`cov(d::MultivariateMixture)` assumes overly specific typing on component covariance

[btmit]

The current cov method for MixtureModel assumes that the cov method for each component returns a type that has strides is defined for it. This is very specific and breaks in the following simple case.

MWE:

using Distributions, LinearAlgebra

struct Gaussian{T,M<:AbstractMatrix{T},V<:AbstractVector{T}} <: AbstractMvNormal
    μ::V
    Σ::Symmetric{T,M}
end

Distributions.mean(g::Gaussian) = g.μ
Distributions.cov(g::Gaussian) = g.Σ

Base.length(g::Gaussian) = length(mean(g))
Base.eltype(::Gaussian{T}) where {T} = T

g1 = Gaussian(rand(3), Symmetric(rand(3,3)))
g2 = Gaussian(rand(3), Symmetric(rand(3,3)))

gm = MixtureModel([g1, g2])

cov(gm)

ERROR: MethodError: no method matching strides(::Symmetric{Float64, Matrix{Float64}})

The issue is that this cov(d::MultivariateMixture) calls directly into BLAS.axpy!(pi, cov(c), V) without any sort of checks on cov(c). It seems there is a layer missing here.

[devmotion]

This is fixed on the master branch (by #1968):

...
julia> cov(gm)
3×3 Matrix{Float64}:
 0.634358  0.406286  0.414077
 0.406286  0.607326  0.658475
 0.414077  0.658475  0.532549

[btmit]

Apologies. Didn't see this issue because it was closed.