House allocation problem

There are n people (also called: agents), and m objects (also called: houses). The agents may have different preferences over the houses. They may express their preferences in various ways:

• Binary valuations: each agent values each house at either 1 (which means that the agent likes the house), or 0 (which means that the agent dislikes the house).
• Preference ranking: each agent ranks the houses from best to worst. The ranking may be strict (no indifferences) or weak (indifferences allowed).
• Cardinal utility: each agent assigns a non-negative numeric value to each house.

Several considerations may be important in designing algorithms for house allocation.

• Pareto efficiency (PE) - no other allocation is better for some agents and not worse to all agents.
• Fairness - can be defined in various ways, for example, envy-freeness (EF) - no agent should envy another agent.
• Strategyproofness (SP) - each agent has an incentive to report his/her true preferences to the algorithm.
• Individual rationality (IR) - no agent should lose from participating in the algorithm.

In economics, the primary efficiency requirement in house allocation is PE. There are various algorithms attaining a PE allocation in various settings.

Probably the simplest algorithm for house allocation is serial dictatorship: the agents are ordered in some arbitrary order (e.g. by seniority), and each agent in turn picks the best remaining house by his/her preferences. This algorithm is obviously SP. If the agents' preferences are strict, then it finds a PE allocation. However, it may be very unfair towards the agents who pick last. It can be made fairer (in expectation) by choosing the order uniformly at random; this leads to the mechanism called random serial dictatorship. The mechanism is PE ex-post, but it is not PE ex-ante; see fair random assignment for other randomized mechanisms which are ex-ante PE.

When each agent already owns a house, fairness considerations are less important, it is more important to guarantee to agents that they will not lose from participating (IR). The top trading cycle algorithm is the unique algorithm which guarantees IR, PE and SP. With strict preferences, TTC finds the unique core-stable allocation.[3]

Abdulkadiroglu and Sönmez[1] consider an extended setting in which some agents already own a house while some others are house-less. Their mechanism is IR, PE and SP. They present two algorithms that implement this mechanism.

Ergin[4] considers rules that are also consistent, that is, their predictions do not depend of the order in which the assignments are realized.

In computer science and operations research, the primary efficiency requirement is maximizing the sum of utilities. Finding a house allocation maximizing the sum of utilities is equivalent to finding a maximum-weight matching in a weighted bipartite graph; it is also called the assignment problem.

Algorithmic problems related to fairness of the matching have been studied in several contexts.

When agents have binary valuations, their "like" relations define a bipartite graph on the sets of agents and houses. An envy-free house allocation corresponds to an envy-free matching in this graph. The following algorithmic problems have been studied.

• Deciding whether a complete EF allocation exists. This holds iff there exists a matching that saturates all the agents; this can be decided in polynomial time by just finding a maximum cardinality matching in the bipartite graph.
• Finding a partial EF allocation of maximum cardinality. This can be done in polynomial time.[5]^{: Thm.1.6(a)}
• Finding a partial EF allocation of maximum cardinality and minimum cost (where each edge has a pre-specified cost for society). This too can be done in polynomial time.[5]^{: Thm.1.6(b)}
• Finding a complete EF allocation, in which the number of non-envious agents is maximized. This problem is NP-hard. The proof is by reduction from the minimum coverage (aka bipartite expansion) problem.[6]^: Thm.3.5

When agents have cardinal valuations, the graph of agents and houses becomes a weighted bipartite graph. The following algorithmic problems have been studied.

• Deciding whether a complete EF allocation exists.
  • When m=n, all houses must be assigned, so an allocation is EF iff each agent gets a house with a highest value. Therefore, it is possible to reduce the original graph to an unweighted graph, in which each agent is adjacent only to his highest-valued houses, and look for a perfect matching in this graph.
  • When m>n, the above algorithm may not work, since not all houses must be assigned: even if a single house is most-preferred by all agents, there may exist a complete EF allocation in which this specific house is not assigned. Gan, Suksompong and Voudouris[7] present a polynomial-time algorithm that decides, in polynomial time, whether a complete EF allocation exists for any m ≥ n. Their algorithm uses, as a subroutine, an algorithm for finding an inclusion-minimal Hall violator.
• Deciding whether a complete local-envy-free allocation exists. Local envy-freeness means that the agents are located on a social network, and they only envy their neighbors in that network. Beynier, Chevaleyre, Gourves, Harutyunyan, Lesca, Maudet and Wilczynski[8] study the problem of deciding whether a complete local-envy-free allocation exists for m=n, for various network structures.
• Finding a complete EF allocation, in which the number of non-envious agents is maximized. Under common complexity theoretic assumption, this problem is hard to approximate. In particular, if NP cannot be solved in subexponential time, then it cannot be approximated to within a factor of ${\displaystyle n^{\gamma }}$ for some ${\displaystyle \gamma >0}$; if the small set expansion hypothesis is true, then it cannot be approximated to within a factor of ${\displaystyle n^{\gamma }}$ for any ${\displaystyle \gamma <1}$ (for ${\displaystyle \gamma =1}$ it is trivial to approximate: just give one agent his favorite house, and allocate the other agents arbitrarily). The proof is by reduction from the maximum balanced biclique problem.[6]^: Thm.3.1
• Deciding whether a complete proportional allocation exists. This problem is NP-complete, by reduction from exact 3-set cover.[6]^: Thm.4.1
• Deciding whether a complete equitable allocation exists. This problem can be solved in polynomial time.[6]^: Thm.5.1
• Finding a partial EF allocation of maximum cardinality. The runtime complexity of this problem is open.[5]^{: open question 2}

• Assignment problem - each agent must get a single object. The goal is to maximize the sum of valuations, or minimize the sum of costs.
• Fair random assignment - each agent must get a single object. Randomization is allowed. The allocation should be fair and efficient in expectation.
• Rental harmony - each agent must get a single object and pay a price; the allocation of objects+prices should be envy-free.
• Envy-free matching - some agents may remain unallocated, as long as they do not like any of the allocated houses.
• Fair item allocation - each agent may get any number of objects.