Closed cgranade closed 7 years ago
I have run into this problem, too.
I seem to recall that SVD algorithms are generally more stable than eigendecomposition algorithms, but I can't find a solid source, just things like the last page of these notes. I don't know if it matters in this context.
When covariance matrices are dominated by uncertainty in a low-dimensional subspace,
scipy.linalg.sqrtm
can sometimes report infinite error even ifn_ess
is large. I suspect that this can be solved by replacingsqrtm
with a eigendecomposition that uses the hermicity of the covariance matrix, but this should be tested for accuracy.