[ https://issues.apache.org/jira/browse/MATH1384?page=com.atlassian.jira.plugin.system.issuetabpanels:alltabpanel
]
Arman Bilge updated MATH1384:

Description:
For certain edge cases, HypergeometricDistribution.logProbability() will return NaN.
To compute the hypergeometric log probability, three binomial log probabilities are computed
and then combined accordingly. The implementation is essentially the same as in BinomialDistribution.logProbability()
and uses the SaddlePointExpansion. However, the Binomial implementation includes an extra
check for the edge case of 0 trials which the HyperGeometric lacks.
An example call which fails is:
new HypergeometricDistribution(null, 11, 0, 1).logProbability(0)
which returns NaN instead of 0.0.
Note that
`new HypergeometricDistribution(null, 10, 0, 1).logProbability(0)`
returns 0 as expected.
Possible fixes:
1. Check for the edge cases and return appropriate values. This would make the code somewhat
more complex.
2. Instead of duplicating the implementation use BinomialDistribution.logProbability(). This
is much simpler/more readable but will reduce performance as each call to BinomialDistribution.logProbability()
makes redundant checks of validity of input parameters etc.
I am happy to submit a PR at the GitHub repo implementing either 1 or 2 with the necessary
tests.
was:
For certain edge cases, HypergeometricDistribution.logProbability() will return NaN.
To compute the hypergeometric log probability, three binomial log probabilities are computed
and then combined accordingly. The implementation is essentially the same as in BinomialDistribution.logProbability()
and uses the SaddlePointExpansion. However, the Binomial implementation includes an extra
check for the edge case of 0 trials which the HyperGeometric lacks.
An example call which fails is:
new HypergeometricDistribution(null, 11, 0, 1).logProbability(0)
which returns NaN instead of 0.0.
Note that
new HypergeometricDistribution(null, 10, 0, 1).logProbability(0)
returns 0 as expected.
Possible fixes:
1. Check for the edge cases and return appropriate values. This would make the code somewhat
more complex.
2. Instead of duplicating the implementation use BinomialDistribution.logProbability(). This
is much simpler/more readable but will reduce performance as each call to BinomialDistribution.logProbability()
makes redundant checks of validity of input parameters etc.
I am happy to submit a PR at the GitHub repo implementing either 1 or 2 with the necessary
tests.
> HypergeometricDistribution logProbability() returns NaN for edge cases
> 
>
> Key: MATH1384
> URL: https://issues.apache.org/jira/browse/MATH1384
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.0, 4.0
> Reporter: Arman Bilge
> Priority: Minor
>
> For certain edge cases, HypergeometricDistribution.logProbability() will return NaN.
> To compute the hypergeometric log probability, three binomial log probabilities are computed
and then combined accordingly. The implementation is essentially the same as in BinomialDistribution.logProbability()
and uses the SaddlePointExpansion. However, the Binomial implementation includes an extra
check for the edge case of 0 trials which the HyperGeometric lacks.
> An example call which fails is:
> new HypergeometricDistribution(null, 11, 0, 1).logProbability(0)
> which returns NaN instead of 0.0.
> Note that
> `new HypergeometricDistribution(null, 10, 0, 1).logProbability(0)`
> returns 0 as expected.
> Possible fixes:
> 1. Check for the edge cases and return appropriate values. This would make the code somewhat
more complex.
> 2. Instead of duplicating the implementation use BinomialDistribution.logProbability().
This is much simpler/more readable but will reduce performance as each call to BinomialDistribution.logProbability()
makes redundant checks of validity of input parameters etc.
> I am happy to submit a PR at the GitHub repo implementing either 1 or 2 with the necessary
tests.

This message was sent by Atlassian JIRA
(v6.3.4#6332)
