JOURNAL OF ECONOMIC THEORY 25, 92-100 (1981)

Games of Perfect Information,

Predatory Pricing and the Chain-Store Paradox

ROBERT W. ROSENTHAL

Bell Telephone Laboratories, Murray Hill, New Jersey 07974

Received April 8, 1980; revised July 2, 1980

1. I~-i-R00uCn0N

The thesis of this paper is that finite, noncooperative games possessing

both complete and perfect information ought to be treated like one-player
decision problems. That is, players ought to assign at every move subjective
probabilities to every subsequent choice in the game and ought to make
decisions via backward induction. This view is in contrast with the game-
theoretic approach of Nash equilibrium.
After expanding on this view for games in the abstract in Sections 2 and 3,
attention is turned in Section 4 to an example due to Reinhard Selten, called
the chain-store paradox, which possesses the flavor of a situation involving a
predatory-pricing monopolist. It is argued that for the chain-store game the
decision-analytic approach leads, under certain assumptions, to more
realistic outcomes than the standard Nash-equilibrium approach.

2. AN ILLUSTRATIVE EXAMPLE

Consider the game tree in Fig. 1 to be played only once in which the
outcomes are assumed to be expressed in U.S. dollars and x and y are dollar
values known to both players. In words, if Player 1 chooses Left, he receives
x dollars and 2 receives y dollars; if Player 1 chooses Right, then the
outcome is either (0,O) or (1 million, 1) depending on Player 2’s choice.
Assuming that each player’s von Neumann-Morgenstern utilities for the
outcomes are ordered in the same way as the dollar values and assuming
that the game is played under conditions of complete information (i.e., each
player knows the rules, the dollar payoffs for both, von
Neumann-Morgenstern utility images of both players’ payoffs, the fact that
the other player knows all of this, the fact that the other knows that he
knows, etc.) what should Player 1 do? More directly, for what values of x
and y would you as Player 1 be indifferent between your two choices? For
92
0022-0531/81/040092-09$02.00/O
Copyright 0 1981 by Academic Press. Inc.
All rights of reproduction in any form reserved.
 GAMES OF PERFECT INFORMATION 93

(0,O) (f MILLION, II

\/

FIG. 1. Example 1.

most of those to whom I have posed this question, the answer depends on
(among other things) what is known about Player 2. When I reply that
Player 2 is known to be a very sensible person whose identity will always
remain secret and to whom Player l’s identity will always remain secret, the
answer invariably involves some x-value strictly less than 1 million.
I interpret this answer to mean that when the stakes are sufficiently high
people will usually not act as though the hypothesis that other people are
utility maximizers is absolutely dependable. Moves of other players simply
can never be anticipated with certainty. If that is the view of real game
players, what are its implications for the theory of games in which nearly all
of the models and solution concepts seem to be based on the idea that all
players can be assumed to be acting in their own best interests?
One might respond that the hypothesis of complete information is too
strong to be ever realized in life, that utility payoffs of others (for example)
can never be known with certainty and that l’s selection of Left over Right
when x ( 1 million simply indicates a natural equilibrium strategy in the
imagined incomplete-information game. My feeling is that while probably no
games are played under conditions of complete information in the real world,
there are some situations in which the information is complete enough that
the situations can reasonably be modeled as complete-information games;
and that for any extensive-form model it should be possible to imagine that
and reason as though the model corresponded closely enough to some
hypothetical game situation. In my view, therefore, appealing to the lack of
realism of the complete-information game would not be sufficient to explain
preferences of Left over Right in Example 1.
It might also be pointed out that the Left strategy for Player 1 is in fact
part of a Nash equilibrium whenever x > 0 and that game theory therefore
admits considerations of the aforementioned sort. But l’s Left is only in
equilibrium against 2’s Left strategy; a situation I interpret to mean that 1
believes 2’s (possibly nonexplicit) threat to play Left, a threat which can
only be believed if 2 is possibly not a utility maximizer. Note that this
strategy combination is in equilibrium whenever x > 0 and is otherwise
insensitive to the magnitude of x. (Nash equilibria containing threats which
do not get carried out are examples of what are called imperfect equilibria.
In some recent game-theoretic work it has been suggested that attention
should focus only on perfect equilibria or special subsets of the set of perfect
equilibria. See [ 3, 7, 81.)
94 ROBERT W.ROSENTHAL

The point of view which I shall adopt is that Player 1 should consider his
Right strategy to be a risky one and that the risk should be treated in the
same way that it would be if Player 2 were replaced by chance in this game.
That is, Player 1 should arrive at some subjective probability distribution
over 2’s strategy choice and should play his Right strategy if the expected
utility to him from the lottery exceedshis certain utility for his Left strategy.
Moreover, I will argue in Section 3 that such an approach is appropriate for
the analysis of all finite gameswith perfect information. Note that the uncer-
tainty should concern Player 2’s choice, not his utilities for the various
outcomes. Although this distinction is not significant in Example 1, where l’s
positing appropriate uncertainty over Player 2’s payoffs and assuming he
does not carry out threats can be seento be equivalent to l’s positing uncer-
tainty over 2’s choice, the distinction does become significant in multi-move
games.

3. PERFECT-INFORMATION GAMES

A finite n-person extensive-form game with perfect information consistsof:

1. A tree with a distinguished node (called the origin). Node b follows
a if there is a path from the origin to b through a. A terminal node is a node
with no followers.
2. A partition of the nonterminal nodes into sets labeled 0, 1,2,..., n. A
node in set i > 0 corresponds to a move by Player i. A node in set 0 is a
move by chance.
3. For every node in set 0, a probability distribution over the
immediate followers of that node. The interpretation should be clear.
4. For every terminal node, a vector in R”. The ith component of the
vector is interpreted as Player i’s von Neumann-Morgenstern utility for the
outcome represented by the terminal node.

For gameswith perfect information, no extra informational complications

arise. When a player is to choose an immediate follower from any of his
nodes, he knows for which node he is making the choice. Thus, he knows
perfectly all the choices which have been made at preceding nodes. There are
no possibilities of secret or simultaneousmoves. The assumption of complete
information is that everything about the extensive form is known to all the
players of the game.
A pure strategy for Player i (>0) is a function which assignsto each of
Player i’s nodes one of its immediate followers. A Nash equilibrium in pure
strategies is a pure strategy combination (one strategy for each player) with
the property that no player can improve his expected utility payoff by
 GAMES OF PERFECT INFORMATION 95

switching to some other pure strategy if the strategies of all other players are
held fixed. It is well known (see, e.g., [ 11) that every finite perfect-
information game possesses at least one Nash equilibrium in pure strategies
and that such an equilibrium can always be constructed by working
backward through the tree, selecting a best (defined inductively) immediate
follower at every personal node and taking expected utilities at every chance
node. Nash equilibria obtained in this way will be termed principal
equilibria. Nonprincipal pure-strategy equilibria may exist, as has already
been observed in Example 1. Nash equilibria in randomized strategies (i.e.,
probability distributions over pure strategies), which have been widely
studied, are of lesser interest for games with perfect information; since there
seem to be no intuitive reasons to randomize consciously when the outcome
of the randomization will be evident to all players who have subsequent
moves.
Although a player may not consciously intend to randomize and may in
fact intend to reason inductively in order to select his strategy in a game
with perfect information, it may be that for some reason his intention is not
realized. If players with preceding moves realize that this may be the case,
they would be wise to consider it when making their choices.
Why might players believe that players moving subsequently might deviate
from their principal-equilibrium strategies? (Incidentally, I do not rule out
the possibility that a player may view his own subsequent choices as
random.) One possibility is a mistake. Another is that subsequent players
may decide not to analyze the tree completely because the time and effort
required are not justified by the potential gains. (In this case it could be
argued that the game in question is only part of the more complicated
problem in which decisions about time and effort are also made. Such
considerations quickly lead to unmanageably large models and even
questions about infinite regressions. It seems far simpler to study the game of
perfect information with appropriate allowance made for moves which are
not part of principal equilibria.) A third possibility (though related to the
previous one) is that the player in question recognizes that the game is
actually played under conditions of incomplete information, but that a model
incorporating the uncertainties would be unmanageable for him. To keep
matters simple, he assumes a game with complete information but modifies
his principal-equilibrium calculations in a manner he feels to be intuitively
appropriate. Still a fourth possibility is the belief that other players may be
acting according to some other view of what is rational. This view may not
known with certainty.
Consider Example 2 (see Fig. 2), where the origin is at the left. If Player 2
is actually called upon to move in this game, he has clear evidence that
Player 1 is not playing according to any Nash-equilibrium strategy. Should
he ignore this evidence, or should he consider the possibility that Player 1

h42/25/1-7
96 ROBERT W. ROSENTHAL

FIG. 2. Example 2.

might not act in his own (and hence Player 2’s) interest at his next move?
(See [2, pp. 80-811 for a related discussion.)
Because the various reasons given for imputing estimates of the subse-
quent move probabilities involve a considerable degree of subjectivity, I shall
not suggest a universal scheme whereby players may logically deduce these
estimates. Rather, I appeal to the decision-analysis paradigm in which each
player arrives at subjective probability estimates by considering what
objective lotteries he views as indifferent to the move in question by the
player in question (see, e.g., [5]), where the outcomes of the objective
lotteries are the same, respectively, as the outcomes of the personal moves.
This is admittedly a major incomplete aspect of the method of analysis
which I am proposing. On the other hand, the arguments which I put forth in
this paper may be viewed as an appeal to include, at least in situations with
perfect information, whatever aspects of a real situation are left unmodeled
by the extensive form. Without making very special assumptions about the
situation in question, it is difficult to imagine that a useful theory is possible.
Indeed, in Example 2 it would seem to be a hopeless task to say anything
sensible about Player 2’s decision without extra unmodeled information. (It
can be argued that all aspects of a situation including, for example, perceived
psychological atitudes of other players should be expressed in the extensive
form. Although possible in principle, this would seem prohibitively
complicated in practice.)
In some cases, it is possible to gain new qualitative insights from this
approach without being too specific about subjective probabilities. Consider
Example 3 (see Fig. 3). In Example 3 the leftmost node is the origin, and
each player’s choices at each node are called Right and Down, respectively.
All the Nash equilibria in this example involve Player 1 picking Down on his
first move. The principal equilibrium has both players selecting Down on
their first move. Suppose now as an illustration that both players have iden-
tical views about each others’ propensities to deviate from principal Nash
behavior and that the views of both are common knowledge (i.e., each knows

FIG. 3. Example 3.
 GAMES OF PERFECT INFORMATION 97

that the other knows, etc.). These views are that the probability assigned to
the better of two choices, at any stage, is min( 1,O.S + 0.40), where D is the
utility difference between the two choices. (Of course, a von
Neumann-Morgenstem utility construction does not fix a scale for
measurement. One must imagine that the coefficient (0.4) is the one that
applies for the utility scale chosen in the representation of the game given
above.) Thus, at the last node, although Player 2 would (no doubt) like to
choose Down with probability 1; it is predicted that he will choose Down
with probability 0.9 and right with probability 0.1. At the penulimate node,
therefore, Player 1 can expect utility of 8 from Down and 7.3 from Right. He
would therefore also elect Down but is predicted as choosing Down with
probability 0.78 and Right with probability 0.22. At the preceding move 2
expects 9 from Down and (0.78(B) + 0.22(11)) from Right if he believes
himself incapable of subsequent error or (0.78(B) + (0.22)( 10.9)) from Right
if he ascribes the common view of error to his own moves. Under either
hypothesis, assuming common knowledge, the probabilities of Right continue
to increase as the induction proceeds back through the tree, becoming one
before the origin is reached.
Although it might be difficult to believe in the specific construction for the
subjective probabilities given above and also in the common-knowledge
assumption; it is clear that if all players consider the subjective probabilities
to be sufficiently dependent on the utility differences and if the form of these
considerations is common knowledge (even if the specifics of the other
players’ functions are not common knowledge) then there is a class of games
for which this type of compounding effect occurs. As I shall argue in the
next section, some important economic models fall into such a class. Of
course, even if subjective probabilities are not related to the size of the utility
differences but depend on other features, the moves prescribed by a decision-
analysis approach may differ from those of every Nash equilibrium.
It is the perfect-information assumption which enables us to use the logic
of backward induction. Still, it is possible that similar reasoning could be
used (to some extent at least) in extensive-form games with imperfect infor-
mation. If we believe that Nash equilibria are not reasonable or realistic in
perfect-information games, the same could be true for the same reasons in
games with imperfect information. If a game is “deep” (in the sense that
some paths from the origin to terminal nodes contain many arcs) but only a
small part of the tree is affected by information imperfections, then it might
be possible to look at only those small parts of the tree as subproblems in
traditional game-theoretic ways and to look at the grand problem more
according to the decision-analytic mode.
98 ROBERT W.ROSENTHAL

4. PREDATORY PRICING AND THE CHAIN-STORE PARADOX

Consider a single-product industry containing a single firm setting a

monopoly price and making corresponding profits. If another firm enters the
industry, there is enough business for both to survive, but profits for the
monopolist will be severely cut. A possible strategy for the monopolist is to
cut price whenever a new firm enters to a level low enough that the entrant
will sustain losses in every period. This strategy presumably reduces the
monopolist’s profits even further for those periods when the low price is in
effect (perhaps even causing lossesfor the monopolist). If the monopolist
drives the entrant out, and if no new firm enters, however, the monopolist
can resume his high-price, high-profit position. Price cutting to drive other
firms out of an industry or to deter entry is called predatory pricing and is a
subject of considerable interest to economists. The proper role of anti-trust
procedures in controlling predatory pricing is a matter of some debate
among both economists and law makers (see, for example, [4,6]).
The question of whether or not and to what extent the threat of predatory
pricing is likely to succeed in deterring entry under various conditions is a
little-understood but important aspect of the debate. It is also a question with
a strong game-theoretic flavor.
In [7], Reinhard Selten describes an example of a game, called the chain-
store paradox (version one), which seemsto capture the essentialelementsof
the deterrence question. In Selten’s game there is one chain store, Player A,
with branches in 20 towns. In each town there is one potential entrant i
(i = l,..., 20). Furthermore, for each i there is only one date at which he can
enter. These dates are sequential with Player l’s date first, then 2’s, etc. If
Player i enters on his date, A with full knowledge has the choice of whether
to respond aggressively or cooperatively. The payoffs to Players i and A
(although it is only a partial payoff to A who is concerned with the sum of
his 20 payoffs) are given below:

A’s decision
i’s decision in period i i’s payoff A’s payoff

In Cooperate 2 2
In Don’t cooperate 0 0
out - 1 5

At Player i’s entry date, he is fully aware of all previous moves.

There are two features of this game which deserve comment with respect
to the predatory-pricing interpretation. One is the assignmentof unique entry
dates to each firm. This assumption seemsto have been made to keep the
 GAMES OF PERFECT INFORMATION 99

specification of the game as simple as possible and does not appear to distort
seriously the deterrence question. The second feature is the finite-time
horizon assumption. Firms are often assumed to plan for the infinite horizon
(with discounting where appropriate). Furthermore, it is well known that in
many repeated games there is a qualitative difference between the structure
of the set of Nash equilibria in the finitely-repeated case (even with a
sufficiently large number of repetitions) and that in the infinitely-repeated
case. Since all of the discussion thus far has been directed at finite-move
games, it would be improper to argue that any insights gained by reasoning
inductively on the finite, chain-store game are valid for a situation in which
players perceive an infinite horizon. But it would also seem that there should
be very little difference between what a firm would actually do early in a
large-but-finite horizon predatory-pricing situation and what that same firm
would do in a similar infinite-horizon situation.
The thrust of Selten’s analysis of the chain-store game is that while the
principal Nash equilibrium of the game requires all Players i to enter and A
to cooperate, one’s intuition suggests perhaps that the threat of the aggressive
response might suffice to deter entry, at least until late in the game. Selten
then proceeds to an interesting discussion of the game from many different
points of view.
The point which I wish to make is that there is a considerable similarity
between the chain-store game and Example 3. If the discussion of the previous
sections has been convincing, then the approach advocated here for games
with perfect information may be of use in resolving the “paradox” and in
analyzing predatory pricing and other deterrence situations.
Let us look at the chain-store game under the assumption that the size of
utility differences affects subjective probabilities in the same direction as in
Example 3. At each of Player A’s last nodes he gains two units of utility by
picking the cooperative response. Suppose, however, that Player 20 views
that choice as less than certain. At 20’s move, therefore, his expected utility
gain is less than one from choosing In. Player A therefore has an expected
gain of less than 2 if he responds aggressively to 19’s choice of In. If 19
considers the aggressive response to him to have higher probability than 20
considered the aggressive response, then he has even less to gain than 20 by
choosing In. As in Example 3, the effects may quickly compound if the
probabilities are appropriate monotone functions of the expected utility
differences. The reader will be spared any further details. It should be
obvious that for a wide class of methods for forming subjective probabilities,
the outcome of the game will be that entry is deterred for a large portion of
the potential entrants.
100 ROBERT W.ROSENTHAL

REFERENCES

1. H. W. KUHN, Extensive games and the problem of information, in “Contributions to the

Theory of Games, Vol. II” (Kuhn and Tucker, Eds.), Annals of Mathematics Studies,
No. 28, pp. 193-216, Princeton Univ. Press, Princeton, N. J.
2. R. D. LUCE AND H. RAIFFA, “Games and Decisions,” Wiley, New York, 1957.
3. R. B. MYERSON, Refinements of the Nash equilibrium concept, Inrernnr. J. Game Theory 7
(1978) 73-80.
4. R. A. POSNER, “Anti-Trust Law: An Economic Perspective,” Chap. 8, Univ. of Chicago
Press, Chicago, 1976.
5. H. RAIFFA, “Decision Analysis,” Addison-Wesley, Reading, Mass., 1968.
6. F. M. SCHERER,“Industrial Market Structure and Economic Performance,” Chap. 8, Rand
McNally, Chicago, 1970.
7. R. SELTEN, The chain-store paradox, Theory and Decision 9 (1978), 127-159.
8. R. SELTEN, Reexamination of the perfectness concept for equilibrium points in extensive
games, Intern. J. Game Theory 4 (1975), 25-55.