See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/284519837

Game Theory and the Law: Ready for Prime Time?

Article in Michigan Law Review · May 1996

DOI: 10.2307/1289973

CITATIONS READS
9 3,307

5 authors, including:

Stephen Walter Salant Douglas G. Baird

University of Maryland, College Park University of Chicago
122 PUBLICATIONS 4,564 CITATIONS 70 PUBLICATIONS 3,145 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Stephen Walter Salant on 15 May 2017.

The user has requested enhancement of the downloaded file.

Review: Game Theory and the Law: Ready for Prime Time?
Reviewed Work(s): Game Theory and the Law by Douglas G. Baird, Robert H. Gertner and
Randal C. Picker
Review by: Stephen W. Salant and Theodore S. Sims
Source: Michigan Law Review, Vol. 94, No. 6, 1996 Survey of Books Relating to the Law
(May, 1996), pp. 1839-1882
Published by: The Michigan Law Review Association
Stable URL: http://www.jstor.org/stable/1289973
Accessed: 15-05-2017 17:12 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

The Michigan Law Review Association is collaborating with JSTOR to digitize, preserve and extend
access to Michigan Law Review

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 GAME THEORY AND THE LAW: READY
FOR PRIME TIME?

Stephen W. Salant*
and Theodore S. Sims**

GAME THEORY AND THE LAW. By Douglas G. Baird, Robert

Gertner, and Randal C. Picker. Cambridge: Harvard Universit
Press. 1994. Pp. xii, 330. $45.

With the advent of imperfect competition theory in the 1930s

there appeared a profusion of oligopoly theories derived from a host
of plausible assumptions, few of them readily testable by empirical
evidence. It sometimes seemed as if every theorist aimed to produce
a model of his own before breakfast each morning, limbering up with
mental calisthenics.

No wonder that distinguished reviewers such as Leonid Hurwicz

and Richard Stone eagerly welcomed the appearance in 1944 of von
Neumann and Morgenstern's great work, The Theory of Games and
Economic Behaviour. At last economists were properly equipped
with powerful and elegant methods of tackling a subject that had be-
come increasingly baroque ....
Surprisingly and embarrassingly and for reasons hard to fathom,
over the next twenty years game theory failed to live up to its promise
for economics.... [E]ven though as early as 1950 Nash had developed
an appealing concept of solution for non-cooperative games, econo-
mists still shied away from applying game theory to strategic eco-
nomic behaviour, its natural home in our discipline....
[T]he analysis of strategic behaviour ... languished mightily until
the mid-1960s, when important papers by Harsanyi and by Selten and
others appeared. By the 1970s this trickle of articles had become a
swiftly moving stream, and by the 1980s a roaring flood that
threatened to engulf the rest of microeconomics ... .

* Professor of Economics, University of Michigan. A.B. 1967, Columbia College; Ph.D.

1973, University of Pennsylvania. - Ed.
** Professor of Law, George Washington University and Visiting Professor of Law, Uni-
versity of Michigan. A.B. 1967, Columbia College; J.D. 1970, University of Chicago; Ph.D.
1995, Massachusetts Institute of Technology. - Ed. The authors are grateful to Linda T.M.
Bui, Alan Gerber, Peter Hammer, Chip Lupu, Eric Talley, and participants in a faculty work-
shop at the University of Michigan Law School for their helpful comments and suggestions.
1. THE NEW PALGRAVE: GAME THEORY xi-xii (John Eatwell et al. eds., 1989) [hereinaf-
ter NEW PALGRAVE: GAME THEORY].

1839

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1840 Michigan Law Review [Vol. 94:1839

INTRODUCTION

The foundations of noncooperative game theory, the formal

study of nonconsensual (or "strategic") interaction among self-
interested rational actors, were laid largely during the middle of this
century.2 Since then, game-theoretic work has threatened to engulf
not only microeconomics, but virtually every branch of social scien-
tific study. The law rests on slightly higher ground, but certainly is
not above the crest. The use of game-theoretic methods to study
legal problems dates at least to the early 1970s,3 and has been grow-
ing - albeit at a more measured pace than in economics proper -
ever since. Almost surely, the differential rate of diffusion of game
theory into law is due partly to its higher entry fee.4 It has only
been with increased numbers of individuals formally trained in both
law and economics, as well as increased lawyer/economist collabo-
ration, that game-theoretic analysis of legal problems has started to
take off.
So it is only possibly surprising that Game Theory and the Law
(hereinafter GTL) marks the first general effort to bring game-
theoretic insights to a specifically legal audience. Even so, GTL
might well encounter differing expectations among different seg-
ments of its potential audience. Some aspects of law and econom-
ics, in some quarters at least, are viewed as resting on undue
simplification, whether in such forms as the core insight of The
Problem of Social Cost,5 or simply the relentless application to legal
problems of the relentlessly economic weltanschauung of such econ-
omists as Gary Becker.6 To be sure, the insights brought to bear on

2. Apart from JOHN VON NEUMANN & OSKAR MORGENSTERN, THEORY OF GAMES AND
ECONOMIC BEHAVIOR (Science Editions 1964) (1944), pivotal founding contributions also
include the mid-century papers by John F. Nash, Jr., including Non-Cooperative Games, 54
ANNALS OF MATHEMATICS 286 (1951) [hereinafter Nash, Games]; The Bargaining Problem,
18 ECONOMETRICA 155 (1950) [hereinafter Nash, The Bargaining Problem]; and Equilibrium
Points in N-Person Games, 36 PROC. NATL. ACAD. SCI. 48 (1950) [hereinafter Nash, N-Per-
son Games]. The slightly later contributions by Reinhart Selten and John Harsanyi include:
John C. Harsanyi, Games with Incomplete Information Played by "Bayesian" Players (pts. 1-
3), 14 MGMT. SCI. 159, 320, 486 (1967-68); Reinhart Selten, Spieltheoretische Behandlung
eines Oligopolmodells mit Nachfragetragheit, 121 ZEITSCHRIFT FOR GESAMTE STAATSSWISEN-
SCHAFT (pts. 1&2), 301, 667 (1965) [hereinafter Selten, Spieltheoretische]; R. Selten, Reexami-
nation of the Perfectness Concept for Equilibrium Points in Extensive Games, 4 INTL. J. GAME
THEORY 25 (1975) [hereinafter Selten, Perfectness Concept].
3. See, e.g., John Prather Brown, Toward an Economic Theory of Liability, 2 J. LEGAL
STUD. 323 (1973).
4. See Ian Ayres, Playing Games with the Law, 42 STAN. L. REV. 1291, 1315-18 (1990)
(reviewing ERIC RASMUSEN, GAMES AND INFORMATION: AN INTRODUCTION TO GAME THE-
ORY (1989)), who suggests that a secondary reason for the slow diffusion of game theory into
legal analysis may be found in its generally less sanguine account of the economic outcomes
of private interaction.
5. See R.H. Coase, The Problem of Social Cost, 3 J.L. & ECON. 1 (1960).
6. See GARY S. BECKER, THE ECONOMIC APPROACH TO HUMAN BEHAVIOR 3-14 (1976);
RICHARD A. POSNER, ECONOMIC ANALYSIS OF LAW (4th ed. 1992).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1841

legal problems have been distinguished economic insights, and they

have richly illuminated our thinking about the law. But it hardly
requires citation to say that their application has often been contro-
versial.7 So, we conjecture at the outset, there are those who will
approach both game theory generally and GTL particularly in a
skeptical frame of mind, vigilant for the possibility that this may
constitute an instance of cutting perhaps too wide a swath through
the law with perhaps too simple, or too simplified, an economic
thought.
We are happy to begin by reporting, then, that such apprehen-
sions are without foundation. GTL decidedly does not bring to the
law a single bright punch-line, distilled from a more ambiguous eco-
nomic account. In part that is in the nature of the beast. But it also
reflects the authors' aspirations, to which they can speak best
themselves:
First, we wanted to introduce the formal tools of modern game theory
to a wide audience using a number of classic legal problems ....
Second, and as important, we wanted to show how modern game the-
ory allows us to sharpen our intuitions and provides us with new ways
of looking at familiar problems. In short, we have tried to write a
book that offers those interested in law a new way of thinking about
legal rules, and a book that shows those interested in game theory a
fertile and largely unexplored domain in which its tools have many
applications.
Much of the analysis in this book makes extensive use of concepts
that have been developed only within the last decade, and we have
not compromised on the rigor that these cutting-edge concepts de-
mand. Nevertheless, we have been able to apply these concepts to the
law without requiring the reader to know calculus, probability theory,
or any other formal mathematical tools beyond simple algebra ....
We depend only on the reader's willingness to think through hard
problems logically and carefully. [pp. xi-xii].
This suggests that GTL will not be found erring on the side of un-
due or convenient simplification. Indeed that is the case. This is a
broad foray, largely unburdened by doctrinaire predispositions, not
only into the range of techniques deployed in contemporary game-
theoretic work, but across the spectrum of law-related problems to
which such techniques, in principle, can be applied.
The authors' prefatory observations nicely frame the issues on
which we wish to dwell. Those are, first, how well chosen are their

7. We offer a selection anyhow. On The Problem of Social Cost, see William J. Baumol,
On Taxation and the Control of Externalities, 62 AM. ECON. REV. 307 (1972); Robert Cooter,
The Cost of Coase, 11 J. LEGAL STUD. 1 (1982); Donald H. Regan, The Problem of Social
Cost Revisited, 15 J.L. & ECON. 427 (1972); R.H. COASE, Notes on the Problem of Social Cost,
in THE FIRM, THE MARKET, AND THE LAW 157-85 (1988). On ECONOMIC ANALYSIS OF LAW,
see Arthur Allen Leff, Economic Analysis of Law: Some Realism About Nominalism, 60 VA.
L. REV. 451 (1974).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 1842 Michigan Law Review [Vol. 94:1839

aspirations, and second, how well realized are the chosen aspira-
tions? Given our venue, we take them up with a readership of law-
yers curious about game theory - rather than the other way
around - in mind. We will, however, say at the beginning that, for
the other segment of GTL's target audience - game theorists in-
terested in applications of their tools to the law - this will be a very
interesting book indeed, one that may well suggest a range of inter-
esting topics on which to work. As for potentially interested law-
yers, we imagine that they fall into one of roughly two groups:
those already familiar with game theory, including those who do
game-theoretic work themselves; and those with just a passing ac-
quaintance, perhaps enough to understand the prisoner's dilemma
and its moral, or possibly no more than enough to know that "stra-
tegic interaction" is a term that with increasing frequency they have
heard. We assume that we can be of service primarily to the latter
group, since those already conversant with game theory and the as-
sociated legal literature will be familiar with much of what is devel-
oped in GTL.
For readers interested in learning a little, or a little more, about
both game theory and its relevance to the law, what might consti-
tute a reasonable set of preliminary questions? Essential elements
would seem clearly to include an introduction to what game theory
is, an introduction to its principal analytic methods and illustrations
of how they can be deployed to illuminate aspects of the law. GTL
aspires to all of this and more. It introduces the reader to the more
elementary game-theoretic techniques.8 Most conspicuously, it
surveys the ways in which the tools of game theory and information
economics can be applied to aspects of the law. The survey takes in
a lot of ground. It extends not only to advanced applications of
game theory and information economics, but to such related topics
as optimal contracting (pp. 109-18), mechanism design (Chapter
Six), and bargaining (Chapters Seven and Eight). Unavoidably,
then, it must introduce the reader to more-advanced techniques,
and that is a third important feature of the book.9 Finally, reflect-
ing its multiplicity of authors and their current research interests,
GTL incorporates detailed treatment of problems on which they
separately have worked.10 By any standard the undertaking is am-
bitious. It is also overdue. Game-theoretic work currently is scat-

8. Pp. 6-78, 159-87; see infra section II.A. For purposes of discussion we adopt the fol-
lowing taxonomy: by "game theory" we mean the underlying mathematical results on which
the discipline rests; by game-theoretic "techniques" or "methods" we have in mind the col-
lection of means, to which that theory gives rise, of "solving" non-cooperative games; and by
game-theoretic "applications" we have in mind the application of those methods to modeling
concrete problems.
9. See, e.g., chapter 3; see infra section II.B.
10. See, e.g., pp. 147-53 (drawing on lan Ayres & Robert Gertner, Strategic Contractual
Inefficiency and the Optimal Choice of Legal Rules, 101 YALE L.J. 729 (1992)); pp. 232-37

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1843

tered throughout the literature, at varying levels of accessibility and

technical detail. There is a need for a general introduction to these
materials.
At the extremes, the organizing thread of GTL - its survey of
applications - could be spun out in one of two quite different ways.
At each stage it could describe for the reader the techniques that
were involved, and then proceed to articulate, above the level of
gritty technical detail, how those techniques might be applied in
particular legal settings, and with what insights and implications-
relegating the details to footnotes, appendices, or references to
other texts.1' This facilitates covering a lot of ground. Alterna-
tively, it could take the time genuinely to school the reader in
game-theoretic methods, developing the survey as a set of "worked-
out" applications.12 That would require a heavier investment by the
reader and almost surely would limit the range of things surveyed.
For the additional investment, however, the reader would end up
with a better grasp of what actually is going on. With the former,
the reader is more likely to end up with a sense of the insights that
others have obtained by applying the techniques of game theory to
the law. Either could produce a very useful book, although they
would be very different books.
In important respects, GTL pursues the former course. It does
not, as the preface immediately makes clear, require the reader to
know - or to learn - any calculus (there is very little and, really,
very little need) (pp. 112-16 & nn.34-38) or probability (of which
there is a lot, none even remotely "theoretical," almost all of it as
elementary as can be) (e.g., pp. 244-67 & nn.6-29). It does, by and
large, develop its account of both the theory and its applications
almost entirely in words. Those constraints effectively foreclose
GTL from offering a genuinely systematic account of the range of
game-theoretic techniques. Nevertheless, GTL does repeatedly ex-
pose the reader to game-theoretic applications in detail. Explana-
tory passages are often interspersed with pages of detailed
verification of whether some set of actions is or is not a solution to
some game. That, to be sure, is an integral feature of what game
theory is about, and it is that to which we take the authors to be
referring when they ask for the "reader's willingness to think
through hard problems logically and carefully" (p. xii). But it is a
form of exposition perhaps better suited to a book devoted to

(drawing on Douglas G. Baird & Randall C. Picker, A Simple Noncooperative Bargaining

Model of Corporate Reorganizations, 20 J. LEGAL STUD. 311 (1991)).
11. One model for this basic approach would be STEVEN SHAVELL, ECONOMIC ANALYSIS
OF ACCIDENT LAW (1987).
12. For comparison, see ROBERT GIBBONS, GAME THEORY FOR APPLIED ECONOMISTS
(1992); ERIC RASMUSEN, GAMES AND INFORMATION: AN INTRODUCTION TO GAME THEORY
(2d ed. 1994), the first edition of which was reviewed in Ayres, supra note 4.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1844 Michigan Law Review [Vol. 94:1839

teaching lawyers how to do game theory than to a book about game

theory and the law. At bottom, however, GTL is about game the-
ory and the law. As such, one might question its allocation of so
much space to the details.
This problem is not, we should immediately add, especially seri-
ous in the introductory chapters of the book. There, the reader is
more carefully introduced to the techniques, and the techniques
themselves are not that difficult to grasp. But GTL also devotes
considerable space to more esoteric matters - those recently de-
veloped "cutting-edge" concepts to which the prefatory remarks re-
fer - developed to deal with problems of "asymmetric"
information.13 Its doing so reflects the authors' considered judg-
ment, with which many would agree, that "[i]ncomplete informa-
tion is the central problem in game theory and the law" (p. 2).
Here, however, GTL courts two more serious difficulties, both
traceable to its concomitant decision to do it (almost) all in words.
First, there inevitably is some compromise, not merely on the rigor
but more importantly on the kind of systematic development that
learning those techniques demands. Modeling uncertainty and in-
formational asymmetry inescapably entails the use of probability,
the very thing that GTL has chosen to eschew. No wealth of detail
is a substitute for the essential analytic foundations of a substantial
fraction of what GTL sets out to do.

Second, to the extent the theory does emerge, it does so with

some important details obscured. Game theory is currently in
state of flux, beset by serious problems that have not been satisfac
torily resolved.14 Chief among them are the extreme sensitivity o
its predictions to aspects of formulations that a modeler canno
hope to specify, and a not insignificant body of evidence suggesting
that even when these aspects are experimentally controlled for, th
theory often fails to predict how people actually do behave. Given
GTL's preoccupation with illustrating applications of game theory
to the law, little space remains for sustained attention to the short
comings of the theory,15 or for the more detailed exposition of th
theory that would form the essential predicate for that.
A book so intent on taking its readers to the frontiers, so willing
to subject them to so much detail, ought perhaps be willing to insis
on just a little additional effort, and to offer in return a more com
plete account of, and more importantly a better perspective on, jus

13. Problems of "asymmetric" information posit the existence of private information

known to some but not to all strategic actors. See infra text accompanying notes 84-93.
14. See infra text accompanying notes 94-95, section II.C.
15. We do not suggest that GTL completely overlooks the issues; only that, by dint of an
overwhelming emphasis on applications, these important questions are dealt with in a frag
mentary fashion and are too substantially submerged. See infra section II.C.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1845

what is being done. For that, we believe, is the important final en-
try in our list of preliminary questions for the seriously interested
reader. One not already acquainted with the theory, its conceptual
difficulties, and its mixed success in controlled experiments, is likely
to come away inadequately informed about the limits to what game
theory may currently and realistically have to offer to the law.
Whether game theory really is ready for prime time, and if so just
which aspects are ready to be aired, are questions GTL does not
systematically explore.
The net of it is this. GTL does not teach lawyers how to do
game theory, but that is something it does not set out to do. GTL
does succeed in offering interested lawyers a nontechnical introduc-
tion to game theory, together with a rich and varied survey of the
ways in which game-theoretic insights can be brought to bear on the
design of legal rules. That is no small accomplishment in itself.
Game-theoretic reasoning focuses on what actors can observe, what
they can infer, and what, in light of that, they do. As such it is
recursive and complex. In developing an account in words, GTL
took on an extraordinarily demanding task. In the chapters in
which the trickier analyses are performed, however, it is possible
that the otherwise uninitiated reader will be able to do little more
than watch. Still, it is interesting to watch. GTL takes up the most
current features of applied game theory and information economic
and sets out to explore their relevance to the law.
In the balance of this review, we will say more about what GTL
actually does. But first we need a methodological preface, since
without that, whatever we might have to say may be unmean
ingfully abstract. So, in Part I, we begin with a brief, nontechnica
survey of the terminology and methodology of noncooperative
game theory. For readers already acquainted with the basics - i
specifically you know what a "subgame perfect equilibrium" is
Part I will tell you little new and should be skipped. If, moreover
you are unacquainted but not (or not yet) interested in the details
feel free in any event to proceed directly to Part II.
There we turn to GTL itself and survey the range of things the
book sets out to do. In lieu of making the catalogue too detailed
we instead focus on one central feature of the exposition. That is
the authors' decision to introduce after a mere eighty pages, and
thereafter to dwell on in detail, games of "incomplete" information
and the solution concept known as the "perfect Bayesian equilib-
rium."l6 Judging by the language from the Preface, we suspect tha
this is an aspect of their endeavors of which the authors are most
proud. But it puts most severely to the test their prior decision to

16. If you do not already know, please do not worry for the moment just what any of
those terms might mean. We will fill you in.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1846 Michigan Law Review [Vol. 94:1839

dispense with mathematical formalities. Together, those decisions

induce a self-inflicted inability to lay out for the reader in adequate
detail either the operation of, or the central difficulty with, the solu-
tion concept that is key to any real grasp of what GTL is doing in a
substantial fraction of the book.
It is not entirely clear what prompted the first of these decisions.
The obvious possibility is that the authors felt it essential to their
accessibility to a wider audience. More specifically, they may have
been motivated by the thought that lawyers simply cannot be ex-
pected to deal with probability, or mathematics at any level of ab-
straction, or even, really, with what ultimately amounts to no more
than simple numerical computations. But it is surely striking that
these three individuals, inhabitants of the community that is literally
the cradle of contemporary law and economics, should have set for
themselves the task they did and then elected to do it all in words.

I. A GAME THEORY PRIMER

What is "game theory"? At a high level of generality, it stu

formally what might be regarded as the social scientific quest
How do, or should, individuals conduct themselves when each
izes that the consequences of his individual acts will depend in
on what other independent actors do?17 How, in other words,
he do best at pursuing his own objectives, whatever they migh
in an interdependent milieu? Game theory formalizes that qu
tion by positing players and endowing them with moves, specif
what information is available to each player at each point at wh
he may be called upon to move, assigning payoffs to the way
players' moves can be combined to play the game, and then inv
gating the kinds of conduct that might plausibly arise.
More specifically, noncooperative game theory - the fiel
most active interest the past twenty-five years - might be def
as follows.18 It is the formal study of conduct by two or more "pl
ers," each of whom must choose what to do at an explicitly enu
ated set of "decision points" or "moves," at each of which it h
specified set of available "actions" from which to choose.19
"payoffs" to each player are determined both by what that p

17. See, e.g., Robert J. Aumann, Game Theory, in NEW PALGRAVE: GAME THE
supra note 1, at 1; John C. Harsanyi, Games with Incomplete Information, 85 AM. E
REV. 291, 292 (1995).
18. In general, noncooperative game theory is concerned with individual conduct,
the object of interest in "cooperative" game theory is with what can be obtained by "gr
or "coalitions." We will use the term "game theory" synonymously with "noncooper
game theory" throughout.
19. Each decision point is called more generally an "information set," which specifi
information available to the player whose move it is to make. See infra text accompa
notes 26 and 58-59.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1847

does and by what the other players do. Each player is assumed to
know the "rules" of the game, including not only all the moves
available to each player and precisely what past choices are observ-
able by each player at each move, but also the payoffs associated
with every possible way in which the players' choices might be com-
bined to "play" the game. What is more, each player is taken to
assume that each other player also knows the rules, usually de-
scribed by saying the structure of the game is "common knowl-
edge."20 So it is the formal study of conduct by actors whose
payoffs are jointly determined by the interaction of their individual
choices, within the confines of a completely specified, commonly
known set of rules.

Once the rules of a game are made concrete - by specifying all

the possible moves, who may make them, and what information is
available and what actions may be taken at each - there will be
different ways in which each player might combine the actions
available to it at its different moves, and by so doing to specify com-
pletely, for that player, one possible way to play the game. Any one
such combination is called a "strategy" for that player.21 The col-
lection of all the strategies open to a given player in a given game
- every possible way that player might conceivably play the game
- is called the player's "strategy set." Any collection of strategies,
formed by choosing one for each player in the game, specifies one
possible play of the entire game, and is called a "strategy profile" of
the game. In a well-defined game, payoffs to each player are speci-
fied for every strategy profile, that is, for every possible way the
elements of the players' strategy sets can be combined to play the
game.
That is all a bit abstract. So, we make the abstractions concrete
using two very simple two-person games. But there are two differ-
ent ways of representing a noncooperative game. To those with just
a passing acquaintance with the subject, the most familiar is proba-
bly the "normal" or "strategic" form, in which the strategies of the
players are arrayed along the margins of a matrix, each cell of
which is defined by one strategy for each of the two players, and
hence corresponds to one strategy profile of the game. The entries
in each cell give the payoffs to each player from the strategy profile
that defines that cell. We begin, however, with the "extensive

20. See infra text accompanying notes 37-38.

21. See, e.g., GIBBONS, supra note 12, at 93; R. DUNCAN LUCE & HOWARD RAIFFA,
GAMES AND DECISIONS: INTRODUCTION AND CRITICAL SURVEY 51-53 (Dover Publications
1989) (1957); MARTIN J. OSBORNE & ARIEL RUBINSTEIN, A COURSE IN GAME THEORY 92-
93 (1994); Eric van Damme, Extensive Form Games, in NEW PALGRAVE: GAME THEORY,
supra note 1, at 139, 140-41. We emphasize here that a "strategy" specifies the actions a
player will take at each point in the game at which she possibly might be called upon to move,
including points in the game that never will be reached. See infra text accompanying note 28.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1848 Michigan Law Review [Vol. 94:1839

form,"22 which provides a more detailed description of a game and

is often the more convenient form to use. Any game can be repre-
sented in extensive form and then reduced to a corresponding and
unique normal form.

A. The Description of Games in Extensive and Normal Form

We base our first example ("Game One") on a situation rou-

tinely encountered by lawyers and their clients. Imagine, as one
possible illustration, the reroofing of a residential building, during
which a series of leaks develop, causing substantial damage to a
resident of the building's topmost floor. The roofer had exclusive
control of the roof, so there is no serious question about liability,
nor is there any doubt about the costs to repair the physical damage
that was sustained. The resident also claims, however, that she was
constructively evicted from her apartment by the repeated water
infiltrations, for which she seeks compensation beyond the cost of
the repairs. The latter claim is a bit out of the ordinary, and the
roofer's Insurer (Player I) is reluctant to accede to it. Even so, the
belief of both the Insurer and the Resident (Player R) is that, on the
evidence, it almost surely would be sustained if the dispute were to
go to court. Suppose the cost of physical repairs is $7,23 that the
claimed damages for constructive eviction are another $6, and that
the resident may choose to litigate in response to the compensation
she is offered. The cost of litigation will be $4 to each side.
How will the Insurer and the Resident proceed?24 If you accept
the rules as they are given, you will probably have an intuition. Our
objective, however, is to formalize the game, after which we will see
if it has a "solution" that corresponds with your intuition. Suppose,
to be both simple and specific, we imagine that the Insurer can
make one of two offers: "Full" compensation ($13), or "Partial"
compensation ($10).25 Confronted by either of these offers, the
Resident may choose either to "settle" or go to "court." Formu-
lated as a noncooperative game, Player I has a single move, at
which it has two "actions" from which to choose. Player R, in con-

22. See generally LUCE & RAIFFA, supra note 21, at 39-55; H.W. Kuhn, Extensive Games
and the Problem of Information, in 2 CONTRIBUTIONS TO THE THEORY OF GAMES 193 (1953);
JEAN TIROLE, THE THEORY OF INDUSTRIAL ORGANIZATION 423-26 (1988); van Damme,
supra note 21, at 139-44.
23. We have in mind $7000, but for simplicity we suppress the zeroes.
24. We will refer, interchangeably, to "the Resident," "Player R," and "R"; and likewise
for "the Insurer."

25. We have, in the interests of simplicity, restricted ourselves to offers of compensation

that are integers. Our specification of the offer of Partial compensation is as the smallest
integral offer that, in the solution to the game, does not provoke the Resident to go to court.
See infra note 52 and accompanying text.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1849

trast, has two possible moves, depending on the Insurer's offer, at

either of which she must choose to settle or go to court.

1. The Extensive Form

The "extensive form" representation of our simple game con

sists of the "game tree" in Figure I-A, in which a move by eith
player is denoted by a solid circled "choice node." To indic
whose move it is, we have located a circled I or R beside each s
node. The arrows emanating from each choice node denote the
tions open to the player whose move it is to make. Player I mo
first and may choose F or P. Either choice leads to a move
Player R. The choice by Player I of offering Partial compensat
for example, leads to a node at which Player R chooses whether
settle or go to court. As we have described it, Player R might g
court even if offered Full compensation, depicted in the tree by
choices c' and s'. There are four possible sequences of moves
and R that can be formed from the choices available to them: P-
P-c, F-s, and F-c'. In Figure I-A, the termination of each result
"branch" of the tree - denoted by an arrow that does not lead
move by another player, but to a "terminal node" instead - cor
sponds to exactly one of those four sequences, and to each
quence associates a set of payoffs, here just two, one for I and
for R.

FIGURE I-A
EXTENSIVE FORM OF GAME ONE

s A c s' / \ c'

(-10,10) (-17,9) (-13,13) (-17,9)

Payoffs: (I, R)

Thus, if Player I offers Full Compensation and Player R settles, the

Insurer pays $13 to the Resident, and their respective payoffs are
denoted by (-13, 13); if instead R chooses to go to court, she will
recover $13, but each party must pay $4 in legal fees, and the
payoffs will be (-17, 9). If, on the other hand, I offers Partial com-
pensation, the payoffs will be (-10, 10) or (-17, 9), depending on
whether the Resident settles or goes to court. (We remind the
reader which payoff belongs to which player at the bottom left of
the game tree.)

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1850 Michigan Law Review [Vol. 94:1839

Game One begins with the Insurer's move; the Resident, when
called upon to move, can observe what the Insurer has actually
done. So, each Player, at each point at which he is called upon to
move, knows everything that has thus far happened in the game.26
As a matter of terminology, a game that has the property that each
player at each move has observed the entire history of the play to
date is called a game of "perfect" information. Checkers and chess
are illustrations.

2. Reduction to Normal (or Strategic) Form

A strategy for a player, as we have noted, specifies one way in
which that player might combine the actions available at its differ-
ent moves and hence one way for it to play the game.27 For Player I
that specification is simple. It has but a single move, at which there
are but two actions it might choose. So a complete list of Player I's
strategies consists simply of its choices (F or P). For Player R, on
the other hand, the matter is a little more complex. Since a strategy
is a complete specification of one way in which a player might play
the entire game, it must prescribe a choice at every move that
player might be called upon to make.28 Here, there are two situa-
tions in which Player R might be called upon to act, and a strategy
must specify what she will do at each. In other words, a strategy for
the Resident must specify whether she will go to court or settle, and
must do so separately for the eventuality that the Insurer offers
either P or F. There are thus four possible combinations, and hence
four strategies available to Player R. For example, the combination
(c, s') denotes the strategy in which R goes to court if I offers Par-
tial compensation, but settles if I offers F.
Note now that Game One has a finite number of players, each
of whom has a finite number of strategies from which to choose.
Consequently, there exists a finite number of ways in which their
strategies may be combined to form a strategy profile for the entire
game. Observe also that, even though each of Player R's strategies
specifies what she will do for either choice by Player I, each of
Player I's strategies is to do just one of those two things. Hence,
each strategy profile corresponds uniquely to a path through the
game tree leading to just one of its four terminal nodes. For exam-

26. In the extensive form, this property generally obtains whenever each player, at each
move, knows exactly where he is in the game tree. As a technical matter it is depicted by the
fact that each decision point contains exactly one choice node and is therefore called a "sin-
gleton information set." See, e.g., GIBBONS, supra note 12, at 115-22; see supra note 19; infra
text accompanying notes 58-59.
27. Strictly speaking, a specification of one way to play the game is a "pure" strategy. A
"mixed" strategy, in contrast, involves play that is randomized among more than one pure
strategy, according to some probabilistic rule. See, e.g., GIBBONS, supra note 12, at 30-31.
28. See supra text accompanying note 21. We will say more about just why in a moment.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 May 1996] Game Theory 1851

pie, if I's strategy is F and the Resident's strategy is (c, s') - she
takes an offer of Partial compensation to court but settles if offered
F- the play of the game will be F-s'. The strategy profile [F, (c, s')]
leads uniquely down the path along which the Insurer offers F and
R plays s' and to payoffs of (-13, 13). More generally, as long as
there are no "chance moves" reflecting uncertainty, every strategy
profile leads to one, and only one, terminal node of the game tree.
So the information captured by the extensive representation of Fig-
ure I-A may also be expressed by associating with each strategy
profile the payoffs at the terminal node of the extensive form to
which it corresponds. That is characteristic of the extensive form,
however complicated it might become,29 of every finite game (not
entailing chance moves) in which the players have discrete and fi-
nite strategy sets.30 When, moreover, as in Game One, there are
only two players both of whom have discrete and finite strategy
sets, the extensive representation may be reduced to matrix form.
That is the "normal form" of the same game.31
In our rudimentary example, the transformation from extensive
to normal form is also relatively simple. Player I's entire strategy
set - offer Partial or Full compensation - is listed to the left of
the entries in the matrix, each possibility corresponding to one row.
Player R's strategy set - all four possible combinations of how she
might respond, respectively, to offers of F or P - is listed above the
entries in the matrix, with each strategy corresponding to a single
column. So the normal-form representation of a two-player game is

29. It can rapidly become quite complicated. See, e.g., Thomas C. Schelling, What is
Game Theory?, in CONTEMPORARY POLITICAL ANALYSIS 212, 224-32 (James C. Charles-
worth ed., 1967), reprinted in THOMAS C. SCHELLING, CHOICE AND CONSEQUENCE 213, 226-
34 (1984). A dramatic example, described in Harold Kuhn, Introduction to Montmort, in
PRECURSORS IN MATHEMATICAL ECONOMICS: AN ANTHOLOGY - (W.J. Baumol & S. Gold-
field eds., 1968), reprinted in SCARCE WORKS IN POLITICAL ECONOMY NO. 19, is a 1713 anal-
ysis by James Waldegrave of a two-person card game in which each player was dealt a single
card, after which each player could exchange it for another card just once. The normal form
of the game is a square matrix with 213 rows and 213 columns, or a total of 226 pure strategy
profiles. Waldegrave's analysis was the first of a noncooperative game, the first example of a
Nash equilibrium, and the first example of a unique Nash equilibrium in "mixed" strategies.
Id.

30. That characteristic is described formally in LUCE & RAIFFA, supra note 21, at 51-53;
van Damme, supra note 21, at 139. In DAVID M. KREPS, A COURSE IN MICROECONOMIC
THEORY 363, 365-66 (1990), it suggestively is called an "arboresence." But it is easier just to
think about the nearest tree. Start with any leaf: there is a unique path connecting it to the
ground.
31. The normal form, in fact, is more general. When the extensive form involves chance
moves reflecting uncertainty, each strategy profile results in the various terminal nodes being
reached with known probabilities, and what is reported as the entry in the normal form cor-
responding to a given strategy profile is the "expected payoff," see infra note 101, to each
player from the lottery induced by that strategy profile. In addition, the normal form is
confined neither to games with two players nor to those with finite strategy sets, see, e.g.,
GIBBONS, supra note 12, at 3-4, although its simple visual representation as a matrix usually
is.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1852 Michigan Law Review [Vol. 94:1839

a matrix whose dimensions correspond to the dimensions of their

respective strategy sets - in this particular instance two by four.
The entries in each cell are the payoffs to each player for each strat-
egy profile, as given by the payoffs at the terminal node of the ex-
tensive form to which that strategy profile uniquely corresponds.32
In the extensive representation of Figure I-A, there are, as we al-
ready noted, four possible sequences of moves by I and R. In the
normal-form matrix representation of Figure I-B, those four possi-
ble sequences correspond to eight possible strategy profiles of the
game.33 More generally, to any extensive form of any non-
cooperative game, there corresponds a unique normal form.34 A
virtue of the matrix representation in games with two players is that
it provides a simple means of visual verification, separately for each
player, of how each of their pure35 strategies will fare.

FIGURE I-B
NORMAL FORM OF GAME ONE

Player R's Strategy

(s, s' (s, c') (c, s') (c, c')

p (-10, 10)* (-10, 10)** (-17,9) (-17,9)

Player I's
Strategy

F (-13, 13) (-17,9) (-13, 13)*** (-17,9)

Payoffs: (I, R)

In writing Game One, in both extensive and normal form, it is

assumed that each player knows the structure of the game, includ-
ing the payoffs to each player for each strategy profile of the
game.36 It is also assumed that each player knows that the other
players know it, that each knows that the others know that they
know it, and to which is usually added "and so on ad infinitum."37

32. As at the terminal nodes of the extensive form depicted in Figure I-A, the entries (I,
R) in the cells of Figure I-B give the payoffs to (respectively) Player I and Player R.
33. Since each of the Resident's four strategies specifies what she will do for each possible
choice by the Insurer, each can form a constituent part of two different strategy profiles,
leading to one of two different terminal nodes, depending on the Insurer's actual choice.
34. The normal form may, however, correspond to several different extensive representa-
tions. See, e.g., DAVID M. KREPS, GAME THEORY AND ECONOMIC MODELLING 24-25 (1990).
35. See supra note 27.
36. See, e.g., LUCE & RAIFFA, supra note 21, at 49.
37. See, e.g., Robert J. Aumann, Agreeing to Disagree, 4 ANNALS OF STAT. 1236 (1976);
John Geanakoplos, Common Knowledge, J. ECON. PERSP., Fall 1992, at 53; see also DREW
FUDENBERG & JEAN TIROLE, GAME THEORY 4 (1991); GIBBONS, supra note 12, at 7.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1853

This infinitely recursive awareness - sometimes called "ultrara-

tionality"38 - is usually denoted by saying that the structure of the
game is "common knowledge." When, as in Game One, what is
common knowledge is not subject to uncertainty about the nature
of the other players or their payoffs, the "information structure" of
the game is said to be "complete." As we already have pointed out,
a game in which each player at each move also knows everything
that has previously happened in the game is called a game of perfect
information. Game One, then, is a game of complete and perfect
information.

B. Solution Concepts and Solutions

What we are most interested in, however, is how we might rea-
sonably expect the Insurer and the Resident to behave, and how the
"game" represented by Figure I is likely to be "played." It is to
how we should go about fashioning answers to that question - to
determining what we should reasonably regard as a (or the) "solu-
tion" to a game - that the most basic research in game theory is
devoted. The fundamental postulate that underlies the inquiry is
that each player in the game is "rational," in the sense that his ob-
jective is to maximize his payoff, and that it is common knowledge
that every other player has a corresponding goal.39 In the context
of the normal form, which lists every possible strategy profile of the
game, the search for a "solution" amounts to winnowing that list by
eliminating strategy profiles that rational players are unlikely, or
less likely, to play. A "solution concept" is simply a criterion that is
more or less useful for that task.

1. Rational Players and Strictly Dominated Strategies

Some games have strategies that a player would never rationally
play. Game One is such a game. Keep in mind that Player R's
payoff is the second entry in each cell in Figure I-B, and look at the
rightmost column of the matrix, in which R's strategy is to take
either offer by Player I to court - i.e., the strategy (c, c'). Since, by
the rules of the game, she always wins but must always pay legal
fees, R's payoff is always 9. Compare that to the leftmost column,
in which R's strategy (s, s') is to settle for any offer she observes.
Her payoff will be 10 or 13, depending on whether the Insurer of-
fers P or F. Irrespective of what the Insurer might choose to do,
then, the Resident can do strictly better by always settling than by
always going to court. That property is described by saying that, for

38. See Howard Raiffa, Game Theory at the University of Michigan, 1948-1952, in TO-
WARD A HISTORY OF GAME THEORY 165, 175 (E. Roy Weintraub ed., 1992).
39. See, e.g., LUCE & RAIFFA, supra note 21, at 50; GIBBONS, supra note 12, at 7.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1854 Michigan Law Review [Vol. 94:1839

Player R, the strategy (c, c') is "strictly dominated" by the strategy

(s, s'). Player I can therefore eliminate the strategy (c, c') as one
that Player R might play, and thus immediately rule out two of the
eight strategy profiles as possible solutions to the game.
There are games in which each player may have strictly domi-
nated strategies. Occasionally the elimination of all the dominated
strategies of a player leaves that player with just one remaining
strategy - denoted his "strictly dominant" strategy. If every player
has a dominant strategy, there is a best way for every player to pro-
ceed, irrespective of what the other players do. When that does
occur, the "elimination of strictly dominated strategies" defines the
only strategies rational players would play, and identifies the result-
ing strategy profile as the solution to the game.40 But many games
of interest cannot be solved that way, and Game One is a case in
point. Only the strategy (c, c') is ruled out by elimination of domi-
nated strategies, still leaving us with six strategy profiles as possible
solutions to the game. To make further progress other solution
concepts are required.

2. Nash Equilibrium
We proceed then, to the best-known and most fundamental so-
lution concept, the "Nash equilibrium," one of John Nash's several
founding contributions to the theory of games.41 Turn to the north-
west cell of Figure I-B, containing the payoff "(-10, 10)" and de-
noted by "*." That cell corresponds to the strategy profile
[P, (s, s')], in which the Insurer offers Partial compensation, and the
Resident's strategy is to settle whether the Insurer offers P or F.
Keep in mind that the Insurer's payoff is the first of the two num-
bers in the cell. Fix Player R's strategy at (s, s') - that is, restrict
your attention to the first column of the matrix - and ask whether
Player I can do better than play P. It can't. Its only alternative is F,
in which event, by paying 13 rather than 10 to the Resident, its pay-
off declines to -13. Now fix I's strategy at P - restrict your atten-
tion to the first row - and observe that the Resident can also do no
better than play (s, s'). Given I's having offered P, her payoff
changes only if she chooses a strategy - (c, c') or (c, s') - in which
she takes a Partial offer to court. If she does, however, her recov-
ery of 13 will be reduced by litigation costs of 4, and she will end up
with 9, less than if she had settled for the Partial offer. Given that I
chooses P, R can do no better than by playing (s, s'); given that R

40. The "Prisoner's Dilemma" is one such game. See infra note 65 and accompanying
text. Sometimes a strategy that is not strictly dominated will become dominated after the
elimination of a dominated strategy of some other player. That is called the "iterated elimi-
nation of strictly dominated strategies." See, e.g., GIBBONS, supra note 12, at 6-7.
41. Nash, Games, supra note 2, at 289-95; Nash, N-Person Games, supra note 2, at 48-49.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1855

plays (s, s'), I does best by choosing P. That, by definition, is a

Nash equilibrium. It is the tie that binds contemporary non-
cooperative game theory.
More generally, a (pure strategy) Nash equilibrium is a strategy
profile with the property that, given the strategies of all the other
players, no player can do strictly better by unilaterally choosing a
different strategy than he has.42 In the shorthand commonly em-
ployed, if the strategy choice of each player in the game is a "best
response" to the strategies of all the others, that strategy profile is a
Nash equilibrium of the game. It is often regarded as a minimal
requirement of equilibrium in a noncooperative game, since, if a
strategy profile were not a Nash equilibrium, at least one player
could unilaterally do better by altering his move. There is at least
one Nash equilibrium of every finite noncooperative game.43
Many games have more. That property is exhibited by Game
One. The other cells denoted by asterisks are also Nash equilibria
of the game. One - the profile [P, (s, c')] (denoted by "**"), in
which R's strategy is to take an offer of Full compensation to court
- differs from that already identified only in terms of the payoff to
Player I if it should deviate and offer F instead; but I's best re-
sponse is still to offer P.44 The third - [F, (c, s')] (denoted by
"***"), in which the Resident settles for F but takes P to court,
while Player I offers F - is a more problematic equilibrium, to
which we shall return.45 For the moment, however, note three
things. First, use of the Nash equilibrium has ruled out five of eight
possible strategy profiles as solutions to Game One, as contrasted
with the dominance argument, which, of those five profiles, elimi-
nated only two.46 In that sense, the Nash equilibrium is a more
powerful solution concept. What follows from the fact that the
Nash equilibrium has ruled out only five of eight strategy profiles,

42. See, e.g., GIBBONS, supra note 12, at 8-9; KREPS, supra note 34, at 28.
43. See Nash, N-Person Games, supra note 2, at 48. What Nash actually proved was that
every finite has at least one Nash equilibrium, although it may involve mixed strategies. See
supra note 27. By "finite" we mean that the game has a finite number of players, each with a
finite strategy set. See, e.g., GIBBONS, supra note 12, at 45-47.
44. Each Nash equilibrium identified in the normal form can also be found in the exten-
sive form, as can be seen from Figure I-A. To take as an example the Nash equilibrium
denoted by "*," given the Resident's strategy of settling for either F or P, a comparison of
the payoffs to the Insurer at the first and third terminal nodes shows that the Insurer does
best to offer P. Given that choice, the Resident can do no better by any strategy combination
other than (s, s) - though she can do as well by playing (s, c'). In general, we will, except
where otherwise necessary, identify Nash equilibria in the normal form of the games we use
as illustrations.

45. See infra text accompanying notes 48-50.

46. If each player has a strictly dominant strategy for any noncooperative game, that
strategy profile must obviously form a Nash equilibrium of the game. More generally, every
Nash equilibrium strategy profile will survive the iterated elimination of strictly dominated
strategies. See supra note 40; GIBBONS, supra note 12, at 12.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 1856 Michigan Law Review [Vol. 94:1839

however, is that it may not be powerful enough. That unfortunately

pervasive fact undermines the predictive power of Nash equilibrium
solution(s) to a game. The set of Nash equilibria in our example
implies that the dispute will always be settled, but offers no predic-
tion about whether Full or Partial compensation will be paid. More
generally, multiple equilibria can blunt, often seriously, the extent
to which theory offers testable restrictions on how people actually
do behave.

3. Incredible Threats and Subgame Perfection (Reinhart Selten)

The problem of multiple equilibria is among the most serious
afflictions of game theory. It has touched off intensive efforts to
"refine" the notion of a Nash equilibrium into more powerful solu-
tion concepts, capable of eliminating "less plausible" elements from
a set of equilibria. The most widely accepted of these refinements
is the subgame perfect Nash equilibrium. Subgame perfection has
the happy property of eliminating all but a single strategy profile as
a solution to our illustrative game. It yields a unique solution to
Game One.
For reasons that eventually will become apparent, it is with th
strategic representation of Figure I-B to which we first turn. In t
of its three Nash equilibria (those denoted "*" and "**"), R's str
egy is to settle whenever I offers P, and I's strategy is to offer
The intuition for these equilibria is straightforward. Given
strategy of settling for P, I does best to offer P. Given I's strate
of offering P, R does best to settle, since by going to court, an
additional compensation is more than eaten up by legal fees. It
the third Nash equilibrium (denoted "***") on which we wish
dwell. There R threatens to go to court if I offers P, and I respon
by offering F. Given R's threat of going to court for less, offering F
is the only route by which I can avoid the legal fees; given t
strategy choice by I, R does as well to threaten litigation as anything
else. A key difficulty with the [F, (c, s')] equilibrium cannot be d
tected in the normal form (Figure I-B).47

47. As we explain immediately below, the [F, (c, s')] strategy profile is not "subga
perfect," a defect that can be identified only in the extensive form but which renders it u
likely that the strategy profile actually will be played. There is a distinct reason why t
profile might not be played. It requires R's playing the strategy (c, s') when she ha
alternative (s, s') that provides her at least as high a payoff no matter what Player I does a
a strictly higher payoff if Player I chooses P. For that reason the strategy (c, s') is said to
"weakly dominated," and the Nash equilibrium profile [F, (c, s')] to be "supported b
weakly dominated strategy." The latter is a defect that can be detected in either the nor
or the extensive form. For an example of an equilibrium profile that is supported b
weakly dominated strategy but is nevertheless "subgame perfect," see the equilibrium prof
denoted (***) of Game Two below. See infra notes 62-69 and accompanying text. Fo
useful discussion of what "defects" can be identified in both the normal and the extensive
form and which only in the more detailed extensive form, see van Damme, supra note 21, at
143.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1857

The problem is that R's "threat" is not "credible," as can readily

be seen in the extensive representation of Figure I-A.48 Note, first,
that R's threat is "off the equilibrium path" of play: she threatens
to litigate a Partial offer, while I offers Full compensation, so in the
[F, (c, s')] equilibrium R is never called upon to carry out the threat.
But if Player I were to offer only Partial compensation, R in fact
would not do best by acting on the threat: because of legal fees, the
payoff from taking a Partial offer to court would be less than the
payoff from settling instead. Assuming (as we do) that R is ra-
tional, she would not, if actually presented with the opportunity,
carry out the threat, something that, by the assumption of common
knowledge,49 Player I can see as plainly as can we.
But what, more generally, defines an incredible threat? It be-
gins with the notion of a (proper) subgame, defined loosely as a
subset of the game tree, beginning at a choice node (other than the
initial node), that forms a self-contained game from that point for-
ward in the tree.50 In isolation, a subgame has the same strategic
properties as any other game. So it seems natural to insist that con-
duct in every subgame comport with the postulate of rationality to
the same extent as in any other game. That, in turn, implies that,
given the strategies of the other players in every subgame, each
player should play his best response. A strategy entails an incredi-
ble threat when it calls for conduct that does not induce a Nash
equilibrium in some subgame. Such conduct is not rational in th
subgame, since by definition the threat (if carried out) would no
the player's best response.
The refinement, due to Reinhart Selten,51 that each playe
conduct form a Nash equilibrium, not only for the game in its e
tirety, but also in every proper subgame - that it be "subgame p
fect" - effectively rules out equilibria of games of perf
information that rest on incredible threats. In Game One, the se
quence of moves following I's choice of P forms a proper subga
Subgame perfection then requires that, in both that subgame a
the game as a whole, R choose her "best response," and that
sponse is to settle. Consequently the Nash equilibrium of the ori
nal game in which R's strategy is (c, s') - she settles if offered F

48. On threats generally, the best introduction is perhaps still to be found in THOMAS
SCHELLING, THE STRATEGY OF CONFLICT 35-52 (1980).
49. See supra text accompanying notes 37-38.
50. A "subgame" is a subset of the game tree that begins at a "singleton" information
-that is, an information set that contains a single choice node. See supra note 26.
subset, to be a subgame, must (1) include the entire balance of the game from that info
tion set to all terminal nodes that follow it and (2) contain all information sets that follow
in their entirety. See, e.g., GIBBONS, supra note 12, at 122-24. The adjective "proper" si
means a "subgame" other than the entire game itself.
51. See Selten, Spieltheoretische, supra note 2.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1858 Michigan Law Review [Vol. 94:1839

compensation but threatens to go to court if she is not - is not a

Nash equilibrium in the subgame and is not a subgame perfect equi-
librium. The same is true of the [P, (s, c')] Nash equilibrium (de-
noted "**") of the original game, in which R threatens to litigate an
offer of Full (but not Partial) compensation; going to court when
offered Full compensation self-evidently is not her best response.
In contrast, the equilibrium ("*") in which R's strategy is to set-
tle if offered either F or P and I's strategy is to offer P is a Nash
equilibrium in the game and in every proper subgame. It is the only
Nash equilibrium of Game One that is. Now, finally, we have ar-
rived at a solution - a unique subgame perfect Nash equilibrium
- to Game One. We suspect, moreover, that the implied predic-
tion comports with your intuition: The Insurer's equilibrium offer
is just enough to deter the resident from going to court, and the
Resident's (possibly reluctant) equilibrium strategy is to settle.52
Is this happy outcome - Game One has a subgame perfect
equilibrium (in pure strategies) and the equilibrium is unique -
more than just a fluke? Just how powerful is the refinement of sub-
game perfection? In the class of games - extensive games of per-
fect information - to which our game belongs, it works quite
wonderful results. At minimum, every game of perfect information
having different payoffs at each terminal node has a unique, pure
strategy subgame perfect equilibrium.53

C. Games of Imperfect Information and the Limits of

Subgame Perfection

Subgame perfection is, however, not the Holy Grail. While es-
pecially effective at solving games of complete and perfect informa-
tion, it does nothing to rule out multiple equilibria - including
implausible equilibria - in many other games. That has been a
major source of difficulty for noncooperative game theory, a diffi-
culty reflected in what ultimately strikes us as most problematic
about GTL. So we close by turning to games of "imperfect infor-
mation," and to a simple illustration of the difficulties such games
can entail.

52. It also would correspond with intuition that, if the litigation cost was only 2, the
unique subgame perfect equilibrium strategy profile would be for the Resident to settle for a
Full offer but take a Partial offer to court, and for the Insurer to offer F.
53. See, e.g., van Damme, supra note 21, at 139, 141. The result is perhaps too wonderful,
in the sense that some games of perfect information, including several well-known examples
involving repeated but finite interactions, have unique subgame perfect equilibria that do not
seem to be the evident way to play the game. See, e.g., pp. 163-65; KREPS, supra note 34, at
77-82; Richard D. McKelvey & Thomas R. Palfrey, An Experimental Study of the Centipede
Game, 60 ECONOMETRICA 803 (1992); Robert W. Rosenthal, Games of Perfect Information,
Predatory Pricing and the Chain-Store Paradox, 25 J. ECON. THEORY 92 (1981).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1859

In a game of imperfect information, at least one player must

make at least one move without having observed what some other
player has done.54 We can illustrate how that feature of a game is
formalized using a second game ("Game Two"), obtained by a sim-
ple modification to Game One. Suppose that, in lieu of cash as
Partial compensation, the Insurer offers to repair the physical dam-
age using a contractor of its choosing and to pay the Resident $4 in
cash as well. That offer is worth $11 to the Resident, if the Insurer
selects the same sort of "High quality" contractor to carry out the
repairs (at a cost of $7) that the Resident would have hired herself.
It introduces, however, the possibility that the Insurer might try to
save money by using a "Lower" quality contractor, in which event
the offer will be worth less.

Our objectives here are illustrative and limited, so we single out

a rather trivial alternative to the High quality choice.55 Suppose
that the selection of a modestly Lower quality contractor will cost
the Insurer only $6, that the Resident cannot distinguish between
High and Low quality contractors ex ante, and cannot cheaply ver-
ify the quality of the work ex post.56 Then, when the Insurer offers
Partial compensation, consisting of $4 cash and the repairs, it may
either be using a High-quality contractor (P-H), at a total cost of
$11 to the Insurer and worth that to the Resident if the Resident
accepts; or using a Low quality contractor (P-L), worth at most $10
to the Resident if she accepts. In Game Two Player I has three
moves, P-L, P-H and F - not just two as in Game One - each of
which we depict in extensive form in Figure II-A below. For each
possible choice by the Insurer, the Resident again must choose
either to settle or go to court. The payoffs are (-10, 10) if the Resi-
dent settles for the P-L offer; (-11, 11) if she settles for the P-H
offer; and (as before) (-17, 9) if in either event she goes to court.57
In Game Two, however, the Resident cannot distinguish among
all three of the Insurer's moves. She can observe whether the In-
surer offers Full or Partial compensation. But when she is co
fronted by a Partial offer, she cannot observe whether it involv
High or Low quality work. That "imperfection" in Player R's info
mation is captured by combining her choice nodes following the P
L and P-H moves into one "information set," defined by the pro
erty that the Resident does not know which of the two choice nod

54. See, e.g., GIBBONS, supra note 12, at 118-22.

55. Our reasons for this choice will become apparent. See infra text accompanying not
67-69 and 104-05.

56. The lower quality work might, for example, affect longevity, and might be more costly
to the Resident than the $1 difference to verify on completion of the work.
57. The payoffs when the Insurer offers Full compensation are identical to the corre-
sponding payoffs in Game One. See supra Figure I-B.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1860 Michigan Law Review [Vol. 94:1839

she is actually at,58 and depicted by connecting those nodes (as in

Figure II-A) by a dotted line.

FIGURE II-A
EXTENSIVE FORM OF GAME Two

O_

s /\c s /\c s'\c'

(-10,10) (-17,9) (-11,11) (-17,9) (-13,13) (-17,9)

Payoffs: (I, R)

More generally, an information set is used formally

what each player can observe when it is that player's tur
When all preceding actions are observable - as when P
fers Full compensation in Game Two - the Resident's in
set reduces to the single choice node (called a "singleton
tion set) to which the immediately preceding action lea
clusion, in contrast, of both the P-L and P-H offers in
information set specifies both that the Insurer can make
Partial offers and that the Resident cannot distinguish b
two. That is how a game of imperfect information is d
extensive form.
The net effect is that although Player I has three possible moves,
there are still only two possibilities for which each of Player R's
strategies must specify what she will do. In Game Two, she has,
accordingly, the same four possible strategies as in Game One. The
normal form of Game Two is thus a three-by-four matrix.

58. To satisfy this requirement, all choice nodes in any information set must be assigned
to the same player and must offer that player the same possible actions at each choice node
that it contains. See, e.g., GIBBONS, supra note 12, at 119-20; KREPS, supra note 30, at 367-69.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1861

FIGURE II-B
NORMAL FORM OF GAME Two
Player R's Strategy

(s, s') (, (c') (c,c')

P-L (-10, 10)* (-10, 10)** (-17,9) (-17,9)

Player I's (-11, 11) (-11, ) , 11) (-17,9) (-17,9)

Strategy

F (-13, 13) (-17, 9) (-13, 13)*** (-17, 9)

Payoffs: (I, R)

Given the payoffs to the P-H and P-L offers in Gam

new Nash equilibrium strategy profiles have been introd
indicated by the asterisks in Figure II-B, the same thre
profiles remain Nash equilibria of the game.60 As in
moreover, the strategy profile [P-L, (s, s')] is a subg
equilibrium; conversely, also as in Game One, the strat
[P-L, (s, c')] - in which the Resident (implausibly) th
litigate an offer of Full compensation - is a Nash bu
game perfect equilibrium.61
Game Two, however, in contrast with Game One, do
a unique subgame perfect equilibrium. The problem, ag
the [F, (c, s')] equilibrium, in which the Resident threa
gate either Partial offer and the Insurer offers F. That
to the suspect equilibrium that did not satisfy the req
subgame perfection and was eliminated as a solutio
One.62 In Game Two the Resident's threat to litigate ei
offer remains implausible. But, in contrast with Game
game perfection cannot be used to eliminate the Nash
supported by that threat. As a practical matter, a "gam
gan with the Partial offers in Game Two would not ha
starting move. By the same token, the portion of the g
Figure II-A beginning with those moves is not, as a tec

59. At an intuitive level, the reason should be clear. In the face of either Pa
Resident does best to settle; so the Insurer always does better by offering
Hence, P-H - the newly introduced middle row to the normal form matrix
can never be part of a Nash equilibrium strategy profile of Game Two.
60. Compare supra Figure I-B. The payoffs to the strategy profiles in
which the Insurer makes the P-L offer correspond to those in Figure I-B in w
simply offers P.
61. See supra text accompanying notes 51-52; Figure I-B.
62. See supra text accompanying note 52.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 1862 Michigan Law Review [Vol. 94:1839

ter, a (proper) subgame of Game Two.63 In all events, however im-

plausible her threat to litigate either Partial offer, the [F, (c, s')]
profile remains a "subgame perfect" equilibrium of Game Two.64
So we are left with multiple equilibria in this simple imperfect infor-
mation game and with no equilibrium prediction about whether
Full or Partial compensation will be paid.
Multiple equilibria are not inevitable in games of imperfect in-
formation. For example, simultaneous-move games are games of
imperfect information, and the best-known of them, the "Prisoner's
Dilemma," has a unique solution.65 At the same time, many such
games do have multiple equilibria, none of them inherently implau-
sible or, indeed, superior to any of the others.66

63. It does not, in particular, satisfy the requirement that a subgame begin at a "single-
ton" information set. See supra note 50.
64. But see infra notes 68-69, 104 and accompanying text.
65. A sequential, two-person game in which the second player cannot observe the action
of the first is conceptually equivalent to a game in which the two move simultaneously. See,
e.g., KREPS, supra note 34, at 16-18; KREPS, supra note 30, at 371-72. In the Prisoner's Di-
lemma, each player has two choices, "cooperate" or "defect." The extensive form is given in
Figure III-A.
FIGURE III-A

c/ \ D

V dN\ Y/ d

(3,3) (0,5) (5,0) (1,1)

Payoffs: (1,2)
Since each player has but a single move with just tw
two matrix.

FIGURE III-B
THE PRISONER'S DILEMMA

Player 2
c d

C (3,3) (0,5)
Player 1

D (5, 0) (1,1)

Payoffs: (1, 2)
The reader easily can verify that, although the jointly optimal strategy is to cooperate (C, c),
defecting is a strictly dominant strategy for each player individually, and (D, d) is the unique
Nash equilibrium of the game.
66. This is true, for example, of "coordination games," the usual example of which is the
problem of where each of two friends, having agreed to meet in New York City but both

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1863

The problem of multiple equilibria is generally quite serious,

and the trivial introduction of imperfect information into Game
Two has brought us to the edge of a morass.67 In Game Two, how-
ever, there is a way out that is not especially difficult to grasp. Note
that, regardless of which Partial offer the Resident thinks the In-
surer has actually made, she always does better to settle. If, then,
she acts rationally when she observes a partial offer,68 it seems plau-
sible that she will settle. Hence, while technically subgame perfect,
it is implausible that the [F, (c, s')] strategy profile is an equilibrium
solution to Game Two. One refinement that formalizes this more
stringent notion of plausibility is the "perfect Bayesian equilib-
rium."69 In the present context it rules out solutions involving irra-
tional conduct by the Resident at the Partial offer information set,
and produces a unique solution to Game Two. But the present con-
text was constructed so that the perfect Bayesian equilibrium not
only eliminated an implausible equilibrium but also could be trans-
parently applied. Matters are not always that simple, a point to
which we will return.70
First, however, it is time to turn to GTL.

II. GAME THEORY AND THE LAW

Game Theory and the Law is in some respects structured al

the lines of a conventional game theory text. We would group
eight substantive chapters as follows. Chapters One, Two, and
deal with elementary topics, including games in normal and ext
sive form, Nash and subgame perfect equilibria (pp. 6-78), and
additional insights obtained from modeling the repeated pla
simple games (pp. 159-87). Chapters Three and Four deal w
more advanced matters, species of games of imperfect informa
that can exhibit a multiplicity of subgame perfect equilibria,
which more elaborate solution concepts are required. Chapter S
is organized thematically around the subject of "collective acti
problems, while Chapters Seven to Eight turn to bargaining, un
both perfect and imperfect information.

having forgotten exactly where, ought to go. Each going to any same location in New
City is a Nash equilibrium. See pp. 38-40; SCHELUNG, supra note 48, at 54-58.
67. See infra text accompanying notes 94-95.
68. This requirement is usually called "sequential rationality." See, e.g., KREPS,
note 30, at 427-29. In effect, it generalizes to every information set - whether or n
beginning of a subgame - the same kind of rationality requirement that subgame perf
imposes on every proper subgame of a game.
69. See, e.g., FUDENBERO & TIROLE, supra note 37, at 321-26; KREPS, supra note 3
429-31. A closely related equilibrium refinement is the "sequential equilibrium." See
M. Kreps & Robert Wilson, Sequential Equilibria, 50 ECONOMETRICA 863 (1982).
70. See infra note 104 and accompanying text.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1864 Michigan Law Review [Vol. 94:1839

That would be an agenda of more than ordinary scope for any

introduction to game theory, but this is an atypical introductory ac-
count. GTL's principal objective is to survey the ways that game-
theoretic models can shed light on legal rules. Its ambitions extend
beyond examining strategic interaction in the shadow of the law, to
studying the implications for the structure of the law of strategic
interaction, especially in the presence of asymmetric information.
So the game-theoretic models function logically as a backdrop
against which to explore how strategic behavior might influence the
choice of legal rules. GTL does, of course, instruct the reader along
the way. But the instruction, while essential, is in important re-
spects subordinate to the larger agenda, which dictates the range of
topics GTL takes on, and influences not only the way those topics
are assembled into chapters but the order in which the chapters are
arranged.
GTL thus differs from conventional texts in ways that shape its
account of game theory itself. Many chapters are organized around
legal themes rather than by game-theoretic topic, often juxtaposing
problems that require quite different analytic methods.71 Perhaps
most importantly, GTL departs from most game theory texts in the
order in which topics are introduced. Texts devoted to game theory
are about how to do game theory, and generally adopt a line of
exposition that takes the student progressively from elementary to
more difficult matters, adding at each stage to the foundation for
what is yet to come. Driven by its organization around legal
themes, GTL departs from that kind of progressive development,
beginning very early in the book. In Chapters Three and Four, in
particular, it embarks on what proves to be an extensive explora-
tion of matters that are both intricate and problematic, considera-
tion of which most game theory books more judiciously postpone.
The net effect is that part of GTL - Chapters One, Two, Five,
and Seven - is essentially an introduction to the basics of game
theory, more or less along conventional lines, fashioned expressly
for lawyers. The balance - Chapters Three, Four, Six, and Eight
- is a selection of more advanced applications of game theory and
information economics to the law; and, at least as far as the game-
theoretic materials are concerned, proceeds in an atypical and far
less systematic way. We take up these two very different aspects of
GTL in turn.

A. The Rudiments

Chapters One ("Simultaneous Decisionmaking and the Nor

Form Game") and Two ("Dynamic Interaction and the Exten

71. We have in mind, particularly, chapters 3 and 6.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1865

Form Game") of GTL cover in substantially more detail the basic

aspects of game theory that we outlined in Part I. Chapter One
deals with games in normal form, and solutions based on domi-
nance arguments as well as Nash equilibria, while Chapter Two
takes up subgame perfect equilibria. In these chapters as well as
Chapter Five, GTL parallels closely conventional game theory
texts.

The differences between GTL and game theory texts are evi-
dent, to be sure, from the very outset of Chapter One. It does not
start by developing solution concepts using standard game theory
paradigms. It begins, instead, with one of the first explicit, and by
now classic, contributions of game theory to the law, the analysis of
civil liability rules. That analysis, which originated with John
Prather Brown,72 establishes that a broad class of damage rules for
tort liability can create incentives for both an injurer and a victim to
exercise "appropriate" or "due" care.73 GTL uses that framework
to introduce the normal form game and solutions based on domi-
nance, culminating with the Nash equilibrium and the insight that,
under a quite general class of compensatory damage rules, it is a
Nash equilibrium (even if not strictly dominant) for both parties to
exercise cost-minimizing care.74
In its introductory development, GTL dwells on the way in
which legal rules can influence the conduct of strategic actors. In
the abstract, they can do so in one or more of at least three basic
ways: (1) by altering the set of available strategies; (2) by altering
the payoffs; or (3) by altering the information structure of the game.
In its treatment of liability rules, for example, it emphasizes that the
choice of legal rule operates by reallocating the payoffs to strategic
interaction. GTL is also careful to emphasize that players can be
expected to adopt equilibrium strategies only if each can rely on the
rationality of the other. Recognizing that this may not always be a
plausible assumption, GTL concludes by exploring other liability
rules, including comparative negligence and rules in which a player
who exercises due care is reimbursed for the costs of precaution by

72. Brown, supra note 3; see also Peter A. Diamond, Single Activity Accidents, 3 J. LEGAL
STUD. 107 (1974); Steven Shavell, Strict Liability Versus Negligence, 9 J. LEGAL STUD. 1
(1980). These explicit formulations built on foundations laid by Guido Calebresi and Rich-
ard Posner; see, e.g., GUIDO CALEBRESI, THE COSTS OF ACCIDENTS: A LEGAL AND ECO-
NOMIC ANALYSIS (1970); Guido Calebresi and Jon T. Hirschoff, Toward a Test for Strict
Liability in Torts, 81 YALE L.J. 1055 (1972); Richard A. Posner, A Theory of Negligence, 1 J.
LEGAL STUD. 29 (1972); Richard A. Posner, Strict Liability: A Comment, 2 J. LEGAL STUD.
205 (1973). A very accessible account, using the game-theoretic framework implicitly, can be
found in SHAVELL, supra note 12.
73. "Due" care is defined as care that minimizes the sum of expected accident and acci-
dent prevention costs. See p. 13.
74. The class of damage rules that have this property includes strict liability with a de-
fense of contributory negligence, negligence (with or without a defense of contributory negli-
gence), and comparative negligence. See, e.g., p. 24.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1866 Michigan Law Review [Vol. 94:1839

a player who does not, which make it strictly dominant for both
injurer and victim to exercise due care.75 That is, under the final
class of liability rules that GTL surveys, it is optimal for both in-
jurer and victim to take cost-minimizing care irrespective of what
they expect the other players will do. Thus, informed by the differ-
ent demands of different game-theoretic solution concepts, the in-
troductory exposition concludes on a skeptical note regarding the
degree of reliance on the rationality of one's adversaries so typically
assumed in game-theoretic models (pp. 28-31).
Only then does GTL revert to a more conventional introduc-
tion, surveying the standard introductory game theory paradigms,
including the Prisoner's Dilemma, "coordination games" with mul-
tiple Nash equilibria, and games with no "pure strategy" equilibria
(pp. 37-43). In this, and the other, introductory chapters, GTL is
careful in qualifying the results that it describes. It is generally sen-
sitive to the problem of multiple equilibria (pp. 38-41). It is good
also at modifying the standard stories whenever possible to frame
its illustrations in legal settings. As with other chapters, Chapter
One concludes with a set of bibliographic notes (pp. 46-49) for the
reader interested in pursuing some topic in more detail.
There are, to be sure, questions that might reasonably be raised
with respect to Chapter One. Of these, what strikes us as most
problematic is GTL's choice of what to emphasize in text. A bit
more than half way through its introductory treatment of civil dam-
age rules, GTL undertakes to prove - but only after assuring the
reader that "we can set [the proof] out quickly" (p. 25) - that when
injurer and victim strategically interact in the shadow of a compen-
satory damage rule, both exercising due care is a Nash equilibrium
of the game. The ensuing proof, however, is rendered entirely in
words, consumes several pages of text, and might well bring the av-
erage reader to a more-than-momentary halt. For one thing, the
requisite assumptions are not explicitly collected; nor is it obvious
on first reading that what is being proved is that "(Due care, Due
care)" is a Nash equilibrium, or the unique Nash equilibrium, of the
game. Wading through a lengthy verbal proof is work enough; the
absence of a clear description of just what is being assumed, and
what is being proved, makes matters genuinely tough.
Overall, however, Chapter One provides a fine introduction,
clearly done, and well motivated by legal examples. A careful
reader would, we think, come away with approximately the same
basic grasp of the normal form, the standard game theory para-
digms, and solutions to static games of complete information using

75. See, e.g., Samuel A. Rea, Jr., The Economics of Comparative Negligence, 7 INTL. REV.
LAW & ECON. 149 (1987); Daniel Orr, The Superiority of Comparative Negligence: Another
Vote, 20 J. LEGAL STUD. 119 (1991).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1867

dominance arguments and Nash equilibria, as he might obtain from

a standard game theory text, together with an appreciation for both
the usefulness and the limits of game-theoretic methods in illumi-
nating the effects of legal rules.
Very much the same may be said of Chapters Two and Five
which also hew relatively closely to the structure of a conventional
game theory text. Incredible threats, subgame perfection, and the
role in making threats credible of prior conduct that is both observ-
able and costly to reverse (covered in Chapter Two) were originally
developed partly in the context of entry deterrence and are of as
much natural relevance to the structure of the antitrust laws as to
industrial organization itself. The same can be said of tacit collu-
sion over time, which can be modeled as the indefinitely repeated
play of the prisoner's dilemma (as it is in Chapter Five).76 So mod
els of entry deterrence and tacit collusion, as well as of reputation
and predation, are all part of the standard game theory exposition,
and they are developed in GTL. As in Chapter One, GTL goes
beyond the standard exposition and fashions examples of particular
interest to lawyers. As with Chapter One, the careful reader will
come away with an exposure very much like that obtained from an
elementary text, together with an enriched diet of applications of
relevance to the law. Here, too, the authors preserve a cautionary
stance, pausing often to put the implications of the models in
perspective.
Special mention should be made of Chapter Seven, where GTL
takes up bargaining, a matter of particular importance both to eco-
nomic theory and to law and economics. Its importance lies partly
in the fact that, despite the central role of bargains and exchange in
economic conduct and economic analysis, modeling the process of
bilateral bargaining has proved to be extraordinarily elusive.77 One
of the more remarkable recent advances in game theory is due to
Ariel Rubinstein, who showed that a two-person model of the bar-
gaining process as a noncooperative game of perfect information,
involving an indefinitely repeated set of alternating offers, has a

76. When the prisoner's dilemma is played repeatedly, equilibrium solutions involving
cooperation arise if players are sufficiently concerned about future payoffs. Although "de-
fection" still generates an immediate gain against a cooperating opponent, that gain is out-
weighed by the anticipated loss in the benefits of future cooperation which that defection
would induce. See, e.g., pp. 167-72; GIBBONs, supra note 12, at 88-97. This, we also note, is
an area in which the problem of multiple equilibria is especially acute. Id.
77. See, e.g., KREPS, supra note 34, at 91-97; KREPS, supra note 30, at 551-56. Another of
John Nash's signal contributions was a model of bargaining. In lieu of specifying a particular
bargaining process in extensive form, Nash formulated a set of axioms that any "reasonable"
solution to the bargaining problem should satisfy. He then established that these axioms
imply a unique solution. See Nash, The Bargaining Problem, supra note 2. For discussions of
the relevance of attitudes towards bargaining to law and economics generally, see Avery
Katz, The Strategic Structure of Offer and Acceptance: Game Theory and the Law of Contract
Formation, 89 MICH. L. REV. 215, 217-19 (1990); lan Ayres, supra note 4, at 1315-17.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1868 Michigan Law Review [Vol. 94:1839

unique subgame perfect equilibrium.78 Chapter Seven develops for

the reader the Rubinstein bargaining model together with a series
of interesting illustrations.79 It is one of the nicer chapters in GTL.
It also is, we might add, the locus of one of the more striking omis-
sions from the book. One of the few unambiguous claims in The
Problem of Social Cost is that, when unimpeded by transactions
costs, economic actors will arrive at mutually advantageous bar-
gains.80 The matter is debated to this day.81 GTL devotes two full
chapters to bargaining without so much as mentioning its relevance
to the work of Ronald Coase.

In sum, however, these chapters give an accessible elementary

account, expressly tailored to an audience of lawyers, of gam
theoretic work. If what one is looking for is a nontechnical intr
duction to basics of game theory and its use in legal settings, Ch
ters One, Two, Five, and Seven of GTL - in that order, we sugge
- would be a useful place to start.

78. See Ariel Rubinstein, Perfect Equilibrium in a Bargaining Model, 50 ECONOMETRI

97 (1982). The alternating offer model of bargaining is originally due to INGOLF STXH
BARGAINING THEORY (1972). Rubinstein's pivotal contribution was to establish that a v
sion of the model had a unique subgame perfect equilibrium. For a less technical introd
tion to the subject see, for example, KREPS, supra note 30, at 556-65. See also Ken Binm
et al., The Nash Bargaining Solution in Economic Modelling, 17 RAND J. ECON. 176 (19
John Sutton, Non-Cooperative Bargaining Theory: An Introduction, 53 REV. ECON. STU
709 (1986). A more complete treatment of bargaining can be found in MARTIN J. OSBOR
& ARIEL RUBINSTEIN, BARGAINING AND MARKETS (1990).
79. In the simplest version of Rubinstein's model, in which the payoff to both player
zero in the absence of agreement, they (approximately) divide the surplus from cooperat
If, however, one of them has an outside option worth more than he would obtain from
equilibrium solution in the absence of the option, his equilibrium share is the value of
outside option. GTL focuses on this feature of the Rubinstein model in several of the ex
ples it develops in chapter 7.
80. See Coase, supra note 5, at 3, 15.
81. The important early contributions include Robert Cooter, The Cost of Coase, 11
LEGAL STUD. 1 (1982); Donald H. Regan, The Problem of Social Cost Revisited, 15 J.
ECON. 427 (1972); see also Robert Cooter, Stephen Marks, & Robert Mnookin, Bargainin
the Shadow of the Law: A Testable Model of Strategic Behavior, 11 J. LEGAL STUD.
(1982); Joseph Farrell, Information and the Coase Theorem, J. ECON. PERSP., Fall 1987
113. Recent manifestations include Ian Ayres & Eric Talley, Distinguishing Between Con
sual and Nonconsensual Advantages of Liability Rules, 105 YALE L.J. 235 (1995); Ian Ay
& Eric Talley, Solomonic Bargaining: Dividing a Legal Entitlement To Facilitate Coas
Trade, 104 YALE L.J. 1027 (1995); Louis Kaplow & Steven Shavell, Do Liability Rules Faci
tate Bargaining? A Reply to Ayres and Talley, 105 YALE L.J. 221 (1995). From the expe
mental literature on bargaining, compare Glenn W. Harrison & Michael McKee,
Experimental Evaluation of the Coase Theorem, 28 J.L. & ECON. 653 (1985) and Elizab
Hoffman & Matthew L. Spitzer, The Coase Theorem: Some Experimental Tests, 25 J.
ECON. 73 (1982) with Ken Binmore et al., Hard Bargains and Lost Opportunities, (unp
lished manuscript on file with authors). For a recent review of that literature see RICHA
H. THALER, THE WINNER'S CURSE: PARADOXES AND ANOMALIES OF ECONOMIC LIFE 21
(1992).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1869

B. The Central Problem in Game Theory and the Law and the
Demands of Complex Models
We have more serious reservations, however, beginning with
Chapters Three and Four, in which GTL introduces a set of related
topics that make use of more intricate games of imperfect informa-
tion (including games of "incomplete" information) and more ad-
vanced solution concepts that refine the notion of a subgame
perfect equilibrium.82 The decision to introduce this material at an
early point reflects GTL's animating belief that asymmetric infor-
mation is the central problem in game theory and the law. That
belief is widely shared. In principle, the introduction of asymmetric
information brings economic analysis an important step closer to
the kinds of problems with which the law must actually deal. In
practice, much current work in game theory and the law involves
asymmetric information games. What at this point is rather less
clear is that game theory has evolved to a point at which it can
furnish reliable solutions. As illustrated at the end of Part I, the
very simplest games of imperfect information can have multiple
subgame perfect equilibria.83 That multiplicity is, unfortunately,
pervasive. What is more, the various equilibrium refinements that
have been proposed to eliminate implausible equilibria are both
controversial and intricate with which to work. So GTL's decision
to devote extended consideration to games of incomplete inform
tion leads to significant problems of exposition and elevates mater
ally the difficulty of the book. After describing the sort of game
that GTL takes up, we turn to its introductory account.

1. Games of Incomplete Information (John Harsanyi)

The imperfect information in Game Two, analyzed at the con

clusion to Part I, arose from a prior action by a human player tha
another player could not observe. There is, however, an importan
class of games of imperfect information that is derived from gam
of "incomplete" information, in which some characteristic of
player is what some other player cannot observe. The unobserved
characteristic is usually called the player's "type";84 different typ
of players have different motivations; and those different motiva
tions can always be represented by ascribing distinct payoffs

82. Subgame perfect equilibria form a subset of the set of Nash equilibria. By "refine
ments" of subgame perfection we mean a further narrowing of that subset of all Nash equil
ria that survive the requirement that they be subgame perfect, i.e., a subset of subgam
perfect equilibria.
83. See supra text accompanying notes 59-64.
84. See FUDENBERG & TIROLE, supra note 37, at 209-11; GIBBONS, supra note 12, at 146
47, 174; OSBORNE & RUBINSTEIN, supra note 21, at 231-32; Harsanyi, supra note 2; see als
Harsanyi, supra note 17, at 294-95.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 1870 Michigan Law Review [Vol. 94:1839

players of the different types.85 Among John Harsanyi's pivotal

contributions was in showing that, by positing a fictitious player
usually called "nature" - any game of incomplete information
could be reformulated as a game that opens with a "chance move"
at which, in accordance with a posited probability distribution -
assumed to be common knowledge86 - nature randomly "chooses"
the player's type.87 Since what is unobserved by the "uninformed"
player in the reformulated game is nature's opening move, the Har-
sanyi transformation converts a game of incomplete information
into a game of imperfect information, which may then be studied
along the lines outlined in Part I.88 It is that class of games on
which Chapters Three and Four extensively dwell.
An illustration may serve to make this point concrete. For sim-
plicity, regard us as of a single mind in all material respects, and
consider the game that you, the Reader, and we, the Reviewers,
might easily be viewed as playing. Our move is our assessment of
GTL, and it is something you can observe: we praise the book or
we do not. You cannot, however, observe our motivation, which
(you may assume) affects our payoff from the way in which we fash-
ion this review. Are we disinterested, guided only by the objective
of providing an honest evaluation, content with the satisfaction of a
job well done; or do we have a hidden agenda that we wish dis-
creetly to pursue? Perhaps the book is wonderful, but we have our
own book on game theory and the law that happens to be just com-
ing off the press;89 or, possibly, the book itself is really very bad, but
we wish to curry favor in Chicago. How you interpret what we say
in deciding whether to invest in reading GTL - your payoff, after
all, is what you might derive from reading it, net of time and cost, or
from the ways you might otherwise deploy your money and your
time - should take account of your assessment of which we are.
It is likely, of course, that you in general have some prior belief
about the fraction of honest reviewers in the general population.
Suppose, possibly naively, you put that fraction at as high as thirty-
five percent, the balance in your view being opportunists with their
own books in the works. Using the Harsanyi transformation, a
game theorist would regard us, Reader and Reviewer, as playing a

85. See Harsanyi, supra note 17, at 293.

86. See supra text accompanying note 37.
87. See Harsanyi, supra note 17.
88. It is sometimes stressed that a game of "incomplete" information is not sufficiently
well structured to be a game at all; and that, once the structure has been completed using the
Harsanyi transformation, it is simply a game of imperfect information. See, e.g., K.G.
BINMORE, FUN AND GAMES: A TEXT ON GAME THEORY 502 (1992); RASMUSEN, supra note
12, at 55. While this point is well taken, we adhere to the conventional if abused terminology
of referring to such games as "games" of incomplete information.
89. For the record, we don't. If we did, admittedly we might strategically deny it.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1871

game of imperfect information, in which "nature" opens the game

by randomly drawing an opportunistic or disinterested reviewer -
choosing the reviewer's type - with the probability of the latter
being thirty-five percent.90 We then make our move by writing the
review, and, after observing our move but not nature's, you - the
"uninformed" player - move by deciding whether to invest in
reading GTL or to wait, perhaps for some some future book.
As you might imagine, strategic actors are frequently uncertain
about the nature of the other players. A buyer may be uninformed
about the quality of goods offered by sellers of different types;91 an
employer may be uncertain about the "quality" of prospective
workers;92 a defendant may be uncertain about the true nature of
the injury to a plaintiff from whom it has received a settlement de-
mand.93 With the Harsanyi transformation, each of these, and the
myriad of analogous possibilities, can be modeled as a game of im-
perfect information.
The structure of such models, while formally like that of other
imperfect information games, is often more complex. They always
involve chance moves and require the use of probability, features
that lend intricacy to the analysis. As with our Game Two, more
elaborate solution concepts must typically be invoked to eliminate
implausible equilibria of the kind exhibited by that game. Those
refined solution concepts, although extensively invoked by GTL,
are more open to debate than one might ever gather from GTL
itself. For all these reasons, most accounts of game theory of which
we are aware treat games of incomplete information as a more ad-
vanced topic, to be taken up only after a careful grounding in the
rudiments,94 and treat equilibrium refinements as a subject for

90. We offer this example for its accessibility. As the reader may have noticed, it is actu-
ally more complicated than we have suggested in the text. At the beginning of the game,
nature chooses both the reviewer's type (disinterested or opportunistic) and the quality (high
or low) of the book. Hence, there are four possible states of the world amongst which nature
must choose. In the simpler case more typically studied, nature chooses between just two
states.

91. See, e.g., George A. Akerlof, The Market for "Lemons": Quality Uncertainty and t
Market Mechanism, 84 Q.J. ECON. 488 (1970); see also pp. 79-121; infra text accompanyi
notes 96-97.

92. See, e.g., pp. 122-58; A. MICHAEL SPENCE, MARKET SIGNALING: INFORMATIONAL
TRANSFER IN HIRING AND RELATED SCREENING PROCESSES (1974).
93. See, e.g., pp. 244-67; Barry Nalebuff, Credible Pretrial Negotiation, 18 RAND. J. ECON.
198 (1987); I.P.L. P'ng, Strategic Behavior in Suit, Settlement, and Trial, 14 BELL J. ECON. 539
(1983); Jennifer F. Reinganum & Louis L. Wilde, Settlement, Litigation, and the Allocation of
Litigation Costs, 17 RAND J. ECON. 557 (1986); Stephen W. Salant, Litigation of Settlement
Demands Questioned by Bayesian Defendants. 516 CAL. INST. TECH. SOC. SCI. WORKING
PAPER 1 (1984).
94. For comparison, a newly published micro-economic theory text, with extensive treat-
ment of game theory, broaches these topics in the most preliminary way only after an eighty
page introduction devoted entirely to game theory. ANDREU MAS-COLELL ET AL.,
MICROECONOMIC THEORY 217-301, 282-96 (1995). See also FUDENBERG & TIROLE, supra

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1872 Michigan Law Review [Vol. 94:1839

which a healthy dose of skepticism is required. As one eminent

game theorist put it, at the outset of a graduate textbook discussion
of the subject, "there is no settled conventional wisdom about re-
finements of Nash equilibrium. Indeed, there is much controversy
about which refinements are reasonable and which not."95

2. Solutions to Games of Incomplete Information and Chapter

Three of GTL
From that perspective, GTL's early introduction of, and sus-
tained subsequent preoccupation with, games of incomplete infor-
mation and equilibrium refinements is a problematic expositional
decision. That is especially so in a book that has devoted more of
its first eighty pages to illustrating applications of game theory to
the law than to schooling readers in the solution of noncooperative
games. Nevertheless, GTL delves deeply into these topics in Chap-
ters Three and Four, and returns to them again in Chapters Six and
Eight. That is something it did not have to do. But given its point
of view, and given also the delicate handling that Nash equilibrium
refinements require, one would naturally expect GTL to be espe-
cially meticulous in introducing an audience of uninitiated readers
to the solutions of such games, and to the elements of probability

note 37, at 319-434; GIBBONS, supra note 12, at 173-255; KREPS, supra note 30, at 417-450;
OSBORNE & RUBINSTEIN, supra note 21, at 197-254.
95. KREPS, supra note 30, at 418. In the more outspoken words of another game theorist
in an undergraduate text, "the subject of refinements of Nash equilibrium . . . is the most
controversial . . . within game theory .... [T]his is an area in which I recommend not
believing anyone until the experts achieve some sort of consensus." BINMORE, supra note 88,
at 544-45. Kreps' remarks are especially germane, since he is responsible for two of the more
prominent equilibrium refinements, the "sequential equilibrium," formalized in Kreps & Wil-
son, supra note 69, at 863, and the "intuitive criterion" of In-Koo Cho & David M. Kreps,
Signaling Games and Stable Equilibria, 102 Q.J. ECON. 179 (1987). Other prominent equilib-
rium refinements include "trembling-hand perfection," see Selten, Perfectness Concept, supra
note 2; "proper" equilibria, Roger B. Myerson, Incentive Compatibility and the Bargaining
Problem, 47 ECONOMETRICA 61 (1979); the set of "stable" equilibria," Elon Kohlberg &
Jean-Francois Mertens, On the Strategic Stability of Equilibria, 54 ECONOMETRICA 1003
(1986); the set of "perfect sequential equilibria," Sanford J. Grossman & Motty Perry, Perfect
Sequential Equilibrium, 39 J. ECON. THEORY 97 (1986); and the set of "universally divine"
equilibria, Jeffrey S. Banks & Joel Sobel, Equilibrium Selection in Signaling Games, 55
ECONOMETRICA 647 (1987). Many are deployed at one point or another by GTL. See, e.g.,
pp. 86-90 (perfect Bayesian equilibrium), 132-33 (the intuitive criterion); 252-59 (perfect se-
quential equilibrium).
Many of these refinements narrow the set of equilibria by restricting the "beliefs" that
players may hold when they encounter eventualities that are off the equilibrium path. See,
e.g., JEFFREY S. BANKS, SIGNALING GAMES IN POLITICAL SCIENCE 14-15 (1991); see also infra
note 104. In other words, they attempt to impose restrictions on what players should be
permitted or taken to believe when they observe events that, in equilibrium, they expect
never to observe at all. An excellent non-technical account of the problems generally can be
found in KREPS, supra note 34, at 108-22. For more technical surveys, see, for example, MAS-
COLELL ET AL., supra note 94, at 282-96; KREPS, supra note 30, at 417-43. A thoughtful
discussion of the relevance of the issues to game theory and the law can be found in Ayres,
supra note 4, at 1291, 1311-13; see also Ian Ayres, Three Approaches to Modeling Corporate
Games: Some Observations, 60 CIN. L. REV. 419, 421-22 (1991).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1873

that such solutions routinely require. Those, however, are things

that GTL simply does not do.
From a substantive perspective, Chapter Three focuses on infor-
mation disclosure and disclosure laws. GTL chooses for its organiz-
ing thread what has been called the "unraveling result," due to
Sanford Grossman and Paul Milgrom.96 The central insight of that
work is that if private information, held for example by a seller of
goods, can be verified and misrepresentation remedied costlessly ex
post, accurate disclosure of favorable information is optimal for
every type of seller, and silence effectively discloses "bad news."
Although its assumption of costless verifiability is relatively strong,
in principle it has intriguing implications for the desirability of dis-
closure laws. In its substantive dimension, Chapter Three begins by
describing that result, and exploring its implications in a variety of
settings. Ultimately, it settles into extended consideration of a
model, due to Steven Shavell,97 that goes beyond the unraveling
result to explore the impact of disclosure laws on incentives to en-
gage in costly acquisition of information. Shavell's work suggests
that, even when voluntary disclosure may privately be optimal,
mandatory disclosure may be useful in deterring socially excessive
investments in acquiring information.98
GTL weaves into this story its introduction to games with infor-
mational asymmetries. Our concerns begin with its exposition of
the refinements of subgame perfection that forms the analytic cen-
terpiece of Chapter Three. The Chapter does begin on a note of
intuition. But it proceeds immediately to "formalize" the intuition
by introducing that refinement of subgame perfection called the
perfect Bayesian equilibrium.99 That move will confront any untu-
tored reader with something for which she will not be adequately
prepared. Any real grasp of that solution concept, and others like
it, requires a facility with simple expected values and an acquain-

96. Sanford J. Grossman, The Informational Role of Warranties and Private Disclosure
About Product Quality, 24 J.L. & ECON. 461 (1981); Paul R. Milgrom, Good News and Bad
News: Representation Theorems and Applications, 12 BELL J. ECON. 380 (1981).
97. See Steven Shavell, Acquisition and Disclosure of Information Prior to Sale, 25 RAND
J. ECON. 20 (1994); see also Anthony T. Kronman, Mistake, Disclosure, Information, and the
Law of Contracts, 7 J. LEGAL STUD. 1 (1978).
98. Shavell's work draws on a distinction between information of social and that of purely
private value, first apparently articulated in Jack Hirshleifer, The Private and Social Value of
Information and the Reward to Inventive Activity, 61 AM. ECON. REV. 561 (1971).
99. See supra text accompanying notes 68-70; infra note 104. The implication in the text
that the set of "perfect Bayesian equilibria" is a subset of the set of subgame perfect equilib-
ria is, strictly speaking, correct for games with no proper subgames. Games with proper
subgames can, however, have perfect Bayesian equilibria that are not subgame perfect. See,
e.g., MAS-COLELL ET AL., supra note 94, at 290, Figure 9.C.5. In such cases an accurate
characterization is that solutions that are perfect Bayesian equilibria in every subgame are
always subgame perfect.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 1874 Michigan Law Review [Vol. 94:1839

tance with both "conditional probabilities" and Bayes' Law.100 If,

like GTL, your operating premise is that asymmetric information is
what counts, those are matters you simply cannot avoid. The eco-
nomics of asymmetric information cannot be done without expected
values. Except in the very simplest of settings, you cannot think
about a perfect Bayesian equilibrium without knowing how to work
with probabilities, and how to use Bayes' Law to update probabili-
ties in response to things that are observed. Those, however, are
matters that GTL never develops for its readers.101 Instead, it
launches headlong into games of asymmetric information with none
of those prerequisites in hand.
Having done so, moreover, GTL proceeds to a formal introduc-
tion using a game of imperfect information that is neither especially
well chosen nor instructive. GTL begins by asserting that its illus-
trative game has multiple equilibria and no proper subgames, so
that subgame perfection is no help. It goes on to claim that, using
the perfect Bayesian equilibrium to "incorporate beliefs into the so-
lution concept, we can eliminate some Nash equilibria as solutions
to the game" (p. 86). GTL then consumes several difficult pages
establishing that the game has two perfect Bayesian equilibria, but
it never identifies just which Nash equilibria the exercise managed
to rule out. The answer, it happens, is "none." A feature of the
illustration - which GTL never mentions - is that the two strat-
egy profiles that survive the application of the perfect Bayesian
equilibrium are the only two pure-strategy Nash equilibria of th
game. GTL thus devotes ten pages to introducing an intricate ne
solution concept, the purpose of which is to eliminate implausib
equilibria, using an example in which it has no effect at all! It i
insufficient to suggest that GTL manages to obscure the point. W

100. Bayes' Law is an elementary but important proposition about conditional probab
ties that offers a natural method by which to update beliefs, in the sense that it offers a way
evaluating the probability of an unobserved event in light of some observed "signal" that h
a bearing on its probability. See, e.g., PAUL L. MEYER, INTRODUCTORY PROBABILITY A
STATISTICAL APPLICATIONS 39-41 (2d ed. 1970). Its principal (and controversial) appe
ances in legal literature have been in connection with evidence. See, e.g., Lawrence H. Tr
Trial by Mathematics: Precision and Ritual in the Legal Process, 84 HARV. L. REV. 1
(1971); Richard D. Friedman, A Close Look at Probative Value, 66 B.U. L. REV. 733 (19
see also Steven C. Salop, Evaluating Uncertain Evidence with Sir Thomas Bayes: A Note
Teachers, J. ECON. PERSP., Summer 1987, at 155.
101. In the interests of completeness, we note that a "probability distribution" over t
discrete events is just a pair of non-negative numbers that sum to one; and, that the "
pected value" of a random experiment whose outcomes are numbers associated with the t
events is the average of those numbers, weighted by the associated probabilities. Althou
GTL routinely computes expected values throughout the book - nearly every footnote
chapter 8, for example, contains an expected value calculation (pp. 285-87) - it never ta
the time to explain to the reader what they are, and the index does not even contain an ent
for the topic. There appear to be no references to "conditional probability" anywhere in
book. Of Bayes' Law GTL says that it "provides a means to capture formally the way r
tional people should update their beliefs in the wake of new information" and little else
83.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 May 1996] Game Theory 1875

know of no other introduction to this material that is quite so thor-

oughly opaque.102 At its end, the reader is still left to discern for
herself what useful purpose, if any, this innovation is really sup-
posed to serve.
Finally, GTL casually invokes yet another equilibrium refine-
ment to rule out one of the two equilibria, thereby producing a "so-
lution" to the game. It does so, moreover, without extended
description or discussion, leaving the reader not a clue to just what
it is invoking, whether it is original with GTL or not, and whether it
is technically a controversial or commonplace thing to do.103 It
never does tell the reader just what, if anything, Bayes' Law has to
do with any of this.
With that as its prologue, GTL proceeds immediately to the un-
raveling result and into the heart of Chapter Three, as though its
ideal reader, having stopped to think about the matter "logically
and carefully," would thereafter be equipped to follow what was
going on. When, another ten pages later, GTL takes up Shavell's
model of the consequences of disclosure, we suspect it will rapidly
become quite difficult for any reader not already well versed in
game theory to keep up. For, despite the mystifying introduction,
GTL plunges into the solution of a series of games using equilib-
rium refinements in excruciating detail. It does so, moreover, with-
out ever formally specifying a single game in either extensive or
normal form. It is, to be sure, possible to follow the details, and
perhaps to discern the moral of the stories being told. But, given
GTL's self-imposed commitment to do it almost all in words, there
is little possibility that an otherwise unschooled reader of Chapters
Three and Four will end up with a real grasp of what is going on.
We do not mean to suggest that games with asymmetric infor-
mation require formal training to understand. To the contrary, our
view is that they do not. In fact, the solution that we sketched to
Game Two at the conclusion of Part I was itself a perfect Bayesian
equilibrium, used in a setting in which that solution concept actually

102. Games typically used to introduce a rarefied refinement of subgame perfection like
this are those in which (a) the refinement operates to eliminate some equilibria and (b) the
eliminated equilibria are arguably implausible. See, e.g., GIBBONS, supra note 12, at 175-80;
KREPS, supra note 30, at 425-28; MAS-COLELL ET AL., supra note 94; see also supra text
accompanying notes 67-69; infra note 104 and accompanying text.
103. See pp. 88-89. The added refinement does not appear in the bibliographic notes,
which tell the reader only that "perfect Bayesian equilibrium" is related to "sequential equi-
librium." P. 119. A trip to the index entry for "sequential equilibrium" refers the reader to
"equilibrium," whereupon he encounters a dead end. The invoked refinement is sometimes
referred to as "forward induction," or "iterated elimination of weakly dominated strategies,"
and its merits are open to dispute. See, for example, MAS-COLELL ET AL., supra note 94, at
292-95, where, after canvassing arguments both for and against the refinement's plausibility,
the authors conclude: "Clearly, the issues here, although interesting and important, are also
tricky." See also OSBORNE & RUBINSTEIN, supra note 21, at 110-14, 244.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1876 Michigan Law Review [Vol. 94:1839

ruled an implausible equilibrium out.104 It was structured, however,

to be transparently simple. It requires, in particular, no obvious
Bayesian updating or expected value calculations, things that will
not be true in general.105 But, when they are needed, the expected
value calculations that GTL never systematically explains are often
little more demanding than adding two and two; and the uses to
which Bayes' Law is actually put, while requiring more careful
groundwork, are only modestly more difficult and no stranger to

104. See supra text accompanying notes 67-69. Speaking loosely, equilibrium refinements
like the perfect Bayesian equilibrium restrict the set of subgame perfect equilibria in a game
of imperfect information by requiring that the play of the game be "sequentially rational" at
every information set, including information sets that are off the equilibrium path of play.
That condition extends in a natural way to every non-singleton information set a requirement
of out-of-equilibrium rationality comparable to what subgame perfection requires in proper
subgames (which, by definition, begin at singleton information sets). See supra notes 50-51,
62-64 and accompanying text. At a nonsingleton information set, however, a prerequisite to
assessing sequential rationality is a system of "beliefs," in the form of a set of probabilities,
conditional on being in that information set, of being at the different choice nodes that it
contains. The beliefs are used to compute the expected values of the actions available to the
player whose move it is to make.
More formally, a perfect Bayesian equilibrium consists of both a strategy profile and a
system of beliefs, requiring (wherever possible) the beliefs to be consistent with the strategy
profile (which is where Bayes' Law comes into play) and the strategy profile to be sequen-
tially rational. The strategy profile is sequentially rational if each player chooses the action at
each of his information sets that maximizes his expected payoff, computed using the beliefs.
The various refinements differ in the restrictions they impose on beliefs at information sets
that are off the equilibrium path. See, e.g., BANKS, supra note 95, at 14-15. The weakest form
of perfect Bayesian equilibrium, in particular, imposes no restriction on beliefs at any infor-
mation set that would never be encountered if the equilibrium strategy profile were played.
See, e.g., KREPS, supra note 30, at 427-31: BINMORE, supra note 88, at 536-38 & n.40; MAS-
COLELL ET AL., supra note 94, at 285-95.
In Game Two, however, irrespective of whether the Insurer makes the High or Low qual-
ity Partial offer, the Resident's payoff from settling exceeds her payoff from going to court.
See supra text accompanying notes 56-57 and 67-69. So for any system of beliefs - any set of
probabilities that she is faced with the P-H or P-L offer - the expected value of settling
always exceeds the expected value of going to court, and it is always sequentially rational for
the Resident to settle. Given that, the Insurer always should make a P-L offer; the Resident
should always observe a Partial offer which, while of unobserved quality, the Resident should
always believe to be of Low quality; and the Resident's sequentially rational action should be
to settle. Given the equilibrium strategy profile, the equilibrium beliefs are unique.
What makes our example seem easy is the absence of any dependence of the sequential
rationality of the equilibrium strategy profile on the beliefs, which permits us to determine
the beliefs implied by the strategy profile almost as an afterthought. More typically, the
sequential rationality of a strategy profile depends on the beliefs, but the beliefs depend in
turn on the strategy profile: it is the joint determination of the two that can make matters so
complex.
Nevertheless, while Game Two illustrates the use of the perfect Bayesian equilibrium in a
deceptively simple setting, it is a setting in which the concept has some bite. It eliminates as a
solution the implausible equilibrium in which the Resident's strategy involves a threat to go
to court, despite the fact that it is always best for her to settle.
105. That is why we originally stipulated that the payoff to the Resident from settling,
even for the P-L offer, exceeded the $9 payoff from taking it to court. See supra text accom-
panying note 55. If, in contrast, the payoff from settling for P-L had been less than $9 while
the payoff from settling for the P-H offer remained $11, it would then be sequentially rational
for the Resident to settle only if, given a set of probabilities that the Insurer had made the P-
H or P-L offers, respectively, the expected (i.e., probability-weighted) payoff from settling
exceeded the $9 payoff from going to court. See supra note 104.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1877

legal discourse in any event.106 If the reader were adequately pre-

pared and the materials more judiciously positioned and methodi-
cally presented, there is no reason why interested lawyers could not
master them. The economics of disclosure, with which Chapter
Three is substantively concerned, makes for a fascinating story in
and of itself. A case for including a discussion of the topic some-
where in GTL is compelling, after the requisite groundwork has
been laid.
The matter of positioning aside, at least two distinct courses
were open to GTL as ways of presenting these materials. One
would have been to provide a more soundly conceived introductory
account of equilibrium refinements - including the prerequisites
- treating disclosure regimes (on perhaps a less ambitious scale) as
an illustrative application. It might alternatively have structured
the chapter as an essay on the economics of disclosure, describing
the technical results to the extent necessary to lend real content to
the story. At different points Chapter Three reads as though it had
both of these in mind. What it failed to do was choose. Chapter
Three oscillates between introducing and then using advanced solu-
tion concepts and exploring the economics of disclosure, leaving the
reader no single thread with which to negotiate the maze. The
study of disclosure leads to a fragmentary account of the theory but
then itself disappears in a sea of details, as GTL investigates more
than a dozen possible equilibria of nearly a half-dozen different
games. The novice reader - however attentive he might be
could easily emerge confused about the exact message for the struc-
ture of disclosure laws, and a sense of the game theory that is surely
incomplete, could easily be incorrect, and might be lacking alto-
gether in coherence.

C. Plausibility, Game Theory, and the Law

As far as asymmetric information is concerned, the conse-
quences linger for the balance of the book. The headlong exposi-
tion more generally persists. One further consequence of that form
of exposition is that it leaves little space in which to reflect on the
operational plausibility of the theory GTL so extensively surveys.
While relatively scrupulous about such matters at the outset,107
GTL's concerns become attenuated as the book proceeds. Model-
ing problems of asymmetric information has become a large-scale
industry, as GTL's existence itself reflects. But more sophisticated
models are increasingly sensitive not merely to details of the ways
that they are specified, but also to what they assume of human ra-

106. See supra note 100.

107. See supra text accompanying notes 74-75.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1878 Michigan Law Review [Vol. 94:1839

tionality, as reflected in the solution concepts they employ.108 At

the same time, in part because their predictions are so frequently
imprecise, game theoretic models are notoriously hard to falsify.
As it ventured into the realm of more demanding models, one
might reasonably have wished for GTL to become more attentive
to questions of verisimilitude, not less.
We offer two examples, the first of which underlies Chapter
Three itself. Both the basic unraveling result and Shavell's model
of acquisition and disclosure assume that affirmative representa-
tions must be truthful and, therefore, will be believed.109 In an en-
tire chapter spun together with the unraveling result, the authors
devote no space to what might happen were that important assump-
tion to be relaxed; indeed, they do not even mention that the as-
sumption is crucial to the chapter's core result.
We draw our second from among our favorite chapters in the
book. It does not even involve asymmetric information. Rubin-
stein's model"0 opened a new vista on bargaining, eliciting a wave
of new theoretical work. An entire chapter of GTL is fashioned
around the basic prediction of that model. But the prediction is
wildly sensitive to assumptions, a well-known fact to which GTL
only indirectly alludes.1'l At the same time, bargaining is more
than ordinarily amenable to controlled experiment and has elicited
a body of empirical evidence against which the theory may be
gauged. Even where the extensive form is entirely straightforward
- two-person, two-move games of perfect information - the re-
sults persistently conflict with what the theory would predict.12 It
is possible that such findings are the product of defects in experi-
mental design, so it remains open whether it is the design or the
theory that is flawed.113 Even so, such persistent, disquieting find-

108. This point is addressed in Ayres, supra note 4, at 1311-15.

109. See, e.g., Grossman, supra note 96, at 464-65; Shavell, supra note 97, at 22. GTL
recasts this as an assumption that misrepresentation - as contrasted with simple silence -
can be costlessly remedied ex post (p. 90).
110. See supra text accompanying notes 97-100.
111. See pp. 224, 239-41. Compare, e.g., KREPS, supra note 30, at 561-65. When incom-
plete information is introduced, to which GTL turns in chapter 8, the problem of multiple
equilibria intrudes in a serious way. See, e.g., Sutton, supra note 78, at 717-21.
112. One such game, the "ultimatum game," involves the simple division of a monetary
stake, in which the first player makes an offer that the second observes and may accept or
reject. If the offer is rejected, each player gets nothing. Theory predicts that the first player
should make the smallest possible positive offer, which the second player should accept. But
first players typically make substantial offers, and small offers are frequently rejected. See,
e.g., Colin Camerer & Richard H. Thaler, Ultimatums, Dictators and Manners, J. ECON.
PERSP., Spring 1995, at 209; RICHARD H. THALER, THE WINNER'S CURSE 21-31 (1992); K.
Binmore et al., Testing Noncooperative Bargaining Theory: A Preliminary Study, 75 AM.
ECON. REV. 1178 (1985).
113. See, e.g., Elizabeth Hoffman et al., Preferences, Property Rights and Anonymity in
Bargaining Games, 7 GAMES & ECON. BEHAVIOR 346 (1994).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1879

ings, in the very simplest of settings, counsel caution in bringing the

insights of game theory to bear on the design of legal rules. They
strike us as meriting more space and emphasis than the two biblio-
graphic sentences (p. 242) GTL devotes to them.
In this respect, to be fair about it, GTL is surely not alone.
More energy is currently devoted to doing game theory in legal set-
tings than to musing about its shortcomings, and their implications
for the real usefulness of game theory to the law. Still, this is a full-
length book devoted entirely to game theory and the law. Just how
useful game theory truly is, or ultimately may become, is something
GTL might have taken more time to explore.

CONCLUSION

GTL, in sum, is really two quite different books. The more ele-
mentary chapters provide an accessible, nontechnical introduction
to the basics of game theory, well illustrated in the context of the
law. For beginners, however, those that dwell on informational
asymmetries will require more than just care and logic to get
through. In all, GTL has performed an unquestionable service, in
rendering the basic insights of game theory in terms that will be
both companionable and accessible to nonspecialists whose primary
training is in law. Judiciously approached, it will provide rewarding
insights to those prepared to do some work.
The real questions, we suppose, are whether you are interested
in learning more game theory; if so, how much; and where would
you be well-advised to start. If your answer to the first is "yes,"
GTL is one obvious answer to the last. If that is what you choose,
we do encourage you to bypass Chapters Three, Four, and Six on
your first pass through the book.114 There are, however, three
other possibilities - complements, really, not substitutes for GTL
- that we might suggest. One, of course, is Thomas Schelling's
classic, The Strategy of Conflict. If you are interested in strategic
thinking generally, even if not formally in game theory, it should
not in any event be missed.115 An updated account along similar
lines is Avinash Dixit and Barry Nalebuff's Thinking Strategi-
cally,ll6 which provides a chatty, nontechnical introduction to the
subject and its underlying intuition.
But for a current, nontechnical introduction to noncooperative
game theory, there is a third possibility that genuinely stands out.

114. Chapter 6 is a somewhat free-standing essay on collective action problems, but it

draws on some of the solution concepts developed in chapter 3.
115. SCHELLING, supra note 48. See also the essays collected in CHOICE AND CONSE-
QUENCE, supra note 29, especially What is Game Theory, at 213.
116. AVINASH K. DIXIT & BARRY J. NALEBUFF, THINKING STRATEGICALLY: THE COM-
PETITIVE EDGE IN BUSINESS, POLITICS, AND EVERYDAY LIFE (1991).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
1880 Michigan Law Review [Vol. 94:1839

David M. Kreps's Game Theory and Economic Modelling17 offers

an honest, critical account of the field, with both fond praise for its
accomplishments and unflinching attention to its shortcomings.
Although written more than seven years ago, everything Kreps said
then continues to ring true. It is not intended for an audience of
lawyers, but it will be at least as accessible to them as GTL. It
reflects at length on the plausibility of Nash equilibrium analysis,
and the extent to which game-theoretic models can reasonably be
expected to provide insight into how real people actually do be-
have. That, as we have said, is a dimension of the subject on which
GTL does not dwell. For a serious, critical account of game theory
for the novice, Kreps's book is the place that we would start. Then
we would turn to GTL, primarily for its more introductory chap-
ters, and their insights into the relevance of the more basic aspects
of game theory to the law.
If, however, you are interested in the kinds of asymmetric infor-
mation problems that GTL takes up in Chapters Three and Four,
additional investment will be required. A short introduction to ex-
pected values, conditional probability, and Bayes' Law is essential;
there are many places one can go for that.118 After that there is a
growing list of possibilities, all of which assume at least an acquain-
tance with mathematical notation.119 To understand equilibrium re-
finements and solutions to games of imperfect information, you will
in all events have to read beyond GTL.
An important prior question, however, is whether you should
want to extend your horizons quite that far. One indisputable con-
tribution of game theory, as GTL frequently points out, is that,
even at what Martin Shubik has called its "conversational" level,120
it helps us hone our instincts about how the possibility of strategic
interaction might influence our thinking about legal rules. Or as
Thomas Schelling put it:

117. KREPS, supra note 34.

118. One good choice for the nonspecialist is EDITH STOKEY & RICHARD ZECKHAUSER,
A PRIMER FOR POLICY ANALYSIS 209-11, 221-29 (1978). Those extracts are included in an
introductory chapter on statistical decision theory which, for anyone interested in non-
cooperative game theory, is more generally worthwhile. Conditional probabilities and Bayes'
Law are also covered early in any introductory probability text.
119. Two of the most popular introductory accounts are GIBBONS, supra note 12, which
contains a wealth of worked-out applications; and RASMUSEN, supra note 12, which devotes
considerable space to intuition. The review of the first edition of GAMES AND INFORMATION
by Ian Ayres remains worth reading for its own interest. Ayres, supra note 4. Graduate-level
introductory microeconomic theory texts increasingly contain extended treatments of game
theory. One of the best introductory accounts will be found in MAS-COLELL ET AL., supra
note 94, at chapters 7-8; another excellent choice is KREPS, supra note 30, at 355-719. Yet
another recent possibility is JAMES D. MORROW, GAME THEORY FOR POLITICAL SCIENTISTS
(1994).
120. Martin Shubik, Game Theory, Law, and the Concept of Competition, 60 U. CIN. L.
REV. 285, 289-300 (1991).

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
May 1996] Game Theory 1881

For the social scientist, what is rudimentary and conceptual about

game theory will be, for a long time, the most valuable. And it will be
valuable not as "instant theory" just waiting to be applied but as a
framework - one with a great deal of thought now behind it - on
which to build his own theory in his own field.121
But, for applications to the law, more is required than merely that.
If game-theoretic insights are properly to influence our choice of
legal rules it must also be because formal game-theoretic models
provide us with credible insights into how human actors really do
behave. Doubts remain about the extent to which even simple
models of strategic interaction offer reliable predictions. Those
doubts become more acute when the interactions involve asymmet-
ric information or are otherwise complex. They have provoked
skepticism, even among those who do game theory for a living,
sometimes more pessimistic than the measured self-assessment that
will be found in Kreps. One of them, Kenneth Binmore, recently
put the matter this way:
It was not until the early 1970s that it was fully realized what a power-
ful tool Nash had provided in formulating the equilibrium concept
that bears his name. Game theory then enjoyed a renaissance as
economists applied the idea to a wide range of problems. However, a
fly in the ointment was awaiting discovery. Games typically have
many Nash equilibria .... At first it was thought that the problem
could be tackled by refining the Nash equilibrium concept. Despite
Nash's remarks in his thesis about a possible evolutionary interpreta-
tion of the idea of a Nash equilibrium, attention at that time was fo-
cused almost entirely on its interpretation as the only viable outcome
of careful reasoning by ideally rational players. Various bells and
whistles were therefore appended to the definition of rationality.
These allowed some Nash equilibria to be discarded as inadequately
rational according to whatever new definition of rationality was being
proposed. However, different game theorists proposed so many dif-
ferent rationality definitions that the available set of refinements of
Nash equilibrium became embarrassingly large. Eventually, almost
any Nash equilibrium could be justified in terms of someone or
other's refinement. As a consequence a new period of disillusion-
ment with game theory seemed inevitable by the late 1980s.122
Or, as the editors of the New Palgrave continued, in the quotation
with which we began, "[s]ometimes, it seems that we are back again
at our early morning calisthenics, only this time using high-tech
equipment."123

121. SCHELLING, supra note 29, at 241.

122. Ken Binmore, Foreword to JORGEN W. WEIBULL, EVOLUTIONARY GAME THEORY
vi (1995); Binmore's criticism of the "refinements literature" should not be offered without
context. It is from the forward to a book devoted to another of game theory's research fron-
tiers, so-called "evolutionary" game theory. See, e.g., Eric Van Damme, 38 EUR. ECON. REV.
847, 856 (1994).
123. NEW PALGRAVE: GAME THEORY, supra note 1, at xii.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
All use subject to http://about.jstor.org/terms
 1882 Michigan Law Review [Vol. 94:1839

GTL does well at introducing its readers to

agenda is dominated by a fascination with preci
the field that, while unquestionably of greatest
law and economics scholars, have simultaneously
among such talented and thoughtful game
Binmore and David Kreps. Even so, GTL is a
could, we think, have been a better book had i
preoccupied with the current frontiers of gam
systematically attentive to the possible limits o
of its development, game theory realistically mi
the law.

This content downloaded from 141.211.4.224 on Mon, 15 May 2017 17:12:34 UTC
View publication stats
All use subject to http://about.jstor.org/terms