MATH 595 Hypergraph Container Method notes

1 Lecture 8: Random Turán Theorem (2/4/19)

In this section, we shall deduce the sparse random analogue of a classical theorem by
Turán [7]. We start by asking the following question.
Question 1.1. For what values of p is it true that the maximum size of a Kr -free subgraph
1
+ o(1) p · n2 ?



of G(n, p) ≈ 1 − r−1
We first consider what happens for certain choices of p. When p = 1, then G(n, 1) = Kn , so
we can use Turán’s theorem directly. What is the situation for other values of p?
If the expected number of Kr ’s is much less than the expected number of edges in the random
graph, then we can remove one edge from each Kr and still get a large Kr -free subgraph
with ≈ (1 − o(1))p · n2 edges. For what value of p will this be the case?

Expected number of Kr ’s  Expected |E(G(n, p))|

 
n (r2)
p  pn2
r
2 −r)/2
nr p(r  pn2
p  n−2/(r+1) .

This raises the following question. Is Question 1.1 true for p > C · n−2/(r+1) ? This theorem
was already proved by Conlon and Gowers [2] and Schacht [5].
Theorem 1.2 ([2]). Given  > 0 and a strictly 2-balanced graph H, there exists a positive
constant C such that

lim P (G(n, p) is (H, ) − Turán) = 1, if p > Cn−1/m2 (H) ,

n→∞

where m2 (H) = (|E(H)| − 1)/(|V (H)| − 2).

Recall that a graph H is said to be strictly 2-balanced if, for every subgraph K ⊂ H,
|E(H)| − 1 |E(K)| − 1
> ,
|V (H)| − 2 |V (K)| − 2
so clearly Kr is strictly 2-balanced. In addition, we define a graph G to be (H, )-Turán if
every subgraph of G with at least
 
1
1− +  |E(G)|
χ(H) − 1

1
edges contains a copy of H. We prove the following version of the above theorem for H := Kr
using the Hypergraph Container Lemma.

Theorem 1.3 (Random Turán Theorem [1]). Given a Kr -free subgraph F ⊆ G(n, p), if
p > C · n−2/(r+1) , then max |F | ≈ 1 − r−1
1
+ o(1) p · n2 .

We recall the statement of the Hypergraph Container Lemma.

Theorem 1.4 (Hypergraph Container Lemma [1]). For every k ∈ N and c, ε > 0, there
exists a positive constant C for which the following holds. Let H be a k-uniform hypergraph
and F ⊆ 2V (H) be an upset such that |A| > ε · v(H) for all A ∈ F. Suppose H is (F, ε)-dense
and q ∈ (0, 1) such that for every ` ∈ [k],

e(H)
∆` (H) 6 c · q `−1 .
v(H)
V (H)


Then there exists a family S ⊆ 6Cq·v(H) and functions f : S → F and g : I(H) → S such
that for every I ∈ I(H),
g(I) ⊆ I and I ⊆ f (g(I)) ∪ g(I).

Proof of Theorem 1.3. We build the hypergraph Hn starting from Kn and the copies of Kr
in that complete graph. Define V (Hn ) := E(Kn ) and E(Hn ) := the edge sets of the copies
of Kr . Note that |V (Hn )| = n2 and |E(Hn )| = nr . Finally, if we let k = 2r , we know that

Hn is k-uniform.

To apply the Hypergraph Container Lemma, we need to find the number q that bounds the
codegrees of the hypergraphs with the following proposition.
Proposition 1.5. Let n be an integer and let H be a 2-uniform hypergraph. Set k = e(H)
and let H be the k-uniform hypergraph of copies of H in Kn . There exists a positive constant
c such that, letting q = n−1/m2 (H) ,

e(H)
∆` (H) 6 c · q `−1
v(H)

for every ` ∈ [k].

v(H)!
Proof of Proposition 1.5. Note that v(H) = n2 = Θ(n2 ) and that e(H) = n



|Aut(H)|
· v(H)
=
Θ(nv(H) ). By the definition of q and m2 (H), we have
0 0
q e(H )−1 nv(H )−2 > 1

2
for every H 0 ⊆ H. Now, for each ` ∈ [k],
n 0
o
∆` (H) 6 c0 · max nv(H)−v(H ) : H 0 ⊆ H with e(H 0 ) = `

for some positive constant c0 . Since e(H)/v(H) > c00 · nv(H)−2 for some constant c00 , it follows
that
−1
nv(H)
  
`−1 e(H) 0 v(H)
∆` (H) · q 6c · · max
v(H) e(H) H 0 ⊆H : e(H 0 )=2 q e(H 0 )−1 nv(H 0 )
c0 c0
 
1 v(H 0 )−2
6 00 · 0 max 0 n 6 ,
c H ⊆H : e(H )=2 q e(H 0 )−1 c00
0 0
where the last inequality follows since q e(H )−1 nv(H )−2 > 1.

So we let q = n−1/m2 (Kr ) = n−2/r+1 . We now compute the codegrees of the hypergraphs.
That is, we find how many copies of Kr will contain any fixed ` vertices. We have that
 
n−t
∆` (Hn ) 6 ,
r−t
where we choose t to be the minimum such that ` 6 2t . Note that it is always the case that

1 6 ` 6 2r . So, for ` = 1, we need t = 2, which implies ∆1 (Hn ) 6 n−2

r−2
≈ nr−2 . Is there a
|E(Hn )|
= Cr · nr / n2 ? In this case, the answer



constant Cr such that ∆1 (Hn ) 6 nr−2 6 Cr |V (Hn )|
is yes. Therefore, there is no restriction on q.

Consider another case. When ` = 2r , then t = r. That is, ∆` (Hn ) = 1, and we want

1 6 Cr · q `−1 nr−2 . Therefore, we choose q = n−2/(r+1) by Proposition 1.5.

We may now apply the Hypergraph Container Lemma. This implies that there exists some
S ⊆ V (Hn ) such that |S| 6 C · n2−2/(r+1) . For such S, there exists an associated graph f (S)
with at most  nr copies of Kr . We apply the following supersaturation argument (with

stability) to f (S).

For small γ > 0 and β > 0, then |f (S)| 6 (1 − 1/(r − 1) + β) n2 by supersaturation (i.e. if

the number of Kr ’s in a graph is at most q nr , then the graph can only be a small, constant

proportion bigger than Turán’s bound). By stability, either |f (S)| 6 (1 − 1/(r − 1) − β) n2

or, by removing at most γn2 edges from f (S), the graph can be made (r − 1)-partite.

Consider the counting statement which follows from the Hypergraph Container Lemma: the
number of Kr -free graphs is at most (the number of choices of S)·2max |f (S)| . We obtain the

3
following stability version of the same counting statement: almost all Kr -free graphs are
‘almost’ (r − 1)-partite. So we need to prove that the number of Kr -free graphs that are
n
not (r − 1)-partite  (# of choices of S) · 2(1−1/(r−1))( 2 ) . This is, in fact true, proven earlier
using the Regularity Lemma.

1
) · p · n2 , where F is Kr -free. What we


Now let F ⊆ G(n, p) with |E(F )| > (1 + γ)(1 − r−1
hope is for the number of such subgraphs F to be small. First note that all such subsets
form an independent set in the hypergraph Hn . Therefore, there exists S ⊆ F and f (S)
such that F ⊆ S ∪ f (S).

1 n


We know that |f (S)| < (1 − r−1
+ o(1)) 2
because it is a container. Therefore
  
1 n
E(|G(n, p) ∩ f (S)|) < p · (1 − + o(1)) .
r−1 2

By Chernoff bound, we obtain the following.

 2
n2
     
1 n γ
P |G(n, p) ∩ f (S)| < (1 + γ/2) 1 − ·p· < exp − · p · .
r−1 2 8 2

Since F ⊆ S ∪ f (S), F ⊆ G(n, p), and F has more edges than the expected size of
G(n, p) ∩ f (S), we get |F ∩ f (S) ∩ G(n, p)| > (1 + γ/2) · p · (1 − 1/(r − 1)) n2 .

By the first approach, we get that

n2 /2
  2 2
0 2− r+1 2− r+1
number of choices for S = 2
2− r+1
· eC n ≈ 2log n·n .
C ·n
We want this number to be o(1). We apply the counting lemma to show that this is true for
the desired p.
!!!!!MISSING COMPUTATION!!!!! TO BE DONE AT ONE POINT!!!!!!
For a second approach, we consider S ⊆ G(n, p) and count the number of choices for S by
summing over the possible sizes of S as follows.
 2 
|S| n /2
2
0 2− r+1
X
p · eC n ,
2
|S|
2− r+1
|S|6C·n

where again, the desired p will make this sum equal to o(1).

4
References
[1] J. Balogh, R. Morris and W. Samotij, Independent sets in hypergraphs, J. Amer. Math.
Soc. 28 (2015), 669–709.

[2] D. Conlon and W. T. Gowers, Combinatorial theorems in sparse random sets, Ann.
Math., Second Series, 184 no. 2 (2016), 367–454.

[3] P. Erdős and A.H. Stone, On the structure of linear graphs, Bull. Amer. Math. Soc. 52
(1946), 1087–1091.

[4] B. Green and T. Tao, The primes contain arbitrarily long arithmetic progressions, Ann.
Math., 167 no. 2 (2008), 481–547.

[5] M. Schacht, Extremal results for random discrete structures, Ann. Math., Second Series,
184 no. 2 (2016), 333–365.

[6] E. Szemerédi, On sets of integers containing no k elements in arithmetic progression,

Acta Arith. 27 (1975), 199–234.

[7] P. Turán, Eine Extremalaufgabe aus der Graphentheorie, Mat. Fiz. Lapok 48 (1941),
436–452.

[8] P. Varnavides, On certain sets of positive density, J. London Math. Soc. 34 (1959), 358–
360.