Shifted extended global Lanczos processes for trace estimation with application to network analysis

In memory of Mohammed Bellalij

Abstract

The need to estimate upper and lower bounds for matrix functions of the form ${\mathrm{trace}}(W^Tf(A)V)$, where the matrix $A\in {{\mathbb {R}}}^{n\times n}$ is large and sparse, $V,W\in {{\mathbb {R}}}^{n\times s}$ are block vectors with $1\le s\ll n$ columns, and f is a function arises in many applications, including network analysis and machine learning. This paper describes the shifted extended global symmetric and nonsymmetric Lanczos processes and how they can be applied to approximate the trace. These processes compute approximations in the union of Krylov subspaces determined by positive powers of A and negative powers of $A-\sigma I_n$, where the shift $\sigma$ is a user-chosen parameter. When A is nonsymmetric, transposes of these powers also are used. When A is symmetric and $W=V$, we describe how error bounds or estimates of bounds for the trace can be computed by pairs of Gauss and Gauss-Radau quadrature rules, or by pairs of Gauss and anti-Gauss quadrature rules. These Gauss-type quadrature rules are defined by recursion coefficients for the shifted extended global Lanczos processes. Gauss and anti-Gauss quadrature rules also can be applied to give estimates of error bounds for the trace when A is nonsymmetric and $W\ne V$. Applications to the computation of the Estrada index for networks and to the nuclear norm of a large matrix are presented. Computed examples show the shifted extended symmetric and nonsymmetric Lanczos processes to produce accurate approximations in fewer steps than the standard symmetric and nonsymmetric global Lanczos processes, respectively.