Accelerating Block-Jacobi Methods (CROSBI ID 501187)
Prilog sa skupa u zborniku | sažetak izlaganja sa skupa | međunarodna recenzija
Podaci o odgovornosti
Hari, Vjeran
engleski
Accelerating Block-Jacobi Methods
Until recently QR methods and their variations have been considered as the fastest methods for computing the singular value and the symmetric eigenvalue problems. Lately, Divide and Conquer methods have developed to a stage at which for larger matrices they overcome the performance of the QR methods. However, the both types of methods, require reducing the initial full matrix to bidiagonal (tridiagonal in case of eigenvalue problem) form. This preprocessing can deteriorate accuracy of smaller singular values (eigenvalues). In addition, reduction to bidiagonal (tridiagonal) form is getting costly since the later part of the QR and DC methods is getting faster through newer variations (e.g.\ DQD+inverse iteration). The third type methods are diagonalization or Jacobi methods. They are very accurate as they can deliver all output data to the accuracy that is allowed by the problem. However, for larger matrices, Jacobi methods are generally considered to be several times slower than the appropriate QR methods. The aim of this communication is to give a closer look at the latest developments in accelerating one-sided Jacobi methods and to show that the newest variations of one-sided Jacobi methods are as fast as the fastest QR methods. In order to further enhance efficiency, block versions of these methods have been designed. It is shown how CS decomposition and some other tricks can be used to reduce the flop count in the slowest part of the algorithm -- updating the block columns.
Block-Jacobi methods
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
Podaci o prilogu
6-6-x.
2003.
nije evidentirano
objavljeno
Podaci o matičnoj publikaciji
Guaranteed Error-Bounds for the Solution of Nonlinear Problems in Applied Mathematics.
Herzberger J.
München: GAMM
Podaci o skupu
Guaranteed Error-Bounds for the Solution of Nonlinear Problems in Applied Mathematics.
predavanje
01.09.2003-05.09.2003
München, Njemačka