David Widmann

As discussed in https://github.com/JuliaStats/StatsFuns.jl/pull/114, we should not add a circular dependency on Distributions to StatsFuns. It's seems it is not actually needed and you could just call `normcdf` etc. I...

> I can replace the Distributions.jl call to Normal() with `normcdf(x) = 0.5*erfc(-x*invsqrt2)` That's precisely what I was referring to since `normcdf` is already defined in StatsFuns (https://github.com/JuliaStats/StatsFuns.jl/blob/36e48bbbac12355906438ce1993736757a0c6568/src/distrs/norm.jl#L45).

Distributions only wants and implements `quantile(::UnivariateDistribution, ::Real)`, and similarly only `cquantile(::UnivariateDistribution, ::Real)` is of interest there. But methods such as `quantile(::AbstractArray, p)` (with all its keyword arguments) do not belong...

I tend to think in general it would be better to move things from Statistics to StatsBase (or some other package). For posterity a copy of the discussion on Slack:...

Caused by https://github.com/JuliaDiff/ChainRules.jl/pull/559 it seems (similar to issues in Distributions, DistributionsAD, NNlib etc.). Should be fixed by the new ChainRules release (bugfix was https://github.com/JuliaDiff/ChainRules.jl/pull/562).

This seems to be yet another AD issue with PDMats. Maybe about time to add CR to that repo.

Regardless, it might still be useful and possibly more efficient to add a CR definition for `kl_divergence` directly.

All of them are PDMats issues, it seems.

The PDMats issues (https://github.com/JuliaStats/PDMats.jl/issues/159) are not fixed yet.

`mapreduce` with multiple arrays just computes the resulting array up-front: https://github.com/JuliaLang/julia/blob/d16f4806e9389dbc92c463efc5b96f67a7aebf22/base/reducedim.jl#L324-L325 (added in https://github.com/JuliaLang/julia/pull/31532) I think a better approach (that seems to break AD with basically all backends though according...