David Widmann
David Widmann
> So we allow one (but not both) of the parameters in beta_inc to be zero but in that case the integrated density degenerate. Isn't the case where both `a`...
It seems there is some problem with the MPFR library on Windows: `beta(big(1.0e8), big(1 // 2))` returns `NaN` only on Windows. The problem did only show up now since previously...
The [documentation of `mpfr_beta`](https://www.mpfr.org/mpfr-current/mpfr.html#index-mpfr_005fbeta) says: > Note: the current code does not try to avoid internal overflow or underflow, and might use a huge internal precision in some cases. Maybe...
Test errors are unrelated and exist on master as well.
They don't, but I didn't want to commit type piracy and hence removed the error message overloads for arrays for the functions implemented in StatsFuns.
It shouldn't. I just tested it on a GPU, and the sigmoid function seems to be fine. The softplus function errored due to the use of `log1p` ("... called `log1p(x::Float32)`...
> Any hope StatFuns could ditch its Rmath dependency? I don't know the exact plans of the maintainers, I think the plan is to remove the dependency eventually at some...
BTW StatsFuns already depends on ChainRulesCore implicitly via SpecialFunctions, so it seems custom ChainRules-based adjoints could be added to StatsFuns without introducing any additional dependencies.
Yes, this issue was one motivation for moving the functions to LogExpFunctions 🙂
FYI recently I added the ChainRules definitions to LogExpFunctions.