No label defined (Q3976415)

The singular value decomposition (SVD) of an \(n\times n\) bidiagonal matrix \(B\) is studied by analyzing the flow of a Hamiltonian system which interpolates the SVD algorithm at integer times. The SVD generates a sequence \(B_ j\) of bidiagonal matrices and thus the algorithm can be formulated as a mapping from \(R^{2n-1}\) to \(R^{2n-1}\): the matrices generated by the SVD satisfy \(b_{j+1}=F(b_ j)\) where \(b_ j\) is a \((2n-1)\times 1\) vector comprised of the logarithms of the entries of \(B_ j\). The Jacobian matrix \(M(j,i)\) of the mapping \(F^{j-i}(b_ i)\) is studied since the properties of \(M(j,i)\) determine the growth of relative errors in the SVD algorithm. It is shown that: the eigenvalues of \(M(j,i)\) occur in reciprocal pairs \(\lambda\) and \(\lambda^{-1}\); furthermore the eigenvalues of \(M(i+1,i)\) are simple, lie on the unit circle and approach 1 for \(i\) sufficietly large; finally \(\| M(j,i)\|\) grows linearly in \((j-i)\). These facts are proved using the symplectic structure of the underlying Hamiltonian flow. The facts substantiate the observation that the SVD of bidiagonal matrix can be computed far more accurately than that of a general matrix. For the entire collection see [Zbl 0752.17022].