Game Theory

Normal-Form Games
Tetsuya Hoshino January 10, 2023

How should we formalize a strategic environment—i.e., a game? There are multiple ways to
represent the game. Among them, a normal-form game is the simplest one. It highlights
fundamental elements to describe a game.
We begin with possibly the most famous example in game theory:

Example 1 (Prisoners’ Dilemma). Two prisoners, called 1 and 2, are being interrogated.
They are in separate rooms, and no communication is allowed. Each prisoner chooses either
to stay silent or to testify that they have committed the crime. We refer to staying silent as
action C (which stands for cooperation with the other prisoner) and to testifying their crime
as action D (which stands for deviation). The penalties on the prisoners depends on their
actions:
• If prisoners 1 and 2 take C and C, they serve 1- and 1-year-long imprisonment.
• If prisoners 1 and 2 take C and D, they serve 3- and 0-year-long imprisonment.
• If prisoners 1 and 2 take D and C, they serve 0- and 3-year-long imprisonment.
• If prisoners 1 and 2 take D and D, they serve 2- and 2-year-long imprisonment.
How would we abstract this story? The minimum element is “who does what for what.” In
this example, two prisoners 1 and 2 each choose either C or D, to minimize the penalties. □

1 Normal-Form Game
To represent a strategic environment in a normal form, it suffices to list up the minimum
elements—i.e., who does what for what?

Definition 1. A normal-form game, or the normal-form representation of a game, is a

tuple G = 〈I, (Ai , ui )i∈I 〉 such that:
1. I is the set of players.
• i is a generic player, and −i is all players but i.
2. Ai is the (non-empty) set of actions, or strategies, available to player i.
• ai ∈ Ai is an action or (pure) strategy available to player i.
3. ui : A → R is player i’s payoff function.1
!
• A= i∈I Ai is the set of action profiles, or (pure) strategy profiles.
• a = (ai )i∈I ∈ A is a strategy profile, which specifies every player’s strategy.
• ui (a) is player i’s payoff when players take strategy profile a.
1
Payoffs are sometimes referred to as utilities. In most cases, payoffs are synonymous with utilities, but they
are a more general notion that includes profits, etc.

1
Example 2. Example 1 can be formalized as the normal-form game G such that:
1. I = {1, 2} is the set of players.
2. Ai = {C, D} is the set of strategies available to player i.
3. ui : A → R is player i’s payoff function defined by the following table:

C D
C −1, −1 −3, 0
D 0, −3 −2, −2

Table 1: prisoners’ dilemma

Player 1’s strategies C and D are identified as the rows, and player 2’s strategies C
and D are identified as the columns. That is, players 1 chooses a row and player 2
chooses a column, and thus they are called the row player and the column player
respectively. In each cell, the first and second numbers denote the payoffs for players 1
and 2 respectively. For example, “−3, 0” means that players 1 and 2, respectively, receive
payoffs −3 and 0 when they play strategy profile (C, D).
This table specifies all the elements required for the normal-form representation. Hence, we
often use this payoff-table representation for a normal-form game. □

Simutaneity of Decision-Making In a normal-form game, each player i chooses strategy ai

simultaneously. Here, “simultaneous” does not necessarily mean that the strategies must be
taken at the same point in time. In Example 1, the two prisoners need not choose their
strategies at the same point in time. The key feature is that neither prisoner knows which
strategy the other prisoner is choosing. That is, each player makes a decision individually,
without being informed of the choices of any other players. A normal-form game is often
called as a simultaneous-move game.

1.1 Examples
Let us see several examples.

Example 3 (Braess’s paradox). Recall Braess’s Paradox in the previous note: There are 10
people, each of whom chooses either route SAG or SBG (where road AB is unavailable). Each
person wants to minimize her travel time. Her travel time when she chooses SAG is 11 + x,
where x is the number of people choosing SAG (including herself); her travel time when she
chooses SBG is 11 + y, where y is the number of people choosing SBG (including herself).
This situation has the normal-form representation G such that:
1. I = {1, 2, . . . , 10} is the set of peoples.
2. Ai = {1, 0} is person i’s strategy space, where strategy 1 is SAG and strategy 0 is SBG.

2
 3. ui : A → R is person i’s payoff function defined by
"
#−11 − % aj if ai = 1
ui (a) =
j∈I
$−11 − %
j∈I (1 − aj ) if ai = 0,
% %
where j aj is the number of people choosing SAG and j (1 − aj ) is the number of
people choosing SBG. □

Example 4 (Cournot game). Firms 1 and 2 produce a homogeneous product. Each firm i
chooses its quantity qi ≥ 0. The price of the product depends on the total quantity Q = q1 +q2 ,
and the inverse demand function P is such that P (Q) = max{1−Q, 0}. Each firm i’s production
cost is zero regardless of qi , and it maximizes its profit P (Q)qi .
The Cournot game is the normal-form game G = 〈I, (Ai , πi )i∈I 〉 such that:
1. I = {1, 2} is the set of firms.
2. Ai = R+ is the set of firm i’s strategies.
3. πi : A → R is firm i’s payoff function defined by

πi (qi , q−i ) = P (Q)qi . □

Example 5 (Bertrand game). Firms 1 and 2 produce a homogeneous product. Each firm i
chooses its price pi ≥ 0. Assume that there is a unit mass Q = 1 of consumers. They buy
the products from the firm with the lowest price, or if the prices are equal, the consumers are
equally split between the two firms. Each firm i’s production cost is zero regardless of the
demanded quantity, and it wants to maximizes its profit.
The Bertrand game is the normal-form game G = 〈I, (Ai , πi )i∈I 〉 such that:
1. I = {1, 2} is the set of firms.
2. Ai = R+ is the set of firm i’s strategies.
3. πi : A → R is firm i’s payoff function defined by
"
&
&
& pi
& if pi < p−i
#
πi (pi , p−i ) = pi
2 if pi = p−i □
&
&
&
&
$0 if pi > p−i .

Coffee Break (Cournot or Bertrand?). Which is the “right” model of oligopoly competition,
Cournot or Bertrand? Which one should we use? It depends on the situation. Neither is
“right.” Neither is “wrong.” Some real-world situations are closer to Cournot competition,
while others are closer to Bertrand competition.
1. An example that is close to Bertrand competition is the competition between two neigh-
boring gas stations. They sell a homogeneous product (gasoline), and each gas station i
offers a price of pi (per liter).

3
 2. An example that is close to Cournot competition is the competition in a local fish market.
Fishing boats (as players) go fishing to catch fish (i.e., chooses their production) and land
them in a market near the local port. Once landed, the fish are sold in the market at the
market price.
3. How about electronics such as TVs? This is a complicated case, but since producers
cannot choose the retail price and retailers compete with each other to set the price, we
can think of competition (among producers) as similar to Cournot competition, where
each producer ships a certain quantity of product and then the market sets the price. □

2 “Solution” to Normal-Form Games

We have set up a normal-form game. The next step is to “solve” it. But, what does “solving a
game” mean? It means to predict how the game will be played. The prediction is referred to as
a “solution” that describes which strategies players will adopt and thus the result of the game.
The rule to make a prediction is called a “solution concept” or an “equilibrium concept.”
To this end, we describe what players may play in a given game. There are (at least) three
approaches to predicting play:
1. Self-enforcement: Suppose that players have reached a non-binding agreement about
how they play a game, where “non-binding” means that no player is compelled (by any
external forces) to obey the agreement. Are all the players willing to obey the agreement?
If so, we can expect that they will play the game as agreed; that is, the agreement is
self-enforcing. In contrast, if there is at least one player who wants to deviate from the
agreement, we cannot expect that they will play the game as agreed. This approach
leads to the following solution concepts, for example:
• Nash equilibrium
• Correlated equilibrium
2. Introspection: Before choosing strategies, players reason how they should play a game.
This approach examines how they reason their way to playing the game. Suppose that
they never play “bad” strategies. Then, the fact that player 1 never plays any bad
strategies may change player 2’s way of playing the game, and vice versa. This approach
leads to the following solution concepts, for example:
• Iterated strict dominance
• Rationalizability
3. Learning: As players play a given game many times, their play may converge to a
particular play through a process by which they learn how they should play and how
others play. This approach leads to the following solution concept, for example:
• Evolutionary equilibrium
• Self-confirming equilibrium
While these approaches are quite different, we will see that the predictions that they entail

4
are closely related. In terms of predictions of play, these approaches (at least for normal-form
games) are not very different, but the interpretations of how agents arrive to play according
to the predictions is quite different.