An Introduction

to
Real Analysis

Cesar O. Aguilar
May 4, 2022
 Contents

1 Preliminaries 1
1.1 Sets, Numbers, and Proofs . . . . . . . . . . . . . . . . . 1
1.2 Mathematical Induction . . . . . . . . . . . . . . . . . . 8
1.3 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Countability . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 The Real Numbers 31

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Algebraic and Order Properties . . . . . . . . . . . . . . 35
2.3 The Absolute Value . . . . . . . . . . . . . . . . . . . . . 39
2.4 The Completeness Axiom . . . . . . . . . . . . . . . . . 47
2.5 Applications of the Supremum . . . . . . . . . . . . . . . 58
2.6 Nested Interval Theorem . . . . . . . . . . . . . . . . . . 63

3 Sequences 69
3.1 Limits of Sequences . . . . . . . . . . . . . . . . . . . . . 69
3.2 Limit Theorems . . . . . . . . . . . . . . . . . . . . . . . 82
3.3 Monotone Sequences . . . . . . . . . . . . . . . . . . . . 91
3.4 Bolzano-Weierstrass Theorem . . . . . . . . . . . . . . . 99
3.5 limsup and liminf . . . . . . . . . . . . . . . . . . . . . . 106
3.6 Cauchy Sequences . . . . . . . . . . . . . . . . . . . . . . 112

3
 3.7 Infinite Series . . . . . . . . . . . . . . . . . . . . . . . . 118

4 Limits of Functions 137

4.1 Limits of Functions . . . . . . . . . . . . . . . . . . . . . 137
4.2 Limit Theorems . . . . . . . . . . . . . . . . . . . . . . . 150

5 Continuity 157
5.1 Continuous Functions . . . . . . . . . . . . . . . . . . . . 157
5.2 Combinations of Continuous Functions . . . . . . . . . . 163
5.3 Continuity on Closed Intervals . . . . . . . . . . . . . . . 167
5.4 Uniform Continuity . . . . . . . . . . . . . . . . . . . . . 174

6 Differentiation 181
6.1 The Derivative . . . . . . . . . . . . . . . . . . . . . . . 181
6.2 The Mean Value Theorem . . . . . . . . . . . . . . . . . 191
6.3 Taylor’s Theorem . . . . . . . . . . . . . . . . . . . . . . 196

7 Riemann Integration 205

7.1 The Riemann Integral . . . . . . . . . . . . . . . . . . . 205
7.2 Riemann Integrable Functions . . . . . . . . . . . . . . . 216
7.3 The Fundamental Theorem of Calculus . . . . . . . . . . 223
7.4 Riemann-Lebesgue Theorem . . . . . . . . . . . . . . . . 225

8 Sequences of Functions 229

8.1 Pointwise Convergence . . . . . . . . . . . . . . . . . . . 230
8.2 Uniform Convergence . . . . . . . . . . . . . . . . . . . . 238
8.3 Properties of Uniform Convergence . . . . . . . . . . . . 245
8.4 Infinite Series of Functions . . . . . . . . . . . . . . . . . 255

9 Metric Spaces 267

9.1 Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . 267
 9.2 Sequences and Limits . . . . . . . . . . . . . . . . . . . . 275
9.3 Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . 286
9.4 Completeness . . . . . . . . . . . . . . . . . . . . . . . . 293
9.5 Compactness . . . . . . . . . . . . . . . . . . . . . . . . 305
9.6 Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . 311

10 Multivariable Differential Calculus 317

10.1 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . 317
10.2 Differentiation Rules and the MVT . . . . . . . . . . . . 328
10.3 The Space of Linear Maps . . . . . . . . . . . . . . . . . 334
10.4 Solutions to Differential Equations . . . . . . . . . . . . 334
10.5 High-Order Derivatives . . . . . . . . . . . . . . . . . . . 334
10.6 Taylor’s Theorem . . . . . . . . . . . . . . . . . . . . . . 345
10.7 The Inverse Function Theorem . . . . . . . . . . . . . . . 348
 Preface

The material in these notes constitute my personal notes that are used
in the course lectures for MATH 324 and 325 (Real Analysis I, II).
You will find that the lectures and these notes are very closely aligned.
The notes highlight the important ideas and examples that you should
master as a student. You may find these notes useful if:

• you miss a lecture and need to know what was covered,

• you want to know what material you are expected to master,

• you want to know the level of difficulty of questions that you

should expect in a test, and

• you want to see more worked out examples in addition to those

worked out in the lectures.

If you find any typos or errors in these notes, no matter how small,
please email me a short description (with a page number) of the typo/error.
Suggestions and comments on how to improve the notes are also wel-
comed.

Cesar O. Aguilar
SUNY Geneseo
 1

Preliminaries

In this short chapter, we will briefly review some basic set notation,
proof methods, functions, and countability. The presentation of these
topics is intentionally brief for two reasons: (1) the reader is likely
familar with these topics, and (2) we include only the necessary material
needed to start doing real analysis.

1.1 Sets, Numbers, and Proofs

Let S be a set. If x is an element of S then we write x ∈ S, otherwise
we write that x ∈/ S. A set A is called a subset of S if each element of
A is also an element of S, that is, if a ∈ A then also a ∈ S. To denote
that A is a subset of S we write A ⊂ S.
Now let A and B be subsets of S. If A ⊂ B and B ⊂ A then A and
B are said to be equal and we write that A = B. The union of A and
B is the set
A ∪ B = {x ∈ S | x ∈ A or x ∈ B}

and the intersection of A and B is the set

A ∩ B = {x ∈ S | x ∈ A and x ∈ B}.

1
 1.1. SETS, NUMBERS, AND PROOFS

S S
A B A B
A∩B
A∪B

Figure 1.1: Set intersection A ∩ B and union A ∪ B

A graphical representation of set unions and intersections are shown in

Figure 1.1.
The empty set is the set that does not contain any elements and
is denoted by ∅. We note that ∅ ⊂ S for any set S. The sets A and B
are disjoint if A ∩ B = ∅. The complement of A in S is the set

S\A = {x ∈ S | x ∈
/ A},

in other words, S\A consists of the elements in S not contained in A.

We sometimes use the shorter notation Ac for S\A when it is clear that
it is the complement of A relative to S.
The Cartesian product of A and B, denoted by A × B, is the set
of ordered pairs (a, b) where a ∈ A and b ∈ B, in other words,

A × B = {(a, b) | a ∈ A, b ∈ B}.

A partition of a set S is a set Π whose elements are subsets of S

such that Π does not contain the empty set, the union of the elements
of Π equals S, and any two distinct elements of Π are disjoint.
Lastly, for any set S, the power set of S is the set of all subsets of
S, and is denoted by P(S).

2
 1.1. SETS, NUMBERS, AND PROOFS

Example 1.1.1. Let A and B be subsets of a set S. Show that

(A ∪ B)c = Ac ∩ B c .

Solution. We first show that (A ∪ B)c ⊂ Ac ∩ B c . If x ∈ (A ∪ B)c then

by definition x ∈
/ (A ∪ B) and therefore x ∈
/ A and x ∈/ B. Thus, x ∈ Ac
and x ∈ B c , that is, x ∈ Ac ∩ B c .
Now suppose that x ∈ Ac ∩ B c , that is, suppose that x ∈ Ac and
x ∈ B c . Thus, x ∈/ A and x ∈ / B and thus x ∈
/ (A ∪ B). By definition,
x ∈ (A ∪ B)c and this proves that Ac ∩ B c ⊂ (A ∪ B)c.

We use the symbol N to denote the set of natural numbers, that

is,
N = {1, 2, 3, 4, . . .}.
The set of integers is denoted by Z so that

Z = {. . . , −3, −2, −1, 0, 1, 2, 3, . . .}.

The set of rational numbers is denoted by

 
p
Q= | p, q ∈ Z, q 6= 0 .
q
Notice that we have the following chain of set inclusions:

N ⊂ Z ⊂ Q.

We now review the most commonly used methods of proof. To that

end, recall that a logical statement is a declarative sentence that can
be unambiguously decided to be either true or false. A theorem is
a logical statement that has been proved to be true using a sequence
of true statements and deductive reasoning. Many theorems are usu-
ally written as a conditional statement of the form “if P then Q” or

3
 1.1. SETS, NUMBERS, AND PROOFS

symbolically “P ⇒ Q”. The statement P is called the hypothesis

or assumption and Q is called the conclusion. Below are the main
techniques used to prove the statement “P ⇒ Q”:

• Direct Proof : To prove the statement “P ⇒ Q”, assume that

the statement P is true and show by combining axioms, defini-
tions, and earlier theorems that Q is true. This should be the
first method you attempt.

• Mathematical Induction: Covered in Section 1.2.

• By Contraposition: Proving the statement “P ⇒ Q” by prov-

ing the logically equivalent statement “not Q ⇒ not P ”. Do not
confuse this with proof by contradiction.

• By Contradiction: To prove the statement “P ⇒ Q” by con-

tradiction, one assumes that “P ⇒ Q” is false and then show that
some contradiction results. Assuming that “P ⇒ Q” is false is
to assume that P is true and Q is false. Using these to latter as-
sumptions, one attempts to derive at a contradiction of the form
“R and not R”, where R is some statement. One disadvantage
with proof by contradiction is that the logical contradiction that
one is seeking “R and not R” is not known in advance so that
the goal of the proof is unclear. Proof by contradiction frequently
gets confused with proof by contraposition in the following way
(do not do this): To prove that “P ⇒ Q”, assume that P is true
and suppose that Q is not true. After some work using only the
assumption that Q is not true you show that P is not true and
thus you say that there is a contradiction because you assumed
that P is true. What you have really done is proved the contra-
positive statement. Thus, if you believe that you are proving a

4
 1.1. SETS, NUMBERS, AND PROOFS

statement by contradiction, take a close look at your proof to see

if what you have is a proof by contraposition.

In the following example, we use both proof by contradiction and

proof by contraposition.

Example 1.1.2. Prove that if x and y are consecutive integers then

x + y is odd.

Solution. Assume that x and y are consecutive integers (i.e. assume

P ) and assume that x + y is not odd (i.e. assume not Q). Since x + y
is not odd then x + y 6= 2n + 1 for all integers n. However, since x and
y are consecutive, and assuming without loss of generality that x < y,
we have that x + y = 2x + 1. Thus, we have that x + y 6= 2n + 1 for all
integers n and also that x + y = 2x + 1. Since x is an integer we have
reached a contradiction. Hence, if x and y are consecutive integers then
x + y is odd.
Now we prove the statement by contraposition. Without loss of
generality, suppose that x < y. Assume that x + y is even. Then there
exists an integer n such that x + y = 2n and therefore x = 2n − y.
Consequently,
y − x = y − (2n − y) = 2(y − n).
Hence, since (y − n) is an integer, y − x 6= 1 and consequently x and y
are not consecutive integers.

In general, if the statement “P ⇒ Q” is true then the converse

conditional statement “Q ⇒ P ” is not necessarily true. For example,
the converse conditional statement in Example 1.1.2 is “if x + y is
odd then x and y are consecutive integers” is easily shown to be false
(e.g., x = 2 and y = 5). The conjoined statement “P ⇒ Q and
Q ⇒ P ”, alternatively written as “P if and only if Q” or symbolically

5
 1.1. SETS, NUMBERS, AND PROOFS

“P ⇔ Q”, is called a biconditional statement. Thus, to prove that

the biconditional statement “P ⇔ Q” is true one must prove that both
“P ⇒ Q” and “Q ⇒ P ” are true.

Example 1.1.3. Let A, B, and C be subsets of a set S. Prove that

(A ∪ B) ⊂ C if and only if A ⊂ C and B ⊂ C.

6
 1.1. SETS, NUMBERS, AND PROOFS

Exercises

Exercise 1.1.1. Let A and B be subsets of a set S. Show that A ⊂ B

if and only if B c ⊂ Ac

Exercise 1.1.2. Find the power set of S = {x, y, z, w}.

Exercise 1.1.3. Let A = {α1, α2 , α3 } and let B = {β1 , β2}. Find

A × B.

Exercise 1.1.4. Let x ∈ Z. Prove that if x2 is even then x is even. Do

not use proof by contradiction.

Exercise 1.1.5. Prove that if x and y are even natural numbers then
xy is even. Do not use proof by contradiction.

Exercise 1.1.6. Prove that if x and y are rational numbers then x + y

is a rational number. Do not use proof by contradiction.

Exercise 1.1.7. Let x and y be natural numbers. Prove that x and y

are odd if and only if xy is odd. Do not use proof by contradiction.

7
 1.2. MATHEMATICAL INDUCTION

1.2 Mathematical Induction

Mathematical induction is a powerful proof technique that relies on the
following property of N.

Axiom 1.2.1: Well-Ordering Principle

Every non-empty subset of N contains a smallest element.

In other words, if S is a non-empty subset of N then there exists a ∈ S

such that a ≤ x for all x ∈ S. The smallest element of S is denoted
by min(S). Thus, min(S) ∈ S and min(S) ≤ x for all x ∈ S. We now
state and prove the principle of Mathematical Induction.

Theorem 1.2.2: Mathematical Induction

Suppose that S is a subset of N with the following properties:

(i) 1 ∈ S

(ii) If k ∈ S then also k + 1 ∈ S.

Then S = N.

Proof. Suppose that S is a subset of N with properties (i) and (ii) and
let T = N\S. Proving that S = N is equivalent to proving that T is
the empty set. Suppose then by contradiction that T is non-empty.
By the well-ordering principle of N, T has a smallest element, say it
is a ∈ T . Because S satisfies property (i) we know that 1 ∈ / T and
therefore a > 1. Now since a is the least element of T , then a − 1 ∈ S
(we know that a − 1 > 0 because a > 1). But since S satisfies property
(ii) then (a − 1) + 1 ∈ S, that is, a ∈ S. This is a contradiction because
we cannot have both a ∈ T and a ∈ S. Thus, T is the empty set, and

8
 1.2. MATHEMATICAL INDUCTION

therefore S = N.

Mathematical induction is frequently used to prove formulas or in-

equalities involving the natural numbers. For example, consider the
validity of the formula

1 + 2 + 3 + · · · + n = 12 n(n + 1) (1.1)

where n ∈ N. In words, the identity (1.1) says that the sum of all the
integers from 1 to n equals 21 n(n + 1). We use induction to show that
this formula is true for all n ∈ N. Let S be the subset of N consisting
of the natural numbers that satisfy (1.1), that is,

S = {n ∈ N | 1 + 2 + 3 + · · · + n = 21 n(n + 1)}.

If n = 1 then
1
2 n(n + 1) = 12 (1 + 1) = 1.
Thus, 12 n(n + 1) is equal to the sum of all the integers from 1 to n = 1.
Hence, (1.1) is true when n = 1 and thus 1 ∈ S. Now suppose that
some k ∈ N satisfies (1.1), that is, suppose that k ∈ S. Then we may
write that
1 + 2 + · · · + k = 21 k(k + 1). (1.2)
We will prove that the integer k + 1 also satisfies (1.1). To that end,
adding k + 1 to both sides of (1.2) we obtain

1 + 2 + · · · + k + (k + 1) = 12 k(k + 1) + (k + 1).

Now notice that we can factor (k + 1) from the right-hand side and
through some algebraic steps we obtain that

1 + 2 + · · · + k + (k + 1) = 21 k(k + 1) + (k + 1)
= (k + 1)[ 12 k + 1]
= 21 (k + 1)(k + 2).

9
 1.2. MATHEMATICAL INDUCTION

Hence, (1.1) also holds for n = k + 1 and thus k + 1 ∈ S. We have

therefore proved that S satisfies properties (i) and (ii), and therefore
by mathematical induction S = N, or equivalently that (1.1) holds for
all n ∈ N.

Example 1.2.3. Use mathematical induction to show that 2n ≤ (n+1)!

holds for all n ∈ N.

Example 1.2.4. Let r 6= 1 be a constant. Use mathematical induction

to show that
2 n 1 − rn+1
1 + r + r + ··· + r =
1−r
holds for all n ∈ N.

Example 1.2.5 (Bernoulli’s inequality). Prove that if x > −1 then

(1 + x)n ≥ 1 + nx for all n ∈ N.

Proof. The statement is trivial for n = 1. Assume that for some k ∈ N

it holds that (1 + x)k ≥ 1 + kx. Since x > −1 then x + 1 > 0 and
therefore

(1 + x)k (1 + x) ≥ (1 + kx)(1 + x)

= 1 + (k + 1)x + kx2

≥ 1 + (k + 1)x.

Therefore, (1 + x)k+1 ≥ 1 + (k + 1)x, and the proof is complete by

induction.

There is another version of mathematical induction called the Prin-

ciple of Strong Induction which we now state.

10
 1.2. MATHEMATICAL INDUCTION

Theorem 1.2.6: Strong Induction

Suppose that S is a subset of N with the following properties:

(i) 1 ∈ S

(ii) If {1, 2, . . . , k} ⊂ S then also k + 1 ∈ S.

Then S = N.

Do you notice the difference between induction and strong induction?

It turns out that the two statements are equivalent, in other words, if S
satisies either one of properties (i)-(ii) of induction or strong induction
then we may conclude that S = N. The upshot with strong induction
is that one is able to use the stronger condition that {1, 2, . . . , k} ⊂ S
to prove that k + 1 ∈ S.

11
 1.2. MATHEMATICAL INDUCTION

Exercises

Exercise 1.2.1. Prove that n < 2n for all n ∈ N.

Exercise 1.2.2. Prove that 2n < n! for all n ≥ 4, n ∈ N.

Exercise 1.2.3. Use induction to prove that if S has n elements then

P(S) has 2n elements. Hint: If S is a set with n + 1 elements, for
instance S = {x1, x2, . . . , xn, xn+1}, then argue that P(S) = P(S̃) ∪ T
where S̃ = S\{xn+1} and T consists of subsets of S that contain xn+1.
How many sets are in P(S̃) and how many are in T ? And what is
P(S̃) ∩ T ? Explain carefully.

12
 1.3. FUNCTIONS

1.3 Functions
Let A and B be sets. A function from A to B is a rule that assigns
to each element x ∈ A one element y ∈ B. The set A is called the
domain of f and B is called the co-domain of f . We usually denote
a function with the notation f : A → B, and the assignment of x to y
is written as y = f (x). We also say that f is a mapping from A to B,
or that f maps A into B. The element y assigned to x is called the
image of x under f . The range of f , denoted by f (A), is the set

f (A) = {y ∈ B | ∃ x ∈ A, y = f (x)}.

In the above definition of f (A), we use the symbol ∃ as a short-hand for

“there exsits”. By definition, f (A) ⊂ B but in general we do not have
that f (A) = B, in other words, the range of a function is generally a
strict subset of the function’s co-domain.
Example 1.3.1. Consider the mapping f : Q → Z defined by
(
1, x≥0
f (x) =
−1, x < 0.
The image of x = 1/2 under f is f (1/2) = 1. The range of f is
f (Q) = {1, −1}.
Example 1.3.2. Consider the function f : N → P(N) defined by

f (n) = {1, 2, . . . , n}.

The set S = {2, 4, 6, 8, . . . , } of even numbers is an element of the co-

domain P(N) but is not in the range of f . As another example, the set
N ∈ P(N) itself is not in the range of f .
Function’s whose range is equal to it’s co-domain are given a special
name.

13
 1.3. FUNCTIONS

Definition 1.3.3: Surjection

A function f : A → B is said to be a surjection if for any y ∈ B
there exists x ∈ A such that f (x) = y.

In other words, f : A → B is a surjection if f (A) = B.

Example 1.3.4. The function f : Q → Q defined by f (x) = x2 is not

a surjection. For example, y = −1 is clearly not in the range of f since
f (x) = x2 6= −1 for all x ∈ Q. On the other hand, y = 121 64 is in the
range of f since f (11/8) = 121
64
. Is y = 2 in the range of f ?

Example 1.3.5. Consider the function f : P(N) → N defined by

f (S) = min(S).

Prove that f is a surjection.

Solution. To prove that f is a surjection, we must show that for any

element y ∈ N (the co-domain), there exists S ∈ P(N) (the domain)
such that f (S) = y. Consider then an arbitrary y ∈ N. Let S = {y}
and thus clearly S ∈ P(N). Moreover, it is clear that min(S) = y and
thus f (S) = min(S) = y. This proves that f is a surjection.

Notice that in Example 1.3.5, given any y ∈ N there are many sets
S ∈ P(N) such that f (S) = y. This leads us to the following definition.

Definition 1.3.6: Injection

A function f : A → B is said to be an injection if no two distinct
elements of A are mapped to the same element in B, in other words,
for any x1, x2 ∈ A, if x1 6= x2 then f (x1) 6= f (x2).

14
 1.3. FUNCTIONS

In other words, f is an injection if whenever f (x1) = f (x2) then neces-

sarily x1 = x2.

Example 1.3.7. The function f : Q → Q defined by f (x) = x2 is not

an injection. For example, f (−2) = f (2) = 4.

Example 1.3.8. Consider again the function f : N → P(N) defined

by
f (n) = {1, 2, . . . , n}.
This function is an injection. Indeed, if f (n) = f (m) then {1, 2, . . . , n} =
{1, 2, . . . , m} and therefore n = m. Hence, whenever f (n) = f (m) then
necessarily n = m and this proves that f is an injection.

Example 1.3.9. Consider the function f : P(N) → N defined by

f (S) = min(S)

Is f an injection?

Example 1.3.10. Consider the function f : N → N × N defined by

f (n) = (2n2, n + 1)

Show that f is an injection.

Definition 1.3.11: Bijection

A function f : A → B is said to be a bijection if it is a surjection
and an injection.

Example 1.3.12. Suppose that f : P → Q is an injection. Prove that

the function f˜ : P → f (P ) defined by f˜(x) = f (x) for x ∈ P is a
bijection.

15
 1.3. FUNCTIONS

Solution. By construction, f˜ is a surjection. If f˜(x) = f˜(y) then f (x) =

f (y) and then x = y since f is an injection. Thus, f˜ is an injection and
this proves that f˜ is a bijection.

Example 1.3.13. Prove that f : Q\{0} → Q\{0} defined by f ( pq ) = q

p
is a bijection, where gcd(p, q) = 1.

Suppose that f : A → B is a bijection and define the function g :

B → A as follows: for b ∈ B let g(b) be the (necessarily unique) element
in A such that f (g(b)) = b. Notice that by definition, g(f (a)) = a. The
function g is called the inverse of f and we write instead g = f −1. It
is not hard to show that g is a bijection and that g −1 = f .
Given functions f : A → B and g : B → C, the composition of g
and f is the function (g ◦ f ) : A → C defined as (g ◦ f )(a) = g(f (a)).

Theorem 1.3.14
If f : A → B and g : B → C are injections (surjections) then the
composition (g ◦ f ) : A → C is an injection (surjection).

Proof. Assume that f : A → B and g : B → C are injections. To

prove that (g ◦ f ) is an injection, suppose that (g ◦ f )(x1) = (g ◦ f )(x2).
Then by definition of (g ◦ f ), it follows that g(f (x1)) = g(f (x2)). Now
since g is an injection then necessarily f (x1) = f (x2) and since f is an
injection then necessarily x1 = x2. Thus if (g ◦ f )(x1) = (g ◦ f )(x2)
then x1 = x2 and this proves that (g ◦ f ) is an injection.
Now suppose that f and g are surjections. To prove that (g ◦ f ) :
A → C is a surjection, let z ∈ C be arbitrary. Since g is a surjection,
there exists y ∈ B such that g(y) = z. Since f is a surjection, there
exists x ∈ A such that y = f (x). Thus, for x ∈ A we have that
(g ◦ f )(x) = g(f (x)) = g(y) = z. Hence, for arbitrary z ∈ C there

16
 1.3. FUNCTIONS

exists x ∈ A such that (g ◦ f )(x) = z and this proves that (g ◦ f ) is a

surjection.

The following result is then an immediate application of Theorem 1.3.14

and the definition of a bijection.

Corollary 1.3.15
The composition of two bijections is a bijection.

17
 1.3. FUNCTIONS

Exercises

Exercise 1.3.1. Consider the function f : N → Q defined as f (n) = n1 .

Is f an injection? Is f a surjection?

Exercise 1.3.2. Consider the function f : N × N → N defined as

f (n, m) = nm. Is f an injection? Is f a surjection?

Exercise 1.3.3. Consider the function f : Q → Q defined as f (x) =

(x − 2)(x − 6). Is f an injection? If f a surjection?

Exercise 1.3.4. Let f : N → Z be the function defined as

(
n
, if n is even,
f (n) = 2 (n−1)
− 2 , if n is odd.

Prove that f is a bijection.

Exercise 1.3.5. The sign of a rational number x ∈ Q is defined as

sgn(x) = x/|x| if x 6= 0, where |x| is the absolute value of x, and
sgn(0) = 1. For example, sgn(−3) = −1 and sgn(2) = 1. Prove that
the function f : Z → {−1, 1} × N defined as

f (x) = (sgn(x), |x| + 1)

is a bijection.

18
 1.4. COUNTABILITY

1.4 Countability
A non-empty set S is said to be finite if there is a bijection from
{1, 2, . . . , n} onto S for some n ∈ N. In this case, we say that S
contains n elements and we write |S| = n. If S is not finite then we say
that S is infinite.

Example 1.4.1. Let f : P → Q be an injection. If P is an infinite set

then f (P ) is an infinite set.

Solution. The proof is by contraposition. Suppose that f (P ) is a

finite set containing n elements. Then there exists a bijection g :
{1, 2, . . . , n} → f (P ). The function f˜ : P → f (P ) defined as f˜(x) =
f (x) for x ∈ P is a bijection (Example 1.3.12) and therefore (f˜−1 ◦ g) :
{1, 2, . . . , n} → P is a bijection. Thus P is a finite set and completes
the proof.

We now introduce the notion of a countable set.

Definition 1.4.2: Countability

Let S be a set.

(i) The set S is countably infinite if there is a bijection from

N onto S.

(ii) The set S is countable if S is either finite or countably infi-

nite.

(iii) The set S is uncountable if S is not countable.

Roughly speaking, a set S is countable if the elements of S can be

listed or enumerated in a systematic manner. To see how, suppose

19
 1.4. COUNTABILITY

that S is countably infinite and let f : N → S be a bijection. Then the

elements of S can be listed as

S = {f (1), f (2), f (3), f (4), . . .}.

Hence, although sets have no predetermined order, the elements of a

countable set can be ordered.
Example 1.4.3. The set S of odd natural numbers is countable. Recall
that n ∈ N is an odd positive integer if n = 2k − 1 for some k ∈ N. A
bijection f : N → S from N to S is f (k) = 2k − 1. The function f can
be interpreted as a listing of the odd natural numbers in the natural
way:
S = {f (1), f (2), f (3), . . .} = {1, 3, 5, . . .}.
Example 1.4.4. The natural numbers S = N are countable. A bijec-
tion f : N → S is f (n) = n, i.e., the identity mapping.
Example 1.4.5. The set of integers Z is countable. A bijection f from
N to Z can be defined by listing the elements of Z as follows:
N : 1 2 3 4 5 6 7 ...
↓ ↓ ↓ ↓ ↓ ↓ ↓ ...
Z : 0 1 −1 2 −2 3 −3 . . .
To be more precise, f is the function
(
n
, if n is even,
f (n) = 2 (n−1)
− 2 , if n is odd.
It is left as an exercise to show that f is indeed a bijection.
Example 1.4.6. The set N×N is countable. There are many bijections
from N to N×N but a particularly interesting one is the function defined
as follows. Consider the family of lines

Lk = {(x, y) ∈ N × N | y = −x + k + 1}

20
 1.4. COUNTABILITY

for k ∈ N. There are k points on the line Lk , namely (j, k + 1 − j) for

1 ≤ j ≤ k, and we say that (j, k + 1 − j) is the jth point on the line Lk .
The point (x, y) ∈ N×N is contained in the line Lk where k = x+y −1.
Thus, one way to enumerate the points in N × N is to assign to each
(x, y) ∈ Lk the number ρ(x, y) obtained by adding all points on the
lines L1, . . . , Lk−1 and adding the position of (x, y) on line Lk , namely
x. Thus,

ρ(x, y) = 1 + 2 + . . . + (k − 1) + x
1
= (k − 1)k + x
2
1
= (x + y − 2)(x + y − 1) + x.
2
The function ρ is called the Cantor pairing function. Alternatively,
we use a modified version of ρ which we call τ : N × N → N and defined
as (
ρ(x, y), if (x + y − 1) is odd,
τ (x, y) =
ρ(y, x), if (x + y − 1) is even.
We now find a formula for the inverse of τ which we call the Cantor
snake. To write down the formula for τ −1 , we first let for each n ∈ N
p !
−1 + 1 + 8(n − 1)
m = floor
2

and we note that m ≥ 0 is the smallest integer such that 12 m(m+1) < n.
We then set p(n) = n − 21 m(m + 1) a then
(
(p(n), −p(n) + m + 2), if m is even
τ −1 (n) =
(−p(n) + m + 2, p(n)), if m is odd.
We note that the point (x, y) = τ −1(n) is on the line Lk with k = m+1.
The range of τ −1 is

τ −1(N) = {(1, 1), (2, 1), (1, 2), (1, 3), (2, 2), (3, 1), . . .}.

21
 1.4. COUNTABILITY

55
10

37 54
9

36 38 53
8

22 35 39 52
7

21 23 34 40 51
6

11 20 24 33 41 50
5

10 12 19 25 32 42 49
4

4 9 13 18 26 31 43 48
3

3 5 8 14 17 27 30 44 47
2

1 2 6 7 15 16 28 29 45 46
1

1 2 3 4 5 6 7 8 9 10

Figure 1.2: The image of the Cantor snake

The sequence of pairs (x, y) ∈ N × N generated by τ −1 for 1 ≤ n ≤ 55

are shown in Figure 1.2.

Example 1.4.7. Suppose that f : T → S is a bijection. Prove that T

is countable if and only if S is countable.

Solution. Suppose that T is countable. Then by definition there exists

a bijection g : N → T . Since g and f are bijections, the composite
function (f ◦ g) : N → S is a bijection. Hence, S is countable.
Now suppose that S is countable. Then by definition there exists
a bijection h : N → S. Since f is a bijection then the inverse function
f −1 : S → T is also a bijection. Therefore, the composite function
(f −1 ◦ h) : N → T is a bijection. Thus, T is countable.

As the following theorem states, countability, or lack thereof, can

22
 1.4. COUNTABILITY

be inherited via set inclusion.

Theorem 1.4.8: Inheriting Countability

Let S and T be sets and suppose that T ⊂ S.

(i) If S is countable then T is also countable.

(ii) If T is uncountable then S is also uncountable.

Proof. To prove (i), let S be a countable set and let f : S → N be

a bijection. Define the mapping f˜ : T → f (T ) by f˜(x) = f (x). By
Example 1.3.12, f˜ is a bijection. Therefore, since f (T ) is a subset of
N, and thus countable, then T is countable. The proof of (ii) is left as
an exercise.

Example 1.4.9. Let S be the set of odd natural numbers. In Exam-

ple 1.4.3, we proved that the odd natural numbers are countable by
explicitly constructing a bijection from N to S. Alternatively, since N
is countable and S ⊂ N then by Theorem 1.4.8 S is countable. More
generally, any subset of N is countable.

If S is known to be a finite set then by Definition 1.4.2 S is countable,

while if S is infinite then S may or may not be countable (we have yet to
encounter an uncountable set but soon we will). To prove that a given
infinite set S is countable we could use Theorem 1.4.8 if it is applicable
but otherwise we must use Definition 1.4.2, that is, we must show that
there is a bijection from S to N, or equivalently from N to S. However,
suppose that we can only prove the existence of a surjection f from N
to S. The problem might be that f is not an injection and thus not a
bijection. However, the fact that f is a surjection from N to S somehow
says that S is no “larger” than N and gives evidence that perhaps S is

23
 1.4. COUNTABILITY

countable. Could we use a surjection f : N → S to construct a bijection

g : N → S? Or, what if instead we had an injection g : S → N; could
we use g to construct a bijection f : S → N? The following theorem
says that it is indeed possible to do both.

Theorem 1.4.10: Countability Relaxations

Let S be a set.

(i) If there exists an injection g : S → N then S is countable.

(ii) If there exists a surjection f : N → S then S is countable.

Proof. (i) Let g : S → N be an injection. Then the function g̃ : S →

g(S) defined by g̃(s) = g(s) for s ∈ S is a bijection. Since g(S) ⊂ N
then g(S) is countable. Therefore, S is countable also.
(ii) Now let f : N → S be a surjection. For s ∈ S let f −1(s) = {n ∈
N | f (n) = s}. Since f is a surjection, f −1(s) is non-empty for each
s ∈ S. Consider the function h : S → N defined by h(s) = min f −1(s).
Then f (h(s)) = s for each s ∈ S. We claim that h is an injection.
Indeed, if h(s) = h(t) then f (h(s)) = f (h(t)) and thus s = t, and the
claim is proved. Thus, h is an injection and then by (i) we conclude
that S is countable.

We must be careful when using Theorem 1.4.10; if f : N → S is

known to be an injection then we cannot conclude that S is countable
and similarly if f : S → N is known to be a surjection then we cannot
conclude that S is countable.

Example 1.4.11. In this example we will prove that the union of

countable sets is countable. Hence, suppose that A and B are count-
able. By definition, there exist bijections f : N → A and g : N → B.

24
 1.4. COUNTABILITY

Consider the function h : N → A ∪ B defined as follows:

(
f ((n + 1)/2), if n is odd,
h(n) =
g(n/2), if n is even.

We claim that h is a surjection (Loosely speaking, if A = {a1, a2 , . . . , }

and B = {b1, b2, . . . , }, then the function h lists the elements of A ∪ B
as A ∪ B = {a1 , b1, a2, b2, a3 , b3, . . . , }.). To see this, let x ∈ A ∪ B. If
x ∈ A then x = f (k) for some k ∈ N. Then h(2k − 1) = f (k) = x. If
on the other hand x ∈ B then x = g(k) for some k ∈ N. Then h(2k) =
g(k) = x. In either case, there exists n ∈ N such that h(n) = x, and
thus h is a surjection. By Theorem 1.4.10, the set A ∪ B is countable.
This example can be generalized as follows. Let A1, A2, A3, . . . , be
S
countable sets and let S = ∞ k=1 Ak . Then S is countable. To prove
this, we first enumerate the elements of each Ak as follows:

A1 = {a1,1, a1,2, a1,3 , . . .}

A2 = {a2,1, a2,2, a2,3 , . . .}
A3 = {a3,1, a3,2, a3,3 , . . .}
··· ··· ···

Formally, we have surjections fk : N → Ak for each k ∈ N. Consider

the mapping ϕ : N × N → S defined by

ϕ(m, n) = am,n = fm (n).

It is clear that ϕ is a surjection. Since N × N is countable, there is a

surjection φ : N → N×N, and therefore the composition (ϕ◦φ) : N → S
is a surjection. Therefore, S is countable.

The following theorem is perhaps surprising.

25
 1.4. COUNTABILITY

Theorem 1.4.12
The set of rational numbers Q is countable.

Proof. Let Q>0 be the set of all positive rational numbers and let Q<0 be
the set of all negative rational numbers. Clearly, Q = Q<0 ∪ {0} ∪ Q>0,
and thus it is enough to show that Q<0 and Q>0 are countable. In fact,
we have the bijection h : Q>0 → Q<0 given by h(x) = −x, and thus if
we can show that Q>0 is countable then this implies that Q<0 is also
countable. In summary, to show that Q is countable it is enough to
show that Q>0 is countable. To show that Q>0 is countable, consider
the function f : N × N → Q>0 defined as
p
f (p, q) = .
q

By definition, any rational number x ∈ Q>0 can be written as x = pq

for some p, q ∈ N. Hence, x = f (p, q) and thus x is in the range of f .
This shows that f is a surjection. Now, because N × N is countable,
there is a surjection g : N → N × N and thus (f ◦ g) : N → Q>0 is a
surjection. By Theorem 1.4.10, this proves that Q>0 is countable and
therefore Q is countable.

We end this section with Cantor’s theorem named after mathemati-

cian Georg Cantor (1845-1918). Cantor is considered as the creator
of set theory. Originally interested in analytical problems having as
their root problems in physics, and in particular in characterizing solu-
tions to equations describing heat conduction, Cantor discovered that
infinite sets come in many possible sizes. One of Cantor’s fascinating
discoveries, which initially were very controversial at the time, led to
the following theorem [1].

26
 1.4. COUNTABILITY

Theorem 1.4.13: Cantor’s Theorem (1891)

For any set S, there is no surjection of S onto the power set P(S).

Proof. Suppose by contradiction that f : S → P(S) is a surjection. By

definition, for each a ∈ S, f (a) is a subset of S. Consider the set

C = {a ∈ S | a ∈
/ f (a)}.

Since C is a subset of S then C ∈ P(S). Since f is a surjection, there

exists x ∈ S such that C = f (x). One of the following must be true:
either x ∈ C or x ∈/ C. If x ∈ C then x ∈ / f (x) by definition of C. But
C = f (x) and thus we reach contradiction. If x ∈ / C then by definition
of C we have x ∈ f (x). But C = f (x) and thus we reach a contradiction.
Hence, neither of the possibilities are true, and thus we have reached an
absurdity. Hence, we conclude that there is no such surjection f .

Cantor’s theorem implies that P(N) is uncountable. Indeed, if we

take S = N in Cantor’s Theorem then there is no surjection from N to
P(N), and thus certainly no bijection from N to P(N). In summary:

Corollary 1.4.14
The set P(N) is uncountable.

27
 1.4. COUNTABILITY

Exercises

Exercise 1.4.1. Let P and Q be infinite sets. Prove that if f : Q → P

is a bijection then Q is uncountable if and only if P is uncountable. Do
not use proof by contradiction.

Exercise 1.4.2. In this exercise you will provide another proof that
N × N is countable.

(a) Prove that 3n is odd for each n ∈ N.

(b) Consider the function f : N × N → N defined as f (p, q) = 2p 3q .

Prove that f is an injection. Hint: Use part (a) in some way.

(c) Explain how part (b) proves that N × N is countable.

Exercise 1.4.3. Prove that if A and B are countably infinite then

A × B is countable.

Exercise 1.4.4. Recall that a sequence a = {ak }∞ k=1 of numbers is an

infinite list
a = (a1 , a2, a3 , a4, . . .)
where each element ak is a number (We will cover sequences in detail
but you are already familiar with them from calculus.). Let Q be the
set of sequences whose elements are either 0 or 1, in other words,

Q = {{ak }∞
k=1 | ak = 0 or ak = 1} .

For example, the following sequences are elements of the set Q:

a = (0, 0, 0, 0, 0, 0, . . .) ∈ Q
b = (1, 0, 1, 0, 1, 0, . . .) ∈ Q
c = (0, 0, 1, 1, 0, 0, . . .) ∈ Q
d = (1, 1, 1, 1, 1, 1, . . .) ∈ Q

28
 1.4. COUNTABILITY

(a) Give two non-trivial examples of functions from N to Q.

(b) Consider the function f : Q → P(N) defined as

f (a) = {k ∈ N | ak = 1}

where P(N) is the power set of N. Hence, f takes a sequence

a ∈ Q and outputs the indices where a is equal to 1. For example:

f (0, 0, 0, 0, 0, 0, 0, 0, . . .) = ∅

f (1, 0, 1, 0, 1, 0, 1, 0, . . .) = {1, 3, 5, 7, . . .}

f (0, 0, 1, 1, 0, 0, 0, 0, . . .) = {3, 4}

f (1, 1, 1, 1, 1, 1, 1, 1, . . .) = {1, 2, 3, 4, 5, 6, . . .} = N

Prove that f : Q → P(N) is a bijection.

(c) Combine part (b), Exercise 1.4.1, and Cantor’s theorem to thor-
oughly explain whether Q is countable or uncountable.

29
1.4. COUNTABILITY

30
 2

The Real Numbers

2.1 Introduction
Recall that a rational number is a number x that can be written in
the form x = pq where p, q are integers with q 6= 0. The rational
number system is all you need to accomplish most everyday tasks. For
instance, to measure distances when building a house it suffices to use
1
a tape measure with an accuracy of about 16 of an inch. However, to
do mathematical analysis the rational numbers have some very serious
shortcomings; here is a an example.

Theorem 2.1.1

If x2 = 2 then x is not a rational number.

Proof. Suppose by contradiction that there exists x ∈ Q such that

x2 = 2. We may write x = pq for some integers p, q, and we can assume
that p and q have no common factor other than 1 (that is, p and q are
relatively prime). Now, since x2 = 2 then p2 = 2q 2 and thus p2 is an
even number. This implies that p is also even. Since p is even, we may
write p = 2k for some k ∈ N and therefore (2k)2 = 2q 2, from which
it follows that 2k 2 = q 2 . Hence, q 2 is even and thus q is also even.

31
 2.1. INTRODUCTION

Thus, both p and q are even, which contradicts the fact that p and q
are relatively prime.

The previous theorem highlights that the set of rational numbers are in
some sense incomplete, or that there are gaps in Q, and that a larger
number system is needed to enlarge the set of math problems that
can be analyzed and solved. Although mathematicians in the 1700s
were using the real number system and resorting to limiting processes
to analyze problems in physics, it was not until the late 1800s that
mathematicians gave a rigorous construction of the real number system.
Motivated by Theorem 2.1.1, we might be tempted to define the real
numbers as the set of solutions of all polynomial equations with integer
coefficients. As it turns out, however, this definition of the reals would
actually leave out almost all the real numbers, including some of our
favorites like π and e. In fact, the set of all numbers that are solutions
to polynomial equations with rational coefficients is countable!
There are two standard ways to construct the set of real numbers.
One standard method to construct R uses the notion of Cauchy se-
quences of rational numbers and is attributed to Georg Cantor [2].
We will cover Cauchy sequences in Section 3.6 and therefore postpone
describing some of the details of the construction until then. The sec-
ond standard method to construct the reals relies on the notion of a
Dedekind cut and is attributed to Richard Dedekind (1831-1916). A
Dedekind cut is a partition {A, B} of Q such that both A and B are
non-empty and
(i) if b ∈ A and a < b then a ∈ A, and

(ii) for any a ∈ A there exists b ∈ A such that a < b.

The set of real numbers R is then defined to be the set of all Dedekind
cuts. For example, let A = {x ∈ Q | x2 < 2 or x < 0} and thus

32
 2.1. INTRODUCTION

B = Q\A. Then one can show that {A, B} is a Dedekind cut of Q

and the idea is that x = {A, B} represents the real number x such that
√
x2 = 2, that is, the irrational number 2. Having defined R as the set
of Dedekind cuts we then proceed to define all the usual operations of
arithmetic and arrive at the familiar model of R [2]. Additionally, if
x = {A, B} and y = {C, D} then we write that x ≤ y if A ⊂ C and
write x < y if A ⊂ C and A 6= C. Refer to [2] for further details.
In this book, we instead adopt the familiar viewpoint that the real
numbers R are in a one-to-one correspondence with the points on an
infinite line:

ℝ
⋯ -4 -3 -2 -1 0 1 2 3 4 ⋯

Figure 2.1: The real numbers are in a one-to-one correspondence with

points on an infinite line

The essential feature that we want to capture with this view of R is that
there are no “holes” in the real number system. This view of R allows us
to quickly start learning the properties of R instead of focusing on the
details of constructing a model for R. Naturally, the rational numbers
Q are a subset of R and we say that a real number x ∈ R is irrational
if it is not rational. As we saw in Thereom 2.1.1, the positive number
x ∈ R such that x2 = 2 is irrational.

33
 2.1. INTRODUCTION

Exercises

Exercise 2.1.1. Let x ∈ Q be fixed. Prove the following statements

without using proof by contradiction.

(a) Prove that if y ∈ R\Q then x + y ∈ R\Q.

(b) Suppose in addition that x > 0. Prove that if y ∈ R\Q then

xy ∈ R\Q.

Exercise 2.1.2. Prove that if 0 < x < 1 then xn < x for all natural
numbers n ≥ 2. Do not assume that x is rational.

34
 2.2. ALGEBRAIC AND ORDER PROPERTIES

2.2 Algebraic and Order Properties

We will soon see the main difference between Q and R from an analysis
point of view but in this section we will discuss one important thing
that Q and R have in common, namely, both are ordered fields. We
begin then with the definition of a field.

Definition 2.2.1
A field is a set F with two binary operations + : F × F → F
and × : F × F → F, the former called addition and the latter
multiplication, satisfying the following properties:

(i) a + b = b + a for all a, b ∈ F

(ii) (a + b) + c = a + (b + c) for all a, b, c ∈ F

(iii) a × (b + c) = a × b + a × c

(iv) There exists an element 1 ∈ F such that a × 1 = 1 × a = a for

all a ∈ F

(v) There exists an element 0 ∈ F such that a + 0 = 0 + a = a for

all a ∈ F

(vi) For each a ∈ F, there exists an element −a ∈ F such that

a + (−a) = (−a) + a = 0.

(vii) For each a ∈ F, there exists an element a−1 ∈ F such that

a × a−1 = a−1 × a = 1.

Example 2.2.2. It is not hard to see that N and Z are not fields. In
each case, what property of a field fails to hold?

35
 2.2. ALGEBRAIC AND ORDER PROPERTIES

Example 2.2.3. Both Q and R are fields.

Besides being fields, both Q and R are totally ordered sets. By

totally ordered we mean that for any a, b ∈ R either a = b, a < b, or
b < a. This property of R is referred to as the Law of Trichotomy. For
a, b ∈ R, the relation a ≤ b means that either a < b or a = b. Similarly,
a ≥ b means that b ≤ a. From our number line viewpoint of R, if a < b
then a is on the left of b.
We now present some very important rules of inequalities that we
will use frequently in this book.

Theorem 2.2.4
Let a, b, c ∈ R.

(i) If a < b and b < c then a < c. (transitivity)

(ii) If a < b then a + c < b + c.

(iii) If a < b and c > 0 then ac < bc.

(iv) If a < b and c < 0 then ac > bc.

(v) If ab > 0 then either a, b > 0 or a, b < 0.

(vi) If a 6= 0 then a2 > 0.

The two inequalities a ≤ b and b ≤ c as sometimes combined as

a ≤ b ≤ c.

Example 2.2.5. Suppose that a > 0 and b > 0, or written more

compactly as a, b > 0. Prove that if a < b then 1b < a1 .

36
 2.2. ALGEBRAIC AND ORDER PROPERTIES

Example 2.2.6. Suppose that a ≤ x and b ≤ y. Prove that a + b ≤

x + y. Deduce that if a ≤ x ≤ ξ and b ≤ y ≤ ζ then

a + b ≤ x + y ≤ ζ + ξ.

We will encounter situations where we will need to prove that if two

numbers a, b ∈ R satisfy a certain property then a = b. Proving that
a = b is equivalent to proving that x = a − b = 0. In such situations,
the following theorem will be very useful.

Theorem 2.2.7
Let x ∈ R be non-negative, that is, x ≥ 0. If for every ε > 0 it
holds that x < ε then x = 0.

Proof. We prove the contrapositive, that is, we prove that if x > 0 then
there exists ε > 0 such that ε < x. Assume then that x > 0 and let
ε = x2 . Then ε > 0 and clearly ε < x.

The next few examples will give us practice with working with in-
equalities.

Example 2.2.8. Let ε = 0.0001. Find a natural number n ∈ N such

that
1
<ε
n+1
Example 2.2.9. Let ε = 0.0001. Find analytically a natural number
n ∈ N such that
n+2
<ε
n2 + 3
Example 2.2.10. Let ε = 0.001. Find analytically a natural number
n ∈ N such that
5n − 4
<ε
2n3 − n

37
 2.2. ALGEBRAIC AND ORDER PROPERTIES

Exercises

Exercise 2.2.1. Let ε = 0.0001 and find analytically a natural number

n ∈ N such that
3n − 2
< ε.
n3 + 2n
Exercise 2.2.2. Let ε = 0.0001 and find analytically a natural number
n ∈ N such that
3n + 2
< ε.
4n3 − n
Exercise 2.2.3. Let ε = 0.0001 and find analytically a natural number
n ∈ N such that
cos2 (3n) + 1
< ε.
arctan(n) + n

38
 2.3. THE ABSOLUTE VALUE

2.3 The Absolute Value

To solve problems in calculus you need to master differentiation and
integration. To solve problems in analysis you need to master inequali-
ties. The content of this section, mostly on inequalities, is fundamental
to everything else that follows in this book.
Given any a ∈ R, we define the absolute value of a as the number
(
a, if a ≥ 0
|a| :=
−a, if a < 0.

Clearly, |a| = 0 if and only if a = 0, and 0 ≤ |a| for all a ∈ R. Below

we record some important properties of the absolute value function.

Theorem 2.3.1
Let a, b ∈ R and c ≥ 0.

(i) |ab| = |a| · |b|

(ii) |a|2 = a2

(iii) |a| ≤ c if and only if −c ≤ a ≤ c

(iv) −|a| ≤ a ≤ |a|

Proof. Statements (i) and (ii) are trivial. To prove (iii), first suppose
that |a| ≤ c. If a > 0 then a ≤ c. Hence, −a ≥ −c and since a > 0
then a > −a ≥ −c. Hence, −c ≤ a ≤ c. If a < 0 then −a ≤ c, and
thus a ≥ −c. Since a < 0 then a < −a ≤ c. Thus, −c ≤ a ≤ c. Now
suppose that −c ≤ a ≤ c. If 0 < a ≤ c then |a| = a ≤ c. If a < 0
then from multiplying the inequality by (−1) we have c ≥ −a ≥ −c
and thus |a| = −a ≤ c.

39
 2.3. THE ABSOLUTE VALUE

To prove part (iv), notice that |a| ≤ |a| and thus applying (iii) we
get −|a| ≤ a ≤ |a|.
a |a|
Example 2.3.2. If b 6= 0 prove that b = |b| .

Example 2.3.3. From Theorem 2.3.1 part (i), we have that |a2 | =
|a · a| = |a| · |a| = |a|2 . Therefore, |a2 | = |a|2 . Similarly, one can show
that |a3 | = |a|3 . By induction, for each n ∈ N it holds that |an | = |a|n .

Below is the most important inequality in this book.

Theorem 2.3.4: Triangle Inequality

For any x, y ∈ R it holds that

|x + y| ≤ |x| + |y|.

Proof. We have that

−|x| ≤ x ≤ |x|
−|y| ≤ y ≤ |y|

from which it follows that

−(|x| + |y|) ≤ x + y ≤ |x| + |y|

and thus
|x + y| ≤ |x| + |y|.

By induction, one can prove the following corollary to the Triangle

inequality.

40
 2.3. THE ABSOLUTE VALUE

Corollary 2.3.5
For any x1, x2, . . . , xn ∈ R it holds that

|x1 + x2 + · · · + xn | ≤ |x1 | + |x2 | + · · · + |xn|.

A compact way to write the triangle inequality using summation nota-

tion is
Xn Xn
xi ≤ |xi|.
i=1 i=1
Here is another consequence of the Triangle inequality.

Corollary 2.3.6
For x, y ∈ R it holds that

(i) |x − y| ≤ |x| + |y|

(ii) ||x| − |y|| ≤ |x − y|

Proof. For part (i), we have

|x − y| = |x + (−y)|
≤ |x| + | − y|
= |x| + |y|.

For part (ii), consider

|x| = |x − y + y| ≤ |x − y| + |y|

and therefore |x| − |y| ≤ |x − y|. Switching the role of x and y we

obtain |y| − |x| ≤ |y − x| = |x − y|, and therefore multiplying this last
inequality by −1 yields −|x − y| ≤ |x| − |y|. Therefore,

−|x − y| ≤ |x| − |y| ≤ |x − y|

41
 2.3. THE ABSOLUTE VALUE

which is the stated inequality.

Example 2.3.7. For a, b ∈ R prove that |a + b| ≥ |a| − |b|.

Example 2.3.8. Let f (x) = 2x2 − 3x + 7 for x ∈ [−2, 2]. Find a

number M > 0 such that |f (x)| ≤ M for all −2 ≤ x ≤ 2.

Solution. Clearly, if −2 ≤ x ≤ 2 then |x| ≤ 2. Apply the triangle

inequality and the properties of the absolute value:

|f (x)| = |2x2 − 3x + 7|
≤ |2x2| + |3x| + |7|
= 2|x|2 + 3|x| + 7
≤ 2(2)2 + 3(2) + 7
= 21.

Therefore, if M = 21 then |f (x)| ≤ M for all x ∈ [−2, 2].

2
Example 2.3.9. Let f (x) = 2x 1−2x
+3x+1
. Find a number M > 0 such that
|f (x)| ≤ M for all 2 ≤ x ≤ 3.

Solution. It is clear that if 2 ≤ x ≤ 3 then |x| ≤ 3. Using the proper-

ties of the absolute value and the triangle inequality repeatedly on the

42
 2.3. THE ABSOLUTE VALUE

numerator:
2x2 + 3x + 1
|f (x)| =
1 − 2x

|2x2 + 3x + 1|
=
|1 − 2x|

|2x2| + |3x| + |1|

≤
|1 − 2x|

2|x|2 + 3|x| + 1|
=
|1 − 2x|

2 · 32 + 3 · 3 + 1
≤
|1 − 2x|

28
= .
|2x − 1|

Now, for 2 ≤ x ≤ 3 we have that −5 ≤ 1 − 2x ≤ −3 and therefore

1
3 ≤ |1 − 2x| ≤ 5 and then |2x−1| ≤ 13 . Therefore,

28 28
|f (x)| ≤ ≤ .
|1 − 2x| 3
28
Hence, we can take M = 3.

Example 2.3.10. Let f (x) = sin(2x)−3

x2 +1
. Find a number M > 0 such
that |f (x)| ≤ M for all −3 ≤ x ≤ 2.

In analysis, the absolute value is used to measure distance between

points in R. For any a ∈ R, the absolute value |a| is the distance
from a to 0. This interpretation of the absolute value can be used to
measure the difference (in magnitude) between two points. That is,

43
 2.3. THE ABSOLUTE VALUE

|x − y| = |y − x|
ℝ
y x

Figure 2.2: The number |x − y| is the distance between x and y.

given x, y ∈ R, the distance between x and y is |x − y|. From the

properties of the absolute value, this distance is also |y − x|.
We will often be concerned with how close a given number x is
to a fixed number a ∈ R. To do this, we introduce the notion of
neighborhoods based at a.

Definition 2.3.11: Neighborhoods

Let a ∈ R and let ε > 0. The ε-neighborhood of a of radius is
the set

Bε (a) = {x ∈ R | |x − a| < ε} = (a − ε, a + ε)

Notice that if ε1 < ε2 then Bε1 (a) ⊂ Bε2 (a).

Example 2.3.12. Let f (n) = 3n+1

2n+3 and let ε = 0.0001. From calculus,
we know that lim f (n) = 32 . Find a natural number N such that
n→∞
|f (n) − 32 | < ε for every n ≥ N .

Solution. The inequality |f (n) − 23 | < ε means that f (n) is within ε of

3
2 . That is,
3 3
− ε < f (n) < + ε.
2 2
This inequality does not hold for all n, but it will eventually hold for
some N ∈ N and for all n ≥ N . For example, f (1) = 54 = 0.8 and
|f (1)− 23 | = 0.7 > ε, and similarly for f (2) = 77 = 1 and |f (2)− 23 | = 0.5.

44
 2.3. THE ABSOLUTE VALUE

In fact, |f (20) − 32 | = 0.081 > ε. However, because we know that

limn→∞ f (n) = 23 , eventually |f (n) − 23 | < ε for large enough n. To find
out how large, let’s analyze the magnitude |f (n) − 32 |:

3n + 1 3
|f (n) − 23 | = −
2n + 3 2

6n + 2 − 6n − 9
=
2(2n + 3)
7
=
2(2n + 3)

Hence, |f (n) − 23 | < ε if and only if

7
<ε
2(2n + 3)
which after re-arranging can be written as
7 3
n> − .
4ε 2
With ε = 0.0001 we obtain

n > 17, 498.5.

Hence, if N = 17, 499 then if n ≥ N then |f (n) − 23 | < ε.

Example 2.3.13. Let ε1 > 0 and ε2 > 0, and let a ∈ R. Show that
Bε1 (a) ∩ Bε2 (a) and Bε1 (a) ∪ Bε2 (a) are ε-neighborhoods of a for some
appropriate value of ε.

45
 2.3. THE ABSOLUTE VALUE

Exercises

Exercise 2.3.1. Prove that if a < x < b and a < y < b then |x − y| <
b−a. Draw a number line with points a, b, x, y satisfying the inequalities
and graphically interpret the inequality |x − y| < b − a.

Exercise 2.3.2. Let a0 , a1, a2 , . . . , an be positive real numbers and

consider the polynomial

f (x) = a0 + a1 x + a2 x2 + · · · + an xn .

Prove that
|f (x)| ≤ f (|x|)
for all x ∈ R. Hint: For example, if say f (x) = 2 + 3x2 + 7x3 then you are asked
to prove that
|2 + 3x2 + 7x3 | ≤ 2 + 3|x|2 + 7|x|3 .
| {z } | {z }
|f (x)| f (|x|)

However, prove the claim for a general polynomial f (x) = a0 + a1 x + a2 x2 + · · · + an xn

with ai > 0.

Exercise 2.3.3. Let f (x) = 3x2 − 7x + 11 for x ∈ [−4, 2]. Find

analytically a number M > 0 such that |f (x)| ≤ M for all x ∈ [−4, 2].
Do not use calculus to find M.
x−1
Exercise 2.3.4. Let f (x) = 2 for x ∈ [0, 10]. Find analytically
x +7
a number M > 0 such that |f (x)| ≤ M for all x ∈ [0, 10]. Do not use
calculus to find M.
3 cos(πx)
Exercise 2.3.5. Let f (x) = 2 for x ∈ [0, 2]. Find analyti-
x − 2x + 3
cally a number M > 0 such that |f (x)| ≤ M for all x ∈ [0, 2]. Do not
use calculus to find M. (Hint: Complete the square.)

Exercise 2.3.6. Let a, b ∈ R be distinct points. Show that there exists

neighborhoods Bε (a) and Bδ (b) such that Bǫ(a) ∩ Bδ (b) 6= ∅.

46
 2.4. THE COMPLETENESS AXIOM

2.4 The Completeness Axiom

In this section, we introduce the Completeness Axiom of R. Recall
that an axiom is a statement or proposition that is accepted as true
without justification. In mathematics, axioms are the first principles
that are accepted as truths and are used to build mathematical theories;
in this case real analysis. Roughly speaking, the Completeness Axiom
is a way to say that the real numbers have “no gaps” or “no holes”,
contrary to the case of the rational numbers. As you will see below,
the Completeness Axiom is centered around the notions of bounded
sets and least upper bounds; let us begin then with some definitions.

Definition 2.4.1: Boundedness

Let S ⊂ R be a non-empty set.

(i) We say that S is bounded above if there exists u ∈ R such

that x ≤ u for all x ∈ S. We then say that u is an upper
bound of S.

(ii) We say that S is bounded below if there exists w ∈ R such

that w ≤ x for all x ∈ S. We then say that w is a lower
bound of S.

(iii) We say that S is bounded if it is both bounded above and

bounded below.

(iv) We say that S is unbounded if it is not bounded.

Example 2.4.2. For each case, determine if S is bounded above, bounded

below, bounded, or unbounded. If the set is bounded below, determine
the set of lower bounds, and similarly if it is bounded above.

47
 2.4. THE COMPLETENESS AXIOM

(i) S = [0, 1]

(ii) S = (−∞, 3)

(iii) N = {1, 2, 3, . . . , }

(iv) S = { x21+1 | − ∞ < x < ∞}

(v) S = {x ∈ R | x2 < 2}

Example 2.4.3. Let A and B be sets and suppose that A ⊂ B.

(a) Prove that if B is bounded above (below) then A is bounded

above (below).

(b) Give an example of sets A and B such that A is bounded below

but B is not bounded below.

We now come to an important notion that will be at the root of

what we do from now.

Definition 2.4.4: Supremum and Infimum

Let S ⊂ R be non-empty.

(i) Let S be bounded above. An upper bound u of S is said to

be a least upper bound of S if u ≤ u′ for any upper bound
u′ of S. In this case we also say that u is a supremum of S
and write u = sup(S).

(ii) Let S be bounded below. A lower bound w of S is said to be

a greatest lower bound of S if w′ ≤ w for any lower bound
w′ of S. In this case we also say that w is an infimum of S
and write w = inf(S).

48
 2.4. THE COMPLETENESS AXIOM

It is straightforward to show that a set that is bounded above (bounded

below) can have at most one supremum (infimum). At the risk of being
repetitive, when it exists, sup(S) is a number that is an upper bound of
S and is the smallest possible upper bound of S. Therefore, x ≤ sup(S)
for all x ∈ S, and any number less than sup(S) is not an upper bound
of S (because sup(S) is the least upper bound!). Similarly, when it
exists, inf(S) is a number that is a lower bound of S and is the largest
possible lower bound of S. Therefore, inf(S) ≤ x for all x ∈ S and any
number greater than inf(S) is not a lower bound of S (because inf(S)
is the greatest lower bound of S!).

Remark 2.4.5. In some analysis texts, sup(S) is written as lub(S)

and inf(S) is written as glb(S). In other words, sup(S) = lub(S) and
inf(S) = glb(S).

Does every non-empty bounded set S have a supremum/infimum?

You might say “Yes, of course!!” and add that “It is a self-evident
principle and needs no justification!”. Is not that what an axiom is?

Axiom 2.4.6: Completeness Axiom

Every non-empty subset of R that is bounded above has a least
upper bound (a supremum) in R. Similarly, every non-empty subset
of R that is bounded below has a greatest lower bound (an infimum)
in R.

As you will see in the pages that follow, The Completeness Axiom is
the key notion upon which the theory of real analysis depends on.

Example 2.4.7. Determine sup(S) and inf(S), if they exist.

(a) S = {−5, −9, 2, −1, 11, 0, 4}

49
 2.4. THE COMPLETENESS AXIOM

(b) S = [0, ∞)

(c) S = (−∞, 3)
The Completeness Axiom is sometimes called the supremum prop-
erty of R or the least upper bound property of R. The Complete-
ness property makes R into a complete ordered field. The following
example shows that Q does not have the completeness property.
Example 2.4.8. The set of rational numbers is an ordered field but it
is not complete. Consider the set S = {x ∈ Q | x2 < 2}. By definition,
S ⊂ Q. Clearly, S is bounded above, for example u = 10 is an upper
√
bound of S, but the least upper bound of S is u = 2 which is not
a rational number. Therefore S ⊂ Q does not have a supremum in Q
and therefore Q does not have the Completeness property. From the
point of view of analysis, this is the main distinction between Q and R
(note that both are ordered fields).
In some cases, it is obvious what sup and inf are, however, to do
analysis rigorously, we need systematic ways to determine sup(S) and
inf(S). To start with, we first need to be a bit more rigorous about
what it means to be the least upper bound, or at least have a more
concrete description that we can work with, i.e., using inequalities.
The following lemma does that and, as you will observe, it is simply a
direct consequence of the definition of the supremum.

Lemma 2.4.9
Let S ⊂ R be non-empty and suppose that u ∈ R is an upper bound
of S. Then u is the least upper bound of S if and only if for any
ε > 0 there exists x ∈ S such that

u − ε < x ≤ u.

50
 2.4. THE COMPLETENESS AXIOM

Proof. Suppose that u is the supremum of S, that is, u is the least

upper bound of S. Since u − ε < u then u − ε is not an upper bound
of S. Thus, there exists x ∈ S such that u − ε < x.
Now suppose that for any ε > 0 there exists x ∈ S such that
u − ε < x ≤ u. Now let v ∈ R be such that v < u. Then there exists
ε > 0 such that v = u − ε, and thus by assumption, there exists x ∈ S
such that v < x. Hence, v is not an upper bound of S and this shows
that u is the least upper bound of S, that is, u = sup(S).

Example 2.4.10. If A ⊂ B ⊂ R, and B is bounded above, prove that

A is bounded above and that sup(A) ≤ sup(B).

Solution. Since B is bounded above, sup(B) exists by the Completeness

property of R. Let x ∈ A. Then x ∈ B and therefore x ≤ sup(B). This
proves that A is bounded above by sup(B) and therefore sup(A) exists.
Since sup(A) is the least upper bound of A we must have sup(A) ≤
sup(B). For example, if say B = [1, 3] and A = [1, 2] then sup(A) <
sup(B), while if A = [2, 3] then sup(A) = sup(B).

Example 2.4.11. Let A ⊂ R be non-empty and bounded above. Let

c ∈ R and define the set

cA = {y ∈ R | ∃ x ∈ A s.t. y = cx}.

Prove that if c > 0 then cA is bounded above and sup(cA) = c sup(A).

Show by example that if c < 0 then cA need not be bounded above
even if A is bounded above.

Proof. Let y ∈ cA be arbitrary. Then there exists x ∈ A such that

y = cx. By the Completeness property, sup(A) exists and x ≤ sup(A).
If c > 0 then cx ≤ c sup(A) which is equivalent to y ≤ c sup(A). Since
y is arbitrary, this shows that c sup(A) is an upper bound of the set

51
 2.4. THE COMPLETENESS AXIOM

cA and thus cA is bounded above and thus sup(cA) exists. Because

sup(cA) is the least upper bound of cA then

sup(cA) ≤ c sup(A). (2.1)

Now, by definition, y ≤ sup(cA) for all y ∈ cA. Thus, cx ≤ sup(cA)

for all x ∈ A and therefore x ≤ 1c sup(cA) for all x ∈ A. Therefore,
1
c
sup(cA) is an upper bound of A and consequently sup(A) ≤ 1c sup(cA)
because sup(A) is the least upper bound of A. We have therefore proved
that
c sup(A) ≤ sup(cA). (2.2)

Combining (2.1) and (2.2) we conclude that

sup(cA) = c sup(A).

Example 2.4.12. Suppose that A and B are non-empty and bounded

above. Prove that A ∪ B is bounded above and that

sup(A ∪ B) = sup{sup(A), sup(B)}.

Proof. Let u = sup{sup(A), sup(B)}. Then clearly sup(A) ≤ u and

sup(B) ≤ u. We first show that A ∪ B is bounded above by showing
that u is an upper bound of A ∪ B. Let x ∈ A ∪ B. Then either z ∈ A
or x ∈ B (or both). If z ∈ A then z ≤ sup(A) ≤ u and if z ∈ B then
z ≤ sup(B) ≤ u. Hence, A ∪ B is bounded above and u is an upper
bound of A ∪ B. Consequently, sup(A ∪ B) exists by the Completeness
axiom and moreover sup(A ∪ B) ≤ u, that is,

sup(A ∪ B) ≤ sup{sup(A), sup(B)}. (2.3)

52
 2.4. THE COMPLETENESS AXIOM

Now, by definition of the supremum, z ≤ sup(A ∪ B) for all z ∈ A ∪ B.

Now since A ⊂ A ∪ B this implies that x ≤ sup(A ∪ B) for all x ∈ A
and y ≤ sup(A ∪ B) for all y ∈ B. In other words, sup(A ∪ B) is an
upper bound of both A and B and thus sup(A) ≤ sup(A ∪ B) and
sup(B) ≤ sup(A ∪ B). Then clearly

sup{sup(A), sup(B)}} ≤ sup(A ∪ B) (2.4)

and combining (2.3) and (2.4) we have proved that sup(A ∪ B) =

sup{sup(A), sup(B)}.

Example 2.4.13. Suppose that A and B are non-empty and bounded

below, and suppose that A ∩ B is non-empty. Prove that A ∩ B is
bounded below and that

sup{inf(A), inf(B)} ≤ inf(A ∩ B).

Proof. If x ∈ A ∩ B then x ∈ A and therefore inf(A) ≤ x, and also

x ∈ B and thus inf(B) ≤ x. Therefore, both inf(A) and inf(B)
are lower bounds of A ∩ B, and by definition of inf(A ∩ B) we have
that inf(A) ≤ inf(A ∩ B) and inf(B) ≤ inf(A ∩ B), and consequently
sup{inf(A), inf(B)} ≤ inf(A ∩ B).

Example 2.4.14. For any set A define the set

−A = {y ∈ R | ∃ x ∈ A s.t. y = −x}.

Prove that if A ⊂ R is non-empty and bounded then sup(−A) =

− inf(A).

Proof. It holds that inf(A) ≤ x for all x ∈ A and therefore, − inf(A) ≥

−x for all x ∈ A, which is equivalent to − inf(A) ≥ y for all y ∈ −A.
Therefore, − inf(A) is an upper bound of the set −A and therefore

53
 2.4. THE COMPLETENESS AXIOM

sup(−A) ≤ − inf(A). Now, y ≤ sup(−A) for all y ∈ −A, or equiv-

alently −x ≤ sup(−A) for all x ∈ A. Hence, x ≥ − sup(−A) for all
x ∈ A. This proves that − sup(−A) is a lower bound of A and there-
fore − sup(−A) ≤ inf(A), or sup(−A) ≥ − inf(A). This proves that
sup(−A) = − inf(A).

Example 2.4.15. Let A ⊂ R be non-empty and bounded below. Let

c ∈ R and define the set

c + A = {y ∈ R | ∃ x ∈ A s.t. y = c + x}.

Prove that c + A is bounded below and that inf(c + A) = c + inf(A).

Proof. For all x ∈ A it holds that inf(A) ≤ x and therefore c+inf(A) ≤

c + x. This proves that c + inf(A) is a lower bound of the set c + A,
and therefore c + inf(A) ≤ inf(c + A). Now, inf(c + A) ≤ y for all
y ∈ c + A and thus inf(c + A) ≤ c + x for all x ∈ A, which is the same
as inf(c + A) − c ≤ x for all x ∈ A. Hence, inf(c + A) − c ≤ inf(A) or
equivalently inf(c + A) ≤ c + inf(A). This proves the claim.

Example 2.4.16. Let A and B be non-empty subsets of R>0 = {x ∈

R : x > 0}, and suppose that A and B are bounded below. Define the
set AB = {xy : x ∈ A, y ∈ B}.

(a) Prove that AB is bounded below.

(b) Prove that inf(AB) = inf(A) · inf(B). Hint: Consider two cases,
when say inf(A) inf(B) = 0 and when inf(A) inf(B) 6= 0.

(c) How do things change if we do not assume A and B are subsets

of R>0 .

Example 2.4.17. Give an example of a non-empty set A ⊂ R>0 such

that inf(A) = 0.

54
 2.4. THE COMPLETENESS AXIOM

Example 2.4.18. For any two non-empty sets P and Q of R let us

write that P ≤ Q if, for each x ∈ P , there exists y ∈ Q such that
x ≤ y.

(a) Prove that if P ≤ Q then sup(P ) ≤ sup(Q).

(b) Show via an example that if P ≤ Q then it is not necessarily true

that inf(P ) ≤ inf(Q).

Example 2.4.19. Let A and B be non-empty bounded sets of positive

real numbers. Define the set
A n x
o
= z ∈ R | ∃ x ∈ A, ∃ y ∈ B s.t. z = y .
B
Assume that inf(B) > 0. Prove that

A

sup(A)
sup B
= .
inf(B)

Proof. Since A is bounded above, sup(A) exists and x ≤ sup(A) for all
x ∈ A. Since B is bounded below, inf(B) exists and inf(B) ≤ y for all
A
y ∈ B. Let z = x/y be an arbitrary point in B . Then since inf(B) ≤ y
and y > 0 and inf(B) > 0 we obtain that

1 1
≤ .
y inf(B)

Since x ≤ sup(A) and x > 0 (and then clearly sup(A) > 0) we obtain

x x sup(A)
≤ ≤ .
y inf(B) inf(B)
sup(A) A
This proves that z ≤ inf(B) and since z ∈ B was arbitrary we have
sup(A) A
proved that inf(B) is an upper bound of the set B. This proves that

55
 2.4. THE COMPLETENESS AXIOM

A
sup( B ) exists. Moreover, by the definition of the supremum, we also
have that
A

sup(A)
sup B ≤ . (2.5)
inf(B)
A A
Now, z ≤ sup( B ) for all z ∈ B and thus xy ≤ sup( BA
) for all x ∈ A and
A
all y ∈ B. If y is held fixed then x ≤ y · sup( B ) for all x ∈ A and thus
A A
sup(A) ≤ y · sup( B ). Therefore, sup(A)/ sup( B ) ≤ y, which holds for
A
all y ∈ B. Therefore, sup(A)/ sup( B ) ≤ inf(B) and consequently

sup(A) A


≤ sup B
. (2.6)
inf(B)

Combining (2.5) and (2.6) completes the proof.

56
 2.4. THE COMPLETENESS AXIOM

Exercises

Exercise 2.4.1. Let S ⊂ R be a non-empty set and suppose that u is

an upper bound of S. Prove that if u ∈ S then necessarily u = sup S.

Exercise 2.4.2. Let A and B be non-empty subsets of R, and suppose

that A ⊂ B. Prove that if B is bounded below then inf B ≤ inf A.

Exercise 2.4.3. If P and Q are non-empty subsets of R such that

sup P = sup Q and inf P = inf Q does it follow that P = Q? Support
your answer with either a proof or a counterexample.

Exercise 2.4.4. Let A ⊂ R be a bounded set, let x ∈ R be fixed, and

define the set

x + A = {y ∈ R | ∃ a ∈ A s.t. y = x + a}.

Prove that sup(x + A) = x + sup A.

Exercise 2.4.5. Let A, B ⊂ R be bounded above. Let

A + B = {z ∈ R | ∃ x ∈ A, ∃ y ∈ B s.t. z = x + y}.

Prove that A+B is bounded above and that sup(A+B) = sup A+sup B.

Exercise 2.4.6. Let R>0 denote the set of all positive real numbers
and let A, B ⊂ R>0 be bounded. Assume that inf(B) > 0. Define the
set
A n x
o
= z ∈ R | ∃ x ∈ A, ∃ y ∈ B s.t. z = y .
B
Prove that
A

sup(A)
sup B = .
inf(B)

57
 2.5. APPLICATIONS OF THE SUPREMUM

2.5 Applications of the Supremum

Having defined the notion of boundedness for a set, we can define a
notion of boundedness for a function.

Definition 2.5.1: Bounded Functions

Let D ⊂ R be a non-empty set and let f : D → R be a function.

(i) f is bounded below if the range f (D) = {f (x) | x ∈ D} is

bounded below.

(ii) f is bounded above if the range f (D) = {f (x) | x ∈ D} is

bounded above.

(iii) f is bounded if the range f (D) = {f (x) | x ∈ D} is bounded.

Hence, boundedness of a function f is boundedness of its range.

Example 2.5.2. Suppose f, g : D → R are bounded functions and

f (x) ≤ g(x) for all x ∈ D. Show that sup(f (D)) ≤ sup(g(D)).

Solution. Since g is bounded, sup(g(D)) exists and is by definition an

upper bound for the set g(D), that is, g(x) ≤ sup(g(D)) for all x ∈ D.
Now, by assumption, for all x ∈ D it holds that f (x) ≤ g(x) and
therefore f (x) ≤ sup(g(D)). This shows that sup(g(D)) is an upper
bound of the set f (D), and therefore by definition of the supremum,
sup(f (D)) ≤ sup(g(D)).

Example 2.5.3. Let f, g : D → R be bounded functions and suppose

that f (x) ≤ g(y) for all x, y ∈ D. Show that sup(f (D)) ≤ inf(g(D)).

58
 2.5. APPLICATIONS OF THE SUPREMUM

Solution. Fix x∗ ∈ D. Then, f (x∗) ≤ g(y) for all y ∈ D. Therefore,

f (x∗) is a lower bound of g(D), and thus f (x∗) ≤ inf(g(D)) by definition
of the infimum. Since x∗ ∈ D was arbitrary, we have proved that
f (x) ≤ inf(g(D)) for all x ∈ D. Hence, inf(g(D)) is an upper bound of
f (D) and thus by definition of the supremum we have that sup(f (D)) ≤
inf(g(D)).

The following sequence of results will be used to prove an im-

portant property of the rational numbers Q as seen from within R.

Theorem 2.5.4: Archimedean Property

If x ∈ R then there exists n ∈ N such that x ≤ n.

Proof. Suppose not. Hence, n ≤ x for all n ∈ N, and thus x is an

upper bound for N, and therefore N is bounded. Let u = sup(N). By
definition of u, u − 1 is not an upper bound of N and therefore there
exists m ∈ N such that u − 1 < m. But then u < m + 1 and this
contradicts the definition of u.

Corollary 2.5.5

If S = { n1 | n ∈ N} then inf(S) = 0.

Proof. Since 0 < n1 for all n then 0 is a lower bound of S. Now suppose
that 0 < y. By the Archimedean Property, there exists n ∈ N such that
1
y
< n and thus n1 < y. Hence, y is not a lower bound of S. Therefore
0 is the greatest lower bound of S, that is, inf(S) = 0.

59
 2.5. APPLICATIONS OF THE SUPREMUM

Corollary 2.5.6
1
For any y > 0 there exists n ∈ N such that n < y.

Corollary 2.5.7
Given y > 0 there exists n ∈ N such that n − 1 ≤ y < n.

Proof. Let E = {k ∈ R | y < k}. By the Archimedean property, E is

non-empty. By the Well-Ordering Principle of N, E has a least element,
say that it is m. Hence, m − 1 ∈
/ E and thus m − 1 ≤ y < m.

We now come to an important result that we will use frequently.

Theorem 2.5.8: Density of the Rationals

If x, y ∈ R and x < y then there exists r ∈ Q such that x < r < y.

Proof. We first prove the claim for the case that 0 < x < y. Suppose
that y − x > 1 and thus x + 1 < y. There exists m ∈ N such that
m − 1 ≤ x < m and thus m ≤ x + 1. Therefore,

x<m≤x+1<y

and thus x < m < y and we may take r = m. In general, if y − x > 0

then there exists n ∈ N such that n1 < y − x and thus 1 + nx < ny.
Since ny − nx > 1 there exists m ∈ N such that 1 + nx < m < ny and
thus
nx < 1 + nx < m < ny
and dividing by n yields x < m
n
< y and thus we may take r = m n
.
This proves the claim when both x and y are positive. If x < 0 < y

60
 2.5. APPLICATIONS OF THE SUPREMUM

then take r = 0 and if x < y < 0 then apply the previous arguments to
0 < −y < −x.

Hence, between any two distinct real numbers there is a rational num-
ber. This implies that any irrational number can be approximated by
a rational number to within any degree of accuracy.

Example 2.5.9. Let ζ ∈ R\Q be an irrational number and let ε > 0

be arbitrary. Prove that there exists a rational number x ∈ Q such
that ζ − ε < x < ζ + ε, that is, x is in the ε-neighborhood of ζ.

Corollary 2.5.10: Density of the Irrationals

If x, y ∈ R and x < y then there exists ξ ∈ R\Q such that x < ξ < y.

√ √ √
Proof. We have that 2x < 2y. By the Density Theorem, 2x <
√ √
r < 2y for some r ∈ Q. Then ξ = r/ 2.

61
 2.5. APPLICATIONS OF THE SUPREMUM

Exercises

Exercise 2.5.1. Let D ⊂ R be non-empty and let f, g : D → R be

functions. Let f + g denote the function defined by

(f + g)(x) = f (x) + g(x)

for any x ∈ D. If f (D) and g(D) are bounded above prove that (f +
g)(D) is also bounded above and that

sup(f + g)(D) ≤ sup f (D) + sup g(D)

Exercise 2.5.2. If y > 0 prove that there exists n ∈ N such that

1 1
2n < y. (Note: If you begin with 2n < y and solve for n then you are
assuming that such an n exists. You are not asked to find an n, you
are asked to prove that such an n exists.)

62
 2.6. NESTED INTERVAL THEOREM

2.6 Nested Interval Theorem

If a, b ∈ R and a < b define

(a, b) := {x ∈ R | a < x < b}.

The set (a, b) is called an open interval from a to b. Define also

[a, b] := {x ∈ R | a ≤ x ≤ b}

which we call the closed interval from a to b. The following are called
half-open (or half-closed) intervals:

[a, b) := {x ∈ R | a ≤ x < b}

(a, b] := {x ∈ R | a < x ≤ b}.

If a = b then (a, a) = ∅ and [a, a] = {a}. Infinite intervals are

(a, ∞) = {x ∈ R | x > a}
(−∞, b) = {x ∈ R | x < b}
[a, ∞) = {x ∈ R | x ≥ a}
(−∞, b] = {x ∈ R | x ≤ b}.

Below is a characterization of intervals, we will omit the proof.

Theorem 2.6.1
Let S ⊂ R contain at least two points. Suppose that if x, y ∈ S and
x < y then [x, y] ⊂ S. Then S is an interval.

A sequence I1, I2, I3, I4, . . . of intervals is nested if

I1 ⊇ I2 ⊇ I3 ⊇ I4 ⊇ · · ·

63
 2.6. NESTED INTERVAL THEOREM

As an example, consider In = [0, n1 ] where n ∈ N. Then I1, I2, I3, . . . is

nested:
[0, 1] ⊇ [0, 21 ] ⊇ [0, 13 ] ⊇ [0, 41 ] · · ·
Notice that since 0 ∈ In for each n ∈ N then
∞
\
0∈ In .
n=1
T∞ T∞
Is there another point in n=1 In ? Suppose that x 6= 0 and x ∈ n=1 In .
1
Then necessarily x > 0. Then there exists m ∈ N such that m
< x.
T
Thus, x ∈ / ∞
/ [0, m1 ] and therefore x ∈ n=1 In . Therefore,
∞
\
In = {0}.
n=1

In general, we have the following.

Theorem 2.6.2: Nested Interval Property

Let I1 , I2, I3, . . . be a sequence of nested closed bounded intervals.
Then there exists ξ ∈ R such that ξ ∈ In for all n ∈ N, that
T
is, ξ ∈ ∞ n=1 In . In particular, if In = [an , bn ] for n ∈ N, and
a = sup{a1 , a2 , a3, . . .} and b = inf{b1, b2, b3, . . .} then
∞
\
{a, b} ⊂ In .
n=1

Proof. Since In is a closed interval, we can write In = [an , bn] for some
an , bn ∈ R and an ≤ bn for all n ∈ N. The nested property can be
written as
[a1 , b1] ⊇ [a2, b2] ⊇ [a3 , b3] ⊇ [a4, b4] ⊇ · · ·
Since [an, bn ] ⊆ [a1 , b1] for all n ∈ N then an ≤ b1 for all n ∈ N.
Therefore, the set S = {an | n ∈ N} is bounded above. Let ξ = sup(S)

64
 2.6. NESTED INTERVAL THEOREM

and thus an ≤ ξ for all n ∈ N. We will show that ξ ≤ bn for all n ∈ N

also. Let n ∈ N be arbitrary. If k ≤ n then [an , bn ] ⊆ [ak , bk ] and
therefore ak ≤ an ≤ bn . On the other hand, if n < k then [ak , bk ] ⊂
[an , bn] and therefore an ≤ ak ≤ bn . In any case, ak ≤ bn for all
k ∈ N. Hence, bn is an upper bound of S, and thus ξ ≤ bn . Since
n ∈ N was arbitrary, we have that ξ ≤ bn for all n ∈ N. Therefore,
T
an ≤ ξ ≤ bn for all n ∈ N, that is ξ ∈ ∞ n=1 [an , bn ]. The proof that
T∞
inf{b1, b2, b3, . . .} ∈ n=1 In is similar.
T
The following theorem gives a condition when ∞ n=1 In contains a single
point.

Theorem 2.6.3
Let In = [an , bn] be a sequence of nested closed bounded intervals.
If
inf{bn − an | n ∈ N} = 0
T
then nn=1 In is a singleton set.

 
Example 2.6.4. Let In = 1 − n1 , 1 + n1 for n ∈ N.
(a) Prove that I1 , I2, I3, . . . is a sequence of nested intervals.
T
(b) Find ∞ n=1 In .

Using the Nested Interval property of R, we give a proof that R is

uncountable

Theorem 2.6.5: Reals Uncountable

The real numbers R are uncountable.

Proof. We will prove that the interval [0, 1] is uncountable. Suppose by

contradiction that I = [0, 1] is countable, and let I = {x1, x2, x3, . . . , }

65
 2.6. NESTED INTERVAL THEOREM

be an enumeration of I (formally this means we have a bijection f :

N → [0, 1] and f (n) = xn ). Since x1 ∈ I = [0, 1], there exists a closed
and bounded interval I1 ⊂ [0, 1] such that x1 ∈ / I1 . Next, consider x2.
There exists a closed and bounded interval I2 ⊂ I1 such that x2 ∈ / I2 .
Next, consider x3. There exists a closed and bounded interval I3 ⊂ I2
such that x3 ∈/ I3. By induction, there exists a sequence I1, I2, I3, . . . of
closed and bounded intervals such that xn ∈ / In for all n ∈ N. Moreover,
T
by construction the sequence In is nested and therefore ∞ n=1 In is non-
empty, say it contains ξ. Clearly, since ξ ∈ In ⊂ [0, 1] for all n ∈ N
T
then ξ ∈ [0, 1]. Now, since xn ∈/ In for each n ∈ N then xn ∈ / ∞ n=1 In .
Therefore, ξ 6= xn for all n ∈ N and thus ξ ∈ / I = {x1, x2, . . .} = [0, 1],
which is a contradiction since ξ ∈ [0, 1]. Therefore, [0, 1] is uncountable
and this implies that R is also uncountable.

We now give an alternative proof that R is uncountable. To that

end, consider the following subset S ⊂ R:

S = {0.a1a2 a3 a4 · · · ∈ R | ak = 0 or ak = 1, k ∈ N}.

In other words, S consists of numbers x ∈ [0, 1) whose decimal ex-

pansion consists of only 0’s and 1’s. For example, some elements of S
are

x = 0.000000 . . .
x = 0.101010 . . .
x = 0.100000 . . .
x = 0.010100 . . .

If we can construct a bijection f : S → P(N), then since P(N) is

uncountable then by Example 1.4.7 this would show that S is uncount-
able. Since S ⊂ R then this would show that R is uncountable (by

66
 2.6. NESTED INTERVAL THEOREM

Theorem 1.4.8). To construct f , given x = 0.a1a2 a3 . . . in S define

f (x) ∈ P(N) as

f (0.a1a2 a3 · · · ) = {k ∈ N | ak = 1}.

In other words, f (x) consists of the decimal places in the decimal ex-
pansion of x that have a value of 1. For example,

f (0.000000 . . .) = ∅
f (0.101010 . . .) = {1, 3, 5, 7, . . .}
f (0.100100 . . .) = {1, 4}
f (0.011000 . . .) = {2, 3}

It is left as an exercise to show that f is a bijection (see Exercise 1.4.4).

67
 2.6. NESTED INTERVAL THEOREM

Exercises
  T
Exercise 2.6.1. Let In = 0, n1 for n ∈ N. Prove that ∞
n=1 In = {0}.

68
 3

Sequences

In the tool box used to build analysis, if the Completeness property of

the real numbers is the hammer then sequences are the nails. Almost
everything that can be said in analysis can be, and is, done using se-
quences. For this reason, the study of sequences will occupy us for the
next foreseeable future.

3.1 Limits of Sequences

A sequence of real numbers is a function X : N → R. Informally,
the sequence X can be written as an infinite list of real numbers as
X = (x1, x2, x3, . . .), where xn = X(n). Other notations for sequences
are (xn) or {xn}∞n=1 ; we will use (xn ).
Some sequences can be written explicitly with a formula such as
xn = n1 , xn = 21n , or

xn = (−1)n cos(n2 + 1),

or we could be given the first few terms of the sequence, such as

X = (3, 3.1, 3.14, 3.141, 3.1415, . . .).

69
 3.1. LIMITS OF SEQUENCES

Some sequences may be given recursively. For example,

xn
x1 = 1, xn+1 = , n ≥ 1.
n+1
Using the definition of xn+1 and the initial value x1 we can in principle
find all the terms:
1 1/2 1/6
x2 = , x3 = , x4 = , ...
2 3 4
A famous sequence given recursively is the Fibonacci sequence which
is defined as x1 = 1, x2 = 1, and

xn+1 = xn−1 + xn , n ≥ 2.

Then
x3 = 2, x4 = 3, x5 = 5, . . .
The range of a sequence (xn) is the set

{xn | n ∈ N},

that is, the usual range of a function. However, the range of a sequence
is not the actual sequence (the range is a set and a sequence is a func-
tion). For example, if X = (1, 2, 3, 1, 2, 3, . . .) then the range of X is
{1, 2, 3}. If xn = sin( nπ
2 ) then the range of (xn ) is {1, 0, −1}.
Many concepts in analysis can be described using the long-term or
limiting behavior of sequences. In calculus, you undoubtedly developed
techniques to compute the limit of basic sequences (and hence show
convergence) but you might have omitted the rigorous definition of the
convergence of a sequence. Perhaps you were told that a given sequence
(xn) converges to L if as n → ∞ the values xn get closer to L. Although
this is intuitively sound, we need a more precise way to describe the
meaning of the convergence of a sequence. Before we give the precise
definition, we will consider an example.

70
 3.1. LIMITS OF SEQUENCES

Example 3.1.1. Consider the sequence (xn) whose nth term is given
by xn = 3n+2
n+1 . The values of (xn ) for several values of n are displayed
in Table 3.1.
n xn
1 2.50000000
2 2.66666667
3 2.75000000
4 2.80000000
5 2.83333333
50 2.98039216
101 2.99019608
10,000 2.99990001
1,000,000 2.99999900
2,000,000 2.99999950
3n+2
Table 3.1: Values of the sequence xn = n+1

The above data suggests that the values of the sequence (xn) become
closer and closer to the number L = 3. For example, suppose that
ε = 0.005 and consider the ε-neighborhood of L = 3, that is, the
interval (3 − ε, 3 + ε) = (2.995, 3.005). Not all the terms of the sequence
(xn) are in the ε-neighborhood, however, it seems that all the terms of
the sequence from x101 and onward are inside the ε-neighborhood. In
other words, IF n ≥ 101 then 3 − ε < xn < 3 + ε, or equivalently
|xn − 3| < ε. Suppose now that ε = 0.00001 and thus the new ε-
neighborhood is (3−ε, 3+ε) = (2.99999, 3.000001). Then it is no longer
true that |xn − 3| < ε for all n ≥ 101. However, it seems that all the
terms of the sequence from x1000000 and onward are inside the smaller
ε-neighborhood, in other words, |xn − 3| < ε for all n ≥ 1, 000, 000.
We can extrapolate these findings and make the following hypothesis:
For any given ε > 0 there exists a natural number K ∈ N such that if
n ≥ K then |xn − L| < ε.

71
 3.1. LIMITS OF SEQUENCES

The above example and our analysis motivates the following defini-
tion.

Definition 3.1.2: Convergence of Sequences

The sequence (xn) is said to converge if there exists a number
L ∈ R such that for any given ε > 0 there exists K ∈ N such that
|xn − L| < ε for all n ≥ K. In this case, we say that (xn) has limit
L and we write
lim xn = L.
n→∞
If (xn) is not convergent then we say that it is divergent.

Hence, xn converges to L if for any given ε > 0 (no matter how small),
there exists a point in the sequence xK such that |xK − L| < ε, |xK+1 −
L| < ε, |xK+2 − L| < ε, . . ., that is, |xn − L| < ε for all n ≥ K. We will
sometimes write lim xn = L simply as lim xn = L or (xn) → L.
n→∞

Example 3.1.3. Using the definition of the limit of a sequence, prove

that limn→∞ n1 = 0.

Proof. Let ε > 0 be arbitrary but fixed. By the Archimedean property

of R, there exists K ∈ N such that K1 < ε. Then, if n ≥ K then
1
n
≤ K1 < ε. Therefore, if n ≥ K then

1
|xn − 0| = n
−0
1
=
n
1
≤
K
< ε.

This proves, by definition, that limn→∞ n1 = 0.

72
 3.1. LIMITS OF SEQUENCES

Example 3.1.4. Using the definition of the limit of a sequence, prove

that limn→∞ 3n+2
n+1 = 3.

Proof. Given an arbitrary ε > 0, we want to prove that there exists

K ∈ N such that
3n + 2
− 3 < ε, ∀ n ≥ K.
n+1
Start by analyzing |xn − L|:
3n + 2
|xn − L| = −3
n+1
−1
=
n+1
1
= .
n+1
Now, the condition that
3n + 2 1
−3 = <ε
n+1 n+1
it is equivalent to
1
− 1 < n.
ε
Now let’s write the formal proof.
Let ε > 0 be arbitrary and let K ∈ N be such that 1ε − 1 < K.
1 1 1
Then, K+1 < ε. Now if n ≥ K then n+1 ≤ K+1 and thus if n ≥ K then
3n + 2 −1
−3 =
n+1 n+1
1
=
n+1
1
≤
K +1
< ε.

By definition, this proves that (xn) → 3.

73
 3.1. LIMITS OF SEQUENCES

Example 3.1.5. Using the definition of the limit of a sequence, prove

3
that limn→∞ 4nn3+3n
+6 = 4.

3
Proof. Let xn = 4nn3+3n
+6 . We want to show that for any given ε > 0,
there exists K ∈ N such that if n ≥ K then

4n3 + 3n
|xn − 4| = − 4 < ε.
n3 + 6

Start by analyzing |xn − 4|:

4n3 + 3n
|xn − 4| = −4
n3 + 6
3n − 24
=
n3 + 6

In this case, it is difficult to explicitly isolate for n in terms of ε. Instead

we take a different approach; we find an upper bound for |xn − 4| =
3n−24
n3 +6 :

3n − 24 3n + 24
≤
n3 + 6 n3 + 6

27n
≤
n3 + 6
27n
<
n3
27
= 2.
n
27
Hence, if n2< ε then also |xn − 4| < ε by the transitivity property
q of
inequalities. The inequality n272 < ε holds true if and only if 27
ε < n.
Now that we have done a detailed preliminary analysis, we can proceed
with the proof.

74
 3.1. LIMITS OF SEQUENCES

q
Suppose that ε > 0 is given and let K ∈ N be such that 27
ε
< K.
Then 27 2 27 27 27
ε < K , and thus K 2 < ε. Then, if n ≥ K then n2 ≤ K 2 and
therefore
3n − 24
|xn − 4| =
n3 + 6

3n + 24
≤
n3 + 6
27n
≤ 3
n +6
27n
< 3
n
27
= 2
n
27
≤ 2
K
< ε.

This proves that limn→∞ xn = 4.

Example 3.1.6 (Important). Prove that for any irrational number ζ

there exists a sequence of rational numbers (xn) converging to ζ.

Proof. Let (δn) be any sequence of positive numbers converging to zero,

for example, δn = n1 . Now since ζ − δn < ζ + δn for each n ∈ N, then
by the Density theorem there exists a rational number xn such that
ζ − δn < xn < ζ + δn . In other words, |xn − ζ| < δn . Now let ε > 0 be
arbitrary. Since (δn) converges to zero, there exists K ∈ N such that
|δn − 0| < ε for all n ≥ K, or since δn > 0, then δn < ε for all n ≥ K.
Therefore, if n ≥ K then |xn − ζ| < δn < ε. Thus, for arbitrary ε > 0
there exists K ∈ N such that if n ≥ K then |xn − ζ| < ε. This proves
that (xn) converges to ζ.

75
 3.1. LIMITS OF SEQUENCES

cos(n)
Example 3.1.7. Let xn = n2 −1 where n ≥ 2. Prove that limn→∞ xn =
0.
Proof. We want to prove given any ε > 0 there exists K ∈ N such that
cos(n)
< ε, n ≥ K.
n2 − 1
Now, since | cos(x)| ≤ 1 for all x ∈ R we have that
cos(n) | cos(n)|
=
n2 − 1 n2 − 1
1
≤ 2
n −1
1
≤ 2 1 2
n − 2n
2
= 2.
n
2 cos(n)
Thus, if n2
< ε then n2 −1
.
q
Let ε > 0 be arbitrary. Let K ∈ N be such that 2ε < K. Then
2
K2
< ε. Therefore, if n ≥ K then n22 ≤ K22 and therefore
cos(n) | cos(n)|
=
n2 − 1 n2 − 1
1
≤ 2
n −1
1
≤ 2 1 2
n − 2n
2 2
= 2≤ 2
n K
<ε

This proves that limn→∞ xn = 0.

(−1)n n
Example 3.1.8. Does the sequence (xn) defined by xn = n+1
con-
verge?

76
 3.1. LIMITS OF SEQUENCES

A useful tool for proving convergence is the following.

Theorem 3.1.9
Let (xn) be a sequence and let L ∈ R. Let (an ) be a sequence of
positive numbers such that lim an = 0. Suppose that there exists
n→∞
M ∈ N such that

|xn − L| ≤ an , ∀ n ≥ M.

Then lim xn = L.
n→∞

Proof. Let ε > 0 be arbitrary. Since an → 0, there exists K1 ∈ N such

that an < ε for all n ≥ K1 . Let K = max{M, K1}. Then, if n ≥ K
then an < ε and |xn − L| ≤ an . Thus, if n ≥ K then

|xn − L| ≤ an < ε.

Example 3.1.10. Suppose that 0 < r < 1. Prove that limn→∞ rn = 0.

Proof. We first note that
1 1
r= 1 =
r
1+x
where x = 1r − 1 and since r < 1 then x > 0. Now, by Bernoulli’s
inequality (Example 1.2.5) it holds that (1 + x)n ≥ 1 + xn for all n ∈ N
and therefore
1
rn =
(1 + x)n
1
≤
1 + nx
1
< .
nx

77
 3.1. LIMITS OF SEQUENCES

1
Now since limn→∞ nx = 0 then it follows by Theorem 3.1.9 that

lim rn = 0.
n→∞

n2 −1
Example 3.1.11. Consider the sequence xn = 2n2 +3 . Prove that
limn→∞ xn = 12 .

Proof. We have that

n2 − 1 1 5 1
− =
2n2 + 3 2 2 (2n2 + 3)
5/2
< 2.
2n
Using the definition of the limit of a sequence, one can show that
limn→∞ 4n5 2 = 0 and therefore limn→∞ xn = 12 .

Notice that in the definition of the limit of a sequence, we wrote

“there exists a number L”. Could there be more than one number L
satisfying the definition of convergence of a sequence? Before we go
any further, we prove that if a sequence converges then it has a unique
limit.

Theorem 3.1.12: Uniqueness of Limits

A convergent sequence can have at most one limit.

Proof. Suppose that (xn) → L1 and that (xn) → L2. Let ε > 0 be
arbitrary. Then there exists K1 such that |xn − L1 | < ε/2 for all
n ≥ K1 and there exists K2 such that |xn − L2| < ε/2 for all n ≥ K2.
Let K = max{K1, K2}. Then for n ≥ K it holds that |xn − L1| < ε/2

78
 3.1. LIMITS OF SEQUENCES

and also |xn − L2 | < ε/2 and therefore

|L1 − L2| = |L1 − xn + xn − L2|

< |xn − L1| + |xn − L2 |
< ε/2 + ε/2
= ε.

Hence, |L1 − L2| < ε for all ε > 0, and therefore by Theoreom 2.2.7 we
conclude that |L1 − L2| = 0, that is, L1 − L2 = 0.

The ultimate long-time behavior of a sequence will not change if

we discard a finite number of terms of the sequence. To be pre-
cise, suppose that X = (x1, x2, x3, . . .) is a sequence and let Y =
(xm+1, xm+2, xm+3, . . .), that is, Y is the sequence obtained from X by
discarding the first m terms of X. In this case, we will call Y the m-tail
of X. The next theorem states, not surprisingly, that the convergence
properties of X and Y are the same.

Theorem 3.1.13: Tails of Sequences

Let X : N → R be a sequence and let Y : N → R be the sequence
obtained from X by discarding the first m ∈ N terms of X, in other
words, Y (n) = X(m + n). Then X converges to L if and only if Y
converges to L.

79
 3.1. LIMITS OF SEQUENCES

Exercises

Exercise 3.1.1. Write the first three terms of the recursively defined
sequence x1 = 1, xn+1 = 21 (xn + x2n ) for n ≥ 1.

Exercise 3.1.2. Use the definition of the limit of a sequence to establish

the following limits:
n+1 1
(a) lim =
n→∞ 3n 3
3n2 + 2 3
(b) lim =
n→∞ 4n2 + 1 4
(−1)nn
(c) lim 2 =0
n→∞ n + 1

Exercise 3.1.3.
(a) Prove that limn→∞ |xn | = 0 if and only if limn→∞ xn = 0.
(b) Combining the previous result and Example 3.1.10, prove that if
1 < r < 0 then limn→∞ rn = 0.
(c) Conclude that for any real number r ∈ R, if |r| < 1 then limn→∞ rn =
0.

Exercise 3.1.4. Let m ∈ N and assume that m ≥ 2.

(a) Prove that m1n < n1 for all n ∈ N.
(b) Use Theorem 3.1.9 to show that limn→∞ m1n = 0.
Note: Do not use Example 3.1.10 to show that limn→∞ m1n = 0.

Exercise 3.1.5. Suppose that S ⊂ R is non-empty and bounded above

and let u = sup S. Show that there exists a sequence (xn) such that
xn ∈ S for all n ∈ N and limn→∞ xn = u. Hint: If ε > 0 then clearly
u − ε < u. Since u = sup(S) there exists x ∈ S such that u − ε < x < u.
Example 3.1.6 is similar.

80
 3.1. LIMITS OF SEQUENCES

Exercise 3.1.6. Let (xn) be the sequence defined as

(
2n2 + 1, n < 50
xn = sin(2n)
n2 +1 , n ≥ 50.

Using the definition of the limit of a sequence, find limn→∞ xn.

81
 3.2. LIMIT THEOREMS

3.2 Limit Theorems

Proving that a particular number L ∈ R is the limit of a given sequence
is usually not easy because there is no systematic way to determine
a candidate limit L for a given arbitrary sequence. Instead, we are
frequently interested in just knowing if a given sequence converges or
not, and not so much on finding the actual limit. The theorems in this
section help us do just that. We begin with a definition.

Definition 3.2.1: Boundedness

A sequence (xn) is said to be bounded if there exists R ≥ 0 such
that |xn| ≤ R for all n ∈ N.

Example 3.2.2. Prove that (xn) is bounded if and only if there exists
numbers R1 and R2 such that R1 ≤ xn ≤ R2 for all n ∈ N.

Theorem 3.2.3: Convergence implies Boundedness

A convergent sequence is bounded.

Proof. Suppose that (xn) converges to L. Then there exists K ∈ N

such that |xn − L| < 1 for all n ≥ K. Let

R = 1 + max{|x1 − L|, |x2 − L|, . . . , |xK−1 − L|},

and we note that R ≥ 1. Then for all n ≥ 1 it holds that |xn − L| ≤ R.

Indeed, if n ≥ K then |xn − L| < 1 ≤ R and if 1 ≤ n ≤ K − 1 then

|xn − L| ≤ max{|x1 − L|, . . . , |xK−1 − L|} ≤ R.

Thus, for all n ≥ 1 it holds that L − R ≤ xn ≤ R + L and this proves

that (xn) is bounded.

82
 3.2. LIMIT THEOREMS

Theorem 3.2.4: Convergence under Absolute Value

If (xn) → L then (|xn|) → |L|.

Proof. Follows by the inequality ||xn | − |L|| ≤ |xn − L| (see Corol-

lary 2.3.6). Indeed, for any given ε > 0 there exists K ∈ N such that
|xn − L| < ε for all n ≥ K and therefore ||xn | − |L|| ≤ |xn − L| < ε for
all n ≥ K.

The following theorem describes how the basic operations of arith-

metic preserve convergence.

Theorem 3.2.5: Limit Laws

Suppose that (xn) → L and (yn) → M.

(a) Then (xn + yn ) → L + M and (xn − yn ) → L − M.

(b) Then (xnyn ) → LM.

 
xn L
(c) If yn 6= 0 and M 6= 0 then yn → M.

Proof. (i) By the triangle inequality

|xn + yn − (L + M)| = |xn − L + yn − M|

< |xn − L| + |yn − M|.

Let ε > 0. There exists K1 such that |xn − L| < ε/2 for n ≥ K1
and there exists K2 such that |yn − M| < ε/2 for n ≥ K2. Let K =
max{K1, K2}. Then for n ≥ K

|xn + yn − (L + M)| ≤ |xn − L| + |yn − M|

< ε/2 + ε/2
= ε.

83
 3.2. LIMIT THEOREMS

The proof for (xn − yn ) → L − M is similar.

(ii) We have that

|xnyn − LM| = |xnyn − yn L + yn L − LM|

≤ |xnyn − yn L| + |yn L − LM|
= |yn ||xn − L| + |L||yn − M|.

Now, (yn ) is bounded because it is convergent, and therefore |yn | ≤ R

for all n ∈ N for some R > 0. By convergence of (yn ) and (xn), there
ε ε
exists K ∈ N such that |xn − L| < 2R and |yn − M| < 2(|L|+1) for all
n ≥ K. Therefore, if n ≥ K then

|xn yn − LM| < |yn ||xn − L| + |L||yn − M|

< R|xn − L| + (|L| + 1)|yn − M|

ε ε
<R + (|L| + 1)
2R 2(|L| + 1)

= ε.
 
(iii) It is enough to prove that y1n → M1 and then use (ii). Now,
since M 6= 0 and yn 6= 0 then |yn | is bounded below by some positive
number, say R > 0. Indeed, (|yn|) → |M| and |yn | > 0. Thus, |y1n | < R1
for all n ∈ N. Now,

1 1 1
− = |yn − M|
yn M |yn ||M|

1
< |yn − M|.
R|M|

For ε > 0, there exists K ∈ N such that |yn − M| < R|M|ε for all

84
 3.2. LIMIT THEOREMS

n ≥ K. Therefore, for n ≥ K we have that

1 1 1
− < |yn − M|
yn M R|M|

1
< R|M|ε
R|M|

= ε.

Corollary 3.2.6

Suppose that (xn) → L. Then (xkn) → Lk for any k ∈ N.

The next theorem states that the limit of a convergent sequence of

non-negative terms is non-negative.

Theorem 3.2.7
Suppose that (xn) → L. If xn ≥ 0 for all n ∈ N then L ≥ 0.

Proof. We prove the contrapositive, that is, we prove that if L < 0 then
there exists K ∈ N such that xK < 0. Suppose then that L < 0. Let
ε > 0 be such that L + ε < 0. Since (xn) → L, there exists K ∈ N such
that xK < L + ε, and thus by transitivity we have xK < 0.

Corollary 3.2.8: Comparison

Suppose that (xn) and (yn) are convergent and suppose that there
exists M ∈ N such that xn ≤ yn for all n ≥ M. Then lim xn ≤
n→∞
lim yn .
n→∞

85
 3.2. LIMIT THEOREMS

Proof. Suppose for now that M = 1, that is, xn ≤ yn for all n ∈ N.

Consider the sequence zn = yn − xn . Then zn ≥ 0 for all n ∈ N and
(zn ) is convergent since it is the difference of convergent sequences. By
Theorem 3.2.7, we conclude that limn→∞ zn ≥ 0. But

lim zn = lim (yn − xn)

n→∞ n→∞
= lim yn − lim xn
n→∞ n→∞

and therefore limn→∞ yn −limn→∞ xn ≥ 0, which is the same as limn→∞ yn ≥

limn→∞ xn. If M > 1, then we can apply the theorem to the M-tail of
the sequences of (xn) and (yn ) and the result follows.

Corollary 3.2.9
Suppose that a ≤ xn ≤ b and lim xn = L. Then a ≤ L ≤ b.
n→∞

Proof. We have that 0 ≤ xn − a ≤ b − a. The sequence yn = b − a is

constant and converges to b − a. The sequence zn = xn − a converges
to L − a. Therefore, by the previous theorem, 0 ≤ L − a ≤ b − a, or
a ≤ L ≤ b.

Theorem 3.2.10: Squeeze Theorem

Suppose that yn ≤ xn ≤ zn for all n ∈ N. Assume that (yn) → L
and also (zn ) → L. Then (xn) is convergent and (xn) → L.

Proof. Let ε > 0 be arbitrary. There exists K1 ∈ N such that L −

ε < yn < L + ε for all n ≥ K1 and there exists K2 ∈ N such that
L − ε < zn < L + ε for all n ≥ K2. Let K = max{K1, K2}. Then if
n ≥ K then
L − ε < yn ≤ xn ≤ zn < L + ε.

86
 3.2. LIMIT THEOREMS

Therefore, for n ≥ K we have that

L − ε < xn < L + ε

and thus limn→∞ xn = L.

Remark 3.2.11. Some people call the Squeeze Theorem the Sandwich
Theorem; we are not those people.

Example 3.2.12. Let 0 < a < b and let xn = (an + bn )1/n. Prove that
lim xn = b.
n→∞

Theorem 3.2.13: Ratio Test

Let (xn) be a sequence such that xn > 0 for all n ∈ N and sup-
xn+1
pose that L = lim exists. If L < 1 then (xn) converges and
n→∞ xn
lim xn = 0.
n→∞

Proof. Let r ∈ R be such that L < r < 1 and set ε = r − L. There

exists K ∈ N such that
xn+1
<L+ε=r
xn
for all n ≥ K. Therefore, for all n ≥ K we have that

0 < xn+1 < rxn.

Thus, xK+1 < rxK , and therefore xK+2 < rxK+1 < r2 xK , and induc-
tively for m ≥ 1 it holds that

xK+m < rm xK .

87
 3.2. LIMIT THEOREMS

Hence, the tail of the sequence (xn) given by (ym ) = (xK+1, xK+2, . . . , )
satisfies
0 < ym < rm xK .
Since 0 < r < 1 it follows that limm→∞ rm = 0 and therefore limm→∞ ym =
0 by the Squeeze theorem. This implies that (xn) converges to 0
also.

88
 3.2. LIMIT THEOREMS

Exercises

Exercise 3.2.1. Use the Limit Theorems to prove that if (xn) converges
and (xn + yn ) converges then (yn ) converges. Give an example of two
sequences (xn) and (yn ) such that both (xn) and (yn) diverge but (xn +
yn ) converges.

Exercise 3.2.2. Is the sequence yn = (−1)nn4 convergent? Explain.

Exercise 3.2.3. Let (xn) and (yn) be sequences in R. Suppose that

limn→∞ xn = 0 and that (yn ) is bounded. Prove that limn→∞ xnyn = 0.

Exercise 3.2.4. Show that if (xn) and (yn ) are sequences such that
(xn) and (xn + yn ) are convergent, then (yn) is convergent.

Exercise 3.2.5. Give examples of the following:

(a) Divergent sequences (xn) and (yn ) such that zn = xnyn converges.

(b) Divergent sequences (xn) and (yn ) such that zn = xnyn diverges.

(c) A divergent sequence (xn) and a convergent sequence (yn) such

that zn = xn yn converges.

(d) A divergent sequence (xn) and a convergent sequence (yn) such

that zn = xn yn diverges.

Exercise 3.2.6. Let (xn) and (yn ) be sequences and suppose that (xn)
converges to L. Assume that for every ε > 0 there exists M ∈ N such
that |xn − yn | < ε for all n ≥ M. Prove that (yn) also converges to L.

Exercise 3.2.7. Let (xn) be a sequence and define a sequence (yn ) as

x1 + x2 + · · · + xn
yn =
n
for n ∈ N. Show that if limn→∞ xn = 0 then limn→∞ yn = 0.

89
 3.2. LIMIT THEOREMS

Exercise 3.2.8. Let (xn) be a convergent sequence with limit L. Let

f (x) = a0 + a1 x + · · · + ak xk be a polynomial. Use the Limit Theorems
to prove that the sequence (yn ) defined by yn = f (xn) is convergent
and find the limit of (yn ).

Exercise 3.2.9. Apply the Limit Theorems to find the limits of the
following sequences:
r
2n2 + 3
(a) xn =
n2 + 1
(b) xn = (2 + 1/n)2
n+1
(c) xn = √
n n
(d) xn = 2n/n!

Exercise 3.2.10. Let (xn) be a sequence such that xn 6= 0 for all n ∈ N.

Suppose that limn→∞ xn = L and L > 0. Let w = inf{|xn | : n ∈ N}.
Prove that w > 0.

Exercise 3.2.11. Let (xn) be a sequence of positive numbers such that

xn+1
lim = L > 1. Show that (xn) is not bounded and hence is not
n→∞ xn
convergent.

90
 3.3. MONOTONE SEQUENCES

3.3 Monotone Sequences

As we have seen, a convergent sequence is necessarily bounded, and it

is straightforward to construct examples of sequences that are bounded
but not convergent, for example, (xn) = (1, 0, 1, 0, 1, 0, . . .). In this
section, we prove the Monotone Convergence Theorem which says that
a bounded sequence whose terms increase (or decrease) must necessarily
converge.

Definition 3.3.1: Monotone Sequences

Let (xn) be a sequence.

(i) We say that (xn) is increasing if xn ≤ xn+1 for all n ∈ N.

(ii) We say that (xn) is decreasing if xn+1 ≤ xn for all n ∈ N.

(iii) We say that (xn) is monotone if (xn) is either increasing or

decreasing.

Example 3.3.2. Prove that if (xn) is increasing then (xn) is bounded

below. Similarly, prove that if (xn) is decreasing then (xn) is bounded
above.

91
 3.3. MONOTONE SEQUENCES

Theorem 3.3.3: Monotone Convergence Theorem

If (xn) is bounded and monotone then (xn) is convergent. In par-
ticular:

(i) if (xn) is bounded above and increasing then

lim xn = sup{xn : n ∈ N},

n→∞

(ii) if (xn) is bounded below and decreasing then

lim xn = inf{xn : n ∈ N}.

n→∞

Proof. Suppose that (xn ) is bounded above and increasing. Let u =

sup{xn | n ∈ N} and let ε > 0 be arbitrary. Then by the properties of
the supremum, there exists xK such that u − ε < xK ≤ u. Since (xn)
is increasing, and u is an upper bound for the range of the sequence, it
follows that xK ≤ xn ≤ u for all n ≥ K. Therefore, u − ε < xn ≤ u for
all n ≥ K. Clearly, this implies that u − ε < xn < u + ε for all n ≥ K.
Since ε > 0 was arbitrary, this proves that (xn) converges to u.

Suppose now that (xn) is bounded below and decreasing. Let w =

inf{xn | n ∈ N} and let ε > 0 be arbitrary. Then by the properties of
the infimum, there exists xK such that w ≤ xK < w + ε. Since (xn)
is decreasing, and w is a lower bound for the range of the sequence, it
follows that w ≤ xn ≤ xK for all n ≥ K. Therefore, w ≤ xn < w + ε
for all n ≥ K. Hence, w − ε < xn < w + ε for all n ≥ K. Since ε > 0
was arbitrary, this proves that (xn) converges to w.

The Monotone Convergence Theorem (MCT) is an important tool

in real analysis and we will use it frequently; notice that it is more-

92
 3.3. MONOTONE SEQUENCES

or-less a direct consequence of the Completeness Axiom. In fact, we

could have taken as our starting axiom the MCT and then proved the
Completeness property of R.

Example 3.3.4. By the MCT, a bounded sequence that is also mono-

tone is convergent. However, it is easy to construct a convergent se-
quence that is not monotone. Provide such an example.

The MCT can be used to show convergence of recursively defined

sequences. To see how, suppose that (xn) is defined recursively as
x1 = a and
xn+1 = f (xn)

where f is some given function. For example, say x1 = 2 and

1
xn+1 = 2 + .
xn

Hence, in this case f (x) = 2 + x1 . If (xn) is bounded and increasing

then by the MCT (xn) converges, but we do not know what the limit
is. However, for example, if f is a polynomial/rational function of x
then we can conclude that L = limn→∞ xn must satisfy the equation

L = f (L).

Indeed, if f is a polynomial/rational function then by the Limit Laws

we have
lim f (xn) = f ( lim xn) = f (L).
n→∞ n→∞

But xn+1 = f (xn) and therefore limn→∞ xn+1 = f (L), which is equiva-
lent to limn→∞ xn = f (L) since (xn+1) is just the 1-tail of the sequence
(xn). Therefore, L = f (L) as claimed. From the equation L = f (L) we
can solve for L if possible.

93
 3.3. MONOTONE SEQUENCES

Example 3.3.5. Consider the sequence (xn) defined recursively as x1 =

1 and
1
xn+1 = 21 xn + , ∀ n ≥ 1.
4
Prove that (xn) converges and find the limit.

Proof. We prove by induction that 21 ≤ xn for all n ∈ N, that is, (xn) is

bounded below by 12 . First of all, it is clear that 12 ≤ x1 . Now assume
that 21 ≤ xn for some n ∈ N. Then
1
xn+1 = 21 xn +
4
1 1 1
≥ · +
2 2 4
1
= .
2
Hence, (xn) is bounded below by 12 . We now prove that (xn) is decreas-
ing. We compute that x2 = 21 + 14 = 34 , and thus x2 < x1. Assume now
that xn < xn−1 for some n ∈ N. Then
1 1
xn+1 < xn−1 + = xn .
2 4
Hence, by induction we have shown that (xn) is decreasing. By the
MCT, (xn) is convergent. Suppose that (xn) → L. Then also (xn+1) →
L and the sequence yn = 12 xn + 41 converges to 12 L + 14 . Therefore,
L = 12 L + 14 and thus L = 1/2.

Before we embark on the next example, we recall that

2 n−1 1 − rn
1 + r + r + ··· + r =
1−r
and if 0 < r < 1 then 0 < rn < r < 1 and therefore
1 − rn 1
< .
1−r 1−r
94
 3.3. MONOTONE SEQUENCES

Example 3.3.6. Consider the sequence

1 1 1
xn = 1 + + + ··· + .
1! 2! n!
Note that this can be defined recursively as x1 = 1 and xn+1 = xn +
1
(n+1)! . Prove that (xn ) converges.

Proof. We will prove by the MCT that (xn) converges. By induction,

one can show that 2n−1 < n! for all n ≥ 3. Therefore,
1 1 1
xn < 1 + + + · · · + n−1
1 2 2
n
1 − (1/2)
=1+
1 − (1/2)
1
<1+
1 − (1/2)
= 3.

Hence, xn < 3 and therefore (xn) is bounded. Now, since xn+1 =

1
xn + (n+1)! then clearly xn < xn+1. Thus (xn) is increasing. By the
MCT, (xn) converges. You might recognize that limn→∞ xn = e =
2.71828 . . ..
√
Example 3.3.7. Let x1 = 0 and let xn+1 = 2 + xn for n ≥ 1. Prove
that (xn) converges and find its limit.
√
Proof. Clearly, x1 < x2 = 2. Assume by induction that xk > xk−1 for
some k ∈ N. Then
√
xk+1 = 2 + xk
p
> 2 + xk−1
= xk .

Hence, (xn) is an increasing sequence. We now prove that (xn) is

bounded above. Clearly, x1 < 2. Assume that xk < 2 for some k ∈ N.

95
 3.3. MONOTONE SEQUENCES

√ √
Then xk+1 = 2 + xk < 2 + 2 = 2. This proves that (xn) is bounded
above. By the MCT, (xn) converges, say to L. Moreover, since xn ≥ 0
√
(as can be proved by induction), then L ó 0. Therefore, L = 2 + L
and then L2 − L − 2 = 0. Hence, L = 1± 21+8 = 1±3
2 . Since L ≥ 0 then
L = 2.

n
Example 3.3.8. Consider the sequence xn = 1 + n1 . We will show
that (xn) is bounded and increasing, and therefore by the MCT (xn)
convergent. The limit of this sequence is the number e. From the
binomial theorem
n  
X n 1
xn =
k nk
k=0
n 1 n(n − 1) 1 n(n − 1)(n − 2)
=1+ + + + ···
n 2! n2 3! n3
1 n(n − 1)(n − 2) · · · (n − (n − 1))
+
n! nk
1 1
= 1+1+ (1 − n1 ) + (1 − n1 )(1 − n1 ) + · · ·
2! 3!
1
+ (1 − n1 )(1 − n1 ) · · · (1 − (n − 1)/n)
n!
1 1 1
< 1+1+ + + ··· +
2! 3! n!
1 1 1
< 1+1+ + 2 + · · · + n−1
2 2 2
1 − (1/2)n
=1+
1 − 1/2

<3

where we used that 2n−1 < n! for all n ≥ 3. This shows that (xn ) is

96
 3.3. MONOTONE SEQUENCES

bounded. Now, for each 1 ≤ k ≤ n, we have that

 
n 1 n(n − 1)(n − 2) · · · (n − (k − 1))
=
k nk nk
= (1 − n1 )(1 − n2 ) · · · (1 − k−1
n ).

And similarly,
 
n+1 1 1 2 k−1
= (1 − n+1 )(1 − n+1 ) · · · (1 − n+1 ).
k (n + 1)k


j
It is clear that (1 − nj ) < (1 − n+1 ) for all 1 ≤ j ≤ n. Hence, nk n1k <
n+1

1
k (n+1)k
. Therefore, xn < xn+1, that is, (xn) is increasing. By the
MCT, (xn) converges to sup{xn : n ∈ N}.

97
 3.3. MONOTONE SEQUENCES

Exercises

Exercise 3.3.1. Let (xn) be an increasing sequence, let (yn ) be a de-

creasing sequence, and assume that xn ≤ yn for all n ∈ N. Prove that
lim xn and lim yn exist, and that lim xn ≤ lim yn . Note: Recall that
n→∞ n→∞ n→∞ n→∞
a sequence (xn ) is bounded if there exist constants R1 , R2 > 0 (independent of n)
such that R1 ≤ xn ≤ R2 for all n ∈ N.

Exercise 3.3.2. Let x1 = 8 and let xn+1 = 12 xn + 2 for n ≥ 1. Prove

that (xn) is bounded and monotone. Find the limit of (xn).
n 2
Exercise 3.3.3. Let x1 = 1 and let xn+1 = x for n ≥ 1. Prove
n+1 n
that (xn) is bounded and monotone. Find the limit of (xn). Hint: Using
induction to prove that (xn ) is monotone will not work with this sequence. Instead,
work with xn+1 directly to prove that (xn ) is monotone.

Exercise 3.3.4. True or false, a convergent sequence is necessarily

monotone? If it is true, prove it. If it is false, give an example.

98
 3.4. BOLZANO-WEIERSTRASS THEOREM

3.4 Bolzano-Weierstrass Theorem

We can gather information about a sequence by studying its subse-
quences. Loosely speaking, a subsequence of (xn) is a new sequence
(yk ) such that each term yk is from the original sequence (xn) and the
term yk+1 appears to the “right” of the term yk in the original sequence
(xn). Let us be precise about what we mean to the “right”.

Definition 3.4.1: Subsequences

Let (xn) be a sequence. A subsequence of (xn) is a sequence of
the form (xn1 , xn2 , xn3 , . . .) where n1 < n2 < n3 < · · · is a sequence
of strictly increasing natural numbers. A subsequence of (xn) will
be denoted by (xnk ).

The notation (xnk ) of a subsequence indicates that the indexing variable

is k ∈ N. The selection of the elements of (xn) to form a subsequence
(xnk ) does not need to follow any particular well-defined pattern but
only that n1 < n2 < n3 < · · · . Notice that for any increasing sequence
n1 < n2 < n3 < · · · of natural numbers, we have

k ≤ nk

for all k ≥ 1.

Example 3.4.2. An example of a subsequence of xn = n1 is the se-

quence (yk ) = (1, 31 , 15 , 71 , . . .). Here we have chosen the odd terms of the
sequence (xn) to create (yk ). In other words, if we write that yk = xnk
then nk = 2k − 1. Another example of a subsequence of (xn) is ob-
tained by taking the even terms to get the subsequence ( 21 , 41 , 61 , . . .), so
that here nk = 2k. In general, we can take any increasing selection

99
 3.4. BOLZANO-WEIERSTRASS THEOREM

n1 < n2 < n3 < · · · , such as

1 1 1
( 11 , 303 , 2000 , . . .)

to form a subsequence of (xn).

Example 3.4.3. Two subsequences of

(xn) = (1, −1, 21 , −1, 31 , −1, 14 , −1, . . .)

(−1, −1, −1, . . . , ) and (1, 21 , 31 , . . . , ). Both of these subsequences con-

verge to distinct limits.

Example 3.4.4 (Important). We proved that Q is countable and hence

there is a bijection f : N → Q. The bijection f defines a sequence

(xn) = (x1, x2, x3, x4, . . .)

where xn = f (n). Let L ∈ R be arbitrary. By the density of Q in

R, there exists xn1 ∈ Q such that xn1 ∈ (L − 1, L + 1). Now consider
the interval (L − 12 , L + 21 ). It has infinitely many distinct rational
numbers (by the Density Theorem). Therefore, there exists n2 > n1
such that xn2 ∈ (L − 12 , L + 12 ). Consider now the interval (L − 31 , L +
1
3
). It has infinitely many rational numbers, and therefore there exists
n3 > n2 such that xn3 ∈ (L − 13 , L + 13 ). By induction, there exists
a subsequence (xnk ) of (xn) such that |xnk − L| < k1 for all k ≥ 1.
Therefore, limk→∞ xnk = L. We proved the following: For any real
number L there exists a sequence of rational numbers that converges
to L.

The following theorem is a necessary condition for convergence.

100
 3.4. BOLZANO-WEIERSTRASS THEOREM

Theorem 3.4.5
If (xn) → L then every subsequence of (xn) converges to L.

Proof. Let ε > 0. Then there exists K ∈ N such that |xn − L| < ε for
all n ≥ K. Since nK ≥ K and nK < nK+1 < . . ., then |xnk − L| < ε for
all k ≥ K.

The contrapositive of the previous theorem is worth stating.

Theorem 3.4.6
Let (xn) be a sequence.

(i) If (xn) has two subsequences converging to distinct limits then

(xn) is divergent.

(ii) If (xn) has a subsequence that diverges then (xn) diverges.

The following is a very neat result that will supply us with a very
short proof of the main result of this section, namely, the Bolzano-
Weierstrass Theorem.

Theorem 3.4.7
Every sequence has a monotone subsequence.

Proof. Let (xn) be an arbitrary sequence. We will say that the term xm
is a peak if xm ≥ xn for all n ≥ m. In other words, xm is a peak if it is
an upper bound of all the terms coming after it. There are two possible
cases for (xn), either it has an infinite number of peaks or it has a finite
number of peaks. Suppose that it has an infinite number of peaks, say
xm1 , xm2 , . . ., and we may assume that m1 < m2 < m3 < · · · . Then,

101
 3.4. BOLZANO-WEIERSTRASS THEOREM

xm1 ≥ xm2 ≥ xm3 ≥ · · · , and therefore (xmk ) is a decreasing sequence.

Now suppose that there are only a finite number of peaks and that xm
is the last peak. Then xn1 = xm+1 is not a peak and therefore there
exists n2 > n1 such that xn2 ≥ xn1 . Similarly, xn2 is not a peak and
therefore there exists n3 > n2 such that xn3 ≥ xn2 . Hence, by induction,
there exists a subsequence (xnk ) that is increasing.

Theorem 3.4.8: Bolzano-Weierstrass

Every bounded sequence contains a convergent subsequence.

Proof. Let (xn ) be an arbitrary bounded sequence. By Theorem 3.4.7,

(xn) has a monotone subsequence (xnk ). Since (xn) is bounded then
so is (xnk ). By the MCT applied to (xnk ) we conclude that (xnk ) is
convergent.

We will give a second proof of the Bolzano-Weierstrass Theorem that

is more “hands-on”.

Another proof of Bolzano-Weierstrass. If (xn) is a bounded sequence,

then there exists a1 , b1 ∈ R such that a1 ≤ xn ≤ b1 for all n ∈ N. We
will apply a recursive bisection algorithm to hunt down a converging
subsequence of (xn). Let m1 = (a1 +b 2
1)
be the mid-point of the interval
[a1 , b1]. Then at least one of the subsets I1 = {n ∈ N : a1 ≤ xn ≤ m1 }
or J1 = {n ∈ N : m1 ≤ xn ≤ b1} is infinite; if it is I1 then choose
some xn1 ∈ [a1 , m1] and let a2 = a1 and b2 = m1 ; otherwise choose some
xn1 ∈ [m1 , b1] and let a2 = m1 , b2 = b1 . In any case, it is clear that
(b2 − a2 ) = (b1 −a
2
1)
, that a1 ≤ a2 and that b1 ≥ b2 . Now let m2 = (a2 +b
2
2)

be the mid-point of the interval [a2 , b2] and let I2 = {n ∈ N : a2 ≤

xn ≤ m2 } and let J2 = {n ∈ N : m2 ≤ xn ≤ b2}. If I2 is infinite then
choose some xn2 ∈ [a2 , m2] and let a3 = a2 , b3 = m2 ; otherwise choose

102
 3.4. BOLZANO-WEIERSTRASS THEOREM

some xn2 ∈ [m2 , b2] and let a3 = m2 and b3 = b2 . In any case, it is

clear that (b3 − a3 ) = (b12−a2
1)
, that a2 ≤ a3 and b3 ≤ b2. By induction,
there exists sequences (ak ), (bk ), and (xnk ) such that ak ≤ xnk ≤ bk
−a1 )
and (bk − ak ) = (b21k−1 , (ak ) is increasing and (bk ) is decreasing. It
is clear that ak ≤ b and a ≤ bk for all k ∈ N. Hence, by the MCT,
−a1 )
(ak ) and (bk ) are convergent. Moreover, since (bk − ak ) = (b21k−1 then
lim(bk − ak ) = 0 and consequently lim ak = lim bk = L. By the Squeeze
theorem we conclude that limk→∞ xnk = L.

Notice that the proofs of the Bolzano-Weierstrass Theorem rely on

the Monotone Convergence Theorem and the latter relies on the Com-
pleteness Axiom. We therefore have the following chain of implications:

Completeness =⇒ MCT =⇒ Bol-Wei

It turns out that if we had taken as our starting axiom the Bolzano-
Weierstrass theorem then we could prove the Completeness property
and then of course the MCT. In other words, all three statements are
equivalent:
Completeness ⇐⇒ MCT ⇐⇒ Bol-Wei

103
 3.4. BOLZANO-WEIERSTRASS THEOREM

Exercises

Exercise 3.4.1. Prove that the following sequences are divergent.

(a) xn = 1 − (−1)n + 1/n

(b) xn = sin(nπ/4)

(Hint: Theorem 3.4.6)

Exercise 3.4.2. Suppose that xn ≥ 0 for all n ∈ N and suppose that

lim (−1)nxn = L exists. Prove that L = 0 and that also lim xn = L.
n→∞ n→∞
(Hint: Consider subsequences of (−1)nxn .)

Exercise 3.4.3. Let (xn) be a sequence.

(a) Suppose that (xn) is increasing. Prove that if (xn) has a sub-
sequence (xnk ) that is bounded above then (xn) is also bounded
above.

(b) Suppose that (xn) is decreasing. Prove that if (xn) has a sub-
sequence (xnk ) that is bounded below then (xn) is also bounded
below.

Exercise 3.4.4. True or false: If (xn) is bounded and diverges then

(xn) has two subsequences that converge to distinct limits. Explain.

Exercise 3.4.5. Give an example of a sequence (xn) with the following


property: For each number L ∈ 1, 12 , 31 , 41 , . . . there exists a subse-
quence (xnk ) such that xnk → L. Hint: If you are spending a lot of
time on this question then you have not been reading this textbook
carefully.

104
 3.4. BOLZANO-WEIERSTRASS THEOREM

Exercise 3.4.6. Suppose that x1 and y1 satisfy 0 < x1 < y1 and define
√ xn + yn
xn+1 = xnyn , yn+1 =
2
for n ≥ 1. Prove that (xn) and (yn) are convergent and that lim xn =
√
lim yn . (Hint: First show that xy ≤ (x + y)/2 for any x, y > 0.)

Exercise 3.4.7. Let (xn) be a sequence and define the sequence (yn)
as
x1 + x2 + · · · + xn
yn =
n
Prove that if (xn) → L then (yn) → L.

105
 3.5. LIMSUP AND LIMINF

3.5 limsup and liminf

The behavior of a convergent sequence is easy to understand. Indeed,
if (xn) → L then eventually the terms of (xn) will be arbitrarily close to
L for n sufficiently large. What else is there to say? In this section, we
focus on bounded sequences that do not necessarily converge. The idea
is that we would like to develop a limit concept for these sequences,
and in particular, a “limiting upper bound”.
Let (xn) be an arbitrary sequence and introduce the set S defined
as the set of all the limits of convergent subsequences of (xn), that is,

S = {L ∈ R | (xnk ) → L} .

We will call S the subsequences limit set of (xn).

Example 3.5.1. If (xn) is bounded and S is the subsequences limit set

of (xn) explain why S is non-empty.

Example 3.5.2. Here are six examples of sequences and the corre-
sponding subsequences limit set. Notice that in the cases where (xn)

(xn) S
1 1
(1, 2 , 3 , . . .) {0}
(1, −1, 1, −1, . . .) {1, −1}
(1, 2, 3, 4, . . .) ∅
3 1 5 1 7
(1, 2 , 3 , 4 , 5 , 6 , . . .) {0, 1}
1 1 3 1 7
(0, 1, 2 , 4 , 4 , 8 , . . . , 8 , . . .) [0, 1]
(rn) enumeration of Q R
Table 3.2: Limits of subsequences

is bounded, the set S is also bounded, which is as expected since if

a ≤ xn ≤ b then for any convergent subsequence (xnk ) of (xn) we nec-
essarily have a ≤ xnk ≤ b and therefore a ≤ limk→∞ xnk ≤ b.

106
 3.5. LIMSUP AND LIMINF

In general, we have seen that for a general set S, sup(S) and inf(S)
are not necessarily in S. This, however, is not the case for the subse-
quences limit set.

Lemma 3.5.3
Let (xn) be a bounded sequence and let S be its subsequences limit
set. Then sup(S) ∈ S and inf(S) ∈ S. In other words, there exists a
subsequence (xnk ) of (xn) such that lim xnk = sup(S) and similarly
k→∞
there exists a subsequence (ynk ) of (xn) such that lim ynk = inf(S).
k→∞

Proof. Let u = sup(S). If ε > 0 then there exists s ∈ S such that

u − ε < s ≤ u. Since s ∈ S, there exists a subsequence (xnk ) of
(xn) that converges to s. Therefore, there exists K ∈ N such that
u − ε < xnk < u + ε for all k ≥ K. Hence, for each ε, the inequality
u − ε < xn < u + ε holds for infinitely many n. Consider ε1 = 1. Then
there exists xn1 such that u − ε1 < xn1 < u + ε1 . Now take ε2 = 12 .
Since u − ε2 < xn < u + ε holds for infinitely many n, there exists xn2
such that u − ε2 < xn2 < u + ε2 and n2 > n1 . By induction, for εk = k1 ,
there exists xnk such that u − εk < xnk < u + εk and nk > nk−1. Hence,
the subsequence (xnk ) satisfies |xnk − u| < k1 and therefore (xnk ) → u.
Therefore, u = sup(S) ∈ S.

By Lemma 3.5.3, if (xn) is a bounded sequence then there exists

a convergent subsequence of (xn) whose limit is larger than any other
limit of a convergent subsequence of (xn). This leads to the following
definition.

107
 3.5. LIMSUP AND LIMINF

Definition 3.5.4
Let (xn) be a bounded sequence and let S be its subsequences limit
set. We define the limit superior of (xn) as

lim sup xn = sup S

and the limit inferior of (xn) as

lim inf xn = inf S.

By Lemma 3.5.3, lim sup xn is simply the largest limit of all convergent
subsequences of (xn) while lim inf xn is the smallest limit of all con-
vergent subsequences of (xn). Notice that by definition it is clear that
lim inf xn ≤ lim sup xn . The next theorem gives an alternative charac-
terization of lim sup xn and lim inf xn. The idea is that lim sup xn is a
sort of limiting supremum and lim inf xn is a sort of limiting infimum
of a bounded sequence (xn).

Theorem 3.5.5
Let (xn) be a bounded sequence and let L∗ ∈ R. The following are
equivalent:

(i) L∗ = lim sup xn

(ii) If ε > 0 then there are at most finitely many xn such that
L∗ + ε < xn and infinitely many xn such that L∗ − ε < xn .

(iii) Let um = sup{xn : n ≥ m}. Then L∗ = lim um = inf{um :

m→∞
m ∈ N}.

Proof. (i)→(ii) Let S denote the subsequences limit set of (xn). By

108
 3.5. LIMSUP AND LIMINF

definition, L∗ = lim sup xn = sup(S) and by Lemma 3.5.3 we have that

L∗ ∈ S. Hence, there exists a subsequence of (xn) converging to L∗ and
thus L∗ − ε < xn < L∗ + ε holds for infinitely many n. In particular
L∗ − ε < xn holds for infinitely many n. Suppose that L∗ + ε < xn
holds infinitely often. Now xn ≤ M for all n and some M > 0. Since
the inequality L∗ + ε < xn holds infinitely often, there exists a sequence
n1 < n2 < · · · such that L∗ +ε < xnk ≤ M for all k ∈ N. We can assume
that (xnk ) is convergent (because it is bounded and we can pass to a
subsequence by the MCT) and thus L∗ + ε ≤ lim xnk ≤ M. Hence we
n→∞
have proved that the subsequence (xnk ) converges to a number greater
than L∗ which contradicts the definition of L∗ = sup(S).

(ii)→(iii) Let ε > 0. Since L∗ + ε/2 < xm holds for finitely many
m, there exists M such that xm ≤ L∗ + ε/2 for all m ≥ M. Hence,
L∗ + ε/2 is an upper bound of {xn | n ≥ m} and thus um < L∗ + ε.
Since (um) is decreasing, we have that um < L∗ + ε for all m ≥ M.
Now, L∗ − ε/2 < xn holds infinitely often and thus L∗ − ε < um for all
m ∈ N. Hence, L∗ − ε < um < L∗ + ε for all m ≥ M. This proves the
claim.

(iii)→(i) Let (xnk ) be a convergent subsequence. Since nk ≥ k, by

definition of uk , we have that xnk ≤ uk . Therefore, lim xnk ≤ lim uk =
L∗. Hence, L∗ is an upper bound of S. By definition of uk , there
exists xn1 such that u1 − 1 < xn1 ≤ u1. By induction, there exists
a subsequence (xnk ) such that uk − k1 < xnk ≤ uk . Hence, by the
Squeeze Theorem, L∗ = lim xnk . Hence, L∗ ∈ S and thus L∗ = sup S =
lim inf xn .

Example 3.5.6. Let pn denote the nth prime number, that is p1 = 2,

p2 = 3, p3 = 5, and so on. The numbers pn and pn+1 are called twin

109
 3.5. LIMSUP AND LIMINF

primes if pn+1 − pn = 2. The Twin Prime Conjecture is that

lim inf(pn+1 − pn ) = 2

In other words, the Twin Prime Conjecture is that there are infinitely
many pairs of twin primes.

We end this section with the following interesting theorem that

says that if the subsequences limit set of a bounded sequence (xn)
consists of a single number L then the sequence (xn) also converges to
L.

Theorem 3.5.7
Let (xn) be a bounded sequence and let L ∈ R. If every convergent
subsequence of (xn) converges to L then (xn) converges to L.

Proof. Suppose that (xn) does not converge to L. Then, there exists
ε > 0 such that for every K ∈ N there exists n ≥ K such that |xn −L| ≥
ε. Let K1 ∈ N. Then there exists n1 ≥ K1 such that |xn1 − L| ≥ ε.
Then there exists n2 > n1 + 1 such that |xn2 − L| ≥ ε. By induction,
there exists a subsequence (xnk ) of (xn) such that |xnk − L| ≥ ε for all
k ∈ N. Now (xnk ) is bounded and therefore by Bolzano-Weierstrass
has a convergent subsequence, say (zk ), which is also a subsequence
of (xn). By assumption, (zk ) converges to L, which contradicts that
|xnk − L| ≥ ε for all k ∈ N.

Another way to say Theorem 3.5.7 is that if (xn) is bounded and L =

lim sup xn = lim inf xn then (xn) → L. The converse, by the way,
has already been proved: if (xn) → L then every subsequence of (xn)
converges to L and therefore L = lim sup xn = lim inf xn .

110
 3.5. LIMSUP AND LIMINF

Exercises

Exercise 3.5.1. Determine the lim sup xn and lim inf xn for each case:
n→∞ n→∞

(a) xn = 3 + (−1)n(1 + 1/n)

(b) xn = 1 + sin(nπ/2)

(c) xn = (2 − 1/n)(−1)n

Exercise 3.5.2. Let (xn) and (yn) be bounded sequences. Let (zn ) be
the sequence zn = xn + yn . Show that

lim sup zn ≤ lim sup xn + lim sup yn

n→∞ n→∞ n→∞

In other words, prove that

lim sup(xn + yn ) ≤ lim sup xn + lim sup yn

n→∞ n→∞ n→∞

Exercise 3.5.3. Let (xn) and (yn ) be bounded sequences. Show that
if xn ≤ yn for all n then

lim sup xn ≤ lim sup yn .

n→∞ n→∞

111
 3.6. CAUCHY SEQUENCES

3.6 Cauchy Sequences

Up until now, the Monotone Convergence theorem is our main tool for
determining that a sequence converges without actually knowing what
the the limit is. It is a general sufficient condition for convergence. In
this section, we prove another groundbreaking general sufficient condi-
tion for convergence known as the Cauchy criterion. Roughly speak-
ing, the idea is that if the terms of a sequence (xn) become closer and
closer to one another as n → ∞ then the sequence ought to converge.
A sequence whose terms become closer and closer to one another is
called a Cauchy sequence.

Definition 3.6.1: Cauchy Sequences

A sequence (xn) is said to be a Cauchy sequence if for every
ε > 0 there exists a natural number K such that if n, m ≥ K then
|xn − xm| < ε.

In other words, (xn) is a Cauchy sequence if the difference |xn − xm| is

arbitrarily small provided that both n and m are sufficiently large.

Example 3.6.2. Prove directly using the definition of a Cauchy se-

quence that if (xn) and (yn ) are Cauchy sequences then the sequence
zn = |xn − yn | is a Cauchy sequence.

Not surprisingly, a convergent sequence is indeed a Cauchy se-

quence.

Lemma 3.6.3
If (xn) is convergent then it is a Cauchy sequence.

112
 3.6. CAUCHY SEQUENCES

Proof. Suppose that (xn) → L. Let ε > 0 and let K be sufficiently

large so that |xn − L| < ε/2 for all n ≥ K. If n, m ≥ K then by the
triangle inequality,

|xn − xm| = |xn − L + L − xm |

≤ |xn − L| + |xm − L|
< ε/2 + ε/2
= ε.

This proves that (xn ) is Cauchy.

A Cauchy sequence is bounded.

Lemma 3.6.4: Cauchy implies Boundedness

If (xn) is a Cauchy sequence then (xn) is bounded.

Proof. The proof is similar to the proof that a convergent sequence is

bounded.

Theorem 3.6.5: Cauchy Criterion

The sequence (xn) is convergent if and only if (xn) is a Cauchy
sequence.

Proof. In Lemma 3.6.3 we already showed that if (xn) converges then

it is a Cauchy sequence. To prove the converse, suppose that (xn) is a
Cauchy sequence. By Lemma 3.6.4, (xn) is bounded. Therefore, by the
Bolzano-Weierstrass theorem there is a subsequence (xnk ) of (xn) that
converges, say it converges to L. We will prove that (xn) also converges
to L. Let ε > 0 be arbitrary. Since (xn) is Cauchy there exists K ∈ N
such that if n, m ≥ K then |xn − xm | < ε/2. On the other hand, since

113
 3.6. CAUCHY SEQUENCES

(xnk ) → L there exists nM ≥ K such that |xnM − L| < ε/2. Therefore,

if n ≥ K then

|xn − L| = |xn − xnM + xnM − L|

≤ |xn − xnM | + |xnM − L|

< ε/2 + ε/2

= ε.

This proves that (xn ) converges to L.

Example 3.6.6. Let 0 < r < 1 and suppose that |xn − xn+1| ≤ rn
1
for all n ∈ N. Using the fact that 1 + r + r2 + · · · + rk < 1−r for all
n
r
k ∈ N prove that if m > n then |xn − xm | ≤ 1−r . Deduce that (xn) is a
Cauchy sequence.

When the MCT is not applicable, the Cauchy criterion is another

possible tool to show convergence of a sequence.

Example 3.6.7. Consider the sequence (xn) defined by x1 = 1, x2 = 2,

and
xn = 21 (xn−2 + xn−1)
for n ≥ 2. One can show that (xn) is not monotone and therefore the
MCT is not applicable.

(a) Prove that 1 ≤ xn ≤ 2 for all n ∈ N.

1
(b) Prove that |xn − xn+1| = 2n−1 for all n ∈ N.

(c) Prove that if m > n then

1
|xn − xm| <
2n−2
Hint: Use part (b) and the Triangle inequality.

114
 3.6. CAUCHY SEQUENCES

(d) Deduce that (xn) is a Cauchy sequence and thus convergent.

(e) Show by induction that x2n+1 = 1 + 23 1 − 41n and deduce that
lim xn = 53 .

Notice that the main result used in the Cauchy Criterion is the
Bolzano–Weierstrass (B–W) theorem. We therefore have the following
chain of implications:

Completeness =⇒ MCT =⇒ B-W =⇒ Cauchy

A close inspection of the Cauchy Criterion reveals that it is really a

statement about the real numbers not having any gaps or holes. In
fact, the same can be said about the MCT and the Bolzano-Weierstrass
theorem. Regarding the Cauchy Criterion, if (xn) is a Cauchy sequence
then the terms of (xn) are clustering around a number and that number
must be in R if R has no holes. It is natural to ask then if we could
have used the Cauchy Criterion as our starting axiom (instead of the
Completeness Axiom) and then prove the Completeness property, and
then the MCT and the Bolzano-Weierstrass theorem. Unfortunately,
the Cauchy Criterion is not enough and we also need to take as an
axiom the Archimedean Property.

Theorem 3.6.8
Suppose that every Cauchy sequence in R converges to a number in
R and the Archimedean Property holds in R. Then R satisfies the
Completeness property, that is, every non-empty bounded above
subset of R has a least upper bound in R.

Proof. Let S ⊂ R be a non-empty set that is bounded above. If u is

an upper bound of S and u ∈ S then u is the least upper bound of S

115
 3.6. CAUCHY SEQUENCES

and since u ∈ R there is nothing to prove. Suppose then that no upper

bound of S is an element of S. Let a1 ∈ S be arbitrary and let b1 ∈ R
be an upper bound of S. Then a1 < b1 , and we set M = b1 − a1 > 0.
Consider the mid-point m1 = a1 +b 2
1
of the interval [a1 , b1]. If m1 is an
upper bound of S then set b2 = m1 and set a2 = a1 , otherwise set
b2 = b1 and a2 = m1 . In any case, we have |a2 − a1 | ≤ M2 , |b2 − b1| ≤ M2 ,
and |b2 − a2 | = M2 . Now consider the mid-point m2 = a2 +b 2
2
of the
interval [a2, b2]. If m2 is an upper bound of S then set b3 = m2 and
a3 = a2 , otherwise set b3 = b2 and a3 = m2 . In any case, we have
|a3 − a2 | ≤ M22
, |b3 − b2 | ≤ M
22
, and |b3 − a3 | = M
22
. By induction, there
exists a sequence (an ) such that an is not an upper bound of S and a
sequence (bn) such that bn is an upper bound of S, and |an −an+1 | ≤ M 2n ,
|bn − bn+1| ≤ M M
2n , and |bn − an | = 2n−1 . By Exercise 3.6.6, and using the
fact that limn→∞ rn = 0 if 0 < r < 1 (this is where the Archimedean
property is needed), it follows that (an ) and (bn ) are Cauchy sequences
and therefore by assumption both (an ) and (bn) are convergent. Since
|bn − an | = 2Mn−1 it follows that u = lim an = lim bn . We claim that u is

the least upper bound of S. First of all, for fixed x ∈ S we have that
x < bn for all n ∈ N and therefore x ≤ lim bn, that is, u is an upper
bound of S. Since an is not an upper bound of S, there exists xn ∈ S
such that an < xn < bn and therefore by the Squeeze theorem we have
u = lim xn . Given an arbitrary ε > 0 then there exists K ∈ N such
that u − ε < xK and thus u − ε is not an upper bound of S. This proves
that u is the least upper bound of S.

116
 3.6. CAUCHY SEQUENCES

Exercises

Exercise 3.6.1. Show that if (xn) is a Cauchy sequence then (xn) is

bounded. (Note: Do not use the fact that a Cauchy sequence converges
but show directly that if (xn) is Cauchy then (xn) is bounded.)

Exercise 3.6.2. Show that if (xn) converges then it is a Cauchy se-

quence.
n+1
Exercise 3.6.3. Show by definition that xn = n is a Cauchy se-
quence.
cos(n2 +1)
Exercise 3.6.4. Show by definition that xn = n is a Cauchy
sequence.

Exercise 3.6.5. Suppose that (xn) and (yn ) are sequences such that
|xm − xn| ≤ |ym − yn | for all n, m ∈ N. Show that if the sequence (yn)
is convergent then so is the sequence (xn).

Exercise 3.6.6. Suppose that 0 < r < 1. Show that if the sequence
(xn) satisfies |xn − xn−1| < rn−1 for all n ≥ 2 then (xn ) is a Cauchy
sequence and therefore convergent. Hint: If m > n then

|xm − xn| = |xm − xm−1 + xm−1 − xm−2 + xm−2 − · · · + xn+1 − xn |

Pk j 1−rk+1 1
Also, if 0 < r < 1 then j=0 r = 1−r < 1−r .

117
 3.7. INFINITE SERIES

3.7 Infinite Series

Informally speaking, a series is an infinite sum:

x1 + x2 + x3 + · · ·

Using summation notation:

∞
X
xn = x1 + x2 + x3 + · · ·
n=1
P∞
The series n=1 xn can be thought of as the sum of the sequence (xn) =
(x1, x2, x3, . . .). It is of course not possible to actually sum an infinite
number of terms and so we need a precise way to talk about what it
means for a series to have a finite value. Take for instance
X∞
( 32 )n−1 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 + · · ·
n=1

so that the sequence being summed is xn = ( 23 )n−1. Let’s compute the

first 10 terms of the sequence of partial sums {sn }∞
n=1 = (s1 , s2 , s3 , s4 , s5 , . . .)
defined as follows:

s1 = 1
s2 = 1 + ( 32 ) = 1.6666
s3 = 1 + ( 23 ) + ( 23 )2 = 2.1111
s4 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 = 2.4074
s5 = 1 + ( 23 ) + ( 23 )2 + ( 23 )3 + ( 32 )4 = 2.6049
s6 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 + ( 32 )4 + ( 23 )5 = 2.7366
s7 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 + ( 32 )4 + ( 23 )5 + ( 32 )6 = 2.8244
s8 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 + ( 32 )4 + ( 23 )5 + ( 32 )6 + ( 23 )7 = 2.8829
s9 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 + ( 32 )4 + ( 23 )5 + ( 32 )6 + ( 23 )7 + ( 32 )8 = 2.9219
s10 = 1 + ( 32 ) + ( 23 )2 + ( 23 )3 + ( 32 )4 + ( 23 )5 + ( 32 )6 + ( 23 )7 + ( 32 )8 + ( 23 )9 = 2.9479

118
 3.7. INFINITE SERIES

With the help of a computer, one can compute

20
X
s20 = ( 23 )k−1 = 2.999097813 . . .
k=1

50
X
s50 = ( 23 )k−1 = 2.999999994 . . .
k=1

100
X
s100 = ( 32 )k−1 = 2.999999998 . . . .
k=1

It seems as though the sequence (sn ) is converging to L = 3, that is,

limn→∞ sn = 3. It is then reasonable to say that the infinite series sums
or converges to

∞
X
( 32 )n−1 = 3 = lim sn .
n→∞
n=1

We now introduce some definitions to formalize our example.

119
 3.7. INFINITE SERIES

Definition 3.7.1
Let (xn) be a sequence. The infinite series generated by (xn) is
the sequence (sn) defined by

sn = x1 + x2 + · · · + xn

or recursively,

s1 = x1
sn+1 = sn + xn+1, n ≥ 1.

The sequence (sn ) is also called the sequence of partials sums

generated by (xn). The nth term of the sequence of partial sums
(sn ) can instead be written using summation notation:
n
X
sn = x1 + x2 + · · · + xn = xk .
k=1

Example 3.7.2. Let (xn) be the sequence xn = n3 . The first few terms
of the sequence of partials (sn) is

s1 = x1 = 3
3 9
s2 = x1 + x2 = 3 + 2 = 2
9 11
s3 = x1 + x2 + x3 = 2 +1= 2
11 3 25
s4 = s3 + x4 = 2 + 4 = 4

In both examples above, we make the following important observa-

tion: if (xn) is a sequence of non-negative terms then the sequence of
partials sums (sn ) is increasing. Indeed, if xn ≥ 0 then

sn+1 = sn + xn+1 ≥ sn

120
 3.7. INFINITE SERIES

and thus sn+1 ≥ sn .

Example 3.7.3. Find the sequence of partial sums generated by xn =

(−1)n.

Solution. We compute:

s1 = x1 = −1
s2 = x1 + x2 = −1 + 1 = 0
s3 = s2 + x3 = 0 − 1 = −1

Hence, (sn ) = (−1, 0, −1, 0, −1, 0, . . .).

The limit of the sequence (sn ), if it exists, makes precise what it

means for an infinite series
∞
X
xn = x1 + x2 + x3 + · · ·
n=1

to converge.

Definition 3.7.4: Convergence of Series

Let (xn ) be a sequence and let (sn ) be the sequence of partial sums
generated by (xn). If limn→∞ sn exists and equals L then we say
that the series generated by (xn) converges to L and we write
that ∞ n
X X
xn = L = lim sn = lim xk .
n→∞ n→∞
n=1 k=1

P∞
The notation n=1 xn is therefore a compact way of writing
n
X
lim xk ,
n→∞
k=1

121
 3.7. INFINITE SERIES

P
and the question of whether the series ∞ n=1 xn converges is really about
whether the limit limn→∞ sn exists. Often we will write a series such as
P∞ P
n=1 x n simply as xn when either the initial value of n is understood
or is unimportant. Sometimes, the initial n value may be n = 0, n = 2,
or some other n = n0 .

Example 3.7.5 (Geometric Series). The geometric series is perhaps

the most important series we will encounter. Let xn = rn where r ∈ R
is a constant. The generated series is
∞
X ∞
X
xn = rn = 1 + r + r2 + r3 + · · ·
n=0 n=0

and is called the geometric series. The nth term of the sequence of
partial sums is
sn = 1 + r + r2 + · · · + rn .
If r = 1 then limn→∞ sn does not exist (why?), so suppose that r 6= 1.
n+1
Using the fact that (1 − r)(1 + r + r2 + · · · + rn ) = 1−r
1−r we can write

1 − rn+1
sn = .
1−r
Now if |r| < 1 then limn→∞ rn = 0, while if |r| > 1 then limn→∞ rn does
not exist. Therefore, if |r| < 1 then

1 − rn+1 1
lim sn = lim = .
n→∞ n→∞ 1 − r 1−r
Therefore,
∞
X 1
rn = lim sn = .
n=0
n→∞ 1−r
P
In summary: The series ∞ n
n=0 r is called the geometric series and
1
converges if and only if |r| < 1 and in this case converges to 1−r .

122
 3.7. INFINITE SERIES

P (−1)n 2n
Example 3.7.6. Consider the series ∞ n=0 3n . The series can be
P∞ −2
n
written as n=0 3 and thus it is a geometric series with r = − 23 .
Since |r| = | − 23 | < 1, the series converges and it converges to
∞
X (−1)n2n 1 3
= =
n=0
3n 1 − (−2/3) 5

Example 3.7.7. Use the geometric series to show that

0.999999 . . . = 1.

Solution. We can write

0.999999 . . . = 0.9 + 0.09 + 0.009 + · · ·

9 9 9
= + + + ···
10 100 1000
9 9 9
= + 2 + 3 + ···
10 10 10
 
9 1 1
= 1+ + + ···
10 10 102
∞
9 X 1
=
10 n=0 10n
 
9 1
= 1
10 1 − 10

= 1.

123
 3.7. INFINITE SERIES

P∞ 1 1
Example 3.7.8 (Telescoping Series). Consider n=1 n(n+1) = 1·2 +
1 1
2·3 + 3·4 + · · · . Using partial fraction decomposition

1 1 1
= − .
n(n + 1) n n + 1
Therefore, the nth term of the sequence of partial sums (sn) is
n n  
X 1 X 1 1
sn = = −
k(k + 1) k k+1
k=1 k=1

This is a telescoping sum because all terms in the middle cancel and
only the first and last remain. For example:
1
s1 = 1 − 2
s2 = 1 − + 12 −
1
2
1
3 =1− 1
3
1
s3 = 1 − + 12 − + 13 −
2
1
3
1
4
= 1 − 41 .
1
By induction one can show that sn = 1− n+1 and therefore limn→∞ sn =
P
1. Therefore, the given series converges and it converges to ∞ 1
n=1 n(n+1) =
limn→∞ sn = 1.

P
Example 3.7.9 (Harmonic Series). Consider the series ∞ 1
n=1 n = 1 +
1 1 1
2 + 3 + 4 + · · · . We are going to analyze a subsequence (snk ) of the
sequence of partial sums (sn ). We will show that (snk ) is unbounded
and thus (sn ) is unbounded and therefore divergent. Consequently, the
P
series ∞ 1
n=1 n is divergent. Consider


s4 = 1 + 21 + 1
3 + 1
4 > 1 + 12 + 1
2 = 1 + 2 21 .

Now consider

s8 = s4 + 51 + 16 + 17 + 1
8 > 1 + 2 21 + 4 18 = 1 + 3 12 .

124
 3.7. INFINITE SERIES

Lastly consider,

s16 = s8 + 91 + 1
10 + ··· + 1
16 > 1 + 3 12 + 8 16
1
= 1 + 4 21 .

In general, one an show by induction that for k ≥ 2 we have

k
s2k > 1 + .
2
Therefore, the subsequence (s4, s8, s16, . . .) is unbounded and thus the
P
series ∞ 1
n=1 n is divergent.

We now present some basic theorems on the convergence of series;

most of them are a direct consequence of results from limit theorems
for sequences. The first theorem we present can be used to show that
a series diverges.

Theorem 3.7.10: Series Divergence Test

P
If xn converges then lim xn = 0. Equivalently, if lim xn 6= 0
P n→∞ n→∞
then xn diverges.

P
Proof. By definition, if xn converges then the sequence of partial
P
sums (sn ) is convergent. Suppose then that L = limn→∞ sn = xn .
Recall that (sn ) has the recursive definition sn+1 = sn + xn+1. The
sequences (sn+1) and (sn ) both converge to L and thus

lim xn+1 = lim (sn+1 − sn )

n→∞ n→∞
=L−L
= 0.

125
 3.7. INFINITE SERIES

P∞ 3n+1
Example 3.7.11. The series n=1 2n+5 is divergent because limn→∞ 3n+1
2n+5 =
3
2 > 1.

Example 3.7.12. The Series Divergence Test can only be used to

show that a series diverges. For example, consider the harmonic series
P∞ 1 1
P∞ 1
n=1 n . Clearly limn→∞ n = 0. However, we already know that n=1 n
is a divergent series. Hence, in general, the condition limn→∞ xn = 0 is
P
not sufficient to establish convergence of the series ∞ n=1 xn .

P
Example 3.7.13. A certain series ∞ k=1 xk has sequence of partial sums
2n
(sn ) whose general nth term is sn = n+1 .
P
(a) What is 10 k=1 xk ?
P
(b) Does the series ∞ k=1 xk converge? If yes, what does it converge
to? Explain.

(c) Does the sequence (xn) converge? If yes, what does it converge
to? Explain.

The next theorem is just an application of the Cauchy criterion

for convergence of sequences to the sequence of partial sums (sn ).

Theorem 3.7.14: Cauchy Criterion

P
The series xn converges if and only if for every ε > 0 there exists
K ∈ N such that if m > n ≥ K then

|sm − sn | = |xn+1 + xn+2 + · · · + xm | < ε.

The following theorem is very useful and is a direct application of

the Monotone convergence theorem.

126
 3.7. INFINITE SERIES

Theorem 3.7.15
Suppose that (xn) is a sequence of non-negative terms, that is, xn ≥
P
0 for all n ∈ N. Then xn converges if and only if (sn ) is bounded.
In this case,
X∞
xn = sup{sn | n ≥ 1}.
n=1

P
Proof. Clearly, if xn = limn→∞ sn exists then (sn ) is bounded. Now
suppose that (sn ) is bounded. Since xn ≥ 0 for all n ∈ N then sn+1 =
sn + xn+1 ≥ sn and thus sn+1 ≥ sn shows that (sn ) is an increasing
sequence. By the Monotone convergence theorem, (sn) converges and
limn→∞ sn = sup{sn | n ≥ 1}.
P
Example 3.7.16. Consider the series ∞ 1 1
n=1 n2 . Since n2 > 0, to prove
that the series converges it is enough to show that the sequence of
partial sums (sn ) is bounded. We will consider a subsequence of (sn ),
namely, (snk ) where nk = 2k − 1 for k ≥ 2. We have

1 1
s3 = 1 ++
22 32
1 1
<1+ 2 + 2
2 2
1
=1+
2
and
1 1 1 1
s7 = s3 + + + +
42 52 62 72
1 4
<1+ + 2
2 4
1 1
= 1 + + 2.
2 2

127
 3.7. INFINITE SERIES

By induction, one can show that

1 1 1
snk < 1 + + 2 + · · · + k−1
2 2 2

and therefore using the geometric series with r = 1/2 we have

1 1 1
snk < 1 + + 2 + · · · + k−1
2 2 2
∞
X 1
< n
= 2.
n=0
2

This shows that the subsequence (snk ) is bounded. In general, the

existence of a bounded subsequence does not imply that the original
sequence is bounded but if the original sequence is increasing then it
does. In this case, (sn) is indeed increasing, and thus since (snk ) is
a bounded subsequence then (sn ) is also bounded. Therefore, (sn)
P
converges, that is, ∞ 1
n=1 n2 is a convergent series.

Theorem 3.7.17
P P
Suppose that xn and yn are convergent series.
P P
(i) Then (xn + yn ) and (xn − yn ) are also convergent and
X X X
(xn ± yn ) = xn ± yn .

P P
(ii) For any constant c ∈ R, cxn is also convergent and cxn =
P
c xn .

128
 3.7. INFINITE SERIES

Proof. Using the limit laws:

n n n
!
X X X
lim (xk ± yk ) = lim xk ± yk
n→∞ n→∞
k=1 k=1 k=1

n
X n
X
= lim xk ± lim yk
n→∞ n→∞
k=1 k=1

∞
X ∞
X
= xn ± yn .
n=1 n=1

Hence,
∞
X ∞
X ∞
X
(xn ± yn ) = xn ± yn .
n=1 n=1 n=1

If c is a constant then by the Limit Laws,

n
X n
X ∞
X
lim cxk = c lim xk = c xn .
n→∞ n→∞
k=1 k=1 n=1

Therefore,
∞
X ∞
X
cxn = c xn .
n=1 n=1

Once establishing the convergence/divergence of some sequences,

we can use comparison tests to the determine convergence/divergence
properties of new sequences.

129
 3.7. INFINITE SERIES

Theorem 3.7.18: Comparison Test

Let (xn) and (yn ) be non-negative sequences and suppose that xn ≤
yn for all n ≥ 1.
P P
(i) If yn converges then xn converges.
P P
(ii) If xn diverges then yn diverges.

P P
Proof. Let tn = nk=1 xn and sn = nk=1 yk be the sequences of partial
P
sums. Since xn ≤ yn then tn ≤ sn . To prove (a), if yn converges then
(sn ) is bounded. Thus, (tn ) is also bounded. Since (tn ) is increasing and
P
bounded it is convergent by the MCT. To prove (b), if xn diverges
then (tn ) is necessarily unbounded and thus (sn ) is also unbounded and
therefore (sn ) is divergent.

Example 3.7.19. Let p ≥ 2 be an integer and let (dn ) be a sequence

of integers such that 0 ≤ dn ≤ p − 1. Use the comparison test to show
P
that the series ∞ dn
n=1 pn converges and converges to a point in [0, 1].

Example 3.7.20. Determine whether the given series converge.

P∞ n2 2
(a) First compute limn→∞ 3nn2 +n = 1/3. Therefore, by
n=1 3n2 +n :
the divergence test, the series diverges.

P∞ P
(b) 1
n=1 np where p ≥ 2: We know that ∞ 1
n=1 n2 is convergent. Now,
P
if p ≥ 2 then n1p ≤ n12 . Therefore by the Comparison test, ∞ 1
n=1 np
is also convergent. This is called a p-series and actually converges
for any p > 1.

130
 3.7. INFINITE SERIES

P∞ n+7
(c) n=1 n3 +3 : We will use the comparison test. We have

n+7 n + 7n
≤
n3 + 3 n3 + 3
8n
< 3
n
8
= 2.
n
P P∞ n+7
The series ∞ 8
n=1 n2 converges and thus the original series n=1 n3 +3
also converges.

P
Example 3.7.21. Suppose that 0 ≤ xn ≤ 1 for all n ∈ N. If xn
P
converges prove that (xn)2 also converges. Does the claim hold if we
only assume that xn > 0?
P
Solution. Since 0 ≤ xn ≤ 1 then 0 ≤ x2n ≤ xn. Since xn converges
P
then by the Comparison test then (xn)2 also converges. More gener-
P
ally, suppose that xn > 0 and xn converges. Then (xn) converges to
zero and thus there exists K ∈ N such that 0 < xn < 1 for all n ≥ K.
P P∞ 2
Then since the series ∞ x
n=1 n+K converges it follows that n=1 xn+K
P∞ 2
converges and consequently n=1 xn converges.

The following two tests can be used for series whose terms are not
necessarily non-negative.

Theorem 3.7.22: Absolute Convergence

P P
If the series |xn | converges then xn converges.

Proof. From −|xn| ≤ xn ≤ |xn | we obtain that

0 ≤ xn + |xn | ≤ 2|xn |.

131
 3.7. INFINITE SERIES

P P
If |xn| converges then so does 2|xn|. By the Comparison test
P
3.7.18, the series (xn + |xn |) converges also. Then the following series
is a difference of two converging series and therefore converges:
X X X
(xn + |xn |) − |xn | = xn

and the proof is complete.

P P
In the case that |xn| converges we say that xn converges
absolutely. We end the section with the Ratio test for series.

Theorem 3.7.23: Ratio Test for Series

P
Consider the series xn and let L = limn→∞ |x|xn+1
n|
|
. If L < 1 then
P
the series xn converges absolutely and if L > 1 then the series
diverges. If L = 1 or the limit does not exist then the test is
inconclusive.

Proof. Suppose that L < 1. Let ε > 0 be such that r = L + ε < 1.

There exists K ∈ N such that |x|xn+1 n|
|
< L + ε = r for all n ≥ K
and thus |xn+1| < |xn |r for all n ≥ K. By induction, it follows that
P∞
|xK+m| < |xK |rm for all m ≥ 1. Since m=1 |xk |r
m
is a geometric
P
series with r < 1 then, by the comparison test, the series ∞ m=1 |xK+m |
P
converges. Therefore, the series |xn | converges and by the Absolute
P
convergence criterion we conclude that xn converges absolutely. If
P
L > 1 then a similar argument shows that xn diverges. The case
L = 1 follows from the fact that some series converge and some diverge
when L = 1.

132
 3.7. INFINITE SERIES

Exercises
P
Exercise 3.7.1. Suppose that xn is a convergent series. Is it true
P P
that if yn is divergent then (xn + yn ) is divergent? If it is true,
prove it, otherwise give an example to show that it is not true.

Exercise 3.7.2. Suppose that xn ≥ 0 for all n ∈ N. Prove that

P∞ P∞ xn
if n=1 xn converges then n=1 n also converges. Is the converse
P∞ xn
true? That is, if n=1 n converges then does it necessarily follow
P∞
that n=1 xn converges?

Exercise 3.7.3. Using only the tests derived in this section, determine
whether the given series converge or diverge:
∞
X 2n2 + 3
(a) √
n=1
n2 + 3n + 2
∞
X cos(nπ)3n
(b)
n=1
2n
∞
X n−3
(c)
n=1
n3 + 1

Exercise 3.7.4. Suppose that (xn) and (yn ) are non-negative sequences.
P P P
Prove that if xn and yn are convergent then xnyn is convergent.
P
(Hint: Recall that if zn ≥ 0 then zn converges iff (sn ) is bounded,
Pn
where sn = k=1 zk is the sequence of partial sums. Alternatively, use
the identity (x + y)2 = x2 + 2xy + y 2 .)
P∞
Exercise 3.7.5. Suppose that xn > 0 for all n ∈ N and n=1 xn
P √
converges. You will prove that ∞n=1 xn xn+1 converges.
P∞
(a) Let yn = xn + xn+1. Prove that n=1 yn converges.

133
 3.7. INFINITE SERIES

√
(b) Using the fact that (a+b)2 = a2 +2ab+b2 , prove that ab ≤ a+b
if a, b > 0. Deduce that
√
xnxn+1 ≤ xn + xn+1 = yn .

P∞ √
(c) Deduce that n=1 xnxn+1 converges.

Exercise 3.7.6. Any number of the form x = 0.a1a2 a3 a4 . . . can be

written as
a1 a2 a3 a4
x= + 2 + 3 + 4 + ···
10 10 10 10
Using this fact, and a geometric series, prove that
23
0.25555555555555555 . . . . . . . . . = .
90
Exercise 3.7.7. Show that the following series converge and find their
sum.
∞
X (−1)n
(a)
n=0
e2n
∞
X 3
(b)
j=2
2j

∞
X 1
(c)
3k
k=−2

Exercise 3.7.8. Using the fact that 2k−1 < k! for all k ≥ 3, prove that
P
the series ∞ 1
n=0 n! converges.

Exercise 3.7.9. Let X : N → R be a decreasing sequence of non-

negative terms. Let (sn) be the sequence of partial sums of the series
P∞ k
n=1 X(n), let nk = 2 − 1 for k ∈ N, and consider the subsequence
(snk ).

134
 3.7. INFINITE SERIES

(a) Show by induction that

k−1
X
snk < 2n X(2n)
n=0

for all k ∈ N.
P∞ n
P∞
(b) Conclude that if n=0 2 X(2n ) converges then n=1 X(n).

(Note: This is a generalization of Example 3.7.16.)

135
3.7. INFINITE SERIES

136
 4

Limits of Functions

In this chapter, we study another notion of convergence that is surely

familiar to the reader, namely, the limit of a function at a given point.
After introducing the precise definition of the limit of a function, and
working through some examples, we will relate limits of functions with
limits of sequences resulting in the Sequential Criterion for Limits (The-
orem 4.1.11). In this chapter, when not explicitly stated, the letter A
will denote a subset of R.

4.1 Limits of Functions

Before we can give the definition of the limit of a function, we need the
notion of a cluster point of a set.

Definition 4.1.1: Cluster Point

A number c ∈ R is called a cluster point of A if for any given
δ > 0 there exists at least one point x ∈ A, with x 6= c, such that
|x − c| < δ.

Hence, c is a cluster point of A if there are points in A that are ar-

bitrarily close to c. In general, a cluster point of A is not necessarily

137
 4.1. LIMITS OF FUNCTIONS

an element of A. Naturally, cluster points can be characterized using

limits of sequences.

Lemma 4.1.2
A point c is a cluster point of A if and only if there exists a sequence
(xn) in A such that xn 6= c and lim xn = c.
n→∞

Proof. Let c be a cluster point of A and let δn = n1 for n ∈ N. Then

by definition of a cluster point, there exists xn ∈ A, xn 6= c, such that
|xn − c| < δn . Since δn → 0 then xn → c.
To prove the converse, suppose that xn → c, xn 6= c and xn ∈ A for
n ∈ N. Then by convergence of (xn) to c, for any δ > 0 there exists
K ∈ N such that |xK − c| < δ. Since x = xK ∈ A, this proves that c is
a cluster point of A.

Example 4.1.3. Below are some examples of cluster points for a given
set:

• Consider the set A = [0, 1]. Every point c ∈ A is a cluster point

of A.

• On the other hand, for A = (0, 1], the point c = 0 is a cluster

point of A but does not belong to A.

• For A = n1 | n ∈ N , the only cluster point of A is c = 0.

• A finite set does not have any cluster points.

• The set A = N has no cluster points.

• Consider the set A = Q ∩ [0, 1]. By the Density theorem, every

point c ∈ [0, 1] is a cluster point of A.

138
 4.1. LIMITS OF FUNCTIONS

We now give the definition of the limit of a function f : A → R at

a cluster point c of A.

Definition 4.1.4: Limit of a Function

Consider a function f : A → R and let c be a cluster point of A.
We say that f has a limit at c, or converges at c, if there exists
a number L ∈ R such that for any given ε > 0 there exists δ > 0
such that if x ∈ A and 0 < |x − c| < δ then |f (x) − L| < ε. In this
case, we write that
lim f (x) = L
x→c
and we say that f converges to L at c, or that f has limit L at
c. If f does not converge at c then we say that f diverges at c.

Another short-hand notation to denote that f converges to L at c is

f (x) → L as x → c.
By definition, if limx→c f (x) = L, then for any ε > 0 there exists a
δ > 0 such that for all x ∈ (c − δ, c + δ) ∩ A not equal to c it holds that
f (x) ∈ (L − ε, L + ε).

Theorem 4.1.5: Uniqueness of Limits

A function f : A → R can have at most one limit at c.

Proof. Suppose that f (x) → L and f (x) → L′ as x → c, and let ε > 0.

Then there exists δ > 0 such that |f (x)−L| < ε/2 and |f (x)−L′| < ε/2,
for all x ∈ A satisfying 0 < |x − c| < δ. Then if 0 < |x − c| < δ then

|L − L′ | ≤ |f (x) − L| + |f (x) − L′ |
< ε/2 + ε/2
= ε.

139
 4.1. LIMITS OF FUNCTIONS

Since ε > 0 is arbitrary, Theorem 2.2.7 implies that L = L′ .

Example 4.1.6. Consider the function f (x) = 5x + 3 with domain

A = R. Prove that

lim f (x) = 13.

x→2

Proof. We begin by analyzing the quantity |f (x) − 13|:

|f (x) − 13| = |5x + 3 − 13|

= |5x − 10|
= 5|x − 2|.

Hence, if 0 < |x − 2| < ε/5 then

|f (x) − 13| = |5x + 3 − 13|

= 5|x − 2|
< 5(ε/5)
= ε.

Thus, given ε > 0 we let δ = ε/5 and thus if 0 < |x − c| < δ then
|f (x) − 13| < ε. Thus, by definition, limx→2 f (x) = 13.

x+1
Example 4.1.7. Consider the function f (x) = x2 +3 with domain A =
R. Prove that
1
lim f (x) = .
x→1 2
140
 4.1. LIMITS OF FUNCTIONS

Proof. We have that

x+1 1
|f (x) − 12 | = −
x2 + 3 2

x2 − 2x + 1
=
2(x2 + 3)

|x − 1|2
=
2(x2 + 3)

< |x − 1|2.
√
Let ε > 0 be arbitrary and let δ = ε. Then if 0 < |x − 1| < δ then
|x − 1|2 < δ 2 = ε. Hence, if 0 < |x − 1| < δ then
x+1 1
|f (x) − 21 | = 2
−
x +3 2

< |x − 1|2
< ε.
Thus, by definition, limx→1 f (x) = 21 .

Example 4.1.8. Consider the function f (x) = x2 with domain A = R.

Prove that for any c ∈ R,
lim f (x) = c2 .
x→c
Proof. We first note that
|f (x) − c2 | = |x2 − c2 | = |x + c||x − c|.
By the triangle inequality, |x + c| ≤ |x| + |c| and therefore
|f (x) − c2 | = |x + c||x − c|

≤ (|x| + |c|)|x − c|.

141
 4.1. LIMITS OF FUNCTIONS

We now need to analyze how large |x| can become when x is say within
δ > 0 of c. To be concrete, suppose that δ = 1/2. Hence, if 0 < |x−c| <
δ then

|x| = |x − c + c|
≤ |x − c| + |c|
< δ + |c|
< 1 + |c|.

Therefore, if 0 < |x − c| < δ it holds that

|f (x) − c2 | ≤ (|x| + c)|x − c|

< (1 + c)|x − c|.

ε
Now suppose that ε > 0 is arbitrary and let δ = min{δ, 1+|c| }. Then if
0 < |x − c| < δ then |x| < 1 + |c| and therefore

|f (x) − c2 | = |x2 − c2 |

= |x + c||x − c|

≤ (|x| + |c|) · δ
ε
< (1 + |c|) ·
1 + |c|

= ε.

This proves, by definition, that limx→c x2 = c2 for any c ∈ R.

x2 −3x
Example 4.1.9. Consider the function f (x) = x+3
with domain A =
R\{−3}. Prove that
lim f (x) = 2.
x→6

142
 4.1. LIMITS OF FUNCTIONS

Proof. We first note that c = −3 is indeed a cluster point of A =

R\{−3}. Now,

x2 − 3x
|f (x) − 2| = −2
x+3

x2 − 5x − 6
=
x+3

(x + 1)(x − 6)
=
(x + 3)

|x + 1|
= |x − 6|.
|x + 3|

We now obtain a bound for |x+1|

|x+3|
when x is close to 6. Suppose then
that |x − 6| < 1. Then 5 < x < 7 and therefore, 6 < x + 1 < 8, which
implies that |x + 1| < 8. Similarly, if |x − 6| < 1 then 8 < x + 3 < 10
1
and therefore 8 < |x + 3|, which implies that |x+3| < 18 . Therefore, if
|x − 6| < 1 then
|x + 1| 1
< 8 · = 1.
|x + 3| 8
Suppose now that ε > 0 is arbitrary and let δ = min{1, ε}. If 0 <
|x − 6| < δ then from our analysis above it follows that |x+1|
|x+3| < 1.
Therefore, if 0 < |x − 6| < δ then

x2 − 3x
|f (x) − 2| = −2
x+3

|x + 1|
= |x − 6|
|x + 3|

<1·δ
≤ ε.

143
 4.1. LIMITS OF FUNCTIONS

This proves that limx→6 f (x) = 2.

Example 4.1.10. Consider the function f : R → R defined as

(
(x − 1) arctan(x), x ∈ Q
f (x) = 3(x−1)
1+x2 , x∈/ Q.

Prove that limx→1 f (x) = 0.

Proof. If x ∈ Q then

|f (x)| = |(x − 1) arctan(x)|

= |x − 1|| arctan(x)|
π
= |x − 1| · 2

and if x ∈
/ Q then
3(x − 1)
|f (x)| =
1 + x2
3|x − 1|
=
1 + x2
≤ 3|x − 1|.

Therefore, for all x ∈ R it holds that |f (x)| ≤ 3|x − 1| since π/2 < 3.
Thus, given ε > 0 let δ = ε/3 and thus if 0 < |x − 1| < δ then

|f (x)| ≤ 3|x − 1|
<3·δ
= 3 · ε/3
= ε.

This proves that limx→1 f (x) = 0.

The following important result states that limits of functions can

be studied using limits of sequences.

144
 4.1. LIMITS OF FUNCTIONS

Theorem 4.1.11: Sequential Criterion for Limits

Let f : A → R be a function and let c be a cluster point of A.
Then limx→c f (x) = L if and only if for every sequence (xn) in A
converging to c (with xn 6= c for all n ∈ N) the sequence (f (xn))
converges to L.

Proof. Suppose that limx→c f (x) = L. Let (xn) be a sequence in A

converging to c, with xn 6= c for all n ∈ N. We must prove that the
sequence (f (xn)) converges to L. To that end, let ε > 0 be arbitrary.
Then, by convergence of f to L at c, there exists δ > 0 such that if
0 < |x − c| < δ then |f (x) − L| < ε. Now, since (xn) → c, there exists
K ∈ N such that |xn − c| < δ for all n ≥ K. Therefore, for n ≥ K we
have that |f (xn) − L| < ε. This proves that limn→∞ f (xn) = L.
To prove the converse, we prove the contrapositive. Hence, we must
show that if f does not converge to L then there exists a sequence (xn)
in A (with xn 6= c) converging to c but the sequence (f (xn)) does not
converge to L. Assume then that f does not converge to L. Then,
negating the definition of the limit of a function, there exists ε > 0
such for all δ > 0 there exists x ∈ A such that 0 < |x − c| < δ and
|f (x) − L| ≥ ε. Then, let δn = n1 for n ∈ N. Then there exists xn 6= c
such that 0 < |xn − c| < δn and |f (xn) − L| ≥ ε. Since δn → 0 then
(xn) → c but clearly f (xn) does not converge to L. This ends the
proof.

The following theorem follows immediately from Theorem 4.1.11.

145
 4.1. LIMITS OF FUNCTIONS

Corollary 4.1.12
Let f : A → R be a function and let c be a cluster point of A and
let L ∈ R. Then f does not converge to L at c if and only if there
exists a sequence (xn) in A converging c, with xn 6= c for all n ∈ N,
and such that (f (xn)) does not converge to L.

Note that in Corollary 4.1.12, if the sequence (f (xn)) diverges then by

definition it does not converge to any L ∈ R and then f does not have
a limit at c. When applicable, the following corollary is a useful tool
to prove that a limit of a function does not exist.

Corollary 4.1.13
Let f : A → R be a function and let c be a cluster point of A.
Suppose that (xn) and (yn ) are sequences in A converging to c, with
xn 6= c and yn 6= c for all n ∈ N. If f (xn) and f (yn) converge but

lim f (xn) 6= lim f (yn)

n→∞ n→∞

then f does not have a limit at c.

Example 4.1.14. Prove that limx→0 x1 does not exist.

Proof. Consider xn = n1 , which clearly converges to c = 0 and xn 6= 0

for all n ∈ N. Then f (xn) = n which is unbounded and thus does not
converge. Thus, by Corollary 4.1.12, limx→0 x1 does not exist.


Example 4.1.15. Prove that limx→0 sin x1 does not exist.


Proof. Let f (x) = sin x1 with domain A = R\{0}. Consider the
1
sequence xn = π/2+nπ . It is clear that (xn) → 0 and xn 6= 0 for all
n ∈ N. Now (f (xn)) = (−1, 1, −1, −1, . . .) and therefore (f (xn)) does

146
 4.1. LIMITS OF FUNCTIONS

not converge. Therefore, f has no limit at c = 0. In fact, for each

1
α ∈ [0, 2π), consider the sequence xn = α+2nπ . Clearly (xn) → 0 and
xn 6= 0 for all n ∈ N. Now, f (xn) = sin(α + 2nπ) = sin(α). Hence,
(f (xn)) converges to sin(α). This shows that f oscillates within the
interval [−1, −1] as x approaches c = 0.

Example 4.1.16. The sign function, denoted by sgn : R → R, is

defined as (
1, x≥0
sgn(x) =
−1, x < 0
Prove that limx→0 sgn(x) does not exist.
n
Proof. Consider the sequence xn = (−1)n . Then (xn ) → 0 and xn 6= 0
for all n ∈ N. Now, yn = sgn(xn) = (−1)n and thus (yn) does not
converge. Therefore, by Corollary 4.1.12, the function sgn has no limit
at c = 0.

147
 4.1. LIMITS OF FUNCTIONS

Exercises

Exercise 4.1.1. Use the definition of the limit of a function to prove

that the following limits do indeed hold.
2x + 3
(a) lim =3
x→3 4x − 9

x2 − 3x
(b) lim =2
x→6 x + 3

(c) lim |x − 3| = 1
x→4

Exercise 4.1.2. Let A ⊂ R, let f : A → R, and suppose that c is a

cluster point of A. Suppose that there exists a constant K > 0 such
that |f (x) − L| ≤ K|x − c| for all x ∈ A. Prove that lim f (x) = L.
x→c

Exercise 4.1.3. Consider the function

(
x2 sin(1/x), x ∈ Q\{0}
f (x) = x2
1+x2
, x∈/ Q.

Prove that limx→0 f (x) = 0.

Exercise 4.1.4. Let f : R → R be defined as follows:

(
x, if x ∈ Q
f (x) =
−x, if x ∈ R\Q.

(a) Prove that f has a limit at c = 0.

(b) Now suppose that c 6= 0. Prove that f has no limit at c.

(c) Define g : R → R by g(x) = (f (x))2. Prove that g has a limit at

any c ∈ R.

148
 4.1. LIMITS OF FUNCTIONS

Hint: The Density Theorem will be helpful for (b). In particular, the
Density Theorem implies that for any point c ∈ R, there exists a se-
quence (xn) of rational numbers such that (xn) → c, and that there
exists a sequence (yn) of irrational numbers such that (yn ) → c.

Exercise 4.1.5. Use any applicable theorem to explain why the fol-
lowing limits do not exist.

(a) limx→0 x12

(b) limx→0 (x + sgn(x))

(c) limx→0 sin(1/x2)

Recall that the function sgn : R → R is defined as follows:

(
1, x≥0
sgn(x) =
−1, x < 0

149
 4.2. LIMIT THEOREMS

4.2 Limit Theorems

In this section, we establish basic limit theorems for limits of functions.
The reader should compare the results of this section with Section 3.2
where we established limit theorems for sequences. In fact, thanks to
the sequential criterion for limits of functions (Theorem 4.1.11), all of
the theorems in this section can be proved using limits of sequences.
To begin, we first show that if f has a limit at c then f satisfies a
local boundedness property at c. Let us first define then what it means
for a function to be locally bounded at a given point.

Definition 4.2.1: Local Boundedness

Consider a function f : A → R and let c be a cluster point of A.
We say that f is bounded locally at c if there exists δ > 0 and
M > 0 such that if x ∈ (c − δ, c + δ) ∩ A then |f (x)| ≤ M.

Theorem 4.2.2
Consider a function f : A → R and let c be a cluster point of A. If
lim f (x) exists then f is bounded locally at c.
x→c

Proof. Let L = limx→c f (x) and let ε > 0 be arbitrary. Then there
exists δ > 0 such that |f (x) − L| < ε for all x ∈ A such that 0 <
|x − c| < δ. Therefore, for all x ∈ A and 0 < |x − c| < δ we have that

|f (x)| = |f (x) − L + L|
≤ |f (x) − L| + |L|
< ε + |L|.

If c ∈ A then let M = max{|f (c)|, ε + |L|} and if c ∈

/ A then let

150
 4.2. LIMIT THEOREMS

M = ε + |L|. Then |f (x)| ≤ M for all x ∈ A such that 0 < |x − c| < δ,

that is, f is bounded locally at c.

Example 4.2.3. Consider the function f (x) = x1 defined on the set

A = (0, ∞). Clearly, c = 0 is a cluster point of A. For any δ > 0
and any M > 0 let x ∈ A be such that 0 < x < min{δ, M1 }. Then
0 < x < M1 , that is, M < x1 = f (x). Since M was arbitrary, this proves
that f is unbounded at c = 0 and consequently f does not have a limit
at c = 0.

We now state and prove some limit laws for functions. Let f, g :
A → R be functions and define the functions (f + g), (f − g), f g, and
f /g on A as follows:

(f ± g)(x) = f (x) ± g(x)

(f g)(x) = f (x)g(x)
 
f f (x)
(x) =
g g(x)
where for f /g we require that g(x) 6= 0 for all x ∈ A.

Theorem 4.2.4: Limit Laws

Let f, g : A → R be functions and let c be a cluster point of A.
Suppose that limx→c f (x) = L and limx→c g(x) = M. Then

(i) lim(f ± g)(x) = L ± M

x→c

(ii) lim(f g)(x) = LM

x→c
 
f L
(iii) lim (x) = , if M 6= 0
x→c g M

151
 4.2. LIMIT THEOREMS

The proofs are left as an exercises. (To prove the results, use the
sequential criterion for limits and the limits laws for sequences).

Corollary 4.2.5
Let f1, . . . , fk : A → R be functions and let c be a cluster point of
A. If limx→c fi (x) exists for each i = 1, 2, . . . , k then
k
X k
X
(i) lim fi(x) = lim fi(x)
x→c x→c
i=1 i=1

k
Y k
Y
(ii) lim fi(x) = lim fi (x)
x→c x→c
i=1 i=1

Example 4.2.6. If f (x) = a0 + a1 x + a2x2 + · · · + an xn is a polynomial

function then limx→c f (x) = f (c) for every c ∈ R. If g(x) = b0 + b1x +
b2x + · · · + bm xm is another polynomial function and g(x) 6= 0 in a
neighborhood of x = c and limx→c g(x) = g(c) 6= 0 then

f (x) f (c)
lim = .
x→c g(x) g(c)

x2 − 4
Example 4.2.7. Prove that lim = 4.
x→2 x − 2

Proof. We cannot use the Limit Laws directly since limx→2 (x − 2) = 0.

2
−4
Instead, notice that if x 6= 2 then xx−2 = x + 2. Hence, the func-
2
−4
tions f (x) = xx−2 and g(x) = x + 2 are equal at every point in
R\{0}. It is clear that limx→2 g(x) = 4 and therefore it follows that
also limx→2 f (x) = 4.

152
 4.2. LIMIT THEOREMS

Theorem 4.2.8
Let f : A → R be a function and let c be a cluster point of A.
Suppose that f has limit L at c. If f (x) ≥ 0 for all x ∈ A then
L ≥ 0.

Proof. We prove the contrapositive. Suppose then that L < 0. Let

ε > 0 be such that L + ε < 0. Then since limx→c f (x) = 0, there exists
δ > 0 such that if 0 < |x − c| < δ then f (x) < L + ε < 0. Hence,
f (x) < 0 for some x ∈ A.
We give another proof using the sequential criterion for limits. To
that end, if f converges to L at c then for any sequence (xn) converging
to c, xn 6= 0, we have that f (xn) → L. Now f (xn) ≥ 0 and therefore
L ≥ 0 from our results on limits of sequences (Theorem 3.2.7).

Theorem 4.2.9
Let f : A → R be a function and let c be a cluster point of A.
Suppose that M1 ≤ f (x) ≤ M2 for all x ∈ A and suppose that
limx→c f (x) = L. Then M1 ≤ L ≤ M2 .

Proof. We have that 0 ≤ f (x) − M1 and therefore by Theorem 4.2.8 we

have that 0 ≤ L − M1 . Similarly, from 0 ≤ M2 − f (x) we deduce that
0 ≤ M2 − L. From this we conclude that M1 ≤ L ≤ M2 . An alternative
proof: Since f → L at c, for any sequence (xn) → c with xn 6= 0, we
have that f (xn) → L. Clearly, M1 ≤ f (xn) ≤ M2 and therefore M1 ≤
L ≤ M2 from our results on limits of sequences (Theorem 3.2.7).

The following is the Squeeze Theorem for functions.

153
 4.2. LIMIT THEOREMS

Theorem 4.2.10: Squeeze Theorem

Let f, g, h : A → R be functions and let c be a cluster point of A.
Suppose that lim g(x) = L and lim h(x) = L. If g(x) ≤ f (x) ≤ h(x)
x→c x→c
for all x ∈ A, x 6= c, then lim f (x) = L.
x→c

Proof. Let (xn) be a sequence in A converging to c with xn 6= c for all

n ∈ N. Then, by the sequential criterion,

L = lim g(xn ) = lim h(xn ).

n→∞ n→∞

By assumption, it holds that g(xn) ≤ f (xn) ≤ h(xn ) for all n ∈ N,

and therefore by the Squeeze Theorem for sequences, we have that
limn→∞ f (xn) = L. This holds for every such sequence and therefore
limx→c f (x) = L.

Example 4.2.11. Let


2
x sin(1/x), x ∈ Q\{0}

f (x) = x2 cos(1/x), x ∈
/Q


0, x = 0.

Show that limx→0 f (x) = 0.

We end this section with the following theorem.

Theorem 4.2.12
Let f : A → R be a function and let c be a cluster point of A.
Suppose that limx→c f (x) = L. If L > 0 then there exists δ > 0
such that f (x) > 0 for all x ∈ (c − δ, c + δ), x 6= c.

Proof. Choose ε > 0 so that L − ε > 0, take for example ε = L/2.

Then there exists δ > 0 such that L − ε < f (x) < L + ε for all

154
 4.2. LIMIT THEOREMS

x ∈ (c−δ, c+δ), x 6= c, and thus by transitivity it follows that 0 < f (x)

for all x ∈ (c − δ, c + δ), x 6= c.

155
 4.2. LIMIT THEOREMS

Exercises

Exercise 4.2.1. Let f, g : A → R and suppose that c ∈ R is a cluster

point of A. Suppose that at c, f converges to L and g converges to M.
Prove that f g converges to LM at c in two ways: (1) using the definition
of the limit of a function, and (2) using the sequential criterion for
limits.

Exercise 4.2.2. Give an example of a set A ⊂ R, a cluster point c of

A, and two functions f, g : A → R such that lim f (x)g(x) exists but
x→c
lim f (x) does not exist.
x→c

Exercise 4.2.3. Give an example of a function f : R → R that is

bounded locally at c = 0 but does not have a limit at c = 0. Your
answer should not be in the form of a graph.

Exercise 4.2.4. Let f : R → R be a function that is bounded locally

at c and suppose that g : R → R converges to L = 0 at c. Prove that
lim f (x)g(x) = 0.
x→c

156
 5

Continuity

Throughout this chapter, A is a non-empty subset of R and f : A → R

is a function.

5.1 Continuous Functions

Definition 5.1.1: Continuity
The function f is continuous at c ∈ A if for any given ε > 0 there
exists δ > 0 such that if x ∈ A and |x−c| < δ then |f (x)−f (c)| < ε.
If f is not continuous at c then we say that f is discontinuous at
c. The function f is continuous on A if f is continuous at every
point in A.

Suppose that c ∈ A is a cluster point of A and f is continuous at

c. Then from the definition of continuity, lim f (x) exists and equal to
x→c
f (c). If c is not a cluster point of A then there exists δ > 0 such that
(c − δ, c + δ) ∩ A = {c} and continuity of f at c is immediate. In either
case, we see that f is continuous at c if and only if

lim f (x) = f (c)

x→c

157
 5.1. CONTINUOUS FUNCTIONS

The following is then immediate.

Theorem 5.1.2: Sequential Criterion for Continuity

The function f is continuous at c ∈ A if and only if for every
sequence (xn) in A converging to c, f (xn) converges to f (c):

lim f (xn) = f ( lim xn) = f (c).

n→∞ x→∞

Notice that here (xn) is allowed to take on the value c. The following
is immediate.

Theorem 5.1.3
The function f is discontinuous at c if and only if there exists a
sequence (xn) in A converging to c but f (xn) does not converge to
f (c).

Example 5.1.4. If f (x) = a0 + a1 x + a2x2 + · · · + an xn is a polynomial

function then limx→c f (x) = f (c) for every c ∈ R. Thus, f is continuous
everywhere. If g(x) = b0 + b1 x + b2 x + · · · + bm xm is another polynomial
function and g(c) 6= 0 then
f (x) f (c)
lim = .
x→c g(x) g(c)
Hence, h(x) = f (x)/g(x) is continuous at every c where g is non-zero.

Example 5.1.5. Determine the points of continuity of

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

Solution. Suppose that c 6= 0. Then limx→c f (x) = 1c = f (c). Hence,

f is continuous at c ∈ R\{0}. Consider now c = 0. The sequence

158
 5.1. CONTINUOUS FUNCTIONS

xn = n1 converges to c = 0 but f (xn) = n does not converge. Hence,

limx→0 f (x) does not exist. Thus, even though f (0) = 0 is well-defined,
f is discontinuous at c = 0.

Example 5.1.6 (Dirichlet Function). The following function was con-

sidered by Peter Dirichlet in 1829:
(
1, x ∈ Q
f (x) = (5.1)
0, x ∈ R\Q

Prove that f is discontinuous everywhere.

Proof. Let c be irrational and let ε = 1/2. Then for all δ > 0, there
exists x ∈ Q ∩ (c − δ, c + δ) (by the Density theorem) and therefore
|f (x) − f (c)| = 1 > ε. Hence, f is discontinuous at c. A similar
argument shows that f is discontinuous c ∈ Q. Alternatively, if c ∈ Q
then there exists a sequence of irrational numbers (xn) converging to c.
Now f (xn) = 0 and f (c) = 1, and this proves that f is discontinuous
at c. A similar arguments holds for c irrational.

Example 5.1.7 (Thomae Function). Let A = {x ∈ R : x > 0} and

define f : A → R as
(
0, x ∈ R\Q,
f (x) = 1 m
n , x = n ∈ Q, gcd(m, n) = 1, n ∈ N

The graph of f is shown in Figure 5.1. Prove that f is continuous

at every irrational number in A and is discontinuous at every rational
number in A.

Proof. Let c = m n ∈ Q with gcd(m, n) = 1. There exists a sequence

(xn) of irrational numbers in A converging to c. Hence, f (xn) = 0
while f (c) = n1 . This shows that f is discontinuous at c. Now let c be

159
 5.1. CONTINUOUS FUNCTIONS

0.5

0.4

0.3

0.2

-1 -0.75 -0.50 -0.25 0.25 0.50 0.75 1

Figure 5.1: Thomae’s function is continuous at each irrational and

discontinuous at each rational

irrational and let ε > 0 be arbitrary. Let N ∈ N be such that N1 < ε.

In the interval (c − 1, c + 1), there are only a finite number of rationals
m mk
n with n < N , otherwise we can create a sequence nk with nk < N , all
the rationals m mk
nk distinct and thus necessarily nk is unbounded. Hence,
k

there exists δ > 0 such that the interval (c − δ, c + δ) contains only

rational numbers x = m m
n with n > N . Hence, if x = n ∈ (c − δ, c + δ)
then f (x) = n1 < N1 and therefore |f (x) − f (c)| = n1 < N1 < ε. On
the other hand, if x ∈ (c − δ, c + δ) is irrational then |f (x) − f (c)| =
|0 − 0| < ε. This proves that f is continuous at c.

Suppose that f has a limit L at c but f is not defined at c. We can

extend the definition of f by defining
(
f (x), x 6= c
F (x) =
L, x = c.

Now, limx→c F (x) = limx→c f (x) = L = F (c), and thus F is continuous

at c. Hence, functions that are not defined at a particular point c but

160
 5.1. CONTINUOUS FUNCTIONS

have a limit at c can be extended to a function that is continuous at c.

Points of discontinuity of this type are called removal singularities.
On the other hand, the function f (x) = sin(1/x) is not defined at c = 0
and has not limit at c = 0, and therefore cannot be extended at c = 0
to a continuous function.

161
 5.1. CONTINUOUS FUNCTIONS

Exercises

Exercise 5.1.1. Let

(
x2 sin(1/x), x 6= 0
f (x) =
0, x=0

Prove that f is continuous at c = 0.

Exercise 5.1.2. Let

(
(1/x) sin(1/x2), x 6= 0
f (x) =
0, x=0

Prove that f is discontinuous at c = 0.

Exercise 5.1.3. This is an interesting exercise.

(a) Suppose that h : R → R is continuous on R and that h(r) = 0

for every rational number r ∈ Q. Prove that in fact h(x) = 0 for
all x ∈ R.

(b) Let f, g : R → R be continuous functions on R such that f (r) =

g(r) for every rational number r ∈ Q. Prove that in fact f (x) =
g(x) for all x ∈ R. Hint: Part (a) will be useful here.

Exercise 5.1.4. Suppose that f : R → R is a continuous function such

that f (p + q) = f (p) + f (q) for every p, q ∈ Q. Prove that in fact
f (x + y) = f (x) + f (y) for every x, y ∈ R.

162
 5.2. COMBINATIONS OF CONTINUOUS FUNCTIONS

5.2 Combinations of Continuous Functions

Not surprisingly, the set of continuous functions is closed under the
basic operation of arithmetic.

Theorem 5.2.1
Let f, g : A → R be continuous functions at c ∈ A and let b ∈ R.
Then

(i) f + g, f − g, f g, and bf are continuous at c.

(ii) If h : A → R is continuous at c ∈ A and h(x) 6= 0 for all x ∈ A

then fh is continuous at c.

Proof. Let ε > 0 be arbitrary. By continuity of f and g at c, there

exists δ1 > 0 such that |f (x) − f (c)| < ε/2 for all x ∈ A such that
0 < |x − c| < δ1, and there exists δ2 > 0 such that |g(x) − g(c)| < ε/2
for all x ∈ A such that 0 < |x − c| < δ2. Let δ = min{δ1, δ2}. Then for
x ∈ A such that 0 < |x − c| < δ we have that

|f (x) + g(x) − (f (c) + g(c))| ≤ |f (x) − f (c)| + |g(x) − g(c)|

< ε/2 + ε/2
= ε.

This proves that f + g is continuous at c. A similar proof holds for

f − g.
Consider now the function bf . If b = 0 then bf (x) = 0 for all x ∈ A
and continuity is trivial. So assume that b 6= 0. Let ε > 0 be arbitrary.
Then there exists δ > 0 such that if x ∈ A ∩ (c − δ, c + δ), x 6= c, then
|f (x) − f (c)| < ε/(|b|). Therefore, for x ∈ A ∩ (c − δ, c + δ), x 6= c, we

163
 5.2. COMBINATIONS OF CONTINUOUS FUNCTIONS

have that

|bf (x) − bf (c)| = |b||f (x) − f (c)|

< |b|ε/(|b|)
= ε.

We now prove continuity of f g. Let (xn) be any sequence in A

converging to c. Then yn = f (xn) converges to f (c) by continuity of f
at c, and zn = g(xn ) converges to g(c) by continuity of g at c. Hence the
sequence wn = yn zn converges to f (c)g(c). Hence, for every sequence
(xn) converging to c, f (xn)g(xn) converges to f (c)g(c). This shows that
f g is continuous at c.

Corollary 5.2.2
Let f, g : A → R be continuous functions on A and let b ∈ R. Then

(i) f + g, f − g, f g, and bf are continuous on A.

(ii) If h : A → R is continuous on A and h(x) 6= 0 for all x ∈ A

then fh is continuous on A.

Example 5.2.3. Prove that f (x) = x is continuous on R.

Proof. Let ε > 0 be arbitrary. Let δ = ε. If 0 < |x − c| < δ then

|f (x) − f (c)| = |x − c| < δ = ε.

Example 5.2.4. All polynomials p(x) = a0 + a1 x + · · · + an xn are

continuous everywhere.

Example 5.2.5. Rational functions f (x) = p(x)/q(x), with q(x) 6= 0

on A ⊂ R, are continuous on A.

164
 5.2. COMBINATIONS OF CONTINUOUS FUNCTIONS

Lemma 5.2.6: Continuity Under Shifting

If f : R → R is continuous then g(x) = f (x + α) is continuous,
where α ∈ R is arbitrary.

Proof. Let ε > 0 be arbitrary. Then there exists δ > 0 such that
|f (y) − f (d)| < ε for all 0 < |y − d| < δ. Therefore, if 0 < |x − c| =
|(x + α) − (c + α)| < δ then |f (x + α) − f (c + α)| < ε and therefore

|g(x) − g(c)| = |f (x + α) − f (c + α)| < ε.

To prove continuity of sin(x) and cos(x) we use the following facts.

For all x ∈ R, | sin(x)| ≤ |x| and | cos(x)| ≤ 1, and for all x, y, ∈ R
1

1


sin(x) − sin(y) = 2 sin 2
(x − y) cos 2
(x + y) .

Example 5.2.7. Prove that sin(x) and cos(x) are continuous every-
where.

Proof. We have that

| sin(x) − sin(c)| ≤ 2| sin( 12 (x − c))|| cos( 21 (x − c))|

1
≤ 2 |x − c|
2
= |x − c|.

Hence given ε > 0 we choose δ = ε. The proof that cos(x) is continuous

follows from the fact that cos(x) = sin(x + π/2) and Lemma 5.2.6.
sin(x)
Example 5.2.8. The functions tan(x) = cos(x) , cot(x) = cos(x)
sin(x)
, sec(x) =
1 1
cos(x) , and csc(x) = sin(x) are continuous on their domain.

165
 5.2. COMBINATIONS OF CONTINUOUS FUNCTIONS

√
Example 5.2.9. Prove that f (x) = x is continuous on A = {x ∈
R | x ≥ 0}.
√ √ √
Proof. For c = 0, we must consider | x − 0| = x. Given ε > 0 let
√
δ = ε2. Then if x ∈ A and x < δ = ε2 then x < ε. This shows that f
is continuous at c = 0. Now suppose that c 6= 0. Then
√ √
√ √ √ √ x+ c
| x − c| = | x − c| · √ √
x+ c

|x − c|
=√ √
x+ c

1
≤ √ |x − c|.
c
√ √ √
Hence, given ε > 0, suppose that 0 < |x − c| < cε. Then | x − c| <
ε.

Example 5.2.10. Prove that f (x) = |x| is continuous everywhere.

Proof. Follows from the inequality ||x| − |c|| ≤ |x − c|.

The last theorem of this section is concerned with the composition

of continuous functions.

Theorem 5.2.11: Continuity of Composite Functions

Let f : A → R and let g : B → R be continuous functions and
suppose that f (A) ⊂ B. Then the composite function (g ◦ f ) : A →
R is continuous.

Proof. Let ε > 0 be given. Let c ∈ A and let d = f (c) ∈ B. Then there
exists δ1 > 0 such that if 0 < |y − d| < δ1 then |g(y) − g(d)| < ε. Now
since f is continuous at c, there exists δ2 > 0 such that if 0 < |x−c| < δ2

166
 5.3. CONTINUITY ON CLOSED INTERVALS

then |f (x)−f (c)| < δ1. Therefore, if 0 < |x−c| < δ2 then |f (x)−d| < δ1
and therefore |g(f (x))−g(d)| < ε. This proves that (g ◦f ) is continuous
at c ∈ A. Since c is arbitrary, (g ◦ f ) is continuous on A.

Corollary 5.2.12
If f : A → R is continuous then g(x) = |f (x)| is continuous. If
p
f (x) ≥ 0 for all x ∈ A then h(x) = f (x) is continuous.

5.3 Continuity on Closed Intervals

In this section we develop properties of continuous functions on closed
intervals.

Definition 5.3.1
We say that f : A → R is bounded on A if there exists M > 0
such that |f (x)| ≤ M for all x ∈ A.

If f is not bounded on A then for any given M > 0 there exists x ∈ A

such that |f (x)| > M.

Example 5.3.2. Consider the function f (x) = x1 defined on the interval

A = (0, ∞). Let M > 0 be arbitrary. Then if 0 < x < M1 then
f (x) = x1 > M. For instance, take x = M1+1 . However, on the interval
[2, 3], f is bounded by M = 12 .

Theorem 5.3.3
Let f : A → R be a continuous function. If A = [a, b] is a closed
and bounded interval then f is bounded on A.

167
 5.3. CONTINUITY ON CLOSED INTERVALS

Proof. Suppose that f is unbounded. Then for each n ∈ N there exists

xn ∈ [a, b] such that |f (xn)| > n. Now the sequence (xn ) is bounded
since a ≤ xn ≤ b. By the Bolzano-Weierstrass theorem, (xn) has a
convergent subsequence, say (xnk ), whose limit u = limk→∞ xnk satisfies
a ≤ u ≤ b. Since f is continuous at u then limk→∞ f (xnk ) exists and
equal to f (u). This is a contradiction since |f (xnk )| > nk ≥ k implies
that f (xnk ) is unbounded.

Definition 5.3.4: Extrema of Functions

Let f : A → R be a function.

(i) The function f has an absolute maximum on A if there

exists x∗ ∈ A such that f (x) ≤ f (x∗) for all x ∈ A. We call
x∗ a maximum point and f (x∗) the maximum value of f
on A.

(ii) The function f has an absolute minimum on A if there

exists x∗ ∈ A such that f (x∗) ≤ f (x) for all x ∈ A. We call
x∗ a minimum point and f (x∗) the minimum value of f
on A.

Suppose that f : [a, b] → R is continuous. By Theorem 5.3.3, the

range of f , that is S = {f (x) | x ∈ [a, b]}, is bounded and therefore
inf(S) and sup(S) exist. In this case, we want to answer the question
as to whether inf(S) and sup(S) are elements of S. In other words, as
to whether f achieves its maximum and/or minimum value on [a, b].
That is, if there exists x∗, x∗ ∈ [a, b] such that f (x∗) ≤ f (x) ≤ f (x∗)
for all x ∈ [a, b].

1
Example 5.3.5. The function f (x) = x is continuous on A = (0, 1].

168
 5.3. CONTINUITY ON CLOSED INTERVALS

However, f is unbounded on A and never achieves a maximum value

on A.
Example 5.3.6. The function f (x) = x2 is continuous on [0, 2), is
bounded on [0, 2) but never reaches its maximum value on [0, 2), that
is, if S = {x2 : x ∈ [0, 2)} then sup(S) = 4 ∈
/ S.

Theorem 5.3.7: Extreme Value Theorem

Let f : A → R be a continuous function. If A = [a, b] is a closed
and bounded interval then f has a maximum and minimum point
on [a, b].

Proof. Let S = {f (x) | x ∈ [a, b]} be the range of f . By Theorem 5.3.3,

sup(S) exists; set M = sup(S). By the definition of the supremum,
for each ε > 0 there exists x ∈ [a, b] such that M − ε < f (x) ≤
M. In particular, for εn = 1/n, there exists xn ∈ [a, b] such that
M − εn < f (xn) ≤ M. Then limn→∞ f (xn) = M. The sequence (xn)
is bounded and thus has a convergent subsequence, say (xnk ). Let
x∗ = limk→∞ xnk . Clearly, a ≤ x∗ ≤ b. Since f is continuous at x∗, we
have that M = limk→∞ f (xnk ) = f (x∗). Hence, x∗ is a maximum point.
A similar proof establishes that f has a minimum point on [a, b].

By Theorem 5.3.7, we can replace sup{f (x) | x ∈ [a, b]} with

max{f (x) | x ∈ [a, b]}, and inf{f (x) | x ∈ [a, b]} with min{f (x) | x ∈
[a, b]}. When the interval [a, b] is clear from the context, we will simply
write max(f ) and min(f ). The following example shows the importance
of continuity in achieving a maximum/minimum.
Example 5.3.8. The function f : [−1, 1] → R defined by
(
3 − x2 , 0 < x ≤ 1
f (x) =
x2 , −1 ≤ x ≤ 0

169
 5.3. CONTINUITY ON CLOSED INTERVALS

does not achieve a maximum value on the closed interval [−1, 1].

The next theorem, called the Intermediate Value Theorem, is the

main result of this section, and one of the most important results in
this course.

Theorem 5.3.9: Intermediate Value Theorem

Consider the function f : [a, b] → R and suppose that f (a) < f (b).
If f is continuous then for any ξ ∈ R such that f (a) < ξ < f (b)
there exists c ∈ (a, b) such that f (c) = ξ.

Proof. Let S = {x ∈ [a, b] : f (x) < ξ}. The set S is non-empty

because a ∈ S. Moreover, S is clearly bounded above. Let c = sup(S).
By the definition of the supremum, there exists a sequence (xn) in S
such that limn→∞ xn = c. Since a ≤ xn ≤ b it follows that a ≤ c ≤ b.
By definition of xn , f (xn) < ξ and since f is continuous at c we have
that f (c) = limn→∞ f (xn) ≤ ξ, and thus f (c) ≤ ξ. Now let δn > 0 be
such that δn → 0 and c + δn < b. Then zn = c + δn converges to c and
ξ ≤ f (zn) because zn 6∈ S. Therefore, since limn→∞ f (zn ) = f (c) we
have that ξ ≤ f (c). Therefore, f (c) ≤ ξ ≤ f (c) and this proves that
ξ = f (c). This shows also that a < c < b.

The Intermediate Value Theorem has applications in finding points

where a function is zero.

Corollary 5.3.10
Let f : [a, b] → R be a function and suppose that f (a)f (b) < 0. If
f is continuous then there exists c ∈ (a, b) such that f (c) = 0.

Example 5.3.11. A hiker begins his climb at 7:00 am on a marked

trail and arrives at the summit at 7:00 pm. The next day, the hiker

170
 5.3. CONTINUITY ON CLOSED INTERVALS

begins his trek down the mountain at 7:00 am, takes the same trail
down as he did going up, and arrives at the base at 7:00 pm. Use the
Intermediate Value Theorem to show that there is a point on the path
that the hiker crossed at exactly the same time of day on both days.
Proof. Let f (t) be the distance traveled along the trail on the way up
the mountain after t units of time, and let g(t) be the distance remaining
to travel along the trail on the way down the mountain after t units of
time. Both f and g are defined on the same time interval, say [0, 12] if
t is measured in hours. If M is the length of the trail, then f (0) = 0,
f (12) = M, g(0) = M and g(12) = 0. Let h(t) = g(t) − f (t). Then
h(0) = M and h(12) = −M. Hence, there exists t∗ ∈ (0, 12) such that
h(t∗ ) = 0. In other words, f (t∗) = g(t∗), and therefore t∗ is the time of
day when the hiker is at exactly the same point on the trail.
Example 5.3.12. Prove by the Intermediate Value Theorem that f (x) =
xex − 2 has a root in the interval [0, 1].
Proof. The function f is continuous on [0, 1]. We have that f (0) =
−2 < 0 and f (1) = e − 2 > 0. Therefore, there exists x∗ ∈ (0, 1) such
that f (x∗) = 0, i.e., f has a zero in the interval (0, 1).

The next results says, roughly, that continuous functions preserve

closed and bounded intervals. In the following theorem, we use the
short-hand notation f ([a, b]) = {f (x) | x ∈ [a, b]} for the range of f
under [a, b].

Theorem 5.3.13
If f : [a, b] → R is continuous then f ([a, b]) = [min(f ), max(f )].

Proof. Since f achieves its maximum and minimum value on [a, b], there
exists x∗, x∗ ∈ [a, b] such that f (x∗) ≤ f (x) ≤ f (x∗) for all x ∈ [a, b].

171
 5.3. CONTINUITY ON CLOSED INTERVALS

Hence, f ([a, b]) ⊂ [f (x∗), f (x∗)]. Assume for simplicity that x∗ < x∗.
Then [x∗, x∗] ⊂ [a, b]. Let ξ ∈ R be such that f (x∗) < ξ < f (x∗). Then
by the Intermediate Value Theorem, there exists c ∈ (x∗, x∗) such that
ξ = f (c). Hence, ξ ∈ f ([a, b]), and this shows that [f (x∗), f (x∗)] ⊂
f ([a, b]). Therefore, f ([a, b]) = [f (x∗), f (x∗)] = [min(f ), max(f )].

It is worth noting that the previous theorem does not say that
f ([a, b]) = [f (a), f (b)].

172
 5.3. CONTINUITY ON CLOSED INTERVALS

Exercises

Exercise 5.3.1. Let f : A → R be any function. Show that if −f

achieves its maximum at x0 ∈ A then f achieves its minimum at x0.

Exercise 5.3.2. Let f and g be continuous functions on [a, b]. Suppose

that f (a) ≥ g(a) and f (b) ≤ g(b). Prove that f (x0) = g(x0) for at least
one x0 in [a, b].

Exercise 5.3.3. Let f : [0, 1] → R be a continuous function and sup-

pose that f (x) ∈ [0, 1] for all x ∈ [0, 1]. Show that there exists x0 ∈ [0, 1]
such that f (x0) = x0. Hint: Consider the function g(x) = f (x) − x on
the interval [0, 1].

Exercise 5.3.4. Let f : [a, b] → R be a continuous function. Prove that

if f (x) ∈ Q for all x ∈ [a, b] then f is a constant function. Hint: You
will need the Density Theorem and the Intermediate Value Theorem.

173
 5.4. UNIFORM CONTINUITY

5.4 Uniform Continuity

In the definition of continuity of f : A → R at c ∈ A, the δ will in
general not only depend on ε but also on c. In other words, given two
points c1 , c2 and fixed ε > 0, the minimum δ1 and δ2 needed for c1
and c2 in the definition of continuity are generally going to be different.
To see this, consider the continuous function f (x) = x2. Then it is
straightforward to verify that if c2 − ε > 0 then |x2 − c2 | < ε if and only
if
p p
c2 − ε − c < x − c < c2 + ε − c.

Let c1 = 1 and c2 = 3, and let ε = 1/2. Then |x − 1| < δ1 implies

√
that |f (x) − f (1)| < ε if and only δ1 ≤ 1 + ε − 1 ≈ 0.2247. On the
other hand, |x − c| < δ2 implies that |f (x) − f (3)| < ε if and only if
√
δ2 ≤ 9 + ε − 9 ≈ 0.0822. The reason that a smaller delta is needed
at c = 3 is that the slope of f at c = 3 is larger than that at c = 1.
On the other hand, consider the function f (x) = sin(2x). For any c it
holds that

|f (x) − f (c)| = | sin(2x) − sin(2c)|

≤ 2|x − c|.

Hence, given any ε > 0 we can set δ = ε/2 and then |x − c| < δ implies
that |f (x) − f (c)| < ε. The punchline is that δ = ε/2 will work for any
c. These motivating examples lead to the following definition.

Definition 5.4.1: Uniform Continuity

The function f : A → R is said to be uniformly continuous on
A if for each ε > 0 there exists δ > 0 such that for all x, u ∈ A
satisfying |x − u| < δ it holds that |f (x) − f (u)| < ε.

174
 5.4. UNIFORM CONTINUITY

Example 5.4.2. Let k 6= 0 be any non-zero constant. Show that

f (x) = kx is uniformly continuous on R.

Proof. We have that |f (x)−f (c)| = |kx−kc| = |k||x−c|. Hence, for any
ε > 0 we let δ = ε/|k|, and thus if |x−c| < δ then |f (x)−f (c)| < ε.

Example 5.4.3. Prove that f (x) = sin(x) is uniformly continuous.

Proof. We have that

| sin(x) − sin(c)| ≤ 2| sin( 12 (x − c))|

≤ |x − c|.

Hence, for ε > 0 let δ = ε and if |x−c| < δ then | sin(x)−sin(c)| < ε.
1
Example 5.4.4. Show that f (x) = is uniformly continuous on
1 + x2
R.

Proof. We have that

1 1
|f (x) − f (c)| = −
1 + x2 1 + c2

1 + c2 − 1 − x2
=
(1 + x2 )(1 + c2 )

x+c
= |x − c|
(1 + x2 )(1 + c2 )

x c
= + |x − c|.
(1 + x2 )(1 + c2 ) (1 + x2)(1 + c2 )
|x|
Now, |x| ≤ 1 + x2 implies that 1+x2
≤ 1 and therefore

|x| 1
≤ ≤ 1.
(1 + x2)(1 + c2 ) 1 + c2
175
 5.4. UNIFORM CONTINUITY

It follows that |f (x) − f (c)| ≤ 2|x − c|. Hence, given ε > 0 we let
δ = ε/2, and if |x − c| < δ then |f (x) − f (c)| < ε.

The following is a simple consequence of the definition of uniform

continuity.

Theorem 5.4.5
Let f : A → R be a function. The following are equivalent:

(i) The function f is not uniformly continuous on A.

(ii) There exists ε0 > 0 such that for every δ > 0 there exists
x, u ∈ A such that |x − u| < δ but |f (x) − f (c)| ≥ ε0.

(iii) There exists ε0 > 0 and two sequences (xn) and (un) such that
lim (xn − un) = 0 and |f (xn) − f (un)| ≥ ε0.
n→∞

Example 5.4.6. Let f (x) = x1 and let A = (0, ∞). Let xn = n1 and let
1
un = n+1 . Then lim(xn − un ) = 0. Now |f (xn) − f (un)| = | − 1| = 1.
Hence, if ε = 1/2 then |f (xn) − f (un)| > ε. This proves that f is not
uniformly continuous on A = (0, ∞).

Theorem 5.4.7: Uniform Continuity on Intervals

Let f : A → R be a continuous function with domain A = [a, b]. If
f is continuous on A then f is uniformly continuous on A.

Proof. We prove the contrapositive, that is, we will prove that if f is

not uniformly continuous on [a, b] then f is not continuous on [a, b].
Suppose then that f is not uniformly continuous on [a, b]. Then there

176
 5.4. UNIFORM CONTINUITY

exists ε > 0 such that for δn = 1/n, there exists xn, un ∈ [a, b] such
that |xn − un| < δn but |f (xn) − f (un)| ≥ ε. Clearly, lim(xn − un ) = 0.
Now, since a ≤ xn ≤ b, by the Bolzano-Weierstrass theorem there is a
subsequence (xnk ) of (xn) that converges to a point z ∈ [a, b]. Now,

|unk − z| ≤ |unk − xnk | + |xnk − z|

and therefore also limk→∞ unk = z. Hence, both (xnk ) and (unk ) are
sequences in [a, b] converging to z but |f (xnk ) − f (unk )| ≥ ε. Hence
f (xnk ) and f (unk ) do not converge to the same limit and thus f is not
continuous at z. This completes the proof.

The following example shows that boundedness of a function does

not imply uniform continuity.

Example 5.4.8. Show that f (x) = sin(x2) is not uniformly continuous

on R.
√ √ √
Proof. Consider xn = πn and un = πn+ 4√πn . Clearly lim(xn −un) =
0. On the other hand
√ √
|f (xn) − f (un)| = sin(πn) − sin( πn + 4√πn )2

π π


= sin πn + 2
+ 16n

π


= cos πn + 16n

π π
= cos(nπ) cos( 16n ) − sin(nπ) sin( 16n )

= |(−1)n cos( 16n

π
)|

π
= cos( 16n )

π
≥ cos( 16 )

177
 5.4. UNIFORM CONTINUITY

The reason that f (x) = sin(x2) is not uniformly continuous is that

f is increasing rapidly on arbitrarily small intervals. Explicitly, it does
not satisfy the following property.

Definition 5.4.9
A function f : A → R is called a Lipschitz function on A if there
exists a constant K > 0 such that |f (x) − f (u)| ≤ K|x − u| for all
x, u ∈ A.

Suppose that f is Lipschitz with Lipschitz constant K > 0. Then

for all x, u ∈ A we have that
f (x) − f (u)
≤ K.
x−u
Hence, the secant line through the points (x, f (x)) and (u, f (u)) has
slope no larger than K in magnitude. Hence, a Lipschitz function has
a constraint on how quickly it can change (measured by |f (x) − f (u)|)
relative to the change in its inputs (measured by |x − u|).

Example 5.4.10. Since | sin(x) − sin(u)| ≤ |x − u|, f (x) = sin(x) is a

Lipschitz function on R with constant K = 1.

Example 5.4.11. If a 6= 0, the function f (x) = ax + b is Lipschitz

on R with constant K = |a|. When a = 0, f is clearly Lipschitz with
arbitrary constant K > 0.

Theorem 5.4.12: Lipschitz and Uniform Continuity

If f : A → R is a Lipschitz function on A then f is uniformly
continuous on A.

178
 5.4. UNIFORM CONTINUITY

Proof. By assumption, |f (x) − f (u)| ≤ K|x − u| for all x, u ∈ A for

some constant K > 0. Let ε > 0 be arbitrary and let δ = ε/K. Then
if |x − u| < δ then |f (x) − f (u)| ≤ K|x − u| < ε. Hence, f is uniformly
continuous on A.

The following example shows that a uniformly continuous function

is not necessarily Lipschitzian.
√
Example 5.4.13. Consider the function f (x) = x defined on A =
[0, 2]. Since f is continuous, f is uniformly continuous on [0, 2]. To show
that f is not Lipschitzian on A, let u = 0 and consider the inequality
√ √
|f (x)| = x ≤ K|x| for some K > 0. If x ∈ [0, 2] then x ≤ K|x| if
and only if x ≤ K 2x2 if and only if x(K 2x−1) ≥ 0. If x ∈ (0, 1/K 2)∩A,
it holds that K 2x − 1 < 0, and thus no such K can exist. Thus, f is
not Lipschitzian on [0, 2].

179
 5.4. UNIFORM CONTINUITY

Exercises

Exercise 5.4.1. Let f : A → R and let g : A → R be uniformly

continuous functions on A. Prove that f + g is uniformly continuous
on A

Exercise 5.4.2. Let f : A → R and let g : A → R be Lipschitz

functions on A. Prove that f + g is a Lipschitz function on A.

Exercise 5.4.3. Give an example of a function that is uniformly con-

tinuous on R but is not bounded on R.

Exercise 5.4.4. Prove that f (x) = x2 is not uniformly continuous on

R.

Exercise 5.4.5. Give an example of distinct functions f : R → R and

g : R → R that are uniformly continuous on R but f g is not uniformly
continuous on R. Prove that your resulting function f g is indeed not
uniformly continuous.

Exercise 5.4.6. A function f : R → R is said to be T -periodic on

R if there exists a number T > 0 such that f (x + T ) = f (x) for all
x ∈ R. Prove that a T -periodic continuous function on R is bounded
and uniformly continuous on R. Hint: First consider f on the interval
[0, T ].

180
 6

Differentiation

6.1 The Derivative

We begin with the definition of the derivative of a function.

Definition 6.1.1: The Derivative

Let I ⊂ R be an interval and let c ∈ I. We say that f : I → R is
differentiable at c or has a derivative at c if
f (x) − f (c)
lim
x→c x−c
exists. We say that f is differentiable on I if f is differentiable at
every point in I.

By definition, f has a derivative at c if there exists a number L ∈ R

such that for every ε > 0 there exists δ > 0 such that if |x − c| < δ then

f (x) − f (c)
− L < ε.
x−c
181
 6.1. THE DERIVATIVE

f (x) − f (c)
If f is differentiable at c, we will denote lim by f ′ (c), that
x→c x−c
is,
f (x) − f (c)
f ′ (c) = lim .
x→c x−c
The rule that sends c to the number f ′ (c) defines a function on a pos-
sibly smaller subset J ⊂ I. The function f ′ : J → R is called the
derivative of f .

Example 6.1.2. Let f (x) = 1/x for x ∈ (0, ∞). Prove that f ′ (x) =
− x12 .

Example 6.1.3. Let f (x) = sin(x) for x ∈ R. Prove that f ′ (x) =

cos(x).

Proof. Recall that

x−c

x+c


sin(x) − sin(c) = 2 sin 2 cos 2

and that limx→0 sin(x)

x = 1. Therefore,
x−c

x+c


sin(x) − sin(c) 2 sin 2
cos 2
lim = lim
x→c x−c x→c x−c
x−c

!
sin 2 x+c


= lim x−c cos 2
x→c
2

= 1 · cos(c) = cos(c).

Hence f ′(c) = cos(c) for all c and thus f ′ (x) = cos(x).

x
Example 6.1.4. Prove by definition that f (x) = 1+x2
is differentiable
on R.

182
 6.1. THE DERIVATIVE

Proof. We have that

x c
f (x) − f (c) 1+x2 − 1+c 2
=
x−c x−c

x(1 + c2 ) − c(1 + x2 )
=
(1 + x2)(1 + c2 )(x − c)

1 − cx
= .
(1 + c2 )(1 + x2)
Now
f (x) − f (c) 1 − c2
lim = .
x→c x−c (1 + c2 )2
Hence, f ′ (c) exists for all c ∈ R and the derivative function of f is

1 − x2
f ′(x) = .
(1 + x2 )2

Example 6.1.5. Prove that f ′(x) = α if f (x) = αx + b.

Proof. We have that f (x) − f (c) = αx − αc = α(x − c). Therefore,

limx→c f (x)−f
x−c
(c)
= α. This proves that f ′(x) = α for all x ∈ R.

Example 6.1.6. Compute the derivative function of f (x) = |x| for

x ∈ R.

Solution. If x > 0 then f (x) = x and thus f ′ (x) = 1 for x > 0. If x < 0
then f (x) = −x and therefore f ′(x) = −1 for x < 0. Now consider
c = 0. We have that
f (x) − f (c) |x|
= .
x−c x
We claim that the limit limx→0 |x|
x
does not exist and thus f ′ (0) does not
exist. To see this, consider xn = 1/n. Then (xn) → 0 and f (xn) = 1

183
 6.1. THE DERIVATIVE

for all n. On the other hand, consider yn = −1/n. Then (yn ) → 0 and
f (yn) = −1. Hence, limn→∞ f (xn) 6= limn→∞ f (yn), and thus the claim
holds by the Sequential criterion for limits. The derivative function f ′
of f is therefore defined on A = R\{0} and is given by
(
1, x>0
f ′(x) =
−1, x > 0.

Hence, even though f is continuous at every point in its domain R,

it is not differentiable at every point in its domain. In other words,
continuity is not a sufficient condition for differentiability.

Theorem 6.1.7: Differentiability Implies Continuity

Suppose that f : I → R is differentiable at c. Then f is continuous
at c.

Proof. To prove that f is continuous at c we must show that limx→c f (x) =

f (c). By assumption limx→c f (x)−f
x−c
(c)
= f ′(c) exists, and clearly limx→c (x−
c) = 0. Hence we can apply the Limits laws and compute

lim f (x) = lim(f (x) − f (c) + f (c))

x→c x→c
 
f (x) − f (c)
= lim (x − c) + f (c)
x→c (x − c)
= f ′(c) · 0 + f (c)
= f (c)

and the proof is complete.

184
 6.1. THE DERIVATIVE

Theorem 6.1.8: Combinations of Differentiable Functions

Let f : I → R and g : I → R be differentiable at c ∈ I. The
following hold:

(i) If α ∈ R then (αf ) is differentiable and (αf )′(c) = αf ′ (c).

(ii) (f + g) is differentiable at c and (f + g)′ (c) = f ′(c) + g ′ (c).

(iii) f g is differentiable at c and (f g)′(c) = f ′ (c)g(c) + f (c)g ′(c).

(iv) If g(c) 6= 0 then (f /g) is differentiable at c and

 ′
f f ′ (c)g(c) − f (c)g ′(c)
(c) =
g g(c)2

Proof. Parts (i) and (ii) are straightforward. We will prove only (iii)
and (iv). For (iii), we have that

f (x)g(x) − f (c)g(c) f (x)g(x) − f (c)g(x) + f (c)g(x) − f (c)g(c)

=
x−c x−c
f (x) − f (c) g(x) − g(c)
= g(x) + f (c) .
x−c x−c

Now limx→c g(x) = g(c) because g is differentiable at c. Therefore,

f (x)g(x) − f (c)g(c) f (x) − f (c) g(x) − g(c)

lim = lim g(x) + lim f (c)
x→c x−c x→c x−c x→c x−c

= f ′(c)g(c) + f (c)g ′ (c).

To prove part (iv), since g(c) 6= 0, then there exist a δ-neighborhood

185
 6.1. THE DERIVATIVE

J = (c − δ, c + δ) such that g(x) 6= 0 for all x ∈ J. If x ∈ J then

f (x) f (c)
g(x)
− g(x) f (x)g(c) − g(x)f (c)
=
x−c g(x)g(c)(x − c)

f (x)g(c) − f (c)g(c) + f (c)g(c) − g(x)f (c)

=
g(x)g(c)(x − c)
f (x)g(c)−f (c)g(c)
x−c − f (c)g(x)−f
x−c
(c)g(c)
=
g(x)g(c)

Since g(c) 6= 0, it follows that

f (x) f (c)
g(x) − g(x) f ′ (c)g(c) − f (c)g ′(c)
lim =
x→c x−c g(c)2

and the proof is complete.

We now prove the Chain Rule.

Theorem 6.1.9: Chain Rule

Let f : I → R and g : J → R be functions such that f (I) ⊂ J and
let c ∈ I. If f ′ (c) exists and g ′ (f (c)) exists then (g ◦ f )′(c) exists
and (g ◦ f )′(c) = g ′ (f (c))f ′(c).

Proof. Suppose that there exists a neighborhood of c where f (x) 6=

f (c). Otherwise, the composite function (g ◦ f )(x) is constant in a
neighborhood of c, and then clearly differentiable at c. Consider the
function h : J → R defined by
(
g(y)−g(f (c))
y−f (c) , y 6= f (c)
h(y) =
g ′ (f (c)), y = f (c).

186
 6.1. THE DERIVATIVE

Now
g(y) − g(f (c))
lim h(y) = lim
y→f (c) y→f (c) y−c

= g ′ (f (c))′

= h(f (c)).

Hence, h is differentiable at f (c) and therefore h is at f (c). Now,

g(f (x)) − g(f (c)) f (x) − f (c)
= h(f (x))
x−c x−c
and therefore
g(f (x)) − g(f (c)) f (x) − f (c)
lim = lim h(f (x))
x→c x−c x→c x−c
′
= h(f (c))f (c)
= g ′ (f (c))f ′(c).

Therefore, (g ◦ f )′(c) = g ′ (f (c))f ′(c) as claimed.

Example 6.1.10. Compute f ′(x) if

(
x2 sin( x1 ), x 6= 0
f (x) =
0, x = 0.

Where is f ′(x) continuous?

Solution. When x 6= 0, f (x) is the composition and product of differ-

entiable functions at x, and therefore f is differentiable at x 6= 0. For
instance, on A = R\{0}, the functions 1/x, sin(x) and x2 are differen-
tiable at every x ∈ A. Hence, if x 6= 0 we have that

f ′ (x) = 2x sin( x1 ) − cos( x1 ).

187
 6.1. THE DERIVATIVE

Consider now c = 0. If f ′ (0) exists it is equal to

f (x) − f (c) x2 sin( x1 )

lim = lim
x→0 x−c x→0 x
= lim x sin( x1 ).
x→0

Using the Squeeze Theorem, we deduce that f ′(0) = 0. Therefore,

(
′ 2x sin( x1 ) − cos( x1 ), x 6= 0
f (x) =
0, x = 0.

From the above formula obtained for f ′ (x), we observe that when x 6= 0
f ′ is continuous since it is the product/difference/composition of con-
tinuous functions. To determine continuity of f ′ at x = 0 consider
1
limx→0 f ′(x). Consider the sequence xn = nπ , which clearly converges
2
to c = 0. Now, f ′ (xn) = nπ sin(nπ) − cos(nπ). Now, sin(nπ) = 0 for all
n and therefore f ′(xn) = − cos(nπ) = (−1)n+1. The sequence f ′(xn)
does not converge and therefore limx→0 f ′(x) does not exist. Thus, f ′
is not continuous at x = 0.

Example 6.1.11. Compute f ′(x) if

(
x3 sin( x1 ), x 6= 0
f (x) =
0, x=0

Where is f ′(x) continuous?

Solution. When x 6= 0, f (x) is the composition and product of differen-

tiable functions, and therefore f is differentiable at x 6= 0. For instance,
on A = R\{0}, the functions 1/x, sin(x) and x3 are differentiable on
A. Hence, if x 6= 0 we have that

f ′ (x) = 3x2 sin( x1 ) − x cos( x1 ).

188
 6.1. THE DERIVATIVE

Consider now c = 0. If f ′ (0) exists it is equal to

f (x) − f (c) x3 sin( x1 )

lim = lim
x→0 x−c x→0 x
2
= lim x sin( x1 )
x→0

and using the Squeeze Theorem we deduce that f ′(0) = 0. Therefore,

(
′ 3x2 sin( x1 ) − x cos( x1 ), x =
6 0
f (x) =
0, x = 0.

When x 6= 0, f ′ is continuous since it is the product/difference/composition

of continuous functions. To determine continuity of f ′ at c = 0 we
consider the limit limx→0 f ′ (x). Now limx→0 3x2 sin( x1 ) = 0 using the
Squeeze Theorem, and similarly limx→0 x cos( x1 ) = 0 using the Squeeze
Theorem. Therefore, limx→0 f ′ (x) exists and is equal to 0, which is
equal to f ′(0). Hence, f ′ is continuous at x = 0, and thus continuous
everywhere.

Example 6.1.12. Consider the function


2 1
x sin( x ), x ∈ Q\{0}

f (x) = x2 cos( x1 ), x ∈
/Q


0, x = 0.

Show that f ′(0) = 0.

189
 6.1. THE DERIVATIVE

Exercises

Exercise 6.1.1. Use the definition of the derivative of a function to

3x + 4
find f ′ (x) if f (x) = . Clearly state the domain of f ′(x).
2x − 1
Exercise 6.1.2. Use the definition of the derivative of a function to
find f ′ (x) if f (x) = x|x|. Clearly state the domain of f ′ (x).

Exercise 6.1.3. Let f : R → R be defined by

(
x2 , x ∈ Q
f (x) =
0, x ∈ R\Q

1. Show that f is differentiable at c = 0 and find f ′(0).

2. Prove that if c 6= 0 then f is not differentiable at c.

Exercise 6.1.4. Let g(x) = |x3| for x ∈ R. Determine whether g ′ (0),

g (2) (0), g (3)(0) exist and if yes find them. Hint: Consider writing g as a
piecewise function and use the definition of the derivative.

Exercise 6.1.5. If f : R → R is differentiable at c ∈ R, explain why

f ′ (c) = lim [n(f (c + 1/n) − f (c))]

n→∞

Give an example of a function f and a number c such that

lim [n(f (c + 1/n) − f (c))]

n→∞

exists but f ′(c) does not exist.

190
 6.2. THE MEAN VALUE THEOREM

6.2 The Mean Value Theorem

Definition 6.2.1: Relative Extrema
Let f : I → R be a function and let c ∈ I.

(i) We say that f has a relative maximum at c if there exists

δ > 0 such that f (x) ≤ f (c) for all x ∈ (c − δ, c + δ).

(ii) We say that f has a relative minimum at c if there exists δ

such that f (c) ≤ f (x) for all x ∈ (c − δ, c + δ).

A point c ∈ I is called a critical point of f : I → R if f ′(c) =

0. The next theorem says that a relative maximum/minimum of a
differentiable function can only occur at a critical point.

Theorem 6.2.2: Critical Point at Extrema

Let f : I → R be a function and let c be an interior point of I.
Suppose that f has a relative maximum (or minimum) at c. If f is
differentiable at c then c is a critical point of f , that is, f ′ (c) = 0.

Proof. Suppose that f has a relative maximum at c; the relative min-

imum case is similar. Then for x 6= c, it holds that f (x) − f (c) ≤ 0
for x ∈ (c − δ, c + δ) and some δ > 0. Consider the function h :
(c − δ, c + δ) → R defined by h(x) = f (x)−f
x−c
(c)
for x 6= c and h(c) = f ′(c).
Then the function h is continuous at c = 0 because limx→c h(x) = h(c).
Now for x ∈ A = (c, c + δ) it holds that h(x) ≤ 0 and therefore
f ′(c) = limx→c h(x) ≤ 0. Similarly, for x ∈ B = (c − δ, c) it holds that
h(x) ≥ 0 and therefore 0 ≤ f ′ (c). Thus f ′ (c) = 0.

191
 6.2. THE MEAN VALUE THEOREM

Corollary 6.2.3
If f : I → R has a relative maximum (or minimum) at c then either
f ′(c) = 0 or f ′(c) does not exist.

Example 6.2.4. The function f (x) = |x| has a relative minimum at

x = 0, however, f is not differentiable at x = 0.

Theorem 6.2.5: Rolle

Let f : [a, b] → R be continuous on [a, b] and differentiable on (a, b).
If f (a) = f (b) then there exists c ∈ (a, b) such that f ′(c) = 0.

Proof. Since f is continuous on [a, b] it achieves its maximum and min-

imum at some point x∗ and x∗, respectively, that is f (x∗) ≤ f (x) ≤
f (x∗) for all x ∈ [a, b]. If f is constant then f ′ (x) = 0 for all x ∈ (a, b).
If f is not constant then f (x∗) < f (x∗). Since f (a) = f (b) it follows
that at least one of x∗ and x∗ is not contained in {a, b}, and hence by
Theorem 6.2.2 there exists c ∈ {x∗, x∗} such that f ′(c) = 0.

We now state and prove the main result of this section.

Theorem 6.2.6: Mean Value

Let f : [a, b] → R be continuous on [a, b] and differentiable on (a, b).
Then there exists c ∈ (a, b) such that f ′(c) = f (b)−f (a)
b−a .

Proof. If f (a) = f (b) then the result follows from Rolle’s Theorem
(f ′(c) = 0 for some c ∈ (a, b)). Let h : [a, b] → R be the line from
(a, f (a)) to (b, f (b), that is,
f (b) − f (a)
h(x) = f (a) + (x − a)
(b − a)
192
 6.2. THE MEAN VALUE THEOREM

and define the function

g(x) = f (x) − h(x)

for x ∈ [a, b]. Then g(a) = f (a) − f (a) = 0 and g(b) = f (b) − f (b) = 0,
and thus g(a) = g(b). Clearly, g is continuous on [a, b] and differentiable
on (a, b), and it is straightforward to verify that g ′ (x) = f ′(x)− f (b)−f
b−a
(a)
.
By Rolle’s Theorem, there exists c ∈ (a, b) such that g ′ (c) = 0, and
therefore f ′(c) = f (b)−f (a)
b−a .

Theorem 6.2.7
Let f : [a, b] → R be continuous on [a, b] and differentiable on (a, b).
If f ′(x) = 0 for all x ∈ (a, b) then f is constant on [a, b].

Proof. Let y ∈ (a, b]. Now f restricted to [a, y] satisfies all the as-
sumptions needed in the Mean Value Theorem. Therefore, there ex-
ists c ∈ (a, y) such that f ′ (c) = f (y)−f (a) ′
y−a . But f (c) = 0 and thus
f (y) = f (a). This holds for all y ∈ (a, b] and thus f is constant on
[a, b].

Example 6.2.8. Show by example that Theorem 6.2.7 is not true for
a function f : A → R if A is not a closed and bounded interval.

Corollary 6.2.9
If f, g : [a, b] → R are continuous and differentiable on (a, b) and
f ′(x) = g ′ (x) for all x ∈ (a, b) then f (x) = g(x) + C for some
constant C.

193
 6.2. THE MEAN VALUE THEOREM

Example 6.2.10. Use the Mean Value theorem to show that −x ≤

sin(x) ≤ x for all x ∈ R.
Proof. Suppose that x > 0 and let g(x) = sin(x) so that g ′ (x) = cos(x).
By the MVT, there exists c ∈ (0, x) such that cos(c) = sin(x) x , that is
sin(x) = x cos(c). Now | cos(c)| ≤ 1 and therefore | sin(x)| ≤ |x| =
x. Therefore, −x ≤ sin(x) ≤ x. The case x < 0 can be treated
similarly.

Definition 6.2.11: Monotone Functions

The function f : I → R is increasing if f (x1) ≤ f (x2) whenever
x1 < x2. Similarly, f is decreasing if f (x2) ≤ f (x1) whenever
x1 < x2. In either case, we say that f is monotone.

The sign of the derivative f ′ determines where f is increasing/decreasing.

Theorem 6.2.12
Suppose that f : I → R is differentiable.

(i) Then f is increasing if and only if f ′(x) ≥ 0 for all x ∈ I.

(ii) Then f is decreasing if and only if f ′(x) ≤ 0 for all x ∈ I.

Proof. Suppose that f is increasing. Then for all x, c ∈ I with x 6= c

it holds that f (x)−f
x−c
(c)
≥ 0 and therefore f ′(c) = limx→c f (x)−f
x−c
(c)
≥ 0.
Hence, this proves that f ′(x) ≥ 0 for all x ∈ I.
Now suppose that f ′(x) ≥ 0 for all x ∈ I. Suppose that x < y. Then
by the Mean Value Theorem, there exists c ∈ (x, y) such that f ′ (c) =
f (y)−f (x)
y−x
. Therefore, since f ′(c) ≥ 0 it follows that f (y) − f (x) ≥ 0.
Part (ii) is proved similarly.

194
 6.2. THE MEAN VALUE THEOREM

Exercises

Exercise 6.2.1. Use the Mean Value Theorem to show that

| cos(x) − cos(y)| ≤ |x − y|.

In general, suppose that f : [a, b] → R is such that f ′ exists on [a, b]

and f ′ is continuous on [a, b]. Prove that f is Lipschitz on [a, b].

Exercise 6.2.2. Give an example of a uniformly continuous function on

[0, 1] that is differentiable on (0, 1) but whose derivative is not bounded
on (0, 1). Justify your answer.

Exercise 6.2.3. Let I be an interval and let f : I → R be differentiable

on I. Prove that if f ′ (x) > 0 for x ∈ I then f is strictly increasing on
I.

Exercise 6.2.4. Let f : [a, b] → R be continuous on [a, b] and differ-

entiable on (a, b). Show that if limx→a f ′(x) = A then f ′ (a) exists and
equals A. Hint: Use the definition of f ′ (a), the Mean Value Theorem,
and the Sequential Criterion for limits.

Exercise 6.2.5. Let f : [a, b] → R be continuous and suppose that f ′

exists on (a, b). Prove that if f ′(x) > 0 for x ∈ (a, b) then f is strictly
increasing on [a, b].

Exercise 6.2.6. Suppose that f : [a, b] → R is continuous on [a, b] and

differentiable on (a, b). We proved that if f ′ (x) = 0 for all x ∈ (a, b)
then f is constant on [a, b]. Give an example of a function f : A → R
such that f ′(x) = 0 for all x ∈ A but f is not constant on A.

Exercise 6.2.7. Let f : [a, b] → R be differentiable. Prove that if

f ′(x) 6= 0 on [a, b] then f is injective.

195
 6.3. TAYLOR’S THEOREM

6.3 Taylor’s Theorem

Taylor’s theorem is a higher-order version of the Mean Value Theo-
rem and it has abundant applications in numerical analysis. Taylor’s
theorem involves Taylor polynomials which you are familiar with from
calculus.

Definition 6.3.1: Taylor Polynomials

Let x0 ∈ [a, b] and suppose that f : [a, b] → R is such that the
derivatives f ′(x0), f (2)(x0), f (3) (x0),. . .,f (n)(x0) exist for some posi-
tive integer n. Then the polynomial
1 (2)
Pn (x) = f (x0) + f ′ (x0)(x − x0) + f (x0)(x − x0)2 +
2!
1 (n)
··· + f (x0)(x − x0 )n
n!
is called the nth order Taylor polynomial of f based at x0.
Using summation convention, Pn (x) can be written as
n
X f (k) (x0)
Pn (x) = (x − x0)k .
k!
k=0

By construction, the derivatives of f and Pn up to order n are

identical at x0 (verify this!):

Pn (x0) = f (x0)
Pn(1) (x0) = f (1) (x0)
.. .
. = ..
P (n) (x0) = f (n) (x0).

It is reasonable then to suspect that Pn (x) is a good approximation to

f (x) for points x near x0 . If x ∈ [a, b] then the difference between f (x)

196
 6.3. TAYLOR’S THEOREM

and Pn (x) is
Rn (x) = f (x) − Pn (x)

and we call Rn (x) the nth order remainder based at x0. Hence, for
each x∗ ∈ [a, b], the remainder Rn (x∗) is the error in approximating
f (x∗) with Pn (x∗). You may be asking yourself why we would need
to approximate f (x) if the function f is known and given. For exam-
ple, if say f (x) = sin(x) then why would we need to approximate say
f (1) = sin(1) since any basic calculator could easily compute sin(1)?
Well, what your calculator is actually computing is an approximation to
sin(1) using a (rational) number such as Pn (1) and using a large value
of n for accuracy (although modern numerical algorithms for comput-
ing trigonometric functions have superseded Taylor approximations but
Taylor approximations are a good start). Taylor’s theorem provides an
expression for the remainder term Rn (x) using the derivative f (n+1).

Theorem 6.3.2: Taylor’s Theorem

Let f : [a, b] → R be a function such that for some n ∈ N the func-
tions f, f (1), f (2), . . . , f (n) are continuous on [a, b] and f (n+1) exists
on (a, b). Fix x0 ∈ [a, b]. Then for any x ∈ [a, b] there exists c
between x0 and x such that

f (x) = Pn (x) + Rn (x)

where
f (n+1)(c)
Rn (x) = (x − x0)n+1.
(n + 1)!

Proof. If x = x0 then Pn (x0) = f (x0) and then c can be chosen arbi-

trarily. Thus, suppose that x 6= x0, let m = f(x−x
(x)−Pn (x)
0)
n+1 , and define the

197
 6.3. TAYLOR’S THEOREM

function g : [a, b] → R by

g(t) = f (t) − Pn (t) − m(t − x0)n+1.

Since f (n+1) exists on (a, b) then g (n+1) exists on (a, b). Moreover,
since P (k) (x0) = f (k) (x0) for k = 0, 1, . . . , n then g (k) (x0) = 0 for
k = 0, 1, . . . , n. Now g(x) = 0 and therefore since g(x0) = 0 by Rolle’s
theorem there exists c1 in between x and x0 such that g ′ (c1 ) = 0. Now
we can apply Rolle’s theorem to g ′ since g ′ (c1) = 0 and g ′ (x0) = 0, and
therefore there exists c2 in between c1 and x0 such that g ′′ (c2 ) = 0. By
applying this same argument repeatedly, there exists c in between x0
and cn−1 such that g (n+1) (c) = 0. Now,

g (n+1) (t) = f (n+1)(t) − m(n + 1)!

and since g (n+1) (c) = 0 then

0 = f (n+1) (c) − m(n + 1)!

from which we conclude that

f (n+1)(c)
f (x) − P (x) = (x − x0)n+1
(n + 1)!

and the proof is complete.

Example 6.3.3. Consider the function f : [0, 2] → R given by f (x) =

ln(1 + x). Use P4 based at x0 = 0 to estimate ln(2) and give a bound
on the error with your estimation.

Solution. Note that f (1) = ln(2) and so the estimate of ln(2) using P4
is ln(2) ≈ P4 (1). To determine P4 we need f (0), f (1)(0), . . . , f (4)(0). We

198
 6.3. TAYLOR’S THEOREM

compute
1
f (1) (x) = f (1) (0) = 1
1+x
−1
f (2) (x) = f (2) (0) = −1
(1 + x)2
2
f (3) (x) = f (3) (0) = 2
(1 + x)3
−6
f (4) (x) = f (4) (0) = −6.
(1 + x)4

Therefore,
P4 (x) = x − 21 x2 + 13 x3 − 41 x4.
Now P4 (1) = 1 − 21 + 13 − 1
4
= 7
12
and therefore
7
ln(2) ≈ P4 (1) = .
12
The error is R4 (1) = f (1) − P4 (1) which is unknown but we can ap-
proximate it using Taylor’s theorem. To that end, by Taylor’s theorem,
for any x ∈ [0, 2] there exists c in between x0 = 0 and x such that
f (5) (c) 5
R4(x) = x
5!
1 24
= 5
x4
5! (1 + c)
1
= .
5(1 + c)5
Therefore, for x = 1, there exists 0 < c < 1 such that
1
R4(1) = .
5(1 + c)5
Therefore, a bound for the error is
1 1
|R4 (1)| = ≤
5(1 + c)5 5
199
 6.3. TAYLOR’S THEOREM

since 1 + c > 1.

Example 6.3.4. Let f : R → R be the sine function, that is, f (x) =

sin(x).

(a) Approximate f (3) = sin(3) using P8 centered at x0 = 0 and give

a bound on the error.

(b) Restrict f to a closed and bounded interval of the form [−R, R].
Show that for any ε > 0 there exists K ∈ N such that if n ≥ K
then |f (x) − Pn (x)| < ε for all x ∈ [−a, a].

Solution. (a) It is straightforward to compute that

1 3 1 5 1 7
P8 (x) = x − x + x − x
3! 5! 7!
and f (9) (x) = sin(x). Thus, by Taylor’s theorem for any x there exists
c in between x = 0 and x such that
sin(c) 9
f (x) − P8 (x) = R8 (x) = x.
9!
The estimate for f (3) = sin(3) is

sin(3) ≈ P8 (3)
1 1 1
= 3 − 33 + 35 − 37
3! 5! 7!
51
=
560
= 0.0910714286

By Taylor’s theorem, there exists c such that 0 < c < 3 and

sin(c) 9
sin(3) − P8 (3) = R8 (3) = 3
9!
200
 6.3. TAYLOR’S THEOREM

Now since | sin(c)| ≤ 1 for all c ∈ R, we have

| sin(c)| 9
|R8 (3)| = 3
9!
39
=
9!
= 0.054241 . . .

(b) Since f (x) = sin(x) has derivatives of all orders, for any n ∈ N we
have by Taylor’s theorem that

|f (x) − Pn (x)| = |Rn (x)|

f (n+1)(c) n+1
= x
(n + 1)!
|f (n+1)(c) n+1
= |x|
(n + 1)!
where c is in between x0 = 0 and x. Now, the derivative of f (x) = sin(x)
of any order is one of ± cos(x) or ± sin(x), and therefore |f (n+1(c)| ≤ 1.
Since x ∈ [−a, a] then |x| ≤ a and therefore |x|n+1 ≤ an+1. Therefore,
for all x ∈ [−a, a] we have
1
|Rn (x)| ≤ an+1
(n + 1)!
an+1
= .
(n + 1)!
an
Consider the sequence xn = n! . Applying the Ratio test we obtain

xn+1 an+1 n!
lim = lim
n→∞ xn n→∞ (n + 1)!an
a
= lim = 0.
n→∞ n + 1

Therefore, by the Ratio test limn→∞ xn = 0. Hence, for any ε > 0 there
exists K ∈ N such that |xn − 0| = xn < ε for all n ≥ K. Therefore, for

201
 6.3. TAYLOR’S THEOREM

all n ≥ K we have that

an+1
|Rn (x)| ≤ <ε
(n + 1)!

for all x ∈ [−a, a].

Taylor’s theorem can be used to derive useful inequalities.

Example 6.3.5. Prove that for all x ∈ R it holds that

1
1 − x2 ≤ cos(x).
2
Solution. Let f (x) = cos(x). Applying Taylor’s theorem to f at x0 = 0
we obtain
1
cos(x) = 1 − x2 + R2 (x)
2
where
f (3) (c) 3 sin(c) 3
R2(x) = x = x
3! 6
and c is in between x0 = 0 and x. Now, if 0 ≤ x ≤ π then 0 < c < π and
then sin(c) > 0, from which it follows that R2 (x) ≥ 0. If on the other
hand −π ≤ x ≤ 0 then −π < c < 0 and then sin(c) < 0, from which it
follows that R2 (x) ≥ 0. Hence, the inequality holds for |x| ≤ π. Now if
|x| ≥ π > 3 then
1
1 − x2 < −3 < cos(x).
2
Hence the inequality holds for all x ∈ R.

202
 6.3. TAYLOR’S THEOREM

Exercises

Exercise 6.3.1. Use Taylor’s theorem to prove that if x > 0 then

1 1 √ 1
1 + x − x2 ≤ 1 + x ≤ 1 + x
2 8 2
√ √
Then use these inequalities to approximate 1.2 and 2, and for each
case determine a bound on the error of your approximation.

Exercise 6.3.2. Let f : R → R be such that f (k) (x) exists for all x ∈ R
and for all k ∈ N (such a function is called infinitely differentiable on
R). Suppose further that there exists M > 0 such that |f (k) (x)| ≤ M for
all x ∈ R and all k ∈ N. Let Pn (x) be the nth order Taylor polynomial
of f centered at x0 = 0. Let I = [−R, R], where R > 0. Prove that for
any fixed ε > 0 there exists K ∈ N such that for n ≥ K it holds that

|f (x) − Pn (x)| < ε

for all x ∈ [−R, R]. Hint: f (n) is continuous on [−R, R] for every
n ∈ N.

Exercise 6.3.3. Euler’s number is approximately e ≈ 2.718281828 . . ..

Use Taylor’s theorem at x0 = 0 on f (x) = ex and the estimate e < 3 to
show that, for all n ∈ N,
 
1 1 1 3
0 < e − 1 + 1 + + + ··· + <
2! 3! n! (n + 1)!

Exercise 6.3.4. Let f : R → R be the cosine function f (x) = cos(x).

Approximate f (2) = cos(2) using P8 centered at x0 = 0 and give a
bound on the error of your estimation.

203
6.3. TAYLOR’S THEOREM

204
 7

Riemann Integration

7.1 The Riemann Integral

We begin with the definition of a partition.

Definition 7.1.1: Partitions

Let a, b ∈ R and suppose a < b. By a partition of the interval [a, b]
we mean a collection of intervals

P = {[x0, x1], [x1, x2], . . . , [xn−1, xn]}

such that a = x0 < x1 < x2 < · · · < xn = b and where n ∈ N.

Hence, a partition P defines a finite collection of non-overlapping in-

tervals Ik = [xk−1, xk ], where k = 1, . . . , n. The norm of a partition P
is defined as

kPk = max{x1 − x0, x2 − x1 , . . . , xn − xn−1}.

In other words, kPk is the maximum length of the intervals in P. To

ease our notation, we will denote a partition as P = {[xk−1, xk ]}nk=1.
Let P = {[xk−1, xk ]}nk=1 be a partition of [a, b]. If tk ∈ Ik = [xk−1, xk ]

205
 7.1. THE RIEMANN INTEGRAL

then we say that tk is a sample of Ik and the set of ordered pairs

Ṗ = {([xk−1, xk ], tk )}nk=1

will be called a sampled partition.

Example 7.1.2. Examples of sampled partitions are mid-points, right-
end points, and left-end points partitions.
Now consider a function f : [a, b] → R and let Ṗ = {([xk−1, xk ], tk )}nk=1
be a sampled partition of the interval [a, b]. The Riemann sum of f
corresponding to Ṗ is the number
n
X
S(f ; Ṗ) = f (tk )(xk − xk−1).
k=1

When f (x) > 0 on the interval [a, b], the Riemann sum S(f ; Ṗ) is
the sum of the areas of the rectangles with height f (tk ) and width
(xk − xk−1).
We now define the notion of Riemann integrability.

Definition 7.1.3: Riemann Integrability

The function f : [a, b] → R is said to be Riemann integrable if
there exists a number L ∈ R such that for every ε > 0 there exists
δ > 0 such that for any sampled partition Ṗ that satisfies kṖk < δ
it holds that |S(f ; Ṗ) − L| < ε.

The set of all Riemann integrable functions on the interval [a, b] will be
denoted by R[a, b].

Theorem 7.1.4
If f ∈ R[a, b] then the number L in the definition of Riemann
integrability is unique.

206
 7.1. THE RIEMANN INTEGRAL

Proof. Let L1 and L2 be two real numbers satisfying the definition of

Riemann integrability and let ε > 0 be arbitrary. Then there exists
δ > 0 such that |S(f ; Ṗ) − L1| < ε/2 and |S(f ; Ṗ) − L2 | < ε/2, for all
sampled partitions Ṗ with kṖk < δ. Then, if kṖk < δ it holds that

|L1 − L2| ≤ |S(f ; Ṗ) − L1| + |S(f ; Ṗ) − L2|

< ε.

By Theorem 2.2.7 this proves that L1 = L2.

If f ∈ R[a, b], we call the number L the integral of f over [a, b] and
we denote it by Z b
L= f
a

Example 7.1.5. Show that a constant function on [a, b] is Riemann

integrable.

Proof. Let f : [a, b] :→ R be such that f (x) = C for all x ∈ [a, b] and
let Ṗ = {([xk−1, xk ], tk )} be a sampled partition of [a, b]. Then
n
X
S(f ; Ṗ) = f (tk )(xk − xk−1)
k=1
n
X
=C (xk − xk−1)
k=1
= C(xn − x0)
= C(b − a).

Hence, with L = C(b − a), we obtain that |S(f ; Ṗ) − L| = 0 < ε for any
Rb
ε > 0 and therefore a f = C(b − a). This proves that f is Riemann
integrable.

Example 7.1.6. Prove that f (x) = x is Riemann integrable on [a, b].

207
 7.1. THE RIEMANN INTEGRAL

Proof. We consider the special case that [a, b] = [0, 1], the general case
is similar. Let Q̇ = {([xk−1, xk ], qk )} be a sampled partition of [0, 1]
chosen so that qk = 21 (xk + xk−1), i.e., qk is the midpoint of the interval
[xk−1, xk ]. Then

n
X
S(f ; Q̇) = f (qk )(xk − xk−1)
k=1

n
X
1
= 2 (xk + xk−1)(xk − xk−1)
k=1

n
1X 2
= (xk − x2k−1)
2
k=1

1
= (x2n − x20)
2
1
= (12 − 02)
2
1
= .
2

Now let Ṗ = {([xk−1, xk ]), tk }nk=1 be an arbitrary sampled partition of

[0, 1] and suppose that kṖk < δ, so that (xk − xk−1) ≤ δ for all k =
1, 2, . . . , n. If Q̇ = {([xk−1, xk ], qk )}nk=1 is the corresponding midpoint

208
 7.1. THE RIEMANN INTEGRAL

sampled partition then |tk − qk | < δ. Therefore,

n
X
|S(f ; Ṗ) − S(f ; Q̇)| = tk (xk − xk−1) − qk (xk − xk−1)
k=1

n
X
≤ |tk − qk |(xk − xk−1)
k=1

< δ(1 − 0)

= δ.

Hence, we have proved that for arbitrary Ṗ that satisfies kṖk < δ it
holds that |S(f ; Ṗ) − 1/2| < δ. Hence, given ε > 0 we let δ = ε and
R1
then if kṖk < δ then |S(f ; Ṗ) − 1/2| < ε. Therefore, 0 f = 12 .

The next result shows that if f ∈ R[a, b] then changing f at a finite

Rb
number of points in [a, b] does not affect the value of a f .

Theorem 7.1.7
Let f ∈ R[a, b] and let g : [a, b] → R be a function such that
g(x) = f (x) for all x ∈ [a, b] except possibly at a finite number of
Rb Rb
points in [a, b]. Then g ∈ R[a, b] and in fact a g = a f .

Rb
Proof. Let L = a f . Suppose that g(x) = f (x) except at one point
x = c. Let Ṗ = {([xk−1, xk ], tk )} be a sampled partition. We consider
mutually exclusive cases. First, if c 6= tk and c 6= xk for all k then
S(f ; Ṗ) = S(g; Ṗ). If c = tk ∈
/ {x0, x1, . . . , xn} for some k then

S(g; Ṗ) − S(f ; Ṗ) = (f (c) − g(c))(xk − xk−1).

If c = tk = tk−1 for some k then necessarily c = xk−1 and then

S(g; Ṗ)−S(f ; Ṗ) = (f (c)−g(c))(xk −xk−1)+(f (c)−g(c))(xk−1 −xk−2).

209
 7.1. THE RIEMANN INTEGRAL

Hence, in any case, by the triangle inequality

|S(g; Ṗ) − S(f ; Ṗ)| ≤ 2(|f (c)| + |g(c)|)kṖk

= MkṖk.

where M = 2(|f (c)| + |g(c)|). Let ε > 0 be arbitrary. Then there exists
δ1 > 0 such that |S(f ; Ṗ) − L| < ε/2 for all partitions Ṗ such that
kṖk < δ1 . Let δ = min{δ1, ε/(2M)}. Then if kṖk < δ then

kS(g; f ) − Lk ≤ |S(g; Ṗ) − S(f ; Ṗ)| + |S(f ; Ṗ) − L|

< MkṖ| + ε/2

< Mε/(2M) + ε/2

= ε.

Rb Rb
This proves that g ∈ R[a, b] and a g = L = a f . Now suppose by
induction that if g(x) = f (x) for all x ∈ [a, b] except at a j ≥ 1 number
Rb Rb
of points in [a, b] then g ∈ R[a, b] and a g = a f . Now suppose that
h : [a, b] → R is such that h(x) = f (x) for all x ∈ [a, b] except at the
points c1 , c2, . . . , cj , cj+1. Define the function g by g(x) = h(x) for all
x ∈ [a, b] except at x = cj+1 and define g(cj+1 ) = f (cj+1). Then g
and f differ at the points c1 , . . . , cj . Then by the induction hypothesis,
Rb Rb
g ∈ R[a, b] and a g = a f . Now g and h differ at the point cj+1 and
Rb Rb Rb
therefore h ∈ R[a, b] and a h = a g = a f . This ends the proof.

We now state some properties of the Riemann integral.

210
 7.1. THE RIEMANN INTEGRAL

Theorem 7.1.8: Properties of the Riemann Integral

Suppose that f, g ∈ R[a, b]. The following hold.
Rb Rb
(i) If k ∈ R then (kf ) ∈ R[a, b] and a kf = k a f .
Rb Rb Rb
(ii) (f + g) ∈ R[a, b] and a (f + g) = a f + a g.
Rb Rb
(iii) If f (x) ≤ g(x) for all x ∈ [a, b] then a f ≤ a g.

Rb
Proof. If k = 0 then (kf )(x) = 0 for all x and then clearly a kf = 0,
so assume that k 6= 0. Let ε > 0 be given. Then there exists δ > 0 such
Rb
that if kṖk < δ then |S(f ; Ṗ) − a f k < ε/|k|. Now for any partition
Ṗ, it holds that S(kf ; Ṗ) = kS(f ; Ṗ). Therefore, if kṖk < δ then

Z b Z b
S(kf ; Ṗ) − k f = |k| S(f ; Ṗ) − f
a a
< |k|(ε/|k|)
= ε.

To prove (b), it is easy to see that S(f + g; Ṗ) = S(f ; Ṗ) + S(g; Ṗ).
Rb
Given ε > 0 there exists δ > 0 such that |S(f ; Ṗ) − a f | < ε/2 and
Rb
|S(g; Ṗ) − a g| < ε/2, whenever kṖk < δ. Therefore, if kṖk < δ we
have that
Z b Z b  Z b Z b
S(f + g; Ṗ) − f+ g ≤ S(f ; Ṗ) − f + S(g; Ṗ) − g
a a a a
< ε.

To prove (c), let ε > 0 be arbitrary and let δ > 0 be such that if

211
 7.1. THE RIEMANN INTEGRAL

kṖk < δ then

Z b Z b
f − ε/2 < S(f ; Ṗ) < f + ε/2
a a
Z b Z b
g − ε/2 < S(g; Ṗ) < g + ε/2
a a

Now, by assumption, S(f ; Ṗ) ≤ S(g; Ṗ) and therefore

Z b Z b
f − ε/2 < S(f ; Ṗ) ≤ S(g; Ṗ) < g + ε/2.
a a

Therefore,
Z b Z b
f< g + ε.
a a

Since ε is arbitrary, we can choose εn = 1/n and then passing to the

Rb Rb
limit we deduce that a f ≤ a g.

Properties (i), (ii), and (iii) in Theorem 7.1.8 are known as homogeneity,
additivity, and monotonicity, respectively.
We now give a necessary condition for Riemann integrability.

Theorem 7.1.9: Integrable Functions are Bounded

If f ∈ R[a, b] then f is bounded on [a, b].

Rb
Proof. Let f ∈ R[a, b] and put L = a f . There exists δ > 0 such that
if kṖk < δ then |S(f ; Ṗ) − L| < 1 and therefore |S(f ; Ṗ)| < |L| + 1.
Suppose by contradiction that f is unbounded on [a, b]. Let P be a
partition of [a, b], with sets I1, . . . , In, and with kPk < δ. Then f is
unbounded on some Ij , i.e., for any M > 0 there exists x ∈ Ij =
[xj−1, xj ] such that f (x) > M. Choose samples in P by asking that

212
 7.1. THE RIEMANN INTEGRAL

tk = xk for k 6= j and tj is such that

X
|f (tj )(xj − xj−1)| > |L| + 1 + f (tk )(xk − xk−1) .
k6=j

Therefore, (using |a| = |a + b − b| ≤ |a + b| + |b| implies that |a + b| ≥

|a| − |b|)

X
|S(f ; Ṗ)| = f (tj )(xj − xj−1) + f (tk )(xk − xk−1)
k6=j

X
≥ |f (tj )(xj − xj−1)| − f (tk )(xk − xk−1)
k6=j

> |L| + 1.

This is a contradiction and thus f is bounded on [a, b].

Example 7.1.10 (Thomae). Consider Thomae’s function h : [0, 1] →

R defined as h(x) = 0 if x is irrational and h(m/n) = 1/n for every
rational m/n ∈ [0, 1], where gcd(m, n) = 1. In Example 5.1.7, we
proved that h is continuous at every irrational but discontinuous at
every rational. Prove that h is Riemann integrable.

Proof. Let ε > 0 be arbitrary and let E = {x ∈ [0, 1] : h(x) ≥ ε/2}.

By definition of h, the set E is finite, say consisting of n elements. Let
δ = ε/(4n) and let Ṗ be a sampled partition of [0, 1] with kṖk < δ.
We can separate the partition Ṗ into two sampled partitions Ṗ1 and
Ṗ2 where Ṗ1 has samples in the the set E and Ṗ2 has no samples in E.
Then the number of intervals in Ṗ1 can be at most 2n, which occurs
when all the elements of E are samples and they are at the endpoints

213
 7.1. THE RIEMANN INTEGRAL

of the subintervals of Ṗ1. Therefore, the total length of the intervals

in Ṗ1 can be at most 2nδ = ε/2. Now 0 < h(tk ) ≤ 1 for every sample
tk in Ṗ1 and therefore S(f ; Ṗ1) ≤ 2nδ = ε/2. For samples tk in Ṗ2
we have that h(tk ) < ε/2. Therefore, since the sum of the lengths of
the subintervals of Ṗ2 is ≤ 1, it follows that S(f ; Ṗ2) ≤ ε/2. Hence
Rb
0 ≤ S(f ; Ṗ) = S(f ; Ṗ1) + S(f ; Ṗ2) < ε. Thus a h = 0.

214
 7.1. THE RIEMANN INTEGRAL

Exercises

Exercise 7.1.1. Suppose that f, g ∈ R[a, b] and let α, β ∈ R. Prove

by definition that (αf + βg) ∈ R[a, b].

Exercise 7.1.2. If f is Riemann integrable on [a, b] and |f (x)| ≤ M

Rb
for all x ∈ [a, b], prove that | a f | ≤ M(b − a). Hint: The inequality
|f (x)| ≤ M is equivalent to −M ≤ f (x) ≤ M. Then use the fact that
constants functions are Riemann integrable whose integrals are easily
computed. Finally, apply a theorem from this section.

Exercise 7.1.3. If f is Riemann integrable on [a, b] and (Ṗn ) is a se-

quence of tagged partitions of [a, b] such that kṖn k → 0 prove that
Z b
f = lim S(f ; Ṗn)
a n→∞

Hint: For each n ∈ N we have the real number sn = S(f ; Ṗn), and we
Rb
therefore have a sequence (sn). Let L = a f . We therefore want to
prove that lim sn = L.
n→∞

Exercise 7.1.4. Give an example of a function f : [0, 1] → R that

is Riemann integrable on [c, 1] for every c ∈ (0, 1) but which is not
Riemann integrable on [0, 1]. Hint: What is a necessary condition for
Riemann integrability?

215
 7.2. RIEMANN INTEGRABLE FUNCTIONS

7.2 Riemann Integrable Functions

To ease our notation, if I is a bounded interval with end-points a < b
we denote by µ(I) the length of I, that is µ(I) = b − a. Hence, if
I = [a, b], I = [a, b), I = (a, b], or I = (a, b) then µ(I) = b − a.
Thus far, to establish the Riemann integrability of f , we computed
Rb
a candidate integral L and showed that in fact L = a f . The following
theorem is useful when a candidate integral L is unknown. The proof
is omitted.

Theorem 7.2.1: Cauchy Criterion

A function f : [a, b] → R is Riemann integrable if and only if for
every ε > 0 there exists δ > 0 such that if Ṗ and Q̇ are sampled
partitions of [a, b] with norm less than δ then

|S(f ; Ṗ) − S(f ; Q̇)| < ε.

Using the Cauchy Criterion, we show next that the Dirichlet func-
tion is not Riemann integrable.

Example 7.2.2 (Non-Riemann integrable function). Let f : [0, 1] → R

be defined as f (x) = 1 if x is rational and f (x) = 0 if x is irrational.
Show that f is not Riemann integrable.

Proof. To show that f is not in R[0, 1], we must show that there exists
ε0 > 0 such that for all δ > 0 there exists sampled partitions Ṗ and Q̇
with norm less than δ but |S(f ; Ṗ) − S(f ; Q̇)| ≥ ε0. To that end, let
ε0 = 1/2, and let δ > 0 be arbitrary. Let n be sufficiently large so that
1/n < δ. Let Ṗ be a sampled partition of [0, 1] with intervals all of equal
length 1/n < δ and let the samples of Ṗ be rational numbers. Similarly,
let Q̇ be a partition of [0, 1] with intervals all of equal length 1/n and

216
 7.2. RIEMANN INTEGRABLE FUNCTIONS

with samples irrational numbers. Then S(f ; Ṗ) = 1 and S(f ; Q̇) = 0,
and therefore |S(f ; Q̇) − S(f ; Q̇)| ≥ ε0.

We now state a sort of squeeze theorem for integration.

Theorem 7.2.3: Squeeze Theorem

Let f be a function on [a, b]. Then f ∈ R[a, b] if and only if for
every ε > 0 there exist functions α and β in R[a, b] with α(x) ≤
Rb
f (x) ≤ β(x) for all x ∈ [a, b] and a (β − α) < ε.

Rb
Proof. If f ∈ R[a, b] then let α(x) = β(x) = f (x). Then clearly a (β −
α) = 0 < ε for all ε > 0. Now suppose the converse and let ε > 0
be arbitrary. Let α and β satisfy the conditions of the theorem, with
Rb ε
a (β − α) < 3 . Now, there exists δ > 0 such that if kṖk < δ then
Z b Z b
ε ε
α − < S(α; Ṗ) < α+
a 3 a 3
and Z Z
b b
ε ε
β − < S(β; Ṗ) < β+ .
a 3 a 3
For any sampled partition Ṗ it holds that S(α; Ṗ) ≤ S(f ; Ṗ) ≤ S(β; Ṗ),
and therefore Z b Z b
ε ε
α − < S(f ; Ṗ) < β+ . (7.1)
a 3 a 3
If Q̇ is another sampled partition with kQ̇k < δ then also
Z b Z b
ε ε
α − < S(f ; Q̇) < β+ . (7.2)
a 3 a 3
Subtracting the two inequalities (7.1)-(7.2), we deduce that
Z b Z b
ε ε
− (β − α) − 2 < S(f ; Ṗ) − S(f ; Q̇) < (β − α) + 2 .
a 3 a 3
217
 7.2. RIEMANN INTEGRABLE FUNCTIONS

Rb ε
Therefore, since a (β − α) < 3 it follows that

−ε < S(f ; Ṗ) − S(f ; Q̇) < ε.

By the Cauchy criterion, this proves that f ∈ R[a, b].

Step-functions, defined below, play an important role in integration

theory.

Definition 7.2.4
A function s : [a, b] → R is called a step-function on [a, b] if
there is a finite number of disjoint intervals I1, I2, . . . , In contained
S
in [a, b] such that [a, b] = nk=1 Ik and such that s is constant on
each interval.

In the definition of a step-function, the intervals Ik may be of any

form, i.e., half-closed, open, or closed.

Lemma 7.2.5
Let J be a subinterval of [a, b] and define ϕJ on [a, b] as ϕJ (x) = 1
Rb
if x ∈ J and ϕJ (x) = 0 otherwise. Then ϕJ ∈ R[a, b] and a ϕJ =
µ(J).

Theorem 7.2.6
If ϕ : [a, b] → R is a step function then ϕ ∈ R[a, b].

Proof. Let I1, . . . , In be the intervals where ϕ is constant, and let c1 , . . . , cn

be the constant values taken by ϕ on the intervals I1, . . . , In, respec-
Pn
tively. Then it is not hard to see that ϕ = k=1 ck ϕIk . Then ϕ is

218
 7.2. RIEMANN INTEGRABLE FUNCTIONS

the sum of Riemann integrable functions and therefore is also Riemann

Rb P
integrable. Moreover, a ϕ = nk=1 ck µ(Ik ).

We will now show that any continuous function on [a, b] is Riemann

integrable. To do that we will need the following.

Lemma 7.2.7: Continuity and Step-Functions

Let f : [a, b] → R be a continuous function. Then for every ε > 0
there exists a step-function s : [a, b] → R such that |f (x)−s(x)| < ε
for all x ∈ [a, b].

Proof. Let ε > 0 be arbitrary. Since f is uniformly continuous on [a, b]

there exists δ > 0 such that if |x − u| < δ then |f (x) − f (u)| < ε. Let
n ∈ N be sufficiently large so that (b − a)/n < δ. Partition [a, b] into n
subintervals of equal length (b−a)/n, and denote them by I1, I2, . . . , In,
where I1 = [x0, x1] and Ik = (xk−1, xk ] for 1 < k ≤ n. Then for x, u ∈ Ik
it holds that |f (x) − f (u)| < ε. For x ∈ Ik define s(x) = f (xk ).
Therefore, for any x ∈ Ik it holds that |f (x)−s(x)| = |f (x)−f (xk )| < ε.
S
Since nk=1 Ik = [a, b], it holds that |f (x)−s(x)| < ε for all x ∈ [a, b].

We now prove that continuous functions are integrable.

Theorem 7.2.8: Continuous Functions are Integrable

A continuous function on [a, b] is Riemann integrable on [a, b].

Proof. Suppose that f : [a, b] → R is continuous. Let ε > 0 be arbitrary

and let ε̃ = (ε/4)/(b − a). Then there exists a step-function s : [a, b] →
R such that |f (x) − s(x)| < ε̃ for all x ∈ [a, b]. In other words, for all
x ∈ [a, b] it holds that

s(x) − ε̃ < f (x) < s(x) + ε̃.

219
 7.2. RIEMANN INTEGRABLE FUNCTIONS

The functions α(x) := s(x) − ε̃ and β(x) := s(x) + ε̃ are Riemann

Rb
integrable integrable on [a, b], and a (β − α) = 2ε̃(b − a) = ε/2 < ε.
Hence, by the Cauchy criterion, f is Riemann integrable.

Recall that a function is called monotone if it is decreasing or in-

creasing.

Theorem 7.2.9: Monotone Functions are Integrable

A monotone function on [a, b] is Riemann integrable on [a, b].

Proof. Assume that f : [a, b] → R is increasing and that M = f (b) −

f (a) > 0 (if M = 0 then f is the zero function which is clearly inte-
grable). Let ε > 0 be arbitrary. Let n ∈ N be such that M (b−a) n < ε.
Partition [a, b] into subintervals of equal length ∆x = (b−a) n , and as
usual let a = x0 < x1 < · · · < xn−1 < xn = b denote the resulting
points of the partition. On each subinterval [xk−1, xk ], it holds that
f (xk−1) ≤ f (x) ≤ f (xk ) for all x ∈ [xk−1, xk ] since f is increasing.
Let α : [a, b] → R be the step-function whose constant value on the
interval [xk−1, xk ) is f (xk−1) and similarly let β : [a, b] → R be the
step-function whose constant value on the interval [xk−1, xk ) is f (xk ),
for all k = 1, . . . , n. Then α(x) ≤ f (x) ≤ β(x) for all x ∈ [a, b]. Both
α and β are Riemann integrable and
Z b n
X
(β − α) = [f (xk ) − f (xk−1)]∆x
a k=1
= (f (xn) − f (xn−1)∆x
(b − a)
=M
n
< ε.
Hence by the Squeeze theorem for integrals (Theorem 7.2.3), f ∈
R[a, b].

220
 7.2. RIEMANN INTEGRABLE FUNCTIONS

Our las theorem is the additivity property of the integral, the proof
is omitted.

Theorem 7.2.10: Additivity Property

Let f : [a, b] → R be a function and let c ∈ (a, b). Then f ∈ R[a, b]
if and only if its restrictions to [a, c] and [c, b] are both Riemann
integrable. In this case,
Z b Z c Z b
f= f+ f
a a c

221
 7.2. RIEMANN INTEGRABLE FUNCTIONS

Exercises

Exercise 7.2.1. Suppose that f : [a, b] → R is continuous and assume

Rb
that f (x) > 0 for all x ∈ [a, b]. Prove that a f > 0. Hint: A contin-
uous function on a closed and bounded interval achieves its minimum
value.

Exercise 7.2.2. Suppose that f is continuous on [a, b] and that f (x) ≥

0 for all x ∈ [a, b].
Rb
(a) Prove that if a f = 0 then necessarily f (x) = 0 for all x ∈ [a, b].

(b) Show by example that if we drop the assumption that f is con-

tinuous on [a, b] then it may not longer hold that f (x) = 0 for all
x ∈ [a, b].

Exercise 7.2.3. Show that if f : [a, b] → R is Riemann integrable then

|f | : [a, b] → R is also Riemann integrable.

222
 7.3. THE FUNDAMENTAL THEOREM OF CALCULUS

7.3 The Fundamental Theorem of Calcu-

lus
Theorem 7.3.1: FTC Part I
Let f : [a, b] → R be a function. Suppose that there exists a
finite set E ⊂ [a, b] and a function F : [a, b] → R such that F is
continuous on [a, b] and F ′ (x) = f (x) for all x ∈ [a, b]\E. If f is
Rb
Riemann integrable then a f = F (b) − F (a).

Proof. Assume for simplicity that E := {a, b}. Let ε > 0 be arbitrary.
Rb
Then there exists δ > 0 such that if kṖk < ε then |S(f ; Ṗ) − a f | < ε.
For any Ṗ, with intervals Ik = [xk−1, xk ] for k = 1, 2, . . . , n, there exists,
by the Mean Value Theorem applied to F on Ik , a point uk ∈ (xk−1, xk )
such that F (xk ) − F (xk−1) = F ′ (uk )(xk − xk−1). Therefore,
n
X
F (b) − F (a) = F (xk ) − F (xk−1)
k=1
n
X
= f (uk )(xk − xk−1)
k=1
= S(f ; Ṗu)

where Ṗu has the same intervals as Ṗ but with samples uk . Therefore,
if kṖk < δ then
Z b Z b
F (b) − F (a) − f = S(f ; Ṗu) − f
a a
< ε.
Rb
Hence, for any ε we have that F (b) − F (a) − a f < ε and this shows
Rb
that a f = F (b) − F (a).

223
 7.3. THE FUNDAMENTAL THEOREM OF CALCULUS

Definition 7.3.2: Indefinite Integral

Let f ∈ R[a, b]. The indefinite integral of f with basepoint a is
the function F : [a, b] → R defined by
Z x
F (x) := f
a

for x ∈ [a, b].

Theorem 7.3.3
Let f ∈ R[a, b]. The indefinite integral F : [a, b] → R of f : [a, b] →
R is a Lipschitz function on [a, b], and thus continuous on [a, b].

Proof. For any w, z ∈ [a, b] such that w ≤ z it holds that

Z z
F (z) = f
Za w Z z
= f+ f
a w
Z z
= F (w) + f
w
Rz
and therefore F (z) − F (w) = w f . Since f is Riemann integrable
on [a, b] it is bounded and therefore |f (x)| ≤ K for all x ∈ [a, b]. In
particular, −K ≤ f (x) ≤ K for all x ∈ [w, z] and thus −K(z − w) ≤
Rz
w f ≤ K(z − w), and thus
Z z
|F (z) − F (w)| = f
w
≤ K|z − w|

224
 7.4. RIEMANN-LEBESGUE THEOREM

Under the additional hypothesis that f ∈ R[a, b] is continuous, the

indefinite integral of f is differentiable.

Theorem 7.3.4: FTC Part 2

Let f ∈ R[a, b] and let f be continuous at a point c ∈ [a, b]. Then
the indefinite integral F of f is differentiable at c and F ′ (c) = f (c).

7.4 Riemann-Lebesgue Theorem

In this section we present a complete characterization of Riemann in-
tegrability for a bounded function. Roughly speaking, a bounded func-
tion is Riemann integrable if the set of points were it is discontinuous
is not too large. We first begin with a definition of “not too large”.

Definition 7.4.1
A set E ⊂ R is said to be of measure zero if for every ε > 0 there
exists a countable collection of open intervals Ik such that
∞
[ ∞
X
E⊂ Ik and µ(Ik ) < ε.
k=1 k=1

Example 7.4.2. Show that a subset of a set of measure zero also has
measure zero. Show that the union of two sets of measure zero is a set
of measure zero.

Example 7.4.3. Let S ⊂ R be a countable set. Show that S has

measure zero.

225
 7.4. RIEMANN-LEBESGUE THEOREM

Solution. Let S = {s1 , s2, s3, . . .}. Consider the interval

 ε ε 
Ik = sk − k+1 , sk + k+1 .
2 2
S
Clearly, sk ∈ Ik and thus S ⊂ Ik . Moreover,
∞ ∞
X X ε
µ(Ik ) = = ε.
2k
k=1 k=1

As a corollary, Q has measure zero.

However, there exists uncountable sets of measure zero.

Example 7.4.4. The Cantor set is defined as follows. Start with I0 =

[0, 1] and remove the middle third J1 = ( 31 , 32 ) yielding the set I1 =
I0\J1 = [0, 31 ] ∪ [ 32 , 1]. Notice that µ(J1) = 13 . Now remove from each
subinterval of I1 the middle third resulting in the set


I2 = I1\ ( 19 , 29 ) ∪ ( 97 , 89 ) = [0, 91 ] ∪ [ 92 , 39 ] ∪ [ 69 , 97 ] ∪ [ 89 , 1]

The two middle thirds J2 = ( 19 , 92 ) ∪ ( 97 , 89 ) removed have total length

µ(J2) = 2 91 . By induction, having constructed In which consists of the
union of 2n closed subintervals of [0, 1], we remove from each subinterval
of In the middle third resulting in the set In+1 = In \Jn+1, where Jn+1
is the union of the 2n middle third open intervals and In+1 now consists
of 2n+1 disjoint closed-subintervals. By induction, the total length of
2n
Jn+1 is µ(Jn+1) = 3n+1 . The Cantor set is defined as
∞
\
C= In .
n=1

We now state the Riemann-Lebesgue theorem.

226
 7.4. RIEMANN-LEBESGUE THEOREM

Theorem 7.4.5: Riemann-Lebesgue

Let f : [a, b] → R be a bounded function. Then f is Riemann
integrable if and only if the points of discontinuity of f forms a set
of measure zero.

227
7.4. RIEMANN-LEBESGUE THEOREM

228
 8

Sequences of Functions

In the previous sections, we have considered real-number sequences,

that is, sequences (xn) such that xn ∈ R for each n ∈ N. In this sec-
tion, we consider sequences whose terms are functions. Sequences of
functions arise naturally in many applications in physics and engineer-
ing. A typical way that sequences of functions arise is in the problem
of solving an equation in which the unknown is a function f . In many
of these types of problems, one is able to generate a sequence of func-
tions (fn ) = (f1, f2, f3, . . .) through some algorithmic process with the
intention that the sequence of functions (fn ) converges to the solution
f . Moreover, it would be desirable that the limiting function f inherit
as many properties possessed by each function fn such as, for example,
continuity, differentiability, or integrability. We will see that this latter
issue is rather delicate. In this section, we develop a notion of the limit
of a sequence of functions and then investigate if the fundamental prop-
erties of boundedness, continuity, integrability, and differentiability are
preserved under the limit operation.

229
 8.1. POINTWISE CONVERGENCE

8.1 Pointwise Convergence

Let A ⊂ R be a non-empty subset and suppose that for each n ∈ N we
have a function fn : A → R. We then say that (fn ) = (f1, f2, f3, . . . , )
is a sequence of functions on A.

Example 8.1.1. Let A = [0, 1] and let fn (x) = xn for n ∈ N and

x ∈ A. Then (fn) = (f1, f2, f3, . . .) is a sequence of functions on A. As
another example, for n ∈ N and x ∈ A let gn (x) = nx(1 − x2 )n. Then
(gn ) = (g1, g2, g3, . . .) is a sequence of functions on A. Or how about

fn (x) = an cos(nx) + bn sin(nx)

where an , bn ∈ R and x ∈ [−π, π].

Let (fn) be a sequence of functions on A. For each fixed x ∈ A

we obtain a sequence of real numbers (xn) by simply evaluating each
fn at x, that is, xn = fn (x). For example, if fn (x) = xn and we fix

n
x = 43 then we obtain the sequence xn = fn ( 43 ) = 43 . If x ∈ A is
fixed we can then easily talk about the convergence of the sequence of
numbers (fn(x)) in the usual way. This leads to our first definition of
convergence of function sequences.

Definition 8.1.2: Pointwise Convergence

Let (fn) be a sequence of functions on A ⊆ R. We say that (fn)
converges pointwise on A to the function f : A → R if for
each x ∈ A the sequence (fn(x)) converges to the number f (x), that
is,
lim fn(x) = f (x).
n→∞
In this case, we call the function f the pointwise limit of the
sequence (fn).

230
 8.1. POINTWISE CONVERGENCE

By uniqueness of limits of sequences of real numbers (Theorem 3.1.12),

the pointwise limit of a sequence (fn) is unique. Also, when the domain
A is understood, we will simply say that (fn) converges pointwise to f .

Example 8.1.3. Consider the sequence (fn) defined on R by fn(x) =

(2xn + (−1)nx2)/n. For fixed x ∈ R we have

2xn + (−1)nx2
lim fn (x) = lim = 2x.
n→∞ n→∞ n
Hence, (fn ) converges pointwise to f (x) = 2x on R. In Figure 8.1, we
graph fn(x) for the values n = 1, 2, 3, 4 and the function f (x) = 2x.
Notice that fn′ (x) = (2n + 2(−1)nx)/n and therefore limn→∞ fn′ (x) = 2,
and for the limit function f (x) = 2x we have f ′(x) = 2. Hence, the
sequence of derivatives (fn′ ) converges pointwise to f ′ . Also, after some
basic computations,
Z 1 Z 1
2xn + (−1)nx2
fn (x) dx = dx
−1 −1 n
2(−1)n
=
3n
and therefore
Z 1
2(−1)n
lim fn(x) dx = lim
n→∞ −1 n→∞ 3n
= 0.
R1
On the other hand it is clear that −1 f (x) dx = 0.

Before considering more examples, we state the following result

which is a direct consequence of the definition of the limit of a sequence
of numbers and the definition of pointwise convergence.

231
 8.1. POINTWISE CONVERGENCE
f1

f2
20
f3

f4

10

0
−4 −2 0 2 4

−10

−20

−30

2xn+(−1)n x2
Figure 8.1: Graph of fn(x) = n for n = 1, 2, 3, 4 and f (x) = 2x

Lemma 8.1.4
Let (fn) be a sequence of functions on A. Then (fn) converges
pointwise to f : A → R if and only if for each x ∈ A and each ε > 0
there exists K ∈ N such that |fn (x) − f (x)| < ε for all n ≥ K.

As the following example shows, it is important to note that the K in

Lemma 8.1.4 depends not only on ε > 0 but in general will also depend
on x ∈ A.

Example 8.1.5. Consider the sequence (fn) defined on A = [0, 1] by

fn (x) = xn. For all n ∈ N we have fn (1) = 1n = 1 and therefore
limn→∞ fn(1) = 1. On the other hand if x ∈ [0, 1) then

lim fn (x) = lim xn = 0.

n→∞ n→∞

232
 8.1. POINTWISE CONVERGENCE

Therefore, (fn) converges pointwise on A to the function

(
0, if x ∈ [0, 1)
f (x) =
1, if x = 1.

In Figure 8.2, we graph fn(x) = xn for various values of n. Consider

a fixed x ∈ (0, 1). Since limn→∞ xn = 0 it follows that for ε > 0 there
exists K ∈ N such that |xn − 0| < ε for all n ≥ K. For ε < 1, in order
for |xK | = xK < ε we can choose K > ln(ε)/ ln(x). Notice that K
clearly depends on both ε and x, and in particular, as x get closer to 1
then a larger K is needed. We note that each fn is continuous while f
is not.
f1
1.0
f3

f8

f12

f25

0.8

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0

Figure 8.2: Graph of fn(x) = xn for n = 1, 3, 8, 12, 25

Example 8.1.5 also illustrates a weakness of pointwise convergence,

namely, that if (fn) is a sequence of continuous functions on A and (fn)

233
 8.1. POINTWISE CONVERGENCE

converges pointwise to f on A then f is not necessarily continuous on

A.

Example q 8.1.6. Consider the sequence (fn) defined on A = [−1, 1] by

2
fn (x) = nxn+1 . For fixed x ∈ A we have
r
nx2 + 1
lim fn(x) = lim
r n
n→∞ n→∞

1
= lim x2 +
n→∞
√ n
= x2
= |x|.

Hence, (fn) converges pointwise on A to the function f (x) = |x|. Notice

that each function fn is continuous on A and the pointwise limit f is
also continuous. After some basic calculations we find that
x
fn′ (x) = q
nx2 +1
n

and fn′ (x) exists for each x ∈ [−1, 1], in other words, fn is differentiable
on A. However, f (x) = |x| is not differentiable on A since f does not
have a derivative at x = 0. In Figure 8.3, we graph fn for various values
of n.

Example 8.1.6 illustrates another weakness of pointwise conver-

gence, namely, that if (fn) is a sequence of differentiable functions on
A and (fn) converges pointwise to f on A then f is not necessarily
differentiable on A.

Example 8.1.7. Consider the sequence (fn) on A = [0, 1] defined by

2
fn (x) = 2nxe−nx . For fixed x ∈ [0, 1] we find (using l’Hôpital’s rule)

234
 8.1. POINTWISE CONVERGENCE

1.4

1.2

1.0

0.8

0.6

0.4

f1

f3
0.2
f8

f12

f25
0.0 f
−1.00 −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00

q
nx2 +1
Figure 8.3: Graph of fn(x) = n
for n = 1, 3, 8, 12, 25 and f (x) =
|x|

that
2nx
lim fn (x) = lim 2 = 0.
n→∞ n→∞ enx

Hence, (fn ) converges pointwise to f (x) = 0 on A. Consider

Z 1 Z 1
2
fn (x) dx = 2nxe−nx dx
0 0
2 1
= −e−nx
0
−n
=1−e

and therefore
Z 1
lim fn(x) = lim (1 − e−n ) = 1.
n→∞ 0 n→∞

235
 8.1. POINTWISE CONVERGENCE

R1
On the other hand, 0 f (x) dx = 0. Therefore,
Z 1 Z 1
f (x) dx 6= lim fn (x) dx
0 n→∞ 0

or another way to write this is

Z 1 Z 1
lim fn (x) dx 6= lim fn (x) dx.
n→∞ 0 0 n→∞

Examples 8.1.5-8.1.6 illustrate that the pointwise limit f of a se-

quence of functions (fn) does not always inherit the properties of con-
tinuity and/or differentiability, and Example 8.1.7 illustrates that un-
expected (or surprising) results can be obtained when combining the
operations of integration and limits, and in particular, one cannot in
general interchange the limit operation with integration.

236
 8.1. POINTWISE CONVERGENCE

Exercises

Exercise 8.1.1. Suppose that fn : [a, b] → R is a sequence of func-

tions such that fn is increasing for each n ∈ N. Suppose that f (x) =
limn→∞ fn(x) exists for each x ∈ [a, b]. Is f an increasing function?

Exercise 8.1.2. Let (an ) be a sequence of positive numbers and define

fn : [0, 1] → R as

2nan x,
 0 ≤ x ≤ 1/(2n),
fn (x) = 2an − 2nan x, 1/(2n) ≤ x ≤ 1/n,


0, 1/n ≤ x ≤ 1.

(a) Find the pointwise limit f : [0, 1] → R of the sequence (fn ).

R1
(b) Find 0 f (x) dx.
(c) If an = 4n, find Z 1
lim fn (x) dx.
n→∞ 0

Exercise 8.1.3. Recall that Q is countable and thus there exists a

bijection r : N → Q. Define the sequence (rn) by letting rn = r(n).
Now define fn : R → R as
(
1, x ∈ {r1, r2, . . . , rn}
fn (x) =
0, otherwise.

(a) Find the pointwise limit f : R → R of the sequence (fn).

(b) Is fn Riemann integrable? Explain.
(c) Is f Riemann integrable? Explain.

237
 8.2. UNIFORM CONVERGENCE

8.2 Uniform Convergence

In the previous section we saw that pointwise convergence is a rather
weak form of convergence since the limiting function will not in general
inherit any of the properties possessed by the terms of the sequence.
Examining the concept of pointwise convergence one observes that it is
a very localized definition of convergence of a sequence of functions; all
that is asked for is that (fn(x)) converge for each x ∈ A. This allows the
possibility that the “speed” of convergence of (fn(x)) may differ wildly
as x varies in A. For example, for the sequence of functions fn(x) = xn
and x ∈ (0, 1), convergence of (fn(x)) to zero is much faster for values
of x near 0 than for values of x near 1. What is worse, as x → 1
convergence of (fn(x)) to zero is arbitrarily slow. Specifically, recall in
Example 8.1.5 that |xK − 0| < ε if and only if K > ln(ε)/ ln(x). Thus,
for a fixed ε > 0, as x → 1 we have K → ∞. Hence, there is no single
K that will work for all values of x ∈ (0, 1), that is, the convergence is
not uniform.

Definition 8.2.1: Uniform Convergence

Let (fn) be a sequence of functions on A ⊆ R. We say that (fn)
converges uniformly on A to the function f : A → R if for any
ε > 0 there exists K ∈ N such that if n ≥ K then |fn (x) − f (x)| < ε
for all x ∈ A.

Notice that in Definition 8.2.1, the K only depends on the given

(but fixed) ε > 0 and the inequality |fn (x) − f (x)| < ε holds for all
x ∈ A provided n ≥ K. The inequality |fn (x) − f (x)| < ε for all x ∈ A
is equivalent to
f (x) − ε < fn (x) < f (x) + ε

238
 8.2. UNIFORM CONVERGENCE

for all x ∈ A and can therefore be interpreted as saying that the graph
of fn lies in the tube of radius ε > 0 and centered along the graph of
f , see Figure 8.4.

4
fn

3
f(x) + ε

2
f(x) − ε

A
1

0.0 0.5 1.0 1.5 2.0 2.5 3.0

−1

Figure 8.4: ε-tubular neighborhood along the graph of f ; if |fn (x) −

f (x)| < ε for all x ∈ A then the graph of fn is within the ε-tubular
neighborhood of f

The following result is a direct consequence of the definitions but it

is worth stating anyhow.

Proposition 8.2.2
If (fn) converges uniformly to f then (fn) converges pointwise to f .

Example 8.2.3. Let A = [−5, 5] and let (fn) be the sequence of func-
tions on A defined by fn (x) = (2xn + (−1)nx2)/n. Prove that (fn)
converges uniformly to f (x) = 2x.

239
 8.2. UNIFORM CONVERGENCE

Solution. We compute that

2xn + (−1)nx2
lim fn(x) = lim = 2x
n→∞ n→∞ n

and thus (fn) converges pointwise to f (x) = 2x on A. To prove that

the convergence is uniform, consider

2xn + (−1)nx2
|fn (x) − f (x)| = − 2x
n
(−1)nx2
=
n
2
|x|
=
n
52
≤ .
n

52
For any given ε > 0 if K ∈ N is such that K < ε then if n ≥ K then
for any x ∈ A we have

52
|fn (x) − f (x)| ≤
n
52
≤
K
< ε.

This proves that (fn ) converges uniformly to f (x) = 2x on A = [−5, 5].

Note that a similar argument will not hold if we take A = R.

√
Example 8.2.4. Show that the sequence of functions fn (x) = sin(nx)/ n
converges uniformly to f (x) = 0 on R.

240
 8.2. UNIFORM CONVERGENCE

Solution. We compute
sin(nx)
|fn (x) − f (x)| = √
n
| sin(nx)|
= √
n
1
≤√
n

and therefore if K ∈ N is such that √1K < ε then if n ≥ K then

|fn (x) − 0| < ε for all x ∈ R. Hence, (fn) converges uniformly to f = 0
on R.

On a close examination of the previous examples on uniform con-

vergence, one observes that in proving that (fn) converges uniformly to
f on A, we used an inequality of the form:

|fn(x) − f (x)| ≤ Mn , ∀x ∈ A

for some sequence (Mn ) of non-negative numbers such that limn→∞ Mn =

0. It follows that
sup |fn(x) − f (x)| ≤ Mn .
x∈A
This observation is worth formalizing.

Theorem 8.2.5
Let fn : A → R be a sequence of functions. Then (fn) converges
uniformly to f on A if and only if there exists a sequence (Mn) of
non-negative numbers converging to zero such that supx∈A |fn (x) −
f (x)| ≤ Mn for n sufficiently large.

Proof. Suppose that (fn) converges uniformly on A to f . There exists

N ∈ N such that |fn(x) − f (x)| < 1 for all n ≥ N and x ∈ A. Hence,

241
 8.2. UNIFORM CONVERGENCE

Mn = supx∈A |fn (x) − f (x)| ≥ 0 is well-defined for all n ≥ N . Define

Mn ≥ 0 arbitrarily for 1 ≤ n ≤ N − 1. Given an arbitrary ε > 0,
there exists K ∈ N such that if n ≥ K then |fn (x) − f (x)| < ε for
all x ∈ A. We can assume that K ≥ N . Therefore, if n ≥ K then
Mn = supx∈A |fn (x) − f (x)| ≤ ε. This prove that limn→∞ Mn = 0.
Conversely, suppose that there exists Mn ≥ 0 such that limn→∞ Mn =
0 and supx∈A |fn (x) − f (x)| ≤ Mn for all n ≥ N . Let ε > 0 be arbitrary.
Then there exists K ∈ N such that if n ≥ K then Mn < ε. Hence, if
n ≥ K ≥ N then supx∈A |fn (x) − f (x)| ≤ Mn < ε. This implies that if
n ≥ K then |fn (x) − f (x)| < ε for all x ∈ A, and thus (fn) converges
uniformly to f on A.

Example 8.2.6. Let f be a continuous function on [a, b]. Prove that

there exists a sequence of step functions (sn ) on [a, b] that converges
uniformly to f on [a, b].

We end this section by stating and proving a Cauchy criterion for

uniform convergence.

Theorem 8.2.7: Cauchy Criterion for Uniform Convergence

The sequence (fn) converges uniformly on A if and only if for every
ε > 0 there exists K ∈ N such that if n, m ≥ K then |fm (x) −
fn (x)| < ε for all x ∈ A.

Proof. Suppose that (fn) → f uniformly on A and let ε > 0. There

exists K ∈ N such that if n ≥ K then |fn (x) − f (x)| < ε/2 for all

242
 8.2. UNIFORM CONVERGENCE

x ∈ A. Therefore, if n, m ≥ K then for all x ∈ A we have

|fn (x) − fm (x)| = |fn (x) − f (x) + f (x) − fm (x)|

≤ |fn (x) − f (x)| + |f (x) − fm(x)|
< ε/2 + ε/2
= ε.

To prove the converse, suppose that for every ε > 0 there exists K ∈ N
such that if n, m ≥ K then |fm (x)−fn(x)| < ε for all x ∈ A. Therefore,
for each x ∈ A the sequence (fn(x)) is a Cauchy sequence and therefore
converges. Let f : A → R be defined by f (x) = limn→∞ fn(x). If
ε > 0 let K ∈ N be such that |fm(x) − fn (x)| < ε for all x ∈ A and
n, m ≥ K. Fix m ≥ K and consider the sequence zn = |fm (x) − fn(x)|
and thus zn < ε. Now since limn→∞ fn (x) = f (x) then lim zn exists and
lim zn ≤ ε, that is,

lim zn = lim |fm (x) − fn (x)|

n→∞ n→∞
= |fm (x) − f (x)|
≤ ε.

Therefore, if m ≥ K then |fm (x) − f (x)| ≤ ε for all x ∈ A.

243
 8.2. UNIFORM CONVERGENCE

Exercises

Exercise 8.2.1. Let fn : A → R be a sequence of functions converging

uniformly to f : A → R. Let g : A → R be a function and let gn = gfn
for each n ∈ N. Under what condition on g does the sequence (gn)
converge uniformly? Prove it. What is the uniform limit of (gn)?

Exercise 8.2.2. Prove that if (fn) converges uniformly to f on A and

(gn ) converges uniformly to g on A then (fn + gn ) converges uniformly
to f + g on A.

Exercise 8.2.3. Let fn : [0, 1] → be the sequence defined in Exer-

cise 8.1.2. Show that if limn→∞ an = 0 then (fn ) converges uniformly.
√
Exercise 8.2.4. Let fn (x) = sin(nx)/ n for x ∈ R. Prove that (fn)
converges uniformly on R.

244
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

8.3 Properties of Uniform Convergence

A sequence (fn) on A is said to be uniformly bounded on A if there
exists a constant M > 0 such that |fn (x)| < M for all x ∈ A and for
all n ∈ N.

Theorem 8.3.1: Uniform Boundedness

Suppose that (fn ) → f uniformly on A. If each fn is bounded on A
then the sequence (fn) is uniformly bounded on A and f is bounded
on A.

Proof. By definition, there exists K ∈ N such that

|f (x)| ≤ |fn (x) − f (x)| + |fn (x)|

< 1 + |fn (x)|

for all x ∈ A and all n ≥ K. Since fK is bounded, then |f (x)| ≤

1 + maxx∈A |fK (x)| for all x ∈ A and thus f is bounded on A with
upper bound M ′ = 1 + maxx∈A |fK (x)|. Therefore, |fn (x)| ≤ |fn −
f (x)| + |f (x)| < 1 + M ′ for all n ≥ K and all x ∈ A. Let Mn
be an upper bounded for fn on A for each n ∈ N. Then if M =
max{M1 , . . . , MK−1, 1 + M ′ } then |fn(x)| < M for all x ∈ A and all
n ∈ N.

Example 8.3.2. Give an example of a set A and a sequence of functions

(fn) on A such that fn is bounded for each n ∈ N, (fn ) converges
pointwise to f but (fn) is not uniformly bounded on A.

Unlike the case with pointwise convergence, a sequence of contin-

uous functions converging uniformly does so to a continuous function.

245
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

Theorem 8.3.3: Uniform Convergence and Continuity

Let (fn ) be a sequence of functions on A converging uniformly to f
on A. If each fn is continuous on A then f is continuous on A.

Proof. To prove that f is continuous on A we must show that f is

continuous at each c ∈ A. Let ε > 0 be arbitrary. Recall that to prove
that f is continuous at c we must show there exists δ > 0 such that if
|x − c| < δ then |f (x) − f (c)| < ε. Consider the following:

|f (x) − f (c)| = |f (x) − fn(x) + fn (x) − fn(c) + fn (c) − f (c)|

≤ |f (x) − fn(x)| + |fn (x) − fn (c)| + |fn (c) − f (c)|.

Since (fn) → f uniformly on A, there exists K ∈ N such that |f (x) −

fK (x)| < ε/3 for all x ∈ A. Moreover, since fK is continuous there
exists δ > 0 such that if |x − c| < δ then |fK (x) − fK (c)| < ε/3.
Therefore, if |x − c| < δ then

|f (x) − f (c)| ≤ |f (x) − fK (x)| + |fK (x) − fK (c)| + |fK (c) − f (c)|

< ε/3 + ε/3 + ε/3

= ε.

This proves that f is continuous at c ∈ A.

A direct consequence of Theorem 8.3.3 is that if (fn ) → f pointwise and

each fn is continuous then if f is discontinuous then the convergence
cannot be uniform.
x
Example 8.3.4. Let gn (x) = q for x ∈ [−1, 1] and n ∈ N.
2 1
x +n
Each function gn is clearly continuous. Now gn (0) = 0 and thus

246
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

limn→∞ gn (0) = 0. If x 6= 0 then

x
lim gn (x) = lim q
n→∞ n→∞ 1
x2 + n
x
=√
x2
x
=
|x|
(
1, x>0
=
−1, x < 0,

Therefore, (gn) converges pointwise to the function


−1,
 −1 ≤ x < 0
g(x) = 0, x=0

1, 0 < x ≤ 1.

The function g is discontinuous and therefore, by Theorem 8.3.3, (gn)

does not converge uniformly to g.

The next property that we can deduce from uniform convergence is

that the limit and integration operations can be interchanged. Recall
from Example 8.1.7 that if (fn) → f pointwise then it is not necessarily
true that Z Z
lim fn = f
n→∞ A A

Since limn→∞ fn (x) = f (x), then it in general we can say that

Z Z
lim fn 6= lim fn
n→∞ A A n→∞

However, when the convergence is uniform we can indeed interchange

the limit and integration operations.

247
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

Theorem 8.3.5: Uniform Convergence and Integration

Let (fn ) be a sequence of Riemann integrable functions on [a, b]. If
(fn) converges uniformly to f on [a, b] then f ∈ R[a, b] and
Z b Z b
lim fn = f.
n→∞ a a

Proof. Let ε > 0 be arbitrary. By uniform convergence, there exists

K ∈ N such that if n ≥ K then for all x ∈ [a, b] we have
ε
|fn (x) − f (x)| <
4(b − a)
or
ε ε
fn (x) − < f (x) < fn (x) + .
4(b − a) 4(b − a)
ε
By assumption, fn ± 4(b−a) is Riemann integrable and thus if n ≥ N
then Z b
ε
[(fn + ε/4(b − a)) − (fn − ε/4(b − a))] = < ε.
a 2
By the Squeeze Theorem of Riemann integration (Theorem 7.2.3), f is
Riemann integrable. Moreover, if n ≥ N then
ε ε
− < fn(x) − f (x) <
4(b − a) 4(b − a)
implies (by monotonicity of integration)
Z b Z b
ε ε
− < fn − f<
4 a a 4
and thus Z Z
b b
ε
fn − f < .
a a 4
Rb Rb
This proves that the sequence a fn converges to a f.

The following corollary to Theorem 8.3.5 is worth noting.

248
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

Corollary 8.3.6: Uniform Convergence and Integration

Let (fn) be a sequence of continuous functions on the interval [a, b].
If (fn) converges uniformly to f then f ∈ R[a, b] and
Z b Z b
lim fn = f.
n→∞ a a

Proof. If each fn is continuous then fn ∈ R[a, b] and Theorem 8.3.5

applies.

Example 8.3.7. Consider the sequence of functions (fn) defined on

[0, 1] given by

2 1
(n + 1) x,
 0 ≤ x ≤ n+1
2

1 2
fn (x) = −(n + 1)2 x − n+1 , n+1 ≤ x ≤ n+1

0, 2
n+1
< x ≤ 1.

(a) Draw a typical function fn .

(b) Prove that (fn ) converges pointwise.

(c) Use Theorem 8.3.5 to show that the convergence is not uniform.

We now consider how the operation of differentiation behaves under

uniform convergence. One would hope, based on the results of Theo-
rem 8.3.3, that if (fn ) → f uniformly and each fn is differentiable then
f ′ is also differentiable and maybe even that (fn′ ) → f ′ at least point-
wise and maybe even uniformly. Unfortunately, the property of dif-
ferentiability is not generally inherited under uniform convergence. An
p
example of this occurred in Example 8.1.6 where fn(x) = (nx2 + 1)/n
and (fn) → f where f (x) = |x| for x ∈ [−1, 1]. The convergence in

249
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

this case is uniform on [−1, 1] but although each fn is differentiable the

limit function f (x) = |x| is not. It turns out that the main assumption
needed for all to be well is that the sequence (fn′ ) converge uniformly.

Theorem 8.3.8: Uniform Convergence and Differentiation

Let (fn) be a sequence of differentiable functions on [a, b]. Assume
that fn′ is Riemann integrable on [a, b] for each n ∈ N and suppose
that (fn′ ) converges uniformly to g on [a, b]. Suppose there exists
x0 ∈ [a, b] such that (fn(x0)) converges. Then the sequence (fn)
converges uniformly on [a, b] to a differentiable function f and f ′ =
g.

Proof. Let x ∈ [a, b] be arbitrary but with x 6= x0. By the Mean Value
theorem applied to the differentiable function fm − fn , there exists y in
between x and x0 such that
(fm(x) − fn (x)) − (fm(x0) − fn(x0)) ′
= fm (y) − fn′ (y)
x − x0
or equivalently
′
fm(x) − fn (x) = fm (x0) − fn (x0) + (x − x0)(fm (y) − fn′ (y))

Therefore,
′
|fm (x) − fn(x)| ≤ |fm (x0) − fn (x0)| + (b − a)|fm (y) − fn′ (y)|.

Since (fn(x0)) converges and (fn′ ) is uniformly convergent, by the Cauchy

criterion, for any ε > 0 there exists K ∈ N such that if n, m ≥ K then
′
|fm (x0) − fn (x0)| < ε/2 and |fm (y) − fn′ (y)| < (ε/2)/(b − a) for all
y ∈ [a, b]. Therefore, if m, n ≥ K then
′
|fm (x) − fn (x)| ≤ |fm (x0) − fn (x0)| + (b − a)|fm (y) − fn′ (y)|
<ε

250
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

and this holds for all x ∈ [a, b]. By the Cauchy criterion for uniform
convergence, (fn) converges uniformly. Let f be the uniform limit of
(fn). We now prove that f is differentiable and f ′ = g. By the Funda-
mental theorem of Calculus (FTC), we have that
Z x
fn (x) = fn(a) + fn′ (t) dt
a

for each x ∈ [a, b]. Since (fn ) converges to f and (fn′ ) converges uni-
formly to g we have

f (x) = lim fn (x)

n→∞
 Z x 
= lim fn (a) + fn′ (t) dt
n→∞ a
Z x
= lim fn (a) + lim fn′ (t) dt
n→∞ n→∞ a
Z x
= f (a) + g(t) dt.
a
Rx
Thus f (x) = f (a) + a g(t) dt and by the FTC we obtain f ′ (x) =
g(x).

Notice that in the statement of Theorem 8.3.8, all that is required

is that (fn (x0)) converge for one x0 ∈ [a, b]. The assumption that
(fn′ ) converges uniformly then guarantees that in fact (fn) converges
uniformly.

Example 8.3.9. Consider the sequence (fn) defined on [−1, 1] by fn (x) =

(2xn + (−1)nx2 )/n. We compute that fn′ (x) = (2n + 2(−1)nx)/n and
clearly fn′ is continuous on [−1, 1] for each n ∈ N. Now limn→∞ fn′ (x) =
2 for all x and thus (fn′ ) converges pointwise to g(x) = 2. To prove that

251
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

the convergence is uniform we note that

x
|fn′ (x) − g(x)| = |2 + (−1)n − 2|
n
|x|
=
n
1
≤ .
n
Therefore, (fn′ ) converges uniformly to g on [−1, 1]. Now fn(0) = 0
and thus (fn(0)) converges to 0. By Theorem 8.3.8, (fn) converges
uniformly to say f with f (0) = 0 and f ′ = g. Now by the FTC,
R
f (x) = g(x) dx + C = 2x + C and since f (0) = 0 then f (x) = 2x.

252
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

Exercises

Exercise 8.3.1. Give an example of a set A and a sequence of functions

(fn) on A such that fn is bounded for each n ∈ N, (fn ) converges
pointwise to f but (fn) is not uniformly bounded on A.

Exercise 8.3.2. Suppose that (fn) → f uniformly on A and (gn) → g

uniformly on A. Prove that if (fn) and (gn) are uniformly bounded on
A then (fn gn ) converges uniformly to f g on A. Then give an example
to show that if one of (fn) or (gn ) is not uniformly bounded then the
result is false.

Exercise 8.3.3. Let

nx2
fn (x) =
1 + nx2
for x ∈ R and n ∈ N.
(a) Show that (fn) converges pointwise on R.
(b) Show that (fn) does not converge uniformly on any closed interval
containing 0.
(c) Show that (fn) converges uniformly on any closed interval not con-
taining 0. For instance, take [a, b] with 0 < a < b.

Exercise 8.3.4. Suppose that f : R → R has that property that

|f (x) − f (y)| ≤ K|x − y| for all x, y, ∈ R and some K > 0. Prove that if
(gn ) converges uniformly on R to g then the sequence (f ◦ gn ) converges
uniformly to f ◦ g on R. Note: f ◦ gn and f ◦ g are compositions of
functions and not function multiplication.

Exercise 8.3.5. Let fn (x) = nx/(nx + 1) for n ∈ N and x ∈ [a, 1]

where 0 < a < 1.
(a) Prove directly that the sequence (fn ) is uniformly Cauchy.
R1
(b) If f is the uniform limit of (fn ), find a f without computing f .

253
 8.3. PROPERTIES OF UNIFORM CONVERGENCE

Exercise 8.3.6. Consider the sequence of functions (fn) on A = [0, ∞)

defined as follows:
(
1/n, 0 ≤ x ≤ n2,
fn (x) =
0, x > n2 .

(a) Prove that (fn) converges uniformly to f = 0 on A.

(b) For each fixed n ∈ N, find the improper integral
Z ∞
fn
0

and show that Z ∞

lim fn = ∞.
n→∞ 0
(c) The results above seem to contradict Theorem 8.3.5. Explain why
there is no contradiction.

254
 8.4. INFINITE SERIES OF FUNCTIONS

8.4 Infinite Series of Functions

In this section, we consider series whose terms are functions. You have
already encountered such objects when studying power series in Calcu-
lus. An example of an infinite series of functions (more specifically a
power series) is
∞
X (−1)nxn
.
n=0
(2n)!
n n
In this case, if we set fn (x) = (−1) x
(2n)! then the above infinite series is
P∞
n=0 fn (x). Let us give the general definition.

Definition 8.4.1: Infinite Series of Functions

Let A be a non-empty subset of R. An infinite series of functions
P
on A is a series of the form ∞ n=1 fn (x) for each x ∈ A where (fn )
is a sequence of functions on A. The sequence of partial sums
P
generated by the series fn is the sequence of functions (sn ) on A
defined as sn (x) = f1(x) + · · · + fn (x) for each x ∈ A.

P
Recall that a series of numbers xn converges if the sequence of partial
sums (tn), defined as tn = x1 + x2 + · · · + xn, converges. Hence, conver-
P
gence of an infinite series of functions fn is treated by considering
the convergence of the sequence of partial sums (sn ) (which are func-
P
tions). For example, to say that the series fn converges uniformly to
a function f we mean that the sequence of partial sums (sn ) converges
uniformly to f , etc. It is now clear that our previous work in Sections
8.1-8.3 translate essentially directly to infinite series of functions. As
an example:

255
 8.4. INFINITE SERIES OF FUNCTIONS

Theorem 8.4.2
P
Let (fn) be a sequence of functions on A and suppose that fn
converges uniformly to f . If each fn is continuous on A then f is
continuous on A.
P
Proof. By assumption, the sequence of functions sn (x) = nk=1 fk (x)
for x ∈ A converges uniformly to f . Since each function fn is continu-
ous, and the sum of continuous functions is continuous, it follows that
sn is continuous. The result now follows by Theorem 8.3.3.

The following translate of Theorem 8.3.5 is worth explicitly writing out.

Theorem 8.4.3: Term-by-Term Integration

P
Let (fn) be a sequence of functions on [a, b] and suppose that fn
converges uniformly to f . If each fn is Riemann integrable on [a, b]
then f ∈ R[a, b] and
Z b X ∞
! ∞ Z b
X
fn (x) dx = fn (x) dx.
a n=1 n=1 a

Proof. By assumption, the sequence (sn ) defined as sn (x) = f1(x)+· · ·+

fn (x) converges uniformly to f . Since each fn is Riemann integrable
P
then sn is Riemann integrable and therefore f = lim sn = fn is
Riemann integrable by Theorem 8.3.5. Also by Theorem 8.3.5, we have
Z b Z b
f = lim sn
a n→∞ a

or written another way is

Z bX ∞ n
Z bX
fn = lim fk
a n=1 n→∞ a k=1

256
 8.4. INFINITE SERIES OF FUNCTIONS

or
∞
Z bX n Z
X b
fn = lim fk
a n=1 n→∞ a
k=1
or
∞
Z bX ∞ Z
X b
fn = fn
a n=1 n=1 a

We now state the derivative theorem (similar to Theorem 8.3.8) for

infinite series of functions.

Theorem 8.4.4: Term-by-Term Differentiation

Let (fn) be a sequence of differentiable functions on [a, b] and sup-
P
pose that fn converges at some point x0 ∈ [a, b]. Assume further
P ′
that fn converges uniformly on [a, b] and each fn′ is continuous.
P
Then fn converges uniformly to some differentiable function f on
P
[a, b] and f ′ = fn′ .

We now state a useful theorem for uniform convergence of infinite

series of functions.

Theorem 8.4.5: Weierstrass M-Test

Let (fn) be a sequence of functions on A and suppose that there
exists a sequence of non-negative numbers (Mn ) such that |fn (x)| ≤
P P
Mn for all x ∈ A, and all n ∈ N. If Mn converges then fn
converges uniformly on A.

P
Proof. Let ε > 0 be arbitrary. Let tn = nk=1 Mk be the sequence of
P
partial sums of the series Mn . By assumption, (tn ) converges and
thus (tn ) is a Cauchy sequence. Hence, there exists K ∈ N such that

257
 8.4. INFINITE SERIES OF FUNCTIONS

|tm − tn | < ε for all m > n ≥ K. Let (sn ) be the sequence of partial
P
sums of fn . Then if m > n ≥ K then for all x ∈ A we have

|sm (x) − sn (x)| = |fm (x) + fm−1(x) + · · · + fn+1(x)|

≤ |fm(x)| + |fm−1(x)| + · · · + |fn+1(x)|

≤ Mm + Mm−1 + · · · + Mn+1

= |tm − tn |
< ε.

Hence, the sequence (sn ) satisfies the Cauchy Criterion for uniform
convergence (Theorem 8.2.7) and the proof is complete.

Example 8.4.6. Prove that

Z ∞
!
π X n sin(nx) 2e
=
0 n=1
en e2 − 1

Proof. For any x ∈ R it holds that

n sin(nx) n
≤ .
en en

P
A straightforward application of the Ratio test shows that ∞ n
n=1 en is
a convergent series. Hence, by the M-Test, the given series converges

258
 8.4. INFINITE SERIES OF FUNCTIONS

uniformly on A = R, and in particular on [0, π]. By Theorem 8.4.3,

Z πX∞ ∞ Z π
n sin(nx) X n sin(nx)
dx = dx
0 n=1 en n=1 0 en

∞
X cos(nx) π
= −
n=1
en 0

∞
X  1
n

−1 n
= e − e
n=1
   
1 1
= 1−1/e −1 − 1+1/e −1
2e
=
e2 − 1

Example 8.4.7 (Riemann (1853)). Consider the function r(x) whose

graph is given in Figure 8.5; one can write down an explicit expression
for r(x) but the details are unimportant. Consider the series
∞
X r(nx)
.
n=1
n2

Since
r(nx) 1/2
≤
n2 n2
P
and ∞ 1
n=1 2n2 converges, then by the M-test the above series converges
uniformly on any interval [a, b]. Let f be the function defined by the
series on [a, b]. Now, on [a, b], the function fn (x) = r(nx)
n2 has only a finite
number of discontinuities and thus fn is Riemann integrable. Therefore,
by Theorem 8.3.5, the function f is Riemann integrable. The graph of
f is shown in Figure 8.6. One can show that f has discontinuities at
the rational points x = 2qp where gcd(p, q) = 1.

259
 8.4. INFINITE SERIES OF FUNCTIONS

0.6

0.4

0.2

0.0
−3 −2 −1 0 1 2 3

−0.2

−0.4

−0.6

Figure 8.5: The function r(x)

0.6

0.4

0.2

0.0
0.0 0.5 1.0 1.5 2.0 2.5

−0.2

−0.4

−0.6

P∞ r(nx)
Figure 8.6: The function f (x) = n=1 n2

Example 8.4.8 (Power Series). Recall that a power series is a series

of the form
X∞
cn (x − a)n
n=0

where cn ∈ R and a ∈ R. Hence, in this case if we write the series as

P
fn (x) then fn (x) = cn (x − a)n for each n ∈ N and f0(x) = c0 . In
calculus courses, the main problem you were asked to solve is to find
the interval of convergence of the given power series. The main tool is

260
 8.4. INFINITE SERIES OF FUNCTIONS

to apply the Ratio test (Theorem 3.7.23):

|cn+1||x − a|n+1 |cn+1|
lim = |x − a| lim .
n→∞ |cn ||x − a|n n→∞ |cn |

Suppose that limn→∞ |c|cn+1n |

|
exists and is non-zero and limn→∞
|cn+1 | 1
|cn | = R
(a similar argument can be done when the limit is zero). Then by
the Ratio test, the power series converges if |x − a| R1 < 1, that is, if
|x − a| < R. The number R > 0 is called the radius of convergence
and the interval (a − R, a + R) is the interval of convergence (if
limn→∞ |c|cn+1
n|
|
is zero then R > 0 can be chosen arbitrarily and the
argument that follows is applicable). Let ρ < r < R and consider the
closed interval [a − ρ, a + ρ] ⊂ (a − R, a + R). Then if x ∈ [a − ρ, a + ρ]
then

|fn (x)| = |cn ||x − a|n

n
n |x − a|
= |cn |r n
 ρrn
≤ |cn |rn .
r
Now if x = a + r ∈ (a − R, a + R) then by assumption the series
P P
cn (x − a)n = cn rn converges, and in particular the sequence |cn |rn
is bounded, say by M. Therefore,
 ρ n
|fn (x)| ≤ M .
r
P ρ
n
Since ρ/r < 1, the geometric series r converges. Therefore, by
P
the M-test, the series cn (x − a)n converges uniformly on the interval
P
[a − ρ, a + ρ]. Let f (x) = fn (x) for x ∈ [a − ρ, a + ρ]. Now consider
the series of the derivatives
∞
X ∞
X
fn′ (x) = cn n(x − a)n−1.
n=1 n=1

261
 8.4. INFINITE SERIES OF FUNCTIONS

Applying the Ratio test again we conclude that the series of the deriva-
tives converges for each x ∈ (a − R, a + R) and a similar argument
as before shows that the series of derivatives converges uniformly on
any interval [a − ρ, a + ρ] where ρ < R. It follows from the Term-
by-Term Differentiation theorem that f is differentiable and f ′ (x) =
P
cn n(x − a)n−1. By the Term-by-Term Integration theorem, we can
also integrate the series and
Z X  XZ
fn(x) dx = fn (x) dx
I I

where I ⊂ (a − R, a + R) is any closed and bounded interval.

Example 8.4.9. Consider the power series
∞ ∞
X (−1)nx2n+1 X (−1)nx2n
and .
n=0
(2n + 1)! n=0
(2n)!

(a) Prove that the series converges at every x ∈ R.

(b) Let f denote the function defined by the series on the left and let
g denote the function defined by the series on the right. Justifying
each step, show that f ′ exists and that f ′ = g.

(c) Similarly, show that g ′ exists and g ′ = −f .

Example 8.4.10. A Fourier series is a series of the form
∞
a0 X
+ (an cos(nx) + bn sin(nx))
2 n=1

where an , bn ∈ R.
(a) Suppose that for a given (an ) and (bn), the associated Fourier
series converges pointwise on [−π, π] and let f be the pointwise
limit. Prove that in fact the Fourier series converges on R. Hint:
For any y ∈ R there exists x ∈ [−π, π] such that y = x + 2π.

262
 8.4. INFINITE SERIES OF FUNCTIONS

P P
(b) Prove that if |an | and |bn | are convergent series then the
associated Fourier series converges uniformly on R.

(c) Suppose that for a given (an ) and (bn), the associated Fourier
series converges uniformly on [−π, π] and let f be the uniform
limit. Prove the following:
Z π
a0 = f (x)dx
−π
Z π
1
an = f (x) cos(nx)dx
π −π
Z π
1
bn = f (x) sin(nx)dx
π −π

You will need the following identities:

Z π
sin(nx) cos(mx)dx = 0, ∀ n, m ∈ N
−π

Z Z (
π π
π, m = n
sin(nx) sin(mx)dx = cos(nx) cos(mx)dx =
−π −π 6 n
0, m =

Example 8.4.11 (Dini). (a) By an open cover of an interval [a, b],

we mean a collection of open intervals {Iµ }µ∈X such that Iµ ∩
[a, b] 6= ∅ for each µ ∈ X and
[
[a, b] ⊂ Iµ .
µ∈X

Here X is some set, possibly uncountable. Prove that if {Iµ}µ∈X

is any open cover of [a, b] then there there exists finitely many
µ1 , µ2, . . . , µN ∈ X such that
N
[
[a, b] ⊂ I µk .
k=1

263
 8.4. INFINITE SERIES OF FUNCTIONS

(b) Let fn be continuous and suppose that fn+1(x) ≤ fn(x) for all
x ∈ [a, b] and all n ∈ N. Suppose that (fn) converges pointwise to
a continuous function f . Prove that the convergence is actually
uniform. Give an example to show that if f is not continuous
then we only have pointwise convergence.

Proof. We first prove (a). For convenience, write Iµ = (aµ , bµ ) for each
µ ∈ X and assume without loss of generality that {bµ | µ ∈ X} is
bounded above. Let b0 = a and let Iµ1 be such that b0 ∈ Iµ1 = (a1 , b1)
and

b1 = sup{bµ | b0 ∈ Iµ }.

If b1 > b then we are done because then [a, b] ⊂ Iµ1 . By induction,

having defined bk−1 ∈ [a, b], let Iµk = (ak , bk ) be such that bk−1 ∈ Iµk
and bk = sup{bµ | bk−1 ∈ Iµ }. We claim that bN > b for some N ∈ N
and thus [a, b] ⊂ ∪N
k=1 Iµk . To prove the claim, suppose that bk ≤ b for all
k ∈ N. Then the increasing sequence (bk ) converges by the Monotone
Convergence theorem, say to L = sup{b1, b2, . . .}. Since L ∈ [a, b] then
L ∈ Iµ for some µ ∈ X and thus by convergence there exists k ∈ N
such that bk ∈ Iµ = (aµ , bµ). However, by definition of bk+1 we must
have that L < bµ ≤ bk+1 which is a contradiction to the definition of L.
This completes the proof.
Now we prove (b). First of all since fm (x) ≤ fn (x) for all m ≥ n it
holds that f (x) ≤ fn (x) for all n ∈ N and all x ∈ [a, b]. Fix x̃ ∈ [a, b]
and let ε > 0. By pointwise convergence, there exists N ∈ N such that
|fn (x̃) − f (x̃)| < ε/3 for all n ≥ N . By continuity of fN and f , there
exists δx̃ > 0 such that |fN (x) − fN (x̃)| < ε/3 and |f (x) − f (x̃)| < ε/3

264
 8.4. INFINITE SERIES OF FUNCTIONS

for all x ∈ Ix̃ = (x̃ − δx̃ , x̃ + δx̃ ). Therefore, if n ≥ N then

|fn (x) − f (x)| = fn(x) − f (x)

≤ fN (x) − f (x)
≤ |fN (x) − fN (x̃)| + |fN (x̃ − f (x̃)| + |f (x̃) − f (x)|
<ε

for all x ∈ Ix̃ . Hence, (fn) converges uniformly to f on the interval

Ix̃ . It is clear that {Ix̃ }x̃∈[a,b] is an open cover of [a, b]. Therefore, by
part (a), there exists x̃1 , x̃2, . . . , x̃k such that [a, b] ⊂ Ix̃1 ∪ · · · ∪ Ix̃k .
Hence, for arbitrary ε > 0 there exists Nj ∈ N such that if n ≥ Nj then
|fn (x) − f (x)| < ε for all x ∈ Ix̃j . If N = max{N1 , N2, . . . , Nk } then
|fn (x) − f (x)| < ε for all n ≥ N and all x ∈ [a, b]. This completes the
proof. The sequence fn(x) = xn on [0, 1] satisfies fn+1(x) ≤ fn(x) for
all n and all x ∈ [0, 1], and (fn) converges to f (x) = 0 if x ∈ [0, 1) and
f (1) = 1. Since f is not continuous on [0, 1], the convergence is not
uniform.

Example 8.4.12. Let gn be continuous and suppose that gn ≥ 0 for

P
all n ∈ N. Prove that if ∞ n=1 gn converges pointwise to a continuous
function f on [a, b] then in fact the convergence is uniform.

265
 8.4. INFINITE SERIES OF FUNCTIONS

Exercises
P
Exercise 8.4.1. Show that ∞ n
n=0 x converges uniformly on [−a, a] for
every a such that 0 < a < 1. Then show that the given series does not
converge uniformly on (−1, 1). Hint: This is an important series and
you should know what function the series converges uniformly to.
P∞ P∞
Exercise 8.4.2. If n=1 |an | < ∞ prove that n=1 an sin(nx) con-
verges uniformly on R.

Exercise 8.4.3. Prove, justifying each step, that

Z 2 X ∞
!
e
ne−nx dx = 2
1 n=1
e −1

Exercise 8.4.4. For any number q ∈ R let χq : R → R be the func-

tion defined as χq (x) = 1 if x = q and χq (x) = 0 if x 6= q. Let
{q1, q2, q3, . . .} = Q be an enumeration of the rational numbers. Define
fn : R → R as
fn = χq1 + χq2 + · · · + χqn .
Find the pointwise limit f of the sequence (fn). Is the convergence
uniform? Explain.

266
 9

Metric Spaces

9.1 Metric Spaces

The main concepts of real analysis on R can be carried over to a general

set M once a notion of “distance” d(x, y) has been defined for points
x, y ∈ M. When M = R, the distance we have been using all along is
d(x, y) = |x − y|. The set R along with the distance function d(x, y) =
|x − y| is an example of a metric space.

Definition 9.1.1: Metric Space

Let M be a non-empty set. A metric on M is a function d :
M × M → [0, ∞) satisfying the following properties:

(i) d(x, y) = 0 if and only if x = y

(ii) d(x, y) = d(y, x) for all x, y ∈ M (symmetry)

(iii) d(x, y) ≤ d(x, z) + d(z, y) for all x, y, z ∈ M (triangle inequal-

ity)

A metric space is a pair (M, d) where d is a metric on M.

267
 9.1. METRIC SPACES

If the metric d is understood, then we simply refer to M as a metric

space instead of formally referring to the pair (M, d).

Example 9.1.2. The set M = R and function d(x, y) = |x − y| is a

metric space. To see this, first of all |x − y| = 0 iff x − y = 0 iff x = y.
Second of all, |x − y| = | − (y − x)| = |y − x|, and finally by the usual
triangle inequality on R we have

d(x, y) = |x − y| = |x − z + y − z| ≤ |x − z| + |z − y| = d(x, z) + d(z, y)

for all x, y, z ∈ R.

Example 9.1.3. Let B([a, b]) denote the set of bounded functions on
the interval [a, b], that is, f ∈ B([a, b]) if there exists M > 0 such that
|f (x)| ≤ M for all x ∈ [a, b]. For f, g ∈ B([a, b]) let

d(f, g) = sup |f (x) − g(x)|.

x∈[a,b]

We claim that (B([a, b]), d) is a metric space. First of all, if f, g ∈

B([a, b]) then using the triangle inequality it follows that (f − g) ∈
B([a, b]). Therefore, d(f, g) is well-defined for all f, g ∈ B([a, b]). Next,
by definition, we have that 0 ≤ d(f, g) and it is clear that d(f, g) =
d(g, f ). Lastly, for f, g, h ∈ B([a, b]) since

|f (x) − g(x)| ≤ |f (x) − h(x)| + |h(x) − g(x)|

268
 9.1. METRIC SPACES

then

d(f, g) = sup |f (x) − g(x)|

x∈[a,b]

≤ sup (|f (x) − h(x)| + |h(x) − g(x)|)

x∈[a,b]

≤ sup (|f (x) − h(x)|) + sup (|h(x) − g(x)|)

x∈[a,b] x∈[a,b]

= d(f, h) + d(h, g).

This proves that (B[a, b], d) is a metric space. It is convention to denote

the metric d(f, g) as d∞ (f, g), and we will follow this convention.

Example 9.1.4. Let M be a non-empty set. Define d(x, y) = 1 if x 6= y

and d(x, y) = 0 if x = y. It is straightforward to show that (M, d) is a
metric space. The metric d is called the discrete metric and (M, d)
is a discrete space.

Example 9.1.5. Let (M, d) be a metric space and let M ′ ⊂ M be

a non-empty subset. Let d′ be the restriction of d onto M ′ , that is,
d′ : M ′ × M ′ → [0, ∞) is defined as d′ (x, y) = d(x, y) for x, y ∈ M ′ .
Then (M ′ , d′) is a metric space. We may therefore say that M ′ is a
metric subspace of M.

Example 9.1.6. Let C([a, b]) denote the set of continuous functions
on the inteval [a, b]. Then C([a, b]) ⊂ B([a, b]) and thus (C([a, b]), d∞)
is a metric subspace of (B([a, b]), d∞).
Rb
Example 9.1.7. For f, g ∈ C([a, b]) let d(f, g) = a |f (t) − g(t)| dt.
Prove that d defines a metric on C([a, b]).

269
 9.1. METRIC SPACES

Example 9.1.8. Let Rn×n denote the set of n × n matrices with real
entries. For A, B ∈ Rn×n define

d(A, B) = max |ai,j − bi,j |.

1≤i,j≤n

It is clear that d(A, B) = d(B, A) and d(A, B) = 0 if and only if A = B.

For A, B, C ∈ Rn×n we have that

d(A, B) = max |ai,j − bi,j |

1≤i,j≤n

= max |ai,j − ci,j + ci,j − bi,j |

1≤i,j≤n

≤ max (|ai,j − ci,j | + |ci,j − bi,j |)

1≤i,j≤n

≤ max |ai,j − ci,j | + max |ci,j − bi,j |

1≤i,j≤n 1≤i,j≤n

= d(A, C) + d(C, B).

Hence, (Rn×n , d) is a metric space.

An important class of metric spaces are normed vector spaces.

270
 9.1. METRIC SPACES

Definition 9.1.9: Norms

Let V be a vector space over R (or C). A norm on V is a function
ψ : V → [0, ∞) satisfying the following properties:

(i) ψ(x) = 0 if and only if x = 0,

(ii) ψ(αx) = |α|ψ(x) for any scalar α ∈ R and any x ∈ V , and

(iii) ψ(x + y) ≤ ψ(x) + ψ(y) for all x, y ∈ V .

The number ψ(x) is called the norm of x ∈ V . A vector space V

together with a norm ψ is called a normed vector space.

Instead of using the generic letter ψ to denote a norm, it is convention to

use instead k · k. Hence, using kxk to denote the norm ψ(x), properties
(i)-(iii) are:

(i) kxk = 0 if and only if x = 0,

(ii) kαxk = |α| kxk for any scalar α ∈ R and any x ∈ V , and

(iii) kx + yk ≤ kxk + kyk for all x, y ∈ V .

Let (V, k·k) be a normed vector space and define d : V × V → [0, ∞)

by
d(x, y) = kx − yk .

It is a straightforward exercise (which you should do) to show that

(V, d) is a metric space. Hence, every normed vector space induces a
metric space.

Example 9.1.10. The real numbers V = R form a vector space over R

under the usual operations of addition and multiplication. The absolute

271
 9.1. METRIC SPACES

value function x 7→ |x| is a norm on R. The induced metric is then

(x, y) 7→ |x − y|.

Example 9.1.11. The Euclidean norm on Rn is defined as

q
kxk2 = x21 + x22 + · · · + x2n

for x = (x1, x2, . . . , xn) ∈ Rn . It can be verified that k·k2 is indeed a

norm on Rn . Hence, we define the distance between x, y ∈ Rn as
p
kx − yk2 = (x1 − y1)2 + (x2 − y2)2 + · · · + (xn − yn )2.

Notice that when n = 1, k·k2 is the absolute value function since kxk2 =
√
x2 = |x| for x ∈ R. When not specified otherwise, whenever we refer
to Rn as a normed vector space we implicitly assume that the norm is
k·k2 and simply use the notation k·k.

Example 9.1.12 (Important). The set B([a, b]) of bounded functions

forms a vector space over R with addition defined as (f + g)(x) =
f (x) + g(x) for f, g ∈ B([a, b]) and scalar multiplication defined as
(αf )(x) = αf (x) for α ∈ R and f ∈ B([a, b]). For f ∈ B([a, b]) let

kf k∞ = sup |f (x)|.
a≤x≤b

It is left as an (important) exercise to show that k·k∞ is indeed a norm

on B([a, b]). The induced metric is

d∞ (f, g) = kf − gk∞ = sup |f (x) − g(x)|.

a≤x≤b

The norm kf k∞ is called the sup-norm of f . Notice that the metric

in Example 9.1.3 is induced by the norm k·k∞ .

272
 9.1. METRIC SPACES

Example 9.1.13. Two examples of norms on C([a, b]) are

Z b
kf k1 = |f (x)| dx
a

and Z 1/2
b
2
kf k2 = |f (x)| dx .
a
These norms are important in the analysis of Fourier series.

Example 9.1.14. For A ∈ Rn×n let

kAk∞ = max |ai,j |.

1≤i,j≤n

It is left as an exercise to show that k·k∞ defined above is a norm on

Rn×n .

Let (M, d) be a metric space. For x ∈ M and r > 0, the open ball
centered at x of radius r is by definition the set

Br (x) = {y ∈ M | d(x, y) < r}.

Example 9.1.15. Interpret geometrically the open balls in the normed

spaces (Rn , k·k) for n ∈ {1, 2, 3}.

Example 9.1.16. Give a graphical/geometric description of the open

balls in the normed space (C([a, b]), k·k∞ ).

A subset S of a metric space M is called bounded if S ⊂ Br (x) for

some x ∈ M and r > 0.

Example 9.1.17. Let (M, d) be a metric space. Prove that if S is

bounded then there exists y ∈ S and r > 0 such that S ⊂ Br (y).

273
 9.1. METRIC SPACES

Exercises

Exercise 9.1.1. Let H be the set of all real sequences x = (x1, x2, x3, . . .)
such that |xn| ≤ 1 for all n ∈ N. For x, y ∈ H let
∞
X
d(x, y) = 2−n|xn − yn |.
n=1

Prove that d is a metric on H. Note: Part of what you have to show is

P
that d(x, y) is well-defined which means to show that ∞ −n
n=1 2 |xn − yn |
converges if x, y ∈ H.

274
 9.2. SEQUENCES AND LIMITS

9.2 Sequences and Limits

Let M be a metric space. A sequence in M is a function z : N → M.
As with sequences of real numbers, we identify a sequence z : N → M
with the infinite list (zn ) = (z1, z2 , z3, . . .) where zn = z(n) for n ∈ N.

Definition 9.2.1: Convergence of Sequences

Let (M, d) be a metric space. A sequence (zn ) in M is said to
converge if there exists p ∈ M such that for any given ε > 0 there
exists K ∈ N such that d(zn, p) < ε for all n ≥ K. In this case, we
write
lim zn = p
n→∞
or (zn) → p, and we call p the limit of (zn ). If (zn ) does not converge
then we say it is divergent.

One can indeed show, just as in Theorem 3.1.12 for sequences of real
numbers, that the point p in Definition 9.2.1 is indeed unique.

Remark 9.2.2. Suppose that (zn ) converges to p and let xn = d(zn , p) ≥

0. Hence, (xn) is a sequence of real numbers. If (zn ) → p then for any
ε > 0 there exists K ∈ N such that xn < ε. Thus, xn = d(zn , p) → 0.
Conversely, if d(zn, p) → 0 then clearly (zn ) → p.

Example 9.2.3 (Important). Prove that a sequence (fn ) converges to f

in the normed vector space (B([a, b], k · k∞) if and only if (fn) converges
uniformly to f on [a, b].

Several of the results for sequences of real numbers carry over to

sequences on a general metric space. For example, a sequence (zn ) in
M is said to be bounded if the set {zn | n ∈ N} is bounded in M.
Then:

275
 9.2. SEQUENCES AND LIMITS

Lemma 9.2.4
In a metric space, a convergent sequence is bounded.

Proof. Suppose that p is the limit of (zn). There exists K ∈ N such that
d(zn , p) < 1 for all n ≥ K. Let r = 1+max{d(z1, p), . . . , d(zK−1, p)} and
we note that r ≥ 1. Then {zn | n ∈ N} ⊂ Br (p). To see this, if n ≥ K
then d(zn , p) < 1 ≤ r and thus zn ∈ Br (p). On the other hand, for zj ∈
{z1 , . . . , zK−1} we have that d(zj , p) ≤ max{d(z1, p), . . . , d(zK−1, p)} <
r, and thus zj ∈ Br (p). This proves that (zn ) is bounded.

A subsequence of a sequence (zn) is a sequence of the form (yk ) =

(znk ) where n1 < n2 < n3 < · · · . Then (compare with Theorem 3.4.5):

Lemma 9.2.5
Let M be a metric space and let (zn) be a sequence in M. If (zn) → p
then (znk ) → p for any subsequence (znk ) of (zn).

A sequence (zn ) in M is called a Cauchy sequence if for any given

ε > 0 there exists K ∈ N such that d(zn , zm) < ε for all n, m ≥ K.
Then (compare with Lemma 3.6.3-3.6.4):

276
 9.2. SEQUENCES AND LIMITS

Lemma 9.2.6
Let M be a metric space and let (zn ) be a sequence in M. The
following hold:

(i) If (zn ) is convergent then (zn ) is a Cauchy sequence.

(ii) If (zn ) is a Cauchy sequence then (zn) is bounded.

(iii) If (zn ) is a Cauchy sequence and if (zn) has a convergent sub-

sequence then (zn ) converges.

Proof. Proofs for (i) and (ii) are left as exercises (see Lemma 3.6.3-
3.6.4). To prove (iii), let (znk ) be a convergent subsequence of (zn), say
converging to p. Let ε > 0 be arbitrary. There exists K ∈ N such
that d(zn , zm ) < ε/2 for all n, m ≥ K. By convergence of (znk ) to p,
by increasing K if necessary we also have that d(znk , p) < ε/2 for all
k ≥ K. Therefore, if n ≥ K, then since nK ≥ K then

d(zn , p) ≤ d(zn , znK ) + d(znK , p)

< ε/2 + ε/2
= ε.

Hence, (zn ) → p.

The previous lemmas (that were applicable on a general metric

space) show that some properties of sequences in R are due entirely to
the metric space structure of R. There are, however, important results
on R, most notably the Bolzano-Weierstrass theorem and the Cauchy
criterion for convergence, that do not generally carry over to a general
metric space. The Bolzano-Weierstrass theorem and the Cauchy crite-
rion rely on the completeness property of R and there is no reason to

277
 9.2. SEQUENCES AND LIMITS

believe that a general metric space comes equipped with a similar com-
pleteness property. Besides, the completeness axiom of R (Axiom 2.4.6)
relies on the order property of R (i.e., ≤) and there is no reason to be-
lieve that a general metric space comes equipped with an order. We
will have more to say about this in Section 9.4. For now, however, we
will consider an important metric space where almost all the results for
sequences in R carry over (an example of a result not carrying over is
the Monotone convergence theorem), namely, the normed vector space
(Rn , k·k).
Denoting a sequence in Rn is notationally cumbersome. Formally,
a sequence in Rn is a function z : N → Rn . How then should we
denote z(k) as a vector in Rn ? One way is to simply write z(k) =
(z1 (k), z2(k), . . . , zn(k)) for each k ∈ N and this is the notation we will
adopt. It is clear that a sequence (z(k)) in Rn induces n sequences
in R, namely, (zi (k)) for each i ∈ {1, 2, . . . , n} (i.e., the component
sequences). The following theorem explains why Rn inherits almost all
the results for sequences in R.

Theorem 9.2.7: Convergence in Euclidean Spaces

Let (z(k)) = (z1 (k), z2(k), . . . , zn (k)) be a sequence in the normed
vector space (Rn , k·k). Then (z(k)) converges if and only if for
each i ∈ {1, 2, . . . , n} the component sequence (zi (k)) converges.
Moreover, if (z(k)) converges then

lim z(k) = ( lim z1 (k), lim z2 (k), . . . , lim zn (k)).

k→∞ k→∞ k→∞ k→∞

Proof. Suppose first that (z(k)) converges, say to p = (p1, p2, . . . , pn).
For any i ∈ {1, 2, . . . , n} it holds that
p
|zi (k) − pi | ≤ (z1(k) − p1 )2 + (z2 (k) − p2)2 + · · · + (zn (k) − pn )2

278
 9.2. SEQUENCES AND LIMITS

in other words, |zi (k) − pi| ≤ kz(k) − pk. Since (z(k)) → p then
limk→∞ kz(k) − pk = 0 and consequently limk→∞ |zi (k) − pi | = 0, that
is, limk→∞ zi (k) = pi .
Conversely, now suppose that (zi (k)) converges for each i ∈ {1, 2, . . . , n}.
Let pi = limk→∞ zi (k) for each i ∈ {1, 2, . . . , n} and let p = (p1, p2, . . . , pn).
By the basic limit laws of sequences in R, the sequence

xk = kz(k) − pk
p
= (z1(k) − p1 )2 + (z2 (k) − p2)2 + · · · + (zn (k) − pn )2

converges to zero since limk→∞ (zi (k) − pi)2 = 0 and the square root
√
function x 7→ x is continuous. Thus, limk→∞ z(k) = p as desired.

Corollary 9.2.8
Every Cauchy sequence in Rn is convergent.

Proof. Let (z(k)) be a Cauchy sequence in Rn . Hence, for ε > 0 there

exists K ∈ N such that kz(k) − z(m)k < ε for all k, m ≥ K. Thus, for
any i ∈ {1, 2, . . . , n}, if k, m ≥ K then

|zi (k) − zi (m)| ≤ kz(k) − z(m)k < ε.

Thus, (zi (k)) is a Cauchy sequence in R, and is therefore convergent

by the completeness property of R. By Theorem 9.2.7, this proves that
(z(k)) is convergent.

Corollary 9.2.9: Bolzano-Weierstrass in Euclidean Space

Every bounded sequence in (Rn , k·k) has a convergent subsequence.

279
 9.2. SEQUENCES AND LIMITS

Proof. Let (z(k)) be a bounded sequence in Rn . There exists x =

(x1, . . . , xn) ∈ Rn and r > 0 such that z(k) ∈ Br (x) for all k ∈ N, that
is, kz(k) − xk < r for all k ∈ N. Therefore, for any i ∈ {1, 2 . . . , n} we
have
|zi (k) − xi | ≤ kz(k) − xk < r, ∀ k ∈ N.

This proves that (zi (k)) is a bounded sequence in R for each i ∈

{1, 2, . . . , n}. We now proceed by induction. If n = 1 then (z(k)) is just
a (bounded) sequence in R and therefore, by the Bolzano-Weierstrass
theorem on R, (z(k)) has a convergent subsequence. Assume by in-
duction that for some n ≥ 1, every bounded sequence in Rn has a
convergent subsequence. Let (z(k)) be a bounded sequence in Rn+1.
Let (z̃(k)) be the sequence in Rn such that z̃(k) ∈ Rn is the vector
of the first n components of z(k) ∈ Rn+1 . Then (z̃(k)) is a bounded
sequence in Rn (why?). By induction, (z̃(k)) has a convergent subse-
quence, say it is (z̃(kj )). Now, the real sequence yj = zn+1(kj ) ∈ R is
bounded and therefore by the Bolzano-Weierstrass theorem on R, (yj )
has a convergent subsequence which we denote by (uℓ) = (yjℓ ), that
is, uℓ = zn+1(kjℓ ). Now, since wℓ = z̃(kjℓ ) is a subsequence of the con-
vergent sequence (z̃(kj )), (wℓ) converges in Rn . Thus, each component
of the sequence (z(kjℓ )) in Rn+1 is convergent and since (z(kjℓ )) is a
subsequence of the sequence (z(k))) the proof is complete.

Definition 9.2.10
Let M be a metric space.
(a) A subset U of M is said to be open if for any x ∈ U there exists
ε > 0 such that Bε(x) ⊂ U .
(b) A subset E of M is closed if E c = M\E is open.

280
 9.2. SEQUENCES AND LIMITS

Example 9.2.11. Prove that an open ball Bε(x) ⊂ M is open. In

other words, prove that for each y ∈ Bε(x) there exists δ > 0 such that
Bδ (y) ⊂ Bε (x).

Example 9.2.12. Below are some facts that are easily proved; once
(a) and (b) are proved use DeMorgan’s Laws to prove (c) and (d).
Tn
(a) If U1, . . . , Un is a finite collection of open sets then k=1 Uk is
open.
S
(b) If {Uk } is collection of open sets indexed by a set I then k∈I Uk
is open.
Sn
(c) If E1, . . . , En is a finite collection of closed sets then k=1 Ek is
closed.
T
(d) If {Ek } is collection of closed sets indexed by a set I then k∈I Ek
is closed.

Below is a characterization of closed sets via sequences.

Theorem 9.2.13: Closed Sets via Sequences

Let M be a metric space and let E ⊂ M. Then E is closed if and
only if every sequence in E that converges does so to a point in E,
that is, if (xn) → x and xn ∈ E then x ∈ E.

Proof. Suppose that E is closed and let (xn) be a sequence in E. If

x ∈ E c then there exists ε > 0 such that Bε (x) ⊂ E c . Hence, xn ∈/
Bε (x) for all n and thus (xn) does not converge to x. Hence, if (xn)
converges then it converges to a point in E. Conversely, assume that
every sequence in E that converges does so to a point in E and let
x ∈ E c be arbitrary. Then by assumption, x is not the limit point of

281
 9.2. SEQUENCES AND LIMITS

any converging sequence in E. Hence, there exists ε > 0 such that

Bε (x) ⊂ E c otherwise we can construct a sequence in E converging to
x (how?). This proves that E c is open and thus E is closed.

Example 9.2.14. Show that C([a, b]) is a closed subset of B([a, b]).

Example 9.2.15. Let M be an arbitrary non-empty set and let d be

the discrete metric, that is, d(x, y) = 1 if x 6= y and d(x, y) = 0 if
x = y. Describe the converging sequences in (M, d). Prove that every
subset of M is both open and closed.

Example 9.2.16. For a ≤ b, prove that [a, b] = {x ∈ R | a ≤ x ≤ b} is

closed.

282
 9.2. SEQUENCES AND LIMITS

Exercises

Exercise 9.2.1. Let M be a metric space and suppose that (zn ) con-
verges in M. Prove that the limit of (zn ) is unique. In other words,
prove that if p and q satisfy the convergence definition for (zn ) then
p = q.

Exercise 9.2.2. Let (M, d) be a metric space.

(a) Let y ∈ M be fixed. Prove that if (zn ) converges to p then

lim d(zn , y) = d(p, y).
n→∞

(b) Prove that if (zn ) converges to p and (yn ) converges to q then

lim d(zn , yn ) = d(p, q).
n→∞

Exercise 9.2.3. Let M be a metric space. Prove that if U1 , . . . , Un ⊂

T
M are open then nk=1 Uk is open in M.

Exercise 9.2.4. Let (M, d) be a metric space and let E ⊂ M. A point

x ∈ M is called a limit point (or cluster point) of E if there exists a
sequence (xn) in E, with xn 6= x for all n, converging to x. The closure
of E, denoted by cl(E), is the union of E and the limit points of E. If
cl(E) = M then we say that E is dense in M. As an example, Q is
dense in R since every irrational number is the limit of a sequence of
rational numbers.

(a) Prove that E is dense in M if and only if E ∩ U 6= ∅ for every

open set U of M.

(b) Let E be the set of step functions on [a, b]. Then clearly E ⊂
B([a, b]). Prove that the set of continuous function C([a, b]) is
contained in the closure of E. (Hint: See Example 8.2.6)

283
 9.2. SEQUENCES AND LIMITS

(c) Perform an internet search and find dense subsets of (C([a, b]), k·k∞ )
(you do not need to supply proofs).

Exercise 9.2.5. For x ∈ Rn define kxk∞ = max1≤i≤n |xi | and kxk1 =

Pn n
i=1 |xi |. It is not hard to verify that these are norms on R . Prove
that:

(a) kxk∞ ≤ kxk2 ≤ kxk1,

(b) kxk1 ≤ n kxk∞ , and

√
(c) kxk1 ≤ n kxk2

Two metrics d and ρ on a set M are equivalent if they generate the

same convergent sequences, in other words, (xn) converges in (M, d) if
and only if (xn) converges in (M, ρ). Prove that k·k1 , k·k2 , k·k∞ are
equivalent norms on Rn .

Exercise 9.2.6. How do we (Riemann) integrate functions from [a, b]

to Rn ? Here is how. First, we equip Rn with the standard Euclidean
norm k · k2. For any function F : [a, b] → Rn and any tagged partition
Ṗ = {([tk−1, tk ], ck )}nk=1 of [a, b], define the Riemann sum
n
X
S(F ; Ṗ) = F (ck )(xk − xk−1).
k=1

We then say that F is Riemann integrable if there exists v ∈ Rn such

that for any ε > 0 there exists δ > 0 such that for any tagged partition
Ṗ of [a, b] with kṖk < δ it holds that

kS(F ; Ṗ) − vk2 < ε.

Rb
We then also write that a F = v. If F has component functions
F = (f1, f2, . . . , fn), prove that F is Riemann integrable on [a, b] if and

284
 9.2. SEQUENCES AND LIMITS

only if the component functions f1, f2, . . . , fn are Riemann integrable

on [a, b], and in this case,
Z b Z b Z b Z b 
F = f1 , f2 , . . . , fn .
a a a a

Hint: Recall that for any x = (x1, x2, . . . , xn) ∈ Rn it holds that |xi| ≤
kxk2.

Exercise 9.2.7. Let R∞ denote the set of infinite sequences in R. It is

not hard to see that R∞ is a R-vector space with addition and scalar
multiplication defined in the obvious way. Let ℓ1 ⊂ R∞ denote the
P
subset of sequences x = (x1, x2, x3, . . .) such that ∞
n=1 |xn | converges,
that is, ℓ1 denotes the set of absolutely convergent series.

(a) Prove that ℓ1 is a subspace of R∞ , i.e., prove that ℓ1 is closed

under addition and scalar multiplication.
P∞
(b) For x ∈ ℓ1 let kxk1 = n=1 |xn |. Prove that k·k1 defines a norm
on ℓ1 .

285
 9.3. CONTINUITY

9.3 Continuity
Using the definition of continuity for a function f : R → R as a guide,
it is a straightforward task to formulate a definition of continuity for a
function f : M1 → M2 where (M1, d1) and (M2, d2) are metric spaces.

Definition 9.3.1: Continuity

Let (M1, d1) and (M2, d2) be metric spaces. A function f : M1 → M2
is continuous at x ∈ M1 if given any ε > 0 there exists δ > 0 such
that if d1(y, x) < δ then d2(f (y), f (x)) < ε. We say that f is
continuous if it is continuous at each point of M1 .

Using open balls, f : M1 → M2 is continuous at x ∈ M1 if for any given

ε > 0 there exists δ > 0 such that f (y) ∈ Bε (f (x)) whenever y ∈ Bδ (x).
We note that Bε (f (x)) is an open ball in M2 while Bδ (x) is an open
ball in M1 .
Below we characterize continuity using sequences (compare with
Theorem 5.1.2).

Theorem 9.3.2: Sequential Criterion for Continuity

Let (M1, d1) and (M2, d2) be metric spaces. A function f : M1 → M2
is continuous at x ∈ M1 if and only if for every sequence (xn) in M1
converging to x the sequence (f (xn)) in M2 converges to f (x).

Proof. Assume that f is continuous at x ∈ M1 and let (xn) be a se-

quence in M1 converging to x. Let ε > 0 be arbitrary. Then there exists
δ > 0 such that f (y) ∈ Bε (f (x)) for all y ∈ Bδ (x). Since (xn) → x,
there exists K ∈ N such that xn ∈ Bδ (x) for all n ≥ K. Therefore, for

286
 9.3. CONTINUITY

n ≥ K we have that f (xn) ∈ Bε(f (x)). Since ε > 0 is arbitrary, this

proves that (f (xn)) converges to f (x).
Suppose that f is not continuous at x. Then there exists ε∗ > 0
such that for every δ > 0 there exists y ∈ Bδ (x) with f (y) ∈
/ Bε∗ (f (x)).
Hence, if δn = n1 then there exists xn ∈ Bδn (x) such that f (xn) ∈ /
Bε∗ (f (x)). Since d(xn, x) < δn then (xn) → x. On the other hand, it is
clear that (f (xn)) does not converge to f (x). Hence, if f is not contin-
uous at x then there exists a sequence (xn) converging to x such that
(f (xn)) does not converge to f (x). This proves that if every sequence
(xn) in M1 converging to x it holds that (f (xn)) converges to f (x) then
f is continuous at x.

Example 9.3.3. A level set of a function f : M → R is a set of the

form E = {x ∈ M | f (x) = k} for some k ∈ R. Prove that if f is
continuous then the level sets of f are closed sets.

As a consequence of Theorem 9.3.2, if f is continuous at p and

limn→∞ xn = p then limn→∞ f (xn) = f (p) can be written as

lim f (xn) = f ( lim xn)

n→∞ n→∞

The sequential criteria for continuity can be conveniently used to show

that the composition of continuous function is a continuous function.

Lemma 9.3.4
Let f : M1 → M2 and let g : M2 → M3 , where M1 , M2 , and M3 are
metric spaces. If f is continuous at x ∈ M1 and g is continuous at
f (x) then the composite mapping (g ◦ f ) : M1 → M3 is continuous
at x ∈ M1 .

287
 9.3. CONTINUITY

Proof. If limn→∞ xn = p then by Theorem 9.3.2, and using the fact that
f is continuous at p, and g is continuous at g(p):

(g ◦ f )(p) = g(f (p))

= g(f ( lim xn))

n→∞

= g( lim f (xn))
n→∞

= lim g(f (xn))

n→∞

= lim (g ◦ f )(xn)
n→∞

In general, given functions f, g : M1 → M2 on metric spaces M1 and

M2 , there is no general way to define the functions f ± g or f g since M2
does not come equipped with a vector space structure nor is it equipped
with a product operation. However, when M2 = R then f (x) and g(x)
are real numbers which can therefore be added/subtracted/multiplied.

Proposition 9.3.5
Let (M, d) be a metric space and let f, g : M → R be continuous
functions, where R is equipped with the usual metric. If f and g
are continuous at x ∈ M then f + g, f − g, and f g are continuous
at x ∈ M.

Proof. In all cases, the most economical proof is to use the sequential
criterion. The details are left as an exercise.

Recall that for any function f : A → B and S ⊂ B the set f −1(S)

288
 9.3. CONTINUITY

is defined as
f −1(S) = {x ∈ A | f (x) ∈ S}.

Example 9.3.6. For any function f : A → B prove that (f −1(S))c =

f −1(S c) for any S ⊂ B.

Proposition 9.3.7: Continuity via Open and Closed Sets

For a given function f : (M1, d1) → (M2, d2) the following are equiv-
alent:

(i) f is continuous on M1 .

(ii) f −1(U ) is open in M1 for every open subset U ⊂ M2.

(iii) f −1(E) is closed in M1 for every closed subset E ⊂ M2 .

Proof. (i) =⇒ (ii): Assume that f is continuous on M1 and let U ⊂ M2

be open. Let x ∈ f −1(U ) and thus f (x) ∈ U . Since U is open, there
exists ε > 0 such that Bε (f (x)) ⊂ U . By continuity of f , there exists
δ > 0 such that if y ∈ Bδ (x) then f (y) ∈ Bε (f (x)). Therefore, Bδ (x) ⊂
f −1(U ) and this proves that f −1(U ) is open.
(ii) =⇒ (i): Let x ∈ M1 and let ε > 0 be arbitrary. Since Bε (f (x)) is
open, by assumption f −1(Bε(f (x))) is open. Clearly x ∈ f −1(Bε(f (x)))
and thus there exists δ > 0 such that Bδ (x) ⊂ f −1(Bε(f (x))), in other
words, if y ∈ Bδ (x) then f (y) ∈ Bε (f (x)). This proves that f is
continuous at x.
(ii) ⇐⇒ (iii): This follows from the fact that (f −1(U ))c = f −1(U c)
for any set U . Thus, for instance, if f −1(U ) is open for every open set
U then if E is closed then f −1(E c ) is open, that is, (f −1(E))c is open,
i.e., f −1(E) is closed.

289
 9.3. CONTINUITY

Example 9.3.8. Use Proposition 9.3.7 to prove that the level sets of
a function f : M → R on a metric space M are closed sets.

Example 9.3.9. A function f : (M1, d1) → (M2 , d2) is called Lipschitz

on M1 if there exists K > 0 such that d2(f (x), f (y)) ≤ Kd1(x, y) for
all x, y ∈ M1 . Prove that a Lipschitz function is continuous.

Example 9.3.10. For A ∈ Rn×n recall that kAk∞ = sup1≤i,j≤n |ai,j |.

Let tr : Rn×n → R be the trace function on Rn×n , that is, tr(A) =
Pn
i=1 ai,i . Show that tr is Lipschitz and therefore continuous.

Let ℓ∞ denote the set of all real sequences (xn) that are bounded,
that is, {|xn | : n ∈ N} is a bounded set. If x = (xn) ∈ ℓ∞ , it is
straightforward to verify that kxk∞ = supn∈N |xn | defines a norm on ℓ∞
with addition and scalar multiplication defined in the obvious way. Let
ℓ1 be the set of absolutely summable sequences (xn), that is, (xn) ∈ ℓ1
P
if and only if ∞ n=1 |xn | converges. It is not too hard to verify that ℓ1
P
is a normed vector space with norm defined as kxk1 = ∞ n=1 |xn |. If
P∞
n=1 |xn | converges then (|xn |) converges to zero and thus (xn ) ∈ ℓ∞ ,
thus ℓ1 ⊂ ℓ∞ .

Example 9.3.11. Fix y = (yn)∞ n=1 ∈ ℓ∞ and let h : ℓ1 → ℓ1 be defined

as h(x) = (xnyn )n=1 for x = (xn)∞
∞
n=1. Verify that h is well-defined and
prove that h is continuous.

Example 9.3.12. Let det : Rn×n → R denote the determinant func-

tion. Prove that det is continuous; you may use the formula
n
!
X Y
det(A) = sgn(σ) ai,σ(i)
σ∈Sn i=1

where Sn is the set of permutations on {1, 2, . . . , n} and sgn(σ) = ±1

is the sign of the permutation σ ∈ Sn .

290
 9.3. CONTINUITY

Exercises

Exercise 9.3.1. Let (M, d) be a metric space. Fix y ∈ M and define

the function f : M → R by f (x) = d(x, y). Prove that f is continuous.

Exercise 9.3.2. Let (V, k·k) be a normed vector space. Prove that f :
V → R defined by f (x) = kxk is continuous. Hint: kak ≤ ka − bk+kbk
for all a, b, c ∈ V .

Exercise 9.3.3. Consider C([a, b]) with norm k·k∞ . Define the functionΨ :
Rb
C([a, b]) → R by Ψ(f ) = a f (x) dx. Prove that Ψ is continuous in two
ways, using the definition and the sequential criterion for continuity.

Exercise 9.3.4. Let M be a metric space and let f : M → R be

continuous. Prove that E = {x ∈ M | f (x) = 0} is closed.

Exercise 9.3.5. Consider Rn×n as a normed vector space with norm

kAk∞ = sup1≤i,j≤n |ai,j | for A ∈ Rn×n .

(a) Let (A(k))∞ k=1 be a sequence in R

n×n
and denote the (i, j) entry
of the matrix A(k) as ai,j (k). Prove that (A(k))∞ k=1 converges to
B ∈ Rn×n if and only if for all i, j ∈ {1, 2, . . . , n} the real sequence
(ai,j (k))∞
k=1 converges to bi,j ∈ R.

(b) Given matrices X, Y ∈ Rn×n , recall that the entries of the prod-
P
uct matrix XY are (XY)i,j = nℓ=1 xi,ℓyℓ,j . Let (A(k))∞ k=1 be a
sequence in Rn×n converging to B and let (C(k))∞ k=1 be the se-
quence whose kth term is C(k) = A(k)A(k) = [A(k)]2. Prove
that (C(k))∞ 2
k=1 converges to B . Hint: By part (a), it is enough
to prove that the (i, j) component of C(k) converges to the (i, j)
component of B2 .

(c) Deduce that if C(k) = [A(k)]m where m ∈ N then the sequence

(C(k))∞ m
k=1 converges to the matrix B .

291
 9.3. CONTINUITY

(d) A polynomial matrix function is a function f : Rn×n → Rn×n

of the form

f (A) = cm Am + cm−1Am−1 + · · · + c1 A + c0 I

where cm , . . . , c0 are constants and I denotes the n × n identity

matrix. Prove that a polynomial matrix function is continuous.

Exercise 9.3.6. According to the sequential criterion for continuity, if

(zn ) and (wn) are sequences in M converging to the same point p ∈ M
and f : M → R is a function such that sequences (f (zn)) and (f (wn))
do not have the same limit f (p) (or worse one of them is divergent!)
then f is discontinuous at p. Consider f : R2 → R given by
(
2xy
x2 +y 2 , (x, y) 6= (0, 0)
f (x, y) =
0, (x, y) = (0, 0).

Show that f is discontinuous at p = (0, 0).

292
 9.4. COMPLETENESS

9.4 Completeness
Consider the space P[a, b] of polynomial functions on the interval [a, b].
Clearly, P[a, b] ⊂ C([a, b]), and thus (P[a, b], k·k∞ ) is a metric space.
P
The sequence of functions fn (x) = nk=0 k!1 xk is a sequence in P[a, b]
and it can be easily verified that (fn ) converges in the metric k·k∞ , that
is, (fn) converges uniformly in [a, b] (see Example 8.4.8). However, the
limiting function f is not an element of P[a, b] because it can be verified
that f ′(x) = f (x) and the only polynomial equal to its derivative is the
zero polynomial, however, it is clear that f (0) = limn→∞ fn (0) = 1, i.e.,
f is not the zero function (you may recognize, of course, that f (x) = ex ).
We do know, however, that f is in C([a, b]) because the uniform limit of
a sequence of continuous functions is continuous. The set P[a, b] then
suffers from the same “weakness” as do the rationals Q relative to R,
namely, there are sequences in P[a, b] that converge to elements not
in P[a, b]. On the other hand, because (fn) converges it is a Cauchy
sequence in C([a, b]) and thus also in P[a, b] (the Cauchy condition
only depends on the metric) and thus (fn) is a Cauchy sequence in
P[a, b] that does not converge to an element of P[a, b]. The following
discussion motivates the following definition.

Definition 9.4.1: Complete Metric Space

A metric space M is called complete if every Cauchy sequence in
M converges in M.

This seems like a reasonable starting definition of completeness since

in R it can be proved that the Cauchy criterion (plus the Archimedean
property) implies the Completeness property of R (Theorem 3.6.8).
Based on our characterization of closed sets via sequences, we have

293
 9.4. COMPLETENESS

the following first theorem regarding completeness.

Theorem 9.4.2
Let (M, d) be a complete metric space and let P ⊂ M. Then (P, d)
is a complete metric space if and only if P is closed.

Proof. If (zn ) is a Cauchy sequence in (P, d) then it is also a Cauchy

sequence in (M, d). Since (M, d) is complete then (zn ) converges. If P
is closed then by Theorem 9.2.13 the limit of (zn ) is in P . Hence, (P, d)
is a complete metric space.
Now suppose that (P, d) is a complete metric space and let (zn)
be a sequence in P that converges to z ∈ M. Then (zn ) is a Cauchy
sequence in M and thus also Cauchy in P . Since P is complete then
z ∈ P . Hence, by Theorem 9.2.13 P is closed.

We now consider how to formulate the Bolzano-Weierstrass (BW)

property in a general metric space. The proof in Theorem 3.6.8 can
be easily modifield to prove that the BW property, namely that ev-
ery bounded sequence in R has a convergent subsequence, implies the
completeness property of R. We therefore want to develop a BW-type
condition in a general metric space M that implies that M is complete.
Our first order of business is to develop the correct notion of bounded-
ness. We have already defined what it means for a subset E ⊂ M to
be bounded, namely, that there exists r > 0 such that E ⊂ Br (x) for
some x ∈ M. However, this notion is not enough as the next example
illustrates.

Example 9.4.3. Consider P[0, 1] with induced metric k·k∞ and let
E = {f ∈ P[0, 1] : kf k∞ < 3}, in other words, E is the open ball of
radius r = 3 centered at the zero function. Clearly, E is bounded and

294
 9.4. COMPLETENESS

P
thus any sequence in E is bounded. The sequence fn (x) = nk=0 k!1 xk
is in E, that is, kfn k∞ < 3 for all n (see Example 3.3.6). However,
as already discussed, (fn) converges in C[0, 1] but not to a point in
P[0, 1]. On the other hand, (fn) is a Cauchy sequence in P[0, 1] and
thus (fn ) cannot have a converging subsequence in P[0, 1] by part (iii)
of Lemma 9.2.6. Thus, (fn) is a bounded sequence in P [0, 1] with no
converging subsequence in P [0, 1].
The correct notion of boundedness that is needed is the following.

Definition 9.4.4
Let (M, d) be a metric space. A subset E ⊂ M is called totally
bounded if for any given ε > 0 there exists z1 , . . . , zN ∈ E such
S
that E ⊂ N i=1 Bε (zi ).

Example 9.4.5. Prove that a subset of a totally bounded set is also

totally bounded.
Example 9.4.6. A totally bounded subset E of a metric space M is
bounded. If E ⊂ Bε (x1) ∪ · · ·∪ Bε(xk ) then if r = max2≤j≤k d(x1, xj ) + ε
then if x ∈ E ∩ Bε(xj ) then d(x1, x) ≤ d(x1, xj ) + d(xj , x) < r. Hence,
E ⊂ Br (x1).
The following shows that the converse in the previous example does
not hold.
Example 9.4.7. Consider ℓ1 and let E = {e1 , e2, e3 , . . .} where ek is
the infinite sequence with entry k equal to 1 and all other entries zero.
Then kek k1 = 1 for all k ∈ N and therefore E is bounded, in particular
E ⊂ Br (0) for any r > 1. Now, kek − ej k1 = 2 for all k 6= j and thus if
ε ≤ 2 then no finite collection of open balls Bε (ek1 ), Bε(ek2 ), . . . , Bε(ekN )
can cover E. Hence, E is not totally bounded.

295
 9.4. COMPLETENESS

Example 9.4.8. Prove that a bounded subset E of R is totally bounded.

Theorem 9.4.9: Bolzano-Weierstrass

Let M be a metric space. Then M is complete if and only if every
infinite totally bounded subset of M has a limit point in M.

Proof. Suppose that (M, d) is a complete metric space. Let E be an

infinite totally bounded subset of M. Let εn = 21n for n ∈ N. For ε1
S 1
there exists z1 , . . . , zm1 ∈ E such that E ⊂ m j=1 Bε1 (zj ). Since E is
infinite, we can assume without loss of generality that E1 = E ∩ Bε1 (z1)
contains infinitely many points of E. Let then x1 = z1 . Now, E1
is totally bounded and thus there exists w1, . . . , wm2 ∈ E1 such that
S 2
E1 ⊂ m j=1 Bε2 (wj ). Since E1 is infinite, we can assume without loss
of generality that E2 = E1 ∩ Bε2 (w1) contains infinitely many points
of E1. Let x2 = w1 and therefore d(x2, x1) < ε1. Since E2 is totally
S 3
bounded there exists u1, . . . , um3 ∈ E2 such that E2 ⊂ m j=1 Bε3 (uj ). We
can assume without loss of generality that E3 = E2 ∩ Bε3 (u1) contains
infinitely many elements of E2 . Let x3 = u1 and thus d(x3, x2) < ε2. By
induction, there exists a sequence (xn) in E such that d(xn+1, xn) < 21n .
Therefore, if m > n then by the triangle inequality (and the geometric
series) we have

d(xm, xn) ≤ d(xm, xm−1) + · · · + d(xn+1, xn)

1 1
< m−1 + · · · + n
2 2
1
< n−1 .
2
Therefore, (xn) is a Cauchy sequence and since M is complete (xn)
converges in M. Thus E has a limit point in M.

296
 9.4. COMPLETENESS

Conversely, assume that every infinite totally bounded subset of M

has a limit point in M. Let (xn) be a Cauchy sequence in M and let
E = {xn | n ∈ N}. For any given ε > 0 there exists K ∈ N such that
|xn − xK | < ε for all n ≥ K. Therefore, xn ∈ Bε (xK ) for all n ≥ K
and clearly xj ∈ Bε(xj ) for all j = 1, 2, . . . , K1. Thus, E is totally
bounded. By assumption, E has a limit point, that is, there exists a
subsequence of (xn) that converges in M. By part (iii) of Lemma 9.2.6,
(xn) converges in M. Thus, M is a complete metric space.

The proof in Theorem 9.4.9 that completeness implies that every

infinite totally bounded subset has a limit point is reminiscent of the
bisection method proof that a bounded sequence in R contains a conver-
gent subsequence. Also, the proof showed the following.

Lemma 9.4.10
If E is an infinite totally bounded subset of (M, d) then E contains
a Cauchy sequence (xn) such that xn 6= xm for all n 6= m.

A complete normed vector space is usually referred to as a Banach

space in honor of Polish mathematician Stefan Banach (1892-1945)
who, in his 1920 doctorate dissertation, laid the foundations of these
spaces and their applications in integral equations. An important ex-
ample of a Banach space is the following. Let X be a non-empty set and
let B(X) be the set of bounded functions from X to R with sup-norm
kf k∞ = supx∈X |f (x)|. Then convergence in (B(X), k·k∞ ) is uniform
convergence (Example 9.2.3). We have all the tools necessary to prove
that B(X) is a Banach space.

297
 9.4. COMPLETENESS

Theorem 9.4.11
Let X be a non-empty set. The normed space (B(X), k·k∞ ) is a
Banach space.

Proof. First of all, it is clear that B(X) is a real vector space and thus
we need only show it is complete. The proof is essentially contained
in the Cauchy criterion for uniform convergence for functions on R
(Theorem 8.2.7). Let fn : X → R be a Cauchy sequence of bounded
functions. Then for any given ε > 0 there exists K ∈ N such that if
n, m ≥ K then kfn − fmk∞ < ε. In particular, for any fixed x ∈ X it
holds that
|fn (x) − fm (x)| ≤ kfn − fm k∞ < ε.
Therefore, the sequence of real numbers (fn(x)) is a Cauchy sequence
and thus f (x) = limn→∞ fn (x) exists for each x ∈ X. Now, since (fn)
is a Cauchy sequence in B(X) then (fn) is bounded in B(X). Thus,
there exists M > 0 such that kfnk∞ ≤ M for all n ≥ 1. Thus, for all
x ∈ X and n ≥ 1 it holds that

|fn (x)| ≤ kfnk∞ ≤ M

and by continuity of the absolute value function it holds that

|f (x)| = lim |fn (x)| ≤ M.

n→∞

Thus, f is a bounded function, that is, f ∈ B(X). Now, for any fixed
ε > 0, let K ∈ N be such that |fn (x) − fm (x)| < ε/2 for all x ∈ X and
n, m ≥ K. Therefore, for any x ∈ X we have that

lim |fn (x) − fm (x)| = |fn (x) − f (x)|

m→∞
≤ ε/2.

298
 9.4. COMPLETENESS

Therefore, kfn − f k∞ < ε for all n ≥ K. This proves that (fn ) con-
verges to f in (B(X), k·k∞ ).

Corollary 9.4.12
The space of continuous functions C([a, b]) on the interval [a, b] with
sup-norm is a Banach space.

Proof. A continuous function on the interval [a, b] is bounded and thus

C([a, b]) ⊂ B([a, b]). Convergence in B([a, b]) with sup-norm is uniform
convergence. A sequence of continuous functions that converges uni-
formly on [a, b] does so to a continuous function. Hence, Theorem 9.2.13
implies that C([a, b]) is a closed subset of the complete metric space
B([a, b]) and then Theorem 9.4.2 finishes the proof.

Example 9.4.13. Prove that ℓ∞ and ℓ1 are complete and hence Banach
spaces.

In a Banach space, convergence of series can be decided entirely

from the convergence of real series.

Theorem 9.4.14: Absolute Convergence Test

Let (V, k·k) be a Banach space and let (zn ) be a sequence in V . If the
P P∞
real series ∞ n=1 kzn k converges then the series n=1 zn converges
in V .

P∞
Proof. Suppose that n=1 kzn k converges, that is, suppose that the
P
sequence of partial sums tn = nk=1 kzn k converges (note that (tn ) is
increasing). Then (tn ) is a Cauchy sequence. Consider the sequence of

299
 9.4. COMPLETENESS

Pn
partial sums sn = k=1 zk . For n > m we have
n
X
ksn − sm k = zn
k=m+1
Xn
≤ kzn k
k=m+1
= tn − tm
= |tn − tm |

and since (tn ) is Cauchy then |tn − tm | can be made arbitrarily small
provided n, m are sufficiently large. This proves that (sn ) is a Cauchy
sequence in V and therefore converges since V is complete.

Remark 9.4.15. We make two remarks. The converse of Theorem 9.4.14

P
is also true, that is, if every series zn in (V, k·k) converges whenever
P
kzn k converges in R then V is a Banach space. Notice that the proof
of Theorem 9.4.14 is essentially the same as the proof of the Weierstrass
M-test.

Example 9.4.16. Consider the set of matrices Rn×n equipped with the
2-norm !1/2
Xn
kAk2 = a2i,j
i,j=1

The norm kAk2 is called the Frobenius norm or the Hilbert-Schmidt

norm.

(a) Prove that (Rn×n , k·k2) is complete.

(b) Use the Cauchy-Schwarz inequality

N
!2 N
! N !
X X X
xℓyℓ ≤ x2ℓ yℓ2
ℓ=1 ℓ=1 ℓ=1

300
 9.4. COMPLETENESS

to prove that kABk2 ≤ kAk2 kBk2. Conclude that Ak 2

≤
(kAk2 )k for all k ∈ N.

P
(c) Let f (x) = ∞ k
k=1 ck x be a power series converging on R. Define
the function f : Rn×n → Rn×n as

∞
X
f (A) = ck Ak .
k=1

Prove that f is well-defined and that if ck ≥ 0 then kf (A)k2 ≤

f (kAk2 ), that is, that

∞
X ∞
X
ck A k
≤ ck kAkk2
k=1 2 k=1

Proof. (a) The norm kAk2 is simply the standard Euclidean norm on
(RN , k·k2) with N = n2 and identifying matrices as elements of
RN . Hence, (Rn×n, k·k2 ) is complete.

(b) From the Cauchy-Schwarz inequality we have

n
!2
X
(AB)2i,j = ai,ℓbℓ,j
ℓ=1
n
! n
!
X X
≤ a2i,ℓ b2ℓ,j
ℓ=1 ℓ=1

301
 9.4. COMPLETENESS

and therefore

!1/2
X
kABk2 = (AB)2i,j
1≤i,j≤n

n
! n
!!1/2
X X X
≤ a2i,ℓ b2ℓ,j
1≤i,j≤n ℓ=1 ℓ=1

 1/2  1/2
n
X n
X
= a2i,ℓ   b2ℓ,j 
i,ℓ=1 j,ℓ=1

= kAk2 kBk2

P
(c) We first note that for any power series ∞ k
k=1 ak x that converges
P
in (−R, R), the power series ∞ k
k=1 |ak |x also converges in (−R, R).
P
The normed space (Rn×n, k·k2 ) is complete and thus ∞ k=1 ck A
k
P
converges whenever ∞ k
k=1 ck A 2 converges. Now by part (b),
P
ck Ak 2 = |ck | Ak 2 ≤ |ck | kAkk2 and since ∞ k
k=1 |ck | kAk2 con-
verges then by the comparison test for series in R, the series
P∞ k
k=1 ck A 2 converges. Therefore, f (A) is well-defined by The-
orem 9.4.14. To prove the last inequality, we note that the norm
function on a vector space is continuous and thus if ck ≥ 0 then

m
X m
X
ck A k
≤ |ck | kAkk2
k=1 2 k=1
Xm
≤ ck kAkk2
k=1

302
 9.4. COMPLETENESS

and therefore
∞
X m
X
k
ck A = lim ck Ak
m→∞
k=1 2 k=1 2
m
X
≤ lim ck kAkk2
m→∞
k=1
∞
X
= ck kAkk2 ,
k=1

in other words, kf (A)k2 ≤ f (kAk2 ).

Example 9.4.17. In view of the previous example, we can define for

A ∈ Rn×n the following:
∞
A
X 1 k
e = A
k!
k=0
∞
X (−1)n
sin(A) = A2n+1
(2n + 1)!
k=0
∞
X (−1)n
cos(A) = A2n
(2n)!
k=0
∞
X (−1)n 2n+1
arctan(A) = A
(2n + 1)
k=0

and for instance eA 2

≤ ekAk2 , etc.

303
 9.4. COMPLETENESS

Exercises

Exercise 9.4.1. Let (M1 , d1) and (M2 , d2) be metric spaces. There
are several ways to define a metric on the Cartesian product M1 × M2.
One way is to imitate what was done in R2 = R × R. We can define
d : M1 × M2 → [0, ∞) as
p
d((x, u), (y, v)) = d1 (x, y)2 + d2 (u, v)2

(a) Prove that d is a metric on M1 × M2 .

(b) Prove that ((xn, un))∞n=1 converges in M1 × M2 if and only if

(xn)∞ ∞
n=1 and (yn )n=1 converge in M1 and M2 , respectively. (Hint:
Theorem 9.2.7)

(c) Prove that M1 × M2 is complete if and only if M1 and M2 are

complete. (Hint: Corollary 9.2.8)

Exercise 9.4.2. Let M be a complete metric space and let (zn ) be a

sequence in M such that d(zn , zn+1) < rn for all n ∈ N for some fixed
0 < r < 1. Prove that (zn ) converges. (See Exercise 3.6.6.)
P
Exercise 9.4.3. Let ∞ n=1 xn be a convergent series in a normed vector
P
space (V, k·k) and suppose that ∞ n=1 kxn k converges. Show that
∞
X ∞
X
xn ≤ kxnk.
n=1 n=1

Note: The △-inequality can only be used on a finite sum. (See Exer-
cise 9.3.2.)

Exercise 9.4.4. Consider the normed space (B(X), k · k∞ ) where X

is a non-empty set. Let K(X) be the set of constant functions on X.
Prove that (K(X), k · k∞ ) is a Banach space. (Hint: Theorem 9.4.2)

304
 9.5. COMPACTNESS

9.5 Compactness

Important results about continuous functions, such as the Extreme

Value Theorem (Theorem 5.3.7) and uniform continuity (Theorem 5.4.7),
depended heavily on the domain being a closed and bounded interval.
On a bounded interval, any sequence (xn) contains a Cauchy subse-
quence (xnk ) (use the bisection algorithm), and if the interval is also
closed then we are guaranteed that the limit of (xnk ) is contained in the
interval. We have already seen that a totally bounded subset E of a
metric space M contains a Cauchy sequence (Lemma 9.4.10) and thus
if E is complete then Cauchy sequences converge in E. This motivates
the following definition.

Definition 9.5.1: Compactness

A metric space M is called compact if M is totally bounded and
complete.

A closed and bounded subset E of R is compact. Indeed, E is complete

because it is closed (Theorem 9.4.2) and it is easy to see how to cover
E with a finite number of open intervals of any given radius ε > 0.
Conversely, if E ⊂ R is compact then E is bounded and E is closed
by Theorem 9.4.2. A similar argument shows that E ⊂ Rn is compact
if and only if E is closed and bounded. This is called the Heine-Borel
theorem.

Theorem 9.5.2: Heine-Borel

A subset E ⊂ Rn is compact if and only if E is closed and bounded.

305
 9.5. COMPACTNESS

Example 9.5.3. The unit n-sphere Sn in Rn+1 is the set

Sn = {x = (x1, x2, . . . , xn, xn+1) ∈ Rn | x21 + x22 + · · · + x2n+1 = 1}.

Explain why Sn is compact subset of Rn+1.

P 1/2
n×n 2
Example 9.5.4. Consider R with norm kAk2 = i,j ai,j . Then
Rn×n is complete. A matrix Q ∈ Rn×n is called orthogonal if QT Q = I,
where QT denotes the transpose of Q and I is the identity matrix.
Prove that the set of orthogonal matrices, which we denote by O(n), is
compact.

A useful characterization of compactness is stated in the language

of sequences.

Theorem 9.5.5: Sequential Criterion for Compactness

A metric space M is compact if and only if every sequence in M
has a convergent subsequence.

Proof. Assume that M is compact. If (xn) is a sequence in M then

{xn | n ∈ N} is totally bounded and thus has a Cauchy subsequence
which converges by completeness of M.
Conversely, assume that every sequence in M has a convergent sub-
sequence. If (xn) is a Cauchy sequence then by assumption it has a
convergent subsequence and thus (xn) converges. This proves M is
complete. Suppose that M is not totally bounded. Then there exists
ε > 0 such that M cannot be covered by a finite number of open balls
of radius ε > 0. Hence, there exists x1, x2 ∈ M such that d(x1, x2) ≥ ε.
By induction, suppose x1, . . . , xk are such that d(xi, xj ) ≥ ε for i 6= j.
Then there exists xk+1 such that d(xi, xk+1) ≥ ε for all i = 1, . . . , k.
By induction then, there exists a sequence (xn) such that d(xi, xj ) ≥ ε

306
 9.5. COMPACTNESS

if i 6= j. Clearly, (xn) is not a Cauchy sequence and therefore cannot

have a convergent subsequence.

Example 9.5.6. Let M be a metric space.

(a) Prove that if E ⊂ M is finite then E is compact.

(b) Is the same true if E is countable?

(c) What if M is compact?

We now describe how compact sets behave under continuous func-

tions.

Theorem 9.5.7
Let f : M1 → M2 be a continuous mapping. If E ⊂ M1 is compact
then f (E) ⊂ M2 is compact.

Proof. We use the sequential criterion for compactness. Let yn = f (xn)

be a sequence in f (E). Since E is compact, by Theorem 9.5.5, there is
a convergent subsequence (xnk ) of (xn). By continuity of f , the subse-
quence ynk = f (xnk ) of (yn) is convergent. Hence, f (E) is compact.

We now prove a generalization of the Extreme value theorem 5.3.7.

Theorem 9.5.8: Extreme Value Theorem

Let (M, d) be a compact metric space. If f : M → R is continuous
then f achieves a maximum and a minimum on M, that is, there
exists x∗ , x∗ ∈ M such that f (x∗) ≤ f (x) ≤ f (x∗) for all x ∈ M. In
particular, f is bounded.

307
 9.5. COMPACTNESS

Proof. By Theorem 9.5.7, f (M) is a compact subset of R and thus

f (M) is closed and bounded. Let y∗ = inf f (M) and let y ∗ = sup f (M).
By the properties of the supremum, there exists a sequence (yn ) in f (M)
converging to y ∗ . Since f (M) is closed, then y ∗ ∈ f (M) and thus y ∗ =
f (x∗) for some x∗ ∈ M. A similar argument shows that y∗ = f (x∗) for
some x∗ ∈ M. Hence, inf f (M) = f (x∗) ≤ f (x) ≤ f (x∗) = sup f (M)
for all x ∈ M.

Let M be a metric space and let E ⊂ M. A cover of E is a

collection {Ui}i∈I of subsets of M whose union contains E. The index
set I may be countable or uncountable. The cover {Ui }i∈I is called an
open cover if each set Ui is open. A subcover of a cover {Ui} of E
is a cover {Uj }j∈J of E such that J ⊂ I.

Theorem 9.5.9: Compactness via Open Covers

A metric space M is compact if and only if every open cover of M
has a finite subcover.

Proof. Assume that M is compact. Then by Theorem 9.5.5, every

sequence in M has a convergent subsequence. Let {Ui} be an open cover
of M. We claim there exists ε > 0 such that for each x ∈ M it holds
that Bε (x) ⊂ Ui for some Ui. If not, then then for each n ∈ N there
exists xn ∈ M such that B1/n(xn) is not properly contained in a single
set Ui . By assumption, the sequence (xn) has a converging subsequence,
say it is (xnk ) and y = lim xnk . Hence, for each k ∈ N, B1/nk (xnk ) is
not properly contained in a single Ui. Now, y ∈ Uj for some j, and
thus since Uj is open there exists δ > 0 such that Bδ (y) ⊂ Uj . Since
(xnk ) → y, there exists K sufficiently large such that d(xnK , y) < δ/2
and n1K < δ/2. Then B1/nK (xnK ) ⊂ Uj which is a contradiction. This
proves that such an ε > 0 exists. Now since M is totally bounded,

308
 9.5. COMPACTNESS

there exists z1 , z2, . . . , zp ∈ M such that Bε (z1 ) ∪ · · · ∪ Bε(zp ) = M, and

since Bε (zj ) ⊂ Uij for some Uij it follows that {Ui1 , Ui2 , . . . , Uip } is a
finite subcover of {Ui}.
For the converse, we prove the contrapositive. Suppose then that M
is not compact. Then by Theorem 9.5.5, there is a sequence (xn) in M
with no convergent subsequence. In particular, there is a subsequence
(yk ) of (xn) such that all yk ’s are distinct and (yk ) has no convergent
subsequence. Then there exists εi > 0 such that Bεi (yi ) contains only
the point yi from the sequence (yk ), otherwise we can construct a sub-
sequence of (yk ) that converges. Hence, {Bεk (yk )}k∈N is an open cover
of the set E = {y1 , y2, y3, . . . , } that has no finite subcover. The set E is
clearly closed since it consists entirely of isolated points of M. Hence,
{Bεk (yk )}k∈N ∪ M\E is an open cover of M with no finite subcover.

Definition 9.5.10: Uniform Continuity

A function f : (M1 , d1) → (M2, d2) is uniformly continuous if
for any ε > 0 there exists δ > 0 such that if d1 (x, y) < δ then
d2(f (x), f (y)) < ε.

Example 9.5.11. A function f : (M1, d1) → (M2 , d2) is Lipschitz if

there is a constant K > 0 such that d2 (f (x), f (y)) ≤ Kd(x, y). Show
that a Lipschitz map is uniformly continuous.

Example 9.5.12. If f : M1 → M2 is uniformly continuous and (xn) is

a Cauchy sequence in M1 , prove that (f (xn)) is a Cauchy sequence in
M2 .

309
 9.5. COMPACTNESS

Theorem 9.5.13
If f : (M1 , d1) → (M2, d2) is continuous and M1 is compact then f
is uniformly continuous.

Proof. Let ε > 0. For each x ∈ M1 , there exists rx > 0 such that if y ∈
Brx (x) then f (y) ∈ Bε/2(f (x)). Now {Brx /2(x)}x∈M1 is an open cover
of M1 and since M1 is compact there exists finite x1 , x2, . . . , xN such
that {Bδi (xi)}Ni=1 is an open cover of M1 , where we have set δi = rxi /2.
Let δ = min{δ1, . . . , δN }. If d1 (x, y) < δ, and say x ∈ Bδi (xi), then
d1(y, xi) ≤ d1 (y, x) + d1 (x, xi) < δ + δi < rxi and thus

d2(f (x), f (y)) ≤ d2(f (x), f (xi)) + d2(f (xi), f (y))

< ε/2 + ε/2
= ε.

This proves that f is uniformly continuous.

310
 9.6. FOURIER SERIES

Exercises

Exercise 9.5.1. Prove that if E ⊂ R is compact then sup(E) and

inf(E) are elements of E.

Exercise 9.5.2. Recall that ℓ∞ is the set of sequences in R that are

bounded and equipped with the norm k(xn)k∞ = supn∈N |xn |. Show
that the unit ball B = {(xn) : k(xn)k∞ ≤ 1} (which is clearly bounded)
is not compact in ℓ∞ . (see Example 9.4.7)

Exercise 9.5.3. Let (xn) be a sequence in a metric space M and sup-

pose that (xn) converges to p. Prove that S = {p} ∪ {xn | n ∈ N} is a
compact subset of M.

Exercise 9.5.4. Prove that if M is compact then there exists a count-

able subset E ⊂ M that is dense in M.

Exercise 9.5.5. Let E be a compact subset of M and fix p ∈ M. Prove

that there exists z ∈ E such that d(z, p) ≤ d(x, p) for all x ∈ E.

9.6 Fourier Series

Motivated by problems involving the conduction of heat in solids and
the motion of waves, a major problem that spurred the development of
modern analysis (and mathematics in general) was whether an arbitrary
function f can be represented by a series of the form
∞
a0 X
+ (an cos(nx) + bn sin(nx))
2 n=1

for appropriately chosen coefficients an and bn. A central character

in this development was mathematician and physicist Joseph Fourier
and for this reason such a series is now known as a Fourier series.

311
 9.6. FOURIER SERIES

Fourier made the bold claim (Theorie analytique de la Chaleur, 1822 )

that “there is no function f (x) or part of a function, which cannot be
expressed by a trigonometric series”. Fourier’s claim led B. Riemann
(1854) to develop what we now call the Riemann integral. After Rie-
mann, Cantor’s (1872) interest in trigonometric series led him to the
investigation of the derived set of a set S (which nowadays we call the
limit points of S) and he subsequently developed set theory. The gen-
eral problem of convergence of a Fourier series led to the realization that
by allowing “arbitrary” functions into the picture the theory of integra-
tion developed by Riemann would have to be extended to widened the
class of “integrable functions”. This extension of the Riemann integral
was done by Henri Lebesgue (1902) and spurred the development of
the theory of measure and integration. The Lebesgue integral is widely
accepted as the “official” integral of modern analysis.
In this section, our aim is to present a brief account of Fourier series
with the tools that we have already developed. To begin, suppose that
f : [−π, π] → R is Riemann integrable and can be represented by a
Fourier series, that is,
∞
a0 X
f (x) = + (an cos(nx) + bn sin(nx)) (9.1)
2 n=1

for x ∈ [−π, π]. In other words, the series on the right of (9.1) converges
pointwise to f on [−π, π]. The first question we need to answer is what
are the coefficients an and bn in terms of f ? To that end, we use the
following facts. Let n, m ∈ N:

(i) For all n:

Z π Z π
sin(nx)dx = cos(nx)dx = 0.
−π −π

312
 9.6. FOURIER SERIES

(ii) If n 6= m then
Z π Z π
sin(nx) sin(mx)dx = cos(nx) cos(mx)dx = 0.
−π −π

(iii) For all n and m:

Z π
sin(nx) cos(mx)dx = 0.
−π

(iv) For all n and m:

Z π Z π
2
sin (nx)dx = cos2 (nx)dx = π.
−π −π

Then, using these facts, and momentarily ignoring the interchange of

the integral and infinite sum,
Z π Z πh
a0
f (x) cos(nx)dx = cos(nx)
−π −π 2
∞
X i
+ (ak cos(kx) cos(nx) + bk sin(kx) cos(nx)) dx
k=1

Z π ∞ Z π
a0 X
= cos(nx)dx + ak cos(kx) cos(nx)dx
2 −π −π
k=1
∞
X Z π
+ bk sin(kx) cos(nx)dx
k=1 −π

= an π.

Therefore, Z
1 π
an = f (x) cos(nx)dx.
π −π
A similar calculation shows that
Z
1 π
bn = f (x) sin(nx)dx.
π −π
313
 9.6. FOURIER SERIES

Finally,

Z π Z ∞ Z π ∞ Z π
a0 π X X
f (x)dx = dx + ak cos(kx)dx + bk sin(kx)dx
−π 2 −π −π −π
k=1 k=1
a0
= 2π
2
= a0 π

and therefore
Z π
1
a0 = f (x)dx.
π −π

Of course, the above calculations are valid provided that the Fourier se-
ries converges uniformly to f on [−π, π] since if all we have is pointwise
convergence then in general we cannot interchange the integral sign and
the infinite sum. Since the functions fn (x) = an cos(nx) + bn sin(nx)
in the Fourier series are clearly continuous, and if we insist that the
convergence is uniform, then we have restricted our investigation of
Fourier series to continuous functions! Relaxing this restriction led to
the development of what we now call the Lebesgue integral; Lebesgue
was interested in extending the notion of integration beyond Riemann’s
development so that a wider class of functions could be integrated and,
more importantly, this new integral would be more robust when it came
to exchanging limits with integration, i.e., uniform convergence would
not be needed. A full development of Lebesgue’s theory of integra-
tion is beyond the scope of this book, however, we can still say some
interesting things about Fourier series.
Motivated by our calculations above, suppose that f ∈ C[−π, π]

314
 9.6. FOURIER SERIES

and define
Z
1 π
a0 = f (x)dx
π −π
Z
1 π
an = f (x) cos(nx)dx
π −π
Z
1 π
bn = f (x) sin(nx)dx.
π −π
Assume that the Fourier series of f converges uniformly on C[−π, π]
and let ∞
a0 X
g(x) = + (an cos(nx) + bn sin(nx)).
2 n=1
Then g is continuous on [−π, π]. Does f = g? To answer this question,
our computations above show that
Z Z
1 π 1 π
g(x) cos(nx)dx = f (x) cos(nx)dx
π −π π −π
and therefore Z π
[f (x) − g(x)] cos(nx)dx = 0
−π
for all n ∈ N ∪ {0}. Similarly, for all n ∈ N ∪ {0} we have
Z π
[f (x) − g(x)] sin(nx)dx = 0.
−π
Pn
Let sn (x) = a20 + k=1 ak cos(kx) + bk sin(kx) and recall that (sn) con-
verges uniformly to g. Consider for the moment
Z π Z π Z π Z π
[f (x) − sn(x)]2dx = f (x)2dx − 2 f (x)sn(x)dx + s2n (x)dx.
−π −π −π −π

A straightforward computation shows that

Z π " n
#
a20 X 2
f (x)sn(x)dx = π + (ak + b2k )
−π 2
k=1

315
 9.6. FOURIER SERIES

and " #
Z π 2 n
a X
s2n (x)dx = π 0 + (a2k + b2k ) .
−π 2
k=1
Therefore,
Z Z π Z π
1 π 1 1
[f (x) − sn (x)]2dx = f (x)2dx − s2n (x)dx
π −π π −π π −π
Rπ
Now since −π [f (x) − sn (x)]2dx ≥ 0 it follows that
Z Z
1 π 2 1 π
sn (x)dx ≤ f (x)2dx,
π −π π −π
or equivalently that
n Z π
a20 X 2 1
+ (ak + b2k ) ≤ f (x)2dx.
2 π −π
k=1

a20 P∞ 2
This proves that the series 2 + k=1 (ak + b2k ) converges and
∞ Z π
a20 X 2 1
+ (ak + b2k ) ≤ f (x)2dx.
2 π −π
k=1

In particular, limk→∞ (a2k +b2k ) = 0 and thus limk→∞ ak = limk→∞ bk = 0.

316
 10

Multivariable Differential Calculus

In this chapter, we consider the differential calculus of mappings from

one Euclidean space to another, that is, mappings F : Rn → Rm . In
first-year calculus, you considered the case n = 2 or n = 3 and m = 1.
Examples of functions that you might have encountered were of the
type F (x1, x2) = x21 − x22, F (x1, x2, x3) = x21 + x22 + x23, or maybe even
F (x1, x2) = sin(x1) sin(x2), etc. If now F : Rn → Rm with m ≥ 2 then
F has m component functions since F (x) ∈ Rm . We can therefore
write
F (x) = (f1(x), f2(x), . . . , fm(x))

and fj : Rn → R is called the jth component of F .

In this chapter, unless stated otherwise, we equip Rn with the Eu-
p
clidean 2-norm kxk2 = x21 + x22 + · · · + x2n. For this reason, we will
omit the subscript in kxk2 and simply write kxk.

10.1 Differentiation
Let U ⊂ Rn and let F : U → Rn be a function. How should we define
differentiability of F at some point a ∈ U ? Recall that for a function

317
 10.1. DIFFERENTIATION

f : I → R, where I ⊂ R, we say that f is differentiable at a ∈ I if

f (x) − f (a)
lim
x→a x−a
exists. In this case, we denote f ′ (a) = limx→a f (x)−f
x−a
(a)
and we call f ′ (a)
the derivative of f at a. As it is written, the above definition does not
make sense for F since division of vectors is not well-defined (or at least
we have not defined it). An equivalent definition of differentiability of
f at a is that there exists a number m ∈ R such that
f (x) − f (a) − m(x − a)
lim =0
x→a x−a
which is equivalent to asking that
|f (x) − f (a) − m(x − a)|
lim = 0.
x→a |x − a|
The number m is then denoted by m = f ′(a) as before. Another way to
think about the derivative m is that the affine function g(x) = f (a)+mx
is a good approximation to f (x) for points x near a. The linear part
of the affine function g is ℓ(x) = mx. Thought of in this way, the
derivative of f at a is a linear function.

Definition 10.1.1: The Derivative

Let U be a subset of Rn . A mapping F : U → Rm is said to be
differentiable at a ∈ U if there exists a linear mapping L : Rn →
Rm such that
kF (x) − F (a) − L(x − a)k
lim = 0.
x→a kx − ak

In the definition of differentiability, the expression L(x − a) denotes the

linear mapping L applied to the vector (x − a) ∈ Rn . An equivalent

318
 10.1. DIFFERENTIATION

definition of differentiability is that

kF (x + h) − F (a) − L(h)k
lim =0
h→0 khk

where again L(h) denotes h ∈ Rn evaluated at L. It is not hard to show

that the linear mapping L in the above definition is unique when U ⊂
Rm is an open set. For this reason, we will deal almost exclusively with
the case that U is open without further mention. We therefore call L
the derivative of F at a and denote it instead by L = DF (a). Hence,
by definition, the derivative of F at a is the unique linear mapping
DF (a) : Rn → Rm satisfying

kF (x) − F (a) − DF (a)(x − a)k

lim = 0.
x→a kx − ak
Applying the definition of the limit, given arbitrary ε > 0 there exists
δ > 0 such that if kx − ak < δ then

kF (x) − F (a) − DF (a)(x − a)k

<ε
kx − ak
or equivalently

kF (x) − F (a) − DF (a)(x − a)k < ε kx − ak .

If F : U → Rm is differentiable at each x ∈ U then x 7→ DF (x) is a

mapping from U to the space of linear maps from Rn to Rm . In other
words, if we denote by L(Rn ; Rm ) the space of linear maps from Rn to
Rm then we have a well-defined mapping DF : U → L(Rn ; Rm ) called
the derivative of F on U which assigns the derivative of F at each
x ∈ U.
We now relate the derivative of F with the derivatives of its com-
ponent functions. To that end, we need to recall some basic facts from

319
 10.1. DIFFERENTIATION

linear algebra and the definition of the partial derivative. For the lat-
ter, recall that a function f : U ⊂ Rn → R, has partial derivative at
a ∈ U with respect to xi, if the following limit exists

f (a1 , . . . , ai−1, ai + t, ai+1, . . . , an) − f (a)

lim
t→0 t

or equivalently, if there exists a number mi ∈ R such that

f (a + ei t) − f (a) − mi t
0 = lim
t→0 t

where ei = (0, . . . , 0, 1, 0, . . . , 0) denotes the ith standard basis vector in

∂f
Rn . We then denote mi = ∂x i
(a). Now, given any linear map L : Rn →
Rm , the action of L on vectors in Rn can be represented as matrix-vector
multiplication once we choose a basis for Rn and Rm . Specifically, if we
choose the most convenient bases in Rn and Rm , namely the standard
bases, then

L(x) = Ax

where A ∈ Rm×n and the the (j, i) entry of the matrix A is the jth
component of the vector Aei ∈ Rm . We can now prove the following.

320
 10.1. DIFFERENTIATION

Theorem 10.1.2: Jacobian Matrix

Let U ⊂ Rn be open and suppose that F : U → Rm is differentiable
at a ∈ U , and write F = (f1, f2, . . . , fm). Then the partial deriva-
∂f
tives ∂xji (a) exist, and the matrix representation of DF (a) in the
standard bases in Rn and Rm is
 ∂f1 ∂f1 ∂f1

∂x1 ∂x2
· · · ∂xn
 ∂f2 ∂f2 · · · ∂f2 
 ∂x1 ∂x2 ∂xn 
 .. .
.. . . . ... 
 . 
∂fm ∂fm ∂fm
∂x1 ∂x2 · · · ∂xn

where all partial derivatives are evaluated at a. The matrix above

is called the Jacobian matrix of F at a.

Proof. Let mj,i denote the (j, i) entry of the matrix representation of
DF (a) in the standard bases in Rn and Rm , that is, mj,i is the jth
component of DF (a)ei . By definition of differentiability, it holds that
kF (x) − F (a) − DF (a)(x − a)k
0 = lim .
x→a kx − ak
Let x = a + tei where ei ∈ Rn is the ith standard basis vector. Since U
is open, x ∈ U provided t is sufficiently small. Then since kx − ak =
ktei k = |t| → 0 iff kx − ak → 0 we have
kF (a + tei ) − F (a) − DF (a)ei tk
0 = lim
t→0 |t|
1 
= lim F (a + tei ) − F (a) − DF (a)eit .
t→0 t

It follows that each component of the vector 1t F (a + tei ) − F (a) −

DF (a)ei t tends to 0 as t → 0. Hence, for each j ∈ {1, 2, . . . , m} we
have
1
0 = lim (fj (a + tei ) − fj (a) − mj,i t).
t→0 t

321
 10.1. DIFFERENTIATION

∂fj ∂fj
Hence, ∂xi
(a) exists and mj,i = ∂xi
(a) as claimed.

It is customary to write
 ∂f1 ∂f1 ∂f1

∂x1 ∂x2 ··· ∂xn
 ∂f2 ∂f2
··· ∂f2 
DF (a) =  ∂x. 1 ∂x2 ∂xn 

 .. .. ... .. 
. . 
∂fm ∂fm ∂fm
∂x1 ∂x2 ··· ∂xn

since for any x ∈ Rn the vector DF (a)x is the Jacobian matrix of

F at a multiplied by x (all partials are evaluated at a). When not
explicitly stated, the matrix representation of DF (a) will always mean
the Jacobian matrix representation.
We now prove that differentiability implies continuity. To that end,
we first recall that if A ∈ Rm×n and B ∈ Rn×p then

kABk2 ≤ kAk2 kBk2 .

The proof of this fact is identical to the one in Example 9.4.16. In

particular, if x ∈ Rn then kAxk2 ≤ kAk2 kxk.

Theorem 10.1.3: Differentiability implies Continuity

Let U ⊂ Rn be an open set. If F : U → Rm is differentiable at
a ∈ U then F is continuous at a.

Proof. Let ε1 = 1. Then there exists δ1 > 0 such that if kx − ak < δ1

then

kF (x) − F (a) − DF (a)(x − a) + DF (a)(x − a)k < 1 · kx − ak .

322
 10.1. DIFFERENTIATION

Then if kx − ak < δ1 then

kF (x) − F (a)k = kF (x) − F (a) − DF (a)(x − a) + DF (a)(x − a)k

≤ kF (x) − F (a) − DF (a)(x − a)k + kDF (a)(x − a)k

≤ kx − ak + kDF (a)k2 kx − ak

and thus kF (x) − F (a)k < ε provided

kx − ak < min{δ1 , ε/(1 + kDF (a)k2 )}.

Hence, F is continuous at a.

Notice that Theorem 10.1.2 says that if DF (a) exists then all the
relevant partials exist. However, it does not generally hold that if all
the relevant partials exist then DF (a) exists. The reason is that partial
derivatives are derivatives along the coordinate axes whereas, as seen
from the definition, the limit used to define DF (a) is along any direction
that x → a.

Example 10.1.4. Consider the function f : R2 → R defined as

(
2xy
x2 +y 2 , (x, y) 6= (0, 0)
f (x, y) =
0, (x, y) = (0, 0)

∂f ∂f
We determine whether ∂x
(0, 0) and ∂y
(0, 0) exist. To that end, we
compute

f (x + t, 0) − f (0, 0) 0
lim = lim = 0
t→0 t t→0 t

f (0, y + t) − f (0, 0) 0
lim = lim = 0
t→0 t t→0 t

323
 10.1. DIFFERENTIATION

Therefore, ∂f ∂f
∂x (0, 0) and ∂y (0, 0) exist and are both equal to zero. It is
straightforward to show that f is not continuous at (0, 0) and therefore
not differentiable at (0, 0).

The previous examples shows that existence of partial derivatives is

a fairly weak assumption with regards to differentiability, in fact, even
with regards to continuity. The following theorem gives a sufficient
condition for DF to exist in terms of the partial derivatives.

Theorem 10.1.5: Condition for Differentiability

Let U ⊂ Rn be an open set and consider F : U → Rm with F =
∂f
(f1, f2, . . . , fm ). If each partial derivative function ∂xji exists and is
continuous on U then F is differentiable on U .

We will omit the proof of Theorem 10.1.5.

Example 10.1.6. Let F : R2 → R3 be defined by

F (x) = (x1 sin(x2), x1x22, ln(x21 + 1) + 2x2).

Explain why DF (x) exists for each x ∈ R2 and find DF (x).

Solution. It is clear that the component functions of F that are given

by f1 (x) = x1 sin(x2), f2(x) = x1 x22, and f3 (x) = ln(x21 + 1) + 2x2
have partial derivatives that are continuous on all of R2 . Hence, F is
differentiable on R2 . Then
 
sin(x2) x1 cos(x2)
DF (x) =  x22 2x1x2 
2x1
x2 +1
2
1

324
 10.1. DIFFERENTIATION

Example 10.1.7. Prove that the given function is differentiable on R2 .

( x2 y2
√ 2 2 , (x, y) 6= (0, 0)
f (x, y) = x +y
0, (x, y) = (0, 0)
Solution. We compute
√ 0
f (0 + t, 0) − f (0, 0) 2
lim = lim t = 0
t→0 t t→ t

and thus ∂f
∂x (0, 0) = 0. A similar computations shows that
∂f
∂y (0, 0) = 0.
On the other hand, if (x, y) 6= (0, 0) then
∂f xy 2 (x2 + 2y 2)
(x, y) =
∂x (x2 + y 2 )3/2
∂f x2y(2x2 + y 2 )
(x, y) = .
∂y (x2 + y 2 )3/2
To prove that Df (x, y) exists for any (x, y) ∈ R2 , it is enough to show
that ∂f ∂f 2
∂x and ∂y are continuous on R (Theorem 10.1.5). It is clear that
∂f ∂f 2
∂x and ∂y are continuous on the open set U = R \{(0, 0)} and thus Df
exists on U . Now consider the continuity of ∂f∂x at (0, 0). Using polar
coordinates x = r cos(θ) and y = r sin(θ), we can write
∂f xy 2(x2 + 2y 2 )
(x, y) =
∂x (x2 + y 2 )3/2

r3 cos(θ) sin2 (θ)(r2 cos2(θ) + 2r2 sin2 (θ)

=
r3
= r2 cos(θ) sin2(θ)(cos2 (θ) + 2 sin2 (θ))

Now (x, y) → (0, 0) if and only if r → 0 and thus

∂f  
lim (x, y) = lim r2 cos(θ) sin2(θ)(cos2 (θ) + 2 sin2 (θ))
(x,y)→(0,0) ∂x r→0

=0

325
 10.1. DIFFERENTIATION

In other words, lim(x,y)→(0,0) ∂f ∂f ∂f

∂x (x, y) = ∂x (0, 0) and thus ∂x is contin-
uous at (0, 0). A similar computation shows that ∂f ∂y is continuous at
2
(0, 0). Hence, by Theorem 10.1.5, Df exists on R .

If F : U ⊂ Rn → Rm is differentiable on U and m = 1, then DF is

called the gradient of F and we write ∇F instead of DF . Hence, in
this case,
 ∂F ∂F ∂F

∇F (x) = ∂x 1 ∂x2 · · · ∂xn

On the other hand, if n = 1 and m ≥ 2 then F : U ⊂ R → Rm is

a curve in Rm . In this case, it is customary to use lower-case letters
such as c, α, or γ instead of F , and use I for the domain instead of
U . In any case, since c : I ⊂ R → Rm is a function of one variable we
use the notation c(t) = (c1 (t), c2(t), . . . , cm (t)) and the derivative of c
is denoted by
dc
= c′ (t) = (c′1(t), c′2(t), . . . , c′m (t))
dt
where all derivatives are derivatives of single-variable-single-valued func-
tions.

326
 10.1. DIFFERENTIATION

Exercises

Exercise 10.1.1. Let f, g : U ⊂ Rn → Rm be differentiable functions

at a ∈ U . Prove by definition that h = f + g is differentiable at a and
that Dh = Df + Dg.

Exercise 10.1.2. Recall that a mapping F : Rn → Rm is said to be

linear if F (x + y) = F (x) + F (y) and F (αx) = αF (x), for all x, y ∈ Rn
and α ∈ R. Prove that if F is linear then DF (a) = F for all a ∈ Rn .

Exercise 10.1.3. Let F : Rn → Rm and suppose that there exists

M > 0 such that kF (x)k ≤ M kxk2 for all x ∈ Rn . Prove that F is
differentiable at a = 0 ∈ Rn and that DF (a) = 0.

Exercise 10.1.4. Determine if the given function is differentiable at

(x, y) = (0, 0).
( xy
√ 2 2 , (x, y) 6= (0, 0)
f (x, y) = x +y
0, (x, y) = (0, 0)

Exercise 10.1.5. Compute DF (x, y, z) if F (x, y, z) = (z xy , x2, tan(xyz)).

327
 10.2. DIFFERENTIATION RULES AND THE MVT

10.2 Differentiation Rules and the MVT

Theorem 10.2.1: Chain Rule
Let U ⊂ Rn and W ⊂ Rm be open sets. Suppose that F : U → Rm
is differentiable at a, F (U ) ⊂ W , and G : W → Rp is differentiable
at F (a). Then (G ◦ F ) : U → Rp is differentiable at a and

D(G ◦ F )(a) = DG(F (a)) ◦ DF (a)

Example 10.2.2. Verify the chain rule for the composite function H =
G ◦ F where F : R3 → R2 and G : R2 → R2 are
 
x1 − 3x2
F (x1, x2, x3) =
x1 x2 x3
 
2y1 + y2
G(y1, y2 ) = .
sin(y2)

An important special case of the chain rule is the composition of a

curve γ : I ⊂ R → Rn with a function f : U ⊂ Rn → R. The composite
function f ◦ γ : I → R is a single-variable and single-valued function.
In this case, if γ ′(t) is defined for all t ∈ I and ∇f (x) exists at each
x ∈ U then

D(f ◦ γ)(t) = ∇f (γ(t)) · γ ′(t)

 
γ1′ (t)
i ′ 
∂f  γ2 (t) 
h
∂f ∂f
= ∂x1 ∂x2 · · · ∂xn  .. 
.
′
γn (t)
n
X ∂f
= (γ(t))γi′(t).
i=1
∂xi

328
 10.2. DIFFERENTIATION RULES AND THE MVT

In the case that γ(t) = a + te and e ∈ Rn is a unit vector, that is,

kek = 1, then
f (a + te) − f (a)
lim = D(f ◦ γ)(0)
t→0 t
= ∇f (γ(0)) · γ ′(0)
= ∇f (a) · e

is called the directional derivative of f at a in the direction

e ∈ Rn .

Example 10.2.3. Let f : R → R and F : R2 → R be differentiable and

∂F/∂x
suppose that F (x, f (x)) = 0. Prove that if ∂F ′
∂y 6= 0 then f (x) = − ∂F/∂y
where y = f (x).

Below is a version of the product rule for multi-variable functions.

Theorem 10.2.4: Product Rule

Let U ⊂ Rn be open and suppose that F : U → Rm and g : U → R
are differentiable at a ∈ U . Then the function G = gF : U → Rm
is differentiable at a ∈ U and

D(gF )(a) = F (a) · ∇g(a) + g(a)DF (a).

Example 10.2.5. Verify the product rule for G = gF if g : R3 → R

and F : R3 → R3 are

g(x1, x2, x3) = x21x3 − ex2

 
x1 x2
F (x1, x2, x3) =  ln(x23 + 1) 
3x1 − x2 − x3
329
 10.2. DIFFERENTIATION RULES AND THE MVT

Example 10.2.6. Let f, g : Rn → R be differentiable functions. Find

an expression of ∇(f g) in terms of f, g, ∇f , and ∇g.

Example 10.2.7. Let f : U ⊂ Rn → R be a differentiable function.

Suppose that γ : [a, b] → Rn is differentiable. Prove that f (γ(t)) =
f (γ(a)) for all t ∈ [a, b] if and only if ∇f (γ(t)) · γ ′(t) = 0 for all
t ∈ [a, b].

Recall the mean value theorem (MVT) on R. If f : [a, b] → R

is continuous on [a, b] and differentiable on (a, b) then there exists c ∈
(a, b) such that f (b) − f (a) = f ′(c)(b − a). The MVT does not generally
hold for a function F : U ⊂ Rn → Rm without some restrictions on U
and, more importantly, on m. For instance, consider f : [0, 1] → R2
defined by f (x) = (x2, x3). Then f (1) − f (0) = (1, 1) − (0, 0) = (1, 1)
while f ′ (c)(1 − 0) = (2c, 3c2) and there is no c ∈ R such that (1, 1) =
(2c, 3c2). With regards to the domain U , we will be able to generalize
the MVT for points a, b ∈ U provided all points on the line segment
joining a and b are contained in U . Specifically, the line segment
joining x, y ∈ U is the set of points

{z ∈ Rn | z = (1 − t)x + ty, t ∈ [0, 1]}.

Hence, the image of the curve γ : [0, 1] → Rn given by γ(t) = (1−t)x+ty

is the line segment joining x and y. Even if U ⊂ Rn is open, the line
segment joining x, y ∈ U may not be contained in U (see Figure 10.1).

330
 10.2. DIFFERENTIATION RULES AND THE MVT

y
x
U

Figure 10.1: Line segment joining x and y not in U

Theorem 10.2.8: Mean Value Theorem

Let U ⊂ Rn be open and assume that f : U → R is differentiable on
U . Let x, y ∈ U and suppose that the line segment joining x, y ∈ U
is contained entirely in U . Then there exists c on the line segment
joining x and y such that f (y) − f (x) = Df (c)(y − x).

Proof. Let γ(t) = (1 − t)x + ty for t ∈ [0, 1]. By assumption, γ(t) ∈ U

for all 0 ≤ t ≤ 1. Consider the function h(t) = f (γ(t)) on [0, 1].
Then h is continuous on [0, 1] and by the chain rule is differentiable
on (0, 1). Hence, applying the MVT on R to h there exists t∗ ∈ (0, 1)
such that h(1) − h(0) = h′ (t∗)(1 − 0). Now h(0) = f (γ(0)) = f (x) and
h(1) = f (γ(1)) = f (y), and by the chain rule,

h′ (t∗ ) = Df (γ(t∗))γ ′(t∗)

= Df (γ(t∗))(y − x).

Hence,
f (y) − f (x) = Df (γ(t∗))(y − x)
and the proof is complete.

331
 10.2. DIFFERENTIATION RULES AND THE MVT

Corollary 10.2.9
Let U ⊂ Rn be open and assume that F = (f1, f2, . . . , fm) : U → Rm
is differentiable on U . Let x, y ∈ U and suppose that the line
segment joining x, y ∈ U is contained entirely in U . Then there
exists c1 , c2, . . . , cm ∈ U on the line segment joining x and y such
that fi(y) − fi(x) = Dfi (ci )(y − x) for i = 1, 2, . . . , m.

Proof. Apply the MVT to each component function fi : U → R

Example 10.2.10. A set U ⊂ Rn is said to be convex if for any

x, y ∈ U the line segment joining x and y is contained in U . Let
F : U → Rm be differentiable. Prove that if U is an open convex set
and DF = 0 on U then F is constant on U .

332
 10.2. DIFFERENTIATION RULES AND THE MVT

Exercises

Exercise 10.2.1. Let U ⊂ Rn be an open set satisfying the following

property: for any x, y ∈ U there is a continuous curve γ : [0, 1] → Rn
such that γ is differentiable on (0, 1) and γ(0) = x and γ(1) = y.

(a) Give an example of a non-convex set U ⊂ R2 satisfying the

above property.

(b) Prove that if U satisfies the above property and f : U → R is

differentiable on U with Df = 0 then f is constant on U .

333
 10.3. THE SPACE OF LINEAR MAPS

10.3 The Space of Linear Maps

Let U be an open subset of Rn . Recall that if F : U → Rn is differen-
tiable at each x ∈ U then DF : U → L(Rn ; Rn ) denotes the derivative
of F on U . The space of linear maps L(Rn ; Rn ) is a vector space which
after

10.4 Solutions to Differential Equations

A differential equation on Rn is an equation of the form

x′ (t) = F (x(t)) (10.1)

where F : Rn → Rn is a given function and x : R → Rn is the unknown

in (10.1). A solution to (10.1) is a curve γ : I → Rn such that

γ ′(t) = F (γ(t))

where I ⊂ R is an interval, possibly infinite. If F is defined

Theorem 10.4.1
Let U ⊂ Rn be an open set and let F : U → Rn be a differentiable
function with a continuous derivative

10.5 High-Order Derivatives

In this section, we consider high-order derivatives of a differentiable
mapping F : U ⊂ Rn → Rm . To do this, we will need to make an
excursion into the world of multilinear algebra. Even though we will
discuss high-order derivatives for functions on Euclidean spaces, it will
be convenient to first work with general vector spaces.

334
 10.5. HIGH-ORDER DERIVATIVES

Definition 10.5.1: Multilinear Maps

Let V1 , V2, . . . , Vk and W be vector spaces. A mapping T : V1 × V2 ×
· · · × Vk → W is said to be a k-multilinear map if T is linear in
each variable separately. Specifically, for any i ∈ {1, 2, . . . , k}, and
any vj ∈ Vj for j 6= i, the mapping Ti : Vi → W defined by

Ti(x) = T (v1, v2, . . . , vi−1, x, vi+1, . . . , vk )

is a linear mapping.

A 1-multilinear mapping is just a linear mapping. A 2-multilinear map-

ping is called a bilinear mapping. Hence, T : V1 × V2 → W is bilinear
if

T (αu + βv, w) = T (αu, w) + T (βv, w)

= αT (u, w) + βT (v, w)

and

T (u, αw + βy) = T (u, αw) + T (u, βy)

= αT (u, w) + βT (u, y)

for all u, v ∈ V1 , w ∈ V2 , and α, β ∈ R. Roughly speaking, a multi-

linear mapping is essentially a special type of polynomial multivariable
function. We will make this precise after presenting a few examples.

Example 10.5.2. Consider T : R × R → R defined as T (x, y) = 2xy.

As can be easily verified, T is bilinear. On the other hand, if T (x, y) =
x2 + y 2 then T is not bilinear since for example T (αx, y) = α2 x2 + y 2 6=
αT (x, y) in general, or T (a + b, y) = (a + b)2 + y 2 6= T (a, y) + T (b, y) in
general. What about T (x, y) = 2xy + y 3 ?

335
 10.5. HIGH-ORDER DERIVATIVES

Example 10.5.3. Let {v1, v2, . . . , vp} be a set of vectors in Rn and

P P
suppose that x = pi=1 xivi and y = pi=1 yi vi. If T : Rn × Rn → Rm
is bilinear then expand T (x, y) so that it depends only on xi, yj and
T (vi, vj ) for 1 ≤ i, j ≤ p.
Example 10.5.4. Let M be a n×n matrix and define T : Rn ×Rn → R
as T (u, v) = uT Mv. Show that T is bilinear. For instance, if say
M = [ 21 −3
1 ] then
  
1 −3 v1
T (u, v) = [u1 u2]
0 1 v2
= u1v1 − 3u1v2 + u2v2 .

Notice that T (u, v) is a polynomial in the components of u and v.

Example 10.5.5. The function that returns the determinant of a ma-
trix is multilinear in the columns of the matrix. Specifically, if say
A = [a1 + b1 a2 · · · an ] ∈ Rn×n then

det(A) = det([a1 a2 · · · an ]) + det([b1 a2 · · · an ])

and if A = [αa1 a2 · · · an ] then

det(A) = α det([a1 a2 · · · an ]).

These facts are proved by expanding the determinant along the first
column. The same is true if we perform the same computation with a
different column of A. In the case of a 2 × 2 matrix A = [ xx21 yy21 ] we have

det(A) = x1 y2 − y1 x2

and if A is a 3 × 3 matrix with columns x = (x1, x2, x3), y = (y1 , y2, y3),
and z = (z1, z2 , z3) then

det(A) = det([x y z])

= x1y2 z3 − x1y3 z2 − x2y1 z3 + x2y3 z1 + x3 y1z2 − x3 y2z1 .

336
 10.5. HIGH-ORDER DERIVATIVES

We now make precise the statement that a multilinear mapping is

a (special type of) multivariable polynomial function. For simplicity,
and since this will be the case when we consider high-order derivatives,
we consider k-multilinear mappings T : Rn × Rn × · · · × Rn → Rm .
For a positive integer k ≥ 1 let (Rn )k = Rn × Rn × · · · × Rn where
on the right-hand-side Rn appears k-times. Let Lk (Rn , Rm ) denote the
space of k-multilinear maps from (Rn )k to Rm . It is easy to see that
Lk (Rn , Rm ) is a vector space under the natural notion of addition and
scalar R-multiplication. In what follows we consider the case k = 3, the
general case is similar but requries more notation. Hence, suppose that
T : (Rn )3 → Rm is a multilinear mapping and let x = (x1, x2, . . . , xn),
P
y = (y1, y2, . . . , yn), and z = (z1, z2 , . . . , zn ). Then x = ni=1 xi ei where
ei is the ith standard basis vector of Rn , and similarly for y and z.
Therefore, by multilinearity of T we have

n n n
!
X X X
T (x, y, z) = T xiei , yi ei , . . . , zi ei
i=1 i=1 i=1

n X
X n X
n
= xiyj zk · T (ei, ej , ek ).
i=1 j=1 k=1

Thus, to compute T (x, y, z) for any x, y, z ∈ Rn , we need only know

the values T (ei , ej , ek ) ∈ Rm for all triples (i, j, k) with 1 ≤ i, j, k ≤ n.
If we set

T (ei, ej , ek ) = (A1i,j,k , A2i,j,k , . . . , Am

i,j,k )

where the superscripts are not exponents but indices, then from our

337
 10.5. HIGH-ORDER DERIVATIVES

computation above
 n

X
 A1i,j,k · xiyj zk 
 
i,j,k=1 
X n 
 2 
 Ai,j,k · xiyj zk 
T (x, y, z) = 
i,j,k=1
.


 .
.. 

 n 
X m 
 Ai,j,k · xiyj zk 
i,j,k=1

Notice that the component functions of T are multilinear, specifically,

the mapping
n
X
(x, y, z) 7→ Tr (x, y, z) = Ari,j,k · xiyj zk
i,j,k=1

is multilinear for each r = 1, 2, . . . , m. The n3m numbers Ari,j,k ∈ R

for 1 ≤ i, j, k ≤ n and 1 ≤ r ≤ m completely determine the multilin-
ear mapping T , and we call these the coefficients of the multilinear
mapping T in the standard bases.

Remark 10.5.6. The general case k ≥ 1 is just more notation. If T :

(Rn )k → Rm is k-multilinear then there exists nk m unique coefficients
Ari1 ,i2 ,...,ik , where 1 ≤ i1 , i2, . . . , ik ≤ n and 1 ≤ r ≤ m, such that for any
vectors u1 , u2, . . . , uk ∈ Rn it holds that
m n X n n
!
X X X
T (u1, u2, . . . , uk ) = ··· Ari1,i2 ,...,ik · u1,i1 u2,i2 · · · uk,ik er
r=1 i1 =1 i2 =1 ik =1

where e1 , e2, . . . , em are the standard basis vectors in Rm .

A multilinear mapping T ∈ Lk (Rn , Rm ) is said to be symmetric if

the value of T is unchanged after an arbitrary permutation of the inputs

338
 10.5. HIGH-ORDER DERIVATIVES

to T . In other words, T is symmetric if for any v1, v2, . . . , vk ∈ Rn it

holds that

T (v1, v2, . . . , vk ) = T (vσ(1) , vσ(2) , . . . , vσ(n) )

for any permutation σ : {1, 2, . . . , n} → {1, 2, . . . , n}. For instance, if

T : (Rn )3 → Rm is symmetric then for any u1, u2, u3 ∈ Rn it holds that

T (u1, u2, u3) = T (u1, u3, u2)

= T (u2, u1, u3)
= T (u2, u3, u2)
= T (u3, u1, u2, )
= T (u3, u2, u1).

Example 10.5.7. Consider T : R2 × R2 → R defined by

T (x, y) = 2x1y1 + 3x1y2 + 3y1x2 − x2y2.

Then

T (y, x) = 2y1x1 + 3y1x2 + 3x1y2 − y2 x2

= T (x, y)

and therefore T is symmetric. Notice that

  
2 3 y1
T (x, y) = [x1 x2 ]
3 −1 y2
= xT My

and the matrix M = [ 23 −1

3 ] is symmetric.

Having introduced the very basics of multilinear mappings, we can

proceed with discussing high-order derivatives of vector-valued multi-
variable functions. Suppose then that F : U → R is differentiable on

339
 10.5. HIGH-ORDER DERIVATIVES

the open set U ⊂ Rn and as usual let DF : U → L(Rn , Rm ) denote

the derivative. Now L(Rn , Rm ) is a finite dimensional vector space and
can be equipped with a norm (all norms on a given finite dimensional
vector space are equivalent). Thus, we can speak of differentiability
of DF , namely, DF is differentiable at a ∈ U if there exists a linear
mapping L : Rn → L(Rn , Rm ) such that
kDF (x) − DF (a) − L(x − a)k
lim = 0.
x→a kx − ak
If such an L exists then we denote it by L = D(DF )(a). To simplify
the notation, we write instead D(DF )(a) = D2 F (a). Hence, DF is
differentiable at a ∈ U if there exists a linear mapping D2 F (a) : Rn →
L(Rn , Rm )) such that
DF (x) − DF (a) − D2 F (a)(x − a)
lim = 0.
x→a kx − ak
To say that D2 F (a) is a linear mapping from Rn to L(Rn , Rm ) is to say
that
D2 F (a) ∈ L(Rn , L(Rn , Rm )).
Let us focus our attention on the space L(Rn , L(Rn, Rm )). If L ∈
L(Rn , L(Rn , Rm )) then L(v) ∈ L(Rn , Rm ) for each v ∈ Rn , and more-
over the assignment v 7→ L(v) is linear, i.e., L(αv + βu) = αL(v) +
βL(u). Now, since L(v) ∈ L(Rn , Rm ), we have that

L(v)(αu + βw) = αL(v)(u) + βL(v)(w).

In other words, the mapping

(u, v) 7→ L(u)(v)

is bilinear! Hence, L defines (uniquely) a bilinear map T : Rn × Rn →

Rm by
T (u, v) = L(u)(v)

340
 10.5. HIGH-ORDER DERIVATIVES

and the assignment L 7→ T is linear. Conversely, to any bilinear map

T : Rn × Rn → Rm we associate an element L ∈ L(Rn , L(Rn , Rm ))
defined as
L(u)(v) = T (u, v)

and the assignment T 7→ L is linear. We have therefore proved the

following.

Lemma 10.5.8
Let V and W be vector spaces. The vector space L(V, L(V, W )) is
isomorphic to the vector space L2 (V, W ) of multilinear maps from
V × V to W .

The punchline is that D2 F (a) ∈ L(Rn , L(Rn , Rm )) can be viewed in

a natural way as a bilinear mapping D2 F (a) : Rn × Rn → Rm and thus
from now on we write D2 F (a)(u, v) instead of the more cumbersome
D2 F (a)(u)(v).
We now determine a coordinate expression for D2 F (a)(u, v). First
P
of all, if F = (f1, f2, . . . , fm) then F (x) = m j=1 fj (x)ej where {e1 , e2 ,
. . . , em } is the standard basis of Rm . By linearity of the derivative and
P
the product rule of differentiation, we have that DF = m j=1 Dfj (x)ej
P
and also D2 F = m 2
j=1 D fj (x)ej . Therefore,

m
X
2
D F (a)(u, v) = D2 fj (a)(u, v)ej .
j=1

This shows that we need only consider D2 f for R-valued functions

f : U ⊂ Rn → R. Now,
h i
∂f ∂f ∂f
Df = ∂x1 ∂x2 ..., ∂xn

341
 10.5. HIGH-ORDER DERIVATIVES

and thus the Jacobian of Df : U → Rn is (Theorem 10.1.2)

 
∂2f ∂2f ∂2f
···
 ∂x∂ 21fx1 ∂x2 x1
∂2f
∂xn x1 
∂2f 
 ···
D2 f =  . ∂x1 x2 ∂x2 x2
.. ...
∂xn x2  .
.. 
 .. . . 
∂2f ∂2f ∂2f
∂x1 xn ∂x2 xn ··· ∂xn xn

Therefore,
2 ∂ 2f
D f (a)(ei, ej ) = (a).
∂xj xi
Therefore, for any u = (u1, u2, . . . , un) and v = (v1, v2, . . . , vn), by mul-
tilinearity we have
n X n
2
X ∂ 2f
D f (a)(u, v) = (a)uivj .
i=1 j=1
∂x i x j

Now, if all second order partials of f are defined and continuous on U

we can say more. Let us first introduce some terminology. We say that
f : U ⊂ Rn → R is of class C k if all partial derivatives up to and
including order k of f are continuous functions on U .

Theorem 10.5.9: Symmetry of Partial Derivatives

Let U ⊂ Rn be an open set and suppose that f : U → R is of class
C 2. Then
∂ 2f ∂ 2f
=
∂xixj ∂xj xi
on U for all 1 ≤ i, j ≤ n. Consequently, D2 f (a) is a symmetric
bilinear map on Rn × Rn .

If we now go back to a multi-valued function F : U → Rm with com-

342
 10.5. HIGH-ORDER DERIVATIVES

ponents F = (f1, f2, . . . , fm ), then if D2 F (a) exists at a ∈ U then

Pn ∂ 2 f1

i,j=1 ∂xi xj (a)u v
i j
Pn ∂ 2
f 
 i,j=1 ∂xi xj
2
(a)u i v j 
D2 F (a)(u, v) =   ..


 . 
Pn ∂ 2 fm
i,j=1 ∂xi xj (a)ui vj

Higher-order derivatives of F : U → Rm can be treated similarly. If

Dk−1 F : U → Lk−1(Rn , Rm ) is differentiable at a ∈ U then we denote
the derivative at a by D(Dk−1)F (a) = Dk F (a). Then Dk F (a) : Rn →
Lk−1 (Rn, Rm ) is a linear map, that is,

Dk F (a) ∈ L(Rn , Lk−1(Rn, Rm )).

The vector space L(Rn , Lk−1(Rn , Rm)) is isomorphic to the space of k-

multilinear maps Lk (Rn , Rm ). The value of Dk F (a) at u1 , u2, . . . , uk ∈
Rn is denoted by Dk F (a)(u1, u2, . . . , uk ). Moreover, Dk F (a) is a sym-
metric k-multilinear map at each a ∈ U if F is of class C k . If f : U ⊂
Rn → R is of class C k then for vectors u1 , u2, . . . , uk ∈ Rn we have

k
X ∂kf
D f (a)(u1, u2, . . . , uk ) = (a)u1,i1 u2,i2 · · · uk,ik
1≤i1 ,i2 ,...,ik ≤n
∂x i 1 ∂x i 2 · · · ∂x i k

where the summation is over all k-tuples (i1, i2, . . . , ik ) where ij ∈

{1, 2, . . . , n}. Hence, there are nk terms in the above summation. In
the case that u1 = u2 = · · · = uk = x, the above expression takes the
form

k
X ∂kf
D f (a)(x, x, . . . , x) = (a)xi1 xi2 · · · xik
1≤i1 ,i2 ,...,ik ≤n
∂x i 1
∂x i 2
· · · ∂x i k

Example 10.5.10. Compute D3 f (a)(u, v, w) if f (x, y) = sin(x − 2y),

a = (0, 0), and u, v, w ∈ R2 . Also compute D2 f (a)(u, u, u).

343
 10.5. HIGH-ORDER DERIVATIVES

Solution. We compute that f (0, 0) = 0 and

fx = cos(x − 2y)
fy = −2 cos(x − 2y)
and then
fxx = − sin(x − 2y)
fxy = fyx = 2 sin(x − 2y)
fyy = −4 sin(x − 2y)
and then
fxxx = − cos(x − 2y)
fyyy = 8 cos(x − 2y)
fxxy = fxyx = fyxx = 2 cos(x − 2y)
fxyy = fyxy = fyyx = −4 cos(x − 2y)
Then,
D3 f (a)(u, v, w) = fxxx (a)u1v1w1 + fxxy (a)u1v1w2 + fxyx (a)u1v2w1
+ fxyy (a)u1v2w2 + fyxx (a)u2v1w1 + fyxy (a)u2v1 w2
+ fyyx(a)u2v2w1 + fyyy (a)u2v2w2

= −u1v1w1 + 2u1v1w2 + 2u1v2 w1 − 4u1v2w2

+ 2u2v1w1 − 4u2v1w2 − 4u2v2w1 + 8u2v2w2

= −u1v1w1 + 2(u1v1 w2 + u1v2 w1 + u2v1 w1)

− 4(u1v2w2 + u2v1w2 + u2v2w1 ) + 8u2v2w2
If u = v = w then
D3 f (a)(u, u, u) = −u31 + 6u21u2 − 12u1u22 + 8u32

344
 10.6. TAYLOR’S THEOREM

10.6 Taylor’s Theorem

Taylor’s theorem for a function f : Rn → R is as follows.

Theorem 10.6.1: Taylor’s Theorem

Let U ⊂ Rn be an open set and suppose that f : U → R if of class
C r+1 on U . Let a ∈ U and suppose that the line segment between
a and x ∈ U lies entirely in U . Then there exists c ∈ U on the line
segment such that
r
X 1 k
f (x) = f (a) + D f (a)(x − a, x − a, . . . , x − a) + Rr (x)
k!
k=1

where
1
Rr (x) = Dr+1 f (c)(x − a, x − a, . . . , x − a).
(r + 1)!
Furthermore,
Rr (x)
lim r = 0
x→a kx − ak

If x = a + h in Taylor’s theorem then

r
X 1 k
f (a + h) = f (a) + D f (a)(h, h, . . . , h)
k!
k=1
1
+ Dr+1 f (c)(h, h, . . . , h)
(r + 1)!
and
Rr (h)
lim r = 0.
h→0 khk
We call
r
X 1 k
Tr (x) = f (a) + D f (a)(x − a, x − a, . . . , x − a)
k!
k=1

345
 10.6. TAYLOR’S THEOREM

the rth order Taylor polynomial of f centered at a and

1
Rr (x) = Dr+1 f (c)(x − a, x − a, . . . , x − a)
(r + 1)!
the rth order remainder term. Hence, Taylor’s theorem says that

f (x) = Tr (x) + Rr (x)

Since limx→a Rr (x) = 0, for x close to a we get an approximation

f (x) ≈ Tr (x).

Moreover, since Dr+1 f is continuous, there is a constant M > 0 such

that if x is sufficiently close to a then the remainder term satisfies the
bound
|Rr (x)| ≤ M kx − akr+1 .
From this it follows that
Rr (x)
lim r =0
x→a kx − ak

Example 10.6.2. Compute the third-order Taylor polynomial of f (x, y) =

sin(x − 2y) centered at a = (0, 0).
Solution. Most of the work has been done in Example 10.5.10. Evalu-
ating all derivatives at a we find that

Df (a)(u) = fx (a)u1 + fy (a)u2 = u1 − 2u2

D2 f (a)(u, u) = 0

D3 f (a)(u, u, u) = −u31 + 6u21u2 − 12u1u22 + 8u32

Therefore,

Tr (u) = u1 − 2u2 − u31 + 6u21u2 − 12u1u22 + 8u32.

346
 10.6. TAYLOR’S THEOREM

Exercises

Exercise 10.6.1. Find the 2nd order Taylor polynomial of the function
f (x, y, z) = cos(x + 2y)ez centered at a = (0, 0, 0).

Exercise 10.6.2. A function L : Rn → R is called a homogeneous

function of degree k ∈ N if for all α ∈ R and x ∈ Rn it holds that
L(αx) = αk L(x). Prove that if f : Rn → R is differentiable at a ∈ Rn
then the mapping

L(x) = Dk f (a)(x, x, . . . , x)

is a homogeneous function of degree k ∈ N.

347
 10.7. THE INVERSE FUNCTION THEOREM

10.7 The Inverse Function Theorem

A square linear system

a1,1 x1 + a1,2 x2 + · · · + a1,n xn = y1

a2,1 x1 + a2,2 x2 + · · · + a2,n xn = y2
.. .. .
. . = ..
an,1 x1 + an,2 x2 + · · · + an,n xn = yn

or in vector form
Ax = y,
where the unknown is x = (x1, x2, . . . , xn) ∈ Rn , has a unique solution
if and only if A−1 exists if and only if det(A) 6= 0. In this case, the
solution is y = A−1x. Another way to say this is that the mapping
F (x) = Ax has a global inverse given by F −1(x) = A−1x. Hence,
invertibility of DF = A completely determines whether F is invertible.
Consider now a system of equations

F (x) = y

where F : Rn → Rn is nonlinear. When is it possible to solve for x

in terms of y, that is, when does F −1(x) exists? In general, this is a
difficult problem and we cannot expect global invertibility even when
assuming the most desirable conditions on F . Even in the 1D case,
we cannot expect global invertibility. For instance, f (x) = cos(x) is
not globally invertible but is so on any interval where f ′(x) 6= 0. For
instance, on the interval I = (0, π), we have that f ′(x) = sin(x) 6= 0
and f −1(x) = arcsin(x). In any neighborhood where f ′ (x) = 0, for
instance, at x = 0, f (x) = cos(x) is not invertible. However, having a

348
 10.7. THE INVERSE FUNCTION THEOREM

non-zero derivative is not necessary for invertibility. For instance, the

function f (x) = x3 has f ′ (0) = 0 but f (x) has an inverse locally around
x = 0; in fact it has a global inverse f −1(x) = x1/3.
Let’s go back to the 1D case and see if we can say something about
the invertibility of f : R → R locally about a point a such that f ′(a) 6=
0. Assume that f ′ is continuous on R (or on an open set containing a).
Then there is an interval I = [a − δ, a + δ] such that f ′(x) 6= 0 for all
x ∈ I. Now if x, y ∈ I and x 6= y, then by the Mean Value Theorem,
there exists c in between x and y such that

f (y) − f (x) = f ′(c)(y − x).

Since f ′(c) 6= 0 and (y − x) 6= 0 then f (y) 6= f (x). Hence, if x 6= y then

f (y) 6= f (x) and this proves that f is injective on I = [c − δ, c + δ].
Therefore, the function f : I → R has an inverse f −1 : J → R where
J = f (I). Hence, if f (a) 6= 0, f has a local inverse at a. In fact, we
can say even more, namely, one can show that f −1 is also differentiable.
Then, since f −1(f (x)) = x for x ∈ I, by the chain rule we have

(f −1)′(f (x)) · f ′(x) = 1

and therefore since f ′ (x) 6= 0 for all x ∈ I we have

1
(f −1)′(f (x)) = .
f ′(x)

The following theorem is a generalization of this idea.

349
 10.7. THE INVERSE FUNCTION THEOREM

Theorem 10.7.1: Inverse Function Theorem

Let V ⊂ Rn be an open set and let F : V → Rn be of class C 1.

Suppose that det(DF (a)) 6= 0 for a ∈ V . Then there exists an open
set U ⊂ Rn containing a such that W = F (U ) is open and F : U →
W is invertible. Moreover, the inverse function F −1 : W → U is
also C 1 and for y ∈ W and x = F −1(y) we have

DF −1 (y) = [DF (x)]−1 .

Example 10.7.2. Prove that F (x, y) = (f1(x, y), f2(x, y)) = (x2 −
y 2 , 2xy) is locally invertible at all points a 6= (0, 0).

Proof. Clearly, DF (x, y) exists for all (x, y) since all partials of the
components of F are continuous on R2 . A direct computation gives
 
2x −2y
DF (x, y) =
2y 2x

and thus det(DF (x, y)) = 2x2 + 2y 2 . Clearly, det(DF (x, y)) = 0 if and
only if (x, y) = (0, 0). Therefore, by the Inverse Function theorem, for
each non-zero a ∈ R2 there exists an open set U ⊂ R2 containing a such
that F : U → F (U ) is invertible. In this very special case, we can find
the local inverse of F about some a ∈ R2 . Let (u, v) = F (x, y), that is,

x2 − y 2 = u
2xy = v
2
v v
If x 6= 0 then y = 2x and therefore x2 − 4x 4 2
2 = u and therefore 4x − v =

4ux2 or
4x4 − 4ux2 − v 2 = 0.

350
 10.7. THE INVERSE FUNCTION THEOREM

By the quadratic formula,

√
4u ± 16u2 + 16v 2
x2 =
8
Since x ∈ R we must take
s √
4u + 16u2 + 16v 2
x=
8
s
√
u+ u2 + v 2
=
2

and therefore √
v 2v
y= = p √
2x 2 u + u2 + v 2
Hence, provided u 6= 0 and v 6= 0 then
q √ 
u+ u2 +v 2
F −1
(u, v) =  √ 2 .
√ √2v
2 u+ u2 +v 2

351
 10.7. THE INVERSE FUNCTION THEOREM

Exercises

Exercise 10.7.1. Let F : R2 → R2 be defined by

F (x, y) = (f1(x, y), f2(x, y)) = (ex cos(y), ex sin(y))

for (x, y) ∈ R2 .

(a) Prove that the range of F is R2 \{0}. Hint: Think polar coordi-
nates.

(b) Prove that F is not injective.

(c) Prove that F is locally invertible at every a ∈ R2 .

Exercise 10.7.2. Can the system of equations

x + xyz = u
y + xy = v
z + 2x + 3z 2 = w

be solved for x, y, z in terms of u, v, w near (0, 0, 0)?

352
 Bibliography

[1] Georg Cantor. Über eiene elementare Frage der Mannigfaltigskeit-

slehre, Jahresber. der DMV, 1: 75-78. 26

[2] Derek Goldrei. Classical Set Theory: For Guided Independent

Study. Chapman and Hall/CRC, 1996. 32, 33

353