scientific article; zbMATH DE number 4106165

From MaRDI portal
Publication:3830441





The convergence properties of descent methods without exact line searches for unconstrained optimization are considered. Under assumptions that the objective function f(x) is continuously differentiable and that \(g^{(k)T}s^{(k)}\geq \epsilon \| g^{(k)}\| \cdot \| s^{(k)}\|,\) \(\epsilon\in (0,1]\) holds for all sufficiently large k, where \(g^{(k)}\) is the gradient of f(x) at iterate \(x^{(k)}\), and \(s^{(k)}\) a descent direction generated by a descent method, it is proved that if the sequence \(\{x^{(k)}\}\) generated by the method is not terminated at a point \(x^{(1)}\), at which \(g(x^{(1)})=0\), then either \(f(x^{(k)})\to -\infty\) or \(\lim_{\overline{k\to \infty}}\| g^{(k)}\| =0.\) In the convergence case, when f(x) is twice continuously differentiable in a neighborhood of a minimum \(x^*\) and \(G(x^*)\), the second derivative matrix, is positive definite, then the convergence rate of the descent method is at least linear iff \(g^{(k)T}s^{(k)}\geq \epsilon \| g^{(k)}\| \| s^{(k)}\|\) holds for all sufficiently large k. In particular, if f(x) is a positive definite quadratic function and the descent method is a quasi-Newton method, especially the DFP or BFGS method, the convergence rate is superlinear. A further analysis shows that the DFP method with appropriate modifying is also superlinearly convergent for general differentiable convex functions.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3830441)