Information Complexity of Inner Product and Matrix Multiplication via Shannon-to-Min-Entropy Conversion

Abstract

We determine the information complexity of matrix multiplication over finite fields under independent uniform inputs, up to a universal additive constant. The binary inner-product case resolves the open question of Braverman, Garg, Pankratov, and Weinstein on the optimal dependence of information complexity on the error probability. For inputs in $\mathbb F_2^m$ and error at most $\varepsilon\in[0,1/2]$, the information complexity with output at Bob lies between $(1-2\varepsilon)m-c_0$ and $(1-2\varepsilon)m$, where $c_0<4.772$ is a universal constant. For independent uniform matrices $X\in\mathbb F_q^{t\times m}$ and $Y\in\mathbb F_q^{s\times m}$, where $t\le s$, computing $XY^{\mathsf T}$ with output at Bob and error at most $\varepsilon\in[0,1]$ has information complexity $(1-q^{ts}\varepsilon/(q^{ts}-1))^+mt\log_2q$, up to an additive gap of less than $9.544$ bits. This gap is independent of the matrix dimensions, field size, and error. At zero error, the information complexity is exactly $mt\log_2q$. The information lower bounds for both inner product and matrix multiplication follow from a new Shannon-to-min-entropy theorem for convolution on discrete groups. Let $Q_1,\ldots,Q_k$ be the finitely supported distributions of independent random variables taking values in a discrete group, and let $s$ be the largest support size. We show that if $k\ge\lceil\lfloor\log_2s\rfloor\ln2\rceil$, then the min-entropy of their ordered product is at least the average Shannon entropy of the $Q_i$ minus $c_0$ bits. This holds for arbitrary discrete groups, including infinite and noncommutative groups, even when the input distributions differ. The leading coefficient of the logarithmic threshold is optimal.