Stability Analysis of Inexact Solves in Moment Matching based Model Reduction

Abstract: Recently a new algorithm for model reduction of second order linear dynamical systems with proportional damping, the Adaptive Iterative Rational Global Arnoldi (AIRGA) algorithm (Bonin et. al., 2016), has been proposed. The main computational cost of the AIRGA algorithm is solving a sequence of linear systems. Usually, direct methods (e.g., LU) are used for solving these systems. As model sizes grow, direct methods become prohibitively expensive. Iterative methods (e.g., Krylov) scale well with size, and hence, are a good choice with an appropriate preconditioner. Preconditioned iterative methods introduce errors in linear solves because they are not exact. They solve linear systems up to a certain tolerance. We prove that, under mild conditions, the AIRGA algorithm is backward stable with respect to the errors introduced by these inexact linear solves. Our first assumption is use of a Ritz-Galerkin based solver that satisfies few extra orthogonality conditions. Since Conjugate Gradient (CG) is the most popular method based upon the Ritz-Galerkin theory, we use it. We show how to modify CG to achieve these extra orthogonalities. Modifying CG with the suggested changes is non-trivial. Hence, we demonstrate that using Recycling CG (RCG) helps us achieve these orthogonalities with no code changes. The extra cost of orthogonalizations is often offset by savings because of recycling. Our second and third assumptions involve existence, invertibility and boundedness of two matrices, which are easy to satisfy. While satisfying the backward stability assumptions, by numerical experiments we show that as we iteratively solve the linear systems arising in the AIRGA algorithm more accurately, we obtain a more accurate reduced system. We also show that recycling Krylov subspaces helps satisfy the backward stability assumptions (extra-orthogonalities) at almost no extra cost.