tpapp/AltDistributions.jl

LKJL with TransformVariables

Closed this issue · 3 comments

cscherrer commented

I'd like to adapt LKJL for MeasureTheory.jl. I have a proof of concept, but @sethaxen found an inconsistency (JuliaMath/MeasureTheory.jl#100), and I'm not sure of the right way to work around it.

LKJL in this package works with the lower Cholesky factor. TransformVariables has CorrCholeskyFactor that seems like it would work, but it works with the upper Cholesky factor.

Do you have a way of using TransformVariables with LKJL?

sethaxen commented

The lower and upper cholesky factors are just transposes of each other, so the CorrCholeskyFactor transformation can be used if followed by a transpose. Since transpose has a logabsdetjac of 0, there's no need to adjust the scoring function. Unless I've missed something.

  • 👍1
cscherrer commented

Good point @sethaxen , it should be very easy to do this and create a new transformation. @tpapp , I'll leave this open in case you have anything to add. If not, feel free to close the issue.

tpapp commented

@sethaxen is correct. I will just close this, but feel free to ping me here if anything comes up and I will reopen.

  • 👍1