David Widmann

I made a PR for AbstractFFTs: https://github.com/JuliaMath/AbstractFFTs.jl/pull/58

> While it's not too hard to contrive a Hermitian matrix with exactly repeating eigenvalues (e.g. a diagonal matrix), I've had no luck so far coming up with a non-Hermitian...

My intuition is that we would want `getproperty(x::Tangent{

As a side remark: As noted above, everything works if one uses the NaN-safe setting (https://juliadiff.org/ForwardDiff.jl/dev/user/advanced/#Fixing-NaN/Inf-Issues). I think one should benchmark if and to what extent performance is affected by...

This seems suprising since the PR does not touch `rand`, I would assume removing rules for `rand!` should not affect something like `sum(rand(args...))`? Edit: Ah, probably `rand(Float32, 1, 1)` calls...

Support for complex-valued functions with real arguments was added in https://github.com/JuliaDiff/ForwardDiff.jl/pull/583: ```julia julia> using ForwardDiff julia> ForwardDiff.derivative(x->exp(im*x), 1.0) -0.8414709848078965 + 0.5403023058681398im ```

I agree, these are points we could/should probably work on. Regarding `display`: A simple alternative to the current approach would be to not perform any statistical evaluations when displaying a...

I thought AxisKeys or NamedDims are the most common replacements for and improvements over AxisArrays these days but I didn't keep track of recent developments. There are many different approaches...

@xukai92 is that correct? BTW a bit confusing that there's both `log_density` and `lp` - are they guaranteed to be equal for every sample?

> I agree. I believe long time ago Turing simply extracted log_density as lp which avoids this issue. What do you think the best way to make it consistent between...