Density Decomposition in Dual-Modular Optimization: Markets, Fairness, and Contracts

TL;DR

This paper introduces a unified framework for dual-modular optimization problems, revealing structural insights into fairness, market equilibria, and contract strategies through density decomposition, with broad applications and convergence guarantees.

Contribution

It develops a novel density decomposition concept for dual-modular instances, connecting fairness, market equilibrium, and contract design, and provides algorithms with convergence guarantees.

Findings

Density decomposition characterizes market equilibria and fair allocations.

Convex programs with various divergences recover the density decomposition.

The framework determines the structure and critical values in combinatorial contracts.

Abstract

We study a unified framework for optimization problems defined on dual-modular instances, where the input comprises a finite ground set $V$ and two set functions: a monotone supermodular reward function $\f$ and a strictly monotone submodular cost function $\g$. This abstraction captures and generalizes classical models in economics and combinatorial optimization, including submodular utility allocation (SUA) markets and combinatorial contracts. At the core of our framework is the notion of density decomposition, which extends classical results to the dual-modular setting and uncovers structural insights into fairness and optimality. We show that the density decomposition yields a canonical vector of reward-to-cost ratios (densities) that simultaneously characterizes market equilibria, fair allocations -- via both lexicographic optimality and local maximin conditions -- and best-response strategies in contract design. Our main result proves the equivalence of these fairness notions and guarantees the existence of allocations that realize the decomposition densities. Our technical contributions include the analysis of a broad family of convex programs -- parameterized by divergences such as quadratic, logarithmic, and hockey-stick functions -- whose minimizers recover the density decomposition. We prove that any strictly convex divergence yields the same canonical density vector, and that locally maximin allocations act as universal minimizers for all divergences satisfying the data processing inequality. As an application of our framework, we determine the structure and number of critical values in the combinatorial contracts problem. Additionally, we generalize a Frank-Wolfe-type iterative method for approximating the dual-modular density decomposition, establishing both convergence guarantees and practical potential through efficient gradient oracle design.