Fast Interpolation-based Globality Certificates for Computing Kreiss Constants and the Distance to Uncontrollability

TL;DR

This paper introduces a fast, interpolation-based method for accurately computing the Kreiss constant and the distance to uncontrollability, significantly improving efficiency and reliability over previous eigenvalue-based techniques.

Contribution

The authors develop a novel interpolation-based globality certificate approach that reduces computational complexity and enhances numerical stability for these matrix analysis problems.

Findings

Achieves $ ext{O}(kn^3)$ work complexity, much faster than previous methods.

Uses $ ext{O}(n^2)$ memory, enabling large-scale computations.

Demonstrates orders of magnitude speedup over state-of-the-art algorithms.

Abstract

We propose a new approach to computing global minimizers of singular value functions in two real variables. Specifically, we present new algorithms to compute the Kreiss constant of a matrix and the distance to uncontrollability of a linear control system, both to arbitrary accuracy. Previous state-of-the-art methods for these two quantities rely on 2D level-set tests that are based on solving large eigenvalue problems. Consequently, these methods are costly, i.e., $\mathcal{O}(n^6)$ work using dense eigensolvers, and often multiple tests are needed before convergence. Divide-and-conquer techniques have been proposed that reduce the work complexity to $\mathcal{O}(n^4)$ on average and $\mathcal{O}(n^5)$ in the worst case, but these variants are nevertheless still very expensive and can be numerically unreliable. In contrast, our new interpolation-based globality certificates perform level-set tests by building interpolant approximations to certain one-variable continuous functions that are both relatively cheap and numerically robust to evaluate. Our new approach has a $\mathcal{O}(kn^3)$ work complexity and uses $\mathcal{O}(n^2)$ memory, where $k$ is the number of function evaluations necessary to build the interpolants. Not only is this interpolation process mostly "embarrassingly parallel," but also low-fidelity approximations typically suffice for all but the final interpolant, which must be built to high accuracy. Even without taking advantage of the aforementioned parallelism, $k$ is sufficiently small that our new approach is generally orders of magnitude faster than the previous state-of-the-art.