MvStudentT incorrect with Cholesky matrix argument

Dear all,

when processing with my project (see previous question here), I noticed a problem with the MvStudentT distribution.
A notebook with a condensed example is here:

  • When trying to use MvStudentT with chol = cholesky_matrix, inference is wrong.
  • However, using the signature cov = covariance_matrix works fine.

By “inference is wrong” I mean that the correlations of the multivariate parameters are not well estimated. From the traceplot, I would guess the the sampling does not take the observed data into account. Everything looks much like Normal in pymc3/distributions/, but I did not follow through to the superclass.

Is this a bug? I could not find hints in the documentation, nor in the code.



That’s fairly surprising, as cov = covariance_matrix is also decomposing the cov to chol internally.
Could you please raise an issue on Github?

Dear Junpeng,
my apologies, but I cannot log in to github to raise an issue. I deleted my github account with their acquisition by Microsoft.


No problem - I will post the issue.


1 Like

Thank you!

Should be fixed on master now thanks to @gBokiau :slight_smile:

1 Like