On the Number of Almost Envy-Free Allocations

05/30/2020 ∙ by Warut Suksompong, et al. ∙ 0

Envy-freeness is a standard benchmark of fairness in resource allocation. Since it cannot always be satisfied when the resource consists of indivisible items even when there are two agents, the relaxations envy-freeness up to one item (EF1) and envy-freeness up to any item (EFX) are often considered. We establish tight lower bounds on the number of allocations satisfying each of these benchmarks in the case of two agents. In particular, while there can be as few as two EFX allocations for any number of items, the number of EF1 allocations is always exponential in the number of items. Our results apply a version of the vertex isoperimetric inequality on the hypercube and help explain the large gap in terms of robustness between the two notions.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

The allocation of scarce resources to interested agents is a task that arises commonly in our everyday lives. Indeed, course slots need to be allocated to university students, grant funding to researchers, and personnel to organizations, to name but a few examples. A chief concern when allocating resources is fairness—we want all agents to feel that they receive a fair share of the resources. Several benchmarks of fairness have been proposed in the fair division literature. One of the most fundamental benchmarks is envy-freeness, which means that every agent should receive their first choice among all of the allocated bundles. Put differently, no agent should have reason to envy any other agent.

While envy-freeness is a compelling fairness benchmark, it suffers from the setback that it cannot always be satisfied when the resource consists of indivisible items such as books, course slots, or personnel. This is evident in the simple instance where there are two agents and only one indivisible item to be allocated. Consequently, two natural relaxations of envy-freeness have been considered. Envy-freeness up to one item (EF1) demands that any envy that one agent has towards another agent can be eliminated by removing a single item of our choice from the latter agent’s bundle, while the stronger notion of envy-freeness up to any item (EFX) requires that the envy can be eliminated by removing any single item from the latter agent’s bundle. An allocation that satisfies EF1 and EFX always exists when the items are divided between two agents with arbitrary valuations over subsets of items (Lipton et al., 2004; Plaut and Roughgarden, 2018).111The valuations are assumed to be monotonic, i.e., an agent’s value for a set cannot decrease when an item is added to the set. This assumption is usually made without loss of generality, since agents can freely dispose of items that they do not like.

The notions of EF1 and EFX serve as fairness criteria that can always be fulfilled when allocating indivisible items between two agents. On the one hand, given that EFX provides a stronger fairness guarantee than EF1, one might view it to be the appropriate criterion in this setting. On the other hand, a series of recent work has shown that EF1 is a more robust criterion than EFX in several ways:

  • An EF1 allocation can be computed using a number of queries that is only logarithmic in the number of items (Oh et al., 2019). On the contrary, computing an EFX allocation takes an exponential number of queries in the worst case (Plaut and Roughgarden, 2018), and a linear number of queries even when the valuations are additive (Oh et al., 2019).

  • There always exists a balanced EF1 allocation (that is, the numbers of items that the two agents receive differ by no more than ), while an EFX allocation may necessarily be highly unbalanced (Kyropoulou et al., 2019). Moreover, when the agents’ valuations are additive, an EF1 allocation fulfilling a set of cardinality constraints can also be found (Biswas and Barman, 2018).

  • With additive valuations, an EF1 allocation satisfying the economic efficiency condition of Pareto optimality always exists, i.e., no other allocation makes one agent better off without making the other agent worse off (Barman et al., 2018; Caragiannis et al., 2019). On the other hand, a Pareto-optimal EFX allocation may not exist (Plaut and Roughgarden, 2018). In addition, requiring EF1 leads to a smaller social welfare loss than EFX in the worst case (Bei et al., 2019b).

  • If the items lie on a line, an EF1 allocation in which each agent receives a contiguous block of items is guaranteed to exist (Bilò et al., 2019; Suksompong, 2019); this is not the case for EFX. More generally, when the items lie on a graph, the existence of an EF1 allocation in which every agent receives a connected subgraph can be guaranteed for a large class of graphs (Bilò et al., 2019), whereas an EFX allocation always exists only if the graph is complete (Bei et al., 2019a).

  • Beyond the setting of two agents, an EF1 allocation can always be found for any number of agents (Lipton et al., 2004), whereas the existence question for EFX remains intriguingly open when there are at least four agents with additive valuations, or three agents with non-additive valuations (Plaut and Roughgarden, 2018; Caragiannis et al., 2019; Amanatidis et al., 2020a, b; Chaudhury et al., 2020).

In this note, we provide an explanation to the robustness of EF1 by deriving tight lower bounds on the number of EF1 and EFX allocations for arbitrary valuations of the two agents. Specifically, while there can be as few as two EFX allocations regardless of the number of items, the number of EF1 allocations is—quite surprisingly—always exponential in the number of items.222We also remark that several extensions of EF1 have been recently studied, including weighted EF1 (Chakraborty et al., 2020), typewise EF1 (Benabbou et al., 2019), and democratic EF1 (Segal-Halevi and Suksompong, 2019).

2 The Bounds

In our formal model, there are two agents, who we sometimes call Alice and Bob, and a set of items. A subset of is referred to as a bundle. For , agent has a nonnegative value for each bundle . We assume without loss of generality that and, as is commonly done, that the valuations are monotonic, i.e., for any . A bundle is said to be

  • envy-free up to one item (EF1) for agent , if either , or there exists an item such that ;

  • envy-free up to any item (EFX) for agent , if for each item , we have .

An allocation is an ordered partition of into two sets , where bundle is given to agent . The allocation is said to be EF1 (resp. EFX) if for each , bundle is EF1 (resp. EFX) for agent . Moreover, we say that an unordered partition of into two sets is EF1 (resp. EFX) for agent if both and are EF1 (resp. EFX) for the agent.

We now derive the lower bounds on the number of EF1 and EFX allocations. We start with the stronger notion, EFX. For any number of agents with identical valuations, Plaut and Roughgarden (2018) showed that an EFX allocation always exists. This implies that in our setting with two agents, each agent has an EFX partition of the items, a fact that will be useful for our proof.

Theorem 1.

For two agents with arbitrary monotonic valuations, the number of EFX allocations is at least . Moreover, this bound is tight for any number of items .

Proof.

We let each agent propose an (unordered) EFX partition. If the two proposed partitions coincide, we have two EFX allocations by assigning either bundle to either agent. If they differ, letting Bob choose a preferred bundle from Alice’s partition yields an EFX allocation, and letting Alice choose a preferred bundle from Bob’s partition yields another EFX allocation different from the first one. Hence we can find two EFX allocations in both cases.

To show that the bound is tight, assume that the agents have identical valuations with value for each of the first items, and value for the last item. The valuations are additive: the value for a bundle of items is simply the sum of the values of the individual items in the bundle. It is clear that all of the first items must be given to the agent who does not receive the last item in order for the allocation to be EFX. Hence there are only two EFX allocations in this instance. ∎

Next, we move on to EF1, for which deriving a tight bound is considerably more challenging.

Theorem 2.

For two agents with arbitrary monotonic valuations, the number of EF1 allocations is at least , where

Moreover, this bound is tight for any number of items .

To prove this theorem, we will rely on two combinatorial results. The first result is a version of the vertex isoperimetric problem on the hypercube, where we want to choose a certain number of vertices on a hypercube so that the “vertex boundary” of this set is as small as possible. For a finite set , denote its power set by . A set system is a subset of . The Hamming distance between two sets is the size of their symmetric difference:

and the Hamming distance between two nonempty set systems is

The Hamming ball of center and radius is

A set system is called a Hamming ball of center and radius if

Note the difference between “the Hamming ball” and “a Hamming ball”.

Lemma 3 (Bollobás (1986); Calabro (2004)).

Let be a finite set, and let be nonempty set systems. There exist a Hamming ball with center and a Hamming ball with center such that , , and .

The second result that we will use in our proof concerns Sperner families, i.e., set systems in which no set contains another. A famous theorem of Sperner (1928) states that if , a Sperner family in can have size at most . Björner (1986, Thm. 2.2) provided more precise information regarding the number of sets of different sizes that can appear in a Sperner family. For any positive integers and , there is a unique way of writing

so that are integers. We define

and let .

Lemma 4 (Björner (1986)).

Let be a set with , and let be nonnegative integers such that at least one is strictly positive. There exists a Sperner family in with exactly members of size for if and only if

where is the smallest index such that .

With Lemmas 3 and 4 in hand, we are now ready to derive the lower bound on the number of EF1 allocations.

Proof of Theorem 2.

We first show tightness of the bound. Assume that the agents have identical valuations, and the value for a bundle of items is simply the sum of the values of the individual items in the bundle. If is even, suppose that the agents have value for every item. The EF1 allocations are exactly the allocations that assign items to each agent, so there are exactly EF1 allocations. For odd, suppose that the agents have value for each of the first items, and value for the last item. The first items must be split equally between the two agents, while the last item can go to either agent. Hence there are exactly EF1 allocations.

Next, we proceed to establish the bound. We claim that for each agent, at least partitions are EF1. Observe that this claim immediately yields the desired result. To see this, list partitions that are EF1 for Alice, and partitions that are EF1 for Bob. For a partition that appears in both lists, we can allocate either part to either agent, giving rise to two EF1 allocations. For a partition that appears in only one list, we let the other agent choose a preferred bundle from the partition, giving rise to one EF1 allocation. Hence the number of EF1 allocations is at least .

We now restrict our attention to Alice and prove that at least partitions are EF1 for her. Recall that denotes the set of all items, and call a bundle “good” if the partition is EF1 for Alice, and “bad” otherwise. Our goal is equivalent to showing that at least bundles are good. A bundle can be bad for one of two reasons: either it is “too large” (i.e., it is EF1, but its complement is not), or it is “too small” (i.e., it is not EF1, but its complement is). Note that a bundle is too small if and only if its complement is too large. In addition, it is easy to check from the definition of EF1 that if we remove an item from a bundle that is too small, then the resulting bundle is, as one should expect, also too small. Likewise, if we add an item to a bundle that is too large, the resulting bundle remains too large.

Consider bundles of items as elements of . We claim that if bundle is too small and bundle is too large, then . To prove this claim, it suffices to show that if a bundle is too small, then adding one item to it cannot make it too large. Assume that bundle is too small, and let . Denoting Alice’s valuation by , since is too small, we have . This is equivalent to , which means that is not too small. Hence is not too large, which establishes our claim.

Let (resp. ) be a set system consisting of all bundles that are too small (resp. too large). It follows from the preceding paragraph that . By Lemma 3, there exist a Hamming ball with center and a Hamming ball with center such that , , and . Moreover, symmetry implies that . The number of good bundles is .

Assume first that is even. Let be the unique integer such that . Since , we have . Note that contains all bundles of size at least , and contains all bundles of size at most . If , then since , we have that contains at least one bundle of size . Moreover, since contains , it contains all bundles of size . At least one of these bundles has Hamming distance from a bundle of size in , a contradiction. So , which means that no bundle of size belongs to . It follows that the number of good bundles is

For the rest of this proof, assume that is odd. As in the previous paragraph, let be such that and . If , then contains all bundles of size and contains all bundles of size , contradicting . If , then no bundle of size or belongs to . In this case, the number of good bundles is

Suppose now that . It suffices to prove that contains at most bundles of size , and contains at most bundles of size . Indeed, this would imply that the number of good bundles is

Assume for contradiction that contains bundles of size . By symmetry, contains bundles of size . Since , none of these bundles in can be a subset of any of the bundles in . This means that the bundles together form a Sperner family in . Applying Lemma 4, we find that

or . Observe that . Since , in order to obtain the desired contradiction, we only need to show that is a non-decreasing function of for any fixed .

To prove this statement, we proceed by induction on . The base case is trivial. Assume that the statement holds up to , and let be two positive integers. Write

with , and

with . If , the claim follows from the induction hypothesis applied to the two integers . If , we have

Else, . A similar chain of inequalities as above shows that , which is impossible. It follows that , concluding the induction and our proof. ∎

Acknowledgments

The author thanks Dominik Peters for useful discussions, the MathOverflow community for technical help, as well as an anonymous reviewer for helpful comments, and acknowledges support from the European Research Council (ERC) under grant number 639945 (ACCORD).

References

  • Amanatidis et al. [2020a] Georgios Amanatidis, Georgios Birmpas, Aris Filos-Ratsikas, Alexandros Hollender, and Alexandros A. Voudouris. Maximum Nash welfare and other stories about EFX. In

    Proceedings of the 29th International Joint Conference on Artificial Intelligence (IJCAI)

    , 2020.
    Forthcoming.
  • Amanatidis et al. [2020b] Georgios Amanatidis, Apostolos Ntokos, and Evangelos Markakis. Multiple birds with one stone: Beating 1/2 for EFX and GMMS via envy cycle elimination. In Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI), 2020. Forthcoming.
  • Barman et al. [2018] Siddharth Barman, Sanath Kumar Krisnamurthy, and Rohit Vaish. Finding fair and efficient allocations. In Proceedings of the 19th ACM Conference on Economics and Computation (EC), pages 557–574, 2018.
  • Bei et al. [2019a] Xiaohui Bei, Ayumi Igarashi, Xinhang Lu, and Warut Suksompong. The price of connectivity in fair division. CoRR, abs/1908.05433, 2019.
  • Bei et al. [2019b] Xiaohui Bei, Xinhang Lu, Pasin Manurangsi, and Warut Suksompong. The price of fairness for indivisible goods. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), pages 81–87, 2019.
  • Benabbou et al. [2019] Nawal Benabbou, Mithun Chakraborty, Edith Elkind, and Yair Zick. Fairness towards groups of agents in the allocation of indivisible items. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), pages 95–101, 2019.
  • Bilò et al. [2019] Vittorio Bilò, Ioannis Caragiannis, Michele Flammini, Ayumi Igarashi, Gianpiero Monaco, Dominik Peters, Cosimo Vinci, and William S. Zwicker. Almost envy-free allocations with connected bundles. In Proceedings of the 10th Innovations in Theoretical Computer Science Conference (ITCS), pages 14:1–14:21, 2019.
  • Biswas and Barman [2018] Arpita Biswas and Siddharth Barman. Fair division under cardinality constraints. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), pages 91–97, 2018.
  • Björner [1986] Anders Björner. Face numbers of complexes and polytopes. In Proceedings of the International Congress of Mathematicians (ICM), pages 1408–1418, 1986.
  • Bollobás [1986] Béla Bollobás.

    Combinatorics: Set Systems, Hypergraphs, Families of Vectors, and Combinatorial Probability

    .
    Cambridge University Press, 1986.
  • Calabro [2004] Chris Calabro. Harper’s theorem. http://cseweb.ucsd.edu/~ccalabro/essays/harper.pdf, 2004.
  • Caragiannis et al. [2019] Ioannis Caragiannis, David Kurokawa, Hervé Moulin, Ariel D. Procaccia, Nisarg Shah, and Junxing Wang. The unreasonable fairness of maximum Nash welfare. ACM Transactions on Economics and Computation, 7(3):12:1–12:32, 2019.
  • Chakraborty et al. [2020] Mithun Chakraborty, Ayumi Igarashi, Warut Suksompong, and Yair Zick. Weighted envy-freeness in indivisible item allocation. In Proceedings of the 19th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), pages 231–239, 2020.
  • Chaudhury et al. [2020] Bhaskar Ray Chaudhury, Jugal Garg, and Kurt Mehlhorn. EFX exists for three agents. CoRR, abs/2002.05119, 2020.
  • Kyropoulou et al. [2019] Maria Kyropoulou, Warut Suksompong, and Alexandros A. Voudouris. Almost envy-freeness in group resource allocation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), pages 400–406, 2019.
  • Lipton et al. [2004] Richard J. Lipton, Evangelos Markakis, Elchanan Mossel, and Amin Saberi. On approximately fair allocations of indivisible goods. In Proceedings of the 5th ACM Conference on Electronic Commerce (EC), pages 125–131, 2004.
  • Oh et al. [2019] Hoon Oh, Ariel D. Procaccia, and Warut Suksompong. Fairly allocating many goods with few queries. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI), pages 2141–2148, 2019.
  • Plaut and Roughgarden [2018] Benjamin Plaut and Tim Roughgarden. Almost envy-freeness with general valuations. In Proceedings of the 29th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 2584–2603, 2018.
  • Segal-Halevi and Suksompong [2019] Erel Segal-Halevi and Warut Suksompong. Democratic fair allocation of indivisible goods. Artificial Intelligence, 277:103167, 2019.
  • Sperner [1928] Emanuel Sperner. Ein Satz über Untermengen einer endlichen Menge. Mathematische Zeitschrift, 27(1):544–548, 1928.
  • Suksompong [2019] Warut Suksompong. Fairly allocating contiguous blocks of indivisible items. Discrete Applied Mathematics, 260:227–236, 2019.