error using multivariate Gaussian number generator (mvnrnd)

Hello,

I'm trying to generate N points in a n dimension with a multivariate distribution.

I'm using the "mvnrnd" matlab funtion, however i keep getting the following error:

Error using mvnrnd (line 110) SIGMA must be a symmetric positive semi-definite matrix.

From my understanding, if the eigen values of the covariance matrix are all >= 0, then sigma is indeed positive semi-definite.

The problem is I'm getting this error even when the eigen values are all positive, and I don't know why.

I've uploaded a .mat with a mean and covariance matrix with which i get this error even though the eigen values are >0:

https://dl.dropbox.com/u/15692375/maltab_mvnrnd_example.mat

Am i overlooking something? Does anyone have idea what i'm doing wrong?

Thanks, Alex

Answer 1

[Peter Perkins]

In this case, it's not the positive semi-definiteness, but the symmetry. You have relative differences that are on the order of single precision:

>> (Pp - Pp')./abs(Pp)
ans =
            0   1.7455e-07   9.2053e-07
  -1.7455e-07            0   2.2278e-10
  -9.2053e-07  -2.2278e-10            0