Structural backward stability in rational eigenvalue problems solved via block Kronecker linearizations

TL;DR

This paper investigates the backward stability of eigenstructure computations for rational matrices using block Kronecker linearizations, establishing bounds and a scaling method to ensure stability and structure preservation.

Contribution

It introduces a structured backward stability analysis for eigenvalue problems of rational matrices via block Kronecker linearizations, including bounds and a scaling technique.

Findings

Backward stable eigenstructure solvers lead to a perturbed rational matrix close to the original.

A method to restore structure via strict equivalence preserves the rational form.

Scaling can make bounds on perturbations satisfactorily small.

Abstract

We study the backward stability of running a backward stable eigenstructure solver on a pencil $S(\lambda)$ that is a strong linearization of a rational matrix $R(\lambda)$ expressed in the form $R(\lambda)=D(\lambda)+ C(\lambda I_\ell-A)^{-1}B$, where $D(\lambda)$ is a polynomial matrix and $C(\lambda I_\ell-A)^{-1}B$ is a minimal state-space realization. We consider the family of block Kronecker linearizations of $R(\lambda)$, which are highly structured pencils. Backward stable eigenstructure solvers applied to $S(\lambda)$ will compute the exact eigenstructure of a perturbed pencil $\widehat S(\lambda):=S(\lambda)+\Delta_S(\lambda)$ and the special structure of $S(\lambda)$ will be lost. In order to link this perturbed pencil with a nearby rational matrix, we construct a strictly equivalent pencil $\widetilde S(\lambda)$ to $\widehat S(\lambda)$ that restores the original structure, and hence is a block Kronecker linearization of a perturbed rational matrix $\widetilde R(\lambda) = \widetilde D(\lambda)+ \widetilde C(\lambda I_\ell- \widetilde A)^{-1} \widetilde B$, where $\widetilde D(\lambda)$ is a polynomial matrix with the same degree as $D(\lambda)$. Moreover, we bound appropriate norms of $\widetilde D(\lambda)- D(\lambda)$, $\widetilde C - C$, $\widetilde A - A$ and $\widetilde B - B$ in terms of an appropriate norm of $\Delta_S(\lambda)$. These bounds may be inadmissibly large, but we also introduce a scaling that allows us to make them satisfactorily tiny. Thus, for this scaled representation, we prove that the staircase and the $QZ$ algorithms compute the exact eigenstructure of a rational matrix $\widetilde R(\lambda)$ that can be expressed in exactly the same form as $R(\lambda)$ with the parameters defining the representation very near to those of $R(\lambda)$. This shows that this approach is backward stable in a structured sense.