Professional Documents
Culture Documents
Game5360 - Chapter 1 - 2 PDF
Game5360 - Chapter 1 - 2 PDF
Contents
1 Two-person zero sum games 2
1.1 Game matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 2 × 2 games . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Games reducible to 2 × 2 games . . . . . . . . . . . . . . . . . 12
1.4 2 × n and m × 2 games . . . . . . . . . . . . . . . . . . . . . . 14
The first condition means that when the row player uses the k-th strategy,
then the payoff to the row player is not less than akl no matter how the column
player plays. The second condition means that when the column player uses
the l-th strategy, then the payoff to the row player is not larger than akl no
matter how the row player plays. Consequently we have
Theorem 1.1.3. If A has a saddle point akl , then the row player may guar-
antee that his payoff is not less than akl by using the k-th strategy and the
column player may guarantee that the payoff to the row player is not larger
than akl by using the l-th strategy.
Suppose A is a matrix which has a saddle point akl . The above theo-
rem implies that the corresponding row and column constitute the optimal
strategies for the players. To find the saddle points of a matrix, first write
down the row minima of the rows and the column maxima of the columns.
Then find the maximum of row minima which is called the maximin, and
Proof? the minimum of the column maxima which is called the minimax. If the
maximin is equal to the minimax, then the entry in the corresponding row
and column is a saddle point. If the maximin and minimax are different,
then the matrix has no saddle point.
Example 1.1.4.
min
1 2 0 0
3 5 2 2
0 −4 −3 −4
−2 4 1 −2
max 3 5 2
Both the maximin and minimax are 2. Therefore the entry a23 = 2 is a
saddle point.
Example 1.1.5.
Two-person zero sum games 4
min
2 −1 3 1 −1
−4 2 0 3 −4
0 1 −2 4 −2
max 2 2 3 4
The maximin is −1 while the minimax is 2 which are not equal. Therefore
the matrix has no saddle point.
Saddle point of a matrix may not be unique. However the values of saddle
points are always the same.
Theorem 1.1.6. The values of the saddle points of a matrix are the same.
That is to say, if akl and apq are saddle points of a matrix, then akl = apq .
Furthermore, we have apq = apl = akq = akl .
Proof. We have
Therefore
akl = akq = apq = apl
We have seen that if A has a saddle point, then the two players have
optimal strategies by using one of their strategies constantly (Theorem 1.1.3).
If A has no saddle point, it is expected that the optimal ways for the players
to play the game are not using one of the strategies constantly. Take the
rock-paper-scissors game as an example.
R P S
R 0 −1 1
P 1 0 −1
S −1 1 0
Two-person zero sum games 5
Here we use the order rock(R), paper(P), scissors(S) to write down the game
matrix.
1. xi ≥ 0 for any i = 1, 2, · · · , m
m
P
2. xi = 1
i=1
= xAyT
3. pAqT = ν
In the above theorem, the real number ν = ν(A) is called the value, or the
security level, of the game matrix A. The strategy p is called a maximin
strategy for the row player and the strategy q is called a minimax strategy
for the column player. The value ν of a matrix is unique. However maximin
strategy and minimax strategy are in general not unique.
The maximin strategy p and the minimax strategy q are the optimal
strategies for the row player and the column player respectively. It is because
the row player may guarantee that his payoff is at least ν no matter how the
column player plays by using the maximin strategy p. This is also the reason
why the value ν is called the security level. Similarly, the column player may
guarantee that the payoff to the row player is at most ν, and thus his payoff
is at least −ν, no matter how the row player plays by using the minimax
strategy q. We will prove the minimax theorem in Section 2.4.
1.2 2 × 2 games
In this section, we study 2 × 2 game matrices closely and see how one can
solve them, that means finding the maximin strategies for the row player,
minimax strategies for the column player and the values of the game. First
we look at a simple example.
S R
R 1 0
P −1 1
It is easy to see that the game matrix does not have a saddle point. Thus
there is no pure maximin or minimax strategy. To solve the game, suppose
Two-person zero sum games 8
This shows that when the row player uses mixed strategy x = (x, 1 − x), then
his payoff is 2x − 1 if the column player uses his 1st strategy scissors(S) and
is 1 − x if the column player uses his 2nd strategy rock(R). Now we solve the
equation 2x − 1 = 1 − x and get x = 23 . One may see that if 0 ≤ x < 23 ,
then 2x − 1 < x − 1 and a rational column player would use his 1st strategy
scissors(S) and the payoff to the row player would be 2x − 1 < 13 . On the
other hand, if 23 < x ≤ 1, then 2x − 1 > 1 − x and a rational column
player would use his 2nd strategy rock(R) and the payoff to the row player
would be 1 − x < 13 . Now if x = 23 , that is if the row player uses the mixed
strategy ( 23 , 13 ), then he may guarantee that his payoff is 1 − x = 2x − 1 = 13
no matter how the column player plays. This is the largest payoff he may
guarantee and therefore the mixed strategy p = ( 23 , 31 ) is the maximin strategy
for the row player. Similarly, suppose the column player uses mixed strategy
y = (y, 1 − y). Consider
T 1 0 y y y
Ay = =
−1 1 1−y −y + (1 − y) 1 − 2y
If 0 ≤ y < 13 , then y < 1 − 2y and a rational row player would use his 2nd
strategy paper(P) and his payoff would be 1 − 2y > 13 . If 31 < y ≤ 1, then
y > 1 − 2y and a rational row player would use his 1st strategy rock(R) and
his payoff would be y > 13 . If y = 31 , then the payoff to the row player is always
1
3
no matter how he plays. Therefore q = ( 31 , 23 ) is the minimax strategy for
the column player and he may guarantee that the payoff to the row player is
1
3
no matter how the row player plays. Moreover the value of the game is
ν = 31 . We summarize the above discussion in the following statements.
1. The row player may use his maximin strategy p = ( 23 , 13 ) to guarantee
that his payoff is ν = 13 no matter how the column player plays.
Proof. Suppose the row player uses mixed strategy x = (x, 1 − x). Consider
a b
xA = (x, 1−x) = (ax+c(1−x), bx+d(1−x)) = ((a−c)x+c, (b−d)x+d)
c d
Since A has no saddle point, we have a − c and b − d are of different sign and
the maximum of the above minimum is obtained when
(a − c)x + c = (b − d)x + d
d−c
⇒ x =
a−b−c+d
a−b
Note that x and 1 − x = a−b−c+d must be of the same sign because A has no
saddle point. We must have 0 < x < 1 and we conclude that the maximin
strategy for the row player is
d−c a−b
p= ,
a−b−c+d a−b−c+d
Two-person zero sum games 10
Similarly suppose the column player uses mixed strategy y = (y, 1 − y).
Consider
T a b y aq + b(1 − y) (a − b)y + b
Ay = = =
c d 1−y cq + d(1 − y) (c − d)y + d
The column player may guarantee that the payoff to the row player is at
most
max{(a − b)y + b, (c − d)y + d}
The above quantity attains it minimum when
(a − b)y + b = (c − d)y + d
d−b
⇒ y =
a−b−c+d
and the minimax strategy for the column player is
d−b a−c
q= ,
a−b−c+d a−b−c+d
By calculating
ad−bc
ad − bc ad − bc T a−b−c+d
pA = , and Aq = ad−bc
a−b−c+d a−b−c+d a−b−c+d
we see that the maximum payoff that the row player may guarantee to him-
self and the minimum payoff to the row player that the column player may
ad−bc
guarantee are both a−b−c+d . In fact the minimax theorem (Theorem 1.1.10)
says that these two values must be equal. We conclude that the value of A
ad−bc
is ν = a−b−c+d .
Note that the above formulas work only when A has no saddle point. If
A has a saddle point, the vectors p and q obtained using the formulas may
not be probability vectors.
The game matrix has no saddle point. By Theorem 1.2.2, the value of the
game is
ad − bc 1 × 1 − 0 × (−1) 1
ν= = =
a−b−c+d 1 − 0 − (−1) + 1 3
the maximin strategy for the row player is
d−c a−b
p = ,
a−b−c+d a−b−c+d
1 − (−1) 1−0
= ,
1 − 0 − (−1) + 1 1 − 0 − (−1) + 1
2 1
= ,
3 3
and the minimax strategy for the column player is
d−b a−c
q = ,
a−b−c+d a−b−c+d
1−0 1 − (−1)
= ,
1 − 0 − (−1) + 1 1 − 0 − (−1) + 1
1 2
= ,
3 3
Example 1.2.4. In a game, each of the two players Andy and Bobby calls
out a number simultaneously. Andy may call out either 1 or −2 while Bobby
may call out either 1 or −3. Then Bobby pays p dollars to Andy where p
is the product of the two numbers (Andy pays Bobby −p dollars when p is
negative). The game matrix of the game is
1 −3
A=
−2 6
The value of the game is
1 × 6 − (−2) × (−3)
ν= =0
1 − (−3) − (−2) + 6
the maximin strategy for Andy is
6 − (−2) 1 − (−3) 2 1
p= , = ,
1 − (−3) − (−2) + 6 1 − (−3) − (−2) + 6 3 3
Two-person zero sum games 12
We say that a two-person zero sum game is fair if its value is zero. The
game in Example 1.2.4 is a fair game.
1. We say that the k-th row is dominated by the r-th row if akj ≤ arj for
any j = 1, 2, · · · , n.
2. We say that the l-th column is dominated the s-th column if ail ≥ ais
for any i = 1, 2, · · · , m.
If the k-th row of A is dominated by the r-th row, then for the row player,
playing the r-th strategy is at least as good as playing the k-th strategy.
Thus the k-th row can be ignored in finding the maximin strategy for the
row player. Similarly the column player may ignore a dominated column
when finding his minimax strategy.
Proof. Suppose the k-th row of A is dominated by the r-th row and A0
is obtained by deleting the k-th row from A. Let ν 0 be the value of A0
and q ∈ P n be a minimax strategy of A0 . For any mixed strategy x =
(x1 , · · · , xm ) ∈ P m , define x0 = (x01 , · · · , x0k−1 , x0k+1 , · · · , x0m ) ∈ P m−1 by
(
xi if i 6= r
x0i =
xk + xr if i = r
and we have
xAqT ≤ x0 A0 qT ≤ ν 0
Here the first inequality holds because the k-th is dominated by the r-th
row and the second inequality holds because q is a minimax strategy of A0 .
Thus the value of A is less than or equal to ν 0 . On the other hand, let
p0 = (p1 , · · · , pk−1 , pk+1 , · · · , pm ) ∈ P m−1 be a maximin strategy of A0 and
let p = (p1 , · · · , pk−1 , 0, pk+1 , · · · , pm ) ∈ P m . Then we have
pAyT = p0 A0 yT ≥ ν 0
by another row if all its entries are strictly smaller than the corresponding
entries of the other. Similarly a column is strictly dominated by another
column if all its entries are strictly larger than the corresponding entries of
the other.
Example 1.3.3. To solve the game matrix
3 −1 4
A = 2 −3 1
−2 4 0
we may delete the second row since it is dominated by the first row and get
the reduced matrix
0 3 −1 4
A =
−2 4 0
Then we may delete the third column since is dominated by the first column.
Hence the matrix A is reduced to the 2 × 2 matrix
00 3 −1
对吗? A =
−2 4
The value of this 2 × 2 matrix is 0.7. The maximin and minimax strategies
are (0.6, 0.4) and (0.5, 0.5) respectively. Therefore the value of A is 0.7, a
maximin strategy for the row player is (0.6, 0, 0.4) and a minimax strategy for
the column player is (0.5, 0.5, 0). Note that we need to insert the zeros to the
dominated rows and columns when writing down the maximin and minimax
strategies for the players.
Step 1.
For each 1 ≤ j ≤ n, draw the graph of
The graph shows the payoff to the row player if the column player uses
the j-th strategy.
Step 2.
Draw the graph of
Step 3.
Suppose (p, ν) is a maximum point of the lower envelope. Then ν is
the value of the game and p = (p, 1 − p) is a maximin strategy for the
row player.
Step 4.
The solutions for y ∈ P n to the equation
AyT = ν1T
where 1 = (1, 1), give the minimax strategy for the column player.
Solution.
Two-person zero sum games 16
we obtain the maximum point (p, ν) = (0.4, 1.8) of the lower envelope.
Step 4. Find the minimax strategies for the column player by solving
0 6 y2 1.8
=
3 −1 y4 1.8
Two-person zero sum games 17
Therefore the value of the game is ν = 1.8. The maximin strategy for the
row player is p = (0.4, 0.6) and the minimax strategy for the column player
is q = (0, 0.7, 0, 0.3).
By solving
C1 : v = x − (1 − x)
C3 : v = 2(1 − x)
C4 : v = −x + 5(1 − x)
Two-person zero sum games 18
we see that the maximum point of the lower envelope is (p, ν) = (0.75, 0.5).
Thus the maximin strategy for the row player is (0.75, 0.25) and the value of
the game is ν = 0.5. To find minimax strategies for the column player, we
solve
1 1 1 y1 1
1 0 −1 y3 = 0.5
−1 2 5 y4 0.5
Note that we have added the equation y1 +y3 +y4 = 1 to exclude the solutions
which are not probability vectors. (Explain why we didn’t do it in Example
1.4.1.) Using row operation, we obtain the row echelon form
1 1 1 1 1 0 −1 0.5
1 0 −1 0.5 −→ 0 1 2 0.5
−1 2 5 0.5 0 0 0 0
The maximum points of the lower envelope are points lying on the line seg-
ment joining (0.25, −1) and (0.5, −1). Thus the value of the game is ν = −1.
The maximin strategies for the row player are
q = (0, 1, 0, 0, 0)
Suppose the minimum point of the upper envelope is (q, ν). Then the
value of the game is ν and the minimax strategy for the column player
Two-person zero sum games 20
is q = (q, 1 − q). Moreover the maximum strategies for the row player
are the solutions for x ∈ P m to the equation
xA = ν1 = (ν, ν)
Method 2.
Solve the game with 2 × m game matrix −AT . Then
−3 4
Solution.
Method 1.
Let y = (y, 1 − y), 0 ≤ y ≤ 1, be the strategy of the column player.
The upper envelope is
Two-person zero sum games 21
Solving (
R2 : v = 3(1 − y)
R4 : v = −3y + 4(1 − y)
the minimum point of the upper envelope is (q, ν) = (0.4, 1.2). Now
the row player would only use the 2nd and 4th strategy and we solve
3 0
(x2 , x4 ) = (1.2, 1.2)
−3 4
which gives (x2 , x4 ) = (0, 7, 0.3). Therefore the value of the game is
ν = 1.2, the maximin strategy for the row player is p = (0, 0.7, 0, 0.3)
and the minimax strategy for the column player is q = (0.4, 0.6).
Method 2.
Consider
T −4 −3 1 3
−A =
2 0 −1 −4
Draw the lower envelope
Two-person zero sum games 22
We see that the value of −AT is -1.2 and the maximin strategy of −AT
is (0.4, 0.6). Solving
−3 3 x2 −1.2
=
0 −4 x4 −1.2
Theorem 1.4.5 (Principle of indifference). Let A be an m × n game matrix.
Suppose ν is the value of A, p = (p1 , · · · , pm ) be a maximin strategy for the
row player and q = (q1 , · · · , qn ) be a minimax strategy for the column player.
Pn
For any k = 1, 2, · · · , m, if pk > 0, then akj qj = ν. In particular, when
j=1
the column player uses his minimax strategy q, then the payoff to the row
Two-person zero sum games 23
player are indifferent among all his k-th strategies with pk > 0. Similarly,
m
P
for any l = 1, 2, · · · , n, if ql > 0, then ail pi = ν. In particular, when the
i=1
row player uses his maximin strategy p, then the payoff to the row player are
indifferent among all the l-th strategies of the column player with ql > 0.
Proof. For each k = 1, 2, · · · , m, we have
n
X
akj qj ≤ ν
j=1
since q is a minimax strategy for the column player. On the other hand,
m n
! m
X X X
ν = pAqT = pk akj qj ≤ pk ν = ν
k=1 j=1 k=1
Thus we have n
X
pk akj qj = pk ν
j=1
2. Solve the following game matrix, that is, find the value of the game,
a maximin strategy for the row player and a minimax strategy for the
column.
Two-person zero sum games 24
1 7 5 −3
(a)
2 −2 −3 5
(e)
2 −1
3 −1 4 0
(b)
−2 4
5 −2 4
(f) 3 −3 1
3 2 4 0
(c) 0 3 2
−2 1 −4 5
5 1 −2 6
1 0 4 2 (g) −1 0 1 −2
(d)
0 2 −3 −2 3 2 5 4
3. Raymond holds a black 2 and a red 9. Calvin holds a red 3 and a black
8. Each of them chooses one of the cards from his hand and then two
players show the chosen cards simultaneously. If the chosen cards are
of the same colour, Raymond wins and Calvin wins if the cards are of
different colours. The loser pays the winner an amount equal to the
number on the winner’s card. Write down the game matrix, find the
value of the game and the optimal strategies of the players.
4. Alex and Becky point fingers to each other, with either one finger or
two fingers. If they match with one finger, Becky pays Alex 3 dollars.
If they match with two fingers, Becky pays Alex 11 dollars. If they
don’t match, Alex pays Becky 1 dollar.
(a) Find the optimal strategies for Alex and Becky.
(b) Suppose Alex pays Becky k dollars as a compensation before the
game. Find the value of k to make the game fair.
5. Player I and II choose integers i and j respectively where 1 ≤ i, j ≤ 7.
Player II pays Player I one dollar if |i − j| = 1. Otherwise there is no
payoff. Write down the game matrix of the game, find the value of the
game and the optimal strategies for the players.
6. Use the principle of indifference to solve the game with game matrix
1 −2 3 −4
0 1 −2 3
0 0 1 −2
0 0 0 1
Two-person zero sum games 25
(a) Suppose λ1 ≤ 0 and λn > 0. Find the value of the zero sum game
with game matrix D.
(b) Suppose λ1 > 0. Solve the zero sum game with game matrix D.
13. Let
1 −1 0 0 0
−1 −1 1 0 0
A=
0 1 1 −1 0 .
0 0 −1 −1 1
0 0 0 1 −1
(a) Solve Ak , that is, find the maximin strategy, minimax strategy
and value of Ak in terms of k.
Two-person zero sum games 27
0 0 0 · · · r1n
max v
subject to a11 p1 + a21 p2 + · · · + am1 pm ≥ v
a12 p1 + a22 p2 + · · · + am2 pm ≥ v
..
.
a1n p1 + a2n p2 + · · · + amn pm ≥ v
p1 + p2 + · · · + pm = 1
p1 , p2 , · · · , pm ≥ 0
Linear programming and minimax theorem 29
Similarly, to find a minimax strategy for the column player, we need to solve
the following minimax problem
min v
subject to a11 q1 + a12 q2 + · · · + a1n qn ≤ v
a21 q1 + a22 q2 + · · · + a2n qn ≤ v
..
.
am1 q1 + am2 q2 + · · · + amn qn ≤ v
q1 + q2 + · · · + qn = 1
q1 , q2 , · · · , qn ≥ 0
To solve the maximin and minimax problems, first we transform them to a
pair of primal and dual problems.
Definition 2.1.1 (Primal and dual problems). A linear programming prob-
lem in the following form is called a primal problem.
n
P
max f (y1 , · · · , yn ) = cj y j + d
j=1
n
P
subject to aij yj ≤ bi , i = 1, 2, · · · , m
j=1
y1 , y2 , · · · , yn ≥ 0
The dual problem associated to the above primal problem is
m
P
min g(x1 , · · · , xm ) = bi x i + d
i=1
m
P
subject to aij xi ≥ cj , j = 1, 2, · · · , n
i=1
x1 , x2 , · · · , xm ≥ 0
Here x1 , · · · , xm , y1 , · · · , yn are variables, and aij , bi , cj , d, i = 1, 2, · · · , m,
j = 1, 2, · · · , n, are constants. The linear functions f and g are called ob-
jective functions. The primal problem and the dual problem can be written
in the following matrix forms
Primal problem max f (y) = cyT + d
subject to AyT ≤ bT
y≥0
Dual problem min g(x) = xbT + d
subject to xA ≥ c
x≥0
Linear programming and minimax theorem 30
The theorem above has a simple and important consequence that the
primal problem is bounded if the dual problem associated with it has a
feasible vector, and vice verse.
Linear programming and minimax theorem 31
3. If both problems are feasible, then both problems are solvable, that is,
there exists optimal vectors p and q for the dual and primal problems
respectively. Moreover we have f (p) ≤ g(q).
Proof. For the first statement, suppose the primal problem has a feasible
vector q. Then for any feasible vector x of the dual problem, we have g(x) ≥
f (q) by Theorem 2.1.3. Hence the dual problem is bounded. The proof of the
second statement is similar. For the third statement, suppose both problems
are feasible. Then both problems are bounded by the first two statements.
Observe that the set of feasible vectors is closed. It follows that the optimal
values of the objective functions f and g are attainable. Therefore there
exists optimal vectors p and q for the dual and primal problems respectively
and f (q) ≤ g(p) by Theorem 2.1.3.
Furthermore we have the following important theorem in linear program-
ming concerning the solutions to the primal and dual problems.
Theorem 2.1.5. Suppose both the dual problem and the primal problem are
feasible. Then there exist optimal vectors p and q for the dual and primal
problem respectively, and we have
f (q) = g(p)
Proof. We have proved the solvability of the problems. The equality f (q) =
g(p) can be proved using minimax theorem and we omit the proof here.
maximin problem.
max v
subject to a11 p1 + a21 p2 + · · · + am1 pm ≥ v
a12 p1 + a22 p2 + · · · + am2 pm ≥ v
..
.
a1n p1 + a2n p2 + · · · + amn pm ≥ v
p1 + p2 + · · · + pm = 1
p1 , p2 , · · · , pm ≥ 0
max v
subject to pA ≥ v1
p1T = 1
p≥0
is equivalent to
a1j x1 + a2j x2 + · · · + amj xm ≥ 1
and the maximin problem would become a dual problem. We summarize the
above procedures as follows.
2. Let
pi
xi = , for i = 1, 2, · · · , m
v
3. Write down the dual problem
min g(x1 , x2 , · · · , xm ) = x1 + x2 + · · · + xm
subject to a11 x1 + a21 x2 + · · · + am1 xm ≥ 1
a12 x1 + a22 x2 + · · · + am2 xm ≥ 1
..
.
a1n x1 + a2n x2 + · · · + amn xm ≥ 1
To find the minimax strategy for the column player, we need to solve the
following minimax problem.
min v
subject to AqT ≤ v1T
1qT = 1
q≥0
y1 ··· yn −1
x1 a11 ··· a1n b1 = −yn+1
.. .. ... .. .. ..
. . . . .
xm am1 ··· amn bm = −yn+m
−1 c1 ··· cn −d =f
q ··· q q
xm+1 ··· xm+n g
Step 2.
Step 4. Pivot on the entry akl and swap the variables at the pivot row
with the variables at the pivot column. The pivoting operation is
performed as follows.
yl yj yn+k yj
xk a∗ b = −yn+k xm+l 1
a
b
a
= −yl
xi c d = −yn+i −→ xi − ac d − bca = −yn+i
q q q q
xm+l xm+j xk xm+j
Linear programming and minimax theorem 36
Step 5. Go to Step 2.
Definition 2.3.1 (Basic form). A basic form of a pair of primal and dual
problems is a problem of the form
Primal basic form max f (y) = cyT + d
subject to AyT − bT = −(yn+1 , · · · , yn+m )T
y≥0
Dual basic form min g(x) = xbT + d
subject to xA − c = (xm+1 , · · · , xm+n )
x≥0
Theorem 2.3.2. The basic forms before and after a pivot operation are
equivalent.
yl yj
xk a∗ b = −yn+k
xi c d = −yn+i
q q
xm+l xm+j
is equivalent to the system of equations
axk + cxi = xm+l ayl + byj = −yn+k
and
bxk + dxi = xm+j cyl + dyj = −yn+i
−xm+l + cxi = −axk yn+k + byj = −ayl
⇔ and
bxk + dxi = xm+j cyl + dyj = −yn+i
( 1 c 1 b
(
xm+l − xi = xk yn+k + yj = −yl
⇔ a a and a a
bxk + dxi = xm+j cyl + dyj = −yn+i
1 c
a xm+l − a xi = xk
⇔
1 c
b xm+l − xi + dxi = xm+j
a a
1 b
a yn+k + a yj = −yl
and
1 b
c yn+k + yj + dyj = −yn+i
a a
1 c
a xm+l − a xi = xk
⇔
b bc
xm+l + d − xi = xm+j
a a
1 b
a yn+k + a yj = −yl
and
c bc
− yn+k + d − yj = −yn+i
a a
which is equivalent to the tableau
yn+k yj
1 b
xm+l a a
= −yl
xi − ac d− bc
a
= −yn+i
q q
xk xm+j
Linear programming and minimax theorem 38
The above calculation shows that the constraints before and after a pivot
operation are equivalent, and the values of the objective functions f and
g for any given x1 , · · · , xm , xm+1 , · · · , xm+n and y1 , · · · , yn , yn+1 , · · · , yn+m
satisfying the constraints remain unchanged.
For each pair of basic forms, there associates a pair of basic solutions
which will be defined below. Note that the basic solutions are not really
solutions to the primal and dual problems because basic solutions are not
necessarily feasible.
The basic solutions are obtained by setting the independent variables, that
is the variables at the top and at the left, to be 0 and then solving for the
basic variables, that is the variables at the bottom and at the right, by the
constraints.
The basic solutions always satisfy the equalities in the constraints, but
they may not be feasible since some variables may have negative values.
However if both the dual and primal basic solutions are feasible, then they
must be optimal.
1. The basic solution to the primal basic form is feasible if and only if
b1 , b2 , · · · , bm ≥ 0.
Linear programming and minimax theorem 39
2. The basic solution to the dual basic form is feasible if and only if
c1 , c2 , · · · , cn ≤ 0.
2. The basic solution to the dual basic form is (x1 , · · · , xm , xm+1 , · · · , xm+n ) =
(0, · · · , 0, −c1 , · · · , −cn ). Thus it is is feasible if and only if all the vari-
ables are non-negative which is equivalent to c1 , c2 , · · · , cn ≤ 0.
f (0, · · · , 0) = d = g(0, · · · , 0)
Therefore f attains its maximin and g attains its minimum at the basic
solutions.
In practice, we do not write down the basic variables. We would swap the
variables at the left and at the top when preforming pivot operation. One
may find the basic and independent variables by referring to the following
table.
Linear programming and minimax theorem 40
Left Top
xi is independent variable xi is basic variable
xi
yn+i is basic variable yn+i is independent variable
yj is basic variable yj is independent variable
yj
xm+j is independent variable xm+j is basic variable
Solution. Set up the tableau and perform pivot operations successively. The
Linear programming and minimax theorem 41
y1 y2 y3 −1 x1 y2 y3 −1
x1 2∗ 1 1 180 y1 1
2
1
2
1
2
90
x2 1 2 3 300 −→ x2 − 12 3
2
5
2
210
∗
x3 2 2 1 240 x3 −1 1 0 60
−1 6 4 5 −150 −1 −3 1 2 −690
x1 x3 y3 −1 x1 x3 x2 −1
y1 1 − 12 21 60 y1 4
5
− 15 −5 1
36
5∗ 2
−→ 3
x2 1 − 2 2 120 −→ y3 5
− 35 2
5
48
y2 −1 1 0 60 y2 −1 1∗ 0 60
−1 −2 −1 2 −750 −1 − 14
5
1
5
4
− 5 −846
x1 y 2 x 2 −1
3 1 1
y1 5 5
− 5
48
1 3 2
−→ y3 − 5 5 5
84
x3 −1 1 0 60
13 1 4
−1 − 5 − 5 − 5 −858
y2 = y4 = y5 = 0, y1 = 48, y3 = 84, y6 = 60
From the last tableau, the independent variables are x3 , x4 , x6 and the basic
variables are x1 , x2 , x5 . The basic solution is
13 4 1
x3 = x4 = x6 = 0, x1 = , x2 = , x5 =
5 5 5
Linear programming and minimax theorem 42
2. The value of the number in the lower right corner is always equal to
the sum of those entries in the lower row which associate with xi ’s at
the top row (and similarly equal to the sum of those entries at the
rightmost column associate with yj ’s at the leftmost column.)
3. The value of the number in the lower right corner never increases.
Finally, one may also check that the result should satisfy the following two
conditions.
Example 2.3.6. Solve the two-person zero sum game with game matrix
−1 5 3 2
6 −1 0 4
y1 y2 y3 y4 −1 x2 y2 y3 y4 −1
x1 0 6 4 3 1 x1 0 6∗ 4 3 1
−→
x2 7∗ 0 1 5 1 y1 1
7
0 17 75 1
7
−1 1 1 1 1 0 −1 − 17 1 7 7 − 17
6 2
x2 x 1 y3 y4 −1 x2 x1 y2 y4 −1
1 2∗ 1 1 1 3 3 1
y2 0 y3 0
−→ 6 3 2 6 −→ 4 2 4 4
y1 1
7
0 1
7
5
7
1
7
y1 − 17 1
− 28 3
− 14 17
28
3
28
−1 − 17 − 16 4
21
3
− 14 − 13
42
−1 − 17 3
− 14 − 72 5
− 14 5
− 14
5
The optimal value is d = . Therefore a maximin strategy for the row
14
player is
1 14 3 1 3 2
p = (x1 , x2 ) = , = ,
d 5 14 7 5 5
A minimax strategy for the column player is
1 14 3 1 3 7
q = (y1 , y2 , y3 , y4 ) = , 0, , 0 = , 0, , 0
d 5 28 4 10 10
Example 2.3.7. Solve the two-person zero sum game with game matrix
2 −1 6
A= 0 1 −1
−2 2 1
y1 y2 y3 −1 x1 y2 y3 −1
1 1 1
x1 4∗ 1 8 1 y1 4 4
2 4
5∗
x2 2 3 1 1 −→ x2 − 21 2
−3 12
x3 0 4 3 1 x3 0 4 3 1
−1 1 1 1 0 −1 − 41 3
4
−1 − 14
x1 x2 y3 −1
3
y1 10
− 10 23
1
10
1
5
1 2 6 1
−→ y2 − 5 5
− 5 5
4
x3 5
− 58 39
5
1
5
1 3 1
−1 − 10 − 10 − 10 − 25
Linear programming and minimax theorem 45
1
One sees that the row player may guarantee that his payoff is at least by
2
1 3
using p = , , 0 and the column player may guarantee that the payoff
4 4
1 1 1
to the row player is at most by using q = , ,0 .
2 2 2
Linear programming and minimax theorem 46
The row value νr (A) of a game matrix A is the largest payoff of the
row player that he may guarantee himself. The column value νc (A) of A is
the least payoff that the column player may guarantee that the row player
cannot surpass. The strategies for the players to achieve these goals are
called maximin and minimax strategies.
Definition 2.4.2 (Maximin and minimax strategies). Let A be an m × n
matrix.
1. A maximin strategy is a strategy p ∈ P m for the row player such
that
minn pAyT = maxm minn xAyT = νr (A)
y∈P x∈P y∈P
It can be seen readily that we always have νr (A) ≤ νc (A) for any matrix
A and we give a rigorous proof here.
1
Note that since the payoff function π(x, y) = xAyT is continuous and the sets P m , P n
are compact, that is closed and bounded, the payoff function attains its maximum and
minimum by extreme value theorem.
Linear programming and minimax theorem 47
νr (A) ≤ νc (A)
= minn pAyT
y∈P
≤ pAqT
≤ maxm xAqT
x∈P
Conv({x1 , x2 , · · · , xk })
X k k
X
n
= {x ∈ R : x = λi xi with λi ≥ 0 for all i and λi = 1}
i=1 i=1
Linear programming and minimax theorem 48
The convex hull of a set of vectors can also be defined as the smallest
convex set which contains all vectors in the set.
To prove the minimax theorem, we prove a lemma concerning properties
of convex sets. Recall that the standard inner product and the norm on Rn
are defined as follows. For any x = (x1 , x2 , · · · , xn ), y = (y1 , y2 , · · · , yn ) ∈
Rn ,
1. hx, yi = x1 y1 + x2 y2 + · · · + xn yn
p p
2. kxk = hx, xi = x21 + x22 + · · · + x2n
The following lemma says that we can always use a plane to separate the
origin and a closed convex set C not containing the origin. It is a special
case of the hyperplane separation theorem2 .
We are going to prove that the two statements in the theorem correspond to
the two cases 0 6∈ C and 0 ∈ C.
Case 1. Suppose 0 6∈ C. Then by Lemma 2.4.6, there exists z =
(z1 , z2 , · · · , zm ) ∈ Rm such that
hz, yi > 0 for any y ∈ C
In particular, we have
hz, ei i = zi > 0 for any i = 1, 2, · · · , m
Then we may take
z
x= ∈ Pm
z1 + z2 + · · · + zm
and we have
hz, aj i
hx, aj i = > 0 for any j = 1, 2, · · · , n
z1 + z2 + · · · + zm
which means xA > 0. Let α > 0 be the smallest coordinate of the
vector xA and we have
νr (A) ≥ minn xAyT ≥ α > 0
y∈P
and we have
λn+1
1 ..
AyT = − . ≤0
λ1 + λ2 + · · · + λn
λn+m
which implies
vc (A) ≤ maxm xAyT ≤ 0
x∈P
Now we give the proof of the minimax theorem (Theorem 1.1.10) which
can be stated in the following form.
Theorem 2.4.8 (Minimax theorem). For any matrix A, the row value and
columns value of A are equal. In other words, we have
νr (A) = νc (A)
Proof. It has been proved that νr (A) ≤ νc (A) for any matrix A (Theorem
2.4.3). We are going to prove that νc (A) ≤ νr (A) by contradiction. Suppose
there exists matrix A such that νr (A) < νc (A). Let k be a real number such
that νr (A) < k < νc (A). Let A0 be the matrix obtained by subtracting every
entry of A by k. Then νr (A0 ) = νr (A) − k < 0 and νc (A0 ) = νc (A) − k > 0
which is impossible by applying Theorem 2.4.7 to A0 . The contradiction
shows that νc (A) ≤ νr (A) for any matrix A and the proof of the minimax
theorem is complete.
Exercise 3
1. Solve the following primal problems. Then write down the dual prob-
lems and the solutions to the dual problems.
(a)
max f = 3y1 + 5y2 + 4y3 + 12
subject to 3y1 + 2y2 + 2y3 ≤ 15
4y2 + 5y3 ≤ 24
(b)
max f = 2y1 + 4y2 + 3y3 + y4
subject to 3y1 + y2 + y3 + 4y4 ≤ 12
y1 − 3y2 + 2y3 + 3y4 ≤ 7
2y1 + y2 + 3y3 − y4 ≤ 10
Linear programming and minimax theorem 53
2. Solve the zero sum games with the following game matrices, that is
find the value of the game, a maximin strategy for the row player and
a minimax strategy for the column player.
2 −3 3 2 0 −2
(a) −2 3 1 (d) −1 −3 3
1 1 5 −2 2 0
1 −1 1
3 1 −5 −2 0 −1
(e)
(b) −1 −2 6 1 −2 2
−2 −1 3 −1 1 −2
−3 2 0
3 0 1 1 −2 −1
(f)
(c) −1 2 −2 −1 0 2
0 1 −1 1 1 −3
3. Prove that if C1 and C2 are convex sets in Rn , then the following sets
are also convex.
(a) C1 ∩ C2
(b) C1 + C2 = {x1 + x2 : x1 ∈ C1 , x2 ∈ C2 }
C = Conv({a1 , a2 , · · · , an , e1 , e2 , · · · , em })
objective function, 29
54