logpdf(Binomial(0,0), x=0) returns NaN instead of 0 when x is Int32 or lower

[BenjamenSimon]

logpdf(Binomial(0, 0.), 0) # = 0.00

logpdf(Binomial(0, 0.), 0.) # = 0.00

logpdf(Binomial(0, 0.), Int32(0)) # = NaN

logpdf(Binomial(0, 0.), Int16(0)) # = NaN

logpdf(Binomial(0, 0.), Float32(0)) # = NaN

logpdf(Binomial(0, 0.), Float16(0)) # = NaN

[andreasnoack]

I believe this will be fixed by JuliaStats/StatsFuns.jl#113. Binomial(0, 0.) is a really odd distribution, though.

[devmotion]

This is also fixed by JuliaStats/StatsFuns.jl#125 (BTW would be great if you can find time for a review @andreasnoack - otherwise, if you are busy maybe someone else could review it?):

julia> logpdf(Binomial(0, 0.), 0) # = 0.00
0.0

julia> logpdf(Binomial(0, 0.), 0.) # = 0.00
0.0

julia> logpdf(Binomial(0, 0.), Int32(0)) # = NaN
0.0

julia> logpdf(Binomial(0, 0.), Int16(0)) # = NaN
0.0

julia> logpdf(Binomial(0, 0.), Float32(0)) # = NaN
0.0

julia> logpdf(Binomial(0, 0.), Float16(0)) # = NaN
0.0

[BenjamenSimon]

Thank you both. Do you have a timeline of when those will be implemented/fixed?

For context, its one term in the likelihood of a chain-binomial simulation. So it represents the probability that no events occur when there are no events that can occur (which is 1) which is a valid situation in the model of interest.

[devmotion]

As a temporary workaround you can install JuliaStats/StatsFuns.jl#125 manually by running

] add StatsFuns#dw/rmath_real+julia_logpdf

in the Julia REPL. Unfortunately, the PR is missing a more detailed review (should work fine though) and it is difficult to estimate when it will be approved, merged, and released.

[devmotion]

The issue will be fixed in StatsFuns 0.9.11 which will be available when JuliaRegistries/General#45672 is merged. Once it is available you should revert the manual installation suggested in the last comment (run ] free StatsFuns in the Julia REPL) and install the latest version of StatsFuns (e.g. by running ] up).