I am generating correlation matrix by some new algorithm. Generated matrix is non positive semi-definite matrix.
I'm getting a few negative eigenvalues. The rest of eigenvalues are quite equal to the ideal matrix.
Can I use that non positive semi-definite matrix? If not, why?
If my estimated correlation matrix has all positive but complex value and imaginary terms are close to zero then is it possible?
A correlation matrix is really the covariance matrix of a bunch of variables which have been rescaled to have variance one.
But every population covariance matrix is positive semi-definite, and if we rule out weird cases (such as with missing data, or "numerical fuzz" turning a small eigenvalue to a negative one), so is every sample covariance matrix.
So if a matrix is supposed to be a correlation matrix, it should be positive semi-definite.
Note that the semi-definite is important here. In the bivariate case, take your two variables to be perfectly positively correlated and then the correlation matrix is $\pmatrix{1 & 1 \\ 1& 1}$ which has eigenvalues of $2$ and $0$: the zero eigenvalue means it is not positive definite.
Negative eigenvalues would imply that by the diagonalizing transformation the random vector would have negative variance in some components. Negative variances do never exist.
A correlation matrix is positive semi-definite, period. Numerics, however, might refuse to acknowledge that mathematical fact depending on how you arrive at the numeric representation of the correlation matrix.
The solution is to choose a representation of your matrix that cannot fail to be positive semi-definite by representing the matrix in a suitable decomposed form. I am not up to scratch, but there are things like LUD decompositions or square root forms that essentially are unable to represent anything but truly positive semi-definite matrices and which you can usually update incrementally similarly to how you would update the full matrix, possibly even easier.
