search

JuliaStats / Distributions.jl

A Julia package for probability distributions and associated functions.
Other
1.11k stars 418 forks source link

Adding WishartCholesky and InverseWishartCholesky #1787

Open sethaxen opened 1 year ago

sethaxen commented 1 year ago

The logpdf function of Wishart computes the Cholesky factorization of the input, while its rand function constructs a Cholesky factor before computing the full matrix, so, like LKJCholesky, it would be convenient to have an implementation of WishartCholesky. This would e.g. allow Turing users to perform inference on parameters with Wishart priors without computing the Cholesky factorization.

Similarly, we could implement an InverseWishartCholesky to avoid Cholesky factorizing the input in logpdf. However, rand does not benefit in this case, since it would be implemented in terms of Wishart, and the Cholesky factorization of the inverse of a matrix is not related to the Cholesky factorization of the matrix.

Side benefits are that random sampling in InverseWishart and MatrixBeta could be sped up using WishartCholesky, though for the latter, this could be a breaking change (currently the struct stores Wishart distributions).

devmotion commented 1 year ago

Completely agree. I have been thinking about it for a few months but never found the time to put together a PR.

sethaxen commented 1 year ago

I have some local implementations I'll polish for a PR. I also looked at the same for MatrixBeta, but there's no way to use the factorization for its logpdf, just rand I think.

sethaxen commented 1 year ago

Similarly, we could implement an InverseWishartCholesky to avoid Cholesky factorizing the input in logpdf. However, rand does not benefit in this case, since it would be implemented in terms of Wishart, and the Cholesky factorization of the inverse of a matrix is not related to the Cholesky factorization of the matrix.

I worked out an algorithm for directly sampling the Cholesky factor of Inverse-Wishart without going through Wishart: https://arxiv.org/abs/2310.15884