Fair by Design: Multidimensional Envy-Free Mechanisms Ahuva Mu’alem∗ December 25, 2013

Abstract The model we present addresses the following common scenario: a group of agents wants to divide a set of items fairly, and at the same time seeks to optimize a global goal. In this paper we focus on a natural task-scheduling scenario, in which each item is a task and we want to find an allocation that minimizes the completion time of the last task in an envy-free manner, where no individual agent prefers anyone else’s allocated task bundle over its own. In the scheduling literature, this optimization goal is called makespan minimization, and the agents are treated as machines. We give tight deterministic bounds on the approximation factors achievable for the following standard scenarios: (1) two unrelated machines and (2) m ≥ 2 related machines. A natural question to ask is whether envy-free pricing techniques can improve the current approximability and inapproximability bounds for truthful mechanisms for the task-scheduling problem studied in the seminal paper of Nisan and Ronen [26]. Here, we present several bounds for envy-freedom in-expectation, and give a partial answer to this question. We find that in multidimensional settings for two unrelated machines, truthful in-expectation is a far stronger constraint (i.e. more restrictive) than envy-free in-expectation.

1

Introduction

Traditionally fairness analysis focuses on the individual performance each participant receives, while allocation algorithms consider optimizing overall criteria. This paper formulates the fair by design approach to combine these two point of views. Consider a project with different tasks to be assigned to a heterogeneous group of employees. As a motivating scenario assume that Alice is a technician who specializes in repairing antennas and Bob specializes in repairing battery charges. Suppose that Carol the customer ∗

Software Engineering Dept. Ort Braude College of Engineering, Karmiel, Israel. This research was done while the author was a Post-Doctoral Fellow at the Social and Information Sciences Laboratory, California Institute of Technology. Email: [email protected], Web: sites.google.com/site/ahuvamualem

1

has two antennas and two battery chargers for repair. Consider two possible allocations: (1) the allocation in which Alice receives two antennas and Bob receives two battery charges; (2) the allocation in which each technician receives an antenna and a battery charger. If the goal of the manager is to complete the repair as soon as possible to please Carol, then the first allocation is preferable. However, if the manager’s goal is to expand the expertise of Alice and Bob then the second allocation is preferable. Observe that both allocations are fair from the point of view of Alice and Bob. This is not the case in general. A natural challenge is determining a fair allocation such that the last task completes as soon as possible. In general, no such allocation may exist if the tasks are indivisible. For instance, in a project with a single task, the fastest employee should be assigned the task. However, this allocation would not be considered fair from the perspective of the fastest employee. This suggests that some (additional) reward should be allowed to guarantee a fair division of the tasks. It is convenient to assume that rewards are granted in the form of monetary payments. In the scheduling literature, the above optimization goal is called makespan minimization of unrelated machines. In the economics literature, an allocation algorithm coupled with a price function is called a mechanism. We treat each machine as a distinct economic agent and say that an allocation is envyfree if no agent would prefer to exchange its assigned set of tasks with those of any other agent. Informally, a heterogeneous group of machines is called unrelated in the scheduling literature and multidimensional in the economics literature, while a homogeneous group is called related and single-dimensional, respectively. The design of mechanisms for our general problem can be regarded as solving an optimization problem with binding envy-freedom constraints. However, even if we allow monetary payments, it could be that no feasible solution exists: we are faced with an inherent clash between global optimization goals and envy-free pricing constraints (regardless of any computational considerations). Thus the need to consider approximations motivates the following definition: we say that an allocation a is a ρ-approximation if the makespan of allocation a is no more than a factor of ρ times the makespan of an optimal allocation, where ρ ≥ 1.1 Our main question is how well can fundamental global goals be approximated in an envy-free manner? Relation to Truthful Mechanisms The design of envy-free mechanisms is intimately connected to the well-studied class of truthful mechanisms [8, 10, 30, 15]. A mechanism is truthful if no agent can ever improve its utility by misreporting his valuation or cost. In many interesting settings, truthful mechanisms are essentially equivalent to mechanisms that select envy-free allocations with the 1

Additionally, if we consider a maximization problem (such as profit maximization) we say that an allocation a is a ρ-approximation if the value of allocation a is at least a factor of ρ1 times the value of an optimal allocation, where ρ ≥ 1. A fair-design problem exhibits an upper bound of ρU and a lower bound of ρL if there exists an envy-free ρU -approximation mechanism and if an envy-free (ρL − ²)-approximation mechanism is impossible for every ² > 0, respectively. If ρU = ρL we say that the bounds are tight.

2

smallest supporting price vectors [8]. The seminal paper of Nisan and Ronen [26] initiates the study of truthful mechanisms in computerized settings, examining the problem of how well the makespan goal on unrelated machines can be approximated. They showed that truthful mechanisms for two machines exhibit a tight bound of 2 [26], and they circumvent this impossibility result by employing randomization. In a truthful in-expectation mechanism, each bidder prefers to truthfully report his value to the mechanism since this gives him higher expected utility. Nisan and Ronen presented a truthful in-expectation mechanism for two machines that exhibits an upper bound of 74 (their upper bound was later improved to 1.5963 by [22]). The lower bound known for truthfulness in-expectation is 32 [24]. As tightening the bounds for unrelated machines is a central problem of algorithmic mechanism design, this raises the question of whether envy-free techniques can tighten the current upper and lower bounds (1.5963 vs. 32 ) known for truthful in-expectation mechanisms for two unrelated machines. In this paper we study envy-free in-expectation mechanisms to minimize the makespan. We present a 43 -approximation envy-free in-expectation mechanism for two unrelated machines. By the simple fact that the envy-free in-expectation upper bound of 43 is smaller than the truthful in-expectation lower bound of 32 [24], we get that our 34 -approximation envy-free in-expectation mechanism is not truthful in-expectation. Intuitively, this suggests that envyfree bounding techniques cannot be applied straightforwardly to tighten the current truthful in-expectation bounds for minimizing the makespan on two unrelated machines. We conclude that in multidimensional settings, truthful in-expectation is a far more restrictive constraint than envy-free in-expectation. Overview and Results In this paper we study two canonical objectives over multidimensional domains: profitmaximizing combinatorial auctions for general bidders and makespan-minimizing scheduling for unrelated machines. We start by formally defining the notion of an envy-free allocation mechanism. In Section 2, we briefly state a known characterization of envy-free allocation mechanisms in terms of locally-efficient bundle assignments [13]. Importantly, this characterization does not involve price functions. In Section 3, we study envy-free combinatorial auctions for general bidders. In this scenario, a profit-maximizing auctioneer has a collection of items for sale, and bidders compete for subsets of items. Envy-free prices can be interpreted as anonymous nondiscriminatory prices. We describe √ an envy-free mechanism that requires polynomial communication and achieves (min{n, O( k log(min{k, n}))})-approximation with respect to the maximum profit, where k is the number of items and n is the number of bidders. On the negative side, we show that any envy-free profit-maximizing mechanism with approximation ratio strictly better than n requires exponential communication. In Section 4, we study envy-free scheduling mechanisms. There are k tasks that are to 3

be scheduled on m unrelated machines. The total cost of a subset of tasks on machine i is the additive sum of the costs of the individual tasks on that machine. The global goal is minimizing the makespan of the chosen schedule; i.e., assigning the tasks to the machines in a way that minimizes the finishing time of the last task. This canonical optimization problem was extensively studied by [20]. We consider minimizing the makespan in the context of envy-free design. Specifically, using the characterization reviewed in Section 2, we derive general bounds on the approximability of deterministic envy-free mechanisms that seek to minimize the makespan on unrelated machines. We exhibit a lower bound of 2 − m1 of the best approximation ratio achievable by any deterministic envy-free mechanism for m unrelated machines. We present a deterministic ( m+1 + ²)-approximation envy-free mechanism. This mechanism is polyno2 mial time computable when m is a constant. Observe that for m = 2 we obtain a tight result of ( 32 + ²). Similar deterministic upper and lower bounds for unrelated machines were achieved independently in a recent work by Hartline et al. [14]. In section 5, we focus on the case of two unrelated machines and show that envy-free in-expectation mechanisms are more powerful than randomized envy-free mechanisms. Informally, a randomized envy-free mechanism specifies a probability distribution over deterministic envy-free mechanisms. While an envy-free in-expectation mechanism essentially specifies a probability distribution over allocations (so that each agent cares about his expected load) assuming that the agents are risk neutral. We first show that randomized envy-free mechanisms are not more powerful than their deterministic counterparts. That is, no randomized envy-free mechanism for two unrelated machines can achieve an approximation ratio better than the deterministic lower bound of 23 . To circumvent this impossibility we show a polynomial time computable ( 43 + ²)-approximation envy-free in-expectation mechanism for two unrelated machines. We show that this mechanism is optimal under the assumption that the mechanism is symmetric. Section 6 focuses on related machines [16] which is a special interesting case of the unrelated model. As opposed to our inapproximability results overviewed so far for the multidimensional settings, we show that the envy-freedom constraint does not impose any further burden if the machines are related. Specifically, there exists a polynomial-time computable deterministic envy-free mechanism that achieves the approximation ratio of 1+² with respect to the optimal makespan. Additionally, we characterize envy-free pricing in single-dimensional environments.2 Related Work The notion of envy-free allocations was introduced by Foley [11]. A recent survey on cakecutting and related models of divisible goods appears in [29]. Envy-free allocations without money of indivisible goods were studied from a computational point of view by Lipton et al. [21]. In this setting, envy-free allocations might not 2

The current paper supersedes "On Multidimensional Envy-Free Mechanisms" that appeared as an extended abstract in [23]. The current version includes a new section about envy-free in-expectation and randomized mechanisms.

4

exist, and thus they consider approximations for the minimum possible envy. Profit Maximization Envy-free profit maximization approximations for combinatorial auctions were first studied by Guruswami et al. [12]. For unit-demand bidders with limited supply they showed an O(log n)-approximation algorithm, and an O(log n + log k)approximation for single-minded bidders with unlimited supply, where n is the number of bidders and k is the number of distinct types of items.3 The latter result was extended to an O(log n + log k)-approximation for general bidders with unlimited-supply, √ by Balcan et al. [2]. Cheung and Swamy [5] used an LP-based technique to obtain an O( k log umax )approximation for single-minded bidders with limited-supply, where umax is the maximum number of item supply. None of these papers studies profit-maximizing envy-free pricing of combinatorial auctions for general bidders when supply is limited. Unrelated Machines Makespan minimization is NP-hard even on identical machines [4, 27]. This fundamental scheduling problem was extensively studied by Lenstra, Shmoys and Tardos [20]. They presented a 2-approximation polynomial-time algorithm for minimizing the makespan of unrelated machines. They also showed that the problem cannot be approximated in polynomial-time within a factor of less than 32 . Horowitz and Sahni presented an FPTAS4 for any fixed number of unrelated machines [17]. Recently, Cohen et al. [7] improved our deterministic results for envy-free makespan minimization for m ≥ 3 unrelated machines. They presented a polynomial time computable deterministic envy-free O(log m)-approximation mechanism. Additionally, they showed that no envy-free mechanism can achieve a better bound than O( logloglogmm ). A paper by Nisan and Ronen [26] defines the notion of algorithmic mechanism design [27]. In this paper each machine is treated as a strategic agent. They consider the unrelated machine setting. Their paper proves that not only is it impossible to minimize the makespan in a truthful manner, but that any approximation ratio strictly better than 2 cannot be achieved by a truthful deterministic mechanism (their lower bound was later improved from 2 to 2.61 for m ≥ 3 by Koutsoupias and Vidali [18]). They also showed that there is a computationally efficient truthful mechanism that achieves an approximation ratio of m. A lower bound of 2 − m1 for truthful in-expectation mechanisms for unrelated machines was given by Mu’alem and Schapira [24] while a randomized truthful upper bound of m+5 was 2 given by Lu and Yu [22], who also provide a randomized truthful upper bound of 1.5963 for the two machine case [22]. Related Machines In this setting the type of each related machine can be described easily by a single positive number (associated with its speed). Hochbaum and Shmoys describe a 3

A unit demand-bidder is essentially interested in every item, but would like to buy at most one item. A single-minded bidder would like to buy a specific subset of items. In the unlimited-supply setting, the number of copies of each item is as large as the number of bidders. 4 An FPTAS algorithm is an (1 + ²)-approximation algorithm that runs in polynomial time in the size of the input and 1² .

5

PTAS for minimizing the makespan of related machines [16].5 This canonical problem was first studied from an algorithmic mechanism design perspective in [1]. Archer and Tardos designed a 3-approximation mechanism based on a randomized rounding of the optimal fractional solution [1]. Recently, [9, 6] closed this gap and showed truthful mechanisms that achieve an approximation ratio of 1 + ².

2

Characterizing Envy-Free Multidimensional Mechanisms

This section characterizes envy-free mechanisms. The characterization is stated in terms of the local efficiency of the allocation rule. We begin with the general framework.

2.1

The Setting

We consider a finite set K of k indivisible items and a set N of n agents. We assume that agents value combinations of items. Formally, each agent i ∈ N has a valuation function vi () (or vi , for short) that describes its nonnegative valuation for each subset S of items, i.e. vi (S) is the maximum finite amount of money agent i is willing to pay for S. A subset S of items is sometimes called a bundle. Every valuation vi ∈ Vi satisfies the following three conditions: (1) No externalities meaning that the valuation of agent i depends only on its allocated bundle. (2) Free disposal - meaning that the valuation is nondecreasing with the set of allocated items (for every S and T , S ⊆ T implies vi (S) ≤ vi (T )). (3) Normalization - meaning that the value of the empty bundle is always zero. The domain of all possible valuations of agent i is denoted by Vi , where V = V1 × V2 × · · · × Vn . An allocation a = (a1 , . . . , an ) is a partition of items among the agents, where ai denotes the bundle allocated to agent i, a1 ∪ a2 ∪ · · · ∪ an ⊆ K (observe that not all items need to be allocated), and ai ∩ aj = ∅, whenever i 6= j. The set of all possible allowed allocations is denoted by A. An allocation rule f : V → A maps an n-tuple of valuations v = (v1 , v2 , . . . , vn ) ∈ V to an allocation a ∈ A. A mechanism specifies an allocation and a set of prices for every possible valuation of the agents. Formally, a mechanism is a tuple M = (f, p1 , p2 , . . . , pn ) (or M (f, p), for short) where f is an allocation rule and the pricing function pi : V → R assigns a price to each agent i ∈ N . A mechanism M (f, p) is individually rational if agents always receive a nonnegative utility. Formally, if for every v = (vP 1 , v2 , . . . , vn ) we have that vi (f (v)) − pi (v) ≥ 0. The profit collected by the mechanism is ni=1 pi (v). Definition 1 (Envy-Free Mechanism) Let M = (f, p) be a mechanism. Let i, j ∈ N and let v = (v1 , . . . , vn ) ∈ V be any n-tuple of valuations. Denote by a ∈ A the allocation 5

A PTAS is an (1 + ²)-approximation algorithm that runs in polynomial time in the size of the input, assuming ² is a fixed constant.

6

that f outputs for v. A mechanism M is said to be envy-free if for every pair of agents i, j it holds that: vi (ai ) − pi (v) ≥ vi (aj ) − pj (v). We say that an allocation rule f : V → A is envy-free achievable if there exists a price function p such that the mechanism M = (f, p) is envy-free.6 Example 1 Consider a mechanism for a single item with two agents, where the agent with the highest value wins the item and pays the average of both values. The other agent pays zero. By symmetry, it is enough to consider the case where v1 ≥ v2 . By the fact that 2 2 v1 − v1 +v ≥ 0 ≥ v2 − v1 +v , we immediately obtain that agent 1 does not envy agent 2, and 2 2 vice versa. Therefore the mechanism is envy-free. Definition 2 For arbitrary allocation a ∈ A and valuation v ∈ V, let Ψ(v, a) = Σni=1 vi (ai ). We call Ψ(v, a) the social welfare of the allocation a with respect to v.

2.2

Locally Efficient Allocations

In order to be able to state the characterization theorem we start with some definitions. Definition 3 (Bundle Allocation based on a and β) Let β : N → N be an arbitrary function. Let a = (a1 , . . . , an ) ∈ A be an arbitrary allocation. We say that the allocation aβ is the bundle-allocation based on a and β, if agent i in aβ is allocated all bundles ak with β(k) = i. We also consider the special case where the function β : N → N is a permutation. Definition 4 (Locally-Efficient Bundle Assignment) An allocation a = (a1 , . . . , an ) is said to be a locally-efficient bundle assignment with respect to v = (v1 , . . . , vn ), if for every permutation π : N → N it holds that: Ψ(v, a) ≥ Ψ(v, aπ ). Clearly, the allocation rule f ∗ (v) ∈ argmaxa∈A Ψ(v, a) which maximizes the social welfare, produces locally-efficient bundle assignments. Specifically, locally-efficient allocations always exist. 6

The agents in our setting are non-strategic, they always report their true valuations. Observe also that we use the notion of "achievable" rather than the notion of "dominant-strategy implementable".

7

2.3

Characterizing Envy-Freedom

We now state a known characterization of envy-free bundle pricing mechanisms in terms of local efficiency by Haake et al. [13]. We provide the proof (based on a shortest path argument) in the appendix. Theorem 1 (Haake et al. [13]) A deterministic allocation rule f : V → A is envy-free achievable if and only if the allocation f (v) is a locally-efficient bundle assignment with respect to v, for every v ∈ V . Intuitively, the above characterization allows us to focus on the allocation rule alone in order to prove or disprove envy-free achievability rather than considering the interplay between the allocation rule and the price function, Based on this characterization, the supporting prices can be calculated in polynomial time. More formally: Claim 1 Given any allocation a ∈ A and values vi (aj ), i, j = 1..n, it can be decided in polynomial time whether supporting envy-free prices exist. Furthermore, if individually-rational nonnegative envy-free prices exist they can be computed in polynomial time.

2.4

Cost Minimization Problems

The above characterization theorem is stated for value maximization problems (such as profit maximization). It also applies to cost minimization problems (such as makespan minimization). The Setting. As before, we will still have a finite set K of k indivisible items and a set N of n agents. Each agent i ∈ N now has a nonnegative cost function ci () ∈ Ci (or ci , for short) that describes its cost incurred by each subset S of items. A mechanism M (f, p) for a cost minimization setting is individually rational if agents always receive a nonnegative utility. That is, if for every c = (c1 , c2 , . . . , cn ) we have that −ci (f (c)) − pi (c) ≥ 0. For notational simplicity, we define pbi (c) = −pi (c). In particular, pbi is called a reward and it specifies the monetary transfer that agent i receives from the mechanism, while pi refers to the price that agent i gives to the mechanism. Therefore, the nonnegative utility requirement is equivalent to pbi (c) − ci (f (c)) ≥ 0. Similarly, the inequality in Definition 1 is equivalent to pbi (c) − ci (ai ) ≥ pbj (c) − ci (aj ). An allocation a = (a1 , . . . , an ) is said to be a locally-efficient bundle assignment with respect to c = (c1 , . . . , cn ), if for every permutation π : N → N it holds that: Ψ(c, a) ≤ Ψ(c, aπ ), 8

where Ψ(c, a) = Σni=1 ci (ai ) and Ψ(c, aπ ) = Σni=1 ci (aπi ). Proposition 1 A deterministic allocation rule f : C → A is envy-free achievable if and only if the allocation f (c) is a locally-efficient bundle assignment with respect to c, for every c ∈ C. Furthermore, given any allocation a ∈ A and costs ci (aj ), i, j = 1..n, it can be decided in polynomial time whether a supporting envy-free reward function pb exists. Additionally, if individually-rational nonnegative envy-free reward function exists it can be computed in polynomial time.

3

Profit-Maximizing Combinatorial Auctions

In this section we consider the problem of maximizing the auctioneer’s profit in a combinatorial auction. Here, we relate to the agents as bidders; each bidder can have a different value for each bundle of items and thus it is a multidimensional setting. Envy-free prices can be interpreted as anonymous non-discriminatory prices. We quantify the number of bits required to be communicated to determine a profitable allocation.7 We show √ an envy-free mechanism that requires polynomial communication and achieves (min{n, O( k·log(min{n, k}))})-approximation with respect to the maximal envy-free profit. However, we show that an envy-free mechanism with approximation ratio strictly better than 2 (with respect to the optimal profit) requires exponential communication in the worst case. We then show that a similar impossibility result applies to approximation ratios strictly better than n, where n is the number of bidders. Based on the work of Guruswami et al. [12], Lehmann et al. [19] and Nisan and Segal [28] we can state the main results of this section. Proposition 2 Any envy-free profit-maximizing mechanism for Combinatorial Auctions that achieves an approximation ratio better than 2 requires exponential communication. Proposition 3 Any envy-free profit-maximizing mechanism for Combinatorial Auctions that achieves an approximation ratio better than n requires exponential communication. Proposition 4 There exists an envy-free mechanism for combinatorial auctions for general √ bidders which achieves a (min{n, O( k ·log(min{n, k}))})-approximation for maximizing the profit and which requires polynomial communication.

4

Deterministic Envy-Free Approximability of Unrelated Machines

This section focuses on minimizing the makespan for unrelated machines. For this model we prove that not only is it impossible to minimize the makespan in an envy-free manner, but 7

An introduction to communication complexity for combinatorial auctions can be found in [27, Chapter

11].

9

that any approximation ratio better than 2− m1 cannot ¡ m+1 be achieved ¢ by an envy-free deterministic mechanism. We then present an envy-free · (1 + ²) -approximation mechanism. 2 Specifically, for m = 2 we get the tight deterministic result of 23 + ². Similar deterministic upper and lower bounds for unrelated machines were achieved independently in a recent work by Hartline et al. [14].

4.1

The Setting

The unrelated machine scheduling setting (R||Cmax ) is a special case of the combinatorial auction setting: There are k tasks to be scheduled on m machines.8 Every machine i is an agent with a nonnegative cost function ci (). Formally, ci ({j}) (or simply ci (j)) specifies the cost of task j on machine i. One can think of the cost of task j on machine i as the time it takes i to complete j. The total cost of a set of tasks S on machine i is the additive sum of the costs of the individual tasks on that machine. Formally, ci (S) = Σj∈S ci (j) for every S. In the unrelated machines setting these costs can be arbitrary (every (k · m)-tuple of nonnegative costs is feasible), and thus it is a multidimensional scheduling problem. Let a ∈ A be an arbitrary allocation of tasks to the machines ("scheduling"), where all tasks must be assigned, and each task is assigned to exactly one machine. The load of machine i is its cost ci (ai ) = Σj∈ai ci (j). The makespan denoted by r(a, c) is the maximum load of all machines, that is r(a, c) = max{c1 (a1 ), c2 (a2 ), . . . , cm (am )}. Given c and m, the makespan minimization problem involves finding an allocation a that minimizes the term r(a, c). We shall use the notation r(a) instead of r(a, c), when c is clear from the context.

4.2

A Deterministic Lower Bound

The optimal allocation with respect to makespan may not be envy-free achievable. In this subsection, we will use the characterization theorem to prove a lower bound on the approximability of envy-free mechanisms. Theorem 2 No envy-free mechanism for m unrelated machines can achieve an approximation ratio better than 2 − m1 with respect to the optimal makespan. Proof: Suppose not. Let M (f, pb) be a deterministic envy-free mechanism that achieves an approximation factor of 2− m1 −². Let ²0 < ². We shall consider two cases. For m = 2, consider the following instance with two tasks: c1 (1) = 1, c1 (2) = 0.5, c2 (1) = 1.5 − ²0 , c2 (2) = 1, (see the matrix below): 8

We chose m to be the number of machines to be consistent with the formulation of Nisan and Ronen. Recall that we used n previously to denote the number of agents, whereas here the agents are the machines.

10

µ

1 0.5 0 1.5 − ² 1

¶

The first column represents the costs of running the first task on the first and the second machines, respectively. Similarly, the second column represents the costs of the second task on each of the machines. Clearly, there exist exactly 4 possible allocations. The makespan of allocating the first task to the first machine and the second task to the second machine is 1. However, this allocation is not a locally-efficient bundle assignment and thus by Proposition 1 is not envyfree achievable, as one can easily verify that: 2 = c1 (1) + c2 (2) > c1 (2) + c2 (1) = 2 − ²0 (recall that this is a cost minimization setting). Furthermore, any other allocation has a makespan ≥ 1.5 − ²0 > 1.5 − ²; this contradicts the assumption that M (f, pb) is a deterministic envy-free mechanism that achieves an approximation factor of 1.5 − ². For m ≥ 3, consider the following matrix   ∞ ∞ ··· ∞ 1 1 − m1     1   ∞ 1 1 − ∞ · · · ∞   m       .. .. .. . .   . . . .       1   ∞ ∞ · · · ∞ 1 1 − m        2 − m1 − ²0  ∞ ··· ∞ ∞ 1 First, consider the allocation b = ({1}, {2}, . . . , , {m}), where task i is assigned to machine i. The makespan of this allocation is 1. Observe that the optimal makespan cannot be strictly smaller than 1, since the cost of the first task on each machine is at least 1. Therefore, b is an optimal allocation with respect to the makespan. However, b is not a locally-efficient bundle assignment: To see this, let bb = ({2}, {3}, {4}, . . . , {m}, {1}) be the assignment marked with italics in the above matrix. Now, Ψ(c, b) = m > (m − 1)(1 − 1 ) + 2 − m1 − ²0 = m − ²0 = Ψ(c, bb). m Let a be an allocation with an approximation ratio of at most 2− m1 −². First, no machine in a is assigned two or more tasks (since the sum of any two elements at each row is at least 2 − m1 ). Additionally, the first task must be assigned to the first machine (otherwise, the makespan would be ≥ 2 − m1 − ²0 > 2 − m1 − ²). Now, the second task must be assigned to the second machine (since assigning it to any other machine would results in a makespan ≥ 2 − m1 ). Repeating the last argument, task i must be assigned to the i0 th machine for every i = 3..m. This implies that a = b and that 11

every allocation 6= b has an approximation ratio strictly larger than 2 − m1 − ² with respect to the optimal makespan. This contradicts the assumption that M (f, pb) is a deterministic envy-free mechanism that achieves an approximation factor of 2 − m1 − ².

4.3

A Deterministic Upper Bound

In this subsection we show how to convert any ρ-approximation algorithm into an envy-free ( m+1 · ρ)-approximation mechanism. Applying the conversion technique to the deterministic 2 FPTAS for any fixed number of machines presented by Horowitz and Sahni [17] yields a polynomial-time-computable ¡ ¢deterministic envy-free mechanism that achieves an approximation ratio of m+1 · (1 + ²) with respect to the optimal makespan. This result is nearly 2 optimal for a small number of machines. Specifically, for m = 2 we get a tight result of 32 + ². We first consider several procedures to reconfigure a given allocation into a locally-efficient bundle assignment. Let [m] = {1, 2, . . . , m} denote the set of machines. Definition 5 (The Function β ∗ ) Let a = (a1 , . . . , am ) ∈ A be an arbitrary allocation. Define the function β ∗ : [m] → [m] as follows: β ∗ (j) ∈ argmin i=1,...,m ci (aj ), j ∈ [m]. That is, β ∗ (j) is the machine with the minimal cost for the bundle aj (breaking ties arbitrarily). ∗

The allocation aβ can be constructed by the following "history-independent" procedure: at step j = 1..m, re-assign bundle aj to a machine with the minimal cost for this bundle. Note that this construction may result in a single machine receiving more than one bundle - or even all bundles. Definition 6 (The Permutation π ∗ ) Let a = (a1 , . . . , am ) ∈ A be an arbitrary allocation. ∗ Define π ∗ to be a permutation such that aπ is a locally-efficient bundle assignment. If there is more than one such permutation, then arbitrarily choose one. ∗

∗

Fact 1 If the cost of each machine is additive, then b = aβ and aπ are locally-efficient ∗ bundle assignments. In particular, Ψ(c, aπ ) ≤ Ψ(c, a), and Ψ(c, b) ≤ Ψ(c, bπ ) for every permutation π. We are now ready to state the algorithm. Algorithm 1 (Bundle-Local-Search) Input: c = c1 , c2 , . . . , cm and an allocation a. ∗

• If the makespan of aπ is at most

m+1 2

∗

times the makespan of a, then output aπ .

∗

• Otherwise, output aβ . ∗

Informally, if the makespan of aπ is not too far from the makespan of a, we output π∗ a . Otherwise, we re-assign each bundle ai independently to the fastest machine for that particular bundle.

12

Theorem 3 Algorithm 1 always ¡ m+1 outputs ¢ a locally-efficient bundle assignment and guarantees an approximation ratio of 2 · r(a) with respect to the makespan in polynomial time. Proof: Algorithm 1 always outputs a locally-efficient bundle assignment and thus by Proposition 1 is envy-free achievable. ∗ Assume without loss of generality, that the highest loaded machine in allocation aπ is ∗ ∗ machine 1. If r(aπ ) ≤ m+1 · r(a) then Algorithm 1 outputs the allocation aπ . Otherwise, 2 we get for the highest loaded machine: ∗

∗

c1 (aπ ) = r(aπ ) > Additionally,

∗

m+1 · r(a). 2

∗

Ψ(c, aβ ) ≤ Ψ(c, aπ ) ≤ Ψ(c, a) ≤ m · r(a). Putting these together, ∗

Ψ(c, aβ ) ≤ r(a) + m · r(a) −

m+1 m+1 · r(a) = · r(a), 2 2

where r(a) upper-bounds the value of the machine with the minimum cost for the bundle ∗ ∗ aπ1 , and m · r(a) − m+1 · r(a) upper-bounds the total minimum cost of the bundles aπi , i ≥ 2. 2 In the worst case, all bundles will be allocated to the same machine. We thus have the following: ∗

∗

r(aβ ) ≤ Ψ(c, aβ ) ≤

m+1 · r(a). 2

Theorem 4 (Horowitz and Sahni [17]) For any fixed number of machines and ² > 0 there exists an (1 + ²)-approximation algorithm for minimizing the makespan on unrelated machines. The running time of this algorithm is polynomial in the number of tasks k and in 1 . ² Theorem 5 ¢For any fixed number of machines and ² > 0 there exists an envy-free ¡ m+1 · (1 + ²) -approximation mechanism for minimizing the makespan on unrelated ma2 chines. The running time of this mechanism is polynomial in the number of tasks k and in 1 . ² Proof: For every input and ² > 0 we can simulate the (1 + ²)-approximation algorithm in Theorem 4 to get a nearly optimal allocation a with respect to the makespan, and we then can simulate ¡Algorithm 1 on ¢ allocation a. By Theorems 3 and 4 and Proposition 1 this yields m+1 an envy-free 2 · (1 + ²) -approximation mechanism whose running time is polynomial in k and 1² .

13

5

Envy-Free in-Expectation Mechanisms for Unrelated Machines

In the previous section we showed a lower bound of 32 for two machines. Here we circumvent this impossibility by employing randomization. Agents in an envy-free in-expectation setting care about their expected load (rather than the actual assigned load). The main result in this section is the existence of an envy-free inexpectation mechanism with an approximation ratio of 43 + ² for two unrelated machines. We then show that this bound is tight under the assumption that the mechanism is essentially symmetric from the perspective of agents.

5.1

The Setting

Any envy-free in-expectation mechanism can be regarded as a mechanism that produces a probability distribution over possible allocations for every instance of a given problem. In turn, for each agent this implies agent-specific probability distribution over tasks. The cost function of each of the agents in such mechanisms can therefore be viewed as assigning costs to probability distributions over tasks, assuming risk neutrality as defined below. Definition 7 (Risk Neutrality) Let ci be a cost function. For every probability distribution D over the tasks 1..k we define the extended cost function Eci as follows: Eci (D) = Σt=1..k P rD [t] · ci (t), where P rD [t] is the probability of task t in D. Definition 8 (Envy-Free in-Expectation Mechanism) Let ∆(A) be the set of all possible allowed probability distributions over allocations. Let f : C → ∆(A) be an allocation rule that maps a tuple of costs c = (c1 , . . . , cn ) ∈ C to a probability distribution over possible allocations ∈ ∆(A). A mechanism M (f, pb) is said to be envy-free in-expectation if for every pair of agents i, j and every c it holds that: pbi (c) − Eci (Di ) ≥ pbj (c) − Eci (Dj ), where pbi , pbj are the the reward that agents i, j receive, and Di , Dj are the associated agentspecific probability distributions over tasks produced by f for c.

5.2

An Upper Bound of

4 3

for Two Unrelated Machines

We now describe our allocation algorithm for the envy-free in-expectation of two machines. Algorithm 2 (Random-Bundle-Local-Search) Input: c = c1 , c2 . 14

1. Simulate the (1 + ²)-approximation algorithm in Theorem 4 to get a nearly optimal allocation a with respect to the makespan. 2. If a is a locally-efficient bundle assignment, then output the allocation a and the rewards pb as in Proposition 1. 3. Let b0 be the allocation that allocates all the tasks to machine 1. Let b00 be the allocation that allocates all the tasks to machine 2. If min(r(b0 ), r(b00 )) ≤ 34 r(a) then output the allocation argmin(r(b0 ), r(b00 )) and the rewards pb as in Proposition 1. 4. Otherwise, let aR be the reverse allocation that allocates the bundle a2 to machine 1 and a1 to machine 2. Output the allocation a with probability 12 and the allocation aR with probability 12 . The reward pb to each machine will be max(r(b0 ), r(b00 )). Theorem 6 Algorithm 2 yields an envy-free in-expectation ( 34 +²)-approximation mechanism for minimizing the makespan on two unrelated machines whose running time is polynomial in the number of tasks k and 1² . Proof: In steps 2 and 3 Algorithm 2 deterministically outputs locally-efficient bundle assignments, and thus by Proposition 1 is envy-free achievable. Let a∗ be the optimal allocation with respect to the makespan. If Algorithm 2 terminates at the third step then clearly min(r(b0 ), r(b00 )) ≤ 43 r(a) ≤ 43 · r(a∗ ) · (1 + ²). In the last step, each machine gets each task with probability 21 and therefore envyfreedom holds by using exactly the same reward for each machine. It remains to show that the makespan of the last step is within 34 (1 + ²) of the optimal makespan. Clearly, if Algorithm 2 terminates at the last step, then a 6= b0 , b00 . Additionally, a1 , a2 6= ∅, min(r(b0 ), r(b00 )) > 4 r(a), and aR is a locally-efficient bundle assignment. 3 We first claim that 5 ci (aR · r(a), i = 1, 2. i ) < 3 By using the facts that min(r(b0 ), r(b00 )) > 43 · r(a) and ci (ai ) ≤ r(a), i = 1, 2, we get that 1 R R R ci (aR i ) > 3 · r(a), i = 1, 2. Now, the local efficiency of a translates to c1 (a1 ) + c2 (a2 ) < 5 c1 (a1 ) + c2 (a2 ) ≤ 2 · r(a). Putting these together we have that ci (aR i ) < 3 · r(a), i = 1, 2. Finally, we get the desired approximability bound on the expected makespan: 1 1 1 5 4 4 · r(a) + · r(aR ) ≤ (1 + ) · r(a) = · r(a) ≤ · r(a∗ ) · (1 + ²). 2 2 2 3 3 3

5.3

A Tight Lower Bound for Balanced Mechanisms

We now show that our envy-free in-expectation mechanism is optimal (regardless of computational considerations) under the assumption that the mechanism is balanced. Formally: Definition 9 (Balanced Mechanism) Let Di (c) denote the associated agent-specific probability distribution over tasks for agent i produced by a mechanism M (f, pb) for the input 15

c = (c1 , . . . , cn ). A distribution Di (c) is called truly random if there exists a task t such that 0 < P rDi (c) [t] < 1. A mechanism M (f, pb) is balanced if a truly random Di (c) for c = (c1 , . . . , cn ) and agent i implies that P rDi (c) [t0 ] = P rDj (c) [t0 ] for every agent j and task t0 = 1..k. Intuitively, the assumption of balance implies that any non-deterministic allocation must be symmetrical from the perspective of the agents: all agents’ truly random probability distributions over tasks must be identical. However, observe that it still allows a rather large degree of freedom in choosing the probability distribution over possible allocations. Theorem 7 No balanced envy-free in-expectation mechanism for two unrelated machines can achieve an approximation ratio strictly better than 43 with respect to the makespan. Proof: To show this bound we solve a linear program based on the the following instance with two machines and two tasks: 1 5 c1 (1) = 1, c1 (2) = , c2 (1) = − ², c2 (2) = 1. 3 3 The optimal makespan of this instance is 1, however the optimal allocation with respect to the makespan is not locally-efficient. Additionally, it is easy to check that the makespan of any other deterministic allocation is at least 43 . The optimal approximation ratio achievable by any balanced envy-free in-expectation mechanism for the above instance can be computed by the following linear program: Minimize

x1 + ( 53 − ²)x2 + 43 x3 + ( 83 − ²)x4

s.t.

x1 + x2 + x3 + x4 = 1 −x1 + x2 − 2x3 + 2x4 + 1.5x5 − 1.5x6 ≥ 0 ( 32 − ²)x1 − ( 23 − ²)x2 + ( 83 − ²)x3 − ( 83 − ²)x4 − x5 + x6 ≥ 0 x5 − x6 = 0 xi ≥ 0, i = 1, 2, 3, 4

The variables x1 , x2 , x3 , x4 represent the probabilities of allocations a1 , a2 , a3 , a4 in the optimal envy-free in-expectation mechanism, where a1 = ({1}, {2}), a2 = ({2}, {1}), a3 = ({1, 2}, ∅), and a4 = (∅, {1, 2}). The variables x5 , x6 represent the reward to the first and the second machines, respectively. The objective function being minimized is the expected makespan. The first and the last constraints imply that x1 , . . . , x4 is a probability distribution over the set of allocations. Now, if the first machine does not envy the second machine, then: 16

x5 − x1 − 13 x2 − 43 x3 ≥ x6 − 13 x1 − x2 − 43 x4 . By rearranging we get the second constraint. Similarly, if the second machine does not envy the first machine, then x6 − x1 − ( 53 − ²)x2 − ( 83 − ²)x4 ≥ x5 − ( 53 − ²)x1 − x2 − ( 83 − ²)x3 ; by rearranging we get the third constraint. Finally, the fourth constraint x5 = x6 is implied by the assumption that the mechanism is balanced. The optimal solution is (x1 , x2 , x3 , x4 , x5 , x6 ) = (0.5, 0.5, 0, 0, 0, 0), with objective value 34 . This probability distribution induces a balanced envy-free in-expectation mechanism, where each machine gets exactly one task uniformly at random. Clearly, the solution is feasible. To verify the optimality of this solution observe that: 4 3

− ² ≤ ( 43 − ²)(x1 + x2 + x3 + x4 ) + 12 (x5 − x6 ) + 13 (−x1 + x2 − 2x3 + 2x4 + 1.5x5 − 1.5x6 ) ≤ x1 + ( 35 − ²)x2 + 43 x3 + ( 83 − ²)x4 .

5.4

Inapplicability of Randomized Envy-Free Mechanisms

A randomized mechanism is a probability distribution over deterministic mechanisms. Practically, for every c the mechanism M (f, pb) produces a distribution DM (c) over deterministic mechanisms and outputs a deterministic mechanism drawn from this distribution (observe that Definition 7 is redundant here). We next show that no randomized envy-free mechanism can provide a strictly better lower bound than any deterministic envy-free mechanism. Formally, Proposition 5 No randomized envy-free mechanism for two machines can achieve an approximation ratio strictly better than 32 (in expectation). Proof: Consider the following instance for two machines c1 (1) = 1, c1 (2) = 0.5, c2 (1) = 1.5 − ²0 , c2 (2) = 1, taken from the proof of Theorem 2. Now, any realization of a randomized envy-free mechanism must output a deterministic envy-free mechanism. Recall that a deterministic envy-free mechanism must output a locally-efficient bundle assignment. However, as we have seen in the proof of Theorem 2, the optimal allocation with respect to the makespan for this instance is not locally-efficient. Furthermore, any other allocation has a makespan of at least 32 times the optimal. Therefore, no convex combination of locally-efficient bundle assignments can produce an approximation ratio strictly better than 32 .

6

Near-Optimality of Related Machines

This section focuses on related machines so as to minimize the makespan, and shows that the envy-freedom constraint does not impose any further burden. Specifically, we show how to convert any ρ-approximation algorithm into an envy-free ρ-approximation mechanism in polynomial time. Applying the conversion technique to the deterministic PTAS

17

by Hochbaum and Shmoys [16] yields a polynomial time computable deterministic envyfree mechanism that achieves the approximation ratio of 1 + ² with respect to the optimal makespan.

6.1

The Setting

The related-machines setting is a special case of the unrelated model. This problem is denoted Q||Cmax in the scheduling literature, and is NP-hard [4], although a PTAS [16] exists. In this model, each task j has a load lj > 0, and every machine i has a type ti > 0. The running time of task j on machine i is ti ·lj . Additionally, machine i’s cost for performing the task j is ci ({j}) = ti ·lj . Since the costs here are derived from the single parameter ti ∈ R, it is considered to be a single-dimensional scheduling problem [1]. The total cost of a set of tasks on machine i is the additive sum of the costs of the individual tasks on that machine. For convenience we use the notation l(S) = Σj∈S lj , to denote the total load of a subset of tasks S.

6.2

Characterizing Envy-Free Single-Dimensional Mechanisms

The following simple condition characterizes envy-freedom in single-dimensional environments. As noted by [15], this characterization is nearly identical to that of the analogous characterization of truthful mechanisms for single-dimensional environments by Myerson [25]. Definition 10 An allocation a = (a1 , . . . , am ) is called aligned if l(ai ) < l(aj ) implies that ti ≥ tj for every pair of machines i, j. Proposition 6 An allocation rule f is envy-free achievable if and only if f outputs an aligned allocation for every c. 6.2.1

Proof of Proposition 6 and Frugality

We first show the "if" part of Proposition 6. Lemma 1 If f is envy-free achievable then f outputs an aligned allocation for every c. Proof: Suppose that f is envy-free achievable but f (c) = a, l(ai ) < l(aj ) and ti < tj for some c, i and j. Now, (l(aj ) − l(ai )) · (tj − ti ) > 0 is equivalent to l(ai )ti + l(aj )tj > l(ai )tj + l(aj )ti . But then exchanging the loads between machines i and j strictly decreases the overall social cost, contradicting Proposition 1. To show the "only-if" part of Proposition 6 it is convenient to directly use the following reward function (rather than to use Proposition 1 to show the local efficiency of aligned allocations).

18

Definition 11 (Allocation-Specific Frugal Reward) Suppose that f (c) = a, where allocation a is aligned. Assume that the machines are numbered from 1 to m in order of decreasing type, breaking ties in favor of the smallest load with respect to a, so that t1 ≥ t2 ≥ · · · ≥ tm and l(a1 ) ≤ l(a2 ) ≤ · · · ≤ l(am ). The allocation-specific frugal reward function is defined recursively as follows: pb1 (c) = t1 · l(a1 ) pbi (c) = pbi−1 (c) + ti · (l(ai ) − l(ai−1 )) i = 2, . . . , m. Lemma 2 If f outputs an aligned allocation for every c then f is envy-free achievable. Proof: We will show that f , coupled with the above allocation-specific frugal reward, yields an envy-free mechanism. Suppose that f (c) = a. Using Definition 11 we show that a slow machine cannot envy a fast machine, and vice versa. We have that pbi (c) − ti · l(ai ) ≥ pbd (c) − ti · l(ad ) for i < d by rearranging the following: ti · l(ad ) − ti · l(ai ) = ti · [(l(ad ) − l(ad−1 )) + (l(ad−1 ) − l(ad−2 )) + · · · + (l(ai+1 ) − l(ai ))] ≥ td · (l(ad ) − l(ad−1 )) + · · · + ti+1 · (l(ai+1 ) − l(ai )) = pbd (c) − pbi (c). Similarly, we have that pbi (c) − ti · l(ai ) ≥ pbd0 (c) − ti · l(ad0 ) for d0 < i by rearranging the following: pbi (c) − pbd0 (c) = ti · (l(ai ) − l(ai−1 )) + · · · + td0 +1 · (l(ad0 +1 ) − l(ad0 )) ≥ ti · [(l(ai ) − l(ai−1 )) + · · · + (l(ad0 +1 ) − l(ad0 ))] = ti · li (ai ) − ti · li (ad0 ). The next proposition shows that the allocation-specific frugal reward function produces the cheapest total reward with respect to a given allocation. Proposition 7 The allocation-specific frugal reward provides the cheapest individually rational envy-free total reward supporting a given allocation. Proof: Suppose that f (c) = a and a is aligned. Clearly by definition the allocation-specific frugal reward is individually rational (that is, pbi (c) − ti · l(ai ) ≥ 0). Recall that the allocation-specific frugal reward for the slowest machine is pb1 (c) = l(a1 )·t1 . This covers exactly the cost of the first machine, and thus it is the cheapest for that machine among all individually rational rewards. 19

Now, let pb0i−1 (c) be an arbitrary reward for machine i − 1. The envy-free reward to the i’th machine must be larger than pb0i−1 (c) + (l(ai ) − l(ai−1 )) · ti . Otherwise machine i would envy machine i − 1. Observe that the allocation-specific frugal reward for machine i satisfies this constraint with equality if pb0j (c) = pbj (c), where j = 1, 2, . . . , i − 1. This completes the proof.

6.3

A Tight Envy-Free Mechanism for Related Machines

Lemma 3 Let f be a deterministic ρ-approximation algorithm. There exists a ρ-approximation algorithm f 0 that outputs aligned allocations. Furthermore, if f is polynomial time computable then so is f 0 . Proof: For every input, algorithm f 0 will simulate f and then reassign the loads in a "sorted" manner to the machines. More formally, let f (c) = a. We can gradually shift from a to an aligned allocation a0 by the following procedure: Exchange loads between any two machines that violate the alignment condition in Definition 10. Similar to the classic Bubble-Sort algorithm, the shifting procedure can be implemented in polynomial-time. It remains to show that the resulting allocation a0 is a ρ-approximation. To do this, number the machines with respect to the aligned allocation a0 and Definition 11. For the two machine case m = 2: if a 6= a0 then clearly r(a) = t1 · l(a1 ) and after the exchange r(a0 ) = max{t1 · l(a2 ), t2 · l(a1 )} ≤ r(a). For the general case, observe that similarly to the previous case the makespan weakly improves in each exchange in the above procedure. Proposition 8 Any polynomial-time computable ρ-approximation deterministic algorithm with respect to the optimal makespan for the related machines model can be converted to a polynomial time computable ρ-approximation envy-free mechanism (using the allocationspecific frugal reward) in polynomial time. Proof: Let f be a polynomial-time-computable deterministic ρ-approximation algorithm. By Lemma 3 there exists a polynomial-time-computable deterministic ρ-approximation algorithm f 0 that outputs aligned allocations. By Lemma 2, f 0 coupled with the allocationspecific frugal reward yields an envy-free mechanism. Since the allocation-specific frugal reward can be computed in polynomial time this concludes the proof. Theorem 8 (Hochbaum and Shmoys [16]) There exists a polynomial-time computable (1 + ²)-approximation algorithm for minimizing the makespan on related machines for every fixed ² > 0. Putting everything together, we can state the main result of this section: Theorem 9 There exists a polynomial-time computable envy-free (1 + ²)-approximation mechanism for minimizing the makespan on related machines for every fixed ² > 0. Proof: The theorem is an immediate consequence of Theorem 8 and Proposition 8. 20

7

Research Directions

This paper formulates the fair by design approach and shows several tight bounds on envyfree mechanisms for makespan minimization. We now briefly consider few future research directions. Can we use randomization to remove the need of rewards? Specifically, is there an envy-free mechanism without money with a reasonable approximation ratio for makespan minimization? Are the determinstic and in-expectation truthful lower-bounds of any reasonable multidimensional problem always bigger than its corresponding envy-free upper bounds?

Acknowledgements I would like to thank Liad Blumrosen, Federico Echenique, Jason D. Hartline, David Kempe, John Ledyard, Debasis Mishra, Mohamed Mostagir, Mahyar Salek, Michael Schapira and anonymous referees for helpful discussions and suggestions.

References ´ Tardos. Truthful mechanisms for one-parameter agents. In IEEE [1] Aaron Archer and Eva Annual Symposium on Foundations of Computer Science (FOCS), pages 482–491, 2001. [2] Maria-Florina Balcan, Avrim Blum, and Yishay Mansour. Item pricing for revenue maximization. In ACM Conference on Electronic Commerce (EC), pages 50–59, 2008. [3] Liad Blumrosen and Noam Nisan. On the computational power of demand queries. SIAM J. Comput., 39(4):1372–1391, 2009. [4] John L. Bruno, Edward G. Coffman Jr., and Ravi Sethi. Scheduling independent tasks to reduce mean finishing time. Commun. ACM, 17(7):382–387, 1974. [5] Maurice Cheung and Chaitanya Swamy. Approximation algorithms for single-minded envy-free profit-maximization problems with limited supply. In IEEE Annual Symposium on Foundations of Computer Science (FOCS), pages 35–44, 2008. [6] George Christodoulou and Annam´aria Kov´acs. A deterministic truthful ptas for scheduling related machines. SIAM J. Comput., 42(4):1572–1595, 2013. [7] Edith Cohen, Michal Feldman, Amos Fiat, Haim Kaplan, and Svetlana Olonetsky. Envy-free makespan approximation. SIAM J. Comput., 41(1):12–25, 2012. [8] Gabrielle Demange and David Gale. The strategy structure of two-sided matching markets. Econometrica, 53(4):873–88, July 1985.

21

[9] Peerapong Dhangwatnotai, Shahar Dobzinski, Shaddin Dughmi, and Tim Roughgarden. Truthful approximation schemes for single-parameter agents. SIAM J. Comput., 40(3):915–933, 2011. [10] Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords. American Economic Review, 97(1):242–259, March 2007. [11] D. Foley. Resource allocation and the public sector. Yale Economics Essays, 7:45–98, 1967. [12] Venkatesan Guruswami, Jason D. Hartline, Anna R. Karlin, David Kempe, Claire Kenyon, and Frank McSherry. On profit-maximizing envy-free pricing. In Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1164–1173, 2005. [13] Claus-Jochen Haake, Matthias G. Raith, and Francis Edward Su. Bidding for envyfreeness: A procedural approach to n-player fair-division problems. Social Choice and Welfare, 19(4):723–749, 2002. [14] Jason D. Hartline, Sam Ieong, Michael Schapira, and Aviv Zohar. Private communication. 2008. [15] Jason D. Hartline and Qiqi Yan. Envy, truth, and profit. In ACM Conference on Electronic Commerce (EC), pages 243–252, 2011. [16] Dorit S. Hochbaum and David B. Shmoys. A polynomial approximation scheme for scheduling on uniform processors: Using the dual approximation approach. SIAM J. Comput., 17(3):539–551, 1988. [17] Ellis Horowitz and Sartaj Sahni. Exact and approximate algorithms for scheduling nonidentical processors. J. ACM, 23(2):317–327, April 1976. [18] Elias Koutsoupias and Angelina Vidali. A lower bound of 1 + φ for truthful scheduling mechanisms. Algorithmica, 66(1):211–223, 2013. [19] Daniel Lehmann, Liadan O’Callaghan, and Yoav Shoham. Truth revelation in approximately efficient combinatorial auctions. Journal of the ACM, 49(5):577–602, 2002. ´ Tardos. Approximation algorithms for [20] Jan Karel Lenstra, David B. Shmoys, and Eva scheduling unrelated parallel machines. Math. Program., 46:259–271, 1990. [21] Richard J. Lipton, Evangelos Markakis, Elchanan Mossel, and Amin Saberi. On approximately fair allocations of indivisible goods. In ACM Conference on Electronic Commerce (EC), pages 125–131, 2004.

22

[22] Pinyan Lu and Changyuan Yu. Randomized truthful mechanisms for scheduling unrelated machines. In Christos Papadimitriou and Shuzhong Zhang, editors, Internet and Network Economics, Lecture Notes in Computer Science, pages 402–413. Springer Berlin / Heidelberg, 2008. [23] Ahuva Mu’alem. On multi-dimensional envy-free mechanisms. In Francesca Rossi and Alexis Tsoukias, editors, Algorithmic Decision Theory, volume 5783 of Lecture Notes in Computer Science, pages 120–131. Springer Berlin / Heidelberg, 2009. [24] Ahuva Mu’alem and Michael Schapira. Setting lower bounds on truthfulness. In Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1143–1152, 2007. [25] R. B. Myerson. Optimal auction design. Mathematics of Operation Research, 6:58–73, 1981. [26] Noam Nisan and Amir Ronen. Algorithmic mechanism design. Games and Economic Behavior, 35:166–196, 2001. [27] Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V. Vazirani (eds.). Algorithmic Game Theory. Cambridge University Press, 2007. [28] Noam Nisan and Ilya Segal. The communication requirements of efficient allocations and supporting prices. Journal of Economic Theory, 129(1):192–224, July 2006. [29] Ariel D. Procaccia. Cake cutting: not just child’s play. Commun. ACM, 56(7):78–87, 2013. [30] Hal R. Varian. Position auctions. International Journal of Industrial Organization, 25(6):1163–1178, December 2007.

A A.1

Appendix Multidimensional Characterization: Proof of Theorem 1

Let v be an arbitrary valuation tuple. Denote by a ∈ A the allocation that f outputs for v. For simplicity, we add a special null agent, with the null valuation function v0 (S) = 0 for every bundle S. Moreover, we also assume that f allocates the empty bundle to the null agent, that is a0 = ∅. In order to construct envy-free prices we define the following finite directed graph Gf,v , − → with the vertices V (Gf,v ) = {0, 1, 2, . . . , n} and the edges E (Gf,v ) = {(i, j) | i, j ∈ V (Gf,v ), i 6= j}. Intuitively, each agent has a corresponding vertex in the graph, and each pair of distinct agents {i, j} has a directed edge from i to j and a directed edge from j to i in the graph. Note that the set of vertices V (Gf,v ) includes a vertex corresponding to the null agent. Finally, the length of the directed edge (i, j) is defined as l(i, j) = vi (ai ) − vi (aj ). 23

Definition 12 (Canonical Prices) The canonical price for agent i is pi (v) = ∆i,0 , where ∆i,0 is the length of the shortest path from vertex i ∈ N to vertex 0 in the directed graph Gf,v . Lemma 4 If f (v) is a locally-efficient bundle assignment with respect to v for every v ∈ V , then f : V → A is an envy-free achievable allocation rule. Proof: We shall verify that the canonical price induces the envy-free achievability of f . Clearly ∆i,0 ≤ l(i, 0) < ∞. We need to show that p is well defined (that is, ∆i,0 > −∞) and that M (f, p) is an envy-free mechanism. It is a well known fact that if a graph has no negative cycles then ∆i,0 > −∞, for every i ∈ N (e.g. by the correctness of the Bellman-Ford Algorithm). We first claim that every directed cycle through {1, 2, . . . n} in the graph Gf,v has a nonnegative length. Suppose not. Clearly every negative length cycle can be decomposed into simple cycles (in the sense that each vertex appears at most once in each of the simple cycles) such that at least one of them is negative. Without loss of generality assume that (1, 2, . . . , d) is a simple negative directed cycle. That is l(1, 2) + l(2, 3) + · · · + l(d, 1) < 0. Define the permutation π(1) = 2, π(2) = 3, . . . , π(d) = 1 and π(i) = i for i > d. Now: Ψ(v, a) − Ψ(v, aπ ) = l(1, 2) + l(2, 3) + · · · + l(d, 1) < 0, contradicting the local efficiency of f . Now we show that there are no negative cycles through vertex 0: Suppose that there is such a cycle. Without loss of generality assume that (0, 1, 2, . . . , d) is a simple negative directed cycle. That is l(0, 1) + l(1, 2) + l(2, 3) + · · · + l(d, 0) < 0. Then we know by the above that l(1, 2) + l(2, 3) + · · · + l(d, 1) ≥ 0. Thus, l(0, 1) + l(1, 2) + l(2, 3) + · · · + l(d, 0) < l(1, 2) + l(2, 3) + · · · + l(d, 1). Equivalently, l(0, 1) + l(d, 0) < l(d, 1). By definition, l(d, 1) = vd (ad ) − vd (a1 ), l(d, 0) = vd (ad ) ≥ 0 and l(0, 1) = 0. Putting together and rearranging we have: vd (a1 ) < 0 contradicting the the assumption that the valuation function vd is nonnegative. We show now that agent i > 0 cannot envy agent j 6= i. Suppose it does, then vi (ai ) − ∆i,0 < vi (aj ) − ∆j,0 . By rearranging we get l(i, j) + ∆j,0 = vi (ai ) − vi (aj ) + ∆j,0 < ∆i,0 . The left-hand side represents a length of a directed path from i to 0 (through j) which is strictly smaller than ∆i,0 , a contradiction to the minimality of ∆i,0 . Lemma 5 If the allocation rule f : V → A is envy-free achievable then the allocation f (v) is a locally-efficient bundle assignment with respect to v. Proof: Suppose that f is envy-free achievable. Denote by a ∈ A the allocation that f outputs for v. Let π : N → N be an arbitrary permutation. By the achievability of f , there exists a price function p such that: vi (ai ) − pi (v) ≥ vi (aπ(i) ) − pπ(i) (v), for every i ∈ N . By rearranging we get vi (ai ) − vi (aπ(i) ) ≥ pi (v) − pπ(i) (v). The local efficiency then follows from summing the inequalities over all agents. Formally: Ψ(v, a) − Ψ(v, aπ ) = Σi∈N vi (ai ) − Σi∈N vi (aπ(i) ) ≥ Σi∈N pi (v) − Σi∈N pπ(i) (v) = 0. 24

Theorem 1 follows from combining Lemma 4 and Lemma 5. We now show that the canonical prices satisfy individual rationality. Claim 2 Let f : V → A be an envy-free achievable allocation rule. Let p1 , p2 , . . . , pn be the canonical prices. Then the mechanism M (f, p) is individually rational. Moreover, pi (v) ≥ 0, for every i. Proof: By Definition 12 we have vi (ai ) = vi (ai ) − vi (a0 ) = l(i, 0) ≥ ∆i,0 = pi (v). That is, vi (ai ) ≥ pi (v) for every i ≥ 1. This shows the individual rationality. Recall that l(0, i) = 0, for every i ≥ 1. Now, pi = ∆i,0 = ∆i,0 + l(0, i) ≥ 0, since by the proof of Lemma 4 all cycles in Gf,v are nonnegative.

A.2

Proof of Claim 1

Claim 1 Given any allocation a ∈ A and values vi (aj ), i, j = 1..n, it can be decided in polynomial time whether supporting envy-free prices exist. Furthermore, if individuallyrational nonnegative envy-free prices exist, they can be computed in polynomial time. Proof: Consider the following decision algorithm: For the input pair (a, v) construct the complete bipartite graph Hv,a with vertices {1, 2, . . . , n} and {a1 , a2 , . . . , an }, where each nondirected edge (i, aj ) has weight vi (aj ). Compute a maximum weighted bipartite matching a∗ for Hv,a . Output ’yes’ if and only if Ψ(a, v) = Ψ(a∗ , v). By definition Ψ(a, v) = Ψ(a∗ , v) if and only if a is locally-efficient bundle assignment. By Theorem 1, allocation a is locally-efficient bundle assignment if and only if it can be supported with envy-free prices. This shows the correctness of the decision algorithm. Now, by Claim 2 the canonical prices pi (v) = ∆i,0 , i = 1..n, if they exist, are individually rational nonnegative envy-free prices. Let GR f,v be the reverse of the graph Gf,v , that is a complete directed graph on the same set of vertices as Gf,v , where the length of the edge R (j, i) in the graph GR f,v is l (j, i) = vi (ai ) − vi (aj ) = l(i, j). The canonical prices can be computed by solving the single source shortest paths problem, specifically, by running the Bellman-Ford algorithm on the input pair (GR f,v , 0) where the single source vertex is the vertex 0.

A.3

Proof of Proposition 1

We now briefly outline how to adjust the proof of Theorem 1 and Claims 1 and 2. Let c be an arbitrary valuation tuple. Denote by a ∈ A the allocation that f outputs for c. We define a new finite directed graph Gf,c , with vertices V (Gf,c ) = {1, 2, . . . , n} and edges − → E (Gf,c ) = {(i, j) | i, j ∈ V (Gf,c ), i 6= j}. The length of the directed edge (i, j) is defined as l(i, j) = ci (aj ) − ci (ai ). The canonical reward for agent i is pbi (c) = ∆1,i + α, where ∆1,i is the length of the shortest path from vertex 1 to vertex i ∈ N in Gf,c and α ≥ 0 is a large enough constant to be defined later. 25

(⇒) Suppose that f (c) is a locally-efficient bundle assignment with respect to c. We shall verify that the canonical reward induces the envy-free achievability of f . Clearly ∆1,i ≤ l(1, i) < ∞. We first claim that every directed cycle through {1, 2, . . . n} in the graph Gf,c has a nonnegative length. Suppose not. Without loss of generality assume that (1, 2, . . . , d) is a simple negative directed cycle. That is l(1, 2) + l(2, 3) + · · · + l(d, 1) < 0. Define the permutation π(1) = 2, π(2) = 3, . . . , π(d) = 1 and π(i) = i for i > d. Now: Ψ(c, aπ ) − Ψ(c, a) = l(1, 2) + l(2, 3) + · · · + l(d, 1) < 0, contradicting the local efficiency of f . We show now that agent i cannot envy agent j 6= i. Suppose it does, then ∆1,i + α − ci (ai ) < ∆1,j + α − ci (aj ). By rearranging we get ∆1,i + l(i, j) = ∆1,i + ci (aj ) − ci (ai ) < ∆1,j . The left-hand side represents a length of a directed path from 1 to j (through i) which is strictly smaller than ∆1,j , a contradiction to the minimality of ∆1,j . We can choose a large enough α ≥ 0 such that ∆1,i + α ≥ ci (ai ) for every i ≥ 1. Clearly, since ci (ai ) ≥ 0 we have ∆1,i + α ≥ 0. This shows that the canonical rewards are nonnegative and satisfy individual rationality. (⇐) Suppose that f is envy-free achievable. Denote by a ∈ A the allocation that f outputs for c. Let π : N → N be an arbitrary permutation. By the achievability of f , there exists a reward function pb such that: pbi (c) − ci (ai ) ≥ pbπ(i) (c) − ci (aπ(i) ), for every i ∈ N . By rearranging we get ci (aπ(i) ) − ci (ai ) ≥ pbπ(i) (c) − pbi (c). The local efficiency then follows from summing the inequalities over all agents. Formally: Ψ(c, aπ ) − Ψ(c, a) = Σi∈N ci (aπi ) − Σi∈N ci (ai ) ≥ Σi∈N pbπ(i) (c) − Σi∈N pbi (c) = 0. Now, the envy-free achievability can be decided in polynomial time by solving minimum weighted bipartite matching as follows: For the input pair (a, c) construct the complete bipartite graph Hc,a with vertices {1, 2, . . . , n} and {a1 , a2 , . . . , an }, where each non-directed edge (i, aj ) has weight ci (aj ). Compute a minimum weighted bipartite matching a∗ for Hc,a . Output ’yes’ if and only if Ψ(a, c) = Ψ(a∗ , c). The canonical rewards can be computed in polynomial time by solving the single source shortest paths problem. Specifically, by running the Bellman-Ford algorithm on the input pair (Gf,c , 1) where the single source vertex is the vertex 1, and by choosing α = max{0, c1 (a1 ) − ∆1,1 , . . . , cn (an ) − ∆1,n }.

A.4 A.4.1

Communication Bounds for Profit Maximization Lower Bounds: Proof of Propositions 1 and 2

To show the lower bounds we reduce the social welfare maximizing problem studied by [28] to the profit maximizing problem. Recall that the problem of maximizing the socialP welfare is to find an allocation that maximizes the sum of valuation (formally, argmax{a∈A} n1 vi (a)). Theorem 10 (Nisan and Segal [28]) Any algorithm for maximizing the social welfare in combinatorial auctions that achieves an approximation ratio better than 2 requires exponential communication. 26

Theorem 11 (Nisan and Segal [28]) Any algorithm for maximizing the social welfare in combinatorial auctions that achieves an approximation ratio better than n requires exponential communication. Proposition 2 Any envy-free profit-maximizing mechanism for combinatorial auctions that achieves an approximation ratio better than 2 requires exponential communication. The idea is to reduce the social welfare maximizing problem in Theorem 10 to the profitmaximizing problem. Let v = (v1 , v2 , . . . , vn ) ∈ V1 ×V2 ×· · ·×Vn = V . We introduce the set D = {d1 , d2 , . . . , dn } of n additional items and define the following valuation:   vi (S \ D) di ∈ S vbi (S) =  0 Otherwise. Observe that vbi is a valid valuation satisfying No externalities, Free disposal and Normalization. Let W ∗ (v, K) be the maximum social welfare with respect to v and the original set of items K. Let P ∗ (b v , K ∪ D) be the maximum envy-free profit with respect to vb and the extended set of items K ∪ D. Claim 3 W ∗ (v, K) ≤ P ∗ (b v , K ∪ D) for every v. P Proof: Let b be an allocation of the items in K such that n1 vi (bi ) = W ∗ (v, K). Let bd = (b1 ∪ {d1 }, . . . , bn ∪ {dn }) be an allocation based on the allocation b and the new items. It is enough to show that the maximum envy-free profit of vb using bd is at least Pn bi (S), to see the envy-freedom observe that vbi (bi ∪ {di }) − pi (bi ∪ 1 vi (bi ). Let pi (S) = v {d bi (bj ∪ {dj }) − pj (bj ∪ {dj }). Additionally, i ) − vi (bi ) = 0 ≥ 0 − vj (bj ) = v Pin}) = vi (bP n 1 vi (bi ) = 1 pi (bi ∪ {di }). Let W (v, a, K) be the social welfare of allocation a with respect to v and K. Let P (b v, b a, K ∪ D) be the maximum envy-free individually-rational profit obtainable from a locally-efficient allocation b a with respect to vb and K ∪ D. Claim 4 There exists an allocation a0 such that P (b v, b a, K ∪ D) ≤ W (v, a0 , K). Proof: To see this, let a0i = b ai \ D, observe that pi (b ai ) ≤ vbi (b ai ) ≤ vi (b ai \ D) = vi (a0i ). Proof (of Proposition): Assume, for the sake of contradiction, that there exists an envyfree profit-maximizing (2 − ²)-approximation mechanism that requires sub-exponential communication. We show that this implies a (2−²)-approximation algorithm for maximizing the social welfare in combinatorial auctions whose communication complexity is sub-exponential. For the input v and K our algorithm will simulate the envy-free profit mechanism on the input vb and K ∪ D. Let b a be the allocation selected by the envy-free profit mechanism. The output of our algorithm will be the allocation (b a1 \ D, . . . , b an \ D). 27

Clearly, the above algorithm requires sub-exponential communication. By Claims 3 and 4 we have that the above algorithm is a (2 − ²)-approximation algorithm with respect to the social welfare. Specifically: W ∗ (v, K) P ∗ (b v , K ∪ D) ≤ ≤ P (b v, b a, K ∪ D) ≤ W (v, a0 , K), 2−² 2−² where a0 = (b a1 \ D, . . . , b an \ D), this contradicts Theorem 10. Proposition 3 Any envy-free profit-maximizing mechanism for combinatorial auctions that achieves an approximation ratio better than n requires exponential communication. Proof: We can apply the proof idea of Proposition 2 to get a contradiction to Theorem 11. Observe that the number of bidders in the input v and K is identical to the number of bidders in the input vb and K ∪ D, and therefore: W ∗ (v, K) P ∗ (b v , K ∪ D) ≤ ≤ P (b v, b a, K ∪ D) ≤ W (v, a0 , K). n−² n−² A.4.2

Upper Bound: Proof of Proposition 3

We observe that a result by Guruswami et al. [12] can be applied to convert an approximation algorithm for maximizing the social welfare for general bidders into an envy-free approximation mechanism for profit maximization. Guruswami et al. [12] studied envy-free profit maximization mechanisms for unit-demand bidders.9 Applying the conversion technique to the algorithm by Lehmann et al. [19] yields an envy-free profit-maximizing mechanism for combinatorial auctions for general bidders that requires polynomial communication. Theorem 12 (Guruswami et al. [12]) There exists a polynomial time computable (2 ln(min{n, k}))-approximation algorithm for maximizing the envy-free profit of n unitdemand bidders with k items. √ Theorem 13 (Lehmann et al. [19, 3]) There exists a (min{n, c · k})-approximation algorithm for maximizing the social welfare in combinatorial auctions that requires polynomial communication, where c > 0 is some fixed constant. Proposition 4 There exists an envy-free √ mechanism for combinatorial auctions for general 0 bidders which achieves a (min{n, c · k · ln(min{n, k})})-approximation for maximizing the profit and which requires polynomial communication, where c0 > 0 is some fixed constant. Proof: Consider the following mechanism that computes two allocations and supporting prices and then selects the allocation with the higher profit: 9

We say that agent i is a unit-demand bidder if vi (S) = max{s∈S} vi ({s}), that is the agent would like to buy at most one item.

28

Step 1: allocate the whole bundle K of k items to the bidder with the highest value. The price for this bundle is maxi vi (K), while the price for an empty bundle is zero. Step 2.1: simulate the algorithm in Theorem 13. Denote by a ∈ A the allocation that the algorithm outputs for v. ∗ Step 2.2: convert a into a locally-efficient bundle assignment aπ , where π ∗ is the permu∗ tation such that Ψ(v, aπ ) ≥ Ψ(v, aπ ) for every π. Step 2.3: apply the envy-free pricing algorithm in Theorem 12 on the input v and the n ∗ ∗ indivisible bundles aπ1 , . . . , aπn . Step 3: output the allocation with the higher profit. Let P ∗ (v) be the maximum envy-free profit, and let P (v) be the profit of the mechanism. The first step provides an n-approximation for maximizing the envy-free profit. Specifically: P ∗ (v) ≤ n · max vi (K) = n · P (v). i

The pricing method of [12] requires that the input allocation has a maximum social ∗ welfare. Since aπ is locally-efficient, it is easy to see that this allocation maximizes the ∗ social welfare when each bundle aπi is treated as indivisible, and the bidders are treated as unit-demand bidders with respect to the new bundled-items. Furthermore, it can be verified ∗ Ψ(v,aπ ) that P (v) ≥ 2 ln(min{n,k}) (see the proof of Theorem 3.5 in [12]). ∗ let W (v) be the maximum social welfare. Clearly, P ∗ (v) is at most W ∗ (v), and therefore: P ∗ (v) W ∗ (v) ∗ √ ≤ √ ≤ Ψ(v, a) ≤ Ψ(v, aπ ) ≤ 2 ln(min{n, k}) · P (v). c· k c· k √ Thus, P ∗ (v) ≤ min{n, 2c · k · ln(min{n, k})} · P (v).

A.5

Item Prices

Item pricing is a special case of bundle pricing, where the price of a bundle is the total of the individual prices of the items in the bundle. Thus an agent can envy a bundle or any sub-bundle allocated to some other agent. Despite their appeal, item prices severely restrict the approximability of mechanisms. Formally: Claim 5 No individually-rational envy-free mechanism for profit maximization with supporting item prices can achieve approximation ratio better than k. Proof: Consider two bidders and k ≥ 2 identical items. Bidder 1 is a single minded bidder with v1 (S) = k − ² if |S| = k, otherwise v1 (S) = 0 for any |S| < k. Bidder 2 is a unit-demand bidder with v2 (S) = 1 if |S| ≥ 1. If bundle prices are allowed, we can allocate the k items to the first bidder. This is a locally-efficient bundle assignment, and the profit of k − ² can be extracted in an envy-free manner. However, if we require item prices the maximum profit is at most 1.

29

Claim 6 No individually-rational envy-free mechanism for minimizing the makespan with supporting item prices (on unrelated or related machines) can achieve approximation ratio better than m. Proof: To show the envy-free inapproximability, consider m identical tasks and m machines. Suppose that machine 1’s cost for each task is 1 − ², and machine i’s cost for each task is 1, where i = 2..m. Clearly, a reward of < 1 − ² for any task violates individual rationality. Additionally, a reward of > 1 for any task violates envy-freedom. The only possible envy-free allocation is to assign all the tasks to machine 1 for a reward ∈ [1 − ², 1] for each task. This gives a schedule with a makespan of m · (1 − ²), while the optimal schedule with respect to the makespan is to allocate exactly one task to each machine.

30