GAME THEORY

Game: A game is a set of rules which players obey while playing it.

n-Person Game: A game played by n players is an n-person game.

2-Person Game: A game played by 2 players is a 2-person game.

e.g. Badminton, Hockey, Squash, Cricket, Football etc.

(Note: A game played by two teams is also called 2-person game)

1-Person Game: A game played by one player is an n-person game.

e.g. Sudoku, Solitaire etc.

Finite Game: A game having finite number of moves is called a finite game.

e.g. 20-20 Cricket Match, Test match, A game of Hockey.

Zero Sum Game: Consider an n-person game with 𝑃1 , 𝑃2 , …, 𝑃𝑛 as its

players. Suppose 𝑝𝑖 , 1  i ≤ n is the payment made to the player 𝑃𝑖 at the end
of the game.

If ∑𝑛𝑖=1 𝑝𝑖 = 0, then this game is called zero-sum game.

In this chapter, we deal with Two-person Zero-Sum Games.

Rectangular Games:

Consider a 2-person game in which the first player chooses a number (row of
a matrix) from {1, 2, …, m} and the second player, without the knowledge
what choice the first player has made, chooses a number (column of the
matrix) from {1, 2, …, n}. The two numbers are then compared and the 2nd
Players pays the 1st Player an amount by the pre-specified rules of the game.
Such games are referred to as Rectangular Games.

A rectangular game is described by the pay-off matrix.

Consider the m  n rectangular game whose pay-off matrix is

𝑃2

[1] [2] … [𝑛]

(1) 𝑎11 𝑎12 ⋯ 𝑎1𝑛
(2) 𝑎21 𝑎22 … 𝑎2𝑛
𝑃1 ( ... … … … )
..
(𝑚) 𝑎𝑚1 𝑎𝑚2 … 𝑎𝑚𝑛

If player 𝑃1 chooses row number (i), then 𝑃1 is sure to get at least minimum
of ith row i.e. at least min 𝑎𝑖𝑗
𝑗

Now, 𝑃1 would like to make min 𝑎𝑖𝑗 as large as possible.

𝑗

Thus, there is a choice for 𝑃1 so that he gets at least 𝑚𝑎𝑥 min 𝑎𝑖𝑗
𝑖 𝑗

And if player 𝑃2 chooses column number (j), then 𝑃2 is sure to pay at most
maximum of jth row i.e. at least max 𝑎𝑖𝑗
𝑖

But 𝑃2 would like to make min 𝑎𝑖𝑗 as small as possible.

𝑗
Thus, there is a choice for 𝑃2 so that he has to pay at most 𝑚𝑖𝑛 max 𝑎𝑖𝑗
𝑗 𝑖

Thus, in totality, 𝑃1 can get at least 𝑚𝑎𝑥 min 𝑎𝑖𝑗

𝑖 𝑗

and 𝑃2 can restrict 𝑃1 to get more than 𝑚𝑖𝑛 max 𝑎𝑖𝑗

𝑗 𝑖

Result: Let A = (𝑎𝑖𝑗 )𝑚×𝑛 be the pay-off matrix of a rectangular game.

Then 𝑚𝑎𝑥 min 𝑎𝑖𝑗  𝑚𝑖𝑛 max 𝑎𝑖𝑗

𝑖 𝑗 𝑗 𝑖

Definition: If 𝑚𝑎𝑥 min 𝑎𝑖𝑗 = 𝑚𝑖𝑛 max 𝑎𝑖𝑗 = v

𝑖 𝑗 𝑗 𝑖

then 𝑃1 can get v units and simultaneously 𝑃2 can be restrict 𝑃1 to get more
than v units..

Thus, in this case, v is the value of the game.

Definition. Saddle Point of a Game with given pay-off matrix:

Let A = (𝑎𝑖𝑗 )𝑚×𝑛 be the pay-off matrix of a rectangular game.

The cell (r, s) is a saddle point of the matrix A if

(i) 𝑎𝑖𝑠 ≤ 𝑎𝑟𝑠  i = 1, 2, …, m

(ii) 𝑎𝑟𝑠 ≤ 𝑎𝑟𝑗  j = 1, 2, …, n.

i.e. The cell (r, s) is a saddle point of A if 𝑎𝑟𝑠 is the maximum element of
sth column as well as 𝑎𝑟𝑠 is the minimum element of rth row.
Theorem: Let A = (𝑎𝑖𝑗 )𝑚×𝑛 be the pay-off matrix of a rectangular game. A

necessary and sufficient condition for (r, s) is a saddle point of A, is that

𝑚𝑎𝑥 min 𝑎𝑖𝑗 = min max 𝑎𝑖𝑗 = 𝑎𝑟𝑠

𝑖 𝑗 𝑗 𝑖

Note: When 𝑚𝑎𝑥 min 𝑎𝑖𝑗 = min max 𝑎𝑖𝑗 = 𝑎𝑟𝑠

𝑖 𝑗 𝑗 𝑖

Then, r is the optimal strategy for player 𝑃1

and s is the optimal strategy for the player 𝑃2

with 𝑎𝑟𝑠 is the value of the Game

Note: Max Min is the maximum of minima of all rows.

whereas Min Max is the Minimum of maxima of all columns.

Exercise 1. Find max min and min max of the following pay-off matrices, and
find the saddle points, if any

 2 2 2 2
(i) A =  
 1 2 3 4

𝑚𝑎𝑥 min 𝑎𝑖𝑗 = max {min 𝑎1𝑗 , min 𝑎2𝑗 }

𝑖 𝑗 𝑗 𝑗

= max { min(𝑎11 , 𝑎12 , 𝑎13 , 𝑎14 ), min(𝑎21 , 𝑎22 , 𝑎23 , 𝑎24 )}

= max { min(2, 2, 2, 2), min (1, 2, 3, 4)}

= max {2, 1}

= 2 = 𝑎11 = 𝑎12 = 𝑎13 = 𝑎14

min max 𝑎𝑖𝑗 = min {max 𝑎𝑖1 , max 𝑎𝑖2 , max 𝑎𝑖3 , max 𝑎𝑖4 }
𝑗 𝑖 𝑖 𝑖 𝑖 𝑖

= min {max(𝑎11 ,𝑎21 ),max(𝑎12 ,𝑎22 ),max(𝑎13 ,𝑎23 ),max(𝑎14 ,𝑎24 )}

 = min {max(2, 1), max(2, 2), max(2, 3), max(2, 4)}

= min {2, 2, 3, 4}

= 2 = 𝑎11 = 𝑎12 = 𝑎22

Hence 𝑚𝑎𝑥 min 𝑎𝑖𝑗 = min max 𝑎𝑖𝑗 = 2 = 𝑎11 = 𝑎12

𝑖 𝑗 𝑗 𝑖

Thus, (1, 1) and (1, 2) are saddle points of A and the value of the game is 2.

1 3 6
(ii) A = ‖2 1 3‖
6 2 1

𝑚𝑎𝑥 min 𝑎𝑖𝑗 = max { min 𝑎1𝑗 , min 𝑎2𝑗 , min 𝑎3𝑗 }
j𝑖 j j j

= max { min(𝑎11 , 𝑎12 , 𝑎13 ), min(𝑎21 , 𝑎22 , 𝑎23 ), min(𝑎31 , 𝑎32 , 𝑎33 )}

= max { min(1, 3, 6), min (2, 1, 3), min(6, 2, 1)}}

= max {1, 1, 1}

= 1 = 𝑎11 = 𝑎22 = 𝑎33

𝑚𝑖𝑛 max 𝑎𝑖𝑗 = min { max 𝑎𝑖1 , max 𝑎𝑖2 , max 𝑎𝑖3 }
𝑗 i i i i

= min { max(𝑎11 ,𝑎21 ,𝑎31 ), max(𝑎12 ,𝑎22 ,𝑎32 ), max(𝑎13 ,𝑎23 ,𝑎33 )}

= min {max(1, 2, 6), max(3, 1, 2), max(6, 3, 1)}

= min {6, 3, 6}

= 3 = 𝑎12 = 𝑎23

𝑚𝑎𝑥 min 𝑎𝑖𝑗  𝑚𝑖𝑛 max 𝑎𝑖𝑗

𝑖 j 𝑗 i

Hence there is no saddle point of the game.

Exercise 2. Find the range of p and q such that (2, 2) will be a saddle point of
the game whose pay-off matrix is

2 4 5
A = (10 7 𝑞)
4 𝑝 6

The point (2, 2) will be a saddle point for the game with pay-off matrix A if

𝑎𝑖2 ≤ 𝑎22  i = 1, 2, 3

𝑎22 ≤ 𝑎2𝑗  j = 1, 2, 3

i.e. if 𝑎12 ≤ 7, 𝑎32 ≤ 7

and 7 ≤ 𝑎21 , 7 ≤ 𝑎23

i.e. if 4 ≤ 7, p ≤ 7

and 7 ≤ 10, 7 ≤ q

i.e. if p ≤ 7 ≤ q

Hence, the range of values of p and q which will render (2, 2) a saddle point
of A is

p ≤ 7 ≤ q.
Mixed Strategies, Expectation Function:

Consider an m  n rectangular game with pay-off matrix

𝑃2

[1] [2] … [𝑛]

(1) 𝑎11 𝑎12 ⋯ 𝑎1𝑛
(2) 𝑎21 𝑎22 … 𝑎2𝑛
𝑃1 ( ... … … … )
..
(𝑚) 𝑎𝑚1 𝑎𝑚2 … 𝑎𝑚𝑛

Suppose the game doesn’t have any saddle point.

Let the player 𝑃1 plays row number (1) with frequency 𝑥1 , row number (2)
with frequency 𝑥2 , …, row number (m) with frequency 𝑥𝑚 and the player
𝑃2 plays column number [1] with frequency 𝑦1 , column number [2] with
frequency 𝑦2 , …, column number [n] with frequency 𝑦𝑛 .

𝑖=1 𝑥𝑖 = 1 and 𝑥𝑖 ≥ 0 ,  i ; is
An m-tuple X = (𝑥1 , 𝑥2 , … , 𝑥𝑚 ) where ∑𝑚
called a mixed strategy for the player 𝑃1 .

The set of all mixed strategies for the player 𝑃1 , denoted by, 𝑆𝑚 is

𝑖=1 𝑥𝑖 = 1 , 𝑥𝑖 ≥ 0 ,  i = 1, 2, …, m }.
𝑆𝑚 = { X = (𝑥1 , 𝑥2 , … , 𝑥𝑚 ) : ∑𝑚

An n-tuple Y = (𝑦1 , 𝑦2 , … , 𝑦𝑛 ) where ∑𝑛𝑗=1 𝑦𝑗 = 1 and 𝑦𝑗 ≥ 0 ,  j ; is

called a mixed strategy for the player 𝑃2 .

The set of all mixed strategies for the player 𝑃1 , denoted by, 𝑆𝑚 is

𝑆𝑛 = { Y = (𝑦1 , 𝑦2 , … , 𝑦𝑛 ) : ∑𝑛𝑗=1 𝑦𝑗 = 1 , 𝑦𝑗 ≥ 0 ,  j = 1, 2, …, n }.
The expectation of the player 𝑃1 , denoted by E(X, Y), is given by

E(X, Y) = ∑𝑚 𝑛
𝑖=1 ∑𝑗=1 𝑎𝑖𝑗 𝑥𝑖 𝑦𝑗 .

E(X, Y) is called the expectation function of the game.

Result: Suppose for some 𝑋 ∗  𝑆𝑚 and 𝑌 ∗  𝑆𝑛

E(X, 𝑌 ∗ )  E(𝑋 ∗ , 𝑌 ∗ )  E(𝑋 ∗ , Y)  X  𝑆𝑚 and  Y  𝑆𝑛

Then 𝑋 ∗ is an optimal strategy for the player 𝑃1 , 𝑌 ∗ is an optimal strategy

for the player 𝑃2 and E(𝑋 ∗ , 𝑌 ∗ ) is the value of the game.

In this case, the point (𝑋 ∗ , 𝑌 ∗ ) is called strategic saddle point of the game.

Pure Strategy:

A mixed strategy X = (𝑥1 , 𝑥2 , … , 𝑥𝑚 ) is called a pure strategy for the player

𝑃1 if

𝑥𝑖 = 1 for some i

and 𝑥𝑗 = 0  j ≠ i.

For the pure strategy 𝑋 𝑖 = (0, …, 0, 𝑥𝑖 = 1, 0, …, 0)

E(i, Y) = E(𝑋 𝑖 , Y) = ∑𝑚 𝑛
𝑖=1 ∑𝑗=1 𝑎𝑖𝑗 𝑥𝑖 𝑦𝑗

= ∑𝑛𝑗=1 𝑎𝑖𝑗 𝑦𝑗
A mixed strategy Y = (𝑦1 , 𝑦2 , … , 𝑦𝑛 ) is called a pure strategy for the player
𝑃2 if

𝑦𝑗 = 1 for some j

and 𝑦𝑖 = 0  i ≠ j.

For the pure strategy 𝑌𝑗 = (0, …, 0, 𝑦𝑗 = 1, 0, …, 0)

E(X, j) = E(X, 𝑌𝑗 ) = ∑𝑚 𝑛
𝑖=1 ∑𝑗=1 𝑎𝑖𝑗 𝑥𝑖 𝑦𝑗

= ∑𝑚
𝑖=1 𝑎𝑖𝑗 𝑥𝑖

Theorem. Let E(X, Y) be the expectation function of the rectangular game.

Let 𝑋 ∗  𝑆𝑚 , 𝑌 ∗  𝑆𝑛 and v be a real number . Then, a necessary and
sufficient condition for 𝑋 ∗ to be an optimal strategy for the player 𝑃1 and
for 𝑌 ∗ to be an optimal strategy for the player 𝑃2 ; and v to be the value of
the game is

E(i, 𝑌 ∗ )  v  E(𝑋 ∗ , j)  i = 1, 2, …, m ,  j = 1, 2, …, n.

2 1
Exercise: Solve the game whose pay-off matrix is ( )
0 3

2 1
Solution. The pay-off matrix is ( )
0 3

Here, max min 𝑎𝑖𝑗 = max{1, 0} = 1

min max 𝑎𝑖𝑗 = min {2, 3} = 2

max min 𝑎𝑖𝑗 ≠ min max 𝑎𝑖𝑗

 The game has no saddle point.
𝑃2
[1] [2]
y 1–y
(1) 𝑥 2 1
𝑃1 ( )
(3) 1 − 𝑥 0 3

Let 𝑃1 plays (1) with frequency x and (3) with frequency 1 – x. Then the
expectation of 𝑃1 is

E(x, y) = 2xy + x (1 – y) + 3(1 – x)(1 – y)

= 4xy – 2x – 3y + 3

1 3
= 4(xy − x – y) + 3
2 4

3 1 3
= 4(x – )( y − ) − + 3
4 2 2

3 1 3
= 4(x – )( y − ) +
4 2 2

3 3
Now, by playing with x = , 𝑃1 is sure to get at least while on the other hand
4 2
1 3
by playing with y = , 𝑃2 can restrict 𝑃1 to get more than .
2 2

3
Thus, is the value of the game.
2

3 1
An optimal strategy for 𝑃1 is ( , )
4 4

1 1
An optimal strategy for 𝑃2 is ( , )
2 2

3
and value of the game is .
2
Shortcut formula to find the optimal strategy for the game with 22 pay-
off matrix

Consider a rectangular game with 22 pay-off matrix

𝑃2
𝑎11 𝑎12
𝑃1 (𝑎 𝑎22 )
21

For this game,

The optimal strategy for the player 𝑃1 is given by

𝑎22 −𝑎21 𝑎22 −𝑎21 𝑎22 −𝑎21 𝑎11 −𝑎12
( ,1 − )=( , )
𝛥 𝛥 𝛥 𝛥

The optimal strategy for the player 𝑃2 is given by

𝑎22 −𝑎12 𝑎22 −𝑎12 𝑎22 −𝑎12 𝑎11 −𝑎21
( ,1 − )=( , )
𝛥 𝛥 𝛥 𝛥

The value of the game is given by

𝑎11 𝑎22 −𝑎12 𝑎21

𝛥

where  = 𝑎11 − 𝑎12 − 𝑎21 + 𝑎22

Equivalence of Rectangular Game and Linear Programming Problem:

The pay-off matrix is

3 −2 4
′
𝐴 = ( −1 4 2)
3 2 −3

To make the value of game positive, adding 4 to the elements of the matrix B,
we solve the game with pay-off matrix

7 2 8
A = (3 8 6)
7 6 1

Let 𝑣 ′ and v be the values of the games with pay-off matrices 𝐴′ and A
respectively. Then
v = 𝑣′ + 4

Solving the game with pay-off matrix A means to find 𝑋 ∗ = (𝑥1 , 𝑥2 , 𝑥3 ) and
𝑌 ∗ = (𝑦1 , 𝑦2 , 𝑦3 ) and a real number v such that

E(i,𝑌 ∗ ) ≤ v ≤ E(𝑋 ∗ , j)  i, j = 1, 2, 3

i.e. to find 𝑥1 , 𝑥2 , 𝑥3 , 𝑦1 , 𝑦2 , 𝑦3 and v such that

7𝑥1 + 3𝑥2 + 7𝑥3  v 7𝑦1 + 2𝑦2 + 8𝑦3 ≤ v

2𝑥1 + 8𝑥2 + 6𝑥3  v and 3𝑦1 + 8𝑦2 + 6𝑦3 ≤ v

8𝑥1 + 6𝑥2 + 𝑥3  v 7𝑦1 + 6𝑦2 + 𝑦3 ≤ v

𝑥1 + 𝑥2 + 𝑥3 = v 𝑦1 + 𝑦2 + 𝑦3 = v

𝑥1 , 𝑥2 , 𝑥3  0 𝑦1 , 𝑦2 , 𝑦3  0
 𝑥𝑖 𝑦𝑗
Dividing the above relations by v ( > 0) and letting = 𝑥′𝑖, = 𝑦′𝑗 , i,
𝑣 𝑣

j = 1, 2, 3; we get

7𝑥 ′1 + 3𝑥 ′ 2 + 7𝑥 ′ 3  1 7𝑦 ′1 + 2𝑦 ′ 2 + 8𝑦 ′ 3 ≤ 1

2𝑥 ′1 + 8𝑥 ′ 2 + 6𝑥 ′ 3  1 and 3𝑦 ′1 + 8𝑦 ′ 2 + 6𝑦 ′ 3 ≤ 1

8𝑥 ′1 + 6𝑥 ′ 2 + 𝑥 ′ 3  1 7𝑦 ′1 + 6𝑦 ′ 2 + 𝑦 ′ 3 ≤ 1
1 1
𝑥 ′1 + 𝑥 ′ 2 + 𝑥 ′ 3 = 𝑦 ′1 + 𝑦 ′ 2 + 𝑦 ′ 3 =
𝑣 𝑣

𝑥 ′1 , 𝑥 ′ 2 , 𝑥 ′ 3  0 𝑦1 , 𝑦 ′ 2 , 𝑦 ′ 3  0
1
Since the player 𝑃1 would like to maximize v i.e to minimize = 𝑥 ′1 + 𝑥 ′ 2 + 𝑥 ′ 3
𝑣

 The problem of the player 𝑃1 is

Minimize z = 𝑥 ′1 + 𝑥 ′ 2 + 𝑥 ′ 3

subject to

7𝑥 ′1 + 3𝑥 ′ 2 + 7𝑥 ′ 3  1

2𝑥 ′1 + 8𝑥 ′ 2 + 6𝑥 ′ 3  1 (A)

8𝑥 ′1 + 6𝑥 ′ 2 + 𝑥 ′ 3  1

𝑥 ′1 , 𝑥 ′ 2 , 𝑥 ′ 3  0

Again,
1
since the player 𝑃2 would like to minimize v i.e. to maximize = 𝑦 ′1 + 𝑦 ′ 2 + 𝑦 ′ 3
𝑣
 The problem of the player 𝑃2 is

Maximize z = 𝑦 ′1 + 𝑦 ′ 2 + 𝑦 ′ 3

subject to

7𝑦 ′1 + 2𝑦 ′ 2 + 8𝑦 ′ 3 ≤ 1

3𝑦 ′1 + 8𝑦 ′ 2 + 6𝑦 ′ 3 ≤ 1 (B)

7𝑦 ′1 + 6𝑦 ′ 2 + 𝑦 ′ 3 ≤ 1

𝑦1 , 𝑦 ′ 2 , 𝑦 ′ 3  0

The problems (A) and (B) are a pair of dual problems equivalent to the game
problem.