A geometric approach to approximating the limit set of eigenvalues for banded Toeplitz matrices

TL;DR

This paper introduces a geometric method to approximate the limit set of eigenvalues for banded Toeplitz matrices, using polygonal approximations and spectral analysis, with an implementation that outperforms some existing algorithms.

Contribution

The paper presents a novel geometric approach for approximating the eigenvalue limit set of banded Toeplitz matrices, including explicit bounds and an efficient algorithm.

Findings

Polygonal approximations converge to the limit set in Hausdorff metric.

The algorithm performs comparably or better than existing methods.

Approximation error decreases at a rate of O(1/√k).

Abstract

This article is about finding the limit set for banded Toeplitz matrices. Our main result is a new approach to approximate the limit set $\Lambda(b)$ where $b$ is the symbol of the banded Toeplitz matrix. The new approach is geometrical and based on the formula $\Lambda(b) = \cap_{\rho \in (0, \infty)} \text{sp } T(b_\rho)$, where $\rho$ is a scaling factor, i.e. $b_\rho(t) := b(\rho t)$, and $\text{sp }(\cdot)$ denotes the spectrum. We show that the full intersection can be approximated by the intersection for a finite number of $\rho$'s, and that the intersection of polygon approximations for $\text{sp } T(b_\rho)$ yields an approximating polygon for $\Lambda(b)$ that converges to $\Lambda(b)$ in the Hausdorff metric. Further, we show that one can slightly expand the polygon approximations for $\text{sp } T(b_\rho)$ to ensure that they contain $\text{sp } T(b_\rho)$. Then, taking the intersection yields an approximating superset of $\Lambda(b)$ which converges to $\Lambda(b)$ in the Hausdorff metric, and is guaranteed to contain $\Lambda(b)$. Combining the established algebraic (root-finding) method with our approximating superset, we are able to give an explicit bound on the Hausdorff distance to the true limit set. We implement the algorithm in Python and test it. It performs on par to and better in some cases than existing algorithms. We argue, but do not prove, that the average time complexity of the algorithm is $O(n^2 + mn\log m)$, where $n$ is the number of $\rho$'s and $m$ is the number of vertices for the polygons approximating $\text{sp } T(b_\rho)$. Further, we argue that the distance from $\Lambda(b)$ to both the approximating polygon and the approximating superset decreases as $O(1/\sqrt{k})$ for most of $\Lambda(b)$, where $k$ is the number of elementary operations required by the algorithm.