Multiplying two 4×4 matrices the textbook way takes 64 multiplications. Fast algorithms reduce this — Strassen's method, recursively applied, achieves sub-cubic scaling by trading multiplications for additions. But every multiplication you eliminate increases the number of additions, and additions accumulate rounding error. Fast matrix multiplication is fast but inaccurate.
Cenk, Hasan, and Kaltofen (arXiv:2603.18699) present a variant of 4×4 matrix multiplication using 48 multiplications — the same count as existing fast algorithms — but with improved numerical stability. The error bound exponent drops from approximately 2.577 to 2.386 in the maximum norm. Same speed, less error.
The improvement comes not from a new algorithmic structure but from choosing better coefficients within the existing framework. The 48-multiplication recipe is a system of bilinear forms over a ring containing an inverse of 2. Different coefficient choices yield the same algebraic result but different error propagation characteristics. The authors search this design space for variants that minimize accumulated rounding error while preserving the multiplication count.
This matters because fast matrix multiplication has been theoretically optimal for decades but practically abandoned for numerical work due to error accumulation. The standard wisdom: use Strassen-type methods only when precision doesn't matter, or when matrices are large enough that the asymptotic speedup overwhelms the accuracy cost. Improving accuracy within the same complexity class narrows the gap between theory and practice.
The shortcut was always there. What changes is how carefully you walk it. The number of steps stays the same; the stumbling decreases.