0% found this document useful (0 votes)
8 views

Chapter 5_ Multivariate unconstrained Optimization

Chapter V discusses multivariate unconstrained optimization, focusing on the definitions of convexity and concavity for functions. It outlines the first-order and second-order conditions for determining maxima and minima in both univariate and bivariate cases, including examples from economics. The chapter also explains the role of the Hessian matrix in identifying the nature of stationary points in optimization problems.

Uploaded by

Mostafa Allam
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
8 views

Chapter 5_ Multivariate unconstrained Optimization

Chapter V discusses multivariate unconstrained optimization, focusing on the definitions of convexity and concavity for functions. It outlines the first-order and second-order conditions for determining maxima and minima in both univariate and bivariate cases, including examples from economics. The chapter also explains the role of the Hessian matrix in identifying the nature of stationary points in optimization problems.

Uploaded by

Mostafa Allam
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 15

Chapter V: Multivariate unconstrained Optimization

In the univariate case where y f x , the curvature of the function is


//
described by using the second-order derivative f x that explains how the
//
first derivative changes. The sign of f x indicates it f x is concave or
convex according to the following:
//
Definition (1): A twice differential function f x is convex if f x 0 at
all points on its domain.

Definitions (2): A twice differential function f x is strictly convex if


f // x 0 at all points on its domain except possibly at a single point.
//
Definition (3): A twice differential function f x is concave if f x 0 at
all points on its domain.

Definition (4): A twice differential function f x is strictly concave if


f // x 0 at all points on its domain except possibly at a single point.

Single variable optimization: Maxima and Minima:


To determine the optimum of a univariate function f x , (a minimum or a
maximum) at a point x* , we have:
/
1) The first derivative f x 0 first order condition (necessary).

122
//
2) If f x 0, x* is a maximum point.
//
If f x* 0, x* is a minimum point.

Second-order condition (sufficient).

& &

Inflection point: it is the point where the function f x changes its type of

curvature. f
//
x changes its sign at x* , where f // x* 0

Ex: A manufacturer estimates that when q thousand units of a particular


commodity are produced each month, the total cost will be
C q 0.4q2 3q 40
thousand dollars, and all q units can be sold at a price
of p(q) 22.2 1.2q dollars per unit. Determine the level of production that
results in max profit. What is the maximum profit?

123
Sol: The profit q R q C q ,

Where R q p q 22.2 1.2q q 22.2q 1.2q 2

Therefore, q 22.2q 1.2q 2 0.4q 2 3q 40

1.6q 2 19.2q 40

To determine the value of q that maximizes q :


/
q 3.2q 19.2 0

3.2q 19.2 , q* 6 thousand units


//
q 3.2 0 so q* 6 is the level of production that

maximizes q as it is a strictly concave function.


The maximum profit is:
2
6 1.6 6 19.2 6 40 17.6 thousand dollars

Bivariate unconstrained optimization:

The idea of optimization is fundamental in economics, and the mathematical


methods of optimization underline most economic models. We have already
studied optimization for functions in one variable and we extend the idea of
optimization for functions of two variables.

Consider y f x1 , x2 , the gradient vector f and the Hessian matrix H


are given by:

f1 f11 f12
f , H
f2 f 21 f 22

124
* *
For a stationary point x1 , x2 , the first-order condition for an optimum (max or
min) is that:

f1 x1* , x2* 0
f
f 2 x1* , x2* 0

* *
The stationary point x1 , x2 could be min, max or saddle point (it is neither a
max nor a min as it a max with respect to changes in some of the xi values
values and minimum to the others).
For univariate functions, the second-order derivative plays an important role in
determining whether the point at which the first derivative is zero would be a
max, min or inflection point. In multivariate optimization problems, the Hessian
matrix plays a similar role. The second-order condition depends on the Hessian
matrix and whether it is positive or negative (semi) definiteness.
First-order conditions:
Let y f x1* , x2* , then the first-order condition (FOCs) for a stationary point is

f1 x1* , x2* 0
f
f 2 x1* , x2* 0

Ex 1: Let y f x1 , x2 10 x12 x22 , the FOCs are


f1 20 x1 0
f2 2 x2 0
* *
Thus, the point x1 , x2 0,0 is a stationary point of f ( x1 , x2 ) and using the
second-order conditions later on, it is confirmed that the point 0,0 is a
maximum point.

Ex 2: Let y f x1 , x2 x12 4 x22 2 x1 16 x2 x1 x2


The FOCs are: f1 2 x1 2 x2 0 ……. (1)

f2 8x2 16 x1 0 ……. (2)


125
Solving the two equations simultaneously, we have from (1) x2 2 2 x1 and
substituting in (2).

8 2 2 x1 16 x1 0, 15x1 0, x1 0 and x2 2 2 0 2
* *
Thus, the point x1 , x2 0,2 is a stationary point.

Ex 3: Multiproduct monopoly:

A monopoly produces two outputs, x1 and x2 , with linear demand functions:

x1 100 2 p1 p2

x2 120 3 p1 5 p2
These demand functions may be expressed as prices functions in quantities:
2 p1 p2 100 x1
3 p1 5 p2 120 x2
2 1 p1 100 x1
Thus
3 5 p2 120 x2

Solving for p1 and p2 by using the inverse of a matrix or Cramér rule we have

p1 88.57 0.71x1 0.14 x2

p2 77.14 0.43x1 0.29 x2


These equations are called the inverse demand functions. Since each output
enters with a negative sign in the inverse demand function of the other, the two
good are substitutes.
Assuming the firm's cost function is

C x1, x2 50 10 x1 20 x2

Thus, the firm's profit function is

x1, x2 R x1 , x2 C x1 , x2 p1 x1 p2 x2 C x1 , x2

126
Thus,
x1 , x2 88.57 0.71x1 0.14 x2 x1
77.14 0.43x1 0.29 x2 x2 50 10 x1 0.29 x2

78.57 x1 57.14 x2 0.71x12 0.29 x22 0.57 x1 x2 50


The FOCs are given by:

1 x1 , x2 78.57 1.42 x1 0.57 x2 0

2 x1 , x2 57.14 0.58x2 0.57 x1 0


These two equations are solved simultaneously to give the optimal output values
x1* , x2* :

x1* 26.069 x2* 72.897


To find the corresponding prices, we substitute into the inverse demand functions
to obtain:

p1* 52.86 p2* 44.79


* * * *
The maximum profit is obtained by substituting the values of x1 , x2 , p1 and p2
*
into the profit function to get 3056.81$
We assume that the stationary point x1 , x2 maximizes the profit function as we
got only the SOCs. We have to check using the second-order conditions that will
be discussed later on. The maximum profit takes the following form:

26.069

127
Ex 4: Cournot Duopoly:
Two firms produce identical output and sell into a market with the linear demand
function in the two outputs q1 and q2 :

P 100 q1 q2
Assuming that each firm's production cost is zero, each firm wants to maximize
its profit, given by

i pqi 100qi q1 q2 qi , i 1,2

Notice that the market price p depends on the total output of the two firms
q1 q2 , thus the profit of each firm depends on how much the other firm
produces and also on its own level of output. This duopoly is a special case of
oligopolies where there is a close interdependence between the profit of one firm
and the output of the others.
Now the Cournot model is:

(1) pq1 100q1 q1 q2 q1

2
pq2 100q2 q1 q2 q2

The difficulty her is that each firm needs to figure out how much the other firm
will produce since the level of output of each firm will affect the profit of the
other firm.
The French economist Augustin Cournot solved the problem by assuming that
each firm takes the output of the other firm as a given parameter when choosing
its own output, and then market equilibrium is obtained by solving the resulting
pair of simultaneous equations.
Note that the assumption of Cournot is consistent with the notion of the partial
derivative. Therefore, each firm will have its own FOC.

d 1
1 1 100 2q1 q2 0
dq1

128
d 2
2 2 100 2q2 q1 0
dq2
Thus, solving the two equations yields:

q1* q2* 33.33

and p* 100 q1* q2*

100 66.67 33.33$


Therefore, the firms end up equally sharing the market and this is expected from
the symmetry of the problem.

* 100 q2* 100 q1


Note that the solution gives the equations q 1 & q2*
2 2
These two equations give the output at one firm which is optimal (profit-
maximizing) for each level of output of the other firm. These are known as Best
response or Reaction functions for firm (1) and (2).
The point at which these two lines intersect gives the Cournot equilibrium
illustrated in the following chart:

129
Second order conditions (SOCs)

The first-order conditions cannot in themselves distinguish between maximum


values, minimum values and saddle points, because they hold at each of these.
The second-order conditions can tell us if a point satisfying the first-order
conditions (FOCs) is a maximum, minimum or a saddle point.
The second-order conditions (SOCs) relate to whether the Hessian matrix (the
matrix of second-order partial derivatives) is positive or negative definite.
Thus, for unconstrained optimization problems, we first find all the stationary
points of the function using the first-order conditions (FOCs) and then evaluate
the second-order conditions (SOCs) at each of the stationary points to determine
the type of this point; Max (Local-Global), Min (Local - Global) or saddle point
using the following theorem:

Theorem: Let f x be a twice differentiable function of n variables.


Suppose that x* is a stationary point of f x

1) If the Hessian matrix at x x* is negative definite, then the function f x is


*
concave and x is a local maximum of f x . Further, if f x is strictly
*
concave function, then x is a global max of f x .

2) If the hessian matrix at x x* is positive definite then the function f x is


*
convex and x is a local minimum of f x . Further if f x is a strictly
*
convex function, then x is a global min of f x .
* *
3) If the Hessian matrix at x is indefinite, then x is neither a local max nor a local
*
min of f x and x is a saddle point

The following theorem provides conditions for a function f x to be concave,


convex, strictly concave and strictly convex.

131
Theorem:
For any function y f x , x R n , which is twice differentiable with Hessian
matrix H or 2 f it follows that:

1) The function f x is concave if and only if the Hessian 2 f is negative


semi-definite for all x .
2) The function f x is strictly concave if the Hessian 2 f is negative definite
for all x .
3) The function f x is convex if and only if the Hessian 2 f is positive semi-
definite for all x .
4) The function f x is strictly convex if the Hessian 2 f is positive definite
for all x .
Applying the two theorems, one can determine if a function f x has at a
stationary point x* a Max (Local or Global), a Min (Local or Global) or a saddle
point.
Determination of Max and Min (Local - Global) or saddle Point:
The Hessian The function The stationary
Pattern's Example
Matrix f x point x*
Negative semi
H1 0, H 2 0, H 3 0... Concave Local Max
definite
Negative H1 0, H 2 0, H 3 0...
Strictly concave Global Max
Definite for all x
Positive semi-
H1 0, H 2 0, H3 0... Convex Local Min
Definite
H1 0, H 2 0, H3 0...
Positive Definite Strictly convex Global Min
for all x
H1 0, H 2 0, H 3 0...
Indefinite - Saddle Point
H1 0, H 2 0, H 3 0...

131
Ex 1: For each of the following functions, determine the stationary point (s) and
specify if it is a max, min or saddle point.

a) y f x1 , x2 2 x12 x22

We start by deriving the FOCs:

df df
f1 4 x1 0, f2 2 x2 0
dx1 dx2
* *
Thus, the stationary point is x1 , x2 0,0 .

To determine the nature of this point, the Hessian matrix is derived:

f11 f12 4 0
H 2 f
f 21 f 22 0 2

The Hessian matrix is positive definite for all x R n as


4 0
H1 4 0 H2 8 0
0 2
Since the Hessian matrix is positive definite at all values of x, the function is
strictly convex and the point 0,0 is a minimum (global) for f x .

b) y f x1, x2 4 x1 2 x2 x12 x22 x1 x2

The FOCs are:

f1 4 2 x1 x2 0

f2 2 2 x2 x1 0

Thus, 2 x1 x2 4

x1 2 x2 2

* * 10 8
Solving the two equations we have x1 , x2 , .
3 3

132
The Hessian matrix is

f11 f12 2 1
H 2 f
f 21 f 22 1 2
Therefore

2 1
H1 2 2 0 H2 3 0
1 2
The Hessian matrix is negative definite for all values of x, so the point
10 8
x1* , x2* , is a maximum (global).
3 3

c) y f x, y x3 y 3 9 xy

To get the stationary point(s) we get the FOCs

fx 3x 2 9y 0 ………. (1)

fy 3y2 9x 0 ………. (2)

To solve both equations simultaneously:

1 2
From equation (1) we have 9 y 3x 2 , y x .
3
Substituting into equation (2), we have:

2
1 2 1 4
3 x 9 x 0, x 9x 0
3 3

Therefore, x4 27 x 0

x x3 27 0, x x3 27 0

133
Thus, x 0, x 3

and the two stationary points are.

1 2
x 0, y 0 0 x* , y * 0,0
3

1 2
x 3, y 3 3 x* , y * 3, 3
3

To determine the nature of these two stationary points, we calculate the Hessian
matrix.

f xx f xy 6x 9
H 2 f
f yx f yy 9 6y

H is not the same for all and y .


Notice that x

Every point will be examined separately.

For (0,0)

0 9
H 2 f
9 0

Thus H 1 0 ,H2 81

Since the signs of the leading principal minors do not follow the recognized
patterns for positive (negative) (semi) definiteness, so H is indefinite and the
point 0,0 is neither a max nor a min, it is a saddle point.

134
For (3, -3)

18 9
H 2f
9 18

Thus H1 18 0 , H2 243 0

Therefore, the Hessian matrix is positive definite at that point, thus 3, 3 is a


local min for f x .

We have no information that the Hessian matrix is positive definite for all values
of x, y , therefore, this function is not strictly convex.

d) f x, y, z 2x2 y2 4z2 x 2z

The FOCs are

fx 4x 1 0

fy 2y 0

fz 8z 2 0
Which implies the following solution (stationary point):

1 1
x y 0 z
4 4
1 1
x* , y * , z * ,0,
4 4

To check the nature of this point, the Hessian matrix is calculated:

f xx f xy f xz 4 0 0
H 2 f f yx f yy f yz 0 2 0
f zx f zy f zz 0 0 8

135
4 0
Since H 1 4 0 H2 8 0
0 2

4 0 0
H3 0 2 6 64 0
0 0 8

1 1
Therefore, H is positive definite at the point ,0, and this point is a
4 4
minimum. H also is strictly convex as it is positive definite for all x, y.z , then
this point is a global min.

136

You might also like