Skip to content
ResolvedAI × Math·2025

AlphaEvolve × Strassen

AI broke Strassen's 56-year record for 4×4 matrix multiplication: 48 vs 49 scalar ops.

Formula

Strassen's algorithm
M(2,2,2)=7(vs. 8 classical)M(2,2,2)=7\quad\bigl(\text{vs.}\ 8\text{ classical}\bigr)

Strassen (1969) showed 2×2 matrix multiplication needs only 7 scalar multiplications, launching fast matrix multiplication.

AlphaEvolve discovery
MC(4,4,4)48(vs. Strassen’s 49)M^{\mathbb{C}}(4,4,4)\le 48\quad\bigl(\text{vs. Strassen's }49\bigr)

DeepMind's AlphaEvolve found an algorithm for 4×4 complex matrix multiplication using 48 multiplications, beating 49.

Matrix multiplication exponent
M(n,n,n)=O(nω),2ω<2.372M(n,n,n)=O(n^\omega),\qquad 2\le\omega< 2.372

The exponent ω is the infimum such that n×n matrices can be multiplied in O(nω) operations; ω = 2 is conjectured.

Summary

Matrix multiplication is the most fundamental operation in computing. Strassen (1969) showed 7 multiplications suffice for 2×2 (not 8), giving O(n^2.807). Applied recursively to 4×4, this gives 49. AlphaEvolve found a scheme using only 48 — the first improvement in 56 years.

Sources

Videos