Communication and Money in Repeated Games on

Networks
Alexander Wolitzky
Stanford University and Microsoft Research
June 16, 2012
Abstract
A key obstacle to coordination and cooperation in many networked environments
is that behavior in each bilateral relationship is not observable to individuals outside
that relationship: that is, information is local. This paper investigates when players
can use communication to replicate any outcome that would have been sustainable
were this information public. A benchmark result is that public information can
always be replicated if public communication is possible, but that if only local com-
munication is possible then public information can only be replicated if the network is
2-connected. The main result is that public information can always be replicated with
local communication if the players also have access to at money, which is modeled
as undierentiated tokens that can be freely transferred among linked players. It is
shown that such money is essential if the network contains a nice subnetwork: in
leading applications, this condition reduces to the property that the network contains
a subtree of size at least three.
Preliminary. Comments Welcome.
1 Introduction
Consider three people1, 2, and 3arranged on a line: 1 and 2 have a relationship, and 2
and 3 have a relationship, but 1 and 3 do not. In this situation, 1 and 3 might hope to keep
2 on good behavior by threatening community enforcement: if 2 cheats 1, then 3 cheats
2. But if 2 cheats 1, how does 3 nd out? She doesnt have a relationship with 1, and 2
clearly cant be trusted to tell her. So the group has a problem.
Their problem is quite similar to the classical problem of trade absent a double coincidence
of wants. Suppose 1 is a university, 2 is an economist, and 3 is a barber. Again, community
enforcement looks promising: if the economist doesnt give a lecture, she doesnt get a
haircut. But the barber doesnt know if she gave the lecture, and he cant trust her to
reveal that information to him.
1
In both of these examples, the obstacle to sustaining cooperation is that information
about individuals past behavior in a bilateral relationship is local : it is common knowl-
edge within the relationship, but is not observable to outsiders. In addition, letting the
players communicate locally does not allow them to sustain certain outcomeslike (lec-
ture,haircut)that would have been sustainable if this information were public to all players.
In the language of this paper, local communication does not replicate public information in
these examples.
The goal of this paper is to compare dierent communication technologies in terms of
their ability to replicate public information. A benchmark result is that public information
can always be replicated if public cheap talk is possible, but that if only private cheap
talk with ones neighbors is possible then public information can be replicated only if the
network is 2-connected (i.e., it remains connected after any node is removed).
2
The intuition
is simple: With public cheap talk, players can be asked to report their observations, and can
be punished if their reports conict.
3
As long as information is common knowledge within
each bilateral relationshipas I assumean individual player cannot benet from lying.
1
This example is adapted from Mishkin (1998), quoted in Wallace (2001).
2
The former result is reminiscent of the theorem of Ben-Porth and Kahneman (1996), and the latter
result is reminiscent of the theorem of Renault and Tomala (1998). The precise connection between these
results is discussed later.
3
I restrict attention to games with a mutual-minmax Nash equilibrium, so designing punishments is
trivial.
1
With private cheap talk, 2-connectedness implies that there are at least two independent
paths through which a piece information can reach each player, so again a single player
cannot gain from lying.
In general, however, public information cannot be replicated with only private cheap talk
when the network is not 2-connected, as the above examples show. I therefore consider
the case where players can not only talk privately but can also privately exchange evidence.
I focus on a very specic form of evidence: the players are endowed with undierentiated,
divisible tokens that they can freely transfer to their neighbors. The crucial dierence
between tokens and talk is that one player cannot send another more tokens than she has,
while a player can always send any cheap talk message.
I study tokens rather than some other form of evidence for two reasons. First, I view
the tokens as a natural model of at money, and indeed they are technologically very similar
to money as modelled in the literature on the microfoundations of money dating back
to Kiyotaki and Wright (1993) (though, crucially, I allow the players to use the tokens in
unrestricted ways, rather than requiring that they trade tokens for goods).
4
I therefore
henceforth refer to the tokens as money. Second, the tokens are intuitively a fairly minimal
form of evidence. Allowing more sophisticated forms of evidence, like tokens that are tagged
with dierent colors, or letters with signatures that cannot be forged, would only make the
positive results of this paper easier to prove.
The main result of the paper is that public information can always be replicated with
private cheap talk and money. Very roughly speaking, the idea is to endow leaf players
like 1 and 3 in the exampleswith money, and to endow non-leaf playerslike 2with
none. Non-leaf players must then obtain money from leaf players in order to convince others
that they have behaved well, which disciplines their behavior. The result is presented in quite
a general setting, however, which necessitates the use of somewhat complicated sequences of
money transfers to ensure that non-leaf players cannot misrepresent their information.
5
I then apply this result to study when money is essential, in that the equilibrium payo set
4
For example, Wallace (2001) denes money as a tangible useless object. See Section 6 for further
discussion.
5
Public information can also always be replicated with money alone: simply replace cheap talk messages
with transfers of tiny amounts of money that nonetheless convey a large amount of information.
2
is strictly larger with private cheap talk and money than with neither, or strongly essential,
in that the equilibrium payo set is strictly larger with private cheap talk and money than
with cheap talk alone.
6
I show that a sucient condition for money to be essential is
that the network contains a nice subnetwork, which is a subtree in which every bilateral
relationship has a product structure (Fudenberg and Levine, 1994) and in which there is
some payo vector that can be sustained in equilibrium with public monitoring that cannot
be sustained in a locally public equilibrium (a generalization of perfect public equilibrium)
with private monitoring.
Finally, I show that in two leading applicationsrepeated partnership games and favor
trading gamesthe condition that the network contains a nice subnetwork reduces to the
property that it contains a subtree of size at least three. Therefore, money is (strongly)
essential in these models if the network contains a subtree of size at least three.
From an applied perspective, there are (at least) two ways of interpreting the results of
the paper. The rst is to take them at face value and conclude that in many models money
can indeed be used to replicate public information and is therefore often essential. The
second is to take them as a puzzle: what constraints on players information or rationality
must be added to the model to get a more realistic model of money? I hope that both
interpretations may have value, and briey discuss the second in the conclusion.
The paper proceeds as follows: Section 2 relates the paper to the literatures on repeated
games, networks, and the microfoundations of money. Section 3 presents an example in
which a simple and realistic way of using of money can expand the equilibrium payo set, and
argues that one must allow for more abstract and complicated uses of money to understand
the limits of what money can do in networks. Section 4 then presents the model. Section 5
contains the benchmark results on replicating public information with cheap talk. Section
6 presents the main result on replicating public information with money. Section 7 shows
how this result can be applied to show that money is essential. Section 8 applies the results
of Sections 6 and 7 to partnership games, and Section 9 applies these results to favor trading
games. Section 10 concludes. Proofs are contained in the appendix.
6
Cf Lagos and Wright (2008), who write, Money is said to be essential when the set of allocations that
can be supported as equilibria is larger (or, sometimes, better) with money than without it. They attribute
this notion to Hahn (1973).
3
2 Related Literature
This paper is distinguished from the literature by its focus on when dierent communication
technologies can replicate public information and on comparing cheap talk and money in a
repeated game. Nonetheless, it is related to the literatures on community enforcement and
communication in repeated games, repeated games on networks, and the microfoundations
of money.
The seminal paper on community enforcement in repeated games is Kandori (1992),
who shows that cooperation is sustainable in the repeated prisoners dilemma with random
matching when players carry labels (such as guilty or innocent) that are exogenously
determined by their past play.
7
However, most of the subsequent literature on community
enforcement has not considered hard evidence.
8
There is also a literature on the folk theorem
in general private monitoring repeated games with communication, dating back to Compte
(1998) and Kandori and Matsushima (1998). The main dierences between my paper and
this literature is that I restrict attention to repeated games on networks and compare the
equilibrium payo set with dierent communication technologies for a xed discount factor.
The is also a rapidly growing literature on repeated games on networks. The folk the-
orems of Ben-Porath and Kahneman (1996) and Renault and Tomala (1998) are related to
the benchmark results of Section 5 and are discussed there. Most of the rest of the literature
studies more specic games. For example, Ahn and Suominen (2001) and Balmaceda and
Escobar (2011) study how local communication among buyers can dissuade a seller from pro-
viding a low-quality good, and Lippert and Spagnolo (2011) and Ali and Miller (2012) study
how local communication can help sustain cooperation in a repeated prisoners dilemma.
Finally, this paper relates to the large literature on the microfoundations of money. I
provide sucient conditions for money to be essential in games with a nite, non-anonymous
population of players interacting on a xed network, relative to what could be achieved
with cheap talk alone. In contrast, most of the existing literature considers games with a
7
This is closely related to the norm equilibrium of Okuno-Fujiwara and Postlewaite (1995). Milgrom,
North, and Weingast (1990) argue that law merchants in medieval Europe may have played a role in
providing such labels. Ellison (1994) shows that cooperation is also sustainable without labels.
8
A recent exception is Fujiwara-Greve, Okuno-Fujiwara, and Suzukis (2012) model of a prisoners
dilemma with reference letters.
4
continuum of anonymous players interacting at random, and does not compare money with
cheap talk; for example, this is the setting in Kiyotaki and Wright (1993).
9
A prominent paper in this literature that shares with mine the goal of comparing at
money with other information technologies is Kocherlakota (1998). Kocherlakotas main
result is that money is often inessential in the presence of a form of public information.
10
In contrast, I show that money is often essential when only private information is available.
Kocherlakota also gives an example in which (indivisible) money can replicate public in-
formation and an example in which it cannot. In my model, money can replicate public
information quite generally; the primary reason for this dierence is that players in my model
are non-anonymous, which makes moneyand repeated game eects more generallymuch
more powerful than in Kocherlakotas model.
In a follow-up paper, Kocherlakota (2002) allows money to be innitely divisible and
shows that undierentiated money can then replicate public information if money holdings
are observable (the one-money theorem), and that tagged money (e.g., red money and
black money) can replicate public information even ifas in my modelplayers can conceal
money (the two-money theorem). In contrast, in my model plain, undierentiated money
can replicate public information even if money is concealable.
11
They key dierence is again
that my players are non-anonymous and can make more complicated sequences of transfers.
12
3 A Simple Example
I begin by eshing out the three players on a line example, which illustrates the need for
the more general analysis of the rest of the paper.
9
There are some exceptions, however. Araujo (2004) adopts the arguments of Kandori and Ellison to
show that money is essential in suciently large nite games with random matching. Aliprantis, Camera,
and Puzzello (2007) present a model with an innite but non-anonymous population where money is essential
even though players occasionally meet in centralized markets. Kocherlakota and Wallace (1998) show that
money is essential with a continuum of players and random matching even in the presence of suciently
unreliable public monitoring of individual actions.
10
More precisely, Kocherlakotas notion of memory is perfect information about ones partners past play,
their partners past play, and so on.
11
I also show that innite divisibility of money can be replaced by a form of targeted ination in my model.
12
However, even with tagged money my results would not follow from Kocherlakotas. Among other
dierences, Kocherlakotas model invovles trading mechanisms, while there are no mechanisms or contracts
in my model (i.e., all money transfers are completely voluntary).
5
To keep the example as simple as possible, suppose the players interact in continuous
time. Players 1 and 2 are linked and players 2 and 3 are linked, but players 1 and 3 are not.
If players i and , are linked, then at rate 1 player i gets the opportunity to do a favor for
player ,, and this opportunity is observed by i and , but not by the other player. Doing a
favor costs c and yields benet / c to the recipient, and the players have discount rate :.
For this example only, I also let the players access a public randomizing device.
It is clear that, if the players are suciently patientin particular, if : _
bc
c
they can
sustain cooperation with bilateral grim trigger strategies: if i fails to do a favor for ,, then
, never again does a favor for i. But if the players are less patientif :
bc
c
then it is
shown in the appendix that no cooperation is possible. However, the appendix also shows
that the players can sustain some cooperation for some :
bc
c
if they have access to a single
dollar bill, by playing as follows:
Initially endow any player with the dollar. At any time, if player 2 (the middle player)
has the dollar and player 1 or 3 gets the opportunity to do a favor for him, she does the
favor in exchange for the dollar. If player 1 or 3 has the dollar, player 2 does a favor for
her when he gets the opportunity, but when he does the favor he gets the dollar only with
probability c

=
b+2c
2b+c
< 1. No one does favors for a player who does not have the dollar,
and no one does favors for or transfers money to anyone who takes an o-path action.
This example has the appealing feature that the players use money in a realistic way:
1 and 3 only do favors in exchange for the dollar, and 2 only does favors in exchange
for a chance of receiving the dollar.
13
But it also has several limitations: the game and
the network structure are both very special, and the constructed equilibrium is still very
inecient. This illustrates a general tradeo: as one requires that money is used more
realistically, money becomes less powerful. This paper studies one extreme point on this
frontier: while I maintain a standard technological denition of moneymoney is a tangible
useless object, like in the exampleI place no restrictions on the strategic role of money.
Studying what can be achieved with a restricted class of strategiesor in models where
13
I conjecture that one could construct a very similar example where 2 does favors in exchange for a
fraction of a dollar, so that the price of buying a favor is lower for 1 and 3 than for 2. The current example
is simpler to analyze because the vector of money holdings takes only three values, given by the location of
the dollar.
6
complicated strategies are infeasible, due to limits on players information or rationalityis
left as an intriguing direction for future research.
4 Model
This section describes the repeated game without communication and the notion of repli-
cating public information. I add cheap talk to the model in Section 5 and add money in
Section 6.
Players: There is a nite set of players ` = 1. . . . . : arranged on an undirected
and connected network 1 _ 1
2
(`), the set of 2-element subsets of `, where i. , 1
denotes a link between players i and ,. The network will determine the structure of players
actions, payos, information, andin subsequent sectionscommunication. In particular,
it will become clear that the assumption that 1 is connected is essentially without loss of
generality, as the fact that players only interact with their neighbors implies that if 1
is not connected one can replicate the analysis on each connected component of 1. Let
`
i
= , : i. , 1 be the set of player is neighbors.
Stage game: Player is stage-game action set is
i
=

j2N
i

i;j
, where the
i;j
are
arbitrary nite sets interpreted as player is possible actions toward player ,. There is a
set of signal proles 2 =

fi;jg2L
2
i;j
, where the 2
i;j
= 2
j;i
are arbitrary nite sets in-
terpreted as the signals that can be generated by the interaction between players i and ,.
It is assumed that there are probability distributions :
i;j
([c
i;j
. c
j;i
) such that the proba-
bility of signal .
i;j
conditional on action pair (c
i;j
. c
j;i
) is :
i;j
(.
i;j
[c
i;j
. c
j;i
), independent of
the signal realizations for other pairs of players. That is, the probability of signal prole
. = (.
i;j
)
fi;jg2L
given action prole c = (c
i
)
i2N
is given by : (.[c) =

fi;jg2L
:
i;j
(.
i;j
[c
i;j
. c
j;i
).
Hence, the signal .
i;j
is locally public, in that it is identically equal to .
j;i
but is com-
pletely uninformative about any other .
i
0
;j
0 . Player is stage-game expected payo is
n
i
(c) =

j2N
i

z
i;j
2Z
i;j
:
i;j
(.
i;j
[c
i;j
. c
j;i
) n

i;j
(.
i;j
), where n

i;j
: 2
i;j
R gives player is
realized payo from her interaction with player ,. To save on notation, let n
i;j
(c
i;j
. c
j;i
) =

z
i;j
2Z
i;j
:
i;j
(.
i;j
[c
i;j
. c
j;i
) n

i;j
(.
i;j
), and note that n
i
(c) =

j2N
i
n
i;j
(c
i;j
. c
j;i
). Thus, n
i;j
:

i;j

j;i
R gives player is expected payo from her interaction with player ,. For
7
i. , 1, I will refer to the two-player game (
i;j
.
j;i
. 2
i;j
. :
i;j
. n
i;j
. n
j;i
), which captures
the direct relationship between i and ,, as the (i. ,)-game.
I assume throughout the paper that each (i. ,)-game has a mutual-minmax Nash equi-
librium. That is, I assume that every mixed action set (
i;j
) contains an element c

i;j
such that the mixed action prole c

= (c

i
)
i2L
=
_
_
c

i;j
_
j2N
i
_
i2N
is a stage-game Nash
equilibrium and
n
i;j
_
c

i;j
. c

j;i
_
= min

j;i
2(A
j;i
)
max

i;j
2(A
i;j
)
n
i;j
(c
i;j
. c
j;i
) for all i. , 1.
This assumption ensures that the worst possible punishments can be delivered link by link,
and thus do not require punishers to coordinate. It is needed for my results, because an
outsider will generally be able to tell when a deviation occurs in the relationship between
two players but will not be able to tell which one of them deviated.
Repeated game: The players play a repeated game in discrete time. At the beginning
of period t 0. 1. . . ., each player i chooses an action c
i;t

i
. The signal .
t
is then
drawn from : ([c), payos are realized, and player i observes (.
i;j;t
)
j2N
i
.
14
Letting /
i;t
=
_
c
i;t
. (.
i;j;t
)
j2N
i
_
, player is time-t history is /
t
i
= (/
i;
)
t1
=0
for t _ 1, and every player
has trivial initial history /
0
i
= /
0
. Letting H
t
i
be the set of player is time-t histories, a
behavior strategy of player is is a map o
i
: H
t
i
(
i
). Players have common discount
factor o (0. 1). Denote the resulting repeated game by
PRI
, where the subscript 111
emphasizes that signal .
i;j
is private to the pair of players i. , (though it is locally public
between i and ,). I will study the sequential equilibria (SE) of game
PRI
.
15
Let 1
PRI
be
the set of SE payos of game
PRI
.
Replicating public information: Let
PUB
be the game in which the entire signal
. is public. That is,
PUB
is derived from
PRI
by letting /
i;t
equal (c
i;t
. .
t
) rather than
_
c
i;t
. (.
i;j;t
)
j2N
i
_
. Let 1
PUB
be the set of SE payos of game
PUB
. Below, I will dene
games
PUBCT
PRI
,
PRICT
PRI
, and
$
PRI
by adding public cheap talk, private (i.e., local) cheap talk,
or money to the game
PRI
, and will denote the corresponding SE payo sets by 1
PUBCT
PRI
,
14
Thus, player i observes her own payo.
15
As usual, sequential equilibrium in repeated games with nite information sets in every period is dened
by putting the product topology on the space of beliefs.
8
1
PRICT
PRI
, and 1
$
PRI
. I will say that public cheap talk (resp., private cheap talk, money) can
replicate public information if 1
PUBCT
PRI
_ 1
PUB
(resp., 1
PRICT
PRI
_ 1
PUB
, 1
$
PRI
_ 1
PUB
).
16
Informally, communication replicates public information if any payo vector that can be
attained in equilibrium when all local information is made public can also be attained with
communication.
5 Replicating Public Information with Cheap Talk
This section studies when public or private cheap talk may be used to replicate public
information. The results of this section are broadly similar to results in the literature, and
are included both for completeness and because there is a natural progression from these
results to the main analysis of Section 6.
5.1 Public Cheap Talk
A game with public cheap talk
PUBCT
PRI
(1 ) is derived by augmenting the game
PRI
with a -
nite message set 1 = (1
1
. . . . . 1
n
) such that after players observe their private signals they si-
multaneously send public messages
i
1
i
. Formally, letting /
i;t
=
_
c
i;t
. (.
i;j;t
)
j2N
i
. (
j;t
)
j2N
_
,
there are now two kinds of histories for every time t, denoted /
t
i
= (/
i;
)
t1
=0
(called action
histories) and /
t+
i
=
_
(/
i;
)
t1
=0
. c
i;t
. (.
i;j;t
)
j2N
i
_
(called communication histories), and a
strategy maps action histories to (
i
) and maps communication histories to (1
i
). Let
1
PUBCT
PRI
(1 ) be the SE payo set of
PUBCT
PRI
(1 ), and let 1
PUBCT
PRI
=

Y
1
PUBCT
PRI
(1 ), where
the union is taken over all nite sets 1 .
The rst benchmark result is that public cheap talk can always replicate public infor-
mation. The proof is very simple: Take a SE o
PUB
in the public monitoring game
PUB
.
Specify that after every round of play the players publicly report what signals they observe
and then play according to o
PUB
, taking the reported signals as the true ones. If any reports
disagree, play the mutual-minmax prole c

forever. This strategy prole yields the same

payos as o
PUB
, and it can be shown to be a SE.
17
16
There are games for which these inclusions are strict. I omit the proof of this fact, since it is not used
in the paper.
17
This argument clearly relies heavily on the assumption that signal .
i;j
is locally public between i and ,.
9
Theorem A Public cheap talk can replicate public information (i.e., 1
PUBCT
PRI
_ 1
PUB
).
In addition, the proof of Theorem A shows that there is a single nite message set 1 such
that 1
PUBCT
PRI
(1 ) _ 1
PUB
. The same will be true of the later results on communication
through private cheap talk or through money. Thus, in all cases there is no need to tailor
the message set (or the initial endowment of money) to the desired payo vector.
Theorem A is related to Theorem 1 of Ben-Porath and Kahneman (1996), which estab-
lishes the folk theorem for repeated games with public communication where each player is
perfectly observed by at least two others. Here, it is enough that i and , observe the same
.
i;j
, because there is a mutual-minmax Nash equilibrium. Also, Theorem A is not a folk
theorem but rather a result about replicating public information for xed o.
5.2 Private Cheap Talk
In this paper, private cheap talk means communication along the links of the network.
That is, with private cheap talk players can communicate directly with their neighbors
but not with other players. However, players can communicate indirectly with players to
whom they are not linked by passing information from one link to another. This requires
multiple rounds of communication after every round of play, which I allow. Indeed, in any
communication round a player may learning something that she would like to pass on. To
accommodate this, I allow for innitely many rounds of communication after each round of
play.
A game with private cheap talk
PRICT
PRI
(1 ) is derived by augmenting the game
PRI
with a nite message set 1 = (1
i;j
)
fi;jg2L
such that after players observe their private signals
they have innitely many opportunities to simultaneously send private messages
k
i;j
1
i
to their neighbors, where the subscript denotes a message from i to , and the superscript
/ N denotes the number of the communication round. Formally, the stage game is now
a long cheap talk game as modeled by Aumann and Hart (2003). That is, letting /
i;t
=
If signals were not even locally public, then one would be in the setting of general repeated games with private
monitoring, and public information could not be replicated even with public communication. However, it
may be the case that if signals are almost locally public then public communication can almost replicate
public information. I do not pursue this question here.
10
_
c
i;t
. (.
i;j;t
)
j2N
i
.
_

k
i;j;t
.
k
j;i;t
_
j2N
i
;k2N
_
, there are now innitely many kinds of histories for
every period t, denoted /
t
i
= (/
i;
)
t1
=0
(action histories), /
t;0
i
=
_
(/
i;
)
t1
=0
. c
i;t
. (.
i;j;t
)
j2N
i
_
,
and /
t;k
i
=
_
(/
i;
)
t1
=0
. c
i;t
. (.
i;j;t
)
j2N
i
.
_

k
0
i;j;t
.
k
0
j;i;t
_
j2N
i
;k
0
2f1;:::;kg
_
, for / N (communication
histories). A strategy o
i
is now a measurable function that maps action histories /
t
i
to
(
i
) and maps communication histories /
t;k
i
to
_
(1
i;j
)
j2N
i
_
. Let 1
PRICT
PRI
(1 ) be the SE
payo set of
PRICT
PRI
(1 ), and let 1
PRICT
PRI
=

Y
1
PRICT
PRI
(1 ).
18
The second benchmark result is that private cheap talk can replicate public information
if the network 1 is 2-connected. Recall that a network is 2-connected if there are at least
two independent paths (i.e., two paths with disjoint sets of internal nodes) between every
pair of nodes. The main idea is again quite simple: Start with a SE o
PUB
in game
PUB
.
Specify that after every round of play there are several rounds of communication in which
players report both the signals they have observed directly and the signals that have been
reported to them in earlier rounds, until all signals have been reported to all players. The
players then play according to o
PUB
, taking the reported signals as the true ones. If a player
sends or receives an inconsistent report, she then reports that there has been a deviation,
and the news of the deviation spreads throughout the network and leads all players to play
the mutual-minmax prole c

. The assumption that the network is 2-connected implies

that no player can deceive another about the signals: if a player i lies about a signal to one
of her neighbors, the neighbor will eventually receive a conicting report via a path that
does not include i, and will then revert to c

.
Theorem B Private cheap talk can replicate public information (i.e., 1
PRICT
PRI
_ 1
PUB
) if
the network 1 is 2-connected.
Theorem B is related to Theorem 2.6 of Renault and Tomala (1998), which gives a Nash
folk theorem for repeated games with a 2-connected monitoring network without commu-
nication. Theorem B avoids some complications that emerge in their paper by allowing
communication and assuming a mutual-minmax Nash equilibrium (though Theorem B is for
18
Aumann and Hart prove that in a single long cheap talk game the induced mapping from strategies to
payos is measurable, so that the game is well-dened. Their proof immediately extends to the current
repeated game model with nite players, actions, signals, and messages. It also extends to the model of
Section 6, where introducing divisible money makes the message sets countably innite.
11
sequential equilibrium rather than Nash). Moreover, the results dier in that Theorem B
is about replicating public information for xed o.
19
6 Replicating Public Information with Money
I now turn to the main part of the analysis, where players have access to at money.
A game with money
$
PRI
is similar to a game with private cheap talk, except that in
addition to sending cheap talk messages players can also transfer quantities of money to each
other.
20
The dierence between cheap talk and money is that a player can send any cheap
talk message she wants, but can only send money that she is currently holding: for example,
any player can say message number 5, but only a player with at least 5 dollars can make a
$5 transfer. This is quite consistent with the modern view of what money is. For example,
Wallace (2001) dierentiates money from an arbitrary state variable as follows:
. . . misrepresentation of holdings of a tangible object is limited by the possibility
that others can at least say show me. For an intangible state variable, there
are no limits on misrepresentation if there is no monitoring.
Formally, a game with money
$
PRI
(1. :
0
) is derived from the game with private cheap
talk
PRICT
PRI
(1 ) by specifying an initial vector of money holdings :
0
= (:
0
1
. . . . . :
0
n
), with
:
0
i
Q
+
for all i ` (where Q
+
denotes the non-negative rationals), and allowing players
to transfer money concurrently with their messages. That is, at every history in
PRICT
PRI
(1 )
where player i chooses a message
i;j
1
i;j
to send to player ,, she now chooses a pair
(
i;j
. :
i;j
) 1
i;j
Q
+
to send to player ,, subject to the constraint that

j2N
i
:
i;j
_ :
i
,
where :
i
is player is current money holding, and the vector of money holding is then
updated to
:
0
i
= :
i
+

j2N
i
(:
j;i
:
i;j
) .
21
19
There is also a general literature on secure information transmission in networks with private cheap talk,
in which conditions similar to 2-connectedness are often critical (e.g., Renault, Renou, and Tomala, 2012).
20
As indicated in the introduction, it would be essentially equivalent to let players transfer only money
and not also cheap talk messages. The current approach is slightly easier to exposit.
21
The point of allowing players to transfer only rational quantities of money is to ensure that the strategy
space remains countable.
12
The following is the main result of the paper. Here, 1
$
PRI
=

(Y;m
0
)
1
$
PRI
(1. :
0
), where
1
$
PRI
(1. :
0
) is the SE payo set in
$
PRI
(1. :
0
).
Theorem 1 Money can replicate public information (i.e., 1
$
PRI
_ 1
PUB
).
Theorem 1 does not require the strong assumption that the network is 2-connected. The
fact that money can replicate public information even when the network is not 2-connected
will form the basis of the later results on the essentiality of money (Theorems 2-4).
To see the overall approach of the proof, suppose for simplicity that 1 is a tree (so that,
in particular, 1 is not 2-connected). Let o
PUB
be a SE strategy prole in game
PUB
.
Initially, endow each of the leaf players in 1 (i.e., players with only one neighbor) with a
large amount of money, and endow non-leaf players with none. Have players initially play
as in o
PUB
. After each round of play, rst have players repeatedly report their signals to
each other as in the model with private cheap talk: this is called the reporting subphase in
the proof. Then have players use money to check that no players have misreported signals:
this is called the conrmation subphase and is described below. In the next round, players
play according to o
PUB
, taking the reported signals as the true ones.
As in Theorems A and B, if players can be induced to report their signals truthfully,
the above construction yields an equilibrium with the same payos as o
PUB
. Thus, the
key insight behind Theorem 1 is that the conrmation subphase can be constructed so as to
ensure that no player can mislead another about the value of any signal. The construction is
as follows: Assign a natural number / to every possible vector of signals .. At the beginning
of the period t conrmation subphase, each leaf player i thinks that the true vector of period
t signals is some ^ .
i
. The conrmation subphase starts with some leaf playersay, player
1sending /
1
dollars down the path toward another leaf playercall him player 2where
/
1
is the number assigned to ^ .
1
. The non-leaf players on this path are supposed to pass
the money on to player 2. When player 2 receives the transfer, he checks whether it equals
the number /
2
assigned to ^ .
2
, the vector of reports he has received. If it does, he adds an
additional /
2
dollars to the transfer he received, and passes this new larger pot of money
on to the next leaf player, player 3. This process continues until each leaf player get the
chance to add money to the pot, and the pot is then returned to player 1. Finally, player
13
1 then makes an additional large transfer toward each leaf player, who then returns this
transfer to player 1.
While it is hard to give a complete intuition for why this construction works without just
proving the theorem, it is worth noting the following three important facts. First, if some
player fails to pass the correct amount of money toward the next leaf player, then player 1
does not get back a pot of the correct size, and thus does not make the nal conrmatory
transfers to the other leaf players (who can then infer that a deviation occurred). Second,
if any leaf players disagree as to the vector of period t signals, then player 1 again does
not get back a pot of the correct size. Finally, the transfers can be constructed so that a
player can never save enough money in an earlier round to mislead another player in a later
round. Broadly speaking, the construction ensures that no player can aord to deviate
in a protable way.
22
I conclude this section with two remarks on Theorem 1. First, the choice of :
0
is
not crucial. As a consequence, even if one takes the view that initial money holdings are
exogenously given, public information can still be replicated for a wide range of initial money
holdings. The idea is that if any non-leaf players are endowed with money, they can in eect
be forced to transfer all of their money to player 1 at the beginning of the game. Formally,
one can show the following result:
23
Proposition 1 Suppose the initial vector of money holdings :
0
is exogenously given. If
there exists a spanning tree 1
0
_ 1 such that :
0
i
0 for all leaf players i in 1
0
, then
1
$
PRI
(1. :
0
) _ 1
PUB
, where 1 is the message set from the proof of Theorem 1.
Second, Theorem 1 relies on the assumption that money is innitely divisible. This
assumption serves two roles in the proof. First, it lets one ensure that leaf players never
run out of money. This could potentially be addressed by instead rebalancing money
22
Another issue in the proof of Theorem 1 is that explicitly describing o-path play is intractable. This
is because when player i sees player , deviate, player i may wish to conceal this information if passing it on
would lead the other palyers to punish her (and, if the network is not 2-connected, she may be able to do
this). Therefore, I specify only that when player i observes a deviation by player ,, she minmaxes player ,
himself as well as those other players who lie on a path from i to , in 1. The rest of o-path play is left
unspecied, and I appeal to Kakutanis xed point theorem to show that a sequential equilibrium o
PRI
with
the desired on-path properties exists.
23
A spanning tree is a connected subnetwork with no cycles that contains all the nodes in the original
network.
14
holdings between rounds, although it is not completely clear how to do this. Secondand
more importantlyit allows the size of the nal conrmation transfer in each period to
increase over time. This ensures that a player who deviates by saving some money cannot
use it to mimic a later conrmation transfer. Both of these roles of innite divisibility could
instead be lled by simply disbursing more money to the leaf players every period, if this
were allowed (contrary to my assumptions). For example, Theorem 1 would go through if
dollars are indivisible but [2[ dollars are disbursed from a bank to each leaf player in every
period. I do not know if Theorem 1 holds with indivisible dollars without allowing this type
of targeted ination.
7 From Replication to Essentiality
Theorem 1 itself, which shows that money facilitates replicating public information in re-
peated games on networks, is one of the two main contributions of this paper. The second
main contribution is using Theorem 1 to show that money is essentialin that the SE payo
set is larger with money than without itin a broad class of games. I use the following
denition:
Denition 1 Money is essential if 1
$
PRI
) 1
PRI
.
Money is strongly essential if 1
$
PRI
) 1
PRICT
PRI
.
The latter property is indeed stronger because 1
PRICT
PRI
(1 ) _ 1
PRI
for every message set
1 , as messages can always be ignored.
How can one tell whether money is essential in a particular game? A rst observation
is that 1
$
PRI
_ 1
PRI
and 1
$
PRI
_ 1
PRICT
PRI
are trivially true: any SE is
PRI
or
PRICT
PRI
can be turned into a payo-equivalent SE in
$
PRI
by specifying that players never make
transfers and ignore transfers if they are made (in particular, 1
$
PRI
(1. :
0
) _ 1
PRI
and
1
$
PRI
(1. :
0
) _ 1
PRICT
PRI
(1 ) for any (1. :
0
)). Combining this observation with Theorem 1
yields the following corollary:
Corollary 1 Money is essential if 1
PUB
1
PRI
,= O. Money is strongly essential if 1
PUB
1
PRICT
PRI
,=
O.
15
Proof. By Theorem 1, 1
$
PRI
_ 1
PUB
. So 1
PUB
1
PRI
,= O (resp., 1
PUB
1
PRICT
PRI
,= O)
implies that 1
$
PRI
1
PRI
,= O (resp., 1
$
PRI
1
PRICT
PRI
,= O). The observation that 1
$
PRI
_ 1
PRI
completes the proof for essential, and the observation that 1
$
PRI
_ 1
PRICT
PRI
completes the
proof for strongly essential.
While Corollary 1 is quite general, it is not a very useful tool for determining when money
is essential because it can be hard to know when 1
PUB
1
PRI
,= O or 1
PUB
1
PRICT
PRI
,= O. The
diculty is that the set 1
PRI
of all sequential equilibrium payos in the private monitoring
game
PRI
is usually impossible to characterize for xed discount factors, as is the set
1
PRICT
PRI
. However, I will show that one can often establish that money is essential or
strongly essential while restricting attention to the following much more tractable class of
strategies:
Denition 2 A locally public strategy o
i
is a strategy in
PRI
where o
i;j
depends only on
(.
i;j;
)
t1
=0
, for all , `
i
. A locally public equilibrium (LPE) in
PRI
is a SE in
PRI
in
locally public strategies. Denote the LPE payo set in
PRI
by 1
LPE
PRI
.
A local cheap talk strategy o
i
is a strategy in
PRICT
PRI
where o
i;j
depends only on
_
.
i;j;
.
_

k
i;j;
.
k
j;i;
_
k2N
_
t1
=0
, for all , `
i
. A local cheap talk equilibrium (LCTE) in

PRICT
PRI
is a SE in
PRICT
PRI
in local cheap talk strategies. Denote the LCTE payo set in

PRICT
PRI
by 1
LCTE
PRI
.
Thus a locally public strategy is one where player i conditions her play in her relationship
with player , only on the history of locally public signals between i and ,, and a local cheap
talk strategy is one where player i conditions her play in her relationship with player ,
(including the messages she sends to ,) only on the history of locally public signals and
cheap talk between i and ,. Locally public equilibrium is the natural analog of the standard
public perfect equilibrium (PPE) in repeated games with imperfect public monitoring, and
local cheap talk equilibrium is the natural analog of PPE when players can send messages
only about mutually public information. Note that with local cheap talk strategies players
have very little to talk about, since they do not condition their messages on information
that the receiver does not already have. In particular, one can show that the set of LCTE
payos is simply the set of LPE payos in the auxiliary game where each pair of players is
16
given access to a public randomizing device.
I now show that the condition that 1
PUB
1
PRI
,= O (resp., 1
PUB
1
PRICT
PRI
,= O) in
Corollary 1 may be replaced with something like 1
PUB
1
LPE
PRI
,= O (resp., 1
PUB
1
LCTE
PRI
,= O).
To do this, I introduce the notion of a nice subnetwork.
For any subnetwork ` _ 1, let 1[
M
be the SE payo set in the game where ` is the
original network, or equivalently the SE payo set in the game where all links i. , , ` are
deleted (so that 1
PRI
[
M
is the SE payo set in this game with private monitoring, 1
PUB
[
M
is the SE payo set in this game with public monitoring, etc.). For future reference, the
game where ` is the original network will be denoted [
M
. Finally, for any set A, let
co (A) denote the convex hull of A. I now introduce a key denition:
24
Denition 3 A subnetwork ` _ 1 is nice if it has the following three properties:
1. ` is a subtree of 1. That is, for any players i. , `, there is a unique path from i
to , in 1, and every node in this path is contained in `.
25
2. For all i. , `, the (i. ,)-game has a product structure. That is, 2
i;j
= 2
i
i;j
2
j
i;j
and :
i;j
(.
i;j
[c
i;j
. c
j;i
) = :
i
i;j
_
.
i
i;j
[c
i;j
_
:
j
i;j
_
.
j
i;j
[c
j;i
_
.
3. 1
PUB
[
M
co
_
1
LPE
PRI
[
M
_
,= O.
In addition, ` is truly nice if the last condition can be strengthened to 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,=
O.
The following theorem is the key tool for determining when money is essential.
Theorem 2 Money is essential if 1 contains a nice subnetwork. Money is strongly essential
if 1 contains a truly nice subnetwork.
For example, if 1 is a tree, all (i. ,)-games in 1 have a product structure, and 1
PUB
co
_
1
LPE
PRI
_
,=
O, then Theorem 2 says that money is essential.
26
However, Theorem 2 is much more gen-
eral that this because 1 itself need not be nice. In particular, the condition that 1 is a
24
I slightly abuse notation here by letting ' stand for both a subnetwork of 1 and the set of nodes in
that subnetwork.
25
This is stronger than the condition that ' is itself a tree: it is not enough that there is a unique path
from i to , in '.
26
Technically, one can show that 1
PUB
1
LPE
PRI
,= O is sucient here.
17
tree is very strong, but the condition that 1 contains a subtree is trivial. However, not
any subtree will do: in particular, if every (i. ,)-game in ` has a product structure, then
1
PUB
[
M
co
_
1
LPE
PRI
[
M
_
,= O can hold only if ` contains at least three players. In the appli-
cations of Sections 8 and 9, the approach will be to show that any subtree of size at least
three is truly nice, and conclude that money is strongly essential whenever 1 contains a
subtree of size at least three.
The intuition for Theorem 2 is as follows: If ` is a tree and all (i. ,)-games in `
have a product structure, then in the game where ` is the original network it is without
loss of generality to restrict attention to LPE. If in addition ` is a subtree of 1, then
the equilibrium payo set on 1 equals the sum of the equilibrium payo set on ` and the
equilibrium payo set on 1` (in
PRI
). So if money expands the equilibrium payo set
on ` while restricting attention to LPE, then it also expands the (unrestricted) equilibrium
payo set on 1.
8 Application 1: Partnership with Perfect Local Mon-
itoring
This section shows that money is strongly essential in a standard repeated partnership
game/prisoners dilemma on a network, as long as players are patient enough to sustain
some eort in equilibrium and the network contains a subtree of size at least three.
Assume that linked players each choose a real-valued action for their relationship
interpreted as their eort levelsand observe each others action: formally,
i;j
is a nite
subset of R
+
containing 0, with
i;j
=
i
0
;j
0 for i. , . i
0
. ,
0
1, 2
i;j
=
i;j

j;i
, and
:
i;j
((c
i;j
. c
j;i
) [c
i;j
. c
j;i
) = 1. Assume also that
n
i;j
(c
i;j
. c
j;i
) =
1
2
, (c
i;j
. c
j;i
) c
i;j
,
where , : R
2
+
R
+
is a symmetric joint production function satisfying , (0. 0) = 0, ,
i

(0. 2), ,
i
(0. 0) 1, and , concave, where the subscripts here denote partial derivatives.
27
I
27
The assumption that ) is symmetric serves only to simplify notation.
18
also assume that there exist ^ c
i;j
0 such that , (^ c
i;j
. ^ c
i;j
) ^ c
i;j
= 0. Note that this game
is a prisoners dilemma, in that n
i
is decreasing in each c
i;j
and increasing in each c
j;i
, and
that setting c
i;j
= 0 for all , `
i
is dominant for player i in the stage game. In particular,
c = 0 is a mutual-minmax Nash equilibrium.
The assumption that
i;j
is nite is necessary to apply Theorem 2.
28
However, since
essentiality requires that money strictly expands the SE payo set, I will need to assume that

i;j
is suciently dense; for example, if
i;j
= 0. 1000000, then the unique SE payo
vector may be 0 with or without money. To rule this out, I say that
i;j
is -dense if for all
c
i;j
[0. ^ c
i;j
] there exists c
0
i;j

i;j
such that

c
i;j
c
0
i;j

< , and I will assume that

i;j
is
-dense for small .
The result is the following:
Theorem 3 In the partnership game, there exists 0 such that if
i;j
is -dense, o
2f
i
(0;0)
f
i
(0;0)
, and 1 contains a subtree of size at least three, then money is strongly essential.
The proof (in the Appendix) proceeds by showing that any subtree ` = 1. 2 . 2. 3
which must exist if 1 contains a subtree of size at least threeis truly nice, and then applying
Theorem 2. The intuition for why such a subtree is truly nice is that with money one can
specify that the equilibrium in the (2. 3) game is favorable to player 2 and that players 2
and 3 stop cooperating if 2 fails to cooperate with 1. This creates slack in player 2s in-
centive constraint in the (1. 2) game, which allows equilibrium in the (1. 2) game to be more
favorable for 1 than is any equilibrium without money.
29
Thus, in this application money
expands the equilibrium payo set by allowing the pooling of incentive constraints across
relationships.
9 Application 2: Trading Favors
This section shows that money is strongly essential in a networked version of a standard
continuous-time trading favors game (Mbius, 2000; Hauser and Hopenhayn, 2010), as long
28
The only place where niteness of
i;j
is used in Theorems 1 and 2 is the appeal to existence of sequential
equilibrium in nite games in the proof of Theorem 1.
29
The assumption that c
2fi(0;0)
fi(0;0)
ensures that some eort can be sustained in a bilateral relationship,
which is clearly necessary for this intuition to apply.
19
as the network contains a subtree of size at least three.
The game
PRI
is now as follows: Time runs continuously from 0 to . If i. , 1,
then at rate 1 player i gets the opportunity to do a favor for player ,, independently across
i. , `. Only i observes when she has an opportunity to do a favor for ,, though both i
and , see when i actually does a favor for ,. Doing a favor costs c 0, and receiving a favor
gives benet / c. The players have common discount rate : 0. Thus, if o
i
is the set
of times at which player i does a favor and 1
i
is the set of times at which player i receives a
favor, then player is payo is
_

t2T
i
c
rt
/
_

t2S
i
c
rt
c
_
.
I also let the players observe the outcome of a public randomizing device .
t
[0. 1] at every
time t.
30
The game
PUB
is identical to
PRI
, except that now all players observe when i does a
favor for ,.
Finally, the game
$
PRI
is identical to
PRI
, except that now, after i gets an opportunity
to do a favor for , (but before the realization of the public randomizing device .
t
), there is
a phase of communication and transfers, described below. That is, the timing is:
1. Opportunities to do favors are determined, and are observed by the players who can
do them.
2. Communication takes place.
3. Players observe the outcome of the public randomizing device.
4. Players with the opportunity to do favors choose whether or not to do them.
In the communication phase, players other than i learn that a communication phase has
begun only if they receive a report, message, or transfer from a player who received one
30
The public randomizing device could be dispensed with by having the players play jointly controlled
lotteries in the communication phase (e.g., Aumann and Hart, 2003). Its only purpose is to minimize
the number of the changes that have to be made to the communication phase strategies from the proof of
Theorem 1.
20
from a player who received one from. . . player i. Formally, if i gets the opportunity to do
a favor for , at time t, then the period t communication phase is exactly as in Section 6,
with the modication that in round 1 only i is allowed to send messages or transfers to their
neighbors, and in subsequent round only players who have received a message or transfer at
some earlier round (in addition to i herself) are allowed to send messages or transfers.
31
The goal of this section is to use Theorems 1 and 2 to show that money is strongly
essential in the trading favors game. However, the trading favors game does not formally
t into the model of Section 3 (e.g., it is not a repeated game), so Theorems 1 and 2 cannot
be immediately applied. The following proposition adapts Theorems 1 and 2 to the trading
favors game:
32
Proposition 2 In the trading favors game, money is strongly essential if 1 contains a sub-
tree ` with 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,= O.
The main result of this section is the following:
Theorem 4 In the trading favors game with unobservable opportunities, if : _
bc
c
and 1
contains a subtree of size at least three, then money is strongly essential.
The proof (in the Appendix) again shows that any subtree ` = (1. 2) . (2. 3) is truly
nice and applies Proposition 2. The proof that any such subtree is truly nice uses the
dierential characterization of the PPE payo set in the two-player trading favors game
with unobservable opportunities due to Hauser and Hopenhayn (2010) (which in turn is
an application of the results of Abreu, Pearce, and Stacchetti (1990)), and also requires
some new results for that game. The intuition is that the source of ineciency in the
two-player game comes from hitting the individual rationality constraints: at some point,
player 2 cannot do more favors for player 1, because player 1 would then owe so many
31
There is also a dierence in notational convention: in Section 6, the period t communication phase occurs
after the period t action phase, whereas here the time t communication phase occurs before players decide
whether or not to do favors at time t.
32
Note that the requirement that every i, , game in ' has a product structure does not appear in
Proposition 2. This is because the property of repeated games with a product structure that is needed for
Theorem 2 is that player is beliefs about player ,s private history conditional on player is actions and the
public history between i and , equals her beliefs about player ,s private history conditional only on the
public history, and this property holds in the trading favors game.
21
favors that she would prefer to revert to autarky. In the three-player game with public
monitoringor moneyone can specify that when player 2 does a favor for player 1, some
of the repayment for the favor has to come from player 3. In other words, player 1 and
3s individual rationality constraints can be partially pooled, which reduces ineciency.
Finally, note that money is also essential whenever 1 contains a subtree of size at least
threefor an open interval of discount ratesin the model where both i and , can observe
when i has an opportunity to do a favor for ,. This is precisely the example of Section 3.
10 Conclusion
This paper has compared cheap talk and at money as means of replicating public informa-
tion in repeated games on networks. The main result is that public information can always
be replicated when money is availablein contrast, it can only be replicated when the net-
work is 2-connected if money is unavailable. This leads to a simple sucient condition for
the essentiality of money: money is essential if the network contains a nice subnetwork, i.e., a
subtree on which replicating public information may be shown to be valuable while restrict-
ing attention to locally public equilibria. In leading applications, this condition reduces to
the property that the network contains a subtree of size at least three.
As I have mentioned above, an important direction for future research is to impose limits
on players information or rationality and see what this implies for what payos can be
achieved with money and for how money is used. For example, it would be interesting
to see how the results of this paper extend to the case where the network is uncertain:
recalling the example of three players in a line, if 1 and 3 do not have a relationship, then
assuming that each knows that the other has a relationship with 2 seems strong. Similarly,
it would be interesting to see how the results extend when the initial distribution of money
is unknown. Other promising directions include studying models where players are more
anonymous than in this paper but less anonymous than in standard continuum agentrandom
matching models of money, or models where players use maxmin optimal strategies or other
boundedly rational rules in the face of uncertainty about the distribution of money.
22
11 Appendix
11.1 Details for Section 3
I rst show that no cooperation is possible without money if :
bc
c
(even if private cheap
talk is allowed), and then show that the proposed strategy prole with money is an equilib-
rium if : _
b

c
c
.
For the game without money, a simple extension of the proof of Theorem 2 (which may
be found in an earlier version of the paper) shows that 1
PRICT
PRI
= 1
LCTE
PRI
. Let c
i
be player
is expected discounted future action toward player , ,= i, for i. , 1. 2:
c
i
= :
_
1
s=t
c
r(st)
1
_
c
i;s
[ (.
i;j;
)
t
=0

d:,
where c
i;t
[0. 1] is the probability that player i does a favor for player , when she gets the
opportunity to do so at time t.
33
So when player i is supposed to do a favor for player , at
time t, her payo in the (i. ,)-game from conforming is
1
r
(/ c
j
c c
i
) c, and her payo in
the (i. ,)-game from deviating to playing c
i
= 0 forever is 0. Hence, in any LCTE in which
c
i
0,
: _
/ c
j
c c
i
c
.
So in any such LCTE it must also be that c
j
0. Suppose, toward a contradiction, that
such an LCTE exists. Then without loss of generality c
i
_ c
j
. So it must be that
: _
_
/ c
c
_
c
i
.
Since c
i
[0. 1] and :
bc
c
, this is impossible. So there is no LCTE with c
i
0 for any i,
so 1
LCTE
PRI
= 1
PRICT
PRI
is just the 0 vector.
For the strategy prole with money, let \
i
be player 2s continuation value when player
33
I impose here that a
i;t
is measurable over intervals where no favors are done, which implies that the
integral is well-dened.
23
i has the dollar. Noting that \
1
= \
3
, we have
:\
2
= 2 (\
1
\
2
+ /) ,
:\
1
= c(\
2
\
1
) c. (1)
Player 2s strategy if optimal if and only if
c\
2
+ (1 c) \
1
_ c,
since if he does a favor he gets \
2
with probability c and gets \
1
with probability 1 c, and
if he fails to do a favor he gets 0 (as he never again gets the dollar). By (1), this inequality
holds if and only if
\
1
_ 0,
or equivalently
\
2
_ c.
Some algebra shows that this is the case if and only if
c _
c
/
_
1 +
:
2
_
.
Next, let \
i
be player 1s continuation value when player i has the dollar. Then
:\
1
= c(\
2
\
1
) + /
:\
2
= \
1
\
2
c + \
3
\
2
(2)
:\
3
= c(\
2
\
3
) . (3)
Player 1s strategy is optimal if and only if
\
1
_ c,
since if she does a favor she gets \
1
and if she fails to do a favor she gets 0. By (2) and
24
(3), this holds if and only if
\
2
_ 0.
Some algebra shows that this is the case if and only if
c _
/
c
:.
By symmetry, this inequality is also necessary and sucient for player 3s strategy to be
optimal. Finally,
b
c
: _
c
b
_
1 +
r
2
_
if and only if : _
b

c
c
. For any : in this range, setting
c = c

satises both incentive constraints.

11.2 Proof of Theorem A
Let 1
i
=

j2N
i
2
i;j
. I show that 1
PUBCT
PRI
(1 ) _ 1
PUB
.
Let o
PUB
be a SE strategy prole in game
PUB
. Dene a strategy prole o
PRI
in game

PUBCT
PRI
(1 ) as follows:
Initially, play as in o
PUB
(i.e., o
PRI
i
(/
0
) = o
PUB
i
(/
0
)).
Always report all signals truthfully (i.e., o
PRI
i
(/
t+
) = (.
i;j;t
)
j2N
i
for all /
t+
).
Let ^ .
i;j;t
be the i. , coordinate of
i;t
. At history /
t
i
, if ^ .
i
0
;j
0
;
= ^ .
j
0
;i
0
;
for all i
0
. ,
0

1 and all t < t, then let o
PRI
i
_
/
t
i
_
= o
PUB
i
_
^
/
t
i
_
, where
^
/
t
i
=
_
c
i;
. (^ .
i
0
;j
0
;
)
fi
0
;j
0
g2L
_
t1
=0
.
If instead ^ .
i
0
;j
0
;
,= ^ .
j
0
;i
0
;
for some i
0
. ,
0
1 and t < t, then let o
PRI
i
_
/
t
i
_
= c

i
.
Clearly, o
PRI
yields the same payo vector as o
PUB
. It remains only to show that o
PRI
is a SE prole. To see this, note rst that player i does not have a protable deviation at an
action history /
t
i
with ^ .
i
0
;j
0
;
= ^ .
j
0
;i
0
;
for all i
0
. ,
0
1 and all t < t. This follows because
her continuation payo from playing any action c
i
at history /
t
i
under o
PRI
is the same as
her continuation payo from playing c
i
at history
^
/
t
i
under o
PUB
, and o
PUB
is a SE. Second,
player i does not have a protable deviation at an action history /
t
i
with ^ .
i
0
;j
0
;
,= ^ .
j
0
;i
0
;
for some i. , 1 and t < t, because starting from such a history her opponents play
c

i
forever and c

i
is a best response to c

i
. Finally, player i does not have a protable
deviation at a communication history /
t+
. In the case where ^ .
i;j;
= ^ .
j;i;
for all i. , 1
25
and all t < t, this follows because her continuation payo when she conforms to o
PRI
equals
her expected continuation payo under o
PUB
conditional on reaching history
^
/
t
i
, playing
c
i;t
, and observing signals (.
i;j;t
)
j2N
i
, while her continuation payo when she deviates equals
n
i
(c

), which is weakly less. In the case where ^ .

i;j;
,= ^ .
j;i;
for some i. , 1 and t < t,
this follows because her continuation payo equals n
i
(c

) whether she conforms or deviates.

Hence, player i does not have a protable deviation at any history, so o
PRI
is a SE prole.
11.3 Proof of Theorem B
I rst introduce one unorthodox piece of terminology. Throughout the appendix, I will say
that an (action or communication) history /
t
i
is on-path under strategy prole o if it reached
with positive probability when all players follow o or if there exists another history
~
/
t
i
that
diers from /
t
i
only in player is past actions (c
i;
)
t
=0
such that
~
/
t
i
is reached with positive
probability when all players follow o. A history is o-path otherwise. The point of this
terminology is that if player i trembles at an action history under a SE o but nonetheless
an on-path signal is generated, then player i will want to forget about the deviation.
By calling the resulting history on-path, it will be possible to insist that player i plays
her mutual-minmax action c

i
at o-path histories, which is convenient for constructing
equilibria.
Let 1
i;j
=

fi
0
;j
0
g2L
(2
i
0
;j
0 ' 1
i
0
;j
0 )'0, where 0 and 1
i
0
;j
0 are arbitrary disjoint messages
not contained in any 2
i
0
;j
0 . If a message
i;j
,= 0 and the i
0
. ,
0
coordinate of
i;j
is an
element of 2
i
0
;j
0 (rather than 1
i
0
;j
0 ), then I refer to it as an i
0
. ,
0
report. I show that
1
PRICT
PRI
(1 ) _ 1
PUB
.
Let o
PUB
be a SE strategy prole in game
PUB
. I now construct a strategy prole o
PRI
in game
PRICT
PRI
(1 ) which will be shown to be a SE prole with the same payo vector as
o
PUB
. I rst describe play at action histories, then describe play at on-path communication
histories, and nally describe play at o-path communication histories.
Action Histories: Initially, play as in o
PUB
(i.e., o
PRI
i
(/
0
) = o
PUB
i
(/
0
)). At subse-
quent on-path action histories, o
PRI
i
_
/
t
i
_
= o
PUB
i
_
^
/
t
i
_
, where
^
/
t
i
=
_
c
i;
. (^ .
i
0
;j
0
;
)
fi
0
;j
0
g2L
_
t1
=0
and ^ .
i
0
;j
0
;
is the i
0
. ,
0
report player i received in the period t communication phase.
26
If player i received conicting (i
0
. ,
0
) reports in some period t < t, or did not receive an
(i
0
. ,
0
) report in some period t < t, then /
t
i
is an o-path history (as will become clear
from the description of the communication phase below). At o-path action histories,
o
PRI
i
_
/
t
i
_
= c

i
.
On-Path Communication Histories: In round 1, each player i sends message
_
(.
i;j;t
)
j2N
i
. (1
i
0
;j
0
;t
)
fi
0
;j
0
g6 =fi;jg;j2N
i
_
to every player ,
0
`
i
. In subsequent rounds,
if all i
0
. ,
0
reports that player i has sent or received so far equal ^ .
i
0
;j
0
;t
, then player i
sends every , `
i
the message with i
0
. ,
0
report ^ .
i
0
;j
0
;t
for those i
0
. ,
0
for which she
has received a report and with i
0
. ,
0
coordinate 1
i
0
;j
0
;t
for those i
0
. ,
0
for which she
has not yet received a report. Consequently, if player i has sent or received conicting
i
0
. ,
0
reports for some i
0
. ,
0
, or has sent or received message 0, then her history is
o-path.
O-Path Communication Histories: Send message 0 to all , `
i
.
Note that if all players follows o
PRI
, then for every player i ` and every on-path
history /
t
i
, o
PRI
i
_
/
t
i
_
= o
PUB
i
_
^
/
t
i
_
, where
^
/
t
i
=
_
c
i;
. (.
i
0
;j
0
;
)
fi
0
;j
0
g2L
_
t1
=0
. Therefore, o
PRI
yields the same payo vector as o
PUB
. It remains to show that o
PRI
is a SE prole.
34
I rst claim that if any player i deviates from o
PRI
at any communication history /
t;k
i
,
then every player , ,= i plays c

j
in all subsequent periods.
The rst step in proving the claim is showing that if player i deviates from o
PRI
at any
communication history /
t;k
i
, then some other player reaches an o-path history during the
period t communication phase. This is clearly true if player i deviates by sending message
0, as message 0 is never sent on-path. It is also true if player i deviates by sending (to some
, `
i
) a message with i
0
. ,
0
coordinate 1
i
0
;j
0
;t
rather than sending an i
0
. ,
0
report, or by
sending an i
0
. ,
0
report rather than 1
i
0
;j
0
;t
, as player , knows at what rounds player i has
received a i
0
. ,
0
report on-path. The only remaining possibility is that player i deviates by
sending an i
0
. ,
0
report ^ .
i
0
;j
0
;
,= .
i
0
;j
0
;
to some , `
i
. Assume without loss of generality
that ,
0
,= i. Let (,
0
. ,
1
. . . . . ,
m
. ,) be a path from ,
0
to , that does not include i, which exists
by 2-connectedness. Then in round 1 player ,
0
sends i
0
. ,
0
report .
i
0
;j
0
;
to player ,
1
, and
34
As in the proof of Theorem A, I will show that
_
o
PRI
, j
_
is a SE for any consistent belief system j.
27
by induction in round :
0
+ 1 player ,
m
0 either sends i
0
. ,
0
reports .
i
0
;j
0
;
to player ,
m
0
+1
or
sends message 0 to player ,
m
0
+1
. In either case, player , receives either i
0
. ,
0
report .
i
0
;j
0
;
or message 0 in round : + 1, so /
t;maxfk+1;m+1g
j
is an o-path history.
The secondand nalstep in proving the claim is showing that if all players except
possibly i conform to o
PRI
and some player , ,= i reaches an o-path history during the
period t communication phase, then every player ,
0
,= i plays c

j
0 in all subsequent periods.
To see this, note that at the rst o-path history reached by a player , during the period
t communication phase (call it /
t;k
j
), player , sends message 0 to all of his neighbors. By
induction, each player ,
0
,= i receives message 0 in round / +d, where d is the length of the
shortest path between , and ,
0
that does not include i (which exists by 2-connectedness).
Hence every player ,
0
,= i reaches an o-path history during the period t communication
phase. Therefore, every subsequent action history is o-path for all ,
0
,= i, so all ,
0
,= i play
c

j
0 in all subsequent periods.
It follows from the claim that no player has a protable deviation at an on-path history.
For at any on-path action history /
t
i
player is continuation payo from playing any action
c
i
is the same as her continuation payo from playing c
i
at history
^
/
t
i
under o
PUB
. And at
any on-path communication history /
t;k
i
player is continuation payo from conforming to
o
PRI
equals her continuation payo under o
PUB
conditional on reaching history
^
/
t
i
, playing
c
i;t
, and observing some subset of the period t signals (.
i;j;t
)
fi;jg2L
, while her continuation
payo from deviating equals n
i
(c

), which is weakly less.

Finally, I argue that no player has a protable deviation at an o-path history. The key
observation is that if player i is at an o-path history then regardless of her future play all
of her opponents will play c

in every subsequent period. This is immediate from the claim

if player i is the only player that has deviated from o
PRI
and player i has deviated at a
communication history. If player i deviated from o
PRI
at an action history and an o-path
signal .
i;j
was generated, then player , is at an o-path history.
35
Similarly, if some player
, ,= i has deviated from o
PRI
, then that player is at an o-path history. In either of these
cases, the second paragraph of the proof of the claim then implies that all players ,
0
,= i will
35
If player i deviated at an action history and an on-path signal was generated, then the history is classied
as on-path.
28
reach an o-path history by the end of the communication phase and will subsequently play
c

forever. Therefore, if i conforms to o

PRI
her continuation payo is n
i
(c

), while if she
deviates her continuation payo is weakly less.
11.4 Proof of Theorem 1
Let 1
i;j
=

fi
0
;j
0
g2L
(2
i
0
;j
0 ' 1
i
0
;j
0 ) ' 0, as in the proof of Theorem B. To dene :
0
,
rst let 1
0
be an arbitrary spanning tree of 1 (i.e., a connected subnetwork with no cycles
that contains all the nodes in 1), and let `
0
i
_ `
i
be the set of player is neighbors in 1
0
.
Renumber the players such that the leaves of 1
0
(i.e., the players with only one neighbor in
1
0
) are numbered 1. 2. . . . . :
0
; these players are subsequently referred to as leaf players. Now
dene :
0
by letting :
0
i
= 4:
0
[2[ for all i 1. . . . . :
0
and :
0
i
= 0 for all i :
0
+ 1. . . . . :.
In particular, only leaf players start with money. In addition, number the elements of 2
from 1 to [2[. I show that 1
$
PRI
(1. :
0
) _ 1
PUB
.
Let o
PUB
be a SE prole in
PUB
. I construct a prole o
PRI
in
$
PRI
(1. :
0
) which will
be shown to be an SE prole with the same payos as o
PUB
. I rst describe on-path play in
the action phase and then describe on-path play in the communication phase (which is now
broken into a reporting subphase followed by a conrmation subphase), o-path play,
and o-path beliefs.
Actions: Initially, player as in o
PUB
(i.e., o
PRI
i
(/
0
) = o
PUB
i
(/
0
)). In subsequent pe-
riods, o
PRI
i
_
/
t
i
_
= o
PUB
i
_
^
/
t
i
_
, where
^
/
t
i
=
_
c
i;
. (^ .
i
0
;j
0
;
)
fi
0
;j
0
g2L
_
t1
=0
and ^ .
i
0
;j
0
;
is the
i
0
. ,
0
report player i received in period t. If player i received conicting i
0
. ,
0
re-
ports or did not receive an i
0
. ,
0
report in some period t < t, then /
t
i
is an o-path
history (as will become clear from the description of the communication phase below)
and o
PRI
i
_
/
t
i
_
is therefore given by the description of o-path play below.
Reporting Subphase (On-Path): The reporting subphase consists of the rst :1 rounds
of the communication phase, during which the players report all signals that they have
been informed of and do not transfer money. In particular, in round 1 player i sends
message
_
(.
i;j;t
)
j2N
i
. (1
i;j
)
j = 2N
i
_
to every player ,
0
`
0
i
. In rounds 1 through : 1,
player i sends message with i
0
. ,
0
report ^ .
i
0
;j
0
;t
if all i
0
. ,
0
reports she has sent or
29
received in earlier rounds equal ^ .
i
0
;j
0
;t
, and with i
0
. ,
0
coordinate 1
i
0
;j
0 if she has not yet
received an i
0
. ,
0
report. Note that if all players conform then they are all informed
of the true signals in the reporting subphase.
Conrmation Subphase (On-Path): The conrmation subphase consists of all but the
rst : 1 communication rounds. All players send message (1
i;j
)
fi;jg2L
to all of their
neighbors in every round of the conrmation subphase. I rst describe what money
transfers are made in the conrmation subphase at time t = 0, and then describe how
this diers at time t _ 1.
If player 1 has received consistent (i.e., non-conicting) i. , reports for all i. , 1,
and these reports equal ^ .
1
=
_
^ .
1
i;j
_
fi;jg2L
, then in the rst round of the conrmation
subphase (round :) player 1 sends $/
1
(i.e., /
1
dollars) to the next player on the path
from player 1 to player 2 in 1
0
(note that this path is unique because 1
0
is a tree), where
/
1
is the number between 1 and [2[ assigned to ^ .
1
. In the next round, this player
passes all money received from player 1 on to the next player on the path from player
1 to 2 in 1
0
, and so on until the money reaches player 2. When player 2 receives $/
2
(where /
2
= /
1
on-path), then if player 2 has received consistent i. , reports for all
i. , 1 and in addition /
2
is the number assigned to ^ .
2
=
_
^ .
2
i;j
_
fi;jg2L
, player 2 then
sends $2/
2
to the next player on the path from player 2 to player 3 in 1
0
. Continue
this process: when player i 2. . . . . :
0
1 receives $/
i
, she sends $
i
i1
/
i
down the
path to player i +1 if
1
i1
/
i
is the number assigned to ^ .
i
. When player :
0
receives $/
n
0 ,
she sends $
n
0
n
0
1
/
n
0 down the path back to player 1 if
1
n
0
1
/
n
0 is the number assigned to
^ .
n
0
.
Finally, if player 1 gets back $:
0
/
1
(where $/
1
is the amount of money she initially sent
toward player 2), she then sends $2:
0
[2[ down the path to player 2. As above, this
money is passed along the path from player 1 to player 2. If player 2 receives $2:
0
[2[,
he sends the same amount back toward player 1. If player 1 receives $2:
0
[2[ back,
she then sends $2:
0
[2[ down the path to player 3, and so on until she sends $2:
0
[2[
toward player :
0
and player :
0
sends it back. No further transfers are made.
30
At times t _ 1, the conrmation subphase is as above, except that throughout $/
i
is
replaced with $/
i
,2
t
, and $2:
0
[2[ is replaced with $ (4 1,2
t1
) :
0
[2[.
36
O-Path Play: Dene the i. ,-connectors of network 1, denoted C
i;j
, to be the set of
players that lie on a path from i to , in 1 (including i and , themselves). For players i
and , `
i
, say that player i punishes player , at history /
t
i
if player i plays c

i;j
0 for all
players ,
0
`
i
C
i;j
at all subsequent action histories, sends message 0 to all players
,
0
`
i
C
i;j
at all subsequent communication histories, and never again transfers
money to any player ,
0
`
i
C
i;j
. I specify the following aspects of o-path play:
1. If player i ever receives an o-path signal .
i;j
, i
0
. ,
0
report ^ .
i
0
;j
0 or 1
i
0
;j
0 , or transfer
:
j;i
from player , at an on-path history /
t
i
(i.e., a signal, report, or transfer that
is never received from , at /
t
i
under o
PRI
), then player i punishes player ,.
2. If player i ever sends an o-path i
0
. ,
0
report ^ .
i
0
;j
0 or 1
i
0
;j
0 or transfer :
i;j
to
player , at an on-path history /
t
i
(i.e., a report or transfer that is never sent to ,
at /
t
i
under o
PRI
), then player i punishes player ,.
3. If player i receives message 0 from player , or sends message 0 to player , at any
history /
t
i
(on or o-path), then player i punishes player ,.
4. If player i receives a transfer :
j;i
0 from a player , , `
0
i
or sends a transfer
:
i;j
0 to a player , , `
0
i
at any history /
t
i
(on or o-path), then player i
punishes player ,.
5. If player i receives a transfer :
j;i
,
_
0.
1
2
t
.
2
2
t
. . . . .
n
0
jZj
2
t
.
_
4
1
2
t1
_
:
0
[2[
_
from
player , or sends a transfer :
i;j
,
_
0.
1
2
t
.
2
2
t
. . . . .
n
0
jZj
2
t
.
_
4
1
2
t1
_
:
0
[2[
_
to player
, at any period t history /
t
i
(on or o-path), then player i punishes player ,.
6. Player i never sends a transfer :
i;j
0 to a player , , `
0
i
. That is, players never
transfer money to players to whom they are not linked in 1
0
, even o-path.
7. Player i never sends a transfer :
i;j
,
_
0.
1
2
t
.
2
2
t
. . . . .
n
0
jZj
2
t
.
_
4
1
2
t1
_
:
0
[2[
_
to any
player , at any period t history. That is, at o-path period t histories the players
36
Replacing $/
i
with $/
i
,2
t
prevents players 2, . . . , :
0
from running out of money, which ensures that the
prescribed actions are feasible. Replacing $2:
0
[7[ with $2 (2 1,2
t
) :
0
[7[ ensures that the nal conrmation
phase transfers get larger every period. This is important for sequential rationality, as will become clear.
31
only make transfers that they make at some on-path period t history.
All other aspects of o-path play are dened implicitly as follows. Consider the auxiliary
game in which player i receives payo n
i
(c

) 1 if she deviates from the on-path play

or the aspects of o-path play specied above. This game has a sequential equilibrium
strategy prole by standard arguments.
37
Let the other aspects of o-path play be
given by any such strategy prole.
O-Path Beliefs: I specify that whenever player i receives an o-path report, transfer, or
message from player , at an on-path communication history, she believes that player
, just deviated, and moreover that player , also just sent message 0 to all players
,
0
`
j
i. Consistent belief systems with this property exist, since players may
believe that deviations in later rounds are innitely more likely than deviations in
earlier rounds, and that deviations where player , sends message 0 to all [`
j
[ of his
neighbors are innitely more likely than deviations where he sends message 0 to [`
j
[1
of his neighbors, and that these deviations are in turn innitely more likely than
deviations where he sends message 0 to fewer than [`
j
[ 1 of his neighbors.
It is clear that o
PRI
yields the same payos as o
PUB
. I will show that o
PRI
together
with any consistent belief system with the specied property is a SE.
A preliminary observation, which I will use repeatedly, is that leaf players never send
transfers at o-path histories. To see this, x a leaf player i. At any o-path history /
t
i
,
player i is punishing some player , (as i either sent or received an o-path signal, transfer,
report, or message to or from some ,). Since 1
0
is connected and spans 1 and i is a leaf
player, `
i
_ C
i;j
, so i never again transfers money to any player in `
i
.
37
A subtlety here is that players action sets in the communication phase are countably innite, as :
i;j
can take on any rational value less than :
i
. However, the specication of o-path play implies that any
period t transfer :
i;j
,
_
0,
1
2
t
,
2
2
t
, . . . ,
n
0
jZj
2
t
,
_
4
1
2
t1
_
:
0
[7[
_
yields continuation value n
i;j
0
_
c

i;j
0 , c

j
0
;i
_
for
all ,
0
C
(i;j)
. In addition, player is continuation value against players ,
0
, C
(i;j)
is non-decreasing in player
is money holding, as a player with more money can mimic any strategy that is feasible for a player with
less money. So any period t transfer :
i;j
,
_
0,
1
2
t
,
2
2
t
, . . . ,
n
0
jZj
2
t
,
_
4
1
2
t1
_
:
0
[7[
_
is weakly dominated by
:
i;j
= 0 (in that it never yields a strictly higher continuation value than :
i;j
= 0). Therefore, players action
sets in the auxilliary game have only nitely many undominated actions, and such games admit sequential
equilibria (by applying Kakutanis xed point theorem to the game with only undominated actions).
32
I now prove a key lemma, which says that if player i deviates from o
PRI
then each of her
neighbors either minmaxes her or plays as if she had conformed to o
PRI
.
Lemma 1 For every pair of players i and , `
i
, every strategy o
i
, and every action history
/
t+1
j
reached under strategy prole
_
o
i
. o
PRI
i
_
, o
PRI
j;i
_
/
t+1
j
_

_
c

j;i
. o
PUB
j;i
_
^
/
t+1
j
__
, where
^
/
t+1
j
=
_
c
j;
. (.
i
0
;j
0
;
)
fi
0
;j
0
g2L
_
t
=0
.
Proof. Suppose, toward a contradiction, that o
PRI
j;i
_
/
t+1
j
_
,
_
c

j;i
. o
PUB
j;i
_
^
/
t+1
j
__
for some
, `
i
. Note that if , ever received an o-path signal, transfer, report, or message, then
the player ,
0
whom he received it from must be in C
i;j
(since only i deviates from o
PRI
), so
, plays c

j;i
(as , minmaxes every player in C
j;j
0 , and since , `
i
there is a path of the form
(,. i. . . . . ,
0
)). Hence, history /
t+1
j
must be on-path, and there must be a period t
0
_ t such
that in the period t
0
communication phase , received a consistent vector of i
0
. ,
0
reports
that does not equal (.
i
0
;j
0
;t
0 )
fi
0
;j
0
g2L
. I consider three cases, deriving a contradiction in each:
Case 1: Player i is a leaf player. For player , to be at an on-path history with incorrect
reports, it must be that player i sent an o-path signal, transfer, report, or message to
some player ,
0
`
i
, in some period t
0
_ t (note that it is not possible that i sent an
incorrect but on-path report to ,
0
, because the fact that , is is only neighbor in 1
0
implies
that all reports sent from i to ,
0
are o-path). Since i is a leaf player, , C
i;j
0 . So player
, receives message 0 from some player in C
i;j
0 during the period t
0
communication phase.
This contradicts the fact that /
t+1
j
is on-path.
Case 2: Player i is not a leaf player, and history /
t
0
;0
1
is o-path. Let t
0
_ t
0
be the
rst time t such that history /
;0
1
is o-path. I will show that player , does not receive the
$ (4 1,2
t
0
1
) :
0
[2[ transfer in period t
0
, which contradicts hypothesis that history /
t+1
j
is
on-path.
First note that if any non-leaf player has any money at the beginning of period t
0
1, then
/
t
0
1;0
1
is o-path, because if any non-leaf player did not pass on some of any of the transfers
he received in some period t _ t
0
2 then player 1 did not get back either the $:
0
/
1
,2

transfer or the $ (4 1,2

1
) :
0
[2[ transfer in period t (recalling that all leaf players are
following o
PRI
). So by denition of t
0
no non-leaf has any money at the beginning of period
t
0
1.
33
I now claim that the joint money holdings of all non-leaf players at the beginning of
period t
0
is at most $ (4 1,2
t
0
2
) :
0
[2[ (i.e.,

n
j=n
0
+1
:
j
_ (4 1,2
t
0
2
) :
0
[2[). To see
this, suppose that the non-leaf players collectively try to maximize their joint money holdings
in the period t
0
1 communication phase. Note that every dollar that the non-leaf players
do not pass on to a leaf player out of any on-path transfer they receive reduces the size of the
next on-path transfer sent by a leaf player by more than one dollar, and that leaf players do
not send transfers at o-path histories. So the joint money holdings of the non-leaf players
is maximized when they pass on all on-path transfers except the last one, which is of size
$ (4 1,2
t
0
2
) :
0
[2[.
Now if player , receives a transfer of size $ (4 1,2
t
0
1
) :
0
[2[ in period t
0
it must be
that the joint money holdings of the non-leaf players (including player , if he is a non-
leaf player) reaches $ (4 1,2
t
0
1
) :
0
[2[ at some point during period t
0
. However, it can
be seen that the joint money holdings of the non-leaf players at any point in period t
0
is
no more than $ (4 1,2
t
0
2
) :
0
[2[ + (:
0
1) [2[ ,2
t
0
, since they start the period with at
most $ (4 1,2
t
0
2
) :
0
[2[ and can obtain at most $ (:
0
1) [2[ ,2
t
0
more in the course of
the communication phase (by sending $ [2[ ,2
t
0
to player 2 in the appropriate round and
eventually receiving $:
0
[2[ ,2
t
0
from player :
0
). Finally,
_
4 1,2
t
0
2
_
:
0
[2[ + (:
0
1) [2[ ,2
t
0
= 4:
0
[2[
_
1,2
t
0
_
(3:
0
+ 1) [2[
< 4:
0
[2[
_
1,2
t
0
_
2:
0
[2[
=
_
4 1,2
t
0
1
_
:
0
[2[ .
So player , does not receive the $ (4 1,2
t
0
1
) :
0
[2[ transfer in period t
0
.
Case 3: Player i is not a leaf player, and history /
t
0
;0
1
is on-path. If player 1 does not
receive a consistent vector of reports in the period t
0
reporting subphase, then the argument
is as in Case 2. So suppose that she does, and denote this vector by (^ .
i
0
;j
0 )
fi
0
;j
0
g2L
. Note that
it is not possible for all players other than i to have the same consistentbut incorrect
vector of reports at the start of the period t
0
conrmation phase (as if ^ .
i
0
;j
0 ,= .
i
0
;j
0 then
players i
0
and ,
0
cannot have consistent vector (^ .
i
0
;j
0 )
fi
0
;j
0
g2L
). So there is some player i
0
34
who at the start of conrmation phase is either o-path or is on-path with consistent vector
(~ .
i
0
;j
0 )
fi
0
;j
0
g2L
,= (^ .
i
0
;j
0 )
fi
0
;j
0
g2L
. Let / be number associated to (^ .
i
0
;j
0 )
fi
0
;j
0
g2L
, let /
0
,= /
number associated to (~ .
i
0
;j
0 )
fi
0
;j
0
g2L
. Consider two cases:
1. / < /
0
: I claim that the rst transfer player i
0
receives in the period t
0
conrmation
phase is o-path. To see this, note that no non-leaf player begins period t
0
with
money (because /
t
0
;0
1
is on-path), so for no joint strategy of the non-leaf players is their
joint money holding at the round where player i
0
receives her rst on-path transfer
greater than /,2
t
0
times the number of leaf players who send transfers prior to this
round. Since i
0
expects a transfer of /
0
,2
t
0
times this number (unless she is o-path
already), the rst transfer she receives is o-path. Consequently, player i
0
punishes
the player ,
0
who sends the o-path transfer, and in particular never again transfers
money to him or to any other player in C
i
0
;j
0 . Since there are no transfers along links
not in 1
0
even o-path, this implies that the player 1 does not receive her expected
$:
0
/,2
t
0
transfer, and therefore player 1 does not send the $
_
4 1,2
t
0
1
_
:
0
[2[ transfer.
Finally, as argued in Case 2, the non-leaf players can collectively obtain no more than
$ (:
0
1) [2[ ,2
t
0
< $
_
4 1,2
t
0
1
_
:
0
[2[ in the course of the period t
0
conrmation
phase, so it follows that player , does not receive the $
_
4 1,2
t
0
1
_
:
0
[2[ transfer, a
contradiction.
2. / /
0
: If the rst transfer player i
0
receives in the period t
0
conrmation phase is
o-path, the argument is as in Case 2. If it is on-path, then it equals /
0
,2
t
0
times
the number of leaf players who send transfers prior to this round. Now let | be a leaf
player such that i
0
is on unique path from | to 1 in 1
0
, with | = i
0
if i
0
is a leaf player
herself. Then the rst transfer | receives in the period t
0
conrmation phase is at most
/
0
,2
t
0
times the number of leaf players who send transfers prior to this round. So |
then sends a transfer that is at most /
0
,2
t
0
greater than the transfer she received in
the preceding round. I now claim that player 1 does not receive her expected $:
0
/,2
t
0
transfer in the period t
0
conrmation phase. For the non-leaf players begin period t
0
with no money, no leaf player sends a transfer that is more than /,2
t
0
greater than
what she received, and some leaf player (player |) sends a transfer that is only at most
35
/
0
,2
t
0
greater than what she received. The rest of the argument is as in Case 2.
Lemma 1 is not quite enough to rule out on-path deviations. The following lemma will
also be needed.
Lemma 2 Suppose that under strategy prole
_
o
i
. o
PRI
i
_
an o-path action history /
t
j
is
reached for some , `
i
. Then o
PRI
j
0
;i
_
/
t
0

j
0
_
= c

j
0
;i
for all t
0
t and all ,
0
`
i
.
Proof. I rst note that it suces to show that if /
t
j
is o-path for some , `
i
then the
next action history /
t+1
j
0
is o-path for all ,
0
`
i
. For if /
t+1
j
0
is o-path then the rst
o-path signal, transfer, report, or message received by player ,
0
must have come from a
player ,
00
C
i;j
0 , and since ,
0
`
i
it follows that i C
j
0
;j
00 , and hence o
PRI
j
0
;i
_
/
t
0

j
0
_
= c

j
0
;i
for all t
0
t.
I now show that /
t+1
j
0
is o-path for all ,
0
`
i
, considering three cases.
Case 1: Player i is a leaf player. Since /
t
j
is o-path, player i sent an o-path signal,
transfer, report, or message to some player ,
0
`
i
at some time t
0
< t. Since i is a leaf
player, `
i
_ C
i;j
0 , so every player ,
00
`
i
receives message 0 in the period t
0
communication
phase. So /
t+1
j
00
is o-path for all ,
00
`
i
.
Case 2: Player i is not a leaf player and history /
t;0
1
is o-path. The same argu-
ment as in Case 2 of the proof of Lemma 1 implies that no player ,
0
`
i
receives the
$ (4 1,2
t
0
1
) :
0
[2[ transfer in the rst period t
0
at which /
t
0
;0
1
is o-path, so /
t+1
j
0
is
o-path for all ,
0
`
i
.
Case 3: Player i is not a leaf player and history /
t;0
1
is on-path. Since i is the lone
deviator and an o-path history /
t
j
is reached, i must have sent an o-path signal, message,
report, or transfer to some player ,
0
`
i
at some time t
0
< t. Note that 1 , C
i;j
0 , as
otherwise 1 would have received message 0 in the period t
0
communication phase and hence
/
t;0
1
would be o-path. Now since 1
0
is a spanning tree of 1 (so that in particular 1 does
not lie on a path from i to ,
0
in 1
0
) there is a leaf player | such that the unique path from 1
to | in 1
0
includes i and ,
0
(where it may be that ,
0
= |). Now no transfer from | reaches 1
in the period t communication phase, as if i comes before ,
0
in the path from 1 to | then any
transfer sent by | does not reach 1, and if ,
0
comes before i then no transfer reaches | (as no
36
transfers are made outside of 1
0
and ,
0
does not transfer money to any player in C
i;j
0 ). So,
recalling that no non-leaf player has money at the start of period t and that no leaf players
deviate, it follows that player 1 does not receive her expected $:
0
/,2
t
transfer in period t.
Hence, no one receives the $ (4 1,2
t1
) :
0
[2[ transfer in period t, and therefore /
t+1
j
00
is
o-path for all ,
00
`, and hence for all ,
00
`
i
.
Together, Lemmas 1 and 2 imply that there are no protable deviations at on-path histo-
ries: It is clear that there are no protable deviations at on-path action histories, as playing
any action c
i
at an on-path action history /
t
i
under o
PRI
yields the same continuation payo
as does playing action c
i
at history
^
/
t
i
under o
PUB
. Now suppose, toward a contradiction,
that a player i has a protable deviation at an on-path period t communication history. By
Lemmas 1 and 2, such a deviation must lead some of is neighbors to start minmaxing i in
period t + 1 and lead the rest of them to play o
PUB
j;i
_
^
/
t+1
j
_
in period t + 1 and then start
minmaxing i in period t + 2. Now such a deviation is weakly worse for i than conforming
to o
PRI
in the period t communication phase, deviating to her myopic best response in the
period t + 1 action phase, and playing c

i
from period t + 2 on, since the latter deviation
yields a weakly higher payo in period t + 1 (as best-responding to an arbitrary mixed ac-
tion gives a weakly higher payo than best-responding to the minmax mixed action) and the
same payo in all subsequent periods. But the latter deviation is not protable, since there
are no protable deviations at on-path action histories, so the proposed deviation cannot be
protable, either.
Finally, I argue that there are no protable deviations at o-path histories. Note that
the only path in 1 from a player in `
i
C
i;j
to a player in `
i
C
i;j
is the one through i, so if a
(i. ,)-game strategy o
i;j
maximizes player is payo in the (i. ,)-game for all ,
0
`
i
C
i;j
and
does not reduce the money transfer she receives from any player ,
0
`
i
C
i;j
at any history
then it maximizes her payo overall (for xed o
i;j
). Now at an o-path action-phase history
/
t
i
where o
i;j
is specied, player i plays c

i;j
and believes that every player ,
0
`
i
C
i;j
has
received message 0 from some player in C
i;j
0 (regardless of player is strategy), so she expects
every player ,
0
`
i
C
i;j
to play c

j
0
;i
and make no transfers to her foreverhence, c

i;j
is
optimal. At an o-path communication-phase history where o
i;j
is specied, player i sends
message 0 and does not transfer money to player , and believes that every player ,
0
`
i
C
i;j
37
will receive message 0 from some player ,
0
C
i;j
0 during the communication phase regardless
of player is strategy, by denition of C
i;j
0 and the specication of o-path beliefs, so sending
message 0 and not transferring money to player , is optimal. Finally, transferring :
i;j
0
to a player , , `
0
i
or transferring :
i;j
,
_
0.
1
2
t
.
2
2
t
. . . . .
n
0
jZj
2
t
.
_
4
1
2
t1
_
:
0
[2[
_
to any player ,
in period t leads all ,
0
C
i;j
to play c

j
0
;i
and make no transfers to i forever, so it is optimal
for player i to never make such a transfer.
11.5 Proof of Proposition 1 (Sketch)
Let 1
0
be such a spanning tree, and renumber the leaf players in 1
0
by 1. . . . . :
0
, as in the proof
of Theorem 1. Let = min
i2f1;:::;n
0
g
:
0
i
. Add a new redistribution subphase to the start
of the period 0 communication phase. In it, all non-leaf players rst pass all their money to
player 1. Let r =

n
i=n
0
+1
:
0
i
be the joint initial money holding of the non-leaf players, so
that player 1 receives $r. The rest of the strategy prole is as in the proof of Theorem 1, ex-
cept that throughout $/
i
,2
t
is replaced with $
k
i
2
t
_
"
4n
0
jZj
_
and $ (4 1,2
t1
) :
0
[2[ is replaced
with $r + (1 1,2
t+1
) , reecting the fact that players 2. . . . . :
0
now end the redistribu-
tion subphase with as little as $ rather than $4:
0
[2[ and player 1 ends the redistribution
subphase with as little as $r + rather than $4:
0
[2[.
The proof that this is a SE prole is as in the proof of Theorem 1. In particular,
the facts that non-leaf players end the redistribution subphase with no money and that the
conrmation transfer $r+(1 1,2
t+1
) is greater than $r and increases each period imply
that no player can mislead another about the history of signals.
11.6 Proof of Theorem 2
I rst prove the result for essential, and then describe how it must be modied for strongly
essential.
I start by introducing the notion of an `-local public equilibrium (`-LPE), where ` is
an arbitrary subnetwork of 1. This is dened to be a SE in
PRI
in which o
i;j
(/
t
i
) depends
only on (.
i;j;
)
t1
=0
for all i. , `, and o
i;j
(/
t
i
) depends only on
_
(c
i;j
0
;
. .
i;j
0
;
)
fi;j
0
g2LnM
_
t1
=0
for all i ` and ,
0
, `. That is, a `-LPE is a SE in which players in ` condition their
38
play in a relationship with another player in ` only on past play in that relationship, and
condition their play in a relationship with a player outside ` only on past play with players
outside `. Denote the `-LPE payo set in
PRI
by 1
MLPE
PRI
.
For the rest of the proof, assume that ` is a nice subnetwork of 1.
First, I claim that 1
PRI
= 1
MLPE
PRI
. The argument adapts the proof of Theorem 5.2 of
Fudenberg and Levine (1994), which shows that the SE payo set and PPE payo set coincide
in repeated games with imperfect public monitoring and a product structure. In particular,
x a SE o in
PRI
, any let i. , `. Because ` is a subtree of 1, player is beliefs at
history /
t
i
about player ,s private history depend only on (c
i;j;
. .
i;j;
)
t1
=0
; this follows from
consistency, because it holds for all completely mixed strategy proles (by Bayes rule) and
equalities are preserved in the limit. Given this, the fact that
PRI
has a product structure
implies that player is beliefs about player ,s private history depend only on (.
i;j;
)
t1
=0
(again by consistency, since for any completely mixed strategy prole player is posterior
over (c
j;i;
)
t1
=0
given (c
i;j;
. .
i;j;
)
t1
=0
depends only on (.
i;j;
)
t1
=0
, by Bayes rule). Now replace
o
i;j
with a strategy that depends only on (.
i;j;
)
t1
=0
but has the same marginals over
i;j
conditional on (.
i;j;
)
t1
=0
as does o
i;j
. Do this for every i. , `. In addition, again
because ` is a subtree of 1, for any i. , 1 with i ` and , , `, player is beliefs at
history /
t
i
about player ,s private history depends only on
_
(c
i;j
0
;
. .
i;j
0
;
)
fi;jg2LnM
_
t1
=0
. For
any such i. ,, replace o
i;j
with a strategy that depends only on
_
(c
i;j
0
;
. .
i;j
0
;
)
fi;jg2LnM
_
t1
=0
but has the same marginals over
i;j
conditional on
_
(c
i;j
0
;
. .
i;j
0
;
)
fi;jg2LnM
_
t1
=0
as does o
i;j
.
Then the resulting strategy prole (after both kinds of replacements) is a `-LPE with the
same payos as o; this is because for every pure strategy of any player i, she faces the same
distribution over outcomes whether her opponents follows the original strategy prole or the
modied strategy prole.
Second, I claim that 1
MLPE
PRI
= 1
LPE
PRI
[
M
+ 1
PRI
[
LnM
.
38
To see this, given a LPE o
0
in

PRI
[
M
and a SE o
00
in
PRI
[
LnM
, dene a strategy prole o in
PRI
by letting o
i;j
(/
t
i
) =
o
0
i;j
_
(.
i;j;
)
t1
=0
_
if i. , ` and o
i;j
= o
00
i;j
_
_
(c
i;j
0
;
. .
i;j
0
;
)
fi;jg2LnM
_
t1
=0
_
if i. , 1`.
Then it is straightforward to check that o is a ` -LPE in
PRI
, so 1
MLPE
PRI
_ 1
LPE
PRI
[
M
+
1
PRI
[
LnM
. Similarly, given a `-LPE in
PRI
, o, dene strategy proles o
0
in
PRI
[
M
and o
00
38
The notation here is that for sets , 1 _ R
n
, +1 = a +/ : a , / 1.
39
in
PRI
[
LnM
by o
0
i;j
(/
t
i
) = o
i;j
(/
t
i
) for all i. , ` and o
00
i;j
(/
t
i
) = o
i;j
(/
t
i
) for all i. , , `.
Then o
0
is a LPE in
PRI
[
M
and o
00
is a SE in
PRI
[
LnM
, so 1
MLPE
PRI
_ 1
LPE
PRI
[
M
+1
PRI
[
LnM
.
Combining the inclusions yields 1
MLPE
PRI
= 1
LPE
PRI
[
M
+ 1
PRI
[
LnM
.
Third, 1
$
PRI
[
M
_ 1
PUB
[
M
by Theorem 1. In addition, 1
PUB
[
M
co
_
1
LPE
PRI
[
M
_
,= O
because ` is nice, so 1
$
PRI
[
M
co
_
1
LPE
PRI
[
M
_
,= O. In addition, 1
$
PRI
[
M
_ 1
PRI
[
M
by the
observation preceding Corollary 1, so 1
$
PRI
[
M
_ 1
LPE
PRI
[
M
.
Finally, given any message set and vector of initial money holdings
_
~
1 . ~ :
0
_
in
PRI
[
M
,
dene message set and initial money holdings (1. :
0
) in
PRI
by 1
i;j
=
~
1
i;j
if i. , `,
1
i;j
= O if i. , , `, :
0
i
= ~ :
0
i
if i `, and :
0
i
= 0 if i , `. Then 1
$
PRI
(1. :
0
) _
1
$
PRI
_
~
1 . ~ :
0
_
[
M
+1
PRI
[
LnM
, as given a SE o
0
in 1
$
PRI
_
~
1 . ~ :
0
_
[
M
and a SE o
00
in 1
PRI
[
LnM
one can construct a SE o in 1
$
PRI
(1. :
0
) by letting o
i;j
(/
t
i
) = o
0
i;j
_
~
/
t
i
_
if i. , `,
where
~
/
t
i
is derived from /
t
i
by deleting actions, signals, messages, and transfers along links
i. , , `, and letting o
i;j
(/
t
i
) = o
00
i;j
_
^
/
t
i
_
if i. , , `, where
^
/
t
i
is derived from /
t
i
by
deleting actions and signals along links i. , ` and deleting all messages and transfers.
Hence, 1
$
PRI
_ 1
$
PRI
[
M
+ 1
PRI
[
LnM
.
Combining all the inclusions derived in the course of the proof, we have
1
$
PRI
_ 1
$
PRI
[
M
+ 1
PRI
[
LnM
) 1
LPE
PRI
[
M
+ 1
PRI
[
LnM
= 1
MLPE
PRI
= 1
PRI
,
where the strict inclusion uses the fact that, for any sets A, A
0
, and \, if A _ A
0
and
A co(A
0
) ,= O then A + \ ) A
0
+ \ (as can be seen from a separating hyperplane
argument). Therefore, 1
$
PRI
) 1
PRI
.
The proof for strongly essential is almost identical. In place of a `-local public equi-
librium, dene a `-local cheap talk equilibrium to be a SE in
PRICT
PRI
in which players
in ` condition their play (including messages) in a relationship with another player in `
only on past play in that relationship, and condition their play in a relationship with a
player outside ` only on past play with players outside `. Let 1
MLCTE
PRI
be the set of
`-local cheap talk equilibrium payos in
PRICT
PRI
. Then 1
PRICT
PRI
= 1
MLCTE
PRI
by the same
argument as for 1
PRI
= 1
MLPE
PRI
, with the addition that strategies about which message
to send may also need to be replaced by `-local cheap talk strategies with the same mar-
40
ginals. Next, 1
MLCTE
PRI
= 1
LCTE
PRI
[
M
+ 1
PRI
[
LnM
by the same argument as for 1
MLPE
PRI
=
1
LPE
PRI
[
M
+1
PRI
[
LnM
, and 1
$
PRI
[
M
co
_
1
LCTE
PRI
[
M
_
,= O and 1
$
PRI
[
M
_ 1
LCTE
PRI
[
M
by the same
argument as for 1
$
PRI
[
M
co
_
1
LPE
PRI
[
M
_
,= O and 1
$
PRI
[
M
_ 1
LPE
PRI
[
M
(where the statement
that 1
PUB
[
M
co
_
1
LPE
PRI
[
M
_
,= O is strengthened to 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,= O, which is
possible when ` is truly nice). Combining these inclusions with 1
$
PRI
_ 1
$
PRI
[
M
+1
PRI
[
LnM
as in the essential case yields 1
$
PRI
) 1
PRICT
PRI
.
11.7 Proof of Theorem 3
It is straightforward to verify that any network containing a subtree of size at three also
contains a subtree of size exactly three. I claim that if
i;j
is -dense for small enough
then any subtree of size exactly three is truly nice.
39
The result then follows from Theorem
2.
To prove the claim, x a subtree ` of size three, which can always be written as ` =
(1. 2) . (2. 3) for 1. 2. 3 `. Note that perfect monitoring games have a product structure,
so to show that ` is truly nice I must only show that 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,= O. I will
do this by showing that player 1s greatest SE payo in
PUB
[
M
is greater than her greatest
LCTE payo in
PRI
[
M
.
I rst derive an upper bound on player 1s greatest LCTE payo in
PRI
[
M
, which I
denote by n
PRI
1
; that is, n
PRI
1
= sup n
1;2
, where the supremum is taken over LCTE in

PRI
[
M
. I claim that n
PRI
1
is bounded from above by the solution to the program
max
a
1
;a
2
1
2
, (c
1
. c
2
) c
1
:.t.
1
2
(, (c
1
. c
2
) (1 o) , (c
1
. 0)) c
2
_ 0,
which is well-dened under our assumptions on ,. Call this Program 1. To see this, let c
i
39
In fact, if
i;j
is --dense for small enough - then any subtree of size at least three is truly nice. But it
is slightly simpler to focus on subtrees of size exactly three.
41
be player is expected discounted future action toward player , ,= i, for i. , 1. 2:
c
i
= (1 o)
1

s=0
o
s
1
_
c
i;t+s
[ (c
i;
. c
j;
)
t1
=0

.
40
By concavity of , and Jensens inequality, player is continuation payo in the (i. ,)-game
from conforming is at most
1
2
, ( c
i
. c
j
) c
i
. In addition, at some history player is contin-
uation payo in the (i. ,)-game from deviating to playing c
i
= 0 forever must be at least
(1 o)
1
2
, (0. c
j
).
41
Hence, in any LCTE it follows that
(1 o)
1
2
, (0. c
j
) _
1
2
, ( c
i
. c
j
) c
i
,
or
1
2
(, ( c
i
. c
j
) (1 o) , (0. c
j
)) c
i
_ 0.
Consequently, n
PRI
1
is bounded from above by the solution to Program 1.
Next, I show that player 1s greatest SE payo in
PUB
[
M
is greater than the solution to
Program 1 if
i;j
is -dense for small enough . I consider grim-trigger strategies proles
in
PUB
[
M
: given c
1
. c
2;1
. c
2;3
. c
3
R
+
, let
o
1;2
_
/
t
1
_
=
_
_
_
c
1
if (c
1;2;
= c
1
. c
2;1;
= c
2;1
. c
2;3;
= c
2;3
. c
3;2;
= c
3;2
) for all t < t,
0 otherwise
o
2;1
_
/
t
2
_
=
_
_
_
c
2;1
if (c
1;2;
= c
1
. c
2;1;
= c
2;1
. c
2;3;
= c
2;3
. c
3;2;
= c
3;2
) for all t < t,
0 otherwise
o
2;3
_
/
t
2
_
=
_
_
_
c
2;3
if (c
1;2;
= c
1
. c
2;1;
= c
2;1
. c
2;3;
= c
2;3
. c
3;2;
= c
3;2
) for all t < t,
0 otherwise
o
3;2
_
/
t
3
_
=
_
_
_
c
3
if (c
1;2;
= c
1
. c
2;1;
= c
2;1
. c
2;3;
= c
2;3
. c
3;2;
= c
3;2
) for all t < t,
0 otherwise
.
It is straightforward to check that this strategy prole is a SE if and only if c
1
. c
2;1
. c
2;3
. c
3

40
Note that this expectation is conditioned on information that is public to i and ,.
41
Technically, it must be at least (1 c) ) ( a
2
) j, for arbitrary j 0. This holds because a
2
is bounded
above in any SE. In any case, the following displayed inequalities are correct.
42

i;j
and
1
2
(, (c
1
. c
2;1
) (1 o) , (0. c
2;1
)) c
1
_ 0
1
2
(, (c
3
. c
2;3
) (1 o) , (0. c
2;3
)) c
3
_ 0
1
2
(, (c
2;1
. c
1
) + , (c
2;3
. c
3
) (1 o) (, (0. c
1
) + , (0. c
3
))) (c
2;1
+ c
2;3
) _ 0.
I now maximize n
1;2
=
1
2
, (c
1
. c
2;1
) c
1
over the inequality constraints, temporarily ignoring
the constraint that c
1
. c
2;1
. c
2;3
. c
3

i;j
. It is clear that this is achieved by rst maximizing
1
2
(, (c
2;3
. c
3
) (1 o) , (0. c
3
))c
2;3
over (c
2;3
. c
3
) subject to
1
2
(, (c
2;3
. c
3
) (1 o) , (c
2;3
. 0))
c
3
_ 0, and then maximizing
1
2
, (c
1
. c
2;1
) c
1
over (c
2;1
. c
1
) subject to the remaining in-
equality constraints. Denote the solution to the rst maximization problem by
_
c

2;3
. c

3
_
,
and let =
1
2
_
,
_
c

2;3
. c

3
_
(1 o) , (0. c

3
)
_
c

2;3
be the slack in player 2s incen-
tive constraint coming from his relationship with player 3. Note that o
2f
i
(0;0)
f
i
(0;0)
is
a sucient condition for 0, as it implies the existence of a number c
i;j
such that
1
2
(, (c
i;j
. c
i;j
) (1 o) , (0. c
i;j
)) c
i;j
0. The reduced program is
max
a
1
;a
2;1
1
2
, (c
1
. c
2;1
) c
1
:.t.
1
2
(, (c
1
. c
2;1
) (1 o) , (c
1
. 0)) c
2;1
+ _ 0.
Call this Program 2. It is now clear that the value of Program 2 is greater than the value of
Program 1, and hence greater than n
PRI
1
; in particular, letting (c

1
. c

2
) denote the solution
to Program 1, setting c
1
= c

1
and c
2;1
= c

2
+ satises the constraints of Program 2 and
yields a strictly higher value in the objective. Finally, letting
_
c

1
. c

2;1
. c

2;3
. c

3
_
denote
the solution to Program 2, it is clear that there is a set of vectors (c
1
. c
2;1
. c
2;3
. c
3
) includ-
ing
_
c

1
. c

2;1
. c

2;3
. c

3
_
with non-empty interior such that for any such vector
1
2
, (c
1
. c
2;1
)
c
1
n
PRI
1
and the resulting grim-trigger strategy prole satises the inequality constraints.
Hence, if
i;j
is -dense for small enough , then, noting that c

1
. c

2;1
. c

2;3
. c

3
_ ^ c
i;j
(as can
be checked), there exists a vector (c
1
. c
2;1
. c
2;3
. c
3
) with c
1
. c
2;1
. c
2;3
. c
3

i;j
such that the
resulting grim-trigger strategy prole is a SE and
1
2
, (c
1
. c
2;1
) c
1
n
PRI
1
.
43
11.8 Proof of Proposition 2
I start by showing that there exists (1. :
0
) such that 1
$
PRI
(1. :
0
) _ 1
PUB
, that is, showing
that Theorem 1 applies to the trading favors game. The proof is almost the same as proof
of Theorem 1, but with dierent play in the communication phase.
Let 1
i;j
= (`
2
O. 0. 1) ' 0. Let :
0
i
= 8:
0
[`
2
[ for all i 1. . . . . :
0
and :
0
i
= 0
for all i :
0
+ 1. . . . . :. Play in the communication phase is as follows:
Reporting Subphase: There are up to 3 (: 1) rounds of reporting in which players re-
port whether the last favor was done and who currently has the opportunity to do
a favor. Suppose the current opportunity is for i to do a favor for ,. Let t

i
be
last time at which player i observed an opportunity to do a favor or sent or received
a message, and suppose that player i observed or was told that the opportunity was
for i
0
i
to do a favor for ,
0
i
; if player i did not receive a report on this, player i is at
an o-path history and play is described below. In round 1 player i sends message
_
,
i;j
. .
fi
0
;j
0
g
i
;t

i
' O
_
, where she sends .
fi
0
;j
0
g
i
;t

i
0. 1 if i
0
. ,
0

i
i. , . (,. i)
and sends O otherwise; here, ,
i;j
indicates that the current opportunity is for i to do
a favor for ,, . = 1 indicates that the previous favor was done, and . = 0 indicates
that it was not done. In future rounds, each player i
00
passes on
^
,
i;j
, adds report
.
fi
0
;j
0
g
i
00 ;t

i
00
if i
0
. ,
0

i
00
(i
00
. ,
00
) . (,
00
. i
00
), and passes on ^ .
fi
0
;j
0
g
i
00 ;t

i
00
if he has received
consistent reports about .
fi
0
;j
0
g
i
00 ;t

i
00
. Each player sends a message according to this
protocol 2 (: 1) times (so there at most 3 (: 1) rounds total). Note that on-path
all players are informed of whether the last favor was done and of the identity of the
current favor opportunity. Also, all players can infer the time and identity of the
previous favor opportunity, as all players receive messages whenever there is a favor
opportunity. In particular, all players are informed of everything they would observe
in
PUB
.
Conrmation Subphase: Assign a number in 1. . . . . [2`
2
[ to every element of `
2
0. 1,
corresponding to whether the previous favor was done and who currently has the
opportunity to do a favor. Play in the conrmation subphase is as in the proof of
Theorem 1, with transfers now corresponding to elements of `
2
0. 1 rather than
44
elements of 2.
O-Path Play: As in Theorem 1, but replace play c

i;j
0 in every subsequent period with
never do another favor for ,
0
.
The proof that this is a SE prole is similar to the proof of Theorem 1 and is omitted.
In particular, note that the only action histories of a player , are ones where he has the
opportunity to do a favor, and player , initiates a communication phase at these histories.
So if player , does not receive a message in some communication phase, he learns about this
before doing another favor. Hence, Lemmas 1 and 2 continue to hold: for any deviator i
and player ,, , either plays toward i as he would had i conformed or he minmaxes i, and if
an o-path history for , is reached then no player does another favor for i.
Now, given the result that 1
$
PRI
_ 1
PUB
, the proof that money is strongly essential if 1
contains a truly nice subnetwork follows is nearly identical to the proof of Theorem 2, and
is omitted.
11.9 Proof of Theorem 4
I claim that 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,= O for any subtree ` = (1. 2) . (2. 3) of size three.
The theorem then follows from Proposition 2.
I show that 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,= O by building on the analysis of the two-player
trading favors game due to Hauser and Hopenhayn (2010). They develop a dierential
characterization of the PPE set in this game. In particular, they show that the Pareto
frontier in the two-player game, \ () (where is player 1s continuation value) is given by
:\ () = max
a
1
;a
2
;v
1
;v
2
; _ v
c
1
/ c
2
c + \
_

1
_
\ () + \
_

2
_
\ () + \
0
() _ (4)
subject to
: = c
2
/ c
1
c +
1
+
2
+ _
=
1
c
1
c
\ () = \
_

2
_
c
2
c.
45
In this formulation,
1
is player 1s continuation value when she does a favor for player 2 (or
reveals an opportunity to do a favor but the public randomization is such that she does not
actually do it),
2
is player 1s continuation value when player 2 does a favor for her, _ is the
drift in player 1s continuation value, and c
i
is the probability that player i does a favor when
she gets an opportunity to do so. The rst constraint is the promise-keeping constraint, and
the second and third constraints are player 1s and 2s incentive constraints, which Hauser
and Hopenhayn show hold with equality at a solution. Hauser and Hopenhayn also show
that \ is weakly concave, that \
0
()
_

b
c
.
c
b

almost everywhere, and that at a solution

c
1
and c
2
are given by
c
1
= min
_

c
. 1
_
c
2
= min
_
\ (0) \ ()
c
. 1
_
,
where
h
= \
1
(0).
42
I now show that 1
PUB
[
M
co
_
1
LCTE
PRI
[
M
_
,= O. I use the following technical lemma about
the two-player game:
Lemma 3 \
0
() <
c
b
for some (0. /).
This lemma would be immediate if \ were strictly concave, but Hauser and Hopenhayn
only establish weak concavity (while conjecturing strict concavity).
Proof of Lemma 3. Suppose, toward a contradiction, that \
0
() =
c
b
for all (0. /).
By symmetry of the two-player game, \
0
_

h
_
=
b
c
. In particular, by weak concavity of
\ there exists

_
/.
h
_
such that \
0
() =
c
b
for all <

and \
0
() <
c
b
for almost
all

. Now let satisfy (0.

) and + c
_

.
h
_
. Then, by (4) and the fact
that \
0
() =
c
b
,
:\ () = c
1
/ c
2
c + \
_

1
_
\ () + \
_

2
_
\ ()

c
/
_
: c
2
/ + c
1
c
1
+
2
+
_
.
42
To be precise, there is a single solution where the incentive constraints bind and r
1
and r
2
are given by
these equations.
46
Now since \ (
0
) =
c
b
for all
0
< , and
2
< , it follows that \ (
2
) \ () =

c
b
(
2
). In addition, c
1
c =
1
. So the above equation becomes
\ () =
c
1
/ + \ (
1
) \ ()
:

c
/
. (5)
Now since + c <
h
, it follows from Hauser and Hopenhayn that c
1
= 1 for all
0
in
a neighborhood of (for some solution), and hence that
01
=
0
+ c

for all such

0
. Since \
0
(
0
) =
c
b
for all
0
<

and \
0
(
0
) <
c
b
for all
0

, it follows that
\ (
01
) \ (
0
) = \ (
0
+ c) \ (
0
) is strictly decreasing with non-vanishing derivative
in a neighborhood of . Therefore, it follows from (5) that \ cannot be dierentiable at
with \
0
() =
c
b
, completing the proof of the lemma.
Turning to the three player game, note that \ (
1
) + \ (
3
) is the Pareto frontier in

PRI
[
M
(restricting to LCTE), and let \ (
1
.
3
) be the Pareto frontier in
PUB
[
M
. I show
that there exist
1
.
3
such that \ (
1
.
3
) \ (
1
)+\ (
3
), which completes the proof. To
prove this, I rst introduce some notation. Let c
1
be the probability that 1 does a favor for
2 and let c
21
be the probability that 2 does a favor for 1, and similarly for c
3
and c
23
. Let
1
1
be player 1s continuation value when 1 does a favor (for 2), let
21
1
be player 1s continuation
value when 2 does a favor for 1, and so on for
23
1
.
3
1
.
1
3
.
21
3
.
23
3
.
3
3
. Now let
~
\ (
1
.
3
) =
\ (
1
) +\ (
3
), ~ c
1
= min
_
v
h
v
1
c
. 1
_
, ~ c
21
= min
_
2W(0)W(v
1
)W(v
2
)
c
. 1
_
, ~
1
1
=
1
+~ c
1
c, ~
3
1
=

1
, ~
21
1
= max \
1
(\ (
1
) + ~ c
21
c) . 0, ~
23
1
= \
1
(\ (
1
) + ~ c
23
c (\ (~
23
3
) \ (
3
))),
and symmetrically for ~ c
3
, ~ c
23
, ~
3
3
, ~
1
3
, ~
23
3
, and ~
21
3
, and nally let
d
~
\ (
1
.
3
) = \
0
(
1
)
_
:
1
~ c
21
/ + ~ c
1
c ~
1
1
~
21
1
~
23
1
~
3
1
+ 4
1
_
+\
0
(
3
)
_
:
3
~ c
23
/ + ~ c
3
c ~
1
3
~
21
3
~
23
3
~
3
3
+ 4
3
_
.
I now claim that there exist
1
.
3
such that
:
~
\ (
1
.
3
) < (~ c
1
+ ~ c
3
) / (~ c
21
~ c
23
) c +
~
\
_
~
1
1
. ~
1
3
_
+
~
\
_
~
21
1
. ~
21
3
_
+
~
\
_
~
23
1
. ~
23
3
_
+
~
\
_
~
3
1
. ~
3
3
_
4
~
\ (
1
.
3
) + d
~
\ (
1
.
3
) . (6)
47
To see this, recall that
:
~
\ (
1
.
3
) = : (\ (
1
) + \ (
3
)) = (c
1
+ c
3
) / (c
21
c
23
) c + \
_

1
1
_
+ \
_

21
1
_
2\ (
1
)
+\
_

23
3
_
+ \
_

3
3
_
2\ (
3
) + \
0
(
1
) _
1
+ \
0
(
3
) _
3
, (7)
by (4). So the claim follows if the right-hand side of (6) is greater than the right-hand side
of (7). Working through the denitions (including observing that ~ c
1
= c
1
, ~
1
1
=
1
1
, ~
3
1
=
3
1
,
and ~
21
1
=
21
1
, and similarly for player 3), this is the case if

\
0
(
1
)
_
~ c
21
/ + ~
23
1
_
+ \
0
(
3
)
_
~ c
23
/ + ~
21
3
_

[\
0
(
1
) (c
2
/ +
1
) + \
0
(
3
) (c
23
/ +
3
)[ .
(8)
Now ~ c
21
_ c
2
, with strict inequality if
1
< / and
3
0, and similarly for ~ c
23
. Let ~ c
23
= c
23
,
and dierentiate the left-hand side of (8) with respect to ~ c
21
at ~ c
21
= c
2
. The derivative
equals
\
0
(
1
) / + \
0
(
3
)
_
c
\
0
(
3
)
_
= \
0
(
1
) / + c.
So taking any
3
0 and any
1
(0. /) such that \
0
(
1
) <
c
b
which exists, by Lemma
3implies that this derivative is negative and that ~ c
21
c
2
, and hence that (8)and
therefore (6)is satised. Therefore, Hauser and Hopenhayns optimality equation (4) (or,
more precisely, its immediate generalization to the three-player game
PUB
[
M
) is not satised
by
~
\, while the constraints are satised are satised by
~
\. So there exist
1
.
3
such that
\ (
1
.
3
) \ (
1
) + \ (
3
), completing the proof.
References
[1] Abreu, D., D. Pearce, and E. Stacchetti (1990), Toward a Theory of Discounted Re-
peated Games with Imperfect Monitoring, Econometrica, 58, 1041-1063.
[2] Ahn, I. and M. Suominen (2001), Word-Of-Mouth Communication and Community
Enforcement, International Economic Review, 42, 399-415.
[3] Ali, S.N. and D. Miller (2012), Ostracism, mimeo.
[4] Aliprantis, C.D., G. Camera, and D. Puzzello (2007), Anonymous Markets and Mon-
etary Trading, Journal of Monetary Economics, 54, 1905-1928.
48
[5] Araujo, L. (2004), Social Norms and Money, Journal of Monetary Economics, 51,
241-256.
[6] Aumann, R.J. and S. Hart (2003), Long Cheap Talk, Econometrica, 71, 1619-1660.
[7] Balmaceda F. and J. Escobar (2011), Self Governance in Social Networks of Information
Transmission, mimeo.
[8] Ben-Porath, E. and Kahneman, M. (1996), Communication in Repeated Games with
Private Monitoring, Journal of Economic Theory, 70, 281-297.
[9] Compte, O. (1998), Communication in Repeated Games with Imperfect Private Mon-
itoring, Econometrica, 66, 597-626.
[10] Ellison, G. (1994), Cooperation in the Prisoners Dilemma with Anonymous Random
Matching, Review of Economic Studies, 61, 567-588.
[11] Fudenberg, D. and D.K. Levine (1994), Eciency and Observability with Long-Run
and Short-Run Players, Journal of Economic Theory, 62, 103-135.
[12] Fujiwara-Greve, T., M. Okuno-Fujiwara, and N. Suzuki (2012), Voluntarily Separable
Prisoners Dilemma with Reference Letters, Games and Economic Behavior, 74, 504-
516.
[13] Hahn, F. (1973), On the Foundations of Monetary Theory, in Essays in Modern
Economics, edited by M. Parkin with A. R. Nobay, New York: Barnes and Noble.
[14] Hauser, C. and H. Hopenhayn (2010), Trading Favors: Optimal Exchange and For-
giveness, mimeo.
[15] Kandori, M. (1992), Social Norms and Community Enforcement, Review of Economic
Studies, 59, 63-80.
[16] Kandori, M. and H. Matsushima (1998), Private Observation, Communication and
Collusion, Econometrica, 66, 627-652.
[17] Kiyotaki, N. and R. Wright (1993), A Search-Theoretic Approach to Monetary Eco-
nomics, American Economic Review, 83, 63-77.
[18] Kocherlakota, N.R. (1998), Money is Memory, Journal of Economic Theory, 81, 232-
251.
[19] Kocherlakota, N.R. (2002), The Two-Money Theorem, International Economic Re-
view, 43, 333-346.
[20] Kocherlakota, N. and N. Wallace (1998), Incomplete Record-Keeping and Optimal
Payment Arrangements, Journal of Economic Theory, 81, 272-289.
[21] Lagos, R. and R. Wright (2008), When is Money Essential? A Comment on Aliprantis,
Camera, and Puzzello, mimeo.
49
[22] Lippert, S. and G. Spagnolo (2011), Networks of Relations and Word-of-Mouth Com-
munication, Games and Economic Behavior, 72, 202-217.
[23] Milgrom, P.R., D.C. North, and B.R. Weingast (1990), The Role of Institutions in
the Revival of Trade: The Law Merchant, Private Judges, and the Champagne Fairs,
Economics and Politics, 2, 1-23.
[24] Mishkin, F.S. (1998), The Economic of Money, Banking, and Financial Markets, 5th
ed., Reading: Addison-Wesley.
[25] Mbius, M.M. (2001), Trading Favors, mimeo.
[26] Okuno-Fujiwara, M. and A. Postlewaite (1995), Social Norms and Random Matching
Games, Games and Economic Behavior, 9, 79-109.
[27] Renault, J., Renou, L., and T. Tomala (2012), Secure Message Transmission on Di-
rected Networks, mimeo.
[28] Renault, J. and T. Tomala (1998), Repeated Proximity Games, International Journal
of Game Theory, 27, 539-559.
[29] Wallace, N. (2001), Whither Monetary Economics? International Economic Review,
42, 847-869.
50