Working Paper No.

61

Hold-Up and the Evolution of Investment and

Bargaining Norms

by

Herbert Dawid and W. Bentley McLeod

University of Bielefeld
Department of Economics
Center for Empirical Macroeconomics
P.O. Box 100 131
33501 Bielefeld, Germany
 Hold-up and the Evolution of Investment and Bargaining Norms∗
Herbert Dawid W. Bentley MacLeod†
Department of Economics Department of Economics and the Law School
University of Bielefeld University of Southern California
P.O. Box 100131 Los Angeles, CA 90089-0253
Bielefeld 33501, Germany wmacleod@usc.edu
hdawid@wiwi.uni-bielefeld.de

Abstract

The purpose of this paper is to explore the evolution of bargaining norms in a simple team production
problem with two sided relationship specific investments and competition. The puzzle we wish to address
is why efficient bargaining norms do not evolve even though there exist efficient sequential equilibria?
Conditions under which stochastically stable bargaining conventions exist are characterized, and it is
shown that the stochastically stable division rule is independent of the long run investment strategy.
Hence, efficient sequential equilibria are not in general stochastically stable, a result that may help us
understand why institutions, such as firms, may be needed to ensure efficient exchange in the context of
relationship specific investments. We also find that increasing competition, while enhancing incentives,
may also destabilize existing bargaining norms, and may explain why the introduction of market in
transition economics may initially result in reduced output.

JEL Classification: C78, D83, Z13

Keywords: hold-up, evolution of norms, fair division, stochastic stability

∗ We would like to thank Mark Armstrong, Jack Robles, Debraj Ray, Luca Anderlini, Leonardo Felli and the anonymous
referees for helpful comments, and Alexey Ravichev for research assistance. We also thank the Industrial Relations Section,
Princeton University and the National Science Foundation, SES 0095606, for research support.
† Current address: Visiting Professor, Industrial Relations Section and Department of Economics, Princeton University,

Princeton, NJ 08540.

1
1 Introduction
A starting point for modern contract theory is the Coase (1960) conjecture stating that in the absence of
transactions costs, individuals should be able to bargain to efficient allocations, regardless of the original
allocation of property rights. However, when individuals make investments that are both relationship specific
and non-contractible before trade, then, as Grossman and Hart (1986) show, the allocation of property rights
can affect the returns from these investments, and hence the efficiency of the relationship. This observation
has led to the “property rights” view of the firm, in which ownership allocation and firm boundaries are
viewed as a mechanisms that enhance productive efficiency (see Hart (1995)).
This view is not uncontroversial. Maskin and Tirole (1999) observe that incomplete contracts and hold-
up do not preclude the existence of efficient contracts. They show that regardless of the property rights
allocation, if agents have sufficient foresight, then there exist contracts that provide efficient investment
incentives. By itself, this does not imply that agents necessarily choose efficient contracts from the set of
incentive compatible contracts. Such a step typically relies upon some model of negotiation or equilibrium
selection. Both Moore (1992) and Tirole (1999) worry that efficient contracts may be too complex to use
in practice, and hence to explain observed contracting arrangements one should rely upon a more realistic
model of behavior.1
One approach to this problem that has been particularly influential in the legal academy are evolutionary
models of selection.2 For example Ellickson (1991) has studied the system of property rights allocation in
Shasta County, and concludes that over time individuals are likely to evolve efficient norms of behavior.
Moreover, even when there is an existing inefficient legal rule, individuals are able to evolve rules that are
more efficient and supersede the legally enforceable rule.
In this paper we use the evolutionary learning model of Young (1993a) and Kandori, Maillath, and Rob
(1993) to evaluate the claim that through a process of experimentation and learning individuals select an
efficient equilibrium in a simple bilateral trade model, in which both parties have an opportunity to make
non-contractible, relationship specific investments. Specifically, we consider a situation in which each agent
makes a non-contractible investment of cost c before entering a market. After entering the market, they are
randomly matched, at which point they observe each other’s productivity, and then play a Nash demand
game to divide the gains from trade.3 An important feature of the Nash demand game is that any division
of the gains from trade is a potential Nash equilibrium, and hence the division rule selected can depend
upon each party’s contribution to the productivity of the match. This ensures that whenever it is efficient
1 See discussion on page 773 of Tirole (1999).
2 See Alchian (1950) for an account that is still well worth reading. Kim and Sobel (1995) make the point explicitly that even
if one allows for communication one cannot be assured that an efficient allocation will be selected. They show that efficient
allocations are selected in pure coordination games. When the common interest assumption fails (as in this paper), evolution
with communication has no unique equilibrium.
3 The rules of the Nash demand game are as follows. Given the gains from trade S, each person makes a demand d , and if
i
d1 + d2 ≤ S, then they recieve their demand. Otherwise they receive zero.

2
for both parties to invest, there exists a sequential equilibrium implementing the efficient allocation.4
Our model captures a phenomenon that is observed in many contractual settings. For example Macauley
(1963) documents in the case of form contracts for the sale of goods that trading partners do not explicitly
agree upon all the terms and conditions for trade. This is evidenced by that fact that the terms and conditions
of the standard purchase and acceptance orders are often inconsistent with each other. In the event of a
dispute contracting parties typically depend upon past practice. For example, if a supplier provides a good
of substandard quality, then in the absence of specific contract term, the Uniform Commercial Code of the
United States simply requires the two parties to reach an amicable adjustment to the price.
The problem is that if the resulting decrease in price is small, then the seller has reduced incentives
to take precautions. Similarly, the buyer may not be able to accept the goods, in which case buyer and
seller need to renegotiate the terms of the agreement to either delay shipment or terminate all together the
agreement. In either case, there is a mutually observed event that results in a rent (or loss) that must be
allocated between the two parties. In some cases the buyer and seller may have contract terms in place
that explicitly allocate liability. Given that it is very common for there to be no contract term in place,
the question then is whether or not over time an there will evolve an efficient bargaining norm governing
the ex post division of the gains from trade. By bargaining norm, we mean an implicit agreement on how
to divide the ex post gains from trade or renegotiation as a function of information available at the time of
negotiation.
The first issue we address is the existence of a stable bargaining norm in a two sided version of models
by Ellingsen and Robles (2002) and Troger (2002). They consider a model in which only one party makes
an investment, followed by play of the Nash demand game. They show that there is an efficient stable
equilibrium, with the feature that the investing party appropriates most of the gains from trade. This
is a very interesting result, however we feel that the interpretation provided by the authors is somewhat
misleading. These authors claim that their results demonstrate that holdup cannot be considered a stable
feature of an environment with boundedly rational individuals. In particular, they claim that their results
imply that stable division rules have the feature that individuals are rewarded according to their contribution.
Troger (2002) states explicitly that their results support equity theory, described by Rabin (1998) as the
theory which predicts “that people feel that those who have put more effort into creating resources have
more claim on those resources.”5
While equity might be correct, it does not follow from the results of Ellingsen and Robles (2002) and
Troger (2002). In their model, like all evolutionary models based upon Young (1993a), experimentation
ensures that all possible histories and strategies occur with positive probability. In particular, one eventually
reaches an outcome corresponding to the investing player getting all the rents and making an efficient
investment. To destabilize this outcome, it must be the case that the investing player can profitably deviate,
6
but since the outcome is efficient, and she appropriates all the rents, this is not possible.
4 In a related model, Carmichael and MacLeod (2003) show that the efficient allocation rule is unique when there is sufficient
diversity in preferences. In the subsequent discussion, when we use the term “equilibrium” by itself, we mean the sequential
equilibrium of the game. Similarly, the term ”stable equilibrium” refers to Peyton Young’s notion of a stochastically stable
equilibrium, which we define formally in section 4.
5 As cited in Troger (2002), page 376. The orginal source is Rabin (1998), page 18.
6 The fact that strategies are discrete implies that the actual results is a bit more complex. However, this arguement captures

3
 Thus, what they have shown is not that each player is rewarded according to their contribution, but
that the allocation of all the rents to one party is stable when that party is making all the investments. In
particular, the additional value added by the investment may be relatively small compared to the gains from
trade, but never the less, the stable outcome entails giving the investing party all the rents. Our preferred
interpretation of their results is that they demonstrate that one does not need new organizational forms,
such as firms, to enhance efficiency when relationship specific investments are one-sided.
This is a very interesting result, and can help us understand the conditions under which efficient property
rights can evolve. It also provides a new way to interpret the evidence in Demsetz (1967) and Ellickson (1991)
that documents the evolution of efficient property right norms. In both cases (fur trapping and the fencing
of land), efficient allocations could be achieved when only one party makes an investment, and hence one
observes the efficient evolution of property rules in this particular case.
The holdup model of Grossman and Hart (1986) is quite different. They are concerned with the case
in which the allocation of bargaining power to one party creates a trade-off, namely it enhances incentives
for one party at the expense of the other.7 This issue goes back to the Alchian and Demsetz (1972) view
of the firm as a problem of team production, namely one faces the problem of coordinating the efforts of
several individuals, not just one person. We pay particular attention to complementary inputs, namely
inputs for which the marginal product of investment by one person is much less than investment by two
individuals. Our first result in the case of two sided relationship specific investments is that there exists no
stable bargaining norm when investments are observable, but not contractible.8
The reason such a norm does not exist is that at an efficient equilibrium both parties make high in-
vestments, and hence there are no high-low matches in the long run equilibrium. This implies that beliefs
regarding the outcome of high-low matches can drift due to the noise inherent in the evolutionary learning
model, until eventually it is in one party’s interest to enter the market with a low investment. A similar
argument applies to the low-low equilibrium.
We solve this problem by observing that it is not reasonable to suppose that there is a deterministic
relationship between investment and performance. For example, if a firm invests in worker training, there
is always a chance that some of the workers do not acquire the skill. We formally model this effect by
supposing there is a small probability that high investment results in low productivity, and conversely that
a low investment may (with low probability) result in high output. This, seemingly small, modification of
the investment game now ensures the existence of a stable bargaining norm, regardless of the investment
strategies. This bargaining norm corresponds to the equal division rule, and hence we can conclude that
equity theory is not consistent with a stable equilibrium when both parties contribute to the gains from
trade, and there is some uncertainty regarding the link between investment and productivity. Moreover,
the model predicts that individuals ignore sunk costs at the bargaining stage, as assumed in the standard
the essential intuition when the set of investment strategies is close to a continuous choice.
7 Grossman and Hart (1986) state in their abstract that “When residual rights are purchased by one party, they are lost by a

second party, and this inevitably creates distortions.” Note that such distortions cannot arise in the case of one-sided investment
when the investing party buys the asset. As they note, this observation goes back to the work of Alchian and Demsetz (1972),
who emphasize the team nature of firm production.
8 This result is suggested in Troger (2002) discussion his model, and we first proved it in Dawid and MacLeod (2001), a

journal this is now unfortunately out of print.

4
hold-up model.
In our model investment is discrete, thus even if the bargaining norm is unique and given by the equal
split rule, multiple investment equilibria are possible. In particular, when investments are complementary,
there is a large set of parameter values for which it is an equilibrium for either both parties to make high
investments or both to make low investments. The learning model can select between these equilibria, and
we find that when the cost of investment is greater than half of the potential gains from going to the high
investment equilibrium, then the low investment equilibrium is stable.
Therefore, we can conclude that in a simple holdup problem with two sided investment, when investments
are complements and costs are sufficiently high, then the efficient equilibrium is not stable. Hence learning
and experimentation over time does not necessarily lead to efficient bargaining norms. This result is consistent
with the view articulated by Jean-Philippe Platteau (2000), who explicitly argues that the egalitarian norm,
very common in Africa, can impede development. He cites the example of a fishery project at lake Kivu in
southern Congo, where a poor village was given access to a new gill net technology.9 Within the village,
there was pressure to allow many individuals to use the new nets, resulting in inefficient maintenance and
care of the equipment. In this case, even though the technology clearly improved total output, the existing
norms made it impossible for the villagers to implement an efficient system to utilize this new technology.
It appears that the egalitarian norm is so stable, that it is very difficult to introduce organizational forms
for which there is a strong respect for property (or “equity theory”). A standard economic prescription
in these cases is the introduction of more competition. We consider the implications of this possibility by
supposing that if trade does not occur, the individual may re-enter the market with their investments the
next period, with a discount factor of δ. Varying δ from 0 to 1 parameterizes the model between the case of
pure holdup and perfect competition.
We find that the introduction of an outside option does indeed enhance efficiency. However, there is a
cost. When δ is close to 1, then individuals with high productivities currently in high-low matches, may
prefer not to trade and re-enter the market next period in the hopes of meeting another high productivity
individual. We find that merely the possibility of being better off is sufficient to destabilize the evolution of
a stable bargaining norm.
Suppose it is an equilibrium for all individuals to make a low investment, and hence an individual with
high productivity has a low probability of meeting a high productivity individuals the next period and hence
should trade as soon as she enters the market, regardless of the productivity of her partner. Now, suppose
she is better off not trading if she were sure to meet a high productivity person the next period, then we can
show that no stable bargaining norm can evolve. The reason is similar to our previous non-existence results.
When high-low matches are rare, then it is possible for beliefs to drift, with the consequence that there is
always a chance that the high type believes she is better off waiting to trade with another high type.
This result provides an illustration of how the introduction of markets can have a potentially destabilizing
effect upon bargaining norms. This observation appears to be consistent with the recent transition experience
in Eastern Europe, as documented by Roland (2000). He notes that the period of transition from a planned to
market economy in Eastern Europe entailed a great deal of disorganization and a decline in output as people
9 See Mellard, Platteau, and Watongoka (1998). The case is discussed in detail in Platteau (2000), pages 200-201.

5
learn how to engage in trade in a new market environment. Our model provides a potential explanation for
the decline in output. Namely, the mere possibility of better trades in the future, can lead to high types
delaying trade, particularly if low types attempt to expropriate too much of the gains from trade created by
the high types.
One of the messages of this paper is that a necessary condition for the evolution of norms is experience
with matches at which these norms are exercised. Hence, our results suggest that even though there may
exist efficient incentive contracts, if they are sensitive to the way payments are structured in events that are
very infrequent, then they may not be stable, and may not work as expected in practice.
The agenda of the paper is as follows. The next section introduces the basic model, and it is shown that
whenever high investment is efficient, there is a sequential equilibrium implementing the efficient allocation
(which we simply call an equilibrium). Sections 4 introduces the formal stochastic learning model that is
used to define the notion of a stochastically stable equilibrium, which simple call a stable equilibrium. This
is followed by a discussion of how adding two sided investment to the model results in the non-existence of
a stable equilibrium. A preliminary analysis of the stable equilibrium for our model is carried out in section
6. Section 7 considers the case of substitutes, where the marginal return from the first investment is greater
than the second investment, while section 8 presents our results for complementary investments. The paper
concludes with a discussion of the results.

2 The Model
We are interested in the kind of bargaining and investment norms which are developed endogenously in a
population of adaptive agents. To examine this, we use an evolutionary bargaining model similar to Young
(1993b) and Kandori, Maillath, and Rob (1993) as extended to incorporate investment by Troger (2002)
and Ellingsen and Robles (2002). The basic idea underlying this approach is that individuals anonymously
interacting in a population use a random sample of observed past behavior to build beliefs about current
actions of their opponent. With a large probability they then choose the optimal strategy given their beliefs.
Consider a single population of identical agents who are repeatedly matched randomly in pairs to engage
in joint production (or in a joint project). Every agent can make an investment, either high (H) or low (L),
before entering the population that influences his type, and accordingly the joint surplus of the project.
This investment can be thought of as human capital, such as the acquisition of special skills needed for
a project, though the framework is sufficiently general that any type of project specific investment might
be considered. It can also be applied to situations where there are explicit contracts in place, but they are
incomplete. In that interpretation an event occurs that is not covered by the contract, for example there is a
defect in the product. Given that the product is defective, then the buyer is not longer under an obligation
to accept delivery, but it may never the less be efficient to trade with the defective product, rather than
order a new product. The question then is how is the new price determined. Notice, that this is a two sided
investment problem because the sellers efforts affect the probability of a defect, while specific investments
by the buyer, create a need to trade now, rather than wait for a replacement product.
Before partners start joint production or trade they bargain over the allocation of the joint surplus. If

6
the bargaining does not lead to an agreement they split without carrying out the project and look for new
partners. The effect of an investment stays intact as long as the agent has not carried out the project, it is
however assumed that the investment is project specific and creates no additional revenue after the project
has been carried out. Looking for a new partner for the project needs time and therefore payoffs from the
next matching are discounted by a factor δ ∈ [0, δ̄]. The more specific the project at hand is the longer is
the search time and the smaller is δ. Hence, we interpret δ as a parameter measuring the project specificity,
however it could be induced by any type of market frictions leading to search times. The value of trade t
periods after the initial investment is δ t U, where U is the agents share of the gains from trade. When δ = 0
the investment can only generate revenues in the current period and the model corresponds to one with
purely relationship specific investment.
The sequence of decisions facing an individual are:

1. The agent, i, decides about her investment level Ii ∈ {h, l} , where the cost of investment is
(
c, if I = h,
c (I) =
0, if I = l.

After the investment has been made the type Ti ∈ {H, L} of the agent is determined. It is assumed
that the probability of being a high type after having invested I is pI , where ph > pl .

2. The agent is randomly matched with some partner and both observe each other’s type. The types
determine the size of the surplus, STi Tj , where when convenient SH ≡ SHH , SA ≡ SHL = SLH and
SL ≡ SLL , and satisfies SH ≥ SA ≥ SL > 0.

3. Individual i makes a demand conditional upon her type and that of her partner j, denoted by xTi Tj ∈
© ª
XTi Tj (k) = 0, αTi Tj , 2αTi Tj , ..., kαTi Tj , αTi Tj = STi Tj /k, k is some large even number.

4. The payoff to individual i in this period is given by the rules of the Nash demand game:
( )
i xiTi Tj , if xiTi Tj + xjTj Ti ≤ SIi Ij ¡ ¢
U = i j
− c Ii
0, if xTi Tj + xTj Ti > SIi Ij

and similarly for player j. Agents are assumed to be risk neutral.

5. If agent i has traded in this period she leaves the population and is replaced by another individual.
If there was no trade the individual stays in the population and goes again through steps 2 - 5 in the
following period where future payoffs are discounted by a factor δ per period.

Throughout the analysis SH and SL are assumed fixed, while the degree of complementarity in investment,
SA , the cost of investment, c, and the discount rate δ are parameters that determine the nature of the
investment problem.
Furthermore, we assume that the probability that the type differs from the investment level is symmetric
and small, namely: 1 − ph = pl = λ for some small positive λ. This latter assumption plays an important role
in the analysis because it ensures that even if all individuals carry out high investment, there is a strictly

7
positive probability of having low types in the population. Hence each period there is the potential for
trade between H and L types. As we shall see, the existence of such trades is a necessary condition for the
evolution of a bargaining norm.
This is a one-population model where the only difference between individuals stems from their investment.
Accordingly, in any uniform equilibrium where all individuals use identical strategies, the surplus has to be
split equally in matches of partners with identical investments. We are concerned with the evolution of norms
which are uniform equilibria, and hence in any norm the surplus has to be split equally between partners
with identical investment10 . Therefore, to simplify the analysis it is assumed here that when two high types
SH SL
meet or two low types meet they split the gains from trade equally if they trade, i.e. xiHH = i
2 , xLL = 2 ∀i.
Although this has to hold true in any norm, our assumption is not completely innocent. In the absence of
such an assumption we may also have a cyclical long-run phenomena where all individuals keep switching in a
coordinated fashion between demanding more or less than half of the surplus in equal investment matchings.
This would result in disagreement for half of the periods and a waste of parts of the surplus for the other half.
Ruling out such phenomena makes the model much more tractable and allows us to focus on the question
we are mainly interested in, namely the allocation of surplus in matches between partners with different
productivities and the implication for investment incentives.
For most of the current analysis it shall be assumed that the discount factor δ is sufficiently small that
it is always efficient to trade, regardless of the type of your partner, rather than wait. Hence the option to
wait will act as a constraint on the current trade, an assumption that is discussed in more detail in the next
section.
These assumptions greatly simplify the strategy space. When a player first enters the game she chooses
I ∈ {h, l}, after which point she learns her type T ∈ {H, L} . Given her type, each period she needs to
formulate only her demand when faced with a partner of a different type, since she adopts the equal split
rule when faced with a partner of the same type. Formally, a strategy of the stage game is given by
(I, xHL , xLH ) ∈ {h, l} × X(k)2 , where X (k) = XLH (k) = XHL (k), but in every period other then the
period she enters an agent only has to determine one action, namely xHL if she is of type H or xLH if she
is of type L. For convenience let xH = xHL , denote the strategy of the high type when paired with a low
type, while xL = xLH is the strategy of a low type when paired with a high type. In what follows we will
refer to the pair (xH , xL ) as the bargaining strategy of an agent.

3 Equilibrium Analysis
Our goal is to understand the structure of the stochastically stable equilibria as a function of the cost
of investment, c, the degree of investment complementarity, SA , and the degree of investment specificity,
modeled by δ. The purpose of this section is to characterize the uniform sequential equilibria in stationary
10 Young (1993b) has shown in a two population model that the equal split is stochastically stable when both populations have
identical characterisitcs. In his model contrary to ours there exist however conventions where the surplus is not split equally
between the partners from the two populations although they have identical characteristics.

8
strategies of the population game11 that result in high investment. It will turn out that if stochastically
stable equilibria exist they are indeed in this class of equilibria.
Note that in the Nash demand game any strategy profile (xH , xL ) such that xL + xH = SA is a Nash
equilibrium. By a bargaining norm we mean a situation where all individuals have identical bargaining
strategies of the form (SA − x̂L , x̂L ) for some x̂L ∈ [0, SA ].
Since the focus of this paper lies on the bargaining behavior in matches of different types we will make
assumptions that guarantee that equal split trades always occur between equal types. Given our assumption
that surplus is split equally between equal types if trade occurs, we only have to be concerned about the
question whether equal types want to trade at all or rather wait for a different type. The maximal payoff a
SL
low type can get in the next period is SA and therefore 2 > δSA is sufficient to guarantee trade between
SH
low types. For high types we must have 2 > δSA which clearly is a weaker condition. Hence we will assume
throughout the paper that
SL
(1) δ< .
2SA
Considering High-Low pairings and observe that for relatively high discount factors and strong comple-
mentarity between investments, even if a bargaining norm exists, one of the two partners would rather wait
for a partner of identical type than to trade according to the bargaining norm. Given that in a High-Low
pairing the high type expects a low bid of x̂L , the low type expects a high bid of SA − x̂L and given that
both partners believe that they will meet an identical type in the following period, they will be willing to
trade if

SA − x̂L > δSH /2,

x̂L > δSL /2.

The first condition ensures that the high type prefers trading with a low type, rather than waiting one
period and trading with a high type. The second condition is the corresponding requirement for the low
type. Adding these inequalities together implies the following necessary condition for trade to occur for HL
matches:
2SA
(2) > δ.
SL + SH
Put differently, (2) implies that there exists a bargaining norm xL such that individuals always trade in
High-Low matchings regardless of their beliefs concerning the distribution of types in the population. Notice
that condition (2) can not be binding, if investments are substitutes. Investments are substitutes if the
marginal return from the first investment is greater than from the second investment:

SA − SL > SH − SA ,
2SA
> 1.
SL + SH
11 This means that we consider scenarios where all individuals use identical strategies of the stage game every period and
these strategies are constant over time.

9
Conversely, investments are complements if the marginal return from the second investment is larger:

SA − SL < SH − SA ,
2SA
< 1.
SL + SH
In this case, when δ is large it may be more efficient for HL pairs not to trade, and instead to delay trade
until they meet a partner of the same type. For further reference, the requirement that there is a bargaining
norm that implies trade in HL pairings regardless of the individual beliefs about the type distribution is
summarized as the trade condition:

2SA
Definition 1 The discount rate δ satisfies the trade condition if δ < SL +SH .

It shall be shown below that this is a necessary condition for the existence of a stochastically stable
bargaining norm when investments are complements. By a norm we mean a pair {I, x̂L } , with the interpre-
tation that each agent selects the investment I upon entering the market, the low type demands x̂L , while
the high type demands x̂H = SA − x̂L . To economize on writing out the full set of strategies and payoffs,
the notion of a self-enforcing norm is be defined as follows.

Definition 2 A norm {H, x̂L } is self-enforcing if:

¡ SL
¢
1. (1 − λ) (SH /2 − x̂L ) + λ (SA − x̂L ) − 2 ≥ c/ (1 − 2λ) ,
(1−λ)
2. SA − x̂L ≥ δ (1−δλ) SH /2
λ
3. x̂L ≥ δ (1−δ(1−λ)) SL /2.

The expected payoff of a person making a high investment assuming that trade is immediate and she
meets a high type is (1 − λ) SH /2+λx̂L , while the result of no investment is λSH /2+(1 − λ) x̂L . If she meets
a low type, the expected payoffs are (1−λ)(SA − x̂L )+λSL /2 if she invests high and λ(SA − x̂L )+(1−λ)SL /2
if she invests low. Given the expected equilibrium fraction of high types in the market in any period is (1−λ)
a simple calculation yields condition 1. The second condition is the requirement that a person who is a high
type prefers to trade with a low type, rather than wait until meeting a high type. The final condition requires
the low type to prefer trading with a high type, rather then waiting until meeting a low type. This places
a lower bound on x̂L . It is a straightforward exercise to show that for every self-enforcing norm there is a
sequential equilibrium yielding this outcome for the trading game outlined above. A self-enforcing norm,
{L, x̂L } , for low investment is defined in a similar fashion.
For much of the analysis the parameter λ is positive, but small. In the limit when λ = 0, a sufficient
condition for the existence of a self-enforcing norm with high investment is that it is efficient.

Proposition 1 Suppose it is strictly efficient for all agents to select high investment, SH −2c > max {SA − c, SL } ,
then for all δ satisfying the trade condition a bargaining norm, x̂L , exists such that {H, x̂L } is a self-enforcing
norm for λ sufficiently small.

10
 This result demonstrates that when noise is small it is possible to support as an equilibrium high invest-
ment whenever it is efficient to do so. In contrast, the literature on the holdup problem assumes that the
ex post division of the surplus is determined by the Nash bargaining solution, which in some cases induces
inefficient investment. However the division implied by the Nash bargaining solution is only one among many
sequential equilibria of the game. In general, one is able to conclude that for this game there are a large
number of sequential equilibria, many of which induce efficient investment. The question then is whether or
not the efficient equilibria are (stochastically) stable.

4 Learning Dynamics
Consider now the kind of bargaining and investment norms that are developed endogenously in a population
of adaptive agents. Following Young (1993a) and Kandori, Maillath, and Rob (1993) it is assumed that
agents sample the previous periods trades to build an empirical distribution regarding the investment and
bargaining behavior of the other individuals in the population (see Young (1993b) for the application of this
approach to the Nash bargaining game). Regarding the value of the outside option, agents believe that the
distribution of low and high types in the economy is time stationary, a hypothesis that is consistent with the
assumption that agents base current actions on past observations of the frequency of high types. It is also
assumed that with a small probability they make mistakes in executing their optimal strategy given their
beliefs regarding the play of the game described in section 2.
Our model consists of a single population of individuals who choose investment from {h, l} upon entering
the population and afterwards have to choose their action from the space X (k) every period until they
trade and leave the population. This choice is based on beliefs about distribution of types and bargaining
behavior of the other individuals in the population. Each period every individual independently takes a
random sample of m individuals from the previous period observing the type and the demand made at
the bargaining stage. This sample is added to the memory of the individual thereby replacing some old
observations12 .
Using the data in her memory each individual generates beliefs about the fraction of types H in the
population and the distribution of demands made by other individuals in HL and LH pairings. Each of
these beliefs is based on m observations, hence there is a finite set of possible beliefs we denote by B. For each
β ∈ B we denote by p̂ (β) the estimated proportion of high types, by F̂H (xH , β) the estimated probability
that xH or less is demanded by a high type in a HL pairing and by F̂L (xL , β) the estimated probability that
xL or less is demanded by a high type in a LH pairing. Put differently, F̂H and F̂L are empirical distribution
functions given the observations in the memory of the individual. It will turn out to be convenient to denote
by P(z) the distribution function of point expectations z, i.e. P(z)(x) = 0 for x < z and P (z)(x) = 1 for
x ≥ z. When an agent leaves the market, her beliefs are passed on to the new agent entering the market to
replace this agent. Beliefs in the first period are arbitrary.
The structure and time-line of the game with adaptive dynamics is summarized as follows (see also
12 An exact mathematical description of the belief formation and learning dynamics considered as well as the associated belief
and state spaces is given in Appendix A

11
Appendix A):

(i) At the beginning of the game beliefs are random, but when an individual leaves she is replaced by
another agent with the same beliefs, say β.

(ii) Investment decisions are only made by agents entering the population in the current period. Given
her beliefs, an agent chooses to invest if the expected gain from investment exceeds investment costs c
under the assumption of optimal behavior on the bargaining stage. Then she draws her type, which is
equal to her investment with probability 1 − λ.

(iii) Each period the following steps are repeated until exit occurs:

1. At the beginning of every period t the individual randomly samples the types of m individuals
from the previous period. This is used to update beliefs bit ∈ B.
2. With probability ε > 0 the individual selects an action randomly from X (k) , under the uniform
distribution. This noise process is i.i.d. between individuals and periods. With probability 1 − ε
the individual determines which demand maximizes the expected payoff under her beliefs if she
is matched with a different type.
3. Agents are randomly paired, and their payoffs are determined. If the partners are of identical
type, there is an equal split, otherwise they chose the actions determined at stage 2.
4. If trade occurs, both agents leave and are replaced with identical agents who begin at step (ii).
If not, step (iii) is restarted.

Given that an agent’s action is completely determined by her beliefs bit ∈ B, and type T i ∈ {H, L}13 , the
state at time t is characterized by a distribution over beliefs and types, and accordingly there is a finite state
space we call S. The learning process described above defines a time homogeneous Markov process {σ t }∞
t=0
on the state space S. Although, even for ² > 0, the transition matrix is not positive, the following lemma
shows that the process is irreducible and aperiodic.

Lemma 1 For ² > 0 the Markov process {σ t }∞

t=0 as defined above is irreducible and aperiodic.

Hence, for ² > 0 there exists a unique limit distribution π ∗ (²) over S, where π ∗s (²) denotes the probability
of state s. Following a standard approach in evolutionary game theory we consider the limit distribution
for small values of ² and in particular characterize the states whose weight in the limit distribution stays
positive as the mutation probability ² goes to 0. Such states are called stochastically stable:

Definition 3 A state s ∈ S is called stochastically stable if lim²→0 π ∗s (²) > 0. We say that a set is stochas-
tically stable if all his elements are stochastically stable.
13 We look at the process after all incoming agents have made their investment decisions, but before they are paired and
therefore the type of all agents is determined.

12
 The reason why this concept is of interest is that for small ² the process spends almost all the time in
stochastically stable sets. Hence, characterizing the stochastically stable outcome means characterizing the
long run properties of the evolutionary process. To identify stochastically stable states it is necessary to first
identify the minimal absorbing sets of the process for ² = 0. It is well known that the set of stochastically
stable states is a subset of the union of these so called limit sets. Formally, a limit set is defined as follows:

Definition 4 A set Ω ⊆ S is called a limit set of the process if for ² = 0 the following statements hold:

∀s ∈ Ω IP(σ t+1 ∈ Ω|σ t = s) = 1

∀s, s̃ ∈ Ω ∃z > 0 s.t. IP(σ t+z = s̃|σ t = s) > 0.

In the following sections we will characterize the stochastically stable sets and discuss the implied invest-
ment and bargaining norms.
The question we address is the emergence of a unique, efficient and stable bargaining norm in which all
individuals follow the same investment strategy, and have the same expectations regarding how to divide
the gains from trade. This is formally defined by:

Definition 5 A state s induces the bargaining norm xL if all individuals have beliefs β ∈ B that place
probability one on the demand by their partner being xL or SA − xL , depending upon their type in HL
matches.14 If all stochastically stable states induce the same bargaining norm we say that this bargaining
norm is stable.

Therefore, we shall say that a bargaining norm does not exist at a state s if there is heterogeneity in
the beliefs of the agents regarding the terms of trade between high and low types. Let us now consider the
constraints that the outside options place upon feasible bargaining norms.

5 Deterministic Investment Effects

Before we explore the stochastically stable norms of the model described above we discuss briefly the im-
portance of our assumption that investment effects are stochastic (λ > 0) for the evolution of bargaining
norms. This is particularly important since the results of Ellingsen and Robles (2002) and Troger (2002)
show that in cases of one-sided investment the stable bargaining norm always induces efficient investment
when λ = 015 .
Dawid and MacLeod (2001) study the two-sided investment model presented for the case λ = δ = 0. Their
findings concerning the evolution of bargaining norms can be summarized as follows (compare Proposition
4 and Proposition 7 in Dawid and MacLeod (2001)):

Proposition 2 For deterministic investment effects (λ = 0) and relationship specific investments (δ = 0)

the stochastically stable set always includes states where individuals have heterogenous beliefs about bargaining
behavior. Hence, there is no stable bargaining norm.
14 Formally F̂H (β) = P(SA − xL ), and F̂L (β) = P(xL )).
15 Also, they only consider the case of relationship specific investments (δ = 0).

13
 The assumption that δ = 0 is not crucial for this finding and the result would still hold for δ > 0. The
intuition is that, once the investment strategies in the population are uniform, any bargaining norm which
might exist at that point will be slowly destroyed. Under identical investment strategies and deterministic
investment effects the only way a pairing between a high and a low type might occur is that at least one of
the two has mutated. A mutant may not follow the bargaining norm and hence at least half of the demands
in high low pairings are completely random and do not follow any bargaining norm. Since all individuals
use these demands to update their beliefs, any uniform consistent point beliefs which might have existed in
the population will be destroyed, and beliefs about bargaining behavior between high and low types keeps
drifting around in the space of possible beliefs. Therefore, stable bargaining norms between high and low
types cannot evolve for deterministic investment effects.
This drift of beliefs is also present in the scenario with one-sided investment and eventually leads to an
outcome where investors who invest efficiently get a sufficiently large part of the surplus that they have no
incentive to change investment regardless of their beliefs about the allocation of surplus for other investment
levels. In the case of two-sided investments one of the two partners will always have incentives to change
her investment level if she believes that such a change increases her fraction of the surplus by a sufficient
amount. Therefore, for the case of two sided deterministic investments there are not only no stochastically
stable bargaining norms but also investment levels are in general inefficient.

6 Stochastically Stable norms

The findings reported in the previous section suggest that the uncertainty of investment effects should have
an important positive role for the evolution of stochastically stable bargaining norms. For λ > 0 high-low
matches occur with positive probability even after an investment norm has been established and therefore
the drift of beliefs which is responsible for the continuous destruction of norms in the deterministic case
cannot occur. In this section we return to the case λ > 0 and explore whether we always get stochastically
stable bargaining norms.
A necessary condition for the evolution of a norm is that the terms of trade between high and low types
result in outcomes that are better than their respective outside options. By simply waiting for a partner with
the same type an agent can guarantee a non-negative expected payoff, where the size of the expected payoff
depends on the agents’ beliefs about the distribution of types and the value of δ. We say that a bargaining
norm is compatible with p̂ and δ if both parties are better off than their respective expected outside option.
Denote the set of all bargaining norms which are compatible with all p̂ ∈ [0, 1] for a certain discount factor
by C(δ). We get the following lemma, where α = αLH = αHL is the minimum unit of account for dividing
the surplus, as defined in the game form of section 2.

Lemma 2 The set C(δ) is non-empty for sufficiently small α if and only if the trade condition holds.

Let us now characterize the limit sets in this framework. Once a bargaining norm xL , which is compatible
with δ, is reached, in the absence of mutations all low types always demand xL against high types and high
types always demand SA − xL against low types. Hence, beliefs can never change once such a state has been

14
reached. If beliefs are heterogeneous there is always a positive probability that all agents observe identical
samples and beliefs become homogeneous and compatible. However, after a bargaining norm has been
reached the distribution of agent types may change between two periods, even if the investment behavior is
constant. The randomness of the outcome from investment implies that all distributions of H and L types
are possible. Hence, if there is a bargaining norm that is not compatible with all p ∈ P, eventually it will be
disrupted. This suggests that the limit sets correspond to norms in C(δ), when it is not empty. Given that
the trade condition holds C(δ) is never empty and we always have bargaining norms as the possible long run
outcome of our process. This is formalized in the following lemma.

Lemma 3 Suppose that k is sufficiently large and the trade condition holds. Then for each xL ∈ C(δ) there
exists a limit set Ω(xL ) consisting of all s ∈ S such that sζ > 0 only if ζ = (T, β) for some some T ∈ {H, L}
and some β such that F̂H (·, β) = P(SA − xL ) and F̂L (·, β) = P(xL ).

If the trade condition does not hold then C(δ) = ∅, and the outside option of waiting for an equal type
always becomes binding for some p̂. In the following proposition we show that in such a scenario bids never
settle down at a compatible norm but there occur persistent fluctuations driven by the fluctuations in the
p̂i . Long run bargaining behavior is then characterized by ergodic behavior on a set of different bids.

Proposition 3 Suppose that the trade condition does not hold, then for sufficiently large m, n, and k there
is a unique stochastically stable set L where beliefs about demands as well as induced actual demands do
not coincide for all individuals in all states contained in L. Accordingly, there exist no stochastically stable
bargaining norms.

When investments are complements, there is a δ̄ < 1, such that for all δ > δ̄ the trade condition is not
satisfied. This demonstrates that if the market is sufficiently competitive and investments are complements,
then it is not possible for a bargaining norm to evolve. This does not imply that increasing market competi-
tion results in inefficiency. When the trade condition does not hold, then LL and HH matches are the most
likely trades, and hence if high investment is strictly efficient, (SH /2 − c > SL /2) individuals often find it in
their interests to invest.
The question now concerns the nature of the fair division rule when it is efficient for HL pairs to trade,
and therefore the remainder of the paper assumes that the trade condition is satisfied. Under this assumption
we always have a whole set of potential long run norms to be reached. Clearly the investment incentives
depend on which of these norms are reached in the long run. The characterization of the limit sets in
lemma 3 shows that productivity types will keep fluctuating even after bargaining behavior has settled down
at a norm. Investment incentives however do not only depend on the bargaining norm but also on the
distribution of productivity types in the population. So, in order to understand the long-run evolution of
investment behavior of the individuals we first have to examine which of the possible bargaining norms is
selected in the long run and then given this bargaining behavior have to study the dynamics of productivity
type distributions under our investment rule. The long run properties of the productivity type distribution
together with the bargaining norm then determine which investment decisions are made in the long run. In
the following two sections this analysis is carried out separately for the case where investments are substitutes
and complements.

15
7 The Case of Substitutes
7.1 Stochastically Stable Bargaining norms
We know already that the trade condition always holds if investments are substitutes which implies that in
the case of investment substitutes there always exist bargaining norms as a potential long-run outcome of
our dynamic process.
The fact that individuals cannot perfectly determine the productivity of their investments has two im-
portant implications. First, every period there is a strictly positive probability of both types existing in
the market, and thus there are always with positive probability HL trades occurring in the market which
can be used to update the believes of individuals. Second, regardless of the investment decisions of indi-
viduals, any distribution of types has strictly positive probability. On the other hand, transitions between
bargaining norms have to be triggered by (in general multiple simultaneous) mutations. Hence, for small
mutation probabilities bargaining norms adjust more slowly, and are more stable than the realized distribu-
tion of types. This implies that the stochastically stable bargaining norm is independent of the long run
investment behavior and therefore also independent of investment costs c. The next proposition provides
a rigorous proof of this fact and derives the properties of the bargaining norm that arises in the long run.
It turns out that the qualitative properties of the process depend on the degree of substitutability between
investments. We say that investments are weak substitutes if 12 (SH + SL ) ≤ SA ≤ S̄ := SH − 2δ (SH − SL ).
For S̄ < SA ≤ SH investments are strong substitutes.

Proposition 4 For sufficiently large m, n the limit of the stochastically stable sets of the process {σ t } for
k → ∞ can be characterized as follows:

(a) If investments are weak substitutes every stochastically stable state induces the bargaining norm

SA δ
x̂SA = − (SA − SL ).
2 2(2 − δ)

(b) If investments are strong substitutes every stochastically stable state induces the bargaining norm

SA δ
x̂SA = − (SH − SL ).
2 4
Following definition 5 we will refer to the bargaining norm induced by all stochastically stable states as
the stable bargaining norm. In the absence of outside options (δ = 0) the equal split rule is the unique,
stable bargaining norm, regardless of investment levels.

7.2 Investment Norms with Substitutes

From proposition 4 we infer that the level of long run investment does not affect the outcome of bargains.
This greatly simplifies the analysis of investment behavior in the long run. We can determine investment
behavior as a function of the bargaining norm and then simply insert the stable bargaining norm. For a

16
given bargaining norm x̂L investment is optimal iff
µ ¶ µ ¶
SH SL
(1 − λ) p̂ + (1 − p̂)(SA − x̂L ) + λ p̂x̂L + (1 − p̂) −c
2 2
µ ¶ µ ¶
SL SH
≥ (1 − λ) p̂x̂L + (1 − p̂) + λ p̂ + (1 − p̂)(SA − x̂L ) .
2 2
Taking into account that (SH + SL ) /2 − SA < 0 this gives the following condition for high investment to be
optimal:
SA − x̂L − SL /2 − c/(1 − 2λ)
(3) p̂ ≤ p∗ (x̂L ; λ) := .
SA − SH /2 − SL /2
As pointed out above, to analyze the dynamics of investment for a given bargaining norm we consider the
evolution of type distributions over time. Given a current distribution of types the distribution of types in
the following period in general depends on the outcome of the stochastic sampling procedure for all agents,
which gives the beliefs p̂(bit ) and therefore influences the investment decisions, and the actual realization of
types given the investment decision. This can be described by a Markov process {σ̃ t }∞
t=0 on the state space
S̃ = {0, 1/n, 2/n, . . . , 1}. For λ > 0 the process is irreducible and aperiodic. The unique limit distribution is
denoted by π̃ ∗ (λ). The following lemma characterizes the limit distribution for small values of λ.

Lemma 4 When investments are substitutes, then given a bargaining norm x̂L , the long run distribution of
types for sufficiently small λ can be characterized as follows:

(a) p∗ (x̂L ; 0) ≤ 0: no individual ever invests and limλ→0 π̃ ∗0 (λ) = 1.

(b) p∗ (x̂L ; 0) > 1: all individuals always invest and limλ→0 π̃ ∗1 (λ) = 1.

(c) p∗ (x̂L ; 0) ∈ (0, 1): limλ→0 π̃ ∗1 (λ) = limλ→0 π̃ ∗0 (λ) = 0.5.

In case (a) we say that x̂L induces a no-investment norm, in (b) x̂L induces a full investment norm, and
in case (c) we say that x̂L induces cyclical investment. By cyclical investment we mean that in one period
everybody invests, and in the next period nobody invests. What is happening is when all individuals invest,
it is optimal not to invest, and vice verso.16 It should also be pointed out here that even if we consider the
limit of long run stochastically stable sets for λ → 0 this should be interpreted as the long run properties of
the process if both ² and λ are small but λ is of an order of magnitude larger than the mutation probability
². In other words, the transition between bargaining norms is always assumed to be much slower than the
transition between investment patterns. Using the previous lemma it is straight forward to describe the
investment behavior which is induced by the stochastically stable bargaining norms. We just have to insert
the stochastically stable bargaining norm x̂sL into p∗ and apply lemma 4.
16 Note that for substitutes in cases where p∗ (x̂L ; 0) > 1 the action H is dominant at the investment stage for small λ and all
heterogeneity in types is created by deviations of the actual type from investment. Therefore, it is easy to see that a bargaining
convention x̂L induces an investment convention if and only if there is a λ∗ > 0 such that the convention {H, x̂L } is stable for
all λ < λ∗ .

17
Proposition 5 Assume that m, n and k are sufficiently large.

(a) If investments are weak substitutes the stochastically stable bargaining norm induces full-investment for
c < c1 , no-investment for c > c2 and cyclical investment for c ∈ [c1 , c2 ], where

1
c1 = 2(2−δ) (δ(SA − SL ) + (2 − δ)(SH − SA ))
1
c2 = 2−δ (SA − SL ).

(b) If investments are strong substitutes the stochastically stable bargaining norm induces full-investment for
c < c3 and cyclical investment for c ≥ c3 , where
1
c3 = (δ(SH − SL ) + 2(SH − SA )).
4
Notice that when δ = 0, then c1 = (SH − SA ) /2, but in the case of substitutes it is efficient for both
parties to invest whenever c < (SH − SA ) . Therefore we obtain under-investment in some cases.
In the case of weak substitutes the gain from investing at the bargaining norm is:

(SH − SA ) δ
SH /2 − xL = + (SA − SL ) ,
2 2 (2 − δ)
(SH − SA )
≥ .
2
Therefore, the outside option increases the gains from investing, regardless of whether it is binding at the
equilibrium. However, for weak substitutes, it never increases incentives to the point that the gains from
investing are equal to the full marginal gains, given by (SH − SA ) . On the other hand, if investments are
strong substitutes and the gains from the second investment are very small (case (b) above) the stochastically
stable norm indeed induces full investment whenever this is efficient. This is formalized in the following
corollary.

Corollary 1 If investments are strong substitutes the stochastically stable bargaining norm induces full-
investment for all values of c where full investment is efficient.

Proof. It is straight-forward to check that c3 ≤ SH − SA under these assumptions.

In the summary of results below we will also provide a graphical illustration of these findings.

7.3 Impact of the Outside Option

Considering proposition 4, the effect of the outside options on the bargaining norm might be quite surprising.
Notice, that in this model the outside option is introduced only as a constraint on the set of possible
bargaining agreements, and hence one might expect the outside option principle to apply (see Binmore,
Rubinstein, and Wolinsky (1986)). In that case if xL > δSL /2 and SA − xL > δSH /2, then xL should not
depend on either SH or SL , yet we find that that for all δ > 0 the stochastically stable bargaining norm
depends upon at least one of the outside options, and that the low types share is always strictly increasing in
SL , a result that is consistent with Binmore, Proulx, Samuelson, and Swierzbinski (1995) who report results

18
from a bargaining game with drift. This might raise the question whether the efficiency result of corollary
1 is a simple implication of the difference in threat point payoffs of the two types.
To address this question let us denote by x̂N
L the allocation consistent with the Nash bargaining solution
between a high and a low type where both have beliefs p̂ = 1 and the expected payoffs in the following period
are treated as a threat point. This allocation has to satisfy
µ ¶
1 SH
x̂N
L = δ x̂N
L + S A − δ x̂N
L − δ ,
2 2
and therefore we get
SA δ(SH − SA )
(4) x̂N
L = − .
2 2(2 − δ)
Comparing this expression with the stochastically stable bargaining norms from proposition 4, simple cal-
culations show that under our assumption of SA > (SH + SL )/2 we always have x̂N S
L > xA . Accordingly,
the investment incentives in a population of investors under the stochastically stable bargaining norm are
not only larger than under the equal split rule but also larger than under Nash bargaining with the outside
options as threat points.
To understand this result intuitively we have to realize that the long run stability of the bargaining norms
are determined by their resistance to change in scenarios where deviations from the norm have the highest
chance of altering the norm. Bargaining norms are more easily destabilized in scenarios with low investment
in the population since the expected losses from disagreement when not giving in to demands of deviators
from the norm are the largest under this investment pattern. If investments are substitutes a high type has a
lot of bargaining power in an environment of low types and hence the stochastically stable bargaining norm
gives a larger part of the surplus to the high types than they would get if the norm had been evolved in a
population of mostly high types. Hence, the stochastically stable norm allocates more to the high types than
the Nash bargaining solution in an environment of high types would. Although developed in low investment
scenarios, the stochastically stable bargaining norm is adhered to even if in the long run everyone invests,
and hence it facilitates the development of full investment norms.
This discussion implies that the evolutionary learning facilitates the development of full investment. In
the following corollary we compare the stochastically stable outcome to the notion of a self-enforcing norm
under the equal split rule and the Nash bargaining solution:

Corollary 2 (a) If c satisfies

1
c≤ (SH − SA ),
2
then for λ sufficiently small the stochastically stable norm induces full investment, and {H, SA /2} as
© ª
well as H, x̂N
L are self-enforcing norms.

(b) If c satisfies
1 1
(SH − SA ) < c ≤ (SH − SA )
2 2−δ
© ª
then for λ sufficiently small the stochastically stable norm induces full investment, H, x̂N
L is a self-
enforcing norm but {H, SA /2} is not a self-enforcing norm.

19
 (c) If c satisfies
1 δ 1
(SH − SA ) < c < (SH − SL ) + (SH − SA ),
2−δ 4 2
then for λ sufficiently small the stochastically stable norm induces full investment, but neither {H, SA /2}
© ª
nor H, x̂NL are self-enforcing norms.

Note that it follows from this corollary that the Nash bargaining solution x̂N
L never implies efficient
investment in the sense that there is always a range of cost values c where high investment is efficient but
© ª
H, x̂N
L is not a self enforcing norm. As we know from corollary 1, the stochastically stable norm does induce
efficient investment if investments are strong substitutes. These results demonstrate that the decrease of
the size of the hold-up region under the evolutionary dynamics compared to the equal split rule is not a
simple implication of the existence of outside options. Rather, the dynamic interplay between investment
and bargaining decisions is responsible for the increased long run investment in our evolutionary setting17 .

8 The Case of Complements

Consider now the case of complementary investments, where (SH + SL ) /2 ≥ SA ≥ SL . We know already
from proposition 3 that no norms evolve if the trade condition is violated. Therefore, we assume throughout
this section that the trade condition holds.
Under this condition there is again a unique stochastically stable bargaining norm which is independent
of beliefs regarding the fraction of high types in the market. The properties of this bargaining norm depend
on the degree of complementarity of investments. Analogous to the substitutes case we call investments weak
complements if (SH + SL ) /2 ≥ SA ≥ S := 2δ ((2 − δ)SH + SL ), and strong complements if S > SA ≥ SL .

Proposition 6 Suppose the trade condition holds, then for sufficiently large m, n the limit of the stochas-
tically stable sets of the process {σ t } for k → ∞ can be characterized as follows:

(a) If investments are strong complements the stochastically stable bargaining norm is
SH
x̂SA = SA − δ .
2
(b) If investments are weak complements the stochastically stable bargaining norm is
SA δ
x̂SA = − (SA − SL ).
2 2(2 − δ)
Case (a) occurs when the outside option for the high type is binding for p̂ = 1. This can only happen if
S > SL . A necessary and sufficient condition for this to apply is:
SL
δ≥ .
SH
One of the implications of the stochastic stability criteria is that, as in the case of substitutes, the existence
of an outside options always increases the payoff for the high type relative to the equal division solution. On
17 If we restricted the model to only a single possible investment level the resulting stochastically stable bargaining convention
would exactly match the Nash bargaining solution with the outside option as threat point.

20
the other hand, it can be easily verified that the Nash bargaining solution with the outside option as threat
point, given by (4), gives a smaller allocation of the surplus to low types compared to the stochastically
stable norm and therefore provides higher investment incentives.
One can explore the maximum incentives possible, while ensuring the existence of a bargaining norm by
2SA
supposing that the trade condition is satisfied with equality, namely δ = SL +SH . In this case

δ
S = ((2 − δ)SH + SL )
2
δ (1 − δ)
= SH + δ(SH + SL )/2
2
δ (1 − δ)
= SH + SA > S A ,
2
and hence we are in the case of strong complements, and the stable bargaining norm is given by:
SH
x̂SA = SA − δ ,
2
SL
= SA .
SL + SA
This result illustrates the effect that the low payoff plays in determining the bargaining norm. When
SL is close to zero (the payoff in the absence of trade), then with sufficient competition one obtains first
2SA
best incentives, while ensuring the existence of a bargaining norm. When δ = SL +SH and c < SA , then low
investment is not an equilibrium for SL = 0, and we have high investment in this case. However, in other
cases, both high and low investment choices may be equilibria given the bargaining norm. The next sections
explores which these equilibria is stable.

8.1 Investment Norms with Complements

Taking the bargaining norm as fixed, investment decisions have the structure of a coordination game when
investments are complements. Incentives are larger for p̂ = 1 than for p̂ = 0. This implies that if the
bargaining norm induces no investment for p̂ = 1, no individual will invest any more, once the bargaining
norm has been established regardless of their beliefs p̂ – a no-investment norm is induced. On the other
hand, if investment is optimal at p̂ = 0, everyone invests under the stochastically stable bargaining norm
– a full investment norm is induced. However, if investment is optimal for p̂ = 1 and no investment is
optimal for p̂ = 0, both the homogeneous state corresponding to full investment and the homogeneous state
corresponding to no investment are locally stable states in the sense that the process never leaves each of
these states as long as high investment always implies high types and low investment always implies low
types. In such a scenario the threshold p∗ defined in (3) is in (0, 1) and investment is optimal if and only if
p̂ ≥ p∗ (x̂L ).
Investment effects are assumed to be stochastic, and therefore there is always a positive probability that
the process wanders from a non-investment to a full investment state and vice versa. As in the case of
substitutes, for a given bargaining norm the evolution of the distribution of high and low types is described
∗
by a Markov process {σ̃ t }∞
t=0 on the state space S̃. This process has a unique limit distribution π̃ (λ). We are
again interested in the the properties of this limit distribution if the probability of deviations of productivity

21
types from investment becomes small with the understanding that λ is still an order of magnitude larger
than ².

Lemma 5 Assume that 0 < λ < 0.5 and m and n are and sufficiently large and a bargaining norm xL is
P P
given. Then, for p∗ (xL , λ) > (<)0.5 we have i<n/2 π̃ ∗i/n (λ) > (<) i>n/2 π̃ ∗i/n (λ). Furthermore, we have
limλ→0 π̃ ∗0 (λ) = 1 if p∗ (xL , 0) > 0.5 and limλ→0 π̃ ∗1 (λ) = 1 if p∗ (xL , 0) < 0.5.

According to this lemma p∗ (xL , 0) < 0.5 implies that in the long run the probability to have a majority of
high types is larger than the probability to have a majority of low types and as λ goes to zero the probability
to see only high types goes to one. Since investment decisions have the structure of a coordination game
here this lemma basically rephrases well known results by Kandori, Maillath, and Rob (1993). We say that a
no-investment norm is induced if the threshold p∗ (xL , 0), is larger than 0.5 and that a full investment norm
is induced if this inequality holds the other way round. Using this we get the following characterization of
the investment norms induced by stochastically stable bargaining norms.

Proposition 7 Assume that m, n and k are sufficiently large, the trade condition holds, and investments
are complements, then the stochastically stable bargaining norm induces full investment if c < c4 (SA , δ) and
no-investment for c > c4 (SA , δ), where
(
1 1
4 4 (SH − SL ) + 2 (δSH − SA ) if SA ≤ S,
(5) c (SA , δ) = 1 δ
4 (SH − SL ) + 2(2−δ) (SA − SL ) if SA > S.

In the case of investment complements high investment is efficient (for small λ) whenever c < (SH −SL )/2,
but is only stochastically stable for c < c4 < (SH − SL )/2. If investments are complements there always
remains a hold-up region with inefficient investments in the long run. It follows from the coordination game
structure of the investment stage that a bargaining norm x̂L does not necessarily induce a high investment
norm even if {H, x̂L } is a self-enforcing norm. An implication of this, especially when compared to the
case of substitutes, is that the set of parameters for which a full investment norm is self-enforcing under
the equal split rule might be larger than the set of parameter values for which high investment is part of
1
a stochastically stable equilibrium. To see this, notice that if SA > δSH and c < 2 (SH − SA ) the norm
{H, SA /2} is self-enforcing for sufficiently small λ. Comparing this bound on investment costs with c4 we
get

Corollary 3 For · ¸
(2 − δ)SH + (2 + δ)SL )
SA < min SL + 2(1 − δ)SH ,
4
© ª
the interval [c4 (SA , δ), (SH − SA )/2] is non-empty and if c belongs to this interval H, S2A is a self-enforcing
norm for sufficiently small λ, but there is no stochastically stable norm with full investment.

The fact that there is a self-enforcing norm with high investment norm does not imply uniqueness of the
equilibrium. Even if a high investment norm is self-enforcing there might be a coexisting low investment
equilibrium with the corresponding equilibrium selection problem. To deal with the equilibrium selection
problem we could use the concept of risk dominance as an equilibrium selection device and say that a

22
bargaining norm induces high investment only if the full investment is the risk dominant equilibrium at
the investment stage18 . It is straightforward to check then that the maximal investment costs inducing
high investment under the equal split rule is always below c4 . So, taking into account the coordination
problems arising at the investment stage, the equal split rule again provides less investment incentives than
the stochastically stable norm. The Nash bargaining solution, on the other hand, provides for the case of
investment complements always larger incentives than the stochastically stable norm19 .
Overall these results illustrate that in the case of complements, stochastic stability implies that the
holdup problem is even more severe than would be predicted with in a standard incomplete contract model
with renegotiation.

9 Summary of the Results

Our results can be summarized with figures illustrating the relationship between the cost and productivity
parameters, {SA , c} and the equilibrium investment level. Consider first the case in which investments are
purely specific, namely δ = 0. In this case the equilibrium bargaining norm is always the equal split rule.
Hence, the only possible candidate equilibria are those investment levels that are also equilibria when agents
use the Nash bargaining solution. The possible outcomes are illustrated in figure 1 below, where SL = 1 and
SH = 2.

—– Figure 1 Goes Here ——

The illustrated trapezoid region represents all the {SA , c} combinations for which it is efficient for
there to be full investment. Holdup occurs in the region above the diagonal line between {SH , 0} and
{SL , (SH − SL ) /2} , because for these values high investment is not a stable outcome (SA /2 > SH /2 − c) .
Notice that when investments are substitutes then in the region between the lines C1 and C2 investments
cycles between high and low.
In the region below the diagonal line, even though all parameter combinations have a self-enforcing
high investment norm, when investments are complements, then only the region below line C4 entails high
investment at the stochastically stable equilibria. These results are in contrast to the results of Ellingsen and
Robles (2002) and Troger (2002) who obtain the efficient outcome when there is investment by one party
alone. Our results suggest that the allocations result in even less investment than suggested by Grossman
and Hart (1986).20 Suppose that one could allocated all bargaining power to one party, then the condition
18 The use of risk dominance as the selection criterium is appropriate because it is well known that the risk dominant
equilibrium coincides with the stochastically stable one in coordination games.
19 This can be easily checked by realizing that the condition for HH to be the risk dominant equilibrium is p∗ (x̂N , λ) < 0.5
L
and inserting the corresponding expressions in order to calculate the upper bound on c.
20 See also Che and Hausch (1999) who show that with two sided investment and efficient renegotiation it is not possible to

implement the first best. Their model is different than ours since it precludes the possibility of agents inefficiently refusing
to trade when demands are not compatible. In our model such efficient renegotiation is an implication of the model, not an
assumption.

23
for high investment by the owner is given by:

SA − c ≥ SL .

The region for which allocating all bargaining power to one party is more efficient than the stable equilibria
in the market with endogenous bargaining norms is illustrated in figure 2. Thus, we may conclude that
the allocation of stable property rights to one party, even though both parties can make investments, can
enhance efficiency.

—– Figure 2 Goes Here ——

The alternative solution to the allocation of bargaining power is to increase competition. The effect of the
outside option on investment is illustrated in figure 3 for the case in which the discount factor is δ = 0.5, and
we continue to suppose SL = 1 and SH = 2. In this case line C4 now moves up, and hence the set of cases
for which investment is high increases. Moreover, in the case of substitutes one now has high investment in
a region where there would not be high investment in either the standard holdup model (the lower line) or
under Nash bargaining with an outside option (the upper line), as discussed in corollary 2. Also notice that
there is the possibility of overinvestment when investments are highly substitutable (SA is close to SH ).

—– Figure 3 Goes Here ——

The model also provides some insights into the interplay between market competition and fair division
norms. In our model, individuals who are similar use the equal split rule, that is HH and LL matches. The
issue of what constitutes a fair division becomes an interesting question precisely when individuals vary in
observable ways. When shielded from competition we find that the equal split rule remains the stochastically
stable norm, however, when the trade condition is not satisfied, then high types prefer not to associate with
low types, and there is no opportunity for a bargaining norm to evolve. In this case, the outside option
increases the incentives to invest, but also results in a breakdown of agreements between low and high types,
which may leave unmatched individuals in the market.21
One should probably not take the analogy too far, but this result does seem consistent with the impact of
markets on agrarian economies. In the absence of markets all family members work on the farm, and hence
the notion of unemployment is simply not well defined. This is no longer the case when there are markets
because individuals spend time looking for better matches, resulting in observed unemployment. In those
cases, market prices are substituted for the norms that have evolved to divide the gains from family labor.
21 See MacLeod and Malcomson (1993), Felli and Roberts (2000) and Cole, Mailath, and Postlewaite (2000) who provide
conditions under which competition, even if imperfect, may provide incentives for efficient two party investment.

24
10 Discussion
In the literature on contracts and organization there is tension between research on mechanism design
demonstrating that efficient allocations can be implemented under a wide variety of situations, and the
fact that in practice one rarely observes many of these mechanisms.22 An example of this tension is the
recent contribution of Maskin and Tirole (1999) demonstrating that one can achieve the first best in the
hold-up model of Grossman and Hart (1986). Yet, the work of Grossman and Hart (1986) built upon a long
literature based upon the idea that ownership and the allocation of property rights do matter for economic
performance.23
The issue then is how can one formally capture the intuition some contractual arrangements - allocation
of bargaining power to one agent, the equal division rule - seem stable, while others, such as the threat to
punish an individual for a small deviation from an agreement, seem inherently unstable. The purpose of this
paper is to begin to address this question by applying the evolutionary learning approach of Young (1993a)
and Kandori, Maillath, and Rob (1993) to a simple model in which individuals make relationship specific
investments, bargain over the terms of trade and then trade or move back into the market.
Our results shed light on how information and potential competition can affect the stability of equilibria.
Our first observation is that the previous work by Troger (2002) and Ellingsen and Robles (2002) can be
interpreted as showing that allocating all bargaining power to one agent is stochastically stable in the case of
one sided investment. As we have shown in Dawid and MacLeod (2001), one can make this an if and only if
statement. When one adds two sided investment to their model, there is no stable equilibrium. The reason
is that the efficient equilibrium depends upon off equilibrium actions that occur with zero probability, and
hence stable beliefs regarding the consequences of those actions do not evolve.
This result may shed light on the tension between the economic approach to contract and legal practice.
Economists have long argued that courts should enforce contracts as written based upon the idea that
voluntary contracts are an expression of individual intent. In practice, the courts often over-rule explicit
contract term in favor of terms consistent with past practice, or as a function of past judicial decisions for
similar cases. In addition, parties may consciously leave a contract incomplete and instead relying upon
courts to fill in missing terms should there be litigation.24
The common principal in both cases is that past practice is often a more binding constraint upon a
contract than express terms. Moreover, when deciding upon which terms to include in a contract, parties
can be more certain of terms that have been previously tested in court. For this reason the American Institute
of Architects publishes a guide to contract cases that arise in their use of the AIA form contracts, called the
Legal Citator, and highlights cases in which the terms of the contract have been enforced as written.25
22 See Moore (1992) a review of the mechanism design literature, and Tirole (1999) for an evaluation of its relevance for
contract theory.
23 See for example the work of Coase (1937), Alchian and Demsetz (1972), Klein, Crawford, and Alchian (1978), and

Williamson, Wachter, and Harris (1975).

24 See MacNeil (1974) industry practice, and its role in the law.
25 Stein, Steven G. M. ed. The American Institute of Architect Legal Citator. New York: Matthew Bender, 2003 - updated

annually. See www.aia.org for more information regarding the legal citator and the forms that the AIA sells for use in the
construction industry.

25
 The Legal Citator plays a role that is not too dissimilar to the role of beliefs in our model. The citator is
merely a subsample of past cases, which attorneys can use to help their clients write enforceable contracts.
Notice, that in this context a contract is enforceable because at some point in the past a court ruled that it is
enforceable. This is quite different from the way an economist thinks about a contract, who includes a clause
for a contingency because she anticipates the consequence of that contingency. In practice, a lawyer includes
a contingency because the event has occurred in the past, and she wishes to protect her client should that
event occur again in the future.
We show that norms can evolve for the division of the gains (or losses) in any event only if they occur
with sufficient frequency. In our model we achieved this by supposing investment has a stochastic effect on
productivity and find that this implies the existence of a stable bargaining norm that is independent of the
value of the match.
This implies that when investments are complementary, the level of investment is lower than even the
level predicted by the hold-up model. The fact that the low investment equilibrium is stable implies that
experimentation and learning are not likely to lead one to a more efficient equilibrium. In the case of
substitutes, we find that the level of investment is in general higher than is predicted by the hold-up model.
In either case, we find that the egalitarian bargaining norm is stable, and hence results are consistent with
Platteau (2000)’s claim that the egalitarian norms that are prevalent in rural Africa may be a barrier to
the progress in that region. In particular, the results emphasize the idea that the egalitarian norm is most
problematic in a production setting where one depends upon the complementary inputs of several individuals.
The economist’s solution to this problem is to increase competition. If investments are interpreted as
human capital investments, then increased competition means that individuals should have several potential
buyers for any skills that they acquire. We explore the implications of increased competition by giving
individuals the option to delay trade into the future in the expectation of finding a better match. Thus by
increased competition we mean thicker markets ex post, as opposed to increased competition at the ex ante
stage, when individuals make their investment decisions.26
We find that adding such competition does indeed enhance efficiency, however when the market is very
competitive, there is a tendency towards associative matching (only high-high and low-low matches trade).
As a consequence, we again have a breakdown in the existence of stable bargaining norms for the high-low
trades, and it is possible for the market to be less efficient than it would be in the absence of competition.
This result provides another way to think about the downturn in productivity observed during the transition
in Eastern Europe, as documented in Roland (2000).
If increased competition is expected to endure, then individuals have an incentive to find ways to match
more efficiently. We do not explore this possibility in this paper, and focus only upon the question of the
evolution of efficient bargaining norms. We find that this questions by itself leads to a rich theory of implicit
agreements and their potential breakdown that may help better understand how contracts and organizations
are likely to work in practice with boundedly rational agents. It would be interesting in future work to
explore the evolution of markets themselves so we may better understand the conditions under which specific
26 MacLeod and Malcomson (1993), Cole, Mailath, and Postlewaite (2001) and Felli and Roberts (2000) have shown in a
hold-up model imperfect competiton may in some case be sufficient to implement the first best.

26
institutions, such as privately owned firms, can allow the economy to evolve better functioning markets.

References
Alchian, A. and H. Demsetz (December, 1972). Production, information costs, and economic organization,.
American Economic Review, 62 (5), 777–795.
Alchian, A. A. (June, 1950). Uncertainty, evolution and economic theory. Journal of Political Econ-
omy 58 (3), 211–21.
Binmore, K., C. Proulx, L. Samuelson, and J. Swierzbinski (1995). Hard Bargains and Lost Opportunities.
University of Wisconsin.
Binmore, K. G., A. Rubinstein, and A. Wolinsky (1986). The nash bargaining solution in economic mod-
eling. Rand Journal of Economics 17, 176–88.
Carmichael, H. L. and W. B. MacLeod (2003, Spring). Caring about sunk costs: A behavioral solution to
hold-up problems with small stakes. Journal of Law, Economics and Organization 19 (1), 106–118.
Che, Y.-K. and D. B. Hausch (March, 1999). Cooperative investments and the value of contracting.
American Economic Review 89 (1), 125–47.
Coase, R. (1937). The nature of the firm. Economica 4, 386–405.
Coase, R. A. (1960). The problem of social cost. Journal of Law and Economics 3, 1–44.
Cole, H. L., G. J. Mailath, and A. Postlewaite (2000). Efficient non-contractible investments in a finite
economy. Technical report, University of Pennsylvannia.
Cole, H. L., G. J. Mailath, and A. Postlewaite (2001). Efficient non-contractible investments in finite
economies. Advances in Theoretical Economics 1 (1), na.
Dawid, H. and W. B. MacLeod (2001). Hold-up and the evolution of bargaining conventions. European
Journal of Economic and Social Systems 15 (3), 153–169.
Demsetz, H. (May, 1967). Toward a theory of property rights. American Economic Review 57 (2), 347–59.
Ellickson, R. C. (1991). Order Without Law: How Neighbors Settle Disputes. Cambridge, MA: Harvard
University Press.
Ellingsen, T. and J. Robles (2002). Does evolution solve the hold-up problem? Games and Economic
Behavior 39 (1), 28–53.
Ellison, G. (2000). Basins of attraction, long run stability and the speed of step-by-step evolution. Review
of Economic Studies 67, 17–45.
Felli, L. and K. Roberts (2000, April). Does competition solve the hold-up problem? Working paper,
University of Pennsylvania Center for Analytic Research in Economics and Social Science (CARESS)
Working Paper: 00/04.
Grossman, S. J. and O. D. Hart (August, 1986). The costs and benefits of ownership: A theory of vertical
and lateral integration. Journal of Political Economy 94 (4), 691–719.

27
Hart, O. D. (1995). Firms, Contracts and Financial Structure. Oxford, UK: Oxford University Press.
Kandori, M., G. Maillath, and R. Rob (1993). Learning, mutation and long run equilibria. Economet-
rica 61, 27–56.
Kim, Y.-G. and J. Sobel (1995). An evolutionary approach to pre-play communication. Econometrica 63,
1181–1193.
Klein, B., R. Crawford, and A. Alchian (October, 1978). Vertical integration, appropriable rents, and the
competitive contracting process. Journal of Law and Economics 21, 297–326.
Macauley, S. (1963). Non-contractual relations in business: A preliminary study. American Sociological
Review 55, 55–69.
MacLeod, W. B. and J. M. Malcomson (September, 1993). Investments, holdup, and the form of market
contracts. American Economic Review 83 (4), 811–837.
MacNeil, I. R. (1974). The many futures of contracts. Southern California Law Review 47 (688), 691–816.
Maskin, E. and J. Tirole (1999). Unforeseen contingencies and incomplete contracts. Reveiw of Economic
Studies 66, 83–114.
Mellard, C., J.-P. Platteau, and H. Watongoka (1998). Etude des incidences socio-économiques de
l’introduction de la technique de pêche au filet maillant au lac kivu. Technical report, FUCID, Univer-
sity of Namur.
Moore, J. (1992). Implementation, contracts, and renegotiation in environments with complete informa-
tion. In J.-J. Laffont (Ed.), Advances in Economic Theory: Sixth World Congress, Volume I, pp.
182–282. Cambridge, UK: Cambridge University Press.
Platteau, J. P. (2000). Institutions, social norms, and economic development. Fundamentals of develop-
ment economics ; v. 1. Amsterdam, The Netherlands: Harwood Academic Publishers. Jean-Philippe
Platteau. ill. ; 24 cm.
Rabin, M. (March, 1998). Pyschology and economics. Journal of Economic Literature 36, 11–46.
Roland, G. (2000). Transition and economics politics, markets, and firms. Comparative institutional anal-
ysis ; 2. Cambridge, Mass.: MIT Press. [electronic resource] : Gérard Roland. ill. ; 24 cm. Electronic
reproduction. Palo Alto, Calif. : ebrary, 2002. Available via World Wide Web. Access may be limited
to ebrary affiliated libraries.
Tirole, J. (July, 1999). Incomplete contracts: Where do we stand? Econometrica 67 (4), 741–782.
Troger, T. (2002). Why sunk costs matter for bargaining outcomes: An evolutionary approach. Journal
of Economic Theory 102 (2), 375–402.
Williamson, O. E., M. L. Wachter, and J. E. Harris (Spring, 1975). Understanding the employment
relation: The analysis of idiosyncratic exchange. Bell Journal of Economics 6 (1), 250–278.
Young, H. P. (1993a). The evolution of conventions. Econometrica 61, 57–84.
Young, H. P. (1993b). An evolutionary model of bargaining. Journal of Economic Theory 59, 145–168.

28
Appendix A: Description of the Learning Dynamics
Sampling, memory and belief formation:
The memory of an individual consists of the following data:

1. m observations of types of individuals where all these observations stem from t − 1. Let p̂it ∈ P :=
{0, 1/m, 2/m, . . . , 1} denote the fraction of individuals in this sample with Ti,t−1 = H.

2. m observations of demands made by high and low types in HL matches. Since in general the sample
taken in period t consists of fewer than m HL matches some older observations may remain in the
sample. The oldest data is dropped as new observations are inserted. This sample is used to estimate
the empirical distribution functions F̂H (.) and F̂L (.). Both of these empirical distribution functions
are elements from the finite set

F = {F : X (k) → {0, 1/m, 2/m, ..., 1} |F (x) is increasing, F (SA ) = 1} .

The set of all possible beliefs of an agent is then given by B = P × F 2 .

Expected payoffs:
The expected payoff of an agent with type H or L choosing a ∈ X (k) under beliefs β ∈ B, is given recursively
by:
³ ³ ´ ´
UL (a, β) = (1 − p̂ (β)) SL /2 + p̂ (β) F̂H (SA − a, β) a + δ 1 − F̂H (SA − a, β) UL (a, β) ,
³ ³ ´ ´
UH (a, β) = p̂ (β) SH /2 + (1 − p̂ (β)) F̂L (SA − a, β) a + δ 1 − F̂L (SA − a, β) UH (a, β) .

Investment Decision
Given beliefs bit ∈ B agent i entering the population in period t chooses to invest if:
¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢
max 2
(1 − λ) UH xH , bit + λUL xL , bit − c ≥ max 2
(1 − λ) UL xL , bit + λUH xH , bit .
(xL ,xH )∈X(k) (xL ,xH )∈X(k)

Determination of demands in step 2. of the timeline

¡ ¢
Individual i chooses ait ∈ X (k) to maximize UT i ait , bit , given her type T i ∈ {L, H} and beliefs bit ∈ B.
When indifferent over demands she chooses the smallest demand. The agent’s strategy is uniquely defined
¡ ¢
by her beliefs and type. Hence, we write ait = α T i , bit .

State space of the process

The state space is given by all possible distributions of n individuals over the set C = {H, L} × B:
X
S = {s ∈ [0, 1]|C| | sc = 1, nsc ∈ IN0 ∀c ∈ C},
c∈C

Appendix B: Proofs
Proof of Proposition 1:
Efficiency implies SH − c > SA > 0, therefore if one sets x̂L = 0, then conditions 1 and 3 for a self-enforcing

29
norm are strictly satisfied for λ = 0. The trade condition implies that SA > δ (SL + SH ) /2 > δSH /2 and
therefore condition 2 is strictly satisfied. Given that the expressions in the definition of a self-enforcing norm
are continuous for small λ, these conditions are satisfied for small λ. 2

Proof of Lemma 1:
Let s and s0 be two arbitrary states in S. We show that there is a positive multi-step transition probability
from s to s0 and a positive one-step transition probability from s0 to s0 . This then implies that the process
is irreducible and aperiodic.
Assume that σ t = s. With positive probability the bargaining strategy of all agents at time t is such
that all agents carry out the project (some mutations of bargaining strategies might be needed) and leave
the population. Hence, with positive probability in period t + 1 the types of all agents in the population are
determined anew and with a positive probability the resulting distribution of types matches exactly the one
in s0 . Every period there is positive probability that the distribution of types stays like that. If there are both
high and low types in s0 it is straight-forward to see that any set of observations needed to create empirical
distribution functions which have positive weight in s0 can be created by multiple mutations of bargaining
behavior of the agents given the type distribution. In case there are only high or only low types in s0 consider
the transition where first all but one agent get the type required in s0 , then all the observations needed to
create all the beliefs in s0 are created by mutations and finally the single agent with a different type leaves
the population and changes her type. In any case there is a positive probability that s0 is reached in multiple
steps. Furthermore, since there is always a positive probability that all agents only observe matches between
the same types during a period and therefore do not change their beliefs, there is a positive probability that
the process stays in s0 once it has reached s0 . Hence, the process is irreducible and aperiodic. 2.

Proof of Lemma 2:
p̂(β) SH
By waiting for an partner of same type, an agent with beliefs β expects a payoff of 1−δ(1−p̂(β)) 2 if she is
1−p̂(β) SL
of type H and 1−δ p̂(β) 2 if she is of type L. Hence, a bargaining norm (xL ) is compatible with p̂ and δ if
xL ∈ [xL (p̂), x̄L (p̂)], where xL (p̂) ∈ X(k) such that

δ(1 − p̂) SL
xL (p̂) − α < ≤ xL (p̂)
1 − δ p̂ 2

and x̄L (p̂) ∈ X such that

δ p̂ SH
x̄L (p̂) ≤ SA − < x̄L (p̂) + α.
1 − δ(1 − p̂) 2
Considering the monotonicity of these expressions with respect to p̂ we get C(δ) = [xL (0), x̄L (1)]. Simple
2SA
calculations now establish that C(δ) 6= ∅ for sufficiently small α if and only if δ < SL +SH . This is exactly
the trade condition, hence the Lemma. 2

Proof of Lemma 3:

30
First we show that all the sets given in the Lemma are limit sets, i.e. we have to show that for ² = 0 they
are absorbing and for each pair of states in such a set there is a positive (multi-step) transition probability.
It follows from the definition of C(δ) that if x ∈ C(δ) and all individuals have point beliefs β such that
F̂L (·, β) = P(x), F̂H (·, β) = P(SA − x), all individuals have the optimal bargaining strategy xL = x, xH =
SA − x. Therefore, in the absence of mutations these point beliefs can never be altered and therefore Ω(x) is
absorbing. Furthermore, since in every period every distribution of types has a positive probability regardless
of the actual investment behavior, and so also for every p̂ ∈ P there is a positive probability that a sample
yielding such an estimator is observed, all possible distributions of types and p̂ can be reached with positive
probability. Hence, the set Ω(x) is connected, which implies that it is a limit set.
To prove that these are the only limit sets, we show that from every state which is not in one of the
limit sets described above there is a positive probability to reach one of these sets. This comes down to
showing that a homogeneous bargaining norm which is consistent with all p̂ ∈ P can always be reached with
positive probability. The transition can go as follows: assume σ t = s for some arbitrary state s ∈ S. With
positive probability there are at least m low types in σ t+1 and with positive probability at t + 2 there is
some pairing of a low type agent aL and a high type agent aH with bids x̃H , x̃L , where aL has beliefs β such
δSL
that p̂(β) = 0 and accordingly x̃L ≥ 2 . With positive probability this pairing is repeated m times from
period t + 2 till t + m − 1 and one agent, we call him bH , in the population samples all these pairings but no
other high-low pairings. Accordingly, at t + m she has beliefs such that F̂L (·, β t+m ) = P(x̃L ). Furthermore,
there is a positive probability that the beliefs of aL (or the agent who replaces her) only observes high-high
meetings during this period and her beliefs stay unchanged. Furthermore there is a positive probability
that aL and bH are matched in periods t + m till t + 2m − 1. In each such matching the two bids are
x̃L of aL and SA − x̃L of bH . Again, there is a positive probability that all individuals sample only these
high/low pairings during periods t + m to t + 2m − 1. Then in t + 2m all agents have beliefs such that
δSH
F̂L (dot, β t+2m ) = P(x̃L ), F̂H (·, β t+2m ) = P(SA − x̃L ). If SA − x̃L ≥ 2 we have x̃L ∈ C(δ) and the proof
of (a) is complete.
δSH
If SA − x̃L < 2 ,
there is a positive probability that in period t + 2m + 2 there is a high type with
˜H such that x̃
p̂ = 1. This agent then makes a bid x̃ ˜H − α < δSH ≤ x̃ ˜H and the same arguments as above
2
imply that there is a positive probability that a homogeneous state will evolve where all agents hold beliefs
˜H ), F̂H (·, β) = P(x̃
β such that F̂L (·, β) = P(SA − x̃ ˜H ). Since δ < 2SA implies δSH > SA − δSL , we have
SH +SL 2 2
˜H ∈ C(δ) for sufficiently small α.
SA − x̃

Proof of Proposition 3:
Define

BL (δ) = {xL ∈ X|∃p ∈ P s.t. xL = xL (p) or ∃p ∈ P s.t. xL = x̄L (p)}

BH (δ) = {xH ∈ X|∃p ∈ P s.t. xH = SA − x̄L (p)
or ∃p ∈ P s.t. xH = SA − xL (p)}

as the set of all demands which lie just above the outside option for some p̂ ∈ P and the best responses
to that. The larger m is the larger these sets are and for sufficiently large m we simply have BH (δ) =

31
 £ δSL δSH
¤ £ δSH δSL
¤
X ∩ SA − 2 , 2 and BL = X ∩ SA − 2 , 2 .
To prove our claim we show that for sufficiently large m, n and k the unique stochastically stable set of
the process {σ t } is a set L where for all states s ∈ L we have sζ > 0 only if ζ = (T, β) for some T ∈ {H, L}
and some β such that supp(F̂H (·, β)) ∈ BH (δ), supp(F̂L (·, β)) ∈ BL (δ).
Assume σ t = s for an arbitrary state s ∈ S. Assume further that there are at least m low types and
at least m high types in the population (if this is not true, there is a positive probability that at least m
low and high types will be in the population within two periods). Then, there is a positive probability
that in period t + 1 all low types have beliefs p̂i = 0 and at least m are matched with high types. The
resulting demands at t + 1 of these low types are larger or equal to xL (0). There is a positive probability
that at least m high types observe these m demands in t + 2 and that the same m high types in period
t + 3 have beliefs such that p̂(β i ) = 1 and are matched with low types. Since for these individuals we have
F̂L (·, β) = P(xL (0)) and xL (0) > x̄L (1) the outside option is binding for all these high types and they
demand xH = SA − x̄L (1) in period t + 3. With positive probability these m demands are sampled by
all agents in t + 4 and hence all agents have beliefs such that F̂H (·, β) = P(SA − x̄L (1)). With positive
probability these beliefs stay unchanged till t + 5 whereas the belief about the type distribution changes to
p̂(β) = 0. With positive probability in t + 5 now at least m low type agents are matched with high types
and since xL (0) > x̄L (1) their outside option is binding and their demands are xL = xL (0). With positive
probability all agents sample the demands of these m low types in t + 6 and hence all agents have beliefs β
such that (p̂(β) = 0, F̂H (·, β) = P(SA − x̄L (1)), F̂L (·, β) = P(xL (0))). We denote this state by s̃. The fact
that there exists a positive multi-step transition probability from every state to s̃ implies that the Markov
chain has a single limit set which includes s̃. Obviously, this single limit set consists of all states which can
be reached with positive probability from s̃. Taking into account that every demand of a high type where
the outside option is binding has to be in BH and that the best response of a high type with some beliefs
F̂L with support in BL and p̂ ∈ P must lie in BH as well, shows that all demands of high types have to be
in BH once s̃ has been reached. Similarly for a low type. Accordingly, given that ² = 0, any observation
outside BH × BL has probability zero once s̃ has been reached before. This shows that L is the only limit
set in the state space which implies that this set has to be stochastically stable. 2

Proof of Proposition 4:
We have to determine which of the limit sets characterized in Lemma 3 are stochastically stable. We use the
radius modified coradius criterion introduced in Ellison (2000). For a union of limit sets Ω the radius R(Ω)
is defined as the minimum number of mutations needed to get to a state outside the basin of attraction of
Ω with positive probability. The modified coradius CR∗ (Ω) is defined as follows: consider an arbitrary state
x 6∈ Ω and a path (z1 , z2 , . . . , zT ) from x to Ω where L1 , L2 , . . . , Lr ⊂ Ω is the sequence of limit sets the path
goes through (this implies Lr ⊆ Ω). We define the modified costs of this path by
r−1
X
c∗ (z1 , . . . , cT ) = c(z1 , . . . , zT ) − R(Li ),
i=2

where c(z1 , . . . , zT ) gives the number of mutations needed on the path (x1 , . . . , zT ). Denoting by c∗ (x, Ω)

32
the minimal modified costs for all paths from x to Ω we define the modified coradius as

CR∗ (Ω) = max c∗ (x, Ω).

x6∈Ω

Ellison (2000) proves that every union of limit sets Ω with R(Ω) < CR∗ (Ω) contains all stochastically stable
states.
In what follows we calculate the radius and modified coradius of the bargaining norms described in
Lemma 3. In the case of substitutes the limit sets are of the form Ω(xL ) for xL ∈ C(δ). Let x̃L be an
arbitrary bargaining norm with x̃L ∈ C(δ). To destabilize the norm upwards either a sufficient number of
high types have to mutate to a xH smaller than SA − x̃L , in the extreme case xH = 0, such that the best
response of a high type who has sampled all these mutants becomes xL = SA , or a sufficient number of low
types have to mutate to xL = x̃L + α such that the best response of a high type who has sampled all these
SA
mutants becomes xH = SA − x̃L − α, where α = k . As has been demonstrated in Young (1993b), for
sufficiently small α the second of these two possibilities yields transitions with a lower number of mutations
(the number goes to zero as α goes to zero). Similar arguments hold for a downwards destabilization and
therefore in order to leave a norm x̃L with the minimal necessary number of mutations either the path to
x̃L + α or the path to x̃L − α has to be taken. We define by c+ (xL ) the minimal number of mutations needed
to get to x̃L + α and by c− (xL ) the minimal number of mutations needed to get to x̃L − α. We first calculate
c+ (x̃L ).
The number of mutations needed to destabilize a norm also depends on the beliefs p̂. We first show that
the minimal number of mutants either occurs at p̂ = 0 or at p̂ = 1. Consider a low type whose beliefs F̂H
attach probability q to xH = SA − x̃L + α and 1 − q to xH = SA − x̃L . Denote by v the expected discounted
payoff of this individual given that he faces a high type and bids xL = x̃L whenever facing a high type.
Taking into account that he will always trade immediately when he meets another low type we get
µ ¶
SL
v = (1 − q)x̃L + δq p̂v + (1 − p̂)
2

and
(1 − q)x̃L + δq(1 − p̂)SL /2
v(q; p̂) := .
1 − δq p̂
Note that this expression is monotonic in p̂ for p̂ ∈ [0, 1] (increasing or decreasing). The minimal number of
mutations needed to destabilize the norm is given by dmq̃e, where q̃ is the minimal q such that:

v(q; p̂) < x̃L − α

holds for some p̂ ∈ [0, 1]. Since the right hand side is constant in q and p̂ and the left hand side is monotonous
in p̂ for all q the minimal q is either attained at p̂ = 0 or at p̂ = 1.
With p̂ = 0 we get
SL
v(q; 0) = (1 − q)x̃L + δq ,
2
which gives
α
q > q1− (x̃L ) := .
x̃L − δ S2L

33
 For p̂ = 1 we have
1−q
v(q, 1) = x̃L .
1 − δq
Accordingly, the norm can be destabilized downwards if
α
q < q2− (x̃L ) := .
x̃L (1 − δ) + δα

Comparing the two we see that q1− (x̃L ) < q2− (x̃L ) if and only if x̃L > S2L + α. All-together we have

 q1− (x̃L ), if x̃L ≥ S2L + α,
c− (x̃L ) =
 q (x̃ ), if x̃ < SL + α.
2− L L 2

Similar reasoning for destabilizations upwards shows that for a high type, who is matched with a low
type and who believes that a fraction q of low types demands xL = x̃L + α and a fraction 1 − q of low types
demands xL = x̃L , has the following expected payoff from demanding xH = SA − x̃L :
1−q
w(q; 0) = 1−δq (SA − x̃L )

w(q, 1) = (1 − q)(SA − x̃L ) + δq S2H .

This implies 
 q1+ (x̃L ), if x̃L ≥ SA − SH
2 − α,
c+ (x̃L ) =
 q (x̃ ) if x̃L < SA − SH
− α.
2+ L 2

where
α
q1+ =
(SA − x̃L )(1 − δ) + δα
α
q2+ = .
SA − x̃L − δ S2H

The function c− is decreasing in x̃L whereas c+ is increasing in this variable which implies that they
have a unique intersection. We denote this intersection point by x̂L . Clearly at this point min[c− , c+ ] is
maximized. For
2(SH − SA )
(6) δ≤
SH − SL
x̂L lies on the intersection of q1− and q1+ and is given by
SA δ
(7) x̂L = − (SA − SL − 2α).
2 2(2 − δ)

To establish (a) we first observe that under the assumptions made in (a) the condition (6) holds and
£ ¤
ˆL ∈ C(δ) that maximizes min[c+ , c− ] over C(δ)
x̂L ∈ δS2L , SA − δS2H for small α. Hence, there exists a x̂
and whose distance from x̂L is smaller than α. Taking into account Lemma 3 this in particular implies that
ˆL .
there is a limit set corresponding to the bargaining norm x̂
From the arguments above it follows that for every xL ∈ C(δ) with xL < x̂ ˆL we have for the radius of
ˆL we have R(Ω(xL )) =
the limit set Ω(xL ): R(Ω(xL )) = dmc+ (xL )e and for every xL ∈ C(δ) with xL > x̂

34
 ˆL ) along a graph g which connects every limit
dmc− (xL )e. From every limit set Ω(xL ) there is a path to Ω(x̂
ˆ
set Ω(xL ) where xL < x̂L with Ω(xL + α), and every limit set Ω(xL ) where xL > x̂ ˆL with Ω(xL − α). This
implies that
ˆL )) ≤
CR∗ (Ω(x̂ max R(Ω(xL )).
ˆL }
xL ∈C(δ)\{x̂

For sufficiently large m we have R(Ω(x̂ ˆL )) > R(Ω(xL )) for all xL ∈ C(δ) \ {x̂
ˆL } and therefore R(Ω(x̂
ˆL )) >
CR∗ (Ω(x̂ˆL )). Using the radius-modified coradius criterion we can conclude that the limit set corresponding
ˆL is stochastically stable. For k → ∞ we have x̂
to x̂ ˆL → x̂L and get (a). Exactly the same arguments
establish (b), where it has to be taken into account that in this case x̂L lies at the intersection of q1− and
q2+ which is given by
SA δ
x̂L = − (SH − SL )
2 4
2

Proof of Lemma 4:
Parts (a) and (b) are trivial. To prove part (c) we denote by Q(λ) = [qij (λ)]i, j ∈ S̃ the one-step transition
matrix of the Markov process {σ̃ t }. It can then easily be established that limλ→0 qi0 + qi1 > 0 for all i ∈ S̃.
Furthermore, at state i = 0 no individual can sample any high types and hence we have p̂(β) = 0 for all
individuals and accordingly all choose high investment. Therefore limλ→0 q01 = 1 and by the same reasoning
limλ→0 q10 = 1. Therefore, the only limit set for λ → 0 is {0, 1} which implies that limλ→0 π̃ ∗i (λ) = 0 for all
i ∈ S̃ \ {0, 1}. Using this we get from the Chapman-Kolmogoroff equation at state 0
 
X X
π̃ ∗0 (λ)  q0,i  = qi,0 π̃ ∗i (λ)
i∈S̃\{0} i∈S̃\{0}

that limλ→0 π̃ ∗0 = limλ→0 π̃ ∗1 = 0.5. 2

Proof of Proposition 6:
The proof of (b) is identical to the proof of part (a) of proposition 4. To proof (a) we again follow the proof
of proposition 4 but observe that for SA < 2δ ((2 − δ)SH + δSL ) we have x̂L > SA − δ S2H . Therefore the point
ˆL where x̂
which maximizes min[c+ , c− ] over C(δ) is given by x̂ ˆL ≤ SA − δ SH < x̂ˆL + α. Stochastic stability of
2
ˆL ) is established analogous to the proof of proposition 4 but here we have x̂
the limit set Ω(x̂ ˆL → SA − δ SH
2
for k → ∞. 2

Proof of Lemma 5:
We show the proposition for p∗ (xL ) > 0.5, the other case analogous.
We denote again the one-step transition matrix of the process{σ̃ t } by Q = [qij (λ)]i, j ∈ S̃ We can write
these transition probabilities as µ ¶
n j
qij = β (1 − β i )n−j ,
j i
where
β i = (1 − λ)s(mp∗ (xL ); i) + λ(1 − s(mp∗ (xL ); i))

35
is the probability that a randomly chosen individual is of high type. Note that for a given bargaining norm
the investment decision only depends on the number of high types sampled by an individual in the current
period. We denote by
X µm¶ µ i ¶k µ i
¶m−k
∗
s(mp (xL ); i) = 1−
∗
k n n
k≥mp

the probability that an individual samples more than mp∗ (xL ) high types in a population with i high types.
Since we are dealing with the case of investment complements here, this is the probability of high investment.
Note first that
X µm¶ µ i ¶k µ i
¶m−k
∗
1 − s(mp ; i) = 1−
k n n
k<mp∗
X µ ¶ µ ¶m−k µ ¶k
m i i
= 1−
∗
k n n
k>m(1−p )

X µm¶ µ n−i
¶m−k µ
n−i
¶k
> 1−
∗
k n n
k>mp
= s(mp∗ ; n − i),

where the inequality follows from p∗ > 0.5. Using this we get that for λ < 0.5

1 − β n−i = 1 − (1 − λ)s(mp∗ ; n − i) − λ(1 − s(mp∗ ; n − i))

= 1 − λ − s(mp∗ ; n − i)(1 − 2λ)
> 1 − λ − (1 − s(mp∗ ; i))(1 − 2λ)
= (1 − λ)s(mp∗ ; i) + λ(1 − s(mp∗ ; i))
= βi.

This means that in a population with i high types the probability that an individual becomes a high type
is smaller than the probability that an individual becomes a low type in a population with i low types.
i
In particular, this implies that the probability that at least z individuals become high types in state n is
smaller than the probability that at least z individuals become low types in state n−i for all z. We denote by
© 1 n−2
ª © n+2 n−1
ª ©nª © n ªn
L = 0, n , . . . , 2n , H = 2n , . . . , n , 1 , L̃ = L ∪ 2 and by H̃ = H ∪ 2 . Furthermore we denote
by qiL the transition probability from state i into the set L and analogous the transition probabilities into
the other sets defined above. The arguments above imply that

qiL > qn−iH and qiL̃ > qn−iH̃ ∀i.

Note that both (L, H̃) and (L̃, H) are partitions of the state space, therefore under the stationary distribution
the flows between L and H̃ must be identical in both directions and so have to be the flows between L̃ and
H. This gives
n/2−1 n/2
X X
(8) π̃ ∗i qiH̃ = π̃ ∗n−i qn−iL
i=0 i=0
n/2 n/2−1
X X
(9) π̃ ∗i qiH = π̃ ∗n−i qn−iL̃
i=0 i=0

36
 Pn/2−1 Pn/2−1
Our claim can now be easily shown by contradiction. If i=0 π̃ ∗i ≤ i=0 π̃ ∗n−i we can use qiH < qn−iL
for all i to derive that
n/2 n/2
X X
π̃ ∗i qiH < π̃ ∗n−i qn−iL
i=0 i=0

Since (8, 9) have to hold this implies

n/2−1 n/2−1
X X
π̃ ∗n−i qn−iL̃ < π̃ ∗i qiH̃
i=0 i=0

But we also have qiH̃ < qn−iL̃ for all i and therefore this inequality contradicts our assumption that
Pn/2−1 ∗ Pn/2−1 ∗ Pn/2−1 Pn/2−1
i=0 π̃ i ≤ i=0 π̃ n−i . Accordingly, we must have i=0 π̃ ∗i > i=0 π̃ ∗n−i
To show that limλ→0 π̃ ∗0 = 1 we can again apply the radius-modified coradius criterion. For λ = 0 there
are two limit sets, namely {0} and {1}. In order to invest high an individual has to sample at least dmp∗ e
high types. Therefore the radius of {0} is given by R({0}) = dmp∗ e. On the other hand, the state where
maximal the number of mutations is needed to have a positive transition probability into {0} is the state
1 and therefore we have CR∗ ({0}) = dm − mp∗ e. For p∗ > 0.5 this implies that R({0}) > CR∗ ({0}) for
sufficiently large m and therefore limλ→0 π̃ ∗0 = 1. 2

37
 Boundaries of Investment Regions
Discount Factor (*) is Zero

(S - S )/2
H L

Low Investment Investment is Low at Stochastically Stable Equilibrium.

but High Investment is an Equilibrim
Under Nash Bargaining

C2
Cost of Investment

C4
(S - S )/4
H L
Investment Cycles between
High and Low

High Investment
C1

S Investments are S
L Investments are (S + S )/2 H
H L Substitutes
Complements
Value of HL Matches (SA)
Figure 1
 Boundaries of Investment Regions
Discount Factor (*) is Zero

Investment Cycles between

(S - S )/2 High and Low
H L

Outcome under One Party Ownership

Low Investment is More Efficient
C4 is Stable Equilibrium that Stable Equilibrium
Cost of Investment

(S - S )/4
H L

High Investment
is Stable Equilibrium

S Investments are S
L Investments are (S + S )/2 H
H L Substitutes
Complements
Value of HL Matches (SA)
Figure 2
 Boundaries of Investment Regions
Discount Factor (*) is 1/2

Investment is Low at Stochastically Stable Equilibrium (SSE)

(S - S )/2 but High Investment is an Equilibrim
H L Under Nash Bargaining

Low Investment Cycles between

Investment High and Low at Stabel Equilibrium
Cost of Investment

C4
Over Investment
(S - S )/4 at Stable Equlibrium
H L

High Investment

S Investments are (S + S )/2 Investments are S

L H L H
Complements Substitutes
Value of HL Matches (SA)
Figure 3