Mathematics For Economics (ECON 104)

Download as pdf or txt
Download as pdf or txt
You are on page 1of 46

Mathematics for Economics (ECON

104)

Lecture 8: Unconstrained Optimization III (Ch. 13) and


Optimization with Equality Constraints (Ch. 14)
Example [Price discrimination]: Suppose a airline firm has
two types of customers, one for business trips (market 1) and
the other for tourists (market 2). Tourists are more sensitive to
price, and thus the demand curves are

q1 = 100 − p1,
q2 = 100 − 2p2.
The cost function of the firm is

C(Q) = 20Q, where Q = q1 + q2.

2
Rewrite the demand curves as
1
p1 = 100 − q1 and p2 = 50 − q2
2
The profit is then

π(q1, q2) = |p1q1 + p q − 20(q1{z+ q2)}


{z 2 2} |
revenue cost
1
 
= (100 − q1) q1 + 50 − q2 q2 − 20(q1 + q2).
2

3
The first-order condition is

π10 (q1, q2) = |100 {z


− 2q1} − |{z}
20 = 0
M R1 MC
π20 (q1, q2) = |50 {z
− q2} − |{z}
20 = 0.
M R2 MC

This gives us

q1∗ = 40, q2∗ = 30


p∗1 = 60, p∗2 = 35.

The maximal profit is

π(q1∗ , q2∗ ) = p∗1q1∗ + p∗2q2∗ − 20(q1∗ + q2∗ )


= (60)(40) + (35)(30) − 20(40 + 30) = 2050.

4
The second-order conditions are
00 = −2, π 00 = π 00 = 0, π 00 = −1.
π11 12 21 22
Thus
!
−2 0
H(q1, q2) = .
0 −1
00
and so π11 = −2 < 0 and |H(q1, q2)| = 2 > 0.

This means that H(q1∗ , q2∗ ) is negative definite (ND) and thus we
obtained the local maximum.

5
Sufficient Conditions for Global Optimum (Ch. 13.2)

Theorem (Single Variable): Suppose that f is a twice differ-


entiable function defined on an interval I, and that a stationary
point x∗ is in the interior of I.

(1) If f is concave on I (or equivalently f 00(x) ≤ 0 for all x ∈ I),


then x∗ is a “global” maximum point of f in I.

(2) If f is convex on I (or equivalently f 00(x) ≥ 0 for all x ∈ I),


then x∗ is a “global” minimum point of f in I.

6
Sufficient Conditions for Global Optimum (continued)

Theorem (Two Variables): Suppose that the function f has


continuous partial derivatives of first- and second-order in a con-
vex set S. Let (x∗, y ∗) be a stationary point of f in the interior
of S.

• If f is concave in S (or equivalently if H(x, y) is negative


semidefinite (NSD) for all (x, y) ∈ S), then (x∗, y ∗) is a global
maximum point of f in S.

• If f is convex in S (or equivalently if H(x, y) is positive


semidefinite (PSD) for all (x, y) ∈ S), then (x∗, y ∗) is a global
minimum point of f in S.

7
Convex set: When we draw a line segment connecting any two
points in a set, if all the points in the segment belong to the set,
then the set is called “convex.”

Note that concave or convex functions need to be defined over


a convex set.

8
Summary: Conditions for Global max and Global min

• Single Variable:
Condition Max Min
1st-order Necessary f 0(x∗) = 0 f 0(x∗) = 0
f is concave ⇔ f is convex ⇔
2nd-order Sufficient
f 00(x) ≤ 0 for all x f 00(x) ≥ 0 for all x

• Two Variables: (x, y)


Cond. Max Min
0 0
Nec f10 (x∗, y ∗) = f2(x∗, y ∗) = 0 f10 (x∗, y ∗) = f2(x∗, y ∗) = 0
f is concave ⇔ f is convex ⇔
Suff
H(x, y) is NSD for all (x, y) H(x, y) is PSD for all (x, y)

9
Example: Once again, we revisit f (x, y) = −2x2 − 2xy − 2y 2 +
36x + 42y − 158.

The first-order conditions are

f10 (x, y) = −4x − 2y + 36 = 0


f20 (x, y) = −2x − 4y + 42 = 0.
Hence, the stationary point is (x∗, y ∗) = (5, 8).

10
The second-order conditions are
00 = −4, f 00 = f 00 = −2, f00
f11 12 21 22 = −4,
which generates
!
−4 −2
H(x, y) = .
−2 −4
00
Thus f11 = −4 < 0 and |H(x, y)| = 12 > 0 for all (x, y).

So, f is concave. The point (5, 8) is a global max point.

11
Optimization with Equality Constraint (Ch. 14.1)

Until now, we have considered only the optimization problems


“without” constraints.

Most cases of economic problems, however, involve some con-


straints.

For example, consumers maximize their utility given the budget


constraint. Producers minimize their production costs given the
quantities they produce.

12
The Lagrangian Method (14.1, 14.4)

Consider the problem of maximizing a function f (x, y) when x


and y are restricted to satisfy an equality constraint: g(x, y) = c;

max f (x, y) subject to g(x, y) = c.


(x,y)∈R2

The constraint set is the set of all pairs (x, y) for which g(x, y) =
c:
n o
2
(x, y) ∈ R g(x, y) = c

13
Heuristic Derivation of the Lagrangian Method (Ch. 14.4)

Consider
max f (x, y) subject to g(x, y) = c.
(x,y)∈R2

Let (x∗, y ∗) be a local maximum point of f to the above con-


strained optimization problem. So, we must have g(x∗, y ∗) = c.

We consider a pair of “small” numbers (∆x, ∆y) ∈ R2 such that


g(x∗ + ∆x, y ∗ + ∆y) = g(x∗, y ∗). Then, we have

∆g = g(x∗ + ∆x, y ∗ + ∆y) − g(x∗, y ∗)


0 0
∗ , y ∗ )∆x + g (x∗ , y ∗ )∆y = 0.

|{z} g 1 (x 2
linear approx
14
0
Assuming that g1(x∗, y ∗) 6= 0, we derive
0
g2(x∗, y ∗)
∆x = − 0 ∆y. (∗)
g1(x∗, y ∗)

Since (x∗, y ∗) is a local maximum point to the constrained opti-


mization problem,

0 ≥ f (x∗ + ∆x, y ∗ + ∆y) − f (x∗, y ∗)


0 0

|{z} f1(x , y )∆x + f2(x∗, y ∗)∆y
∗ ∗
linear approx
0
 
∗ ∗
f (x , y ) 0 ∗ ∗ 0
= − 1
0 g2(x , y ) + f2(x∗, y ∗) ∆y. (∵ (∗))
g1(x∗, y ∗)

15
Since ∆y could be positive or negative, we must have
0
f1(x∗, y ∗) 0 0
− 0 g2(x , y ) + f2(x∗, y ∗) = 0.
∗ ∗ (∗∗)
g1(x∗, y ∗)
Define
0
f1(x∗, y ∗)
λ∗ ≡ 0 .
g1(x∗, y ∗)
Then, (∗∗) can be translated into:
0 0
f1(x∗, y ∗) = λ∗g1(x∗, y ∗),
0 0
f2(x , y ) = λ g2(x∗, y ∗).
∗ ∗ ∗

16
Define the Lagrangian function:

L(x, y, λ) = f (x, y) − λ (g(x, y ) − c),


where λ ∈ R is called the Lagrange multiplier.

Notice that the necessary conditions are the FOCs of L:

L01(x, y, λ) = f10 (x, y) − λg10 (x, y) = 0


L02(x, y, λ) = f20 (x, y) − λg20 (x, y) = 0
L03(x, y, λ) = g(x, y) − c = 0 (constraint).
First, two equations are the 1st-order derivatives with x and y,
and the last equation is the constraint.

Hence, the FOCs (necessary conditions) in max problem with


equality constraints can be obtained by using Lagrangian method.
17
Example: Consider

max xy subject to 2x + y = 100.


(x,y)∈R2
+

We set
L(x, y, λ) = xy − λ(2x + y − 100).
Compute the FOCs of L:
0
L1(x, y, λ) = y − 2λ = 0,
0
L2(x, y, λ) = x − λ = 0,
0
L3(x, y, λ) = 2x + y − 100 = 0.

The first two equations imply y = 2λ and x = λ so that y = 2x.


Plugging this into the constraint set, we obtain x = 25, y = 50,
and λ = x = 25.
18
Theorem [Necessity for Lagrangian Method]: Suppose that
f and g are continuously differentiable functions in a domain S,
and that (x∗, y ∗) is both an interior point of S and a local max
or min point of f (x, y) subject to g(x, y) = c. Assume also that
the following constraint qualification is satisfied: for any point
(x, y) satisfying g(x, y) = c (we call it a feasible point),

(g10 (x, y ) , g20 (x, y )) 6= (0, 0).


Then, there is a unique number λ ∈ R such that the Lagrangian

L(x, y, λ) = f (x, y) − λ (g (x, y ) − c)


has a stationary point at (x∗, y ∗): i.e. there exists λ such that

L01 = f10 (x∗, y ∗) − λg10 x∗, y ∗ = 0,




L02 = f20 (x∗, y ∗) − λg20 x∗, y ∗ = 0.




19
What does this mean? The first-order necessary conditions
in max or min problem with an equality constraint are exactly
the first-order conditions in Lagrangian. Thus, we can use La-
grangian to find the stationary points in max or min problem.

Constraint Qualification: The above result would not be pos-


sible to derive if both g10 and g20 were zero. For this reason, we
need to make the assumption that, at any feasible point (x, y)
satisfying g(x, y) = c,
0 0
 
g1(x, y), g2(x, y) =6 (0, 0).

Remark: If the constraint is linear, as in consumer’s problem


(pxx+py y = I), then this constraint qualification is automatically
satisfied.
20
What if (g10 , g20 ) = (0, 0)?

Example: Consider the problem:

max f (x, y) = −y subject to g(x, y) = y 3 − x2 = 0.


x,y

The constraint implies

y 3 = x2 ≥ 0 ⇒ y ≥ 0
⇒ −y ≤ 0.
Thus, f is maximized at y ∗ = 0 and from the constraint, x∗ = 0.
Note that (g10 , g20 ) = (−2x, 3y 2) = (0, 0) when (x∗, y ∗) = (0, 0).

Clearly, the objective function has the maximum at (x∗, y ∗) =


(0, 0).
21
Example (continued): The solution “cannot” be found through
the Lagrangian method. To see this, form

L = −y − λ(y 3 − x2).
First-order conditions are

L01 = 2λx = 0,
L02 = −1 − 3λy 2 = 0,
y 3 − x2 = 0.
From the first equation, (i) λ = 0 or (ii) x = 0. (i) If λ = 0, then
the second equation does not hold. (ii) If x = 0, then the third
equation implies that y = 0, and once again the second equation
does not hold.

22
Example: Solve the following problem:
max or min x2y s.t. 2x2 + y 2 = 3
(x,y)∈R2 (x,y)∈R2

The objective function is continuous because it is the product of


polynomials. x and y are drawn from the set D:
D = {(x, y) ∈ R2 : 2x2 + y 2 = 3}.
First, D is closed. Second, define
B = {(x, y) ∈ R2| x2 + y 2 ≤ 3}.

We claim that D ⊆ B, (i.e., D is bounded). This is because for


any (x, y) ∈ D,
x2 + y 2 = 2x2 + y 2 − x2 = 3 − x2 ≤ 3 ⇒ (x, y) ∈ B
So, D is closed and bounded so that it is compact.
23
Thus, by the extreme-value theorem, there exist both the min
point and max point.

Set up
L(x, y, λ) = x2y − λ(2x2 + y 2 − 3).
First-order conditions are

L01 = 2xy − 4λx = 2x(y − 2λ) = 0,


L02 = x2 − 2λy = 0,
L0λ = −(2x2 + y 2 − 3) = 0.

24
From the first equation, we get

x = 0 or y = 2λ.

√ √
Case 1: If x = 0, by the constraint, y = 3 (λ = 0) or − 3
(λ = 0) from the second equation. Thus we have two solution
candidates:

x = 0, y = 3 and λ = 0

x = 0, y = − 3 and λ = 0.

Case 2: if y = 2λ, then λ = y/2. Plug this into the second


equation so that x2 = y 2.

25
Plugging this into the third equation (i.e., x2 = 1), we get
1
x = 1, y = 1 and λ =
2
1
x = −1, y = −1 and λ = −
2
1
x = 1, y = −1 and λ = −
2
1
x = −1, y = 1 and λ = .
2

26
Thus, we have six solution candidates that satisfy the first-order
conditions:
 √   √ 
(x, y, λ) = 0, 3, 0 with f 0, 3 = 0
 √   √ 
(x, y, λ) = 0, − 3, 0 with f 0, − 3 = 0
1
 
(x, y, λ) = 1, 1, with f (1, 1) = 1
2 
1

(x, y, λ) = 1, −1, with f (1, −1) = −1
2
1
 
(x, y, λ) = −1, 1, with f (−1, 1) = 1
2
1
 
(x, y, λ) = −1, −1, − with f (−1, −1) = −1.
2

27
Constraint qualification: g10 = g20 = 0 occurs “only” at (0, 0).
But (0, 0) is not a feasible point.

Thus, the constraint qualification is satisfied.

By the necessity of the Lagrangian method and extreme value


theorem, the max points are (1, 1) and (−1, 1), and the min
points are (1, −1) and (−1, −1).

28
Interpretation of Lagrange multiplier (λ) (Ch. 14.2)

Consider the problem

max or min f (x, y) subject to g(x, y) = c.


x,y x,y
The solutions x∗ and y ∗ depend on the parameter c : x∗(c) and
y ∗(c).

Assume that x∗(c) and y ∗(c) are differentiable functions of c.

Then the value of f at the solution is


∗ ∗ ∗ 
f (c) = f x (c), y (c) .

29
Differentiate f ∗(c) with respect to c :

df ∗ (c) 0 ∗ ∗  dx∗(c) 0 ∗ ∗  dy ∗(c)


= f1 x (c), y (c) + f2 x (c), y (c)
dc dc dc
∗ ∗
" #
0 ∗ ∗  dx (c) 0 ∗ ∗  dy (c)
= λ g1 x (c), y (c) + g2 x (c), y (c) .
dc dc

The second equality follows from the FOC of the Lagrangian L.

30
Now look at the constraint. Since
∗ ∗ 
g x (c), y (c) = c for all c,

the derivatives of both sides are


0 ∗ ∗  dx∗(c) 0 ∗ ∗  dy ∗(c)
g1 x (c), y (c) + g2 x (c), y (c) = 1.
dc dc

Hence,
df ∗ (c)
= λ.
dc

31
The value of the Lagrange multiplier at the solution is the rate
of change in the optimal value of the objective function as
the constraint is relaxed by one unit.

For example, in a utility maximization problem, the optimal value


of the Lagrange multiplier measures the marginal utility of in-
come.

32
Example: Consider a simple problem:

max f (x) = x2 s.t. x = c.


x
The solution is x∗ = c. Thus the value at the solution is

f ∗(c) = f (x∗ = c) = c2.

We must have the following:


df ∗(c)
= 2c.
dc

33
To show this using the Lagrangian method, set up

L = x2 − λ(x − c).
First-order conditions are

2x − λ = 0 and x = c.
Thus we obtain λ = 2c, which equals
df ∗(c)
.
dc

34
Sufficient Conditions for Local Optimum (Ch. 14.5)

Consider the problem

max f (x, y ) subject to g (x, y ) = c.


x,y

Assume that g20 (x, y) 6= 0. Then, by the implicit function theorem,


the equation g (x, y ) = c defines y implicitly as a differentiable
function of x: So, we can set y = y(x).

We thus have
dy g10
=− 0.
dx g2

36
The objective function can then be written as a function of x:

max F (x) = f (x, y(x)).


x

Now from the first-order condition, we get a stationary point of


F, which is x∗: F 0(x∗) = 0.

A sufficient condition for x∗ to be a local max point of F is

F 00(x∗) < 0.

37
In other words, given F (x) = f (x, y(x)), it follows that, at the
point (x∗, y(x∗)) = (x∗, y ∗),

0 0 0 dy 0 0 g10
F = f1 + f2 = f1 − f2 0 .
dx g2
 0
dy dy g1

F 00 = f11
00 + f 00
12 − f 00 + f 00
21 22
dx dx g20
 h i
00
i h
g 0 g 00 + g (dy/dx) − g 0 g 00 + g 00 (dy/dx)
 2 11 12 1 21 22
−f20 

 2 
0
g2

38
00 1 h
2 0
00 0 0 00 0 2 00 i
F = 0 2 (g2 ) f11 − 2g1 g2 f12 − (g1 ) f22
(g2)
0 0
 
λg2  0 00 00 0 0 00 0 00 g1
− 0 g2g11 − g12g1 − g1g12 + g1g22 0 
(g2)2 g2
1 h 0 2 00 0 0 00 0 2 00 i
= 0 2 (g2 ) f11 − 2g1 g2 f12 − (g1 ) f22
(g2)
λ h 0 2 00 0 0 00 0 2 00 i
− 0 (g2) g11 − 2g1g2g12 + (g1) g22 ,
(g2) 2

where FOCs require that f10 = λg10 and f20 = λg20 . By Young’s
00 = f 00 and g 00 = g 00 .
theorem, f12 21 12 21

39
We then obtain
    2   
00 − λg 00 0 − 2 f12 00 − λg 00 g 0 g 0
00 ∗ 1  f11 11 g2 12 1 2 
F (x ) =  2     2 
0 00
+ f22 − λg22 g100 0
g2
    2 
00
 00 
1  L11 g20 − 2 L12 g10 g20 
=  2   00   2 
0
g20 + L22 g1
H̄(x∗, y ∗)

= −  2 ,
g20

where H̄(x∗, y ∗) is defined in the next slide.

40
The term H̄(x∗, y ∗) is the determinant of a bordered Hessian

matrix at (x∗, y ∗):


g10 (x∗, y ∗) g20 (x∗, y ∗)


0

|H̄(x∗, y ∗)| = g10 (x∗, y ∗) L0011(x∗, y ∗, λ∗) L0012(x∗, y ∗, λ∗)


0
g (x∗, y ∗) L00 (x∗, y ∗, λ∗) L00 (x∗, y ∗, λ∗)

2 21 22

0 00 g 0 L00

0 g L 0
= −g1 10 12 + g2 1 11

00 0 00
2 L22 g2 L21
g

00
2 0
00 0 0 00 0 2
= −L11(g2) + 2L12g1g2 − L22(g1) .

Note that
H̄(x∗, y ∗)

F 00(x∗) < 0 ⇔ −  2 < 0 ⇔ H̄(x∗, y ∗) > 0.

g20

41
Theorem [Sufficiency for Local Optimum]: Consider the
problems:

max or min f (x, y) subject to g(x, y) = c.


x,y x,y
Suppose that (x∗, y ∗) and λ∗ satisfy the first-order conditions:
0
L01 = f10 (x∗, y ∗) − λ∗g1(x∗, y ∗) = 0
0
L02 = f20 (x∗, y ∗) − λ∗g2(x∗, y ∗) = 0
g(x∗, y ∗) = c.
∗ ∗ ∗ ∗ ∗ ∗

If H̄(x , y ) > 0, then (x , y ) is a local max point. If H̄(x , y ) <

0, then (x∗, y ∗) is a local min point.

42
Example: Consider the problem:

max xy subject to x + y = 6.
x,y
The Lagrangian is

L = xy − λ (x + y − 6)
The first-order conditions are

L01 = y − λ = 0
L02 = x − λ = 0
x + y = 6.
Thus (x∗, y ∗, λ∗) = (3, 3, 3).

43
Now we can check if (3, 3) is at least a local max point.

The determinant of the bordered Hessian is



0 1 1

H̄(3, 3) = 1 0 1


1 1 0

1 1 1 0
= 1 · (−1)1+2 1+3
+ 1 · (−1)

1 0 1 1


= 1 + 1 = 2 > 0.

Thus, the point (3, 3) is a local max point and the local maximum
value is 3 × 3 = 9.

44
Example: Revisit the previous example:

max or min x2y s.t. 2x2 + y 2 = 3.


x,y x,y
We had six solutions that satisfy the first-order conditions:
 √   √ 
(x, y, λ) = 0, 3, 0 with f 0, 3 = 0
 √   √ 
(x, y, λ) = 0, − 3, 0 with f 0, − 3 = 0
1
 
(x, y, λ) = 1, 1, with f (1, 1) = 1 (global max)
2 
1

(x, y, λ) = 1, −1, with f (1, −1) = −1 (global min)
2
1
 
(x, y, λ) = −1, 1, with f (−1, 1) = 1 (global max)
2
1
 
(x, y, λ) = −1, −1, − with f (−1, −1) = −1 (global min).
2

45
Recall that the first-order conditions were

L01 = 2xy − 4λx = 2x(y − 2λ) = 0,


L02 = x2 − 2λy = 0,
L0λ = −(2x
|
2 + y 2 −3) = 0.
{z }
=g(x,y)
Thus, we get

g10 = 4x
g20 = 2y
L0011 = 2y − 4λ
L0012 = L0021 = 2x
L0022 = −2λ.

46
The determinant of H̄(x, y) is then

0 4x 2y

H̄(x, y) = 4x 2y − 4λ 2x

−2λ

2y 2x

4x 2x 4x 2y − 4λ
= 4x(−1) 1+2 1+3
+ 2y(−1)

2y −2λ 2y 2x

= −4x(−8xλ − 4xy) + 2y(8x2 − 2y(2y − 4λ)).

 √  √ √
Thus if (x, y, λ) = 0, 3, 0 , then |H̄(0, 3)| = −8 · 3 3 =
√  √ 
−24 3 < 0; thus 0, 3 is a local min point.

 √  √ √ √
If (x, y, λ) = 0, − 3, 0 , then |H̄(0, − 3)| = 8·3 3 = 24 3 > 0;
 √ 
thus 0, − 3 is a local max point.

47

You might also like