Extensive Form Games:

Perfect and Imperfect Information

Raghul S Venkatesh

Raghul S Venkatesh Extensive Form Games 1 / 57

 Extensive Form Games - I
Histories, Information Sets, and Strategies

Raghul S Venkatesh Extensive Form Games 2 / 57

Motivation

In many strategic settings, players must choose actions dynamically/sequentially

Parlor games (monopoly, take-away) and chess are some prominent examples where there
is sequential decisions

A number of other applications across various fields of economics come to mind:

▶ wage bargaining in labor

▶ price competition or market entry in IO
▶ policy and valence competition in elections
▶ agenda setting and bargaining in legislatures (political economy)

Raghul S Venkatesh Extensive Form Games 3 / 57

Example: Incumbent-Entrant Game

Two firms compete in an industry

One of them (say, Incumbent or firm 2) is currently producing while the other (say, Entrant
or firm 1) wishes to enter

If the entrant decides to opt not to enter, then status quo prevails with realized payoffs
(1, 5) respectively

If the entrant chooses to enter, the incumbent decides whether to fight (e.g., predatory
pricing and undercutting) or cooperate

Fighting results in payoff (0, 0) and cooperating entails payoffs (2, 2) respectively to the
firms

Raghul S Venkatesh Extensive Form Games 4 / 57

Example: Incumbent-Entrant Game

Consider the following market entry game:

Incumbent Out In

Fight Cooperate 2
1, 5
Fight Cooperate
Out 1, 5 1, 5
Entrant
In 0, 0 2, 2 0, 0 2, 2
(a) Normal Form (b) Extensive form

Raghul S Venkatesh Extensive Form Games 5 / 57

Example: Incumbent-Entrant Game

The incumbent would always prefer the entrant to stay out and enjoy monopoly profits (5
in this example)

Whether the entrant stays out depends on the incumbent’s strategy: If incumbent decides
to fight, then it is better for the entrant to stay out

Suppose both firms make their decisions simultaneously. The normal form game has two
pure strategy Nash equilibria:

(In, Cooperate) and (Out, Fight)

Raghul S Venkatesh Extensive Form Games 6 / 57

Example: Incumbent-Entrant Game

In the normal form, an incumbent firm cannot reconsider her decision to fight or cooperate
in case an entrant firm chooses to enter

Notice that if the game is sequential, then the incumbent would prefer to cooperate with an
entrant rather than fight

Therefore the equilibrium (Out, Fight) is due to a “non-credible threat” by incumbent to fight
irrespective of the entry decision of firm 1

The equilibrium corresponding to this non-credible threat is unreasonable; the extensive

form provides an alternate way to analyze such dynamic play

Raghul S Venkatesh Extensive Form Games 7 / 57

Other Examples of Dynamic Play
Consider the following sequence of moves:
▶ Player 1 chooses between two actions A1 = {L, R}

▶ After observing Player 1’s action, player two chooses between her actions A2

▶ Payoffs are realized

Or, the following:

▶ Player 1 chooses between two actions A1 = {L, R}

▶ If Player 1’s action is L, then player two chooses between actions {A, B}; if player 1’s
action is R, then both player 1 and 2 again choose one of two actions simultaneously:
player 1 between {X, Y} and player 2 between {C, D}

▶ Payoffs are then realized

Raghul S Venkatesh Extensive Form Games 8 / 57

Extensive Form Game: Formal Definition

An finite horizon extensive form game consists of,

▶ A set of players and set of actions

▶ A set of possible histories (terminal and non-terminal)

▶ Rules of the game (i.e., order of play, information sets of each player)

▶ Players’ preferences over terminal histories (end of the game)

Raghul S Venkatesh Extensive Form Games 9 / 57

Extensive Form Game: Formal Definition
Formally, an extensive form game denoted by Γ:

1. A set of players N and set of all possible actions A

2. A set of nodes or histories, X, each of which is a finite sequence of actions; X consists of

the initial node, call it x0 or ∅

▶ Each subsequent node x ∈ X is of the form x = (a1 , a2 , ..., ak ) for finitely many actions
ai ∈ A

▶ If (a1 , a2 , ..., at ) ∈ X then so too is (a1 , a2 , ..., at−1 ), for all t > 1

Node or history is simply the set of actions taken so far in the game. Consequently,

A(x) ≡ {a ∈ A | (x, a) ∈ X}

denotes the set of actions available to player who moves at node x

Raghul S Venkatesh Extensive Form Games 10 / 57

Extensive Form Game: Formal Definition

3. A set of terminal histories or end nodes E where,

E ≡ {x ∈ X | (x, a) ̸∈ X for all a ∈ A}

Each end node therefore represents one complete play of the game from starting to the
end

4. A player function P : X \ E → N that indicates who takes an action at each decision

node in X, where
Xi ≡ {x ∈ X \ E | P(x) = i}
denotes the set of decision nodes belonging to i.

Raghul S Venkatesh Extensive Form Games 11 / 57

Extensive Form Game: Formal Definition

5. A partition, I , of decision nodes X \ E such that if two nodes x′ and x′′ are in the same
element of the partition, then,

P(x′ ) = P(x′′ ) A(x′ ) = A(x′′ )

Therefore I partitions the decision nodes into information sets

Information set containing node x is denoted by I(x)

When this decision node is reached, player P(x) takes an action under the information
that the history is one of elements of I(x) (in the case of imperfect information, I(x)
may not be a singleton!)

Let Ii denote the information sets belonging to player i

Raghul S Venkatesh Extensive Form Games 12 / 57

Extensive Form Game: Formal Definition

6. (Payoffs) For every player i ∈ N, a von Neumann-Morgenstern utility function

ui : E → R that provides a payoff to each player at every possible terminal node.

Together,
Γ = ⟨(N, A), (X, E), (P, I), (ui )i∈N ⟩
is called a finite extensive form game

To fix ideas, let us go through a few examples.

Raghul S Venkatesh Extensive Form Games 13 / 57

Example 1

Out In

2
(1, 5)
Fight Cooperate

(0, 0) (2, 2)

Raghul S Venkatesh Extensive Form Games 14 / 57

Example 2

Player 1

L R
Player 2 Player 2
a b a b

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

Raghul S Venkatesh Extensive Form Games 15 / 57

Example 3

L R
2 2

A B C D
1
3, 2 6, 4
X Y X Y

3, 0 8, 5 4, 6 2, 1

Raghul S Venkatesh Extensive Form Games 16 / 57

Strategies

A pure strategy in an extensive form game is a “complete description” of decisions that a

player makes in the game

In other words, (pure) strategy specifies actions for every contingency of the game, i.e., an
action at every information set of player i

Formally, a pure strategy in game Γ is a mapping si : Ii → A such that si (I(x)) ∈ A(x) for
all x such that P(x) = i

Si denotes the set of all pure strategies for player i

Raghul S Venkatesh Extensive Form Games 17 / 57

Example 4

Player 1

L R
Player 3 Player 2
a b a b
    Player 3 Player 3
2 −1 r r
0  5  l l
1 6       
3 5 0 −2
 1   4   − 1  2 
2 4 7 0

Raghul S Venkatesh Extensive Form Games 18 / 57

Strategies in Example 4

First we define what the information sets of players are:

I1 = { ∅ } I2 = {(R)} I3 = {(L), (R, a), (R, b)}

Since P1 and P2 have only one information set it is straightforward, s1 = L or R and

s2 = a or b
S1 = {L, R} S2 = {a, b}
For player 3 a strategy consists of an action at each of the three information sets

 action a if (L)
s3 = action l if (R, a) ≡ all
action l if (R, b)


Raghul S Venkatesh Extensive Form Games 19 / 57

Strategies in Example 4

Similarly, a strategy s3 = brl is simply the following instruction:


 action b if (L)
s3 = action r if (R, a)
action l if (R, b)


Consequently, the set of (pure) strategies available to P3 is given by,

S3 = {all, alr, arl, arr, bll, blr, brl, brr}

What if P3 did not observe the actions of P2 in Example 3? How does it change the
strategies of the players?

Raghul S Venkatesh Extensive Form Games 20 / 57

Example 4.1

Player 1

L R
Player 3 Player 2
a b a b
3
    Player 3 Player 3
2 −1 r r
0  5  l l
1 6       
3 5 0 −2
 1   4   − 1  2 
2 4 7 0

Raghul S Venkatesh Extensive Form Games 21 / 57

Strategic form game

Notice that we have inadvertently converted an extensive form game into strategic form by
specifying the strategy set Si (which is of course different from action set Ai )

Together the tuple (Si , ui )i∈N is the strategic form of Γ

This means we can apply the notion of Nash equilibria to finite extensive form games – all
we require to do is specify strategy set of each player

A Nash equilibrium for the extensive form game Γ is simply a “joint strategy” that is a
Nash equilibrium of the strategic form of Γ

Raghul S Venkatesh Extensive Form Games 22 / 57

Example 5
1
A B
2
C D 3, 3
1
E F 1, 1

2, 0 0, 2

Here, P(∅) = 1. The information sets of the two players are,

I1 = {∅, (A, C)} I2 = {(A)}

Correspondingly the strategy set for the two players are,

S1 = {AE, AF, BE, BF} S2 = {C, D}

Raghul S Venkatesh Extensive Form Games 23 / 57

Example 5
C D
1
A B AE 2,0 1,1
2
C D 3,3
AF 0,2 1,1
1
E F 1,1
BE 3,3 3,3

2,0 0,2
BF 3,3 3,3

The strategic form game has 4 Nash equilibria:

{ (BE, C) , (BE, D) , (BF, C) , (BF, D) }

Are these equilibria consistent with rational behavior?

Raghul S Venkatesh Extensive Form Games 24 / 57
Recap

▶ We first partition the decision nodes into information sets, and assign them to players
making the decision at any information set

▶ Depending on whether the extensive game has perfect information or imperfect

information, the information sets are singleton or non-singleton respectively

▶ For each player, a pure strategy specifies an action at every information set where the
player is the decision maker

▶ Once strategies are specified, we can convert the extensive game into a normal
(strategic) form game and solve for the Nash equilibria

▶ What about other notions of equilibrium in extensive form games?

Raghul S Venkatesh Extensive Form Games 25 / 57

 Extensive Form Games - II
Backwards Induction and Subgame Perfection

Raghul S Venkatesh Extensive Form Games 26 / 57

Perfect Information and Backward Induction

When players are perfectly informed of all actions taken previously then the information
sets are all singleton nodes (i.e., consisting of a single history of actions)

Specifically, I(x) = x for all x ∈ X \ E

Backward induction (BI) procedure premises that when player clearly observes the future
order of play, they will choose an action that leads to the highest payoff

Intuitively, this implies solving the game backwards one step at a time starting from
penultimate nodes that lead to the terminal history

Raghul S Venkatesh Extensive Form Games 27 / 57

Perfect Information and Backward Induction

In order to describe the BI process, we require a notion of diameter of any history/node x

Definition of Diameter of node x:

The diameter of a node x is the largest K such that there are actions a1 , a2 , ..., aK that lead to a
(terminal) node (x, a1 , a2 , ..., aK )

The diameter of a terminal node is zero; diameter of the null history (starting node) is the
number of actions in the longest history of the game

Raghul S Venkatesh Extensive Form Games 28 / 57

Perfect Information and Backward Induction

BI procedure first specifies an action for any player that moves at all of the nodes with
diameter 1

Suppose the total number of nodes of diameter 1 is n1 . Then the nodes of distance 1 can be
written as (xi1 )ni=1 1

At each of these nodes xi1 , the player P(xi1 ) is supposed to make a decision. Denote this
strategy as sP(xi ) (xi1 ) ∈ A(xi1 )
1

sP(xi ) (xi1 ) is the strategy of player P(xi1 ) at the history xi1 . This strategy (action) maximizes
1
the player’s payoff

We denote this utility as ui1 , i.e. the utility profile of node i of the ones that are at a diameter
of 1

Raghul S Venkatesh Extensive Form Games 29 / 57

Perfect Information and Backward Induction

Once the payoff vector for all the penultimate nodes are determined, u1 = (u11 , ..., un1 1 ),
remove those nodes and the actions that follow them.

And assign the payoff vector u1 to the respective nodes. This now shortens the length of
the game tree and makes this payoff vector the end nodes of Γ

Repeat the same process by assessing the nodes that are at a diameter 1 from these new
terminal nodes, until an action has been assigned to every decision node

Raghul S Venkatesh Extensive Form Games 30 / 57

Backward Induction

1
A B
2
C D 3, 3
1
E F 1, 1

2, 0 0, 2

We can apply the BI procedure by starting at node where player 1 moves for the second
time

Can you compute the strategy profile generated by BI procedure?

Raghul S Venkatesh Extensive Form Games 31 / 57

Example 6

L R
2 2

A B A B
3 3 3 3

a b a b a b a b

              
8 6 5 2 7 3 1 0
3  5  8 7 6 4 0  1 
4 10 3 5 8 2 2 −1

Raghul S Venkatesh Extensive Form Games 32 / 57

Centipede Game

Description of the game –

Two players sit across a table with two pots in front of them. In the beginning, one pot
contains two gold coins and other contains nothing.

Player 1 moves first. Each player has two actions available: continue (C) or stop (S).

If a player continues then a gold coin is added to each of the pots. If a player stops, he/she
gets the bigger pot for themselves.

At the end of three moves by each player, if the game has not been stopped, player 1 gets
the larger pot.

Raghul S Venkatesh Extensive Form Games 33 / 57

Centipede Game

1 C 2 C 1 C 2 C 1 C 2 C
8, 6

S S S S S S

2, 0 1, 3 4, 2 3, 5 6, 4 5, 7

Raghul S Venkatesh Extensive Form Games 34 / 57

Centipede Game
Backward Induction Reasoning:

▶ If we were to reach the last round of the game, Player 2 would do better by choosing S
instead of C

▶ However, given that 2 will choose S, player 1 should choose S in the second to last
round, receiving 6 instead of 5

▶ Given that 1 would choose S in the second to last round, 2 would choose S in the third
to last round, receiving 5 instead of 4

▶ And so on.....

▶ At the beginning, Player 1 should choose to stop and receive 2 instead of getting 1 in
the second round

The SPE of the game is therefore S by both players at every history of the game. The
payoffs are (2, 0) respectively.
Raghul S Venkatesh Extensive Form Games 35 / 57
Imperfect Information and Subgame Perfection

OUT IN
1
3, 3

a b
2

A B A B

1, 3 0, 0 0, 0 3, 1

Clearly this is a game of imperfect information, i.e., there is at least one information set that
is a non-singleton. How do we analyze this game?

Raghul S Venkatesh Extensive Form Games 36 / 57

Imperfect Information and Subgame Perfection

To understand the idea of Subgame Perfect equilibrium (SPE) we need an idea of what a
subgame is

In order to define a subgame we require the following:

Node y strictly follows a node x if y = (x, a1 , ..., am ) for some sequence of actions
a1 , a2 , ..., am ∈ A

y immediately follows x if m = 1; weakly follows if either y = x or y strictly follows x

Raghul S Venkatesh Extensive Form Games 37 / 57

Imperfect Information and Subgame Perfection

Definition
A node x is said to define a subgame of Γ if whenever y weakly follows x and another node
z is in the IS containing y, then z also weakly follows x.

From the definition it is clear that for a node to begin a subgame, it must be a singleton, i.e.,
I(x) = {x}

Further, the definition also implies that every player at every turn subsequently knows
whether x has been reached in the game

Every game Γ has at least one subgame, the game itself!

Raghul S Venkatesh Extensive Form Games 38 / 57

Example 6
Count the Subgames

L R
2 2

A B A B
3 3 3 3

a b a b a b a b

              
8 6 5 2 7 3 1 0
3  5  8 7 6 4 0  1 
4 10 3 5 8 2 2 −1

Raghul S Venkatesh Extensive Form Games 39 / 57

Modified Example 6
Count the Subgames

L R
2 2
3 3
A B A B

a b a b a b a b

              
8 6 5 2 7 3 1 0
3  5  8 7 6 4 0  1 
4 10 3 5 8 2 2 −1

Raghul S Venkatesh Extensive Form Games 40 / 57

Imperfect Information and Subgame Perfection
Definition

A strategy profile s = (si )i∈N is a subgame perfect equilibrium of Γ if for every player i ∈ N
and every non-terminal node x for which P(x) = i,

ui (x, s) ≥ ui (x, (s−i , ri )) for all ri ∈ Si

where, the node (x, s) is a terminal node that follows from node x and strategy s

The difference between SPE and NE is the line “and every non-terminal node x for which
P(x) = i”

This implies that SPE requires a player’s strategy to be optimal after every history of the
game (given other players’ strategies)

NE on the other hand requires player’s strategies to be optimal only at the beginning of the
game

Raghul S Venkatesh Extensive Form Games 41 / 57

Imperfect Information and Subgame Perfection

Requiring rationality at every stage of decision-making implies SPE is optimal even when a
history is not reached under a strategy profile

SPE idea is a refinement of the Nash equilibrium. Since the whole game is a subgame, a pure
strategy SPE is also a pure strategy NE

Intuitively, it follows that every subgame perfect equilibrium is a NE, but the converse is
not true (can you prove this?)

Definition

A joint pure strategy profile s is a subgame perfect equilibrium of Γ if s induces a Nash

equilibrium in every subgame of Γ

Raghul S Venkatesh Extensive Form Games 42 / 57

Imperfect Information and Subgame Perfection

OUT IN
1
3, 3

a b
2

A B A B

1, 3 0, 0 0, 0 3, 1

We can still use backward induction in this game! We treat the subgame with simultaneous
moves as if it were a game in its own right

Raghul S Venkatesh Extensive Form Games 43 / 57

Example 3 Revisited

Z W
2 2

D C A B
1
3, 2 6, 4
X Y X Y

3, 0 8, 5 4, 6 2, 1

This game has similar features and can be analyzed using techniques used in the previous
problem. The only difference is that there is an additional subgame starting at player 2 on
the left node

Raghul S Venkatesh Extensive Form Games 44 / 57

Imperfect Information and Subgame Perfection
1

OUT IN
2
 
1
2
L R
1

l r l r
       
0 3 −1 0
0 1 3 0

The subgame starting at player 2 has an unique pure strategy equilibrium (does it have a
mixed strategy equilibrium?): (L, r)

Backward induction leads to 1 taking action IN at the beginning of the game.

Raghul S Venkatesh Extensive Form Games 45 / 57
Imperfect Information and Mixed/Behavioral Strategy

Definition: A pure strategy for player i is a mapping si : Ii → Ai that assigns an action

si (ιi ) ∈ Ai (ιi ) for every information set ιi ∈ Ii . Si is the set of all pure-strategy mappings
where si ∈ Si

Pure strategy SPE may not always exist in extensive games. In order to characterize SPE in
the absence of pure strategies, one has to rely on randomization. A mixed strategy for
player i is a probability distribution over his pure strategies si ∈ Si

Definition: A behavioral strategy specifies for each information set ιi ∈ Ii , an

independent probability distribution over Ai (ιi ) and is denoted by σi : Ii → ∆Ai (ιi )

where σi (ai (ιi )) is the probability that player i plays action ai (ιi ) ∈ Ai (ιi ) in information set ιi

Raghul S Venkatesh Extensive Form Games 46 / 57

Imperfect Information and Mixed/Behavioral Strategy

A subgame may not have a NE in pure strategies. In this case randomization by players
would allow for the existence of a SPE.

Definition: Let Γ be an n-player extensive-form game. A behavioral strategy profile

σ∗ = (σ1∗ , ..., σn∗ ) is a SPE if for every proper subgame G of Γ the restriction of σ∗ to the
subgame G is a Nash equilibrium in G

SPE requires not only that a NE profile of strategies be a combination of best responses on
the equilibrium path, but also consist of mutual best responses off the equilibrium path

Intuitively, randomization is applied whenever it is a player’s turn to move in the game,

instead of everything in the beginning approach

This means during the course of the game players can choose different randomization
depending on which turn it is

Raghul S Venkatesh Extensive Form Games 47 / 57

Imperfect Information and Mixed/Behavioral Strategy

OUT IN
2
 
1
2
L R
1

l r l r
       
0 4 2 0
2 1 1 3

In this subgame, there are no pure strategies. However, there is a mixed strategy
equilibrium in the subgame and this can be computed to solve the game backwards

Raghul S Venkatesh Extensive Form Games 48 / 57

Imperfect Information and Mixed/Behavioral Strategy

OUT IN
2
 
1
2
q 1−q
1
p 1−p p 1−p
       
0 4 2 0
2 1 1 3

2 1 4 5



The mixed strategy solution is (p, q) = 3, 3 . The equilibrium payoff is 3, 3

Raghul S Venkatesh Extensive Form Games 49 / 57

Imperfect Information and Mixed/Behavioral Strategy

OUT IN
 ∗∗
1, 2 4 5
3, 3

The SPE is the following:

▶ Player 1 plays IN with probability 1

1
▶ Player 2 plays L (resp. R) with probability 3 (resp. 23 )

2
▶ Player 1 to play l (resp. r) with probability 3 (resp. 13 )

Raghul S Venkatesh Extensive Form Games 50 / 57

Imperfect Information and Mixed/Behavioral Strategy

(0) (1)
2
 
1
2
   
1 2
3 3
1

       
2 1 2 1
3 3 3 3

       
0 4 2 0
2 1 1 3

Raghul S Venkatesh Extensive Form Games 51 / 57

Imperfect Information and Mixed/Behavioral Strategy

NO
YES
1
 
1.5
1.5
O B
2

o b o b
       
2 0 0 1
1 0 0 2

Dynamic version of BoS game with voluntary participation by player 1

Raghul S Venkatesh Extensive Form Games 52 / 57

Imperfect Information and Mixed/Behavioral Strategy

L R
M
2 2

A B a b a b

1
       
3 1 1 3
l r l r 0 1 1 0

       
0 4 2 0
2 1 1 3

Raghul S Venkatesh Extensive Form Games 53 / 57

Stackelberg Duopoly

▶ It was introduced and analyzed by Heinrich von Stackelberg (1934). It illustrates an

important point: the order of moves might matter, and rational actors will take this
into account

▶ Consider the familiar “Cournot game” with demand p = 100 − q, q = q1 + q2 , and

ci (qi ) = 10qi for i ∈ {1, 2}

▶ We have already solved for the best-response functions of each firm by maximizing
their profit functions when each takes the other’s quantity as fixed (the
simultaneous-move game)

▶ We now change a small but important detail of the game and assume that player 1 will
choose q1 first, and that player 2 will observe the choice made by player 1 before it
makes its choice of q2 . What is the backward induction solution of this game?

Raghul S Venkatesh Extensive Form Games 54 / 57

Mutually Assured Destruction

▶ Game theoretic analysis to represent the Cuban missile crisis of 1962. The crisis started
with the United States’ discovery of Soviet nuclear missiles in Cuba, after which the
United States escalated the crisis by quarantining Cuba

▶ The USSR then backed down, agreeing to remove its missiles from Cuba, which
suggests that the United States had a credible threat along the lines of “if you don’t
back off we both pay dearly.” Could this indeed be a credible threat?

Raghul S Venkatesh Extensive Form Games 55 / 57

Mutually Assured Destruction

▶ To analyze this as a game, consider two superpowers, player 1 (the United States) and
player 2 (the USSR), that have engaged in a provocative incident

▶ The game starts with player 1’s choice to either ignore the incident (I), resulting in
maintenance of the status quo with payoffs (0, 0), or escalate the situation (E).
Following escalation by player 1, player 2 can back down (B), causing it to lose face
and resulting in payoffs of (10, −10), or it can choose to proceed to a nuclear
confrontation (N).

▶ Upon the choice of N, the players play a simultaneous-move game in which they can
either retreat (R for player 1, r for player 2) or choose Doomsday (D for player 1, d for
player 2), in which the world is all but destroyed. If both call things off and retreat
then they suffer a small loss due to the mobilization process and payoffs are (−5, −5),
while if either party chooses Doomsday then the world destructs and payoffs are
(−100, −100)

Raghul S Venkatesh Extensive Form Games 56 / 57

Thank You