do not convert covariance/precision matrices of AbstractMvNormal subtypes to dense Matrix (#1373)

Author: st--

src/multivariate/mvnormal.jl

@@ -106,14 +106,10 @@ function kldivergence(p::AbstractMvNormal, q::AbstractMvNormal)
     # This is the generic implementation for AbstractMvNormal, you might need to specialize for your type
     length(p) == length(q) ||
         throw(DimensionMismatch("Distributions p and q have different dimensions $(length(p)) and $(length(q))"))
-    # logdetcov is used separately from _cov for any potential optimization done there
-    return (tr(_cov(q) \ _cov(p)) + sqmahal(q, mean(p)) - length(p) + logdetcov(q) - logdetcov(p)) / 2
+    # logdetcov is used for any potential optimization done there
+    return (tr(cov(q) \ cov(p)) + sqmahal(q, mean(p)) - length(p) + logdetcov(q) - logdetcov(p)) / 2
 end
 
-# This is a workaround to take advantage of the PDMats objects for MvNormal and avoid copies as Matrix
-# TODO: Remove this once `cov(::MvNormal)` returns the PDMats object
-_cov(d::AbstractMvNormal) = cov(d)
-
 """
     invcov(d::AbstractMvNormal)
 
@@ -256,10 +252,9 @@ params(d::MvNormal) = (d.μ, d.Σ)
 @inline partype(d::MvNormal{T}) where {T<:Real} = T
 
 var(d::MvNormal) = diag(d.Σ)
-cov(d::MvNormal) = Matrix(d.Σ)
-_cov(d::MvNormal) = d.Σ
+cov(d::MvNormal) = d.Σ
 
-invcov(d::MvNormal) = Matrix(inv(d.Σ))
+invcov(d::MvNormal) = inv(d.Σ)
 logdetcov(d::MvNormal) = logdet(d.Σ)
 
 ### Evaluation

src/multivariate/mvnormalcanon.jl

@@ -157,8 +157,8 @@ params(d::MvNormalCanon) = (d.μ, d.h, d.J)
 Base.eltype(::Type{<:MvNormalCanon{T}}) where {T} = T
 
 var(d::MvNormalCanon) = diag(inv(d.J))
-cov(d::MvNormalCanon) = Matrix(inv(d.J))
-invcov(d::MvNormalCanon) = Matrix(d.J)
+cov(d::MvNormalCanon) = inv(d.J)
+invcov(d::MvNormalCanon) = d.J
 logdetcov(d::MvNormalCanon) = -logdet(d.J)