The k-Clique hypothesis is dead, for good

Virginia Vassilevska Williams, Or Zamir

For all sufficiently large fixed kk, Vassilevska Williams and Zamir report a word-RAM algorithm that counts kk-cliques in an nn-vertex graph in time O(n(ω−δ)k/3)O(n^{(\omega-\delta)k/3}). Here ω\omega is the square matrix-multiplication exponent and δ>0\delta>0 is independent of kk and ω\omega.

What was already known

Nešetřil and Poljak reduced clique detection to triangle detection using matrix multiplication. For k divisible by three this gives the standard exponent ωk/3\omega k/3, up to arbitrarily small overhead in the exponent.

AI involvement AI-centered

The authors credit ChatGPT-6 with an initial five-clique improvement and with the algebraic decomposition distilled from OpenAI’s earlier ETH proof. They developed most of the framework that applies the decomposition to nine-vertex patterns and then to larger cliques.

Importance Major

The algorithm beats the predicted exponent for large fixed clique sizes even if the matrix-multiplication exponent is two.

Technical verification Reported

Reported is the default when evidence for a higher level is insufficient. The paper gives a proof in the stated computation model. The reviewed sources provide no documented completed independent check or machine verification of the full running-time claim.

Human understanding Explained

Vassilevska Williams and Zamir explain the decomposition and how it lets their algorithm replace dense nine-vertex patterns by sparser ones. They give a self-contained proof of the decomposition and derive the clique-counting bound.

Correction and publication history