Professional Documents
Culture Documents
Vectors
6.1 What is a Vector?
For most purposes economists think of vectors as an array of real numbers
(r
1
, r
2
.....r
q
). The set of all : vectors is written as
q
, pronounced R n. You
have to distinguish between row and column vectors when doing matrix and
vector algebra. In this book all vectors are column vectors.
x =
3
E
E
C
r
1
r
2
:
r
q
4
F
F
D
.
The corresponding row vector is written as:
x
0
= (r
1
, r
2
.....r
q
) .
The vector x
0
is sometimes pronounced as "x prime". Pure mathematicians
work with a more abstract denition of a vector and a vector space. Physicists
think of vectors as things that have both size and direction, for example velocity
and force.
6.2 Vector Addition and Multiplication
6.2.1 Vector Addition
If x and y are two vectors in
q
x +y =
3
E
E
C
r
1
r
2
:
r
q
4
F
F
D
+
3
E
E
C
j
1
j
2
:
j
q
4
F
F
D
=
3
E
E
C
r
1
+ j
1
r
2
+ j
2
:
r
q
+ j
q
4
F
F
D
.
47
48 CHAPTER 6. VECTORS
0 x
y
x + y
Figure 6.1: Vector Addition
Figure 6.1 illustrates vector addition with the vectors x and y shown nose to
tail. The alternative is to show vector addition as a parallelogram as in Figure
6.2.
Vector addition is commutative, that is x +y = y +x. This is because
x +y =
3
E
E
E
E
C
r
1
+ j
1
r
2
+ j
2
.
.
r
q
+ j
q
4
F
F
F
F
D
=
3
E
E
E
E
C
j
1
+ r
1
j
2
+ r
2
.
.
j
q
+ r
q
4
F
F
F
F
D
= y +x
If x
q
and y
p
and : and : are dierent you cannot add x and y. In
mathematical language the sum of x and y is undened.
6.2.2 Scalar Multiplication
If / is a real number (called a scalar in this context)
/x = /
3
E
E
C
r
1
r
2
:
r
q
4
F
F
D
=
3
E
E
C
/r
1
/r
2
:
/r
q
4
F
F
D
= x/.
Figure 6.3 illustrates a vector x multiplied by 2, 1 and 2. Multiplying a
vector by a positive number keeps it pointing in the same direction but makes
it longer or shorter. Multiplying a vector by a negative number makes it point
in the opposite direction.
Scalar multiplication is distributive over vector addition, that is
6.2. VECTOR ADDITION AND MULTIPLICATION 49
0 x
y
x + y
Figure 6.2: Vector Addition Illustrated with a Parallelogram
0
x
- x
- 2x
2x
Figure 6.3: Multiplying a Vector by a Scalar
50 CHAPTER 6. VECTORS
/(x +y) = /x + /y for all : vectors x and y.
To see why observe that
/ (x +y) = /
3
E
E
E
E
C
r
1
+ j
1
r
2
+ j
2
.
.
r
q
+ j
q
4
F
F
F
F
D
=
3
E
E
E
E
C
/ (r
1
+ j
1
)
/ (r
2
+ j
2
)
.
.
/ (r
q
+ j
q
)
4
F
F
F
F
D
= /
3
E
E
E
E
C
r
1
r
2
.
.
r
q
4
F
F
F
F
D
+ /
3
E
E
E
E
C
j
1
j
2
.
.
j
q
4
F
F
F
F
D
= /x+/y.
The denition of vector addition also implies that if / and / are scalars and x
a vector
(/ + /) x =/x + /x
that is scalar multiplication is distributive over scalar addition.
6.2.3 Inner product
The inner product of two vectors x and y in
q
is
x
0
y =
q
X
l=1
r
l
j
l
.
The inner product is sometime called the "scalar product" or "dot product"
and written as x y. The properties of the inner product are:
1. The inner product is commutative, that is
x
0
y = y
0
x
for all : vectors x and y. This is because x
0
y =
P
q
l=1
r
l
j
l
=
P
q
l=1
j
l
r
l
=
y
0
x.
2. If x and y are : vectors and / is a scalar
/(x
0
y) = (/x)
0
y = x
0
(/y)
because /(x
0
y) = / (
P
q
l=1
r
l
j
l
) =
P
q
l=1
(/r
l
)j
l
= (/x)
0
y =
P
q
l=1
r
l
(/j
l
) =
x
0
(/y) .
6.3. LENGTH AND NORM 51
3. If t, u, v and w are : vectors
t
0
(u +v) = (u +v)
0
t = t
0
u +t
0
v
because t
0
(u +v) = (u +v)
0
t =
P
q
l=1
t
l
(n
l
+
l
) =
P
q
l=1
t
l
n
l
+
P
q
l=1
t
l
l
=
t
0
u +t
0
v.
4. Also
(t +w)
0
(u +v) = t
0
u +w
0
u +t
0
v +w
0
v = u
0
t +u
0
w+v
0
t +v
0
w
because (t +w)
0
(u +v) =
P
q
l=1
(t
l
+ n
l
) (n
l
+
l
) =
P
q
l=1
(t
l
n
l
+ n
l
n
l
+ t
l
l
+ n
l
l
) .
5. Similarly If t, u, v and w are : vectors and a, /, c, d are scalars
(at+/w)
0
(cu+dv) = (at)
0
(cu+dv) + (/w)
0
(cu+dv)
= (at)
0
(cu) + (at)
0
(dv) +(/w)
0
(cu) +(/w)
0
(dv)
= act
0
u+adt
0
v+/cw
0
u+/dw
0
v. (6.1)
6.3 Length and Norm
6.3.1 Denition
For all : vectors x, x
0
x =
P
q
l=1
r
2
l
. As r
2
l
_ 0 for all r
l
and r
2
l
= 0 if and
only if r
l
= 0,
x
0
x _ 0 for all x and x
0
x = 0 if and only if x = 0
where 0 is an : vector with every element 0. Dene |x| as
|x| = (x
0
x)
1
2
.
so |x| _ 0 and |x| = 0 if and only if x = 0. Pythagoras Theorem implies
that in 2 and 3 dimensions |x| is the length of x. Pure mathematicians call
|x| the norm of x. It is also called the length of x.
6.3.2 Properties of the Length |x|
These are things you need to remember. The rst property follows directly from
the denition. The others are proved in sections 6.5 and 6.6 of this chapter.
1. If r is a scalar |rx| = [r[ |x| where [r[ is the absolute value of r, so [r[ = r
if r _ 0 and [r[ = r if r < 0.
2. x
0
y = |x| |y| cos 0 where 0 is the angle between x and y.
3. Cauchy-Schwarz inequality:
[x
0
y[ _ |x| |y| .
52 CHAPTER 6. VECTORS
4. Triangle inequality
|x +y| _ |x| +|y|
|x y| _ |x| +|y| .
5. If x
0
y = 0, x ,= 0 and y ,= 0 the angle between x and y is 90
because
cos 90
= 1.
6.4 The Least Squares Problem
6.4.1 The Least Squares Problem With Vector Notation
You will become very familiar with least squares problems in econometrics. This
section of notes covers the simplest case. It also generates inequality 6.3 which
is an essential step on the way to the Cauchy-Schwarz inequality. The problem
here is nding the value of / that minimizes
1(/, x, y) =
W
X
w=1
(j
w
/r
w
)
2
.
You will come to think of this as the regression of j onto a single variable r
without an intercept. Suppose x and y are T vectors. Given the denition of
the norm |y /x|
1(/, x, y) =
W
X
w=1
(j
w
/r
w
)
2
= |y /x|
2
= (y /x)
0
(y /x)
Expanding the brackets using the result in equation 6.1 on inner products gives
1(/, x, y) = (y /x)
0
(y /x)
= y
0
y /x
0
y /y
0
x + /
2
x
0
x
= y
0
y 2/x
0
y + /
2
x
0
x
because x
0
y = x
0
y.
Assume that r
w
,= 0 for some t so x ,=0. This implies that x
0
x 0. Then
completing the square
1(/, x, y) = y
0
y2/x
0
y + /
2
x
0
x (6.2)
= x
0
x
/ (x
0
x)
1
(x
0
y)
2
+y
0
y(x
0
x)
1
(x
0
y)
2
_ y
0
y(x
0
x)
1
(x
0
y)
2
6.4. THE LEAST SQUARES PROBLEM 53
for all / and
1(/, x, y)= y
0
y(x
0
x)
1
(x
0
y)
2
if and only if / = (x
0
x)
1
(x
0
y). Thus 1(/, x, y) is minimized by setting
/ = (x
0
x)
1
(x
0
y)
and has a minimum value of
y
0
y(x
0
x)
1
(x
0
y)
2
.
Recall that 1(/, x, y) is a sum of squares, so 1(/, x, y) _0 for all / includ-
ing / = (x
0
x)
1
(x
0
y) at which point 1(/, x, y) =y
0
y(x
0
x)
1
(x
0
y)
2
. Thus
y
0
y _ (x
0
x)
1
(x
0
y)
2
. Recall that by assumption x ,= 0 so x
0
x 0. Multiply-
ing by x
0
x and rearranging gives
(x
0
y)
2
_ (x
0
x) (y
0
y) . (6.3)
6.4.2 The Least Squares Problem Without Vector Nota-
tion
You can also solve this simple least squares problem without using vectors. The
problem is to nd / that minimizes the sum of squares.
q
X
w=1
(j
w
/r
w
)
2
.
Expanding the brackets
q
X
w=1
(j
w
/r
w
)
2
=
q
X
w=1
j
2
w
2/j
w
r
w
+ /
2
r
2
w
=
q
X
w=1
j
2
w
2/
q
X
w=1
r
w
j
w
!
+ /
2
q
X
w=1
r
2
w
!
.
Assume that at least one of the r
0
w
: is not zero, so
P
q
w=1
r
2
w
0. Completing
the square
q
X
w=1
(j
w
/r
w
)
2
=
q
X
w=1
r
2
w
!
3
C
/
q
X
w=1
r
2
w
!
1
q
X
w=1
r
w
j
w
!
4
D
2
+
q
X
w=1
j
2
w
q
X
w=1
r
2
w
!
1
q
X
w=1
r
w
j
w
!
2
54 CHAPTER 6. VECTORS
which is minimized by setting
/ =
q
X
w=1
r
2
w
!
1
q
X
w=1
r
w
j
w
!
= (x
0
x)
1
(x
0
y) .
and has a minimum value of
q
X
w=1
j
2
w
q
X
w=1
r
2
w
!
1
q
X
w=1
r
w
j
w
!
2
= y
0
y(x
0
x)
1
(x
0
y)
2
.
Notice that expressions like
P
q
w=1
r
w
j
w
take longer to write down and are harder
to work with than the corresponding vector expression x
0
y. The algebra
becomes even worse if you are working with the general least squares problem
of minimizing
q
X
w=1
j
w
n
X
n=1
r
wn
/
n
!
2
.
If you are to survive your econometrics course you must make the switch from
the notation
P
q
w=1
r
w
j
w
to vector notation x
0
y as rapidly as possible.
6.5 Inequalities for Vectors
6.5.1 The Cauchy-Schwarz Inequality
The Cauchy-Schwarz inequality says that for any vectors x and y with the same
number of elements
[x
0
y[ _|x| |y| . (6.4)
Here |x| and |y| are the norm or length of x and y, so |x|
2
= x
0
x and
|y|
2
= y
0
y. The expression [x
0
y[ is the absolute value of x
0
y. (The absolute
value of a number j is j if j _ 0 and j if j < 0. Note that this implies that
j
2
= [j[
2
)
The Cauchy-Schwarz inequality is derived from inequality 6.3 which states
that (x
0
y)
2
_ (x
0
x) (y
0
y). From the denitions of absolute value and norm
[x
0
y[
2
= (x
0
y)
2
, (x
0
x) = |x|
2
and y
0
y =|y|
2
so the inequality (x
0
y)
2
_
(x
0
x) (y
0
y) can be written as
[x
0
y[
2
_|x|
2
|y|
2
(6.5)
which implies the Cauchy-Schwarz inequality [x
0
y[ _|x| |y| . Inequality 6.3
was proved under the assumption that x ,= 0, but as both sides of the Cauchy-
Schwarz inequality are zero when x =0 it also holds when x =0.
6.5. INEQUALITIES FOR VECTORS 55
x
y
x + y
Figure 6.4:
6.5.2 The Triangle Inequality
The triangle inequality states that
|x +y| _ |x| +|y| .
Geometrically you can think of x +y, x and y as three sides of a triangle. The
inequality simply states that one side of a triangle cannot be longer than the
sum of the other two sides, as illustrated in Figure 6.4.
The triangle inequality is a direct consequence of the Cauchy-Schwarz in-
equality. From the denition of |x +y|
|x +y|
2
= (x +y)
0
(x +y) = x
0
x +2x
0
y +y
0
y
56 CHAPTER 6. VECTORS
T
O bx x
C
y
B
Figure 6.5:
From the Cauchy-Schwarz inequality [x
0
y[ _|x| |y|. As . _ [.[ for any num-
ber . this implies that x
0
y _|x| |y| so
|x +y|
2
= x
0
x +2x
0
y +y
0
y
_ |x|
2
+2|x| |y| +|y|
2
= (|x| +|y|)
2
.
As |x +y| , |x| and |y| are all non negative this implies the triangle inequality
|x +y| _ |x| + |y| . Replacing y by y and noting that |y| = |y| the
triangle inequality implies that
|x y| _ |x| +|y| .
6.6 The Angle Between Two Vectors
6.6.1 Showing That x
0
y = |x| |y| cos
Think of x and y as non zero vectors and / as a scalar. In Figure 6.5 /x is
a scalar multiple of the vector x. The least squares problem is nding / to
minimize |y/x| which is the distance between the point C at the end of the
vector y and the horizontal straight line from the origin 0 that goes through
the end of the vector x. Geometrically you nd the point /x on the line 0x that
6.6. THE ANGLE BETWEEN TWO VECTORS 57
0 90 180 270 360
T
cosT
1
- 1
Figure 6.6: The Graph of cos 0
minimizes the length of y/x by drawing the line from C to 1 that is at 90
1
2
(x
0
y) . Ele-
mentary trigonometry and Figure 6.5 imply that if 0 is the angle between the
vectors x and y
cos 0 =
01
0C
=
(x
0
x)
1
2
(x
0
y)
|y|
=
x
0
y
|x| |y|
(6.6)
where cos 0 is the cosine of the angle 0. Equation 6.6 implies that
x
0
y =|x| |y| cos 0.
In words this says that the inner product of the vectors x and y is the product
of their lengths and the cosine of the angle between the two vectors. The
relationship also holds when / is negative, so the angle between the two vectors
is more than 90
_ 0 _ 360
and 180
58 CHAPTER 6. VECTORS
T
360 - T
x
y
Figure 6.7:
and
1 < cos 0 < 0 if 90
.
x
0
y < 0 if the angle between x and y is between 90
and 180
.
6.6.2 Orthogonal Vectors
Suppose x and y are non zero vectors, so |x| 0 and |y| 0. Suppose
also that x
0
y = 0. This implies that x
0
y = |x| |y| cos 0 = 0 where 0 is the
angle between the vectors x and y. As |x| 0 and |y| 0 this implies that
cos 0 = 0. Given the graph of cos 0 this implies that 0 = 90
or 0 = 270
.
As Figure 6.8 shows you can measure the angle between x and y as 90
or 270
. In mathematical
language the vectors x and y are said to be orthogonal if x
0
y = 0 so the vectors
are at an angle of 90
is
called a "right angle".
6.6.3 Parallel vectors
If
x
0
y = |x| |y| cos 0 = |x| |y|
then cos 0 = 1 so 0 = 0. The angle between x and y is 0
j
1
j
2
r
1
so the slope of the budget line is
s
1
s2
. If j
2
= 0 the line is vertical. In vector
notation equation 6.7 can be written as
p
0
x = :.
Another way at looking at this equation is to choose any point x
0
on the line
so p
0
x
0
=: and write the equation as p
0
x = p
0
x
0
, or rearranging
p
0
(x x
0
) = 0.
Recall that the inner product p
0
(x x
0
) = |p| |x x
0
| cos 0 where 0 is the
angle between the two vectors p and x x
0
. Thus as cos 90
= 0, if the inner
product is 0, the angle is 90
, so from Proposition 1 p
0
(x
1
x
0
) < 0
and q(x
1
x
0
) 0. The intuition is obvious from the diagram, but the
result is so important that it merits a formal proof.
Proof. Let
r = q (p
0
p)
1
(p
0
q) p (6.8)
and note that
p
0
r = 0. (6.9)
so as shown in Figure 6.12 r is orthogonal to p. Thus from equations 6.8 and
6.9 r
0
r = r
0
q (p
0
p)
1
(p
0
q) r
0
p = r
0
q so
r
0
r = r
0
q = q
0
r. (6.10)
Let
x
1
= x
0
+r p
where j 0. Then
p
0
x
1
= p
0
x
0
+p
0
r p
0
p = p
0
x
0
p
0
p (6.11)
because from equation 6.9 p
0
r = 0. By assumption j 0 and p ,= 0, so p
0
p 0.
Thus from equation 6.11
p
0
(x
1
x
0
) < 0. (6.12)
64 CHAPTER 6. VECTORS
x
1
x
2
p
q
p( x - x
0
) = 0
q( x - x
0
) = 0
x
1
0
x
0
Figure 6.11:
x
1
x
2
p
q
p( x - x
0
) = 0
q( x - x
0
) = 0
r
r
x
1
0
x
0
q r = Dp
-Pp
Figure 6.12:
6.8. FIRST THOUGHTS ON CONSTRAINED MAXIMIZATION 65
Also
q
0
x
1
= q
0
x
0
+q
0
r q
0
p = q
0
x
0
+r
0
r q
0
p
because from equation 6.10 q
0
r = r
0
r. Thus
q
0
(x
1
x
0
) = r
0
r q
0
p. (6.13)
As j 0 the right hand side of equation 6.13 can be made strictly positive if
either r ,= 0 so r
0
r 0 and j is su!ciently small, or r = 0 and q
0
p < 0. In
either case
q
0
(x
1
x
0
) 0. (6.14)
If r = 0 and q
0
p = p
0
q _ 0 equation 6.8 implies that
q =`p
where ` = (p
0
p)
1
(p
0
q). As p
0
p 0 and p
0
q _ 0, ` _ 0. Thus either there
is a vector x
1
satisfying conditions 6.12 and 6.14 or there is a number ` _ 0
such that q =`p.
Proposition 5 implies that if p ,= 0, p
0
x
0
= c, and there is no number
` _ 0 such that q = `p, then x
0
cannot solve the problem of maximizing q
0
x
subject to the constraint p
0
x _ c, because p
0
x
1
< p
0
x
0
= c and q
0
x
1
q
0
x
0
.
Proposition 4 states that if p
0
x
0
= c and there is a real number ` _ 0 such that
q = `p then x
0
solves the constrained maximization problem. Taken together
these results imply that
Proposition 6 If p and x
0
are : vectors, p ,= 0 and p
0
x
0
= c then x
0
solves
the problem of maximizing q
0
x subject to p
0
x _ c if and only if there is a number
` _ 0 such that q = `p.
The number ` is of course a Lagrange multiplier.
66 CHAPTER 6. VECTORS
Chapter 7
Open Sets, Closed Sets and
Continuous Functions
7.1 Why This Matters to Economists
The mathematics in this chapter matters to economists for several reasons.
There is a big result, that a function that is continuous on a closed and bounded
subset of
q
has a maximum and a minimum, which is important because maxi-
mization and minimization crop up very frequently in economics. You will meet
the mathematics discussed in this chapter very early in your microeconomics
course in the form of the continuity assumptions for consumer preferences and
its implication, the existence of a continuous utility functions that captures the
preferences. Continuity is also central to theorems on the existence of general
competitive equilibrium.
Sections 1-4 are very useful. The rest of the chapter is harder going and less
widely used in economics. Section 9 is the background you need to understand
the material on continuity in consumer theory. This is quite hard; if you
are struggling it would be reasonable to skip this and accept that this bit of
consumer theory will not mean much to you.
7.2 Vector Length and Open Balls
Remember from the previous chapter that in
q
we have a concept of the length
or norm of a vector
|x| =
q
X
l=1
r
2
l
!1
2
.
In one dimensional space : = 1, and |r| = [r[ , the absolute value of r. In two
and three dimensional space Pythagoras Theorem implies that this is indeed
67
68CHAPTER7. OPENSETS, CLOSEDSETS ANDCONTINUOUS FUNCTIONS
the length of the vector. The norm has the property that
|x| _ 0 for all x (7.1)
|x| = 0 if and only if x = 0. (7.2)
The distance between two vectors x and x
0
in
q
is dened as
|x x
0
| =
q
X
l=1
(r
l
r
0l
)
2
!1
2
.
In one dimensional space this reduces to the absolute value [r r
0
[ which is the
distance between the numbers r and r
0
. In two and three dimensional space
Pythagoras Theorem implies that |x x
0
| is the distance between x and x
0
.
I therefore refer to |x x
0
| as distance, even when : 3. Properties 7.1
and 7.2 imply that
|x x
0
| _ 0 for all x
|x x
0
| = 0 if and only if x = x
0
.
Now think about the set of points that are closer to x
0
than some strictly
positive number c. In one dimensional space x
0
is a number and this is the
open interval (r
0
c, r
0
+ c) . In two dimensional space it is a circle centred
on x
0
with radius c. In three dimensional space this is a sphere or ball. In
fact we use the term ball regardless of the dimension of the space, using the
following denition.
Denition 7 If c is a strictly positive number and x
0
is an element of
q
the
open ball 1(x
0
, c) centred on x
0
with radius c is the set of points that are
closer than c to x
0
, that is
1 (x
0
, c) = x : x
q
, |x x
0
| < c .
Note that the inequality in this denition < is strict. I have used the
word "open" many times before when writing about "open intervals" and I
will explain shortly what the word "open" means in general. In order to do
this I will introduce some more terms that mathematicians have borrowed from
elsewhere.
7.3 Boundaries
Imagine yourself standing on the equator. Take a tiny step to the north; you
are in the northern hemisphere. Take a tiny step to the south; you are in the
southern hemisphere. The equator is the boundary of the two hemispheres.
Mathematicians have borrowed the term boundary from geographers and use
it much the same sense, applying it to the : dimensional space of real number
q
.
7.4. OPEN AND CLOSED SETS AND BOUNDARIES 69
Denition 8 A point x
0
in
q
is an element of the boundary of a subset o
of
q
if for any number c 0 the open ball 1 (x
0
, c) contains both points that
are elements of o and points that are in the complement o
F
of o.
You may recall that the denition of complement:
Denition 9 The complement o
F
of a subset o of
q
is the set of elements
of
q
that are not in o, that is
o
F
= x : x
q
, x , o .
The fact that the radius c of the open ball 1 (x
0
, ) can be a very small posi-
tive number captures the idea that a step to the north, however tiny, moves you
from the equator to the northern hemisphere. The denitions of the boundary
and the complement of a set immediately imply a result that is useful later.
Proposition 10 The boundary of a set o and the boundary of its complement
o
F
are the same.
The idea of a boundary may become clearer with an example.
Example 11 The boundary of the set
o =
(r
1
, r
2
) : (r
1
, r
2
)
2
, r
1
r
2
(r
1
, r
2
) : (r
1
, r
2
)
2
, r
1
0, r
2
0
(r
1
, r
2
) : (r
1
, r
2
)
2
, r
1
_ 0, r
2
_ 0
(r
1
, r
2
) : (r
1
, r
2
)
2
, r
1
0, r
2
_ 0
q
is open.
Proof. If x is an element of the union it is an element of one of the open sets
o, so there is an open ball 1 (x, c) that is a subset of o, and thus a subset of
the union of the sets. Thus the union is open.
The second result is:
Proposition 17 The intersection of a nite number of open sets of
q
is
open.
Proof. Let the subsets be o
1>
o
2
....o
q
, and let x be a point in their intersection.
As o
l
is open there is a number c
l
0 such that the open ball 1 (x, c
l
) is a
subset of o
l
, that is
1 (x, c
l
) o
l
(7.3)
Let
c = min(c
1
, c
2
, ....c
q
) .
Thus 0 < c _ c
l
for all i so the open ball 1 (x, c) is a subset of the open ball
1 (x, c
l
) for all i, that is 1 (x, c) 1 (x, c
l
). Given relationship 7.3 this
implies
1 (x, c) o
l
for i = 1, 2...:
implying that 1 (x, c) is a subset of the intersection of o
1>
o
2
....o
q
, so the inter-
section is itself open.
Note that the fact that this result applies to a nite number of open sets
is important. For example that sets 1
0,
1
q
r,
%
2
.
Inequality 7.7 then implies that 1
r,
%
2
x,
1
l
a
/
e
d+e
+
a
/
d
d+e
n
d
d+e
1
n
e
d+e
2
j
1
d+e
.
(You can nd this result on page 54 of H.R.Varian, Microeconomic Analysis.)
This reduces to the cost function j
2
if a = / =
1
4
and = 2
2
n
1
2
1
n
1
2
2
.
However, having acknowledged that costs depend on n
1
, n
2
and j, you now
have to use the partial derivative
0c (n
1
, n
1
, j)
0j
=
1
a + /
1
d+e
a
/
e
d+e
+
a
/
d
d+e
n
d
d+e
1
n
e
d+e
2
j
1
d+e
1
for marginal cost. When you have a function of many variables ) (r
1
, r
2
, .r
l
..r
q
)
the partial derivative is dened analogously to the derivative of a function of a
85
86 CHAPTER 8. INTRODUCTION TO MULTIVARIATE CALCULUS
x 0
y
x
0
y = f(x)
y = f(x
0
) + f(x
0
) (x x
0
)
Figure 8.1: A function and its tangent
single variable as
0) (r
1
, r
2
, .r
l
..r
q
)
0r
l
= lim
k$0
) (r
1
, r
2
, .r
l
+ /..r
q
) ) (r
1
, r
2
, .r
l
..r
q
)
/
.
You calculate the partial derivative
0) (r
1
, r
2
, .r
l
..r
q
)
0r
l
in exactly the same way
as you calculate the derivative of a function of a single variable, treating all the
variables except r
l
as parameters of the function.
In fact the distinction between a parameter and a variable depends entirely
on what you are interested in. Varian writes c (n
1
, n
1
, j) treating , a and
/ as parameters. But for some purposes, for example modelling technical
change you might want to think of , a and / as variables, and write costs as
q (, a, /, n
1
, n
2
, j) .
8.2 Derivatives and Approximations
8.2.1 When Can You Approximate
Think about a function of one variable j = ) (r) illustrated by the curved
line in Figure 8.1 and its tangent at r
0
which is a straight line with equation
j = ) (r
0
)+)
0
(r
0
) (r r
0
). The denition of the derivative )
0
(r) implies that
close to r the tangent line is a good approximation to the original function, that
is
) (r) - ) (r
0
) + )
0
(r
0
) (r r
0
) when r is close to r
0
.
8.2. DERIVATIVES AND APPROXIMATIONS 87
5
2.5
0
-2.5
-5
5
2.5
0
-2.5
-5
120
100
80
60
x y
z
x y
z
Figure 8.2: Appoximating the function at (1, 2) ) (r) = 100 r
2
1
r
2
2
at (1, 2)
by the plane j (r) = 95 2 (r
1
1) 4 (r
2
2) .
Now think about the function of two variables j = ) (r
1
, r
2
) illustrated in Figure
8.2. The graph of the function is now a curved surface. Just as the curved line
in two dimensional space corresponds to a curved surface in three dimensional
space, the tangent line in two dimensional space corresponds to a tangent plane
in three dimensional space. In the chapter on vectors I showed that any plane
through the point (r
01
, r
02
, j
0
) can be written in vector notation as
(j
1
, j
2
, j
3
)
3
C
3
C
r
1
r
2
j
4
D
3
C
r
01
r
02
j
0
4
D
4
D
= 0
or equivalently
j
1
(r
1
r
01
) + j
2
(r
2
r
02
) + j
3
(j j
0
) = 0
or if j
3
,= 0 as
j = j
0
+ a
1
(r
1
r
01
) + a
2
(r
2
r
02
)
where a
1
=
s1
s
3
and a
2
=
s2
s
3
. By analogy with the equation of the tangent
line as j = ) (r
0
) +)
0
(r
0
) (r r
0
) the equation of the tangent plane should be
j = ) (r
01
, r
02
) +
0) (r
01
, r
02
)
0r
1
(r
1
r
01
) +
0) (r
01
, r
02
)
0r
2
(r
2
r
02
)
and the approximation should be
) (r
1
, r
2
) - ) (r
01
, r
02
) +
0) (r
01
, r
02
)
0r
1
(r
1
r
01
) +
0) (r
01
, r
02
)
0r
2
(r
2
r
02
) .
88 CHAPTER 8. INTRODUCTION TO MULTIVARIATE CALCULUS
This suggests that the extension to : variables should be the approximation
) (x) - ) (x
0
) +
q
X
l=1
0) (x
0
)
0r
l
(r
l
r
l0
) when x is close to x
0
.
The argument I have just given is in no way rigorous, but it can be made so.
This is done by assuming that the function ) (x) can be approximated by a
function of the form
) (x) - ) (x
0
) +
q
X
l=1
a
l
(r
l
r
l0
)
and then showing that a
l
must be the partial derivative
Ci(x0)
C{l
.
You have to be slightly careful with this, because unlike the situation with
functions of a single variable it is possible that the derivatives exist but the
approximation does not work. As an example consider the function ) :
2
given by
) (r
1
, r
2
) = 0 when r
1
r
2
= 0
) (r
1
, r
2
) = 1 when r
1
r
2
,= 0.
This function is 0 when one or both of r
1
and r
2
are zero, and 1 otherwise.
Thus ) (r
1
, 0) = 0 for all values of r
1
and ) (0, r
2
) = 0 for all values of r
2
,
implying that the partial derivatives
0) (0, 0)
0r
1
=
0) (0, 0)
0r
2
= 0.
This suggests that when (r
1
, r
2
) is close to (0, 0)
) (0, 0) +
0) (0, 0)
0r
1
r
1
+
0) (0, 0)
0r
2
r
2
(8.1)
should be a good approximation to ) (r
1
, r
2
) . However
) (0, 0) =
0) (0, 0)
0r
1
=
0) (0, 0)
0r
2
= 0
so this implies that 0 is a good approximation to ) (r
1
, r
2
) when (r
1
, r
2
) is
close to 0. But this is very far from being a good approximation because
) (r
1
, r
2
) = 1 whenever r
1
r
2
,= 0.
An assumption that ensures that the approximation works for a function
) : o is that o is open, that ) has partial derivatives on o, and that the
partial derivatives are continuous. In order to avoid writing out the conditions
many times I will use the following denition.
Denition 36 If o is an open subset of
q
the function ) (x) : o is a well
behaved if it has partial derivatives, and the partial derivatives are continuous
on o.
8.3. LEVEL SETS AND VECTORS OF PARTIAL DERIVATIVES 89
x
0
x
1
x
2
Df(x
0
)
Df(x
0
)( x - x
0
) = 0
f(x) = f(x
0
)
Figure 8.3: The Level Sets, Tangent and Partial Derivative Vector for a Function
If a function is well behaved the approximation
) (x) - ) (x
0
) +
q
X
l=1
0) (x
0
)
0r
l
(r
l
r
l0
) when x is close to x
0
works. Making this argument rigorous requires a precise denition of what is
meant by approximation, and then a proof that existence and continuity of the
partial derivatives implies that the approximation is valid.
8.3 Level Sets and Vectors of Partial Derivatives
8.3.1 The Level Set and Tangent for One Function
In the previous chapter I gave:
Denition 37 If o is a subset of
q
and the function ) : o , the set
x : x o, ) (x) = j
of elements of o for which ) (x) = j is called a level set.
A level set of a utility function is an indierence curve; a level set of a
production function is an isoquant. Now think about two points x and x
0
that
lie in the same level set so
) (x) = ) (x
0
)
and assume that the function ) : o is well behaved so close to x
0
) (x) - ) (x
0
) +
q
X
l=1
0) (x
0
)
0r
l
(r
l
r
l0
) .
90 CHAPTER 8. INTRODUCTION TO MULTIVARIATE CALCULUS
As ) (x) = ) (x
0
) this implies that
q
X
l=1
0) (x
0
)
0r
l
(r
l
r
l0
) - 0. (8.2)
If I use notation
1) (x
0
) =
3
E
E
E
E
E
C
Ci(x0)
C{
1
Ci(x
0
)
C{
2
.
.
Ci(x
0
)
C{
q
4
F
F
F
F
F
D
so 1) (x
0
) is the : vector of partial derivatives equation 8.2 becomes
1) (x
0
)
0
(x x
0
) - 0.
Thus when x is very close to x
0
and ) (x) = ) (x
0
) x very nearly satises the
equation.
1) (x
0
)
0
(x x
0
) = 0.
This is an equation of the type that I discussed at some length in section 7
of the chapter on vectors. If : = 2 this is a straight line, if : = 3 it is
a plane, and for : 3 it is a hyperplane. In every case the vector of partial
derivatives 1) (x
0
) is orthogonal (at 90