Advanced Topics in Microeconomics

Game Theory

Duozhe Li
Department of Economics
Chinese University of Hong Kong

Spring 2021
1 Introduction

• Game theory originated as a branch of applied math-

ematics. It has become a standard analytical tool of
modern economics.

• Game theory studies strategic interactions, which

may occur within or between families, local commu-
nities, firms, organizations, or nations.

• A crucial feature of strategic interaction is that the

outcome for one party depends not only on what that
party does, but also very largely on how other parties
act or react. For example, a firm that decreases its
price to attract more customers will not succeed in
this strategy if the other major firms in the market
use the same strategy.

• A simple example: rock-paper-scissors.

2 Strategic Games

• Definition A strategic game consists of:

1. a set of players ;

2. for each player , a set of pure strategies (actions)

;

3. for each player , a payoff function  :  →

R, where  ≡ ×∈  is the set of strategy
profiles.

• Definition A mixed strategy  of player  is a

probability distribution over pure strategies. The set
of mixed strategies is denoted by Σ (Σ ≡ ×∈ Σ).

— Independent randomization

— Expected payoff:
X ³ ³ ´´
 () = 
Π=1    () 
∈
• Assumptions

— Each rational player  has a vNM preference º

over the set of lotteries over , which is repre-
sented by the expected payoff function  (·).

— The game structure is common knowledge.

• Definition A strategy ³ is strictly

´ dominated by
0 if  ( −)   0 − for every −.

— A pure strategy may be dominated by a mixed

strategy, and vice versa.

— A mixed strategy that assigns positive probability

to a dominated pure strategy must be dominated.

— A rational player will never play a strictly domi-

nated strategy.

• Some (very few) games can be “solved” by iterated

elimination of strictly dominated strategies.
Example Original Game
Left Center Right
Up 4 3 5 1 6 2
Middle 2 1 8 4 3 6
Down 3 0 9 6 2 8
Center is strictly dominated for player 2.
Left Right
Up 4 3 6 2
Middle 2 1 3 6
Down 3 0 2 8
Middle & Down are strictly dominated for player 1.
Left Right
Up 4 3 6 2
Right is strictly dominated for player 2.
Left
.
Up 4 3
• The procedure of iterated elimination of strictly dom-
inated strategies is based on the assumption of com-
mon knowledge of rationality among players. An
important feature of the procedure is order inde-
pendence, i.e., the order of the elimination has no
effect on the set of strategies that remain in the end.

• Definition A strategy ³ is weakly

´ dominated by
0 if  ( −) ≤  0 − for every −, and
the inequality is strict for some −.

— A strategy that is weakly dominated cannot be

ruled out based only on principles of rationality.

— Iterated elimination of weakly dominated strate-

gies is order dependent, for example,
L C R
T 1 1 1 1 0 0 
B 0 0 1 2 1 2
• A game is dominance solvable if all weakly domi-
nated strategies of each player are eliminated at each
stage, and all the players are indifferent among all re-
maining strategy profiles.

• Definition A mixed-strategy profile  ∗ ∈ Σ is a

Nash Equilibrium (NE) if for all ,
³ ´ ³ ´
∗ ∗ ∗
    − ≥    − for all  ∈ 

— The definition does not require equilibrium choices

to be strictly better than other available choices.

— In a mixed-strategy NE, each player is indifferent

among all those pure strategies that he chooses
with positive probability. (Why?)

— A pure-strategy NE is a degenerate mixed-strategy

NE. There may exist no pure-strategy NE.
Examples

1. Prisoners’ Dilemma
Not Confess Confess
Not Confess −1 −1 −150 
Confess 0 −15 −10−10
The unique NE is (Confess, Confess)

Many real life situations can be modelled as a Pris-

oners’ Dilemma game, for example, the arms race
in the cold war era, price competition between two
firms producing the same product, etc.

2. Battle of Sexes
Opera Football
Opera 31 0 0 
Football 0 0 13
There are two pure-strategy NEs: (Opera, Opera)
and (Football, Football).
 The mixed-strategy NE in BOS: player 1 (2 resp.)
chooses Opera with probability 34 (14 resp.).

Computing the mixed NE:

Let player 1 (2 resp.) choose Opera with probability

 ( resp.). If   ∈ (0 1), both players should
be indifferent between his two pure strategies. That
is, given player 2 chooses Opera with probability ,
player 1 should obtain the same expected payoff from
choosing Opera or Football, and vice versa. Thus,

Player 1: 3 ×  = 1 × (1 − )
Player 2: 1 ×  = 3 × (1 − ) 
by which we have  = 34 and  = 14.

3. Matching Pennies
Head Tail
Head 1 −1 −11
Tail −11 1 −1
No pure NE. A unique mixed NE: both choose Head
with probability 12.
4. Hawk-Dove
Hawk Dove
Hawk −2 −2 2 0 
Dove 0 2 1 1
Two pure NEs: (Hawk, Dove) and (Dove, Hawk). A
mixed NE: both choose Hawk with probability 13.

5. Stag-Hunt or Coordination Game

Hunt Stag Chase Hare
Hunt Stag 33 0 2 
¯¯
Chase Hare 2 0 11
¯¯
Canteen A Canteen B
Canteen A 33 0 0 
¯¯
Canteen B 0 0 11
¯¯
Both games have two pure NEs, in which two players
choose the same action.

There is also a mixed NE: (1) In the Stag-Hunt game,

both choose “Hunt Stag” with probability 12. (2)
In the pure coordination game, both choose Canteen
A with probability 14.
6. Games with Non-strict NE

Accept Fight
Out 0 4 0 4
In 1 2 −2 1
Two pure NEs: (In, Accept) and (Out, Fight).

When player 1 chooses Out, 2 is indifferent between

his two strategies. Hence, there is a continuum of
mixed NE: player 1 chooses “Out”, while 2 chooses
“Fight” with a probability at least 13. (The pure
NE (Out, Fight) belongs to this family of mixed NE.)
L R
T 1 1 0 0
B 0 0 0 0
(T, L) is a NE, (B, R) is also a NE. No mixed NE.

• Theorem (Nash 1951) Every finite strategic game

has a NE in mixed strategies.
7. Cournot Model (Strategy is a continuous variable.)

Two firms produce an identical product, and they set

output levels (1 2) simultaneously.

Inverse demand function:  =  − (1 + 2).

Constant unit production cost  ≤ .

As a profit maximizer, firm  ∈ {1 2} solves

h³ ³ ´´ i
max   −  +  −
 ≥0

Best responses from the FOCs:

 −  − 2  −  − 1
1 = and 2 = .
2 2
Cournot-Nash equilibrium:

1 = 2 = ( − ) 3

Extension: NE of the Cournot game with  firms:

 = ( − )  ( + 1) 
3 Bayesian Games

• Example BoS with incomplete information

Prob. 12 Prob. 12
F M F M
F 2 1 0 0 F 2 0 0 2
M 0 0 1 2 M 0 1 1 0
2 wishes to meet 1 2 wishes to avoid 1

— Player 2 has two types. One type wishes to meet

1, and another type wishes to avoid 1.

— Player 1 does not know player 2’s type, but holds

a belief (equal probabilities) about it.

— Player 1’s expected payoffs:

F, F F, M M, F M, M
F 2 1 1 0 
M 0 12 12 1

— Pure strategy NE: (F, (F, M)).

• A Bayesian game is a list

 = {; ; ; }

=1
where

—  is the action space of  ( ∈ )

—  is the type space of  ( ∈ )

—  (−|) is ’s belief about other players’ types

—  (1     ; 1     ) is ’s payoff.

• The type of a player embodies any private informa-

tion that is relevant to the player’s decision making.
This may include, in addition to the player’s payoff
function, his beliefs about other players’ payoff func-
tions, his beliefs about what other players believe his
beliefs are, and so on.
• A Bayesian Nash equilibrium (BNE) is a Nash
equilibrium of a Bayesian game. Given any Bayesian
game
 = {; ; ; }
=1 
a strategy of a player  is any function  :  → .
A strategy profile

∗ = (∗1     ∗)
is a BNE iff ∗ () is a best response to ∗− for each
. More precisely, ∗ () solves
X ³ ´
max ∗
 (−|) ×   − (−) ; t 
∈
−∈−
• Public good provision: Two players decide simul-
taneously whether or not to contribute to a public
good. Each player  derives a commonly known value
 if at least one of them contributes and 0 if none
of them does. Player ’s cost of contribution is ,
only known to himself. It is common knowledge that
1 and 2 are i.i.d. on [  ], and the continuous
distribution function is  (·).
The Bayesian game:

— Two players,  = 1 2

— Player ’s action space: {0 1}, where 1 stands

for “contribute”, and 0 stands for “don’t”.

— Player ’s type space: [  ] 

³ ´
— Player ’s belief: Pr  ≤  p  =  () 

— Player ’s payoff function:

⎧
⎪
⎨ 0 if 1 = 2 = 0
 (1 2; 1 2) =  −  if  = 1 
⎪
⎩  if 
 = 0 and  = 1
Bayesian Nash equilibrium

1. BNE exhibits a monotonicity property. That is,

if player  of type ∗ contributes, then player 
of type   ∗ must also contribute; if player 
of type ∗ does not contribute, then player  of
type   ∗ must not contribute.

2. Assume that  = 2, 1 and 2 are uniformly dis-

tributed on [1 3]  There is a symmetric BNE,
in which ∗ () = 1 iff  ≤ ∗

3. To find ∗ note that player  of type ∗ must

be indifferent between the two actions, provided
that player  uses the same equilibrium strategy.
If  () = 1, i.e., he chooses to contribute, then
his payoff is 2 − ∗.
If  () = 0, i.e., he chooses not to contribute,
then his expected payoff is
³ ´ ∗ − 1
∗
Pr  ≤  ·  = · 2 = ∗ − 1
3−1
Let 2− ∗ = ∗ − 1, we obtain ∗ = 32
4. In this case, there are two asymmetric equilib-
ria, in which one player never contributes and the
other player contributes for all  ≤  = 2.
The existence of such asymmetric equilibria de-
pends on the common value  and the distribu-
tion of  More specifically, asymmetric equilibria
exist when

 () ≥  − 
that is, it is optimal for a player with the low-
est cost not to contribute if he believes that the
other player contributes whenever that player’s
cost does not exceed .
For example, if  = 1 1 and 2 are uniformly
distributed on [0 2]  then there does not exist
such asymmetric equilibria.
4 Extensive Games

4.1 Games with Perfect Information

• Definition An extensive game with perfect infor-

mation is a game tree containing a starting node,
other decision nodes, terminal nodes, and branches
linking each decision node to successor nodes.

— A set of  players,  ≡ {1 2  } 

— For each decision node, the identity of the player

entitled to move.

— For each player  a specification of his action set

at each node that he is entitled to move.

— A specification of the payoff to each player at

each terminal node.
• Examples

1. Entry game
Entrant

Stay Out Enter

Incumbent
0
Fight Accommodate
4

-2 1
1 2

2. Modified BoS
Wife

Opera Football

Husband Husband

Opera Football Opera Football

3 0 0 1
1 0 0 3
• Definition A strategy of player  in an extensive
form game with perfect information is a complete
list of actions, one action for each decision node that
player  is entitled to move.

— How many strategies does Ann (Bob) have in the

following game?
Ann
out in
Bob
1 out in
0
Bob
-1
8 Left Right

Ann Ann

Left Right Left Right

12 13 3 0
13 1 5 -10

• Definition A strategy profile  (one strategy 

for each player ) determines a sequence of actions
leading to a terminal node, namely, a path of play.
We refer to this path of play as the outcome of .
• Strategic form representation

— Entry game
accommodate fight
out 0 4 0 4
in 1 2 −2 1

— Modified BoS
OO FF OF FO
O 3 1 0 0 3 1 0 0
F 0 0 1 3 1 3 0 0

• Nash equilibrium

— A finite game of perfect information has a pure

strategy Nash equilibrium.

— In general extensive form games, some of the NEs

are “more reasonable” than others.
• Sequential rationality: A player is sequentially ra-
tional iff, at each node he is to move, he maximizes
his expected utility conditional on that he is at the
node — even if this node is precluded by his own
strategy.

• Backward induction
Ann

Bob
Ann
1
0
Bob
Bob
-1
1
8
0
Bob
Ann
Ann -1
8

12 13 3 0 13 3
13 1 5 -10 1 5

Ann

Ann
B ob
1
0
1 -1
-1 3
8 5 0 8
• Remarks

— In a finite game of perfect information, com-

mon knowledge of sequential rationality yields
the backward induction outcome.

— There are Nash equilibria that are different from

the backward induction outcome.

— Backward induction always yields a Nash Equi-

librium.

— Sequential rationality is stronger than rationality

— Backward induction sometimes leads to a coun-

terintuitive outcome, for example, in the cen-
tipede game:
A nn B ob A nn B ob A nn B ob A nn B ob 103
101

1 0 3 2 5 98 101 100
0 2 1 4 3 100 99 102

4.2 Games with Imperfect Information

• Imperfect information There are situations where

players do not observe actions taken by other player(s)
in earlier stages of the game, for example, a game
involving simultaneous move.

• An information set for a player is a collection of

decision nodes in which he has to move. In a perfect
information game, each information set contains a
single decision node.

• A sequence of decision nodes starting from the initial

decision node and connected by actions taken by the
players is often referred to as a history. A terminal
history is simply a complete path of play (an out-
come), a nonterminal history ends with a decision
node where one player is to move.
• Definition A pure strategy of player  in an exten-
sive form game is a complete list of actions, one
action for each information set that player  is en-
titled to move.

• If player  has  information sets, and at the 

information set, the number of actions is  then
the number of pure strategies is

# = 1 × 2 ×  ×  

• Randomization: mixed and behavioral strategies

— A mixed strategy is a probability distribution on

the set of pure strategies of the strategic form
representation.

— A behavioral strategy of player  specifies a prob-

ability distribution on the set of actions at each
information set of player  (i.e., it is a profile of
probability distributions).
• Definition A (proper) subgame is a part of the orig-
inal game tree with following properties:

— it begins with an information set containing a

single decision node;

— it contains all the successor nodes, their informa-

tion sets, and connecting branches, up to all the
relevant terminal nodes;

— if a subgame contains one node in an informa-

tion set, it must contain all the nodes in that
information set.

• Examples
Ann
out in
Bob
1 out in
0
Bob
-1
8 L e ft R ig h t

Ann Ann

L eft R ig h t L eft R ig h t

12 13 3 0
13 1 5 -1 0
 Wife

Shopping Go Home

Wife
2
0 Opera Football

Husband

Opera Football Opera Football

3 0 0 1
1 0 0 3

A B

2 2
L R L R
3
C D C D C D C D
• Subgame Perfect Equilibrium (SPE) A strategy
profile is a SPE if it induces a NE in every subgame.

• Every finite game of perfect information has a pure

strategy SPE. Moreover, if no player is indifferent
(has the same payoff) at any two terminal nodes,
then there is a unique SPE, which can be derived by
backward induction.

• Generalized backward induction procedure

— Identify all NEs of the final subgames (i.e., those

that have no other subgames nested within them).

— Select one NE in each final subgame, and replace

the subgame with a terminal node with the pay-
offs of the selected NE.

— Go backwards until a strategy profile of the orig-

inal game is determined.
• Example Stackelberg model

Two firms produce the same product. Production is

costless.

Firm 1 chooses 1 first, and then firm 2, after ob-

serving 1 chooses 2

Demand Function:  = 100 − 

Finding SPE by backward induction

Second period: firm 2

max  2 (1 2) = (100 − 1 − 2) 2

2
FOC:
2
= 100 − 1 − 22 = 0
2
100 − 1
2 = 2(1) =
2
This should be player 2’s strategy in the SPE.
First period: firm 1 foresees firm 2’s choice,

max 1 = (100 − 1 − 2) 1

1
µ ¶
100 − 1
= 100 − 1 − 1
2
FOC,
1
= 100 − 21 = 0
1
⇒ 1 = 50
SPE strategy profile:
½ ¾
100 − 1
1 = 50 2 = 
2
SPE outcome: 1 = 50 and 2 = 25  = 25,
 1 = 1250 and 2 = 625

Remark Compared with the Cournot model, firm 1

(the leader) makes higher profit, firm 2 (the follower)
makes lower profit, and the total profit is lower.
4.3 Sequential Equilibrium

• Example SPE needs further refinement.

E n tra n t

O ut In
E n tra n t
0
In - 1 In - 2
2

In c u m b e n t

F ig h t A c c o m m o d a te F ig h t A c c o m m o d a te

-1 3 -1 2
-1 0 -1 1

The unique SPE is {In→In-1, Accommodate}.

Entrant

Out In-1 In-2

Incumbent
0
2 Fight Accommodate Fight Accommodate

-1 3 -1 2
-1 0 -1 1

After we merge player 1’s two moves, {Out, Fight}

also becomes a SPE, which seems unreasonable.
• Beliefs () of an agent at a given information set is
a probability distribution on the information set.

• An assessment ( ) in an extensive game consists

of a behavioral strategy profile and a belief system.

• Sequential Rationality A player is sequentially ra-

tional iff, at each of his information set, he maximizes
his expected payoff given his beliefs.

• Weak Consistency Given any strategy profile  and

any information set  on the path of play of , a
player’s beliefs at  is weakly consistent with  iff
the beliefs are derived using the Bayes’ rule and .
There is no restriction on the beliefs at information
sets that are off the path of play.
• Example

L (p) M (1-p) R (0)

2 2 2
A (q) B (1-q) A (r) B (1-r) C (s) D (1-s)

[x] [y] [1-x] [1-y] [z] 3-c [1-z]

3-a 3-b
E F E F E F E F G H G H

Information sets 3-a and 3-b are on the path of play,

and 3-c is off the path of play. Thus, weak consis-
tency requires that

 = and
 + (1 − )
(1 − )
 =
(1 − ) + (1 − )(1 − )
and it does not put any restriction on .
• Definition An assessment ( ) in an extensive game
is a weak perfect Bayesian equilibrium (WPBE) if
it satisfies both sequential rationality and weak con-
sistency. (It is also referred to as weak sequential
equilibrium.)

• Note the link between strategies and beliefs: The

beliefs are consistent with the strategies, which are
optimal given the beliefs. Due to this circularity,
WPBE cannot be determined by backward induction.

• A WPBE is a NE, but not every NE is a WPBE.

• To find all WPBEs of an extensive game, we can

first find all NEs from its strategic form, and then
for each NE strategy profile , check whether there
is a system of belief  such that ( ) satisfies both
sequential rationality and weak consistency.
• Example A SPE may not be a WPBE.
Entrant

Out In-1 In-2

Incumbent
0
2 Fight Accommodate Fight Accommodate

-1 3 -1 2
-1 0 -1 1

(Out, Fight) is a NE and SPE, but not a WPBE.

• Example A WPBE may not be a SPE.

E n tr a n t

O ut In
E n tr a n t
0
In - 1 In -2
2

[1 ] In c u m b e n t [0 ]

F ig h t A c c o m m o d a te F ig h t A c c o m m o d a te

-3 1 -2 3
-1 -2 -1 1
• We need reasonable restrictions about the beliefs at
the information sets which are off the path of play.

• Consistency Given  and an information set , even

if  is off the path of play, the beliefs must be derived
using the Bayes’ rule and  “whenever possible,” e.g.,
if players tremble with very small probability so that
 is on the path, the beliefs must be very close to
the ones derived using the Bayes’ rule and .

• Formally, an assessment
³ ´ ( ) is consistent if there
is a sequence     of assessments s.t.
³ ´
—    → ( ),

— each  is completely mixed, and

—  is derived from   using Bayes’ rule.

• Definition An assessment ( ) is a sequential
equilibrium (SE) if it satisfies both sequential ra-
tionality and consistency.

• A SE is both a SPE and a WPBE.

• Example The assessment specified below is not a

SE because consistency is not satisfied.
E n tra n t

O ut In
E n tra n t
0
In -1 In -2
2

[1 ] In c u m b e n t [0 ]

F ig h t A cco m m o d a te F ig h t A cco m m o d a te

-3 1 -2 3
-1 -2 -1 1
The strategy profile specified in this example cannot
be a SE, that is, we cannot find a belief system for
the strategy profile such that the assessment satisfies
both sequential rationality and consistency.

To see this, suppose that, on a sequence of com-

pletely mixed strategy profile, the Entrant chooses
In with probability  and then chooses In-1
with probability   .

On the Incumbent’s information set, using Bayes’

rule, his belief should be that the history In, In-1
has occurred with a probability of
 ×  
=   → 0
 ×   +  × (1 −   )

However, if Fight is a rational choice of the in-

cumbent on his off-the-equilibrium-path information
set, then he must be assigning a probability at least
23 on the occurrence of the history In, In-1.
• Example
Player 1

a b

1 2
2 c Player 2 c 1
0 3
d d
Player 3

e f e f

-1 0 0 1
1 3 3 -1
0 1 2 0

NE in pure strategies: (  ) and (   ).

WPBE in pure strategies:

1. (   |  =  = 0)
2. (   |  = 0  ≥ 23)

SE: The first WPBE is also a SE. The second WPBE

is not a SE because consistency is not satisfied. (why?)
5 Signaling Game

• We consider interactions between two players with

asymmetric information.

— Player 1 (sender) has private information about

his type  ∈ Θ and chooses an action 1 ∈ 1
to signal his type.

— Player 2 (receiver), whose type is common knowl-

edge for simplicity, observes 1 and chooses 2 ∈
2.

— Both players’ payoffs depend on  1 and 2

— Before the game begins, it is common knowledge

that player 2 knows the distribution of .

— If different types of player 1 choose different ac-

tion 1, then player 2, who observes 1, should
update his belief about player 1’s type, and base
his choice of 2 on the updated belief.
• We focus on sequential equilibria in pure strategies.

— Separating equilibrium: Each type of sender

sends a different signal, so that upon observing
the signal, the receiver knows the sender’s type.

— Pooling equilibrium: All types of the sender

send the same signal, so that the signal gives the
receiver no information about the sender’s type.

• The beer-quiche game (Cho and Kreps 1987)

(0, 1 ) (1, 1 )
D ue l D ue l
b e er 1 -W q u ich e

D o n’t 0 .1 D on ’t (3, 0 )
(2, 0 )

2 N
2

(1, 0 ) 0 .9 (0, 0 )
D u el D ue l

b eer q u ich e
1 -T
D on ’t D o n’t (2, 1 )
(3, 1 )

— No separating equilibrium.

— Two pooling equilibria, which are also sequential.

• Intuitive Criterion: Refinement based on strategic
stability.
(0, 1 ) (1, 1 )
D ue l D ue l
b e er 1 -W q u ich e

D o n’t 0 .1 D on ’t (3, 0 )
(2, 0 )

2 N
2

(1, 0 ) 0 .9 (0, 0 )
D u el D ue l

b eer q u ich e
1 -T
D on ’t D o n’t (2, 1 )
(3, 1 )

The pooling equilibrium in which both types choose

quiche is not reasonable. It can be eliminated by
the so-called intuitive criterion. In this equilibrium,
the weak type is getting his highest possible payoff,
thus has no incentive to deviate from the equilibrium
path; in other words, the deviation is dominated by
the equilibrium payoff. If player 2 observes such a
deviation, that is, a player 1 chooses beer, he should
believe that it cannot be a weak type. Hence player
2 should not duel with such a deviator, which gives
the tough type the incentive to deviate.