Kevin Milner created STATISTICS-92:
--------------------------------------
Summary: TruncatedNormalDistribution rejection-sampler uses
incorrect bounds for mirroring check
Key: STATISTICS-92
URL: https://issues.apache.org/jira/browse/STATISTICS-92
Project: Commons Statistics
Issue Type: Bug
Components: distribution
Affects Versions: 1.0
Reporter: Kevin Milner
The rejection sampler implementation of TruncatedNormalDistribution incorrectly
uses the original upper and lower bounds rather than the standard normal bounds
for the mirroring check; this leads to inaccurate sampling if the original
bounds are fully-positive or fully-negative, as well as infinite loops if the
bounds are positive but fully below the mean value or are negative but fully
above the mean value.
For example, this will only return samples >1 (even though they should be in
the range [0.7, 1.3]:
{code:java}
ContinuousDistribution dist = TruncatedNormalDistribution.of(1, 0.1, 0.7,
1.3).createSampler(RandomSource.XO_RO_SHI_RO_128_PP.create(123456l));{code}
...and this will get stuck in an infinite loop on the first sampling attempt:
{code:java}
ContinuousDistribution dist = TruncatedNormalDistribution.of(1, 0.1, 0.7,
0.99).createSampler(RandomSource.XO_RO_SHI_RO_128_PP.create(123456l));{code}
Similarly, this will only return samples <1:
{code:java}
ContinuousDistribution dist = TruncatedNormalDistribution.of(-1, 0.1, -1.3,
-0.7).createSampler(RandomSource.XO_RO_SHI_RO_128_PP.create(123456l));{code}
...and this will get stuck in an infinite loop on the first sampling attempt:
{code:java}
ContinuousDistribution dist = TruncatedNormalDistribution.of(-1, 0.1, -0.99,
-0.7).createSampler(RandomSource.XO_RO_SHI_RO_128_PP.create(123456l));{code}
The bug seems to have been present since v1.0. It's a simple fix and I will
submit a PR imminently.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
