The extended global Lanczos method for matrix function approximation (Q1716855)

In problems arising in network analysis, machine learning and electronic structure computation among others, there is a need to compute the trace of a large symmetric matrix. In the literature, the global Lanczos method or an extended Lanczos method is applied to compute the trace approximately. The present work describes an extended global Lanczos method and discusses properties of the corresponding Gauss-Laurent quadrature rules. Numerical illustrations are presented, showing that, while the (standard) global Lanczos method is faster, the new procedure requires fewer iterations.