2.13 does not tell us the whole story. When χ( H ) = 2, i.e.
H is bipartite, the theorem implies that ex(n, H ) = o (n2 ), which com-
pels us to ask if we may obtain more precise bounds. For example, if
we write ex(n, H ) as a function of n, what its growth with respect to
n? This is an open problem for most bipartite graphs (for example,
K4,4 ) and the focus of the remainder of the chapter.
Let Ks,t be the complete bipartite graph where the two parts of the
bipartite graph have s and t vertices respectively. In this section, we
consider ex(n, Ks,t ), and seek to answer the following main question:

An example of a complete bipartite

graph K3,5 .
 forbidding subgraphs 29

Question 2.16 (Zarankiewicz problem). For some r, s ≥ 1, what is

the maximum number of edges in an n-vertex graph which does not
contain Ks,t as a subgraph.
Every bipartite graph H is a subgraph of some complete bipartite
graph Ks,t . If H ⊆ Ks,t , then ex(n, H ) ≤ ex(n, Ks,t ). Therefore, by
understanding the upper bound on the extremal number of complete
bipartite graphs, we obtain an upper bound on the extremal number
of general bipartite graphs as well. Later, we will give improved
bounds for several specific biparite graphs.
Kővári, Sós and Turán gave an upper bound on Ks,t :
Theorem 2.17 (Kővári–Sós–Turán). For every integers 1 ≤ s ≤ t, there Kővári, Sós, and Turán (1954)
exists some constant C, such that
1
ex(n, Ks,t ) ≤ Cn2− s .
There is an easy way to remember
the name of this theorem: “KST”, the
Proof. Let G be a Ks,t -free n-vertex graph with m edges. initials of the authors, is also the letters
First, we repeatedly remove all vertices v ∈ V ( G ) where d(v) < for the complete bipartite graph Ks,t .

s − 1. Since we only remove at most (s − 2)n edges this way, it suffices

to prove the theorem assuming that all vertices have degree at least
s − 1.
We denote the number of copies of Ks,1 in G as #Ks,1 . The proof
establishes an upper bound and a lower bound on #Ks,1 , and then
gets a bound on m by combining the upper bound and the lower
bound.
Since Ks,1 is a complete bipartite graph, we can call the side with s
vertices the ‘left side‘, and the side with 1 vertices the ‘right side‘.
On the one hand, we can count #Ks,1 by enumerating the ‘left side‘.
For any subset of s vertices, the number of Ks,1 where these s vertices
form the ‘left side‘ is exactly the number of common neighbors of
these s vertices. Since G is Ks,t -free, the number of common neigh-
bors of any subset of s vertices is at most t − 1. Thus, we establish
that #Ks,1 ≤ (ns)(t − 1).
On the other hand, for each vertex v ∈ V ( G ), the number of copies
of Ks,1 where v is the ‘right side‘ is exactly (d(sv)). Therefore, Here we regard ( xs) as a degree s poly-
1 nomial in x, so it makes sense for x to
∑ v ∈V ( G ) d ( v ) be non-integers. The function ( xs) is
    
d(v) 2m/n
#Ks,1 = ∑ ≥n n =n , convex when x ≥ s − 1.
v ∈V ( G )
s s s

where the inequality step uses the convextiy of x 7→ ( xs).

Combining the upper bound and lower bound of #Ks,1 , we obtain
n x
that n(2m/n s ) ≤ ( s )(
t − 1). For constant s, we can use ( s ) = (1 +
s s
o (1)) ns! to get n 2m
n ≤ (1 + o (1))ns (t − 1). The above inequality
simplifies to  
1 1
m≤ + o (1) (t − 1)1/s n2− s .
2
30 k ővári–sós–turán theorem: forbidding a complete bipartite graph

Let us discuss a geometric application of Theorem 2.17.

Question 2.18 (Unit distance problem). What is the maximum num- Erdős (1946)
ber of unit distances formed by n points in R2 ?

For small values of n, we precisely know the answer to the unit

n=3 n=4 n=5
distance problem. The best configurations are listed in Figure 2.1.
It is possible to generalize some of these constructions to arbitrary
n.

• A line graph has (n − 1) unit distances. n=6 n=7

Figure 2.1: The configurations of points
··· for small values of n with maximum
number of unit distances. The edges
between vertices mean that the distance
• A chain of triangles has (2n − 3) unit distances for n ≥ 3. is 1. These constructions are unique up
··· to isomorphism except when n = 6.

• There is also a recursive construction. Given a configuration P P

P0
1
with n/2 points that have f (n/2) unit distances, we can copy P
and translate it by an arbitrary unit vector to get P0 . The configu-
ration P ∪ P0 have at least 2 f (n/2) + n/2 unit distances. We can
solve the recursion to get f (n) = Ω(n log n).

The current best lower bound on the maximum number of unit dis-
tances is given by Erdős.

Proposition 2.19. There exists a set of n points in R2 that have at least Erdős (1946)
n1+c/ log log n unit distances for some constant c.
√ √
Proof sketch. Consider a square grid with b nc × b nc vertices. We
√
can scale the graph arbitrarily so that r becomes the unit distance
for some integer r. We can pick r so that r can be represented as a
sum of two squares in many different ways. One candidate of such
r is a product of many primes that are congruent to 1 module 4. We
can use some number-theoretical theorems to analyze the best r, and
get the n1+c/ log log n bound.
Figure 2.2: An example grid graph
Theorem 2.17 can be used to prove an upper bound on the number where n = 25 and r = 10.
of unit distances.

Theorem 2.20. Every set of n points in R2 has at most O(n3/2 ) unit

distances.

Proof. Given any set of points S ⊂ R2 , we can create the unit distance
graph G as follows:

• The vertex set of G is S,

• For any point p, q where d( p, q) = 1, we add an edge between p

and q.
 forbidding subgraphs 31

p q

r=1

The graph G is K2,3 -free since for every pair of points p, q, there are at
Figure 2.3: Two vertices p, q can have at
most 2 points that have unit distances to both of them. By applying most two common neighbors in the unit
Theorem 2.17, we obtain that e( G ) = O(n3/2 ). distance graph.

Remark 2.21. The best known upper bound on the number of unit Spencer, Szemerédi and Trotter (1984)
distances is O(n4/3 ). The proof is a nice application of the crossing
number inequality which will be introduced later in this book.
Here is another problem that is strongly related to the unit dis-
tance problem:
Question 2.22 (Distinct distance problem). What is the minimum
number of distinct distances formed by n points in R2 ?
Example 2.23. Consider n points on the x-axis where the i-th point
has coordinate (i, 0). The number of distinct distances for these
points is n − 1.
The current best construction for minimum number of distinct
√
distances is also the grid graph. Consider a square grid with b nc ×
√
b nc vertices. Possible distances between two vertices are numbers
that can be expressed as a sum of the squares of two numbers that
√
are at most b nc. Using number-theoretical methods, we can obtain
that the number of such distances: Θ(n/ log n).
p

The maximum number of unit distances is also the maximum

number that each distance can occur. Therefore, we have the follow-
ing relationship between distinct distances and unit distances:
(n2 )
#distinct distances ≥ .
max #unit distances
If we apply Theorem 2.20 to the above inequality, we immediately get
an Ω(n0.5 ) lower bound for the number of distinct distances. Many
mathematicians successively improved the exponent in this lower
bound over the span of seven decades. Recently, Guth and Katz gave
the following celebrated theorem, which almost matches the upper
p
bound (only off by an O( log n)) factor).
Theorem 2.24 (Guth–Katz). Every set of n points in R2 has at least Guth and Katz (2015)
cn/ log n distinct distances for some constant c.
The proof of Theorem 2.24 is quite sophisticated: it uses tools
ranging from polynomial method to algebraic geometry. We won’t
cover it in this book.

2.6 Lower bounds: randomized constructions

It is conjectured that the bound proven in Theorem 2.17 is tight. In

other words, ex(n, Ks,t ) = Θ(n2−1/s ). Although this still remains
32 lower bounds: randomized constructions

open for arbitrary Ks,t , it is already proven for a few small cases,
and in cases where t is way larger than s. In this and the next two
sections, we will show techniques for constructing H-free graphs.
Here are the three main types of constructions that we will cover:

• Randomized construction. This method is powerful and general,

but introducing randomness means that the constructions are
usually not tight.

• Algebraic construction. This method uses tools in number theory

or algebra to assist construction. It gives tighter results, but they
are usually ‘magical’, and only works in a small set of cases.

• Randomized algebraic construction. This method is the hybrid of

the two methods above and combines the advantages of both.

This section will focus on randomized constructions. We start with a

general lower bound for extremal numbers.
Theorem 2.25. For any graph H with at least 2 edges, there exists a con-
stant c > 0, such that for any n ∈ N, there exists an H-free graph on n
v( H )−2
2− e( H )−1
vertices with at least cn edges. In other words,
v( H )−2
2− e( H )−1
ex(n, H ) ≥ cn .

Proof. The idea is to use the alteration method: we can construct a

graph that has few copies of H in it, and delete one edge from each
copy to eliminate the occurrences of H. The random graph G (n, p) is called
the Erdős–Rényi random graph, which
Consider G = G (n, p) as a random graph with n vertices where
appears in many randomized construc-
each edge appears with probability p (p to be determined). Let #H be tions.
the number of copies of H in G. Then,
n ( n − 1) · · · ( n − v ( H ) + 1) e ( H )
E[#H ] = p ≤ pe( H ) nv( H ) ,
| Aut( H )|
where Aut( H ) is the automorphism group of graph H, and
 
n
E[e( G )] = p .
2
− ve((H
H )−2
Let p = 12 n )−1 , chosen so that
1
E[#H ] ≤ E[e( G )],
2
which further implies
 
1 n 1 2− v( H )−2
E[e( G ) − #H ] ≥ p ≥ n e( H)−1 .
2 2 16
Thus, there exists a graph G, such that the value of (e( G ) − #H ) is
at least the expectation. Remove one edge from each copy of H in G,
and we get an H-free graph with enough edges.
 forbidding subgraphs 33

Remark 2.26. For example, if H is the following graph

then applying Theorem 2.25 directly gives

ex(n, H ) & n11/7 .

However, if we forbid H’s subgraph K4 instead (forbidding a sub-

graph will automatically forbid the original graph), Theorem 2.25
actually gives us a better bound:

ex(n, H ) ≥ ex(n, K4 ) & n8/5 .

For a general H, we apply Theorem 2.25 to the subgraph of H with

the maximum (e − 1)/(v − 2) value. For this purpose, define the
2-density of H as

e( H 0 ) − 1
m2 ( H ) := max .
H0 ⊆H v( H 0 ) − 2
v( H 0 )≥3

We have the following corollary.

Corollary 2.27. For any graph H with at least two edges, there exists
constant c = c H > 0 such that

ex(n, H ) ≥ cn2−1/m2 ( H ) .

Example 2.28. We present some specific examples of Theorem 2.25.

This lower bound, combined with the upper bound from the Kővári–
Sós–Turán theorem (Theorem 2.17), gives that for every 2 ≤ s ≤ t,
s + t −2
n 2− st−1 . ex(n, Ks,t ) . n2−1/s .

When t is large compared to s, the exponents in the two bounds

above are close to each other (but never equal).
When t = s, the above bounds specialize to
2 s + t −2
n 2 − s +1 . n 2 − st−1 .. n2−1/s .

In particular, for s = 2, we obtain

n4/3 . ex(n, K2,2 ) . n3/2 .

It turns out what the upper bound is close to tight, as we show next a
different, algebraic, construction of a K2,2 -free graph.
34 lower bounds: algebraic constructions

2.7 Lower bounds: algebraic constructions

In this section, we use algebraic constructions to find Ks,t -free graphs,

for various values of (s, t), that match the upper bound in the Kővári–
Sós–Turán theorem (Theorem 2.17) up to a constant factor.
The simplest example of such an algebraic construction is the
following construction of K2,2 -free graphs with many edges.

Theorem 2.29 (Erdős–Rényi–Sós). Erdős, Rényi and Sós (1966)

 
1
ex(n, K2,2 ) ≥ − o (1) n3/2 .
2

Proof. Suppose n = p2 − 1 where p is a prime. Consider the following

graph G (called polarity graph): Why is it called a polarity graph? It
may be helpful to first think about
• V ( G ) = F2p \ {(0, 0)}, the partite version of the construction,
where one vertex set is the set of points
• E( G ) = {( x, y) ∼ ( a, b)| ax + by = 1 in F p }. of of a (projective) plane over F p , and
the other vertex set is the set of lines in
the same plane, and one has an edge
For any two distinct vertices ( a, b) 6= ( a0 , b0 ) ∈ V ( G ), there is at
between point p and line ` if p ∈ `.
most one solution (common neighbour) ( x, y) ∈ V ( G ) satisfying both This graph is C4 -free since no two lines
ax + by = 1 and a0 x + b0 y = 1. Therefore, G is K2,2 -free. intersect in two distinct points.
The construction in the proof of
Moreover, every vertex has degree p or p − 1, so the total number Theorem 2.29 has one vertex set that
of edges identifies points with lines. This duality
pairing between points and lines
   
1 1
e( G ) = − o (1) p3 = − o (1) n3/2 , is known in projective geometry a
2 2 polarity.
which concludes our proof. Most vertices have degree p because
the equation ax + by = 1 has exactly p
If n does not have the form p2 − 1 for some prime, then we let p solutions ( x, y). Sometimes we have to
be the largest prime such that p2 − 1 ≤ n. Then p = (1 − o (1)n subtract 1 because one of the solutions
and constructing the same graph G p2 −1 with n − p2 + 1 isolated might be ( a, b) itself, which forms a
self-loop.
vertices. Here we use that the smallest prime
greater than n has size n + o (n). The
A natural question to ask here is whether the construction above best result of this form says that there
can be generalized. The next construction gives us a construction for exists a prime in the interval [n −
n0.525 , n] for every sufficiently large n.
K3,3 -free graphs.
Baker, Harman and Pintz (2001)
Theorem 2.30 (Brown). Brown (1966)
  It is known that the constant 1/2 in
1 Theorem 2.30 is the best constant
ex(n, K3,3 ) ≥ − o (1) n5/3 possible.
2

Proof sketch. Let n = p3 where p is a prime. Consider the following

graph G:

• V ( G ) = F3p

• E( G ) = {( x, y, z) ∼ ( a, b, c)|( a − x )2 + (b − y)2 + (c − z)2 =

u in F p }, where u is some carefully-chosen fixed nonzero element
in F p
 forbidding subgraphs 35

One needs to check that it is possible to choose u so that the above

graph is K3,3 . We omit the proof but give some intuition. Had we
used points in R3 instead of F3p , the K3,3 -freeness is equivalent to the
statement that three unit spheres have at most two common points.
This statement about unit spheres in R3 , and it can be proved rigor-
ously by some algebraic manipulation. One would carry out a similar
algebraic manipulation over F p to verify that the graph above is K3,3 -
free.
Moreover, each vertex has degree around p2 since the distribution
of ( a − x )2 + (b − y)2 + (c − z)2 is almost uniform across F p as ( x, y, z)
varies randomly over F3p , and so we expect roughly a 1/p fraction of
( x, y, z) to have ( a − x )2 + (b − y)2 + (c − z)2 = u. Again we omit the
details.

Although the case of K2,2 and K3,3 are fully solved, the correspond-
ing problem for K4,4 is a central open problem in extremal graph
theory.

Open problem 2.31. What is the order of growth of ex(n, K4,4 )? Is it

Θ(n7/4 ), matching the upper bound in Theorem 2.17?

We have obtained the Kővári–Sós–Turán bound up to a constant

factor for K2,2 and K3,3 . Now we present a construction that matches
the Kővári–Sós–Turán bound for Ks,t whenever t is sufficiently large
compared to s.

Theorem 2.32 (Alon, Kollár, Rónyai, Szabó). If t ≥ (s − 1)! + 1 then Kollár, Rónyai, and Szabó (1996)
Alon, Rónyai, and Szabó (1999)
1
ex (n, Ks,t ) = Θ(n2− s ).

We begin by proving a weaker version for t ≥ s! + 1. This will be

similar in spirit and later we will make an adjustment to achieve the
desired bound. Take a prime p and n = ps with s ≥ 2. Consider the Notice that we said the image of N
lies in F p rather than F ps . We can
norm map N : F ps → F p defined by
easily check this is indeed the case as
p s −1 N ( x ) p = N ( x ).
2 s −1
N (x) = x · x p · x p · · · x p =x p −1 .

Define the graph NormGraph p,s = (V, E) with

V = F ps and E = {{ a, b}| a 6= b, N ( a + b) = 1}.

Proposition 2.33. In NormGraph p,s defined as above, letting n = ps be the

number of vertices,
1 1
| E | ≥ n 2− s .
2
Proof. Since F× s
ps is a cyclic group of order p − 1 we know that

ps − 1
|{ x ∈ F ps | N ( x ) = 1}| = .
p−1
36 lower bounds: algebraic constructions

Thus for every vertex x (the minus one accounts for vertices with
N ( x + x ) = 1)

ps − 1 1
deg( x ) ≥ − 1 ≥ p s −1 = n 1− s .
p−1

This gives us the desired lower bound on the number of edges.

Proposition 2.34. NormGraph p,s is Ks,s!+1 -free.

We wish to upper bound the number of common neighbors to a

set of s vertices. We quote without proof the following result, which
can be proved using algebraic geometry.

Theorem 2.35. Let F be any field and aij , bi ∈ F such that aij 6= ai0 j for all Kollár, Rónyai, and Szabó (1996)
i 6= i0 . Then the system of equations

( x1 − a11 )( x2 − a12 ) · · · ( xs − a1s ) = b1

( x1 − a21 )( x2 − a22 ) · · · ( xs − a2s ) = b2
..
.
( x1 − as1 )( x2 − as2 ) · · · ( xs − ass ) = bs

has at most s! solutions in Fs .

Remark 2.36. Consider the special case when all the bi are 0. In this
case, since the aij are distinct for a fixed j, we are picking an i j for
which x j = ai j j . Since all the i j are distinct, this is equivalent to
picking a permutation on [s]. Therefore there are exactly s! solutions.
We can now prove Proposition 2.34.

Proof of Proposition 2.34. Consider distinct y1 , y2 , . . . , ys ∈ F ps . We

wish to bound the number of common neighbors x. We can use the
fact that in a field with characteristic p we have ( x + y) p = x p + y p to
obtain
s −1
1 = N ( x + yi ) = ( x + yi )( x + yi ) p . . . ( x + yi ) p
p s −1 p s −1
= ( x + yi )( x p + yi ) . . . ( x p + yi )

for all 1 ≤ i ≤ s. By Theorem 2.35 these s equations have at most

s! solutions in x. Notice we do in fact satisfy the hypothesis since
p p
yi = y j if and only if yi = y j in our field.

Now we introduce the adjustment to achieve the bound t ≥ (s −

1)! + 1 in Theorem 2.32. We define the graph ProjNormGraph p,s =
(V, E) with V = F ps−1 × F× p for s ≥ 3. Here n = ( p − 1) p
s−1 . Define

the edge relation as ( X, x ) ∼ (Y, y) if and only if

N ( X + Y ) = xy.
 forbidding subgraphs 37

Proposition 2.37. In ProjNormGraph p,s defined as above, letting n =

( p − 1) ps−1 denote the number of vertices,
 
1 1
| E| = − o (1 ) n 2− s .
2

Proof. It follows from that every vertex ( X, x ) has degree ps−1 − 1 =

(1 − o (1))n1−1/s since its neighbors are (Y, N ( X + Y )/x ) as Y ranges
over elements of F ps−1 over than − X.

Now that we know we have a sufficient amount of edges we just

need our graph to be Ks,(s−1)!+1 -free.

Proposition 2.38. ProjNormGraph p,s is Ks,(s−1)!+1 -free.

Proof. Once again we fix distinct (Yi , yi ) ∈ V for 1 ≤ i ≤ s and we

wish to find all common neighbors ( X, x ). Then

N ( X + Yi ) = xyi .

Assume this system has at least one solution. Then if Yi = Yj with i 6=

j we must have that yi = y j . Therefore all the Yi are distinct. For each
i < s we can take N ( X + Yi ) = xyi and divide by N ( X + Ys ) = xys to
obtain  
X + Yi y
N = i.
X + Ys ys
Dividing both sides by N (Yi − Ys ) we obtain
 
1 1 yi
N + =
X + Ys Yi − Ys N (Yi − Ys )ys

for all 1 ≤ i ≤ s − 1. Now applying Theorem 2.35 there are at most

(s − 1)! choices for X, which also determines x = N ( X + Y1 )/y1 .
Thus there are at most (s − 1)! common neighbors.

Now we are ready to prove Theorem 2.32.

Proof of Theorem 2.32. By Proposition 2.37 and Proposition 2.38 we

know that ProjNormGraph p,s is Ks,(s−1)!+1 -free and therefore Ks,t -free
  1
and has 21 − o (1) n2− s edges as desired.

2.8 Lower bounds: randomized algebraic constructions

So far we have seen both constructions using random graphs and

algebraic constructions. In this section we present an alternative
1
construction of Ks,t -free graphs due to Bukh with Θ(n2− s ) edges pro- Bukh (2015)
vided t > t0 (s) for some function t0 . This is an algebraic construction
with some randomness added to it.
38 lower bounds: randomized algebraic constructions

First fix s ≥ 4 and take a prime power q. Let d = s2 − s + 2 and

f ∈ Fq [ x1 , x2 , . . . , xs , y1 , y2 , . . . , ys ] be a polynomial chosen uniformly
at random among all polynomials with degree at most d in each
of X = ( x1 , x2 , . . . , xs ) and Y = (y1 , y2 , . . . , ys ). Take G bipartite
with vertex parts n = L = R = Fsq and define the edge relation as
( X, Y ) ∈ L × R when f ( X, Y ) = 0.
Lemma 2.39. For all u, v ∈ Fsq and f chosen randomly as above

1
P[ f (u, v) = 0] = .
q
Proof. Notice that if g is a uniformly random constant in Fq , then
f (u, v) and f (u, v) + g are identically distributed. Hence each of the q
possibilities are equally likely to the probability is 1/q.

Now the expected number of edges is the order we want as

2
E[e( G )] = nq . All that we need is for the number of copies of Ks,t
to be relatively low. In order to do so, we must answer the follow-
ing question. For a set of vertices in L of size s, how many common
neighbors can it have?
√
Lemma 2.40. Suppose r, s ≤ min( q, d) and U, V ⊂ Fsq with |U | = s
and |V | = r. Furthermore let f ∈ Fq [ x1 , x2 , . . . , xs , y1 , y2 , . . . , ys ] be a
polynomial chosen uniformly at random among all polynomials with degree
at most d in each of X = ( x1 , x2 , . . . , xs ) and Y = (y1 , y2 , . . . , ys ). Then

P[ f (u, v) = 0 for all u ∈ U, v ∈ V ] = q−sr .

Proof. First let us consider the special case where the first coordinates
of points in U and V are all distinct. Define

∑
j
g( X1 , Y1 ) = aij X1i Y1
0≤ i ≤ s −1
0≤ j ≤r −1

with aij each uniform iid random variables over Fq . We know that
f and f + g have the same distribution, so it suffices to show for
all buv ∈ Fq where u ∈ U and v ∈ V there exists aij for which
g(u, v) = buv for all u ∈ U, v ∈ V. The idea is to apply Lagrange
Interpolation twice. First for all u ∈ U we can find a single variable
polynomial gu (Y1 ) with degree at most r − 1 such that gu (v) = buv
for all v ∈ V. Then we can view g( X1 , Y1 ) as a polynomial in Y1 with
coefficients being polynomials in X1 , i.e.,

∑
j
g( X1 , Y1 ) = a j ( X1 )Y1 .
0≤ j ≤r −1

Applying the Lagrange interpolation theorem for a second time

we can find polynomials a0 , a1 , . . . , ar−1 such that for all u ∈ U,
g(u, Y1 ) = gu (Y1 ) as polynomials in Y1 .
 forbidding subgraphs 39

Now suppose the first coordinates are not necessarily distinct. It

suffices to find linear maps T, S : Fsq → Fsq such that TU and SV have
all their first coordinates different. Let us prove that such a map T
exists. If we find a linear map T1 : Fsq → Fq that sends the elements of
U to distinct elements, then we can extend T1 to T by using T1 for the
first coordinate. To find T1 pick T1 uniformly among all linear maps.
Then for every pair in U the probability of collision is 1q . So by union
bounding we have the probability of success is at least 1 − (|U2 |) 1q > 0,
so such a map T exists. Similarly S exists.

Fix U ⊂ Fsq with |U | = s. We wish to upper bound the number

of instances of U having many common neighbors. In order to do
this, we will use the method of moments. Let I (v) represent the
indicator variable which is 1 exactly when v is a common neighbor
of U and set X to be the number of common neighbors of U. Then
using Lemma 2.40,

E[ X d ] = E[( ∑s I (v))d ] = ∑ E[ I (v1 ) · · · I (vd )]

v ∈Fq v1 ,...,vd ∈Fsq
 s
q
= ∑ q−rs Mr ≤ ∑ Mr = M,
r ≤d
r r ≤d

where Mr is defined as the number of surjections from [d] to [r ] and

M = ∑r≤d Mr . Using Markov’s inequality we get

E[ X d ] M
P( X ≥ λ ) ≤ ≤ d.
λd λ
Now even if the expectation of X is low, we cannot be certain that
the probability of X being large is low. For example if we took the
1
random graph with p = n− s then X will have low expectation but
a long, smooth-decaying tail and therefore it is likely that X will be
large for some U.
It turns out what algebraic geometry prevents the number of com-
mon neighbors X from taking arbitrary values. The common neigh-
bors are determined by the zeros of a set of polynomial equations,
and hence form an algebraic variety. The intuition is that either we
are in a “zero-dimensional” case where X is very small or a “positive
dimensional” case where X is at least on the order of q.

Lemma 2.41. For all s, d there exists a constant C such that if f 1 (Y ), . . . , f s (Y ) Bukh (2015)
are polynomials on Fsq of degree at most d then

{y ∈ Fsq | f 1 (y) = . . . f s (y) = 0}

√
has size either at most C at least q − C q.
40 forbidding a sparse bipartite graph

The lemma can be deduced from the following important result

from algebraic geometry known as the Lang–Weil bound, which says
that the number of points of an r-dimensional algebraic variety in Fsq
is roughly qr , as long as certain irreducibility hypotheses are satisfied.
s
Theorem 2.42 (Lang–Weil bound). If V = {y ∈ Fq | g1 (y) = g2 (y) = Lang and Weil (1954)
. . . = gm (y)} is irreducible and gi has degree at most d, then
1
|V ∩ Fsq | = qdim V (1 + Os,m,d (q− 2 )).

Now we can use our bound from Markov’s Inequality along with
Lemma 2.41. Let the s polynomials f 1 (Y ), . . . , f s (Y ) in Lemma 2.41 be
the s polynomials f (u, Y ) as u ranges over the s elements of U. Then
for large enough q there exists a constant C from Lemma 2.41 such
√
that having X > C would imply X ≥ q − C q > q/2, so that
 q M
P( X > C ) = P X > ≤ .
2 (q/2)d

Thus the number of subsets of L or R with size s and more than C

common neighbors is at most
 
n M
2 = O ( q s −2 )
s (q/2)d

in expectation. Take G and remove a vertex from every such subset to

create G 0 . First we have that G 0 is Ks,C+1 -free. Then

n2 n2 1
E[e( G 0 )] ≥ − O(nqs−2 ) = (1 − o (1)) = (1 − o (1))n2− s
q q

and v( G 0 ) ≤ 2n. So there exists an instance of G 0 that obtains the

desired bound.

2.9 Forbidding a sparse bipartite graph

For any bipartite graph H, it is always contained in Ks,t for some s, t.

Therefore by Theorem 2.17,
1
ex(n, H ) ≤ ex(n, Ks,t ) . n2− s .

The first inequality is not tight in general when H is some sparse

bipartite graph. In this section, we will see some techniques that give
a better upper bound on ex(n, H ) for sparse bipartite graphs H.
The first result we are going to see is an upper bound on ex(n, H )
when H is bipartite and the degrees of vertices in one part are
bounded above.
 forbidding subgraphs 41

Theorem 2.43. Let H be a bipartite graph whose vertex set is A ∪ B such Füredi (1991)
that every vertex in A has degree at most r. Then there exists a constant Alon, Krivelevich and Sudakov (2003)

C = CH such that
1
ex(n, H ) ≤ Cn2− r

Remark 2.44. Theorem 2.32 shows that the exponent 2 − 1r is the

best possible as function of r since we can take H = Kr,t for some
t ≤ (r − 1)! + 1.
To show this result, we introduce the following powerful proba-
bilistic technique called dependent random choice. The main idea of
this lemma is the following: if G has many edges, then there exists
a large subset U of V ( G ) such that all small subsets of vertices in U
have many common neighbors.

Lemma 2.45 (Dependent random choice). Let u, n, r, m, t ∈ N, α > 0 be Alon, Krivelevich and Sudakov (2003)
numbers that satisfy the inequality
  
n m t
nαt − ≥ u.
r n

Then every graph G with n vertices and at least αn2 /2 edges contains a
subset U of vertices with size at least u such that every r-element subset S of
U has at least m common neighbors.

Proof. Let T be a list of t vertices chosen uniformly at random from

V ( G ) with replacement (allowing repetition). Let A be the common
neighborhood of T. The expected value of | A| is

E| A | = ∑ P( v ∈ A )
v ∈V
= ∑ P(T ⊆ N (v))
v ∈V
 t
d(v)
= ∑ n
v ∈V
!t
1 d(v)
≥n
n ∑ n
(convexity)
v ∈V

≥ nαt .

For every r-element subset S of V, the event of A containing S oc-

curs if and only if T is contained in the common neighborhood of S,
which occurs with probability
 t
#common neighbors of S
.
n

Call a set S bad if it has less than m common neighbors. Then each
bad r-element subset S ⊂ V is contained in A with probability less
42 forbidding a sparse bipartite graph

than (m/n)t . Therefore by linearity of expectation,

  
n m t
E[the number bad r-element subset of A] < .
r n

To make sure that there are no bad subsets, we can get rid of one
element in each bad subset. The number of remaining elements is at
least | A| − (#bad r-element subset of A), whose expected value is at
least   
t n m t
nα − ≥ u.
r n
Consequently, there exists a T such that there are at least u elements
in A remaining after getting rid of all bad r-element subsets. The set
U of the remaining u elements satisfies the desired properties.

Setting the parameters of Lemma 2.45 to what we need for proving

Theorem 2.43, we get the following corollary.

Corollary 2.46. For any bipartite graph H with vertex set A ∪ B where
each vertex in A has degree at most r, there exists C such that the following
1
holds: Every graph with at least Cn2− r edges contains a vertex subset U
with |U | = | B| such that every r-element subset in U has at least | A| + | B|
common neighbors.

Proof. By Lemma 2.45 with u = | B|, m = | A| + | B|, and t = r, it

suffices to check that there exists C so that
| A| + | B| r
  
1 r n
 
n 2Cn− r − ≥ | B |.
r n

The first term evaluates to (2C )r , and the second term is O H (1).
Therefore we can choose C large enough to make this inequality
hold.

Now we are ready to show Theorem 2.43.

Proof of Theorem 2.43. Let G be a graph with n vertices and at least

1
Cn2− r edges, where C is chosen as in Corollary 2.46. First embed
B into V ( G ) using U from Corollary 2.46. The plan is to extend this
embedding furthermore to A ∪ B ,− → V ( G ). To do this, assume that
we have an embedding φ : A0 ∪ B ,− → V ( G ) already where A0 ⊆ A,
and we want to extend φ to an arbitrary v ∈ A\ A0 . We have to make
sure that φ(v) is a common neighbor of φ( N (v)) in G. Note that by
assumption, |φ( N (v))| = | N (v)| ≤ r, and so by the choice of B,
the set φ( N (v)) has at least | A| + | B| common neighbors. φ(v) can
then be any of those common neighbors, with an exception that φ(v)
cannot be the same as φ(u) for any other u ∈ A0 ∪ B. This eliminates
| A0 | + | B| ≤ | A| + | B| − 1 possibilities for φ(v). Since there are at least
| A| + | B| vertices to choose from, we can just extend φ by setting φ(v)
 forbidding subgraphs 43

to be one of the remaining choices. With this process, we can extend

the embedding to A ∪ B ,−→ V ( G ), which shows that there is a copy of
H in G.

This is a general result that can be applied to all bipartite graphs.

However, for some specific bipartite graph H, there could be room
for improvement. For example, from this technique, the bound we
get for C6 is the same as C4 , which is O(n3/2 ). This is nonetheless not
tight.

Theorem 2.47 (Even cycles). For all integer k ≥ 2, there exists a constant Bondy and Simonovits (1974)
C so that
1
ex(n, C2k ) ≤ Cn1+ k .
 
Remark 2.48. It is known that ex(n, C2k ) = Θ n1+1/k for k = 2, 3, 5. Benson (1966)
However, it is open whether the same holds for other values of k.
Instead of this theorem, we will show a weaker result:

Theorem 2.49. For any integer k ≥ 2, there exists a constant C so that

every graph G with n vertices and at least Cn1+1/k edges contains an even
cycle of length at most 2k.

To show this theorem, we will first “clean up" the graph so that
the minimum degree of the graph is large enough, and also the graph
is bipartite. The following two lemmas will allow us to focus on a
subgraph of G that satisfies those nice properties.

Lemma 2.50. Let t ∈ R and G a graph with average degree 2t. Then G
contains a subgraph with minimum degree greater than r.

Proof. We have e( G ) = v( G )t. Removing a vertex of degree at most

t cannot decrease the average degree. We can keep removing vertices
of degree at most t until every vertex has degree more than t. This
algorithm must terminate before reaching the empty subgraph since
every graph with at most 2t vertices has average degree less than 2t.
The remaining subgraph when the algorithm terminates is then a
subgraph whose minimum degree is more than t.

Lemma 2.51. Every G has a bipartite subgraph with at least e( G )/2 edges.

Proof. Color every vertex with one of two colors uniformly at ran-
dom. Then the expected value of non-monochromatic edges is
e( G )/2. Hence there exists a coloring that has at least e( G )/2 non-
monochromatic edges.

Proof of Theorem 2.49. Suppose that G contains no even cycles of

length at most 2k. By Lemma 2.50 and Lemma 2.51 there exists a
bipartite subgraph G 0 with minimum degree at least δ := Cn1/k /2.
 44 forbidding a sparse bipartite graph

Let A0 = {u} where u is an arbitrary vertex in V ( G 0 ). Let Ai+1 =

NG0 ( Ai )\ Ai−1 . Then Ai is the set of vertices that are distance exactly i
away from the starting vertex u since G 0 is bipartite.
Now for every two different vertices v, v0 in Ai−1 for some 1 ≤
i ≤ k, if they have a common neighbor w in Ai , then there are two
different shortest paths from u to w. The union two distinct paths
(even if they overlap) contains an even-cycle of length at most 2i ≤ A0 A1 A2 A3 ··· At
2k, which is a contradiction. Therefore the common neighbors of any
Figure 2.4: Diagram for Proof of Theo-
two vertices in Ai−1 can only lie in Ai−2 , which implies that | Ai | ≥ rem 2.49
(δ − 1)| Ai−1 |. Hence | Ak | ≥ (δ − 1)k ≥ (Cn1/k − 1)k . If C is chosen
large enough then we get | Ak | > n, which is a contradiction.

If H is a bipartite graph with vertex set A ∪ B and each vertex in

A has degree at most 2, then ex(n, H ) = O(n3/2 ). The exponent 3/2
is optimal since ex(n, K2,2 ) = Θ(n3/2 ) and hence the same holds
whenever H contains K2,2 . It turns out that this exponent can be
improved whenever H does not contain any copy of K2,2 .

Theorem 2.52. Let H be a bipartite graph with vertex bipartition A ∪ B Colon and Lee (2019+)
such that each vertex in A has degree at most 2, and H does not contain
K2,2 . Then there exist c, C > 0 dependent on H such that
3
ex(n, H ) ≤ Cn 2 −c .

To prove this theorem, we show an equivalent statement for-

mulated using the notion of subdivisons. For a graph H, the 1-
subdivision H 1-sub of H is obtained by adding an extra vertex in the
middle of every edge in H. Notice that every H in the setting of The-
orem 2.52 is a subgraph of some Kt1-sub . Therefore we can consider −→
the following alternative formulation of Theorem 2.52.
K4 K41-sub
Theorem 2.53. For all t ≥ 3, there exists ct > 0 such that
Figure 2.5: 1-subdivision of K4
3 −c
ex(n, Kt1-sub ) = O(n 2 t ).

Now we present a proof of Theorem 2.53 by Janzer. As in The- Janzer (2018)

orem 2.49, it is helpful to pass the entire argument to a subgraph
where we have a better control of the degrees of the vertices. To do
so, we are going to use the following lemma (proof omitted) to find
an almost regular subgraph.

Lemma 2.54. For all 0 < α < 1, there exist constants β, k > 0 such that Colon and Lee (2019+)
for all C > 0, n sufficiently large, every n-vertex graph G with ≥ Cn1+α
edges has a subgraph G 0 such that

(a) v( G 0 ) ≥ n β ,

(b) e( G 0 ) ≥ 1 0 1+ α ,
10 Cv ( G )
 forbidding subgraphs 45

(c) max deg( G 0 ) ≤ K min deg( G 0 ),

(d) G 0 is bipartite with two parts of sizes differing by factor ≤ 2.

From now on, we treat t as a constant. For any two vertices u, v ∈

A, we say that the pair uv is light if the number of common neighbors
of u and v is at least 1 and less than (2t ); moreover, we say that the
pair uv is heavy if the number of common neighbors of u and v is at
least (2t ). Note that pairs u, v ∈ A without any common neighbors are
neither light nor heavy. The following lemma gives a lower bound on
the number of light pairs.

Lemma 2.55. Let G be a Kt1-sub -free bipartite graph with bipartition U ∪ B,

d( x ) ≥ δ for all x ∈ U, and |U | ≥ 4| B|t/δ. Then there exists u ∈ U in
Ω(δ2 |U |/| B|) light pairs in U.

Proof. Let S be the set of {({u, v}, x )|u, v ∈ U, x ∈ B} where {u, v}

is an unordered pair of vertices in U and x is a common neighbor of
{u, v}. We can count this by choosing x ∈ B first:

2
e( G )/| B| | B| δ |U | δ 2 |U | 2
    
d( x )
|S| = ∑ 2
≥ | B|
2
≥
4 | B|
=
4| B |
.
x∈B

Notice that the low-degree vertices in B contributes very little since

δ 2 |U | 2
 
d( x )
∑ 2
≤ 2t2 | B| ≤
8| B |
.
x∈B
d( x )<2t

Therefore

δ 2 |U | 2
 
d( x )
∑ 2
≥
8| B |
.
x∈B
d( x )≥2t

Note that if there are t mutually heavy vertices in U, then we can

choose a common neighbor uij for every pair {vi , v j } with i < j. Since
there are at least (2t ) such neighbors for each pair {vi , v j }, one can
make choices so that all uij are distinct. This then produces a Kt1-sub
subgraph, which is a contradiction. Therefore there do not exist t
mutually heavy vertices in U, and by Turán’s Theorem, the number
of heavy pairs in N ( x ) for x ∈ B is at most e( Td( x),t−1 ). Since any two
vertices in N ( x ) have at least one common neighbor x, they either
form a light pair or a heavy pair. This shows that there are at least
 46 forbidding a sparse bipartite graph

(d(2x)) − e( Td(x),t−1 ) light pairs among N ( x ). If d( x ) ≥ 2t, then

 
d( x )
− e( Td(x),t−1 )
2
t−1 d( x ) 2
    
d( x )
≥ −
2 2 t−1
1 1
= d ( x )2 − d ( x )
2( t − 1) 2
& d ( x )2 .

If we sum over x ∈ B, then each light pair is only going to be counted

for at most (2t ) times according to the definition. This is constant
since we view t as a constant. Therefore
δ 2 |U | 2
#light pairs in U & ∑ d ( x )2 & | S | & | B|
,
x∈B

and by pigeon hole principle there exists a vertex u ∈ U that is in

Ω(δ2 |U |/| B|) light pairs.

With these lemmas, we are ready to prove Theorem 2.53.

Proof of Theorem 2.53. Let G be any Kt1-sub -free graph. First pick G 0 by
Lemma 2.54 with α = (t − 2)/(2t − 3), and say that the two parts
are A and B. Set δ to be the minimum degree of G 0 . We will prove
that δ ≤ Cv( G 0 )(t−2)/(2t−3) for some sufficiently large constant C by
contradiction. Suppose that δ > Cv( G 0 )(t−2)/(2t−3) . Our plan is to
pick v1 , v2 , . . . , vt such that vi v j are light for all i < j, and no three
of v1 , . . . , vt have common neighbors. This will give us a Kt1-sub and
hence a contradiction.
We will do so by repeatedly using Lemma 2.55 and induction on a
stronger hypothesis: For each 1 ≤ i ≤ t, there exists A = U1 ⊇ U2 ⊇
· · · ⊇ Ui and v j ∈ Uj such that

(a) v j is in at least Θ(δ2 |Uj |/v( G 0 )) light pairs in Uj for all 1 ≤ j ≤

i − 1,
v1
(b) v j is light to all vertices in Uj+1 for all 1 ≤ j ≤ i − 1. v2

(c) no three of v1 , . . . , vi have common neighbors,

v3

(d) |Uj+1 | & δ 2 |U j |/v( G 0 ) for all 1 ≤ j ≤ i − 1,

U3
This statement clearly holds when i = 1 by choosing v1 to be the U2
vertex found by Lemma 2.55. Now suppose that we have constructed
A = U1 ⊇ · · · ⊇ Ui−1 with v j ∈ Uj for all j = 1, . . . , i − 1. To construct U1

Ui , let Ui0 be the set of vertices that form light pairs with vi−1 . Then Figure 2.6: Repeatedly applying
|Ui0 | & δ2 |Ui−1 |/v( G 0 ) by the inductive hypothesis (a). Now we get Lemma 2.55 to obtain vi ’s and Ui ’s

rid of all the vertices in Ui0 that violate (c) to get Ui . It suffices to look
 forbidding subgraphs 47

at each pair v j vk , look at their common neighbors u and delete all

the neighbors of u from Ui0 . There are (i−2 1) choices v j vk , and they
have at most (2t ) common neighbors since they form a light pair, and
each such neighbors has degree at most Kδ. Therefore the number of
vertices removed is at most
i−1
  
t
Kδ = O(δ)
2 2

since t and K are constants. Therefore after this alteration, (d) will
still hold as long as |Ui0 | = Ω(δ) and C is chosen sufficiently large.
This is true since
 i −1
δ2

|Ui0 | & | A| & δ2t−2 V ( G 0 )t−2 = Θ(δ)
V (G0 )

given that i ≤ t. Therefore (d) holds for i, and we just need to choose
a vertex vi from Lemma 2.55 in Ui and (a), (b), (c) follow directly.
Therefore by induction, this also holds for i = t. Now by (b) and (c),
there exists a copy of Kt1-sub in G 0 , which is a contradiction.
The above argument shows that δ ≤ Cv( G 0 )(t−2)/(2t−3) , and so
the maximum degree is at most KCv( G 0 )(t−2)/(2t−3) . Hence e( G 0 ) ≤
KCv( G 0 )1+α , and by the choice of G 0 , we know that e( G ) ≤ 10KCn1+α ,
as desired.
MIT OpenCourseWare
https://ocw.mit.edu

18.217 Graph Theory and Additive Combinatorics

Fall 2019

For information about citing these materials or our Terms of Use, visit: https://ocw.mit.edu/terms.