John Nachbar

Washington University in St. Louis

September 30, 2019

Game Theory Basics II:

Extensive Form Games1
1 Introduction.
The extensive form provides a mathematical language for describing and analyz-
ing strategic environments (i.e., games) with a non-trivial temporal structure. A
rigorous treatment of the formalism is cumbersome. My discussion, therefore, is
informal. For a formal treatment, see Fudenberg and Tirole (1991).
To keep things simple, I start with games in which each player acts in turn, fully
informed of the previous actions in the game. Such games are called games of perfect
information. I then introduce a more elaborate formalism that allows players to be
only partially informed of previous play in the game. Such games are called games
of imperfect information. Last, I briefly discuss environments in which players are
asymmetrically informed about details of the game even at the start of the game;
such games are called games of incomplete information.

2 Games of Perfect Information.

2.1 The basic formalism.
As an example of a game in extensive form, consider Figure 1. There are two players,
an industry incumbent (player 2) and a potential entrant (player 1). Player 1 moves
first, at the decision node labeled d0 , and chooses one of two actions, I (“in”: enter
the industry) or O (“out”: don’t enter the industry). If player 1 chooses I, then play
moves to the decision node labeled d1 and it is player 2’s turn. If player 2 chooses
F (“fight”: launch a price war), then play ends at the terminal node labeled d4 and
a payoff profile of (−1, −1). If player 2 chooses A (“accommodate”: don’t launch a
price war), then play ends at the terminal node labeled d3 and a payoff profile of
(15, 15). Finally, if player 1 chooses O, then play ends at the terminal node labeled
d2 and a payoff profile of (0, 35).
The immediate successors of d0 are d1 and d2 . The successors of d0 are d1 , d2 ,
d3 , and d4 . The immediate predecessor of d4 is d1 . The predecessors of d4 are d1
and d0 . And so on.
More generally, a finite extensive form game of perfect information consists of
the following components.
cbna. This work is licensed under the Creative Commons Attribution-NonCommercial-
1

ShareAlike 4.0 License.

1
 Player 1 Player 2 d4 (-1, -1)
d0 I d1 F
O A

d2 d3
(0, 35) (15, 15)

Figure 1: A Basic Entry Deterrence Game.

• There is a finite set I = {1, . . . , N } of players.

• There is a finite set D of decision nodes, which are the points in the game at
which players take action.
D is partitioned into disjoint sets Di , one for each player. The set Di contains
the decision nodes for player i. Di could be empty, in which case i is a dummy
player who gets payoffs but otherwise plays no role in the game.

• There is a finite non-empty set Dτ of special nodes called terminal nodes.

• For player i and each decision node d ∈ Di , there is a finite non-empty set
Aid , the actions available to player i at decision node d.
S
Let Ai = d∈Di Aid be the set of all of player A’s possible actions. Ai could
be empty if player i is a dummy player (in which case there is no d ∈ Di ).
For a given decision node d, each action at d is associated with a unique
immediate successor node, which can be either a decision node (but not d
itself) or a terminal node. If dˆ is an immediate successor of d then d is the
ˆ
immediate predecessor of d.

• There is a unique decision node d0 , the initial node, that has no predecessor
and that is the predecessor of every other node. As a matter of labeling
convention, d0 is typically (but not universally) a decision node for player 1.

• Every decision node has at least one successor and every node, decision or
terminal, except d0 , has exactly one immediate predecessor.

• To each terminal node, I associate a payoff profile.

The assumptions guarantee that play cannot cycle. Rather, play starts at d0 and
ends, after a finite number of decision nodes and actions, at some terminal node.

2
 Every terminal node is uniquely identified with its play path (or path of play),
which are the nodes that are passed through if that terminal node is reached. In
the example, d3 is identified with the play path (d0 , d1 , d3 ). I sometimes use play
path and terminal node interchangeably. One can also identify a play path by the
sequence of actions taken. Thus, in the example, d3 is identified by (I, A). The play
path taken in a chess match is reported in this way. The formal representation of
games that never end, discussed in Section 4, uses play paths in place of terminal
nodes.
In some games, randomness is an intrinsic part of the game. For example, in
many board games, there are points at which a player tosses a die. I model this
by introducing a new notional player, player 0, called Nature; the toss of a die is
then a randomization by Nature. For technical reasons, I defer further discussion of
Nature to Section 2.4.
Just as strategic form game boxes are convenient for small games but useless for
large games, so pictures like Figure 1 are convenient for small extensive form games
but useless for large or complicated extensive form games. The abstract formalism
does not require that one be able to draw a picture to have a well-defined extensive
form.
Remark 1. Is chess a finite game? It depends on what rules are being enforced for
ties/draws. Suppose that we declare the game a draw if the same board position
repeats three times. Since there are only a finite number pieces (even taking pawn
promotion into consideration) and only a finite number of squares on the board,
there is a finite upper bound on how long any chess game can last before either
there is checkmate, stalemate, or a board position repeats three times. Hence chess
is, under this rule, finite. Without such a rule, then it is possible that a game of
chess could last forever. 

2.2 Pure Strategies.

Definition 1. In a game of perfect information, a pure strategy is a function

s i : D i → Ai

such that for any d ∈ Di , si (d) ∈ Aid .

In words, a pure strategy is a complete contingent plan of action: no matter what

node you are at, the strategy specifies an action. One way to think of a strategy is
that it is a set of instructions for your agent, say a parent, who insists that you be
thorough and give instructions for every possible decision node that you might face,
including decision nodes that could be reached only if the agent makes a mistake
(because, after all, agents in general and parents in particular do sometimes make
mistakes).

3
 Let Si be the set of player i’s pure strategies. The cardinality of Si , which I
denote |Si |, is then Y
|Si | = |Aid |.
d∈Di

Example 1. In the entry deterrence game, Figure 1, S1 = {I, O}, S2 = {A, F }.

Thus, each player has two strategies. .
Example 2. Consider two-move chess, in white moves, then player 2 black, and then
the game ends in a draw; it isn’t a very exciting game.

• White has 20 strategies: each pawn can move either one or two spaces, and
each knight has two possible moves.

• Similarly, black has 20 actions at each decision node, but black has a total of
2020 ≈ 1026 strategies, because a strategy for black has to specify an action
for black at each of black’s 20 decision nodes (one for each of white’s actions).

In contrast, there are only 400 = 20 × 20 terminal nodes: there are only 400 ways
the game can be played. Note, in particular, that the set of black’s strategies is
much larger than the set of terminal nodes.
And one can go on in this way. For three-move chess, in which white moves, then
black, then white again, and then the game is a draw, white has far (far, far) more
strategies than could be physically represented: white has many, many orders of
magnitude more strategies than the estimated 1080 baryons (neutrons and protons)
in the observable universe.
As this example illustrates, the pure strategy sets can be very large even for
“small” games. 
Exactly as for games in strategic form, for games in extensive form, once one
has the sets ofQindividual pure strategies, one can construct the set of pure strategy
profiles, S = i Si , the individual
Q sets of mixed strategies, Σi = ∆(Si ), the set of
mixed strategy profiles, Σ = i Σi , and the set of correlated profiles, which are
elements of ∆(S). Notation such as (si , s−i ) and (σi , σ−i ) is exactly the same as in
games in strategic form.
Any pure strategy profile s ∈ S determines a play path/terminal node. Any (pos-
sibly mixed) strategy profile σ ∈ Σ determines a distribution over play paths/terminal
nodes. A decision node is on the play path iff it is reached with positive probability
(i.e., it is on a play path that is played with positive probability).
Remark 2. For completeness, I also need to describe what happens for a dummy
player, meaning a player for whom Di = ∅. For such a player, let Si consist of a
single notional strategy si . 

4
2.3 Mixed and Behavior Strategies.
Informally, in a mixed strategy, a player randomizes over pure strategies before the
start of play, and then executes that pure strategy. There is, however, another
obvious way to randomize, which is at each decision node individually.

Definition 2. A behavior strategy (also called a behavioral strategy) is a function

[
σib : Di → ∆(Aid )
d∈Di

such that for each decision node d ∈ Di , the mixed action σib (d) ∈ ∆(Aid ).

Although mixed strategies and behavior strategies are formally different (they
are defined on different spaces), they are closely connected in the following respect.
Say that any two strategies (pure or mixed or behavior) are Kuhn equivalent (named
in honor of Harold Kuhn) iff, for any given profile of strategies of the other players,
the two strategies induce the same probability distributions over play paths.

Theorem 1 (Kuhn’s Theorem for Finite Games of Perfect Information). For any
mixed strategy σi there is a Kuhn-equivalent behavior strategy, and vice versa.

Proof. Omitted. See, for example, Fudenberg and Tirole (1991). 

Example 3. To illustrate the basic idea, consider the two-player game in Figure 2 in
which player 1 moves first followed by player 2; I have omitted payoffs, since these
do not matter for the example.

Player 1 d0
A B

Player 2 d1 d2 Player 2

a b U M D

d3 d4 d5 d6 d7

Figure 2: A game to illustrate Kuhn’s Theorem.

This game has two decision nodes for player 2, d1 and d2 . There are two actions
available at d1 and three at d2 . Therefore, there are 2 × 3 = 6 pure strategies for
player 2 (even though there are only five terminal nodes): (a, U ), (a, M ), (a, D),
(b, U ), (b, M ), (b, D).

5
 Consider the behavior strategy in which player 2 randomizes ( 14 , 34 ) at node d1
and ( 61 , 36 , 62 ) at node d2 . This behavior strategy is Kuhn equivalent to the mixed
strategy (using the enumeration of pure strategies given above),
 
1 3 2 3 9 6
, , , , , .
24 24 24 24 24 24

For example, if player 1 plays the pure action A then under either player 2’s
mixed strategy or her behavior strategy, the induced distribution over the five ter-
minal nodes is ( 14 , 34 , 0, 0, 0). 
Because behavior strategies are so natural, it is common to refer to mixed strate-
gies even when, strictly speaking, one is using behavior strategies. By Kuhn’s The-
orem, there is not much, if any, harm in this, although it is wise to be aware of the
distinction between mixed and behavior strategies.

2.4 Nature.
To model certain types of randomness that can arise in a game, such as the roll of
a die, I introduce a new notional player, player 0, called Nature. By assumption,
Nature chooses a behavior strategy σ0b that is part of the extensive form.
Assume that Nature assigns positive probability to each of her available actions,
since, if any action by Nature were to receive zero probability, then all nodes that
follow that action would never be reached no matter what the actual players did and
hence any actions that would have been taken at such nodes would be strategically
irrelevant.

2.5 The Strategic Form.

Given strategies and payoffs, we can re-represent the extensive form game as a game
in strategic form. For example, the basic entry deterrence game in Figure 1 can also
be represented by the strategic form game box in Figure 3, a game that I already
discussed in the Game Theory I notes.

A F
I 15, 15 −1, −1
O 0, 35 0, 35

Figure 3: The Entry Deterrence Game in Strategic Form.

If Nature is a player in the game (Section 2.4), then the convention is to use
Nature’s behavior strategy to compute the expectations of player payoffs and record
those expectations in the strategic form. This assumes, of course, that players are
expected payoff maximizers.

6
 To take a trivial example, consider a one-player gambling game in which the
player can either stay out, for a payoff of 0, or go in (take the bet), in which case
a fair coin is tossed with payoffs of either 100 or -99. In the strategic form, the
payoff from staying out is 0 and the payoff from going in is 100/2 − 99/2 = 1/2.
The strategic form is given in Figure 4.

In 1/2
Out 0

Figure 4: The Strategic Form for the Simple Gambling Game.

2.6 Nash Equilibrium.

The Nash equilibria (NE) for a game in extensive form are the Nash equilibria of
the associated strategic form game.
Consider again the Entry Deterrence Game of Figure 1 and Figure 3. This game
has infinitely many NE: (I, A) and a continuum of partly mixed NE of the from
(O, (q, 1 − q)), where q is the probability of A and satisfies q ≤ 1/16. In words, there
are a continuum of NE in which the entrant stays out and the incumbent threatens
to fight with high enough probability.
As I discussed in the notes on Strategic Form Games, there is a precise sense
in which “typical” finite strategic form games have only a finite (and odd) number
of equilibria. So the game of Figure 3 is not typical in this sense. However, there
is nothing especially atypical about the extensive form Figure 1. The moral is that
for extensive form games, we should expect that the set of Nash equilibria may be
infinite.
A related point is that while there are infinitely many Nash equilibria in this
example, there are only two distributions over terminal nodes that occur in equi-
librium: either d3 is reached for sure or d2 is reached for sure. This turns out be
general: for “typical” extensive form games, the set of Nash equilibrium distribu-
tions over terminal nodes is finite (but not, as the example illustrates, necessarily
odd); see Theorem 2 in Kreps and Wilson (1982) and the discussion immediately
following.

2.7 Alternate Extensive Forms for the Same Game.

Consider the one-player game in Figure 5.
There are four strategies, (O, A), (O, B), (I, A, ), (I, B), but strategies (O, A)
and (O, B) generate the same play path and one can argue, therefore, they should
be collapsed into the single strategy O. Put differently, one can argue that the game
is “really” as in Figure 6; the representation in Figure 5 adds strategically irrelevant

7
 1 I 1 B
1000
d0 d1
O A

999 0

Figure 5: A One-Player Game.

clutter. The classic cites for extensive form simplifications of this kind are Dalkey

B
d0 1000
A
O

999 0

Figure 6: A Modification of the One-Player Game.

(1953) and Thompson (1952). For more on the connection between this issue and
the reduced form in strategic form games, see Mailath et al. (1993) and Mailath
et al. (1994).
My own view is that I don’t think it is obvious that the game of Figure 5
should be modified to that of Figure 6. First, the extensive form as originally given
may be a descriptively accurate representation of the environment: in the actual
environment, it may be that the player must literally move twice to get the payoff
of 1000. Second, the difference between Figure 5 and Figure 6 could matter: a
player who didn’t trust herself not to make a mistake might opt instead for O. The
classical position in game theory is that if the player has a potential for making an
error at d1 , then that should be represented in the game form, say by adding moves
by Nature: if the player chooses B, then Nature moves next and chooses whether B
is actually implemented or whether instead A is implemented. An alternate view,
to which I currently lean, is that the possibility of player error should instead be
modeled outside of the extensive form, perhaps as a strategy restriction: the player
cannot choose the pure strategy (I, B) but must instead choose a behavior strategy
that involves a small amount of randomization (mistakes) at d1 .

8
2.8 Interpreting Payoffs.
In the notes on strategic form games, I said that payoffs encode preferences over
outcomes, which record both prizes and strategy profiles. In particular, players can
rank outcomes differently even if the prizes are the same but the strategy profiles
are different.
The extensive form framework is less flexible than this in the following sense.
By assigning payoffs to terminal nodes, we are implicitly saying that two different
strategy profiles that play the same, and hence must generate the same prize, gen-
erate the same payoffs. For example, consider again the game in Figure 2. The
profiles (B, aD) and (B, bD) both reach terminal node d7 and therefore generate
the same payoffs, but the profiles are different because Player 2’s strategy dictates
a different action contingent on Player 1 playing A. A strategic form game, in
principle, would allow payoffs to differ in this case. This observation implies that
strategic form games that are derived from extensive form games display structure
that reflects details of the extensive form. I won’t pursue this point further but it
has been important to the literature on whether, or to what extent, analysis of the
extensive form can be done in the strategic form; see Mailath et al. (1993), cited
earlier.

2.9 Nim, Tic-Tac-Toe, and Chess.

Consider parlor games like Nim and Tic-Tac-Toe: 2-player finite games of perfect
information, with no moves by Nature, in which either one player or the other wins,
or there is a tie (ties can happen in Tic-Tac-Toe but not in Nim). Assume that
players care only about who wins and that a win is better than a tie is better than
a loss. Games with this sort of payoff structure are called strictly competitive.
In a game like Nim, where there can be no ties, one can show that one of the
players has a strategy that guarantees a win regardless of what the other player
does; which player wins depends on the details of the game. Informally, the idea
is to start at the terminal nodes and work backwards. Consider any decision node
whose successors are all terminal nodes. For a decision node for player 1, either
all successor terminal nodes have outcome (L, W ), in which case label this decision
node (L, W ), or at least one successor terminal node has outcome (W, L), in which
case label this decision node (W, L). Now modify the game with the decision node of
the original game changed into a terminal node in the new game, with the outcome
as labeled. And symmetrically if the decision node belongs to player 2. Working
backward in this way, we eventually (because the original game was finite) end up
with a modified game consisting of an initial node, which for convenience I assume is
a decision node for player 1, followed by terminal nodes. If all terminal nodes have
outcome (L, W ), then this argument implies that there is a strategy for player 2 in
the original game that guarantees that she wins, regardless of what player 1 does.
If, on the other hand, at least one terminal node in the modified game has outcome

9
(W, L), then there is a strategy for player 1 in the original game that guarantees
that he wins.
The analysis of games like Tic-Tac-Toe is similar but slightly trickier because of
the possibility of ties. The conclusion for Tic-Tac-Toe is that one of the players has
a strategy that guarantees no worse than a tie; again, which player this is depends
on the details of the game.
The original study of such games dates back to Zermelo (1913). (Yeah, it’s in
German; a number of classic game theory papers are. And, no, I don’t read German.
But a translation, along with a detailed discussion, can be found here.) Zermelo
used chess as his leading example.2 With chess, there is an issue as to whether it
is finite; see Remark 1. Zermelo did not, in fact, assume that chess is finite. If,
however, we make chess finite, as in Remark 1, then one can show that chess is
weak dominance solvable in the strategic form with two rounds of weak dominance
deletion; see Ewerhart (2000). In this sense, chess is “simple;” the strategic form,
however, is mind-bogglingly gigantic; see Example 2 above.

3 Games of Imperfect Information.

3.1 Information Sets.
To capture strategic environments in which players may be only partially informed
of what has happened in the game to date, I assume that, for each player i, if Di 6= ∅,
then Di is partitioned into disjoint, non-emptySinformation sets. Let Hi denote the
set of information sets for player i. Thus, Di = h∈Hi h and, for any h ∈ Hi , h ⊆ Di ,
h 6= ∅ (assuming Di 6= ∅), and if h 6= ĥ then h ∩ ĥ = ∅. The intended interpretation
is that if d, dˆ ∈ h ∈ Hi , then player i can observe that she is at information set h,
but she cannot observe whether the decision node is d or d. ˆ The initial node, d0 ,
is assumed to be in a singleton information set (an information set containing just
d0 ). And if Nature is include as player 0, as discussed in Section 2.4, then all of
Nature’s “decision” nodes are singletons.
For any two distinct decision nodes d and dˆ in an information set, the action sets
Aid and Aidˆ are the same. If Aid and Aidˆ were different, then player i would be able
to infer which node they were at merely by observing which action set was available.
This is contrary to the intended interpretation of information sets. I therefore let
Aih denote the set of actions available at every node d ∈ h.
If all information sets are singletons then the game is one of perfect information,
as in Section 2, but with the addition of nuisance notation about singleton informa-
tion sets; I ignore the extra notation in this case and refer to such games as games
of perfect information. If the game has non-singleton information sets then I refer
to the game as being one of imperfect information.
2
And there is this brilliant clip from Futurama (Season 2, Episode 2)

10
 As an illustration, see Figure 7. To avoid clutter, I have stopped labeling terminal
nodes.
r2 d4 (-1, -1)
d0 Player 1
F
H T
A
d1 d2 Player 2

, 15) H T H T

(1, -1) (-1, 1) (1, -1) (-1, 1)

Figure 7: An extensive form version of Matching Pennies.

Player 1 can choose either H or T . In either case, player 2 learns that she is
at an information set containing the nodes d1 and d2 (where the information set
is represented by the oval) but not which decision node she is at: she does not
know what action player 1 chose. In this example, it is as if the players were acting
simultaneously. Player 1 nominally moves first, but player 2 moves in ignorance of
what Player 1 has done. With the payoffs as written, this game is an extensive form
version of Matching Pennies.
Remark 3. One might argue that a better representation of a game like Matching
Pennies (or Rock-Paper-Scissors) is one in which there is randomness as to which
player nominally moves first. This can be handled by giving an initial decision
node to Nature, who tosses a fair coin as to which player nominally goes first. But
if neither player knows who nominally goes first, then the strategic form is just
standard Matching Pennies. And even if players do learn who nominally goes first,
but nothing else, the reduced strategic form ends up being just Matching Pennies.
For example, in this modified game, the strategy, go H if first and T if second is
equivalent to randomizing 50:50 in standard Matching Pennies. 

3.2 Strategies, the Strategic Form, and Nash Equilibrium Revis-

ited.
To keep things short: the rest of the construction for imperfect information is much
like the construction for perfect information with two caveats.
First, a pure strategy is now a function from player i’s information sets to her
actions:
si : Hi → Ai

11
such that for each h ∈ Hi , si (h) ∈ Aih . Similarly, a behavior strategy becomes
[
σib : Hi → ∆(Aih )
h∈Hi

such that for each h ∈ Hi , σii (h) ∈ ∆(Aih ). The count of player i’s pure strategies
becomes, Y
|Si | = |Aih |.
h∈Hi

An information set is on the play path iff it contains a decision node that is on the
play path (a decision node belonging to a play path that is played with positive
probability).
Second, Kuhn’s Theorem (Theorem 1) is not true in general but the games for
which it fails look pathological (see Section 3.3) and so I ignore this issue.

3.3 Pathologies.
The formal definition of a game of imperfect information allows for some games that
arguably, should be excluded.
One issue involves a property called perfect recall. Informally (I won’t give a
formal definition), perfect recall means that information sets encode the property
that each player can observe (or “knows”) her own previous information, including
her own previous actions. It may be easier to understand this with an example of
imperfect recall.
Example 4. The one-person game in Figure 8 is called the “absent-minded driver”
game and was introduced in Piccione and Rubinstein (1997); see also Aumann et al.
(1997). The story for the game is that a driver is entering a city and wants to exit

C C
1
d0 d1 d2

E E

0 4

Figure 8: The Absent Minded Driver Game.

at the second city exit. This is labeled d2 in the extensive form, but the driver can’t
see the exit labels (and there is no GPS), and he is absent-minded (perhaps tired),
so he does not remember, after entering the city, whether or not he has already
passed the first exit, labeled d1 in the extensive form. The payoff from taking the

12
first exit is low (bad part of town) while the payoff from continuing out of the city
is intermediate (he’ll have to turn around and try again).
This game of imperfect recall spurred a lively debate as to how analyze it (see
the two cited articles), but as a practical matter, economics almost always works
only with games of perfect recall. My own take is that modifying the extensive form
is the wrong way to model bounded rationality issues such as forgetfulness. 
In games of imperfect information, Kuhn’s Theorem (Theorem 1) holds for games
of perfect recall but not necessary for games with imperfect recall.
Another extensive form pathology concerns “playability.” The imperfect infor-
mation formalism allows for the possibility of games that do not admit a natural
interpretation in terms of players moving either simultaneously or in sequence. See
Mailath et al. (1994), previously cited. Arguably, this possibility is a defect in the
definition of a game of imperfect information.

4 Games Without Terminal Nodes.

We will frequently be interested in games without terminal nodes. The leading
example is a repeated game: the same players play the same game, called the stage
game, over and over again, resulting in a new, much larger and more complicated
game. It is often convenient to model the repeated game as a game that can last
literally forever, with payoffs for the repeated game equal to some aggregation of
the stage game payoffs; the aggregation is most commonly a discounted sum but
there are other approaches, such as the limsup or liminf of the payoff averages over
finite numbers of periods.
In terms of the extensive form formalism, there is a simple way to represent
payoffs in such games. Recall that there is an 1-1 relationship between paths of
play and terminal nodes. Therefore, to represent payoffs in extensive forms without
terminal nodes, simply attach payoffs to the paths of play rather than the (missing)
terminal nodes.
Note, by the way, that an infinitely repeated game will have infinitely many pure
strategies (in fact a continuum of pure strategies, if the stage game has at least two
actions for that player).

5 Games of Incomplete Information.

The implicit assumption thus far has been that players start the game from a posi-
tion of informational symmetry: every player knows the game structure and every
player’s payoffs (and every player knows that every player knows this, etc). Games
in which this is not true are called games of incomplete information. The classic
investigation of how to formalize such strategic environments in general are the
three papers in Harsanyi (1967). But games of incomplete information had been

13
investigated earlier, most notably in Vickrey (1962), which introduced incomplete
information to study auctions. And there has been much subsequent work; indeed
incomplete information has been a main focus, perhaps the main focus, of game
theory research for decades. For a brief survey that includes some more recent
references, see Myerson (2004).
I investigate games of incomplete information in more detail later in the course.
But here is a brief look at some of the foundational issues. In a game of incomplete
information, which I also call a Bayesian game, each player’s private information is
encoded as that player’s type. In the Bayesian game, a strategy, called a Bayesian
strategy, is a function from a player’s type to her strategy in the underlying game.
For example, consider a card game in which players initially see only their own card
hand and assume that that hand is their type. The implicit assumption is that the
player has an overall strategy, her Bayesian strategy, that encodes her actions not
only for her actual realized card hand but also for any other hand she might have
been dealt. A Bayesian equilibrium is then a profile of Bayesian strategies with the
property that each player’s Bayesian strategy is a best response to the Bayesian
strategies of the other players.
There is, you may have noticed, a problem with this construction: in order to
compute their own expected payoffs, players have to have probability distributions
over the type profiles of the other players (and those probabilities can depend on
the player’s own type, since that could contain information about the types of the
other players). Moreover, two players could, in principle, disagree about the type
probabilities of a third player. It all gets very complicated very quickly.
In practice, game theory has typically adopted the simplifying assumption that a
game of incomplete information can be modeled as a game of complete but imperfect
information in which Nature moves first and selects types. The Bayesian equilibrium
is then just the Nash equilibrium of this game. When I adopt this way of modeling
incomplete information (which is nearly all the time), I often don’t bother with the
term Bayesian equilibrium.
Because Nature’s behavior strategy is part of the description of the game, the
imperfect information approach to modeling incomplete information implies that
players agree on the joint type distribution. This informational assumption is called
a common prior. Harsanyi (1967) advocated for this assumption, a philosophical
position called the Harsanyi Doctrine. I find the common prior approach to be com-
pelling for standard card games; arguably, players “should” agree on the probability
distribution over card deals (although even here you can imagine players disagreeing
about whether the dealer is a cheater and, if so, exactly how the dealer is cheat-
ing). For many other settings, the common prior assumption is more controversial,
although still widely adopted if only for convenience.
The common prior assumption has substantive implications. Perhaps most fa-
mously, it lies at the heart of “no trade” theorems, which say (very loosely) that
under a common prior, players cannot agree to trade based solely on differing in-

14
formation/beliefs about the state of the world. For example, I cannot get you to
buy my shares in company X based on our differing views of company X’s prospects
(say, because I have received inside information). See Milgrom and Stokey (1982)
and also Morris (1994).
Finally, here are two other, rather technical, comments, both of which relate to
Bayesian games with a continuum of types (which frequently happens in applica-
tions; for example, in auctions, it is common for the type space to be the set of
possible buyer values, which is often taken to be an interval in R).
First, there are measure theoretic issues with defining behavior strategies in such
settings. Look at Milgrom and Weber (1985) if you are interested.
Second, there is some ambiguity as to what optimization means in such settings.
Strictly speaking, optimization means choosing the Bayesian strategy that maxi-
mizes expected payoffs. But two Bayesian strategies that differ on a set of types of
probability zero will generate the same expected payoffs, even if one (or both) of the
Bayesian strategies do not, in fact, optimize conditional on one of the zero proba-
bility types (of which there may be an uncountable infinity). It is not uncommon to
see authors implicitly assume that optimization means optimization for every type
and not just for a probability 1 set of types.

References
Aumann, R., S. Hart, and M. Perry (1997): “The Absent-Minded Driver,”
Games and Economic Behavior, 20.

Dalkey, N. (1953): “Equivalence of Information Patterns and Essentially Determi-

nate Games,” in Contributions to the Theory of Games, Volume II, ed. by H. W.
Kuhn and A. W. Tucker, Princeton University Press.

Ewerhart, C. (2000): “Chess-like Games Are Dominance Solvable in at Most Two

Steps,” Games and Economic Behavior, 33, 41–47.

Fudenberg, D. and J. Tirole (1991): Game Theory, Cambridge, MA: MIT

Press.

Harsanyi, J. (1967): “Games with Incomplete Information Played by Bayesian

Players, Parts I, II, III,” Management Science, 14, 159–182, 320–334, 486–502.

Kreps, D. and J. Wilson (1982): “Sequential Equilibria,” Econometrica, 50,

863–894.

Mailath, G. J., L. Samuelson, and J. M. Swinkels (1993): “Extensive Form

Reasoning in Normal Form Games,” Econometrica, 61, 273–302.

——— (1994): “Normal form structures in extensive form games,” Journal of Eco-
nomic Theory, 64, 325–371.

15
Milgrom, P. and N. Stokey (1982): “Information, Trade and Common Knowl-
edge,” Journal of Economic Theory, 26, 177–227.

Milgrom, P. and R. Weber (1985): “Distributional Strategies for Games with

Incomplete Information,” Mathematics of Operations Research, 10, 619–632.

Morris, S. (1994): “Trade with Heterogeneous Prior Beliefs and Asymmetric In-
formation,” Econometrica, 52, 1327–1347.

Myerson, R. B. (2004): “Comments on “Games with Incomplete Information

Played by ‘Bayesian’ Players, I-III”,” Management Science, 50, 1818–1824.

Piccione, M. and A. Rubinstein (1997): “The Absent-Minded Driver’s Paradox:

Synthesis and Responses,” Games and Economic Behavior, 20, 121–130.

Thompson, F. B. (1952): “Equivalence of Games in Extensive Form,” Research

Memorandum 759.

Vickrey, W. (1962): “Auctions and Bidding Games,” in Recent Advances in Game

Theory, Princeton University Press, no. 29 in Princeton Conference Series, 15–27.

Zermelo, E. (1913): “Über eine Anwendung der Mengenlehre auf die Theorie des
Schachspiels,” in Proceedings of the the Fifth Congress Mathematicians, Cam-
bridge University Press, 501–504.

16