On Mon, Nov 24, 2008 at 03:28:11PM -0500, Nathan Bell wrote: > Try QR first. It's reasonably stable and SVD is considerably more expensive. +1. SVD is the sledge hammer of numeric matrix factorisation. (In other cases, options may involve 'np.linalg.eigh(np.dot(A.T, A))', or other trick to avoid the costly SVD). Gaël