Régis LEBRUN

@jschueller Do you think its time for me to (try to) implement my trick on using an approximation of log(link) to avoid overflows?

@m-balesdent I am not sure that there is a bug here. The ignoreFailures flag is set to false by default, and its name tells everything: ignore the failures. Not only...

@m-balesdent You are right! Either we make the optimization algorithms more robust regarding convergence (and we generalize setIgnoreFailure) or we make FORM/SORM more cautious about the optimization result, and we...

There are several tricks to solve this issue: 1) Add a Uniform distribution over the range of the mixture with a weight of order a few CDFEpsilon_. This way, the...

The current distribution, while being probably misnamed, is *not* a categorical distribution, but a discrete distribution over $\mathbb{R}^n$. As such, it has a well defined expectation, a covariance matrix, a...

Something like FiniteSupportDiscreteDistribution?

Do you mean, hide it from the Python interface, or remove it from the C++ interface? I am dubious regarding the first choice (it should be well documented and looks...

There are two bugs in DistributionImplementation::computeQuantile: * the test against quantileEpsilon_ should be moved after the test for dimension==1. In the current code, Binomial::computeScalarQuantile is not called. * cdfEpsilon_ should...

@mbaudin47 I don't have time to review this work right now. Here are my general comments: 1) The conditional expectation part of this PR looks very promising. 2) I agree...

I disagree with the following sentence in the description of this PR: *In the current implementation, when we create a polynomial chaos expansion based on linear regression, we get a...