Andreas Noack

It's `Normal` that is the odd one out here. The element type of a distribution specifies the type of the parameters, not the type of the variates produced by `rand`....

The overhead I'm referring to is overhead for the programmer. The complexity of managing your methods grows a lot in the number of type parameters that you have to support....

The original design of Distributions was ended up being too strict as it only allowed parameters and variates to be either `Int` or `Float64`. To support automatic differentiation, we've moved...

I believe this issue belongs in StaticArrays. The inverse of the `Cholesky` here isn't symmetric ```julia julia> issymmetric(MvNormal(μ, Σ).Σ.mat) true julia> issymmetric(inv(MvNormal(μ, Σ).Σ.chol)) false ``` This is probably because it's...

> I think the main issue with the current implementation is that the Gamma sampler returns incorrect values: Why do you conclude that the values are incorrect? The examples seems...

Doh. I misread your test. Indeed, it doesn't look right. Would be good to figure out why that happens.

The problem is that the probability of observing a value smaller than the smallest (positive) `Float64` is pretty large. ```julia julia> cdf(Gamma(0.01, 1.0), floatmin(Float64)) 0.0008432274068068664 ``` so I'm not sure...

Looks like this carries over to the Beta ```julia julia> cdf(Beta(0.01, 1.01), floatmin(Float64)) 0.0008385787538914535 ``` so I'm not really sure what an be done.

It is probably worth mentioning that the Gamma cdf is a slightly complicated function and the Beta cdf is an extremely complicated function which would require a lot of work...

Two comments: 1. We should probably just have a diff rule for `cdf` that returns the `pdf` 2. The more tricky case that is usually hit when fitting models is...