ECON0027 Lecture notes

Section 2: Games in extensive form

with perfect information
Reading: Osborne: Ch 5, 6, 7

2.1 Example: tic-tac-toe

Consider a well-know game called “tic-tac-toe”. It does not fit well into our model of Normal
form games (or does it?) because the players do not make their decisions simultaneously—they
take turns. Here is a representation of the first two rounds of the game showing all possible
scenarios:

P layer 1

x x
P layer 2 x

x o x x o x x o o x o x x x x
P layer 1 o o x x o o
o o o

In this section we develop tools to study games in which players take turns. The tic-tac-toe
example suggests that a convenient tool to represent the structure of such games is a graph (or,
rather a special case of a graph—a tree).

2.2 Definition
A game in extensive form is =< N, A, X, E, ◆, (ui )i2N >, where
• set of players N = {1, 2, ..., N }
• set of actions A
• set of nodes (or histories) X, including x0 – the initial node or empty history
• any node x = (a1 , a2 , .., ak ) results from this sequence of actions
• A(x) is the set of actions available at x i.e. the player playing after x needs to choose from A(x)

• set of terminal nodes E, where A(z) = ; for z 2 E

• ◆ : X\E ! N tells us which player chooses an action at node x i.e. the player active at history x

• ui : E ! R is i’s payo↵ function

A game is of perfect recall if each player remembers what he did in prior moves, and each
player remembers everything that he knew before. A game is finite if A and X are finite.

1
Assumptions of a game in extensive form:
• The agents posses common knowledge of rationality (i.e. each agent is rational, each agent knows that his opponents are rational, each
agent knows that his opponents know that he is rational, and so on).
• The agents posses common knowledge of the game (i.e. each agent understand the rules of the game, knows that his opponents
knows the rules, knows that his opponents know that he knows the game, and so on).
• The agents posses perfect recall (if the agent knew some information in the past, he still knows these information in the future).
2.3 Example of a game tree
Consider =< N, A, X, E, ◆, (ui )i2N >, where
• N = {1, 2};

• A = {L, R, a, b, C, D};

• X = {;, L, R, (R, a), (R, b), (R, b, C), (R, b, D)};

8
>{L, R} if x = x0
>
>
<{a, b} if x = R
• A(x) =
>
> {C, D} if x = (R, b)
>
:
; if x 2 {L, (R, a), (R, b, C), (R, b, D)}
8
>
<1 if x 2 {x0 , (R, b)}
• ◆(x) = 2 if x = R
>
:
; if x 2 {L, (R, a), (R, b, C), (R, b, D)}
8 8
>
> 3 if x=L >
> 1 if x = L
>
<6 >
<4 if x = (R, a)
if x = (R, a)
• u1 (x) = and u2 (x) =
>
> 5 if x = (R, b, C) >
> 3 if x = (R, b, C)
>
: >
:
0 if x = (R, b, D) 14 if x = (R, b, D)
As you can see from the description above, even a very simple game can have a long and
tedious description. Luckily, we can draw the game tree, which is a more user-friendly represen-
tation of the same game. The game tree is without loss—it contains all the information from
the formal description of the game.

1
L R

3,1 2
a b

6,4 1
C D

5,-3 0,14

Note that we have not talked about strategies in the definition of the games in extensive
form. That’s because the strategy in such a game is a more complex, derivative object.

2
A extensive form game can be written as a normal form (or strategic form) game:

The set of nodes for which player 1 is active is given by

A pure strategy for player i is a mapping from X_i to A

The condition says that player 1’s strategy at node x must belong to his available actions at x, which is A(x)

A strategy profile is
Each strategy profile gives a terminal history

The payoff at the terminal history is given by

2.4 Strategies
A pure strategy for player i is a function

si : Xi ! A satisfying si (x) 2 A(x).

As before, Si is the set of pure strategies for i.

If we fix a pure strategy profile s (one pure strategy per each player), this results in a
terminal node z 2 E and it gives rise to payo↵ for each player (Why?). Note that (Si , ui )i2I is
the strategic (or normal) form of . We can therefore analyze by analysing its strategic form

2.5 Mixed and behavioral strategies

1
L R

3,1 2
a b

6,4 1
C D

5,-3 0,14

In the example above:

• A set of pure strategies for player 1: {LC, LD, RC, RD}

• A mixed strategy is a vector of probabilities =( LC , LD , RC , RD )

• A behavioral strategy is two vectors of probabilities: ( L , R ), ( C , D ). Randomization

between L and R is independent from randomization between C and D

With perfect recall mixed and behavioral strategies are equivalent. We will use behavioral
strategies because they are more convenient (Why?). Whenever, in the contest of an extensive
form game, we use term “mixed strategy”, what we really mean is “behavioral strategy”.
Formally, a behavioral strategy of player i is

i : Xi ! (A) satisfying supp( i (x)) ⇢ A(x)

3
2.6 Example: entry deterrence game
E
OUT IN

0,2 I
A F

1,1 -1,-1

Entrant has one decision node and two choices. Two strategies corresponding to two choices,
SE = {IN, OU T }
Incumbent has one decision node and two choices. Two strategies corresponding to two
choices, SE = {A, F }. The normal form of this game is depicted below

A F
IN 1,1 -1,-1
OUT 0,2 0,2

There are two (pure strategy) Nash equilibria: (IN, A) and (OU T, F ). Consider these two
equilibria in the context of the original game, not its normal form. Equilibrium (OU T, F ) makes
little sense. To exclude equilibria like (OU T, F ) (which are called equilibria with non-credible
threats), we need an equilibrium notion that takes into account that the past cannot be changed.

2.7 Backward induction

A node x is penultimate if every action in A(x) results in a node belonging to E
Backward induction: i.e. results in a terminal node
Find a penultimate node x in the initial game
• At any penultimate node x, let ◆(x) select an action that maximizes her payo↵s

• Let ux be the resulting payo↵ vector. Delete all the branches following x, and assign payo↵
vector ux to x i.e. delete all A(x) and make x the terminal node

1
• This gives rise to new game tree . Repeat this procedure until every node has been
assigned an action.

This procedure results in backward induction (pure) strategy profile. i.e. BISP
Let us come back to our example. The strategy profile highlighted in red is the backward
induction one.

4
 E
OUT IN

0,2 I
A F

1,1 -1,-1

2.8 Backwards induction and NE

Theorem 1 If s is a backward induction strategy profile of a game of perfect information , then
s is a Nash equilibrium of . Moreover, pure strategy NE exist in games of perfect information.

Proof By contradiction suppose s is BISP but not NE. Then 9i 2 N and ŝi 2 Si such that

ui (si , s i ) < ui (ŝi , s i )

Start from penultimate nodes and see if ŝi and si are di↵erent there (in terms of payo↵s):

• If yes, we arrived at a contradiction because si selects the highest payo↵ at penultimate

nodes.

• If not remove all branches from penultimate nodes and raise the payo↵s generated by si
(or ŝi since they are the same) to these nodes.

This procedure obtains a new game. Repeat the same steps until we arrive at a contradiction.
Q.E.D.

2.9 Simultaneous moves

Consider the quality choice game from the previous section. This is a normal form game:

Firm
H(igh) L(ow)
Consumer B(uy) 2, 2 1, 3
N (o) 0, 0 0, 1

Can we rewrite it in a extensive form without loosing the feature of simultaneous moves?
We can, but we need to introduce one more tool: an information set. Let us start with the
following extensive form:

5
 C

B N

F F
H L H L

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

Clearly, this is not an accurate representation of the quality choice game because the firm can
condition its choice on the choice of the consumer in the extensive form above. However, if we
assume that the firm cannot distinguish between histories B and N , we would get an accurate
representation. To do that we put two histories—B and N —into a single information set.
The firm knows that the history is either B or N , but cannot tell which one exactly:

B N

F F
H L H L

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

Can we draw a di↵erent, yet equivalent representation? Does the order of players matter in
this case?

2.10 Incorporating simultaneous moves into an extensive form game

Here is a more complex example: a quality choice game with investment:

6
 2,2 -1,3 0,0 0,1

H L H L
F F

B N

C
Subgame 1

NI

Subgame 3
F

Subgame 2
C

B N

F F
H L H L

2,2 -1,2 0,0 0,1

2.11 Subgames
An information set is the collection of nodes that satisfy the following three conditions:

1. In all these nodes the same player (denote him by i) is taking an action;

2. The player i does not know in which node of the information set he is;

3. Naturally, the set of actions that is available in two nodes of the same information set is
the same: if x1 and x2 belong to the same information set then A(x1 ) = A(x2 ).

Take a node (a history) x and consider a subtree of the game that follows x. If all the

7
A game with perfect information means that every information set is a singleton (i.e. contains a single history), which means that
everything is observable so there is no simultaneous moves.
The theorem means that SPNE and BISP gives the same result in a game with perfect information. For infinite game or games
without perfect information, BISP cannot be used. In other words, SPNE is a extension that can be applied in more games.

information sets that intersect this subtree are contained in this subtree, then node x gives rise
to a subgame.

2.12 Subgame perfect Nash equilibrium

A subgame perfect Nash equilibrium of a game is a strategy profile s⇤ that is a Nash equilibrium
in every subgame of . i.e. SPNE

Theorem 2 A pure strategy profile s⇤ is a subgame perfect Nash equilibrium of a game of perfect
information if and only if it is a backward induction strategy profile of this game.

2.13 One-step deviation principle

One-step deviation principle If there exists a profitable deviation from a strategy profile s
for player i, there exists a subgame in which there exists a ”one-step” profitable deviation from
a strategy profile s for player i—i.e., a deviation that involves changing only one action.

1
A B

3,1 2
a b

6,4 1
C D

5,-3 0,14

In this example, player 1 has a profitable deviation from a strategy profile ((A, D), b): player
1 can increase his payo↵ from 3 to 5 by deviating to a strategy (B, C). The one-step deviation
principle tells us that there is a subgame induced by history (B, b) that has a profitable “one-
step” deviation from action D to action C.

2.14 Example: Stackelberg competition

Two firms move sequentially.
Demand: P = 90 Q.
The second firm observes the choice of the first firm and makes a decision using this extra
information:
max (90 q1 q2 )q2
q2

8
 First order condition gives us:
90 q1
q2 (q1 ) =
2
The first firm anticipates that firm 1 will react this way and takes this reaction into account:

max (90 q1 q2 (q1 ))q1

q1

If we plug in the best reply of firm 2 in firm 1’s problem we gets

⇢
90 q1
max q1
q1 2
First order condition gives us:
q1⇤ = 45
and
q2⇤ = 22.5

2.15 Example: Vote Buying

Legislature with k members, k is odd for simplicity
Two rival bills, X and Y
Two interest groups:
• X favours X: it attaches value VX to X being passed and value 0 to Y being passed

• Y favours Y : it attaches value 0 to X being passed and value VY to Y being passed

A game:
Each group sequentially o↵ers individual payments to legislators conditional on tehm voting
for the preferred bill:
• X announces schedule of contingent payments to each legislator (x1 , x2 , ..., xk )

• Y observes the o↵ers made by X and announces a schedule (y1 , y2 , ...., yk )

2.15.1 Vote Buying: Payo↵s

Legislator i votes for X if xi > yi , and votes for Y otherwise. i.e. votes for Y if y_i ≥ x_i (the legislator i would vote for Y
Payo↵ to X is when there is a tie).

( k )
This term says that X receives the X Xk
k
payoff V_x if and only if more than VX I i > xi i
half of the legislators vote for X. i=1
2 i=1

where i = 1 if i voted for X, and zero if i voted for Y Note that X and Y need to make the payment they offered
If bill Y is passed, payo↵ to Y is whether or not the bill they prefer gets passed.

( k ) k
X k X
VY I i < yi [1 i]
i=1
2 i=1

9
Consider the case when k = 3.

The common knowledge (i.e. information known by both X and Y) are the following:
1. The number of legislators (i.e. k = 3)
2. The value of V_x
3. The value of V_y

The actions for X and Y are

To find the best reply for X, we make the following claims:

1. X has to make a payment such that every legislator gets exactly the same amount.
2. If X wants to win, the group X has to drive the cost of winning for group Y all the way to V_y. In other words, X has to make it too
expensive for group Y to win.

To verify the first claim, consider the following offer made by group X:

In this case, group X can reduce x_1 by the shadowed amount to make a
pro table deviation. Doing this would not change the order x_1 ≥ x_2 ≥ x_3.

The key thing that matters is the sum of x_2 and x_3. Speci cally, group X would
be guaranteed to win if the sum is greater than V_y. However, the allocation
across x_2 and x_3 is not important.

It follows that the sum of x_2 and x_3 would not change if group X removes the
shadowed part of x_2 to x_3. Doing this would allow x_2 to be equal to x_3.

Then, group X can make the pro table deviation by removing the shadowed part
of x_1.

The only pro le in which a pro table deviation does not exist for group X is when
they pay the legislator exactly the same amount. Thus, statement 1 holds in a
equilibrium (i.e. group X must pay uniformly all the legislator in equilibrium).
Consider the general case with k legislator.
2.15.2 Vote Buying: k = 3, VY = 300
At stage 2, Y can win by buying o↵ 2 cheapest legislators
Let x1 , x2 be the lowest two values of payments by X
Y buys if and only if x1 + x2 < 300.
At stage 1, X can win most cheaply by setting

x1 = x2 = x3 = 150
X will do so only if VX > 450.
To summarize:
• If VX > 450, X promises 150 to each legislator and wins. Y o↵ers nothing.

• If VX < 450, X promises nothing. Y pays nothing and bill Y gets passed.

2.15.3 Vote Buying: General case

Let µ denote the bare majority of the legislators
k+1
µ=
2
Since Y will try to undercut µ most under-paid by X legislators, the best strategy for X is
to make equal payments.
If X makes an equal payment of c to each legislator, then X can win only if

µc VY
or
VY
c
µ
But this must be profitable for X, i.e.

kc  VX
or
kVY
 VX
µ
If the latter condition is satisfied, X pays VµY to each legislator, and Y pays nothing, and X
gets passed.
If not, then X pays nothing, Y pays nothing, and Y gets passed.

2.16 Rubinstein-Stahl bargaining solution

Two players are dividing a dollar using the following protocol: This is a game with in nitely many histories and
none of them is penultimate. The game highlights
• At odd periods, player 1 is proposing a division (x,1-x). the difference between SPNE and NE. Speci cally,
the game has many NE and a unique SPNE.
• Player 2 either accepts it or not.

10
 • If the division is accepted, the game ends and players get their shares of the dollar.
• If the division is not accepted the game continues to the next (even) period in which the
players switch their roles: Player 2 is proposing the division and player 1 either accepts or
rejects it.

Players disounts their future payo↵s with discount factors 1 and 2.

2.16.1 How to find SPNE?

• Infinitely many subgames: sort them in classes
• Each subgame is infinite: use one-step deviation principle
• Many possible strategy profiles: use payo↵s as a state variable (value functoins)

2.16.2 Value functions

Two classes of subgames: Player 1 is a proposer and player 2 is a proposer. Subgames within a
class are strategically identical to each other.

• Let v i be the maximum (supremum) equilibrium payo↵ of a player i in a subgame in which

he is a proposer and
• let wi be his maximum (supremum) eq. payo↵ in a subgame in which his opponent is
proposing.
• Let v i be the minimum (infinum) equilibrium payo↵ of a player i in a subgame in which
he is a proposer and
• let wi be his minimum (infinum) eq. payo↵ in a subgame in which his opponent is propos-
ing.

Clearly, v i v i and wi wi .

2.16.3 Proposals
• The sum of the payo↵ cannot exceed 1 because the players are dividing 1 dollar:

v1 + w2  1 (1)

v2 + w1  1 (2)

• Player 1 can always o↵er 2v2 and this o↵er will be accepted:

v1 1 2v2 (3)

• Player 1 can always reject all o↵ers

w1 1v1 (4)

11
2.16.4 Bounds on v and v
• Start with inequality (3) and plug in (1) and then (4):
v1 1 2v2 1 2 + 2 w1 1 2 + 2 1v1

Solving for v 1 we get

1 2
v1 (5)
1 1 2

• Start with inequality (1) and plug in (4) and then (5):
1 1
v1  1 w2  1 2v2 1 2
1 1 2

Solving for v 1 we get

1 2
v1  (6)
1 1 2

• Combine (5) and (6) together to get

1 2 1 2
 v1  v1 
1 1 2 1 1 2

from which we conclude that

1 2
v1 = v1 = (7)
1 1 2

• Similarly for agent 2

1 1
v2 = v2 = (8)
1 1 2

2.16.5 Bounds on w1 and w1

• From (4) and (7) :
1 (1 2)
w1 (9)
1 1 2

• From (2) and (8):

1 (1 2)
w1  1 v2 = (10)
1 1 2

• Combine (9) and (10) together to get

1 (1 2) 1 (1 2)
 w1  w1 
1 1 2 1 1 2

from which we conclude that

1 (1 2)
w1 = w1 =
1 1 2

• Similarly for agent 2

2 (1 1)
w2 = w2 =
1 1 2

12
2.16.6 Unique equilibrium
• Since w1 + v2 = w2 + v1 = 1, there is no delay and the dollar is divided immediately

• When agent i is a proposer, he proposes a division (vi , 1 vi )

• A receiver j accepts if and only if the o↵er is weakly above wj .

13
Part 1: division of sub-games into classes

Player 1 is the proposer in both of the periods highlighted in yellow. Both periods are
the same (e.g. the actions available are the same) except that the payoffs are
discounted. Every payoff that happens down the road has its equivalent counterpart
above, except for discounting. For example, the payoffs in the second highlighted (in
yellow) period are multiplied by delta_1^2 and delta_2^2. Since the payoffs are
always discounted by a positive number, this gives a positive af ne transformation,
which does not change the strategic nature of the sub-games. For the sub-game
highlighted in blue, where player 2 is proposing, there will be a equivalent (in terms
of strategic nature) sub-game somewhere down below.

The sub-games can be divided into the following classes:

Sub-games where player 1 is proposing; and
Sub-games where player 2 is proposing.

Every sub-game from the same class is identical to each other. Solving the game
requires nding a NE that is consistent across the different sub-games from the same
class.

Part 2: one step deviation principle

Suppose that there is a game and I have found a strategy pro le. I can check whether
this pro le is a SPNE by looking for other strategies for each player and show that
none of the players would want to deviate. However, achieving this is very dif cult in
this game because the strategy pro les are very complex.

The one step deviation principle says that to check whether a strategy pro le is a
SPNE, it is suf cient to check the following: at every history, ask the active player
whether he wants to change his action today assuming that the rest of actions are
xed; if at every history, every agent does not have the incentive to change his action,
then this strategy pro le is a SPNE.

The power of this principle is that instead of looking for deviation at the domain of
strategy, I can look for deviations in the domain of actions that are available at a
particular history.

Part 3: use payoffs as a state variable

Plot a diagram for a sub-game in which player 1 is the proposer. Let the y-axis be the
equilibrium payoffs for the proposer (player 1 in this case) and denote this variable by
V_1; and let the x-axis be the equilibrium payoffs for the receiver (player 2 in this
case) and denote this variable by W_2.

Plot a similar graph for a sub-game in which player 2 is the proposer (change the
variables on the axes and the subscript of the variables).

In principle, a game can have many SPNE. Suppose that there is a set of equilibrium
payoffs (i.e. every point in the set is a equilibrium payoff for some sub-game). The
extreme points of the set (i.e. the maximum and minimum payoffs) are shown by the
dashed lines and are denoted using upper and lower bars (upper bar indicates
maximum and lower bar indicates minimum). These points give the bounds on the
equilibrium payoffs.
Part 4: nd a set of inequalities

4.1 nd inequalities using feasibility

The key idea of feasibility in this game is that the sum of the equilibrium payoffs must be smaller than or equal to 1 since there
is only one dollar to be divided.

Situation 1:
Consider an equilibrium situation in which player 1 is the proposer and obtains the highest amount possible. Suppose that
player 2 receives W_2 which may or may not be his maximum or minimum payoff possible. Then, the sum of the payoffs of the
two players is given by the following:

Since W_2 must be weakly larger than min W_2 possible, the inequality above implies the following:

Situation 2:
Consider an equilibrium situation in which player 1 is the proposer and receives an amount V_1. Suppose that player 2 receives
the maximum W_2 possible, then the sum of the payoffs is the following:

Since V_1 must be weakly larger than min V_1 possible, the inequality above implies the following:

In cases where player 2 is the proposer, the following inequalities can be obtained:

4.2 sub-games in which player 1 is the proposer

Claim: there is a proposal that player 1 can make in which player 2 will always accept.
Player 1 can call for the following division:

Player 2 will always accept this because he knows that if he rejects, he will go to the next sub-game. The best equilibrium
payoff for him in the next sub-game would be maxV_2 possible. From the point of view of today, the present value of maxV_2 is
equal to delta_2 * maxV_2. In other words, player 2 would not be able to get more than the amount proposed by player 1 if he
rejects. Therefore, he will always accept.

Under this division, player 1 will get the following amount:

It must be the case that even in the worst equilibrium in this sub-game, the agent is guaranteed this amount

This is because player 1 can always follow this strategy by suggesting this division and this division been accepted. In other
words, even in the worse case scenario, player 1 will not do less than 1 - delta_2 * maxV_2.

The counterpart for player 2 is the following:

4.3 sub-games in which player 2 is the proposer

One viable strategy for player 1 in sub-games where player 2 is the proposer is to reject any division made. By doing this,
player 1 guarantees himself a payoff that is at least

This is because when he rejects an offer, he continues to a sub-game where he becomes the proposal. It follows that he will
receive an equivalent payoff that is weakly larger than minV_1. From the point view of the today, the payoff that player 1
receives from the next period, in which he proposes, is equivalent to at least delta_1 * minV_1. To conclude, player 1 would
accept an offer in the current sub-game (with player 2 is the proposer) if and only if the following inequality is satis ed:

The counterpart for player 2 is the following:

Part 5: solving the inequalities

The LHS shows all the inequalities obtained and the RHS shows the process of solving the inequalities. Note that the last four
inequalities on the left relies on the one step deviation principle.
Part 6: convert the equilibrium payoffs into strategy pro les

The sum of equilibrium payoffs for player 1 and 2 in a game in which player 1 is the proposer (i.e. the sum of V_1 and W_2) is 1.
Similarly, the sum of equilibrium payoffs of the two players in a game in which player 2 is the proposer (i.e. the sum of V_2 and
W_1) is also 1. Since the sum of the payoffs would be smaller than 1 if the players delay the division of the one dollar (due to
discounting), the fact that the sum of equilibrium payoffs is always 1 suggests there must be no delays. In other words, the
equilibrium strategy is that player 1 is going to propose some division and this division is going to be accepted in the rst
round in equilibrium in every sub-game.

The results obtained suggest two features:

The equilibrium payoffs are unique and the SPNE of the game is unique. One thing to note is that this feature means that the
set of equilibrium payoffs is a single point.
The division is going to the following:

The equilibrium strategy is the mapping from histories into actions. The action for player 1 is to propose the division above.
The action for player 2 is the following:

Part 7: intuition of the game

When the agents play this game, they realize that if the proposal gets rejected twice, they would experience a new round of
game that is same as what had happen except that the payoffs are discounted. If there is an equilibrium in which they reach at
agreement at certain time t, where t is not today, then there is a pro table deviation. The deviation is for them to arrive at a
division today, which avoids the costly delay. Thus, logically, in the equilibrium, it must be the case that the agreement is made
immediately.