The Ramsey Number R(3, t) has Order of Magnitude t2 / log t

Jeong Han Kim

Mathematical Sciences Research Center

AT&T Bell Laboratories, Murray Hill, NJ 07974
(e-mail addresses : [email protected])

Abstract
The Ramsey number R(s, t) for positive integers s and t is the minimum integer
n for which every red-blue coloring of the edges of a complete n-vertex graph
induces either a red complete graph of order s or a blue complete graph of order
t. This paper proves that R(3, t) is bounded below by (1 − o(1))t2 / log t times a
positive constant. Together with the known upper bound of (1 + o(1))t2 / log t, it
follows that R(3, t) has asymptotic order of magnitude t2 / log t.

1 Introduction
Throughout this paper, logarithms are natural logarithms, c denotes a positive constant, s, t
(3)
and n are positive integers, Kn and Gn denote respectively the complete graph and a triangle-
free (K3 -free) graph on n vertices, and α(G) and χ(G) are respectively the independence
number and the chromatic number of graph G. Our graph theory terminology follows [8], [5].
The Ramsey number R(s, t) is the minimum n such that every red-blue coloring of the
edges of Kn induces either a red Ks or a blue Kt . Equivalently, R(s, t) is the smallest n such
that every n-vertex graph has either an s-vertex clique or a t-vertex independent set. We focus
on
(3) (3)
R(3, t) := min{n : α(Gn ) ≥ t for every Gn } .

Our main result is an upper bound on independence numbers of triangle-free graphs.

(3)
Theorem 1.1 Every sufficiently large n has a Gn for which
(3) p
α(Gn ) ≤ 9 n log n .

Since χ(G) ≥ n/α(G) for every graph G on n vertices, we have the following corollary.

1
 (3)
Corollary 1.2 Every sufficiently large n has a Gn for which
1 n
r
(3)
χ(Gn ) ≥ .
9 log n

An easy consequence of Theorem 1.1 is

t2
c(1 − o(1)) ≤ R(3, t)
log t

with c = 1/162 = 1/(2 · 92 ), where o(1) goes to 0 as t goes to infinity. (We make no attempt
here to find the tightest possible constants.) Because it is known [1], [2], [32] also that

t2
R(3, t) ≤ (1 + o(1)) , (1)
log t

we now know that t2 / log t is the correct asymptotic order of magnitude of R(3, t). Also, (1)
(3)
easily gives an upper bound of χ(Gn ) which, together with Corollary 1.2, yields
1
r
n (3) √ r
n
(1 − o(1)) ≤ max χ(Gn ) ≤ (1 + o(1))2 2 .
9 log n (3)
Gn log n

The asymptotic behavior of R(3, t) has been a major open problem in Ramsey theory for
many years (see e.g [15], Appendix B of [3]). In 1961, Erdős [10] obtained a lower bound
from a lovely probabilistic argument that has become a cornerstone of probabilistic methods
in Combinatorics (see e.g. [3] or [6]). Graver and Yackel [16] found an upper bound in 1968
which, in conjunction with Erdős’s bound, gave

t2 t2 log log t
c1 2
≤ R(3, t) ≤ c2 .
(log t) log t

Ajtai, Komlós and Szemerédi [1], [2] removed the log log t factor in the upper bound, and
Shearer [32] (see also [33]) reduced the constant and simplified the proof to obtain (1).
Meanwhile, the lower bound c1 (t/ log t)2 defied improvement although Spencer [34], Bol-
lobás [6], Erdős, Suen and Winkler [12] and Krivelevich [25] simplified its proof and/or in-
creased c1 through refined probabilistic arguments. We consider parts of their arguments
later. More to the point of the present paper, Spencer [37] showed very recently that c1 can
be arbitrarily large, i.e.
R(3, t)
lim =∞.
t→∞ (t/ log t)2

He introduced also a differential equation which first suggested c t2 / log t as a lower bound and
which inspired the present contribution. We consider this in Section 2.
Our approach uses the so-called “semirandom method” or “Rödl’s nibble method”, a ver-
sion of which may have been used first in [2], the paper that removed the log log t factor in the
upper bound. More refined applications of the semirandom method were subsequently used to
settle intriguing conjectures on hypergraph packings, colorings and list colorings, see e.g. [30],

2
[13], [28], [14], [29], [18], [19]. The present author [22] used a similar method to make progress
on Vizing’s old problem [38] of upper bounds for the chromatic number of a triangle-free graph
of maximum degree D. We refer [18] and [22] for more on the history of the method.
The next section describes our random block construction of a triangle free graph along with
Spencer’s differential equation. Section 3 introduces basic parameters and proves Theorem 1.1
modulo the proof of our main lemma (Lemma 2.1) on the behavior of those parameters under
the block construction. Section 4 presents tools used in the proof of the main lemma, including
Azuma-Hoeffding type martingale inequalities that lead to proofs of the high concentrations
of our parameters near their expected values. The main lemma is then proved in Section 5.

1.1 Block Construction

(3)
One of the most natural ways to construct a random Gn is the following one-by-one construc-
tion. We assume henceforth that every graph has vertex set V := {v1 , ..., vn } and describe a
graph G by its edge set E(G).

One-by-One Construction
(OC 1) Set m := n2 and choose a random permutation π := (e1 , ..., em ) from the uniform

distribution on the m! permutations of the edges of Kn ;

(OC 2) Let G0 = ∅, so E(G0 ) is empty;

(OC 3) Suppose we have Gi . For j > i we say that edge ej “survives” (see [36]) if the graph
E(Gi ) ∪ {ej } has no triangle. Define
(
E(Gi ) ∪ {ei+1 } if ei+1 survives
E(Gi+1 ) :=
E(Gi ) otherwise.

It is obvious that Gm has no triangle. However, it seems hard to find any tight upper
bound on α(Gm ). The main obstacle is the fact that the event “ei+1 survives” depends highly
on the ordering e1 , ..., ei . There is no apparent property that all surviving edges share and
this makes it difficult to find a small upper bound on the variance of the random variable
|E(Gi+1 )| − |E(Gi )|.
Erdős, Suen and Winkler [12] modified the preceding notion of surviving so that an edge
ei+1 survives only if the graph {e1 , ..., ei+1 } has no triangle. Their new notion and a real time
(3)
version of the above construction enable them to prove the existence of a Gn such that
(3) √
α(Gn ) ≤ (3/2 + o(1)) n log n ,

which automatically yields

t2
R(3, t) ≥ (1 − o(1)) .
9(log t)2

3
 Our block construction uses a combination of the above two notions of surviving, but is
closer to the original. For our new construction, we write evw for edge {v, w} ∈ E(Kn ). With
e, f, g ∈ E(Kn ), ef and ef g denote the sets {e, f } and {e, f, g} respectively.

Block Construction with Parameter θ

(BC 1) Let G0 = ∅ and E0 = ∅.

(BC 2) Suppose we have a set Ei of edges and a triangle-free graph Gi with E(Gi ) ⊆ Ei . We
now say that edge e ∈ E(Kn ) \ Ei survives (after stage i) if e cannot be extended to a triangle
using edges from Ei . For small θ > 0 (our θ will be (log n)−2 ) the random set Xi+1 is defined
by ( √
θ/ n if e survives
Pr (e ∈ Xi+1 ) :=
0 otherwise
where all events “e ∈ Xi+1 ” are mutually independent.

(BC 3) We set Ei+1 := Ei ∪Xi+1 and define the set of forbidden pairs of edges in Xi+1 = Ei+1 \Ei
by
Λ(Xi+1 ) := {euv evw : euv evw ⊆ Xi+1 , ewu ∈ Ei } . (2)

Thus ef ∈ Λ(Xi+1 ) means that the pair ef makes a triangle with an edge g in Ei . So we do
not want both e and f in Gi+1 regardless whether g is actually in Gi . The set of forbidden
triples of edges is
∆(Xi+1 ) := {euv evw ewu : euv evw ewu ⊆ Xi+1 } . (3)

We now take a maximal disjoint collection Fi+1 of elements of Λ(Xi+1 ) ∪ ∆(Xi+1 ) and define

E(Gi+1 ) := E(Gi ) ∪ (Xi+1 \ E(Fi+1 )),

[
where E(Fi+1 ) := F.
F ∈Fi+1
This removes all edges which are parts of elements of the maximal collection Fi+1 . Because
of maximality, Gi+1 has no triangles. This method, used by Krivelevich [25], will make the
analysis simpler than the usual one that deletes an edge from each element of Λ(Xi+1 ) ∪
∆(Xi+1 ).
We would like our block construction to do the following. When θ is small enough, the
number of edges in ∪ij=0 E(Fj ) should be small relative to |Ei | so that Gi is approximately Ei .
This is a big advantage because Ei has more random structure than Gi . Moreover, the notion
of surviving in our block construction is now almost as loose as possible since we can not use
any edge that makes a triangle with two edges in Gi (≈ Ei ). A similar idea was first used in
[2], [24].

4
1.2 Differential Equation
As mentioned earlier, Spencer [37] introduced a differential equation to analyze the behavior of
a graph constructed by one-by-one construction. His differential equation also nicely explains
block construction, which suggests that our constructions are similar. We use “≈” to mean
approximately equal and emphasize that what follows is only an aid to our intuition.
Ψ(iθ)n3/2
Suppose Gi has edges, where Ψ is an unknown function. This might occur because
2

Ψ(iθ)n3/2 Ψ(iθ)
Pr (e ∈ Gi ) ≈ n
≈ √ for all e ∈ E(Kn ) . (4)
2 2 n

Recall that “e survives” (e 6∈ Ei ) if and only if there is no pair f, g ∈ Ei which together with
e makes a triangle. Since n − 2 triangles in E(Kn ) contain e, if Gi ≈ Ei and all events are
independent, then (4) would yield
  Ψ(iθ) 2 n−2
Pr ( “ e survives after stage i” ) ≈ 1 − √ ≈ exp(−Ψ2 (iθ)) , (5)
n

where, as usual, Ψ2 (iθ) := (Ψ(iθ))2 . So we expect that the number of surviving edges after
stage i is about n2 exp(−Ψ2 (iθ)).

Therefore, in expectation,

|E(Gi+1 )| ≈ |Ei+1 | = |Ei | + |Xi+1 |

!
Ψ(iθ)n3/2 θ exp(−Ψ2 (iθ)) n
≈ + √
2 n 2
  n3/2
≈ Ψ(iθ) + θ exp(−Ψ2 (iθ)) ,
2
which means that
Ψ((i + 1)θ) ≈ Ψ(iθ) + θ exp(−Ψ2 (iθ)) .

When iθ remains constant, we might obtain

Ψ((i + 1)θ) − Ψ(iθ)

Ψ0 (iθ) = lim ≈ exp(−Ψ2 (iθ)) .
θ→0 θ
This yields the differential equation

Ψ0 (x) = exp(−Ψ2 (x)) .

Since Ψ(0) must be 0, Ψ(x) may be defined implicitly by

Z Ψ(x) 2
x= eξ dξ . (6)
0
√
Thus Ψ(x) is very close to log n for sufficiently large x.

5
 To see where this leads, set a0 := 0, b0 := 1 and
i−1 i−1
0 2
Ψ0 (jθ)θ ≈ Ψ(iθ) for i = 1, 2, · · · .
X X
bi := Ψ (iθ) = exp(−Ψ (iθ)) , ai := bj θ = (7)
j=0 j=0

|T |

We might expect from (5) that every big subset T of V contains bi 2 ≈ bi |T |2 /2 surviving
edges. To be more precise, let
p
t := d9 n log n e , Γi (T ) :≈ {evw ⊆ T : evw survives after stage i}

(cf. (9) in Section 3). In particular, Γ0 (T ) :≈ {evw : evw ⊆ T }. What we expect is

|Γi (T )| ≈ bi t2 /2 for all i and T with |T | = t. (8)

Suppose Gi ≈ Ei . Then (8) and sufficient independence might give that

i−1 i−1
Y θ bj t2 /2  X bj θt2   a t2 
i
Pr (E(Gi ) ∩ Γ0 (T ) = ∅) ≈ 1− √ ≤ exp − √ = exp − √
j=0
n j=0
2 n 2 n

for each fixed t-subset T . Let Ti := {T : |T | = t, Γ0 (T ) ∩ E(Gi ) = ∅}. Then, in expectation,

!
n  a t2  ai t2   √ √
i
 
|Ti | <
∼ exp − √ ≤ exp t log n − √ = exp 9 n (log n)3/2 − (81/2)ai n log n .
t 2 n 2 n

On the other hand, for io := nδ /θ,

io −1
Ψ0 (jθ)θ ≈ Ψ(io θ) = Ψ(nδ ) = (1 + o(1)) δ log n .
X p
aio =
j=0

If δ is not too small, then the expectation of |Tio | is almost 0, which implies α(Gio ) ≤ t with
probability almost 1. Because Gio is triangle-free, we might be done.

This suggests that Theorem 1.1 could follow easily from (8). Our main goal, in fact, is to
derive a property (see Property 7 in Section 3) that is essentially the same as (8). To achieve
it we define a subset Γi of the set of all surviving edges after stage i and adjoin fixed Ei and
Gi to satisfy properties that allows us to continue to the next stage. The triple (Ei , Γi , Gi ) is
no longer random when we begin stage i + 1. A small modification in Γi that gives up a few
surviving edges to gain more regularity is noted early in Section 5.

2 Parameters and Main Lemma

This section presents our parameters and their desired properties. Throughout, (E, Γ, G)
denotes a triple of two subsets of E(Kn ) and a triangle-free graph. We set E0 = G0 = ∅ and

6
Γ0 := E(Kn ). In addition, ∆0 := {ef g ⊆ E(Kn ) : ef g makes a triangle}. Triple (Ei , Γi , Gi ) for
each i is required to satisfy

E(Gi ) ⊆ Ei and Γi ⊆ {e ∈ E(Kn ) \ Ei : ef g 6∈ ∆0 for all f, g ∈ Ei } . (9)

Our parameters depend only on (Ei , Γi , Gi ). We denote the forbidden pairs and triples of
edges in Γi (cf. (2) and (3)) by

Λi := {ef ⊆ Γi : ∃ g ∈ Ei · 3 · ef g ∈ ∆0 }
∆i := {ef g ⊆ Γi : ef g ∈ ∆0 } .

Given v ∈ V , denote its neighborhood and degree in Ei by

NEi (v) := {w ∈ V : evw ∈ Ei } , dEi (v) := |NEi (v)| ,

and the set of edges in Γi containing v as well as its neighborhood and degree in Γi by

NΓi (v) := {evw : evw ∈ Γi }

NΓi (v) := {w ∈ V : evw ∈ Γi } = {w ∈ V : evw ∈ NΓi (v)}
dΓi (v) := |NΓi (v)| = |NΓi (v)| .

For v ∈ V and evw ∈ Γi , we denote the set of incident edges euv of v which together with evw
form forbidden pairs in Γi , and the set of such vertices u, and their (same) size by

NΛi (evw , v) := {euv ∈ Γi : euv evw ∈ Λi }

NΛi (evw , v) := {u ∈ V : euv evw ∈ Λi } = {u ∈ V : euv ∈ NΛi (evw , v)}
dΛi (evw , v) := |NΛi (evw , v)| = |NΛi (evw , v)| .

Set

NΛi (evw ) := NΛi (evw , v) ∪ NΛi (evw , w)

NΛi (evw ) := NΛi (evw , v) ∪ NΛi (evw , w) .

Finally, for evw ∈ Γi , denote the set of vertices each of which together with v, w forms a triangle
in Γi by
N∆i (evw ) := {u ∈ V : euv evw ewu ⊆ Γi , euv evw ewu ∈ ∆0 } .

Our desired properties use the definitions of ai and bi in (7), i.e.,

i−1 i−1
bi := Ψ0 (iθ) = exp(−Ψ2 (iθ)), Ψ0 (jθ)θ = ai−1 + bi−1 θ,
X X
ai := bj θ =
j=0 j=0

where θ := (log n)−2 and Ψ is the function defined in (6). Also, for A, B ⊆ V , let

Γi (A, B) := {evw ∈ Γi : v ∈ A, w ∈ B}, and Γi (A) := Γi (A, A) .

7
 We define eight properties based on the preceding concepts.
√
Property 1. For all v ∈ V , dEi (v) ≤ ai n + in1/4 log n.
Property 2. For all v ∈ V , dΓi (v) ≤ bi n.
Property 3. For all v 6= w in V , |NEi (v) ∩ NEi (w)| ≤ 3i log n.
√
Property 4. For all evw ∈ Γi , dΛi (evw , v) ≤ bi (ai + 5θ) n.
Property 5. For all e ∈ Γi , d∆i (e) ≤ b2i n.
√
Property 6. For all disjoint subsets A, B of V with |A|, |B| ≥ θ2 b2i n,

|Γi (A, B)| ≤ bi |A||B| .

√
Similarly, for all A ⊆ V with |A| ≥ θ2 b2i n
!
|A|
|Γi (A)| ≤ bi .
2
√
The final two properties use t := d9 n log n e,
ai bi−1 θ
µi := 1 − 18ai θ − √ = µi−1 − 18bi−1 θ2 − √
3 log n 3 log n
and
Ti := {T ⊆ V : |T | = t, E(Gi ) ∩ Γ0 (T ) = ∅} .

Property 7. (cf. (8)) For all T ∈ Ti ,

!
t
|Γi (T )| ≥ bi µi .
2

Property 8.
i−1
! !
i n  X bj µj θ t 
|Ti | ≤ n exp − (1 − ) √ ,
t j=0
n 2

where  := (log log n)−1/4 .

If (4) and (5) held (recall ai ≈ Ψ(iθ)) and all events were independent, then all properties
√
would seem quite natural except for terms such as in1/4 log n and 5θbi n, which are basically
error term estimates. For example, we would expect
X
dΛi (evw , v) = 1(ewu ∈ Ei and euv survives)
u∈V \evw
≈ nPr (ewu ∈ Ei and euv survives) in expectation
≈ nPr (ewu ∈ Ei )Pr (ewu survives) assuming independence
√ √
≈ n(ai / n)bi = ai bi n assuming (4) and (5).

8
We note also that all properties are automatic for i = 0, and that Properties 1-6 for i are
needed to guarantee Property 7 for i + 1. Property 8 is a consequence of Property 7, as we
roughly saw in the previous section.

Theorem 1.1 easily follows from our main lemma.

Lemma 2.1 (Main Lemma) Let δ := 1/17−10−5 and 0 ≤ k ≤ bnδ /θc. Suppose (Ek , Γk , Gk )
satisfies (9) and Properties 1 through 8 for i = k. Then some triple (Ek+1 , Γk+1 , Gk+1 ) satisfies
(9) and Properties 1 through 8 for i = k + 1.

The rest of this section examines parameters ai , bi and µi , and proves Theorem 1.1 using
Lemma 2.1. The following lemma presents upper and/or lower bounds of various terms in-
volving the parameters which, while not best possible, are ideally suited to their frequent use
in the rest of the paper.

Lemma 2.2 The following inequalities hold for iθ ≤ nδ and sufficiently large n:

0 ≤ ai − Ψ(iθ) ≤ θ, bi > bi+1 , bi (ai + 5θ) < 1/2 and bi (ai + 5θ)2 < 1/2 ; (10)
q q
log(iθ) − 1 ≤ ai ≤ log(iθ) + 1 + θ , for all iθ ≥ 1 ; (11)

n−1/17 (log n)3 ≤ bi ≤ 1, 1/2 ≤ µi ≤ 1 ; (12)

bi+1
|(1 − 2ai bi θ) − | ≤ 4bi θ2 ; (13)
bi
b2i ≤ bi+1 bi + min{bi+1 bi θ, 2b2i θ(ai + 5θ)} ; (14)
i−1
−4
aj bj ≤ (1/2 + θ1/5 ) log(iθ) i ≥ n10 /θ .
X
θ for all (15)
j=0

Proof. Observe that

Z iθ i−1 Z (j+1)θ
0
Ψ0 (ξ) dξ .
X
Ψ(iθ) = Ψ (ξ) dξ =
0 j=0 jθ

Since Ψ0 (ξ) = exp(−Ψ2 (ξ)) is strictly decreasing and at most 1, clearly bi+1 < bi and
i−1 i−1 Z (j+1)θ i−1
Ψ0 ((j + 1)θ) ≤ Ψ(iθ) = Ψ0 (ξ) dξ ≤ θ Ψ0 (jθ) = ai .
X X X
ai − θ ≤ θ
j=0 j=0 jθ j=0

This verifies the first part of (10). The other parts of (10) follow easily from the first part and
the fact that both y exp(−y 2 ) and y 2 exp(−y 2 ) are less than 0.43.
For (11) and (12) it is enough to point out that
p p
log x − 1 ≤ Ψ(x) ≤ log x + 1 for x ≥ 1 (16)

9
since this implies

Ψ0 (x) = exp(−Ψ2 (x)) = exp(−(1 + o(1)) log x) = x−1+o(1) .

We apply Taylor’s theorem to Ψ0 to obtain

Ψ0 ((i + 1)θ) = Ψ0 (iθ) + θΨ00 (iθ) + (θ2 /2)Ψ000 (ξ) ,

for some ξ with iθ ≤ ξ ≤ (i + 1)θ. Inequalities (13) and (14) follow routinely from this.
For (15) it is enough to consider
i−1 Z iθ i−1
Ψ(ξ)Ψ0 (ξ) dξ + θ2 bj ≤ (1/2 + 2θ)Ψ2 (iθ) ≤ (1/2 + θ1/5 ) log(iθ) ,
X X
θ aj bj ≤ (1 + 2θ)
j=0 0 j=0

where (10) is used in the first inequality, and (16) and θ = (log n)−2 in the last inequality.

Proof of Theorem 1.1 modulo Lemma 2.1. Let ko := bnδ /θc + 1. Then by Lemma
2.1 we have a triangle-free graph Gko that satisfies Property 8. Thus it is enough to show that
kXo −1
! !
ko n  bj µj θ t 
n exp − (1 − ) √ <1 (17)
t j=0
n 2
√
(recall t = 9 n log n.) Inequalities (15) and (11) give

o −1
kX o −1
kX  kXo −1
 1
θ bj µj = θ bj − 18θ + √ θ aj bj
j=0 j=0
3 log n j=0

log(nδ + θ)
≥ ako − (1/6 + θ1/5 ) √
log n
δ log n − 1 − δ(1/6 + 2θ1/5 ) log n
p p
≥
p
≥ 0.23 log n .

Hence
s
kXo −1
! ! !

ko n  bj µj θ t  log n t
log n exp − (1 − ) √ ≤ ko log n + t log n − 0.23(1 − )
t j=0
n 2 n 2
q
≤ n2δ + (9 − (0.23 · 81/2)(1 − 2)) n(log n)3
q
≤ −0.3 n(log n)3 < 0 ,

and (17) follows.

10
3 Tools
3.1 Azuma-Hoeffding type martingale inequalities
Most applications of the semirandom method involve Azuma-Hoeffding type martingale in-
equalities (from [17], [4]), which are very useful in showing that many events can happen
simultaneously. Indeed, Azuma-Hoeffding type martingale inequalities, followed by Lovász’s
local lemma, have become the most popular way to prove the existence of certain packings,
colorings and list colorings mentioned in Section 1. (See [9] for general treatment of probabil-
ity, [35] for Lovász’s local lemma, [27], [7], [26], [21], [19], [22] for more on Azuma-Hoeffding
type inequalities, and [31], [20], [23] for simple applications.)
We need a simple version of the Azuma-Hoeffding type martingale inequality. Let Φ =
Φ(τ1 , ..., τm ), where τ1 , ..., τm are independent identically distributed (i.i.d) Bernoulli random
variables with probability p:

τi ∈ {0, 1} and Pr (τi = 1) = p .

For τ = (τ1 , ..., τm ), τ 0 = (τ10 , ..., τm0 ) and j = 1, 2, ..., m we write

τ ≡j τ 0 if τl = τl0 for all l 6= j.

The following lemma is due to Kahn [19] (Proposition 3.8). We present here a simpler version,
a proof of which can be found in the Appendix.

Lemma 3.1 Suppose

|Φ(τ ) − Φ(τ 0 )| ≤ cj for j = 1, 2, ..., m and τ ≡j τ 0 . (18)

Then for all ρ > 0

 m 
2
c2j exp(ρcj ) .
X
Pr (|Φ − E[Φ]| ≥ λ) ≤ 2 exp − ρλ + (ρ /2)p(1 − p)
j=1

Proof. A proof is given in the Appendix.

We have the following corollaries.

Corollary 3.2 Suppose the hypotheses of Lemma 3.1 hold and ρ > 0 satisfies

ρ max{cj : j = 1, ..., m} ≤ log 2 .

Then m
 
Pr (|Φ − E[Φ]| ≥ λ) ≤ 2 exp − ρλ + ρ2 p c2j .
X

j=1

11
Proof. Since (1 − p) exp(ρcj ) ≤ 2, for all j, we are done.
2
m
X
Corollary 3.3 Let Φ = 1Aj where each Aj is an event that depends only on τj . Then
j=1
 
Pr (|Φ − E[Φ]| ≥ λ) ≤ 2 exp − ρλ + (ρ2 /2)E[Φ] exp(ρ) .

Proof. It is easy to see for all τ ≡j τ 0 that

(
0 0 if Pr (Aj )(1 − Pr (Aj )) = 0
|Φ(τ ) − Φ(τ )| ≤ cj :=
1 otherwise.
Lemma 3.1 says that
m
2
c2j ) .
X
Pr (|Φ − E[Φ]| ≥ λ) ≤ 2 exp(−ρλ + (ρ /2)p(1 − p) exp(ρ)
j=1

Thus it is enough to show

m
c2j ≤ E[Φ] .
X
p(1 − p)
j=1
But since Pr (Aj ) ∈ {0, p, 1 − p, 1},
m m m
c2j .
X X X
E[Φ] = Pr (Aj ) ≥ p(1 − p)cj = p(1 − p)
j=1 j=1 j=1
2

3.2 Disjointness Lemma

Erdős and Tetali introduced a useful lemma in [11]. We will use their proof idea rather than
the lemma itself in Section 5.9, so we simply state the lemma without proof.
Let A be a collection of events in a probability space. We are interested in the simultaneous
occurrence of many independent events in A. Let

In(A) := {B ⊆ A : all events {A}A∈B are mutually independent} .

We want to have a nice upper bound on

[ \
Pr ( A) .
B ∈ In(A) A∈B
|B| = l
X
Lemma 3.4 (Disjointness Lemma). If Pr (A) ≤ η, then
A∈A
[ \ ηl
Pr ( A) ≤ for l = 1, 2, · · · .
l!
B ∈ In(A) A∈B
|B| = l
2

12
3.3 Almost Disjoint Covering Lemma
Suppose A1 , ..., Al are subsets of a set B. In the proof of the Main Lemma (see Sections 5.7
and 5.8), we need good upper bounds on l and j |Aj | when the sets Aj are not too small but
P

their pairwise intersections Aj ∩ Aj 0 are small enough. Notice that when |Aj | ≥ |B|1/2 and
Aj ∩ Aj 0 = ∅ for all j 6= j 0 , we easily have l ≤ |B|1/2 and j |Aj | ≤ |B|. The following lemma
P

considers more general cases.

Lemma 3.5 Let B be a set of size at least 4, A1 , ..., Al ⊆ B, and 1 ≤ β, γ ≤ |B|1/2 /2. If

|Aj | ≥ 2βγ|B|1/2 and |Aj ∩ Aj 0 | ≤ β 2 for all j 6= j 0 ∈ [l] := {1, ..., l}

then
l  1 −1
l ≤ β −1 γ −1 |B|1/2
X
and |Aj | ≤ 1 − |B| .
j=1
2γ 2

Proof. Suppose to the contrary of the first part that l ≥ l0 := bβ −1 γ −1 |B|1/2 c + 1. Then
l0 l0
[ X X
|B| ≥ | Aj | ≥ |Aj | − |Aj ∩ Aj 0 |
j=1 j=1 1≤j<j 0 ≤l0

≥ 2βγ|B|1/2 l0 − (l0 )2 β 2 /2
≥ 2|B| − |B|/2 > |B| .

This is a contradiction.
For the second part, set

A0j := Aj \
[
Aj for all j ∈ [l].
j 0 6=j

Then A01 , ..., A0l are mutually disjoint and

 1 
|A0j | ≥ |Aj | − |Aj ∩ Aj 0 | ≥ |Aj | − β −1 γ −1 |B|1/2 β 2 ≥ 1 −
X
|Aj | .
j 0 6=j
2γ 2

Therefore,
l l l
X  1 −1 X 0
 1 −1 [ 0
 1 −1
|Aj | ≤ 1 − |Aj | = 1 − | Aj | ≤ 1 − |B| .
j=1
2γ 2 j=1
2γ 2 j=1
2γ 2

4 Proof of the Main Lemma

4.1 Preliminaries
Suppose (Ek , Γk , Gk ) as stated in Lemma 2.1. We first modify Γk and Λk to obtain exact
equality in Property 4.

13
Modification. For each pair (evw , v) with evw ∈ Γi , let U (evw , v) be a set of bbk (ak +
√
5θ) n c − dΛk (evw , v) new vertices so that U (evw , v) ∩ V = ∅ and all U (evw , v) are mutually
disjoint. Define

V ∗ := V ∪
[
U (evw , v)
(evw ,v)

Γ∗k := Γk ∪
[
{euv : u ∈ U (evw , v)}
(evw ,v)

and
Λ∗k := Λk ∪
[
{euv evw : u ∈ U (evw , v)} .
(evw ,v)

We also define N ∗ (e, v), N ∗ (e, v), · · · analogously as in Section 3. Notice that
( √
bk (ak + 5θ) n if evw ∈ Γk
d Λ∗k (evw , v) = (19)
1 if evw ∈ Γ∗k \ Γk .

√
Given θ := (log n)−2 and p := θ/ n , we now define a random subset Xk+1
∗ of Γ∗k (cf. (BC
2) of Section 2):
∗
Pr (e ∈ Xk+1 ) = p for all e ∈ Γ∗k
∗ ” mutually independent. Let
with all events “e ∈ Xk+1

∗
Xk+1 := Xk+1 ∩ Γk
Ek+1 := Ek ∪ Xk+1 .

Notice that for all e ∈ Γk ,

∗ √
Pr (e ∈ Xk+1 ) = Pr (e ∈ Xk+1 ) = p = θ/ n , (20)

and all events “e ∈ Xk+1 ” are mutually independent. Using (BC 3) of Section 2 we may also
define a triangle-free graph Gk+1 . Finally, set

∗
Yk+1 := {e ∈ Γk : ∃ f ∈ Xk+1 · 3 · ef ∈ Λ∗k }
∗
Zk+1 := {e ∈ Γk : ∃ f, g ∈ Xk+1 · 3 · ef g ∈ ∆k }

and
Γk+1 := Γk \ (Xk+1 ∪ Yk+1 ∪ Zk+1 ) . (21)

The (random) triple (Ek+1 , Γk+1 , Gk+1 ) obviously satisfies (9). It remains to verify

Pr ((Ek+1 , Γk+1 , Gk+1 ) satisfies Properties 1-8) > 0 .

14
We prove this by showing that (Ek+1 , Γk+1 , Gk+1 ) satisfies each property with probability at
least 1 − 1/n, that is,

Pr ((Ek+1 , Γk+1 , Gk+1 ) does not satisfy Property l) ≤ 1/n for l = 1, ..., 8 . (22)

Three preliminary lemmas will be needed. Henceforth, we fix k ≤ bnδ /θc, omit subscript
k (Γ = Γk , a = ak , dΛ∗ (v) = dΛ∗k (v), · · ·) and write Γ0 , a0 , · · · for Γk+1 , ak+1 , · · ·. Also, we just
write X 0 for (X ∗ )0 and generally omit the asterisk if there is another superscript.

Lemma 4.1 For all e ∈ Γ

b0
−14 b θ2 ≤ Pr (e 6∈ Y 0 ) − ≤ −5bθ2 .
b

Proof. By (19) and NΛ∗ (evw , v) ∩ NΛ∗ (evw , w) = ∅

√
Pr (e 6∈ Y 0 ) = Pr (f 6∈ X ∗ ) = (1 − p)2b(a+5θ) n
Y
for all e ∈ Γ .
f ∈NΛ∗ (e)
√
Also, since p = θ/ n and

1 − hx ≤ (1 − x)h ≤ 1 − hx + h2 x2 /2 for all 0 ≤ x ≤ 1, h ≥ 2 ,

(10) gives

1 − 2abθ − 10bθ2 ≤ Pr (e 6∈ Y 0 ) ≤ 1 − 2bθ(a + 5θ) + 2θ2 b2 (a + 5θ)2 ≤ 1 − 2abθ − 9bθ2 .

The upper and lower bounds of the lemma follow from (13).

Lemma 4.2 Let A ⊆ Γ. Then

X √
|NΛ∗ (e) ∩ A| = 2b(a + 5θ) n |A| .
e∈Γ∗

Proof. By (19),

1(ef ∈ Λ∗ )
X X X
|NΛ∗ (e) ∩ A| =
e∈Γ∗ e∈Γ∗ f ∈A

1(ef ∈ Λ∗ )
X X
=
f ∈A e∈Γ∗
X
= dΛ∗ (f )
f ∈A
√
= 2b(a + 5θ) n |A| . 2

15
 We will prove (22) one property at a time. In most cases we first compute the expectations
of corresponding random variables, then derive good concentration results using the Azuma-
Hoeffding type inequalities of Section 4. Unless specified otherwise, our i.i.d Bernoulli random
variables are {τe }e∈Γ∗ with τe = 1 if and only if e ∈ X ∗ :
√
Pr (τe = 1) = p = θ/ n .

Applications of Corollary 3.2 will simply note the ce in the hypotheses of Lemma 3.1. If we
do not mention ce for some edges, then those edges are irrelevant.

The following lemma completes our preliminaries. Let

NX 0 (v) := {w ∈ V : evw ∈ X 0 } for v ∈ V .

Lemma 4.3 The following three conditions hold simultaneously with probability at least 1 −
3/n2 :
√
(i) For all v ∈ V , |NX 0 (v)| ≤ bθ n + n1/4 log n;
(ii) For all v 6= w ∈ V , |NE (v) ∩ NX 0 (w)| ≤ log n;
(iii) For all v 6= w ∈ V , |NX 0 (v) ∩ NX 0 (w)| ≤ log n.

Proof. We show that each of (i), (ii), and (iii) occurs with probability at least 1 − 1/n2 .
For (i), let
1(evw ∈ X 0 ) .
X
Φv := |NX 0 (v)| =
w∈NΓ (v)

Property 2 (for i = k, of course) and (20) give

√
Pr (evw ∈ X 0 ) ≤ pbn = bθ n .
X
E[Φv ] =
w∈NΓ (v)

Thus it is enough to show

Pr (Φv − E[Φv ] ≥ n1/4 log n) ≤ 1/n3 for all v∈V ,

which gives

Pr (∃ v ∈ V · 3 · Φv − E[Φv ] ≥ n1/4 log n) ≤ n(1/n3 ) = 1/n2 .

Corollary 3.3 with ρ = 4n−1/4 yields

√
Pr (Φv − E[Φv ] ≥ n1/4 log n) ≤ 2 exp(−ρn1/4 log n + ρ2 bθ n )
≤ exp(−3 log n) = 1/n3 .

For (ii), let

Φ(1) 1(ewu ∈ X 0 ) .
X
v,w := |NE (v) ∩ NX 0 (w)| =
u∈NE (v)

16
Property 1 and (11) give
E[Φ(1)
v,w ] ≤ p|NE (v)| ≤ θ
1/2
≤1,
and Corollary 3.3 with ρ = 5 yields

Pr (Φ(1) (1) (1)

v,w ≥ log n) ≤ Pr (Φv,w − E[Φv,w ] ≥ log n − 1)

≤ 2 exp(−ρ(log n − 1) + ρ2 )
≤ exp(−4 log n) = 1/n4 .

Because there are at most n2 pairs of v, w, we are done.

For (iii), let
Φ(2) 1(euv ewu ⊆ X 0 ) .
X
v,w := |NX 0 (v) ∩ NX 0 (w)| =
u∈V
Our i.i.d Bernoulli random variables in this case are τu := τeuv τewu , indexed by u ∈ V . Because

Pr (τu = 1) ≤ p2 = θ2 /n ,

we have
E[Φ(2) 2 2
v,w ] ≤ p n = θ < 1 .

Let ρ = 5. Then

Pr (Φ(2) (2) (2)

v,w ≥ log n) ≤ Pr (Φv,w − E[Φv,w ] ≥ log n − 1)

≤ 2 exp(−ρ(log n − 1) + ρ2 )
≤ exp(−4 log n) = 1/n4 .

This completes the proof.

We now consider each property separately to prove (22). The definitions of Φv , Φv,w , · · ·
vary from property to property and therefore hold only within subsections. Also, to prove (22)
for the properties 2,4 and 5, we show that each vertex or pair of vertices violates the properties
with probability at most 1/n2 or 1/n3 , respectively. This is enough because there are at most
n vertices and n2 pairs.

4.2 Property 1
Since
NE 0 (v) = NE (v) ∪ NX 0 (v) ,
Property 1 and (i) of Lemma 4.3 give
√ √
dE 0 (v) = dE (v) + |NX 0 (v)| ≤ a n + kn1/4 log n + bθ n + n1/4 log n
√
= a0 n + (k + 1)n1/4 log n

with probability at least 1 − 3/n2 > 1 − 1/n.

17
4.3 Property 2
Let
1(e 6∈ Y 0 ) .
X
Φv :=
e∈NΓ (v)

Then dΓ0 (v) ≤ Φv by (21). Property 2 and Lemma 4.1 imply

Pr (e 6∈ Y 0 )
X
E[Φv ] =
e∈NΓ (v)

≤ bn(b0 /b − 5bθ2 ) ≤ b0 n − b2 θ2 n .

We take
√
ce = |NΛ∗ (e) ∩ NΓ (v)| ≤ 2b(a + 5θ) n

for all e ∈ Γ∗ . Lemma 4.2 and Property 2 then give

√ X
c2e ≤ 2b(a + 5θ) n ce ≤ 4b2 (a + 5θ)2 n · bn ≤ n2 .
X

e e

Together with (12) and Corollary 3.2 with ρ = n−3/4 , this yields
√
Pr (Φv − E[Φv ] ≥ b2 θ2 n) ≤ 2 exp(−b2 θ2 n1/4 + (θ/ n )n1/2 )
≤ exp(−n1/4−2/17 ) ≤ 1/n2 .

4.4 Property 3
Since

|NE 0 (v)∩NE 0 (w)| ≤ |NE (v)∩NE (w)|+|NX 0 (v)∩NE (w)|+|NE (v)∩NX 0 (w)|+|NX 0 (v)∩NX 0 (w)| ,

we are done by Property 3 and (ii), (iii) of Lemma 4.3.

4.5 Property 4
For evw ∈ Γ \ X 0 let
Φ(1) 1(e 6∈ Y 0 )
X
v,w :=
e∈NΛ (evw ,v)

and
Φ(2) 1(ewu ∈ X 0 ) .
X
v,w :=
u∈N∆ (evw )
(1) (2)
/ X 0 . Clearly,
Note that Φv,w and Φv,w are considered under the condition evw ∈

dΛ0 (evw , v) ≤ Φ(1) (2)

v,w + Φv,w for all evw ∈ Γ0 . (23)

Because events “e ∈ X 0 ” are mutually independent, Property 5 yields

0 2 √
E[Φ(2) (2)
v,w |evw 6∈ X ] = E[Φv,w ] = pd∆ (evw ) ≤ b θ n .

18
Thus (12) and Corollary 3.3 with ρ = n−1/4 imply that
2 √ √ √
Pr (Φ(2) 2 2 (2) (2) 2 2
v,w − b θ n ≥ b θ (a + 5θ) n ) ≤ Pr (Φv,w − E[Φv,w ] ≥ b θ (a + 5θ) n )

≤ 2 exp(−b2 θ2 (a + 5θ)n1/4 + b2 θ)
≤ exp(−n1/4−2/17 ) ≤ 1/n4 . (24)
(1)
We now consider Φv,w . Since

Pr (e 6∈ Y 0 |evw 6∈ X 0 ) = Pr (e 6∈ Y 0 )(1 − p)−1 for all e ∈ NΛ (evw , v)

Lemma 4.1 and Property 4 give

0
E[Φ(1)
v,w ] ≤ (1 + 2p)Pr (e 6∈ Y )dΛ (evw , v)
√ √
≤ (1 + 2θ/ n )(b0 /b − 5bθ2 ) b(a + 5θ) n
√ √
≤ b0 (a + 5θ) n − 4b2 θ2 (a + 5θ) n . (25)

Let
ce := |NΛ (evw , v) ∩ NΛ∗ (e)| for e 6= evw .

Since
NΛ (euv , v) ⊆ NE (u) for all euv ∈ Γ, (26)

Property 3 implies for euv ∈ Γ that

|NΛ (evw , v) ∩ NΛ∗ (euv , v)| = |NΛ (evw , v) ∩ NΛ (euv , v)|

≤ |NE (w) ∩ NE (u)| ≤ 3k log n .

Hence, for all e ∈ Γ∗ \ {evw },

ce ≤ 1 + 3k log n .

Thus Lemma 4.2 and Property 4 yield

c2e ≤ (1 + 3k log n) ce = 2(1 + 3k log n)b2 (a + 5θ)2 n .

X X

e∈Γ∗ \{evw } e∈Γ∗ \{evw }

Let ρ = n−1/4 . Then

√
Pr (Φ(1) (1) 2 2
v,w − E[Φv,w ] ≥ b θ (a + 5θ) n )
√
≤ 2 exp(−b2 θ2 (a + 5θ)n1/4 + 2(θ/ n )(1 + 3k log n)b2 (a + 5θ)2 n1/2 )
≤ exp(−n1/4−2/17 ) ≤ 1/n4 . (27)

It now follows from (25), (27), (24), and (14) that, with probability at least 1 − 2/n4 ,
0 √ √ 2 √ 0 0 √
Φ(1) (2) 2 2
v,w + Φv,w ≤ b (a + 5θ) n − 2b θ (a + 5θ) n + b θ n ≤ b (a + 5θ) n .

By (23) we are done.

19
4.6 Property 5
Let
1(euv ewu ∩ Y 0 = ∅)
X
Φvw :=
u∈N∆ (evw )

for each evw ∈ Γ. Then

d∆0 (evw ) ≤ Φvw . (28)

Because (26) yields

|NΛ (euv , u) ∩ NΛ (ewu , u)| ≤ |NE (v) ∩ NE (w)| ≤ 3k log n ,

Lemma 4.1 gives

Pr (euv ewu ∩ Y 0 = ∅) = Pr (euv 6∈ Y 0 )Pr (ewu 6∈ Y 0 |euv 6∈ Y 0 )

≤ Pr (euv 6∈ Y 0 )Pr (ewu 6∈ Y 0 )(1 − p)−3k log n
√
≤ (b0 /b − 5bθ2 )2 (1 + 4θk(log n)/ n )
≤ (b0 /b)2 − bθ2 .

Property 5 easily yields

Pr (euv ewu ∩ Y 0 = ∅)
X
E[Φvw ] =
u∈N∆ (evw )

≤ ((b0 /b)2 − bθ2 )b2 n ≤ (b0 )2 n − b3 θ2 n . (29)

Let
 √
 b(a + 5θ) n
 if NΛ∗ (e) ∩ N∆ (evw ) 6= ∅ and evw ∩ e 6= ∅
ce := 2 if NΛ∗ (e) ∩ (N∆ (evw ) ∪ evw ) 6= ∅ and evw ∩ e = ∅
 0

otherwise

for e ∈ Γ∗ . (Actually, we may take ce = 1 for the second case.) Clearly, ce ≥ 2 for at most
√
2b(a + 5θ) n · 2b2 n edges (cf. Lemma 4.2), and ce > 2 for at most 2n edges. Hence

c2e ≤ 4 · 4b3 (a + 5θ)n3/2 + 2nb2 (a + 5θ)2 n ≤ 3b2 (a + 5θ)2 n2 .

X

Then (28), (29), (12), and Corollary 3.2 with ρ = n−5/8 imply that

Pr (d∆0 (evw ) ≥ (b0 )2 n) ≤ Pr (Φvw − E[Φvw ] ≥ b3 θ2 n)

√
≤ 2 exp(−b3 θ2 n3/8 + 3(θ/ n )b2 (a + 5θ)2 n3/4 )
≤ exp(−n3/8−3/17 ) ≤ 1/n3 .

20
4.7 Property 6
We prove only the first part. An analogous proof holds for the other part.
√
It is enough to prove the property for all A, B with |A| = |B| = b2 θ2 n . (Of course, we
√
should really write bb2 θ2 n c here.) Let

L := {(e, v) ∈ Γ∗ × V ∗ : v ∈ e}

and for A, B ⊆ V set

L(1) (A, B) := {(e, v) ∈ L : |NΛ∗ (e, v) ∩ Γ(A, B)| < 4((k + 1) log n)1/2 |A ∪ B|1/2 } . (30)

Because
Y0 =
[
NΛ∗ (e, v) ∩ Γ , (31)
e∈ X∗
(e, v) ∈ L

we further define

Y (1) (A, B) := 1(e 6∈ Y (1) (A, B)) .

[ X
NΛ∗ (e, v) ∩ Γ(A, B) and ΦA,B :=
e∈ X∗ e∈Γ(A,B)
(e, v) ∈ L(1) (A, B)

Then
|Γ0 (A, B)| ≤ ΦA,B . (32)

In regard to the expectation of ΦA,B , we claim that, for all evw ∈ Γ(A, B),
1/4
Pr (evw 6∈ Y (1) (A, B)) ≤ Pr (evw 6∈ Y 0 )(1 − p)−2n . (33)

This holds if
|{u ∈ NΛ∗ (evw , v) : (euv , v) 6∈ L(1) (A, B)}| ≤ n1/4 (34)

and
|{u ∈ NΛ∗ (evw , w) : (ewu , w) 6∈ L(1) (A, B)}| ≤ n1/4 . (35)

By (26), (euv , v) 6∈ L(1) (A, B) implies that

4((k + 1) log n)1/2 |A ∪ B|1/2 ≤ |NΛ∗ (euv , v) ∩ Γ(A, B)|

≤ |NΛ (euv , v) ∩ (A ∪ B)| ≤ |NE (u) ∩ (A ∪ B)| .

This inequality, Property 3, and Lemma 3.5 with β = (3(k + 1) log n)1/2 , γ = 1 then imply
that there are at most (3(k + 1) log n)−1/2 |(A ∪ B)|1/2 such u ∈ V ∗ . Since

(3(k + 1) log n)−1/2 |(A ∪ B)|1/2 ≤ n1/4 ,

(34) follows. By the same method (35) also holds.

21
 √
Lemma 4.1, Property 6, (33), and our condition of |A| = |B| = b2 θ2 n imply

Pr (evw 6∈ Y (1) (A, B)) ≤ (b0 /b − 5bθ2 )(1 + n−1/4 ) ≤ b0 /b − bθ2

and
E[ΦA,B ] ≤ (b0 /b − bθ2 )|Γ(A, B)| ≤ b0 |A||B| − b6 θ6 n . (36)
Let

ce := min{4((k + 1) log n)1/2 |A ∪ B|1/2 , |NΛ∗ (e) ∩ Γ(A, B)|} ≤ 6((k + 1) log n)1/2 bθn1/4

for all e ∈ Γ∗ . Then Lemma 4.2 and Property 6 give

c2e ≤ 6((k + 1) log n)1/2 bθn1/4

X X
ce
≤ 6((k + 1) log n)1/2 bθn1/4 · 2b(a + 5θ)n1/2 |Γ(A, B)|
≤ 12((k + 1) log n)1/2 b7 θ5 (a + 5θ)n7/4 .

It follows from (12) and Corollary 3.2 with ρ = n−1/4−εo , where εo := 1/4 − 4/17 = 1/68, that

Pr (ΦA,B − E[ΦA,B ] ≥ b6 θ6 n)
 √ 
≤ 2 exp − b6 θ6 n3/4−εo + 12(θ/ n )((k + 1) log n)1/2 b7 θ5 (a + 5θ)n5/4−2εo
2
≤ exp(−b θ2 (log n)4 n3/4−4/17−εo /2)
2
≤ exp(−b θ2 (log n)2 n1/2 ) .

This result along with (32) and (36) then gives

 √ 
Pr ∃ disjoint A, B ⊆ V with |A| = |B| = b2 θ2 n · 3 · |Γ0 (A, B)| > b0 |A||B|
 √ 
≤ Pr ∃ disjoint A, B ⊆ V with |A| = |B| = b2 θ2 n · 3 · ΦA,B − E[ΦA,B ] ≥ b6 θ6 n
!2
n
≤ exp(−b2 θ2 n1/2 (log n)2 )
b2 θ2 n1/2

≤ exp(2b2 θ2 n1/2 log n − b2 θ2 n1/2 (log n)2 ) ≤ 1/n .

4.8 Property 7
√
Recall for T ∈ T that |T | = t = d9 n log n e. Let T ∈ T . (Actually, our proof works for all T
of size t.) We know by (21) that

|Γ0 (T )| ≥ |Γ(T )| − |X 0 ∩ Γ(T )| − |Y 0 ∩ Γ(T )| − |Z 0 ∩ Γ(T )| . (37)

We verify Property 7 by first proving

Pr (∃ T ∈ T · 3 · |X 0 ∩ Γ(T )| ≥ b0 θ2 |Γ(T )|/2 ) ≤ 1/n2 . (38)

22
Proof of (38). Set
(0)
ΦT := |X 0 ∩ Γ(T )| = 1(e ∈ X 0 ) .
X

e∈Γ(T )

Then
(0) √
E[ΦT ] = |Γ(T )|Pr (e ∈ X 0 ) = (θ/ n )|Γ(T )| ,

and Corollary 3.3 with ρ = n−5/17 , Property 7 and (12) give

(0) (0) (0)
Pr (ΦT ≥ b0 θ2 |Γ(T )|/2) ≤ Pr (ΦT − E[ΦT ] ≥ b0 θ2 |Γ(T )|/4)
√
≤ 2 exp(−ρb0 θ2 |Γ(T )|/4 + ρ2 (θ/ n )|Γ(T )|)
≤ exp(−ρbb0 θ2 n)
≤ exp(−n1−5/17−2/17 ) = exp(−n10/17 ) .

Thus
!
(0) 0 2 n
Pr (∃ T ∈ T ·3· ΦT ≥ b θ |Γ(T )|/2 ) ≤ exp(−n10/17 )
t
√
≤ exp(9 n (log n)3/2 − n10/17 ) ≤ 1/n2 .

We now divide |Y 0 ∩ Γ(T )| into two parts (cf. (30)). Let

L(1) (T ) := {(e, v) ∈ L : |NΛ∗ (e, v) ∩ Γ(T )| < 4((k + 1) log n)1/2 |T |1/2 },
L(2) (T ) := L \ L(1) (T ) ,

and

Y (1) (T ) :=
[
NΛ∗ (e, v) ∩ Γ(T ) ,
e∈ X∗
(e, v) ∈ L(1) (T )

Y (2) (T ) :=
[
NΛ∗ (e, v) ∩ Γ(T ) .
e ∈ X∗
(e, v) ∈ L(2) (T )

The corresponding random variables are

(1) (2)
ΦT := |Γ(T )| − |Y (1) (T )| = 1(e 6∈ Y (1) (T )) and ΦT := |Y (2) (T )| .
X

e∈Γ(T )

By (31),
(1) (2)
|Γ(T )| − |Y 0 ∩ Γ(T )| ≥ ΦT − ΦT .

We claim that
(1) 2
Pr (∃ T ∈ T · 3 · ΦT ≤ (b0 /b − 15b0 θ )|Γ(T )| ) ≤ 1/n2 (39)

23
and that !

(2) 2b0 bθ(1 + 6θ) t 
Pr ∃ T ∈ T · 3 · ΦT > √ ≤ 3/n2 , (40)
9 log n 2
which imply directly that, with probability at least 1 − 4/n2 ,
!
0 0 2b0 bθ(1 + 6θ) t
0 2
|Γ(T )| − |Y ∩ Γ(T )| ≥ (b /b − 15b θ )|Γ(T )| − √ . (41)
9 log n 2

Proof of (39). Lemma 4.1 (the lower bound) gives

(1)
Pr (e 6∈ Y (1) (T )) ≥ Pr (e 6∈ Y 0 ) ≥ (b0 /b − 14bθ2 )|Γ(T )| .
X X
E[ΦT ] =
e∈Γ(T ) e∈Γ(T )

Since (13) yields b = (1 + O(θ))b0 , we have

(1) 2 (1) (1) 2
Pr (ΦT ≤ (b0 /b − 15b0 θ )|Γ(T )|) ≤ Pr (ΦT − E[ΦT ] ≤ −b0 θ |Γ(T )|/2) . (42)

Set
ce := min{|NΛ∗ (e) ∩ Γ(T )|, 8((k + 1) log n)1/2 t1/2 } .

Then Lemma 4.2 yields

c2e ≤ 8((k + 1) log n)1/2 t1/2

X X
ce
1/2 1/2 √
≤ 16((k + 1) log n) t b(a + 5θ) n |Γ(T )| .

Let ρ = n−1/4−1/17 . Then Property 7 and (12) imply that

(1) (1) 2
Pr (ΦT − E[ΦT ] ≤ −b0 θ |Γ(T )|/2)
 √ √ 
≤ 2 exp − ρb0 θ2 |Γ(T )|/2 + 16ρ2 (θ/ n )((k + 1) log n)1/2 t1/2 b(a + 5θ) n |Γ(T )|
√
≤ exp(−n1−1/4−3/17 ) ≤ exp(− n (log n)2 ) .

We combine this with (42) and the inequality

!
√ n
exp(− n (log n)2 ) ≤ 1/n2
t

to obtain (39).

Proof of (40). It is enough to show that (i), (ii), and (iii) of Lemma 4.3 imply (40).

Take NE (w, T ) := NE (w) ∩ T and let

WT := {w ∈ V : |NE (w, T )| ≥ 4((k + 1) log n)1/2 t1/2 } .

24
We show first that
Y (2) (T ) ⊆
[
Γ(NE (w, T ), NX 0 (w, T )) (43)
w∈WT

where, as usual, NX 0 (w, T ) := NX 0 (w) ∩ T .

Suppose f ∈ Y (2) (T ) ⊆ Γ(T ). Then there exists (e, v) ∈ L(2) (T ), say e = evw , such that
f ∈ NΛ∗ (evw , v) and evw ∈ X 0 . In particular, v ∈ f ⊆ T , and the other vertex u of f is in
NE (w, T ). Moreover, since evw ∈ X 0 , we have v ∈ NX 0 (w, T ). Since (evw , v) ∈ L(2) (T ), we
know also by (26) that

4((k + 1) log n)1/2 t1/2 ≤ |NΛ∗ (evw , v) ∩ Γ(T )| = |NΛ (evw , v) ∩ T | ≤ |NE (w, T )| .

Thus w ∈ WT , and the proof of (43) is complete.

Let (
NE (w, T ) if |NE (w, T )| ≥ |NX 0 (w, T )|
A(w, T ) :=
NX 0 (w, T ) otherwise
and for summations let
X0 X X 00 X
:= and := .
w ∈ WT √ w ∈ WT √
min{|NE (w, T )|, |NX 0 (w, T )|} ≤ b2 θ2 n min{|NE (w, T )|, |NX 0 (w, T )|} > b2 θ2 n

Note that

|A(w, T )| ≥ |NE (w, T )| ≥ 4((k + 1) log n)1/2 t1/2 for all w ∈ WT . (44)

Property 6, (i) of Lemma 4.3, and (43) yield

|Y (2) (T )| ≤
X
|Γ(NE (w, T ), NX 0 (w, T ))|
w∈WT
X0 X 00
≤ |NE (w, T )||NX 0 (w, T )| + b |NE (w, T )||NX 0 (w, T )|
√ X0 √ X 00
≤ b2 θ2 n |A(w, T )| + (1 + θ)b2 θ n |NE (w, T )| .
p
Moreover, Lemma 3.5 (with β = 3(k + 1) log n for both of the following inequalities, γ = 1
for the first inequality, and γ = log n for the second), Property 3, (ii) and (iii) of Lemma 4.3,
and (44) imply
X0 X 00
|A(w, T )| ≤ 2t and |NE (w, T )| ≤ (1 + θ)t .

Therefore, by (14),
!
(2) 2 2√ 2 2 √2b0 bθ(1 + 6θ) t
|Y (T )| ≤ 2b θ n t + (1 + θ) b θ n t ≤ √ .
9 log n 2
2

25
 We return to (37) and divide |Z 0 ∩ Γ(T )| into two parts. We recall that NX 0 (v, T ) =
NX 0 (v) ∩ T and note that

Z 0 ∩ Γ(T ) =
[ [ [
Γ(NX 0 (v)) ∩ Γ(T ) = Γ(NX 0 (v) ∩ T ) = Γ(NX 0 (v, T )) .
v∈V v∈V v∈V

Let h := d4((k + 1) log n)1/2 t1/2 e. Given NX 0 (v, T ) = {vj , ..., vjr } with j1 < · · · < jr , let
1
(
NX 0 (v, T ) if r ≤ h
N̂X 0 (v, T ) :=
{vj , ..., vj } if r > h.
1 h

Then
Z 0 ∩ Γ(T ) =
[ [
Γ(N̂X 0 (v, T )) ∪ Γ(NX 0 (v, T )) .
v∈V v∈V
|NX 0 (v)| ≥ h

Let
(3) X (4) X
ΦT := |Γ(N̂X 0 (v, T ))|, and ΦT := |Γ(NX 0 (v, T ))| .
v∈V v∈V
|NX 0 (v, T )| ≥ h

Clearly
(3) (4)
|Z 0 ∩ Γ(T )| ≤ ΦT + ΦT . (45)

We now claim that

(3) √
Pr (ΦT ≥ 2b0 θ2 |Γ(T )| ) ≤ 2 exp(− n (log n)2 ) for all T ∈ T , (46)

which implies
(3)
Pr (∃T ∈ T · 3 · ΦT ≥ 2b0 θ2 |Γ(T )| ) ≤ 1/n2 . (47)

Proof of (46). Property 5 gives, for expectation,

(3) X
E[ΦT ] ≤ E[ |Γ(NX 0 (v, T ))| ]
v∈V

Pr (euv evw ⊆ X 0 )
X X
=
v∈V ewu ∈ Γ(T )
v ∈ N∆ (ewu )

= p2
X X
1
ewu ∈Γ(T ) v∈N∆ (ewu )

≤ p2 b2 n|Γ(T )| = b2 θ2 |Γ(T )| ≤ (3/2)b0 θ2 |Γ(T )| .

Thus
(3) (3) (3)
Pr (ΦT ≥ 2b0 θ2 |Γ(T )| ) ≤ Pr (ΦT − E[ΦT ] ≥ b0 θ2 |Γ(T )|/2) .

We obtain a concentration result by taking

(5) X (6) X
ΦT := |Γ(N̂X 0 (v, T ))| and ΦT := |Γ(N̂X 0 (v, T ))| .
v∈T v∈V \T

26
Since
(3) (5) (6)
ΦT = ΦT + ΦT ,
it is enough to show that
(5) (5) √
Pr (ΦT − E[ΦT ] ≥ b0 θ2 |Γ(T )|/4) ≤ exp(− n (log n)2 ) (48)

and
(6) (6) √
Pr (ΦT − E[ΦT ] ≥ b0 θ2 |Γ(T )|/4) ≤ exp(− n (log n)2 ) . (49)
Consider (48). Let (
2h if e ∈ Γ(T )
ce :=
0 otherwise.
Then
c2e = 4h2 |Γ(T )| .
X

We take ρ = n−1/4−1/17 to obtain

(5) (5) √
Pr (ΦT − E[ΦT ] ≥ b0 θ2 |Γ(T )|/4) ≤ 2 exp(−ρb0 θ2 |Γ(T )|/4 + 4ρ2 (θ/ n )h2 |Γ(T )|)
√
≤ exp(−n1−1/4−3/17 ) ≤ exp(− n (log n)2 ) .

Consider (49). All random variables |Γ(N̂X 0 (v, T ))| for v ∈ V \T are mutually independent,
so
(6) Y
E[exp(ρΦT )] = E[exp(ρ|Γ(N̂X 0 (v, T ))|)] . (50)
v∈V \T

Our aim is to find a good upper bound of E[exp(ρ|Γ(N̂X 0 (v, T ))|)] so that we may apply
Markov’s inequality with the function exp(ρx) (see (53)). Let

φv (ρ) := E[ exp(ρ|Γ(N̂X 0 (v, T ))|) ] .

Then there is ρ∗ with 0 ≤ ρ∗ ≤ ρ such that

φv (ρ) = 1 + ρ E[ |Γ(N̂X 0 (v, T ))| ] + (ρ2 /2)φ00v (ρ∗ ) .

We prove the following claim.

Claim.
φ00v (ρ∗ ) ≤ 1 if ρ∗ ≤ h−1 . (51)

Proof. Let B = B(v, T ) := NΓ (v) ∩ T . Then, since ρ∗ ≤ h−1 ,

φ00v (ρ∗ ) = E[ |Γ(N̂X 0 (v, T ))|2 exp(ρ∗ |Γ(N̂X 0 (v, T ))|) ]

h−1
! !2 ! |B| ! !2 !
|B| l 
∗ l  l |B| h  h  l
p (1 − p)|B|−l + ∗
p (1 − p)|B|−l
X X
≤ exp ρ exp ρ
l=0
l 2 2 l=h
l 2 2
|B| !
1 X |B|
≤ (l + 2)(l + 1)l(l − 1) exp(l/2)pl (1 − p)|B|−l . (52)
4 l=2 l

27
 Let
|B| !
|B| |B|
2
xl+2 (1 − p)|B|−l .
X
ω(x) := x (1 − p + x) =
l=0
l

Then it is clear that the last term of (52) is exactly (p2 exp(1)/4)ω (4) (p exp(1/2)), where ω (4)
is the fourth derivative of ω. Therefore, with

p|B| ≤ pt ≤ (log n)−1 ,

we easily have
(p2 /4)ω (4) (p exp(1/2)) ≤ 1 .

We use 51 and set ρ = n−1/4−1/17 (notice that ρ ≤ h−1 ) to obtain

φv (ρ) ≤ exp(E[ |Γ(N̂X 0 (v, T ))| ] ρ + ρ2 /2) .

Then, by (50),
(6) (6)
E[ |Γ(N̂X 0 (v, T ))| ] + ρ2 n/2) = exp(ρE[ΦT ] + ρ2 n/2) .
X
E[exp(ρΦT )] ≤ exp(ρ
v∈V \T

Thus Markov’s inequality gives

(6) (6) (6) (6)
Pr (ΦT − E[ΦT ] ≥ b0 θ2 |Γ(T )|/4 ) = Pr (exp(ρ(ΦT − E[ΦT ])) ≥ exp(ρb0 θ2 |Γ(T )|/4) )
(6) (6)
≤ exp(−ρb0 θ2 |Γ(T )|/4)E[ exp(ρ(ΦT − E[ΦT ])) ]
≤ exp(−ρb0 θ2 |Γ(T )|/4 + ρ2 n/2)
√
≤ exp(−n1−1/4−3/17 ) ≤ exp(− n (log n)2 ) , (53)

which completes the proof of (49) and therefore of (46).

Finally, we claim that

!

(4) b0 bθ(1 + 6θ) t 
Pr ∃T ∈T ·3· ΦT > √ ≤ 3/n2 . (54)
9 log n 2

Proof of (54). It is enough to show that (i) and (iii) in Lemma 4.3 imply
!
(4) b0 bθ(1 + 6θ) t
ΦT ≤ √ for all T ∈T .
9 log n 2

28
 Let
X∗ X X ∗∗ X
:= and := .
v∈V √ v∈V √
h ≤ |NX 0 (v, T )| < b2 θ2 n |NX 0 (v, T )| ≥ b2 θ2 n

Then Property 6 and (i) of Lemma 4.3 give

(4) X
ΦT = Γ(NX 0 (v, T ))
v∈V
|NX 0 (v, T )| ≥ h
1 X∗ b X ∗∗
≤ |NX 0 (v, T )|2 + |NX 0 (v, T )|2
2 √ 2 √
b2 θ2 n X ∗ (1 + θ)b2 θ n X ∗∗
≤ |NX 0 (v, T )| + |NX 0 (v, T )| .
2 2
√
Also, (iii) of Lemma 4.3 and Lemma 3.5 (with β = log n for both of the following inequalities,
γ = 1 for the first inequality, and γ = 2 log n for the second) give
X∗ X ∗∗
|NX 0 (v, T )| ≤ 2t and |NX 0 (v, T )| ≤ (1 + θ/2)t .

Thus (14) yields

!
(4) 2 2√ b0 bθ(1 + 6θ) t
2 √
ΦT ≤b θ n t + (1/2 + θ)b θ n t ≤ √ .
9 log n 2
2

Therefore, combining (37), (38), (41), (45), (47), (54) and Property 7, we have that, with
probability at least 1 − 1/n,
!
0 0 3b0 bθ(1 + 6θ) t
0 2
|Γ (T )| ≥ (b /b − 35b θ /2)|Γ(T )| − √
9 log n 2
!
0
 bθ(1 + 6θ)  t 2
≥ b µ − 35bµθ /2 − √
3 log n 2
! !
0
 bθ  t
2 t
0 0
≥ b µ − 18bθ − √ =bµ .
3 log n 2 2

4.9 Property 8
We will show that
k
! !
0 k n  X bj µj θ t 
E[ |T | ] ≤ n exp − (1 − ) √ , (55)
t j=0
n 2

which together with Markov’s inequality implies

k
! !
 n  bj µj θ t 
Pr |T 0 | ≥ nk+1
X
exp − (1 − ) √
t j=0
n 2
k
! !
 n  bj µj θ t −1
k+1
E[ |T 0 | ] ≤ 1/n .
X
≤ n exp − (1 − ) √
t j=0
n 2

29
Proof of (55). Since
E[ |T 0 | ] = Pr (E(G0 ) ∩ Γ(T ) = ∅) ,
X

T ∈T

it is enough to show that for each T ∈ T

!
0 (1 − )bµθ t 

Pr (E(G ) ∩ Γ(T ) = ∅) ≤ exp − √ (56)
n 2

(recall that b and µ actually mean bk and µk ). But notice that for

F 0 (T ) := {F ∈ F 0 : F ∩ Γ(T ) 6= ∅}

(again recall that F 0 (= Fk+1 ) is a maximal disjoint collection of forbidden pairs and triples in
X 0 : see (BC 3) in Section 2.1), we clearly have

Pr (E(G0 ) ∩ Γ(T ) = ∅) ≤ Pr (|X 0 ∩ Γ(T )| ≤ 3|F 0 (T )|)

 32 θ 2 θ 
≤ Pr |X 0 ∩ Γ(T )| ≤ √ |Γ(T )| or |F 0 (T )| ≥ √ |Γ(T )|
n n
 2
3 θ   2 θ 
≤ Pr |X 0 ∩ Γ(T )| ≤ √ |Γ(T )| + Pr |F 0 (T )| ≥ √ |Γ(T )| . (57)
n n
Set
ΦT := |X 0 ∩ Γ(T )| = 1(e ∈ X 0 ) .
X

e∈Γ(T )

Then, since all events “e ∈ X 0 ” are mutually independent and Pr (e ∈ X 0 ) = p, we have

Y    θ(1 − exp(ρ)) 
E[exp(ρΦT )] = 1 − p(1 − exp(ρ)) ≤ exp − √ |Γ(T )| .
e∈Γ(T )
n

Markov’s inequality with ρ = −−1 /4 then gives

 32 θ    3ρ2 θ 
Pr ΦT ≤ √ |Γ(T )| = Pr exp(ρΦT ) ≥ exp √ |Γ(T )|
n n
 3ρ2 θ 
≤ exp − √ |Γ(T )| E[exp(ρΦT )]
n
  1 − exp(ρ) 3ρ2  
≤ exp − θ √ +√ |Γ(T )|
n n
!
1  (1 − )bµθ t 
≤ exp − √ .
2 n 2

Therefore
 32 θ   32 θ 
Pr |X 0 ∩ Γ(T )| ≤ √ |Γ(T )| = Pr ΦT ≤ √ |Γ(T )|
n n
!
1  (1 − )bµθ t 
≤ exp − √ . (58)
2 n 2

30
 √
On the other hand, with l := b2 θ|Γ(T )|/ n c, we know that
 2 θ  
Pr |F 0 (T )| ≥ √ |Γ(T )| ≤ Pr ∃ F1 , ..., Fl ∈ Λ ∪ ∆ · 3 ·
n

Fj ⊆ X 0 , Fj ∩ Γ(T ) 6= ∅ and Fj ∩ Fj 0 = ∅ ∀ j 6= j 0 .

Let
X (l) X
:= .
{F1 , ..., Fl } ⊆ Λ ∪ ∆
Fj ∩ Γ(T ) 6= ∅ ∀j ∈ [l]
Fj ∩ Fj 0 = ∅ ∀j 6= j 0

Then
 2 θ  X (l)
Pr |F 0 (T )| ≥ √ |Γ(T )| ≤ Pr (Fj ⊆ X 0 ∀j ∈ [l])
n
l
X (l) Y
= Pr (Fj ⊆ X 0 )
j=1
1  l
Pr (F ⊆ X 0 )
X
≤
l!
F ∈Λ∪∆
F ∩ Γ(T ) 6= ∅

(cf. Lemma 3.4). Properties 4 and 5, and (10) yield

Pr (F ⊆ X 0 ) = Pr (F ⊆ X 0 ) + Pr (F ⊆ X 0 )
X X X

F ∈Λ∪∆ F ∈Λ F ∈∆
F ∩ Γ(T ) 6= ∅ F ∩ Γ(T ) 6= ∅ F ∩ Γ(T ) 6= ∅
√
≤ 2b(a + 5θ) n |Γ(T )| p2 + b2 n|Γ(T )| p3
θ2 θ3
≤ √ |Γ(T )| + √ |Γ(T )|
n n
2θ 2
≤ √ |Γ(T )| .
n

2θ2
Thus with η := √ |Γ(T )|, we have
n
!
 2 θ  ηl 1  θ|Γ(T )|  1  bµθ t 
Pr |F 0 (T )| ≥ √ |Γ(T )| ≤ ≤ exp − √ ≤ exp − √ , (59)
n l! 2 n 2 n 2

ηl  η exp(1) l
where the second inequality uses ≤ . Hence (57), (58) and (59) yield (56).
l! l
Acknowledgments. I thank Joel Spencer and Noga Alon for helpful discussions. Especially,
I appreciate Spencer’s nice explanation of his differential equation which interested me in the
Ramsey number R(3, t).

31
Appendix
To prove Lemma 3.1, set

Ωj := E[Φ|τ1 , ..., τj ] − E[Φ|τ1 , ..., τj−1 ] for j = 1, 2, ...m.

We first verify two more lemmas. The first is from [31].

Lemma A.1 The hypotheses of Lemma 3.1 imply

|Ωj | ≤ cj for j = 1, 2, ..., m .

Proof. Note first for fixed κ = (κ1 , ..., κm ) that

X
E[Φ|τ1 , ..., τj−1 ](κ) = Φ(κ1 , · · · , κj−1 , γj , · · · , γm )Pr (τj = γj , · · · , τm = γm ),
γj ,···,γm

X
and Pr (τj = γj · · · , τm = γm ) = Pr (τj+1 = γj+1 , · · · , τm = γm ) yields
γj

X
E[Φ|τ1 , ..., τj ](κ) = Φ(κ1 , · · · , κj , γj+1 , · · · , γm )Pr (τj+1 = γj+1 , · · · , τm = γm )
γj+1 ,···,γm
X
= Φ(κ1 , · · · , κj , γj+1 , · · · , γn )Pr (τj = γj , · · · , τm = γm ) .
γj ,···,γm

Thus (18) implies

|Ωj (κ)| = |(E[Φ|τ1 , ..., τj ] − E[Φ|τ1 , ..., τj−1 ])(κ)|

X
≤ |Φ(κ1 , · · · , κj , γj+1 , · · · , γm ) − Φ(κ1 , · · · , κj−1 , γj , · · · , γm )|
γj ,···,γm

× Pr (τj = γj , · · · , τm = γm )
X
≤ cj Pr (τj = γj , · · · , τm = γm ) = cj .
γj ,···,γm

Lemma A.2 The hypotheses of Lemma 3.1 imply

E[(Ωj )2 |τ1 , ..., τj−1 ] ≤ p(1 − p)c2j for j = 1, 2, ..., m .

Proof. Jensen’s inequality gives

(Ωj )2 = (E[Φ|τ1 , ..., τj ] − E[Φ|τ1 , ..., τj−1 ])2

= (E[Φ − E[Φ|τ1 , ..., τj−1 , τj+1 , ..., τm ]|τ1 , ..., τj ])2
≤ E[(Φ − E[Φ|τ1 , ..., τj−1 , τj+1 , ..., τm ])2 |τ1 , ..., τj ] .

32
Hence

E[(Ωj )2 |τ1 , ..., τj−1 ]

≤ E[ (Φ − E[Φ|τ1 , ..., τj−1 , τj+1 , ..., τm ])2 |τ1 , ..., τj−1 ]
= E[ E[ (Φ − E[Φ|τ1 , ..., τj−1 , τj+1 , ..., τm ])2 |τ1 , ..., τj−1 , τj+1 , ..., τm ] | τ1 , ..., τj−1 ] .

Thus it is enough to show that

E[ (Φ − E[Φ|τ1 , ..., τj−1 , τj+1 , ..., τm ])2 | τ1 , ..., τj−1 , τj+1 , ..., τm ] ≤ p(1 − p)c2j .

For fixed τ1 , ..., τj−1 , τj+1 , ..., τm let

x := Φ(τ1 , ..., τj−1 , 0, τj+1 , ..., τm ) and y := Φ(τ1 , ..., τj−1 , 1, τj+1 , ..., τm )

Then (18) gives

|x − y| ≤ cj .
Since Pr (τj = 1) = p, this implies

E[ (Φ − E[Φ|τ1 , ..., τj−1 , τj+1 , ..., τm ])2 | τ1 , ..., τj−1 , τj+1 , ..., τm ]
= (1 − p)(x − ((1 − p)x + py))2 + p(y − ((1 − p)x + py))2
= p(1 − p)(x − y)2 ≤ p(1 − p)c2j .

Proof of Lemma 3.1. (cf. Lemma 5.3 of [22]) It is enough to show that
m
Pr (Φ − E[Φ] ≥ λ) ≤ exp(−ρλ + (ρ2 /2)
X
cj exp(ρcj )) ,
j=1

because the same argument gives

m
2
X
Pr (−Φ − E[−Φ] ≥ λ) ≤ exp(−ρλ + (ρ /2) cj exp(ρcj )) .
j=1

We claim first that

E[exp(ρΩj )|τ1 , ..., τj−1 ] ≤ exp((ρ2 /2)p(1 − p)c2j exp(ρcj )) . (60)

For fixed τ1 , ..., τj−1 let

φ(ρ) := E[exp(ρΩj )|τ1 , ..., τj−1 ] .
Then Taylor’s theorem and Lemma A.1 imply for some 0 ≤ ρ∗ ≤ ρ that
φ00 (ρ∗ ) 2
φ(ρ) = φ(0) + φ0 (0)ρ + ρ
2
= 1 + (ρ2 /2)E[(Ωj )2 exp(ρ∗ Ωj )|τ1 , ..., τj−1 ]
≤ 1 + (ρ2 /2)p(1 − p)c2j exp(ρcj ) ≤ exp((ρ2 /2)p(1 − p)c2j exp(ρcj )) ,

33
where the second equality uses E[Ωj ] = 0.
Next we show that
m
E[exp(ρ(Φ − E[Φ]))] ≤ exp((ρ2 /2)p(1 − p) c2j exp(ρcj )) .
X
(61)
j=1
Pm
Since Φ − E[Φ] = j=1 Ωj , it is enough to show that
l l
Ωj )] ≤ exp((ρ2 /2)p(1 − p) c2j exp(ρcj ))
X X
E[exp(ρ
j=1 j=1

by induction on l = 1, 2, ..., m. For l = 1, (60) gives

E[exp(ρΩ1 )] ≤ exp((ρ2 /2)p(1 − p)c21 exp(ρc1 )) .

For l > 1, (60) and the induction hypothesis yield

l
X l
X
E[exp(ρ Ωj )] = E[E[exp(ρ Ωj )|τ1 , ..., τl−1 ]]
j=1 j=1
l−1
X
= E[exp(ρ Ωj )E[exp(ρΩl )|τ1 , ..., τl−1 ]]
j=1
l−1
Ωj )] exp((ρ2 /2)p(1 − p)cl exp(ρcl ))
X
≤ E[exp(ρ
j=1
l
2
c2j exp(ρcj )) .
X
≤ exp((ρ /2)p(1 − p)
j=1

Finally, Markov’s inequality and (61) give

Pr (Φ − E[Φ] ≥ λ) = Pr (exp(ρ(Φ − E[Φ])) > exp(ρλ))

≤ exp(−ρλ)E[exp(ρ(Φ − E[Φ]))]
m
≤ exp(−ρλ + (ρ2 /2)p(1 − p) c2j exp(ρcj )) .
X

j=1

References
[1] M. Ajtai and J. Komlós and E. Szemerédi, A note on Ramsey numbers, J. Combi. Th.
Series A, 29 (1980), 354–360.

[2] M. Ajtai and J. Komlós and E. Szemerédi, A dense infinite Sidon sequence, Europ. J.
Combinatorics 2 (1981), 1-11.

[3] N. Alon and J. Spencer, The probabilistic method, Wiley”, New York, 1992.

34
 [4] K. Azuma, Weighted sums of certain dependent random variables, Tokuku Math. J. 19
(1967), 357–367

[5] B. Bollobás”, Graph Theory, Springer-Verlag, New York, 1979.

[6] B. Bollobás, Random Graphs, Academic Press, London, 1985.

[7] B. Bollobás, Martingales, isoperimetric inequalities and random graphs in Combinatorics

A. Hajnal and L. Lovász and V.T. Sós, Eds. (1998), North Holland.

[8] J. A. Bondy and U. S. R. Murty Graph Theory with Applications, North-Holland, New
York 1976.

[9] L. Breiman, Probability, Addison-Wesley, Reading, Mass., 1968.

[10] P. Erdős Graph theory and probability II, Canad. J. Math 13 (1961), 346-352.

[11] P. Erdős and P. Tetali, Representations of integers as the sum of k terms, Random Struc-
tures and Algorithms 1, (1990), 245–261.

[12] P. Erdős and S. Suen and P. Winkler On the Size of a Random Maximal Graph, Preprint
(1993).

[13] P. Frankl and V. Rödl, Near perfect coverings in graphs and hypergraphs, Europ. J.
Combinatorics 6 (1985), 317–326.

[14] Z. Füredi, Matchings and Covers in Hypergraphs, Graphs and Combinatorics 4 (1988),
115–206.

[15] R. Graham and B. Rothschild and J. Spencer, Ramsey Theory, Wiley, New York, 1990.

[16] J. E. Graver and J. Yackel, Some graph theoretic results associated with Ramsey’s theo-
rem, J. of Combinatorial Theory 4 (1968), 125–175.

[17] W. Hoeffding, Probability inequalities for sums of bounded random variables, J. Amer.
Stat. Assoc. 27 (1963), 13–30.

[18] J. Kahn, Recent results on some not-so-recent hypergraph matching and covering prob-
lems, Proc. 1st Int’l Conference on Extremal Problems for Finite Sets, Visegrád, June
(1991).

[19] J. Kahn, Asymptotically good list colorings, J. of Combinatorial Th. (A), to appear.

[20] J. Kahn and J. H. Kim, Random Matchings in Regular Graphs, In preparation.

[21] J. Kahn and E. Szemerédi, tThe second eigenvalue of a random regular graph, Manuscript,
1988.

35
[22] J. H. Kim, On Brooks’ Theorem For Sparse Graphs, Combinatorics, Probability & Com-
puting, to appear.

[23] J. H. Kim and J. Spencer and P. Tetali, Certificates for Random Tournaments, In prepa-
ration.

[24] J. Komlós and J. Pintz and E. Szemerédi”, On Heilbronn’s triangle problem, J. London
Math. Soc. 25 (1982), 13–24.

[25] M. Krivelevich, Bounding Ramsey numbers through large deviation inequalities, Preprint
(1994).

[26] C. McDiarmid, On the method of bounded differences in Surveys in Combinatorics 1989,

editor J. Siemons (1989) Cambridge Univ. Pr., 148–188.

[27] V. Milman and G. Schechtman, Asymptotic Theory of Finite Dimensional Normed Spaces,
Springer, Berlin, 1980.

[28] N. Pippenger, Unpublished.

[29] N. Pippenger and J. Spencer, Asymptotic behavior of the chromatic index for hyper-
graphs, J. of Combinatorial Th. (A) 51 (1989), 24–42.

[30] V. Rödl, On a packing and covering problem, Europ. J. Combinatorics 5 (1985), 69–78.

[31] E. Shamir and J. Spencer, Sharp concentration of the chromatic number on random
graphs Gn,p , Combinatorica 7 (1980), 121–129.

[32] J. Shearer, A note on the independence number of triangle-free graphs, Discrete Mathe-
matics 46 (1983), 83–87.

[33] J. Shearer, A note on the independence number of triangle-free graphs II, J. of Combina-
torial Th. (B) 53 (1991), 300–307.

[34] J. Spencer, Asymptotic lower bounds for Ramsey functions, Discrete Mathematics 20
(1977), 69–76.

[35] J. Spencer, Ten Lectures on the Probabilistic Method, Society for Industrial and Applied
Mathematics, 1987.

[36] J. Spencer, A Continuous Proof of the Erdős-Hanani Conjecture, Preprint, 1994.

[37] J. Spencer, Maximal Triangle-free Graphs and the Ramsey number R(3, k), Preprint,
1994.

[38] V. Vizing, Some unsolved problems in graph theory (Russian), Usp. Mat. Nauk. 23 (1968),
117–134. English Translation in Russian Math. Surveys, 23, 125–141.

36