CHAPTER 2

Extensive Form Games

Again, we begin our discussion of extensive form games without defining what
one is, but giving some examples.

1. Examples of extensive form games

4, 3, 7, 11 2, 9, 16, 3 10, 1, 9, 8 1, 10, 8, 9

X Y X Y
1, 1, 1, 1 1, 1, 1, 1
4
U D A B

2 3

L R

Figure 1. An extensive form game.

Look at Figure 1. We interpret this as follows. Each point where a player gets
to move in the game or at which the game ends is called a node. Nodes at which
players move are shown by small black dots in Figure 1 and are called decision
nodes. The game starts at a particular node, called the initial node or root. In
this case we assume that the lowest node where Player 1 moves is the initial node.
Player 1 chooses between L or R. If Player 1 chooses L then Player 2 moves and
chooses between U or D. If Player 1 chooses R then Player 3 moves and chooses
between A and B. If Player 2 chooses U then the game ends. If Player 2 chooses D
then player 4 moves. If player 3 chooses B then the game ends. If player 3 chooses
A then Player 4 moves. When it’s Player 4’s turn to move, he doesn’t see whether
he is called upon to move because Player 1 chose L and Player 2 chose D or because
Player 1 chose R and Player 3 chose A. We say that the two nodes at which player
4 moves are in the same information set and we represent this by joining them with
a dotted line as in Figure 1. If it’s player 4’s turn to move, he chooses between X
and Y , after which the game ends. The nodes at which the game ends are called
terminal nodes. To each terminal node we associate a payoff to each player. These
payoffs tell us how the player evaluates the game ending at that particular node,
that is they tell us the players’ preferences over the terminal nodes, as well as their
preferences over randomisations over those nodes.
23
24 2. EXTENSIVE FORM GAMES

−2, 2 2, −2 2, −2 −2, 2

S D S D
1, −1 −1, 1
1
OUT IN In Out

2 1

1 1
2 2

Nature

Figure 2. Another extensive form game. This game does not

have perfect recall.

As another example, consider the extensive form game shown in Figure 2. In

this game, the first mover is not a player but “Nature”. That is, at the beginning of
the game, there is a random selection of whether Player 1 or Player 2 gets to move,
each being chosen with probability 12 . (I shall indicate such moves of Nature by an
open circle, while nodes at which real players move are indicated by closed dots.
The reader is warned that such a convention is not universally followed elsewhere.)
If Player 1 is chosen, Player 1 gets to choose between In and Out, and if Player 2 is
chosen, Player 2 gets to choose between IN and OUT. If either Player 1 chooses Out
or Player 2 chooses OUT if chosen to move then the game ends. If Player 1 made
the choice then Player 1 gets −1 and Player 2 gets 1 while if Player 2 made the
choice, Player 1 gets 1 and Player 2 gets −1. On the other hand, if either Player 1
chose In or Player 2 chose IN then Player 1 gets to move again, but Player 1 does
not know if it was Player 1 or Player 2 that moved previously. This may seem a bit
bizarre because it means that Player 1 forgets whether he moved previously in the
game or not. Thus we say that this is a game without perfect recall (see Section 3
below). One way to rationalise this is to think of Player 1 as a ‘team’. So the first
time Player 1 gets to move (if Player 1 was chosen by Nature), it is one person or
agent in the “Player 1 team” moving, and the second time it is another agent in
the “Player 1 team”, and neither agent in the Player 1 team can communicate with
the other. Finally, if Player 1 gets to move for the second time, he gets to choose
between S or D, and the appropriate payoffs are given as in Figure 2.
Let’s summarise the terminology we have introduced. In a game such as those
in Figures 1 or 1, we call nodes at which a player moves decision nodes. A node
at which some exogenous randomisation occurs is also called a decision node and
we say that “Nature” moves at this node. The first node of the game is called
the initial node or simply the root. Nodes at which the game ends and payoffs are
specified are called terminal nodes. When a player cannot distinguish between two
nodes at which he or she gets to move, as is the case for the second move of Player 1
in the game in Figure 2, we say that these two nodes are in the same information
set of that player.
 2. DEFINITION OF AN EXTENSIVE FORM GAME 25

2. Definition of an extensive form game

One of the essential building blocks of an extensive form game is the game
tree, g. Consider Figure 3. Formally, a game tree is a finite connected graph
with no loops and a distinguished initial node. What does this mean? A fi-
nite graph is a finite set of nodes, X = {x1 , x2 , . . . , xK } and a set of branches
connecting them. A branch is a set of two different nodes, {xi , xj } where xi 6=
xj . In Figure 3, the set of nodes is X = {x1 , . . . , x9 } and the branches are
{x1 , x2 } , {x2 , x5 } , {x6 , x2 } , {x1 , x9 } , {x9 , x3 } , {x9 , x4 } , {x3 , x7 } , and {x3 , x8 }. The
initial node, or root, is x1 .

x6 x7 x8 x9

x3 x4 x5

x2

x1

Figure 3. A game tree.

In a game tree, we want to rule out trees that look like those in Figures
4 and 5. That is, we don’t want to have trees with loops (as in Figure 4) or
trees with more than one initial node (as in Figure 4). A path is a sequence of
branches ({x1 , x2 } , {x2 , x3 } , . . . , {xT −1 , xT }). For example, in Figure 3, one path
is ({x7 , x3 } , {x3 , x8 }).1 Then, to avoid problem trees like those in Figures 4 and 5,
we make the following assumption: Every node except the root is connected to the
root by one and only one path.

x4

x2 x3

x1

Figure 4. Not a game tree.

1Notice that in our definition of a path, we didn’t say that we are only moving in one direction
in the game tree.
26 2. EXTENSIVE FORM GAMES

x3 x4

x1 x2

Figure 5. Not a game tree either.

As an alternative way of avoiding game trees like those in Figures 4 and 5,

given a finite set of nodes X, we define the immediate predecessor function p :
X → X ∪ {∅}, to be the function that gives the node that comes immediately
before any given node in the game tree. For example, in Figure 3, p (x2 ) = x1 ,
p (x3 ) = x9 , p (x7 ) = x3 , and so on. To make sure that a game tree has only one
initial node, we require that there is a unique x0 ∈ X such that p (x0 ) = ∅, that is,
that there is only one node with no immediate predecessor. To prevent loops in the
game tree, we require that for every x 6= x0 , either p (x) = x0 or p (p (x)) = x0 or
. . . p (p (p . . . p (x))) = x0 . That is, by applying the immediate predecessor function
p to any node except for the initial node, eventually we get to the initial node x 0 .
The terminal nodes in Figure 3 are x5 , x6 , x7 , x8 , and x9 . We denote the
set of terminal nodes by T . A terminal node is a node such that there is a path
({x0 , x1 } , {x1 , x2 } , . . . , {xt−1 , xt }) and there are no paths that extend this. Alter-
natively, we can say that xt is a terminal node if there is no node y ∈ X with
p (y) = xt .
Now we can give a formal definition of an extensive form game.
Definition 2.1. An extensive form game consists of:
(1) N = {1, 2, . . . , N } a finite set of players.
(2) X a set of nodes.
(3) X is a game tree.
(4) A set of actions, A, and a labelling function α : X\ {x0 } → A where α (x)
is the action at the predecessor of x that leads to x. If p (x) = p (x0 ) and
x 6= x0 then α (x) 6= α (x0 ).
(5) H a collection of information sets and H : X\T → H a function that
assigns for every node, except the terminal ones, which information set
the node is in.
(6) A function n : H → N ∪ {0}, where player 0 denotes Nature. That
is, n (H) is the player who moves at information set H. Let Hn =
{H ∈ H|n (H) = n} be the information sets controlled by player n.
(7) ρ : H0 × A → [0, 1] giving the probability that action a is taken at the
information set H of Nature. That is, ρ (H, a).
(8) (u1 , . . . , uN ) with un : T → R being the payoff to player n.
This completes the definition of an extensive form game. Just a few more
comments are in order at this stage.
First, let C (x) = {a ∈ A|a = αx0 for some x0 with p (x0 ) = x}. That is, C (x)
is the set of choices that are available at node x. Note that if x is a terminal node
then C (x) = ∅. If two nodes x and x0 are in the same information set, that is,
 3. PERFECT RECALL 27

if H (x) = H (x0 ), then the same choices must be available at x and x0 , that is,
C (x) = C (x0 ).
To illustrate this, consider the “game” shown in Figure 6. This figure illustrates
the “forgetful driver”: A student is driving home after spending the evening at a
pub. He reaches a set of traffic lights and can choose to go left or right. If he
goes left, he falls off a cliff. If he goes right, he reaches another set of traffic lights.
However when he gets to this second set of traffic lights, since he’s had a few drinks
this evening, he cannot remember if he passed a set of traffic lights already or not.
At the second traffic lights he can again either go left or right. If he goes left he gets
home and if he goes right he reaches a rooming house. The fact that he forgets at
the second set of traffic lights whether he’s already passed the first set is indicated
in Figure 6 by the nodes x0 and x1 being in the same information set. That is,
H (x0 ) = H (x1 ). Under our definition, this is not a proper extensive form game.
Under our definition, if two nodes are in the same information set, neither should
be a predecessor of the other.

Rooming
Home House

L R
x1
Cliff

L R

x0

Figure 6. The forgetful driver.

Finally, we’ll assume that all H ∈ H0 are singletons, that is, sets consisting of
only one element. In other words, if H ∈ H0 and H (x) = H (x0 ) then x = x0 . This
says that Nature’s information sets always have only one node in them.

3. Perfect recall
An informal definition is that a player has perfect recall if he remembers ev-
erything that he knew and everything that he did in the past. This has a few
implications:
• If the player has only one information set then the player has perfect recall
(because he has no past!).
• If it is never the case that two different information sets of the player both
occur in a single play of the game then the player has perfect recall.
• If every information set is a single node we say the game has perfect
information. In this case a player when called upon to move sees exactly
what has happened in the past. If the game has perfect information then
each player has perfect recall.
28 2. EXTENSIVE FORM GAMES

Now let us give a formal definition of perfect recall. We define it in terms of

a player, and if every player in the game has perfect recall, we say that the game
itself has perfect recall.
Definition 2.2. Given an extensive form game, we say that player n in that
game has perfect recall if whenever H (x) = H (x0 ) ∈ Hn , that is, whenever x and
x0 are in the same information set and player n moves at that information set, with
x00 a predecessor of x with H (x00 ) ∈ Hn , that is, x00 comes before x and player
n moves at x00 , and a00 is the action at x00 on the path to x, then there is some
x000 ∈ H (x00 ) a predecessor of x0 with a00 the action at x000 on the path to x0 .

4. Equilibria of extensive form games

To find the Nash equilibria of an extensive form game, we have two choices.
First, we could find the equivalent normal form game and find all the equilibria
from that game, using the methods that we learned in the previous section. The
only trouble with this is that two different extensive form games can give the same
normal form. Alternatively, we can find the equilibria directly from the extensive
form using a concept (to be explained in subsection 9 below) called subgame perfect
equilibrium (SPE). Note that we cannot find the SPE of an extensive form game
from its associated normal form; we must find it directly from the extensive form.
The two approaches that we can take to finding the equilibria of an extensive form
game are shown in Figure 7.

Extensive Normal
form form Solutions
games games
*
-
1 Equilibrium 1
Normal form correspondence
function
Subgame Perfect
Equilibrium correspondence

Figure 7. Methods for defining equilibria of extensive form

games. Note that two different extensive form games may have
the same normal form and that a single game may have multiple
equilibria, some of which may be subgame perfect and some of
which may not.

5. The associated normal form

Let us first consider the method for finding equilibria of extensive form games
whereby we find the Nash equilibria of the associated normal form. Consider the
extensive form game shown in Figure 8. To find the associated normal form of this
game, we first need to know what a strategy of a player is. As we said before, a
 5. THE ASSOCIATED NORMAL FORM 29

strategy for a player is a complete set of instructions as to how to play in the game.
More formally, we have the following definition.

3, 3 0, 0

U D
1, 2 1

L R
2, 1 2

T B

Figure 8. Yet another extensive form game.

Definition 2.3. A strategy for player n is a function that assigns to each

information set of that player a choice at that information set.
Note that this definition doesn’t take account of the fact that certain choices
by the players may render some of their information sets irrelevant. For example,
in Figure 8, if Player 1 plays T at her first information set, her second information
set is never reached. However, a strategy for Player 1 must still specify what she
would do if her second information set were reached. Thus, strictly speaking, T U
and T D are distinct strategies for Player 1. However, two strategies that for any
fixed choice of the other player lead to the same outcome are said to be equivalent.
In the game in Figure 8, the strategies T U and T D are equivalent for Player 1.
Now that we know what a strategy in an extensive form game is, we can set
about deriving the associated normal form. Figure 9 shows the associated normal
form game of the extensive form game in Figure 8. Player 1’s strategies are T U ,
T D, BU and BD and Player 2’s strategies are L and R. We then construct the
associated normal form by simply writing all the strategies of Player 1 down the
left hand side and all the strategies of Player 2 across the top, and then filling in
each cell in the matrix with the appropriate payoff according to the strategies that
are played by each player.
Exercise 2.1. Find all the Nash equilibria of the game in Figure 9.
As another example, consider the extensive form game in Figure 2. In this
game, Player 2’s strategy set is S2 = {IN, OUT} and Player 1’s strategy set is
S1 = {(In,S) , (In,D) , (Out,S) , (Out,D)}. In this game, because there is a random
move by Nature at the beginning of the game, a profile of pure strategies of the
player generates a probability distribution over the terminal nodes. To generate the
associated normal form of the game, we then use this probability distribution to
calculate the expected value of the terminal payoffs. Thus the associated normal
form of the extensive form game in Figure 2 is shown in Figure 10. As an example
of how the payoffs were calculated, suppose Player 1 plays (In,S) and Player 2
plays IN. In this case, the payoff is (−2, 2) with probability 12 and (2, −2) with
30 2. EXTENSIVE FORM GAMES

Player 2
L R
TU 2, 1 2, 1
Player 1 TD 2, 1 2, 1
BU 1, 2 3, 3
BD 1, 2 0, 0

Figure 9. The normal form game corresponding to the extensive

form game in Figure 8.

probability 21 . Thus the expected payoffs are 12 · (−2, 2) + 21 · (2, −2) = (0, 0). If
instead Player 1 plays (In,S) and
Player 2 plays OUT, the expected payoffs are
1 1 1 1
2 · (1, −1) + 2 · (2, −2) = 1 2 , −1 2 . The other payoffs are calculated similarly.

Player 2
IN OU T
3 3
In, S 0, 0 2, −2

Player 1 In, D 0, 0 − 12 , 21
Out, S − 32 , 32 0, 0
1 1
Out, D 2, −2 0, 0

Figure 10. The normal form game corresponding to the extensive

form game in Figure 2.

Let us complete this example by finding all the Nash equilibria of the nor-
mal form game in Figure 10. Note that for Player 1, the strategies (In,
D) and
(Out, S) are strictly dominated by, for example, the strategy 21 , 0, 0, 21 . To see
this, suppose that Player 1 plays (In, D) and Player 2 plays a mixed strategy
(y, 1
− y) where 0 ≤ y ≤ 1. Then Player 1’s expected payoff from (In, D) is 0 · y +
− 21 · (1
− y) = − 21 + 12 y. Suppose instead that Player 1 plays the mixed strategy

1 1
2 , 0, 0, 2 . Player 1’s

expected payoff from this strategy is 21
0 · y + 1 12 · (1 − y) +
1 1 3 1 3 1 1 1 5
2 2 · y + 0 · (1 − y)
= 4 − 2 y. Since 4 − 2 y − − 2 + 2 y = 4 − y ≥ 0 for any
1 1
0 ≤ y ≤ 1, 2 , 0, 0, 2 gives Player 1 a higher expected payoff than (In,D) whatever
Player 2 does. Similar calculations show that (Out,S) is also strictly dominated
for Player 1. Thus we know that Player 1 will never play (In,D) or (Out,S) in a
Nash equilibrium. This means that we can reduce the game to the one found in
Figure 11.
From Figure 11, we see that the game looks rather like that of matching pennies
from Figure 6. We know that in such a game there are no pure strategy equilibria.
In fact, there is no equilibrium in which either player plays a pure strategy. So
we only need to check for mixed strategy equilibria. Suppose that Player 1 plays
(In,S) with probability x and (Out,D) with probability 1 − x, and Player 2 plays
IN with probability y and OUT with probability 1 − y. Then x must be such that
Player 2 is indifferent between IN and OUT, which implies



0 · x + − 12 · (1 − x) = −1 12 · x + 0. (1 − x)
 6. BEHAVIOUR STRATEGIES 31

Player 2
IN OU T
3 3
Player 1 In, S 0, 0 2, −2
1 1
Out, D 2, −2 0, 0

Figure 11. The normal form game corresponding to the exten-

sive form game in Figure 2, after eliminating Player 1’s strictly
dominated strategies.

which implies x = 41 . Similarly, y must be such that Player 1 is indifferent between

(In,S) and (Out,D). This implies that
0 · y + 1 21 · (1 − y) = 1
2 · y + 0 · (1 − y)
which implies y = 34 . So, the only Nash equilibrium of the game in Figure 11, and
hence of the game in Figure 10 is
 1 3

3 1

4 , 0, 0, 4 , 4 , 4 .
Exercise 2.2. Consider the normal form game in Figure 10 and the Nash
equilibrium strategy that we have just found.
(1) If the players play their equilibrium strategies, what is the expected payoff
to each player?
(2) If Player 1 plays his equilibrium strategy, what is the worst payoff that
he can get (whatever Player 2 does)?

6. Behaviour strategies
Consider the extensive form game shown in Figure 12 and its associated normal
form shown in Figure 13. In this game, we need three independent numbers to
describe a mixed strategy of Player 2, i.e., (x, y, z, 1 − x − y − z). Suppose that
instead Player 2 puts off her decision about which strategy to use until she is called
upon to move. In this case we only need two independent numbers to describe
the uncertainty about what Player 2 is going to do. That is, we could say that at
Player 2’s left-hand information set she would choose L with probability x and R
with probability 1 − x, and at her right-hand information set she would choose W
with probability y and E with probability 1 − y. We can see that by describing
Player 2’s strategies in this way, we can save ourselves some work. This efficiency
increases very quickly depending on the number of information sets and the number
of choices at each information set. For example, suppose that the game is similar
to that of Figure 12, except that Player 1’s choice is between four strategies, each
of which leads to a choice of Player 2 between three strategies. In this case Player 2
has 3·3·3·3 = 81 different pure strategies and we’d need 80 independent numbers to
describe the uncertainty about what Player 2 is going to do! If instead we suppose
that Player 2 puts off her decision about which strategy to use until she is called
upon to move, we only need 4 · 2 = 8 independent numbers. This is really great!
When we say that a player “puts off” his or her decision about which strategy
to use until he or she is called upon to move, what we really mean is that we
are using what is called a behaviour strategy to describe what the player is doing.
Formally, we have the following definition.
Definition 2.4. In a given extensive form game with player set N , a behaviour
strategy for player n ∈ N is a rule, or function, that assigns to each information set of
32 2. EXTENSIVE FORM GAMES

3, 3 0, 0 2, 2 1, 4

L R W E

2 2
T B

Figure 12. An extensive form game.

Player 2
LW LE RW RE
Player 1 T 3, 3 3, 3 0, 0 0, 0
B 2, 2 1, 4 2, 2 1, 4

Figure 13. The normal form game corresponding to the extensive

form game in Figure 12.

that player a probability distribution over the choices available at that information
set.

7. Kuhn’s theorem
Remember we motivated behaviour strategies in subsection 6 as a way of re-
ducing the amount of numbers we need compared to using mixed strategies. You
might be wondering whether we can always do this, that is, if we can represent
any arbitrary mixed strategy by a behaviour strategy. The answer is provided by
Kuhn’s theorem.
Theorem 2.1. (Kuhn) Given an extensive form game, a player n who has per-
fect recall in that game, and a mixed strategy σn of player n, there exists a behaviour
strategy bn of player n such that for any profile of strategies of the other players
(x1 , . . . , xn−1 , xn+1 , . . . , xN ) where xm , m 6= n, is either a mixed strategy of a be-
haviour strategy of player m, the strategy profiles (x1 , . . . , xn−1 , σn , xn+1 , . . . , xN )
and (x1 , . . . , xn−1 , bn , xn+1 , . . . , xN ) give the same distribution over terminal nodes.
In other words, Kuhn’s theorem says that given what the other players are
doing, we can get the same distribution over terminal nodes from σn and bn , as
long as player n has perfect recall. To see Kuhn’s theorem in action, consider
the game in Figure 14. Suppose Player 1 plays T with probability p and B with
probability 1 − p. Player 2’s pure strategies are XA, XB, Y A, and Y B. Suppose
Player 2 plays a mixed strategy of playing XA and Y B with probability 21 and XB
and Y A with probability 0. Thus, with probability 21 , Player 2 plays XA and we
get to terminal node x with probability p and to terminal node a with probability
1−p, and with probability 21 , Player 2 plays Y B and we get to terminal node y with
probability p and terminal node b with probability 1 − p. This gives a distribution
over terminal nodes as shown in the table in Figure 15.
 8. MORE THAN TWO PLAYERS 33

x y a b

X Y A B

2 2
T B

Figure 14. An extensive form game.

Terminal node Probability

x p/2
y p/2
a (1 − p) /2
b (1 − p) /2

Figure 15. Distributions over terminal nodes in the game of Fig-

ure 14 for the strategy profile given in the text.

Now suppose that Player 2 plays a behaviour strategy of playing X with prob-
ability 12 at his left-hand information set and A with probability 21 at his right-hand
information set. Thus we get to terminal node x with probability p · 12 , to terminal
node y with probability p · 21 , to terminal node a with probability (1 − p) · 12 and
to terminal node b with probability (1 − p) · 12 . Just as Kuhn’s theorem predicts,
there is a behaviour strategy that gives the same distribution over terminal nodes
as the mixed strategy.

8. More than two players

3, 3, 3 1, 1, 1

U D
6, 2, 7 3

L R
2, 7, 1 2

T B

Figure 16. An extensive form game with three players.

34 2. EXTENSIVE FORM GAMES

It is easy to draw extensive form games that have more than two players, such
as the one shown in Figure 16, which has three players. How would we find the
associated normal form of such a game? Recall that a normal form game is given
by (N, S, u). For the game in Figure 16, we have N = {1, 2, 3}, S1 = {T, B},
S2 = {L, R} and S3 = {U, D}, hence
S = {(T, L, U ) , (T, L, D) , (T, R, U ) , (T, R, D) , (B, L, U ) , (B, L, D) ,
(B, R, U ) , (B, R, D)}.
Finally, the payoff functions of each player are shown in Figure 17.

S T, L, U T, L, D T, R, U T, R, D B, L, U B, L, D B, R, U B, R, D
u1 2 2 2 2 6 6 3 1
u2 7 7 7 7 2 2 3 1
u3 1 1 1 1 7 7 3 1

Figure 17. One way of presenting the normal form game associ-
ated with the extensive form game of Figure 16.

The form shown in Figure 17 is a perfectly acceptable exposition of the asso-

ciated normal form. However, it’s more convenient to represent this three player
game as shown in Figure 18, where Player 3 chooses the matrix.

Player 3
U D
Player 2 Player 2
L R L R
Player 1 Player 1
T 2, 7, 1 2, 7, 1 T 2, 7, 1 2, 7, 1
B 6, 2, 7 3, 3, 3 B 6, 2, 7 1, 1, 1

Figure 18. A more convenient representation of the normal form

of the game in Figure 16.

9. Subgame perfect equilibrium

Subgame perfect equilibrium is an equilibrium concept that relates directly to
the extensive form of a game. The basic idea is that equilibrium strategies should
continue to be an equilibrium in each subgame (subgames will be defined formally
below).
Definition 2.5. A subgame perfect equilibrium is a profile of strategies such
that for every subgame the parts of the profile relevant to the subgame constitute
an equilibrium of the subgame.
To understand subgame perfect equilibrium, we first need to know what a
subgame is. Consider the game in Figure 16. This game has two proper subgames.
Strictly speaking, the whole game is a subgame of itself, so we call subgames that
are not the whole game proper subgames. The first is the game that begins from
Player 2’s decision node and the second is the game that begins from player 3’s
decision node. However, it would be a mistake to think that every game has proper
 9. SUBGAME PERFECT EQUILIBRIUM 35

subgames. Consider the extensive form game in Figure 19, which is the extensive
form version of matching pennies. In this game it doesn’t make sense to say there is
a subgame that begins at either of Player 2’s decision nodes. This is because both
of these nodes are in the same information set. If we said that a subgame started
at either one of them, then we would be saying that Player 2 knew which node she
is at, which violates the original structure of the game.

not a subgame
1, −1 −1, 1 −1, 1 1, −1

H T H T

2
h t

Figure 19. The extensive form version of matching pennies. This

game has no proper subgames.

As another example, consider the extensive form game shown in Figure 20.
In this game, the game starting from Player 2’s decision node is not a subgame,
because it splits player 3’s information set.

not a subgame
0, 0, 0 3, 2, 2 0, 0, 1 4, 4, 0

U D U D

3 1, 1, 1

L R

T B

Figure 20. The game beginning at Player 2’s decision node is not
a subgame because it splits Player 3’s information set.

Let us now give a formal definition of a subgame.

36 2. EXTENSIVE FORM GAMES

Definition 2.6. A subgame of an extensive form game Γ is some node in

the tree of Γ and all the nodes that follow it, with the original tree structure but
restricted to this subset of nodes, with the property that any information set of Γ
is either completely in the subgame or completely outside the subgame. The rest
of the structure is the same as in Γ but restricted to the new (smaller) tree.
Notice that under this definition, Γ is a subgame of Γ. We call subgames that
are not Γ itself proper subgames of Γ.
And we are now in a position to give a more formal definition of subgame
perfect equilibrium.
Definition 2.7. A subgame perfect equilibrium of an extensive form game Γ
with perfect recall is a profile of behaviour strategies (b1 , b2 , . . . , bN ) such that, for
every subgame Γ0 of Γ, (b01 , b02 , . . . , b0N ) is a Nash equilibrium of Γ0 where b0n is bn
restricted to Γ0 .
9.1. Finding subgame perfect equilibria. An example of the simplest type
of extensive form game in which subgame perfect equilibrium is interesting is shown
in Figure 21. Notice that in this game, (T, R) is a Nash equilibrium (of the whole
game). Given that Player 1 plays T , Player 2 is indifferent between L and R. And
given that Player 2 plays R, Player 1 prefers to play T . There is something strange
about this equilibrium, however. Surely, if Player 2 was actually called upon to
move, she would play L rather than R. The idea of subgame perfect equilibrium is
to get rid of this silly type of Nash equilibrium. In a subgame perfect equilibrium,
Player 2 should behave rationally if her information set is reached. Thus Player 2
playing R cannot form part of a subgame perfect equilibrium. Rather, if Player 2 is
called upon to move, the only rational thing for her to do is to play L. This means
that Player 1 will prefer B over T , and the only subgame perfect equilibrium of
this game is (B, L). Note that (B, L) is also a Nash equilibrium of the whole game.
This is true in general: Subgame perfect equilibria are always Nash equilibria, but
Nash equilibria are not necessarily subgame perfect.

2, 2 0, 0

L R
1, 1 2

T B

Figure 21. An example of the simplest type of game in which

subgame perfect equilibrium is interesting.

As another example, consider the extensive form game in Figure 22. In this
game, players 1 and 2 are playing a prisoners’ dilemma, while at the beginning of
the game player 3 gets to choose whether players 1 and 2 will actually play the
prisoners’ dilemma or whether the game will end. The only proper subgame of this
game begins at Player 1’s node. We claim that (C, C, L) is a Nash equilibrium of
the whole game. Given that player 3 is playing L, players 1 and 2 can do whatever
they like without affecting their payoffs and do not care what the other is playing.
 10. SEQUENTIAL EQUILIBRIUM 37

And given that players 1 and 2 are both playing C, player 3 does best by playing
L. However, although (C, C, L) is a Nash equilibrium of the game, it is not a
subgame perfect equilibrium. This is because (C, C) is not a Nash equilibrium of
the subgame beginning at Player 1’s decision node. Given that Player 1 is playing
C, Player 2 would be better off playing D, and similarly for Player 1. The only
subgame perfect equilibrium is (D, D, A).

9, 9, 0 0, 10, 10 10, 0, 10 1, 1, 10

C D C D

C D
10, 10, 1 1

L A

Figure 22. Player 3 chooses whether or not Players 1 and 2 will

play a prisoners’ dilemma game.

In the previous example, we have seen that a extensive form games can have
equilibria that are Nash but not subgame perfect. You might then be wondering
whether it’s possible to have an extensive form game that has no subgame perfect
equilibria. The answer is no. Selten in 1965 proved that every finite extensive form
game with perfect recall has at least one subgame perfect equilibrium.
9.2. Backwards induction. Backwards induction is a convenient way of find-
ing subgame perfect equilibria of extensive form games. We simply proceed back-
wards through the game tree, starting from the subgames that have no proper
subgames of themselves, and pick an equilibrium. We then replace the subgame by
a terminal node with payoff equal to the expected payoff in the equilibrium of the
subgame. Then repeat, as necessary.
As an illustration, consider the game in Figure 23. First Player 1 moves and
chooses whether she and Player 2 will play a battle of the sexes game or a coordi-
nation game. There are two proper subgames of this game, the battle of the sexes
subgame and the coordination subgame. Following the backwards induction pro-
cedure, one possible equilibrium of the battle of the sexes subgame is (F, F ), and
one possible equilibrium of the coordination subgame is (B, b). We then replace
each subgame by the appropriate expected payoffs, as shown in Figure 24. We can
then see that at the initial node, Player 1 will choose S. Thus one subgame perfect
equilibrium of this game is ((S, F, B) , (F, b)).

10. Sequential equilibrium

Consider the game shown in Figure 25. This game is known as “Selten’s Horse”.
In this game, there are no proper subgames, hence all Nash equilibria are subgame
perfect. Note that (T, R, D) is a Nash equilibrium (and hence subgame perfect).
But in this equilibrium, Player 2 isn’t really being rational, because if player 3 is
38 2. EXTENSIVE FORM GAMES

3, 1 0, 0 0, 0 1, 3 4, 4 0, 0 0, 0 1, 1

F B F B a b a b
2 2

F B A B
1 1

S C
1

Figure 23. Player 1 chooses battle of the sexes or coordination

game.

3, 1 1, 1

S C
1

Figure 24. Replacing the subgames with the equilibrium ex-

pected payoff in each subgame.

really playing D then if Player 2 actually got to move, she would be better off
playing L rather R. Thus we have another “silly” Nash equilibrium, and subgame
perfection is no help to us here to eliminate it. This kind of game lead to the
development of another equilibrium concept called sequential equilibrium.

0, 0, 0 3, 2, 2 0, 0, 1 4, 4, 0

U D U D

3 1, 1, 1

L R

T B

Figure 25. “Selten’s Horse”.

Exercise 2.3. Find all the Nash equilibria of the game in Figure 25.
A sequential equilibrium is a pair, (b, µ) where b is a profile of behaviour strate-
gies and µ is a system of beliefs. That is, µ is a profile of probability distributions,
one for each information set, over the nodes in the information set. The system of
 10. SEQUENTIAL EQUILIBRIUM 39

beliefs summarises the probabilities with which each player believes he is at each
of the nodes within each of his information sets.
To be a sequential equilibrium, the pair (b, µ) must satisfy two properties:
(1) At each information set, b puts positive weight only on those actions that
are optimal given b and µ. This is called sequential rationality.
(2) µ and b should be consistent. This means that if anything can be inferred
about µ from b then µ should be that.
If we drop the consistency requirement, we get another equilibrium concept
called Perfect Bayesian Equilibrium, which we won’t talk about in this course but
you might come across in other books.

−2, 2 2, −2 2, −2 −2, 2

L R L R
1, −1 −1, 1
2
OUT IN IN OUT

1 1
2 2

Nature

Figure 26. Illustrating consistent beliefs.

More formally, consistency requires that there is a sequence of completely mixed

behaviour strategy profiles bt → b and µt → µ with µt the conditional distribution
over the nodes induced by bt . To see how this works, consider the game in Fig-
ure 26. Each player has two nodes in their (only) information set. Under sequential
equilibrium we must specify the probabilities with which each player believes they
are at each node in each of their information sets. Furthermore, these beliefs must
be consistent with the equilibrium behaviour strategy. Suppose that in this game,
b = {(0, 1) , (1, 0)}, that is, that Player 1 plays OUT and Player 2 plays L. The
profile of beliefs is µ = ((µ (x) , µ (y)) , (µ (a) , µ (b))). Clearly, since the only move
before Player 1’s information set is that of Nature, Player 1’s beliefs must be that
µ (x) = 12 and µ (y) = 21 . These beliefs will be consistent with any behaviour
strategy, since whatever behaviour strategy is being played will not affect the prob-
ability that Player 1 is at either node in his information set. To find Player 2’s
beliefs µ (a) and µ (b) we need to find a completely mixed behaviour strategy close
to b. Actually, we only need to worry about Player 1’s behaviour strategy, since
there are no
moves of Player 2 before Player 2’s information set. Suppose we use
bt = 1t , t−1
t for Player 1. Then,
1 1 1
Pr (a) = · =
2 t 2t
and
1 1 1
Pr (b) = · = .
2 t 2t
40 2. EXTENSIVE FORM GAMES

Thus,
1
1
µt (a) = 1
2t
1 = .
2t + 2t
2
Similarly, we can show that µt (b) = 12 also.
Why couldn’t we just use Player 1’s behaviour strategy (0, 1)? Because in this
case Player 1 plays OUT and thus Player 2’s information set is not reached. Math-
ematically, we would have
1
·0
µ (a) = 1 2 1
2 ·0+ 2 ·0
which is not defined. So we must use completely mixed behaviour strategies close
to b to find the (consistent) beliefs, since if we use a completely mixed behaviour
strategy, every information set in the game will be reached with strictly positive
probability (even if the probability is very small).

11. Signaling games

Signaling games are a special type of extensive form game that arise in many
economic models. The basic structure of a signaling game is as follows:
(1) Nature moves first, and picks one “state of nature” or “type of Player 1”.
(2) Player 1 sees the result of Nature’s move, and makes his choice. It is as-
sumed that Player 1’s available actions do not depend on Nature’s choice.
(3) Player 2 sees Player 1’s choice but not Nature’s, and makes her choice.
(4) Payoffs are determined based on the choices of the two players and Nature,
and the game ends.
The extensive form of a simple signaling game is shown in Figure 27. In this
game, Player 1 is either type a or b with probability 12 . Player 1 can choose T or B.
Player 2 then observes Player 1’s choice but not Nature’s choice. If Player 1 chose
T then Player 2 chooses between L and R and if Player 1 chose B then Player 2
chooses between X, Y , and Z (this illustrates the fact that Player 2’s available
choices can depend on Player 1’s choice, but neither player’s available choices can
depend on Nature’s choice).

2, 2 3, 3 2, 2 2, 2 2, 2
1, 1 0, 0 1, 1 1, 1 1, 0
X Y Z L R
L R X Y Z

2 2
T B T B

1 1

1 1
2 2

Nature

Figure 27. The extensive form of a simple signaling game.

 11. SIGNALING GAMES 41

Typically, signaling games are not presented in the form of Figure 27, however.
Instead, a presentation due originally to Kreps and Wilson (1982) is used, in which
we suppress Nature’s move and instead have a number of initial nodes with associ-
ated probabilities. This alternative presentation of the game in Figure 27 is shown
in Figure 28. As you can see, it is much neater to present signaling games in this
way.

4, 3 L X 3, 2
T 1a B Y 0, 1
[ 12 ]
2, 2 R Z 0, 0
2 2
2, 0 L X 3, 0
T [ 12 ] B Y 0, 0
1b
1, 2 R Z 0, 1

Figure 28. An alternative, and tidier, presentation of the exten-

sive form signaling game in Figure 27.

One economic example of a signaling game is as follows. In some market there

is an incumbent firm that sets its price in the time period before a potential entrant
enters the market. The incumbent knows if the market demand conditions are good
or bad. The potential entrant does not know the demand conditions but does see
the price chosen by the incumbent. The entrant then decides whether to enter or
not. If the entrant enters, it finds out the market conditions and then there is
some competition in the second period. The fixed costs are such that the entrant
will make positive profits only if market conditions are good. A representation of
this game is shown in Figure 29, where it is assumed that the incumbent can only
choose a ‘high’ or ‘low’ price.

2, 1 E E 3, 1
L 1G H
[pG ]
9, 0 S S 10, 0
2 2
2, −1 E E 1, −1
L [pB ] H
1B
4, 0 S S 3, 0

Figure 29. A simple price signaling game.

A famous signaling game is shown in Figure 30. The story behind this game is
as follows. There are two players, Player A and Player B. At the beginning of the
game, Nature selects Player A to be either a wimp (with probability 0.1) or surly
(with probability 0.9). At the start of the game, Player A knows whether he is a
wimp or surly. Player A then has to choose what to have for breakfast. He has
42 2. EXTENSIVE FORM GAMES

two choices: beer or quiche.2 If Player A is surly, he prefers beer for breakfast and
if Player A is a wimp he prefers quiche for breakfast, everything else equal. That
is, if Player A has his preferred breakfast, his incremental payoff is 1, otherwise
0. After breakfast, Player A meets Player B. At the meeting, Player B observes
what Player A had for breakfast (perhaps by seeing the bits of quiche stuck in
Player A’s beard if Player A had quiche for breakfast or smelling the alcohol on
Player A’s breath if A had beer for breakfast), but Player B does not know whether
Player A is a wimp or surly. Having observed Player A’s breakfast, Player B must
then choose whether or not to duel (fight) with Player A. Then the game ends and
payoffs are decided. Regardless of whether Player A is a wimp or surly, Player A
dislikes fighting. Thus Player A’s incremental payoff is 2 if Player B chooses not to
duel and 0 if Player B chooses to duel. So, for example, if Player A is surly and has
his preferred breakfast of beer and then Player B chooses not to duel, Player A’s
payoff is 1 + 2 = 3. Or, if Player A is a wimp and has his less preferred breakfast
of beer and Player B chooses to duel, Player A’s payoff is 0 + 0 = 0. Player B, on
the other hand, only prefers to duel if Player A is a wimp. If Player A is surly and
Player B chooses to duel, Player B’s payoff is 0, while if Player B chooses not to
duel Player B’s payoff is 1. If Player A is a wimp and Player B chooses to duel,
Player B’s payoff is 2, while if Player B chooses not to duel his payoff is 1.3

don’t don’t
3, 1 duel duel 2, 1
beer AS quiche
[.9]
1, 0 duel duel 0, 0
don’t BB BQ don’t
2, 1 duel duel 3, 1
[.1]
beer AW quiche
0, 2 duel duel 1, 2

Figure 30. The beer quiche game of Cho and Kreps.

Exercise 2.4. Consider the signaling game in Figure 30.

(1) Find the normal form of this game.
(2) Find all the pure strategy Nash equilibria (from the normal form).
(3) Find all the mixed strategy Nash equilibria (from the normal form).
Since the beer-quiche game in Figure 30 is a game of asymmetric information
(that is, Player B doesn’t know whether Player A is surly or a wimp), it makes sense
to look for a sequential equilibrium. In this game a sequential equilibrium will be
an assessment ((bA , bB ) , µ) that is consistent and sequentially rational, where
bA = ((bAS (beer) , bAS (quiche)) , (bAW (beer) , bAW (quiche)))
bB = ((bBB (don’t) , bBB (duel)) , bBQ (don’t) , bBQ (duel))
µ = ((µbeer (S) , µbeer (W )) , (µquiche (S) , µquiche (W ))) .
2For those of you who don’t know, quiche is a sort-of egg pie type thing that only wimps and
Australian rugby players eat.
3Note that Player B’s payoff only depends on his own decision about whether to duel or not,
and whether Player A is surly or a wimp. That is, Player B’s payoff does not depend on what
Player A had for breakfast.
 11. SIGNALING GAMES 43

In such a game as this, there are two basic types of equilibrium that can arise. In
the first type of equilibrium, called a pooling equilibrium, both types of Player A
choose the same probabilities for having beer and quiche for breakfast. In this case,
Player A’s breakfast gives Player B no information about Player A’s type and thus
Player B’s best estimate of Player A’s type is that Player A is surly with probability
0.9 and a wimp with probability 0.1. In the second type of equilibrium, called a
separating equilibrium, the two types of Player A choose different probabilities for
the breakfasts. In this case, Player B gets some information about Player A’s type
from Player A’s breakfast, and will be able to modify his estimate of Player A’s
type.
You will analyse this game more fully in tutorials. So as not to spoil the fun,
we shall content ourselves with just answering one question about this game here:
Are there any equilibria in which both beer and quiche are played with positive
probability? Before answering this question, note the following facts about the
game:
(1) If bB is such that the surly Player A is willing to have quiche for breakfast
then the wimp Player A strictly wants to have quiche for breakfast (since
quiche is the wimp’s preferred breakfast).
(2) Similarly, if bB is such that the wimp Player A wants to have beer for
breakfast then the surly Player A strictly wants to have beer for breakfast.
Now suppose that in equilibrium the surly pPlayer A plays quiche with positive
probability. Then by fact (1), the wimp Player A plays quiche with probability 1.
Thus either surly Player A plays quiche with probability 1, in which case beer is
not played with positive probability, or surly Player A plays beer with positive
probability. But then µbeer (S) = 1 and thus Player B chooses not to duel after
observing Player A having beer for breakfast, that is, bBB (duel) = 0. But then
surly Player A will not want to play quiche with positive probability.
Suppose, on the other hand, that surly Player A plays beer with probability
1, i.e., bAS (beer) = 1. Then µbeer (S) ≥ 0.9. Then sequential rationality implies
bBB ( duel) = 0. So if wimp Player A plays quiche with positive probability then
µquiche (S) = 0 and µquiche (W ) = 1 and so bBQ (duel) = 1. Thus wimp Player A
chooses beer for sure and quiche is not played with positive probability.
So, we have established that there are no sequential equilibria in which both
beer and quiche are played with strictly positive probability. As you will discover in
tutorials, there are sequential equilibria in which both types of Player A play beer
and there are sequential equilibria in which both types of Player A play quiche.