[
https://issues.apache.org/jira/browse/MATH-611?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13060478#comment-13060478
]
Christopher Nix commented on MATH-611:
--------------------------------------
With regard to speed, I've tried it on random arrays with up to 1,000,000
elements and it's definitely faster than the current SVD implementation.
The current SVD implementation is noticeably slowed by the implementations of
AbstractRealMatrix and TriDiagonalTransformer that use setEntry and getEntry on
matrices within loops. As such, matrix indices are checked to be in bounds at
every iteration of a loop, when what we really should do is check the max and
min indices only, if at all. (I've a further patch yet to submit on this).
If we remove the use of getEntry and setEntry in methods called by the current
SVD implementation, and instead use direct array access, then the speed of the
current implementation is improved significantly, however the stability remains
unchanged. Even with this change, the JAMA code is about twice as fast to
converge (according to my profiler) and it is demonstrably more accurate.
Chris.
> A fast and stable SVD implementation from JAMA
> ----------------------------------------------
>
> Key: MATH-611
> URL: https://issues.apache.org/jira/browse/MATH-611
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 3.0, Nightly Builds
> Reporter: Christopher Nix
> Labels: patch
> Fix For: 3.0, Nightly Builds
>
> Attachments: SingularValueDecompositionImpl.java
>
>
> Common numerical stability issues with the current SVD implementation, ie
> MATH-327, MATH-383, MATH-465, MATH-583 can all be solved by co-opting JAMA
> code that is within the public domain.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
