Three improved lower bounds for hypergraph Ramsey numbers

Explicit group-invariant colourings improving three entries of §7.1(a) of the dynamic survey Small Ramsey Numbers. Each bound is carried by an object small enough to print, not by a search log.

Hanyu Yang · ORCID 0009-0005-0419-4070 · 10 September 2026 · Repository
R(4,4,4;3) ≥ 84   R(4,6;3) ≥ 64   R(5,5;4) ≥ 36

Under the survey's convention that R is the least n forcing a monochromatic clique, a good colouring on n points proves R ≥ n+1.

cellsurvey rev. #18this workwitness
R(4,4,4;3)≥ 79≥ 8483 points, 3 colours
R(4,6;3)≥ 63≥ 6463 points, 2 colours
R(5,5;4)≥ 35≥ 3635 points, 2 colours

The objects, not the search

Each result is a group-invariant colouring given as a small table over group orbits rather than as solver output. For R(4,4,4;3): let p = 83, let a = 9 generate the order-41 subgroup of Z83*, and let G = {x → aix + b}, of order 3403. G acts on the 91,881 triples of Z83 with exactly 27 orbits, each free of size 3403; 27 numbers give the colours, nine per colour, so the three colour classes are equal at 30,627 each.

That equality is not forced. Regularity forces each class to be a multiple of 3403, and enumerating every invariant good colouring of this class gives 36 of them — twelve splitting 9/9/9 and twenty-four splitting 8/9/10. The other two witnesses use 120 and 107 orbits, under actions that are not free. Every construction regenerates its whole colouring from its table alone, with no solver.

Verification

All twelve deposited witnesses are checked by complete enumeration, each under its own hash-bound run. For the three headline objects that is 1,837,620 four-subsets at n = 83, 324,632 five-subsets at n = 35, and all four- and six-subsets of a 63-set. The R(4,6;3) check runs both ways round: the intended reading gives zero monochromatic cliques and the inverse reading gives 102,627, which is what makes the first number meaningful.

Negative results carry certificates too: an independent RUP/RAT proof checker re-certifies five UNSAT symmetry classes of the n = 63 tower, and its discriminating control weakens a real instance until it becomes satisfiable and requires the checker to refuse the original proof.

An obstruction theorem

If an affine group satisfies |G| = C(p,2) and its multiplier subgroup omits -1 (equivalently p ≡ 3 mod 4), it acts regularly on unordered pairs. Adjoining a fixed point then forces every triple through that point into a single orbit, and hence one colour; no base triple may take that colour, so for the diagonal problem the base must be coloured with m-1 colours while avoiding a monochromatic K4(3) — impossible only once the base reaches or exceeds the (m-1)-colour Ramsey number. The threshold is a hypothesis, not a detail: at m = 3 it is R(4,4;3) = 13 and both towers clear it; at m = 4 it is R(4,4,4;3) itself, and p = 79 does not. Both group conditions are necessary too: for p ≡ 1 (mod 4) the conclusion genuinely fails, and the corresponding classes are satisfiable at p = 13 and p = 37.

What is not claimed

Reproducing

git clone https://github.com/SeverinVisionary/hypergraph-ramsey-bounds
cd hypergraph-ramsey-bounds
pip install -r requirements.txt
bash reproduce.sh --full

The harness re-derives every claim above and refuses to report success for a step that produced no evidence.