A characterization of simultaneous optimization, majorization, and (bi)submodular polyhedra

TL;DR

This paper explores the conditions under which solutions that simultaneously optimize entire classes of functions exist, linking these solutions to (bi)submodular polyhedra and providing new characterizations relevant to resource allocation and other applications.

Contribution

It introduces a generalized majorization concept and characterizes feasible sets with simultaneous optimization solutions, extending classical polyhedral characterizations.

Findings

Characterization of feasible sets with simultaneous optimization solutions

Extension of classical base and bisubmodular polyhedra characterizations

Theoretical explanation for the rarity of such solutions in practice

Abstract

Motivated by resource allocation problems (RAPs) in power management applications, we investigate solutions to optimization problems that simultaneously minimize an entire class of objective functions. It is straightforward to show empirically that such solutions do not exist for most optimization problems. However, little is known on why this is the case and whether a characterization exists of problems that do have such solutions. In this article, we answer these questions by linking the existence of solutions that simultaneously optimize the class of Schur-convex functions, called least majorized elements, to (bi)submodular functions and the corresponding polyhedra. For this, we introduce a generalization of majorization and least majorized elements, called $(a,b)$-majorization and least $(a,b)$-majorized elements, and characterize the feasible sets of problems that have such elements in terms of these polyhedra. Hereby, we also obtain new characterizations of base and bisubmodular polyhedra that extend classical characterizations of these sets in terms of optimal greedy algorithms for linear optimization from the 1970s. We discuss the implications of our results for RAPs in power management applications and use the results to derive a new characterization of convex cooperative games and new properties of optimal estimators of specific regularized regression problems. In general, our results highlight the combinatorial nature of simultaneously optimizing solutions and, at the same time, provide a theoretical explanation for the observation that such solutions generally do not exist.