An inexact Krylov-Schur algorithm for the unitary eigenvalue problem (Q935390)

The authors develop a structure-preserving implicitly-restarted process for large unitary eigenvalue problems. The restarts are done by \textit{G.W. Stewart}'s Krylov-Schur method [SIAM J. Matrix Anal. Appl. 23, No.3, 601--614 (2001); addendum ibid. 24, No.2, 599--601 (2002; Zbl 1003.65045)] which amounts to the thick restart method of \textit{K. Wu} and \textit{H. Simon} in the symmetric case [SIAM J. Matrix Anal. Appl. 22, No.2, 602--616 (2000; Zbl 0969.65030)]. The algorithm preserves the unitary structure by employing a variant of \textit{W.B. Gragg}'s isometric Arnoldi process [J. Comput. Appl. Math. 16, 1--8 (1986; Zbl 0623.65041) and J. Comput. Appl. Math. 46, No.1-2, 183--198 (1993; Zbl 0777.65013)] to generate the approximating Krylov space. An upper Hessenberg matrix \(H_m\) is produced in the form of Schur parameters, and all manipulations are done in terms of the Schur parameters. Since \(H_n\) is not quite unitary, the Arnoldi process is modified to produce a unitary matrix that is used in the filtering process for the implicit restart. Numerical results for diagonal matrices confirm that the algorithm produces accurate results and is superior to a conventional Krylov-Schur algorithm.