Addition counts for rank-23 3×33\times 3 matrix multiplication

A rank-23 scheme multiplies two 3×33\times 3 matrices with 23 products of linear forms. How many additions and subtractions does it need at each recursive step when changes into and out of alternative bases can be shared across steps?

Why it matters and what counts as progress

Why it matters: The multiplication count gives only one part of the cost. A low addition count makes a recursive scheme cheaper, particularly when each entry is itself a matrix block.

What counts as progress: Give a correct scheme with fewer additions under the same accounting or prove a lower bound with a precise scope. State whether the scheme works over arbitrary rings and count basis conversion separately.

How the problem developed

  1. 26 Sept 2026 · Earlier 51-addition scheme [3]

    Joshua Stapleton and Andrew Perminov Gave a rank-23 alternative-basis kernel with 51 additions and five more for conversion.

  2. 17 Oct 2026 · Separate ring-valid scheme [1][2]

    Greg Sidebottom with Claude Reported another 51-addition recursive kernel with a Lean proof over arbitrary rings.

AI contribution

51 Additions per Recursive Step for Rank-23 3×33\times 3 Matrix Multiplication, Machine-Checked over Arbitrary Rings

Greg Sidebottom

Sidebottom gives a rank-23 scheme for 3×33\times 3 matrix multiplication over arbitrary rings. It uses 51 additions or subtractions per recursive step in the alternative-basis model, with 11 further additions for conversion in a one-shot computation.

AI involvementAI-centered
ImportanceNotable
Human understandingFew
Paper date:

Sources

  1. 51 Additions per Recursive Step for Rank-23 3×33\times 3 Matrix Multiplication, Machine-Checked over Arbitrary Rings
  2. Exact verification and Lean proof
  3. Stapleton and Perminov’s earlier 51-addition kernel
  4. Nielsen’s 52-addition scheme
Publication history