Information Design Multistage Game

INFORMATION DESIGN IN MULTI-STAGE GAMES
MILTIADIS MAKRIS AND LUDOVIC RENOU
A BSTRACT. This paper generalizes the concept of Bayes correlated equilib-

rium (Bergemann and Morris, 2016) to multi-stage games. We apply our
characterization results to a number of illustrative examples and applica-
tions.
K EYWORDS: Multi-stage games, information design, communication equi-

librium, sequential communication equilibrium, information structures, Bayes
correlated equilibrium, revelation principle.
JEL C LASSIFICATION: C73, D82.
Date: September 29, 2022.

Ludovic Renou gratefully acknowledges the support of the Agence Nationale pour la
Recherche under grant ANR CIGNE (ANR-15-CE38-0007-01) and through the ORA Project
“Ambiguity in Dynamic Environments” (ANR-18-ORAR-0005). We thank Laura Doval,
Stephen Morris, Sujoy Mukerji, Peter Norman, Alessandro Pavan, and Alex Wolitzky for
insightful comments and the audiences at the many seminars we have given. We are par-
ticularly indebted to Tristan Tomala for his generosity with time, pointed discussions and
perspective comments.
1
2
1. I NTRODUCTION
This paper generalizes the concept of Bayes correlated equilibrium (Berge-

mann and Morris, 2016) to multi-stage games. In a multi-stage game, a set
of players interact over several stages and, at each stage, players receive
private signals about past and current (payoff-relevant) states, past actions
and past signals, and choose actions. Repeated games and, more generally,
stochastic games are examples of multi-stage games.
Consider an analyst, who postulates a multi-stage game, which we call the

base game, but also acknowledges that players may receive additional sig-
nals, which can depend on past and current states, past actions, and past
and current signals (including the past additional ones). Define an expan-
sion of the base game to be a multi-stage game that differs from the base
game only in that players receive additional signals. Which predictions can
the analyst make if he does not want to hypothesize a fixed expansion? Al-
ternatively, consider an information designer who can design the additional
information players receive. Which outcomes can the information designer
achieve?
Bergemann and Morris (2016) address the above questions within the class
of static games. These authors show that the Bayes correlated equilibria
of the (static) base game characterize all the predictions the analyst can
make or, equivalently, the outcomes the information designer can achieve.
(See below for an informal definition of a Bayes correlated equilibrium.)
In many economic applications, however, the interaction between the eco-
nomic agents is best modeled as a dynamic game, where the agents receive
information over time and have the opportunity to make multiple decisions.
As an example, consider the refinancing operations of central banks. Typi-

cally, central banks organize weekly tender auctions to provide short-term
liquidities to financial institutions. While extensive regulations carefully
INFORMATION DESIGN IN MULTI-STAGE GAMES 3
specify the auction formats central banks use, the information the financial
institutions and the central banks receive over time as well as the commu-
nication between them are substantially harder to model. An analyst may
thus want to postulate a base game, which captures all that is known to
him – auction format, public annoucements, public statistics – and to re-
main agnostic about the private information the financial institutions and
central banks have. In other words, the analyst considers all possible expan-
sions of the base game. Recent contributions in the econometrics literature
on partial identification have adopted such an approach. See Bergemann,
Brooks and Morris (2019), Gualdani and Sinha (2021), Magnolfi and Ron-
coroni (2017), Syrgkanis, Tamer, and Ziani (2018).
Our main contribution is methodological. We derive several generaliza-

tions of the concept of Bayes correlated equilibrium, where each general-
ization corresponds to a solution concept for multi-stage games. We focus
primarily on the concept of Bayes-Nash equilibrium. While refinements are
frequently used in applications, we do so for a simple reason: the logical
arguments do not differ from one solution concept to another. Our main
theorem (Theorem 1) states an equivalence between (i) the set of all distri-
butions over states, signals in the base game, and actions induced by all
Bayes-Nash equilibria of all expansions of the base game, and (ii) the set of
all distributions over states, signals in the base game, and actions induced
by all Bayes correlated equilibria of the base game.
At a Bayes correlated equilibrium of the base game, at each stage, an “om-

niscient” mediator, who knows everything that has occurred, makes pri-
vate recommendations of actions to players, conditional on past and cur-
rent states and signals, past actions and past recommendations. In other
words, the mediator makes recommendations at each history of the base
game. Moreover, at each stage, players have an incentive to be obedient,
if they have never disobeyed in the past, and expect others to have been
obedient in the past and to continue to be in the future. We stress here
4 MILTIADIS MAKRIS AND LUDOVIC RENOU
that the “omniscient” mediator is a metaphor, an abstract entity, which only

serves as a tool to characterize all the equilibrium outcomes we can obtain
by varying the information structures. Whenever we refer to the media-
tor making recommendations at a given history, it should be understood as
the information structure –the statistical experiment– generating the addi-
tional signals conditional on the given history.
The logical arguments are simple. Fix an expansion of the base game and
a (Bayes-Nash) equilibrium. We show that we can emulate the equilibrium
of the expansion as an equilibrium of an auxiliary mediated game, where a
dummy (additional) player makes reports to a mediator and the mediator
sends messages to the original players. In that auxiliary mediated game,
the dummy player knows the actions, signals and states and the messages
the mediator sends are the additional signals of the expansion. We can then
apply the classical revelation principle of Myerson (1986) and Forges (1986)
to replicate the equilibrium of the auxiliary mediated game as a canonical
equilibrium of the “direct” game, i.e., as an equilibrium of a mediated game
where players report their private information to the mediator, the medi-
ator recommends actions and players are truthful and obedient, provided
they have been in the past. At that canonical equilibrium, the mediator is
“omniscient” at truthful histories and players are obedient provided they
have been in the past: we have a Bayes correlated equilibrium. The very
same logic generalizes to a variety of other solution concepts. All we need is
a revelation principle.
Finally, we provide two illustrations of the broad applicability of our results.

In particular, we extend the characterization of de Oliveira and Lamba
(2019).
The closest paper to ours is Bergemann and Morris (2016), henceforth BM.
These authors characterize the set of distributions over actions, signals in
the base game, and states induced by all Bayes-Nash equilibria of all expan-
sions of static base games, and show the equivalence with the distributions
induced by the Bayes correlated equilibria of the static base games. The
present paper generalizes their work to dynamic problems.1 Two main in-
sights emerge from our generalization.
The first insight of our analysis is that we genuinely need the mediator to
make recommendations at all histories. To understand the need for this,
note that even in dynamic games where all the states and signals about
the states are drawn ex-ante, it would not be enough to have the media-
tor recommend strategies as a function of the realized states and signals
at the first stage only. The reason is that players’ signals at interim stages
may also provide private information about the actions taken by players
in earlier stages. For instance, if the base game is a repeated game with
imperfect monitoring, a possible expansion is to perfectly inform players of
past actions. As a result, if the mediator could not react to deviations that
are unobserved by some players, it might not be able to induce the appro-
priate continuation play. In fact, as the introductory example (Section 2)
demonstrates, applying the definition of BM on the strategic form of even
the simplest multi-stage games does not characterize what we can obtain by
considering all equilibria of all expansions of the base game – an extensive-
form game. The need to work with games in extensive form in order to
characterize their Bayes correlated equilibria echoes a similar observation
in Myerson (1986) regarding the gradual release of information in commu-
nication games. This is not coincidental: Bayes correlated equilibria are
communication equilibria (Forges, 1986; Myerson, 1986) of mediated games
with an “omniscient mediator.”
The second insight is that our characterization of Bayes correlated equilib-

rium in multi-stage games generalizes to any solution concept for which a
1
Our work is related in spirit to Penta (2015) and Penta and Zuazo-Garin (2021). The first
paper extends the belief-free approach to robust mechanism design to dynamic problems,
in which agents obtain information over time about payoff-relevant states, but where ac-
tions are public. The second paper considers the question of robust predictions when com-
mon knowledge assumptions about players’ information on each other moves are relaxed;
payoffs remain common knowledge. In contrast to our paper, where we consider all ex-
pansions, Penta and Zuazo-Garin (2021) incorporate only “local” perturbations of the belief
hierarchies.
revelation principle holds. In particular, this is true for the two versions
of perfect Bayesian equilibrium we consider (Section 5). This is particu-
larly important for many economic applications. Bargaining problems (e.g.,
Bergemann, Brooks and Morris, 2015), allocation problems with aftermar-
kets (e.g., Calzolari and Pavan, 2006, Giovannoni and Makris, 2014, and
Dworczak, 2017), dynamic persuasion problems (Ely, 2017 and Renault,
Solan and Vieille, 2017) are all instances of dynamic problems, where se-
quential rationality is a natural requirement.
Doval and Ely (2020) is another generalization of the work of BM and nicely
complements our own generalization. These authors take as given states,
consequences and state-contingent payoffs over the consequences, and char-
acterize all the distributions over states and consequences consistent with
the players playing according to some extensive-form game. Our work dif-
fers from theirs in two important dimensions. First, we take as given the
base game (and, thus, the order of moves). In some economic applications, it
is a reasonable assumption. For instance, if we think about the refinancing
operations of central banks, the auction format and their frequencies define
the base game. If a first-price auction is used to allocate liquidities, it would
not make sense to consider games, where another auction format is used. In
other applications, this is less reasonable. Second, unlike Doval and Ely, we
are able to accommodate dynamic problems, where the evolution of states
and signals is controlled by the players through their actions. This is a nat-
ural assumption in many economic problems, such as mergers with ex-ante
match-specific investments or inventory problems.
Finally, this paper contributes to the literature on correlated equilibrium

and its generalizations, e.g., communication equilibrium (Myerson, 1986,
Forges, 1986), extensive-form correlated equilibrium (von Stengel and Forges,
2008), or Bayesian solution (Forges, 1993, 2006).2 The concept of Bayes cor-
related equilibrium is a generalization of all these notions. Solan (2001) is
a notable exception. In stochastic games with players perfectly informed
of past actions and past and current states, Solan considers general com-
munication devices, where the mediator sends messages to the players as
a function of past messages sent and received, and the history of the game,
i.e., the past actions, and the past and current states. Solan’s mediator is
omniscient. For that class of games, Solan shows that the set of Bayes cor-
related equilibrium payoffs is equal to the set of extensive-form correlated
equilibrium payoffs. As we show in the introductory example, this equiva-
lence does not hold if players are not perfectly informed of past actions. See
Forges (1985) for a related result.
2. A N I NTRODUCTORY E XAMPLE
This section illustrates our main results with the help of a simple example.
The example highlights a novel and distinctive aspect of (unconstrained)
information design in dynamic games: In addition to providing informa-
tion about payoff-relevant states, the designer can choose the information
players have about the past actions of others. E.g., in voting problems, the
designer can choose how much information to reveal about past votes.
Example 1. There are two players, labelled 1 and 2, and two stages. Player
1 chooses either T or B at the first stage, while player 2 chooses either L or
R at the second stage. In the base game, player 2 has no information about
player 1’s choice. Figure 1 depicts the base game, with the payoff of player
1 as the first coordinate.
Suppose that the information designer wants to maximize player 1’s payoff.
Which information structure(s) should it design?
2
The concept of extensive-form correlated equilibrium was first introduced in Forges (1986).
The concept introduced in von Stengel and Forges (2008) differs from the one in Forges
(1986).
1
T B
2
L R L R
(2, 2) (0, 1) (3, 0) (1, 1)
F IGURE 1. The base game
To address that question, we generalize the work of Bergemann and Mor-

ris (2016) to dynamic games. Bergemann and Morris (2016) characterize
the distributions over outcomes we can induce by varying the information
players have in static games of incomplete information–that is, as we vary
the information a player has about the payoff-relevant states and the in-
formation of others. They prove that the Bayes correlated equilibria of the
static base game characterize the distributions we can induce. At a Bayes
correlated equilibrium, an omniscient mediator recommends actions, and
the players have an incentive to be obedient.
Thus, a naive idea is to apply the concept of Bayes correlated equilibrium to

the strategic form of the base game. In our example, the unique Bayes corre-
lated equilibrium of the strategic form is (B, R) with a payoff profile of (1, 1).
Working with the strategic form is, however, too restrictive. E.g., if the de-
signer perfectly informs player 2 of player 1’s action, the induced game has
an equilibrium with outcome (T, L) and associated payoff (2, 2). (In static
games with complete information, the set of Bayes correlated equilibria co-
incides with the set of correlated equilibria.)
Our approach is to have the omniscient mediator recommending actions to

the players not only at the initial history, but at each history of the dynamic
game. In addition, the players must have an incentive to be obedient, pro-
vided they have been obedient in the past. This approach generalizes the
definition of Bayes correlated equilibria of Bergemann and Morris (2016) to
multi-stage games and illustrates the need to work on the extensive-form
games. (Myerson, 1986, has already pointed out the insufficiency of the
strategic form; see Section 2 of his paper.)
We now illustrate how our approach works in our example. Since the medi-
ator is omniscient and makes recommendations at all histories, we need to
consider two recommendation kernels. The first kernel specifies the proba-
bility of recommending an action to player 1 at the first stage. The second
kernel specifies the probability of recommending an action to player 2 at
the second stage as a function of the action recommended and chosen at the
first stage. Players must have an incentive to be obedient. We claim that
there exist such recommendation kernels with a payoff profile of (5/2, 1).
To see this, assume that the mediator recommends with probability 1/2
player 1 to play T and with the complementary probability to play B at
the first stage, and recommends player 2 to play L at the second stage if,
and only if, player 1 was obedient at the first stage. (Otherwise, the media-
tor recommends R.) We now prove that the players have an incentive to be
obedient.
If player 2’s recommendation is L, he believes that player 1 has played T

with probability 1/2 and, thus, expects a payoff of 1 if he plays L. He there-
fore has an incentive to be obedient. If player 2’s recommendation is R, we
are off the equilibrium path and any conjecture that puts probability of at
least 1/2 on player 1 having played B makes R optimal. As for player 1, he
clearly has an incentive to be obedient when his recommendation is B since
he gets his highest payoff. When his recommendation is T , a deviation to B
is unprofitable because this leads player 2 to play R. Thus, we have a Bayes
correlated equilibrium with a payoff profile of (5/2, 1).
We now argue that no information structures can give player 1 a payoff

higher than 5/2, hence answering our initial question. Since player 2 can
always play R, player 2’s payoff cannot be lower than 1. Therefore, within
the set of feasible payoff profiles, conditional on player 2’s getting a payoff
of at least 1, player 1’s highest payoff is 5/2.
Finally, we now explain how we can use the Bayes correlated equilibrium to
design an information structure, whose associated expansion generates an
equilibrium payoff of (5/2, 1). The idea is simple: Think of recommendations
as signals. Accordingly, suppose that there are two equally likely signals, t
and b, at the first stage, and two signals l and r at the second stage. Player
1 privately observes the first signal, while player 2 observes the signal l if
only if either (t, T ) or (b, B) is the profile of signal and action at the first
stage. With such information structure, players have an incentive to play
according to their signals and, thus, we obtain the payoff profile (5/2, 1).
As a final observation, consider the alternative base game in Figure 2,

where we reverse the order of play. (The first coordinate of a payoff vec-
tor refers to player 1’s payoff.)
2
L R
1
T B T B
(2, 2) (3, 0) (0, 1) (1, 1)
F IGURE 2. An alternative base game: reversing the order of play.
Note that the two base games have the same strategic form. Yet, their set
of Bayes correlated equilibria differ. The alternative base game has the
unique payoff profile (1, 1) as Bayes correlated equilibrium’s payoff. Indeed,
on path, player 1 must play B since T is strictly dominated. And since
player 2 can guarantee a payoff of 1 by playing R, the unique outcome is
(B, R).
3. M ULTI - STAGE G AMES AND E XPANSIONS
The model follows closely Myerson (1986). There is a set I of n players, who
interact over T < +∞ stages, numbered 1 to T . (With a slight abuse of nota-
tion, we denote T the set of stages.) At each stage, a payoff-relevant state is
drawn, players receive private signals about past and current states, past
private signals and actions, and choose an action. We are interested in char-
acterizing the joint distributions over profiles of states, actions and signals,
which arise as equilibria of “expansions” of the game, i.e., games where
players receive additional signals.
3.1. The base game. We first define the base game Γ, which corresponds
to the game being played if no additional signals are given to the players.
At each stage t, a state ωt ∈ Ωt is drawn, player i ∈ I receives the private
signal si,t ∈ Si,t , which may depend probabilistically on the current and past
states, past signals and actions, and then chooses an action ai,t ∈ Ai,t . All
sets are non-empty and finite.
The description of the base game is very flexible. It can accommodate games
with perfect or imperfect observation, repeated games with incomplete in-
formation, or stochastic games, among others. We stress, however, that the
base game puts certain restrictions on what an information designer can do
or what robustness exercises an analyst can conduct. For instance, once the
order of moves is fixed, we cannot have a player receiving information (orig-
inal or additional) about the moves of players who, according to the base
game, move at a later stage.
We now introduce some notation. We write At := ×i∈I Ai,t for the set of
actions at stage t and A := ×t∈T ×i∈I Ai,t for the set of profiles of actions.
We let Hi,t = Ai,t−1 × Si,t be the set of player i’s new information at the
beginning of stage t ∈ {2, . . . , T }, Hi,1 = Si,1 the set of initial information,
and Hi,T +1 = Ai,T the set of new information at the end of the last stage.
We denote p1 (h1 , ω1 ) the joint probability of (h1 , ω1 ) at the beginning of the

first stage and pt+1 (ht+1 , ωt+1 |at , ht , ω t ) the joint probability of (ht+1 , ωt+1 ) at
stage t + 1 given that at is the profile of actions played at stage t and (ht , ω t )
is the history of actions played, signals received and states realized at the
beginning of stage t. We assume perfect recall and, therefore, impose that
pt+1 ((bt , st+1 ), ωt+1 |at , ht , ω t ) = 0 if bt 6= at .
We denote HΩ the subset of ×Tt=1

+1
(×i∈I Hi,t × Ωt ) that consists of all terminal
histories of the game, with generic element (h, ω).3 The history (h, ω) is in
HΩ if and only if there exists a profile of actions a ∈ A such that
Y
pa (h, ω) := p1 (h1 , ω1 ) · pt+1 (ht+1 , ωt+1 |at , ht , ω t ) > 0.
t∈T
For any vector (h, ω), we can denote various sub-vectors: hi = (hi,1 , . . . , hi,t , . . . , hi,T +1 )
the private (terminal) history of player i, hti = (hi,1 , . . . , hi,t ) the private his-
tory of player i at stage t, ht = (h1,t , . . . , hn,t ) the profile of actions played at
stage t − 1 and signals received at stage t, ht = (h1 , . . . , ht ) the history of
signals and actions at stage t, ω = (ω1 , . . . , ωT ) the profile of realized states,
and ω t = (ω1 , . . . , ωt ) the profile of states realized up to stage t, with cor-
responding sets Hi = {hi : (h, ω) ∈ HΩ for some ω}, Hit = {hti : (h, ω) ∈
HΩ for some ω}, Ht = {ht : (h, ω) ∈ HΩ for some ω}, H t = {ht : (h, ω) ∈
HΩ for some ω}, Ω = {ω : (h, ω) ∈ HΩ for some h}, Ωt = {ω t : (h, ω) ∈
HΩ for some h}. We write H t Ωt for the restriction of HΩ to the first t stages.
b t := ×i∈I H t . Similar notation will apply to other
b := ×i∈I Hi and H
We let H i
sets. If there is no risk of confusion, we will not formally define these addi-
tional notation.
The payoff to player i is ui (h, ω) when the terminal history is (h, ω) ∈ HΩ.
We assume that payoffs do not depend on the signal realizations, i.e., for any
two histories h = (a, s) and h0 = (a0 , s0 ) such that a = a0 , ui (h, ω) = ui (h0 , ω) for
all ω, for all i.4 Throughout, we refer to the signals in S as the base signals.
3
The sets ST +1 and ΩT +1 are defined to be a singleton.
4
This is without loss of generality as we can always redefine the states to include the
signals.
3.2. Expansions. In an expansion of the base game, at each stage, play-

ers receive additional signals, which may depend probabilistically on past
and current states, past and current signals (including the past additional
ones), and past actions. Thus, players can receive additional information
not only about the realization of current and past (payoff-relevant) states
(such as the valuations for objects in auction problems), but also about the
past realization of actions (as in repeated games with imperfect monitor-
ing). Throughout, we use the same notation as in the base game to denote
relevant sub-vectors and their corresponding sets.
Formally, an expansion is a collection of sets of additional private signals

(Mi,t )i,t and probability kernels (ξt )t such that all sets of additional signals
are non-empty and finite, ξ1 : H1 × Ω1 → ∆(M1 ), and ξt : H t × M t−1 ×
Ωt → ∆(Mt ) for all t ≥ 2.5 Intuitively, at each stage t, player i receives the
additional private signal mi,t ∈ Mi,t , with
ξt (mt |ht , mt−1 , ω t )
the probability of mt when (ht , mt−1 , ω t ) is the history of actions, base signals,
states, and past additional signals at the beginning of stage t. We write M
for the collection (Mi,t )i,t and ξ for (ξt )t .
Together with the base game Γ, an expansion (M, ξ) induces a multi-stage

game Γπ , where at each stage t, a payoff-relevant state ωt is realized, player
i receives the private signal (si,t , mi,t ) and takes an action ai,t . To complete
the description of the induced multi-stage game, we let π1 (h1 , m1 , ω1 ) :=
ξ1 (m1 |h1 , ω1 )p1 (h1 , ω1 ) be the probability of (h1 , m1 , ω1 ) at the first stage and
πt+1 (ht+1 , mt+1 , ωt+1 |at , ht , mt , ω t ) := ξt+1 (mt+1 |ht+1 , mt , ω t+1 )pt+1 (ht+1 , ωt+1 |at , ht , ω t )
the probability of (ht+1 , mt+1 , ωt+1 ), when at is the profile of actions played
at stage t and (ht , mt , ω t ) is the history of actions, signals and states at the
beginning of stage t. With a slight abuse of language, we use the word
5
The set Mi,T +1 is a singleton.
“expansion” to refer to the collection of additional signals and kernels (M, ξ)

as well as to the multi-stage game Γπ induced by it.
It is worth stressing that our definition of expansions implies that the evo-
lution of future states and base signals, as a function of earlier actions,
states and base signals, is known to the analyst/designer in the sense of
being pinned down by the base game (and, hence, remains constant across
all expansions). Additional signals are just information and do not cause
future states.
We denote HM Ω the set of all terminal histories, with (h, m, ω) ∈ HM Ω if

and only if (h, ω) ∈ HΩ. We do not require, however, histories (h, m, ω) ∈
HM Ω to have strictly positive probability for some profile of actions, i.e., for
all (h, m, ω) ∈ HM Ω, we do not require the existence of a ∈ A such that
Y
π a (h, m, ω) := π1 (h1 , m1 , ω1 ) · πt+1 (ht+1 , mt+1 , ωt+1 |at , ht , mt , ω t ) > 0.
t∈T
In other words, we do not prune out the additional messages, even when
they have zero probability under all action profiles. (Note that (h, ω) ∈ HΩ
corresponds to m π a (h, m, ω) > 0 for some a.)
P
In closing, it is worth noting that an expansion ξ induces a collection of

kernels π, with the property that margHΩ π a = pa for all a ∈ A, that is,
X Y
π1 (h1 , m1 , ω1 ) · πt+1 (ht+1 , mt+1 , ωt+1 |at , ht , mt , ω t ) =
(m1 ,...,mT ) t∈T
Y
p1 (h1 , ω1 ) · pt+1 (ht+1 , ωt+1 |at , ht , ω t ). (†)
t∈T
We call this property consistency. As explained in the introduction, in static

problems, the converse is also true, that is, any consistent kernel π induces
an expansion ξ. This equivalence breaks down in dynamic problems. This is
the case when the additional signals are not just signals, but also cause the
realization of future states and base signals, so that the base game stops
pinning down their evolution (see Example 2 below).
4. A F IRST E QUIVALENCE T HEOREM AND AN A PPLICATION
This section contains our first characterization theorem–other character-

izations will differ by the solution concepts adopted. In Section 4.1, we
consider the concept of Bayes-Nash equilibrium. This allows us to present
our first characterization theorem in the simplest possible terms, without
cluttering the analysis with issues such as consistency of beliefs, sequential
rationality, or truthfulness and obedience at off-equilibrium path histories.
As we will see, the main arguments extend almost verbatim to other solu-
tion concepts. In addition, if we are interested in proving an impossibility
result, e.g., whether efficiency obtains, the weaker the solution concept, the
stronger the result. Section 5 extends our analysis to two refinements of the
concept of Bayes-Nash equilibrium, which all impose sequential rationality.
4.1. A First Equivalence Theorem. We first define the concepts of Bayes-

Nash equilibrium and Bayes correlated equilibrium. Throughout, we fix an
expansion Γπ of Γ. A behavioral strategy σi is a collection of maps (σi,t )t∈T ,
with σi,t : Hit Mit → ∆(Ai,t ).
Definition 1 (BNE): A profile σ of behavioral strategies is a Bayes-Nash

equilibrium of Γπ if
X X
ui (h, ω)Pσ,π (h, m, ω) ≥ ui (h, ω)P(σi0 ,σ−i ),π (h, m, ω),
h,m,ω h,m,ω
for all σi0 , for all i, with Pσ̃,π ∈ ∆(HM Ω) denoting the distribution over pro-
files of actions, signals and states induced by σ̃ and π.
We let BN E(Γπ ) be the set of distributions over HΩ induced by the Bayes-

Nash equilibria of Γπ .
We now state formally the main objective of our paper: we want to provide a
S
characterization of the set Γπ an expansion of Γ BN E(Γπ ), i.e., we want to char-
acterize the distributions over the outcomes HΩ of the base game Γ that
we can induce by means of some expansion Γπ of the base game, without
any reference to particular expansions. To do so, we need to introduce the

concept of Bayes correlated equilibrium of Γ.
Bayes correlated equilibrium. Consider the following mediated exten-

sion of Γ, denoted M(Γ). At each period t, player i observes the private
signal hi,t , receives a private recommendation âi,t from an “omniscient” me-
diator – a mediator who knows everything that has occurred in M(Γ) – and
chooses an action ai,t . We let τi,t : Hit × Ati → ∆(Ai,t ) be an action strategy at
∗
period t and write τi,t for the obedient strategy. We write τi for (τi,t )t and τ
for (τi )i .
Definition 2 (BCE): A Bayes correlated equilibrium of Γ is a collection

of recommendation kernels µt : H t Ωt × At−1 → ∆(At ) such that τ ∗ is an
equilibrium of the mediated game M(Γ), that is,
X X
ui (h, ω)Pµ◦τ ∗ ,p (h, ω, â) ≥ ui (h, ω)Pµ◦(τi ,τ−i
∗ ),p (h, ω, â)
h,ω,â h,ω,â
for all τi , for all i, with Pµ◦τ̃ ,p denoting the distribution over profiles of ac-
tions, base signals, states and recommendations induced by µ ◦ τ̃ and p.
We let BCE(Γ) be the set of distributions over HΩ induced by the Bayes

correlated equilibria of Γ. The set BCE(Γ) is convex.
It is instructive to compare the concept of Bayes correlated equilibrium and

communication equilibrium (Forges, 1986, Myerson, 1986). In a communi-
cation equilibrium, the mediator relies on the information provided by the
players to make recommendations, while in a Bayes correlated equilibrium
it is as if the mediator knows the realized states, actions and base signals
prior to making recommendations–the mediator is omniscient. Let CE(Γ) be
the distributions over HΩ induced by the communication equilibria of Γ. For
all multi-stage games Γ, we have that CE(Γ) ⊆ BCE(Γ) since the omniscient
mediator can always replicate the Forges-Myerson mediator. Since we also
have that BN E(Γ) ⊆ CE(Γ), we have the inclusion BN E(Γ) ⊆ BCE(Γ). How-
ever, it is a priori unclear whether BN E(Γπ ) ⊆ BCE(Γ) for all expansions
Γπ of Γ since players have additional signals in Γπ , while the omniscient

mediator of Γ has no additional signals. A consequence of our main result,
Theorem 1, is that it is indeed the case.
Theorem 1. We have the following equivalence:

[
BCE(Γ) = BN E(Γπ ).
Γπ an expansion of Γ
Theorem 1 states an equivalence between (i) the set of distributions over

actions, base signals and states induced by all Bayes correlated equilibria
of Γ and (ii) the set of distributions over actions, base signals and states we
can obtain by considering all Bayes-Nash equilibria of all expansions of Γ.6
Theorem 1 generalizes the work of BM to multi-stage games. As the intro-

ductory example demonstrates, our definition of a Bayes correlated equi-
librium is, in general, weaker than applying the definition of BM to the
strategic form of the base game, which would amount to making recommen-
dations of strategies at the first stage, as a function of the realized states
and base signals. Yet, in multi-stage base games where (i) all the states and
base signals are drawn at the first stage, (ii) players observe their signals
sequentially over time, and (iii) past actions are perfectly observable, the
two formulations are equivalent. To see this, we first define the concept of
“ex-ante” Bayes-correlated equilibrium. To ease notation, assume that the
profile (s, ω) of base signals and states is drawn with probability p(s, ω) at
the first stage, and no states and base signals are drawn at later stages. An
“ex-ante” Bayes-correlated equilibrium is a kernel µ : S × Ω → ∆(Σ) (where
6
In a supplementary document, we also prove the equivalence with the set of distribu-
tions over actions, base signals and states we can obtain by considering all communication
equilibria of all expansions of Γ.
Σ is the set of strategies of Γ), which satisfies

X
p(s, ω)µ(σ|s, ω)P(σi ,σ−i ) (a|s)ui (a, ω) ≥
ω,s,σ−i ,a
X
p(s, ω)µ(σ|s, ω)P(σi0 ,σ−i ) (a|s)ui (a, ω),
ω,s,σ−i ,a
for all σi0 , σi , i. In words, players must have an incentive to be obedient.
Clearly, any distribution induced by an “ex-ante” Bayes-correlated equilib-

rium is in BCE(Γ). Indeed, we can interpret the kernel µ as an expansion,
where M1 = Σ and ξ1 = µ, and an equilibrium of that expansion is to play
according to one’s additional signal–the recommended strategy.
Conversely, consider a Bayes-correlated equilibrium µ of such multi-stage

games, where µt (ât |ât−1 , at−1 , s, ω) is the probability of recommending ât ,
when the profile of past recommendations and actions is (ât−1 , at−1 ) and
the profile of states and signals is (s, ω). The intuition is best understood
when there is a single stage. At a Bayes-correlated equilibrium, the om-
niscient mediator, who knows (s, ω), recommends ai to player i with prob-
P
ability µ(ai |s, ω) = a−i µ((ai , a−i )|s, ω), and player i has an incentive to be
obedient when his signal is indeed si . An alternative, but equivalent, formu-
[ai ,si ]
lation would be to recommend a strategy σi to player i, which stipulates
to play ai when the signal is si . Player i would indeed have an incentive to
follow that strategy when his signal is si . When there are more stages, a
similar construction is possible. We now sketch it and refer the reader to
the supplementary materials for the details.
The idea is to associate with (µt (·|·, ·, s, ω))t an outcome-equivalent distribu-

tion over feedback rules f = (ft : At−1 × At−1 → At )t (as in Kuhn’s theorem),
and with each feedback rule f and signal s, a strategy σ [f,s] such that players
have an incentive to be obedient. (When there is a single stage, feedback
rules are actions.) Formally, for each feedback rule f , define recursively:
f 1 (∅) = f1 (∅, ∅) and f t (at−1 ) = (f t−1 (at−2 ), ft (at−1 , f t−1 (at−2 ))) for all at−1 , for
[f,s]
all t. We now associate a pure strategy σi with the feedback rule f and the
[f,s]
profile of signals s, as follows: σi,t (sti , at−1 ) = fi,t (at−1 , f t−1 (at−2 )) for all at−1 ,
for all t, with fi,t (at−1 , f t−1 (at−2 )) the i-th component of ft (at−1 , f t−1 (at−2 )). In
[f,s]
words, σi dictates player i to play as f dictates when his signal’s realiza-
[f,s]
tions are (si ). Let σ [f,s] be the profile (σi )i . It is important to stress that
the strategy is well-defined because all past actions are perfectly observed.
Without that assumption, this wouldn’t be the case. Finally, if the mediator
recommends σ [f,s] with probability
µ(σ [f,s] |s, ω) = ××

t at−1 ,ât−1
µt (ft (at−1 , ât−1 )|at−1 , ât−1 , s, ω),
then it is routine to verify that the players have an incentive to be obedi-

ent and the desired distribution over actions, signals and states, is imple-
mented. The equivalence, however, breaks down if either actions are not
perfectly observed (as in the introductory example) or states and signals
are not drawn at the first stage, e.g., because the players control their evo-
lutions through their actions.
S
As in BM, the proof of BCE(Γ) ⊆ BN E(Γπ ) is constructive.
S
However, unlike BM’s constructive proof, our proof of Γπ an expansion of Γ BN E(Γπ ) ⊆
BCE(Γ) is non-constructive: it utilizes the revelation principle of Forges
(1986) and Myerson (1986). This approach has two main advantages: (i)
it reveals the main logical arguments, which are somewhat hidden in con-
structive proofs, and (ii) its generalization to many other solution concepts
is straightforward. The central arguments are the following. Consider an
expansion Γπ of Γ and an equilibrium distribution µd ∈ BN E(Γπ ). By def-
inition, there exists a Bayes-Nash equilibrium σ of Γπ , which induces µd .
The main idea is to replicate the expansion Γπ and its equilibrium σ as a
Bayes-Nash equilibrium of an auxiliary mediated game M∗ (Γ), which we
now describe. The game M∗ (Γ) has one additional dummy player, called
player 0, and a Forges-Myerson mediator, who receives reports by and sends
messages to the players. At the first stage, Nature draws (h1 , ω1 ) with prob-
ability p1 (h1 , ω1 ), player i observes hi,1 and player 0 observes (h1 , ω1 ). Player
0 then reports (ĥ1 , ω̂1 ) to the mediator; all other players do not make reports
(their sets of reports are singletons). The mediator then draws the message
m1 with probability ξ1 (m1 |ĥ1 , ω̂1 ) and sends mi,1 to player i. Player 0 does
not receive a message. Finally, player i chooses an action ai,1 ; player 0 does
not take an action. Consider now a history (at−1 , ht−1 , ω t−1 ) of past actions,
signals and states and a history ((ĥt−1 , ω̂ t−1 ), mt−1 ) of reports and messages.
Stage t unfolds as follows:
- Nature draws (ht , ωt ) with probability pt (ht , ωt |at−1 , ht−1 , ω t−1 ).

- Player i ∈ I observes the signal hi,t and player 0 observes (ht , ωt ).
- Player 0 reports (ĥt , ω̂t ) to the mediator. All other players do not
make reports.
- The mediator draws the message mt with probability ξt (mt |ĥt , mt−1 , ω̂ t )
and sends the message mi,t to player i. Player 0 does not receive a
message.
- Player i takes an action ai,t . Player 0 does not take an action.
If player 0 is truthful and each player i ∈ I follows σi , we have a Bayes-Nash

equilibrium of the mediated game M∗ (Γ), with equilibrium distribution µd .
From the revelation principle of Forges (1986) and Myerson (1986), there
exists a canonical equilibrium which implements µd , i.e., an equilibrium
where players report their private information to the mediator, the medi-
ator recommends actions and players are truthful and obedient provided
that they have been in the past.7 At truthful histories, the mediator is om-
niscient and players have an incentive to be obedient provided they have
been in the past: this is the Bayes correlated equilibrium.
Before applying Theorem 1, two additional remarks are worth making.

First, the above arguments are not limited to the concept of Bayes-Nash
equilibrium. The same arguments apply to all solution concepts, such as
weak perfect Bayesian equilibrium or conditional probability perfect Bayesian
equilibrium, which admit a revelation principle. We formally state these
equivalences below. Second, the above arguments clearly demonstrate the
7
Proposition 2 in Sugaya and Wolitzky (2021) is a restatement of the revelation principle of
Forges (1986) and Myerson (1986). It also applies to mediated games, where the mediator
receives private signals in addition to the players’ reports. With such restatement, we can
dispense with the dummy player and have the mediator directly learn (ht , ωt ).
role our definition of an expansion plays. It makes it possible for the me-
diator to replicate any expansion as the kernels ξt are assumed measur-
able with respect to the mediator’s histories. With the alternative and
weaker definition of an expansion as a consistent information structure,
i.e., marg π a = pa for all a, it is no longer guaranteed that the mediator,
despite being omniscient, can simulate any expansion, as the next example
illustrates.
Example 2. We first define the base game. There are a single player, two
stages, two actions A1 = {0, 1} at the first stage, two states Ω2 = {0, 1} at
the second stage, and all other sets are singletons. The probabilities are:
p2 (ω2 = 1|a1 = 1) = 5/6 and p2 (ω2 = 1|a1 = 0) = 1/2. The player’s payoff is
one (resp., zero) if the second-stage state is zero (resp., one), regardless of
his action.
Consider now the following information structure: M1 = {0, 1}, M2 is a

singleton, π1 (m1 = 1) = 1/2, π2 (ω2 = 1|a1 = 1, m1 = 1) = 2/3, π2 (ω2 = 1|a1 =
0, m1 = 1) = 1, π2 (ω2 = 1|a1 = 1, m1 = 0) = 1, and π2 (ω2 = 1|a1 = 0, m1 = 0) =
0. This information structure is consistent, but there are no kernels (ξ1 , ξ2 )
that induce this information structure from the base game. The issue is
that action a1 and additional signal m1 jointly cause the second-stage state
ω2 . In other words, the signal m1 is not just additional information; it also
determines the realization of the second-period state.
Player’s optimal payoff is 2/3 in the game Γπ : the optimal strategy con-
sists in playing a1 = 1 (resp., a1 = 0) when m1 = 1 (resp., m1 = 0). The
player’s optimal strategy consists in choosing the action that maximizes
the likelihood of the second-stage state being 0. The induced distribution
µ over actions and states is µ(a1 = 0, ω2 = 0) = 1/2, µ(a1 = 0, ω2 = 1) = 0,
µ(a1 = 1, ω2 = 0) = 1/6, µ(a1 = 1, ω2 = 1) = 1/3. This is not a Bayes corre-
lated distribution. In any Bayes correlated equilibrium, the probability of
(a1 , ω2 ) is µ1 (a1 )p2 (ω2 |a1 ) and there is no µ1 that induces the distribution µ.
4.2. Application I: Rationalizing Dynamic Choices. Suppose that an

analyst observes the choices of a decision-maker over a finite number of
periods, but does not observe the information the decision-maker had. Sup-
pose, furthermore, that the analyst assumes that the state does not change
over time. Which profiles of choices can be rationalized? This question was
recently addressed by de Oliveira and Lamba (2019), under the assump-
tion that the information the decision-maker receives over time is indepen-
dent of his actions. With our notation, this is equivalent to requiring that
ξt (·|at−1 , mt−1 , ω) = ξt (·|ãt−1 , mt−1 , ω) for all (at−1 , ãt−1 ), for all mt−1 , for all t.8
We call such expansions autonomous.
We now show how Theorem 1 makes it possible to extend their result to

all expansions. Throughout, we follow the terminology of de Oliveira and
Lamba.
We say that the profile of choices a∗ := (a∗1 , . . . , a∗T ) is rationalizable if there

exist a probability p ∈ ∆(Ω), sets of signals Mt and kernels ξt : At−1 × M t−1 ×
Ω → ∆(Mt ) such that the decision-maker chooses optimally and (a∗1 , . . . , a∗T )
is optimal for some realizations (ω, m) of states and signals. As de Oliveira
and Lamba, we assume the decision-maker payoff function u is known to
the analyst.
From Theorem 1, the profile of choices if there exists a probability p ∈ ∆(Ω)

and a Bayes correlated equilibrium µ such that Pµ◦τ ∗ ,p (a∗ ) > 0. Recall that
µ is a Bayes correlation equilibrium if:
X X
u(a, ω)Pµ◦τ ∗ ,p (a, ω, â) ≥ u(a, ω)Pµ◦τ,p (a, ω, â),
a,ω,â a,ω,â
for all τ . The objective is to derive conditions on the primitives, which guar-
antee the existence of such a Bayes correlated equilibrium.
8
As all sets of base signals are singletons, we do not denote them. Similarly, since the state
does not change over time, we write ω for the fully persistent state.
We need some additional notation. We say that D : A → ∆(A) is a deviation

plan if there exists a behavioral strategy τ such that
T
Y
D(a1 , . . . , aT |â1 , . . . , âT ) := τt (at | (â1 , . . . , ât ) , (a1 , . . . , at−1 ))
t=1
| {z } | {z }
recommendations choices
for all (â, a). A deviation plan specifies what the decision-maker would do if
he were to face a fixed sequence of recommendations.
We say that γ : A → A is a recommendation plan if there exist maps γt :

At−1 → At such that

γ(a) = γ1 (∅), γ2 (a1 ), . . . , γt (a1 , . . . , at−1 ), . . . , γT (a1 , . . . , aT −1 )
for all a, with the convention that A0 = {∅}. A recommendation plan spec-
ifies a recommendation for each fixed sequence of choices. We write γ >t (a)
for (γt (a1 , . . . , at−1 ), . . . , γT (a1 , . . . , aT −1 )).
Finally, we define the sets: Ba1 := (A1 \{a1 })×A2 ×· · ·×AT , Bat := {(a1 , . . . , at−1 )}×
(At \{at })×At+1 ×· · ·×AT for all t ∈ {2, . . . , T −1}, and BaT := {(a1 , . . . , aT −1 )}×
AT . (By convention, if T = 1, then Ba1 = A1 .) For all t < T , the set Bat is the
set of all profiles of choices, which coincide with a up to period t−1 and differ
from a at period t. Note that the sets are pairwise disjoint and Tt=1 Bat = A
S
for all a. We are now ready to state our main definition.
Definition 1. The profile a∗ is surely dominated if there exists a deviation

plan D such that for all ω, for all a, for all recommendation plans γ:
T
X X
∗
u(a , ω) < u(b, ω)D(b|a∗1 , . . . , a∗t , γ >t+1 (b)), (D1)
t=1 b∈B t ∗
a
T X
X
u(a, ω) ≤ u(b, ω)D(b|a1 , . . . , at , γ >t+1 (b)). (D2)
t=1 b∈Bat
Intuitively, a profile of choices is surely dominated if the decision-maker

has a deviation plan, which guarantees an improvement regardless of the
state ω, the period t at which the decision-maker is first disobedient, and
the subsequent recommendations γ >t+1 (b), which may depend on the past
choices. (By convention, the profile (a1 , . . . , aT , γ >T +1 (b)) = (a1 , . . . , aT ) for all
b.) We have the following characterization.
Theorem 2. The profile of choices a∗ is rationalizable if, and only if, it is not
surely dominated.
To get some intuition for Theorem 2, let us first consider the special case of
a single decision (T = 1). When the decision-maker makes a single decision,
it is well-known that a∗ is rationalizable if, and only if, it is not strictly
dominated, that is, there does not exist a mixed action α ∈ ∆(A) such that
u(a∗ , ω) <
P
b u(b, ω)α(b) for all ω. We now argue that sure dominance is
equivalent to strict dominance. To see this, if a∗ is strictly dominated by α,

then choosing D(·|a∗ ) = α and D(·|a) = 1(a) for all a 6= a∗ guarantees that a∗
is surely dominated. Conversely, choosing α = D(·|a∗ ) guarantees that a∗ is
strictly dominated.
To understand the role of (D2), we need to consider genuine dynamic prob-

lems. So, let us assume that the decision-maker has to choose twice (T = 2).
To play a∗ = (a∗1 , a∗2 ) with positive probability, the decision-maker must find
it optimal to play a∗1 at the first period given that he will play a∗2 with some
probability at the second period and other actions with the complementary
probability. Condition (D2) guarantees that the decision-maker would not
find it optimal to play a∗1 at the first period, regardless of what he would play
at the second period. Finally, the role of γ is to capture the fact that the rec-
ommendations and, therefore, the information the decision-maker receives
may depend on his past actions.
As already mentioned, de Oliveira and Lamba (2019) were the first to con-
sider the problem of rationalizing dynamic choices, restricting attention
to autonomous expansions. We extend their work in that we consider all
expansions. It is instructive to compare their characterization with ours.
Their main result states that the profile a∗ is rationalizable if, and only if, it
is not truly dominated. An action profile is truly dominated if there exists

a deviation plan that strictly increases the payoff along the action sequence
without worsening payoffs at other parts of the decision tree, regardless of
the state. Formally, the profile a∗ is truly dominated if there exists a devia-
tion rule D such that:
u(a∗ , ω) < u(b, ω)D(b|a∗ ),

P
b
P
u(a, ω) ≤ b u(b, ω)D(b|a),
for all ω, for all a.
Clearly, if a profile a∗ is surely dominated, then it is truly dominated. In-

deed, if we choose γ >t+1 (b) to be equal to (at+1 , . . . , aT ) for all (a, b, t), then we
recover the conditions for true dominance.
To understand the differences, recall that de Oliveira and Lamba (2019)

restrict attention to autonomous expansions. With such a restriction, it is
without loss of generality to assume that all signals are drawn ex-ante and
then they are gradually released to the decision-maker, independently of
what he does. It is easy to prove that this is equivalent to restricting atten-
tion to recommendation kernels, which depend on past recommendations,
base signals and states, but not on past actions, i.e., µt (·|at−1 , st , ω t , ât−1 ) =
µt (·|bt−1 , st , ω t , ât−1 ) for all (at−1 , bt−1 ). (See the supplementary material for a
proof.) We therefore have:
Pµ◦τ,p (a, ω, â) = Pµ◦τ,p (ω, â)Pµ◦τ,p (a|ω, â)
= Pµ◦τ ∗ ,p (ω, â)[τ1 (a1 |â1 ) × · · · × τt (at |ât , at−1 ) × · · · × τT (aT |âT , aT −1 )]
= Pµ◦τ ∗ ,p (ω, â)D(a|â),
where D is a deviation plan. The equality Pµ◦τ,p (ω, â) = Pµ◦τ ∗ ,p (ω, â) follows
from the fact that the recommendations made and the state realized are
independent of the decision-maker’s choices and, thus, do not depend on τ .
We can thus rewrite the obedience constraint as:

X X
Pµ◦τ ∗ ,p (ω, â)u(â, ω) ≥ Pµ◦τ ∗ ,p (ω, â)D(a|â)u(a, ω),
â,ω â,a,ω
for all deviation plans D. To recover the characterization of de Oliveira

and Lamba, it then suffices to follow the same steps as in the proof below,
starting with the above rewriting of the obedience constraint.
We conclude this discussion with an example, which demonstrates that a

profile can be truly dominated and not surely dominated. There are two
states, ω and ω 0 , three actions, `(eft), c(enter), r(ight), and two periods. The
inter-temporal payoff is the sum of the per-period payoff in Table 1.
T ABLE 1. (`, c) is rationalizable and truly dominated
` c r
ω 0 1 0
ω0 0 0 1
We now argue that (`, c) is truly dominated. Intuitively, since ` is strictly

dominated, the decision-maker benefits from playing a mixture of c and
r instead of ` in the first period. More formally, consider the behavioral
strategy τ given by τ1 (c|`) = τ1 (r|`) = 1/2, τ1 (r|r) = τ1 (c|c) = 1 and τ2 = τ2∗ .
The induced deviation rule is D(c`|``) = D(r`|``) = D(cc|`c) = D(rc|`c) =
D(cr|`r) = D(rr|`r) = 1/2 and D(a1 a2 |â1 â2 ) = 1 for all other profiles (a1 , a2 )
and (â1 , â2 ) such that (a1 , a2 ) = (â1 , â2 ). It is then easy to verify that (`, c) is
indeed truly dominated.
Yet, it is not surely dominated and, therefore, is rationalizable. Intuitively,

if the decision-maker learns the state after playing ` in the first period but
does not get any additional information otherwise, he has an incentive to
play `. A Bayes correlated equilibrium is as follows: the mediator recom-
mends ` at the first period, regardless of the state, and recommends c (resp.,
r) at the second period if and only if the decision-maker has been obedient
and the state is ω (resp., ω 0 ). If the decision-maker disobeys the recommen-

dation, the mediator recommends then either c or r, independently of the
state.
Proof of Theorem 2. We first rewrite the obedience constraint. Let f =

(f1 , . . . , fT ) be a feedback rule, with ft : At−1 × At−1 × Ω → A. A feedback rule
specifies a deterministic recommendation at each history of past actions,
recommendations and states. A feedback rule is a pure strategy of the me-
diator. Let F be the finite set of all feedback rules and Fω∗ be the non-empty
subset of feedback rules, which recommend a∗ on path when the state is ω.
That is, f ∗ ∈ Fω∗ if f1∗ (ω) = a∗1 and
ft∗ ((a∗1 , . . . , a∗t−1 ), ((a∗1 , . . . , a∗t−1 )), ω) = a∗t ,
for all t ≥ 2. We write F ∗ Ω for ∪ω∈Ω (Fω∗ × {ω}).
Similarly, we associate a pure strategy τ with an action rule g = (g1 , . . . , gT ),

where gt : At−1 × At−1 × A → A. The action rule g specifies a pure action
at each history of past actions and past and current recommendations. We
associate τ ∗ with the rule g ∗ , where gt∗ (at−1 , ât−1 , â) = â. Let G be the set of
action rules.
Thanks to Kuhn’s theorem, we can rewrite the condition for rationalization

as: there exists µ ∈ ∆(F × Ω) such that µ(F ∗ Ω) > 0 and
XX
u(a, ω) Pf,ω,g∗ (a, â) − Pf,ω,g (a, â) µ(f, ω)ν(g) ≥ 0,
f,ω,g a,â
for all ν ∈ ∆(G), where Pf,ω,g is the degenerate distribution over actions and
recommendations induced by the feedback rule f and the action rule g when
the state is ω.9
We first prove necessity. We prove that if a∗ is not rationalizable, then a∗ is

surely dominated. So, assume that a∗ is not rationalizable. For all µ such
9
E.g., â = (f1 (ω), f2 (f1 (ω), g1 (f1 (ω)), ω), ...) and a = (g1 (f1 (ω)), g2 (f2 (f1 (ω), g1 (f1 (ω))), ω), ...)
that µ(F ∗ Ω) > 0, there exists ν such the obedience constraint is violated,
i.e.,
XX
sup min u(a, ω) Pf,ω,g∗ (a, â) − Pf,ω,g (a, â) µ(f, ω)ν(g) < 0.
µ:µ(F ∗ Ω)>0 ν
f,ω,g a,â
Since the set of µ such that µ(F ∗ Ω) > 0 is non-empty and convex (but not
compact) and the objective is bi-linear in (µ, ν), we can apply Proposition
I.1.3 from Mertens, Sorin and Zamir, (2015, p. 6) to obtain:
XX
min sup u(a, ω) Pf,ω,g∗ (a, â) − Pf,ω,g (a, â) µ(f, ω)ν(g) < 0.
ν µ:µ(F ∗ )>0 f,ω,g a,â
Hence, there exists ν such that for all µ with µ(F ∗ Ω) > 0,
XX
u(a, ω) Pf,ω,g∗ (a, â) − Pf,ω,g (a, â) µ(f, ω)ν(g) < 0.
f,ω,g a,â
Necessity then follows by first constructing the behavioral strategy τ in-

duced by ν and, therefore, its associated deviation plan D and second con-
sidering all (f, ω).
More precisely, fix an arbitrary ω and any feedback rule f ∗ such that f1∗ (ω) =
a∗1 and
ft∗ ((a∗1 , . . . , a∗t−1 ), ((a∗1 , . . . , a∗t−1 )), ω) = a∗t ,
for all t ≥ 2. Let µ be degenerate on (f ∗ , ω). Note that (f ∗ , ω) ∈ F ∗ Ω. We

have that:
XX
u(a, ω)Pf,ω,g∗ (a, â)µ(f, ω)ν(g) = u(a∗ , ω),
f,ω,g a,â
while
XX XX
u(a, ω)Pf,ω,g (a, â)µ(f, ω)ν(g) = u(a, ω)Pf ∗ ,ω,g (a, â)ν(g)
f,ω,g a,â g a,â
X
= u(a, ω)Pf ∗ ,τ (a, â)
a,â
X
= u(a, ω)D(a|γ ∗ (a)),
a
where γ ∗ is the recommendation plan induced by γ1∗ (∅) = f1∗ (ω) and
γt∗ (a1 , . . . , at−1 ) = ft∗ ((a1 , . . . , at−1 ), (γ1 (∅), . . . , γt−1 (a1 , . . . , at−2 )), ω).
for all t ≥ 2. Since γt∗ (a∗1 , . . . , a∗t−1 ) = a∗t , this is readily seen to be equivalent
to the right-hand side of (D1).
We now prove the necessity of (D2). The arguments are nearly identical to
the above ones. Fix an arbitrary ω, an arbirary profile a and a feedback rule
f such that f1 (ω) = a1 and
ft ((a1 , . . . , at−1 ), ((a1 , . . . , at−1 )), ω) = at ,
for all t ≥ 2. If f ∈ F ∗ Ω, we can repeat the above arguments. If, however,

/ F ∗ Ω, choose µ such that µ(f, ω) = 1 − ε and µ(f ∗ , ω) = ε, where 1 > ε > 0
f∈
and f ∗ is the feedback rule defined above. From the above steps, for all
ε > 0, we have that
X X
u(a, ω) < (1 − ε) u(b, ω)D(b|γ(b)) + ε u(b, ω)D(b|γ ∗ (b)),
b b
where γ is the recommendation plan induced by γ1 (∅) = f1 (ω) and
γt (b1 , . . . , bt−1 ) = ft ((b1 , . . . , bt−1 ), (γ1 (∅), . . . , γt−1 (b1 , . . . , bt−2 )), ω).
for all t ≥ 2. Taking the limit as ε → 0 and noting that γt (a1 , . . . , at−1 ) = at ,
we obtain the condition (D2).
The proof of sufficiency is immediate and left to the reader.
To conclude, this application illustrates how we can apply our results to

derive testable implications in dynamic decision problems. We stress that
our results apply equally to dynamic games, including games with evolving
states, and thus offer a wide scope for applications.
5. A DDITIONAL E QUIVALENCE T HEOREMS AND ANOTHER APPLICATION
The objective of this section is to enrich our analysis by requiring rational

behavior on and off the equilibrium path. The main message is that Theo-
rem 1 generalizes to stronger solution concepts. All we need is a revelation
principle for these solution concepts. While the definitions of these revela-
tion principles are rather complex, they share some salient features. First,
these revelation principles require players to be obedient and truthful at
other histories than the on-path histories, but not at all histories. A player
is required to be obedient and truthful only if he has not lied in the past.
That player may have disobeyed past recommendations, however. In ad-
dition, not even all these histories are considered. A further consistency
requirement is imposed, the so-called consistency with mediation ranges.
Second, these revelation principles postulate that players assign probabil-
ity zero to the event that others have lied to the mediator. As Myerson
(1986, p. 342) put it:
This begs the question of whether we could get a larger set of

sequentially rational communication equilibria if we allowed
players to assign positive probability to the event that oth-
ers have lied to the mediator. Fortunately, by the revelation
principle, this set would not be any larger. (Myerson, 1986, p.
342)
In what follows, we do not provide the reader with a restatement of these

revelation principles–this would take too much space. We refer to Myerson
(1986) and the recent work of Sugaya and Wolitzky (2021). We start with
the concept of weak perfect Bayesian equilibrium, one of the most widely
used solution concepts in applications and, probably, the easiest refinement
to state.
5.1. Weak Perfect Bayesian Equilibrium. Throughout, we fix an expan-

sion Γπ of Γ. We denote Pσ,π (·|ht , mt , ω t ) the distribution over HM Ω induced
by the profile of behavioral strategies σ and the expansion π, given the

history (ht , mt , ω t ). The distribution Pσ,π (·|ht , mt , ω t ) is well-defined even if
(ht , mt , ω t ) has zero probability under Pσ,π , and it is equal to Pσ,π (·|ht , mt , ω t )
when Pσ,π (ht , mt , ω t ) > 0. Intuitively, this distribution represents the beliefs
an outside observer has at (ht , mt , ω t ) if it is conjectured that players con-
tinue to follow their equilibrium strategies even after deviations. We adopt
the convention that Pσ,π (h, m, ω) := Pσ,π (h, m, ω|h0 , m0 , ω 0 ). At any given
history (ht , mt , ω t ), player i’s expected payoff is
X
Ui (σ|ht , mt , ω t ) := ui (h, ω)Pσ,π (h, m, ω|ht , mt , ω t ).
h,m,ω
To complete the description, we need to specify the belief player i has at

any private history (hti , mti ). To do so, we specify a belief system β. Player i
believes that the history is (ht , mt , ω t ) with probability β(ht , mt , ω t |hti , mti ) at
the private history (hti , mti ). At the private history (hti , mti ), player i’s expected
payoff is therefore:
X
Ui (σ, β|hti , mti ) := Ui (σ|ht , mt , ω t )β(ht , mt , ω t |hti , mti ).
ht ,mt ,ω t
Definition 3 (wPBE): A profile σ of behavioral strategies is a weak perfect

Bayesian equilibrium of Γπ if there exists a belief system β on HM Ω such
that:
(i) Sequential rationality: For all t, for all i, for all (hti , mti ),
Ui (σ, β|hti , mti ) ≥ Ui ((σi0 , σ−i ), β|hti , mti ),
for all σi0 .

(ii) Belief consistency: The belief system β is consistent with σ, that is,
for all (h, m, ω) ∈ HM Ω, for all (i, t),
Pσ,π (ht , mt , ω t )
β(ht , mt , ω t |hti , mti ) = ,
Pσ,π (hti , mti )
whenever Pσ,π (hti , mti ) > 0.

We let wPBE(Γπ ) be the set of distributions over HΩ induced by the weak

perfect Bayes equilibria of Γπ .
S
As before, the objective is to characterize the set Γπ an expansion of Γ wPBE(Γπ ),
i.e., we want to characterize the distributions over the outcomes HΩ of the
base game Γ that we can induce by means of some expansion Γπ of the base
game, without any reference to particular expansions. To do so, we need to
introduce the concept of weak perfect Bayes correlated equilibrium of Γ.
Weak Perfect Bayes Correlated Equilibrium. We consider mediated

extensions M(Γ) of the game Γ, where at each stage the set of recom-
mendations made to a player may be a strict subset of the set of actions
available to the player. Formally, for each private history (hti , ât−1
i ) of past
and current signals sti , past actions at−1

i and past recommendations ât−1
i ,
Ri,t (hti , ât−1

i ) ⊆ Ai,t is the set of possible recommendations to player i. We
refer to the function Ri,t as the mediation range of player i at stage t. We

denote H (R) the set of all terminal histories consistent with the mediation
ranges in the mediated extension M(Γ), i.e., (h, ω, â) ∈ H (R) if and only if
(h, ω) ∈ HΩ and âi,t ∈ Ri,t (hti , ât−1
i ) for all i, for all t.
We denote Pµ◦τ,p (·|ht , ω t , ât ) the distribution over H (R) induced by the pro-
file of strategies τ , the recommendation kernels µ and the kernels p, given
the history (ht , ω t , ât ). At any history (ht , ω t , ât ), player i’s expected payoff is
X
Ui (µ ◦ τ |ht , ω t , ât ) := ui (h, ω)Pµ◦τ,p (h, ω, â|ht , ω t , ât ).
h,ω,â
Finally, at any private history (hti , âti ), player i’s expected payoff is:
X
Ui (µ ◦ τ, β|hti , âti ) := Ui (µ ◦ τ |ht , ω t , ât )β(ht , ω t , ât |hti , âti ),
ht ,ω t ,ât
where β is a belief system. We write Ti∗,t for the subset of action strategies
of player i, where player i is obedient up to (including) stage t. We are now
ready to define the concept of weak perfect Bayes correlated equilibrium.
Definition 4 (wPBCE): A weak perfect Bayes correlated equilibrium of Γ is

a collection of mediation ranges Ri,t : Hit × At−1
i → 2Ai,t \ {∅} for all (i, t), a col-
lection of recommendation kernels µt (ht , ω t , ât−1 ) : ×i∈I Ri,t (hti , ât−1 i ) → [0, 1],
where ât ∈×i∈I Ri,t (ht ,ât−1 ) µt (ht , ω t , ât−1 )[ât ] = 1, for all (ht , ω t , ât−1 ) in H (R)
P
i i
and a belief system β such that:
(i) Obedience: For all t, for all i, for all private histories (hti , âti ) such that
0 0
âi,t0 ∈ Ri,t0 (hti , âit −1 ) for all t0 ≤ t,
Ui (µ ◦ τ ∗ , β|hti , âti ) ≥ Ui (µ ◦ (τi , τ−i

∗
), β|hti , âti ),
for all τi ∈ Ti∗,t−1 .

(ii) Belief consistency: β is consistent with (τ ∗ , µ, p), that is, for all (h, ω, â) ∈
H (R), for all (i, t),
Pµ◦τ ∗ ,p (ht , ω t , ât )

β(ht , ω t , ât |hti , âti ) = ,
Pµ◦τ ∗ ,p (hti , âti )
whenever Pµ◦τ ∗ ,p (hti , âti ) > 0.
We let wPBCE(Γ) be the set of distributions over HΩ induced by the weak

perfect Bayes correlated equilibria of Γ.
It is worth pausing over the role of the mediation ranges. A weak perfect
Bayes correlated equilibrium constrains the mediator to only recommend
actions consistent with the mediation ranges, that is, the only recommen-
dations the mediator can make to player i are in Ri,t (hti , ât−1
i ) at history
(ht , ω t , ât−1 ).10 In addition, players must have an incentive to be obedient

at all histories consistent with the mediation ranges. The role of mediation
ranges is precisely to ensure that players can be obedient at all histories of
the mediated game. Without constraining the recommendations the media-
tor can make, it wouldn’t be possible to ensure that players are obedient at
all histories. E.g., no player would ever have an incentive to play a strictly
dominated action. An equivalent formulation is to consider weak perfect
10
This explains why the domain of µt (ht , ω t , ât−1 ) is ×i∈I Ri,t (hti , ât−1
i ) in our definition.
Bayesian equilibria of the mediated game M(Γ), where the mediator is om-
niscient and unconstrained in its recommendations, and players are obedi-
ent on path. The drawback of this alternative formulation is that players
do not have to be obedient off path and, therefore, requires to explore all
possible behaviors off path. The advantage is that no mediation ranges are
required.
With all these preliminaries done, we can now state our second equivalence
result.

[
wPBCE(Γ) = wPBE(Γπ ).
Theorem 3 states an equivalence between (i) the set of distributions over ac-
tions, base signals and states induced by all weak perfect Bayes correlated
equilibria of Γ, and (ii) the set of distributions over actions, base signals and
states we can obtain by considering all weak perfect Bayesian equilibria of
all expansions of Γ.
The logic behind Theorem 3 is identical to the the one behind Theorem 1.
We can replicate any weak perfect Bayesian equilibrium of Γπ as a weak per-
fect Bayesian equilibrium of the auxiliary mediated game M∗ (Γ) and then
invoke the revelation principle for weak perfect Bayesian equilibria, which
was recently proved by Sugaya and Wolitzky (2018, Proposition 2). More
precisely, their revelation principle states that it is without loss of general-
ity to assume that players report their private information to the mediator,
that the mediator recommends actions to the players, and that players have
an incentive to be truthful and obedient provided they have been truthful
in the past. The mediator cannot recommend actions outside the mediation
ranges and a belief system gives the player’s beliefs. Thus, unlike the clas-
sical revelation principle of Forges (1986) and Myerson (1986), players are
required to continue to be truthful and obedient even if they have disobeyed

in the past, so long as they have been truthful in the past.
We conclude with few additional remarks. First, the set wPBCE(Γ) is con-
vex.11 Second, despite its theoretical shortcomings, we have considered the
concept of weak perfect Bayesian equilibrium as our solution concept.12 We
did so for two two main reasons. First, it is simple and indeed widely used
in applications. Second, it generalizes to continuous games, a common as-
sumption in applications. In what follows, we present another solution con-
cept, which alleviates some of the theoretical shortcomings of weak perfect
Bayesian equilibrium. However, it comes at a cost: it is “harder” to state
and to use in applications.
5.2. Conditional Probability Perfect Bayesian Equilibrium. An im-

portant tool in modeling off-equilibrium path beliefs is the concept of condi-
tional probability systems (henceforth, CPS). Fix a finite non-empty set X .
A conditional probability system β on X is a function from 2X × 2X \ {∅} to
[0, 1], which satisfies three properties: for all X, Y, Z with X ⊆ X , Y ⊆ X
and ∅ =
6 Z ⊆ X,
(i) β(Z|Z) = 1 and β(X |Z) = 1,

(ii) if X ∩ Y = ∅, then β(X ∪ Y |Z) = β(X|Z) + β(Y |Z),
(iii) if X ⊆ Y ⊆ Z and Y 6= ∅, then β(X|Z) = β(X|Y )β(Y |Z).
Conditional probability systems capture the idea of “conditional beliefs”

even after zero-probability events. In particular, if X is the set of termi-
nal histories of a game, a conditional probability system induces a belief
system, i.e., a belief over histories at each information set of a player. A
conditional probability system also captures the beliefs players have about
the strategies and beliefs of others. Finally, using a conditional probability
11
See the working paper version for a proof.
12
It is well-known that weak perfect Bayesian equilibria may not be subgame perfect, may
rely on “irrational” beliefs, and may not satisfy the one-shot deviation principle.
system to represent the players’ beliefs imposes that all differences in be-
liefs come from differences in information. We refer the reader to Myerson
(1986) for more on conditional probability systems.13
We now define the concept of conditional probability perfect Bayesian equi-

librium, a concept introduced by Sugaya and Wolitzky (2021). We first give
an informal definition. A conditional probability perfect Bayesian equilib-
rium is a profile of strategies and a conditional probability system such
that (i) sequential rationality holds given the belief system induced by the
conditional probability system and (ii) the conditional probability system
is consistent with the profile of strategies and the data of the game. It is
a stronger concept than the concept of weak perfect Bayesian equilibrium
and a weaker concept than the concept of sequential equilibrium. We now
turn to a formal definition.
In what follows, we use notation, which parallel the one used in previous
definitions, and thus do not rehash formal definitions.
Definition 5 (CPPBE): A conditional probability perfect Bayesian equilib-

rium of Γπ is a profile σ of behavioral strategies and a CPS β on HM Ω, which
satisfy:
(i) Sequential rationality: For all t, for all i, for all (hti , mti ),
Ui (σ, β|hti , mti ) ≥ Ui ((σi0 , σ−i ), β|hti , mti ),
for all σi0 .

13
Myerson shows that for any conditional probability system β, there exists a sequence of
probability measures Pn on X such that (i) Pn ({x}) > 0 for all x ∈ X and (ii) β = limn Pn ,
n
that is, β(X|Y ) = limn P P(X∩Y
n (Y )
)
for all X, for all Y 6= ∅.
(ii) CPS consistency: The CPS β is consistent with (σ, p, ξ), that is, for all
(h, m, ω) ∈ HM Ω, for all (i, t),
Y
β(at |ht , mt , ω t ) = σi,t (ai,t |hti , mti ),
i∈I
β(ht+1 , ωt+1 |at , ht , mt , ω t ) = pt+1 (ht+1 , ωt+1 |at , ht , ω t ),
β(mt+1 |ht+1 , mt , ω t+1 ) = ξt+1 (mt+1 |ht+1 , mt , ω t+1 ).
We let CPPBE(Γπ ) be the set of distributions over HΩ induced by the condi-

tional probability perfect Bayesian equilibria of Γπ .
Few comments are worth making. First, to ease notation, we have written
β(at |ht , mt , ω t ) for
n o
β (h, m, ω) ∈ HM Ω : (at , ht , mt , ω t ) =(at , ht , mt , ω t )
n o
(h, m, ω) ∈ HM Ω : (ht , mt , ω t ) = (ht , mt , ω t ) .

We use similar abuse of notation throughout. Second, the consistency of the

CPS implies that
Pσ,π (ht , mt , ω t )
β(ht , mt , ω t |hti , mti ) = ,
Pσ,π (hti , mti )
whenever Pσ,π (hti , mti ) > 0. Third, a conditional probability perfect Bayesian
equilibrium is subgame perfect. Fourth, since the belief a player has is
induced by the CPS, two players with the same information have the same
belief. However, the CPS does not impose a “don’t signal what you don’t
know” condition. To do so, we would need to require the CPS to maintain
the relative likelihood of any two histories before and after players taking
actions.
S
As before, the objective is to characterize the set Γπ an expansion of Γ CPPBE(Γπ ),
i.e., we want to characterize the distributions over the outcomes HΩ of the
base game Γ that we can induce by means of some expansion Γπ of the base
game, without any reference to particular expansions. To do so, we need to
introduce the concept of sequential Bayes correlated equilibrium of Γ.
Sequential Bayes correlated equilibrium. As in the previous section,

we consider mediated extensions M(Γ) of the game Γ, where at each stage
the set of recommendations made to a player may be a strict subset of the
set of actions available to the player. We use the same notation and do not
rehash them.
A feedback rule f := (f1 , . . . , fT ) is a deterministic recommendation kernel,

which recommends the action ft (ht , ω t ) at history (ht , ω t ) ∈ H t Ωt . Note that
given f , the history (ht , ω t ) encodes the profile of recommendations ât as
(f1 (h1 , ω 1 ), f2 (h2 , ω 2 ), . . . , ft (ht , ω t )). A feedback rule f is consistent with the
mediation ranges R if fi,t (ht , ω t ) ∈ Ri,t (hti , ât−1 t t
i ) for all i, for all (h , ω ), for all
t, where ât−1 is the profile of recommendations encoded by f at (ht−1 , ω t−1 ).

We let F be the set of feedback rules and F(R) the subset of feedback rules
consistent with the mediation ranges R.
We denote Pf ◦τ,p (·|ht , ω t ) the distribution over H (R) induced by the profile of
strategies τ , the feedback rule f and the kernels p, given the history (ht , ω t ).
At any history (ht , ω t ), player i’s expected payoff is
X
Ui (f ◦ τ |ht , ω t ) := ui (h, ω)Pf ◦τ,p (h, ω|ht , ω t ),
h,ω
when the feedback rule is f . Finally, at any private history (hti , âti ), player
i’s expected payoff is:
X
Ui (τ, β|hti , âti ) := Ui (f ◦ τ |ht , ω t )β(f, ht , ω t |hti , âti ),
ht ,ω t ,f
where β is a CPS on F(R) × HΩ. Here, we write β(f, ht , ω t |hti , âti ) for:
n on o
t t t t 1 1 t t t t t
β (f , h, ω) : (f , h , ω = f, h , ω ) (f , h, ω) : (fi,1 (h , ω ), . . . , fi,t (h , ω )) = âi , hi = hi

Definition 6 (SBCE): A communication mechanism µ ∈ ∆(F) is a sequen-

tial Bayes correlated equilibrium of Γ if there exist mediation ranges R and
a conditional probability system β on F(R) × HΩ such that:
(i) Obedience: For all t, for all i, for all private histories (hti , âti ) such that
0 0
âi,t0 ∈ Ri,t0 (hti , âti −1 ) for all t0 ≤ t,
Ui (τ ∗ , β|hti , ati ) ≥ Ui ((τi , τ−i

∗
), β|hti , ati )
for all τi ∈ Ti∗,t−1 .

(ii) CPS consistency: For all f, h, ω, t,
β(f, h, ω) = µ(f )Pf ◦τ ∗ ,p (h, ω)
β(f, h, ω|(f1 , . . . , ft ), (ht , ω t )) = β(f |(f1 , . . . , ft ), (ht , ω t ))Pf ◦τ ∗ ,p (h, ω|ht , ω t ).
We let SBCE(Γ) be the set of distributions over HΩ induced by the sequen-

tial Bayes correlated equilibria of Γ. The set SBCE(Γ) is convex.
Few remarks are worth making. First, in a sequential Bayes correlated

equilibrium, players have an incentive to be obedient at all histories con-
sistent with the mediation ranges. Second, unlike previous definitions, the
definition asserts that the omniscient mediator selects a feedback rule f
with probability µ, i.e., as if the mediator chooses a mixed strategy (and
not a behavioral strategy). In addition, the conditional probability system
is required to be consistent with µ. Third, we may wonder whether an
equivalent formulation exists where the mediator chooses recommendation
kernels (µt )t (behavioral strategies) and consistency is imposed with respect
to (µt )t , as we did in the definition of a weak perfect Bayes correlated equi-
librium. As Sugaya and Wolitzky (2021) show, the answer is unfortunately
no. Intuitively, the current formulation allows more flexibility in choosing
beliefs, which is needed for a revelation principle to hold. Lastly, sequential
Bayes correlated equilibria are sequential communication equilibria (Myer-
son, 1986) of mediated games, where the mediator is omniscient.14 Read-
ers should not confuse these concepts with that of sequential equilibrium,
which does not involve a device such as a mediator.
14
Sequential Bayes correlated equilibria are the subsets of Bayes correlated equilibria,
where the mediator never recommends co-dominated actions, a generalization of the con-
cept of dominance. We refer the reader to Myerson (1986) for more detail.

[
SBCE(Γ) = CPPBE(Γπ ).
Theorem 4 states an equivalence between (i) the set of distributions over

actions, base signals and states induced by all sequential Bayes correlated
equilibria of Γ, and (ii) the set of distributions over actions, base signals
and states we can obtain by considering all conditional probability perfect
Bayesian equilibria of all expansions of Γ. The logic behind Theorem 4 and
its proof are the same as in previous sections. (More precisely, the revela-
tion principle we invoke is stated in Proposition 8 in Sugaya and Wolitzky
(2021).)
5.3. Application II: Bilateral Bargaining. We consider a variation on

the work of Bergemann, Brooks and Morris (2013). There are one buyer
and one seller. The seller makes an offer a1 ∈ A1 ⊂ R+ to the buyer, who
observes the offer and either accepts (a2 = 1) or rejects (a2 = 0) it. If the
buyer accepts the offer a1 , the payoff to the buyer is ω − a1 , while the payoff
to the seller is a1 , with ω being the buyer’s valuation (the payoff-relevant
state). We assume that ω ∈ Ω ⊂ R++ . If the buyer rejects the offer, the
payoff to both the seller and the buyer is normalized to zero. The buyer and
the seller are symmetrically informed and believe that the state is ω with
probability p(ω) > 0. We assume that the set of offers the seller can make
is finite, but as fine as needed. For future reference, we write ωL for the
lowest state, ωL− for the largest offer a1 strictly smaller than ωL , and ωH for
the highest state.
This model differs from Bergemann, Brooks and Morris (2013) in one im-
portant aspect. In our model, both the seller and the buyer have no initial
private information about the state, while Bergemann, Brooks and Mor-
ris assume that the buyer is privately informed of the state ω. The base
game of Bergemann, Brooks and Morris thus corresponds to a particular
expansion of our base game. Similarly, Roesler and Szentes (2017) consider
all information structures, where the buyer has some signals about his own
valuation (and the seller is uninformed).15 Unlike these papers, we consider
all information structures. In particular, the information the buyer receives
may depend on the information the seller has received as well as the offer
made. In addition, the seller can be better informed than the buyer in our
model.
We characterize the set of sequential Bayes correlated equilibria. A commu-

nication system µ is a sequential Bayes correlated equilibrium if there exist
mediation ranges (R1 , R2 ) and a conditional probability system β, which
jointly satisfy the following constraints. First, if the omniscient mediator
recommends f1 (ω) ∈ R1 to the seller, the seller must have an incentive to be
obedient, i.e.,
X X
f1 (ω)f2 (f1 (ω), ω)β(f, ω|f1 (ω)) ≥ a1 f2 (a1 , ω)β(f, ω|f1 (ω))
f,ω f,ω
for all a1 . Second, if the offer made to the buyer is a1 and the mediator rec-
ommends f2 (a1 , ω) ∈ R2 (a1 ) to the buyer, the buyer must have an incentive
to be obedient, i.e.,
X X
(ω − a1 )f2 (a1 , ω)β(f, ω|a1 , f2 (a1 , ω)) ≥ (ω − a1 )(1 − f2 (a1 , ω))β(f, ω|a1 , f2 (a1 , ω)).
f,ω f,ω
Third, the conditional probability system must be consistent, that is, for all
f ∈ F(R), for all a1 , a2 , ω,
β(f, a1 , a2 , ω) = µ(f )p(ω)1{f1 (ω) = a1 , f2 (f1 (ω), ω) = a2 },
β(f, a1 , a2 , ω|f1 , a1 , ω) = β(f |f1 , a1 , ω)1{(f2 (a1 , ω), ω) = a2 }.
There are immediate bounds on the equilibrium payoffs: the sum of the
P
buyer and seller’s payoffs is bounded from above by E(ω) = ω p(ω)ω, the
15
In related work, Kartik and Zhang (2019) characterize all wPBE of the bilateral bargain-
ing model with interdependent values as one varies the information structure. They do
so under three different scenarios regarding base signals: one like ours, one like that in
Bergemann, Brooks and Morris, and one where the buyer is better informed than the seller.
Under the first scenario, they also find that the set of implementable payoffs is the one that
satisfies the immediate bounds on equilibrium payoffs we discuss prior to Proposition 1.
buyer’s payoff is bounded from below by 0, and the seller’s payoff is bounded
from below by ωL− . The following proposition states that there are, in fact,
no other restrictions on equilibrium payoffs.
Proposition 1. The set of sequential Bayes correlated equilibrium payoffs

is
co {(0, ωL− ), (0, E(ω)), (E(ω) − ωL− , ωL− )}.
The set of equilibrium payoffs is depicted in the figure below.
seller’s payoff
(0, E(ω))
(0, ωL− ) (E(ω) − ωL− , ωL− )

buyer’s payoff
F IGURE 3. Payoffs at all sequential Bayes correlated equilibria
Notice that the set of BCE here differs from the one in Bergemann, Brooks
and Morris (2013) in that the lowest seller payoff is ωL− , and not the monop-
oly profit.
We prove this proposition in what follows. As a preliminary observation,

note that the conditional probability system puts no restriction on the buyer’s
P
beliefs after observing an off-path offer a1 , i.e., an offer such that f,ω µ(f )p(ω)1{f1 (ω) =
a1 } = 0. To see this, for any conditional probability system, β(a1 , ω) =
β(ω, a1 |a1 )β(a1 ). Moreover, from the consistency of β, we have that β(a1 , ω) =
P P
f β(f, a1 , ω) = f µ(f )p(ω)1{f1 (ω) = a1 } = 0. Since β(a1 ) = 0, β(ω, a1 |a1 ) is
arbitrary and, thus, we can assume that the buyer believes that the state
is ωL with probability one. We refer to those beliefs as the most pessimistic
beliefs. Similarly, there are no restrictions on the buyer’s beliefs after ob-
serving an off-path offer a1 and a recommendation f2 (a1 , ω).
We are now ready to state how to obtain the payoff profile (E(ω) − ωL− , ωL− ).
We first start with an informal description. The mediator recommends the
seller to offer ωL− , regardless of the state. If the offer ωL− is made, the me-
diator recommends the buyer to accept, regardless of the state. If any offer
a1 > ωL− is made, the mediator recommends the buyer to reject the offer,
regardless of the state. Since any such offer is off-path, the buyer has an
incentive to be obedient when he believes that the state is ωL with proba-
bility one. As we have just argued, we can choose a well-defined conditional
probability system capturing such beliefs. Finally, if any offer a1 < ωL− is
made, the mediator recommends the buyer to accept, regardless of the state.
Formally, the communication system puts probability one to f , given by
f1 (ω) = ωL− , f2 (a1 , ω) = 0 if a1 > ωL− and f2 (a1 , ω) = 1 if a1 ≤ ωL− for all ω. The
mediation ranges are R1 = {ωL− }, R2 (a1 ) = {1} if a1 < ωL , R2 (ωL ) ⊆ {0, 1},
and R2 (a1 ) = {0} if a1 > ωL .
We now turn our attention to the two other payoff profiles (0, E(ω)) and
(0, ωL− ). The profile (0, E(ω)) corresponds to full surplus extraction, which
can be obtained with f1 (ω) = ω for all ω and f2 (a1 , ω) = 1 whenever a1 ≤ ω
(and zero, otherwise). The mediation ranges are R1 = Ω, R2 (a1 ) = {0} if
a1 > ωH , R2 (a1 ) = {1} if a1 < ωL , and R2 (a1 ) = {0, 1} if a1 ∈ Ω.
Lastly, when E(ω) ∈ A1 (which we assume), the profile (0, ωL− ) is imple-
mentable as follows. Consider two feedback rules f and f 0 such that for
all ω, f1 (ω) = f10 (ω) = E(ω), f2 (a1 , ω) = f20 (a1 , ω) = 0 if a1 > E(ω), f2 (a1 , ω) =
f20 (a1 , ω) = 1 if a1 < E(ω), f2 (E(ω), ω) = 1 while f20 (E(ω), ω) = 0. Assume that
µ(f ) = ωL− /E(ω), µ(f 0 ) = 1 − µ(f ), and that R1 = {E(ω)}, R2 (a1 ) = {1} if
a1 < ωL , R2 (a1 ) = {0, 1} if a1 = E(ω), and R2 (a1 ) = {0}, otherwise. In effect,
the mediator recommends the seller to offer E(ω), regardless of the state,
and the buyer to accept that offer with probability ωL− /E(ω), on path. Off-
path, we again use the most pessimistic beliefs to give the seller a payoff
of zero, if he deviates. To complete the proof of Proposition 1, it is enough
to invoke the bounds on the payoff profiles and the convexity of the set of
sequential Bayes correlated equilibrium payoffs.
6. C ONCLUSION
This paper generalizes the concept of Bayes correlated equilibrium to multi-

stage games and offers two applications, which are suggestive of the useful-
ness of our characterization results. The main contribution is methodologi-
cal.
The reader may wonder why we have not considered the concept of sequen-
tial equilibrium. The main reason is that a revelation principle does not
hold for this concept. To be more precise, Sugaya and Wolitzky (2021) show
that the set of sequential communication equilibria of a multi-stage game
characterizes the set of equilibrium distributions we can obtain by consid-
ering all mediated extensions of the multi-stage game, where the solution
concept is sequential equilibrium. However, their definition of a sequential
equilibrium treats the mediator as a player and, thus, allows for the media-
tor to tremble. When we consider an expansion and its emulation by a me-
diator with the mediated game M∗ (Γ), players do not expect the mediator
to tremble. If a player observes an unexpected additional signal, that player
must believe with probability one that one of his opponents has deviated.
He cannot believe that none of his opponents deviated, but the mediator
did. This would be inconsistent with the expansion being the game actually
played. Extending the analysis to other solution concepts such as sequen-
tial equilibrium or rationalizability or to general extensive-form games is
challenging and left for future research.
A PPENDICES
S
A. Proof of Theorem 1. (⇐ .) We first prove that Γπ an expansion of Γ BN E(Γπ ) ⊆
BCE(Γ). Throughout, we fix an expansion Γπ of Γ. Recall that there exist
kernels (ξt )t such that:
πt+1 (ht+1 , mt+1 , ωt+1 |at , ht , mt , ω t ) = ξt+1 (mt+1 |ht+1 , mt , ω t+1 )pt+1 (ht+1 , ωt+1 |at , ht , ω t ),
for all (ht+1 , mt+1 , ω t+1 ), for all t.
Let σ ∗ be a Bayes-Nash equilibrium of Γπ . We now construct an auxiliary

mediated game M∗ (Γ), which emulates the distribution Pσ∗ ,π as an equilib-
rium distribution.
The game M∗ (Γ) has one additional player, labelled player 0, and a (Forges-
Myerson) mediator. Player 0 is a dummy player: his payoff is identically
zero.
The game unfolds as follows: At stage t = 1,
- Nature draws (h1 , ω1 ) with probability p1 (h1 , ω1 ).

- Player i ∈ I observes the signal hi,1 and player 0 observes (h1 , ω1 ).
- Player 0 reports (ĥ1 , ω̂1 ) to the mediator. All other players do not
make reports.
- The mediator draws the message m1 with probability ξ1 (m1 |ĥ1 , ω̂1 )
and sends the message mi,1 to player i. Player 0 does not receive a
message.
- Player i takes an action ai,1 . Player 0 does not take an action.
Consider now a history (at−1 , ht−1 , ω t−1 ) of past actions, signals and states
and a history ((ĥt−1 , ω̂ t−1 ), mt−1 ) of reports and messages. At stage t:
- Nature draws (ht , ωt ) with probability pt (ht , ωt |at−1 , ht−1 , ω t−1 ).

- Player i ∈ I observes the signal hi,t and player 0 observes (ht , ωt ).
- Player 0 reports (ĥt , ω̂t ) to the mediator. All other players do not
make reports.
- The mediator draws the message mt with probability ξt (mt |ĥt , mt−1 , ω̂ t )
and sends the message mi,t to player i. Player 0 does not receive a
message.
- Player i takes an action ai,t . Player 0 does not take an action.
In the above description, when we say that player i does not make a report,
we implicitly assume that the set of reports player i can make to the medi-
ator is a singleton. Similarly, when we say that player 0 does not take an
action. In the rest of the proof, we omit these trivial reports and actions.
We restrict attention to the histories of M∗ (Γ), where (h, ω) ∈ HΩ. At stage

t, player i’s private history is (hti , mti ), which is also player i’s private history
in Γπ . In addition, any private history (hi,t , mi,t ) in M∗ (Γ) is also a private
history in Γπ . Thus, σi∗ is a well-defined strategy for player i in M∗ (Γ).
Moreover, if player 0 truthfully reports his private information (ht , ωt ) at all
histories ((ht , ωt ), (ht−1 , ω t−1 ), (ĥt−1 , ω̂ t−1 )), the conditional probability of the
message mt is the same as in Γπ . It follows immediately that σ ∗ together
with the truthful strategy for player 0 is a Bayes-Nash equilibrium of the
auxiliary mediated game M∗ (Γ).
From the revelation principle of Forges (1986) and Myerson (1986), there
exists a canonical equilibrium µ, where the mediator recommends actions
and players are truthful and obedient, provided they have been in the past.
At truthful histories, the mediator recommends ât with probability
µt (ât | (ht , ω t ), (ht1 , . . . , htn ), ât−1

|{z} ).
| {z } | {z }
player 0 players in I past recommendations
It is then routine to verify that we have a Bayes-correlated equilibrium with

the recommendation kernel µt given by
µt (ât |ht , ω t , ât−1 ) := µt (ât |(ht , ω t ), (ht1 , . . . , htn ), ât−1 ),
for all (ht , ω t , ât−1 ) for all t.

S
(⇒). We now prove that BCE(Γ) ⊆ Γπ an expansion of Γ BN E(Γπ ).
Let µ be a Bayes correlated equilibrium with distribution Pµ◦τ ∗ ,p . We now

construct an expansion Γπ and a Bayes-Nash equilibrium σ ∗ of Γπ , with the
property that margHΩ Pσ∗ ,π = margHΩ Pµ◦τ ∗ ,p .
The expansion is as follows. Let Mi,t = Ai,t for all (i, t),
π1 (h1 , m1 , ω1 ) = p1 (h1 , ω1 )µ1 (â1 |h1 , ω1 ),
with m1 = â1 , for all (h1 , m1 , ω1 ), and
πt+1 (ht+1 , mt+1 , ωt+1 |at , ht , mt , ω t ) = pt+1 (ht+1 , ωt+1 |at , ht , ω t )µt+1 (ât+1 |ht+1 , ω t+1 , ât ),
with (mt , mt+1 ) = (ât , ât+1 ), for all (at , ht , mt , ω t , ht+1 , mt+1 , ωt+1 ). Clearly, the
expansion is well-defined: ξ1 (m1 |h1 , ω1 ) = µ1 (â1 |h1 , ω1 ) with m1 = â1 , and,
for t > 1, ξt+1 (mt+1 |ht+1 , mt , ω t+1 ) = µt+1 (ât+1 |ht+1 , ω t+1 , ât ) with (mt , mt+1 ) =
(ât , ât+1 ).
By construction, any strategy τt : H t × At → ∆(At ) of M(Γ) is equivalent to

a strategy σt : H t × M t → ∆(At ) of Γπ , i.e., σt (at |ht , mt ) := ×i σi,t (ai,t |hti , mti ) =
×i τi,t (ai,t |hti , âti ) with mt = ât , with the property that Pσ,π (ht , mt , ω t ) = Pµ◦τ,p (ht , ât , ω t )
when mt = ât , for all (ht , mt , ω t ), for all t.
To see this last point, note that the definition of π1 is clearly equivalent
to Pσ,π (h1 , m1 , ω1 ) = Pµ◦τ,p (h1 , ω1 , â1 ) with m1 = â1 , for all (h1 , m1 , ω1 ). By
induction, assume that Pσ,π (ht , mt , ω t ) = Pµ◦τ,p (ht , ω t , ât ) with mt = ât , for all
(ht , mt , ω t ). We now compute the probability of (ht+1 , mt+1 , ω t+1 ). We have
that
Pσ,π (ht+1 , mt+1 , ω t+1 ) = Pσ,π (ht+1 , mt+1 , ωt+1 |ht , mt , ω t )Pσ,π (ht , mt , ω t )
= πt+1 (ht+1 , mt+1 , ωt+1 |at , ht , mt , ω t )σt (at |ht , mt )Pσ,π (ht , mt , ω t )
= pt+1 (ht+1 , ωt+1 |at , ht , ω t )µt+1 (ât+1 |ht+1 , ω t+1 , ât )τt (at |ht , ât )
×Pµ◦τ,p (ht , ât , ω t )
= Pµ◦τ,p (ht+1 , ω t+1 , ât+1 ),
with ât+1 = mt+1 . Finally, since µ is a Bayes correlated equilibrium of M(Γ),

the strategy σ ∗ ≡ τ ∗ is a Bayes-Nash equilibrium of Γπ and, thus,
[
BCE(Γ) ⊆ BN E(Γπ ).
This completes the proof.

B. Proof of Theorem 3. The proof is nearly identical to the proof of The-

orem 1 and is, therefore, omitted. We only sketch the minor differences.
(⇐.) Fix an expansion Γπ and a weak perfect Bayesian equilibrium (σ ∗ , β)

of Γπ . We need to construct a weak perfect Bayesian equilibrium of the
auxiliary game M∗ (Γ), which replicates the distribution Pσ,π . To do so, we
define a belief system β ∗ of the auxiliary game M∗ (Γ) as follows:
β ∗ (ht , mt , ω t , (ht , ω t ), (ht , ω t )|hti , mti ) := β(ht , mt , ω t |hti , mti )
for all (ht , mt , ω t ), for all (i, t). Note that the above formulation implies that
player i believes with probability 1 that player 0 truthfully report (ht , ω t ).
(In M∗ (Γ), player i also has beliefs about the signals (ht , ω t ) player 0 receives
and the reports (ĥt , ω̂ t ) by player 0 to the mediator.) It is immediate to verify
that (σ0∗ , σ ∗ , β ∗ ) is a weak perfect Bayesian equilibrium of M∗ (Γ), where σ0∗ is
the truthful reporting strategy of player 0. The proof then follows from the
revelation principle for weak perfect Bayesian equilibrium. Three remarks
are worth making. First, the mediation ranges and the belief system come
from the revelation principle–the revelation principle precisely states the
existence of mediation ranges and belief system such that players have an
incentive to be obedient and truthful, provided they have been truthful in
the past. Second, since M∗ (Γ) is a mediated extension of Γ, it is a common
belief that states and base signals evolve according to (pt )t . At the truthful
histories of the direct mediated extension of Γ, players have an incentive to
be obedient at all recommendations consistent with the mediation ranges,
the mediator is omniscient and beliefs are as in the base game Γ. (Beliefs
of player i about the action of player 0 are trivial–player 0 has no actions.)
Third, the revelation principle asserts that players assign probability zero
to the event others have lied to the mediator, hence we can use the belief β
such that:
β(ht , ω t , ât |hti , âti ) := β ∗∗ (ht , ω t , ât , (htj )j , (ht , ω t )|hti , hti , âti ).
for all ht , ω t , ât , t, to sustain obedience, where β ∗∗ is the belief system inher-
ited from the revelation principle.
We therefore have a weak perfect Bayes correlated equilibrium.
(⇒.) We construct the expansions as in the the proof of Theorem 1, i.e.,

defining the additional signals as the recommendations. Since the addi-
tional signals player i can receive are the recommendations, player i can
only receive additional signals consistent with the mediation ranges. Thus,
we can use the belief system of the weak perfect Bayes correlated equilib-
rium to construct the weak perfect Bayesian equilibrium of Γπ .
C. Proof of Theorem 4. The proof is yet again nearly identical to the proof
of Theorem 1. We only sketch the main differences.
(⇐). Fix an expansion Γπ and a conditional probability perfect Bayesian

equilibrium (σ ∗ , β) of Γπ . As in the previous proofs, we construct a condi-
tional probability perfect Bayesian equilibrium of the mediated game M∗ (Γ),
which replicates the distribution Pσ∗ ,π . As in the proof of Theorem 3, we con-
struct a conditional probability system β ∗ of the mediated game M∗ (Γ) from
the conditional probability system β of the game Γπ such that ((σ0∗ , σ ∗ ), β ∗ )
is a conditional probability perfect Bayesian equilibrium of M∗ (Γ), with
player 0, the dummy player, truthfully reporting his private information
(ht , ω t ) at each stage t. Since β is a conditional probability system, there
exists a sequence β n of fully supported probabilities such that β(X|Y ) =
β n (X∩Y )
limn β n (Y )
for all X and all non-empty Y . Consider now the sequence of
fully supported kernels γ n : HM Ω → ∆(HΩ × HΩ), where γ n converges
to γ((h, ω), (h, ω)|(h, m, ω)) = 1 for all (h, m, ω). The interpretation is that
player 0 learns and truthfully report (h, ω), when the profile of actions, sig-
nals, and states is (h, m, ω). Let β ∗ be the CPS resulting from taking the
limit of β n × γ n . By construction, β ∗ (ht , mt , ω t |hti , mti ) = β(ht , mt , ω t |hti , mti ),
so that σi∗ remains sequentially rational for player i. Moreover, the newly
constructed conditional probability system is consistent with the kernels p
and ξ. The rest of the proof follows from the revelation principle.
(⇒). Let (µ, R, β) be a sequential Bayes correlated equilibrium. The differ-

ence with the previous proofs is that the definition of a sequential Bayes cor-
related equilibrium does not specify recommendation kernels (µ1 , . . . , µT ),
which can then be used as expansions. However, as in the proof of Kuhn’s
theorem, we can construct such recommendation kernels from µ.
The construction is iterative. In the sequel, we slightly abuse notation and

write (µ1 , . . . , µT ) for the kernels. For all (h1 , ω 1 , â0 ) such that p1 (h1 , ω1 ) > 0,
X
µ1 (â1 |â0 , h1 , ω 1 ) := µ(f )1{f1 (h1 , ω 1 ) = â1 }.
f
Note that µ1 (â1 |h1 , ω 1 , â0 ) > 0 and µ1 (â1 |h1 , ω 1 , â0 ) = 1. The kernel µ1 is
P
â1
thus well-defined. (Recall that (h1 , ω 1 ) = (h1 , ω1 ) and that â0 is a singleton.)
We proceed iteratively. For all (ht , ω t , ât−1 ) such that
p1 (h1 , ω1 )µ1 (â1 |â0 , h1 , ω 1 ) × · · · ×
pt−1 (ht−1 , ωt−1 |at−1 , ht−2 , ω t−2 )µt−1 (ât−1 |ât−2 , ht−1 , ω t−1 t)pt (ht , ωt |at , ht−1 , ω t−1 ) > 0
for some (a1 , . . . , at ),
µ(f )1{f1 (h1 , ω 1 ) = â1 , . . . , ft (ht , ω t ) = ât }

P
t−1 t t f
µt (ât |â , h , ω ) := P 1 1 t−1 , ω t−1 ) = â
.
f µ(f )1{f1 (h , ω ) = â1 , . . . , ft−1 (h t−1 }
It is immediate to verify that the kernel is well-defined.
Two remarks are in order. First, since we consider histories (h, ω) ∈ HΩ, we
already have that
p1 (h1 , ω1 )µ1 (â1 |â0 , h1 , ω 1 )×· · ·×pt−1 (ht−1 , ωt−1 |at−1 , ht−2 , ω t−2 )pt (ht , ωt |at , ht−1 , ω t−1 ) > 0
for some (a1 , . . . , at ). Second, since we only consider feedback rules in the
support of µ, all the recommendations with positive probabilities are con-
sistent with the mediation ranges. Hence, players are obedient at these
recommendations.
We define the conditional probability system on HΩA as

X
β(h, ω, â) = β(f, h, ω)1{f (h, ω) = â}.
f
To complete the proof, we repeat the same steps as in the proof of Theorem
1, that is, the additional messages are the recommendations, the kernels
(ξ1 , . . . , ξT ) are the recommendation kernels (µ1 , . . . , µT ), and the conditional
probability system is the one on HΩA defined above. Since we consider
the restriction to recommendations with positive probabilities, the recom-
mendations are consistent with the mediation ranges and players have an
incentive to play according to their signals.
R EFERENCES
[1] Bergemann, Dirk and Stephen Morris, “Bayes Correlated Equilibrium and The Com-
parison of Information Structures in Games,” 2016, Theoretical Economics, 11, pp.
487-522.
[2] Bergemann, Dirk, Benjamin Brooks and Stephen Morris, “The Limits of Price Dis-
crimination,” 2015, American Economic Review, 105, pp. 921-957.
[3] Bergemann, Dirk, Benjamin Brooks, and Stephen Morris, “Counterfactuals with La-
tent Information,” 2019, Cowles Foundation Discussion Paper 2162.
[4] Calzolari, Giacomo and Alessandro Pavan, “Monopoly with Resale,” 2006, Rand Jour-
nal of Economics, 37, pp. 362-375.
[5] de Oliveira, and Rohit Lamba, “Rationalizing Dynamic Choices,” 2019, mimeo.
[6] Doval, Laura and Jeffrey C., Ely, “Sequential Information Design,” 2020, Economet-
rica, 86, pp. 2575-2608.
[7] Dworczak, Piotr, “Mechanism Design with Aftermarkets: Cutoff Mechanisms,” 2017,
mimeo.
[8] Ely, Jeffrey C., “Beeps.” 2017, American Economic Review, 107, pp. 31-53.
[9] Forges, Françoise, “Correlated Equilibria in a Class of Repeated Games with Incom-
plete Information,” 1985, International Journal of Game Theory, 14(3), pp. 29-150.
[10] Forges, Françoise, “An Approach to Communication Equilibria,” 1986, Econometrica,
54, pp. 1375-1385.
[11] Forges, Françoise, “Five Legitimate Definitions of Correlated Equilibrium in Games
with Incomplete Information,” 1993, Theory and Decision, 35, pp. 277-310.
[12] Forges, Françoise, “Correlated Equilibrium in Games with Incomplete Information

Revisited, 2006, Theory and Decision, 61, pp. 329-344.
[13] Gualdani, Christina and Sruti Sinha, “Identification in Discrete Choice Models with
Imperfect Information,” 2021, mimeo.
[14] Giovannoni, Francesco, and Miltiadis Makris, “Reputational Bidding,” 2014, Interna-
tional Economic Review, 55, pp. 693-710.
[15] Forges, Françoise, “Correlated Equilibria in a Class of Repeated Games with Incom-
plete Information,” 1985, International Journal of Game Theory, 14(3), pp. 29-150.
[16] Forges, Françoise, “An Approach to Communication Equilibria,” 1986, Econometrica,
54, pp. 1375-1385.
[17] Forges, Françoise, “Five Legitimate Definitions of Correlated Equilibrium in Games
with Incomplete Information,” 1993, Theory and Decision, 35, pp. 277-310.
[18] Forges, Françoise, “Correlated Equilibrium in Games with Incomplete Information
Revisited, 2006, Theory and Decision, 61, pp. 329-344.
[19] Magnolfi, L., and C. Roncoroni, “Estimation of Discrete Games with Weak Assump-
tions on Information,” 2017, mimeo.
[20] Mertens, Jean-François, and Sylvain Sorin and Shmuel Zamir, “Repeated Games,”
2015, Cambridge University Press.
[21] Myerson, Roger, “Multistage Games with Communication,” 1986, Econometrica, 54,
pp. 323-358.
[22] Penta, Antonio, “Robust Dynamic Implementation,” 2015, Journal of Economic The-
ory, 160, pp. 280-316.
[23] Penta, Antonio, and Peio Zuazo-Garin, “Rationalizability, Observability and Common
Knowledge,” Review of Economic Studies, forthcoming.
[24] Renault, Jérôme, Elon Solan, and Nicolas Vieille, “Optimal Dynamic Information Pro-
vision,” 2017, Games and Economic Behavior, 104, pp. 329-349.
[25] Roesler, Anne-Katrin, and Balász Szentes, “Buyer-optimal Learning and Monopoly
Pricing,” 2017, American Economic Review, 107, pp. 2072-2080.
[26] Syrgkanis, Vasilis, Elie Tamer, and Juba Ziani, “Inference on Auctions with Weak
Assumptions on Information,” 2018, arXiv:1710.03830.
[27] Solan, Elon, “Characterization of Correlated Equilibria in Stochastic Games”, 2001,
International Journal of Game Theory, 30, pp. 1259-277.
[28] Sugaya, Takuo, and Alexander Wolitzky, “Revelation Principles in Multistage Games,”
2021, The Review of Economic Studies, 88, pp. 1503-1540.
[29] Sugaya, Takuo, and Alexander Wolitzky, “Revelation Principles in Multistage Games,”
2018, mimeo.
[30] von Stengel, Bernhard, and Françoise Forges, “Extensive form Correlated Equilib-
rium: Definition and Computational Complexity,” 2008, Mathematics of Operations
Research, 33, pp. 1002-1022.
M ILTIADIS M AKRIS, D EPARTMENT OF E CONOMICS, U NIVERSITY OF E SSEX , UK
E-mail address: mmakris.econ(at)gmail.com
L UDOVIC R ENOU, Q UEEN M ARY U NIVERSITY OF L ONDON , CEPR AND U NIVERSITY OF

A DELAIDE , M ILES E ND, E1 4NS, L ONDON, UK
E-mail address: lrenou.econ(at)gmail.com

Information Design Multistage Game

Uploaded by

Copyright:

Available Formats

Information Design Multistage Game

Uploaded by

Document Information

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

Information Design Multistage Game

Uploaded by

Copyright:

Available Formats

INFORMATION DESIGN IN MULTI-STAGE GAMES

MILTIADIS MAKRIS AND LUDOVIC RENOU

A BSTRACT. This paper generalizes the concept of Bayes correlated equilib-

K EYWORDS: Multi-stage games, information design, communication equi-

JEL C LASSIFICATION: C73, D82.

Date: September 29, 2022.

This paper generalizes the concept of Bayes correlated equilibrium (Berge-

Consider an analyst, who postulates a multi-stage game, which we call the

As an example, consider the refinancing operations of central banks. Typi-

Our main contribution is methodological. We derive several generaliza-

At a Bayes correlated equilibrium of the base game, at each stage, an “om-

that the “omniscient” mediator is a metaphor, an abstract entity, which only

Finally, we provide two illustrations of the broad applicability of our results.

The second insight is that our characterization of Bayes correlated equilib-

Finally, this paper contributes to the literature on correlated equilibrium

(2, 2) (0, 1) (3, 0) (1, 1)

F IGURE 1. The base game

To address that question, we generalize the work of Bergemann and Mor-

Thus, a naive idea is to apply the concept of Bayes correlated equilibrium to

Our approach is to have the omniscient mediator recommending actions to

If player 2’s recommendation is L, he believes that player 1 has played T

We now argue that no information structures can give player 1 a payoff

As a final observation, consider the alternative base game in Figure 2,

(2, 2) (3, 0) (0, 1) (1, 1)

F IGURE 2. An alternative base game: reversing the order of play.

3. M ULTI - STAGE G AMES AND E XPANSIONS

We denote p1 (h1 , ω1 ) the joint probability of (h1 , ω1 ) at the beginning of the

We denote HΩ the subset of ×Tt=1

3.2. Expansions. In an expansion of the base game, at each stage, play-

Formally, an expansion is a collection of sets of additional private signals

ξt (mt |ht , mt−1 , ω t )

Together with the base game Γ, an expansion (M, ξ) induces a multi-stage

“expansion” to refer to the collection of additional signals and kernels (M, ξ)

We denote HM Ω the set of all terminal histories, with (h, m, ω) ∈ HM Ω if

In closing, it is worth noting that an expansion ξ induces a collection of

We call this property consistency. As explained in the introduction, in static

4. A F IRST E QUIVALENCE T HEOREM AND AN A PPLICATION

This section contains our first characterization theorem–other character-

4.1. A First Equivalence Theorem. We first define the concepts of Bayes-

Definition 1 (BNE): A profile σ of behavioral strategies is a Bayes-Nash

We let BN E(Γπ ) be the set of distributions over HΩ induced by the Bayes-

any reference to particular expansions. To do so, we need to introduce the

Bayes correlated equilibrium. Consider the following mediated exten-

Definition 2 (BCE): A Bayes correlated equilibrium of Γ is a collection

We let BCE(Γ) be the set of distributions over HΩ induced by the Bayes

It is instructive to compare the concept of Bayes correlated equilibrium and

Γπ of Γ since players have additional signals in Γπ , while the omniscient

Theorem 1. We have the following equivalence:

Theorem 1 states an equivalence between (i) the set of distributions over

Theorem 1 generalizes the work of BM to multi-stage games. As the intro-

Σ is the set of strategies of Γ), which satisfies

for all σi0 , σi , i. In words, players must have an incentive to be obedient.

Clearly, any distribution induced by an “ex-ante” Bayes-correlated equilib-

Conversely, consider a Bayes-correlated equilibrium µ of such multi-stage

The idea is to associate with (µt (·|·, ·, s, ω))t an outcome-equivalent distribu-

µ(σ [f,s] |s, ω) = ××

then it is routine to verify that the players have an incentive to be obedi-

- Nature draws (ht , ωt ) with probability pt (ht , ωt |at−1 , ht−1 , ω t−1 ).

If player 0 is truthful and each player i ∈ I follows σi , we have a Bayes-Nash

Before applying Theorem 1, two additional remarks are worth making.