Optimality conditions in terms of alternance: two approaches (Q462998)

The article of V.F. Demyanov and V.N. Malozemov is a valuable contribution to the area of nondifferentiable or nonsmooth calculus and, especially, optimization. This investigation takes place in a wide set-theoretical and analytic setting, and very elegantly generalizes and, as special cases, includes, basic and famous numerical approaches of continuous optimization, namely, gradient-type methods and Newton-type methods, while at the same time benefitting from superb relations to convex analysis, approximation theory, etc. This paper is very deep, well-structured and well-written. It stimulates important future research and application. In optimization theory, optimality conditions -- both of necessary and sufficient ones -- play a key role. Firstly, they permit a check of whether a point under investigation satisfies these conditions and, secondly, in case it does not, to find a ``better'' candidate solution. In the class of directionally differentiable functions, a necessary condition for an (unconstrained) minimizer needs the directional derivative to be nonnegative with respect to all directions. For special classes of directionally differentiable functions this condition becomes efficient; e.g., in the classes of convex and max-type functions, the necessary condition for a minimizer becomes an inclusion. The verification problem of this condition reduces to that of finding the element of some convex and compact set C which is closest to the origin. If the origin is not contained in C, we obtain the steepest descent direction and are able to construct a numerical algorithm. In the famous Chebychev polynomial approximation problem, the necessary optimality conditions are represented in the way of sign alternation of certain values. In this article, an extension of the alternance approach to a general optimization problem is presented. Two forms of the alternance condition (the so-called inside form and the outside form), both being equivalent, are carefully discussed. Sometimes, it can be more convenient to use the conditions in inclusion form, whereas in some other cases the condition in alternance form is advisable , just as in Chebychev approximation. Usually, algorithmic techniques which base on the inclusion condition are gradient-type, whereas methods employing the alternance form often are Newton-type. The authors express their hope that in some cases it will be possible to refine the existing toolbox of techniques of optimization with a new family of efficient methods. In this article, they only discuss unconstrained optimization programs in finite dimensions. In various further cases, a constrained optimization program can be reduced (e.g., by exact penalization techniques) to an unconstrained program, and further extensions can be imagined. The seven sections of this article are as follows: 1. Introduction, 2. Statement of the Problem, 3. Auxiliary Results, 4. The First (Inside) Approach, 5. The Second (Outside) Approach, 6. Examples and 7. Conclusions. In the future, further extensions and strong analytical results and numerical algorithms can be expected, initialized in the research community by the present research paper. These advances could motivate, foster and support important achievements in science, engineering, economics, decision making, finance and OR, in bioinformatics, medicine and healthcare, and, eventually, for improvements of living conditions of the people on earth. Each of those achievements will also serve for a commemoration of Professor Vladimir F. Demyanov whom the author of this report and the family of optimizers, of OR researchers and of mathematicians worldwide remembers with cordial thankfulness and with love.