Alex D Herbert created RNG-52:
---------------------------------
Summary: PoissonSampler allows mean above Integer.MAX_VALUE
Key: RNG-52
URL: https://issues.apache.org/jira/browse/RNG-52
Project: Commons RNG
Issue Type: Bug
Affects Versions: 1.1
Reporter: Alex D Herbert
The {{PoissonSampler}} is limited to returning an integer by the interface of
the {{DiscreteSampler}}. As it stands an input mean above {{Integer.MAX_VALUE}}
is allowed although it makes no sense as the Poisson distribution is
significantly truncated.
The algorithm of the {{SmallMeanPoissonSampler}} sets a limit on the returned
sample of {{Integer.MAX_VALUE}}. The algorithm is valid although run-time would
be impractical due to the nature of the algorithm. However at high mean (>40)
the end user is expected to use either the {{LargeMeanPoissonSampler}} directly
or the {{PoissonSampler}} which chooses the appropriate large mean algorithm.
However the current {{LargeMeanPoissonSampler}} uses {{(int)Math.floor(mean)}}
during initialisation and any mean above {{Integer.MAX_VALUE}} would therefore
be unsupported.
I propose to add this to the constructor of each Poisson sampler:
{code:java}
if (mean > Integer.MAX_VALUE) {
throw new IllegalArgumentException(mean + " > " + Integer.MAX_VALUE);
}
{code}
with documentation
{code:java}
* @throws IllegalArgumentException if {@code mean <= 0} or {@code mean >
}{@link Integer.MAX_VALUE}.
{code}
It is noted that the limit of {{Integer.MAX_VALUE}} would allow the samples to
reflect the Poisson distribution below that level but truncate it above that
level to represent the remaining cumulative histogram at the single point of
{{Integer.MAX_VALUE}}. This maintains the functionality of the sampler within
the contract of the integer value returned by {{DiscreteSampler}}.
In practice the Poisson distribution is unlikely to be used at such a high
mean; in this case it is appropriate to use a Gaussian approximation to the
Poisson.
Note: Currently there is no code coverage from tests for the
\{{LargeMeanPoissonSampler}} checking if the mean is <= 0. Tests should be
added to check the constructor does throw when a bad mean is used.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
