pdf for ChineseRestaurantProcess gives a stack overflow

[phipsgabler]

julia> ChineseRestaurantProcess(DirichletProcess(1.0), Int[])
ChineseRestaurantProcess{DirichletProcess{Float64},Array{Int64,1}}(
rpm: DirichletProcess{Float64}(1.0)
m: Int64[]
)

julia> pdf(ChineseRestaurantProcess(DirichletProcess(1.0), Int[]), 1)
ERROR: StackOverflowError:
Stacktrace:
 [1] pdf(::ChineseRestaurantProcess{DirichletProcess{Float64},Array{Int64,1}}, ::Int64) at /home/philipp/.julia/packages/Distributions/dTXqn/src/univariates.jl:333 (repeats 80000 times)

Appearently this is because Distributions contains a recursive definition of pdf for discrete univariate distributions that does not fall back on logpdf:

pdf(d::DiscreteUnivariateDistribution, x::Real) = isinteger(x) ? pdf(d, round(Int, x)) : zero(partype(d))

(logpdf works as it should.)

CC @trappmartin

[devmotion]

This already came up with other discrete distributions (see e.g. #1089). I really think one should change the definition in Distributions if possible.

[torfjelde]

This also caught me by surprise when implementing TransformedDistribution in Bijectors.jl. For multivariate distributions it's sufficient to implement logpdf because pdf falls back to it. But in the case of Univariate distributions, it's the reverse: https://github.com/JuliaStats/Distributions.jl/blob/09f2bbdf9faca686d72ff8d30d0299137d601842/src/univariates.jl#L342-L343

[trappmartin]

This is completely insane. I think we should open a PR on the Distributions.jl side.

[phipsgabler]

Done.

[trappmartin]

Thanks!

[devmotion]

JuliaStats/Distributions.jl#1171

[devmotion]

The PR was merged into Distributions and released as part of Distributions 0.24.

[devmotion]

As mentioned in the last comment, this is fixed in recent versions of Distributions.