Professional Documents
Culture Documents
Game1bb Nash Eqm Mixed-1
Game1bb Nash Eqm Mixed-1
In this note, we study the notion of mixed strategies, which randomize a player’s choice of
action and then define Nash equilibrium in mixed strategies (Nash, 1950).
H T
H +1, −1 −1, +1
T −1, +1 +1, −1
This game has no pure-strategy Nash equilibrium. However, each player i may randomize her
choice of action. She chooses actions H and T with probabilities αi (H) and αi (T ) respectively.
This randomization makes sense if and only if αi (H), αi (T ) ≥ 0 and αi (H) + αi (T ) = 1.
We will define a mixed-strategy version of Nash equilibrium, which we call mixed-strategy
Nash equilibrium or Nash equilibrium in mixed strategies. By extending the notion of Nash
equilibrium to mixed strategies, we will be able to find a (mixed-strategy) Nash equilibrium in
the matching-pennies game. □
1 Mixed Strategies
We introduce a notation: For any (non-empty) set X, we write ∆(X) for the set of all proba-
bility distributions over the set X.1
Definition 1. In a normal-form game G = 〈I, (Ai , ui )i∈I 〉, player i’s mixed strategy αi is
a probability distribution over the set Ai of her actions, and we write Ai = ∆(Ai ) for the set
of all her mixed strategies. Let α = (αi )i∈I ∈ A be a mixed strategy profile, and we write
!
A= i∈I Ai for the set of all mixed strategy profiles.
1
How should we define players’ payoffs when they use a mixed strategy profile α = (αi )i ?
Since each player randomizes her choice of actions on her own, it is natural to assume that
each player’s randomization is independent of any other player’s randomization. We then define
player i’s payoff function ui : A → R such that ui (α) is the expectation of her payoffs:2
# %
" $
ui (α) ≡ αi (ai ) ui (a). (1)
a∈A i∈I
ui (αi∗ , α−i
∗
) ≥ ui (αi , α−i
∗
), ∀αi ∈ Ai . (2)
2
For the sake of simplicity, we assume that the set I of players and the set Ai of each player i’s actions are
countable. It is straightforward to extend this notion to the uncountable case.
2
Box (Mixed Extension): Some references often define mixed-strategy Nash equilibrium
in a seemingly different but equivalent way. Here we mention these definitions.
As in the case of pure-strategy Nash equilibrium, Definitions 3 and 5 are equivalent to each
other. To see this, we note that inequality (2) says that strategy αi∗ is optimal for player i
when players −i play strategies α−i
∗ . Formally, α∗ ∈ argmax u (α , α∗ ) ≡ BR (α∗ ).
i αi i i −i i −i
We offer a memorable paraphrase of Nash equilibrium.
Example 2. Penalty kicks in soccer involve two players: Kicker (K) and Goalkeeper (G). As
shots by soccer professionals are fast, we assume that K and G must move simultaneously.
Since it is rare to kick at the middle of the goal, we assume that K kicks either to the left
(L) or to the right (R). We also assume that G defends either the left side (L) or the right
side (R). That is, their action spaces are AK = AG = {L, R}. At action profile a = (aK , aG ),
K’s payoff is probability π(a) that K makes a goal and G’s payoff is probability 1 − π(a) that
G saves.3 To specify the payoffs, we can use the real data of professional soccer games. For
example, we estimate the goal probability π(L, R) as the likelihood of the goal when K kicks
3
This kind of game is called a zero-sum game since the sum of players’ payoffs is zero: uK (a) + uG (a) = 0
for every a.
3
pG
1 BRG
0.4199 BRK
NE
O 0.3854 1 pK
K\G L R
L 58.30, 41.70 94.97, 5.03
R 92.92, 7.08 69.92, 30.08
This game has a unique Nash equilibrium. Since player i has two actions, we define his
mixed strategy by choosing probability pi of playing L. To find the Nash equilibrium, we find
K’s best response. Given probability pG that G plays L, K’s best response BRK is such that:4
&
'
'
'
' {1} if pG < 0.4199
(
BRK (pG ) = [0, 1] if pG = 0.4199
'
'
'
'
){0} if pG > 0.4199.
Similarly, we find G’s best response BRG . The two players’ best responses are illustrated in
Figure 1. The Nash equilibrium is, by Definition 5, the intersection of the two best responses.
We summarize the Nash equilibrium in Table 3 (row 1).
4 835
Here “0.4199” is an approximation of 1989
.
4
aG = L aG = R aK = L aK = R
theory 0.4199 0.5801 0.3854 0.6146
data 0.4231 0.5769 0.3998 0.6002
In Table 3, we compare the Nash equilibrium (row 1) with the real data of soccer professionals’
play (row 2). The Nash equilibrium prediction fits the data well!5,6 □
Coffee Break . Many studies examine and compare how people play a game in the field and
in the laboratory. Interestingly, many of them suggest that the field play is often different from
the laboratory play. For example, Walker & Wooders (2001) study the field play by tennis
professionals, while Chiappori et al. (2002) soccer professionals.7 More recent studies compare
professionals’ play in the field and their play in the laboratory (e.g., Palacios-Huerta & Volij,
2008; Levitt et al., 2010; Wooders, 2010). □
In a Nash equilibrium, each player is indifferent between all her actions that are played
with non-zero probabilities.
Proof. Suppose, by contradiction, that there exist ai , a′i ∈ Ai such that αi∗ (ai ), αi∗ (a′i ) > 0 but
ui (ai , α−i
∗ ) > u (a′ , α∗ ). Then, player i would increase her payoff by increasing the probability
i i −i
of playing ai . This contradicts the assumption that αi∗ is a best response to α−i
∗ . "
Example 3. We compute the Nash equilibrium, using Proposition 1. Let pG be the probability
that G plays aG = L. K’s payoff from playing aK = L is 58.30pG + 94.97(1 − pG ), while K’s
5
This result is robust regardless of whether a player is right- or left-footed. In general, a right-footed kicker
will find it easier to kick to his left than his right, and vice versa for a left-footed kicker.
6
Palacios-Huerta (2022) reports that Maradona’s play is consistent with the Nash equilibrium prediction,
using the data of his penalty kicks during his lifetime professional career.
7
These results are also robust regardless of whether a player is right-handed/footed or left-handed/footed.
5
payoff from playing aK = R is 92.92pG + 69.92(1 − pG ). By Proposition 1, it must be that at
the Nash equilibrium,
This equation yields pG = 0.4199, which coincides with the above result. Similarly, we can
compute the probability pK that K plays aK = L. □
Remark 4. You may be puzzled by the aspect of mixed-strategy Nash equilibria that each
player is indifferent between all actions to which he assigns non-zero probabilities, but he
randomizes them in order to make other players indifferent. In Example 2, G randomizes his
actions to make K indifferent between L and R, while K randomizes his actions to make G
indifferent between L and R. This puzzle is justified by Harsanyi’s Purification Theorem. See
also the discussion in Osborne & Rubinstein (1994, Section 3.2). □
Definition 8. A normal-form game G = 〈I, (Ai , ui )i 〉 is finite if I is finite and Ai is finite for
each i ∈ I, and is infinite otherwise.
For example, the games of Examples 1 and 2 are finite, while the Cournot game and
Bertrand game, covered in the previous notes, are infinite.
Theorem 1 (Nash’s Existence Theorem). Every finite normal-form game G has a mixed-
strategy Nash equilibrium.
Proof. Let BR : A → 2A be a correspondence such that for each α ∈ A, BR(α) = (BRi (α−i ))i .
In this view, we obtain a fixed-point interpretation of mixed-strategy Nash equilibrium:
This proof is reduced to the existence of a fixed point of BR. To this end, it suffices to show
that BR satisfies all the conditions for Kakutani’s Fixed-Point Theorem 2 for BR. We show
them one by one:
1. Ai is non-empty, compact, and convex.
2. BRi is non-empty-valued.
Since ui is continuous in αi and Ai is compact, it follows that for any α−i , BRi (α−i ) ≡
argmaxαi ui (αi , ) is non-empty.
6
Box (Kakutani’s Fixed-Point Theorem):
Remark 6. The closed graph property is, in general, not equivalent to upper hemiconti-
nuity. However, these are equivalent if Y is compact and F is closed-valued. □
3. BRi is convex-valued.
Fix any α−i ∈ A−i , and take any αi , αi′ ∈ BRi (α−i ). By Proposition 1, all ai ∈ supp(αi )
and all a′i ∈ supp(αi′ ) give the same payoff: ui (ai , α−i ) = ui (a′i , α−i ).8 Hence, player i is
indifferent to any randomization over supp(αi ) ∪ supp(αi′ ).
4. BRi has a closed graph.
Let {(αk , α̂k )}k ⊂ Gr(BR) be a sequence such that limk αk = α∗ and limk α̂k = α̂∗ . For
each k ∈ N,
ui (α̂ik , α−i
k
) ≥ ui (αi′ , α−i
k
), ∀αi′ ∈ Ai .
Hence,
ui (α̂i∗ , α−i
∗
) = lim ui (α̂ik , α−i
k
) ≥ lim ui (αi′ , α−i
k
) = ui (αi′ , α−i
∗
), ∀αi′ ∈ Ai . "
k→∞ k→∞
8
For a finite set X, the support of a distribution µ ∈ ∆(X) is the set, supp(µ) = {x ∈ X : µ(x) > 0}.
7
Theorem 3. Suppose that a finite-player normal-form game G = 〈I, (Ai , ui )i 〉 is such that:
1. Ai is a compact convex subset of a Euclidean space for each i ∈ I.
2. ui is continuous for each i ∈ I.
3. ui is quasi-concave in ai for each i ∈ I.
Then, game G has a pure-strategy Nash equilibrium.
Example 4. Theorem 3 guarantees that the following Cournot game G = 〈I, (Ai , πi )i∈I 〉 has
a pure-strategy Nash equilibrium:
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
We cannot apply Theorem 3 to this game as is, because Ai is not compact. However, we can
avoid this issue by considering the following modified game G′ = 〈I, (A′i , πi′ )i∈I 〉:
1. I = {1, 2} is the set of firms.
2. A′i = [0, 1] is the set of firm i’s strategies.
3. πi′ : A′ → R is firm i’s payoff function (with the set A′ = A′1 × A′2 ) defined by
This game satisfies all the conditions for Theorem 3, and thus we have a pure-strategy Nash
equilibrium q ∗ = (q1∗ , q2∗ ) ∈ A′ .
What we want to prove is that game G has a pure-strategy Nash equilibrium. To show this,
it suffices to show that this Nash equilibrium q ∗ (in game G′ ) remains to be a Nash equilibrium
in game G. To see this, consider firm i’s deviation to any qi > 1.9 This deviation yields profit
0 (since qi + q−i
∗ > 1 and thus the price is zero), whereas strategy q ∗ gives profit π (q ∗ ) ≥ 0.
i i
Since this deviation is never profitable, we can say that game G has the pure-strategy Nash
equilibrium q ∗ . □
8
To show that a game G = 〈I, (Ai , ui )i 〉 has a Nash equilibrium with property #, the
following approach may be useful:
1. Consider a “restricted” game G# = 〈I, (A# #
i , ui )i 〉 such that:
• A#
i is a subset of player i’s strategies that satisfy property #.
!
• u# # #
i : A → R is the restricted payoff function, with domain A =
#
i Ai .
2. Show that game G# has a Nash equilibrium a# ∈ A# (which satisfies property #).
• The existence is established possibly by applying a fixed-point theorem.
3. Show that player i has no profitable deviation in the set Ai , given strategies a#
−i .
• This step shows that a# is a Nash equilibrium in the original game G.
Corollary 1. Suppose that a finite-player normal-form game G = 〈I, (Ai , ui )i 〉 is such that:
1. Ai is a compact convex subset of a Euclidean space for each i ∈ I.10
2. ui is continuous for each i ∈ I.
Then, game G has a mixed-strategy Nash equilibrium.
References
Chiappori, P.-A., Levitt, S., & Groseclose, T. (2002). Testing mixed-strategy equilibria when
players are heterogeneous: The case of penalty kicks in soccer. American Economic Review,
92 (4), 1138–1151.
Levitt, S. D., List, J. A., & Reiley, D. H. (2010). What happens in the field stays in the field:
Exploring whether professionals play minimax in laboratory experiments. Econometrica,
78 (4), 1413–1434.
Nash, J. F. (1950). Equilibrium points in n-person games. Proceedings of the National Academy
of Sciences, 36 (1), 48–49.
Osborne, M. J., & Rubinstein, A. (1994). A course in game theory. MIT Press.
10
A subset of a Euclidean space is compact if and only if it is closed and bounded.
9
Palacios-Huerta, I. (2003). Professionals play minimax. Review of Economic Studies, 70 (2),
395–415.
Palacios-Huerta, I., & Volij, O. (2008). Experientia docet: Professionals play minimax in
laboratory experiments. Econometrica, 76 (1), 71–115.
Walker, M., & Wooders, J. (2001). Minimax play at Wimbledon. American Economic Review,
91 (5), 1521–1538.
Wooders, J. (2010). Does experience teach? Professionals and minimax play in the lab.
Econometrica, 78 (3), 1143–1154.
10