An improved lower bound on multicolor Ramsey numbers

Yuval Wigderson∗

December 11, 2020

Abstract
A recent breakthrough of Conlon and Ferber yielded an exponential improvement on the
lower bounds for multicolor diagonal Ramsey numbers. In this note, we modify their construc-
tion and obtain improved bounds for more than three colors.

1 Introduction
For positive integers t and ℓ, let r(t; ℓ) denote the ℓ-color Ramsey number of Kt , i.e. the least integer
N such that every ℓ-coloring of E(KN ) contains a monochromatic Kt . The most well-studied case
is that of ℓ = 2, where the bounds
2t/2 ≤ r(t; 2) ≤ 22t
were proved by Erdős [6] and Erdős–Szekeres [7] in 1947 and 1935, respectively. Despite decades
of effort, only lower-order improvements have been made to these bounds [13, 4, 12].
For larger values of ℓ, even less is known. The Erdős–Szekeres [7] argument yields that r(t; ℓ) ≤
ℓℓt . For the lower bound, Erdős’s random construction [6] shows that r(t; ℓ) ≥ ℓt/2 . This was
improved substantially by Lefmann [9], who used an iterated product coloring to show that r(t; ℓ) ≥
2tℓ/4 . Thus, we see that the dependence on the clique size t is exponential, and the dependence
on the number of colors ℓ is somewhere between exponential and super-exponential, i.e. between
2Ω(ℓ) and 2O(ℓ log ℓ) . It is a major open problem to determine the correct ℓ-dependence. Already
for the case t = 3, Erdős offered $100 for the determination of whether r(3; ℓ) is exponential or
super-exponential in ℓ, and this question is closely related to a number of other questions in graph
theory, coding theory, and beyond; see e.g. [1, 11] for more.
In a recent breakthrough, Conlon and Ferber [5] improved Lefmann’s lower bound on r(t; ℓ) for
fixed ℓ > 2 and t → ∞. To do so, they introduced a new construction that mixes algebraic and
probabilistic approaches, and which does better than the random construction for ℓ = 3 and ℓ = 4.
Then, they use Lefmann’s iterated product trick to obtain better bounds for all larger values of ℓ
as well. Their result is that  7ℓ t−o(t)
r(t; ℓ) ≥ 2 24 +C ,

for some constant C that depends only on the residue of ℓ modulo 3. In this note, we use a variant
of the Conlon–Ferber construction to improve the lower bounds on r(t; ℓ) for fixed ℓ and large t.
∗
Department of Mathematics, Stanford University, Stanford, CA 94305, USA. Email: [email protected].
Research supported by NSF GRFP Grant DGE-1656518.

1
Theorem 1. For any fixed ℓ ≥ 2,
 3ℓ 1 t−o(t)
r(t; ℓ) ≥ 2 8 − 4 .

Theorem 1 gives the best known bound for all ℓ ≥ 4, and for large ℓ, improves the constant in
the exponent by roughly a factor of 9/7. It is interesting to note that for our bound, we do not use a
product coloring at all, and instead obtain the bound in Theorem 1 directly from the construction.
The bound also matches the best known exponential constant for ℓ = 2 (due to Erdős [6]) and for
ℓ = 3 (due to Conlon and Ferber [5]). This is because our construction specializes for ℓ = 2, 3 to
these earlier constructions.
At a high level, our construction differs from the Conlon–Ferber construction by replacing their
random induced subgraph by a number of independent random blowups. Such an approach to
proving lower bounds for multicolor Ramsey problems goes back to work of Alon and Rödl [2].
Moreover, it was observed in [8], combining ideas of Alon–Rödl with those of Mubayi–Verstraëte
[10], that for such problems random induced subgraphs and random blowups are closely related,
and are both part of a more general framework of random homomorphisms.

2 Proof of Theorem 1
We begin by recalling the basics of the Conlon–Ferber construction, in the special case of q = 2.
Let t be even and let V ⊂ Ft2 denote the set of vectors of even Hamming weight, so that |V | = 2t−1 .
We define
P a graph G0 with vertex set V by letting {u, v} ∈ E(G0 ) if and only if u · v = 1, where
u · v = ti=1 ui vi denotes the scalar product over F2 .
Lemma 2 (Conlon–Ferber [5]). G0 has no clique of order t.
Proof. This is a simple variant of the Oddtown theorem [3]. Since V consists of vectors of even
Hamming weight, we see that v · v = 0 for all v ∈ V . Therefore, it is simple to show that every
clique in G0 of order t consists of linearly independent vectors, since t is even. Since dim V = t − 1,
this gives the desired result.
5t2
+o(t2 )
Lemma 3 (Conlon–Ferber [5]). G0 has at most 2 8 independent sets of order at most t.
In their paper, Conlon and Ferber only state this bound for the number of independent sets of
size exactly t, but their proof actually yields Lemma 3.
We now fix a non-negative integer m, and define an (m + 2)-coloring χ of E(KN ) for every N .
3mt t
We will eventually take N = 2 8 + 2 −o(t) ; in particular, one should think of N as much larger than
|V | = 2t−1 . We pick m uniformly random functions f1 , . . . , fm : [N ] → V , all independent of one
another. For two distinct vertices x, y ∈ [N ], we define their color χ(x, y) as follows. First, if there
is some index i ∈ [m] such that {fi (x), fi (y)} ∈ E(G0 ), then we let χ(x, y) be the minimum such
index i; note that in particular, χ(x, y) = i implies that fi (x) 6= fi (y). If there is no such i, then
we pick χ(x, y) ∈ {m + 1, m + 2} uniformly at random, with these choices made independently over
all pairs x, y.
In other words, the coloring of KN is obtained by overlaying m random blowups of G0 to N
vertices, and then randomly coloring all the remaining pairs with the two unused colors. We now
claim that for an appropriate choice of N , this coloring will contain no monochromatic cliques of
order t.

2
 3mt t
Theorem 4. For every non-negative integer m, if N = 2 8 + 2 −o(t) , then the coloring χ will contain
3mt t
no monochromatic clique of order t with positive probability. In particular, r(t; m+2) ≥ 2 8 + 2 −o(t) .
Remark. By letting m = ℓ − 2, one obtains the bound in Theorem 1.
Proof. We fix a set S ⊂ [N ] with |S| = t, and will bound the probability that S spans a monochro-
matic clique under χ. First, we observe that S cannot be a monochromatic clique in any of the
first m colors, since blowing up a graph cannot increase its clique number. More formally, if S were
a monochromatic clique in color i ∈ [m], then the set of vertices fi (S) ⊂ V would form a clique in
G0 of order t, which cannot exist by Lemma 2.
Now we bound the probability that S is monochromatic in one of the last two colors. To do
so, we will first compute the probability that no pair in S receives one of the first m colors, i.e.
the probability that all the functions f1 , . . . , fm map S into an independent set of G0 . If T is some
independent set of G0 with |T | ≤ t, then the probability that fi (S) ⊆ T is precisely (|T |/|V |)t ,
since each vertex of S has a |T |/|V | chance of being mapped into T by fi . Therefore,
|T | t
   t
t 2 2
Pr(fi (S) = T ) ≤ ≤ t−1
= 2−t +o(t ) .
|V | 2
2 2
By Lemma 3, the number of choices for such a T is at most 25t /8+o(t ) . Therefore, by the union
2 2
bound, the probability that fi (S) is an independent set in G0 is at most 2−3t /8+o(t ) . Since these
events are independent over all i ∈ [m], we conclude that
3mt2
+o(t2 )
Pr(fi (S) is independent in G0 for all i ∈ [m]) ≤ 2− 8 .
Now, for S to be monochromatic in one of the last two colors, we must first have that fi (S) is
independent in G0 for all i, and then that all the pairs in S receive the same color under the
random assignment of the colors m + 1 and m + 2. In other words,
t
Pr(S is monochromatic) = 21−(2) Pr(fi (S) is independent in G0 for all i ∈ [m])
t2 3mt2
+o(t2 )
≤ 2− 2 − 8 .
Finally, we can apply the union bound over all choices of S, and conclude that
  2
N − t − 3mt2 +o(t2 )
Pr(KN has a monochromatic clique of order t) ≤ 2 2 8
t
 t 3mt
t
≤ N 2− 2 − 8 +o(t)
= o(1),
3mt
+ 2t −o(t)
by our choice of N = 2 8 .

For m = 1, our construction is actually identical to the Conlon–Ferber construction, so of course

3
Acknowledgments. I would like to thank David Conlon and Asaf Ferber for helpful comments,
and Xiaoyu He for introducing me to the method of random homomorphisms. I am also extremely
grateful to Jacob Fox for many insights and for carefully reading an earlier draft of this paper.

References
[1] N. Alon, Lovász, vectors, graphs and codes, in I. Bárány, G. Katona, and A. Sali (eds.),
Building Bridges II, Bolyai Soc. Math. Stud., vol. 28, Springer, 2019.

[2] N. Alon and V. Rödl, Sharp bounds for some multicolor Ramsey numbers, Combinatorica 25
(2005), 125–141.

[3] E. R. Berlekamp, On subsets with intersections of even cardinality, Canad. Math. Bull. 12
(1969), 471–474.

[4] D. Conlon, A new upper bound for diagonal Ramsey numbers, Ann. of Math. (2) 170 (2009),
941–960.

[5] D. Conlon and A. Ferber, Lower bounds for multicolor Ramsey numbers, 2020. Preprint avail-
able at arXiv:2009.10458.

[6] P. Erdös, Some remarks on the theory of graphs, Bull. Amer. Math. Soc. 53 (1947), 292–294.

[7] P. Erdös and G. Szekeres, A combinatorial problem in geometry, Compositio Math. 2 (1935),
463–470.

[8] X. He and Y. Wigderson, Multicolor Ramsey numbers via pseudorandom graphs, Electron. J.
Combin. 27 (2020), Article No. P1.32.

[9] H. Lefmann, A note on Ramsey numbers, Studia Sci. Math. Hungar. 22 (1987), 445–446.

[10] D. Mubayi and J. Verstraëte, A note on pseudorandom Ramsey graphs, 2019. Preprint available
at arXiv:1909.01461.

[11] J. Nešetřil and M. Rosenfeld, I. Schur, C. E. Shannon and Ramsey numbers, a short story,
Discrete Math. 229 (2001), 185–195.

[12] A. Sah, Diagonal Ramsey via effective quasirandomness, 2020. Preprint available at arXiv:
2005.09251.

[13] J. Spencer, Ramsey’s theorem—a new lower bound, J. Combin. Theory Ser. A 18 (1975),
108–115.