Bounding the Menu-Size of Approximately Optimal Auctions via Optimal-Transport Duality

08/29/2017
by   Yannai A. Gonczarowski, et al.
UV Web Design, Inc.
0

The question of the minimum menu-size for approximate (i.e., up-to-ε) Bayesian revenue maximization when selling two goods to an additive risk-neutral quasilinear buyer was introduced by Hart and Nisan (2013), who give an upper bound of O(1/ε^4) for this problem. Using the optimal-transport duality framework of Daskalakis et al. (2013, 2015), we derive the first lower bound for this problem - of Ω(1/√(ε)), even when the values for the two goods are drawn i.i.d. from "nice" distributions, establishing how to reason about approximately optimal mechanisms via this duality framework. This bound implies, for any fixed number of goods, a tight bound of Θ(1/ε) on the minimum deterministic communication complexity guaranteed to suffice for running some approximately revenue-maximizing mechanism, thereby completely resolving this problem. As a secondary result, we show that under standard economic assumptions on distributions, the above upper bound of Hart and Nisan (2013) can be strengthened to O(1/ε^2).

READ FULL TEXT VIEW PDF

page 1

page 2

page 3

page 4

03/03/2022

Numerical method for feasible and approximately optimal solutions of multi-marginal optimal transport beyond discrete measures

We propose a numerical algorithm for the computation of multi-marginal o...
09/18/2017

A Note on Tight Lower Bound for MNL-Bandit Assortment Selection Models

In this note we prove a tight lower bound for the MNL-bandit assortment ...
09/30/2019

On the Complexity of Approximating Multimarginal Optimal Transport

We study the complexity of approximating the multimarginal optimal trans...
02/10/2019

Set Cover in Sub-linear Time

We study the classic set cover problem from the perspective of sub-linea...
05/11/2022

Numerical method for approximately optimal solutions of two-stage distributionally robust optimization with marginal constraints

We consider a general class of two-stage distributionally robust optimiz...
02/18/2022

Is Selling Complete Information (Approximately) Optimal?

We study the problem of selling information to a data-buyer who faces a ...
11/06/2017

The menu complexity of "one-and-a-half-dimensional" mechanism design

We study the menu complexity of optimal and approximately-optimal auctio...

1 Introduction

One of the high-level goals of the field of Algorithmic Mechanism Design is to understand the tradeoff between the economic efficiency and the simplicity of mechanisms, with a central example being auction mechanisms. One of the most fundamental scenarios studied in this context is that of revenue-maximization by a single seller who is offering for sale two or more goods to a single buyer. Indeed, while classic economic analysis (Myerson, 1981) shows that for a single good, the revenue-maximizing mechanism is extremely simple to describe, it is known that the optimal auction for even two goods may be surprisingly complex and unintuitive (McAfee and McMillan, 1988; Thanassoulis, 2004; Manelli and Vincent, 2006; Hart and Reny, 2015; Daskalakis et al., 2013, 2015; Giannakopoulos and Koutsoupias, 2014, 2015), eluding a general description to date.

In this paper we study, for a fixed number of goods, the tradeoff between the complexity of an auction and the extent to which it can approximate the optimal revenue. While one may choose various measures of auction complexity (Hart and Nisan, 2013; Dughmi et al., 2014; Morgenstern and Roughgarden, 2015), we join several recent papers by focusing on the simplest measure, the menu-size suggested by Hart and Nisan (2013). While previous lower bounds on the menu-size as a function of the desired approximation to the revenue all assume a coupling between the number of goods and the desired approximation (so that the former tends to infinity simultaneously with the latter tending to optimal; e.g., setting ), in this paper we focus on the behavior of the menu-size as a function only of the desired approximation to the revenue, keeping the number of goods fixed and uncoupled from it. In particular, we obtain the first lower-bound on the menu-size that is not asymptotic in the number of goods, thereby quantifying the degree to which the menu-size of an auction really is a bottleneck to extracting high revenue even for a fixed number of goods.

Revenue Maximization

We consider the following classic setting. A risk-neutral seller has two goods for sale. A risk-neutral quasilinear buyer has a valuation (maximum willingness to pay) for each of these goods ,111Having rather than any other upper bound is without loss of generality, as the units are arbitrary. and has an additive valuation (i.e., values the bundle of both goods by the sum of the valuations for each good). The seller has no access to the valuations 

, but only to a joint distribution

from which they are drawn. The seller wishes to devise a (truthful) auction mechanism for selling these goods, which will maximize her revenue among all such mechanisms, in expectation over the distribution . (The seller has no use for any unsold good.) We denote the maximum obtainable expected revenue by .

Menu-Size

Hart and Nisan (2013) have introduced the menu-size

of a mechanism as a measure of its complexity: this measure counts the number of possible outcomes of the mechanism (where an outcome is a specification of an allocation probability for each good, coupled with a price).

222By the Taxation Principle, any mechanism is essentially described by the menu of its possible outcomes, as the mechanism amounts to the buyer choosing from this menu an outcome that maximizes her utility. Daskalakis et al. (2013) have shown that even in the case of independently distributed valuations for the two goods, precise revenue maximization may require an infinite menu-size; Daskalakis et al. (2015) have shown this even when the valuations for the two goods are drawn i.i.d. from “nice” distributions. In light of these results, relaxations of this problem, allowing for mechanisms that maximize revenue up-to-, were considered.

Approximate Revenue Maximization

Hart and Nisan (2013) have shown that a menu-size of  suffices for maximizing revenue up to an additive :

Theorem 1 (Hart and Nisan, 2013).

There exists such that for every and for every distribution , there exists a mechanism with menu-size at most such that . ( is the expected revenue of from .)

While the above-described results of Daskalakis et al. imply that the menu-size required for up-to- revenue maximization tends to infinity as tends to , no lower bound whatsoever was known on the speed at which it tends to infinity. (I.e., all that was known was that the menu-size is  as a function of .) Our main result, which we prove in Section 2, is the first lower bound on the required menu-size for this problem,333In fact, we do not know of any previous menu-size bound, for any problem, that lower-bounds the menu-size as a function of without having the number of goods also tend to infinity (e.g., by setting ). showing that a polynomial dependence on  is not only sufficient, but also required, hence establishing that the menu-size really is a nontrivial bottleneck to extracting high revenue, even for two goods and even when the valuations for the goods are drawn i.i.d. from “nice” distributions:

Theorem 2 (Menu-Size: Lower Bound).

There exist and a distribution , such that for every it is the case that for every mechanism  with menu-size at most .

Our proof of Theorem 2 uses the optimal-transport duality framework of Daskalakis et al. (2013, 2015). To the best of our knowledge, this is the first use of this framework to reason about approximately optimal mechanisms, thereby establishing how to leverage this framework to do so.

Communication Complexity

As Babaioff et al. (2017) show, the logarithm (base , rounded up) of the menu-size of a mechanism is precisely the deterministic communication complexity (between the seller and the buyer) of running this mechanism, when the description of the mechanism itself is common knowledge. Therefore, by Theorems 2 and 1, we obtain a tight bound on the minimum deterministic communication complexity guaranteed to suffice for running some up-to- revenue-maximizing mechanism, thereby completely resolving this problem:

Corollary 3 (Communuication Complexity: Tight Bound).

There exists such that for every it is the case that ) is the minimum communication complexity that satisfies the following: for every distribution there exists a mechanism  such that the deterministic communication complexity of running  is ) and such that . This continues to hold even if is guaranteed to be a product of two independent identical distributions.

In Section 4, we extend this tight communication-complexity bound to any fixed number of goods, as well as derive analogues of our results for multiplicative up-to- approximation.

While our lower bound completely resolves the open question of whether a polynomial menu-size is necessary (and not merely sufficient), and while it tightly characterizes the related communication complexity, it does not yet fully characterize the precise polynomial dependence of the menu-size on . While the proof of our lower bound (Theorem 2) makes delicate use of a considerable amount of information regarding the optimal mechanism via the optimal-transport duality framework of Daskalakis et al. (2013, 2015), the proof of the upper bound of Hart and Nisan (2013) (Theorem 1) makes use of very little information regarding the optimal mechanism. (As noted above, indeed very little is known regarding the structure of general optimal mechanisms.) Our secondary result, which we prove in Section 3, shows that under standard economic assumptions on valuation distributions, the upper bound of Hart and Nisan (2013) can be tightened by two orders of magnitude. This suggests that it may well be possible to use more information regarding the structure of optimal mechanisms, as such will be discovered, to unconditionally improve the upper bound of Hart and Nisan (2013).

Theorem 4 (Menu-Size: Conditional Upper Bound).

There exists such that for every and for every distribution satisfying the McAfee-McMillan hazard condition (see Definition 1 in Section 3), there exists a mechanism with menu-size at most such that .

The question of a precise tight bound on the menu-size required for up-to- revenue maximization remains open, for correlated as well as product (even i.i.d.) distributions. We further discuss our results and their connection to other results and open problems in the literature in Section 5.

2 Lower Bound

We prove our lower bound (i.e., Theorem 2) by considering the setting in which Daskalakis et al. (2015)

show that precise revenue maximization requires infinite menu-size, that is, the case of two items with i.i.d. valuations each drawn from the Beta distribution

, i.e., the distribution over with density . We first present, in Section 2.1, a very high-level overview of the main proof idea in a way that does not go into any technical details regarding the duality framework of Daskalakis et al. (2013, 2015). We then present, in Section 2.2, only the minimal amount of detail from the extensive analysis of Daskalakis et al. (2015) that is required to follow our proof. Finally, in Section 2.3 we prove Theorem 2.

2.1 Proof Idea Overview

We start by presenting a very high-level overview of the main idea of the proof of Theorem 2 in a way that does not go into any technical details regarding the duality framework of Daskalakis et al. (2013, 2015).

Fix a concrete distribution from which the values of the goods are drawn. Let us denote the set of all truthful mechanisms by and for each , let us denote its expected revenue by .444As will become clear momentarily, and stand for “solution,” stands for “primal,” and stands for “objective.” The revenue maximization problem is to find a solution for which the value of the objective function is maximal. Daskalakis et al. (2013, 2015) identify a dual problem to the revenue maximization problem: this is a minimization problem, i.e., a problem where the goal is to find a solution from a specific set of feasible solutions that they identify, that minimizes the value of a specific objective function that they identify. This problem is an instance of a class of problems called optimal-transport problems, and it is a dual problem to the revenue maximization problem in the sense that for every pair of solutions for the primal (revenue maximization) problem and dual (optimal-transport) problem respectively, it holds that the value of the primal objective function for the primal solution is upper-bounded by the value of the dual objective function for the dual solution, that is,

(1)

(See Equation 5 below for the full details.) This property is called weak duality. Daskalakis et al. (2015) also show that this duality is strong in the sense that there always exists a pair of solutions for the primal and dual problems respectively such that

(2)

A standard observation in duality frameworks is that such certifies that is an optimal solution for the primal problem, since by Equation 1 the value of any primal solution is bounded by . Indeed, Daskalakis et al. (2015) use their framework to identify and certify optimal primal solutions (revenue-maximizing mechanisms) by identifying such pairs . To facilitate finding such pairs of solutions, they identify complementary slackness conditions, that is, conditions on and that are necessary and sufficient for the inequality in Equation 1 to in fact be an equality as in Equation 2. In particular, for the revenue maximization problem where the two items are sampled i.i.d. from the Beta distribution , they identify such a pair of solutions . They in fact show the that complementary slackness conditions (for this distribution) uniquely define the optimal primal solution and that this solution has an infinite menu-size.

The main idea of our proof is, with the optimal dual that Daskalakis et al. (2015) identify in hand, to carefully show that for every primal solution that is a mechanism with small menu-size, these complementary slackness conditions not only fail to hold for (as follows from the result of Daskalakis et al., 2015), but in fact are sufficiently violated to yield the required separation, that is,

(3)

where is the optimal primal solution and the equality is by Equation 2. (See Equation 7 below for the full details.)

In slightly more detail, Daskalakis et al. (2015) show that for two items sampled i.i.d. from , the complementary slackness conditions dictate that in a certain part of the value space, the set of values of buyers to which Good (say) is allocated with probability  by the optimal mechanism and the set of values of buyers to which it is allocated with probability in that mechanism are separated by a strictly concave curve (the curve in Figure 1 below) and that this implies that the optimal mechanism has an infinite menu-size. Another way to state this conclusion is to observe that it follows from known properties that in a mechanism with a finite menu-size such a curve must be piecewise-linear rather than strictly concave, and so to conclude that the optimal mechanism cannot have a finite menu-size. Roughly speaking, we relate the loss in revenue (compared to the optimal mechanism) of a given mechanism to a certain metric (see below) of the region between the separating curve of the optimal mechanism and an analogue of this curve (see below for the precise definition) in the given mechanism. We then observe that this analogue of is a piecewise-linear curve with number of pieces at most the menu-size of the given mechanism, and use this to appropriately lower-bound this metric (see Proposition 5 below) for mechanisms with small menu-size. A lower bound on the loss in revenue for mechanisms with such a menu-size follows.

To present our analysis in further detail, we must now first dive into some of the details of the optimal-transport duality framework of Daskalakis et al. (2015).

2.2 Minimal Needed Essentials of the Optimal-Transport Duality Framework of Daskalakis et al. (2015), and Commentary

We now present only the minimal amount of detail from the extensive analysis of Daskalakis et al. (2015) that is required to follow our proof; the interested reader is referred to their paper or to the excellent survey of Daskalakis (2015), whose notation we follow, for the full details that lie beyond the scope of this paper. (See also Giannakopoulos and Koutsoupias (2014) for a slightly different duality approach, and Kash and Frongillo (2016) for an extension.)

In their analysis, Daskalakis et al. (2015) identify a signed Radon measure555To understand our proof there is no need to be familiar with the general definition of a signed Radon measure. It suffices to know that signed Radon measures generalize distributions that are defined by a combination of atoms and a density function, and allow in particular for   a) densities (and atoms) that can also be negative (hence the term signed), and  b) the overall measure not necessarily summing up to . on with ,666That is, the overall measure sums up to . such that for a mechanism with utility function777The utility function of a mechanism maps each buyer type (i.e., pair of valuations) to the utility that a buyer of this type obtains from participating in the mechanism. , the expected revenue of this mechanism from is equal to888This Lebesgue integral is the measure-theoretic analogue of the expectation (or average) of with respect to a given distribution (but as is not a distribution, when “averaging,” some values are taken with negative weights, and weights do not sum up to ).

(4)

They show that (the utility function of) the revenue-maximizing mechanism is obtained by maximizing Equation 4 subject to , to being convex, and to , where . (Rochet (1987) has shown that the utility function of any truthful mechanism satisfies the latter two properties as well as . An equality as in the first property may be assumed without losing revenue or changing the menu-size.)

We comment that while one could have directly attempted prove Theorem 2 by analyzing how much revenue is lost in Equation 4 due to restricting attention only to that corresponds to a mechanism with a certain (small) menu-size (in particular, the graph of such is a maximum of planes, where is the menu-size), such an analysis, even if successful, would have been hard and involved, and immensely tailored to the specifics of the distribution , due to the complex definition of . For this reason we base our analysis on the duality-based framework of Daskalakis et al. (2013, 2015), which they have developed to help find and certify the optimal , and we show, for the first time to the best of our knowledge, how to use this framework to quantitatively reason about the revenue loss from suboptimal mechanisms. The resulting approach is principled, general, and robust.999For example, readers familiar with the definition of exclusion set mechanisms (Daskalakis et al., 2015) may notice that our analysis of below can be readily applied with virtually no changes also to other distributions for which the optimal mechanism is derived from an exclusion set that is nonpolygonal (as is in the analysis below).

Daskalakis et al. (2015) show that for every utility function  of a (truthful) mechanism for valuations in and for every coupling101010Informally (and sufficient to understand our proof), a coupling of two unsigned Radon measures and both having the same overall measure is a recipe for rearranging the mass of into the mass of by specifying where each piece of (positive) mass is transported. Formally, a coupling of two unsigned Radon measures and on with is an unsigned Radon measure on whose marginals are and , i.e., for every measurable set , it holds that and . of and , where is any measure that convex-dominates111111A distribution convex-dominates a distribution if is obtained from by shifting mass to coordinate-wise larger points and by performing mean-preserving spreads of positive mass. To follow our paper only a single property of convex dominance is needed — see below. As Daskalakis et al. (2015) show (but not required for our proof), of interest in this context are in fact only cases where is obtained from  by mean-preserving spreads of positive mass. , it is the case that121212Once again, the Lebesgue integral on the right-hand side is the measure-theoretic analogue of the average of the values , where, informally, each pair is taken with weight equal to the amount of (positive) mass transported by from to .

(5)

(This is precisely Equation 1 in full detail.) They identify the optimal mechanism for the distribution by finding a measure131313For this specific distribution , the measure that they identify in fact equals .  and a coupling of and , such that Equation 5 holds with an equality for (the utility function of ) and (this is precisely Equation 2 in full detail).141414In fact, Daskalakis et al. (2015) prove a beautiful theorem that states that this (i.e., finding suitable and such that Equation 5 holds with an equality for the optimal ) can be done for any underlying distribution, i.e., that this duality is strong. To find and , they make use of complementary slackness conditions that they identify, and make sure they are all completely satisfied. In our proof below, we will claim that for any utility function that corresponds to a mechanism with small menu-size, the complementary slackness conditions with respect to and will be sufficiently violated so as to give sufficient separation between the left-hand side of Equation 5 for (that is, the revenue of the mechanism with small menu-size) and the right-hand side of Equation 5 for  (that is, the optimal revenue).

To better understand the complementary slackness conditions identified by Daskalakis et al. (2015), let us review their proof for Equation 5:

(6)

where the first inequality is since is convex (this inequality is the only property of convex dominance that is needed to follow our paper), the second equality is by the definition of a coupling, and the second inequality is due to the third property of as defined above following Equation 4. Daskalakis et al. (2015) note that if it would not be the case that -almost everywhere151515Informally (and sufficient to understand our proof), for a condition to hold -almost everywhere means for that condition to hold for every and such that the coupling  transports (positive) mass from to . we would have , then the second inequality in Equation 6 would be strict, and so the same proof would give a strict inequality in Equation 5; they use this insight to guide their search for the optimal and its tight dual . (They also perform a similar analysis with respect to the first inequality in Equation 6, which we skip as we do not require it.) In our proof below we will show that for the coupling that they identify, and for any that corresponds to a mechanism with small menu-size, this constraint (i.e., that -almost everywhere ) will be significantly violated, in a precise sense. To do so, we first describe the measure and the coupling that they identify.

Examine Figure 1.

0

Figure 1: The optimal coupling identified by Daskalakis et al. (2015) for , illustrated in the region . Figure adapted from Daskalakis et al. (2015) with permission.

For our proof below it suffices to describe the measure and the coupling , both restricted to a region for an appropriate .161616We choose as the horizontal-axis coordinate of the right boundary of the region denoted by in Daskalakis et al. (2015). The measure has a point mass of measure  in , and otherwise in every has density

In the region , the coupling sends positive mass downward, from positive-density points to negative-density points. In the region , the coupling sends positive mass from upward and rightward to all points (the density is indeed negative throughout ). No other positive mass is transported inside or into . (Some additional positive mass from is transported out of .)

The optimal mechanism that Daskalakis et al. (2015) identify does not award any good (nor does it charge any price) in the region , while awarding Good  with probability  and Good  with varying probabilities (and charging varying prices) in the region .

As Daskalakis et al. (2015) note, indeed -almost everywhere the complementary slackness condition holds for this mechanism: in the region , coupled points  have and , and in this region, ; in the region , coupled points have and . In fact, this reasoning shows that given the optimal coupling , the utility function  is uniquely defined by the complementary slackness conditions, and so is the unique revenue-maximizing utility function. Since it is well known that wherever a utility function of a truthful mechanism is differentiable, its derivative in the direction of is the allocation probability of Good (indeed, by examining Good  one can verify using this property that the mechanism corresponding to indeed awards Good with probability  in the region  and with probability in the region ), then by examining Good , since the curve (see Figure 1) that separates the regions and  is strictly concave, Daskalakis et al. (2015) conclude that there is a continuum of allocation probabilities of Good in the mechanism corresponding to (which is the unique revenue-maximizing mechanism), thus concluding that the unique revenue-maximizing mechanism for the distribution has an infinite menu-size.

2.3 Proof of Theorem 2

An alternative way to state the conclusion of the argument of Daskalakis et al. (2015) that any revenue-maximizing mechanism for has an infinite menus-size is as follows: let be the utility function of a mechanism with a finite menu-size. It is well known that the graph of is the maximum of a finite number of planes (each corresponding to one entry in the menu of ). Therefore, since is strictly concave, it is impossible for to have derivative  beneath the curve and derivative above the curve , and so the complementary slackness conditions must be violated, and hence is not optimal. In our proof we will quantify the degree of violation of the complementary slackness conditions as a function of the finite menu-size of such . We would like to reason as follows: for such with a finite menu-size, define the corresponding curve  that is the analogue for of the curve , and then show that since must be piecewise-linear, quantifiable revenue is lost due to the complementary slackness conditions not holding in the region between and . It is not immediately clear how to define , though.

Intuitively we would have liked to define to be the curve on above which awards Good with probability and below which awards Good with probability , but what if also awards Good with fractional probability? How should we define in such cases? (Remember that all that we know about is that it has small menu-size.) As we will see below, to show that we indeed have quantifiably sufficient revenue loss from any deviation of from , we will define as the curve above which awards Good with probability more than one half, and below which awards Good  with probability less than or equal to one half. As will become clear from our calculations, the constant one half could have been replaced here with any fixed fraction,171717Similarly, the direction of tie breaking with respect to the region where Good is awarded with probability precisely one half could have been flipped. but crucially it could not have been replaced with (i.e., defining as the curve above which awards Good with positive probability and below which does not award Good ) or with (i.e., defining as the curve above which awards Good with probability and below which awards Good with probability strictly less than ).

We are now finally ready to prove Theorem 2, with the help of the following Proposition, which may be of separate interest. (We prove Proposition 5 following the proof of Theorem 2.)

Proposition 5.

Let be a concave function with radius of curvature at most  everywhere, for some . For small enough , the following holds. For any piecewise-linear function  composed of at most linear segments, the Lebesgue measure of the set of coordinates with is at least .

Proof of Theorem 2.

The curve (see Figure 1) that separates the regions and is given by (where ) (Daskalakis et al., 2015). Therefore, it is strictly concave, having radius of curvature at most everywhere, for some fixed . We note also that there exists a constant and a neighborhood  of the curve  in  in which the density of is (negative and) smaller than .

Let and set . Assume without loss of generality that is sufficiently small so that both  i) the -neighborhood of in is contained in the neighborhood  of , and  ii) Proposition 5 holds with respect to . Let .

Let be the utility function of a mechanism with menu-size at most , and let be defined as follows:

It is well known that the graph of is the maximum of planes. Therefore, is a piecewise-linear function composed of at most linear segments.

Let with . Let and let be such that transfers positive mass from to . (All mass transferred to by  is from points of this form.) We note that by definition of ,

Similarly, let with . Let . Note that  transfers positive mass from to . (All mass transferred to by  is from the point .) We note that by definition of ,

By Proposition 5, the Lebesgue measure of the set of coordinates with is at least . Similarly to Equation 6, we therefore obtain

(7)

as required (this is precisely Equation 3 in full detail). ∎

Proof of Proposition 5.

We will show that from each linear segment of , at most a Lebesgue measure of coordinates satisfy

(8)

This implies the Proposition since this means that from all linear segments of together, at most a Lebesgue measure of coordinates satisfy Equation 8, and hence at least a Lebesgue measure of coordinates satisfy , as required.

For a Lebesgue measure of coordinates from a single linear segment of to satisfy Equation 8, we note that a necessary condition is that be at most the length of a chord of sagitta at most  in a circle of radius at most . (See Figure 2.)

Figure 2: In the extreme case where is an arc of radius , a line segment with the maximum Lebesgue measure of coordinates satisfying Equation 8 is a horizontal chord of sagitta  in a circle (the circle of the top dotted arc) of radius .

We claim that this implies that , as required. Indeed, in the extreme case where is the length of a chord of sagitta precisely  in a circle of radius precisely , we have by a standard use of the Intersecting Chords Theorem181818The Intersecting Chords Theorem states that when two chords of the same circle intersect, the product of the lengths of the two segments (that are delineated by the intersection point) of one chord equals the product of the lengths of the two segments of the other. that . Solving for , we have that (in the extreme case) , as claimed. ∎

3 Upper Bound

Recall that Theorem 1, due to Hart and Nisan (2013), provides an upper bound of on the menu-size of some mechanism that maximizes revenue up to an additive . Their proof uses virtually no information regarding the structure of the optimal mechanism: it starts with a revenue-maximizing mechanism, and cleverly rounds two of the three coordinates (allocation probability of Good , allocation probability of Good , price) of every outcome, to obtain a mechanism with small menu-size without significant revenue loss. We will follow a similar strategy, but will only round one of these three coordinates (namely, the price), using a result by Pavlov (2011) that shows that under an assumption on distributions that is standard in the economics literature on multidimensional mechanism design (see, e.g., McAfee and McMillan, 1988; Manelli and Vincent, 2006; Pavlov, 2011), for an appropriate choice of revenue-maximizing mechanism, one of the other (allocation) coordinates is in fact already rounded (specifically, it is either zero or one).

Definition 1 (McAfee and McMillan, 1988).

A distribution is said to satisfy the McAfee-McMillan hazard condition if it has a differentiable density function satisfying

for every .

Theorem 6 (Pavlov, 2011).

For every distribution satisfying the McAfee-McMillan hazard condition (for ), there exists a mechanism that maximizes the revenue from and has no outcome for which both allocation probabilities are in the open interval .

Our upper bound follows by combining the argument of Hart and Nisan (2013) with Theorem 6.

Proof of Theorem 4.

Let be a revenue-maximizing mechanism for as in Theorem 6. Let , and for every real number , denote by the rounding-down of to the nearest integer multiple of . Let be the mechanism whose menu is comprised of all outcomes of the form for every outcome of (where is the allocation probability of Good , and is the price charged in this outcome). We claim that .

The main idea of the “nudge and round” argument of Hart and Nisan (2013) is that while the rounding (which is performed to reduce the menu-size — see below), by itself (without the discounting, which is the “nudge” part), could have hypothetically constituted the “last straw” that causes some buyer type to switch from preferring a very expensive outcome to preferring a very cheap one (thus significantly hurting the revenue), since more expensive outcomes are more heavily discounted, then this compensates for any such “last straw” effects. More concretely, while the rounding, before the -discounting, can cause a buyer’s utility from any outcome to shift by at most (which could be the “last straw”), and since for any outcome whose price is cheaper by more than an compared to the buyer’s original outcome of choice the given discount is smaller by at least , this smaller discount more than eliminates any potential utility gain due to rounding, so such an outcome would not become the most-preferred one. We will now formally show this.

Fix a type for the buyer. Let be the outcome according to  when the buyer has type . It is enough to show that the buyer pays at least according to when she has type . (We denote the price of, e.g., by .) Let be a possible outcome of , and let  be the outcome of that corresponds to it. We will show that if , then a buyer of type strictly prefers the outcome of that corresponds to over  (and so does not choose in ). Indeed, since in this case , we have that

so in , a buyer of type pays at least

How many menu entries does really have (i.e., how many menu entries ever get chosen by the buyer)? The number of menu entries with is at most ), since for every price (there are such options) we can assume without loss of generality that only the menu entry with highest will ever be chosen.191919If a maximum such is not attained, then we can add a suitable menu entry with the supremum of such ; see Babaioff et al. (2017) for a full argument. A similar argument for the cases , , and (by Theorem 6, no more cases exist beyond these four) shows that in total there really are at most menu entries in , as required. ∎

We note that both Pavlov (2011) and Kash and Frongillo (2016) conjecture that the conclusion of Theorem 6 holds under more general conditions than the McAfee-McMillan hazard condition. An affirmation of (either of) these conjectures would, by the above proof, immediately imply that the conclusion of Theorem 4 holds under the same generalized assumptions.

Hart and Nisan (2013) also analyze the scenario of a two-good distribution supported on for any given , and give an upper bound of on the menu-size that suffices for revenue maximization up to a multiplicative . Using the above techniques, their argument could be similarly modified to give an improved upper bound of in that setting for distributions satisfying the McAfee-McMillan hazard condition (or any generalized condition under which the conclusion of Theorem 6 holds).

4 Extensions

4.1 More than Two Goods

Recall that Corollary 3 concludes, from our menu-size lower bound (Theorem 2) and the menu-size upper bound of Hart and Nisan (2013) (Theorem 1), a tight bound on the minimum deterministic communication complexity guaranteed to suffice for running some up-to- revenue-maximizing mechanism for selling two goods, thereby completely resolving this problem. In fact, since Hart and Nisan (2013) prove an upper bound of (later strengthened to by Dughmi et al., 2014) on the menu-size required for revenue maximization up to an additive when selling any number of goods , we obtain our tight communication-complexity bound not only for two goods, but for any fixed number of goods :

Corollary 7 (Communication Complexity: Tight Bound for Any Number of Goods).

Fix . There exists such that for every it is the case that ) is the minimum communication complexity that satisfies the following: for every distribution there exists a mechanism  for selling goods such that the deterministic communication complexity of running  is ) and such that . This continues to hold even if is guaranteed to be a product distribution.

For the case of one good, the seminal result of Myerson (1981) shows that there exists a (precisely) revenue-maximizing mechanism with only two possible outcomes (and hence deterministic communication complexity of ), which simply offers the good for a suitably chosen take-it-or-leave-it price. Corollary 7 therefore shows a precise dichotomy in the asymptotic communication complexity of up-to- revenue maximization, between the case of one good (Myerson’s result; bit of communication) on the one hand, and the case of any other fixed number of goods ( bits of communication) on the other hand.

4.2 Multiplicative Approximation

In a scenario where the valuations may be unbounded, i.e., for all , Hart and Nisan (2013) have shown that no finite menu-size suffices for maximizing revenue up to a multiplicative202020For unbounded valuations, it makes no sense to consider additive guarantees, as the problem is invariant under scaling of the currency. , and consequently Hart and Nisan (2014) asked212121Hart and Nisan (2014) is a manuscript combining Hart and Nisan (2013) with an earlier paper. whether this impossibility may be overcome for the case of independently distributed valuations for the goods. Babaioff, Gonczarowski, and Nisan (2017) gave a positive answer, showing that for every and , a finite menu-size suffices, and moreover gave an upper bound of on the sufficient menu-size. Since for valuations in , revenue maximization up to a multiplicative  is a stricter requirement than revenue maximization up to an additive , our lower bound from Theorem 2 immediately provides a lower bound for this scenario as well.

Corollary 8 (Menu-Size for Multiplicative Approximation: Lower Bound).

There exist and a distribution , such that for every it is the case that for every mechanism with menu-size at most .

By an argument similar to that yielding Corollary 7, using the above upper bound of Babaioff et al. (2017) in lieu of that of Hart and Nisan (2013) / Dughmi et al. (2014), we therefore obtain an analogue of Corollary 7 for this setting as well.

Corollary 9 (Communication Complexity of Multiplicative Approximation: Tight Bound).

Fix . There exists such that for every it is the case that is the minimum communication complexity that satisfies the following: for every product distribution there exists a mechanism  for selling goods such that the deterministic communication complexity of running  is ) and such that . This continues to hold even if each of the distributions whose product is is guaranteed to be supported in .

Corollary 9 shows that the dichotomy between one good and any other fixed number of goods that is shown by Corollary 7 also holds for multiplicative up-to- revenue maximization.

5 Discussion

In a very recent paper, Saxena et al. (2018) analyze the menu-size required for approximate revenue maximization in what is known as the FedEx problem (Fiat et al., 2016). Interestingly, they also use piecewise-linear approximation of concave functions to derive their bounds. Nonetheless, there are considerably many differences between their analysis and ours, e.g.: the effects of bad piecewise-linear approximation on the revenue,222222There: even a single point having large distance can cause quantifiable revenue loss; Here: at least a certain measure of points having large distance causes quantifiable revenue loss. the approximated/approximating object,232323There: revenue curves, with vertices corresponding to menu entries; Here: contour lines of the allocation function, with edges corresponding to menu entries. the mathematical features of the approximated object,242424There: piecewise-linear; Here: strictly concave. the geometric/analytic proof of the bound on the number of linear segments, and finally, the argument that uses the piecewise-linear approximation and whether or not it couples the desired approximation with another parameter of the problem.252525There: lower bound achieved by coupling the desired approximation with the number of possible deadlines  (setting ); Here: desired approximation uncoupled from the number of goods . It therefore seems to be unlikely that both analyses are special cases of some general analysis, and so it would be interesting to see whether piecewise-linear approximations of concave (or other) functions “pop up” in the future in any additional contexts in connection with bounds on the menu-size of mechanisms.262626Incidentally, other known derivations of menu-size bounds, such as those in Hart and Nisan (2013), Dughmi et al. (2014), and Babaioff et al. (2017), do not use (even implicitly, to the best of our understanding) piecewise-linear approximations.

As mentioned in Section 4, Babaioff et al. (2017) prove an upper bound on the menu-size required for multiplicative up-to- approximate revenue maximization when selling goods to an additive buyer with independently distributed valuations. For any fixed , this bound is polynomial in , and the lower bound that we establish in Corollary 8 shows in particular that a polynomial dependence cannot be avoided here (e.g., it cannot be reduced to a logarithmic or lower dependence) even for bounded distributions, yielding a tight communication-complexity bound (see Corollary 9 in Section 4). Alternatively to fixing and analyzing the menu-size and communication complexity as functions of as we do, one may fix and analyze these quantities as functions of . For any fixed , the upper bound of Babaioff et al. (2017) is exponential in ; therefore, another question left open by that paper is whether this exponential dependence may be avoided. (In terms of communication complexity, this question asks whether for every fixed , the communication complexity can be logarithmic or even polylogarithmic in .) Some progress on this question has been made already by Babaioff et al. (2017), who show that on the one hand, a polynomial dependence on  suffices for some values of  (namely, ), and that on the other hand, an exponential dependence on is required when coupling with the number of goods by setting ;272727This result also shows that the exponential dependence on in the upper bounds of Hart and Nisan (2013) and Dughmi et al. (2014) for additive up-to- approximation (see Section 4) is required even for bounded product distributions. However, one may claim that the “more interesting” question when keeping  fixed and letting vary is that of a multiplicative approximation, as asking for at most an additive loss while increasing the total value in the market (by adding more and more items) is quite a harsh requirement. however, as noted above, they leave the general case of arbitrary fixed  (uncoupled from , e.g., ) as their main remaining open question. While the current state-of-the-art literature seems to be a long way from identifying very-high-dimensional optimal mechanisms, and especially from identifying their duals (indeed, it took quite some impressive effort for Giannakopoulos and Koutsoupias (2014) to identify the optimal mechanism for goods whose valuations are i.i.d. uniform in ), one may hope that with time, it may be possible to do so. It seems plausible that if one could generate high-dimensional optimal mechanisms (and corresponding duals) for which the high-dimensional analogue of the curve that we denote by in Section 2 has large-enough measure (while maintaining a small-enough radius of curvature, etc.), then an argument similar to the one that we give in the proof of Theorem 2 may be used to show that an exponential dependence on  in the above bound is indeed required for sufficiently small, yet fixed, , and thereby resolve the above open question. Whether one can generate such mechanisms with large-enough high-dimensional analogues of , however, remains to be seen.

Acknowledgments

Yannai Gonczarowski is supported by the Adams Fellowship Program of the Israel Academy of Sciences and Humanities. His work is supported by ISF grant 1435/14 administered by the Israeli Academy of Sciences and by Israel-USA Bi-national Science Foundation (BSF) grant number 2014389. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 740282). I thank Costis Daskalakis, Kira Goldner, Sergiu Hart, Ian Kash, Noam Nisan, Christos Tzamos, and an anonymous referee for helpful feedback. I thank Costis Daskalakis, Alan Deckelbaum, and Christos Tzamos for permission to base Figure 1 on a Figure from their paper.

References

  • Babaioff et al. (2017) M. Babaioff, Y. A. Gonczarowski, and N. Nisan. The menu-size complexity of revenue approximation. In

    Proceedings of the 49th Annual ACM Symposium on Theory of Computing (STOC)

    , pages 869–877, 2017.
  • Daskalakis (2015) C. Daskalakis. Multi-item auctions defying intuition? ACM SIGecom Exchanges, 14(1):41–75, 2015.
  • Daskalakis et al. (2013) C. Daskalakis, A. Deckelbaum, and C. Tzamos. Mechanism design via optimal transport. In Proceedings of the 14th ACM Conference on Electronic Commerce (EC), pages 269–286, 2013.
  • Daskalakis et al. (2015) C. Daskalakis, A. Deckelbaum, and C. Tzamos. Strong duality for a multiple-good monopolist. In Proceedings of the 16th ACM Conference on Economics and Computation (EC), pages 449–450, 2015.
  • Dughmi et al. (2014) S. Dughmi, L. Han, and N. Nisan. Sampling and representation complexity of revenue maximization. In Proceedings of the 10th Conference on Web and Internet Economics (WINE), pages 277–291, 2014.
  • Fiat et al. (2016) A. Fiat, K. Goldner, A. R. Karlin, and E. Koutsoupias. The FedEx problem. In Proceedings of the 17th ACM Conference on Economics and Computation (EC), pages 21–22, 2016.
  • Giannakopoulos and Koutsoupias (2014) Y. Giannakopoulos and E. Koutsoupias.

    Duality and optimality of auctions for uniform distributions.

    In Proceedings of the 15th ACM Conference on Economics and Computation (EC), pages 259–276, 2014.
  • Giannakopoulos and Koutsoupias (2015) Y. Giannakopoulos and E. Koutsoupias. Selling two goods optimally. In Proceedings of the 42nd International Colloquium on Automata, Languages, and Programming (ICALP), pages 650–662, 2015.
  • Hart and Nisan (2013) S. Hart and N. Nisan. The menu-size complexity of auctions. In Proceedings of the 14th ACM Conference on Electronic Commerce (EC), page 565, 2013.
  • Hart and Nisan (2014) S. Hart and N. Nisan. How good are simple mechanisms for selling multiple goods? Discussion Paper 666, Center for the Study of Rationality, The Hebrew University of Jerusalem, 2014.
  • Hart and Reny (2015) S. Hart and P. J. Reny. Maximal revenue with multiple goods: Nonmonotonicity and other observations. Theoretical Economics, 10(3):893–922, 2015.
  • Kash and Frongillo (2016) I. A. Kash and R. M. Frongillo. Optimal auctions with restricted allocations. In Proceedings of the 17th ACM Conference on Economics and Computation (EC), pages 215–232, 2016.
  • Manelli and Vincent (2006) A. M. Manelli and D. R. Vincent. Bundling as an optimal selling mechanism for a multiple-good monopolist. Journal of Economic Theory, 127(1):1–35, 2006.
  • McAfee and McMillan (1988) R. P. McAfee and J. McMillan. Multidimensional incentive compatibility and mechanism design. Journal of Economic Theory, 46(2):335–354, 1988.
  • Morgenstern and Roughgarden (2015) J. Morgenstern and T. Roughgarden. On the pseudo-dimension of nearly optimal auctions. In Proceedings of Advances in Neural Information Processing Systems 28 (NIPS), pages 136–144, 2015.
  • Myerson (1981) R. Myerson. Optimal auction design. Mathematics of Operations Research, 6(1):58–73, 1981.
  • Pavlov (2011) G. Pavlov. A property of solutions to linear monopoly problems. The BE Journal of Theoretical Economics, 11(1):#4, 2011.
  • Rochet (1987) J.-C. Rochet. A necessary and sufficient condition for rationalizability in a quasi-linear context. Journal of Mathematical Economics, 16(2):191–200, 1987.
  • Saxena et al. (2018) R. R. Saxena, A. Schvartzman, and S. M. Weinberg. The menu complexity of “one-and-a-half-dimensional” mechanism design. In Proceedings of the 29th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 2026–2035, 2018.
  • Thanassoulis (2004) J. Thanassoulis. Haggling over substitutes. Journal of Economic Theory, 117(2):217–245, 2004.