scientific article; zbMATH DE number 3979878

From MaRDI portal
Publication:3744775





[For the entire collection see Zbl 0595.00017.] The paper overviews the theory necessary to construct optimization algorithms based on convex analysis, namely on the use of approximate subdifferentials. Let \(f: {\mathbb{R}}^ n\to {\mathbb{R}}\) be a convex function. Given \(x\in {\mathbb{R}}^ n\) (the current approximation to a minimizer of f), the problem is to find \(h\in {\mathbb{R}}^ n\), giving the next approximation \(x+h\). The tools that are used from convex analysis are the approximate directional derivative \(f'_{\epsilon}(x;p)\) and its associated approximate subdifferential \(\partial_{\epsilon}f(x)\) [for the definitions, see \textit{J. B. Hiriart-Urruty}, Nonlinear Anal., Theory Methods Appl. 6, 1309-1326 (1982; Zbl 0536.26007)]. Some information concerning these notions is given at the beginning of the paper. Next, three possible approaches to the problem of defining h are described. The first ot them consists in finding the so-called steepest \(\epsilon\)-descent directions and was given first by \textit{D. P. Bertsekas} and \textit{S. K. Mitter} [SIAM J. Control 11, 637-652 (1973; Zbl 0243.49012)]. In the second approach, a certain second-order development of f is used to find the so-called \(\sigma\)-Newton directions [these results are described in detail in: the author and \textit{J. Zowe}, Appl. Math. Optimization 10, 51-68 (1983; Zbl 0527.65042)]. The third approach makes use of the directional derivative of \(f'_{\epsilon}(\cdot;p)\); concerning this approach, many fundamental questions are still open. Finally, because the approximate subdifferential is not a computable set, a section is devoted to implementable approximations of it.











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 Q3744775)