qedbot

Algorithms·matmul:omega

Exponent of matrix multiplication

record target minimiserecord held by a machine F0 no formal statement

No formal proof attached. Any claim here rests on a write-up or a report.

Fidelity F0: Absent. No formal statement is attached to the result.

The smallest ω such that two n×n matrices can be multiplied in about n^ω operations. The schoolbook method gives 3, and ω is at least 2; the record is the best proven upper bound.

Source

Record history

15 steps·exponent (upper bound), lower is better
2.42.62.81980200020202.8074, Volker Strassen, 19692.796, Victor Pan, 19782.7799, Dario Bini, Milvio Capovani, Grazia Lotti and Francesco Romani, 19792.522, Arnold Schönhage, 19812.517, Francesco Romani, 19822.496, Don Coppersmith and Shmuel Winograd, 19822.479, Volker Strassen, 19862.3755, Don Coppersmith and Shmuel Winograd, 19902.3729, Virginia Vassilevska Williams, 20122.3728639, François Le Gall, 2014-012.3728596, Josh Alman and Virginia Vassilevska Williams, 2020-102.371866, Ran Duan, Hongxun Wu and Renfei Zhou, 2022-112.371552, Virginia Vassilevska Williams, Yinzhan Xu, Zixuan Xu and Renfei Zhou, 2023-072.371339, Josh Alman, Ran Duan, Virginia Vassilevska Williams, Yinzhan Xu, Zixuan Xu and Renfei Zhou, 2024-042.371177, Emilien Dupont and nine co-authors, with AlphaEvolve, 2026-08-17
Ink marks are steps by people and clay marks steps where AI took part; grey marks did not move the record, and hollow marks are candidates. How frontiers are drawn

Strassen's 1969 algorithm first brought the exponent below 3. By 1990 the methods of Coppersmith, Winograd and Strassen had taken it to 2.3755, and refinements since 2012 have moved it in the third and fourth decimal places. The 2026 step came from a new optimisation algorithm for the combination-loss analysis, refined with AlphaEvolve.

2.371177

Emilien Dupont and nine co-authors, with AlphaEvolve·17 Aug 2026·Source

A reformulated optimisation problem and a new optimisation algorithm, refined with AlphaEvolve.

AI took partbest known
2.371339

Josh Alman, Ran Duan, Virginia Vassilevska Williams, Yinzhan Xu, Zixuan Xu and Renfei Zhou·Apr 2024·Source

More asymmetry yields faster matrix multiplication.

moved the record
2.371552

Virginia Vassilevska Williams, Yinzhan Xu, Zixuan Xu and Renfei Zhou·Jul 2023·Source

From alpha to omega.

moved the record
2.371866

Ran Duan, Hongxun Wu and Renfei Zhou·Nov 2022·Source

Asymmetric hashing.

moved the record
2.3728596

Josh Alman and Virginia Vassilevska Williams·Oct 2020·Source

A refined laser method.

moved the record
2.3728639

François Le Gall·Jan 2014·Source

Powers of tensors and fast matrix multiplication.

moved the record
2.3729

Virginia Vassilevska Williams·2012·Source

Multiplying matrices faster than Coppersmith–Winograd.

moved the record
2.3755

Don Coppersmith and Shmuel Winograd·1990·Source

Matrix multiplication via arithmetic progressions; the record for two decades.

moved the record
2.479

Volker Strassen·1986·Source

The laser method.

moved the record
2.496

Don Coppersmith and Shmuel Winograd·1982·Source

On the asymptotic complexity of matrix multiplication.

moved the record
2.517

Francesco Romani·1982·Source

Disjoint sums of tensors.

moved the record
2.522

Arnold Schönhage·1981·Source

Partial and total matrix multiplication, and the asymptotic sum inequality.

moved the record
2.7799

Dario Bini, Milvio Capovani, Grazia Lotti and Francesco Romani·1979·Source

Approximate algorithms and border rank.

moved the record
2.796

Victor Pan·1978·Source

Trilinear aggregating.

moved the record
2.8074

Volker Strassen·1969·Source

Gaussian elimination is not optimal: seven multiplications for 2×2, applied recursively.

moved the record

AI activity

How grades work
AlphaEvolve

17 Aug 2026·with Emilien Dupont, Marvin Eisenberger, Borislav Kozlovskii, Abbas Mehrabian, Francisco J. R. Ruiz, Abigail See, Renfei Zhou, Josh Alman, Virginia Vassilevska Williams, Matej Balog

An upper bound of 2.371177 on the exponent of matrix multiplication, improving 2.371339.

record A1 V1 F0
Reasoning and sources

Autonomy

The authors reformulated the optimisation problem and designed a new algorithm for it; AlphaEvolve played a supporting role, refining that algorithm, as the abstract says.

F0 no formal statement. Absent. No formal statement is attached to the result.

Follow and discuss

All discussion

Discussion and bounties for this problem load here.

Something wrong or missing here? Request a correction or add a claim, with its sources.

Sources

Cite this record

qed.bot, “Exponent of matrix multiplication”, https://qed.bot/t/matmul-omega, as of 30 Sep 2026.

This record as plain text, with each claim, its grades and its sources.