4555 38402 1 PB

Theoretical Economics 19 (2024), 61–93 1555-7561/20240061
Worst-case equilibria in first-price auctions

Vitali Gretschko
School of Business and Economics, University of Muenster
Helene Mass
Institute for Microeconomics, University of Bonn
The usual analysis of bidding in first-price auctions assumes that bidders know
the distribution of valuations. We analyze first-price auctions in which bidders
do not know the precise distribution of their competitors’ valuations, but only the
mean of the distribution. We propose a novel equilibrium solution concept based
on worst-case reasoning. We find an essentially unique and efficient worst-case
equilibrium of the first-price auction that has appealing properties from both the
bidders’ and the seller’s point of view.
Keywords. Auctions, worst-case equilibria, uncertainty.
JEL classification. D44, D81, D82.
1. Introduction
Consider a bidder preparing a bid for a first-price auction. If her valuation for the auc-
tioned object is private and independently distributed from the valuation of her oppo-
nents, the only relevant information for an optimal bid is the bid distribution of her
competitors. Ideally, the bidder has access to data from prior auctions to estimate the
bid distribution. However, many important auction environments are either one shot
or infrequent. For example, auctions in mergers and acquisitions are usually one shot.
Spectrum auctions and auctions for sports rights take place infrequently in a changing
market environment. Moreover, data on prior auctions is not always revealed by the
sellers.
Without data, an alternative way for a bidder to prepare for the auction is to build a
model. Within the model, she optimizes her bid against the bid distribution generated
by her competitors’ bidding strategies and the distribution of their valuations. Typical
analyses of first-price auctions assume that the distributions of valuations are common
knowledge. Bidding strategies can then be derived by assuming that all bidders choose
Vitali Gretschko: [email protected]

Helene Mass: [email protected]
We would like to thank four anonymous referees, Dirk Bergemann, Daniil Larionov, Vitor Farinha Luz,
Achim Wambach, Philip Zilke, and seminar participants at the University of Bonn, Cologne University,
Duesseldorf University, the Econometric Society European Winter Meeting, the EARIE Conference, and the
University of Mannheim for their helpful comments and suggestions. We would also like to gratefully ac-
knowledge the financial support from the Leibniz Association through the project “Market Design for Public
Authorities” within the Leibniz Cooperative Excellence competition and from the CRC TR 224 (project B04).
© 2024 The Authors. Licensed under the Creative Commons Attribution-NonCommercial License 4.0.
Available at https://econtheory.org. https://doi.org/10.3982/TE4555
62 Gretschko and Mass Theoretical Economics 19 (2024)
optimal bids given the valuation distributions and the other bidders’ strategies, that is,
assuming that the competitors play a (Bayes–) Nash equilibrium.
In this paper, we analyze first-price auctions without a commonly known distribu-
tion of valuations. Bidders consider all distributions of their competitors’ valuations
with the same mean and lower and upper bound on the support, as feasible. A bid-
der preparing for such an auction faces two sources of uncertainty: the distribution of
valuations and the bidding strategies. Thus, the bidder cannot directly apply Nash equi-
librium reasoning, since this merely resolves the uncertainty about bidding strategies
given a unique distribution of valuations.
We assume that the bidder resolves the two uncertainties jointly by building a men-
tal model. Specifically, she first resolves the uncertainty about the valuation distribution
by choosing a feasible belief about it. To resolve the uncertainty about bidding strate-
gies, she additionally chooses a symmetric bidding strategy for her competitors. Then,
given the belief and the assumed bidding strategy, she can calculate a (hypothetical) bid
distribution for her competitors and choose a payoff-maximizing bid. We assume that
the choice of bidding strategy in the mental model is not arbitrary: we require that if ev-
eryone were to share the beliefs of the bidder whose mental model we are considering,
the bidding strategy would form a symmetric Nash equilibrium.
Given that the set of feasible distributions is large, so is the set of feasible mental
models. To choose between mental models, a bidder prepares for the worst case: she
considers the mental model that gives her the worst possible payoff, provided that she
chooses the optimal bid in that mental model.
We take an equilibrium perspective on worst-case mental models. That is, we re-
quire that if all bidders choose worst-case mental models and bids within their models,
the bidders’ best-reply to their competitors. We call the resulting equilibrium worst-case
equilibrium.
Our main finding is a characterization of a worst-case equilibrium in the first-price
auction. In particular, we show that a worst-case equilibrium exists. The payoff of a bid-
der is minimal in a mental model, in which her competitors bid as high as possible. In
this case, it is a best reply of this bidder to also bid as high as possible. Bidders, there-
fore, share the same beliefs about the strategies of their competitors in the worst case.
Overall, existence is then a consequence of the structure of best replies in a first-price
auction. Bidders with valuations below the mean of the distribution bid their valuation,
while bidders with valuations above the mean bid below their valuation and outbid any
bidder with a lower valuation. Thus, the worst-case equilibrium is efficient.
The worst-case equilibrium described above has several appealing properties. First,
it is unique for bidders with valuations above the mean. That is, in any other worst-case
equilibrium, those bidders must have the same beliefs and strategies. As bidders with
valuations below the mean always earn zero payoff in any worst-case equilibrium, we
say that the worst-case equilibrium is essentially unique, i.e., it is payoff unique. Sec-
ond, if a bidder bids according to the worst-case equilibrium strategy, she will outbid
all bidders with lower valuations, irrespective of the mental models they choose. Third,
the bid distributions generated from the strategies in the worst-case equilibrium cou-
pled with any true distribution of valuations first-order stochastically dominate the bid
Theoretical Economics 19 (2024) Equilibria in first-price auctions 63
distributions generated in the same way from any other mental model with an efficient
equilibrium. Fourth, from the seller’s point of view, the worst-case equilibrium maxi-
mizes revenue over all mental models. Moreover, seller revenue in the worst-case equi-
librium of the first-price auction is higher than revenue in a second-price auction for
any true distribution of valuations.
1.1 Related literature

There are several approaches to analyzing the behavior of a bidder whose information
about the distribution of her competitors’ valuations is consistent with multiple priors.
One way is to model the situation as a game of incomplete information with a non-
common prior endowing the bidders with subjective beliefs from the set of feasible dis-
tributions. A justification for such an approach is that bidders learn equilibrium play
from frequent interaction and from their observation of the empirical bid distributions.
Thus, Nash equilibrium with a correct common prior emerges as a steady state. Fuden-
berg and Levine (1993) conceptualize such learning by introducing a self-confirming
equilibrium (SCE), in which a bidder’s strategy is a best response to her beliefs about
her competitors, and beliefs are consistent with observed bids. Indeed, Esponda (2008)
shows that if, in a private-value first-price auction, the seller reveals the two largest bids,
then all SCE are Nash equilibria. However, if only the highest bid is revealed, the set of
SCE is larger than the set of Nash equilibria. Similarly, Dekel, Fudenberg, and Levine
(2004) show that without a common prior, not all SCE are Nash equilibria. Moreover,
learning takes time; thus, SCE is not a good approximation for infrequent or one-shot
auctions.
Another strand of the literature assumes that bidders are ambiguity averse and have
maximin preferences. They evaluate each bid with respect to their competitors’ strate-
gies and the valuation distribution that gives them the worst payoff given their bid (Bose,
Ozdenoren, and Pape (2006), Bodoh-Creed (2012), Di Tillio, Kos, and Messner (2016),
Lang and Wambach (2013), Lo (1998)). For example, Auster and Kellner (2022) analyze
the Dutch auction with ambiguity and ambiguity-averse bidders. They use maximin
preferences to evaluate bidding strategies. To account for the dynamic setting, they as-
sume prior-by-prior updating of the beliefs, which may lead to dynamic inconsistencies.
Thus, ambiguity aversion solves the problem of a multiplicity of priors by focusing on
the maximin, giving rise to sharp predictions about bidding behavior. By contrast, in this
paper we consider Bayesian bidders who maximize expected utility within their mental
models. Thus, our contribution is to provide a useful solution for first-price auctions
under ambiguity without dropping the assumption of expected utility maximization.
Knightian preferences, as introduced by Bewley (2002), have recently been used in
the analysis of auctions (Chiesa, Micali, and Zhu (2015)). In particular, Koçyiğit, Iyen-
gar, Kuhn, and Wiesemann (2020) introduce the concept of a Knightian Nash equilib-
rium. A profile of strategies constitutes a Knightian Nash equilibrium if, given the bid-
ding strategies of her competitors, the equilibrium bid of a bidder yields a higher pay-
off than any other bid coupled with any of the feasible beliefs about her competitors’
valuation distributions. Knightian Nash equilibria are not necessarily unique. Koçyiğit
et al. (2020) sidestep this issue by analyzing the equilibrium that gives the lowest seller
revenue. Thus, their analysis can be viewed as a robustness analysis for different mech-
anisms. By contrast, we take the viewpoint of the bidders rather than that of the seller.
Our worst-case concept yields an essentially unique equilibrium.
All the papers described above consider uncertainty only with respect to valuations,
not with respect to strategies. Three papers also consider first-price auctions with strate-
gic uncertainty: Kasberger and Schlag (2023), Kasberger (2020), and Mass (2023). They
derive strategies that minimize the maximal loss for any admissible strategy of the com-
petitors. However, unlike in our work, the resulting strategies do not form an equilib-
rium.
Bergemann, Brooks, and Morris (2017) analyze the first-price auction when the seller
is uncertain about the distribution of the bidders’ valuations. Bidders have a common
prior and know the information structure. Bergemann, Brooks, and Morris (2017) derive
bounds on the seller revenue under any information structure among the bidders. In
Bergemann, Brooks, and Morris (2019), the authors extend their results to all standard
auctions. In contrast to their work, we do not assume that the bidders have a common
prior or analyze the seller’s problem.
A recent literature focuses on the problem of a mechanism designer who does not
have precise beliefs about the participants in the mechanism. The papers in this litera-
ture either assume that the participants in the mechanism have a common prior (Azar,
Chen, and Micali (2012), Bergemann, Brooks, and Morris (2017)), analyze mechanisms
with a single participant (Bergemann and Schlag (2008, 2011), Carrasco, Luz, Kos, Mess-
ner, Monteiro, and Moreira (2018), Carroll (2015), Pınar and Kızılkale (2017)), or focus on
dominant-strategy incentive-compatible mechanisms (Allouah and Besbes (2020)). All
of them sidestep the issue of how participants behave if they face uncertainty about their
competitors. By contrast, we focus on bidder behavior under uncertainty and provide a
solution concept for such situations. However, we do not consider optimal mechanism
design; rather we focus on the first-price auction.
The idea that bidders form a valuation-dependent belief and best reply to their com-
petitors’ strategies given this belief within a mental model is related to the work of
Gagnon-Bartsch, Pagnozzi, and Rosato (2021), who consider bidders with beliefs biased
toward their own valuations. This assumption captures the idea that bidders overesti-
mate how similar their tastes are to others’. In Gagnon-Bartsch, Pagnozzi, and Rosato
(2021), each bidder forms a mental model by assigning her own biased belief to all of
her competitors, who she assumes play a Nash equilibrium in this hypothetical game of
incomplete information. Unlike in our work, the bidders’ beliefs about their competi-
tors’ strategies are not generically correct. However, the authors find, as we do, that the
first-price auction generates a higher revenue than the second-price auction.
2. Model
2.1 Environment
There are n risk-neutral and ex ante symmetric bidders competing in a first-price
sealed-bid auction for one indivisible object. Before the auction starts, each bidder
i ∈ {1, , n} = I privately observes her valuation θi ∈ = {0 = θ1 , θ2 , , θm−1 , 1 =

θm }, with θk > θl for k > l. For any vector (v1 , , vn ), we denote by v−i the vector
(v1 , , vi−1 , vi+1 , , vn ). The valuations are independently distributed with a common
distribution. We call this distribution the true distribution of valuations; we assume that
it is unknown to the bidders. However, it is common knowledge that the mean of this
distribution is μ. Hence, every bidder knows that the probability mass function of a
bidder’s valuations is an element of

m k m

Fμ = f : → [0, 1] f θ = 1 and θk f θk = μ .
k=1 k=1
In other words, Fμ is the set of all probability mass functions over the set with mean μ.
We focus on the symmetric case. That is, from the point of view of a bidder, all her
competitors have the same distribution.
The assumption that the mean characterizes the set of feasible distributions is fre-
quently used to analyze mechanisms under uncertainty about the distribution of val-
uations (distributional uncertainty). (For example, see Carrasco et al. (2018), Wolitzky
(2016), Azar and Micali (2013), or Pınar and Kızılkale (2017).) There are several ways
in which this assumption is interpreted in the literature. First, bidders may have only
a limited amount of data for a nonparametric estimation of the true distribution. Sec-
ond, bidders may acquire information about their own valuations before the auction
and may be uncertain about each other’s information acquisition technologies. Third,
the restriction of the set of possible distributions may be viewed as what makes the prob-
lem interesting. Fourth, bidders may learn that their competitors’ valuations lie in some
neighborhood, but may not be able to quantify the error.
In a first-price auction, the bidders submit bids b ∈ R+ ; the bidder with the highest
bid wins the object and pays her bid. Ties are broken in favor of bidders with higher
valuations (efficient tie-breaking). Thus, the payoff of bidder i with valuation θi and bid
bi , given that the other bids are b−i and valuations are θ−i , is denoted by
⎧
⎪
⎪θi − bi
⎪
⎪
if bi > max bj
j=i
⎪
⎪
⎪
⎪
⎪
⎪θ i − b i if b i = max bj and θi > max{θj | bj = bi }
⎪
⎪ j=i j=i
⎨
u(θi , θ−i , bi , b−i ) = 0 if b i = max b j and θ i < max {θj | bj = bi }
⎪ j=i j=i
⎪
⎪
⎪
⎪ 1
⎪
⎪ (θi − bi ) if bi = max bj and θi = max{θj | bj = bi }
⎪
⎪ p j=i j=i
⎪
⎪
⎪
⎩0 if bi < max bj ,
j=i
where θj denotes the valuation of bidder j with bid bj for j ∈ {1, , n} and p = #{j | θj =
θi ∧ bj = bi }.
We assume an efficient tie-breaking rule, since this simplifies notation. With a ran-
dom tie-breaking rule, all results are similar; the main difference is that we need to as-
sume a discrete bid grid (which may be arbitrarily fine) in order to ensure equilibrium
existence. With such a bid grid, the equilibrium strategies and beliefs under a random
tie-breaking rule differ from those under an efficient tie-breaking rule by at most one
step in the bid grid.
A symmetric (mixed) strategy β maps each bidder’s valuations to a distribution of
bids,
β : → R+

θk → β θk = Gk ,
where R+ is the set of all cumulative distribution functions on R+ and Gk denotes the
cumulative distribution function of bids from a bidder with valuation θk . That is, Gk (s)
is the probability that a bidder with valuation θk bids s or lower.
A pure strategy for a bidder with valuation θk is a mapping
β : → R+

θk → β θk
from the set of valuations to the set of bids. A pure strategy can be interpreted as a dis-
tribution of bids that puts probability weight 1 on a single bid. These definitions involve
an abuse of notation, since, in the case of a pure strategy, β(θk ) denotes an element in
R+ , while in the case of a (mixed) strategy, β(θk ) denotes an element in R+ . However,
in the following discussion, it will always be clear whether β is a pure or a mixed strategy.
In addition, we use the notation Gk instead of β(θk ) in the case of mixed strategies.
2.2 Worst-case mental models

To choose a bid, a bidder builds a mental model of the situation. A mental model is a
vehicle for a bidder to form a belief about the bid distribution of her competitors and
choose an optimal bid. For her mental model, bidder i with valuation θk chooses a fea-
sible belief f k ∈ Fμ about her competitors’ valuations. To reason about her competitors’
behavior, she assumes that they employ a symmetric bidding strategy β. Given the be-
lief f k and the assumed bidding strategy β, bidder i can calculate a (hypothetical) bid
distribution for her competitors and thereby determine the expected payoff of each bid
she may place.
β
Denote by b̃j the random variable representing the bid of bidder j = i given β. That
β m
is, Prob( b̃j ≤ s) = k l
l=1 f (θ )Gl (s). Denote by θ̃−i the random vector of the valuations
of bidders other than bidder i. The expected payoff of bidder i with valuation θk , belief
f k , and bid b, given that her competitors employ the symmetric bidding strategy β, is
then given by
β
U θk , f k , b, β = Eθ̃ , b̃β u θk , θ̃−i , b, b̃−i .
−i −i
The bidder can now choose a bid b to maximize U(θk , f k , b, β).

In a mental model, the bidder does not choose an arbitrary symmetric β for her
competitors. The chosen bidding strategy β reflects that for all possible valuations, the
competitors also best-reply to β and a feasible belief. Thus, a bidder with valuation
θk not only chooses a feasible belief f k about the valuations of her competitors, but
also assigns a belief to every valuation. If these beliefs are the same across bidders, β
constitutes a symmetric Nash equilibrium. Formally, a bidder building a mental model
considers a mapping
ϕ : → Fμ
θl → f l ,
assigning to each valuation θl a feasible belief ϕ(θl ) = f l , which describes the symmetric
and valuation-dependent belief of all bidders. Denote by Gϕ = (I, R+ , ϕ, u) the resulting
game of incomplete information. Note that the beliefs are the same across bidders, but
not necessarily the same across valuations. That is, bidder i and bidder j with valua-
tion θk will have the same belief, but bidder i will not necessarily have the same belief
with valuation θk as with valuation θl . The assumption that every bidder maximizes her
utility given these beliefs yields that β is a symmetric Nash equilibrium in Gϕ .
Definition 1. A bidding strategy β constitutes a symmetric Nash equilibrium in un-

dominated strategies of Gϕ if for every k ∈ {1, , m} and for every bid b ∈ supp(β(θk )),
it holds that

b ∈ arg max U θk , ϕ θk , b̂, β and b ≤ θk .
b̂∈R+
We call β an efficient equilibrium if for all θk , θl ∈ with θk > θl , for all b ∈

supp(β(θk )), and for all b ∈ supp(β(θl )), it holds that b ≥ b . Together with the efficient
tie-breaking rule, an efficient equilibrium implies that the bidder with the highest valu-
ation will win the object in equilibrium.
Any bid strictly above the own valuation induces a payoff of at most zero and is,
therefore, weakly dominated by bidding the valuation. Allowing for dominated bids may
lead to implausible equilibria. For example, a bidder with a valuation strictly below μ
can believe that all other bidders have a strictly higher valuation. In this case, there
exist equilibria in which such a bidder wins with probability 0 when placing a bid above
her valuation. To simplify exposition, in what follows we write “Nash equilibrium” as
shorthand for “Nash equilibrium in undominated strategies.”
We define a mental model as a (hypothetical) game of incomplete information with
beliefs ϕ and a Nash equilibrium β of this game.
Definition 2. A pair ( Gϕ , β) is a mental model if β is a Nash equilibrium of Gϕ . If β is an

efficient equilibrium, we speak of a mental model with an efficient equilibrium. Denote
by U k ( Gϕ , β) the equilibrium expected payoff of a bidder with valuation θk in ( Gϕ , β);
that is, U k ( Gϕ , β) = U(θk , ϕ(θk ), β(θk ), β).
Given that the set of feasible beliefs Fμ is large, so is the set of feasible mental models.
To choose between mental models, the bidder prepares for the worst case. That is, she
is pessimistic and constructs a mental model that gives her the worst possible expected
payoff given her valuation and provided that she best-replies in her mental model.
Definition 3. A mental model ( Gϕw , βw )k is a worst-case mental model for a bidder

with valuation θk for k ∈ {1, , m} if

U k Gϕw , βw k = min U k ( Gϕ , β),
( Gϕ ,β) is a mental model
that is, if the mental model minimizes the bidder’s expected payoff among all mental
models.
Worst-case mental models are not necessarily unique, as a bidder with a valuation
θk may achieve her worst payoff in different mental models. Moreover, the worst case
is defined with respect to the bidder’s valuation. Thus, bidders with different valuations
may, in principle, arrive at different worst-case mental models.
2.3 Worst-case equilibria

Up to this point we have taken the perspective of a single bidder building a mental model
of the situation, conditional on her valuation, making certain assumptions about her
competitors. In particular, she assumes beliefs for her competitors, that they choose
a symmetric bidding strategy, and that this bidding strategy forms a Nash equilibrium
given the beliefs. The bidder is pessimistic and constructs a worst-case mental model.
Next we take an equilibrium perspective on worst-case mental models. That is, we
require that if all bidders choose worst-case mental models and bids within their mod-
els, the bidders best-reply to their competitors. As bidding strategies within mental
models constitute a Nash equilibrium, this can be achieved by requiring that all bidders
choose the same worst-case mental model.
Definition 4. A mental model ( Gϕ∗ , β∗ ) is a symmetric worst-case equilibrium if it is

the worst-case mental model for all θk ∈ . That is, it minimizes the expected payoff
over all mental models ( Gϕ , β) for all valuations:

U k Gϕ∗ , β∗ = min U k ( Gϕ , β).
( Gϕ ,β) is a mental model
Worst-case equilibrium captures the idea that all bidders assume that they prepare
for the auction in the same way, that is, by considering worst-case mental models given
their valuations. The worst-case equilibrium is not self-defeating: if bidders were in-
formed about all of their competitors’ beliefs and bidding strategies, they would not
change their bids.
Definition 4 differs from Definition 3 in that the equilibrium worst-case mental
model does not depend on the bidders’ valuations. Bidders with all valuations choose
the same worst-case mental model. Therefore, as bidding strategies within a mental
model form a Nash equilibrium, all bidders best-reply to their competitors’ bidding
strategies.
It is not straightforward that a worst-case equilibrium exists, since bidders with dif-
ferent valuations could obtain their worst payoff in different mental models. A more
permissive equilibrium definition would merely require that each bidder best-replies to
the bidding strategies of her competitors given her beliefs in some worst-case mental
model. Such a definition would allow bidders with different valuations to choose differ-
ent worst-case mental models as long as each bidder best-replies to her beliefs and the
strategies of the other bidders. However, as we will show below, a worst-case equilib-
rium in the sense of Definition 4 always exists, and, therefore, it is not limiting to require
all bidders to choose the same worst-case mental model.
We illustrate the concept by means of a simple example.
Example 1. Consider two bidders with valuations = {0, 12 , 1} and a set of feasible
probability mass functions containing two functions {fa = ( 18 , 14 , 58 ), fb = (0, 12 , 12 )}. This
setup leads to 23 = 8 different mental models. However, as a bidder with valuation 0 al-
ways bids 0 in any mental model, only four of the models lead to different behaviors and,
thus, different payoffs. Consider first the mental model in which ϕ(0) = fa , ϕ( 12 ) = fa ,
and ϕ(1) = fa , which results in a standard symmetric first-price auction with a common
belief fa . The unique symmetric Nash equilibrium bidding function has the following
properties. A bidder with valuation 0 bids 0, a bidder with valuation 12 mixes her bids on
[0, 13 ], and a bidder with valuation 1 mixes her bids on [ 13 , 34 ]. The expected payoffs of
1
the bidders are 0 for a bidder with valuation 0, 16 for a bidder with valuation 12 , and 14
for a bidder with valuation 1.
The beliefs and the payoffs in the three other relevant mental models are
1 3
ϕ = (fa , fb , fa ), U 0 = 0, U 2 = 0, U1 =
16
1 1
ϕ = (fa , fb , fb ), U 0 = 0, U 2 = 0, U1 =
4
1 1 1
ϕ = (fa , fa , fb ), U 0 = 0, U2 = , U1 = .
16 3
For a bidder with valuation 0, each of these mental models is a worst-case mental model.
For a bidder with valuation 12 , the models with ϕ = (fa , fb , fa ) and ϕ = (fa , fb , fb ) are
worst-case mental models. For a bidder with valuation 1, the model with ϕ = (fa , fb , fa )
is a worst-case mental model. Thus, the worst-case equilibrium in the example is the
mental model with ϕ = (fa , fb , fa ). ♦
For our general setup, we show below that a worst-case equilibrium exists (Proposi-
tion 1). It is efficient and unique for bidders with valuations above the mean (Proposi-
tion 2). Moreover, the worst-case equilibrium has some appealing properties compared
with other mental models a bidder might consider. If a bidder bids according to the
worst-case equilibrium bidding strategy, she will outbid all bidders with lower valua-
tions, irrespective of the mental models they choose (Proposition 3). The bid distribu-
tions generated from the strategies in the worst-case equilibrium and some true val-
uation distribution first-order stochastically dominate the bid distributions generated
in the same way from any other mental model with an efficient equilibrium (Propo-
sition 4). From the point of view of the seller, the worst-case equilibrium maximizes
revenue over all mental models. Moreover, seller revenue in the worst-case equilibrium
of the first-price auction is higher than revenue in a second-price auction for any true
distribution of valuations (Proposition 5).
3. Worst-case equilibrium of the first-price auction

In this section, we characterize the beliefs and strategies in a worst-case equilibrium and
∗
show that it is essentially unique. We denote by [b∗k , bk ] the support of the bid distribu-
tion of a bidder with valuation θk in the worst-case strategy β∗ . We denote the worst-
case belief of a bidder with valuation θk by ϕ∗ (θk ) = f k,∗ = (f k,∗ (θ1 ), , f k,∗ (θm )).
More generally, for any belief ϕ, we use the notation ϕ(θk ) = f k = (f k (θ1 ), , f k (θm )),
with f k (θl ) denoting the probability that a bidder with valuation θk assigns to one of the
other n − 1 bidders having a valuation θl .
We start with a heuristic construction of the equilibrium to explain the underlying
intuition. We proceed in two steps, considering (i) bidders with valuations below μ and
(ii) bidders with valuations above μ. To organize the argument and highlight the main
insights we will present “observations” without formal proofs. At the end of the section,
we will summarize the results in our main proposition, Proposition 1. The formal proof
of the proposition can be found in the Appendix.
(i) Bidders with valuations below μ If a bidder’s valuation is below the mean, it is fea-
sible for her to believe that she has the lowest valuation. In any equilibrium with such
a belief, the bidder places a bid equal to her valuation. In this case, she earns a payoff
of 0 in equilibrium. Therefore, the belief that one has the lowest valuation must lead
to the worst case. To calculate a feasible belief for bidders with valuations θk ≤ μ, con-
sider θz , the lowest valuation strictly greater than μ. The belief that puts strictly positive
weight only on θk and θz induces a best reply of β(θk ) = θk in every mental model, i.e.,
for θk ≤ μ, it holds that bk = bk = θk . The probability weight is determined by

f k,∗ θk + f k,∗ θz = 1

f k,∗ θk θk + f k,∗ θz θz = μ.
The unique solution is given by
θz − μ μ − θk
f k,∗ θk = z , f k,∗ θz = z .
θ − θk θ − θk
Observation 1. Whenever θk ≤ μ, the worst-case bidding strategy is β∗ (θk ) = θk . Let

θz denote the lowest valuation strictly larger than μ. A worst-case belief is
f k,∗ θk = z , f k,∗ θz = z , f k,∗ θl = 0 for all l = k, z.
(ii) Bidders with valuations above μ Now consider the belief and bidding strategy of a
bidder with valuation θk > μ. It is infeasible for such a bidder to believe that she has the
lowest valuation. Thus, she earns a positive payoff in any mental model. We will show
below that bidders whose valuation θk is above μ play a mixed strategy without atoms
∗
on the interval [b∗k , bk ] according to a continuous bid distribution G∗k . Given the bid-
ding strategy, there are two levers to minimize a bidder’s payoff: reducing the winning
probability and increasing the best reply to her competitors’ equilibrium bid.
To minimize the winning probability, it is optimal to maximize the probability
weight on θk while respecting that the expected value of the valuations is μ. To see
that this minimizes the equilibrium payoff of a bidder with valuation θk , observe that in
a mixed-strategy equilibrium, she is indifferent between all bids in her bidding interval.
In particular, her equilibrium payoff is
n−1 k ∗
f k,∗ θ1 + · · · + f k,∗ θk−1 θ − bk−1 .
This expression is minimized given μ if the probability weight on θk is maximized in her

belief. Putting probability weight on types higher than θk would require an increase in
probability weight on types below θk in order to preserve the mean. However, if we sim-
ply maximized the probability weight on θk , we would distribute the probability weight
between valuations θk and 0. This would cause a bidder with valuation θk to bid 0.
Thus, to increase the best reply, the worst-case belief should put just enough probability
weight on valuations strictly below θk . This is achieved by making a bidder with valua-
tion θk exactly indifferent between her equilibrium bids and the bids of the bidders with
valuations lower than θk . This observation leads to the following system of equations.
Observation 2. The conditions

k

k

f k θl = 1 and f k θl θl = μ
l=1 l=1
and the equations

k−1 n−1
k ∗
f k θl θ − bk−1
l=1
n−1

h
∗
= f θl
k
θ k − bh for all h ∈ {1, , k − 2} (1)
l=1
constitute a system of linear equations. The worst-case belief (f k,∗ (θ1 ), , f k,∗ (θm )) is
the unique nonnegative solution to this system.
It remains to characterize the bidding functions and their support. The upper end-
point of the bidding interval of a bidder with valuation θk is obtained from the equation
n−1 k ∗
f k,∗ θ1 + · · · + f k,∗ θk−1 θ − b̄∗k−1 = θk − bk .
Recall that bidders with valuations below μ bid their valuation. Thus, we can calculate
the endpoints of the intervals inductively, starting with bidders with valuation 0. In par-
ticular, the system of equations in (1) depends only on the bid intervals of bidders with
lower valuations. Thus, there is no circularity between beliefs and bid distribution in the
derivation. The bid distribution G∗k is defined to be such that every bidder with valua-
tion θk is indifferent between every bid in her bidding interval given her belief and the
∗ ∗
other bidders’ strategies; i.e., for every s ∈ [bk−1 , bk ], it holds that
n−1 k
f k,∗ θ1 + · · · + f k,∗ θk−1 + f k,∗ θk G∗k (s) θ −s
n−1 k
= f k,∗ θ1 + · · · + f k,∗ θk−1 θ − b̄∗k−1 .
Rearranging yields our last observation.

∗ ∗
Observation 3. For every s ∈ [bk−1 , bk ], G∗k (s) is given by
k,∗ 1 k ∗ 1 1
f θ + · · · + f k,∗ θk−1 θ − bk−1 n−1 − θk − s n−1
G∗k (s) = 1 . (2)
f k,∗ θk θk − s n−1
∗
That is, the bid distribution and bk make a bidder with valuation θk indifferent between
∗
all bids in [b∗k , bk ].
We are now in a position to state our main proposition.
Proposition 1. The mental model ( Gϕ∗ , β∗ ) constitutes a symmetric worst-case equilib-

rium of the first-price auction. It holds that
• β∗ (θk ) = θk whenever θk ≤ μ
• β∗ (θk ) = G∗k whenever θk > μ.
That is, all bidders with valuations below μ bid their valuation and all bidders with valu-
∗ ∗
ations above μ play a mixed strategy on the interval [b∗k , bk ] with b∗k = bk−1 , according to
a continuous bid distribution G∗k given by (2). In particular, the worst-case equilibrium is
efficient.
Let θz > μ be the lowest valuation strictly larger than μ. A worst-case belief of a bidder
with valuation θk ≤ μ is given by
f k,∗ θk = z , f k,∗ θz = z , f k,∗ θl = 0 for all l = k, z,
while that of a bidder with valuation θk > μ is given by

f k,∗ θl > 0 for all l ≤ k, f k,∗ θl = 0 for all l > k,
with f k,∗ (θl ) for l ≤ k being the unique solution to the system of equations (1).
All proofs can be found in the Appendix. We illustrate the worst-case equilibrium
with two bidders for = {0, 0.25, 0.5, 0.75, 1} and μ = 0.5 in Table 1.
Table 1. Worst-case equilibrium with two bidders for = {0, 0.25, 0.5, 0.75, 1} and μ = 0.5.
Valuation f0k,∗ k,∗

f0.25 k,∗
f0.5 k,∗
f0.75 f1k,∗ Bid
1 2
0 3 0 0 3 0 0
1 1 1
0.25 0 2 0 2 0 4
1
0.5 0 0 1 0 0 2
0.75 2
11
1
11
3
11
5
11 0 G0.75 = (12 − 24s)(20s − 15)−1 on [ 12 , 27
44 ]
1 102
353
34
353
68
353
60
353
89
353 G1 = (162 − 264s)(89s − 89)−1 on [ 27
44 ,
251
353 ]
4. Properties of the worst-case equilibrium

In this section, we derive several properties of the worst-case equilibrium. We argue why
these properties are appealing and reinforce the worst-case equilibrium as a solution
concept. A solution concept is most useful if it selects a unique equilibrium. Thus, we
start by showing that the worst-case equilibrium is unique in the following sense.
Proposition 2. For any mental model ( Gϕ , β) such that for all 1 ≤ k ≤ m,

U k ( Gϕ , β) = U k Gϕ∗ , β∗ , (3)
it holds that β(θk ) = β∗ (θk ) and ϕ(θk ) = ϕ∗ (θk ) for all θk > μ.
The worst-case equilibrium yields unique bidding strategies and unique beliefs for
all bidders with valuations above the mean. For bidders with valuations strictly below
the mean, there are several beliefs and strategies that induce the worst-case payoff of 0.
In particular, if such a bidder believes that the lowest possible valuation in her mental
model is strictly larger than her own valuation, many different bidding strategies can
constitute an equilibrium. One can get rid of this multiplicity by assuming that a bidder
needs to place a positive probability on her own valuation in her mental model. In this
case, in every worst-case equilibrium, a bidder with a valuation below the mean believes
that her competitors with the same valuation bid their valuation; otherwise, she would
not obtain the worst-case payoff of 0.
In what follows, we first compare the mental models that arise as part of the worst-
case equilibrium to all other potential mental models. We then restrict attention to men-
tal models with efficient Nash equilibria. We close the section by considering properties
of the worst-case equilibrium from the point of view of the seller.
4.1 Comparison to all potential mental models

If the bidder is uncertain whether she has constructed the correct mental model, she
may want to make sure that she will at least win the object whenever all her competi-
tors’ valuations are lower than hers. This is an appealing property. Ex post, a bidder may
be remorseful after losing to a bidder with a lower valuation, as she could have imitated
that bidder’s strategy ex ante and won the auction. Losing to bidders with a higher val-
uation is different, because the imitation of their strategies is not always feasible. The
following proposition establishes that if a bidder follows the bidding strategy from the
worst-case equilibrium, she will win against all competitors with lower valuations, even
if her competitors follow bidding strategies from other mental models.
Proposition 3. For any mental model ( Gϕ , β), it holds for all θk , θl ∈ with θk > θl , for
all b ∈ supp(β∗ (θk )), and for all b ∈ supp(β(θl )) that
b ≥ b .
That is, if each of the competitors bids according to the bidding strategy in some mental
model, a bidder who follows the strategy from the worst-case equilibrium always outbids
all bidders with lower valuations.
To establish the result, we demonstrate that the upper bound of the support of the
bidding strategy for each valuation is weakly higher in the worst-case equilibrium than
any other mental model. We arrive at the result because the worst-case equilibrium
is efficient and the upper bound of the bid support of a bidder with valuation θk−1 is
equal to the lower bound of the bid support of a bidder with valuation θk . Moreover,
this shows that the worst-case equilibrium is, in a sense, the unique mental model that
satisfies Proposition 3.
4.2 Comparison to mental models with efficient equilibria

In this section, we compare the worst-case equilibrium to other mental models with
efficient equilibria. Restricting our attention to efficient equilibria allows us to derive
sharper results about the worst-case equilibrium. The focus on efficient equilibria is
similar in spirit to the earlier restriction to symmetric Nash equilibria: to narrow down
the possibilities when building a mental model, a bidder may reasonably conjecture that
her competitors will choose from efficient equilibria.
We focus on the empirical bid distributions that arise from equilibrium bidding
strategies in the mental models and the true valuation distribution. As we show in
the following proposition, the worst-case equilibrium generates a bid distribution that
(first-order stochastically) dominates any bid distribution generated by any efficient
equilibrium bidding function and any true valuation distribution.
Proposition 4. For every mental model ( Gϕ , β) with an efficient equilibrium and every
f = (f (θ1 ), , f (θm )) ∈ Fμ , denote by
β
Bf (s) = f θ1 G1 (s) + · · · + f θm Gm (s) (4)
the empirical bid distribution generated if bidders bid according to β and the true valu-
ation distribution is f , where Gk is the bid distribution of a bidder with valuation θk for
β∗ β
all 1 ≤ k ≤ m. It holds that Bf first-order stochastically dominates Bf .
We show that the bid distribution of every valuation in the worst-case equilibrium
bidding strategy β∗ hazard-rate dominates the bid distribution in the bidding strategy
β of any other mental model. First-order stochastic dominance follows directly for the
respective empirical bid distributions evaluated at the same true valuation distribution.
There are mental models that lead to inefficient equilibria of the first-price auction.
If we allow for inefficient equilibria, there is no result similar to Proposition 4. In par-
ticular, there exist inefficient equilibria and true valuation distributions such that the
worst-case equilibrium bid distribution does not dominate the empirical bid distribu-
tion in those equilibria. However, for any inefficient equilibrium, there exists a true val-
uation distribution such that the worst-case equilibrium bid distribution dominates the
bid distribution in that equilibrium.
4.3 Properties from the seller’s point of view

Now taking the viewpoint of the seller, we compare the revenue from the worst-case
equilibrium with the revenue of a first-price auction in any other mental model, as well
as with the expected revenue of the second-price auction. Revenue is a function of the
true valuation distribution and the bidding strategies of the bidders. That is, for a true
valuation distribution f and bidding strategies β, define n independent random vari-
β
ables B1 , , Bn with Bi distributed according to Bf . Denote by B(1) the first-order statis-
tic of B1 , , Bn . The revenue in a first-price auction is given by E[B(1) ]. In a second-
price auction, we focus on the dominant-strategy equilibrium, in which bidders bid
their valuations. In this case, the revenue is the second-order statistic of the valuations
implied by the true valuation distribution f .
Proposition 5. Let ( Gϕ , β) be a mental model with an efficient equilibrium. The rev-

enue in the worst-case equilibrium of the first-price auction is higher than
(i) the revenue generated in ( Gϕ , β) in a first-price auction and
(ii) the efficient equilibrium of the second-price auction.
The revenue comparison is a direct consequence of Proposition 4: if the bid distri-

bution of the worst-case equilibrium dominates the bid distribution that arises from any
mental model with an efficient equilibrium, the worst-case equilibrium also generates
higher revenue than any other profile of mental models with an efficient equilibrium.
The comparison with the second-price auction works in a similar fashion. In a second-
price auction, it is a weakly dominant strategy to bid one’s valuation. Thus, the revenue
of the second-price auction is independent of the bidders’ beliefs but is determined by
the true distribution of valuations. However, if every bidder in the first-price auction has
the true distribution as her belief, the resulting mental model is revenue-equivalent to
the efficient equilibrium of the second-price auction. Again by way of Proposition 4, this
equilibrium must yield a lower revenue than the worst-case equilibrium.
5. Conclusion
We analyze first-price auctions with bidders whose information about their competitors
is consistent with all distributions having the same mean. To deal with such uncertainty,
we introduce the notion of a worst-case equilibrium, which is based on the idea that
bidders build mental models of the situation and are pessimistic. If all bidders reason in
the same way, they evaluate the situation with the same mental model.
Our results can be extended to the case in which the bidders’ information is con-
sistent with all distributions satisfying E(h(θ)) = μ as long as h is a strictly increasing
function. In particular, the results can be extended to the case where the bidders’ in-
formation is consistent with some moment of the distribution other than the mean. In
such a case, there is a threshold, depending on h and μ, such that bidders with valua-
tions below this threshold bid their valuation. The proofs for this case follow the same
steps as the proofs of our original results, from which they can easily be adapted.
The notion of a worst-case equilibrium can be adapted to any game of incomplete
information. We chose the first-price auction because it is arguably a leading example
of a game of incomplete information, appearing widely in the literature and textbooks.
An interesting question for further research is which games of incomplete information
admit a worst-case equilibrium.
Appendix
A.1 Proof of Proposition 1
The proof proceeds through four lemmas. Lemma 1 shows that the system of linear
equations defining the worst-case beliefs has a unique solution. Lemma 2 establishes
that the proposed strategies form a Nash equilibrium. Lemma 3 introduces a useful
technical tool for the proof that the proposed equilibrium is worst case. Lemma 4 estab-
lishes that the proposed equilibrium is worst case.
Lemma 1. For every θk > μ, the system of linear equations defined by (1) and

k

k

f k,∗ θl = 1 and f k,∗ θl θl = μ
l=1 l=1
has a unique solution. In the unique solution, all coordinates are strictly positive.
Proof. We show for every θk > μ that the matrix corresponding to the system of equa-
tions has rank k by applying Gaussian elimination and obtaining a row echelon form.
The conditions in (1) can also be summarized as

k,∗ 1 n−1 ∗ n−1 k ∗ n−1 ∗
f θ + · · · + f k,∗ θh θ k − bh − θ − bh+1 − f k,∗ θh+1 θk − bh+1 = 0
for all h ∈ {1, , k − 2}. In order to obtain an upper triangular matrix, we eliminate the
variables f k,∗ (θ2 ), , f k,∗ (θk ). We eliminate the variable f k,∗ (θk ) by multiplying the
equation

k

f k,∗ θl = 1
l=1
by −θk and adding it to

k

f k,∗ θl θl = μ.
l=1
Multiplying the resulting equation by −1 gives

k−1

f k,∗ θl θk − θl = θk − μ,
l=1
which eliminates the variable f k,∗ (θk ). Moreover, the coefficients (θk − θl ) are strictly
positive. Next, we use the transformed conditions given by

k,∗ 1 n−1 ∗ n−1 k ∗ n−1 ∗
f θ + · · · + f k,∗ θh θk − bh − θ − bh+1 − f k,∗ θh+1 θk − bh+1 = 0
for all h ∈ {1, , k − 2} to eliminate the variables f k,∗ (θ2 ), , f k,∗ (θk−1 ). We show by
induction that in every elimination step, all coefficients are strictly positive. In particu-
lar, this implies that none of the coefficients is equal to 0. We therefore obtain an upper
triangular matrix after applying Gaussian elimination. We start the induction by show-
ing that in the equation that is obtained after eliminating f k,∗ (θk−1 ), all coefficients are
strictly positive. The variable f k,∗ (θk−1 ) is eliminated by multiplying the condition given
by
k,∗ 1
n−1
∗ n−1 k ∗ n−1 ∗
f θ + · · · + f k,∗ θk−2 θk − bk−2 − θ − bk−1 − f k,∗ θk−1 θk − bk−1 = 0
by the factor
θk − θk−1

n−1 k ∗
θ − bk−1
and adding it to the equation

k−1

f k,∗ θl θk − θl = θk − μ.
l=1
This gives the equation

k ∗

∗

k−2

k−2
θ − θ k−1 n−1 k
θ − bk−2 −
n−1
θk − bk−1
f k,∗ θl θk − θl + f k,∗ θl = θk − μ,
n−1 k ∗
l=1 l=1 θ − bk−1
where all coefficients are strictly positive. Now turning our attention to the induction
step, we assume that the variables f k,∗ (θk−1 ), , f k,∗ (θh+1 ) have been eliminated and
that in the resulting equation

h

cl f k,∗ θl = c,
l=1
all coefficients c and cl for 1 ≤ l ≤ h are strictly positive. Now we have to eliminate the
variable f k,∗ (θh ) using the condition

k,∗ 1 n−1 ∗ n−1 k ∗ n−1 ∗
f θ + · · · + f k,∗ θh−1 θk − bh−1 − θ − bh − f k,∗ θh θk − bh = 0.
We multiply this equation by the factor

ch

n−1 ∗
θ k − bh
and add it to the equation

h

cl f k,∗ θl = c. (5)
l=1
This gives the equation

n−1 ∗ ∗

n−1 k
h−1
h−1
c
k,∗ l h
θ − bh−1 −
k θ − bh
f k,∗ l
θ cl + f θ = c, (6)
n−1 k ∗
l=1 l=1 θ − bh
in which all coefficients are strictly positive. We conclude that the system of k linear
equations with k variables given by

k

f k,∗ θl = 1
l=1

k

f k,∗ θl θl = μ
l=1
k−1 n−1 h n−1
k ∗ k ∗
k,∗ l k,∗ l
f θ θ − bk−1 = f θ θ − bh for all h ∈ {1, , k − 2}
l=1 l=1
can be rearranged into a system of linear equations such that the resulting matrix has
rank k and, therefore, this system of equations has a unique solution.
We use an inductive argument to show that all coordinates in the unique solution
are strictly positive. Since, in the linear equation obtained after performing Gaussian
elimination, all coefficients are strictly positive, it holds that f k,∗ (θ1 ) is strictly positive.
Assume it has been shown that f k,∗ (θ1 ), , f k,∗ (θh−1 ) are strictly positive. It follows
from (5) that

h−1

ch f k,∗ θh = c − cl f k,∗ θl .
l=1
Since we have already established that ch is strictly positive, we need to show that

h−1

c− cl f k,∗ θl > 0.
l=1
It follows from (6) that

h−1
h−1
c− cl f k,∗ θl = f k,∗ θl cl
l=1 l=1

n−1 ∗ ∗

n−1 k
h−1
ch θk − bh−1 − θ − bh h−1

+ f k,∗ θl − cl f k,∗ θl
n−1 k ∗
l=1 θ − bh l=1

n−1 k ∗ n−1 k ∗
h−1
ch θ − b h−1 − θ − bh
= f k,∗ l
θ > 0,
n−1 k ∗
l=1 θ − bh
where the last inequality follows from the induction hypothesis. We conclude that all
coordinates in the unique solution are strictly positive.
Lemma 2. The bidding strategy β∗ , as specified in Proposition 1, constitutes a Nash equi-

librium of the first-price auction given by Gϕ∗ .
Proof. We have to check, for every valuation θk ∈ , that there does not exist a bid
b∈/ supp(G∗k ) that induces a higher expected payoff for a bidder with valuation θk than
the equilibrium payoff. Fix a valuation θk . If θk ≤ μ, bidding above the equilibrium bid
θk is not feasible since it implies bidding above the own valuation. Bidding below θk
does not constitute a profitable deviation since in the belief of a bidder with valuation
θk , the lowest bid placed by her competitors with positive probability is θk . If θk > μ, we
will consider three different sets of possible bids outside the support of G∗k .
∗
First, we consider all bids above the upper bound of the support of G∗k . Let b > bk .
Since f k,∗ (θl ) = 0 for all l > k, it holds that
∗
U θk , ϕ∗ θk , b, β∗ = θk − b < θk − bk = U k Gϕ∗ , β∗ .
Thus, bids above the upper bound of the support of G∗k can be excluded as deviating
bids.
∗
Second, we consider all bids of the form bl for l ∈ {1, , k}. By construction of the
worst-case equilibrium, β∗ makes a bidder indifferent between all these bids and the
∗
bids in the support of G∗k , as can be seen in (1). Thus, none of the bids of the form bl
induces a higher expected payoff than the equilibrium payoff. Third, we can exclude
bids b with θl−1 = β∗ (θl−1 ) < b < β∗ (θl ) = θl for 0 < θl ≤ μ, since there is no probability
mass between β∗ (θl−1 ) and β∗ (θl ).
∗ ∗
Finally, we consider bids b ∈ (bl−1 , bl ) for l < k and θl > μ, i.e., bids that are in the
bidding interval of a bidder with a strictly lower valuation but are not endpoints. We will
∗ ∗
proceed in two steps. First, we show that the payoff from bidding b ∈ (bl−1 , bl ) has a
unique critical point. It then follows that the payoff from bidding b is either less than (or
∗ ∗ ∗
equal to) the payoff from bidding bl−1 for all b ∈ (bl−1 , bl ) or greater than (or equal to)
∗ ∗ ∗
the payoff from bidding bl−1 for all b ∈ (bl−1 , bl ). This is a direct consequence of the fact
∗ ∗
that the payoff from bidding bl−1 is the same as the payoff from bidding bl and the payoff
∗ ∗
function has a unique critical point. Second, we show for a particular b ∈ (bl−1 , bl ) that
∗ ∗
the payoff from bidding b is less than or equal to the payoff from bidding bl−1 or bl .
∗ ∗
We start by showing that the payoff from bidding some b ∈ (bl−1 , bl ) has a unique
∗ ∗
critical point. The payoff for a bidder with valuation θk from bidding b ∈ (bl−1 , bl ) is
n−1 k
f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl G∗l (b) θ −b . (7)
∗ ∗ ∗ ∗
The payoff as a function of b is continuous on [bl−1 , bl ] and differentiable on (bl−1 , bl ).
∗ ∗
Thus, it attains a maximum and minimum in [bl−1 , bl ]. Whereas, by construction of
∗ ∗
the equilibrium beliefs, the payoffs at bl−1 and at bl are the same, the derivative of the
payoff necessarily is 0 at each critical point. Denote by gl∗ the density of G∗l . We analyze
the solution of
n−2 k,∗ l ∗ k
(n − 1) f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl G∗l (b) f θ gl (b) θ − b
n−1
− f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl G∗l (b) = 0. (8)
This is equivalent to
n−1
k f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl G∗l (b)
θ −b− n−2 k,∗ l ∗ = 0,
(n − 1) f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl G∗l (b) f θ gl (b)
and, therefore, to

f k,∗ θ1 + · · · + f k,∗ θl−1 G∗l (b)
θk − b − − = 0. (9)
(n − 1)f k,∗ θl gl∗ (b) (n − 1)gl∗ (b)
Thus, the left-hand side of (8) has the same number of zeros as the left-hand side of
(9). We now show that the left-hand side of (9) has a unique zero. To do so, we take the
derivative of the left-hand side of (9) with respect to b and show that it does not change
signs. Using
l ∗ 1 1
f l,∗ θ1 + · · · + f l,∗ θl−1 θ − bl−1 n−1 − θl − b n−1
G∗l (b) = 1 (10)
f l,∗ θl θl − b n−1
and
∗ 1
f l,∗ θ1 + · · · + f l,∗ θl−1 θl − bl−1 n−1
gl∗ (b) = n , (11)
f l,∗ θl θl − b n−1 (n − 1)
we get the following expression for the left-hand side of (9):
n
k f k,∗ θ1 + · · · + f k,∗ θl−1 f l,∗ θl θl − b n−1 (n − 1)
θ −b− l,∗ 1
(n − 1)f k,∗ θl ∗ n−1
1
f θ + · · · + f l,∗ θl−1 θl − b l−1
l,∗ 1 l ∗ 1 1
f θ + · · · + f l,∗ θl−1 θ − bl−1 n−1 − θl − b n−1
− 1
f l,∗ θl θl − b n−1 (n − 1)
n
f l,∗ θl θl − b n−1 (n − 1)
· ∗ 1
f l,∗ θ1 + · · · + f l,∗ θl−1 θl − bl−1 n−1
l n
k θ − b n−1 f k,∗ θ1 + · · · + f k,∗ θl−1 f l,∗ θl
=θ −b−
∗ 1
θl − bl−1 n−1 f l,∗ θ1 + · · · + f l,∗ θl−1 f k,∗ θl
l n
l θ − b n−1
− θ −b + .
∗ 1
θl − bl−1 n−1
This gives a derivative of

l 1 l 1
n θ − b n−1 f k,∗ θ1 + · · · + f k,∗ θl−1 f l,∗ θl n θ − b n−1
n − 1 θl − b∗ n−1
1 − n−1 l ∗ 1
l−1 f l,∗ θ1 + · · · + f l,∗ θl−1 f k,∗ θl θ − bl−1 n−1
l 1
n θ − b n−1 f k,∗ θ1 + · · · + f k,∗ θl−1 f l,∗ θl − f l,∗ θ1 + · · · + f l,∗ θl−1 f k,∗ θl
= l ∗ n−11 .
n−1 θ −b f l,∗ θ1 + · · · + f l,∗ θl−1 f k,∗ θl
l−1
∗ ∗
Since θl − b > 0 for all b ∈ (bl−1 , bl ), the derivative does not change signs. Thus, the
∗ ∗
payoff from bidding b ∈ (bl−1 , bl ) has a unique critical point. It remains to show that
this critical point is a minimum.
∗ ∗
Choose b ∈ (bl−1 , bl ) such that

n−1 ∗ n−1 n−1 n−1 ∗
θk − bl−1 − θk − b = θk − b − θk − bl . (12)
Suppose for the sake of contradiction that the critical point is a maximum. In this case,
it holds that
k,∗ 1 n−1 k
f θ + · · · + f k,∗ θl−1 + f k,∗ θl Gl (b) θ −b
n−1 k ∗
≥ f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl θ − bl (13)
and
n−1 k
f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl Gl (b) θ −b
n−1 k ∗
≥ f k,∗ θ1 + · · · + f k,∗ θl−1 θ − bl−1 . (14)
Both inequalities (13) and (14) hold with equality if k = l.

Rearranging (13) gives

n−1
f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl Gl (b) θk − b

n−1 ∗
≥ f k,∗ θ1 + · · · + f k,∗ θl−1 + f k,∗ θl θ k − bl

k,∗ 1 n−1 n−1 ∗
⇔ f θ + · · · + f k,∗ θl−1 θk − b − θk − bl

l n−1
k,∗ ∗ n−1
≥f θ θk − bl − f k,∗ θl Gl (b) θk − b

⇔ f k,∗ θ1 + · · · + f k,∗ θl−1

n−1 ∗ n−1
f k,∗ θl θk − bl − f k,∗ θl Gl (b) θk − b
≥ . (15)
n−1 n−1 ∗
θk − b − θk − bl
Rearranging (14) in the same way gives

n−1
k,∗
1
k,∗
l−1
f k,∗ θl Gl (b) θk − b
⇔f θ + ··· + f θ ≤ . (16)
n−1 k ∗ n−1
θ − bl−1 − θk − b
Again, both inequalities (15) and (16) hold with equality if k = l.

If we show that

n−1 n−1 ∗ n−1
f k,∗ θl Gl (b) θk − b f k,∗ θl θk − bl − f k,∗ θl Gl (b) θk − b
< (17)
n−1 k ∗ n−1 n−1 n−1 ∗
θ − bl−1 − θk − b θk − b − θ k − bl
for k > l, we get a contradiction to inequalities (13) and (14).

Using (12) and rearranging, we get

n−1 n−1 ∗
2Gl (b) θk − b < θ k − bl . (18)
Using (18) with equality for k = l, rearranging for Gl (b), substituting in (17) for k > l,
and rearranging yields
k k ∗
θ −b θ − bl
l < l ∗ ,
θ −b θ − bl
which is equivalent to
∗ ∗
θ k bl − b > θ l bl − b .
The last inequality is obviously true as θk > θl . Thus, the assumption that the critical
point of the payoff function is a maximum leads to a contradiction. Summing up, for a
∗ ∗
bidder with valuation θk , it is not a profitable deviation to choose b ∈ (bl−1 , bl ).
To establish that the proposed equilibrium is worst case, we use the following
lemma.
Lemma 3. Let f ∈ Fμ and let f˜ be a function f˜ : → [0, 1] such that m ˜ l

l=1 f (θ ) = 1. Let
(δ1 , , δm ) be a vector of real numbers such that f˜(θl ) = f (θl ) + δl for all 1 ≤ l ≤ m and
it holds for at least one 1 ≤ l ≤ m that δl = 0. Assume that for all 1 ≤ l ≤ m, it holds that
l m l ˜ l ˜ / Fμ .
h=1 δh ≤ 0. Then l=1 θ f (θ ) > μ, i.e., f ∈
Proof. Let f , f˜, and (δ1 , , δm ) be as in Lemma 3. It holds that

l m

m

m
m−1

θl f˜ θl − f θl = θl δl = θl − θl+1 δh + θm δl > 0.
l=1 l=1 l=1 h=1 l=1
m ˜ l
Since l=1 f (θ ) = 1 and f˜(θl ) = f (θl ) + δl , it follows that m
=1 δl = 0. Therefore,
m l ˜ l
l=1 θ f (θ ) > μ.
Intuitively, the statement is based on the fact that if the conditions of the lemma are
fulfilled, the distribution f˜ first-order stochastically dominates the distribution f and,
therefore, they cannot have the same mean.
Lemma 4. For every mental model ( Gϕ , β) of the first-price auction, it holds for all k ∈
{1, , m} that

U k ( Gϕ , β) ≥ U k Gϕ∗ , β∗ (19)
∗
and bk ≤ bk .
Proof. As preparation for the proof of this lemma, we will show the following claim.
We will then use the claim in an inductive argument.
Claim 1. Let ( Gϕ , β) be a mental model such that

U k ( Gϕ , β) ≤ U k Gϕ∗ , β∗ (a)
for some k with θk > μ and

∗
bl ≤ bl (b)
for all 1 ≤ l ≤ k − 1, where bl = sup{b | b ∈ supp(Gl )}. For 1 ≤ l ≤ m, define δl by f k (θl ) =
f k,∗ (θl ) + δl . Then δl = 0 for all 1 ≤ l ≤ m. That is, f k (θl ) = f k,∗ (θl ) for all 1 ≤ l ≤ m.
Proof. Let ( Gϕ , β) be a mental model that fulfills the conditions (a) and (b). Recall that
the infimum of the support of β(θk ) is denoted by bk . The expected payoff of a bidder
with valuation θk in the mental model ( Gϕ , β) is the expected payoff of bidding bk , which
is given by
m n−1
k
k l
f θ Gl (bk ) θ − bk (20)
l=1
(some of the expressions of the form Gl (bk ) may be zero). Note that there may be an
atom in the bid distribution Gl at bk . If f k (θl ) = 0, this does not matter for the expression
in (20). If f k (θl ) > 0, a bidder with valuation θk would not tie with positive probability
at bk . In this case, the probability that a bidder with valuation θk wins against a bidder
with valuation θl is given by the limit b → bk from above and is equal to Gl (bk ).
Recall that in the worst-case equilibrium ( Gϕ∗ , β∗ ), the lower endpoint of the bidding
interval of a bidder with valuation θk coincides with the upper endpoint of the bidding
∗
interval of a bidder with valuation θk−1 , denoted by bk−1 . Thus, the expected payoff of a
bidder with valuation θk in the worst-case equilibrium ( Gϕ∗ , β∗ ) is given by
k−1 n−1
l ∗
k,∗
f θ θk − bk−1 .
l=1
Let (δ1 , , δm ) be such that f k (θl ) = f k,∗ (θl ) + δl for 1 ≤ l ≤ m.

Since β in ( Gϕ , β) is an equilibrium bidding function, it must hold that the expected
payoff in (20) is at least as high as the expected payoff of deviating to a bid outside the
support of the bidder’s bidding strategy. In particular, it must hold that deviating to
bidding 0 does not induce a higher payoff. Thus,
n−1

m
n−1 k
f k θl Gl (bk ) θ k − bk ≥ f k θ 1 θ .
l=1
It follows that
m n−1
n−1 n−1 k
f k,∗ θ1 + δ1 θk = f k θ1 θk ≤ f k θl Gl (bk ) θ − bk
l=1
k−1 n−1
∗ n−1 k
≤ f k,∗ θl θk − bk−1 = f k,∗ θ1 θ ,
l=1
where the second inequality follows from the fact that U k ( Gϕ , β) ≤ U k ( Gϕ∗ , β∗ ) and the
second equality follows from (1). We conclude that δ1 ≤ 0. We now extend the argument
and show for all 1 ≤ h ≤ k − 1 that hl=1 δl ≤ 0. Let 1 ≤ h ≤ k − 1 and let
bt := max bl .
l≤h
Since we assume an efficient tie-breaking rule and k > h, it follows that a bidder with
valuation θk wins against all valuations less than or equal to θh if she deviates to bt . It
follows that the expected utility from deviating equals
h n−1

m

f θl +
k
f k θl Ĝl (bt ) θ k − bt ,
l=1 l=h+1
where Ĝl (bt ) denotes the probability that a bidder with valuation θk wins against a bid-
der with valuation θl at bt (which may be different from Gl (bt ) if there is an atom in the
bid distribution for valuation θl and l > h).
Since deviating to bid bt cannot yield a higher payoff for a bidder with valuation θk ,
it holds that
m n−1 h n−1
k
m
k
k l k l k l
f θ Gl (bk ) θ − bk ≥ f θ + f θ Ĝl (bt ) θ − bt , (21)
l=1 l=1 l=h+1
from which it follows that

n−1

h

m

f k,∗ θl + δl + f k θl Ĝl (bt ) θ k − bt
l=1 l=h+1
h n−1

m

= f θl +
k
f k θl Ĝl (bt ) θk − bt
l=1 l=h+1
n−1 k−1 n−1

m
k n−1 k ∗
≤ f θl Gl (bk )
k
θ − bk ≤ f k,∗ l
θ θ − bk−1
l=1 l=1
n−1

t
∗
= f k,∗ θl θk − bt ,
l=1
where the first inequality follows from (21) and the second inequality follows from con-
∗
dition (a). Since h ≤ k − 1 and 1 ≤ t ≤ h ≤ k − 1, it follows by assumption that bt ≤ bt .
Hence, it must hold that

h

m

t

f k,∗ θl + δl + f k θl Ĝl (bt ) ≤ f k,∗ θl
l=1 l=h+1 l=1

h
h

m

⇔ δl + f k,∗ θl + f k θl Ĝl (bt ) ≤ 0.
l=1 l=t+1 l=h+1
Since

h

m

f k,∗ θl + f k θl Ĝl (bk ) ≥ 0,
l=t+1 l=h+1
it follows that

h
δl ≤ 0.
l=1
We have shown for all 1 ≤ h ≤ k − 1 that hl=1 δl ≤ 0. Let h ≥ k. By the construction

of (f k,∗ (θ1 ), , f k,∗ (θm )), it holds that f k,∗ (θl ) = 0 for all l > k. Hence, δl ≥ 0 for all
k + 1 ≤ l ≤ m and, therefore, m l=h+1 δl ≥ 0. Since
m
l=1 δl = 0, it follows that
h
l=1 δl ≤ 0.
h
Thus, there does not exist any 1 ≤ h ≤ m such that l=1 δl > 0. It follows from Lemma 3
that δl = 0 for all 1 ≤ l ≤ m, as otherwise the belief (f k (θ1 ), , f k (θm )) would violate
the mean constraint kl=1 f k (θl )θl = μ.
Having proved the claim, we proceed with the proof of the statement in (19). Let
( Gϕ , β) be a mental model. We will show the statement by proving the following two
statements simultaneously by induction:
(i) For every k ∈ {1, , m}, it holds that

U k Gϕ∗ , β∗ ≤ U k ( Gϕ , β).
∗
(ii) For every k ∈ {1, , m}, it holds that bk ≤ bk .
Recall that we only consider equilibria in which bidders never bid above their own
valuation and, hence, a bidder with valuation zero bids 0 (Definition 1). Thus, both
statements are trivially true for k = 1, since θ1 = 0 and, therefore,

0 = U 1 Gϕ∗ , β∗ ≤ U 1 ( Gϕ , β) = 0
∗
and 0 = b1 ≤ b1 = 0. Assume that both statements have been shown for k − 1; we have to
prove both statements for k. We begin with statement (i). If θk ≤ μ, the statement is triv-
ially true, because then a bidder with valuation θk obtains the lowest possible payoff of
0 in the worst-case equilibrium. Statement (ii) is also true, since a bidder with valuation
θk with θk ≤ μ bids her valuation in the worst-case equilibrium, and this is the highest
possible undominated bid. Assume that θk > μ and statement (i) is true for all l < k.
Let (δ1 , , δm ) be such that f k (θl ) = f k,∗ (θl ) + δl for 1 ≤ l ≤ m. In this case, the con-
ditions (a) and (b) of Claim 1 are fulfilled, and it holds that δl = 0 for all 1 ≤ l ≤ m. Thus,
given the belief f k , a bidder with valuation θk believes that she has the highest valua-
∗
tion. By the induction hypothesis, bl ≤ bk−1 for l < k. Thus, if a bidder with valuation θk
∗
bids bk−1 instead of bk , she wins against all lower valuations. Therefore, she obtains at
least a payoff of
k,∗ 1 ∗
f θ + · · · + f k,∗ θk−1 θk − bk−1 . (22)
Her payoff from bidding her equilibrium bid bk must be at least as high. By the construc-
∗
tion of the worst-case equilibrium, it holds that bk−1 = b∗k . Thus, the expression (22) is
equal to the payoff of a bidder with valuation θk in the worst-case equilibrium. Thus,
statement (i) is true for k. Hence, we have shown the induction step for statement (i).
It is left to show the induction step for statement (ii). It holds that
n−1 n−1

m

m

f θl Gl (bk )
k k
θ − bk = f θl Ĝl (bk )
k
θ k − bk
l=1 l=1
and
k−1 n−1 k n−1
l ∗ l ∗ ∗
k,∗ k k,∗
f θ θ − bk−1 = f θ θk − bk = θk − bk ,
l=1 l=1
where, as above, Ĝl (bk ) denotes the probability with which a bidder with valuation θk
wins against a bidder with valuation θl at bk .
Since we have shown that
m n−1 k−1 n−1
k k ∗
k l k,∗ l
f θ Gl (bk ) θ − bk ≥ f θ θ − bk−1 ,
l=1 l=1
it follows that
m n−1
∗
k
θ − bk ≤ f θl Ĝl (bk )
k
θk − bk .
l=1
Since

m

f k θl Ĝl (bk ) ≤ 1,
l=1
∗
we conclude that bk ≤ bk .

We prove that whenever

U k Gϕ∗ , β∗ = U k ( Gϕ , β) (23)
for all k ∈ {1, , m}, it follows that Gk (b) = G∗k (b) and f k = f k,∗ for all k ∈ {1, , m}
with θk > μ.
Let ( Gϕ , β) be a mental model such that

U k Gϕ∗ , β∗ = U k ( Gϕ , β)
for all k ∈ {1, , m}. First, we show that f k = f k,∗ for all k ∈ {1, , m} with θk > μ. Let
∗
k ∈ {1, , m} with θk > μ. We have shown that bl ≤ bl for all 1 ≤ l ≤ m. Thus, conditions
(a) and (b) of Claim 1 are fulfilled and it holds that f k = f k,∗ .
∗
Second, we show for all k ∈ {1, , m} with θk > μ that bk ≥ bk in order to conclude
∗ ∗
that bk = bk . Assume that bk < bk for some k ∈ {1, , m} with θk > μ. Since, in the
∗
equilibrium (Gϕ∗ , β∗ ), a bidder with valuation θk wins with probability 1 if bidding bk
and a bidder with valuation θk obtains the same utility as in the mental model ( Gϕ , β),
it must hold that a bidder with valuation θk does not win with probability 1 if bidding
bk in the mental model ( Gϕ , β). Since there is no probability weight on valuations above
k, there must exist a valuation θt with t < k and bk < bt . If there would be an atom at
bk in the bid distribution of a type θl ≤ θk , it would be possible that bk = bt . However,
since f k = f k,∗ , in the belief of type θk , there is positive probability weight on every type
θl ≤ θk . Thus a bidder with valuation θk would slightly overbid this atom.
Let U(θk , ϕ(θk ), b, β) denote the expected utility of a bidder with valuation θk in the
equilibrium ( Gϕ , β) if she bids b (where b may be a deviating bid). We use the notation
U(θk , ϕ∗ (θk ), b, β∗ ) analogously for the equilibrium ( Gϕ∗ , β∗ ).
By assumption,
∗
U θk , ϕ θk , bk , β = U k ( Gϕ , β) = U k Gϕ∗ , β∗ = U θk , ϕ∗ θk , bk , β∗ .
Note that if bk is the supremum but not the maximum of the support of the bid distribu-
tion for valuation θk , the expected utility at bk may not be the equilibrium utility if there
is an atom at bk in the bid distribution of some other bidder with valuation θl . If l > k,
an atom does not influence the expected utility for valuations θk , since a bidder with
valuation θk assigns zero probability to type θl . The case l ≤ k has been excluded above.
Bidding above bk is not a dominated bid since bk < bt ≤ θt < θk . By the construction of
the worst-case equilibrium, it holds that
∗ ∗
U θk , ϕ∗ θk , bk , β∗ = U θk , ϕ∗ θk , bt , β∗ .
Combined, this gives

∗
U θk , ϕ θk , bk , β = U θk , ϕ∗ θk , bt , β∗ .
∗
Since bl ≤ bl for all 1 ≤ l ≤ m, a bidder with valuation θk would win not only against all
∗
bidders with lower valuations when deviating to bt , but also against all other bidders
∗ ∗
with valuation θk , since bk < bt ≤ bt . This implies that bidding bt in the mental model
( Gϕ , β) induces a higher winning probability than in the worst-case equilibrium. Since
f k = f k,∗ , it also induces a higher expected payoff. It follows that
∗ ∗
U θk , ϕ θk , bk , β ≥ U θk , ϕ θk , bt , β > U θk , ϕ∗ θk , bt , β∗ ,
which leads to a contradiction. The first inequality holds since the expected payoff in
∗
equilibrium cannot be lower than the expected payoff from deviating to bt . We conclude
∗
that bk = bk .
∗
Third, we show that for all k ∈ {1, , m} with θk > μ, it holds that Gk (bk−1 ) = 0; i.e.,
in the mental model ( Gϕ , β), in the bid distribution of a bidder with valuation θk there
∗ ∗
does not exist a positive mass of bids that lies below bk−1 . Assume that Gk (bk−1 ) > 0 for
∗
some k ∈ {1, , m} with θk > μ. If a bidder with valuation θk bids bk−1 in the equilib-
rium ( Gϕ , β), she wins against all bidders with valuations lower than θk with probabil-
ity 1. Thus, her winning probability is given by
∗
f k θ1 + · · · + f k θk−1 + f k θk Gk bk−1 ,
and since we have established that f k,∗ = f k for all k with θk > μ, this is equal to
∗
f k,∗ θ1 + · · · + f k,∗ θk−1 + f k,∗ θk Gk bk−1 .
We conclude that U k ( Gϕ , β) > U k ( Gϕ∗ , β∗ ), which leads to a contradiction.

Fourth, we show that bk ≤ b∗k for all k ∈ {1, , m} with θk > μ. Assume that bk > b∗k
∗
for some k ∈ {1, , m} with θk > μ. It follows that bk > b∗k = bk−1 = bk−1 . Valuations
∗ ∗
below θk cannot bid above bk−1 , and valuations above θk bid between bk−1 and bk with
∗
probability 0, since we have established that Gl (bk ) = Gl (bk ) = 0 for l > k. Thus, there is
∗
a gap in the bid distribution between bk−1 and bk given the strategy β. This cannot occur
in equilibrium since for a sufficiently small , a bidder with valuation θk could deviate
∗
from bk + to a bid b with bk−1 < b < bk and strictly increase her expected payoff since
for any > 0, one can find > 0 such that the winning probability decreases by less than
when deviating. It follows that the assumption b > b∗ leads to a contradiction.
k k
Finally we show that Gk (b) = G∗k (b) for all k ∈ {1, , m} with θk > μ. Observe that
Gk cannot have an atom at bk , as otherwise a bidder with valuation θk would gain by
∗
slightly overbidding bk . Let b ∈ [bk , bk ]. We have established that bk = bk and f k = f k,∗
for all θk ∈ with θk > μ, from which it follows that
k,∗ 1
f θ + · · · + f k,∗ θk−1 + f k,∗ θk Gk (b) n−1 θk − b
n−1
= f k,∗ θ1 + · · · + f k,∗ θk θ k − bk
n−1
∗
= f k,∗ θ1 + · · · + f k,∗ θk θ k − bk

= f k,∗ θ1 + · · · + f k,∗ θk−1 + f k,∗ θk G∗k (b) n−1 θk − b .
Therefore, Gk (b) and G∗k (b) are solutions of the same linear equation and, hence, are
equal.

Proof. Let ( Gϕ , β) be a mental model. Fix a valuation θk ∈ . Let b ∈ supp(β∗ (θk )) and
∗
let b ∈ supp(β(θl )) for 1 ≤ l < k. By definition, b ≥ bk−1 and b ≤ bl . As shown in the
∗
proof of Lemma 4, it holds that bl ≤ bk−1 , from which it follows that
∗
b ≤ bl ≤ bk−1 ≤ b.

Proof. Let ( Gϕ , β) be a mental model and let f = (f (θ1 ), , f (θm )) ∈ Fμ . We have to
show that
β∗
Bf (s) = f θ1 G∗1 (s) + · · · + f θm G∗m (s) (24)
first-order stochastically dominates

β
Bf (s) = f θ1 G1 (s) + · · · + f θm Gm (s). (25)
∗ ∗
Let s be an arbitrary bid. By Lemma 4, bm ≥ bk for all k ∈ {1, , m}. Therefore, if s > bm ,
β β∗
then Bf (s) = Bf (s) = 1. If s ∈ [θl−1 , θl ) for θl ≤ μ, then
β∗
Bf (s) = f θ1 + · · · + f θl−1 .
Since θl−1 is the highest possible bid for a bidder with valuation θl−1 , it holds that
Gk (θl ) = 1 for all k ≤ l − 1. It follows that
β β
Bf (s) ≥ Bf θl−1 ≥ f θ1 G1 θl−1 + · · · + f θl−1 Gj−1 θl−1
β∗
= f θ1 + · · · + f θl−1 = Bf (s).
∗
Thus, we can assume that s ∈ [b∗k , bk ] for some k ∈ {1, , m} with θk > μ. Since ( Gϕ∗ , β∗ )
is a mental model with an efficient equilibrium, it follows that
β∗
Bf (s) = f θ1 + · · · + f θk−1 + f θk G∗k (s).
∗
Since bm is the highest possible bid in any mental model, three cases are relevant:
(i) s ∈ ( b̄m , b̄∗m ], (ii) s ∈ ( b̄h , bh+1 ), and (iii) s ∈ [bh−1 , bh ] for some 1 ≤ h ≤ m. Case (i) is
∗
immediate. For case (ii), observe that by Lemma 2, bl ≤ bl for all 1 ≤ l ≤ m and, therefore,
β β∗
h ≥ k. We can immediately conclude that Bf (s) ≥ Bf (s). Thus, consider case (iii). Since
β is an efficient Nash equilibrium, it holds that bh ≤ bh−1 and
β
Bf (s) = f θ1 + · · · + f θh−1 + f θh Gh (s).
∗
As shown in Lemma 4, bl ≤ bl for all 1 ≤ l ≤ m. It follows that h ≥ k. If h > k, we can im-
β∗
mediately conclude that Bf (s) ≥ Bf (s). If h = k, we have to show that Gk (s) ≥ G∗k (s). In
β
order to do so, we will show that the bid distribution G∗k dominates the bid distribution
Gk in terms of the reverse hazard rate g/G.
It holds that
k ∗ 1 1
f k,∗ θ1 + · · · + f k,∗ θk−1 θ − bk−1 n−1 − θk − s n−1
G∗k (s) = 1 ,
from which it follows that the reverse hazard rate of this bid distribution is given by
1
k,∗ 1 k ∗ 1 1
f θ + · · · + f k,∗ θk−1 θ − bk−1 n−1 − θk − s n−1
k,∗ 1 ∗ 1
f θ + · · · + f k,∗ θk−1 θk − bk−1 n−1
· n
f k,∗ θk θk − s n−1 (n − 1)
∗ 1
θk − bk−1 n−1
= 1 .
∗ 1
θk − bk−1 n−1
− θk − s n−1 θk − s (n − 1)
For Gk consider two cases. First, f k (θk ) = 0. In this case, it follows that b̄k = b̄k−1 and we
β β∗
can immediately conclude that Bf (s) ≥ Bf (s). Second, f k (θk ) > 0. In this case, Gk is
atomless. If Gk were to have an atom and f k (θk ) > 0, a bidder with valuation θk would
slightly outbid the atom instead of mixing her bids. Thus, we obtain that the reverse
hazard rate of the bid distribution Gk is given by
1
θk − bk−1 n−1
1 1 .
θk − bk−1 n−1
− θk − s n−1 θk − s (n − 1)
We therefore have to show that

∗ 1 k 1
θk − bk−1 n−1 θ − bk−1 n−1
k ∗ 1 1 ≥ k 1 1
θ − bk−1 n−1 − θk − s n−1 θ − bk−1 n−1 − θk − s n−1
∗ 1 1 1
⇔ θk − bk−1 n−1
θk − bk−1 n−1
− θk − s n−1
1 ∗ 1 1
≥ θk − bk−1 n−1 θk − bk−1 n−1 − θk − s n−1 ,
which reduces to
∗ 1 1 1 1
θk − bk−1 n−1
θk − s n−1
≤ θk − bk−1 n−1 θk − s n−1 . (26)
∗
Since bk−1 ≥ bk−1 by the first part of the proof, the last statement is obviously true.

Proof. Both properties are a direct consequence of Proposition 4.
β∗
(i) Following the same steps as in the proof of Proposition 4, we can show that Bf
β
first-order stochastically dominates Bf . It follows directly that the expectation of
β∗
the first-order statistic of n random variables drawn from Bf is larger than the
β
expectation of the first-order statistic of n variables drawn from Bf .
(ii) Riley (1989) provides a revenue-equivalence principle that holds for first- and
second-price auctions with discrete valuations. His argument does not rely on
any tie-breaking rule and is based on two properties of any equilibrium of the
first-price auction: monotonicity (that is, that each bidder’s bids are weakly in-
creasing in her valuation) and continuity (that is, that there are no gaps or atoms
in the bid distribution). Both properties hold with an efficient tie-breaking rule.
It follows that for any true valuation distribution, the revenue in the second-price
auction is equal to the revenue in the first-price auction where the true valuation
distribution is a common prior. Since a first-price auction with a common prior
together with the unique efficient symmetric equilibrium constitutes a profile of
mental models, the statement follows from part (i).
References
Allouah, Amine and Omar Besbes (2020), “Prior-independent optimal auctions.” Man-
agement Science, 66, 4417–4432. [64]
Auster, Sarah and Christian Kellner (2022), “Robust bidding and revenue in descending
price auctions.” Journal of Economic Theory, 199, 105072. [63]
Azar, Pablo, Jing Chen, and Silvio Micali (2012), “Crowdsourced Bayesian auctions.” In
Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, 236–248.
[64]
Azar, Pablo Daniel and Silvio Micali (2013), “Parametric digital auctions.” In Proceedings
of the 4th Conference on Innovations in Theoretical Computer Science, 231–232. [65]
Bergemann, Dirk, Benjamin Brooks, and Stephen Morris (2017), “First-price auctions
with general information structures: Implications for bidding and revenue.” Economet-
rica, 85, 107–143. [64]
Bergemann, Dirk, Benjamin Brooks, and Stephen Morris (2019), “Revenue guarantee
equivalence.” American Economic Review, 109, 1911–1929. [64]
Bergemann, Dirk and Karl Schlag (2011), “Robust monopoly pricing.” Journal of Eco-
nomic Theory, 146, 2527–2543. [64]
Bergemann, Dirk and Karl H. Schlag (2008), “Pricing without priors.” Journal of the Eu-
ropean Economic Association, 6, 560–569. [64]
Bewley, Truman F. (2002), “Knightian decision theory. Part I.” Decisions in Economics
and Finance, 25, 79–110. [63]
Bodoh-Creed, Aaron L. (2012), “Ambiguous beliefs and mechanism design.” Games and
Economic Behavior, 75, 518–537. [63]
Bose, Subir, Emre Ozdenoren, and Andreas Pape (2006), “Optimal auctions with ambi-
guity.” Theoretical Economics, 1, 411–438. [63]
Carrasco, Vinicius, Vitor Farinha Luz, Nenad Kos, Matthias Messner, Paulo Monteiro,
and Humberto Moreira (2018), “Optimal selling mechanisms under moment condi-
tions.” Journal of Economic Theory, 177, 245–279. [64, 65]
Carroll, Gabriel (2015), “Robustness and linear contracts.” American Economic Review,
105, 536–563. [64]
Chiesa, Alessandro, Silvio Micali, and Zeyuan Allen Zhu (2015), “Knightian analysis of
the Vickrey mechanism.” Econometrica, 83, 1727–1754. [63]
Dekel, Eddie, Drew Fudenberg, and David K. Levine (2004), “Learning to play Bayesian
games.” Games and Economic Behavior, 46, 282–303. [63]
Di Tillio, Alfredo, Nenad Kos, and Matthias Messner (2016), “The design of ambiguous
mechanisms.” The Review of Economic Studies, 84, 237–276. [63]
Esponda, Ignacio (2008), “Information feedback in first price auctions.” The RAND Jour-
nal of Economics, 39, 491–508. [63]
Fudenberg, Drew and David K. Levine (1993), “Self-confirming equilibrium.” Economet-
rica, 61, 523–545. [63]
Gagnon-Bartsch, Tristan, Marco Pagnozzi, and Antonio Rosato (2021), “Projection of
private values in auctions.” American Economic Review, 111, 3256–3298. [64]
Kasberger, Bernhard (2020), “An equilibrium model of the first-price auction with strate-
gic uncertainty: Theory and empirics.” Report. [64]
Kasberger, Bernhard and Karl H. Schlag (2023), “Robust bidding in first-price auctions:
How to bid without knowing what others are doing.” Management Science, https://doi.
org/10.1287/mnsc.2023.4899. [64]
Koçyiğit, Çağıl, Garud Iyengar, Daniel Kuhn, and Wolfram Wiesemann (2020), “Distribu-
tionally robust mechanism design.” Management Science, 66, 159–189. [63, 64]
Lang, Matthias and Achim Wambach (2013), “The fog of fraud—Mitigating fraud by
strategic ambiguity.” Games and Economic Behavior, 81, 255–275. [63]
Lo, Kin Chung (1998), “Sealed-bid auctions with uncertainty averse bidders.” Economic
Theory, 12, 1–20. [63]
Mass, Helene (2023), “First-price auctions under uncertainty - maximin selection from
rationalizable strategies.” Working paper. [64]
Pınar, Mustafa Ç. and Can Kızılkale (2017), “Robust screening under ambiguity.” Mathe-
matical Programming, 163, 273–299. [64, 65]
Riley, John G. (1989), “Expected revenue from open and sealed bid auctions.” Journal of
Economic Perspectives, 3, 41–50. [91]
Wolitzky, Alexander (2016), “Mechanism design with maxmin agents: Theory and an
application to bilateral trade.” Theoretical Economics, 11, 971–1004. [65]
Co-editor Ran Spiegler handled this manuscript.
Manuscript received 7 October, 2020; final version accepted 30 January, 2023; available online 23
February, 2023.

4555 38402 1 PB

Uploaded by

Copyright:

Available Formats

4555 38402 1 PB

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

4555 38402 1 PB

Uploaded by

Copyright:

Available Formats

Theoretical Economics 19 (2024), 61–93 1555-7561/20240061

Worst-case equilibria in first-price auctions

Vitali Gretschko: [email protected]

1.1 Related literature

i ∈ {1, , n} = I privately observes her valuation θi ∈ = {0 = θ1 , θ2 , , θm−1 , 1 =

2.2 Worst-case mental models

The bidder can now choose a bid b to maximize U(θk , f k , b, β).

Definition 1. A bidding strategy β constitutes a symmetric Nash equilibrium in un-

We call β an efficient equilibrium if for all θk , θl ∈ with θk > θl , for all b ∈

Definition 2. A pair ( Gϕ , β) is a mental model if β is a Nash equilibrium of Gϕ . If β is an

Definition 3. A mental model ( Gϕw , βw )k is a worst-case mental model for a bidder

2.3 Worst-case equilibria

Definition 4. A mental model ( Gϕ∗ , β∗ ) is a symmetric worst-case equilibrium if it is

3. Worst-case equilibrium of the first-price auction

The unique solution is given by

Observation 1. Whenever θk ≤ μ, the worst-case bidding strategy is β∗ (θk ) = θk . Let

This expression is minimized given μ if the probability weight on θk is maximized in her

Observation 2. The conditions

and the equations

Rearranging yields our last observation.

We are now in a position to state our main proposition.

Proposition 1. The mental model ( Gϕ∗ , β∗ ) constitutes a symmetric worst-case equilib-

• β∗ (θk ) = G∗k whenever θk > μ.

Valuation f0k,∗ k,∗

4. Properties of the worst-case equilibrium

Proposition 2. For any mental model ( Gϕ , β) such that for all 1 ≤ k ≤ m,

4.1 Comparison to all potential mental models

4.2 Comparison to mental models with efficient equilibria

4.3 Properties from the seller’s point of view

Proposition 5. Let ( Gϕ , β) be a mental model with an efficient equilibrium. The rev-

(i) the revenue generated in ( Gϕ , β) in a first-price auction and

(ii) the efficient equilibrium of the second-price auction.

The revenue comparison is a direct consequence of Proposition 4: if the bid distri-

by −θk and adding it to

Multiplying the resulting equation by −1 gives

This gives the equation

We multiply this equation by the factor

and add it to the equation

This gives the equation

It follows from (6) that

Lemma 2. The bidding strategy β∗ , as specified in Proposition 1, constitutes a Nash equi-

This gives a derivative of

Both inequalities (13) and (14) hold with equality if k = l.

Rearranging (14) in the same way gives

Again, both inequalities (15) and (16) hold with equality if k = l.

for k > l, we get a contradiction to inequalities (13) and (14).

Lemma 3. Let f ∈ Fμ and let f˜ be a function f˜ : → [0, 1] such that m ˜ l

Proof. Let f , f˜, and (δ1 , , δm ) be as in Lemma 3. It holds that

Claim 1. Let ( Gϕ , β) be a mental model such that

for some k with θk > μ and

Let (δ1 , , δm ) be such that f k (θl ) = f k,∗ (θl ) + δl for 1 ≤ l ≤ m.

from which it follows that

We have shown for all 1 ≤ h ≤ k − 1 that hl=1 δl ≤ 0. Let h ≥ k. By the construction

(i) For every k ∈ {1, , m}, it holds that

A.2 Proof of Proposition 2

Combined, this gives

We conclude that U k ( Gϕ , β) > U k ( Gϕ∗ , β∗ ), which leads to a contradiction.

A.3 Proof of Proposition 3

A.4 Proof of Proposition 4