Existence of Equilibria

for Strategic Games

LUISS

1/34
(LUISS) Existence of Equilibria for Strategic Games 1 / 34
 Games & Strategies Existence of Equilibria

Topics

Mixed extension of finite games

Best response maps
Nash existence theorem
Examples of n-person games
Hotelling game
Cournot competition
Braess paradox
El Farol bar
Auctions

2/34
(LUISS) Existence of Equilibria for Strategic Games 2 / 34
 Games & Strategies Mixed strategies

Nash equilibrium in pure strategies might not exist

Consider the game  

(4, 0) (3, 1)
(3, 5) (5, 0)
There is no equilibrium... in pure strategies.

A player cannot use the same strategy all the time; this would be observed and
the other player could take advantage from this.
It makes sense to play strategies according to some probability scheme.
But these probabilities must be chosen strategically!

3/34
(LUISS) Existence of Equilibria for Strategic Games 3 / 34
 Games & Strategies Mixed strategies

Simplexes and mixed strategies

Definition
Let A be a finite strategy set with d elements (also called pure strategies). The
set of mixed strategies over the set A is the d-dimensional simplex

ΣA = {(xa )a∈A ∈ RA : xa ≥ 0, a∈A xa = 1}.

P

A vector x = (xa )a∈A ∈ ΣA defines a probability distribution on the set A with

xa = P(playing the pure strategy a)
4/34
(LUISS) Existence of Equilibria for Strategic Games 4 / 34
 Games & Strategies Mixed strategies

Mixed extension of 2-player games

Consider a 2-person game with strategy sets I = {1, . . . , n} and J = {1, . . . , m},
and payoff matrices (A, B). In the mixed extension of the game player 1 chooses a
probability distribution x ∈ ΣI and player 2 a probability distribution y ∈ ΣJ .

The probability of observing the outcome ij is the product xi yj and then the
expected payoffs for both players are respectively:
Pn Pm
Player 1: f (x, y ) = i=1 j=1 aij xi yj = x t Ay
Pn Pm
Player 2: g (x, y ) = i=1 j=1 bij xi yj = x t By

5/34
(LUISS) Existence of Equilibria for Strategic Games 5 / 34
 Games & Strategies Mixed strategies

Mixed extension of 2-player games

Alternatively:
The expected payoff for player 1 when playing the pure strategy i ∈ I against
the mixed strategy y ∈ ΣJ of player 2 is
Pm
ui (y ) = j=1 aij yj

and then Pn
f (x, y ) = i=1 xi ui (y )

The expected payoff for player 2 when playing the pure strategy j ∈ J against
the mixed strategy x ∈ ΣI of player 1 is
Pn
vj (x) = i=1 bij xi

and then Pm
g (x, y ) = j=1 yj vj (x)

6/34
(LUISS) Existence of Equilibria for Strategic Games 6 / 34
 Games & Strategies Best response maps

Best responses

Nash equilibrium relies on the assumption that players maximize their payoff with
respect to their own variable, taking for granted the choice of the other player.
Pn Pn
Player 1: max i=1 xi ui (y ) ⇒ BR1 (y ) = Argmax i=1 xi ui (y )
x∈ΣI x∈ΣI
Pm Pm
Player 2: max j=1 yj vj (x) ⇒ BR2 (x) = Argmax j=1 yj vj (x)
y ∈ΣJ y ∈ΣJ

BR1 (y ) is the set of all x’s that maximize f (·, y ) for a fixed y ∈ ΣJ .
BR2 (x) is the set of all y ’s that maximize g (x, ·) for a fixed x ∈ ΣI .

Hence, a Nash equilibrium in mixed strategies is a pair (x̄, ȳ ) ∈ ΣI × ΣJ such that


x̄ ∈ BR1 (ȳ )
ȳ ∈ BR2 (x̄)

7/34
(LUISS) Existence of Equilibria for Strategic Games 7 / 34
 Games & Strategies Best response maps

Example: finding Nash equilibria

Consider the game  

(4, 0) (3, 1)
(3, 5) (5, 0)
Player 1 selects Top or Bottom with probabilities (p, 1 − p).
Player 2 chooses Left or Right with probabilities (q, 1 − q).
Then their expected payoffs are respectively

f (p, q) = p uT (q) + (1−p) uB (q)

= p [4q + 3(1−q)] + (1−p) [3q + 5(1−q)]

g (p, q) = q vL (p) + (1−q) vR (p)

= q [5(1−p)] + (1−q) [1p]

8/34
(LUISS) Existence of Equilibria for Strategic Games 8 / 34
 Games & Strategies Best response maps

Example: finding Nash equilibria

BR1 (q) = Argmax p uT (q) + (1−p) uB (q)

p∈[0,1]

uT (q) = 4q + 3(1−q) ; uB (q) = 3q + 5(1−q)

2
q> 3 ⇒ uT (q) > uB (q) ⇒ it is optimal for player 1 to choose p = 1.
2
q< 3 ⇒ uT (q) < uB (q) ⇒ it is optimal for player 1 to choose p = 0.
2
q= 3 ⇒ uT (q) = uB (q) ⇒ every p ∈ [0, 1] is equally good for player 1.

9/34
(LUISS) Existence of Equilibria for Strategic Games 9 / 34
 Games & Strategies Best response maps

Example: finding Nash equilibria

BR2 (p) = Argmax q vL (p) + (1−q) vR (p)

q∈[0,1]

vL (p) = 5(1−p) ; vR (p) = p

5
p< 6 ⇒ vL (p) > vR (p) ⇒ it is optimal for player 2 to choose q = 1.
5
p> 6 ⇒ vL (p) < vR (p) ⇒ it is optimal for player 2 to choose q = 0.
5
p= 6 ⇒ vL (p) = vR (p) ⇒ every q ∈ [0, 1] is equally good for player 2.

10/34
(LUISS) Existence of Equilibria for Strategic Games 10 / 34
 Games & Strategies Best response maps

Example: finding Nash equilibria

2

 {1} if q > 3
2
BR1 (q) = {0} if q < 3
2
[0, 1] if q =

3
5

 {1} if p < 6
5
BR2 (p) = {0} if p > 6
5
[0, 1] if p =

6
5 2
Unique Nash equilibrium: p̄ = 6 and q̄ = 3

2
3

5 1 p
6
11/34
(LUISS) Existence of Equilibria for Strategic Games 11 / 34
 Games & Strategies Nash theorem

General strategic games

Consider anQn-player game with strategy sets Xi and payoffs fi : X → R where as

n
usual X = j=1 Xj is the set of strategy profiles.

Each player i = 1, . . . , n maximizes her payoff with respect to her ownQvariable

xi ∈ Xi while taking for granted the choice of the other players x-i ∈ j6=i Xj .

Define the best response map BRi : X-i → Xi as

BRi (x-i ) = {xi ∈ Xi : fi (xi , x-i ) ≥ fi (zi , x-i ) ∀zi ∈ Xi }

which associates to each possible strategies x-i of the other players the set of xi ’s
that maximize the payoff fi (·, x-i ).
Then: (x̄i )ni=1 is a Nash equilibrium if and only if for each player i = 1, . . . , n we
have x̄i ∈ BRi (x̄-i ).

12/34
(LUISS) Existence of Equilibria for Strategic Games 12 / 34
 Games & Strategies Nash theorem

The Nash theorem

Theorem
Qn game with strategy sets Xi and payoff functions fi : X → R
Given a n-player
where X = i=1 Xi . Suppose:
each Xi is a closed bounded convex subset in a finite dimensional space Rdi
each fi : X → R is continuous
xi 7→ fi (xi , x-i ) is a (quasi) concave function for each fixed x-i ∈ X-i
Then there exists at least one Nash equilibrium.

13/34
(LUISS) Existence of Equilibria for Strategic Games 13 / 34
 Games & Strategies Nash theorem

A simple 2-player game

Let X = Y = [0, 10] and let the payoffs of the players be

f (x, y ) = −x 2 − 2xy + 12x + 1

g (x, y ) = −y 2 + 2xy + 8y + 7

The common strategy set is a closed interval, thus a closed convex bounded set.
The functions are continuous, and concave in one variable when the other is fixed.
We get BR1 (y ) = {max(0, 6 − y )} and BR2 (x) = {min(10, 4 + x)}.
To find an equilibrium we solve
(
x = max(0, 6 − y )
y = min(10, 4 + x)

with unique solution (x̄, ȳ ) = (1, 5).

14/34
(LUISS) Existence of Equilibria for Strategic Games 14 / 34
 Games & Strategies Existence of equilibria in mixed strategies

Mixed equilibria for 2-player finite games

Corollary
Every 2-player finite game has at least one Nash equilibrium in mixed strategies.

In this case
ΣI ⊆ RI and ΣJ ⊆ RJ are simplexes, hence closed bounded and convex
f (x, y ) = x t Ay and g (x, y ) = x t By are jointly continuous w.r.t (x, y )
Pn
f (x, y ) = Pi=1 xi ui (y ) is linear with respect to x (for fixed y ) and
m
g (x, y ) = j=1 yi vj (x) is linear with respect to y (for fixed x)
and thus the assumptions of Nash theorem are fulfilled.

15/34
(LUISS) Existence of Equilibria for Strategic Games 15 / 34
 Games & Strategies Existence of equilibria in mixed strategies

Mixed equilibria for n-player finite games

Consider an n-person finite game with strategy sets Ai and payoffs fi (a1 , . . . , an ).
In the mixed extension each player i chooses P a probability distribution x i ∈ ΣAi ,
i i
that is to say, xai ≥ 0 for all ai ∈ Ai and ai ∈Ai xai = 1.
Qn
Denote A = i=1 Ai the set of pure strategyQprofiles. The probability of observing
n
an outcome (a1 , . . . , an ) ∈ A is the product i=1 xai i and the expected payoffs are:

X n
Y X
f¯i (x 1 , . . . , x n ) = fi (a1 , . . . , an ) xaj j = xai i ui (ai , x −i )
(a1 ,...,an )∈A j=1 ai ∈Ai

X Y
ui (ai , x −i ) = fi (a1 , . . . , an ) xaj j
aj ∈Aj ,j6=i j6=i

Corollary
Every n-player finite game has at least one Nash equilibrium in mixed strategies.

16/34
(LUISS) Existence of Equilibria for Strategic Games 16 / 34
 Games & Strategies Computing mixed equilibria

The statement of the problem

For every fixed mixed strategy y of the second player, the utility of the first player
is, the following function, where x must be considered as a variable and y as a
parameter

Pn
f (x, y ) = i=1 xi ui (y ).

Thus the first player faces the problem of maximizing f (·, y ), a linear function for
every fixed y , over the simplex of his mixed strategies.
How can this be done?
Let us see it in general.

17/34
(LUISS) Existence of Equilibria for Strategic Games 17 / 34
 Games & Strategies Computing mixed equilibria

Maximizing over a simplex

Consider the problem of maximizing, over the simplex in R4 , the weighted average

maximize 15x1 + 18x2 + 23x3 + 23x4

subject to:
x1 + x2 + x3 + x4 = 1
x1 , . . . , x4 ≥ 0

Clearly the optimal value is 23 and is attained by putting all the weight on the
variables x3 and x4 , since they have larger coefficients.
In particular (0, 0, 1, 0) and (0, 0, 0, 1) are optimal, as well as (0, 0, 12 , 21 ).
In general the optimal solutions are all vectors (x1 , x2 , x3 , x4 ) with x1 = x2 = 0 and
x3 , x4 ≥ 0 with x3 + x4 = 1.

18/34
(LUISS) Existence of Equilibria for Strategic Games 18 / 34
 Games & Strategies Computing mixed equilibria

Maximizing over a simplex

Let us go back and consider a player with a finite set A of pure strategies. Let
ua = ua (x-i ) be the expected payoff when playing action a ∈ A given that the
other player(s) use mixed strategies x-i . Then the best response requires to solve
X
max xa ua
x∈ΣA
a∈A
P
In order to render the weighted average a∈A xa ua as large as possible one simply
has to put all the weight on the variables with largest coefficients ua .
Thus, setting v = maxa∈A ua we actually have
X
max xa ua = v
x∈ΣA
a∈A

and x ∈ ΣA is a best response (optimal solution) if and only if

(∀a ∈ A) xa > 0 ⇒ ua = v .

Note that there is always a best response in pure strategies: choose any a with
maximal ua and set xa = 1 and xa0 = 0 for a0 6= a. 19/34
(LUISS) Existence of Equilibria for Strategic Games 19 / 34
 Games & Strategies Computing mixed equilibria

Maximizing over a simplex

It follows that, given y a mixed strategy of the second player, the first one must
look at the rows, i.e. the pure strategies, which give him the highest expected
utility. So these pure strategies play a crucial role in finding Nash equilibria.

Definition
For a non-negative vector (x1 , . . . , xn ) ≥ 0 we define its support as the set of
indexes of its strictly positive entries, that is

spt(x) = {i : xi > 0}

The support of a mixed strategy then is the set of the pure strategies played with
positive probability.

20/34
(LUISS) Existence of Equilibria for Strategic Games 20 / 34
 Games & Strategies Computing mixed equilibria

Mixed equilibria in 2-player finite games

In a 2-player finite game, the pair (x̄, ȳ ) ∈ ΣI × ΣJ is a Nash equilibrium in mixed

strategies iff there exists v , w ∈ R such that
 Pm
j=1 aij ȳj = v for all i ∈ spt(x̄)
Pm
j=1 aij ȳj ≤ v for all i 6∈ spt(x̄)
 Pn
Pni=1 bij x̄i = w for all j ∈ spt(ȳ )
i=1 bij x̄i ≤ w for all j 6∈ spt(ȳ )

“Linear” system of equations and inequalities in n + m + 2 unknowns: x̄i , ȳj , v , w

In words: Player 1 should play with positive probability only the rows with
maximal expected payoff ui (ȳ ) = v , and symmetrically Player 2 should assign
positive probability only to the columns with maximal payoff vj (x̄) = w .
v = equilibrium payoff for Player 1
w = equilibrium payoff for Player 2

21/34
(LUISS) Existence of Equilibria for Strategic Games 21 / 34
 Games & Strategies Computing mixed equilibria

Nash equilibria with given support

When finding a Nash equilibrium, a system of inequalities and equalities like that
one above must be solved.
Where are the difficulties in solving such a system?
The key fact is that the support of a (potential) Nash equilibrium is not a priori
known.
So the above system can be easily used only when one looks for a Nash
equilibrium with given support.

22/34
(LUISS) Existence of Equilibria for Strategic Games 22 / 34
 Games & Strategies Computing mixed equilibria

Full support
In particular, the above system of equalities/inequalities simplifies and becomes a
linear system of equalities if one looks for fully mixed (meaning that all strategies
are played with positive probability) Nash equilibria.

Suppose (x̄, ȳ ) is such a Nash equilibrium profile. Then it holds that

ai1 ȳ1 + ai2 ȳ2 + · · · + aim ȳm = ak1 ȳ1 + ak2 ȳ2 + · · · + akm ȳm

for all i, k = 1, . . . n, and similarly

b1r x̄1 + b2r x̄2 + · · · + bnr x̄n = b1s x̄1 + b2s x̄2 + · · · + bns x̄n
for all r , s = 1, . . . m with the further conditions

n
X m
X
x̄j , ȳj ≥ 0, x̄i = 1, ȳj = 1
i=1 j=1

In this case we speak about Indifference principle. 23/34

(LUISS) Existence of Equilibria for Strategic Games 23 / 34
 Games & Strategies Computing mixed equilibria

Brute force algorithm

Finding Nash equilibria in general is a more complicated issue:
1 Guess the supports of the equilibria spt(x̄) and spt(ȳ )
2 Ignore the inequalities and find x, y , v , w by solving the linear system of
n + m + 2 equations
 Pn
 Pi=1 xi = 1
m
j=1 aij yj = v for all i ∈ spt(x̄)
xi = 0 for all i 6∈ spt(x̄)


 Pm
 Pj=1 yj = 1
n
i=1 bij xi = w for all j ∈ spt(ȳ )
yj = 0 for all j 6∈ spt(ȳ )


3 Check whether the Pmignored inequalities

Pnare satisfied.
If xi ≥ 0, yj ≥ 0, j=1 aij yj ≤ v and i=1 bij xi ≤ w then Stop: we have
found a mixed equilibrium. Otherwise, go back to step 1 and try another
guess of the supports.
24/34
(LUISS) Existence of Equilibria for Strategic Games 24 / 34
 Games & Strategies Computing mixed equilibria

Lemke-Howson Algorithm

Enumerating all the possible supports in the brute force algorithm quickly becomes
computationally prohibitive: there are potentially (2n − 1)(2m − 1) options!
For n × n games the number of combinations grow very quickly
n # of potential supports
2 9
3 49
4 225
5 961
10 1.046.529
20 1.099.509.530.625

Lemke-Howson proposed a more efficient algorithm... though still with

exponential running time in the worst case.

25/34
(LUISS) Existence of Equilibria for Strategic Games 25 / 34
 Games & Strategies Interesting Examples

Hotelling game

Several icecream vendors must place their cart in a beach 1 kilometer long.
People are distributed uniformly on the beach and choose the closest cart to get
icecream. People having several carts at the same distance are split evenly.

The following results are known:

1 If there are only two icecream vendors, the unique equilibrium is when they
both stay in the center of the beach.
2 When there are three vendors, no Nash equilibrium exists.
3 When they are four or five, there is one equilibrium (up to permutations).
4 With six or more carts there are infinitely many equilibria.

26/34
(LUISS) Existence of Equilibria for Strategic Games 26 / 34
 Games & Strategies Interesting Examples

Cournot competition

Two firms choose produce a certain good, with a unit production cost of c > 0.
Firm 1 produces a quantity q1 , whereas firm 2 produces a quantity q2 .
The selling price of the good decreases linearly with the total quantity q produced

p(q) = [a − q]+ = max{a − q, 0}

where the maximum price a is larger than the production cost c.

Therefore, the utilities of both firms are respectively

u1 (q1 , q2 ) = q1 p(q1 +q2 ) − c q1

u2 (q1 , q2 ) = q2 p(q1 +q2 ) − c q2

27/34
(LUISS) Existence of Equilibria for Strategic Games 27 / 34
 Games & Strategies Interesting Examples

The monopolist

Suppose that Firm 2 is temporarily out of business, so that q2 = 0. Then Firm 1

becomes a monopolist and maximizes its utility

max u1 (q1 ) = q1 [a − q1 ]+ − cq1 .

q1 ≥0

Clearly it makes no sense to produce more than a, so the optimal production lies
in the interval [0, a]. Hence the firm maximizes the quadratic (a − c)q1 − q12 ,from
which we get the optimal production level as well as the resulting price and utility

qM = 12 (a − c); pM = 12 (a + c); uM (qM ) = 14 (a − c)2

Note that if we had a < c it is optimal to produce qM = 0, so in general we have

qM = 12 [a − c]+ .

28/34
(LUISS) Existence of Equilibria for Strategic Games 28 / 34
 Games & Strategies Interesting Examples

The duopoly
Suppose now that Firm 2 is back and produces q2 . Then Firm 1 solves

max q1 [(a − q2 ) − q1 ]+ − cq1

q1 ≥0

as in the monopolist case but with a replaced by a − q2 . A symmetric problem is

solved by Firm 2. Thus, equilibrium is characterized by the equations
1
q1 = 2 [a − q2 − c]+
1
q2 = 2 [a − q1 − c]+

Assuming that both firms produce a positive quantity, we get the unique solution
q1 = q2 = 13 (a − c). The total production and the resulting price are

qD = 32 (a − c) > qM ; pD = 13 (a + 2c) < pM

and each firm makes an utility

1
uD = 19 (a − c)2 < uM .
2
29/34
(LUISS) Existence of Equilibria for Strategic Games 29 / 34
 Games & Strategies Interesting Examples

Braess Paradox

4.000 people travel from Torino to Milano, each one wants to minimize travel
time. N is the number of people driving in the corresponding road.

What are the Nash equilibria if the Up-Down street between the two small
cities is closed for maintenance work?

What if the Up-Down street is available with 5 minutes travel time?

What if the Up-Down street can also be used Down-Up?

30/34
(LUISS) Existence of Equilibria for Strategic Games 30 / 34
 Games & Strategies Interesting Examples

El Farol bar

In Santa Fe there are 500 young people, happy to go to the El Farol bar. More
people in the bar, happier they are, till they reach 300 people. They can also
choose to stay at home. So utility function can be assumed to be 0 if they stay at
home, u(x) = x if x ≤ 300, u(x) = 300 − x if x > 300.

Essentially there is a Nash equilibrium, where 300 young people are in the bar, the
other stay at home. An asymmetric situation, notwithstanding the players are
symmetric.

A mixed symmetric equilibrium is also present in this case.

31/34
(LUISS) Existence of Equilibria for Strategic Games 31 / 34
 Games & Strategies Interesting Examples

Auctions

Several types of auctions since ancient times: sequential offers, sealed offers, first
price, second price,. . .
There are n bidders, each one has a valuation v for the object, which is kept
as private information. We assume that v1 > v2 > · · · > vn .
Each bidder proposes a (non-negative) bid bi , seen as strategy for the player.
Thus the strategy space of the players is [0, ∞).
An assignment rule and the payment must be decided, including the rule for
handling ties.
Player i gets an utility vi − bi if he wins the auction, and 0 otherwise.

We consider only auctions where the winner is the highest bidder. In case of tie in
the highest bid the winner is the one who values more the object.

32/34
(LUISS) Existence of Equilibria for Strategic Games 32 / 34
 Games & Strategies Interesting Examples

First price auction

In a first price auction the rule is: the player i offering the highest bid bi gets the
object and pays exactly her bid. The other players pay nothing.

1 For player i bidding more than vi is weakly dominated.

2 One Nash equilibrium is (v2 , v2 , v3 , . . . , vn ).
3 In all equilibria the winner is player 1.
4 The two highest bids are the same and one is made by player 1. The highest
bid b1 satisfies v2 ≤ b1 ≤ v1 . All such bid profiles are Nash equilibria.

33/34
(LUISS) Existence of Equilibria for Strategic Games 33 / 34
 Games & Strategies Interesting Examples

Second price auctions

In a second price auction the rule is: the player i offering the highest bid bi gets
the object and pays the second highest bid. The other players pay nothing.

1 One Nash equilibrium is (v1 , v2 , v3 , . . . , vn )

2 Other equilibria: (v1 , 0, 0, . . . , 0), (v2 , v1 , v3 , . . . , vn )
3 A player’s bid equalizing her evaluation is a weakly dominant strategy

34/34
(LUISS) Existence of Equilibria for Strategic Games 34 / 34