(Approximate) Matrix Multiplication via Convolutions

A longstanding open question in algorithm design is whether "combinatorial" matrix multiplication algorithms -- avoiding Strassen-like divide-and-conquer -- can achieve truly subcubic runtime $n^{3-\delta}$. We present an $O(n^{2.89})$-time exact algorithm, which only sums convolutions in $\mathbb{Z}_m^k$ (multivariate polynomial multiplications) via FFT, building on the work of Cohn, Kleinberg, Szegedy and Umans (CKSU'05). While the algorithm avoids recursion, the asymptotic speedup arises only for impractically large matrices. Motivated by practical applications, we use this baseline to develop a new framework for fast approximate matrix multiplication (AMM), via low-degree approximations of the CKSU polynomials. We show that combining the aforementioned algorithm with black-box linear sketching already breaks the longstanding linear speed-accuracy tradeoff for AMM (Sarlos'06, Clarkson-Woodruff'13 ,Pagh'11, Cohn-Lewis'00), achieving $\frac{1}{r^{1.1}}\|\mathbf{A}\|_F^2\|\mathbf{B}\|_F^2$ error in $O(rn^2)$-time. Our main result is a low-degree approximation scheme for the CKSU polynomials, based on a Fourier-concentration lemma, yielding substantially smaller error in the distributional setting where $\mathbf{A},\mathbf{B}$ come from an i.i.d product-distribution; For random Gaussian matrices, this practical AMM algorithm attains smaller error than the best rank-$r$ SVD of the output matrix $\mathbf{A}\mathbf{B}$, in time $O(rn^2)$. This is a substantial improvement over iterative Krylov subspace methods for low-rank approximation. Our theoretical and empirical results suggest the possibility of replacing MatMuls with sums of convolutions in LLM training and inference.