Matrix probing and its conditioning

TL;DR

This paper investigates the conditions under which matrix probing techniques are well-conditioned, demonstrating their effectiveness in approximating matrices and their inverses, especially for operators with smooth symbols like in seismic imaging.

Contribution

It establishes probabilistic bounds for the conditioning of matrix probing systems based on basis matrix properties and applies these results to seismic imaging and preconditioning.

Findings

Well-conditioned systems require n proportional to p log^2 n.

Matrix probing effectively approximates operators with smooth pseudodifferential symbols.

Numerical results show good preconditioning for elliptic operators in variable media.

Abstract

When a matrix A with n columns is known to be well approximated by a linear combination of basis matrices B_1,..., B_p, we can apply A to a random vector and solve a linear system to recover this linear combination. The same technique can be used to recover an approximation to A^-1. A basic question is whether this linear system is invertible and well-conditioned. In this paper, we show that if the Gram matrix of the B_j's is sufficiently well-conditioned and each B_j has a high numerical rank, then n {proportional} p log^2 n will ensure that the linear system is well-conditioned with high probability. Our main application is probing linear operators with smooth pseudodifferential symbols such as the wave equation Hessian in seismic imaging. We demonstrate numerically that matrix probing can also produce good preconditioners for inverting elliptic operators in variable media.