Say I've got a mxn matrix that I'd like to factor into an mxk and a kxn matrix. The factorization doesn't have to be exact; I'd just like to minimize some measure of error.

--Take the SVD, then zero out the smallest singular values until only the k largest are left.