A Lanczos-type method for solving nonsymmetric linear systems with multiple right-hand sides -- matrix and polynomial interpretation (Q1300798)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | A Lanczos-type method for solving nonsymmetric linear systems with multiple right-hand sides -- matrix and polynomial interpretation |
scientific article |
Statements
A Lanczos-type method for solving nonsymmetric linear systems with multiple right-hand sides -- matrix and polynomial interpretation (English)
0 references
25 January 2001
0 references
An apparently new Lanczos type method for solving linear (symmetric and not) systems with several right hand sides is examined both theoretically and numerically. The method's aim is to avoid the use of multiple Hankel matrices that would arise with the usual implementations of Lanczos' method, because they are determined by the Krylov spaces associated with each of the initial residues. Instead, the Krylov spaces considered here are generated by two arbitrary initial vectors, so that there is one Hankel matrix per iteration to solve each of the systems associated with each of the right hand sides. One of the consequences is that usually, unlike in Lanczos method, the matrix polynomials yielding the successive approximations to each solution are not orthogonal. The paper first examines in detail two implementations of the method corresponding to the basic implementations of Lanczos method known as Orthodir and Orthomin, but discards them because of bad convergence behaviour, although they possess nice implementation properties regarding memory requirements and computational cost per iteration. A transpose-free variant of the Orthomin implementation, along with another implementation based on \textit{H. A. van der Vorst's} stabilized bi-conjugate gradients method (BiCGSTAB) [SIAM J. Sci. Stat. Comput. 13, No.~2, 631-644 (1992; Zbl 0761.65023)] are compared regarding convergence, on six examples, three symmetric and three nonsymmetric. In the symmetric case, the former behaves decidedly better than the BiCGSTAB implementation in one example, while there is a similar behaviour in both methods for the other two. The situation is less simple for the nonsymmetric case: in one example, with low condition number, the Orthomin implementation behaves quite better than BiCGSTAB, but the latter decidedly beats the former in the other two examples, one of them from a benchmark set and both having large condition numbers. This should not come as a surprise, because BiCGSTAB was conceived in order to deal with numerically unstable nonsymmetric problems. It has to be pointed out however, that the general drawback, also remarked in the paper, is that there is an overall bad behaviour of the residues, which suggests the need of further improvement of the method.
0 references
Lanczos method
0 references
Krylov subspaces
0 references
multiple right-hand sides
0 references
Hankel matrix
0 references
orthogonal polynomials
0 references
BiCGSTAB
0 references
nonsymmetric linear systems
0 references
Orthodir
0 references
Orthomin
0 references
bi-conjugate gradients method
0 references
convergence
0 references
0 references
0 references