home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: sci.math.num-analysis
- Path: sparky!uunet!math.fu-berlin.de!unidui!rrz.uni-koeln.de!Germany.EU.net!ifado!wb
- From: wb@arb-phys.uni-dortmund.de (Wilhelm B. Kloke)
- Subject: Re: 32-bit vs 64-bit precision ?
- Message-ID: <1992Nov19.071505.5501@arb-phys.uni-dortmund.de>
- Organization: Institut f. Arbeitsphysiologie a.d. Uni Dortmund
- References: <1992Nov17.115111.23520@csrd.uiuc.edu>
- Date: Thu, 19 Nov 92 07:15:05 GMT
- Lines: 16
-
- In article <1992Nov17.115111.23520@csrd.uiuc.edu> hblim@sp1.csrd.uiuc.edu (Hock-Beng Lim) writes:
- >For example, are there any numerical properties which might make an
- >algorithm stable when implemented in 64-bit precision ? I heard that Monte
- >Carlo techniques are more stable when implemented in 64-bit precision. Is
- >that true ? What about some other such differences for other well-known
- >algorithms ? The algorithms I am primarily interested in are the linear
- >system solvers and the eigensolvers.
- At least, it is well known that dot products which occur in Matrix
- multiplication should be carried out in higher precision than that of
- the stored elements. Dot products (Sum over A_i * B_i) tend to
- cancellation if the vectors are nearly orthogonal. This will cause
- problems if the matrices are ill-conditioned.
- --
- Dipl.-Math. Wilhelm Bernhard Kloke,
- Institut fuer Arbeitsphysiologie an der Universitaet Dortmund
- Ardeystrasse 67, D-4600 Dortmund 1, Tel. 0231-1084-257
-