Professional Documents
Culture Documents
00
00
Simon RubinsteinSalzedo
Fall 2005
0.1
Introduction
These notes are based on a graduate course on game theory I took from Professor
Rod Garratt in the Fall of 2005. The primary textbooks were Game Theory by Drew
Fudenberg and Jean Tirole and Game Theory for the Applied Economist by Robert
Gibbons. Also recommended were A Course in Microeconomic Theory by David
Kreps and The Winners Curse by Richard Thaler.
Chapter 1
Games in Strategic Form
We seek to model and make predictions about the outcome in situations where
1. each decision maker, in a group, makes choices (or takes actions) independently
and simultaneously.
2. the payoff to each player is determined by the profile of actions taken.
L
4,11
1,3
R
4,1
1,13
A
7,7
0,8
B
8,0
9,9
1.1
u
m
d
C
5,1
8,1
9,6
R
6,2
3,6
2,8
The column player ought not to play C as it is dominated by R. Then the row player
ought not to play m or d as the remaining possibilities are dominated by u. Then the
column player ought not to play R as it is dominated by L. Hence (u, L) is the only
rational strategy. This process is called iterated elimination of strictly dominated
strategies. It does not always produce a solution, but the order in which we eliminate
solutions does not matter.
Now consider the following game:
u
m
d
L
2,0
0,0
-1,0
R
-1,0
0,0
2,0
This game has no strictly dominated strategies in pure strategies, however, we can
mixed strategies as well. Let p, q, r be the probabilities the row player places on
3
u
d
R
-1,0
2,0
1.2
Mixed Strategies
j=1
: 1/2
4,0
2,8
r: 1/2
3,9
1,5
si = Q
(s1 , . . . , si1 , si+1 , . . . , sI ) and ui (i , si ) > ui (si , si ) for all si Si , where
Si = jP\{i} Sj .
L
1.01,5
1,.25
R
-500,.4
.9,.2
Then u, L is the optimal play, but if the row player is not fully confident in the
rationality of the column player, then u is a very risky play.
1.3
Nash Equilibria
ui (i , i
) ui (si , i
) for all si Si .
In a Nash equilibrium, player is payoff from following i is at least as great as the
payoff she would obtain from any one of her pure strategies. In fact, this implies she
cannot do better by using some other mixed strategy.
ui (i , i ) for all i i .
ui (i , i
)=
i (si )ui (si , i
)
si Si
i (si )ui (i , i
)
si Si
= ui (i , i
),
as desired.
1.4
1.4.1
Example.
FB B
fb 2,1 0,0
b 0,0 1,2
The pure strategy Nash equilibria are (f b, F B) and (b, B).
Example. Cournot competition. Firms 1 and 2 each simultaneously choose nonnegative outputs q1 and q2 . There is an inverse market demand P (q) = 1 q, where
q = q1 + q2 . The cost is given by ci (qi ) = cqi , where c < 1 is the constant marginal
cost. ui (qi , qi ) = qi P (q) ci (qi ) = qi (1 q1 q2 ) cqi . Firm 1 solves
max(1 q1 q2 )q1 cq1 .
q1
1.4.2
3. Nash equilibria are also the result of adaptive learning. If we start with an
arbitrary point on the reaction curve and choose the optimal strategy and then
continually refine the process, we converge to a Nash equilibrium.
1.5
h
1,1
1,1
t
1,1
1,1
We can find the mixed-strategy Nash equilibrium using reaction functions. The
curves r1 (2 (h)) and r2 (1 (H)) intersect at (1/2, 1/2), so this is the Nash equilibrium.
1.5.1
The following fact is very useful for finding mixed-strategy Nash equilibria.
Claim. If ui (i , i ) ui (si , i ) for all si Si , then ui (i , i ) = ui (si , i ) for all
si Si with i (si ) > 0.
This means that you must be indifferent in order to randomize.
Proof. (By contradiction.) Suppose ui (i , i ) ui (si , i ) for all si Si , but that
ui (i , i ) > ui (si , i ) for some si Si with i (si ) > 0. Then
X
ui (i , i ) =
i (si )ui (si , i )
si Si
i (si )ui (i , i )
si Si
= ui (i , i ),
a contradiction.
7
Example.
FB B
fb 2,1 0,0
b 0,0 1,2
Suppose 1 (f b) = p and 2 (F B) = q. To get player 1 to randomize, we require
2q + 0(1 q) = 0q + 1(1 q), or q = 31 . For player 2 to randomize, we require
p 1 + (1 p) 0 = p 0 + (1 p)2, or p = 32 . The induced probability distribution
over strategy profiles is
fb
b
FB
pq
(1 p)q
B
=
p(1 q)
(1 p)(1 q)
2
9
2/9
1/9
4/9
2/9
2 + 92 1 = 32 .
L
1,0
0,1
R
1,0
0,0
(u, L) is the unique pure strategy Nash equilibrium. There is also a mixed strategy
Nash equilibrium.
Other facts.
1. Theorem. (Nash, 1950) Every finite strategic (normal) form game has a Nash
equilibrium.
2. Theorem (Wilson, 1971) Almost all finite strategic form games have a finite and
odd number of Nash equilibria.
u
d
L
1,1
0,0
R
0,0
0,0
(u, L) and (d, R) are the only Nash equilibria. Thus the above game is nongeneric. Note that one can think of a 2 2 game (for example) as a point in
R8 .
3. If a unique strategy profile s survives iterated elimination of strictly dominated
strategies, then s is the unique Nash equilibrium.
4. Only those strategies which survive iterated elimination of strictly dominated
strategies are played with positive probability in a Nash equilibrium.
8
1.6
Correlated Equilibria
Premise. Suppose that the players of a game can build a signaling device which
makes to each player a non-binding recommendation of what action to choose.
Example.
Lq R
p u 5,1 0,0
d 4,4 1,5
(u, L) and (d, R) are Nash equilibria. There is also a mixed-strategy Nash equilibrium
with p = 12 and q = 12 ; i.e. 1 (u) = 2 (L) = 12 . Moreover, ui (1 , 2 ) = 2.5 for i = 1, 2.
Denote a signaling device by which maps strategy profiles into [0, 1].
Device 1: (u, L) = (d, R) = 12 . Consider player 1. Suppose he is recommended
to play u. Then he can calculate (L | u) = (L|u)
= 1. Likewise, (R | u) = 0.
(u)
Thus his expected utility of playing u (or d) is u1 (u | u) = 5, u1 (d | u) = 4 assuming
the other player follows her recommendation! Hence player 1 optimally obeys his
recommendation. One can show that the same holds for player 2. Device 1 is therefore
a correlated equilibrium.
Device 2: (u, L) = (d, R) = (d, L) = 13 . Consider player 1. Suppose he is told
to play d. Then he calculates (L | d) = 21 , (R | d) = 12 . Thus, assuming player 2
follows her recommendation, then player 1 calculates u1 (d | d) = 2.5, u1 (u | d) = 2.5,
u1 (u | u) = 5, and u1 (d | u) = 4. Hence the device works. The expected payoff to
each player under this correlated equilibrium is 10
.
3
Device 3: (d, L) = 1. This doesnt work.
Definition. A correlated equilibrium is any probability distribution () over pure
strategy profiles in S such that for each player i and every si Si with (si ) > 0,
X
X
(si | si )ui (si , si )
(si | si )ui (si , si )
si Si
si Si
for all si Si .
Chapter 2
Games in Extensive Form
1
L
2
5
4
4
5
3
0
2
1
R
10
3
This game is one of perfect information, which means that whenever a player moves,
she knows exactly where she is in the game tree.
Alternatively, we can represent games of imperfect information using non-singleton
information sets.
1
L
2
5
4
4
5
3
0
2
1
10
R
10
3
2.1
2.2
A strategy for a player is a plan of action for every information set controlled
by the
S
player. Let Hi denote the set of player is information sets, and let Ai = hi Hi A(hi ),
where A(hi ) = A(x) is the set of actions available at node x for x hi .
Definition. A pure (behavioral) strategy for player i is a map si : Hi Ai with
si (hi ) A(hi ) for all hi Hi .
Definition. A pure (behavioral) strategy profile s of an extensive form game is a
subgame perfect equilibrium if the restriction of s to G is a Nash equilibrium of G
for every G.
2.3
2.4
Informally, no player ever forgets any information he once knew, and all players know
the actions they have chosen previously. We may add the following condition to (5)
above:
5. (d) if x h(x ), x 7 x , and i(x) = i(x ), then there is an x
b (possibly x itself)
with x
b h(x), x
b 7 x , x (x ) = xb(x ).
R
x
L
x
Example.
12
1x
b
L
2
1
R
x
2.5
[1/2]
2
Strategies
[1/2]
2.6
w2
2
0
[1/2]
1 h1
w1
x1
h2
x 2 w2
x2
0
2
0
2
y2
4
2
2
0
[1/2]
h1 1
y1
z1
h2
z2 y2
z2
2
4
2
4
4
2
w2 y 2
3,1
2,2
2,2
1,3
w2 z2
2,2
3,1
1,3
2,2
x2 y2
2,2
1,3
3,1
2,2
x2 z2
1,3
2,2
2,2
3,1
It is easy to see that A = (1 = (0, 1/2, 1/2, 0), 2 = (0, 1/2, 1/2, 0)) is a Nash
equilibrium. It is also easy to see that B = (1 = (1/2, 0, 0, 1/2), 2 = (1/2, 0, 0, 1/2))
is a Nash equilibrium. In fact, any convex combination of A and B is also a Nash
equilibrium. All these Nash equilibria are equivalent in the following sense: consider
the strategy
e1 = (1/4, 1/4, 1/4, 1/4). What is the probability that player 1s move
1 ,h1 ]
= 1/21/2
= 12 . Similarly,
is w1 at information set h1 ? Prob[w1 | h1 ] = Prob[w
Prob[h1 ]
1/2
Prob[y1 | h1 ] = 12 . Thus we can view the strategy
e1 as player 1 playing each of his
actions with equal probability at each information set. In fact, all Nash equilibrium
strategies for player 1 generate the same behavioral strategy 1 (w1 | h1 ) = (y1 |
h1 ) = 21 . In this game, there is a unique Nash equilibrium in behavioral strategies.
2.7
14
Q
Example. Suppose H1 = {h1 , h1 , h1 }. Then hi Hi (A(hi )) denotes the set of
all triples h1 ( | h1 ), 1 ( | h1 ), 1 ( | h1 )i, where 1 ( | h1 ) (A(h1 )), 1 ( |
h1 ) (A(h1 )), and 1 ( | h1 ) (A(h1 )). Typically a Nash equilibrium is more
compactly represented by a profile of behavioral strategies.
Note. Not all Nash equilibria of extensive form games are reasonable.
1 h1
L
h2
R
h2
1
9
0
0
2
1
1
9
This is a game of perfect information. The Nash equilibria are (L, ), (R, r), (L, r),
and (R, rr). Not all of these are reasonable. By backward induction, (R, rr) and
(R, r) are the solutions.
1 h1
L
R
2 h2
2
2
r
1 h1
H
3
1
2 h2
2
2
0
0
2
2
1
1
2
2
1
1
2
2
3
3
If we use the red payoffs, then there are multiple subgame perfect equilibria. A Nash
equilibrium in behavioral strategies is 1 (L | h1 ) = 1, 1 (H | h1 ) = 21 , 2 (r | h2 ) = 1,
2 (h | h2 ) = 21 . However, the move r by play 2 in this Nash equilibrium is not
reasonable.
2.8
At any point during play of the game, each player should expect that play proceeds
according to some Nash equilibrium of the subgame that remains. Given a behavioral
15
Again looking at the example: consider the smallest subgame G. The unique Nash
equilibrium is
1
1
2 (H | h1 ) = ,
2 (h | h2 ) = .
()
2
2
Now consider the second smallest subgame G . The unique Nash equilibrium that is
consistent with () is
1
1
1 (H | h1 ) = ,
1 (h | h2 ) = ,
2 ( | h2 ) = 1.
2
2
2.9
Firm 1 chooses an outputQq1 . Firm 2 observes q1 and then chooses its own output
q2 . Payoffs are given by i (qi , qi ) = qi (1 q1 q2 ) cqi . A strategy for player 1
is a number qi [0, Q1 ]. A strategy for player 2 is a function r2 : [0, Q1 ] [0, Q2 ].
Consider subgame q1 : Firm 2 solves maxq2 q2 (1 q1 q2 ) cq2 , so
(
1q1 c
q1 1 c,
2
r2 (q1 ) =
0
q1 > 1 c.
Now we can consider the overall game. Given q1 1 c, firm 1s profit is
1 q1 c
max q1 1 q1
cq1 ,
q1
2
q1 =
1c
,
2
q2 =
1c
,
2
1c
.
4
Perfect information means you always know where you are in the game tree (no nonsingleton information sets). Complete information means you know your opponent,
payoffs, etc.
17
Chapter 3
Static Games of Incomplete
Information
Incomplete means some players do not know the payoffs (or types) of other players.
Example. Cournot competition under asymmetric information. Two firms simultaneously choose levels of output q1 and q2 . P (q) = a q,
( where q = q1 + q2 . The cost of
cH q2 with probability ,
firm 1 is c1 (q1 ) = cq1 . The cost of firm 2 is c2 (q2 ) =
cL q2 with probability 1 ,
c H > cL .
3.1
Firm 2 knows its true cost function. Firm 1 knows Firm 2s costs is cH with probability
, etc. Both firms know Firm 1s cost function. We first solve Firm 2s problem. If
Firm 2 is high cost, then q2 H solves
max[a q1 q2 cH ]q2 .
q2
aq1 cL
.
2
Hence
a [q2H + (1 )q2L ] c
.
2
So we have three equations and three variables.
(cH cL ) a cL
2
c ,
a+
q1 =
3
2
2
q1 =
3.2
Bayesian Games
Note. p is the same for all players, i.e. players have common priors.
1 ,...,I )
Then pi (i | i ) = p(p(
is the probability that player i assigns to the event
i)
that the profile of types for everyone else is i conditional on his own type
being i .
Let s() denote a strategy profile. Let si (i ) = (s1 (1 ), . . . , si1 (i1 ), si+1 (i+1 ), . . . , sI (I )).
Assume that for each player, each of her types occurs with strictly positive probability.
Definition. A strategy profile s() is a Bayesian equilibrium if for each player i and
all ,
X
p(i | i )ui (si , si (i ), ).
si (i ) arg max
si Si
i i
3.3
The buyer and seller simultaneously propose prices Pb and Ps , respectively. If Pb > Ps ,
then they trade at the price of 12 (Pb + Ps ). If Pb < Ps , then they dont trade.
Assume that the buyer and seller have values Ve1 and Ve2 that are drawn independently
from the uniform distribution on [0, 1]. Formally, the game is given by
1. Players: P = {b, s}.
2. Strategies: A strategy for the buyer is a mapping from his valuation to his bid,
i.e. Pb () : [0, 1] [0, 1]. Similarly for the seller.
3. Payoffs:
(
Vb
ub (Pb , Ps , Vs , Vb ) =
0
(
us (Pb , Ps , Vs , Vb ) =
Pb +Ps
2
Pb +Ps
2
Pb Ps ,
Pb < Ps ,
Vs
Pb Ps ,
Pb < Ps .
P
]
b
s s
s s
b
max Prob[Ps (Ves ) < Pb ] Vb
.
Pb
2
20
We also need
Pb s
Pb s
e
e
e
=
.
Prob[Ps Pb ] = Prob[s + cs Vs Pb ] = Prob Vs
cs
cs
Myerson and Sattherthwaite showed that there is no expost efficient Bayesian equilibrium for this game.
3.4
vi bi
vi bi
2
bi > b3i ,
bi = b3i ,
bi < b3i .
The strategies are functions bi () : [0, 1] [0, 1]. (b1 (), b2 ()) is a Bayesian equilibrium
if for each bidder i and Vi [0, 1], bi (Vi ) solves
1
max(Vi bi ) Prob[bi (Vei ) < bi ] + (Vi bi ) Prob[bi (Vei ) = bi ].
bi
2
Claim. There is a unique symmetric Bayesian equilibrium in strictly increasing and
differentiable strategies.
Proof. Suppose (b(), b()) is a Bayesian equilibrium and b() is strictly increasing
and differentiable. If b() is part of a Bayesian equilibrium, then for each player i and
each Vi [0, 1], b(Bi ) solves
max(Vi bi ) Prob[b(Vi ) < bi ].
bi
Now
22
1
b (b1 (bi ))
= 0.
()
This is an implicit solution for player is best response to b(). If (b(), b()) is a
Bayesian equilibrium, we require that bi = b(Vi ) satisfy (). Substituting b() into ()
gives
1
=0
b (b1 (b(Vi )))
1
= 0.
Vi + (Vi b(Vi ))
b (Vi )
()
d[b(Vi ) Vi ]
= b (Vi )Vi + b(Vi ),
dVi
so
d[b(Vi ) Vi ]
= Vi .
dVi
V2
Integrating both sides, we have b(Vi )Vi = 2i + K, or b(Vi ) = V2i + VKi . Suppose K > 0.
Then for Vi small enough, b(Vi ) > Vi , which is never optimal. Suppose K < 0. Then
for Vi small enough, b(Vi ) < 0, which is not allowed. Hence K = 0, and so b(Vi ) = V2i
is the solution, i.e. (b(), b()) = (1/2V1 , 1/2V2 ) is the unique symmetric Bayesian
equilibrium in increasing and differentiable strategies.
3.5
Second-Price Auction
(1)
High bidder wins but pays the second-highest bid. Let bi denote max{b1 , . . . , bi1 , bi+1 , . . . , bn }.
(1)
(1)
bi > Vi
(1)
bi > Vi
(1)
bi = Vi bi < Vi
win: Vi bi < 0
lose: 0
(1)
0Vi bi > 0
(1)
bi = Vi 0
0Vi bi > 0
(1)
win: Vi bi > 0
0
lose: 0
bi < Vi 0
We see that bi = Vi is a dominating strategy.
23
3.6
24
The economic environment and the mechanism yield a Bayesian game. For each player
i, denote a pure strategy by i : i Mi . Given a strategy profile = (1 , . . . , I ),
player is expected utility when her type profile is i is
X
(i | i )ui (y < (i , i ), (i , i )).
i i
3.7
i i
i i
i (i | i )ui (y(i (i ), mi ), (i , i ))
>
i i
which contradicts the fact that is a Bayesian equilibrium of the original mechanism.
We now return to the trading game. What can we say about the mechanisms we
could potentially design?
By the revelation principle, we can restrict attention to a direct mechanism in which
it is a Bayesian equilibrium for each player to reveal her type truthfully and every
player is willing to participate in the mechanism. Set Mi = i . Let ts denote the
25
transfer from buyer to seller when seller reports type s . Let xs be the probability
that the good is exchanged when the seller reports type s .
We have the following incentive compatibility constraints:
ICh: th 40xh t 40x ,
IC: t 20x th 20xh .
We also have the following individual rationality constraints:
IRh: th 40xh 0,
IR: t 20x 0,
IRB : .2[50xh th ] + .8[30x t ] 0.
Proposition. There is no incentive compatible and individually rational mechanism
that is expost efficient.
Proof. Expost efficiency requires xh = x = 1.
ICh th t
IC t th
th = t th = t 40.
IRh th 40
IR t 20
We also require
26
Chapter 4
Perfect Bayesian Equilibria
Consider the following game of complete but imperfect information:
1
R
M
1
3
2
L
2
1
R
0
0
0
2
0
1
There are two pure strategy Nash equilibria, namely (L, L ) and (R, R ). Both are
also subgame perfect equilibria, as there is only one subgame. However, (R, R ) is
unreasonable, in the sense that if player 2 gets to move she would always go L . We
rule out such unreasonable Nash equilibria or subgame perfect equilibria by adding
the following requirements:
1. At each information set, the player who moves must have a belief about which
node in the information set has been reached. (A belief at an information set
is a probability distribution over the nodes in the information set.)
2. Given their beliefs, the players strategies must be sequentially rational, which
means that for each information set controlled by a player, the action taken
by the player must be optimal given the players beliefs and the other players
strategies.
There are no beliefs for player 2 for which playing R would be optimal.
27
1
2
L
[p]
2
0
0
D
R
1
2
1
3
3
3
0
1
2
0
1
1
The Nash equilibrium of the red subgame is (L, R ). Thus (D, L, R ) is a subgame
perfect equilibrium. If we add belief p = 1, then (D, L, R ) and p = 1 is a perfect
Bayesian equilibrium.
Now consider (A, L, L ). This is a Nash equilibrium. If we add the belief p = 0, then
(A, L, L ) satisfies requirements 13 but violates 4.
Example
28
R
L
M
[p]
2 [1 p]
L
R L
R
1
4
1
0
0
3
0
2
2
0
1
2 2 (L ) 3
2
We assign belief p = 0. Then 2 (L ) = 0, 1 (R) = 1, p = 0 is a perfect Bayesian
equilibrium. 1 (R) = 1, 2 (L ) = 21 , p = 21 is also a perfect Bayesian equilibrium.
Now we look for a perfect Bayesian equilibrium in which 1 (R) < 1. In this case
2 (L ) > 0 since this is the only way player 1 can gain from choosing L or M . But if
2 (L ) > 0, then 1 (M ) = 0 since 42 (L ) > 32 (L ) (i.e. L is always better than M
for player 1). Since 1 (R) < 1 and 1 (M ) = 0, we have 1 (L) > 0. But if 1 (L) > 0
1 (L)
and 1 (M ) = 0, then by Bayess Rule, p = 1 (L)+
= 1, which means that 2 (L ) =
1 (M )
(sequential rationality). But then 1 (L) = 1 (also by sequential rationality). Hence
1 (L) = 1, 2 (L ) = 1, p = 1 is a perfect Bayesian equilibrium.
4.1
Signaling Games
Consider a dynamic game involving two players: the sender (S) and the receiver (R).
Nature draws a type ti for the sender from T = {t1 , . . . , tI } according to a probability
distribution p : T [0, 1]. The sender observes ti and chooses a message mj from
M = {m1 , . . . , mJ }. The receiver observes mj (but not ti ) and then chooses an action
ak from {a1 , . . . , aK }. Payoffs are given by uS (ti , mj , ak ) and uR (ti , mj , ak ).
Consider a 2 2 2 structure, with T = {t1 , t2 }, M = {m1 , m2 }, A = {a1 , a2 }, and
Prob[t1 ] = p.
29
a1
t1
m1
m2
a2
a1
a2
p
Nature
1p
a1
m1
t2
a2
m2
a1
a2
A strategy for the sender is a function m(ti ) = mj . A strategy for the receiver is a
function a(mj ) = ak .
We now restate the conditions for perfect Bayesian equilibria in such a way that they
apply to signaling games.
1. After observing any message mj M , the receiver must have a belief about
which types could have sent mj . (The belief isPa probability distribution ( |
mj ), where (ti | mj ) 0 for each ti T , and ti T (ti | mj ) = 1.)
2R. For each mj M , the receivers action a (mj ) must maximize her expected
utility given the belief ( | mj ), i.e. a (mj ) solves
X
(ti | mj )uR (ti , mj , ak ).
max
ak A
ti T
2S. For each ti T , the senders message m (ti ) must maximize his expected utility
given the receivers strategy a (mj ), i.e. m (ti ) solves
max uS (ti , mj , a (mj )).
mj M
Given the senders strategy m (), let Tj denote the set of types that send the message
mj , i.e. ti Tj iff m (ti ) = mj .
30
p(ti )
.
ti Ti p(ti )
Example.
1,3
4,0
[p] L
[1R q]
2,1
0,0
t1 [.5]
2
Nature
2
t2 [.5]
2,4
0,1
[1 Lp]
R [q]
1,0
1,2
4. If the sender plays (m(t1 ), m(t2 )) = (R, L), then the receiver must have beliefs
p = 0 and q = 1. Hence the receivers best response is (a(L), a(R)) = (u, u),
and in both cases the sender gets a payoff of 2. If a type 1 sender plays L, the
receiver plays u, so senders payoff is 1, so there is no incentive to deviate. If
a type 2 sender plays R, the receiver plays u, and the senders payoff would
be 1, so there is no incentive to deviate. Thus [(R, L), (u, u), p = 0, q = 1] is a
separating perfect Bayesian equilibrium.
2. a (R) = d and a (L) = u regardless of p. So either type of sender would deviate.
Thus there is no perfect Bayesian equilibrium with pooling on R.
3. (m (t1 ), m (t2 )) = (L, R), p = 1, q = 0. Then a (L) = u and a (R) = d. Thus
a t1 sender would not deviate, but a t2 sender would. Thus there is no perfect
Bayesian equilibrium with this type.
4.2
duel
wimpyBeerduel
Quiche
[p]
[q]
not
not
[.1]
Nature
receiver
0,1
2,0
receiver
[.9]
0,0
duel
Quichesurly Beerduel
1,0
2,1
not
3,1
not
32
We argue that the first equilibrium is not sensible. The intuitive criterion says that
a wimpy type can never gain by sending message Beer, but a surly type can. So if I
see Beer, I must put probability 1 on type surly.
Definition. A message mj 6= m (ti ) M is equilibrium-dominated for type
ti T if type ti s equilibrium payoff uS (ti , m (ti ), a (mj )) is greater than ti s highest
possible payoff from mj :
uS (ti , m (ti ), a (mj )) > max uS (ti , mj , ak ).
ak A
Definition. The (Weak) Intuitive Criterion. If the information set following mj is off
the equilibrium path and mj is equilibrium-dominated for type ti , then (if possible)
the receivers belief (ti | mj ) must be 0. This is possible so long as mj is not
equilibrium-dominated for all types ti T .
4.3
Moral Hazard
We will consider the case where signals are the principals realized gross profit, so
B(an ) =
M
X
mn sm
m=1
is the principals gross expected profit. Let xm = u(W (sm )). Let v = u1 , so
v(xm ) = W (sm ). The principals problem is to solve
max
an ,x1 ,...,xm
B(an ) =
M
X
mn v(xm )
m=1
such that
1.
M
X
m=1
mn xm d(an ) 0 ,
M
X
m=1
mn xm d(an )
M
X
m=1
mn xm d(an )
x1 ,...,xm
M
X
m=1
34
mn v(xm )
2. We solve for the action which maximizes the principals expected net profit:
max B(an ) C(an ).
an A
Example. (Varian 25.4) Two actions and two signals. A = {a1 , a2 }, where a1 and
a2 are effort levels, with a2 > a1 , and S = {success,failure}. Let sn = Prob[success|
a = an ], f n [failure| a = an ] = 1 sn , n = 1, 2. Assume that s2 > s1 , and hence
s2
> ff 21 .
f 1 > f 2 . The implication is that
s1
We solve
min sn v(xs ) + f n v(xf )
xs ,xf
u + a1
s1
xs .
f 1
f 1
(1)
s2
u + a2
xs .
f 2
f 2
s2 s1
a2 a1
xs
.
f 2 f 1
f 1 f 2
s2 s1 = (1 f 2 ) (1 f 1 ) = f 1 f 2 ,
so
xf = xs
a2 a1
.
f 1 f 2
35
(2)
We now work out the isocost line for the principal. If the agent takes action a1 , then
contracts that cost c are denoted by s1 v(xs ) + f 1 v(xf ) = c. Differentiate totally to
get s1 v (xs ) dxs + f 1 v (xf ) dxf = 0, or
dxf
s1 v (xs )
.
=
dxs
f 1 v (xf )
dx
Note that if xs = xf , then dxfs = fs11 . This means action a1 isocost curves are
tangent to action a1 isocost curves along the 45 line. The optimal contract for
implementing xf and action a1 has xs = xf .
Example. Sam is employed by Clampett Oil Company. He gives recommendations
on where to drill. There are two effort levels (where h denotes the number of hours
worked). If h = 400, then the company makes $250 000 with probability .7 and $ 50
000 with probability .3. If h = 0, then the company makes $250 000 with probability
.5 and $ 50
000 hwith probability .5. Sams utility over wages and ours worked is
u(w, h) = w 200
(in thousands). Sams alternative is to be a professional dart
thrower, which gives utility 8.
1. Suppose Clampett Oil cannot monitor Sams hours. Find the minimum cost of
inducing h = 400 or h = 0.
(a) Let us solve the problem of inducing h = 400. We must solve minxs ,xf .7x2s +
.3x2f such that
.7xs + .3xf 2 8
(1)
and
.7xs + .3xf 2 .5xs + .5xf .
(2)
From the graph, we know that the solution occurs where (1) and (2) intersect. We therefore have .7xs + .3xf = 10 and .2xs .2xf = 2, so x2 = 13
and xs = 13 and xf = 3, i.e. the minimum cost of implementing h = 400
is $121 000.
(b) Let us solve the problem of implementing h = 0. We have .5xs + .5xf = 8
and xs = xf , so xs = xf = 8. So the minimum cost is $64 000.
2. Which contract should Clampett Oil give Sam? The profit for inducing h = 400
is .7(250000) + .3(50000) 121000 = 39000, whereas the profit for inducing
h = 0 is .5(250000) + .5(50000) 64000 = 36000. So the contract would be
to offer Sam ws =169 000 in case of success and wf =9000 in case of failure.
3. What contract would Clampett offer if they could observe h?
36
(a) The first-best solution with h = 400 would be obtained by solving .7xs +
.3xf 2 = 8 and xs = xf . Hence we have xs = xf = 10. Thus the cost is
$100 000. The profit is .7(250000) + .3(50000) 100000 = 60000.
(b) We solved the other case before. The profit is only $36 000. Hence Clampett Oil should induce h = 400 by offering Sam $100 000 in any case.
37
Chapter 5
Problems
1. There are two baseball teams that are preparing for a three-game series. Each
team has three pitchers. Team 1 has an Ace, a mediocre pitcher, and a scrub.
Team 2 has two mediocre pitchers and a scrub. A pitcher can only be used once
in the series. The probabilities of winning the game depending on the pitcher
match-ups are as follows:
Match up
Ace versus mediocre player
Ace versus scrub
Mediocre pitcher versus scrub
Same versus same
Outcome
Ace wins with probability .8
Ace wins with probability .9
Mediocre pitcher wins with probability .6
Each wins with probability .5
Assume a win is worth 1 and a loss is worth 0. Suppose the three games
are played sequentially. The pitchers for each game are chosen simultaneously,
but after a game is played it is common knowledge what pitchers were used.
Remember each team can only use each pitcher once. Write out the extensive
form of the game. (The extensive form has 18 terminal nodes.)
2. One day, long ago, two women and a baby were brought before King Salmon.
Each claimed to be the true mother of the baby. King Salmon pondered dividing
the baby in half and giving an equal share to each mother, but decided that in
the interest of the baby he would use game theory to resolve the dispute instead.
Here is what the king knew. One of the women is the true mother but he does
not know which one. It is common knowledge among the two women who the
true mother is. The true mother places a value of 10 on having the baby and 0
on not having it. The woman who is not the true mother places a value of 5 on
having the baby and 0 on not having it. The king proposed the following game
38
to determine who would get to keep the baby. One woman is chosen to go first.
She is player 1, and the other woman is player 2. She must announce mine if
she wants to claim to be the real mother and hers if she wants to claim that
the other woman (player 2) is the true mother. If she says hers the baby goes
to player 2 and the game ends. If she says mine, player two gets to move.
Player 2 announces MINE or HERS. If player 2 says MINE, she gets the
baby and pays King Salmon an amount v. Meanwhile, player 1 also has to pay
King Salmon 1. If player 2 says HERS, Player 1 gets the baby at no cost and
player 2 makes no payments to King Salmon.
(a) Draw the extensive form for each game:
i. The true mother is player 1.
ii. The true mother is player 2.
(b) Assume that the mothers play the (unique) subgame perfect equilibrium
in each game. Find a value v such that the true mother receives the baby
regardless of which mother is player 1.
(c) Draw the normal form representation of each of the extensive form (using
your value of v from part (b).) Are there any Nash equilibria in which the
true mother does not receive the baby? If so, why does this happen under
Nash equilibria but not under subgame perfect equilibria?
3. Two firms, A and B, produce a homogeneous good that they sell in two markets,
1 and 2. Each firm has the cost function
C(q1j + q2j ) =
(q1j + q2j )2
,
2
where q1j and q2j are the quantities firm j = A, B sells in market 1 and market
2, respectively. The inverse demand function is the same in each market and is
given by P (Qi ) = 20 Qi , where Qi = qiA + qiB for i = 1, 2.
(a) Write firm As profit (from sales in both markets) as a function of q1A , q2A , q1B , q2B .
(b) Consider the case where the firms sell the good in both markets simultaneously. Write down the strategy set for firm j. Find the Nash equilibrium.
(c) Consider the case where the two markets are open sequentially. Both
firms sell output in market 1 and then, after seeing the sales in market
1, both firms sell output in market 2. Define a strategy for firm j. Find
the unique subgame perfect equilibrium. Provide intuition as to why in
subgame perfect equilibrium both firms do better in this case than in the
previous case.
39
Rock
0,0
1, 1
1, 1
Paper
1, 1
0,0
1, 1
Scissors
1, 1
1, 1
0,0
6. Players 1 and 2 are bargaining over how to split one dollar. Both players
simultaneously name shares they would like to have, s1 and s2 , where 0
s1 , s2 1. If s1 + s2 1, then the dollar is divided with each player receiving
the share named. If s1 + s2 > 1, then each player receives $0. What are the
pure-strategy Nash equilibria of this game? Is the dollar divided in every Nash
equilibrium?
7. For the extensive form game below (which has only one
player), consider the
1
1
1
4
mixed strategy 1 = 3 1 2 3 , 18 r1 2 3 , 6 r1 r2 3 , 9 r1 r2 r3 , where r1 2 3 , for example, denotes the pure strategy in which player 1 chooses r1 at his first decision
node, 2 at his second decision node, and 3 at his third decision node.
1
1
r1
1
2
r2
1
3
r3
1
L1
R1
2
L2
2
5
R2
r1
0
1
0
0
r1
0
2
5
2
9. Consider the following two-player game, . Each player can either work hard
(h) or shirk (s). If both work hard they each get a payoff of 2. If both shirk
they each get a payoff of 0. If one shirks and the other works hard, the person
who works hard gets a payoff of 1 and the person who shirks gets a payoff
of 0. The game is preceded by another simultaneous game in which the two
players announce whether they wish to Play or Not Play the game . If
they both announce Play then they play the game . Otherwise, each gets a
payoff of 1, i.e., strategy profiles (Play, Not Play), (Not Play, Play), and (Not
Play, Not Play) give each player a payoff of 1.
(a) Compute all the Nash equilibria of the subgame .
(b) Is there a subgame-perfect equilibrium in which both players choose shirk
with positive probability in the subgame ?
10. Consider the following static Bayesian Game:
(1) Nature determines whether the payoffs are as in Game 1 or as in Game 2,
each game being equally likely.
(2) Player 1 (the row player) learns whether nature has drawn Game 1 or
Game 2, but player 2 does not.
(3) Player 1 chooses either T or B and Player 2 simultaneously chooses either
L or R.
L
10,10
0,0
T
B
R
0,0
4,4
T
B
L
4,4
0,0
R
0,0
8,8
L
2,2
2,1
R
1,2
0,0
T
B
42
L
0,1
4,4
R
1,0
6,6
T
B
L
2,4
0,1
R
0,2
1,0
T
B
L
2,2
4,0
R
0,0
1,2
The type profile determines which of the four matrices maps actions to payoffs.
If, for example, t1 = 0 and t2 = 1, then the upper right matrix shows how
player payoffs depend on actions. In that case, if player 1 chooses T and player
2 chooses R then their payoffs are 1 and 0, respectively.
(a) Show that (s2 (0), s2 (1)) = (R, L) is NOT part of a Bayesian equilibrium.
(b) Find a pure strategy Bayesian equilibrium.
12. Two individuals each have $1 deposited in a bank and must simultaneously
decide whether to withdraw their $1 or leave it deposited in the bank. The
bank has invested all money in a risky, illiquid asset. If one or both players
attempt to withdraw their $1, the bank will have to liquidate the asset for onehalf the initial value and the proceeds will be divided among the claimants.
Since there are two players that means $2 would be liquidated for $1, which
would be divided 1 way (if one claimant) or 2 ways (if two claimants). If neither
player withdraws their $1, the money will remain invested and each player will
either gain 20 cents or lose 20 cents depending on the success or failure of the
banks risky investment. Success occurs with probability p and failure occurs
with probability 1 p. The game is described below. Player 1 chooses either
withdraw or not withdraw (w or n, respectively) and player 2 simultaneously
chooses either withdraw or not withdraw (W or N , respectively.)
w
n
W
.5,.5
0,1
N
1,0
1.2,1.2
w
n
W
.5,.5
0,1
N
1,0
.8,.8
(a) Describe the set of pure-strategy Nash equilibria for each value of p in the
interval [0, 1]. Identify which equilibria are efficient.
(b) Suppose now that player 1 is informed about the outcome of the banks investment before she decides whether to withdraw or not withdraw. Player
2 still only knows the probability of success or failure. These facts are
common knowledge. Suppose the game is still played simultaneously. Describe the set of pure-strategy Bayesian equilibria for each value of p in
ther interval [0, 1]. Briefly comment on how private information affects the
outcome of this game.
43
13. Consider a first-price, sealed-bid auction in which the bidders valuations are
independently and uniformly distributed on [0, 1]. Show that if there are n bidders, then the strategy of bidding (n1)/n times ones valuation is a symmetric
Bayesian equilibrium of the auction.
14. Consider the following static Bayesian game:
(i) Nature determines whether the payoffs are as in game 1 or as in game 2,
each game being equally likely.
(ii) Player 1 (the row player) learns whether nature has drawn game 1 or game
2, but player 2 does not.
(iii) Player 1 chooses either T or B, player 2 observes player 1s choice, and
then chooses either L or R.
T
B
L
0,0
6,6
R
10,4
0,2
T
B
L
2,4
6,0
R
0,4
2,6
Find all the pure-strategy pooling and separating perfect Bayesian equilibria in
the resulting signaling game. Be sure to specify the equilibrium strategies and
beliefs in each case.
15. An evil dictator might have weapons of mass destruction (WMD). A neighboring
country regards each possibility as being equally likely and must decide whether
or not to invade. You must analyze the signaling game in which the evil dictator
makes an announcement about his weapons status and the neighboring country
decides whether or not to invade. Find all the pure-strategy Bayesian equilibria.
Are there any separating equilibria? Are there any pooling equilibria?
0,50
100,0
Invade
No
Evil Dictator
Yes
Not
Invade
Not
0,50
100,0
[.5] WMD
Neighbor
Nature
Neighbor
[.5] No WMD
0,0
100,100
Invade
No
Evil Dictator
Not
Yes
Invade
Not
44
0,50
120,80
16. Once there was an evil lord that visited a poor serf each fall and demanded half
of his harvest. Both the lord and the serf knew that a good harvest produced
10 bushels of potatoes and that a bad harvest produced only 6. In addition, it
was common knowledge that the probability of a good harvest was 0.6. The serf
knew which type of harvest had occurred. The lord was told the type of harvest
by the serf and had to decide whether or not to believe him. If he believed the
serf he took the potatoes he was offered and left. If he didnt believe the serf,
he took what the serf offered and punished him for lying by burning down his
house.
The serf kept potatoes in the cellar of his house. Hence, the truth would be
revealed once the house had burnt down and his potato holdings were exposed.
Punishing a dishonest serf gave the lord an added boost of pleasure equal to
c > 0. On the other hand, if the fire revealed that the serf had been telling
the truth, the lords happiness was diminished by c. While the lord felt bad
when he was wrong, he would not compensate the serf by giving the serf some
of his crop back. Also, neither the serf nor the lord liked baked potatoes, so any
potatoes that were hidden in the cellar were worthless if the house was burned.
The payoffs of this signaling game are shown below.
(a) Find all the pure-strategy perfect Bayesian equilibria. Are there any separating equilibria? Are there any pooling equilibria?
(b) Do any of the pooling equilibria fail the (weak) intuitive criterion?
(c) Suppose the lord can commit to a strategy in advance. Namely, the lord
can announce in advance what he will do in response to each message by
the serf, and the announcement is credible. What strategy for the lord
maximizes his expected payoff? Set c = 1.
17. Two identical firms each simultaneously choose a nonnegative quantity of output. There are no costs. The payoff to firm i as a function of the outputs is
i (qi , qi ) = (100 qi qi )qi . Show that for each firm, any level of output
greater than 50 is strictly dominated.
18. On the hit TV show Survivor Thailand two tribes competed in a game they
called Thai 21. The game starts with 21 flags in a circle. The teams take turns
removing flags from the circle until they are all gone. The team that removes
the last flag wins immunity. At each turn teams must remove 1, 2, or 3 flags.
Team 1 moves first. How many flags should Team 1 take on its first move?
45
19. Consider the two-player Bayesian game in which S1 = {T, B} and S2 = {L, R},
each player has two types 1 = 2 = {0, 1}, and each type profile is equally
likely, i.e. p(0, 0) = p(0, 1) = p(1, 0) = p(1, 1) = 14 . Payoffs are given below:
t2 = 0
t1 = 0
t1 = 1
T
B
L
0,0
2,1
R
1,2
0,0
T
B
L
2,0
0,1
R
0,2
1,0
t2 = 1
T
B
L
0,1
2,0
R
1,0
0,2
T
B
L
2,1
0,0
R
0,0
1,2
The type profile determines which of the four matrices maps actions to payoffs.
If, for example, t1 = 0 and t2 = 1, then the upper right matrix shows how
player payoffs depend on actions. In that case, if player 1 chooses T and player
2 chooses R then their payoffs are 1 and 0, respectively.
(a) Is (s1 (0), s1 (1)) = (T, B) part of a Bayesian equilibrium?
(b) Is (s2 (0), s2 (1)) = (L, L) part of a Bayesian equilibrium?
20. Find all the pure-strategy perfect Bayesian equilibria of the following game in
extensive form.
Player 1 R
L
10
30
M
Player 2
20
10
0
0
0
20
0
10
21. Once upon a time a signaling game was played between a princess and a frog.
The frog was the Sender. He could either say he was a prince or a frog.
The princess was the receiver. She could either kiss the frog, in which case he
might turn into a prince. Or, she could eat him. It was well known that 10%
of the frogs in the kingdom would turn into princes when kissed by a princess.
For reasons that arise from commonly known facts about frogs and princesses,
the payoffs of the signaling game are as presented below.
46
10,100
Kiss
0,5
Eat
Frog
Sender
Prince
Kiss
10,100
Eat
10,5
[.1] Prince
Princess
Nature
Princess
[.9] Frog
10,10
Kiss
0,5
Eat
Frog
Sender
Prince
Kiss
5,5
Eat
10,5
(a) Find all the pure-strategy perfect Bayesian equilibria. Are there any separating equilibria? Are there any pooling equilibria?
(b) Do any of the pooling equilibria fail the (weak) intuitive criterion?
22. An agent has effort levels a1 = and a2 = 3. The probability distribution over
sales for each level of effort is given below.
Sales
Sales
Effort s1 =$100 s2 =$150
a1 = 1
.8
.2
a2 = 3
.4
.6
The expected profit for the principal when the agent takes action ai is expected
sales conditional on ai minus expected costs conditional on ai . The agent has
utility over wages and actions given by U (w, ai ) = w1/2 ai and a reservation
utility of 1. Contracts can be written in terms of sales, but not actions.
(a) Find the minimum cost of implementing each action.
(b) Which action does the principal implement to maximize profit?
47