David Widmann

Sure, I understand, I was just referring to the fact that `MvNormalCanon` is the distribution type in Distributions for multivariate normal distributions that are parameterized by precision matrices.

> The issue is that even with a custom adjoint defined for the Cholesky decomposition, the rrule still has to calculate the factorization--which, as long as it is done via...

Ah OK, I think I understand now. I'm a bit surprised that ForwardDiff doesn't support `cholesky` of sparse matrices but I've never checked. Did you try to use Zygote instead...

Sorry, it's not clear to me what exactly you tested here. Is it the Turing model above? That might still require some additional work but I was just referring to...

> For what it's worth, the pullback and pushforward of cholesky with a sparse argument appears to work: That's exactly what I assumed and why I thought that for Zygote...

I am wondering if the type parameter `N` can be dropped and the batch computation issue solved in a different way, e.g. by wrapping inputs similar to ColVecs and RowVecs...

IMO a macro seems to complicated and leads to un-julian syntax. I also don't think it is necessarily easier for users to figure out how to write the macro than...

> Sidenote: I think I realized a way we can avoid this code-duplication for making things work when we look at "batches" (i.e. define the (::Bijector{0})(x::AbstractArray{

> logabsdetjac(b::Softplus, x::Real) = x - log(1 + exp(x)) # I THINK this is right, haven't written it down Assuming it's correct (haven't thought about it), one should probably implement...

IMO the broadcasting machinery is quite heavy and all sorts of tricks and hacks are used in AD backends to (try to) deal with its complexity, so I try to...