Sam

Good to hear! I'm interested in generating meshes for Markov random spatial fields, to use in spatio-temporal statistical models. My use case is to take a set of geographical boundaries...

I was recently looking at the method in [this paper](https://arxiv.org/abs/0901.4107), which is what VoronoiDelaunay.jl uses and seems like a fast and robust approach. I don't think it would be that...

This is also something I would find really valuable. From trying out a toy model, am I correct that Distributions is using PDMats which, when it constructs a `PDSparseMat`, calls...

Reading the stacktrace for my error above more carefully, I the error is actually coming from the `Sparse` constructor, not the Cholesky decomposition, so hopefully https://github.com/JuliaDiff/ChainRules.jl/pull/246 will solve the problem.

I've done more investigating and think that it might be easier to implement a custom `TuringMvNormalCanon` type in this package, at least as long as `Distributions` uses `PDMats`. The issue...

In my example above, the precision matrix `Q` is defined in terms of the unknown parameter `ρ`, so when applying any autodiff routine to the model, `Q` ends up being...

Zygote doesn't work either, it fails with a `Mutating arrays is not supported` error. Some of the linear algebra functions build up their results in-place. Traceback ```julia Mutating arrays is...

Yes I just ran the Turing model above after setting the AD backend to Zygote...sorry for not being clearer. However, I also get the same error with a hand-coded likelihood...

> It seems logabsdet is not defined for sparse matrices (and its LU decomposition)?! Weird that we'd be the first people to notice this, but that seems to be the...

Some progress on this: `logabsdet` will [now work](https://github.com/JuliaLang/julia/pull/40716) for sparse matrices (on Julia master, should be released in v1.7). Once the AD issues get sorted out, this issue can probably...