Advanced Microeconomics

HKU Business School

Masaki Miyashita

Chapter 4:
Games with Compete Information
 What is Game Theory?

• Decision Theory: a single decision maker

• Game Theory: multiple decision makers

• Rationality: As in decision theory, each person (“player”) acts to maximize

one’s objective, captured by preferences or utility functions.

• Interdependent Preferences: A player’s preference or utility depends not

only one’s choice but also on others’ choices.

• Strategic reasoning is important to play the best strategy.

 What is Game Theory?

• Two strands: non-cooperative or cooperative game theory

• Non-cooperative (“Strategic”) Games: each player acts independently

to achieve one’s own objective.

• Cooperative Games: binding contracts between players are possible. Study

axiomatic characterization of solutions with desirable properties.

• This course: non-cooperative game theory.

• Goal: Make predications of the strategic interactions among rational players.

• Many applications: biology, political science, economics, management

sciences, operations research, and yes, actual games, occasionally.
 Example 1: Partnership

• Two partners simultaneously and independently choose between:

effort (E) or shirk (S).

• The output is

x = a × (# of partenrs excerting E), where a > 0.

• The output is equally shared among partners.

• Cost of effort is c > 0, borne by each partner individually.

• Suppose a > c > a/2.

 Example 2: Cournot Duopoly

• Two rms competing in an industry.

• Each rm i ∈ {1, 2} chooses output qi ≥ 0, incurring cost cqi .

• c > 0 represents marginal cost.

• The market clearing price P is determined by a downward sloping demand

P = max {a − b(q1 + q2), 0} .

• Suppose a > c and b > 0.

fi
fi
 Example 3: Second-price Auction

• n bidders compete for one unit of an indivisible good.

• Each bidder has a valuation vi ≥ 0 for the good, which is unknown to the others.

• They simultaneously submit bids bi ≥ 0.

• The highest bidder wins the object and pays the second highest bit.

• (Analogously, we can also consider rst-price or third-price auctions etc.)

fi
Normal-form Game
 Normal-form Game

• The three basic ingredients:

• N = {1,…, n} is a nite set of players.

• For each player i ∈ N, there is a non-empty set Ai of actions (or strategies).

• For each player i ∈ N, a payoff function

ui : A1 × ⋯ × An → ℝ .

• A normal-form game is summarized as the pro le

G = ⟨N, (Ai, ui)i∈N⟩ .

fi
fi
 Normal-form Game
Notation.
A := A1 × ⋯ × An .

A−i := A1 × ⋯ × Ai−1 × Ai+1 × ⋯ × An .

a := (a1, …, an) .

a−i := (a1, …, ai−1, ai+1, …, an) .

(ai, a−i) := (a1, …, ai−1, ai, ai+1, …, an) .

• An n-tuple a = (a1, …, an) ∈ A is called a strategy pro le.

• Each player i has a utility function, de ned over the set of strategy pro les A .

• Alternatively, we can consider a preference relation ≿i over A .

fi
fi
fi
 Normal-form Game

• Timeline:

• Each player i simultaneously chooses action ai from Ai without knowing

others’ action choices.

• Payoff ui(a1, …, an) materializes to each player. The game ends right away.
 Example 1: Partnership
N = {1,2}

A1 = A2 = {E, S}

ui(E, E) = a − c, ui(E, S) = a/2 − c, ui(S, E) = a/2, ui(S, S) = 0

• When # of players/actions are small, a payoff matrix is useful to describe the

normal-form game.

Player 1 / Player 2 Effort Shirk

Effort a-c , a-c a/2-c , a/2

Shirk a/2 , a/2-c 0,0
 Example 2: Cournot Duopoly

N = {1,2}

A1 = A2 = ℝ+

ui(qi, qj) = (P − c)qi where P = max{a − b(qi + qj), 0}

• Note that P ≥ 0 if and only if (qi + qj) ≤ a/b .

• If (qi + qj) > a/b, then ui(qi, qj) = − cqi .

• If (qi + qj) ≤ a/b, then

ui(qi, qj) = (a − b(qi + qj) − c) ⋅ qi .

 Example 3: Second-price Auction

N = {1,…, n}

A1 = ⋯ = An = ℝ+
vi − maxj≠i bj if bi > maxj≠i bj
{0
ui(bi, b−i) =
if bi < maxj≠i bj

• (For simplicity we ignore tie cases.)

• In this game, reporting own valuation truthfully —- i.e., bi = vi —- turns out to be

optimal, no matter how other bidders behave.
 Comments

• A normal-form game rests on several key assumptions.

• Static & Simultaneous Choice: players choose their actions only once and do
so without knowing the choices of other players.

• Many strategic situations are inherently dynamic, where players make

decisions at multiple (and possibly different) points in time.

• Extensive-form games are used to study such dynamic interactions.

 Comments

• Complete Information: Each player has full knowledge of the game’s details,
including one’s own action set Ai and payoff function ui, as well the action sets Aj

and payoff functions uj of other players.

• Moreover, each player knows the fact that the other players also fully
understand these details.

• This assumption will be relaxed in incomplete information games, where

players may lack full knowledge about certain aspects of the game.
 Comments

• Common Knowledge of Rationality: Each player behaves rationally, aiming

to maximize their own payoff, and this is common knowledge among all players.

• 1st order: All agents are rational.

• 2nd order: All agents know “all agents are rational.”

• 3rd order: All agents know “all agents know “all agents are rational.””

⋮
 Comments

Muddy Children Puzzle.

• A group of n children is told that “some” of them have muddy faces.

• Each child can see the faces of others but cannot tell if one’s own face is muddy.

• The children are told that those with muddy faces must step forward (but anyone
with a clean face should not step forward).

• Suppose there are k children with muddy faces. Can they eventually gure out
that they are among those with muddy faces?

fi
 Equilibrium

• The outcome of a game is predicted in terms of an equilibrium.

• An action pro le forms an equilibrium when each player’s choice is, in some
sense, “optimal.”

• However, we must clarify what is meant by “optimal,” as each player’s payoff

depends not only on their own action but also on the actions of others.

• There are several different equilibrium concepts.

• Two of the most important are:

Dominant Strategy Equilibrium and Nash equilibrium.

fi
Dominant Strategy Equilibrium
 Dominant Strategy Equilibrium

• One action ai dominates another action bi when:

the player i obtains a higher payoff by choosing ai than bi

no matter what other players do.

• In such cases, there is no reason to choose bi .

• In contrast, if there exists an action that dominates all other actions for a player,
that action should be choose de nitely.
fi
 Dominant Strategy Equilibrium

De nition.

• Let ai and bi be actions of agent i .

• We say ai strictly dominates bi if ui(ai, a−i) > ui(bi, a−i) for all a−i ∈ A−i .

• Moreover, an action ai is called strictly dominant if it strictly dominates

all other actions bi ∈ Ai∖{ai} .
fi
 Dominant Strategy Equilibrium

De nition.

• Let ai and bi be actions of agent i .

• We say ai weakly dominates bi if ui(ai, a−i) ≥ ui(bi, a−i) for all a−i ∈ A−i
and ui(ai, a′−i) > ui(bi, a′−i) for some a′−i ∈ A−i .

• Moreover, an action ai is called weakly dominant if it weakly dominates

all other actions bi ∈ Ai∖{ai} .



fi
 Dominant Strategy Equilibrium

• Consider a normal-form game G = ⟨N, (Ai, ui)i∈N⟩ .

• Dominant Strategy Equilibrium (DSE) refers to a situation where every player

chooses their dominant strategy (whether in the weak or strict sense).

De nition.
• A strategy pro le a* = (a*
1
, …, a*
n ) ∈ A is a strict DSE in G if a*
i
is strictly

dominant for every i ∈ N . Similarly, a* is a weak DSE in G if a*

i
is weakly

dominant for every i ∈ N .

fi
fi
 Dominant Strategy Equilibrium

• A few observations about DSE:

1. In general, neither weak nor strict DSE need exist.

2. If a strict or weak DSE exists, it must be unique.

 Example 1: Partnership

• Recall the payoff matrix:

Player 1 / Player 2 Effort Shirk

Effort a-c , a-c a/2-c , a/2

Shirk a/2 , a/2-c 0,0

Claim. Assume a > c > a/2.

Then, (a*, a*) = (S, S) is a strict DSE.

1 2

• The equilibrium payoffs are 0 for both agents.

• In contrast, both get strictly better off if they choose move to (E, E) .

• This is a typical instance of coordination failure.

 Example 2: Cournot Duopoly
• Each rm’s pro t is given as follows:

• If (qi + qj) > a/b, then ui(qi, qj) = − cqi .

• If (qi + qj) ≤ a/b, then

ui(qi, qj) = (a − b(qi + qj) − c) ⋅ qi .

• Taking opponent’s choice qj as given, the rm i ’s best-response strategy

q*
i
(qj) is derived as the output level that maximizes ui( ⋅ , qj) .

Claim. The rm i’s best-response strategy is given as follows:

a−c qj
− if qj ≤ (a − c)/b,
q*
i
(qj) = 2b 2
0 if qj > (a − c)/b .
fi
fi
fi
fi
 Example 3: Second-price Auction

Claim. In 2PA, truth-telling is the unique weak dominant strategy.

• Fix any b−i ∈ A−i and let x ≡ maxj≠i bj denote the strongest opponent’s bid.

1. Any downward deviation bi < vi cannot be pro table.

• If x < bi or vi < x, then the payoff remains unchanged between bi and vi .
• In contrast, if bi < x < vi, the player receives vi − x > 0 by bidding vi, while
the payoff would be 0 with bi .

2. Any upward deviation bi > vi cannot be pro table.

• If x < vi or x < bi, then the payoff remains unchanged between bi and b*
i
.
• In contrast, if vi < x < bi, the player receives 0 by bidding vi, while the payoff
wold be negative vi − bi < 0 with bi .
fi
fi
 Dominant Strategy Equilibrium

• Partnership Game: (S, S) is the unique strict DSE.

• Cournot Duopoly: any output choice qi ∈ [0, (a − c)/2b] can be a unique best
response against some qj ≥ 0. So, there exists no weak DSE.

• Second-price Auction: truth-telling is the unique weak DSE.

• The dominance principle can provide strong predictions when a game has a DSE.

• However, it may be too stringent for some games; for instance, in the Cournot
duopoly, no strategy pro le survives the dominance criterion.
fi
Nash Equilibrium
 Nash Equilibrium

• The most commonly used equilibrium concept in game theory is Nash equilibrium.

• Captures a steady state of the play in a game, where each player holds the
correct belief about others’ behavior and acts rationally.

De nition.

• A strategy pro le a* = (a*

1
, …, a*
n ) is a Nash equilibrium in normal-form

game G = ⟨N, (Ai, ui)i∈N⟩ if

ui(a*
i
, a*
−i
) ≥ ui(ai, a*
−i
), ∀i ∈ N, ∀ai ∈ Ai .
fi
fi
 Example 2: Cournot Duopoly
• Each rm’s best-response strategy was:

a−c qj
− if qj ≤ (a − c)/b,
q*
i
(qj) = 2b 2
0 if qj > (a − c)/b .

• A pro le (q*, q*) constitutes a Nash equilibrium when (i) q*

1 2 1
best responds to q*
2

and (ii) q*
2
best responds to q*
1
.

• The system of equations

a − c q2 a − c q1
q1 = − and q2 = −
2b 2 2b 2
has a unique solution

( 3 3 )
a−c a−c
(q*, q*) =
1 2
, .

• This is the NE in the Cournot duopoly.

fi
fi
 Nash Equilibrium

• A few observations about Nash equilibrium (NE):

1. NE need not be unique; multiple Nash equilibria can exist in a game.

2. NE is weaker than DSE in the sense that if a strategy pro le forms a Nash
equilibrium, then it is a weak DSE (and thus, a strict DSE as well).

3. In general, a “pure strategy” NE may not exist; however, as we will discuss

soon, the existence of NE is guaranteed when the de nition is extended to
mixed strategies.
fi
fi
 Nash Equilibrium

• Each ai available in Ai is called a pure strategy.

• A player i’s mixed strategy is given as any probability distribution σi ∈ Δ(Ai) .

• In other words, σi is a lottery over pure strategies, where σi(ai) represents the
probability that player i chooses ai according to σi .

• How to interpret mixed strategies?

• Naive interpretation: Each player’s actual action randomization

• Alternative interpretation: Mixed strategies represent population frequencies,

with different individuals choosing various pure strategies.
 Nash Equilibrium

Notation.

• Σi := Δ(Ai) the set of player i’s mixed strategies.

• Σ := Σ1 × ⋯ × Σn the set of mixed strategy pro les.

• For a mixed strategy pro le σ = (σ1, …, σn) ∈ Σ, we extend each player’s payoff
functions by considering expected utility:

∑
ui(σ) = σ(a)ui(a) where σ(a) = σ1(a1) × ⋯ × σ(an) .
a∈A

• Behavioral assumption: Each player’s preferences over lotteries satisfy the

vNM axioms.
fi
fi
 Nash Equilibrium

• Since pure strategies are a special type of mixed strategies, the next de nition
generalizes the previous one.

De nition.

• A strategy pro le σ* = (σ*

1
, …, σ*
n ) ∈ Σ is a Nash equilibrium in in

normal-form game G = ⟨N, (Ai, ui)i∈N⟩ if

ui(σ*
i
, σ*
−i
) ≥ ui(σi, σ*
−i
), ∀i ∈ N, ∀σi ∈ Σi .

• Equivalently, σ* is a NE if

ui(σ* , σ* ) ≥ ui(ai, σ* ), ∀i ∈ N, ∀ai ∈ Ai . (Why?)

i −i −i
fi
fi
fi
 Indifference Principle

• By de nition, each player’s payoff function is af ne.

• In other words, for any σi, σ′i ∈ Σi and λ ∈ [0, 1], it holds that

ui (λσi + (1 − λ)σ′i , σ−i) = λui (σi, σ−i) + (1 − λ)ui (σ′i , σ−i) .

• This fact implies the following indifference principle: For σ* = (σ*

1
, …, σ*
n ) to

form a NE, each player must be indifferent among all pure strategies assigned
with positive probabilities.



fi
fi
 Indifference Principle

• The support of a mixed strategy σi is de ned by

supp(σi) := {ai ∈ Ai : σi(ai) > 0} .

Lemma (Indifference Principle).

• A strategy pro le σ* ∈ Σ forms a Nash equilibrium if and only if

ui(σ*
i
, σ*
−i
) = ui(ai, σ*
−i
), ∀ai ∈ supp(σ*
i
),

ui(σ*
i
, σ*
−i
) ≥ ui(ai, σ*
−i
), ∀ai ∉ supp(σ*
i
).
fi
fi
 Indifference Principle

Example (Coordination Game).

Player 1 / Player 2 Movie Hiking

Movie 5,2 0,0

Hiking 0,0 3,4

• Derive all Nash equilibria (including mixed ones) in this game.

 Existence of Nash Equilibrium

• Now we arrive at one of the most famous results in game theory.

• A norma-form game is said to be nite when the set of players N and players’
action sets A1, …, An are all nite.

Theorem 1 (Nash 1950).

• Every nite normal-form game has at least one Nash equilibrium.

• The proof is based on Fixed Point Theorem; see Appendix for details.

Proof
fi
fi
fi
 Existence of Nash Equilibrium

• The theorem guarantees the existence of NE for any nite game, but it does not
address the issue of uniqueness.

• Also, the theorem does not provide any ef cient way to calculate NE.

• The calculation can be dif cult even in simple games.

• For nite games, we can utilize LP algorithms; however, the number of

iterations grows exponentially with the number of actions.

• No general method for solving games with continuous actions even with two
players, while some classes of games are known to have ef cient computation
algorithms.
fi
fi
fi
fi
fi
 Discussion: Nash Equilibrium

• How do we interpret NE?

• Steady state: NE captures a stable situation, where each player has no

incentives for unilateral deviation.

• Bayesian rationality: NE captures the situation, where each player trie to

maximize their subjective expected utility, based on their belief about the
behavior of opponents.

• See Chapter 3.2 of Course of Game Theory by Osborne and Rubinstein (1995)
for a comprehensive discussion.
 NE as Steady State

• Consider the following best-response dynamics:

• Player 1 chooses an initial strategy σ1 .

• Player 2 chooses the strategy σ′2 that best responds to σ1 .

• Player 1 chooses the strategy σ′1′ that best responds to σ′2 .

• Player 2 chooses the strategy σ′2′′ that best responds to σ′1′ .

• NE is the steady state of this best-response dynamics.

• I.e., once the players reach NE, the dynamics cease to evolve.

• However, the dynamics do not necessarily convergent.










 NE as Steady State

Example (Matching Penny).

Player 1 / Player 2 Head Tail

Head 1,0 0,1

Tail 0,1 1,0
 NE as Bayesian Rationality

• Each player faces strategic uncertainty regarding the opponents’ behavior.

• Set Si := A−i and call it player i’s personal state space.

• A subjective belief is a probability distribution μi ∈ Δ(Si) .

• Each action ai is literally interpreted as an “act” that determines the payoff

relevant outcome (ai, a−i) contingently on the realized “state” a−i ∈ Si .

• A Bayesian rational player chooses a (possibly randomized) action to maximize

the following expected payoff:

∑
μi(a−i)u(σi, a−i) .
a−i∈Si

Belief about
strategic uncertainty
 NE as Bayesian Rationality

• In NE σ* = (σ*
1
, …, σ*
n) …

• Each player forms the subjective belief μ*

i
that is consistent with the actual
behavior of their opponents:

μ*
i
(a−i) = σ*(a ) × ⋯ × σ*
1 1
(a ) × σ*
i−1 i−1
(a ) × ⋯ × σ*
i+1 i+1 n (an) .

• In response to μ*
i
, each player chooses the optimal strategy σ*
i
that
maximizes the subjective expected utility:

∑
σ*
i
maximizes μ*
i
(a−i)u(σi, ai) .
a−i∈Si

• The strong assumption: Each player’s belief must coincide perfectly with the
“actual” action distribution, resulting from the play of opponents.

• NE would not be reliable when players may “misspecify” their beliefs.

 DSE as Bayesian Rationality

• DSE can also be interpreted in a belief-based way.

• It can be easily seen that: If a* = (a*

1
, …, a*
n ) is a (weak) DSE, then for every

belief μi ∈ Δ(A−i), we have

∑
a*
i
∈ arg max μi(a−i)ui(σi, a−i) .
σi∈Σi
a−i∈Si

• I.e., a*
i
is optimal not only for the equilibrium belief but also for all possible beliefs
(akin to Bewley’s representation under Knightian uncertainty).

• Robustness against belief misspeci cation

fi
 Example 4: O’Neill Game

• Each of two players i ∈ {R, B}, say Red and Black, simultaneously chooses one
of four cards ai ∈ {1, 2, 3, J} (Ace, Two, Three, and Joker).

• The payoff matrix is given as follows:

Red / Black 1 2 3 J
1 0,1 1,0 1,0 0,1
2 1,0 0,1 1,0 0,1
3 1,0 1,0 0,1 0,1
J 0,1 0,1 0,1 1,0

• In words:

• Red wins iff both players choose J, or they choose different numbers.

• Black wins iff exactly one player chooses J, or both choose the same number.
 Example 4: O’Neill Game

• This game has a unique mixed strategy Nash equilibrium:

• Each player chooses J with probability 40 % and each of 1, 2, and 3 with

probability 20 % respectively.

• The winning probability is asymmetric: 40 % for Red and 60 % for Black.

 Example 4: O’Neill Game

Black

Experiment

Theoretical prediction

Red
 Summary: NE and DSE
• NE and DSE can be related through the lens of Bayesian rationality:

In Nash equilibrium…

players optimize their strategic choices with respect to their equilibrium beliefs,
and those beliefs are consistent with the actual choices of opponents.

In DSE…

players optimize their strategic choices with respect to all possible beliefs.

• A trade off between robustness and existence:

• DSE provides a robust prediction, but it need not exist in several games.

• NE always exists, but it is not robust against belief misspeci cation.

fi
Extensive-form Game
 Extensive-form Game

• Many strategic situations are inherently dynamic, where players make decisions at
multiple (and possibly different) points in time.

• Let’s revisit the coordination game:

Player 1 / Player 2 Movie Hiking

Movie 5,2 0,0

Hiking 0,0 3,4

• But now, we assume:

• P1 moves rst.

• After observing P1’s choice, P2 moves.

fi
 Extensive-form Game

• Their dynamic decision making can be summarized as the following game tree:

• Each node is labeled with a player, indicating who makes a decision at that point.

• Each edge (arrow) represents an action available to the player at each node.

• At each terminal node, the resulting payoffs for all players are speci ed.

fi
 Extensive-form Game

• Each player’s (pure) strategy in an extensive-form game is a contingent plan

that speci es the action chosen by the player for every decision node where it is
their turn to move.

• For example, in the coordination game:

• The set of Player 1’s strategies is A1 = {M, H} .

• The set of Player 2’s strategies is

A2 = {(M, M), (M, H), (H, M), (H, H)},

where (M, H) describes the P2’s contingent plan to take M when P1 takes M
and H when P1 takes H, etc.
fi
 NE in Extensive-form Game

• The reduced normal form of an extensive-form game is the normal-form game,

where each player’s set of actions is exactly given as the set of the player’s pure
strategies in the original extensive-form game.

P1 / P2 (M,M) (M,H) (H,M) (H,H)

M 5,2 5,2 0,0 0,0

H 0,0 3,4 0,0 3,4

• In the coordination game, there are three (pure strategy) Nash equilibria.

• Which one is the most plausible?

NE in Extensive-form Game

(M, (M, M))

P2: “I’ll take M if P1 takes H.”

Is this credible? — No, it is an empty threat !

If P2 reaches this decision node, he would

actually choose H instead.
 NE in Extensive-form Game

(H, (H, H))

P2: “I’ll take H if P1 takes M.”

…Again, this is an empty threat.

 Subgame Perfect Equilibrium

• A subgame of the extensive-form game is a portion of the original game that

starts at a particular decision node and includes all subsequent nodes and edges
stemming from that node.

• A strategy pro le in the extensive-form game is said to be a subgame perfect

equilibrium (SPE) if it forms a Nash equilibrium in every subgame.
fi
 Subgame Perfect Equilibrium
(M, (M, H))

• There are three subgames (including the original one itself).

• (M, (M, H)) forms a NE in every subgame ➡ it is an SPE!

• In general, SPE of extensive-form games can be derived using

backward induction.
 Example 2’: Stackelberg Competition
• A dynamic version of Cournot Competition

• Two rms, each trying to maximize the pro t:

ui(qi, qj) = (P − c) ⋅ qi where P = max {a − b(q1 + q2), 0} .

• But now, we assume:

• First, Firm 1 decides on the output level q1 .

• After observing Firm 1’s choice, Firm 2 decides on q2 .

Exercise.

• Derive the SPE of Stackelberg Competition.

• For each rm, compare the equilibrium production levels with those in the NE of
Cournot Competition; are they higher or lower? What about pro ts?
fi
fi
fi
fi
 Example 3’: Ascending-price Auction

• The auction begins with a starting price (let’s say $0).

• The auctioneer continuously increases the price, and at each moment, participants
must decide whether to “stay” in the auction or “leave.”

• In practice, the minimum increment (e.g., $1) is usually determined.

• As long as multiple participants choose to stay, the price keeps increasing. The
auction terminates when only one participant remains. This last participant is the
winner, who receives the good & pays the nal price.

• In a sense, this auction is strategically equivalent to 2PA. (Why?)

• Practical difference: Unlike sealed-bid auctions (1PA/2PA), the ascending-price

auction is “open.” The decisions of all bidders are fully revealed to each other
during the auction.
fi
 (Im)perfect Information

• In the previous examples, actions are observable & players are perfectly informed
about all past decisions.

• In other situations, each player may have only partial information about actions
taken previously when making a decision.

• We model this imperfect information by introducing an information set.

 (Im)perfect Information

• An information set for a player is a set of decision nodes at which it is that

player’s turn to move, typically represented by a dashed line in the game tree.

• When multiple nodes belong to the same information set, it means that the player
cannot distinguish between these nodes.
 (Im)perfect Information

• A player’s (pure) strategy in an extensive-form game with imperfect

information assigns one action to each information set.

• Note that the player has to choose the same action for all decisions nodes
within the same information set.

• A subgame of the extensive-form game with imperfect information is a portion of

the original game with the following properties:

• The subgame starts at a particular decision node and includes all subsequent
nodes and edges stemming from that node.

• Any information set intersecting the subgame must lie entirely within it.

• A strategy pro le is said to be a subgame perfect equilibrium (SPE) if it

forms a Nash equilibrium in every subgame.
fi
 (Im)perfect Information

Exercise. How many subgames does the following game have? Write down the set
of strategies for each player and derive SPE.
 (Im)perfect Information

Remark. Any normal-form game (i.e., a simultaneous-move game) can also be

represented as an extensive-form game with imperfect information.
 Existence of SPE

• An extensive-form game is said to be nite when the set of nodes is nite. It has
perfect information if every information set consists of a single decision node.

Theorem 2 (Kuhn’s Theorem).

• Every nite extensive-form game with perfect information has an SPE.

• Proof is based on backward induction.

• For each of the last decision nodes, we choose an optimal action for the player
acting at that node and replace the node with a terminal node that re ects the
resulting payoff pro le from the chosen optimal action.

• Repeat the same procedure, working backward through the game tree, until
reaching the initial node.
fi
fi
fi
fl
fi
 Existence of SPE

• Consider any nite two-person game with perfect information, where:

• Two players, A and B, move alternately.

• The game ends in one of three outcomes: “A Wins,” “B Wins,” or “Draw.”

Corollary (Zermelo’s Theorem).

• In any game described as above, at least one player has a drawing

strategy (i.e., can at least force a draw). In particular, if the game cannot
end in a draw, at least one player has a winning strategy.
fi
 Existence of SPE

• Two players, Black and White, alternately put one stone on a rectangle table.

• All stones are the same size and shape.

• Players can put a stone anywhere on the table as long as it doesn’t overlap
with any existing stones.

• The game continues until there is no space left to place a new stone. The winner is
the player who placed the last stone.

Question: Which player has a winning strategy, and what is the winning strategy?
 Appendix:
Proof of Theorem 1

Go Back
 Existence of Nash Equilibrium

• A best-response correspondence is de ned as

BRi(σ−i) := arg max ui(σi, σ−i) .

σi∈Σi

• In other words, BRi(σ−i) is the set of mixed strategies that maximize i ’s expected
payoff when other players are expected to play σ−i .

• Note that σ* = (σ*

1
, …, σ*
n ) is a Nash equilibrium iff σ*
i
∈ BRi(σ*
−i
) for all i ∈ N .
fi
 Existence of Nash Equilibrium

• An Euclidean subset X ⊆ ℝn is compact if it is closed and bounded.

• Recall that a function f : X → X is said to be continuous if for any sequence

{xn}∞
n=1 in X,

xn → x ⟹ f(xn) → f(x) .

Brouwer’s Fixed Point Theorem.

• Suppose X ⊆ ℝn is a nonempty, convex, compact set.

• Then, any continuous function f : X → X admits at least one xed point, i.e.,

∃x* ∈ X such that x* = f(x*) .

fi
 Existence of Nash Equilibrium
• A correspondence is a set-valued function F : X → 2X that assigns a subset
F(x) ⊆ X to each point x ∈ X .

• F is nonempty-valued if F(x) ≠ ∅ for all x ∈ X .

• F is convex-valued if F(x) is a convex set for all x ∈ X .

• F has a closed graph if for any sequence {xn} and {yn} in X such that
yn ∈ F(xn), if xn → x and yn → y, then y ∈ F(x) .

Kakutani’s Fixed Point Theorem.

• Suppose X ⊆ ℝn is a nonempty, convex, compact set.

• Then, any nonempty & convex-valued correspondence F : X → 2X with

a closed graph admits at least one xed point, i.e.,

∃x* ∈ X such that x* ∈ F(x*) .

fi
Proof of Nash’s Theorem.

• Consider the best-response correspondence F : Σ → 2Σ given by

F(σ) = BR1(σ−1) × ⋯ × BRi(σ−i) × ⋯ × BRn(σ−n),

where

BRi(σ−i) = arg max ui(σi, σ−i) .

σi∈Σi

• Observe that σ* constitutes a Nash equilibrium if and only if it is a xed point of

the correspondence BR, i.e., σ* ∈ BR(σ*) .

• Let’s check the Kakutani’s assumptions to apply the xed point theorem!
fi
fi
• It is straightforward to check that:

☑ Σ is a nonempty convex compact set.

• This is true because Σ = Δ(A1) × ⋯ × Δ(An), where each Δ(Ai) is nonempty,

convex, closed, and compact.

☑ F is non-empty & convex-valued.

• Recall that F = BR1 × ⋯ × BRn, where

BRi(σ−i) = arg max ui(σi, σ−i) .

σi∈Δ(Ai)

• Since Σi is compact and ui is continuous, BRi(σ−i) is nonempty.

(by Weierstrass Theorem)

• Moreover, since ui is af ne, BRi(σ−i) is convex.

fi
• It remains to show F has a closed graph.

• Consider any sequences {σ k} and {ρ k} of mixed strategy pro les such that

i) σ k = (σ1k, …, σnk) → σ and ρ k = (ρ1k, …, ρnk) → ρ;

ii) ρ k ∈ BR(σ k) for every k ∈ ℕ .

• We want to show that ρ ∈ BR(σ) .

• By (ii), each ρik best responds to σ−i

k
, namely, we have

ui(ρik, σ−i
k k
) ≥ ui(ai, σ−i ), ∀k ≥ 1, ∀i ∈ N, ∀ai ∈ Ai .

• Since ui is continuous, by (i), taking the limit k → ∞ yields

ui(ρi, σ−i) ≥ ui(ai, σ−i), ∀i ∈ N, ∀ai ∈ Ai .

• This shows that ρi ∈ BRi(σ−i), as desired.

fi
• We have con rmed that:

• Σ is nonempty, convex, and compact.

• F is nonempty & convex-valued and has a closed graph.

• Therefore, Kakutani’s xed point theorem implies there exists σ* ∈ Σ such that

σ* ∈ BR(σ*),
which is a Nash equilibrium.

Go Back
fi
fi