On One-Point Iterations and DIIS

Publikation: Bidrag til tidsskrift/Konferencebidrag i tidsskrift /Bidrag til avisKonferenceartikelForskningpeer review

  • Datalogisk Institut
We analyze various iteration procedures in many dimensions inspired by the SCF iteration used in first principles electronic structure calculations. We show that the simple mixing of densities can turn a divergent (or slowly convergent) iteration into a (faster) convergent process provided all the eigenvalues of the Jacobian are real and less than one. The problem of determining good mixing parameters are solved with the DIIS method of Pulay. We show that the large condition numbers often encountered in the linear system has no detrimental effect on the solution, i.e., the weights. Large weights can occur if the residual vectors are (nearly) linearly dependent. We show how to remove this linear dependence using the singular value decomposition (SVD).
OriginalsprogEngelsk
TidsskriftA I P Conference Proceedings Series
Sider (fra-til)468-472
Antal sider5
ISSN0094-243X
DOI
StatusUdgivet - 2009
BegivenhedInternational Conference on Numerical Analysis and Applied Mathematics - Rethymno, Grækenland
Varighed: 18 sep. 200922 sep. 2009
Konferencens nummer: 7

Konference

KonferenceInternational Conference on Numerical Analysis and Applied Mathematics
Nummer7
LandGrækenland
ByRethymno
Periode18/09/200922/09/2009

Bibliografisk note

Volume title: Numerical analysis and Applied Mathematics.
ISBN 978-0-7354-0705-3
Volumne: 1168

Se relationer på Aarhus Universitet Citationsformater

ID: 17711430