Zero-Sum Games

LUISS

1/26
(LUISS) Zero-Sum Games 1 / 26
Contents of the week

Zero sum games

Conservative values
Von Neumann theorem
Fair games

2/26
(LUISS) Zero-Sum Games 2 / 26
 Zero sum games

Zero sum games

An interesting case of non cooperative game is is when there are two players, with
opposite interests.

Definition
A two player zero sum game in strategic form is given by strategy sets X and Y
and a payoff function f : X × Y → R

Conventionally f (x, y ) is what Player I gets from Player II, when they play x, y
respectively. The payoff for Player II is −f (x, y ) (she pays f (x, y ) to Player I).

3/26
(LUISS) Zero-Sum Games 3 / 26
 Zero sum games

Finite zero-sum games

In the finite case X = {1, 2, . . . , n}, Y = {1, 2, . . . , m} the game is described by a

payoff matrix P

Example
 
4 3 1
P= 7 5 8 
8 2 0

Player I selects row i, Player II selects column j.

4/26
(LUISS) Zero-Sum Games 4 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

Conservative values

 
4 3 1
 7 5 8 
8 2 0
Player I can guarantee herself to get at least

v1 = max min pij = max{1, 5, 0} = 5

i j

Player II can guarantee himself to pay no more than

v2 = min max pij = min{8, 5, 8} = 5

j i

Rational outcome: i ∗ = 2, j ∗ = 2, Value v1 = v2 = 5

5/26
(LUISS) Zero-Sum Games 5 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

Alternative idea of solution

Suppose
v1 = v2 := v
i ∗ is the row attaining the maxi minj pij = v so that pi ∗j ≥ v for all j
j ∗ is the column attaining the minj maxi pij = v so that pij ∗ ≤ v for all i

Then pi ∗j ∗ = v is the rational outcome of the game.

Remark
i ∗ is an optimal strategy for Player I, because he cannot get more than v ,
since v is the conservative value of Player II
j ∗ is an optimal strategy for Player II, because he cannot pay less than v ,
since v is the conservative value of Player I

6/26
(LUISS) Zero-Sum Games 6 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

The equality v1 = v2 need not hold

Example
 
0 1 -1
P =  -1 0 1 
1 -1 0

v1 = −1, v2 = 1

Nothing unexpected. . .

7/26
(LUISS) Zero-Sum Games 7 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

General zero-sum games

f :X ×Y →R

The players can guarantee to themselves the conservative values:

Player I: v1 = maxx miny f (x, y )

Player II: v2 = miny maxx f (x, y )

We always have v1 ≤ v2

8/26
(LUISS) Zero-Sum Games 8 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

v1 ≤ v2

Proposition

v1 = max min f (x, y ) ≤ min max f (x, y ) = v2

x y y x

Proof Observe that, for all x, y ,

min f (x, y ) ≤ f (x, y ) ≤ max f (x, y )

y x

Thus
min f (x, y ) ≤ max f (x, y )
y x

Since the left hand side does not depend on y and the right hand side does not
depend on x, the thesis follows.

9/26
(LUISS) Zero-Sum Games 9 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

Saddle points and Nash equilibria of zero-sum games

Suppose that we have strategies x̄ and ȳ such that for all y ∈ Y and x ∈ X

f (x, ȳ ) ≤ f (x̄, ȳ ) ≤ f (x̄, y ).

Such a pair (x̄, ȳ ) is called a saddle point.

Let v = f (x̄, ȳ ). Then
The rational outcome of the game is v1 = v2 = v
x̄ is an optimal strategy for Player I
ȳ is an optimal strategy for Player II

10/26
(LUISS) Zero-Sum Games 10 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

The Nash equilibria of a zero sum game

Theorem
Let X , Y be nonempty strategy sets and f : X × Y → R. Then the following are
equivalent:
1 The pair (x̄, ȳ ) is a Nash equilibrium, i.e. fulfills

f (x, ȳ ) ≤ f (x̄, ȳ ) ≤ f (x̄, y ) ∀x ∈ X , ∀y ∈ Y

2 The following conditions are satisfied:

(i) miny maxx f (x, y ) = maxx miny f (x, y ): the two conservative values agree
(ii) miny f (x̄, y ) = maxx miny f (x, y ): x̄ is optimal for Player I
(iii) maxx f (x, ȳ ) = miny maxx f (x, y ): ȳ is optimal for Player II

11/26
(LUISS) Zero-Sum Games 11 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

Proof

Proof 1) implies 2). From 1) we get:

min max f (x, y ) ≤ max f (x, ȳ ) = f (x̄, ȳ ) = min f (x̄, y ) ≤ max min f (x, y )
y x x y x y

Since v1 ≤ v2 always holds, all above inequalities are equalities

Conversely, suppose 2) holds. Then

(iii) (ii)
min max f (x, y ) = max f (x, ȳ ) ≥ f (x̄, ȳ ) ≥ min f (x̄, y ) = max min f (x, y )
y x x y x y

Because of (i), all inequalities are equalities and the proof is complete

12/26
(LUISS) Zero-Sum Games 12 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

Mixed extension of a zero-sum game

Zero-sum finite game: n × m matrix P.

Mixed strategy space for Player I:
n
X
X = Σn = {x = (x1 , . . . , xn ) : xi ≥ 0, xi = 1}
i=1

Mixed strategy space for Player II:

m
X
Y = Σm = {y = (y1 , . . . , ym ) : yj ≥ 0, yj = 1}
j=1

Expected payoff:
n X
X m
f (x, y ) = pij xi yj = x t Py
i=1 j=1

13/26
(LUISS) Zero-Sum Games 13 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

To prove existence of a rational outcome

What must be proved to have existence of a rational outcome:

1 v1 = v2
2 there exists x̄ such that

v1 = max min f (x, y ) = min f (x̄, y )

x y y

3 there exists ȳ such that

v2 = min max f (x, y ) = max f (x, ȳ )

y x x

Observe that if one considers a finite set of strategies x̄ and ȳ fulfilling 2) and 3)
always exist; thus in this case it suffices to establish 1). This means that in order
to check if there are equilibria in pure strategies, only 1) must be checked.

14/26
(LUISS) Zero-Sum Games 14 / 26
 Zero sum games Ignoring the idea of Nash equilibrium

The von Neumann theorem

Theorem

A two player, finite, zero sum game as described by a payoff matrix P has a
rational outcome in mixed strategies: the two conservative values of the players
coincide and there are optimal mixed strategies x̄, ȳ for the players.

Remark
We remind that when the two conservative values agree the strategy x̄ is optimal
for Player I if and only if it guarantees her to get the (common conservative) value
no matter what Player II does; dually the strategy ȳ is optimal for Player II if and
only if it guarantees him to get the (common conservative) value no matter what
Player I does.

15/26
(LUISS) Zero-Sum Games 15 / 26
 How to find optimal strategies

Finding optimal strategies: Player I

Player I must choose a probability distribution Σn 3 x = (x1 , . . . , xn ):

p11 x1 + . . . + pn1 xn ≥ v1
...
p1j x1 + . . . + pnj xn ≥ v1
...
p1m x1 + . . . + pnm xn ≥ v1

where v1 must be as large as possible



 max v1
 t
P x ≥ v1 1m
(P1 )

 x1 + . . . + xn = 1
x ≥ 0, v1 ∈ R


16/26
(LUISS) Zero-Sum Games 16 / 26
 How to find optimal strategies

Finding optimal strategies: Player II

Player II must choose a probability distribution Σm 3 y = (y1 , . . . , ym ):

p11 y1 + . . . + p1m ym ≤ v2
...
pi1 y1 + . . . + pim ym ≤ v2
...
pn1 y1 + . . . + pnm ym ≤ v2

where v2 must be as small as possible



 min v2
Py ≤ v2 1n

(P2 )

 y1 + . . . + ym = 1
y ≥ 0, v2 ∈ R


17/26
(LUISS) Zero-Sum Games 17 / 26
 How to find optimal strategies

In matrix form

 

 max v1 
 min v2
 t
P x ≥ v1 1 m Py ≤ v2 1n

(P1 ) (P2 )

 x1 + . . . + xn = 1 
 y1 + . . . + ym = 1
x ≥ 0, v1 ∈ R y ≥ 0, v2 ∈ R
 

These linear programs are dual to each other!

Both are feasible ⇒ no duality gap v1 = v2 and there are optimal solutions

18/26
(LUISS) Zero-Sum Games 18 / 26
 How to find optimal strategies

The complementarity conditions

The complementarity conditions become

Pm
xi > 0 =⇒ j=1 pij yj = v2
Pn
yj > 0 =⇒ i=1 pij xi = v1
Pn
Since i=1 pij xi is the expected value for Player II if she plays column j and
Player I the mixed strategy x = (x1 , . . . , xn ), the complementarity conditions
show, one more time, that a Player must give a positive probability only to those
pure strategies that have optimal expected payoff.

19/26
(LUISS) Zero-Sum Games 19 / 26
 How to find optimal strategies

Summarizing

A finite zero sum game has always rational outcome in mixed strategies.

The set of Nash equilibria can be found by solving a pair of dual linear
programming problems.

The outcome, at each pair of optimal strategies, is the common conservative

value v = v1 = v2 of the players.

The set of optimal strategies for the players is a nonempty closed convex set, the
smallest convex set containing a finite number of points, called the extreme points
of the set.

20/26
(LUISS) Zero-Sum Games 20 / 26
 How to find optimal strategies

As a consequence of the theorem

Every Nash equilibrium (x̄, ȳ ) of the zero sum game provides optimal
strategies for the players
Any pair of optimal strategies for the players provides a Nash equilibrium for
the zero sum game

Thus Nash theorem is a generalization of von Neumann theorem

21/26
(LUISS) Zero-Sum Games 21 / 26
 How to find optimal strategies

A comment

Remark
Von Neumann approach with conservatives values shows that, in the particular
case of the zero sum game:
Each player can find an optimal strategy independently of the other player.
Any pair of optimal strategies provides a Nash equilibrium; this implies no
need of coordination to reach an equilibrium.
Every Nash equilibrium provides the same utility (payoff) to the players:
multiplicity of solutions does not create problems.
Nash equilibria are easy to be found in zero sum games.

22/26
(LUISS) Zero-Sum Games 22 / 26
 How to find optimal strategies

Fair games

Definition
A square matrix n × n P = (pij ) is said to be anti-symmetric provided pij = −pji
for all i, j = 1, . . . , n. A finite zero sum game is said to be fair if the associated
matrix is antisymmetric.

Example: Rock-Scissors-Paper.
In fair games there is no favorite player.

23/26
(LUISS) Zero-Sum Games 23 / 26
 How to find optimal strategies

Fair outcome

Proposition
In a fair game
the value is 0
x̄ is an optimal strategy for Player I if and only if it is optimal for Player II

Proof Since
x t Px = (x t Px)t = x t P t x = −x t Px,

f (x, x) = 0 for all x, thus v1 ≤ 0, v2 ≥ 0

Then v = 0.

If x̄ is optimal for the Player I, x̄ t Py ≥ 0 for all y

Thus y t P x̄ ≤ 0 for all y ∈ Σn , and x̄ is optimal for Player II

24/26
(LUISS) Zero-Sum Games 24 / 26
 How to find optimal strategies

Finding optimal strategies in a fair game

Need to solve the system of inequalities

p11 x1 + . . . + pn1 xn ≥ 0
...
p1j x1 + . . . + pnj xn ≥ 0
...
p1m x1 + . . . + pnm xn ≥ 0

with the extra conditions:

n
X
xi ≥ 0, xi = 1
i=1

25/26
(LUISS) Zero-Sum Games 25 / 26
 How to find optimal strategies

A proposed exercise

Exercise
Find the optimal strategies of the players in the rock,scissors, paper game and in
the following fair game:
 
0 3 -2 0
 -3 0 0 4 
P=  2 0

0 -3 
0 -4 3 0

26/26
(LUISS) Zero-Sum Games 26 / 26