Closed jishnub closed 8 months ago
Most certainly caused by the recent PDMats updates. Possibly I introduced a new bug and/or numerical instability in these bugfixes.
All but the last issue should be fixed by https://github.com/JuliaStats/PDMats.jl/pull/190.
Indeed, with the new PDMats release only the cholesky
factorization issue shows up in https://github.com/JuliaStats/Distributions.jl/actions/runs/6338587838.
The remaining test error seems to be caused by https://github.com/JuliaStats/PDMats.jl/commit/a97533d1b2c162d6354834be774beff769f7e8eb: With this performance improvement it becomes more likely that the covariance matrix of the scaled mv normal distribution in https://github.com/JuliaStats/Distributions.jl/blob/e407fa5fd098e50df51801c6d062946eac7a7d0f/src/multivariate/mvnormal.jl#L298 is numerically not exactly symmetric. I think we should wrap the output of X_A_Xt
in a Symmetric
or Hermitian
, either in PDMats or Distributions, to avoid such errors.
This fixed the remaining issues, tests on the master branch pass again.
Thanks for the rapid response and the fixes!
I just cloned the package (v0.25.102), and I found the following tests to be failing on Julia v1.9.3:
Since the CI was green on the commit, perhaps the issue lies upstream? It'll be good to identify this anyway