Superlinear convergence of the rational Arnoldi method for the approximation of matrix functions (Q431305)

To approximate the vector \(f(A)b\) with \(A\) an Hermitian matrix and \(b\) a vector of size \(N\), and \(f\) a function, one computes in the rational Arnoldi method this vector as \(V_nf(A_n)V_n^*b\) with (the columns of) \(V_n\) the orthogonal basis of a rational Krylov space of dimension \(n\) and \(A_n=V_n^*AV_n\) a much smaller matrix of size \(n\) (\(N\gg n\)). In this paper, the authors derive bounds for the superlinear convergence when \(f\) has a Cauchy-Stieltjes representation \(f(z)=\int _\Gamma (x-z)^{-1}d\gamma(x)\) with \(\gamma\) a complex measure supported on a closed set \(\Gamma=\mathbb{C}\cup\{\infty\}\setminus [\lambda_{\min},\lambda_{\max}]\) and where the (finite or infinite) poles of the rational function underlying the rational Krylov space are appropriately chosen outside the spectrum of \(A\). The analysis uses potential theory where the equilibrium problem is solved with an external field imposed by the poles. These results rely strongly on their previous paper with \textit{R. Vandebril} [SIAM J. Matrix Anal. Appl. 31, No. 4, 1740--1774 (2010; Zbl 1213.15009)].