Honours Analysis 2 (Math 255) Tutorial Notes

Edward Chernysh

These notes may continue to be updated when the TA-strike ends. I hope to be back and
teaching soon!

McGill University, Montréal, Québec, Canada.

Email: [email protected]

1
Contents

1 Metric Spaces and a Topological Introduction 4

1.1 Distance and Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Nearness and Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 More on Continuous Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 A Quick Look at the Subspace Topology . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Characterizations of continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.6 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.1 Finite Intersection Property . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.7 Connectedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.8 Homeomorphy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.9 Completeness and Compactness in Metric Spaces . . . . . . . . . . . . . . . . . . 21
1.10 Baire Category Theorem and Basic Applications . . . . . . . . . . . . . . . . . . . 24
1.10.1 Some Examples and Applications . . . . . . . . . . . . . . . . . . . . . . . 26
1.11 Lebesgue Measure Zero in One-Dimension . . . . . . . . . . . . . . . . . . . . . . 29

2 Differentiability 31
2.1 Examples and Elementary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 The Mean Value Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 A Word on the Lipschitz Condition . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4 Applications of the Mean Value Theorem and Exercises . . . . . . . . . . . . . . . 40
2.5 Uniform Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6 A Basic Form of l’Hôpital’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.7 The Intermediate Properties of Derivatives: Darboux’s Theorem . . . . . . . . . . 47

3 The Riemann Integral 49

3.1 Approximating by Riemann Sums . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2 Cauchy’s Criterion for Riemann Integrability . . . . . . . . . . . . . . . . . . . . . 57
3.3 The Squeeze Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4 Additivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Examples with Riemann Sums: Even and Odd Functions . . . . . . . . . . . . . . 63
3.6 Integrating Continuous Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.7 Darboux Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.8 Lebesgue Criterion & The Fundamental Theorem of Calculus . . . . . . . . . . . . 69
3.8.1 The Fundamental Theorem of Calculus . . . . . . . . . . . . . . . . . . . . 70

2
4 Sequences of Functions 72
4.1 Technical Examples of Non-Uniform Convergence . . . . . . . . . . . . . . . . . . 74
4.2 A Squeeze type Theorem for Uniform Convergence . . . . . . . . . . . . . . . . . 76
4.3 Continuity of the Uniform Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 Uniform Limits and Bounded functions . . . . . . . . . . . . . . . . . . . . . . . . 80
4.5 More on Uniform Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.6 Uniform Limits of Uniformly Continuous Functions . . . . . . . . . . . . . . . . . 83

5 Infinite Series 85
5.1 The Cauchy Condensation Test & Examples . . . . . . . . . . . . . . . . . . . . . 87
5.2 The Comparison & Limit-Comparison Tests . . . . . . . . . . . . . . . . . . . . . . 88
5.3 Examples of Convergence/Divergence of Series. . . . . . . . . . . . . . . . . . . . 91
5.4 Abel’s Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.5 Dirichlet’s Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.6 Weierstass 𝑀-Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

3
1 Metric Spaces and a Topological Introduction
Let us begin with a brief recall of the key concepts seen in class. A metric space is a pair (𝑋, 𝑑)
where 𝑋 is a non-empty set and 𝑑 : 𝑋 × 𝑋 → [0, ∞) is a metric (or distance-function) satisfying
the following:

(i) 𝑑 (𝑥, 𝑦) = 0 if and only if 𝑥 = 𝑦;

(ii) 𝑑 (𝑥, 𝑦) = 𝑑 (𝑦, 𝑥) for all 𝑥, 𝑦 ∈ 𝑋 ;

(iii) 𝑑 (𝑥, 𝑦) ≤ 𝑑 (𝑥, 𝑧) + 𝑑 (𝑧, 𝑦) for all 𝑥, 𝑦, 𝑧 ∈ 𝑋 .

We note that (iii) is called the triangle inequality. As seen in class, any normed vector space 𝑉
inherits a metric structure from its norm. Indeed, 𝑉 is automatically a metric space when equipped
with the metric 𝑑 (𝑥, 𝑦) = ∥𝑥 − 𝑦 ∥. The most frequently encountered metric space is R𝑛 with its
usual norm and induced metric. However, as seen in class, there are more “exotic” instances of
such spaces.

Example 1.1 (The Discrete Metric). Given any non-empty set 𝑋 , we can define the discrete metric
on 𝑋 by putting (
1 if 𝑥 ≠ 𝑦,
𝑑 (𝑥, 𝑦) :=
0 if 𝑥 = 𝑦.
It is left as a straightforward exercise to verify that 𝑑 is indeed a valid metric on 𝑋 .

Example 1.2. Consider the space 𝑋 = RN of real sequences. That is, an element 𝑥 ∈ 𝑋 is a
sequence 𝑥 = (𝑥𝑛 )𝑛∈N of real numbers. We show that
∞
∑︁ |𝑥𝑛 − 𝑦𝑛 |
𝑑 (𝑥, 𝑦) = 2−𝑛
𝑛=1
1 + |𝑥𝑛 − 𝑦𝑛 |

is a metric on 𝑋 .

(i) It is clear that 𝑑 (𝑥, 𝑥) = 0 for any 𝑥 ∈ 𝑋 . Conversely, if 𝑑 (𝑥, 𝑦) = 0 then

|𝑥𝑛 − 𝑦𝑛 |
2−𝑛 =0
1 + |𝑥𝑛 − 𝑦𝑛 |

for each 𝑛 ∈ N. Thus |𝑥𝑛 − 𝑦𝑛 | = 0 for all 𝑛 ∈ N so 𝑥𝑛 = 𝑦𝑛 at each index 𝑛; i.e. 𝑥 = 𝑦.

(ii) The equality 𝑑 (𝑥, 𝑦) = 𝑑 (𝑦, 𝑥) for all 𝑥, 𝑦 ∈ 𝑋 is trivially true by symmetry of the absolute
value function.

4
 (iii) Finally, to establish the triangle inequality, we first make some observations on the function
𝑡 1
𝑓 : [0, ∞) → [0, 1), 𝑓 (𝑡) = =1− .
1+𝑡 1+𝑡
Clearly, 𝑓 is an increasing function. Moreover, given 𝑠, 𝑡 ≥ 0 there holds
𝑠 𝑡 𝑠 𝑡 𝑠 +𝑡
𝑓 (𝑠) + 𝑓 (𝑡) = + ≥ + = = 𝑓 (𝑠 + 𝑡). (1.1)
1+𝑠 1+𝑡 1+𝑠 +𝑡 1+𝑠 +𝑡 1+𝑠 +𝑡
Using these observations, we see that for arbitrary 𝑥, 𝑦, 𝑧 ∈ 𝑋 there holds
∞
∑︁
𝑑 (𝑥, 𝑦) = 2−𝑛 𝑓 (|𝑥𝑛 − 𝑦𝑛 |)
𝑛=1
∞
∑︁
≤ 2−𝑛 𝑓 (|𝑥𝑛 − 𝑧𝑛 | + |𝑧𝑛 − 𝑦𝑛 |) since 𝑓 increases
𝑛=1
∞
∑︁
≤ 2−𝑛 (𝑓 (|𝑥𝑛 − 𝑧𝑛 |) + 𝑓 (|𝑧𝑛 − 𝑦𝑛 |)) by (1.1)
𝑛=1
∞
∑︁ ∞
∑︁
= 2−𝑛 𝑓 (|𝑥𝑛 − 𝑧𝑛 |) + 2−𝑛 𝑓 (|𝑧𝑛 − 𝑦𝑛 |)
𝑛=1 𝑛=1
= 𝑑 (𝑥, 𝑧) + 𝑑 (𝑧, 𝑦).

1.1 Distance and Bases

As seen in class, the metric 𝑑 induces a topological structure on 𝑋 , i.e. a notion of open sets. That
is, we declare a subset 𝑈 of 𝑋 to be open if, for each 𝑥 ∈ 𝑈 , there exists 𝜀 > 0 such that

𝐵(𝑥, 𝜀) := {𝑦 ∈ 𝑋 : 𝑑 (𝑥, 𝑦) < 𝜀} ⊆ 𝑈 .

These “𝜀-balls” constitute a basis for the topology induced by the metric 𝑑. That is, every open
subset of 𝑋 is the union of 𝜀-balls.

Definition 1.1. Let 𝑋 be a non-empty set. A given collection 𝔗 of subsets of 𝑋 is called a topology
on 𝑋 provided the following conditions are met

1. ∅, 𝑋 ∈ 𝔗;

2. 𝔗 is closed under arbitrary unions;

3. 𝔗 is closed under finitely many intersections.

The pair (𝑋, 𝔗) is then called a topological space. We say that the topological space (𝑋, 𝔗) is
metrizable when there exists a metric 𝑑 on 𝑋 whose induced topology is 𝔗.

5
Definition 1.2. Let (𝑋, 𝔗) be a topological space. A subset of 𝑋 is said to be closed if its com-
plement is open.

In a topological space, open balls are generalized to the concept of a basis.

Definition 1.3. Let 𝑋 be a non-empty set. A family B of subsets from 𝑋 is called a basis on 𝑋
provided each of the following hold
1. For every 𝑥 ∈ 𝑋 there exists some 𝐵 ∈ B such that 𝑥 ∈ 𝐵;

2. If 𝐵 1, 𝐵 2 ∈ B and 𝑥 ∈ 𝐵 1 ∩ 𝐵 2 , there exists 𝐵 3 ∈ B such that 𝑥 ∈ 𝐵 3 ⊆ 𝐵 1 ∩ 𝐵 2 .

If B is a basis for 𝑋 , we define the topology generated by B on 𝑋 as the set

𝔗 := {𝑈 ⊆ 𝑋 : ∀𝑥 ∈ 𝑈 , ∃𝐵 ∈ B s.t. 𝑥 ∈ 𝐵 ⊆ 𝑈 } . (1.2)

That is, 𝔗 consists of all 𝑈 ⊆ 𝑋 such that, for every 𝑥 ∈ 𝑈 , one can find 𝐵 ∈ B with 𝑥 ∈ 𝐵 ⊆ 𝑈 .

Put otherwise, given a basis B we see that the topology generated by B is the collection of
arbitrary unions of basis elements. That is, a set 𝑈 is open in the topology generated by B if
and only if 𝑈 is a (possibly uncountable) union of basis elements. In a metric space, this is the
statement that open sets are arbitrary unions of balls. For instance, in R every open set is an
arbitrary union of open intervals.

Example 1.3. The discrete metric on a non-empty set 𝑋 induces a topology where every subset
of 𝑋 is open. Indeed, with the discrete metric, we have

𝐵(𝑥, 1/2) = {𝑥 }

for any 𝑥 ∈ 𝑋 . Thus, singletons are open. Since arbitrary unions of open sets (or open balls) are
open, every subset of 𝑋 is open. Equivalently, every subset of 𝑋 is closed.
So, we see that the discrete metric induces the topology (𝑋, P (𝑋 )) where P (𝑋 ) denotes the
power set of 𝑋 . This is known as the discrete topology.

Problem 1. Let B be a basis on 𝑋 . Show that the topology 𝔗 generated by B is precisely the inter-
section of all topologies containing B. Put otherwise, 𝔗 is the smallest topology containing B.

Proof. Let B be a basis for a topology 𝔗 on 𝑋 . We claim that 𝔗 = 𝔚 𝔚 where the intersection
Ñ
is taken over all topologies 𝔚 containing B. Since B is a basis for 𝔗, we know that every element
of 𝔗 is simply the union of elements in B. Consequently, if 𝔚 is a topology containing B, we
have 𝔗 ⊆ 𝔚 since 𝔚 is closed with respect to unions. Especially, 𝔗 ⊆ 𝔚 𝔚. Conversely, 𝔗 is
Ñ
a topology containing B and therefore appears as one of the 𝔚 indexing the intersection. This
yields 𝔚 𝔚 ⊆ 𝔗.
Ñ
□

Bases are sometimes useful for simplifying proofs. For instance, let us recall the definition of
continuity:

6
Definition 1.4. Let (𝑋, 𝔗) and (𝑌 , 𝔚) be topological spaces and consider a function 𝑓 : 𝑋 → 𝑌 .
We say that 𝑓 is continuous if 𝑓 −1 (𝑉 ) is open in 𝑋 for every open set 𝑉 in 𝑌 .

As seen in class, in presence of a basis continuity need not be verified for all open sets. We
make this statement precise in the case where 𝑌 is a metric space:

Proposition 1.1. Let 𝑋 be a topological space and 𝑌 a metric space. A function 𝑓 : 𝑋 → 𝑌 is

continuous if and only if for every ball open ball 𝐵 ⊆ 𝑌 the set 𝑓 −1 (𝐵) is open in 𝑋 .

Proof. If 𝑓 is continuous, there is nothing to show. Conversely, suppose that for any open ball
𝐵 ⊆ 𝑌 , the set 𝑓 −1 (𝐵) is open in 𝑋 . Let 𝑈 ⊆ 𝑌 be an arbitrary open set. Since balls generate open
sets, 𝑈 is a union (possibly uncountable) of open balls. That is, we may write
Ø
𝑈 = 𝐵𝑖
𝑖 ∈I

where I is an index set and each 𝐵𝑖 is open. Then,

Ø
𝑓 −1 (𝑈 ) = 𝑓 −1 (𝐵𝑖 ).
𝑖 ∈I

Since each 𝑓 −1 (𝐵𝑖 ) is open, it follows that 𝑓 −1 (𝑈 ) is open as an arbitrary union of open sets. □

Remark 1.1. A similar statement can be made if 𝑌 is a topological space whose topology is gen-
erated by a basis. Modifying the statement/proof of the above proposition for this more general
setting is left as an exercise.
A similar statement can be made for open maps. We begin with a definition:

Definition 1.5. Let 𝑋, 𝑌 be topological spaces and 𝑓 : 𝑋 → 𝑌 a function.

1. We say that 𝑓 is an open map if 𝑓 (𝑂) is open in 𝑌 for all open sets 𝑂 ⊆ 𝑋 .

2. We say that 𝑓 is a closed map if 𝑓 (𝐶) is closed in 𝑌 for all closed sets 𝐶 ⊆ 𝑋 .

Note that continuous functions need not be open or closed. For instance, consider the function
1
𝑓 : R → R, 𝑓 (𝑥) = .
1 + 𝑥2
Clearly, 𝑓 is continuous. However, the image of the clopen1 set R is 𝑓 (R) = (0, 1], which is
neither open nor closed. Hence, 𝑓 is neither an open map nor a closed map.

Proposition 1.2. Let 𝑋 be a metric space space and 𝑌 a topological space. A function 𝑓 : 𝑋 → 𝑌 is
open if and only if for every ball open ball 𝐵 ⊆ 𝑋 the set 𝑓 (𝐵) is open in 𝑌 .
1A clopen set is a set which is both open and closed.

7
Proof. If 𝑓 is open, there is nothing to show. Conversely, suppose that for any open ball 𝐵 ⊆ 𝑋 ,
the set 𝑓 (𝐵) is open in 𝑌 . Let 𝑂 ⊆ 𝑋 be an arbitrary open set. Since balls generate open sets, 𝑂 is
a union (possibly uncountable) of open balls. That is, we may write
Ø
𝑂= 𝐵𝑖
𝑖 ∈I

where I is an index set and each 𝐵𝑖 is open. Then,

Ø
𝑓 (𝑂) = 𝑓 (𝐵𝑖 ).
𝑖 ∈I

Since each 𝑓 (𝐵𝑖 ) is open, it follows that 𝑓 (𝑂) is open as an arbitrary union of open sets. □
Remark 1.2. The above statement cannot be made for closed maps if we replace open sets by closed
sets. This is because arbitrary unions of closed sets need not be closed. For instance, consider the
function
1
𝑓 : R → R, 𝑓 (𝑥) = .
1 + 𝑥2
Note that 𝑓 is continuous. Hence, by the preservation of intervals theorem, every closed bounded
interval [𝑎, 𝑏] (i.e. the closed balls in R) maps to a closed interval. However, the closed (and open)
set R maps to (0, 1] which is not closed.
A similar statement can be made if 𝑋 is a topological space whose topology is generated by
a basis. Modifying the statement/proof of the above proposition for this more general setting is
left as an exercise.

1.2 Nearness and Separation

In a general topological space, the open sets are what constitute our notion of “nearness”. In fact,
given a topological space (𝑋, 𝔗) and a point 𝑥 ∈ 𝑋 , a “neighbourhood” of 𝑥 refers to any open
set containing 𝑥. Adhering to this picture, we can use the open sets to “separate points” as well.
Definition 1.6. Let (𝑋, 𝔗) be a topological space. Two points 𝑥 and 𝑦 are said to be separated
if there exist disjoint open sets containing 𝑥 and 𝑦, respectively. A topological space is called
Hausdorff if all distinct points in 𝑋 are separated.
We remark that all metric spaces are automatically Hausdorff. Indeed, let (𝑋, 𝑑) be a metric
space and fix two distinct points 𝑥, 𝑦 ∈ 𝑋 . Put 𝜀 := 𝑑 (𝑥, 𝑦)/2 and note that 𝜀 > 0 because
𝑥 ≠ 𝑦. Next, consider the two 𝜀-neighbourhoods 𝐵(𝑥, 𝜀) and 𝐵(𝑦, 𝜀). Obviously, these are both
open subsets of 𝑋 containing 𝑥 and 𝑦, respectively. Furthermore, these two balls are disjoint.
Otherwise, there exists 𝑧 ∈ 𝐵(𝑥, 𝜀) ∩ 𝐵(𝑦, 𝜀) whence the triangle inequality yields

𝑑 (𝑥, 𝑦) ≤ 𝑑 (𝑥, 𝑧) + 𝑑 (𝑧, 𝑦) < 2𝜀 = 𝑑 (𝑥, 𝑦),

which is a contradiction. Thus, (𝑋, 𝑑) is Hausdorff.

8
Proposition 1.3. Let (𝑋, 𝔗) be a Hausdorff topological space (e.g. a metric space). All finite subsets
of 𝑋 are closed in 𝑋 .

Proof. Since finite unions of closed sets are again closed, it suffices to show that any singleton is
closed. Let 𝑥 ∈ 𝑋 be given and consider {𝑥 }. The complement of {𝑥 } is the set 𝑈 = 𝑋 \ {𝑥 }. For
each 𝑦 ∈ 𝑈 , since (𝑋, 𝔗) is a Hausdorff space and 𝑥 ≠ 𝑦, we may select disjoint open sets 𝑉𝑦 ,𝑊𝑦
with 𝑥 ∈ 𝑉𝑦 and 𝑦 ∈ 𝑊𝑦 . In particular, 𝑊𝑦 is an open set containing 𝑦 but not intersecting 𝑥. Put
otherwise,
𝑦 ∈ 𝑊𝑦 ⊆ 𝑋 \ {𝑥 } = 𝑈 .
Therefore, 𝑈 is open as the union of the open sets {𝑊𝑦 : 𝑦 ∈ 𝑈 }.
□

We now provide an example of a topology that is not Hausdorff.

Example 1.4. Let 𝑋 be a non-empty set, and endow 𝑋 with the finite complement topology:

𝔗 := {𝑈 ⊆ 𝑋 : 𝑈 c is finite or 𝑈 = ∅} .

It is easy to check that (𝑋, 𝔗) is indeed a topological space. Certainly, it follows from the definition
that ∅, 𝑋 ∈ 𝔗. Furthermore, given a family {𝑈𝛼 }𝛼 ∈𝐼 of sets from 𝔗, we have for each 𝛼 ∈ 𝐼 that
either 𝑈𝛼 = ∅ or 𝑈𝛼c is finite. In the event that every 𝑈𝛼 = ∅, we see ∅ = 𝛼 ∈𝐼 𝑈𝛼 ∈ 𝔗 by
Ð
construction. If some 𝑈 𝛽 ≠ ∅, then
!c
Ø Ù
𝑈𝛼 = 𝑈𝛼c ⊆ 𝑈 𝛽c
𝛼 ∈𝐼 𝛼 ∈𝑈

implies that the complement of 𝛼 ∈𝐼 𝑈𝛼 is finite, and hence 𝔗 is closed under arbitrary unions.
Ð
Finally, let 𝑈 1, 𝑈 2 ∈ 𝔗. If either is empty, it follows at once that 𝑈 1 ∩ 𝑈 2 ∈ 𝔗. Otherwise, 𝑈 1c and
𝑈 2c are finite whence
(𝑈 1 ∩ 𝑈 2 ) c = 𝑈 1c ∪ 𝑈 2c
is finite and 𝑈 1 ∩𝑈 2 ∈ 𝔗. By induction, it readily follows that 𝔗 is closed under finite intersections.
We now show that 𝑋 is not Hausdorff. Let 𝑥, 𝑦 be distinct points in 𝑋 and suppose, by way
of contradiction, that there exist disjoint open sets 𝑈𝑥 and 𝑈 𝑦 containing 𝑥 and 𝑦, respectively.
Since 𝑈𝑥 and 𝑈 𝑦 are both open, they each contain all but finitely many points in 𝑥. That is,

𝑈𝑥 = 𝑋 \ {𝑥 1, . . . , 𝑥𝑘 } and 𝑈 𝑦 = 𝑋 \ {𝑦1, . . . , 𝑦𝑙 }

for the appropriate 𝑥𝑖 , 𝑦 𝑗 ∈ 𝑋 . Since 𝑋 is infinite, we can choose a point 𝜉 ∈ 𝑋 distinct from
each 𝑥𝑖 and every 𝑦 𝑗 . Then, 𝜉 ∈ 𝑈𝑥 ∩ 𝑈 𝑦 , which is absurd. In fact, we have shown that any two
non-empty open sets must intersect!

This last example actually shows that the finite complement topology on an infinite set is not
induced by any metric. Indeed, the metric topology is always Hausdorff!

9
Remark 1.3. The closure of the open ball need not be the closed ball of the same radius and center!
Indeed, consider any set 𝑋 with at least two points equipped with the discrete metric. Then, for
any 𝑥 ∈ 𝑋 , we have by definition that

𝐵(𝑥, 1) = {𝑦 ∈ 𝑋 : 𝑑 (𝑥, 𝑦) < 1} = {𝑦 ∈ 𝑋 : 𝑑 (𝑥, 𝑦) = 0} = {𝑥 }

where this set is both open and closed (why?) in 𝑋 . Consequently, cl (𝐵(𝑥, 1)) = 𝐵(𝑥, 1) = {𝑥 }.
On the other hand, the closed ball of radius 1 about 𝑥 is the entire space 𝑋 !

1.3 More on Continuous Functions

Continuity is not always intuitive, particularly in non-metrizable spaces.

Problem 2. Let 𝑋 be an infinite set equipped with the finite complement topology and let 𝑌 be a
Hausdorff space (e.g. a metric space). Then, every continuous function 𝑋 → 𝑌 is constant.

Proof. Let 𝑓 : 𝑋 → 𝑌 be a continuous mapping. By way of contradiction, assume that 𝑓 is not

constant, i.e. there exist distinct points 𝑦1, 𝑦2 ∈ 𝑌 with 𝑦1 = 𝑓 (𝑥 1 ), 𝑦2 = 𝑓 (𝑥 2 ) for suitable 𝑥 1, 𝑥 2 ∈
𝑋 . Using that 𝑌 is a Hausdorff space, we may select disjoint open sets 𝑉1, 𝑉2 ⊆ 𝑌 with 𝑦1 ∈ 𝑉1 and
𝑦2 ∈ 𝑉2 . By continuity, both 𝑈 1 = 𝑓 −1 (𝑉1 ) and 𝑈 2 = 𝑓 −1 (𝑉2 ) are open in 𝑋 . Furthermore, they
are non-empty since they contain 𝑥 1 and 𝑥 2 , respectively. Consequently, by a previously seen
property of the finite complement topology, 𝑈 1 and 𝑈 2 must necessarily intersect. Thus, there is
a point 𝑧 ∈ 𝑈 1 ∩ 𝑈 2 whence 𝑓 (𝑧) ∈ 𝑉1 ∩ 𝑉2 , contradicting our choice of 𝑉1 and 𝑉2 . □

Thankfully, in metric spaces, much of the intuition developed from studying functions from
the real line to itself will remain useful. For instance, let us recall the definition of density:

Definition 1.7. Let (𝑋, 𝔗) be a topological space. A subset 𝐷 of 𝑋 is said to be dense in 𝑋 provided
cl (𝐷) = 𝑋 . Or, equivalently, if every non-empty open subset of 𝑋 intersects 𝐷.

In analysis I, it is seen that if two continuous functions 𝑓 , 𝑔 : R → R correspond on a dense

set then they are in fact equal. This fact remains true for functions between metric spaces. Even
more generally, we have the following result:

Example 1.5. Let 𝑋 be a topological space and 𝑌 a Hausdorff space (e.g. a metric space). Let
𝑓 , 𝑔 : 𝑋 → 𝑌 be continuous functions.

1. Show that the set {𝑥 ∈ 𝑋 : 𝑓 (𝑥) = 𝑔(𝑥)} is closed in 𝑋 .

2. Show that if 𝑓 ≡ 𝑔 on a dense subset 𝐷 of 𝑋 , then 𝑓 ≡ 𝑔 on all of 𝑋 .

Proof.

10
 1. We achieve this by proving that the complement {𝑥 ∈ 𝑋 : 𝑓 (𝑥) ≠ 𝑔(𝑥)} is open in 𝑋 . Choos-
ing 𝑥 from this set, we have 𝑓 (𝑥) ≠ 𝑔(𝑥). Because 𝑌 is Hausdorff, we may select dis-
joint neighbourhoods 𝑈 and 𝑉 containing 𝑓 (𝑥) and 𝑔(𝑥), respectively. Then, 𝑓 −1 (𝑈 ) and
𝑔 −1 (𝑉 ) are both open (by continuity) sets containing the point 𝑥. Consequently, 𝑊 :=
𝑓 −1 (𝑈 ) ∩ 𝑔 −1 (𝑉 ) is a neighbourhood of 𝑥. Furthermore, 𝑊 ⊆ {𝑥 ∈ 𝑋 : 𝑓 (𝑥) ≠ 𝑔(𝑥)} by
construction. It follows that {𝑥 ∈ 𝑋 : 𝑓 (𝑥) ≠ 𝑔(𝑥)} is the union of open sets and thence
open, as was asserted.

2. Denote by 𝐸 the set of points in 𝑋 on which 𝑓 = 𝑔. By hypothesis, we have 𝐷 ⊆ 𝐸 ⊆ 𝑋 .

Since 𝐸 is closed, it follows that cl (𝐷) ⊆ 𝐸 ⊆ 𝑋 . Density of 𝐷 then reduces our last
equation to 𝑋 ⊆ 𝐸 ⊆ 𝑋 whence 𝐸 = 𝑋 .
□

1.4 A Quick Look at the Subspace Topology

We first recall the subspace topology

Definition 1.8. Let (𝑋, 𝔗) be a topological space and 𝑌 ⊆ 𝑋 a subset. Then, the subspace topol-
ogy on 𝑌 is given by
𝔗𝑌 = {𝑉 ∩ 𝑌 : 𝑉 ∈ 𝔗} .
It is easy to show (𝑌 , 𝔗𝑌 ) is indeed a topological space.

In a similar vein, we can see that closed sets in a subspace 𝑌 of a topological space (𝑋, 𝔗) are
all of the form 𝐴 ∩ 𝑌 where 𝐴 is closed in 𝑋 . Indeed, we quickly verify this below.

Proposition 1.4. Let (𝑋, 𝔗) be a topological space and let 𝑌 be a subspace of 𝑋 . Then, a subset 𝐴
of 𝑌 is closed in 𝑌 if and only if it takes the form 𝐴 = 𝐹 ∩ 𝑌 for some closed subset 𝐹 of 𝑋 .

Proof. Let 𝐴 ⊆ 𝑌 be a closed subset of 𝑌 . That is, 𝑌 \ 𝐴 is open in 𝑌 . By definition of the subspace
topology, it follows that 𝑌 \ 𝐴 = 𝑈 ∩ 𝑌 for some open subset 𝑈 of 𝑋 (i.e. 𝑈 ∈ 𝔗). Clearly, we may
write 𝐴 = 𝑌 ∩ (𝑋 \ 𝑈 ) where 𝑋 \ 𝑈 is closed in 𝑋 . Hence, every arbitrary closed subset of 𝑌 is
necessarily of the asserted form. It remains only to show that every such set is closed in 𝑌 . To
this end, let 𝐹 ⊆ 𝑋 be closed in 𝑋 . That is, 𝑋 \ 𝐹 is open in 𝑋 (i.e. 𝑋 \ 𝐹 ∈ 𝔗). Then, the set

𝐴 := 𝑌 ∩ 𝐹

is closed in 𝑌 . Indeed, it is clear that 𝐴 ⊆ 𝑌 . Furthermore, 𝑌 \ 𝐴 = 𝑌 ∩ (𝑋 \ 𝐹 ) where 𝑋 \ 𝐹 is open

in the parent topology of 𝑋 . Thus, by definition, 𝑌 \ 𝐴 is open in 𝑌 whence 𝐴 is closed in 𝑌 . This
completes the proof. □

Proposition 1.5. Let (𝑋, 𝔗) be a topological space and 𝑌 is a subspace of 𝑋 .

1. Suppose 𝑌 is open in 𝑋 . If 𝑈 is open in 𝑌 then 𝑈 is open in 𝑋 .

11
 2. Suppose 𝑌 is closed in 𝑋 . If 𝐴 is closed in 𝑌 then 𝐴 is closed in 𝑋 .

Proof. We only prove the first point. Since 𝑈 is open in 𝑌 , it must be of the form 𝑈 = 𝑉 ∩ 𝑌
for some 𝑉 that is open in 𝑋 . However, 𝑌 is open in 𝑋 and topologies are closed under finite
intersections. It follows that 𝑈 is open in 𝑋 . □

Consider the real line R with the usual topology (induced by the metric 𝑑 (𝑥, 𝑦) = |𝑥 − 𝑦|).
Consider then 𝐼 = [0, 1] along with the subspace topology inherited from R. Observe that
[0, 1], [0, 1), (0, 1] and (0, 1) are all open in 𝐼 . However, only the interval (0, 1) is open in R.
On the other hand, by the above proposition, any set that is open in 𝐽 = (0, 1) with respect to the
subspace topology is also open with respect to the parent R-topology.

1.5 Characterizations of continuity

We now return to the concept of continuity; with the aim of developing several characterizations
of continuity, As a primer, recall that a function 𝑓 : 𝑋 → 𝑌 is said to be continuous provided

𝑓 −1 (𝑈 ) is open in 𝑋 for all open sets 𝑈 in 𝑌 .

Taking the complement, we see that

c
𝑓 −1 (𝑈 ) = 𝑓 −1 (𝑈 c ) is closed in 𝑋 for all open sets 𝑈 in 𝑌 .

Thus, an equivalent characterization of continuity is as follows: 𝑓 : 𝑋 → 𝑌 is continuous provided

𝑓 −1 (𝐶) is closed in 𝑋 for all closed sets 𝐶 in 𝑌 .

Theorem 1.6. Let (𝑋, 𝔗) and (𝑌 , 𝔚) be topological spaces and fix a function 𝑓 : 𝑋 → 𝑌 . The
following statements are equivalent.

1. 𝑓 is continuous.

2. For every 𝐴 ⊆ 𝑋 , there holds 𝑓 (cl (𝐴)) ⊆ cl (𝑓 (𝐴)).

3. If 𝐵 is closed in 𝑌 , then 𝑓 −1 (𝐵) is closed in 𝑋 .

4. For every 𝑥 ∈ 𝑋 and every neighbourhood 𝑉 of 𝑓 (𝑥) in 𝑌 , there exists a neighbourhood 𝑈 of

𝑥 such that 𝑓 (𝑈 ) ⊆ 𝑉 .

Proof. We begin with the implication (1) =⇒ (2). Let 𝐴 ⊆ 𝑋 be given and suppose that
𝑦 ∈ 𝑓 (cl (𝐴)). This means that there exists 𝑥 ∈ cl (𝐴) such that 𝑦 = 𝑓 (𝑥). Let now 𝑉 be a
neighbourhood of 𝑦 = 𝑓 (𝑥) and notice that 𝑓 −1 (𝑉 ) is a neighbourhood of 𝑥, by continuity of 𝑓 .
Since 𝑥 belongs to the closure of 𝐴, this means that 𝑓 −1 (𝑉 ) intersects 𝐴. Given that 𝑓 (𝑓 −1 (𝑉 )) ⊆

12
𝑉 , we conclude that 𝑉 intersects 𝑓 (𝐴). As 𝑉 was an arbitrary neighbourhood of 𝑦, we have
𝑦 ∈ cl (𝑓 (𝐴)).
Let us now argue for (2) =⇒ (3). Let 𝐵 ⊆ 𝑌 be a closed subset and let 𝐴 denote 𝑓 −1 (𝐵). We
claim that 𝐴 = cl (𝐴), and for this the only non-trivial inclusion is cl (𝐴) ⊆ 𝐴. Let 𝑥 ∈ cl (𝐴) so
that, by hypothesis,
𝑓 (𝑥) ∈ 𝑓 (cl (𝐴)) ⊆ cl (𝑓 (𝐴)) ⊆ cl (𝐵) = 𝐵.
This certainly gives 𝑥 ∈ 𝑓 −1 (𝐵) = 𝐴. Since 𝑥 was arbitrary, we conclude that cl (𝐴) ⊆ 𝐴 as was
required.
We show (3) =⇒ (1). Let 𝑈 ⊆ 𝑌 be open. Then, 𝑓 −1 (𝑈 c ) is closed as the pre-image of a
closed set. Put otherwise,
𝑓 −1 (𝑈 c ) = 𝑓 −1 (𝑈 ) c
is closed. That is, 𝑓 −1 (𝑈 ) is open in 𝑋 .
It now only remains to check that (1) ⇐⇒ (4). First, assume that 𝑓 is continuous and fix
𝑥 ∈ 𝑋 and a neighbourhood 𝑉 of 𝑓 (𝑥). By continuity, 𝑓 −1 (𝑉 ) is open and contains 𝑥. Thus, taking
𝑈 := 𝑓 −1 (𝑉 ) gives one implication. Conversely, fix an open set 𝑉 in 𝑌 . For any 𝑥 ∈ 𝑓 −1 (𝑉 ), we
can choose an open set 𝑈𝑥 containing 𝑥 such that 𝑓 (𝑈𝑥 ) ⊆ 𝑉 . But, this gives 𝑈𝑥 ⊆ 𝑓 −1 (𝑉 ). Since
𝑈𝑥 ∋ 𝑥, it is then easy to see that
Ø
𝑈𝑥 = 𝑓 −1 (𝑉 ).
𝑥 ∈ 𝑓 −1 (𝑉 )

Since unions of open sets are open, 𝑓 −1 (𝑉 ) is open and this concludes the proof. □

1.6 Compactness
We now turn towards the next pivotal topic of study: the notion of compactness. This begins with
the simple definition of an open covering.

Definition 1.9. Let (𝑋, 𝔗) be a topological space and let 𝐾 be a subset of 𝑋 . A covering of 𝐾 is a
family of subsets of 𝑋 whose union contains 𝐾. An open cover of 𝐾 is a family U of open sets in
𝑋 such that 𝐾 ⊆ 𝑈 ∈U 𝑈 .
Ð

We now come to the definition of compactness.

Definition 1.10. Let (𝑋, 𝔗) be a topological space and 𝐾 a subset of 𝑋 . The set 𝐾 is called compact
in 𝑋 if for every open cover U of 𝐾 by open subsets of 𝑋 , there exists a finite sub-collection of U
whose union contains 𝐾. When the context is understood, we will simply say that 𝐾 is compact.

A topological space (𝑋, 𝔗) is called compact if 𝑋 is compact as a subset of 𝑋 . In this case, we

would simply say that 𝑋 is a compact topological space (or subspace). Equivalently, we have the
following.

13
Definition 1.11. Let (𝑋, 𝔗) be a topological space. We say that 𝑋 is a compact space if for every
open cover U of 𝑋 , by elements of 𝔗, there exists a finite sub-collection of U whose union is equal
to 𝑋 .

Theorem 1.7. Let (𝑋, 𝔗) be a compact topological space. If 𝐹 ⊆ 𝑋 is closed, then 𝐹 is compact in
𝑋 . Hence, 𝐹 is a compact subspace of 𝑋 .

Proof. Let U be an open covering of 𝐹 by subsets of 𝑋 . Since 𝐹 c is open in 𝑋 , the family U ∪ {𝐹 c }

is itself an open covering of 𝑋 , by subsets of 𝑋 . Using that (𝑋, 𝔗) is compact, we may extract a
finite sub-collection
𝑈 1, . . . , 𝑈𝑛 , 𝐹 c, with 𝑈 𝑗 ∈ U,
such that 𝐹 c ∪ 𝑛1 𝑈𝑛 contains 𝑋 , and hence 𝐹 . Since 𝐹 and 𝐹 c are disjoint, we get that 𝐹 ⊆ 𝑛1 𝑈 𝑗 .
Ð Ð
This means that we have found a finite sub-collection of U that covers 𝐹 . Since U was arbitrary,
we conclude that 𝐹 is compact in 𝑋 . □

Theorem 1.8. Let (𝑋, 𝔗) be a Hausdorff topological space. If 𝐾 ⊆ 𝑋 is compact, then 𝐾 is closed in
𝑋.

Proof. We will prove that 𝐾 c = 𝑋 \ 𝐾 is open in 𝑋 . To this end, let us first fix a point 𝑥 0 ∈ 𝐾 c .
For each 𝑦 ∈ 𝐾, we may choose disjoint open sets 𝑈 𝑦 ∋ 𝑦 and 𝑉𝑦 ∋ 𝑥 0 in 𝑋 . Notice then that the
collection
U := 𝑈 𝑦 : 𝑦 ∈ 𝐾


is an open cover of 𝐾 in 𝑋 . By compactness, may choose finitely many points 𝑦1, . . . , 𝑦𝑛 in 𝐾 so

that 𝐾 ⊆ 𝑛1 𝑈 𝑦 𝑗 . Now, the set
Ð
𝑉𝑥 0 := 𝑉𝑦1 ∩ · · · ∩ 𝑉𝑦𝑛
is open in 𝑋 and contains 𝑥 0 . By construction, 𝑉𝑥 0 does not intersect 𝑛1 𝑈 𝑗 . Since 𝑛1 𝑈 𝑗 ⊇ 𝐾,
Ð Ð
we have found an open set 𝑉 containing 𝑥 0 with the property thats 𝑉𝑥 0 ⊆ 𝐾 c . But then, 𝐾 c =
𝑥 0 ∈𝐾 c 𝑉𝑥 0 whence 𝐾 is open.
Ð c □

Proposition 1.9. Let (𝑋, 𝔗) and (𝑌 , 𝔚) be topological spaces. Assume further that 𝑋 is compact
and 𝑌 is Hausdorff. Any continuous function 𝑋 → 𝑌 is closed.

Proof. Let 𝑓 be a continuous function 𝑋 → 𝑌 and fix a closed subset 𝐴 of 𝑋 . Since 𝑋 is compact,
𝐴 is itself compact. By continuity, the image 𝑓 (𝐴) is compact in 𝑌 . Using that 𝑌 Hausdorff, we
see that 𝑓 (𝐴) is closed in 𝑌 . □

Example 1.6. Consider once again any infinite set 𝑋 equipped with the co-finite (finite comple-
ment topology) we have seen in the past, i.e. give 𝑋 the topology

𝔗 := {𝑈 ∈ P (𝑋 ) : 𝑋 \ 𝑈 is finite or 𝑈 = ∅} .

We have seen that the resulting space (𝑋, 𝔗) is “strange” in the sense that it defies much of our
common intuition. In particular, we have seen that any two non-empty open sets 𝑈 , 𝑉 ⊆ 𝑋

14
must necessarily intersect. We concluded from this that (𝑋, 𝔗) was not Hausdorff nor metrizable.
However, even when the underlying set 𝑋 is infinite, every subset of 𝑋 is compact. Certainly, let
𝐴 ⊆ 𝑋 be given and let {𝑈𝛼 }𝛼 ∈𝐼 be an arbitrary open cover of 𝐴. Note that if 𝐴 is empty then
it is automatically contained in any one of these covering sets. Consequently, we may assume
without loss of generality that 𝐴 ≠ ∅. Next, choose any non-empty 𝑈 𝛽 ∈ {𝑈𝛼 }𝛼 ∈𝐼 and note that
𝑈 𝛽 is open. Now, if 𝐴 ⊆ 𝑈 𝛽 then we are done as we hav exhibited a finite subcover. If 𝐴 ⊈ 𝑈 𝛽
then 𝐴 ∩ 𝑈 𝛽c ≠ ∅. In fact, 𝐴 ∩ 𝑈 𝛽c consists of finitely many points, say, 𝑥 1, . . . , 𝑥𝑘 ∈ 𝑋 since 𝑈 𝛽
is open with respect to the co-finite topology. Because every 𝑥𝑖 ∈ 𝐴, there exists some member
𝑈𝛼𝑖 from our cover {𝑈𝛼 }𝛼 ∈𝐼 containing 𝑥𝑖 . However, this means that 𝐴 ⊆ 𝑈 𝛽 ∪ 𝑈𝛼 1 ∪ · · · ∪ 𝑈𝛼𝑘
whence 𝐴 is compact.

1.6.1 Finite Intersection Property

Definition 1.12. A topological space (𝑋, 𝔗) is said to have the finite intersection property if every
family ℭ of closed subsets of 𝑋 having the property that

𝐶 1 ∩ 𝐶 2 ∩ · · · ∩ 𝐶𝑛 ≠ ∅,

for all finite sub-collections {𝐶 1, . . . , 𝐶𝑛 } ⊆ ℭ, also satisfies 𝐶 ∈ℭ 𝐶 ≠ ∅.

Ñ

We have just shown that compact spaces have this property. We now prove that all topological
spaces having this property are, in fact, compact.

Theorem 1.10. Let (𝑋, 𝔗) be a topological space having the finite intersection property. Then, (𝑋, 𝔗)
is compact.

Proof. We shall prove the contrapositive. If (𝑋, 𝔗) is not compact, then one can find an open
covering U of 𝑋 having no finite sub-covering of 𝑋 . Consider now the set

ℭ := {𝑋 \ 𝑈 : 𝑈 ∈ U }

which is a family of closed subsets in 𝑋 . Since no finite sub-collection of U can cover 𝑋 , finite
intersections of elements in ℭ are non-empty. However, 𝑈 ∈U 𝑈 = 𝑋 which means that 𝐶 ∈ℭ 𝐶
Ð Ñ
is empty. Hence, (𝑋, 𝔗) does not have the finite intersection property. □

1.7 Connectedness
Lemma 1.11. Let (𝑋, 𝔗) be a topological space and 𝑌 a connected subspace of 𝑋 . If (𝐴, 𝐵) forms a
separation of 𝑋 , then 𝑌 is contained in one of 𝐴 or 𝐵.

Proof. Let 𝐴 and 𝐵 be two non-empty disjoint open sets whose union is 𝑋 . In the subspace topol-
ogy of 𝑌 , the sets 𝐴 ∩ 𝑌 and 𝐵 ∩ 𝑌 are open. Of course,

(𝐴 ∩ 𝑌 ) ∪ (𝐵 ∩ 𝑌 ) = (𝐴 ∪ 𝐵) ∩ 𝑌 = 𝑌

15
and (𝐴 ∩𝑌 ) ∩ (𝐵 ∩𝑌 ) = ∅ since 𝐴 and 𝐵 are disjoint. This means that 𝐴 ∩𝑌 and 𝐵 ∩𝑌 are disjoint
open subsets of 𝑌 whose union is 𝑌 . Since 𝑌 is connected, one of these is empty. Without harm,
assume that 𝐵 ∩ 𝑌 is empty. This implies that 𝑌 ⊆ 𝐴, as was required. □

Proposition 1.12. Let (𝑋, 𝔗) be a topological space. Then 𝑋 is connected if and only if the only
clopen sets in 𝑋 are 𝑋 and ∅.

Proof. Let 𝐴 ⊆ 𝑋 be a clopen set that is neither empty nor the whole space 𝑋 . Since 𝐴 is closed,
its complement 𝐴c is open in 𝑋 . Since 𝐴 and 𝐴c are disjoint, we have found a separation of the
space 𝑋 . This means that (𝑋, 𝔗) is disconnected. Conversely, assume that (𝑋, 𝔗) is disconnected.
We may then choose non-empty disjoint open sets 𝐴 and 𝐵 whose union is 𝑋 . This means that
𝐴c = 𝐵 so that 𝐴 is clopen. Since 𝐴 ≠ 𝑋 and 𝐴 ≠ ∅, the proof is complete. □

Theorem 1.13. Let (𝑋, 𝔗) be a topological space and let {𝑌𝛼 }𝛼 ∈𝐼 be an indexed family of connected
Ñ Ð
subspaces of 𝑋 . If 𝛼 ∈𝐼 𝑌𝛼 is non-empty, then 𝛼 ∈𝐼 𝑌𝛼 is connected.

Proof. The claim amounts to proving that 𝛼 ∈𝐼 𝑌𝛼 does not admit a separation. By way of contra-
Ð
diction, assume we can find two non-empty disjoint open sets 𝐴 and 𝐵 whose union is 𝛼 ∈𝐼 𝑌𝛼 .
Ð
Let 𝑝 be a point belonging to 𝛼 ∈𝐼 𝑌𝛼 . Without loss of generality, assume that 𝑝 ∈ 𝐴. Since 𝐴 and
Ñ
𝐵 are disjoint, 𝑝 ∉ 𝐵. Since every 𝑌𝛼 is connected, it must lie entirely within one of 𝐴 or 𝐵, by the
previous lemma. As 𝑝 ∈ 𝐴, we get that 𝑌𝛼 ⊆ 𝐴 for each 𝛼 ∈ 𝐼 . That is, 𝛼 ∈𝐼 𝑌𝛼 ⊆ 𝐴. This leaves
Ð
us with 𝐵 = ∅, which is a contradiction. □

Problem 3. Let 𝑋 be a space and 𝑌 a connected subspace of 𝑋 . Will Int (𝑌 ) and 𝜕𝑌 necessarily be
connected? Need the converse also hold true?

Proof. This is a tricky question that is useful to keep in mind. Both implications do not hold, but
it will require some contrived examples to demonstrate this fact.

1. Let 𝐵 1 (1, 0) denote the closed ball of radius 1 centered at (1, 0) in R2 . Let 𝐵 1 (−1, 0) denote
the closed ball of radius 1 centered at (−1, 0). These two connected sets share the point
(0, 0), and hence their union is connected. On the other hand, the interior of this set is the
union of two disjoint open balls, which is disconnected. In a similar vein, the connected set
(0, 1) has {0, 1} as its boundary, which is clearly disconnected.

2. The converse direction also need not hold true. Recall that Q is a totally disconnected sub-
space of R. Since Qc is dense in R, it is clear that Int (Q) = ∅. Thus, Int (Q) is connected.
By definition, one has 𝜕Q = cl (Q) ∩ cl (R \ Q) = R, which is also connected.

This completes the problem. □

Problem 4. Let 𝐴 ⊆ 𝑋 and assume that 𝐶 is a connected subspace of 𝑋 that intersects both 𝐴 and
𝐴c . Prove that 𝐶 also intersects 𝜕𝐴.

16
Proof. We begin by showing that

𝑋 = Int (𝐴) ⊔ Int (𝐴c ) ⊔ 𝜕𝐴.

Note that the use of ‘⊔’ is justified because 𝜕𝐴 never intersects Int (𝐴) (this is Problem ??). Let
now 𝑥 ∈ 𝑋 but assume that 𝑥 ∉ Int (𝐴) ⊔ Int (𝐴c ); we will show that 𝑥 ∈ 𝜕𝐴. Since 𝑥 ∉ Int (𝐴),
any neighbourhood of 𝑥 has non-empty intersection with 𝐴c . Similarly, 𝑥 ∉ Int (𝐴c ) means that
any neighbourhood of 𝑥 intersects 𝐴. However, both of these statements mean that

𝑥 ∈ cl (𝐴) ∩ cl (𝑋 \ 𝐴) = 𝜕𝐴.

Having now verified the aforementioned identity, the proof is easily within reach. If 𝐶 does not
intersect 𝜕𝐴, then it is contained within the union Int (𝐴) ⊔ Int (𝐴c ), which therefore forms a
separation of 𝐶. □

Example 1.7. Let (𝑋, 𝔗) be any infinite set 𝑋 endowed with its co-finite topology. Then, 𝑋 is
connected. By way of contradiction, let us suppose that 𝑋 is disconnected. Then, there exist
disjoint non-empty open sets 𝑈 , 𝑉 ⊆ 𝑋 such that 𝑋 = 𝑈 ⊔ 𝑉 .2 However, we have seen (in earlier
examples) that any two non-empty open sets must necessarily intersect (since 𝑋 is infinite with
the co-finite topology). Thus, 𝑈 ∩ 𝑉 ≠ ∅ and we have our contradiction.
If instead 𝑋 is a finite set with the co-finite topology, then it follows by inspecting the defini-
tion of the co-finite topology that every subset of 𝑋 is open. Put otherwise, when 𝑋 is finite, the
co-finite topology is precisely the discrete topology P (𝑋 ) which always produces a disconnected
space. Thus, imposing the co-finite topology on a set 𝑋 results in a connected space if and only if
𝑋 is infinite.

By popular demand, we also elaborate upon Hatchers’ proof that every closed and bounded
interval in R is compact. Observe that, as a consequence of Theorem 1.7, it follows that every
closed and bounded subset of R is compact.

Theorem 1.14. A closed and bounded interval 𝐼 = [𝑎, 𝑏] is compact in R (with the usual topology).

Proof. Since the case 𝑎 = 𝑏 is trivial, we may assume 𝑎 < 𝑏. Let U = {𝑈𝛼 }𝛼 ∈𝐼 be an open cover of
𝐼 and define

𝐽 := {𝑥 ∈ [𝑎, 𝑏] : [𝑎, 𝑥] is contained in finitely many elements of U } .

Since U covers 𝐼 , let us observe that 𝑎 ∈ 𝐼 implies the existence of some member 𝑈𝛼 ∈ U such
that 𝑎 ∈ 𝑈𝛼 . Furthermore, since 𝑈𝛼 is open, there exists some 𝛿 > 0 such that (𝑎 − 𝛿, 𝑎 + 𝛿) ⊆ 𝑈𝛼 .
In particular, every 𝑥 ∈ [𝑎, 𝑎 + 𝛿) is an element of 𝐽 and so 𝐽 is non-empty.
Define 𝑠 := sup 𝐽 and note that, by definition, 𝑎 < 𝑠 ≤ 𝑏. Next, let us show that 𝑠 ∈ 𝐽 . Indeed,
since 𝑠 ∈ 𝐼 , there is some member 𝑈 𝛽 ∈ U having 𝑠 ∈ 𝑈 𝛽 . Using that 𝑈 𝛽 is open, we can find some
2 We use the notation ‘⊔’ rather than the usual ‘∪’ to emphasize the disjoint nature of the union.

17
𝜀 > 0 such that (𝑠 − 𝜀, 𝑠 + 𝜀) ⊆ 𝑈 𝛽 . Without loss of generality, we may assume that 𝜀 > 0 is so
small that
𝑠 − 𝜀 > 𝑎. (1.3)
In fact, since 𝑠 − 𝜀 is not an upperbound of 𝐽 , there also exists 𝑥 ∈ [𝑎, 𝑏] such that 𝑠 − 𝜀 < 𝑥 ≤ 𝑠
and the interval [𝑎, 𝑥] is contained in the union of finitely many members of U; let us call this
finite sub-collection V. Then,
[𝑎, 𝑠] ⊆ [𝑎, 𝑠 − 𝜀] ∪ (𝑠 − 𝜀, 𝑠 + 𝜀) ⊆ [𝑎, 𝑥] ∪ (𝑠 − 𝜀, 𝑠 + 𝜀)
is contained in the finite sub-collection of U given by V ∪ {𝑈 𝛽 }. Thus, we have shown that 𝑠 ∈ 𝐽
and the proof will be complete if we can show that 𝑏 = 𝑠. Arguing by contradiction, let us assume
that 𝑎 < 𝑠 < 𝑏. Then, by following the argument used above3 , we will have that 𝑠 + 𝜀2 ∈ 𝐽 ,
contradicting the fact that 𝑠 is by definition an upperbound of 𝐽 . This proves that 𝑏 ∈ 𝐽 and so 𝐼 is
contained in the union of finitely many open sets from the covering U. Since U was an arbitrary
open cover, this proves that 𝐼 is compact. □
Lemma 1.15. Give R the usual topology and let 𝑌 ⊆ 𝑋 be a connected subspace. Then, 𝑌 is an
interval.4
Proof. We shall establish the contrapositive, i.e. we show that if 𝑌 is not an interval then it is
disconnected. If 𝑌 ≠ ∅ is not an interval, then there exist points 𝑥, 𝑦 ∈ 𝑌 (with 𝑥 < 𝑦) such that
[𝑥, 𝑦] ⊈ 𝑌 . That is, there exists some point 𝑐 ∈ (𝑥, 𝑦) such that 𝑐 ∉ 𝑌 . Then, 𝑌 = ((−∞, 𝑐) ∩ 𝑌 ) ⊔
((𝑐, ∞) ∩ 𝑌 ). □
Theorem 1.16. Let 𝑛 ≥ 2 be an integer. There does not exist a continuous injection R𝑛 → R. In
particular, R𝑛 is homeomorphic to R if and only if 𝑛 = 1.
Proof. We argue by contradiction; let 𝑓 : R𝑛 → R be an injective continuous function. Note that
since R𝑛 is convex, it is path connected, and therefore connected. By continuity, the image 𝑓 (R𝑛 )
is a connected subspace of R. Invoking the previous lemma, it follows that 𝑓 (R𝑛 ) is an interval.
In fact, since 𝑓 is injective, the interval 𝑓 (R𝑛 ) has end-points 𝑎, 𝑏 with 𝑎 < 𝑏 (otherwise, 𝑎 = 𝑏
and 𝑓 (R𝑛 ) would contain only one point which contradicts injectivity). Thus, we may choose a
point 𝑦0 = 𝑓 (𝑥 0 ) in the interior of 𝑓 (R𝑛 ), i.e. 𝑎 < 𝑦0 < 𝑏, it is clear that the subspace 𝑓 (R𝑛 ) \ {𝑦0 }
is disconnected.
However, the restriction
𝑓 | R𝑛 \{𝑥 0 } : R𝑛 \ {𝑥 0 } → 𝑓 (R𝑛 ) \ {𝑦0 } = 𝑓 (R𝑛 ) \ {𝑓 (𝑥 0 )}
is again a continuous function. Since R𝑛 \ {𝑥 0 } is connected (why?), its continuous image under
𝑓 must also be connected. That is, 𝑓 (R𝑛 ) \ {𝑦0 } is connected – which contradicts our argument
above. □
3 It is the exact same argument, but in (1.3) choose 𝜀 > 0 so small that 𝑎 < 𝑠 − 𝜀 < 𝑠 + 𝜀 < 𝑏. Note that
this is possible in the case because 𝑠 < 𝑏 by assumption.
4 We use the following definition: a subset 𝐼 ⊆ R is an interval if and only if, for each 𝑥, 𝑦 ∈ 𝐼 with

𝑥 < 𝑦, one has [𝑥, 𝑦] ⊆ 𝐼 .

18
1.8 Homeomorphy
Let us now take a brief side-step to introduce an important concept that will bring with it some
very convenient mathematical language. Although we will not really spend time on this during
the tutorials themselves, these notes will hopefully be of use to those who read this section.

Definition 1.13. Let (𝑋, 𝔗) and (𝑌 , 𝔚) be topological spaces. A bijection 𝑓 : 𝑋 → 𝑌 is called a

homeomorphism provided both 𝑓 : 𝑋 → 𝑌 and 𝑓 −1 : 𝑌 → 𝑋 are continuous.

Remark 1.4. We remark that a homeomorphism is automatically both an open map and a closed
map. Indeed, let 𝑓 : 𝑋 → 𝑌 be a homeomorphism of topological spaces (𝑋, 𝔗) and (𝑌 , 𝔚). Then,
given any open set 𝑂 ⊆ 𝑋 , we have that 𝑓 (𝑂) = (𝑓 −1 ) −1 (𝑂) is open in 𝑌 by continuity of 𝑓 −1 .
Similarly, we see that 𝑓 is closed.
We remark that the inverse function 𝑓 −1 : 𝑌 → 𝑋 exists because 𝑓 (in the definition above)
is assumed to be bijective, a priori. Having introduced a new category of function, we now use
this to determine a notion of equivalence (or isomorphism!) for topological spaces.

Definition 1.14. Two topological spaces (𝑋, 𝔗) and (𝑌 , 𝔚) are said to be homeomorphic if there
exists a homeomorphism 𝑋 → 𝑌 . In this case, we often abbreviate this by simply writing 𝑋  𝑌 .5

Before proceeding, several key remarks are in order. We try to present these quickly by sum-
marizing them below and most of their proofs are left as straight-forward exercises to the reader.

(i) Any space (𝑋, 𝔗) is homeomorphic to itself because the identity map id𝑋 : 𝑋 → 𝑋 given
by id𝑋 (𝑥) = 𝑥 is always a homeomorphism.

(ii) If 𝑓 : 𝑋 → 𝑌 is a homeomorphism, then so is its inverse 𝑓 −1 : 𝑌 → 𝑋 . Thus, 𝑋  𝑌 ⇐⇒

𝑌  𝑋.

(iii) If 𝑓 : 𝑋 → 𝑌 and 𝑔 : 𝑌 → 𝑍 are homeomorphisms, then so is the composite mapping

𝑔 ◦ 𝑓 : 𝑋 → 𝑍 (with inverse 𝑓 −1 ◦ 𝑔 −1 : 𝑍 → 𝑋 ). Thus, if 𝑋  𝑌 and 𝑌  𝑍 then 𝑋  𝑍 .

(iv) These last three properties (resp. reflexivity, symmetry, transitivity) imply that being home-
omorphic defines an equivalence relation on topological spaces. That is, we declare two
topological spaces (𝑋, 𝔗) and (𝑌 , 𝔚) to be “the same” when they are homeomorphic. 6

(v) In what follows let (𝑋, 𝔗) and (𝑌 , 𝔚) be two homeomorphic topological spaces.

• 𝑋 is (path) connected if and only if 𝑌 is (path) connected;

• 𝑋 is compact if and only if 𝑌 is compact;
5 For those using LAT X, the symbol ‘’ is the congruence symbol given by the command ‘\cong’.
E
6 We warn the reader that these spaces are deemed equivalent only in a context that is purely topological;

and 𝑋 may carry additional structure that 𝑌 does not – or vice versa.

19
 • 𝑋 is metrizable if and only if 𝑌 is metrizable;
• 𝑋 is completely metrizable if and only if 𝑌 is completely metrizable;
• 𝑋 is Hausdorff if and only if 𝑌 is Hausdorff.
However, not everything is preserved via homeomorphism. For instance, if (𝑋, 𝑑𝑋 ) and
(𝑌 , 𝑑𝑌 ) are metric spaces with a homeomorphism 𝑓 : 𝑋 → 𝑌 , we cannot guarantee that
𝑑𝑋 (𝑥 1, 𝑥 2 ) = 𝑑𝑌 (𝑓 (𝑥 1 ), 𝑓 (𝑥 2 )) for 𝑥 1, 𝑥 2 ∈ 𝑋 (i.e. 𝑓 may not be distance preserving7 ). A
simple example of such a case is the homeomorphism
𝑓 : [0, 1] → [0, 2], 𝑥 ↦→ 2𝑥,
where [0, 1], [0, 2] inherit the metric of R.
Theorem 1.17. Let (𝑋, 𝔗) be a compact topological space and (𝑌 , 𝔚) a Hausdorff topological space.
Let 𝑓 : 𝑋 → 𝑌 be a bijective continuous function. Then, 𝑓 is a homeomorphism.
Proof. The only non-trivial property to prove is the continuity of 𝑓 −1 . By Theorem 1.6, this is
equivalent to showing that 𝑓 maps closed sets to closed sets (that is, 𝑓 is a closed map). This is
true by Proposition 1.9. Since the proof is short, we re-state it here:
Fix a closed subset 𝐴 of 𝑋 . Since 𝑋 is compact, 𝐴 is itself compact. By continuity, the image
𝑓 (𝐴) is compact in 𝑌 . Using that 𝑌 Hausdorff, we see that 𝑓 (𝐴) is closed in 𝑌 . □
Example 1.8. If we forgo the compactness assumption on the domain space, all bets are off in the
previous theorem’s conclusion. For instance, let 𝑋 = R with the discrete topology and 𝑌 = R with
the usual (Euclidean) topology. Consider the identity map 𝑓 (𝑥) = id𝑋 (𝑥) = 𝑥 from 𝑋 → 𝑌 . Since
every subset of 𝑋 is open with respect to its discrete topology, it is automatic that 𝑓 is continuous.
Furthermore, 𝑓 is
−1obviously a bijection. However, 𝑓 : 𝑌 → 𝑋 is not continuous. To see this,
−1

notice that 𝑓 −1 ({0}) = 𝑓 ({0}) = {0} is not open in 𝑌 , whereas {0} is open in 𝑋 . It follows
that 𝑓 is discontinuous and is thus 𝑓 is not a homeomorphism.
−1

Theorem 1.18. Let 𝑋 1, . . . , 𝑋𝑛 be connected topological spaces and let 𝑋 denote the product 𝑋 1 ×
· · · × 𝑋𝑛 with the product topology. Then, 𝑋 is connected.
Proof. By induction, it suffices to show that 𝑋 1 × 𝑋 2 is connected whenever 𝑋 1,2 are themselves
connected. To this end, let us first fix a point (𝑎, 𝑏) ∈ 𝑋 1 × 𝑋 2 . The reader may easily verify that
𝑋 1 × {𝑏} is homeomorphic to 𝑋 1 . Similarly, for every 𝑥 ∈ 𝑋 1 it can be shown that {𝑥 } × 𝑋 2  𝑋 2 .
Thus, by the previous theorem, both 𝑋 1 × {𝑏} and {𝑥 } × 𝑋 2 will be connected, for each 𝑥 ∈ 𝑋 1 .
Noticing that
(𝑥, 𝑏) ∈ (𝑋 1 × {𝑏}) ∩ ({𝑥 } × 𝑋 2 ) ,
we apply Theorem 1.13 to deduce that the “slice”
Γ𝑥 := (𝑋 1 × {𝑏}) ∪ ({𝑥 } × 𝑋 2 )
is also connected. Finally, by observing that (𝑎, 𝑏) ∈ 𝑥 ∈𝑋1 Γ𝑥 , we see that this same theorem
Ñ
implies that 𝑥 ∈𝑋1 Γ𝑥 = 𝑋 1 × 𝑋 2 is connected.
Ð
□
7A continuous function between two metric spaces that preserves distance is called an isometry.

20
1.9 Completeness and Compactness in Metric Spaces
Let us recall that a metric space (𝑋, 𝑑) is called complete when all of its Cauchy sequences con-
verge. Furthermore, a contraction of a metric space is a mapping 𝑓 : 𝑋 → 𝑋 such that there exists
𝑐 ∈ [0, 1) having the property that 𝑑 (𝑓 (𝑥), 𝑓 (𝑦)) ≤ 𝑐𝑑 (𝑥, 𝑦) for all 𝑥, 𝑦 ∈ 𝑋 . In the lectures, the
following major result was proven, via iterative means:

Theorem 1.19 (Banach Fixed Point Theorem). Let (𝑋, 𝑑) be a complete metric space and 𝑓 : 𝑋 → 𝑋
be a contraction. Then, there exists a unique fixed point of 𝑓 in 𝑋 . That is, there exists a unique point
𝜉 ∈ 𝑋 such that 𝑓 (𝜉) = 𝜉.

Example 1.9. Let (𝑋, 𝑑) be a complete metric space and let 𝑓 : 𝑋 → 𝑋 be given. Given 𝑘 ∈ N,
let 𝑓 (𝑘 ) denote the 𝑘-fold composition

𝑓 (𝑘 ) (𝑥) := (𝑓 ◦ · · · ◦ 𝑓 ) (𝑥).
| {z }
𝑘 −𝑡𝑖𝑚𝑒𝑠

Suppose there exists 𝑘 ∈ N such that 𝑓 (𝑘 ) is a contraction. Then, 𝑓 still possesses a unique fixed
point in 𝑋 . Indeed, to see the existence of such a point, observe that 𝑓 (𝑘 ) has a fixed point 𝜉 ∈ 𝑋
by virtue of the Banach Fixed Point Theorem. But then,
 
𝑓 (𝑘 ) (𝜉) = 𝜉 =⇒ 𝑓 𝑓 (𝑘 ) (𝜉) = 𝑓 (𝜉) ⇐⇒ 𝑓 (𝑘 ) (𝑓 (𝜉)) = 𝑓 (𝜉),

meaning that 𝑓 (𝜉) is a fixed point for 𝑓 (𝑘 ) whenever 𝜉 is. By the uniqueness of the fixed point,
we infer that 𝑓 (𝜉) = 𝜉. Hence, 𝑓 has a fixed point. Let now 𝜂 ∈ 𝑋 be any fixed point of 𝑓 , i.e.
𝑓 (𝜂) = 𝜂. Clearly,

𝑓 (𝑘 ) (𝜂) = 𝑓 (𝑘 −1) (𝑓 (𝜂)) = 𝑓 (𝑘 −1) (𝜂) = · · · = 𝑓 (𝜂) = 𝜂

whence 𝜂 is a fixed point of 𝑓 (𝑘 ) . Since 𝑓 (𝑘 ) has a unique fixed point, it follows that 𝑓 has a unique
fixed point.

Complete metric spaces also satisfy the following intersection property, which is highly rem-
iniscent of the finite intersection property we previously encountered (both within these notes
and your assignments) for compact topological spaces.

Theorem 1.20 (Cantor’s Intersection Theorem for Complete Metric Spaces). Let (𝑋, 𝑑) be a com-
plete metric space and let {𝐹𝑛 }𝑛∈N be a decreasing family of non-empty closed subsets of 𝑋 . If, in
addition,
lim diam(𝐹𝑛 ) = 0,
𝑛→∞

where diam(𝐴) = sup {𝑑 (𝑥, 𝑦) : 𝑥, 𝑦 ∈ 𝐴}, then the intersection 𝑛∈N 𝐹𝑛 is non-empty and consists
Ñ
of exactly one point.

21
Proof. First, we show that 𝑛∈N 𝐹𝑛 is non-empty via a constructive argument. Since each 𝐹𝑛 ≠ ∅
Ñ
by assumption, we may extract some 𝑥𝑛 from each 𝐹𝑛 , thereby forming a sequence (𝑥𝑛 ). Further-
more, since 𝐹𝑛 ⊇ 𝐹𝑛+1 for each 𝑛 ∈ N, it follows that 𝑥𝑚 ∈ 𝑥𝑛 if 𝑚 ≥ 𝑛. Thus, if 𝑚 ≥ 𝑛 there
holds
𝑑 (𝑥𝑛 , 𝑥𝑚 ) ≤ diam(𝐹𝑛 ).
Since diam(𝐹𝑛 ) → 0, it readily follows that the sequence (𝑥𝑛 ) is Cauchy. Using that (𝑋, 𝑑) is a
complete space, there exists a point 𝑥 ∈ 𝑋 such that 𝑥𝑛 → 𝑥. Furthermore, 𝑥 is a limit of every
tail sequence (𝑥𝑚 )𝑚≥𝑛 . Put otherwise, for each 𝑛 ∈ N there is a sequence in 𝐹𝑛 converging to 𝑥.
Since 𝐹𝑛 is closed, it must contain its limit points and so 𝑥 ∈ 𝐹𝑛 . As 𝑛 ∈ N is arbitrary, this implies
that 𝑥 ∈ 𝑛∈N 𝐹𝑛 .
Ñ
Next, we show that this intersection can only contain a single point. Let 𝑥, 𝑦 ∈ 𝑛∈N 𝐹𝑛 be
Ñ
given; we claim that 𝑦 = 𝑥. Indeed, if 𝑥, 𝑦 ∈ 𝑛∈N 𝐹𝑛 then 𝑥, 𝑦 ∈ 𝐹𝑛 for all 𝑛 ∈ N. Hence,
Ñ
𝑑 (𝑥, 𝑦) ≤ diam(𝐹𝑛 ) for all 𝑛 ∈ N. Passing to the limit as 𝑛 → ∞, we infer that 0 ≤ 𝑑 (𝑥, 𝑦) ≤ 0
whence 𝑥 = 𝑦. □

We now provide a practical example of a complete metric space.

Proposition 1.21. Let (𝑋, 𝑑) be a compact metric space and denote by 𝐶 (𝑋 ) the real vector space
of continuous functions 𝑋 → R. Then, 𝐶 (𝑋 ) is a complete metric space when given the norm

∥ 𝑓 ∥ := sup |𝑓 (𝑥)| .
𝑥 ∈𝑋

Proof. We leave it as an exercise to check that (𝐶 (𝑋 ), ∥·∥) is indeed a normed-space. To see that
it is complete as a metric space, we must show that every Cauchy sequence is convergent. To this
end, let (𝑓𝑛 ) be a Cauchy sequence in 𝐶 (𝑋 ). By definition, given 𝜀 > 0, there exists 𝑁 ∈ N such
that
∥ 𝑓𝑛 − 𝑓𝑚 ∥ = sup |𝑓𝑛 (𝑥) − 𝑓𝑛 (𝑥)| < 𝜀, ∀𝑛, 𝑚 ≥ 𝑁 (1.4)
𝑥 ∈𝑋

In particular, for any 𝑥 ∈ 𝑋 and every 𝑛, 𝑚 ≥ 𝑁 we find that |𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥)| < 𝜀. Since 𝜀 > 0 was
arbitrary, this shows that (𝑓𝑛 (𝑥))𝑛∈N is Cauchy in R, for each fixed 𝑥 ∈ 𝑋 . Since R is complete,
every Cauchy sequence is convergent. Especially, there exists some number, which we denote
𝑓 (𝑥), such that
𝑛→∞
𝑓𝑛 (𝑥) −−−−→ 𝑓 (𝑥).
This defines a function 𝑓 : 𝑋 → R. Furthermore, given 𝜀 > 0 is 𝑁 is as in equation (1.4) then

|𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥)| < 𝜀 ∀𝑥 ∈ 𝑋 and 𝑛 ≥ 𝑁 .

Taking the limit as 𝑚 → ∞ and using that 𝑓 (𝑥) := lim𝑚→∞ 𝑓𝑚 (𝑥), we find that

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ 𝜀, ∀𝑥 ∈ 𝑋 and 𝑛 ≥ 𝑁 .

22
Put otherwise, we have shown that there exists a function 𝑓 : 𝑋 → R such that, given any 𝜀 > 0,
we can find 𝑁 ∈ N such that
∥ 𝑓𝑛 − 𝑓 ∥ = sup |𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ 𝜀. (1.5)
𝑥 ∈𝑋

The proof will now be compete if we can show that 𝑓 ∈ 𝐶 (𝑋 ), i.e. that 𝑓 is continuous on 𝑋 . To
this end, fix a point 𝑐 ∈ 𝑋 and let 𝜀 > 0. Using (1.5), we can find 𝑁 ∈ N such that
𝜀
∥𝑓𝑛 − 𝑓 ∥ < , ∀𝑛 ≥ 𝑁 .
3
Now, since 𝑓𝑁 is continuous on 𝑋 , there exists 𝛿 > 0 such that 𝑑 (𝑥, 𝑐) < 𝛿 implies
𝜀
|𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑐)| < .
3
Combining all of this, we see that for any 𝑛 ≥ 𝑁 and any 𝑥 ∈ 𝑋 with 𝑑 (𝑥, 𝑐) < 𝛿 there holds
|𝑓 (𝑥) − 𝑓 (𝑐)| ≤ |𝑓 (𝑥) − 𝑓𝑁 (𝑥)| + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑐)| + |𝑓𝑁 (𝑐) − 𝑓 (𝑐)|
≤ 2 ∥ 𝑓𝑁 − 𝑓 ∥ + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑐)|
2𝜀 𝜀
< +
3 3
= 𝜀.
Hence, 𝑓 ∈ 𝐶 (𝑋 ) and the proof is complete. □
Theorem 1.22. Let 𝑋 be compact metric space. Suppose 𝑓 : 𝑋 → 𝑋 is a function such that
𝑑 (𝑓 (𝑥), 𝑓 (𝑦)) ≥ 𝑑 (𝑥, 𝑦) for all 𝑥, 𝑦 ∈ 𝑋 .8 Then, 𝑓 is a surjective isometry.
Proof. Let 𝑥 ∈ 𝑋 and consider the sequence 𝑥 1 = 𝑓 (𝑥) and 𝑥𝑛+1 = 𝑓 (𝑥𝑛 ). Put otherwise, if 𝑓 (𝑛)
denotes the composition of 𝑓 with itself 𝑛 times, then 𝑥𝑛 = 𝑓 (𝑛) (𝑥). Now, suppose that (𝑥𝑛 )𝑛 has
a convergent subsequence (𝑥𝑛𝑘 )𝑘 . In particular, since the latter sequence is Cauchy,

𝑘→∞
𝑑 𝑥𝑛𝑘+1 , 𝑥𝑛𝑘 −−−−→ 0.
On the other hand, observe that
   
𝑑 𝑥𝑛𝑘+1 , 𝑥𝑛𝑘 = 𝑑 𝑓 (𝑛𝑘+1 ) (𝑥), 𝑓 (𝑛𝑘 ) (𝑥) ≥ 𝑑 𝑓 (𝑛𝑘+1 −1) (𝑥), 𝑓 (𝑛𝑘 −1) (𝑥)

≥ ···
 
≥ 𝑑 𝑓 (𝑛𝑘+1 −𝑛𝑘 ) (𝑥), 𝑥


= 𝑑 𝑥𝑛𝑘+1 −𝑛𝑘 , 𝑥 .
In particular, since 𝑑 𝑥𝑛𝑘+1 , 𝑥𝑛𝑘 → 0, we see that 𝑥𝑛𝑘+1 −𝑛𝑘 → 𝑥 as 𝑘 → ∞. We summarize our

work so far in the form of a claim.

8 Such a mapping is called an expansion of 𝑋 . Alternatively, 𝑓 may be called an expanding map.

23
 Claim: Given 𝑥 ∈ 𝑋 , let (𝑥𝑛 )𝑛 be the sequence defined by 𝑥 1 = 𝑓 (𝑥) and 𝑥𝑛+1 =
𝑓 (𝑥𝑛 ), i.e. 𝑥𝑛 = 𝑓 (𝑛) (𝑥). If (𝑥𝑛𝑘 )𝑘 is a convergent subsequence of (𝑥𝑛 )𝑛 then (𝑥𝑛𝑘+1 −𝑛𝑘 )𝑘
converges to 𝑥.

The above claim has some immediate consequences. First, since 𝑋 is compact, any sequence
has a convergence subsequence. In particular, given 𝑥 ∈ 𝑋 the associated sequences (𝑥𝑛 )𝑛 defined
by 𝑥𝑛 = 𝑓 (𝑛) (𝑥) has a convergent subsequence (𝑥𝑛𝑘 )𝑘 . Next, let 𝑦 ∈ 𝑋 be another element and
consider also the sequence (𝑦𝑛 )𝑛 given by 𝑦𝑛 = 𝑓 (𝑛) (𝑦). Then, its subsequence (𝑦𝑛𝑘 )𝑘 has a con-
vergent subsequence (𝑦𝑛𝑘 𝑗 ) 𝑗 . Then, using that subsequences of convergent sequences converge,
we see that
(𝑥𝑛𝑘 𝑗 ) 𝑗 , (𝑦𝑛𝑘 𝑗 ) 𝑗
are both convergent subsequences of, respectively, (𝑥𝑛 )𝑛 and (𝑦𝑛 )𝑛 . For notational convenience,
we re-label these sequences to
(𝑥𝑛 𝑗 ) 𝑗 , (𝑦𝑛 𝑗 ) 𝑗
By our claim, 𝑥𝑛 𝑗 +1 −𝑛 𝑗 → 𝑥 and 𝑦𝑛 𝑗 +1 −𝑛 𝑗 → 𝑦 as 𝑗 → ∞. Thus,
   
𝑑 (𝑥, 𝑦) = lim 𝑑 𝑥𝑛 𝑗 +1 −𝑛 𝑗 , 𝑦𝑛 𝑗 +1 −𝑛 𝑗 = lim 𝑑 𝑓 (𝑛 𝑗 +1 −𝑛 𝑗 ) (𝑥), 𝑓 (𝑛 𝑗 +1 −𝑛 𝑗 ) (𝑦)
𝑗→∞ 𝑗→∞

≥ 𝑑 (𝑓 (𝑥), 𝑓 (𝑦))

On the other hand, since we have assumed that 𝑑 (𝑥, 𝑦) ≤ 𝑑 (𝑓 (𝑥), 𝑓 (𝑦)) we conclude that, in fact,
equality holds. That is, 𝑑 (𝑥, 𝑦) = 𝑑 (𝑓 (𝑥), 𝑓 (𝑦)). Since 𝑥, 𝑦 ∈ 𝑋 were arbitrary, this means that 𝑓
is an isometry. In particular, 𝑓 is (uniformly) continuous.
It remains only to verify that 𝑓 is surjective. Put otherwise, we must show that 𝑓 (𝑋 ) = 𝑋 . To
see this, first observe that 𝑓 (𝑋 ) is dense in 𝑋 . Indeed, for any 𝑥 ∈ 𝑋 we consider the sequence
(𝑥𝑛 )𝑛 ⊆ 𝑓 (𝑋 ) given by 𝑥𝑛 = 𝑓 (𝑛) (𝑥). Then, since 𝑋 is compact, we may extract a convergent
subsequence (𝑥𝑛𝑘 )𝑘 . By our claim, 𝑥𝑛𝑘+1 −𝑛𝑘 → 𝑥. That is, we found a sequence in 𝑓 (𝑋 ) converging
to 𝑥 so, indeed, 𝑓 (𝑋 ) is dense in 𝑋 . Put otherwise, the closure of 𝑓 (𝑋 ) is all of 𝑋
On the other hand, we have shown that 𝑓 is continuous. Thus, since 𝑋 is compact, its image
𝑓 (𝑋 ) under 𝑓 is also compact. Then, because 𝑋 is Hausdorff, we see that 𝑓 (𝑋 ) ⊆ 𝑋 is closed.
Thus, the closure of 𝑓 (𝑋 ) is itself. But, by density, the closure of 𝑓 (𝑋 ) is 𝑋 . Ergo, 𝑓 (𝑋 ) = 𝑋 and
we are done. □

1.10 Baire Category Theorem and Basic Applications

Theorem 1.23 (Baire Category Theorem). Let (𝑋, 𝔗) be a completely metrizable space and suppose
Ñ
{𝑈𝑛 }𝑛 is a countable collection of open dense sets in 𝑋 . Then 𝑛 𝑈𝑛 is dense in 𝑋 .

Proof. Let 𝑑 be a complete metric on 𝑋 inducing the topology 𝔗. Let 𝑊 ⊆ 𝑋 be a non-empty

open set; the statement amounts to showing that 𝑊 ∩ 𝑛 𝑈𝑛 is non-empty. To this end, notice
Ñ

24
that 𝑊 ∩ 𝑈 1 is non-empty (since 𝑈 1 is dense in 𝑋 ) and thus contains a point, say 𝑥 1 . There then
exists 0 < 𝑟 1 < 2−1 such that
𝐵(𝑥 1, 𝑟 1 ) ⊆ 𝑈 1 ∩ 𝑊 .
We now proceed inductively as follows:
Given 𝑥𝑘 and 𝑟𝑘 , we consider the intersection 𝐵(𝑥𝑘 , 𝑟𝑘 ) ∩ 𝑈𝑘+1 , which is non-empty by
assumption. Thus, there exists a point
𝑥𝑘+1 ∈ 𝐵(𝑥𝑘 , 𝑟𝑘 ) ∩ 𝑈𝑘+1
where this intersection is an open set. Once again, we may choose 𝑟𝑘+1 > 0 such that
𝐵(𝑥𝑘+1, 𝑟𝑘+1 ) ⊆ 𝐵(𝑥𝑘 , 𝑟𝑘 ) ∩ 𝑈𝑘+1 .
Without harm, we may choose 𝑟𝑘+1 < 2− (𝑘+1) .
Now, for every 𝑛 > 1 we have found
𝑥𝑛 ∈ 𝐵(𝑥𝑛−1, 𝑟𝑛−1 ) ∩ 𝑈𝑛 ⊆ [𝐵(𝑥𝑛−2, 𝑟𝑛−2 ) ∩ 𝑈𝑛−1 ] ∩ 𝑈𝑛
𝑛
Ù
⊆ 𝐵(𝑥 1, 𝑟 1 ) ∩ 𝑈𝑘
𝑘=2
𝑛
Ù
⊆𝑊 ∩ 𝑈𝑘 .
𝑘=1

which means that we are done if the sequence of sets {𝑈𝑛 } is finite. Otherwise, we have con-
structed a sequence (𝑥𝑛 )𝑛=1
∞ of points in 𝑋 . We now claim that this sequence is Cauchy. Indeed,

let 𝜀 > 0 be given and let 𝑁 ∈ N be such that 2−𝑁 < 𝜀/2. If 𝑛, 𝑚 ≥ 𝑁 then
𝑥𝑛 , 𝑥𝑚 ∈ 𝐵(𝑥 𝑁 , 𝑟 𝑁 )
whence it follows that 𝑑 (𝑥𝑛 , 𝑥𝑚 ) < 𝜀. Since (𝑋, 𝑑) is complete, there exists a point 𝑥 ∈ 𝑋 such
that lim𝑛→∞ 𝑥𝑛 = 𝑥. Now, for every 𝑁 ∈ N our construction gives
𝑥𝑛 ∈ 𝐵(𝑥 𝑁 , 𝑟 𝑁 ), ∀𝑛 ≥ 𝑁 .
Passing to the limit, we find that
𝑥 ∈ 𝐵(𝑥 𝑁 , 𝑟 𝑁 ) ⊆ 𝐵(𝑥 𝑁 −1, 𝑟 𝑁 −1 ) ∩ 𝑊 ⊆ 𝑈 𝑁 ∩ 𝑊 .
This implies that 𝑥 ∈ 𝑊 ∩ 𝑛 𝑈𝑛 as was required.
Ñ
□
Remark 1.5. This result remains valid if (𝑋, 𝔗) is instead a locally compact Hausdorff space (ab-
breviated LCH). That is, if (𝑋, 𝔗) is a Hausdorff space such that, at each point 𝑥 ∈ 𝑋 , there
exists an open set 𝑈 ∋ 𝑥 such that cl(𝑈 ) is compact. The proof strategy is similar, but requires
modifications (what are Cauchy sequences without a metric?). One cannot deduce the result for
completely metrizable spaces from that for LCH-spaces.

25
1.10.1 Some Examples and Applications
Before we delve into some interesting applications of the Baire Category theorem, we require
some terminology.

Definition 1.15. Let (𝑋, 𝔗) be a topological space and let 𝐸 ⊆ 𝑋 . We say that 𝐸 is nowhere dense
in 𝐸 if cl (𝐸) has empty interior. That is, if for each 𝑥 ∈ cl (𝐸) and every neighbourhood 𝑈 ∋ 𝑥,
one has 𝑈 ⊈ cl (𝐸).

With this language, the Baire-Category theorem may be rephrased:

Corollary 1.24. Let (𝑋, 𝔗) be a completely metrizable space. Then, 𝑋 is not the countable union of
nowhere dense sets.

Proof. We argue by contradiction. Suppose that one may write 𝑋 = 𝑛 𝐸𝑛 , where this union is
Ð
countable and each 𝐸𝑛 ⊆ 𝑋 is nowhere dense. Then, cl (𝐸𝑛 ) has empty interior which implies that
cl (𝐸𝑛 ) c is open and dense in 𝑋 . Indeed, cl (𝐸𝑛 ) c is open by definition. To see that this set is dense,
let 𝑈 ≠ ∅ be an open subset of 𝑋 and note that 𝑈 ⊈ cl (𝐸𝑛 ) because cl (𝐸𝑛 ) has empty interior.
Therefore, 𝑈 ∩ cl (𝐸𝑛 ) c ≠ ∅. Put otherwise, cl (𝐸𝑛 ) c intersects every non-empty open subset of
𝑋 , i.e. cl (𝐸𝑛 ) c is dense in 𝑋 . Finally, appealing to the Baire Category Theorem ensures that the
countable intersection of open dense sets 𝑛 cl (𝐸𝑛 ) c remains dense in 𝑋 . Especially, 𝑛 cl (𝐸𝑛 ) c
Ñ Ñ
is non-empty whence
" #c
Ø Ø Ù
𝑋 = 𝐸𝑛 ⊆ cl (𝐸𝑛 ) = cl (𝐸𝑛 ) c
≠ 𝑋.
𝑛 𝑛 𝑛

which is a contradiction. □

We obtain as a direct consequence of this an alternative proof that R is uncountable.

Corollary 1.25. Let (𝑋, 𝔗) be a completely metrizable space such that singletons are not open.9
Then, 𝑋 is uncountable. In particular, R is uncountable.

Proof. Observe that if a singleton {𝑥 } is not open then it must have empty interior. Otherwise,
there would exist a non-empty open set 𝑂 ⊆ {𝑥 } whence 𝑂 = {𝑥 } and {𝑥 } is open.
Now, we address the statement itself: we argue by contradiction and assume that 𝑋 is count-
able. Then, we may enumerate its elements {𝑥𝑛 }𝑛 . In particular, 𝑋 = 𝑛 {𝑥𝑛 } where each {𝑥𝑛 } is
Ð
closed and with empty interior. Put otherwise, each {𝑥𝑛 } is nowhere dense in 𝑋 . Thus, we have
written 𝑋 as the countable union of nowhere dense sets which contradicts the Baire Category
Theorem. □

Some surprising results can even be deduced about the rationals Q in R.

9A topological space (𝑋, 𝔗) such that no singleton {𝑥 } is open is called perfect. That is, a perfect space
is one with no isolated points. In some texts, these spaces are also referred to as crowded spaces.

26
Lemma 1.26. There do not exist countably many open sets10 {𝑈𝑛 }𝑛 in R such that
Ñ
𝑛 𝑈𝑛 = Q.

Proof. We proceed by way of contradiction: assume that there exists a countable family {𝑈𝑛 }𝑛
of open subsets of R such that Q = 𝑛 𝑈𝑛 . By allowing for repetitions in our collection, we
Ñ
may assume without loss of generality that the family {𝑈𝑛 }𝑛 = {𝑈𝑛 }𝑛∈N is countably infinite.
Furthermore, it is clear that each 𝑈𝑛 contains Q whence each 𝑈𝑛 is open and dense in R. Let now
{𝑟𝑛 }𝑛∈N be an enumeration of Q (which is possible because Q is countably infinite) and define,
for each 𝑛 ∈ N, the open set
𝑉𝑛 := 𝑈𝑛 \ {𝑟𝑛 } = 𝑈𝑛 ∩ {𝑟𝑛 }c .
We note that 𝑉𝑛 is also dense in R. Indeed, given a non-empty open set 𝑊 ⊆ R, the set 𝑊 \ {𝑟𝑛 }
is both open and non-empty. Since 𝑈𝑛 ⊇ Q is dense, it follows that

(𝑊 \ {𝑟𝑛 }) ∩ 𝑈𝑛 ≠ ∅

thereby implying that 𝑊 ∩ 𝑉𝑛 = 𝑊 ∩ (𝑈𝑛 \ {𝑟𝑛 }) is non-empty. Hence, 𝑉𝑛 is dense in R for each
𝑛 ∈ N. By virtue of the Baire Category Theorem, the countable intersection 𝑛∈N 𝑉𝑛 is dense in
Ñ
R. In particular, this intersection is non-empty. On the other hand,
! !
Ù Ù Ù Ø
𝑉𝑛 = (𝑈𝑛 \ {𝑟𝑛 }) = 𝑈𝑛 \ {𝑟𝑛 }
𝑛∈N 𝑛∈N 𝑛∈N 𝑛∈N
!
Ù
= 𝑈𝑛 \ Q
𝑛∈N
=Q\Q
= ∅,

which is an obvious contradiction. □

Let 𝑓 : R → R be an arbitrary function. We define the continuity set of 𝑓 (labelled C (𝑓 )) as

the set of points in R at which the function 𝑓 is continuous. Symbolically, we define

C (𝑓 ) := {𝑥 ∈ R : 𝑓 is continuous at 𝑥 } .

As it turns out, C (𝑓 ) is always the countable intersection of open sets (i.e. a 𝐺𝛿 set). Indeed, it
follows from the 𝜀 − 𝛿 definition of continuity that
Ù
C (𝑓 ) = 𝐴𝑛 (1.6)
𝑛∈N

10 Let (𝑋, 𝔗) be a topological space

Ñ and 𝐴 ⊆ 𝑋 . We say that 𝐴 is a 𝐺𝛿 set if there exist countable many
open sets {𝑈𝑛 }𝑛 such that 𝐴 = 𝑛 𝑈𝑛 . That is, a set is called 𝐺𝛿 if it can be written as the countable
intersection of open sets. Analogously, 𝐴 is called an 𝐹𝜎 set provided it is the countable union of closed
sets. In short, this lemma states that Q is not 𝐺𝛿 in R.

27
where
1
 
𝐴𝑛 := 𝑥 ∈ R : ∃𝛿 > 0 s.t. |𝑓 (𝑠) − 𝑓 (𝑡)| < for all 𝑠, 𝑡 ∈ R with (𝑥 − 𝛿, 𝑥 + 𝛿) . (1.7)
𝑛
The proof of (1.6) is left as an exercise to the reader.11 However, we shall prove that that each 𝐴𝑛
is open in R. Certainly, fix a point 𝑥 ∈ 𝐴𝑛 . By definition of the set 𝐴𝑛 , there exists 𝛿 > 0 such that
1
|𝑓 (𝑠) − 𝑓 (𝑡)| < (1.8)
𝑛
for all 𝑠, 𝑡 ∈ (𝑥 − 𝛿, 𝑥 + 𝛿). Since we wish to show that 𝐴𝑛 is open, we seek to exhibit some 𝜀-
neighbourhood (𝑥 − 𝜀, 𝑥 + 𝜀) of 𝑥 such that (𝑥 − 𝜀, 𝑥 + 𝜀) ⊆ 𝐴𝑛 . To see that such an inclusion is
possible, notice that, since (𝑥 − 𝛿, 𝑥 + 𝛿) is an open set, there exists for each 𝑦 ∈ (𝑥 − 𝛿, 𝑥 + 𝛿)
some 𝑟 > 0 such that (𝑦 − 𝑟, 𝑦 + 𝑟 ) ⊆ (𝑥 − 𝛿, 𝑥 + 𝛿). Furthermore, because (1.8) holds for all
𝑠, 𝑡 ∈ (𝑥 −𝛿, 𝑥 +𝛿), this same identity must be valid in (𝑦 −𝑟, 𝑦 +𝑟 ) ⊆ (𝑥 −𝛿, 𝑥 +𝛿) whence 𝑦 ∈ 𝐴𝑛 .
Using that 𝑦 ∈ (𝑥 − 𝛿, 𝑥 + 𝛿), we infer that (𝑥 − 𝛿, 𝑥 + 𝛿) ⊆ 𝐴𝑛 . In summary, we have shown that
each 𝐴𝑛 is an open set.
Corollary 1.27. There does not exist a function 𝑓 : R → R such that C (𝑓 ) = Q. That is, no function
R → R can have Q as its continuity set.
Proof. We proceed by way of contradiction. Let 𝑓 : R → R be a function with C (𝑓 ) = Q. By our
discussion above, C (𝑓 ) is the countable intersection of open sets (i.e. the 𝐴𝑛 ’s from (1.7)) which
contradicts Lemma 1.26. □
Definition 1.16. Let (𝑋, 𝑑) be a metric space. The metric 𝑑 is said to be proper if every closed
ball
𝐵𝑐 (𝑥, 𝜀) := {𝑦 ∈ 𝑋 : 𝑑 (𝑥, 𝑦) ≤ 𝜀}
is compact in 𝑋 .
Theorem 1.28. A proper metric space is both complete and locally compact Hausdorff.
Proof. Any metric space is Hausdorff so we need only verify that 𝑋 is locally compact to establish
that 𝑋 is LCH. Indeed, given any point 𝑥 ∈ 𝑋 , the open ball 𝐵(𝑥, 1) is by definition a neighbour-
hood of 𝑥. Furthermore, since 𝐵(𝑥, 1) ⊆ 𝐵𝑐 (𝑥, 1), we find that cl (𝐵(𝑥, 1)) ⊆ 𝐵𝑐 (𝑥, 1), where
𝐵𝑐 (𝑥, 1) is compact. Since closed subsets of compact sets remain compact, we have exhibited a
neighbourhood of 𝑥 with compact closure. Since 𝑥 was arbitrary, we infer that 𝑋 is LCH.
Next, we show that (𝑋, 𝑑) is complete. To this end, let (𝑥𝑛 ) be a Cauchy sequence in 𝑋 .
Since (𝑥𝑛 ) is bounded, for any 𝑎 ∈ 𝑋 , there exists 𝑀 > 0 such that 𝑑 (𝑥𝑛 , 𝑎) ≤ 𝑀 for all 𝑛 ∈ N.
That is, (𝑥𝑛 ) is contained within the closed ball 𝐵𝑐 (𝑎, 𝑀). But, since 𝑑 is a proper metric, this set
is (sequentially) compact. Hence, (𝑥𝑛 ) has a convergent subsequence (𝑥𝑛𝑘 ). By an assignment
result, (see your assignment 4), a Cauchy sequence with a convergent subsequence is necessarily
convergent. Thus, (𝑥𝑛 ) is seen to convergence and (𝑋, 𝑑) is complete. □
11 Recall 1
that, for each 𝜀 > 0, there exists 𝑛 ∈ N with the property that 𝑛 < 𝜀.

28
1.11 Lebesgue Measure Zero in One-Dimension
In relevance to your fourth assignment, we now briefly turn our focus towards the very tip of the
measure theory iceberg. That is, we introduce what is means for a subset of R to be “negligible”
in a volumetric sense. We would also like to point out that, although we will be discussing sets
of Lebesgue measure zero, we shall not be constructing the full Lebesgue measure or 𝜎-algebra.
Let now 𝐸 ⊆ R be an arbitrary set. Intuitively, we know how to measure the “length” of
an interval (𝑎, 𝑏). Indeed, we define its length (or measure) to be the value 𝑏 − 𝑎. Now, let us
imagine that we have some countable collection {𝐼𝑘 } of open intervals whose union covers 𝐸, i.e.
𝐸 ⊆ 𝑘 𝐼𝑘 . Then, although we haven’t yet given a definition for the measure or “length” of 𝐸, any
Ð
reasonable definition would have to satisfy the following:
∑︁
measure of 𝐸 ≤ (𝑏𝑘 − 𝑎𝑘 ), where 𝐼𝑘 = (𝑎𝑘 , 𝑏𝑘 ).
𝑘

That is, if 𝐸 ⊆ 𝑘 𝐼𝑘 , then the measure or length of 𝐸 should be no-greater than the combined
Ð
lengths of all the intervals covering it. With this, the following definition of having measure
zero is better motivated:

Definition 1.17 (Lebesgue Measure Zero). Let 𝐸 ⊆ R. We say that 𝐸 has Lebesgue measure 0
(or, alternatively, a null set) provided, for each 𝜀 > 0, there exists a countable collection of open
intervals {𝐼𝑘 } such that Ø ∑︁
𝐸⊆ 𝐼𝑘 and |𝐼𝑘 | ≤ 𝜀.
𝑘 𝑘

Here, |𝐼 | = 𝑏 − 𝑎 denotes the length of any open interval 𝐼 = (𝑎, 𝑏).

Loosely speaking, the set 𝐸 is said to have measure 0 if it can be covered by countably many
open intervals of arbitrarily small combined length.

Example 1.10. Every countable subset of R has Lebesgue measure 0.

Proof. Let 𝐸 = {𝑒𝑘 }𝑘 be a countable set and let 𝜀 > 0 be given. For each index 𝑘, put
 𝜀 𝜀 
𝐼𝑘 := 𝑒𝑘 − 𝑘+1 , 𝑒𝑘 + 𝑘+1
2 2
so that {𝐼𝑘 }𝑘 forms a countable collection open intervals whose union includes the set 𝐸, i.e.
𝑘 {𝑒𝑘 } = 𝐸. Furthermore,
Ð Ð
𝑘 𝐼𝑘 ⊇

∑︁ 1 ∞
∑︁ ∑︁ 𝜀 ∑︁ 1
|𝐼𝑘 | = = 𝜀 ≤ 𝜀 = 𝜀.
𝑘 𝑘
2𝑘
𝑘
2𝑘
𝑘=1
2𝑘

29
Remark 1.6 (Cantor Set has Measure Zero). Although it may be tempting at first, one cannot think
of measure zero sets as being countable. Indeed, countable sets are necessarily measure zero sets
but the converse need not hold. As a “simple” example, one can show that the Cantor set is a
measure zero set, despite being uncountable.

Proposition 1.29. The countable union of measure zero sets is again a measure zero set.

Proof. Let 𝐸 𝑗 𝑗 be a countable family of (Lebesgue) measure zero sets in R. That is, for each

n o
𝜀 > 0 and every 𝐸 𝑗 , there exists a countable family of open intervals 𝐼𝑘( 𝑗 ) such that
𝑘
Ø ∑︁ 𝜀
𝐸𝑗 ⊆ 𝐼𝑘( 𝑗 ) and 𝐼𝑘( 𝑗 ) ≤ . (1.9)
2𝑗
𝑘 𝑘

Next, we consider then countable

o family of open intervals formed by collecting all of the afore-
(𝑗)
mentioned intervals: 𝐼𝑘 . Clearly,
𝑗,𝑘
Ø Ø
𝐼𝑘( 𝑗 ) ⊇ 𝐸𝑗
𝑗,𝑘 𝑗

and, moreover,
∑︁ ∑︁ ∑︁ ∑︁  𝜀  ∑︁ 1
𝐼𝑘( 𝑗 ) = 𝐼𝑘( 𝑗 ) ≤ = 𝜀
𝑗 𝑗
2𝑗 𝑗
2𝑗
𝑗,𝑘 𝑘
∞
∑︁ 1
≤𝜀
𝑗=1
2𝑗
= 𝜀.

Since 𝜀 > 0 was taken to be arbitrary, this concludes the proof. □

30
2 Differentiability
As your third assignment involves some questions with derivatives, it is likely beneficial for us to
give a brief overview of the definition(s). In what follows, unless otherwise stated, we consider
functions defined on non-trivial intervals 𝐼 ⊂ R.

Definition 2.1 (Differentiability). Let 𝐼 be an interval and 𝑐 ∈ 𝐼 . We say that a function 𝑓 : 𝐼 → R

is differentiable at the point 𝑐 provided the the following limit exists and converges:

𝑓 (𝑥) − 𝑓 (𝑐)
lim .
𝑥→𝑐 𝑥 −𝑐
In this case, we define
𝑓 (𝑥) − 𝑓 (𝑐)
𝑓 ′ (𝑐) := lim
𝑥→𝑐 𝑥 −𝑐
and call it the derivative of 𝑓 at 𝑐. If 𝑓 is differentiable at all 𝑐 ∈ 𝐼 , we say that 𝑓 is differentiable
on 𝐼 .

Much like with our treatment of spaces, it is nice to understand how different definitions relate
to each other (e.g. normed spaces are metric spaces and metric spaces are topological spaces). In
this case, it is an elementary fact that all differentiable functions are continuous. However, there
are continuous functions that are nowhere differentiable (see the Weierstrass function).

Proposition 2.1. Let 𝐼 ⊆ R be an interval and 𝑐 ∈ 𝐼 . If 𝑓 : 𝐼 → R is differentiable at 𝑐, then it is

continuous at 𝑐.

Proof. For all 𝑥 ≠ 𝑐 we may write

𝑓 (𝑥) − 𝑓 (𝑐)
𝑓 (𝑥) − 𝑓 (𝑐) = · (𝑥 − 𝑐).
𝑥 −𝑐
Thus, by the limit laws,
 
𝑓 (𝑥) − 𝑓 (𝑐)
lim (𝑓 (𝑥) − 𝑓 (𝑐)) = lim · (𝑥 − 𝑐)
𝑥→𝑐 𝑥→𝑐 𝑥 −𝑐
 
𝑓 (𝑥) − 𝑓 (𝑐)
= lim · lim (𝑥 − 𝑐)
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐

= 𝑓 ′ (𝑐) · 0
= 0.

Put otherwise, we have shown that lim𝑥→𝑐 𝑓 (𝑥) = 𝑓 (𝑐) which establishes continuity at 𝑐. □

31
Remark 2.1. Let 𝑓 : 𝐼 → R be differentiable at a point 𝑐 ∈ 𝐼 . Then, by the sequential criterion, we
have
𝑓 (𝑐𝑛 ) − 𝑓 (𝑐)
𝑓 ′ (𝑐) = lim ,
𝑛→∞ 𝑐𝑛 − 𝑐
for any sequence (𝑐𝑛 ) in the domain 𝐼 with 𝑐𝑛 → 𝑐 as 𝑛 → ∞. Thus, to evaluate the value of
𝑓 ′ (𝑐), we need only evaluate the limit along a convenient chosen sequence of points. However,
this approach is only valid when 𝑓 is assumed to be differentiable a priori. Indeed, consider the
function
1 if 𝑥 ∈ 𝑛1 : 𝑛 ∈ N
( 
𝑓 : R → R, 𝑥 ↦→
0 otherwise.
Then, 𝑓 is fails the sequential criterion for continuity at 𝑥 = 0 because although 𝑛1 → 0 as 𝑛 → ∞,
we have by inspection that

1
 
lim 𝑓 = lim 1 = 1 ≠ 𝑓 (0) = 0.
𝑛→∞ 𝑛 𝑛→∞

Therefore, 𝑓 cannot be differentiable at 0. Despite this, however, we can still come up with a
sequence (𝑐𝑛 ) converging to 0 such that
𝑓 (𝑐𝑛 ) − 𝑓 (𝑐)
lim
𝑛→∞ 𝑐𝑛 − 𝑐
√
2
converges. Indeed, let us put 𝑐𝑛 := 𝑛 so that 𝑐𝑛 ∉ Q for all 𝑛 ∈ N. Clearly, 𝑐𝑛 → 0 as 𝑛 → ∞ and
√ 
2
𝑓 (𝑐𝑛 ) − 𝑓 (𝑐) 𝑓 𝑛 − 𝑓 (0)
lim = lim √ = lim 0 = 0.
𝑐𝑛 − 𝑐 2
𝑛→∞ 𝑛→∞
𝑛 − 0 𝑛→∞

Theorem 2.2. Let 𝐼 ⊆ R be an interval and 𝑓 , 𝑔 : 𝐼 → R be differentiable at a point 𝑐 ∈ 𝐼 . Then,

1. 𝑓 + 𝑔 is differentiable at 𝑐 and

(𝑓 + 𝑔) ′ (𝑐) = 𝑓 ′ (𝑐) + 𝑔′ (𝑐).

2. 𝑓 𝑔 is differentiable at 𝑐 and

(𝑓 𝑔) ′ (𝑐) = 𝑓 ′ (𝑐)𝑔(𝑐) + 𝑓 (𝑐)𝑔′ (𝑐).

3. 𝛼 𝑓 is differentiable at 𝑐, for every 𝛼 ∈ R and

(𝛼 𝑓 ) ′ (𝑐) = 𝛼 𝑓 ′ (𝑐).

Proof.

32
1. Using that, by assumption, both 𝑓 and 𝑔 are differentiable at 𝑐, we have both limits below
𝑓 (𝑥) − 𝑓 (𝑐)
𝑓 ′ (𝑐) = lim , (2.1)
𝑥→𝑐 𝑥 −𝑐
𝑔(𝑥) − 𝑓 (𝑐)
𝑔′ (𝑐) = lim (2.2)
𝑥→𝑐 𝑥 −𝑐
are well defined and exist. Then, we observe that for all 𝑥 ∈ 𝐼 with 𝑥 ≠ 𝑐 there holds
(𝑓 + 𝑔) (𝑥) − (𝑓 + 𝑔) (𝑐) 𝑓 (𝑥) + 𝑔(𝑥) − (𝑓 (𝑐) − 𝑔(𝑐))
=
𝑥 −𝑐 𝑥 −𝑐
𝑓 (𝑥) − 𝑓 (𝑐) 𝑔(𝑥) − 𝑔(𝑐)
= + .
𝑥 −𝑐 𝑥 −𝑐
By the limit laws from Math 242, and the fact that the limits in (2.1) exist, we infer that
(𝑓 + 𝑔) (𝑥) − (𝑓 + 𝑔) (𝑐)
lim
𝑥→𝑐 𝑥 −𝑐
exists and is equal to
 
(𝑓 + 𝑔) (𝑥) − (𝑓 + 𝑔) (𝑐) 𝑓 (𝑥) − 𝑓 (𝑐) 𝑔(𝑥) − 𝑔(𝑐)
lim = lim + .
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐 𝑥 −𝑐
𝑓 (𝑥) − 𝑓 (𝑐) 𝑔(𝑥) − 𝑔(𝑐)
= lim + lim
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐
= 𝑓 ′ (𝑐) + 𝑔′ (𝑐).
This proves (1).
2. To establish the product rule, we employ the same approach but using a clever trick. Notice
that, for all 𝑥 ≠ 𝑐 with 𝑥 ∈ 𝐼 there holds
(𝑓 𝑔) (𝑥) − (𝑓 𝑔) (𝑐) 𝑓 (𝑥)𝑔(𝑥) − 𝑓 (𝑐)𝑔(𝑐)
=
𝑥 −𝑐 𝑥 −𝑐
=0
z }| {
𝑓 (𝑥)𝑔(𝑥) −𝑓 (𝑥)𝑔(𝑐) + 𝑓 (𝑥)𝑔(𝑐) −𝑓 (𝑐)𝑔(𝑐)
=
𝑥 −𝑐
𝑓 (𝑥)𝑔(𝑥) − 𝑓 (𝑥)𝑔(𝑐) 𝑓 (𝑥)𝑔(𝑐) − 𝑓 (𝑐)𝑔(𝑐)
= +
𝑥 −𝑐 𝑥 −𝑐
𝑔(𝑥) − 𝑔(𝑐) 𝑓 (𝑥) − 𝑓 (𝑐)
= 𝑓 (𝑥) + 𝑔(𝑐) .
𝑥 −𝑐 𝑥 −𝑐
Now, because 𝑓 and 𝑔 are differentiable at the point 𝑐, we know that the limits below
𝑓 (𝑥) − 𝑓 (𝑐)
𝑓 ′ (𝑐) = lim ,
𝑥→𝑐 𝑥 −𝑐
𝑔(𝑥) − 𝑓 (𝑐)
𝑔′ (𝑐) = lim
𝑥→𝑐 𝑥 −𝑐

33
 exist. Furthermore, since 𝑓 is differentiable at 𝑐, it must be continuous at 𝑐. Namely, one
has that lim𝑥→𝑐 𝑓 (𝑥) = 𝑓 (𝑐). Therefore, by applying the limit laws to the expression above,
we see that the following limit exists and is equal to:
 
(𝑓 𝑔) (𝑥) − (𝑓 𝑔) (𝑐) 𝑔(𝑥) − 𝑔(𝑐) 𝑓 (𝑥) − 𝑓 (𝑐)
lim = lim 𝑓 (𝑥) + 𝑔(𝑐)
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐 𝑥 −𝑐
   
𝑔(𝑥) − 𝑔(𝑐) 𝑓 (𝑥) − 𝑓 (𝑐)
= lim 𝑓 (𝑥) + lim +𝑔(𝑐)
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐
   
h i 𝑔(𝑥) − 𝑔(𝑐) 𝑓 (𝑥) − 𝑓 (𝑐)
= lim 𝑓 (𝑥) lim + 𝑔(𝑐) lim
𝑥→𝑐 𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐
= 𝑓 (𝑐)𝑔′ (𝑐) + 𝑔(𝑐)𝑓 ′ (𝑐).

3. This follows from the product rule with 𝑔 = 𝛼.

2.1 Examples and Elementary Results

For the sake of clarity and completeness, we now illustrate some explicit examples of common
functions that are verifiably differentiable via the definition.
√
Example 2.1. Consider the function 𝑓 (𝑥) := 𝑥 defined on [0, ∞). We show that this function
is differentiable on (0, ∞) with derivative
1
𝑓 ′ (𝑥) = √ .
2 𝑥

Fixing 𝑐 ∈ (0, ∞) and letting 𝑥 ≠ 𝑐 be positive we calculate

√ √ √ √ √ √
𝑓 (𝑥) − 𝑓 (𝑐) 𝑥− 𝑐 𝑥− 𝑐 𝑥+ 𝑐
= = ·√ √
𝑥 −𝑐 𝑥 −𝑐 𝑥 −𝑐 𝑥+ 𝑐
𝑥 −𝑐
= √ √
(𝑥 − 𝑐) ( 𝑥 + 𝑐)
1
=√ √ .
𝑥+ 𝑐
√ √
So, since 𝑥 → 𝑐 as 𝑥 → 𝑐, the limit laws yield

𝑓 (𝑥) − 𝑓 (𝑐) 1
lim = √ , ∀𝑐 ∈ (0, ∞).
𝑥→𝑐 𝑥 −𝑐 2 𝑐
1
Therefore 𝑓 is differentiable on (0, ∞) with 𝑓 ′ (𝑥) = √
2 𝑥
.

34
Example 2.2. Consider the function 𝑓 (𝑥) := √1 on (0, ∞). Then, 𝑓 is differentiable on its entire
𝑥
domain with derivative given by
1
𝑓 ′ (𝑥) = − .
2𝑥 3/2
Indeed, let 𝑐 > 0 be fixed and let 𝑥 > 0 with 𝑥 ≠ 𝑐. A straightforward calculation yields:
1
√ − √ 1
𝑓 (𝑥) − 𝑓 (𝑐) 𝑥 𝑐
=
𝑥 −𝑐 𝑥√− 𝑐 √
𝑐− 𝑥
= √ √
(𝑥 − 𝑐) 𝑥 𝑐
√ √ √ √
𝑐− 𝑥 𝑥+ 𝑐
= √ √ ·√ √
(𝑥 − 𝑐) 𝑥 𝑐 𝑥+ 𝑐
𝑐 −𝑥
= √ √ √ √
(𝑥 − 𝑐) 𝑥 𝑐 ( 𝑥 + 𝑐)
1
= −√ √ √ √ .
𝑥 𝑐 ( 𝑥 + 𝑐)
√ √
Thus, applying the limit laws as above and using that 𝑥 → 𝑐 as 𝑥 → 𝑐, we deduce that
1
 
𝑓 (𝑥) − 𝑓 (𝑐)
lim = lim − √ √ √ √
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 𝑐 ( 𝑥 + 𝑐)
1
=− √ √ √ √
lim𝑥→𝑐 𝑥 𝑐 ( 𝑥 + 𝑐)
1
= −√ √ √

𝑐 𝑐 2 𝑐
1
= − 3/2 .
2𝑐
Example 2.3. Consider the function 𝑓 : R → R given by the following conditional rule:
(
𝑥 2 if 𝑥 ∈ Q,
𝑓 (𝑥) :=
0 if 𝑥 ∉ Q.

We claim that 𝑓 is differentiable at 𝑥 = 0 and not differentiable at all other points. To see that 𝑓 is
not differentiable away from 0, it suffices to show that 𝑓 is discontinuous at all 𝑐 ≠ 0. Indeed, given
𝑐 ≠ 0, we may select (by density) a sequence (𝑟𝑛 ) in Q such that 𝑟𝑛 → 𝑐 as 𝑛 → ∞. Analogously,
since R \ Q is dense in R, there also exists a sequence (𝜉𝑛 ) from R \ Q such that 𝑟𝑛 → 𝑐 as 𝑛 → ∞.
However,
lim 𝑓 (𝑟𝑛 ) = lim 𝑟𝑛2 = 𝑐 2, lim 𝑓 (𝜉𝑛 ) = lim 0 = 0.
𝑛→∞ 𝑛→∞ 𝑛→∞ 𝑛→∞

Since 𝑐 2 ≠ 0, it follows that 𝑓 fails the sequential criterion for continuity away from 0. As differ-
entiability always implies continuity, we infer that 𝑓 is not differentiable at 𝑐 ≠ 0.

35
 Next, we show that 𝑓 ′ (0) exists and equals 0. Certainly, for any 𝑥 ∈ R, we have that |𝑓 (𝑥)| ≤
𝑥 2 . In particular, for every 𝑥 ≠ 0, we have
𝑓 (𝑥) − 𝑓 (0) |𝑓 (𝑥)| 𝑥2
−0 = ≤ = |𝑥 | .
𝑥 −0 |𝑥 | |𝑥 |
Next, let 𝜀 > 0 be given and set 𝛿 := 𝜀. If 𝑥 ∈ R is such that |𝑥 | < 𝛿, then
𝑓 (𝑥) − 𝑓 (0)
− 0 ≤ |𝑥 | < 𝛿 = 𝜀.
𝑥 −0
It follows that
𝑓 (𝑥) − 𝑓 (0)
0 = lim
𝑥→𝑐 𝑥 −0
whence 𝑓 ′ (0) = 0.
Proposition 2.3. Let 𝑓 : R → R be differentiable at a point 𝑐 ∈ R with 𝑓 (𝑐) = 0. Then. |𝑓 | is
differentiable at 𝑐 if and only if 𝑓 ′ (𝑐) = 0.12
Proof. Put 𝑔(𝑥) := |𝑓 (𝑥)| and note that 𝑔(𝑐) = 0. First, suppose that 𝑓 ′ (𝑐) = 0. Then, 𝑔 is
differentiable at 𝑐 with 𝑔′ (𝑐) = 0. Indeed, notice that (by the reverse triangle inequality)
𝑔(𝑥) − 𝑔(𝑐) ||𝑓 (𝑥)| − |𝑓 (𝑐)|| |𝑓 (𝑥) − 𝑓 (𝑐)|
0≤ = ≤
𝑥 −𝑐 |𝑥 − 𝑐 | |𝑥 − 𝑐 |
𝑓 (𝑥) − 𝑓 (𝑐) 𝑥→𝑐
= − 0 −−−→ 0.
𝑥 −𝑐
Thus, 𝑔′ (𝑐) exists and equals 0 by definition. Conversely, let us assume that 𝑔′ (𝑐) exists; we seek
to show that 𝑓 ′ (𝑐) = 0. Indeed, it is obvious that 𝑔 ≥ 0 on R. Thus, since 𝑔(𝑐) = 0, we see that
𝑐 ∈ R is a local minimum for the function 𝑔 on R. Since 𝑔′ (𝑐) exists by assumption, we must have
𝑔′ (𝑐) = 0. This implies that
𝑔(𝑥)
lim = 0.
𝑥→𝑐 𝑥 − 𝑐
But then,
𝑓 (𝑥) − 𝑓 (𝑐) 𝑓 (𝑥) |𝑓 (𝑥)| 𝑔(𝑥) |𝑔(𝑥)| 𝑔(𝑥) 𝑥→𝑐
−0 = = = = = −−−→ 0.
𝑥 −𝑐 𝑥 −𝑐 |𝑥 − 𝑐 | |𝑥 − 𝑐 | |𝑥 − 𝑐 | 𝑥 −𝑐
Hence, 𝑓 is differentiable at 𝑥 = 𝑐 with 𝑓 ′ (0) = 0. □
Lemma 2.4 (Straddle Lemma). Let 𝑓 : 𝐼 → R be differentiable at 𝑐 ∈ 𝐼 . For each 𝜀 > 0, there exists
𝛿 > 0 such that
|𝑓 (𝑥) − 𝑓 (𝑦) − (𝑥 − 𝑦)𝑓 ′ (𝑐)| ≤ 𝜀 (𝑥 − 𝑦)
for all 𝑥, 𝑦 ∈ 𝐼 with 𝑐 − 𝛿 < 𝑦 ≤ 𝑐 ≤ 𝑥 < 𝑐 + 𝛿.
12 Our proof shows that the “ ⇐= ” implication holds without 𝑓 (𝑐) = 0.

36
Proof. Let 𝜀 > 0 be given. Since 𝑓 ′ (𝑐) exists (i.e. 𝑓 is differentiable at 𝑐), the 𝜀 − 𝛿-definition of the
limit ensures the existence of some 𝛿 > 0 having the property that

𝑓 (𝑥) − 𝑓 (𝑐)
− 𝑓 ′ (𝑐) < 𝜀
𝑥 −𝑐
whenever 𝑥 ∈ 𝐼 satisfies 0 < |𝑥 − 𝑐 | < 𝛿. In fact, multiplying through by |𝑥 − 𝑐 | > 0 shows that

|𝑓 (𝑥) − 𝑓 (𝑐) − 𝑓 ′ (𝑐) (𝑥 − 𝑐)| < 𝜀 |𝑥 − 𝑐 | .

Let now 𝑥, 𝑦 ∈ 𝐼 be as in the statement and observe that |𝑓 (𝑥) − 𝑓 (𝑦) − (𝑥 − 𝑦) 𝑓 ′ (𝑐)| equals

|𝑓 (𝑥) − 𝑓 (𝑦) + 𝑓 (𝑐) − 𝑐 𝑓 ′ (𝑐) − 𝑓 (𝑐) + 𝑐 𝑓 ′ (𝑐) − (𝑥 − 𝑦) 𝑓 ′ (𝑐)|

≤ |𝑓 (𝑥) − 𝑓 (𝑐) − (𝑥 − 𝑐) 𝑓 ′ (𝑐)| + |−𝑓 (𝑦) + 𝑓 (𝑐) + (𝑦 − 𝑐) 𝑓 ′ (𝑐)|
≤ |𝑓 (𝑥) − 𝑓 (𝑐) − (𝑥 − 𝑐) 𝑓 ′ (𝑐)| + |𝑓 (𝑦) − 𝑓 (𝑐) − (𝑦 − 𝑐) 𝑓 ′ (𝑐)|
≤ 𝜀 (𝑥 − 𝑐) + 𝜀 (𝑐 − 𝑦)
= 𝜀 (𝑥 − 𝑦).

2.2 The Mean Value Theorem

Lemma 2.5. Let 𝐼 ⊆ R be an interval and 𝑓 : 𝐼 → R be differentiable at an interior point 𝑐 ∈ 𝐼 . If
𝑐 is a relative extremum, then 𝑓 ′ (𝑐) = 0.

Proof. Replacing 𝑓 with −𝑓 , we may assume without loss of generality that 𝑐 is an interior relative
maximum. This means that 𝑐 ∈ Int(𝐼 ). Hence, there exists 𝜀 1 > 0 such that (𝑐 −𝜀 1, 𝑐 +𝜀 1 ) ⊆ Int(𝐼 ) ⊆
𝐼 . Furthermore, since 𝑐 is a relative maximum, there exists 𝜀 2 > 0 such that 𝑓 (𝑥) ≤ 𝑓 (𝑐) for all
𝑥 ∈ 𝐼 ∩ (𝑥 −𝜀 2, 𝑥 +𝜀 2 ). Define 𝛿 := min (𝜀 1, 𝜀 2 ) and observe that (𝑐 −𝜀, 𝑐 +𝜀) ⊆ 𝐼 and that 𝑓 (𝑥) ≤ 𝑓 (𝑐)
on all of (𝑐 − 𝛿, 𝑐 + 𝛿). Furthermore, since 𝑓 is differentiable at 𝑐,
𝑓 (𝑥) − 𝑓 (𝑐) 𝑓 (𝑥) − 𝑓 (𝑐) 𝑓 (𝑥) − 𝑓 (𝑐)
𝑓 ′ (𝑐) = lim = lim− = lim+
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐
where both one-sided limits must agree because 𝑓 ′ (𝑐) is well defined. Notice, however, that
𝑓 (𝑥) − 𝑓 (𝑐)
lim ≥0
𝑥→𝑐 − 𝑥 −𝑐
since 𝑓 (𝑥) ≤ 𝑓 (𝑐) for all 𝑥 ∈ (𝑐 − 𝜀, 𝑐 + 𝜀) and 𝑥 < 𝑐 for all 𝑥 ∈ (𝑐 − 𝜀, 𝑐). Similarly, one finds that
𝑓 (𝑥) − 𝑓 (𝑐)
lim− ≤0
𝑥→𝑐 𝑥 −𝑐
whence 0 ≤ 𝑓 ′ (𝑐) ≤ 0, thereby concluding our proof. □

37
Theorem 2.6 (Rolle’s Theorem). Let 𝑓 : [𝑎, 𝑏] → R be continuous and assume that 𝑓 is differen-
tiable on (𝑎, 𝑏). Suppose further that 𝑓 (𝑎) = 𝑓 (𝑏). Then, there exists a point 𝑐 ∈ (𝑎, 𝑏) such that
𝑓 ′ (𝑐) = 0.

Proof. By replacing 𝑓 with 𝑓 − 𝑓 (𝑎) if necessary, we may assume without loss of generality that
𝑓 (𝑎) = 𝑓 (𝑏) = 0. Since 𝑓 is continuous on the compact interval [𝑎, 𝑏], it achieves a global
minimum 𝑚 and a global maximum 𝑀 on [𝑎, 𝑏]. If 𝑀 = 𝑚 = 0, then 𝑓 ≡ 0 on [𝑎, 𝑏] and so
𝑓 ′ (𝑐) = 0 for all points 𝑐 ∈ (𝑎, 𝑏). Otherwise, one of 𝑀 or 𝑚 is non-trivial. In particular, 𝑓
achieves a non-zero global extremum in [𝑎, 𝑏]; let 𝑐 ∈ [𝑎, 𝑏] be a point where this extremum is
achieved. Since 𝑓 (𝑎) = 𝑓 (𝑏) = 0, we see that 𝑐 cannot be the endpoints of the interval [𝑎, 𝑏], i.e.
out extremum is achieved in the open interval (𝑎, 𝑏). Thus, 𝑓 has a local extremum at some point
𝑐 ∈ (𝑎, 𝑏) whence 𝑓 ′ (𝑐) = 0, as was asserted. □

Theorem 2.7 (Cauchy Mean Value Theorem). Let 𝑓 , 𝑔 : [𝑎, 𝑏] → R be continuous on [𝑎, 𝑏] and
differentiable on the open interval (𝑎, 𝑏). Assume in addition that 𝑔′ (𝑥) ≠ 0 for all 𝑥 ∈ (𝑎, 𝑏). Then,
there exists a point 𝑐 ∈ (𝑎, 𝑏) such that
𝑓 ′ (𝑐) 𝑓 (𝑏) − 𝑓 (𝑎)
=
𝑔′ (𝑐) 𝑔(𝑏) − 𝑔(𝑎)
Proof. By virtue of Rolle’s theorem, we must have 𝑔(𝑏) ≠ 𝑔(𝑎). Now, let us consider the function
𝑓 (𝑏) − 𝑓 (𝑎)
ℎ(𝑥) := (𝑔(𝑥) − 𝑔(𝑎)) − (𝑓 (𝑥) − 𝑓 (𝑎)) .
𝑔(𝑏) − 𝑔(𝑎)
Clearly, ℎ is continuous on [𝑎, 𝑏] and differentiable on (𝑎, 𝑏). Furthermore, by inspection, we have
that ℎ(𝑎) = 0 and ℎ(𝑏) = 0. Thus, there exists a point 𝑐 ∈ (𝑎, 𝑏) where
𝑓 (𝑏) − 𝑓 (𝑎) ′
0 = ℎ ′ (𝑐) = 𝑔 (𝑐) − 𝑓 ′ (𝑐)
𝑔(𝑏) − 𝑔(𝑎)
whence a rearrangement of terms yields the assertion. □

Corollary 2.8 (Mean Value Theorem). Let 𝑓 : [𝑎, 𝑏] → R be continuous on [𝑎, 𝑏] and differentiable
on the open interval (𝑎, 𝑏). Then, there exists a point 𝑐 ∈ (𝑎, 𝑏) such that
𝑓 (𝑏) − 𝑓 (𝑎)
𝑓 ′ (𝑐) = .
𝑏 −𝑎
Proof. This follows by taking 𝑔(𝑥) := 𝑥 in the previous statement. □

2.3 A Word on the Lipschitz Condition

Let us recall an important definition. Let 𝐴 ⊆ R be non-empty and let 𝑓 : 𝐴 → R be a func-
tion. We say that 𝑓 is Lipschitz continuous on 𝐴 if there exists some constant 𝐿 > 0 such that

38
|𝑓 (𝑥) − 𝑓 (𝑦)| ≤ 𝐿 |𝑥 − 𝑦| for all 𝑥, 𝑦 ∈ 𝐴. It is an easy exercise to check that every Lipschitz func-
tion is automatically (uniformly) continuous on its domain. The next theorem links the Lipschitz-
property to the boundedness of the derivative 𝑓 ′ , provided your given function is a priori assumed
to be differentiable.

Theorem 2.9. Let 𝐼 ⊆ R be an interval and let 𝑓 : 𝐼 → R be differentiable on 𝐼 . Then, 𝑓 is Lipschitz

continuous on 𝐼 if and only if 𝑓 ′ is bounded on 𝐼 .

Proof. Assume that 𝑓 is Lipschitz continuous and fix a point 𝑦 ∈ 𝐼 . Since 𝑓 ′ (𝑦) exists, we know
that the limit
𝑓 (𝑥) − 𝑓 (𝑦)
𝑓 ′ (𝑦) = lim
𝑥→𝑦 𝑥 −𝑦
is defined. Therefore,
𝑓 (𝑥) − 𝑓 (𝑦)
|𝑓 ′ (𝑦)| = lim ≤ lim 𝐿 = 𝐿.
𝑥→𝑦 𝑥 −𝑦 𝑥→𝑦

Since 𝑦 ∈ 𝐼 was arbitrary, the assertion follows.

We now establish the converse, i.e. we assume that 𝑓 ′ is bounded on 𝐼 and deduce that 𝑓 is
Lipschitz continuous there. Since 𝑓 ′ is bounded on 𝐼 , there exists 𝐿 > 0 such that |𝑓 ′ (𝑥)| ≤ 𝐿 for
all 𝑥 ∈ 𝐼 . Fix now two points 𝑥, 𝑦 ∈ 𝐼 ; we want to show that

|𝑓 (𝑥) − 𝑓 (𝑦)| ≤ 𝐿 |𝑥 − 𝑦| .

Note that this inequality is trivial when 𝑥 = 𝑦. Therefore, we may assume without loss of general-
ity that 𝑥 < 𝑦. Clearly, we then have [𝑥, 𝑦] ⊆ 𝐼 . Thus, 𝑓 is differentiable on (𝑥, 𝑦) and continuous
on [𝑥, 𝑦]. By the mean value theorem, there exists a point 𝑐 ∈ (𝑥, 𝑦) ⊆ 𝐼 such that
𝑓 (𝑦) − 𝑓 (𝑥)
= 𝑓 ′ (𝑐)
𝑦 −𝑥
whence
𝑓 (𝑦) − 𝑓 (𝑥)
= |𝑓 ′ (𝑐)| ≤ 𝐿.
𝑦 −𝑥
This gives |𝑓 (𝑥) − 𝑓 (𝑦)| ≤ 𝐿 |𝑥 − 𝑦| which completes the proof. □

Remark 2.2. The differentiability assumption imposed on the function 𝑓 can be marginally weak-
ened if we are careful in how we craft our argument. Indeed, the argument used in our proof
boils down to applying the mean value theorem on subintervals [𝑥, 𝑦] ⊆ 𝐼 , where one would only
require that 𝑓 ′ exist inside the open interval (𝑥, 𝑦). Therefore, the last result remains valid if we
assume merely that 𝑓 is continuous on 𝐼 and differentiable in the interior of 𝐼 .

Corollary 2.10. For all 𝑥, 𝑦 ∈ R one has

|sin 𝑥 − sin 𝑦| ≤ |𝑥 − 𝑦| and |cos 𝑥 − cos 𝑦| ≤ |𝑥 − 𝑦| .

39
Proof. Since sin 𝑥 and cos 𝑥 are differentiable on all of R with (sin 𝑥) ′ = cos 𝑥 and (cos 𝑥) ′ =
− sin(𝑥), where |sin 𝑥 | ≤ 1 and |cos 𝑥 | ≤ 1, the two inequalities above readily follow from the
previous theorem. □

Corollary 2.11. Let [𝑎, 𝑏] be an interval and 𝑓 : [𝑎, 𝑏] → [𝑎, 𝑏] be continuous on [𝑎, 𝑏] and
differentiable on (𝑎, 𝑏). Assume in addition that |𝑓 ′ (𝑥)| ≤ 𝑐 < 1 for all 𝑥 ∈ (𝑎, 𝑏). Then, 𝑓 possesses
a unique fixed point in [𝑎, 𝑏].

Proof. Citing Theorem 2.9 and Remark 2.2, it is automatic that |𝑓 (𝑥) − 𝑓 (𝑦)| ≤ 𝑐 |𝑥 − 𝑦| for all
𝑥, 𝑦 ∈ 𝐼 . Because 0 ≤ 𝑐 < 1, this implies that our mapping 𝑓 is a contraction of 𝐼 . Then, appealing
to the Banach Fixed Point Theorem (Theorem 1.19) yields the existence of a unique fixed point. □

Remark 2.3. This theorem offers a nice characterization of Lipschitz continuous functions when
the function in question is a priori assumed to be differentiable. Nonetheless, one should not
make the assumption that every Lipschitz function is differentiable. Indeed, by the inequality

||𝑥 | − |𝑦|| ≤ |𝑥 − 𝑦| , ∀𝑥, 𝑦 ∈ R,

the function 𝑓 (𝑥) = |𝑥 | is Lipschitz on R. However, 𝑓 is not differentiable at 0.

2.4 Applications of the Mean Value Theorem and Exercises

Example 2.4. Let 𝑓 : R → R be differentiable and assume that 𝑓 ′ (𝑥) → 𝑏, with 𝑏 ∈ R, as
𝑥 → ∞.13

(i) Show that, for each ℎ > 0, there holds

𝑓 (𝑥 + ℎ) − 𝑓 (𝑥)
lim = 𝑏.
𝑥→∞ ℎ

(ii) Assume now that lim𝑥→∞ 𝑓 (𝑥) = 𝑎 ∈ R. Then, prove that 𝑏 = 0.

Solution.

(i) Fix ℎ > 0 and let 𝜀 > 0. Since 𝑓 ′ (𝑥) → 𝑏 as 𝑥 → ∞, there exists some 𝑁 > 0 so large that

|𝑓 ′ (𝑥) − 𝑏 | < 𝜀, ∀𝑥 ≥ 𝑁 . (2.3)

Now, let 𝑥 ≥ 𝑁 be arbitrary and consider the subinterval [𝑥, 𝑥 + ℎ] ⊂ R. By assumption on

𝑓 , we have that 𝑓 is continuous on this compact interval and differentiable on the interior
13 For the sake of clarity, we write 𝑥 → ∞ to mean 𝑥 → +∞. When “approaching” ±∞, we write 𝑥 → ±∞
or |𝑥 | → ∞.

40
 of the interval. Consequently, 𝑓 satisfies the requirements of the Mean Valu Theorem on
[𝑥, 𝑥 + ℎ], Thus, there exists a point14 𝑐 𝑥 ∈ (𝑥, 𝑥 + ℎ) such that

𝑓 (𝑥 + ℎ) − 𝑓 (𝑥) 𝑓 (𝑥 + ℎ) − 𝑓 (𝑥)
𝑓 ′ (𝑐 𝑥 ) = = . (2.4)
(𝑥 + ℎ) − 𝑥 ℎ
Furthermore, since 𝑐 𝑥 ∈ (𝑥, 𝑥 + ℎ), we have 𝑐 𝑥 > 𝑥 ≥ 𝑁 . But then, by combining (2.3)-(2.4)
we infer that
𝑓 (𝑥 + ℎ) − 𝑓 (𝑥)
− 𝑏 = |𝑓 ′ (𝑐 𝑥 ) − 𝑏 | < 𝜀.
ℎ

Since 𝑥 ≥ 𝑁 was arbitrary, it follows by definition that

𝑓 (𝑥 + ℎ) − 𝑓 (𝑥)
lim = 𝑏.
𝑥→∞ ℎ

(ii) With the additional assumption that 𝑓 (𝑥) → 𝑎 as 𝑥 → ∞, we shall show that 𝑏 is neces-
sarily 0. For each 𝑛 ∈ N, let us consider the subinterval (𝑛, 2𝑛) ⊂ R. Once again, the given
function 𝑓 satisfies the requirements of the Mean Value Theorem on [𝑛, 2𝑛]. Consequently,
for each 𝑛 ∈ N, we may select a point 𝑐𝑛 ∈ (𝑛, 2𝑛) such that

𝑓 (2𝑛) − 𝑓 (𝑛) 𝑓 (2𝑛) − 𝑓 (𝑛)

𝑓 ′ (𝑐𝑛 ) = = .
2𝑛 − 𝑛 𝑛
Now, by the sequential criterion, we have

lim 𝑓 ′ (𝑥𝑛 ) = 𝑏 and lim 𝑓 (𝑥𝑛 ) = 𝑎

𝑛→∞ 𝑛→∞

for all sequences (𝑥𝑛 ) in R with 𝑥𝑛 → ∞ as 𝑛 → ∞. Since 𝑐𝑛 ∈ (𝑛, 2𝑛) for all 𝑛 ∈ N, it is
obvious that 𝑐𝑛 → ∞ (squeeze theorem!). In particular,

lim 𝑓 ′ (𝑐𝑛 ) = 𝑏, lim 𝑓 (𝑛) = 𝑎, and lim 𝑓 (2𝑛) = 𝑎.

𝑛→∞ 𝑛→∞ 𝑛→∞

But then, the limit laws tell us that

𝑓 (2𝑛) − 𝑓 (𝑛)
𝑏 = lim 𝑓 ′ (𝑐𝑛 ) = lim
𝑛→∞ 𝑛→∞ 𝑛
1
 
= lim (𝑓 (2𝑛) − 𝑓 (𝑛)) ·
𝑛→∞ 𝑛
1
= lim (𝑓 (2𝑛) − 𝑓 (𝑛)) · lim
𝑛→∞ 𝑛→∞ 𝑛
= (𝑎 − 𝑎) · 0 = 0.
14 We use the notation 𝑐 𝑥 to emphasize how 𝑐 𝑥 is selected in a way that depends on 𝑥.

41
 □
Proposition 2.12. Let 𝑎 < 𝑏 be real numbers and assume that 𝑓 : [𝑎, 𝑏] → R is differentiable on
(𝑎, 𝑏) and continuous on [𝑎, 𝑏]. If
lim+ 𝑓 ′ (𝑥) = 𝐿1
𝑥→𝑎
exists, then 𝑓 ′ (𝑎) = 𝐿1 . Similarly, if
lim 𝑓 ′ (𝑥) = 𝐿2
𝑥→𝑏 −
then 𝑓 ′ (𝑏) = 𝐿2 .
Proof. Assume that 𝑓 ′ (𝑥) → 𝐿1 as 𝑥 → 𝑎 + . We want to show that
𝑓 (𝑥) − 𝑓 (𝑎)
lim+ = 𝐿1 .
𝑥→𝑎 𝑥 −𝑎
Let 𝜀 > 0 be given; using that lim𝑥→𝑎+ 𝑓 ′ (𝑥) = 𝐿1 , we can find 𝛿 > 0 such that
|𝑓 ′ (𝑥) − 𝐿1 | < 𝜀
whenever 0 < 𝑥 − 𝑎 < 𝛿.15 For any such 𝑥, since 𝑓 is continuous on [𝑎, 𝑥] ⊆ [𝑎, 𝑏] and differ-
entiable on (𝑎, 𝑥), the mean value theorem guarantees the existence of a point 𝑐 𝑥 ∈ (𝑎, 𝑥) such
that
𝑓 (𝑥) − 𝑓 (𝑎)
= 𝑓 ′ (𝑐 𝑥 ).
𝑥 −𝑎
Since 0 < 𝑐 𝑥 − 𝑎 < 𝑥 − 𝑎 < 𝛿, we obtain
𝑓 (𝑥) − 𝑓 (𝑎)
− 𝐿1 = |𝑓 ′ (𝑐 𝑥 ) − 𝐿1 | < 𝜀.
𝑥 −𝑎
We have therefore shown that
𝑓 (𝑥) − 𝑓 (𝑎)
− 𝐿1 < 𝜀
𝑥 −𝑎
whenever 0 < 𝑥 − 𝑎 < 𝛿. This proves that 𝑓 ′ (𝑎) = 𝐿1 . The second part can be verified by a
symmetric argument. □

2.5 Uniform Differentiability

Definition 2.2. Let 𝐼 ⊆ R be an interval. A differentiable function 𝑓 : 𝐼 → R is said to be
uniformly differentiable on 𝐼 if, for each 𝜀 > 0, there exists 𝛿 > 0 such that
𝑓 (𝑥) − 𝑓 (𝑦)
− 𝑓 ′ (𝑦) < 𝜀
𝑥 −𝑦
for all 𝑥, 𝑦 ∈ 𝐼 with 0 < |𝑥 − 𝑦| < 𝛿.16
15 As we are taking the limit as 𝑥 → 𝑎 from “above”, we are only considering points 𝑥 > 𝑎. Therefore,
no absolute values are needed here!
16 Uniform differentiability has some nice applications in numerical analysis.

42
 The definition of uniform differentiability can be compared to that of uniform continuity.
Indeed, in the above, both 𝑥 and 𝑦 are allowed to “vary” when taking the limit. Namely, we are
asking that the 𝛿 > 0 obtained above be independent of the point 𝑦 at which we are taking the
derivative.
Proposition 2.13. Let 𝑓 : 𝐼 → R be uniformly differentiable on 𝐼 . Then, 𝑓 ′ is uniformly continuous
on 𝐼 . Especially, 𝑓 is continuously differentiable on 𝐼 .17
Proof. We must show that 𝑓 ′ is uniformly continuous on 𝐼 . Let 𝜀 > 0 be given. By assumption,
there exists 𝛿 > 0 such that
𝑓 (𝑥) − 𝑓 (𝑦) 𝜀
− 𝑓 ′ (𝑦) < . (2.5)
𝑥 −𝑦 2
for all 𝑥, 𝑦 ∈ 𝐼 with 0 < |𝑥 − 𝑦| < 𝛿. For all such 𝑥, 𝑦 we have, by the triangle inequality,
𝑓 (𝑥) − 𝑓 (𝑦) 𝑓 (𝑥) − 𝑓 (𝑦)
|𝑓 ′ (𝑥) − 𝑓 ′ (𝑦)| = 𝑓 ′ (𝑥) − + − 𝑓 ′ (𝑦)
𝑥 −𝑦 𝑥 −𝑦
𝑓 (𝑥) − 𝑓 (𝑦) 𝑓 (𝑥) − 𝑓 (𝑦)
≤ − 𝑓 ′ (𝑥) + − 𝑓 ′ (𝑦)
𝑥 −𝑦 𝑥 −𝑦
𝑓 (𝑦) − 𝑓 (𝑥) 𝑓 (𝑥) − 𝑓 (𝑦)
= − 𝑓 ′ (𝑥) + − 𝑓 ′ (𝑦)
𝑦 −𝑥 𝑥 −𝑦
𝜀 𝜀
< + = 𝜀;
2 2
where we have used (2.5) in this last step. We note that this also trivially holds when 𝑥 = 𝑦.
Summarizing, we have shown that |𝑓 ′ (𝑥) − 𝑓 ′ (𝑦)| < 𝜀 whenever 𝑥, 𝑦 ∈ 𝐼 are such that |𝑥 − 𝑦| < 𝛿.
That is, we have that 𝑓 ′ is uniformly continuous on the interval 𝐼 . □
Remark 2.4. There exist uniformly continuous functions, differentiable at all points in the interior,
√
whose derivative is unbounded. Indeed, let us consider the function 𝑓 (𝑥) := 𝑥 on [0, 1]. Clearly,
𝑓 is continuous on the compact set [0, 1] and thus uniformly continuous there as well. However,
by a previous example, we know that 𝑓 is differentiable at all 𝑥 > 0 with
1
𝑓 ′ (𝑥) = √ , ∀𝑥 > 0.
2 𝑥
Hence, 𝑓 ′ is unbounded on the interval (0, 1). Since uniformly continuous functions on bounded
intervals are bounded 18 , we see that 𝑓 ′ cannot be uniformly continuous on (0, 1).
17 A partial converse to this result is proven in your third assignment. Indeed, you are asked to prove
that a continuously differentiable function 𝑓 : [𝑎, 𝑏] → R is uniformly differentiable on [𝑎, 𝑏]. The key
difference is that we are assuming the interval in question is compact. If we remove the compactness
assumption on [𝑎, 𝑏] (i.e. replace [𝑎, 𝑏] with half-open, open, or unbounded intervals) then we must also
ask that 𝑓 ′ be uniformly continuous on the aforementioned interval.
18 This is because uniformly continuous functions map Cauchy sequences to Cauchy sequences! Use this

in a proof by contradiction (with the Bolzano–Weierstrass theorem) to show that a uniformly continuous
function on a bounded set is bounded. Alternatively, consult your Analysis 1 notes!

43
2.6 A Basic Form of l’Hôpital’s Rule
The intuition behind l’Hôpital’s Rule is straightforward and comes directly from the definition of
the derivative. As a result, we begin with the following basic result.

Proposition 2.14 (Preliminary Form). Let 𝐼 ⊆ R be an interval and 𝑓 , 𝑔 : 𝐼 → R be differentiable

at 𝑐 ∈ 𝐼 with 𝑓 (𝑐) = 𝑔(𝑐) = 0. Assume also that 𝑔(𝑥) ≠ 0 for all 𝑥 ∈ 𝐼 \ {𝑐} and that 𝑔′ (𝑐) ≠ 0. Then,
𝑓 (𝑥) 𝑓 ′ (𝑐)
lim = .
𝑥→𝑐 𝑔(𝑥) 𝑔′ (𝑐)
Proof. The proof is elementary: using that 𝑓 (𝑐) = 𝑔(𝑐) = 0 we may write, for all 𝑥 ≠ 𝑐,
𝑓 (𝑥 ) −𝑓 (𝑐 )
𝑓 (𝑥) 𝑓 (𝑥) − 𝑓 (𝑐) 𝑥 −𝑐
= = 𝑔 (𝑥 ) −𝑔 (𝑐 )
. (2.6)
𝑔(𝑥) 𝑔(𝑥) − 𝑔(𝑐)
𝑥 −𝑐

Now, because
𝑓 (𝑥) − 𝑓 (𝑐) 𝑔(𝑥) − 𝑔(𝑐)
lim = 𝑓 ′ (𝑐) and lim = 𝑔′ (𝑐) ≠ 0,
𝑥→𝑐 𝑥 −𝑐 𝑥→𝑐 𝑥 −𝑐
the limit-laws applied to the expression in (2.6) yield
𝑓 (𝑥 ) −𝑓 (𝑐 ) 𝑓 (𝑥 ) −𝑓 (𝑐 )
𝑓 (𝑥) 𝑥 −𝑐 lim𝑥→𝑐 𝑥 −𝑐 𝑓 ′ (𝑐)
lim = lim = = .
𝑥→𝑐 𝑔(𝑥) 𝑥→𝑐 𝑔 (𝑥 ) −𝑔 (𝑐 ) 𝑔 (𝑥 ) −𝑔 (𝑐 )
lim𝑥→𝑐 𝑥 −𝑐 𝑔′ (𝑐)
𝑥 −𝑐

This concludes the proof. □

However, this elementary form is often lackluster in terms of practical applications. For in-
stance, it is hard to “successively” apply the result above to multiple derivatives. To counteract
this, we refine the statement (and proof!) to obtain a more general result which uses the limit in-
stead of requiring that 𝑓 , 𝑔 be differentiable at the point 𝑐 (with 𝑔′ (𝑥) ≠ 0). In fact, we developed
the Cauchy Mean Value Theorem for this proof.

Theorem 2.15 (One-Sided l’Hôpital’s Rule: Case 0/0 at a point). Let 𝑎 < 𝑏 be real numbers and
𝑓 , 𝑔 be differentiable functions on (𝑎, 𝑏) and assume that 𝑔(𝑥), 𝑔′ (𝑥) ≠ 0 for all 𝑥 ∈ (𝑎, 𝑏). Suppose
in addition that
lim+ 𝑓 (𝑥) = lim+ 𝑔(𝑥) = 0. (2.7)
𝑥→𝑎 𝑥→𝑎
If
𝑓 ′ (𝑥)
lim+ = 𝐿,
𝑥→𝑎 𝑔′ (𝑥)
then
𝑓 (𝑥)
lim+ = 𝐿.
𝑥→𝑎 𝑔(𝑥)

44
Proof. Let 𝜀 > 0 be given. By assumption, we can find 𝛿 > 0 such that 𝑥 ∈ (𝑎, 𝑏) with 0 < 𝑥 −𝑎 < 𝛿
implies
𝑓 ′ (𝑥)
− 𝐿 < 𝜀.
𝑔′ (𝑥)
Or, equivalently,
𝑓 ′ (𝑥)
𝐿 −𝜀 < < 𝐿 + 𝜀,
𝑔′ (𝑥)
for all 𝑥 ∈ (𝑎, 𝑏) with 0 < 𝑥 − 𝑎 < 𝛿. We may choose 𝛿 > 0 so small that 𝑎 + 𝛿 < 𝑏. Now,
let 𝑎 < 𝛼 < 𝛽 < 𝑎 + 𝛿 and notice that the assumptions of the Cauchy Mean Value Theorem are
satisfied on [𝛼, 𝛽]. Thus, there exists a point 𝑐 ∈ (𝛼, 𝛽) such that
𝑓 ′ (𝑐) 𝑓 (𝛽) − 𝑓 (𝛼)
= .
𝑔′ (𝑐) 𝑔(𝛽) − 𝑔(𝛼)
Since 𝑐 ∈ (𝛼, 𝛽) ⊆ (𝑎, 𝑎 + 𝛿), we have that 0 < 𝑐 − 𝑎 < 𝛿. Consequently,
𝑓 ′ (𝑐)
𝐿 −𝜀 < < 𝐿 +𝜀
𝑔′ (𝑐)
whence
𝑓 (𝛽) − 𝑓 (𝛼)
𝐿 −𝜀 < < 𝐿 +𝜀
𝑔(𝛽) − 𝑔(𝛼)
for all 𝑎 < 𝛼 < 𝛽 < 𝑎 + 𝛿. Passing to the limit in the above as 𝛼 → 𝑎 + and using (2.8), it follows
that
𝑓 (𝛽) − 𝑓 (𝛼) 𝑓 (𝛽)
𝐿 − 𝜀 ≤ lim+ = ≤ 𝐿 + 𝜀.
𝛼→𝑎 𝑔(𝛽) − 𝑔(𝛼) 𝑔(𝛽)
That is, we have
𝑓 (𝛽)
−𝐿 ≤ 𝜀
𝑔(𝛽)
for all 𝑎 < 𝛽 < 𝑎 + 𝛿. This proves, by definition, that
𝑓 (𝑥)
lim = 𝐿.
𝑥→𝑎 + 𝑔(𝑥)
□

Remark 2.5. Using a symmetric argument, an analogous result holds for the “right endpoint”.
Indeed, by following the same proof-strategy as above, one can show that if

lim 𝑓 (𝑥) = lim− 𝑔(𝑥) = 0

𝑥→𝑏 − 𝑥→𝑏

and
𝑓 ′ (𝑥)
lim− = 𝐿,
𝑥→𝑏 𝑔′ (𝑥)

45
then
𝑓 (𝑥)
lim− = 𝐿.
𝑥→𝑏 𝑔(𝑥)
In summary, l’Hôpital’s rule also holds at the right end point.
Finally, by combining both the left and right hand versions of this rule, we obtain the following
two-sided result:
Corollary 2.16 (l’Hôpital Rule: Case 0/0 at a Point). Let 𝑎 < 𝑐 < 𝑏 be real numbers and 𝑓 , 𝑔 be
differentiable functions on (𝑎, 𝑏) \ {𝑐}. Assume that 𝑔(𝑥), 𝑔′ (𝑥) ≠ 0 for all 𝑥 ∈ (𝑎, 𝑏) \ {𝑐}. Suppose
in addition that
lim 𝑓 (𝑥) = lim 𝑔(𝑥) = 0. (2.8)
𝑥→𝑐 𝑥→𝑐
If
𝑓 ′ (𝑥)
lim = 𝐿,
𝑥→𝑐 𝑔′ (𝑥)
then
𝑓 (𝑥)
lim = 𝐿.
𝑥→𝑐 𝑔(𝑥)
We remark that there are many forms of l’Hôpital’s rule, as well as several indeterminate
forms that one can reduce to cases the l’Hôpital rules can handle. The multiple forms of these
rules along with their detailed proofs can be found in §6.3 of Bartle-Sherbert’s Introducion to Real
Analysis. Rather than state and prove these, we supply the reader with an interesting example.
Example 2.5. Let 𝑓 : (0, ∞) → R be differentiable and assume that

lim (𝑓 (𝑥) + 𝑓 ′ (𝑥)) = 𝐿.

𝑥→∞

Prove that
lim 𝑓 (𝑥) = 𝐿 and lim 𝑓 ′ (𝑥) = 0.
𝑥→∞ 𝑥→∞

Proof. Notice that we may write

𝑒 𝑥 𝑓 (𝑥) 𝑔(𝑥)
𝑓 (𝑥) = = , ∀𝑥 > 0,
𝑒𝑥 ℎ(𝑥)
where we obviously set 𝑔(𝑥) := 𝑒 𝑥 𝑓 (𝑥) and ℎ(𝑥) = 𝑒 𝑥 . Then, we find ourselves in the case of
l’Hôpital’s Rule II from Bartle-Sherbert (one can verify by inspection that the requirements of
l’Hôpital’s are satisfied). Consequently, we may evaluate
𝑔(𝑥) l’Hôpital 𝑔′ (𝑥) 𝑒 𝑥 𝑓 (𝑥) + 𝑒 𝑥 𝑓 ′ (𝑥)
lim 𝑓 (𝑥) = lim = lim = lim
𝑥→∞ 𝑥→∞ ℎ(𝑥) 𝑥→∞ ℎ ′ (𝑥) 𝑥→∞ 𝑒𝑥
= lim (𝑓 (𝑥) + 𝑓 ′ (𝑥))
𝑥→∞
= 𝐿.

46
Finally, observe that for each 𝑥 > 0 one has

𝑓 ′ (𝑥) = 𝑓 ′ (𝑥) − 𝑓 (𝑥) + 𝑓 (𝑥)

whence

lim 𝑓 ′ (𝑥) = lim (𝑓 ′ (𝑥) − 𝑓 (𝑥) + 𝑓 (𝑥))

𝑥→∞ 𝑥→∞
= lim (𝑓 ′ (𝑥) − 𝑓 (𝑥)) + lim 𝑓 (𝑥)
𝑥→∞ 𝑥→∞
= lim 𝑓 (𝑥) − lim (𝑓 (𝑥) − 𝑓 ′ (𝑥))
𝑥→∞ 𝑥→∞
= 𝐿 − 𝐿 = 0.

2.7 The Intermediate Properties of Derivatives: Darboux’s Theorem

Let 𝐼 ⊆ R be an interval and 𝑓 : 𝐼 → R be a function. We say that 𝑓 has the intermediate value
property if, for any 𝑥, 𝑦 ∈ 𝐼 with 𝑥 ≠ 𝑦, any value between 𝑓 (𝑥) and 𝑓 (𝑦) is achieved at least once
by 𝑓 at some point between 𝑥 and 𝑦.
Darboux’s theorem asserts that derivatives satisfy the intermediate value property. More
precisely, let us recall the following class result:

Theorem 2.17 (Darboux). Let 𝐼 = [𝑎, 𝑏] and 𝑓 : 𝐼 → R be differentiable. If 𝑦 is any point between
𝑓 ′ (𝑎) and 𝑓 ′ (𝑏), there exists a point 𝑥 between 𝑎 and 𝑏 such that 𝑓 ′ (𝑥) = 𝑦.

Darboux’s theorem can sometimes make it easy to show that certain functions cannot be
derivatives of any differentiable function. Put otherwise, Darboux’s theorem sometimes implies
that a given a function has no antiderivative. We provide such an example below.

Example 2.6. Consider the function

(
0 if 𝑥 < 0,
ℎ(𝑥) :=
1 if 𝑥 ≥ 0.

We claim that there does not exist a differentiable function 𝑓 : R → R such that 𝑓 ′ = ℎ on R.
Arguing by contradiction, suppose such a function 𝑓 does indeed exist. Take 𝑎 = −1 and 𝑏 = 1.
Then, 𝑓 ′ (𝑎) = ℎ(𝑎) = 0 and 𝑓 ′ (𝑏) = ℎ(𝑏) = 1. By Darboux’s theorem, there must exist a point
𝑥 ∈ (−1, 1) such that
1
𝑓 ′ (𝑥) = ℎ(𝑥) = .
2
Clearly, no such point exists and we therefore have a contradiction. However, although ℎ is not
the derivative of a function 𝑓 : R → R that is everywhere differentiable, there exist uncountably

47
many functions 𝑓 : R → R, differentiable away from 0, such that 𝑓 ′ (𝑥) = ℎ(𝑥) for all 𝑥 ≠ 0.
Indeed, for each 𝑥 0 ∈ R, (
𝑥 0 if 𝑥 < 0,
𝑓 (𝑥) :=
𝑥 if 𝑥 ≥ 0
is differentiable at all points 𝑐 ≠ 0 and 𝑓 ′ (𝑐) = ℎ(𝑐) at all such 𝑐.

Let us also provide a less trivial example utilizing the result of Darboux:

Example 2.7. Let 𝑓 : [0, 2] → R be continuous on [0, 2] and differentiable on (0, 2). Suppose
that the given function 𝑓 satisfies the following:

• 𝑓 (0) = 0;

• 𝑓 (1) = 1;

• 𝑓 (2) = 1.

Prove that there exists a point 𝑐 ∈ (0, 2) such that 𝑓 ′ (𝑐) = 1/3.

Proof. First, consider the subinterval [0, 1] of [0, 2]. Clearly, 𝑓 is continuous on this compact
interval and differentiable in the interior, i.e. on (0, 1) ⊂ (0, 2). Thus, the requirements of the
Mean Value Theorem are satisfied for 𝑓 on [0, 1] whence there exists a point 𝑐 1 ∈ (0, 1) such that

𝑓 (1) − 𝑓 (0)
𝑓 ′ (𝑐 1 ) = = 𝑓 (1) − 𝑓 (0) = 1 − 0 = 1. (2.9)
1−0
In a similar vein, one can see by inspection that the conditions of the mean value theorem are
also satisfied for 𝑓 on [1, 2] ⊂ [0, 2]; therefore, there exists a point 𝑐 2 ∈ (1, 2) with

𝑓 (2) − 𝑓 (1)
𝑓 ′ (𝑐 2 ) = = 𝑓 (2) − 𝑓 (1) = 1 − 1 = 0. (2.10)
2−1
To summarize, we have found points 𝑐 1, 𝑐 2 ∈ (0, 2) such that 𝑓 ′ (𝑐 1 ) = 1 and 𝑓 ′ (𝑐 2 ) = 0. Appealing
to Darboux’s theorem, we see that 𝑓 ′ has the intermediate value property. Furthermore, (2.9)-
(2.10) yield
1
𝑓 ′ (𝑐 2 ) < < 𝑓 ′ (𝑐 1 ).
3
It then follows that there exists a point 𝑐 ∈ (𝑐 1, 𝑐 2 ) such that 𝑓 ′ (𝑐) = 13 . □

48
3 The Riemann Integral
Let us now address the topic of integration, which begins with the following question:

For which functions can one systematically assign a notion of “area” to the region
bounded by its graph?

Although exhaustive methods have a rich history (dating back to the work of Archimedes), it
wasn’t until the times of Newton and Leibniz that integration was formally introduced. Further-
more, it wouldn’t be until to the contributions of Riemann and Darboux for there to be a rigor-
ous mathematical treatment of the theory.19 Despite all this effort, the Riemann integral would
nonetheless turn out to be “lacking” in certain areas. Namely, it only allows one to integrate over
intervals (and by extension cubes in higher dimensions) and even so, places significant restriction
on the function. To help rectify this, Lebesgue would later developed the Lebesgue integral.
Consider a closed and bounded interval [𝑎, 𝑏] ⊆ R and let 𝑓 : [𝑎, 𝑏] → R be an arbitrary
function. A partition of the interval [𝑎, 𝑏] is a finite ordered set of points

P = {𝑥 0, . . . , 𝑥𝑛 }

in [𝑎, 𝑏] such that 𝑎 = 𝑥 0 < 𝑥 1 < · · · < 𝑥𝑛 = 𝑏. Especially, [𝑎, 𝑏] = 𝑛𝑗=1 [𝑥 𝑗 −1, 𝑥 𝑗 ]. Informally, P
Ð
describes a unique way of breaking the interval [𝑎, 𝑏] into non-overlapping (except at the end-
points) compact intervals [𝑥 𝑗 −1, 𝑥 𝑗 ] ⊆ 𝐼 . A tagged partition is a partition P together with a set of
points, called tags,  ∗
𝑥 1 , . . . , 𝑥𝑛∗
such that 𝑥 ∗𝑗 ∈ [𝑥 𝑗 −1, 𝑥 𝑗 ] for each index 𝑗 = 1, . . . , 𝑛. To emphasize the fact that a partition P is
equipped with a set of tags, we will write P. ¤
Given a (possibly un-tagged) partition P of an interval [𝑎, 𝑏], the mesh of P is defined to be
the length of the largest sub-interval defined by P. More precisely, we define

∥P ∥ := max 𝑥 𝑗 − 𝑥 𝑗 −1 > 0.


1≤ 𝑗 ≤𝑛

Definition 3.1. Let [𝑎, 𝑏] ⊂ R be a compact interval and 𝑓 : [𝑎, 𝑏] → R a function. Given a
tagged partition P¤ of [𝑎, 𝑏] as above, we define the Riemann sum of 𝑓 over P¤ to be the sum
𝑛
∑︁
𝑆 (𝑓 ; P)
¤ := 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ). (3.1)
𝑗=0

19 It
turns out that both treatments of the integral are equivalent, i.e. a bounded function is Riemann
integrable if and only if it is integrable according to Darboux’s theory. In our notes, we will mainly use
Riemann’s integral, whilst offering in the final subsection a brief overview of how the Darboux integral
may be constructed.

49
Often times, we will write Δ𝑥 𝑗 = (𝑥 𝑗 − 𝑥 𝑗 −1 ) so that 𝑆 (𝑓 ; P)
¤ := Í𝑛 𝑓 (𝑥 ∗ )Δ𝑥 𝑗 . The function 𝑓
𝑗=0 𝑗
is said to be Riemann integrable on [𝑎, 𝑏] if there exists Λ ∈ R such that, for each 𝜀 > 0, there is
𝛿 > 0 with the property that
𝑆 (𝑓 ; P)
¤ −Λ <𝜀

for all tagged partitions P¤ of [𝑎, 𝑏] with ∥ P¤ ∥ < 𝛿. In this case, we call Λ the Riemann integral of
𝑓 on [𝑎, 𝑏] and denote this quantity by
ˆ 𝑏 ˆ ˆ 𝑏
𝑓, 𝑓, or 𝑓 (𝑥) d𝑥 .
𝑎 [𝑎,𝑏 ] 𝑎

Let us take a moment to unpack the definition of Riemann integrability. The Riemann sum
¤ should be thought of as a rectangular approximation of the area under the “curve” of the
𝑆 (𝑓 , P)
function 𝑓 . Indeed, the step function
𝑛
∑︁
𝜑 (𝑥) := 𝑓 (𝑥 ∗𝑗 )1 [𝑥 𝑗 −1,𝑥 𝑗 ) (𝑥)
𝑗=1

is precisely an approximation of 𝑓 by a function that takes the constant values 𝑓 (𝑥 ∗𝑗 ) on each

subinterval [𝑥 𝑗 −1, 𝑥 𝑗 ) of [𝑎, 𝑏]. Here, 1𝐴 (𝑥) denotes the indicator function (or characteristic func-
tion) of the set 𝐴.20 Clearly, the classical area under the graph of 𝜑 is equal to the Riemann sum
𝑛
∑︁
¤ =
𝑆 (𝑓 , P) 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ).
𝑗=0

Note that our choice of tags directly influences the approximation of 𝑓 we obtain. Luckily, the
definition of Riemann integrability guarantees that these Riemann sums 𝑆 (𝑓 , P) ¤ converge to a
21
meaningful real number, independently of our choice of tags , so long as we ensure that ∥ P¤ ∥ is
sufficiently small, i.e. provided we refine our approximation of 𝑓 sufficiently.
We now recall some basic properties of the Riemann integral that will/have been seen in the
lectures.

Theorem 3.1. Let [𝑎, 𝑏] ⊂ R be a compact interval and denote by R( [𝑎, 𝑏]) the set of all Riemann
integrable functions on [𝑎, 𝑏]. The following properties hold.
20 Given a set 𝑋 and a subset 𝐴 ⊆ 𝑋 , we define 1𝐴 (𝑥) to be the function given by the rule
(
1 if 𝑥 ∈ 𝐴,
1𝐴 (𝑥) :=
0 otherwise.

Some authors will instead write 𝜒𝐴 (𝑥) to denote this function.

21 The mesh function ∥·∥ does not take into account the tag points, and only thinks about the lengths of

the subintervals of [𝑎, 𝑏] that P creates.

50
 (i) If 𝑓 , 𝑔 ∈ R( [𝑎, 𝑏]) then 𝑓 + 𝑔 ∈ R( [𝑎, 𝑏]) and
ˆ 𝑏 ˆ 𝑏 ˆ 𝑏
(𝑓 + 𝑔) = 𝑓 + 𝑔.
𝑎 𝑎 𝑎

(ii) Given 𝑓 ∈ R( [𝑎, 𝑏]) and 𝛼 ∈ R we have 𝛼 𝑓 ∈ R( [𝑎, 𝑏]) with

ˆ 𝑏 ˆ 𝑏
(𝛼 𝑓 ) = 𝛼 𝑓.
𝑎 𝑎

Let us now give a detailed example in which we verify the Riemann integrability of an explicit
function.

Example 3.1. Fix a number 𝑐 ∈ (0, 1) and define 𝑓 : [0, 1] → R by

(
0 if 0 ≤ 𝑥 < 𝑐,
𝑓 (𝑥) :=
1 if 𝑐 ≤ 𝑥 ≤ 1.

We claim that 𝑓 is Riemann integrable on [0, 1] with integral equal to (1 − 𝑐). Let 𝜀 > 0 and take

𝛿 = 𝜀.

If P¤ is a tagged partition of [0, 1] with ∥ P¤ ∥ < 𝛿 then every subinterval of [0, 1] created by P¤ has
length no larger than 𝛿. Let us now enumerate the elements of P¤ as

0 = 𝑥 0 < 𝑥 1 < · · · < 𝑥𝑛 = 1

with tags 𝑥 ∗𝑗 ∈ [𝑥 𝑗 −1, 𝑥 𝑗 ] for every 𝑗 = 1, . . . , 𝑛. Let 𝑘 ≥ 1 be the unique integer such that
𝑥𝑘 −1 < 𝑐 ≤ 𝑥𝑘 . By definition of the function 𝑓 , we have 𝑓 (𝑥 ∗𝑗 ) = 0 for all 𝑗 < 𝑘.22 There are now
two cases that we will distinguish:

(1) Assume that 𝑥𝑘 −1 ≤ 𝑥𝑘∗ < 𝑐 ≤ 𝑥𝑘 . If 𝑘 = 𝑛 then 𝑓 (𝑥 ∗𝑗 ) = 0 for all 𝑗 = 1, . . . , 𝑛 so that

𝑆 (𝑓 ; P)
¤ − (1 − 𝑐) = |1 − 𝑐 | = |𝑥𝑘 − 𝑐 | ≤ P¤ < 𝛿.

In this last step we have used that both 𝑐 and 𝑥𝑘 belong to the same subinterval of P.
¤ If
22 Indeed, if 1 ≤ 𝑗 < 𝑘 then 𝑥 ∗𝑗 ≤ 𝑥𝑘 −1 < 𝑐 so that 𝑓 (𝑥 ∗𝑗 ) = 0.

51
 instead 𝑘 < 𝑛 then
𝑛
∑︁
𝑆 (𝑓 ; P)
¤ − (1 − 𝑐) = 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) − (1 − 𝑐)
𝑗=1
𝑛
∑︁
= 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) − (1 − 𝑐)
𝑗=𝑘+1

𝑛
∑︁
= (𝑥 𝑗 − 𝑥 𝑗 −1 ) − (1 − 𝑐)
𝑗=𝑘+1

= |(𝑥𝑛 − 𝑥𝑘 − (1 − 𝑐)|
= |(1 − 𝑥𝑘 ) − (1 − 𝑐)|
= |𝑥𝑘 − 𝑐 |
< 𝛿,

where we have once again used the fact that 𝑥𝑘 and 𝑐 belong to the same subinterval of P¤

(2) Otherwise, we have 𝑥𝑘 −1 < 𝑐 ≤ 𝑥𝑘∗ ≤ 𝑥𝑘 . Then, 𝑓 (𝑥𝑘∗ ) = 1 so that, as above,

𝑛
∑︁
𝑆 (𝑓 ; P)
¤ − (1 − 𝑐) = 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) − (1 − 𝑐)
𝑗=1
𝑛
∑︁
= 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) − (1 − 𝑐)
𝑗=𝑘

𝑛
∑︁
= (𝑥 𝑗 − 𝑥 𝑗 −1 ) − (1 − 𝑐)
𝑗=𝑘

= |(𝑥𝑛 − 𝑥𝑘 −1 − (1 − 𝑐)|
= |𝑥𝑘 −1 − 𝑐 |
< 𝛿.

In either case we have that

𝑆 (𝑓 ; P)
¤ − (1 − 𝑐) < 𝛿 = 𝜀.

Since P¤ was an arbitrary tagged partition of [0, 1] with ∥ P¤ ∥ < 𝛿, we see that 𝑓 is Riemann
integrable on [0, 1] and that ˆ 1
𝑓 = (1 − 𝑐).
0

52
 In addition to Example 3.1, let us also show that the “classical” function 𝑓 (𝑥) = 1 − 𝑥 is
Riemann integrable on [0, 1].

Example 3.2. Let 𝑓 : [0, 1] → R be given by 𝑓 (𝑥) = 1 − 𝑥. We claim that 𝑓 ∈ R( [0, 1]) and that
ˆ 1
1
𝑓 = .
0 2

Let 𝜀 > 0 be given and define 𝛿 := 𝜀. Let P¤ be a tagged partition of [0, 1] with ∥ P¤ ∥ < 𝛿. Denote
the partition points of P¤ by
0 = 𝑥 0 < 𝑥 1 < · · · < 𝑥𝑛 = 1
and let {𝑥 1∗, . . . , 𝑥𝑛∗ } be the tags of P.
¤ Let Q¤ be the tagged partition of [0, 1] formed by taking the
same partition points as P with tags
¤

𝑥 𝑗 + 𝑥 𝑗 −1
𝑦 ∗𝑗 := , 𝑗 = 1, . . . , 𝑛.
2
An easy calculation then shows that
𝑛 𝑛 h
∑︁ ∑︁ 𝑥 𝑗 + 𝑥 𝑗 −1 i
𝑆 (𝑓 ; Q)
¤ = 𝑓 (𝑦 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) = 1− (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1 𝑗=1
2

1 ∑︁  2
𝑛 𝑛 
𝑥 𝑗 − 𝑥 2𝑗 −1
∑︁
= (𝑥 𝑗 − 𝑥 𝑗 −1 ) −
𝑗=1
2 𝑗=1
1
=1−
2
1
= .
2
On the other hand, using the inequality |𝑎 1 + · · · + 𝑎𝑛 | ≤ |𝑎 1 | + · · · + |𝑎𝑛 |,
𝑛 
∑︁ 
𝑆 (𝑓 ; P)
¤ − 𝑆 (𝑓 ; Q)
¤ = 𝑓 (𝑥 ∗𝑗 ) − 𝑓 (𝑦 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) (3.2)
𝑗=1
𝑛
∑︁
≤ 𝑓 (𝑥 ∗𝑗 ) − 𝑓 (𝑦 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) (3.3)
𝑗=1
𝑛
∑︁ 𝑥 𝑗 + 𝑥 𝑗 −1
= 𝑥 ∗𝑗 − (𝑥 𝑗 − 𝑥 𝑗 −1 ). (3.4)
𝑗=1
2

𝑥 𝑗 +𝑥 𝑗 −1
Since 𝑥 ∗𝑗 ∈ 𝑥 𝑗 −1, 𝑥 𝑗 and is the midpoint of this same interval, we find that
 
2

𝑥 𝑗 + 𝑥 𝑗 −1
𝑥 ∗𝑗 − ≤ (𝑥 𝑗 − 𝑥 𝑗 −1 ) ≤ P¤ < 𝛿,
2

53
for each index 𝑗 = 1, . . . , 𝑛. Returning to (3.4) gives
𝑛
∑︁
𝑆 (𝑓 ; P)
¤ − 𝑆 (𝑓 ; Q)
¤ <𝛿 (𝑥 𝑗 − 𝑥 𝑗 −1 ) = 𝛿 = 𝜀.
𝑗=1

¤ = 1 , we see that
Finally, recalling that 𝑆 (𝑓 , Q) 2

1
𝑆 (𝑓 ; P)
¤ − = 𝑆 (𝑓 ; P)
¤ − 𝑆 (𝑓 ; Q)
¤ <𝜀
2

for all tagged partitions P¤ of [0, 1] with P¤ < 𝛿.

We finalize this section by providing a direct example of a function that will turn out to not
be Riemann integrable. We will provide a direct proof of this fact later, when we treat the Cauchy
Criterion for Riemann Integrability.

Example 3.3. Consider the Dirichlet function 𝔡 : R → R given by the rule

(
1 if 𝑥 ∈ Q,
𝔡(𝑥) := (3.5)
0 otherwise.

In other words, 𝔡(𝑥) = 1Q (𝑥). Then, as previously stated, we will later prove that 𝔡 is not inte-
grable on any compact interval [𝑎, 𝑏] ⊂ R despite being bounded.

3.1 Approximating by Riemann Sums

Given the striking similarities between the 𝜀 − 𝛿 definition of the limit and our definition of Rie-
´𝑏
mann integrability (see Definition 3.1), it is reasonable to hope that one can approximate 𝑎 𝑓 by
simply taking a limit along a sequence of Riemann sums with the mesh of the partitions tending to
zero. Informally, by sufficiently refining our tagged partitions, we should be able to approximate
the area under the graph of 𝑓 . This is confirmed by the following:

Lemma 3.2. Let 𝑓 : [𝑎, 𝑏] → R be Riemann integrable on [𝑎, 𝑏]. If P¤𝑛 is a sequence of tagged

partitions of [𝑎, 𝑏] such that

lim P¤𝑛 = 0,
then ˆ 𝑏
lim 𝑆 (𝑓 ; P¤𝑛 ) = 𝑓.
𝑎

Proof. Let 𝜀 > 0 be given. Since 𝑓 is Riemann integrable on [𝑎, 𝑏], there exists 𝛿 > 0 such that
ˆ 𝑏
𝑆 (𝑓 ; P)
¤ − 𝑓 <𝜀
𝑎

54
whenever P¤ is a tagged partition of [𝑎, 𝑏] with P¤ < 𝛿. Now, because P¤𝑛 → 0 as 𝑛 → ∞, we
can find 𝑁 ∈ N such that
P¤𝑛 < 𝛿
for all 𝑛 ≥ 𝑁 . It follows that
ˆ 𝑏
𝑆 (𝑓 ; P¤𝑛 ) − 𝑓 <𝜀
𝑎
´𝑏
for all 𝑛 ≥ 𝑁 . This shows that 𝑆 (𝑓 ; P¤𝑛 ) → 𝑎 𝑓 as 𝑛 → ∞. □

Remark 3.1. This lemma can sometimes make it relatively easy to extend properties of Riemann
sums to the Riemann integral. Informally speaking, if a certain property holds for Riemann sums
and this property is preserved by limits, one might use the lemma above to extend this property
to the integral. See the next example for an application of such an argument.
Example 3.4. Let 𝑓 , 𝑔 ∈ R( [𝑎, 𝑏]) be such that 𝑓 ≤ 𝑔 on [𝑎, 𝑏]. Let (P𝑛 ) be a sequence of tagged
partitions of [𝑎, 𝑏] such
that ∥ P¤𝑛 ∥ → 0 as 𝑛 → ∞. Since we have 𝑓 ≤ 𝑔 on all of [𝑎, 𝑏], one clearly
has 𝑆 (𝑓 ; P¤𝑛 ) ≤ 𝑆 𝑔, P¤𝑛 for each 𝑛 ∈ N. By virtue of Lemma 3.2 applied to both 𝑓 and 𝑔, we have
´𝑏 ´𝑏
that 𝑆 (𝑓 ; P¤𝑛 ) → 𝑎 𝑓 and 𝑆 𝑔, P¤𝑛 → 𝑎 𝑔 as 𝑛 → ∞. Since non-strict inequalities are preserved

by limits, we infer that

ˆ 𝑏 ˆ 𝑏
𝑓 ≤ 𝑔.
𝑎 𝑎
Note that this argument will fail if we do not know a priori that 𝑓 , 𝑔 ∈ R( [𝑎, 𝑏]).
Example 3.5. Let 𝑓 ∈ R( [𝑎, 𝑏]) be such that |𝑓 | is also Riemann integrable on [𝑎, 𝑏].23 Then,
ˆ 𝑏 ˆ 𝑏
𝑓 ≤ |𝑓 | .
𝑎 𝑎

Indeed, this follows from the previous example by observing that − |𝑓 | ≤ 𝑓 ≤ |𝑓 | on all of [𝑎, 𝑏].
Thus, ˆ ˆ ˆ ˆ
𝑏 𝑏 𝑏 𝑏
− |𝑓 | = (− |𝑓 |) ≤ 𝑓 ≤ |𝑓 |
𝑎 𝑎 𝑎 𝑎
´𝑏 ´𝑏
which gives that 𝑎 𝑓 ≤ 𝑎 |𝑓 |.

Example 3.6. Consider the function 𝑓 : [0, 1] → R given by

1
(
if 𝑥 ∈ Q \ {0},
𝑓 (𝑥) := 𝑥
0 if 𝑥 ∉ Q or 𝑥 = 0.
23 The integrability assumption on |𝑓 | turns out to be superfluous. Indeed, |𝑓 | is Riemann integrable
whenever 𝑓 is. However, we do not yet possess the tools required to show that 𝑓 ∈ R(𝑎, 𝑏]) implies that
|𝑓 | ∈ R( [𝑎, 𝑏]). We therefore have no qualms about making this assumption in our example.

55
Let (𝑥𝑛 ) be the sequence in [0, 1] defined by 𝑥𝑛 := 𝑛1 . Clearly, 𝑓 (𝑥𝑛 ) = 𝑛 for each 𝑛 ∈ N whence
𝑓 is unbounded on the interval [0, 1]. Consequently, 𝑓 cannot be Riemann integrable on [0, 1].
Nonetheless, one can find a sequence P¤𝑛 of tagged partitions of [0, 1] such that ∥ P¤𝑛 ∥ → 0 and

lim 𝑆 (𝑓 ; P¤𝑛 ) exists. Indeed, given 𝑛 ∈ N, divide [0, 1] into 𝑛-subintervals of equal length and
choose from each subinterval an irrational tag (this can be done because R \ Q is dense in R). The
resulting tagged partition P¤𝑛 will be such that ∥ P¤𝑛 ∥ ≤ 𝑛1 and 𝑆 (𝑓 ; P¤𝑛 ) = 0.

We also have the following sequential condition for non-integrability:

Corollary 3.3. Let 𝑓 : [𝑎, 𝑏] → R be a function. Let P¤𝑛 and Q¤ 𝑛 two sequences of tagged

partitions of [𝑎, 𝑏] such that

lim P¤𝑛 = lim Q¤ 𝑛 = 0.
If lim 𝑆 (𝑓 ; P¤𝑛 ) ≠ lim 𝑆 (𝑓 ; Q¤𝑛 ), then 𝑓 is not Riemann integrable on [𝑎, 𝑏]. Moreover, if one of these
limits does not exist, then 𝑓 ∉ R( [𝑎, 𝑏]).

Proof. By way of contradiction, let us assume that 𝑓 ∈ R( [𝑎, 𝑏]). Citing Lemma 3.2, we must
have ˆ 𝑏
lim 𝑆 (𝑓 ; P𝑛 ) =
¤ 𝑓 = lim 𝑆 (𝑓 ; Q¤𝑛 )
𝑎
which is a contradiction. □

Consider once more the Dirichlet function 𝑓 : R → R given by (3.5). We have already seen
that 𝑓 is discontinuous at every point 𝑐 ∈ R. Using Corollary 3.3, we will show that 𝑓 is not
Riemann integrable on any interval [𝑎, 𝑏] ⊆ R. Indeed, let (P𝑛 ) be any sequence of partitions of
[𝑎, 𝑏] with
𝑏 −𝑎
∥P𝑛 ∥ ≤
𝑛
for each 𝑛.24 From each subinterval of P𝑛 , we choose (by density) a rational tag and denote the
resulting tagged partition by P¤𝑛 . Similarly, from every subinterval of P𝑛 we choose an irrational
tag (again by density) and let Q¤ 𝑛 be the corresponding tagged partition of [𝑎, 𝑏]. Since P¤𝑛 and
Q¤ 𝑛 have the same partition points as P𝑛 , it is easy to see that

lim P¤𝑛 = lim Q¤ 𝑛 = 0.

On the other hand, because every tag of P¤𝑛 is rational, we have 𝑆 (𝑓 ; P¤𝑛 ) = (𝑏 −𝑎) for each 𝑛 ∈ N.
But, as every tag of Q¤𝑛 is irrational, we instead have 𝑆 (𝑓 ; Q¤ 𝑛 ) = 0 for each 𝑛 ∈ N. Citing Corollary
3.3, we infer that 𝑓 ∉ R( [𝑎, 𝑏]).
24 Such a sequence (P𝑛 ) can be obtained by dividing [𝑎, 𝑏] into 𝑛-subintervals of equal length.

56
3.2 Cauchy’s Criterion for Riemann Integrability
Let us recall the Cauchy Criterion for Riemann integrability, which characterizes the Riemann
integrability of a function without relying upon the “value” of its integral. Much like the case of
the real line, it is often far easier to practically deduce integrability through the Cauchy criterion
rather than the definition; this is especially true when there is no obvious “guess” for the integral.

Theorem 3.4 (Cauchy’s Criterion). Let 𝑓 : [𝑎, 𝑏] → R be a function. Then, 𝑓 is Riemann integrable
on [𝑎, 𝑏] if and only if for every 𝜀 > 0 there exists 𝛿 > 0 such that

𝑆 (𝑓 ; P)
¤ − 𝑆 (𝑓 ; Q)
¤ <𝜀

¤ Q¤ of [𝑎, 𝑏] such that ∥ P¤ ∥ < 𝛿 and ∥ Q∥

for all tagged partitions P, ¤ < 𝛿.

Negating the condition given above, we see that a function 𝑓 : [𝑎, 𝑏] → R is not Riemann
integrable on [𝑎, 𝑏] if and only if there exists 𝜀 0 > 0 such that, for every 𝛿 > 0, one can find tagged
partitions P¤ 𝛿 and Q¤ 𝛿 of [𝑎, 𝑏], each having mesh strictly less than 𝛿, such that

𝑆 (𝑓 ; P¤ 𝛿 ) − 𝑆 (𝑓 ; Q¤ 𝛿 ) ≥ 𝜀 0 .
1
In this case, taking 𝛿 := 𝑛 gives two tagged partitions P¤𝑛 and Q¤ 𝑛 such that

1 1
P¤𝑛 < and Q¤ 𝑛 <
𝑛 𝑛
with 𝑆 (𝑓 ; P¤𝑛 ) − 𝑆 (𝑓 ; Q¤ 𝑛 ) ≥ 𝜀 0 . Conversely, assume that we are given two sequences of tagged
partitions P¤𝑛 and Q¤ 𝑛 such that
lim P¤𝑛 = lim Q¤ 𝑛 = 0
but 𝑆 (𝑓 ; P¤𝑛 ) − 𝑆 (𝑓 ; Q¤ 𝑛 ) ≥ 𝜀 0 > 0 for each 𝑛 ∈ N. Can we conclude from this that 𝑓 ∉ R( [𝑎, 𝑏])?
Indeed, for any 𝛿 > 0, there exists 𝑁 ∈ N such that

P¤𝑛 < 𝛿 and Q¤ 𝑛 < 𝛿

for all 𝑛 ≥ 𝑁 . In particular, for 𝑛 = 𝑁 we have

P¤ 𝑁 < 𝛿 and Q¤ 𝑁 < 𝛿.

On the other hand,

𝑆 (𝑓 ; P¤ 𝑁 ) − 𝑆 (𝑓 ; Q¤ 𝑁 ) ≥ 𝜀 0 .
Taking P¤ 𝛿 := P¤ 𝑁 and Q¤ 𝛿 := Q¤ 𝑁 , we see that the Cauchy criterion fails and 𝑓 is not Riemann
integrable on [𝑎, 𝑏]. To summarize, we have proven the following:

Corollary 3.5. Let 𝑓 : [𝑎, 𝑏] → R be a function. Then, 𝑓 is not Riemann integrable on [𝑎, 𝑏] if and
only if there exists 𝜀 0 > 0 and two sequences ( P¤𝑛 ), ( Q¤ 𝑛 ) of tagged partitions of [𝑎, 𝑏] such that

57
 (i) ∥ P¤𝑛 ∥ → 0 and ∥ Q¤ 𝑛 ∥ → 0 as 𝑛 → ∞;
(ii) 𝑆 (𝑓 ; P¤𝑛 ) − 𝑆 (𝑓 ; Q¤ 𝑛 ) ≥ 𝜀 0 for all 𝑛 ∈ N.
Example 3.7. This criterion makes it possible to show that a large class of functions are not
Riemann integrable. Let 𝑔 : Q → R be a function such that 𝑔(𝑥) ≥ 𝜀 0 > 0 for all 𝑥 ∈ Q and some
𝜀 0 > 0. Define 𝑓 : [𝑎, 𝑏] → R by
(
𝑔(𝑥) if 𝑥 ∈ Q,
𝑓 (𝑥) :=
0 if 𝑥 ∉ Q.
Note that by taking 𝑔 = 1 we recover the Dirichlet function in (3.5). We now claim that 𝑓 is not
Riemann integrable on [𝑎, 𝑏]. To prove this, we will make use of the criterion proven in Corollary
3.5. For each 𝑛 ∈ N we can create an untagged partition P𝑛 of [𝑎, 𝑏] by dividing [𝑎, 𝑏] into
𝑛-subintervals of equal length. Clearly, this gives us a sequence of partitions of [𝑎, 𝑏] such that
𝑏 −𝑎
∥P𝑛 ∥ = →0
𝑛
as 𝑛 → ∞. For fixed 𝑛 ∈ N we can choose by density a rational tag from each subinterval of P𝑛 ;
doing so gives us a tagged partition P¤𝑛 of [𝑎, 𝑏] having only rational tags and the same partition
points as P𝑛 . Similarly, we build from P𝑛 a tagged partition Q¤ 𝑛 of [𝑎, 𝑏] having only irrational
tags and the same partition points as P𝑛 . Since P¤𝑛 and Q¤𝑛 have the same partition points as P𝑛 ,
it is obvious that
lim P¤𝑛 = lim Q¤ 𝑛 = 0.
On the other hand, because 𝑓 vanishes at every irrational number, 𝑆 (𝑓 ; Q¤ 𝑛 ) = 0 for all 𝑛 ∈ N.
n o𝑘
Consequently, letting 𝑥 0, . . . , 𝑥𝑘 and 𝑥 ∗𝑗 be the partition points and tags (respectively) of P¤𝑛
𝑗=1
for fixed 𝑛, we infer that
𝑘
∑︁
𝑆 (𝑓 ; P¤𝑛 ) − 𝑆 (𝑓 ; Q¤ 𝑛 ) = 𝑆 (𝑓 ; P¤𝑛 ) = 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1
𝑘
∑︁
= 𝑔(𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1
𝑘
∑︁
= 𝑔(𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1
𝑘
∑︁
≥ 𝜀0 (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1

= 𝜀 0 (𝑏 − 𝑎).
Since 𝑛 ∈ N was arbitrary, we see from Corollary 3.5 that 𝑓 cannot be Riemann integrable on
[𝑎, 𝑏].

58
3.3 The Squeeze Theorem
In addition to the Cauchy Criterion in Theorem 3.4, we possess the following criterion for Riemann
integrability:

Theorem 3.6 (Squeeze Theorem). Let 𝑓 : [𝑎, 𝑏] → R be a function. Then, 𝑓 is Riemann integrable
on [𝑎, 𝑏] if and only if, for each 𝜀 > 0, there exists two Riemann integrable functions 𝛼𝜀 , 𝜔𝜀 : [𝑎, 𝑏] →
R such that
𝛼𝜀 (𝑥) ≤ 𝑓 (𝑥) ≤ 𝜔𝜀 (𝑥), ∀𝑥 ∈ [𝑎, 𝑏],
and ˆ 𝑏
(𝜔𝜀 − 𝛼𝜀 ) < 𝜀.
𝑎

Remark 3.2. In practice, one will often use step-functions to “build” these 𝛼𝜀 , 𝜔𝜀 . This has the
´𝑏 ´𝑏
added benefit that the integrals 𝑎 𝛼𝜀 and 𝑎 𝜔𝜀 are easy to evaluate, for each 𝜀 > 0. Then, since
the Squeeze Theorem produces functions 𝛼𝜀 , 𝜔𝜀 with
ˆ 𝑏
𝛼𝜀 ≤ 𝑓 ≤ 𝜔𝜀 on [𝑎, 𝑏], and (𝜔𝜀 − 𝛼𝜀 ) < 𝜀.
𝑎

we have that ˆ ˆ ˆ
𝑏 𝑏 𝑏
𝛼𝜀 ≤ 𝑓 ≤ 𝜔𝜀
𝑎 𝑎 𝑎
by monotonicity (Example 3.4). Furthermore,
(ˆ ˆ ˆ ˆ )
𝑏 𝑏 𝑏 𝑏
0 ≤ max 𝑓 − 𝛼𝜀 , 𝜔𝜀 − 𝑓 < 𝜀.
𝑎 𝑎 𝑎 𝑎

It therefore follows that ˆ ˆ ˆ

𝑏 𝑏 𝑏
𝑓 = lim 𝜔𝜀 = lim 𝛼𝜀 .
𝑎 𝜀↘0 𝑎 𝜀↘0 𝑎

To see how Theorem 3.6 can be used in practice, let us consider the following example.

Example 3.8. Consider the function 𝑓 : [0, 1] → R given by 𝑥´↦→ 𝑥. Using the Squeeze Theorem,
1
we will prove that 𝑓 is Riemann integrable on [0, 1] and that 0 𝑥 = 21 . Given a 𝑛 ∈ N, let P𝑛 be
the partition of [0, 1] obtained by dividing [0, 1] into 𝑛-subintervals of equal length. Clearly,
1
lim ∥P𝑛 ∥ = lim = 0.
𝑛
Fix any 𝑛 ∈ N. The partition points of P𝑛 are given by
𝑗
𝑥𝑗 = , 𝑗 = 0, . . . , 𝑛
𝑛

59
Define two functions 𝛼𝑛 , 𝜔𝑛 on [0, 1] by
( (
𝑗 −1
if 𝑗 𝑛−1 ≤ 𝑥 < 𝑛𝑗 , 𝑗
if 𝑗 𝑛−1 ≤ 𝑥 < 𝑛𝑗 ,
𝛼𝑛 (𝑥) := 𝑛 and 𝜔𝑛 (𝑥) := 𝑛
1 if 𝑥 = 1, 1 if 𝑥 = 1.

and note that 𝛼𝑛 ≤ 𝑓 ≤ 𝜔𝑛 on [0, 1]. Since step functions are always integrable, 𝛼𝑛 and 𝜔𝑛 are
both Riemann integrable on [0, 1]. Now, a direct calculation shows that
1
(
if 𝑗 𝑛−1 ≤ 𝑥 < 𝑛𝑗 ,
(𝜔𝑛 − 𝛼𝑛 ) (𝑥) = 𝑛
0 if 𝑥 = 1

whence ˆ 1
1 𝑗 𝑗 −1 1 ∑︁ 1
𝑛
∑︁   𝑛
(𝜔𝑛 − 𝛼𝑛 ) = − = 2 1= .
0 𝑗=1
𝑛 𝑛 𝑛 𝑛 𝑗=1 𝑛
1
In particular, for any 𝑛 such that 𝑛 < 𝜀, one has
ˆ 1
𝛼𝑛 ≤ 𝑓 ≤ 𝜔 𝑛 and (𝜔𝑛 − 𝛼𝑛 ) < 𝜀.
0
´1
It follows from the Squeeze theorem that 𝑓 ∈ R( [0, 1]) and that 0 𝑓 exists. Now, note that
ˆ 1
𝑗 −1 1 1 ∑︁ 1 𝑛(𝑛 + 1)
𝑛
∑︁ 𝑛  
𝛼𝑛 = · = 2 ( 𝑗 − 1) = 2 −𝑛
0 𝑗=1
𝑛 𝑛 𝑛 𝑗=1 𝑛 2
1 1
= − .
2 𝑛
Similarly,
ˆ 1
𝑗 1 1 𝑛(𝑛 + 1) 1 1
𝑛
∑︁  
𝜔𝑛 = · = 2 = + .
0 𝑗=1
𝑛 𝑛 𝑛 2 2 2𝑛

By monotonicity of the integral,

ˆ 1 ˆ 1 ˆ 1
1 1 1 1
− = 𝛼𝑛 ≤ 𝑓 ≤ 𝜔𝑛 = + .
2 𝑛 0 0 0 2 2𝑛
Since 𝑛 ∈ N was arbitrary,
´1 we can take the limit as 𝑛 → ∞ and deduce from the Squeeze Theorem
(for sequences) that 0 𝑓 = 12 .

3.4 Additivity
You will have seen in the lectures (alternatively, see Bartle-Sherbert §7.2) that the Riemann inte-
gral is additive in the following sense:

60
Theorem 3.7 (Additivity of the Integral). Let 𝑓 : [𝑎, 𝑏] → R be a function. Then, 𝑓 is Riemann
integrable on [𝑎, 𝑏] if and only if, for each 𝑎 < 𝑐 < 𝑏, 𝑓 is Riemann integrable on [𝑎, 𝑐] and on [𝑐, 𝑏].
In this case, one has ˆ 𝑏 ˆ 𝑐 ˆ 𝑏
𝑓 = 𝑓 + 𝑓.
𝑎 𝑎 𝑐

The additivity theorem makes it much easier to construct suitable functions 𝛼𝜀 , 𝜔𝜀 for use in
the application of the Squeeze Theorem; we illustrate this below by way of an explicit example.

Example 3.9. We prove (using the Squeeze Theroem) that

sin 𝑥1
(
if 𝑥 ≠ 0,


𝑓 (𝑥) :=
0 if 𝑥 = 0.
1 𝜀 25
is Riemann integrable
 1 on
 [0, 1]. Fix an arbitrary 𝜀 > 0 and let 𝑁 ≥ 2 be such that 𝑁 < 2 . Since
 1 𝑓
is continuous on 𝑁 , 1 , it must be integrable there. Now, it is clear that 0 ≤ 𝑓 (𝑥) ≤ 1 on 0, 𝑁 .
Furthermore,
ˆ 1
𝑁 1
(1 − 0) = <𝜀
0 𝑁
Consider the functions

if 0 ≤ 𝑥 < 𝑁1 , if 0 ≤ 𝑥 < 𝑁1 ,
( (
0 1
𝛼𝜀 (𝑥) := and 𝜔𝜀 (𝑥) :=
𝑓 (𝑥) if 𝑁1 ≤ 𝑥 ≤ 1 𝑓 (𝑥) if 𝑁1 ≤ 𝑥 ≤ 1.

Clearly, 𝛼𝜀 ≤ 𝑓 ≤ 𝜔𝜀 on all of [0, 1].

 1 Since 𝑓 is Riemann integrable on 𝑁1 , 1 and 𝛼𝜀 = 𝑓 there,
 

𝛼𝜀 is also Riemann integrable on 𝑁 , 1 . On the other hand, 𝛼𝜀 is equal to a constant

 1  function

1 1
on 0, 𝑁 except at the one point 𝑥 = 𝑁 . Thus, 𝛼𝜀 is integrable on the interval 0, 𝑁 . By the
Additivity Theorem, we infer that 𝛼𝜀 ∈ R( [0, 1]). Similarly, 𝜔𝜀 ∈ R( [0, 1]). Finally, another
application of the Additivity Theorem gives
ˆ 1 ˆ 1/𝑁 ˆ 1
(𝜔𝜀 − 𝛼𝜀 ) = (𝜔𝜀 − 𝛼𝜀 ) + (𝜔𝜀 − 𝛼𝜀 )
0 0 1/𝑁
ˆ 1/𝑁 ˆ 1
= (1 − 0) + 0
0 1/𝑁
1
=
𝑁
< 𝜀.

By the Squeeze Theorem, we infer that 𝑓 ∈ R( [0, 1]).

25 I 1
only want 𝑁 ≥ 2 to guarantee that 𝑁 < 1. This way, both [0, 1/𝑁 ] and [1/𝑁 , 1] are non-trivial
intervals.

61
Remark 3.3. Let 𝑓 : [𝑎, 𝑏] → R be a function and assume that 𝑓 ∈ R( [𝑐, 𝑏]) for all 𝑐 ∈ (𝑎, 𝑏). Does
it follow that 𝑓 ∈ R( [𝑎, 𝑏])? Unfortunately, one cannot conclude that 𝑓 is Riemann integrable on
[𝑎, 𝑏]. Consider the function
1
(
if 𝑥 ≠ 0,
𝑓 (𝑥) := 𝑥
0 if 𝑥 = 0.
Since 𝑓 is continuous on [𝑐, 1] for every 0 < 𝑐 < 1, we see that 𝑓 is Riemann integrable on
[𝑐, 1] for each such 𝑐. On the other hand, because 𝑓 is unbounded on [0, 1], it cannot be Riemann
integrable there.
Despite this, one can recover the phenomenon from Example 3.9 if we assume that 𝑓 is
bounded and belongs to R( [𝑐, 𝑏]) for all 𝑐 ∈ (𝑎, 𝑏). Indeed, this is confirmed by the following
integral analogue to Proposition 2.12

Proposition 3.8. Let 𝑓 : [𝑎, 𝑏] → R be a bounded function such that, for each 𝑐 ∈ (𝑎, 𝑏), there
holds 𝑓 ∈ R( [𝑐, 𝑏]). Then, 𝑓 ∈ R( [𝑎, 𝑏]).

Proof. We adapt the argument used in the solution of Example 3.9. Let 𝜀 > 0 be given; since 𝑓 is
bounded, there exists some 𝑀 > 0 such that −𝑀 ≤ 𝑓 (𝑥) ≤ 𝑀 for all 𝑥 ∈ [𝑎, 𝑏]. Next, let 𝛿 > 0 be
so small that
𝜀
𝛿< and 𝑎 + 𝛿 < 𝑏.
2𝑀
We construct functions 𝛼𝜀 and 𝜔𝜀 on [𝑎, 𝑏] as follows:
( (
−𝑀 if 𝑎 ≤ 𝑥 < 𝑎 + 𝛿, 𝑀 if 𝑎 ≤ 𝑥 < 𝑎 + 𝛿,
𝛼𝜀 (𝑥) := 𝜔𝜀 (𝑥) :=
𝑓 (𝑥) if 𝑎 + 𝛿 ≤ 𝑥 ≤ 𝑏, 𝑓 (𝑥) if 𝑎 + 𝛿 ≤ 𝑥 ≤ 𝑏,

By assumption, 𝛼𝜀 , 𝜔𝜀 is integrable on [𝑎 + 𝛿, 𝑏] because 𝑓 is assumed to be integrable there.

Furthermore, 𝛼𝜀 , 𝜔𝜀 are equal to a constant function on all of [𝑎, 𝑎 +𝛿). Hence, 𝛼𝜀 , 𝜔𝜀 are integrable
on [𝑎, 𝑎 + 𝛿]. By the additivity theorem, we see that 𝛼𝜀 , 𝜔𝜀 ∈ R( [𝑎, 𝑏]). Furthermore, it is obvious
that
𝛼𝜀 ≤ 𝑓 ≤ 𝜔𝜀 , on [𝑎, 𝑏].
Finally, notice that, by the additivity theorem,
ˆ 𝑏 ˆ 𝑎+𝛿 ˆ 𝑏
(𝜔𝜀 − 𝛼𝜀 ) = (𝜔𝜀 − 𝛼𝜀 ) + (𝜔𝜀 − 𝛼𝜀 )
ˆ ˆ
𝑎 𝑎 𝑎+𝛿
𝑎+𝛿 𝑏
= (𝑀 − (−𝑀)) + 0
𝑎 𝑎+𝛿
= 2𝑀𝛿
< 𝜀.

Thus, it follows from the Squeeze Theorem that 𝑓 ∈ R( [𝑎, 𝑏]). □

62
3.5 Examples with Riemann Sums: Even and Odd Functions
Let 𝑓 : [𝑎, 𝑏] → R be Riemann integrable and let P¤𝑛 be a sequence of tagged partitions of [𝑎, 𝑏]

such that ∥ P¤𝑛 ∥ → 0, as 𝑛 → ∞. Lemma 3.2 states that the Riemann sums 𝑆 (𝑓 ; P¤𝑛 ) converge to
´𝑏
𝑎 𝑓 , i.e. ˆ 𝑏
lim 𝑆 (𝑓 ; P𝑛 ) =
¤ 𝑓.
𝑎
As demonstrated in Example 3.5, this can sometimes help us extend properties of Riemann sums
to the Riemann integral. This type of argument is very common in analysis and is extremely
useful. We provide another application of this argument below:

Proposition 3.9. Let 𝑎 > 0 and 𝑓 : [−𝑎, 𝑎] → R be Riemann Integrable. In particular, 𝑓 is Riemann
integrable on every closed subinterval of [−𝑎, 𝑎]. If 𝑓 is even, i.e. if 𝑓 (𝑥) = 𝑓 (−𝑥) for all 𝑥 ∈ [−𝑎, 𝑎],
then ˆ 𝑎 ˆ 𝑎
𝑓 =2 𝑓.
−𝑎 0

Proof. For each 𝑛 ∈ N we construct a partition P𝑛 of [0, 𝑎] by dividing this interval in to 𝑛-

subintervals of equal length. Clearly, ∥P𝑛 ∥ = 𝑛𝑎 for each 𝑛 ∈ N. Let 𝑥 1∗, . . . , 𝑥𝑘∗ be any set of tags
for P𝑛 and let P¤𝑛 denote the resulting tagged partition of [𝑎, 𝑏]. Let

0 = 𝑥 0 < · · · < 𝑥𝑛 = 𝑎

be the partition points of P¤𝑛 . Let Q𝑛 be the tagged partition of [−𝑎, 𝑎] with partition points

−𝑎 = −𝑥𝑛 < −𝑥𝑛−1 < · · · < −𝑥 1 < 𝑥 0 < · · · < 𝑥𝑛 = 𝑎.

Finally, we give Q𝑛 the tags

(
𝑥 ∗𝑗 ∈ 𝑥 𝑗 −1, 𝑥 𝑗 1 ≤ 𝑗 ≤ 𝑛,
 

−𝑥 ∗𝑗 ∈ −𝑥 𝑗 , −𝑥 𝑗 −1 1 ≤ 𝑗 ≤ 𝑛.
 

Then, Q¤ 𝑛 is tagged partition of [−𝑎, 𝑎] with ∥ Q¤ 𝑛 ∥ = ∥ P¤𝑛 ∥ = 𝑛𝑎 . Since 𝑓 is an even function,

𝑛
∑︁ 𝑛
∑︁
𝑆 (𝑓 ; Q¤ 𝑛 ) = (−𝑥 ∗𝑗 ) 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )


𝑓 −𝑥 𝑗 −1 − (−𝑥 𝑗 ) +
𝑗=1 𝑗=1
𝑛
∑︁ 𝑛
∑︁
= 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) + 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1 𝑗=1

= 2𝑆 (𝑓 ; P¤𝑛 ).

63
As 𝑛 ∈ N was arbitrary, this gives us a sequence of tagged partitions P¤𝑛 of [0, 𝑎], and a related
sequence Q¤ 𝑛 of tagged partitions of [−𝑎, 𝑎], such that
𝑎
lim Q¤ 𝑛 = lim P¤𝑛 = lim = 0.
𝑛
Consequently, two applications of Lemma 3.2 implies
ˆ 𝑎 ˆ 𝑎
𝑓 = lim 𝑆 (𝑓 ; Q𝑛 ) = 2 lim 𝑆 (𝑓 ; P𝑛 ) = 2
¤ ¤ 𝑓.
−𝑎 0

A similar argument will allow us to obtain the following:

Proposition 3.10. Let 𝑎 > 0 and 𝑓 : [−𝑎, 𝑎] → R be Riemann integrable. In particular, 𝑓 is

Riemann integrable on every closed subinterval of [−𝑎, 𝑎]. If 𝑓 is odd, i.e. if −𝑓 (𝑥) = 𝑓 (−𝑥) for all
𝑥 ∈ [−𝑎, 𝑎], then ˆ 𝑎
𝑓 = 0.
−𝑎

Proof. For each 𝑛 ∈ N we defined tagged partitions P¤𝑛 and Q¤ 𝑛 as in the proof of the Proposition
3.9. However, in this case, we obtain
𝑛
∑︁ 𝑛

∑︁
𝑆 (𝑓 ; Q¤ 𝑛 ) = 𝑓 (−𝑥 ∗𝑗 ) −𝑥 𝑗 −1 − (−𝑥 𝑗 ) + 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1 𝑗=1
𝑛
∑︁ 𝑛
∑︁
=− 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 ) + 𝑓 (𝑥 ∗𝑗 ) (𝑥 𝑗 − 𝑥 𝑗 −1 )
𝑗=1 𝑗=1

= 0.

By the same argument as before, Lemma 3.2 yields

ˆ 𝑎
𝑓 = lim 𝑆 (𝑓 ; Q¤ 𝑛 ) = 0.
−𝑎

3.6 Integrating Continuous Functions

We know that all monotone increasing functions [𝑎, 𝑏] → R are Riemann integrable. Perhaps
even more importantly, it has been proven that continuous functions are always Riemann inte-
grable. The Riemann integral enjoys much nicer properties when restricted to continuous func-
´𝑏
tions. Namely, much more can be said about 𝑎 𝑓 when 𝑓 is assumed to be continuous on [𝑎, 𝑏].

64
 ´𝑏
Proposition 3.11. Let 𝑓 : [𝑎, 𝑏] → R be continuous and assume that 𝑓 ≥ 0 on [𝑎, 𝑏]. Then 𝑎 𝑓 = 0
if and only if 𝑓 = 0 on [𝑎, 𝑏].
´𝑏
Proof. Clearly, if 𝑓 = 0 on [𝑎, 𝑏] then 𝑎 𝑓 = 0. Conversely, we show that if 𝑓 (𝑐) > 0 at some
´𝑏
point 𝑐 ∈ [𝑎, 𝑏] then 𝑎 𝑓 > 0. Indeed, by continuity, there exists 𝛿 > 0 such that

𝑓 (𝑐)
|𝑓 (𝑥) − 𝑓 (𝑐)| <
2
for all 𝑥 ∈ [𝑎, 𝑏] with |𝑥 − 𝑐 | < 𝛿. Therefore, for all such 𝑥, the above implies that

𝑓 (𝑐) 3𝑓 (𝑐)
< 𝑓 (𝑥) < .
2 2
Let now 𝐼 = [𝑐, 𝑑] ⊆ [𝑎, 𝑏] be the interval (try drawing a picture) given by
 
𝛿 𝛿
[𝑎, 𝑏] ∩ 𝑐 − , 𝑐 + .
2 2
𝑓 (𝑐 )
If 𝑥 ∈ [𝑐, 𝑑] then |𝑥 − 𝑐 | < 𝛿 so that 𝑓 (𝑥) > 2 by the above. By the additivity theorem, we
know that ˆ 𝑏 ˆ 𝑐 ˆ ˆ ˆ
𝑑 𝑑 𝑑
𝑓 = 𝑓 + 𝑓 + 𝑓 ≥ 𝑓.
𝑎 𝑎 𝑐 𝑏 𝑐
Here we have used that 𝑓 (𝑥) ≥ 0 on [𝑎, 𝑏] and that the integral is monotone (by your assignment
problem). It follows that
ˆ 𝑏 ˆ 𝑑 ˆ 𝑑
𝑓 (𝑐) 𝑓 (𝑐)
𝑓 ≥ 𝑓 ≥ = (𝑑 − 𝑐) > 0.
𝑎 𝑐 𝑐 2 2

This completes the proof. □

Remark 3.4. It is easy to see that one cannot drop the continuity assumption on 𝑓 . Indeed, the
function 𝑓 : [0, 1] → R given by 𝑓 (𝑥) = 0 for all 0 ≤ 𝑥 < 1 and 𝑓 (1) = 1. Then, as 𝑓 differs
´𝑏
from the constant map 𝑥 ↦→ 0 at only a single point, we have that 𝑓 ∈ R( [0, 1]) with 𝑎 𝑓 = 0.
However, 𝑓 . 0.
Since compositions of continuous functions are continuous, |𝑓 | is Riemann integrable on [𝑎, 𝑏]
whenever 𝑓 is a continuous function on [𝑎, 𝑏]. Recalling Example 3.5, we obtain the following:

Proposition 3.12. Let 𝑓 : [𝑎, 𝑏] → R be continuous. Then |𝑓 | is Riemann integrable on [𝑎, 𝑏] and
ˆ 𝑏 ˆ 𝑏
𝑓 ≤ |𝑓 | .
𝑎 𝑎

65
 Let us now discuss the mean value theorem for integrals. As in the case of differentiable
functions, there is a generalized “Cauchy” Mean Value Theorem to explore.

Theorem 3.13. Let 𝑓 , 𝑔 : [𝑎, 𝑏] → R be continuous and assume that 𝑔(𝑥) > 0 on all of [𝑎, 𝑏]. Then,
´𝑏
𝑎 𝑔 > 0 and there exists a point 𝑐 ∈ [𝑎, 𝑏] such that
ˆ 𝑏 ˆ 𝑏
𝑓 𝑔 = 𝑓 (𝑐) 𝑔. (3.6)
𝑎 𝑎
´𝑏
Proof. First, we argue that 𝑎 𝑔 > 0. Indeed, since 𝑔 is continuous on the interval [𝑎, 𝑏], it achieves
an absolute minimum 𝜇 on [𝑎, 𝑏]. In fact, since 𝑔(𝑥) > 0 on [𝑎, 𝑏], we must have 𝜇 > 0. Conse-
quently, by monotonicity,
ˆ 𝑏 ˆ 𝑏
𝑔≥ 𝜇 = 𝜇 (𝑏 − 𝑎) > 0.
𝑎 𝑎
We now establish the equality in (3.6). Using that 𝑓 is continuous on [𝑎, 𝑏], the extreme value
theorem implies that 𝑓 achieves an absolute minimum 𝑚 and an absolute maximum 𝑀 on [𝑎, 𝑏].
Then, since
´𝑏 ´𝑏 ´𝑏
𝑎 (𝑚𝑔) 𝑎 𝑓𝑔 (𝑀𝑔)
𝑚 = ´𝑏 ≤ ´ 𝑏 ≤ 𝑎´ 𝑏 = 𝑀.
𝑎 𝑔 𝑎 𝑔 𝑎 𝑔

Then, the Intermediate Value Theorem ensures the existence of a point 𝑐 ∈ [𝑎, 𝑏] such that
´𝑏
𝑓𝑔
𝑓 (𝑐) = ´𝑎 𝑏
𝑎 𝑔

which completes the proof. □

Remark 3.5. Observe that one cannot drop the requirement that 𝑔(𝑥) be positive for all 𝑥 ∈ [𝑎, 𝑏].
Indeed, consider the functions 𝑓 (𝑥) = 𝑔(𝑥) = 𝑥 on [−1, 1]. Clearly, each function is continuous
(even Lipschitz) on [−1, 1]. Furthermore, we will later be able to prove (using the fundamental
theorem of calculus – see §3.8) that
ˆ 1
𝑥3 2
𝑥=1
2
𝑥 = = .
−1 3 𝑥=−1 3

Similarly26 , ˆ 1
𝑥 = 0.
−1
2
Clearly, there cannot exist a point 𝑐 ∈ [−1, 1] such that 3 = 𝑓 (𝑐) · 0.
26 This also follows from Proposition 3.10.

66
Theorem 3.14 (Mean Value Theorem for Integrals). Let 𝑓 : [𝑎, 𝑏] → R be continuous. Then, there
exists a point 𝑐 ∈ [𝑎, 𝑏] such that
ˆ 𝑏
1
𝑓 = 𝑓 (𝑐).
𝑏 −𝑎 𝑎
Proof Using Theorem 3.13. Apply Theorem 3.13 with 𝑔 ≡ 1. □

Direct Proof. Since [𝑎, 𝑏] is compact and 𝑓 is continuous, 𝑓 achieves an absolute minimum 𝑚 and
an absolute maximum 𝑀 on [𝑎, 𝑏]. By the monotonicity of the integral, we must have
ˆ 𝑏
𝑚(𝑏 − 𝑎) ≤ 𝑓 ≤ 𝑀 (𝑏 − 𝑎).
𝑎

Consequently,
ˆ
1 𝑏
𝑚≤ 𝑓 ≤ 𝑀.
𝑏 −𝑎 𝑎
By the intermediate value theorem, 𝑓 must achieve every value in [𝑚, 𝑀]. In particular, there
1
´𝑏
exists a point 𝑐 ∈ [𝑎, 𝑏] such that 𝑓 (𝑐) = 𝑏 −𝑎 𝑎 𝑓 . This completes the proof. □
´𝑏 ´𝑏
Corollary 3.15. Let 𝑓 , 𝑔 : [𝑎, 𝑏] → R be continuous. If 𝑎 𝑓 = 𝑎 𝑔, there exists a point 𝑐 ∈ [𝑎, 𝑏]
such that 𝑓 (𝑐) = 𝑔(𝑐).

Proof. Applying the Mean Value Theorem for Integrals to 𝑓 − 𝑔 implies the existence of a point
𝑐 ∈ [𝑎, 𝑏] such that
ˆ ´𝑏 ´𝑏
1 𝑏 𝑓 − 𝑎𝑔
(𝑓 − 𝑔) (𝑐) = 𝑓 (𝑐) − 𝑔(𝑐) = (𝑓 − 𝑔) = 𝑎
= 0.
𝑏 −𝑎 𝑎 𝑏 −𝑎

That is, 𝑓 (𝑐) = 𝑔(𝑐). □

In Example 3.8, we proved that the function 𝑓 (𝑥) = 𝑥 is Riemann integrable on [0, 1] using
the Squeeze Theorem. There, we approximated 𝑓 above and below by conveniently chosen step
functions. More precisely, we divided [0, 1] into 𝑛-subintervals {𝐼 𝑗 }𝑛𝑗=1 of equal length and defined
two step functions 𝛼𝑛 and 𝜔𝑛 by taking 𝛼𝑛 to be the minimum of 𝑓 over 𝐼 𝑗 and 𝜔𝑛 to be the
maximum. In this sense, we approximating 𝑓 from above and below by step functions taking on
the extrema of 𝑓 (on these 𝐼 𝑗 ). This is precisely the idea behind the Darboux integral, which we
briefly discuss.

67
3.7 Darboux Integration
Let 𝐼 = [𝑎, 𝑏] be a compact interval and let P be a partition of [𝑎, 𝑏]. Let

𝑎 = 𝑥 0 < · · · < 𝑥𝑛 = 𝑏

be the partition points of P. Given a bounded function 𝑓 : [𝑎, 𝑏] → R, we can define the upper
and lower Darboux sums of 𝑓 on P, respectively, by
𝑛
∑︁
𝐿(𝑓 ; P) := 𝑚 𝑗 (𝑥 𝑗 − 𝑥 𝑗 −1 ), 𝑚 𝑗 := inf 𝑓 (𝑥), (3.7)
𝑥 ∈ [𝑥 𝑗 −1 ,𝑥 𝑗 ]
𝑗=1
𝑛
∑︁
𝑈 (𝑓 ; P) := 𝑀 𝑗 (𝑥 𝑗 − 𝑥 𝑗 −1 ), 𝑀 𝑗 := sup 𝑓 (𝑥). (3.8)
𝑗=1 𝑥 ∈ [𝑥 𝑗 −1 ,𝑥 𝑗 ]

Since 𝑚 𝑗 ≤ 𝑀 𝑗 for all 𝑗 = 1, . . . , 𝑛, we see that 𝐿(𝑓 ; P) ≤ 𝑈 (𝑓 ; P). Furthermore, let 𝑀 > 0 be
such that −𝑀 ≤ 𝑓 (𝑥) ≤ 𝑀 on [𝑎, 𝑏]. Then, 𝐿(𝑓 ; P) is bounded from above independently of P.
Indeed,
𝑛
∑︁ 𝑛
∑︁
𝐿(𝑓 ; P) = 𝑚 𝑗 (𝑥 𝑗 − 𝑥 𝑗 −1 ) ≤ 𝑀 (𝑥 𝑗 − 𝑥 𝑗 −1 ) = 𝑀 (𝑏 − 𝑎).
𝑗=1 𝑗=1

Similarly,
𝑛
∑︁
𝑈 (𝑓 ; P) ≥ −𝑀 (𝑥 𝑗 − 𝑥 𝑗 −1 ) = −𝑀 (𝑏 − 𝑎).
𝑗=1

Hence, 𝑈 (𝑓 ; P) is bounded from below independently of P.

Definition 3.2. Let 𝐼 = [𝑎, 𝑏] and 𝑓 : 𝐼 → R be a bounded function. Denote by P (𝐼 ) the

collection of all partitions P of [𝑎, 𝑏]. The lower and upper Darboux integrals of 𝑓 are defined,
respectively, by

𝐿(𝑓 ) := sup 𝐿(𝑓 ; P) and 𝑈 (𝑓 ) := inf 𝑈 (𝑓 ; P).

P ∈P (𝐼 ) P ∈P (𝐼 )

Note that these quantities exist by our previous argument. We say that 𝑓 is Darboux integrable
on [𝑎, 𝑏] if 𝑈 (𝑓 ) = 𝐿(𝑓 ). In this case the Darboux integral of 𝑓 on [𝑎, 𝑏] is defined as:
ˆ 𝑏
𝑓 := 𝑈 (𝑓 ) = 𝐿(𝑓 ).
𝑎

One can show (see Bartle §7.4) that a function is Darboux integrable on [𝑎, 𝑏] if and only if it is
Riemann integrable.

68
 Unlike the Riemann integral, the Darboux integral declares a function 𝑓 to be integrable if it
can be approximated from above and below by step functions. Put otherwise, a bounded function
𝑓 is considered integrable if and only if if can be “squeezed” between two step functions. This
bears a notable resemblance to the statement of the squeeze theorem 3.6.
Although the Riemann and Darboux integrals are equivalent, the Riemann one is “better” in
the sense that it more easily generalizes. By making a minor modification to the definition of the
Riemann integral (replacing 𝛿 with something called a gauge), one obtains the so-called gauge
integral. As with the Riemann integral, the gauge integral is defined on intervals. In fact, the
gauge integral can be defined on unbounded intervals. On an interval 𝐼 , it turns out that the
gauge integral is more general than the Lebesgue integral. However, the Lebesgue integral can
be defined in much more general settings than the gauge integral. For instance, the Lebesgue
integral can be defined on subsets of R that are not intervals.

3.8 Lebesgue Criterion & The Fundamental Theorem of Calculus

Before presenting both parts of Fundamental Theorem of Calculus, we briefly review the Lebesgue
integrability criterion. A proof of this result may be found in the Appendix of Bartle-Sherbert, as
this proof is relatively technical. Regardless, it is an important result which provides a necessary
and sufficient condition for a bounded function to be Riemann integrable, purely in terms of its
continuity points. We reiterate the formal statement below.
Theorem 3.16 (Lebesgue Integrability Criterion). Let 𝑓 : [𝑎, 𝑏] → R be a bounded function. Then,
𝑓 is Riemann integrable on [𝑎, 𝑏] if and only if the set
{𝑥 ∈ [𝑎, 𝑏] : 𝑓 is discontinuous at 𝑥 }
has Lebesgue measure zero.
This result is very useful as it shows that continuous functions map Riemann integrable func-
tions to Riemann integrable functions. In particular, it enables us to show that |𝑓 | is Riemann
integrable wherever 𝑓 is.
Corollary 3.17. Let 𝑓 : [𝑎, 𝑏] → R be Riemann integrable and 𝜑 a bounded continuous real-valued
function whose domain contains the image of 𝑓 . Then, 𝜑 ◦ 𝑓 is Riemann integrable on [𝑎, 𝑏].
Proof. Put 𝑔(𝑥) := (𝜑 ◦ 𝑓 ) (𝑥). Observe that 𝑓 ∈ R( [𝑎, 𝑏]) implies that 𝑓 is bounded; since 𝜑
is bounded, so is 𝑔. Furthermore, if 𝑓 is continuous at 𝑥 ∈ [𝑎, 𝑏], then so is 𝑔 by composition.
Consequently, defining
Zℎ := {𝑥 ∈ [𝑎, 𝑏] : ℎ is discontinuous at 𝑥 }
for any function ℎ : [𝑎, 𝑏] → R, it follows from the Lebesgue criterion that Z 𝑓 has Lebesgue
measure 0. By our argument above, we find that Z𝑔 ⊆ Z 𝑓 whence Z𝑔 has Lebesgue measure zero.
Applying the Lebsegue criterion for integrability, we infer that 𝑔 is also Riemann integrable on
[𝑎, 𝑏]. □

69
 As stated previously, the afore Corollary allows us to strengthen the conclusions of Example
3.5. Indeed, we need no longer assume that 𝑓 is continuous in the context of Proposition 3.12

Proposition 3.18. Let 𝑓 : [𝑎, 𝑏] → R be Riemann integrable. Then |𝑓 | is Riemann integrable on

[𝑎, 𝑏] and
ˆ 𝑏 ˆ 𝑏
𝑓 ≤ |𝑓 | .
𝑎 𝑎

3.8.1 The Fundamental Theorem of Calculus

There are two forms of the Fundamental Theorem of Calculus, both of which are proven in §7.3
of Bartle-Sherbert. For the sake of completeness, we reiterate their statements below. For the
proofs, we of course refer the reader to our aforementioned reference.

Theorem 3.19 (Fundamental Theorem of Calculus – Form 1). Let [𝑎, 𝑏] ⊂ R be a compact interval
and 𝐸 ⊂ [𝑎, 𝑏] a finite set. Let 𝑓 , 𝐹 : [𝑎, 𝑏] → R be functions such that the following hold:

(a) 𝐹 is continuous on [𝑎, 𝑏];

(b) 𝐹 ′ (𝑥) = 𝑓 (𝑥) for 𝑥 ∈ [𝑎, 𝑏] \ 𝐸;

(c) 𝑓 is Riemann integrable on [𝑎, 𝑏].

Then, ˆ 𝑏
𝑓 = 𝐹 (𝑏) − 𝐹 (𝑎). (3.9)
𝑎

Typically, students are quite familiar with a basic form of this result (when 𝐸 = ∅ most often).
Indeed, this is main result one uses in a standard Calculus course to calculate integrals using
antiderivatives!

Theorem 3.20 (Fundamental Theorem of Calculus – Form 2). Let 𝑓 : [𝑎, 𝑏] → R be Riemann
integrable and suppose that 𝑓 is continuous at a point 𝑐 ∈ [𝑎, 𝑏]. Then, the function
ˆ 𝑥
𝐹 : [𝑎, 𝑏] → R, 𝑥 ↦→ 𝑓 (3.10)
𝑎

is differentiable at 𝑐. Moreover, 𝐹 ′ (𝑐) = 𝑓 (𝑐). The mapping 𝐹 is called the indefinite integral of 𝑓
with basepoint 𝑎.

Example 3.10. Suppose 𝑓 : [𝑎, 𝑏] → R is a Riemann integrable function. Then, the function 𝐹
defined by (3.10) is continuous. In fact 𝐹 is even Lipschitz continuous.

70
Solution. To see this, first let 𝑀 > 0 be such that

|𝑓 (𝑥)| ≤ 𝑀 ∀𝑥 ∈ [𝑎, 𝑏].

Given 𝑥, 𝑦 ∈ [𝑎, 𝑏] with 𝑥 < 𝑦 there holds

ˆ 𝑦 ˆ 𝑥 ˆ 𝑦
|𝐹 (𝑦) − 𝐹 (𝑥)| = 𝑓 − 𝑓 = 𝑓 .
𝑎 𝑎 𝑥

By Proposition 3.18, ˆ 𝑦
|𝐹 (𝑦) − 𝐹 (𝑥)| ≤ |𝑓 | .
𝑥
Using that |𝑓 | ≤ 𝑀, we conclude that
ˆ 𝑦
|𝐹 (𝑦) − 𝐹 (𝑥)| ≤ 𝑀 = 𝑀 (𝑦 − 𝑥) = 𝑀 |𝑦 − 𝑥 | .
𝑥

In other words, 𝐹 is Lipschitz continuous. □

Example 3.11. Suppose 𝑓 : [𝑎, 𝑏] → R is a continuous function satisfying

ˆ 𝑥 ˆ 𝑏
𝑓 = 𝑓 ∀𝑥 ∈ [𝑎, 𝑏].
𝑎 𝑥
´𝑥
Then, define 𝐹 : [𝑎, 𝑏] → R by 𝐹 (𝑥) := 𝑎 𝑓 and observe that
ˆ 𝑥 ˆ 𝑏 ˆ 𝑏 ˆ 𝑥
𝐹 (𝑥) = 𝑓 = 𝑓 = 𝑓 − 𝑓 = 𝐹 (𝑏) − 𝐹 (𝑥).
𝑎 𝑥 𝑎 𝑎

By the fundamental theorem of calculus, taking the derivative on either side of the above equation
yields
𝑓 (𝑥) = 0 − 𝑓 (𝑥) = −𝑓 (𝑥).
Thus, 𝑓 (𝑥) = 0 for all 𝑥 ∈ [𝑎, 𝑏].

71
4 Sequences of Functions
We now turn towards the study of function sequences. More precisely, rather than simply study
the convergence of sequences consisting of real (or complex) numbers, we analyze the conver-
gence of sequences of functions. Formally, given a set 𝑋 , let us denote by R𝑋 the set of all functions
𝑋 → R. A sequence of functions defined on 𝑋 is a function 𝑓 : N → R𝑋 . That is, we have for
each 𝑛 ∈ N an associated function 𝑓 (𝑛) : 𝑋 → R.27 Typically, since each 𝑓 (𝑛) is itself a function
𝑋 → R, we write 𝑓𝑛 to denote the function 𝑓 (𝑛) so that 𝑓𝑛 (𝑥) = (𝑓 (𝑛)) (𝑥) for all points 𝑥 ∈ 𝑋 .

Definition 4.1 (Pointwise Convergence). Let 𝐴 ⊆ R and (𝑓𝑛 ) be a sequence of functions defined
on 𝐴. That is, for each 𝑛 ∈ N we have an associated function 𝑓𝑛 : 𝐴 → R. We say that (𝑓𝑛 )
converges pointwise to a function 𝑓 : 𝐴 → R (as 𝑛 → ∞) if

lim 𝑓𝑛 (𝑥) = 𝑓 (𝑥)

𝑛→∞

for every fixed 𝑥 ∈ 𝐴. Here, as 𝑥 is fixed, the limit above is understood as the limit of a sequence
of real numbers.

For example, let us consider the sequence of functions on [0, 1] defined by

𝑓𝑛 (𝑥) := 𝑥 𝑛 .

At 𝑥 = 0, one has 𝑓𝑛 (𝑥) = 𝑓𝑛 (0) = 0 for each 𝑛 ∈ N. Therefore, 𝑓𝑛 (0) is the constant sequence
𝑓𝑛 (0) = 0 and so lim𝑛→∞ 𝑓𝑛 (0) = 0. Similarly, one has lim𝑛→∞ 𝑓𝑛 (1) = lim𝑛→∞ 1 = 1. Now, if
0 < 𝑥 < 1, we have seen in analysis 1 that 𝑥 𝑛 → 0 as 𝑛 → ∞. Therefore,

lim 𝑓𝑛 (𝑥) = lim 𝑥 𝑛 = 0.

𝑛→∞ 𝑛→∞

Combining all possible cases for 𝑥, we see that

(
0 if 0 ≤ 𝑥 < 1,
lim 𝑓𝑛 (𝑥) = 𝑓 (𝑥) :=
𝑛→∞ 1 if 𝑥 = 1.

This means that 𝑓𝑛 → 𝑓 pointwise on [0, 1].

Definition 4.2. Let 𝐴 ⊆ R and (𝑓𝑛 ) a sequence of functions defined on 𝐴. We say that 𝑓𝑛 con-
verges uniformly to a function 𝑓 : 𝐴 → R (as 𝑛 → ∞) if for each 𝜀 > 0, there exists 𝑁 ∈ N such
that for all 𝑥 ∈ 𝐴 and all 𝑛 ≥ 𝑁 there holds

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| < 𝜀.

27 Notice that this is similar to the definition employed for sequences of real numbers. Indeed, a sequence
of real numbers (𝑥𝑛 ) is simply a function N → R where we write 𝑥𝑛 := 𝑥 (𝑛).

72
Note that here the natural number 𝑁 is not allowed to depend on the point 𝑥. Equivalently, we
say that 𝑓𝑛 → 𝑓 uniformly on 𝐴 provided, for each 𝜀 > 0, one can find 𝑁 ∈ N such that

sup |𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ 𝜀

𝑥 ∈𝐴

for all 𝑛 ≥ 𝑁 .

Example 4.1. Define on R the sequence of functions

1
√︂
𝑓𝑛 (𝑥) := 𝑥2 + .
𝑛2
Clearly, each 𝑓𝑛 is continuous (even differentiable) on all of R. Now, for each fixed 𝑥 ∈ R, the
standard limit laws show that

1 √
√︂
lim 𝑓𝑛 (𝑥) = lim 𝑥 2 + 2 = 𝑥 2 = |𝑥 | .
𝑛→∞ 𝑛→∞ 𝑛
Letting 𝑓 (𝑥) := |𝑥 |, we see that 𝑓𝑛 → 𝑓 pointwise on R. We now inspect the possible uniform
convergence of this sequence. First, we observe that
√︂
√ 1 2 |𝑥 | 1
√︂
|𝑥 | = 𝑥 2 ≤ 𝑓𝑛 (𝑥) = 𝑥 2 + 2 ≤ |𝑥 | 2 + + 2
𝑛 𝑛 𝑛
√︄  2
1
= |𝑥 | +
𝑛
1
= |𝑥 | + .
𝑛
In particular, for each 𝑥 ∈ R and all 𝑛 ∈ N,
1 1
|𝑥 | − ≤ 𝑓𝑛 (𝑥) ≤ |𝑥 | +
𝑛 𝑛
whence
1
|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ , ∀𝑥 ∈ R and ∀𝑛 ∈ N.
𝑛
1
Given 𝜀 > 0, we can find 𝑁 ∈ N such that 𝑁 < 𝜀. Then, for all 𝑛 ≥ 𝑁 and all 𝑥 ∈ R,

1 1
|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ ≤ < 𝜀.
𝑛 𝑁
This shows that 𝑓𝑛 → 𝑓 uniformly on R.

73
4.1 Technical Examples of Non-Uniform Convergence
Despite the simplicity the last example might suggest, you will frequently encounter sequences
of functions that converge pointwise but not uniformly. Nevertheless, establishing non-uniform
convergence can be just as tricky. For this reason, our next several examples will include cases in
which the sequence will not converge uniformly.

Example 4.2. Consider the sequence of functions 𝑓𝑛 defined by

𝑛𝑥
𝑓𝑛 (𝑥) :=
1 + 𝑥 2𝑛 2
on R. Clearly, for every fixed 𝑥 ≠ 0, we have

𝑛 |𝑥 | 𝑛 |𝑥 | 1 1
0 ≤ |𝑓𝑛 (𝑥)| = ≤ = ·
1 + 𝑥 2𝑛 2 𝑥 2𝑛 2 |𝑥 | 𝑛

where 𝑛1 → 0 as 𝑛 → ∞. By the Squeeze Theorem for sequences, we see that 𝑓𝑛 (𝑥) → 0 as

𝑛 → ∞. This is simply the statement that 𝑓𝑛 → 0 pointwise on R \ {0}. Since 𝑓𝑛 (0) = 0 for all 𝑛
and 𝑓𝑛 (0) → 0 as 𝑛 → ∞, we conclude that 𝑓𝑛 → 0 pointwise on all of R.
Let 𝑎 > 0 be given; we claim that 𝑓𝑛 → 0 uniformly on [𝑎, ∞). Indeed, for all 𝑥 ∈ [𝑎, ∞) and
every 𝑛 ≥ 1 one has
𝑛𝑥 𝑛𝑥 1 1
|𝑓𝑛 (𝑥) − 0| = 𝑓𝑛 (𝑥) = 2 2
≤ 2 2 = ≤ .
1+𝑥 𝑛 𝑛𝑥 𝑛𝑥 𝑛𝑎
1
By the Archimedean property, there exists 𝑁 ∈ N such that 𝑁𝑎 < 𝜀, where 𝜀 > 0 is arbitrary but
fixed. Then, for all 𝑛 ≥ 𝑁 and all 𝑥 ∈ [𝑎, ∞),
1 1
|𝑓𝑛 (𝑥) − 0| ≤ ≤ < 𝜀.
𝑛𝑎 𝑁𝑎
It follows that 𝑓𝑛 → 0 uniformly on [𝑎, ∞) for all 𝑎 > 0. Nonetheless, 𝑓𝑛 does not converge
uniformly to 0 on [0, ∞). To see this, assume by contradiction that 𝑓𝑛 → 0 uniformly on [0, ∞).
By definition, one could find 𝑁 ∈ N such that
𝑛𝑥 1
|𝑓𝑛 (𝑥) − 0| = 2 2
<
1+𝑥 𝑛 2
for all 𝑛 ≥ 𝑁 and every 𝑥 ∈ [0, ∞). In particular, for each 𝑛 ≥ 𝑁 ,

1 1 1
 
𝑛(1/𝑛)
= 2 2
= 𝑓𝑛 −0 <
2 1 + (1/𝑛) 𝑛 𝑛 2

which is a contradiction. This shows that 𝑓𝑛 does not converge to 0 uniformly on [0, ∞).

74
Example 4.3. Consider once more the sequence of functions (𝑓𝑛 ) given in Example 4.1. We note
that each
1
√︂
𝑓𝑛 (𝑥) = 𝑥 2 + 2
𝑛
converged uniformly to 𝑓 (𝑥) = |𝑥 | on R. By the chain rule and quotient rule, each 𝑓𝑛 is continu-
ously differentiable with derivative given by
𝑥
𝑓𝑛′ (𝑥) = √︃ .
1
𝑥2 + 𝑛2

Clearly,


 −1 if 𝑥 < 0
𝑓𝑛 (𝑥) → 𝑔(𝑥) := 0
′


if 𝑥 = 0,

1

if 𝑥 > 0.

Despite uniform convergence of (𝑓𝑛 ) and the pointwise convergence of (𝑓𝑛′ ), the sequence of
derivatives (𝑓𝑛′ ) does not converge uniformly on any interval of the form [−𝑎.𝑎]! Indeed, pro-
ceeding by way of contradiction, let us assume that 𝑓𝑛′ → 𝑔 uniformly on [−𝑎, 𝑎]. Then, taking
𝜀 := 12 we can find 𝑁 ∈ N such that

1
𝑓𝑛′ (𝑥) − 𝑔(𝑥) < , ∀𝑛 ≥ 𝑁 ,
2
and all 𝑥 ∈ [−𝑎, 𝑎]. In particular, for every 𝑥 ∈ (0, 𝑎],
1
𝑓𝑁′ (𝑥) − 𝑔(𝑥) < .
2
However, this implies that

𝑥 1
−1 < , ∀𝑥 ∈ (0, 𝑎].
2
√︃
1
𝑥2 + 𝑁2

Consider now the sequence (𝑥𝑘 ) given by 𝑥𝑘 := 𝑘1 . Since 𝑥𝑘 → 0 as 𝑘 → ∞, there exists 𝐾 ∈ N

such that 𝑥𝑘 ∈ (0, 𝑎] ⊆ [−𝑎, 𝑎] for all 𝑘 ≥ 𝐾.28 Therefore,

𝑥𝑘 1
− 1 = 𝑓𝑁′ (𝑥𝑘 ) − 𝑔(𝑥𝑘 ) <
2
√︃
1
𝑥𝑘2 + 𝑁2

28 Bythe 𝜀 − 𝑁 definition of convergence for sequences, there exists 𝐾 ∈ N such that |𝑥𝑘 | < 𝑎 for all
𝑘 ≥ 𝐾. Thus, 𝑥𝑘 ∈ (0, 𝑎) for all 𝑘 ≥ 𝐾.

75
for all 𝑘 ≥ 𝐾. Taking the limit as 𝑘 → ∞, we find that

𝑥𝑘 𝑥𝑘 1
|0 − 1| = lim √︃ − 1 = lim √︃ −1 ≤
1 1 2
𝑥𝑘2 + 𝑥𝑘2 +
𝑘→∞ 𝑘→∞
𝑁2 𝑁2

which is a contradiction. Therefore, 𝑓𝑛′ does not converge to 𝑔 uniformly.

4.2 A Squeeze type Theorem for Uniform Convergence

Recall that a given sequence of functions 𝑓𝑛 : 𝐴 → R converges uniformly to a function 𝑓 : 𝐴 →
R, as 𝑛 → ∞, if for each 𝜀 > 0 there exists 𝑁 ∈ N such that

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| < 𝜀

for all 𝑥 ∈ 𝐴 and all 𝑛 ≥ 𝑁 . In practice, to show that (𝑓𝑛 ) converges uniformly to 𝑓 on 𝐴, we try
to find a sequence (𝑎𝑛 ) converging to 0 such that

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ 𝑎𝑛

for all 𝑥 ∈ 𝐴 and all 𝑛 ∈ N. This is sufficient because, for each 𝜀 > 0, one can find 𝑁 ∈ N such
that 𝑎𝑛 < 𝜀 for all 𝑛 ≥ 𝑁 . Consequently, for all 𝑛 ≥ 𝑁 and all 𝑥 ∈ 𝐴 there holds

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ 𝑎𝑛 < 𝜀

whence 𝑓𝑛 → 𝑓 uniformly on 𝐴. We restate this below in the form of a lemma:

Lemma 4.1. Let 𝐴 ⊆ R and (𝑓𝑛 ) be a sequence of functions on 𝐴 and fix a function 𝑓 : 𝐴 → R.
Assume that there exists a sequence (𝑎𝑛 ) such that

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| ≤ 𝑎𝑛

for all 𝑛 ∈ N and every 𝑥 ∈ 𝐴. If 𝑎𝑛 → 0 as 𝑛 → ∞, then 𝑓𝑛 → 𝑓 uniformly as 𝑛 → ∞.

Example 4.4. Consider the sequence of functions (𝑓𝑛 ) on R defined by

𝑥
𝑓𝑛 (𝑥) := .
1 + 𝑛𝑥 2
To identify the uniform limit of these 𝑓𝑛 ’s (if it exists), we should first try and determine their
pointwise limit. Clearly, for every fixed 𝑥 ≠ 0 we have

|𝑥 | |𝑥 | 1 𝑛→∞
|𝑓𝑛 (𝑥)| ≤ 2
≤ 2 = −−−−→ 0.
1 + 𝑛𝑥 𝑛𝑥 𝑛 |𝑥 |

76
Thus, if 𝑥 ≠ 0 is fixed, then the Squeeze Theorem for sequences implies that
lim 𝑓𝑛 (𝑥) = 0.
𝑛→∞

Moreover,
lim 𝑓𝑛 (0) = lim 0 = 0.
𝑛→∞ 𝑛→∞
We infer that 𝑓𝑛 (𝑥) → 0, as 𝑛 → ∞, for each 𝑥 ∈ R. This is precisely the statement that 𝑓𝑛 → 0
pointwise on R. We now assert that 𝑓𝑛 → 0 uniformly on R. For this, we will require the following
identity:
|𝑡 | 1
2
≤ , ∀𝑡 ∈ R. (4.1)
1+𝑡 2
To prove (4.1), it suffices to check that
𝑡 2 − 2 |𝑡 | + 1 ≥ 0, ∀𝑡 ∈ R.
However, 𝑡 2 − 2 |𝑡 | + 1 = (|𝑡 | − 1) 2 ≥ 0 for all 𝑡 ∈ R. This verifies (4.1). Returning to the example,
√
we can apply (4.1) with 𝑡 = 𝑛𝑥 to obtain the following uniform estimate:
√
|𝑥 | 1 𝑛 |𝑥 |
|𝑓𝑛 (𝑥) − 0| = 2
=√ ·
1 + 𝑛𝑥 𝑛 1 + 𝑛𝑥 2
√
1 𝑛𝑥
=√ · √
2
𝑛 1 + 𝑛𝑥
(4.1) 1
≤ √
2 𝑛
1
for all 𝑥 ∈ R and every 𝑛 ∈ N. Applying Lemma 4.1 with 𝑎𝑛 := √
2 𝑛
then shows that 𝑓𝑛 → 0
uniformly.
Remark 4.1. We have seen in the lectures and tutorials that uniform convergence does not imply
the uniform convergence of the derivatives. As is turns out, counter examples to such a statement
are typically neither rare nor contrived. In fact, the example above is yet another instance of such
a sequence. Indeed, by the chain rule and quotient rule, each 𝑓𝑛 is continuously differentiable
with derivative given by
(1 + 𝑛𝑥 2 ) − 2𝑛𝑥 2 1 − 𝑛𝑥 2
𝑓𝑛′ (𝑥) = 2 2
= .
(1 + 𝑛𝑥 ) (1 + 𝑛𝑥 2 ) 2
Clearly,
lim 𝑓 ′ (0) = lim 1 = 1.
𝑛→∞ 𝑛 𝑛→∞
On the other hand, if 𝑥 ≠ 0 is fixed,
1 − 𝑛𝑥 2 1 + 𝑛𝑥 2 1 1
2 2
≤ 2 2
= 2
≤ 2
(1 + 𝑛𝑥 ) (1 + 𝑛𝑥 ) 1 + 𝑛𝑥 𝑛𝑥

77
which tends to 0 as 𝑛 → ∞. Hence, 𝑓𝑛′ (𝑥) → 0, as 𝑛 → ∞, for all 𝑥 ≠ 0. Put otherwise,
(
1 if 𝑥 = 0,
𝑓𝑛′ (𝑥) → 𝑔(𝑥) :=
0 if 𝑥 ≠ 0.

However, defining 𝑓 (𝑥) := 0 on R, we have shown above that 𝑓𝑛 → 𝑓 uniformly. Despite this,
𝑓𝑛′ (0) ̸→ 𝑓 ′ (0). Thus, 𝑓𝑛′ ̸→ 𝑓 ′ pointwise on R even though 𝑓𝑛′ does converge pointwise. Further-
more, (𝑓𝑛′ ) does not converge uniformly to 𝑔 on R since 𝑔 is discontinuous at 0.
It should now be clear that uniform convergence does not imply the uniform convergence of
the derivatives. In fact, if (𝑓𝑛 ) is a sequence of differentiable functions converging uniformly to
a differentiable function 𝑓 , the sequence (𝑓𝑛′ ) may not even converge pointwise to 𝑓 ′ . Therefore,
deducing properties about the convergence of (𝑓𝑛′ ) using the uniform convergence of (𝑓𝑛 ) can in
general be quite tricky.

4.3 Continuity of the Uniform Limit

We also recall the following theorem. In general, this result makes it much easier to show that
certain pointwise convergences cannot be uniform.

Theorem 4.2. Let 𝐴 ⊆ R and let (𝑓𝑛 ) be a sequence of functions defined on 𝐴 converging uniformly
to a function 𝑓 : 𝐴 → R on the set 𝐴. If each 𝑓𝑛 is continuous at a point 𝑐 ∈ 𝐴, then so is 𝑓 .

Proof. Let 𝜀 > 0. Since 𝑓𝑛 → 𝑓 uniformly on 𝐴, there exists 𝑁 ∈ N such that

𝜀
|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| <
3
for all 𝑥 ∈ 𝐴 and every 𝑛 ≥ 𝑁 . Now, 𝑓𝑁 is continuous at the point 𝑐 ∈ 𝐴. Thus, there exists 𝛿 > 0
such that
𝜀
|𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑐)| <
3
for all 𝑥 ∈ 𝐴 with |𝑥 − 𝑐 | < 𝛿. Finally, note that the triangle inequality implies

|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)|
2𝜀
< + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑐)| .
3
If in addition |𝑥 − 𝑐 | < 𝛿, then
2𝜀
|𝑓 (𝑥) − 𝑓 (𝑐)| < + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑐)| < 𝜀.
3
This completes the proof. □

78
Corollary 4.3. Let 𝐴 ⊆ R and let (𝑓𝑛 ) be a sequence of functions defined on 𝐴 converging uniformly
to a function 𝑓 : 𝐴 → R on the set 𝐴. If each 𝑓𝑛 is continuous on 𝐴, then so is 𝑓 .

Equipped with this result, we provide an alternative approach to Example 4.3. Recall Example
4.1 where we showed that the sequence

1
√︂
𝑓𝑛 (𝑥) = 𝑥 2 + 2
𝑛
converges uniformly to 𝑓 (𝑥) = |𝑥 | on R. Then, in Example 4.3 we showed however that (𝑓𝑛′ ) does
not converge uniformly.

Example 4.5. By the chain rule, every function 𝑓𝑛 defined above is differentiable on R with
derivative given by
𝑥
𝑓𝑛′ (𝑥) = √︃ .
𝑥 2 + 𝑛12

However, the uniform limit of these 𝑓𝑛 ’s is not differentiable at 0. Indeed, the function 𝑓 (𝑥) = |𝑥 |
does not have a derivative at 𝑥 = 0. Despite this, the functions 𝑓𝑛′ still converge pointwise on R.
More precisely,


 0 if 𝑥 = 0,
𝑥
lim 𝑓 (𝑥) = lim √︃
′


= 1 if 𝑥 > 0,
𝑛→∞ 𝑛
𝑥 2 + 𝑛12 
𝑛→∞ 
 −1 if 𝑥 < 0.

But, is it true that 𝑓𝑛′ converges uniformly to this function described above? To answer this, denote
by 𝑔 the pointwise limit of (𝑓𝑛′ ), i.e.



 0 if 𝑥 = 0,
𝑔(𝑥) := 1


if 𝑥 > 0,

 −1 if 𝑥 < 0.


We claim that 𝑓𝑛′ does not converge uniformly to 𝑔 on any compact interval [−𝑎, 𝑎]. To see this,
we argue by way of contradiction. Assume that 𝑓𝑛′ → 𝑔 uniformly on [−𝑎, 𝑎]. Since every 𝑓𝑛′ is
continuous on [−𝑎, 𝑎], the previous Corollary implies that the uniform limit 𝑔 must be continuous
on [−𝑎, 𝑎] as well. However, 𝑔 is clearly discontinuous at 0. Hence, the convergence cannot be
uniform.

We have proven that uniform limits of continuous functions are continuous. More precisely,
we proved in Theorem 4.2 that if 𝑓𝑛 → 𝑓 uniformly and every 𝑓𝑛 is continuous at a point 𝑐, then
so is the limit 𝑓 . It is therefore reasonable to ask whether uniform limits also preserve points of
discontinuitiy. This answer is negative and this is justified by the following example:

79
Example 4.6. For each 𝑛 ∈ N consider the function
1
(
if 𝑥 ∈ Q,
𝑓𝑛 : R → R, 𝑥 ↦→ 𝑛
0 if 𝑥 ∉ Q.

We claim that every 𝑓𝑛 is discontinuous on all of R. Indeed, if 𝑓𝑛 were continuous at a point 𝑐 ∈ R,

then so would be the function 𝑛𝑓𝑛 . However, 𝑛𝑓𝑛 is precisely the Dirichlet function 1Q : R → R,
which is everywhere discontinuous. Hence, 𝑓𝑛 is discontinuous at every point 𝑐 ∈ R.
Despite this, 𝑓𝑛 → 0 uniformly on R, where the constant function 𝑥 ↦→ 0 is everywhere
continuous (even infinitely differentiable). To see this, note that for each 𝑛 ∈ N one has
1
|𝑓𝑛 (𝑥)| = 𝑓𝑛 (𝑥) ≤
𝑛
for all 𝑥 ∈ R. By Lemma 4.1, it follows that 𝑓𝑛 → 0 uniformly on R.

4.4 Uniform Limits and Bounded functions

At this point it is quite convincing that uniform limits enjoy nicer properties than simple point-
wise limits. This section explores more of these nice properties, the first of which is an elegant
boundedness result.

Proposition 4.4. Let 𝐴 ⊆ R and (𝑓𝑛 ) be a sequence of functions converging uniformly to a function
𝑓 : 𝐴 → R on the set 𝐴. If each 𝑓𝑛 is bounded, then so is 𝑓 .29 That is, uniform limits of bounded
functions are bounded.

Proof. Let 𝜀 = 1. Since 𝑓𝑛 → 𝑓 uniformly on 𝐴, there exists 𝑁 ∈ N such that

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| < 1

for all 𝑛 ≥ 𝑁 and every 𝑥 ∈ 𝐴. Since 𝑓𝑁 is bounded, one has

|𝑓 (𝑥)| = |𝑓 (𝑥) − 𝑓𝑁 (𝑥) + 𝑓𝑁 (𝑥)| ≤ |𝑓 (𝑥) − 𝑓𝑁 (𝑥)| + |𝑓𝑁 (𝑥)|

< 1 + 𝑀𝑁

for any 𝑥 ∈ 𝐴. This is precisely the statement that 𝑓 is bounded on 𝐴. □

We also offer a slight improvement of our previous statement:

29 Here we assume that, for each 𝑛 ∈ N, there exists 𝑀 > 0 such that |𝑓 (𝑥)| ≤ 𝑀 for all 𝑥 ∈ 𝐴. It is
𝑛 𝑛 𝑛
allowed that these 𝑀𝑛 depend on the index 𝑛 of the sequence!

80
Corollary 4.5. Let (𝑓𝑛 ) be a sequence of functions, each defined on a set 𝐴 ⊆ R, converging uni-
formly to a function 𝑓 : 𝐴 → R on 𝐴. If each 𝑓𝑛 is bounded, then the sequence (𝑓𝑛 ) is uniformly
bounded. More precisely, if there exists for each 𝑛 ∈ N some 𝑀𝑛 > 0 such that

|𝑓𝑛 (𝑥)| ≤ 𝑀𝑛 , ∀𝑥 ∈ 𝐴,

then there exists 𝑀 > 0 such that

|𝑓𝑛 (𝑥)| ≤ 𝑀
for all 𝑥 ∈ 𝐴 and all 𝑛 ∈ N. Moreover,
|𝑓 (𝑥)| ≤ 𝑀
for every 𝑥 ∈ 𝐴.

Proof. Citing the previous proposition, the uniform limit 𝑓 is bounded on 𝐴. Thus, there exists a
constant 𝐶 > 0 such that |𝑓 (𝑥)| ≤ 𝐶 for all 𝑥 ∈ 𝐴. Now, as 𝑓𝑛 → 𝑓 uniformly on 𝐴, there exists
𝑁 ∈ N such that
|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| < 1
for all 𝑛 ≥ 𝑁 and each 𝑥 ∈ 𝐴. In particular, we have

|𝑓𝑛 (𝑥)| ≤ |𝑓𝑛 (𝑥) − 𝑓 (𝑥)| + |𝑓 (𝑥)| < 1 + 𝐶

for every 𝑛 ≥ 𝑁 and any 𝑥 ∈ 𝐴. Since |𝑓𝑛 | ≤ 𝑀𝑛 on 𝐴 for every 𝑛 ∈ N, it follows from the above
that, given any 𝑛 ∈ N,

|𝑓𝑛 (𝑥)| ≤ max {𝑀1, 𝑀2, . . . , 𝑀𝑁 −1, 1 + 𝐶} =: 𝑀.

for every point 𝑥 ∈ 𝐴. It follows that |𝑓𝑛 (𝑥)| ≤ 𝑀 for every 𝑥 ∈ 𝐴 and all 𝑛 ∈ N. Since
|𝑓 (𝑥)| ≤ 𝐶 < 1 + 𝐶 ≤ 𝑀 on 𝐴, we are done. □

Example 4.7. Consider the sequence of functions (𝑓𝑛 ) given by

𝑛 1
𝑓𝑛 : (0, ∞) → R, 𝑓𝑛 (𝑥) = = .
𝑛𝑥 + 1 𝑥 + 𝑛1

Observe that this is a sequence of bounded functions. Indeed, each function 𝑓𝑛 is bounded since
𝑛 𝑛
0 ≤ 𝑓𝑛 (𝑥) = ≤ = 𝑛.
𝑛𝑥 + 1 1
for any 𝑥 > 0. Moreover, 𝑓𝑛 → 𝑓 pointwise where 𝑓 : (0, ∞) → R is the function 𝑓 (𝑥) = 1/𝑥.
However, 𝑓𝑛 does not converge to 𝑓 uniformly on (0, ∞). Indeed, if this were the case then 𝑓
would be bounded. However, 𝑓 (𝑥) = 1/𝑥 is an unbounded function.

By a similar argument to the proof of Proposition 4.4, we also obtain the following:

81
Proposition 4.6. Let 𝐴 ⊆ R and (𝑓𝑛 ) be a sequence of functions converging uniformly to a function
𝑓 : 𝐴 → R on the set 𝐴. If every 𝑓𝑛 is unbounded, then 𝑓 is also unbounded.

Proof. We argue by contradiction. Assuming that 𝑓 is bounded on 𝐴, there exists a constant 𝑀 > 0
such that |𝑓 (𝑥)| ≤ 𝑀 for all 𝑥 ∈ 𝐴. For 𝜀 := 1, we can also find 𝑁 ∈ N such that

|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| < 1

for all 𝑛 ≥ 𝑁 and every 𝑥 ∈ 𝐴. In particular,

|𝑓𝑁 (𝑥)| = |𝑓𝑁 (𝑥) − 𝑓 (𝑥) + 𝑓 (𝑥)| ≤ |𝑓𝑁 (𝑥) − 𝑓 (𝑥)| + |𝑓 (𝑥)| < 1 + 𝑀

for every 𝑥 ∈ 𝐴. This implies that 𝑓𝑁 is bounded on 𝐴, which is a contradiction. □

4.5 More on Uniform Convergence

We recall from class that uniform limits preserve Riemann integrability:

Theorem 4.7. Let (𝑓𝑛 ) be a sequence of Riemann integrable functions on [𝑎, 𝑏] converging uniformly
to a function 𝑓 : [𝑎, 𝑏] → R on [𝑎, 𝑏]. Then, 𝑓 is Riemann integrable on [𝑎, 𝑏] and
ˆ 𝑏 ˆ 𝑏
lim 𝑓𝑛 = 𝑓.
𝑛→∞ 𝑎 𝑎

In fact, one can say slightly more:

Corollary 4.8. If (𝑓𝑛 ) is a sequence of Riemann integrable functions on [𝑎, 𝑏] converging uniformly
to a function 𝑓 : [𝑎, 𝑏] → R on [𝑎, 𝑏], then 𝑓 is Riemann integrable on [𝑎, 𝑏] and
ˆ 𝑏
lim |𝑓𝑛 − 𝑓 | = 0.
𝑛→∞ 𝑎

Proof. Since 𝑓𝑛 is Riemann integrable for every 𝑛 ∈ N and 𝑓 ∈ R( [𝑎, 𝑏]) by the previous theorem,
(𝑓𝑛 − 𝑓 ) is a sequence of Riemann integrable functions on [𝑎, 𝑏]. By Proposition ??, |𝑓𝑛 − 𝑓 | is also
Riemann integrable on [𝑎, 𝑏]. Because 𝑓𝑛 → 𝑓 uniformly if and only if |𝑓𝑛 − 𝑓 | → 0 uniformly,
another application of the previous theorem ensures that
ˆ 𝑏 ˆ 𝑏
lim |𝑓𝑛 − 𝑓 | = 0 = 0.
𝑛→∞ 𝑎 𝑎

We have seen that uniform convergence preserves many properties of functions. But what
operations preserve uniform convergence? One example, is composition through continuous
functions. More precisely, we have the following result:

82
Proposition 4.9 (Composition Theorem for Uniform Convergence). Let 𝐴 ⊆ R and let (𝑓𝑛 ) be a
sequence of functions converging uniformly to a function 𝑓 on 𝐴. Assume that every 𝑓𝑛 is bounded
on 𝐴. If 𝑔 : R → R is continuous, then

𝑔 ◦ 𝑓𝑛 → 𝑔 ◦ 𝑓

uniformly on 𝐴.

Proof. By virtue of Corollary 4.5, there exists 𝑀 > 0 such that |𝑓𝑛 (𝑥)| ≤ 𝑀 and |𝑓 (𝑥)| ≤ 𝑀 for
every 𝑛 ∈ N and each 𝑥 ∈ 𝐴. Let 𝜀 > 0 be given. Since [−𝑀, 𝑀] is compact, 𝑔 is uniformly
continuous on [−𝑀, 𝑀]. Thus, there exists 𝛿 > 0 such that

|𝑔(𝑦) − 𝑔(𝑣)| < 𝜀 (4.2)

for all 𝑦, 𝑣 ∈ [−𝑀, 𝑀] with |𝑦 − 𝑣 | < 𝛿. On the other hand, because 𝑓𝑛 → 𝑓 uniformly on 𝐴, there
exists 𝑁 ∈ N such that
|𝑓𝑛 (𝑥) − 𝑓 (𝑥)| < 𝛿
for all 𝑛 ≥ 𝑁 . For any such 𝑛 and all 𝑥 ∈ 𝐴, we see from (4.2) (with 𝑦 := 𝑓𝑛 (𝑥) and 𝑣 := 𝑓 (𝑥)) that

|(𝑔 ◦ 𝑓𝑛 ) (𝑥) − (𝑔 ◦ 𝑓 ) (𝑥)| = |𝑔(𝑓𝑛 (𝑥)) − 𝑔(𝑓 (𝑥))| < 𝜀.

Since this holds for all 𝑛 ≥ 𝑁 and every 𝑥 ∈ 𝐴, it follows that 𝑔 ◦ 𝑓𝑛 → 𝑔 ◦ 𝑓 uniformly on 𝐴. □

Remark 4.2. Note that the result above continues to hold if 𝑔 is merely assumed to be continuous
on the interval [−𝑀, 𝑀].

4.6 Uniform Limits of Uniformly Continuous Functions

Having shown that uniform limits preserve certain local properties (e.g. continuity), we now
ask which global properties survive limits. Having already seen that boundedness is preserved
by uniform limits, we now check that uniform limits of uniformly continuous functions are still
uniformly continuous.

Proposition 4.10. Let 𝐴 ⊆ R and 𝑓 : 𝐴 → R be a function. Let (𝑓𝑛 ) be a sequence of uniformly

continuous functions defined on 𝐴 and assume that 𝑓𝑛 → 𝑓 uniformly on 𝐴. Then, 𝑓 is uniformly
continuous on 𝐴.

Proof. We can use the same proof as in Theorem 4.2. Given 𝜀 > 0, there exists 𝑁 ∈ N such that,
for all 𝑛 ≥ 𝑁 and each 𝑡 ∈ 𝐴,
𝜀
|𝑓𝑛 (𝑡) − 𝑓 (𝑡)| < .
3

83
For any 𝑥, 𝑢 ∈ 𝐴 there holds

|𝑓 (𝑥) − 𝑓 (𝑢)| = |𝑓 (𝑥) − 𝑓𝑁 (𝑥) + 𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑢) + 𝑓𝑁 (𝑢) − 𝑓 (𝑢)|

≤ |𝑓 (𝑥) − 𝑓𝑁 (𝑥)| + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑢)| + |𝑓𝑁 (𝑢) − 𝑓 (𝑢)|
2𝜀
≤ + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑢)| .
3
Now, 𝑓𝑁 is uniformly continuous on 𝐴. Thus, there exists 𝛿 > 0 such that
𝜀
|𝑓 (𝑥) − 𝑓 (𝑢)| <
3
whenever |𝑥 − 𝑢 | < 𝛿. So, if |𝑥 − 𝑢 | < 𝛿, our calculations above indicate that
2𝜀
|𝑓 (𝑥) − 𝑓 (𝑢)| ≤ + |𝑓𝑁 (𝑥) − 𝑓𝑁 (𝑢)| < 𝜀
3
whence 𝑓 is uniformly continuous on 𝐴. □

84
5 Infinite Series
Since we will spend a significant portion of this tutorial reviewing the midterm exam solutions,
these tutorial notes will be significantly shorter than usual. However, the midterm solutions
will be posted (shortly) within a self contained document on MyCourses. These notes shall only
contain the tutorial content related to the convergence of series.

Definition 5.1. Let (𝑎𝑛 ) be a sequence of real numbers and define for each 𝑁 ∈ N the partial
sum
𝑁
∑︁
𝑆 𝑁 := 𝑎𝑛
𝑛=1

is a well defined real number. Furthermore, the partial sums themselves form a sequence (𝑆 𝑁 )𝑁 ∈N .
With this in mind, we say that the series 𝑛=1 𝑎 𝑁 converges provided the sequence of partial sums
Í∞
(𝑆 𝑁 )𝑁 ∈N is convergent in R. In this case, we define
∞
∑︁
𝑎𝑛 := lim 𝑆 𝑁 .
𝑁 →∞
𝑛=1

Although we do not yet possess many tools related to the convergence of series, we have
already seen a few useful results in class, the first of which is the straightforward divergence test:

Theorem 5.1 (Divergence Test). Let (𝑎𝑛 ) be a sequence of real numbers. If 𝑛=1
Í∞
𝑎𝑛 converges, then
we must have
lim 𝑎𝑛 = 0.
𝑛→∞

Equivalently, if lim𝑛→∞ 𝑎𝑛 ≠ 0, then the series 𝑛=1

Í∞
𝑎𝑛 is divergent.

Using this as our only series-test, it is already possible for us to prove that the series
∞
∑︁
sin(𝑛)
𝑛=1

is divergent. We justify this in the following example:

Example 5.1. Using the divergence test stated above, we shall prove that the series 𝑛=1 sin 𝑛 is
Í∞
divergent. Thus, we are reduced to proving that

lim sin 𝑛 ≠ 0,
𝑛→∞

which requires some work. Recall the sin addition formula which states that

sin(𝑥 + 𝑦) = sin(𝑥) cos(𝑦) + sin(𝑦) cos(𝑥), ∀𝑥, 𝑦 ∈ R. (5.1)

85
In particular, since sin(−𝑦) = − sin(𝑦), we obtain

sin(𝑥 − 𝑦) = sin(𝑥) cos(𝑦) − sin(𝑦) cos(𝑥), ∀𝑥, 𝑦 ∈ R. (5.2)

By way of contradiction, let us assume that

lim sin 𝑛 = 0.
𝑛→∞

Then,
lim sin(𝑛 + 1) = lim sin 𝑛 = lim sin(𝑛 − 1) = 0.
𝑛→∞ 𝑛→∞ 𝑛→∞
Now, observe that by combining (5.1)-(5.2) we obtain the following expression:

sin(𝑛 + 1) − sin(𝑛 − 1) = [sin(𝑛) cos(1) + sin(1) cos(𝑛)] − [sin(𝑛) cos(1) − sin(1) cos(𝑛)]
= 2 sin(1) cos(𝑛).

But, by limit laws, this implies that

1
lim cos(𝑛) = lim lim (sin(𝑛 + 1) − sin(𝑛 − 1))
𝑛→∞ 𝑛→∞ 2 sin(1) 𝑛→∞
1  
= lim sin(𝑛 + 1) − lim sin(𝑛 − 1)
2 sin(1) 𝑛→∞ 𝑛→∞
1
= (0 − 0)
2 sin(1)
= 0.

Consequently, sin(𝑛) → 0 and cos(𝑛) → 0 as 𝑛 → ∞. However, this implies that sin2 (𝑛), cos2 (𝑛) →
0 as 𝑛 → ∞, which is a contradiction because

1 = sin2 (𝑛) + cos2 (𝑛), ∀𝑛 ∈ N.

It follows that sin(𝑛) ̸→ 0 as 𝑛 → ∞ whence we see from the divergence test that 𝑛=1 sin(𝑛)
Í∞
diverges.

A second example involving convergence is also in order:

Example 5.2. Given 𝛼 > 0, we will show that the series

∞
∑︁ 1
𝑛=0
(𝛼 + 𝑛) (𝛼 + 𝑛 + 1)

converges and determine its value. In this case, the “trick” is to realize that the summand term
1 1 1
= −
(𝛼 + 𝑛) (𝛼 + 𝑛 + 1) 𝛼 + 𝑛 𝛼 + 𝑛 + 1

86
admits a convenient partial fractions decomposition. Consequently, the partial sums for the series
1
𝑛=1 (𝛼+𝑛) (𝛼+𝑛+1) can also be written in the following way:
Í∞

1 1 1
𝑁
∑︁ 𝑁 
∑︁ 
= −
𝑛=0
(𝛼 + 𝑛) (𝛼 + 𝑛 + 1) 𝑛=0 𝛼 + 𝑛 𝛼 + (𝑛 + 1)

1 1
𝑁
∑︁ 𝑁
∑︁
= −
𝑛=0
𝛼 + 𝑛 𝑛=0 𝛼 + (𝑛 + 1)
𝑁 +1
1 ∑︁ 1
𝑁
∑︁
= −
𝑛=0
𝛼 + 𝑛 𝑛=1 𝛼 + 𝑛
1 1
= − .
𝛼 𝛼 +𝑁 +1
Therefore,

1 1 1 1
𝑁
∑︁  
lim = lim − = .
𝑁 →∞
𝑛=0
(𝛼 + 𝑛) (𝛼 + 𝑛 + 1) 𝑁 →∞ 𝛼 𝛼 + 𝑁 + 1 𝛼

By definition, this means that

∞
∑︁ 1 1
= .
𝑛=0
(𝛼 + 𝑛) (𝛼 + 𝑛 + 1) 𝛼

5.1 The Cauchy Condensation Test & Examples

Before providing a few worked out examples, let us recall the precise statement of the Cauchy
condensation test:
Theorem 5.2 (Cauchy Condensation Test). Let (𝑎𝑛 ) be a decreasing sequence of non-negative real
𝑎𝑛 converges if and only if the condensed series 𝑛=1 𝑎 2𝑛 2𝑛 converges.
Í∞ Í∞
numbers. Then, the series 𝑛=1
Remark 5.1. Although this almost certainly has/will be addressed in class, I would like to point
out that the Cauchy condensation test provides a simple way to show that the series
∞
∑︁ 1
𝑛=1
𝑛𝑝

converges if and only if 𝑝 > 1.

Example 5.3. Using the Cauchy condensation test, we shall show that the series
∞
∑︁ 1
𝑛=2
𝑛 ln(𝑛)

87
diverges. First, note that if 𝑎𝑛 := 𝑛 ln1 𝑛 then (𝑎𝑛 ) is non-negative and decreasing (since both 𝑛 and
ln(𝑛) are non-negative and increasing with respect to 𝑛). Furthermore, the terms of our condensed
series are given by
1 1 1
𝑎 2𝑛 = = 𝑛 =⇒ 2𝑛 𝑎 2𝑛 = .
2𝑛 ln(2 ) 2 𝑛 ln(2)
𝑛 𝑛 ln(2)
1
Since the harmonic series diverges, it follows from the above that our condensed series
Í∞
𝑛=1 𝑛

∞
∑︁ 1
𝑛=1
𝑛 ln(2)

diverges and hence so does our original series.

Example 5.4. Let 𝑐 > 1 be given an consider the series

∞
∑︁ 1
.
𝑛=2
𝑛 ln(𝑛)𝑐

Using the Cauchy condensation test, we will show that this series is actually convergent. Indeed,
1
as above, if we set 𝑎𝑛 := 𝑛 ln(𝑛) 𝑐 we find that

1 1 1 1
2𝑛 𝑎 2𝑛 = 2𝑛 · = = · .
2𝑛 (ln(2𝑛 ))𝑐 (𝑛 ln(2))𝑐 ln(2)𝑐 𝑛𝑐
Now, since 𝑐 > 1 and the 𝑝-series
∞
∑︁ 1
𝑛=1
𝑛𝑝
converges if and only if 𝑝 > 1, it follows that the series
∞ ∞
∑︁ ∑︁ 1 1 ∑︁ 1
2𝑛 𝑎 2𝑛 = =
𝑛=2 𝑛=2
ln(2)𝑐 𝑛𝑐 ln(2)𝑐 𝑛=2 𝑛𝑐

converges. Hence, by the condensation test, our original series must also converge.

5.2 The Comparison & Limit-Comparison Tests

We devote this tutorial to the proofs of the limit comparison test and the ratio test (both the limit-
free and limit versions). Since the proofs of these respective results will fundamentally boil down
to a clever application of this comparison test, we will reiterate this result for completeness.

Theorem 5.3 (Comparison Test). Let 𝑁 ∈ N and suppose that (𝑎𝑛 ), (𝑏𝑛 ) are non-negative sequences
such that 𝑎𝑛 ≤ 𝑏𝑛 for all 𝑛 ≥ 𝑁 .

88
 Í Í
1. If 𝑏𝑛 converges then 𝑎𝑛 converges.
Í Í
2. If 𝑎𝑛 diverges then 𝑏𝑛 diverges.

Equipped with this, the following limit comparison test is within reach:

Proposition 5.4 (Limit Comparison Test). Let (𝑎𝑛 ) be a non-negative sequence and (𝑏𝑛 ) a strictly
positive sequence such that
𝑎𝑛
lim = 𝐿.
𝑛→∞ 𝑏𝑛

Then,

1. If 0 < 𝐿 < ∞ then 𝑎𝑛 converges if and only if 𝑏𝑛 converges.

Í Í

2. If 𝐿 = 0 then 𝑎𝑛 converges whenever 𝑏𝑛 converges.

Í Í

Í Í
3. If 𝐿 = ∞, then 𝑎𝑛 diverges whenever 𝑏𝑛 diverges.

Proof.

1. Choose 𝜀 > 0 such that 𝐿 − 𝜀 > 0 and observe that there exists 𝑁 ∈ N such that
𝑎𝑛
− 𝐿 < 𝜀, ∀𝑛 ≥ 𝑁 .
𝑏𝑛
Equivalently, one has
𝑎𝑛
𝐿 −𝜀 < < 𝐿 + 𝜀, ∀𝑛 ≥ 𝑁
𝑏𝑛
whence it follows that

(𝐿 − 𝜀)𝑏𝑛 < 𝑎𝑛 < (𝐿 + 𝜀)𝑏𝑛 , ∀𝑛 ≥ 𝑁 .

It therefore follows from the comparison test that 𝑎𝑛 converges if and only if 𝑏𝑛 con-
Í Í
verges.

2. In the event that 𝐿 = 0, for 𝜀 = 1, there exists 𝑁 ∈ N such that

𝑎𝑛
0≤ < 1, ∀𝑛 ≥ 𝑁 .
𝑏𝑛
Thus, 𝑎𝑛 < 𝑏𝑛 for all 𝑛 large whence by comparison we see that 𝑎𝑛 converges whenever
Í
𝑏𝑛 does.
Í

𝑎𝑛
3. Since lim = ∞, there exists 𝑁 ∈ N such that 𝑎𝑛 /𝑏𝑛 > 1 for all 𝑛 ≥ 𝑁 . By comparison,
𝑛→∞ 𝑏𝑛
the assertion follows.

89
 We now come to one of the most famous tests used in the analysis of series: the ratio test. We
begin with a statement/version of this test that does not involve limits. Although the statement
of the limit-free version is slightly more technical that its counterpart, it has the added benefit of
being stronger.

Theorem 5.5 (Limit-Free Ratio Test). Let (𝑎𝑛 ) be a sequence of positive numbers. Then

1. If there exists 𝑟 ∈ [0, 1), 𝑁 ∈ N such that 𝑎𝑎𝑛+1

Í
𝑛
≤ 𝑟 for all 𝑛 ≥ 𝑁 then 𝑎𝑛 converges.

2. If there exists 𝑁 ∈ N such that 𝑎𝑎𝑛+1 ≥ 1 for all 𝑛 ≥ 𝑁 then 𝑎𝑛 diverges.

Í
𝑛

Proof. In the first case, we infer via induction that 0 < 𝑎 𝑁 +𝑘 ≤ 𝑟 𝑘 𝑎 𝑁 for all 𝑘 ≥ 1. Therefore,
one has 0 < 𝑎𝑛 ≤ 𝑟 𝑛−𝑁 𝑎 𝑁 for all 𝑛 ≥ 𝑁 whence the series 𝑎𝑛 converges by comparison with a
Í
geometric series.
In the second case, we have 𝑎𝑛 ≥ 𝑎 𝑁 > 0 for all 𝑛 ≥ 𝑁 whence, by the divergence test, the
series diverges. □

As a consequence, we recover the ratio test in a slightly more familiar form from calculus:

Corollary 5.6 (Ratio Test). Let (𝑎𝑛 ) be a sequence of positive numbers such that
𝑎𝑛+1
lim = 𝐿.
𝑛→∞ 𝑎𝑛
1. If 𝐿 < 1 then
Í
𝑎𝑛 converges.

2. If 𝐿 > 1 then
Í
𝑎𝑛 diverges.

Proof. If 𝐿 < ∞ then, given any 𝜀 > 0, there exists 𝑁 ∈ N such that
𝑎𝑛+1
𝐿 −𝜀 < < 𝐿 + 𝜀, ∀𝑛 ≥ 𝑁 .
𝑎𝑛
If 𝐿 < 1, we can choose 𝜀 > 0 so small that 𝐿 + 𝜀 < 1 whence we have 𝑎𝑛+1 ≤ (𝐿 + 𝜀)𝑎𝑛 for all
𝑛 ≥ 𝑁 . It then follows from the limit free ratio test that the series 𝑎𝑛 is convergent. If 𝐿 > 1,
Í
we choose 𝜀 > 0 such that 𝐿 − 𝜀 > 1 whence it follows by the limit free ratio test that the given
series is divergent.
Similarly if 𝐿 = ∞, there exists 𝑁 ∈ N such that
𝑎𝑛+1
> 1 ∀𝑛 ≥ 𝑁 .
𝑎𝑛
Once, again, it follows from the ratio-free limit test that 𝑎𝑛 diverges.
Í
□

90
5.3 Examples of Convergence/Divergence of Series.
Having dispensed with the proofs of the limit comparison test and the ratio tests, we are ready to
tackle some more examples of series. Note that although we will not have the time to cover all of
these in the tutorial, the full solutions to each example is included nonetheless.
Example 5.5. Let 𝑎 1 = 1 and, given 𝑎𝑛 , define
2 + (−1)𝑛
𝑎𝑛+1 = 𝑎𝑛 .
4
By induction, observe that (𝑎𝑛 ) is a sequence of positive numbers. To determine whether
Í
𝑎𝑛
converges, we could try to apply the ratio test. However, the limit of
𝑎𝑛+1 2 + (−1)𝑛
=
𝑎𝑛 4
does not exist. Despite this, observe that
𝑎𝑛+1 2 + (−1)𝑛 1+2 3
= ≤ = .
𝑎𝑛 4 4 4
Therefore, by the limit-free ratio test 𝑎𝑛 converges.
Í
Í∞ 𝑛2 +𝑛
Example 5.6. Consider the series 𝑛=6 𝑛 4 −𝑛
.
𝑛 2 +𝑛 1
Solution. Define 𝑎𝑛 := 𝑛 4 −𝑛
and put 𝑏𝑛 := 𝑛2
. Then, for each 𝑛 ∈ N, one has

𝑎𝑛 𝑛 2 + 𝑛 2 𝑛 4 + 𝑛 3
= ·𝑛 = 4 .
𝑏𝑛 𝑛 4 − 𝑛 𝑛 −𝑛
It follows that
𝑎𝑛
=1lim
𝑛→∞ 𝑏𝑛

whence by the limit comparison test the series 𝑎𝑛 is convergent.

Í
□
Example 5.7. Consider the series (2𝑛)!2𝑛 .
Í

Solution. Define
(2𝑛)!
𝑎𝑛 :=
2𝑛
so that
(2𝑛 + 2)!
𝑎𝑛+1 = .
2𝑛+1
Then, for each 𝑛 ∈ N, we have
𝑎𝑛+1 (2𝑛 + 2)! 2𝑛 (2𝑛 + 2) (2𝑛 + 1) 𝑛→∞
= · = −−−−→ ∞.
𝑎𝑛 2𝑛+1 (2𝑛)! 2
By the ratio test, the series 𝑎𝑛 is divergent.
Í
□

91
 (2𝑛)!
Example 5.8. Consider the series 𝑛=8 [ (𝑛−1)!] 3 .
Í∞

Solution. Let us define (𝑎𝑛 ) by the rule

(2𝑛)!
𝑎𝑛 := , ∀𝑛 ∈ N
[(𝑛 − 1)!] 3
and observe that
(2𝑛 + 2)!
𝑎𝑛+1 := , ∀𝑛 ∈ N.
[𝑛!] 3
Clearly,

𝑎𝑛+1 (2𝑛 + 2)! [(𝑛 − 1)!] 3 (2𝑛 + 2) (2𝑛 + 1)

= · = .
𝑎𝑛 [𝑛!] 3 (2𝑛)! 𝑛3
Thus,
𝑎𝑛+1
lim
=0
𝑛→∞ 𝑎𝑛

whence our series 𝑎𝑛 is convergent by the ratio test.

Í
□
Í∞ 𝑛5+𝑛
Example 5.9. Consider the series 𝑛=1 (5𝑛+1) 𝑛
.

𝑛 5+𝑛
Solution. Let us define 𝑎𝑛 := (5𝑛+1) 𝑛
so that

𝑛 5/𝑛 · 𝑛
|𝑎𝑛 | 1/𝑛 = 𝑎𝑛1/𝑛 =
𝑛
= (𝑛 1/𝑛 ) 5 ·
5𝑛 + 1 5𝑛 + 1
1
→ .
5
□

Example 5.10. Consider the series (1 − cos(1/𝑛)).

Í∞
𝑛=1

Solution. Set
1
 
2
𝑎𝑛 = (1 − cos(1/𝑛)) , 𝑏𝑛 = 1 − cos
𝑛
Observe that
𝑏𝑛 = (1 − cos(1/𝑛)) (1 + cos(1/𝑛)) = (1 + cos(1/𝑛)) 𝑎𝑛 .
Hence,
𝑎𝑛 1 1 1
= → = .
𝑏𝑛 1 + cos(1/𝑛) 1 + cos(0) 2
By the limit comparison test, 𝑎𝑛 converges if and only 𝑏𝑛 converges.
Í Í
On the other hand,
2 1 2 1
   
𝑏𝑛 = 1 − cos = sin
𝑛 𝑛

92
We have seen in the second tutorial that sin(𝑥) ≤ |𝑥 |. Therefore,
 2
1 12 1

𝑏𝑛 = sin ≤ = 2.
𝑛 𝑛 𝑛

By the 𝑝-test, 𝑏𝑛 converges. We conclude that 𝑛=1 (1 − cos(1/𝑛)) is a convergent series.

Í Í∞
□
ln(𝑛)
Example 5.11. Consider the series 𝑛 .
Í∞
𝑛=1

Solution. Set 𝑎𝑛 = ln(𝑛)/𝑛. When a series contains logarithms, we have previously used the
Cauchy condensation test. However, this test only applies if (𝑎𝑛 ) is a decreasing sequence. Since
it is not clear whether (𝑎𝑛 ) is decreasing, let us try a different method. Consider 𝑏𝑛 = 𝑛1 . Observe
that
𝑎𝑛
lim = lim ln(𝑛) = ∞.
𝑛→∞ 𝑏𝑛 𝑛→∞

Therefore, by the limit comparison test, since 𝑏𝑛 diverges 𝑎𝑛 also diverges.

Í Í
□
Í √
Example 5.12. Consider the series 𝑒 −1 .
𝑛

𝑛

√
Solution. By the root test, this series converges. Indeed, if we define 𝑎𝑛 := 𝑛 𝑒 − 1 we see that

𝑛
√ √
𝑎𝑛 ≥ 0 and that 𝑎𝑛1/𝑛 = 𝑛 𝑒 −1. Since 𝑛 𝑒 → 1 as 𝑛 → ∞, it follows that lim𝑛→∞ 𝑎𝑛1/𝑛 = 1−1 = 0. □
Í∞ √
Example 5.13. Consider 𝑛=1 𝑒 − 1 . To determine if the series converges or diverges, you
𝑛

can use that the derivative of 𝑓 (𝑥) = 𝑒 𝑥 is 𝑓 ′ (𝑥) = 𝑒 𝑥 .

Solution. First, we claim that 𝑒 𝑥 ≥ 𝑥 + 1 for all 𝑥 ≥ 0. To this end, consider ℎ : [0, ∞) → R given
by ℎ(𝑥) = 𝑒 𝑥 − 𝑥 − 1. We have ℎ ′ (𝑥) = 𝑒 𝑥 − 1 ≥ 0 so ℎ is an increasing function. Since, ℎ(0) = 0,
we obtain ℎ(𝑥) ≥ 0 or, equivalently, 𝑒 𝑥 ≥ 𝑥 + 1 for all 𝑥 ≥ 0.
Using this inequality, we may write

√ 1 1
 
1/𝑛
𝑛
𝑒 −1=𝑒 −1 ≥ +1 −1= .
𝑛 𝑛
Í∞ √
Since 𝑛1 diverges, the comparison test implies that 𝑛=1 𝑒 − 1 diverges.
Í

𝑛
□
Í∞ 2𝑛 +𝑛2
 
Example 5.14. Consider the series 𝑛=1 𝑛 2 ·2𝑛
.

Solution. Observe that

2𝑛 + 𝑛 2 1 1
2
= 2 + 𝑛.
𝑛 ·2 𝑛 𝑛 2
1 1 Í∞  𝑛2 +2𝑛 
Since and converge, the limit laws imply that 𝑛=1 converges.
Í Í
2𝑛 𝑛2 𝑛 2 ·2𝑛
□

Í∞  2𝑛 +𝑛 
Example 5.15. Consider the series 𝑛=1 𝑛·2𝑛 .

93
Solution. We can write
𝑛 + 2𝑛 1 1
= + 𝑛.
𝑛 · 2𝑛 𝑛 2
1 Í∞  2𝑛 +𝑛 
Since diverges, the comparison test implies that 𝑛=1 𝑛·2𝑛 diverges.
Í
𝑛 □

2𝑛 𝑛 7
Example 5.16. Consider the series 𝑛=1 3𝑛 .
Í∞

2𝑛 𝑛 7
Solution. Setting 𝑎𝑛 = 3𝑛 we see that
7
𝑎𝑛+1 2𝑛+1 (𝑛 + 1) 7 3𝑛 2 𝑛+1

= · 7 𝑛 = .
𝑎𝑛 3 ( 𝑛 + 1) 𝑛 ·2 3 𝑛
By the limit laws,
𝑎𝑛+1 2
lim = < 1.
𝑛→∞ 𝑎𝑛 3
2𝑛 𝑛 7
By the ratio test, the series converges.
Í∞
𝑛=1 3𝑛 □
𝑛!+(−1) 𝑛
Example 5.17. Consider 𝑛+5 .
Í∞
𝑛=1

Solution. By the divergence test, this series diverges. Indeed,

𝑛! + (−1)𝑛 𝑛! (−1)𝑛 𝑛 (−1)𝑛
= + = (𝑛 − 1)! +
𝑛+5 𝑛+5 𝑛+5 𝑛+5 𝑛+5
does not converge to 0. □
∞
∑︁ 1 · 3 · 5 · · · (2𝑛 − 1)
Example 5.18. Consider the series .
𝑛=1
2 · 4 · 6 · · · (2𝑛)

(a) Show that the ratio test is inconclusive when applied to this series.

(b) By establishing the following inequality:

1 · 3 · 5 · · · (2𝑛 − 1) 1
≥ , ∀𝑛 ∈ N.
2 · 4 · 6 · · · (2𝑛) 2𝑛

Solution.
Letting (𝑎𝑛 ) be the sequence of summands given, we see that for all 𝑛 ≥ 1
𝑎𝑛+1 1 · 3 · 5 · · · (2𝑛 − 1) · (2𝑛 + 1) 2 · 4 · 6 · · · (2𝑛)
= ·
𝑎𝑛 2 · 4 · 6 · · · (2𝑛) · (2𝑛 + 2) 1 · 3 · 5 · · · (2𝑛 − 1)
2𝑛 + 1
= .
2𝑛 + 2
Therefore, 𝑎𝑛+1 /𝑎𝑛 → 1 which is inconclusive.

94
We establish the stated inequality via induction. Plugging in 𝑛 = 1, the identity reduces to

2(1) − 1 1 1
= ≥
2(1) 2 2
which holds true. Assuming the identity holds for 𝑛 ∈ N, observe that by our inductive hypothesis

1 · 3 · 5 · · · (2𝑛 − 1) (2𝑛 + 1) 1 · 3 · 5 · · · (2𝑛 − 1) 1 2𝑛 + 1

 
= ≥ ·
2 · 4 · 6 · · · (2𝑛) (2𝑛 + 2) 2 · 4 · 6 · · · (2𝑛) 2𝑛 2𝑛 + 2
1 2𝑛 + 1
≥ ·
2𝑛 2𝑛 + 2
1
≥ .
2𝑛
Hence, we need only compare with the harmonic series. □

5.4 Abel’s Test

We now treat examples relating to Abel’s test, which we recall below from the class notes.

Theorem 5.7 (Abel’s Test). Let 𝑛=1

Í∞
𝑎𝑛 be a convergent series and let (𝑏𝑛 ) be a bounded monotone
sequence. Then, the series
∞
∑︁
𝑎𝑛 𝑏 𝑛
𝑛=1
is convergent.

Example 5.19. Consider the series

∞
1
 𝑛
∑︁ (−1)𝑛
· 1+ .
𝑛=1
𝑛 𝑛

This can be written as

∞
∑︁
𝑎𝑛 𝑏 𝑛
𝑛=1

where
1
 𝑛
(−1)𝑛
𝑎𝑛 := and 𝑏𝑛 := 1 + .
𝑛 𝑛
Now, the sequence
1
|𝑎𝑛 | =
𝑛
is monotone decreasing and converges to 0. Therefore, the Alternating Series Test ensures that
𝑛=1 𝑎𝑛 is convergent. If we can show that (𝑏𝑛 ) is bounded and monotone, it will follow from
Í∞

95
Abel’s test that the original series 𝑛=1 𝑎𝑛𝑏𝑛 converges. Luckily for us, this is already a known
Í∞
fact! Certainly, it was proven Analysis 1 that
1
 𝑛
𝑏𝑛 = 1 +
𝑛
is monotone increasing with lim 𝑏𝑛 = 𝑒.
Example 5.20. We prove that the series
∞
!
2𝑛 2 − 3𝑛 + 1 ∑︁ 1
∑︁ 𝑛

𝑛=1
4𝑛 5 − 3 𝑘2
𝑘=1

is convergent. With the hope of applying Abel’s test, we will write this series as where
Í∞
𝑛=1 𝑎𝑛𝑏𝑛

2𝑛 2 − 3𝑛 + 1 1
𝑛
∑︁
𝑎𝑛 := and 𝑏𝑛 := .
4𝑛 5 − 3 𝑘2
𝑘=1

To successfully apply Abel’s test, the following must be verified:

(i) 𝑛=1 𝑎𝑛 converges;
Í∞

(ii) (𝑏𝑛 ) is monotone;

(iii) (𝑏𝑛 ) is bounded.
Let us first verify (ii)-(iii). Clearly, for each 𝑛 ∈ N,

1 1
𝑛
∑︁ 𝑛+1
∑︁
𝑏𝑛 = ≤ = 𝑏𝑛+1
𝑘2 𝑘2
𝑘=1 𝑘=1
Í∞ 1
whence (𝑏𝑛 ) is monotone increasing. Now, every 𝑏𝑛 is a partial sum for the series 𝑘=1 𝑘2
which
was shown to be convergent in the previous section (or tutorial). Consequently, (𝑏𝑛 ) is convergent
and hence bounded. Therefore, it only remains to establish (i). Here we will make use of the Limit
Comparison Test. By the 𝑝-test, we know that
∞
∑︁ 1
𝑛=1
𝑛3
is convergent. Since
2𝑛 2 −3𝑛+1
𝑛 3 2𝑛 2 − 3𝑛 + 1 2𝑛 5 − 3𝑛 4 + 1


4𝑛 5 −3
lim = lim = lim
1/𝑛 3 4𝑛 5 − 3 4𝑛 5 − 3
1
= >0
2
it follows from the Limit Comparison Test that 𝑛=1 𝑎𝑛 converges. By our previous remarks and
Í∞
Abel’s test, we infer that 𝑛=1 𝑎𝑛𝑏𝑛 is convergent.
Í∞

96
5.5 Dirichlet’s Test
Theorem 5.8 (Dirichlet’s Test). Let (𝑎𝑛 ) be a decreasing sequence of real numbers such that lim 𝑎𝑛 =
0. Let (𝑏𝑛 ) be a sequence such that there exists 𝑀 > 0 with the property that
𝑁
∑︁
𝑏𝑛 ≤ 𝑀
𝑛=1

for all 𝑁 ≥ 1. Then,

∞
∑︁
𝑎𝑛 𝑏 𝑛
𝑛=1
is convergent.

Example 5.21. We will prove that the series

∞
∑︁ cos(𝜋𝑛)
𝑛=1
ln 𝑛

is convergent using Dirichlet’s test. For this, we define

1
𝑎𝑛 := and 𝑏𝑛 := cos(𝜋𝑛).
ln 𝑛
Clearly, (𝑎𝑛 ) is monotone decreasing and converges to 0 as 𝑛 → ∞. Therefore, to apply Dirichlet’s
test, we need only check that there exists 𝑀 > 0 such that
𝑁
∑︁ 𝑁
∑︁
𝑏𝑛 = cos(𝜋𝑛) ≤ 𝑀
𝑛=1 𝑛=1

for all 𝑁 ∈ N. To see this, we note that

Í1
• 𝑛=1 cos(𝜋𝑛) = cos(𝜋) = −1;
Í2
• 𝑛=1 cos(𝜋𝑛) = cos(𝜋) + cos(2𝜋) = −1 + 1 = 0;
Í3
• 𝑛=1 cos(𝜋𝑛) = cos(𝜋) + cos(2𝜋) + cos(3𝜋) = −1 + 1 − 1 = −1;
Í4
• 𝑛=1 cos(𝜋𝑛) = cos(𝜋) + cos(2𝜋) + cos(3𝜋) + cos(4𝜋) = −1 + 1 − 1 + 1 = 0.

By induction, it follows that

𝑁
∑︁ 𝑁
∑︁
𝑏𝑛 = cos(𝜋𝑛) ≤ 1
𝑛=1 𝑛=1

for all 𝑁 ∈ N. Hence, Dirichlet’s test applies.

97
Example 5.22. Consider the series
∞
∑︁ 𝑐𝑛
√
𝑛=1 𝑛
where (𝑐𝑛 ) = (1, −4, 1, 2, 1, −4, 1, 2, 1, . . . ). Using Dirichlet’s test, we will prove that this series is
convergent. Clearly, √1𝑛 is a decreasing sequence of real numbers converging to 0. It remains to
show that the partial sums 𝑛=1 𝑐𝑛 are bounded in 𝑁 . As before, we try to notice a pattern in
Í𝑁
these partial sums:
• For 𝑁 = 1 we have 𝑛=1 𝑐𝑛 = 1;
Í𝑁

• For 𝑁 = 2 we have 𝑛=1 𝑐𝑛 = 1 − 4 = −3;

Í𝑁

• For 𝑁 = 3 we have 𝑛=1 𝑐𝑛 = 1 − 4 + 1 = −2;

Í𝑁

• For 𝑁 = 4 we have 𝑛=1 𝑐𝑛 = 1 − 4 + 1 + 2 = 0;

Í𝑁

• For 𝑁 = 5 we have 𝑛=1 𝑐𝑛 = 1 − 4 + 1 + 2 + 1 = 1;

Í𝑁

and so forth. Therefore,

𝑁
∑︁
𝑐𝑛 ≤ 3
𝑛=1
for all 𝑁 ∈ N. Dirichlet’s test thus yields the convergence of our series.

Example 5.23. Consider the series

1 1 1 1 1 1 1 1
1+ − + + − + + − +···
2 3 4 5 6 7 8 9
Namely, we consider the harmonic series where each 3rd term is negative. By way of contradic-
tion, let us assume that this series converges. Let (𝑆 𝑁 ) denote the sequence of partial sums. By
assumption, (𝑆 𝑁 ), and hence (𝑆 3𝑁 ), is convergent. Now, each 𝑆 3𝑁 is given by

1 1 1
𝑁 
∑︁ 
𝑆 3𝑁 = + − .
𝑛=1
3𝑛 − 2 3𝑛 − 1 3𝑛

Since lim 𝑆 3𝑁 exists by assumption, we see that the series

∞  ∞
1 1 1 9𝑛 2 − 2
∑︁  ∑︁
+ − = .
𝑛=1
3𝑛 − 2 3𝑛 − 1 3𝑛 𝑛=1
3𝑛(3𝑛 − 2) (3𝑛 − 1)

must be convergent. Note that every term of the series

∞
∑︁ 9𝑛 2 − 2
𝑛=1
3𝑛(3𝑛 − 2) (3𝑛 − 1)

98
is non-negative. On the other hand,
9𝑛 2 −2
3𝑛 (3𝑛−2) (3𝑛−1) 9𝑛 3 − 2𝑛 1
lim 1
= lim = > 0.
𝑛
3𝑛(3𝑛 − 2) (3𝑛 − 1) 3
1
Using the Limit Comparison Test with the divergent harmonic series 𝑛=1 𝑛 , we infer that the
Í∞
series
∞
∑︁ 9𝑛 2 − 2
𝑛=1
3𝑛(3𝑛 − 2) (3𝑛 − 1)
is divergent. This contradiction shows that (𝑆 𝑁 ) cannot be convergent.

5.6 Weierstass 𝑀-Test

Theorem 5.9. Let (𝑓𝑛 ) be a sequence of real valued functions defined on a set 𝐴 ⊆ R and assume
that for each 𝑛 ∈ N there exists 𝑀𝑛 > 0 with the property that |𝑓𝑛 (𝑥)| ≤ 𝑀𝑛 for all 𝑥 ∈ 𝐴. Then, if
∞
∑︁
𝑀𝑛
𝑛=1

converges, the series

∞
∑︁
𝑓𝑛 (𝑥)
𝑛=1

converges uniformly and absolutely on 𝐴 as 𝑛 → ∞.

Example 5.24. Prove that the following series

∞
∑︁ 𝑛2 + 𝑥 4
𝑓 (𝑥) :=
𝑛=1
𝑛4 + 𝑥 2

converges to a continuous function 𝑓 : R → R.

Proof. Let 𝑅 > 0 be given and consider the interval 𝐼 = [−𝑅, 𝑅]. Given any 𝑥 ∈ 𝐼 , observe that

𝑛2 + 𝑥 4 𝑛2 + 𝑥 4 𝑛2 + 𝑅4
≤ ≤
𝑛4 + 𝑥 2 𝑛4 𝑛4
1 𝑅4
= 2 + 4.
𝑛 𝑛
If we define
1 𝑅4
𝑀𝑛 := +
𝑛2 𝑛4

99
, then it automatically follows from the 𝑝-test that
∞ ∞ ∞
∑︁ ∑︁ 1 4
∑︁ 1
𝑀𝑛 = 2
+ 𝑅
𝑛=1 𝑛=1
𝑛 𝑛=1
𝑛4

is convergent. Consequently, it follows from the Weierstass 𝑀-test that the series defining 𝑓
converges absolutely and uniformly on 𝐼 . Since the partial sums are continuous and uniform
limits preserve continuity, we infer that 𝑓 is continuous on 𝐼 = [−𝑅, 𝑅]. Since continuity is a local
property and 𝑅 > 0 was arbitrary, we see that 𝑓 is continuous R → R. □

Example 5.25. We claim that the series

∞ ∞
∑︁ sin(𝑛) ∑︁ cos(𝑛)
and
𝑛=1
𝑛 𝑛=1
𝑛

are convergent.

Solution. We only handle the first series, as a similar argument applies to the latter. Observe that
2 sin(1) ≠ 0 and so, for each 𝑁 ∈ N,
𝑁
∑︁ 𝑁
∑︁
2 sin(1) sin(𝑛) = 2 sin(1) sin(𝑛)
𝑛=1 𝑛=1
𝑁
∑︁
= [cos(1 − 𝑛) − cos(1 + 𝑛)]
𝑛=1
𝑁
∑︁
= [cos(𝑛 − 1) − cos(1 + 𝑛)] ,
𝑛=1

where we have used the identity

2 sin(𝛼) sin(𝛽) = cos(𝛼 − 𝛽) − cos(𝛼 + 𝛽).

Consequently,
𝑁
∑︁
2 sin(1) sin(𝑛) = cos(0) − cos(1 + 𝑁 )
𝑛=1

whence
|cos(0) − cos(1 + 𝑁 )| 2 1
𝑁
∑︁
sin(𝑛) = ≤ = .
𝑛=1
2 sin(1) 2 sin(1) sin(1)
is bounded independently of 𝑁 . Consequently, the series converges by Dirichlet’s test. Similarly,
one can handle the series with cos. □

100