A faster matrix multiplication algorithm through structured optimization

Abstract

We present a new asymptotic matrix multiplication construction yielding $ω<2.37115924$, and therefore an algorithm that multiplies two $n\times n$ matrices in $O(n^{2.37115924})$ arithmetic operations over the rationals, reals, or complex numbers. The construction improves the bound $2.371177$ of Dupont et al. within the combination-loss framework for the Coppersmith--Winograd tensor. Our approach combines a shared recursive parameter family with a conditioned optimization procedure that addresses two obstacles: weak gradients at low-mass constituents and competing bottlenecks in the retained exponent. A mass- and probability-dependent rescaling supports quasi-Newton optimization in $212\,820$ coordinates, while a six-control correction balances the three extraction roles at each relevant stage. Positive rational entropy-dual factors and an independent ordinary-string verifier turn the resulting construction into an exact certificate. The verified upper-bound expression lies in $[2.3711592385931075066123,\,2.3711592385931075066124]$. The complete witness and an offline verifier using only the Python standard library accompany the paper.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.