Numerical precision of primal not respected for SpecialFunctions.jl

[torfjelde]

Several methods from SpecialFunctions.jl are not respecting the numerical precision of the primal:

julia> using ChainRules, SpecialFunctions

julia> y, ȳ = ChainRules.rrule(SpecialFunctions.erf, 1.0f0)
(0.8427008f0, SpecialFunctions.var"#erf_pullback#12"{Float32}(1.0f0))

julia> ȳ(1.0f0)[2]
0.41510750774498784

I'm guessing this is because the rules were originally copy-pasted from DiffRules.jl which suffers from the same issue: JuliaDiff/DiffRules.jl#55.

I think I'm currently of the opinion that the best way to fix this is to copy-paste the constants from e.g. StatsFuns.jl to remove most of the promotions, and then I guess we'll have to use oftype for the remainders.

Is there a better solution?

Btw, I know this related to other issues, e.g. #232, but to me it seems like this particular case is non-ambiguous and can be easily fixed so might as well do it asap:)

[devmotion]

The differentials were moved to SpecialFunctions (JuliaMath/SpecialFunctions.jl#238) and are only loaded from ChainRules if an old version of SpecialFunctions without them is used. So all problems have to be fixed there as well.

[devmotion]

I think I'm currently of the opinion that the best way to fix this is to copy-paste the constants from e.g. StatsFuns.jl to remove most of the promotions

Instead of copy-pasting the definitions one could just use oftype (as you suggested) or move the constants to a separate package (MathConstants.jl? StatsConstants.jl?). BTW log-related constants live in LogExpFunctions now and are only reexported by StatsFuns.

[torfjelde]

The differentials were moved to SpecialFunctions

Ah, my bad! Should I just close this and open one over there then? Though I guess the rules also needs to be changed here.

Instead of copy-pasting the definitions one could just use oftype (as you suggested) or move the constants to a separate package (MathConstants.jl? StatsConstants.jl?).

I was sort of surprised to find that these were in fact not all in one package, so I would be cool with that. Would be suuuper-lightweight too. Initially I thought "eh, it's dependent on the stats packages actually using this", but I guess we could just start using it AD-frameworks immediately.

BTW log-related constants live in LogExpFunctions now and are only reexported by StatsFuns.

Cool 👍 But StatsFuns also has its own constants, etc. so it's a bit spread around.

[mzgubic]

We are removing the SpecialFunctions rules in this package, have you transferred the issue?