Using cross-product matrices to compute the SVD (Q2502231)

Computing singular values of a real matrix \(A\) as the square roots of the eigenvalues of \(A^TA\) generally produces satisfactory accuracy only for the largest singular values. The author shows that, if the \(k\) smallest singular values are sufficiently well separated from all the other singular values, then these \(k\) smallest singular values may be computed, with accuracy of the order of the unit round, as the square roots of the eigenvalues of \(V^TA^TAV\), where the \(k\) columns of \(V\) are computed orthonormal eigenvectors of \(A^TA\) corresponding to its \(k\) smallest eigenvalues.