Use un-pivoted Cholesky triangle to sample from sparse MvNormalCanon (#…

…1218)

* Use unpivoted Cholesky triangle to generate sparse MvNormalCanon samples

* Remove stray `end`

* Check that PDSparseMat is defined in tests

src/multivariate/mvnormalcanon.jl

@@ -177,9 +177,9 @@ unwhiten_winv!(J::AbstractPDMat, x::AbstractVecOrMat) = unwhiten!(inv(J), x)
 unwhiten_winv!(J::PDiagMat, x::AbstractVecOrMat) = whiten!(J, x)
 unwhiten_winv!(J::ScalMat, x::AbstractVecOrMat) = whiten!(J, x)
 if isdefined(PDMats, :PDSparseMat)
-    unwhiten_winv!(J::PDSparseMat, x::AbstractVecOrMat) = x[:] = J.chol.U \ x
+    unwhiten_winv!(J::PDSparseMat, x::AbstractVecOrMat) = x[:] = J.chol.PtL' \ x
 end
-    
+
 _rand!(rng::AbstractRNG, d::MvNormalCanon, x::AbstractVector) =
     add!(unwhiten_winv!(d.J, randn!(rng,x)), d.μ)
 _rand!(rng::AbstractRNG, d::MvNormalCanon, x::AbstractMatrix) =

test/mvnormal.jl

@@ -1,6 +1,9 @@
 # Tests on Multivariate Normal distributions
 
 import PDMats: ScalMat, PDiagMat, PDMat
+if isdefined(PDMats, :PDSparseMat)
+    import PDMats: PDSparseMat
+end
 
 using Distributions
 using LinearAlgebra, Random, Test
@@ -178,28 +181,51 @@ end
 end
 
 ##### Random sampling from MvNormalCanon with sparse precision matrix
-@testset "Sparse MvNormalCanon random sampling" begin
-    # Random samples from MvNormalCanon and MvNormal diverge as
-    # 1) Dimension of cov/prec matrix increases (n)
-    # 2) Determinant of J increases
-    # ...hence, the relative tolerance for testing their equality
-    n = 10
-    k = 0.1
-    rtol = n*k^2
-    seed = 1234
-    J = sprandn(n, n, 0.25) * k
-    J = J'J + I
-    Σ = inv(Matrix(J))
-    J = PDSparseMat(J)
-    μ = zeros(n)
-    d = MvNormalCanon(μ, J*μ, J)
-    d1 = MvNormal(μ, PDMat(Symmetric(Σ)))
-    r = rand(MersenneTwister(seed), d)
-    r1 = rand(MersenneTwister(seed), d1)
-    @test all(isapprox.(r, r1, rtol=rtol))
-    @test mean(abs2.(r .- r1)) < rtol
+if isdefined(PDMats, :PDSparseMat)
+    @testset "Sparse MvNormalCanon random sampling" begin
+        n = 20
+        nsamp = 100_000
+        Random.seed!(1234)
+
+        J = sprandn(n, n, 0.25)
+        J = J'J + I
+        Σ = inv(Matrix(J))
+        J = PDSparseMat(J)
+        μ = zeros(n)
+
+        d_prec_sparse = MvNormalCanon(μ, J*μ, J)
+        d_prec_dense = MvNormalCanon(μ, J*μ, PDMat(Matrix(J)))
+        d_cov_dense = MvNormal(μ, PDMat(Symmetric(Σ)))
+
+        x_prec_sparse = rand(d_prec_sparse, nsamp)
+        x_prec_dense = rand(d_prec_dense, nsamp)
+        x_cov_dense = rand(d_cov_dense, nsamp)
+
+        dists = [d_prec_sparse, d_prec_dense, d_cov_dense]
+        samples = [x_prec_sparse, x_prec_dense, x_cov_dense]
+        tol = 1e-16
+        se = sqrt.(diag(Σ) ./ nsamp)
+        #=
+        The cholesky decomposition of sparse matrices is performed by `SuiteSparse.CHOLMOD`,
+        which returns a different decomposition than the `Base.LinearAlgebra` function (which uses
+        LAPACK). These different Cholesky routines produce different factorizations (since the
+        Cholesky factorization is not in general unique).  As a result, the random samples from
+        an `MvNormalCanon` distribution with a sparse precision matrix are not in general
+        identical to those from an `MvNormalCanon` or `MvNormal`, even if the seeds are
+        identical.  As a result, these tests only check for approximate statistical equality,
+        rather than strict numerical equality of the samples.
+        =#
+        for i in 1:3, j in 1:3
+            @test all(abs.(mean(samples[i]) .- μ) .< 2se)
+            loglik_ii = [logpdf(dists[i], samples[i][:, k]) for k in 1:100_000]
+            loglik_ji = [logpdf(dists[j], samples[i][:, k]) for k in 1:100_000]
+            # test average likelihood ratio between distribution i and sample j are small
+            @test mean((loglik_ii .- loglik_ji).^2) < tol
+        end
+    end
 end
 
+
 ##### MLE
 
 # a slow but safe way to implement MLE for verification