Which eigenvalues are found by the Lanczos method? (Q2706258)

\textit{L. N. Trefethen} and \textit{D. B. Bau} [Numerical linear algebra (1997; Zbl 0874.65013)] stated the thumb rule that the Lanczos iteration tends to converge to eigenvalues of real symmetric matrices that lie in regions of ``too little charge'' for an equilibrium distribution, so that outliers are well approximated, whereas eigenvalues in the bulk of the spectrum are poorly approximated.NEWLINENEWLINENEWLINEThe paper has the goal to provide a quantitative version of this rule, asymptotically for matrix size \(N\) and number of iterations \(n\). It uses results of \textit{E. A. Rakhmanov} [Sb. Math. 187, No. 8, 1213-1228 (1996; Zbl 0873.42014)] to obtain a potential-theoretic characterization of the problem as an extremal problem.