Parijat Biswas created MATH-1696:
------------------------------------
Summary: EnumeratedRealDistribution computes incorrect quantiles
for unsorted input
Key: MATH-1696
URL: https://issues.apache.org/jira/browse/MATH-1696
Project: Commons Math
Issue Type: Bug
Components: legacy
Affects Versions: 3.3
Reporter: Parijat Biswas
h3. Problem
EnumeratedRealDistribution allows unsorted singletons and duplicate values, but
inverseCumulativeProbability accumulates their probabilities in input order
rather than numerical order. This returns incorrect quantiles, including
incorrect support endpoints.
{code:java}
EnumeratedRealDistribution d = new EnumeratedRealDistribution(
new double[] {7, 3, -1, 3},
new double[] {0.25, 0.25, 0.25, 0.25});
System.out.println(d.inverseCumulativeProbability(0)); // actual 7, expected -1
System.out.println(d.inverseCumulativeProbability(0.5)); // actual 3, expected 3
System.out.println(d.inverseCumulativeProbability(1)); // actual 3, expected 7
{code}
The data-only constructor is affected too: new EnumeratedRealDistribution(new
double[] {7, 3, -1, 3}). The cumulative distribution itself correctly combines
duplicate masses regardless of their input order.
Reproduced on current master 36cefd500d2b6a193bd81654665944db1bb1aaed and the
published Commons Math 3.3 jar. This is distinct from MATH-1065, which fixed
returning points outside the sample set, and MATH-1533, which concerns seeded
sampling order.
h3. Proposed fix
Sort the copied PMF returned by getPmf() only inside
inverseCumulativeProbability. This leaves the stored input order and seeded
sampling unchanged. Add three regression tests covering both constructors,
duplicate masses, zero-probability outliers, p=0 and p=1, and unchanged seeded
sampling after a quantile query.
h3. Validation
* Both unsorted-input regression tests fail on unchanged production code; the
fixed class has 15 passing tests.
* Complete split local test validation on Linux, JDK 17: 3,826 test cases, zero
final failures/errors, 33 existing skips. The repository's configured retry
policy was unchanged; 48 cases were reported as flakes and passed on configured
retries.
* Maven verification phases completed by module, including zero Checkstyle
violations. Existing warning-only PMD and coverage diagnostics remain; no claim
of 100% coverage.
* Javadoc generation passes with UTF-8 locale.
* The initial uninterrupted clean verify command exceeded the execution time
limit. The default legacy test execution completed, then the separate
configured flaky-test execution was rerun to completion. This is split
validation, not a claim that one uninterrupted default build completed. No
test, assertion or retry setting was changed.
AI disclosure: This report and the proposed fix/tests were generated with AI
assistance (Instinct) and checked with the validation above.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)