Infinite GMRES for parameterized linear systems

TL;DR

This paper introduces a novel approach combining GMRES and nonlinear eigenvalue problem algorithms to efficiently solve large, parameter-dependent linear systems by constructing a cheap-to-evaluate approximate solution across many parameters.

Contribution

It develops a new method that integrates Krylov subspace techniques with NEP algorithms, exploiting low-rank structures for efficient parameterized system solutions.

Findings

Convergence bounds similar to GMRES for linear systems.

Method is competitive for large-scale problems.

Numerical experiments demonstrate effectiveness.

Abstract

We consider linear parameter-dependent systems $A(\mu) x(\mu) = b$ for many different $\mu$, where $A$ is large and sparse, and depends nonlinearly on $\mu$. Solving such systems individually for each $\mu$ would require great computational effort. In this work we propose to compute a partial parameterization $\tilde{x} \approx x(\mu)$ where $\tilde{x}(\mu)$ is cheap to compute for many different $\mu$. Our methods are based on the observation that a companion linearization can be formed where the dependence on $\mu$ is only linear. In particular, we develop methods which combine the well-established Krylov subspace method for linear systems, GMRES, with algorithms for nonlinear eigenvalue problems (NEPs) to generate a basis for the Krylov subspace. Within this new approach, the basis matrix is constructed in three different ways, using a tensor structure and exploiting that certain problems have low-rank properties. We show convergence factor bounds obtained similarly to those for the method GMRES for linear systems. More specifically, a bound is obtained based on the magnitude of the parameter $\mu$ and the spectrum of the linear companion matrix, which corresponds to the reciprocal solutions to the corresponding NEP. Numerical experiments illustrate the competitiveness of our methods for large-scale problems. The simulations are reproducible and publicly available online.