josephmure
josephmure
I think there are only a few comments that still need to be taken into account, and then I will approve this PR.
Does PR #2402 really fix this issue as presented? @sofianehaddad wanted ```python print(ot.Beta(1., 2., 0, 1) == ot.Uniform(0, 1)) ``` to return `True`. @jschueller Does your commit make this happen?...
Restoring the numerical bound is a workaround IMO, not a fix. We need to understand why a NaN is produced. Note that when you make the correlation positive the result...
To investigate further, let us simplify the code provided by @mbaudin47: ```python import openturns as ot L = ot.Uniform(0.0, 1.0) R = ot.CorrelationMatrix(2) R[0, 1] = -0.1 copula = ot.NormalCopula(ot.NormalCopula.GetCorrelationFromSpearmanCorrelation(R))...
The corresponding code to check this: ```python sam = ot.Sample([[0,0], [1,1], [1,0], [0,1], [0.5, 0], [0.5, 1], [0, 0.5], [1, 0.5]]) print(sam) print(f(sam)) ``` ``` 0 : [ 0 0...
Bottom line: if the linear evaluation were smarter and used the triangular structure in its computations, we would get rid of many NaNs, but not all, because : 1. we...
Why is it transposed? This is really counterintuitive. Shouldn't we change the behavior of LinearFunction rather than the doc?
In practice, what I do when I want linear levels discretization is call `np.linspace` but it is clunky. I agree we need this setting.
Ah, I did not see that, thanks! Maybe we could add warnings, telling users to check if the arguments of their InverseGamma distribution are in the correct order? But even...
@jschueller this is ready to be merged IMO