**Roberto Serrano and Oscar Volij** Department of
Economics, Brown University, Providence, RI 02912, USA

**Abstract:**We study a decentralized matching model
in a large exchange economy, in which trade takes place through
non-cooperative bargaining in coalitions of finite size. Under
essentially the same conditions of core equivalence, we show that the
strategic equilibrium outcomes of our model coincide with the Walrasian
allocations of the economy. Our method of proof makes use of the theory
of the core. With respect to previous work, our model relaxes
differentiability and convexity of preferences, and also admits
indivisible goods.

**Keywords:** Finite coalitions; strategic
bargaining; core; Walrasian equilibrium.

**JEL:** D51; D41; C78.

This version: July 1998

The Walrasian or competitive equilibrium is the central solution
concept in economics. However, from its definition it is not clear
what trading procedures lead to Walrasian outcomes. In contrast to
what many economists may think, the original account of the theory
(Walras, (1874)) does not rely on the existence
of the so-called
Walrasian auctioneer.^{1} Nonetheless,
the usual formal presentation of the model includes the auctioneer
implicitly due to a lack of explanation for the formation of equilibrium
prices.

Negishi (1989) distinguishes two major schools in the analysis of markets. One of them considers prices as part of the economic mechanism in and out of equilibrium. This school is attributed to Cournot (1838) and Walras (1874). A second school, associated with Jevons (1879) and Edgeworth (1881), attempts to consider decentralized trading mechanisms and answer the question of whether equilibrium prices will emerge as the consequence of agents' trading actions.

We can distinguish at least two major approaches in the Jevons-Edgeworth school of decentralized trading. One of them, which today is referred to as the core equivalence literature, has its origins in Edgeworth (1881). This approach finds conditions under which core and Walrasian allocations are equivalent. If one conceives the core as a decentralized mechanism, these results give insights about the Walrasian allocations that were not provided by Walras. The shortcoming of this approach is that, although the core captures a natural idea of coalitional stability, it does not specify the trading procedure either. Here is relevant the work on the non-cooperative implementation of the core in finite games and economies (see, for example, Perry and Reny (1994), Serrano (1995) and Serrano and Vohra (1997)).

A second approach models the trading procedure more explicitly and
studies its strategic equilibria. As part of a recent literature that
starts with Rubinstein (1982) and
Rubinstein and Wolinsky (1985)
[henceforth, RW], many researchers have turned to
models where a decentralized trading procedure is made explicit in a
bargaining extensive form.^{2} RW (1985)
analyze an assignment market and claim that
in a frictionless economy the strategic equilibria need not be Walrasian.
This claim was challenged later by the classic paper of Gale (1986a), who
constructed an alternative bargaining procedure in a continuum economy in
which strategic and Walrasian equilibria coincide. Gale's work was
generalized in some respects by McLennan and
Sonnenschein (1991)
[McLS in the sequel].

Combining the above strands of the literature, in this paper we
study a decentralized matching model in a large exchange economy,
in which trade takes place through non-cooperative bargaining
in coalitions of finite size. Under essentially the conditions of core
equivalence, we show that the strategic equilibrium outcomes of
our model coincide with the Walrasian allocations of the economy.
Our model relaxes differentiability and convexity of preferences,
and also admits indivisible goods.^{3}
By considering multilateral meetings in which trade takes place, we
are able to use the full power of the theory of the core. In doing so,
it becomes apparent that the driving force behind core equivalence and
Gale-like results is similar: core equivalence takes now a
non-cooperative dimension.

From an implementation theoretic point of view, we construct a mechanism that fully implements the Walrasian allocations of an economy with a continuum of agents. In contrast to all previous work, though, implementation is achieved over a substantially larger class of economies. Moreover, our mechanism can be seen as an ε-perturbation of Gale's, since meetings involving more than two agents can be assumed to have negligible probability.

Free of the differentiability assumption, our model differs from the
preceding ones in allowing for trade to take place in coalitions with
any finite number of participants, as opposed to only pairwise meetings.
This trade-off in the modelling choice leads to two distinct methods of
proof and captures different insights. The proofs of the existing results
(Gale (1986a, c), McLS (1991)) rely crucially on the existence of the
marginal rate of substitution at every bundle in order to get the unique
(to that bundle) supporting price. Since in the strategic equilibrium of
these models there cannot be any pair of agents with positive gains
from trade, by differentiability, at the equilibrium bundles every
two agents have the same marginal rate of substitution. Consequently,
this must hold for all agents. Feasibility of the equilibrium outcome
(zero aggregate excess demand) and some extra technical assumptions then
take care of the rest of the argument.^{4}
In contrast to this method of proof, Theorem 1
uses the
theory of the core to show that every strategic equilibrium yields
a Walrasian allocation. The existence of marginal rates of
substitution is not necessary. Since we allow for finite
coalitions to meet, in our strategic equilibria there cannot
be any such coalition with positive gains from trade. Then,
a powerful lemma due to Hammond, Kaneko and Wooders
(1989) and core
equivalence take us the rest of the way without the intermediate
aid of differentiability.

Unlike our first result just referred to, its converse, Theorem 2, does not use the theory of the core. That is, although we allow for finite coalitions, it turns out that all equilibrium outcomes of our game can be supported by strategies according to which all trade takes place in pairs. In this sense, pairwise trade arises as an endogenous feature of the procedure instead of being imposed as an exogenous restriction. On the other hand, our trading procedure is indeed decentralized, since finite coalitions are negligible in the continuum. Moreover, if all gains from trade in the economy can be exhausted in coalitions bounded by a given finite size k (e.g. assignment markets, like in RW (1985) in which k=2), one can restrict the matching process to meetings of size k+1.

The paper is organized as follows: Section 2 describes the underlying economic model. The non-cooperative bargaining game is described in Section 3. The main result is presented in Section 4. In Section 5 we show that every Walrasian allocation can be supported by a strategic equilibrium of the game, and Section 6 concludes and compares our results with the previous literature.

Let (A, A*, μ) be a measure space, where
A is the set of agents, A*
is the set of measurable subsets of A and μ
is an atomless measure. We denote by C the set of agents'
characteristics. An element c∈C is a
pair c=(u,e),
where u:X→ℜ is a utility
function, X
is the consumption set and e∈X is an endowment. The
consumption set X is assumed to be identical for
all agents and is of the form
ℜ_{+}^{D}×**N**^{I},
where **N** is the set of non-negative integers.
The consumption set includes |D| divisible goods
and |I| indivisible goods; we assume
that |D|≥1.

An economy E is a measurable map
E:A→C. Let S⊆A
be a coalition. An S-allocation f
is a measurable map f:S→X that satisfies
∫_{S}f(a)dμ≤∫_{S}e(a)dμ.
Allocations are A-allocations. We denote by
F the set of allocations of the economy.

From now on, and whenever there is no danger of confusion, the domain of integration A of the set of all agents will be omitted.

A coalition S can improve upon an allocation
f if S has a positive measure and there exists
an S-allocation g:S→X such
that almost everywhere in
S u_{a}(g(a))>u_{a}(f(a)).
The core of an economy is the set of all allocations that no
coalition can improve upon (see Aumann (1964)).

We shall make the following assumptions on the utility functions of the agents. The assumptions can be grouped into two classes. First, we need the assumptions of Hammond, Kaneko and Wooders (1989) that guarantee the validity of their core equivalence theorem.

**A0**All the commodities are present in the economy: ∫e(a)dμ>>0.- For all c=(u,e)∈C:
**A1**The utility function u is continuous, strictly increasing in the divisible commodities, and non-decreasing in the indivisible commodities;**A2**for all (x_{D},x_{I})∈ℜ_{+}^{D}×**N**^{I}, there exists y_{D}∈ℜ_{+}^{D}such that u(y_{D},0_{I})>u(x_{D},x_{I}); and**A3**for all x_{I}∈**N**^{I}, u(e)>u(0_{D},x_{I}).

In addition, we need assumptions that are necessary to deal with the possibly random outcome of the strategic bargaining process.

**A4**The utility functions are bounded: For all c=(u,e)∈C: there exists a number k_{u}such that u(x)≤k_{u}for all x∈X.**A5**The economy satisfies strong aggregate risk aversion in the individually rational domain (defined in the following paragraphs.)

Gale (1986a, b, c) assumes that all individuals have strictly concave utility functions. This assumption is used in order to prove that the outcome of a strategic equilibrium of the market is a degenerate lottery. This assumption excludes the possibility that the consumption sets include indivisible goods. This makes the comparison of his model with the earlier ones (e.g. RW (1985)) difficult. The work of McLS (1991) dispenses with the convexity assumption, but requires differentiability. Therefore, also in their framework indivisible goods cannot be studied. On the other hand, one should expect that the convexification effects of large numbers may help relax the assumption of individual risk aversion of preferences. These arguments motivate our assumption of aggregate risk aversion.

A lottery on allocations is a probability measure on the
Borel σ-algebra of allocations. We shall say that the economy
satisfies weak aggregate risk aversion if for every lottery
L on allocations there exists an allocation g
such that for almost all agents
u_{a}(g(a))≥∫_{F}u_{a}(f(a))dL.

We say that the economy satisfies weak aggregate risk aversion in the individually rational domain if it satisfies weak aggregate risk aversion when considering only lotteries that assign an expected utility at least as high as the utility of the endowment.

A lottery L is degenerate if for almost every
agent a there exists a constant k_{a}
such that
u_{a}(f(a))=k_{a} for L-almost
all f∈F.
The economy satisfies strong aggregate risk aversion
if it satisfies weak aggregate risk aversion and for every lottery which
is not degenerate there exists an allocation g that satisfies
for almost all a∈A
u_{a}(g(a))≥∫_{F}u_{a}(f(a))dL,
and there exists a set of agents of positive measure for whom the
inequality is strict.

Strong aggregate risk aversion in the individually rational domain is defined just like its weak counterpart. Aggregate risk aversion means that society cannot gain from lotteries over allocations. The strong version of the property means that if the lottery is non-degenerate, society actually will lose from having the lottery.

Although this is a property on the aggregate, we find conditions on individual preferences that are weaker than concavity of the utility functions that imply the corresponding aggregate risk aversion properties.

First we consider an individual with a utility function
u:X→ℜ. We define the quasiconcave cover
of u as __u__:__X__→ℜ,
where __X__ is the convex hull of X:

__u__(x)=*max*{u(y): x∈__R__(y)}

where __R__(y) is the convex hull of the set of all
bundles that are weakly preferred to y. The assumption that
the consumption set X is bounded below and assumption A1 on
the utility function u ensure the existence of
__u__, as shown in Einy and Shitovitz
(1997, Lemma 3.3).
See also Starr (1969, Appendix 3) for an early
treatment of non-convex preferences.

The first condition we consider is simply that __u__
is concave for almost all agents.

**Proposition 1**- If the quasiconcave covers of the utility functions of almost all agents are concave, then the economy satisfies weak aggregate risk aversion.

A slightly stronger assumption on individual preferences is that the
quasiconcave cover of the utility function is almost strictly concave.
A function v:X→ℜ is almost strictly concave
if for all x_{1},x_{2}∈X such that
v(x_{1})≠v(x_{2}) and
for all λ∈(0,1) we have that
v(λx_{1}+(1-λ)x_{2}) >
λv(x_{1})+(1-λ)v(x_{2}).
The difference between almost strict concavity
and strict concavity is that the former requires that
v(x_{1})≠v(x_{2}) and not simply that
x_{1}≠x_{2}.
This difference is crucial for our purposes, because
if a cover is strictly concave, the original function is strictly
concave as well, whereas functions that are not even quasiconcave may
have an almost strictly concave cover.

**Proposition 2**- If the quasiconcave covers of the utility functions of almost all agents are almost strictly concave, then the economy satisfies strong aggregate risk aversion.

The two propositions are proved in the appendix. It follows that if the quasiconcave covers (restricted to the individually rational domain) of the utility functions of almost all agents are almost strictly concave, then the economy satisfies strong aggregate risk aversion in the individually rational domain, which is the assumption we need.

**Example** An example will be useful
to clarify our last assumption A5. Consider an individual who may consume
two goods: one of them is perfectly divisible and the other is indivisible,
(like in an assignment market of RW (1985)). The agent wants to
consume at most one unit of the indivisible good and his reservation
price in terms of the divisible commodity for the first unit of the
indivisible good is 1. More formally, the commodity space is
ℜ_{+}×**N**. His utility function is
u(x_{1},x_{2})= v(x_{1}+I(x_{2}))
where I(x_{2})=1 if x_{2}≥1
and I(x_{2})=0 otherwise, and v is a
strictly increasing, strictly concave and bounded function from
ℜ_{+} to ℜ. Suppose agents
differ only in their endowments, i.e., each buyer holds only one unit of
x_{1} and each seller holds
one unit of x_{2}. Since the consumption set
contains indivisible goods, there is no clear notion of risk aversion of
preferences in this setting. However, the quasiconcave cover of
u, restricted to the individually rational domain, is almost
strictly concave:
__u__(x_{1},x_{2})=v(x_{1}+*min*{x_{2},1}).
This implies, for example, that if there is a continuum of such agents
with the same preferences, the economy cannot gain from introducing
lotteries over bundles.

The set A of all agents is present from the outset.
Time runs discretely from 1 to infinity. In each round the
agents are matched at random into coalitions of finite size. At
every round t there is a proportion α∈(0,1)
that is left unmatched. For each round t and each size
n≥1, there is
a positive probability p(n) of being matched in an n-person
coalition. Thus, p(1)=α. Matches are made randomly and
for a fixed n≥2 the probability of being matched
to any n-1 agents chosen from n-1 sets is
proportional to the product of the measures of these
sets.^{5}

When a coalition S meets, there is
a 'cheap talk' phase in which every agent announces a
bundle.^{6} In addition, an order
is chosen at random with equal probability. Denote by
x_{j} the bundle held by agent j∈S
at the beginning of this meeting. The first agent
in the order becomes the proposer, who makes a public offer
consisting of a trade (z_{j})_{j∈S}
in which ∑_{j∈S}z_{j}=0 and for
all j∈S, x_{j}+z_{j}∈X.
Responses are also public and occur sequentially following the order.
They can be one of two possible actions: 'yes,' and 'no.'
The trade proposed to the coalition takes place if and only if
every responder agrees to it. Every agent who is matched
in round t can, if he so wishes, leave the
market and consume his bundle after the bargaining session ends. Agents
who are not matched in round t cannot leave the market in
that round. In the next round, all agents who
chose to leave abandon the market place and consume their
current bundle. All other agents continue as active traders ready
to be matched again.

In each round each agent recognizes the economic characteristics of the agents with whom he is matched. These consist of the current bundle each of them holds and their utility functions. However, they do not have information about their histories: each agent remembers only his own, but not the others'. They do not know anything about meetings that do not include them.

The restrictive information available to traders requires us to endow agents with beliefs about what happens elsewhere in the market. This must be done in order to have well defined expected utility computations. We will take care of these details in the next section, where we present our equilibrium notion.

The payoff to a typical trader in this market is the utility of the bundle with which he leaves the market. Thus, there is no discounting. On the other hand, if an agent never leaves the market, his utility is the utility corresponding to the zero bundle. All agents are expected utility maximizers when evaluating lotteries over bundles.

A strategic equilibrium is a particular type of perfect Bayesian equilibrium (see Fudenberg and Tirole (1991) for the general notion of PBE), i.e., it is a profile of strategies, one for each agent, such that, given the beliefs explained below, every agent plays a best response to the others at every information set. On the equilibrium path, we shall assume that beliefs are derived from the equilibrium strategies using Bayes' rule. On the other hand, off the equilibrium path we shall assume that each agent believes that a full measure of the agents continues to play according to the equilibrium. This equilibrium concept is motivated and formally defined in the sequel.

Since each agent is an entity of measure 0 in the
continuum and since each of them has met only a
finite number of agents in all the rounds up to round t,
we can define the variable of the "state of the market."
That is, a fixed profile of strategies played by the continuum of
agents determines the state of the market in round t as a
distribution over characteristics. This happens with independence of
the actions of a set of measure 0
(the history of an agent at a given point). Notice
that the distribution over characteristics that we refer to as the
'state of the market' need not be supported by an allocation
of the economy. Such an example can be constructed following the
one found in Kannai (1970). However, a
distribution is all an expected
utility maximizing agent needs in order to make his
calculations.^{7}

Now we can state more formally the equilibrium concept based on Osborne and Rubinstein (1990) as follows:

**Definition**- A strategic equilibrium is a pair of functions
(σ,β), that assign to each agent a strategy
and a belief, respectively, such that:
- The beliefs β are the "state of the market" beliefs induced by σ both on and off the equilibrium path,
- For each agent a and for each of his information sets, σ(a) prescribes a best response to σ(A\a) given the beliefs β(a).

Our main result follows.

**Theorem 1**- Suppose that the economy satisfies assumptions A0-A5. In every strategic equilibrium there exists a core allocation f such that almost every agent a eventually leaves the market with a bundle g(a) such that u(g(a))=u(f(a)).

Note that in equilibrium an agent may receive a lottery over different bundles (due to the random matching process and possibly to mixed strategies, the same equilibrium comprises different paths). However, all of these bundles belong to the same indifference surface. If there were several core allocations that assign the same utility to almost every agent, then the outcome could be a lottery over these core allocations.

**Proof** Consider a strategic equilibrium. All of our
statements are relative to this equilibrium and to
histories in which at most a set of agents of measure 0
has deviated. Since each agent's history is private
information and as the matching process treats all
agents alike, two agents with the same characteristics and beliefs
must get the same payoff regardless of their histories.
If not, the agent with the lower payoff would simply
imitate the behavior of the other and get the same probability
distribution over outcomes. Recall that all agents have the
same beliefs about the "state of the market" independent of
their histories.

All agents at the beginning of round t before their match has been determined believe that the "state of the market" corresponds to the equilibrium. Thus in the equilibrium all such agents that in addition share the same characteristics have the same expected utility. We denote this utility by V(c,t). For each c=(u,e), we define w(c)=u(e).

*Step 1*: V(c,t)≥w(c) for all values
c and t.

To see this, notice that every agent with characteristics c in period t can adopt the following strategy: whenever matched, propose the zero trade, reject any trade, and leave the market. Since with probability 1 he will be matched in finite time, this strategy guarantees him a payoff of w(c).

*Step 2*: V(c,t)≥V(c,t+1) for all values of
c and t.

This assertion follows from the fact that by proposing the null trade and rejecting every offer and staying in the market, any agent in the market in round t is sure to be in the market in round t+1 with the same bundle as in round t.

We shall say that an agent is 'about to leave the market' if he has already reached an information set at which his strategy tells him to leave.

*Step 3*: For an agent of characteristic c
who is about to leave the market in round t, we have that
V(c,t+1)=w(c).

By Step 1, we have that V(c,t+1)≥w(c). If V(c,t+1)>w(c) and given that this agent is about to leave the market, he would be better off by deviating and staying in the market until round t+1.

*Step 4*: At some round t there is a
positive measure of agents who are about to leave the market.

We argue by contradiction. Suppose that no positive measure of agents ever leaves the market. In this case the utility of almost all agents is that of the zero bundle. On the other hand, at any point in time there is a positive measure of agents who hold a bundle different from the zero bundle. This contradicts Step 1.

At this point our method of proof separates crucially from Gale's. Instead of proceeding through a sequence of lemmas based on the existence of the marginal rates of substitution, we employ the insights of the theory of the core.

*Step 5*: There does not exist a coalition
S∈A* with μ(S)>0,
that has an S-allocation g for which
u_{a}(g(a))>V(c(a),1) for almost all
a∈S, where c(a) is the initial
characteristics of agent a.

Assume there exist such a coalition and such an allocation.
Then, by Hammond, Kaneko and Wooders
(1989, Claim 1) there exists
a partition of this coalition into h+1 coalitions
S_{0},S_{1},…,S_{h}
such that μ(S_{1})=μ(S_{2})=…=μ(S_{h})>0
and a list of trades z_{1},…,z_{h}
such that for all a∈S_{m}, m=1,…,h
u_{a}(e(a)+z_{m})>u_{a}(g(a))
and ∑_{m}z_{m}≤0. Informally, this
means that there are 'many' h-person (finite) coalitions
that can improve upon the S-allocation g.

By step 4, at some round t a positive measure of agents
is about to leave the market. We will show now that, under the
contradiction hypothesis we are making, i.e., the
existence of the improving coalition S, Step 3 would be
violated. Recall that in every round t a positive measure
of agents did not trade yet and thus keep their initial endowment.
In particular, there is a positive measure of agents that can be chosen
from each of the above coalitions S_{1},…,S_{h}.
It follows from Step 2 that in every round t the probability
of an agent to be matched in an h+1-person coalition such
that his h partners constitute an improving coalition is
positive. Now each person who is about to leave the market can adopt the
following strategy: to stay in the market and whenever being a proposer
in such an improving coalition, to offer them an improving trade z
which gives the proposer higher amounts of some divisible goods
without giving away any amount of the others; in all other situations, he
holds on to his bundle and leaves the market at some finite date.
The proposal z will be unanimously accepted by the members
of the improving coalition since for each of them with characteristic c=(u,e)
we have that:

V((u,e+z_{c}),t+1)≥u(e+z_{c})>V(c,1)≥V(c,t+1),

where the first inequality follows from Step 1, the second from the existence
of the improving trade z for the group of responders and the
third from step 2. Clearly, this deviation gives the deviating agent
a higher expected utility than the utility of his current bundle, which
contradicts Step 3.

*Step 6*: In a strategic equilibrium, there exists a core
allocation f such that almost every agent a
eventually leaves the market with a bundle g(a) such that
u(g(a))=u(f(a)).

By the previous step applied to the set of all agents A, it is not true that A can improve upon the strategic equilibrium outcome. In addition, this must be such that almost every agent is receiving a degenerate lottery over bundles. Otherwise, since the economy satisfies strong aggregate risk aversion, Step 5 would be violated. In addition, the equilibrium outcome satisfies all the core conditions. In order to choose the core allocation that is utility equivalent to the equilibrium outcome, note that almost every path in the extensive form associated with the equilibrium strategies constitutes a core allocation. Thus, it suffices to choose any such path.

**Corollary 1**- Suppose that the economy satisfies assumptions A0-A5. In every strategic equilibrium there exists a Walrasian price p such that almost all agents eventually leave the market with a bundle that maximizes their utility on the budget set corresponding to the price p and to their initial endowment e.

**Proof:** By
Hammond, Kaneko and Wooders (1989,
Theorem 2), A0-A3 imply that core allocations are Walrasian.

**Remark:** If agents are allowed to
observe the state of the market every period,
Theorem 1 and its proof
go through without change when the solution concept used is the
unrestricted set of perfect Bayesian equilibria.

In Corollary 1 we have shown that all strategic equilibrium outcomes are Walrasian. Next we show the converse. That is, for every Walrasian outcome we find strategies that support it as a strategic equilibrium. In the proof below, no use will be made of the fact that the Walrasian allocation f is also a core allocation. This departs from the papers on implementation of the core in finite games and economies (e.g., Perry and Reny (1994), Serrano (1995) and Serrano and Vohra (1997)). There, the strategies that support each core allocation are of the following form: the allocation in question is proposed to the grand coalition, and it is unanimously accepted. Obviously, such a construction is not possible in a decentralized model of trade among a continuum of agents.

**Theorem 2**- Let f be a Walrasian allocation corresponding to an equilibrium price p and suppose that all agents have a maximizer over every budget set corresponding to the price vector p. Then, there exists a strategy profile that supports f as a strategic equilibrium outcome of the game.

The assumption in Theorem 2 is needed in our model as we can have Walrasian equilibria with some prices equal to 0. In this case, it could be that some agents (that constitute a set of measure 0) do not have an optimal bundle in their budget sets even in the equilibrium allocation. Of course, if equilibrium prices are all positive, or the model is a standard assignment market, (e.g. like in RW (1985)), the assumption in Theorem 2 is satisfied. In the models of Gale (1986a, b, c) and McLS (1991), such an assumption is also needed since they work with open consumption sets.

**Proof:** For each agent a let
h(a,e) be a function that assigns to each agent
a with holdings e a bundle selected from
the agents' demand correspondence with respect to the price p
and the income pe. The selection h is
restricted so that h(a,e)=f(a) if pe=pe(a).
Now consider the following strategy profile:

- In all meetings every trader announces during the 'cheap talk' phase the bundle assigned to him by the selection h.
- In multilateral meetings (those with at least three agents), the proposer offers the 0 trade.
- In bilateral meetings, the proposer offers a trade according to the trading rule g defined below (which is based on Gale (1986b)).
- In all meetings, every responder a who currently holds e that did not achieve the bundle h(a,e) accepts a trade if and only if his income (the value of his holdings evaluated at the prices p) does not decrease. If he already achieved the bundle h(a,e), then he accepts if and only if his income increases.
- Every agent leaves the market as soon as he achieves the bundle h(a,e).

On the equilibrium path and as in Gale (1986b), we shall distinguish between the behavior of one of the divisible commodities (say, commodity 1) and that of the other |D|+|I|-1 goods. While according to the trading rule g an agent's excess demand in commodities other than 1 is non-increasing, commodity 1 serves to balance the budget whenever there is no pure coincidence of wants.

In order to define the trading rule g, we shall denote the
proposer by a_{0} and the responder by
a_{1}. Subscripts denote agents and superscripts
denote commodities. Let z_{0} be
h(a_{0},e(a_{0},t))-e(a_{0},t),
where e(a_{0},t) are the holdings of a_{0}
in round t, and similarly let z_{1}
be h(a_{1},e(a_{1},t))-e(a_{1},t).
Define the set B(z_{0},z_{1}) as the set of
vectors x∈ℜ^{D}×**Z**^{I}
satisfying the following conditions (where **Z** denotes
the set of integers):

- |x
^{k}|≤|z^{k}_{i}|, i=0,1, k≥2 - 0≤(-1)
^{i}x^{k}z^{k}_{i}, i=0,1, k≥2 - e(a
_{i})+(-1)^{i}x∈ℜ_{+}^{D}×**N**^{I}and px=0 i=0,1.

If the net trade x is proposed and accepted, the
proposer's new endowment e(a_{0},t+1)=e(a_{0},t)+x
and the responder's e(a_{1},t+1)=e(a_{1},t)-x.

Now we are ready to present the trading rule proposed in all bilateral meetings. The trade proposed is denoted by:

g(z_{0},z_{1})=argmax{-∑_{k≥2}exp(-|x^{k}|):
x∈B(z_{0},z_{1})}.

Denote by u^{*}(c,p) the maximum utility
that an agent with characteristics c can achieve over
the budget set determined by his endowment and the prices p.
We will show that if every agent behaves according to the specified
strategies, almost every agent of characteristic c achieves
a bundle (corresponding to the allocation f) that yields
u^{*}(c,p) in finite time.

Notice first, as in Gale (1986b), that if agents follow the specified strategies, it is not possible for an agent to increase his income as evaluated by the prices p. Next we will show that an agent a ends up at the bundle f(a) in finite time with probability 1. This will show that the proposed strategies are a strategic equilibrium. That is, given that there is no way to increase one's income, the proposed strategies induce a random path that takes each agent to his chosen maximizer f(a) over the budget set determined by e and p.

For each agent a and for prices p define the excess demand as follows (for convenience and given that the Walrasian prices p are fixed throughout, we shall drop p from the expressions below): φ(a,t)=f(a)-e(a,t), where e(a,t) are the holdings of agent a at round t. Notice that, given the strategies specified above, every agent travels along the frontier of his budget set which means that f(a) continues to be a utility maximizer for agent a.

As we said above, we shall distinguish between the behavior of
commodity 1 and that of the others. The trading rule
g is constructed so that the absolute value of
the excess demand of every agent in all goods but 1
does not increase. On the other hand, good 1 serves to
balance the budget whenever there is no pure coincidence of wants.
Thus we define for each agent a the following number:
β(a,t)=∑_{k≥2}|φ^{k}(a,t)|.
That is, for each agent a the statistic β(a,t)
indicates the sum of absolute values of excess demands in all goods but
1 in round t. We will next show that over time
the distribution of β(a,t) converges weakly in measure
to a degenerate distribution concentrated on 0.

Recall that μ denotes the measure of characteristics in the economy, i.e., the measure of characteristics of all agents who are active in the market plus that of the agents who already left the market. The random matching process and the specified strategies lead to new distributions of characteristics at every round, and hence to new distributions of the statistic β(a,t). We shall concentrate on an arbitrary path determined by a particular realization of the different random variables at play (the coalitions that meet and the roles of each agent in each meeting). We show then that, along this path, the distribution of β(a,t) converges in measure to the degenerate distribution concentrated on 0.

The space of characteristics C at each round t
is the Cartesian product of a fixed space of utility functions
with ℜ_{+}^{D}×**N**^{I}.
The evolution of the economy is thus described by a sequence
of measurable maps from the set of agents A to the set
C. Given that the set of utility functions is fixed
throughout the model, any Cauchy sequence of such measurable maps must
converge to a measurable map from A to C.
To see this, notice that, after having fixed the utility functions,
the marginal of characteristics c on agents' endowments e
allows us to consider a Cauchy sequence of integrable maps from A
to ℜ_{+}^{D}×**N**^{I}.
Endowing this space of integrable maps with the topology induced by the
supremum norm, it is easy to see that such a sequence converges to an
integrable map into ℜ_{+}^{D}×**N**^{I}
as this is a complete space. Notice that the marginal of the measure
μ_{t} on utility functions is constant. We can
then abuse notation slightly and denote by μ_{t}
round t's measure on the agents' endowments and not on
characteristics. Thus, using Hildenbrand (1974,
p.50) the sequence of
measures {μ_{t}} is tight and has a convergent
subsequence to μ^{*}. Without loss of generality,
suppose the sequence itself converges to μ^{*}.

By the properties of the trading rule g, we must have that
for a given constant τ>0, ∫_{β(a,t)≥τ}μ(a,t)
converges to 0 as time goes to infinity. To see this, we
argue by contradiction. Suppose that the limiting measure μ^{*}
is not the one concentrated at 0. Since μ^{*}
is the limiting measure, it must be the case that the measure of agents trading positive amounts
of goods when the distribution of β(a,t) is approximately
mu;^{*} must be arbitrarily close to 0:
For all ε>0 there exists a T such
that for all t>T we have that
∫[μ_{t}-μ^{*}]<ε.
If μ^{*} is not the one concentrated at
0, there must exist a positive measure of agents whose
characteristics satisfy that φ^{k}(c)>0
for some good k. By Walras' law which holds at each step
of this time path, there must also exist a positive fraction
of agents whose characteristics satisfy that φ^{k}(c)<0.
Since the matching process is random, there exists a positive probability
that agents in these two situations will meet. Finally, given the trading
rule g, these agents will trade at least in good k,
which is a contradiction, i.e., there exists ε>0
such that for all T there exists t>T with
∫[μ_{t}-μ^{*}]>ε
as g(z_{0},z_{1}) stays bounded away from
0 for a positive fraction of meetings.

As for convergence in finite time, the arguments are identical to those in Gale (1986b, section 7).

By using the insights of the theory of the core, this paper presents a model of decentralized trade through bargaining in coalitions of finite size. This allows us to obtain equivalence results among core, Walrasian and strategic equilibrium allocations for a wide class of large exchange economies, including non-differentiable non-convex preferences and indivisible goods.

Our results are robust to several extensions of the model. First, we could allow for any entry process (not necessarily one time entry) as long as the measure of the total entering population is finite. Second, different bargaining procedures in the coalitional meetings could be adopted: for example, veto power can be given only to those responders who are offered a non-zero trade vector. Third, we only need to assume that an agent's probability of meeting a coalition of size n be positive. In particular, we could assume that the probability of meeting a coalition of more than two agents be arbitrarily small and all our results would go through. Thus, Gale's model can be viewed as the "limit" of ours as the probabilities of multilateral meetings vanish. This poses the important open question of lower hemicontinuity of the equilibrium payoff correspondence, i.e., which of the extra assumptions made by Gale are really needed to obtain the result using only trade in pairs.

A separate dimension along which our results are more robust than the previous ones found in the literature is the class of economies to which they apply. We discuss this in length in the following paragraphs, especially comparing our results to the important works of Gale (1986a, b, c) and McLS (1991).

**Existence of equilibrium** Thanks to considering finite
coalitions in the procedure, our paper yields the equivalence
between strategic and Walrasian equilibria under essentially
the same assumptions as those needed for the core equivalence theorem.
Moreover, our assumptions also guarantee the existence of a Walrasian
equilibrium, as opposed to Gale (1986a,
b, c)
and McLS (1991), which
deal with open consumption sets.

**Limited applicability of the previous models** As
discussed in the introduction, we regard the relaxation of
differentiability as a crucial conceptual departure from Gale's
and McLS's work. From an applied view-point, the differentiability
assumption by itself that Gale (1986a,
c) and McLS (1991) make
is not very restrictive: many models in economics incorporate it
in order to allow for a closed solution and for the performance of
comparative statics exercises. However, the proofs of the above
mentioned authors rely on additional strong assumptions, that exclude
most applied models. Gale (1986a) assumes that
for each utility
function the support of the endowments compatible with it
is the entire consumption set. This assumption excludes the
possibility of a finite type economy. Gale
(1986c), who assumes a finite
number of types, uses a bounded curvature assumption,
thereby excluding, for example, Cobb-Douglas utility functions on the
non-negative orthant or its interior. McLS (1991)
make either a bounded curvature assumption similar to
Gale's (1986a) or a restriction on the
equilibrium which seems to require an assumption similar to
Gale's (1986a) on the primitives of the economy.
In contrast, our model, which applies to very general economies, also
applies to these standard cases.

**Feasibility in and out of equilibrium** We assume, like
Gale (1986a, b,
c), that the flow of agents entering the market
constitutes an economy, i.e., they sum up to a finite measure.
In addition, we also assume that short sales are not allowed. These two
assumptions together ensure that the flow of agents out of the
market is consistent with the feasibility constraint of the economy.
Suppose, like McLS (1991), that the total measure of
agents is finite,
but short sales are allowed. In this case, nothing assures that feasibility is
met. Consider an arbitrary assignment of bundles to agents, and
the following strategies (that do not constitute an equilibrium
in McLS's game with short sales). Each proposer asks for the bundle assigned
arbitrarily to him and each responder accepts any proposal; agents leave
the market as soon as they reach their assigned bundle. Clearly, these
strategies guarantee that each agent will get with probability 1
the assigned bundle. The problem stays even if we restrict attention
to the equilibria of their game. Indeed, the strategic equilibrium
that McLS propose (pp. 1395-1396) to support a Walrasian equilibrium is
a strategic equilibrium for any prices. That is,
for an arbitrary price vector, their strategic equilibrium
gives the outcome that every agent maximizes over the corresponding
budget set, but the market clearing conditions may be violated. This casts
doubt on the validity of such a model as a foundation of Walrasian
outcomes, since the IOUs are eventually consumed by the
agents. As we perceive market clearing conditions as an essential part of
the Walrasian concept, we do not allow for short sales and adopt Gale's approach,
which ensures feasibility in and out of equilibrium. See
Dagan, Serrano and Volij (1998) for other related
criticisms of the McLS model.

In the case where the sum of the measures does not constitute an economy, it is not clear what are the feasibility constraints. McLS's Theorem 3 deals with the case when the inflow and the outflow of agents have long run averages. Following this approach, we can define the flow of agents out of the market to be 'feasible' if it is consistent with the long run average of the inflow of agents. However, it is not clear to us how one can construct a model in which this kind of constraint is met in and out of equilibrium. Therefore, interpreting the result of RW (1985) as consistent with Walrasian allocations (as done by McLS's Theorem 3) is not sound: the outcome of their strategic equilibrium is consistent with a notion of feasibility, but behavior different from this equilibrium may violate the same feasibility notion.

**Strict concavity and indivisible goods**
McLS (1991) note the restrictiveness of
Gale's (1986a, b,
c) assumption
of strictly concave utility functions in a continuum setting. One should
expect that the convexifying effects of large numbers could be helpful
to relax this assumption. However, we believe that McLS's solution to the
problem is inadequate and provide an alternative treatment of the issue.
One difference between the underlying economy and the strategic model is
that in the latter the outcome (at least for an individual agent) may be
random and thus preferences on random outcomes must be specified. McLS
do not make any assumption regarding the concavity of utility functions;
instead, they allow for short sales, which enables them to prove
that outcomes of the strategic equilibria are not random. As explained
above, assuming short sales in this framework may be problematic.
In addition, a separate shortcoming of their treatment of
non-convexities is that they maintain the differentiability
assumption, which precludes non-convexities arising
from the existence of indivisible commodities.

Gale (1986a) uses the strict concavity assumption only to ensure that the introduction of lotteries does not enlarge the set of possible utilities of the agents. Thus, what is needed is a property of risk aversion in the aggregate. We impose a condition on the quasiconcave covers of the utility functions that ensures the sufficient degree of aggregate risk aversion. This assumption is compatible with having indivisible commodities as well as other kinds of non-convexities. Thus, our assumptions allow for a unified treatment of assignment markets à la RW (1985) and classical exchange economies à la Gale (1986a, b, c). We should stress that our assumption of aggregate risk aversion is sufficient to obtain Gale's results as well (of course, within his restrictive subdomain).

This appendix contains the proofs of Propositions 1 and 2.

**Proof of
Proposition 1** Let L
be a lottery on allocations. We define
EU_{a}(L)=∫_{F}u_{a}(f(a))dL.
We also define E_{a}(L)=&int_{F}f(a)dL and
h:A→ℜ_{+}^{D∪I} as the
function that assigns to each agent a the bundle
E_{a}(L). It is easy to see that h
is integrable. Now let
$φ(a)={x∈ℜ_{+}^{D∪I}:__u___{a}(x)≥__u___{a}(h(a))}.
Clearly, ∫h(a)∈∫φ(a)
since it is true for every a. Let
ψ(a)={x∈X:u_{a}(x)≥__u___{a}(h(a))}.
It follows from the definition of __u__ that __ψ__(a)=φ(a),
where __ψ__(a) is the convex hull of ψ(a).
Now consider ∫ψ(a). It follows from Liapunov's
theorem that this integral is convex, and thus
∫ψ(a)=∫φ(a). Therefore, ∫h(a)∈∫ψ(a).
Thus, there exists an allocation g such that
u_{a}(g(a))≥__u___{a}(h(a)) for
almost all a∈A. It follows from the fact that all the
covers __u__ are concave that almost all agents weakly
prefer the allocation g to the lottery L.

**Proof of Proposition 2**
Let L be a non-degenerate lottery over allocations.
First it follows from Proposition 1
that there exists an allocation g that
satisfies that u_{a}(g(a))≥__u___{a}(h(a))
for almost all a∈A.
Since the lottery is non-degenerate, there is a positive measure of
agents such that each of them is not indifferent among almost all
bundles in the support of L. Now for each agent
a in this set, we can have two cases:

1. There does not exist k such that for
L-almost all f∈F,
__u___{a}(f(a))=k. In this case,
it follows from almost strict concavity of __u__ that
__u___{a}(h(a))>EU_{a}(L).

2. There existsk such that for
L-almost all f∈F,
__u___{a}(f(a))=k. Since the agent is not
indifferent among almost all bundles, there exists a set of allocations
that are assigned positive probability by L such that
__u___{a}(f(a))>u_{a}(f(a)).
Thus, __u___{a}(h(a))=k>EU_{a}(L).

- Aumann, R. J. (1964), Markets with a Continuum of Traders, Econometrica 32, 39-50.
- Cournot, A. (1838), Researches into the Mathematical Principles of the Theory of Wealth. English edition (ed. N. Bacon) Macmillan, 1897.
- Dagan, N., R. Serrano and O. Volij (1998), Comment on McLennan and Sonnenschein "Sequential Bargaining as a Non-Cooperative Foundation for Walrasian Equilibrium," Econometrica, forthcoming.
- Edgeworth, F. Y. (1881), Mathematical Psychics, Kegan Paul Publishers, London.
- Einy, E. and B. Shitovitz (1997), Von Neumann-Morgenstern Stability of the Symmetric Pareto Optimal Allocations in Pure Exchange Economies, mimeo, Ben-Gurion University of the Negev.
- Fudenberg, D. and J. Tirole (1991), Game Theory, M.I.T. Press, Cambridge, MA.
- Gale, D., (1986a), Bargaining and Competition Part I: Characterization, Econometrica 54, 785-806.
- Gale, D., (1986b), Bargaining and Competition Part II: Existence, Econometrica 54, 807-818.
- Gale, D., (1986c), A Simple Characterization of Bargaining Equilibrium in a Large Market without the Assumption of Dispersed Characteristics, working paper 86-05, CARESS, University of Pennsylvania, Philladelphia, PA.
- Hammond, P., M. Kaneko and M. H. Wooders (1989), Continuum Economies with Finite Coalitions: Core, Equilibria and Widespread Externalities, Journal of Economic Theory 49, 113-134.
- Hart, S., W. Hildenbrand and E. Kohlberg (1974), On Equilibrium Allocations as Distributions on the Commodity Space, Journal of Mathematical Economics 1, 159-166.
- Hildenbrand, W. (1974), Core and Equilibria of a Large Economy, Princeton University Press, Princeton, NJ.
- Jevons, W. S. (1879), The Theory of Political Economy, Mcmillan, London.
- Kannai, Y. (1970), Continuity Properties of the Core of a Market, Econometrica 38, 791-815.
- McLennan, A. and H. Sonnenschein (1991), Sequential Bargaining as a Foundation of Competitive Equilibrium, Econometrica 59, 1395-1424.
- Negishi, T. (1989), History of Economic Theory, North Holland, Amsterdam.
- Osborne, M. J. and A. Rubinstein, (1990), Bargaining and Markets, Academic Press, San Diego, CA.
- Perry, M. and P. J. Reny (1994), A Non-Cooperative View of Coalition Formation and the Core, Econometrica 62, 795-817.
- Rubinstein, A., (1982), Perfect Equilibrium in a Bargaining Model, Econometrica 50, 97-109.
- Rubinstein, A. and A. Wolinsky, (1985), Equilibrium in a Market with Sequential Bargaining, Econometrica 53, 1133-1150.
- Serrano, R. (1995), A Market to Implement the Core, Journal of Economic Theory 67, 285-294.
- Serrano, R. and R. Vohra (1997), Non-Cooperative Implementation of the Core, Social Choice and Welfare 14, 513-525.
- Starr, R. (1969), Quasi-Equilibria in Markets with Non-Convex Preferences, Econometrica 37, 25-38.
- Walras, L. (1874), Elements of Pure Economics, or the Theory of Social Wealth, English edition (ed. W. Jaffé), Reprinted by Orion Editions, Philadelphia, PA, 1984.

- 1

(Walras, 1874, pp.83-84)."The markets which are best organized from a competitive standpoint are those in which purchases and sales are made by auctions ... Besides these markets, there are others, such as the fruit, vegetables and poultry markets, where competition, though not so well organized, functions fairly effectively and satisfactorily. City streets with their stores and shops of all kinds ---baker's, butcher's, grocer's, taylor's, shoemaker's, etc.--- are markets where competition, though poorly organized, nevertheless operates quite adequately... ".

- 2
- Osborne and Rubinstein (1990, chapter 6) and the references therein for earlier models of decentralized trade in pairwise meetings where each pair uses the Nash bargaining solution to split the gains from trade.
- 3
- We will need an additional assumption beyond those needed for core equivalence to guarantee that the outcome of a strategic equilibrium is not a lottery over allocations. This extra assumption is weaker than Gale's strict concavity and is compatible with the presence of indivisible goods.
- 4
- These assumptions are either uniformly bounded curvature or dispersed endowments. See our final section for details.
- 5
- The problem of existence of such random matching processes is already present in models with only pairwise meetings. For possible treatments of this problem, see McLS (1991, footnote 4) and the references therein.
- 6
- The arguments in the characterization theorem (Theorem 1) are entirely independent from this phase. However, its introduction simplifies greatly the proof of Theorem 2, which could be complicated due to the fact that our model allows for demand correspondences.
- 7
- Alternatively, we could take the approach based on distributions like in Hart, Hildenbrand and Kohlberg (1974) instead of the name-based approach. We should then assume (like Gale (1986a, b, c), Osborne and Rubinstein (1990) and McLS (1991) that any two agents with identical characteristics and histories play the same strategy. This would enable us to employ the machinery developed by McLS (1991, section 3.3) in order to establish that for any given strategies the "state of the market" in round t in the sense of distribution of agents' characteristics is deterministic. See also Osborne and Rubinstein (1990, pp. 160-161), who show this for the finite type pure strategy case.

Nir Dagan / Contact information / Last modified: July 28, 1998.