An optimization problem for continuous submodular functions

TL;DR

This paper studies an optimization problem for entropy-like continuous submodular functions, providing bounds on minimal cost under geometric constraints and exploring uniqueness and open questions.

Contribution

It introduces a new optimization framework for entropy-like functions with geometric constraints, deriving tight bounds and analyzing uniqueness.

Findings

Derived a lower bound for minimal cost based on surface normals.

Proved the bound is tight for linear and certain convex/concave surfaces.

Showed the optimal solution may not be unique.

Abstract

Real continuous submodular functions, as a generalization of the corresponding discrete notion to the continuous domain, gained considerable attention recently. The analog notion for entropy functions requires additional properties: a real function defined on the non-negative orthant of $\mathbb R^n$ is entropy-like (EL) if it is submodular, takes zero at zero, non-decreasing, and has the Diminishing Returns property. Motivated by problems concerning the Shannon complexity of multipartite secret sharing, a special case of the following general optimization problem is considered: find the minimal cost of those EL functions which satisfy certain constraints. In our special case the cost of an EL function is the maximal value of the $n$ partial derivatives at zero. Another possibility could be the supremum of the function range. The constraints are specified by a smooth bounded surface $S$ cutting off a downward closed subset. An EL function is feasible if at the internal points of $S$ the left and right partial derivatives of the function differ by at least one. A general lower bound for the minimal cost is given in terms of the normals of the surface $S$. The bound is tight when $S$ is linear. In the two-dimensional case the same bound is tight for convex or concave $S$. It is shown that the optimal EL function is not necessarily unique. The paper concludes with several open problems.