Add commonly used priors over PSD matrices

[jessegrabowski]

Tell us about it

I'm thinking specifically about LKJCholeskyCov and/or LKJCholeskyCorr, but also Wishart/InverseWishart.

I use these often enough but don't have strong intuitions about them. Would be very handy to be able to explore these. For example, I have read that the marginals of the LKJ priors are quite sensitive to the eta parameter, although the joint is not. These are the types of questions I look to preliz for help.

Thoughts on implementation

Happy to open a PR to help, let me know i these are in-scope for the project.

[aloctavodia]

Short answer. Thanks for offering help. Yes, please add them.

Long answer. The rule is that any distribution (and parameterization) that is already implemented in a PPLs in Python should also be in PreliZ. In practice this is more a guide than a rule, because we have been focusing on PyMC, due to time constraints and familiarity. Also we have a couple of distributions that are not in PyMC like beta scaled (and one more than I don't remember). Also as you noticed, there are a couple missing, I think all the missing ones are multivariate. In general I would say that the support for multivariate distribution is not as good as it should be in PreliZ. So this proposal is very in the scope of the project.

[jessegrabowski]

Awesome, I'll open a PR in the next couple days!