Lecture VI: Existence of Nash equilibrium

Markus M. Mobius
February 26, 2004
Gibbons, sections 1.3B
Osborne, chapter 4

Nashs Existence Theorem

When we introduced the notion of Nash equilibrium the idea was to come
up with a solution concept which is stronger than IDSDS. Today we show
that NE is not too strong in the sense that it guarantees the existence of at
least one mixed Nash equilibrium in most games (for sure in all finite games).
This is reassuring because it tells that there is at least one way to play most
games.1
Lets start by stating the main theorem we will prove:
Theorem 1 (Nash Existence)Every finite strategic-form game has a mixedstrategy Nash equilibrium.
Many game theorists therefore regard the set of NE for this reason as the
lower bound for the set of reasonably solution concept. A lot of research has
gone into refining the notion of NE in order to retain the existence result
but get more precise predictions in games with multiple equilibria (such as
coordination games).
However, we have already discussed games which are solvable by IDSDS
and hence have a unique Nash equilibrium as well (for example, the two
thirds of the average game), but subjects in an experiment will not follow
those equilibrium prescription. Therefore, if we want to describe and predict
1

Note, that a pure Nash equilibrium is a (degenerate) mixed equilibrium, too.

the behavior of real-world people rather than come up with an explanation

of how they should play a game, then the notion of NE and even even IDSDS
can be too restricting.
Behavioral game theory has tried to weaken the joint assumptions of
rationality and common knowledge in order to come up with better theories
of how real people play real games. Anyone interested should take David
Laibsons course next year.
Despite these reservation about Nash equilibrium it is still a very useful
benchmark and a starting point for any game analysis.
In the following we will go through three proofs of the Existence Theorem
using various levels of mathematical sophistication:
existence in 2 2 games using elementary techniques
existence in 2 2 games using a fixed point approach
general existence theorem in finite games
You are only required to understand the simplest approach. The rest is for
the intellectually curious.

Nash Existence in 2 2 Games

Let us consider the simple 2 2 game which we discussed in the previous

lecture on mixed Nash equilibria:

1,1

0,4

0,2

2,1

We next draw the best-response curves of both players. Recall that player
1s strategy can be represented by a single number such that 1 = U +
(1 )D while player 2s strategy is 2 = L + (1 )R.
2

Lets find the best-response of player 2 to player 1 playing strategy :

u2 (L, U + (1 )D) = 2
u2 (R, U + (1 )D) = 1 + 3

(1)

Therefore, player 2 will strictly prefer strategy L iff 2 > 1 + 3 which

implies < 14 . The best-response correspondence of player 2 is therefore:

if < 14
1
[0, 1] if = 14
BR2 () =
(2)

0
if > 14
We can similarly find the best-response correspondence of player 1:

if < 23
0
[0, 1] if = 23
BR1 () =

1
if > 23

(3)

We draw both best-response correspondences in a single graph (the graph is

in color - so looking at it on the computer screen might help you):

BR2()

BR1()

2/3

1/4

We immediately see, that both correspondences intersect in the single point

= 14 and = 23 which is therefore the unique (mixed) Nash equilibrium of
the game.
3

Whats useful about this approach is that it generalizes to a proof that

any two by two game has at least one Nash equilibriu, i.e. its two best
response correspondences have to intersect in at least one point.
An informal argument runs as follows:
1. The best response correspondence for player 2 maps each into at
least one . The graph of the correspondence connects the left and
right side of the square [0, 1] [0, 1]. This connection is continuous
- the only discontinuity could happen when player 2s best response
switches from L to R or vice versa at some . But at this switching
point player 2 has to be exactly indifferent between both strategies hence the graph has the value BR2 ( ) = [0, 1] at this point and there
cannot be a discontinuity. Note, that this is precisely why we need
mixed strategies - with pure strategies the BR graph would generally
be discontinuous at some point.
2. By an analogous argument the BR graph of player 1 connects the upper
and lower side of the square [0, 1] [0, 1].
3. Two lines which connect the left/right side and the upper/lower side
of the square respectively have to intersect in at least one point. Hence
each 2 by 2 game has a mixed Nash equilibrium.

Nash Existence in 2 2 Games using Fixed

Point Argument

There is a different way to prove existence of NE on 2 2 games. The

advantage of this new approach is that it generalizes easily to general finite
games.
Consider any strategy profile (U + (1 )D, L + (1 )R) represented
by the point (, ) inside the square [0, 1][0, 1]. Now imagine the following:
player 1 assumes that player 2 follows strategy and player 2 assumes that
player 1 follows strategy . What should they do? They should play their
BR to their beliefs - i.e. player 1 should play BR1 () and player 2 should
play BR2 (). So we can imagine that the strategy profile (, ) is mapped
onto (BR1 (), BR2 ()). This would describe the actual play of both players
if their beliefs would be summarizes by (, ). We can therefore define a
4

giant correspondence BR : [0, 1] [0, 1] [0, 1] [0, 1] in the following way:

BR(, ) = BR1 () BR2 ()

(4)

The following figure illustrates the properties of the combined best-response

map BR:

BR2()

BR1()

(2,2)
2/3

(1,1)

(BR1(1),BR2(1))
1/4

1 (BR1(2),BR2(2))

The neat fact about BR is that the Nash equilibria are precisely the
fixed points of BR, i.e. BR( ). In other words, if players have beliefs
then should also be a best response by them. The next lemma follows
directly from the definition of mixed Nash equilibrium:
Lemma 1 A mixed strategy profile is a Nash equilibrium if and only if it
is a fixed point of the BR correspondence, i.e. BR ( ).
We therefore look precisely for the fixed points of the correspondence
BR which maps the square [0, 1] [0, 1] onto itself. There is well developed
mathematical theory for these types of maps which we utilize to prove Nash
existence (i.e. that BR has at least one fixed point).

3.1

Kakutanis Fixed Point Theorem

The key result we need is Kakutanis fixed point theorem. You might have
used Browers fixed point theorem in some mathematics class. This is not
5

sufficient for proving the existence of nash equilibria because it only applies
to functions but not to correspondences.
Theorem 2 Kakutani A correspondence r : X X has a fixed point
x X such that x r (x) if
1. X is a compact, convex and non-empty subset of <n .
2. r (x) is non-empty for all x.
3. r (x) is convex for all x.
4. r has a closed graph.
There are a few concepts in this definition which have to be defined:
Convex Set: A set A <n is convex if for any two points x, y A the
straight line connecting these two points lies inside the set as well. Formally,
x + (1 ) y A for all [0, 1].
Closed Set: A set A <n is closed if for any converging sequence

{xn }
n=1 with xn x as n we have x A. Closed intervals such
as [0, 1] are closed sets but open or half-open intervals are not. For example
(0, 1] cannot be closed because the sequence n1 converges to 0 which is not in
the set.
Compact Set: A set A <n is compact if it is both closed and bounded.
For example, the set [0, 1] is compact but the set [0, ) is only closed but
unbounded, and hence not compact.
Graph: The graph of a correspondence r : X Y is the set {(x, y) |y r (x)}.
If r is a real function the graph is simply the plot of the function.
Closed Graph: A correspondence has a closed graph if the graph of the
correspondence is a closed set. Formally, this implies that for a sequence of

point on the graph {(xn , yn )}

n=1 such that xn x and yn y as n

2
we have y r (x ).
It is useful to understand exactly why we need each of the conditions in
Kakutanis fixed point theorem to be fulfilled. We discuss the conditions by
looking correspondences on the real line, i.e. r : < <. In this case, a fixed
point simply lies on the intersection between the graph of the correspondence
and the diagonal y = x. Hence Kakutanis fixed point theorem tells us that
2

If the correspondence is a function then the closed graph requirement is equivalent to

assuming that the function is continuous. Its easy to see that a continuous function has
a closed graph. For the reverse, youll need Baires category theorem.

a correspondence r : [0, 1] [0, 1] which fulfills the conditions above always

intersects with the diagonal.
3.1.1

Kakutani Condition I: X is compact, convex and non-empty.

Assume X is not compact because it is not closed - for example X = (0, 1).
Now consider the correspondence r(x) = x2 which maps X into X. However,
it has no fixed point. Now consider X non-compact because it is unbounded
such as X = [0, ) and consider the correspondence r(x) = 1 + x which
maps X into X but has again no fixed point.
If X is empty there is clearly no fixed point. For convexity of X look at
the example X = [0, 13 ] [ 32 , 1] which is not convex because the set has a hole.
Now consider the following correspondence (see figure below):
3
if x [0, 31 ]
4
(5)
r(x) =
1
if x [ 23 , 1]
4
This correspondence maps X into X but has no fixed point again.

3/4

1/4

1/3

2/3

From now on we focus on correspondences r : [0, 1] [0, 1] - note that [0, 1]

is closed and bounded and hence compact, and is also convex.
7

3.1.2

Kakutani Condition II: r(x) is non-empty.

If r(x) could be empty we could define

such as the following:
3
4 if
if
r(x) =
1
if
4

a correspondence r : [0, 1] [0, 1]

x [0, 13 ]
x [ 31 , 32 ]
x [ 23 , 1]

(6)

As before, this correspondence has no fixed point because of the hole in the
middle.
3.1.3

Kakutani Condition III: r(x) is convex.

If r(x) is not convex, then the graph does not have to have a fixed point as
the following example of a correspondence r : [0, 1] [0, 1] shows:

1 if x < 12
0, 13 23 , 1 if x = 21
r(x) =
(7)

0
if x > 12

1/2

The graph is non-convex because r( 12 ) is not convex. It also does not have a
fixed point.
8

3.1.4

Kakutani Condition IV: r(x) has a closed graph.

This condition ensures that the graph cannot have holes. Consider the following correspondence r : [0, 1] [0, 1] which fulfills all conditions of Kakutani
except (4):
1
if x < 12
2

1 1
,
if x = 21
r(x) =
(8)
14 2
1
if
x
>
4
2

1/2

Note, that r( 12 ) is the convex set 14 , 12 but that this set is not closed. Hence
the graph is not closed. For example, consider the sequence xn = 21 and
1
yn = 12 n+2
for n 1. Clearly, we have yn r(xn ). However, xn x = 12
/ r(x ). Hence the graph is not closed.
and yn y = 12 but y

3.2

Applying Kakutani

We now apply Kakutani to prove that 2 2 games have a Nash equilibrium,

i.e. the giant best-response correspondence BR has a fixed point. We denote
the strategies of player 1 with U and D and the strategies of player 2 with
L and R.
9

We have to check (a) that BR is a map from some compact and convex
set X into itself, and (b) conditions (1) to (4) of Kakutani.
First note, that BR : [0, 1] [0, 1] [0, 1] [0, 1]. The square X =
[0, 1] [0, 1] is convex and compact because it is bounded and closed.
Now check condition (2) of Kakutani - BR() is non-empty. This is
true if BR1 (2 ) and BR2 (1 ) are non-empty. Lets prove it for BR1 the proof for BR2 is analogous. Player 1 will get the following payoff
u1, () from playing strategy if the other player plays :
u1, () = u1 (U, L) + (1 )u1 (U, R) +
+ (1 )u1 (D, L) + (1 )(1 )u1 (D, R)

(9)

The function u1, is continuous in . We also know that [0, 1] which

is a closed interval. Therefore, we know that the continuous function
u1, reaches its maximum over that interval (standard min-max result
from real analysis - continuous functions reach their minimum and maximum over closed intervals). Hence there is at least one best response
which maximizes player 1s payoff.
Condition (3) requires that if player 1 has tow best responses 1 U +
(1 1 )D and 2 U + (1 2 )D to player 2 playing L + (1 )R then
the strategy where player 1 chooses U with probability 1 + (1 )2
for some 0 < < 1 is also a best response (i.e. BR1 () is convex).
But since both the 1 and the 2 strategy are best responses of player
1 to the same strategy of player 2 they also have to provide the same
payoffs to player 1. But this implies that if player 1 plays strategy 1
with probability and 2 with probability 1 she will get exactly
the same payoff as well. Hence the strategy where she plays U with
probability 1 +(1)2 is also a best response and her best response
set BR1 () is convex.
The final condition (4) requires that BR has a closed graph. To show
this consider a sequence n = (n , n ) of (mixed) strategy profiles and

n = (
n , n ) BR ( n ). Both sequences are assumed to converge to
= ( , ) and
= (
, ), respectively. We now want to show
that
BR () to prove that BR has a closed graph.
We know that for player 1, for example, we have
u1 (
n , n ) u1 (0 , n )
10

for any 0 [0, 1]. Note, that the utility function is continuous in both
arguments because it is linear in and . Therefore, we can take the
limit on both sides while preserving the inequality sign:
u1 (
, ) u2 (0 , )
for all 0 [0, 1]. This shows that BR1 () and therefore

BR ( ). Hence the graph of the BR correspondence is closed.

Therefore, all four Kakutani conditions apply and the giant best-response
correspondence BR has a fixed point, and each 2 2 game has a Nash
equilibrium.

Nash Existence Proof for General Finite

Case

Using the fixed point method it is now relatively easy to extend the proof
for the 2 2 case to general finite games.The biggest difference is that we
cannot represent a mixed strategy any longer with a single number such as .
If player 1 has three pure strategies A1 ,A2 and A3 , for example, then his set
of mixed strategies is represented by two probabilities - for example, (1 , 2 )
which are the probabilities that A1 and A2 are chosen. The set of admissible
1 and 2 is described by:
1 = {(1 , 2 )|0 1 , 2 1 and 1 + 2 1}

(10)

The definition of the set of mixed strategies can be straightforwardly extended to games where player 1 has a strategy set consisting of n pure
strategies A1 ,..,An . Then we need n 1 probabilities 1 ,..,n1 such that:
1 = {(1 , .., n1 )|0 1 , .., n1 1 and 1 + .. + n1 1}

(11)

So instead of representing strategies on the unit interval [0, 1] we have to

represent as elements of the simplex 1 .
Lemma 2 The set 1 is compact and convex.
Proof: It is clearly convex - if you mix between two mixed strategies you get
another mixed strategy. The set is also compact because it is bounded
i
)
(all |i | 1) and closed. To see closedness take a sequence (1i , .., n1

of
elements of 1 which converges to (1 , ..). Then we have i 0 and
Pn1

i=1 i 1 because the limit preserves weak inequalities. QED

11

We can now check that all conditions of Kakutani are fulfilled in the general finite case. Checking them is almost 1-1 identical to checking Kakutanis
condition for 2 2 games.
Condition 1: The individual mixed strategy sets i are clearly nonempty because every player has at least one strategy. Since i is compact
= 1 ...I is also compact. Hence the BR correspondence BR :
acts on a compact and convex non-empty set.
Condition 2: For each player i we can calculate his utiltiy ui,i (i ) for
i i . Since i is compact and ui,i is continuous the set of payoffs is also
compact and hence has a maximum. Therefore, BRi (i ) is non-empty.
Condition 3: Assume that i1 and i2 are both BR of player i to i .
Both strategies have to give player i equal payoffs then and any linear combination of these two strategies has to be a BR for player i, too.
Condition 4: Assume that n is a sequence of strategy profiles and
n

BR ( n ). Both sequences converge to and

, respectively. We
know that for each player i we have
n n

n
ui
i , i ui i0 , i
for all i0 i . Note, that the utility function is continuous in both arguments
because it is linear.3 Therefore, we can take the limit on both sides while
preserving the inequality sign:

ui
i , i ui i0 , i

for all i0 i . This shows that
i BRi i
and therefore
BR ( ).
Hence the graph of the BR correspondence is closed.
So Kakutanis theorem applies and the giant best-response map BR has
a fixed point.

It is crucial here that the set of pure strategies is finite.

12