The folk theorem for repeated games - Semantic Scholar

28 downloads 0 Views 847KB Size Report
The Fudenberg and Maskin (1986) folk theorem for discounted re- peated games assumes that the set of feasible payoffs is full dimensional. We obtain the same ...
MIT LIBRARIES

9080

3

00832

6610

mH mBka

BHNsW m

I

£5$$$5&»hs^

mm

•-

-----

•-

sa«B&

«

ffi»

it

W&m mm m mmmm, ^P^nBH

'

-

1

-

'.'

Illllllll Jft'*filsi v* "'? . 11111

11111

;-

m sm m

:

iH

'''
R. The game G satisfies NEU if for i and j, there do not exist scalars c, d where d > such that 717(a) = c + dnj(a) for all a G A. Let Mi be the set of player i's mixed strategies, and let = x"=1 M,. Abusconsider a finite n-player

game

in

where 7T,-

:

M

ing notation, (jii,...,Hn)

(n



= we write 7r,(/x) for i's expected payoff under the mixed strategy £ M. For any n-element vector v = (ui,...,v„), the corresponding /j,

l)-element vector with element v

{

missing

is

denoted

v_;.

Let

=

7r* (//_,)

be player i's best response payoff against the mixed profile /i_,. Denote by m = (m\, ,m'n ) e a mixed strategy profile which satisfies m'_ € argmin _ 7r*(^_,) and m\ e argmax^.Tr^/i^mL,). In words, ml,- is an (n — l)-profile

max,,; 7T;(a;,/x_,-) 1

.

/J

3

.

.

M

t

j

In the context of

Nash equilibrium and

its

refinements, only single person deviations need be

deterred. Hence, the focus on pairs (an original deviant as opposed to coalitions of players.

and a

single subsequent deviant) of players

mixed

of

=

We

is

An

2.2

player

i,

and m\

:

is

have adopted the normalization

e M}, so that the set of feasible and = F* {w 6 F w > for all i}.

co{7r(//)

payoffs

minimax

strategies which

being minimaxed.

=

7T,(m')

for all

when

i

i.

Let

F

(strictly) individually rational

fi

:

a best response for

{

Equivalence Result

NEU has two quite powerful and equivalent representations that are developed in the lemmas below. For j ^ let F,j, denote the projection of F on the i-j coordinate i,

plane and dimF.j the dimension of

The

Definition (a) dimF.-fc all

^

k

i,j.

=

1

profiles.

Then

diva

^

or

i,

dimF,j

(b)

NEU holds

Suppose that

F

=

1 for

and

that

is

some j

^

and

i

i,

either

dimF^ =

2 for

a line with negative slope.

no player

is indifferent

over

all

action

satisfies the projection condition.

Since no player

F^ >

satisfies the projection condition if for all players

k

all

Furthermore, in the latter case, F^

Lemma Proof:

F

set

2 for

Fij.

1 for all i

^

is

j.

indifferent over all possible action profiles,

Now

suppose the dimF.y

=

1

dimi* ,*

=

1

it

follows that

some j ^ k. and k) implies

for

NEU applied to the payoffs of players i and j (and similarly players i that the payoffs are perfectly negatively correlated. This in turn implies that the payoffs of players j and k are perfectly positively correlated.

k have equivalent payoffs, in violation of

Definition

Lemma ble

The

vectors {v 1

2 Suppose that

F

,

.

.

.

,v

n }

That

is,

players j and

NEU.



Vj

3

k, there exist v].

.

some feasible payoff

For each player

i,

order the

^ k) in increasing size (break ties arbitrarily) from and assign these ordered vectors strictly decreasing weights 6h,h = 1,2, .. ,n(n — 1), summing to one. Let v' be the resulting convex combination of the payoff vectors vjk Note that in defining the u"s we use the same weights 6h for all i. Then by construction, v\ < vj for all i ^ j, establishing payoff asymmetry. ^ Finally, it is straightforward to see that the existence of asymmetric payoffs implies NEU (and rules out universal indifference for any player). In other words, NEU, the projection condition, and the existence of asymmetric payoffs are equivalent assumptions (absent universal indifference). n(n

1)

payoff vectors v

the point of view of player

(Vj i,

.

.

THE MAIN THEOREMS

3.

We

game with

will analyze the infinitely repeated

discounting that

is

associated

with the stage game G. We assume condition his action in period t on the past actions of all players. In addition, we permit public randomization. That is, in every period players publicly observe the perfect monitoring; that

random

realization of an exogenous continuous

variable

is

each player can

and can condition on

its

outcome. This assumption can be made without loss of generality in the infinitely repeated game; a result due to Fudenberg and Maskin (1991) shows explicitly how public randomization can be replaced by "time-averaging" (see also, Sorin (1986)). ,a«, Denote by o, = (a,i, .) a (behavior) strategy for player i and by 7r,«(a) his .

.

.

.

expected payoff in period function

is

.

given the strategy profile a.

t

Each player

his infinite-horizon average expected discounted payoff (1

under the (common) discount factor

— 5)

i's

objective

£2° ^ t7r «t( a )

Let V(5) denote the set of subgame perfect

5.

equilibrium payoffs.

3.1 Sufficient

We

Conditions for a Folk Theorem

will establish here that

that for a wide class of games,

NEU it is

and later on a necessary condition for the folk theorem

sufficient for the folk theorem,

is

also

to hold.

Theorem game

is

Under NEU, any (strictly) individually rational payoff in the stage a subgame perfect equilibrium payoff of the infinitely repeated game when 1

players are sufficiently patient. That


x\


j (payoff asymmetry). For small enough e

(strict individual rationality).

itj

we must have

i

define

3

the definition of w* and v\

then x x-

=

r?),/?2

G F}. Now

/W + fov* + P u

=

x*

where

(u_;,u,)

:

even

if v\

>

u,-.

0,

Finally, for small

5 6 '

Strategies Let a (respectively,

game

payoff

with m',

The (for

is

a')

denote the publicly randomized action vector whose stage

u (respectively, x

minimaxing player

i.e.

%

Further, let

).

and

i,

recall

v'

be the payoff vector associated

from section

2.1 that u)

=

0.

strategy vector that will generate the target payoff u as an equilibrium payoff

appropriate choice of parameters) can be defined in Markov strategy terminology

as follows:

1.

When a'{ 7^

2.

3.

a,-

When and

in "state" u, play a.

and

a'_ :

-

=

a_;,

go to "state"

a'_j

=



When

a'.

in state a'_j,

x

: ,

play

go to "state"

first

q)

If

^

m'

proceed to "state"

x'.

the observed action vector

a' satisfies

a'j

7^

but

a'j

y?. Else, stay in "state" x'.

In words, the strategy says:

the

Else, stay in "state" u.

m]_j, go to "state" v*. Else, with probability q stay in "state" v\ (1

=

v'.

a' satisfies

in "state" v', play m'. If the observed action vector a' satisfies a'

while with probability

a'_j

the observed (mixed) action vector

If

and continue to play this action till i). Then, minimax player i for one the event of no observed deviation) continue

Start with a

single-player deviation (say by player

period (with probability one) and the minimaxing with probability

minimaxing and play

a' until

(in

With

q.

the remaining probability, terminate the

further deviations.

Treat players symmetrically and

subject every single player deviation to this (stochastic) punishment schedule. 4

in

This construction together with

F* C K". Then a simple

Lemma 2

choice of x'

is

has a nice geometric intuition. Consider u as a point F* with the smallest i coordinate on the

the point in

Be (u) about u, for small enough e > 0. Indeed, Lemma 1 implies that the projection of B £ {u) n F* onto any two players' coordinate plane is either an ellipse or a line with negative slope. e-sphere

and x k lie at different locations on the ellipse, while in the second they reside opposite ends of a line segment. 5 A referee's comments helped clarify the specification of the /?,'s above. 6 x" We remark in passing that there is a minor error in the construction of the vectors x Fudenberg and Maskin (1986): They implicitly assume (as pointed out to us by Peter Sorensen MIT) that because u £ F*, it does not lie on the lower boundary of F. This is false. One way patch up their construction is suggested by a closer reading of footnote 4.

In the at

first case, x*

,

in

of to

.

.

,

Choice of Parameter The only parameter

in the

feasible payoff for player

above strategy

Choose q to

i.

rrj

>

by equation

(1),

we can

q.

Let

6,

be the best

satisfy:

< l^-x\ \-q

bi

Since

the probability

is

(4)

find such q



(0, 1).

Verification of Equilibrium

We show that no one-shot deviation by any player from any state is profitable. Hence the strategy proposed is unimprovable and consequently a subgame perfect equilibrium.

State

v*:

Player

(discounted average) payoff in state v\ denoted

"lifetime"

i's

Li(v'), satisfies

=

Litf)

S (qLiivj)

+

(1

- q)x])

so that

Ate')

Note that

—>

=

f^*- >

0.

(5)

as 6 f 1- Player i will not deviate in state v' since the maximal payoff to one-shot deviation is SLi(v ). Player j ^ i will not deviate for high 5, since his maximal payoffs are bounded by (1 — S)bj + 8Lj(y?), which is less L)(v')

x\,

l

than

by equation

x^,

State x

1 '.

From

(3).

the definitions

and conformity

to one-shot deviation

[{I

-

it is

5)bi

is

is

that (6)

bounded above by

is

+ SL^))- x\ =

where we have substituted from defining q

clear that the difference in the lifetime payoffs

is

(5).

{I

-

5)

^

i

l

+ 5-6q \ (6)

An immediate

strictly negative for all

immediate that players j

(

implication of inequality (4) 5 close to 1. Since x > x-j, j ^ i, it x

-

do not have a profitable one-shot deviation

either.

State u: Since, by target payoff domination (2), u, > x\, the arguments above also imply that (1 — *(a )mi(a

)+

t

+ 1-9(8),

ay.

choices ay and

iff

a'-,

is

zero

between his action minimaxing strategy mlindifferent

State

The arguments

c'-':

+p

ij

i

(aj )c

f

E &*(«*)«*(«*) -py

(«j)

XJ

(

g)

term and the two others involving p ,J'(aj), is Hence, the difference in lifetime payoffs, from the

except the

independent of the choice

)e?

kjiij a k

I

Everything in

fc

first

equation

(7) is satisfied.

In that event, player j

choices. In particular, a best response

that no one-shot deviation

is

is

is

to play the

profitable in state

7

c'-

is

arguments that have established that no one-shot deviation is proffrom state x'. (Note that these last arguments do not change at all since no mixed strategies are played in state x f ). identical to the

itable

3.2

Necessary Conditions for the Folk Theorem

We

turn

for all j}

now

to the necessity of

be player

We

f's

NEU.

Let

/,•

=

min{vi

|

v



F

and

Vj

>

worst payoff in the set of weakly individually rational payoff

minimal attainable payoff. 7 The necessity of payoff asymmetry is shown for games in which no two (maximizing) players can be simultaneously held at or below their minimal attainable payoff. In stage games where every player's minimal attainable payoff is indeed his minimax payoff (/, = 0), the condition stated below (which uses the term minimuing) is equivalent to the restriction that no pair of players can be simultaneously minimaxed. 8 Say that a subset of players S C {1, n} can be simultaneously minimized if there exists a strategy profile, fi such that 7r*(^,) < /,- for all i € S. vectors.

will refer to

/,•

as player

i's

.

.

,

.

Definition G satisfies no simultaneous minimizing [NSM] can be simultaneously minimized.

Under

this

Theorem

if

no two players of

G

assumption we obtain a complete characterization. 2 Suppose

NSM

obtains.

Then

NEU

is

necessary for the conclusion of

the folk theorem. 7

Note how

8

A game in

from w\. which the minimally attainable payoff is not the minimax payoff for every player is the two-player game specified by the following payoffs: co[(0, -1), (1, 0), (2, 1)] with (0, —1) a payoff at which player 1 is minimaxed and (1,0) a payoff where player 2 is minimaxed. /, differs

Proof:

To

establish necessity,

that for

all

i

^ j, x\
/,-, as 6 ->

1.

By

By

the folk theorem hypothesis,

playing his myopic best response in period one and

— 7T*(7l,(