VibeMathedMath problems solved with AI

The Matrix Multiplication Exponent

The matrix multiplication exponent ω\omega is the infimum of all tt for which two n×nn \times n matrices can be multiplied in O(nt)O(n^t) arithmetic operations. Strassen showed in 1969 that ω<3\omega < 3, and sixty years of work has driven the upper bound down without anyone knowing the true value. Whether ω=2\omega = 2 is one of the central open questions of algebraic complexity.

The current bounds come from the laser method as refined by combination loss analysis. This paper attacks the optimization problem at the core of that refinement, reformulating it so it can be solved in a larger setting, designing a new optimization algorithm for it, and then refining that algorithm with AlphaEvolve.

The result is ω<2.371177\omega < 2.371177, improving the previous best of 2.3713392.371339.

Result
Proved(see note)
Status
Partial result
AI contribution
AI co-developed
Method
Computation
Field
Algebraic complexity
Posed by
Volker Strassen
Year posed
1969
Years open
57y
Solved
2026-08-17
Model
AlphaEvolve
Vendor
Google DeepMind
Collaborators
Emilien Dupont, Marvin Eisenberger, Borislav Kozlovskii, Abbas Mehrabian, Francisco J. R. Ruiz, Abigail See, Renfei Zhou, Josh Alman, Virginia Vassilevska Williams, Matej Balog
Verification
Unreviewed
Publication
Preprint
Significance
55 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

A record, not a resolution, and a small one by design. The bound moves from 2.3713392.371339 to 2.3711772.371177, about 1.6×1041.6 \times 10^{-4}, and the authors describe it as a small step. Whether ω=2\omega = 2 is untouched, and nothing here suggests the laser method can reach it.

The interesting claim is methodological rather than numerical. The bottleneck in this line of work is a hard optimization problem, and the paper reports progress by reformulating that problem and then improving the optimizer, with AlphaEvolve doing the final refinement. That is a different kind of contribution from a new mathematical identity, and it is why the entry is filed as computation.

What the AI did

The paper describes three improvements to the optimization problem at the heart of combination loss analysis, and AlphaEvolve is the third of them. In the authors' own order: they reformulate the problem so it can be solved in a larger setting than was previously possible, they leverage recent advances in machine learning to design a new optimization algorithm for it, and then they "refine the resulting optimization algorithm with AlphaEvolve".

Co-developed rather than discovered. The model improves a component of a human-designed pipeline rather than being handed the problem, and the reformulation that made the larger setting tractable is the authors' own. It is more than tooling, though, because the refined optimizer is what produces the bound.

Verification

A two-day-old arXiv preprint, unrefereed, and nothing was checked here. Bounds of this kind are not the sort of claim a reader can spot-check: the number falls out of a large optimization over laser-method parameters, so reproducing it means re-running the optimization rather than verifying a certificate.

What the author list is worth saying: Josh Alman and Virginia Vassilevska Williams are authors of the prior bounds this improves on, which is unusual and cuts against the main risk with an automated search, namely that it optimizes something subtly different from the quantity everyone means by ω\omega.

Sources

Changelog1 change
  • Rasmus Lindahlapproved this entry

Discussion