On the Impossibility of Black-Box Transformations in Mechanism Design
OOn the Impossibility of Black-Box Transformations in MechanismDesign
Shuchi Chawla ∗ Nicole Immorlica † Brendan Lucier ‡ Abstract
We consider the problem of converting an arbitrary approximation algorithm for a single-parameter optimization problem into a computationally efficient truthful mechanism. We askfor reductions that are black-box, meaning that they require only oracle access to the givenalgorithm and in particular do not require explicit knowledge of the problem constraints. Sucha reduction is known to be possible, for example, for the social welfare objective when the goalis to achieve Bayesian truthfulness and preserve social welfare in expectation. We show that ablack-box reduction for the social welfare objective is not possible if the resulting mechanism isrequired to be truthful in expectation and to preserve the worst-case approximation ratio of thealgorithm to within a subpolynomial factor. Further, we prove that for other objectives such asmakespan, no black-box reduction is possible even if we only require Bayesian truthfulness andan average-case performance guarantee. ∗ Department of Computer Science, University of Wisconsin - Madison. † Department of Electrical Engineering and Computer Science, Northwestern University. ‡ Department of Computer Science, University of Toronto. a r X i v : . [ c s . G T ] S e p Introduction
Mechanism design studies optimization problems arising in settings involving selfish agents, withthe goal of designing a system or protocol whereby agents’ individual selfish optimization leads toglobal optimization of a desired objective. A central theme in algorithmic mechanism design isto reconcile the incentive constraints of selfish participants with the requirement of computationaltractability and to understand whether the combination of these two considerations limits algorithmdesign in a way that each one alone does not.In the best-case scenario, one might hope for a sort of equivalence between the considerations ofalgorithm design and mechanism design. In particular, recent research has explored general reduc-tions that convert an arbitrary approximation algorithm into an incentive compatible mechanism.Ideally, these reductions have an arbitrarily small loss in performance, and are black-box in thesense that they need not understand the underlying structure of the given algorithm or problemconstraints. A big benefit of this approach is that it allows a practitioner to ignore incentive con-straints while fine-tuning his algorithm to the observed workload. Of course the feasibility of theapproach depends heavily on the objective at hand as well as the incentive requirements. The goalof this paper is to understand what scenarios enable such black-box reductions.The classic result of Vickrey, Clarke and Groves provides a positive result along these lines forsocial welfare maximization. Social welfare maximization is a standard objective in mechanismdesign. Here, a central authority wishes to assist a group of individuals in choosing from amonga set of outcomes with the goal of maximizing the total outcome value to all participants. TheVickrey, Clarke and Groves result demonstrates that for any such problem, there exists a mechanismthat maximizes social welfare with very robust incentive properties (namely, it is ex post incentivecompatible ). This construction requires that the mechanism optimize social welfare precisely, andso can be thought of as a reduction from incentive compatible mechanism design to exact algorithmdesign. This result can be extended to more general scenarios beyond social welfare. In the single-parameter setting, where the preferences of every selfish agent can be described by a single scalarparameter, (Bayesian or ex post) incentive compatibility is essentially equivalent to a per-agentmonotonicity condition on the allocation returned by the mechanism. Therefore, for objectivefunctions that are “monotone” in the sense that exact optimization of the objective leads to amonotone allocation function, there is a reduction from mechanism design to exact algorithmdesign along the lines of the VCG mechanism for social welfare.However, many settings of interest involve constraints which are computationally infeasible tooptimize precisely, and so exact algorithms are not known to exist. Recent work [12, 11, 4] showsthat in Bayesian settings of partial information, the reduction from mechanism design to algorithmdesign for social welfare can be extended to encompass arbitrary approximate algorithms witharbitrarily small loss in expected performance. These reductions work even in multi-parametersettings. Moreover, these reductions are black-box, meaning that they require only oracle accessto the prior type distributions and the algorithm, and proceed without knowledge of the feasibilityconstraints of the problem to be solved.In light of this positive result, two natural questions arise: • Are there black-box reductions transforming approximation algorithms for social welfare into ex post incentive compatible mechanisms with little loss in worst-case approximation ratio? • Does every monotone objective admit a black-box reduction that transforms approximationalgorithms into (Bayesian or ex post) incentive compatible mechanisms with little loss in1worst-case or average) approximation ratio?In this paper we answer both of these questions in the negative. Our impossibility results apply tothe simpler class of single-parameter optimization problems.The first question strengthens the demands of the reduction beyond those of the aforementionedpositive results [12, 11, 4] in two significant ways. First, it requires the stronger solution concept ofex post incentive compatibility, rather than Bayesian incentive compatibility. Second, it requiresthat the approximation factor of the original algorithm be preserved in the worst case over allpossible inputs, rather than in expectation. It is already known that such reductions are not possiblefor general multi-parameter social welfare problems. For some multi-parameter problems, no expost incentive compatible mechanism can match the worst-case approximation factors achievableby algorithms without game-theoretic constraints [16, 8]. Thus, for general social welfare problems,the relaxation from an ex post, worst-case setting to a Bayesian setting provably improves one’sability to implement algorithms as mechanisms. However, prior to our work, it was not knownwhether a lossless black-box reduction could exist for the important special case of single-parameterproblems. We show that any such reduction for single-parameter problems must sometimes degradean algorithm’s worst-case performance by a factor that is polynomial in the problem size.The second question asks whether there are properties specific to social welfare that enable thecomputationally efficient reductions of the above results [12, 11, 4]. One property that appearsto be crucial in their analysis is the linearity of the objective in both the agents’ values and thealgorithm’s allocations. For our second impossibility result we consider the (highly non-linear butmonotone) makespan objective for scheduling problems. In a scheduling problem we are given anumber of (selfish) machines and jobs; our goal is to schedule the jobs on the machines in such a waythat the load on the most loaded machine (namely, the makespan of the schedule) is minimized.The sizes of jobs on machines are private information that the machines possess and must beincentivized to share with the algorithm.Ashlagi et al. [2] showed that for the makespan objective in multi-parameter settings (thatis, when the sizes of jobs on different machines are unrelated), ex post incentive compatibilityimposes a huge cost: while constant factor approximations can be obtained in the absence ofincentive constraints, no “anonymous” mechanism can obtain a sublinear approximation ratio underthe requirement of incentive compatibility. The situation for single-parameter settings is quitedifferent. In single-parameter (a.k.a. related) settings, each machine has a single private parameter,namely its speed, and each job has a known intrinsic size; the load that a job places on a machineis its size divided by the speed of the machine. In such settings in the absence of feasibilityconstraints, deterministic PTASes are known both with and without incentives. Given this positiveresult one might expect that for single-parameter makespan minimization there is no gap betweenalgorithm design and mechanism design, at least in the weaker Bayesian setting where the goal isto achieve Bayesian incentive compatibility and match the algorithm’s expected performance. Weshow that this is not true—any black-box reduction that achieves Bayesian incentive compatibilitymust sometimes make the expected makespan worse by a factor polynomial in the problem size.Finally, while makespan is quite different from social welfare, one might ask whether there existobjectives that share some of the nice properties of social welfare that enable reductions in the styleof [11] and others. At a high level, the black-box reductions for social welfare perform“ironing”operations for each agent independently fixing non-monotonicities in the algorithm’s output in alocal fashion without hurting the overall social welfare. One property of social welfare that enablessuch an approach is that it is additive across agents. In our final result we show that even restricting2ttention to objectives that are additive across agents, for almost any objective other than socialwelfare no per-agent ironing procedure can simultaneously ensure Bayesian incentive compatibilityas well as a bounded loss in performance. The implication for mechanism design is that anysuccessful reduction must take a holistic approach over agents and look very different from thoseknown for social welfare.
Our results and techniques.
As mentioned earlier, the existence of a black-box reduction frommechanism design to algorithm design can depend on the objective function we are optimizing,the incentive requirements, as well as whether we are interested in a worst-case or average-caseperformance guarantee. We distinguish between two kinds of incentive requirements (see formaldefinitions in Section 2). Bayesian incentive compatibility (BIC) implies that truthtelling formsa Bayes-Nash equilibrium under the assumption that the agents’ value distributions are commonknowledge. The stronger notion of ex post incentive compatibility (EPIC), a.k.a. universal truth-fulness, implies that every agent maximizes her utility by truthtelling regardless of others’ actionsand the mechanism’s coin flips. For randomized mechanisms there is a weaker notion of truthful-ness called truthfulness in expectation (TIE) which implies that every agent maximizes her utilityin expectation over the randomness in the mechanism by truthtelling regardless of others’ actions.We further distinguish between social welfare and arbitrary monotone objectives, and between theaverage performance of the algorithm and its worst case performance.Table 1 below summarizes our findings as well as known results along these three dimensions.Essentially, we find that there is a dichotomy of settings: some allow for essentially lossless trans-formations whereas others suffer an unbounded loss in performance.
Objective: social welfare Objective: arbitrary monotone (e.g. makespan)Avg-case approx Worst-case approxBIC Yes [12, 4, 11] ?TIE ? No (Section 3) Avg-case approx Worst-case approxBIC No (Section 4) NoTIE No No
Table 1: A summary of our results on the existence of black-box transformations. A “yes” indicatesthat a reduction exists and gives an arbitrarily small loss in performance; a “no” indicates thatevery reduction satisfying incentive constraints suffers an arbitrarily large loss in performance.One way to establish our impossibility results would be to demonstrate the existence of single-parameter optimization problems for which there is a gap in the approximating power of arbitraryalgorithms and ex post incentive compatible algorithms. This is an important open problem whichhas resisted much effort by the algorithmic mechanism design community, and is beyond the scopeof our work. Instead, we focus upon the black-box nature of the reductions with respect to,in particular, the feasibility constraint that they face. Note that for single-parameter problems,(Bayesian or ex post) incentive compatibility is essentially equivalent to a per-agent monotonicitycondition on the allocation returned by the mechanism. We construct instances that contain “hid-den” non-monotonicities and yet provide good approximations. In order for the transformationto be incentive compatible while also preserving the algorithm’s approximation factor, it must fixthese non-monotonicities by replacing the algorithm’s original allocation with very specific kindsof “good” allocations. However, in order to determine which of these good allocations are alsofeasible the transformation must query the original algorithm at multiple inputs with the hope of3nding a good allocation. We construct the algorithm in such a way that any single query of thetransformation is exponentially unlikely to find a good allocation.
Related Work.
Reductions from mechanism design to algorithm design in the Bayesian set-ting were first studied by Hartline and Lucier [12], who showed that any approximation algorithmfor a single-parameter social welfare problem can be converted into a Bayesian incentive com-patible mechanism with arbitrarily small loss in expected performance. This was extended tomulti-parameter settings by Hartline, Kleinberg and Malekian [11] and Bei and Huang [4].Some reductions from mechanism design to algorithm design are known for prior-free settings,for certain restricted classes of algorithms. Lavi and Swamy [15] consider mechanisms for multi-parameter packing problems and show how to construct a (randomized) β -approximation mech-anism that is truthful in expectation, from any β -approximation that verifies an integrality gap.Dughmi, Roughgarden and Yan [10] extend the notion of designing mechanisms based upon ran-domized rounding algorithms, and obtain truthful in expectation mechanisms for a broad class ofsubmodular combinatorial auctions. Dughmi and Roughgarden [9] give a construction that convertsany FPTAS algorithm for a social welfare problem into a mechanism that is truthful in expectation,by way of a variation on smoothed analysis.Babaioff et al. [3] provide a technique for turning a β -algorithm for a single-valued combinatorialauction problem into a truthful β (log v max /v min )-approximation mechanism, when agent values arerestricted to lie in [ v min , v max ]. This reduction applies to single-parameter problems with downward-closed feasibility constraints and binary allocations (each agent’s allocation can be either 0 or 1).Many recent papers have explored limitations on the power of deterministic ex post incentivecompatible mechanisms to approximate social welfare. Papadimitriou, Schapira and Singer [16] gavean example of a social welfare problem for which constant-factor approximation algorithms exist,but any polytime ex post incentive compatible mechanism attains at best a polynomial approxi-mation factor. A similar gap for the submodular combinatorial auction problem was establishedby Dobzinski [8]. For the general combinatorial auction problem, such gaps have been establishedfor the restricted class of max-in-range mechanisms by Buchfuhrer et al. [5].Truthful scheduling on related machines to minimize makespan was studied by Archer andTardos [1], who designed a truthful-in-expectation 3-approximation. Dhangwatnotai et al. [7] gavea randomized PTAS that is truthful-in-expectation, which was then improved to a deterministictruthful PTAS by Christodoulou and Kov´acs [6], matching the performance of the best possibleapproximation algorithm [13]. Our work on makespan minimization differs in that we consider thegoal of minimizing makespan subject to an arbitrary feasibility constraint.A preliminary version of this work [14] proved an impossibility result for EPIC black-box re-ductions for single-parameter social welfare problems. In this paper we extend that result to applyto (the broader class of) TIE reductions. Optimization Problems.
In a single-parameter real-valued optimization problem we are givenan input vector v = ( v , v , . . . , v n ). Each v i is assumed to be drawn from a known set V i ⊆ R , sothat V = V × · · · × V n is the set of possible input vectors. The goal is to choose some allocation x ∈ F ⊆ R n from among a set of feasible allocations F such that a given objective function φ : F × V → R is optimized (i.e. either maximized or minimized, depending on the nature of the4roblem). We think of the feasibility set F and the objective function φ as defining an instance ofthe optimization problem. We will write x = ( x , x , . . . , x n ), where each x i ∈ R .An algorithm A defines a mapping from input vectors v to outcomes x . We will write A ( v ) forthe allocation returned by A as well as the value it obtains; the intended meaning should be clearfrom the context. In general an algorithm can be randomized, in which case A ( v ) is a randomvariable.Given an instance F of the social welfare problem, we will write OP T F ( v ) for the allocation in F that maximizes φ ( x , v ), as well as the value it obtains. Given algorithm A , let approx F ( A ) denotethe worst-case approximation ratio of A for problem F . That is, approx F ( A ) = min v ∈ V A ( v ) OP T F ( v ) for a maximization problem; here φ is implicit and should be clear from the context. Note that approx F ( A ) ≤ F and A .We also consider a Bayesian version of our optimization problem, in which there is publicly-known product distribution F on input vectors. That is, F = F × . . . × F n and each v i is distributedaccording to F i . Given F , the expected objective value of a given algorithm A is given by φ ( A ) = E v ∼ F [ φ ( A ( v ) , v )]. The goal of the optimization problem in this setting is to optimize the expectedobjective value. Mechanisms.
We will consider our optimization problems in a mechanism design setting with n rational agents, where each agent possesses one value from the input vector as private information.We think of an outcome x as representing an allocation to the agents, where x i is the allocation toagent i . A (direct-revelation) mechanism for our optimization problem then proceeds by elicitingdeclared values b ∈ R n from the agents, then applying an allocation algorithm A : R n → F thatmaps b to an allocation x , and a payment rule that maps b to a payment vector p . We will write x ( b ) and p ( b ) for the allocations and payments that result on input b . The utility of agent i , giventhat the agents declare b and his true private value is v i , is taken to be v i x i ( b ) − p i ( b ).A (possibly randomized) mechanism is truthful in expectation (TIE) if each agent maximizes itsexpected utility by reporting its value truthfully, regardless of the reports of the other agents, whereexpectation is taken over any randomness in the mechanism. That is, E [ v i x i ( v i , b − i ) − p i ( v i , b − i )] ≥ E [ v i x i ( b i , b − i ) − p i ( b i , b − i )] for all i , all v i , b i ∈ V i , and all b − i ∈ V − i . We say that an algorithm isTIE if there exists a payment rule such that the resulting mechanism is TIE. It is known that analgorithm is TIE if and only if, for all i and all v − i , E [ x i ( v i , v − i )] is monotone non-decreasing as afunction of v i , where the expectation is over the randomness in the mechanism.We say that a (possibly randomized) mechanism is Bayesian incentive compatible (BIC) fordistribution F if each agent maximizes its expected utility by reporting its value truthfully, giventhat the other agents’ values are distributed according to F (and given any randomness in themechanism). That is, E v − i [ v i x i ( v i , v − i ) − p i ( v i , v − i )] ≥ E v − i [ v i x i ( b i , v − i ) − p i ( b i , v − i )] for all i andall v i , b i ∈ V i , where the expectation is over the distribution of others’ values and the randomnessin the mechanism. We say that an algorithm is BIC if there exists a payment rule such thatthe resulting mechanism is BIC. It is known that an algorithm is BIC if and only if, for all i , E v − i [ x i ( v i , v − i )] is monotone non-decreasing as a function of v i . Transformations. A polytime transformation T is an algorithm that is given black-box accessto an algorithm A . We will write T ( A , v ) for the allocation returned by T on input v , given thatits black-box access is to algorithm A . Then, for any A , we can think of T ( A , · ) as an algorithmthat maps value vectors to allocations; we think of this as the algorithm A transformed by T .5e write T ( A ) for the allocation rule that results when A is transformed by T . Note that T isnot parameterized by F ; informally speaking, T has no knowledge of the feasibility constraint F being optimized by a given algorithm A . However, we do assume that T is aware of the objectivefunction φ , the domain V i of values for each agent i , and (in Bayesian settings) the distribution F over values.We say that a transformation T is truthful in expectation (TIE) if, for all A , T ( A ) is a TIEalgorithm. In a Bayesian setting with distribution F , we say that transformation T is Bayesianincentive compatible (BIC) for F if, for all A , T ( A ) is a BIC algorithm. Note that whether or not T is TIE or BIC is independent of the objective function φ and feasibility constraint F . In this section we consider the problem of maximizing social welfare. For this problem, BICtransformations that approximately preserve expected performance are known to exist. We provethat if we strengthen our solution concept to truthfulness in expectation and our performancemetric to worst-case approximation, then such black-box transformations are not possible.
The social welfare objective is defined as φ ( x , v ) = v · x .Our main result is that, for any TIE transformation T , there is a problem instance F andalgorithm A such that T degrades the worst-case performance of A by a polynomially large factor. Theorem 3.1.
There is a constant c > such that, for any polytime TIE transformation T , thereis an algorithm A and problem instance F such that approx F ( A ) approx F ( T ( A )) ≥ n c . The high-level idea behind our proof of Theorem 3.1 is as follows. We will construct an algorithm A and input vectors v and v (cid:48) such that, for each agent i in some large subset of the players, v i (cid:48) > v i but A i ( v (cid:48) ) < A i ( v ). This does not immediately imply that A is non-truthful, but we willshow that it does imply non-truthfulness under a certain feasibility condition F , namely that anyallocation is constant on the players i with v i (cid:48) > v i . Thus, any TIE transformation T must alterthe allocation of A either on input v or on input v (cid:48) . However, we will craft our algorithm in such away that, on input v , the only allocations that the transformation will observe given polynomiallymany queries of A will be A ( v ), plus allocations that have significantly worse social welfare than A ( v ), with high probability. Similarly, on input v (cid:48) , with high probability the transformation willonly observe allocation A ( v (cid:48) ) plus allocations that have significantly worse social welfare than A ( v (cid:48) ). Furthermore, we ensure that the transformation can not even find the magnitude of theallocation to players i in v (cid:48) when presented with input v , thereby preventing the transformation fromrandomizing between the high allocation of A ( v ) and an essentially empty allocation to simulatethe A ( v (cid:48) ) allocation without directly observing it. Instead, in order to guarantee that it generatesan TIE allocation rule, the transformation will be forced to assume the worst-case and offer players i the smallest possible allocation on input v . This signifcantly worsens the worst-case performanceof the algorithm A . 6 𝑆 | = 𝑟 𝑡𝑈 = 𝑆∩𝑇 | 𝑈 | = 𝑟 𝑡 | 𝑇 | = 𝑟 𝑡 | 𝑉 | = 𝑟𝑡𝑇 𝑆𝑉 𝑈 = 𝑆∩𝑇
𝒜(𝑇) = 𝒙 α 𝑇 𝑇 𝑆𝑉 𝒜(𝑆) = 𝒙 𝑆 𝒜(𝑉) = 𝒙 𝑆 𝒜(𝑈) = 𝒙 α 𝑇 (a) (b)Figure 1: (a) Visualization of typical admissible sets of bidders V , S , and T , given size parameters r and t ,and (b) the corresponding allocations of algorithm A = A V,S,T,α . In the instances we consider, each private value v i is chosen from { v, } , where 0 < v < V i = { v, } for all i ∈ [ n ]. We can thereforeinterpret an input vector as a subset y ⊆ [ n ], corresponding to those agents with value 1 (theremaining agents have value v ). Accordingly we define A ( y ), OP T F ( y ), etc., for a given subset y ⊆ [ n ]. Also, for a ≥ y ⊆ [ n ], we will write x ay for the allocation in which each agent i ∈ y is allocated a , and each agent i (cid:54)∈ y is allocated 0. Feasible Allocations.
We now define a family of feasibility constraints. Roughly speaking, wewill choose α, γ ∈ (0 ,
1) with γ < α and sets
S, T ⊆ [ n ] of agents. The feasible allocations will be x γ [ n ] , x S , and x αT . That is, we can allocate γ to every agent, 1 to all agents in S , or α to all agentsin T . We will also require that S and T satisfy certain properties, which essentially state that S and T are sufficiently large and have a sufficiently large intersection.More formally, define parameters γ ∈ (0 , α ∈ (0 , r ≥
1, and t ≥ n ), such that t (cid:29) r (cid:29) γ − (cid:29) α − , r t ≤ n , and tγn (cid:28)
1. We think of t as abound on the size of “small” sets, and we think of r as a ratio between the sizes of “small” and“large” sets.Suppose that V , S , and T are subsets of [ n ]. We say that the triple V , S , T is admissible if thefollowing conditions hold:1. | S | = | T | = r t ,2. | S ∩ T | = r t ,3. V ⊂ S ∩ T , and,4. | V | = rt .In general, for a given admissible V , S , and T , we will tend to write U = S ∩ T for notationalconvenience. See Figure 1(a) for an illustration of the relationship between the sets in an admissibletriple. In order to hide the feasibility constraint F from the transformation, we will pick the sets7 , S , and T uniformly at random from all admissible triples, and the value α from an appropriaterange. For each admissible tuple V , S , T , and value α , we define a corresponding feasibilityconstraint F V,S,T,α = { x S , x αT , x γ [ n ] } . Note that F V,S,T,α does not depend on V ; we include set V purely for notational convenience. Weremark that all of the feasible allocations allocate the same amount to agents in U .Recall that agents have values chosen from { v, } . We will choose v = tγn , where we recall thatour parameters have been chosen so that tγn (cid:28) The Algorithm.
We now define the algorithm A V,S,T,α corresponding to an admissible tuple
V, S, T and value α . We think of A V,S,T,α as an approximation algorithm for the social welfareproblem F V,S,T,α and later show that there is no TIE transformation of A V,S,T,α without a significantloss in worst-case approximation for some value of α .Given y ⊂ [ n ], we define n T ( y ) = | y ∩ T | + | y ∩ U | and n S ( y ) = | y ∩ S | + 2 | y ∩ V | . That is, n T ( y ) is the number of elements of y that lie in T , with elements of U counted twice.Likewise, n S ( y ) is the number of elements of y that lie in S , with elements of V counted thrice.The algorithm A V,S,T,α is then described as Algorithm 1.
Algorithm 1:
Allocation Algorithm A V,S,T,α
Input : Subset y ∈ [ n ] of agents with value 1 Output : An allocation x ∈ F V,S,T,α if n S ( y ) ≥ t , n S ( y ) ≥ γ | y | , and n S ( y ) ≥ n T ( y ) then return x S else if n T ( y ) ≥ t , n T ( y ) ≥ γ | y | , and n T ( y ) ≥ n S ( y ) then return x αT else return x γ [ n ] end In this section, we derive the key lemmas for the proof of Theorem 3.1. First, we bound theapproximation factor of algorithm A V,S,T,α for problem F V,S,T,α . Lemma 3.2. approx F V,S,T,α ( A V,S,T,α ) ≥ α/ .Proof. Choose y ⊆ [ n ] and consider the three cases for the output of A V,S,T,α . Case 1: n S ( y ) ≥ t , n S ( y ) ≥ γ | y | , and n S ( y ) ≥ n T ( y ) . Our algorithm returns allocation x S and obtains welfare at least | y ∩ S | . Note that | y ∩ S | ≥ n S ( y ) ≥ t | y ∩ S | ≥ n S ( y ) ≥ n T ( y ) ≥ | y ∩ T | . The allocation x αT obtains welfare at most α ( | y ∩ T | + | T \ y | v ) ≤ α ( | y ∩ T | + nvγ ) ≤ | y ∩ T | + t ≤ | y ∩ S | . Note that here we used | T | ≤ nγ , which follows since r > γ − .The allocation x γ [ n ] obtains welfare at most γ | y | + t ≤ n S ( y ) ≤ | y ∩ S | . So we obtain at leasta 1 / Case 2: n T ( y ) ≥ t , n T ( y ) ≥ γ | y | , and n T ( y ) ≥ n S ( y ) . Our algorithm returns allocation x αT and obtains welfare at least α | y ∩ T | . The same argument as case 1 shows that our approximationfactor is at least α/ Case 3: n S ( y ) ≤ t and n T ( y ) ≤ t . Our algorithm returns allocation x γ [ n ] for a welfare of atleast γ ( | y | + v ( n − | y | )) ≥ t . The allocation x S obtains welfare at most | y ∩ S | + t ≤ n S ( y ) + t ≤ t ,and allocation x αT obtains welfare at most 2 αt ≤ t . So our approximation factor is at least 1 / Case 4: n S ( y ) ≤ γ | y | and n T ( y ) ≤ γ | y | . Our algorithm returns allocation x γ [ n ] for a welfare of atleast γ ( | y | + v ( n −| y | )) ≥ γ | y | . The allocation x S obtains welfare at most | y ∩ S | + t ≤ n S ( y ) ≤ γ | y | ,and allocation x αT obtains welfare at most | y ∩ T | + t ≤ αγ | y | ≤ γ | y | . So our approximation factoris at least 1 / A (cid:48) is any algorithm for problem F V,S,T,α that is TIE. We will show that A (cid:48) is then very restricted in the allocations it can return on inputs y = V and y = U . Furthermore,we note that if A (cid:48) has a good enough approximation ratio, then its allocations on inputs y = V and y = U are restricted further still. In particular, the optimal allocation on both V and U is x S ;so to obtain a good approximation factor, on both U and V , the algorithm should allocate a largeenough amount to agents in U . As any TIE transformation of A is itself an algorithm for problem F V,S,T,α , these observations will later play a key role in our impossibility result.
Claim 3.3.
Suppose A (cid:48) is a truthful-in-expectation algorithm for problem F V,S,T,α . Then the ex-pected allocation to each agent in U must be at least as large in A (cid:48) ( U ) as in A (cid:48) ( V ) .Proof. Take any set W with V ⊆ W ⊆ U , | W | = | V | +1. Then, on input W , the expected allocationto the agent in W \ V must not decrease. Since all allocations are constant on U , this means thatthe expected allocation to each agent in U must not decrease. By the same argument, A (cid:48) returnsan allocation at least this large for all W such that V ⊆ W ⊆ U , and in particular for W = U .In light of these claims, our strategy for proving Theorem 3.1 will be to show that a polytimetransformation T is unlikely to encounter the allocation x αT during its sampling when the input is V , given that the sets V , S , and T are chosen uniformly at random over all admissible tuples. Thismeans the transformation will be unable to learn the value of α . This is key since it prevents thetransformation from using the value of α to appropriately randomize between the allocation of x S and the essentially empty allocation of x γ [ n ] to acheive an effective allocation of α for agents in U oninput V thereby satisfying the conditions of Claim 3.3. Similarly, a transformation is unlikely toencounter the allocation x S during its sampling on input U , and therefore can not satisfy Claim 3.3by allocating 1 to agents in U on input U . Lemma 3.4.
Fix V and S satsifying the requirements of admissibility. Then for any y ⊆ [ n ] , Pr T [ A V,S,T,α ( y ) = x αT ] ≤ e − O ( tr +1 ) , with probability taken over all choices of T that are admissiblegiven V and S . roof. Fix any y . Write n V = | y ∩ V | , n S = | y ∩ ( S − V ) | , and n ∗ = | y ∩ ([ n ] − S ) | . Notethat | y | = n V + n S + n ∗ . Define the random variables m U and m T by m U = | y ∩ ( U − V ) | and m T = | y ∩ ( T \ S ) | .The event [ A V,S,T,α ( y ) = x αT ] occurs precisely if the following are true: m T + 2 n V + 2 m U ≥ t, (1) m T + 2 n V + 2 m U ≥ γ ( n V + n S + n ∗ ) , (2) m T + 2 n V + 2 m U ≥ n S + 3 n V . (3)We will show that the probability of these three inequalities being true is exponentially small.To see this, note that (3) implies that m T + 2 m U ≥ n V . Thus, (1) implies that m T + 2 m U ≥ t/ m T + m U ≥ t/
6. Now each element of y counted in n S will count toward m U withprobability r +1 , and each element of y counted in n ∗ will count toward m T with probability r +1 .Since t (cid:29) r , Chernoff bounds imply that with probability at least 1 − e − O ( t/r ) , we will have n ∗ + n S ≥ r ( m T + m U ). Then m T + 2 n V + 2 m U n V + n S + n ∗ < m T + 2 m U ) n S + n ∗ < r (cid:28) γ contradicting (2). Lemma 3.5.
Fix U and T satsifying the requirements of admissibility. Then for any y ⊆ [ n ] , Pr S [ A V,S,T,α ( y ) = x S ] ≤ e − O ( tr +1 ) , with probability taken over all choices of V and S that areadmissible given U and T .Proof. Fix any y . Write n U = | y ∩ U | , n T = | y ∩ ( T − U ) | , and n ∗ = | y ∩ ([ n ] − T ) | . Note that | y | = n U + n T + n ∗ . Define the random variables m V and m S by m V = | y ∩ V | and m S = | y ∩ ( S \ T ) | .The event [ A V,S,T,α ( y ) = x S ] occurs precisely if the following are true: m S + n U + 2 m V ≥ t, (4) m S + n U + 2 m V ≥ γ ( n U + n T + n ∗ ) , (5) m S + n U + 2 m V ≥ n T + 2 n U . (6)We will show that the probability of these three inequalities being true is exponentially small.To see this, first note that we can assume n T = 0, as this only loosens the requirements of theinequalities. We then have that (6) implies m S + 2 m V ≥ n U . Thus, (4) implies that m S + 2 m V ≥ t/
2, and hence m S + m V ≥ t/
4. Now each element of y counted in n U will count toward m V with probability r , and each element of y counted in n ∗ will count toward m S with probability r +1 . Since t (cid:29) r , Chernoff bounds imply that with probability at least 1 − e O ( t/r ) , we will have n ∗ + n T ≥ r ( m S + m V ). Then m S + n U + 2 m V n U + n ∗ < m S + m V ) n U + n ∗ < r (cid:28) γ contradicting (5). 10 .4 Proof of Main Theorem We can now set our parameters t , r , α , and γ . We will choose t = n / , r = n / , and γ = n − / .The values of α we will be considering are 1 and n − / . Note that t (cid:29) r (cid:29) γ − (cid:29) α − for eachchoice of α . Note also that v = tγ − /n = n − / (cid:28) α on input V (by Lemma 3.4), and since it can not find the “good” allocation of x S oninput U (by Lemma 3.5), it must be pessimistic and allocate the minimum possible value of α toagents in V on input V in order to guarantee that the resulting allocation rule is TIE (by Claim3.3). This implies a bad approximation on input V and hence a bad worst-case approximation. Proof of Theorem 3.1 :
For each admissible
V, S, T and α ∈ { , n − / } , write A (cid:48) V,S,T,α for T ( A V,S,T,α ). Lemma 3.5 implies that, with all but exponentially small probability, A (cid:48) V,S,T,α will notencounter allocation x S on input U . Thus, on input U , it can allocate at most α to each agent in U in expectation (using the fact that α > γ ). Then, since A (cid:48) V,S,T,α is incentive compatible, Claim3.3 implies that A (cid:48) V,S,T,α must allocate at most α to each agent in U on input V .Now Lemma 3.4 implies that, with all but exponentially small probability, A (cid:48) V,S,T,α will notencounter allocation x αT on input V , and thus is unaware of the value of α on input V . Thus, toensure incentive compatibility, A (cid:48) V,S,T, ( V ) must allocate at most n − / to each agent in U . Ittherefore obtains a welfare of | V | n − / + t ≤ n / + n / < n / , whereas a total of | V | = n / is possible with allocation x S . Thus A (cid:48) V,S,T, has a worst-case approximation of n − / , whereas A V,S,T, has an approximation factor of 1 / F .We would like to prove that, when agents’ values are drawn according to a distribution F , anyTIE transformation necessarily degrades the average welfare of some algorithm by a large factor.The difficulty with extending our techniques to this setting is that a transformation may usethe distribution F to “guess” the relevant sets V and U (i.e. if the distribution is concentratedaround the sets V and U in our construction). One might hope to overcome this difficulty in ourconstruction by hiding a “true” set V (that generates a non-monotonicity) in a large sea of setsthat could potentially take the role of V . Then, if the transformation is unlikely to find a goodallocation on input U , and unlikely to determine the value of α on any of these potential sets, andis further unable to determine which set is the “true” V , then it must be pessimistic and allocatethe minimum potential value of α on any of these potential sets in order to guarantee truthfulness.Unfortunately, our construction assumes that all allocations are constant on U , and this makes itdifficult to hide a set V while simultaneously making it difficult to discover a good allocation oninput U . We feel that, in order to make progress on this interesting open question, it is necessaryto remove the assumption that all allocations are constant on U which, in hand, seems to make itmuch more difficult to derive necessary conditions for a transformation to be TIE. We now consider an objective function, namely makespan, that differs from the social welfareobjective in that it is not linear in agent values or allocations. Informally we show that black-box reductions for approximation algorithms for makespan are not possible even if we relax thenotion of truthfulness to Bayesian incentive compatibility and relax the measure of performance to11xpected makespan, where both the notions are defined with respect to a certain fixed and known distribution over values. As in the previous section, our impossibility result hinges on the fact thatthe transformation is not aware of the feasibility constraint that an allocation needs to satisfy andcan learn this constraint only by querying the algorithm at different inputs.
We consider the following minimization problem in a Bayesian setting. In this problem n selfishmachines (a.k.a. agents) are being allocated jobs. Each agent has a private value v i representingits speed. If machine i is allocated jobs with a total length x i , the load of machine i is x i /v i . Themakespan of allocation x to machines with speeds v is the maximum load of any machine: φ ( x , v ) = max i x i v i . An instance of the (Bayesian) makespan problem is given by a feasibility constraint F and adistribution over values F ; the goal is to map every value vector to allocations so as to minimizethe expected makespan: E v ∼ F [ φ ( x ( v ) , v )] subject to x ( v ) ∈ F for all v . Given an algorithm A , we use φ ( A ) to denote its expected makespan.Our main result is the following: Theorem 4.1.
Let n be large enough and T be any black-box BIC transformation that makesat most e n / / black-box queries to the given algorithm on each input. There exists an instance ( F , F ) of the makespan problem and a deterministic algorithm A such that T ( A ) either returns aninfeasible allocation with positive probability, or has makespan φ ( T ( A )) = Ω( n / ) φ ( A ) . Here F isthe uniform distribution over { , α } n for an appropriate α and is known to T . We note that the algorithm A in the statement of Theorem 4.1 is deterministic. A BIC transfor-mation T must therefore degrade the makespan of some algorithms by a polynomially large factoreven when we limit ourselves to deterministic algorithms. For simplicity of exposition, we provea gap of Ω( n / ), however, our construction can be tweaked to obtain a gap of Ω( n / − δ ) for any δ > Problem Instance.
We now describe the problem instance ( F , F ) in more detail. Let α < n / be a parameter to be determined later. As mentioned earlier, F is the uniform distribution over { , α } n . That is, every value (i.e. speed) v i is 1 or α with equal probability. There are 2 n jobs in all, n of length α and n of length 1. Our feasibility constraint will have the property that each machinecan be assigned at most one job. So a valid allocation will set the allocation to each machine to avalue in { , , α } . Of all such allocations (i.e. all x ∈ { , , α } n ), all but one will be feasible. This one infeasibleallocation is thought of as a parameter of the problem instance. Given x ∈ { , , α } n , we will write F x as the set { , , α } n \ x , and Γ( x ) = ( F x , F ) as the corresponding problem instance in which x is infeasible. We will use x bad to denote the forbidden allocation in the remainder of this section. A makespan assignment must allocate each job to a machine, but we will sometimes wish to specify an allocationin which not all jobs are allocated. For ease of exposition, we will therefore assume that there is an extra agent withvalue n ( α + 1); this agent will always be allocated all jobs not allocated to any other agent. Note that the load ofthis machine is always at most 1. he algorithm. We will first describe a randomized algorithm A ( x bad ) (Algorithm 2 below)which we think of as an approximation algorithm for problem instance Γ( x bad ). Algorithm 2:
Allocation Algorithm A ( x bad ) Input : Vector v ∈ { , α } n Output : An allocation x (cid:54) = x bad H ← { i : v i = α } ; if n − n / ≤ | H | ≤ n + n / then Choose set S ⊂ H with | S | = n − / | H | uniformly at random; for i ∈ S do x i ← α ; for i ∈ H \ S do x i ← for i (cid:54)∈ H do x i ← else Choose set S ⊂ [ n ] with | S | = n / uniformly at random; for i ∈ S do x i ← α ; for i (cid:54)∈ S do x i ← end if x = x bad then go to line return x We first note that A ( x bad ) must terminate. Claim 4.2.
For all v , algorithm A ( x bad ) terminates with probability .Proof. This follows from noting that at least two distinct allocations can be chosen by A ( x bad ) oneach branch of the condition on line 2, so A ( x bad ) must eventually choose an allocation that is not x bad .We now use A ( x bad ) to define a set of deterministic algorithms . Let D ( x bad ) (or D for short)denote the set of deterministic algorithms in the support of A ( x bad ). That is, for every A ∈ D , A ( v )is an allocation returned by A ( x bad ) on input v with positive probability for every v ∈ { , α } n .Moreover, for every combination of allocations that can be returned by A ( x bad ) on each inputprofile, there is a corresponding deterministic algorithm in D .For any v , let H ( v ) = { i : v i = α } be the set of high speed agents. Let C denote the eventthat n − n / ≤ | H ( v ) | ≤ n + n / , over randomness in v . We think of C as the event thatthe number of high-speed agents is concentrated around its expectation. We note the followingimmediate consequence of Chernoff bounds. Observation 4.3. Pr v [ C ] ≥ − e − n / / . This allows us to bound the expected makespan of each
A ∈ D . Lemma 4.4.
For each
A ∈ D , φ ( A ) ≤ αe − n / / where the expectation is taken over v . More precisely, we will define a set of deterministic allocation rules that map type profiles to allocations; inparticular we will not be concerned with implementations of these allocation rules. roof. If event C occurs, then A returns an allocation in which each agent i with v i = 1 receivesallocation at most 1. Since each agent with v i = α also receives allocation at most α , we concludethat if event C occurs then the makespan of A ( v ) is at most 1. Otherwise, the makespan of A ( v ) istrivially bounded by α . Since Observation 4.3 implies that this latter case occurs with probabilityat most 2 e − n / / , the result follows. We now present a proof of Theorem 4.1. Let T denote a BIC transformation that can make at most e n / / black-box queries to an algorithm for makespan. Write T ( A ) for the mechanism inducedwhen T is given black-box access to an algorithm A .We first note that if T ( A ) returns only feasible allocations with probability 1, then it can onlyreturn an allocation that it observed during black-box queries to algorithm A . This is true even ifwe consider only algorithms of the form A ∈ D ( x bad ) for some choice of x bad . Claim 4.5.
Suppose that for some
A ∈ D , with positive probability, T returns an allocation notreturned by a black-box query to A . Then there exists an algorithm A (cid:48) such that T A (cid:48) returns aninfeasible allocation with positive probability.Proof. Suppose that with positive probability T A returns the allocation x (cid:48) on A ∈ D ( x ) withoutencountering it in a black box query to the algorithm A . Then there exists A (cid:48) ∈ D ( x ) such that A (cid:48) and A agree on each input queried by T in some instance where T A returns x (cid:48) , and furthermore A (cid:48) never returns allocation x (cid:48) on any input. Note, then, that T A (cid:48) also returns allocation x (cid:48) withpositive probability. But A (cid:48) ∈ D ( x (cid:48) ), so if we set x bad = x (cid:48) then we conclude that T A (cid:48) returns theinfeasible allocation x (cid:48) with positive probability.In the remainder of the analysis we assume that T only returns observed allocations. We willnow think of x bad as being fixed, and A as being drawn from D ( x bad ) uniformly at random. Let B denote the (bad) event, over randomness in v , T , and the choice of A ∈ D , that T ( A ) returns anallocation with makespan α . Our goal will be to show that if T ( A ) is BIC for every A ∈ D , thenPr[ B ] must be large; this will imply that the expected makespan of T ( A ) will be large for some A ∈ D .Intuitively, the reason that an algorithm A ∈ D may not be truthful is because low-speedagents are often allocated 1 while high-speed agents are often allocated 0. In order to fix such non-monotonicities, T must either increase the probability with which 1 is allocated to the high-speedagents, or increase the probability with which 0 is allocated to the low-speed agents. To this end,let U ( v ) be the event that, on input v , T ( A ) returns an allocation x in which at least n / agentssatisfy v i = 1 and x i = 0.As the following lemma shows, the event U ( v ) is unlikely to occur unless B also occurs. Then tofix the non-monotonicity while avoiding B , T must rely on allocating 1 more often to the high-speedagents. However this would require T to query A on speed vectors v (cid:48) that are near-complementsof v , and in turn imply a large-enough probability of allocating α to low-speed agents, i.e. event B . We now make this intuition precise. Lemma 4.6.
For each v , Pr[ U ( v ) ∧ ¬ B | v ] < (ln 4) e − n / / .Proof. Fix input v , and suppose that event U ( v ) ∧ ¬ B occurs. Recall that T only returns anallocation that A outputs on a query v (cid:48) . We will refer to a query of A on input v (cid:48) as a successful x that satisfies U ( v ) ∧ ¬ B . Let us bound the probability of a single querybeing successful. Let t denote the number of agents with v i = 1. Then U ( v ) implies t ≥ n / .First, suppose that v (cid:48) does not satisfy the event C , that is, the number of high speed agents in v (cid:48) is far from its mean n/
2. Then each of the t agents has an n − / probability of being allocated α (taken over the choice of A from D ). The probability that none of the t agents is allocated aload of α is at most (1 − n − / ) t < e − n / .On the other hand, suppose that v (cid:48) satisfies the event C . Then U ( v ) implies that at least t (cid:48) ≥ n / agents satisfy v i = 1 and v i (cid:48) = α . In this case, each of these t (cid:48) agents has probability n − / of being allocated α . The probability that none of them is allocated a load of α is at most(1 − n − / ) t (cid:48) < e − n / .In either case, the probability that a single query is successful is at most e − n / . Transformation T can make at most e n / / v . We will now bound the probability that any oneof them is successful. First note that since A is deterministic, we can assume that T does notquery A more than once on the same input. Furthermore, we can think of the choice of A ∈ D asindependently selecting the behaviour of A for each input profile, so that the allocations returnedby A on different input profiles are independent with respect to the choice of A from D . We cantherefore think of the e n / / e − n / . Thus, the probability that at least one of these queries is successful is at most1 − (1 − e − n / ) e n / / = 1 − (1 − e − n / ) e n / e − n / / < − (1 / e − n / / = 1 − (1 /e ) (ln 4) e − n / / < (ln 4) e − n / / as required.We now consider the specific probabilities with which T returns high or low allocations on highor low values. For agent i , value v , and allocation x , we will write p xi ( v ) = Pr v − i , A , T [ x i ( v, v − i ) = x ].That is, p xi ( v ) is the probability that conditioned on agent i ’s value being v , T ( A ) allocates x tothe agent; Here the probability is over the values of the other agents, any randomness in T , andthe choice of A ∈ D . Observation 4.7. (cid:80) i p xi ( v ) = 2 (cid:80) i Pr v [( v i = v ) ∧ ( x i ( v ) = x )] . We can express the fact that T satisfies BIC in terms of conditions on these probabilities(Lemma 4.8 below): either p i (1) should be large, i.e. low-speed agents get a low allocation, or oneof p i ( α ) and p αi ( α ) should be large, i.e. high-speed agents get a high allocation. On the other hand,in Lemmas 4.9, 4.10, and 4.11 we show that on average over all agents, each of these probabilitiesis small if the probability of the bad event B is small. The proofs of these lemmas are deferred tothe end of this section. In Lemma 4.12 we put these results together to argue that B occurs witha large probability. Lemma 4.8.
Let x be the allocation rule of T ( A ) . Then if x i (1) ≤ x i ( α ) , p i (1) < / , and p i ( α ) < / , then p αi ( α ) > /α . roof. Since p i (1) < , we have x i (1) > . Since p i ( α ) < , we have x i ( α ) < + p αi ( α ) α . Weconclude that < + p αi ( α ) α which implies the desired result. Lemma 4.9. n (cid:80) i p αi ( α ) ≤ n − + Pr[ B ] n − / + 2 e − n / / n − / + (ln 4) e − n / / n − / . Lemma 4.10. n (cid:80) i p i (1) ≤ n − / + Pr[ B ] + (ln 4) e − n / / . Lemma 4.11. n (cid:80) i p i ( α ) ≤ n − / + Pr[ B ] + (ln 4) e − n / / + 2 e − n / / . The above lemmas put together give a lower bound for Pr[ B ]. Set α = n / . Lemma 4.12.
Pr[ B ] ≥ n − / − e − n / / − (ln 4) e − n / / .Proof. We know that, for each i , either p i (1) ≥ / p i ( α ) ≥ /
3, or p αi ( α ) > /α . So one of theseinequalities must be true for at least one third of the agents, and hence one of the following mustbe true: 12 n (cid:88) i p i (1) ≥ / n (cid:88) i p i ( α ) ≥ / n (cid:88) i p αi ( α ) ≥ / α. Suppose the first inequality is true. Then by Lemma 4.10 we knowPr[ B ] ≥ / − n − / − (ln 4) e − n / / which implies the desired result for sufficiently large n (as the right hand side is at least a constantfor large n , whereas n − / − e − n / / − (ln 4) e − n / / , from the statement of the lemma, vanishesas n grows).Suppose the second inequality is true. Then we know from Lemma 4.11Pr[ B ] ≥ / − n − / − (ln 4) e − n / / − e − n / / which again implies the desired result.Finally, suppose the third inequality is true. Then we know from Lemma 4.9 n − + Pr[ B ] n − / + 2 e − n / / n − / + (ln 4) e − n / / n − / ≥ / α which implies (recalling α = n / )Pr[ B ] ≥ (2 n − − n − ) n / − e − n / / − (ln 4) e − n / / = n − / − e − n − / / − (ln 4) e − n / / as required.We can now prove our main result. 16 roof of Theorem 4.1. Write Pr[ B | A ] for the probability of event B given that T is given black-box access to algorithm A , with probability over the choice of input profile v . Choose A (cid:48) ∈ argmax A∈ D { Pr[ B | A ] } . Then in particular Pr[ B | A ] ≥ Pr[ B ], where recall that Pr[ B ] is theprobability of event B when A is chosen uniformly at random from D .Recall that we set α = n / . By Lemma 4.4 the expected makespan of A (cid:48) is at most 1 +2 αe − n / / = 1 + n / e − n / / < n . Using Lemma 4.12, the expected makespanof T A (cid:48) is at least1 + α Pr[ B |A (cid:48) ] ≥ α Pr[ B ] ≥ α ( n − / − e − n / / − (ln 4) e − n / / )= 1 + 14 n / − n / e − n / / − (cid:18) ln 44 (cid:19) n / e − n / / ≥ n / as required. Proofs of bounds on the allocation probabilities.
To conclude the analysis, we now presentproofs of Lemmas 4.9, 4.10, and 4.11.
Proof of Lemma 4.9.
Let us first condition on the event ¬ B ∧ ¬ U ( v ) ∧ C . That is, we consider theoutput of T ( A ) on a value vector v that satisfies the concentration event C , and further assumethat the makespan of T ( A ) is small ( ¬ B ) and few agents with v i = 1 have a 0 allocation ( ¬ U ( v )). C implies that many of the agents (at least n − n / ) in v are low-speed agents. Along with ¬ B and ¬ U ( v ) this implies that most of these agents have an allocation of 1; Call this set of agents L .Then | L | > n − n / . In particular L is non-empty. Now suppose that T ( A ) returns an allocation x that is returned by A on input v (cid:48) . Then, since L is non-empty, v (cid:48) must satisfy the condition online 2 of A ( x bad ) (as this is the only way in which any agent can be allocated 1, regardless of thechoice of A ∈ D ). This implies that at most n / agents get an allocation of α because S is of sizeat most n / .We conclude that for fixed v satisfying C and conditioning on ¬ B ∧ ¬ U ( v ),1 n (cid:88) i Pr[( v i = α ) ∧ ( x i ( α, v − i ) = α )] ≤ n n = n − . For the case that events B , ¬ C , or U ( v ) occur, we note that every allocation returned by T ( A )allocates α to at most n / agents. So, conditioning on either of these events, we have1 n (cid:88) i Pr[( v i = α ) ∧ ( x i ( α, v − i ) = α )] ≤ n − / . We conclude, taking probabilities over all v , that12 n (cid:88) i p i (1) ≤ n − + Pr[ B ] n − / + Pr[ ¬ C ] n − / + Pr[ U ( v ) ∧ ¬ B ] n − / ≤ n − + Pr[ B ] n − / + 2 e − n / / n − / + (ln 4) e − n / / n − / as required. 17 roof of Lemma 4.10. For each input vector v , either event ¬ U ( v ) occurs, event B occurs, or event U ( v ) ∧ ¬ B occurs. Event ¬ U ( v ) by definition gives us a bound on the number of agents with value1 that receive an allocation of 0. So conditioning on this event (and keeping v fixed) we have1 n (cid:88) i Pr[( v i = 1) ∧ ( x i (1 , v − i ) = 0)] ≤ n n / = n − / . Thus, taking probabilities over all v , we have12 n (cid:88) i p i (1) ≤ n − / + Pr[ B ] + Pr v [ U ( v ) ∧ ¬ B | v ] ≤ n − / + Pr[ B ] + (ln 4) e − n / / as required. Proof of Lemma 4.11.
Let us first fix v and condition on the event ¬ B ∧ ¬ U ( v ) ∧ C . Suppose that T ( A ) returns an allocation x that is returned by A on input v (cid:48) .As in the proof of Lemma 4.9, event ¬ B ∧ ¬ U ( v ) ∧ C implies that T ( A ) returns an allocationin which some agents are allocated 1. We therefore conclude that v (cid:48) satisfies the condition on line2 of A ( x bad ). This implies − n / ≤ | H ( v ) | − | H ( v (cid:48) ) | ≤ n / . Furthermore, ¬ U ( v ) ∧ ¬ B implies that | H ( v (cid:48) ) \ H ( v ) | ≤ n / . Combining with the inequalitiesabove, we conclude that | H ( v ) \ H ( v (cid:48) ) | ≤ n / + n / . Note that this is a bound on the numberof agents such that v i = α and v i (cid:48) = 1, which is also a bound on the number of agents such that v i = α and x i = 1.We conclude that, conditioning on event ¬ B ∧ ¬ U ( v ) ∧ C and keeping v fixed, we have1 n (cid:88) i Pr[( v i = α ) ∧ ( x i ( α, v − i ) = 1)] ≤ n (3 n / ) . Thus, taking probabilities over all v and all events, we have12 n (cid:88) i p i (1) ≤ n − / + Pr[ B ] + Pr v [ U ( v ) ∧ ¬ B | v ] + Pr[ ¬ C ] ≤ n − / + Pr[ B ] + (ln 4) e − n / / + 2 e − n / / as required. In Section 4 we showed that no BIC approximation-preserving transformations are possible forthe makespan objective. One of the properties of the social welfare objective that allows a BICtransformation where one cannot exist for makespan is that the objective function is additive acrossagents. This allows a transformation to focus on each agent individually while taking an aggregateview over other agents and preserving the performance with respect to the respective componentof the objective function alone. [12] and [11] formalize this idea as follows: for each agent i theyconstruct a mapping g i from the value space of i to itself, and on input v return the output of thealgorithm on g ( v ) = ( g ( v ) , g ( v ) , · · · ). The mappings g i ensure the following three properties:18P.1) the mapping preserves the distribution over values of i ,(P.2) the expected allocation of agent i upon applying the mapping, i.e. x i ( g i ( v i )), is monotonenon-decreasing in v i , and,(P.3) the contribution of agent i to the overall social welfare is no worse than in the originalalgorithm.The benefit of this approach is that if each agent’s value space is single-dimensional or well struc-tured in some other way, the computational problem of finding such a mapping becomes easy.Given this construction, it is natural to ask whether there are other objectives that are additiveacross agents and for which such a per-agent ironing procedure works. We show in this section thatfor almost any objective other than social welfare, such an approach cannot work. In particular,given an objective satisfying some mild properties, we construct an instance such that for anymapping g i from agent i ’s value space to itself that satisfies properties (P.1) and (P.2) above,property (P.3) fails to hold by an arbitrarily large factor.We focus first on maximization problems. Note that for an objective function of the formmax E v [ (cid:80) i x i ( v ) h i ( v i )] where h i s are non-decreasing functions, the approach of [12] and [11] worksas-is to give a BIC approximation preserving transformation. In the sequel, we consider objectivesof the form max E v [ (cid:80) i h i ( x i ( v )) v i ] where h i is an arbitrary non-linear continuous function. Theorem 5.1.
Consider the objective max E v [ (cid:80) i h i ( x i ( v )) v i ] where each h i is an arbitrary in-creasing function. Suppose that there exists an agent i for which h i is a continuous super-linearfunction (i.e. h i ( x ) = ω ( x ) ) or a continuous sub-linear function (i.e. h i ( x ) = o ( x ) ). Then for any (cid:15) ∈ (0 , , there exists a distribution over agent values and an algorithm A such that any transfor-mation that performs a per-agent ironing of the allocation function of A achieving properties (P.1)and (P.2) above, must violate property (P.3) by a factor of Ω(1 /(cid:15) ) .Proof. We focus on a single agent i and drop the subscript i to improve readability. Our algorithmmakes non-zero allocations only to agent i so that the contribution of other agents to the objectiveis 0. Let h be the corresponding continuous increasing function and assume wlog that h (0) = 0and h (1) = 1. In the remainder of the proof we assume that h is super-linear. The proof for thesub-linear case is similar.With respect to this agent, our goal is to maximize the objective E [ h ( x ( v )) v ]. Fix any (cid:15) > − (cid:15) and1 with probability (cid:15) . At 0, the algorithm always allocates an amount 1 + (cid:15) (cid:48) . At 1, the allocationis 0 with probability 1 − /k and k with probability 1 /k . Here we pick (cid:15) (cid:48) > k such that h (1 + (cid:15) (cid:48) ) < / (1 − (cid:15) ) and h ( k ) > k/(cid:15) . The existence of (cid:15) (cid:48) and k follows from the fact that h iscontinuous and superlinear.Now, the expected allocation at 1 is 1, so in order to produce a BIC output, the mapping g must map each of the values to the other with some probability. Suppose that 0 gets mapped to 1with probability z/ (1 − (cid:15) ) for some z . Then, to preserve the distribution over values, we must have z ≤ (cid:15) and 1 must get mapped to 0 with probability z/(cid:15) .How large does z have to be to fix the non-monotonicity? The new expected allocation at 0 is z/ (1 − (cid:15) ) + (1 − z/ (1 − (cid:15) ))(1 + (cid:15) (cid:48) ), while the new expected allocation at 1 is (1 − z/(cid:15) ) + z/(cid:15) (1 + (cid:15) (cid:48) ).Setting the former to be no larger than the latter, and rearranging terms, we get z/ (1 − (cid:15) ) + z/(cid:15) ≥ z ≥ (cid:15) (1 − (cid:15) ).Let us now compute the objective function value. The original objective function value is (cid:15)h ( k ) /k . The new objective function value is given by (cid:15) ( z/(cid:15)h (1 + (cid:15) (cid:48) ) + (1 − z/(cid:15) ) h ( k ) /k ) < z/ (1 − (cid:15) ) + ( (cid:15) − z ) h ( k ) /k< h ( k ) /k ( (cid:15) − z + (cid:15)z/ (1 − (cid:15) ))= h ( k ) /k ( (cid:15) − (1 − (cid:15) ) z/ (1 − (cid:15) )) < (cid:15) h ( k ) /k Here the first inequality follows from h (1 + (cid:15) (cid:48) ) < / (1 − (cid:15) ), the second from 1 < (cid:15)h ( k ) /k and thefourth from z ≥ (cid:15) (1 − (cid:15) ). This implies that any mapping g that satisfies properties (P.1) and (P.2)must violate property (P.3) by a factor of at least 1 / (cid:15) .A similar example can be constructed for sub-linear continuous h , and we skip the details.Next we consider minimization problems of the form min E v [ (cid:80) i h i ( x i ( v )) h (cid:48) i ( v i )] where h i s arenon-decreasing functions and h (cid:48) i s are non-increasing functions. Once again, if there exists an i suchthat h i is non-linear, we get a gap. Theorem 5.2.
Consider the objective max E v [ (cid:80) i h i ( x i ( v )) h (cid:48) i ( v i )] where each h i is an arbitraryincreasing function and each h (cid:48) i is an arbitrary continuous decreasing function. Suppose that thereexists an agent i for which h i is a continuous super-linear function (i.e. h i ( x ) = ω ( x ) ) or acontinuous sub-linear function (i.e. h i ( x ) = o ( x ) ). Then for any (cid:15) > , there exists a distributionover agent values and an algorithm A such that any transformation that performs a per-agentironing of the allocation function of A achieving properties (P.1) and (P.2) above, must violateproperty (P.3) by a factor of Ω(1 /(cid:15) ) .Proof. Once again we focus on the agent i and present the proof for the case where the function h (the subscript i being implicit) is continuous and super-linear. The proof for the sub-linear caseis similar. Assume without loss of generality that h (0) = 0 and h (1) = 1. Let k ≥ h ( k ) > k/(cid:15) . Let v = h (cid:48)− (1) and v = h (cid:48)− ( (cid:15)/k ). Agent i ’s value distribution is uniform over { v , v } .The algorithm A is defined as follows. At v , the algorithm returns x with h ( x ) = 1 + (cid:15) . Notethat x >
1. At v , the algorithm returns k with probability 1 /k and 0 otherwise. The expectedallocation is 1, and therefore the allocation is non-monotone. Suppose that the tranformation maps v to v with probability z and vice versa. Then it is easy to see that z > / / (cid:15) ) + 1 / (cid:15)/k ( h ( k ) /k ) = 1 / (cid:15)/ h ( k ) /k ) < (cid:15)h ( k ) / k . We can boundfrom below the objective function value of the transformed mechanism by considering the termcorresponding to value v when the allocation is k (an event that happens with probability 1 / z times 1 /k ). The new objective function value is therefore at least zh ( k ) / k ≥ h ( k ) / k ≥ / (cid:15) + h ( k ) / k = 1 / (cid:15) (1 + (cid:15)h ( k ) / k )Here the second inequality follows by using h ( k ) > k/(cid:15) and k > /
2. This implies that anytransformation that satisfies properties (1) and (2) must violate property (3) by a factor of at least1 / (cid:15) . 20o conclude this section we note that for minimization problems with objectives of the form E v [ (cid:80) i x i ( v ) h (cid:48) i ( v i )] where h (cid:48) i s are decreasing functions, an approach similar to the ironing approachof [11] gives a BIC approximation-preserving transformation. In particular, if the type space foreach agent is finite then we can find a mapping by finding the min-cost perfect matching whereedge costs between v i and v i (cid:48) are h (cid:48) i ( v i ) E v − i [ x i ( v i (cid:48) , v − i )]. References [1] A. Archer and E. Tardos. Truthful mechanisms for one-parameter agents. In
Proc. 42nd IEEESymp. on Foundations of Computer Science , 2001.[2] Itai Ashlagi, Shahar Dobzinski, and Ron Lavi. An optimal lower bound for anonymous schedul-ing mechanisms. In
Proceedings of the 10th ACM conference on Electronic commerce , pages169–176, 2009.[3] M. Babaioff, R. Lavi, and E. Pavlov. Single-value combinatorial auctions and algorithmicimplementation in undominated strategies.
Journal of the ACM , 2009.[4] X. Bei and Z. Huang. Bayesian incentive compatibility via fractional assignments. In
Proc.22nd ACM Symp. on Discrete Algorithms , 2011.[5] D. Buchfuhrer, S. Dughmi, H. Fu, R. Kleinberg, E. Mossel, C. Papadimitriou, M. Schapira,Y. Singer, and C. Umans. Inapproximability for vcg-based combinatorial auctions. In
Proc.21st ACM Symp. on Discrete Algorithms , 2010.[6] George Christodoulou and Annam´aria Kov´acs. A deterministic truthful ptas for schedulingrelated machines. In
Proceedings of the Twenty-First Annual ACM-SIAM Symposium onDiscrete Algorithms , SODA ’10, pages 1005–1016, Philadelphia, PA, USA, 2010. Society forIndustrial and Applied Mathematics.[7] P. Dhangwatnotai, S. Dobzinski, S. Dughmi, and T. Roughgarden. Truthful approximationschemes for single-parameter agents. In
Proc. 49th IEEE Symp. on Foundations of ComputerScience , 2008.[8] S. Dobzinski. An impossibility result for truthful combinatorial auctions with submodularvaluations. In
Proc. 42nd ACM Symp. on Theory of Computing , 2011.[9] S. Dughmi and T. Roughgarden. Black-box randomized reductions in algorithmic mechanismdesign. In
Proc. 51st IEEE Symp. on Foundations of Computer Science , 2010.[10] S. Dughmi, T. Roughgarden, and Q. Yan. From convex optimization to randomized mech-anisms: Toward optimal combinatorial auctions. In
Proc. 42nd ACM Symp. on Theory ofComputing , 2011.[11] J. Hartline, R. Kleinberg, and A. Malekian. Bayesian incentive compatibility via matchings.In
Proc. 22nd ACM Symp. on Discrete Algorithms , 2011.[12] J. Hartline and B. Lucier. Bayesian algorithmic mechanism design. In
Proc. 41st ACM Symp.on Theory of Computing , 2010. 2113] Dorit s. Hochbaum and David B. Shmoys. A polynomial approximation scheme for schedulingon uniform processors: Using the dual approximation approach.
SIAM J. Comput. , 17:539–551, June 1988.[14] N. Immorlica and B. Lucier. On the impossibility of black-box truthfulness without priors. In
Workshop on Bayesian Mechanism Design , 2011.[15] R. Lavi and C. Swamy. Truthful and near-optimal mechanism design via linear programming.In
Proc. 46th IEEE Symp. on Foundations of Computer Science , 2005.[16] C. Papadimitriou, M. Schapira, and Y. Singer. On the hardness of being truthful. In