Hello. > > > > I've encountered a failure of one the tests which you introduced in > > revision 1363105. Here is an excerpt of the error: > > --- > > > > testNormalDistributionUnsymmetricMatrix(org.apache.commons.math3.linear.EigenDecompositionTest) > > Time elapsed: 0.044 sec <<< FAILURE! > > java.lang.AssertionError: The norm of (X-Y) is too large > > at org.junit.Assert.fail(Assert.java:93) > > at org.junit.Assert.assertTrue(Assert.java:43) > > at > > org.apache.commons.math3.linear.EigenDecompositionTest.checkUnsymmetricMatrix(EigenDecompositionTest.java:422) > > at > > org.apache.commons.math3.linear.EigenDecompositionTest.testNormalDistributionUnsymmetricMatrix(EigenDecompositionTest.java:404) > > [... etc ...] > > --- > > > > For tests that uses a RNG, the conclusion was reached that it is better > > (for > > unit tests) to select a seed for which the test succeeds. > > > > ok thanks for the hint. The random test principle is flawed but I wanted to > check with lots of different input data if the whole algorithm is stable. > Now there was not yet a way to quickly display the corresponding matrix, > but I added a RealMatrixFormat for that purpose. So I will update the test > and use also a fixed seed to make it predictable (and do more investigation > on failing cases).
Maybe that we should create an informal set of validation tests in the same way that there are performance tests. I.e. such classes would have a name that ends with "...TestValidation". Not ending in "...Test", they would not be run automatically. [Cf. the "FastMathTestPerformance" class.] Regards, Gilles --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
