Symmetries and Optimal Multi-Dimensional Mechanism Design

CONSTANTINOS DASKALAKIS, Massachusetts Institute of Technology

S. MATTHEW WEINBERG, Massachusetts Institute of Technology

We efficiently solve the optimal multi-dimensional mechanism design problem for independent additive bid-
ders with arbitrary demands when either the number of bidders is held constant or the number of items is
held constant. In the first setting, we need that each bidder’s values for the items are sampled from a pos-
sibly correlated, item-symmetric distribution, allowing different distributions for each bidder. In the second
setting, we allow the values of each bidder for the items to be arbitrarily correlated, but assume that the
distribution of bidder types is bidder-symmetric. These symmetric distributions include i.i.d. distributions,
as well as many natural correlated distributions. E.g., an item-symmetric distribution can be obtained by
taking an arbitrary distribution, and “forgetting” the names of items; this could arise when different mem-
bers of a bidder population have various sorts of correlations among the items, but the items are “the same”
with respect to a random bidder from the population.
For all  > 0, we obtain a computationally efficient additive -approximation, when the value distribu-
tions are bounded, or a multiplicative (1 − )-approximation when the value distributions are unbounded,
but satisfy the Monotone Hazard Rate condition, covering a widely studied class of distributions in Eco-
nomics. Our running time is polynomial in max{#items,#bidders}, and not the size of the support of the
joint distribution of all bidders’ values for all items, which is typically exponential in both the number of
items and the number of bidders. Our mechanisms are randomized, explicitly price bundles, and in some
cases can also accommodate budget constraints.
Our results are enabled by several new tools and structural properties of Bayesian mechanisms, which
we expect to find applications beyond the settings we consider here; indeed, there has already been follow-
up research [Cai et al. 2012; Cai and Huang 2012] making use of our tools in both symmetric and non-
symmetric settings. In particular, we provide a symmetrization technique that turns any truthful mechanism
into one that has the same revenue and respects all symmetries in the underlying value distributions. We
also prove that item-symmetric mechanisms satisfy a natural strong-monotonicity property which, unlike
cyclic-monotonicity, can be harnessed algorithmically. Finally, we provide a technique that turns any given
-BIC mechansism (i.e. one where incentive constraints are violated by ) into a truly-BIC mechanism at the
√
cost of O( ) revenue.
Categories and Subject Descriptors: F.2.0 [Theory of Computation]: Analysis of Algorithms and Problem
Complexity
General Terms: Algorithms, Design, Economics, Theory
Additional Key Words and Phrases: Multi-Dimensional Mechanism Design, Revenue Optimization, Pricing,
Auction Theory, Symmetries

1. INTRODUCTION
How can a seller auction off a set of items to a group of interested buyers to maxi-
mize profit? This problem, dubbed the optimal mechanism design problem, has gained

Authors’ addresses and support: C. Daskalakis, [email protected]; Supported by a Sloan Foundation Fellowship
and NSF Awards CCF-0953960 (CAREER) and CCF-1101491.
S.M. Weinberg, [email protected]; Supported by a NSF Graduate Research Fellowship and a NPSC Graduate
Fellowship.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights
for components of this work owned by others than ACM must be honored. Abstracting with credit is per-
mitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component
of this work in other works requires prior specific permission and/or a fee. Permissions may be requested
from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or [email protected].
EC’12, June 4–8, 2012, Valencia, Spain. Copyright 2012 ACM 978-1-4503-1415-2/12/06...$10.00.

370
central importance in mathematical Economics over the past decades. The seller could
certainly auction off the items sequentially, using her favorite single-item auction, such
as the English auction. But this is not always the best idea, as it is easy to find exam-
ples where this approach leaves money on the table.1 The chief challenge that the
auctioneer faces is that the values of the buyers for the items, which determine how
much each buyer is willing to pay for each item, is information that is private to the
buyers, at least at the onset of the auction. Hence, the mechanism needs to provide
the appropriate incentives for the buyers to reveal “just enough” information for the
optimal revenue to be extracted.
Viewed as an optimization problem, the optimal mechanism design problem is of a
rather intricate kind. First, it is a priori not clear how to evaluate the revenue of an
arbitrary mechanism because it is not clear how rational bidders will play. One way to
cope with this is to only consider mechanisms where rational bidders are properly in-
centivized to tell the designer their complete type, i.e. how much they would value each
possible outcome of the mechanism (i.e. each bundle of items they may end up getting).
Such mechanisms can be Incentive Compatible (IC), where each bidder’s strategy is to
report a type and the following worst-case guarantee is met: regardless of the reports
of the other bidders, it is in the best interest of a bidder to truthfully report her type.
Or the mechanism can be Bayesian Incentive Compatible (BIC), where it is assumed
that the bidders’ types come from a known distribution and the following average-case
guarantee is met: in expectation over the other bidders’ types, it is in the best interest
of a bidder to truthfully report her type, if the other bidders report truthfully. See Sec 2
for formal definitions. We only note here that, under very weak assumptions, restrict-
ing attention to IC/BIC mechanisms in the aforementioned settings of without or with
prior information over bidders’ types is without loss of generality due to the revelation
principle [Nisan et al. 2007].
But even once it is clear how to evaluate the revenue of a given mechanism, it is not
necessarily clear what benchmark to compare it against. For example, it is not hard to
see that the social welfare, i.e. the sum of the values of the buyers for the items they are
allocated, is not the right benchmark to use, as in general one cannot hope to achieve
revenue that is within any constant factor of the optimal social welfare: why would
a buyer with a large value for an item pay an equally large price to the auctioneer
to get it, if there is no competition for this item? Given the lack of a useful revenue
benchmark (i.e. one that upper bounds the revenue that one may hope to achieve but
is not too large to allow any reasonable approximation), the task of the mechanism
designer can only be specified in generic terms as follows: come up with an IC/BIC
auction whose revenue is at least as large as the revenue of any other IC/BIC auction.
Finally, even after restricting the search space to IC/BIC auctions and only compar-
ing to the optimal revenue achievable by any IC/BIC auction, it is still easy to show
that it is impossible to guarantee any finite approximation if no prior is known over
the bidders’ types. Instead, many solutions in the literature adopt a Bayesian view-
point, assuming that a prior does exist and is known to both the auctioneer and the
bidders, and targeting the optimal achievable expected revenue. Once the leap to the
Bayesian setting is made the goal is typically this: Design a BIC, possibly randomized,

1A simple example is this: Suppose that an auctioneer is selling a Picasso and a Dali painting and there are
two bidders of which one loves Picasso and does not care about Dali and vice versa. Running a separate En-
glish auction for each painting will result in small revenue since there is going to be no serious competition
for either painting. But bundling the paintings together will induce competition and drive the auctioneer’s
revenue higher.

371
mechanism whose expected revenue is optimal among all BIC, possibly randomized,
mechanisms.2
One of the most celebrated results in this realm is Myerson’s optimal auction [My-
erson 1981], which achieves optimal revenue via an elegant design that spans several
important settings. Despite its significance, Myerson’s result is limited to the case
where bidders are single-dimensional. In simple terms, this means that each bidder
can be characterized by a single number (unknown to the auctioneer), specifying the
value of the bidder per item received. This is quite a strong assumption when the
items are heterogeneous, so naturally, after Myerson’s work, a large body of research
has been devoted to the multi-dimensional problem, i.e. the setting where the bidders
may have different values for different items/bundles of items. Even though progress
has been made in certain restricted settings, it seems that we are far from an optimal
mechanism, generalizing Myerson’s result; see survey [Manelli and Vincent 2007] and
its references for work on this problem in Economics.
Algorithmic Game Theory has also studied this problem, with an extra eye on the
computational efficiency of mechanisms. Chawla et al. [2007] study the case of a single
(multidimensional) unit-demand bidder with independent values for the items. They
propose an elegant reduction of this problem to Myerson’s single-dimensional setting,
resulting in a mechanism that achieves a constant factor approximation to the opti-
mal revenue among all BIC, possibly randomized [Chawla et al. 2010b], mechanisms.
For the same problem, Cai and Daskalakis [2011] recently closed the constant ap-
proximation gap against all deterministic mechanisms by obtaining polynomial-time
approximation schemes for optimal item-pricing. As for the case of correlated values,
it had been known that finding the optimal pricing (i.e. deterministic mechanism) is
highly inapproximable by [Briest and Krysta 2007], although no hardness results are
known for randomized mechanisms. In the multi-bidder setting, Chawla et al. [2010a],
Bhattacharya et al. [2010] and recently Alaei [2011] obtain constant factor approxima-
tions in the case of additive bidders or unit-demand bidders and matroidal constraints
on the possible allocations.
While our algorithmic understanding of the optimal mechanism design problem is
solid, at least as far as constant factor approximations go, there has been virtually no
result in designing computationally efficient revenue-optimal mechanisms for multi-
dimensional settings, besides the single-bidder result of [Cai and Daskalakis 2011]. In
particular, one can argue that the previous approaches [Alaei 2011; Bhattacharya et al.
2010; Chawla et al. 2007, 2010a] are inherently limited to constant factor approxima-
tions, as ultimately the revenue of these mechanisms is compared against the optimal
revenue in a related single-dimensional setting [Chawla et al. 2007, 2010a], or a con-
vex programming relaxation of the problem [Alaei 2011; Bhattacharya et al. 2010].
Our focus in this work is to fill this important gap in the algorithmic mechanism de-
sign literature, i.e. to obtain computationally efficient near-optimal multi-dimensional
mechanisms, coming -close to the optimal revenue in polynomial time, for any desired
accuracy  > 0. We obtain a Polynomial-Time Approximation Scheme (PTAS) for the
following two important cases of the general problem.
The BIC k-items problem. Given as input an arbitrary (possibly correlated) dis-
tribution F over valuation vectors for k items, a demand bound C, and an integer m,
the number of bidders, output a BIC mechanism M whose expected revenue is optimal
relative to any other, possibly randomized, BIC mechanism, when played by m addi-

2 Inview of the results of [Briest et al. 2010; Chawla et al. 2010b], to achieve optimal, or even near-optimal,
revenue in correlated settings, or even i.i.d. multi-item settings, we are forced to explore randomized mech-
anisms.

372
tive bidders with demand C whose valuation vectors are sampled independently from
F.
The BIC k-bidders problem. Given as input an integer n representing the number
of items, k item-symmetric3 distributions F1 , . . . , Fk , and demand bounds C1 , . . . , Ck
(one for each bidder), output a BIC mechanism M whose expected revenue is optimal
relative to any other, possibly randomized, BIC mechanism, when played by k additive
bidders with demands C1 , . . . , Ck respectively whose valuation vectors for the n items
are sampled independently from F1 , . . . , Fk .
In other words, the problems we study are where either the number of bidders is
large, but they come from the same population, i.e. each bidder’s value vector is sam-
pled from the same, arbitrary, possibly correlated distribution, or the number of items
is large, but each bidder’s value distribution is item-symmetric (possibly different for
each bidder). While these do not capture the problem of Bayesian mechanism design
in its complete generality, they certainly represent important special cases of the gen-
eral problem and indeed the first interesting cases for which computationally efficient
near-optimal mechanisms have been obtained. Before stating our main result, it is
worth noting that:
• When the number of bidders is large, it does not make sense to expect that the auc-
tioneer has a separate prior distribution for the values of each individual bidder for
the items. So our assumption in the k-items problem that the bidders are drawn from
the same population of bidders is a realistic one, and—in our opinion—the practically
interesting case of the general problem. Indeed, there are hardly any practical ex-
amples of auctions using bidder-specific information (think, e.g., eBay, Sotheby’s etc.)
A reasonable extension of our model would be to assume that bidders come from a
constant number of different sub-populations of bidders, and that the auctioneer has
a prior for each sub-population. Our results extend to this setting.
• When the number of items is large, it is still hard to imagine that the auction-
eer has a distribution for each individual item. In the k-bidders problem, we assume
that each bidder’s value distribution is item-symmetric. This certainly contains the
case where each bidder has i.i.d. values for the items, but there are realistic appli-
cations where values are correlated, but still item-symmetric. Consider the following
scenario. Suppose that the auctioneer has the same number of Yankees, Red Sox, and
White Sox baseball caps to sell. Moreover, suppose that k bidders are sampled from
a large population of bidders with the following characteristics: Each bidder in the
population is a fan of one of the three teams and has non-zero value for exactly one
of the three kinds of caps, but it is unknown to the auctioneer which kind that is and
what the value of the bidder for that kind is. (Hence, the values of a random bidder
for the caps are certainly non-i.i.d., as if the bidder likes a Red Sox cap then she will
equally like another Red Sox cap, but will have zero value for a Yankees cap.) Suppose
also that we are willing to make the assumption that all teams have approximately
the same number of fans in the population and those fans have statistically the same
passion for their team. Then a random bidder’s values for the items is drawn from an
item-symmetric distribution, or close to one, so we can handle such distributions. In
this case too, our techniques still apply if we deviate from the item-symmetric model
to models where there is a constant number of types of objects, e.g. caps and jerseys,
and symmetries do not permute types, but permute objects within the same type.

3 A distribution over Rn is symmetric if, for all ~

v ∈ Rn , the probability it assigns to ~v is equal to the proba-
bility it assigns to any permutation of ~v .

373
 T HEOREM 1.1. (Additive approximation) For all k, if F samples values from [0, 1]k
there exists a PTAS with additive error  for the BIC k-items problem. For all k, if Fi
samples vectors from [0, 1]n , there exists a PTAS with additive error  · max{Ci } for the
BIC k-bidders problem.

Remark 1.2. Some qualifications on Theorem 1.1 are due.

— The mechanism output by our PTAS is truly BIC, not -BIC, and there are no extra
assumptions necessary to achieve this.
— We make no assumptions about the size of the support of Fi or F, as the runtime
of our algorithms does not depend on the size of the support. This is an important
distinction between our work and the literature where it is folklore knowledge that if
one is willing to pay computational time polynomial in the size of the support of the
value distribution, then the optimal mechanism can be easily computed via an LP
(see, e.g., [Bhattacharya et al. 2010; Briest et al. 2010; Dobzinski et al. 2011]). How-
ever, exponential size supports are easy to observe. Take, e.g., our k-bidders prob-
lem and assume that every bidder’s value for each of the items is i.i.d. uniform in
{$5, $10}. The naı̈ve LP based approach would result in time polynomial in 2n , while
our solution needs time polynomial in n.
— If we are willing to replace BIC by -BIC (or -IC) in Thm 1.1 and compare our rev-
enue to the best revenue achievable by any BIC (or respectively IC) mechanism, then
we can also accommodate budget constraints for our bidders. The only step of our
algorithm that does not respect budgets or does not apply to IC mechanisms is the
-BIC to BIC reduction (Thm 3.3). For space considerations, we restrict our attention
to BIC throughout the main body of the paper and prove the related claims for IC in
Appendix K of the full version [Daskalakis and Weinberg 2011].
— If the value distributions are discrete and every marginal has constant-size support,
then our algorithms achieve exactly optimal revenue in polynomial time, even though
the support of such a distribution may well be exponential. For instance, in the exam-
ple given in the second bullet our algorithm obtains exactly optimal revenue in time
polynomial in n. In these cases, we can find optimal truly BIC or truly IC mechanisms
that also accommodate budget constraints.
— The mechanisms produced by our techniques satisfy the demand bounds of each bid-
der in a strong sense (and not in expectation). Moreover, the user of our theorem is
free to choose whether they want to satisfy ex-interim individual rationality, that
the expected value of a bidder for the received bundle of items is larger than the ex-
pected price she pays, or ex-post individual rationality, where this constraint is true
with probability 1 (and not just in expectation). We focus the main presentation on
producing mechanisms that are ex-interim IR. In Appendix D of the full version we
explain the required modification for producing ex-post IR mechanisms without any
loss in revenue.
— The assumption that Fi , F sample from [0, 1]n as opposed to some other bounded set
is w.l.o.g. and previous work has made the same assumption on the input distribu-
tions [Hartline and Lucier 2010; Hartline et al. 2011].
— The point of the additive error of  · max{Ci } is not to set  so small that it cancels
out the factor of max{Ci }, but rather to accept the factor of max{Ci } as lost revenue.
Notice that, if each vij has expectation that is bounded away from 0, the maximum
attainable revenue scales with max{Ci }. In this case, an additive error of  · max{Ci }
corresponds to a (1 − O())-factor approximation to the optimal revenue. As another
example, when each marginal satisfies the Monotone Hazard Rate condition, an ad-
ditive error of  · max{Ci } as obtained by Theorem 1.1 can be readily used to obtain

374
 a multiplicative (1 − O())-factor approximation. (For details see the proof of Corol-
lary 1.3 below in Appendix J of the full version.)
One might prefer to assume that the value distributions are not upper bounded, but
satisfy some tail condition, such as the Monotone Hazard Rate (MHR) condition. (The
class of MHR distributions includes such familiar distributions as the Normal, Ex-
ponential and Uniform distributions, and is widely used in Mechanism Design. See
Appendix J of the full version for a precise definition.) Using techniques from [Cai and
Daskalakis 2011], we can extend our results to MHR distributions. All the relevant
remarks still apply.
C OROLLARY 1.3. (Multiplicative approximation for MHR distributions) For all k,
if the k marginals of F all satisfy the MHR condition, there exists a PTAS obtaining
at least a (1 − )-fraction of the optimal revenue for the BIC k-items problem (whose
runtime does not depend on F or C). Likewise, for all k, if every marginal of Fi is MHR
for all i, there exists a PTAS obtaining at least a (1 − )-fraction of the optimal revenue
for the BIC k-bidders problem (whose runtime does not depend on Fi or Ci ).
1.1. Tools and Techniques
Our main result (Theorem 1.1) is enabled by establishing several new tools and struc-
tural results for multi-dimensional mechanism design, which are of applicability far
beyond the symmetric auction-settings we focus on. (See, e.g., next section.)
Our first contribution is a symmetrization technique, stated informally below and
formally as Theorem 3.1. This technique enables one to save on the description com-
plexity of the optimal mechanism by exploiting any kind of symmetry in the underlying
joint distribution D from which the bidders’ values for the items are sampled (no mat-
ter how pervasive that symmetry may be in D). In the highly symmetric settings of
Theorem 1.1, our symmetrization technique bears witness to a succinct (i.e. poly-size)
description of the revenue-optimal mechanism.
I NFORMAL T HEOREM 1. For all distributions D and all mechanisms M , M can be
symmetrized into a new mechanism M 0 , such that M 0 obtains the same expected revenue
as M , retains the truthfulness of M (such as IC, BIC, -IC, or -BIC.), and satisfies all
symmetries that D has.
In addition, we prove a structural property of item-symmetric mechanisms which
we call strong-monotonicity.4 This structure enables one to efficiently find an optimal
mechanism, as well as efficiently implement the -BIC to BIC reduction (described be-
low) in item-symmetric settings. The theorem is stated informally below, and formally
as Theorem 3.2.
I NFORMAL T HEOREM 2. Every BIC, item-symmetric mechanism for an item-
symmetric distribution D is strongly-monotone. That is, each bidder receives a higher
valued item more probably than a lower valued item, where the probability is with
respect to the randomness in the other bidders’ values and the mechanism.
Finally, we provide an -BIC to BIC reduction that turns an -BIC mechanism into
a BIC mechanism, while approximately preserving the revenue of the mechanism. We
note here that the rounding procedure attributed to Nisan in [Chawla et al. 2007]
doesn’t suffice in multi-bidder scenarios. Simply put, incentivizing a single bidder to
purchase a more expensive option affects other bidders as well, and this procedure
may completely destroy any truthfulness the original mechanism had. Therefore, a

4 Theterm strong monotonicity has also been used in other contexts. There is no connection between our use
and other usages.

375
new technique that considers the interaction between the bidders is necessary. We
prove the following theorem, stated informally below and formally as Theorem 3.3.

I NFORMAL T HEOREM√ 3. Any -BIC mechanism can be converted to a BIC mecha-

nism at a cost of O( ) revenue. Furthermore, the conversion runs in polynomial time
for inputs to the k-bidders or k-items problem.

1.2. Related Work

As discussed earlier, several polynomial-time constant-factor approximation algo-
rithms have been recently discovered for multi-dimensional settings [Bhattacharya
et al. 2010; Chawla et al. 2010a; Alaei 2011]. Our work differs from these in that we
obtain nearly-optimal revenue, i.e. a PTAS.
Other recent results have also obtained a PTAS for settings related to ours [Cai
and Daskalakis 2011; Cai and Huang 2012; Cai et al. 2012]. However, the settings
covered here are more general in that we allow arbitrary demand bounds and possibly
budget constraints. Instead, the results in [Cai and Daskalakis 2011] apply to a single,
unit-demand bidder with no budget constraints, while those of [Cai and Huang 2012]
and [Cai et al. 2012] apply to additive bidders with no demand or budget constraints.
In terms of techniques, these papers provide tools that interface nicely with the tools
developed here. Specifically, Corollary 1.3 in our paper makes use of an extreme value
theorem from [Cai and Daskalakis 2011]. In the other direction, our -BIC to BIC
reduction (Theorem 3.3) is used in both [Cai and Huang 2012] and [Cai et al. 2012].
In addition, [Cai et al. 2012] makes use of our structural theorems (Theorems 3.1 and
Theorem 3.2) as well. In other words, the tools provided in this paper are orthogonal to
the tools developed in [Cai and Daskalakis 2011; Cai and Huang 2012; Cai et al. 2012],
and our results here are a necessary prerequisite for comparable theorems in [Cai and
Huang 2012; Cai et al. 2012].

1.3. Structure
The rest of the paper is organized as follows: Sec 2 provides a few standard definitions
from Mechanism Design. Sec 3 gives an overview of our proof of Thm 1.1, explaining
the different components that get into our proof and guiding through the rest of the
paper. The rest of the main body and the appendix of the full version [Daskalakis and
Weinberg 2011] provide all technical details. Appendix J of the full version provides
the proof of Corollary 1.3.

2. PRELIMINARIES AND NOTATION

We assume that the seller has a single copy of n (heterogeneous) items that she wishes
to auction to m bidders. Each bidder i has some non-negative value for item j which
we denote vij . We can think of bidder i’s type as an n-dimensional vector ~vi , and denote
the entire profile of bidders as ~v , or sometimes (~vi ; ~v−i ) if we want to emphasize its
decomposition to the type ~vi of bidder i and the joint profile ~v−i of all other bidders.
We denote by D the distribution from which ~v is sampled. We also denote by Di the
distribution of types for bidder i, and by D−i the distribution of types for every bidder
except i. The value of an additive bidder with demand C for any subset of items is the
sum of her values for her favorite C items in the subset; that is, we assume bidders
are additive up to their demand.
Since we are shooting for BIC/IC mechanisms, we will only consider (direct revela-
tion) mechanisms where each bidder’s strategy is to report a type. When the reported
bidder types are ~v , we denote the (possibly randomized) outcome of mechanism M as
M (~v ). The outcome can be summarized in: the expected price charged to each bidder

376
(denoted pi (~v )), and a collection of marginal probabilities φ(~ ~ v ) = (φij (~v ))ij , where φij (~v )
denotes the marginal probability that bidder i receives item j.
A collection of marginal probabilities φ(~ ~ v ) := (φij (~v ))ij is feasible iff there exists a
consistent with them joint distribution over allocations of items to bidders so that in
addition, with probability 1, no item is allocated to more than one bidder, and no bidder
receives more items than her demand. A straightforward application of the Birkhoff-
von Neumann theorem [Johnson et al. 1960] reveals that a sufficient condition for the
above to hold is that in expectation no item is given more than once, and all bidders
receive an expected number of items less than or equal to their demand. Note that this
sufficient condition is expressible in terms of the φij ’s only. Moreover, under the same
conditions, we can efficiently sample a joint distribution with the desired φij ’s. (See
Appendix C of the full version [Daskalakis and Weinberg 2011] for details.)
The outcome of mechanism M restricted to bidder i on input ~v is denoted Mi (~v ) =
~ i (~v ), pi (~v )). Assuming that bidder i is additive (up to her demand) and risk-neutral
(φ
and that the mechanism is feasible (so in particular it does not violate the bidder’s de-
mand constraint) the value of bidder i for outcome Mi (w) ~ is just (her expected value) ~vi ·
~ ~ while the bidder’s utility for the same outcome is U (~vi , Mi (w))
φi (w), ~ := ~vi · φ ~ i (w)−p
~ i (w).
~
Such bidders subtracting price from expected value are called quasi-linear. Moreover,
for a given value vector ~vi for bidder i, we write: πij (~vi ) = E~v−i ∼D−i [φij (~vi ; ~v−i )].
We proceed to formally define incentive compatibility of mechanisms in our notation:
Definition 2.1. (BIC/-BIC/IC/-IC Mechanism) A mechanism M is called -BIC iff
the following inequality holds for all i, ~vi , w ~ i:
X
E~v−i ∼D−i [U (~vi , Mi (~v ))] ≥ E~v−i ∼D−i [U (~vi , Mi (w
~ i ; ~v−i ))] − vmax · πij (w
~ i ),
j

where vmax is the maximum possible value of any bidder for any item in the support of
the value distribution. In other words, M is -BIC iff when a bidder lies by reporting w ~i
instead of ~vi , they do not expect to gain more than vmax times the expected number of
items that w ~ i receives. Similarly,
P ~ i , ~v−i : U (~vi , Mi (~v )) ≥
M is called -IC iff for all i, ~vi , w
~ i ; ~v−i )) − vmax · j φij (w
U (~vi , Mi (w ~ i ; ~v−i ). A mechanism is called BIC iff it is 0-BIC
and IC iff it is 0-IC.5
In our proof of Thm 1.1 throughout this paper we assume that vmax = 1. If vmax < 1,
we can scale the value distribution so that this condition is satisfied. We also define
individual rationality of BIC/-BIC mechanisms:
Definition 2.2. A BIC/-BIC mechanism M is called ex-interim individually rational
(ex-interim IR) iff for all i, ~vi :
E~v−i ∼D−i [U (~vi , Mi (~v ))] ≥ 0.
It is called ex-post individually rational (ex-post IR) iff for all i, ~vi and ~v−i ,
U (~vi , Mi (~v )) ≥ 0 with probability 1 (over the randomness in the mechanism).
While we focus the main presentation on obtaining ex-interim IR mechanisms, in Ap-
pendix D of the full version [Daskalakis and Weinberg 2011] we describe how without
any loss in revenue we can turn these mechanisms into ex-post IR.

5 Any feasible mechanism that we call -BIC, respectively -IC, by our definition is certainly an  · max{C }-
i
P respectively  · max{C
BIC, P i }-IC, mechanism by the more standard definition, which omits the factors
~ i ), respectively
j πij (w ~ i ; ~v−i ), from the incentive error. We only include these factors here
j φij (w
for convenience.

377
 For a mechanism M , we denote by RM (D) the expected revenue of the mechanism
when bidders sampled from D play M truthfully. We also let ROP T (D) (resp. ROP T (D))
denote the maximum possible expected revenue attainable by any BIC (resp. -BIC)
mechanism when bidders are sampled from D and play truthfully. For all cases we
consider, these terms are well-defined.
We state and prove our results assuming that we can exactly sample from all input
distributions efficiently and exactly evaluate their cumulative distribution functions.
Our results still hold even if we only have oracle access to sample from the input distri-
butions, as this is sufficient for us to approximately evaluate the cumulative functions
to within the right accuracy in polynomial time (by making use of our symmetry and
discretization tools, described in the next section). The approximation error on eval-
uating the cumulative functions is absorbed into loss in revenue. See discussion in
Appendix A of the full version.
Finally, we denote by Sm , Sn the symmetric groups over the sets [m] := {1, . . . , m}
and [n] respectively. Moreover, for σ = (σ1 , σ2 ) ∈ Sm × Sn , we assume that σ maps
element (i, j) ∈ [m] × [n] to σ(i, j) := (σ1 (i), σ2 (j)). We extend this definition to map
a value vector ~v = (vij )i∈[m],j∈[n] to the vector w ~ such that w
~ σ(i,j) = ~vij , for all i, j.
Likewise, if D is a value distribution, σ(D) is the distribution that first samples ~v from
D and then outputs σ(~v ).

3. OVERVIEW OF OUR APPROACH

A Naı̈ve LP Formulation. Let D be the distribution of all bidders’ values for all
items (supported on a subset of Rm×n , where m is the number of bidders and n is
the number of items). For a mechanism design problem with unit-demand bidders
whose values are distributed according to D, it is folklore knowledge how to write a
linear programming formulation of size polynomial in |supp(D)| optimizing revenue.
The relaxation keeps track of the (marginal) probability φij (~v ) ∈ [0, 1] that item j is
given to bidder i if the bidders’ values are ~v , and enforces feasibility constraints (no
item is given more than once in expectation, no bidder gets more than one item in
expectation), incentive compatibility constraints (in expectation over the other bidders’
values for the items, no bidder has incentive to misreport her values for the items, if
the other bidders don’t), while optimizing the expected revenue of the mechanism.
Notice that all constraints and the objective function can be written in terms of the
marginals φij . Moreover, using the Birkhoff-von Neumann decomposition theorem, it
is possible to convert the solution of this LP to a mechanism that has the same revenue
and satisfies the feasibility constraints strongly (i.e. not just in expectation, but also
with probability 1). We give the details of the linear program in Appendix B of the full
version [Daskalakis and Weinberg 2011], and also describe how to generalize this LP
to incorporate publicly known demand and budget constraints.
Despite its general applicability, the naı̈ve LP formulation has a major drawback in
that |supp(D)| could in general be infinite, and when it is finite it is usually exponential
in both m and n. For the settings we consider, this is always the case. For example, in
the very simple setting where D samples each value i.i.d. uniformly from {$5, $10}, the
support of the distribution becomes 2m×n . Such support size is obviously prohibitive if
we plan to employ the naı̈ve LP formulation to optimize revenue.
A Comparison to Myerson’s Setting. What enables succinct and computationally
efficient mechanisms in the single-item setting of Myerson? Indeed, the curse of dimen-
sionality discussed above arises even when there is a single item to sell; e.g., if every
bidder’s distribution has support 2 and the bidders are independent, then the number
of different bidder profiles is already 2m . What drives Myerson’s result is the realiza-
tion that there is structure in a BIC mechanism coming in the form of monotonicity:

378
 0 0
for all i, for all vi1 ≥ vi1 : E~v−i (φi1 (vi1 ; ~v−i )) ≥ E~v−i (φi1 (vi1 ; ~v−i )), i.e. the expected
probability that bidder i gets the single item for sale in the auction increases with the
value of bidder i, where the expectation is taken over the other bidders’ values. Unfor-
tunately, such a crisp monotonicity property of BIC mechanisms fails to hold if there
are multiple items, and even if it were present it would still not be sufficient in itself
to reduce the naı̈ve LP to a manageable size.
So what next? We argued earlier that the symmetric distributions considered in the
BIC k-items and the BIC k-bidders problems are very natural cases of the general
optimal mechanism design problem. We argue next that they are natural for another
reason: they enable enough structure for (i) the optimal mechanism to have small
description complexity, instead of being an unusable, exponentially long list of what
the mechanism ought to do for every input value vector ~v ; and (ii) the succinct so-
lution to be efficiently computable, bypassing the exponentially large naı̈ve LP. Our
structural results are discussed in the following paragraphs. The first is enabled by
exploiting randomization to transfer symmetries from the value distribution to the op-
timal mechanism. The second is enabled by proving a strong-monotonicity property of
all BIC mechanisms. Our notion of monotonicity is more powerful than the notion of
cyclic-monotonicity, which holds more generally but can’t be exploited algorithmically.
Together our structural results bring to light how the item- and bidder-symmetric set-
tings are mathematically more elegant than general settings with no apparent struc-
ture.
Structural Result 1: The Interplay Between Symmetries and Randomization. Since
the inception of Game Theory scientists were interested in the implications of symme-
tries in the structure of equilibria [Gale et al. 1950; Brown and Neumann 1950; Nash
1951]. In his seminal paper [Nash 1951], Nash showed a rather interesting structural
result, informally reading as follows: “If a game has any symmetry, there exists a Nash
equilibrium satisfying that symmetry.” Indeed, something even more powerful is true:
“There always exists a Nash equilibrium that simultaneously satisfies all symmetries
that the game may have.”
Inspired by Nash’s symmetry result, albeit in our different setting, we show a similar
structural property of randomized mechanisms.6 Our structural result is rather gen-
eral, applying to settings beyond those addressed in Thm 1.1, and even beyond MHR
or regular distributions. The following theorem holds for any (arbitrarily correlated)
joint distribution D.
T HEOREM 3.1. Let D be the distribution of bidders’ values for the items (supported
on a subset of Rm×n ). Let also S ⊆ Sm × Sn be an arbitrary set such that D ≡ σ(D), for
all σ ∈ S; that is, assume that D is invariant under all permutations in S. Then any BIC
mechanism M can be symmetrized into a mechanism M 0 that respects all symmetries
in S without any loss in revenue. I.E. for all bid vectors →
−
v the behavior of M 0 under →
−
v
and σ(→−v ) is identical (up to permutation by σ) for all σ ∈ S. The same result holds if
we replace BIC with -BIC, IC, or -IC.
While we postpone further discussion of this theorem and what it means for M to
behave “identically” to Sec 4, we give a quick example to illustrate the symmetries
that randomization enables in the optimal mechanism. Consider a single unit-demand
bidder and two items. Her value for each item is drawn i.i.d. from the uniform dis-
tribution on {4, 5}. It is easy to see that the only optimal deterministic mechanism
assigns price 4 to one item and 5 to the other. However, there is an optimal randomized

6 Weemphasize ‘randomized’, since none of the symmetries we describe holds for deterministic optimal
mechanisms.

379
mechanism that offers each item at price 4 12 , and the uniform lottery (1/2 chance of
getting item 1, 1/2 chance of getting item 2) at price 4. While item 1 and item 2 need to
be priced differently in the deterministic mechanism to achieve optimal revenue, they
can be treated identically in the optimal randomized mechanism. Thm 3.1 applies in
an extremely general setting: distributions can be continuous with arbitrary support
and correlation, bidders can have budgets and demands, we could be maximizing social
welfare instead of revenue, etc.
Structural Result 2: Strong-Monotonicity. Even though the naı̈ve LP formulation is
not computationally efficient, Thm 3.1 certifies the existence of a compact solution for
the cases we consider. This solution lies in the subspace of Rm×n spanned by the sym-
metries induced by D. Still Thm 3.1 does not inform us how to locate such a symmetric
optimal solution. Indeed, the symmetry of the optimal solution is not a priori capable
in itself to decrease the size of our naı̈ve LP to a manageable one. For this purpose
we establish a strong monotonicity property of item-symmetric BIC mechanisms (an
item-symmetric mechanism is one that respects every item symmetry; see Sec 4 for a
definition).
T HEOREM 3.2. If D is item-symmetric, every item-symmetric BIC mechanism is
strongly monotone:
for all bidders i, and items j, j 0 : vij ≥ vij 0 =⇒ E~v−i (φij (~v )) ≥ E~v−i (φij 0 (~v )).
i.e., if i likes item j more than item j 0 , her expected probability (over the other bidders’
values) of getting item j is higher. We give an analogous monotonicity property of IC
mechanisms in Appendix K of the full version [Daskalakis and Weinberg 2011].
From - to truly-BIC. Exploiting the aforementioned structural theorems we are
able to efficiently compute exactly optimal mechanisms for value distributions D
whose marginals on every item have constant-size support. (D itself can easily have
exponentially-large support if, e.g., the items are independent.) To adapt our solution to
continuous distributions or distributions whose marginals have non-constant support,
we attempt the obvious rounding idea that changes D by rounding all values sampled
from D down to the nearest multiple of some accuracy , and solves the problem on the
resulting distribution D . While we can argue that the optimal BIC mechanism for D
is also approximately optimal for D, we need to also give up on the incentive compat-
ibility constraints, resulting in an approximately-BIC mechanism where bidders may
have an incentive to misreport their values, but the incentive to misreport is always
smaller than some function of . A natural approach to eliminate those incentives to
misreport is to appropriately discount the prices for items or bundles of items charged
by the mechanism computed for D , generalizing the single-bidder rounding idea at-
tributed to Nisan in [Chawla et al. 2007]. Unfortunately, this approach fails to work in
the multi-bidder settings, destroying both the revenue and truthfulness. Simply put,
even though the discounts encourage bidders to choose more expensive options, these
choices affect not only the price they pay us, but the prices paid by other bidders as
well as the incentives of other bidders. Once we start rounding the prices, we could
completely destroy any truthfulness the original mechanism had, leaving us with no
guarantees on revenue.
Our approach is entirely different, comprising a non-trivial extension of the main
technique of [Hartline et al. 2011]. We run simultaneous VCG auctions, one per bidder,
where each bidder competes with make-believe replicas of himself, whose values are
drawn from the same value distribution where his own values are drawn from. The
goods for sale in these per-bidder VCG auctions are replicas of the bidder drawn from
the modified distribution D . These replicas are called surrogates. The intention is that
the surrogates bought by the bidders in the per-bidder VCG auction will compete with

380
each other in the optimal mechanism M designed for the modified distribution D .
Accordingly, the value of a bidder for a surrogate is the expected value of the bidder
for the items that the surrogate is expected to win in M minus the price the surrogate
is expected to pay. This is exactly our approach, except we modify mechanism M to
discount all these prices by a factor of 1 − O(). This is necessary to argue that bidders
choose to purchase a surrogate with high probability, as otherwise we cannot hope
to make good revenue. There are several technical ideas coming into the design and
analysis of our two-phase auction (surrogate sale, surrogate competition). We describe
these ideas in detail in Sec 6.2, emphasizing several important complications departing
from the setting of [Hartline et al. 2011]. Importantly, the approach of [Hartline et al.
2011] is brute force in |supp(Di )|. While this is okay for k-items, this takes exponential
time for k-bidders. In addition to showing the following theorem, we show how to make
use of Thm 3.2 to get the reduction to run in polynomial time in both settings.
T HEOREM 3.3. Consider a generic setting with n items and m bidders who are
additive up to some capacity. Let D := ×i Di and D0 := ×i Di0 be product distributions,
sampling every bidder independently from [0, 1]n . Suppose that, for all i, Di and Di0 can
be coupled so that, with probability 1, a value vector ~vi sampled from Di and a value
vector ~vi0 sampled from Di0 satisfy that vij ≥ vij
0
≥ vij − δ, ∀j. Then, for all η,  > 0,
0
any -BIC mechanism M1 for D can be transformed into a BIC mechanism M2 for D
such that RM2 (D) ≥ (1 − η) · RM1 (D0 ) − +4δη T , where T is the maximum number of
items that can be awarded by a feasible mechanism. Furthermore, if D and D0 are both
valid inputs to the BIC k-bidders or k-items problem, the transformation runs in time
polynomial in n and m. Moreover, for the BIC k-items problem, T = k and, for the BIC
k-bidders problem, T ≤ k maxi Ci , where Ci is the demand of bidder i.
Figure 1 shows how the various components discussed above interact with each other
to prove Theorem 1.1. The modifications required to establish Corollary 1.3 are pro-
vided in Appendix J of the full version [Daskalakis and Weinberg 2011].

Optimal Mechanism Efficient Solution

Theorem 2 rich in symmetries for discrete. D

Section 5
Rich Subgroup of Naïve LP Succinct LP
Symmetries in D. formulation formulation

Surrogate
Theorem 3 Auction
+ Theorem 4
Monotonicity of
Optimal Mechanisms
Efficient Solution
for continuous D .

Fig. 1. Our Proof Structure

4. SYMMETRY THEOREM
We provide the necessary definitions to understand exactly what our symmetry result
is claiming.

381
 Definition 4.1. (Symmetry in a Distribution) We say that a distribution D has sym-
metry σ ∈ Sm × Sn if, for all ~v ∈ Rm×n , P rD [~v ] = P rD [σ(~v )]. We also write D ≡ σ(D).
Definition 4.2. (Symmetry in a Mechanism) We say that a mechanism respects sym-
metry σ ∈ Sm × Sn if, for all ~v ∈ Rm×n , M (σ(~v )) = σ(M (~v )).
Definition 4.3. (Permutation of a Mechanism) For any σ ∈ Sm × Sn , and any mech-
anism M , define the mechanism σ(M ) as [σ(M )](~v ) = σ(M (σ −1 (~v ))).
We proceed to state our symmetry theorem; its proof can be found in Appendix E of the
full version [Daskalakis and Weinberg 2011].
Theorem 3.1. (Restated from Sec 3) For all D, any BIC (respectively IC, -IC, -BIC)
mechanism M can be symmetrized into a BIC (respectively IC, -IC, -BIC) mechanism
M 0 such that, for all σ ∈ Sm × Sn , if D has symmetry σ, M 0 respects σ, and RM (D) =
0
RM (D).
We note that Thm 3.1 is an extremely general theorem. D can have arbitrary cor-
relation between bidders or items, and can be continuous. One might wonder why
we had to restrict our theorem to symmetries in Sm × Sn and not arbitary permuta-
tions of the set [m] × [n]. In fact, after reading through our proof, one can see that the
same inequalities that make symmetries in Sm × Sn work also hold for symmetries in
S[m]×[n] . However, the mechanism resulting from our proof is not a feasible one, since
our transformation can violate feasibility constraints for symmetries σ ∈ / Sm × Sn .
We also emphasize a subtle property of our symmetrizing transformation: the trans-
formation takes as input a set of symmetries satisfied by D and a mechanism, and
symmetrizes the mechanism so that it satisfies all symmetries in the given set of
symmetries. Our transformation does not work if the given set of symmetries is not
a subgroup. Luckily the maximal subset of symmetries in Sm × Sn satisfied by a value
distribution is always a subgroup, and this enables our result.

5. OPTIMAL SYMMETRIC MECHANISMS FOR DISCRETE DISTRIBUTIONS

In this section, we solve the following problem: “Given a distribution D with constant
support per dimension and a subgroup of symmetries S ⊆ Sm × Sn satisfied by D, find
a BIC mechanism M that respects all symmetries in S and maximizes RM (D).” By
Thm 3.1, such M will in fact be optimal with respect to all mechanisms. Intuitively,
optimizing over symmetric mechanisms should require less work than over general
mechanisms, since we should be able to exploit the symmetry constraints in our op-
timization. Indeed, suppose that every bidder can report c different values for each
item, where c is some absolute constant. Then the naı̈ve LP of Section 3/Appendix B
of the full version [Daskalakis and Weinberg 2011] has size polynomial in cmn , where
m, n are the number of bidders and items respectively. In Sec 5.1 we give a simple
observation that reduces the number of variables and constraints of this LP for any
given S. This observation in itself is sufficient to provide an efficient solution to the
BIC k-items problem (in our constant-support-per-dimension setting), but falls short
from solving the BIC k-bidders problem. For the latter, we need another structural
consequence of symmetry, which comes in the form of a strong-monotonicity property
satisfied by all symmetric BIC mechanisms. Strong-monotonicity and symmetry to-
gether enable us to obtain an efficient solution to the BIC k-bidders problem in Sec 5.2
(still for our constant-support-per-dimension setting). We explicitly
P write the LPs that
find the optimal BIC mechanism. Simply tacking on a − · j πij (w ~ i ) to the correct
side of the BIC constraints yields an LP to find the optimal -BIC mechanism for any .
Efficiently solving non-constant/infinite supports per dimension is postponed to Sec 6.

382
5.1. Reducing the LP size for any Subgroup of Symmetries, and Solving the discrete BIC
k-items Problem
We provide an LP formulation that works for any S. Our LP is the same as the naı̈ve
LP of Figure 2 (Appendix B of the full version), except we drop some constraints of that
LP and modify its objective function as follows. Since our mechanism needs to respect
every symmetry in S, it must satisfy
φij (~v ) = φσ(i,j) (σ(~v )), ∀i, j, ~v , σ ∈ S and pi (~v ) = pσ(i) (σ(~v )), ∀i, ~v , σ ∈ S.
Therefore, if we define an equivalence relation by saying that ~v ∼S σ(~v ), for all
σ ∈ S, we only need to keep variables φij (~v ), pi (~v ) for a single representative from
each equivalence class. We can then use the above equalities to substitute for all non-
representative ~v ’s into the naı̈ve LP. This will cause some constraints to become dupli-
cates. If we let E denote the set of representatives, then we are left with the LP of Fig-
ure 3 in Appendix G of the full version, after removing duplicates. In parentheses at the
end of each type of variable/constraint is the number of distinct variables/constraints
of that type.
L EMMA 5.1. The LP of Fig. 3 in Appendix G of the full version has polynomial size
for the BIC k-items problem, if the support of every marginal of the value distribution
is an absolute constant.
5.2. Strong-Monotonicity, and Solving the discrete BIC k-bidders Problem
Unfortunately, the reduction of the previous section is not strong enough to make the
LP polynomial in the number of items n, even if S contains all item permutations and
there is a constant number of bidders. This is because a bidder can deviate to an ex-
ponential number cn of types, and our LP needs to maintain an exponential number
of BIC constraints. To remedy this, we prove that every item-symmetric BIC mech-
anism for bidders sampled from an item-symmetric distribution satisfies a natural
monotonicity property:
Definition 5.2. (Strong-Monotonicity of a BIC mechanism) A BIC or -BIC mecha-
nism is said to be strongly monotone if for all i, j, j 0 , vij ≥ vij 0 ⇒ πij (~
vi ) ≥ πij 0 (~
vi ). That
is, bidders expect to receive their favorite items more often.

Theorem 3.2. (Restated from Sec 3) If M is BIC and D and M are both item-
symmetric, then M is strongly monotone. If M is -BIC and D and M are both item-
symmetric, there exists a -BIC mechanism of the same expected revenue that is strongly
monotone.
The proof of Thm 3.2 can be found in Appendix F of the full version and is a direct con-
sequence of 2-cycle monotonicity and item symmetry. We note again that our notion
of strong-monotonicity is different than the notion of cyclic-monotonicity that holds
more generally, but is not sufficient for obtaining efficient algorithms. Instead strong-
monotonicity suffices due to the following:
O BSERVATION 1. When playing an item-symmetric, strongly monotone BIC mech-
~i with wij > wij 0 unless vij ≥ vij 0 .
anism, bidder v~i has no incentive to report any w
L EMMA 5.3. There exists a polynomial-size LP for the BIC k-bidders problem, if the
support of every marginal of the value distribution is an absolute constant. The LP is
shown in Fig. 4 of Appendix G of the full version.
We note that Theorem 3.2 is also true for IC and -IC mechanisms with the appropri-
ate definition of strong-monotonicity. The definition and proof are given in Appendix K
of the full version.

383
6. EFFICIENT MECHANISMS FOR GENERAL DISTRIBUTIONS
We use the results of Sec 5 to prove Thm 1.1. First, it is not hard to see that discretiz-
ing the value distribution to multiples of δ, for sufficiently small δ = δ(), and applying
Lemmas 5.1 and 5.3 yields an algorithm for computing an -BIC -optimal mechanism
for the k-items and k-bidders problems. The resulting technical difficulty is turning
these mechanisms into being 0-BIC. To do this, we employ a non-trivial modification
of the construction in [Hartline et al. 2011] to improve the truthfulness of the mech-
anism at the cost of a small amount of revenue. We present our construction and its
challenges in Sec 6.2.

6.1. A Warmup: -Truthful Near-Optimal Mechanisms

Discretization:. Let D be a valid input to the BIC k-items or the BIC k-bidders prob-
lem. For each i, create a new distribution Di0 that first samples a bidder from Di , and
rounds every value down to the nearest multiple of δ. Let D0 be the product distribu-
tion of all Di0 ’s. Let also T denote the maximum number of items that can be awarded
by a feasible mechanism. We show the following lemma whose proof can be found in
Appendix H of the full version.

L EMMA 6.1. For all δ, let M 0 be the optimal δ-BIC mechanism for D0 . Then
M0
R (D0 ) ≥ ROP T (D) − δT . Moreover, let M be the mechanism that on input ~v rounds
0
every vij down to the nearest multiple vij of δ and implements the outcome M 0 (~v 0 ). Then
M is 2δ-BIC for bidders sampled from D, and has revenue at least ROP T (D) − δT .

Now notice that our algorithms of Sec 5 allow us to find an optimal δ-BIC mechanism
M 0 for D0 . So an application of Lemma 6.1 allows us to obtain a 2δ-BIC mechanism for
D whose revenue is at least ROP T (D) − δT .

6.2. Truthful Near-Optimal Mechanisms: Proof of Theorems 3.3 and 1.1

We start this section with describing our -BIC to BIC transformation result (Thm 3.3)
arguing that it can be implemented efficiently in the BIC k-items and k-bidders set-
tings.7 Combining our transformation with the results of the previous sections we ob-
tain Thm 1.1 in the end of this section. Our transformation is inspired by [Hartline
et al. 2011], but has several important differences. We explicitly describe our trans-
formation, point out the key differences between our setting and that considered in
[Hartline et al. 2011], and outline the proof of correctness, postponing the complete
proof of Theorem 3.3 to Appendix I of the full version.

Algorithm Phase 1: Surrogate Sale

(1) Recall from the statement of Theorem 3.3 that D and D0 can be coupled so that,
whenever we have ~v sampled from D and ~v 0 sampled from D0 , we have vij ≥ vij 0
≥
0
vij − δ, for all i, j. Moreover, M1 is an -BIC mechanism for D , for some .
(2) Modify M1 to give each bidder a rebate of δ whenever they receive an item. Further
modify M1 to multiply all prices it charges by a factor of (1 − η). Call M the mecha-
nism resulting from these modifications. Interpret the η-fraction of the prices given
back as additonal rebates.

7 We will explicitly describe the transformation for the BIC k-items and k-bidders settings. For an arbi-
trary setting where m bidders sample their valuation vectors for n items independently (but not necessarily
identically) from [0, 1]n (allowing correlation among items), simply employ the BIC k-items transformation,
replacing k with n.

384
(3) For each bidder i, create r − 1 replicas sampled i.i.d. from Di and r surrogates
sampled i.i.d. from Di0 . Use r = ( ηδ )2 · m2 · β̂, where β̂ = ( 1δ + 1)k , for the k-items
transformation, and β̂ = (n + 1)1/δ+1 , for the k-bidders transformation.
(4) Ask each bidder to report v~i . For k-bidders only: Fix a permutation σ such that
viσ( j) ≥ viσ( j+1) , ∀j. For each surrogate and replica w ~i , permute w
~i into w ~i 0 satisfy-
0 0
ing wiσ( j) ≥ wiσ( j+1) , ∀j.
(5) Create a weighted bipartite graph with replicas (and bidder i) on the left and sur-
rogates on the right. The weight of an edge between replica (or bidder i) with type
r~i and surrogate of type s~i is r~i ’s utility for the expected outcome of s~i when play-
ing M (where the expectation is taken over the randomness of M and of the other
0
bidders assuming they are sampled from D−i ).
(6) Compute the VCG matching and prices. If a replica (or bidder i) is unmatched in
the VCG matching, add an edge to a random unmatched surrogate. The surrogate
selected for bidder i is whoever she is matched to.
Algorithm Phase 2: Surrogate Competition
(1) Let s~i denote the surrogate chosen to represent bidder i in phase one, and let ~s
denote the entire surrogate profile. Have the surrogates ~s play M .
(2) If bidder i was matched to their surrogate through VCG, charge them the VCG
price and award them Mi (~s). (Recall that this has both an allocation and a price
component; the price is added onto the VCG price.) If bidder i was matched to a
random surrogate after VCG, award them nothing and charge them nothing.
There are several differences between our transformation and that of [Hartline et al.
2011]. First, observe that, because D0 and M1 are explicitly given as input to our trans-
formation (via an exact sampling oracle from Di0 and explicitly specifying the outcome
awarded to every type ~vi sampled from Di0 , for all i), we do not have to worry about
approximation issues in calculating the edge weights of our VCG auctions in Phase
1. Second, in [Hartline et al. 2011], the surrogates are taking part in an algorithm
rather than playing a mechanism, and every replica has non-negative value for the
outcome of an algorithm because there are no prices charged. Here, however, replicas
may have negative value for the outcome of a mechanism because there are prices
charged. Therefore, some edges may have negative weights, and the VCG matching
may not be perfect. We have modified M to give rebates (phase 1, step 2) so that the
VCG matching cannot be far from perfect, and show that we do not lose too much rev-
enue from unmatched bidders. Finally, in the k-bidders problem, the vanilla approach
that does not permute sampled replicas and surrogates (like we do in phase 1, step
4 of our reduction) would require exponentially many replicas and surrogates to pre-
serve revenue. To maintain the computational efficiency of our reduction, we resort to
sampling only polynomially many replicas/surrogates and permuting them according
to the permutation induced by the bidder’s reported values. This may seem like it is
giving a bidder control over the distribution of replicas and surrogates sampled for
her. We show, exploiting the monotonicity results of Sec 5, that our construction is still
BIC despite our permuting the replicas and surrogates. We overview the main steps of
the proof of Thm 3.3 and give its complete proof in Appendix I of the full version. We
conclude this section with the proof of Thm 1.1.
Proof of Theorem 1.1: Choose D0 to be the distribution that samples from D and
rounds every vij down to the nearest multiple of δ. Let then M1 be the optimal δ-
BIC mechanism for D0 as computed by the algorithms of Section 5. By Lemma 6.1,
RM1 (D0 ) ≥ ROP T (D) − δT . Applying Thm 3.3 we obtain a BIC mechanism M2 such
that

385
 5δ
RM2 (D) ≥ (1 − η) · RM1 (D0 ) − T (1)
η
5δ
≥ ROP T (D) − η · ROP T (D) − (1 − η)δT − T. (2)
η

Notice that ROP T (D)≤ T . Hence, choosing η =  and δ = 2 , (2) gives

RM2 (D) ≥ ROP T (D) − O( · k) (for k-items); and (3)

X
RM2 (D) ≥ ROP T (D) − O( · Ci ) (for k-bidders). (4)
i
P
The proof of Theorem 1.1 is concluded by noticing that i Ci ≤ k maxi Ci and k is an
absolute constant. 2

REFERENCES
A LAEI , S. 2011. Bayesian Combinatorial Auctions: Expanding Single Buyer Mecha-
nisms to Many Buyers. In the 52nd Annual IEEE Symposium on Foundations of
Computer Science (FOCS).
B HATTACHARYA , S., G OEL , G., G OLLAPUDI , S., AND M UNAGALA , K. 2010. Budget
Constrained Auctions with Heterogeneous Items. In the 42nd ACM Symposium on
Theory of Computing (STOC).
B RIEST, P., C HAWLA , S., K LEINBERG, R., AND W EINBERG, S. M. 2010. Pricing Ran-
domized Allocations. In the Twenty-First Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA).
B RIEST, P. AND K RYSTA , P. 2007. Buying Cheap is Expensive: Hardness of Non-
Parametric Multi-Product Pricing. In the Eighteenth Annual ACM-SIAM Sympo-
sium on Discrete Algorithms (SODA).
B ROWN, G. W. AND N EUMANN, J. 1950. Solutions of Games by Differential Equations.
In H. W. Kuhn and A. W. Tucker (editors), Contributions to the Theory of Games.
Vol. 1. Princeton University Press, 73–79.
C AI , Y. AND D ASKALAKIS, C. 2011. Extreme-Value Theorems for Optimal Multidi-
mensional Pricing. In the 52nd Annual IEEE Symposium on Foundations of Com-
puter Science (FOCS).
C AI , Y., D ASKALAKIS, C., AND W EINBERG, S. M. 2012. An Algorithmic Characteri-
zation of Multi-Dimensional Mechanisms. In the 43rd Annual ACM Symposium on
Theory of Computing (STOC).
C AI , Y. AND H UANG, Z. 2012. Simple and Nearly Optimal Multi-Item Auction.
Manuscript.
C HAWLA , S., H ARTLINE , J. D., AND K LEINBERG, R. D. 2007. Algorithmic Pricing via
Virtual Valuations. In the 8th ACM Conference on Electronic Commerce (EC).
C HAWLA , S., H ARTLINE , J. D., M ALEC, D. L., AND S IVAN, B. 2010a. Multi-Parameter
Mechanism Design and Sequential Posted Pricing. In the 42nd ACM Symposium on
Theory of Computing (STOC).
C HAWLA , S., M ALEC, D. L., AND S IVAN, B. 2010b. The Power of Randomness in
Bayesian Optimal Mechanism Design. In the 11th ACM Conference on Electronic
Commerce (EC).
D ASKALAKIS, C. AND W EINBERG, S. M. 2011. On Optimal Multi-Dimensional Mech-
anism Design. arXiv Report.

386
D OBZINSKI , S., F U, H., AND K LEINBERG, R. D. 2011. Optimal Auctions with Cor-
related Bidders are Easy. In the 43rd ACM Symposium on Theory of Computing
(STOC).
G ALE , D., K UHN, H. W., AND T UCKER , A. W. 1950. On Symmetric Games. In
H. W. Kuhn and A. W. Tucker (editors), Contributions to the Theory of Games. Vol. 1.
Princeton University Press, 81–87.
H ARTLINE , J. D., K LEINBERG, R., AND M ALEKIAN, A. 2011. Bayesian Incentive Com-
patibility via Matchings. In the Twenty-Second Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA).
H ARTLINE , J. D. AND L UCIER , B. 2010. Bayesian Algorithmic Mechanism Design. In
the 42nd ACM Symposium on Theory of Computing (STOC).
J OHNSON, D. M., D ULMAGE , A. L., AND M ENDELSOHN, N. S. 1960. On an Algorithm
of G. Birkhoff Concerning Doubly Stochastic Matrices. Canadian Mathematical Bul-
letin 3, 3, 237–242.
M ANELLI , A. M. AND V INCENT, D. R. 2007. Multidimensional Mechanism Design:
Revenue Maximization and the Multiple-Good Monopoly. Journal of Economic The-
ory 137, 1, 153–185.
M YERSON, R. B. 1981. Optimal Auction Design. Mathematics of Operations Re-
search 6, 1, 58–73.
N ASH , J. F. 1951. Non-Cooperative Games. Annals of Mathematics 54, 2, 286–295.
N ISAN, N., R OUGHGARDEN, T., T ARDOS, E., AND VAZIRANI , V. V., Eds. 2007. Algo-
rithmic Game Theory. Cambridge University Press.

387