Math Prep Notes1

Joel Sobel and Joel Watson

2008, under revision

1 2006,
c 2008 by Joel Sobel and Joel Watson. For use only by Econ 205 students at
UCSD in 2008. These notes are not to be distributed elsewhere.
ii

Preface
These notes are the starting point for a math-preparation book, primarily for use by
UCSD students enrolled in Econ 205 (potentially for use by folks outside UCSD
as well). The first draft consists of a transcription of Joel Watson’s handwritten
notes, as well as extra material added by Philip Neary, who worked on the tran-
scription in 2006. Joel Sobel and Joel Watson have revised parts of these notes
and added material, but the document is still rough and disorganized. Surely there
are many mistakes. These material here is incomplete and contain many mistakes.
If you find an error, a notational inconsistency, or other deficiency, please let one
of the Joels know.
Contents

1 Sets, Functions, and the Real Line 3

1.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Real Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Methods of Proof . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Some helpful notes . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Sequences 25
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 Functions and Limits of Functions 37

4 Differentiation 47

5 Taylor’s Theorem 59

6 Univariate Optimization 63

7 Integration 69
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.2 Fundamental Theorems of Calculus . . . . . . . . . . . . . . . . 73
7.3 Properties of Integrals . . . . . . . . . . . . . . . . . . . . . . . . 75
7.4 Computing Integrals . . . . . . . . . . . . . . . . . . . . . . . . 76

8 Basic Linear Algebra 79

8.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
8.2 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
8.2.1 Matrix Algebra . . . . . . . . . . . . . . . . . . . . . . . 83
8.2.2 Inner Product and Distance . . . . . . . . . . . . . . . . . 87
8.3 Systems of Linear Equations . . . . . . . . . . . . . . . . . . . . 90

iii
iv CONTENTS

8.4 Linear Algebra: Main Theory . . . . . . . . . . . . . . . . . . . . 93

8.5 Eigenvectors and Eigenvalues . . . . . . . . . . . . . . . . . . . . 95
8.6 Quadratic Forms . . . . . . . . . . . . . . . . . . . . . . . . . . 98

9 Multivariable Calculus 101

9.1 Linear Structures . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.2 Linear Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.3 Representing Functions . . . . . . . . . . . . . . . . . . . . . . . 105
9.4 Limits and Continuity . . . . . . . . . . . . . . . . . . . . . . . . 108
9.5 Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.6 Partial Derivatives and Directional Derivatives . . . . . . . . . . . 111
9.7 Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
9.8 Properties of the Derivative . . . . . . . . . . . . . . . . . . . . . 116
9.9 Gradients and Level Sets . . . . . . . . . . . . . . . . . . . . . . 120
9.10 Homogeneous Functions . . . . . . . . . . . . . . . . . . . . . . 123
9.11 Higher-Order Derivatives . . . . . . . . . . . . . . . . . . . . . . 124
9.12 Taylor Approximations . . . . . . . . . . . . . . . . . . . . . . . 125

10 Convexity 127
10.1 Preliminary: Topological Concepts . . . . . . . . . . . . . . . . . 127
10.2 Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
10.3 Quasi-Concave and Quasi-Convex Functions . . . . . . . . . . . 130
10.3.1 How to check if a function f is quasiconcave or not . . . . 131
10.3.2 Relationship between Concavity and Quasiconcavity . . . 132
10.3.3 Ordinal “vs” Cardinal . . . . . . . . . . . . . . . . . . . 133

11 Unconstrained Extrema of Real-Valued Functions 137

11.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
11.2 First-Order Conditions . . . . . . . . . . . . . . . . . . . . . . . 138
11.3 Second Order Conditions . . . . . . . . . . . . . . . . . . . . . . 139
11.3.1 S.O. Sufficient Conditions . . . . . . . . . . . . . . . . . 140
11.3.2 S.O. Necessary Conditions . . . . . . . . . . . . . . . . . 140

12 Invertibility and Implicit Function Theorem 143

12.1 Inverse Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 143
12.2 Implicit Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 146
12.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
12.4 Envelope Theorem for Unconstrained Optimization . . . . . . . . 153
CONTENTS v

13 Constrained Optimization 155

13.1 Equality Constraints . . . . . . . . . . . . . . . . . . . . . . . . . 155
13.2 The Kuhn-Tucker Theorem . . . . . . . . . . . . . . . . . . . . . 159
13.3 Saddle Point Theorems . . . . . . . . . . . . . . . . . . . . . . . 164
13.4 Second-Order Conditions . . . . . . . . . . . . . . . . . . . . . . 173
13.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

14 Monotone Comparative Statics 179

vi CONTENTS
CONTENTS 1

Notes on notation and items to be corrected:

1. Sets are typically denoted by italic (math type) upper case letters; elements
are lower case.

2. PN used script letters to denote sets and there may be remnants of this
throughout (such as in some figures, all which must be redrawn anyway).

3. ε (varepsilon) is the preferred epsilon symbol.

4. Standard numerical sets (reals, positive integers, etc.) are written using the
mathbb symbols: R, P, and so on.
2 CONTENTS
Chapter 1

Sets, Functions, and the Real Line

This chapter reviews some basic definitions regarding sets and functions, and it
contains a brief overview of the construction of the real line.

1.1 Sets
The most basic of mathematical concepts is a set, which is simply a collection of
objects. For example, the “days of the week” is a set comprising the following
objects: Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday.
The set of Scandinavian countries consists of: Sweden, Norway, Denmark, Fin-
land and Iceland. Sets can be composed of any objects whatsoever.
We often use capital italic letters to denote sets. For instance, we might let D
denote the set of days of the week and S denote the set of Scandinavian countries.
We will use lowercase italic letters to denote individual objects (called elements
or points) in sets. Using the symbol “∈,” which means “is an element of,” we thus
write x ∈ X to indicate that x is an element of set X. By the way, the symbol “6∈”
means “is not an element of,” so we would write x 6∈ X to mean that X does not
contain x.
To define a set, it is sometimes convenient to list its elements. Formally, we
do this by enclosing the list in curly brackets and separating the elements with
commas. For instance, the set of Scandinavian countries is

S = {Sweden, Norway, Denmark, Finland, Iceland}

and the set of days of the week is

D = {Monday, Tuesday, Wednesday, . . . , Sunday}.

3
4 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

When a set contains many elements, it is useful to define it by making reference

to a generic element x, using the “such as” symbol “|”, and including qualifying
statements (properties that x is required to have). For instance, the set of numbers
between 2 and 5 (including the endpoints) is
{x | 2 ≤ x ≤ 5}.
The set that contains no elements is called the empty set and is denoted ∅. If a
set has at least one element, it is called nonempty.
Here are other common definitions that we will use frequently:
Definition 1. A set Y is called a subset of set X, written Y ⊂ X, if every element
of Y is also an element of X (that is, x ∈ Y implies x ∈ X). If, in addition,
Y 6= X, then we say that Y is a proper subset of X.
Note that the set of “weekend days,” {Saturday, Sunday}, is a proper subset of
the set of days of the week.
Observe that if sets X and Y are both subsets of each other, then they must
be equal. Using the symbol ⇒, which means “implies,” this conclusion can be
expressed as: X ⊂ Y and Y ⊂ X ⇒ X = Y . Sometimes when we want to prove
that two sets are equal, we perform the two steps of showing that each is contained
in the other.
Definition 2. The union of sets X and Y , written X ∪ Y , is the set formed by
combining the distinct elements of the two sets. That is,
X ∪ Y ≡ {x | x ∈ X or x ∈ Y (or both)}.
The symbol “≡”, used first in the definition above, means “is defined as” or
“is equivalent to.”
Definition 3. The intersection of sets X and Y , written X ∩ Y , is the set formed
by collecting only the points that are common to both sets. That is,
X ∩ Y ≡ {x | x ∈ X and x ∈ Y }.
Definition 4. The difference between sets X and set Y , written X \ Y , is the set
formed by removing from X all points in the intersection of X and Y . That is,
X \ Y ≡ {x | x ∈ X and x 6∈ Y }.
When we are analyzing various sets, we typically have in mind a grand set that
contains all objects of interest, such that every set we study is a subset of the grand
set. This grand set is called the space, or universe, and is sometimes denoted U .
The next definition makes reference to the space.
1.2. FUNCTIONS 5

Figure 1.1: Proper subset.

A⊂B

Figure 1.2: Union of two sets.

A∪B

Definition 5. The complement of the set X, denoted X 0 or X c , is defined as

U \ X.
Example 1. Suppose that the space is all of the lowercase letters:

U ≡ {a, b, c, . . . , z}.

Consider X = {a, b, c, d} and Y = {d, e, f, g}. Then we see that X ∪ Y =

{a, b, c, d, e, f, g}, X ∩ Y = {d}, X \ Y = {a, b, c}, and X c = {e, f, g, . . . , z}.
For visual thinkers, Venn diagrams can be useful to represent the relations
between sets. Figure 1.1 represents that set A (the points inside the smaller circle)
is a proper subset of set B (the points inside the larger circle). The shaded region
of Figure 1.2 shows the union of sets A and B, whereas the shaded region in
Figure 1.3 is the intersection.

1.2 Functions
Often we are interested in representing ways in which elements of one set might
be related to, or associated with, elements of another set. For example, for sets X

Figure 1.3: Intersection of two sets

A∩B
6 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

Figure 1.4: A function α..

and Y , we might say that every x ∈ X points to, or “maps to,” a point y ∈ Y . The
concept of a “function” represents such a mapping.
Definition 6. A function f from a set X to a set Y is a specification (a mapping)
that assigns to each element of X exactly one element of Y . Typically we express
that f is such a mapping by writing f : X → Y . The set X (the points one “plugs
into” the function) is called the domain of the function, and the set Y (the items
that one can get out of the function) is called the function’s codomain.
Note that the key property of a function is that every element in X is associated
with exactly one element in Y . If you plug a point from X into the function, it
specifies just one point in Y that is associated with it. It is not the case that some
point x ∈ X maps to both y and y 0 such that y 6= y 0 .
We write f (x) as the point in Y that the function associates with x ∈ X. Also,
for any subset of the domain Z ⊂ X, we define

f (Z) ≡ {f (x) | x ∈ Z}.

We refer to this as the image of the set Z for the function f . The set f (X), which
is clearly a subset of Y , is called the range of the function. Finally, we define the
inverse image of a set W ⊂ Y as

f −1 (W ) ≡ {x ∈ X | f (x) ∈ W }.

This is the set of points in X that map to points in W . Note that the inverse image
does not necessarily define a function from Y to X, because it could be that there
are two distinct elements of X that map to the same point in Y .
Figures 1.4 and 1.5 depict functions from S to T . Figure 1.6 depicts a map-

Figure 1.5: A function β..

ping that is not a function. The mapping does not associate any element in T with
the point y ∈ S. Another violation of the requirements for a function is that the
mapping associates two elements of T (both 1 and 3) with the single point x ∈ S.
1.3. THE REAL LINE 7

Figure 1.6: A mapping that is not a function.

Example 2. With α and β shown in Figures 1.4 and 1.5, we have α({ x, z}) =
{ 2, 3} and β −1 ({1, 2}) = {x, z}.

Definition 7. Consider a function f : X → Y . If f (X) = Y then f is said to be

onto. That is, f is onto if for each y ∈ Y there is at least one x ∈ X that maps to
it.

Note that the function α in Example 2 is onto while the function β is not, since
its image is {1, 3} which is a proper subset of {1, 2, 3}, the codomain of β.

Definition 8. Consider a function f : X → Y . This function is said to be one-to-

one if f (x) = f (x0 ) implies that x = x0 . An equivalent condition is that for every
two points x, x0 ∈ X such that x 6= x0 , it is the case that f (x) 6= f (x0 ).

Note that the function α in Example 2 is one-to-one, wherease the function β

is not one-to-one because β(x) = β(z) and yet x 6= z.

Definition 9. A function f : X → Y is called invertible if the inverse image

mapping is a function from Y to X. Then, for every y ∈ Y , f −1 (y) is defined to
be the point x ∈ X for which f (x) = y.

Theorem 1. A function is invertible if and only if it is both one-to-one and onto.

This section concludes with the definition of a composition function, which

is basically a function formed by performing the operations of two functions in
order.

Definition 10. Suppose we have functions g : X → Y and f : Y → Z. Then the

composition function f ◦ g is a mapping from X to Z (that is, f ◦ g :: X → Z.
Writing h = f ◦ g, this function is defined by h(x) ≡ g(f (x)) for every x ∈ X.

1.3 The Real Line

There are several sets of numbers that we will work with often. Perhaps the most
important for everyday mathematics is the set of real numbers, denoted by R.
8 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

The set of reals comprises all of the ordinary numbers you are used to dealing
with, such as 0, 1, fractions like 1/2, decimals such as 4.23, and more exotic
numbers like the “natural number” e. The set includes both positive and negative
numbers, arbitrarily high numbers, and arbitrarily low (large negative) numbers.
Although many of its elements are quite familiar to most people, the defini-
tion of R is not as straightfoward. In fact, R is defined in relation to the real
number system, which includes the set R, operators of addition and multiplica-
tion, standard “identiity numbers” 0 and 1, and some axioms (assumptions). Here
is a brief description of how the set R is defined. It is okay to not study a complete
treatment, but realize that the real number system has some nuances.
A good place to start is with the set of positive integers:

P ≡ {1, 2, 3, . . .}.

One way to think of this set is that it is defined by the special “multiplicative iden-
tity” number 1 and the idea of addition. The number 1 is called the multiplicative
identify because 1 times any number is the same number. The set of positive in-
tegers comprises the number 1, 1 + 1, 1 + 1 + 1, and all such numbers formed by
adding 1 to itself multiple times.

Example 3. Define functions α and β from the set of positive integers P =

{ 1, 2, 3, . . . } to itself by α(n) = 2n for all n, and

(n + 1)/2, if n is odd,
β=
n/2, if n is even.

These are pictured in Figure 1.7. Note that α is one-to-one but not onto, and β is

Figure 1.7: α and β.

onto but not one-to-one.

Definition 11. If there exists a 1-1 function of X onto Y, we say that X and Y can
be put in a 1-1 correspondence, or that X and Y have the same cardinal number,
or briefly, that X and Y are equivalent, and we write X ∼ Y.

This relation has the following properties:

• It is reflexive: X ∼ X .
1.3. THE REAL LINE 9

• It is symmetric: If X ∼ Y, then Y ∼ X .
• It is transitive: If X ∼ Y and Y ∼ Z, then X ∼ Z.
Any relation with these three properties is called an equivalence relation.
Note: there may be lots of functions from X onto Y that are not 1-1, but for
this condition we only need to be able to find 1 such function!

Definition 12. For any positive integer n, let Pn be the set whose elements are the
integers 1, 2, . . . , n; Let P be the set consisting of all positive integers (which is
the set of natural numbers we already saw in Example 3).
For any set A, we say:
• A is finite if A ∼ Pn for some n.
• A is infinite if A is not finite.
• A is countable if A ∼ P.
• A is uncountable if A is neither finite nor countable.
• A is at most countable if A is finite or countable.
Note: see section ?? for a further discussion of countable versus uncountable.

Definition 13. The rational numbers, Q1 , are informally defined as the set of all
numbers of the form m/n, where m and n are integers and n 6= 0.

Definition 14. The Real Numbers, R2 , are informally defined as the set of all
regular numbers from negative to positive.
R = { x | −∞ < x < ∞}
Formally3 R is defined as a set of objects associated with a set of “operators”
(function from X ×X → X ) “+” and “.”, and special elements 0 and 1, satisfying
certain axioms:
1
Q is an example of a countable set. This may seem very confusing since there must be more
rational numbers than integers (mustn’t there be!), and will be discussed further in section ??
2
R is an example of an uncountable set
3
For mathsy types in the class; really first of all you have to define a field, then an ordered
field, and then R can be defined as the ordered field which has the least upper bound property
(to be defined soon) and which contains Q as a subfield. Finally you must show that R \ Q, the
irrationals, can be written as infinite decimal expansions and are considered “approximated” by
the corresponding finite decimals. If after all this you still care, see Rudin chapter 1.
10 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

• Addition: If x, y ∈ R ⇒ x + y ∈ R.

• Additive Identity: x + 0 = 0 + x, ∀ x ∈ R

• Multiplicative Identity: 1.x = x.1, ∀ x ∈ R

• Associativity: (x + y) + z = x + (y + z), ∀ x, y, z ∈ R

Definition 15. There is a set R++ ⊂ R such that

• If x, y ∈ R++ ⇒ x + y ∈ R++ .

• If x ∈ R ⇒ x ∈ R++ or x = 0 or −x ∈ R++
where -x ≡ so that (-x) + x = 0
It is fairly obvious from this that R++ is just the set of strictly positive real
numbers.

Definition 16. We say

x>y ⇐⇒ x − y ∈ R++

where
−y ≡ +(−y)

Definition 17. For a, b ∈ R where a ≤ b

[a, b] ≡ { x | a ≤ x ≤ b} (closed interval)
(a, b] ≡ { x | a < x ≤ b} (half-open interval)
[a, b) ≡ { x | a ≤ x < b} (half-open interval
(a, b) ≡ { x | a < x < b} (open interval / segment)

Definition 18. Take X ⊂ R

Then we say a ∈ R is an upper bound for X if

a ≥ x, ∀ x ∈ X

And we say b ∈ R is a lower bound for X if

b ≤ x, ∀ x ∈ X
1.3. THE REAL LINE 11

Note that there can be lots of upper bounds and lots of lower bounds.

We say “The set X is bounded” if “X is bounded from above” and “X is bounded

from below”.

Definition 19. We say a ∈ R is a least upper bound of X , or the supremum of X ,

if a ≥ x, ∀ x ∈ X (a is an upper bound), and if a0 is also an upper bound then
a0 ≥ a. We write
a = sup X

Definition 20. We say b ∈ R is a greatest lower bound of X , or the infimum of

X , if b ≤ x, ∀ x ∈ X (b is an lower bound), and if b0 is also an lower bound then
b0 ≤ b. We write
b = inf X

Example 4. The following illustrates the important point that the “sup” and
“inf” of a set do not have to be contained in the set.
For example let
X = [0, 1], and Y = (0, 1)
We can see that

sup X = sup Y = 1

and that

inf X = inf Y = 0

However, obviously the points { 0} and { 1} are not contained in the set Y.

Lemma 1. Any set X ⊂ R, (X 6= φ), that has an upper bound, has a least upper
bound (sup).

Proof. Left as exercise to the reader.4

4
Any reader able to complete this exercise without assistance should have a successful career
as a mathematician. Although the result seems intuitive, it is a deep property of the real numbers.
12 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

Definition 21. We define the following

max X ≡ a number a such that
a ∈ X and a ≥ x, ∀ x ∈ X
and
min X ≡ a number b such that b ∈ X and b ≤ x, ∀ x ∈ X
Referring back to Example 4 we can see that X = [0,1] has a max and min
max X = 1 and min X = 0

While Y = (0,1) does not have a max or min.

Note It should now be clear that the max and min do not always exist even if
the set is bounded, but the sup and the inf do always exist if the set is bounded.

Theorem 2. If max X exists then

max X = sup X
Proof. Let a ≡ max X
We will demonstrate that a is sup X .
To do this we must show that:
(i) a is an upper bound
(ii) Every other upper bound a0 satisfies a0 ≥ a
so
(i) a is an upper bound on X since by the definition of max
a ≥ x, ∀ x ∈ X
(ii) Consider any other upper bound a0 of the set X such that a0 6= a.
Since a0 is an upper bound, and a ∈ X (by the fact that a = max X ), we must have
a0 > a.

Definition 22 (Certain Sets). Positive Integers:

P ≡ { 1, 1 + 1, 1 + 1 + 1, . . . }

Integers:
Z ≡ P ∪ −P ∪ { 0}
1.4. METHODS OF PROOF 13

Rational Numbers: nm o
Q≡ | m, n ∈ Z, n 6= 0
n
It can be shown that
Q⊂R
and
Q 6= R

1.4 Methods of Proof

Proof by Direction Implication/Construction
Example 5. Show that
max X = sup X
when maxX exists.
Proof. Refer back to Theorem 2

Theorem 3 (Direct Implication). Distributive Law for Sets

A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C)
To prove this, let the left hand side and right hand side of the above equation be
denoted by D and E respectively. We will show that
D⊂E and that E ⊂D
Note. Subsets going in each direction imply equality!
Proof. Suppose x ∈ D. Then x ∈ A and x ∈ B ∪ C, that is, x ∈ B or x ∈ C
(possibly both). Hence x ∈ A ∩ B or x ∈ A ∩ C, so that x ∈ E. Thus D ⊂ E.
Next, suppose x ∈ E. Then x ∈ A ∩ B or x ∈ A ∩ C. That is, x ∈ A, and
x ∈ B ∪ C. Hence x ∈ A ∩ (B ∪ C), so that E ⊂ D.

It follows that D = E.

Proof by Contradiction These proofs usually begin with, let’s assume

the statement does not hold and then at the end we find that this cannot be.
14 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

Example 6. Suppose there is a two person world consisting of John and Mary,
and in this world everybody wears properly fitting clothes. Suppose people fit into
the clothes of people the same size or bigger, but do not fit into the clothes of
smaller people. We denote John’s height by x, and Mary’s height by y. We want
to show that John is taller than Mary, i.e. that

x>y

So the way we proceed is that we make the initial assumption that Mary is at least
as tall as John, i.e. y ≥ x. But then we note that John’s clothes are all bigger
than Mary’s clothes. So John must not fit into Mary’s clothes. Thus we must have
that our initial assumption is false. So

x>y

Now this is a pretty simple and stupid example but highlights the way to proceed.

Example 7. Suppose we have 2 sets A and B. And we have that A is a subset of

B. They are both subsets of the universal set U.

FIGURE 9 GOES HERE

The following statements are all equivalent (you should convince yourself of
this)

• If x ∈ A ⇒ x ∈ B,
⇒A⊂B

• If x ∈ B c ⇒ x ∈ Ac ,
⇒ B c ⊂ Ac

• x 6∈ B ⇒ x 6∈ A

Basically this method of proof hinges on showing that if we have 2 statements p

and q. Then

p⇒q ⇐⇒ ¬q ⇒ ¬p

Lemma 2.
1.4. METHODS OF PROOF 15

X ⊂ R, a = supX ,  > 0 (let this be statement p)

⇒ ∃ x ∈ X such that a −  < x ≤ a (let this be statement q)

Proof. Suppose the conclusion (i.e. statement q) does not hold.

Assume there is no x ∈ X such that

a−<x≤a

By the definition of supremum, we know that a is an upper bound.

Since there are no points in X between a −  and a (and obviously nothing above
a), we must conclude that a −  is an upper bound of X .
But this contradicts that a is supX .

Proof by Induction This involves making a statement for n,

S(n) = “claim . . . has to do with n”

To prove this claim for all n, it’s sufficient to prove

• S(1) is true

• S(n) ⇒ S(n + 1)

Example 8. Show that

n
X 1
k 3 = n2 (n + 1)2
k=1
4

Proof. • Observe that

1
S(1) = (1)2 (1 + 1)2
4
=1

• Suppose the claim is true for n = m

m
X 1
k 3 = m2 (m + 1)2
k=1
4
16 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

Now look at n = m + 1
m+1
X m
X
3
k = k 3 + (m + 1)3
k=1 k=1
1 2
= 4
m (m + 1)2 + (m + 1)3
= (m + 1)2 ( 41 m2 + m + 1)
= 14 (m + 1)2 (m2 + 4m + 4)
= 41 (m + 1)2 (m + 2)2

So you can see that if we assume the original statement is true for m and now
everywhere that we had an m we have an m + 1. Thus it must be true for m + 1.
And so we are done.

1.5 Some helpful notes

If and only if
You will have noticed in preceding sections mentioning of things like “if and
only if ” (which is often abbreviated to iff ), and phrases like “A is necessary for
B”, or “x is sufficient for y”. You will also have noticed the symbol ⇐⇒ .
So what the hell do all these things mean? Thankfully they are all very closely
related.
The symbol ⇐⇒ is just the mathematical symbol for “if and only if”5 . But
that’s not particularly helpful if we don’t know what “if and only if” means.
The difference between “if” and “iff” is as follows. Compare the 2 sentences
below:

1. I will drink beer if the beer is a Guinness. (equivalently: If the beer is a

Guinness, then I will drink it)

2. I will drink beer if and only if the beer is a Guinness.

Sentence (1) says only that I will drink Guinness. It does not rule out that I
may also drink Budweiser. Maybe I will, maybe I won’t - there just is not enough
information to determine. All we know for sure is that I will drink Guinness.
5
in fact many symbols are used for if and only if, such as ⇐⇒ , ≡, and ←→, so watch out as
different authors may use different ones.
1.5. SOME HELPFUL NOTES 17

Sentence (2) makes it quite clear that I will drink Guinness and Guinness only.
I won’t drink any other type of beer6 . Also, I will definitely drink the beer if it is
a Guinness.
This may seem confusing so perhaps we should look at how proofs involving
“iff” are special. Consider the First Fundamental Theorem of Asset Pricing due
to Harrison and Kreps7 .
Theorem 4. The finite market model is viable if and only if there exists an equiv-
alent martingale measure.
So perhaps it is better to look at the statement of the theorem more carefully
(with some shortening so as to fit the page better).
Theorem 5. viable
| {z } if and only if equivalent martingale measure.
| {z }
A B

Or in mathematical symbols
Theorem 6. A ⇐⇒ B.
Now you might have noticedn that the ⇐⇒ is sort of a fusion of the symbols
⇐=, and =⇒. This was not an accident. To prove an iff statement, you must
prove the implication both ways. To see what is meant by this let’s look at a mock
proof for the above theorem.
Basically to prove a statement A ⇐⇒ B, you must show
• A =⇒ B

• B =⇒ A
Of course you can also prove that
• A =⇒ B

• ¬A =⇒ ¬B
since recall that proving B ⇒ A is the same as proving ¬A ⇒ ¬B 8 .
Ok, so now let’s run through a mock proof9 of The First Fundamental Theorem
of Asset Pricing
6
I’m not really like this:)
7
Don’t worry if you don’t understand what the Theorem means, the method of proof is what’s
important
8
or now to show we really get it, we could write [B ⇒ A] ⇐⇒ [¬A ⇒ ¬B]!
9
There is not even a hint of how to prove the theorem here. I’m just trying to show you how
you go about proving such types of statements.
18 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

Proof. (⇒) So to prove the implication this direction, we first assume that the
finite market model is viable. Or more simply, we assume A. Thus, having
made this assumption, we now have to show that armed with this assump-
tion, that there exists an equivalent martingale measure. We have to show
B.

(⇐) Now we must go the other way. So we start by assuming there is an equiv-
alent martingale measure, i.e. we assume B. Then using only this assump-
tion, we must some how end up with the statement that the finite market
model is viable (statement A).

Once we do both of these we are done. Make sense??

A word of caution. If you are reading a paper, and it makes a Definition

statement, then any if in this type of statement should be read as iff. This is true
since it is a definition, so by definition the if is an iff (what!).

For example, see back to Definition 18 on page 10. For clarity’s sake I’ll
restate the first part.

Definition 23. Take X ⊂ R

Then we say a ∈ R is an upper bound for X if

a ≥ x, ∀ x ∈ X

So obviously if a ≥ x, ∀ x ∈ X , then we can say a ∈ R is an upper bound

for X . But by the same token, since it is a definition, if a ∈ R is an upper bound
for X , then we can also say that a ≥ x, ∀ x ∈ X . Make sense???

Necessary and Sufficient

Supposing we have a statement P ⇐⇒ Q. Then we can also say P is
“necessary and sufficient” for Q. Since proving P ⇐⇒ Q, requires implications
going 2 different directions, and “necessary” and “sufficient” are a list of 2 words,
you can probably guess that there is a connection between them. In fact they mean
equivalent things. Or to be quite cryptic

[P ⇐⇒ Q] ⇐⇒ [P is necessary and sufficient for Q]

1.5. SOME HELPFUL NOTES 19

Necessary Condition

As usual, maybe it’s best to start with an example.

Example 9. Suppose that I am the greatest football10 player in the world. I am

therefore the best football player in Ireland too. So we could make a statement
like:

“I am the best football player in Ireland if I am the best football player in the
world” (equivalently: If I am the best football player in the world, then I am the
best football player in Ireland).

Note however that this is not an if and only if statement. It does not go both
ways. If I am the best football player in Ireland, we cannot conclude that I am the
best football player in the world (one of those Brazilians might be better than me!).

Thus we would say that being the best football player in Ireland is a necessary
condition for being the best football player in the world, since if I am the best
football player in the world, then I am necessarily the best football player in
Ireland.

Sufficient Condition

Again it’s probably best to start with an example.

Example 10. Supposing you are studying for the mathcamp final while sitting
on a bench on the cliffs overlooking Torrey Pines. Suppose you are becoming
increasingly frustrated with the course and are really getting sick of studying. You
decide to hurl your notes off the cliff so that they land in the sea where they will
never be seen again. Then we would say that the hurling of the notes is a sufficient
condition for the notes to land in the sea. But the notes landing in the sea does
not imply that they were hurled: you might have slipped and dropped them, or a
gust of wind may have come and taken them. Thus the hurl was sufficient but not
necessary for the notes to land in the sea.

10
football refers to soccer not “American Football”
20 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

Relationship between necessity and sufficiency

You guessed it, necessity and sufficiency are dual to one another. Look back to
example 9. We have that the following are all equivalent:

• “(best football player in Ireland) if (best football player in the world)”

• (best football player in the world) ⇒ (best football player in Ireland)

• (best football player in the world) is sufficient for (best football player in
Ireland)

• (best football player in Ireland) is necessary for (best football player in the
world)

Infinity
The notion of infinity is one of the most seemingly simple yet difficult con-
cepts in all of mathematics. The symbol for infinity is ∞. Intuitively, ∞ just
represents a really “big” number that “beats” all other numbers. We got an intro-
duction to how to measure the size of sets on pages 8 and 9.

But before we get to comparing sizes of sets, let’s just look at how infinity is
defined.

Definition 24. The extended real number system, R, consists of R11 and two sym-
bols, +∞ and −∞. We define

−∞ < x < +∞

for every x ∈ R.

We make the following assumptions concerning ∞12 .

11
recall from definition 14 that R is defined as the set { x | −∞ < x < ∞}, where the inequal-
ities are strict.
12
for those who care, the extended real number system does not form a field, but the adding of
+∞ and −∞ does form a 2-point compactification of the topological space of real numbers.
1.5. SOME HELPFUL NOTES 21

• If x ∈ R then
x+∞=∞
x − ∞ = −∞
x x
= =0
+∞ −∞
• If x > 0 then
x · (+∞) = +∞
x · (−∞) = −∞

• If x < 0 then
x · (+∞) = −∞
x · (−∞) = +∞

So point is infinity wins, always. However the following is a very important

assumption:
∞ − ∞ = undefined
It is not equal to zero!13 This is because it is possible for one number to be “more
infinite” than another as we will soon see.

Now recall definitions 11, 12. In fact let’s just state them again

Definition 25. If there exists a 1-1 function of X onto Y, we say that X and Y can
be put in a 1-1 correspondence, or that X and Y have the same cardinal number,
or briefly, that X and Y are equivalent, and we write X ∼ Y.

Note: there may be lots of functions from X onto Y that are not 1-1, but for
this condition we only need to be able to find 1 such function!

Definition 26. For any positive integer n, let Pn be the set whose elements are the
integers 1, 2, . . . , n; Let P be the set consisting of all positive integers (which is
the set of natural numbers we already saw in Example 3).

For any set A, we say:

• A is finite if A ∼ Pn for some n.

• A is infinite if A is not finite.

13
this is why R cannot be a field since every element minus itself must be the “zero” element.
22 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE

• A is countable if A ∼ P.
• A is uncountable if A is neither finite nor countable.
• A is at most countable if A is finite or countable.

The notion of cardinality was always used to compare finite sets, and is very
intuitive for such finite sets. But the notion, as described above to describe infinite
sets was first used by Georg Cantor. He said that any set which could be put in
one-to-one correspondence with the set of natural numbers Z is countably infinite.
Loosely, what this means is that we can label or index all the elements of any
countable set with the integers. Now it turns out that the rational numbers Q can
be put in a one-to-one correspondence with Z and as such are countable. Now this
seems very unintuitive since there are obviously more rationals since they also
include fractions. The point is you can keep going further out along the set of
integers and as such can keep indexing the rationals. Every time you need another
index it is there. This is not an easy concept. In fact the natural numbers Z have
the same cardinality as the positive integers, P even though P ⊂ Z. Confusing
eh?? To make it worse, a countable union of countable sets is a new set that is also
countable! For a proof of the countability of the rationals Z see Rudin Theorem
2.13.

Now the reals, R, are an example of an uncountable set. Loosely, what this
means is that they can not be put in a one-to-one correspondence with Z. There
are just too many of them to index by the integers. Below we list some other sets
of cardinality C.

Example 11 (Sets with cardinality C).

• the real numbers R

• any closed or open interval in R (for example the unit interval [0, 1])
• the irrational numbers (R \ Q)

Yes that’s right, there are more numbers within [0, 1] than there are in Q! And
the sets R and [0, 1] are of the same cardinality even though obviously [0, 1] is a
tiny subset of R.

Finally, to really confuse you, Q is dense in R which means if x ∈ R, y ∈ R,

and x < y, then there exists a z ∈ Q such that x < z < y. So what this says is
that between any two real numbers there is a rational one.
1.5. SOME HELPFUL NOTES 23

But then mustn’t the cardinality of R be at most twice that of Q? And from
the statement above, twice a countable set would definitely be countable! What’s
wrong with this?
24 CHAPTER 1. SETS, FUNCTIONS, AND THE REAL LINE
Chapter 2

Sequences

2.1 Introduction

2.2 Sequences

Definition 27. A sequence is a function f from P to R.

f : P 7−→ R
If f (n) = an , for n ∈ P, it is customary to denote the sequence f by the symbol
{ an } ∞
n=1 , or sometimes by {a1 , a2 , a3 , . . . }. The values of f , that is, the elements
an , are called the terms of the sequence.
Note that the terms a1 , a2 , a3 , . . . of a sequence need not be distinct.

Example 12. Here are some sequences:

an = n { an }∞
n=1 = {1, 2, 3, . . . }
an+1 = an + 2, ∀ n ≥ 1 { an }∞
n=1 = {1, 3, 5, . . . }
n ∞
an = (−1) , n ≥ 1 { an }n=1 = {−1, 1, −1, 1, . . . }

Definition 28. [Convergent Sequence] A sequence { an } is said to converge in R,

if there is a point a ∈ R with the following property: For every 1 > 0, ∃ N ∈ P
such that n ≥ N implies
|an − a| < 
1
This is our first encounter with  (epsilon). It is just the standard symbol in mathematics to
denote a really tiny, tiny, tiny, but still positive amount.

25
26 CHAPTER 2. SEQUENCES

In this case we also say that { an } converges to a, or that a is the limit of { an },

and we write
an −→ a
or
lim an = a
n−→∞

If { an } does not converge, it is said to diverge.

The definition of a convergent sequence in a set X requires that the limit be an

element of X. Hence, convergence depends not only on { an } but also on the the
space in which it is defined. Strictly speaking, it is incorrect to say that a sequence
“converges,” you must also specify the space. For example the sequence n1


converges in R (to 0), but fails to converge in the set of all positive real numbers
R++ .

Theorem 7. Let {an } be a sequence in R.

If b ∈ R, b0 ∈ R, and if {an } converges to b and to b0 , then

b = b0

Let d(x, y) ≡| x − y |.
Proof. Let  > 0 be given. There exist integers N and N 0 such that

n ≥ N =⇒ d(an , b) <
2

n ≥ N 0 =⇒ d(an , b0 ) <
2
0
Hence, if n ≥ max (N, N ), we have

d(b, b0 ) ≤ d(b, an ) + d(b0 , an ) < 

Since  was arbitrary, we conclude that d(b, b0 ) = 0.2

The theorem means that it is appropriate to say the limit of a sequence when
the limit exists. The proof formalizes the intuition that elements of sequence can-
not be arbitrarily close to two different things at the same time.

2
d here is called a metric. You can think of it as being function which takes in any two points
in a set, and then spits out the distance between the two of them. It will be more formally defined
in Definition 72.
2.2. SEQUENCES 27

Definition 29. Formally, a subsequence is a composition mapping3 . A subse-

quence of the function f from P to R.

f : P −→ R

is given by a function
f ◦ g : P −→ R
where
g : P −→ P
and g is assumed to be strictly increasing4 .

FIGURE GOES HERE

Alternatively, given a sequence { an }, consider a sequence { nk } of positive

integers, such that n1 < n2 < n3 < · · ·. Then the sequence { ani } is called a
subsequence of { an }. If { ani } converges, its limit is called a subsequential limit
of { an }.

If you think of a sequence as an ordered list, then one obtains a subsequence

by deleting terms from the sequence (as long as you leave infinitely many terms).

It can be shown that { an } converges to b if and only if every subsequence of

{ an } converges to b.

Definition 30. A sequence of real numbers {an } of real numbers is said to be

1. monotonically increasing if

an ≤ an+1 (n = 1, 2, 3, . . . )

2. monotonically decreasing if

an ≤ an+1 (n = 1, 2, 3, . . . )
3
see back to Definition 10 for what exactly is a composition mapping
4
we will see exactly what it means for a function to be strictly increasing soon, but for now just
think of it as always going up
28 CHAPTER 2. SEQUENCES

Note we say strictly increasing/decreasing if the above weak inequalities are re-
placed with strict inequalities.

Definition 31. A function f is said to be

1. monotonically increasing on (a, b) if
a<x<y<b =⇒ f (x) ≤ f (y)

2. monotonically decreasing on (a, b) if

a<x<y<b =⇒ f (x) ≥ f (y)

Note we say a function f is strictly increasing/decreasing if the above weak in-

equalities are replaced with strict inequalities.

A function is monotonically increasing if it never takes a “step” backwards.

Each time you go further on in the sequence you are at a value at least as high
as the one before. If the function is strictly increasing, then the function not only
never takes a “step” backwards, it always takes a step forwards. In other words,
each time you go further on in the sequence, you always take a step forwards.
Moving to a value the same as the one you are at is no longer allowed.

Theorem 8. Suppose
lim an = b
n−→∞

then all subsequences of {an } converge to b as well.

The converse is not true. That is, if a subsequence has a limit - this does not imply
that the sequence necessarily has a limit5 .
Example 13. Consider the sequence
 
1 1 1 1
{ an } = 1, 1, 1, , 1, , 1, , 1, , 1, . . .
2 3 4 5
So we can tell from this that the subsequence obtained by taking the even elements
of the above sequence has a limit of 06 . But this is clearly not the limit of the orig-
inal sequence since the number 1 keeps appearing! Moreover, the subsequence
5
As a test of whether or not you’ve grasped necessary versus sufficiency, which is which in this
case?
6
thought I didn’t define the space on which the sequence is defined so technically this is not
correct!
2.2. SEQUENCES 29

obtained by taking the odd elements of the original sequence clearly has a limit of
1 (since we are only taking the element 1 over and over again)!! (since recall that
the terms of a sequence need not be distinct, and a subsequence is a sequence in
its own right).

Theorem 9. Suppose {an } and {bn } are sequences, and

lim an = a
n−→∞

lim bn = b
n−→∞

Then
1.
lim (an + bn ) = a + b
n−→∞

2.
lim can = ca, c ∈ R
n−→∞

3.
lim an bn = ab
n−→∞

4.
lim (an )k = ak
n−→∞

5.
1 1
lim =
n−→∞ an a
6 0(n = 1, 2, 3, . . . ), and a 6= 0
provided an =

6.
an a
lim =
n−→∞ bn b
6 0(n = 1, 2, 3, . . . ), and b 6= 0
provided bn =
7
Proof. 1. Given  > 0, ∃ integers N1 , N2 such that

n ≥ N1 =⇒ |an − a| < ,
2
7
this proof really uses the triangle inequality which will not show up til Definition 72, but
which intuitively just states that the shortest distance between any two points is a straight line
between them.
30 CHAPTER 2. SEQUENCES


n ≥ N2 =⇒ |bn − b| < .
2
If we take N = max (N1 , N2 ), then n ≥ N =⇒

|(an + bn ) − (a + b)| ≤ |an − a| + |bn − b| < 

2. This proof is very easy - just factor out the c and then put it in again at the
end

3. This uses a trick

an bn − ab = (an − a)(bn − b) + a(bn − b) + b(an − a) (2.1)

Given  > 0, there are integers N1 , N2 such that

√
n ≥ N1 =⇒ |an − a| < ,
√
n ≥ N2 =⇒ |bn − b| < .
If we take N = max(N1 , N2 ), n ≥ N =⇒

|(an − a)(bn − b)| ≤ ,

so that
lim (an − a)(bn − b) = 0.
n−→∞

We now apply results (1) and (2) to Equation 2.1, and conclude that

lim (an bn − ab) = 0

n−→∞

4. This is left as an exercise to the reader

5. Choosing m such that |an − a| < 21 |a| if n ≥ m, we see that

1
|an | > |a| , (n ≥ m)
2
Given  > 0, ∃ an integer N > m such that n ≥ N =⇒
1 2
|an − a| < |a| .
2
Hence for n ≥ N ,
   
1 1   an − a  2
 − =
 an a   an a  < |a|2 |an − a| < 

2.2. SEQUENCES 31

6. This is really just combining previous results

Definition 32. We say the sequence {an } is bounded above if ∃ m ∈ R such that
an ≤ m, ∀n ∈ P
We say the sequence {an } is bounded below if ∃ m ∈ R such that
an ≥ m, ∀n ∈ P
Note: that a bounded sequence does not necessarily converge, for example the
sequence {0, 1, 0, 1, 0, 1, . . . } just oscillates back forth between {0} and {1} for-
ever.

Theorem 10. If {an } converges then {an } is bounded.

Proof. Suppose an −→ a. There is an integer N such that n > N implies
|an − a| < 1. Put
r = max {1, |a1 − a| , . . . , |aN − a|}
Then |an − a| ≤ r for n = 1, 2, 3, . . . .
Another way to think of this is that after a certain n (i.e. n = 1000) we don’t
have to worry since it must converge - so it’s definitely bounded. So then what
happens before this? Well before this, it is a finite sequence - and a finite sequence
must be bounded since it must have a max and a min. This last result can be proved
by induction:
Is a set of 1 element bounded? Obviously!
Is a set of 2 elements bounded? Obviously!
.. ..
. .
Is a set of n elements bounded? Obviously!

Theorem 11. Let {an }, {bn }, and {cn } be sequences such that
an ≤ bn ≤ cn , ∀n ∈ R
and we have that both
an −→ a, cn −→ a
=⇒ bn −→ a
32 CHAPTER 2. SEQUENCES

Proof. The proof is left as an exercise to the reader8 .

Theorem 12. A monotone sequence is convergent if and only if it is bounded9 .

Proof. This is our first full “if and only if” proof. Recall from section ?? that to
prove this we have to prove 2 things.
only if
• ( =⇒ ) We have a monotone sequence that is convergent, and must prove
that this implies that it’s bounded.
if
• (⇐=) We have a monotone sequence that is bounded and must prove that
this implies convergence.
To reiterate if and only if (iff) statements must be proved going both ways.

only if
1. ( =⇒ ) We already showed that any convergent sequence is bounded in The-
orem 10. So obviously a monotone sequence is bounded!10

if
2. (⇐=) Suppose { an } is bounded. Thus by Lemma 1, we know that the set
X ≡ { an | n ∈ P} has a sup which we will denote a. We must show that a
is the limit of { an }. We need to show that ∀  > 0, ∃ N ∈ P such that

|a − an | < , ∀n ≥ N

Take any  > 0. Since a −  is not an upper bound of X , we know from

Lemma 2 that ∃ N such that aN > a − . Since {an } is increasing, for every
n ≥ N we have
a −  < a N ≤ an ≤ a
since we also have that an ≤ a, ∀ n.
Thus,
|an − a| < , ∀n ≥ N
8
Think about this. If the first sequence converges, and the third sequence converges, and we
are told that every element of the second sequence is “sandwiched” between the corresponding
element of the first and third sequence then this sequence should also converge to the same point.
Makes sense right???!!! So you’ll find that often times you know exactly what you have to prove,
but it can still be difficult to actually do it.
9
compare this to Theorem 10. What’s different? Why has “if” gone to “iff”?
10
Since the class of monotonic sequences is a subset of the class of all sequences, then obviously
this applies to monotone sequences.
2.2. SEQUENCES 33

Theorem 13. Every real sequence has a monotone subsequence.

Proof. Take a sequence {xn }. We must find a subsequence {xnk } that is either
increasing or decreasing.
One of two conditions must hold (it is possible for both to hold):
1. the set {xn | n ≥ N } has a maximum for all N .
And if this is true we should be able to form a decreasing sequence
2. the set {xn | n ≥ N } doesn’t have a maximum for some N ∈ P

Note: that this theorem is not saying that there exists a unique monotone subse-
quence. Indeed, if there is one monotone subsequence, there will be infinitely
many (generate a new one by deleting the first term of the original one).
1. Form a sequence of P by:
xn1 = max xn
n>1

xn2 = max xn
n>n1
..
.
xnk+1 = max xn
n>nk

So we see by construction that this sequence {xnk } is decreasing. (Convince

yourself of this!) Also note that the set {xn | n > N } is decreasing in size
(number of elements) as N increases.
2. For each n > N , we can find m > n such that xm > xn . If we can’t do
such a thing then the set has a maximum so we are in case 1 and there is no
reason to proceed.

Define,
xn1 ≡ xN ,
xn2 ≡ the first term following xn1 for which xn2 > xn1
.. ..
. .
xnk+1 ≡ the first term following xnk for which xnk+1 > xnk .

And we can see that {xnk } is increasing by construction.

34 CHAPTER 2. SEQUENCES

Theorem 14 (Bolzano-Weierstrauss). Every bounded, real sequence has a con-

vergent subsequence.
Proof. This is got by applying the last 2 theorems.
1. Every real sequence has a monotone subsequence
if
2. We know from applying Theorem 12 using the (⇐=) direction case, that
every bounded monotone sequence is bounded.

Definition 33. A sequence {an } is said to be a Cauchy Sequence if for every  > 0
there is an integer N such that
|an − am | < 
if n ≥ N and m ≥ N .
Refer back to the definition of a convergent sequence (Definition 28) to see the
difference between the definition of convergence and the definition of a Cauchy
sequence. The difference is that the limit is explicitly involved in the definition of
convergence but not in that of a Cauchy sequence11 .

Theorem 15. A real sequence {an } is convergent if and only if it is a Cauchy

sequence12 .

Definition 34. X ⊂ R is a compact space if every sequence in X has a convergent

subsequence in X .
Example 14. Consider the sequence defined by an = n1
 
∞ 1 1 1
=⇒ { an }n=1 = 1, , , , . . .
2 3 4
On R it is obvious that this converges to 0. But on the set X = (0, 1) it does not
converge (as will be explained below). Here we say that the set X is not compact.
11
note above that I did not specify a space in which the sequence is Cauchy. This is sloppy.
12
This result is usually called the Cauchy criterion. Also note that {an } above was stated to be
real, so the space was stated this time.
2.2. SEQUENCES 35

Definition 35. Take any set X ⊂ R.

Call a ∈ R a limit point of X if there is a sequence {an }, (where { an | n ∈ P} ⊂
X ) such that
an −→ a.

Note: that a 6∈ X is possible, so a set need not include it’s limit points.

Definition 36. X is called a closed set if it contains all its limit points.

(That is, every sequence that converges in X does so to a point in X .)

Definition 37. X is called open if for every x ∈ X , ∃  > 0 such that

(x − , x + ) ⊂ X

A note on notation:
As you may have seen before that a closed set from the point x to the point y is
denoted with square brackets
[x, y]
while an open set from x to y is denoted with round brackets

(x, y)

So the only difference is that the open set does not contain the points x and y but
does include every point right up next to them!

Recall the set X from example 14. We can see that this set X has a limit point
(namely the point a = 0) but that no point of X is a limit point of X . It’s important
to understand the difference between having a limit point and containing one!

Theorem 16. X ⊂ R is compact if and only if it is closed and bounded.

36 CHAPTER 2. SEQUENCES
Chapter 3

Functions and Limits of Functions

Take any function f ,

f : X −→ R
for X = (c, d) (so note that it is an open set) and take a ∈ X .

Definition 38. y is called the limit of f from the right at a (right hand limit) if,
for every  > 0, there is a δ such that

0<x−a<δ =⇒ |f (x) − y| < 

this is also denoted as

y = lim+ f (x)
x→a

Note:We are only considering x > a, so x is to the right of a. Also x 6= a.

Definition 39. y is called the limit of f from the left at a (left hand limit) if, for
every  > 0, there is a δ > 0 such that

0<a−x<δ =⇒ |f (x) − y| < 

this is also denoted as

y = lim− f (x)
x→a

Note: Now it is a − x since we are coming from the left, so this must be positive.

37
38 CHAPTER 3. FUNCTIONS AND LIMITS OF FUNCTIONS

FIGURE GOES HERE

Definition 40. y is called the limit of f at a if

y = lim− f (x)
x→a
= lim+ f (x)
x→a

and we write
y = lim f (x)
x→a

y is the limit of f at a if for every  > 0, there is a δ > 0 such that

0 < |x − a| < δ =⇒ |f (x) − y| < 

Note: The definition of the limit of a function at the point a does not require
the function to be defined at a.
The use of  and δ in these definitions is a standard part of mathematical anal-
ysis, but takes practice to appreciate. The formal definition of limit at a point
is a way to capture the idea of the behavior of a function “near” the point. If
limx→a f (x) = L these means, roughly, that values of f (x) are near L when x is
near to a.
In order to use the definition, you must answer infinitely many questions: for
each  you must say how close x must be to a to guarantee that f (x) is within 
of L. The way that you do this is to define δ as a function of  and then attempt to
confirm the definition.

Example 15. Consider the function


0, if x 6= 0,
f (x) =
1, if x = 0.

FIGURE GOES HERE

It is clear from the picture that

lim f (x) = 0 = lim f (x)

x→0− x→0+
 39

⇒ lim f (x) = 0
x→0

But the value of the function at the point 0 is not equal to the limit. The example
shows that it is possible for limx→a f (x) to exist but for it to be different from
f (a). Since you can define the limit without knowing the value of f (a), this
observation is mathematically trivial. It highlights a case that we wish to avoid,
because we want a function’s value to be approximated by nearby values of the
function.
Theorem 17. Limits are unique. That is, if limx→a f (x) = L and limx→a f (x) =
L0 , then L = L0 .
Proof. Assume that L 6= L0 and argue to a contradiction. Let  = |L − L0 | /2.
Given this  let δ ∗ > 0 have the property that |f (x) − L| and |f (x) − L0 | are less
than  when 0 < |x − a| < δ ∗ . (This is possible by the definition of limits.) Since

|f (x) − L| + |f (x) − L0 | ≥ |L − L0 | (3.1)

it follows that  ≥ |L − L0 |, which is not possible.

The theorem simply states that the function cannot be close to two different
things at the same time. By the way, the critical inequality (inequality (3.1)) is
called the triangle inequality. You can write it as

|a − b| ≤ |a| + |b|

This expression is the same as |a − b| ≤ |a| + |b| – just replace b by −b. The
left-hand side is the distance between a and b. The right-hand side is the distance
between 0 and a plus the distance between 0 and b. Hence the inequality says that
it is shorter (or equal) to go directly from a to b than to go from a to 0 and then
0 to b. In one dimension you can check that the inequality is true by using the
definition of absolute value, by drawing a simple picture, or by thinking about the
meaning of the three different terms.
If a limit exists, then both limit from the left and limit from the right exist (and
they are equal) and, conversely, if both limit from right and limit from left exist,
then the limit exists. These statements require proofs, but the proofs are simple.
You do not want to use −δ proofs whenever you need to find a limit. To avoid
tedium, you need to collect a few “obvious” limits (for example, if the function f
is constant, then limx→a f (x) exists for all a and is equal to the constant value of
f ), and then some basic results that permit you to compute limits.
Theorem 18. If f and g are functions defined on a set S, a ∈ (α, β) ⊂ S and
limx→a f (x) = M and limx→a g(x) = N , then
40 CHAPTER 3. FUNCTIONS AND LIMITS OF FUNCTIONS

1. limx→a (f + g) (x) = M + N

2. limx→a (f g) (x) = M N
 
3. limx→a fg (x) = M/N provided N 6= 0.

Proof. For the first part, given  > 0, let δ1 > 0 be such that if 0 < |x − a| <
δ1 then |f (x) − M | < /2 and δ2 > 0 be such that if 0 < |x − a| < δ2 then
|g(x) − N | < /2. This is possible by the definition of limit. If δ = min{δ1 , δ2 },
then 0 < |x − a| < δ implies

|f (x) + g(x) − M − N | ≤ |f (x) − M | + |g(x) − N | < /2 + /2 < .

The first inequality follows from the triangle inequality while the second uses the
definition of δ. This proves the first part of the theorem.
For the second part, you can use the√ same type of argument, setting δ1 so that
if 0 < |x − a| < δ1 then |f (x) − M | <  and so on.
For the third part, note that when g(x) and N are not equal to zero

f (x) M g(x) (f (x) − M ) + f (x) (N − g(x))

− = .
g(x) N g(x)N
|f (x)| 2|M |
So given  > 0, find δ so small that if 0 < |x − a| δ, then |g(x)N |
< N2
,
N2
|f (x) − M | < N /2 and |g(x) − N | < 4|M |
/2.
This is possible provided that
 
 f (x) 
N 6= 0. (In this part I constructed an upper bound for  g(x)N  using the fact that
f (x) is near M and g(x) is near N when x is a.

Using this theorem you can generate the limits of many functions by combin-
ing more basic information.

Definition 41. We say f : X 7−→ R is continuous at a if

f (a) = lim f (x)

x→a

Note: f is said to be continuous at the point a. If f is continuous at every point of

X , then f is said to be continuous on X .
Note: the limit exists and the limit of the function is equal to the function of the
limit. It should also be noted that the function has to be defined at a. Note: In
order for a function to be continuous at a, it must be defined “in a neighborhood”
 41

of a – that is, the function must be defined on an interval (α, β) with a ∈ (α, β).
We extend the definition to take into account “boundary points” in a natural way:
We say that f defined on [a, b] is continuous at a (resp. b) if limx→a+ f (x) = f (a)
(resp. limx→b− f (x) = f (b)). Note: Informally one could remember the definition
of continuity as:
f (lim x) = lim f (x).
x→a x→a

Continuity at a point requires two things: a limit exists and the limit is equal
to the right thing. It is easy to think of examples in which the limit exists, but
is not equal to the value of the function. The limit of a function can fail to exist
for two different reasons. It could be that the left and right hand limits exist,
but are different. Alternatively, either left or right-hand limit may fail to exist.
This could happen either because the function is growing to infinity (for example,
f (x) = 1/x for x > 0 is continuous at any point on a > 0, but cannot be
continuous at 0 no matter how f (0) is defined. Alternatively, the function can
oscillate wildly (a standard example is sin( x1 ) near x = 0.
Lemma 3. f : X 7−→ R is continuous at a if and only if for every  > 0, there is
a δ > 0 such that
0 < |a − x| < δ =⇒ |f (x) − f (a)| < 

We say “f is continuous” if f is continuous at every point a ∈ X .

Note: a very informal way of describing a continuous function is that you can
draw the function without lifting your hand off the page, by sweeping the pen in
one fluid, continuous motion!

As in the case of limits, we can identify whether combinations of functions are

continuous without using the definition directly. The sums, products, and ratios of
continuous functions are continuous (in the last case the denominator of the ratio
must be non-negative in the limit).
Theorem 19. For
g : X −→ Y
and
f : Y −→ R
where both X and Y are open intervals of R, if g is continuous at a ∈ X and if f
is continuous at g(a) ∈ Y, then
f ◦ g : X −→ R
42 CHAPTER 3. FUNCTIONS AND LIMITS OF FUNCTIONS

is continuous at a.1

Proof. Let  > 0 be given. Since f is continuous at g(a), there exists γ > 0 such
that
|f (y) − f (g(a))| < 
if
|y − g(a)| < γ
and y ∈ g(X ).
Since g is continuous at a, there exist δ > 0 such that

|g(x) − g(a)| < γ

if
|x − a| < δ
and x ∈ X .
It follows that
|f (g(x)) − f (g(a))| < 
if
|x − a| < δ
and x ∈ X .

Thus f ◦ g is continuous at a.

It is easy to show that constant functions and linear functions are continuous.
By the combining properties, this allows you to conclude polynomials and ratios
of polynomials are continuous.

Theorem 20. The function

f : (a, b) −→ R
is continuous at a point x if and only if, for all sequences {xn } ⊂ (a, b) with
xn −→ x, it is the case that

f (xn ) −→ f (x)
1
again note that this is continuity at a point.
 43

only if
Proof. 1. ( =⇒ ) Let’s use contradiction method of ¬B =⇒ ¬A (to prove
A =⇒ B)
Suppose ∃ a sequence {xn } ⊂ (a, b) with xn → x but f (xn ) 6→ f (x). This
means that ∃  > 0 such that ∀ δ > 0, ∃ n ∈ P (i.e. ∃ {xn }) such that

|xn − x| < δ

but
|f (xn ) − f (x)| ≥ 

But this violates the continuity of f .

if
2. (⇐=) Suppose f is not continuous at x. Then ∃  > 0 such that ∀ δ >
0, ∃ x0 ∈ (a, b) satisfying |x0 − x| < δ and |f (x0 ) − f (x)| > .
We can construct {xn } by letting xn be this x0 for δ = n1 .
Then we get
xn → x but f (xn ) 6→ f (x)

The next result states has two important consequences.

Theorem 21. The continuous image of a closed and bounded interval is a closed
and bounded interval. That is, if

f : [a, b] −→ R

is continuous, then there exists d ≥ c such that f ([a, b]) = [c, d].
The assumptions in the theorem are important. If f is not continuous, then
there generally nothing that can be said about the image. If the domain is an open
interval, the image could be a closed interval (it is a point of f is constant) or it
could be unbounded even if the interval is finite (for example if f (x) = 1/x on
(0, 1)).
The first consequence of the result is the existence of a maximum (and mini-
mum).
Definition 42. We say that x∗ ∈ X maximizes the function f on X if

f (x∗ ) ≥ f (x),

for every x ∈ X .
44 CHAPTER 3. FUNCTIONS AND LIMITS OF FUNCTIONS

Definition 43. We say that x∗ ∈ X minimizes the function f on X if

f (x∗ ) ≤ f (x),

for every x ∈ X .

We write
max f (x) = max f (X )
x∈X

and
max f (X ) = y ∗
is the max value.

Since the image of f is a closed, bounded interval, f attains both its maximum
(the maximum value is d) and its minimum. This means that if a function is
continuous and it is defined on a “nice” domain, then it has a maximum. This
result is much more general than the result above. It applies to all real-valued
continuous functions (defined on arbitrary sets, not just the real numbers) provided
that the domain is “compact.” Bounded closed intervals contained in the real line
are examples of compact sets. More generally, any set that is “closed” (contains
its boundary points) and bounded is compact.

Theorem 22. If the function

f : X −→ R
is continuous and X is compact, then f (X ) is compact.
Proof. Take any {yn } ⊂ f (X ), we must show that ∃ a subsequence that converges
to a point in f (X ). Construct the sequence {xn } ⊂ X such that yn = f (xn ) ∀ n.
Since X is compact, {xn } has a convergent subsequence {xnk }; that is, xnk −→
x ∈ X . By continuity of f and our previous theorem, we have that ynk −→ f (x)
as well.

FIGURES GO HERE

Theorem 23. If the function

f : X −→ R
is continuous and X is compact, then max f (X ) exists.
 45

Proof. Note that f (X ) is bounded because f (X ) is compact due to X being com-

pact and using Theorem 22.

Thus using this result and Lemma 1, we know that sup f (X ) exists. So we
can find a sequence {yn } ⊂ f (X ) such that

yn −→ sup f (X ).

Since f (X ) is compact, it is thus closed by Theorem 16, so it therefore contains

its limit points, and so

lim yn = sup f (X ) ∈ f (X )
n→∞

And now by Theorem 2 when the max exists it equals the sup.

So finally we have that max f (X ) exists and

lim yn = sup f (X )
n→∞
= max f (X )

FIGURE GOES HERE

The second consequence of Theorem 21 is the Intermediate Value Theorem.

Theorem 24. If f : [a, b] −→ R is continuous, and z is between f (a) and f (b),

then there exists c ∈ [a, b] such that f (c) = [a, b]

In the statement of the theorem, “between” means either z ∈ [f (a), f (b)]

(when f (a) ≤ f (b) or z ∈ [f (b), f (a)] if f (b) < f (a).

Proof. From Theorem 21 it follows that the image of f is an interval. Hence if

two points are in the image, then all points between these two points are in the
image.

Theorem 24 is a method of solving showing that equations have solutions. It is

common to conclude that there must be a c for which the continuous function f is
zero (f (c) = 0) from the observation that sometimes f is positive and sometimes
f is negative. The most important existence theorems in micro (the existence of a
46 CHAPTER 3. FUNCTIONS AND LIMITS OF FUNCTIONS

market clearing price system, for example) follow from (harder to prove) versions
of this result.
Economists often prove existence results using fixed-point theorems. The eas-
iest fixed point theorem is a consequence of the Intermediate Value Theorem.
The point x∗ ∈ S is called a fixed point of the function f : S −→ S if
f (x∗ ) = x∗ .

Theorem 25. Let S = [a, b] be a closed bounded interval and f : S −→ S a

continuous function. There exists x∗ ∈ S such that f (x∗ ) = x∗ .

Proof. Consider the function h(x) = f (x) − x. Since f (a) ≥ a, h(a) ≥ 0.

Since f (b) ≤ b, h(b) ≤ 0. Since h is a continuous function on a closed bounded
interval, there must be an x∗ such that h(x∗ ) = 0. It is clear that for this value,
f (x∗ ) = x∗ .
This theorem generalizes: One can show that if S satisfies some technical
properties (convexity and compactness) then a continuous function from S to itself
must have a fixed point. This result does not require that S is one-dimensional
and it is instrumental in many existence proofs. The more general fixed-point
theorems are easy to state and understand, but are hard to prove.
It is not hard to give examples of functions that fail to have fixed points (if the
domain is not right). A simple one is: f (x) = x + 1.
Chapter 4

Differentiation

Calculus works because of two insights. The first is that linear functions are easy
to understand. The second insight is that although not all interesting functions
are linear, there is a large set of functions that can be approximated by a linear
function. The derivative is the best linear approximation to a function. You obtain
a lot of analytic power by taking a general function and studying it by learning
about linear approximations to the function.
How do you approximate a function? The first step is to treat the approxima-
tion as local. Given a point in the domain of the function, a, the idea is to come up
with a function that is easy to deal with and is close to the given function when x
is close to a. A possible attempt to do this is with a zero-th order approximation.
Definition 44. The zero-th order approximation of the function f at a point a in
the domain of f is the function A0 (x) ≡ f (a).
The symbol ≡ indicates an identity. One could also have written A0 (x) = a
for all x.
One way to approximate a function is with the constant function that is equal to
the value of the function at a point. A0 (x) is certainly a tractable function and if f
is continuous at a it is the only function that satisfies limx→a (f (x) − A0 (x)) = 0.
That is the good news. The bad news is that A0 tells you almost nothing about the
behavior of f .
The next step is to try to replace A0 with the best linear approximation to f . In
order to do this, imagine a line that intersects the graph of f at the points (x, f (x))
and the point (x + δ, f (x + δ)). This line has slope given by:
f (x + δ) − f (x) f (x + δ) − f (x)
=
(x + δ) − x δ

47
48 CHAPTER 4. DIFFERENTIATION

FIGURE GOES HERE

When f is linear, the line with this slope is f (just like when f is constant,
A0 ≡ f ). Otherwise, it is not. If we are interesting only in the local behavior of
f , then it makes sense to consider slopes defined when δ is small. The notion of
limit tells us how to do this:
f (x + δ) − f (x) f (x + δ) − f (x)
lim = lim
δ→0 (x + δ) − x δ→0 δ

The denominator of the expression does not make sense when δ = 0, but we do
not need to know this value to evaluate the limit of the ratio as δ approaches zero.
On the other hand, for the limit to make sense the ratio must be defined for all
sufficiently small non-zero δ – that is x must be an interior point of the domain
of f . If the limit exists, we say that f is differentiable at f and call the limit the
derivative of f at x.
As with the two earlier definitions (limit of function and continuity), there are
two “one-sided” versions of this definition.

Take
f : (a, b) −→ R
and x ∈ (a, b).

Define if they exist

f (y) − f (x)
f+0 ≡ lim+
y→x y−x
f (x + δ) − f (x)
( lim+ )
δ→0 δ
and
f (y) − f (x)
f−0 ≡ lim−
y→x y−x
f (x + δ) − f (x)
( lim− )
δ→0 δ

Definition 45. The derivative of a function f is defined at a point x when the left
hand derivative equals the right hand derivative. There are many ways to denote
 49

the derivative of the function f .

f 0 (x) = Df
df
=
dx
f (y) − f (x)
≡ lim
y→x y−x

Example 16. f (x) = x2

f (y) − f (x)
f 0 (x) = lim
y→x y−x
y − x2
2
= lim
y→x y − x

(y − x)(y + x)
= lim
y→x y−x
= lim (y + x)
y→x

= 2x

FIGURE GOES HERE

Example 17. Consider the function


x, if x ≥ 0,
f (x) =
−x, if x < 0.

So we get from this that

f+0 (0) = 1 right hand derivative

f−0 (0) = −1 left hand derivative

So obviously no derivative exists since

f+0 (0) 6= f−0 (0)

50 CHAPTER 4. DIFFERENTIATION

Now we can see how the derivative creates a first-order approximation. As-
sume the function f is defined on an open interval containing x. Let A1 (y) =
f (x) + f 0 (x)(y − x). This is the equation of the line with slope f 0 (x) that passes
through (x, f (x)). It follows from the definition of the derivative that

f (y) − A1 (y)
limy→x = 0. (4.1)
y−x
Equation (4.1) explains why A1 is a better approximation to f than A0 . Not only is
the linear approximation A1 close to f when y is close to x – this would be true if
the limit of the numerator in (4.1) converged to zero as y converged to x. It is also
the case that the linear approximation is close to f if you derive the difference by
something really close to zero (y − x). From this interpretation and the examples,
it is not surprising that it is harder to be differentiable than it is to be continuous.
Theorem 26. Consider the function

f : X −→ R

where X is an open interval. If f is differentiable at a point x (i.e. f 0 (x) exists),

then f is continuous at x.
Proof.
f (y) − f (x)
lim f (y) − f (x) = lim (y − x)
y→x y→x y−x
= f 0 (x) · 0
=0

where the second last equality is got from the fact that if you have 2 functions g
and h and limy→x h(y) exists and limy→x g(y) exists, then
  
lim g(y)h(y) = lim g(y) lim h(y)
y→x y→x y→x

which is the corresponding version of Theorem 9 (3) for functions.

As with continuity, it is tedious to use the definition to confirm that functions

are differentiable. Instead, you note that certain functions are differentiable (con-
stant functions and linear functions) and then apply a higher-level theorem. The
next result should look familiar (compare the related statement about continuity).
 51

Theorem 27. Suppose f and g are defined on an open interval containing point
x and both functions are differentiable at x. Then, (f + g), f · g, and f /g are
differentiable at x (the last of these provided g 0 (x) 6= 0).1

1.
(f + g)0 (x) = f 0 (x) + g 0 (x)

2.
(f · g)0 (x) = f 0 (x)g(x) + f (x)g 0 (x)

3.
f g(x)f 0 (x) − f (x)g 0 (x)
( )0 (x) =
g [g(x)]2

Proof. 1. This should be clear by Theorem 9 (1), though again we are using
the “functional version”.

2. Let h = f g. Then

h(y) − h(x) = f (y)[g(y) − g(x)] + g(x)[f (y) − f (x)]

If we divide this by y − x and note that f (y) −→ f (x), and g(y) −→ g(x)
as y −→ x by Theorem 26, then the result follows.
Then the result follows.

3. Now let h = f /g. Then

 
h(y) − h(x) 1 f (y) − f (x) g(y) − g(x)
= g(x) − f (x)
y−x g(y)g(x) y−x y−x
Now let y −→ x and you are done.

Exercise 1. A good exercise at this stage would be to prove that if

f (x) = xk

⇒ f 0 (x) = kxk−1
Hint Prove it by induction and using property (2) of Theorem 27. We already have
it for the case where k = 2 from Example 16
1
f · g denotes multiplication of functions. Composition, f ◦ g, is different.
52 CHAPTER 4. DIFFERENTIATION

Theorem 28 (Chain Rule). Suppose that

g : X −→ Y

and
f : Y −→ R
that g is differentiable at x and that f is differentiable at y = g(x) ∈ Y.2
Then

(f ◦ g)0 (x) = f 0 (y)g 0 (x)

= f 0 (g(x))g 0 (x)

Proof. Define the function

(
f (g(z))−f (g(x))
g(z)−g(x)
if g(z) − g(x) 6= 0
h(z) = 0
f (g(x)) if g(z) − g(x) = 0

 
f (g(z)) − f (g(x)) g(z) − g(x)
lim = lim h(z) ·
z→x z−x z→x z−x

h i g(z) − g(x)

= lim h(z) lim
z→x z→x z−x

To complete the proof it suffices to show that limz→x h(z) = f 0 (g(x)). We

do this using the definition of the limits. Let  > 0. Because f 0 is differ-
entiable
  find δ1 > 0 such that if 0 < |y − g(x)| < δ1 , then
at g(x) we can
 f (y)−f (g(x)) 0
 y−g(x) − f (g(x)) < . By continuity of g at x, we can find δ > 0 such that


if 0 < |z − x| < 0, then |g(z) − g(x)| < δ1 . It follows

 that if 0 < |z − x| < δ
 f (g(z))−f (g(x)) 0
and g(z) 6= g(x), then  g(z)−g(x) − f (g(x)) < . This means that if 0 <


|z − x| < δ and g(z) 6= g(x), then |h(z) − f (g(x))| < . To complete the proof
just note that if g(z) = g(y) then h(z) = f 0 (g(x)).

2
In this theorem we are talking about f ◦ g.
 53

Definition 46. Recall Definition 42. This is technically the definition of a global
max. We say globally since it maximizes the function over the whole domain.
However, we say x∗ is a local maximizer of the function f if ∃ a segment (a, b)
such that
f (x∗ ) ≥ f (x), ∀ x ∈ (a, b)

Definition 47. We say x∗ is a local minimizer of the function f if ∃ a segment

(a, b) such that
f (x∗ ) ≤ f (x), ∀ x ∈ (a, b)
We are interested in solving optimization problems, which are problems of the
form:
max f (x) subject to x ∈ S.
A solution to such a problem is a point x∗ with two properties: x∗ ∈ S and
if x ∈ S, then f (x∗ ) ≥ f (x). Calculus provides a method for solving these
problems. Note that a maximum that occurs at the boundary of the domain is not

a local maximum, but that a maximum that occurs in the interior of the domain is
a local maximum.
Theorem 29. Suppose f is defined on [a, b]. If f has a local max at c ∈ (a, b) and
if f 0 (c) exists, then
f 0 (c) = 0
Proof. For |x − c| < δ, we have
f (x) − f (c) ≤ 0
Therefore, if x ∈ (c, c + δ), then
f (x) − f (c)
≥0 (4.2)
x−c
while for x ∈ (c − δ, c)
f (x) − f (c)
≤ 0. (4.3)
x−c
Since f is differentiable at c, both left and right derivatives exist and they are
equal. Inequality (4.2) states that the derivative from above must be nonpositive
(if it exists). Inequality (4.3) states that the derivative from below must be non-
negative (if it exists). Since the derivative exists, these observations imply that the
derivative must be both nonpositive and nonnegative. The only possibility is that
f 0 (c) = 0.
54 CHAPTER 4. DIFFERENTIATION

You can use essentially the same argument to conclude that if c is a local
minimum and f 0 (c) exists, then f 0 (c) = 0. The equation f 0 (c) = 0 is called a
first-order condition. The theorem states that satisfying a first-order condition is
a necessary condition for c to be a local maximum or minimum. This observation
may be the most important result in calculus for economics. Economists are in-
terested in optimizing functions. The result says that you can replace solving an
optimization problem (which seems complicated) with solving an equation (which
perhaps is easier to do). The approach is powerful and generalizes. It suffers from
several limitations. One limitation is that it does not distinguish local maxima
from local minima. This is a major limitation of the statement of the theorem. If
you examine the proof carefully, you will see that the calculus does distinguish
maxima from minima. If you attempt to carry out the proof when c is a local
minimum, the inequalities in (4.2) and (4.3) will be reversed. This means, loosely
speaking, for a local maximum f 0 (x) ≥ 0 for x < c and f 0 (x) ≥ 0 for x > c. That
means, it appears that f 0 is decreasing. While for a local minimum the derivative
is negative to the left of c and positive to the right.
A second limitation is that the theorem only applies to local extrema. It is
possible that the maximum occurs at the boundary or there are many local max-
ima. Calculus still has something to say about boundary optima: If f is defined on
[a, b] and a is a maximum, then f (a) ≥ f (x) for all x ∈ [a, b] so that the derivative
from above must be nonpositive. Analogous statements are available for minima
or for the right-hand endpoint. As for identifying which local maximum is a true
maximum, calculus typically does not help. You must compare the values of the
various candidates.
Finally, it is possible that f 0 (c) equals zero, but c is neither a local maximum
nor a local minimum. The standard example of this is f (x) = x3 at the point
x = 0.
In spite of the drawbacks, Theorem 29 gives a procedure for solving an op-
timization problem for a differentiable function on the interval [a, b]: Solve the
equation f 0 (c) = 0. Call the set of solutions Z. Evaluate f (x) for all x ∈
Z ∪ {a, b}. The highest value of f over this set is the maximum. The lowest
value is the minimum. This means that instead of checking the values of f over
the entire domain, you need only check over a much smaller set.
An implication of the algorithm is that if a differentiable function’s derivative
is never zero, then it only has maxima and minima at the boundaries of its domain.
The previous result tells you that there is something special about places where
the derivative is zero. It is also possible to interpret places where the derivative is
positive or negative.

Definition 48. The function f is increasing if x > y implies f (x) ≥ f (y). The
 55

function f is strictily increasing if x > y implies f (x) > f (y). The function is
increasing in a neighborhood of x if there exists δ > 0 such that if y ∈ (x −
δ, x + δ) then x > y implies f (x) ≥ f (y). The function is strictly increasing in a
neighborhood of x if there exists δ > 0 such that if y ∈ (x − δ, x + δ) then x > y
implies f (x) > f (y).

There are analogous definitions for deceasing and strictly decreasing. A func-
tion that is either (strictly) increasing everyone or decreasing everywhere is called
(strictly) monotonic. There is a little bit of ambiguity about the term “increasing.”
Some people use “non-decreasing” to describe a function that we call increasing.

Theorem 30. If f is a real valued function differentiable on (a, b), then

1. If f 0 (x) > 0, then f is strictly increasing in the neighborhood of x.

2. If f 0 (x) > 0 for all x ∈ (a, b) then f is strictly increasing.

3. If f is increasing in the neighborhood of x, then f 0 (x) ≥ 0.

4. If f is increasing, then f 0 (x) ≥ 0 for all x ∈ (a, b).

The theorem almost says that differentiable functions are (strictly) increasing
if and only if the derivative is nonnegative (positive). Alas, you can have functions
that are strictly increasing but whose derivative is sometimes zero: f (x) = x3 is
again the standard example.
The proof of Theorem 30 is a straightforward exercise in the definition of the
derivative. It requires only writing down inequalities similar to (4.2) and (4.3) and
a little bit of care.
The previous two results give you excellent tools for graphing functions. You
can use derivatives to identify when the function is increasing, decreasing, or has
local maxima and minima. If you can figure out when the function crosses zero
and its behavior at infinity, then you have a nice insight into into behavior.

Theorem 31. If
g : R 7−→ R
is the inverse of
f : R 7−→ R
and if f is strictly increasing and differentiable with f 0 > 0, then
1
g 0 (f (x)) =
f 0 (x)
56 CHAPTER 4. DIFFERENTIATION

Note: f absolutely must be strictly increasing since otherwise it’s not invertible
and hence g is not well defined.
Theorem 32 (Mean Value Theorem). If f is real valued and continuous on [a, b],
and differentiable on (a, b), then ∃ a point c ∈ (a, b) such that

f (b) − f (a) = (b − a)f 0 (c)

FIGURE GOES HERE

We can see from the picture above what this is saying

f (b) − f (a)
= f 0 (c)
b−a
That is, there must be a point between a and b such that the derivative at that point
(i.e. slope of the tangent to the curve) is the same as that of the line connecting
our two points.
The mean value theorem is another way to think about how derivatives gen-
erate first-order approximations. Instead of saying that f (x) is approximately
f (x0 ) + f 0 (x0 )(x − x0 ) it says that f (x) is exactly f (x0 ) + f 0 (c)(x − x0 ) for
some value of c between x0 and x. So one approach gives you an approximation,
but to get the approximation you know exactly where to evaluate f 0 . The other
approach gives you an exact expression, but you do not know where to evaluate
the derivative.

Proof. Define  
f (b) − f (a)
g(x) = f (x) − (x − a)
b−a
We know that g(x) is continuous on compact [a, b]. Thus g(x) attains its maximum
and minimum on [a, b]. Note, however, that g(a) = g(b) = f (a). Consequently,
either g(a) ≥ g(x) for all x ∈ [a, b] so that g must attain its minimum on (a, b) or
else g attains its maximum on (a, b). We can conclude that g has a local minimum
or local maximum at some point c ∈ (a, b) and so g 0 (c) = 0. Thus
f (b) − f (a)
g 0 (c) = f 0 (c) −
b−a
= 0.
 57

The Mean Value Theorem has several applications.

Theorem 33. Suppose f is real valued and continuous on [a, b], and differentiable
on (a, b). If f 0 (x) ≡ 0 for x ∈ (a, b), then f is constant.

The result is intuitively obvious and perhaps something that you would im-
plicitly assume. It does require proof. (Notice that the converse is true too: If f is
constant, then it is differentiable and its derivative is always zero.)
Proof. By the Mean Value Theorem, for all x ∈ [a, b],

f (x) − f (a) = f 0 (c)(x − a)

for some c ∈ [a, x]. Since the right-hand side is zero by assumption, we have
f (x) = f (a) for all x, which is the desired result.
This is your first differential equation! That is, you found all of the functions
that satisfy an equation involving a derivative.
Here is another application.

Theorem 34. Suppose that f is a continuous real-valued function defined on

the interval (−1, 1). Further suppose that f 0 (x) exists for all x 6= 0 and that
limx→0 f 0 (x) exists. f is differentiable at x = 0 and f 0 is continuous at x = 0.

Proof. By the Mean Value Theorem,

f (x) − f (0)
= f 0 (c)
x
for c between 0 and x. Since f 0 is continuous at 0, limx→0+ f 0 (x) = limx→0− f 0 (x).
It follows that

f (x) − f (0) f (x) − f (0)

lim+ = lim+ f 0 (c) = lim− f 0 (c) = lim− ,
x→0 x c→0 c→0 x→0 x
which establishes that f 0 (0) exists and is equal to limc→0 f 0 (c).

Theorem 35. (L’Hopital’s Rule) Consider functions f and g differentiable on

[a, b), where g 0 (x) 6= 0 on [a, b).
If either
58 CHAPTER 4. DIFFERENTIATION

1. limx→b− f (x) = 0, and limx→b− g(x) = 0

OR

2. limx→b− f (x) = ∞, and limx→b− g(x) = ∞

and further
f 0 (x)
lim− =L∈R
x→b g 0 (x)
Then
f (x)
lim− =L
x→b g(x)
L’Hopital’s Rule is a useful way to evaluate indeterminate forms (0/0). It
looks a bit magical: How can the ratio of the functions be equal to the ratio of
derivatives? You prove that the rule works by using a variation of the mean value
theorem. In Case 1, what is going on is that f (x) = f (b) + f 0 (c)(x − b) for some
c ∈ (x, b) and similarly g(x) = g(b) + g 0 (d)(x − b). Since f (b) = g(b) = 0
(loosely), the ratio of f to g is the ratio of derivatives of f and g. The trouble is
that these derivatives are evaluated at different points. The good news is that you
can prove a version of the Mean-Value Theorem that allows you to take c = d.
This enables you to prove the theorem.
Warning: The Rule needs the two conditions to hold. If you try to evaluate
x+1
limx→0
x2 + 1
by setting the ratio equal to 1/(2x) and taking the limit you would be making a
big mistake. (The ratio is a rational function and the denominator is positive, so it
is continuous. Therefore the limit is just the function’s value at 0: 1.)
Chapter 5

Taylor’s Theorem

Using the first derivative, we were able to come up with a way to find the best
linear approximation to a function. It is natural to ask whether it is possible to
find higher order approximations. What does this mean? By analogy with zeroth
and first order approximations, we first decide what an appropriate approximating
function is and then what the appropriate definition of approximation is.
First-order approximations were affine functions. In general, an nth-order
approximation is a polynomial of degree n, that is a function of the form

a0 + a1 x + · · · + an−1 xn−1 + an xn .
Technically, the degree of a polynomial is the largest power of n that appears
with non-zero coefficient. So this polynomial has degree n if and only if an 6= 0.
Plainly a zeroth degree polynomial is a constant, a first degree polynomial is
an affine function, a second degree polynomial is a quadratic, and so on. An nth
order approximation of the function f at x is a polynomial of degree at most n,
An that satisfies

f (y) − An (y)
lim = 0.
y→x (y − x)n
This definition generalizes the earlier definition. Notice that the denominator is
a power of y − x. When y approaches x the denominator is really small. If the
ratio has limit zero it must be that the numerator is really, really small. We know
that zeroth order approximations exist for continuous functions and first-order
approximations exist for differentiable functions. It is natural to guess that higher-
order approximations exist under stricter assumptions. This guess is correct.
Definition 49. The nth derivative of a function f , denoted f n , is defined induc-
tively to be the derivative of f (n−1) .

59
60 CHAPTER 5. TAYLOR’S THEOREM

We say f is of class C n on (a, b) (f ∈ C n ) if f (n) (x) exists and is continuous

∀ x.

One can check that, just as differentiability of f implies continuity of f , if f n

exists, then f n−1 is continuous.

Theorem 36 (Taylor’s Theorem). Let f ∈ C n and assume that f n exists on (a, b)

and let c and d be any points in (a, b). Then there exists a point t between c and d
and a polynomial An of degree at most n such that

f (n+1) (t)
f (d) = An (d) + (d − c)n+1 , (5.1)
(n + 1)!
where An is the Taylor Polynomial for f centered at c:
n
X f (k) (c)
An (d) = (d − c)k .
j=0
k!

f (n+1) (t)
The theorem decomposes f into a polynomial and an error term En = (n+1)!
(d−
n+1
c) . Notice that
En
lim n = 0
d→c (d − c)

so the theorem states that the Taylor polynomial is, in fact, the nth order approxi-
mation of f at c.
The form of the Taylor approximation may seem mysterious at first, but the
coefficients can be seen to be the only choices with the property that f (k) (c) =
(k)
An (c) for k ≤ n. As impressive as the theorem appears, it is just a disguised
version of the mean-value theorem.

FIGURE GOES HERE

Proof. Define
n
X f (k) (x)
F (x) ≡ f (d) − (d − x)k
k=0
k!
and  n+1
d−x
G(x) ≡ F (x) − F (c).
d−c
 61

(n+1)
It follows that F (d) = 0 and (lots of terms cancel) F 0 (x) = − f n! (x) (d − x)n .
Also, G(c) = G(d) = 0. It follows from the mean value theorem that there exists
a t between c and d such that G0 (t) = 0. That is, there exists a t such that
n
f (n+1) (x)

n d−t
0=− (d − x) + F (c)
n! d−c
or
f (n+1) (t)
F (c) = (d − c)n+1 .
(n + 1)!
An examination of the definition of F confirms that this completes the proof.

Taylor’s Theorem has several uses. As a conceptual tool it makes precise the
notion that well behaved functions have polynomial approximations. This per-
mits you to understand “complicated” functions like the logarithm or exponential
by using their Taylor’s expansion. As a computational tool, it permits you to
compute approximate values of functions. Of course, doing this is not practical
(because calculators and computers are available). As a practical tool, first- and
second-order approximations permit you to conduct analyses in terms of linear or
quadratic approximations. This insight is especially important for solving opti-
mization problems, as we will see soon.
Next we provide examples of the first two uses.
Consider the logarithm function: f (x) = log x.1 This function is defined for
x > 0. It is not hard to show that f (k) (x) = x−k (−1)k−1 (k − 1)!. So f (k) (1) =
(−1)k−1 (k − 1)! Hence:
N
X (x − 1)k
f (x) = (−1)k−1 + EN
k=1
k

N
where EN = (−1)N (y−1) N +1
for y between 1 and x. Notice that this expansion is
done around x0 = 1. This is a point at which the function is nicely behaved. Next
notice that the function f (k) is differentiable at 1 for all k. This suggests that you
can extend the polynomial for an infinite number of terms. It is the case that
∞
X (x − 1)k
log(x) = (−1)k−1 .
k=1
k
Sometimes this formula is written in the equivalent form:
1
Unless otherwise mentioned, logarithms are always with respect to the base e.
62 CHAPTER 5. TAYLOR’S THEOREM

∞
X yk
log(y + 1) = (−1)k−1 .
k=1
k
The second way to use Taylor’s Theorem is to find approximations. The for-
mula above can let you compute logarithms. How about square roots? The Tay-
lor’s expansion of the square root of x around 1 is:
√
2 = 1 + .5 − .125 + E2 .

E2 = x−2.5 /16 for some x ∈ [1, 2]. Check to √ make sure you know where the
terms come from. The approximation says that 2 = 1.375 up to an error. The
error term is largest when x = 1. Hence the error is no more than .0625. The error
term is smallest when x = 2. I’m not sure what the error is then, but it is certainly
positive. Hence I know that the square root of 2 is at least 1.375 and no greater
than 1.4275. Perhaps this technique will come in handy the next time you need to
compute a square root without the aid of modern electronics.
Chapter 6

Univariate Optimization

Economics is all about optimization subject to constraints. Consumers do it. They

maximize utility subject to a budget constraint. Firms do it. They maximize
profits (or minimize costs) subject to technological constraints. You will study
optimization problems in many forms. The one-variable theory is particularly
easy, but it teaches many lessons that extend to general settings.
There are two features of an optimization problem: the objective function,
which is the real-valued function1 that you are trying to maximize or minimize,
and the constraint set.
You already know a lot about one-variable optimization problems. You know
that continuous functions defined on closed and bounded intervals attain their
maximum and minimum values. You know that local optima of differentiable
functions satisfy the first-order condition.
The next step is to distinguish between local maxima and local minima.
Suppose that f differentiable on an open interval. The point x∗ is a critical
point if f 0 (x∗ ) = 0. We know that local minima and local maxima must be critical
points. We know from examples that critical points need not be local optima. It
turns out that properties of the second derivative of f classify critical points.

Theorem 37 (Second-Order Conditions). Let f be twice continuously differen-

tiable on an open interval (a, b) and let x∗ ∈ (a, b) be a critical point of f . Then
00
1. If x∗ is a local maximum, then f (x∗ ) ≤ 0.
00
2. If x∗ is a local minimum, then f (x∗ ) ≥ 0.
1
Although the domain of the function you are trying to optimize varies in applications, the
range is typically the real numbers. In general, you need the range to have an ordering that enables
you to compare any pair of points.

63
64 CHAPTER 6. UNIVARIATE OPTIMIZATION

00
3. If f (x∗ ) < 0, then x∗ is a local maximum.
00
4. If f (x∗ ) > 0, then x∗ is a local minimum.
Conditions (1) and (3) are almost converses (as are (2) and (4)), but not quite.
00
Knowing that x∗ is a local maximum is enough to guarantee that f (x∗ ) ≤ 0.
00
Knowing that f x∗ ) is not enough to guarantee that you have a local minimum
(you may have a local maximum or you may have neither a minimum nor a max-
imum). (All the intuition you need comes from thinking about the behavior of
f (x) = xn at x = 0 for different values of x = 0.) You might think that you
could improve the statements by trying to characterize strict local maxima. It is
00
true that if f (x∗ ) < 0, then x∗ is a strict local maximum, but it is possible to have
00
a strict local maximum and f (x∗ ) = 0. The conditions in Theorem 37 parallel
the results about first derivatives and monotonicity stated earlier.
Proof. By Taylor’s Theorem we can write:
1 00
f (x) = f (x∗ ) + f 0 (x∗ )(x − x∗ ) + f (t)(x − x∗ )2 (6.1)
2
00 00 00
for t between x and x∗ . If f (x∗ ) > 0, then by continuity of f , f (t) > 0 for t
sufficiently close to x∗ and so, by (6.1), f (x) > f (x∗ ) for all x sufficiently close
00
to x∗ . Consequently, if x∗ is a local maximum, f (x∗ ) ≤ 0, proving (1). (2) is
similar.
00 00
If f (x∗ ) < 0, then by continuity of f , there exists δ > 0 such that if 0 <
00
|x − t| < δ, then f (t) < 0. By (6.1), it follows that if 0 < |x − x∗ | < δ, then
f (x) < f (x∗ ), which establishes (3). (4) is similar.
The theorem allows us to refine our method for looking for maxima. If f is
defined on an interval (and is twice continuously differentiable), the maximum (if
it exists) must occur either at a boundary point or at a critical point x∗ that satisfies
f (x∗ ) ≤ 0. So you can search for maxima by evaluating f at the boundaries and
at the appropriate critical points.
This method still does not permit you to say when a local maximum is really
a global maximum. You can do this only if f satisfies the appropriate global
conditions.
Definition 50. We say a function f is concave over an interval X ⊂ R if ∀x, y ∈
X and δ ∈ (0, 1), we have
f (δx + (1 − δ)y) ≥ δf (x) + (1 − δ)f (y) (6.2)
If f is only a function of one argument you can think of this graph as having an
inverted “u” shape.
 65

Geometrically the definition says that the graph of the function always lies
above segments connecting two points on the graph. Another way to say this is
that the graph of the function always lies below its tangents (when the tangents
exist). If the inequality in (6.2) is strict, then we say that the function is strictly
concave. A linear function is concave, but not strictly concave.
Concave functions have nicely behaved sets of local maximizers. It is an im-
mediate consequence of the definition that if x and y are local maxima, then so
are all of the points on the line segment connecting x to y. A fancy way to get at
this result is to note that concavity implies
f (δx + (1 − δ)y) ≥ δf (x) + (1 − δ)f (y) ≥ min{f (x), f (y)}. (6.3)
Moreover, the inequality in (6.3) is strict if δ ∈ (0, 1) and either (a) f (x) 6= f (y)
or (b) f is strictly concave. Suppose that x is a local maximum of f . It follows
that f (x) ≥ f (y) for all y. Otherwise f (λx + (1 − λ)y) > f (x), for all λ ∈ (0, 1),
contradicting the hypothesis that x is a local maximum. This means that any local
maximum of f must be a global maximum. It follows that if x and y are both local
maxima, then they both must be global maximal and so f (x) = f (y). In this case
it follows from (6.3) that all of the points on the segment connecting x and y must
also be maxima. It further implies that x = y if f is strictly concave.
These results are useful. They guarantee that local extrema are global maxima
(so you know that a critical point must be a maximum without worrying about
boundary points or local minima) and they provide a tractable sufficient condition
for uniqueness. Notice that these nice properties follow from (6.3), which is a
weaker condition than (6.2).2 This suggests that the following definition might be
useful.
Definition 51. We say a function f is quasi concave over an interval X ⊂ R if
∀x, y ∈ X and δ ∈ (0, 1), we have f (δx + (1 − δ)y) ≥ min{f (x), f (y)}.
If quasi-concavity is so great, why bother with concavity? It turns out that
concavity has a characterization in terms of second derivatives.
We can repeat the same analysis with signs reversed.
Definition 52. We say a function f is convex over an interval X ⊂ R if ∀x, y ∈ X
and δ ∈ (0, 1), we have
f (δx + (1 − δ)y) ≤ δf (x) + (1 − δ)f (y)
If f is only a function of one argument you can think of this graph as having an
“u” shape.

2
As an exercise, try to find a function that satisfies (6.3) but not (6.2).
66 CHAPTER 6. UNIVARIATE OPTIMIZATION

Convex functions have interior minima and differentiable convex functions

have positive second derivatives.

Theorem 38. Let f : X −→ R, X an open interval and f ∈ C 2 on X .

f 00 (x) ≤ 0, ∀x ∈ X , if and only if f is concave on X .

Proof. If f is concave, then for all λ ∈ (0, 1)

f (λx + (1 − λ)y) − f (x) f (y) − f (λx + (1 − λ)y)

≥ .
1−λ λ
Routine manipulation demonstrates that the limit of the left-hand side (if it exists)
as λ → 1 is equal to (y − x)f 0 (x) (note that f (λx + (1 − λ)y) − f (x) = f (x +
(1 − λ)(y − x)) − f (x)) and the limit of the right-hand side λ → 1 is equal to
(y − x)f 0 (y). It follows that if f is concave and differentiable, then

(y − x)f 0 (x) ≥ (y − x)f 0 (y),

00
which in turn implies that f 0 is decreasing. Consequently, if f exists, then it is
non-positive.
Conversely, if f is differentiable, then by the Mean Value Theorem

f (λx + (1 − λ)y) − f (x) = (1 − λ)f 0 (c)(y − x)

for some c between x and λx + (1 − λ)y. You can check that this means that if f 0
is decreasing, then

f (λx + (1 − λ)y) − f (x) ≥ (1 − λ)f 0 (λx + (1 − λ)y)(y − x). (6.4)

Similarly,
f (λx + (1 − λ)y) − f (y) = −λf 0 (c)(y − x)
for some c between λx + (1 − λ)y and y and if f 0 is decreasing, then

f (λx + (1 − λ)y) − f (y) ≥ −λf 0 (λx + (1 − λ)y)(y − x). (6.5)

The result follows from adding λ times inequality (6.4) to 1 − λ times inequal-
ity (6.5).

Note
00
• The previous theorem used the fact that f 0 was decreasing (rather than f ≤
0).
 67

• The Mean-Value Theorem says that if f is differentiable, then f (y)−f (x) =

f 0 (c)(y − x) for some c between x and y. You can check that this means
that if f 0 is decreasing, then f (y) − f (x) ≤ f 0 (x)(y − x). This inequality is
the algebraic way of expressing the fact that the tangent line to the graph of
f at (x, f (x)) lies above the graph.
00
• If f < 0 so that f is strictly concave, then f 0 is strictly decreasing. That
00
means that f can have at most one critical point. Since f < 0, this critical
point must be a local maximum. These statements simply reiterate earlier
observations: Local maxima of strictly concave functions are global max-
ima and these must be global maxima.
68 CHAPTER 6. UNIVARIATE OPTIMIZATION
Chapter 7

Integration

FIGURES GO HERE

7.1 Introduction
Integration is a technique that does three superficially different things. First, it acts
an the inverse to differentiation. That is, if you have the derivative of a function
and want to know the function, then the process of “anti-differentiation” is closely
related to the integration. Finding antiderivatives is essential if you want to solve
differential equations.
Second, integration is a way to take averages of general functions. This in-
terpretation is the most natural and insightful one for the study of integration in
probability and statistics.
Third, for non-negative functions, the integral is a way to compute areas.
The connection between the second and third interpretations is fairly straight-
forward. The connection between the second and first interpretation is important
and surprising and is a consequence of “The Fundamental Theorems of Calculus.”
We will motivate the definition of integral as a generalized average. If you are
given a set of N numbers, you average them by adding the numbers up and then
dividing by N . How do you generalize this to a situation in which you are given
an infinite set of numbers to average?
Assume that f : [a, b] −→ R. A naive way to guess the average value of f
would be to pick some x ∈ [a, b] and say that the average is equal to f (x). This
would work great if f happened to be constant. If you could find the smallest and
largest values of f on the interval (this would be possible if f were continuous),

69
70 CHAPTER 7. INTEGRATION

then you could use these are lower and upper bounds for the average. If f did not
vary too much, perhaps you could use these values to get a good estimate of the
value of f ’s average. You could get an every better estimate if you subdivided the
interval [a, b] and repeated the process: The lower bound for the average value of f
would be the average of the minimum value of f on [a, (a+b)/2] and [(a+b)/2, b].
When you do this, the estimate for the lower bound will be higher (at least it
won’t be lower) and the estimate for the upper bound will be no higher than the
original estimate. Maybe if you keep repeating the process the upper and lower
bounds converge to something that would be a good candidate for the average.
The theory of integration is primarily about finding conditions under which this
kind of argument works.

A partition P of [a, b] is a finite set of points x0 , x1 , . . . , xn , where

a = x0 ≤ x1 ≤ · · · ≤ xn−1 ≤ xn ≤ b

Given such a partition, define

n
X
Lf (P ) = mk ∆k
k=1

where

mk ≡ inf f (x)
x∈[xk−1 ,xk ]

= inf {f (x) | x ∈ [xk−1 , xk ]}

and

∆k ≡ xk − xk−1
n
X
Uf (P ) = Mk ∆k
k=1

where

Mk ≡ supx∈[xk−1 ,xk ] f (x)

= sup {f (x) | x ∈ [xk−1 , xk ]}

Since these definitions are defined in terms of “sup” and “inf” they are well
defined even if f is not continuous. (If f is continuous, then the maxima and
7.1. INTRODUCTION 71

minima are attained on each subinterval, so you can replace sup by max and inf
by min in the definitions.) It is clear that for each partition P , Lf (P ) ≤ Uf (P ).
Also, Lf (P ) is an underestimate of the value that we want while U (P ) is an
overestimate.1
Now imagine subdividing the partition P by dividing each subset of P into
two non-empty pieces. This leads to a new partition P 0 and new values Lf (P 0 ) ≥
Lf (P ) and Uf (P 0 ) ≤ Uf (P ). The reason that subdividing increases the lower
sums is that

inf f (x) ≤ inf f (x) + inf f (x) for all z ∈ [xk−1 , xk ]

x∈[xk−1 ,xk ] x∈[xk−1 ,z] x∈[z,xk ]

because the value of x that makes f smallest in the expression on the left may be
in only one of the two subintervals.
So far we have a process that generates an increasing sequence of lower esti-
mates of the average value of f and a process that generates a decreasing sequence
of upper estimates of the average value of f . We know that bounded monotone
sequences converge. This motivates the following definition.

Definition 53. Let f : [a, b] −→ R and suppose that Pr is a sequence of partitions

a = x0 (r) ≤ x1 (r) ≤ · · · ≤ xn−1 (r) ≤ xnr (r) = b

such that ∆k (r) = xk (r) − xk−1 (r) goes to zero as r approaches infinity for all k.
f is integrable if

lim Lf (Pr ) exists and is equal to lim Uf (Pr ). (7.1)

r→∞ r→∞

Rb
If f is integrable then we denote the common limit in (7.1) by a
f (x)dx.

In the definition, there are lots of ways to take “finer and finer” partitions. It
turns out that if the upper and lower limits exist and are equal, then it does not
matter which partition you take. They will all converge to the same limit as the
length of each element in the partition converges to zero. It also turns out that if a
function is integrable, then it does not matter whether you evaluate f inside each
partition element using the sup, the inf or any value in between. All choices will
lead to the same value.
1
The formulas for U and L are the standard ones, but they are not quite right for the “average”
interpretation. If f (x) ≡ c, then we would have Uf (P ) = Lf (P ) = c(b − a) – the length of the
interval appears in the formula. In order to maintain the interpretation as average, you must divide
the formulas by the length of the interval, b − a.
72 CHAPTER 7. INTEGRATION

If a function is integrable, then the integral (divided by b − a) is a perfect

generalization of average value. The next question is: Which functions are inte-
grable? It is not hard to come up with examples of non-integrable functions: The
function that is 1 on the rationals and 0 on the irrationals has the property that the
lower sums is equal to 0 and the upper sum is equal to b − a. One can show that
continuous functions must be integrable (because you can make the difference be-
tween Uf (P ) and Lf (P ) arbitrarily small by making the size of the pieces of the
partition small). It is useful in some situations to look at families of functions that
are discontinuous and still integrable. For example, monotonic functions may be
discontinuous, but bounded monotonic functions are always integrable.
The definition above is the definition of the Riemann Integral. You may hear
about other kinds of integral. A Riemann-Stieltjes Integral is a generalization of
the Riemann Integral in which the “length” of the interval is not uniform. Mea-
sure theory is a mathematical area that allows you to define integrals of functions
defined on abstract sets equipped with a “measure” that gives you the “size” of
various sets. One example of a measure is a probability distribution. The integrals
used in measure theory are called expectations by probability theorists. Lebesgue
Measure is the most common measure. The Lebesgue measure defines the size
of the interval [a, b] to be b − a and, more generally, the size of rectangles to
be their (standard) area. The basic theory is similar to the theory of Riemann
integration and you should not be intimidated when people talk about Lebesgue
integrals.2 One difference between the theory is that typically the class of inte-
grable functions is larger in measure theory. In particular, the function that is 0 on
rationals and 1 on irrationals fails to be Riemann integrable (the lower sum is 0
and the upper sum is 1). This function is Lebesgue integrable (the integral is one),
because measure theory takes into account that while there are a lot of rational
numbers (precisely they are a dense set – you can find a rational number in any
open interval) they are a small set (precisely, they have Lebesgue measure zero).
The theory of stochastic calculus, which is used in continuous-time finance,
uses a different kind of integral. For this theory it is important to compute integrals
of functions that are not integrable in the sense of Riemann. (That is, the class
of functions that includes all continuous things and some other objects just isn’t
big enough.) In this theory it does matter where you evaluate f in the definition
of the approximating sums. This leads to a theory that is different from classic
integration theory in several respects.

2
To define the Lebesgue integral you approximate a function by a sequence of “simple” func-
tions that take on only finitely many values and hence are easy to average.
7.2. FUNDAMENTAL THEOREMS OF CALCULUS 73

Theorem 39. If
f (x) = c, ∀x ∈ [a, b]
Then Z b
f (x)dx = c(b − a)
a

Definition 54. Take f : (a, b) −→ R.

Suppose ∃ F : (a, b) −→ R, continuous and differentiable on (a, b).
If
F 0 (x) = f (x),
∀x ∈ (a, b).
Then F is called the antiderivative of f .

Example 18.
f (x) = x2
1
=⇒ F (x) = x3
3
also could have
1
F (x) = x3 + 6
3

Theorem 40. If F and G are both antiderivatives of f on [a, b], then

G(x) = F (x) + c
Note what this is saying is that any 2 antiderivatives differ only by a constant.

7.2 Fundamental Theorems of Calculus

Theorem 41 (Fundamental Theorem of Calculus). If
f : X −→ R
is continuous and X is an open interval then, for a ∈ X and F defined by
Z x
F (x) ≡ f (t)dt, ∀ x ∈ X
a

it is the case that F is differentiable and

F 0 (x) = f (x), ∀ x ∈ X
74 CHAPTER 7. INTEGRATION

The theorem states that differentiation and integration are inverse operations
in the sense that if you start with a function and integrate it, then you get a dif-
ferentiable function and the derivative of that function is the function you started
with.
Proof. By the definition of the integral,
Z x+h
h sup f (y) ≥ F (x + h) − F (x) = f (x)dx ≥ h inf f (y)
y∈[x,x+h] x y∈[x,x+h]

the result follows from dividing through by h and taking limits.

Theorem 42. If F is an antiderivative of

f : [a, b] −→ R
then,
Z b
f (x)dx = F (b) − F (a)
a
= [F (x)]ba
Proof. By the mean value theorem F (y)−F (x) = f (t)(y−x) for some t between
x and y. It follows that for any partition P
Xn
Uf (P ) ≥ (F (xk ) − F (xk−1 ) ≥ Lf (P ) (7.2)
k=1

because for each element of the partition F (xk ) − F (xk−1 ) = f (tk )(xk − xk−1 )
and f (tk ) ∈ [inf x∈[xk−1 ,xk ] f (x), supx∈[xk−1 ,xk ] f (x)]. Since
n
X
(F (xk ) − F (xk−1 ) = F (b) − F (a)
k=1
the result follows by taking limits of finer and finer partitions in (7.2).
This theorem is a converse of the first result.
A bit of terminology: An antiderivative of f is a function whose derivative is
f . This is
R sometimes called a primitive of f or the indefinite integral of f and is
denoted f (where the limits of integration are not specified). When the limits of
Rb
integration are specified, a f is a number, called the definite integral of f on the
interval [a, b].
Up until now, we have only talked about integration over closed and bounded
intervals. It is sometimes convenient to talk about “improper” integrals in which
the limits of the integrals may be infinity. These integrals are defined to be limits
of integrals over bounded intervals (provided that the limits exist).
7.3. PROPERTIES OF INTEGRALS 75

7.3 Properties of Integrals

Theorem 43. 1.
Z b Z b Z b
[λf (x) + µg(x)]dx = λ f (x)dx + µ g(x)dx
a a a

where λ, µ ∈ R.

2. If f and g are continuous on [a, b] with antiderivatives F and G, then

Z b Z b
f (x)G(x) = [F (x)G(x)]ba − F (x)g(x)dx
a a

3. Suppose π is strictly monotonic on [a, b] and f is continuous on an open

interval containing π([a, b]). Then
Z π(b) Z b
f (x)dx = f (π(t))π 0 (t)dt
π(a) a

4. If f : [a, b] −→ R is continuous, then there exists t ∈ [a, b] such that

Z b
f (x)dx = f (t)(b − a).
a

Proof. 1. A direct consequence of the definitions.

2. There are 2 ways to see this

(a) We know from Theorem 27 (2) that

d
(F G) = f G + F g
dx
d
(F G) − F g
=⇒ f G =
dx
and we can just integrate to get the desired result
(b) Alternatively we can set H(x) = F (X)G(X) an apply Theorem 42.

3. Exercise
76 CHAPTER 7. INTEGRATION

4. This result is a simple consequence of the intermediate value theorem. By

the definition of the integral it cannot be that
Z b
f (x)dx > f (t)(b − a)
a
for all t ∈ [a, b] nor can it be that
Z b
f (x)dx < f (t)(b − a)
a
for all t ∈ [a, b]. Hence
Z b
f (x)dx − f (t)(b − a)
a

is a continuous function of t that changes sign on [a, b] so it must equal zero

for some value of t.

The second formula is called integration by parts and it comes up a lot. The
third formula is the change of variables formula. The fourth result is called the
mean value theorem for integrals and states that at some point in an interval a
function must take on the average value of the function on the interval.

7.4 Computing Integrals

Since every continuous function is integrable, we know a lot of functions that
are integrable. Coming up with formulas for the integrals is not easyR in practice.
Polynomials are easy. One can integrate the exponential function ( ex = ex ) and
the identity Z x
1
dt = log x
1 t
is often taken to be the definition of the log function. Other than that, the change
of variables formula and integration by parts are the primary methods we have to
find integrals.

1. Let F 0 = f .
Z b Z b
xf (x)dx = bF (b) − aF (a) − F (x)dx
a a
7.4. COMPUTING INTEGRALS 77

In applications, F is a CDF, f a density, F (b) = 1, F (a) = 0.

LHS is mean of distribution. So mean is
Z b
b− F (x)dx
a
.

2. Z ∞ Z ∞
n −x
x e dx = n xn−1 e−x dx = n!
0 0

(I simplified using limx→∞ xn e−x = 0).

This suggests a general formula for factorials (valid for n that is not an
integer).
78 CHAPTER 7. INTEGRATION
Chapter 8

Basic Linear Algebra

Many of the elements of one-variable differential calculus extend naturally

to higher dimensions. The definition of continuity is (stated appropriately)
identical. Maxima exist for real-valued continuous functions defined on the
appropriate domain. The derivative still plays the role of linear approxi-
mation. Critical points play the same role in optimization theory. Taylor’s
Theorem and Second-Order Conditions reappear. Before we return to cal-
culus, however, we need to record a few facts about domains that are rich
enough to permit many variables.

8.1 Preliminaries

Definition 55. n-dimensional Euclidean Space:

Rn = R × R × · · · × R × R

Note if X and Y are sets, then

X × Y ≡ {(x, y) | x ∈ X , y ∈ Y}

so
Rn = {(x1 , x2 , . . . , xn ) | xi ∈ R, ∀ i = 1, 2, . . . , n}

There are different ways to keep track of elements of Rn . When doing

calculus, it is standard to think of a point in Rn as a list of n real numbers,

79
80 CHAPTER 8. BASIC LINEAR ALGEBRA

and write x = (x1 , . . . , xn ). When doing linear algebra, it is common to

think of the elements of Rn as column vectors. When we do this we write
 
x1
 x2 
x =  ..  .
 
 . 
xn

Given x ∈ Rn , we understand xi to be the ith coordinate. In these notes, we

will try (and almost certainly fail) to denote vectors using bold face (x) and
elements more plainly (x). Although there may be exceptions, for the most
part you’ll see x in discussions of linear algebra and x in discussions about
calculus.

Definition 56. The zero element:

 
0
 0 
0=
 
.. 
 . 
0

Definition 57 (Vector Addition). For x, y ∈ Rn we have

 
x1 + y 1
 x2 + y 2 
x+y=
 
.. 
 . 
xn + y n
Vector addition is commutative
x+y=y+x

Definition 58 (Scalar Multiplication). For x ∈ Rn , and a ∈ R we have

 
ax1
 ax2 
ax =  .. 
 
 . 
axn
In other words every element of x gets multiplied by a.
8.2. MATRICES 81

You may hear people talk about vector spaces. Maybe they are showing off.
Maybe they really need a more general structure. In any event a vector space
is a general set of V in which the operation of addition and multiplication
by a scalar make sense, where addition is commutative and associative (as
above), there is a special zero vector that is the additive identity (0 + v = v),
additive inverses exist (for each v there is a −v, and where scalar multi-
plication is defined as above. Euclidean Spaces are the leading example of
vector spaces. We will need to talk about subsets of Euclidean Spaces that
have a linear structure (they contain 0, and if x and y are in the set, then so
is x + y and all scalar multiples of x and y). We will call these subspaces
(this is a correct use of the technical term), but we have no reason to talk
about more general kinds of vector spaces.

8.2 Matrices

Definition 59. An m × n matrix is an element of Mm×n written as in the

form  
α11 α12 · · · α1n
 α21 α22 · · · α2n 
A =  .. ..  = [αij ]
 
..
 . . . 
αm1 αm2 · · · αmn
where m denotes the number of rows and n denotes the number of columns.

Note An m × n matrix is just of a collection of nm numbers organized in

a particular way. Hence we can think of a matrix as an element of Rm×n .
The extra notation Mm×n makes it possible to distinguish the way that the
numbers are organized. Note We denote vectors in boldface lower-case let-
ters. Matrices are represented in capital boldface.

Note Vectors are just a special case of matrices. e.g.

 
x1
 x2 
x =  ..  ∈ Mn×1
 
 . 
xn
82 CHAPTER 8. BASIC LINEAR ALGEBRA

In particular real numbers are just another special case of matrices. e.g. the
number 6

6 ∈ R = R1×1

This notation emphasizes that we think of a vector with n components as a

matrix with n rows and 1 columns.

Example 19.
 
0 1 5
A =
2×3 6 0 2

Definition 60. The transpose of a matrix A, is denoted At . To get the trans-

pose of a matrix, we let the first row of the original matrix become the first
column of the new (transposed) matrix. Using Definition 59 we would get
 
α11 α21 ··· α1n
t
 α12 α22 ··· α2n 
A =  = [αji ]
 
.. .. ..
 . . . 
α1m α2m · · · αnm

Definition 61. A matrix A is symmetric if A = At .

So we can see that if A ∈ Mm×n , then At ∈ Mn×m .

Example 20. Using Example 19 we see that

 
0 6
At =  1 0 
3×2
5 2
8.2. MATRICES 83

8.2.1 Matrix Algebra

You have probably guessed then how to add 2 (m × n) matrices - term by
term!
Note as we saw beforehand with vectors, it is totally meaningless to add
matrices that are of different dimensions.

Definition 62 (Addition of Matrices). If

 
α11 α12 · · · α1n
 α21 α22 · · · α2n 
A =  ..  = [αij ]
 
m×n
.. ..
 . . . 
αm1 αm2 · · · αmn
and  
β11 β12 · · · β1n
 β21 β22 · · · β2n 
B =  = [βij ]
 
m×n
.. .. ..
 . . . 
βm1 βn2 · · · βmn
then
 
α11 + β11 α12 + β12 ··· α1n + β1n
 α21 + β21 α22 + β22 ··· α2n + β2n 
A+B= D =  = [δij ] = [αij +βij ]
 
m×n
.. .. ..
m×n  . . . 
αm1 + βm1 αm2 + βm2 · · · αmn + βmn

 
α11 + β11 α12 + β12 ··· α1n + β1n
 α21 + β21 α22 + β22 ··· α2n + β2n 
A + B =  = [αij + βij ]
 
| {z }  .. .. ..
m×n
 . . . 
αm1 + βm1 αm2 + βm2 · · · αmn + βmn

Definition 63 (Multiplication of Matrices). If A and B are given, then

m×k k×n
we define
A · B = C = [cij ]
m×k k×n m×n
such that
k
X
cij ≡ ail blj
l=1
so note above that the only index being summed over is l.
84 CHAPTER 8. BASIC LINEAR ALGEBRA

The above expression may look quite daunting if you have never seen sum-
mation signs before so a simple example should help to clarify.

Example 21. Let  

0 1 5
A =
2×3 6 0 2
and  
0 3
B = 1 0 
3×2
2 3
Then
 
  0 3
0 1 5
A · B = · 1 0 
2×3 3×2 6 0 2
| {z } 2 3
2×2
 
(0 × 0) + (1 × 1) + (5 × 2), (0 × 3) + (1 × 0) + (5 × 3)
=
(6 × 0) + (0 × 1) + (2 × 2), (6 × 3) + (0 × 0) + (2 × 3)
 
11 15
=
4 24

Note that A · B is meaningless. The second dimension of A must be equal

2×3 4×5
to the first dimension of B.

Note further that this brings up the very important point that matrices do
not multiply like regular numbers. They are NOT commutative i.e.
A · B 6= B · A
For example

A · B 6= B · A
2×3 3×4 3×4 2×3
in fact, not only does the LHS not equal the RHS, the RHS does not even
exist. We will see later that one interpretation of a matrix is as a represen-
tation of a linear function. With that interpretation, matrix multiplication
takes on a specific meaning and there will be another way to think about
why you can only multiply certain “conformable” pairs of matrices.
8.2. MATRICES 85

Definition 64. Any matrix which has the same number of rows as columns
is known as a square matrix, and is denoted A .
n×n

For example the matrix

 
0 1 5
A = 6 0 2 
3×3
3 2 1
is a square (3 × 3) matrix.

Definition 65. There is a special square matrix known as the identity matrix
(which is likened to the number 1 (the multiplicative identity) from Defini-
tion 14), in that any matrix multiplied by this identity matrix gives back the
original matrix. The Identity matrix is denoted In and is equal to
 
1 0 ... 0
 0 1 ... 0 
In ==  .. ..  .
 
. .
n×n  . . . 
0 ... 0 1
Definition 66. A square matrix is called a diagonal matrix if aij = 0 when-
ever i 6= j.1
Definition 67. A square matrix is called an upper triangular matrix (resp.
lower triangular if aij = 0 whenever i > j (resp. i < j).

Diagonal matrices are easy to deal with. Triangular matrixes are also some-
what tractable. You’ll see that for many applications you can replace an
arbitrary square matrix with a related diagonal matrix.
For any matrix A we have the results that
m×n

A · In = A
m×n m×n

and
Im · A = A
m×n m×n

Note that unlike normal algebra it is not the same matrix which multiplies
A on both sides to give back A (unless n = m).
m×n m×n

1
The main diagonal is always defined as the diagonal going from top left corner to bottom right
corner i.e. &
86 CHAPTER 8. BASIC LINEAR ALGEBRA

Definition 68. We say a matrix A is invertible or non-singular if ∃ B such

n×n n×n
that
A · B = B · A = In
n×n n×n n×n n×n
| {z } | {z }
n×n n×n

If A is invertible, we denote it’s inverse as A−1 .

So we get
A · A−1 = A−1 · A = In
(n×n) (n×n) n×n n×n
| {z } | {z }
n×n n×n

A square matrix that is not invertible is called singular. Note that this only
applies to square matrices2 .
Note We will see how to calculate inverses soon.

Definition 69. The determinant of a matrix A (written detA = |A|) is defined

inductively.

n=1 A
(1×1)

detA = |A| ≡ a11

n≥2 A
(n×n)

detA = |A| ≡ a11 |A−11 | − a12 |A−12 | + a13 |A−13 | − · · · ± a1n |A−1n |
where A−1j is the matrix formed by deleting the first row and jth col-
umn of A.

Note A−1j is an (n − 1) × (n − 1) dimensional matrix.

Example 22. If  
a11 a12
A = [aij ] =
2×2 a21 a22
=⇒ |A| = a11 a22 − a12 a21
2
You can find one-sided “pseudo inverses” for all matrices, even those that are not square.
8.2. MATRICES 87

Example 23. If
 
a11 a12 a13
A = [aij ] =  a21 a22 a23 
3×3
a31 a32 a33
     
 a22 a23   a21 a23   a21 a22 
=⇒ |A| = a11 
  − a 12
  + a 13
 
a32 a33   a31 a33   a31 a32 

The determinant is useful primarily because of the following result:

Theorem 44. A matrix is invertible if and only if its determinant 6= 0.

Definition 70. The Inverse of a matrix A is defined as

n×n

1
A−1 = · adjA
|A|
where adjA is the adjoint of A and we will not show how to calculate it here.

Example 24. If A is a (2 × 2) matrix and invertible then

 
−1 1 a22 −a12
A = ·
|A| −a21 a11

Note It is worthwhile memorizing the formula for the inverse of a 2 × 2

matrix. Leave inverting higher-order matrices to computers.

8.2.2 Inner Product and Distance

Definition 71 (Inner Product). If x, y ∈ Mn×1 , then the inner product (or
dot product or scalar product) is given by

xt y = x1 y1 + x2 y2 + · · · + xn yx
Xn
= xi y i
i=1
88 CHAPTER 8. BASIC LINEAR ALGEBRA

Note that xt y = yt x. We will have

Pn reason to use this concept when we do
calculus, and will write x · y = i=1 xi yi .

Definition 72 (Distance). We generalize the notion of distance.

For Rn , a function
d : Rn × Rn −→ R
is called a metric if for any two points x and y ∈ Rn

(a) d(x, y) ≥ 0
(b) d(x, y) = 0 ⇐⇒ x = y
(c) d(x, y) = d(y, x)
(d) d(x, y) ≤ d(x, z) + d(z, y), for any z ∈ Rn

The last of these properties is called the triangle inequality.

If you think about an example on a map (e.g. in R2 ), all this is saying is
that it is a shorter distance to walk in a straight line from x to y, than it is
to walk from x to z and then from z to y.

A metric is a generalized distance function. It is possible to do calculus on

abstract spaces that have a metric (they are called metric spaces). Usually
there are many possible ways to define a metric. Even in Rn there are many
possibilities:
Example 25. 
1, if x 6= y,
d(x, y) =
0, if x = y.

This satisfies the definition of a metric. It basically tells you whether x = y.

The metric defined by
Example 26. X
d(x, y) = |xi − yi |
i=1,...,n

states that the distance between two points is the length of the path connect-
ing the two points using segments parallel to the coordinate axes.

We will be satisfied with the standard Euclidean metric:

8.2. MATRICES 89

Example 27.
d(x, y) = kx − yk
where
q
kzk = z12 + z22 + · · · + zn2
v
u n
uX
=t zi2
i=1

Under the Euclidean metric, the distance between two points is the length of
the line segment connecting the points. We call kzk, which is the distance
between 0 and z the norm of z.
Notice that kzk2 = z · z.
When x·y = 0 we say that x and y are orthogonal/at right angles/perpendicular.
It is a surprising geometric property that two vectors are perpendicular if
and only if their inner product is zero. This fact follows rather easily from
“The Law of Cosines.” The law of cosines states that if a triangle has sides
A,B, and C and the angle θ opposite the side c, then

c2 = a2 + b2 − 2ab cos(θ),
where a, b, and c are the lengths of A, B, and C respectively. This means
that:

(x − y) · (x − y) = x · x + y · y − 2 kxk kxk cos(θ),

where θ is the angle between x and y. If you multiply everything out you
get the identity:

kxk kxk cos(θ) = xt y. (8.1)

Equation (8.1) has two nice consequences. First, it justifies the use of the
term orthogonal: The inner product of two non-zero vectors is zero if and
only if the cosine of the angle between them is zero. Second, it gives you an
upper bound of the inner product (because the absolute value of the cosine
is less than or equal to one):
 
kxk kxk ≥ xt y .
90 CHAPTER 8. BASIC LINEAR ALGEBRA

8.3 Systems of Linear Equations

Consider the system of n equations in m variables:

y1 = α11 x1 + α12 x2 + · · · + α1n xn

y2 = α21 x1 + α22 x2 + · · · + α2n xn
..
.
yi = αi1 x1 + αi2 x2 + · · · + αin xn
..
.
ym = αm1 x1 + αm2 x2 + · · · + αmn xn

Here the variables are the xj . This can be written as

y = A · x
(m×1) (m×n) (n×1)
| {z }
(m×1)

where    
y1 x1
 y2   x2 
y = , x =
   
..  .. 
(m×1)  .  (n×1)  . 
ym xn
 
α11 α12 · · · α1n
 α21 α22 · · · α2n 
A =  ..  = [αij ]
 
.. ..
m×n  . . . 
αm1 αm2 · · · αmn
or, putting it all together

y = A · x
(m×1) (m×n) (n×1)
| {z }
(m×1)

     
y1 α11 α12 ··· α1n x1
 y2   α21 α22 ··· α2n   x2 
= ·
     
 .. .. .. .. .. 
 .   . . .   . 
ym αm1 αm2 · · · αmn xn
8.3. SYSTEMS OF LINEAR EQUATIONS 91

Note you should convince yourself that if you multiply out the RHS of the
above equation and then compare corresponding entries of the new (m × 1)
vectors that the result is equivalent to the original system of equations. The
value of matrices is that they permit you to write the complicated system
of equations in a simple form. Once you have written them a system of
equations in this way, you can use matrix operations to solve some systems.

Example 28. In high school you probably solved equations of the form:

3x1 − 2x2 = 7
8x1 + x2 = 25

Well matrix algebra is just a clever way to solve these in one go.
So here we have that
     
3 −2 x1 7
A = , x = , y =
(2×2) 8 1 (2×1) x2 (2×1) 25

And can write this as

A · x = y
(2×2) (2×1) (2×1)

And obviously we want to solve for x.

So we multiply both sides of the equation on the left (recall that it does
matter what side of the equation you multiply) by A−1 ,

−1 −1
| {zA} x = A y
A
I2

⇐⇒ I2 x = A−1 y

⇐⇒ x = A−1 · y
(2×2) (2×1)
| {z }
(2×1)
92 CHAPTER 8. BASIC LINEAR ALGEBRA

and we have that

 
−1 1 1 2
A = ·
|A| −8 3
 
1 1 2
= ·
|(3 × 1) − (−2 × 8)| −8 3
 
1 1 2
= ·
|19| −8 3

So
x = A−1 y
   
1 1 2 7
= · ·
|19| −8 3 25
 
1 57
= ·
19 19
 
3
=
1

The example illustrates some general properties. If you have exactly as

many equations as unknowns (so that the matrix A is square, then the system
has a unique solution if and only if A is invertible. If A is invertible, it is
obvious that the solution is unique (and given by the formula: x = A−1 y).
If A is not invertible, it is the case that there is a nonzero z such that Az = 0.
This means that if you can find one solution to Ax = y,3 then you can find
infinitely many solutions (by adding arbitrary multiples of z to the original
solution. Intuitively, it is hard for a matrix to be singular, so “most” of the
time systems of n equations and n unknowns have unique solutions.
These comments do not tell you about situations where the number of equa-
tions is not equal to the number of unknowns. In most cases, when there are
extra equations, a system of linear equations will have no solutions. If there
are more unknowns than equations, typically the system will have infinitely
many solutions (it is possible for the system to have no solutions, but it is
not possible for the system to have a unique solution).
A system of equations of the form Ax = 0 is called a homogeneous system
of equations. Such a system always has a solution (x = 0). The solution
will not be unique if there are more unknowns than equations.
3
When A is singular, there is no guarantee that a solution to this equation exists.
8.4. LINEAR ALGEBRA: MAIN THEORY 93

The standard way to establish these results is by applying “elementary row

operations” to A to transform a system of equations into an equivalent sys-
tem that is easier to analyze.

8.4 Linear Algebra: Main Theory

Euclidean Spaces (and Vector Spaces more generally) have a nice structure
that permits you to write all elements in terms of a fixed set of elements. In
order to describe this theory, we need a few definitions.
combination of a collection of vectors {x1 , . . . , xk } is a vector of
A linear P
the form ki=1 λi xi for some scalars λ1 , . . . , λk .
Rn has the property that sums and scalar multiples of elements of Rn re-
main in the set. Hence if we are given some elements of the set, all linear
combinations will also be in the set. Some subsets are special because they
contain no redundancies:
Definition 73. A collection of vectors {x1 , . . . , xk } is linearly independent
if ki=1 λi xi = 0 if and only if λi = 0 for all i.
P

Here is the way in which linear independence captures the idea of no redun-
dancies:
Theorem 45. If X = {x1 , . . . , xk } is a linearly independent collection of
vectors
Pk and z ∈ S(X), then there are unique λ1 , . . . , λk such that z =
i=1 λi xi .

Proof. Existence follows from the definition of span. Suppose that there are
two linear combinations that of the elements of X that yield z so that
k
X
z= λi x i
i=1

and
k
X
z= λ0i xi .
i=1
Subtract the equations to obtain:
k
X
0= (λ0i − λi ) xi .
i=1
94 CHAPTER 8. BASIC LINEAR ALGEBRA

By linear independence, λi = λ0i for all i, the desired result.

Next, let us investigate the set of things that can be described by a collection
of vectors.

Pspan of a collection of vectors X = {x1 , . . . , xk } is

Definition 74. The
S(X) = {y : y = ki=1 λi xi for some scalars λ1 , . . . , λk }.

S(X) is the smallest vector space containing all of the vectors in X.

Definition 75. The dimension of a vector space is N , where N is the small-
est number of vectors needed to span the space.

We deal only with finite dimensional vector spaces. We’ll see this defini-
tion agrees with the intuitive notion of dimension. In particular, Rn has
dimension n.
Definition 76. A basis for a vector span V is any collection of linearly
independent vectors than span V .
Theorem 46. If X = {x1 , . . . , xk } is a set of linearly independent vectors
that does not span V , then there exists v ∈ V such that X ∪ {v} is linearly
independent.

Proof. Take v ∈ V such that v 6= 0 and v 6∈ S(X). X ∪ {v} is a linearly

independent set. To see this, suppose that there exists λi i = 0, . . . , k such
that at least one λi 6= 0 and
k
X
λ0 v + λi xi . (8.2)
i=1

If λ0 = 0, then X are not linearly independent. If λ0 6= 0, then equa-

tion (8.2) can be rewritten
k
X λi
v= xi .
λ
i=1 0

In either case, we have a contradiction.

Definition 77. The standard basis for Rn consists of the set of N vectors ei ,
i = 1, . . . , N , where ei is the vector with component 1 in the ith position
and zero in all other positions.
8.5. EIGENVECTORS AND EIGENVALUES 95

You should check that the standard basis really is a linearly independent set
that spans Rn . Also notice that the elements of the standard basis are mutu-
ally orthogonal. When this happens, we say that the basis is orthogonal. It
is also the case that each basis element has unit length. When this also hap-
pens, we say that the basis is orthonormal. It is always possible to find an
orthonormal basis.4 They are particularly useful because it is easy to figure
out how to express an arbitrary element of X in terms of the basis.
It follows from these observations that each vector v has a unique repre-
sentation in terms of the basis, where the representation consists of the λi
used in the linear combination that expresses v in terms of the basis. For the
standard basis, this representation is just the components of the vector.
It is not hard (but a bit tedious) to prove that all bases have the same number
of elements. (This follows from the observation that any system of n homo-
geneous equations and m > n unknowns has a non-trivial solution, which
in turn follows from “row-reduction” arguments.)

8.5 Eigenvectors and Eigenvalues

An eigenvalue of the square matrix A is a number λ with the property A−λI
is singular. If λ is an eigenvalue of A, then any x 6= 0 such that (A − λI)x =
0 is called an eigenvector of A associated with the eigenvalue λ.
Eigenvalues are those values for which the equation
Ax = λx
has a non-zero solution. You can compute eigenvalues (in theory) by solv-
ing the equation det(A − λI) = 0. If A is an n × n matrix, then this
characteristic equation is a polynomial equation of degree n. By the Fun-
damental Theorem of Algebra, it will have n (not necessarily distinct and
not necessarily real) roots. That is, the characteristic polynomial can be
written
P (λ) = (λ − r1 )m1 · · · (λ − rk )mk ,
where r1 , r2 . . . , rk are the distinct roots (ri 6= rj when i 6= j) and mi
are positive integers summing to n. We call mi the multiplicity of root
ri . Eigenvalues and their corresponding eigenvectors are important because
they enable one to relate complicated matrices to simple ones.
4
We have exhibited an orthonormal basis for Rn . It is possible to construct an orthonormal
basis for any vector space.
96 CHAPTER 8. BASIC LINEAR ALGEBRA

Theorem 47. If A is an n × n matrix that has n distinct eigen-values or

is symmetric, then there exists an invertible n × n matrix P and a diagonal
matrix D such that A = PDP−1 . Moveover, the diagonal entries of D are the
eigenvalues of A and the columns of P are the corresponding eigenvectors.

The theorem says that if A satisfies certain conditions, then it is “related”

to a diagonal matrix. It also tells you how to find the diagonal matrix. The
relationship A = PDP−1 is quite useful. For example, it follows from the
relationship that Ak = PDk P−1 . It is much easier to raise a diagonal matrix
to a power than find a power of a general matrix.

Proof. Suppose that λ is an eigenvalue of A and x is an eigenvector. This

means that Ax = λx. If P is a matrix with column j equal to an eigenvector
associated with λi , it follows that AP = PD. The theorem would follow if
we could guarantee that P is invertible.
When A is symmetric, one can prove that A has only real eigenvalues and
that one can find n linearly independent eigenvectors even if the eigenvalues
are not distinct. This result is elementary (but uses some basic facts about
complex numbers).
Suppose that you have
Ax = λx
for a real symmetric matrix A. Taking the complex conjugate of both sides
yields
Ax∗ = λ∗ x∗ .
We have

(x∗ )t Ax − xt Ax∗ = (λ − λ∗ )xt x∗ (8.3)

∗ t
(using (x ) x = xt x∗ . On the other hand, the left-hand side of equation (8.3)

t
xt Ax∗ = xt Ax∗ = (x∗ )t Ax
where the first equation is true because the transpose of a number is equal
to itself and the second equation follows because A is symmetric. It follows
that λ is equal to its transpose, so it must be real.
It is also the case that the eigenvectors corresponding to distinct eigenvalues
of a symmetric matrix are orthogonal. To see this observe that Axi = λi xi
and Axj = λj xj implies that
λi xtj xi = xti Axj = xtj Axi = λj xtj xi .
8.5. EIGENVECTORS AND EIGENVALUES 97

In general, one can prove that the eigenvectors of distinct eigenvalues are
distinct. To see this, suppose that λ1 , . . . , λk are distinct eigenvalues and
x1 , . . . , xk are associated eigenvectors. In order to reach a contradiction,
suppose that the vectors are linearly dependent. Without loss of generality,
we may assume that {x1 , . . . , xk−1 } are linearly independent, but that xk can
be written as a linear combination of the first k − 1 vectors. This means that
there exists αi i = 1, . . . , k − 1 not all zero such that:
k−1
X
αi xi = xk . (8.4)
i=1

Multiply both sides of equation (8.4) by A and use the eigenvalue property
to obtain:
Xk−1
αi λi xi = λk xk . (8.5)
i=1

Multiply equation (8.4) by λk and subtract it from equation (8.5) to obtain:

k−1
X
αi (λi − λk )xi = 0. (8.6)
i=1

Since the eigenvalues are distinct, equation (8.6) gives a non-trivial linear
combination of the first k − 1 xi that is equal to 0, which contradicts linear
independence.

Here are some useful facts about determinants and eigenvalues. (The proofs
range from obvious to tedious.)

(a) det AB = det BA

(b) If D is a diagonal matrix, then det D is equal to the product of its
diagonal elements.
(c) det A is equal to the product of the eigenvalues of A.
(d) The trace of a square matrix A isP equal to the sum of the P
diagonal
elements of A. That is, tr (A) = i=1 aii . Fact: tr (A) = ni=1 λi ,
n

where λi is the ith eigenvalue of A (eigenvalues counted with multi-

plicity).

One variation on the symmetric case is particularly useful in the next sec-
tion. When A is symmetric, then we take take the eigenvectors of A to
98 CHAPTER 8. BASIC LINEAR ALGEBRA

be orthonormal. In this case, the P in the previous theorem has the prop-
erty that P−1 = Pt . Eigenvalues turn out to be important in many different
places. They play a role in the study of stability of difference and differen-
tial equations. They make certain computations easy. They make it possible
to define a sense in which matrices can be positive and negative that allows
us to generalize the one-variable second-order conditions. The next topic
will do this.

8.6 Quadratic Forms

Definition 78. A quadratic form in n variables is any function Q : Rn −→
R that can be written Q(x) = xt Ax where A is a symmetric n × n matrix.

When n = 1 a quadratic form is a function of the form ax2 . When n = 2 it

is a function of the form a11 x21 + 2a12 x1 x2 + a22 x22 (remember a12 = a21 ).
When n = 3, it is a function of the form a11 x21 + a22 x22 + a33 x23 + 2a12 x1 x2 +
2a13 x1 x3 + 2a23 x2 x3 . A quadratic form is second-degree polynomial that
has no constant term.
We will see soon that the second derivative of a real-valued function on Rn
is not a single function, but a collection of n2 functions. Quadratic forms
will be the way in which we study second derivatives. In particular, they are
important for checking second-order conditions and concavity of functions.
Definition 79. A quadratic form Q(x) is

(a) positive definite if Q(x) > 0 for all x 6= 0.

(b) positive semi definite if Q(x) ≥ 0 for all x.
(c) negative definite if Q(x) < 0 for all x 6= 0.
(d) negative semi definite if Q(x) ≤ 0 for all x.
(e) indefinite if there exists x and y such that Q(x) > 0 > Q(y).

The definition provides a notion of positivity and negativity for matrices.

You should check to confirm that the definitions coincide with ordinary no-
tions of positive and negative when Q is a quadratic form of one variable
(that is Q(x) = Ax2 ). In this case the quadratic form is positive definite
in A > 0, negative (and positive) semi-definite when A = 0 and negative
definite when a < 0. When n > 1 it is not hard to find indefinite matrices.
We will see this soon.
8.6. QUADRATIC FORMS 99

If A happened to be a diagonal matrix, then it is easy to classifyP the associ-

ated quadratic form according to the definition. Q(x) = x Ax = ni=1 aii x2i .
t

This quadratic form is positive definite if and only if all of the aii > 0, neg-
ative definite if and only if all of the aii < 0, positive semi definite if and
only if aii ≥ 0, for all i negative semi definite if and only if aii ≤ 0 for all i,
and indefinite if A has both negative and positive diagonal entries.
The theory of diagonalization gives us a way to translate use these results for
all matrices. We know that if A is a symmetric matrix, then it can be written
A = Rt DR, where D is a diagonal matrix with (real) eigenvalues down the
diagonal and R is an orthogonal matrix. This means that the quadratic form:
Q(x) = xt Ax = xt Rt DRx = (Rx)t D (Rx). This expression is useful
because it means that the definiteness of A is equivalent to the definiteness
of its diagonal matrix of eigenvectors, D. (Notice that if I can find an x such
that xt Ax > 0, then I can find an x such that yt Dy (y = Rx) and conversely.)

Theorem 48. The quadratic form Q(x) = xt Ax is

(a) positive definite if λi > 0 for all i.

(b) positive semi definite if λi ≥ 0 for all i.
(c) negative definite if λi < 0 for all i.
(d) negative semi definite if λi ≤ 0 for all i.
(e) indefinite if there exists j and k such that λj > 0 > λk .

There is a computational trick that often allows you to identify definiteness

without computing eigenvalue.

Definition 80. A principal submatrix of a square matrix A is the matrix

obtained by deleting any k rows and the corresponding k columns. The
determinant of a principal submatrix is called the principal minor of A. The
leading principal submatrix of order k of an n × n matrix is obtained by
deleting the last n − k rows and column of the matrix. The determinant of
a leading principal submatrix is called the leading principal minor of A.

Theorem 49. A matrix is

(a) positive definite if and only if all of its leading principal minors are
positive.
(b) negative definite if and only if its odd principal minors and negative
and its even principal minors are positive.
100 CHAPTER 8. BASIC LINEAR ALGEBRA

(c) indefinite if one of its kth order leading principal minors is negative
for an even k or if there are two odd leading principal minors that
have different signs.

The theorem permits you to classify the definiteness of matrices without

finding eigenvalues. It may seem strange at first, but you can remember it
by thinking about diagonal matrices.
Chapter 9

Multivariable Calculus

The goal is to extend the calculus from real-valued functions of a real vari-
able to general functions from Rn to Rm . Some ideas generalize easily, but
going from one dimensional domains to many dimensional domains raises
new issues that need discussion. Raising the dimension of the range space,
on the other hand, raises no new conceptual issues. Consequently, we begin
our discussion to real-valued functions. We will explicitly consider higher
dimensional ranges only when convenient or necessary. (It will be conve-
nient to talk about linear functions in general terms. It is necessary to talk
about the most interesting generalization of the Chain Rule and for the dis-
cussions of inverse and implicit functions.)

9.1 Linear Structures

To do calculus, you need to understand linear “objects” in Rn . In the plane
there are three kinds of subspace (recall that a subspace is a set L with the
property that if x and y are in the set, then so are x + y and λx for all λ).
These sets are: the entire plane, any line through the origin, and the origin.
In higher dimensional spaces, there are more linear subsets.
It turns out that the most useful to us are one dimensional linear subsets,
lines, and n − 1 dimensional subsets, hyperplanes.

Definition 81. A line is described by a point x and a direction v. It can be

represented as {z : there exists t ∈ R such that z = x + tv}.

101
102 CHAPTER 9. MULTIVARIABLE CALCULUS

If we constrain t ∈ [0, 1] in the definition, then the set is the line segment
connecting x to x+v. Two points still determine a line: The line connecting
x to y can be viewed as the line containing x in the direction v. You should
check that this is the same as the line through y in the direction v.
Definition 82. A hyperplane is described by a point x0 and a normal direc-
tion p ∈ Rn , p 6= 0. It can be represented as {z : p · (z − x0 ) = 0}. p is
called the normal direction of the plane.

The interpretation of the definition is that a hyperplane consists of all of the

z with the property that the direction z − x0 is normal to p.
In R2 lines are hyperplanes. In R3 hyperplanes are “ordinary” planes.
Lines and hyperplanes are two kinds of “flat” subset of Rn . Lines are sub-
sets of dimension one. Hyperplanes are subsets of dimension n − 1 or
co-dimension one. You can have a flat subsets of any dimension less than n.
Although, in general, lines and hyperplanes are not subspaces (because they
do not contain the origin) you obtain these sets by “translating” a subspace
that is, by adding the same constant to all of its elements.
Definition 83. A linear manifold of Rn is a set S such that there is a sub-
space V on Rn and x0 ∈ Rn with S = V + {x0 }.

In the above definition, V + {x0 } ≡ {y : y = v + x0 for some v ∈ V }.

The definition is a bit pedantic. Officially lines and hyperplanes are linear
manifolds and not linear subspaces.
It is worthwhile reviewing the concepts of line and hyperplane. Given two
points x and y, you can construct a line that passes through the points. This
line is
{z : z = x + t(y − x) for some t.}
This formulation is somewhat different from the one normally sees, but it is
equivalent. Writing out the two-dimensional version yields:
z1 = x1 + t(y1 − x1 ) and z2 = x2 + t(y2 − x2 ).
If you use the equation for z1 to solve for t and substitute out you get:
(y2 − x2 ) (z1 − x1 )
z2 = x2 +
y1 − x1
or
y2 − x2
z2 − x2 = (z1 − x1 ) ,
y1 − x1
9.1. LINEAR STRUCTURES 103

which is the standard way to represent the equation of a line (in the plane)
through the point (x1 , x2 ) with slope (y2 − x2 )(y1 − x1 ). This means that
the “parametric” representation is essentially equivalent to the standard rep-
resentation in R2 .1 The familiar ways to represent lines do not work in
higher dimensions. The reason is that one linear equation in Rn typically
has an n − 1 dimensional solution set, so it is a good way to describe a one
dimensional set only if n = 2.
You need two pieces of information to describe a line. If the information
consists of a point and a direction, then the parametric version of the line is
immediately available. If the information consists of two points, then you
form a direction by subtracting one point from the other (the order is not
important).
You can describe a hyperplane easily given a point and a (normal) direction.
Note that the direction of a line is the direction you follow to stay on the line.
The direction for a hyperplane is the direction you follow to go away from
the hyperplane. If you are given a point and a normal direction, then you
can immediately write the equation for the hyperplane. What other pieces of
information determine a hyperplane? In R3 , a hyperplane is just a standard
plane. Typically, three points determine a plane (if the three points are all
on the same line, then infinitely many planes pass through the points). How
can you determine the equation of a plane in R3 that passes through three
given points? A mechanical procedure is to note that the equation for the
plane can always be written Ax1 + Bx2 + Cx3 = D and, use the three
points to find values for the coefficients. For example, if the points are
(1, 2 − 3), (0, 1, 1), (2, 1, 1), then we can solve:
A + 2B − 3C = D
B + C = D
2A + B + C = D

Doing so yields (A, B, C, D) = (0, .8D, .2D, D). (If you find one set of
coefficients that work, any non-zero multiple will also work.) Hence an
equation for the plane is: 4x2 + x3 = 5 you can check that the three points
actually satisfy this equation.
An alternate computation technique is to look for a normal direction. A nor-
mal direction is a direction that is orthogonal to all directions in the plane.
1
The parametric representation is actually a bit more general, since it allows you to describe
lines that are parallel to the vertical axis. Because these lines have infinite slope, they cannot be
represented in standard form.
104 CHAPTER 9. MULTIVARIABLE CALCULUS

A direction in the plane is a direction of a line in the plane. You can get
such a direction by subtracting any two points in the plane. A two dimen-
sional hyperplane will have two independent directions. For this example,
one direction can come from the difference between the first two points:
(1, 1, −4) and the other can come from the difference between the second
and third points (−2, 0, 0) (a third direction will be redundant, but you can
do the computation using the direction of the line connecting (1, 2, −3) and
(2, 1, 1) instead of either of directions computed above). Once you have
two directions, you want to find a normal to both of them. That is, a p such
that p 6= 0 and p · (1, 1, −4) = p · (−2, 0, 0) = 0. This is a system of two
equations and three variables. All multiples of (0, 4, 1) solve the equations.2
Hence the equation for the hyperplane is (0, 4, 1)·(x1 −1, x2 −2, x3 +3) = 0.
You can check that this agrees with the equation will found earlier. It also
would be equivalent to the equation you would obtain if you used either of
the other two given points as “the point on the plane.”

9.2 Linear Functions

Definition 84. A function

L : Rn −→ Rm

is linear if and only if

(a) If for all x and y, f (x + y) = x + y and

(b) for all scalars λ, f (λx) = λf (x).

The first condition says that a linear function must be additive. The sec-
ond condition says that it must have constant returns to scale. The condi-
tions generate several obvious consequences. If L is a linear function, then
L(0) = 0 and, more generally, L(x) = −L(−x). It is an important observa-
tion that any linear function can be “represented” by matrix multiplication.
Given a linear function, compute L(ei ), where ei is the ith standard basis
element. Can this ai and let A be the square matrix with ith column equal
to ai . Note that A must have n columns and m rows. Note also that (by the
2
The “cross product” is a computational tool that allows you to mechanically compute a direc-
tion perpendicular to two given directions.
9.3. REPRESENTING FUNCTIONS 105

properties of matrix multiplication and linear functions) L(x) = Ax for all

x. This means that any linear function can be thought of as matrix multipli-
cation (and the matrix has a column for each dimension in the domain and
a row for each dimension in the range). Conversely, every matrix gives rise
to a linear function. Hence the identification between linear functions and
matrices is perfect. When the linear function is real valued, linear functions
can be thought of as taking an inner product.
The identification of linear functions with matrix multiplication supplies
another interpretation of matrix multiplication. If L : Rm −→ Rk and
M : Rk −→ Rn are functions, then we can form the composite function
M ◦ L : Rm −→ Rn by M ◦ L(x) = M (L(x)). If M and L are linear
functions, and A is the matrix that represents M and B is the matrix that
represents L, then AB is the matrix that represents M ◦ L. (Officially, you
must verify this by checking how the composite function transforms the
standard basis elements.)

9.3 Representing Functions

Functions of one variable are easy to visualize because we could draw their
graphs in the plane. In general, the graph of a function f is Gr(f ) =
{(x, y) : y = f (x)}. This means that if f : Rn −→ R, then the graph is
a subset of Rn+1 . If n = 2, then the graph is a subset of R3 , so someone
with a good imagination of a three-dimensional drawing surface could vi-
sualize it. If n > 2 there is no hope. You can get some intuition by looking
at “slices” of the graph obtained by holding the function’s value constant.

Definition 85. A level set is the set of points such that the functions achieves
the same value. Formally it is defined as the set

{x ∈ Rn | f (x) = c} for any c ∈ R

While the graph of the function is a subset of Rn+1 , the level set (actually,
level sets) are subsets of Rn .

Example 29.
f (x) = x21 + x22
So the function is R2 −→ R and therefore the graph is in R3 .
106 CHAPTER 9. MULTIVARIABLE CALCULUS

The level sets of this function are circles in the plane. (The graph is a cone.)
FIGURE GOES HERE

Example 30. A good example to help understand this is a utility function (of
which you will see lots!). A utility function is a function which “measures” a
person’s happiness. It is usually denoted U . In 200A you will see conditions
necessary for the existence of the utility function but for now we will just
assume that it exists, and is strictly increasing in each argument. Suppose
we have a guy Joel whose utility function is just a function of the number of
apples and the number of bananas he eats. So his happiness is determined
solely by the number of apples and bananas he eats, and nothing else. Thus
we lose no information when we think about utility as a function of two
variables:
U : R2 −→ R
U (xA , xB ) where xA is the number of apples he eats, and xB is the number
of bananas he eats.
A level set is all the different possible combinations of apples and bananas
that give him the same utility level i.e. that leave him equally happy!jFor
example Joel might really like apples and only slightly like bananas. So 3
apples and 2 bananas might make him as happy as 1 apple and 10 bananas.
In other words he needs lots of bananas to compensate for the loss of the 2
apples.

If the only functions we dealt with were utility functions, we would call
level sets “indifference curves.” In economics, typically curves that are “iso-
SOMETHING” are level sets of some function.

In fact maybe another example is in order.

Example 31. Suppose we have a guy Joel who only derives joy from teach-
ing mathematics. Nothing else in the world gives him any pleasure, and as
such his utility function UJ is only a function of the number of hours of he
spends teaching mathematics HM . Now we also make the assumption that
Joel’s utility is strictly increasing in hours spend teaching mathematics, the
more he teaches the happier he is3 . So the question is does Joel’s utility
function UJ have any level sets? Since utility is a function of one variable,
3
you may doubt that Joel is really like this but I assure you he is!
9.3. REPRESENTING FUNCTIONS 107

Joel’s level “sets” are zero dimensional objects – points. Since if his utility
1/2
function is defined as UJ (HM ) = HM and Joel teaches for 4 hours (i.e.
HM = 4), then UJ (4) = 41/2 = 2. So is there any other combination of
hours teaching that could leave Joel equally happy? Obviously not, since
his utility is only a function of one argument, and it is strictly increasing, no
two distinct values can leave him equally happy.4

Example 32. Recall Example 29

So plotting level sets of a function f : R2 −→ R is really just a clever way of

representing the information from a 3d graph, but only drawing a 2d graph
(much easier to draw!).

A good example of this is a map of a country. When you spread a map

out in front of you we know that left-right (←→) represents East-West, and
that up-down (l) represents North-South. So how do we represent moun-
tains and valleys i.e. points on the earth of different altitude? The answer
is with level sets! There are many contour lines drawn all over a map and
at each point along these lines the earth is at the same altitude. Of course to
be completely rigorous it should also be obvious which direction altitude is
increasing in.
So in this case our function would take in coordinates (values for east and
west) and spit out the value (altitude) at those coordinates (values).

Definition 86. We define the upper contour set of f as the set

{x ∈ Rn | f (x) ≥ c} c∈R

And we define the lower contour set of f as the set

{x ∈ Rn | f (x) ≤ c} c∈R
4
So you can see that level sets reference the arguments of a function. And functions with two
or more arguments are much more likely to have level sets than functions of one argument, since
you can have many different combinations of the arguments.
108 CHAPTER 9. MULTIVARIABLE CALCULUS

So referring back to our map example. The upper contour set of a point x
would be a set of all the coordinates such that if we plugged those coor-
dinates into our altitude function, it would give out a value greater than or
equal to the value at the point x, i.e. all points that are at a higher altitude
than x.

9.4 Limits and Continuity

Definition 87 (Limit of a Function).

f : Rn −→ R

lim f (x) = c ∈ R
x−→a

⇐⇒ ∀  > 0, ∃ δ > 0 such that

0 < d(x, a) < δ

=⇒ |f (x) − c| < 

This definition agrees with the earlier definition, although there are two
twists. First, a general “distance function” replaces absolute values in the
condition that says that x is close to a. For our purposes, the distance func-
tion will always be the standard Euclidean distance. Second, there we do
not define one-sided continuity.

Definition 88 (Continuity of a Function).

f : Rn −→ R

is called continuous at a point a ∈ Rn if

lim f (x) = f (a)

x−→a

Again, this definition is a simple generalization of the one-variable defini-

tion.
Exercise 2. Prove that f (x) = x1 x2 is continuous at the point (1, 1).
9.4. LIMITS AND CONTINUITY 109

We must show that ∀  > 0, ∃ δ > 0 such that

k(x1 , x2 ) − (1, 1)k < δ

=⇒ |f (x1 , x2 ) − 1| < 

Note that
p
k(x1 , x2 ) − (1, 1)k = (x1 − 1)2 + (x2 − 1)2

Also

|f (x1 , x2 ) − 1| = |x1 x2 − 1|
= |x2 x1 − x1 + x1 − 1|
= |x1 (x2 − 1) + x1 − 1|
4
≤ |x1 (x2 − 1)| + |x1 − 1|
| {z } | {z }
< 12  < 12 

<

where the second last inequality is got using the triangle inequality hence
the 4 superscipt.
1
For any given  > 0 let δ = min 4
, 1 . Then

1
k(x1 , x2 ) − (1, 1)k < 
4
1 1
=⇒ |x1 − 1| <  and |x2 − 1| < 
4 4
Also we have that x1 < 2. Thus
1 1
|x1 (x2 − 1)| + |x1 − 1| < 2 ·  + 
4 4
3
= 
4
implying that
3
|f (x1 , x2 ) − 1| < 
4
<
110 CHAPTER 9. MULTIVARIABLE CALCULUS

9.5 Sequences
(we now use superscript for elements of sequence)
 k ∞
x k=1 sequences of vectors in Rn .

xk ∈ Rn , ∀k
xk = (xk1 , xk2 , . . . , xkn )

 ∞
Definition 89. A sequence xk k=1 converges to a point x ∈ Rn , that is
xk −→ x, if and only if ∀  > 0, ∃ K ∈ P such that k ≥ K =⇒

d(xk , x) < 

Definition 90. For a, b ∈ Rn , we say

a≥b

⇐⇒ ai ≥ bi ∀ i = 1, 2, . . . , n
and
a>b
⇐⇒ ai ≥ bi ∀ i = 1, 2, . . . , n,
and aj > bj for some j

Definition 91. Let X ⊂ R.

X is bounded from above if ∃ M ∈ Rn such that

M ≥ x, ∀x ∈ X

X is bounded from below if ∃ m ∈ Rn such that

m ≤ x, ∀x ∈ X

Now we define what mean by vectors being “greater” or “less” than each
other.
9.6. PARTIAL DERIVATIVES AND DIRECTIONAL DERIVATIVES 111

Definition 92. X is said to be closed if, for every sequence xk from X , if

xk −→ x ∈ Rn

(so x is a limit point of X ), then x ∈ X .

Definition 93. X is said to be a compact space if every sequence from X

has a subsequence that converges to a point in X .

Definition 94. For a metric d, a Ball of Radius  around a point x ∈ Rn is

defined as
B (x) ≡ { y ∈ Rn | d(x, y) < }

FIGURE GOES HERE

Definition 95. X is called open if ∀x ∈ X , ∃ > 0 such that

B (x) ⊂ X

Note the following 2 common misconceptions

X not open, does not imply X closed

X not closed, does not imply X open
For example, Rn is both open and closed.
[0, 1) is neither open nor closed.

9.6 Partial Derivatives and Directional Deriva-

tives
Definition 96. Take f : Rn −→ R. The ith partial derivative of f at x is
defined as
∂f f (x + hei ) − f (x)
(x) ≡ lim .
∂xi h−→0 h
112 CHAPTER 9. MULTIVARIABLE CALCULUS

Treat every other xj as a constant and take the derivative as though f we a

function of just xi .
As in the one variable case, partial derivatives need not exist. If the ith
partial derivative exists, then the function (when viewed as a function of xi
alone) must be continuous.
The definition illustrates the way that we will think about functions of many
variables. We we think of them as many functions of one variable. If it is
too hard to figure out how f is behaving when you move several variables
at once, hold all but one of the variables fixed and analyze the one-variable
function that results. A partial derivative is just an ordinary derivative of a
function of one variable. The one variable is one of the components of the
function. Roughly speaking, if a function is well behaved, knowing how it
changes in every direction allows you to know how it behaves in general.
These considerations suggest a more general concept.
Definition 97. Take f : Rn −→ R and let v be a unit vector in Rn . The
directional derivative of f in the direction v at x is defined as
f (x + hv) − f (x)
Dv (x) ≡ lim .
h−→0 h
It follows from the definition, that
∂f
(x) ≡ Dei (x).
∂xi
That is, the ith partial derivative is just a directional derivative in the direc-
tion ei .
Notice that to compute directional derivatives you are just computing the
derivative of a function of one variable. The one-variable function is the
function of h of the form f (x + hv) with x and v fixed.

9.7 Differentiability
Definition 98. We say a function f : Rn −→ Rm is differentiable at a ∈ Rn
if and only if there is a linear function L : Rn −→ Rm such that
m×1 m×1

n×1

f (x) − f (a) − L(x − a

lim = 0.
x−→a
x − a
n×1
9.7. DIFFERENTIABILITY 113

If L exists we call it the derivative of f at a and denote it by Df (a). In the

case f : Rn −→ R this is equivalent to
 
f (x) − f (a) L(x − a)
lim − =0
x−→a kx − ak kx − ak

This implies that if f is differentiable, then for any directions defined by

y ∈ Rn and a magnitude given by α ∈ R

f (a + αy) − f (a) Df (a)αy

lim = lim
α−→0 |α| kyk α−→0 |α| kyk

Df (a)y
=
kyk

That is, if a function in differentiable at a point, then all directional deriva-

tives exist at the point.

Stated more officially:

Theorem 50. If f : Rn −→ R is differentiable then all of its directional
derivatives exist. The directional derivative of f at a ∈ Rn in direction
v ∈ Rn is given by
Df (a)v

We assume that the direction v in the definition of directional derivative is

of unit length.
The theorem says two things. First, it says the if a function is differentiable,
then the matrix representation of the derivative is just the matrix of partial
derivatives. (This follows because the way that you get the matrix repre-
sentation is to evaluate the linear function on the standard basis elements.
Here, this would give you partial derivatives.) Second, it gives a formula for
computing derivational derivatives is any direction. A directional derivative
is just a weighted average of partial derivatives.
The definition of differentiability should look like the one-variable defini-
tion. In both cases, the derivative is the best linear approximation to the
function. In the multivariable setting a few things change. First, Euclidean
distance replaces absolute values. Notice that the objects inside the k·k are
vectors and not scalars. Taking their norms replaces a difference with a
114 CHAPTER 9. MULTIVARIABLE CALCULUS

non-negative scalar. If we did not do this (at least for the denominator),
then the ratios would not make sense. Second, because the linear function
has the same domain and range as f , it is more complicated than in the
one variable case. In the one variable case, the derivative of f evaluated
at a point is a single function of a single variable. This allows us to think
about f 0 as a real-valued function. When f : Rn −→ Rm , the derivative is a
linear function from Rn into Rm . This means that it can be represented by
matrix multiplication of a matrix with m rows and n columns. That is, the
derivative is described by mn numbers. What are these numbers? The com-
putation after the definition demonstrates that the entries in the matrix that
represents the derivative are the partial derivatives of (the component func-
tions of) f . This is why we typically think of the derivatives of multivariable
functions as “matrices of partial derivatives.”
Sometimes people represent the derivative of a function from Rn to R as a
vector rather than a linear function.

Definition 99. Given

f : Rn −→ R
the gradient of f at x ∈ Rn is

Df (x) = ∇f (x)
 
∂f ∂f ∂f
= (x), (x), . . . , (x)
∂x1 ∂x2 ∂xn

The definition just introduces notation. In terms of the notation, Theorem 50

states that
Dv f (a) = ∇f (a) · v. (9.1)

We have a definition of differentiability. We also have definitions of partial

derivatives. Theorem 50 says that if you know that a function is differen-
tiable, then you know that is has partial derivatives and that these partial
derivatives tell you everything you need to know about the derivative. What
about the converse? That is, if a function has partial derivatives, then do we
know it is differentiable? The answer is “almost”.
Theorem 51. Take f : Rn −→ R, if f ’s partial derivatives exist and are
continuous (f ∈ C 1 ), then f is differentiable.
9.7. DIFFERENTIABILITY 115

Theorem 51 states that if a function has partial derivatives in all directions,

then the function is differentiable provided that these partial derivatives
are continuous. There are standard examples of functions that have partial
derivatives in all directions but still are not differentiable. In these exam-
ples the partial derivatives are discontinuous. These examples also provide
situations in which equation (9.1) fails to be true (because the gradient can
exist although the function is not differentiable).

Example 33.
f (x) = x1 x2

∂f
=⇒ (0, 0) = x2 |(x1 =0,x2 =0)
∂x1
=0

∂f
=⇒ (0, 0) = x1 |(x1 =0,x2 =0)
∂x2
=0

f ((0, 0) + h(1, 1)) − f (0, 0)

lim √
h−→0 h 2
√
from (0, 0) the distance traveled to the point √(1, 1) is 2, so that is why we
normalize and divide in the denominator by 2.

f ((h, h)) − f (0, 0)

= lim √
h−→0 h 2
h2
= lim √
h−→0 h 2

=0

So for this function the derivative in every direction is zero.

116 CHAPTER 9. MULTIVARIABLE CALCULUS

Example 34.
1 1
f (x) = |x1 | 2 |x2 | 2
So
f ((0, 0) + h(1, 1)) − f (0, 0) f ((h, h)) − f (0, 0)
lim √ = lim √
h−→0 h 2 h−→0 h 2
1 1
|h| 2 |h| 2
= lim √
h−→0 h 2
h
= lim √
h−→0 h 2

Why include this example? The computation above √ tells√you that the direc-
tional derivative of f at (0, 0) in the direction (1/
√ 2, 1/ √ 2) does not exist.
(The one sided limit exists and is equal to 1/ 2 or −1/ 2. On the other
hand, you can easily check that both partial derivatives of the function at
(0, 0) exist and are equal to zero. Hence the formula for computing the di-
rection directive as the average of partials fails. Why? Because f is not
differentiable at (0, 0).

Question: What is the direction from x that most increases the value of f ?
Answer: It’s the direction given by the gradient.

Theorem 52. Suppose f : Rn −→ R is differentiable at x, then the direction

that maximizes the directional derivative at x is given by
v = ∇f (x)

One proves this theorem by noting that

|Dv f (x)| = |Df (x)v| ≤ ||∇f (x)|| × ||v||
with equality if and only if v = λDf (x).

9.8 Properties of the Derivative

Theorem 53. g, f : Rn −→ Rm both differentiable at a ∈ Rn

Then
9.8. PROPERTIES OF THE DERIVATIVE 117

(a)
D[cf ](a) = cDf (a)∀c ∈ R
(b)
D[f + g](a) = Df (a) + Dg(a)
For the case m=1
(c)
D[g · f ](a) = g(a) · Df (a) + f (a) · Dg(a)
1×n 1×1 1×n 1×1 1×n

(d)  
f g(a) · Df (a) − f (a) · Dg(a)
D (a) =
g [g(a)]2

Theorem 54 (Chain Rule Theorem). Suppose U ⊂ Rn and V ⊂ Rm

are open sets and suppose g : U −→ Rm is differentiable at x ∈ U and
f : V −→ Rl is differentiable at y ≡ g(x) ∈ V. Then f ◦ g is differentiable
at x and
D[f ◦ g](x) = Df (y)Dg(x)
| {z } l×m m×n
l×n

Proof. Though not a complete proof it is much more intuitive to consider

the case
f : Rm −→ R, and g : R −→ Rm
So we can see that l = n = 1 and

f ◦ g : R −→ R

Then
D(f ◦ g)0 (t) = Df (x)Dg(t) for x = g(t)
That is
∂f1 ∂x1 ∂f2 ∂x2 ∂fm ∂xm
D[f ◦ g](t) = · + · + ··· + ·
∂x1 ∂t ∂x2 ∂t ∂xm ∂t

Example 35. Let the variable t denote the price of oil. This one variable
induces an array of population responses (thus becomes a vector valued
function) like
118 CHAPTER 9. MULTIVARIABLE CALCULUS

(a) What type of car to buy?

(b) How fast to expand your business?
(c) How far to drive on holiday?
(d) etc.

and then these responses in turn have their own effect like determining GNP,
the variable y (which was got by the function f using these population re-
sponses).
g f
t −→ x = g(t) −→ y = f (g(t)) ∈ R
↓ ↓ ↓
price of oil actions taken by individuals GNP

∂y
D[f ◦ g](t) =
∂t
= D(f (g(t))Dg(t)
 dg1

  dt
∂f ∂f ..
= (g(t)), . . . , (g(t)) · 
 
∂x1 ∂xm . 
dgm
dt
m
X ∂y dg1
= ·
i=1
∂x1 dt

Example 36.
g(x) = x − 1
 
2y
f (y) =
y2
So note that
g : R −→ R, and f : R −→ R2
 
2(x − 1)
[f ◦ g](x) =
(x − 1)2
 
2
D[f ◦ g](x) =
2(x − 1)
9.8. PROPERTIES OF THE DERIVATIVE 119

Now let’s see if we get the same answer doing it the chain rule way:

Dg(x) = 1
 
2
Df (y) =
2y
 
2
Df (g(x))Dg(x) =
2(x − 1)

Example 37.

f (y) = f (y1 , y2 )
 2 
y1 + y2
=
y1 − y1 y2

g(x) = g(x1 , x2 )
 2 
x1 − x2
=
x1 x2
=y

So we note here that both g and f take in two arguments and spit out a
(2 × 1) vector, so we must have

g : R2 −→ R2 , and f : R2 −→ R2

Again, it is very very important to keep track of the dimensions!

D[f ◦ g](x) = Df (g(x)Dg(x)

! !
∂f1 ∂f1 ∂g1 ∂g1
∂y1 ∂y2 ∂x1 ∂x2
= ∂f2 ∂f2 · ∂g2 ∂g2
∂y1 ∂y2 ∂x1 ∂x2
   
2y1 1 2x1 −1
= ·
1 − y2 −y1 x2 x1
120 CHAPTER 9. MULTIVARIABLE CALCULUS

and we know that

y1 = x21 − x2
y2 = x1 x2

So

   
2x21 − 2x2 1 2x1 −1
= ·
1 − x1 x2 x2 − x21 x2 x1
 
4x1 (x21 − x2 ) + x2 x1 − 2(x21 − x2 )
=
2x1 (1 − x1 x2 ) + x2 (x2 − x1 ) x1 (x2 − x21 ) + x1 x2
2

9.9 Gradients and Level Sets

EXAMPLES AND FIGURES GO HERE

Example 38.
f : R3 −→ R
so the graph is in R4 (pretty difficult to draw!), but the graph of the level set
is in R3 .
f (x) = x21 + x22 + x23 = 1

Note that this is just a sphere of radius 1.

So at any point along the sphere there is not just one tangent line to the
point - there are lots of them.

In general, a surface in Rn+1 can be viewed as the solution to a system of

equations. For convenience, represent a point in Rn+1 as a pair (x, y), with
x ∈ Rn and y ∈ R. If F : Rn+1 −→ R, then the set {(x, y) : F (x, y) = 0}
is typically an n dimensional set. We can talk about what it means to be
a tangent to this surface. Certainly the tangent at (x0 , y0 ) should be an n
dimensional linear manifold in Rn+1 that contains (x0 , y0 ). It should also
9.9. GRADIENTS AND LEVEL SETS 121

satisfy the approximation property: if (x, y) is a point on the surface that

is close to (x0 , y0 ), then it should be approximated up to first order by a
point on the tangent. One way to think about this property is to think about
directions on the surface. Consider a function G : R −→ Rn+1 such that
G(0) = (x0 , y0 ) and F ◦ G(t) ≡ 0 for t in a neighborhood of 0. G defines a
curve on the surface through (x0 , y0 ). A direction on the surface at (x0 , y0 )
is just a direction of a curve through (x0 , y0 ) or DG(0). By the chain rule it
follows that
∇F (x0 , y0 ) · DG(0) = 0,
which implies that ∇F (x0 , y0 ) is orthogonal to all of the directions on the
surface. This generates a non-trivial hyperplane provided that DF (x0 , y0 ) 6=
0. We summarize these observations below. (You can decide whether this
is a definition or a theorem.)
Definition 100. Let F : Rn+1 −→ R be differentiable at the point (x0 , y0 ).
Assume that F (x0 , y0 ) = 0 and that DF (x0 , y0 ) 6= 0. The equation of the
hyperplane tangent to the surface F (x, y) = 0 at the point (x0 , y0 ) is

∇F (x0 , y0 ) · ((x, y) − (x0 , y0 )) = 0. (9.2)

Now suppose that f : Rn −→ R is differentiable at x ∈ Rn . Consider the

function F (x, y) = f (x) − y. The surface F (x, y) = 0 is exactly the graph
of f . Hence the tangent to the surface is the tangent to the graph of f . This
means that the formula for the equation of the tangent hyperplane given
above can be used to find the formula for the equation of the tangent to the
graph of a function.
Theorem 55. If f : Rn −→ R is differentiable at x0 ∈ Rn then the vector
∇f (x0 ) is normal (perpendicular) to the tangent vector of the level set of f
at value f (x) at point x ∈ Rn and the equation of the hyperplane tangent
to the graph of f at the point (x0 , f (x0 )) is

∇f (x0 ) · (x − x0 )) = y − y0 .

Proof. Substitute ∇F (x0 , y0 ) = (∇f (x0 , −1) into equation (9.2) and re-
arrange terms.

Example 39. Find the tangent plane to {x | x1 x2 − x23 = 6} ⊂ R3 at b x=

(2, 5, 2).
Notice if you let f (x) = x1 x2 − x23 , then this is a level set of f for value 6.
122 CHAPTER 9. MULTIVARIABLE CALCULUS

∂f ∂f ∂f
∇f (x) = ( , , )
∂x1 ∂x2 ∂x3
= (x2 , x1 , −2x3 )

∇f (b
x) = ∇f (x) |x=(2,5,2)
= (5, 2, −4)

Tangent Plane:

{b
x + y | y · ∇f (b
x) = 0} = {(2, 5, 2) + (y1 , y2 , y3 ) | 5y1 + 2y2 − 4y3 = 0}
= {x | 5x1 − 10 + 2x2 − 10 − 4x3 + 8 = 0}
= {5x1 + 2x2 − 4x3 = 12}

Example 40. Suppose f (x, y, z) = 3x2 + 2xy − z 2 . ∇f (x, y, z) = (6x +

2y, 2x, −2z). Notice that f (2, 1, 3) = 7. The level set of f when f (x, y, z) =
7 is {(x, y, z) : f (x, y, z) = 7}. This set is a (two dimensional) surface in
R3 : It can be written F (x, y, z) = 0 (for F (x, y, z) = f (x, y, z) − 7).
Consequently the equation of the tangent to the level set of f is a (two-
dimensional) hyperplane in R3 . At the point (2, 1, 3), the hyperplane has
normal equal to ∇f (2, 1, 3) = (12, 4, −6). Hence the equation of the hy-
perplane to the level set at (2, 1, 3) is equal to:

(12, 4, −6) · (x − 2, y − 1, z − 3) = 0

or
12x + 4y − 6z = 10.

On the other hand, the graph of f is a three-dimensional subset of R4 :

{(x, y, z, w) : w = f (x, y, z)}. A point on this surface is (2, 1, 3, 7) =
(x, y, z, w). The tangent hyperplane at this point can be written:

w − 7 = ∇f (2, 1, 3) · (x − 2, y − 1, z − 3) = 12x + 4y − 6z − 10

or
12x + 4y − 6z − w = 3.
9.10. HOMOGENEOUS FUNCTIONS 123

9.10 Homogeneous Functions

Definition 101. The function F : Rn −→ R is homogeneous of degree k if
F (λx) = λk F (x) for all λ.

Homogeneity of degree one is weaker than linearity: All linear functions

are homogeneous of degree one, but not conversely. For example, f (x, y) =
√
xy is homogeneous of degree one but not linear.
Theorem 56 (Euler’s Theorem). If F : Rn −→ R be a differential at x and
homogeneous of degree k, then ∇F (x) · x = kF (x).

Proof. Fix x. Consider the function H(λ) = F (λx). This is a composite

function, H(λ) = F ◦ G(λ), where G : R −→ Rn , such that G(λ) = λx.
By the chain rule, DH(λ) = DF (G(λ))DG(λ). If we evaluate this when
λ = 1 we have
DH(1) = ∇F (x) · x. (9.3)
On the other hand, we know from homogeneity that H(λ) = λk x. Differ-
entiating the right hand side of this equation yields DH(λ) = kλk−1 F (λx)
and evaluating when λ = 1 yields

DH(1) = kF (x). (9.4)

Combining equations (9.3) and (9.4) yields the theorem.

In economics functions that are homogeneous of degree zero and one arise
naturally in consumer theory. A cost function depends on the wages you
pay to workers. If all of the wages double, then the cost doubles. This
is homogeneity of degree one. On the other hand, a consumer’s demand
behavior is typically homogeneous of degree zero. Demand is a function
φ(p, w) that gives the consumer’s utility maximizing feasible demand given
prices p and wealth w. The demand is the best affordable consumption for
the consumer. The consumptions x that are affordable satisfy p · x ≤ w (and
possibly another constraint like non-negativity). If p and w are multiplied
by the same factor, λ, then the budget constraint remains unchanged. Hence
the demand function is homogeneous of degree zero.
Euler’s Theorem provides a nice decomposition of a function F . Suppose
that F describes the profit produced by a team of n agents, when agent
i contributes effort xi . How such the team divide the profit it generates?
If F is linear, the answer is easy: If F (x) = p · x, then just give agent
124 CHAPTER 9. MULTIVARIABLE CALCULUS

i pi xi . Here you give each agent a constant “per unit” payment equal to
the marginal contribution of her effort. When you do so, you distribute the
entire surplus (and nothing else). When F is non-linear, it is harder to figure
out the contribution of each agent. The theorem states that if you pay each
agent her marginal contribution (Dei f (x)) per unit, then you distribute the
surplus fully if F is homogeneous of degree one. Otherwise it identifies
alternative ways to distribute the surplus.

9.11 Higher-Order Derivatives

If f : Rn −→ R is differentiable, then its partial derivatives, Dei f (x) =

Di f (x) can also be viewed as functions from Rn to R. So these functions
maybe differentiable. This leads to the idea of higher-order partial deriva-
tives, which are the natural generalization of higher-order derivatives of a
function of one variable. You can imagine taking a derivative with respect
to one variable, then other then the first again, and so on, creating sequences
of the form Dik · · · Di2 Di1 f (x). Fortunately, provided that all of the deriva-
tives are continuous in a neighborhood of x, the order in which you take
partial derivatives does not matter. Hence we will denote an kth derivative
Dik1 ,...,in f (x), where k is the total number of derivatives and ij is the num-
ber of partial derivatives P with respect to the jth argument (so each ij is a
non-negative integer and nj=1 ij = x.
Except for Taylor’s Formula, we will have little interest in third or higher
derivatives. Second derivatives are important. A real-valued function of n
variables will have n2 second derivatives, which we sometimes think of as
terms in a square matrix:
 
∂2f ∂2f
∂x21
(x) ... ∂xn ∂x1
(x)

D2 f (x) =  .. .. 
 . . 

n×n ∂2f ∂2f
∂x1 ∂xn
(x) ... ∂x2
(x)
n

We say f ∈ C k if the kth order partials exist and are continuous.

9.12. TAYLOR APPROXIMATIONS 125

9.12 Taylor Approximations

The multivariable version of Taylor’s Theorem looks intimidating, but it

really is a one-variable theorem. Suppose that you know about a function at
the point a ∈ Rn and you want to approximate the function at x. Consider
the function F (t) = f (xt + a(1 − t)). F : R −→ R and so you can use
one-variable calculus. Also F (1) = x and F (0) = a. So if you want to
know about the function f at x using information about the function f at
a, you really just want to know about the one-variable function F at t = 1
knowing something about F at t = 0. Hence to get the multivariable version
of Taylor’s Theorem, you apply the one variable version of the theorem to
F . You need the chain rule to compute the derivatives of F (in terms of
f ) and there are a lot of these derivatives. This means that the formula is
a bit messy, but it is conceptually no different from the one-dimensional
theorem.

1st Order Approx Consider f : Rn −→ R such that f is differentiable. At

a ∈ Rn the 1st degree Taylor Polynomial of f is

P1 (x) ≡ f (a) + ∇f (a) · (x − a)

The first-order approximation should be familiar. Notice that you have

n “derivatives” (the partials).
If we write f (x) = P1 (x, a) + E2 (x, a) for the first-order approxima-
tion with error of f at x around the point a, then we have

|E2 (x, a)| |f (x) − f (a) − Df (a) · (x − a)|

lim = lim
x→a kx − ak x→a kx − ak
=0

Thus as before, as x → a, E2 converges to 0 faster than x to a.

2nd Order Approx If f ∈ C 2 , the 2nd degree Taylor approximation is

1
f (x) = f (a) + ∇f (a)(x − a) + (x − a)0 D2 f (a)(x − a) +E3 (x, a)
1×n n×1 2 1×n n×n n×1
| {z }
P2 (x,a)
126 CHAPTER 9. MULTIVARIABLE CALCULUS

where
  
∂2f ∂2f
∂x21
(a) ... ∂xn ∂x1
(a) x
1 1 .. ..
(x − a)0 D2 f (a)(x − a) = (x1 − a1 , . . . , xn − an ) · 
  
. . ·
2 1×n n×n n×1 2 
∂2f 2
∂ f

| {z } (a) ... (a) x
1×1 ∂x1 ∂xn ∂x2n

n X
n
1 X ∂ 2f
= (xi − ai ) (a)(xj − aj )
2 i=1 j=1
∂xi ∂xj

Notice that the error term is a quadratic form.

The general form is quite messy. The way to make the mess more pleasant is
to let the notation do all of the work. We define Dhk f to be a kth derivative:
X  k  j
k
Dh f = h11 · · · hjnn D1j1 · · · Dnjn f,
j +···+j =k
j1 · · · jn
1 n

where the summation is taken over all n-tuples of j1 , . . . , jn of non-negative

integers that sum to k and the symbol
 
k k!
= .
j1 · · · jn j1 ! · · · jn !
Here is the general form.
Theorem 57 (Taylor’s Theorem). If f is a real-valued function in C k+1
defined on an open set containing the line segment connecting a to x, then
there exists a point η on the segment such that
f (x) = Pk (x, a) + Ek (x, a)
where Pk (x, a) is a kth order Taylor’s Approximation:
k r
X Dx−a (a)
Pk (x, a) =
r=0
r!
and Ek (x, a) is the error term:
k+1
Dx−a (η)
Ek (x, a) = .
(k + 1)!
Moveover, the error term satisfies:
Ek (a + h, a)
lim =0
h→0 khkk
Chapter 10

Convexity

This section contains some basic information about convex sets and func-
tions in Rn .

10.1 Preliminary: Topological Concepts

Let X ⊂ Rn .

Definition 102. x ∈ Rn is a limit point of X if ∀  > 0

B (x) ∩ X 6= φ

Definition 103. x is an interior point of X if for some  > 0

B ⊂ X

Definition 104. x is a boundary point of X if ∀  > 0 both

B (x) ∩ X 6= φ

and
B (x) ∩ [Rn \X ] 6= φ

127
128 CHAPTER 10. CONVEXITY

Definition 105. X is called open if every x ∈ X is an interior point.

Definition 106. X is called closed if it contains all of its limit points

Reminder: A hyperplane in Rn is of dimension Rn−1 and can be expressed

as
x·p=c
where x ∈ Rn , p 6= 0 is in Rn , and c ∈ R. p is the normal to the hyperplane.

FIGURE GOES HERE

10.2 Convex Sets

Definition 107. X ⊂ Rn is called convex if ∀ x, y ∈ X and ∀ α ∈ [0, 1] we
have
αx + (1 − α)y ∈ X

FIGURE GOES HERE

If you draw a line (or in higher dimensions a hyperplane) between any 2

points in X , then every point on the line is also in X .

Theorem 58. Given a nonempty, closed, convex set X ⊂ Rn and x ∈ Rn ,

/ X . There exists p ∈ Rn , p 6= 0, and c ∈ R such that
x∈

X ⊂ {y | y · p ≥ c}

and
x·p<c
10.2. CONVEX SETS 129

That is, y · p = c defines a separating hyperplane for X .

FIGURE GOES HERE

Proof. Consider the problem of minimizing the distance between x and the
set X . That is, find y∗ to solve:

min kx − yk subject to y ∈ X . (10.1)

The norm is a continuous function. While X is not necessarily bounded,

since it is nonempty, we can find some element z ∈ X . Without loss of
generality we can replace X in (10.1) by {y ∈ X : ky − xk ≤ kz − xk}.
This set is compact because X is closed. Hence there is a solution y∗ to
(10.1). Let p = y∗ − x. Since x ∈ / X , p 6= 0. Let c = y∗ · p. Since
2
c − p · x = kpk , c > p · x. To complete the proof we must show that if

y ∈ X , then y · p ≥ c.

This inequality is equivalent to

(y − y∗ ) · (y∗ − x) ≥ 0. (10.2)

Since X is convex and y∗ is defined to solve (10.1), it must be that kty + (1 − t)y∗ − xk2
is minimized when t = 0 so that the derivative of kty + (1 − t)y∗ − xk2
is non-negative at t = 0. Differentiation and simplifying yields inequal-
ity (10.2).

Notice that without loss of generality you can normalize the normal to the
separating hyperplane. That is, you can assume that kpk = 1.
You can refine the separating hyperplane theorem in two ways. First, if x is
an element of the boundary of X , then you can approximate x by a sequence
xk such that each xk ∈ / X . This yields a sequence of pk , which can be taken
to be unit vectors, that satisfy the conclusion of the theorem. A subsequence
of the pk must converge. The limit point p∗ will satisfy the conclusion of
the theorem (except we can only guarantee that c ≥ p∗ · x rather than the
strict equality above). Second, one can check that the closure of any convex
130 CHAPTER 10. CONVEXITY

set is convex. Therefore, given a convex set X and a point x not in the
interior of the set, we can separate x from the closure of X . Taking these
considerations into account we have the following version of the separating
hyperplane theorem.

Theorem 59. Given a convex set X ⊂ Rn and x ∈ Rn , x. If x is not in the

interior of X , then there exists p ∈ Rn , p 6= 0, and c ∈ R such that

X ⊂ {y | y · p ≥ c}

and
x·p≤c

In the typical economic application, the separating hyperplane’s normal has

the interpretation of prices and the separation property states that a particu-
lar vector costs more than vectors in a consumption set.

10.3 Quasi-Concave and Quasi-Convex Functions

Definition 108. A function f : Rn −→ R is called quasiconcave if {x | f (x) ≥ a}
is convex for all a ∈ R. In other words the upper contour set, as defined in
Definition 86, is a convex set.

Example 41. f (x) = −x21 − x22

the upper contour set is the inner shaded part of the sphere.

FIGURE GOES HERE

Definition 109. A function f : Rn −→ R is called quasiconvex if {x | f (x) ≤ a}

is convex for all a ∈ R. In other words the lower contour set, as defined in
Definition 86, is a convex set.
10.3. QUASI-CONCAVE AND QUASI-CONVEX FUNCTIONS 131

Note: In the one-variable setting, we said that a function was quasiconcave

if

for all x, y, and λ ∈ (0, 1), f (λx + (1 − λ)y) ≥ min{f (x), f (y)}. (10.3)
To see that this definition is equivalent to Definition 108 note first that if
a = min{f (x), f (y)}, then Definition 108 implies (10.3). Conversely, if
the condition in the Definition 108 fails, then there exists a, x, y such that
f (x), f (y) ≥ a but f (λx + (1 − λ)y) < a. Plainly condition (10.3) fails for
these values of λ, x, and y.
Quasiconcavity and Quasiconvexity are global properties of a function. Un-
like continuity, differentiability, concavity and convexity (of functions), they
cannot be defined at a point.

Theorem 60. If f is concave and x∗ is a local maximizer of f , then x∗ is a

global maximizer.

Proof. (by contradiction)

Suppose x∗ is not a global maximizer of f . Then there is a point b
x such that
x) > f (x∗ )
f (b
but then we would have that
f (αx∗ + (1 − α)b
x) ≥ αf (x∗ ) + (1 − α)f (b
x)
And this contradicts that x∗ is a local max.

10.3.1 How to check if a function f is quasiconcave or not

There are a few steps to follow.
(a) First solve for the level sets and graph a selection of them
(b) Decide by inspection which side of the plotted level set is the upper
contour set and which side is the lower contour set
(c) Are the inspected upper and lower contour sets convex?
Note we are looking for nice shaped graphs. If the graph bends and twists
alot then it is not quasiconvex or quasiconcave. We are looking for nice
graphs like circles or parabolas.
132 CHAPTER 10. CONVEXITY

10.3.2 Relationship between Concavity and Quasiconcav-

ity
Suppose f : Rn −→ R is concave ∀ x, y ∈ Rn . Then ∀ α ∈ [0, 1]

f (αx + (1 − α)y) ≥ αf (x) + (1 − α)f (y)

FIGURE GOES HERE

Points on the boundary of this level set (convex) give you values of the
function equal to some constant c. Points inside the shaded region give
you values for the function that are greater than or equal to c. We want to
examine the set { z | f (z) ≥ c}. Consider the points x, y ∈ { z | f (z) ≥ c},

=⇒ f (x) ≥ c and f (y) ≥ c

and this means ∀ α ∈ [0, 1] that

f (αx + (1 − α)y) ≥ αf (x) + (1 − α)f (y)

≥ αc + (1 − α)c
=c

=⇒ αx + (1 − α)y ∈ { z | f (z) ≥ c}
This is really providing the intuition for the following Theorem.

Theorem 61. If f : Rn −→ R is concave, then f is quasiconcave.

Similarly if f is convex, then f is quasiconvex.

Note that concavity implies quasiconcavity but that the opposite is not true.
A quasiconcave function may be concave, but certainly does not have to be.

Example 42. ex is quasiconcave but not concave. In fact it is also convex

and quasiconvex!

The easiest way to see this is just to note that f (x) = ex is just a func-
tion of one variable and thus is easy to plot. But we plot the function not
the level sets (Note level sets are points). Since the exponential function
10.3. QUASI-CONCAVE AND QUASI-CONVEX FUNCTIONS 133

is increasing, upper contour sets are of the form [a, ∞). So the function is
quasiconcave. The function is quasiconvex too!

More generally, we have:

Theorem 62. Any increasing or decreasing function is both quasiconvex

and quasiconcave.

10.3.3 Ordinal “vs” Cardinal

Definition 110. We say that the property of functions ∗ is ordinal if f : Rn −→

R has property (∗) and g : R −→ R is strictly increasing

=⇒ g ◦ f has property(∗)

In other words it doesn’t matter how much greater the value f (x) is than
f (y), it just matters that it is greater.
“Doing better” is typically an ordinary property. When you finish a race,
you receive an ordinal ranking (“fifth place”), which would not change if
you applied a monotonic transformation to the times. There are times when
you care about more than ordinal ranking. Saying someone is taller than
someone else is ordinal. Saying some is ‘a lot’ taller is not.

Definition 111. We say that the property of functions ∗ is cardinal if f : Rn −→

R has property (∗) and g : R −→ R is strictly increasing affine function,
then
=⇒ g ◦ f has property(∗)

There concepts are relevant because concavity is not ordinal.

f : (0, ∞) −→ R
1
f (x) = x 2
So if
x > y =⇒ f (x) > f (y)
134 CHAPTER 10. CONVEXITY

This is a concave function. Now suppose

g(y) = y 4

=⇒ g ◦ f (x) = x2
And again we have that

x > y =⇒ g ◦ f (x) > g ◦ f (y)

This is it now a convex function.

On the other hand, quasiconcavity is ordinal. Suppose f : Rn −→ R is
quasiconcave. That is, the set { x | f (x) ≥ c} is convex ∀c ∈ R.
Now take a strictly increasing function

g : R −→ R

So the following 2 sets must be equivalent

{ x | g ◦ f (x) ≥ c} ⇐⇒ x | f (x) ≥ g −1 (c)



and so the upper contour set remains convex =⇒ quasiconcavity is pre-

served.
Example 43. A good example of an ordinal concept is utility. A utility
function is a function which measures the level of happiness (or utility) that
a person gets out of the argument. (In 200A you will see all the conditions
necessary for the existence of such a function - but for now we will just
assume that it exists).
So suppose we have a guy Joel, and the only thing in the world that gives
Joel any joy are apples. Then we can write Joel’s utility function as

UJ : R++ −→ R

So the input into Joel’s utility function is just a positive real number (i.e. the
number of apples he ate) and the output is also just a number (the number
of “utils” that this number of apples gave him). But the important point is
that we do not attach any psychological significance to the value attained
by the function, because Joel only cares about the ranking of how much joy
(how many utils) he gets from different quantities of apples.

Suppose
1
UJ (x) = x 2
10.3. QUASI-CONCAVE AND QUASI-CONVEX FUNCTIONS 135

Then we have that

1 1
UJ (16) = 16 2 = 4 and UJ (9) = 9 2 = 3

So we have that
UJ (16) > UJ (9)
The fact that it is 1 “util” greater is irrelevant - the important thing is that
16 apples gives Joel more utils than 9 apples. So as you would expect he
prefers 16 apples to 9!! He ranks the bundles of goods (in this case apples)
ordinally.

While it seems obvious that Joel prefers 16 apples to 9 apples, it does get
more tricky when Joel’s utility function depends on two commodities: say
apples and bananas. But you will see all this in 200A.

Example 44. Suppose f : Rn −→ R is quasiconcave. That is, the set

{ x | f (x) ≥ c} is convex ∀c ∈ R.
Now take a strictly increasing function

g : R −→ R

Note that since g is strictly increasing, g is invertible.

So the following 2 sets must be equivalent

{ x | g ◦ f (x) ≥ c} ⇐⇒ x | f (x) ≥ g −1 (c)



and so the upper contour set remains convex =⇒ quasiconcavity is pre-

served!!!

conclusion: quasiconcavity and quasiconvexity are ordinal concepts.

Another more technical way to distinguish the two is to say that an ordinal
property is preserved under any monotonic transformation while a cardinal
property is preserved only under a positive affine transformation. Ordi-
nal properties rank points in the domain while cardinal properties rank the
cartesian product of points in the domain.
136 CHAPTER 10. CONVEXITY
Chapter 11

Unconstrained Extrema of
Real-Valued Functions

11.1 Definitions
The following are natural generalizations of the one-variable concepts.

Definition 112. Take f : Rn −→ R.

x∗ is a local maximizer ⇐⇒ ∃ δ > 0 such that ∀ x ∈ Bδ (x∗ ),

f (x) ≤ f (x∗ )

x∗ is a local minimizer ⇐⇒ ∃ δ > 0 such that ∀ x ∈ Bδ (x∗ ),

f (x) ≥ f (x∗ )

x∗ is a global maximizer ⇐⇒ ∀ x ∈ Rn , we have

f (x) ≤ f (x∗ )

x∗ is a global minimizer ⇐⇒ ∀ x ∈ Rn , we have

f (x) ≥ f (x∗ )

137
138CHAPTER 11. UNCONSTRAINED EXTREMA OF REAL-VALUED FUNCTIONS

11.2 First-Order Conditions

Theorem 63 (First Order Conditions). If f is differentiable at x∗ , and x∗ is
a local maximizer or minimizer then

Df (x) = 0.

Consequently,
∂f ∗
(x ) = 0,
∂xi
∀ i = 1, 2, . . . , n.

FIGURE GOES HERE

Proof. We define h : R −→ R by

h(t) ≡ f (x∗ + tv)

for any v ∈ Rn , t ∈ R.
6 0). We have (for a local maximum)
Fix a direction v (kvk =

f (x∗ ) ≥ f (x),

∀x ∈ Bδ (x∗ ), for some δ > 0.

In particular for t small (t < δ kvk) we have

f (x∗ + tv) = h(t)

≤ f (x∗ )

Thus, h is maximized locally by t∗ = 0.

Our F.O.C. from the R −→ R case

=⇒ h0 (0) = 0

So by the chain rule

=⇒ ∇f (x∗ ) · v = 0
And since this must hold for every v ∈ Rn , this implies that

∇f (x∗ ) = 0.

Since Df is equal to the matrix of partial derivatives, the result follows.

11.3. SECOND ORDER CONDITIONS 139

Note We call x∗ satisfying Df (x∗ ) = 0 a critical point.

The theorem mimics the one variable theorem. The proof works like this.
If x∗ is a local maximum, then the one variable function you obtain by
restricting x to move along a fixed line through x∗ (in the direction v) also
must have a local maximum. Hence that function of one variable (f (x∗ +tv)
treated as a function of t) must have derivative zero when t = 0. The
derivative of this function is simply the directional derivative of f in the
direction v. If all of the directional derivatives are zero and the function is
differentiable, then the derivative must be zero.
Just like the one-variable case, the first-derivative test cannot distinguish be-
tween local minima and and local maxima, but an examination of the proof
tells you that at local maxima derivatives decrease in the neighborhood of a
critical point. Just like the one-variable case, critical points may fail to be
minima or maxima. In the one variable case, this problem arises essentially
for one reason: a function decrease if you reduce x (suggesting a local max-
imum) and increase if you increase x (suggesting a local minimum). In the
many variable case, this behavior could happen in any direction. Moreover,
it could be that the function restricted to direction has a local maximum, but
it has a local minimum with respect to another direction. At this point, the
central insight is that it is much less likely than a critical point of a multi-
variable function is a local extremum than in the one variable case.

We want necessary and sufficient conditions.

11.3 Second Order Conditions

For a critical point x∗ we have by Taylor’s Theorem that
1 t 2
f (x∗ + h) − f (x∗ ) 2
h D f (z)h
2 =
khk khk2

We could write h = tv, where kvk = 1. So

1 2 t 2
f (x∗ + tv) − f (x∗ ) 2
t v D f (z)v
2
=
t t2
1
= vt D2 f (z)v
2
140CHAPTER 11. UNCONSTRAINED EXTREMA OF REAL-VALUED FUNCTIONS

for z ∈ Bkhk (x∗ )

The generalization of second-order conditions follows from this expression.

If x∗ is a critical point and the quadratic form vt D2 f (z)v is negative definite,
then continuity implies that vt D2 f (x∗ )v is negative definite and hence x∗ is
a local maximum. Conversely (almost conversely), if x∗ is a local maxi-
mum, then it must be that vt D2 f (x∗ )v is negative semi definite. Finally
note that if the quadratic form vt D2 f (x∗ )v is indefinite, then the critical
point x∗ cannot be either a local maximum or a local minimum because
there are directions in which the function decreases near x∗ and directions
in which the function increases near x∗ .

FIGURE GOES HERE

What follows is an informal summary.

11.3.1 S.O. Sufficient Conditions

If f is twice continuously differentiable and x∗ is a critical point of f , then:

(a) x∗ is a local minimizer whenever the quadratic form vt D2 f (x∗ )v is a

positive definite.
(b) x∗ is a local maximizer whenever the quadratic form vt D2 f (x∗ )v is
negative definite.

11.3.2 S.O. Necessary Conditions

If D2 f (x∗ ) exists, then:

(a) if x∗ is a local maximizer, then vt D2 f (x∗ )v is a negative semi-definite

quadratic form.
(b) if x∗ is a local minimizer, then vt D2 f (x∗ )vt is a positive semi-definite
quadratic form.
11.3. SECOND ORDER CONDITIONS 141

A consequence of the necessary conditions is that when the second-derivative

matrix gives rise to an indefinite quadratic form a critical point cannot be
either a maximum or a minimum.

Also recall our definitions of concave up and concave down.

We also have that

• D2 f (x∗ ) positive definite everywhere ⇔ f concave up everywhere

• D2 f (x∗ ) negative definite everywhere ⇔ f concave down everywhere

Example 45. Let f : R2 −→ R be

f (x, y) = 4x3 + y 2 − 6xy + 6x

Now (x∗ , y ∗ ) is our max. We have then as usual that

∇f (x∗ , y ∗ ) = 0

∂f
(x, y) = 12x2 − 6y + 6
∂x
=0

∂f
(x, y) = 2y − 6x
∂y
=0

from the second equation we get

2y = 6x

And combining this with the first equation we get

12x2 − 18x + 6 = 0
⇒ 2x2 − 6x + 1 = 0
⇒ (2x − 1)(x − 1) = 0
142CHAPTER 11. UNCONSTRAINED EXTREMA OF REAL-VALUED FUNCTIONS

gives us
1
x= 2
or x=1
1 3
x= 2
⇒ y= 2
x=1 ⇒ y=3
So we have 2 critical points
( 21 , 32 ), (1, 3)

Df (x) = ∇f (x)
 
∂f ∂f
= ,
∂x ∂y
 ∂f 
0 ∂x
(Df ) = ∂f
∂y

D f = D(Df )0
2

!
∂2f ∂2f
∂x2
D2 f = ∂2f
∂x∂y
∂2f
∂y∂x ∂x2
 
24x −6
=
−6 2
So  
2 12 −6
D f ( 12 , 32 )
=
−6 2
And we note that the first entry in this matrix (i.e. the a entry) is positive! It
can also be seen that
 2 1 3 
D f ( , ) = −12 < 0
2 2

And from our rule for quadratic forms this gives us an indefinite quadratic
form. So the point ( 12 , 23 ) is neither a maximizer nor a minimizer.
So  
2 24 −6
D f (1, 3) =
−6 2
And we note that the first entry in this matrix (i.e. the a entry) is positive! It
can also be seen that  2 1 3 
D f ( , ) = 12 > 0
2 2
And from our rule for quadratic forms this gives us an positive definite
quadratic form. So the point (1, 3) is a local minimizer.
Chapter 12

Invertibility and Implicit

Function Theorem

The standard motto is: Whatever you know about linear functions is true
locally about differentiable functions. This section discusses two useful
properties that can be understood in these terms.

12.1 Inverse Functions

First, let’s review invertibility for functions of one variable. In the linear
case, f (x) = ax, so the function is invertible if a 6= 0. Notice that if a linear
function is invertible at a point, it is invertible everywhere. More generally,
we have this definition and result:

Definition 113. We say the function f : X −→ Y is one-to-one if

f (x) = f (x0 ) =⇒ x = x0

We have already defined f −1 (S) = {x ∈ X | f (x) ∈ S} for S ⊂ Y.

Now let’s consider f −1 (y) for y ∈ Y. Is f −1 a function?

If f is one-to-one, then f −1 : f (X ) −→ X is a function.

143
144CHAPTER 12. INVERTIBILITY AND IMPLICIT FUNCTION THEOREM

FIGURE GOES HERE

f is generally not invertible as the inverse is not one-to-one. But in the

neighborhood (circle around a point x0 ), it may be strictly increasing so it
is one-to-one and therefore invertible.

Theorem 64. If f : R −→ R is C 1 and

f 0 (x0 ) 6= 0,

then ∃  > 0 such that f is strictly monotone on the open interval (x0 −
, x0 + ).

Note this is just saying that if f 0 (x0 ) is positive and f 0 is continuous, then it
is remains positive over an interval.
So f is locally invertible at x0 , then we can define g on (x0 − , x0 + ) such
that
g(f (x)) = x

That is, in the one-variable case, linear functions with (constant) derivative
not equal to zero are invertible. Differentiable functions with derivative not
equal to zero at a point are invertible locally. For one variable functions,
if the derivative is always non zero, then the inverse can be defined on the
entire range of the function. When you move from functions from Rn to
itself, you can ask whether inverse functions exist. Linear functions can be
represented as multiplication by a square matrix. Invertibility of the func-
tion is equivalent to inverting the matrix. So a linear function is invertible
(globally) if its matrix representation is invertible.

Definition 114. The function f : Rn −→ Rn is locally invertible at x0 if

there is a δ > 0 and a function g : Bδ (f (x0 )) −→ Rn such that f ◦g(y) ≡ y
for y ∈ Bδ (f (x0 )) and g ◦ f (x) ≡ x for x ∈ Bδ (x0 ).

The definition asserts the existence of an inverse defined in a (possibly

small) set containing f (x0 ).
12.1. INVERSE FUNCTIONS 145

Theorem 65. If f : Rn −→ Rn is differentiable at x0 and Df (x0 ) is invert-

ible, then f is locally invertible at x0 . Moreover, the inverse function, g is
differentiable at f (x0 ) and Dg(f (x0 )) = (Df (x0 ))−1 .

The theorem asserts that if the linear approximation of a function is invert-

ible, then the function is invertible locally. In contrast to the one variable
case, the assumption that Df is globally invertible does not imply the exis-
tence of a global inverse.
Recall Theorem 31 that says
1
g 0 (y0 ) =
f 0 (x 0)

so the formula for the derivative of the inverse generalizes the one-variable
formula.
The proof of the inverse function theorem is hard. One standard technique
involves methods that are sometimes used in economics, but the details are
fairly intricate and not worth our time.
The problem of finding an inverse suggests a more general problem. Sup-
pose that you have a function G : Rn+m −→ Rm . Let x ∈ Rn and y ∈ Rm .
We might be interested in whether we can solve the system of equations:
G(x, y) = 0. This is a system of m equations in n + m variables. You
might hope therefore that for every choice of y you could solve the equa-
tion. That is, you might search for a solution to the equation that gives x as
a function of y. The problem of finding an inverse is really a special case
where n = m and G(x, y) = f (x) − y.
The general case is an important problem in economics. In the typical ap-
plication, the system of equations characterize an economic equilibrium.
Maybe they are the equations that determine market clearing price. Maybe
they are the first-order conditions that characterize the solution to an opti-
mization problem. The x variables are parameters. You “solve” a model
for a fixed value of the parameters and you want to know what the solution
to the problem is when the parameters change by a little bit. The implicit
function theorem is a tool for analyzing the problem. This theorem says that
(under a certain condition), if you can solve the system at a give x0 , then
you can solve the system in a neighborhood of x0 . Furthermore, it gives you
expressions for the derivatives of the solution function.
Why call this the implicit function theorem? Life would be great if you
could write down the system of equations and solve them to get an explicit
146CHAPTER 12. INVERTIBILITY AND IMPLICIT FUNCTION THEOREM

representation of the solution function. If you can do this, then you can ex-
hibit the solution (so existence is not problematic) and you can differentiate
it (so you do not need a separate formula for derivatives). In practice, you
may not be able to find an explicit form for the solution. The theorem is the
next best thing.
We will try to illustrate the ideas with some simple examples.

12.2 Implicit Functions

f : R2 −→ R
Suppose f (x, z) = 0 is an identity relating x and z.

So we are looking for how z depends on x at this value of f which is at-

tained with a certain value of x. This is really the point!

Example 46.
f (x, z) = x3 − z = 0

FIGURE GOES HERE

So there is a pair x and z which fits this identity (actually in this exam-
ple there are an infinite number of them)

In this example it is easy to pull out the z to one side and solve for it in
terms of x but supposing you had a much more complicated example like

x2 z − z 2 + sin x ln z + cos x = 0,

then this is not as easy to just find z in terms of x.

The important point is that even this messy function is still expressing z
in terms of x, although it may be difficult solve for this explicitly - but there
are ways around this problem.
12.2. IMPLICIT FUNCTIONS 147

Not only might x0 correspond to a few z-values, but also z0 might corre-
spond to a few x-values.
But if we examine (x0 , z0 ) in a small neighborhood where the function is
increasing...

FIGURE GOES HERE

Using the picture, we can see that increasing the x value slightly, the func-
tion no longer takes the value of zero, so you can see that we are no longer
on the level set!

∂f ∂f
4x = −4z
∂x ∂z

∂f
4z ∂x
= − ∂f
4x ∂z
∂f
∂z ∂x
= − ∂f
∂x ∂z

So if f gives us a function

g : (x0 − , x0 + ) −→ (z0 − , z0 + )

such that
f (x, g(x)) = 0,
∀ x ∈ (x0 − , x0 + )
then we can define
h : (x0 − , x0 + ) −→ R
by
h(x) = f (x, g(x))

So the point of this is that if h(x) = 0 for any x, then h0 (x) = 0 for any
x. But the clever part is that h0 (x) can be written as a function of f 0 (x) and
148CHAPTER 12. INVERTIBILITY AND IMPLICIT FUNCTION THEOREM

g 0 (x) which are not necessarily zero.

If f and g are differentiable, we calculate

   
x x
y= , G(x) =
z g(x)
h(x) = [f ◦ G](x)

h0 (x) = Df (x0 , z0 )DG(x0 )

  
∂f ∂f 1
= (x0 , z0 ) (x0 , z0 )
∂x ∂z g 0 (x0 )
∂f ∂f
= (x0 , z0 ) + (x0 , z0 )g 0 (x0 )
∂x ∂z
=0
This gives us
∂f
0 ∂x
(x0 , z0 )
g (x0 ) = − ∂f
∂z
(x0 , z0 )

Now we will look at the more general case:

f : R × Rm −→ Rm
 
f1 (x, z)
 f2 (x, z) 
f (x, z) = 
 
.. 
 . 
fm (x, z)
 
0
 0 
=
 
.. 
 . 
0
=0
where x ∈ R, and z ∈ Rm .
And we have
g : R −→ Rm
12.2. IMPLICIT FUNCTIONS 149

z = g(x)

Theorem 66. Suppose

f : R × Rm −→ Rm

is C 1 and write f (x, z) where x ∈ R and z ∈ Rm .

Assume
 ∂f
 1 . . . ∂f1


 ∂z1 ∂zm 
|Dz f (x0 , z0 )| =  ... ..
 
 ∂fm . 
∂fm


∂z1
. . . ∂zm 
6= 0

and
f (x0 , z0 ) = 0.
There exists a neighborhood of (x0 , z0 ) and a function g : R −→ Rm defined
on the neighborhood of x0 , such that z = g(x) uniquely solves f (x, z) = 0
on this neighborhood.
Furthermore the derivatives of g are given by

Dg(x0 ) = −[Dz f (x0 , z0 )]−1 Dx f (x0 , z0 )

m×1 m×m m×1

We said that the inverse function theorem was too hard to prove here. Since
the implicit function theorem is a generalization, it too must be too hard
to prove. It turns out that the techniques one develops to prove the inverse
function theorem can be used to prove the implicit function theorem, so the
proof is not much harder. It also is that case that the hard thing to prove
is the existence of the function g that gives z in terms of x. If you assume
that this function exists, then computing the derivatives of g is a simple
application of the chain rule. The following proof describes this argument.

Proof. So we have
f (x, g(x)) = 0
And we define
H(x) ≡ f (x, g(x))
150CHAPTER 12. INVERTIBILITY AND IMPLICIT FUNCTION THEOREM

And thus

Dx H(x) = Dx f (x0 , z0 ) + Dz f (x0 , z0 )Dx g(x0 )

=0

=⇒ Dz f (x0 , z0 )Dx g(x0 ) = −Dx f (x0 , z0 )

Multiply both sides by the inverse:

[Dz f (x0 , z0 )]−1 · [Dz f (x0 , z0 )] ·Dx g(x0 ) = −[Dz f (x0 , z0 )]−1 Dx F (x0 , z0 )
| {z }
=Im

=⇒ Dx g(x0 ) = −[Dz f (x0 , z0 )]−1 Dx F (x0 , z0 )

The implicit function theorem thus gives you a guarantee that you can (lo-
cally) solve a system of equations in terms of parameters. As before, the
theorem really is a local version of a result about linear systems. The theo-
rem tells you what happens if you have one parameter. The same theorem
holds when you have many parameters x ∈ Rn rather than x ∈ R.
Theorem 67. Suppose

f : Rn × Rm −→ Rm

is C 1 and write F (x, z) where x ∈ Rn and z ∈ Rm .

Assume
 ∂f
 1 . . . ∂f1


 ∂z1 ∂zm 
 .. ..
|Dz f (x0 , z0 )| =  .

 ∂fm . 
∂f m


∂z1
. . . ∂zm 
6= 0

and
f (x0 , z0 ) = 0.
There exists a neighborhood of (x0 , z0 ) and a function g : Rn −→ Rm
defined on the neighborhood of x0 , such that z = g(x) uniquely solves
f (x, z) = 0 on this neighborhood.
Furthermore the derivatives of g are given by implicit differentiation (use
chain rule)
Dg(x0 ) = −[Dz f (x0 , z0 )]−1 Dx f (x0 , z0 )
m×n m×m m×n
12.3. EXAMPLES 151

Verifying that this is the correct formula for the derivative is just the chain
rule.
Comments:

(a) It is useful to keep track of the dimensions of the various matrices.

Keep in mind the intuitive idea that you usually need exactly m vari-
ables to solve m equations is helpful. This means that if the domain
has n extra dimensions, typically you will have n parameters – the
solution function will go from Rn into Rm .
(b) The implicit function theorem proves that a system of equations has
a solution if you already know that a solution exists at a point. That
is, the theorem states that if you can solve the system once, then you
can solve it locally. The theorem does not guarantee existence of a
solution. (This is where the general case is much less informative
than the linear case, where the invertibility assumption in the theorem
would be sufficient to show existence.)
(c) The theorem provides an explicit formula for the derivatives of the
implicit function. One could memorize the formula and apply it when
needed. That would be wrong. It turns out that you can compute the
derivatives of the implicit function by “implicitly differentiating” the
system of equations and solving the resulting equations by explicitly.
An example follows.

12.3 Examples
Here is a simple economic example of a comparative-statics computation.
A monopolist produces a single output to be sold in a single market. The
cost to produce q units is C(q) = q + .5q 2 dollars and the monopolist can
5
sell q units for the price of P (q) = 4 − q6 dollars per unit. The monopolist
must pay a tax of one dollar per unit sold.

(a) Show that the output q ∗ = 1 that maximizes profit (revenue minus tax
payments minus production cost).
(b) How does the monopolist’s output change when the tax rate changes
by a small amount?

The monopolist picks q to maximize:

152CHAPTER 12. INVERTIBILITY AND IMPLICIT FUNCTION THEOREM

q5
q(4 − ) − tq − q − .5q 2 .
6

The first-order condition is

q5 + q − 3 + t = 0

and the second derivative of the objective function is −5q 4 − 1 < 0, so there
is at most one solution to this equation, and the solution must be a (global)
maximum. Plug in q = 1 to see that this value does satisfy the first-order
condition when t = 1. Next the question asks you to see how the solution
q(t) to:
q5 + q − 3 + t = 0
varies as a function of t when t is close to one. We know that q(1) = 1
satisfies the equation. We also know that the left-hand side of the equation
is increasing in q, so the condition of the implicit function theorem holds.
Differentiation yields:
1
q 0 (t) = − 4 . (12.1)
5q + 1
In particular, q 0 (1) = − 61 .
In order to obtain the equation for q 0 (t) (12.1), you could use the general
formula or you could differentiate the identity:

q(t)5 + q(t) − 3 + t ≡ 0

with respect to one variable (t) to obtain

5q(t)q 0 (t) + q 0 (t) + 1 = 0,

and solve for q 0 (t). Notice that the equation is linear in q 0 . This technique
of “implicit differentiation” is fully general. In the example you have n =
m = 1 so there is just one equation and one derivative to find. In general,
you will have an identity in n variables and m equations. If you differentiate
each of the equations with respect to a fixed parameter, you will get m linear
equations for the derivatives of the m implicit functions with respect to that
variable. The system will have a solution if the invertibility condition in the
theorem is true.
12.4. ENVELOPE THEOREM FOR UNCONSTRAINED OPTIMIZATION153

12.4 Envelope Theorem for Unconstrained Op-

timization
We are interested in
V (a) = max f (x, a)
x∈X
n n
for a ∈ R, x ∈ R , X ⊂ R open.

We will not concern ourselves with conditions under which the maximum
exists (we cannot apply the standard existence result when X is open), but
just focus on examples where the max value of f does exist i.e. V (a) exists,
and we will ask the question what happens to V (a) as a changes. i.e. What
is V 0 (a)? Suppose that we can find a function g : R −→ Rn such that g(a)
maximizes f (x, a). We have V (a) = f (g(a), a). That is, the value function
V is a real-valued function of a real variable. We can compute its derivative
using the chain rule (assuming that f and g are differentiable). The implicit
function theorem gives us a sufficient condition for g to be differentiable
and a formula for the derivative provided that solutions to the optimization
problem are characterized as solution to the first order condition. That is,
if x∗ solves maxx∈X f (x, a∗ ) if and only if Dx f (x∗ , a∗ ) = 0, then x∗ is im-
plicitly defined as a solution to a system of equations. Applying the chain
rule mechanically we have

V 0 (a∗ ) = Dx f (g(a∗ ), a∗ )Dg(a∗ ) + Da f (g(a∗ ), a∗ ).

1×1 1×n n×1 1×1

The implicit function theorem tells us when Dg(a∗ ) exists and gives us a
formula for the derivative. In order to evaluate V 0 , however, we only need
to know that the derivative exists. This is because at an interior solution to
the optimization problem Dx f (g(a∗ ), a∗ ) = 0. It follows that the change in
the value function is given by V 0 (a∗ ) = Da f (x∗ , a∗ ), where x∗ = g(a∗ ) is
the solution to the optimization problem at a∗ .
The one condition that needs to be checked is the non-singularity condition
in the statement of the implicit function theorem. Here the condition is
that the matrix of second derivatives of f with respect to the components
of x is non-singular. We often make this assumption (in fact, we assume
that the matrix is negative definite), to guarantee that solutions to first-order
conditions characterize a maximum.
The result that the derivative of the value function depends only on the par-
tial derivative of the objective function with respect to the parameter maybe
154CHAPTER 12. INVERTIBILITY AND IMPLICIT FUNCTION THEOREM

somewhat surprising. It is called the envelope theorem (you will see other
flavors of envelope theorem).
Suppose that f is a profit function and a is a technological parameter. Cur-
rently a∗ describes the state of the technology. The technology shifts. You
shift your production accordingly (optimizing by selecting x = g(a)). Your
profits shift. How can you tell whether the technological change makes you
better off? Changing the technology has two effects: a direct one that is
captured by Da f . For example, if Da f (x∗ , a∗ ) < 0 this means that increas-
ing a leads to a decrease in profits if you hold x∗ fixed. The other effect
is indirect: you adjust x∗ to be optimal with respect to the new technology.
The envelope theorem says that locally the first effect is the only thing that
matters. If the change in a makes the technology less profitable for fixed
x∗ , then you will not be able to change x∗ to reverse the negative effect.
The reason is, loosely, that when you optimize with respect to x and change
the technology, then x∗ is still nearly optimal for the new problem. Up to
“first order” you do not gain from changing it. When Da f (x∗ , a∗ ) 6= 0,
changing a leads to a first-order change in profits that dominates the effect
of reoptimizing.
The result is called the envelope theorem because (geometrically) the result
says that V (a), the value function, is tangent to a family of functions f (x, a)
when x = g(a). On the other hand, V (a) ≥ f (x, a) for all a, so the V curve
looks like an “upper envelope” to the f curves.
For a presentation of value functions that is both more simple and coherent
than this, you can read chapters 26 and 27 of Varian. It is a nice treatment
since it does lots of simpler examples of functions of only 1 and 2 variables,
as opposed to the general n variable case above. Really only chapter 27
deals with value functions but chapter 26 is good background reading too.
Chapter 13

Constrained Optimization

Economic theory is all about solving constrained optimization problems.

Consumers maximize utility subject to a budget constraint. Firms mini-
mize cost or maximize profit subject to their outputs being technologically
feasible. Governments may maximize social welfare subject to constraints
on consumer’s independence. The theory of unconstrained optimization is
easy because at an interior optima it must be that directional derivatives in
all directions are zero. This follows from a simple argument. If you are at
a maximum, then moving “forward” in any direction must not increase the
objective function. So a directional derivative must be less than or equal to
zero. However, moving “backward” in that direction must also lower the
function’s value, so the directional derivative must be equal to zero. The
same insight applies in the constrained case, but since movement in some
directions might be ruled out by constraints, first-order conditions come in
the form a set of inequalities that hold for all directions that are not ruled
out by constraints.

13.1 Equality Constraints

The basic problem is:

max f (x) (13.1)

x∈S (13.2)
Here f : Rn −→ R. The function f in (13.1) is called the objective function.
We study maximization problems, but since minimizing a function h is the

155
156 CHAPTER 13. CONSTRAINED OPTIMIZATION

same as maximizing −h, the theory covers minimization problems as well.

(13.2) is the constraint set.
Recall that a solution to the problem is x∗ such that (a) x∗ ∈ S and (b) if
x ∈ S, then f (x∗ ) ≥ f (x).
Calculus does not provide conditions for the existence of an optima (real
analysis does, because continuous real-valued functions on compact sets
attain their maxima). We do not discuss existence in this section. Instead
we concentrate on characterizing optima in terms of equations. When S is
an open set, any maxima must be interior. Hence Df = 0. In general, we
must modify this condition.
We begin with the study of equality constrained maximization. That is,
assume that there are a set of functions gi : Rn −→ R and S = {x : gi (x) =
bi } for all i. Sometimes we restrict gi to be linear functions (so that they
can be written gi (x) = ai · x). In this case, it is important to have arbitrary
constants on the right-hand side of the constraint. When gi is an arbitrary
function, then we can assume without loss of generality that bi = 0 because
otherwise we can replace gi by the function gi − bi ..
If your problem had just one constraint, then there is an intuitive way to
solve it. Use the constraint to write one variable in terms of the others.
Substitute out for that variable and solve the unconstrained problem. This
method really works as long as it is possible to use the constraint to write
one variable in terms of the others. It is also completely general.
Let G = (g1 , . . . , gm ) so that there are m constraints. G : Rn −→ Rm . Typ-
ically, the problem is uninteresting if m ≥ n (because then the constraint
set is likely to contain only one point or be empty).1
We say that the constraints are non-degenerate (or satisfy a regularity con-
dition or a constraint qualification) at a point x∗ if DG(x∗ ) has full rank
({∇gi (x∗ )} is a collection of m linearly independent vectors).
To spell this out, divide the variable x∗ = (u∗ , w∗ ) where u∗ ∈ Rn−m and
w∗ ∈ Rm and Dw G(x∗ ) 6= 0. The nondegeneracy condition guarantees that
there are m variables you can select so that Dw G(x∗ ) is nonsingular. This
means that by the implicit function theorem it is possible to solve the equa-
tions G(x∗ ) = 0 for a function W : Rn−m −→ Rm such that W (u∗ ) = w∗
and G(u∗ , W (u∗ )) = 0 in a neighborhood of (u∗ , w∗ ). With this observa-
tion, solving the constrained optimization problem is equivalent to solving
1
m ≥ n is perfectly fine for the case of inequality constraints.
13.1. EQUALITY CONSTRAINTS 157

the unconstrained optimization problem:

max f (u, W (u)).

In particular, if x∗ solves the original problem, then u∗ must be a local

maximum of the unconstrained problem. Hence the composite function
h(u) = f (u, W (u)) has a critical point at u∗ . Using the chain rule we see
that Dh(u∗ ) = Du f (x∗ ) + Dw f (x∗ )DW (u∗ ). Furthermore, the implicit
function theorem gives you a formula for DW (u∗ ). It follows that

0 = Dh(u∗ ) = Du f (x∗ ) − Dw f (x∗ ) (Dw G(x∗ ))−1 Du G(x∗ ) (13.3)

or
Du f (x∗ ) = Dw f (x∗ ) (Dw G(x∗ ))−1 Du G(x∗ ). (13.4)
Let
λ = Dw f (x∗ ) (Dw G(x∗ ))−1
or
Dw f (x∗ ) = λDw G(x∗ ). (13.5)
It follows that from equation (13.4) that

Du f (x∗ ) = λDu G(x∗ ),

which combined with (13.5) we can write

Df (x∗ ) = λDG(x∗ ).
We summarize the result in the following result.

Theorem 68. Suppose that f : Rn → R and G : Rn → Rm are differ-

entiable at x∗ and that G satisfies a nondegeneracy condition (there exist
m variables w such that Dw G(x∗ ) is non-singular). If x∗ solves max f (x)
subject to G(x) = 0, then there exists λ such that Df (x∗ ) = λDG(x∗ ).

In the theorem, DG(x∗ ) is represented as a matrix with m rows and n

columns and λ is a row vector with m components.
It is standard to write the equation

m
X
∗
∇f (x ) = λi ∇gi (x∗ ).
i=1
158 CHAPTER 13. CONSTRAINED OPTIMIZATION

The λi are often called Lagrange Multipliers.

The beauty of first-order conditions is that they replace solving a maximiza-
tion problem with finding solutions to a system of equations involving first
derivatives. This result demonstrates that the result extends to equality con-
strained optimization problems. An interesting aspect of the construction is
that it provides extra information: the multipliers λ have economic interpre-
tation.
Imagine that the constraints of the original problem are G(x) = b and we
can solve the problem at x∗ when G(x∗ ) = 0. What happens to the solution
of the problem as b changes? In particular, is it possible to compute the
derivatives of the value function: V (b) ≡ f (x∗ (b)) where x∗ (b) solves:
max f (x) subject to G(x) = b. If the same regularity condition we needed
for the theorem holds (there is a subset of m variables – denoted by w – such
that Dw G is invertible at x∗ ), then the implicit function theorem guarantees
that we can solve G(x) = b for W as a function of u and b in a neighborhood
of (u, b) = (u∗ , 0). Call the function that satisfies the first-order condition
x∗ (b) = (u∗ (b), w∗ (b)). If we carry out the substitution and computation as
before we can differentiate V (b) to obtain:

DV (0) = ((Du f (x∗ ) + Dw f (x∗ )) (Du W (x∗ , 0)) Du∗ (0)+Dw f (x∗ )Db W (x∗ , 0).

The first two terms on the right-hand side are precisely those from the first-
order condition (13.3) multiplied by Du∗ (0). These terms are zero (as in
the case of the unconstrained envelope theorem). The final term on the right
is the contribution that comes from the fact that W may change with b. We
can use the implicit function theorem to get:

Dw f (x∗ )Db W (x∗ , 0) = Dw f (x∗ ) (Dw G(x∗ ))−1 I = λ,

where the last equation is just the definition of λ. Hence we have the fol-
lowing equality constrained envelope theorem:

DV (0) = λ. (13.6)

This equation provides a nice interpretation of the Lagrange multipliers.

The question is: How does the value of the problem change if the right-
hand side of the ith constraint increases from 0? The answer is λi . That
is, the Lagrange multipliers give the incremental value of a change in the
available resources on the value of the objective function. Loosely speaking,
the decision maker is willing to pay λi to have one unit more of the ith
13.2. THE KUHN-TUCKER THEOREM 159

resource. When the constraints are equations you cannot state in advance
whether an increase in bi will raise or lower the value of the problem. That
is, you cannot state in advance the sign of λi . There is a variation of this
result that holds in inequality constrained problems. In that case there is a
definite sign to λi (whether the sign is positive or negative depends on how
you write the constraints).

13.2 The Kuhn-Tucker Theorem

This result generalizes the equality constrained case to the general case of
inequality constraints. That is, we assume that the constraint set can be
written S = {x : gi (x) ≥ bi , i ∈ I}.
The intuition is fairly simple. We know how to solve problems without con-
straints. We know how to solve problems with equality constraints. If you
have inequality constraints and you think that x∗ is an optimum, then (lo-
cally) the constraints will divide into the group that are satisfied as equations
(binding or active constraints) and the others. The first group we treat as we
would in an equality constrained problem. The second group we treat as
in an inequality constrained problem. Hence we would expect a first-order
condition like the one in the previous section:
X
∇f (x∗ ) = λi ∇gi (x∗ ) (13.7)
i∈J

where J is the set of binding constraints. Under appropriate conditions, this

turns out to be the appropriate first-order necessary condition for optimality.
The condition is often written as a pair of conditions:
X
∇f (x∗ ) = λi ∇gi (x∗ ) (13.8)
i∈I

and

λi (g(x∗i ) − bi ) = 0 for all i ∈ I. (13.9)

Equation (13.8) differs from equation (13.7) because it includes all con-
straints (not just the binding ones). Equation (13.8) and (13.9) are equiva-
lent to (13.7). Equation (13.9) can only be satisfied (for a given i) if either
λi = 0 or gi (x∗ ) = 0. That is, it is a clever way of writing that either a con-
straints is binding or its associated multiplier is zero. Once we know that
160 CHAPTER 13. CONSTRAINED OPTIMIZATION

the multipliers associated with non-binding constraints are zero, it is clear

that equations (13.7) and (13.8) are equivalent.
Equation (13.9) is called a complementary slackness condition. (Constraints
that do not bind are said to have slack.) It provides a weak interpretation of
Lagrange multipliers as “prices” or values of resources represented in each
constraint. It is useful to think of multiplier λi as the amount that value
would increase if the ith constraint were relaxed by a unit or, alternatively,
how much the person solving the problem would pay to have an “extra” unit
to allocate on the ith constraint. If a constraint does not bind, it should be
the case that this value is zero (why would you pay to relax a constraint that
is already relaxed?). Also, if you are willing to pay to gain more resource
(λi = 0) it must be the case that the constraint is binding. It is possible (but
rare) to have both λi = 0 and gi (x∗ ) = 0. This happens when you have “just
enough” of the ith resource for your purposes, but cannot use anymore.
This treatment follows the exposition of Harold Kuhn “Lectures on Mathe-
matical Economics” (Danzig and Veinot (ed.), Lectures on Applied Mathe-
matics, Volume 2).
To begin the study, we need a notion of a direction that is not ruled out by
constraints.

Definition 115. The vector v enters S from x∗ if there exists  > 0 such that
x∗ + tv ∈ S for all t ∈ [0, ).

If x∗ ∈ S and v enters S from x∗ , then it is possible to start at x∗ , go in

the direction v, and stay in the feasible set. It follows that if x is a local
maximum, then Dv f (x∗ ) ≤ 0.2 To fully describe the set of first-order con-
ditions for a constrained optimization problem, we need to figure out a nice
way to describe which vectors enter S at x and try to get a clean way of
summarizing the resulting inequalities.

Definition 116. S is described by linear inequalities if there is a matrix A

such that S = {x : Ax ≥ b}.

A general constraint set will be written in the form S = {x : gi (x) ≥ bi , i ∈

I}. If the constraint set is described by linear inequalities, then we can take
all of the gi to be linear functions (that is, gi (x) = ai · x, where ai is the ith
row of the matrix A). Note that a set described by inequalities includes an
f (x∗ +tv)−f (x∗ )
2
We will think of Dv f (x∗ ) = limt→0+ t in this section. That is, take the direc-
tional derivative to be one sided.
13.2. THE KUHN-TUCKER THEOREM 161

equality constrained set as a special case (to impose the constraint g(x) = b
substitute two inequalities: g(x) ≥ b and −g(x) ≥ −b). The converse is
not true: Equality constrained problems are typically easier than inequality
constrained problems.

Definition 117. An inward pointing normal to the boundary of the set S =

{x : gi (x) ≥ bi , i = 1, . . . , I} at x∗ is a direction of the form ∇gi (x∗ ) where
gi (x∗ ) = bi .

When gi (x∗ ) = bi the ith constraint is binding. If you move from x∗ in the
“wrong” direction, then the constraint will no longer be satisfied. However,
if you move “into” S, then the constraint will be satisfied. When S is de-
scribed by linear inequalities, then the inward directions are simply the ai
associated with binding constraints.

Theorem 69. Suppose that S is described by linear inequalities. v enters

S from x∗ if and only if v makes a nonnegative inner product with all of the
inward pointing normals to the boundary of the feasible set at x∗ .

Proof. Let x∗ ∈ S and let ak · x∗ = 0 for k ∈ J (J may be empty). Let v

enter S from x∗ . It follows that there exists  > 0 such that ak ·(x∗ + tv) ≥ 0
for k ∈ I and t ∈ [0, ). In particular, if i ∈ J, then ak · x∗ = bk and hence

ak · v ≥ 0 for all k ∈ J.

Conversely, if ak · v ≥ 0 for all k ∈ J. For k ∈ / K, it must be that

∗ ∗
ak · x > bk . Therefore ak (x + tv) > bk for sufficiently small t. For k ∈ J
we have ak · x∗ = bk so that ak · (x∗ + tv) ≥ bk . This means that v enters S
at x∗ .

Now that we have a characterization of what it means for v to enter S, we

can get a condition for optimality.

Theorem 70. If x∗ maximizes f (x) subject to x ∈ S, then ∇f (x∗ ) · v ≤ 0

for all v entering S from v.

Proof. This is the standard one-variable argument. We know that there ex-
ists  > 0 such that x∗ + tv ∈ S for t ∈ [0, ) and therefore f (x∗ ) ≥
f (x∗ + tv) for t ∈ [0, ). Hence Dv f (x∗ ) ≥ 0, which is equivalent to
∇f (x∗ ) · v ≤ 0.
162 CHAPTER 13. CONSTRAINED OPTIMIZATION

Now we can apply a variation of the separating hyperplane theorem. The

argument that we used to prove the separating hyperplane theorem proves
the following.

Theorem 71. Exactly one of the following two things holds:

There exists λ ≥ 0 such that At λ = w.

or

There exists x such that Ax ≥ 0 and w · x < 0.

The first condition in the theorem says that w is in the convex set generated
by the rows of the matrix A. Hence the theorem says that if w fails to be in a
certain convex set, then it can be separated from the set. (In the application
of the result, w = −∇f (x∗ ).) The vector x in the second condition plays the
role of the normal of the separating hyperplane. Geometrically, the second
condition states that it is possible to find a direction that that makes an angle
of less than ninety degrees with all of the rows of A and greater than ninety
degrees with w. This condition is ruled out by Theorem 70, hence the first
condition must hold.

Theorem 72. Suppose x∗ solves: max f (x) subject to x ∈ S when S is

defined by linear inequalities and f is differentiable. Then there exists λ∗k ≥
0 such that X
∇f (x∗ ) + λ∗i ∇ (gi (x∗ ) − bi ) = 0 (13.10)
i∈I

and X
λ∗i (gi (x∗ ) − bi ) = 0. (13.11)
i∈I

This theorem gives us multipliers for an inequality constrained optimization

problem provided that the constraint set is linear. The only place where we
used linearity of the constraint set was in the characterization of inward
pointing normals. It would be nice if this condition held in the non-linear
case as well. It does not.
Consider the problem: max f (x1 , x2 ) = x1 subject to x1 , x2 ≥ 0 and
(1 − x1 )3 − x2 ≥ 0. It is apparent that the solution to this problem is
(x1 , x2 ) = (1, 0). [The second and third constraints force x1 ≤ 1. Since
(1, 0) is feasible, it must be optimal.] On the other hand, if you just try
13.2. THE KUHN-TUCKER THEOREM 163

to apply the formula in Theorem 72 (in particular, the first component of

equation (13.10)), then
1 + λ1 + λ3 −3(1 − x∗ )2 = 0,


(13.12)
which plainly has no non-negative solution when x∗ = 1. On the other hand,
equation (13.11) implies that λ3 = 0 when x∗ < 1, which again means that
equation (13.10 cannot have a solution. We conclude that the conditions in
Theorem 72 need not hold if S is not described by linear inequalities.
The problem comes from the characterization of vectors entering S. The
constraints that define S give inward pointing normals at (1, 0) as (0, 1)
(from the binding constraint that x2 ≥ 0) and (0, −1) (from the binding
constraint that (1 − x1 )3 − x2 ≥ 0). v = (1, 0) makes a zero inner product
with both of these normals, but does not satisfy the definition of an entering
direction. The characterization in Theorem 72 therefore places “too many”
constraints on the derivatives of f so is not a useful characterization. The
example means that we need to do a bit of work to make sure that the the-
orem applies to general problems. The work involves a slight modification
in the definition of inward normal and then an assumption that rules out
awkward cases.
Definition 118. The vector v enters S from x∗ if there is a  > 0 and a curve
σ : [0, ) −→ S such that σ(0) = x∗ , and σ 0 (0) = v.

In the definition, σ 0 (0) denotes the right-hand derivative of σ at 0 (because

the function need not be defined for t < 0). This definition is more gen-
eral than Definition 117 because it allows inward directions by paths that
are not necessarily straight lines. If S is defined by linear inequalities, the
distinction does not matter, but in general directions may enter according to
Definition 118 but not Definition 117.3 It is immediate that ∇f (x∗ ) · v ≥ 0
for all v entering S from x∗ (according to Definition 118 as in Theorem 69).
Definition 119. If S = {x : gi (x) ≥ 0 for all i ∈ I}, gi are differentiable
and x∗ ∈ S, then the inward normals at x∗ are the vectors ∇gi (x∗ ) for i
such that gi (x∗ ) = 0.

Now we simply rule out troublesome cases:

Definition 120. Let S = {x : gi (x) ≥ 0 for all i ∈ I}, gi are differentiable,
and x∗ ∈ S. S satisfies the constraint qualification at x∗ if ni · v ≥ 0 for all
inward normals ni at x∗ implies that v enters S from x∗ .
3
A direction that enters according to Definition 117 must satisfy Definition 118.
164 CHAPTER 13. CONSTRAINED OPTIMIZATION

Plainly the example does not satisfy the constraint qualification. The con-
straint qualification will hold if the set of inward normals is linearly inde-
pendent (check that this condition fails in the example). When linear inde-
pendence holds, it is possible to find a normal direction that strictly enters:
ni · w = 1 for all i and with this you can construct a curve with the property
that ∇gi (x∗ ) · σ 0 (0) > 0 for all i such that gi (x∗ ).
Here is another technical point. The linear independence condition will not
hold if some of the constraints were derived from equality constraints (that
is, one constraint is of the form g(x) ≥ 0 and another constraint is of the
form −g(x) ≥ 0). This is why frequently equality constraints are written
separately and a linear independence condition is imposed on directions
formed by all binding constraints.
We now modify Theorem 72.

Theorem 73. Suppose x∗ solves: max f (x) subject to x ∈ S when S =

{x : gi (x) ≥ 0 for all i ∈ I} and f and g are differentiable. If x∗ satisfies
the constraint qualification at x∗ , then there exists λ∗k ≥ 0 such that
X
∇f (x∗ ) + λ∗i ∇gi (x∗ ) − bi ) = 0 (13.13)
i∈I

and X
λ∗i (gi (x∗ ) − bi ) = 0. (13.14)
i∈I

It is no accident Theorem 73 looks like Theorem 72. All that we did was
assume away the pathologies caused by non-linear constraints.

13.3 Saddle Point Theorems

We have shown that if a “regularity” condition holds then

x∗ solves max f (x) (13.15)

subject to
gi (x) ≥ b̃i , i ∈ I (13.16)
then equations (13.10) and (13.11) must hold. Note that

∇gi (x∗ ) = ∇ (gi (x∗ ) − bi )

13.3. SADDLE POINT THEOREMS 165

so equation (13.16) can be written without the bi .

In particular, this result holds when the constraints are linear (which guar-
antees that the regularity condition will hold).
As before, let us concentrate first on the linear case:

max c · x subject to Ax ≤ b, x ≥ 0. (13.17)

This is the standard linear programming problem. Here f (x) = c · x and
for each i = 1, . . . , m, gi (x) = −ai · x, where ai is a row of A and for
i = m + 1, . . . , m + n, gi (x) = xi . To put (13.17) in the form of the
original problem, set the b̃ in the theorem equal (constraint (13.16)) equal to
(−b1 , . . . , −bm , 0, . . . , 0). b̃ has n + m components.
Notice that ∇f (x) ≡ c when f (x) ≡ c · x and similarly for the constraints.
Hence, if x∗ solves (13.17) there exists y ∗ , z ∗ ≥ 0 such that

c − y∗A + z∗ = 0 (13.18)

and
y ∗ · (b − Ax∗ ) + z ∗ · x∗ = 0. (13.19)

Equations (13.18) and (13.19) divide the multiplier vector λ into two parts.
y ∗ contains the components of λ corresponding to i ≤ m – the constraints
summarized in the matrix A. z ∗ are the components of λ corresponding to
i > m; these are the non-negativity constraints.
It is useful to rewrite the constraints. Let

s(x; y, z) = c · x + y · (b − Ax) + z · x. (13.20)

We have the following theorem.

Theorem 74. If x∗ solves (13.17), then there exists y ∗ and z ∗ ≥ 0 such that

(a) s(x∗ ; y ∗ , z ∗ ) ≥ s(x; y ∗ , z ∗ ) for all x.

(b) s(x∗ ; y ∗ , z ∗ ) ≤ s(x∗ ; y, z) for all y, z ≥ 0.
(c) y ∗ solves min b · y subject to At y ≥ c and y ≥ 0.
(d) b · y ∗ = c · x∗ .

.
166 CHAPTER 13. CONSTRAINED OPTIMIZATION

Proof. It follows from equation (13.18) that

s(x : y ∗ , z ∗ ) = c − At y ∗ +z ∗ x + b · y ∗ = b · y ∗


(13.21)

for all x. Therefore the first claim of the theorem holds as an equation.
Furthermore, equation (13.19) implies that s(x∗ ; y ∗ , z ∗ ) = c · x∗ , which
yields the fourth line.
To prove the second line, we must show that

s(x∗ ; y, z) = c·x∗ +y ∗ ·(b − Ax∗ )+x·z ≥ s(x∗ ; y ∗ , z ∗ ) = c·x∗ for all y, z.

(13.22)
∗ ∗
However, x ≥ 0 and b ≥ Ax by the constraints in (13.17). Therefore,
if y, z ≥ 0, then y (b − Ax∗ ) + z · x ≥ 0 and so the second claim in the
theorem follows.
To prove the third claim, first note that

c − At y ∗ + z ∗ = 0 and z ∗ ≥ 0

implies that c ≤ At y ∗ . Therefore y ∗ satisfies At y ∗ ≥ c and y ∗ ≥ 0. If y also

satisfies At y ≥ c and y ≥ 0, then

s(x∗ ; y, z ∗ ) = c · x∗ + y (b − Ax∗ ) + z ∗ · x∗ (13.23)

= c − At y x∗ + b · y + z ∗ · x∗


(13.24)
≤ c − At y x∗ + b · y


(13.25)
= b · y, (13.26)

where the first equation is the definition of s(·), the second equation is just
algebraic manipulation, the third equation follows because equations (13.19),
y ∗ ≥ 0, and b ≥ Ax∗ imply that z ∗ · x∗ = 0, and the inequality follows be-
cause x∗ ≥ 0 and c ≤ At y. It follows that

b · y ≥ s(x∗ ; y, z ∗ ) ≥ s(x∗ ; y ∗ , z ∗ ) = b · y ∗

so y ∗ solves the minimization problem (3).

We have show that if x∗ solves the Primal:

max c · x subject to Ax ≤ b, x ≥ 0,

then y ∗ solves the Dual:

min b · y subject to At y ≥ c, y ≥ 0.
13.3. SADDLE POINT THEOREMS 167

Notice that the Dual is equivalent to

max −b · y subject to − At y ≤ −c, y ≥ 0

and that this problem has the same general form as the Primal (with c re-
placed by −b and A replaced by −At and b replaced by −c. Hence the Dual
can be written in the form of the Primal and (applying the transformation
one more time), the Dual of the Dual is the Primal.
Summarizing the result we have:

Theorem 75. Associated with any Primal Linear Programming Problem

max c · x subject to Ax ≤ b, x ≥ 0,

there is a Dual Linear Programming Problem:

min b · y subject to At y ≥ c, y ≥ 0.

The Dual of the Dual is the Primal. If the Primal has a solution x∗ , then
the Dual has a solution y ∗ and b · y ∗ = c · x∗ . Moreover, if the Dual has
a solution, then the Primal has a solution and the problems have the same
value.

This result is called “The Fundamental Duality Theorem” (of Linear Pro-
gramming). It frequently makes it possible to interpret the “multipliers”
economically. Mathematically, the result makes it clear that when you solve
a constrained optimization problem you are simultaneously solving another
optimization problem for a vector of multipliers.
There is one yi for every constraint in the primal (in particular, there need
not be the same number of variables in the primal as variables in the dual).
You therefore cannot compare x and y directly. You can compare the values
of the two problems. The theorem shows that when you solve these prob-
lems, the values are equal: b · y ∗ = c · x∗ . It is straightforward to show that if
x satisfies the constraints of the Primal then c·x ≤ c·x∗ and if y satisfies the
constraints of the dual then b · y ≥ b · y ∗ . Consequently any feasible value
of the Primal is less than or equal to any feasible value for the Dual. The
minimization problem provides upper bounds for the maximization problem
(and conversely).
It is a useful exercise to look up some standard linear programming prob-
lems and try to come up with interpretations of dual variables.
168 CHAPTER 13. CONSTRAINED OPTIMIZATION

The arguments above demonstrate that y ∗ (b − Ax∗ ) = 0 and (c − At y ∗ ) x∗ =

0. These equations are informative. Consider y ∗ (b − Ax∗ ) = 0. This equa-
tion is a sum of terms of the form yi∗ (the value of the ith dual variable)
times (b − Ax∗ )i (the difference between the larger and smaller side of the
ith constraint in the primal. Both of these quantities are positive or zero.
When you add all of the terms together, you must get zero. Therefore (since
each term is non-negative), each term must be zero. If you multiply two
things together and get zero, one of them must be zero. What does this
mean? If you know that yi > 0, then you know that the associated con-
straint in binding (holds as an equation). If you know that a constraint is
not binding, then the associated dual variable must be zero. This provides a
weak sense in which the dual variables give an economic value to the given
resources. Think about the Primal as a production problem and the con-
straints resource constraints. That is, you have access to a technology that
transforms the inputs b1 , . . . , bm into outputs x1 , . . . , xn . Suppose that you
solve the problem and there is left over b1 . That is, the first constraint is
not binding. How much would you pay for an extra unit of the first input?
Nothing. Why? Because you cannot profitably turn that resource into prof-
itable output (if you could do this, you would have used up your original
supply). Hence the “economic” value of the first input is 0. Similarly, if
you are willing to pay a positive amount for the second good y2 > 0, then
it must be that you use up your entire supply (second constraint binding). It
turn out that yi operates as a resource price more generally: If you change
the amount of bi by a small amount δ, then your profits change by δyi . This
story provides another interpretation of the equation b · y ∗ = c · x∗ . The
left-hand quantity is the value of your inputs. The right-hand side is the
value of your output.

The duality theorem of linear programming tells you a lot about the dual
if you know that the primal has a solution. What happens if the Primal
fails to have a solution? First, since the duality theorem applies when the
Dual plays the role of the Primal, it cannot be that the Dual has a solution.
That is, the Dual has a solution if and only if the Primal has a solution.
There are only two ways in which a linear programming problem can fail
to have a solution. Either {x : Ax ≤ b, x ≥ 0} is empty. In this case the
problem is infeasible – nothing satisfies the constraints. In this case you
certainly do not have a maximum. Alternatively, the problem is feasible
({x : Ax ≤ b, x ≥ 0} is not empty), but you cannot achieve a maximum.
Of course, since linear functions are continuous, this means that the feasible
set cannot be bounded. It turns out that in the second case you can make
13.3. SADDLE POINT THEOREMS 169

the objective function of the Primal arbitrarily large (so the Primal’s value
is “infinite”). It is not hard to show this. It is even easier to show that it is
impossible for the Primal to be unbounded if the Dual is feasible (because
any feasible value of the Dual is an upper bound for the Primal’s value).
To summarize:
(a) If the primal has a solution, then so does the dual, and the solutions
have the same value.
(b) If the primal has no solution because it is unbounded, then the dual is
not feasible.
(c) If the primal is not feasible, then either the dual is not feasible or the
dual is unbounded.
(d) The three statements about remain true if you interchange the words
“primal” and “dual.”
We began by showing that if x∗ solves the Primal, then there exist y ∗ and z ∗
such that
s(x∗ ; y, z) ≥ s(x∗ ; y ∗ , z ∗ ) ≥ s(x; y ∗ , z ∗ ). (13.27)
Expression (13.27) states that the function s is maximized with respect to
x at x = x∗ , (y, z) = (y ∗ , z ∗ ) and minimized with respect to (y, z) at
(x∗ , y ∗ , z ∗ ), (y, z) ≥ 0. This makes the point (x∗ , y ∗ , z ∗ ) a saddle point of
the function s.
What about the converse? Let us state the problem more generally.
Given the problem

max f (x) subject to gi (x) ≥ bi , i ∈ I

consider the function

s(x, λ) = f (x) + λ · (g(x) − b) (13.28)

where g(x) = (g1 (x), . . . , gI (x)) , b = (b1 , . . . , bI ) .(x∗ , λ∗ ) is a saddle

point if λ∗ ≥ 0 and s is maximized with respect to x and minimized with
respect to λ at (x∗ , λ∗ ). That is,

s(x, λ∗ ) ≤ s(x∗ , λ∗ ) ≤ s(x∗ , λ) (13.29)

for all λ ≥ 0.
Theorem 76. If (x∗ , λ∗ ) is a saddle point of s, then x∗ solves max f (x)
subject to gi (x) ≥ bi .
170 CHAPTER 13. CONSTRAINED OPTIMIZATION

Proof. Since s(x∗ , λ∗ ) ≤ s(x∗ , λ) for all λ ≥ 0,

λ∗ · (g(x∗ ) − b) ≤ λ · (g(x∗ ) − b) for all λ ≥ 0. (13.30)

Therefore (setting λ = 0 in inequality (13.30)) it must be that λ∗ ·(g(x∗ ) − b) ≤

0.
We also claim that gi (x∗ ) − bi ≥ 0 for all i. To prove this, suppose that there
exists a k such that gj (x∗ ) − bj < 0 and let λi = 0 for i 6= k and λk = M .
It follows that

λ∗ · (g(x∗ ) − b) ≤ λ · (g(x∗ ) − b) = M (gk (x∗ ) − bk ) .

Since the right-hand side gets arbitrarily small as M grows large, which
cannot hold. This establishes the claim.
The only way to have λ∗ ≥ 0, gi (x∗ )−bi ≥ 0 for all i, and λ∗ ·(g(x∗ ) − b) ≤
0 is for λ∗ · (g(x∗ ) − b) = 0. It follows that s(x∗ , λ∗ ) ≤ s(x∗ , λ) implies
that x∗ is feasible (g(x∗ ) ≥ b) and that λ∗ , x∗ satisfy the complementary
slackness condition
λ∗ · (g(x∗ ) − b) = 0. (13.31)

To complete the proof, we must show that g(x) ≥ b implies that f (x∗ ) ≥
f (x). However, we know that s(x∗ , λ∗ ) ≥ s(x, λ∗ ) and we just showed that

s(x, λ∗ ) = f (x∗ ) + λ∗ · (g(x∗ ) − b) = f (x∗ ).

On the other hand,

s(x, λ∗ ) = f (x) + λ∗ · (g(x) − b)

so if g(x) − b ≥ 0, λ∗ ≥ 0 implies that s(x, λ∗ ) ≥ f (x). Summarizing we

have

f (x∗ ) = s(x∗ , λ∗ ) ≥ s(x, λ∗ ) ≤ f (x) whenever g(x) ≥ b.

Hence we have shown an equivalence between saddle points and optimal-

ity in linear programming problems and that, in general, the saddle point
property guarantees optimality. Under what conditions is the converse true?
That is, when can we say that if x∗ solves max f (x) subject to g(x) ≥ b,
that there exists λ∗ ≥ 0 such that (x∗ , λ∗ ) is a saddle point?
13.3. SADDLE POINT THEOREMS 171

Theorem 77. If f and gi are concave and x∗ solves max f (x) subject to
g(x) ≥ b and a “constraint qualification” holds, then there exists λ∗ ≥ 0
such that (x∗ , λ∗ ) is a saddle point of s(x, λ) = f (x) + λ · (g(x) − b).

Comments:

(a) We will discuss the “constraint qualification” later. It plays the role
that “regularity” played in the “inward normal” discussion.
(b) This result is often called the Kuhn-Tucker Theorem. λ are called
Kuhn-Tucker multipliers. (Largrange mulitpliers typically refer only
to multipliers from equality constrained problems. Remember: equal-
ity constraints are a special case of inequality constraints and are gen-
erally easier to deal with.)
(c) It is possible to prove essentially the same theorem with weaker as-
sumptions (quasi concavity of g). The proof that we propose does not
generalize directly. See Sydsaeter or the original paper of Arrow and
Enthoven, Econometrica 1961.
(d) The beauty of this result is that it transforms a constrained problem
into an unconstrained one. Finding x∗ to maximize s(x, λ∗ ) for λ∗
fixed is relatively easy. Also, note that if f and gi are differentiable,
then the first-order conditions for maximizing s are the familiar ones:

∇f (x∗ ) + λ · ∇g(x∗ ) = 0.

The first order conditions that characterization the minimization prob-

lem are precisely the complementary slackness conditions. Since s(x, λ)
is concave in x when f and g are concave, it will turn out that

∇f (x∗ ) + λ · ∇g(x∗ ) = 0, λ∗ · (g(x∗ ) − b) = 0, g(x∗ ) ≥ 0

are necessary and sufficient for optimization. This will be an imme-

diate consequence of the (ultimate) discussion of the sufficient condi-
tions for optimality.

Proof. We assumed concavity and we need to find “multipliers.” It looks

like we need a separation theorem. The trick is to find the right set.
Let

K = {(y, z) : there exists x such that g(x) ≥ y and f (x) ≥ z}

172 CHAPTER 13. CONSTRAINED OPTIMIZATION

K is convex since gi and f are concave. In this notation, y ∈ Rm and

z ∈ R. If x∗ solves the maximization problem, then (b, f (x∗ )) ∈ K. In
fact, it must be a boundary point since otherwise there exists  > 0 and x
such that gi (x) ≥ bi + > bi for all i and f (x) ≥ f (x∗ )+. This contradicts
the optimality of x∗ .
Hence there is a separating hyperplane. That is, there exists (λ, µ) 6= 0, λ ∈
Rm , µ ∈ R such that

λ · b + µf (x∗ ) ≥ λ · y ≥ µz

for all (y, z) ∈ K. This is almost what we need. First, observe that (λ, µ) ≥
0. If some component of λ, say λj < 0, you get a contradiction in the
usual way by taking points (y, z) ∈ K where, for example, z = f (x∗ ), yi =
bi , i 6= j, yj = −M . If M is large enough, this choice contradicts

λ · b + µf (x∗ ) ≥ λ · y + µz.

Similarly, if µ < 0 you obtain a contradiction because (b, z) ∈ K for all

arbitrarily small (large negative) z. Hence we have (λ, µ) ≥ 0, (λ, µ) 6= 0
and
λ · b + µf (x∗ ) ≥ λ · y + µz
for all (y, z) ∈ K. It is also true that

λ · (g(x∗ ) − b) = 0.

This fact follows because (g(x∗ ), f (x∗ )) ∈ K. Hence

λ · b + µf (x∗ ) ≥ λ · g(x∗ ) + µf (x∗ ).

Therefore 0 ≥ λ (g(x∗ ) − b) . However, λ ≥ 0 and g(x∗ ) ≥ b (by feasibil-

ity). Hence, 0 = λ (g(x∗ ) − b) .
Now suppose that µ > 0. (This is where the constraint qualification comes
in.) Put λ∗ = λ/µ. We have

f (x∗ ) + λ∗ · b ≥ f (x) + λ∗ · g(x) for all x (13.32)

(since (g(x), f (x)) ∈ K) and λ∗ · (g(x∗ ) − b) = 0.

Now we can confirm the saddle-point property:

s(x∗ , λ) ≥ s(x∗ , λ∗ ) for λ ≥ 0 ≡

f (x∗ ) + λ · (g(x∗ ) − b) ≥ f (x∗ ) + λ∗ · (g(x∗ ) − b) ≡
λ · (g(x∗ ) − b) ≥ 0 for λ ≥ 0.
13.4. SECOND-ORDER CONDITIONS 173

where we use λ∗ ·(g(x∗ ) − b) = 0 and the last line holds because g(x∗ ) ≥ b.

s(x∗ , λ∗ ) ≥ s(x, λ∗ ) ≡
f (x∗ ) + λ∗ · (g(x∗ ) − b) ≥ f (x) + λ∗ · (g(x) − b) ≡
f (x∗ ) + λ∗ · b ≥ f (x) + λ∗ · g(x)

since λ∗ · (g(x∗ ) − b) = 0.
All that remains is to clarify the constraint qualification. We need an as-
sumption that guarantees that µ, the weight on f , is strictly positive. What
happens if µ = 0? We have

λ 6= 0 such that λ · b ≥ λ · y for all y such that (y, z) ∈ K for some z.

(13.33)
One condition that guarantees this cannot happen is:

there exist x∗ such that gi (x∗ ) > bi for all i.

This condition is not consistent with (13.33) because we cannot have λ · b ≥

λ · g(x∗ ) if λ ≥ 0, λ 6= 0, and gi (x∗ ) > bi for all i.
The constraint qualification is fairly weak. It just requires that the constraint
set has a non-empty interior.

13.4 Second-Order Conditions

We know from the theory of unconstrained optimization that the first-order
conditions for optimality do not distinguish local maxima from local min-
ima from points that are neither. We also know that the second-order con-
dition, derived from thinking about how a first derivative changes in the
neighborhood of a critical point, does provide useful information. When
you have an equality constrained maximization problem you derive first-
order conditions by requiring that the objective function be maximized in
all directions that will satisfy the constraints of the problem. The second-
order conditions must hold for exactly these directions. That is, if x∗ is a
maximum and x∗ +tv satisfies the constraints of the problem for t near zero,
00
then h(t) = f (x∗ + tv) has a critical point at t = 0 and h ) (0) ≤ 0. Two
00 00
applications of the chain rule yield that h ) (0) = v t D2 f (x∗ )v, so h (0) < 0
if and only if a quadratic form is negative. If the problem is unconstrained
174 CHAPTER 13. CONSTRAINED OPTIMIZATION

(that is, if x∗ + tv satisfies the constraints of the problem for all v and suf-
ficiently small t, then the second-order conditions require that the matrix
of second derivatives be negative semi-definite. In general, this condition
need only apply in the directions consistent with the constraints. There is a
theory of “Boardered Hessians” that allows you to use some insights from
the theory of quadratic forms to classify when quadratic form restricted to
a set of directions will be positive definite. This theory is ugly and we will
say no more about it.

13.5 Examples
Example 47.
max x21 − x2
x1 ,x2

subject to x1 − x2 = 0
It is pretty obvious then that x1 = x2 and thus the problem is really

max x21 − x1
x1

So this example is so simple that it can be done by direct substitution, but

with harder examples we must do this implicitly.

g(x1 , x2 ) = c
implicitly define x2 = h(x1 ) around x∗1 (the solution).
Recall that h is well defined if
∂g ∗
(x ) 6= 0
∂x2

g(x1 , h(x1 )) = 0
∂g ∗ ∂g ∗ 0 ∗
(x ) + (x )h (x1 ) = 0
∂x1 ∂x2
∂g
0 ∂x1
(x∗ )
h (x1 ) = − ∂g
∂x2
(x∗ )
13.5. EXAMPLES 175

This gives us an unconstrained maximization problem, for x near x∗ .

max f (x1 , h(x1 ))

x1

Because x∗ solves the original maximization problem, it solves this one.

Example 48. The classic example of constrained optimization is the utility

maximization problem subject to a budget constraint. You can think of this
as me walking into Costco (the world’s greatest store) with I dollars in my
pocket, and having surveyed the vast array of wonderful products they have,
I make the selection that will leave me happiest on leaving the store.

max U (x)
x

subject to pt · x ≤ I
where,
p is vector of prices
x is vector of commodities
I is income.
So all the inequality is saying is that a person cannot spend more than the
money in their pocket!
For these problems we just assume that the constraint binds (we will see
why) since people obviously spend all the money in their pockets. You may
question this as people do save, but if people do invest in savings then we
just define saving as an extra commodity and this gets rid of the problem.

Example 49. The other classic example of constrained optimization is the

expenditure minimization problem. Suppose I really on Tuesday afternoon
my girlfriend is very happy. Her utility level is 100 utils. But then on Tues-
day night I do something really stupid, and she gets really annoyed. As a
result her utility level drops. Now I really want to get her back to her Tues-
day afternoon level of happiness again so I decide to buy her some gifts. But
I am a cheapskate and as such want to spend as little as possible getting her
back to the same level of happiness that she was at on Tuesday afternoon.
So now my problem is different: the goal is now to spend as little as possible
subject to a utility constraint. I want to spend as little as possible and get
her back to a level of utility that is at least as great as it was on Tuesday
afternoon.
176 CHAPTER 13. CONSTRAINED OPTIMIZATION

Example 50. max log(x1 + x2 ) subject to: x1 + x2 ≤ 5 and x1 , x2 ≥ 0.

Since log is a strictly increasing function, the problem is equivalent to solv-
ing max(x1 + x2 ) subject to the same constraints. Since the objective func-
tion is increasing in both variables, the constraint x1 + x2 ≤ 5 must bind.
Hence the problem is equivalent to maximizing x1 +x2 subject to x1 +x2 = 5
and x1 , x2 ≥ 0. At this point it should be clear that as long as x1 ∈ [0, 5]
any pair (x1 , 5 − x1 ) solves the problem.
Of course, you could solve the problem using the Kuhn-Tucker Theorem.
Doing so would give you the additional insight of finding multipliers, but
this would not compensate you for the pain.
Example 51. max log(1+x1 )−x22 subject to: x1 +2x2 ≤ 3 and x1 , x2 ≥ 0.
The first-order conditions are
1
= λ1 − λ2 (13.34)
1 + x1
−2x22 = 2λ1 − λ3 (13.35)
λ1 (3 − x1 − 2x2 ) = 0 (13.36)
λ 2 x1 = 0 (13.37)
λ 3 x2 = 0 (13.38)

It should be clear that 3 = x1 + 2x2 and x1 > 0 in the solution. Hence

the third constraint is satisfied, λ2 = 0 by the fourth constraint, and the
first constraint gives: λ1 = 1/(1 + x1 ) and therefore the second constraint
implies that λ3 > 0. It follows from the final constraint that x2 = 0 and
therefore x1 = 3 and λ1 = 1/4.
Again, one could solve this problem by observation.
Example 52. max x1 · · · xk subject to ki=1 xi = K.
P

This problem has an equality constraint. If v = x1 · · · xk , then we can write

the first-order conditions as v = λxi for all i. The constraint then implies
that kv = λK and therefore each xi = K/k.
Example 53. We can show that the P maximum value of the function f (x) =
x1 x2 · · · xn on the unit sphere: {x : ni=1 x2i = 1} is to set xn = (1/n)n for
2 2 2

all n. This is because the solution must satisfy

f (x)
2 = 2λxi ,
xi
13.5. EXAMPLES 177

which implies that the xi are independent of i.4

We can use this solution to deduce an interesting property. Given any n
positive numbers a1 , . . . , an , let

1/2
ai
xi = , for i = 1, . . . , n.
(a1 + · · · + an )1/2
Pn
It follows that i=1 x2i = 1 and so
 1/n
a1 · · · an 1
≤
(a1 + · · · + an )n n

and so
(a1 + · · · + an )1/2
1/n
(a1 + · · · + an ) ≤ .
n
The interpretation of the last inequality is that the geometric mean of n
positive numbers is no greater than their arithmetic mean.
It is possible to use techniques from constrained optimization to deduce
other important results (the triangle inequality, the optimality of least squares,
. . . ).

Example 54. A standard example in consumer theory is the Cobb-Douglas

Utility function, f (x) = xa11 · · · xann , where the coefficients are nonnegative
and sum to one. A consumer will try to maximize f subject to a budget
constraint: p · x ≤ w, where p ≥ 0, p 6= 0 is the vector of prices and
w > 0 is a positive wealth level. Since the function f is increasing in its
arguments, the budget constraint must hold as an equation. The first-order
conditions can be written ai f (x)/xi = λpi or

ai f (x) = λpi xi .

Summing both sides of this equation and using the fact that the ai sum to
one yields:
f (x) = λp · x = λw.
It follows that
4
Careful: We know that this problem has both a maximum and a minimum. How did we know
that this was the maximum? And where did the minimum go? The answer is that it is plain that
you do not want to set any xi = 0 to get a maximum and that we have the only critical point
that does not involve xi = 0. It is also clear that setting one or more xi = 0 does solve the
minimization problem.
178 CHAPTER 13. CONSTRAINED OPTIMIZATION

 a1  an
wai a1 an
xi = and λ = ··· .
pi p1 pn
The previous two examples show illustrate that a bit of cleverness allows
neat solutions to the first-order conditions.
Chapter 14

Monotone Comparative Statics

[following Levin-Segal notes.]

Three of the department’s theorists have been educated at Stanford where
they were trained to think of comparative statics in a different way.
The theory of monotone comparative statics (MCS) provides a general set of
techniques that are different from the implicit function theorem for deciding
when the set of solutions to an optimization problem are monotonic in a
parameter.
The standard approach uses the Implicit Function Theorem. What makes
MCS attractive is that it may permit you to say things even when IFT does
not apply. MCS can provide new insights when (a) the solution “function”
is not a function (that is, for example, when a parameterized optimization
problem has multiple solutions); (b) when the solution is not differentiable;
or (c) when you cannot sign the derivative of the solution function using
IFT.
Perhaps the most important attraction of the theory is that it shifts attention
away from concavity assumptions. Consider a problem of the form:

x∗ (θ) ≡ arg max f (x, θ), subject to θ ∈ Θ; x ∈ S(θ)

This problem is equivalent to

x∗ (θ) ≡ arg max φ(f (x, θ)), subject to θ ∈ Θ; x ∈ S(θ)

for any strictly increasing φ(·) even though the transformation may destroy
smoothness or concavity properties of the objective function.

179
180 CHAPTER 14. MONOTONE COMPARATIVE STATICS

Begin with problems in which S(θ) is independent of θ and both x and θ

are real variables. Assume existence of a solution (we know conditions that
guarantee this), but not necessarily uniqueness (we also know conditions
that guarantee uniqueness, but they are strong). Instead, allow x∗ (θ) it be
set valued.
Once you permit the solution to be multi-valued, you need to generalize the
notion of “greater than” to include sets. One idea is to say that the set of
real numbers A is greater than the set of real numbers B if and only if for
all a ∈ A, b ∈ B, a > b. This definition is too restrictive. We would like
to be able to say that a set A is bigger than B if A is obtained from B by
adding one element that is strictly greater than every element in B. It turns
out that this is the “right” definition.

Definition 121. For two sets of real numbers A and B, we say that A ≥s B
(“A is greater than or equal to B in the strong set order” if for any a ∈ A
and b ∈ B, min{a, b} ∈ B and max{a, b} ∈ A.

According to this definition A = {1, 3} is not greater than or equal to B =

{0, 2}.
Using this definition, we will say that x∗ (·) is non-decreasing in µ if and
only if µ < µ0 implies that x∗ (µ0 ) ≥s x∗ (µ). One immediate implication
of the definitions is that if x∗ (·) is nondecreasing and min x∗ (θ) exists for
all θ, then min x∗ (θ) is non decreasing. (An analogous statement holds for
max x∗ (·).)

Definition 122. The function f : R2 → R is supermodular or has increas-

ing differences in (x; µ) if for all x0 > x, f (x0 ; µ)−f (x; µ) is nondecreasing
in µ.

If f is supermodular in (x; µ), then the incremental gain to choosing a higher

x is greater when µ is higher.
Supermodularity is equivalent to the property that µ0 > µ implies that
f (x; µ0 ) − f (x; µ) is nondecreasing in x. This claim follows because

f (x0 ; µ0 ) − f (x0 ; µ) ≥ f (x; µ0 ) − f (x; µ)

is equivalent to

f (x0 ; µ0 ) − f (x; µ0 ) ≥ f (x0 ; µ) − f (x; µ).

 181

When f is smooth, supermodularity has a characterization in terms of deriva-

tives.

Lemma 4. A twice continuously differentiable function f : R2 → R is su-

permodular in (x; µ) if and only if D12 f (x; µ) ≥ 0 for all (x; µ).

The inequality in the definition of supermodularity is just the discrete ver-

sion of the mixed-partial condition in the lemma.
The next result, Topkis’s Monotonicity Theorem, says that supermodularity
is sufficient to draw comparative statics conclusions in optimization prob-
lems.

Theorem 78 ( Topkis’s Monotonicity Theorem)). If f is supermodular in

(x; µ), then x∗ (µ).

Proof. Suppose µ0 > µ, and that x ∈ x∗ (µ) and x0 ∈ x∗ (µ0 ). x ∈ x∗ (µ) im-
plies f (x; ) − f (min{x, x0 }; µ) ≥ 0. This implies that f (max{x, x0 }; µ) −
f (x0 ; µ) ≥ 0 (you need to check two cases, x > x0 and x0 > x). By
supermodularity, f (max{x, x0 }; µ0 ) − f (x0 ; µ0 ) ≥ 0, which implies that
max{x, x0 } ∈ x∗ (µ0 ).
On the other hand, x0 ∈ x∗ (µ0 ) implies that f (x0 ; µ0 ) − f (max{x, x0 }, µ) ≥
0, or equivalently f (max{x, x0 }, µ) − f (x0 ; µ0 ) ≤ 0. This implies that
f (max{x, x0 }; µ0 )−f (x0 ; µ0 ) ≥ 0, which by supermodularity implies f (x; µ)−
f (min{x, x0 }; µ) ≤ 0 and so min{x, x0 } ∈ x∗ (µ).

Given the differential form of supermodularity and the implicit function

theorem, this result should not be surprising. It would follow from the IFT
whenever the standard full-rank condition in the IFT holds. (Note that at
a maximum, if D11 f (x∗ , µ) 6= 0, if must be negative (by the second-order
condition), hence the IFT tells you that x∗ (µ) is strictly increasing.

Example 55. A monopolist solves max p(q)q − c(q, µ) by picking quantity

q. p(·) is the price function and c(·) is the cost function, parametrized by
µ. Let q ∗ (µ) be the monopolists optimal quantity choice. If −c(q, µ) is
supermodular in (q, µ) then the entire objective function is. It follows that
q ∗ is nondecreasing as long as the marginal cost of production decreases in
µ.

It is sometimes useful to “invent” an objective function in order to apply

the theorem. For example, if one wishes to compare the solutions to two
182 CHAPTER 14. MONOTONE COMPARATIVE STATICS

different maximization problems, maxx∈S g(x) and maxx∈S h(x), then we

can apply the theorem to an artificial function, f
(
g(x) if µ = 0
f (x, µ) =
h(x) if µ = 1
so that if f is supermodular (h(x) − g(x) nondecreasing), then the solution
to the second problem is greater than the solution to the first.
MCS can be extended in several ways. The general ideas extend to more
complicated domains. The supermodularity condition can be replaced by
a weaker single-crossing condition (that appears in many applications in
200C).

Definition 123. The function f : R2 → R satisfies the single-crossing con-

dition in (x; µ) if for all x0 > x, µ0 > µ f (x0 ; µ) − f (x; µ) ≥ 0 implies
f (x0 ; µ0 ) − f (x; µ0 ) ≥ 0 and f (x0 ; µ) − f (x; µ) > 0 implies f (x0 ; µ0 ) −
f (x; µ0 ) > 0.

Theorem 79. If f is single crossing in (x; µ), then x∗ (µ) = arg maxx∈S(µ) f (x; µ)
is nondecreasing. Moreover, if x∗ (µ) is nondecreasing in µ for all choice
sets S, then f is single-crossing in (x; µ).