Fast Computation of Sep$_\lambda$ via Interpolation-based Globality Certificates

TL;DR

This paper introduces a fast, parallelizable interpolation-based algorithm for computing the spectral separation measure between matrices, significantly improving efficiency over previous methods and enabling practical solutions for moderate-sized problems.

Contribution

The paper presents a novel, efficient algorithm for computing sep-lambda, overcoming the computational limitations of earlier methods and including the first algorithm for a related spectral separation measure.

Findings

Algorithm is significantly faster than previous methods.

Supports parallel computation on multi-core systems.

Effective for moderate-sized matrix problems.

Abstract

Given two square matrices $A$ and $B$, we propose a new approach for computing the smallest value $\varepsilon \geq 0$ such that $A+E$ and $A+F$ share an eigenvalue, where $\|E\|=\|F\|=\varepsilon$. In 2006, Gu and Overton proposed the first algorithm for computing this quantity, called $\mathrm{sep}_\lambda(A,B)$ ("sep-lambda"), using ideas inspired from an earlier algorithm of Gu for computing the distance to uncontrollability. However, the algorithm of Gu and Overton is extremely expensive, which limits it to the tiniest of problems, and until now, no other algorithms have been known. Our new algorithm can be orders of magnitude faster and can solve problems where $A$ and $B$ are of moderate size. Moreover, our method consists of many "embarrassingly parallel" computations, and so it can be further accelerated on multi-core hardware. Finally, we also propose the first algorithm to compute an earlier version of sep-lambda where $\|E\| + \|F\|=\varepsilon$.