mvnormalcv takes 20 minutes when it should take under a second

[Mauricio Caceres]

Usually mvnormalcv executes in a fraction of a second but for some reason this specific setup it took 20 minutes (and in a fairly beefy server at that). Any ideas?

set rmsg on
mata
bound = (0.580, 0.970, 0.771, 0.914, 0.673, 1.068, 0.810, 0.592, 1.271)
means = J(1, 9, 0)
sigma = 0.088,
        0.104,
        0.023,
        0.070,
        0.046,
        0.136,
        0.100,
        0.025,
        0.074,
        0.245,
        -0.032,
        0.130,
        0.100,
        0.215,
        0.116,
        -0.015,
        0.143,
        0.155,
        -0.077,
        -0.026,
        -0.004,
        0.027,
        0.024,
        -0.173,
        0.218,
        0.060,
        0.137,
        0.103,
        -0.008,
        0.139,
        0.118,
        0.119,
        0.018,
        -0.037,
        0.094,
        0.297,
        0.106,
        -0.029,
        0.171,
        0.171,
        0.055,
        0.056,
        0.091,
        -0.011,
        0.420
mvnormalcv(-bound, bound, means, sigma)
end

[John Mullahy]

I can report from experience that dimensions beyond 7 or 8 (yours is 9) typically bring execution to a halt for me when using any of the mvnormal(...) functions. Running your code has done so on my not-so-beefy computer.

When you report that it usually executes in a fraction of a second are you referring to 9-dimension problems or to smaller ones?

Either way, have you tried reducing the number of quadrature points, using

mvnormalcvqp(L, U, M, V, q)

instead of

mvnormalcv(L, U, M, V)

[Mauricio Caceres]

John Mullahy Fraction of a second remark: yes, I meant smaller problems fed to mvnormalcv. However, I also fed this problem to R's mvtnorm and it executed fairly quickly, leading me to believe there was an issue here.

Yes, I also tried playing around with the number of curvature points and it was still slow. (Though I did not benchmark it.)