May 18 – 22, 2026
Virginia Tech
America/New_York timezone

The Matrix-Vector Complexity of Ax=b

May 22, 2026, 9:35 AM
25m
McBryde Hall 129 (Virginia Tech)

McBryde Hall 129

Virginia Tech

Minisymposium Talk Polynomials, Krylov Methods and Applications Polynomials, Krylov Methods and Applications

Speaker

Raphael Meyer (UC Berkeley)

Description

Solving linear systems Ax=b is a fundamental pillar of NLA. For over 60 years, iterative methods that access A only through matrix-vector products have been the standard approach for solving large linear systems. While lower bounds exist for many special cases, prior work has not shown that methods like GMRES and MINRES achieve an asymptotically optimal matrix-vector complexity for approximately solving Ax=b. In this talk, we prove that MINRES (and CG on the normal equations) achieves an optimal complexity (up to a factor of 4) over all randomized matrix-vector methods, and GMRES achieves the best possible over all "transpose-free" methods.

Authors

Ethan Epperly (UC Berkeley) Prof. Michał Dereziński (University of Michigan) Raphael Meyer (UC Berkeley)

Presentation materials

There are no materials yet.