TRIANGULARIZATION-BASED METHOD FOR MATRIX FUNCTIONS OF PENCILS AND APPLICATIONS TO GEOMETRY ON SPD MANIFOLD

Read full text Summarise with VIRAT
Authors: Rémy Boyer
Cited by: 0
References: 32

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Matrix functions of the form f (AB -1 ), where A and B are symmetric positive definite (SPD) matrices, arise naturally in several problems of numerical linear algebra and in the geometry of the SPD manifold. Their reliable evaluation is however nontrivial, since the matrix product AB -1 is in general nonsymmetric, and its explicit construction may lead to a significant loss of numerical accuracy, especially when B is ill-conditioned. In this work, we propose a QZ-Parlett strategy that avoids the explicit formation of AB -1 by exploiting a joint triangularization of the pencil (A, B). More precisely, a QZ decomposition reduces the problem to triangular factors (T A , T B ) and yields a Schur-like representation whose core quantity is f (T A T -1 B ). This makes it possible to work entirely with triangular factors: the inversion of B is avoided through backward substitution with T B , while the matrix function is evaluated by a stable Parlett recurrence on the triangular core B . The contributions of the paper are threefold. First, we develop this QZ-Parlett computational framework for matrix functions of the form f (AB -1 ). Second, we provide a stability analysis, including an upper bound that describes the effect of the conditioning of B and certificate-type diagnostics that assess the reliability of the computed results. Third, we show that a closely-related triangular core naturally extends to geometric quantities on the SPD manifold associated with the affine-invariant Riemannian metric, including geodesics and the exponential and logarithm maps. The numerical experiments confirm the practical interest of this approach. In ill-conditioned regimes, the proposed method remains markedly more stable than several state-ofthe-art alternatives, while preserving the expected qualitative behavior in the associated geometric computations.

Fields of Study & Subjects

Matrix Theory and Algorithms Numerical Methods and Algorithms Advanced Optimization Algorithms Research

Publication details

Journal / source
HAL
Publication date
16 March 2026
Language
EN
Document type
Preprint