Lower bounds for multicolor Ramsey numbers

David Conlon∗ Asaf Ferber†

arXiv:2009.10458v2 [math.CO] 24 Sep 2020

Abstract
We give an exponential improvement to the lower bound on diagonal Ramsey numbers for any
fixed number of colors greater than two.

1 Introduction
The Ramsey number r(t; ℓ) is the smallest natural number n such that every ℓ-coloring of the
edges of the complete graph Kn contains a monochromatic Kt . For ℓ = 2, the problem of determining
r(t) := r(t; 2) is arguably one of the most famous in combinatorics. The bounds
√ t
2 < r(t) < 4t
have been known since the 1940s, but, despite considerable interest, only lower order improvements
√ t
[2, 6, 7] have been made to either bound. In particular, the lower bound r(t) > (1 + o(1)) √t2e 2 ,
proved by Erdős [3] as one of the earliest applications of the probabilistic method, has only been
improved [7] by a factor of 2 in the intervening 70 years.
If we ignore lower order terms, the best known upper bound for ℓ ≥ 3 is r(t; ℓ) < ℓℓt , proved
through a simple modification of the Erdős–Szekeres neighborhood-chasing argument [4] that yields
√ t
r(t) < 4t . For ℓ = 3, the best lower bound, r(t; 3) > 3 , again comes from the probabilistic method.
For higher ℓ, the best lower bounds come from the simple observation of Lefmann [5] that
r(t; ℓ1 + ℓ2 ) − 1 ≥ (r(t; ℓ1 ) − 1)(r(t; ℓ2 ) − 1).
To see this, we blow-up an ℓ1 -coloring of Kr(t;ℓ1 )−1 with no monochromatic Kt so that each vertex set
has order r(t; ℓ2 ) − 1 and then color each of these copies of Kr(t;ℓ2 )−1 separately with the remaining
ℓ2 colors so that there is again no monochromatic Kt . By using the bounds r(t; 2) − 1 ≥ 2t/2 and
r(t; 3) − 1 ≥ 3t/2 , we can repeatedly apply this observation to conclude that
r(t; 3k) > 3kt/2 , r(t; 3k + 1) > 2t 3(k−1)t/2 , r(t; 3k + 2) > 2t/2 3kt/2 .
Our main result is an exponential improvement to all these lower bounds for three or more colors.
Our principal contribution is the following theorem, proved via a construction which is partly
deterministic and partly random. The deterministic part shares some characteristics with a con-
struction of Alon and Krivelevich [1], in that we consider a graph whose vertices are vectors over
a finite field where adjacency is determined by the value of their scalar product, while randomness
comes in through both random coloring and random sampling.
∗
Department of Mathematics, California Institute of Technology, CA 91125, USA. Email: [email protected].
†
Department of Mathematics, University of California, Irvine, CA 92697, USA. Email: [email protected]. Research
supported in part by NSF grants DMS-1954395 and DMS-1953799.

1
Theorem 1. For any prime q, r(t; q + 1) > 2t/2 q 3t/8+o(t) .

In particular, the cases q = 2 and q = 3 yield exponential improvements over the previous
bounds for r(t; 3) and r(t; 4), both of which came from the probabilistic method (in fact, Lefmann’s
observation gives an additional polynomial factor in the four-color case, but this is of lower order
than the exponential improvements that are our concern).

Corollary 2. r(t; 3) > 27t/8+o(t) and r(t; 4) > 2t/2 33t/8+o(t) .

For the sake of comparison, we note that the improvement for three colors is from 1.732t to 1.834t ,
while, for four colors, it is from 2t to 2.135t . Improvements for all ℓ ≥ 5 now follow from repeated
applications of Lefmann’s observation, yielding

r(t; 3k) > 27kt/8+o(t) , r(t; 3k + 1) > 27(k−1)t/8+t/2 33t/8+o(t) , r(t; 3k + 2) > 27kt/8+t/2+o(t) ,

where we used, for instance,

r(t; 3k + 1) − 1 ≥ (r(t; 3(k − 1)) − 1)(r(t; 4) − 1) ≥ (r(t; 3) − 1)k−1 (r(t; 4) − 1).

2 Proof of Theorem 1
Let q be a prime. Suppose t 6= 0 mod q and let V ⊆ Ftq be the set consisting of all vectors v ∈ Ftq for
which ti=1 vi2 = 0 mod q, noting that q t−2 ≤ |V | ≤ q t . Here the lower bound follows from observing
P

that we may pick v1 , . . . , vt−2 arbitrarily and, since every element in Fq can be written as the sum of
2 + v2 = −
Pt−2 2
two squares, there must then exist at least one choice of vt−1 and vt such that vt−1 t i=1 vi .
V

We will first color all the pairs 2 and then define a coloring of E(Kn ) by restricting our attention
to a random sample of n vertices in V . Formally:

V V



Coloring all pairs in 2 . For every pair uv ∈ 2 , we define its color χ(uv) according to the
following rules:

• If u · v = i mod q and i 6= 0, then set χ(uv) = i.

• Otherwise, choose χ(uv) ∈ {q, q + 1} uniformly at random, independently of all other pairs.

Mapping [n] into V . Take a random injective map f : [n] → V and define the color of every edge
ij as χ(f (i)f (j)).

Our goal is to upper bound the orders of the cliques in each color class.

Colors 1 ≤ i ≤ q − 1. Note that one cannot find an i-monochromatic clique of order larger than t
for any 1 ≤ i ≤ q − 1. Indeed, suppose that v1 , . . . , vs form an i-monochromatic clique. We wish to
show that they must be linearly independent and, therefore, that there are at most t of them. To
this end, suppose that
X s
u := αj vj = 0̄
j=1

2
and we wish to show that αj = 0 mod q for all j. Observe that since vj · vj = 0 mod q for all j (our
ground set V consists only of such vectors) and vk · vj = i mod q for each k 6= j, by considering all
the products u · vj , we obtain that the vector ᾱ = (α1 , . . . , αs ) is a solution to

M ᾱ = 0̄

with M = iJ − iI, where J is the s × s all 1 matrix and I is the s × s identity matrix. In particular,
we obtain that the eigenvalues of M (over Z) are is − i with multiplicity 1 and −i with multiplicity
s − 1. Therefore, if s 6= 1 mod q, the matrix is also non-singular over Zq , implying that ᾱ = 0, as
required. On the other hand, if s = 1 mod q, we can apply the same argument with v1 , . . . , vs−1 to
conclude that s − 1 ≤ t. But, we cannot have s − 1 = t, since this would imply that t = 0 mod q,
contradicting our assumption. Therefore, we may also conclude that s ≤ t in this case.

Colors q and q + 1. We call a subset X ⊆ V a potential clique if |X| = t and u · v = 0 mod q

for all u, v ∈ X. Given a potential clique X, we let MX be the t × t matrix whose rows consist
of all the vectors in X. Observe that MX · MX T = 0, where we use the fact that each vector is

self-orthogonal. First we wish to count the number of potential cliques and later we will calculate
the expected number of cliques that survive after we color randomly and restrict to a random subset
of order n.
Suppose that X is a potential clique and let r := rank(X) be the rank of the vectors in this clique,
noting that r ≤ t/2, since the dimension of any isotropic subspace of Ftq is at most t/2. By assuming
that the first r elements are linearly independent, the number of ways to build a potential clique X
of rank r is upper bounded by
r−1
!
3r 2
· q (t−r)r = q tr−(2)+tr−r = q 2tr− 2 + 2 .
Y r 2 r
t−i
q
i=0

This function is increasing up to r = 2t 1

3 + 6 , so the maximum occurs at t/2. By plugging this into the
estimate above and summing over all possible ranks, we see that the number Nt of potential cliques
5t2 2
in V is upper bounded by q 8 +o(t ) .
The probability that a potential clique becomes monochromatic after the random coloring is
21−(2) . Suppose now that p is such that p|V | = 2n and observe that p = nq −t+O(1) . If we choose a
t

random subset of V by picking each v ∈ V independently with probability p, the expected number
of monochromatic potential cliques in this subset is, for n = 2t/2 q 3t/8+o(t) ,
t2 5t2
 t 3t t
pt 21−(2) Nt ≤ q −t +o(t ) nt 2− 2 +o(t ) q 8 +o(t ) = 2− 2 q − 8 +o(t) n < 1/2.
t 2 2 2 2

Since our random subset will also contain more than n elements with probability at least 1/2, there
exists a choice of coloring and a choice of subset of order n such that there is no monochromatic
potential clique in this subset. This completes the proof.
√ t
Remark. Our method also gives a new construction which matches Erdős’ bound r(t) > 2 up to
lower order terms. To see this, we set V = F2t 2 and color edges red or blue depending on whether
u · v = 0 or 1 mod q. If we then sample 2t/2+o(t) vertices of V at random, we can show that w.h.p. the
resulting set does not contain a monochromatic clique of order t. It was pointed out to us by Jacob
Fox that one can achieve the same end by starting with any pseudorandom graph on n vertices for

3
which the count of cliques and independent sets of order 2c log 2 n is asymptotically the same as in
G(n, 1/2) and sampling nc vertices. This can be applied, for instance, with the Paley graph.

Acknowledgements. We are extremely grateful to Vishesh Jain and Wojciech Samotij for carefully
reading an earlier draft of this paper and offering several suggestions which improved the presentation.
We also owe a debt to Noga Alon and Anurag Bishnoi, both of whom pointed out the constraint on
the dimension of isotropic subspaces, thereby improving the bound in our original draft.

References
[1] N. Alon and M. Krivelevich, Constructive bounds for a Ramsey-type problem, Graphs Combin.
13 (1997), 217–225.

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

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

[4] P. Erdős and G. Szekeres, A combinatorial problem in geometry, Compos. Math. 2 (1935), 463–
470.

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

[6] A. Sah, Diagonal Ramsey via effective quasirandomness, preprint available at arXiv:2005.09251
[math.CO].

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