Accurate similarity transformation derived from the discrete Lotka-Volterra system for bidiagonal singular values (Q2017988)

The discrete Lotka-Volterra (dLV) system describes the predator-prey evolution of \(2m-1\) different species \(u_k\), \(k=1,\dots,2m-1\). The time discretisation \(\delta^{(n)}\) may vary with time \(n\). A bidiagonal matrix \(B\in\mathbb{R}^{m\times m}\) defines the initial conditions for \(n=0\). The continuous system results as \(\delta^{(n)}\to0\). In this paper, the authors are interested in \(\delta^{(n)}\to\infty\). As shown by \textit{M. Iwasaki} and \textit{Y. Nakamura} [Inverse Probl. 18, No. 6, 1569--1578 (2002; Zbl 1021.35115)], the \(u_{2k-1}^{(n)}\) converge to the eigenvalues of \(B^TB\) as time \(n\to\infty\) while \(u_{2k}^{(n)}\to 0\). The iterative solution describes similarity transformations on \(B\) by application of the QD algorithm on the \(u_k^{(n)}\) values for the \((n,k)\) grid. This paper analyses the iteration in the simplified case where \(\delta^{(n)}\) is replaced by \(\infty\). It is shown that convergence still holds and that bounds for the forward relative rounding errors grow only linearly in \(m\).