Strategic Uncertainty as a Cause of War∗

Adam Meirowitz and Anne E. Sartori

Princeton University and Northwestern University

July 30, 2008

Abstract

This paper shows why states, acting in their own self-interest, may create infor-
mational asymmetries that lead to war. In our models, two actors with no private
information invest in military capacity before engaging in crisis bargaining. If bar-
gaining fails, the states go to war, and the payoffs of a war depend on the two states’
military capacities. We examine a large class of models and show that states have in-
centives to keep each other guessing about their exact levels of military capacity — even
though doing so creates the risk of war. Thus, self interest and strategy are to blame for
the emergence of uncertainty about military strength and war. Our paper explains two
stylized facts: States devote considerable resources to secrecy in the national-security
realm, and often disagree about the balance of capabilities.

∗
An earlier version was presented at the American Political Science Association Annual Meeting, Wash-
ington, DC, September 2005. We thank Ken Schultz, Kris Ramsay, Joanne Gowa, Mark Fey, Scott Ashworth,
Ahmer Tarar, Will Moore, participants in the William Riker seminar at the University of Rochester and the
University of Chicago PIPES seminar, the editors, and anonymous referees for useful comments.
1. Introduction

Many rationalist explanations of war hinge on private information or uncertainty. While the
standard argument stems from a simplified description of the international environment, its
logic is quite compelling. By most accounts, war is costly both in terms of weapons and lives;
it uses up resources and is, thus, inefficient. As a consequence, we may think of international
disputes as concerning the division of a pie that shrinks in the event of a war. At least in
principle, there is a peaceful settlement that all states prefer to a war — a settlement in which
each state gets at least the share of the pie it would have gotten in war, and at least one gets
more. If each state knew others’ values for war precisely, the states would be able to reach
such a peaceful settlement. Thus, a hypothetical world in which states possess all relevant
information is completely peaceful.
Unfortunately, this peaceful description is inaccurate; states hide information about their
militaries, not only about specific programs but also about their overall military budgets
and the strength of their armed forces. For example, the United States Department of State
publishes a book containing estimates of the foreign military expenditures of most foreign
countries, and it does so with this qualification:

A primary aim [of the document being quoted] is to inform the reader of the
main qualifications to the data, much of which is not as accurate as uniform pre-
sentation in statistical tables may imply. This is particularly true of the data on
military expenditures, armed forces, and arms transfers, which in many countries
are subject to severe limitations of incompleteness, ambiguity, or total absence due
to governmental secrecy (emphasis added).1

If the United States, with its large budget and extensive intelligence network, has trouble ob-
taining reliable information on other states’ military expenditures and armed forces, imagine
the trouble faced by other countries.2
1
“Statistical Notes,” downloaded from Department of State web site:
http://www.state.gov/t/vci/rls/rpt/wmeat/1999_2000/.
2
In some cases the State Department also may have better estimates than it acknowledges, or other
branches may have better estimates than they share with the State Department.

1
 Fearon’s (1995) well-known argument shows that extant private information (like that
observed by the State Department) can have pernicious consequences. When private informa-
tion is present, states have incentives to misrepresent their values for war. For this reason,
they may fail to reach a bargain and go to war, even though doing so is inefficient (war
reduces the size of the pie).3
But why are states uncertain about important attributes of their potential opponents
in the first place? Fearon’s argument is not wholly satisfying because it does not explain
the emergence of private information. Instead, it takes as a premise that states begin
their interaction with private information and explains why asymmetries of information
are maintained throughout negotiations. The assumption that states begin with private
information is unproblematic insofar as this asymmetric information is truly exogenous.
However, insofar as states are uncertain about attributes that other states control (e.g.,
military capacity), we must understand why they make choices that lead to uncertainty in
order to understand the origins of military conflict more completely.
The present paper helps to explain the emergence and persistence of asymmetric infor-
mation about military capacity.4 More specifically, we show why states (i) arm, (ii) do so in
a manner that is not predictable and (iii) keep their capacity secret, even though doing so
involves a risk of war. To be clear, Fearon’s and other work speak to point (iii) — keeping the
capacity a secret, given that (i) and (ii) are descriptively accurate. In this paper, we take a
step back and also investigate why states choose to arm and why their armament decisions
remain unpredictable, as in the above excerpt, when it is known that this unpredictabil-
ity will create the possibility of inefficient war. Our explanation for war has an important
normative implication: self-interest and strategy can lead parties to undertake actions that
make war possible even when they start in a world without asymmetric information, which
3
Private information plays an important role in many theories of war. For example, Powell (1999) [97]
shows that states never go to war with complete information; they may, however, go to war when one has
private information. Blainey (1973) [246] writes about mutual optimism as a cause of war. For an argument
against mutual optimism as an explanation for war, see Fey and Ramsay (2005).
4
The literature on international conflict identifies several factors that influence a state’s resolve (Morrow
1989a; Fearon 1992). We follow Morrow (1989a) in focussing on the first of these factors; Morrow explains
the effects of such uncertainty, while we focus on its origins.

2
one might think would be a world of peace.
We find that states create and keep secrets about their overall level of military capacity
and risk war unless there is no net benefit to either state of arming secretly and unilaterally
and starting a war, relative to the settlement that reflects the status-quo levels of military
capacity.
Why do states create uncertainty about military capacity, even though doing so creates
a risk of war? States have two competing incentives when it comes to acquiring military
capacity. First, insofar as states’ military capacities are known, a state with greater military
capacity is likely to do better in negotiations; if a war occurs the capacity also will be
beneficial (Banks 1990). In other words, each state benefits from the acquisition of military
capacity. Military capacity is costly, however, so states do not wish to acquire too much;
this incentive is especially strong if negotiation is likely to lead to a peaceful settlement. In
fact, when peaceful settlement is anticipated, there is a very clear incentive to minimize the
actual investment in capacity but pretend that one has made a large investment (though
this deception does not happen in equilibrium).
In a strategic setting, these individual preferences translate into a conundrum. If each
state knows that an adversary is unarmed, then they will reach a settlement. But if both are
unarmed, then it is likely that either can gain from unilaterally increasing its capacity and
engaging in a surprise attack — a unilateral increase in strength makes victory much more
likely and this offsets the cost of acquiring additional capacity as well as any costs of fighting.
If two states arm and their levels of strengths are known, they again can reach a bargain
that is mutually preferred to war; however, as long as the states expect to reach a bargain,
either benefits from secretly investing in a lower level of capacity. As a consequence of these
competing incentives, we show that states’ decisions about how much military capacity to
accumulate in our model are unpredictable (they are in mixed strategies), except when the
incentive to arm unilaterally is absent, a condition we discuss later. Following Palfrey and
Rosenthal (1985), we call this uncertainty "strategic."5
5
Readers may recall that there are two forms of strategic uncertainty. Players may not observe the
choices of other players, i.e. the game involves imperfect information, and players may not know the costs,
strengths or preferences of other players, i.e. the game involves incomplete information. Our paper purposely

3
 In practice, states create "strategic uncertainty" (make their adversaries uncertain about
their exact level of capabilities) by arming somewhat unpredictably and by keeping military
secrets. Note that keeping purchases or expenditures secret is not in itself enough to
create uncertainty about a state’s military strength — some degree of unpredictability also is
necessary. For example, if a state’s equilibrium strategy were to buy five tanks with certainty,
even adversaries that could not observe the purchase would not be uncertain about the state’s
number of tanks. Thus, when we refer to states "keeping secrets" or "hiding information"
in this paper, we mean this expanded version of secrecy, accumulating military strength in
such a way as to make it possible for their exact levels of strength to be secret, and then not
revealing this exact level of strength.6
Of course, we do not mean to say that states act completely unpredictably in the collo-
quial sense of that phrase, that next year’s military budget in the United States could be
anything. Military budgets often, though not always, change incrementally, and much infor-
mation about military acquisitions is public in democracies. Rather, we mean to say that
Secretaries of Defense, Presidents, and their counterparts elsewhere make decisions about
building their military forces in such a way that adversaries do not know the country’s over-
all level of strength precisely, and then they keep secrets to preserve this uncertainty. For
example, countries often divide their military expenditures among several departments, in
part to make it harder to figure out the budget precisely. In the U.S., these include the
Department of Energy and in Russia the atomic-energy ministry. Leaders also keep par-
ticular programs secret. While there are multiple reasons for states’ secrecy about military
matters, one consequence is that adversaries have difficulty estimating their overall strength
with precision.
By explaining the strategic origins of states’ uncertainty about their adversaries’ overall
military strength, our paper provides a foundation for the literature on war that assumes
the existence of private information. It also explains the corresponding fact that states often
considers a game in which the second form of strategic uncertianty is not present and focuses on the first
type of strategic uncertainty.
6
States in our models keep secrets by not revealing their acquisitions through bargaining; while we do
analyze cheap talk in this paper, it is easy to see that they still keep secrets if given the opportunity to talk.

4
lack information about each other’s military forces. In equilibrium, two states with the
same information cannot disagree about the balance of military forces. Thus, our work also
provides an equilibrium explanation for why states often disagree about the military balance,
as was the case at the start of the Russo-Japanese War (Fearon 1995, 398-400) Finally, our
paper speaks to the literature on international institutions, which argues that institutions
can reduce the risk of war by providing members with information, and thus reducing the
uncertainty (e.g., Keohane (1984) [93-95]). This argument again presumes that states have
private information in the first place. Moreover, our analysis explains why it is difficult to
build institutions that reduce the scope of states’ uncertainty about security (see Keohane
(1984) [94, 247]). If institutions were to do away with informational asymmetries, states still
would have the incentives we identify to increase their capacity and to do so privately (that
is, to create uncertainty).
Our results about the difficulties of avoiding uncertainty and the ensuing risk of war
are quite general; we show that they hold for many bargaining protocols and under different
assumptions about states’ ability to observe or learn about each other’s military investments.
States in each of our models begin with complete information. We begin by examining
models in which states do not observe their adversaries’ investments in military capacity
directly. We show that a version of our results holds for any bargaining protocol satisfying
two very reasonable conditions: (1) under the protocol, if the parties were to know each
other’s capacities, they would reach a bargain short of war in equilibrium, and (2) under
the protocol, it is possible for either party to initiate a war without the consent of the other
state (although this need not occur in equilibrium). That is, we show that despite the fact
that states would reach a settlement if they maintained the complete information they have
at the start of the game, they choose to create uncertainty and to generate a risk of war.
We next show that our substantive conclusions hold as long as purchases of military
capacity are even slightly difficult to observe directly (rather than completely unobserved).
If states can demonstrate their military capacities precisely (and cannot cheat by secretly
acquiring more), then equilibria without war do exist under a wider range of circumstances; in
some of these circumstances, states create uncertainty, and in others, they do not. However,

5
if states’ observations of or other information about their adversaries’ military acquisitions
are even a little noisy, then in equilibrium they create uncertainty and generate a risk of
war under the same conditions as in our original model, subject to a weak assumption about
when states can unilaterally initiate conflict.
These later results again speak to the effects of international institutions. If institutions
provide information that is even slightly imperfect and states are sufficiently patient, then
states create uncertainty and risk war despite the presence of the institution.
Our model differs from most previous formal models of crisis bargaining in that it begins
with complete information (any uncertainty is endogenous) and has a rich bargaining space
with no commitment problems, and yet in equilibrium states sometimes go to war. Most
previous models of bargaining in international crises have been either complete-information
models in which states never go to war (e.g., Fearon (1995), Kydd (2000), Powell (1999)), or
incomplete-information models in which states begin the game possessing private information
and go to war in equilibrium (e.g., Fearon (1995), Morrow (1989a), Powell (1999), Slantchev
(2005), Schultz (1998)). Powell (1993) and Morrow (1989b) contain complete-information
models in which states go to war, in these models states do not have the opportunity to reach
a bargain short of war.7 In the sense that it predicts that wars occur with rich bargaining
protocols and no ex ante asymmetric information, our paper is similar to Slantchev (2003).
However, our explanation for war is quite different. Slantchev shows that in the presence of
multiple, stage-game equilibria and an infinite horizon, war can be supported in the short
run by threats of playing inefficient equilibria in subsequent periods.
Our paper is most closely connected to contemporaneous work by Jackson and Morelli
(2007), who analyze a dynamic game in which states select levels of investment and decide
whether or not to fight in each period. In contrast to the current paper, Jackson and
Morelli assume that investment decisions are perfectly observed. For some assumptions
about the costs of war, they find equilibria in which states randomize in their investments
7
Some authors have created incomplete-information models in which states do not go to war. For example,
in Kydd (2000)[240], an arms race resolves the underlying uncertainty, so the states do not go to war in
equilibrium. Levontoglu and Tarar (2005) show that under some circumstances private information results
in delay in reaching a settlement rather than in war.

6
and eventually go to war. In these equilibria, the forces that lead to randomization in the
investment stage are similar to the forces that lead to randomization in the current paper
and are closely connected to contests in general. Their finding that war can occur after
perfectly observed investments, however, depends on aspects of the dynamic structure in
their game and differs from our findings in the section with observed investments. Jackson
and Morelli do not study a situation in which investment decisions are imperfectly observed.
Our paper also is related to Baliga and Sjöstrom (2007). In their model, private information
is more exogenous than in ours; weak states that try to acquire weapons of mass destruction
may (WMD) or may not succeed in doing so, leading to uncertainty over whether or not
they have WMD. However, when states decide to preserve this uncertainty, they do so to
balance the costs and benefits of arming, analogous to states in our model.
The paper develops as follows. In section 2, we present the basic model of arms accumu-
lation in the setting of a stylized assumption about the bargaining protocol. We characterize
necessary and sufficient conditions for equilibria in which states always avoid war in this
model. In section 3, we generalize the results to a very large class of bargaining protocols. In
section 4 we treat the issue of equilibrium existence, and illustrate that the logic of section 3
applies to a large class of models for which equilibria are known to exist (possibly in mixed
strategies). In Section 5, we relax the assumption that states’ military purchases are not
directly observed. In Section 6, we conclude.

2. The Model

Before turning to the models and analyses, we make a few clarifications about the modeling
strategy and terminology. States in each of our models begin with complete information. In
this section and in section 3, we examine models in which states do not observe their adver-
saries’ investments in military capacity directly.8 (In later sections of the paper, we relax this
assumption about unobserved investments, to rather surprising results.) In allowing states
8
Slantchev (2005) contains a model with endogenous military capacity and exogenous incomplete infor-
mation. Kydd (2000) also contains a model in which states buy capacity, but the purchases are directly
observed.

7
to take hidden actions to acquire capacity, however, we are not simply assuming that states
have private information. In an equilibrium in which the states’ capacity investments are in
pure strategies, each state can figure out the capacity of its opponent and the ensuing equi-
librium bargaining corresponds to equilibrium behavior in a bargaining model with complete
information. This feature is standard in Nash equilibria. Equilibria of our game typically
involve mixed strategies at the capacity-accumulation stage — that is, each state has a proba-
bility distribution over the levels of capacity that it buys. As long as this is the case, after the
states acquire military capacity, each state has beliefs about the other’s chosen capacity, but
knows its own level precisely. Thus, states face uncertainty about their opponents’ levels of
military capacity. In other words, states begin the interaction with symmetric information,
but we show that in most settings they generate asymmetric information, or uncertainty, in
every equilibrium of the game.
This paper does not attempt to provide a complete explanation for how states make
decisions about arming, negotiating, and fighting. Instead we provide very general answers
to questions about the circumstances under which states’ behavior leads to uncertainty about
their military capabilities, and those under which they are able to avoid war with certainty.
A cost of our focus on generality is that we do not answer a natural, follow-up question:
Precisely what happens in equilibrium?
In the model, two states begin by simultaneously investing in military capacity. Each
state’s investment is unobserved by the other player, and each pays a per unit cost for
capacity. After investing in capacity, the states bargain; if bargaining fails, they go to war.
For simplicity’s sake, we assume that the states bargain over a good of size one that does
not begin as the property of either state. Either the states agree on a division of the pie
that totals one, or they go to war. Our analyses would not be affected by changing the size
of the good or by assuming that one or the other state began by owning the good.
Formally, we consider two states, 1 and 2. It often is convenient to refer to them as i
and j. In period 1, each state simultaneously chooses an investment mi ≥ 0 in military
capacity.9 Only state i observes the choice of mi . Investment by i has a marginal cost
9
We use the terms capacity and arms interchangeably in this paper, but capacity in the model can include

8
β i > 0. In period 2, the states bargain. (In later extensions of the game, bargaining may last
many periods; in this case, they begin bargaining in period 2.) Either the states agree to
a settlement ai , aj such that ai + aj = 1 and ai , aj ≥ 0, or they engage in military conflict.
In the event of a conflict, the payoff to state 1 is given by the function w1 (m1 , m2 ) and
the payoff to state 2 is w2 (m2 , m1 ). By m we denote a vector of investment levels. When it
creates no ambiguity we write wi (m).
Two assumptions are imposed on the payoffs to war. First, military investment by country
i can only help country i in the event of war, and it can only hurt the adversary, j. Second,
war is inefficient relative to a peaceful bargain. That is, following Fearon (1995) and others,
we assume that war is costly; people die and equipment is destroyed or used up, and territory
may be devastated. This means that the total pie available to be divided is decreased if the
states go to war. Formally, these assumptions are:
Assumption 1: For i ∈ {1, 2} the function wi (m) : R2+ → [0, 1] is non-decreasing in mi
and non-increasing in mj with j 6= i.
Assumption 2: w1 (m) + w2 (m) < 1 for all m.10
To begin, we focus on a very simple bargaining process, or protocol: State 1 makes an
offer a1 ∈ [0, 1] and state 2 either accepts it (getting 1 − a1 ) or rejects it. Following a
rejection war occurs. This is the protocol analyzed in Fearon (1995). Later in the paper,
we show that the main results hold for a large class of bargaining protocols. In this simple
version of the model, we sometimes refer to the first state as the proposer and the second as
the veto player. We focus on perfect Bayesian equilibria (PBE), which require that at each
information set play is sequentially rational given beliefs and beliefs are updated according
to Bayes’ Rule whenever possible.
For simplicity, we consider two states that start the game possessing capacity of ai = 0.
An exercise in relabeling allows the game to capture a situation in which the states begin the
any factors that make a state more likely to win a war but are costly to accumulate — for example, a new
technology or military strategy.
10
Most formal studies treat war as a costly lottery, with each state’s probability of winning and thus its
expected utility reflecting the distribution of power (Powell 2004; Wagner 2000). This approach satisfies
assumptions 1 and 2, so our results apply to models with this payoff structure.

9
game with known levels of capacity other than 0. In this case, as we discuss later, the choices
ai can be interpreted as net enhancements in capacity, and 0 as each state’s status-quo level
of capacity.
Of course, in practice there may be reasons for investing in military capacity that are
unrelated to relations with the adversary — for example, bureaucratic politics. Insofar as
there is randomness (or symmetric information) in these other investments, then there are
other sources of uncertainty than those that we identify.
It also may be reasonable to think that investments in military capacity that affect rela-
tions with the adversary in our model also have spillover effects. For instance, investments
might support a segment of the economy, satisfy constituencies, or provide for security in
ways unrelated to the dyad that the game focuses on. One way to reinterpret the model
in this context is to recognize that these additional factors decrease the marginal cost of
investment (relative to the other things a state can do with resources). Since the marginal
cost is a parameter, β i , such an interpretation is entirely consistent with our analyses and it
is easy to see how changes in these additional benefits might change the results.

2.1. Results

We first consider the types of pure-strategy equilibria that are possible — that is, under what
circumstances are there equilibria in which states’ actions do not lead to uncertainty? As
we discussed earlier, states in our model begin with complete information about all features
of the game. Since equilibrium strategies are common knowledge, both states will continue
to know all of the relevant information as long as they acquire military capacity in pure
strategies.
Since our model is one of ex ante complete information, one might expect a number of
equilibria in which states arm but reach bargains short of war. We begin by showing that
no such equilibrium exists. To do so, we begin by showing that states do not arm in any
pure-strategy equilibrium of this game. The reason for this is that states will reach a bargain
short of war in any pure-strategy equilibrium, since, as long as states accumulate military
capacity in pure strategies, the capacity levels are known (though not observed) at the time

10
of bargaining. However, if states are avoiding war with certainty, each prefers to secretly
remain unarmed, since acquiring arms is costly. Thus, if states are reaching a bargain and
mi > 0, state i has an incentive to deviate to mi = 0.

Proposition 1. There are no equilibria with pure-strategy accumulations and max{m1 , m2 } >
0.

Proof: Suppose that there is a pure strategy equilibrium with mi > 0 for
some i. In the bargaining stage of the game, sequential rationality requires that
2 accept any offer that satisfies 1 − a1 ≥ w2 (m). This conclusion relies on the
fact that in games of this form it is not possible to support equilibria in which
the second mover rejects offers when she is indifferent. An acceptance strategy
of this form results in an open-set-problem for the proposer and thus equilibrium
requires a selection of the correct rule for how players act when indifferent; we
do not belabor this point in the remainder of the paper. Thus, the maximum
offer a1 that will be accepted is 1 − w2 (m). An offer by 1 above this amount
will be rejected by 2, which leads to the war payoff w1 (m). By assumption (2),
1 − w2 (m) > w1 (m), so sequential rationality requires that 1 offer 1 − w2 (m)
which is accepted. Therefore, war does not occur in this equilibrium and thus an
unobserved deviation to mi = 0 will not change the outcome in the bargaining
stage, but it will increase i’s payoff by β i mi . Thus, this is a profitable deviation.¥

Thus, in any equilibrium in which states do not acquire uncertainty, neither state accu-
mulates any military capacity.
We next show that an equilibrium in which neither state arms and therefore states re-
main certain about each other’s capabilities is the only possible equilibrium in which the
states never go to war.11 Put differently, states go to war with positive probability in any
equilibrium in which they generate uncertainty. We also show that an unarmed, completely
peaceful equilibrium exists only under a very limited set of circumstances. The intuition
11
The following result also implies that there is no equilibrium in which only one state plays a mixed
accumulation strategy (generates uncertainty) and war occurs with probability 0.

11
behind this next result is again simple: Because states do not directly observe each other’s
military capacity, each country can arm heavily without the other knowing. Thus, a situa-
tion in which both states are unarmed and peaceful is hard to sustain, because each has an
incentive to become strong secretly and start a war.
To find the weakest condition that supports an equilibrium in which neither state ac-
quires military capacity (m = 0), we first assume that such an equilibrium exists. In
such an equilibrium, the offer must be a1 = 1 − w2 (0, 0) as this maximizes player 1’s payoff
while keeping player 2 indifferent between war and the offer. This equilibrium requires that
neither party is willing to unilaterally deviate from mi = 0. Since β i > 0, such a devi-
ation is only worthwhile if war occurs (with positive probability) following the deviation.
There are two such deviations that must be ruled out: 1) The veto player unilaterally devi-
ates (accumulating a positive amount) and rejects a1 . This deviation is worthwhile only if,
w2 (m2 , 0)−β 2 m2 > w2 (0, 0). 2) The proposer deviates by accumulating capacity and offering
less to the veto player. This is only worthwhile if w1 (m1 , 0) − β 1 m1 > 1 − w2 (0, 0).

Proposition 2. War occurs with probability zero in a particular equilibrium if and only if
m = 0 with probability one in the equilibrium. Moreover, an equilibrium of this type exists
if and only if

w2 (m2 , 0) − β 2 m2 ≤ w2 (0, 0) for all m2 ∈ (0, ∞) (2.1)

and
w1 (m1 , 0) − β 1 m1 ≤ 1 − w2 (0, 0) for all m1 ∈ (0, ∞). (2.2)

Proof: Step 1: We first establish the first part of the proposition. Suppose that
there is an equilibrium in which war happens with probability 0. This means
that on the path the offer is accepted. Since β i > 0 a deviation from any lottery
putting probability on mi > 0 to another strategy with mi = 0 with probability
one would be desirable. Thus, m = 0 must occur with probability 1 in the
equilibrium. Assume that m = 0 with probability 1 in equilibrium. Given this
and assumption 2 it is well known that in all perfect equilibria of the ultimatum
game a settlement is reached.

12
 Step 2: We now establish that such an equilibrium occurs if and only if the stated
condition is satisfied. Given m = 0, sequential rationality in bargaining requires
that the proposer, 1, offer a1 = 1 − w2 (0, 0) as long as 1 − w2 (0, 0) ≥ w1 (0, 0).
This condition is true by assumption. There are two deviations to consider:
a deviation by the proposer, 1, to increase m1 and make an offer which is not
accepted or a deviation by the veto player, 2, to increase m2 and to reject the
offer. If 2 uses the above threshold strategy then 1’s best possible deviation is
desirable iff (2.2) is not satisfied. Alternatively 2’s deviation is desirable iff (2.1)
is not satisfied.¥

A completely peaceful equilibrium in which states do not hide information can exist —
but only under a particular set of circumstances (parameterizations). When we take the
model literally, assuming that states that do not purchase arms in the model are completely
unarmed, these circumstances seem likely to be satisfied very rarely in the real world. Propo-
sition 2 states that the certain-peace equilibrium exists only when it is better to be a weak
state negotiating with another weak state than a heavily armed state at war with an unarmed
adversary.
To give some sense of the difficulty in obtaining a completely peaceful equilibrium if we
interpret the baseline level of arms as zero, we turn to an example with specific payoffs for
war. One might think of many wars as contests: the state that has the greater military
capabilities wins, and takes the spoils of war, while the other state loses, and gets nothing.
Our next example reflects this idea: In the event of war, the more heavily armed state wins
and gets the prize; the other state loses and gets nothing, and the states split the prize if
they have acquired equal military capacity. Formally, for a number α ∈ (0, 1) we can define
⎧
⎪
⎪ α if mj > mi
⎪
⎨
wj (mj , mi ) = α (2.3)
if mj = mi
⎪
⎪ 2
⎪
⎩ 0 otherwise.

In this winner-take-all example, the conditions of proposition 2 are satisfied if and only if
α = 0. Thus, in this example, the peaceful equilibrium exists only if war provides no goods

13
to the victor.
We would argue, however, that the correct interpretation of Proposition 2 is less pes-
simistic than this example would suggest because states have some (at least partially) known,
baseline level of military capabilities when they make their decisions about armaments. If
this is the case, then the certain-peace equilibrium exists as long as each state’s net payoff
to secretly arming as much as it likes and starting a war is less than its payoff from the
status-quo bargain, or the bargain that reflects the status-quo distribution of power. This
condition is more likely to be satisfied because states sometimes are quite happy with the
status-quo bargain and therefore may not find it worthwhile to pay the costs of secretly
arming and launching a war.12
While one might think that states will be more likely to hide information about their
capacity when military capacity is more expensive, Proposition 2 states that this is not
the case. All else equal, the conditions for the completely peaceful equilibrium are more
likely to hold when the cost parameters are higher, because secretly arming is more costly.
Unfortunately, the interpretation of Proposition 2 therefore is more pessimistic if we believe
that military spending has positive externalities that are unrelated to the dyadic relationship
(e.g. bolstering the economy, pleasing domestic constituencies), which can be thought of as
lowering the cost parameter in the model.
In sum, this section has shown that in every equilibrium of the game, states act somewhat
unpredictably and keep secrets about their military capabilities, thus risking war, unless
each state derives no net benefit from war in which it is heavily armed relative to the status
quo. They do so even though the "strategic uncertainty" that accompanies this behavior
carries with it a risk of war.Empirically, a certain-peace equilibrium in which states do not
act unpredictably seems more plausible if states have some known level of military forces
acquired for other reasons, but less so if military expenditures driven by the adversarial
12
If each state begins with a known, status-quo level of capabilities, qi , then the relevant condition is as
in Proposition 2, but with each 0 replaced with qi . While we have assumed very little about a state’s payoff
from war, a reasonable assumption might be that the war payoff function is concave (exhibiting diminishing
returns) in the country’s own investment so that each state has greater incentive to secretly buy more arms
if both begin unarmed.

14
relationship have positive externalities.

3. Other Bargaining Protocols

We have seen that, when bargaining is characterized by the ultimatum game, states create
uncertainty about military capabilities and risk war if either condition (2.1) or condition (2.2)
is not satisfied. We now investigate whether this result extends to a larger class of bargaining
protocols. We find that it remains the case that if states do not create uncertainty and there is
a zero probability of war in any equilibrium then the states must remain completely unarmed
(or maintain the status quo levels of military capacity) in that equilibrium. Moreover, in all
of these models, rather stringent conditions (which represent natural extensions of conditions
2.1 and 2.2) must be satisfied in order for peaceful equilibria to exist.
The results in this section apply to all bargaining protocols that meet two conditions.
First, as we discussed at the start of the paper, as long as war is costly, there is always some
settlement that both states prefer to war. When states have complete information, they
can identify settlements of this form. Thus, we consider bargaining protocols under which
states reach an efficient settlement (one that fully divides the pie) if they have complete
information. This is a large class of protocols; it includes protocols with veto players and
the alternating-offer Rubinstein bargaining model that often is used to study international
conflict. Second, we would argue that any state eventually has the possibility of opting out
of bargaining and starting a war without the consent of its adversary, though it may never
choose to do so. If instead both states have to consent to war, a state with military capacity
could be forever stuck in a bargaining process it does not like, with no option of unilaterally
using whatever forces it possesses. Thus, we assume that either state can end the bargaining
and start a war, unilaterally and in a finite amount of time (but it need never do so). The
first condition is a bit more subtle as it restricts protocols based on particular aspects of
their equilibrium set; the second condition is more straightforward as it speaks only to the
structure of the protocol and not any description of how states will actually behave in the
protocol. We now give formal definitions.
Formally, the bargaining protocol includes a description of the available offers, sequence of

15
speaking, and how behavior maps into payoffs. Specifically, we define a bargaining protocol
to be a pair of strategy spaces, S1 , S2 and a mapping b(ξ 1 , ξ 2 ) : S1 × S2 → A × N. The
strategy spaces may be finite, countably infinite, or uncountably infinite. The set A is defined
to be the union of the 2 dimensional simplex and a particular outcome that we call war. We
denote this outcome by the vector (−1, −1). Specifically,

A := {(p1 , p2 ) ∈ R2+ : p1 + p2 = 1} ∪ {(−1, −1)}. (3.1)

The mapping b(ξ 1 , ξ 2 ) 7→ (p1 (ξ 1 , ξ 2 )), p2 (ξ 1 , ξ 2 ), t(ξ 1 , ξ 2 )) assigns a vector of allocations to

the two states and a time t ∈ N := {1, 2, ...} at which the decision is reached for any profile
of bargaining positions . If (−1, −1) is the outcome and it is reached in period t, then
we say that bargaining breaks down in period t and war occurs. The payoffs from this
outcome are given by δ t−1 wi (mi , mj ). If a settlement is reached in period t then state i0 s
payoff is δ t−1 pi (ξ 1 , ξ 2 ). The term δ ∈ (0, 1] serves as the common discount factor.13 Since
we are not modeling the details (timing and information sets) in the bargaining protocol
we use a Normal form game refinement instead of PBE. Thus, we require that strategies
form a trembling hand perfect equilibrium to the investing and bargaining game. Because
trembling hand perfect equilibria are sequential equilibria, this requires that strategies in
the bargaining histories are sequentially rational to beliefs about investments and that the
beliefs are consistent.
In order to state the assumptions rigorously, we need to be able to talk about m being
known. We say the accumulation m is known at the beginning of the bargaining protocol if
player 1’s belief about player 2’s accumulation is concentrated at the correct value, player 2’s
belief about player 1’s accumulation is concentrated at the correct value, and both of these
facts are common knowledge. The first assumption states that when the accumulations are
known, sequentially rational play results in a peaceful settlement.
Condition 1 (war is inefficient): Whenever m is known, every profile of sequentially
rational play in the bargaining protocol results in some bargaining solution (p1 , p2 , t) that
13
We are assuming that bargaining costs can be represented adequately by the discount factor — that is, if
settlement is delayed by a period, each state gets a smaller payoff in present value, which is similar to paying
a cost.

16
satisfies the conditions that p1 + p2 = 1 and t is finite.
The second assumption ensures that war is not a consensual outcome; either party can
in a finite period of time initiate a conflict (the outcome (−1, −1)) regardless of the other
player’s bargaining behavior. This is an assumption about what is possible, and not about
what occurs in equilibrium; neither state need ever start a war.
Condition 2 (war is not consensual): There exists a finite t0 such that for either
i ∈ {1, 2} there exists a strategy ξ 0i ∈ Si s.t. for any ξ j ∈ Sj , b(ξ 0i , ξ j ) = (−1, −1, t) for some
t < t0 .
The ultimatum bargaining protocol of the previous section does not satisfy condition 2,
since player 1 must make an offer and in principle player 2 could play the strategy: accept
any offer. However, it is easy to see that if the ultimatum game were modified so that player
1 could just start a war instead of making an offer the equilibrium set would not change.
In the large class of bargaining models that satisfy these two conditions, we find, again,
that in any equilibrium in which war occurs with probability zero, the states remain unarmed
with probability one. For the class of bargaining models, peaceful equilibria exist only when
a condition exists that is analogous to the one we presented earlier. We begin by showing
that states create uncertainty (accumulate military capacity in mixed strategies) in any
equilibrium unless the states remain completely unarmed. That is, we generalize proposition
1.

Proposition 3. There are no equilibria with pure-strategy accumulations satisfying max{mi , mj } >
0.

Proof: Assume that such an equilibrium exists with efforts m. Since the equi-
librium is in pure strategy accumulations, states possess all relevant information
when they begin bargaining. By condition 1 a bargain will be reached with prob-
ability one. Since effort is not observed, a deviation to mi = 0 (and no change in
i’s bargaining actions) would not affect the outcome of the bargaining protocol.
Since β i > 0, for a state with mi > 0 such a deviation would increase its utility.¥

The next result establishes that the states do not go to war if they remain completely
unarmed. However, if states choose to arm and create uncertainty about their military

17
capacities, there is always a positive probability that they go to war. This result rules
out equilibria in which states generate uncertainty (accumulate military capacity in mixed
strategies), but always reach a peaceful settlement through negotiation. It generalizes the
first half of proposition 2.

Proposition 4. War occurs with probability 0 in an equilibrium if and only if m = 0 with

probability one in this equilibrium.

Proof: We proceed in steps.

(⇐=)If the players play pure accumulation strategies, then m is known. Condi-
tion 1 then implies that war does not occur.

(=⇒)Suppose that there is an equilibrium in which war happens with probability

0. This means that on the path some offer ai is accepted. Since β i > 0 a deviation
from any non-degenerate lottery to mi = 0 (and no change in i’s bargaining
behavior) would be desirable. Thus, if we are in an equilibrium and no deviation
is desirable it must be the case that m = 0.¥

Finally, it is possible to present partial analogues to the necessary and sufficient conditions
for the existence of the unarmed, peaceful equilibria (2.1 and 2.2). We have assumed that
the states reach a peaceful bargain if they have complete information. We show next that
— as with the particular bargaining protocol we considered earlier — a sufficient condition
for the peaceful equilibrium is that neither state can benefit from unilaterally arming and
starting a war when the other state is unarmed. Establishing necessary conditions for the
unarmed, peaceful equilibrium is a bit more challenging. In principle, the value to a state
of deviating from its peaceful equilibrium strategy, arming, and starting a war can be less
than its full value for war (net of the costs of arming) because of the costs of delay. We
have limited our consideration to bargaining protocols that allow either state to start a war,
but not necessarily immediately. Thus, with some protocols in the class we are considering,
if a state deviates by arming and starting a war, it will receive only a discounted payoff
from war. This implies that even when the sufficient conditions for the peaceful equilibrium
are not satisfied, neither state may wish to deviate from its peaceful-equilibrium strategy of

18
remaining unarmed. Let Π0 denote the set of expected payoffs pairs that are feasible given
some sequentially rational profile of bargaining behavior when it is known that neither state
has accumulated any military capacity (m = 0).14

Proposition 5. (i) There is an equilibrium with m = 0 if

wj (mj , 0) − β j mj ≤ π j for all mj ∈ (0, ∞) (3.2)

and
wi (mi , 0) − β i mi ≤ π i for all mi ∈ (0, ∞). (3.3)

for some (πi , π j ) ∈ Π0 .

(ii) There is an equilibrium with m = 0 only if

β j mj πj
wj (mj , 0) − t0 −1
≤ 0 for all mj ∈ (0, ∞) (3.4)
δ δ t −1
and
β i mi πi
wi (mi , 0) − t0 −1
≤ t0 −1 for all mi ∈ (0, ∞). (3.5)
δ δ
for some (πi , π j ) ∈ Π0

Proof: (i) By assumption, following m = 0 the payoffs correspond to the right-

hand side of the inequalities for some equilibrium selection. We now consider a
unilateral deviation by i. Given that (π i , πj ) ∈ Π0 the most that i can get from
a peaceful settlement is π i . If this were not true then (π i , πj ) would not be an
equilibrium payoff to bargaining when it is known that m = 0. This implies that
the deviation to mi > 0 is profitable only if the deviation results in a lottery that
puts positive probability on war, and wi (mi , 0) − β i mi > π i . This is not possible
if wi (mi , 0) − β i mi ≤ π i for all mi ∈ (0, ∞). Thus, the result is established.

(ii) Step 1: First note that for every pair (π i , π j ) ∈ Π0 and the supporting
equilibrium strategy profile, given knowledge of m = 0, there is no unilateral
14
Recall, a payoff is the discounted value of an outcome. Thus each coordinate of a payoff vector in this
set is of the form π i = δ t−1 pi (ξ i , ξ j ).

19
 deviation by i from the conjectured equilibrium strategy that involves mi = 0
and different actions in the bargaining game that results in a payoff higher than
π i . Second, consider a unilateral deviation from a particular profile that starts
with m = 0 and has sequentially rational play in the bargaining protocol. Any
such deviation that (i) reaches a settled outcome (not (−1, −1)) with probability
one and (ii) involves mi > 0 could be improved upon by the strategy that has
mi = 0 and mimics this deviation. These two points imply that in considering
equilibria with m = 0 it is sufficient to consider deviations that reach war with
positive probability and yield war payoffs that exceed π i .

Step 2: Assume that an equilibrium with m = 0 exists. The conclusion of step

1 implies that it must be the case that for the t0 defined in condition 2

0
δ t −1 wi (mi , 0) − β i mi ≤ π i for all mi ∈ (0, ∞). (3.6)

If this were not true, then the equilibrium with m = 0 would not exist as i
could unilaterally increase its payoff by selecting some mi > 0 and selecting the
bargaining action ξ 0i defined in condition 2. Since t0 is finite, multiplication by
1
δt
0 −1 is permissible and the conclusion is established.¥

In order to contrast this results with conditions 2.1 and 2.2, consider bargaining protocols
in which Π0 is a singleton. The last result demonstrates that

wi (mi , 0) − β i mi ≤ π i for all mi ∈ (0, ∞) (3.7)

is sufficient but not necessary for peaceful equilibria. Recall that with the particular bar-
gaining protocol considered earlier in the paper, this condition is necessary and sufficient.
Within the larger class of bargaining protocols that we consider here, the necessary and suffi-
cient conditions become analogous to conditions 2.1 and 2.2 if the states become sufficiently
patient. (They converge to the natural analogues of conditions 2.1 and 2.2 in the limit as δ
goes to 1.) Similarly, if the states have the option of starting a war immediately, though they

20
may choose not to exercise it (that is, if t0 = 1), then the peaceful-equilibrium conditions are
exactly analogous to conditions 2.1 and 2.2.
Like propositions 1 and 2, this result shows that states are less likely to hide information
when arming is costly, because they are more likely to remain completely disarmed, or armed
only at the known status-quo level. Looking at the left-hand side of the inequalities, one can
see that the payoff to unilaterally arming and starting a war goes down as the per-unit cost
of arming (β) goes up; thus states are more content to remain peacefully disarmed.
Within this large class of bargaining models, it also remains the case that states risk war
whenever they arm in a way that creates uncertainty. Moreover, if states are sufficiently
patient or have the option of starting a war immediately, the completely peaceful equilibrium
exists only if there is no incentive to secretly arm, unilaterally and perhaps heavily, and to
start a war, thus overturning the status-quo bargain. If states are impatient, the existence
of the completely peaceful equilibrium depends upon the states’ degree of patience and/or
on the time that it takes them to mobilize forces and begin a war.

4. Existence of equilibria

In light of our focus on understanding the circumstances in which states choose not to create
uncertainty, a natural question emerges. Do our models have equilibria in which states
do create uncertainty — that is, equilibria in mixed strategies? If we assume that the set
of possible investment levels and the strategy spaces in the bargaining protocol are finite
(albeit arbitrarily large), then Selten (1975) guarantees that the answer is "yes"; trembling
hand perfect equilibria exist in mixed strategies. We now discuss how the main results of
the previous section hold in finite versions of the games we present there.
Consider a setting in which the set of armaments is a finite set Mi ⊂ R+ with 0 ∈ Mi
(and Mj ⊂ R+ with 0 ∈ Mj ) and in the bargaining protocols, Si and Sj are finite. Assume
that conditions 1 and 2 are satisfied.15 In this context, propositions 3-5 hold and the only
15
Note that the assumption of finite strategy spaces in the bargaining protocol does not limit the possible
bargains in a substantively meaningful way. Since the set of possible armaments also is finite in these versions
of the games, the baragaining space can be much larger than the armament space. Thus, it is still reasonable

21
change to the proofs is that phrases like mi ∈ (0, ∞) are replaced by phrases like mi ∈ Mi .

Proposition 6. Suppose that Mi ⊂ R+ , Mj ⊂ R+ are both finite with 0 ∈ Mi and 0 ∈ Mj ,

and the bargaining protocol is such that Si and Sj are finite.
(i) This game possesses a (possibly mixed strategy) trembling hand perfect equilibrium.
(ii) There are no equilibria with pure-strategy accumulations satisfying max{mi , mj } > 0.
(iii) War occurs with probability 0 in an equilibrium if and only if m = 0 with probability
one in this equilibrium.
(iv) There is an equilibrium with m = 0 if

wj (mj , 0) − β j mj ≤ π j for all mj ∈ Mj (4.1)

and
wi (mi , 0) − β i mi ≤ π i for all mi ∈ Mi . (4.2)

for some (πi , π j ) ∈ Π0 .

(v) There is an equilibrium with m = 0 only if

πj
wj (mj , 0) − β j mj ≤ 0 for all mj ∈ Mj (4.3)
δ t −1
and
πi
wi (mi , 0) − β i mi ≤ t0 −1
for all mi ∈ Mi (4.4)
δ
for some (πi , π j ) ∈ Π0 .

Thus, we know that each finite version of the games we have discussed must have an
equilibrium in which states create uncertainty and go to war with positive probability, unless
the conditions for a completely-peaceful equilibrium are satisfied. In light of this result, we
see that the non-existence of certain-peace equilibria when conditions 4.3 and 4.4 are not
satisfied should not be attributed to a technical pathology of the games that make equilibrium
itself poorly defined. Instead, the non-existence of certain-peace equilibria can be attributed
to an incentive problem that leads countries to create and keep secrets about their military
capabilities.
to assume that with complete information, states will reach a bargain short of war.

22
 Establishing existence of mixed-strategy equilibria in the infinite versions of the games we
examine is non-trivia. Recall that a mixed strategy equilibrium involves both lotteries over
investments and investment-contingent bargaining strategies; the investment lotteries must
solve a fixed-point problem that is induced by the bargaining strategies, and the bargaining
strategies must be sequentially rational (given beliefs). As such, the existence problem is
substantially harder than the one in Jackson and Morelli (2007), because the post-investment
strategies are more clearly pinned down in their model.

5. Acquisitions of Military Capacity Can Be Observed

Thus far, we have considered a large class of models in which states do not observe each
other’s military acquisitions directly, but could choose to maintain complete information by
buying military capacity in pure strategies. We have shown the difficulties in avoiding the
creation of uncertainty about military capacity and an associated risk of war. It is easy
to see that these difficulties do not disappear if states can engage in cheap talk after they
acquire arms.
But do states create uncertainty about their military capacities act somewhat unpre-
dictably in acquiring military capacity even when they can observe each other’s military
acquisitions? We now consider two alternative extensions of our models in which states can
observe each other’s military acquisitions to varying degrees in order to address this question.
These extensions build on the general model of Section 3.

5.1. Perfect Signals about Military Acquisitions

In the first model, states can choose whether or not to demonstrate their military acquisitions
to their adversaries; if a state decides to do so, the adversary learns the information precisely
(and the state cannot secretly acquire additional capacity). This model also can represent
a situation in which a state can choose to allow an international institution to provide an
adversary with information about its military capacity and the institution has the technology
to acquire and provide this information perfectly.

23
 As in our earlier models, the states begin by selecting their capacities, mi , mj , simultane-
ously. Now, however, each state also selects a message si ∈ {mi , φ} prior to the bargaining
process. If a state chooses si = mi , this signal informs the adversary completely about the
state’s military acquisitions. (One can think of the message as a verified statement or a
perfect demonstration). If the state chooses si = φ, it is deciding not to allow the adversary
to observe its military acquisitions. After the messages, the states bargain; if bargaining
fails, they go to war. In this section, we assume that bargaining occurs according to the
ultimatum bargaining protocol, as in the first model in this paper.
In this setting, a state cannot deviate from a conjectured equilibrium strategy profile that
involves informative communication without letting its opponent know that it has deviated.
The requirements for a pure strategy equilibrium m = (m1 , m2 ), s = (m1 , m2 ) are (1) that

m1 ∈ arg max{1 − w2 (m2 , m) − β 1 m} (5.1)

m

m2 ∈ arg max{w2 (m, m1 ) − β 2 m},

m

and (2) that given the strategy profile no player has an incentive to deviate to a pair m0i , φ
with m0i 6= mi .
To see the possibility of pure-strategy equilibria, we consider two examples. The first
follows other works in the international-relations literature (e.g. Powell (1999)[Chapter 2])
in assuming that a state’s payoff from war is increasing in the share of the two states’ forces
that it possesses:
⎧
⎨ pmi
if (mi , mj ) 6= (0, 0)
mi +mj
wi (mi , mj ) = (5.2)
⎩ p
if (mi , mj ) = (0, 0).
2

With p ∈ (0, 1), first order conditions from (1) yield

pm2
= β1 (5.3)
(m1 + m2 )2
pm1
= β2
(m1 + m2 )2

24
and the solution is given by

pβ 2
m1 = (5.4)
(β 2 + β 1 )2
pβ 1
m2 = .
(β 2 + β 1 )2
These values satisfy the second-order conditions at any values of m1 and m2 . So as long
as the payoffs to these accumulations exceed 0, condition (1) is satisfied. The boundary
condition requires

pβ 2 pβ 1 β 2
≥ (5.5)
pβ 2 + pβ 1 (β 2 + β 1 )2
pβ 1 pβ 1 β 2
≥ .
pβ 2 + pβ 1 (β 2 + β 1 )2
Simplifying yields

½ ¾
pβ 1 pβ 2
1 ≥ max , . (5.6)
β2 + β1 β2 + β1
In order to check that condition (2) is satisfied, we need to specify off-the-path beliefs
for bargaining games in which si = φ. The equilibrium is supported if the beliefs assign
probability 1 to mi = 0 given si = φ. To see this, note that we already have seen that the
equilibrium payoff to i exceeds the payoff to bargaining as if mi = 0. It remains to verify
that i does not prefer to arm, announce si = φ , and then have a conflict. For country 1,
the maximum payoff to this deviation solves

pm1
m1 ∈ arg max{ − β 1 m1 }, (5.7)
m1 + m2

which has the same first-order condition as above since the partial derivatives of 1 − mpm 2
1 +m2
pm1
and m1 +m2
with respect to m1 are the same. Since the second state is solving the same
problem, the optimal deviation cannot improve its payoff. It is interesting to note that in
this equilibrium, secrecy is interpreted as a sign of weakness. That is, when states reveal
information about their capacities in equilibrium, but have the option of choosing not to,

25
opponents must treat failure to reveal information as a signal that i is weaker than the equi-
librium level mi . Were this not true, i would benefit from hiding its information (changing
si from mi to φ). The results of our analysis of this example are as follows.

Proposition 7. With the war technology in 5.2 and the ability to demonstrate capacity
perfectly prior to play of the ultimatum game, there is an equilibrium in which the states
select capacities in pure strategies, reveal their capacities, and reach a negotiated settlement;
war occurs with probability 0.

Is the beneficial effect of perfect monitoring limited to the particular military technology
that we have just considered? Earlier in the paper, we considered a situation in which war
payoffs are like a winner take all contest (2.3). With these payoffs, the game also has pure-
strategy equilibria in which the states avoid war completely.
To see the possibility of pure-strategy equilibria, note that an equilibrium of the form
m = (m1 , m2 ), s = (m1 , m2 ) must result in a tie or at least one player must be selecting
mi = 0. In either of these cases, one might think that a player could improve its payoff
by selecting a slightly higher accumulation than the opponent and announcing it. However,
α
equilibria of this sort can exist because a country is willing to accumulate at most mi = βi
α
in order to win a prize worth α, or 2β i
in order to tie. Thus, for example, when β i = β j ,
α
there is a pure-strategy equilibrium in which both states accumulate 2β i
and avoid war.
Overall, the results in this section seem encouraging. While they rely upon a particular
bargaining protocol, a more general version could be developed that allows for a large class
of protocols. In addition, the results from the first example (in which the payoff from war
is increasing in the state’s share of the military forces) can be readily extended to necessary
and sufficient conditions for pure-strategy equilibria under more general differentiable payoff
functions. Extensions of this form are technical. Instead, we conclude the paper with an
important and surprising robustness check.

5.2. Noisy Signals about Military Acquisitions

The previous section showed the benefits of being able to demonstrate military capacity
perfectly: If statements about military acquisitions are known to be correct, states may

26
arm and yet always avoid war. However, in the real world, technologies for monitoring
others’ military capacity and thus verifying their claims about their capabilities are imperfect.
For this reason, we now consider a model in which states obtain information about their
opponent’s choices of military capacity, but this information is imprecise. This model, which
extends the general model of Section 3, differs in two important ways from the one in the
previous section. First, a state does not choose whether or not to give information about its
capacity to the opponent; instead, the opponent exogenously receives a signal. Second, the
signal a state receives about its adversary is noisy.
If instead of demonstrating one’s strength, (si = mi ) states only can observe a noisy
signal of strength, si = mi + ε, where ε might be thought of as white noise, the encouraging
findings of the previous section do not generalize. No matter how small the variance of ε, if
the shocks have full support and players are sufficiently patient then a peaceful equilibrium
exists only under strong conditions like those presented in part (ii) of proposition 5. Thus,
there is a disjuncture between complete demonstration of capacity, which we considered in
the previous section, and noisy signals, which we consider here. Even a little bit of noise
in the information a player gets about its opponent’s capacity is enough to make war a
possibility.
We now assume that after the states have simultaneously chosen their levels m1 and
m2 , but before they bargain, each state i receives an exogenous noisy signal σ i ∈ R1 about
the other’s (j’s) chosen capacity. In the event that a state plays a mixed strategy, the
opponent receives a noisy signal about the realization of the mixed strategy. We assume
nothing about the distribution of the signal F (σi |mj ) except that it has full support for
any level of capacity mj that the state may have chosen and that conditional on mi , mj the
signals are independent.16 The full support assumption states that for two distinct levels
m0j and m00j the conditional distributions F (· | m0j ) and F (· | m00j ) have the same support.
This assumption requires that a state cannot completely "rule out" any level of capacity
because of the signal it has observed. This assumption can be satisfied in settings in which
16
It would be natural to assume that the distributions F (· | m0j ) and F (· | m00j ) are ordered (in terms of
likelihood ratio or first order stochastic dominance) if m0j < m00j . Our analysis is, of course, consistent with
this type of assumption, but this structure is not needed for the following results.

27
the conditional distributions have arbitrarily low variance. For example (though this need
not be the case), it could be that the distribution of the signals that i receives is normally
distributed with a mean of mj and variance ε > 0, where ε is very small.
The reason why many of the results about the difficulties of avoiding uncertainty and
war hold with partial observability is that the addition of the noisy signals σ =(σ 1 , σ 2 )
does not change the fact that if states acquire capacity in pure strategies, they know each
other’s capacity precisely at the beginning of the bargaining stage and will therefore reach
an agreement. Obviously, this is not true for capacity accumulations in mixed strategies.
More technically, if state j acquires mj = m1j with probability one in equilibrium, then by
Bayes’ Rule, state i must believe that mj = m1j with probability one after seeing any σ j ,
since the prior belief is that mj = m1j and F (σ j |mj ) has full support for all mj .
Why are the cases of perfect and imperfect observability different? With perfect ob-
servability, any deviation from a pure-strategy equilibrium is immediately known. With
imperfect observability, a state cannot observe when its adversary has deviated.
We begin again by showing that states create uncertainty (accumulate military capacity
in mixed strategies) in any equilibrium unless the states remain completely unarmed and
that states do not go to war if they remain completely unarmed.

Proposition 8. (i) There are no equilibria in which the accumulations are in pure strate-
gies with max{mi , mj } > 0.(ii) In any equilibrium in which the accumulations are in pure
strategies, war occurs with probability 0.

Proof: (i) Assume that such an equilibrium exists with efforts {m∗i , m∗j }. Since
the equilibrium accumulations are in pure strategies and F (σ i |mj ) has full sup-
port for all mj , states possess all relevant information when they begin bargain-
ing. By condition 1, a bargain will be reached with probability one. Since state
j must believe that mi = m∗i after seeing any σ j , a deviation to mi = 0 would
not affect the outcome of the bargaining protocol. Since β i > 0, for a state with
m∗i > 0 such a deviation would increase its utility.

(ii) Consider an equilibrium with pure strategy accumulations. Since the equi-
librium is in pure strategy accumulations and F (σi |mj ) has full support for all

28
 mj , by Bayes’ Rule, i continues to believe mj = m∗j after seeing any σ i and j
continues to believe mi = m∗i after seeing any σ j . Condition 1 then implies that
war does not occur.¥

Finally, we show that if states generate uncertainty in equilibrium, there is always a

positive probability that they go to war.

Proposition 9. If war occurs with probability 0 in an equilibrium that involves no delay17 or

in any equilibrium to a game with δ = 1, then m = 0 with probability one in this equilibrium.

Proof: We proceed in steps.

Step 1: We consider an equilibrium with pure-strategy accumulations. Suppose

that there is an equilibrium in which both states’ accumulations are in pure
strategies, at least one state’s accumulation is not zero (mi 6= 0), and the states go
to war with probability 0. This means that on the path some offer ai is accepted
with probability 1. Since the equilibrium involves pure strategy accumulations,
by Bayes’ Rule, i continues to believe mj = m∗j after seeing any σ j and j continues
to believe mi = m∗i after seeing any σ i . Since β i > 0, a deviation to mi = 0 would
be desirable. Thus, if we are in an equilibrium with pure strategy accumulations
and no deviation is desirable, it must be the case that m = 0.

Step 2: Now we consider an equilibrium in which player i’s accumulation is

in mixed strategies with a support containing two distinct levels, m0i and m00i .
Supposed that there is such an equilibrium in which the states go to war with
probability 0. This means that on the path an offer is accepted with probability
1 (and we can select m0i and m00i such that war occurs with probability 0 following
these two accumulation levels). Since both m0i and m00i are in the support of i’s
equilibrium strategy, we must have vi0 − vi00 = β i (m0i − m00i ) where vi0 and vi00 denote
the expected discounted payoffs to the equilibrium settlement when i selects m0i
and m00i respectively. This means that vi0 6= vi00 .
17
To be clear, no delay means that following any profile of accumulations m and messages that are possible
in equilibrium, the resuling bargaining strategies reach a settlement before any discounting occurs.

29
 Step 3: Now let ξ i (mi , σ i ) and ξ j (mj , σ j ) denote the mappings from choices in the
accumulation stage into strategies in the bargaining protocol. With this notation,
we can express equilibrium expected discounted payoffs from the bargaining set-
tlement as functions of the form vi (ξ i (mi , σ i ), ξ j (mj , σ j )) and vj (ξ i (mi , σ i ), ξ j (mj , σ j )).
Since war occurs with probability 0 on the path following both m0i and m00i it can-
not be the case that vi (ξ i (m0i , σ i ), ξ j (mj , σ j )) 6= vi (ξ i (m00i , σ i ), ξ j (mj , σ j )). If this
inequality held, say with the former larger than the latter, then player i would
have an incentive to deviate (say by playing ξ i (m0i , σ i ) when mi = m00i ). Recall
that state j would not know that i had deviated (because mi is hidden and and
the support of F (σ j |m0i ) coincides with the support of F (σ j |m00i ), and thus player
j́’s bargaining behavior could not respond to the deviation. Thus we have shown
that vi (ξ i (m0i , σ i ), ξ j (mj , σ j )) = vi (ξ i (m00i , σ i ), ξ j (mj , σ j )).

Step 4: But since vi0 6= vi00 , it must be the case that ξ j (mj , σ 0j ) 6= ξ j (mj , σ 00j )
for some realizations σ 0j and σ 00j . Now since i does not observe σ j and since
σ i and σ j are independent (by construction the mixtures that players use are
independent and thus the signals are unconditionally independent) it cannot be
the case that vj (ξ i (mi , σ i ), ξ j (mj , σ 0j )) 6= vj (ξ i (mi , σ i ), ξ j (mj , σ 00j )) for any values
of mi , σ j , mj . This is true because if the inequality held, j would have an
incentive to deviate (this can be shown by using a very similar argument). So,
holding fixed mj , the expected discounted payoff from bargaining to j cannot
be different under m0i and m00i . But since war happens with probability 0, the
assertion that vi0 6= vi00 contradicts the assumption that equilibrium settlements
sum to 1 and the assumption that δ = 1 or the equilibrium involves no delay.¥

We also note that the necessary and sufficient conditions for the existence of the unarmed,
peaceful equilibrium are exactly as given in Proposition 5. To see this results, remember that
in an equilibrium with m = 0, since F (·|mj ) has full support for all mj and the equilibrium
is in pure strategies, m = 0 remains known after the signals are observed. The proof then
proceeds exactly as given in the previous section. (For the proof of part (ii) with the added
noisy signals, note that any deviation that reaches settlement with probability one and

30
involves mi > 0 could be improved on by a strategy that has mi = 0 and mimics this
deviation because j will believe mi = 0 after any signal it receives.)
Without discounting, we again rule out equilibria in which states act unpredictably and
keep secrets about their military acquisitions, thus making their adversaries uncertain about
their overall levels of military capacity, but never go to war. With discounting, we cannot
rule out such equilibria. However, because we have put so few restrictions on the payoffs from
war, our results apply very generally, even to situations that one might consider to involve
discounting. For example, a situation in which a state must engage in time-consuming
mobilization before starting a war can be represented as one in which the state has a lower
(expected) value for war instead of as one with discounting; in this case, our results apply.
The existence and precise form of any equilibria in which states act unpredictably but
always remain at peace depend on the particular bargaining protocol. However, we can make
several observations about their features. First, if an equilibrium of this form exists, it must
involve states taking some time to reach a bargain (delay). Second, a state’s (i’s) payoffs
cannot be affected by the signal σi that it observes about the adversary’s capabilities; if
it were, the state always would pretend to observe the signal that gave it the best payoff.
Thus, in such an equilibrium, the noisy signals are not valuable. This finding leads to an
interesting conclusion. States would not be willing to devote resources to see a noisy signal
of the form in this model in any equilibrium in which war occurs with probability 0. Thus,
while we have not extended proposition 4 to the case of noisy signals, equilibria that involve
war with probability 0 but nondeterministic arming require that states can observe signals
that are meaningless to them (in the sense that what a state learns about an opponent’s
strength does not affect its payoff). This finding leads us to conclude that a satisfactory
explanation of arms accumulation and bargaining that involves uncertainty and costly spying
must also involve a risk of war.
Third, if an equilibrium in which states act unpredictably but never go to war exists,
the payoff of any state i that generates uncertainty must be affected by the signal that its
adversary observes about its capacity, σ j . For a state to be willing to create uncertainty
about its payoffs, it must expect a better bargain, on average, when it has invested more in

31
military capacity; thus, the adversary’s actions must be tied to the signals it receives. The
second and third points together imply that an equilibrium of the noisy signals game with
uncertainty and without war must take a particular form: The signal a state receives about
its adversary must affect the adversary’s expected payoff without affecting its own. This is
only possible if the probability that a settlement occurs in a given period is affected by the
signals.
Thus, without discounting, our results about the difficulties of avoiding secrets and war
extend to a version of the model that includes noisy signals about the states’ military ac-
cumulations — even when those signals are just a little bit noisy. Even with some ability
to observe each other’s accumulations, states act somewhat unpredictably and keep secrets
about their military acquisitions in every equilibrium except the ones in which they remain
disarmed, or armed at some known, status-quo level. In addition, a completely peaceful
equilibrium exists only as long as each state’s net payoff to secretly arming as much as it
likes and starting a war is less than its payoff from the status-quo bargain. With discounting,
we do not rule out the possibility of a completely peaceful equilibrium in which states create
uncertainty about their military capacities. Such an equilibrium only can exist, however,
when it takes time to reach a bargain, and when (in equilibrium) getting information about
an adversary’s military capacity is not valuable.

6. Conclusion

Much of the literature on war and on international institutions rests on the assumption that
countries have private information about their military capabilities. Yet military capabilities
are not simply an extant feature of the world; they are the result of choices made by states,
or by their leaders. This paper helps to explain war by explaining why states arm in such a
way as to create uncertainty about their military capabilities — that is, why they build their
military strength in such a way that the overall level is not completely predictable, and hide
information about both their budgets and particular programs.
We have presented a simple model in which states acquire military capacity, bargain,
and go to war only if they fail to reach an agreement. The model has no ex ante private

32
information. If states maintain a situation of full information (their investments in military
capacity are in pure strategies or they choose to reveal their military capacity), there is
always a bargain they prefer to war. Yet, unless there is no incentive to arm unilaterally
in secret and start a war, states arm in a way that makes adversaries uncertain about their
capabilities and go to war with positive probability in every equilibrium of our games. States
behave this way because of competing incentives — the incentive to prevail in war, and the
incentive to minimize military expenditures. These conclusions are quite robust; they follow
from a large class of bargaining models, and from an alternative version of the model in
which states get noisy signals about their adversaries’ military acquisitions.
Taking our model literally, the conditions for the equilibrium in which states maintain
complete information and never go to war seem unlikely often to obtain in practice. If states
begin their interaction with no arms, then the certain-peace equilibrium exists only if it
is better for each state to have the bargain reached when both are unarmed than to be
a superpower at war with an unarmed adversary. Assuming instead that states begin with
some known level of armaments, our conclusions are somewhat more optimistic; certain peace
exists if there is no incentive to arm in secret, unilaterally and perhaps heavily and then to
start a war, relative to remaining at some status-quo bargain based on the ex ante distribution
of military power. However, certain peace is less likely when the costs of arming is lower, so
that positive externalities of arming make it more likely that states create uncertainty and
go to war with positive probability in practice.
The drawback of the generality of our models is that we are not able to derive testable
propositions. As Kennan and Wilson (1993) explain, the detailed empirical implications of
bargaining models generally vary with the particulars of the bargaining protocol. We do
not restrict our model to a particular protocol because states use a wide range of bargaining
procedures when negotiating and we want our results to be widely applicable. For a particular
international interaction, one could determine a reasonable protocol and payoffs from war,
and then derive testable propositions. Such an exercise remains for future work.
While the substantive conclusions of the analysis are negative in the sense that completely
peaceful equilibria exist only in circumstances that seem to us to be empirically fairly rare,

33
one important positive interpretation should be noted. The model and analysis broaden the
scope of rationalist studies of security. By moving beyond a taxonomy of the types of asym-
metric information and explaining the emergence of this critical feature of the international
landscape, the model provides a framework from which future studies may learn how bar-
gaining procedures, bilateral agreements, and international institutions can limit the scope
of costly asymmetric information and investment in military capacity.
In the introduction to this paper, we noted that international institutions often are
thought to reduce the likelihood of war in part by providing information to states and less-
ening the extent of harmful uncertainty. Our paper points to unanswered questions about
international institutions. We find that desirable equilibria in which states reveal informa-
tion and do not go to war are possible when the institution allows states to demonstrate
their military capabilities perfectly (with no noise or possibility of cheating). Regrettably,
this conclusion hinges on the term "perfectly;" even a minuscule amount of noise (or doubt
on the part of a state receiving a signal) undermines the effectiveness of the institutions. We
might think of this conclusion as saying that "close-to-perfect" signals and information are
not good enough. An important caveat is that our focus on generality has come at a price.
We have not characterized the equilibrium likelihood of war in these settings. It remains
possible that the "closer-to-perfect" the signals are, the closer to peace the international
order remains. We leave this investigation to future work.
Finally, states in our model begin with no private information but have incentives to ac-
quire it. The existing literature does not explain how international institutions can overcome
states’ incentives to acquire additional uncertainty once the institutions (hypothetically) have
revealed all existing uncertainty. Thus, a logical next step is to investigate whether any vol-
untary monitoring schemes beyond those we consider here can ameliorate the incentives for
arming and secrecy. Overall, we hope that a broader investigation of international orga-
nizations will help explain what types of mechanisms can reduce the likelihood of military
conflict.

34
References

Baliga, S. and T. Sjöstrom (2007). Strategic ambiguity and arms proliferation. Unpub-
lished manuscript.

Banks, J. S. (1990, August). Equilibrium behavior in crisis bargaining games. American

Journal of Political Science 34(3), 599—614.

Blainey, G. (1973). The Causes of War. New York: The Free Press.

Fearon, J. D. (1992). Threats to Use Force: Costly Signals and Bargaining in International
Crises. Ph. D. thesis, University of California, Berkeley.

Fearon, J. D. (1995). Rationalist explanations for war. International Organization 49,

379—414.

Fey, M. and K. W. Ramsay (2005). Mutual optimism and war.

Kennan, J. and R. Wilson (1993). Bargaining with private information. Journal of

Economic Literature 31(1), 45—104.

Keohane, R. O. (1984). After Hegemony. Princeton: Princeton University Press.

Kydd, A. (2000, November). Arms races and arms control: Modeling the hawk perspective.
American Journal of Political Science 44(2), 228—244.

Levontoglu, B. and A. Tarar (2005). War and incomplete information. paper presented at
the Annual Meeting of the American Political Science Association.

Morrow, J. D. (1989a). Capabilities, uncertainty, and resolve: A limited information model

of crisis bargaining. American Journal of Political Science 33(4), 941—972.

Morrow, J. D. (1989b). A twist of truth: A reexamination of the effects of arms races on

the occurrence of war. Journal of Conflict Resolution 33, 500—529.

Palfrey, T. R. and H. Rosenthal (1985). Voter participation and strategic uncertainty.

American Political Science Review 79(1), 62—78.

35
Powell, R. (1993). Guns, butter, anarchy. American Political Science Review 87, 115—132.

Powell, R. (1999). In the Shadow of Power. Princeton: Princeton University Press.

Powell, R. (2004). Bargaining and learning while fighting. American Journal of Political
Science 48(2), 344—361.

Schultz, K. (1998). Domestic opposition and signaling in international crises. American

Political Science Review 92(4), 829—844.

Selten, R. (1974). Reexamination of the perfectness concept for equilibrium points in

extensive games. International Journal of Game Theory 4, 25—55.

Slantchev, B. (2003, February). The power to hurt: Costly conflict with completely in-
formed states. American Political Science Review 97, 123—133.

Slantchev, B. (2005). Military coercion in interstate crises. forthcoming in the American

Political Science Review, 99:4 .

Wagner, R. H. (2000, July). Bargaining and war. American Journal of Political Science 44,
469—84.

36