Infinite (semi-infinite) problems to characterize the optimality of nonlinear optimization problems (Q2469580)

The paper of M. Soleimani-Damaneh is located on the interface and in a common and unifying framework between infinite programming, semi-infinite programming and with a continuous view at finitely constrained optimization. It provides an important theoretical contribution to this common emerging area which is of great importance for very new and modern applications. Besides of classical, i.e., single- (scalar-) valued programming, the important field on multi- (vector-) valued optimization is treated in this article as well. Indeed, after that calculus of variations and optimal control theory widely represented infinite programming (IP) for many centuries and became recently reinterpreted as optimization of (ordinary, partial or stochastic) differential equations, there is a fast growing variety of infinite programming problems now. The latter ones are characterized by the usually infinite dimensionl space underlying and by an infinite number of inequality constraints permitted. Differently from calculus of variations and control, those new IP problems need not be (but may be) defined on Lebesgue or Sobolev and need not (but may) reveal target function(al)s of integral form. In fact, the increasing research done in the theory of inverse problems with its refined regularization techniques developed encourage and allow for more generalized views onto classical programming problems. As a matter of fact, these techniques can serve for a better ``control'' of the increased problem complexity which would otherwise cause instabilities. In this way, we may expect IP problems to become more and more posed, studied and successfully solved. One example is machine learning with an infinite number of kernels, herewith offering a larger wealth of embeddings of not linearly separable class structures into spaces where linear separation is performed. (This approach to IP together with numerous approaches to SIP (semi-infinite programming) were proposed and conducted by the authors of this report with their colleagues.) Considering a generic nonlinear pogramming problem (NLPP), this paper provides a family of linear infinite problems (LIPs) or linear semi-infinite problems (LSIPs), and establishes the connection between optimality in the respective NLPP and that in the provided LIPs and LSIPs, respectively. Whenever the Banach space is finite dimensional, then ``S'' (for ``semi'') enters the abbreviated problem name. The author addresses a Banach space setting, focusing on an open set in that space. For NLPP, there, we presumes a locally Lipschitzian setting of objective function and finitely many inequality contraints, which enables the use of nonsmooth optimization, especially, of Clarke's generalized gradients. Some results are established asserting that an NLPP is optimal if it is optimal to an introduced LIP (LSIP), and vice versa. Provided a corresponding version of the famous Mangasarian-Fromovitz constraint qualification (MFCQ), (local) optimality of the NLPP implies optimality for the LIP that is a linearization of NLPP with the infinite number of elements of the Cartesian product of NLPP's inequality constraint functions representing the inequality constraints in LIP. For the reverse implication, a modified LIP (with for NLPP inactive constraints deleted) is referred to and convexity of the defining functions of NLPP assumed in addition. A similar pair of results is obtained by the author for a generic multi-objective NLPP. For this problem, and for the corresponding multi-objective LIPs he refers to weak Pareto optimality. In this well written, structured and proved paper, the treatment of an NLPP is the main motivation; the L(S)IPs inherited their forms from the NLPP via nonsmooth calculus. Of course, L(S)IPs can be much more general and even nonlinear (S)IPs (N(S)IPs), such as quadratic ones, can and will occur easily and often. Nevertheless, this article is an important step in a line of present and future ones which are initiated in these days and will have a great help for many applications in science, engineering, economy and society.