The e-MoM approach for approximating matrix functionals (Q1989169)

The problem is to evaluate \(X^Tf(A)Y\) for \(A\) being a diagonalizable matrix and \(X\) and \(Y\) block vectors. If \(A=\sum_k \lambda_k p_k q_k^T\) is the eigenvalue decomposition, then the problem is reduced to evaluate many sums of the form \(\sum_k f(\lambda_k) \alpha_k\beta_k\) with \(\alpha_k=x^Tp_k\) and \(\beta_k=q_k^Ty\). The e-MoM of the title stands for extrapolated Method of Moments. That means that \(f(x)\) is approximated by the initial terms of its power series expansion. Using moment matching and extrapolation, one gets relatively good approximations without having to compute the full eigenvalue decomposition. This idea was used for Hermitian matrices by \textit{P. Fika} and \textit{M. Mitrouli} [Linear Algebra Appl. 502, 140--158 (2016; Zbl 1336.65068)] for one or two-term estimates. Here the method is re-derived for non-Hermitian \(A\) and also the formulas for a three-term estimate are added. A Wilkinson-type backward error analysis for each step in the computational procedure is performed. This allows to show that for certain classes one-term estimates are stable and advise whether or not to move to more term estimates can be given. Extensive numerical tests illustrate that for some classes of matrices the method is more stable and more efficient than alternative methods such as those based on the polarization identity or on quadrature formulas.