MVG Multivariate Gaussian random number generator

Generates vectors from the multivariate normal distribution parameterized by specified mean vector a

Description

MVG is a multivariate Gaussian (normal) random number generator. A user can generate a vector from the multivariate normal distribution of any dimension by specifying a mean vector and symmetric positive-definite covariance matrix. A linear transformation based on the Cholesky decomposition of the covariance matrix is applied to a set of realizations from the distribution N(0,I). By applying the linear transformation to those samples, the output is a matrix whose columns are samples drawn from the distribution N(mu,Sigma) where mu is the specified mean vector and Sigma is an SPD covariance matrix. Type help mvg to learn more.

• MATLAB release: MATLAB 7.4 (R2007a)

[John D'Errico]

Why would you use the central limit theorem to generate approximate normal variates, when randn does so much more accurately, as well as doing so much more quickly? I'm sorry, but that is just a downright ridiculous thing to do.

Looking at the help, it is pretty good. There is a reference, there are readable descriptions of the parameters, etc. I found error checks for the errors one might expect to see.

One issue is the H1 line. There was none. The H1 line is the very FIRST line of comment in the help block. The H1 line is used by lookfor, as well in creating a contents file when necessary.

The funny thing is, the author has written a very serviceable H1 line, but left it as the second comment line, with a blank comment line preceding it. This is silly, since it fails to serve its proper utility.

There are other issues, more grievous ones, I am afraid. The code computes the cholesky factor of the covariance matrix, but for no good reason, since then it next computers the eigenvalue decomposition. Since either of these decompositions will serve to transform a set of independent normal variates into the desired structure, the code does extra work for no valid reason.

Finally, there are at least a few other multivariate normal deviate generation codes on the file exchange. I don't see anything that this improves upon. I've given it as high as a 2 rating only because the help was clearly written.

[John D'Errico]

Much better now. Problems I noted previously are repaired. I also did not see a truly n-d normal generator after another check, only the 2-d case. So this is more general.

[Jøger Hansegård]