[
https://issues.apache.org/jira/browse/RNG-159?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17394946#comment-17394946
]
Alex Herbert commented on RNG-159:
----------------------------------
Further investigation using the reference C implementation shows that the
problem exists in the source implementation. It can be avoided by not using the
fast rejection method when sampling the convex overhangs of the ziggurat. This
may invalidate the performance gains of the method.
The issue has been reported to the maintainer of the reference implementation.
> ZigguratSampler.NormalizedGaussian has incorrect Gaussian output
> ----------------------------------------------------------------
>
> Key: RNG-159
> URL: https://issues.apache.org/jira/browse/RNG-159
> Project: Commons RNG
> Issue Type: Bug
> Components: sampling
> Affects Versions: 1.4
> Reporter: Alex Herbert
> Priority: Blocker
> Fix For: 1.4
>
> Attachments: pdf.gauss.modified.ziggurat.txt.png,
> pdf.gauss.ziggurat.txt.png
>
>
> The new ZigguratSampler.NormalizedGaussian does not produce suitable output
> from the examples sampling application:
> !pdf.gauss.modified.ziggurat.txt.png!
> Compare with the current ZigguratNormalizedGaussianSampler:
> !pdf.gauss.ziggurat.txt.png!
> This is is repeatable with several different RNG implementations.
> This issue was not detected with the current test suite in the sampling
> module which uses the deciles of the distribution. The deciles may be too
> coarse grained to detect a problem around the distribution mean.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
