Moritz Schauer

I took ˋ⊕ˋ as symbol (for the independent sum of random variables) in https://github.com/mschauer/GaussianDistributions.jl/blob/master/src/GaussianDistributions.jl .

How can we set it up in a way that https://github.com/TuringLang/DistributionsAD.jl works and we don't need to define it for all `logpdf`s?

> I assume also that in many cases this general definition based on gradlogpdf would be suboptimal since it does not share computations and potentially recomputes many partial results. This...

Even non-API function can need docstrings (including the information that the function is not part of the API protected by SEMVER.)

Because `gradlogpdf` would be required to call an autodiff package to do its work in most cases, which changes the order of dependency.

At least for the Gaussian distribution, I would like to *not* separate the one and the `n`-dimensional case into different objects and hard-code various implementation details (isotropic or not). I'd...

My code to simulate linear stochastic processes is indifferent about the underlying type say as long as `randn(T)` gives a Gaussian increment. In general, if I can write `randn(T)` and...

One would really focus on the case `σ*randn(T) + μ`, which for complex values makes the mixing pseudo-covariance zero.

Static arrays can be supported to a large extend without depending on the StaticArray package I suppose.

The PDMats dependence could be replaced with an `AbstractPDMats`, at the cost of users of MvNormals providing actual factorizations for the covariance matrices themselves, which are then required to define...