Why cov doesn't return a semi definite positive matrix ?

I have a matrix of features and based on that, I need to compute a semi definite positive matrix (covariance matrix) for testing purposes, so I naturally used "cov" but when I tested the semidefinite positiveness of the output, results were not satisfying. Is there another function that can do the job ? or may be another way to compute the desired covariance matrix.

Thank you!

[Mayssa]

For those who may be interested, this can solve the problem: https://www.mathworks.com/matlabcentral/fileexchange/42885-nearestspd

Answer 1

[Matt J]

The non-positive definiteness is probably due to floating point calculation errors. You cannot avoid this if your cov matrix is close to singular. Remove linearly dependent data from your covariance calculations. Or, just set eigenvalues below a certain threshold to zero.