On the complex Falk-Langemeyer method (Q2290913)

The author presents an algorithm for a high relative accuracy computation of the generalized eigenvalue problem \(A x = \lambda B x\), \(x \neq 0\), for the case that the matrices \(A\) and \(B\) build a pair of definite complex Hermitian matrices. A pair of such matrices is definite if there exist real numbers \(c_1\) and \(c_2\) such that \(c_1 A + c_2 B\) is positive definite. The algorithm is an extension of the method of \textit{S. Falk} and \textit{P. Langemeyer} [Elektron. Datenverarb. No. 7, 30--34 (1960; Zbl 0100.33104)] for the simultaneous diagonalization of a pair \(A, B\) of real symmetric positive definite matrices. The author shows by numerical tests that the eigenvalues and eigenvectors are computed with high relative accuracy provided that the condition numbers of \(D_A A D_A\) and \(D_B B D_B\) are small for some diagonal matrices \(D_A\) and \(D_B\). The algorithm works also very well when the starting matrices are ill-conditioned with respect to matrix inversion. Another advantage is the inherent parallelism. The algorithm is very fast if \(A\) and \(B\) are nearly diagonal. Due to its properties the method can be used also as a kernel algorithm for block Jacobi methods and in the computation of generalized singular value decompositions. The complex Falk-Langemeyer algorithm is derived in great detail. A sequence of congruent matrix pairs is generated to solve the problem. The transformation matrices define the so-called pivot pairs which are selected by a pivot strategy. A special algorithm for the computation of the corresponding transformation parameters is presented. Some ideas for the proofs of the global and of the quadratic asymptotic convergence of the method are provided. Another topic of the paper is the analysis of the influence of rounding errors and an extensive section on the choice of numerical tests to validate the theoretical results.