It would be great to have multivariate_normal raising exception

instead of undefined behaviour as the NumPy doc says:

Note that the covariance matrix must be positive semidefinite (a.k.a. nonnegative-definite). Otherwise, the behavior of this method is undefined and backwards compatibility is not guaranteed.

For example, the following

from numpy import errstate
from numpy.random import RandomState
from numpy.linalg import eigh
from numpy import dot

rs = RandomState(3)
m = rs.randn(10)
G = rs.randn(10, 5)
K = dot(G, G.T)

with errstate(all='raise'):
    try:
        rs.multivariate_normal(m, K)
    except Exception as e:
        print("Exception raised!")
    else:
        print("Exception not raised!")

prints

bug.py:13: RuntimeWarning: covariance is not positive-semidefinite.
  rs.multivariate_normal(m, K)
Exception not raised!

As you know, K happens to not be positive-semidefinite in the above example because of floating point arithmetic. This can be easily fixed by adding a small value to its diagonal. If NumPy raised an Exception, I could perform this hack only if necessary. As it is now, I would need to check for positive-semidefiniteness beforehand, or add a small value to its diagonal before hand. Both solutions are not ideal in my opinion. Raising an exception would solve that.

[rgommers]

You said you wanted to use errstate. errstate is for floating point errors (as in IEEE 754). This is different - you basically want a check for whether the covariance matrix provided by a user is valid or not. The normal way to deal with those kinds of user parameter checks is: do a specific check, then raise a ValueError if needed.

[jenishah]

If no one is working, I would like to take this up.