1 Introduction
One of the highlevel 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 revenuemaximization 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 revenuemaximizing 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 menusize suggested by Hart and Nisan (2013). While previous lower bounds on the menusize 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 menusize 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 lowerbound on the menusize that is not asymptotic in the number of goods, thereby quantifying the degree to which the menusize 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 riskneutral seller has two goods for sale. A riskneutral quasilinear buyer has a valuation (maximum willingness to pay) for each of these goods ,^{1}^{1}1Having 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 .MenuSize
Hart and Nisan (2013) have introduced the menusize
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).
^{2}^{2}2By 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 menusize; 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 upto, were considered.Approximate Revenue Maximization
Hart and Nisan (2013) have shown that a menusize 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 menusize at most such that . ( is the expected revenue of from .)
While the abovedescribed results of Daskalakis et al. imply that the menusize required for upto 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 menusize is as a function of .) Our main result, which we prove in Section 2, is the first lower bound on the required menusize for this problem,^{3}^{3}3In fact, we do not know of any previous menusize bound, for any problem, that lowerbounds the menusize 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 menusize 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 (MenuSize: Lower Bound).
There exist and a distribution , such that for every it is the case that for every mechanism with menusize at most .
Communication Complexity
As Babaioff et al. (2017) show, the logarithm (base , rounded up) of the menusize 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 upto revenuemaximizing 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 communicationcomplexity bound to any fixed number of goods, as well as derive analogues of our results for multiplicative upto approximation.
While our lower bound completely resolves the open question of whether a polynomial menusize 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 menusize 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 optimaltransport 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 (MenuSize: Conditional Upper Bound).
There exists such that for every and for every distribution satisfying the McAfeeMcMillan hazard condition (see Definition 1 in Section 3), there exists a mechanism with menusize at most such that .
The question of a precise tight bound on the menusize required for upto 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 menusize, 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 highlevel 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 highlevel 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 .^{4}^{4}4As 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 optimaltransport 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 (optimaltransport) problem respectively, it holds that the value of the primal objective function for the primal solution is upperbounded 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 (revenuemaximizing 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 menusize.
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 menusize, 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 menusize. Another way to state this conclusion is to observe that it follows from known properties that in a mechanism with a finite menusize such a curve must be piecewiselinear rather than strictly concave, and so to conclude that the optimal mechanism cannot have a finite menusize. 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 piecewiselinear curve with number of pieces at most the menusize of the given mechanism, and use this to appropriately lowerbound this metric (see Proposition 5 below) for mechanisms with small menusize. A lower bound on the loss in revenue for mechanisms with such a menusize follows.
To present our analysis in further detail, we must now first dive into some of the details of the optimaltransport duality framework of Daskalakis et al. (2015).
2.2 Minimal Needed Essentials of the OptimalTransport 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 measure^{5}^{5}5To 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 ,^{6}^{6}6That is, the overall measure sums up to . such that for a mechanism with utility function^{7}^{7}7The 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 to^{8}^{8}8This Lebesgue integral is the measuretheoretic 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 revenuemaximizing 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 menusize.)
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) menusize (in particular, the graph of such is a maximum of planes, where is the menusize), 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 dualitybased 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.^{9}^{9}9For 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 coupling^{10}^{10}10Informally (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 convexdominates^{11}^{11}11A distribution convexdominates a distribution if is obtained from by shifting mass to coordinatewise larger points and by performing meanpreserving 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 meanpreserving spreads of positive mass. , it is the case that^{12}^{12}12Once again, the Lebesgue integral on the righthand side is the measuretheoretic 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 measure^{13}^{13}13For 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).^{14}^{14}14In 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 menusize, the complementary slackness conditions with respect to and will be sufficiently violated so as to give sufficient separation between the lefthand side of Equation 5 for (that is, the revenue of the mechanism with small menusize) and the righthand 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 everywhere^{15}^{15}15Informally (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 menusize, 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.
For our proof below it suffices to describe the measure and the coupling , both restricted to a region for an appropriate .^{16}^{16}16We choose as the horizontalaxis 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 positivedensity points to negativedensity 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 revenuemaximizing 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 revenuemaximizing mechanism), thus concluding that the unique revenuemaximizing mechanism for the distribution has an infinite menusize.
2.3 Proof of Theorem 2
An alternative way to state the conclusion of the argument of Daskalakis et al. (2015) that any revenuemaximizing mechanism for has an infinite menussize is as follows: let be the utility function of a mechanism with a finite menusize. 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 menusize of such . We would like to reason as follows: for such with a finite menusize, define the corresponding curve that is the analogue for of the curve , and then show that since must be piecewiselinear, 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 menusize.) 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,^{17}^{17}17Similarly, 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 piecewiselinear 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 menusize at most , and let be defined as follows:
It is well known that the graph of is the maximum of planes. Therefore, is a piecewiselinear 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.)
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 Theorem^{18}^{18}18The 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 menusize 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 revenuemaximizing 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 menusize 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 revenuemaximizing 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 McAfeeMcMillan hazard condition if it has a differentiable density function satisfying
for every .
Theorem 6 (Pavlov, 2011).
For every distribution satisfying the McAfeeMcMillan 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 .
Proof of Theorem 4.
Let be a revenuemaximizing mechanism for as in Theorem 6. Let , and for every real number , denote by the roundingdown 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 menusize — 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 mostpreferred 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.^{19}^{19}19If 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 McAfeeMcMillan 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 twogood distribution supported on for any given , and give an upper bound of on the menusize 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 McAfeeMcMillan 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 menusize lower bound (Theorem 2) and the menusize upper bound of Hart and Nisan (2013) (Theorem 1), a tight bound on the minimum deterministic communication complexity guaranteed to suffice for running some upto revenuemaximizing 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 menusize required for revenue maximization up to an additive when selling any number of goods , we obtain our tight communicationcomplexity 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) revenuemaximizing mechanism with only two possible outcomes (and hence deterministic communication complexity of ), which simply offers the good for a suitably chosen takeitorleaveit price. Corollary 7 therefore shows a precise dichotomy in the asymptotic communication complexity of upto 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 menusize suffices for maximizing revenue up to a multiplicative^{20}^{20}20For 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) asked^{21}^{21}21Hart 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 menusize suffices, and moreover gave an upper bound of on the sufficient menusize. 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 (MenuSize for Multiplicative Approximation: Lower Bound).
There exist and a distribution , such that for every it is the case that for every mechanism with menusize 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 upto revenue maximization.
5 Discussion
In a very recent paper, Saxena et al. (2018) analyze the menusize required for approximate revenue maximization in what is known as the FedEx problem (Fiat et al., 2016). Interestingly, they also use piecewiselinear 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 piecewiselinear approximation on the revenue,^{22}^{22}22There: 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,^{23}^{23}23There: 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,^{24}^{24}24There: piecewiselinear; Here: strictly concave. the geometric/analytic proof of the bound on the number of linear segments, and finally, the argument that uses the piecewiselinear approximation and whether or not it couples the desired approximation with another parameter of the problem.^{25}^{25}25There: 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 piecewiselinear approximations of concave (or other) functions “pop up” in the future in any additional contexts in connection with bounds on the menusize of mechanisms.^{26}^{26}26Incidentally, other known derivations of menusize 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) piecewiselinear approximations.
As mentioned in Section 4, Babaioff et al. (2017) prove an upper bound on the menusize required for multiplicative upto 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 communicationcomplexity bound (see Corollary 9 in Section 4). Alternatively to fixing and analyzing the menusize 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 ;^{27}^{27}27This result also shows that the exponential dependence on in the upper bounds of Hart and Nisan (2013) and Dughmi et al. (2014) for additive upto 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 stateoftheart literature seems to be a long way from identifying veryhighdimensional 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 highdimensional optimal mechanisms (and corresponding duals) for which the highdimensional analogue of the curve that we denote by in Section 2 has largeenough measure (while maintaining a smallenough 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 largeenough highdimensional 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 IsraelUSA Binational 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 menusize 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. Multiitem 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 multiplegood 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 menusize 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 multiplegood 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 pseudodimension 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 quasilinear 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 “oneandahalfdimensional” mechanism design. In Proceedings of the 29th Annual ACMSIAM 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.