Use Cholesky decomposition for correlated_values

[ces42]

Trying to do by hand exactly what correlated_values(nom_values, covariance_mat) does I thought of a different approach than the one used in the current implementation:

n = len(nom_values)
L = cholesky(covariance_mat)
return nom_values + dot(L, uarray(zeros(n), ones(n))

This essentially does the same as correlated_values (except for the tags and it returns a array instead of a tuple). It should be faster than diagonalizing the covariance matrix and some quick tests suggest that it might be numerically more precise:
For nom_values = ones(10) and covariance_mat = scipy.linalg.hilbert(10) the reconstructed covariance matrix using covariance_matrix only differs in 15 using cholesky while the current implementation differs in 78 entries.

Is there a reason why the current implementation uses diagonalization instead of cholesky? (If there is no such reason I would be happy to write a pull request)

P.S. I really like this package and it's saving me a ton of time in my lab course :)

[lebigot]

That's an interesting point: thanks for sharing!

I feel that what you propose is the better way, indeed. A pull request would be much appreciated.

I am glad that you find this library useful!