Fenchel's duality theorem for nearly convex functions (Q946335)

The paper of R. I. Bot, S. M. Grad and G. Wanka is located in the mathematical theory of optimization which bases on convex analysis and is sometimes called generalized convexity. Indeed, it studies near(ly) convexity and concavity, it generalizes Fenchel's Duality Theorem for them, and it briefly studies close(ly) convexity and convexity for which, however, that generalization does not hold which is shown by a counterexample. By their investigation within optimization in finite dimensional Euclidean spaces, the authors not only contribute to a better analytical understanding and an extension of the frontiers of convex optimization, but they also implicitly prepare and justify the future use of numerical methods for that widened problem class. In fact, many so-called primal-dual methods in optimization perform a gradual diminishment of the duality gap until it vanishes. Referring to the difference function between a proper convex and and a proper concave function (the domains of both having elements of their relative interiors in common), Fenchel's Duality Theorem asserts that there is a strong duality between mininizing that difference function and maximizing the difference function given by the Fenchel conjugate of the second and the Fenchel conjugate of the first function. In a suitable setting of ``convexity'', the authors extend this result for nearly convex and a nearly concave functions, respectively. By definition, the property of being nearly convex comes from a relaxation in the definition of convexity (subaddive kind of inequality with convex coefficients multiplied) that is given by turning from the uniform (in these coefficients) version of the inequality to a one where just one such a coefficient value, strictly between 0 and 1, and its value complementary to 1 are sufficient to satisfy the inequality. This all is understood uniformly with respect to the domain. (In an analogous way, the authors define what a nearly convex set is.) The meaning of a nearly concave function follows canonically as near(ly) convexity of the function multiplied by \(-1\). As an example of a nearly convex function, a solution of the discontinuous Cauchy functional equation is given. The main theorem achieved in the paper concludes that the infimum value and the maximum value of the corresponding two problems, respectively, are the same (without a duality gap). Another theorem is a corollary coming from a special application. The reviewers mention that real-world interpretations, applications and generalizations of this paper could come from game theory, the theory of risk functionals, and could be attempted in the presence of contraints and discussed for infinite programming regularized, e.g., for advanced statistical learning and inverse problems.