David Widmann

> Should we memoize `inv(σ)` since there's a good chance it'll be frequently used? No, I don't think it's worth it. In my experience it just causes problems, eg with...

No, `eltype` should (and will) not match the parameters in general. If one wants to e.g. evaluate just the log density it is irrelevant what the `eltype` is - one...

I'm very certain memoization just causes problems here and is not worth it. As an example in a different package, similar issues occurred with memoization of inverses in PDMats and...

> Whatever is wrong with the discrete CDF functions, which are failing the tests; if you know what I'm doing wrong I'm all ears. I don't know which error you...

> I think some things should be designed a bit differently. E.g., it is completely fine to translate `Distribution{

I hope you don't take this personal but I don't think we can merge the PR in its current state. In my opinion it is difficult to review and, in...

> If you'd like, I can break this into two PRs. The first can add a UnivariateAffine which is just LocationScale with a possibly negative scale; this would replace LocationScale....

Sure, it's seemingly simple changes and of course it could be done in fewer PRs - but Distributions has > 300 direct dependents and > 700 indirect ones and hence...

I still wonder why/if the definition of `eltype` is needed (https://github.com/JuliaStats/Distributions.jl/pull/1090/files#r467471768). In general, as far as I remember, the general policy is that there are no guarantees that `eltype(d) ===...

> The decision that eltype(dist) should mean the type of parameters seems strange to me. The parameters can also be of different types. I am very unsure if this is...