Envy-free matchings with cost-controlled quotas

TL;DR

This paper studies the problem of assigning agents to programs with resource-based quotas, focusing on envy-freeness and cost optimization, revealing complexity differences and providing approximation algorithms.

Contribution

It introduces the CCQ setting with cost-controlled quotas, analyzing the complexity of envy-free matchings for minimizing total and maximum costs, and offers approximation algorithms.

Findings

MINMAX is solvable in polynomial time.

MINSUM is NP-hard and hard to approximate.

LP-based approximation algorithms are developed for MINSUM.

Abstract

We consider the problem of assigning agents to programs in the presence of two-sided preferences, commonly known as the Hospital Residents problem. In the standard setting each program has a rigid upper-quota which cannot be violated. Motivated by applications where quotas are governed by resource availability, we propose and study the problem of computing optimal matchings with cost-controlled quotas -- denoted as the CCQ setting. In the CCQ setting we have a cost associated with every program which denotes the cost of matching a single agent to the program. Our goal is to compute a matching that matches all agents, respects the preference lists of agents and programs and is optimal with respect to the cost criteria. We consider envy-freeness as a notion of optimality and study two optimization problems with respect to the costs -- minimize the total cost (MINSUM) and minimize the maximum cost at a program (MINMAX). We show that there is a sharp contrast in the complexity status of these two problems -- MINMAX is polynomial time solvable whereas MINSUM is NP-hard and hard to approximate within a constant factor unless P = NP even under severe restrictions. On the positive side, we present approximation algorithms for the MINSUM for the general case and a special hard case. We chieve the approximation guarantee for the special case via a technically involved linear programming (LP) based algorithm. We remark that our LP is for the general case of the problem.