Sampling from MvNormal with integer arrays fails

[baggepinnen]

When an MvNormal is created using integer arrays, sampling from the distribution fails.

julia> using Distributions

julia> rand(MvNormal([1,1],[1,1]), 10)
ERROR: MethodError: no method matching randn(::Random._GLOBAL_RNG, ::Type{Int64})
Closest candidates are:
  randn(::AbstractRNG, ::Type{T}, ::Tuple{Vararg{Int64,N}} where N) where T at /home/fredrikb/julia-1.3.0-rc4/share/julia/stdlib/v1.3/Random/src/normal.jl:181
  randn(::AbstractRNG, ::Type{T}, ::Integer, ::Integer...) where T at /home/fredrikb/julia-1.3.0-rc4/share/julia/stdlib/v1.3/Random/src/normal.jl:184
  randn(::AbstractRNG) at /home/fredrikb/julia-1.3.0-rc4/share/julia/stdlib/v1.3/Random/src/normal.jl:39
  ...
Stacktrace:
 [1] randn!(::Random._GLOBAL_RNG, ::Array{Int64,2}) at /home/fredrikb/julia-1.3.0-rc4/share/julia/stdlib/v1.3/Random/src/normal.jl:173
 [2] _rand! at /home/fredrikb/.julia/packages/Distributions/2i7UF/src/multivariate/mvnormal.jl:269 [inlined]
 [3] rand at /home/fredrikb/.julia/packages/Distributions/2i7UF/src/multivariates.jl:77 [inlined]
 [4] rand(::MvNormal{Int64,PDMats.PDiagMat{Int64,Array{Int64,1}},Array{Int64,1}}, ::Int64) at /home/fredrikb/.julia/packages/Distributions/2i7UF/src/multivariates.jl:76
 [5] top-level scope at REPL[10]:1

(v1.3) pkg> st Distributions
    Status `~/.julia/environments/v1.3/Project.toml`
  [31c24e10] Distributions v0.21.5
  [1fd47b50] QuadGK v2.0.3
  [2913bbd2] StatsBase v0.32.0

julia> VERSION
v"1.3.0-rc4.1"

[matbesancon]

we could add a conversion method, PR welcome :)

[ben-schulz]

it looks like something similar was marked resolved way back in issue #498 -- but it would appear that change only addressed a type error on invocation of the constructor.

the code above fails because because:

dist = MvNormal([1,1],[1,1])
dist::MvNormal{Int32,PDMats.PDiagMat{Int32,Array{Int32,1}},Array{Int32,1}}

i.e., MvNormal successfully instantiates with element type <: Integer, where as an invocation to rand expects a float of some kind.

should the MvNormal constructor always promote the underlying element type to be usable with rand? that might make sense (seeing as it is a distribution), but it would seem to entail parameterizing MvNormal{T} by a type more restrictive than the current T<:Real

[JobJob]

Related - MvNormal with Array{Int} covariance also gives an unintuitive error:

julia> MvNormal([0, 0.0], [1 0; 0 1])
ERROR: MethodError: no method matching PDMats.PDMat{Int64,Array{Int64,2}}(::Int64, ::Array{Int64,2}, ::LinearAlgebra.Cholesky{Float64,Array{Float64,2}})

[mossr]

Note that the fix for this (which originates in PDMats.jl) was merged and is now part of PDMats v0.10.1

[matbesancon]

so should this be closed?

[mossr]

The fix is part of a PDMats patch release (v0.10.1), but a fresh install of Distributions grabs v0.10 (I just tested this with Julia v1.5)—so it's up to you how to proceed. I doubt putting an explicit [compat] requirement for PDMats v0.10.1 is what you'd want.

[mossr]

But my suggestion would be to—in fact—add the [compat] for "0.10.1" before closing (which would capture [0.10.1, 0.11.0)