Near instance optimality of the Lanczos method for Stieltjes and related matrix functions (Q6936079)

This interesting paper studies near instance optimality of the Lanczos method for Stieltjes and related matrix functions. The setting is the following. Consider the problem of the approximation of a matrix function of the form \(f(A)b\) where the following are given and fixed. \(A\in \mathbb C^{n\times n}\) is Hermitian, the function \(f\) is defined on the spectrum of the matrix \(A\) and \(b\in \mathbb C^n\). This problem has many applications when for example \(A\) is large and sparse. Some are in scientific computing, data science, solutions of fractional differential equations, complex networks, Gaussian process regression, theoretical particle physics and many more. Typically, Krylov methods have been known for some time to perform this approximation well for various \(f\). When \(A\) is Hermitian and positive definite, the Lanczos method is the standard choice of Krylov methods and performs better than more sophisticated methods. Indeed, the error is of the order of best possible approximation from the Krylov space.\N\NThe authors study theoretical guarantees for the deviation of the Lanczos error from the optimal error and prove a rigorous bound for this deviation when the function \(f\) is a Stieltjes function which includes inverse fractional powers, square root and shifted logarithm. Their bounds improve on existing bounds in the literature.\N\NThe paper is very well written with a very good set of references.