Group-theoretic Methods for Bounding the Exponent of Matrix Multiplication

TL;DR

This paper explores group-theoretic techniques to establish bounds on the matrix multiplication exponent, offering a novel approach that could lead to tighter complexity estimates beyond traditional methods.

Contribution

The paper analyzes and proves key results from Cohn and Umans' group-theoretic approach, proposing potential improvements for bounding the matrix multiplication exponent.

Findings

Established bounds on matrix multiplication exponent using group methods

Proposed new avenues for improving exponent estimates

Analyzed the effectiveness of elementary group-theoretic techniques

Abstract

The (asymptotic) complexity of matrix multiplication (over the complex field) is measured by a real parameter w > 0, called the exponent of matrix multiplication (over the complex field), which is defined to be the smallest real number w > 0 such that for an arbitrary degree of precision > 0, two n by n complex matrices can be multiplied using an algorithm using O(n^(w+\epsilon)) number of non-division arithmetical operations. By the standard algorithm for multiplying two matrices, the trivial lower and upper bounds for the exponent w are 2 and 3 respectively. W. Strassen in 1969 obtained the first important result that w < 2.81 using his result that 2 by 2 matrix multiplication could be performed using 7 multiplications, not 8, as in the standard algorithm. In 1984, V. Pan improved this to 2.67, using a variant of Strassen's approach. It has been conjectured that w = 2, but the best known result is that w < 2.38, due to D. Coppersmith and S. Winograd. In all these approaches, estimates for w depend on the number of main running steps in their algorithms. In a recent series of papers in 2003 and 2005, H. Cohn and C. Umans put forward an entirely different approach using fairly elementary methods involving finite groups, group algebras and their representations. The author describes and proves their main results, and suggests possible ways of getting improved estimates for the exponent using their methods.