[
https://issues.apache.org/jira/browse/RNG-183?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alex Herbert resolved RNG-183.
------------------------------
Fix Version/s: 1.6
Resolution: Implemented
Update in commit:
25fa9cd68ab1061c7f914543d69b5e2b3e98056f
Changed the sampler from using nextDouble for the value of (1 - p) to using
nextLong as a source of random bits. This is converted to the range [0, 1) or
(0, 1] as appropriate based on the shape. The switch point is done at shape = 1.
Tests verify that extreme shape parameters return samples from the lower or
upper end of the distribution bounds.
> Pareto distribution sampler to concentrate samples at the lower/upper bounds
> for extreme shape parameters
> ---------------------------------------------------------------------------------------------------------
>
> Key: RNG-183
> URL: https://issues.apache.org/jira/browse/RNG-183
> Project: Commons RNG
> Issue Type: Improvement
> Components: sampling
> Affects Versions: 1.5
> Reporter: Alex Herbert
> Priority: Minor
> Fix For: 1.6
>
>
> The Pareto distribution for scale and shape is bounded by [scale, infinity).
> When shape -> infinity the distribution concentrates values close to scale.
> When 1 / shape -> infinity the distribution concentrates values close to
> infinity.
> The sampler uses inverse transform sampling. This can introduce artefacts in
> the samples when p=0 or p=1 depending on the shape. This has been corrected
> in STATISTICS-59 to sample from p in [0, 1) for large shape and p in (0, 1]
> for small shape. The same correction should be applied within the sampler.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
