ON K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS

Y. Kohayakawa, T. Luczak, and V. Rödl

Abstract. For 0 < γ ≤ 1 and graphs G and H, write G →γ H if any γ-proportion of

the edges of G span at least one copy of H in G. As customary, write K r for the com-
plete graph on r vertices. We show that for every fixed real η > 0 there exists a con-
stant C = C(η) such that almost every random graph Gn,p with p = p(n) ≥ Cn−2/5
satisfies Gn,p →2/3+η K 4 . The proof makes use of a variant of Szemerédi’s regu-
larity lemma for sparse graphs and is based on a certain superexponential estimate
for the number of pseudo-random tripartite graphs whose triangles are not too well
distributed. Related results and a general conjecture concerning H-free subgraphs of
random graphs in the spirit of the Erdős–Stone–Simonovits theorem are discussed.

§0. Introduction
A classical area of extremal graph theory investigates numerical and structural
problems concerning H-free graphs, namely graphs that do not contain a copy of
a given fixed graph H as a subgraph. Let ex(n, H) be the maximal number of
edges that an H-free graph on n vertices may have. A basic question is then to
determine or estimate ex(n, H) for any given H and large n. A solution to this
problem is given by the celebrated Erdős–Stone–Simonovits theorem, which states
that, as n → ∞, we have
  
1 n
ex(n, H) = 1− + o(1) , (1)
χ(H) − 1 2

where as usual χ(H) is the chromatic number of H. Furthermore, as proved in-

dependently by Erdős and Simonovits, every H-free graph G = Gn that has as
many edges as in (1) is in fact ‘very close’ (in a certain precise sense) to the densest
n-vertex (χ(H) − 1)-partite graph. For these and related results, see, for instance,
Bollobás [2].
Here we are interested in a variant of the function ex(n, H). Let G and H be
graphs, and write ex(G, H) for the maximal number of edges that an H-free sub-
graph of G may have. Formally, ex(G, H) = max{e(J): H 6⊂ J ⊂ G}, where e(J)
stands for the size |E(J)| of J. Clearly ex(n, H) = ex(K n , H). As an example of

1991 Mathematics Subject Classification. Primary 05C80, 05C35; Secondary 05D99.

Key words and phrases. Turán’s extremal problem, forbidden subgraphs, extremal subgraphs,
cliques, random graphs.
The first author was partially supported by FAPESP (Proc. 93/0603–1) and by CNPq (Proc.
300334/93–1 and ProTeM–CC–II Project ProComb). Part of this work was done while the second
author was visiting the University of São Paulo, supported by FAPESP (Proc. 94/4276–8). The
third author was partially supported by the NSF grant DMS–9401559.

Typeset by AMS-TEX
1
2 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

a problem involving ex(G, H) with G 6= K n , let us recall that a well-known con-

jecture of Erdős states that ex(Qn , C 4 ) = (1/2 + o(1))e(Qn ), where Qn stands for
the n-dimensional hypercube and C 4 is the 4-cycle. (For several results concerning
this conjecture, see Chung [4].)
Our aim here is to study ex(G, H) when G is a ‘typical’ graph, by which
we mean
a random graph. Let 0 < p = p(n) ≤ 1 and 0 < M = M (n) ≤ N = n2 be given.
The standard binomial random graph Gp = Gn,p has as vertex set a fixed set V (Gp )
of cardinality n and two such vertices are adjacent in Gp with probability p, with all
such adjacencies independent. The random graph GM = Gn,M is simply a graph on

N
a fixed n-element vertex set V (GM ) chosen uniformly at random from all the M
possible candidates. (For concepts and results concerning random graphs not given
in detail below, see e.g. Bollobás [3].) Here we wish to investigate the random
variables ex(Gn,p , H) and ex(Gn,M , H).
Let H be a graph of order |H| = |V (H)| ≥ 3. Let us write d2 (H) for the 2-density
of H, that is  
e(J) − 1
d2 (H) = max : J ⊂ H, |J| ≥ 3 .
|J| − 2
Given a real 0 ≤ ε ≤ 1 and an integer r ≥ 2, let us say that a graph J is ε-quasi
r-partite if J may be made r-partite by the deletion of at most εe(J) of its edges.
A general conjecture concerning ex(Gn,p , H) is as follows. For simplicity, below we
restrict our attention to the binomial random graph Gn,p . Much of what follows
may be restated in terms of Gn,M . As is usual in the theory of random graphs, we
say that a property P holds almost surely or that almost every random graph Gn,p
or Gn,M satisfies P if P holds with probability tending to 1 as n → ∞.
Conjecture 1. Let H be a non-empty graph of order at least 3, and let 0 < p =
p(n) ≤ 1 be such that pn1/d2 (H) → ∞ as n → ∞. Then the following assertions
hold.
(i ) Almost every Gn,p satisfies
 
1
ex(Gn,p , H) = 1− + o(1) e(Gn,p ). (2)
χ(H) − 1

(ii ) Suppose χ(H) ≥ 3. Then for any ε > 0 there is a δ = δ(ε) > 0 such
that almost every Gn,p has the property that any H-free subgraph J ⊂ Gn,p
of Gn,p with e(J) ≥ (1 − δ) ex(Gn,p , H) is ε-quasi (χ(H) − 1)-partite.

Recall that any graph G contains an r-partite subgraph J ⊂ G with e(J) ≥

(1 − 1/r)e(G). Thus the content of Conjecture 1(i ) is that ex(Gn,p , H) is at most
as large as the right-hand side of (2), or, in other words, that Gn,p →γ H holds
almost surely for any fixed γ > 1 − 1/(χ(H) − 1). There are a few results in support
of Conjecture 1(i ).
Any result concerning the tree-universality of expanding graphs or else a simple
application of Szemerédi’s regularity lemma for sparse graphs (see Lemma 4 below)
give Conjecture 1(i ) for forests. The cases in which H = K 3 and H = C 4 are essen-
tially proved in Frankl and Rödl [5] and Füredi [6], respectively, in connection with
problems concerning the existence of some graphs with certain extremal properties.
The case in which H is a general cycle was settled by Haxell, Kohayakawa, and
Luczak [8, 9] (see also Kohayakawa, Kreuter, and Steger [11]). Conjecture 1(ii )
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 3

in the case in which 0 < p < 1 is a constant follows easily from Szemerédi’s regu-
larity lemma [15]. A variant of this lemma for sparse graphs (cf. Lemma 4 below)
and a lemma from Kohayakawa, Luczak, and Rödl [12] concerning induced sub-
graphs of bipartite graphs may be used to verify Conjecture 1 for H = K 3 in full.
(See comments following Conjecture 23 for further details.) Still in the case in
which H = K 3 , for 0 < p < 1 sufficiently close to 1/2, a much stronger result than
Conjecture 1(ii ) was proved by Babai, Simonovits, and Spencer [1]. Finally, let us
note that a result concerning Ramsey properties of random graphs in the spirit of
Conjecture 1 was proved by Rödl and Ruciński [13, 14].
Here we prove Conjecture 1(i ) for H = K 4 . Our results are as follows.
Theorem 2. For any constant 0 < η ≤ 1/3, there is a constant C = C(η) for
which the following holds. If 0 ≤ p = p(n) ≤ 1 is such that p ≥ Cn−2/5 for all large
enough n, then almost every Gp = Gn,p is such that Gp →2/3+η K 4 .
Corollary 3. For any constant 0 < η ≤ 1/3, there is a constant C = C(η) for
which the following holds. If 0 ≤ M = M (n) ≤ n2 is such that M ≥ Cn8/5 for all

large enough n, then almost every GM = Gn,M is such that GM →2/3+η K 4 .

In §6 below, we formulate an auxiliary conjecture (Conjecture 23) that, if proved,
would imply Conjecture 1 in full for all graphs H.
Finally, let us mention that Conjecture 1, if true, would immediately imply
the existence of very ‘sparse’ graphs G satisfying the property that G →γ H for
any γ > 1 − 1/(χ(H) − 1). A simple corollary of Theorem 2 is that, for any η > 0,
there is a graph G = Gη that contains no K 5 but we have G →2/3+η K 4 (see §5).
Erdős and Nešetřil have asked whether such graphs exist.
This note is organised as follows. In Section 1.1 we give a short outline of the
proof of our main result, Theorem 2, and in Section 1.2, some preliminary results
are given. In §2 the distribution of triangles in random and pseudo-random graphs
is studied. In §3 we prove a key lemma in the proof of our main result, Lemma 16.
Theorem 2 is proved in §4. In §5 we discuss a deterministic corollary to Theorem 2
concerning the Erdős–Nešetřil problem. Our last paragraph contains Conjecture 23.

§1. Outline of Proof and Preliminaries

1.1. Outline of the proof of Theorem 2. The proof of our main result is
somewhat long and hence, for convenience, in this section we describe its main
steps. Here we try to avoid being too technical.
The proof of Theorem 2 naturally splits into two parts. Suppose p = p(n) ≥
Cn−2/5 , where C is some large constant, and let H be a spanning subgraph of Gp =
Gn,p with ‘relative density’ e(H)/e(Gp ) ≥ λ. Let us say that two vertices x,
4
y ∈ Gp are K− -connected by H if there are two other vertices z1 , z2 ∈ H such
that both {x, z1 , z2 } and {z1 , z2 , y} induce triangles in H. Sometimes we also say
that such a pair xy is a pivotal pair .
In the first part of our proof, we show that the number of pairs
of vertices x,
y ∈ Gp that are K− 4
-connected by H is roughly at least (2λ − 1) n2 , as long as C is
a large enough constant. The precise statement of this result is given in Section 3.2,
Lemma 16. The second part of the proof consists of deducing our Theorem 2 from
Lemma 16. This part is less technical than the first, and is also considerably shorter.
The method used here was inspired by an argument in Rödl and Ruciński [14], and
4 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

a version of this technique was used in Haxell, Kohayakawa, and Luczak [8]. Let
us give a brief description of this method.
Thus let Gp = Gn,p be the binomial random graph with p = p(n) ≥ Cn−2/5 ,
where C is some large constant, and let a constant 0 < η ≤ 1/3 be fixed. For
simplicity, let us also assume that p = p(n) → 0 as n → ∞. We may write Gp
(j)
as the union of k independent random graphs Gp1 (1 ≤ j ≤ k), where k is some
large constant to be carefully chosen later. Since p = o(1), below we may ignore
(1) (k) (j)
the edges of Gp = Gp1 ∪ · · · ∪ Gp1 that belong to more than one of the Gp1 . Let us
now ask an ‘adversary’ to choose a subgraph H ⊂ Gp of Gp of size at least λe(Gp ),
where λ = 2/3 + η, or, equivalently, let us ask our adversary to choose a set of
edges F ⊂ E(Gp ) with |F | ≥ λe(Gp ). Our aim is to show that such a set F must
span a K 4 .
Instead of asking our adversary to pick F directly, we ask him to pick F ∩
(j) (j ) (j )
E(Gp1 ) for all j. For some j0 , we must have |F ∩ E(Gp10 ))| ≥ λe(Gp10 ). We
(j )
may in fact ask our adversary to pick first j0 and Fj0 = F ∩ E(Gp10 ), and leave
(j)
the choice of F ∩ E(G
p1 ) (j 6= j0n)
for later. By Lemma 16, we know that at
n n
least ∼ (2λ − 1) 2 = (1/3 + 2η) 2 edges of K join pairs of vertices that are
(j)
4
-connected by Fj0 . We now show G0 = j6=j0 Gp1 to our adversary, and ask
S
K−
him to pick F ∩ E(G0 ). Note that, with very high probability, at least 1/3 + η of the
edges of G0 will be formed by K− 4
-connected pairs, and if our adversary puts any
of these edges into F ∩ E(G ), then F will span a copy of K 4 . However, since G0
0

contains an extremely large proportion of the edges of Gp (we choose k very large),
our adversary is forced to pick at least 2/3 of the edges of G0 , and hence he is forced
to ‘close’ a K 4 by picking a K− 4
-connected pair xy for an edge of H.
Let us close this section with a few words on the proof of Lemma 16. Recall
4
that in that lemma we are concerned with estimating the number of K− -connected
pairs induced by subgraphs of random graphs. A very simple lower bound for the
number of such pairs induced by an arbitrary graph H∗ is given in assertion (*)
in Section 3.1. This estimate is far too weak to be of any use when dealing with
subgraphs of random graphs, but a weighted version of this estimate, Lemma 15, is
important in the proof of Lemma 16. Another important and a much deeper ingre-
dient in the proof of Lemma 16 is a version of Szemerédi’s regularity lemma [15]
for sparse graphs; see Lemma 4 in Section 1.2 below. A simple application of Lem-
mas 4 and 15 allows us to focus our attention on certain ε-regular quadruples. The
key lemma concerning such quadruples is Lemma 17 in Section 3.2. The proof of
Lemma 17 is based on certain results concerning the number and the distribution
of triangles in random and pseudo-random graphs. Paragraph 2 is entirely devoted
to those results. The main lemmas in §2 are Lemmas 7 and 10.

1.2. Preliminaries. Let a graph H = H n of order |H| = n be fixed. For U ,

W ⊂ V = V (H) with U ∩ W = ∅, we write E(U, W ) = EH (U, W ) for the set of
edges of H that have one endvertex in U and the other in W . We set e(U, W ) =
eH (U, W ) = |E(U, W )|.
The following notion will be needed in what follows. Suppose 0 < η ≤ 1 and 0 <
p ≤ 1. We say that H is η-upper-uniform with density p if, for all U , W ⊂
V with U ∩ W = ∅ and |U |, |W | ≥ ηn, we have eH (U, W ) ≤ (1 + η)p|U ||W |.
Clearly, if H is η-upper-uniform with density p, then it is also η 0 -upper-uniform
with density p0 for any η ≤ η 0 ≤ 1 and any p ≤ p0 ≤ 1. In the sequel, for any two
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 5

disjoint non-empty sets U , W ⊂ V , let

dH,p (U, W ) = eH (U, W )/p|U ||W |
be the p-relative density or, for short, the p-density of H between U and W . Now
suppose ε > 0, U , W ⊂ V , and U ∩ W = ∅. We say that the pair (U, W ) is
(ε, H, p)-regular if for all U 0 ⊂ U , W 0 ⊂ W with |U 0 | ≥ ε|U | and |W 0 | ≥ ε|W | we
have
|dH,p (U 0 , W 0 ) − dH,p (U, W )| ≤ ε.
We say that a partition P = (Vi )k0 of V = V (H) is (ε, k)-equitable if |V0 | ≤ εn,
and |V1 | = . . . = |Vk |. Also, we say that V0 is the exceptional class of P . When the
value of ε is not relevant, we refer to an (ε, k)-equitable partition as a k-equitable
partition. Similarly, P is an equitable partition of V if it is a k-equitable partition
k
for some k. Finally, we say that
an (ε, k)-equitable partition P = (Vi )0 of V is
k
(ε, H, p)-regular if at most ε 2 pairs (Vi , Vj ) with 1 ≤ i < j ≤ k are not (ε, p)-
regular. We may now state an extension of Szemerédi’s lemma [15] to subgraphs
of η-upper-uniform graphs.
Lemma 4. For any given ε > 0 and k0 ≥ 1, there are constants η = η(ε, k0 ) > 0
and K0 = K0 (ε, k0 ) ≥ k0 that depend only on ε and k0 such that any η-upper-
uniform graph H with density 0 < p ≤ 1 admits an (ε, H, p)-regular (ε, k)-equitable
partition of its vertex set with k0 ≤ k ≤ K0 . 
Using standard estimates for tails of the binomial distribution, it is easy to check
that a.e. Gn,p is η-upper-uniform with density p for any constant 0 < η ≤ 1 if d = pn
is larger than some constant d0 = d0 (η).
Let us introduce a piece of notation before we proceed. If U1 , . . . , U` ⊂ V (J)
are pairwise disjoint sets of vertices of a given graph J, we write J[U1 , . . . , U` ]
for the `-partite subgraph of J naturally defined by the Ui (1 ≤ i ≤ `). Thus,
S`
J[U1 , . . . , U` ] has vertex set 1 Ui and two of its vertices are adjacent if and only
if they are adjacent in J and, moreover, they belong to distinct Ui .
Now suppose we have real numbers 0 < p ≤ 1, 0 < ε ≤ 1, 0 < γ0 ≤ 1 and
an integer m ≥ 1. Suppose the above Ui (1 ≤ i ≤ `) all have cardinality m, and
write γij for the p-density dJ,p (Ui , Uj ) for all distinct i and j. Suppose L is a
graph on [`] = {1, . . . , `} such that, for any 1 ≤ i < j ≤ `, the pair (Ui , Uj ) is
(ε, J, p)-regular and γij ≥ γ0 whenever ij ∈ E(L).
We may now state our next lemma. In what follows, we write O1 (x) for any
term y satisfying |y| ≤ x. Also, as usual, we write ∆ = ∆(L) for the maximal
degree of L and we write ΓJ (x) for the J-neighbourhood of a vertex x ∈ V (J).
Lemma 5. Let J, L, and the sets Ui (1 ≤ i ≤ `) be as above and let ∆ = ∆(L).
Suppose 0 < ε ≤ 1/(2∆ + 1) and put ρ = (2∆ + 1/γ0 )ε and µ = 2∆ε. Then there
are sets Ūi ⊂ Ui with |Ūi | ≥ (1 − µ)m for all 1 ≤ i ≤ ` such that, for all x ∈ Ūi
and any 1 ≤ i ≤ `, we have
dij (x) = |ΓJ (x) ∩ Ūj | = (1 + O1 (ρ))γij pm (3)
for any j with ij ∈ E(L). 
Lemma 5 above is very similar to Lemma 2 in [7], and hence its rather elementary
proof is omitted. We close this section with a very simple large deviation inequality
for the hypergeometric distribution. This inequality will be used in Section 2.1
below.
6 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

Lemma 6. Let 1 ≤ a ≤ n and 1 ≤ t ≤ r ≤ n be integers, and suppose R ⊂ [n] is

an r-element subset of [n] = {1, . . . , n} chosen uniformly at random. Then
 t  

a r
P |R ∩ [a]| ≥ t ≤ . 
n−r+1 t

§2. Triangles in Pseudo-Random and Random Graphs

2.1. The counting lemma. Let m ≥ 1 be an integer. In this section we shall
consider a fixed triple V = (V̄1 , V̄2 , V̄3 ) of pairwise disjoint sets with m/2 ≤ mi =
|V̄i | ≤ m for all i ∈ {1, 2, 3}. We shall also suppose that 0 < p = p(m) ≤ 1
satisfies pm ≥ m1/2+1/ log log log m for all large enough m and, moreover, that p =
o(1) as m → ∞. In this section, all the asymptotic notation refers to m → ∞.
Our aim is to estimate from above the number of certain pseudo-random tripartite
graphs F with tripartition V (F ) = V̄1 ∪ V̄2 ∪ V̄3 that contain unexpectedly few
triangles given the number of edges that they have, or else whose triangles are not
too regularly distributed.
Before we may describe precisely which graphs F are of interest to us, we need
to introduce a few definitions. In what follows, indices will be tacitly taken mod-
ulo 3 when convenient. Let 0 < δ ≤ 1 be given. Suppose e ∈ EF (V̄i−1 , V̄i+1 ) =
E(F [V̄i−1 , V̄i+1 ]) (i ∈ {1, 2, 3}) and let k3 (e) = k3F (e) be the number of triangles
of F that contain e. We shall say that e is (δ, K 3 )-poor if

k3 (e) < (1 − δ)dF,p (V̄i , V̄i−1 )dF,p (V̄i , V̄i+1 )p2 mi .

The graph F is (δ, K 3 )-unbalanced if, for some i ∈ {1, 2, 3}, the number of (δ, K 3 )-
poor edges in EF (V̄i−1 , V̄i+1 ) is at least δeF (V̄i−1 , V̄i+1 ) = δ|EF (V̄i−1 , V̄i+1 )|. For
simplicity, below we write γi = dF,p (V̄i−1 , V̄i+1 ) (i ∈ {1, 2, 3}) and, if x ∈ V̄i and i 6=
j ∈ {1, 2, 3}, we let dij (x) = |ΓF (x) ∩ V̄j |.
Now suppose integers m and T and real constants 0 < ε̄ ≤ 1, 0 < γ̄0 ≤ 1,
and 0 < ρ̄ ≤ 1 are given, and let V = (V̄1 , V̄2 , V̄3 ) and m = (m1 , m2 , m3 ) be
as above. Let us write Fp (ε̄, γ̄0 , ρ̄, T ) = Fp (ε̄, γ̄0 , ρ̄; V, T ) for the set of tripartite
graphs F with tripartition V (F ) = V̄1 ∪ V̄2 ∪ V̄3 that satisfy the following properties:
(i ) (V̄1 , V̄2 , V̄3 ) is an (ε̄, F, p)-regular triple,
(ii ) γ̄0 ≤ γi = dF,p (V̄i−1 , V̄i+1 ) ≤ 2 for all i ∈ {1, 2, 3},
(iii ) dij (x) = |ΓF (x) ∩ V̄j | = (1 + O1 (ρ̄))γk pmj for all x ∈ V̄i and any choice of i,
j, and k such that {i, j, k} = [3],
(iv ) F has size e(F ) = |E(F )| = T .
Moreover, for any given 0 < δ ≤ 1, let Fpδ (ε̄, γ̄0 , ρ̄, T ) = Fpδ (ε̄, γ̄0 , ρ̄; V, T ) be the set
of (δ, K 3 )-unbalanced graphs F in Fp (ε̄, γ̄0 , ρ̄, T ). Put

fpδ (ε̄, γ̄0 , ρ̄; m, T ) = |Fpδ (ε̄, γ̄0 , ρ̄; V, T )|.

Sometimes the labelling of the vertices of the graphs in Fp (ε̄, γ̄0 , ρ̄; V, T ) or in
Fpδ (ε̄, γ̄0 , ρ̄; V, T ) is not relevant, and in that case we may replace V by m in our
notation.
Our crucial counting lemma is as follows.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 7

Lemma 7. Let 0 < α ≤ 1, 0 < γ̄0 ≤ 1, and 0 < δ ≤ 1 be given. Then there is
a constant ε0 = ε0 (α, γ̄0 , δ) > 0 that depends only on α, γ̄0 , and δ for which the
following holds. Suppose 0 < p = p(m) ≤ 1 is such that p ≥ m−1/2+1/ log log log m for
all large enough m and, moreover, p = o(1) as m → ∞. Then, if ρ̄ ≤ ρ̄0 = δ/27,
ε̄ ≤ ε0 , and m is sufficiently large, we have

3m2
 
fpδ (ε̄, γ̄0 , ρ̄; m, T ) = |Fpδ (ε̄, γ̄0 , ρ̄; m, T )| ≤ αT
T

for all T and all m = (m1 , m2 , m3 ) with m/2 ≤ mi ≤ m (i ∈ {1, 2, 3}).

Most of the rest of Section 2.1 is dedicated to the proof of Lemma 7 above.
Our general strategy in this proof is as follows. We randomly generate a tripartite
graph F with tripartition V (F ) = V̄1 ∪V̄2 ∪V̄3 and size e(F ) = |E(F )| = T , and show
that the probability that the graph we obtain will be a member of Fpδ (ε̄, γ̄0 , ρ̄, T )
is suitably small. We generate F in steps: we first generate F [V̄2 , V̄3 ]. We then
generate F [V̄1 , V̄3 ] and analyse the structure of the graph F [V̄1 , V̄3 ] ∪ F [V̄2 , V̄3 ]. We
then finally generate F [V̄1 , V̄2 ] and show that the appropriate probability is indeed
small.
Let us now make precise the process by which we generate F . We first of all
fix a partition T = T1 + T2 + T3 of T such that, putting γi = Ti /pmi−1 mi+1 ,
we have γ̄0 ≤ γi ≤ 2 for all i ∈ {1, 2, 3}. Note that, because of condition (ii )
above for F , we may disregard the T for which such a partition does not exist.
Let us suppose that the bipartite graph F23 = F [V̄2 , V̄3 ] has been fixed, and that
the following properties hold (cf. (i )–(iv ) above): (a) the pair (V̄2 , V̄3 ) is (ε̄, F23 , p)-
regular, (b) dij (x) = |ΓF23 (x)∩ V̄j | = (1+O1 (ρ̄))γ1 pmj for all x ∈ V̄i , where {i, j} =
{2, 3}, and (c) e(F23 ) = T1 .
We now fix the degree sequence for the vertices x ∈ V̄1 in the bipartite graph
F [V̄1 , V̄3 ], and generate this graph respecting this sequence. Thus let (d13 (x))x∈V̄1
P
with x∈V̄1 d13 (x) = T2 and d13 (x) = (1 + O1 (ρ̄))γ2 pm3 for all x ∈ V̄1 be fixed,
and generate the bipartite graph F13 = F [V̄1 , V̄3 ] by selecting the neighbourhoods
ΓF13 (x) ⊂ V̄3 (x ∈ V̄1 ) randomly and independently for all x ∈ V̄1 . Thus, for
every x ∈ V̄1 , all the d13 (x)-element subsets of V̄3 are equally likely to be chosen as
the neighbourhood of x within V̄3 . We now analyse the structure of F13 ∪ F23 =
F [V̄1 , V̄3 ] ∪ F [V̄2 , V̄3 ].
For convenience, let us put dij = Avex∈V̄i dij (x) = Tk /mi = γk pmj for all i, j,
and k with {i, j, k} = [3] and k 6= 3. Put also d13 = γ2 pm3 and d23 = γ1 pm3 .
Thus d13 (x) = (1 + O1 (ρ̄))d13 for all x ∈ V̄1 , and d23 (y) = (1 + O1 (ρ̄))d23 for
all y ∈ V̄2 .
Let 0 < β1 ≤ 1 be given. For all x ∈ V̄1 , put

Ve2 = Ve2 (x, β1 ) = {y ∈ V̄2 : σ(x, y) ≤ −β1 d13 d23 /m3 },

where σ(x, y) = d(x, y) − d13 d23 /m3 = |ΓF13 (x) ∩ ΓF23 (y)| − d13 d23 /m3 . Note that
the set Ve2 (x, β1 ) is defined in such a way that the following fact holds: if e = xy
(x ∈ V̄1 , y ∈ V̄2 ) is an edge of F , then e is a (β1 , K 3 )-poor edge if and only
if y ∈ Ve2 (x, β1 ).
Now let 0 < β2 ≤ 1 be given. Below we say that x ∈ V̄1 is (β1 , β2 )-faulty
if |Ve2 (x, β1 )| ≥ β2 m2 . Note that, clearly, since we are conditioning on F23 =
8 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

F [V̄2 , V̄3 ], the event that a vertex x ∈ V̄1 should be (β1 , β2 )-faulty depends only on
the random set ΓF13 (x) ⊂ V̄3 that is chosen as the neighbourhood of x within V̄3 .
Our next lemma is the key technical result in the proof of the main lemma in
this section, Lemma 7.
Lemma 8. Suppose the constants 0 < β1 ≤ 1, 0 < β2 ≤ 1, 0 < γ̄0 ≤ 1, 0 < ε̄ ≤ 1,
and 0 < ρ̄ ≤ 1 are such β1 β2 ≥ 27ε̄, β2 ρ̄ ≤ ε̄, and β2 ≤ γ̄0 . Then, for all
sufficiently large m, the probability that a given vertex x ∈ V̄1 is (β1 , β2 )-faulty is

d13
at most 5ε̄ε̄/β2 .

Proof. Let us fix x ∈ V̄1 , and assume that |Ve2 | = |Ve2 (x, β1 )| ≥ β2 m2 . Let Ve20 ⊂ Ve2
be such that m e 2 = |Ve20 | = dβ2 m2 e. The following assertion, whose proof we omit,
is very similar to Lemma 5.
Assertion 1. There exist sets V̄20 ⊂ Ve20 and V̄30 ⊂ V̄3 for which we have m02 = |V̄20 | ≥
e 2 and m03 = |V̄30 | ≥ (1 − 2ε̄)m3 , and furthermore
(1 − 2ε̄/β2 )m
  
3ε̄
d023 (y) = |ΓF23 (y) ∩ V̄30 | = 1 + O1 d23 (4)
β2

for all y ∈ V̄20 , and

  
3ε̄
d032 (z) = |ΓF23 (z) ∩ V̄20 | = 1 + O1 β2 d32 (5)
β2

for all z ∈ V̄30 .

Using Assertion 1, we prove next that there exists a small set Y of vertices
from Ve2 whose F23 -neighbourhood uniformly cover essentially all of V̄3 . We need
to introduce some notation. Let us set ω = ω(m) = m1/ log log m , and let dY (z) =
|ΓF23 (z) ∩ Y | for all Y ⊂ V̄2 and all z ∈ V̄3 . Put also
    
4ε̄
V̄300 = V̄300 (Y ) = z ∈ V̄3 : dY (z) = 1 + O1 ω
β2

for all Y ⊂ V̄2 .

Assertion 2. There is a set Y ⊂ Ve2 of cardinality q = |Y | = (1+O1 (3ε̄/β2 ))ωm2 /d32
such that |V̄300 | = |V̄300 (Y )| ≥ (1 − 2ε̄)m3 , and such that
 
3ε̄
d0023 (y) = |ΓF23 (y) ∩ V̄300 | ≥ 1− d23 (6)
β2

for all y ∈ Y .
Let V̄20 and V̄30 be as in Assertion 1. To prove Assertion 2, we construct Y by
randomly selecting its elements from the set V̄20 ⊂ Ve2 . Let pY = pY (m) = ω/β2 d32 .
Note that 0 < pY < 1 for all large enough m. Put the vertices of V̄20 into Y
randomly, each with probability pY , and with all these events independent. The
expected cardinality of Y is then
  
2ε̄ ωm2
E(|Y |) = 1 + O1 ,
β2 d32
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 9

and the expected degree of a vertex z ∈ V̄30 into Y is

  
Y 3ε̄
E(d (z)) = 1 + O1 ω.
β2

From standard bounds for the tail of the binomial distribution, we may deduce that
there is a set Y ⊂ V̄20 ⊂ Ve2 such that q = |Y | is as required in Assertion 2, and such
that every z ∈ V̄30 satisfies dY (z) = (1 + O1 (4ε̄/β2 ))ω. Note that, for such a set Y ,
we have V̄30 ⊂ V̄300 = V̄300 (Y ), and hence |V̄300 | ≥ |V̄30 | ≥ (1 − 2ε̄)m3 . It now suffices to
notice that every y ∈ Y is such that
 
3ε̄
d0023 (y) = |ΓF23 (y) ∩ V̄300 | ≥ |ΓF23 (y) ∩ V̄30 | ≥ 1 − d23 ,
β2

since y ∈ Y ⊂ V̄20 and relation (4) in Assertion 1 holds. This completes the proof
of Assertion 2.
Assertion 3. The probability that our fixed vertex x ∈ V̄1 admits a set Y as in

d13
Assertion 2 is at most 4ε̄ε̄/β2 .
Let us first sketch the idea in the proof of Assertion 3. Roughly speaking, the set Y
is such that the neighbourhoods ΓF23 (y) ∩ V̄300 of the vertices y ∈ Y within V̄300 are
about of the same size, and the vertices in V̄300 are, by definition, covered by those
sets quite uniformly. Now, since the vertices in Y are all in Ve2 , we have that the
neighbourhood ΓF13 (x) of x within V̄3 intersects all the neighbourhoods ΓF23 (y)
(y ∈ Y ) too little. Thus it intersects the sets ΓF23 (y) ∩ V̄300 too little as well, and
this is possible only if it in fact intersects V̄300 in an unexpectedly small set. It then
suffices to estimate the probability that ΓF13 (x) ∩ V̄300 should be as small. Let us
now formalise the argument above.
Assume that a set Y as in Assertion 2 exists. We consider the vectors g = gx =
(gz )z∈V̄300 and f y = (fzy )z∈V̄300 (y ∈ Y ) with entries

if z ∈ ΓF13 (x)

1
gz =
−d13 /m3 otherwise,

and
if z ∈ ΓF23 (y)

1
fzy =
−d23 /m3 otherwise.
fzy for z ∈ V̄300 . For any fixed z ∈ V̄300 , we have
P
We first estimate ξz = y∈Y
 
X d23  d23 d23 q
ξz = fzy Y
= d (z) − Y
q − d (z) = 1 + dY (z) −
m3 m3 m3
y∈Y
   
d23 4ε̄ d23 q
= 1+ 1 + O1 ω− .
m3 β2 m3

We have d23 = γ1 pm3 = o(m3 ), and, since d23 /m3 = T1 /m2 m3 = d32 /m2 , we
have d23 q/m3 = (1 + O1 (3ε̄/β2 ))ω. Therefore
   
4ε̄ 3ε̄ 8ε̄
ξz ≤ (1 + o(1)) 1 + ω− 1− ω≤ ω
β2 β2 β2
10 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

for large enough m, and similarly ξz ≥ −(8ε̄/β2 )ω if m is sufficiently

P large. Given
two vectors a = (az )z∈V̄300 and b = (bz )z∈V̄300 , let ha, bi = z∈V̄ 00 az bz . We now
3
estimate y∈Y hf y , gi. We have
P
* +
X X D E X
hf y , gi = f y , g = (ξz )z∈V̄300 , g = ξz gz .
y∈Y y∈Y z∈V̄300

Therefore
X X 8ε̄ X
hf y , gi ≤ |ξz gz | ≤ ω |gz |.
β2
y∈Y z∈V̄300 00
z∈V̄3

We have from our assumptions on our constants that ρ̄ ≤ ε̄/β2 ≤ β1 /27 ≤ 1/27.
Hence X X 9
|gz | ≤ |gz | ≤ (1 + ρ̄)d13 + d13 ≤ d13 ,
00
4
z∈V̄3 z∈V̄3

with plenty to spare. Thus

X 18ε̄
hf y , gi ≤ ωd13 . (7)
β2
y∈Y

We now give a lower estimate for the left-hand side of (7) in terms of the σ(x, y)
(y ∈ Y ). Let d0013 (x) = |ΓF13 (x) ∩ V̄300 | and recall that we let d0023 (y) = |ΓF23 (y) ∩ V̄300 |
for all y ∈ Y . Put m003 = |V̄300 | and d00 (x, y) = |ΓF13 (x) ∩ ΓF23 (y) ∩ V̄300 | for all y ∈ Y .
Fix y ∈ Y . Clearly d00 (x, y) ≤ d(x, y) = |ΓF13 (x) ∩ ΓF23 (y)|. Thus
d13  00 d23  00
hf y , gi = d00 (x, y) − d23 (y) − d00 (x, y) − d (x) − d00 (x, y)
m3 m3 13
 
d13 d23 00
 00 00 00
+ m 3 − d 13 (x) + d 23 (y) − d (x, y) .
m23
Write d0013 (x) = (1 − τ )d13 . Recalling (6), we see that
d13 d23 d13 d23 m003
hf y , gi ≤ d(x, y) − 2 +
m3 m23
3ε̄d13 d23 d13 d23 d13 d00 (x, y) d23 d00 (x, y)
+ +τ + +
β2 m3 m3 m3 m3
00 00 00
d13 d23 d13 (x) d13 d23 d23 (y) d13 d23 d (x, y)
− − +
m23 m23 m23
 
3ε̄d13 d23 d13 d23 d13 + d23 00 d13 d23
≤ σ(x, y) + +τ + d (x, y) + o ,
β2 m3 m3 m3 m3
where we used that, trivially, m003 ≤ m3 , and that d13 (x), d23 (y), and d00 (x, y) are
all o(m3 ), since by assumption p = p(m) → 0 as m → ∞. However,
d13 + d23 X 00 d13 + d23 X
d (x, y) = |ΓF13 (x) ∩ ΓF23 (y) ∩ V̄300 |
m3 m3
y∈Y y∈Y
 
d13 + d23 X Y 00
≤ d (z): z ∈ ΓF13 (x) ∩ V̄3
m3
 
d13 + d23 4ε̄
≤ (1 + ρ̄) 1 + ωd13 ,
m3 β2
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 11

which is o(d13 d23 q/m3 ). Therefore

 
X
y 4ε̄ d13 d23 q
hf , gi ≤ − β1 − −τ . (8)
β2 m3
y∈Y

Using that q = (1 + O1 (3ε̄/β2 ))ωm2 /d32 and that d23 /m3 = d32 /m2 , we deduce
from inequality (8) above that
   
X 4ε̄ 3ε̄
hf y , gi ≤ − β1 − −τ 1 + O1 ωd13
β2 β2
y∈Y
  
4ε̄ 3ε̄
≤ − β1 − − τ + O1 ωd13 . (9)
β2 β2

We now claim that τ ≥ 2ε̄/β2 . Indeed, otherwise we would have that β1 − 4ε̄/β2 −
τ + O1 (3ε̄/β2 ) ≥ 0, and hence, comparing inequalities (7) and (9), we would have
that β1 − 4ε̄/β2 − τ + O1 (3ε̄/β2 ) ≤ 18ε̄/β2 , and consequently that τ ≥ 2ε̄/β2 , which
is a contradiction.
Recall that d0013 (x) = |ΓF13 (x) ∩ V̄300 | = (1 − τ )d13 and that d13 (x) = |ΓF13 (x)| =
(1 + O1 (ρ̄))d13 . Also, |V̄3 \ V̄300 | ≤ 2ε̄|V̄3 |. Thus we deduce from the existence of the
set Y that at least (τ − ρ̄)d13 ≥ τ d13 /2 elements of ΓF13 (x) ⊂ V̄3 are confined to a
subset of V̄3 of cardinality at most 2ε̄|V̄3 |. Thus, by Lemma 6, the probability that

d13
such a set Y exists is at most 2(1+ρ̄)d13 (3ε̄)τ d13 /2 ≤ 23d13 /2 (3ε̄)ε̄d13 /β2 ≤ 4ε̄ε̄/β2 ,
completing the proof of Assertion 3.
We may now P finish the proof of Lemma 8 combining Assertions 2 and 3 above.
0
Indeed, writing q for the sum over all q satisfying q = (1 + O1 (3ε̄/β2 ))ωm2 /d32 ,
we have from the above two assertions that probability that the vertex x should be
(β1 , β2 )-faulty is, for large enough m, at most

X0 m2  6ε̄ωm2  m2

ε̄/β2 d13

d13
4ε̄ε̄/β2


4ε̄ ≤
q q β2 d32 2ωm2 /d32
 2ωm2 /d32
6ε̄ωm2 ed32
d13
d13
≤ 4ε̄ε̄/β2 ≤ 5ε̄ε̄/β2 ,
β2 d32 2ω

where in the last inequality we used that ωm2 (log m)/d32 = o(d13 ), which follows
easily from our assumption that pm ≥ m1/2+1/ log log log m for all sufficiently large m.
Thus Lemma 8 is proved. 
We are now in position to finish the proof of Lemma 7.
Proof of Lemma 7. Let us first state a simple fact concerning the Ti .
Assertion 1. For all i ∈ {1, 2, 3} we have Ti ≥ (γ̄0 /24)T .
Indeed, T ≤ i∈{1,2,3} γi pmi−1 mi+1 ≤ 6pm2 , and therefore, for any i ∈ {1, 2, 3},
P

we have
γ̄0 γ̄0
Ti = γi pmi−1 mi+1 ≥ pm2 ≥ T,
4 24
as required.
Our next observation is an immediate consequence of Lemma 8 and Assertion 1.
12 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

Assertion 2. Let 0 < β3 ≤ 1 be given. The probability that at least β3 m1 vertices

β3 γ̄0 T /24
in V̄1 are (β1 , β2 )-faulty is at most 6ε̄ε̄/β2 .
Indeed, by Lemma 8 and Assertion 1 above, the probability in question is at most

β3 m1
β3 T1
β3 γ̄0 T /24
2m1 (5ε̄ε̄/β2 )d13 = 2m1 5ε̄ε̄/β2 ≤ 6ε̄ε̄/β2 ,

completing the proof of Assertion 2.

We now describe the last step in the generation of F . Recall that we have
generated F13 ∪ F23 = F [V̄1 , V̄3 ] ∪ F [V̄2 , V̄3 ] so far. Let K[V̄1 , V̄2 ] be the complete
bipartite graph with bipartition V̄1 ∪ V̄2 . To generate F12 = F [V̄1 , V̄2 ], we randomly
pick for E(F [V̄1 , V̄2 ]) a T3 -element subset of E(K[V̄1 , V̄2 ]), uniformly chosen from
all such sets.
Assertion 3. Suppose that fewer than β3 m1 vertices in V̄1 are (β1 , β2 )-faulty. Then
the probability that at least δT3 edges in F12 = F [V̄1 , V̄2 ] are (β1 , K 3 )-poor is no
 T
larger than 4(β2 + β3 )δγ̄0 /24 .
Recall that an edge xy ∈ E(F [V̄1 , V̄2 ]) (x ∈ V̄1 , y ∈ V̄2 ) is (β1 , K 3 )-poor if and only
if y ∈ Ve2 (x, β1 ). The probability in question P0 is the probability that at least δT3
edges xy ∈ E(K[V̄1 , V̄2 ]) (x ∈ V̄1 , y ∈ V̄2 ) with y ∈ Ve2 (x, β1 ) are selected to be
elements of F12 . The number of such ‘potentially poor’ edges xy in K[V̄1 , V̄2 ] is at
most β3 m1 m2 + (1 − β3 )β2 m1 m2 ≤ (β2 + β3 )m1 m2 , and hence, by Lemma 6 and
Assertion 1, we have

δT3 γ̄0 T /24 T

P0 ≤ 2T3 2(β2 + β3 ) ≤ 4(β2 + β3 )δ ≤ 4(β2 + β3 )δγ̄0 /24
  
,

proving Assertion 3.
We may now finish the proof of Lemma 7. Let the constants α, γ̄0 , and δ as
in the statement of our lemma be given. We then let ε0 = ε0 (α, γ̄0 , δ) be such
that 0 < ε0 ≤ δγ̄0 /27 and, moreover,

γ̄0 /24 log log(1/ε̄) 1

6ε̄δ/27 ≤ α
6

and
 δγ̄0 /24
27ε̄ 1 1
4 + ≤ α
δ log log(1/ε̄) 6
for all 0 < ε̄ ≤ ε0 .
We now apply Assertions 2 and 3 above with suitably chosen β1 , β2 , β3 , ε̄, and ρ̄.
Let β1 = δ and fix any 0 < ε̄ ≤ ε0 . Let β2 = 27ε̄/δ ≤ γ̄0 and β3 = 1/ log log(1/ε̄).
Recall that ρ̄0 = δ/27. To complete the proof, we proceed as follows: we suppose
that ρ̄ ≤ ρ̄0 is given and that m is sufficiently large for the inequalities below to
hold. Fix the partition T = T1 + T2 + T3 of T , the bipartite graph F23 = F [V̄2 , V̄3 ],
and the degree sequence (d13 (x))x∈V̄1 as above. Then generate F13 = F [V̄1 , V̄3 ]. The
probability that at least β3 m1 vertices in V̄1 are (β1 , β2 )-faulty is, by Assertion 2,
at most

β3 γ̄0 T /24
β3 γ̄0 T /24  α T 1
6ε̄ε̄/β2 = 6ε̄δ/27 ≤ ≤ αT .
6 6
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 13

Let us now assume that fewer than β3 m1 vertices in V̄1 are (β1 , β2 )-faulty, and let
us generate F12 = F [V̄1 , V̄2 ]. Then the probability that δT3 edges in F12 = F [V̄1 , V̄2 ]
are (δ, K 3 )-poor is, by Assertion 3, at most
(  δγ̄0 /24 )T  
δγ̄0 /24 T 27ε̄ 1 α T 1
≤ αT .

4(β2 + β3 ) ≤ 4 + ≤
δ log log(1/ε̄) 6 6

We now note that the argument above is symmetric with respect to the indices i ∈
{1, 2, 3} (note the factor ‘3’ below), and thus we may conclude that

2 T 3m2
   2
T 3m
|Fpδ (ε̄, γ̄0 , ρ̄; V, T )| ≤3× α ≤α ,
6 T T

completing the proof of Lemma 7. 

Recall that m and T are integers, 0 < ε̄ ≤ 1, 0 < γ̄0 ≤ 1, 0 < δ ≤ 1, and 0 < ρ̄ ≤ 1
are fixed reals, 0 < p = p(m) ≤ 1 is such that pm → ∞ and p = o(1) as m → ∞,
and V = (V̄1 , V̄2 , V̄3 ) and m = (m1 , m2 , m3 ), where the V̄i (i ∈ {1, 2, 3}) are
pairwise disjoint sets with cardinality |V̄i | = mi (i ∈ {1, 2, 3}). Our next lemma
concerns a property of graphs F ∈ Fp (ε̄, γ̄0 , ρ̄, T ) \ Fpδ (ε̄, γ̄0 , ρ̄, T ), namely, graphs F
in Fp (ε̄, γ̄0 , ρ̄, T ) that are (δ, K 3 )-balanced . For a vertex x ∈ V (F ), let k3 (x) =
k3F (x) denote the number of triangles of F that contain x.
Lemma 9. Suppose F is a (δ, K 3 )-balanced graph in Fp (ε̄, γ̄0 , ρ̄; V, T ). Put δ 0 =
(2δ)1/2 . Then, for any i ∈ {1, 2, 3}, there are at most δ 0 mi vertices x in V̄i such
that
k3 (x) = k3F (x) < (1 − ρ̄ − δ 0 )γ1 γ2 γ3 p3 mi−1 mi+1 .

Proof. By symmetry, it suffices to prove the statement for i = 1. In the sequel

we freely use the notation introduced before Lemma 7. Also, let us put dij =
Avex∈V̄i dij (x) = γk pmj for all i, j, and k with {i, j, k} = [3]. Below we say
that a vertex x ∈ V̄1 is bad if at least δd12 /δ 0 = δ 0 d12 /2 edges in E(F [V̄1 , V̄2 ])
incident to x are (δ, K 3 )-poor edges. The number of such bad vertices x ∈ V̄1 is
at most δ 0 m1 , as otherwise the number of (δ, K 3 )-poor edges in E(F [V̄1 , V̄2 ]) would
be more than δm1 d12 = δe(F [V̄1 , V̄2 ]), contradicting the fact that F is (δ, K 3 )-
balanced.
Note that an edge e = x1 x2 ∈ E(F [V̄1 , V̄2 ]) (xi ∈ V̄i , i ∈ {1, 2}) that is
not (δ, K 3 )-poor ‘contributes’ with at least k3 (e) ≥ (1 − δ)γ1 γ2 p2 m3 triangles
to k3 (x1 ). Supposing that x1 ∈ V̄1 is not a bad vertex, summing over all x2 ∈ V̄2
for which x1 x2 ∈ E(F [V̄1 , V̄2 ]) is not a (δ, K 3 )-poor edge, we obtain that

δ0
 
k3 (x1 ) ≥ (1 − ρ̄)d12 − d12 (1 − δ)γ1 γ2 p2 m3
2
δ0
 
≥ 1 − ρ̄ − δ − γ1 γ2 γ3 p3 m2 m3 ≥ (1 − ρ̄ − δ 0 )γ1 γ2 γ3 p3 m2 m3 .
2

Since, as we saw above, at most δ 0 m1 vertices x1 ∈ V̄1 are bad, the proof is com-
plete. 
14 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

2.2. Distribution of triangles in random graphs. In this section we look at

the random graph Gp = Gn,p and study the distribution of the triangles it contains.
The aim will be to prove that the triangles of Gn,p do not unduly concentrate on
any fixed set of edges and vertices. To be precise, let x be any given vertex of Gp
and let E be a set of edges of Gp taken from the subgraph Gp [ΓGp (x)] induced by
the neighbourhood ΓGp (x) of x in Gp . Also, let W be a subset of vertices of Gp
disjoint from {x}∪V (E), where we write V (E) for the set V (Gp [E]) of vertices of Gp
that are incident to at least one edge from E. Our aim is to find an upper bound
G
for the number of triangles k3 (E, W ) = k3 p (E, W ) of Gp that are determined by
an edge from E and a vertex in W . Note that the expected value of this number
is p2 |E||W |. We shall show that this is an upper bound in probability up to 1 + θ
for any fixed θ > 0 as long as E and W are reasonably large and p = p(n) does not
tend to 0 too fast.
Lemma 10. Let c1 , c2 > 0 and 0 < θ ≤ 1 be given. Then there is a con-
stant C0 = C0 (c1 , c2 , θ) that depends only on c1 , c2 , and θ for which the following
holds. Suppose p = p(n) = ωn−2/5 where C0 ≤ ω = ω(n) = o(n1/190 ). Then
almost every Gp = Gn,p is such that, if E ⊂ E(Gp [ΓGp (x)]) and W ⊂ W f =
V (Gp ) \ ({x} ∪ V (E)) for some x ∈ V (Gp ), then

G
k3 (E, W ) = k3 p (E, W ) ≤ (1 + θ)p2 |E||W | (10)

as long as |E| ≥ c1 p3 n2 and |W | ≥ c2 n.

Most of the remainder of this section is devoted to the proof of Lemma 10.
Unfortunately, our proof below is a little indirect and is based on a few auxiliary
lemmas; moreover, this proof makes use of a technical condition that ω should
not be too large. The obvious direct approaches based on simple large deviation
inequalities seem to fail to give Lemma 10. To see why this might be expected,
note that (i ) for the sets E and W of interest, the expected value E(k3 (E, W )) =
p2 |E||W | of k3 (E, W ) is of order O(n) only, while the number of sets W that we
have to handle is exp{Ω(n)}, and (ii ) k3 (E, W ) is a sum of positively correlated
indicator variables and the most common large deviation inequalities for such sums
do not seem to be strong enough for our purposes.
Let us turn to the proof of Lemma 10. For the rest of this section we assume
that p = p(n) = ωn−2/5 , where C0 ≤ ω = ω(n) = o(n1/190 ) and C0 is some large
constant. (The main result for Gp with larger p will be deduced from the small p
case; cf. Lemma 19.)
Let P 3 be the path of length 2 and E k the k-vertex graph with no edges (k ≥ 1).
We write Hk for P 3 ∨ E k (k ≥ 1), the graph on k + 3 vertices we obtain from the
disjoint union of P 3 and E k by adding all the 3k edges between V (P 3 ) and V (E k ).
A little piece of arithmetic shows that almost no Gn,p contains a copy of H12 . Thus
for the rest of this section we may and shall assume that our Gn,p is H12 -free.
We may clearly assume that the degree of any vertex of Gp = Gn,p is (1 +
o(1))pn, and also that any vertex of Gp is contained in at most p3 n2 triangles.
Furthermore, the expected number of common neighbours of any two fixed vertices
of Gp is p2 (n−2) ≤ ω 2 n1/5 . Thus, we may and shall condition on our Gp being such
that dGp (x, y) = |ΓGp (x) ∩ ΓGp (y)| ≤ 2ω 2 n1/5 for any pair of distinct vertices x,
y ∈ V (Gp ). Finally, we may assume that Gp is o(1)-upper-uniform.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 15

Suppose x ∈ V (Gp ) and E ⊂ E(Gp [ΓGp (x)]). For a vertex y ∈ W f = V (Gp ) \

({x} ∪ V (E)), let Ey be the set E ∩ E(Gp [ΓGp (y)]) of edges of E that the neigh-
bourhood of y in Gp induces in Gp . Clearly, k3 (E, y) = k3 (E, {y}) = |Ey |. We shall
say that a vertex y ∈ W
f is (x, E)-bad , or simply E-bad , if Ey is not an independent
set of edges.
Lemma 11. Almost every Gp is such that, for any x ∈ V (Gp ) and any E ⊂
E(Gp [ΓGp (x)]), at most 10eω 8 n4/5 vertices are E-bad.
Proof. Fix a vertex x ∈ V (Gp ). Let us generate Gp as follows: we first choose the
neighbourhood ΓGp (x) of x in Gp , and once this set is fixed, we decide which edges
within ΓGp (x) should be in Gp . Put E (0) = E(Gp [ΓGp (x)]), and let V (E (0) ) ⊂
ΓGp (x) be the set of vertices in ΓGp (x) that are incident to at least one edge
in E(Gp [ΓGp (x)]). For the rest of the proof, we assume that the edges generated
so far are fixed.
Let us now consider a vertex y ∈ Y = V (Gp ) \ ({x} ∪ ΓGp (x)), and let us
decide which edges between y and V (E (0) ) should be in our Gp . Notice that
whether or not y is E (0) -bad depends solely on these y–V (E (0) ) edges. In par-
ticular, the events ‘y is E (0) -bad’ (y ∈ Y ) are all independent. Let us estimate
the probability that a given vertex y ∈ Y turns out to be E (0) -bad. We first
observe that with probability 1 − o(1/n) we have |V (E (0) )| ≤ |ΓGp (x)| ≤ 2pn
and |∆(E (0) )| = |∆(Gp [E (0) ])| ≤ 2p2 n = 2ω 2 n1/5 . In the sequel, we assume that
these two inequalities hold. In particular, the number of copies of P 3 spanned
by E (0) is, crudely, at most |V (E (0) )|∆(E (0) )2 ≤ 8ω 5 n. Thus the probability that y
is E (0) -bad is

P(∆(Ey(0) ) ≥ 2) = P(Gp [Ey(0) ] contains a P 3 )

≤ E(#{P 3 ⊂ Gp [Ey(0) ]}) ≤ 8ω 5 np3 = 8ω 8 n−1/5 ,

(0) (0)
where #{P 3 ⊂ Gp [Ey ]} denotes the number of copies of P 3 in Gp [Ey ]. Now,
from the independence of the events ‘y is E (0) -bad’ (y ∈ Y ), we have that the
probability Pk that at least k such vertices y are E (0) -bad satisfies
k
8eω 8 n4/5
  
n  8 −1/5 k
Pk ≤ 8ω n ≤ . (11)
k k

Thus, if k = b9eω 8 n4/5 c, we have Pk = o(1/n) with plenty to spare.

The above argument proves our lemma with ‘E ⊂ E(Gp [ΓGp (x)])’ replaced
by ‘E = E(Gp [ΓGp (x)])’. To complete the proof, we make the following simple
observation. Suppose x ∈ V (Gp ) is fixed, E ⊂ E (0) = E(Gp [ΓGp (x)]), and y ∈
Y = V (Gp ) \ ({x} ∪ ΓGp (x)). Then, y is necessarily E (0) -bad whenever it is E-bad.
Thus, an E-bad vertex y ∈ W f = V (Gp ) \ (V (E) ∪ {x}) is either contained in ΓG (x)
p

or else it is E (0) -bad. Since we may assume that ∆(Gp ) ≤ 2pn = 2ωn3/5 ≤ ω 8 n4/5 ,
our lemma follows. 
Lemma 11 above tells us that, for any x and any E, we have ∆(Ey ) = ∆(Gp [Ey ]) ≤
1 for most y ∈ W
f = V (Gp ) \ ({x} ∪ V (E)). Of course, since Gp is supposed not to
contain H12 , we have ∆(Ey ) ≤ 11 for any vertex y ∈ Wf.
16 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

For each vertex y ∈ W f , let Xy = |Ey | = k3 (E, y) and let Xy0 be the cardinal-
ity ν(Ey ) = ν(Gp [Ey ]) of a maximum matching in Gp [Ey ]. Since ∆(Ey ) ≤ 11, we
have Xy0 ≥ Xy /2∆(Ey ) ≥ Xy /22 for any y ∈ W f . Moreover, for any y ∈ W f that is
0
not E-bad, we clearly have Xy = Xy .
Let us now fix W ⊂ W f . Put
X X
X = XW = Xw = k3 (E, w) = k3 (E, W ),
w∈W w∈W

and similarly X 0 = X 0 0
P P
PW = w∈W Xw . Let us write b for sum over all w ∈ W
that are E-bad and g for sum over all w ∈ W that are not E-bad. Then
X X
k3 (E, W ) = XW = Xw + Xw
g b
X X X
= Xw0 + 0
Xw ≤ XW + Xw . (12)
g b b

Now our aim is to bound the last two summands in (12).

For any two distinct vertices x and y ∈ V (Gp ), let Yxy be the number of edges in-
duced by the set ΓGp (x, y) = ΓGp (x)∩ΓGp (y) in Gp . Thus Yxy = |E(Gp [ΓGp (x, y)])|,
and E(Yxy ) = n−2

5 5
2 p = (1/2 + o(1))ω .

Lemma 12. For almost every Gp we have Yxy ≤ 22e2 max{log n, ω 5 } for any pair
of distinct vertices x, y ∈ V (Gp ).
0
Proof. Let Yxy be the maximum cardinality ν(Gp [ΓGp (x, y)]) of a matching in
Gp [ΓGp (x, y)]. We claim that
 0 2
P Yxy ≥ e2 max{log n, ω 5 } ≤ n−e = o(n−2 ). (13)

For convenience, put µ = E(Yxy ) ≤ ω 5 . By Lemma 2 in Janson [10], for any a ≥ e2 ,

we have
 
0 1
P(Yxy ≥ aµ) ≤ exp{−µ(a log a + 1 − a)} ≤ exp − a(log a)µ . (14)
2

We now check that (13) follows from (14). Suppose µ = E(Yxy ) ≤ log n. Then we
take a = e2 µ−1 log n ≥ e2 in (14) and note that then
2
0 0
P(Yxy ≥ e2 log n) = P(Yxy ≥ aµ) ≤ exp{−e2 log n} = n−e .

Suppose now that µ = E(Yxy ) > log n. Then we take a = e2 in (14) to obtain
2
0 0
P(Yxy ≥ e2 ω 5 ) ≤ P(Yxy ≥ e2 µ) ≤ exp{−e2 log n} = n−e .
0
Thus the claimed inequality (13) does hold. In particular, almost surely Yxy ≤
2 5
e max{log n, ω } for any distinct x, y ∈ V (Gp ). Our lemma now follows, since
0
∆(Gp [ΓGp (x, y)]) ≤ 11, and hence Yxy ≥ Yxy /22. 
By Lemmas 11 and 12 above, an almost sure upper bound for the last summand
0
in (12) is 220e3 max{log n, ω 5 }ω 8 n4/5 . Our aim now is to estimate XW . Fortu-
nately, the method of proof of Lemma 2 in Janson [10] gives the following result
immediately.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 17

Lemma 13. Suppose 0 ≤ ε ≤ 1/2. Then for any W ⊂ W

f = V (Gp ) \ ({x} ∪ V (E))
we have  
0 2 1 2 2
P(XW ≥ (1 + ε)p |E||W |) ≤ exp − ε p |E||W | .
3

Proof. We follow an argument of Janson [10] (cf. the proof of Lemma 2 in [10]).
Fix W ⊂ W f = V (Gp ) \ ({x} ∪ V (E)) and let w ∈ W . Let F be the family of all
copies J of P 3 with middle vertex w and endvertices coinciding with endvertices of
edges in E in the complete graph on V (Gp ). Thus each such J corresponds to a
triangle determined by w and one edge from E. Let Fm (m ≥ 0) be the collection
of all m-tuples (J1 , . . . , Jm ) of pairwise edge-disjoint elements from F. Let IJ be
the indicator variable ofPthe event that a fixed J ∈ F should be present in Gp .
Thus k3 (E, w) = Xw = J∈F IJ . We have
( ) ( )
0 X X 0  X 1 X m
tXw w
(et − 1)m ≤ E IJ1 . . . IJm et − 1


E(e )=E
m m!
m≥0 m≥0 Fm
( )m
X 1 X m X 1 X t
p2m et − 1 ≤ p2 e t − 1 = eλ(e −1) ,



=
m! m!
m≥0 Fm m≥0 F

where λ = E(Xw ). Now, the Xw0 (w ∈ W ) are independent and therefore

0 P 0 Y 0 t
E(etXW ) = E(et w∈W Xw
)= E(etXw ) ≤ eλ|W |(e −1)
.
w∈W

Let a ≥ 1. Note that then, by Markov’s inequality, we have

0
0
P(XW ≥ aλ|W |) = P(etXW ≥ etaλ|W | )
0 t t
≤ E(etXW )e−taλ|W | ≤ eλ|W |(e −1) −taλ|W |
e = eλ|W |(e −1−at)
.

Taking t = log a, we have

0
P(XW ≥ aλ|W |) ≤ eλ|W |(a−1−a log a) = e−λ|W |(a log a−a+1) .

Our lemma follows by setting a = 1 + ε in this last inequality. 

We are finally ready to prove Lemma 10.
Proof of Lemma 10. Take C0 = C0 (c1 , c2 , θ) = {12/(θ2 c1 c2 )}1/5 . We proceed to
show that this choice for C0 will do. Thus let x, E, and W as in the statement of
our lemma be given. To prove (10), it suffices to put together (12) and Lemmas 11,
12, and 13. Indeed, with the notation as above, by Lemma 13 we have
    (  2 )
0 θ 1 θ
P XW ≥ 1+ p2 |E||W | ≤ exp − p2 |E||W | ≤ e−n ,
2 3 2

where in the last inequality we used that |E| ≥ c1 p3 n2 , |W | ≥ c2 n, and C0 =

{12/(θ2 c1 c2 )}1/5 .
18 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

The number of choices for the triple (x, E, W ) is at most n exp{2ω 3 n4/5 log n}2n .
Indeed, it suffices to notice that we may assume that |E| ≤ u0 = bp3 n2 c = bω 3 n4/5 c,
and hence the number of choices for E is at most
X n 2   2
n
≤2 ≤ exp{2ω 3 n4/5 log n}.
u u0
u≤u0

0
Thus almost every Gp is such that XW ≤ (1 + θ/2)p2 |E||W |.
From (12) and Lemmas 11 and 12, we have, for almost every Gp ,
 
θ
k3 (E, W ) ≤ 1 + p2 |E||W | + 220e3 max{log n, ω 5 }ω 8 n4/5 ,
2

which is at most (1 + θ)p2 |E||W |, as required. 

G
In our next lemma we give a set of conditions that ensures that k3 p (E, W )
(E ⊂ E(Gp ), W ⊂ V (Gp ) \ V (E)) concentrates around its mean.
Lemma 14. Suppose ω = ω(n) → ∞ as n → ∞. Let 0 < p = p(n) ≤ 1 with p ≥
ωn−1/2 log n be given. Then almost every Gp = Gn,p is such that, for any E ⊂
E(Gp ) and W ⊂ V (Gp ) \ V (E) with |E| ≥ ωnp−1 log n and |W | ≥ n/ log n, we
have
G
k3 (E, W ) = k3 p (E, W ) = (1 + o(1))p2 |E||W |. (15)

Proof. Let θ > 0 be fixed. Let M ⊂ E(K n ) be a matching in the complete

graph K n , and set ν = |M |. Suppose that νp2 / log n → ∞ as n → ∞. Let W ⊂
V (Gp ) \ V (M ) be such that w = |W | ≥ n/ log n. Let us write k30 (M, W ) =
G ∪M
k3 p (M, W ) for the number of triangles in Gp ∪ M that are determined by an
edge of M and a vertex of W . Note that k30 (M, W ) has binomial distribution with
parameters νw = |M ||W | and p2 . Thus

k30 (M, W ) = (1 + O1 (θ))p2 |M ||W | (16)

with probability 1 − exp{−Ω(p2 νw)}. The number of matchings M ⊂ E(K n ) of

n

cardinality ν is no
larger than (ν2 ) ≤ n2ν , and the number of sets W as above is
n
no larger than w ≤ nw . Hence, we see that (16) holds a.s. for any matching M ⊂
E(K ) with |M | ≥ ν0 = ω(log n)/4p2 and any W ⊂ V (Gp ) \ V (M ) with |W | ≥
n

n/ log n, since
X X X X  2
w
n2ν+w exp{−p2 νw} ≤ n2ν ne−Ω(p ν)

ν≥ν0 w≥n log n ν≥ν0 w≥n log n

2ν−Ω(p2 νn/ log n)
X
≤ n
ν≥ν0
= o(1).

Thus, let us assume that our Gp does have this property. Clearly, we may also as-
sume that ∆(Gp ) ≤ 2pn. We claim that under these assumptions our Gp necessarily
satisfies (15) with ‘o(1)’ replaced by ‘O1 (θ)’ for all E and W as in the statement of
our lemma. To see this, fix E ⊂ E(Gp ) and W ⊂ V (Gp ) \ V (E) as in the lemma.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 19

We shall now make use of the following simple fact that may be easily deduced from
Vizing’s theorem: if J is a graph of maximum degree ∆(J), then J admits a proper
edge-colouring with at most ∆(J) + 1 colours such that the cardinality of any two
colour classes differ by at most 1. Note that ∆(E) = ∆(Gp [E]) ≤ ∆(Gp ) ≤ 2pn,
and hence, by the observation above, we may write E = E1 ∪ · · · ∪ Eq where
the Ei are matchings satisfying |Ei | − |Ej | ≤ 1 for all i and j, and more-
over q ≤ ∆(E) + 1 ≤ 3pn. In particular, |Ei | ≥ |E|/4pn
P ≥ ν0 for all i. There-
fore (16) applies with M = Ei and, since k3 (E, W ) = 1≤i≤q k3 (Ei , W ), our claim
does hold. Lemma 14 follows by letting θ tend to 0. 

§3. Pivotal Pairs of Vertices

3.1. A weighted Turán type result. Let H∗ be a graph and r ≥ 3 an integer.
for the graph with r vertices and 2r − 1 edges, and let us say
r


Let us write K−
r
that its two vertices of degree r − 2 are the endvertices of K− . Let us say that the
r
unordered pair xy = {x, y} of distinct vertices of H∗ is a K− -connected pair if there
r r
is a copy of K− in H∗ with endvertices x and y. Hence, if xy is a K− -connected
pair of non-adjacent vertices, then the addition of xy to H∗ creates a new copy
of K r in H∗ . Thus, we shall also say that a K− r
-connected pair is K r -pivotal , or
simply pivotal . For technical reasons, let us also say that the vertex x ∈ V (H∗ ) is
r
by itself a K− -connected pair if x lies in a copy of K r−1 in H∗ . The following is an
asymptotic version of Turán’s theorem for K r .
(*) Any graph H∗ with k vertices and e(H∗ ) edges contains at least
 
k+1
(r − 2)e(H∗ ) − (r − 3)
2
r
K− -connected pairs of vertices. 
To check that (*) is indeed an asymptotic form of Turan’s theorem, observe that

−1
if λ = λ(H∗ ) = e(H∗ ) |H2∗ | is the ‘density’ of H∗ = H∗k , then the lower bound
r
in (*) for the number of K− -connected pairs in H∗ is, for large k,
 

k
∼ (r − 2)λ − r + 3 . (17)
2

Note that (17) is bigger than (1/(r − 1)) k2 for λ > 1 − 1/(r − 1). Therefore we

may deduce that any (large) H∗ with λ(H∗ ) > 1 − 1/(r − 1) necessarily contains
a K r , which is, of course, a weak form of Turán’s theorem. Unfortunately, (*) does
not seem to imply Turán’s theorem for K r in its precise form.
In the sequel we shall need, however, a weighted version of (*) for r = 4. To
describe this version we need some technical definitions. Also, to simplify the
notation we restrict ourselves to the case r = 4. (The general case does not present
any further difficulty.) We start with a graph H∗ = H∗k of order k and assume γ =
(γe )e∈E(H∗ ) is an assignment of weights γe ≥ 0 to the edges e ∈ E(H∗ ) of H∗ .
Suppose x = (x1 , . . . , xk ) is an ordering of the vertices of H∗ . For any two not
4
necessarily distinct vertices x, y ∈ V (H∗ ) that form a K− -connected pair in H∗ , we
let
wH∗ ,γ (x, y) = max{γ1 γ2 : ∃z1 , z2 ∈ V (H∗ ) with γi = γzi y (i ∈ {1, 2})
and xz1 , xz2 , z1 z2 , z1 y, z2 y ∈ E(H∗ )}.
20 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

4
For convenience, if x, y do not form a K− -connected pair, we put wH∗ ,γ (x, y) = 0.
Let X
w(H∗ , γ, x) = wH∗ ,γ (xi , xj )
1≤i≤j≤k
P
and γ(H∗ ) = e∈E(H∗ ) γe . Our weighted version of (*) for r = 4 is given in
Lemma 15 below. We remark that to deduce the unweighted case (*) for r = 4
from this lemma, it suffices to take γ̄ = 1 and γe = 1 for all edges e ∈ E(H∗ ).
Lemma 15. Let H∗ = H∗k be a graph of order k and edge weights γ = (γe )e∈E(H∗ )
with 0 ≤ γe ≤ γ̄ for all e ∈ E(H∗ ), where γ̄ ≥ 1. Then there is an ordering x =
(x1 , . . . , xk ) of the vertices of H∗ for which we have
 
k+1
w(H∗ , γ, x) ≥ 2γ(H∗ ) − γ̄ . (18)
2

Proof. Our proof is by induction on k. Since the case k ≤ 3 is trivial, we assume

that k ≥ 4 and that our lemma holds for graphs H∗ with at most 3 vertices. Note
that if γ(H∗ ) ≤ γ̄k 2 /4, then (18) is trivially true, as in this case 2γ(H∗ ) − γ̄ k+1


2 ≤
2 2
γ̄(k /2 − k(k + 1)/2) = −γ̄k/2 ≤ 0. Thus we may assume that γ(H∗ ) > γ̄k /4.
Since γe ≤ γ̄ for all e ∈ E(H∗ ), we have that H∗ has more than k 2 /4 edges.
Therefore, there are three vertices y1 , y2 , and y3 P ∈ V (H∗ ) inducing a triangle in H∗ .
For y ∈ V (H∗ ), let us write dγ (y) = dH∗ ,γ (y) = z∈ΓH (y) γyz for the γ-degree of y.
∗
We may assume that dγ (y1 ) ≤ dγ (y2 ) ≤ dγ (y3 ). Put x1 = y1 , and by induction
let x0 = (x2 , . . . , xk ) be an ordering of the vertices of H∗0 = H∗ − x1 = H∗ − y1 such
that  
0 0 0 k
w(H∗ , γ, x ) ≥ 2γ(H∗ ) − γ̄ .
2
0
For simplicity, γ above stands for
P the restriction (γe )e∈E(H∗0 ) of γ = (γe )E(H∗ ) to H∗ .
Our aim now is to estimate 1≤i≤k wH∗ ,γ (x1 , xi ) from below. For j ∈ {2, 3},
set γi (j) = γyj xi if yj xi ∈ E(H∗ ) and set γi (j) = 0 otherwise. Now observe that,
since γ̄ ≥ 1, for any reals 0 ≤ α ≤ γ̄ and 0 ≤ β ≤ γ̄ we have αβ ≥ α + β − γ̄.
Therefore
X X X 
wH∗ ,γ (x1 , xi ) ≥ γi (2)γi (3) ≥ γi (2) + γi (3) − γ̄
1≤i≤k 1≤i≤k 1≤i≤k

= dγ (y2 ) + dγ (y3 ) − γ̄k ≥ 2dγ (y1 ) − γ̄k = 2dγ (x1 ) − γ̄k.

Hence, with x = (x1 , x2 , . . . , xk ), we have

X
w(H∗ , γ, x) ≥ wH∗ ,γ (x1 , xi ) + wH∗0 ,γ (H∗0 , γ, x0 )
1≤i≤k
   
k k+1
γ
≥ 2d (x1 ) − γ̄k + 2γ(H∗0 ) − γ̄ = 2γ(H∗ ) − γ̄ ,
2 2

as required. 
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 21

3.2. Pivotal pairs in subgraphs of random graphs. In this section we turn

4
to the study of K− -connected pairs in subgraphs of random graphs. Recall that
given a graph H and two distinct vertices x, y ∈ V (H), we say that the unordered
4 4
pair xy = {x, y} is a K− -connected pair if they are the endvertices of a copy of K−
4
in H, and that a single vertex by itself forms a K− -connected pair if it belongs to
a triangle of H. Our main aim here is to prove Lemma 16 below, which roughly
says that, if a subgraph H ⊂ Gp = Gn,p of Gp is such that e(H) ≥ λe(Gp ) for some
4
fixed λ > 0 and p = p(n) is not too small, then the number of K− -connected pairs
in H is, almost surely,  
n
& (2λ − 1) .
2
This result is similar in spirit to assertion (*) in Section 3.1, but note that (*)
applied to H ⊂ Gp above gives nothing if p = p(n) → 0 as n → ∞. Lemma 16
below remedies this situation and recovers essentially the same lower bound for the
4
number of K− -connected pairs.
In the sequel, for any given graph H, it will be convenient to define a graph Π =
ΠH on V (H) by letting two distinct vertices x, y ∈ V (H) = V (Π) be adjacent in Π
4
if and only if they form a K− -connected in H.
Lemma 16. Let a constant 0 < σ ≤ 1 be given. Then there is a constant C0 =
C0 (σ) that depends only on σ for which the following holds. If p = p(n) = ωn−2/5
and C0 ≤ ω = ω(n) = o(n1/190 ), then almost every Gp = Gn,p is such that, for any
subgraph H ⊂ Gp of Gp , we have
 
n
e(ΠH ) ≥ (2λ − 1 − σ) , (19)
2

−1
where λ = e(H) p n2


.
Before we proceed, let us remark that we shall apply Lemma 16 above with σ
much smaller than λ. In fact, we shall be interested in the case in which λ is a little
greater than 2/3 and σ is very small.
The proof of Lemma 16 is based on the results of the previous three sections and
on a further technical lemma, Lemma 17, which we now describe. The context in
which Lemma 17 applies is as follows. Suppose 0 < µ ≤ 1/2, 0 < κ ≤ 1, 0 < δ 0 ≤ 1,
0 < γ̄0 ≤ 1, and 0 < ρ̄ ≤ 1/4 are constants, and let n and m ≥ κn be integers. Let
also 0 < p = p(n) ≤ 1 be given. Let V̄1 , V̄2 , Z̄1 , and Z̄2 be pairwise disjoint sets of
cardinality (1−µ)m ≤ mj = |V̄j | ≤ m and (1−µ)m ≤ m0i = |Z̄i | ≤ m (i, j ∈ {1, 2}).
Suppose F = F1 ∪F2 is a graph with Fj a tripartite graph with tripartition V (Fj ) =
V̄j ∪ Z̄1 ∪ Z̄2 (j ∈ {1, 2}) and such that E(F [Z̄1 , Z̄2 ]) = E(Fj [Z̄1 , Z̄2 ]) for both j ∈
{1, 2}. For simplicity, put γZ = dF,p (Z̄1 , Z̄2 ) and let γij = dF,p (Z̄i , V̄j ) for i,
j ∈ {1, 2}. Suppose further that the following conditions hold:
(i ) We have γZ ≥ γ̄0 and γij ≥ γ̄0 for all i, j ∈ {1, 2}.
(ii ) At least (1 − δ 0 )m1 vertices x in V̄1 are such that

k3F (x) ≥ (1 − ρ̄ − δ 0 )γ11 γ21 γZ p3 m01 m02 ,

where k3F (x) is the number of triangles of F that contain x.

22 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

(iii ) At least (1 − δ 0 )e(F [Z̄1 , Z̄2 ]) edges e in F [Z̄1 , Z̄2 ] are such that

k3F (e, V̄2 ) ≥ (1 − δ 0 )γ12 γ22 p2 m2 , (20)

where k3F (e, V̄2 ) denotes the number of triangles in F determined by the
edge e and some vertex in V̄2 .
(iv ) Let c = γ̄03 κ2 /8. For any x ∈ V̄1 , E ⊂ E(F [ΓF (x)]), and W ⊂ V̄2 with |E| ≥
cp3 n2 and |W | ≥ cn, we have
δ0
 
F
k3 (E, W ) ≤ 1 + p2 |E||W |.
4

(v ) For all E ⊂ E(F [Z̄1 , Z̄2 ]) with |E| ≥ pn2 / log n and W ⊂ V̄1 with |W | ≥
n/ log n, we have k3F (E, W ) ≤ 2p2 |E||W |.
We may now state a key technical lemma in the proof of Lemma 16.
Lemma 17. Let constants 0 < γ̄0 ≤ 1 and 0 < σ ≤ 1 be given. Then, with
the notation above, if δ 0 ≤ δ00 = δ00 (γ̄0 , σ) = σ 2 γ̄0 /128 and µ ≤ µ0 = µ0 (σ) =
4
σ/8, then the number of K− -connected pairs x1 x2 with xj ∈ V̄j (j ∈ {1, 2}) is at
2
least (1 − σ)γ12 γ22 m .
Proof. Let us put α = µ0 = σ/8 and note that then δ00 = σ 2 γ̄0 /128 = α2 γ̄02 /2.
Let 0 < δ 0 ≤ δ0 and 0 < µ ≤ µ0 be given, and suppose that F is as described before
our lemma.
In the sequel, an edge e ∈ E(F [Z̄1 , Z̄2 ]) will be said to be (δ 0 , K 3 ; F2 )-poor
if (20) fails. For x ∈ V̄1 , let us write k3p (x) for the number of (δ 0 , K 3 ; F2 )-poor
edges e ∈ E(F [Z̄1 , Z̄2 ]) induced by the neighbourhood of x in F . Let us say
that x ∈ V̄1 is unusable if k3p (x) ≥ αγ11 γ21 γZ p3 m01 m02 . The proof is now split into
a few claims.
Assertion 1. At most αm1 vertices in V̄1 are unusable.
Suppose the contrary. Let us consider the number N of pairs (x, e) with x a vertex
in V̄1 and e a (δ 0 , K 3 ; F2 )-poor edge in E(F [Z̄1 , Z̄2 ]) such that there is a triangle of F
containing both x and e. Since we are assuming that more than αm1 vertices x ∈ V̄1
are unusable, we have

N > α2 γ11 γ21 γZ p3 m01 m02 m1 ≥ α2 γ̄02 γZ p3 m01 m02 m1 . (21)

We now use condition (v ) above to deduce that

N ≤ 2p2 m1 δ 0 e(F [Z̄1 , Z̄2 ]) = 2δ 0 γZ p3 m01 m02 m1 (22)

Comparing (21) and (22), we obtain that α2 γ̄02 < 2δ 0 , contradicting the fact
that δ 0 ≤ δ00 = α2 γ̄0 /2. Thus Assertion 1 holds.
We now observe that condition (ii ) above immediately implies the following.
Assertion 2. For at least (1 − 2α)m1 vertices x in V̄1 , we have k3F (x) − k3p (x) ≥
c1 p3 n2 , where c1 = (1 − ρ̄ − 2α)(1 − µ)2 κ2 γ̄03 ≥ c.
Now let x ∈ V̄1 be given. Let Ex ⊂ E(F [Z̄1 , Z̄2 ]) be the set of edges e induced by
the neighbourhood of x in F that are not (δ 0 , K 3 ; F2 )-poor. Thus |Ex | = k3F (x) −
k3p (x). Let
Wx = {y ∈ V̄2 : E(F [ΓF (x)]) ∩ E(F [ΓF (y)]) 6= ∅}
4
= {y ∈ V̄2 : xy is a K− -connected pair}.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 23

Assertion 3. For at least (1 − 2α)m1 vertices x ∈ V̄1 , we have |Wx | ≥ (1 −

σ/2)γ12 γ22 m2 .
Let x ∈ V̄1 be such that |Ex | = k3F (x) − k3p (x) ≥ c1 p3 n2 , and suppose that |Wx | <
(1 − σ/2)γ12 γ22 m2 . We now use (iv ) above to deduce that

δ0
 
0 2 F
(1 − δ )γ12 γ22 p m2 |Ex | ≤ k3 (Ex , Wx ) ≤ 1 + p2 |Ex ||Wx |
4
δ0 
 
σ
≤ 1+ 1− γ12 γ22 p2 m2 |Ex |,
4 2

which is a contradiction since 1 − δ 0 > (1 + δ 0 /4)(1 − σ/2). Thus any x ∈ V̄1

with |Ex | = k3F (x) − k3p (x) ≥ c1 p3 n2 is such that |Wx | ≥ (1 − σ/2)γ12 γ22 m2 , and
hence Assertion 3 follows from Assertion 2.
From Assertion 3, we deduce that at least
 σ
(1 − 2α) 1 − γ12 γ22 m1 m2
2  σ
≥ (1 − 2α)(1 − µ)2 1 − γ12 γ22 m2 ≥ (1 − σ)γ12 γ22 m2
2
4
pairs x1 x2 (xj ∈ V̄j , j ∈ {1, 2}) are K− -connected with respect to F , thereby
proving Lemma 17. 
We are now ready to prove Lemma 16. Our proof will make use of several of our
previous lemmas.
Proof of Lemma 16. Let 0 < σ ≤ 1 be given. We now define the many constants
with which we shall apply Lemmas 4, 5, 7, 9, 10, 14, 15, and 17.
Let γ0 = σ/100, γ̄0 = γ0 /2, α = γ0 /24e and k0 = d100/σe. Let δ = (δ 0 )2 /2,
where δ 0 = σ 3 /2 × 106 ≤ δ00 = δ00 (γ̄0 , σ/7) = σ 2 γ̄0 /6272. Note that δ00 (γ̄0 , σ/7) is as
given in Lemma 17. We now let
 
1
ε = min 10−16 σ 6 γ0 , ε0 (α, γ̄0 , δ) ,
2

where ε0 = ε0 (α, γ̄0 , δ) is as defined in Lemma 7. Put ε̄ = 2ε, ρ = 2ε/γ0 and ρ̄ =

5ε/γ0 . For later reference, note that

5ε δ
ρ̄ = ≤ 5 × 10−16 σ 6 ≤ , (23)
γ0 27

and if k ≥ k0 , then, with plenty to spare, we have

2σ 6 σ
+ ≤ , (24)
7(k − 1) k − 1 7

and
3ε 1 2
+ ≤ σ. (25)
γ0 k 7
Set µ = 6ε ≤ µ0 (σ/7), where µ0 (σ/7) = σ/56 is as given in Lemma 17. Now
let K0 = K0 (ε, k0 ) and η = η(ε, k0 ) be as given by Lemma 4. We may assume
24 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

that η ≤ min{σ/7, ε/2K0 }. Put κ = 1/2K0 , and let c = γ̄03 κ2 /8. Finally, let C0 =
C0 (c, c, δ 0 /4), where C0 (c, c, δ 0 /4) is as given by Lemma 10. We claim that this
choice of C0 = C0 (σ) will do in Lemma 16, and proceed to prove this assertion.
Let p = p(n) = ωn−2/5 , where C0 ≤ ω = ω(n) = o(n1/190 ). Let us consider the
following conditions for Gp = Gn,p .
(a) Gp is η-upper-uniform and has size e(Gp ) = (1 + o(1))p n2 .

(b) Suppose m ≥ κn. Then, for any T ≥ (3/4)γ̄0 pm2 and any m = (m1 , m2 , m3 )
with m/2 ≤ mi ≤ m (i ∈ {1, 2, 3}), our Gp contains no copy of any
graph F ∈ Fpδ (ε̄, γ̄0 , ρ̄; m, T ) as a subgraph.
(c) Inequality (10) in Lemma 10 holds for all E and W as in the statement of
that lemma with c1 = c2 = c and θ = δ 0 /4.
(d ) Relation (15) in Lemma 14 holds for all E ⊂ E(Gp ) and W ⊂ V (Gp ) \ V (E)
with |E| ≥ pn2 / log n and |W | ≥ n/ log n.
Claim. Conditions (a)–(d ) hold for almost every Gp .
Proof of the Claim. Condition (a) clearly holds almost surely. We use Lemma 7
to prove that (b) holds almost surely as well. Let m, m, and T be as in (b). Note
that the number of choices for m and m is trivially at most n4 . The number of
copies of a fixed graph F ∈ Fpδ (ε̄, γ̄0 , ρ̄; m, T ) that the complete graph K n on n
vertices contains is clearly at most nn . Thus, since ρ̄ ≤ δ/27 (see (23)) and ε̄ =
2ε ≤ ε0 (α, γ̄0 , δ), Lemma 7 applies to give that the expected number of copies of
elements from Fpδ (ε̄, γ̄0 , ρ̄; m, T ) that Gp contains is, for sufficiently large n, at most
T
3m2 T 3eαpm2
  
4+n
n Fpδ (ε̄, γ̄0 , ρ̄; m, T ) T
p ≤n 4+n T
α p ≤n 4+n
,
T T

which, since T ≥ (3/4)γ̄0 pm2 , is at most

 T
8eα
n 4+n
≤ n4+n 3−T ≤ 2−T .
γ0

Summing over all T ≥ (3/4)γ̄0 pm2 , we obtain that (b) holds almost always. Con-
dition (c) holds almost surely for Gp by Lemma 10 and the choice of C0 . To check
1/2
that (d ) holds for a.e.
 G−1p , by Lemma 14 it suffices to see that pn / log n → ∞
2
and that (pn / log n) np log n → ∞ as n → ∞. This completes the proof of our
claim.
In the remainder of the proof, we show that (19) holds for any subgraph H ⊂ Gp
whenever Gp satisfies conditions (a)–(d ) above. This clearly proves our lemma.
Thus let us assume that H ⊂ Gp is given and that Gp satisfies (a)–(d ). We may
clearly assume that H is a spanning subgraph of Gp .
Let us apply Lemma 4 to the η-upper-uniform graph H, with parameters ε
and k0 . Let V (H) = V0 ∪ · · · ∪ Vk be the (ε, H, p)-regularS(ε, k)-equitable partition
we obtain in this way. Let H 0 be the subgraph of H on 1≤i≤k Vi with e ∈ E(H)
an edge of H 0 if and only if e joins two vertices that belong to distinct classes of
our regular partition, say, Vi and Vj (1 ≤ i < j ≤ k), with (Vi , Vj ) an (ε, H, p)-
regular pair and dH,p (Vi , Vj ) ≥ γ0 . In the sequel, we let m stand for the common

−1
cardinality of the sets Vi (1 ≤ i ≤ k). Recall that λ = e(H) p n2 . We now
check the following simple fact.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 25

n
Assertion 1. We have e(H 0 ) ≥ (λ − σ/7)p


2 provided n is sufficiently large.
From the η-upper-uniformity of H it follows that if W ⊂ V = V (H), |W | ≥ 2ηn
then e(H[W ]) ≤ (1 + η)p |W |

2 2
2 . Hence e(H[V0 ]) ≤ (3/5)ε pn . Also, eH (V0 , V \ V0 )
2
is at most (1 + η)p|V0 ||V \ V0 |
≤ (6/5)εpn . Thus the number ofPedges of H
incident to V0 is at most 4εp n2 for large enough n. Now note that eH (Vi , Vj )
with the sum over all 1 ≤ i < j ≤ k such that (Vi , Vj ) is not (ε, H, p)-regular
is at most ε k2 (1 + η)pm2 ≤ 2εp n2 . Also,


P
eH (Vi , Vj ) with the sum extended
over all
1 ≤ i < j ≤ k such that dH,p (Vi , Vj ) ≤ γ0 is at most (1 +
η)γ0 k2 pm2 ≤

2γ0 p n2 . Finally, we have that m

≤ (2p/k) n2 .
P

1≤i≤k e(H[Vi ])
≤ k(1 + η)p 2
Therefore |E(H) \ E(H 0 )| ≤ (6ε + 2/k + 2/γ0 )p n2 ≤ (σ/7)p n2 if n is sufficiently

large, as required.
We now define a graph H∗ on [k] = {1, . . . , k} by letting ij (1 ≤ i < j ≤ k) be an
edge of H∗ if and only if (Vi , Vj ) is an (ε, H, p)-regular pair and dH,p (Vi , Vj ) ≥ γ0 .
We write γij for dH,p (Vi , Vj ) for all ij ∈ E(H∗ ), and put γ = (γe )e∈E(H∗ ) . Now
put γ̄ = 1 + σ/7, and notice that then the definition of η and (a) above gives
that γe ≤ 1 + η ≤ γ̄ for all e ∈ E(H∗ ).
Lemma 15 now tells us that, suitably adjusting the notation, the ordering x =
(1, . . . , k) of the vertices of H∗ is such that

X k+1
w(H∗ , γ, x) = wH∗ ,γ (i, j) ≥ 2γ(H∗ ) − γ̄ . (26)
2
1≤i≤j≤k

In our next assertion we bound γ(H∗ ).

Assertion 2. We have γ(H∗ ) = e∈E(H∗ ) γe ≥ (λ − σ/7) k2 .
P

Indeed, we have e(H 0 ) = e∈E(H∗ ) γe pm2 = γ(H∗ )pm2 , and hence, by Assertion 1,
P

     2  
1  σ n σ n k  σ k
γ(H∗ ) ≥ λ − ≥ λ − ≥ λ − .
m2 7 2 7 2 n2 7 2

4
Our next step relates the number of K− -connected pairs meeting two fixed
classes Vi and Vj (1 ≤ i < j ≤ k) with the summand wH∗ ,γ (i, j) appearing in (26).
Assertion 3. Suppose ι1 , ι2 ∈ [k] are two distinct vertices of H∗ . Then the number
4
of K− -connected pairs x1 x2 with xj ∈ Vιj (j ∈ {1, 2}) is at least

(1 − σ/7 − 2ε/γ0 )wH∗ ,γ (ι1 , ι2 )m2 .

The above assertion is an easy consequence of Lemma 17, although we shall have
to work a little to check that that lemma does apply here.
4
Let us start by observing that, trivially, if ι1 and ι2 are not K− -connected in H∗ ,
then by definition wH∗ ,γ (ι1 , ι2 ) = 0, and hence there is nothing to prove. Thus let
us assume that this is not the case, and let ι3 , ι4 ∈ [k] be two vertices of H∗ such
that ι1 ι3 , ι1 ι4 , ι3 ι4 , ι2 ι3 , ι2 ι4 ∈ E(H∗ ). Choosing ι3 and ι4 suitably, we may further
assume that wH∗ ,γ (ι1 , ι2 ) = γι2 ι3 γι2 ι4 . We may now restrict our attention to the
4-partite subgraph of H induced by the Vιa (1 ≤ a ≤ 4).
Let J = H[Vι1 , Vι2 , Vι3 , Vι4 ] and write L for the graph on [4] = {1, 2, 3, 4} iso-
4
morphic to K− , with 1 and 2 as the endvertices. We first apply Lemma 5 to J to
26 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

obtain Ūιa ⊂ Vιa (1 ≤ a ≤ 4) such that the following holds. Putting ma = |Ūιa |,
we have (1 − µ)m ≤ ma ≤ m for all 1 ≤ a ≤ 4 and, furthermore, if ab ∈ E(L)
and x ∈ Ūιa , then
dab (x) = |ΓJ (x) ∩ Ūιb | = (1 + O1 (ρ))dJ,p (Vιa , Vιb )pm. (27)
Let F = J[Ūι1 , Ūι2 , Ūι3 , Ūι4 ]. Since dJ,p (Vιa , Vιb ) = dF,p (Ūιa , Ūιb ) + O1 (ε) and
dF,p (Ūιa , Ūιb ) ≥ γ0 − ε ≥ γ̄0 = γ0 /2, we have
dJ,p (Vιa , Vιb ) = (1 + O1 (2ε/γ0 ))dF,p (Ūιa , Ūιb ).
Moreover, as (1−µ)m ≤ mb ≤ m, we have m = (1+O1 (2µ))mb . Thus, relation (27)
gives that, for any x ∈ Ūιa ,
dab (x) = (1 + O1 (ρ̄))dF,p (Ūιa , Ūιb )pmb ,
where, as defined above, ρ̄ = 5ε/γ0 . Note also that γ̄0 ≤ dF,p (Ūιa , Ūιb ) ≤ γ̄ =
1 + σ/7 (cf. condition (a) above).
Our immediate aim now is to apply Lemma 17. To make our current notation the
same as the one used in that lemma, let us put V̄j = Ūιj and Z̄j = Ū2+j for j ∈ {1, 2}
and Fj = J[V̄j , Z̄1 , Z̄2 ] (j ∈ {1, 2}). Clearly, F = J[V̄1 , V̄2 , Z̄1 , Z̄2 ] = F1 ∪ F2
and F [Z̄1 , Z̄2 ] = F1 [Z̄1 , Z̄2 ] = F2 [Z̄1 , Z̄2 ]. Also, let mj = |V̄j | and m0i = |Z̄i | (i,
j ∈ {1, 2}).
Let mj = (mj , m01 , m02 ) and Tj = e(Fj ) ≥ (3/4)γ̄0 pm2 (j ∈ {1, 2}). Observe that
by (b) above we may conclude that

Fj = H[V̄j , Z̄1 , Z̄2 ] ∈ Fp (ε̄, γ̄0 , ρ̄; mj , Tj ) \ Fpδ (ε̄, γ̄0 , ρ̄; mj , Tj )
for both j ∈ {1, 2}. We shall now invoke Lemma 17, but to see that that lemma
does apply we verify the following claim.
Claim. Conditions (i )–(v ) given before the statement of Lemma 17 hold.
Proof of the Claim. Condition (i ) has already been seen to hold. To see that (ii )
holds, we simply apply Lemma 9 to the (δ, K 3 )-balanced graph F1 . Now, condi-
tion (iii ) clearly holds as F2 is (δ, K 3 )-balanced. Finally, condition (iv ) is equivalent
to (c), while condition (v ) follows from (d ). This finishes the proof of our claim.
In view of the above claim and the definitions of δ00 = δ00 (γ̄0 , σ/7) and µ0 =
µ0 (σ/7), we see that we may indeed apply Lemma 17 to deduce that the number
4
of K− -connected pairs x1 x2 with xj ∈ Vιj (j ∈ {1, 2}) is at least
 σ
1− dF,p (Z̄1 , V̄2 )dF,p (Z̄2 , V̄2 )m2 . (28)
7
We now use that, for i ∈ {1, 2}, we have
 
ε
dF,p (Z̄i , V̄2 ) ≥ dH,p (Vι2 , Vιi+2 ) − ε ≥ 1− dH,p (Vι2 , Vιi+2 ),
γ0
since dH,p (Vι2 , Vιi+2 ) ≥ γ0 . Thus the quantity in (28) is at least
 2
 σ ε
1− 1− dH,p (Vι2 , Vι3 )dH,p (Vι2 , Vι4 )m2
7 γ0
 
σ 2ε
≥ 1− − γι2 ι3 γι2 ι4 m2 ,
7 γ0
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 27

which concludes the proof Assertion 3, since ι3 and ι4 were chosen so as to have
wH∗ ,γ (ι1 , ι2 ) = γι2 ι3 γι2 ι4 .
The proof of our lemma is completed in the next assertion. Recall that ΠH
stands for the graph on V (H) with two vertices of H adjacent in ΠH if and only if
4
they are K− -connected in H.
Assertion 4. We have e(ΠH ) ≥ (2λ − 1 − σ) n2 .

By Assertion 3 we have that

X X  σ 2ε

e(ΠH ) ≥ e(ΠH [Vi , Vj ]) ≥ 1− − wH∗ ,γ (i, j)m2 . (29)
7 γ0
1≤i<j≤k 1≤i<j≤k

Now, clearly, we have

X X
wH∗ ,γ (i, j) = w(H∗ , γ, x) − wH∗ ,γ (i, i)
1≤i<j≤k 1≤i≤k

≥ w(H∗ , γ, x) − γ̄ 2 k ≥ w(H∗ , γ, x) − 2k,

since γ̄ 2 = (1 + σ/7)2 ≤ 2. Thus, recalling (26) and using that m ≥ (1 − ε)n/k, we

have from (29) that e(ΠH ) is at least
2
   
σ 2ε 2n
1− − (1 − ε) 2 w(H∗ , γ, x) − 2k
7 γ0 k
σ 3ε n2
     
k+1
≥ 1− − 2γ(H∗ ) − γ̄ − 2k ,
7 γ0 k 2 2
which, by Assertion 2, is at least
      −1  −1  
σ 3ε 1 σ  σ k + 1 k k n
1− − 1− 2 λ− − 1+ −2k ,
7 γ0 k 7 7 2 2 2 2

which is, as one may check using (24) and (25), at least
     
σ 3ε 1 4σ n n
1− − − 2λ − 1 − ≥ (2λ − 1 − σ) ,
7 γ0 k 7 2 2
proving Assertion 4.
The proof of Lemma 16 is complete. 

§4. Proof of the Main Result

We first prove Theorem 2 under the extra hypothesis that p = p(n) should
not be too large. More precisely, we prove the following result. Recall that we
write G →γ H if any subgraph J of G with e(J) ≥ γe(G) contains a copy of H.
Lemma 18. Let a constant η > 0 be given. Then there is a constant C =
C(η) that depends only on η for which the following holds. If 0 ≤ p = p(n) =
ωn−2/5 ≤ 1 and C ≤ ω = ω(n) = o(n1/190 ), then almost every Gp = Gn,p is such
that Gp →2/3+η K 4 .
Proof. Let ε = σ = η/10 and k = d1 + 24(log 2)η −2 e. Let C = kC0 (σ), where C0 (σ)
is as given in Lemma 16. We shall show that this choice of C = C(η) will do in our
28 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

result. Thus let p = p(n) be as in the statement of Theorem 2, and consider the
space G(n, p) of the random graphs Gn,p = Gp . In this proof, we shall write Gp as
a union of sparser, independent random graphs.
Let p1 be such that 1−p = (1−p1 )k , and note that then p/k ≤ p1 = (1+o(1))p/k.
Q (1) (k)
Put Ω = 1≤j≤k G(n, p1 ). We shall write G = (Gp1 , . . . , Gp1 ) for a general random
(j)
element of Ω. Thus the Gp1 (1 ≤ j ≤ k) are independent random graphs, each taken
(1) (k)
from G(n, p1 ). For any given G = (Gp1 , . . . , Gp1 ) ∈ Ω, let us put Gp = Gp (G) =
(1) (k) (j)
Gp1 ∪· · ·∪Gp1 , and note that then the map G = (Gp1 )1≤j≤k ∈ Ω 7→ Gp = Gp (G) ∈
G(n, p) is measure-preserving. We may thus study G(n, p) investigating the random
(j)
elements G ∈ Ω. Let us define Ω0 ⊂ Ω by letting G = (Gp1 )1≤j≤k ∈ Ω belong to Ω0
(j)
if and only if (i ) e(Gp ) = (1 + O1 (ε))p n2 , (ii ) e(Gp1 ) = (1 + O1 (ε))(p/k) n2 for



(j)
all 1 ≤ j ≤ k and, finally, (iii ) for all 1 ≤ j ≤ k, the graph Gp1 has the property
(j)
−1
that if E ⊂ E(Gp1 ) and λ = |E| p1 n2

, then
 
n
e(ΠE ) ≥ (2λ − 1 − σ) . (30)
2

For simplicity, ΠE above stands for ΠH , where H = H(E) is the graph on V (Gp )
with edge set E.
Elementary facts concerning random graphs and Lemma 16 gives that P(Ω0 ) =
1 − o(1). In the sequel, we shall often condition on Ω0 and we shall write P0 (A) for
the conditional probability P(A Ω0 ) for any event A ⊂ Ω.
(1) (k)
Let B ⊂ Ω be the set of G = (Gp1 , . . . , Gp1 ) that admit a set F ⊂ E(Gp )
with |F | ≥ (2/3 + η)e(Gp ) but Gp [F ] 6⊃ K 4 . We need to show that P(B) = o(1),
or, equivalently, that P0 (B) = o(1). Let us put B 0 = B ∩ Ω0 . For each G ∈ B, let
us fix once and for all a set F = F (G) ⊂ E(Gp ) as required in the definition of B.
(j)
Let us also put F (j) = F (j) (G) = F ∩ E(Gp1 ) (1 ≤ j ≤ k) and set f = f (G) = |F |
(j) (j) (j)
and f = f (G) = |F | (1 ≤ j ≤ k).

Now let γ (j) = γ (j) (G) = f (j) /p1 n2 , and note that then we have f ≤ f (1) +
· · · + f (k) , and hence that

n n


 
2 (1) p1 2 (k) p1 2 1 1 + ε  (1) 
+η ≤γ + ··· + γ ≤ γ + · · · + γ (k) ,
3 e(Gp ) e(Gp ) k 1−ε

from which we conclude that

 
1−ε 2 2
γ ∗ = γ ∗ (G) = max γ (j) ≥ Ave γ (j) ≥ +η ≥ (1 + η), (31)
1≤j≤k 1≤j≤k 1+ε 3 3

where the last inequality follows from the choice of ε. Let us also note that

γ ∗ ≤ 1 + ε, (32)

since we are assuming that (ii ) above holds. For each 1 ≤ j ≤ k, let

Bj0 = {G ∈ B 0 : γ (j) (G) = γ ∗ (G)}.

 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 29

Clearly B 0 = 1≤j≤k Bj0 , and hence it suffices to show that P0 (Bj0 ) = o(1) for all j.
S

Thus we now fix j ∈ [k] and proceed to show that Bj0 almost surely does not hold.
We have
X
P0 (Bj0 ) = P0 (Bj0 ∩ {G ∈ Ω : G(j)
p1 = G0 })
G0
X
= P0 (Bj0 G(j) 0 (j)
p1 = G0 )P (Gp1 = G0 )
G0

≤ max P0 (Bj0 G(j)

p1 = G0 ), (33)
G0

where G0 ranges over all graphs on V (Gp ) with e(G0 ) =
(1 + O1 (ε))(p/k) n2 and

such that (30) holds for all E ⊂ E(G0 ) with λ = |E|/p1 n2 . We now fix one such G0
and proceed to show an upper bound for (33). For each F0 ⊂ E(G0 ), let
P 0 (j, G0 , F0 ) = P0 (G ∈ Bj0 , F (j) = F0 G(j)
p1 = G0 ). (34)
Then
p n
P 0 (j, G0 , F0 ) ≤ 2(1+ε) k ( 2 )
X
P0 (Bj0 G(j)
p1 = G0 ) = max P 0 (j, G0 , F0 ).
F0 ⊂E(G0 )
F0 ⊂E(G0 )
(35)
We now fix F0 and estimate the last term in (35) from above. We may of course
assume that P 0 (j, G0 , F0 ) > 0 for the fixed triple (j, G0 , F0 ) under considera-
tion, as otherwise there is nothing to prove. Thus we may in particular assume
n
that |F0 |/p1 2 ≥ (2/3)(1 + η). Let us write B 0 (j, G0 , F0 ) for the set of G ∈ Bj0


(j)
such that Gp1 = G0 and F (j) = F0 . To show that P 0 (j, G0 , F0 ) is small, we argue
(i) (i)
that if G = (Gp1 )1≤i≤k ∈ B 0 (j, G0 , F0 ), then the edges of the graphs Gp1 (i 6= j,
1 ≤ i ≤ k) are distributed in a rather unlikely way.
Let p2 be such that 1 − p2 = (1 − p1 )k−1 , and note that then p2 /(k − 1) ≤
(i) (¬j)
p1 = (1 + o(1))p2 /(k − 1). For any given G = (Gp1 )1≤i≤k ∈ Ω, we write Gp2
S (i) (¬j)
for i Gp1 , where the union is taken over all i 6= j (1 ≤ i ≤ k). Thus Gp2 is a
(j) (¬j)
random element from G(n, p2 ), and Gp = Gp1 ∪ Gp2 . Note also that if G ∈ Ω0 ,
(¬j)
then e(Gp2 ) ≤ (1 + 2ε)p2 n2 .


(i) (¬j)
For any G = (Gp1 )1≤i≤k ∈ Ω, let F 0 = F 0 (G) = E(Gp2 ) ∩ E(ΠF0 ). Also,
let f 0 = f 0 (G) = |F 0 | for any G ∈ Ω. Clearly, f 0 = f 0 (G) (G ∈ Ω) is binomially
distributed with parameters e(ΠF0 ) and p2 . In fact, it is clear that f 0 = f 0 (G)
(j)
has this distribution even if we condition on G being such that Gp1 = G0 . We
now verify the following claim, from which we shall deduce an exponential upper
estimate for P 0 (j, G0 , F0 ). In the sequel, we write EG0 for the expectation in the
(j)
space Ω ∩ {G: Gp1 = G0 }.
Claim. If G ∈ B 0 (j, G0 , F0 ) then f 0 ≤ (1 − η)EG0 (f 0 ).
(j)
Proof of the Claim. Let us fix G = (Gp1 )1≤j≤k ∈ B 0 (j, G0 , F0 ). Let F (¬j) =
(¬j)
F (¬j) (G) = F ∩ E(Gp2 ), and put f (¬j) = f (¬j) (G) = |F (¬j) |. Clearly F (¬j) ∪
(¬j)
F 0 ⊂ E(Gp2 ) and, since F spans no K 4 , we have F (¬j) ∩ F 0 = ∅. Thus we
(¬j)
have f (¬j) + f 0 ≤ e(Gp2 ) ≤ (1 + 2ε)p2 n2 , and hence


 
0 (¬j) n
f ≤ (1 + 2ε − γ )p2 , (36)
2
30 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

where γ (¬j) = γ (¬j) (G) = f (¬j) /p2 n2 . We now show that γ (¬j) is suitably large.

We have f = |F | ≤ f (j) + f (¬j) , and hence

n n


 
2 (j) p1 2 (¬j) p2 2 1 1 + ε  (j) 
+η ≤γ +γ ≤ γ + (k − 1)γ (¬j) .
3 e(Gp ) e(Gp ) k 1−ε

Therefore

2 1 k − 1 (¬j) 1 k − 1 (¬j) 1
(1 + η) ≤ γ (j) + γ ≤ γ∗ + γ ≤ (1 + ε) + γ (¬j) ,
3 k k k k k

(¬j)
where the last inequality follows from (32). Thus we conclude that

γ ≥ 2/3.
0 ∗ n
We now note that µ = EG0 (f ) = p2 e(ΠF0 ) ≥ (2γ − 1 − σ)p
2 2 . Note that, in
n n


particular, by (31) and the choice of σ, we have (1/3)p2 2 ≤ µ ≤ p2 2 . Let b
denote the right-hand side of (36). Then
 

∗ n (¬j)

µ − b ≥ 2γ − 1 − σ − 1 − 2ε + γ p2
2
     
4η n n
≥ − 2ε − σ p2 ≥ ηp2 ≥ ηµ,
3 2 2

and therefore f 0 ≤ b ≤ (1 − η)µ, as claimed.

We now use our claim to bound P 0 (j, G0 , F0 ), which, we recall, is defined in (34)
above. Recall also that f 0 ∼ Bi(e(ΠF0 ), p2 ). We have
   
0 0 0 (j) 1 2
P (j, G0 , F0 ) ≤ P f ≤ (1 − η)EG0 (f ) Gp1 = G0 ≤ exp − η µ ,
2

n
where, as remarked before, µ = EG0 (f 0 ) = e(ΠF0 )p2 ≥ (1/3)p2


2 . Thus, from (35)
we deduce that
    
0 p n
(2) 1 2 1 n
P (Bj0 G(j)
p1
(1+ε) k
= G0 ) ≤ 2 exp − η 1 − p
6 k 2
   
k−1 2 p n
= exp (1 + ε)(log 2) − η
6 k 2
    
1 1 n
≤ exp − 1− η2 p
12 k 2
 
1
≤ exp − η 2 pn2 .
30

We now recall (33) to deduce that P0 (Bj0 ) ≤ exp{η 2 pn2 /30} = o(1), completing the
proof of Lemma 18. 

−1
We next show that, loosely speaking, the quantity ex(Gn,p , H) p n2 is non-
increasing in probability for any fixed graph H. In particular, this shows that
Lemma 18 implies Theorem 2.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 31

Lemma 19. Suppose 0 ≤ p = p(n) ≤ 1, 0 < γ = γ(n) ≤ 1, and 0 < ε = ε(n) ≤ 1

are such that ε2 γpn2 → ∞ as n → ∞. Suppose also that Gn,p →γ H holds almost
surely for some graph H. Then, if 0 ≤ p0 = p0 (n) ≤ 1 is such that p0 ≥ p for all
large enough n, we almost surely have Gn,p0 →γ(1+ε) H.
Proof. Write γ 0 = γ 0 (n) = γ(1 + ε). Let p0 = p0 (n) be as in the statement of
our lemma. Suppose for a contradiction that Gn,p0 →γ 0 H fails with probability
at least θ > 0 for arbitrarily large values of n, where θ is some positive absolute
constant. Put λ = λ(n) = p(n)/p0 (n) ≤ 1. Note that we may generate Gn,p by
first generating Gn,p0 and then randomly removing its edges, each with probabil-
ity 1 − λ, and with all these deletions independent. Looking at this method for
generating Gn,p , we shall deduce below that the probability that (*) Gn,p →γ H
fails is at least θ/3 for arbitrarily large n, which is a contradiction.
Let δ = ε/4. For arbitrarily large n, with probability
at least 2θ/3 we have
that (†) Gn,p0 →γ 0 H fails and e(Gn,p0 ) = (1 + O1 (δ))p0 n2 . Suppose that, when
generating Gn,p by the above method, we first generated a Gn,p0 satisfying (†)
above. Let J = J(Gn,p0 ) ⊂ Gn,p0 be an H-free subgraph of Gn,p0 with e(J) ≥
γ 0 e(Gn,p0 ). Clearly, the H-free subgraph Jλ = J ∩ Gn,p of J is a subgraph of
our Gn,p . We have e(Jλ ) = (1 + O1 (δ))λe(J) and e(Gn,p ) = (1 + O1 (δ))λe(Gn,p0 )
with probability 1 − o(1), and hence we have e(Jλ ) ≥ γe(Gn,p ) with probability 1 −
o(1). Therefore, given that Gn,p0 satisfies (†) above, the probability that we generate
a Gn,p for which (*) fails is 1 − o(1). Since the probability that we generate Gn,p
satisfying (†) is at least 2θ/3 for arbitrarily large n, we conclude that our Gn,p will
fail to satisfy (*) with probability at least θ/3 for arbitrarily large n, which is the
contradiction we were after. 
Proof of Theorem 2. Theorem 2 follows at once from Lemmas 18 and 19. 
A simple variant of the method used in the proof of Lemma 19 gives the following
‘equivalence result’ between the binomial and the uniform models of random graphs
with respect to the property G →γ H.
Lemma 20. Let H be a graph. Consider the following two assertions.
Sbin (γ, p): Gn,p →γ H holds almost surely,
Sunif (γ, M ): Gn,M →γ H holds almost surely,
where 0 < γ = γ(n) ≤ 1, 0 < p = p(n) ≤ 1, and 0 < M = M (n) ≤ n2 are arbitrary

functions. Suppose ω = ω(n) → ∞ as n → ∞. Then the following holds.

(i ) Suppose pn2 → ∞, ω = o(np1/2 ), and h =
h(n) = ωnp1/2 . Let γ 0 =
γ 0 (n) = γ + 2h/M 0 and M 0 = M 0 (n) = dp n2 + he. Then Sbin (γ, p) im-
plies Sunif (γ 0 , M 0 ).
(ii ) Suppose M = M (n) → ∞, ω = o(M 1/2 ), and h = h(n) = ωM 1/2 . Let γ 0 =

−1
γ 0 (n) = γ + 2h/M , and p0 = p0 (n) = (M + h) n2 . Then Sunif (γ, M )
implies Sbin (γ 0 , p0 ).

Proof. Let us prove (i ). Assume Sbin (γ, p) holds. We may generate Gn,M 0 by first
generating Gn,p conditioned on the event E = {e(Gn,p ) ≤ M 0 }, and then randomly
adding M 0 − e(Gn,p ) edges to it so as to have a graph with M 0 edges. For clarity,
let us write GEn,p for our binomial random graph Gn,p conditioned on E. Note
that P(E) = 1 − o(1), and hence the effect of conditioning on E is, so to speak,
negligible. We now claim that Gn,M 0 →γ 0 H holds with probability 1 − o(1).
32 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

Suppose our claim fails, and hence, for arbitrarily large n, there exists with
probability at least θ > 0 an H-free subgraph J ⊂ Gn,M 0 of Gn,M 0 with e(J) ≥
γ 0 e(Gn,M 0 ), where θ is some positive absolute constant. Observe that if J ⊂ Gn,M 0
is an H-free subgraph of Gn,M 0 , then, obviously, J 0 = J ∩ GE E
n,p ⊂ Gn,p is an H-
0
free subgraph of GE E
n,p . Now note that almost surely we have M − e(Gn,p ) ≤ 2h,
E
and hence almost surely ex(Gn,p , H) ≥ ex(Gn,M 0 , H) − 2h. Since we are assuming
that ex(Gn,M 0 , H) ≥ γ 0 e(Gn,M 0 ) with probability θ > 0 for arbitrarily large n, we
have
0 0 0
ex(GE E
n,p , H) ≥ γ e(Gn,M 0 ) − 2h ≥ γ M − 2h ≥ γe(Gn,p )

with probability θ/2 for arbitrarily large n. Since GE n,p is the binomial random
graph Gn,p conditioned on the almost sure event E, we deduce that ex(Gn,p , H) ≥
γe(Gn,p ) with probability at least θ/3 for arbitrarily large n, contradicting Sbin (γ, p).
Thus Sunif (γ 0 , M 0 ) follows. The proof of (ii ) is similar. 
Proof of Corollary 3. Corollary 3 follows easily from Theorem 2 and Lemma 20. 

§5. Deterministic Consequences

In this section we give a few results concerning the existence of very sparse
graphs G = Gη that satisfy G →2/3+η K 4 for any fixed η > 0.
If H is a graph of order |H| ≥ 3 and size e(H) ≥ 1, recall that its 2-density
is d2 (H) = (e(H) − 1)/(|H| − 2). For an integer k ≥ 3, let Hk be the family of all
graphs H with 3 ≤ |H| ≤ k, e(H) ≥ 1, and d2 (H) > d2 (K 4 ). Also, let Forb(Hk ) be
the collection of all graphs G that are H-free for all H ∈ Hk . The following result
may be proved by the so called ‘deletion method’. (For details, see [8, 9].)
Theorem 21. Let 0 < η ≤ 1/3 and k ≥ 3 be fixed. Then there exists a graph G =
Gη,k ∈ Forb(Hk ) such that G →2/3+η K 4 . 
We now single out a corollary to Theorem 21. In Corollary 22 below, the property
that G belongs to Forb(Hk ) in Theorem 21 is replaced by a collection of simpler and
more concrete conditions. For instance, one of these conditions is that G should
not contain a copy of K 5 . To state another condition that appears in Corollary 22,
we need to introduce a definition.
Let G be a graph. Suppose K1 , . . . , Kh (h ≥ 2) are distinct copies of K 4 in G,
S e1 ∈ E(K1 ), . . . , eh−1 ∈ E(Kh−1
and S) are h − 1 edges of G such that E(Ki ) ∩
1≤j<i E(K j ) = {ei−1 } and V (K i )∩ 1≤j<i V (Kj ) = V (ei ) for all 2 ≤ i ≤ h. Then
4
we say that (K1 , . . . , Kh ) is an (h, K )-path in G. Now assume that (K1 , . . . , Kh )
is an (h, K 4 )-path in G and that theSedge e ∈ E(G) joins a vertex in V (K1 ) \
S
1<j≤h V (Kj ) to a vertex in V (Kh ) \ 1≤i<h V (Ki ). Then, (K1 , . . . , Kh ; e) is said
to be an (h, K 4 )-quasi-cycle in G.
4
S It is immediate to check that if (K14, . . . , Kh ) is an (h, K )-path, then H 4=
1≤j≤h Kj has 2-density d2 (H) = d2 (K ). Also, if (K1 , . . . , Kh ; e) is an (h, K )-
quasi-cycle, then H 0 = H + e has 2-density d2 (H 0 ) > d2 (K 4 ).
Corollary 22. For any 0 < η ≤ 1/3 and k ≥ 1, there is a graph G = Gη,k
such that (i ) G contains no K 5 , (ii ) any two copies of K 4 in G share at most
two vertices, (iii ) G contains no (h, K 4 )-quasi-cycles for any 2 ≤ h ≤ k, and
(iv ) G →2/3+η K 4 . 
We remark that Erdős and Nešetřil have raised the question as to whether the
graphs Gη,k as in Corollary 22 exist.
 K 4 -FREE SUBGRAPHS OF RANDOM GRAPHS 33

§6. A Conjecture
In this short paragraph we state a conjecture from which, if true, one may deduce
Conjecture 1. Let H = H h be a graph of order |H| = h ≥ 3 and suppose H has
vertices v1 , . . . , vh . Let 0 < p = p(m) ≤ 1 be given. Let also V = (Vi )hi=1 be a
family of h pairwise disjoint sets, each of cardinality m. Suppose reals 0 < ε ≤ 1
and 0 < γ ≤ 1 and an integer T are given. We say that an h-partite graph F with
h-partition V (F ) = V1 ∪ · · · ∪ Vh and size e(F ) = |F | = T is an (ε, γ, H; V, T )-
graph if the pair (Vi , Vj ) is (ε, F, p)-regular and has p-density γ ≤ dF,p (Vi , Vj ) ≤ 2
whenever vi vj ∈ E(H).
Conjecture 23. Let constants 0 < α ≤ 1 and 0 < γ ≤ 1 be given. Then there are
constants ε = ε(α, γ) > 0 and C = C(α, γ) such that, if p = p(m) ≥ Cm−1/d23 (H) ,
the number of H-free (ε, γ, H; V, T )-graphs is at most
 h
2 
α T 2 m
T

for all T and all sufficiently large m.

If H above is a forest, Conjecture 23 holds trivially, since, in this case, all
(ε, γ, H; V, T )-graphs contain a copy of H. A lemma in Kohayakawa, Luczak,
and Rödl [12] may be used to show that Conjecture 23 holds for the case in
which H = K 3 . In fact, this lemma from [12] is similar in spirit to Lemma 8
above, although much simpler.

Acknowledgements
We are indebted to B. Kreuter for a careful reading of the manuscript.

References
1. Babai, L., Simonovits, M., Spencer, J.H., Extremal subgraphs of random graphs, J. Graph
Theory 14 (1990), 599–622.
2. Bollobás, B., Extremal Graph Theory, Academic Press, London, 1978.
3. , Random Graphs, Academic Press, London, 1985.
4. Chung, F.R.K., Subgraphs of a hypercube containing no small even cycles, J. Graph Theory
16 (1992), 273–286.
5. Frankl, P., Rödl, V., Large triangle-free subgraphs in graphs without K4 , Graphs and Combi-
natorics 2 (1986), 135–244.
6. Füredi, Z., Random Ramsey graphs for the four-cycle, Discrete Math. 126 (1994), 407–410.
7. Haxell, P.E., Kohayakawa, Y., Luczak, T., The induced size-Ramsey number of cycles, Com-
binatorics, Probability, and Computing (to appear).
8. , Turán’s extremal problem in random graphs: forbidding odd cycles, Combinatorica
(to appear).
9. , Turán’s extremal problem in random graphs: forbidding even cycles, J. Combinatorial
Theory, Series B 64 (1995), 273–287.
10. Janson, S., Poisson approximation for large deviations, Random Structures and Algorithms
1 (1990), 221–230.
11. Kohayakawa, Y., Kreuter, B., Steger, A., An extremal problem for random graphs and the
number of graphs with large even-girth (1995), submitted.
12. Kohayakawa, Y., Luczak, T., Rödl, V., Arithmetic progressions of length three in subsets of
a random set, Acta Arithmetica (to appear).
13. Rödl, V., Ruciński, A., Lower bounds on probability thresholds for Ramsey properties, Comb-
inatorics—Paul Erdős is Eighty (Volume 1) (Miklós, D., Sós, V.T., Szőnyi, T., eds.), Bolyai
Soc. Math. Studies, Budapest, 1993, pp. 317–346.
34 Y. KOHAYAKAWA, T. LUCZAK, AND V. RÖDL

14. , Threshold functions for Ramsey properties, J. Amer. Math. Soc. (to appear).
15. Szemerédi, E., Regular partitions of graphs, Problèmes Combinatoires et Théorie des Graphes,
Proc. Colloque Inter. CNRS (Bermond, J.-C., Fournier, J.-C., Las Vergnas, M., Sotteau, D.,
eds.), CNRS, Paris, 1978, pp. 399–401.

Instituto de Matemática e Estatı́stica, Universidade de São Paulo, Rua do Matão

1010, 05508–900 São Paulo, SP, Brazil
E-mail address: <yoshi@[Link]>

Department of Discrete Mathematics, Adam Mickiewicz University, ul. Matejki

48/49, 60–769 Poznań, Poland
E-mail address: <tomasz@[Link]>, <tomasz@[Link]>

Department of Mathematics and Computer Science, Emory University, Atlanta,

Georgia 30322, USA
E-mail address: <rodl@[Link]>