Professional Documents
Culture Documents
Abstract. Selten's concept of perfect equilibrium for normal form games is reviewed, and a new
concept of proper equilibrium is defined. It is shown that the proper equilibria form a nonempty
subset of the perfect equilibria, which in turn form a subset of the Nash equilibria. An example
is given to show that these inclusions may be strict.
1. Introduction
The concept of equilibrium, as defined by Nash [ 1951 ], is one of the most impor-
tant and elegant ideas in game theory. Unfortunately, a game can have many Nash
equilibria, and some of these equilibria may be inconsistent with our intuitive notions
about what should be the outcome of a game. To reduce this ambiguity and to elimi-
nate some of these counterintuitive equilibria, Selten [ 1975] introduced the concept
of a perfect equih'brium. In this paper, we shall define the notion of a proper equilib-
rium, to further refine the equilibrium concept. We will show that, for any game, the
proper equilibria form a nonempty subset of Selten's perfect equilibria, which are
themselves a subset of the Nash equilibria.
To see how these counterintuitive equilibria can arise, consider the game in Figure
1.
F~ : Hayer 2
131 /32
Hayer 1
Ot2
Fig. 1
1) The author acknowledges helpful conversations with Ehud Kalai and David Schmeidler.
2) Prof. R.B. Myerson, Graduate School of Management, Northwestern University, Nathaniel
Leverone Hall, Evanston, Illinois 60201, USA.
74 R.B. Myerson
There are two Nash equilibria in this game, (al, 31 ) and (a2, f12), because in each case
neither player can improve his payoff by unilaterally changing his strategy. But it
would be unreasonable to predict (t~2, f12) as the outcome of this game. If player 1
thought that there was any chance of player 2 using fil, then player 1 would certainly
prefer al. Thus (a2, f12) qualifies as an equilibrium only because Nash's definition
presumes that a player will ignore all parts of the payoff matrix corresponding to
opponents strategies which are given zero probability. The essential idea behind
Selten's perfect equilibria is that no strategy should ever be given zero probability,
since there is always a small chance that any strategy might be chosen, if only by
mistake. So, in our example, al and/}~ always must get at least an infinitesimal prob-
ability weight, which will eliminate (a~, ~2) from the class of perfect (and proper)
equilibria.
Although Selten [ 1975] initially developed his theory of perfect equilibria for ex-
tensive from games, we will limit our attention in this paper to normal form games.
P is an n-person game in normal form if
F = (S 1 . . . . . S n; U 1 , . . . , Un) (1)
where each S i is a nonempty finite set, and each Ui is a real-valued function defined
on the domain $1 X $2 X . . . X S n . We interpret (1, 2 . . . . . n} as the set of players
in the game. For each player i, S i is the set of pure strategies which are available to
player i. Each Ui is the utility function for player i, so that Ui (sl . . . . , Sn) would be
the payoff to player i (measured in some yon Neumann-Morgenstern utility scale) if
(s l , . . . , sn) were the combination of strategies chosen by the players.
For any finite set M, let A (M) be the set of all probability distributions over M.
Thus:
So A (Si) is the set of randomized or mixed strategies which player i could choose
in P.
It is straightforward to extend the utility functions to the mixed strategies, using
the formula:
n
Z • ...x s (i_I'l10 i (Si)) ~ (Sl . . . . . Sn).
(ol . . . . . On) = (s, ..... sn)~s, (3)
That is, U/(ol . . . . . On) is the expected utility which player/would get if each player
i planned to independently randomize his strategy according to o i.
n
Suppose that (ol, 9 9 a n) E z• A (Si) is the combination of mixed strategies
Ref'mements of the Nash Equilibrium Concept 75
which the players are expected to use, but suppose that player ] is considering whether
to switch to one of his pure strategies s7 ES/. Let l~ (s/* I 01, 9 9 9 on) be the expected
utility f o r / i f he makes this switch and all others remain with their o i mixed strategies,
so that:
1 if sj=s*.
where oI (si)= !
0 if si :/: s/*.
We say that sl is a best response (in pure strategies) to 0 1 . . . . . On) for player] iff
t
V]. (s] [ o l .... ,On)----- max V/(sila, ..... On).
s'.~S.
I I
!
ui(o, ..... o,,)>~uj(o, .... ,~,} ..... o) w, voj ~,~(s?. (5)
O1 ~ 9 . . ~O n )-- O1, 9 . . ~ , . . . , On
=. ',%
z ') %
z o~(s? o/'(s}) (~ (~j I o,, .... %) - ~. (sj'1 ol, .... o,)).
The proposition then follows easily from the definition of a Nash equilibrium.
3. Perfect Equilibria
For any finite set M, let A~ (M) be the set of all probability distributions on M
which give positive probability weight to all members of M. So, for any player i,
The members of A~ (Si) are called totally mixed strategies for player i.
Heuristically, a perfect equilibrium could be described as a combination of totally
mixed strategies (ol, 9 a n) such that, for any player/" and any pure strategy
st ES/, if s] is not a best response to (ol . . . . . On) for] then o] (s~) should be infinitesi-
/ . . . ! . .
many small. Thus, in a perfect equilibrium a player cannot ignore any point m
$1 X . . . X S n as "impossible" when he computes his best responses, since every one
of his opponents' pure strategies has a positive probability; and yet no player wants
to make any substantial change in his strategy, since each player is already putting
almost all his probability weight on his best responses.
To make these ideas about "infinitesimally small" probabilites precise, let e be any
small positive number. Then we define an e-perfect equilibrium to be any combination
of totally mixed strategies (ox . . . . . an) E A~ ($1) X . . . X A~ (Sn) such that:
(Notice that every q will have to be in A~ (Si), but o i need not be, since the closure
of A~ (Si) is A (Si).)
Selten has shown that a perfect equilibrium must be a Nash equilibrium [see ].,emma
9 in Selten]. The proof follows easily from the fact that ~ (si I ol . . . . . On) is contin-
uous in (oa . . . . . On). So a perfect equilibrium must satisfy the conditions of our
Proposition 1, because it is the limit of e-perfect equilbria which satisfy (7).
All perfect equilibria are Nash equilibria, but the converse does not hold. For
example, consider the game Fl defined in Figure 1. The only e-perfect equilibria are
those pairs of totally mixed strategies in which ol (~t2) ~<e and o2 (~32)< e. Thus the
Refinements of the Nash Equilibrium Concept 77
only perfect equilibrium is (0/1,/31 ) - or, more precisely, it is (o~', o~'), where
o~ (cq) = 1 and o~ (/31) = 1. (0/2,/32) was a Nash equilibrium, but it is not a perfect
equilibrium.
4. Proper Equilibria
r2 Hayer 2
(u1, v2)
~2 /33
0/1 1, 1 O, 0 -9,-9
Player 1 0/2 O, 0 O, 0 -7,-7
Fig. 2
As in our first example, (0/1, fll ) would seem like the obvious outcome for this game.
There are three Nash equilibria, and all are in pure strategies (or, more precisely, in
mixed strategies which assign all weight to one pair of pure strategies); these equilibria
are (0/1, fll ), (0/2, f12), and (0/3, f13).
Of these three Nash equilibria, (0/3,/33) is not perfect, but (0/1, fll ) and (0/2, f12) are
both perfect equilibria. To check that (0/2, f12) is a perfect equilibrium, define e~ and
o~ by
and observe that (o~, o~) forms an e-perfect equilibrium. (For example
V,(0/11 ol,e 02)
e _ - -- 8e, Vl (0/2 I o~, a~) -- 7e and VI (0/a I e~, 05) = -- 7 - 2e, so
=
0/2 is best response. As required, o~ (0/t) ~< e and o~ (0/3) ~< e.) Then, as e ~ 0, these o~
and o~ converge to the strategies which select 0/2 and/32 with probability 1. In effect,
adding the 0/3 row and/33 column has converted (0/2,/32) into a perfect equilibrium
even though 0/3 and/33 are obviously dominated strategies.
To discriminate between (0/1,31) and (0/2,32) in this example, we need a new
refinement of the equilibrium concept: the proper equilibrium.
For our general normal form game, we define an e-proper equilibrium to be any
combination of totally mixed strategies (el . . . . . On) E A ~ (S~) • . . . • A ~ (Sn) such
that:
78 R.B. Myerson
if ~ ( s ] [ ol . . . . . On)<I~(s}lol,...,On)theno/(sj)<e'o/(s}),
As with perfe:t equilibria, a proper equilibrium need not be totally mixed; it must
only be the limit of totally mixed e-proper equilibria.
Proposition 2: For any game in normal form, the proper equilibria form a subset of
the perfect equilbria, which in turn form a subset of the Nash equilibria. These in-
clusions may both be strict inclusions.
Theorem: For any normal form game P (as in (1)), there exists at least one proper
equilibrium.
Proof: We show first that there exists an e-proper equilibrium, for any e, 0 < e < 1.
Let m = max I S i I. Given e, let ~ = 1 . em" For any player], let
i m
(sp = (S? l (s? 8, V sj Esi .
Observe that A* (Sj) is a nonempty compact subset of A ~ (S]). We now define a point-
n
to-set map t7/:i=• A* (Si) =>A* (Sj) by:
References