Data Parallel Lanczos and Pad E-rayleigh-ritz Methods on the Cm5
نویسنده
چکیده
The projective Pad e-Rayleigh-Ritz method that we developed in 6]consists in approximating the eigenvalues of a hermitian matrix A of order n, by the roots of the polynomial of degree m of the denominator of m ? 1=m] Rx () where m ? 1=m] Rx () is the Pad e approximant of order m of R x () = ((I ? A) ?1 x; x) the mean value of the resolvant of A. The Lanczos and PRR methods are closely linked. This prompts us to ask rst whether PRR has a good stability property compared with the Lanc-zos. The second question which arises is to know if the exploitation of this method can be interesting from the point of view of its computation cost. In this paper we deal more particularly with this practical second question than with the rst. Next, we give performances of the projection phase of these two methods on the massively parallel architecture of Connection Machine 5 (CM5). La m ethode de projection Pad e-Rayleigh-Ritz que nous avons d evelopp e dans 6]consiste a approcher les valeurs propres d'une matrice hermitienne A d'ordre n, par les racines du polyn^ ome du degr e m du d enominateur de m ? 1=m] Rx () o u m ? 1=m] Rx () est l'approximant de Pad e d'ordre m de R x () = ((I ? A) ?1 x; x) la valeur moyenne de la r solvante de A. Les m ethodes de Lanczos et PRR sont tr es etroitement li ees. La premi ere question qui se pose est PRR est-elle une m ethode num eriquement stable relativement a Lanczos? La question suivante est l'exploitation de cette m ethode conduit-elle a des r esultats int eressant au point du vue du co^ ut du calcul? Dans ce papier nous nous con-centrons sur cette seconde question pratique plut^ ot que sur la premi ere. Nous pr esentons, alors, les performances de la phase de projection de ces deux m ethodes sur l'architecture massivement parall ele de la Connection Machine 5 (CM5).
منابع مشابه
The Pad e - Rayleigh - Ritz Method for SolvingLarge
We make use of the Pad e approximants and Krylov's sequence x; Ax; ; : : :; A m?1 x in the projection methods to compute a few Ritz values of a hermitian matrix A of order n. This process consists of approaching the poles of R x () = ((I ?A) ?1 x; x), the mean value of the resolvant of A, by those of m ? 1=m] Rx (), where m ? 1=m] Rx () is the Pad e approximant of order m of the function R x ()...
متن کاملAdjusting the Rayleigh Quotient in Semiorthogonal Lanczos Methods Adjusting the Rayleigh Quotient in Semiorthogonal Lanczos Methods
In a semiorthogonal Lanczos algorithm, the orthogonality of the Lanczos vectors is allowed to deteriorate to roughly the square root of the rounding unit, after which the current vectors are reorthogonalized. A theorem of Simon 4] shows that the Rayleigh quotient | i.e., the tridiagonal matrix produced by the Lanczos recursion | contains fully accurate approximations to the Ritz values in spite...
متن کاملAdjusting the Rayleigh Quotient in Semiorthogonal Lanczos Methods
In a semiorthogonal Lanczos algorithm, the orthogonality of the Lanczos vectors is allowed to deteriorate to roughly the square root of the rounding unit, after which the current vectors are reorthogonalized. A theorem of Simon 4] shows that the Rayleigh quotient | i.e., the tridiagonal matrix produced by the Lanczos recursion | contains fully accurate approximations to the Ritz values in spite...
متن کاملRayleigh-ritz Majorization Error Bounds with Applications to Fem and Subspace Iterations
The Rayleigh-Ritz method finds the stationary values, called Ritz values, of the Rayleigh quotient on a given trial subspace as approximations to eigenvalues of a Hermitian operator A. If the trial subspace is A-invariant, the Ritz values are exactly some of the eigenvalues of A. Given two subspaces X and Y of the same finite dimension, such that X is A-invariant, the absolute changes in the Ri...
متن کاملGeneralized block Lanczos methods for large unsymmetric eigenproblems
Generalized block Lanczos methods for large unsymmetric eigenproblems are presented, which contain the block Arnoldi method, and the block Arnoldi algorithms are developed. The convergence of this class of methods is analyzed when the matrix A is diagonalizable. Upper bounds for the distances between normalized eigenvectors and a block Krylov subspace are derived, and a priori theoretical error...
متن کامل