Extensive-Form Games with Perfect Information

Jeffrey Ely

April 22, 2015

Jeffrey Ely Extensive-Form Games with Perfect Information

Extensive-Form Games With Perfect Information

Recall the alternating offers bargaining game.

It is an extensive-form game. (Described in terms of sequences of
moves.)
In this game, each time a player moves, he knows everything that has
happened in the past.
He can compute exactly how the game will end given any profile of
continuation strategies
Such a game is said to have perfect information.

Jeffrey Ely Extensive-Form Games with Perfect Information

Simple Extensive-Form Game With Perfect Information

1b
  HH
L HR
2  HH 2
r Hr
A @ B C @ D
@ @
r @r r @r
1, 0 2, 3 4, 1 −1, 0
Figure: An extensive game

Jeffrey Ely Extensive-Form Games with Perfect Information

Game Trees

Nodes H
Initial node.
Terminal Nodes Z ⊂ H.
Player correspondence N : H ⇒ {1, . . . , n } (N (h ) indicates who
moves at h. They move simultaneously)
Actions Ai (h ) ̸= ∅ at each h where i ∈ N (h ).
▶ Action profiles can be identified with branches in the tree.
Payoffs ui : Z → R.

Jeffrey Ely Extensive-Form Games with Perfect Information

Strategies

A strategy for player i is a function σi specifying σi (h ) ∈ Ai (h) for each h

such that i ∈ N (h ).

Jeffrey Ely Extensive-Form Games with Perfect Information

Strategic Form

Given any profile of strategies σ, there is a unique terminal node z (σ )

that will be reached.
We define ui (σ) = ui (z (σ )).
We have thus obtained a strategic-form representation of the game.
We can apply strategic-form solution concepts.

Jeffrey Ely Extensive-Form Games with Perfect Information

Example

Rationalizability:
▶ The strategy AD is never a best-reply for player 2.
▶ All other strategies are rationalizable (I think.)
Nash equilibrium:
▶ The profile (L, BD ) is a Nash equilibrium.
▶ The profile (R, BC ) is a Nash equilibrium.
▶ The profile (R, AC ) is a Nash equilibrium.
▶ There are many mixed equilibria.
Iterative elimination of weakly dominated strategies:
▶ Any strategy for 2 that plays either A or D is weakly dominated.
▶ Therefore BC is the only weakly un-dominated strategy for 2.
▶ Next we can eliminate L for 1.
▶ The order of elimination doesn’t matter in this example.

Jeffrey Ely Extensive-Form Games with Perfect Information

Subgames

Associated with every non-terminal node h, there is a subgame consisting

of all the nodes in the tree that follows h.
The game as a whole is a subgame
All other subgames are called proper subgames
Given a strategy profile σ
▶ Denote by σ | the continuation strategy profile in the subgame
h
beginning at h.
▶ Denote by z (σ |h ) the terminal node reached by σ beginning from h.
▶ Denote by ui (σ |h ), the continuation payoff

ui (σ|h ) = ui (z (σ |h ))

Jeffrey Ely Extensive-Form Games with Perfect Information

Subgame Perfect Nash Equilibrium

A strategy profile σ is a subgame perfect Nash equilibrium if for every

non-terminal node h, the continuation strategy profile σ |h is a Nash
equilibrium of the subgame that begins at h, i.e.

ui (σ |h) ≥ ui ((σi′ , σ−i )|h )

for every strategy σi′ .

Jeffrey Ely Extensive-Form Games with Perfect Information

Example

The unique subgame-perfect Nash equilibrium in the example is (R, BC ).

Jeffrey Ely Extensive-Form Games with Perfect Information

Backward Induction

Consider a perfect information game that

1 is finite,
2 has a single player moving at each node (i.e. N (h ) is a singleton),
3 and has no indifferences

z ̸= z ′ =⇒ ui (z ) ̸= ui (z ′ ))

Such a game has a unique SPE which can be found by backward induction.
(Note that we do not need mixed strategies for existence.)

Jeffrey Ely Extensive-Form Games with Perfect Information

The Dukes of Earl

Example
The five Dukes of Earl are scheduled to arrive at the royal palace on each
of the first five days of May. Duke One is scheduled to arrive on the first
day of May, Duke Two on the second, etc. Each Duke, upon arrival, can
either kill the king or support the king. If he kills the king, he becomes the
new king and awaits the next Duke’s arrival. If he supports the king all
subsequent Duke’s cancel their visits. A Duke’s first priority is to remain
alive, and his second priority is to become king. Who is king on May 6?

Draw Extensive Form

Describe the strategies of each player
Identify the subgames
Apply Backward Induction.

Jeffrey Ely Extensive-Form Games with Perfect Information

The Ultimatum Game

There is a dollar to be divided between two players.

Player 1 moves first and offers a split of the dollar giving x to Player
2 and leaving 1 − x for himself.
Player 2 then either accepts or rejects the offer.
An accepted offer is implemented, but if the offer is rejected both
players get zero.
In the unique SPE of this game, Player 1 gets all of the bargaining surplus:

He offers x = 0 to Player 2 which is Player 2’s outside option.

He thus keeps the difference between the total surplus from
agreement and the total surplus from disagreement.
Issue with continuous action space.

Jeffrey Ely Extensive-Form Games with Perfect Information

Infinite Games?

(Infinite in terms of the tree length.)

Obviously there is no backward induction procedure.
We can think of backward induction in a different way.
The backward induction procedure constructs a strategy for each
player that is unimprovable by a one-stage deviation.

Jeffrey Ely Extensive-Form Games with Perfect Information

One-Stage Deviations

Definition
Let σi and σi′ be two distinct strategies for player i and let h be a node at
which i moves. Let h (σi , σi′ ) denote the strategy that coincides with σi at
all nodes except for h where it plays according to σi′ .
{
′ σi′ (h̃ ) if h̃ = h
h (σi , σi )(h̃ ) =
σi (h̃ ) otherwise

Jeffrey Ely Extensive-Form Games with Perfect Information

Unimprovable Strategies

Fix a strategy profile σ−i . Strategy σi is unimprovable by a one-stage

deviation if for every node h at which i moves, and every alternative
strategy σi′ ,
ui ((h (σi , σi′ ), σ−i )|h) ≤ ui (σ |h )

Jeffrey Ely Extensive-Form Games with Perfect Information

The One-Stage Deviation Principle

Proposition
In any finite perfect-information game, if a strategy profile is unimprovable
for all players then it is a SPE.

Jeffrey Ely Extensive-Form Games with Perfect Information

The Marathon Game

1 player game.
The player’s name is Zeno.
H \ Z = {1, 2 . . .}
A(h ) = {quit, continue} for all h.
▶ Quitting at h gives payoff −(1 − 1/h ).
▶ Continuing at h leads to h + 1.

Continuing forever leads to the terminal node ∞ which gives payoff 1.

h = 1 is the initial node.

Jeffrey Ely Extensive-Form Games with Perfect Information

SPE in the Marathon Game

There is a unique SPE in which Zeno completes the Marathon.

Jeffrey Ely Extensive-Form Games with Perfect Information

Unimprovable Strategies

The strategy which quits at every node is unimprovable.

Jeffrey Ely Extensive-Form Games with Perfect Information

Continuity at Infinity

Definition
A perfect information game is continuous at infinity if for every strategy
profile σ−i , player i, history h, pair of strategies σi , σi′ , and real number
ε > 0 there is an integer t such that the strategy σ̃ defined by
{
σi (h̃ ) if h̃ is no more than t moves past h
σ̃i (h̃) =
σi′ (h̃ ) otherwise

earns a continuation payoff within ε of σi in the subgame beginning at h,

i.e.
ui ((σ̃i , σ−i )|h ) > ui (σi , σ−i |h ) − ε

Jeffrey Ely Extensive-Form Games with Perfect Information

Conditions for Continuity at Infinity

A finite game is continuous at infinity

A game with discounting is continuous at infinity

Jeffrey Ely Extensive-Form Games with Perfect Information

The One-Stage Deviation Principle

Proposition
In any perfect information game that is continuous at infinity, a strategy
profile that is unimprovable for all players is a subgame perfect equilibrium.

(Of course the converse holds as well.)

Jeffrey Ely Extensive-Form Games with Perfect Information

Proof of OSDP
We will show that any strategy profile which is not an SPE is improvable
by a one-stage deviation for some player.
Suppose σ is a strategy profile such that at some subgame h, there is a
player i whose continuation strategy is not a best reply, i.e.
ui ((σi′ , σ−i )|h) > ui (σ|h )
for some σi′ . For any integer t > 0 define the following strategy
{
σi′ (h̃ ) if h̃ is no more than t moves past h
σ̃i (h̃ ) =
t
σi (h̃) otherwise.
By continuity at infinity, for any ε there is a t such that
ui ((σ̃it , σ−i )|h ) > ui ((σi′ , σ−i )|h ) − ε.
We take ε small enough so that
ui ((σ̃it , σ−i )|h) > ui (σ |h )
Jeffrey Ely Extensive-Form Games with Perfect Information
Proof of OSDP

Is there any node h̃ which is exactly t moves past h such that

ui ((h̃ (σi , σi′ ), σ−i )|h̃) > ui (σ |h̃ )

holds? If yes, then we have found an improvement by a one-stage

deviation. If not, then we have

ui ((σ̃it −1 , σ−i )|h ) ≥ ui ((σ̃it , σ−i )|h ) > ui (σ |h)

(to understand the first inequality, note that the strategy profile
(σ̃it −1 , σ−i ) either reaches a node h̃ which is t moves past h in which case
the inequality holds by the previous answer, or it terminates before ever
reaching such a node in which case the payoffs are equal.)

Jeffrey Ely Extensive-Form Games with Perfect Information

Proof of OSDP

Continuing backward in this way, we either find a node h̃ which is within t

moves of h such that

ui ((h̃ (σi , σi′ ), σ−i )|h̃) > ui (σ |h̃ )

i.e. we find an improvement by a one-stage deviation, or if we never do,

we obtain

ui ((σ̃i0 , σ−i )|h ) ≥ ui ((σ̃i1 , σ−i )|h) ≥ . . . > ui (σ |h)

but σ̃i0 is a one-stage deviation from σi and so in this case we are done as
well.

Jeffrey Ely Extensive-Form Games with Perfect Information

The Rubinstein Bargaining Game

Two players take turns proposing a division of a unit surplus..

Player 1 is the first to make a proposal.
Player 2 accepts or rejects.
When an offer is accepted the game ends.
When an offer is rejected, the game continues and rejecting player
makes a proposal.
If the tth proposal gives x to player 1 and 1 − x to player 2
(0 ≤ x ≤ 1) and it is accepted, the payoffs are δt −1 x for player 1 and
δt −1 (1 − x ) for player 2, where δ ∈ (0, 1) is the discount factor.

Jeffrey Ely Extensive-Form Games with Perfect Information

Continuity at Infinity

The discounting in the payoffs ensures that the game is continuous at

infinity so that we can apply the OSDP

Jeffrey Ely Extensive-Form Games with Perfect Information

SPE of the Rubinstein game

The following constitutes a SPE of the Rubinstein bargaining game

δ
The proposer offers to give R = 1+ δ to the opponent and keep
(1 − R ) for himself.
The responder accepts all offers that give at least R and rejects all
offers that give less.
(Write this out as a properly specified extensive-form game strategy.)
Note that R = δ(1 − R ). Thus, in parallel with the ultimatum game, the
offering player is extracting the full surplus associated with agreement
today rather than agreement tomorrow.

Jeffrey Ely Extensive-Form Games with Perfect Information

We show that this strategy profile is unimprovable.
Suppose that a responder is offered y < R in some period t. The
strategy profile dictates that he should reject. If he does so, then in
the next period there will be agreement and he will get δ(1 − R ). He
would not deviate and accept y because

y < R = δ (1 − R )

Suppose that a responder is offered y which is at least R. He would

not deviate and reject y due to the same equality.
If the proposer follows the strategy, he gets a payoff of (1 − R ).
Consider a deviation by the proposer.
▶ An offer more than R will be accepted and this is worse than offering R
which will also be accepted.
▶ If he offers less than R, it will be rejected and in the next period an
agreement will be reach giving him δR. Since R = δ(1 − R ) we have
δR = δ2 (1 − R ) < (1 − R )so this is not an improvement either.

Jeffrey Ely Extensive-Form Games with Perfect Information