Honors Calculus

Pete L. Clark

c Pete L. Clark, 2014.

Thanks to Ron Freiwald, Eilidh Geddes, Melody Hine, Nita Jain, Andrew Kane,
Bryan Oakley, Elliot Outland, Didier Piau, Jim Propp, Troy Woo, and Sharon
Yao.

Contents
Foreword
Spivak and Me
What is Honors Calculus?
Some Features of the Text

7
7
10
10

Chapter 1. Introduction and Foundations

1. Introduction
2. Some Properties of Numbers

13
13
17

Chapter 2. Mathematical Induction

1. Introduction
2. The First Induction Proofs
3. Closed Form Identities
4. Inequalities
5. Extending Binary Properties to n-ary Properties
6. The Principle of Strong/Complete Induction
7. Solving Homogeneous Linear Recurrences
8. The Well-Ordering Principle
9. The Fundamental Theorem of Arithmetic

25
25
27
29
31
32
34
35
39
40

Chapter 3. Polynomial and Rational Functions

1. Polynomial Functions
2. Rational Functions

43
43
49

Chapter 4. Continuity and Limits

1. Remarks on the Early History of the Calculus
2. Derivatives Without a Careful Definition of Limits
3. Limits in Terms of Continuity
4. Continuity Done Right
5. Limits Done Right

53
53
54
56
58
62

Chapter 5. Differentiation
1. Differentiability Versus Continuity
2. Differentiation Rules
3. Optimization
4. The Mean Value Theorem
5. Monotone Functions
6. Inverse Functions I: Theory
7. Inverse Functions II: Examples and Applications
8. Some Complements
3

71
71
72
77
81
85
92
97
104

CONTENTS

Chapter 6. Completeness
1. Dedekind Completeness
2. Intervals and the Intermediate Value Theorem
3. The Monotone Jump Theorem
4. Real Induction
5. The Extreme Value Theorem
6. The Heine-Borel Theorem
7. Uniform Continuity
8. The Bolzano-Weierstrass Theorem For Subsets
9. Tarskis Fixed Point Theorem

105
105
113
116
117
119
119
120
122
123

Chapter 7. Differential Miscellany

1. LH
opitals Rule
2. Newtons Method
3. Convex Functions

125
125
128
136

Chapter 8. Integration
1. The Fundamental Theorem of Calculus
2. Building the Definite Integral
3. Further Results on Integration
4. Riemann Sums, Dicing, and the Riemann Integral
5. Lesbesgues Theorem
6. Improper Integrals
7. Some Complements

149
149
152
161
170
176
179
182

Chapter 9. Integral Miscellany

1. The Mean Value Therem for Integrals
2. Some Antidifferentiation Techniques
3. Approximate Integration
4. Integral Inequalities
5. The Riemann-Lebesgue Lemma

185
185
185
187
194
196

Chapter 10. Infinite Sequences

1. Summation by Parts
2. Easy Facts
3. Characterizing Continuity
4. Monotone Sequences
5. Subsequences
6. The Bolzano-Weierstrass Theorem For Sequences
7. Partial Limits; Limits Superior and Inferior
8. Cauchy Sequences
9. Geometric Sequences and Series
10. Contraction Mappings Revisited
11. Extending Continuous Functions

199
200
201
203
203
206
208
211
214
216
217
225

Chapter 11. Infinite Series

1. Introduction
2. Basic Operations on Series
3. Series With Non-Negative Terms I: Comparison
4. Series With Non-Negative Terms II: Condensation and Integration

229
229
233
236
241

CONTENTS

5.
6.
7.
8.

Series With Non-Negative Terms III: Ratios and Roots

Absolute Convergence
Non-Absolute Convergence
Power Series I: Power Series as Series

247
250
254
260

Chapter 12. Taylor Taylor Taylor Taylor

1. Taylor Polynomials
2. Taylors Theorem Without Remainder
3. Taylors Theorem With Remainder
4. Taylor Series
5. Hermite Interpolation

265
265
266
269
271
277

Chapter 13. Sequences and Series of Functions

1. Pointwise Convergence
2. Uniform Convergence
3. Power Series II: Power Series as (Wonderful) Functions

281
281
283
288

Chapter 14. Serial Miscellany

P 1
2
1.
n=1 n2 = 6
2. Rearrangements and Unordered Summation
3. Abels Theorem
4. The Peano-Borel Theorem
5. The Weierstrass Approximation Theorem
6. A Continuous, Nowhere Differentiable Function
7. The Gamma Function

291
291
293
302
306
310
313
315

Chapter 15. Several Real Variables and Complex Numbers

1. A Crash Course in the Honors Calculus of Several Variables
2. Complex Numbers and Complex Series
3. Elementary Functions Over the Complex Numbers
4. The Fundamental Theorem of Algebra

321
321
323
325
327

Chapter 16. Foundations Revisited

1. Ordered Fields
2. The Sequential Completion

333
334
342

Bibliography

351

Foreword
Spivak and Me
The document you are currently reading began its life as the lecture notes for a year
long undergraduate course in honors calculus. Specifically, during the 2011-2012
academic year I taught Math 2400(H) and Math 2410(H), Calculus With Theory,
at the University of Georgia. This is a course for unusually talented and motivated
(mostly first year) undergraduate students. It has been offered for many years at
the University of Georgia, and so far as I know the course text has always been
Michael Spivaks celebrated Calculus [S]. The Spivak texts take on calculus is
sufficiently theoretical that, although it is much beloved by students and practitioners of mathematics, it is seldomed used nowadays as a course text. In fact, the
UGA Math 2400/2410 course is traditionally something of an interpolation between
standard freshman calculus and the fully theoretical approach of Spivak. My own
take on the course was different: I treated it as being a sequel to, rather than an
enriched revision of, freshman calculus.
I began the course with a substantial familiarity with Spivaks text. The summer after my junior year of high school I interviewed at the University of Chicago
and visited Paul Sally, then (and still now, as I write this in early 2013, though
he is 80 years old) Director of Undergraduate Mathematics at the University of
Chicago. After hearing that I had taken AP calculus and recently completed a
summer course in multivariable calculus at Johns Hopkins, Sally led me to a supply closet, rummaged through it, and came out with a beat up old copy of Spivaks
text. This is how we do calculus around here, he said, presenting it to me. During
my senior year at high school I took more advanced math courses at LaSalle University (which turned out, almost magically, to be located directly adjacent to my
high school) but read through Spivaks text. And I must have learned something
from it, because by the time I went on to college of course at the University of
Chicago I placed not into their Spivak calculus course, but the following course,
Honors Analysis in Rn . This course has the reputation of sharing the honor with
Harvards Math 55 of being the hardest undergraduate math course that American
universities have to offer. I cant speak to that, but it was certainly the hardest
math course I ever took. There were three ten week quarters. The first quarter was
primarily taught out of Rudins classic text [R], with an emphasis on metric spaces.
The second quarter treated Lebesgue integration and some Fourier analysis, and
the third quarter treated analysis on manifolds and Stokess theorem.
However, that was not the end of my exposure to Spivaks text. In my second
year of college I was a grader for the first quarter of Spivak calculus, at the (even
then) amazingly low rate of $20 per student for the entire 10 week quarter. Though
7

FOREWORD

the material in Spivaks text was at least a level below the trial by fire that I had
successfully endured in the previous year, I found that there were many truly difficult problems in Spivaks text, including a non-negligible percentage that I still
did not know how to solve. Grading for this course solidified my knowledge of this
elementary but important material. It was also my first experience with reading
proofs written by bright but very inexperienced authors: often I would stare at
an entire page of text, one long paragraph, and eventually circle a single sentence
which carried the entire content that the writer was trying to express. I only graded
for one quarter after that, but I was a drop in tutor for my last three years of
college, meaning that I would field questions from any undergraduate math course
that a student was having trouble with, and I had many further interactions with
Spivaks text. (By the end of my undergraduate career there were a small number
of double-starred problems of which I well knew to steer clear.)
Here is what I remembered about Spivaks text in fall of 2011:
(i) It is an amazing trove of problems, some of which are truly difficult.
(ii) The text itself is lively and idiosyncratic.1
(iii) The organization is somewhat eccentric. In particular limits are not touched
until Chapter 5. The text begins with a chapter basic properties of numbers,
which are essentially the ordered field axioms, although not called such. Chapter
3 is on Functions, and Chapter 4 is on Graphs. These chapters are essentially contentless. The text is broken up into five parts of which the third is Derivatives
and Integrals.
After teaching the 2400 course for a while, I lost no esteem for Spivaks text,
but increasingly I realized it was not the ideal accompaniment for my course. For
one thing, I realized that my much more vivid memories of the problems than the
text itself had some basis in fact: although Spivak writes with a lively and distinct
voice and has many humorous turns of phrase, the text itself is rather spare. His
words are (very) well chosen, but few. When one takes into account the ample
margins (sometimes used for figures, but most often a white expanse) the chapters
themselves are very short and core-minded. When given the chance to introduce a
subtlety or ancillary topic, Spivak almost inevitably defers it to the problems.
I have had many years to reflect on Spivaks text, and I now think its best use
is in fact the way I first encountered it myself: as a source of self study for bright,
motivated students with little prior background and exposure to university level
mathematics (in our day, this probably means high school students, but I could
imagine a student underwhelmed by an ordinary freshman calculus class for which
Spivaks text would be a similarly mighty gift). Being good for self study is a
high compliment to pay a text, and any such text can a fortiori also be used in a
course...but not in a completely straightforward way. For my course, although the
1In between the edition of the book that Sally had given me and the following edition, all
instances of third person pronouns referring to mathematicians had been changed from he to
she. Initially I found this amusing but silly. More recently I have begun doing so myself.

SPIVAK AND ME

students who stuck with it were very motivated and hard-working, most (or all) of
them needed a lot of help from me in all aspects of the course. I had been informed
in advance by my colleague Ted Shifrin that signing on to teach this course should
entail committing to about twice as many office hours as is typical for an undergraduate course, and this did indeed come to pass: I ended up having office hours
four days a week, and I spent the majority of them helping the students make their
way through the problems in Spivaks text. Even the best students who, by the
way, I thought were awfully good could not solve some of the problems unassisted.
There were many beautiful, multi-part problems introducing extra material that I
wanted my students to experience: and eventually I figured out that the best way
to do this was to incorporate the problems into my lectures as theorems and proofs.
All in all I ended up viewing Spivaks book as being something of a deconstruction2 of the material, and much of my teaching time was spent reconstructing it.
For me, the lightness of touch of Spivaks approach was ultimately something I
appreciated aesthetically but could see was causing my students difficulty at various key points. My experience rather convinced me that more is more: students
wanted to see more arguments in complete detail, and also simply more proofs overall. Aside from the (substantial) content conveyed in the course, a major goal is to
give the students facility with reading, understanding, constructing and critiquing
proofs. In this regard being in a classroom setting with other motivated students is
certainly invaluable, and I tried to take advantage of this by insisting that students
present their questions and arguments to each other as well as to me.
But students also learn about how to reason and express themselves mathematically by being repeatedly exposed to careful, complete proofs. I think experienced
mathematicians can forget the extent to which excellent, even brilliant, students
benefit from this exposure. Of course they also benefit immensely, probably more
so, by working things out for themselves. The value of this is certainly not in
question, and there is no competition here: the amount of things to do in honors
calculus is infinite and the repository of problems in Spivaks text is nearly so
so by working out more for herself, the instructor is not leaving less for the students
to do, but only different things for them to do.
This brings me to the current text. As explained above, it is heavily indebted
to [S]. However, it is or at least, I mean it to be a new honors calculus text,
and not merely a gloss of [S]. Indeed:
The text is indebted to [S], but not uniquely so. It is also heavily indebted
to Rudins classic text [R], and it borrows at key points from several other sources,
e.g. [A], [Go], [H], [L].
I do not view this borrowing as being in any way detrimental, and I certainly do
not attempt to suppress my reliance on other texts. The mathematics exposed here
is hundreds of years old and has been treated excellently in many famous texts. An
undergraduate mathematics text which strove to be unlike its predecessors would
2My conception of the meaning of deconstruction in the sense of academic humanities is
painfully vague. I am more thinking of the term in the sense that chefs use it.

10

FOREWORD

almost certainly do so to the detriment of its audience.

Having said this, most undergraduate texts with calculus in the title offer
very little innovation indeed. It is distressing to me how many calculus books are
written which look like nearly exact replicas of some platonic (but far from perfect)
calculus text written circa 1920. Apparently the issue of treating transcendental
functions earlier rather than later was enough to warrant many new editions, if not
wholly new texts.
This text does contain some novelties, up to and including a proof technique,
real induction, which to my knowledge has not appeared before in texts. Spivaks
text is a great innovation, and its distinctiveness some might say eccentricity
has not gone unnoticed by instructors. This text is also eccentric, and in its own
way: if you like Spivaks text, you need not like this one.
What is Honors Calculus?
It was audacious of Spivak to call his text simply calculus, since it is light years
away from any other text of that name currently in print. I have chosen to call this
text: honors calculus. What is honors calculus?
By honors calculus I mean a certain take on undergraduate real analysis.
Some Features of the Text
Our approach is heavily theoretical but not heavily abstract.
The jumping off point for an honors calculus course is a theoretical understanding
of the results seen in freshman calculus, especially rigorous definitions of the various
limiting processes and proofs of the hard theorems of calculus, which inevitably
turn on the completeness of R. We are not interested in calculation for its own sake,
and we do not give any halfway serious applications to any subject outside mathematics. This theoretical perspective is of course a huge change from the practical,
problem-oriented approach of freshman calculus (with which we assume a prior
familiarity). It is certainly not as appealing to as broad an audience as freshman
calculus. At the University of Georgia, it happens that many freshman students
(especially, those in the honors program) are signed up for Math 2400 without any
real idea of what they are getting into. The attrition at the beginning of the course
is therefore significant, and the students that remain are not necessarily the most
talented or experienced but rather those that remain interested in a fully theoretical
approach to calculus. This text assumes that interest on the part of the reader, in
fact probably even more so than Spivaks text.
The idea of presenting a fully theoretical take on real analysis to a young undergraduate audience is hardly new, of course. But most contemporary texts on
real analysis take a more sophisticated approach, making use of concepts from set
theory and topology. In this text we rarely speak explicitly about the connectedness and compactness of intervals on the real line. (More precisely, we never use
these concepts in the main text, but only in some digressions which briefly and
optionally present some advanced material: e.g. 10.10.5 introduces metric spaces
and discussed compactness therein.) Instead we make - arguments which amount
to this. Let us freely admit that this concreteness makes certain arguments longer

SOME FEATURES OF THE TEXT

11

and harder (or phrased more positively: one merit of abstraction is to make certain
arguments shorter and easier ).
I am a fan of abstraction, but I have come to believe that it is much less useful and moreover, much less educational in basic real analysis than in most
other branches of pure mathematics. A turning point in my feelings about this
was a second semester undergraduate real analysis course I taught at McGill University in 2005. When preparing to teach the course I was quite surprised at how
pedestrian the syllabus seemed: an entire year of real analysis was restricted to
the one-dimensional case, and there was no topology or set theory whatsoever.3 I
was strictly forbidden from talking about metric spaces, for instance. The topics of
the course were: infinite series, the Riemann integral, and sequences and series of
functions, including uniform convergence. By the end of the course I had acquired
more respect for the deep content inherent in these basic topics and the efficacy of
treating them using no more than - arguments and the completeness axiom. This
was also the first course I taught in which I typed up lecture notes, and portions of
these notes appear here as parts of Chapters 11 through 14.
Exception: We begin our treatment of the Riemann integral with an axiomatic approach: that is, we list some reasonable properties (axioms) that an area functional
should satisfy, and before we do the hard work of constructing such a functional
we explore the consequences of these axioms. In particular, less than two pages
into our discussion of integration we state and prove (given the axioms!) the Fundamental Theorem of Calculus. This approach to the Riemann Integral resonates
deeply with me, as it addresses concerns that bubbled up over several years of my
teaching this material in the context of freshman calculus. I came up with this
approach in late 2004 while preparing for the McGill analysis course. Just a few
days later I noticed that Lang had done (essentially) the same thing in his text [L].
I was encouraged that this idea had been used by others, and I endorse it still.
Given the choice between pounding out an - argument and developing a more
abstract or softer technique that leads to an easier proof, in this text we usually
opt for the former. Well, thats not strictly true: sometimes we do both.
We often spend time giving multiple proofs and approaches to basic results.
For instance, there are two distinct approaches to the Riemann integral: Riemanns
original approach using tagged partitions and Riemann sums, and Gaston Darbouxs later simplification using upper and lower sums and integrals. Most texts at
this level cover only one of these in detail, but here we cover both: first Darboux,
then Riemann. I got permission to do so while teaching the analysis class at
McGill, since this was done in the official course text of R. Gordon [Go]. Later I
realized that Gordon is in real life an integration theorist!
More significantly, we prove the Interval Theorems using - arguments and later
come back to give much quicker proofs using sequences and the Bolzano-Weierstrass
Theorem. One may well say that this is evidence that sequences should be treated
3In fact, I didnt find out until just after the course ended that the students did not know
about countable and uncountable sets. Without conscious thought I had assumed otherwise.

12

FOREWORD

at the beginning of the course rather than towards the end, and many texts do
take this approach, most notably [R]. However I endorse Spivaks ordering of the
material, in which honors calculus essentially begins with a thorough grappling
with the - definition of limit. Although there are easier and especially, softer
approaches to most of the key theorems, I feel that manipulation of inequalities
that characterized hard analysis is a vitally important skill for students to learn
and this is their best chance to learn it.
We view the completeness axiom as the star of the show. Thus we do not put it
on the stage in Act I, Scene I, when people are still settling into their seats.

CHAPTER 1

Introduction and Foundations

1. Introduction
1.1. The Goal: Calculus Made Rigorous.
The goal of this course is to cover the material of single variable calculus in a
mathematically rigorous way. The latter phrase is important: in most calculus
classes the emphasis is on techniques and applications; while theoretical explanations may be given by the instructor e.g. it is usual to give some discussion of
the meaning of a continuous function the student tests her understanding of
the theory mostly or entirely through her ability to apply it to solve problems.
This course is very different: not only will theorems and proofs be presented in
class by me, but they will also be presented by you, the student, in homework and
on exams. This course offers a strong foundation for a students future study of
mathematics, at the undergraduate level and beyond.
As examples, here are three of the fundamental results of calculus; they are called
by me, at least the three Interval Theorems, because they all concern an
arbitrary continuous function defined on a closed, bounded interval.
Theorem 1.1. (Intermediate Value Theorem) Let f : [a, b] R be a continuous
function defined on a closed, bounded interval. Suppose that f (a) < 0 and f (b) > 0.
Then there exists c with a < c < b such that f (c) = 0.
Theorem 1.2. (Extreme Value Theorem) Let f : [a, b] R be a continuous
function defined on a closed, bonuded interval. Then f is bounded and assumes its
maximum and minimum values: there are numbers m M such that
a) For all x [a, b], m f (x) M .
b) There exists at least one x [a, b] such that f (x) = m.
c) There exists at least one x [a, b] such that f (x) = M .
Theorem 1.3. (Uniform Continuity and Integrability) Let f : [a, b] R be a
continuous function defined on a closed, bounded interval. Then:
a) f is uniformly continuous.1
Rb
b) f is integrable: a f exists and is finite.
Except for the part about uniform continuity, these theorems are familiar results
from freshman calculus. Their proofs, however, are not. Most freshman calculus texts like to give at least some proofs, so it is often the case that these three
theorems are used to prove even more famous theorems in the course, e.g. the
1The definition of this is somewhat technical and will be given only later on in the course.
Please dont worry about it for now.
13

14

1. INTRODUCTION AND FOUNDATIONS

Mean Value Theorem and the Fundamental Theorem of Calculus.

Why then are the three interval theorems not proved in freshman calculus? Because
their proofs depend upon fundamental properties of the real numbers that are not
discussed in such courses. Thus one of the necessary tasks of the present course is
to give a more penetrating account of the real numbers than you have seen before.
1.2. Numbers of Various Kinds.
There are various kinds of numbers. Here is a list of the ones which will be
most important to us:
(1)

Z+ N Z Q R C.

Let me remind you what these various numbers are.

Z+ = {1, 2, 3, . . . , n, . . .}
is the set of positive integers (a.k.a. counting numbers). On them we have
defined the operations of addition + and multiplication . Moreover there is an
identity element for the multiplication, namely 1. There is no additive identity.
N = {0} Z+ = {0, 1, 2, 3, . . . , n, . . .}
is the set of natural numbers. Again we have operations of addition and multiplication, and now we have an additive identity as well as a multiplicative identity.
Clearly Z+ and N are very similar: they differ only as to whether 0 is included
or not. In analysis the subject we are beginning the study of here! the distinction between Z+ and N is not very important, and in fact most analysts I know
use N to denote the positive integers. I am probably showing my stripes as an
algebraically minded mathematician by making the distinction, but so be it.
Recall that the operation of subtraction is nothing else than the inverse operation of addition: in other words, to say a b = c is to say that a = b + c. However
the operation of subtraction is not everywhere defined on N: for instance, there is
a natural number 5 3, but no natural number 3 5.
Z = {. . . , 2, 1, 0, 1, 2, . . .}
is the set of integers. This is formed out of the natural numbers N by formally
allowing all subtractions: for instance, 17 is 0 17. Thus in Z every element n
has an additive inverse n. However, the same cannot be said of multiplicative
inverses. Recall that the operation of division is nothing else than the inverse operation of multiplication: to say a/b = c is to say that a = bc. However the operation
of division is not everywhere defined on Z: thre is an integer 6/3, but no integer 6/4.
a
| a, b Z, b 6= 0}.
b
is the set of rational numbers. This is formed out of the integers Z by formally
allowing all divisions by nonzero integers. Note that one subtlety here is that the
same rational number has many different expressions as the quotient of two intea
c
+
gers: for instance 46 = 32 = 3n
2n for any n Z . So we just need to agree that b = d
a
iff ad = bc. Alternately, any nonzero rational number has a unique expression b in
Q={

1. INTRODUCTION

15

lowest terms, i.e., with a and b not both divisible by any integer n > 1.2 Thus in
Q we have an additive identity 0, every element has an additive inverse, we have a
multiplicative identity 1, and every nonzero element has a multiplicative inverse.
What then are the real numbers R? The geometric answer is that the real numbers
correspond to points on the number line, but this does not make clear why there
are such points other than the rational numbers. An answer that one learns in high
school is that every real number has an infinite decimal expansion, not necessarily
terminating or repeating, and conversely any integer followed by an infinite decimal expansion determines a real number. In fact this is perfectly correct: it gives a
complete characterization of the real numbers, but it is not a cure-all: in order to
pursue the implications of this definition and even to really understand it one
needs tools that we will develop later in the course.
Finally, the complex numbers C are expressions of the form a + bi where a and
b are real numbers and i2 = 1. They are extremely important in mathematics
generally e.g. one needs them in order to solve polynomial equations but in this
course they will play at most a peripheral role.
Back to R: let us nail down the fact that there are real numbers which are not
rational. One way to see this is as follows: show that the decimal expansion of
every rational number is eventually periodic, and then exhibit a decimal expansion
which is not eventually periodic, e.g.
x = 0.16116111611116111116 . . .
where the number of 1s after each 6 increases by 1 each time. But this number x
reeks of contrivance: it seems to have been constructed only to make trouble. The
ancient Pythagoreans discovered a much more natural irrational real number.
Theorem 1.4. The square root of 2 is not a rational number.
Proof. The proof is the most famous (and surely one of the first) instances
of a certain important kind of argument, namely a proof by contradiction. We
assume that what we are trying to prove is false, and from that we reason until we
reach an absurd conclusion. Therefore what we are trying
to prove must be true.

2
is
rational:
then
Here goes: seeking a contradiction,
we
suppose

there are
integers a, b with b > 0 and 2 = ab . Since the defining property of 2 is that its
square is 2, there is really nothing to do but square both sides to get
2=

a2
.
b2

Clearing denominators, we get

2b2 = a2 .
This shows that the integer a2 is even, i.e., divisible by 2. It happens that for any
integer a, if a2 is even, then so is a: let us assume this for now; we can come back
to it later. Thus we may write a = 2A with A Z. Substituting, we get
2b2 = a2 = (2A)2 = 4A2 ,
2Like most of the statements we have made recently, this requires proof! Let me reiterate
that we are not giving proofs here or even careful definitions; rather, we are reminding the reader
of some of her mathematical past.

16

1. INTRODUCTION AND FOUNDATIONS

or
b2 = 2A2 .
Thus b2 is divisible by 2, so as above b = 2B for some B Z. Substituting, we get
4B 2 = (2B)2 = b2 = 2A2 ,
or
2B 2 = A2 .
Thus we are back where we started: assuming that 2b2 = a2 , we found that both
a and b were divisible by 2. This is suspect in the extreme, and we now have our
choice of killing blow. One ending is to observe that everything we have said above
applies to A and B: thus we must also have A = 2A1 , B = 2B2 , and so forth.
We can continue in this way factoring out as many powers of 2 from a and b as we
wish. But the only integer which is arbitrarily divisible by 2 is 0, so our conclusion
is a = b = 0, whereas we assumed b > 0: contradiction.
Alternately and perhaps more simply each rational number may be written
in lowest terms, so we could have assumed this about ab at the outset and, in particular,
that a and b are not both divisible by 2. Either way we get a contradiction,
so 2 must not be a rational number.

1.3. Why do we not do calculus on Q?
To paraphrase the title question, why do we want to use R to do calculus? Is
there something stopping us from doing calculus over, say, Q?
The answer to the second question is no: we can define limits, continuity, derivatives and so forth for functions f : Q Q exactly as is done for real functions. The
most routine results carry over with no change: it is still true, for instance, that
sums and products of continuous functions are continuous. However most of the
big theorems especially, the Interval Theorems become false over Q.
For a, b Q, let [a, b]Q = {x Q | a x b}.
Example: Consider the function f : [0, 2]Q Q given by f (x) = 1 if x2 <
2
and f (x) = 1 if x2 > 2. Note that we do not need to define f (x) at x = 2,
because by the result of the previous section these are not rational numbers. Then
f is continuous in fact it is differentiable and has identically zero derivative. But
f (0) = 1 < 0, f (2) = 1 > 0, and there is no c [0, 2]Q such that f (c) = 0. Thus
the Intermediate Value Theorem fails over Q.
Example: Consider the function: f : [0, 2]Q Q given by f (x) = x212 . Again, this

function is well-defined at all points of [0, 2]Q because 2 is not a rational number.
It is also a continuous function. However it is
not bounded above: by taking rational numbers which are arbitrarily close to 2, x2 2 becomes arbitrarily small
and thus f (x) becomes arbitarily large.3 In particular, f certainly does not attain
a maximum value. Thus the Extreme Value Theorem fails over Q.
Moreover, it can be shown (and will be later) that any function on a closed,
bounded interval which is either uniformly continuous or integrable is bounded, so
3We will be much more precise about this sort of thing later on. This is just an overview.

2. SOME PROPERTIES OF NUMBERS

17

the above function f is neither uniformly continuous nor integrable. If you have
had second semester
freshman calculus, you should think about why the analogous
function f : [0, 2] \ { 2} R
is not improperly Riemann integrable: it builds up
infinite area as we approach 2.
The point of these examples is in order to succeed in getting calculus off the ground,
we need to make use of some fundamental property of the real numbers not possessed by (for intance) the rational numbers. This property, which can be expressed
in various forms, is called completeness, and will play a major role in this course.
2. Some Properties of Numbers
2.1. Axioms for a Field.
In order to do mathematics in a rigorous way, one needs to identify a starting
point. Virtually all mathematical theorems are of the form A = B. That is,
assuming A, B must follow. For instance, in Euclidean geometry one lays down a
set of axioms and reasons only from them. The axioms needed for calculus are a
lot to swallow in one dose, so we will introduce them gradually. What we give here
is essentially a codification of high school algebra, including inequalities.
Specifically, we will give axioms that we want a number system to satisfy. At
this point we will take it for granted that in our number system we have operations
of addition, multiplication and an inequality relation <, and that there are distinguished numbers called 0 and 1. We require the following properties:
(P0) 0 6= 1.
(P1) (Commutativity of +): For all numbers x, y, x + y = y + x.
(P2) (Associativity of +): For all numbers x, y, z, (x + y) + z = x + (y + z).
(P3) (Identity for +): For all numbers x, x + 0 = x.
(P4) (Inverses for +): For all numbers x, there exists y with x + y = 0.
(P5) (Commutativity of ): For all numbers x, y, x y = y x.
(P6) (Associativity of ): For all numbers x, y, z (x y) z = x (y z).
(P7) (Identity for ): For all numbers x, x 1 = x.
(P8) (Inverses for ) For all numbers x 6= 0, there exists a number y with xy = 1.
(P9) (Distributivity of over +): For all numbers x, y, z, x (y + z) = (x y) + (x z).
Although it is not important for us now, the above axioms (P0) through (P9)
are called the field axioms, and a structure which satisfies them is called a field.
Example: Both Q and R satisfy all of the above field axioms. (We take this as
known information.)
Example: The complex numbers C satisfy all of the above field axioms. The only
one which is not straightforward is the existence of multiplicative inverses. For this:
if z = x + iy is a nonzero complex number i.e., the real numbers x and y are not
both zero then if w = xxiy
2 +y 2 , zw = 1.
Example: Let F2 = {0, 1} be a set consisting of two elements, 0 and 1. We define

18

1. INTRODUCTION AND FOUNDATIONS

0 + 0 = 0, 0 + 1 = 1 + 0 = 1, 1 + 1 = 0, 0 0 = 0 1 = 1 0 = 0, 1 1 = 1. Then F2
satisfies all of the above field axioms. It is sometimes called the binary field.
Proposition 1.5. In every system satisfying the field axioms, for every number
x we have x 0 = 0.
Proof. We have x 0 = x (0 + 0) = (x 0) + (x 0). Subtracting (x 0) from
both sides gives 0 = x 0.

Proposition 1.6. In every system satisfying the field axioms:
a) The only additive identity is 0.
b) Every number x has a unique additive inverse. If 1 denotes the additive inverse
of 1, then the additive inverse of x is (1) x.
c) The only multiplicative identity is 1.
b) Every nonzero number has a unique multiplicative inverse.
Proof. a) Note that 0 is an additive identity by (P3). Suppose that z is
another additive identity, and consider 0+z. Since 0 is an additive identity, 0+z = z.
Since z is an additive identity, 0 + z = 0. Thus z = 0.
b) Suppose y and z are both additive inverses to x: x + y = x + z = 0. Adding y
to both sides gives
y = 0 + y = (x + y) + y = (y + x) + y = y + (x + y)
= y + (x + z) = (y + x) + z = (x + y) + z = 0 + z = z,
so y = z. Moreover, for any number x,
(1) x + x = ((1) x) + (1 x) = (1 + 1) x = 0 x = 0.
c), d) The proofs of these are the same as the proofs of parts a) and b) but with
all instances of + replaced by and all instances of 0 replaced by 1.

Proposition 1.7. In every system satisfying the field axioms, (1)2 = 1.
Proof. By Proposition 1.6, 11 is the additive inverse of 1, namely 1.

Proposition 1.8. In every system satisfying the field axioms, if x 6= 0 and

y 6= 0 then xy 6= 0.
Proof. Seeking a contradiction, suppose that xy = 0. Since x 6= 0 it has a
multiplicative inverse x1 and then by Proposition 1.5,
0 = x1 xy = (x1 x)y = 1 y = y,
contradicting the assumption that y 6= 0.

Note that a logically equivalent formulation of Proposition 1.8 is: in any system
satisfying the field axioms, if xy = 0 then x = 0 or y = 0.
2.2. Axioms for an ordered field.
The remaining properties of numbers concern the inequality relation <. Instead
of describing the relation < directly, it turns out to be simpler to talk about the
properties of positive numbers. If we are given the inequality relation <, then we
say that x is positive if x > 0, thus knowing < we know which numbers are positive. Conversely, suppose that we have identified a subset P of numbers as positive.
Then we can define x < y if y x P. Now we want our set of positive numbers

2. SOME PROPERTIES OF NUMBERS

19

to satisfy the following properties.

(P10) (Trichotomy) For all numbers x, exactly one of the following holds: x = 0, x
is positive, x is positive.
(P11) (Closure under +) For all positive numbers x, y, x + y P.
(P12) (Closure under ) For all positive numbers x, y, x y P.
A number system satisfying (P1) through (P12) is called an ordered field.
Proposition 1.9. In every ordered field:
a) 1 > 0 and 1 < 0.
b) For every nonzero x, x2 > 0.
c) It follows that for all x, x2 0.
Proof. a) By (P0), 1 6= 0. Thus by trichotomy, either 1 is positive and 1
is negative, or 1 is positive and 1 is negative. But by (P12) the product of two
positive numbers is positive, so if 1 is positive and 1 is negative then 1 = (1)2
is positive, a contradiction. So it must be that 1 is positive and 1 is negative.
b) Since x is nonzero, either x > 0 or x > 0. If x > 0, then x2 = x x is the
product of two positive numbers, hence positive. If x < 0, then x > 0 and then
x2 = (1)2 x2 = (x) (x) is the product of two positive numbers, hence positive.
c) Since 02 = 0, part c) follows immediately from part b).

Example: The binary numbers F2 satisfy the field axioms (P0) through (P9), but
are they an ordered field? Well, not on the face of it because we have not been
given an inequality relation < satisfying (P10) through (P12). In fact we will now
show that there is no such relation. Indeed, in any ordered field, since 1 > 0, also
1 + 1 > 0, but in F2 1 + 1 = 0. In fancy language, F2 is a field which cannot be
endowed with the structure of an ordered field.
Example: The complex numbers C satisfy the field axioms (P0) through (P9),
but are they an ordered field? As above, we have not been given an inequality
relation. Also as above we can show that there is no such relation. For in the
complex numbers we have an element i with i2 = 1. But the ordered field axioms
imply both that 1 is negative and that any square is non-negative, contradiction.
Proposition 1.10. For any x, y, z in an ordered field:
a) x < 0 0 < x. (We say x is negative.)
b) The trichotomy property may be restated as: for any number x, exactly one of
the following holds: x is positive, x is zero, x is negative.
c) If x is positive, x1 is positive. If x is negative, x1 is negative.
d) If x is positive and y is negative, then xy is negative.
e) If x and y are both negative, then xy is positive.
Proof. a) By definition, x < 0 means 0 x = x is positive. Also 0 < x
means x 0 = x is positive. So there is nothing to show here.
b) No further argument for this is needed; we just state it for future reference.
c) Suppose x is positive. Certainly x1 is not zero, so we need to rule out the
possibility that its negative. But if it were, then by part a) 1
x would be positive
1
and thus by (P12) x x = 1 would be positive, contradicting Proposition 1.9a).
1
If x is negative then x is positive so by what we just showed x
= 1
x is positive,

20

1. INTRODUCTION AND FOUNDATIONS

and thus x1 = ( 1
x ) is negative.
d) Suppose x is positive and y is negative. In particular x and y are not zero,
so xy 6= 0. To show that xy is negative, by part b) it is enough to rule out the
possibility that xy is positive. Suppose it is. Then, by part c), since x is positive,
1
1
x is positive, and thus y = xy x would be positive: contradiction.
e) Suppose x and y are both negative. Then xy 6= 0, and we need to rule out the
possibility that xy is negative. Suppose it is. Then xy is positive, x1 is negative,

so by part d) y = xy x1 is negative and thus y is positive: contradiction.
a)
b)
c)
d)
e)

Proposition 1.11. For all a, b, c, d in an ordered field:

If a < b and c < d, a + c < b + d.
If a < b and c > 0, then ac < bc.
If a < b and c < 0, then ac > bc.
If 0 < a < b, then 0 < 1b < a1 .
If a > 0 and b > 0, then a < b a2 < b2 .

Proof. a) Since a < b and c < d, b a is positive and d c is positive, and

then by (P11) (b a) + (d c) = (b + d) (a + c) is positive, so b + d > a + c.
b) Since a < b, b a is positive. Since c is positive, by (P12) bc ac = (b c)a is
positive, so bc > ac.
c) Left to the reader.
1
1
d) We have a1 1b = (b a) ab
. The hypotheses imply that b a and ab
are both
positive, so by (P12) so is their product.
e) Note that b2 a2 = (b + a)(b a). Since a and b are both positive, b + a is
positive, and therefore b a is positive iff b2 a2 is positive.

2.3. Some further properties of Q and R.
As we have mentioned before, the ordered field axioms (P0) through (P12) are
just a list of some of the useful properties of Q and R. They are not a complete
set of axioms for either Q or R in other words, there are other properties these
fields have that cannot be logically deduced from these axioms alone. In fact this
is already clear because Q and R each satisfy all the ordered field axioms but are
essentially different structures: in Q the element 2 = 1 + 1 is not the square of
another element, but in R it is. Here we want to give some further familiar properties that do not hold for all ordered fields but both of which hold for R and one
of which holds for Q. (We are still far away from the fundamental completeness
axiom for R which is necessary to prove the Interval Theorems.)
The first axiom is called the Archimedean property: it says that for any positive
number x, there is a positive integer n such that x n. This clearly holds for R
according to our description of real numbers as integers followed by infinite decimal
expansions: a positive real number x is of the form
x = n0 .a1 a2 . . . an . . .
+

with n0 Z and ai {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}, and thus x is less than or equal to

the integer n0 + 1.4
4When I first typed this I wrote that x is less than n + 1. But actually this need not be
0
true! Can you think of an example? Beware: decimal expansions can be tricky.

2. SOME PROPERTIES OF NUMBERS

21

Since every positive real number is less than or equal to some integer, and every positive rational number is, in particular, a positive real number, then also
every positive rational number is less than or equal to some integer. That is, Q also
satisfies the Archimedean property. (Or, directly: any positive rational number
may be written in the form ab with a, b Z+ , and then ab a.)
This Archimedean property is so natural and familiar (not to mention useful...)
that the curious student may be well wonder: are there in fact systems of numbers
satisfying the ordered field axioms but not the Archimedean property?!? The answer is yes, there are plenty of them, and it is in fact possible to construct a theory
of calculus based upon them (in fact, such a theory is in many ways more faithful to
the calculus of Newton and Leibniz than the theory which we are presenting here,
which is a 19th century innovation). But we will not see such things in this course!
The next property does provide a basic difference between Q and R.
Theorem 1.12. Let x be a real number and n Z+ .

a) If n is odd, there is a unique real number y such that y n = x. We write y = n x.

b) If n is even and x ispositive, there is a unique positive real number y such that
y n = x. We write y = n x.
c) If n is even and x is negative, then there is no real number y with y n = x.
The first two parts of Theorem 1.12 rely on the Intermediate Value Theorem so are
not accessible to us at this time. (Thus we must guard aginst using the existence of
nth roots of real numbers in any of the theorems that lead up to the Intermediate
Value Theorem. In fact we will not use such things in the proof of any theorem,
but only as examples.) As a supplement tto part b), note that if n is even and
y is a positive real number such that y n = x, then there is exactly one other real
number with nth power equal to x: y. (You might try to prove this as an exercise.)
We can prove part c) now, since it holds in any ordered field. Indeed, if n is
even then n = 2k for an integer k, so if for some negative x we have y n = x then
x = y 2k = (y k )2 , contradicting Proposition 1.9c).
Here is a special case of Theorem 1.12 important enough to be recorded separately.
Corollary 1.13. A real number x is non-negative if and only if it is a square,
i.e., if and only if there exists a real number y with y 2 = x.
Note that Corollary 1.13 does not hold in the number system Q, since 2 = 1 + 1 is
positive but is not the square of any rational number.
Corollary 1.13 leads to a basic strategy for proving inequalities in the real numbers:
for x, y R, x y (y x) = z 2 for some real number z. In the next section
we will see some instances of this strategy in action.
2.4. Some Inequalities.
For an element x of an ordered field, we define the absolute value of x to be
x if x 0 and x if x < 0; it is denoted by |x|. Thus |x| 0 always and x = |x|.
Proposition 1.14. For any number x in an ordered field, x |x|.

22

1. INTRODUCTION AND FOUNDATIONS

Proof. If x 0 then x = |x|. If x < 0 then x < 0 < x = |x|, so x < |x|. 
Theorem 1.15. (Triangle Inequality) For all numbers x, y, |x + y| |x| + |y|.
Proof. Since |x| is defined to be x if x 0 and x if x < 0, it is natural to
break the proof into cases.
Case 1: x, y 0. Then |x + y| = x + y = |x| + |y|.
Case 2: x, y < 0. Then x + y < 0, so |x + y| = (x + y) = x y = |x| + |y|.
Case 3: x 0, y < 0. Now unfortunately we do not know whether |x + y| is
non-negative or negative, so we must consider consider further cases.
Case 3a: x + y 0. Then |x + y| = x + y |x| + |y|.
Case 3b: x + y < 0. Then |x + y| = x y | x| + | y| = |x| + |y|.
Case 4: x < 0, y 0. The argument is exactly the same as that in Case 3. In fact,
we can guarantee it is the same: since the desired inequality is symmetric in x
and y meaning, if we interchange x and y we do not change what we are trying
to show we may reduce to Case 3 by interchanging x and y.5

The preceding argument is definitely the sort that one should be prepared to make
when dealing with expressions involving absolute values. However, it is certainly
not very much fun. Spivak gives an alternate proof of the Triangle Inequality
which is more interesting and thematic. First, since both quantities |x + y| and
|x| + |y| are non-negative, the inequality will hold iff it holds after squaring both
sides (Proposition 1.11e). So it is enough to show
(|x + y|)2 (|x| + |y|)2 .
now (|x+y|)2 = (x+y)2 = x2 +2xy +y 2 , whereas (|x|+|y|)2 = |x|2 +2|x||y|+|y|2 =
x2 + |2xy| + y 2 , so subtracting the left hand side from the right, it is equivalent to
show that
0 (x2 + |2xy| + y 2 ) (x2 + 2xy + y 2 ).
But
(x2 + |2xy| + y 2 (x2 + 2xy + y 2 ) = |2xy| 2xy 0
by Proposition 1.14. So this gives a second proof of the Triangle Inequality.
A similar argument can be used to establish the following variant.
Proposition 1.16. (Reverse Triangle Inequality)
For all numbers x, y, ||x| |y|| |x y|.
Proof. Again, since both quantities are non-negative, it is sufficient to prove
the inequality after squaring both sides:
(||x| |y||)2 = (|x| |y|)2 = |x|2 2|x||y| + |y|2 = x2 |2xy| + y 2
x2 2xy + y 2 = (x y)2 = (|x y|)2 .

Exercise: Let x, y be any numbers.
a) Show that |x| |y| |x y| by writing x = (x y) + y and applying the usual
triangle inequality.
b) Deduce from part a) that ||x| |y|| |x y|.
5Such symmetry arguments can often by used to reduce the number of cases considered.

2. SOME PROPERTIES OF NUMBERS

23

Theorem 1.17. (Cauchy-Bunyakovsky-Schwarz Inequality, n = 2)

a) For all numbers x1 , x2 , y1 , y2 ,
(2)

(x1 y1 + x2 y2 )2 (x21 + x22 )(y12 + y22 ).

b) Moreover equality holds in (2) iff y1 = y2 = 0 or there exists a number such

that x1 = y1 and x2 = y2 .
Proof. By pure brute force, one can prove the two squares identity:
(x21 + x22 )(y12 + y22 ) = (x1 y2 x2 y1 )2 + (x1 y1 + x2 y2 )2 .
Now we need only rewrite it in the form
(x21 + x22 )(y12 + y22 ) (x1 y1 + x2 y2 )2 = (x1 y2 x2 y1 )2 0,
establishing part a). Moreover, equality holds iff x1 y2 = x2 y1 . If in this equality
y1 and y2 are both nonzero, we may divide by them to get xy11 = xy22 = . If y1 = 0
and y2 6= 0 then we must have x1 = 0 and then we may take = xy22 . Similarly,
if y1 6= 0 and y2 = 0, then we must have x2 = 0 and then we may take = xy11 .
Finally, if y1 = y2 = 0 then the equality x1 y2 = x2 y1 also holds.

Theorem 1.18. (Cauchy-Bunyakovsky-Schwarz Inequality) For any n Z+
and numbers x1 , . . . , xn , y1 , . . . , yn we have
(x1 y1 + . . . + xn yn )2 (x21 + . . . + x2n )(y12 + . . . + yn2 ).
Proof. Expanding out the right and left hand sides, we get
n
X
X
x2i yi2 +
x2i yj2 .
RHS =
i=1

LHS =

n
X

i6=j

x2i yi2 + 2

i=1

xi yi xj yj ,

i<j

so
RHS LHS =

X
i6=j

x2i yj2 2

X
i<j

xi yj xj yi =

X
(xi yj xj yi )2 0.
i<j


Theorem 1.19. (Arithmetic-Geometric Mean Inequality, n = 2)
For all numbers 0 < a < b, we have

2
a+b
2
a < ab <
< b2 .
2
Proof. First inequality: Since a > 0 and 0 < a < b, a a < a b.
Second inequality: Expanding out the square and clearing denominators, it is equivalent to 4ab < a2 + 2ab + ab2 , or to a2 2b + b2 > 0. But a2 2ab + b2 = (a b)2 ,
so since a 6= b, (a b)2 > 0.
a+b
Third inequality: Since a+b
2 and b are both positive, it is equivalent to 2 < b and
thus to a + b < 2b. But indeed since a < b, a + b < b + b = 2b.

Later we will use the theory of convexity to prove a signficant generalization of
Theorem 1.19, the Weighted Arithmetic-Geometric Mean Inequality.

CHAPTER 2

Mathematical Induction
1. Introduction
Principle of Mathematical Induction for sets
Let S be a subset of the positive integers. Suppose that:
(i) 1 S, and
(ii) n Z+ , n S = n + 1 S.
Then S = Z+ .
The intuitive justification is as follows: by (i), we know that 1 S. Now apply (ii) with n = 1: since 1 S, we deduce 1 + 1 = 2 S. Now apply (ii) with
n = 2: since 2 S, we deduce 2 + 1 = 3 S. Now apply (ii) with n = 3: since
3 S, we deduce 3 + 1 = 4 S. And so forth.
This is not a proof. (No good proof uses and so forth to gloss over a key point!)
But the idea is as follows: we can keep iterating the above argument as many times
as we want, deducing at each stage that since S contains the natural number which
is one greater than the last natural number we showed that it contained. Now it
is a fundamental part of the structure of the positive integers that every positive
integer can be reached in this way, i.e., starting from 1 and adding 1 sufficiently
many times. In other words, any rigorous definition of the natural numbers (for
instance in terms of sets, as alluded to earlier in the course) needs to incorporate,
either implicitly or (more often) explicitly, the principle of mathematical induction.
Alternately, the principle of mathematical induction is a key ingredient in any axiomatic characterization of the natural numbers.
It is not a key point, but it is somewhat interesting, so let us be a bit more specific.
In Euclidean geometry one studies points, lines, planes and so forth, but one does
not start by saying what sort of object the Euclidean plane really is. (At least
this is how Euclidean geometry has been approached for more than a hundred years.
Euclid himself gave such definitions as: A point is that which has position but
not dimensions. A line is breadth without depth. In the 19th century it was
recognized that these are descriptions rather than definitions, in the same way that
many dictionary definitions are actually descriptions: cat: A small carnivorous
mammal domesticated since early times as a catcher of rats and mice and as a pet
and existing in several distinctive breeds and varieties. This helps you if you are
already familiar with the animal but not the word, but if you have never seen a cat
before this definition would not allow you to determine with certainty whether any
particular animal you encountered was a cat, and still less would it allow you to
reason abstractly about the cat concept or prove theorems about cats.) Rather
25

26

2. MATHEMATICAL INDUCTION

point, line, plane and so forth are taken as undefined terms. They are
related by certain axioms: abstract properties they must satisfy.
In 1889, the Italian mathematician and proto-logician Gisueppe Peano came up
with a similar (and, in fact, much simpler) system of axioms for the natural numbers. In slightly modernized form, this goes as follows:
The undefined terms are zero, number and successor.
There are five axioms that they must satisfy, the Peano axioms. The first four
are:
(P1) Zero is a number.
(P2) Every number has a successor, which is also a number.
(P3) No two distinct numbers have the same successor.
(P4) Zero is not the successor of any number.
Using set-theoretic language we can clarify what is going on here as follows: the
structures we are considering are triples (X, 0, S), where X is a set, 0 is an element
of X, and S : X X is a function, subject to the above axioms.
From this we can deduce quite a bit. First, we have a number (i.e., an element
of X) called S(0). Is 0 = S(0)? No, that is prohibited by (P4). We also have a
number S(S(0)), which is not equal to 0 by (P4) and it is also not equal to S(0),
because then S(0) = S(S(0)) would be the successor of the distinct numbers 0
and S(0), contradicting (P3). Continuing in this way, we can produce an infinite
sequence of distinct elements of X:
(3)

0, S(0), S(S(0)), S(S(S(0)), . . . .

In particular X itself is infinite. The crux of the matter is this: is there any element
of X which is not a member of the sequence (3), i.e., is not obtained by starting at
0 and applying the successor function finitely many times?
The axioms so far do not allow us to answer this question. For instance, suppose
that the numbers consisted of the set [0, ) of all non-negative real numbers, we
define 0 to be the real number of that name, and we define the successor of x to be
x + 1. This system satisfies (P1) through (P4) but has much more in it than just
the natural numbers we want, so we must be missing an axiom! Indeed, the last
axiom is:
(P5) If Y is a subset of the set X of numbers such that 0 Y and such that
x Y implies S(x) Y , then Y = X.
Notice that the example we cooked up above fails (P5), since in [0, ) the subset
of natural numbers contains zero and contains the successor of each of its elements
but is a proper subset of [0, ).
Thus it was Peanos contribution to realize that mathematical induction is an axiom for the natural numbers in much the same way that the parallel postulate is
an axiom for Euclidean geometry.

2. THE FIRST INDUCTION PROOFS

27

On the other hand, it is telling that this work of Peano is little more than one
hundred years old, which in the scope of mathematical history is quite recent.
Traces of what we now recognize as induction can be found from the mathematics
of antiquity (including Euclids Elements!) on forward. According to the (highly
recommended!) Wikipedia article on mathematical induction, the first mathematician to formulate it explicitly was Blaise Pascal, in 1665. During the next hundred
years various equivalent versions were used by different mathematicians notably
the methods of infinite descent and minimal counterexample, which we shall discuss later and the technique seems to have become commonplace by the end of
the 18th century. Not having an formal understanding of the relationship between
mathematical induction and the structure of the natural numbers was not much
of a hindrance to mathematicians of the time, so still less should it stop us from
learning to use induction as a proof technique.
Principle of mathematical induction for predicates
Let P (x) be a sentence whose domain is the positive integers. Suppose that:
(i) P (1) is true, and
(ii) For all n Z+ , P (n) is true = P (n + 1) is true.
Then P (n) is true for all positive integers n.
Variant 1: Suppose instead that P (x) is a sentence whose domain is the natural numbers, i.e., with zero included, and in the above principle we replace (i) by
the assumption that P (0) is true and keep the assumption (ii). Then of course the
conclusion is that P (n) is true for all natural numbers n. This is more in accordance
with the discussion of the Peano axioms above.1
Exercise 1: Suppose that N0 is a fixed integer. Let P (x) be a sentence whose
domain contains the set of all integers n N0 . Suppose that:
(i) P (N0 ) is true, and
(ii) For all n N0 , P (n) is true = P (n + 1) is true.
Show that P (n) is true for all integers n N0 . (Hint: define a new predicate Q(n)
with domain Z+ by making a change of variables in P .)
2. The First Induction Proofs
2.1. The Pedagogically First Induction Proof.
There are many things that one can prove by induction, but the first thing that
everyone proves by induction is invariably the following result.
Proposition 2.1. For all n Z+ , 1 + . . . + n =

n(n+1)
.
2

Proof. We go by induction on n.
Base case (n = 1): Indeed 1 = 1(1+1)
.
2
+
Induction step: Let n Z and suppose that 1 + . . . + n =
IH

1 + . . . + n + n + 1 = (1 + . . . + n) + n + 1 =

n(n+1)
.
2

Then

n(n + 1)
+n+1
2

1In fact Peanos original axiomatization did not include zero. What we presented above is a

standard modern modification which is slightly cleaner to work with.

28

2. MATHEMATICAL INDUCTION

n2 + n 2n + 2
n2 + 2n + 3
(n + 1)(n + 2)
(n + 1)((n + 1) + 1)
+
=
=
=
.
2
2
2
2
2
Here the letters IH signify that the induction hypothesis was used.

=

Induction is such a powerful tool that once one learns how to use it one can prove
many nontrivial facts with essentially no thought or ideas required, as is the case in
the above proof. However thought and ideas are good things when you have them!
In many cases an inductive proof of a result is a sort of first assault which raises
the challenge of a more insightful, noninductive proof. This is certainly the case
for Proposition 2.1 above, which can be proved in many ways.
Here is one non-inductive proof: replacing n by n 1, it is equivalent to show:
(n 1)n
.
(4)
n Z, n 2 : 1 + . . . + n 1 =
2


We recognize the quantity on the right-hand side as the binomial coefficient n2 :
it counts the number of 2-element subsets of an n element set. This raises the
prospect of a combinatorial proof, i.e., to show that the number of 2-element
subsets of an n element set is also equal to 1 + 2 + . . . + n 1. This comes out
immediately if we list the 2-element subsets of {1, 2, . . . , n} in a systematic way:
we may write each such subset as {i, j} with 1 i n 1 and i < j n. Then:
The subsets with least element 1 are {1, 2}, {1, 3}, . . . , {1, n}, a total of n 1.
The subsets with least element 2 are {2, 3}, {2, 4}, . . . , {2, n}, a total of n 2.
..
.
The subset with least element n 1 is {n 1, n}, a total of 1.


Thus the number of 2-element subsets of {1, . . . , n} is on the one hand n2 and
on the other hand (n 1) + (n 2) + . . . + 1 = 1 + 2 + . . . + n 1. This gives a
combinatorial proof of Proposition 2.1.
2.2. The (Historically) First(?) Induction Proof.
Theorem 2.2. (Euclid) There are infinitely many prime numbers.
Proof. For n Z+ , let P (n) be the assertion that there are at least n prime
numbers. Then there are infinitely many primes if and only if P (n) holds for all
positive integers n. We will prove the latter by induction on n.
Base Case (n = 1): We need to show that there is at least one prime number. For
instance, 2 is a prime number.
Induction Step: Let n Z+ , and assume that P (n) holds, i.e., that there are at
least n prime numbers p1 < . . . < pn . We need to show that P (n + 1) holds, i.e.,
there is at least one prime number different from the numbers we have already
found. To establish this, consider the quantity
Nn = p1 pn + 1.
Since p1 pn p1 2, Nn 3. In particular it is divisible by at least one prime
number, say q.2 But I claim that Nn is not divisible by pi for any 1 i n. Indeed,
n
Z. Then kpi = p1 pn + 1 =
if Nn = api for some a Z, then let b = p1 p
pi
2Later in these notes we will prove the stronger fact that any integer greater than one may
be expressed as a product of primes. For now we assume this (familiar) fact.

3. CLOSED FORM IDENTITIES

29

bpi + 1, so (k b)pi = 1 and thus pi = 1, a contradiction. So if we take q to be,

for instance, the smallest prime divisor of Nn , then there are at least n + 1 prime
numbers: p1 , . . . , pn , q.

The proof that there are infinitely many prime numbers first appeared in Euclids
Elements (Book IX, Proposition 20). Euclid did not explicitly use induction, but in
retrospect his proof is clearly an inductive argument: what he does is to explain, as
above, how given any finite list p1 , . . . , pn of distinct primes, one can produce a new
prime which is not on the list. (Euclid does not verify the base case; he must have
regarded it as obvious that there is at least one prime number.) What is strange is
that in our day Euclids proof is generally not seen as a proof by induction. Rather,
it is often construed as a proof by contradiction which it isnt! Rather, Euclids
argument is perfectly constructive. Starting with any given prime number say
p1 = 2 and following his procedure, one generates an infinite sequence of primes.
For instance, N1 = 2 + 1 = 3 is prime, so we take p2 = 3. Then N2 = 2 3 + 1 = 7
is again prime, so we take p3 = 7. Then N3 = 2 3 7 + 1 = 43 is also prime, so we
take p4 = 43. But this time something more interesting happens:
N4 = 2 3 7 43 + 1 = 13 139
is not prime. For definiteness let us take p5 to be the smallest prime factor of N4 ,
so p5 = 13. In this way we generate an infinite sequence of prime numbers so the
proof is unassailably constructive.
By the way, this sequence of prime numbers is itself rather interesting. It
is often called the Euclid-Mullin sequence, after Albert A. Mullin who asked
questions about it in 1963 [Mu63]. The next few terms are
53, 5, 6221671, 38709183810571, 139, 2801, 11, 17, 5471, 52662739, 23003,
30693651606209, 37, 1741, 1313797957, 887, 71, 7127, 109, 23, . . . .
Thus one can see that it is rather far from just giving us all of the prime numbers
in increasing order! In fact, since to find pn+1 we need to factor Nn = p1 pn + 1,
a quantity which rapidly increases with n, it is in fact quite difficult to compute the
terms of this sequence, and as of 2010 only the first 47 terms are known. Perhaps
Mullins most interesting question about this sequence is: does every prime number appear in it eventually? This is an absolutely open question. At the moment
the smallest prime which is not known to appear in the Euclid-Mullin sequence is 31.
Remark: Some scholars have suggested that what is essentially an argument by
mathematical induction appears in the later middle Platonic dialogue Parmenides,
lines 149a7-c3. But this argument is of mostly historical and philosophical interest.
The statement in question is, very roughly, that if n objects are placed adjacent
to another in a linear fashion, the number of points of contact between them is
n 1. (Maybe. To quote wikipedia: It is widely considered to be one of the
more, if not the most, challenging and enigmatic of Platos dialogues.) There is
not much mathematics here! Nevertheless, for a thorough discussion of induction
in the Parmenides the reader may consult [Ac00] and the references cited therein.
3. Closed Form Identities
The inductive proof of Proposition 2.1 is a prototype for a certain kind of
induction proof (the easiest kind!) in which P (n) is some algebraic identity: say

30

2. MATHEMATICAL INDUCTION

LHS(n) = RHS(n). In this case to make the induction proof work you need only
(i) establish the base case and (ii) verify the equality of successive differences
LHS(n + 1) LHS(n) = RHS(n + 1) RHS(n).
We give two more familiar examples of this.
Proposition 2.3. For all n Z+ , 1 + 3 + . . . + (2n 1) = n2 .
Proof. Let P (n) be the statement 1 + 3 + . . . + (2n 1) = n2 . We will show
P (n) holds for all n Z+ by induction on n. Base case (n = 1): indeed 1 = 12 .
Induction step: Let n be an arbitrary positive integer and assume P (n):
(5)

1 + 3 + . . . + (2n 1) = n2 .

Adding 2(n + 1) 1 = 2n + 1 to both sides, we get

(1 + 3 + . . . + (2n 1) + 2(n + 1) 1 = n2 + 2(n + 1) 1 = n2 + 2n + 1 = (n + 1)2 .
This is precisely P (n + 1), so the induction step is complete.
Proposition 2.4. For all n Z+ , 12 + 22 + . . . + n2 =


n(n+1)(2n+1)
.
6

Proof. By induction on n.
Base case: n = 1.
Induction step: Let n Z+ and suppose that 12 + . . . + n2 =
IH

1 + . . . + n2 + (n + 1)2 =

n(n+1)(2n+1)
.
6

Then

n(n + 1)(2n + 1)
+ (n + 1)2 =
6

2n3 + 3n2 + n + 6 + 6n2 + 12n + 1

2n3 + 9n2 + 13n + 7
=
.
6
6
3
2
On the other hand, expanding out (n+1)((n+1)+1)(2(n+1)+1)
, we also get 2n +9n 6+13n+7 .
6

Often a non-inductive proof, when available, offers more insight. Again returning
to our archetypical example: 1 + . . . + n, it is time to tell the story of little Gauss.
As a child of no more than 10 or so, Gauss and his classmates were asked to add up
the numbers from 1 to 100. Most of the students did this by a laborious calculation
and got incorrect answers in the end. Gauss reasoned essentially as follows: put
Sn = 1 + . . . + (n 1) + n.
Of course the sum is unchanged if we we write the terms in descending order:
Sn = n + (n 1) + . . . + 2 + 1.
Adding the two equations gives
2Sn = (n + 1) + (n + 1) + . . . + (n + 1) = n(n + 1),
so
n(n + 1)
.
2
This is no doubt preferable to induction, so long as one is clever enough to see it.
Sn =

Mathematical induction can be viewed as a particular incarnation of a much more

general proof technique: try to solve your problem by reducing it to a previously

4. INEQUALITIES

31

solved problem. A more straightforward application of this philosophy allows us to

deduce Proposition 2.3 from Proposition 2.1:


n
n
n
X
X
X
n(n + 1)
1+3+. . .+(2n1) =
(2i1) = 2
n = n2 +nn = n2 .
i
1=2
2
i=1
i=1
i=1
4. Inequalities
Proposition 2.5. For all n N , 2n > n.
proof analysis For n N, let P (n) be the statement 2n > n. We want to
show that P (n) holds for all natural numbers n by induction.
Base case: n = 0: 20 = 1 > 0.
Induction step: let n be any natural number and assume P (n): 2n > n. Then
2n+1 = 2 2n > 2 n.
We would now like to say that 2n n + 1. But in fact this is true if and only
if n 1. Well, dont panic. We just need to restructure the argument a bit: we
verify the statement separately for n = 0 and then use n = 1 as the base case of
our induction argument. Here is a formal writeup:
Proof. Since 20 = 1 > 0 and 21 = 2 > 1, it suffices to verify the statement
for all natural numbers n 2. We go by induction on n.
Base case: n = 2: 22 = 4 > 2.
Induction step: Assume that for some natural number n 2, 2n > n. Then
2n+1 = 2 2n > 2n > n + 1.

Proposition 2.6. There exists N0 Z+ such that for all n N0 , 2n n3 .
Proof analysis A little experimentation shows that there are several small values
of n such that 2n < n3 : for instance 29 = 512 < 93 = 729. On the other hand, it
seems to be the case that we can take N0 = 10: lets try.
Base case: n = 10: 210 = 1024 > 1000 = 103 .
Induction step: Suppose that for some n 10 we have 2n n3 . Then
2n+1 = 2 2n 2n3 .
Our task is then to show that 2n3 (n + 1)3 for all n 10. (By considering limits
as n , it is certainly the case that the left hand side exceeds the right hand
side for all sufficiently large n. Its not guaranteed to work for n 10; if not, we
will replace 10 with some larger number.) Now,
2n3 (n + 1)3 = 2n3 n3 3n2 3n 1 = n3 3n2 3n 1 0
n3 3n2 3n 1.
Since everything in sight is a whole number, this is in turn equivalent to
n3 3n2 3n > 0.
Now n3 3n2 3n = n(n2 3n 3), so this is equivalent
to n2 3n 3 0.

3 21
2
, so n2 3n 3 > 0 if
The roots of the polynomial x 3x 3 are x =
2

32

2. MATHEMATICAL INDUCTION

n > 4 = 3+2 25 > 3+2 21 . In particular, the desired inequality holds if n 10, so
by induction we have shown that 2n n3 for all n 10.
We leave it to to the student to convert the above analysis into a formal proof.
Remark: More precisely, 2n n3 for all natural numbers n except n = 2, 3, 4, 6, 7, 8, 9.
It is interesting that the desired inequality is true for a little while (i.e., at n = 0, 1)
then becomes false for a little while longer, and then becomes true for all n 10.
Note that it follows from our analysis that if for any N 4 we have 2N N 3 , then
this equality remains true for all larger natural numbers n. Thus from the fact that
29 < 93 , we can in fact deduce that 2n < n3 for all 4 n 8.
Proposition 2.7. For all n Z+ , 1 +

1
4

+ ... +

proof analysis By induction on n.

Base case (n = 1): 1 2 11 .
Induction step: Assume that for some n Z+ , 1 +
1+

1
4

1
n2

2 n1 .

+ ... +

1
n2

2 n1 . Then

1
1
1
1
1
+ ... + 2 +
2 +
.
4
n
(n + 1)2
n (n + 1)2

We want the left hand side to be less than 2

1
n+1 ,

so it will suffice to show

1
1
1
+
<2
.
2
n (n + 1)
n+1
Equivalently, it suffices to show
1
1
1
+
< .
n + 1 (n + 1)2
n
But we have
1
1
n+1+1
n+2
+
=
=
.
n + 1 (n + 1)2
(n + 1)2
(n + 1)2
Everything in sight is positive, so by clearing denominators, the desired inequality
is equivalent to
2

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

which is true! Thus we have all the ingredients of an induction proof, but again we
need to put things together in proper order, a task which we leave to the reader.
P
Remark: Taking limits as n , it follows that n=1 n12 2. In particular,
this argument shows that the infinite series converges. The exact value of the sum
2
is, in fact, 6 . A proof of this requires techniques from advanced calculus.
5. Extending Binary Properties to n-ary Properties
Example: All horses have the same color.
Proposed proof: There are only finitely many horses in the world, so it will suffice
to show that for all n Z+ , P (n) holds, where P (n) is the statement that in any
set of n horses, all of them have the same color.
Base case: In any set S of one horse, all of the horses in S have the same color!
Induction step: We suppose that for some positive integer n, in any set of n horses,

5. EXTENDING BINARY PROPERTIES TO n-ARY PROPERTIES

33

all have the same color. Consider now a set of n + 1 horses, which for specificity
we label H1 , H2 , . . . , Hn , Hn+1 . Now we can split this into two sets of n horses:
S = {H1 , . . . , Hn }
and
T = {H2 , . . . , Hn , Hn+1 }.
By induction, every horse in S has the same color as H1 : in particular Hn has
the same color as H1 . Similarly, every horse in T has the same color as Hn : in
particular Hn+1 has the same color as Hn . But this means that H2 , . . . , Hn , Hn+1
all have the same color as H1 . It follows by induction that for all n Z+ , in any
set of n horses, all have the same color.
Proof analysis: Naturally one suspects that there is a mistake somewhere, and
there is. However it is subtle, and occurs in a perhaps unexpected place. In fact the
argument is completely correct, except the induction step is not valid when n = 1:
in this case S = {H1 } and T = {H2 } and these two sets are disjoint: they have
no horses in common. We have been misled by the dot dot dot notation which
suggests, erroneously, that S and T must have more than one element.
In fact, if only we could establish the argument for n = 2, then the proof goes
through just fine. For instance, the result can be fixed as follows: if in a finite set
of horses, any two have the same color, then they all have the same color.
There is a moral here: one should pay especially close attention to the smallest
values of n to make sure that the argument has no gaps. On the other hand,
there is a certain type of induction proof for which the n = 2 case is the most
important (often it is also the base case, but not always), and the induction step
is easy to show, but uses once again the n = 2 case. Here are some examples of this.
The following is a fundamental fact of number theory, called Euclids Lemma.
Proposition 2.8. Let p be a prime, and let a, b Z+ . If p | ab, p | a or p | b.
Later in this chapter we will give a proof (yes, by induction!). Lets assume it for
now. Then we can swiftly deduce the following useful generalization.
Proposition 2.9. Let p be a prime number, n Z+ and a1 , . . . , an Z+ . If
p | a1 an , then p | ai for some 1 i n.
Proof. This is trivial for n = 1. We show it for all n 2 by induction.
Base case: n = 2: This is precisely Euclids Lemma.
Induction Step: We assume that for a given n Z+ and a1 , . . . , an Z+ , if a prime
p divides the product a1 an , then it divides at least one ai . Let a1 , . . . , an , an+1
Z, and that a prime p divides a1 an an+1 . Then p | (a1 an )an+1 , so by Euclids
Lemma, p | a1 an or p | an+1 . If the latter, were done. If the former, then by
our inductive hypothesis, p | ai for some 1 i n, so we are also done.

Corollary 2.10. Let p be a prime, and let a, n Z+ . Then p | an = p | a.
1

Exercise 5: Use Corollary 2.10 to show that for any prime p, p n is irrational.

34

2. MATHEMATICAL INDUCTION

6. The Principle of Strong/Complete Induction

Problem: A sequence is defined recursively by a1 = 1, a2 = 2 and an = 3an1
2an2 . Find a general formula for an and prove it by induction.
Proof analysis: Unless we know something better, we may as well examine the
first few terms of the sequence and hope that a pattern jumps out at us. We have
a3 = 3a2 2a1 = 3 2 2 1 = 4.
a4 = 3a3 2a2 = 3 4 2 2 = 8.
a5 = 3a4 2a3 = 3 8 2 4 = 16.
a6 = 3a5 2a4 = 3 16 2 8 = 32.
The evident guess is therefore an = 2n1 . Now a key point: it is not possible to
prove this formula using the version of mathematical induction we currently have.
Indeed, lets try: assume that an = 2n1 . Then
an+1 = 3an 2an1 .
By the induction hypothesis we can replace an with 2n1 , getting
an+1 = 3 2n1 2an1 ;
now what?? A little bit of thought indicates that we think an1 = 2n2 . If for
some reason it were logically permissible to make that substitution, then wed be
in good shape:
an+1 = 3 2n1 2 2n2 = 3 2n1 2n1 = 2 2n1 = 2n = 2(n+1)1 ,
which is what we wanted to show. Evidently this goes a bit beyond the type of
induction we have seen so far: in addition to assuming the truth of a statement
P (n) and using it to prove P (n + 1), we also want to assume the truth of P (n 1).
There is a version of induction that allows this, and more:
Principle of Strong/Complete Induction:
Let P (n) be a sentence with domain the positive integers. Suppose:
(i) P (1) is true, and
(ii) For all n Z+ , if P (1), . . . , P (n 1), P (n) are all true, then P (n + 1) is true.
Then P (n) is true for all n Z+ .
Thus, in a nutshell, strong/complete induction allows us to assume not only the
truth of our statement for a single value of n in order to prove it for the next value
n + 1, but rather allows us to assume the truth of the statement for all positive
integer values less than n + 1 in order to prove it for n + 1.
It is easy to see that PS/CI implies the usual principle of mathematical induction. The logical form of this is simply3
(A = C) = (A B = C).
3The symbol denotes logical conjunction: in other words, A B means A and B.

7. SOLVING HOMOGENEOUS LINEAR RECURRENCES

35

In other words, if one can deduce statement C from statement A, then one can
also deduce statement C from A together with some additional hypothesis or hypotheses B. Specifically, we can take A to be P (n), C to be P (n + 1) and B to be
P (1) P (2) . . . P (n 1).
Less obviously, one can use our previous PMI to prove PS/CI. The proof is not
hard but slightly tricky. Suppose we know PMI and wish to prove PS/CI. Let P (n)
be a sentence with domain the positive integers and satisfying (i) and (ii) above.
We wish to show that P (n) holds for all n Z+ , using only ordinary induction.
The trick is to introduce a new predicate Q(n), namely
Q(n) = P (1) P (2) . . . P (n).
Notice that Q(1) = P (1); (ii) above tells us that Q(n) = P (n + 1). But if we
know Q(n) = P (1) . . . P (n) and also P (n + 1), then we know P (1) . . . P (n)
P (n + 1) = Q(n + 1). So Q(1) holds and for all n, Q(n) = Q(n + 1). So by PMI,
Q(n) holds for all n, hence certainly P (n) holds for all n.
Exercise 6: As for ordinary induction, there is a variant of strong/complete induction where instead of starting at 1 we start at any integer N0 . State this explicitly.
Here is an application which makes full use of the strength of PS/CI.
Proposition 2.11. Let n > 1 be an integer. Then there exist prime numbers
p1 , . . . , pk (for some k 1) such that n = p1 pk .
Proof. We go by strong induction on n.
Base case: n = 2. Indeed 2 is prime, so were good.
Induction step: Let n > 2 be any integer and assume that the statement is true for
all integers 2 k < n. We wish to show that it is true for n.
Case 1: n is prime. As above, were good.
Case 2: n is not prime. By definition, this means that there exist integers a, b, with
1 < a, b < n, such that n = ab. But now our induction hypothesis applies to both
a and b: we can write a = p1 pk and b = q1 ql , where the pi s and qj s are all
prime numbers. But then
n = ab = p1 pk q1 ql
is an expression of n as a product of prime numbers: done!

This is a good example of the use of induction (of one kind or another) to give a
very clean proof of a result whose truth was not really in doubt but for which a
more straightforward proof is wordier and messier.
7. Solving Homogeneous Linear Recurrences
Recall our motivating problem for PS/CI: we were given a sequence defined by
a1 = 1, a2 = 2, and for all n 1, an = 3an1 2an2 . By trial and error we
guessed that an = 2n1 , and this was easily confirmed using PS/CI.
But this was very lucky (or worse: the example was constructed so as to be easy
to solve). In general, it might not be so obvious what the answer is, and as above,
this is inductions Kryptonite: it has no help to offer in guessing the answer.

36

2. MATHEMATICAL INDUCTION

Example: Suppose a sequence is defined by x0 = 2, xn = 5xn1 3 for all n 1.

Here the first few terms of the sequence are x1 = 7, x2 = 32, x3 = 157, x4 = 782,
x5 = 3907. Whats the pattern? At least to me, its not evident.
This is a case where more generality brings clarity: it is often easier to detect
a pattern involving algebraic expressions than a pattern involving integers. So
suppose we have a, b, c R, and we define a sequence recursively by
x0 = c; n N, xn+1 = axn + b.
Now lets try again:
x1 = ax0 + b = ac + b.
x2 = ax1 + b = a(ac + b) + b = ca2 + ba + b.
x3 = ax2 + b = a(ca2 + ba + b) + b = ca3 + ba2 + ba + b.
x4 = ax3 + b = a(ca3 + ba2 + ba + b) + b = ca4 + ba3 + ba2 + ba + b.
Aha: it seems that we have for all n 1.
xn = can + ban1 + . . . + ba + b.
Now we have something that induction can help us with: it is true for n = 1.
Assuming it is true for n, we calculate
IH

xn+1 = axn +b = a(can +ban1 +. . .+ba+b)+b) = can+1 +ban + +ba2 +ba+b,

so the desired expression is correct for all n. Indeed, we can simplify it:

 n+1
n
X
a
1
(ab + ac c)an b
n
n
xn = ca + b
ai = ca + b
=
.
a1
a1
i=1
In particular the sequence xn grows exponentially in n.
Let us try our hand on a sequence defined by a two-term recurrence:
F1 = F2 = 1; n 1, Fn+2 = Fn+1 + Fn .
The Fn s are the famous Fibonacci numbers. Again we list some values:
F3 = 2, F4 = 3, F5 = 5, F6 = 8, F7 = 13, F8 = 21, F9 = 34, F10 = 55,
F11 = 89, F12 = 144, F13 = 233, F14 = 377, F15 = 377,
F200 = 280571172992510140037611932413038677189525,
F201 = 453973694165307953197296969697410619233826.
These computations suggest Fn grows exponentially. Taking ratios of successive
values suggests that the base of the exponential lies between 1 and 2, e.g.
F201
= 1.618033988749894848204586834 . . . .
F200
Cognoscenti may recognize this as the decimal expansion of the golden ratio

1+ 5
=
.
2

7. SOLVING HOMOGENEOUS LINEAR RECURRENCES

37

However, lets consider a more general problem and make a vaguer guess. Namely,
for real numbers b, c we consider an recurrence of the form
(6)

x1 = A1 , x2 = A2 , n 1, xn+2 = bxn+1 + cxn .

In all the cases weve looked at the solution was (roughly) exponential. So lets
guess an exponential solution xn = Crn and plug this into the recurrence; we get
Crn+2 = xn+2 = b(Crn+1 ) + c(Crn ),
which simplifies to
r2 br cr = 0.
Evidently the solutions to this are
r=

b2 + 4c
.
2
2

Some cases to be concerned about are the case c = b4 , in which case we have
2
only a single root r = 2b , and the case c < b4 in which case the roots are complex
numbers.
case: b = c = 1. Then
But for the moment lets look at the Fibonacci

r = 12 5 . So we recover the golden ratio = 1+2 5 a good sign! as well as

1 5
= 1 = .618033988749894848204586834 . . . .
2
So we have two different bases what do we do with that? A little thought shows
that if r1n and r2n are both solutions to the recurrence xn+2 = bxn+1 cxn (with any
initial conditions), then so is C1 r1n + C2 r2n for any constants C1 and C2 . Therefore
we propose xn = C1 r1n + C2 r2n as the general solution to the two-term homogeneous linear recurrence (6) and the two initial conditions x1 = A1 , x2 = A2 provide
just enough information to solve for C1 and C2 .
Trying this for the Fibonacci sequence, we get
1 = F1 = C1 + C2 (1 ).
1 = F2 = C1 ()2 + C2 (1 )2 .
Multiplying the first equation by and subtracting it from the second equation
will give us a linear equation to solve for C2 , and then we plug the solution into
either of the equations and solve for C1 . It turns out that
1
1
C1 = , C2 = .
5
5
interlude: This is easily said and indeed involves only high school algebra. But
we can do something slicker. Instead of determining the constants by evaluating
Fn at n = 1 and n = 2, its easier to evaluate at n = 1 and n = 0: then we have
F0 = C1 0 + C2 (1 )0 = C1 + C2 .
But for this to work we need to know F0 , which we have not defined. Can it be
defined in a sensible way? Yes! Writing the basic recurrence in the form Fn+1 =
Fn + Fn1 and solving for Fn1 gives:
Fn1 = Fn+1 Fn .

38

2. MATHEMATICAL INDUCTION

This allows us to define Fn for all integers n. In particular, we have

F0 = F2 F1 = 1 1 = 0.
Thus we get
0 = C1 + C2 ,
whereas plugging in n = 1 gives
1 = C1 () + C2 (1 ) = C1 () C1 (1 ) = (2 1)C1 ,
C1 =

1
1
1
1
=  
= , C2 = .
1+
5
2 1
5
5
2
1
2

Now we are ready to prove the following result.

Theorem 2.12. (Binets Formula) For any n Z, the nth Fibonacci number
is

where =

1+ 5
2 .

1
Fn = (n (1 )n ) ,
5

Proof. We go by strong/complete induction on n. The base cases are n = 1

and n = 2, but we have already checked these: we used them to determine the
constants C1 and C2 . So now assume that n 3 and that the formula is correct
for all positive integers smaller than n + 2. Then, using the identities
2 = + 1,
(1 ) = 1 ,
1 1 = 2 = ()2 ,
we compute
1
Fn+2 = Fn+1 + Fn = (n+1 + n (1 )n+1 (1 )n ))
5
1
= (n ( + 1) (1 )n (1 + 1) =
5
1
(n (2 ) ()n (()1 + 1)
5
1
1
1
= (n+2 ()n ()2 ) = (n+2 ()(n+2) ) = (n+2 (1)n+2 ).
5
5
5

Exercise 7: Find all n Z such that Fn < 0.
By the way, it is not quite true that any solution to (6) must have exponential
growth. For instance, consider the recurrence
x1 = 1, x2 = 2; n 1, xn+2 = 2xn+1 xn .
Then
x3 = 2x2 x1 = 2 2 1 = 3, x4 = 2x3 x2 = 2 3 2 = 4, x5 = 2 4 3 = 5.

8. THE WELL-ORDERING PRINCIPLE

39

It certainly looks as though xn = n for all n. Indeed, assuming it to be true for all
positive integers smaller than n + 2, we easily check
xn+2 = 2xn+1 xn = 2(n + 1) n = 2n + 2 n = n + 2.
The characteristic polynomial is r2 2r + 1 = (r 1)2 : it has repeated roots. One
solution is C1 1n = C1 (i.e., xn is a constant sequence). This occurs iff x2 = x1 , so
clearly there are nonconstant solutions as well. It turns out that in general, if the
characteristic polynomial is (x r)2 , then the two basic solutions are xn = rn and
also xn = nrn . It is unfortunately harder to guess this in advance, but it is not hard
to check that this gives a solution to a recurrence of the form xn+2 = 2r0 xn+1 r02 xn .
These considerations will be eerily familiar to the reader who has studied differential equations. For a more systematic exposition on discrete analogues of calculus
concepts (with applications to the determination of power sums as in 3), see [DC].
8. The Well-Ordering Principle
There is yet another form of mathematical induction that can be used to give what
is, arguably, an even more elegant proof of Proposition 2.11.
Theorem 2.13. (Well-Ordering Principle) Let S be a nonempty subset of Z+ .
Then S has a least element, i.e., there exists s S such that for all t S, s t.
Intutitively, the statement is true by the following reasoning: first we ask: is 1 S?
If so, it is certainly the least element of S. If not, we ask: is 2 S? If so, it is
certainly the least element of S. And then we continue in this way: if we eventually
get a yes answer then we have found our least element. But if for every n the
answer to the question Is n an element of S? is negative, then S is empty!
The well-ordering principle (henceforth WOP) is often useful in its contrapositive form: if a subset S Z+ does not have a least element, then S = .
We claim WOP is logically equivalent to the principle of mathematical induction (PMI) and thus also to the principle of strong/complete induction (PS/CI).
First we will assume PS/CI and show that WOP follows. For this, observe that
WOP holds iff P (n) holds for all n Z+ , where P (n) is the following statement:
P (n): If S Z+ and n S, then S has a least element.
Indeed, if P (n) holds for all n and S Z is nonempty, then it contains some
positive integer n, and then we can apply P (n) to see that S has a least element.
Now we can prove that P (n) holds for all n by complete induction: first, if 1 S,
then indeed 1 is the least element of S, so P (1) is certainly true. Now assume P (k)
for all 1 k n, and suppose that n + 1 S. If n + 1 is the least element of S,
then were done. If it isnt, then it means that there exists k S, 1 k S. Since
we have assumed P (k) is true, therefore there exists a least element of S.
Conversely, let us assume WOP and prove PMI. Namely, let S Z and suppose
that 1 S, and that for all n, if n S then n + 1 S. We wish to show that
S = Z+ . Equivalently, putting T = Z+ \ S, we wish to show that T = . If not,

40

2. MATHEMATICAL INDUCTION

then by WOP T has a least element, say n. Reasoning this out gives an immediate
contradiction: first, n 6 S. By assumption, 1 S, so we must have n > 1, so that
we can write n = m + 1 for some m Z+ . Further, since n is the least element of
T we must have n 1 = m S, but now our inductive assumption implies that
n + 1 = n S, contradiction.
So now we have shown that PMI PS/CI = WOP = PMI.
Let us give another proof of Proposition 2.11 using WOP. We wish to show that
every integer n > 1 can be factored into primes. Let S be the set of integers n > 1
which cannot be factored into primes. Seeking a contradiction, we assume S is
nonempty. In that case, by WOP it has a least element, say n. Now n is certainly
not prime, since otherwise it can be factored into primes. So we must have n = ab
with 1 < a, b < n. But now, since a and b are integers greater than 1 which are
smaller than the least element of S, they must each have prime factorizations, say
a = p1 pk , b = q1 ql . But then (stop me if youve heard this one before)
n = ab = p1 pk q1 ql
itself can be expressed as a product of primes, contradicting our assumption. therefore S is empty: every integer greater than 1 is a product of primes.
This kind of argument is often called proof by minimum counterexample.
These two proofs of Proposition 2.11 are very close: the difference between a proof
by PS/CI and a proof by WOP is more a matter of taste than technique.
9. The Fundamental Theorem of Arithmetic
9.1. Euclids Lemma and the Fundamental Theorem of Arithmetic.
The following are the two most important theorems in beginning number theory.
Theorem 2.14. (Euclids Lemma) Let p be a prime number and a, b be positive
Suppose that p | ab. Then p | a or p | b.
Theorem 2.15. (Fundamental Theorem of Arithmetic) The factorization of
any integer n > 1 into primes is unique, up to the order of the factors: suppose
n = p1 pk = q1 ql ,
are two factorizations of n into primes, with p1 . . . pk and q1 . . . ql . Then
k = l and pi = qi for all 1 i k.
A prime factorization n = p1 pk is in standard form if p1 . . . pk . Every
prime factorization can be put in standard form by ordering the primes from least
to greatest. Dealing with standard form factorizations is a convenient bookkeeping
device, since otherwise our uniqueness statement would have to include the proviso
up to the order of the factors.
Given Proposition 2.11 i.e., the existence of prime factorizations Theorems
2.14 and 15.13 are equivalent: each can be easily deduced from the other.

9. THE FUNDAMENTAL THEOREM OF ARITHMETIC

41

EL implies FTA: Assume Euclids Lemma. As seen above, this implies Proposition 2.9: if a prime divides any finite product of integers it must divide one of
the factors. Our proof will be by minimal counterexample: suppose that there are
some integers greater than one which factor into primes in more than one way, and
let n be the least such integer, so
(7)

n = p1 pk = q1 ql ,

where each of the primes is written in nonincreasing order. Evidently p1 | n =

q1 ql , so by Proposition 2.9, we must have that p1 | qj for some 1 j l. But
since qj is also prime, this means that p1 = qj . Therefore we can cancel them from
the expression, getting
n
= p2 pk = q1 qj1 qj+1 ql .
(8)
p1
But pn1 is less than the least integer which has two different factorizations into
primes, so it must have a unique factorization into primes, meaning that the primes
on the left hand side of (8) are equal, in order, to the primes on the right hand
side of (8). This also implies that p1 = qj is less than or equal to all the primes
appearing on the right hand side, so j = 1. Thus k = l, p1 = qj = q1 and pi = qi
for 2 i j. This means that in (7) the two factorizations are the same after all!
FTA implies EL: Assume every integer greater than one factors uniquely into
a product of primes, let p be a prime, and let a and b be positive integers such that
p | ab. If either a or b is 1, then the other is just p and the conclusion is clear, so
we may assume a, b > 1 and therefore have unique prime factorizations
a = p1 pr , b = q1 qs ;
our assumption that p divides ab means ab = kp for some k Z+ and thus
ab = p1 pr q1 qs = kp.
The right hand side of this equation shows that p must appear in the prime factorization of ab. Since the prime factorization is unique, we must have at least one pi or
at least one qj equal to p. In the first case p divides a; in the second case p divides b.
The traditional route to FTA is via Euclids Lemma, and the traditional route
to Euclids Lemma (Euclids route in Elements) is via a series of intermediate steps
including the Euclidean algorithm and finding the set of all integer solutions to
equations of the form ax+by = 1. This takes some time to develop perhaps a week
in an elementary number theory course. But we can bypass all these intermediate
steps and give direct inductive proofs of both EL and FTA. And we will.
9.2. Rogers Inductive Proof of Euclids Lemma.
Here is a proof of Euclids Lemma using WOP, following K. Rogers [Ro63].
Seeking a contradiction, we suppose there is at least one prime such that Euclids Lemma does not hold for that prime. By WOP there is a least such prime,
say p, so there are a, b Z+ with p | ab but p - a and p - b. By WOP there is a
least a Z+ such that there is at least one b Z+ with p | ab and p - a, p - b.
Now consider the following equation:
ab = (a p)b + pb,

42

2. MATHEMATICAL INDUCTION

which shows that p | ab p | (a p)b. There are three cases:

Case 1: a p is a positive integer. Then, since 0 < a p < a and a was by
assumption the least positive integer such that Euclids Lemma fails for the prime
p, we must have that p | a p or p | b. By assumption p - b, so we must have
p | a p, but then p | (a p) + p = a, contradiction!
Case 2: We have a = p. But then p | a, contradiction.
Case 3: We have a < p. On the other hand, a > 1 if p | 1 b, then p | b so by
Proposition 2.11 a is divisible by some prime q, and q | a < p, so q < p. Therefore q
is a prime which is smaller than the least prime for which Euclids Lemma fails, so
Euclids Lemma holds for q. Since p | ab, we may write pk = ab for some k Z+ ,
and now q | a = q | ab = pk, so by Euclids Lemma for q, q | p or q | k. The first
case
since p is prime and 1 < q < p, so we must have q | k. Therefore
 
 is impossible
p

k
q

a
q

b, so p |

a
q b.

But 1 <

a
q

< a and a is the least positive integer for

which Euclids Lemma fails for p and a, so it must be that p | aq (so in particular
p | a) or p | b. Contradiction. So Euclids Lemma holds for all primes p.
9.3. The Lindemann-Zermelo Inductive Proof of FTA.
Here is a proof of FTA using WOP, following Lindemann [Li33] and Zermelo [Ze34].
We claim that the standard form factorization of a positive integer is unique.
Assume not; then the set of positive integers which have at least two different
standard form factorizations is nonempty, so has a least element, say n, where:
(9)

n = p1 pr = q1 qs .

Here the pi s and qj s are prime numbers, not necessarily distinct from each other.
However, p1 6= qj for any j. Indeed, if we had such an equality, then after relabelling
the qj s we could assume p1 = q1 and then divide through by p1 = q1 to get a smaller
positive integer pn1 . By the assumed minimality of n, the prime factorization of pn1
must be unique: i.e., r 1 = s 1 and pi = qi for all 2 i r. Then multiplying
by p1 = q1 we see that we didnt have two different factorizations after all.
In particular p1 6= q1 . Without loss of generality, assume p1 < q1 . Then, if we
subtract p1 q2 qs from both sides of (9), we get
(10)

m := n p1 q2 qs = p1 (p2 pr q2 qs ) = (q1 p1 )(q2 qs ).

Evidently 0 < m < n, so by minimality of n, the prime factorization of m must

be unique. However, (10) gives two different factorizations of m, and we can use
these to get a contradiction. Specifically, m = p1 (p2 pr q2 qs ) shows that
p1 | m. Therefore, when we factor m = (q1 p1 )(q2 qs ) into primes, at least
one of the prime factors must be p1 . But q2 , . . . , qj are already primes which are
different from p1 , so the only way we could get a p1 factor is if p1 | (q1 p1 ). But
this implies p1 | q1 , and since q1 is also prime this implies p1 = q1 . Contradiction!
9.4. A Generalized Euclids Lemma.
Theorem 2.16. (Generalized Euclids Lemma) Let a, b, c Z+ . Suppose that
x | yz and x and y are relatively prime. Then x | z.
Exercise: a) Prove Theorem 2.16 using the Fundamental Theorem of Arithmetic.
b) Explain why Theorem 2.16 is indeed a generalization of Euclids Lemma.

CHAPTER 3

Polynomial and Rational Functions

1. Polynomial Functions
Using the basic operations of addition, subtraction, multiplication, division and
composition of functions, we can combine very simple functions to build large and
interesting (and useful!) classes of functions. For us, the two simplest kinds of
functions are the following:
Constant functions: for each a R there is a function Ca : R R such that for
all x R, Ca (x) = a. In other words, the output of the function does not depend
on the input: whatever we put in, the same value a will come out. The graph of
such a function is the horizontal line y = a. Such functions are called constant.
The identity function I : R R by I(x) = x. The graph of the identity
function is the straight line y = x.
Recall that the identity function is so-called because it is an identity element for
the operation of function composition: that is, for any function f : R R we have
I f = f I = f.
Example: Let m, b R, and consider the function L : R R by x 7 mx + b.
Then L is built up out of constant functions and the identity function by addition
and multiplication: L = Cm I + Cb .
Example: Let n Z+ . The function mn : x 7 xn is built up out of the identity function by repreated multiplication: mn = I I I (n Is altogether).
The general name for a function f : R R which is built up out of the identity
function and the constant functions by finitely many additions and multiplications
is a polynomial. In other words, every polynomial function is of the form
(11)

f : x 7 an xn + . . . + a1 x + a0

for some constants a0 , . . . , an R.

However, we also want to take at least until we prove it doesnt make a difference a more algebraic approach to polynomials.
Let us define a polynomial
Pn
expression as an expression of the form i=0 ai xi . Thus, to give a polynomial
expression we need to give for each natural number i a constant ai , while requiring
that all but finitely many of these constants are equal to zero: i.e., there exists
some n N such that ai = 0 for all i > n.
43

44

3. POLYNOMIAL AND RATIONAL FUNCTIONS

Pn
Then every polynomial expression f = i=0 ai xi determines a polynomial function x 7 f (x). But it is at least conceivable that two different-looking polynomial
expressions give rise to the same function. To give some rough idea of what I mean
here, consider the two expressions f = 2 arcsin x + 2 arccos x and g = . Now it
turns out for all x [1, 1] (the common domain of the arcsin and arccos functions) we have f (x) = . (The angle whose sine is x is complementary to the
angle whose cosine is x, so arcsin x + arccos x = + = 2 .) But still f and g are
given by different expressions: if I ask you what the coefficient of arcsin x is in
the expression f , you will immediately tell me it is 2. If I ask you what the coefficient of arcsin x is in the expression , you will have no idea what Im talking about.
One special polynomial expression is the zero polynomial. This is the polynomial whose ith coefficient ai is equal to zero for all i 0.
Every nonzero polynomial expression has a degree, which is a natural number,
the largest natural number i such that the coefficient ai of xi is nonzero. Thus in
(11) the degree of f is n if and only if an 6= 0: otherwise the degree is smaller than n.
Although the zero polynomial expression does not have any natural number as
a degree, it is extremely convenient to regard deg 0 as negative, i.e., such that deg 0
is smaller than the degree of any nonzero polynomial. This means that for any
d N the set of polynomials of degree at most d includes the zero polynomial.
We will follow this convention here but try not to make too big a deal of it.
Let us give some examples to solidify this important concept:
The polynomials of degree at most 0 are the expressions f = a0 . The corresponding
functions are all constant functions: their graphs are horizontal lines. (The graph
of the zero polynomial is still a horizontal line, y = 0, so it is useful to include the
zero polynomial as having degree at most zero.)
The polynomials of degree at most one are the linear expressions L = mx + b.
The corresponding functions are linear functions: their graphs are straight lines.
The degree of L(x) is one if m 6= 0 i.e., if the line is not horizontal and 0 if
m = 0 and b 6= 0.
Similarly the polynomials of degree at most two are the quadratic expressions
q(x) = ax2 + bc + c. The degree of q is 2 unless a = 0.
We often denote the degree of the polynomial expression f by deg f .
Theorem 3.1. Let f, g be nonzero polynomial expressions.
a) If f + g 6= 0, then deg(f + g) max(deg f, deg g).
b) We have deg(f g) = deg f + deg g.
Proof. a) Suppose that
f (x) = am xm + . . . + a1 x + a0 , am 6= 0
and
g(x) = bn xn + . . . + b1 x + b0 , bn 6= 0

1. POLYNOMIAL FUNCTIONS

45

so that deg g = m, deg g = n.

Case 1: m > n. Then when we add f and g, the highest order term will be am xm ,
since the polynomial g only smaller powers of x. In particular the degree of f + g
is m = max(m, n).
Case 2: m < n. Similarly, when we add f and g, the highest order term will be
an xn , so the degree of f + g is n = max(m, n).
Case 3: Suppose m = n. Then
(f + g)(x) = (am + bm )xm + . . . + (a1 + b1 )x + (a0 + b0 ).
Thus the degree of f + g is at most m. It will be exactly m unless am + bm = 0,
i.e., unless bm = am ; in this case it will be strictly smaller than m.
b) If f and g are as above, then the leading term of f g will be am bn xm+n , and
since am , bn 6= 0, am bn 6= 0. Thus deg f g = m + n.

Pm
P
n
Exercise: For polynomial expressions f = i=0 ai xi and g = j=0 bj xj , define
(f g)(x) =

m
X
i=0

n
X
ai (
bj xj )n .
j=0

Show that deg(f g) = (deg f )(deg g).

The following is the most important algebraic property of polynomials.
Theorem 3.2. (Polynomial Division With Remainder) Let a(x) be a polynomial expression and b(x) be a nonzero polynomial expression. There are unique
polynomial expressions q(x) and r(x) such that
(i) a(x) = q(x)b(x) + r(x) and
(ii) deg r(x) < deg b(x).
Proof. First note that we really get to choose only q(x), for then r(x) =
a(x) q(x)b(x). Also let us denote the leading coefficient of a(x) by and the
leading coefficient of b(x) by . Step 1: We prove the existence of q(x) and b(x).
Case 1: deg a(x) < deg b(x). Take q(x) = 0: deg r(x) = deg a(x) < deg b(x).
deg a(x)deg b(x)
. The point of this
Case 2: deg a(x) degb(x). Take q1 (x) =
x
is that q1 (x)b(x) has degee deg a(x) deg b(x) + deg b(x) = deg a(x) and leading
coefficient
= , so in r1 (x) = a(x) q1 (x)b(x) the leading terms cancel and
deg r1 (x) deg(a(x))1. If deg r1 (x) < deg b(x) then were done, take q(x) = q1 (x)
and r(x) = r1 (x). On the other hand, if deg r1 (x) deg b(x) we apply the process
again with r1 (x) in place of a(x): let 1 be the leading coefficient of r1 (x) and
take q2 (x) = 1 xdeg r1 (x)deg b(x) , so that in r2 (x) = r1 (x) q2 (x)b(x) the leading
terms cancel to give deg r2 (x) deg(r1 (x)) 1 deg(a(x)) 2. Continue in
this way generating polynomial expressions qi (x) and ri (x) until we reach a k with
deg rk (x) < deg b(x). Then:
rk (x) = rk1 (x) qk (x)b(x)
= rk2 (x) qk1 (x)b(x) qk b(x)
= rk3 (x) qk2 (x)b(x) qk1 (x)b(x) qk b(x)
..
.
= a(x) (q1 (x) + q2 (x) + . . . qk (x))b(x).

46

3. POLYNOMIAL AND RATIONAL FUNCTIONS

Thus we may take q(x) = q1 (x) +. . . +qk (x), so that r(x) = a(x)q(x)b(x) = rk (x)
and deg r(x) = deg rk (x) < deg b(x).
Step 2: We prove the uniqueness of q(x) (and thus of r(x) = a(x) q(x)b(x)).
Suppose Q(x) is another polynomial such that R(x) = a(x) Q(x)b(x) has degree
less than the degree of b(x). Then
a(x) = q(x)b(x) + r(x) = Q(x)b(x) + R(x),
so
(q(x) Q(x))b(x) = R(x) r(x).
Since r(x) and R(x) both have degree less than deg b(x), so does r(x) R(x), so
deg b(x) > deg(R(x)r(x)) = deg((q(x)Q(x))b(x)) = deg(q(x)Q(x))+deg b(x).
Thus deg(q(x) Q(x)) < 0, and the only polynomial with negative degree is the
zero polynomial, i.e., q(x) = Q(x) and thus r(x) = R(x).1

Exercise: Convince yourself that the proof of Step 1 is really a careful, abstract
description of the standard high school procedure for long division of polynomials.
Theorem 3.2 has many important and useful consequences; here are some of them.
Theorem 3.3. (Root-Factor Theorem) Let f (x) be a polynomial expression and
c a real number. The following are equivalent:
(i) f (c) = 0. (c is a root of f .)
(ii) There is some polynomial expression q such that as polynomial expressions,
f (x) = (x c)q(x). (x c is a factor of f .)
Proof. We apply the Division Theorem with a(x) = f (x) and b(x) = x c,
getting polynomials q(x) and r(x) such that
f (x) = (x c)q(x) + r(x)
and r(x) is either the zero polynomial or has deg r < deg x c = 1. In other
words, r(x) is in all cases a constant polynomial (perhaps constantly zero), and its
constant value can be determined by plugging in x = c:
f (c) = (c c)q(c) + r(c) = r(c).
The converse is easier: if f (x) = (x c)q(x), then f (c) = (c c)q(c) = 0.

Corollary 3.4. Let f be a nonzero polynomial of degree n. Then the corresponding polynomial function f has at most n real roots: i.e., there are at most n
real numbers a such that f (a) = 0.
Proof. By induction on n.
Base case (n = 0): If deg f (x) = 0, then f is a nonzero constant, so has no roots.
Induction Step: Let n N, suppose that every polynomial of degree n has at
most n real roots, and let f (x) be a polynomial of degree n + 1. If f (x) has no
real root, great. Otherwise, there exists a R such that f (a) = 0, and by the
Root-Factor Theorem we may write f (x) = (x a)g(x). Moreover by Theorem
3.1, we have n + 1 = deg f = deg(x a)g(x) = deg(x a) + deg g = 1 + deg g, so
deg g = n. Therefore our induction hypothesis applies and g(x) has m distinct real
1If you dont like the convention that the zero polynomial has negative degree, then you
can phrase the argument as follows: if q(x) Q(x) were a nonzero polynomial, these degree
considerations would give the absurd conclusion deg(q(x) Q(x)) < 0, so q(x) Q(x) = 0.

1. POLYNOMIAL FUNCTIONS

47

roots a1 , . . . , am for some 0 m n. Then f has either m + 1 real roots if a is

distinct from all the roots ai of g or m real roots if a = ai for some i, so it has
at most m + 1 n + 1 real roots.

Pn
Lemma 3.5. P
Let f = i=0 ai xi be a polynomial expression. Suppose that the
n
i
function f (x) =
i=0 ai x is the zero function: f (x) = 0 for all x R. Then
ai = 0 for all i, i.e., f is the zero polynomial expression.
Proof. Suppose that f is not the zero polynomial, i.e., ai 6= 0 for some i.
Then it has a degree n N, so by Corollary 3.4 there are at most n real numbers c
such that f (c) = 0. But this is absurd: f (x) is the zero function, so for all (infinitely
many!) real numbers c we have f (c) = 0.

Theorem 3.6. (Uniqueness Theorem For Polynomials) Let
f = an xn + . . . + a1 x + a0 ,
g = bn x n + . . . + b1 x + b0
be two polynomial expressions. The following are equivalent:
(i) f and g are equal as polynomial expressions: for all 0 i n, ai = bi .
(ii) f and g are equal as polynomial functions: for all c R, f (c) = g(c).
(iii) There are c1 < c2 < . . . < cn+1 such that f (ci ) = g(ci ) for 1 i n + 1.
Proof. (i) = (ii): This is clear, since if ai = bi for all i then f and g are
being given by the same expression, so they must give the same function.
(ii) = (iii): This is also immediate: since f (c) = g(c) for all real numbers c, we
may take for instance c1 = 1, c2 = 2, . . . , cn+1 = n + 1.
(iii) = (i): Consider the polynomial expression
h = f g = (an bn )xn + . . . + (a1 b1 )x + (a0 b0 ).
Then h(c1 ) = f (c1 ) g(c1 ) = 0, . . . , h(cn+1 ) = f (cn+1 ) g(cn+1 ) = 0. So h is a
polynomial of degree at most n which has (at least) n + 1 distinct real roots. By
Corollary 3.4, h must be the zero polynomial expression: that is, for all 0 i n,
ai bi = 0. Equivalently, ai = bi for all 0 i n, so f and g are equal as
polynomial expressions.

In particular, Theorem 3.6 says that if two polynomials f (x) and g(x) look different i.e., they are not coefficient-by-coefficient the same expression then they are
actually different functions, i.e., there is some c R such that f (c) 6= g(c).
Finally, we want to prove an arithmetic result about polynomials, the Rational
Roots Theorem. For this we need another number-theoretic preliminary. Two
positive integers a and b are coprime (or relatively prime) if they are not both
divisible by any integer d > 1 (equivalently, they have no common prime factor).
Theorem 3.7. (Rational Roots Theorem) Let a0 , . . . , an be integers, with a0 , an 6=
0. Consider the polynomial
P (x) = an xn + . . . + a1 x + a0 .
Suppose that cb is a rational number, written in lowest terms, which is a root of P :
P ( cb ) = 0. Then a0 is divisible by b and an is divisible by c.

48

3. POLYNOMIAL AND RATIONAL FUNCTIONS

Proof. We know
 
b
b
bn
0=P
= an n + . . . + a1 + a0 .
c
c
c
Multiplying through by cn clears denominators, giving
an bn + an1 bn1 c + . . . + a1 bcn1 + a0 cn = 0.
Rewriting this equation as
an bn = c(an1 bn1 . . . a0 cn1 )
shows that an bn is divisible by c. But since b and c have no prime factors in
common and bn has the same distinct prime factors as does b, bn and c have no
prime factors in common and are thus coprime. So Theorem 2.16 applies to show
that an is divisible by c. Similarly, rewriting the equation as
a0 cn = b(an bn1 an1 bn2 c . . . a1 cn1 )
shows that a0 cn is divisible by b. As above, since b and c are coprime, so are b and
cn , so by Theorem 2.16 a0 is divisible by b.

In high school algebra the Rational Roots Theorem is often used to generate a finite
list of possible rational roots of a polynomial with integer coefficients. This is nice,
but there are more impressive applications. For instance, taking an = 1 and noting
that 1 is divisible by c iff c = 1 we get the following result.
Corollary 3.8. Let a0 , . . . , an1 Z, and consider the polynomial
P (x) = xn + an1 xn1 + . . . + a1 x + a0 .
Suppose c is a rational number such that P (c) = 0. Then c is an integer.
So what? Let p be any prime number, let n 2, and consider the polynomial
P (x) = xn p.
By Corollary 3.8, if c Q is such that P (c) = 0, then c Z. But if c Z is such
that P (c) = 0, then cn = p. But this means that the prime number p is divisible
by the integer c, so c = 1 or c = p. But (1)n = 1 and (p)n = pn , so

cn = p is impossible. So there is no c Q such that cn = p: that is, n p is irrational.

This is a doubly infinite generalization of our first irrationality proof, that 2

is irrational, but the argument is, if anything, shorter and easier to understand.
(However, we did use and prove, by induction the Fundamental Theorem of
Arithmetic, a tool which was not available to us at the very beginning of the
course.) Moral: polynomials can be useful in surprising ways!
Proposition 3.9. For every polynomial P of positive degree there are irreducible polynomials p1 , . . . , pk such that
P = p1 pk .
Exercise: Prove it. (Suggestion: adapt the proof that every integer n > 1 can be
factored into a product of primes.)

2. RATIONAL FUNCTIONS

49

2. Rational Functions
A rational function is a function which is a quotient of two polynomial functions:
P (x)
f (x) = Q(x)
, with Q(x) not the zero polynomial. To be sure, Q is allowed to have
roots; it is just not allowed to be zero at all points of R.
P (x)
The natural domain of Q(x)
is the set of real numbers for which Q(x) 6= 0,
i.e., all but a finite (possibly empty) set of points.

2.1. The Partial Fractions Decomposition.

The goal of this section is to derive the Partial Fractions Decomposition (PFD)
for a proper rational function. The PFD is a purely algebraic fact which is useful
in certain analytic contexts: especially, it is the key technique used to find explicit
antiderivatives of rational functions.
Theorem 3.10. Let P1 , P2 , P be polynomials, with P1 monic irreducible and
P2 not divisible by P1 . Then there are polynomials A, B such that
AP1 + BP2 = P.
Proof. Let S be the set of nonzero polynomials of the form aP1 + bP2 for
some polynomials a, b. Since P1 = 1 P1 + 0 P2 , P2 = 0 P1 + 1 P2 are in S, S =
6 .
By the Well Ordering Principle there is m S of minimal degree: write
aP1 + bP2 = m.
If a polynomial lies in S then so does every nonzero constant multiple of it, so we
may assume m is monic. We claim that for any M = AP1 + BP2 S, we must
have m | M . To see this, apply Polynomial Division to M and m, getting
M = Qm + R, deg R < deg m.
Since also
R = M Qm = (A Qa)P1 + (B Qb)P2 ,
so if R 6= 0 wed have R S of smaller degree than m, contradiction. Thus R = 0
and m | M . Since P1 S, m | P1 . Because P1 is monic irreducible, m = P1 or
m = 1. Since P2 S, also m | P2 , and so by hypothesis m 6= P1 . Thus m = 1 and
aP1 + bP2 = 1.
Multiplying through by P we get
(aP )P1 + (bP )P2 = P,
so we may take A = aP , B = bP .

Corollary 3.11. (Euclids Lemma For Polynomials) Let p be an irreducible

polynomia, and suppose p | P1 P2 . Then p | P1 or p | P2 .
Proof. As is usual in trying to prove statements of the form A implies (B or
C), we assume that B is false and show that C is true. Here this means assuming
p - P1 . By Theorem 3.10, there are polynomials A and B such that
Ap + BP1 = 1.
Multiplying through by P2 gives
ApP2 + BP1 P2 = P2 .

50

3. POLYNOMIAL AND RATIONAL FUNCTIONS

Since p | ApP2 and p | BP1 P2 , p | (Ap2 + BP1 P2 ) = P2 .

Exercise: Show that every monic polynomial of positive degree factors uniquely into
a product of monic irreducible polynomials.
Corollary 3.12. Let n Z+ , and let p, P, Q0 be polynomials, such that p is
irreducible and that p does not divide Q0 .
a) There are polynomials B, A such that we have a rational function identity
A(x)
P (x)
B(x)
+
=
.
p(x)n Q0 (x)
p(x)n
p(x)n1 Q0 (x)

(12)

b) If deg P < deg(pn Q0 ), we may choose A and B such that

deg A < deg(pn1 Q0 ), deg B < deg pn .
Proof. a) Apply Theorem 3.10a) with P1 = p, P2 = Q0 : we get
(13)

P = A1 p + BQ0 .

for some polynomials A1 , B. Thus

P
pn Q0

A1
A1 p + BQ0
B
= n + n1 .
n
p Q0
p
p
Q0

b) By Polynomial Division (Theorem 3.2), there are C, r such that B = Cp + r and

deg r < deg p. Substituting in (13) and putting A = A1 + CQ0 we get
P = A1 p + BQ0 = A1 p + (Cp + r)Q0 = (A1 + CQ0 )p + rq0 = Ap + rQ0 .
We have
deg(rQ0 ) = deg r + deg Q0 < deg p + deg Q0 deg pn + deg Q0 = deg(pn Q0 ).
Since also deg P < deg(pn Q0 ) and Ap = P rQ0 , it follows that
deg A + deg p = deg(Ap) < deg(pn Q0 ) = deg Q0 + n deg p,
so
deg A deg Q0 + (n 1) deg p = deg(pn1 Q0 ).

The identity of Corollary 3.12 can be applied repeatedly: suppose we start with a
P
, with Q a monic polynomial (as is no loss of generality;
proper rational function Q
we can absorb the leading coefficient into P ). Then Q is a polynomial of positive
degree, so we may factor it as
Q = pa1 1 par r ,
where p1 , . . . , pr are distinct monic irreducible polynomials. Let us put Q0 =
pa2 2 par r , so Q = pa1 1 Q0 . Applying Corollary 3.12, we may write
Ba
Aa1
P
= a11 + n1
Q
p1
p1 Q0
with deg Ba1 < deg p1 and deg Aa1 < deg(pn1
Q0 ). Thus Corollary 3.12 applies to
1
Aa1
,
as
well,
giving
us
overall
a1 1
p1

Q0

P
Ba
Ba 1
Aa1 1
= a11 + a111 + n2
.
Q
p1
p1 Q0
p1

2. RATIONAL FUNCTIONS

51

And we may continue in this way until we get

B1
P
Ba
A0
= a11 + . . . +
+
,
Q
p1
p1
Q0
with each deg B < deg P1 and deg A0 < deg Q0 . Recalling that Q0 = pa2 2 par r ,
we may put Q0 = pa2 2 Q1 and continue on. Finally we get an identity of the form
P
B1,1
B1,a1
Br,ar
Br,1
(14)
+ . . . + ar + . . . +
,
ar =
a1
a1 + . . . +
p1 pr
p1
p1
pr
pr
with deg Bi,j < deg pi for all i and j. From an algebraic perspective, (14) is the
Partial Fractions Decomposition. In fact, though it is not the point for us,
everything that we have done works for polynomials with coefficients in any field K
e.g. K = Q, K = C, K = Z/pZ. Conversely, when we work with polynomials over
the real numbers, the expression simplifies, because the classification of irreducible
polynomials over the real numbers is rather simple. Indeed:
Lemma 3.13. The irreducible (real!) polynomials are precisely:
(i) The linear polynomials `(x) = a(x c) for a, c R, a 6= 0; and
(ii) The quadratic polynomisls Q(x) = ax2 + bx + c for a, b, c R, b2 4ac < 0.
Every linear polynomial is irreducible. Further, from the Rational Roots Theorem,
for a polynomial of degree greater than 1 to be irreducible, it cannot have any real
roots. For quadratic (and, for the record, cubic) polynomials this is also sufficient,
since a nontrivial factorization of a degree 2 or 3 polynomial necessarily has a linear
factor and thus a real root. It is a a consequence of the Quadratic Formula that a
real quadratic polynomial Q(x) = ax2 + bx + c has no real roots iff its discriminant
b2 4ac is negative. Thus half of Lemma 3.13 the half that says that a linear
polynomials and quadratic polynomial without real roots are irreducible is easy.
The other half of Lemma 3.13 the half that says that there are no other
irreducible polynomials is far from easy: in fact it lies well beyond our current
means. In several hundred pages time we will return to prove this result as a
consequence of the Fundamental Theorem of Algebra: see Theorem 15.17.
Combining (14) and Lemma 3.13 we get the desired result.
Theorem 3.14. (Real Partial Fractions Decomposition) Let
rational function, with

P (x)
Q(x)

be a proper

Q(x) = (x c1 )m1 (x ck )mk q1 (x)n1 ql (x)nl ;

here c1 , . . . , ck R, with a 6= 0, c1 , . . . , ck distinct; and q1 , . . . , ql are distinct monic
irreducible quadratics. There are real numbers A, , B, , C, such that we have an
identity of rational functions
(15)
mk
nl
m1
n1
X
X
X
P (x) X
A1,i
Ak,i
B1,j x + C1,j
Bl,j x + Cl,j
=
+. . .+
+
+. . .+
.
i
i
j
Q(x)
(x

c
)
(x

c
)
q
(x)
ql (x)j
1
k
1
i=1
j=1
i=1
j=1
Exercise: Show that the constants in (15) are unique.

CHAPTER 4

Continuity and Limits

1. Remarks on the Early History of the Calculus
We have seen that in order to define the derivative f 0 of a function f : R R we
need to understand the notion of a limit of a function at a point. It turns out that
giving a mathematically rigorous and workable definition of a limit is hard really
hard. Let us begin with a quick historical survey.
It is generally agreed that calculus was invented (discovered?) independently by
Isaac Newton and Gottfried Wilhelm von Leibniz, roughly in the 1670s. Leibniz
was the first to publish on calculus, in 1685. However Newton probably could have
published his work on calculus before Leibniz, but held it back for various reasons.1
To say that calculus was discovered by Newton and Leibniz is an oversimplification. Computations of areas and volumes which we can now recognize as
using calculus concepts go back to ancient Egypt, if not earlier. The Greek mathematicians Eudoxus (408-355 BCE) and Archimedes (287-212 BCE) developed
the method of exhaustion, a limiting process which anticipates integral calculus.
Also Chinese and Indian mathematicians made significant achievements. Even in
modern Europe, Newton and Leibniz were not functioning in a complete intellectual vacuum. They were responding to and continuing earlier work by Pierre
de Fermat (on tangent lines) and John Wallis, Isaac Barrow and James Gregory. This should not be surprising: all scientific work builds on work of others.
But the accomplishments of Newton and Leibniz were so significant that after their
efforts calculus existed as a systematic body of work, whereas before them it did not.
How did Newton and Leibniz construe the fundamental concept, namely that of
a limit? Both of their efforts were far from satisfactory, indeed far from making
sense. Newtons limiting concept was based on a notion of fluxions, which is so
obscure as not to be worth our time to describe it. Leibniz, a philosopher and writer
as well as a mathematician, addressed the matter more thoroughly and came up
with the notion of an infinitesimal quantity, a number which is not zero but
vanishingly small, i.e., smaller than any ordinary positive number.
The concept of infinitesimals has been taken up by mathematicians since Leibniz, and eventually with complete mathematical success...but not until the 1960s!2
For instance one has a definition of an infinitesimal element x of an ordered field
K, namely an element x which is positive but smaller than n1 for all n Z+ . It is
1I highly recommend James Gleicks biography of Newton. If I wanted to distill hundreds of
pages of information about his personality into one word, the word I would choose is...weirdo.
2See http://en.wikipedia.org/wiki/Nonstandard analysis for an overview of this story.
53

54

4. CONTINUITY AND LIMITS

easy to see that an ordered field admits infinitesimal elements iff it does not satisfy
the Archimedean axiom, whereas the real numbers R do satisfy the Archimedean
axiom. So at best Lebiniz was advocating a limiting process based on a different
mathematical model of the real numbers than the standard modern one. And
at worst, Lebinizs writing on infinitesimals seems like equivocation: at different
stages of a calculation the same quantity is at one point vanishingly small and at
another point not. The calculus of both fluxions and infnitesimals required, among
other things, some goodwill: if you used them as Newton and Leibniz did in their
calculations then at the end you would get a sensible (in fact, correct!) answer.
But if you wanted to make trouble and ask why infinitesimals could not be manipulated in other ways which swiftly led to contradictions, it was all too easy to do so.
The calculus of Newton and Leibniz had a famous early critic, Bishop George
Berkeley. In 1734 he published The Analyst, subtitled A DISCOURSE Addressed
to an Infidel MATHEMATICIAN. WHEREIN It is examined whether the Object,
Principles, and Inferences of the modern Analysis are more distinctly conceived,
or more evidently deduced, than Religious Mysteries and Points of Faith. Famously, Berkeley described fluxions as the ghosts of departed quantities. I havent
read Berkeleys text, but from what I am told it displays a remarkable amount of
mathematical sophistication and most of its criticisms are essentially valid!
So if the mid 17th century is the birth of systematic calculus it is not the birth
of a satisfactory treatment of the limiting concept. When did this come? More
than 150 years later! The modern definition of limits via inequalities was given by
Bolzano in 1817 (but not widely read), in a somewhat imprecise form by Cauchy
in his influential 1821 text, and then finally by Weierstrass around 1850.
2. Derivatives Without a Careful Definition of Limits
Example 2.1: Let f (x) = mx + b be a linear function. Then f has the following
property: for any x1 6= x2 , secant line between the two points (x1 , f (x1 )) and
(x2 , f (x2 )) is the line y = f (x). Indeed, the slope of the secant line is
mx2 + b (mx1 + b)
m(x2 x1 )
f (x2 ) f (x1 )
=
=
= m.
x2 x1
x2 x1
x2 x1
Thus the secant line has slope m and passes through the point (x1 , f (x1 )), as does
the linear function f . But there is a unique line passing through a given point with
a given slope, so that the secant line must be y = mx + b.
Using this, it is now not at all difficult to compute the derivative of a linear function...assuming an innocuous fact about limits.
Example 2.2: Let f (x) = mx + b. Then
f 0 (x) = lim

h0

f (x + h) f (x)
mh
= lim m(x + h) + b (mx + b)h = lim
= lim m.
h0
h0 h
h0
h

The above computation is no surprise, since we already saw that the slope of any
secant line to a linear function y = mx + b is just m. So now we need to evaluate
the limiting slope of the secant lines. But surely if the slope of every secant line

2. DERIVATIVES WITHOUT A CAREFUL DEFINITION OF LIMITS

55

is m, the desired limiting slope is also m, and thus f 0 (x) = m (constant function).
Let us record the fact about limits we used.
Fact 4.1. The limit of a constant function f (x) = C as x approaches a is C.
Example 2.3: Let f (x) = x2 . Then
f (x + h) f (x)
(x + h)2 x2
= lim
h0
h0
h
h

f 0 (x) = lim

x2 + 2xh + h2 x2
h(2x + h)
= lim
= lim 2x + h.
h0
h0
h0
h
h
Now Leibniz would argue as follows: in computing the limit, we want to take h
infinitesimally small. Therefore 2x + h is infinitesimally close to 2x, and so in the
limit the value is 2x. Thus
f 0 (x) = 2x.
= lim

But these are just words. A simpler and equally accurate description of what we
(x)
until we
have done is as follows: we simplified the difference quotient f (x+h)f
h
got an expression in which it made good sense to plug in h = 0, and then we
plugged in h = 0. If you wanted to give a freshman calculus student practical
instructions on how to compute derivatives of reasonably simple functions directly
from the definition, I think you couldnt do much better than this!
Example 2.4: f (x) = x3 . Then
f (x + h) f (x)
(x + h)3 x3
x3 + 3x2 h + 3xh2 + h3 x3
= lim
= lim
h0
h0
h0
h
h
h

f 0 (x) = lim

h(3x2 + 3xh + h2 )
= lim 3x2 + 3xh + h2 .
h0
h0
h
Again we have simplified to the point where we may meaningfully set h = 0, getting
= lim

f 0 (x) = 3x2 .
Example 2.5: For n Z+ , let f (x) = xn . Then
Pn n
ni i
x h xn
f (x + h) f (x)
(x + h)n xn
f (x) = lim
= lim
= lim i=0 i
h0
h0
h0
h
h
h
Pn n
ni ii
 


n
X n
h i=1 i x h
n n1
= lim
x
+h
xni hi2
= lim
h0
h0 1
h
i
i=2
 
n n1
=
x
= nxn1 .
1
At this point we have seen many examples of a very pleasant algebraic phenomenon.
Namely, for y = f (x) a polynomial function, when we compute the difference
(x)
quotient f (x+h)f
we find that the numerator, say G(h) = f (x + h) f (x),
h
always has h as a factor: thus we can write it as G(h) = hg(h), where g(h) is
another polynomial in h. This is exactly what we need in order to compute the
derivative, because when this happens we get
0

f (x + h) f (x)
hg(h)
= lim
= lim g(h) = g(0).
h0
h0
h0
h
h

f 0 (x) = lim

56

4. CONTINUITY AND LIMITS

Can we guarantee that this factorization f (x + h) f (x) = G(h) = hg(h) always

takes place? Yes, we can! By the Root-Factor Theorem, we may factor out h = h0
from the polynomial G(h) iff G(0) = 0. But G(0) = f (x+0)f (x) = f (x)f (x) =
0. Thus some simple polynomial algebra ensures that we will be able to compute
the derivative of any polynomial function, provided we assume the following fact.
Fact 4.2. For any polynomial function g,
lim g(h) = g(a).

ha

(This generalizes our first fact above, since constant functions are polynomials.)
Differentiating polynomial functions directly from the definition is, evidently, somewhat tedious. Perhaps we can establish some techniques to streamline the process?
For instance, suppose we know the derivative of some function f : what can we say
about the derivative of cf (x), where c is some real number? Lets see:


cf (x + h) cf (x)
f (x + h) f (x)
lim
= lim c
.
h0
h0
h
h
If we assume limxa cf (x) = c limxa f (x), then we can complete this computation:
the derivative of cf (x) is cf 0 (x). Let us again record our assumption about limits.
Fact 4.3. If limxa f (x) = L, then limxa cf (x) = cL.
This tells for instance that the derivative of 17x10 is 17(10x9 ) = 170x9 . More generally, this tells us that the derivative of the general monomial cxn is cnxn1 .
Now what about sums?
Let f and g be two differentiable functions. Then the derivative of f + g is
f (x + h) + g(x + h) f (x) g(x)
f (x + h) f (x) g(x + h) g(x)
lim
= lim
+
.
h0
h0
h
h
h
If we assume the limit of a sum is the sum of limits, we get
(f + g)0 = f 0 + g 0 .
Again, lets record what weve used.
Fact 4.4. If limxa f (x) = L and limxa g(x) = M , then limx (f + g)(x) =
L + M.
Exercise 2.6: Show by mathematical induction that if f1 , . . . , fn are functions with
derivatives f10 , . . . , fn0 , then (f1 + . . . + fn )0 = f10 + . . . + fn0 .
Putting these facts together, we get an expression for the derivative of any polynomial function: if f (x) = an xn + . . . + a1 x + a0 , then f 0 (x) = nan xn1 + . . . + a1 . In
particular the derivative of a degree n polynomial is a polynomial of degree n 1
(and the derivative of a constant polynomial is the zero polynomial).
3. Limits in Terms of Continuity
We have been dancing around two fundmaental issues in our provisional treatment
of derivatives. The first is, of course, the notion of the limit of a function at a point.
The second, just as important, is that of continuity at a point.
In freshman calculus it is traditional to define continuity in terms of limits. A

3. LIMITS IN TERMS OF CONTINUITY

57

true fact which is not often mentioned is that this works just as well the other way
around: treating the concept of a continuous function as known, one can define
limits in terms of it. Since I think most people have at least some vague notion
of what a continuous function is very roughly it is that the graph y = f (x) is a
nice, unbroken curve and I know all too well that many students have zero intuition for limits, it seems to be of some value to define limits in terms of continuity.
Let f : R R be a function. For any x R, f may or may not be continuous at x. We say f is simply continuous if it is continuous at x for every x R.3u
Here are some basic and useful properties of continuous functions. (Of course
we cannot prove them until we give a formal definition of continuous functionsu.)
Fact 4.5. a) Every constant function is continuous at every c R.
b) The identity function I(x) = x is continuous at every c R.
c) If f and g are continuous at x = c, then f + g and f g are continuous at x = c.
d) If f is continuous at x = c and f (c) 6= 0, then f1 is continuous at x = c.
e) If f is continuous at x = c and g is continuous at x = f (c), then g f is
continuous at x = c.
From this relatively small number of facts many other facts follow. For instance,
since polynomials are built up out of the identity function and the constant functions by repeated addition and multiplication, it follows that all polynomials are
continuous at every c R. Similarly, every rational function fg is continuous at
every c in its domain, i.e., at all points c such that g(c) 6= 0.
We now wish to define the limit of a function f at a point c R. Here it is
crucial to remark that c need not be in the domain of f . Rather what we need
is that f is defined on some deleted interval Ic, about c: that is, there is some
> 0 such that all points in (c , c + ) except possibly at c, f is defined. To see
that this a necessary business, consider the basic limit defining the derivative:
f (x + h) f (x)
.
h
Here x is fixed and we are thinking of the difference quotient as a function of h.
Note though that this function is not defined at h = 0: the denominator is equal
to zero. In fact what we are trying to when differentiating is to find the most reasonable extension of the right hand side to a function which is defined at 0. This
brings us to the following definition.
f 0 (x) = lim

h0

Let f be a real-valued function defined on some subset D of R such that (c

, c) (c, c + ) is contained in D. Let L be a real number. Then limxc f (x) = L
if the function f with domain (c , c + ) and defined as f (x) = f (x), x 6= c,
f (c) = L, is continuous at c.
Thus the limit L of a function as x c is a value that if we plug the hole
in the graph of the function y = f (x) by setting f (c) = L, we get a graph which is
3The concept of continuity also makes sense for functions with domain a proper subset of R,
but lets save this for later.

58

4. CONTINUITY AND LIMITS

continuous just think nicely behaved, for now at x = c.

Note that an immediate consequence of the definition is that if f (x) is itself continuous at x = c, then it is already defined at c and limxc f (x) = f (c). Thus the
limit of a continuous function at a point is simply the value of the function at that
point. This is very important!
In fact, we can now give a better account of what we have been doing when we
(x)
which is defined
compute f 0 (x). We start with the difference quotient f (x+h)f
h
for all sufficiently small h but not for h = 0. Then we manipulate / simplify the
difference quotient until we recognize it as being equal, for all h 6= 0, to some new
function g(h) which is continuous at zero. (For instance, when f (x) = x2 , that
new function was g(h) = 2x + h.) Then the limiting value is obtained simply by
plugging in h = 0, i.e., it is g(0).
4. Continuity Done Right
4.1. The formal defininition of continuity.
Let D R and f : D R be a function. For c R we say f is continuous
at c if for all  > 0 there exists > 0 such that (c , c + ) D and for all x with
|x c| < , |f (x) f (c)| < .
Morover we say that f is continuous if it is continuous at c for all c in its domain D.
The bit about the domain D is traditionally left a bit more implicit, but since
we are trying to be precise we may as well be completely precise. The condition
is equivalent to requiring that there be some > 0 such that f is defined on
(c , c + ) and that for all  > 0, whenever we speak of > 0 we always take it
as implicit that .
In general, mathematical statements become more complex and harder to parse
the more alternating quantifiers they have: i.e., statements of the form There exists x such that P (x) or For all x, P (x) have a simple logical structure (if P (x)
is itself something reasonably simple, of course). Statements of the form For all x,
there exists y such that P (x, y) and There exists x such that for all y, P (x, y)
are a bit more complex; the untrained mind must stop to remind itself that they
are not logically equivalent: e.g. if x and y are real numbers and P (x, y) is x > y
then the first statement is true for every real number, there exists a greatest
real number and the second statement is false there is no real number which is
greater than every eal number. The - definition of continuity has three alternating
quantifiers: for all, then there exists, then for all. In general, to fully comphrehend
the meaning of statements this logically complex takes serious mental energy.
Let us first talk about the geometric meaning of the statement: the inequality
|f (x) f (c)| <  means f (c)  < f (x) < f (c) + : that is, it determines a horizontal strip centered at the horizontal line y = f (c) of width 2. Similarly, the
inequality |xc| < means c < x < c+, so determines a vertical strip centered
at the vertical line x = c of width 2. Thus the statement is saying something about

4. CONTINUITY DONE RIGHT

59

approximating both the y-values and the x-values of the function.

Now let us talk about the logical meaning of the statement and the sequence of
quantifiers. We may think of it as a game: the first player chooses any  > 0 and
thereby lays down a horizontal strip bounded above and below by the lines f (c) + 
and f (c) . The second player chooses a > 0. Moreover, the second player wins
if by restricting to x values lying between the vertical lines c and c+, the graph
of the function is trapped between the two vertical lines f (c) ; otherwise the first
player wins. Now the assertion that f is continuous at x = c is equivalent to the
fact that the second player has a winning strategy: in other words, it is possible
for her to win no matter which  > 0 the first player names.
Example 4.1: Constant functins f (x) = C are continuous.
...
Example 4.2: The identity function I(x) = x is continuous.
...
Example 4.3: Linear functions f (x) = mx + b are continuous.
...
Example 4.4: f (x) = x2 is continuous.
...
4.2. Basic properties of continuous functions.
Lemma 4.6. (Upper and Lower Bounds for Continuous Functions)
Let f be continuous at x = c.
a) For any  > 0, there exists > 0 such that |x c| < implies |f (x)| |f (c)| + .
b) Suppose f (c) 6= 0. Then for any (0, 1), there exists > 0 such that |xc| <
implies |f (x)| |f (c)|.
Proof. a) For any  > 0, there exists > 0 such that |x c| < implies
|f (x) f (c)| < . By the Reverse Triangle Inequality,
|f (x)| |f (c)| |f (x) f (c)| < ,
so
|f (x)| |f (c)| + .
b) We will prove the result for = 12 , leaving the general case as an extra credit
problem. There exists > 0 such that |x c| < implies |f (x) f (c)| < |f (c)|
2 .
Again the Reverse Triangle Inequality implies
|f (c)|
|f (c)| |f (x)| |f (x) f (c)| <
,
2
or
|f (c)|
|f (c)|
|f (x)| > |f (c)|
=
.
2
2

Theorem 4.7. Let f and g be functions and c R.
a) If f is continuous at c and A R, then Af is continuous at c.
b) If f and g are both continuous at c then f + g is continuous at c.

60

4. CONTINUITY AND LIMITS

c) If f and g are both continuous at c then f g is continuous at c.

d) If f and g are both continuous at c and g(c) 6= 0, then fg is continuous at c.
e) If f is continuous at c and g is continuous at f (c) then g f is continuous at c.
Proof. For each part we work out the idea of the proof first and then translate
it into a formal - argument.
a) Fix  > 0. We must show that there exists > 0 such that |x c| < implies
|Af (x) Af (c)| < . but |Af (x) Af (c)| = |A||f (x) f (c)|. Moreover, precisely
because f is continuous at c we may make the quantity |f (x) f (c) as small as
we like by taking x sufficiently close to c. A quantity which we can make as small
as we like times a constant can still be made as small as we like!
Now formally: we may assume A 6= 0 for othewise Af is the constantly zero
function, which we have already proved is continuous. For any  > 0, since f is
continuous at x = c there exists > 0 such that |x c| < implies |f (x) f (c)| <

|A| . (Note what is being done here: by continuity, we can make |f (x) f (c)| less
than any positive number we choose. It is convenient for us to make it smaller than

|A| , where  is a previously given positive number.) Then |x c| < implies
|Af (x) Af (c)| = |A||f (x) f (c)| < |A|


= .
|A|

b) Fix  > 0. We must show that there exists > 0 such that |x c| < implies
|f (x) + g(x) (f (c) + g(c))| < . Now
|f (x)+g(x)(f (c)+g(c))| = |(f (x)f (c))+(g(x)g(c))| |f (x)f (c)|+|g(x)g(c)|.
This is good: since f and g are both continuous at c, we can make each of |f (x)
f (c)| and |g(x) g(c)| as small as we like by taking x sufficiently close to c. The
sum of two quantities which can each be made as small as we like can be made as
small as we like!
Now formally: choose 1 > 0 such that |x c| < 1 implies |f (x) f (c)| < 2 .
Choose 2 > 0 such that |x c| < 2 implies |g(x) g(c)| < 2 . Let = min(1 , 2 ).
Then |x c| < implies |x c| < 1 and |x c| < 2 , so


|f (x) + g(x) (f (c) + g(c))| |f (x) f (c)| + |g(x) g(c)| < + = .
2 2
c) Fix  > 0. We must show that there exists > 0 such that |x c| < implies
|f (x)g(x) f (c)g(c)| < . The situation here is somewhat perplexing: clearly we
need to use the continuity of f and g at c, and to do this it stands to reason that we
should be estimating |f (x)g(x)f (c)g(c)| in terms of |f (x)f (c)| and |g(x)g(c)|,
but unfortunately we do not yet see these latter two expressions. So we force them
to appear by adding and subtracting f (x)g(c):
|f (x)g(x) f (c)g(c)| = |f (x)g(x) f (x)g(c) + f (x)g(c) f (c)g(c)|
|f (x)||g(x) g(c)| + |g(c)||f (x) f (c)|.
This is much better: |g(c)||f (x) f (c)| is a constant times something which can
be made arbitrarily small, so can be made arbitarily small. Moreover in the term
|f (x)||g(x) g(c)| we can make |g(x) g(c)| arbitarily small by taking x sufficiently
close to c, and then, by continuity of f , |f (x)| gets arbitrarily close to |f (c)|. So
|f (x)| is nonconstant but bounded, and something which is bounded times something
which can be made arbitarily small can be made arbitrarily small!
Now formally:

4. CONTINUITY DONE RIGHT

61

Using Lemma 4.6a) and taking  = 1, there exists 1 > 0 such that |x c| < 1
implies |f (x)| |f (c)| + 1. There exists 2 > 0 such that |x c| < 2 implies

|g(x) g(c)| < 2(|f (c)|+1)
. Finally, there exists 3 > 0 such that |x c| < 3 implies

|f (x) f (c)| < 2|g(c)| . (Here we are assuming that g(c) 6= 0. If g(c) = 0 then we
simply dont have the second term in our expression and the argument is similar
but easier.) Taking = min 1 , 2 , 3 , for |x c| < then |x c| is less than 1 , 2
and 3 so
|f (x)g(x) f (c)g(c)| |f (x)||g(x) g(c)| + |g(c)||f (x) f (c)|




< (|f (c)| + 1)
+ |g(c)|
= + = .
2(|f (c)| + 1)
2|g(c)|
2 2
d) Since fg = f g1 , in light of part c) it will suffice to show that if g is continuous
at c and g(c) 6= 0 then g1 is continuous at c. Fix  > 0. We must show that there
exists > 0 such that |x c| < implies
1
1
|

| < .
g(x) g(c)
Now
1
|g(c) g(x)|
|g(x) g(c)|
1

|=
=
.
|
g(x) g(c)
|g(x)||g(c)|
|g(x)||g(c)|
Since g is continuous at x = c, we can make the numerator |g(x) g(c)| as small
as we like by taking x sufficiently close to c. This will make the entire fraction as
small as we like provided the denominator is not also getting arbitrarily small as x
approaches c. But indeed, since g is continuous at c and g(c) 6= 0, the denominator
is approaching |g(c)|2 6= 0. Thus again we have a quantity which we can make
arbitarily small times a bounded quantity, so it can be made arbitrarily small!
Now formally:
We apply Lemma 4.6b) with = 12 : there exists 1 > 0 such that |x c| < 1
implies |g(x)| |g(c)|
and thus also
2
2
1

.
|g(x)||g(c)|
|g(c)|2
Also there exists 2 > 0 such that |x c| < 2 implies |g(x) g(c)| <

|g(c)|2
2

.

Taking = min(1 , 2 , |x c| < implies





1
1
1
2
|g(c)|2
|

|=
|g(x) g(c)| <
 = .
g(x) g(c)
|g(x)||g(c)|
|g(c)|2
2
e) Fix  > 0. Since g(y) is continuous at y = f (c), there exists > 0 such that
|y f (c)| < implies |g(y) g(f (c))| < . Moreover, since f is continuous at c,
there exists > 0 such that |xc| < implies |f (x)f (c)| < . Thus, if |xc| < ,
|f (x) f (c)| = |y f (c)| < and hence
|g(f (x)) g(f (c))| = |g(y) g(f (c))| < .

Corollary 4.8. All rational functions are continuous.
Proof. Since rational functions are built out of constant functions and the
identity by repeated addition, multiplication and division, this follows immediately
from Theorem 4.7.


62

4. CONTINUITY AND LIMITS

Other elementary functions: unfortunately if we try to go beyond rational functions to other elementary functions familiar from precalculus, we run into the issue
that we have not yet given complete, satisfactory definitions
of these functions! For

instance, take even the relatively innocuous f (x) = x. We want this function to
have domain [0, ), but this uses the special property of R that every non-negative
number has a square root: we havent proved this yet! If > 0is irrational we
have not given any definition of the power function x . Similarly we do not yet
have rigorous definitions of ax for a > 1, log x, sin x and cos x, so we are poorly
placed to rigorously prove their continuity. However (following Spivak) in order
so as not to drastically limit the supply of functions to appear in our examples
and exercises, we will proceed for now on the assumption that all the above
elementary functions are continuous. We hasten to make two remarks:
Remark 4.5: This assumption can be justified! That is, all the elementary functions above are indeed continuous atevery point of their domain (with the small
proviso that for power functions like x we will need to give a separate definition
of continuity at an endpoint of an interval, coming up soon). And in fact we will
prove this later in the course...much later.
Remark 4.6: We will not use the continuity of the elementary functions as an
assumption in any of our main results (but only in results and examples explicitly
involving elementary functions; e.g. we will use the assumed continuity of the sine
function to differentiate it). Thus it will be clear that we are not arguing circularly
when we finally prove the continuity of these functions.
5. Limits Done Right
5.1. The Formal Definition of a Limit.
In order to formally define limits, it is convenient to have the notion of a deleted
interval Ic, about a point c, namely a set of real numbers of the form
0 < |x c| <
for some > 0. Thus Ic consists of (c , c) together with the points (c, c + ),
or more colloquially it contains all points sufficiently close to c but not equal to c.
Now comes the definition. For real numbers c and L and a function f : D R R,
we say limxc f (x) = L if for every  > 0 there exists > 0 such that for all x in
the deleted interval Ic, i.e., for all x with 0 < |x c| < f is defined at x and
|f (x) L| < .
Among all the many problems of limits, perhaps the following is the most basic
and important.
Theorem 4.9. The limit at a point is unique (if it exists at all): that is, if L
and M are two numbers such that limxc f (x) = L and limxc f (x) = M , then
L = M.
Proof. Seeking a contradiction, we suppose L 6= M ; it is no loss of generality
to suppose that L < M (otherwise switch L and M ) and we do so. Now we take
 = M 2L in the definition of limit: since limxc f (x) = L, there exists 1 > 0 such

5. LIMITS DONE RIGHT

63

that 0 < |x c| < 1 implies |f (x) L| < M 2L ; and similarly, since limxc f (x) =
M , there exists 2 > 0 such that 0 < |x c| < 2 implies |f (x) M | < M 2L .
Taking = min(1 , 2 , then, as usual, for 0 < |x c| < we get both inequalities:
M L
|f (x) L| <
2
M L
.
|f (x) M | <
2
However these inequalities are contradictory! Before we go further we urge the
reader to draw a picture to see that the vertical strips defined by the two inequalities above are disjoint: they have no points in common. Let us now check
this formally: since |f (x) L| < M 2L , f (x) < L + M 2L = M 2+L . On the other
hand, since |f (x) M | < M 2L , f (x) > M M 2L = M 2+L . Clearly there is not
a single value of x such that f (x) is at the same time greater than and less than
M +L
2 , let alone a deleted interval around c of such values of x, so we have reached
a contradiction. Therefore L = M .

We have now given a formal definition of continuity at a point and also a formal
definition of limits at a point. Previously though we argued that each of limits and
continuity can be defined in terms of the other, so we are now in an overdetermined situation. We should therefore check the compatibility of our definitions.
Theorem 4.10. Let f : D R R, and let c R.
a) f is continuous at x = c if and only if f is defined at c and limxc f (x) = f (c).
b) limxc f (x) = L iff f is defined in some deleted interval Ic, around x and,
defining f on (c , c + ) by f (x) = f (x), x 6= c, f (c) = L makes f continuous at
x = c.
Proof. All the pedagogical advantage here comes from working through this
yourself rather than reading my proof, so I leave it to you.

5.2. Basic Properties of Limits.
Most of the basic properties of continuity discussed above have analogues for limits.
We state the facts in the following result.
Theorem 4.11. Let f and g be two functions defined in a deleted interval Ic,
of a point c. We suppose that limxc f (x) = L and limxc g(x) = M .
a) For any constant A, limxc Af (x) = AL.
b) We have limxc f (x) + g(x) = L + M .
c) We have limxc f (x)g(x) = LM .
(x)
L
d) If M 6= 0, then limxc fg(x)
=M
.
We leave the proof to the reader. It is possible to prove any/all of these facts in
one of two ways: (i) by rephrasing the definition of limit in terms of continuity and
appealing to Theorem 4.7 above, or (ii) adapting the proofs of Theorem 4.7 to the
current context.
Now what about limits of composite functions? The natural analogue of Theorem 4.7e) above would be the following:
If limxc f (x) = L and limxL g(x) = M , then limxc g(f (x)) = M .

64

4. CONTINUITY AND LIMITS

Unfortunately the above statement is not always true!

Example 5.1: Let f (x) = 0 (constant function). Let g(x) be equal to 1 for x 6= 0
and g(0) = 0. Take c = 0. Then limx0 f (x) = 0 and limx0 g(x) = 1, but for all
x R, g(f (x)) = g(0) = 0, so limx0 g(f (x)) = 0.
This result can be repaired by requiring the continuity of the outside function
g(x). Indeed the proof of the following result is almost identical to that of Theorem
4.7e) above in which f is also continuous at c: in fact, it shows that the continuity
of the inside function f was not really used.
Theorem 4.12. Let f, g be functions with limxc f (x) = L and g continuous
at L. Then limxc g(f (x)) = g(L).
Proof. Fix  > 0. Since limyf (c) g(y) = g(L), there exists > 0 such that
|y f (c)| < implies |g(y) g(L)| < . Since limxc f (x) = L, there exists
> 0 such that 0 < |x c| < implies |y L| = |f (x) L| < and thus
|g(f (x)) g(L)| = |g(y) g(L)| < .

One may rephrase Theorem 4.12 as: if g is continuous and limxc f (x) exists, then
lim g(f (x)) = g( lim f (x)).

xc

xc

In other words, one can pull a limit through a continuous function. In this form
the result is actually a standard one in freshman calculus.
It happens that one can say exactly when the above statement about limits of
composite funtions holds. I dont plan on mentioning this in class and you neednt
keep it in mind or even read it, but I recently learned that this question has a rather
simple answer so I might as well record it here so I dont forget it myself.
Theorem 4.13. (Marjanovic [MK09]) Suppose limxc f (x) = L and limxL g(x) =
M . The following are equivalent:
(i) limxc g(f (x)) = M .
(ii) At least one of the following holds:
a) g is continuous at L.
b) There exists > 0 such that for all 0 < |x c| < , f (x) 6= L.
Proof. (i) = (ii): We will argue by contradiction: suppose that neither a)
nor b) holds; we will show that limxc g(f (x)) 6= M . Indeed, since b) dooes not
hold, for every > 0 there exists x with 0 < |x c| < such that f (x) = L. For
such x we have g(f (x)) = g(L). But since a) does not hold, g is not continuous at
L, i.e., M 6= g(L). Thus g(f (x)) = g(L) 6= M . Taking  = |g(L) M | this shows
that there is no > 0 such that 0 < |x c| < implies |g(f (x)) M | < , so
limxc g(f (x)) 6= M .
(ii) = (i). The case in which a) holds i.e., g is continuous at L is precisely
Theorem 4.12. So it suffices to assume that b) holds: there exists some > 0 such
that 0 < |x c| < implies f (x) 6= L. Now fix  > 0; since limxL g(x) = M ,
there exists > 0 such that 0 < |y L| < implies |g(y) M | < . Similarly (and
familiarly), since limxc f (x) = L, there exists 1 > 0 such that 0 < |x c| < 1
implies |f (x)L| < . Here is the point: for 0 < |xc| < 1 , we have |f (x)L| < .

5. LIMITS DONE RIGHT

65

If in addition 0 < |f (x) L| < , then we may conclude that |g(f (x)) M | < .
So our only concern is that pehaps f (x) = L for some c with 0 < |x c| < 1 ,
and this is exactly what the additional hypothesis b) allows us to rule out: if we
take = min(1 , ) then 0 < |x c| < implies 0 < |f (x) L| < and thus
|g(f (x)) M | < .

Remark 5.2: The nice expository article [MK09] gives some applications of the
implication (ii)b) = (i) of Theorem 4.13 involving making an inverse change of
variables to evaluate a limit. Perhaps we may revisit this point towards the end of
the course when we talk about inverse functions.
5.3. The Squeeze Theorem and the Switching Theorem.
Theorem 4.14. (Squeeze Theorem) Let m(x), f (x) and M (x) be defined on
some deleted interval I = (c , c + ) {c} about x = c. We suppose that:
(i) For all x I , m(x) f (x) M (x), and
(ii) limxc m(x) = limxc M (x) = L.
Then limxc f (x) = L.
Proof. Fix  > 0. There exists 1 > 0 such that 0 < |x c| < 1 implies
|m(x) L| <  and 2 > 0 such that 0 < |x c| < 2 implies |M (x) L| < . Let
= min(1 , 2 ). Then 0 < |x c| < implies
f (x) M (x) < L + 
and
f (x) m(x) > L ,
so L  < f (x) < L + , or equivalently |f (x) L| < .

sin( x1 ),

x 6= 0 and
Example 5.3: For 0, define f : R R by f (x) = x
f (0) = 0. By our assumption about the continuity of sign and our results on
continuity of rational functions and compositions of continuous functions, f is
continuous at all x 6= 0. We claim that f is continuous at x = 0 iff > 0. . . .
Example 5.4: The function f defined above is differentiable at x = 0 iff > 1. . . .
Example 5.5: limx0 sinx x = 1.
Solution: This will be a 17th century solution to the problem: i.e., we will assume that the trigonometric functions are continuous and use geometric reasoning,
including properties of angles and arclength. (Nevertheless, this solution seems a
lot better than not giving any proof at all until much later in the course...)
Consider the unit circle and the point on it Px = (cos x, sin x). There is a right
triangle T1 with vertices (0, 0), (cos x, 0), Px . This right triangle is contained in
the circular sector determined by all points on or inside the unit circle with angle
between 0 and x. In turn this circular sector is contained in a second right triangle
T2 , with vertices (0, 0), (1, 0), (1, tan x). Now let us write down the inequalities expressing the fact that since T1 is contained in the circular sector which is contained
in T2 , the area of T1 is less than or equal to the area of the circular sector, which
is less than or equal to the area of T2 .
The area of T1 is (one half the base times the height)

1
2

cos x sin x. The area of

66

4. CONTINUITY AND LIMITS

x
the circular sector is 2
times the area of the unit circle, or
1
1 sin x
of T2 is 2 tan x = 2 cos x . This gives us the inequalities

x
2

x
2.

The area

1
1
1 sin x
cos x sin x x
,
2
2
2 cos x
or equivalently, for x 6= 0,
x
1

.
sin x
cos x
Taking reciprocals this inequality is equivalent to
1
sin x

cos x.
cos x
x
Now we may apply the Squeeze Theorem: since cos x is continuous at 0 and takes
the value 1 6= 0 there, we have
1
= lim cos x = 1.
lim
x0
x0 cos x
Therefore the Squeeze Theorem implies
sin x
(16)
lim
= 1.
x0 x
cos x

x
Example 5.6: We will evaluate limx0 1cos
= 0. The idea is to use trigonometric
x
identities to reduce this limit to an expression involving the limit (16). Here goes:


1 cos x
1 cos x 1 + cos x
cos2 x 1
sin2 x
lim
= lim
= lim
= lim
x0
x0
x0 x(1 + cos x)
x0 x(1 + cos x)
x
x
1 + cos x



sin x
sin x
0
= lim
) = 0.
lim
=1(
x0 x
x0 1 + cos x
2
x
Of course we also have limx0 cos xx1 = limx0 1cos
= 0 = 0. In summary:
x

(17)

lim

x0

cos x 1
1 cos x
= lim
= 0.
x0
x
x

Before doing the next two examples we remind the reader of the composite angle formulas from trigonometry: for any real numbers x, y,
sin(x + y) = sin x cos y + cos x sin y,
cos(x + y) = cos x cos y sin x sin y.
Example 5.7: If f (x) = sin x, then we claim f 0 (x) = cos x. Indeed
f (x + h) f (x)
sin(x + h) sin x
sin x cos h + cos x sin h sin x
f 0 (x) = lim
= lim
= lim
h0
h0
h0
h
h
h




1 cos h
sin h
= sin x lim
+ cos x lim
= ( sin x) 0 + (cos x) 1 = cos x.
h0
h0 h
h
Example 5.8: If f (x) = cos x, then we claim f 0 (x) = sin x. Indeed
f (x + h) f (x)
cos(x + h) cos x
= lim
=
h0
h
h
cos x cos h sin x sin h cos x
lim
h0
h

f 0 (x) = lim

h0

5. LIMITS DONE RIGHT


= cos x

67




1 cos h
sin h
sin x lim
= ( cos x)0+( sin x)1 = sin x.
h0
h0 h
h
lim

Theorem 4.15. (Switching Theorem) Consider three functions f, g1 , g2 defined

on some deleted interval Ic, of x = c. We suppose that:
(i) limxc g1 (x) = limxc g2 (x) = L.
(ii) For all x with 0 < |x c| < , either f (x) = g1 (x) or f (x) = g2 (x).
Then limxc f (x) = L.
Proof. Fix  > 0. Let 1 > 0 be such that 0 < |xc| < 1 implies |g1 (x)L| <
, let 2 > 0 be such that 0 < |x c| < 2 implies |g2 (x) L| < , and let
= min(1 , 2 ). Let x be such that 0 < |x c| < . Then either f (x) = g1 (x), in
which case |f (x) L| = |g1 (x) L| < , or f (x) = g2 (x), in which case |f (x) L| =
|g2 (x) L| < . Either way, |f (x) L| < !

Example 5.9: Let f (x) be defined as x for rational x and x for irrational x.
We may apply the Switching Theorem to show that f is continuous at 0. Indeed, put g1 (x) = x and g2 (x) = x. Then limx0 g1 (x) = limx0 g2 (x) = 0
and for all x, f (x) = g1 (x) or f (x) = g2 (x). So by the Switching Theorem,
f (0) = 0 = limx0 f (x).
Remark 5.10: The previous example shows that a function may be very strangely
behaved and still be continuous at a point. It is thus worth emphasizing that we are
not really interested in functions which are continuous at certain points, but rather
at functions which are continuous at every point of their domain. Such functions
have many pleasant properties that we will prove later on in the course.
Remark 5.11: The Switching Theorem is not a standard result. That is to say,
I came up with it myself, inspired by the homework problem asking for a function
which is continuous at a single point. Although I do not claim it is in the same
league as the venerated Squeeze Theorem, I do plan to use it later on to give a
proof of the Chain Rule which is (I think) simpler than the one Spivak gives.
5.4. Variations on the Limit Concept.
Finally we consider three variations on the limit concept: one-sided limits, infinite limits, and limits at infinity. There is nothing really novel going on here, but
we need to be sure that we can adapt our - formalism to the variant notions of
limit that apply in calculus.

By way of introducing the first variant, consider the function f (x) = x. We

have said earlier that we are assuming for now that this function is continuous on
its entire domain. But that statement glossed over a technicality which we now
address. Namely, the domain of f is [0, ). So f cannot be continuous at 0 according to the definition that we gave because it is not defined on any open interval
containing zero. Instead it is defined, for instance, on an interval of the form [0, )
for > 0: i.e., this contains all points sufficiently close to 0 but greater than or
equal to zero.
p
A similar phenomenon arises when we consider the function f (x) = x(1 x),
which has natural domain [0, 1]. The function is defined at 1 but not in any open
interval containing 1: only intervals of the form (1 , 1].

68

4. CONTINUITY AND LIMITS

This brings us to our definition. We say that a function f : D R is right

continuous at x = c if for every  > 0 there exists > 0 such that f is defined on
[c, c + ) and x [c, c + ) = |f (x) f (c)| < . Similarly, we say that a function
f : D R is left continuous at x = c if for every  > 0 there exists > 0 such
that f is defined on (c , c] and x (c , c] = |f (x) f (c)| < .
Finally we make the following definition: let I R be any interval and let f : I R
be a function. We say that a is a left endpoint of I if a I and there is no x I
with x < a; similarly we say b is a right endpoint of I if b I and there is
no x I with x > b. An interval I has at most one endpoint and at most one
endpoint; all four possibilities of having / not having left / right endpoints are of
course possible. Let us say c I is an interior point if it is not a left endpoint or
a right endpoint. Now we say that f : I R is continuous if:
f is continuous at c for each interior point c I,
If I has a left endpoint a, then f is right continuous at a, and
If I had a right endpoint b, then f is left continuous at b.
Example: f (x) =

x is right continuous at x = 0.

As above, it is necessary to require left/right continuity when discussing behavior at right/left endpoints of an interval. On the other hand one may still discuss
left/right continuity at interior points of an interval, and it is sometimes helpful to
do so.
Example: Let f (x) = bxc be the greatest integer function. Then f is continuous at c for all c R \ Z, whereas for any c Z, f is right continuous but not left
continuous at c.
This example suggests the following simple result.
Proposition 4.16. For a function f : D R R and c D, the following
are equivalent:
(i) f is left continuous at c and right continuous at c.
(ii) f is continuous at c.
We leave the proof to the reader.
In a similar way we can define one-sided limits at a point c.
We say limxc f (x) = L and read this as the limit as x approaches c from
the left of f (x) is L if for all  > 0 there exists > 0 such that for all x with
c < x < c + , |f (x) L| < .
We say limxc+ f (x) = L and read this as the limit as x approaches c from
the right of f (x) is L if for all  > 0, there exists > 0 such that for all x with
c < x < c, |f (x) L| < .
Proposition 4.17. For a function f : D R R and c D, the following
are equivalent:

5. LIMITS DONE RIGHT

69

(i) The left hand and right hand limits at c exist and are equal.
(ii) limxc f (x) exists.
Again we leave the proof to the reader.
Example: Let f (x) = bxc be the greatest integer function, and let n Z. Then
limxn f (x) = n 1 and limxn+ f (x) = n, so f is not continuous at n.
There is some terminology here not essential, but sometimes useful. If for a
function f and c R, the left and right hand limits at c both exist but are unequal,
we say that f has a jump discontinuity at c. If the left and right hand limts at
c both exist and are equal i.e., if limxc f (x) = L exists but still f (x) is not
continuous c (this can happen if either f (c) 6= L or, more plausibly, if c is not in
the domain of f ) we say that f has a removable discontinuity at c. This terminology comes from our earlier observation that if we (re)define f at c to be the
limiting value L then f becomes continuous at c. One sometimes calls a discontinuity which is either removable or a jump discontinuity a simple discontinuity: i.e.,
this is the case whenever both one-sided limits exist at c but f is not continuous at c.
Infinite limits: Consider limx0 x12 . This limit does not exist: indeed, if it did,
then there would be some deleted interval I0, on which f is bounded, whereas
just the opposite is happening: the closer x is to 0, the larger f (x) becomes. In
freshman calculus we would say limx0 f (x) = . And we still want to say that,
but in order to know what we mean when we say this we want to give an - style
definition of this. Here it is:
We say limxc f (x) = if for all M R, there exists > 0 such that 0 <
|x c| < = f (x) > M .
Geometrically, this is similar to the - definition of limit, but instead of picking two horizontal lines arbitrarily close to y = L, we pick one horizontal line which
is arbitrarily large and require that on some small deleted interval the graph of
y = f (x) always lie above that line. Similarly:
We say limxc f (x) = if for all m R, there exists > 0 such that 0 <
|x c| < implies f (x) < m.
Example: Let us indeed prove that limx0 x12 = . Fix M R. We need to
find such that 0 < |x| < implies x12 > M . It is no loss of generality to assume
M > 0 (why?). Then x12 > M |x| < 1M , so we may take = 1M .

CHAPTER 5

Differentiation
1. Differentiability Versus Continuity
Recall that a function f : D R R is differentiable at a D if
f (a + h) f (a)
h0
h
lim

exists, and when this limit exists it is called the derivative f 0 (a) of f at a. Moreover, the tangent line to y = f (x) at f (a) exists if f is differentiable at a and is
the unique line passing through the point (a, f (a)) with slope f 0 (a).
Note that an equivalent definition of the derivative at a is
lim

xa

f (x) f (a)
.
xa

One can see this by going to the - definition of a limit and making the substitution h = x a: then 0 < |h| < 0 < |x a| < .
Theorem 5.1. Let f : D R R be a function, and let a D. If f is
differentiable at a, then f is continuous at a.
Proof. We have
f (x) f (a)
(x a)
xa



f (x) f (a) 
lim x a = f 0 (a) 0 = 0.
= lim
xa
xa
xa
lim f (x) f (a) = lim

xa

xa

Thus
0 = lim (f (x) f (a)) =
xa


lim f (x) f (a),

xa

so
lim f (x) = f (a).

xa


The converse of Theorem 5.1 is far from being true: a function f which is continuous at a need not be differentiable at a. An easy example is f (x) = |x| at a = 0.
In fact the situation is worse: a function f : R R can be continuous everywhere yet still fail to be differentiable at many points. One way of introducing
points of non-differentiability while preserving continuity is to take the absolute
value of a differentiable function.
71

72

5. DIFFERENTIATION

Theorem 5.2. Let f : D R R be continuous at a D.

a) Then |f | is continuous at a.
b) The following are equivalent:
(i) f is differentiable at a, and either f (a) 6= 0 or f (a) = f 0 (a) = 0.
(ii) |f | is differentiable at a.
Exercise: Prove Theorem 8.31.
2. Differentiation Rules
2.1. Linearity of the Derivative.
Theorem 5.3. (Constant Rule) Let f be differentiable at a R and C R.
Then the function Cf is also differentiable at a and
(Cf )0 (a) = Cf 0 (a).
Proof. There is nothing to it:
(Cf )(a + h) (Cf )(a)
=C
h0
h

(Cf )0 (a) = lim

f (a + h) f (a)
h0
h
lim

= Cf 0 (a).


Theorem 5.4. (Sum Rule) Let f and g be functions which are both differentiable at a R. Then the sum f + g is also differentiable at a and
(f + g)0 (a) = f 0 (a) + g 0 (a).
Proof. Again, no biggie:
(f + g)(a + h) (f + g)(a)
f (a + h) f (a) g(a + h) g(a)
= lim
+
h0
h
h
h
f (a + h) f (a)
g(a + h) g(a)
= lim
+ lim
= f 0 (a) + g 0 (a).
h0
h0
h
h


(f +g)0 (a) = lim

h0

These results, simple as they are, have the following important consequence.
Corollary 5.5. (Linearity of the Derivative) For any differentiable functions
f and g and any constants C1 , C2 , we have
(C1 f + C2 g)0 = C1 f 0 + C2 g 0 .
2.2. Product Rule(s).
Theorem 5.6. (Product Rule) Let f and g be functions which are both differentiable at a R. Then the product f g is also differentiable at a and
(f g)0 (a) = f 0 (a)g(a) + f (a)g 0 (a).
Proof.

f (a + h)g(a + h) f (a)g(a)
h
f (a + h)g(a + h) f (a)g(a + h) + (f (a)g(a + h) f (a)g(a))
= lim
h0
h





f (a + h) f (a)
g(a + h) g(a)
= lim
lim g(a + h) + f (a) lim
.
h0
h0
h0
h
h
(f g)0 (a) = lim

h0

2. DIFFERENTIATION RULES

73

Since g is differentiable at a, g is continuous at a and thus limh0 g(a + h) =

limx ag(x) = g(a). The last expression above is therefore equal to
f 0 (a)g(a) + f (a)g 0 (a).

Dimensional analysis and the product rule: Leibniz kept a diary in which he
recorded his scientific discoveries, including his development of differential calculus.
One day he got to the product rule and wrote a formula equivalent to
(f g)0 = f 0 g 0 .
But this entire entry of the diary is crossed out. Three days later there is a new
entry with a correct statement of the product rule and the derivation, which Leibniz
says he has known for some time.
I confess that I have taken this story from a third-party account. Every once
in a while I think about trying to verify it, but I am always stopped by the fact
that if it turns out to be apocryphal, I dont really want to know: its too good
a story! It is fundamentally honest about the way mathematics is done and, especially, the way new mathematics is created. That is, to go forward we make
guesses. Often we later realize that our guesses were not only wrong but silly, to
the extent that we try to hide them from others and perhaps even from ourselves.
But the process of silly guesswork is vital to mathematics, and if the great genius Gottfried von Leibniz was not above them, what chance do the rest of us have?
It is natural, like Leibniz, to want to hide our silly guesses. However, they play an
important role in teaching and learning. Many veteran instructors of higher mathematics come to lament the fact that for the sake of efficiency and coverage of the
syllabus we often give the student the correct answer within minutes or seconds of
asking the question (or worse, do not present the question at all!) and thus deprive
them of the opportunity to make a silly guess and learn from their mistake. What
can we learn from Leibnizs guess that (f g)0 = f 0 g 0 ? Simply to write down the
correct product rule is not enough: I want to try to persuade you that Leibnizs
guess was not only incorrect but truly silly: that in some sense the product rule
could have turned out to be any number of things, but certainly not (f g)0 = f 0 g 0 .
For this I want to follow an approach I learned in high school chemistry: dimensional analysis. We can give physically meaningful dimensions (the more
common, but less precise, term is units) to all our quantities, and then our formulas must manipulate these quantities in certain ways or they are meaningless.
For a simple example, suppose you walk into an empty room and find on the
blackboard a drawing of a cylinder with its radius labelled r and its height labelled
h. Below it are some formulas, and this one jumps out at you: 2(rh + r3 ). You
can see right away that something has gone wrong: since r and h are both lengths
say in meters rh denotes a product of lengths say, square meters. But r3 is
a product of three lengths i.e., cubic meters. How on earth will we get anything
physically meaningful by adding square meters to cubic meters? It is much more
likely that the writer meant 2(rh + r2 ), since that is a meaningful quantity of dimension length squared (i.e., area). In fact, it happens to be the correct formula for

74

5. DIFFERENTIATION

the surface area of the cylinder with the top and bottom faces included, although
dimensional considerations wont tell you that.
Something similar can be applied to the formula (f g)0 = f 0 g 0 . Let f = f (t)
and view t as being time, say measured in seconds. Since f 0 (t) is a limit of quo(t)
tients f (t+h)f
, its dimension is the dimension of f divided by time. Now suppose
h
both f (t) and g(t) are lengths, say in meters. Then f 0 (t) and g 0 (t) both have units
meters per second, so f 0 (t) g 0 (t) has units meters squared per second squared.
On the other hand, the units of (f g)0 are meters squared per second. Thus the
formula (f g)0 = f 0 g 0 is asserting that some number of meters squared per second
is equal to some number of meters squared per second squared. Thats not only
wrong, its a priori meaningless.
By contrast the correct formula (f g)0 = f 0 g + f g 0 makes good dimensional
sense: as above, (f g)0 is meters squared per second; so is f 0 g and so is f g 0 , so we
can add them to get a meaningful number of meters squared per second. Thus the
formula at least makes sense, which is good because above we proved it to be correct.
Taking these ideas more seriously suggests that we should look for a proof of the
product rule which explicitly takes into account that both sides are rates of change
of areas. This is indeed possible, but we omit it for now.
Suppose we want to find the derivative of a function which is a product of not
two but three functions whose derivatives we already know, e.g. f (x) = x sin xex .
We can of course? still use the product rule, in two steps:
f 0 (x) = (x sin xex )0 = ((x sin x)ex )0 = (x sin x)0 ex + (x sin x)(ex )0
= (x0 sin x + x(sin x)0 )ex + x sin xex = sin x + x cos xex + x sin xex .
Note that we didnt use the fact that our three differentiable functions were x,
sin x and ex until the last step, so the same method shows that for any three
functions f1 , f2 , f3 which are all differentiable at x, the product f = f1 f2 f3 is also
differentiable at a and
f 0 (a) = f10 (a)f2 (a)f3 (a) + f1 (a)f20 (a)f3 (a) + f1 (a)f2 (a)f30 (a).
The following result rides this train of thought to its final destination.
Theorem 5.7. (Generalized Product Rule) Let n 2 be an integer, and let
f1 , . . . , fn be n functions which are all differentiable at a. Then f = f1 fn is also
differentiable at a, and
(18)

(f1 fn )0 (a) = f10 (a)f2 (a) fn (a) + . . . + f1 (a) fn1 (a)fn0 (a).

Proof. By induction on n.
Base Case (n = 2): This is precisely the ordinary Product Rule (Theorem 5.6).
Induction Step: Let n 2 be an integer, and suppose that the product of any n
functions which are each differentiable at a R is differentiable at a and that the
derivative is given by (18). Now let f1 , . . . , fn , fn+1 be functions, each differentiable
at a. Then by the usual product rule
0
(f1 fn fn+1 )0 (a) = ((f1 fn )fn+1 )0 (a) = (f1 fn )0 (a)fn+1 (a)+f1 (a) fn (a)fn+1
(a).

Using the induction hypothesis this last expression becomes

0
(f10 (a)f2 (a) fn (a) + . . . + f1 (a) fn1 (a)fn0 (a)) fn+1 (a)+f1 (a) fn (a)fn+1
(a)

2. DIFFERENTIATION RULES

75

0
= f10 (a)f2 (a) fn (a)fn+1 (a) + . . . + f1 (a) fn (a)fn+1
(a).

Example: We may use the Generalized Product Rule to give a less computationally
intensive derivation of the power rule
(xn )0 = nxn1
for n Z+ . Indeed, taking f1 = = fn = x, we have f (x) = xn = f1 fn , so
applying the Generalized Power rule we get
(xn )0 = (x)0 x x + . . . + x x(x)0 .
Here in each term we have x0 = 1 multiplied by n 1 factors of x, so each term
evaluates to xn1 . Moreover we have n terms in all, so
(xn )0 = nxn1 .
No need to dirty our hands with binomial coefficients!
Theorem 5.8. (Generalized Leibniz Rule) Let n Z+ , and let f, g be functions
which are each n times differentiable at a R. Then f g is is n times differentiable
at a and
n  
X
n (k) (nk)
(f g)(n) =
f g
.
k
k=0

Exercise: Prove Theorem 5.8.

Theorem 5.9. (Quotient Rule) Let f and g be functions which are both differentiable at a R, with g(a) 6= 0. Then fg is differentiable at a and
 0
f
g(a)f 0 (a) f (a)g 0 (a)
(a) =
.
g
g(a)2
Proof. Step 0: First observe that since g is continuous and g(a) 6= 0, there is
some interval I = (a , a + ) about a on which g is nonzero, and on this interval
f
g is defined. Thus it makes sense to consider the difference quotient
f (a + h)/g(a + h) f (a)/g(a)
h
for h sufficiently close to zero.
Step 1: We first establish the Reciprocal Rule, i.e., the special case of the Quotient Rule in which f (x) = 1 (constant function). Then
 0
1
1
1
g(a+h) g(a)
(a) = lim
h0
g
h



g(a) g(a + h)
g(a + h) g(a)
1
g 0 (a)
= lim
= lim
lim
=
.
h0 hg(a)g(a + h)
h0
h0 g(a)g(a + h)
h
g(a)2
We have again used the fact that if g is differentiable at a, g is continuous at a.
Step 2: We now derive the full Quotient Rule by combining the Product Rule and
the Reciprocal Rule. Indeed, we have
 0

0
 0
f
1
1
1
0
(a) = f
(a) = f (a)
+ f (a)
(a)
g
g
g(a)
g
=

f 0 (a)
g 0 (a)
g(a)f 0 (a) g 0 (a)f (a)
f (a)
=
.
2
g(a)
g(a)
g(a)2


76

5. DIFFERENTIATION

Lemma 5.10. Let f : D R R. Suppose:

(i) limxa f (x) exists, and
(ii) There exists a number L R such that for all > 0, there exists at least one x
with 0 < |x a| < such that f (x) = L.
Then limxa f (x) = L.
Proof. Left as an exercise. (Suggestion: suppose limxa f (x) = M 6= L, and
derive a contradiction by taking  to be small enough compared to |M L|.)

Theorem 5.11. (Chain Rule) Let f and g be functions, and let a R be such
that f is differentiable at a and g is differentiable at f (a). Then the composite
function g f is differentiable at a and
(g f )0 (a) = g 0 (f (a))f 0 (a).
Proof. Motivated by Leibniz notation, it is tempting to argue as follows:
 


f (x) f (a)
g(f (x)) g(f (a))
g(f (x)) g(f (a))
= lim

(g f )0 (a) = lim
xa
xa
xa
f (x) f (a)
xa



g(f (x)) g(f (a))
f (x) f (a)
= lim
lim
xa
xa
f (x) f (a)
xa



g(f (x)) g(f (a))
f (x) f (a)
=
lim
lim
= g 0 (f (a))f 0 (a).
xa
f (x) f (a)
xa
f (x)f (a)
The replacement of limxa . . . by limf (x)f (a) . . . in the first factor above is
justified by the fact that f is continuous at a. However, this argument has a gap:
when we multiply and divide by f (x) f (a), how do we know that we are not
dividing by zero?? The answer is that we cannot rule this out: it is possible for
f (x) to take the value f (a) on arbitarily small deleted intervals around a: again,
this is exactly what happens for the function f (x) of the above example near a = 0.
I maintain that the above gap can be mended so as to give a complete proof.
The above argument is valid unless for all > 0, there is x with 0 < |x a| <
such that f (x) f (a) = 0. In this case, it follows from Lemma 5.10 that if
f (x) f (a)
xa
exists at all, it must be equal to 0. But we are assuming the above limit exists, since
we are assuming f is differentiable at a. So f 0 (a) = 0, and therefore, since the Chain
Rule reads (g f )0 (a) = g 0 (f (a))f 0 (a), here we are trying to show (g f )0 (a) = 0.
For x R we have two possibilities: the first is f (x) f (a) = 0, in which case
also g(f (x)) g(f (a)) = g(f (a)) g(f (a)) = 0, so the difference quotient is zero at
these points. The second is f (x) f (a) 6= 0, in which case
lim

xa

g(f (x)) g(f (a)) =

g(f (x)) g(f (a)) f (x) f (a)

f (x) f (a)
xa

holds, and the above argument shows this expression tends to g 0 (f (a))f 0 (a) = 0
(a))
as x a. So whichever holds, the difference quotient g(f (x))g(f
is close to (or
xa
equal to!) zero. Thus the limit tends to zero no matter which alternative obtains.
Somewhat more formally, if we fix  > 0, then the first step of the argument
shows that there is > 0 such that for all x with 0 < |x a| < such that
(a))
f (x) f (a) 6= 0, | g(f (x))g(f
| < . On the other hand, when f (x) f (a) = 0,
xa

3. OPTIMIZATION

77

(a))
then | g(f (x))g(f
| = 0, so it is certainly less than ! Therefore, all in all we have
xa
(a))
| < , so that
0 < |x a| < = | g(f (x))g(f
xa

lim

xa

g(f (x)) g(f (a))

= 0 = g 0 (f (a))f 0 (a).
xa

3. Optimization

3.1. Intervals and interior points.

At this point I wish to digress to formally define the notion of an interval on
the real line and and interior point of the interval. . . .
3.2. Functions increasing or decreasing at a point.
Let f : D R be a function, and let a be an interior point of D. We say that f is
increasing at a if for all x sufficiently close to a and to the left of a, f (x) < f (a)
and for all x sufficiently close to a and to the right of a, f (x) > f (a). More formally
phrased, we require the existence of a > 0 such that:
for all x with a < x < a, f (x) < f (a), and
for all x with a < x < a + , f (x) > f (a).
We say f is decreasing at a if there exists > 0 such that:
for all x with a < x < a, f (x) > f (a), and
for all x with a < x < a + , f (x) < f (a).
We say f is weakly increasing at a if there exists > 0 such that:
for all x with a < x < a, f (x) f (a), and
for all x with a < x < a + , f (x) f (a).
Exercise: Give the definition of f is decreasing at a.
Exercise: Let f : I R, and let a be an interior point of I.
a) Show that f is increasing at a iff f is decreasing at a.
b) Show that f is weakly increasing at a iff f is weakly decreasing at a.
Example: Let f (x) = mx + b be the general linear function. Then for any a R:
f is increasing at a iff m > 0, f is weakly increasing at a iff m 0, f is decreasing
at a iff m < 0, and f is weakly decreasing at a iff m 0.
Example: Let n be a positive integer, let f (x) = xn . Then:
If x is odd, then for all a R, f (x) is increasing at a.
If x is even, then if a < 0, f (x) is decreasing at a, if a > 0 then f (x) is increasing
at a. Note that when n is even f is neither increasing at 0 nor decreasing at 0
because for every nonzero x, f (x) > 0 = f (0).1
1We do not stop to prove these assertions as it would be inefficient to do so: soon enough

we will develop the right tools to prove stronger assertions. But when given a new definition, it is
always good to find ones feet by considering some examples and nonexamples of that definition.

78

5. DIFFERENTIATION

If one looks back at the previous examples and keeps in mind that we are supposed to be studying derivatives (!), one is swiftly led to the following fact.
Theorem 5.12. Let f : I R. Suppose f is differentiable at a I .
a) If f 0 (a) > 0, then f is increasing at a.
b) If f 0 (a) < 0, then f is decreasing at a.
c) If f 0 (a) = 0, then no conclusion can be drawn: f may be increasing at a, decreasing at a, or neither.
Proof. a) We use the - interpretation of differentiability at a to our advantage. Namely, take  = f 0 (a): there exists > 0 such that for all x with
(a)
f 0 (a)| < f 0 (a), or equivalently
0 < |x a| < , | f (x)f
xa
0<

f (x) f (a)
< 2f 0 (a).
xa

(a)
> 0, so: if x > a, f (x)f (a) >
In particular, for all x with 0 < |xa| < , f (x)f
xa
0, i.e., f (x) > f (a); and if x < a, f (x) f (a) < 0, i.e., f (x) < f (a).
b) This is similar enough to part a) to be best left to the reader as an exercise.2
c) If f (x) = x3 , then f 0 (0) = 0 but f is increasing at 0. If f (x) = x3 , then
f 0 (0) = 0 but f is decreasing at 0. If f (x) = x2 , then f 0 (0) = 0 but f is neither
increasing nor decreasing at 0.


3.3. Extreme Values.

Let f : D R. We say M R is the maximum value of f on D if
(MV1) There exists x D such that f (x) = M , and
(MV2) For all x D, f (x) M .
It is clear that a function can have at most one maximum value: if it had more
than one, one of the two would be larger than the other! However a function need
not have any maximum value: for instance f : (0, ) R by f (x) = x1 has no
maximum value: limx0+ f (x) = .
Similarly, we say m R is the minimum value of f on D if
(mV1) There exists x D such that f (x) = m, and
(mV2) For all x D, f (x) m.
Also, a function can have at most one minimum value but need not have any. The
function f : R \ {0} R by f (x) = x1 has no minimum value: limx0 f (x) = .
Exercise: For a function f : D R, the following are equivalent:
(i) f assumes a maximum value M , a minimum value m, and M = m.
(ii) f is a constant function.
Recall that f : D R is bounded above if there is B R such that for all
x D, f (x) B. A function is bounded below if there is b R such that for all
x D, f (x) b. A function is bounded if it is both bounded above and bounded
2Two ways to go: (i) Revisit the above proof flipping inequalities as appropriate. (ii) Use the
fact that f is decreasing at a iff f is increasing at a and f 0 (a) < 0 (f )0 (a) > 0 to apply
the result of part a).

3. OPTIMIZATION

79

below: equivalently, there exists B 0 such that for all x D, |f (x)| B: i.e.,
the graph of f is trapped between the horizontal lines y = B and y = B.
Exercise: Let f : D R be a function.
a) Show: if f has a maximum value, it is bounded above.
b) Show: if f has a minimum value, it is bounded below.
Exercise: a) If a function has both a maximum and minimum value on D, then it
is bounded on D: indeed, if M is the maximum value of f and m is the minimum
value, then for all x B, |f (x)| max(|m|, |M |).
b) Give an example of a bounded function f : R R which has neither a maximum
nor a minimum value.
We say f assumes its maximum value at a if f (a) is the maximum value
of f on D, or in other words, for all x D, f (x) f (a). Simlarly, we say f
assumes its minimum value at a if f (a) is the minimum value of f on D, or in
other words, for all x D, f (x) f (a).
Example: The function f (x) = sin x assumes its maximum value at x = 2 , because sin 2 = 1, and 1 is the maximum value of the sine function. Note however
that 2 is not the only x-value at which f assumes its maximum value: indeed, the
sine function is periodic and takes value 1 precisely at x = 2 + 2n for n Z.
Thus there may be more than one x-value at which a function attains its maximum
3
value. Similarly f attains its minimum value at x = 3
2 f ( 2 ) = 1 and f takes
3
no smaller values and also at x = 2 + 2n for n Z.
Example: Let f : R R by f (x) = x3 + 5. Then f does not assume a maximum or minimum value. Indeed, limx f (x) = and limx f (x) = .
Example: Let f : [0, 2] R be defined as follows: f (x) = x + 1, 0 x < 1.
f (x) = 1, x = 1.
f (x) = x 1, 1 < x 2.
Then f is defined on a closed, bounded interval and is bounded above (by 2) and
bounded below (by 0) but does not have a maximum or minimum value. Of course
this example of a function defined on a closed bounded interval without a maximum
or minimum value feels rather contrived: in particular it is not continuous at x = 1.
This brings us to one of the most important theorems in the course.
Theorem 5.13. (Extreme Value Theorem) Let f : [a, b] R be continuous.
Then f has a maximum and minimum value, hence is bounded above and below.
The proof will be given in the next chapter, following a discussion of completeness.
Ubiquitously in (pure and applied) mathematics we wish to optimize functions:
that is, find their maximum and or minimum values on a certain domain. Unfortunately, as we have seen above, a general function f : D R need not have a
maximum or minimum value! But the Extreme Value Theorem gives rather mild
hypotheses on which these values are guaranteed to exist, and in fact is a useful
tool for establishing the existence of maximia / minima in other situations as well.

80

5. DIFFERENTIATION

Example: Let f : R R be defined by f (x) = x2 (x 1)(x 2). Note that f

does not have a maximum value: indeed limx f (x) = limx = . However,
we claim that f does have a minimum value. We argue for this as follows: given
that f tends to with |x|, there must exist > 0 such that for all x with |x| > ,
f (x) 1. On the other hand, if we restrict f to [, ] we have a continuous
function on a closed bounded interval, so by the Extreme Value Theorem it must
have a minimum value, say m. In fact since f (0) = 0, we see that m < 0, so in
particular m < 1. This means that the minimum value m for f on [, ] must
in fact be the minimum value for f on all of R, since at the other values namely,
on (, ) and (, ), f (x) > 1 > 0 m.
We can be at least a little more explicit: a sign analysis of f shows that f
is positive on (, 1) and (2, ) and negative on (1, 2), so the minimum value
of f will be its minimum value on [1, 2], which will be strictly negative. But exactly what is this minimum value m, and for which x value(s) does it occur? Stay
tuned:we are about to develop tools to answer this question!
3.4. Local Extrema and a Procedure for Optimization.
We now describe a type of local behavior near a of a very different sort from
being increasing or decreasing at a.
Let f : D R be a function, and let a D. We say that f has a local
maximum at a if the value of f at a is greater than or equal to its values at
all sufficiently close points x. More formally: there exists > 0 such that for all
x D, |x a| < = f (x) f (a). Similarly, we say that f has a local
minimum at a if the vaalue of f at a is greater than or equal to its values at
all sufficiently close points x. More formally: there exists > 0 such that for all
x D, |x a| < = f (x) f (a).
Theorem 5.14. Let f : D R, and let a be an interior point of a. If f is
differentiable at a and has a local extremum i.e., either a local minimum or a
local maximum at x = a, then f 0 (a) = 0.
Proof. Indeed, if f 0 (a) 6= 0 then either f 0 (a) > 0 or f 0 (a) < 0.
If f (a) > 0, then by Theorem 5.12, f is increasing at a. Thus for x slightly smaller
than a, f (x) < f (a), and for x slightly larger than a, f (x) > f (a). So f does not
have a local extremum at a.
Similarly, if f 0 (a) < 0, then by Theorem 5.12, f is decreasing at a. Thus for x
slightly smaller than a, f (x) > f (a), and for x slightly larger than a, f (x) < f (a).
So f does not have a local extremum at a.

0

Theorem 5.15. (Optimization Theorem) Let f : I R be continuous. Suppose

that f attains a minimum or maximum value at x = a. Then a is either:
(i) an endpoint of I,
(ii) a stationary point: f 0 (a) = 0, or
(iii) a point of nondifferentiability.
Proof. This is an immediate consequence of Theorem 5.14.
Exercise: State and prove a version of Theorem 5.15 for the maximum value.

4. THE MEAN VALUE THEOREM

81

Often one lumps cases (ii) and (iii) of Theorem 5.15 together under the term critical point (but there is nothing very deep going on here: its just terminology).
Clearly there are always exactly two endpoints. In favorable circustances there
will be only finitely many critical points, and in very favorable circumstances they
can be found exactly: suppose they are c1 , . . . , cn . (There may not be any critical
points, but that only makes things easier...) Suppose further that we can explicitly
compute all the values f (a), f (b), f (c1 ), . . . , f (cn ). Then we win: the largest of
these values is the maximum value, and the smallest is the minimum value.
Example: Let f (x) = x2 (x 1)(x 2) = x4 3x3 + 2x2 . Above we argued
that there is a such that |x| > = |f (x)| 1: lets find such a explicitly.
We intend nothing fancy here:
f (x) = x4 3x2 + 2x2 x4 3x3 = x3 (x 3).
So if x 4, then
x3 (x 3) 43 1 = 64 1.
On the other hand, if x < 1, then x < 0, so 3x3 > 0 and thus
f (x) x4 + 2x2 = x2 (x2 + 2) 1 3 = 3.
Thus we may take = 4.
Now let us try the procedure of Theorem 5.15 out by finding the maximum and
minimum values of f (x) = x4 3x3 + 2x2 on [4, 4].
Since f is differentiable everywhere on (4, 4), the only critical points will be the
stationary points, where f 0 (x) = 0. So we compute the derivative:
f 0 (x) = 4x3 9x2 + 4x = x(4x2 9x + 4).
The roots are x =

9 17
,
8

or, approximately,

x1 0.6094 . . . , x2 = 1.604 . . . .
f (x1 ) = 0.2017 . . . , f (x2 ) = 0.619 . . . .
Also we always test the endpoints:
f (4) = 480, f (4) = 96.
So the maximum value is 480 and the minimum value is .619 . . ..
4. The Mean Value Theorem
4.1. Statement of the Mean Value Theorem.
Our goal in this section is to prove the following important result.
Theorem 5.16. (Mean Value Theorem) Let f : [a, b] R be continuous on
[a, b] and differentiable on (a, b). Then there is at least one c with a < c < b and
f 0 (c) =

f (b) f (a)
.
ba

82

5. DIFFERENTIATION

Remark:3 I still remember the calculus test I took in high school in which I was
asked to state the Mean Value Theorem. It was a multiple choice question, and I
didnt see the choice I wanted, which was as above except with the subtly stronger
assumption that fR0 (a) and fL0 (b) exist: i.e., f is one-sided differentiable at both
endpoints. So I went up to the teachers desk to ask about this. He thought for
a moment and said, Okay, you can add that as an answer if you want, and so
as not to give special treatment to any one student, he announced to all that he
was adding a possible answer to the Mean Value Theorem question. So I marked
my added answer, did the rest of the exam, and then had time to come back to
this question. After more thought I decided that one-sided differentiability at the
endpoints was not required. So in the end I selected this pre-existing choice and
submitted my exam. As you can see, my final answer was correct. But many other
students figured that if I had successfully lobbied for an additional answer then
my answer was probably correct, so they changed their answer from the correct
answer to my added answer. They were not so thrilled with me or the teacher, but
in my opinion he behaved admirably: talk about a teachable moment!
One should certainly draw a picture to go with the Mean Value Theorem, as it
has a very simple geometric interpretation: under the hypotheses of the theorem,
there exists at least one interior point c of the interval such that the tangent line
at c is parallel to the secant line joining the endpoints of the interval.
And one should also interpret it physically: if y = f (x) gives the position of
(a)
is nothing less than the
a particle at a given time x, then the expression f (b)f
ba
average velocity between time a and time b, whereas the derivative f 0 (c) is the instantaneous velocity at time c, so that the Mean Value Theorem says that there is at
least one instant at which the instantaneous velocity is equal to the average velocity.
Example: Suppose that cameras are set up at checkpoints along an interstate
highway in Georgia. One day you receive timestamped photos of yourself at two
checkpoints. The two checkpoints are 90 miles apart and the second photo is taken
73 minutes after the first photo. You are issued a ticket for violating the speeed
limit of 70 miles per hour. Your average velocity was (90 miles) / (73 minutes)
(60 minutes) / (hour) 73.94 miles per hour. Thus, although no one saw you
violating the speed limit, they may mathematically deduce that at some point your
instantaneous velocity was over 70 mph. Guilt by the Mean Value Theorem!
4.2. Proof of the Mean Value Theorem.
We will deduce the Mean Value Theorem from the (as yet unproven) Extreme
Value Theorem. However, it is convenient to first establish a special case.
Theorem 5.17. (Rolles Theorem) Let f : [a, b] R. We suppose:
(i) f is continuous on [a, b].
(ii) f is differentiable on (a, b).
(iii) f (a) = f (b).
Then there exists c with a < c < b and f 0 (c) = 0.
Proof. By Theorem 5.13, f has a maximum M and a minimum m.
Case 1: Suppose M > f (a) = f (b). Then the maximum value does not occur
3Please excuse this personal anecdote.

4. THE MEAN VALUE THEOREM

83

at either endpoint. Since f is differentiable on (a, b), it must therefore occur at a

stationary point: i.e., there exists c (a, b) with f 0 (c) = 0.
Case 2: Suppose m < f (a) = f (b). Then the minimum value does not occur at
either endpoint. Since f is differentiable on (a, b), it must therefore occur at a
stationary point: there exists c (a, b) with f 0 (c) = 0.
Case 3: The remaining case is m = f (a) = M . Then f is constant and f 0 (c) = 0
at every point c (a, b).

To deduce the Mean Value Theorem from Rolles Theorem, it is tempting to tilt
our head until the secant line from (a, f (a)) to (b, f (b)) becomes horizontal and
then apply Rolles Theorem. The possible flaw here is that if we start a subset in
the plane which is the graph of a function and rotate it too much, it may no longer
be the graph of a function, so Rolles Theorem does not apply.
The above objection is just a technicality. In fact, it suggests that more is true:
there should be some version of the Mean Value Theorem which applies to curves
in the plane which are not necessarily graphs of functions. Nevertheless formalizing
this argument is more of a digression than we want to make, so the official proof
that follows is (slightly) different.
Proof of the Mean Value Theorem: Let f : [a, b] R be continuous on [a, b] and
differentiable on (a, b). There is a unique linear function L(x) such that L(a) = f (a)
and L(b) = f (b): indeed, L is nothing else than the secant line to f between (a, f (a))
and (b, f (b)). Heres the trick: by subtracting L(x) from f (x) we reduce ourselves
to a situation where we may apply Rolles Theorem, and then the conclusion that
we get is easily seen to be the one we want about f . Here goes: define
g(x) = f (x) L(x).
Then g is defined and continuous on [a, b], differentiable on (a, b), and g(a) =
f (a) L(a) = f (a) f (a) = 0 = f (b) f (b) = f (b) L(b) = g(b). Applying Rolles
Theorem to g, there exists c (a, b) such that g 0 (c) = 0. On the other hand, since
(a)
, we compute
L is a linear function with slope f (b)f
ba
0 = g 0 (c) = f 0 (c) L0 (c) = f 0 (c)

f (b) f (a)
,
ba

and thus
f 0 (c) =

f (b) f (a)
.
ba

4.3. The Cauchy Mean Value Theorem.

We present here a generalization of the Mean Value Theorem due to A.L. Cauchy.
Although perhaps not as fundamental and physically appealing as the Mean Value
Theorem, it has its place: later we will use it to prove LHopitals Rule.
Theorem 5.18. (Cauchy Mean Value Theorem) Let f, g : [a, b] R be continuous and differentiable on (a, b). Then there exists c (a, b) such that
(19)

(f (b) f (a))g 0 (c) = (g(b) g(a))f 0 (c).

Proof. Case 1: Suppose g(a) = g(b). By Rolles Theorem, there is c (a, b)

such that g 0 (c) = 0. For this c, both sides of (19) are zero, hence they are equal.

84

5. DIFFERENTIATION

Case 2: Suppose g(a) 6= g(b), and define



f (b) f (a)
h(x) = f (x)
g(x).
g(b) g(a)
Then h is continuous on [a, b], differentiable on (a, b), and
h(a) =

f (a)(g(b) g(a)) g(a)(f (b) f (a))

f (a)g(b) g(a)f (b)
=
,
g(b) g(a)
g(b) g(a)

f (b)(g(b) g(a)) g(b)(f (b) f (a))

f (a)g(b) g(a)f (b)
=
,
g(b) g(a)
g(b) g(a)
so h(a) = h(b).4 By Rolles Theorem there exists c (a, b) with


f (b) f (a)
0 = h0 (c) = f 0 (c)
g 0 (c),
g(b) g(a)
or equivalently,
(f (b) f (a))g 0 (c) = (g(b) g(a))f 0 (c).
h(b) =


Exercise: Which choice of g recovers the ordinary Mean Value Theorem?
The Cauchy Mean Value Theorem can be used to vindicate the head-tilting argument contemplated before the formal proof of the Mean Value Theorem. To do
so takes us a bit outside our wheelhouse (i.e., a seriously theoretical approach to
single variable calculus), but it is interesting enough to be worth sketching here.
First recall (if possible) the notion of a parameterized curve: this is given
by a function f with domain an interval I in R and codomain the plane R2 : we
may write v : t 7 (x(t), y(t)), where t 7 x(t) and y 7 y(t) are each functions
from I to R. We say that v is continuous at t if both x and y are continuous at t.
Further, when both x and y are differentiable at t we decree that v is differentiable
at t and set v 0 (t) = (x0 (t), y 0 (t)), which we think of as being a vector in the plane
with tail at the point (x(t), y(t)). When x0 (t) = y 0 (t) = 0 we get the zero vector,
which we dont want: lets call this a singular point. For any nonsingular point
(x(t), y(t)) we can define the tangent line to be the unique line passing through
the points (x(t), y(t)) and (x(t) + x0 (t), y(t) + y 0 (t)). (When x0 (t) 6= 0, we can
define the tangent line in a more familiar way, as the unique line passing through
0
(x(t), y(t)) with slope xy 0 (t)
(t) ; the above definition is phrased so as to make sense also
if x0 (t) = 0 and y 0 (t) 6= 0, in which case we get a vertical tangent line.)
Example: Let v : [0, 2] R2 by v(t) = (cos t, sin t). This point traces out a
path which winds once around the unit circle counterclockwise. We have v 0 (t) =
( sin t, cos t): since for all t, ( sin t)2 + (cos t)2 = 1, in particular v 0 (t) is never
zero: there are no singular points. At the points t = 0, , 2, x0 (t) = 0 and the
tangent line is vertical at the points (1, 0) (as you would certainly expect from
contemplation of the unit circle). When t = 0, 2 , , 3
2 , 2, one of (x(t), y(t)) and
(x0 (t), y 0 (t)) is horizontal and the other is vertical, so in particular the tangent line
at v(t) is perpendicular to v(t). This holds at all other points as well, e.g. by noting
y(t)
sin t
that the slope of the radial line from (0, 0) to v(t)) is x(t)
= cos
t and the slope of
4Dont be so impressed: we wanted a constant C such that if h(x) = f (x) Cg(x), then
h(a) = h(b), so we set f (a) Cg(a) = f (b) Cg(b) and solved for C.

5. MONOTONE FUNCTIONS

85

cos t
the tangent line to v(t) is xy 0 (t)
(t) = sin t and observing that these two slopes are
opposite reciprocals. Thus we get a derivation using calculus of a familiar (or so I
hope!) fact from elementary geometry.

Here is a version of the Mean Value Theorem for nonsingular parameterized curves.
Theorem 5.19. (Parameteric Mean Value Theorem) Let v : I R2 , t 7
(x(t), y(t)), be a nonsingular parametrized curve in the plane, i.e., for all t I,
x0 (t) and y 0 (t) both exist and are not both 0. Let a < b I, and suppose that either
x(a) 6= x(b) or y(a) 6= y(b). Then there is c (a, b) such that the tangent vector
(x0 (c), y 0 (c)) at c is parallel to the secant line from (a, f (a)) to (b, f (b)).
Proof. We apply the Cauchy Mean Value Theorem with f = x and g = y:
there is c (a, b) such that
(x(b) x(a))y 0 (c) = (y(b) y(a))x0 (c).
By assumption, x0 (c) and y 0 (c) are not both 0.
Case 1: Suppose x0 (c) 6= 0. Then we must have x(b) x(a) 6= 0: if not,
0 = (x(b) x(a))y 0 (c) = (y(b) y(a))x0 (c)
and thus, since x0 (c) 6= 0, y(b) y(a) 6= 0, and then we have x(a) = x(b) and
y(a) = y(b), contrary to our hypothesis. So we may divide to get
y(b) y(a)
y 0 (c)
=
,
x0 (c)
x(b) x(a)
which says that the tangent line to v at c has the same slope as the secant line
between (x(a), y(a)) and (x(b), y(b)), hence these two lines are parallel.
Case 2: Suppose x0 (c) = 0, i.e., the tangent line to v at c is vertical. By our
nonsingularity assumption y 0 (c) 6= 0, so
0 = (y(b) y(a))x0 (c) = (x(b) x(a))y 0 (c)
implies x(a) = x(b), so the secant line from (x(a), x(b)) to (y(a), y(b)) is vertical,
hence the two lines are parallel.

Conversely, it is possible (indeed, similar) to deduce the Cauchy Mean Value Theorem from the Parametric Mean Value Theorem: try it. Thus really we have one
theorem with two moderately different phrasings, and indeed it is common to also
refer to Theorem 5.19 as the Cauchy Mean Value Theorem.
5. Monotone Functions
A function f : I R is monotone if it is weakly increasing or weakly decreasing.
5.1. The Monotone Function Theorems.
The Mean Value Theorem has several important consequences. Foremost of all
it will be used in the proof of the Fundamental Theorem of Calculus, but thats for
later. At the moment we can use it to establish a criterion for a function f to be
monotone on an interval in terms of sign condition on f 0 .

86

5. DIFFERENTIATION

Theorem 5.20. (First Monotone Function Theorem) Let I be an open interval,

and let f : I R be a function which is differentiable on I.
a) Suppose f 0 (x) > 0 for all x I. Then f is increasing on I: for all x1 , x2 I
with x1 < x2 , f (x1 ) < f (x2 ).
b) Suppose f 0 (x) 0 for all x I. Then f is weakly increasing on I: for all
x1 , x2 I with x1 < x2 , f (x1 ) f (x2 ).
c) Suppose f 0 (x) < 0 for all x I. Then f is decreasing on I: for all x1 , x2 inI
with x1 < x2 , f (x1 ) > f (x2 ).
d) Suppose f 0 (x) 0 for all x I. Then f is weakly decreasing on I: for all
x1 , x2 I with x1 < x2 , f (x1 ) f (x2 ).
Proof. a) We go by contraposition: suppose that f is not increasing: then
there exist x1 , x2 I with x1 < x2 such that f (x1 ) f (x2 ). Apply the Mean Value
(x1 )
0.
Theorem to f on [x1 , x2 ]: there exists x1 < c < x2 such that f 0 (c) = f (xx22)f
x1
b) Again, we argue by contraposition: suppose that f is not weakly increasing: then
there exist x1 , x2 I with x1 < x2 such that f (x1 ) > f (x2 ). Apply the Mean Value
(x1 )
Theorem to f on [x1 , x2 ]: there exists x1 < c < x2 such that f 0 (c) = f (xx22)f
< 0.
x1
c),d) We leave these proofs to the reader. One may either proceed exactly as in
parts a) and b), or reduce to them by multiplying f by 1.

Corollary 5.21. (Zero Velocity Theorem) Let f : I R be a differentiable
function with identically zero derivative. Then f is constant.
Proof. Since f 0 (x) 0 for all x I, f is weakly increasing on I: x1 <
x2 = f (x1 ) f (x2 ). Since f 0 (x) 0 for all x I, f is weakly decreasing on
I: x1 < x2 = f (x1 ) f (x2 ). But a function which is weakly increasing and
weakly decreasing satsifies: for all x1 < x2 , f (x1 ) f (x2 ) and f (x1 ) f (x2 ) and
thus f (x1 ) = f (x2 ): f is constant.

Remark: Above, we deduced Corollary 5.21 from Theorem 5.20. If we had argued
directly from the Mean Value Theorem the proof would have been shorter: try it!
Corollary 5.22. Suppose f, g : I R are both differentiable and such that
f 0 = g 0 (equality as functions, i.e., f 0 (x) = g 0 (x) for all x I). Then there exists
a constant C R such that f = g + C, i.e., for all x I, f (x) = g(x) + C.
Proof. Let h = f g. Then h0 = (f g)0 = f 0 g 0 0, so by Corollary 5.21,
h C and thus f = g + h = g + C.

Remark: Corollary 5.22 can be viewed as a uniqueness theorem for differential equations. Let f : I R be a function, and consider the set of all functions F : I R
such that F 0 = f . Then Corollary 5.22 asserts that if there is a function F such
that F 0 = f , then there is a one-parameter family of such functions, and more
specifically that the general such function is of the form F + C.
On the other hand, the existence question lies deeper: namely, given f : I R,
must there exist F : I R such that F 0 = f ? In general the answer is no.
Exercise: Let f : R R by f (x) = 0 for x 0 and f (x) = 1 for x > 0. Show that
there is no function F : R R such that F 0 = f .
In other words, not every function f : R R has an antiderivative, i.e., is

5. MONOTONE FUNCTIONS

87

the derivative of some other function. It turns out that every continuous function
has an antiderivative: this will be proved later. (Much more subtly, there are also
some discontinuous functions which have antiderivatives...)
Corollary 5.23. Let f : I R be a function whose nth derivative f (n) is
identically zero. Then f is a polynomial function of degree at most n 1.
Proof. Exercise. (Hint: use induction.)

The setting of the Increasing Function Theorem is that of a differentiable function

defined on an open interval I. This is just a technical convenience: for continuous functions, the increasing / decreasing / weakly increasing / weakly decreasing
behavior on the interior of I implies the same behavior at an endpoint of I.
Theorem 5.24. Let f : [a, b] R be continuous at x = a and x = b.
a) If f is weakly increasing on (a, b), it is weakly increasing on [a, b].
b) If f is incereasing on (a, b), it is increasing on [a, b].
Proof. Step 1: Suppose that f is continuous at a and weakly increasing on
(a, b). We will show that f is weakly increasing on [a, b). Indeed, assume not: then
there exists x0 (a, b) such that f (a) > f (x0 ). Now take  = f (a) f (x0 ); since
f is (right-)continuous at a, there exists > 0 such that for all a x < a + ,
|f (x) f (a)| < f (a) f (x0 ), which implies f (x) > f (x0 ). By taking a < x < x0 ,
this contradicts the assumption that f is weakly increasing on (a, b).
Step 2: Suppose that f is continuous at a and increasing on (a, b). We will show
that f is increasing on [a, b). Note first that Step 1 applies to show that f (a) f (x)
for all x (a, b), but we want slightly more than this, namely strict inequality. So,
seeking a contradiction, we suppose that f (a) = f (x0 ) for some x0 (a, b). But
now take x1 (a, x0 ): since f is increasing on (a, b) we have f (x1 ) < f (x0 ) = f (a),
contradicting the fact that f is weakly increasing on [a, b).
Step 3: In a similar way one can handle the right endpoint b. Now suppose that
f is increasing on [a, b) and also increasing on (a, b]. It remains to show that f is
increasing on [a, b]. The only thing that could go wrong is f (a) f (b). To see that
this cannot happen, choose any c (a, b): then f (a) < f (c) < f (b).

Exercise: Show that we may replace each instance of increasing in Theorem 5.24
with decreasing (and still get a true statement!).
Theorem 5.25. (Second Monotone Function Theorem) Let f : I R be a
function which is continuous on I and differentiable on the interior I of I (i.e.,
at every point of I except possibly at any endpoints I may have).
a) The following are equivalent:
(i) f is monotone.
(ii) Either we have f 0 (x) 0 for all x I or f 0 (x) 0 for all x I .
b) Suppose f is monotone. The following are equivalent:
(i) f is not increasing or decreasing.
(ii) There exist a, b I with a < b such that the restriction of f to [a, b] is constant.
(iii) There exist a, b I with a < b such that f 0 (x) = 0 for all x [a, b].
Proof. Throughout the proof we restrict our attention to increasing / weakly
increasing functions, leaving the other case to the reader as a routine exercise.
a) (i) = (ii): Suppose f is weakly increasing on I. We claim f 0 (x) 0 for all
x I . If not, there is a I with f 0 (a) < 0. Then f is decreasing at a, so there

88

5. DIFFERENTIATION

exists b > a with f (b) < f (a), contradicting the fact that f is weakly decreasing.
(ii) = (i): Immediate from the Increasing Function Theorem and Theorem 5.24.
b) (i) = (ii): Suppose f is weakly increasing on I but not increasing on I. By
Theorem 5.24 f is still not increasing on I , so there exist a, b I with a < b
such that f (a) = f (b). Then, since f is weakly increasing, for all c [a, b] we have
f (a) f (c) f (b) = f (a), so f is constant on [a, b].
(ii) = (iii): If f is constant on [a, b], f 0 is identically zero on [a, b].
(iii) = (i): If f 0 is identically zero on some subinterval [a, b], then by the Zero
Velocity Theorem f is constant on [a, b], hence is not increasing.

The next result follows immediately.
Corollary 5.26. Let f : I R. Suppose f 0 (x) 0 for all x I, and
f 0 (x) > 0 except at a finite set of points x1 , . . . , xn . Then f is increasing on I.
Example: A typical application of Theorem 5.26 is to show that the function f :
R R by f (x) = x3 is increasing on all of R. Indeed, f 0 (x) = 3x2 which is strictly
positive at all x 6= 0 and 0 at x = 0.
5.2. The First Derivative Test.
We can use Theorem 5.24 to quickly derive another staple of freshman calculus.
Theorem 5.27. (First Derivative Test) Let I be an interval, a an interior
point of I, and f : I R a function. We suppose that f is continuous on I and
differentiable on I \ {a}. Then:
a) If there exists > 0 such that f 0 (x) is negative on (a , a) and is positive on
(a, a + ). Then f has a strict local minimum at a.
b) If there exists > 0 such that f 0 (x) is positive on (a , a) and is negative on
(a, a + ). Then f has a strict local maximum at a.
Proof. a) By the First Monotone Function Theorem, since f 0 is negative on
the open interval (a, a) and positive on the open interval (a, a+) f is decreasing
on (a , a) and increasing on (a, a + ). Moreover, since f is differentiable on its
entire domain, it is continuous at a , a and a + , and thus Theorem 5.24 applies
to show that f is decreasing on [a , a] and increasing on [a, a + ]. This gives
the desired result, since it implies that f (a) is strictly smaller than f (x) for any
x [a , a) or in (a, a + ].
b) As usual this may be proved either by revisiting the above argument or deduced
directly from the result of part a) by multiplying f by 1.

Remark: This version of the First Derivative Test is a little stronger than the
familiar one from freshman calculus in that we have not assumed that f 0 (a) = 0
nor even that f is differentiable at a. Thus for instance our version of the test
applies to f (x) = |x| to show that it has a strict local minimum at x = 0.
5.3. The Second Derivative Test.
Theorem 5.28. (Second Derivative Test) Let a be an interior point of an interval I, and let f : I R. We suppose:
(i) f is twice differentiable at a, and
(ii) f 0 (a) = 0.
Then if f 00 (a) > 0, f has a strict local minimum at a, whereas if f 00 (a) < 0, f has
a strict local maximum at a.

5. MONOTONE FUNCTIONS

89

Proof. As usual it suffices to handle the case f 00 (a) > a.

Notice that the hypothesis that f is twice differentiable at a implies that f is
differentiable on some interval (a, a+) (otherwise it would not be meaningful to
talk about the derivative of f 0 at a). Our strategy will be to show that for sufficiently
small > 0, f 0 (x) is negative for x (a , a) and positive for x (a, a + ) and
then apply the First Derivative Test. To see this, consider
f 0 (x) f 0 (a)
f 0 (x)
= lim
.
xa
xa x a
xa
We are assuming that this limit exists and is positive, so that there exists > 0
0
(x)
is positive. And this gives us exactly
such that for all x (a , a) (a, a + ), fxa
f 00 (a) = lim

what we want: suppose x (a , a). Then

f 0 (x)
xa

> 0 and x a < 0, so f 0 (x) < 0.

0

(x)
> 0 and x a > 0, so
On the other hand, suppose x (a, a + ). Then fxa
0
f (x) > 0. So f has a strict local minimum at a by the First Derivative Test. 

Remark: When f 0 (a) = f 00 (a) = 0, no conclusion can be drawn about the local
behavior of f at a: it may have a local minimum at a, a local maximum at a, be
increasing at a, decreasing at a, or none of the above.
5.4. Sign analysis and graphing.
When one is graphing a function f , the features of interest include number and
approximate locations of the roots of f , regions on which f is positive or negative,
regions on which f is increasing or decreasing, and local extrema, if any. For these
considerations one wishes to do a sign analysis on both f and its derivative f 0 .
Let us agree that a sign analysis of a function g : I R is the determination of regions on which g is positive, negative and zero.
The basic strategy is to determine first the set of roots of g. As discussed before,
finding exact values of roots may be difficult or impossible even for polynomial
functions, but often it is feasible to determine at least the number of roots and
their approximate location (certainly this is possible for all polynomial functions,
although this requires justification that we do not give here). The next step is to
test a point in each region between consecutive roots to determine the sign.
This procedure comes with two implicit assumptions. Let us make them explicit.
The first is that the roots of f are sparse enough to separate the domain I into
regions. One precise formulation of of this is that f has only finitely many roots
on any bounded subset of its domain. This holds for all the elementary functions we
know and love, but certainly not for all functions, even all differentiable functions:
we have seen that things like x2 sin( x1 ) are not so well-behaved. But this is a convenient assumption and in a given situation it is usually easy to see whether it holds.
The second assumption is more subtle: it is that if a function f takes a positive value at some point a and a negative value at some other point b then it must
take the value zero somewhere in between. Of course this does not hold for all
functions: it fails very badly, for instance, for the function f which takes the value

90

5. DIFFERENTIATION

1 at every rational number and 1 at every irrational number.

Let us formalize the desired property and then say which functions satisfy it.
A function f : I R has the intermediate value property if for all a, b I
with a < b and all L in between f (a) and f (b) i.e., with f (a) < L < f (b) or
f (b) < L < f (a) there exists some c (a, b) with f (c) = L.
Thus a function has the intermediate value property when it does not skip values.
Here are two important theorems, each asserting that a broad class of functions
has the intermediate value property.
Theorem 5.29. (Intermediate Value Theorem) Let f : [a, b] R be continuous. Then f has the intermediate value property.
Example of a continuous function f :[0, 2]Q Q failing
the intermediate value
property. Let f (x) be 1 for 0 x < 2 and f (x) = 1 for 2 < x 1.
The point of this example is to drive home the point that the Intermediate Value
Theorem is the second of our three hard theorems in the sense that we have no
chance to prove it without using special properties of the real numbers beyond the
ordered field axioms. And indeed we will not prove IVT right now, but we will use
it, just as we used but did not yet prove the Extreme Value Theorem. (However we
are now not so far away from the point at which we will switch back, talk about
completeness of the real numbers, and prove the three hard theorems.)
The Intermediate Value Theorem (IVT) is ubiquitously useful. Such innocuous
statements as every non-negative real number having a square root contain an implicit appeal to IVT. Further, IVT justifies the following observation.
Let f : I R be a continuous function, and suppose that there are only finitely
many roots, i.e., there are x1 , . . . , xn I such that f (xi ) = 0 for all i and f (x) 6= 0
for all other x I. Then I \ {x1 , . . . , xn } is a finite union of intervals, and on each
of them f has constant sign: it is either always positive or always negative.
So this is how sign analysis works for a function f when f is continuous a very
mild assumption. But as above we also want to do a sign analysis of the derivative
f 0 : how may we justify this?
Well, here is one very reasonable justification: if the derivative f 0 of f is itself
continuous, then by IVT it too has the intermediate value property and thus, at
least if f 0 has only finitely many roots on any bounded interval, sign analysis is
justified. This brings up the following basic question.
Question 5.30. Let f : I R be a differentiable function. Must its derivative
f 0 : I R be continuous?
Let us first pause to appreciate the subtlety of the question: we are not asking
whether f differentiable implies f continuous: we well know this is the case. Rather
we are asking whether the new function f 0 can exist at every point of I but fail to
itself be a continuous function. In fact the answer is yes.

5. MONOTONE FUNCTIONS

91

Example: Let f (x) = x2 sin( x1 ). I claim that f is differentiable on all of R but

that the derivative is discontinuous at x = 0, and in fact that limx0 f 0 (x) does
not exist. . . .
Theorem 5.31. (Darboux) Let f : I R be a differentiable function. Suppose
that we have a, b I with a < b and f 0 (a) < f 0 (b). Then for every L R with
f 0 (a) < L < f 0 (b), there exists c (a, b) such that f 0 (c) = L.
Proof. Step 1: First we handle the special case L = 0, which implies f 0 (a) < 0
and f 0 (b) > 0. Now f is a differentiable hence continuous function defined on
the closed interval [a, b] so assumes its minimum value at some point c [a, b]. If c
is an interior point, then as we have seen, it must be a stationary point: f 0 (c) = 0.
But the hypotheses guarantee this: since f 0 (a) < 0, f is decreasing at a, thus
takes smaller values slightly to the right of a, so the minimum cannot occur at a.
Similarly, since f 0 (b) > 0, f is increasing at b, thus takes smaller values slightly to
the left of b, so the minimum cannot occur at b.
Step 2: We now reduce the general case to the special case of Step 1 by defining
g(x) = f (x) Lx. Then g is still differentiable, g 0 (a) = f 0 (a) L < 0 and g 0 (b) =
f 0 (b) L > 0, so by Step 1, there exists c (a, b) such that 0 = g 0 (c) = f 0 (c) L.
In other words, there exists c (a, b) such that f 0 (c) = L.

Remark: Of course there is a corresponding version of the theorem when f (b) <
L < f (a). Darbouxs Theorem also often called the Intermediate Value Theorem For Derivatives, terminology we will understand better when we discuss
the Intermediate Value Theorem (for arbitrary continuous functions).
Exercise: Let a be an interior point of an interval I, and suppose f : I
a differentiable function. Show that the function f 0 cannot have a simple
continuity at x = a. (Recall that a function g has a simple discontinuity
if limxa g(x) and limxa+ g(x) both exist but either they are unequal to
other or they are unequal to g(a).)

R is
disat a
each

5.5. A Theorem of Spivak.

The following theorem is taken directly from Spivaks book (Theorem 7 of Chapter
11): it does not seem to be nearly as well known as Darbouxs Theorem (and in
fact I think I encountered it for the first time in Spivaks book).
Theorem 5.32. Let a be an interior point of I, and let f : I R. Suppose:
(i) f is continuous on I,
(ii) f is differentiable on I \ {a}, i.e., at every point of I except possibly at a, and
(iii) limxa f 0 (x) = L exists.
Then f is differentiable at a and f 0 (a) = L.
Proof. Choose > 0 such that (a , a + ) I. Let x (a, a + ). Then
f is differentiable at x, and we may apply the Mean Value Theorem to f on [a, x]:
there exists cx (a, x) such that
f (x) f (a)
= f 0 (cx ).
xa

92

5. DIFFERENTIATION

Now, as x a every point in the interval [a, x] gets arbitrarily close to x, so

limxa cx = x and thus
fR0 (a) = lim+
xa

f (x) f (a)
= lim+ f 0 (cx ) = lim+ f 0 (x) = L.
xa
xa
xa

By a similar argument involving x (a , a) we get

fL0 (a) = lim f 0 (x) = L,
x

so f is differentiable at a and f (a) = L.

6. Inverse Functions I: Theory

6.1. Review of inverse functions.
Let X and Y be sets, and let f : X Y be a function between them. Recall
that an inverse function is a function g : Y X such that
g f = 1X : X X, f g = 1Y : Y Y.
Lets unpack this notation: it means the following: first, that for all x X, (g
f )(x) = g(f (x)) = x; and second, that for all y Y , (f g)(y) = f (g(y)) = y.
Proposition 5.33. (Uniqueness of Inverse Functions) Let f : X Y be a
function. Suppose that g1 , g2 : Y X are both inverses of f . Then g1 = g2 .
Proof. For all y Y , we have
g1 (y) = (g2 f )(g1 (y)) = g2 (f (g1 (y))) = g1 (y).

Since the inverse function to f is always unique provided it exists, we denote it by
f 1 . (Caution: this has nothing to do with f1 . Thus sin1 (x) 6= csc(x) = sin1 x .)
We now turn to giving conditions for the existence of the inverse function. Recall that f : X Y is injective if for all x1 , x2 X, x1 6= x2 = f (x1 ) 6= f (x2 ).
In other words, distinct x-values get mapped to distinct y-values. (And in yet other
words, the graph of f satisfies the horizontal line test.) Also f : X Y is surjective if for all y Y , there exists at least one x X such that y = f (x).
Putting these two concepts together we get the important notion of a bijective
function f : X Y , i.e., a function which is both injective and surjective. Otherwise put, for all y Y there exists exactly one x X such that y = f (x). It may
well be intuitively clear that bijectivity is exactly the condition needed to guarantee
existence of the inverse function: if f is bijective, we define f 1 (y) = xy , the unique
element of X such that f (xy ) = y. And if f is not bijective, this definition breaks
down and thus we are unable to define f 1 . Nevertheless we ask the reader to bear
with us as we give a slightly tedious formal proof of this.
Theorem 5.34. (Existence of Inverse Functions) For f : X Y , TFAE:
(i) f is bijective.
(ii) f admits an inverse function.

6. INVERSE FUNCTIONS I: THEORY

93

Proof. (i) = (ii): If f is bijective, then as above, for each y X there

exists exactly one element of X say xy such that f (xy ) = y. We may therefore
define a function g : Y X by g(y) = xy . Let us verify that g is in fact the
inverse function of f . For any x X, consider g(f (x)). Because f is injective,
the only element x0 X such that f (x0 ) = f (x) is x0 = x, and thus g(f (x)) = x.
For any y Y , let xy be the unique element of X such that f (xy ) = y. Then
f (g(y)) = f (xy ) = y.
(ii) = (i): Suppose that f 1 exists. To see that f is injective, let x1 , x2 X
be such that f (x1 ) = f (x2 ). Applying f 1 on the left gives x1 = f 1 (f (x1 )) =
f 1 (f (x2 )) = x2 . So f is injective. To see that f is surjective, let y Y . Then
f (f 1 (y)) = y, so there is x X with f (x) = y, namely x = f 1 (y).

For any function f : X Y , we define the image of f to be {y Y | x X | y =
f (x)}. The image of f is often denoted f (X).5
We now introduce the dirty trick of codomain restriction. Let f : X Y
be any function. Then if we replace the codomain Y by the image f (X), we still
get a well-defined function f : X f (X), and this new function is tautologically
surjective. (Imagine that you manage the up-and-coming band Yellow Pigs. You
get them a gig one night in an enormous room filled with folding chairs. After
everyone sits down you remove all the empty chairs, and the next morning you
write a press release saying that Yellow Pigs played to a packed house. This is
essentially the same dirty trick as codomain restriction.)
Example: Let f : R R by f (x) = x2 . Then f (R) = [0, ), and although
x2 : R R is not surjective, x2 : R [0, ) certainly is.
Since a codomain-restricted function is always surjective, it has an inverse iff it
is injective iff the original functionb is injective. Thus:
Corollary 5.35. For a function f : X Y , the following are equivalent:
(i) The codomain-restricted function f : X f (X) has an inverse function.
(ii) The original function f is injective.
6.2. The Interval Image Theorem.
Next we want to return to earth by considering functions f : I R and their
inverses, concentrating on the case in which f is continuous.
Theorem 5.36. (Interval Image Theorem) Let I R be an interval, and let
f : I R be a continuous function. Then the image f (I) of f is also an interval.
Proof. For now we will give the proof when I = [a, b], i.e., is closed and
bounded. The general case will be discussed later.
Suppose f : [a, b] R is continuous. Then f has a minimum value m, say at
xm and a maximum value M , say at xM . Thus the image f ([a, b]) of f is a subset
of [m, M ]. Moreover, if L (m, M ), then by the Intermediate Value Theorem there
exists c in between xm and xM such that f (c) = L. So f ([a, b]) = [m, M ].

Exercise: Let I be a nonempty interval which is not of the form [a, b]. Let J be any
nonempty interval. Show: there is a continuous function f : I R with f (I) = J.
5This is sometimes called the range of f , but sometimes not. It is safer to call it the image!

94

5. DIFFERENTIATION

6.3. Monotone Functions and Invertibility.

Recall f : I R is strictly monotone if it is either increasing or decreasing.
Every strictly monotone function is injective. Therefore our dirty trick of codomain
restriction works to show that if f : I R is strictly monotone, f : I f (I) is
bijective, hence invertible. Thus in this sense we may speak of the inverse of any
strictly monotone function.
Proposition 5.37. Let f : I f (I) be a strictly monotone function.
a) If f is increasing, then f 1 : f (I) I is increasing.
b) If f is decreasing, then f 1 : f (I) I is decreasing.
Proof. As usual, we will content ourselves with the increasing case, the decreasing case being so similar as to make a good exercise for the reader.
Seeking a contradiction we suppose that f 1 is not increasing: that is, there
exist y1 < y2 f (I) such that f 1 (y1 ) is not less than f 1 (y2 ). Since f 1 is an
inverse function, it is necessarily injective (if it werent, f itself would not be a
function!), so we cannot have f 1 (y1 ) = f 1 (y2 ), and thus the possibility we need
to rule out is f 1 (y2 ) < f 1 (y1 ). But if this holds we apply the increasing function
f to get y2 = f (f 1 (y2 )) < f (f 1 (y1 )) = y1 , a contradiction.

Lemma 5.38. (-V Lemma) Let f : I R. The following are equivalent:
(i) f is not monotone: i.e., f is neither increasing nor decreasing.
(ii) At least one of the following holds:
(a) f is not injective.
(b) f admits a -configuration: there exist a < b < c I with f (a) < f (b) > f (c).
(c) f admits a V -configuration: there exist a < b < c I with f (a) > f (b) < f (c).
Exercise: Prove Lemma 5.38.
Theorem 5.39. If f : I R is continuous and injective, it is monotone.
Proof. We will suppose that f is injective and not monotone and show that
it cannot be continuous, which suffices. We may apply Lemma 5.38 to conclude
that f has either a configuration or a V configuration.
Suppose first f has a configuration: there exist a < b < c I with f (a) <
f (b) > f (c). Then there exists L R such that f (a) < L < f (b) > L > f (c). If f
were continuous then by the Intermediate Value Theorem there would be d (a, b)
and e (b, c) such that f (d) = f (e) = L, contradicting the injectivity of f .
Next suppose f has a V configuration: there exist a < b < c I such that
f (a) > f (b) < f (c). Then there exists L R such that f (a) > L > f (b) < L < f (c).
If f were continuous then by the Intermediate Value Theorem there would be
d (a, b) and e (b, c) such that f (d) = f (e) = L, contradicting injectivity.

6.4. Inverses of Continuous Functions.
Theorem 5.40. (Continuous Inverse Function Theorem) Let f : I R be
injective and continuous. Let J = f (I) be the image of f .
a) f : I J is a bijection, and thus there is an inverse function f 1 : J I.
b) J is an interval in R.
c) If I = [a, b], then either f is increasing and J = [f (a), f (b)] or f is decreasing
and J = [f (b), f (a)].
d) The function f 1 : J I is also continuous.

6. INVERSE FUNCTIONS I: THEORY

95

Proof. [S, Thm. 12.3] Parts a) through c) simply recap previous results. The
new result is part d), that f 1 : J I is continuous. By part c) and Proposition
5.37, either f and f 1 are both increasing, or f and f 1 are both decreasing. As
usual, we restrict ourselves to the first case.
Let b J. We must show that limyb f 1 (y) = f 1 (b). We may write b = f (a)
for a unique a I. Fix  > 0. We want to find > 0 such that if f (a) < y <
f (a) + , then a  < f 1 (y) < a + .
Take = min(f (a + ) f (a), f (a) f (a )). Then:
f (a ) f (a) , f (a) + f (a + ),
and thus if f (a) < y < f (a) + we have
f (a ) f (a) < y < f (a) + f (a + ).
Since f 1 is increasing, we get
f 1 (f (a )) < f 1 (y) < f 1 (f (a + )),
or
f 1 (b)  < f 1 (y) < f 1 (b) + .

Remark: To be honest, I dont find the above proof very enlightening. After reflecting on my dissatisfaction with it, I came up with an alternate proof that I find
conceptually simpler, but which depends on the Monotone Jump Theorem, a
characterization of the possible discontinuities of a monotone function. The proof
uses the completeness of the real numbers, so is postponed to the next chapter.
6.5. Inverses of Differentiable Functions.
In this section our goal is to determine conditions under which the inverse f 1
of a differentiable funtion is differentiable, and if so to find a formula for (f 1 )0 .
Lets first think about the problem geometrically. The graph of the inverse function y = f 1 (x) is obtained from the graph of y = f (x) by interchanging x and
y, or, put more geometrically, by reflecting the graph of y = f (x) across the line
y = x. Geometrically speaking y = f (x) is differentiable at x iff its graph has
a well-defined, nonvertical tangent line at the point (x, f (x)), and if a curve has
a well-defined tangent line, then reflecting it across a line should not change this.
Thus it should be the case that if f is differentiable, so is f 1 . Well, almost. Notice
the occurrence of nonvertical above: if a curve has a vertical tangent line, then
since a vertical line has infinite slope it does not have a finite-valued derivative.
So we need to worry about the possibility that reflection through y = x carries a
nonvertical tangent line to a vertical tangent line. When does this happen? Well,
1
(x b)
the inverse function of the straight line y = mx + b is the straight line y = m
1
i.e., reflecting across y = x takes a line of slope m to a line of slope m . Morever,
it takes a horizontal line y = c to a vertical line x = c, so that is our answer: at
any point (a, b) = (a, f (a)) such that f 0 (a) = 0, then the inverse function will fail
to be differentiable at the point (b, a) = (b, f 1 (b)) because it will have a vertical
tangent. Otherwise, the slope of the tangent line of the inverse function at (b, a) is
precisely the reciprocal of the slope of the tangent line to y = f (x) at (a, b).

96

5. DIFFERENTIATION

Well, so the geometry tells us. It turns out to be quite straightforward to adapt
this geometric argument to derive the desired formula for (f 1 )0 (b), under the assumption that f is differentiable. We will do this first. Then we need to come back
and verify that indeed f 1 is differentiable at b if f 0 (f 1 (b)) exists and is nonzero:
this turns out to be a bit stickier, but we are ready for it and we will do it.
Proposition 5.41. Let f : I J be a bijective differentiable function. Suppose that the inverse function f 1 : J I is differentiable at b J. Then
1
1
is differentiable at b then f 0 (f 1 (b)) 6=
(f 1 )0 (b) = f 0 (f 1
(b)) . In particular, if f
0.
Proof. We need only implicitly differentiate the equation
f 1 (f (x)) = x,
getting
(f 1 )0 (f (x))f 0 (x) = 1,

(20)
or

1
.
f 0 (x)
To apply this to get the derivative at b J, we just need to think a little about our
variables. Let a = f 1 (b), so f (a) = b. Evaluating the last equation at x = a gives
1
1
(f 1 )0 (b) = 0
= 0 1
.
f (a)
f (f (b))
(f 1 )0 (f (x)) =

Moreover, since by (20) we have (f 1 )0 (b)f 0 (f 1 (b)) = 1, f 0 (f 1 (b)) 6= 0.

As mentioned above, unfortunately we need to work a little harder to show the

differentiability of f 1 , and for this we cannot directly use Proposition 5.41 but
end up deriving it again. Well, enough complaining: here goes.
Theorem 5.42. (Differentiable Inverse Function Theorem) Let f : I J be
continuous and bijective. Let b be an interior point of J and put a = f 1 (b).
Suppose that f is differentiable at a and f 0 (a) 6= 0. Then f 1 is differentiable at b,
with the familiar formula
1
1
= 0 1
.
(f 1 )0 (b) = 0
f (a)
f (f (b))
Proof. [S, Thm. 12.5] We have
f 1 (b + h) f 1 (b)
f 1 (b + h) a
= lim
.
h0
h0
h
h
Since J = f (I), every b + h J is of the form
(f 1 )0 (b) = lim

b + h = f (a + kh )
for a unique kh I. Since b + h = f (a + kh ), f 1 (b + h) = a + kh ; lets make this
substitution as well as h = f (a + kh ) f (a) in the limit we are trying to evaluate:
6

a + kh a
kh
= lim
= lim
h0 f (a + kh ) b
h0 f (a + kh ) f (a)
h0

(f 1 )0 (b) = lim

1
f (a+kh )f (a)
kh

6Unlike Spivak, we will include the subscript k to remind ourselves that this k is defined in
h
terms of h: to my taste this reminder is worth a little notational complication.

7. INVERSE FUNCTIONS II: EXAMPLES AND APPLICATIONS

97

We are getting close: the limit now looks like the reciprocal of the derivative of f
at a. The only issue is the pesky kh , but if we can show that limh0 kh = 0, then
we may simply replace the limh0 with limkh 0 and well be done.
But kh = f 1 (b + h) a, so since f 1 is continuous by Theorem 5.40 we
have
lim kh = lim f 1 (b + h) a = f 1 (b + 0) a = f 1 (b) a = a a = 0.

h0

h0

So as h 0, kh 0 and thus
(f 1 )0 (b) =

1
limkh 0 f (a+kkhh)f (a)

1
f 0 (a)

1
f 0 (f 1 (b))

.


7. Inverse Functions II: Examples and Applications

1
n

7.1. x .
In this section we illustrate the preceding concepts by defining and differentiat1
ing the nth root function x n . The reader should not now be surprised to hear that
we give separate consideration to the cases of odd n and even n.
Either way, let n > 1 be an integer, and consider
f : R R, x 7 xn .
Case 1: n = 2k +1 is odd. Then f 0 (x) = (2k +1)x2k = (2k +1)(xk )2 is non-negative
for all x R and not identically zero on any subinterval [a, b] with a < b, so by
Theorem 5.25 f : R R is increasing. Moreover, we have limx f (x) = .
Since f is continuous, by the Intermediate Value Theorem the image of f is all of
R. Moreover, f is everywhere differentiable and has a horizontal tangent only at
x = 0. Therefore there is an inverse function
f 1 : R R
which is everywhere continuous and differentiable at every x R except x = 0. It
1
is typical to call this function x n .
Case 2: n = 2k is even. Then f 0 (x) = (2k)x2k1 is positive when x > 0 and
negative when x < 0. Thus f is decreasing on (, 0] and increasing on [0, ). In
particular it is not injective on its domain. If we want to get an inverse function,
we need to engage in domain restriction. Unlike codomain restriction, which can
be done in exactly one way so as to result in a surjective function, domain restriction brings with it many choices. Luckily for us, this is a relatively simple case: if
D R, then the restriction of f to D will be injective if and only if for each x R,
at most one of x, x lies in D. If we want the restricted domain to be as large as
possible, we should choose the domain to include 0 and exactly one of x, x for
all x > 0. There are still lots of ways to do this, so lets try to impose another
desirable property of the domain of a function: namely, if possible we would like
it to be an interval. A little thought shows that there are two restricted domains
which meet all these requirements: we may take D = [0, ) or D = (, 0].

98

5. DIFFERENTIATION

7.2. L(x) and E(x).

Consider the function l : (0, ) R given by l(x) = x1 . As advertised, we will
soon be able to prove that every continuous function has an antiderivative, so borrowing on this result we define L : (0, ) R to be such that L0 (x) = l(x). More
precisely, recall that when they exist antiderivatives are unique up to the addition
of a constant, so we may uniquely specify L(x) by requiring L(1) = 0.
Proposition 5.43. For all x, y (0, ), we have
(21)

L(xy) = L(x) + L(y).

Proof. Let y (0, ) be regarded as fixed, and consider the function
f (x) = L(xy) L(x) L(y).

We have
f 0 (x) = L0 (xy)(xy)0 L0 (x) =

1
y
1
1
y =
= 0.
xy
x
xy x

By the zero velocity theorem, the function f (x) is a constant (depending, a priori
on y), say Cy . Thus for all x (0, ),
L(xy) = L(x) + L(y) + Cy .
If we plug in x = 1 we get
L(y) = 0 + L(y) + Cy ,
and thus Cy = 0, so L(xy) = L(x) + L(y).

Corollary 5.44. a) For all x (0, ) and n Z+ , we have L(xn ) = nL(x).

b) For x (0, ), we have L( x1 ) = L(x).
c) We have limx L(x) = , limx0+ L(x) = .
d) We have L((0, )) = R.
Proof. a) An easy induction argument using L(x2 ) = L(x) + L(x) = 2L(x).
b) For any x (0, ) we have 0 = L(1) = L(x x1 ) = L(x) + L( x1 ).
c) Since L0 (x) = x1 > 0 for all x (0, ), L is increasing on (0, ). Since
L(1) = 0, for any x > 0, L(x) > 0. To be specific, take C = L(2), so C > 0.
Then by part a), L(2n ) = nL(2) = nC. By the Archimedean property of R, this
shows that L takes arbitaririly large values, and since it is increasing, this implies
limx L(x) = . To evaluate limx0+ L(x) we may proceed similarly: by part
b), L( 12 ) = L(2) = C < 0, so L( 21n ) = nL(2) = Cn, so L takes arbitrarily
small values. Again, combined with the fact that L is increasing, this implies
limx0+ L(x) = . (Alternately, we may evaluate limx0+ L(x) by making the
change of variable y = x1 and noting that as x 0+ , y +. This is perhaps
more intuitive but is slightly tedious to make completely rigorous.)
d) Since L is differentiable, it is continuous, and the result follows immediately from
part c) and the Intermediate Value Theorem.

Definition: We define e to be the unique positive real number such that L(e) = 1.
(Such a number exists because L : (0, ) R is increasing hence injective and
has image (, ). Thus in fact for any real number there is a unique positive
real number such that L() = .)

7. INVERSE FUNCTIONS II: EXAMPLES AND APPLICATIONS

99

Since L(x) is everywhere differentiable with nonzero derivative x1 , the differentiable

inverse function theorem applies: L has a differentiable inverse function
E : R (0, ), E(0) = 1.
0

Lets compute E : differentiating L(E(x)) = x gives

1 = L0 (E(x))E 0 (x) =

E 0 (x)
.
E(x)

In other words, we get

E 0 (x) = E(x).
Corollary 5.45. For all x, y R we have E(x + y) = E(x)E(y).
Proof. To showcase the range of techniques available, we give three different
proofs.
Ey (x)
. Then
First proof: For y R, let Ey (x) = E(x + y). Put f (x) = E(x)
f 0 (x) =

Ey (x)E 0 (x) Ey0 (x)E(x)

E(x + y)E(x) E 0 (x + y)(x + y)0 E(x)
=
E(x)2
E(x)2

E(x + y)E(x) E(x + y) 1 E(x)

= 0.
E(x)2
By the Zero Velocity Theorem, there is Cy R such that for all x R, f (x) =
E(x + y)/E(x) = Cy , or E(x + y) = E(x)Cy . Plugging in x = 0 gives
=

E(y) = E(0)C(y) = 1 C(y) = C(y),

so
E(x + y) = E(x)E(y).
Second proof: We have


E(x + y)
L
= L(E(x + y)) L(E(x)) L(E(y)) = x + y x y = 0.
E(x)E(y)
The unique x (0, ) such that L(x) = 0 is x = 1, so we must have
E(x + y)
= 1,
E(x)E(y)
or
E(x + y) = E(x)E(y).
Third proof: For any y1 , y2 > 0, we have
L(y1 y2 ) = L(y1 ) + L(y2 ).
Put y1 = E(x1 ) and y2 = E(x2 ), so that x1 = L(y1 ), x2 = L(y2 ) and thus
E(x1 )E(x2 ) = y1 y2 = E(L(y1 y2 )) = E(L(y1 ) + L(y2 )) = E(x1 + x2 ).

Note also that since E and L are inverse functions and L(e) = 1, we have E(1) = e.
Now the previous disucssion must suggest to any graduate of freshman calculus
that E(x) = ex : both functions defined and positive for all real numbers, are equal
to their own derivatives, convert multiplication into addition, and take the value 1
at x = 0. How many such functions could there be?

100

5. DIFFERENTIATION

Proposition 5.46. Let f : R R be a differentiable function such that f 0 (x) =

f (x) for all x R. There is a constant C such that f (x) = CE(x) for all x R.
Proof. Define a function g : R R by g(x) =
g 0 (x) =

f (x)
E(x) .

Then for all x R,

E(x)f 0 (x) E 0 (x)f (x)

E(x)f (x) E(x)f (x)
=
= 0.
E(x)2
E(x)2

By the Zero Velocity Theorem g =

f
E

is constant: f (x) = CE(x) for all x.

In other words, if there really is a function f (x) = ex out there with f 0 (x) = ex and
f (0) = 1, then we must have ex = E(x) for all x. The point of this logical maneuver
is that although in precalculus mathematics one learns to manipulate and graph
exponential functions, the actual definition of ax for irrational x is not given, and
indeed I dont see how it can be given without using key concepts and theorems of
calculus. But, with the functions E(x) and L(x) in hand, let us develop the theory
of exponentials and logarithms to arbitrary bases.
Let a > 0 be a real number. How should we define ax ? In the following slightly
strange way: for any x R,
ax := E(L(a)x).
Let us make two comments: first, if a = e this agrees with our previous definition:
ex = E(xL(e))) = E(x). Second, the definition is motivated by the following
desirable law of exponents: (ab )c = abc . Indeed, assuming this holds unrestrictedly
for b, c R and a > 1, we would have
ax = E(x log a) = ex log a = (elog a )x = ax .
But here is the point: we do not wish to assume that the laws of exponents work
for all real numbers as they do for positive integers...we want to prove them!
Proposition 5.47. Fix a (0, ). For x R, we define
ax := E(L(a)x).
If a 6= 1, we define
L(x)
.
L(a)
a) The function ax is differentiable and (ax )0 = L(a)ax .
1
b) The function loga x is differentiable and (loga x)0 = L(a)x
.
x
c) Suppose a > 1. Then a is increasing with image (0, ), loga x is increasing
with image (, ), and ax and loga x are inverse functions.
d) For all x, y R, ax+y = ax ay .
e) For all x > 0 and y R, (ax )y = axy .
f ) For all x, y > 0, loga (xy) = loga x + loga y.
g) For all x > 0 and y R, loga (xy ) = y loga x.
loga (x) =

Proof. a) We have
(ax )0 = E(L(a)x)0 = E 0 (L(a)x)(L(a)x)0 = E(L(a)x) L(a) = L(a)ax .
b) We have
0

(loga (x)) =

L(x)
L(a)

0
=

1
.
L(a)x

7. INVERSE FUNCTIONS II: EXAMPLES AND APPLICATIONS

101

c) Since their derivatives are always positive, ax and loga x are both increasing
functions. Moreover, since a > 1, L(a) > 0 and thus
lim ax = lim E(L(a)x) = E() = ,

L(x)

=
= .
x
x L(a)
L(a)
Thus ax : (, ) (0, ) and loga x : (0, ) (, ) are bijective and thus
have inverse functions. Thus check that they are inverses of each other, it suffices
to show that either one of the two compositions is the identity function. Now
lim loga (x) = lim

loga (ax ) =

L(E(L(a)x))
L(a)x
L(ax )
=
=
= x.
L(a)
L(a)
L(a)

d) We have
ax+y = E(L(a)(x + y)) = E(L(a)x + L(a)y) = E(L(a)x)E(L(a)y) = ax ay .
e) We have
(ax )y = E(L(ax )y) = E(L(E(L(a)x))y) = E(L(a)xy) = axy .
f) We have
loga (xy) =

L(x) + L(y)
L(x) L(y)
L(xy)
=
=
+
= loga x + loga y.
L(a)
L(a)
L(a) L(a)

g) We have
loga xy =

L(xy )
L(E(L(x)y))
L(x)y
=
=
= y loga x.
L(a)
L(a)
L(a)


Having established all this, we now feel free to write ex for E(x) and log x for L(x).
Exercise: Suppose 0 < a < 1. Show that ax is decreasing with image (0, ),
loga x is decreasing with image (0, ), and ax and loga x are inverse functions.
Exercise: Prove the change of base formula: for all a, b, c > 0 with a, c 6= 1,
logc b
loga b =
.
logc a
2

Proposition 5.48. Let f (x) = ex . Then for all n Z+ there exists a polynomial Pn (x), of degree n, such that
2
dn
f (x) = Pn (x)ex .
dxn
Proof. By induction on n.
Base case (n = 1):
2
2
d x2
= 2xex = P1 (x)ex , where P1 (x) = 2x, a degree one polynomial.
dx e

Inductive step: Assume that for some positive integer n there exists Pn (x) of degree
2
dn x2
dn+1 x2
n such that dx
= Pn (x)ex . So dx
=
ne
n+1 e
2
2
2
2
d dn x2 IH d
e =
Pn (x)ex = Pn0 (x)ex + 2xPn (x)ex = (Pn0 (x) + 2xPn (x)) ex .
n
dx dx
dx

102

5. DIFFERENTIATION

Now, since Pn (x) has degree n, Pn0 (x) has degree n 1 and 2xPn (x) has degree
n + 1. If f and g are two polynomials such that the degree of f is different from
the degree of g, then deg(f + g) = max(deg(f ), deg(g)). In particular, Pn+1 (x) :=
Pn0 (x) + 2xPn (x) has degree n + 1, completing the proof of the induction step. 
7.3. Some inverse trigonometric functions.
We now wish to consider inverses of the trigonometric functions: sine, cosine, tangent, and so forth. Right away we encounter a problem similar to the case of xn for
even n: the trigonometric functions are periodic, hence certainly not injective on
their entire domain. Once again we are forced into the art of domain restriction
(as opposed to the science of codomain restriction).
Consider first f (x) = sin x. To get an inverse function, we need to restrict the
domain to some subset S on which f is injective. As usual we like intervals, and a
little thought shows that the maximal possible length of an interval on which the
sine function is injective is , attained by any interval at which the function either
increases from 1 to 1 or decreases from 1 to 1. This still gives us choices to make.
The most standard choice but to be sure, one that is not the only possible one

nor is mathematically consecrated in any particular way is to take I = [

2 , 2 ].
0
We claim that f is increasing on I. To check this, note that f (x) = cos x is indeed

positive on (
2 , 2 ). We have f ([ 2 , 2 ]) = [1, 1]. The inverse function here is often called arcsin x (arcsine of x) in an attempt to distinguish it from sin1 x = csc x.
This is as good a name as any: lets go with it. We have

arcsin : [1, 1] [
, ].
2 2
As the inverse of an increasing function, arcsin x is increasing. Moreover since sin x

has a nonzero derivative on (

2 , 2 ), arcsin x is differentiable there. Differentiating
sin(arcsin x)) = x,
we get
cos(arcsin x) arcsin0 (x) = 1,
or
d
1
arcsin x =
.
dx
cos(arcsin x)
This looks like a mess, but a little trigonometry will clean it up. The key is to
realize that cos arcsin x means the cosine of the angle whose sine is x and that
there must be a simpler description of this. If we draw a right triangle with angle
= arcsin x, then to get the ratio of the opposite side to the hypotenuse to be x we
may take the length of the opposite side to be x and the length of the hypotenuse to
be
of the adjacent side is, by the Pythagorean Theorem,

1, in which case the length

1 x2 . Thus cos = 1 x2 , so finally
d
1
arcsin x =
.
dx
1 x2
Now consider f (x) = cos x. Since f is even, it is not injective on any interval
containing 0 in its interior. Reflecting a bit on the graph of f (x) = cos x one sees
that a reasonable choice for the restricted domain is [0, ]: since f 0 (x) = sin x is

7. INVERSE FUNCTIONS II: EXAMPLES AND APPLICATIONS

103

negative on (0, ) and 0 and 0 and , f (x) is decreasing on [0, ] and hence injective
there. Its image is f ([0, ])) = [1, 1]. Therefore we have an inverse function
arccos : [1, 1] [0, ].
Since cos x is continuous, so is arccos x. Since cos x is differentiable and has zero
derivative only at 0 and , arccos x is differentiable on (1, 1) and has vertical tangent lines at x = 1 and x = 1. Morever, since cos x is decreasing, so is arccos x.
We find a formula for the derivative of arccos just as for arcsin: differentiating
cos arccos x = x
gives
sin(arccos x) arccos0 x = 1,
or

1
.
sin arccos x
Again, this may be simplified. If = arccos
x, then x = cos , so if we are on the
unit circle then the y-coordinate is sin = 1 x2 , and thus
1
arccos0 x =
.
1 x2
Remark: It is hard not to notice that the derivatives of the arcsine and the arccosine
are simply negatives of each other, so for all x [0, 2 ],
arccos0 x =

arccos0 x + arcsin0 x = 0.
By the Zero Velocity Theorem, we conclude
arccos x + arcsin x = C
for some constant C. To determine C, simply evaluate at x = 0:

C = arccos 0 + arcsin 0 = + 0 = ,
2
2
and thus for all x [0, 2 ] we have

arccos x + arcsin x = .
2
So the angle whose sine is x is complementary to the angle whose cosine is x.
sin x
Finally, consider f (x) = tan x = cos
x . The domain is all real numbers for which
cos x 6= 0, so all real numbers except 2 , 3
2 , . . .. The tangent function is periodic with period and also odd, which suggests that, as with the sine function, we
should restrict this domain to the largest interval about 0 on which f is defined and

injective. Since f 0 (x) = sec2 x > 0, f is increasing on (

2 , 2 ) and thus is injective
there. Moreover, limx 2 tan x = , so by the Intermediate Value Theorem

f ((
2 , 2 )) = R. Therefore we have an inverse function



arctan : R
,
.
2 2

Since the tangent function is differentiable with everywhere positive derivative, the
same is true for arctan x. In particular it is increasing but bounded: limx arctan x =
2 . In other words the arctangent has horizontal asymptotes at y = 2 .

104

5. DIFFERENTIATION

8. Some Complements
The Mean Value Theorem and its role in freshman calculus has been a popular
topic of research and debate over the years.
A short paper of W.J. Knight improves upon the Zero Velocity Theorem.
Theorem 5.49. (Right-Handed Zero Velocity Theorem [Kn80]) Let f : [a, b]
0
R be continuous. If the right-hand derivative f+
(x) exists and is 0 for all x (a, b),
then f is constant.
The proof is modelled upon the usual one: one starts with a right-handed Rolles
Theorem, deduces a right-handed Mean Value Inequality, and then Theorem 5.49.

CHAPTER 6

Completeness
1. Dedekind Completeness
1.1. Introducing (LUB) and (GLB).
Gather round, my friends: the time has come to tell what makes calculus work.
Recall that we began the course by considering the real numbers as a set endowed
with two binary operations + and together with a relation <, and satisfying a
longish list of familiar axioms (P0) through (P12), the ordered field axioms. We
then showed that using these axioms we could deduce many other familiar properties of numbers and prove many other identities and inequalities.
However we did not claim that (P0) through (P12) was a complete list of axioms
for R. On the contrary, we saw that this could not be the case: for instance the
rational numbers Q also satisfy the ordered field axioms but as we have taken
great pains to point out most of the big theorems of calculus are meaningful
but false when regarded as results applied to the system of rational numbers. So
there must be some further axiom, or property, of R which is needed to prove the
three Interval Theorems, among others.
Here it is. Among structures F satisfying the ordered field axioms, consider the
following further property:
(P14): Least Upper Bound Axiom (LUB): Let S be a nonempty subset of
F which is bounded above. Then S admits a least upper bound.
This means exactly what it sounds like, but it is so important that we had better
make sure. Recall a subset S of F is bounded above if there exists M R such
that for all x S, x M . (For future reference, a subset S of R is bounded
below if there exists m F such that for all x S, m x.) By a least upper
bound for a subset S of F , we mean an upper bound M which is less than any
other upper bound: thus, M is a least upper bound for S if M is an upper bound
for S and for any upper bound M 0 for S, M M 0 .
There is a widely used synonym for the least upper bound of S, namely the
supremum of S. We also introduce the notation lub S = sup S for the supremum
of a subset S of an ordered field (when it exists).
The following is a useful alternate characterization of sup S: the supremum of
105

106

6. COMPLETENESS

S is an upper bound M for S with the property that for any M 0 < M , M 0 is not
an upper bound for S: explicitly, for all M 0 < M , there exists x S with M 0 < x.
The definition of the least upper bound of a subset S makes sense for any set
X endowed with an order relation <. Notice that the uniqueness of the supremum
sup S is clear: we cannot have two different least upper bounds for a subset, because one of them will be larger than the other! Rather what is in question is the
existence of least upper bounds, and (LUB) is an assertion about this.
Taking the risk of introducing even more terminology, we say that an ordered field
(F, +, , <) is Dedekind complete1 if it satisfies the least upper bound axiom.
Now here is the key fact lying at the foundations of calculus and real analysis.
Theorem 6.1. a) The ordered field R is Dedekind complete.
b) Conversely, any Dedekind complete ordered field is isomorphic to R.
Part b) of Theorem 6.1 really means the following: if F is any Dedekind complete
ordered field then there is a bijection f : F R which preserves the addition,
multiplication and order structures in the following sense: for all x, y F ,
f (x + y) = f (x) + f (y),
f (xy) = f (x)f (y), and
If x < y, then f (x) < f (y).
This concept of isomorphism of structures comes from a more advanced course
abstract algebra so it is probably best to let it go for now. One may take part
b) to mean that there is essentially only one Dedekind complete ordered field: R.
The proof of Theorem 6.1 involves constructing the real numbers in a mathematically rigorous way. This is something of a production, and although in some
sense every serious student of mathematics should see a construction of R at some
point of her career, this sense is similar to the one in which every serious student
of computer science should build at least one working computer from scratch: in
practice, one can probably get away with relying on the fact that many other people
have performed this task in the past. Spivak does give a construction of R and a
proof of Theorem 6.1 in the Epilogue of his text. And indeed, if we treat this
material at all it will be at the very end of the course.
After discussing least upper bounds, it is only natural to consider the dual concept of greatest lower bounds. Again, this means exactly what it sounds like but it
is so important that we spell it out explicitly: if S is a subset of an ordered field F ,
then a greatest lower bound for S, or an infimum of S, is an element m F
which is a lower bound for S i.e., m x for all x S and is such that if m0
is any lower bound for S then m0 m. Equivalently, m = inf S iff m is a lower
bound for S and for any m0 > m there exists x S with x < m0 . Now consider:
(P140 ): Greatest Lower Bound Axiom (GLB): Let S be a nonempty subset
of F which is bounded below. Then S admits a greatest lower bound, or infimum.
1It is perhaps more common to say complete instead of Dedekind complete. I have my
reasons for preferring the lengthier terminology, but I wont trouble you with them.

1. DEDEKIND COMPLETENESS

107

Example 1.1: In any ordered field F , we may consider the subset

SF = {x F | x2 < 2}.
Then SF is nonempty and bounded: indeed 0 SF and if x SF , then |x| 2. Of
course in the previous inequality we could do better: for instance, if |x| > 23 , then
3
x2 > 94 > 2, so also 3
2 is a lower bound for SF and 2 is an upper bound for SF .
Of course we could do better still...
Indeed the bounded set SF will have an infimum and a supremum if and only
if there are best possible inequalities x S = m x M , i.e., for which no
improvement on either m or M is possible. Whether such best possible inequalities
exist depends on the ordered field F . Indeed, it is clear that if M = sup SF exists,
then it must be a positive element of F with M 2 = 2:or in other words, what in
precalculus mathematics one cavalierly writes as M = 2. similarly, if m = inf SF
exists, thenit must be a negative element of F with m2 = 2, or what we usually
write as 2. But heres the point: how do we know that our ordered field F
contains such an element 2?
The answer of course is that depending on F such an element may or may not
exist. As we saw at the beginning of the course, there is no rational number x with
x2 = 2, so if F = Q then our set SQ has neither an infimum nor a supremum. Thus
Q does not satisfy (LUB) or (GLB). On the other hand, we certainly believe that
there is a real number whose square is 2. But...why do we believe this? As we
have seen, the existence of a real square root of every non-negative real number is a
consequence of the Intermediate Value Theorem...which is of course a theorem that
we have
exalted but not yet proved. A more fundamental answer is that we believe
that 2 exists in R because of the Dedekind completeness of R, i.e., according to
Theorem 6.1 every nonempty bounded above subset
of R has a supremum, so in

particular SR has a supremum, which must be 2.

An interesting feature of this example is that we can see that inf SR exists,
even though we have not as yet addressed the issue of whether R satisfies (GLB).
In general, inf SF exists iff there is an element y < 0 in F with y 2 = 2. But okay:
if in F we have a positive element x with x2 = 2, we necessarily must also have a
negative element y with y 2 = 2: namely, y = x.
This turns out to be a very general phenomenon.
Theorem 6.2. Let F be an ordered field.
a) Then F satisfies (LUB) iff it satisfies (GLB).
b) In particular R satisfies both (LUB) and (GLB) and is (up to isomorphism) the
only ordered field with this property.
Proof. a) I know two ways of showing that (LUB) (GLB). Both of
these arguments is very nice in its own way, and I dont want to have to choose
between them. So I will show you both, in the following way: I will use the first
argument to show that (LUB) = (GLB) and the second argument to show that
(GLB) = (LUB). (In Exercise 1.2 below, you are asked to do things the other
way around.)
(LUB) = (GLB): Let S F be nonempty and bounded below by m. Consider
S = {x | x S}.

108

6. COMPLETENESS

Then S is nonempty and bounded above by m. By (LUB), it has a least upper

bound sup(S). We claim that in fact sup(S) is a greatest lower bound for S,
or more symbolically:
inf S = sup S.
You are asked to check this in Exercise 1.2 below.
(GLB) = (LUB): Let S be nonempty and bounded above by M . Consider
U(S) = {x F | x is an upper bound for S.}.
Then U(S) is nonempty: indeed M U(S). Also U(S) is bounded below: indeed
any s S (there is at least one such s, since S 6= !) is a lower bound for U(S). By
(GLB) U(S) has a greatest lower bound inf U(S). We claim that in fact inf U(S) is
a least upper bound for S, or more succinctly,
sup S = inf U(S).
Once again, Exercise 1.2 asks you to check this.
b) By Theorem 6.1a), R satisfies (LUB), and thus by part a) it satisfies (GLB). By
Theorem 6.1b) R is the only ordered field satisfying (LUB), so certainly it is the
only ordered field satifying (LUB) and (GLB).

Exercise 1.2: a) Fill in the details of the proof of Theorem 6.2a).
b) Let F be an ordered field, and let S be a subset of F . Suppose that inf S exists.
Show that sup S exists and
sup S = inf S.
c) Use part b) to give a second proof that (GLB) = (LUB).
d) Let F be an ordered field, and let S be a subset of F . Define
L(S) = {x F | x is a lower bound for S.}.
Suppose that sup L(S) exists. Show that inf S exists and
inf S = sup L(S).
e) Use part d) to give a second proof that (LUB) = (GLB).
The technique which was used to prove (LUB) = (GLB) is very familiar: we
multiply everything in sight by 1. It seems likely that by now we have used this
type of argument more than any other single trick or technique. When this has
come up we have usually used the phrase and similarly one can show... Perhaps
this stalwart ally deserves better. Henceforth, when we wish to multiply by 1 to
convert to , max to min, sup to inf and so forth, we will say by reflection.
This seems more appealing and also more specific than similarly...!
In view of Theorem 6.2 it is reasonable to use the term Dedekind completeness to refer to either or both of (LUB), (GLB), and we shall do so.
Theorem 6.3. A Dedekind complete ordered field is Archimedean.
Proof. We will prove the contrapositive: let F be a non-Archimedean ordered
field: thus there exists x F such that n x for all n Z+ . Then the subset Z+
of F is bounded above by x, so in particular it is nonempty and bounded above.
So, if F were Dedekind complete then sup Z+ would exist.
But we claim that in no ordered field F does Z+ have a supremum. Indeed,

1. DEDEKIND COMPLETENESS

109

suppose that M = sup Z+ . It follows that for all n Z+ , n M . But then it

is equally true that for all n Z+ , n + 1 M , or equivalently, for all n Z+ ,
n M 1, so M 1 is a smaller upper bound for Z+ than sup Z+ : contradiction! 
1.2. Calisthenics With Sup and Inf.
The material and presentation of this section is partly based on [A, 1.3.13].
convention: Whenever sup S appears in the conclusion of a result, the statement should be understood as including the assertion that sup S exists, i.e., that
S is nonempty and bounded above. Similarly for inf S: when it appears in the
conclusion of a result then an implicit part of the conclusion is the assertion that
inf S exists, i.e., that S is nonempty and bounded below.2
Proposition 6.4. Let S be a nonempty subset of R.
a) Suppose S is bounded above. Then for every  > 0, there exists x S such that
sup S  < x sup S.
b) Conversely, suppose M R is an upper bound for S such that for all  > 0,
there exists x S with M  < x M . Then M = sup S.
c) Suppose S is bounded below. Then for every  > 0, there exists x S such that
inf S x < inf S + .
d) Conversely, suppose m R is a lower bound for S such that for all  > 0, there
exists x S with m x m + . Then m = inf S.
Proof. a) Fix  > 0. Since sup S is the least upper bound of S and sup S  <
sup S, there exists y S with sup S  < y. It follows that
sup S  < min(y, sup S) sup S,
so we may take x = min(y, sup S).
b) By hypothesis, M is an upper bound for S and nothing smaller than M is an
upper bound for S, so indeed M = sup S.
c),d) These follow from parts a) and b) by reflection.

Exercise 1.3: Let a, b R. Suppose that for all  > 0, a b + . Show that a b.
Proposition 6.5. Let X, Y be nonempty subsets of R, and define
X + Y = {x + y | x X, y Y }.
a) Suppose X and Y are bounded above. Then
sup(X + Y ) = sup X + sup Y.
b) Suppose X and Y are bounded below. Then
inf(X + Y ) = inf X + inf Y.
Proof. a) Let x X, y Y . Then x sup X and x sup Y , so x + y
sup X+sup Y , and thus sup(X+Y ) sup X+sup Y . Now fix  > 0. By Proposition
6.4 there are x X and y Y with sup X 2 < x, sup Y 2 < y, so
sup X + sup Y x + y + .
Since this folds for all  > 0, by Exercise 1.3 sup X + sup Y sup(X + Y ).
b) This follows from part a) by reflection.
2Notice that a similar convention governs the use of lim
xc f (x), so this is nothing new.

110

6. COMPLETENESS

Let X, Y be subsets of R. We write X Y if for all x X and all y Y , x y.

(In a similar way we define X < Y, X Y , X > Y .)
Exercise 1.4: Let X, Y be subsets of R. Give necessary and sufficient conditions for
X Y and Y X both to hold. (Hint: in the case in which X and Y are both
nonempty, X = Y is necessary but not sufficient!)
Proposition 6.6. Let X, Y be nonempty subsets of R with X Y . Then
sup X inf Y.
Proof. Seeking a contradiction, we suppose that inf Y < sup X. Put
sup X inf Y
=
.
2
By Proposition 6.4 there are x X, y Y with sup X  < x and y < inf Y + .
Since X Y this gives
sup X  < x y < inf Y + 
and thus
sup X inf Y < 2 = sup X inf Y,
a contradiction.

Proposition 6.7. Let X, Y be nonempty subsets of R with X Y . Then:

a) If Y is bounded above, then sup X sup Y .
b) If Y is bounded below, then inf Y inf X.
Exercise 1.5: Prove Proposition 6.7.
1.3. The Extended Real Numbers.
As exciting and useful as this whole business with sup and inf is, there is one
slightly annoying point: sup S and inf S are not defined for every subset of R.
Rather, for sup S to be defined, S must be nonempty and bounded above, and for
inf S to be defined, S must be nonempty and bounded below.
Is there some way around this? There is. It involves bending the rules a bit,
but in a very natural and useful way. Consider the subset N of R. It is not bounded
above, so it does not have a least upper bound in R. Because N contains arbitrarily large elements of R, it is not completely unreasonable to say that its elements
approach infinity and thus to set sup N = +. In other words, we are suggesting
the following definition:
If S R is unbounded above, then we will say sup S = +.
Surely we also want to make the following definition (by reflection!):
If S R is unbounded below, then we will say inf S = .
These definitions come with a warning: are not real numbers! They
are just symbols suggestively standing for a certain type of behavior of a subset of
R, in a similar (but, in fact, simpler) way as when we write limxc f (x) = and
mean that the function has a certain type of behavior near the point c.

1. DEDEKIND COMPLETENESS

111

To give a name to what we have done, we define the extended real numbers
[, ] = R {} to be the real numbers together with these two formal
symbols and . This extension is primarily order-theoretic: that is, we may
extend the relation to the extended real numbers in the obvious way:
x R, < x < .
Conversely much of the point of the extended real numbers is to give the real
numbers, as an ordered set, the pleasant properties of a closed, bounded interval
[a, b]: namely we have a largest and smallest element.
The extended real numbers [, ] are not a field. In fact, we cannot even
define the operations of + and unrestrictedly on them. However, it is useful to
define some of these operations:
x R, + x = , x + = .
x (0, ), x = , x () = .
x (, 0), x = , x () = .
= , () = , () () = .
1
1
=
= 0.

None of these definitions are really surprising, are they? If you think about it,
they correspond to facts you have learned about manipulating infinite limits, e.g.
if limxc f (x) = and limxc g(x) = 17, then limxc f (x) + g(x) = . However,
certain other operations with the extended real numbers are not defined, for similar
reasons. In particular we do not define
,
0 ,

Why not? Well, again we might think in terms of associated limits. The above are
indeterminate forms: if I tell you that limxc f (x) = and limxc g(x) = ,
then what can you tell me about limxc f (x) + g(x)? Answer: nothing, unless you
know what specific functions f and g are. As a simple example, suppose
1
1
f (x) =
+ 2011, g(x) =
.
(x c)2
(x c)2
Then limxc f (x) = , limxc g(x) = , but
lim f (x) + g(x) = lim 2011 = 2011.

xc

xc

So cannot have a universal definition independent of the chosen functions.3 In a similar way, when evaluating limits 0 is an indeterminate form:
if limxc f (x) = 0 and limxc g(x) = , then limxc f (x)g(x) depends on how
fast f approaches zero compared to how fast g approaches infinity. Again, consider
2011

something like f (x) = (x c)2 , g(x) = (xc)

2 . And similarly for .
These are good reasons. However, there are also more purely algebraic reasons:
there is no way to define the above expressions in such a way to make the field
3In the unlikely event you think that perhaps = 2011 always, try constructing another
example...or wait until next semester and ask me again.

112

6. COMPLETENESS

axioms work out. For instance, let a R. Then a + = . If therefore we were

allowed to substract from we would deduce a = , and thus
could be any real number: thats not a well-defined operation.
Remark: Sometimes above we have alluded to the existence of ordered fields F
which do not satisfy the Archimedean axiom, i.e., for which there exist elements x
such that x > n for all n Z+ . In speaking about elements like x we sometimes
call them infinitely large. This is a totally different use of infinity than the extended real numbers above. Indeed, no ordered field F can have a largest element
x, because it follows easily from the field axioms that for any x F , x + 1 > x. The
moral: although we call extended real numbers, one should not think of them
as being elements of a number system at all, but rather limiting cases of such things.
One of the merits of this extended definition of sup S and inf S is that it works
nicely with calculations: in particular, all of the calisthenics of the previous section have nice analogues for unbounded sets. We leave it to the reader to investigate
this phenomenon on her own. In particular though, lets look back at Proposition
6.7: it says that, under conditions ensuring that the sets are nonempty and bounded
above / below, that if X Y R, then
sup X sup Y,
inf Y inf X.
This definition could have motivated our definition of sup and inf for unbounded
sets, as follows: for n Z and X R, put
X n = {x X | x n}, Xn = {x X | x n}.
The idea here is that in defining X n we are cutting it off at n in order to force it
to be bounded above, but in increasingly generous ways. We have
X0 X1 . . . X
and also
X=

X n;

n=0

in other words, every element of X is a subset of X n for some n (this is precisely the
Archimedean property). Applying Proposition 6.7, we get that for every nonempty
subset X of R,
sup X 0 sup X 1 sup X 2 . . . sup X n . . . .
Suppose moreover that X is bounded above. Then some N Z+ is an upper
bound for X, i.e., X = X N = X N +1 = . . ., so the sequence sup X n is eventually
constant, and in particular limn sup X n = sup X. On the other hand, if X
is bounded above, then the sequence sup X n is not eventually constant; in fact it
takes increasingly large values, and thus
lim sup X n = .

Thus if we take as our definition for sup X, limn sup X n , then for X which is
unbounded above, we get sup X = limn sup X n = . By reflection, a similar
discussion holds for inf X.

2. INTERVALS AND THE INTERMEDIATE VALUE THEOREM

113

There is, however, one last piece of business to attend to: we said we wanted
sup S and inf S to be defined for all subsets of R: what if S = ? There is an
answer for this as well, but many people find it confusing and counterintuitive at
first, so let me approach it again using Proposition 6.7. For each n Z, consider
the set Pn = {n}: i.e., Pn has a single element, the integer n. Certainly then
inf Pn = sup Pn = n. So what? Well, I claim we can use these sets Pn along with
Proposition 6.7 to see what inf and sup should be. Namely, to define these
quantities in such a way as to obey Proposition 6.7, then for all n Z, because
{n}, we must have
sup sup{n} = n
and
inf inf{n} = n.
There is exactly one extended real number which is less than or equal to every
integer: . Similarly, there is exactly one extended real number which is greater
than or equal to every integer: . Therefore the inexorable conclusion is
sup = , inf = .
Other reasonable thought leads to this conclusion: for instance, in class I had a lot
of success with the pushing conception of suprema and infima. Namely, if your
set S is bounded above, then you start out to the right of every element of your
set i.e., at some upper bound of S and keep pushing to the left until you cant
push any farther without passing by some element of S. What happens if you try
this with ? Well, every real number is an upper bound for , so start anywhere
and push to the left: you can keep pushing as far as you want, because you will
never hit an element of the set. Thus you can push all the way to , so to speak.
Similarly for infima, by reflection.
2. Intervals and the Intermediate Value Theorem
2.1. Convex subsets of R.
We say that a subset S of R is convex if for all x < y S, the entire interval
[x, y] lies in S. In other words, a convex set is one that whenever two points are in
it, all in between points are also in it.
Example 2.1: The empty set is convex. For any x R, the singleton set {x} is
convex. In both cases the definition applies vacuously: until we have two distinct
points of S, there is nothing to check!
Example 2.2: We claim any interval is convex. This is immediate or it would
be, if we didnt have so many different kinds of intervals to write down and check.
One needs to see that the definition applies to invervals of all of the following forms:
(a, b), [a, b), (a, b], [a, b], (, b), (, b], (a, ), [a, ), (, ).
All these verifications are trivial appeals to things like the transitivity of and .
Are there any nonempty convex sets other than intervals? (Just to be sure, we

114

6. COMPLETENESS

count {x} = [x, x] as an interval.4) A little thought suggests that the answer should
be no. But more thought shows that if so we had better use the Dedekind completeness of R, because if we work over Q with all of the corresponding definitions
then there are nonempty convex sets which are not intervals, e.g.
S = {x Q | x2 < 2}.
This
Q by R we would get an interval, namely
has
a familiar theme: replacing

( 2, 2), but once again 2 6 Q. When one looks carefully at the definitions
it is no trouble to check that working solely in the rational numbers S is a convex
set but is not an interval.
Remark: Perhaps the above example seems legalistic, or maybe even a little silly.
It really isnt: one may surmise that contemplation of such examples led Dedekind
to his construction of the real numbers via Dedekind cuts. This construction
may be discussed at the end of this course. Most contemporary analysts prefer
a rival construction of R due to Cauchy using Cauchy sequences. I agree that
Cauchys construction is simpler. However, both are important in later mathematics: Cauchys construction works in the context of a general metric space (and,
with certain modifications, in a general uniform space) to construct an associated complete space. Dedekinds construction works in the context of a general
linearly ordered set to construct an associated Dedekind-complete ordered set.
Theorem 6.8. Every nonempty convex subset D of R is an interval.
Proof. We are given a nonempty convex subset D of R and we want to show
it is an interval, but as above an interval can have any one of nine basic shapes. It
may be quite tedious to argue that one of nine things must occur!
So we just need to set things up a bit carefully: here goes: let a [, ) be
the infimum of D, and let b (, ] be the supremum of D. Let I = (a, b), and
let I be the closure of I, i.e., if a is finite, we include a; if b is finite, we include b.
Step 1: We claim that I D I. Let x I.
Case 1: Suppose I = (a, b) with a, b R. Let z (a, b). Then, since z > a =
inf D, there exists c D with c < z. Similarly, since z < b = sup D, then there
exists d D with z < d. Since D is convex, z D. Now suppose z D. We must
have inf D = a z b = sup D.
Case 2: Suppose I = (, b), and let z I. Since D is unbounded below,
there exists a D with a < z. Moreover, since z < sup D, there exists b D such
that z < b. Since D is convex, z D. Next, let z D. We wish to show that
z I = (, b]; in other words, we want z b. But since z D and b = sup D,
this is immediate. Thus I D I.
Case 3: Suppose I = (a, ). This is similar to Case 2 and is left to the reader.
Case 4: Suppose I = (, ) = R. Let z R. Since D is unbounded below,
there exists a D with a < z, and since D is unbounded above there exists b D
with z < b. Since D is convex, z D. Thus I = D = I = R.
Step 2: We claim that any subset D which contains I and is contained in I is an
interval. Indeed I and I are both intervals, and the only case in which there is any
subset D strictly in between them is I = (a, b) with a, b R in this case D could
also be [a, b) or (a, b], and both are intervals.

4However, we do not wish to say whether the empty set is an interval. Throughout these
notes the reader may notice minor linguistic contortions to ensure that this issue never arises.

2. INTERVALS AND THE INTERMEDIATE VALUE THEOREM

115

Exercise (R. Freiwald): Let F be an ordered field. For a subset S of F , we define

the associated downset
S = {x F | x s for some s S}.
a) Show that S 6= S 6= .
b) Show that S is bounded above S is bounded above.
c) Show that S has a supremum in F S has a supremum in F , and if so
then we have sup S = sup S .
d) In any ordered field F , we say a subset S of F is convex if x, y S with x < y,
then for all z with x < z < y, z S. Show that for all nonempty subsets S of F ,
the downset S is convex.
e) Deduce the converse to Theorem 6.8: if F is an ordered field in which every
nonempty convex subset is an interval, then F is Dedekind complete.
Recall that a function f : D R satisfies the Intermediate Value Property
(IVP) if for all [a, b] D, for all L in between f (a) and f (b) is of the form f (c)
for some c (a, b). As you may well have noticed, the IVP is closely related to the
notion of a convex subset. The following result clarifies this connection.
Theorem 6.9. For f : D R R, the following are equivalent:
(i) For all [a, b] D, f ([a, b]) is a convex subset of R.
(ii) f satisfies the Intermediate Value Property.
(iii) For any interval I D, f (I) is an interval.
Proof. (i) = (ii): For all [a, b] D, f ([a, b]) is a convex subset containing
f (a) and f (b), hence it contains all numbers in between f (a) and f (b).
(ii) = (iii): Suppose that f satisfies IVP, and let I D be an interval. We want
to show that f (I) is an interval. By Theorem 8.1 it suffices to show that f (I) is
convex. Assume not: then there exists a < b I and some L in between f (a) and
f (b) such that L 6= f (c) for any c I. In particular L 6= f (c) for any c [a, b],
contradicting the Intermediate Value Property.
(iii) = (i): This is immediate: [a, b] is an interval, so by assumption f ([a, b]) is
an interval, hence a convex subset.

2.2. The (Strong) Intermediate Value Theorem.
Theorem 6.10. (Strong Intermediate Value Theorem) If f : I R is continuous, then f satisfies the Intermediate Value Property and thus f (I) is an interval.
Proof. Step 1: We make the following claim: if f : [a, b] R is continuous,
f (a) < 0 and f (b) > 0, then there exists c (a, b) such that f (c) = 0.
proof of claim: Let S = {x [a, b] | f (x) < 0}. Since a S, S is nonempty.
Moreover S is bounded above by b. Therefore S has a least upper bound c = sup S.
It is easy to see that we must have f (c) = 0. Indeed, if f (c) < 0, then as we have
seen several times there exists > 0 such that f (x) < 0 for all x (c , c + ),
and thus there are elements of S larger than c, contradicting c = sup S. Similarly,
if f (c) > 0, then there exists > 0 such that f (x) > 0 for all x (c , c + ), in
which case any element of (c , c) gives a smaller upper bound for S than c. By
the process of elimination we must have f (c) = 0!
Step 2: We will show that f satisfies the Intermediate Value Property: for all
[a, b] I, and any L in between f (a) and f (b), we must find c (a, b) such that

116

6. COMPLETENESS

f (c) = L. If f (a) = f (b) there is nothing to show. If f (a) > f (b), then we may
replace f by f (this is still a continuous function), so it is enough to treat the case
f (a) < L < f (b). Now consider the function g(x) = f (x)L. Since f is continuous,
so is g; moreover g(a) = f (a) L < 0 and g(b) = f (b) L > 0. Therefore by Step
1 there exists c (a, b) such that 0 = g(c) = f (c) L, i.e., such that f (c) = L.
Step 3: By Step 2 and Theorem 6.9, f (I) is an interval.

Remark: Theorem 6.10 is in fact a mild improvement of the Intermediate Value
Theorem we stated earlier in these notes. This version of IVT applies to continuous
functions with domain any interval, not just an interval of the form [a, b], and
includes a result that we previously called the Interval Image Theorem.
2.3. The Intermediate Value Theorem Implies Dedekind Completeness.
Theorem 6.11. Let F be an ordered field such that every continuous function
f : F F satisfies the Intermediate Value Property. Then F is Dedekind complete.
Proof. We will prove the contrapositive: suppose F is not Dedekind complete,
and let S F be nonempty and bounded above but without a least upper bound
in F . Let U(S) be the set of upper bounds of S. We define f : F F by:
f (x) = 1, if x
/ U(S),
f (x) = 1, x U(S).
Then f is continuous on F indeed, a point of discontinuity would occur only at
the least upper bound of S, which is assumed not to exist. Moreover f takes the
value 1 at any element s S, which cannot be an upper bound for S because
then it would be the maximum element of S and the value 1 at any upper bound
for S (we have assumed that S is bounded above so such elements exist), but it
never takes the value zero, so f does not satisfy IVP.

Exercise 2.3: Show in detail that the function f : F F constructed in the proof
of Theorem 6.11 is continuous at every element of F .
3. The Monotone Jump Theorem
Theorem 6.12. (Monotone Jump) Let f : I R be weakly increasing.
a) Let c be an interior point of I. Then limxc f (x) and limxc+ f (x) exist, and
lim f (x) f (c) lim f (x).

xc

xc+

b) Suppose I has a left endpoint a. Then limxa+ f (x) exists and is at least f (a).
c) Suppose I has a right endpoint b. Then limxb f (x) exists and is at most f (c).
Proof. a) Step 0: As usual, we may f is weakly increasing. We define
L = {f (x) | x I, x < c}, R = {f (x) | x I, x > c}.
Since f is weakly increasing, L is bounded above by f (c) and U is bounded below
by f (c). Therefore we may define
l = sup L, r = inf R.
Step 1: For all x < c, f (x) f (c), f (c) is an upper bound for L, so l f (c). For
all c < x, f (c) f (x), so f (c) is a lower bound for R, so f (c) r. Thus
(22)

l f (c) r.

4. REAL INDUCTION

117

Step 2: We claim limxc f (x) = l. To see this, let  > 0. Since l is the least upper
bound of L and l  < l, l  is not an upper bound for L: there exists x0 < c
such that f (x0 ) > l . Since f is weakly increasing, for all x0 < x < c we have
l  < f (x0 ) f (x) l < l + .
Thus we may take = c x0 .
Step 3: We claim limxc+ f (x) = r: this is shown as above and is left to the reader.
Step 4: Substituting the results of Steps 2 and 3 into (22) gives the desired result.
b) and c): The arguments at an endpoint are routine modifications of those of part
a) above and are left to the reader as an opportunity to check her understanding. 
Theorem 6.13. For a monotone function f : I R, TFAE:
(i) f (I) is an interval.
(ii) f is continuous.
Proof. As usual, it is no loss of generality to assume f is weakly increasing.
(i) = (ii): If f is not continuous on all of I, then by the Monotone Jump
Theorem f (I) fails to be convex. In more detail: suppose f is discontinuous at c.
If c is an interior point then either limxc f (x) < f (c) or f (c) < limxc+ f (x).
In the former case, choose any b I, b < c. Then f (I) contains f (b) < f (c) but
not the in-between point limxc f (x). In the latter case, choose any d I, c < d.
Then f (I) contains f (c) < f (d) but not the in-between point limxc+ f (x). Similar
arguments hold if c is the left or right endpoint of I: these are left to the reader.
Thus in all cases f (I) is not convex hence is not an interval.
(ii) = (i): This follows immediately from Theorem 6.10.

With Theorems 6.10 and 6.13 in hand, we get an especially snappy proof of the
Continuous Inverse Function Theorem. Let f : I R be continuous and injective.
By Theorem 6.10, f (I) = J is an interval. Moreover f : I J is a bijection, with
inverse function f 1 : J I. Since f is monotone, so is f 1 . Moreover f 1 (J) = I
is an interval, so by Theorem 6.13, f 1 is continuous!
4. Real Induction
Theorem 6.14. (Principle of Real Induction) Let a < b be real numbers, let
S [a, b], and suppose:
(RI1) a S,
(RI2) for all x S, if x 6= b there exists y > x such that [x, y] S.
(RI3) For all x R, if [a, x) S, then x S.
Then S = [a, b].
Proof. Seeking a contradiction we suppose not: S 0 = [a, b] \ S is nonempty.
It is bounded below by a, so has a (finite!) greatest lower bound inf S 0 . However:
Case 1: inf S 0 = a. Then by (RI1), a S, so by (RI2), there exists y > a such that
[a, y] S, and thus y is a greater lower bound for S 0 then a = inf S 0 : contradiction.
Case 2: a < inf S 0 S. If inf S 0 = b, then S = [a, b]. Otherwise, by (RI2) there
exists y > inf S 0 such that [inf S 0 , y 0 ] S, contradicting the definition of inf S 0 .
Case 3: a < inf S 0 S 0 . Then [a, inf S 0 ) S, so by (RI3) inf S 0 S: contradiction!

Example 4.1: Let us reprove the Intermediate Value Theorem. Recall that the key
special case of IVT, from which the full theorem easily follows, is this: if f : [a, b]

118

6. COMPLETENESS

R is continuous, f (a) < 0 and f (b) > 0, then there exists c (a, b) with f (c) = 0.
We prove this by real induction, as follows. Let S = {x [a, b] | f (x) 0}. We
know that S is proper in [a, b], so applying real induction shows that one of (RI1),
(RI2) and (RI3) must fail. We have a S so (RI1) holds and if a continuous
function is non-negative on [a, c), then it is also non-negative at c: (RI3). So (RI2)
must fail: there exists y (a, b] such that f (y) 0 but there is no  > 0 such that
f is non-negative on [y, y + ). This implies f (y) = 0.
Example 4.1, redux: In class I handled the proof of IVT by Real Induction differently, and in a way which I think gives a better first example of the method (most
Real Induction proofs are not by contradiction). This strategy follows [Ka07].
Namely, IVT is equivalent to: let f : [a, b] R be continuous and nowhere zero. If
f (a) > 0, then f (b) > 0. We prove this by Real Induction. Let
S = {x [a, b] | f (x) > 0}.
Then f (b) > 0 iff b S. We will show S = [a, b] by real induction, which suffices.
(RI1) By hypothesis, f (a) > 0, so a S.
(RI2) Let x S, x < b, so f (x) > 0. Since f is continuous at x, there exists > 0
such that f is positive on [x, x + ], and thus [x, x + ] S.
(RI3) Let x (a, b] be such that [a, x) S, i.e., f is positive on [a, x). We claim
that f (x) > 0. Indeed, since f (x) 6= 0, the only other possibility is f (x) < 0,
but if so, then by continuity there would exist > 0 such that f is negative on
[x, x], i.e., f is both positive and negative at each point of [x, x]: contradiction!
The following result shows that Real Induction does not only uses the Dedekind
completeness of R but actually carries the full force of it.
Theorem 6.15. In an ordered field F , the following are equivalent:
(i) F is Dedekind complete: every nonempty bounded above subset has a supremum.
(ii) F satisfies the Principle of Real Induction: for all a < b F , a subset S [a, b]
satisfying (RI1) through (RI3) above must be all of [a, b].
Proof. (i) = (ii): This is simply a restatement of Theorem 6.14.
(ii) = (i): Let T F be nonempty and bounded below by a F . We will show
that T has an infimum. For this, let S be the set of lower bounds m of T with
a m. Let b be any element of T . Then S [a, b].
Step 1: Observe that b S b = inf T . In general the infimum could be
smaller, so our strategy is not exactly to use real induction to prove S = [a, b].
Nevertheless we claim that S satisfies (RI1) and (RI3).
(RI1): Since a is a lower bound of T with a a, we have a S.
(RI3): Suppose x (a, b] and [a, x) S, so every y [a, x) is a lower bound for T .
Then x is a lower bound for T : if not, there exists t T such that t < x; taking
any y (t, x), we get that y is not a lower bound for T either, a contradiction.
Step 2: Since F satisfies the Principle of Real Induction, by Step 1 S = [a, b] iff S
satisfies (RI2). If S = [a, b], then the element b is a lower bound for T , so it must
be the infimum of T . Now suppose that S 6= [a, b], so by Step 1 S does not satisfy
(RI2): there exists x S, x < b such that for any y > x, there exists z (x, y) such
that z
/ S, i.e., z is not a lower bound for T . In other words x is a lower bound
for T and no element larger than x is a lower bound for T ...so x = inf T .


6. THE HEINE-BOREL THEOREM

119

Remark: Like Dedekind completeness, Real Induction depends only on the ordering relation < and not on the field operations + and . In fact, given any ordered
set (F, <) i.e., we need not have operations + or at all it makes sense to speak
of Dedekind completeness and also of whether an analogue of Real Induction holds.
In [Cl11], I proved that Theorem 6.15 holds in this general context: an ordered set
F is Dedekind complete iff the only it satisfies a Principle of Ordered Induction.
5. The Extreme Value Theorem
Theorem 6.16. (Extreme Value Theorem)
Let f : [a, b] R be continuous. Then:
a) f is bounded.
b) f attains a minimum and maximum value.
Proof. a) Let S = {x [a, b] | f : [a, x] R is bounded}.
(RI1): Evidently a S.
(RI2): Suppose x S, so that f is bounded on [a, x]. But then f is continuous
at x, so is bounded near x: for instance, there exists > 0 such that for all
y [x , x + ], |f (y)| |f (x)| + 1. So f is bounded on [a, x] and also on [x, x + ]
and thus on [a, x + ].
(RI3): Suppose that x (a, b] and [a, x) S. Now beware: this does not say that
f is bounded on [a, x): rather it says that for all a y < x, f is bounded on [a, y].
1
is bounded on [0, y]
These are really different statements: for instance, f (x) = x2
for all y < 2 but it is not bounded on [0, 2). But, as usual, the key feature of this
counterexample is a lack of continuity: this f is not continuous at 2. Having said
this, it becomes clear that we can proceed almost exactly as we did above: since f
is continuous at x, there exists 0 < < x a such that f is bounded on [x , x].
But since a < x < x we know also that f is bounded on [a, x ], so f is
bounded on [a, x].
b) Let m = inf f ([a, b]) and M = sup f ([a, b]). By part a) we have
< m M < .
We want to show that there exist xm , xM [a, b] such that f (xm ) = m, f (xM ) = M ,
i.e., that the infimum and supremum are actually attained as values of f . Suppose
that there does not exist x [a, b] with f (x) = m: then f (x) > m for all x [a, b]
1
and the function gm : [a, b] R by gm (x) = f (x)m
is defined and continuous. By
the result of part a), gm is bounded, but this is absurd: by definition of the infimum,
f (x) m takes values less than n1 for any n Z+ and thus gm takes values greater
than n for any n Z+ and is accordingly unbounded. So indeed there must exist
xm [a, b] such that f (xm ) = m. Similarly, assuming that f (x) < M for all x
1
[a, b] gives rise to an unbounded continuous function gM : [a, b] R, x 7 M f
(x) ,
contradicting part a). So there exists xM [a, b] with f (xM ) = M .

6. The Heine-Borel Theorem
Let S R, and let {XiS
}iI be a family of subsets of R. We say that the family
{Xi } covers S if S iI Xi : in words, this simply means that every element
x S is also an element of Xi for at least one i.
Theorem 6.17. (Heine-Borel) Let {Ui }iI be any covering of the closed,
bounded interval by open intervals Ui . Then the covering {Ui }iI has a finite

120

6. COMPLETENESS

subcovering: there is a finite subset J I such that every x [a, b] lies in Uj for
some j J.
Proof. For an open covering U = {Ui }iI of [a, b], let
S = {x [a, b] | U [a, x] has a finite subcovering}.
We prove S = [a, b] by Real Induction. (RI1) is clear. (RI2): If U1 , . . . , Un covers
[a, x], then some Ui contains [x, x + ] for some > 0. (RI3): if [a, x) S, let ix I
be such that x Uix , and let
S > 0 be such that [x , x] Uix . Since x S,
there is a finite J I with iJ Ui [a, x ], so {Ui }iJ Uix covers [a, x]. 
Exercise: The formulation of the Heine-Borel Theorem given above is superficially
weaker than the standard one. A subset of R is called open if it is a union of open
intervals. In the usual statement of the Heine-Borel Theorem the Ui s are allowed
to be arbitrary open subsets of R. Show that this apparently more general version
can be deduced from Theorem 6.17. (Suggestion: for each x [a, b], there is an
open subset Ux containing x. By definition of open subset, Ux must contain some
open interval Ix containing x. Apply Theorem 6.17 to the open covering {Ix }.)
7. Uniform Continuity
7.1. The Definition; Key Examples.
noindent Let I be an interval and f : I R. Then f is uniformly continuous on I if for every  > 0, there exists a > 0 such that for all x1 , x2 I, if
|x1 x2 | < then |f (x1 ) f (x2 )| < .
In order to show what the difference is between uniform continuty on I and mere
continuity on I i.e., continuity at every point of I let us rephrase the standard
- definition of continuity using the notation above. Namely:
A function f : I R is continuous on I if for every  > 0 and every x1 I, there
exists > 0 such that for all x2 I, if |x1 x2 | < then |f (x1 ) f (x2 )| < .
These two definitions are eerily (and lets admit it: confusingly, at first) similar:
they use all the same words and symbols. The only difference is in the ordering
of the quantifiers: in the definition of continuity, player two gets to hear the value
of  and also the value of x1 before choosing her value of . In the definition of
uniform continuity, player two only gets to hear the value of : thus, her choice of
must work simultaneously or, in the lingo of this subject, uniformly across
all values of x1 I. Thats the only difference. Of course, switching the order of
quantifiers in general makes a big diffference in the meaning and truth of mathematical statements, and this is no exception. Lets look at some simple examples.
Example 6.1: Let f : R R by f (x) = mx + b, m 6= 0. We claim that f is
uniformly continuous on R. In fact the argument that we gave for continuity

. Although we used
long ago shows this, because for every  > 0 we took = |m|
this to show that f is continuous at some arbitrary point c R, evidently the
choice of does not depend on the point c: it works uniformly across all values of
c. Thus f is uniformly continuous on R.

7. UNIFORM CONTINUITY

121

Example 6.2: Let f : R R by f (x) = x2 . This time I claim that our usual
proof did not show uniform continuity. Lets see it in action. To show that f is
continuous at c, we factored x2 c into (x c)(x + c) and saw that to get some
control on the other factor x + c we needed to restrict x to some bounded interval
around c, say [c 1, c + 1]. On this interval |x + c| |x| + |c| |c| + 1 + |c| 2|c| + 1.

) we found that if |x c| < then
So by taking = min(1, 2|c|+1
|f (x) f (c)| = |x c||x + c|


(2|c| + 1) = .
2|c| + 1

But the above choice of depends on c. So it doesnt show that f is uniformly

continuous on R. In fact the function f (x) = x2 is not uniformly continuous on R.
For instance, take  = 1. If it were uniformly continuous, there would have to be
some > 0 such that for all x1 , x2 R with |x1 x2 | < , |x21 x22 | < . But this
is not possible: take any > 0. Then for any x R, x and x + 2 are less than
2
apart, and |x2 (x + 2 )2 | = |x + 4 |. But if I get to choose x after you choose
, this expression can be made arbitrarily large. In particular, if x = 1 , then it is
strictly greater than 1. So f is not uniformly continuous on R.
Remark: In fact a polynomial function f : R R is uniformly continuous on
R if and only if it has degree at most one. The reasoning is similar to the above.
So thats sad: uniform continuity is apparently quite rare. But wait! What if
the domain is a closed, bounded interval I? For instance, by restricting f (x) = x2
to any such interval, it is uniformly continuous. Indeed, we may as well assume
I = [M, M ], because any I is contained in such an interval, and uniform continuity on [M, M ] implies uniform continuity on I. Now we need only use the fact
that we are assuming |c| M to remove the dependence of on c: since |c| M

we have 2|c|+1
) 2M1+1 , so for  > 0 we may take = min(1, 2M1+1 ). This shows
2
that f (x) = x is uniformly continuous on [M, M ].
It turns out that one can always recover uniform continuity from continuity by
restricting to a closed bounded interval: this is the last of our Interval Theorems.
7.2. The Uniform Continuity Theorem.
Let f : I R. For , > 0, let us say that f is (, )-UC on I if for all x1 , x2 I,
|x1 x2 | < = |f (x1 ) f (x2 )| < . This is a sort of halfway unpacking of the
definition of uniform continuity. More precisely, f : I R is uniformly continuous
iff for all  > 0, there exists > 0 such that f is (, )-UC on I.
The following small technical argument will be applied twice in the proof of the
Uniform Continuity Theorem, so advance treatment of this argument should make
the proof of the Uniform Continuity Theorem more palatable.
Lemma 6.18. (Covering Lemma) Let a < b < c < d be real numbers, and let
f : [a, d] R. Suppose that for real numbers 1 , 1 , 2 > 0,
f is (, 1 )-UC on [a, c] and

122

6. COMPLETENESS

f is (, 2 )-UC on [b, d].

Then f is (, min(1 , 2 , c b))-UC on [a, b].
Proof. Suppose x1 < x2 I are such that |x1 x2 | < . Then it cannot be
the case that both x1 < b and c < x2 : if so, x2 x1 > c b . Thus we must
have either that b x1 < x2 or x1 < x2 c. If b x1 < x2 , then x1 , x2 [b, d]
and |x1 x2 | < 2 , so |f (x1 ) f (x2 )| < . Similarly, if x1 < x2 c, then
x1 , x2 [a, c] and |x1 x2 | < 1 , so |f (x1 ) f (x2 )| < .

Theorem 6.19. (Uniform Continuity Theorem) Let f : [a, b] R be continuous. Then f is uniformly continuous on [a, b].
Proof. For  > 0, let S() be the set of x [a, b] such that there exists > 0
such that f is (, )-UC on [a, x]. To show that f is uniformly continuous on [a, b], it
suffices to show that S() = [a, b] for all  > 0. We will show this by Real Induction.
(RI1): Trivially a S(): f is (, )-UC on [a, a] for all > 0!
(RI2): Suppose x S(), so there exists 1 > 0 such that f is (, 1 )-UC on
[a, x]. Moreover, since f is continuous at x, there exists 2 > 0 such that for all
c [x, x+2 ], |f (c)f (x)| < 2 . Why 2 ? Because then for all c1 , c2 [x2 , x+2 ],
|f (c1 ) f (c2 )| = |f (c1 ) f (x) + f (x) f (c2 )| |f (c1 ) f (x)| + |f (c2 ) f (x)| < .
In other words, f is (, 2 )-UC on [x 2 , x + 2 ]. We apply the Covering Lemma to
f with a < x 2 < x < x + 2 to conclude that f is (, min(, 2 , x (x 2 ))) =
(, min(1 , 2 ))-UC on [a, x + 2 ]. It follows that [x, x + 2 ] S().
(RI3): Suppose [a, x) S(). As above, since f is continuous at x, there exists
1 > 0 such that f is (, 1 )-UC on [x 1 , x]. Since x 21 < x, by hypothesis there
exists 2 such that f is (, 2 )-UC on [a, x 21 ]. We apply the Covering Lemma to f
with a < x1 < x 21 < x to conclude that f is (, min(1 , 2 , x 21 (x1 ))) =
(, min( 21 , 2 ))-UC on [a, x]. Thus x S().

8. The Bolzano-Weierstrass Theorem For Subsets
Let S R. We say that x R is a limit point of S if for every > 0, there
exists s S with 0 < |s x| < . Equivalently, x is a limit point of S if every open
interval I containing x also contains an element s of S which is not equal to x.
Proposition 6.20. For S R and x R, the following are equivalent:
(i) Every open interval I containing x also contains infinitely many points of S.
(ii) x is a limit point of S.
Example: If S = R, then every x R is a limit point. More generally, if S R
is dense i.e., if every nonempty open interval I contains an element of S then
every point of R is a limit point of S. In particular this holds when S = Q and
when S = R \ Q. Note that these examples show that a limit point x of S may or
may not be an element of S: both cases can occur.
Example: If S T and x is a limit point of S, x is a limit point of T .
Example: No finite subset S of R has a limit point.
Example: The subset Z has no limit points: indeed, for any x R, take I =
(x 1, x + 1). Then I is bounded so contains only finitely many integers.

9. TARSKIS FIXED POINT THEOREM

123

Example: More generally, let S be a subset such that for all M > 0, S [M, M ]
is finite. Then S has no limit points.
Theorem 6.21. (Bolzano-Weierstrass)
Every infinite subset A of [a, b] has a limit point.
Proof. Let A [a, b], and let S be the set of x in [a, b] such that if A [a, x]
is infinite, it has a limit point. It suffices to show S = [a, b], which we will do by
Real Induction. (RI1) is clear. (RI2) Suppose x [a, b) S. If A [a, x] is infinite,
then it has a limit point and hence so does A [a, b]: thus S = [a, b]. If for some
> 0, A [a, x + ] is finite, then [x, x + ] S. Otherwise A [a, x] is finite but
A [a, x + ] is infinite for all > 0, and then x is a limit point for A and S = [a, b]
as above. (RI3) If [a, x) S, then: either A [a, y] is infinite for some y < x, so
x S; or A [a, x] is finite, so x S; or A [a, y] is finite for all y < x and A [a, x]
is infinite, so x is a limit point of A [a, x] and x S.

Remark: Later we will give a sequential version of Bolzano-Weierstrass and see
that it is equivalent to Theorem 6.21 above.
9. Tarskis Fixed Point Theorem
Mainly as a further showpiece for Real Induction, we give here a special case of
a theorem of the great logician A. Tarski [Ta55]. It is a fixed point theorem:
that is, a theorem which gives conditions on a function f : X X for there to
be a point c X with f (c) = c. Fixed point theorems are ubiquitously useful
throughout mathematics, and further fixed point theorems for functions defined on
subintervals of R will be given when we study infinite sequences later on. However
Tarskis theorem has a quite different flavor from the fixed point theorems to come
in that the function f is not assumed to be continuous!
Theorem 6.22. (Tarski Fixed Point Theorem) Let f : [a, b] [a, b] be a weakly
increasing function. Then f has a fixed point: there is c [a, b] with f (c) = c.
Proof. Seeking a contradiction we suppose f has no fixed point. Let S =
{x [a, b] | f (x) > x}. We will show S = [a, b] by Real Induction. But then b S,
so f (b) > b, contradicting the fact that b is the largest element of [a, b]!
(RI1) Since a = min[a, b], f (a) a; since there is no fixed point, f (a) > a.
(RI2) Let x [a, b) and suppose [a, x] S. Put y = f (x), so y > x by definition
of S. We claim [x, y] S. If not, there is x z y with f (z) < z y; but since
x z, y = f (x) f (z) and we get y f (z) < z y, a contradiction.
(RI3) Let x (a, b] and suppose [a, x) S. We claim x S. If not, y = f (x) < x,
so y S and thus f (y) > y = f (x); but since y < x, f (y) f (x).

Exercise: a) Give another (more standard) proof of Theorem 6.22 by showing that
sup{x [a, b] | f (x) x} is a fixed point of f .
b) Which argument do you prefer?
Exercise: Find an application of Theorem 6.22 to calculus.
I am indebted to J. Propp for alerting me to this application of Real Induction.

CHAPTER 7

Differential Miscellany
1. LH
opitals Rule
We have come to the calculus topic most hated by calculus instructors: LHopitals

.
Rule. This result gives a criterion for evaluating indeterminate forms 00 or
The following succinct formulation is taken from [R, Thm. 5.13], as is the proof.
Theorem 7.1. Let a < b . Let f, g : (a, b) R be differentiable.
a) We suppose that
f 0 (x)
lim 0
= A [, ]
xb g (x)
and also that either
(i) limxb f (x) = limxb g(x) = 0, or
(ii) limxb g(x) = .
(x)
Then limxb fg(x)
= A.
b) The analogous statements with limxb replaced everywhere by limxa+ hold.
Proof. a) Step 1: Suppose that A < , let be any real number which is
greater than A, and let be such that A < < . We will show that there exists
(x)
c (a, b) such that for all x > c, fg(x)
< .
0

(x)
(x)
A < , there is c (a, b) such that for all x > c, fg0 (x)
< .
First, since fg0 (x)
Let c < x < y < b. By Cauchys Mean Value Theorem, there is t (x, y) such that

(23)

f (x) f (y)
f 0 (t)
= 0
<
g(x) g(y)
g (t)

(y)

Suppose first that (i) holds. Then by letting x approach b in (23) we get fg(y)
< for all c < y < b, which is what we wanted to show.
Next suppose that (ii) holds. Fix y in (23) and choose c1 (y, b) such that c1 <
x < b implies g(x) > g(y) and g(x) > 0. Multiplying (23) by g(x)g(y)
gives
g(x)


f (x) f (y)
g(x) g(y)
<
,
g(x)
g(x)

and then a little algebra yields

(24)

f (x)
g(y) f (y)
<
+
.
g(x)
g(x) g(x)

(x)
Letting x approach b, we find: there is c (c1 , b) such that for all x > c, fg(x)
< .
Step 2: Suppose A > . Then arguing in a very similar manner as in Step 1
we may show that for any < A there exists c (a, b) such that for all x > c,
125

126

7. DIFFERENTIAL MISCELLANY

f (x)
g(x)

> . Putting together these two estimates shows limxb

b) This is quite straightforward and left to the reader.1

f (x)
g(x)

= A.


Remark: Perhaps you were expecting the additional hypothesis limxb f (x) =
in condition (ii). As the proof shows, this is not necessary. But it seems to be
very risky to present the result to freshman calculus students in this form!
n

Example 7.1: We claim that for all n Z+ , limx xex = 0. We show this by
induction on n. First we do n = 1: limx exx = 0. Since limx g(x) = 0
and limx

f 0 (x)
g 0 (x)

= limx

1
ex

= 0, condition (ii) of LHopitals Rule applies so

n
limx exx = 0. Induction Step: let n Z+ and suppose limx xex = 0. Then
n

n+1
LH
(n+1)xn
= (n + 1) limx xex = (n + 1) 0 = 0.
limx x ex =
= limx
ex

Why do calculus instructors not like LHopitals Rule? Oh, let us count the ways!
(x)
is of the form 00 . Thus many calcu1) Every derivative f 0 (x) = limh0 f (x+h)f
h
lus students switch to applying LHopitals Rule instead of evaluating derivatives
from the definition. This can lead to painfully circular reasoning. For instance,
what is limx0 sinx x ? Well, both numerator and denominator approach 0 and
0

limx0 (sinx0x) = limx0 cos1 x = cos 0 = 1. Whats wrong with this? Well, how
do we know that (sin x)0 = cos x? Thinking back, we reduced this to computing
the derivative of sin x at x = 0, i.e., to showing that limx0 sinx x = 1!
2) Many limits which can be evaluated using LHopitals Rule can also be evaluated in many other ways, and often just by thinking a bit about how functions
actually behave. For intance, try to evaluate the limit of Example 7.1 above without
using LH
opital. There are any number of ways. For instance:
Lemma 7.2. (Racetrack Principle) Let f, g : [a, ) R be two differentiable
functions such that f 0 (x) g 0 (x) for all x a. Then:
a) We have f (x) f (a) g(x) g(a) for all x a.
b) If f 0 (x) > g 0 (x) for all x > a, then f (x) f (a) > g(x) g(a) for all x > a.
Proof. Put h = f g : [a, ) R.
a) Then h0 (x) 0 for all x a, so h is weakly increasing on (a, ), and thus being
continuous, weakly increasing on [a, ): for all x a, f (x) g(x) = h(x) h(a) =
f (a) g(a), and thus f (x) f (a) g(x) g(a).
b) This is the same as part a) with all instances of replaced by >: details may
be safely left to the reader.

Proposition 7.3. Let f : [a, ) R be twice differentiable such that:
(i) f 0 (a) > 0 and
(ii) f 00 (x) 0 for all x a.
Then limx f (x) = .
Proof. Let g be the tangent line to f at x = a, viewed as a function from
[a, ) to R. Because f 00 = (f 0 )0 is non-negative on [a, ), f 0 is weakly increasing
1In fact, [R] states and proves the result with lim
xa+ instead of limxb . I recast it this
way since a natural class of examples concerns limx .


1. LHOPITALS
RULE

127

on [a, ), and thus for all x a,

f 0 (x) f 0 (a) = g 0 (x).
Applying the Racetrack Principle, we get that for all x a,
f (x) f (a) g(x) g(a),
or
f (x) g(x) + f (a) g(a) = g(x).
Since g is a linear function with positive slope,
lim f (x) lim g(x) = .


x

(i) Let fn (x) = xen . One can therefore establish limx xen = by showing that
fn0 (x), fn00 (x) are both positive for sufficiently large x. It is easy to see that fn0 (x) > 0
for all x > n. The analysis for fn00 is a bit messier; we leave it to the reader and try
something slightly different instead.
(ii) Since fn0 (x) > 0 for all x > n, f is eventually increasing and thus tends either
to a positive limit A or to +. But as x , x + 1 , so
ex+1
ex
= e lim
= eA.
n
x (x + 1)
x (x + 1)n

A = lim

The only A (0, ] which satisfies A = eA is A = .

x
(iii) Take logarithms: if A = limx xen , then
ex
= lim x n log x.
x
x
xn
Now if l(x) = x n log x, then l0 (x) = 1 nx , l00 (x) = xn2 ; both are positive for large
enough x, so by Proposition 7.3, log A = and thus A = .
(iv) When we learn about Taylor series we will have access to a superior expression
P n
for ex , namely n=0 xn! . From this the desired limit follows almost immediately!
log A = lim log

3) The statement of LH
opitals Rule is complicated and easy for even relatively
proficient practitioners of the mathematical craft to misremember or misapply. A
classic rookie mistake is to forget to verify condition (i) or (ii): of course in general
f 0 (x)
f (x)
6= lim 0
;
xa g (x)
xa g(x)
lim

try a random example. But there are subtler pitfalls as well. For instance, even
0
(x)
(x)
under conditions (i) and (ii), limxa fg(x)
= A need not imply that limxa fg0 (x)
exists, so you cannot use LHopitals Rule to show that a limit does not exist.
Example 7.2: Let f, g : R R by f (x) = x2 sin( x1 ) and f (0) = 0 and g(x) = x.
Then f and g are both differentiable (for f this involves going back to the limit
definition of the derivative at x = 0 we have seen this example before), and
(x)
limx0 fg(x)
= limx0 x sin( x1 ) = 0. However,
 
 
 
f 0 (x)
1
1
1
lim
= lim 2x sin
cos
= lim cos
,
x0 g 0 (x)
x0
x0
x
x
x

128

7. DIFFERENTIAL MISCELLANY

and this limit does not exist.

Nevertheless, every once in a while one really does need LHopitals Rule! We
will encounter such a situation in our study of Taylor polynomials. And, you know,
everyone else is doing it, so you should at least know how to do it...
Exercise: Use LH
opitals Rule to give a short proof of Theorem 5.32.
2. Newtons Method
2.1. Introducing Newtons Method.
Newtons Method is an important procedure for approximating roots of differentiable functions. Namely, suppose that y = f (x) is a differentiable function and
that we know perhaps by the methods of calculus! that there is a real number
c such that f (c) = 0. Well, in general there are many, but suppose we restrict ourselves to some small interval in which we believe there is a unique root c. Say we
do not need to know c exactly, but that we wish to approximate to any prescribed
degree of accuracy e.g. we may need to know its value to 100 decimal places.
The first key idea is that Newtons method is one of sucessive approximations. That is, we start with a number x1 which is an approximate root of f ,
i.e., f (x1 ) is rather close to 0 (this is not a precise mathematical statement, but
the closer x1 is to the true root c, the better the method will work. If its too far
away, then it wont work at all.) Then we perform some ameloriation procedure
resulting in a second approximate root x2 , which is (in general) closer to the true
root c than x1 is. And then we continue: performing our amelioration procedure
again we get x3 , performing it a third time we get x4 , and so forth, resulting in an
infinite sequence of approximate roots {xn }
n=1 .
This amelioration procedure is very geometric: let n Z+ , and start with the
approximate root xn . What we do is consider the tangent line ln (x) to y = f (x) at
x = xn . The equation of this line is
y f (xn ) = f 0 (xn )(x xn ),
so
y = ln (x) = f (xn ) + f 0 (xn )(x xn ).
Now we take xn+1 to be x-intercept of the line ln (x), i.e., the unique number such
that ln (xn+1 ) = 0. So lets do it:
0 = ln (xn+1 ) = f (xn ) + f 0 (xn )(xn+1 xn ),
so
xn+1 xn =

f (xn )
f 0 (xn )

xn+1 = xn

f (xn )
.
f 0 (xn )

or
(25)

Note that our expression for xn+1 is undefined if f 0 (xn ) = 0, as well it should be:
if the tangent line at xn is horizontal, then either it coincides with the x-axis (in
which case xn is already a root of f and no amelioration is needed) or it is parallel

2. NEWTONS METHOD

129

to the x-axis, in which case the method breaks down: in a sense we will soon make
precise, this means that xn is too far away from the true root c of f .
2.2. A Babylonian Algorithm.

We can use Newtons method to approximate 2. Consider f (x) = x2 2; straightforward calculus tells us that there is a unique positive number c such that f (c) = 2
and that for instance c [1, 2]. We compute the amelioration formula in this case:


2x2 (x2n 2)
x2 + 2
1
2
x2 2
= n
= n
=
xn +
.
(26)
xn+1 = xn n
2xn
2xn
2xn
2
xn
In other words, to get from xn to xn+1 we take the average of xn and

2
xn .

In this case we are allowed to take any x1 6= 0; if we want to approximate the

positive root of x2 2 = 0, it is more than plausible that we should start with a
positive number. So lets try
x1 = 1.
Then


1
2
3
x2 =
1+
= = 1.5.
2
1
2


1 3 4
17
x3 =
+
= 1.41666 . . .
=
2 2 3
12


577
1 17 24
+
=
= 1.414215686 . . . .
x4 =
2 12 17
408


1 577 816
x5 =
+
= 665857/470832 = 1.41421356237468991 . . . .
2 408 577


1 470832 1331714
+
x6 =
= 886731088897/627013566048 = 1.41421356237309504880 . . . .
2 665857
470832

If I now ask my laptop computer to directly compute 2, then it tells me2 that

2 = 1.414213562373095048801688724 . . . .
Thus x5 is accurate to 11 decimal places and x6 is accurate to 23 decimal places.
Looking more carefully, it seems that each iteration of the amelioration process
xn 7 xn+1 roughly doubles the number of decimal places of accuracy. If this holds
true, it means
that the approximations get close to the true root very fast it we

wanted 2 to 100 decimal places we would only need tocompute x9 .

The formula (26) for successive approximations to 2 was known to the ancient Babylonians, thousands of years before tangent lines, calculus of any kind,
and Isaac Newton. So in practice Newtons method seems to work very well indeed.
Remark: Similarly, for any a > 0, using Newtons method as above with f (x) =
x2 a leads to the recursion


a
1
xn +
,
xn+1 =
2
xn
2Of course one should not neglect to wonder how my computer is doing this computation. I
dont have access to the source code the software I used, so I dont really know, but it is plausible
that it is in fact using some form Newtons method.

130

7. DIFFERENTIAL MISCELLANY

application
of which with x1 = 1 leads to fantastically good numerical approxima
tions to a. If you ever find yourself on a desert island and needing to compute a
to many decimal placees as part of your engineering research to build a raft that will
carry you back to civilization, then this is probably the method you should use.
And now if anyone asks you whether honors calculus contains any practically
useful information, you must tell them that the answer is yes!
2.3. Questioning Newtons Method.
Of course we havent proven anything yet. Here are two natural questions:
Question 7.4. Let f : I R be differentiable, and let c I be such that
f (c) = 0.
a) Is there some subinterval (c , c + ) about the true root c such that starting
Newtons method with any x1 (c , c + ) guarantees that the sequence of approximations {xn } gets arbitrarily close to c?
b) Assuming the answer to part a) is yes, given some x1 (c , c + ) can we give
a quantitative estimate on how close xn is to c as a function of n?
Questions like these are explored in a branch of mathematics called numerical
analysis. Most theoretical mathematicians (e.g. me) know little about it, which
is a shame because the questions its treats are fundamental and closely related to
pure mathematics. (As well as being useful in applications, of course.)
2.4. Introducing Infinite Sequences.
We will give some answers to these questions. First, the business of the xn s getting
arbitrarily close to c should be construed in terms of a limiting process, but one of
a kind which is slightly different and in fact simpler than the limit of a real-valued
function at a point. Namely, a real infinite sequence xn is simply an ordered list
of real numbers x1 , x2 , . . . , xn , . . ., or slightly more formally, is given by a function
from the positive integers Z+ to R, say f (n) = xn . If L R, we say the infinite
sequence {xn } converges to L and write xn L if for all  > 0 there exists
N Z+ such that for all n N , |xn L| < . This is precisely the definition
of limx f (x) = L except that our function f is no longer defined for all (or all
sufficiently large) real numbers but only at positive integers. So it is a very close
cousin of the types of limit operations we have already studied.
Here is one very convenient property of limits of sequences.
Proposition 7.5. Let {xn }
n=1 be a sequence of real numbers, and let f : R
R be a continuous function. Suppose that xn L. Then f (xn ) f (L).
Proof. Fix  > 0. Since f is continuous at L, there exists > 0 such that
|x L| < = |f (x) f (L)| < . Moreover, since xn L, there exists a positive
integer N such that for all n N , |xn L| < . Putting these together: if n N
then |xn L| < , so |f (xn ) f (L)| < . This shows that f (xn ) L.

Remark: a) Proposition 7.5 is a close cousin of the fact that compositions are continuous functions are continuous, and in particular the proof is almost the same.
b) At the moment we are just getting a taste of infinite sequences. Later in the
course we will study them more seriously and show that Proposition 7.5 has a very

2. NEWTONS METHOD

131

important converse: if f : R R is a function such that whenever xn L, we have

also f (xn ) f (L), then f is continuous. Thus preservation of limits of sequences
is actually a characteristic property of continuous functions, and this suggests (correctly!) that sequences can be a powerful tool in studying functions f : R R.
Thus Newtons method starts with x1 R and produces an infinite sequence {xn }
of successive aproximations, and our first question is whether or more precisely,
when, i.e., for which choices of x1 this sequence converges to the true root c.
2.5. Contractions and Fixed Points.
Recall that a fixed point of f : I R is a point c I with f (c) = c.
A function f : I R is contractive (or is a contraction) if there is < 1
such that for all x, y I, |f (x) f (y)| |x y|; a real number for which such
an inequality holds will be called a contraction constant for f .
Exercise: a) Which functions f : I R have contraction constant 0?
b) Show that every contraction f : I R is continuous.
Proposition 7.6. Let f : I R be differentiable, and suppose there is < 1
such that for all x I, |f 0 (x)| . Then is a contraction constant for f .
Proof. Let x < y I. By the Mean Value Theorem, there is c (x, y) such
that
f (x) f (y)
= f 0 (c),
xy
so
|f (x) f (y)| = |f 0 (c)||x y| |x y|.

Lemma 7.7. A contractive function f : I R has at most one fixed point.
Proof. Suppose there are x1 6= x2 in I with f (x1 ) = x1 and f (x2 ) = x2 . Let
< 1 be a contraction constant for f . Then |x1 x2 | = |f (x1 )f (x2 )| |x1 x2 |;
dividing through by |x1 x2 | gives 1 : contradiction.

Suppose f : I R is a contraction with constant , and let c I be a fixed point
of f . Then for any > 0, if x [c , c + ], then
|f (x) c| = |f (x) f (c)| |x c| < |x c| ,
so f : [c , c + ] [c , c + ]. This is a key point, because given any f : I I
and any x1 [c , c + ] I, we may define a sequence
x1 , x2 = f (x1 ), x3 = f (x2 ) = f (f (x1 )), . . . , xn+1 = f (xn ) = (f f )(x1 ), . . . .
We call this the sequence of iterates of x1 under f.
Example: Let R, and consider the function
f : R R, f (x) = x.
Let us study the dynamics of iteration of this very simple function.
Step 1: What are the fixed points? We set c = f (c) = c. Then c = 0 is a fixed
point no matter what is. Conversely, if c 6= 0 then we divide through to get

132

7. DIFFERENTIAL MISCELLANY

= 1, so if 6= 1 then 0 is the only fixed point. Finally, if = 1 then every c R

is a fixed point, i.e., f is the identity function.
Step 2: Let us try to figure out the limiting behavior of the sequences of iterates.
First observe that for any fixed point c, the sequence of iterates will be constant:
c, c, c, c . . . ,
so of course it will converge to c. So really we are interested in the case when
x1 is not a fixed point of f . In fact we can and this is an unusual feature
arising because of the very simple f we choose give an explicit formula for the
general term xn in the sequence of iterates. Starting at x1 we have x2 = x,
x3 = x2 = (x1 ) = 2 x1 , and so on: in general we have xn = n1 x1 .
Case 1: If = 0, then no matter what x1 is, for all n 2, xn = 0. Thus the
sequence is eventually constant, with eventual value the unique fixed point: 0.
Case 2: If = 1, then for all n, xn = x1 , so the sequence is constant, and this
constant value is a fixed point of f .
Case 3: If 0 < || < 1, then xn = n1 x1 0. Here, no matter what the initial
point x1 is, the sequence of iterates converges to the unique fixed point c = 0.
Case 4: If 1 < ||, then for all x1 6= 0, |xn | = |n1 x1 | : the sequence of
iterates grows without bound, in fact with exponential speed. In particular it does
not converge. One can say more about the signs if desired: e.g. when x1 > 0
and > 1, then every xn is positive and the sequence of iterates approaches ,
whereas if x1 > 0 and < 1, then the terms xn alternate in sign while increasing
in absolute value, so they do not approach either or .
Here are three observations about the above example:
1) If x1 is a fixed point of f , then the equence of iterates is constant.
2) Whenever the sequence of iterates converges, it converges to a fixed point of f .
3) The sequence of iterates converges for all initial points x1 R iff || < 1. But
note that |f (x) f (y)| = |||x y|, so f is contractive iff || < 1.
The first observation clearly holds for the iterates of any function f : I I.
What about the last two observations? In fact they also hold quite generally.
Lemma 7.8. Suppose f : I I is continuous, x1 I, and the sequence of
iterates {xn } of x1 under f converges to L I. Then L is a fixed point of f .
Proof. Since f is continuous and xn L, by Proposition 7.5 we have f (xn )
f (L). But f (xn ) = xn+1 , and if xn L then certainly xn+1 L as well. A
sequence can have at most one limit, so f (L) = L.

Lemma 7.9. (Contraction Lemma) For c R and > 0, put I = [c , c + ].
Suppose that f : I R is a contraction with constant and that f (c) = c.
a) For all x I, f (x) I, i.e., f : I I.
b) For any x1 I, define a sequence {xn }
n=1 by xn+1 = f (xn ) for all n 1. Then
for all n Z+ , we have |xn+1 c| n |x1 c|. In particular xn c.
Proof. a) This was established above. We repeat the statement here because
it is a key point: in order to be able to iterate x1 under f we need f : I I.

2. NEWTONS METHOD

133

b) We compute |xn+1 c| = |f (xn ) f (c)| |xn c| = |f (xn1 ) f (c)|

2 |xn1 c| = 3 |xn2 c| = . . . = n |x1 c|.

2.6. Convergence of Newtons Method.
Believe it or not, we are now very close to a convergence theorem for the sequence
of iterates generated by Newtons method. Let f : I R be a C 2 function i.e.,
the second derivative exists and is continuous and let c be a point of the interior
of I such that f (c) = 0 and f 0 (c) 6= 0 a simple root. Note that since f 0 (c) > 0, f
is increasing through c, so there exists > 0 such that f is negative on [c , c) and
positive on (c, c + ]. In particular c is the unique root of f on [c , c + ]. What
we want to show is that possibly after shrinking for any choice of initial approximation x1 [c, c+], the Newtons method sequence converges rapidly to c.
So...what does the material we have just developed have to do with this problem? At first sight it does not seem relevant, because we have been talking about
fixed points and are now interested in roots, we have been talking about iterating
functions on [c, c+] and we probably cannot iterate f on this interval: f (c) = 0
and 0 need not lie in [c , ], and we have been talking about contractions and f
need not be a contraction. What gives?
The answer is that all of our preparations apply not to f but to some auxiliary
function defined in terms of f . To figure out what this function is, consider the
recursive definition of the Newtons method sequence:
xn+1 = xn

f (xn )
.
f 0 (xn )

Thus the sequence {xn }

n=1 is generated by repeatedly applying a certain function...its just not the function f . Rather it is the amelioration function
T (x) = x

f (x)
.
f 0 (x)

Now we have to check that our setup does apply to T , quite nicely. First, observe
that a point x is a root of f if and only if it is a fixed point of T . Since by our
assumption c is the unique root of f in [c , c + ], c is the unique fixed point of
T on this interval.
The next order of business is to show that T is contractive, at least in some
smaller interval around c. For this we look at the derivative:
T 0 (x) = 1

f 0 (x)f 0 (x) f (x)f 00 (x)

f (x)f 00 (x)
=
.
(f 0 (x))2
(f 0 (x))2

This is all kosher, since we have assumed f 0 is nonzero on [c , c + ] and also

that f is C 2 . In fact, since f is C 2 , T 0 is continuous. But now a miracle occurs:
T 0 (c) = 0. Since T 0 is continuous at c, this means that by making smaller we
may assume that |T 0 (x)| for all x [c , c + ], for any positive we want!
Thus not only can we make T contractive, we can make it contractive with any
contractive constant (0, 1) we want! Thus we get the following result.

134

7. DIFFERENTIAL MISCELLANY

Theorem 7.10. (Convergence of Newtons Method) Let f : I R be a C 2

function, i.e., f 00 exists and is continuous. Let c R be such that f (c) = 0 and
f 0 (c) 6= 0. Fix (0, 1). Then there exists > 0 such that for all x1 [c , c + ],
n)
the Newtons method sequence xn+1 = xn ff0(x
(xn ) is well-defined and converges
rapdily to c in the following sense: for all n Z+ , |xn+1 c| n |x1 c|.
Proof. Fix (0, 1). As above, since T 0 (c) = 0, there exists > 0 such that
is a contraction constant for T 0 on [c , c + ], and therefore, by Lemma 7.9a),
T maps [c , c + ] back into [c , c + ], so for any x1 [c , c + ] the sequence
of iterates xn+1 = T (xn ) is well-defined, and note that it is precisely the Newtons
method sequence with initial approximation x1 . By Lemma 7.9b), for all n Z+ ,
|xn+1 c| n |x1 c|. So xn c, and indeed it does so exponentially fast. In
1
and 1, the above estimate ensures that xn approximates
fact, if we take = 10
c to within n decimal places of accuracy.

2.7. Quadratic Convergence of Newtons Method.
Although Theorem 7.10 is a very satisfying result, it is far from the last word
on Newtons method. In fact, if we compare the proven convergence rate of Theorem 7.10 with the empirically observed convergence rate of f (x) = x2 2, x1 = 1,
we see that what we have proved is not as good as what we observed: in our example the number of decimal places of accuracy doubled with each iteration, i.e., is an
exponential function of n, whereas we proved that the number of decimal places of
accuracy must grow at least linearly with n. Thats a big difference! And in fact,
if you look back at exactly what we proved, it seems to itself suggest that more is
true: namely, we get exponential convergence, but the base of the exponential gets
better and better as we get closer to the point, i.e., as we get farther out into the
sequence of iterates. This is faster than exponential decay with any fixed base.
I believe it should be possible to use the idea expressed at the end of the last
paragraph to give a proof of the quadratic convergence of Newtons Method. For
this, an essential tool is Taylors Theorem With Remainder, an important
result which we will unfortunately not treat until much later on. In fact it is possible to give a short, elementary proof of the quadratic convergence which not only
avoids Taylors theorem but also avoids our analysis of the contractive properties
of the amelioration map T . Here it is.
Theorem 7.11. (Quadratic Convergence of Newtons Method)
Let f : I R be twice differentiable, and let c I be a root of f .
a) Suppose there are real numbers , A, B > 0 such that for all x [c , c + ],
|f 0 (x)| A and |f 00 (x)| B. For any x0 [c , c + ], let {xn } be the Newtons
Method sequence with initial value x0 . Then for all n N,
B
(27)
|xn+1 c| |xn c|2 .
A
b) If f 00 is continuous on I and f 0 (c) 6= 0, then there are indeed , A, B > 0 as in
the statement of part a), so that (27) holds for all x0 [c , c + ].
Proof. a) (W. Miller) First, for a, b R, we denote by |[a, b]| the set of all
real numbers on the closed line segment from a to b; precisely |[a, b]| = [a, b] if a b
and |[a, b] = [b, a] if a > b. Similarly, let |(a, b)| = |[a, b]| \ {a, b} be the open line
segment from a to b.

2. NEWTONS METHOD

135

Let x0 [c, c+], and let {xn } be the Newtons Method sequence of iterates.
It will be useful to rewrite the defining recursion as
n N, xn+1 xn =

f (xn )
.
f 0 (xn )

Apply the Mean Value Theorem to f on |[xn , c]|: there is yn |(xn , c)| such that
f (xn )
f (xn ) f (c)
=
= f 0 (yn ).
xn c
xn c
Apply the Mean Value Theorem to f 0 on |[xn , yn ]|: there is zn |(xn , yn ))| such
that
f 0 (xn ) f 0 (yn )
= f 00 (zn ).
xn yn
The rest of the proof is a clever calculation: for n N, we have


 f (xn )
f (xn ) 

+ 0
|xn+1 c| = |(xn+1 xn ) + (xn c)| =  0
f (xn )
f (yn ) 

 

 f (xn )(f 0 (xn ) f 0 (yn ))   f 0 (xn )(xn c)(f 0 (xn ) f 0 (yn )) 
=

= 
 

f 0 (xn )f 0 (yn )
f 0 (xn )f 0 (yn )
 00


 0
 f (zn ) 
 f (xn ) f 0 (yn ) 




=
 |xn c| =  f 0 (yn )  |xn yn ||xn c|
f 0 (yn )
 00

 f (zn ) 

 |xn c|2 B |xn c|2 .
 0
f (yn ) 
A
In the second to the last inequality above, we used the fact that since yn lies between
xn and c, |xn yn | |xn c|.
b) This is left as an exercise for the reader.

Exercise: Prove Theorem 7.11b).
Let c be a real number, C a positive real number, and let {xn }
n=0 be a sequence
of real numbers. We say that the sequence {xn } quadratically converges to c if
(QC1) xn c, and
(QC2) For all n N, |xn+1 c| C|xn c|2 .
Exercise:
a) Show that a sequence to satsify (QC2) but not (QC1).
1
b) Suppose that a sequence {xn }
n=0 satisfies |x0 c| < min(1, C ). Show that {xn }
quadratically converges to c.
c) Deduce that under the hypotheses of Theorem 7.11, there is > 0 such that for
all x0 [c , c + ], the Newtons Method sequence quadratically converges to c.
Exercise: Let {xn } be a sequence which is quadrtically convergent to c R. Viewing xn as an approximation to c, one often says that the number of decimal places
of accuracy roughly doubles with each iteration.
a) Formulate this as a statement about the sequence dn = log10 (|xn c|).
b) Prove the precise statement you formulated in part a).

136

7. DIFFERENTIAL MISCELLANY

2.8. An example of nonconvergence of Newtons Method.

Consider the cubic function f : R R by f (x) = x3 2x + 2, so
x3n 2xn + 2
.
3x2n 2
Take x1 = 0. Then x2 = 1 and x3 = 0, so the sequence of iterates will then
alternate between 0 and 1: one calls this type of dynamical behavior a 2-cycle.
(The unique real root is at approximately 1.769, so we are plenty far away from it.)
xn+1 = xn

This example is the tip of a substantial iceberg: complex dynamics. If you

consider cubic polynomials as functions from the complex numbers C to C, then
you are well on your way to generating those striking, trippy pictures of fractal
sets that have been appearing on tee-shirts for the last twenty years. I recommend
[Wa95] as an especially gentle, but insightful, introduction.
3. Convex Functions
3.1. Convex subsets of Euclidean n-space.
Let n be a positive integer and let Rn be Euclidean n-space, i.e., the set of
all ordered n-tuples (x1 , . . . , xn ) of real numbers. E.g. for n = 2 we get the plane
R2 , for n = 3 one gets three space R3 , and for higher n the visual/spatial intuition is less immediate but there is no mathematical obstruction. For the most
part Rn is the subject of the sequel to this course multivariable mathematics
but lets digress a little bit to talk about certain subsets of Rn . In fact for our
applications we need only n = 1 and n = 2.
Given two points P = (x1 , . . . , xn ), Q = (y1 , . . . , yn Rn we can add them:
P + Q = (x1 + y1 , . . . , xn + yn ),
a generalization of vector addition that you may be familiar with from physics.
Also given any real number , we can scale any point P by , namely
P = (x1 , . . . , xn ) := (x1 , . . . , xn ).
In other words, we simply multiply every coordinate of P by .
Let P 6= Q be points in Rn . There is a unique line passing through P and Q.
We can express this line parameterically as follows: for every R, let
R = (1 )P + Q = ((1 )x1 + y1 , . . . , (1 )xn + yn ).
In particular R0 = P and R1 = Q. Thus the line segment P Q is given as the set
{R | 0 1}.
Now here is the basic definition: a subset Rn is convex if for all P, Q , the
line segment P Q is contained in . In other words, if our physical universe is the
subset and we stand at any two points of , then we can see each other : we have
an unobstructed line of view that does not at any point leave . There are many
convex subsets in the plane: e.g. interiors of disks, ellipses, regular polygons, the
portion of the plane lying on one side of any straight line, and so forth.

3. CONVEX FUNCTIONS

137

Exercise: Let T
1 , . . . , n be convex subsets of Rn .
n
a) Show that i=1 i the
Snset of all points lying in every i is convex.
b) Show by example that i=1 i need not be convex.
When n = 1, convex subsets are quite constrained. Recall we have proven:
Theorem 7.12. For a subset R, the following are equivalent:
(i) is an interval.
(ii) is convex.
3.2. Goals.
In freshman calculus one learns, when graphing a function f , to identify subintervals on which the graph of f is concave up and intervals on which it is concave
down. Indeed one learns that the former occurs when f 00 (x) > 0 and the latter
occurs when f 00 (x) < 0. But, really, what does this mean?
First, where freshman calculus textbooks say concave up the rest of the mathematical world says convex ; and where freshman calculus textbooks say concave
down the rest of the mathematical world says concave. Moreover, the rest of the
mathematical world doesnt speak explicitly of concave functions very much because it knows that f is concave exactly when f is convex.
Second of all, really, whats going on here? Are we saying that our definition
of convexity is that f 00 > 0? If so, exactly why do we care when f 00 > 0 and when
f 00 < 0: why not look at the third, fourth or seventeenth derivatives? The answer is
that we have not a formal definition but an intuitive conception of convexity, which
a good calculus text will at least try to nurture: for instance I was taught that a
function is convex (or rather concave up) when its graph holds water and that
it is concave (concave down) when its graph spills water. This is obviously not
a mathematical definition, but it may succeed in conveying some intuition. In less
poetic terms, the graph of a convex function has a certain characteristic shape that
the eye can see: it looks, in fact, qualitatively like an upward opening parabola or
some portion thereof. Similarly, the eye can spot concavity as regions where the
graph looks, qualitatively, like a piece of a downward opening parabola. And this
explains why one talks about convexity in freshman calculus: it is a qualitative,
visual feature of the graph of f that you want to take into account. If you are
graphing f and you draw something concave when the graph is actually convex,
the graph will look wrong and you are liable to draw false conclusions about the
behavior of the function.
So, at a minimum, our task at making good mathematical sense of this portion
of freshman calculus, comes down to the following:
Step 1: Give a precise definition of convexity: no pitchers of water allowed!
Step 2: Use our definition to prove a theorem relating convexity of f to the second
derivative f 00 , when f 00 exists.
In fact this is an oversimplification of what we will actually do. When we try
to nail down a mathematical definition of a convex function, we succeed all too
well: there are five different definitions, each having some intuitive geometric appeal and each having its technical uses. But we want to be talking about one class

138

7. DIFFERENTIAL MISCELLANY

of functions, not four different classes, so we will need to show that all five of our
definitions are equivalent, i.e., that any function f : I R which satisfies any one
of these definitions in fact satisfies all four. This will take some time.
3.3. Epigraphs.
For a function f : I R, we define its epigraph to be the set of points (x, y) IR
which lie on or above the graph of the function. In fewer words,
Epi(f ) = {(x, y) I R | y f (x)}.
A function f : I R is convex if its epigraph Epi(f ) is a convex subset of R2 .
Example: Any linear function f (x) = mx + b is convex.
Example: The function f (x) = |x| is convex.
Example: Suppose f (x) = ax2 + bx + c. Then Epi(f ) is just the set of points
of R2 lying on or above a parabola. From this picture it is certainly intuitively
clear that Epi(f ) is convex iff a > 0, i.e., iff the parabola is opening upward. But
proving from scratch that Epi(f ) is a convex subset is not so much fun.
3.4. Secant-graph, three-secant and two-secant inequalities.
A function f : I R satisfies the secant-graph inequality if for all a < b I
and all [0, 1], we have
(28)

f ((1 )a + b) (1 )f (a) + f (b).

As ranges from 0 to 1, the expression (1 )a + b is a parameterization of the

line segment from a to b. Similarly, (1 )f (a) + f (b) is a parameterization of
the line segment from f (a) to f (b), and thus
7 ((1 )a + b, (1 )f (a) + f (b))
parameterizes the segment of the secant line on the graph of y = f (x) from
(a, f (a)) to (b, f (b)). Thus the secant-graph inequality is asserting that the graph
of the function lies on or below the graph of any of its secant line segments.
A function f : I R satisfies the three-secant inequality if for all a < x < b,
f (b) f (a)
f (b) f (x)
f (x) f (a)

.
xa
ba
bx
A function f : I R satisfies the two-secant inequality if for all a < x < b,
(29)

(30)

f (x) f (a)
f (b) f (a)

.
xa
ba

Proposition 7.13. For a function f : I R, the following are equivalent:

(i) f satisfies the three-secant inequality.
(ii) f satisfies the two-secant inequality.
(iii) f satisfies the secant-graph inequality.
(iv) f is convex, i.e., Epi(f ) is a convex subset of R2 .

3. CONVEX FUNCTIONS

139

Proof. We will show (i) = (ii) (iii) = (i) and (iii) (iv).
(i) = (ii): This is immediate.
(ii) (iii): The two-secant inequality
f (x) f (a)
f (b) f (a)

xa
ba
is equivalent to

f (x) f (a) +

f (b) f (a)
ba


(x a) = La,b (x),

say. Now La,b (x) is a linear function with La,b (a) = f (a) and La,b b() = f (b), hence
it is the secant line between (a, f (a)) and (b, f (b)). Thus the two-secant inequality
is equivalent to the secant-graph inequality.
(iii) = (i): As above,
 secant line La,b (x) from (a, f (a)) to (b, f (b)) has
 snce the
equation y = f (a) +

f (b)f (a)
ba

(x a), the secant graph inequality implies

f (x) f (a)
f (b) f (a)

.
xa
ba
To get the other half of the three-secant inequality, note that we also have
La,b (x) = f (b) +
and the inequality f (x) f (b) +

f (a) f (b)
(b x),
ba

f (a)f (b)
(b
ba

x) is easily seen to be equivalent to

f (b) f (a)
f (b) f (x)

.
ba
bx
(iii) = (iv): Let P1 = (x1 , y1 ), P2 = (x2 , y2 ) Epi(f ). We want to show Epi(f )
contains the line segment joining P1 and P2 . This is clear if x1 = x2 = x in
this case the line segment is vertical, since since y1 and y2 are both greater than or
equal to f (x), so is every point y in between y1 and y2 . So we may assume x1 6= x2
and then that x1 < x2 (otherwise interchange P1 and P2 ). Seeking a contradiction,
we suppose there is 1 (0, 1) such that (1 1 )P1 + 1 P2
/ Epi(f ): that is,
(1 1 )y1 + 1 y2 < f ((1 1 )x1 + 1 x2 ).
But since f (x1 ) y1 and f (x2 ) y2 , we have
(1 1 )f (x1 ) + 1 f (x2 ) (1 1 )y1 + 1 y2
and thus
(1 1 )f (x1 ) + 1 f (x2 ) < f ((1 1 )x1 + 1 x2 ),
violating the secant-graph inequality.
(iv) = (iii): Let x < y I. Since (x, f (x)) and (y, f (y)) lie on the graph of f ,
they are elements of the epigraph Epi(f ). Since Epi(f ) is convex the line segment
joining (x, f (x)) and (y, f (y)) lies inside Epi(f ). But this line segment is nothing
else than the secant line between the two points, and to say that it lies inside the
epigraph is to say that the secant line always lies on or above the graph of f . 
Corollary 7.14. (Generalized Two Secant Inequality) Let f : I R be a
convex function, and let a, b, c, d I with a < b c < d. Then
(31)

f (b) f (a)
f (d) f (c)

.
ba
dc

140

7. DIFFERENTIAL MISCELLANY

Proof. If b = c then (31) is the Two Secant Inequality, so we may suppose

b < c. Applying the Two Secant Inequality to the points a, b, c gives
f (b) f (a)
f (c) f (b)

ba
cb
and applying it to the points b, c, d gives
f (c) f (b)
f (d) f (c)

.
cb
dc
Combining these two inequalities gives the desired result.

Exercise (Interlaced Secant Inequality): Let f : I R be a convex function, and

let a < b < c < d I. Show that we have
f (d) f (b)
f (c) f (a)

.
ca
db
3.5. Continuity properties of convex functions.
Theorem 7.15. Suppose f : I R is convex, and let [a, b] be any subinterval
of I . Then f is a Lipschitz function on [a, b]: there exists a constant C such that
for all x, y [a, b], |f (x) f (y)| C|x y|.
Proof. Choose u, v, w, z I with u < v < a and b < w < z. Applying the
Generalized Two Secant Inequality twice we get
f (v) f (u)
f (y) f (x)
f (z) f (w)

.
vu
yx
zw
Thus

 

 f (v) f (u)   f (z) f (w) 
|f (x) f (y)|



,
,
max 

|x y|
vu   zw
so

 

 f (v) f (u)   f (z) f (w) 




L = max 
,

vu   zw
is a Lipschitz constant for f on [a, b].

Corollary 7.16. Let I be any open interval, and let f : I R be a convex

function. Then f is continuous.
Exercise: Prove Corollary 7.16.
It turns out that at least according to our definition a convex function need
not be continuous on an endpoint of an interval on which it is defined. The following result analyzes endpoint behavior of convex functions.
Proposition 7.17. Let a R, and let I be an interval of the form (a, b), (a, b]
or (a, ). Let f : I R be convex.
a) limxa+ f (x) (, ].
b) Suppose that L = limxa+ f (x) < . For M R, define a function f : {a}I
R by f(a) = M , f(x) = f (x) for x I. Then f is convex iff M L and continuous
at a iff M = L.
Exercise: Prove Proposition 7.17.

3. CONVEX FUNCTIONS

141

3.6. Differentiable convex functions.

Theorem 7.18. For a differentiable function f : I R, TFAE:
(i) f is convex.
(ii) f 0 is weakly increasing.
Proof. For both directions of the proof it is convenient to consider fixed
a < b I and variable a < x < b.
(i) = (ii): For x (a, b], we define
s(x) =

f (x) f (a)
,
xa

S(x) =

f (b) f (x)
.
bx

and for x [a, b) we define

Since f is convex, the three-secant inequality for a < x < b holds:

s(x) s(b) = S(a) S(x).
Taking limits, we get
f 0 (a) = lim+ s(x) s(b) = S(a) lim S(x) = f 0 (b).
xa

xb

(a)
(ii) = (i): Let g(x) = f (x)f
. We will show that g is increasing on (a, b]: this
xa
gibes the two-secant inequality and thus the convexity of f . Since f is differentiable,
so is g and
(x a)f 0 (x) (f (x) f (a))
g 0 (x) =
.
(x a)2
(a)
= f 0 (y) for some a < y < x. Since f 0 is
By the Mean Value Theorem, f (x)f
xa
increasing,
f (x) f (a)
= f 0 (y) f 0 (x);
xa
equivalently
(x a)f 0 (x) (f (x) f (a)) 0.

Thus g 0 (x) 0 for all x (a, b], so indeed g is increasing on (a, b].

Corollary 7.19. A differentiable convex function is C 1 .

Proof. By Theorem 7.18, f 0 is weakly increasing so has only jump discontinuities. By Darbouxs Theorem, a derivative cannot have jump discontinuities. 
Corollary 7.20. (Kane3 Criterion) For twice differentiable f : I R, TFAE:
(i) f is convex.
(ii) f 00 0.
Proof. By Theorem 7.17, f is convex iff f 0 is weakly increasing. Moreover,
f is weakly increasing iff f 00 = (f 0 )0 0. Were done.

0

3Andrew Kane was a student in the 2011-2012 course who suggested this criterion upon being
prompted in class.

142

7. DIFFERENTIAL MISCELLANY

3.7. An extremal property of convex functions.

The following result secures the importance of convex functions in applied mathematics: for these functions, a critical point must be a global minimum.
Theorem 7.21. Let f : I R be a differentiable convex function. Let c I
be a stationary point: f 0 (c) = 0. Then:
a) f attains its global minimum at c.
b) Unless f is constant on some nontrivial subinterval of I, f attains a strict global
minimum at c: for all x 6= c, f (x) > f (c).
Proof. By Theorem 7.18, f 0 is weakly increasing.
a) Suppose d > c is such that f (d) < f (c). Then the secant line between
(c, f (c)) and (d, f (d)) has negative slope, so by the Mean Value Theorem there is
z (c, d) such that f 0 (z) < 0. But then c z and 0 = f 0 (c) > f 0 (z), contradicting
f 0 being weakly increasing. So f (c) f (x) for all x c.
Similarly, suppose b < c is such that f (b) < f (c). Then the secant line between
(b, f (b)) and (c, f (c)) has positive slope, so there is w (b, c) such that f 0 (w) > 0.
But then b c and f 0 (b) > 0 = f 0 (c), contradicting f 0 being weakly incereasing.
So f (x) f (c) for all x c. That is, f attains a global minimum at c.
b) If e 6= c is such that f 0 (e) = 0, then since f 0 is weakly increasing, f 0 0 on
the subinterval |[c, e]| from c to e, and thus f is constant on |[c, e]|.

Exercise: a) Let f : (a, b) R be a convex function such that limxa+ f (x) =
limxb f (x) = . Show that f assumes a global minimum.
b) State and prove an analogue of part a) with (a, b) replaced by R = (, ).
c) Exhibit a convex function f : R R which does not assume a global minimum.
3.8. Supporting lines and differentiability.
Let f : I R be a function. A supporting line for f at c I is a linear
function ` : R R such that f (c) = `(c) and f (x) `(x) for all x I.
Example: Consider the function f : R R given by f (x) = x2 . Observe that
the horizontal line ` = 0 is a supporting line at c = 0: indeed f (0) = 0 and
f (x) 0 for all x. Notice that ` = 0 is the tangent line to y = f (x) at c = 0.
Example: More generally, let A, B, C R with A 6= 0. We claim that for all c R
the tangent line `c (x) to the parabola f (x) = Ax2 + Bx + C is the unique line passing through (c, f (c)) such that f (x) 6= `c (x) for all x 6= c. To see this, consider the
function g(x) = f (x) `c (x). Then g is a quadratic polynomial with leading coefficient A and g(c) = g 0 (c) = 0, so g(c) = A(x c)2 , and thus f (x) `c (x) = g(x) 6= 0
for x 6= c. On the other hand, let ` be any other line passing through (c, f (c)).
Then h(x) = f (x) `(x) is a degree two polynomial with h(c) = 0. Moreover, since
` is not the tangent line at c, h0 (c) 6= 0, and thus h has a simple root at c, i.e.,
h(x) = A(x c)j(x) with j(x) a linear function with j(c) 6= 0. Therefore j has a
root at some d 6= c and that point d, `(x) = f (x).
We claim that the tangent line `c (x) to f (x) = Ax2 +Bx+C is a supporting line
iff A > 0. Indeed, on both the intervals (, c) and (c, ) the continuous function
f (x)`c (x) is nonzero, so must have constant sign. But limx f (x)`c (x) =
if A > 0 and if A < 0. It follows that if A > 0, f (x) `c (x) > 0 for all x 6= c

3. CONVEX FUNCTIONS

143

so `c is a supporting line and also that if A < 0, f (x) `c (x) < 0 for all x so
`c is not a supporting line. Note that since f 00 (x) = 2A, f is convex iff A > 0.
Example: Consider the function f : R R given by f (x) = |x|. Since Epi(f )
is a convex subset of R2 , f is convex. For every c > 0, the line y = x is a supporting line, and for every c < 0, the line y = x is a supporting line, and in both
cases these supporting lines are unique. For c = 0, y = 0 is a supporting line, but
it is not the only one: indeed y = mx is a supporting line at c = 0 iff 1 m 1.
Note that the smallest slope of a supporting line is the left-hand derivative at zero:
h
f (0 + h) f (0)
= lim
= 1,
h
h
h0
and the largest slope of a supporting line is the right-hand derivative at zero:
0
f
(0) = lim
h0

0
f+
(0) = lim

h0+

h
f (0 + h) f (0)
= lim
= 1.
+
h
h0 h

Lemma 7.22. Convex functions are closed under suprema. More precisely, if
{fi : I R}iI is a family of convex functions, f : I R is a function such that
for all x I, f (x) = supiI fi (x), then f is convex.
Proof. Let a < b I and (0, 1). Then
f ((1 )a + b) = sup fi ((1 )a + b)
iI

(1 ) sup f (a) + sup f (b) = (1 )f (a) + f (b).

ii

iI


Theorem 7.23. Let I be an open interval. For a function f : I R, TFAE:
(i) f is convex.
(ii) f admits a supporting line at each c I.
Proof. (i) = (ii): Neither property (i) or (ii) is disturbed by translating
the coordinate axes, so we may assume that c = 0 and f (0) = 0. Let I \ {0}.
For all 1 , 2 > 0 such that 1 , 2 I, by the secant-graph inequality we have


1
2
0 = (1 + 2 )f
(2 ) +
(1 ) 1 f (2 ) + 2 f (1 ),
1 + 2
1 + 2
or
f (2 )
f (1 )

.
2
1
It follows that sup2

f (2 )
2

inf 1

f (1 )
1 ,

so there is m R with

f (2 )
f (1 )
m
.
2
1
Equivalently, f (t) mt for all t R such that t I. Thus `(x) = mx is a
supporting line for f at c = 0.
(ii) = (i): For each c I, let `c : I R be a supporting line for f at c. Since
for all x I, f (x) `c (x) for all c and f (c) = `c (c), we have f (x) = supcI `c (x).
Since the linear functions `c are certainly convex, f is the supremum of a family of
convex functions, hence convex by Lemma 7.21


144

7. DIFFERENTIAL MISCELLANY

Before stating the next result, we recall the notion of one-sided differentiability:
if f is a function defined (at least) on some interval [c , c], we say f is left
(c)
differentiable at c if limxc f (x)f
exists, and if so, we denote this limit by
xc
0
f (c), the left derivative of f at c. Similarly, if f is defined (at least) on sme
(c)
interval [c, c + ), we say f is right differentiable at c if limxc+ f (x)f
exists,
xc
0
and if so, we denote this limit by f+ (c), the right derivative of f at c. (As usual
0
0
for limits, f is differentiable at c iff f
(c) and f+
(c) both exist and are equal.)
Theorem 7.24. Let I be an interval, c I , and f : I R be convex.
0
0
a) f is both left-differentiable and right-differentiable at c: f
(c) and f+
(c) exist.
0
0
b) For all c (a, b), f (c) f+ (c).
0
0
c) f
, f+
: I R are both weakly increasing functions.
d) A line ` passing through (c, f (c)) is a supporting line for f iff its slope m satisfies
0
0
f
(c) m f+
(c).

Proof. We follow [Go, p. 153] and [Gr, pp. 8-10].

a), b) Define : (a, b) \ {c} R by
(x) =

f (x) f (c)
.
xc

Further, put
A = ((a, c)), B = (c, b).
From the three-secant inequality we immediately deduce all of the following: is
weakly increasing on (a, c), is weakly increasing on (c, b), and A B. Thus
0
0
f
(c) = lim (x) = sup A inf B = lim+ (x) = f+
(c).
xc

xc

c) Let x1 , x2 (a, b) with x1 < x2 , and choose v with x1 < v < x2 . Then by part
b) and the three-secant inequality,
f (v) f (x1 )
f (v) f (x2 )
0
0

f
(x2 ) f+
(x2 ).
v x1
v x2
d) The proof of parts a) and b) shows that for x I,
0
0
f
(x1 ) f+
(x1 )

0
f (x) f (c) + f
(c)(x c), if x c,
0
f (x) f (c) + f+
(c)(x c), if x c.
0
0
Thus if f (c) m f+ (c), we have
0
f (x) f (c) + f
(c)(x c) f (c) + m(x c), if x c,
0
f (x) f (c) + f+
(c)(x c) m(xc ) if x c,
so `(x) = f (c) + m(x c) is a supporting line for f at c. That these are the only
possible slopes of supporting lines for f at c is left as an exercise for the reader. 

Exercise: We supppose the hypotheses of Theorem 7.24, and for m R, put

`(x) = f (c) + m(x c).
0
a) Suppose that m < f
(c). Show that there is > 0 such that for all x (c , c),
`(x) > f (x) and thus ` is not a supporting line for f at c.
0
b) Suppose that m > f+
(c). Show that there is > 0 such that for all x (c, c + ),
`(x) > f (x) and thus ` is not a supporting line for f at c.
Exercise: Assume the hypotheses of Theorem 7.24. Show TFAE:

3. CONVEX FUNCTIONS

145

0
(i) f
is continuous at c.
0
(ii) f+
is continuous at c.
(iii) f is differentiable at c, and the tangent line is a supporting line at c.
(iv) f has a unique supporting line at c.

Remark: Because a weakly increasing function can only have jump discontinuities, it can be shown that such functions are continuous at most points of their
domain. For those who are familiar with the notions of countable and uncountable
sets, we may be more precise: the set of discontinuities of a monotone function
must be countable. Since the union of two countable sets is countable, it follows
that a convex function is differentiable except possibly at a countable set of points.
Remark: One can go further: a convex function is twice differentiable at most
points of its domain. The sense of most here is different (and weaker): the set
of points at which f fails to be twice differentiable has measure zero in the sense
of Lebesgue. This result, which is itself due to Lebesgue, lies considerably deeper
than the one of the previous remark.
3.9. Jensens Inequality.
Theorem 7.25. (Jensens Inequality) Let f : I R be continuous and convex.
For any x1 , . . . , xn I and any 1 , . . . , n [0, 1] with 1 + . . . + n = 1, we have
f (1 x1 + . . . + n xn ) 1 f (x1 ) + . . . + n f (xn ).
Proof. We go by induction on n, the base case n = 1 being trivial. So
suppose Jensens Inequality holds for some n Z+ , and consider x1 , . . . , xn+1 I
and 1 , . . . , n+1 [0, 1] with 1 + . . . + n+1 = 1. If n+1 = 0 we are reduced
to the case of n variables which holds by induction. Similarly if n+1 = 1 then
1 = . . . = n = 0 and we have, trivially, equality. So we may assume n+1 (0, 1)
and thus also that 1 n+1 (0, 1). Now for the big trick: we write


1
n
1 x1 +. . .+n+1 xn+1 = (1n+1 )
x1 + . . . +
xn +n+1 xn+1 ,
1 n+1
1 n+1
so that
1
n
x1 +. . .+
xn )+n+1 xn+1 )
1 n+1
1 n+1


1
n
(1 n+1 )f
x1 + . . . +
xn + n+1 f (xn+1 ).
1 n+1
1 n+1

f (1 x1 + . . . + n xn ) = f ((1n+1 )(

1
n
Since 1
, . . . , 1
are non-negative numbers that sum to 1, by induction
n+1
n+1
the n variable case of Jensens Inequality can be applied to give that the above
expression is less than or equal to


1
n
(1 n+1 )
f (x1 ) + . . . +
f (xn ) + n+1 f (xn+1 )
1 n+1
1 n+1

= 1 f (x1 ) + . . . + n f (xn ) + n+1 f (xn+1 ).



146

7. DIFFERENTIAL MISCELLANY

3.10. Some applications of Jensens Inequality.

Example: For any p > 1, the function f : [0, ) R by x 7 xp is twice differentiable on (0, ) with f 00 (x) > 0 there, so by the Kane Criterion f is convex
on (0, ). Also f is continuous at 0, so f is convex on [0, ).
Example: The function f : R R given by x 7 ex has f 00 (x) = ex > 0 for
all x, so by the Kane Criterion f is convex on R.
By plugging these convex functions into Jensens Inequality and massaging what
we get a bit, we will quickly deduce some very important inequalities.4
Theorem 7.26. (Weighted Arithmetic Geometric Mean Inequality) Let x1 , . . . , xn
[0, ) and 1 , . . . , n [0, 1] be such that 1 + . . . + n = 1. Then:
x1 1 xnn 1 x1 + . . . + n xn .

(32)

1
n,

we get the arithmetic geometric mean inequality:

1
x1 + . . . + xn
(x1 xn ) n
.
n
Proof. We may assume x1 , . . . , xn > 0. For 1 i n, put yi = log xi . Then

Taking 1 = . . . = n =

1
n
x
n )

x1 1 xnn = elog(x1

= e1 y1 +...+n yn 1 ey1 +. . .+n eyn = 1 x1 +. . .+n xn .



Theorem 7.27. (Youngs Inequality)

Let x, y [0, ) and let p, q (1, ) satisfy

1
p

1
q

= 1. Then

yq
xp
+ .
p
q
Proof. When either x = 0 or y = 0 the left hand side is zero and the right hand
side is non-negative, so the inequality holds and we may thus assume x, y > 0. Now
apply the Weighted Arithmetic-Geometric Mean Inequality with n = 2, x1 = xp ,
x2 = y q , 1 = p1 , 2 = 1q . We get
xy

(33)

xy = (xp ) p (y q ) q = x1 1 x2 2 1 x1 + 2 x2 =

xp
yq
+ .
p
q


Theorem 7.28. (H
olders Inequality)
Let x1 , . . . , xn , y1 , . . . , yn R and let p, q (1, ) satisfy
(34)

1
p

1
p

1
q

= 1. Then
1

|x1 y1 | + . . . + |xn yn | (|x1 |p + . . . + |xn |p ) (|y1 |q + . . . + |yn |q ) q .

Proof. As above, the result is clear if either x1 = . . . = xn = 0 or y1 = . . . =

yn = 0, so we may assume that neither of these is the case. For 1 i n, apply
Youngs Inequality with
|yi |
|xi |
x=
1 ,y =
1 ,
p
p
q
p
(|x1 | + . . . + |xn | )
(|y1 | + . . . + |yn |q ) q
4Unfortunately we will not see here why these inequalities are important: youll have to trust
me on that for now. But just as soon as you take graduate level real analysis, youll see these
inequalities again.

3. CONVEX FUNCTIONS

147

and sum the resulting inequalities from i = 1 to n, getting

Pn
1 1
i=1 |xi yi |
+ = 1.
1
1
p
q
p
p
q
q
p
q
(|x1 | + . . . + |xn | ) (|y1 | + . . . + |yn | )

Theorem 7.29. (Minkowskis Inequality)
For x1 , . . . , xn , y1 , . . . , yn R and p 1:
1

(35) (|x1 + y1 |p + . . . + |xn + yn |p ) p (|x1 |p + . . . + |xn |p ) p + (|y1 |p + . . . |yn |p ) p

Proof. When p = 1, the inequality reads
|x1 + y1 | + . . . + |xn + yn | |x1 | + |y1 | + . . . + |xn | + |yn |
and this holds just by applying the triangle inequality: for all 1 i n, |xi + yi |
|xi | + |yi . So we may assume p > 1. Let q be such that p1 + 1q = 1, and note that
then (p 1)q = p. We have
|x1 + y1 |p + . . . + |xn + yn |p
HI

|x1 ||x1 +y1 |p1 +. . .+|xn ||xn +yn |p1 +|y1 ||x1 +y1 |p1 +. . .+|yn ||xn +yn |p1
1

(|x1 |p +. . .+|xn |p ) p (|x1 +y1 |p +. . .+|xn +yn |p ) q +(|y1 |p +. . .+|yn |p ) p (|x1 +y1 |p +. . .+|xn +yn |p ) q


1
1
1
= (|x1 |p + . . . + |xn |p ) p + (|y1 |p + . . . |yn |p ) p (|x1 + y1 |p + . . . + |xn + yn |p ) q .
1

Dividing both sides by (|x1 + y1 |p + . . . + |xn + yn |p ) q and using 1 1q = p1 , we get

the desired result.

Corollary 7.30. (Triangle Inequality in Rn ) For all x1 , . . . , xn , y1 , . . . , yn R
q
q
p
(x1 + y1 )2 + . . . + (xn + yn )2 x21 + . . . + x2n + y12 + . . . + yn2 .

CHAPTER 8

Integration
1. The Fundamental Theorem of Calculus
Having finished with continuity and differentiation, we turn to the third main
theme of calculus: integration. The basic idea is this: for a function f : [a, b] R,
Rb
we wish to associate a number a f , the definite integral. When f is non-negative,
Rb
our intuition is that a f should represent the area under the curve y = f (x), or
more precisely the area of the region bounded above by y = f (x), bounded below
by y = 0, bounded on the left by x = a and bounded on the right by x = b.
Unfortunately this is not yet a formal definition, because we do not have a
formal definition of the area of a subset of the plane! In high school geometry
one learns only about areas of very simple figures: polygons, circles and so forth.
Dealing head-on with the task of assigning an area to every subset of R2 is quite
difficult: it is one of the important topics of graduate level real analysis: measure
theory.
Rb
So we need to back up a bit and give a definition of a f . As you probably know,
Rb
the general idea is to construe a f as the result of some kind of limiting process,
wherein we divide [a, b] into subintervals and take the sum of the areas of certain rectangles which approximate the function f at various points of the interval
(Riemann sums). As usual in freshman calculus, reasonably careful definitions
appear in the textbook somewhere, but with so little context and development that
(almost) no actual freshman calculus student can really appreciate them.
But wait! Before plunging into the details of this limiting process, lets take a
Rb
more axiomatic approach: given that we want a f to represent the area under
y = f (x), what properties should it satisfy? Here are some reasonable ones.
Rb

C = C(b a).
Ra b
Rb
(I2) If f1 (x) f2 (x) for all x [a, b], then a f1 a f2 .
Rc
Rb
Rc
(I3) If a c b, then a f = a f + b f .
(I1) If f = C is a constant function, then

Exercise 1.1: Show (I1) implies: for any f : [a, b] R and any c [a, b],

Rc
c

f = 0.

It turns out that these three axioms already imply many of the other properties we
want an integral to have. Even more, there is essentially only one way to define
Rb
f so as to satisfy (I1) through (I3).
a
Well, almost. One feature that we havent explicitly addressed yet is this: for
149

150

8. INTEGRATION

Rb
which functions f : [a, b] R do we expect a f to be defined? For all functions??
A little thought shows this not to be plausible: there are some functions so pathological that there is no reason to believe that the area under the curve y = f (x)
has any meaning whatsoever, and there are some functions for which this area concept seems meaningful but for which the area is infinite.
So it turns out to be useful to think of integration itself as a real-valued function,
with domain some set of functions {f : [a, b] R}. That is, for each a b we
wish to have a set, say R[a, b], of integrable functions f : [a, b] R and for each
Rb
f R[a, b], we wish to associate a real number a f . As to exactly what this set
R[a, b] of integrable functions should be, it turns out that we have some leeway, but
to get a theory which is useful and not too complicated, lets assume the following:
(I0) For all real numbers a < b:
a) Every continuous f : [a, b] R lies in R[a, b].
b) Every function f R[a, b] is bounded.
By the Extreme Value Theorem, every continuous function f : [a, b] R is
bounded. Thus the class C[a, b] of all continuous functions f : [a, b] R is contained in the class B[a, b] of all bounded functions f : [a, b], and axiom (I0) requires
that the set of integrable functions lies somewhere in between:
C[a, b] R[a, b] B[a, b].
Lets recast the other three axioms in terms of our set R[a, b] of integrable functions:
(I1) If f = C is constant, then f R[a, b] and

Rb
a

C = C(b a).

Rb
Rb
(I2) If for f1 , f2 R[a, b] we have f1 (x) f2 (x) for all x [a, b], then a f1 a f2 .
(I3) Let f : [a, b] R, and let c (a, b). Then f R[a, b] iff f R[a, c] and
Rb
Rc
Rb
f R[c, b]. If these equivalent conditions hold, then a f = a f + c f .
If this business of integrable functions seems abstruse, then on the first pass just
imagine that R[a, b] is precisely the set of all continuous functions f : [a, b] R.
Now we have the following extremely important result.
Theorem 8.1. (Fundamental Theorem of Calculus)
Let f R[a, b] be any
Rx
integrable function. For x [a, b], define F(x) = a f . Then:
a) The function F : [a, b] R is continuous at every c [a, b].
b) If f is continuous at c [a, b], then F is differentiable at c, and F 0 (c) = f (c).
c) If f is continuous and F is any antiderivative of f i.e., a function F : [a, b] R
Rb
such that F 0 (x) = f (x) for all x [a, b], then a f = F (b) F (a).
Proof. By (I0), there exists M R such that |f (x)| M for all x [a, b]. If
M = 0 then f is the constant function 0, and then it follows from (I1) that F is
also the constant function zero, and one sees easily that the theorem holds in this
case.

. Indeed, by (I3)
So we may assume M > 0. For all  > 0, we may take = M
Z x
Z c
Z x
(36)
F(x) F(c) =
f
f=
f.
a

1. THE FUNDAMENTAL THEOREM OF CALCULUS

151

Moreover, let a A B b. Then still M f (x) M for all x [A, B], so by

(I2) and then (I2) we have
Z B
Z B
f M (B A),
(M )
M (B A) =
A

and thus
Z

(37)

f | M (B A).
A

Now suppose |x c| < =


M.

Using (36) and then (37) with A = c, B = x, we get

Z x
  
|F(x) F(c)| = |
f | M |x c| < M
= .
M
c
b) Suppose f is continuous at c. We wish to compute
F(x) F(c)
.
xc
Since f is continuous at c, for all  > 0, there exists > 0 such that |x c| < =
|f (x) f (c)| < , or equivalently
F 0 (x) = lim

xc

f (c)  < f (x) < f (c) + .

Therefore
Rx
Rx
f
f (c) + 
f (c) 
c

c
= f (c) + ,
f (c)  =
xc
xc
xc
and thus
Rx
f
| c
f (c)| .
xc
This shows that F 0 (c) exists and is equalR to f (c).
x
c) By part b), if f is continuous, F(x) = a f is an antiderivative of f . But we have
shown that if antiderivatives exist at all they are unique up to an additive constant.
We have just found an antiderivative F, so if F is any other antiderivative of f we
must have F (x) = F(x) + C for some constant C, and then
Z b
Z a
Z b
F (b) F (a) = (F(b) + C) (F(a) + C) = F(b) F(a) =
f
f=
f.
Rx
c


Remark: Although we introduced the integral axiomatically, as long as we are
only trying to integrate continuous functions we had no choice: the only way to
Rb
assign a value a f to each continuous function f : [a, b] R satisfying the (reaRb
sonable!) axioms (I1) through (I3) is to take a f to be an antidervative F of f
with F (a) = 0, and again, there is at most one such function.
These same considerations answer the conundrum of why the celebrated Theorem 8.1 has such a short and simple proof.1 The theorem assumes that we already
1This is not just florid language. I taught second semester calculus four times as a graduate
student and really did become puzzled at how easy it was to prove the Fundamental Theorem
of Calculus so soon after integration is discussed. I worked out the answer while teaching an
undergraduate real analysis course at McGill University in 2005. The current presentation is an
adaptation of my lecture notes from this older course. Soon after I gave my 2005 lectures I found
that a very similar axiomatic treatment of the integral was given by the eminent mathematician

152

8. INTEGRATION

Rb
have an integral, i.e., an assignment (f : [a, b] R) 7 a f for every continuous
function f . We have shown that there is at most one such integral on the continuous functions, but we have not yet constructed this integral! In other words,
we have settled the problem of uniqueness of the definite integral but (thus far)
assumed a solution to the much harder problem of existence of the definite integral.
And again, this existence problem is equivalent to an existence problem that we
mentioned before, namely that every continuous function has an antiderivative.
Thus: if we could prove by some other means that every continuous function f
is the derivative of some other function F , then by the above we may simply define
Rb
f = F (b) F (a). This is the approach that Newton himself took, although
a
he didnt prove that every continuous function was a derivative but rather merely
assumed it. It is also what freshman calculus students seem to think is taught in
Rb
freshman calculus, namely that the definition of a f is F (b) F (a).2
But I do not know any way to prove that an arbitrary continuous function has
Rb
an antiderivative except to give a constructive definition of a f as a limit of sums
Rx
and then appeal to Theorem 8.1b) to get that a f is an antiderivative of f .
Thus Theorem 8.1 is easy because it diverts the hard work elsewhere: we need to
give a constructive definition of the definite integral via a (new) kind of limiting process and then show from scratch that applied to every continuous f : [a, b] R
Rb
this limiting process converges and results in a well-defined number a f .
2. Building the Definite Integral
2.1. Upper and Lower Sums.
Now we begin the proof of the hard fact lurking underneath the Fundamental Theorem of Calculus: that we may define for every continuous function f : [a, b] R
Rb
a number a f so as to satisfy (I1) through (I3) above. For now, we will make
a simplifying assumption on our class of integrable functions: namely, let us only
consider functions f : [a, b] R such that for every closed subinterval [c, d] [a, b],
f : [c, d] R has a maximum and minimum value. Of course this holds for all
continuous functions, so it will be a good start.
The basic idea is familiar from freshman calculus: we wish to subdivide our interval [a, b] into a bunch of closed subintervals meeting only at the endpoints, and
then we want to consider the lower sum and upper sum associated to f on each
subinterval. Then the lower sum should be less than or equal to the true area
under the curve which should be less than or equal to the upper sum, and by
dividing [a, b] into more and smaller subintervals we should get better and better
Rb
approximations to the true area under the curve, so we should define a f via
some limiting process involving lower sums and upper sums.
Okay, lets do it!
Serge Lang in [L]. So the presentation that I give here is not being given by me for the first time
and was not originated by me...but nevertheless the material is rarely presented this way.
2This is not what the books actually say, but what they actually say they dont say loudly
enough in order for the point to really stick.

2. BUILDING THE DEFINITE INTEGRAL

153

Step 1: We need the notion of a partition of an interval [a, b]: we choose finitely
many same points in [a, b] and use them to divide [a, b] into subintervals. Formally, a partitition P is given by a positive integer n and real numbers
a = a0 a1 . . . an1 an = b.
That is, we require the first sample point a0 to be the left endpoint of the interval,
the last sample point an to be the right endpoint of the interval, and the other
(distinct) points are absolutely arbitrary but written in increasing order.
Let f : [a, b] R be any function admitting a minimum and maximum value on
every closed subinterval of [a, b] (e.g. any continuous function!). For 0 i n 1,
let mi (f ) denote the minimum value of f on the subinterval [xi , xi+1 ] and let Mi (f )
denote the maximum value of f on the subinterval [xi , xi+1 ]. Then we define the
lower sum associated to f : [a, b] R and the partition P = {x0 , . . . , xn } as
L(f, P) =

n1
X

mi (f )(xi+1 xi )

i=0

and also the upper sum associated to f and P as

U (f, P) =

n1
X

Mi (f )(xi+1 xi ).

i=0

These sums have a simple and important geometric interpretation: for any 0 i
n 1, the quantity xi+1 xi is simply the length of the subinterval [xi , xi+1 ]. So
consider the constant function mi (f ) on the interval [xi , xi+1 ]: by definition of mi ,
this is the largest constant function whose graph lies on or below the graph of f
at every point of [xi , xi+1 ]. Therefore the quantity mi (f )(xi+1 xi ) is simply the
area of the rectangle with height mi (f ) and width xi+1 xi , or equivalently the
area under the constant function y = mi (f ) on [xi , xi+1 ].
We say the function f : [a, b] R is integrable if there is a unique I R
such that for every partition P of [a, b] we have
L(f, P) I U (f, P).
This definition, although correct, is not ideally formulated: it underplays the most
important part the uniqueness of I while making it annoying to show the existence of I. (It turn outs that there is always at least one I lying between every lower
sum and every upper sum, but this is as yet far from clear.) Here are some examples.
Example 2.1: If f (x) C is a constant function, then for every partition P on
[a, b] we have L(f, P) = U (f, P) = C(b a). Thus the unique I in question is
celarly C(b a): constant functions are integrable.
Example 2.2: Suppose f (x) is constantly equal to 1 on the interval [a, b] except
for one interior point c, at which f (c) = 0. We claim that despite having a disRb
continuity at c, f is integrable, with a f = b a. To see this, first observe that
for any partition P of [a, b] we have U (f, P) = b a. Indeed this is because on
every subinterval of [a, b] f has 1 as its maximum value. On the other hand, for

154

8. INTEGRATION

any sufficiently small  > 0, we may choose a partition in which c occurs in exactly
one subinterval (i.e., c is not one of the points of the partition). Then the lower
sum on that subinterval is 0, whereas on every other subinterval the minimum is
again 1, so L(f, P) = (b a)(1 ). This shows that the unique number between
Rb
every L(f, P) and U (f, P) is b a, so a f = (b a).
Exercise 2.3: Show that starting with the constant function C on [a, b] and changing
Rb
its value at finitely many points yields an integrable function f with a f = C(ba).
The previous examples have the property that the upper sums U (f, P) are constant. When this happens, one can show f is integrable by finding a sequence of
partitions for which the lower sums approach this common value U (f, P) which
must then be the integral. But constancy of upper sums only occurs in trivial examples. For instance, suppose we want to show that f (x) = x is integrable on [0, 1].
If we partition [0, 1] into n equally spaced subintervals let us call this partition
Pn then since f is increasing its minimum on each subinterval occurs at the left
endpoint and its maximum on each subinterval occurs at the right endpoint. Thus


n1
n1
Xi 1
1 X
1
(n 1)n
1
i= 2
.
= 2
=1
L(f, Pn ) =
n
n
n
n
2
2n
i=0
i=0
and
U (f, Pn ) =

n1
X
i=0

i+1
n



n
1 X
1 n(n + 1)
1
1
i= 2
.
= 2
=1+
n
n i=1
n
2
2n

1
Since
= limx x+1
2x = 2 , the upper and lower sums can both be
1
made arbitrarily close to 2 by taking n to be sufficiently large. Thus if f (x) = x
is integrable on [0, 1], its integral must be 12 . Unfortunately we have not yet shown
that f is integrable according to our definition: to do this we would have to consider
an arbitrary partition P of [0, 1] and show that L(f, P) 12 U (f, P). For this
very simple function f (x) = x it is possible to grind this out directly, but its quite
a bit of work. And thats just to find the area of a right triangle!

limx x1
2x

Example 2.4: Consider the function f : [a, b] R which is 1 at every irrational

point and 0 at every rational point. Because every subinterval contains both rational and irrational numbers, for every partition P of [a, b] we have L(f, P) = 0 and
U (f, P) = 1(ba) = ba. Assuming of course that a < b, this shows that f is not
integrable: rather than the lower sums and the upper sums becoming arbitrarily
close together, there is an unbridgeable gap between them.
The previous example perhaps suggests a solution to the problem. It did so to
the late 19th century mathematician Jean-Gaston Darboux, who came up with an
elegant way to see whether there is an unbridgeable gap between the lower and
upper sums. Darbouxs definition crucially and cleverly relies on the existence of
suprema and infima for subsets of R. In the next section we do things Darbouxs
way, which as we will see is much more pleasant than our current setup.
2.2. Darboux Integrability.

2. BUILDING THE DEFINITE INTEGRAL

155

Let f : [a, b] R be any function. For a partition P = {a = x0 < x1 < . . . <

xn1 < xn = b} and 0 i n 1, let mi (f ) be the infimum of f on [xi , xi+1 ] and
Mi (f ) be the supreumum of f on [xi , xi+1 ]. Thus we have mi (f ) [, ) and
Mi (f ) (, ]. As above, we define the lower and upper sums associated to P:
L(f, P) =

n1
X

mi (f )(xi+1 xi ) [, ),

i=0

U (f, P) =

n1
X

Mi (f )(xi+1 xi ) (, ].

i=0

For any f and P we have

(38)

L(f, P) U (f, P).

Observe though that the lower sum could take the value and the upper sum
could take the value . The following result clarifies when this is the case.
Proposition 8.2. Let f : [a, b] R be any function.
a) The following are equivalent:
(i) For all partitions P of [a, b], L(f, P) = .
(ii) There exists a partition P of [a, b] such that L(f, P) = .
(iii) f is not bounded below on [a, b].
b) The following are equivalent:
(i) For all partitions P of [a, b], U (f, P) = .
(ii) There exists a partition P of [a, b] such that U (f, P) = .
(iii) f is not bounded above on [a, b].
c) The following are equivalent:
(i) For all partitions P of [a, b], L(f, P) > and U (f, P) < .
(ii) f is bounded on [a, b].
Proof. a) (i) = (ii) is immediate.
(ii) = (iii): We prove the contrapositive: suppose that there is m R such that
m f (x) for all x [a, b]. Then for all partitions P = {a = x0 < . . . < xn1 <
xn = b} and all 0 i n 1, we have mi (f ) m > , so L(f, P) > .
(iii) = (i): Suppose f is not bounded below on [a, b], and let P = {a = x0 < . . . <
xn1 < xn = b} be a partition of [a, b]. If mi (f ) > for all 0 i n 1, then
minn1
i=0 mi (f ) is a finite lower bound for f on [a, b], contradicting our assumption.
So there is at least one i such that mi (f ) = , which forces L(f, P) = .
b) This is similar enough to part a) to be left to the reader.
c) If for all partitions P, L(f, P) > and U (f, P) < , then by parts a) and b)
f is bounded above and below on [a, b], so is bounded on [a, b]. Conversely, if f is
bounded on [a, b] then it is bounded above and below on [a, b], so by parts a) and
b), for all partitions P we have L(f, P) > and U (f, P) < .

Let P1 and P2 be two partitions of [a, b]. We say that P2 refines P1 if P2 contains
every point of P1 : i.e., if P1 P2 .
Lemma 8.3. (Refinement Lemma) Let P1 P2 be partitions of [a, b] (i.e., P2
refines P1 ). Then for any bounded function f : [a, b] R we have
L(f, P1 ) L(f, P2 ) U (f, P2 ) U (f, P1 ).

156

8. INTEGRATION

Proof. If P2 refines P1 , P2 is obtained from P1 by a finite number of instances

of the following: choose a subinterval [c, d] of P1 and insert a new point e to
subdivide it into two subintervals. So it is enough to address the case of a single
interval. Let m and M be the infiumum and supremum of f on [c, d]. Futher, let
m1 and M1 be the infimum and supremum of f on [c, e], and m2 and M2 be the
infimum and supremum of f on [e, d]. Then m = inf m1 , m2 and M = sup M1 , M2 ,
so
m m1 , m m2 , M1 M, M2 M.
Then
m(d c) = m(e c) + m(d e) m1 (e c) + m2 (e c)
and
M (d c) = M (e c) + M (d e) M1 (e c) + M2 (e c).
This shows that subdivision causes the lower sum to stay the same or increase and
the upper sum to stay the same or decrease. Since this holds each time we add a
point to get from P1 to P2 , we get L(f, P1 ) L(f, P2 ) U (f, P2 ) L(f, P2 ). 
Lemma 8.4. Let f : [a, b] R be a function, and let P1 , P2 be partitions of
[a, b]. Then L(f, P1 ) U (f, P2 ). That is, any lower sum associated to any partition
is less than or equal to the upper sum associated to any other partition.
Proof. The idea here is simple but important: we choose a common refinement of P1 and P2 , i.e., a partition which refines (contains) both P1 and P2 . Any
two partitions have infinitely many common refinements, but the most economical
choice is simply the union of the two: put P = P1 P2 . Then by Lemma 8.3 we
have
L(f, P1 ) L(f, P) U (f, P) U (f, P2 ).

Now we come to the crux of Darbouxs theory of integrability: for every function
f : [a, b] R we define the lower integral intba f as the supremum of L(f, P) as
Rb
P ranges over all partitions of [a, b] and the upper integral a f as the infimum of
U (f, P) as P ranges over all partitions of [a, b]. Finally, we say that f is Darboux
Rb
Rb
Rb
integrable if a f = a f R, and we denote this common value by a f .
Lemma 8.5. For any function f : [a, b] R, we have
Z

f
a

f.
a

Proof. Recall that if X, Y R are such that x y for all x X and all
y Y , then sup X inf Y . Now, by Lemma 8.4, for any partitions P1 and P2 we
have L(f, P1 ) U (f, P2 ). Therefore
Z

Z
f = sup L(f, P1 ) inf U (f, P2 ) =

P1

P2

f.
a

2. BUILDING THE DEFINITE INTEGRAL

157

Proposition 8.6. Let f : [a, b] R be any function.

Rb
a) We have a f = iff f is unbounded below.
Rb
b) We have a f = iff f is unbounded above.
Rb Rb
c) Therefore a f, a f R iff f is bounded. In particular, if f is Darboux integrable,
then it is bounded.
Rb
Proof. a) If f is unbounded below,
f = supP L(f, P) = supP = .
a
If f is bounded below, then for all P, L(f, P) R and thus supP L(f, P) > .
b) This is very similar to part a) and is left to the reader.
c) This follows immediately from parts a) and b).

In particular the class of Darboux integrable functions satisfies axiom (IOb).
Exercise 2.5: Is there a function f : [a, b] R with

Rb
a

f=

Rb
a

f = ?

Finally, here is the result we really want.

Theorem 8.7. (Darbouxs Integrability Criterion)
For a bounded function f : [a, b] R, the following are equivalent:
(i) f is Darboux integrable.
(ii) For all  > 0 there exists a partition P of [a, b] such that U (f, P) L(f, P) < .
(iii) There is exactly one real number I such that for all partitions P of [a, b], we
have L(f, P) I U (f, P).
Rb
Rb
Proof. (i) = (ii): Fix  > 0. Since a f = f = supP L(f, P), there exists
a
Rb
Rb
Rb
P1 with L(f, P1 ) > a 2 . Similarly, since a f = a f = inf P U (f, P), there exists
Rb
P2 with U (f, P2 ) < a + 2 . Let P = P1 P2 . Since P is a common refinement of
P1 and P2 we have
Z b

(39)
U (f, P) U (f, P2 ) <
f+
2
a
and also
b

Z
L(f, P) L(f, P1 ) >
a


f ,
2

and thus

L(f, P) <
2

(40)

f.
a

Adding (39) and (40) gives

U (f, P) L(f, P) < .
(ii) = (i): By assumption, for every  > 0, there exists a partition P such that
Rb
Rb
U (f, P) L(f, P) < . But by definition, we have a f U (f, P) and
L(f, P)
a
Rb
and thus also f L(f, P). Adding these two inequalities gives
a

f
a

f U (f, P) L(f, P) < .

a

158

8. INTEGRATION

Rb
Rb
Since this holds for all  > 0, we have a f f . On the other hand, by Lemma
a
Rb
Rb
Rb
Rb
8.5 we have f a f , so f = a f R and thus f is Darboux integrable.
a
a
Rb Rb
Rb
(i) = (iii): Suppose f is Darboux integrable, so a = f = a f R. Then for
a
all partitions P we have
Z

L(f, P)

Rb

f U (f, P).
a

Rb

f . Then I is less than the supremum of the lower

Rb
Rb
sums, so there exists a partition P with I < L(f, P). Similarly, if I > a f = a f ,
then I is greater than the infimum of the upper sums, so there exists a partition
Rb
P with U (f, P) < I. This shows that a f is the unique real number which lies in
between every lower sum and every upper sum.
(iii) = (i): We prove the contrapositive. Suppose that f is not Darboux integrable. Then for any partition P we have
a

f=

Z
f

f=
a

Moreover, suppose I <

L(f, P)

Rb
a

f,

Rb
a

f U (f, P),

f<
a

and thus every I [

b
a

f ] lies between every upper sum and every lower sum.



2.3. Verification of the Axioms.

Let R([a, b]) denote the set of Darboux integrable functions on [a, b]. We now
tie together the work of the previous two sections by showing that the assignment
Rb
f R([a, b]) 7 a f satisfies the axioms (I0) through (I3) introduced in 1. In
particular, this shores up the foundations of the Fundamental Theorem of Calculus
and completes the proof that every continuous f : [a, b] R has an antiderivative.
In summary, we wish to prove the following result.
Theorem 8.8. (Main Theorem on Integration)
a) Every continuous function f : [a, b] R is Darboux integrable.
b) The operation which assigns to every Darboux integrable function f : [a, b] R
Rb
the number a f satisfies axioms (I0) through (I3) above.
c) Thus the Fundamental Theorem of Calculus holdsR for the Darboux integral. In
x
particular, for every continuous function f , F (x) = a f is an antiderivative of f .
Proof. a) Let f : [a, b] R be continuous. The key is that f is uniformly
continuous, so for all  > 0, there is > 0 such that for all x1 , x2 [a, b],

|x1 x2 | < = |f (x1 ) f (x2 )| < ba
. Let n Z+ be such that ba
n < , and
let Pn be the partition of [a, b] into n subintervals of equal length ba
.
Then
n
(41)

 
 n1
n1
X
ba
ba X
U (f, Pn ) L(f, Pn ) =
(Mi (f ) mi (f ))

Mi (f ) mi (f ).
n
n
i=0
i=0

2. BUILDING THE DEFINITE INTEGRAL

159

Now for all 0 i < n 1, mi (f ) = f (ci ) and Mi (f ) = f (di ) for some ci , di

[xi , xi+1 ]. Thus |ci di | xi+1 xi = ba
n < , so

(42)
|Mi (f ) mi (f )| = |f (di ) f (ci )| <
.
ba
Combining (41) and (42) gives
 n1
 n1


ba X
ba X 
= .
U (f, Pn ) L(f, Pn )
(Mi (f ) mi (f ))
n
n
ba
i=0
i=0
b) (I0): By part a), every continuous function f : [a, b] R is Darboux integrable.
By Proposition 8.6, every Darboux integrable function on [a, b] is bounded. (I1):
In Example 2.1, we showed that the constant function C is integrable on [a, b] with
Rb
C = C(b a). (I2): If f1 , f2 : [a, b] R are both Darboux integrable and such
a
that f1 (x) f2 (x) for all x [a, b], then for every partition P of [a, b] we have
L(f1 , P) L(f2 , P), and thus
Z b
Z b
f1 = sup L(f1 , P) sup L(f2 , P) =
f2 .
a

(I3): Let f : [a, b] R, and let c (a, b). Suppose first that f : [a, b] R is
Darboux integrable: thus, for all  > 0, there exists a partition P of [a, b] with
U (f, P) L(f, P) < . Let Pc = P {c}. By the Refinement Lemma,
L(f, P) L(f, Pc ) U (f, Pc ) U (f, P),
so U (f, Pc ) L(f, Pc ) U (f, P) L(f, P) < . Let P1 = Pc [a, c] and P2 =
Pc [c, b]. Then
L(f, Pc ) = L(f, P1 ) + L(f, P2 ), U (f, Pc ) = U (f, P1 ) + U (f, P2 ),
and therefore
(U (f, P1 )L(f, P1 ))+(U (f, P2 )L(f, P1 )) = (U (f, P1 )+U (f, P2 ))(L(f, P1 )+L(f, P2 ))
= U (f, Pc ) L(f, Pc ) < ,
so by Darbouxs criterion f : [a, c] R and f : [c, b] R are Darboux integrable.
Conversely, suppose f : [a, c] R and f : [c, b] R are Darboux integrable; let
 > 0. By Darbouxs criterion, there is a partition P1 of [a, c] such that

U (f, P1 ) L(f, P2 ) <
2
and a partition P2 of [c, b] such that

U (f, P2 ) L(f, P2 ) < .
2
Then P = P1 P2 is a partition of [a, b], and
U (f, P) L(f, P) = U (f, P1 ) + U (f, P2 ) (L(f, P1 ) + L(f, P2 ))


= (U (f, P1 ) L(f, P1 )) + (U (f, P2 ) L(f, P2 )) < + = .
2 2
As for the value of the integral: fix  > 0. Let P be any partition of [a, b], Pc =
P {c}, P1 = Pc [a, c], P2 = Pc [c, b]. Then
Z c
Z b
L(f, P) L(f, Pc ) = L(f, P1 ) + L(f, P2 )
f+
f U (f, P1 ) + U (f, P2 )
a

= U (f, Pc ) U (f, P).

160

8. INTEGRATION

Rc

Rb

f is a real number lying in between L(f, P) and U (f, P) for every

Rc
Rb
Rb
partition P of [a, b], so by Theorem 8.7 a f + c f = a f .
c) This is immediate from Theorem 8.1 (the Fundamental Theorem of Calculus!).


Thus

f+

2.4. An Inductive Proof of the Integrability of Continuous Functions.

In this section we will give a proof of the Darboux integrability of an arbitrary
continuous function f : [a, b] R which avoids the rather technical Uniform Continuity Theorem. We should say that we got the idea for doing this from Spivaks
text, which first proves the integrability using uniform continuity as we did above
and then later goes back to give a direct proof.
Theorem 8.9. Let f : [a, b] R be a continuous function on a closed bounded
interval. Then f is Darboux integrable.
Proof. By Darbouxs Criterion, it suffices to show that for all  > 0, there is
a partition P of [a, b] such that U (f, P) L(f, P) < . It is convenient to prove the
following slightly different (but logically equivalent!) statement: for every  > 0,
there exists a partion P of [a, b] such that U (f, P) L(f, P) < (b a).
Fix  > 0, and let S() be the set of x [a, b] such that there exists a partition
Px of [a, b] with U (f, Px ) L(f, Px ) < . We want to show b S(); our strategy
will be to show S() = [a, b] by Real Induction.
(RI1) The only partition of [a, a] is Pa = {a}, and for this partition we have
U (f, Pa ) = L(f, Pa ) = f (a) 0 = 0, so U (f, Pa ) L(f, Pa ) = 0 < .
(RI2) Suppose that for x [a, b) we have [a, x] S(). We must show that there
is > 0 such that [a, x + ] S(), and by the above observation it is enough
to find > 0 such that x + S(): we must find a partition Px+ of [a, x + ]
such that U (f, Px+ ) L(f, Px+ ) < (x + a)). Since x S(), there is a
partition Px of [a, x] with U (f, Px ) L(f, Px ) < (x a). Since f is continuous
at x, we can make the difference between the maximum value and the minimum
value of f as small as we want by taking a sufficiently small interval around x: i.e.,
there is > 0 such that max(f, [x, x + ]) min(f, [x, x + ]) < . Now take the
smallest partition of [x, x + ], namely P 0 = {x, x + }. Then U (f, P 0 ) L(f, P 0 ) =
(x+x)(max(f, [x, x+])min(f, [x, x+])) < . Thus if we put Px+ = Px +P 0
and use the fact that upper / lower sums add when split into subintervals, we have
U (f, Px+ ) L(f, Px+ ) = U (f, Px ) + U (f, P 0 ) L(f, Px ) L(f, P 0 )
= U (f, Px ) L(f, Px ) + U (f, P 0 ) L(f, P 0 ) < (x a) +  = (x + a).
(RI3) Suppose that for x (a, b] we have [a, x) S(). We must show that x S().
The argument for this is the same as for (RI2) except we use the interval [x , x]
instead of [x, x + ]. Indeed: since f is continuous at x, there exists > 0 such that
max(f, [x, x])min(f, [x, x]) < . Since x < x, x S() and thus there
exists a partition Px of [a, x] such that U (f, Px ) = L(f, Px ) = (x a).
Let P 0 = {x , x} and let Px = Px P 0 . Then
U (f, Px ) L(f, Px ) = U (f, Px ) + U (f, P 0 ) (L(f, Px ) + L(f, P 0 ))
= (U (f, Px ) L(f, Px )) + (max(f, [x , x]) min(f, [x , x]))
< (x a) +  = (x a).


3. FURTHER RESULTS ON INTEGRATION

161

Exercise 2.6: Show that if x S() and a y x, then also y S().

Spivak [S, pp. 292-293] gives a different uniform continuity-free proof of Theorem 8.9: he establishes equality of the upper and lower integrals by differentiation.
This sort of proof goes back at least to M.J. Norris [No52].
3. Further Results on Integration
3.1. The oscillation.
Let f : D R R, and let I be an interval contained in the domain D of f .
We define the oscillation of f on I as
(f, I) = sup(f, I) inf(f, I).
Note that (f, I) is in general an extended real number; it is an honest real number
iff f is bounded on I (which will almost always be the case for us).
If J I D, then inf(f, J) inf(f, I) and sup(f, J) sup(f, I), and thus
(43)

(f, J) (f, I).

Suppose now that c is a point in the interior of the domain D of f . We define the
oscillation of f at c to be
(f, c) = lim+ (f, [c , c + ]).
0

In other words, we are considering the oscillation of f on smaller and smaller

intervals centered around c and taking the limit as approaches zero. Because of
(43) the function 7 (f, [c , c + ]) is an increasing function of , so the limit as
approaches zero from the right exists as an element of [0, ] and is simply equal
to the infimum. Whats the point? This:
Proposition 8.10. Let I be an interval, f : I R be a function, and c an
interior point of I. The following are equivalent:
(i) (f, c) = 0.
(ii) f is continuous at c.
Proof. (i) = (ii): If (f, c) = 0, then for all  > 0, there exists > 0 such
that (f, [c , c + ]) = sup(f, [c , c + ]) inf(f, [c , c + ]) < , i.e., there
exists > 0 such that for all x with |x c| , |f (x) f (c)| < . So f is continuous
at c.
(ii) = (i): This is almost exactly the same. If f is continuous at c, then for all
 > 0 there exists > 0 such that for all x with |xc| , |f (x)f (c)| < , and then
sup(f, [c, c+]) f (c)+, inf(f, [c, c+]) f (c), so (f, [c, c+]) 2,
so (f, c) = lim0+ (f, [c , c + ]) = 0.

Remark: If f : I R and c is an endpoint of I, we can still define the oscillation
(f, c), just by taking suitable half-intervals: e.g. if c is the left endpoint we put
(f, c) = lim0+ (f, [c, c + ]). With this definition and our usual standard conventions about continuity at an endpoint of an interval, Proposition 8.10 remains
true even if c is an endpoint of the interval I.

162

8. INTEGRATION

Now let f : [a, b] R and let P = {a = x0 < x1 < . . . < xn1 < xn = b}
be a partition of [a, b]. We define3
(f, P) =

n1
X

(f, [xi , xi+1 ])(xi+1 xi ) = U (f, P) L(f, P).

i=0

Thus this notation is just a way of abbreviating the quantities upper sum minus
lower sum which will appear ubiquitously in the near future. We can restate Darbouxs Criterion especially cleanly with this new notation: a function f : [a, b] R
is integrable iff for all  > 0, there exists a partition P of [a, b] with (f, P) < .
3.2. Discontinuities of Darboux Integrable Functions.
At this point, I want to discuss the result that a bounded function f : [a, b] R
with only finitely many discontinuities is Darboux integrable. So I wrote up a direct proof of this and it was long and messy. Afterwards I realized that a better
argument is by induction on the number of discontinuities. One then has to prove
the result for a function with a single discontinuity (base case), and assuming the
result for every function with n discontinuities, prove it for every function with
n + 1 discontinutities (inductive step). Here the inductive step is especially easy:
if f : [a, b] R has n + 1 points of discontinuity, we can choose c (a, b) such
that f |[a,c] has exactly one discontinuity and f |[c,b] has exactly n discontinuities.
The restricted functions are Darboux integrable by the base case and the induction
hypothesis, and as we know, this implies that f : [a, b] R is Darboux integrable.
So really it is enough to treat the case of a bounded function with a single
discontinuity. It turns out that it is no trouble to prove a stronger version of this.
Theorem 8.11. Let f : [a, b] R be bounded. Suppose that for all c (a, b),
f |[c,b] : [c, b] R is Darboux integrable. Then f is Darboux integrable and
Z b
Z b
lim+
f=
f.
ca

Proof. Let M > 0 be such that |f (x)| M for all x [a, b]. Fix > 0 and
consider partitions P of [a, b] with x1 = a + . For such partitions,
(f, P) = (f, P [a, a + ]) + (f, P [a + , b]).
Since the infimum of f on any subinterval of [a, b] is at least M and the supremum
is at most M , (f, [a, a + ]) 2M , which we can make as small as we wish by
taking small enough. Similarly, having chosen , we may make (f, P [a + , b])
as small as we like with a suitable choice of P, since f is assumed to be Darboux
integrable on [a + , b]. Thus we can make the oscillation at most  for any  > 0,
so f is Darboux integrable on [a, b]. The second statement follows easily:
Z b
Z b
Z c
|
f
f| = |
f | 2M (c a),
a

and the last quantity goes to zero as c a+ .

3For once we do not introduce a name but only a piece of notation. In an earlier course on
this subject I called this quantity the oscillation of f on P, but this is not especially apt. Better
perhaps would be to call (f, P) the discrepancy of f and P, since it is the difference between
the upper and the lower sum. But in fact it is simplest not to call it anything but (f, P)!

3. FURTHER RESULTS ON INTEGRATION

163

Of course there is an analogous result with the roles of a and b reversed.

This result is also telling us that under certain situations we need not bother to
consider improper integrals: the improper integral will exist iff the conventional
Darboux integral exists. This will make a lot more sense in the context of a discussion on improper integrals, so we defer the point until then.
Theorem 8.12. Let f : [a, b] R be a bounded function which is continuous
except at a finite set of points in its domain. Then f is Darboux integrable.
Exercise 3.1: Use Theorem 8.11 (and its reflected version) to prove Corollary 8.12.
Theorem 8.13. A monotone function f : [a, b] R is Darboux integrable.
Proof. By reflection it suffices to deal with the case of a weakly increasing
case. For such functions a miracle occurs: for every partition P = {a = x0 < x1 <
. . . < xn1 < xn = b} of [a, b] and for all 0 i n 1, the infimum of f on
[xi , xi+1 ] is attained at the left endpoint xi and the supremum of f on [xi , xi+1 ] is
attained at the right endpoint xi+1 . Therefore
L(f, P) = f (x0 )(x1 x0 ) + f (x1 )(x2 x1 ) + . . . + f (xn1 )(xn xn1 ),
U (f, P) = f (x1 )(x1 x0 ) + f (x2 )(x2 x1 ) + . . . + f (xn )(xn xn1 ).
Things simplify further if we simply take Pn to be the uniform partition of [a, b]
into n equal parts. Then
ba
(f (x0 ) + . . . + f (xn1 )) ,
L(f, Pn ) =
n
ba
(f (x1 ) + . . . + f (xn )) ,
U (f, Pn ) =
n
so


ba
U (f, Pn ) L(f, Pn ) =
(f (b) f (a)) .
n
Thus taking n sufficiently large we may take U (f, Pn ) L(f, Pn ) arbitrarily small,
so f is Darboux integrable by Darbouxs criterion.

Theorem 8.13 goes beyond Theorem 8.12: there are increasing functions which are
discontinuous at infinitely many points. To construct such things becomes much
easier with some knowledge of infinite sequences and series, so we defer this discussion until later, except to make the following advertisement: for any real numbers
a < b and any injective function s : Z+ [a, b], there exists an increasing function
f : [a, b] R which is discontinuous at precisely the points s(n) [a, b].
Example 3.2 (Thomaes Function): Let f : [0, 1] R be defined by f (0) = 0,
f ( pq ) = 1q , and f of any irrational number is zero. Then f is continuous at 0 and at
all irrational numbers but discontinuous at every rational number. Thus not only
does f have infinitely many points of discontinuity in [0, 1], they are dense: any
nontrivial subinterval contains at least one
R 1point of discontinuity. We claim that
nevertheless f is Darboux integrable and 0 f = 0. First observe that since every
subinterval [xi , xi+1 ] contains an irrational number, the infimum of f on [xi , xi+1 ]
is zero, so for any partition P the lower sum is L(f, P) = 0. It follows then that
R1
= supP L(f, P) = 0, and thus if f integrable, its integral is zero. It remains to
0
show that for each  > 0 we may find a partition P of [0, 1] such that U (f, P) < .

164

8. INTEGRATION

To see this, observe that for any fixed , there are only finitely many nonzero rational numbers pq in [0, 1] with q : indeed there is at most 1 such with denominator
1, at most 2 with denominator 2, and so forth (and in fact there are less than
this because e.g. in our terminology the denominator of 24 is actually 2, since
1
2
4 = 2 in lowest terms). Suppose then that there are N points x in [0, 1] such that
f (x) . Choose a partition P such that each of these points x lies in the interior
of a subinterval of length at most N . Since the maximum value of f on [0, 1] is 1,
the term of the upper sum corresponding to each of these N bad subintervals is

; since there are N bad subintervals over all, this part of the sum
at most 1 2N
is at most N N = , and the remaining part of the sum is at most  times the
length of [a, b] = [0, 1], i.e., at most . Thus U (f, P)  +  = 2. Since of course
lim0 2 = 0, this shows that f is Darboux integrable.
All of our results so far have been in the direction of exhibiting examples of Darboux
integrable functions with increasingly large sets of discontinuities. What about the
other direction: is there, for instance, a Darboux integrable function which is discontinuous at every point? In fact, no:
Theorem 8.14. Let f : [a, b] R be Darboux integrable. Let S be the set of
x [a, b] such that f is continuous at x. Then S is dense in [a, b]: i.e., for all
a x < y b, there exists z (x, y) such that f is continuous at z.
Proof. Step 1: We show that there is at least one c [a, b] such that f is
continuous at c. We will construct such a c using the Nested Intervals Theorem:
recall that if we have a sequence of closed subintervals [an , bn ] such that for all n,
an bn ,
an an+1 ,
bn+1 bn ,
there is at least one c such that an c bn for all n: indeed supn an inf n bn ,
so any c [supn an , inf n bn ] will do. Since f is Darboux integrable, for all n Z+
there is a partition Pn = {a = x0 < x1 < . . . < xn1 < xn = b} of [a, b] such that
(44)

(f, Pn ) =

n1
X
i=0

(f, [xi , xi+1 ])(xi+1 xi ) <

ba
.
n

Now (44) implies that for at least one 0 i n 1 we have (f, [xi , xi+1 ]) <
for, if not, (f, [xi , xi+1 ]) n1 for all i and thus

1
n:

1
1
1
xn x0
ba
(x1 x0 ) + (x2 x1 ) + . . . + (xn xn1 ) =
=
,
n
n
n
n
n
contradiction. We will use this analysis to choose a nested sequence of subintervals.
First we take n = 1 and see that there is some closed subinterval [xi , xi+1 ] of [a, b]
on which (f, [xi , xi+1 ]) < 1. We then define a1 = xi , b1 = xi+1 , and instead of
considering f as defined on [a, b], we now consider it as defined on the subinterval
[a1 , b1 ]. Since f is Darboux integrable on [a, b], we know it is also Darboux integrable
on [a1 , b1 ], so the above argument still works: there exists a partition P2 of [a1 , b1 ]
such that for at least one subinterval [xi , xi+1 ] [a1 , b1 ] we have (f, [xi , xi+1 ]) <
1
2 . We then put a2 = xi (this is not necessarily the same number that we were
calling xi in the previous step, but we will stick with the simpler notation) and
b2 = xi+1 and have defined a sub-subinterval [a2 , b2 ] [a1 , b1 ] [ab ] on which
(f, Pn )

3. FURTHER RESULTS ON INTEGRATION

165

(f, [a2 , b2 ]) < n1 . Now, continuing in this way we construct a nested sequence
[an , bn ] of closed subintervals such that for all n Z+ , (f, [an , bn ]) < n1 . Now
apply the Nested Intervals Theorem: there exists c R such that c [an , bn ] for
all n Z+ . It follows that for all n Z+
1
(f, c) (f, [an , bn ])) < ,
n
i.e., (f, c) = 0 and thus f is continuous at c by Proposition 8.10.
Step 2: To show f has infinitely many points of continuity, its enough to show that
for all N Z+ f is continuous at at least N distinct points, and we can do this by
induction, the base case N = 1 being Step 1 above. So suppose we have already
shown f is continuous at x1 < x2 < . . . < xN in [a, b]. Choose any A, B R with
a x1 < A < B < x2 b. Once again, since f : [a, b] R is Darboux integrable,
the restriction of f to [A, B] is Darboux integrable on [A, B]. Applying Step 1, we
get c [A, B] such that f is continuous at c, and by construction c is different from
all the continuity points we have already found. This completes the induction step,
and thus it follows that f is continuous at infinitely many points of [a, b].

3.3. A supplement to the Fundamental Theorem of Calculus.
Theorem 8.15. Let f : [a, b] R be differentiable and suppose f 0 is Darboux
Rb
integrable. Then a f 0 = f (b) f (a).
Proof. Let P be a partition of [a, b]. By the Mean Value Theorem there is
ti [xi , xi+1 ] such that f (xi+1 ) f (xi ) = f 0 (ti )(xi+1 xi ). Then we have
mi (f 0 )(xi+1 xi ) f 0 (ti )(xi+1 xi ) Mi (f 0 )(xi+1 xi )
and thus
mi (f 0 )(xi+1 xi ) f (xi+1 ) f (xi ) Mi (f 0 )(xi+1 xi ).
Summing these inequalities from i = 0 to n 1 gives
L(f 0 , P) f (b) f (a) U (f 0 , P).
Rb
Since for the integrable function f , a f is the unique number lying in between all
Rb
lower and upper sums, we conclude f (b) f (a) = a f 0 .

How is Theorem 8.15 different from Theorem 8.1c)? Only in a rather subtle way:
in order to apply Theorem 8.1c) to f 0 , we need f 0 to be continuous, whereas in
Theorem 8.15 we are assuming only that f 0 is Darboux integrable. Every continuous function is Darboux integrable but, as we have seen, there are discontinuous Darboux integrable functions. What about discontinuous, Darboux integrable
derivatives? The possible discontinuities of a monotone function are incompatible
with the possible discontinuities of a derivative: if f 0 is monotone, it is continuous.
So we must look elsewhere for examples. In fact, we return to an old friend.
Example 3.3: Let a, b (0, ) and let fa,b be given by x 7 xa sin( x1b ), x 6= 0
and 0 7 0. Then fa,b is infinitely differentiable except possibly at zero. It is continuous at 0, the sine of anything is bounded, and limx0 xa = 0, so the product
approaches zero. To check differentiability at 0, we use the definition:
ha sin( h1b )
f (h) f (0)
1
= lim
= lim ha1 sin( b ).
h0
h0
h0
h
h
h

f 0 (0) = lim

166

8. INTEGRATION

0
This limit exists and is 0 iff a 1 > 0 iff a > 1. Thus if a > 1 then fa,b
(0) = 0. As
0
for continuity of fa,b at zero, we compute the derivative for nonzero x and consider
the limit as x 0:
1
1
0
fa,b
(x) = axa1 sin( b ) bxab1 cos( b ).
x
x
The first term approaches 0 for a > 1. As for the second term, in order for the
0
limit to exist we need a > b + 1. This calculation shows that fa,b
is continuous at
0 iff a > b + 1; so in this case we can apply the first version of the Fundamental
Theorem of Calculus to conclude for instance that
Z x
0
fa,b
= fa,b (x) fa,b (0) = fa,b (x).
0
0
Next, if a < b + 1, then fa,b
is unbounded near 0, hence is not Darboux integrable
on any interval containing zero. But there is a third case: if a = b + 1, then
0
0
limx0 fa,b
does not exist, but fa,b
is bounded on any closed, bounded interval, say
[0, x]. Therefore Theorem 8.15 applies to give
Z x
0
fb+1,b
= fb+1,b (x) fb+1,b (0) = fb+1,b (x)
0

for all b > 0.

This example, delicate though it is, provides the first evidence that the Darboux
integral may not be the last word on integration theory. It is natural to want a
fundamental theorem of calculus in which no hypothesis is needed on f 0 . Thus we
want an integration theory in which every derivative f 0 is integrable. As we have
just seen, the Darboux integral is not such a theory. In graduate real analysis,
one meets a more powerful and general integral, the Lebesgue integral, which
remedies many of the defects of the Darboux integral...but not this one! In fact
0
are not Lebesgue integrable either. There is a
for b > a + 1 the derivatives fa,b
more recent theory which allows every derivative to be integrable (and satisfy the
fundamental theorem of calculus): it is called the Kurzweil-Henstock integral.
3.4. New Integrable Functions From Old.
In this section we show that performing many familiar, elementary operations on
integrable functions yields integrable functions. By showing any interest in these
results we are exploring the extent that the class of Darboux integrable functions
goes beyond the class of all continuous functions on [a, b]. Indeed, we know that
every continuous function is integrable, and all of the operations we are discussing
here will take continuous functions to continuous functions.
Theorem 8.16. Let f, g : [a, b] R be Darboux integrable functions.
Rb
Rb
a) For any constant C, Cf is Darboux integrable and a Cf = C a f .
b) The function f + g is Darboux integrable, and moreover
Z b
Z b
Z b
f +g =
f+
g.
a

Proof. a) The idea here is simply that C may be factored out of the lower
sums and the upper sums. The details may be safely left to the reader.
b) Let I [a, b] be a subinterval, and let mf , mg , mf +g be the infima of f , g

3. FURTHER RESULTS ON INTEGRATION

167

and f + g, respectively on I. Things would certainly be easy for us if we had

mf + mg = mf +g , but observe that this need not be the case: e.g. consider
f (x) = x and g(x) = x on [1, 1]. Then mf = 1, mg = 1 and mf +g = 0, so
mf + mg < mf +g . However there is a true inequality here: we always have
mf + mg mf +g .
Applying this on every subinterval of a partition P gives us
L(f, P) + L(g, P) L(f + g, P).
Similarly, denoting by Mf , Mg and Mf +g the suprema of f , g and f + g on some
subinterval I, we have
Mf + Mg Mf +g
and this implies that for every partition P of [a, b] we have
U (f + g, P) U (f, P) + U (g, P).
Combining these inequalities gives
(45)

L(f, P) + L(g, P) L(f + g, P) U (f + g, P) U (f, P) + U (g, P).

Moreover, subtracting the smallest quantity from the largest gives

0 (f + g, P) (f, P) + (g, P);
since f and g are Darboux integrable, for  > 0, there is P1 such that (f, P1 ) <
and P2 such that (g, P2 ) < 2 . Taking P = P1 P2 , both inequalities hold, so
(f + g, P) (f, P) + (g, P) <


2



+ = .
2 2

By Darbouxs Criterion, f +g is Darboux integrable, and the inequalities (45) imply

Rb
Rb
Rb
f + g = a f + a g.

a
Theorem 8.17. Let f : [a, b] [c, d] be Darboux integrable and g : [c, d] R
be continuous. Then the composite function g f : [a, b] R is Darboux integrable.
Proof. Since g is continuous on [c, d], by the Extreme Value Theorem it is
bounded: there exists M such that |g(x)| M for all x [c, d].
Fix  > 0. By the Uniform Continuity Theorem, there exists > 0 such that

. Shrinking if necessary, we may assume
|x y| = |g(x) g(y)| ba+2M
<


.
b a + 2M

Since f is Darboux integrable, there exists a partition P = {a = x0 < x1 < . . . <

xn1 < xn = b} of [a, b] such that (f, P) < 2 .
We divide the index set {0, 1, . . . , n 1} into two subsets: let S1 be the set of
such i such that (f, [xi , xi+1 ])(xi+1 xi ) , and let S2 be the complementary
set of all i such that (f, [xi , xi+1 ])(xi+1 xi ) > . We have set things up such

that for all i S1 , (g f, [xi , xi+1 ]) ba+2M
. Since S1 {0, . . . , n 1}, we have
X
iS1

(xi+1 xi )

n1
X
i=0

(xi+1 xi ) = b a.

168

8. INTEGRATION

On the other hand, since M f (x) M for all x [a, b], the oscillation of f on
any subinterval of [a, b] is at most 2M . Thus we get
X
X
(f, [xi , xi+1 ])(xi+1 xi )
(f, [xi , xi+1 ])(xi+1 xi ) +
(g f, P) =
iS2

iS1

X

(b a) + 2M
(xi+1 xi ).
b a + 2M
iS1
iS2
iS2
P
(Note that reasoning as above also gives iS2 xi+1 xi (b a), but this is not
good enough: using it would give us a second term of 2M (b a), i.e., not something
that we can make arbitrarily small.) Here is a better estimate:
X
1 X
1 X
(xi+1 xi ) =
(xi+1 xi ) <
(f, [xi , xi+1 ])(xi+1 xi )

1

(xi+1 xi ) + (2M )

iS2

(xi+1 xi )

iS2

iS2

n1
1
1
1X
(f, [xi , xi+1 ])(xi+1 , xi ) = (f, P) < 2 = .
i=0

Using this estimate, we get

(g f, P)


(b a)
2M 
(b a) + 2M <
+
= .
b a + 2M
b a + 2M
b a + 2M


The proof of Theorem 8.17 becomes much easier if we assume that g is not merely
continuous but a Lipschitz function. Recall that f : I R is Lipschitz if there
exists C 0 such that for all x, y I, |f (x) f (y)| C|x y|.
Example: f : R R by x 7 |x| is a Lipschitz function. Indeed, the reverse
triangle inequality reads: for all x, y R,
||x| |y|| |x y|,
and this shows that 1 is a Lipschitz constant for f .
Exercise: a) For which functions may we take C = 0 as a Lipshitz constant?
b) Let I be an interval. Show that for every Lipshitz function f : I R, there is
a smallest Lipschitz constant.
Proposition 8.18. Let f : [a, b] R be a C 1 -function. Then M = maxx[a,b] |f 0 (x)|
is a Lipschitz constant for f .
Proof. Let x < y [a, b]. By the Mean Value Theorem, there is z (x, y)
such that f (x)f (y) = f 0 (z)(xy), so |f (x)f (y)| |f 0 (z)||xy| M |xy|. 
Lemma 8.19. Let f : I [c, d] be bounded and g : [c, d] R a Lipshitz function
with Lipschitz constant C. Then (g f, I) C(f, I).
Exercise: Prove Lemma 8.19.
Theorem 8.20. Let f : [a, b] [c, d] be Darboux integrable, and let g : [c, d]
R be Lipschitz with contant C. Then g f : [a, b] R is Darboux integrable.

3. FURTHER RESULTS ON INTEGRATION

169

Proof. Fix  > 0. Since f is integrable, there is a partition P of [a, b] with

(f, P) =

n1
X

(f, [xi , xi+1 ])(xi+1 xi ) <

i=0

Then by Lemma 8.19 we have (g f, P) =

C


.
C

Pn1

(g f, [xi , xi+1 ])(xi+1 xi )

!
n1

X
(f, [xi , xi+1 ])(xi+1 xi ) < C
= .
C
i=0
i=0


Corollary 8.21. Let f : [a, b] R be Darboux integrable. Then |f | : [a, b]
R is Darboux integrable, and we have the integral triangle inequality
Z b
Z b
f|
|f |.
|
a

Proof. Since g(x) = |x| is a Lipshitz function, by Theorem 8.20 g f = |f | is

Darboux integrable on [a, b]. Moreover, since |f | f |f |, by (I2) we have
Z b
Z b
Z b

|f |
f
|f |,
a

so
Z
|

Z
f|

|f |.
a


Corollary 8.22. Let f1 , f2 : [a, b] R be Darboux integrable. Then the
product f1 f2 : [a, b] R is Darboux integrable.
Proof. It is really just a dirty trick: we have the identity
(f1 + f2 )2 (f1 f2 )2
.
4
Now, by Theorem 8.15, both f1 + f2 and f1 f2 are Darboux integrable. Since f1
and f2 are Darboux integrable, they are both bounded, so there exists M with
f1 ([a, b]), f2 ([a, b]) [M, M ]. The function g(x) = x2 is C 1 on the closed,
bounded interval [M, M ] and thus Lipschitz there. Thus Theorem 8.20 applies
to show that (f1 + f2 )2 and (f1 f2 )2 are Darboux integrable, and then Theorem
8.16 appplies again to show that f1 f2 is Darboux integrable.

Rb Rb
Rb
Warning: It is usually not the case that a f1 f2 = a f1 a f2 !
f1 f2 =

Example 3.6: Let f : [1, 2] [0, 1] be the function which takes the value 1q at
every rational number pq and 0 at every irrational number, and let g : [0, 1] R
be the function which takes 0 to 0 and every x (0, 1] to 1. Then f is Darboux
integrable, g is bounded and discontinuous only at 0 so is Darboux integrable, but
g f : [1, 2] R takes every rational number to 0 and every irrational number to
1, so is not Darboux integrable. Thus we see that the composition of Darboux integrable functions need not by Darboux integrable without some further hypothesis.
Example 3.7: Above we showed that if g is continuous and f is Darboux integrable then the composition g f is Darboux integrable; then we saw that if g

170

8. INTEGRATION

and f are both merely Darboux integrable, g f need not be Darboux integrable.
So what about the other way around: suppose f is continuous and g is Darboux
integrable; must g f be Darboux integrable? The answer is again no; the easiest
counterexample I know is contained in a paper of Jitan Lu [Lu99].
4. Riemann Sums, Dicing, and the Riemann Integral
We now turn to the task of reconciling G. Darbouxs take on the integral with
B. Riemanns (earlier) work. Riemann gave an apparently different construction
Rb
of an integral a f which also satisfies axioms (I0) through (I3). By virtue of the
Rb
uniqueness of the integral of a continuous function, the Riemann integral R a f of
Rb
a continuous function agrees with our previously constructed Darboux integral a f .
But this leaves open the question of how the class of Riemann integrable functions compares with the class of Darboux integrable functions. In fact, although
the definitions look different, a function f : [a, b] R is Riemann integrable iff it
Rb
Rb
is Darboux integrable and then R a f = a f . Thus what we really have is a rival
construction of the Darboux integral, which is in some respects more complicated
but also possesses certain advantages.
It turns out however to be relatively clear that a Riemann integrable function
is necessarily Darboux integrable. This suggests a slightly different perspective: we
view Riemann integrability as an additional property that we want to show that
every Darboux integrable function possesses. This seems like a clean way to go: one
Rb
the one hand, it obviates the need for things like R a f . On the other, it highlights
what is gained by this construction: namely, further insight on the relationship of
Rb
the upper and lower sums U (f, P) and L(f, P) to the integral a f . At the moment
the theory tells us that if f is Darboux integrable, then for every  > 0 there exists
some partition P of [a, b] such that U (f, P ) L(f, P ) < . But this is not very
explicit: how do we go about finding such a P ? In the (few!) examples in which
we showed integrability from scratch, we saw that we could always take a uniform
partition Pn : in particular it was enough to chop the interval [a, b] into a sufficiently
large number of pieces of equal size. In fact, looking back at our first proof of the
integrability of continuous functions, we see that at least if f is continuous, such
uniform partitions always suffice. The key claim that we wish to establish in this
section is that for any integrable function we will have (f, Pn ) <  for sufficiently
large n. In fact we will show something more general than this: in order to achieve
(f, P) < , we do not need P to have equally spaced subintervals but only to have
all subintervals of length no larger than some fixed, sufficiently small constant .
Given a function f : [a, b] R and a partition P of [a, b], we will also introduce
Rb
a more general approximating sum to a f than just the upper and lower sums,
namely we will define and consider Riemann sums. The additional flexibility of
Riemann sums is of great importance in the field of numerical integration (i.e.,
the branch of numerical analysis where we quantitatively study the error between
a numerical approximation to an integral and its true value), and it pays some
modest theoretical dividends as well. But the Riemann sums are little more than
a filigree to the main dicing property of the Darboux integral alluded to in the
last paragraph: the Riemann sums will always lie in between the lower and upper

4. RIEMANN SUMS, DICING, AND THE RIEMANN INTEGRAL

171

sums, so if we can prove that the upper and lower sums are good approximations, in
Rb
whatever sense, to the integral a f , then the same has to be true for the Riemann
sums: they will be carried along for the ride.
4.1. Riemann sums.
Let f : [a, b] R be any function, and let P be a partition of [a, b]. Instead
of forming the rectangle with height the infimum (or supremum) of f on [xi , xi+1 ],
we choose any point xi [xi , xi+1 ] and take f (xi ) as the height of the rectanPn1
gle. In this way we get a Riemann sum i=0 f (xi )(xi+1 xi ) associated to
the function f , the partition P, and the choice of a point xi [xi , xi+1 ] for all
0 i n 1. Given a partition P = {a = x0 < x1 < . . . < xn1 < xn = b},
a choice of xi [xi , xi+1 ] for 0 i n 1 is called a tagging of P and gets
a notation of its own, say = {x0 , . . . , xn1 }. A pair (P, ) is called a tagged
partition, and given any function f : [a, b] R and any tagged partition (P, )
of [a, b], we associate the Riemann sum
R(f, P, ) =

n1
X

f (xi )(xi+1 xi ).

i=0

Let us compare the Riemann sums R(f, ) to the upper and lower sums. Just
because every value of a function f on a (sub)interval I lies between its infimum
and its supremum, we have that for any tagging of P,
L(f, P) R(f, P, ) U (f, P).

(46)

Conversely, if f is bounded then for all  > 0 we can find xi , x

[xi , xi+1 ]
i
such that sup(f, [xi , xi+1 ]) f (xi ) +  and inf(f, [xi , xi+1 ]) f (x
)
, and it
i
follows that the upper and lower sums associated to f and P are the supremum
and infimum of the possible Riemann sums R(f, P, ):
(47)

sup R(f, P, ) = U (f, P), inf R(f, P, ) = L(f, P).

Exercise 4.1: Show that (47) holds even if f is unbounded. More precisely, show:
a) If f is unbounded above, then sup R(f, P, ) = U (f, P) = .
b) If f is unbounded below, then inf R(f, P, ) = L(f, P) = .
From inequalities (46) and (47) the following result follows almost immediately.
Theorem 8.23. For a function f : [a, b] R, the following are equivalent:
(i) f is Darboux integrable.
(ii) For all  > 0, there exists a real number I and a partition P of [a, b] such that
for any refinement P of P and any tagging of P we have
|R(f, P, ) I| < .
Rb
If the equivalent conditions hold, then I = a f .
Proof. (i) = (ii): If f is Darboux integrable, then by Darbouxs Criterion
there is a partition P such that (f, P ) = U (f, P ) L(f, P ) < . For any
refinement P of P we have (f, P) (f, P ), and moreover by integrablity
Z b
L(f, P)
f U (f, P).
a

172

8. INTEGRATION

For any tagging of P we have also

L(f, P) R(f, P, ) U (f, P).
Rb
Thus both R(f, P, ) and a lie in an interval of length less than , and it follows
Rb
that their distance from each other, |R(f, P, ) a f |, is less than .
(ii) = (i): Fix  > 0. Since U (f, P ) = sup R(f, P , ) and L(f, P ) =
inf R(f, P , ), if |R(f, P , ) I| <  for all , then |U (f, P ) I| <  and
|L(f, P ) I| <  and thus U (f, P ) L(f, P ) < 2. Since  was arbitrary, f
is Darboux integrable. Moreover, the only number I such that for all  > 0, there
Rb
is a partition P with |U (f, P ) I| < , |L(f, P ) I| <  is a f .

Theorem 8.23 gives a sense in which a function is Darboux integrable iff the RieRb
mann sums R(f, P, ) converge to a f . However, the proof involves little more
than pushing around already established facts. Riemann considered a different, a
Rb
priori stronger, sense in which the Riemann sums converge to a f . In the next
section we discuss this and show that it holds for every Darboux integrable function.
4.2. Dicing.
For a partition P = {a = x0 < x1 < . . . < xn1 < xn = b} of [a, b], its mesh
|P| is maxi (xi+1 xi ), i.e., the largest length of a subinterval in P. The mesh of
a partition is a better measure of its size than the number of points it contains.
One can think of a kitchen assistant dicing vegetables making a lot of knife cuts
doesnt ensure a good dicing job: you might have some tiny pieces but also some
large pieces. Rather a proper dicing will ensure the mesh is sufficiently small.
Lemma 8.24. Let P1 be a partition of [a, b] and let P2 P1 be a refinement of
P1 . Then |P2 | |P1 |.
Proof. Proving this is merely a matter of absorbing the definitions, so we
leave it to the reader as a good opportunity to stop and think.

Lemma 8.25. (Dicing Lemma) Let f : [a, b] R be bounded. For all  > 0,
there exists > 0 such that for all partitions P of [a, b] with |P| < ,
Z b
Z b
(48)
f L(f, P) <  and U (f, P)
f < .
a

Proof. (Levermore) Let  > 0. There exists a partition P of [a, b] such that
Z b
Z b


f L(f, P ) < , 0 U (f, P )
f< .
0
2
2
a
a
Suppose |f (x)| M for all x [a, b]. Let N be the number of subintervals of P .
Choose > 0 such that 2M N < 2 . We claim (48) holds for any partition P with
|P| < . Indeed, let P be any partition with |P| < , and put P 0 = P P . Now
Z b
Z b
(49)
0
f L(f, P) = ( f L(f, P 0 )) + (L(f, P 0 ) L(f, P)),
a

Z
(50)

0 U (f, P)

b
0

f = (U (f, P )
a

f ) + (U (f, P) U (f, P 0 )).

4. RIEMANN SUMS, DICING, AND THE RIEMANN INTEGRAL

173

We will establish the claim by showing that the two terms on the right hand side of
(49) are each less than 2 and, similarly, that the two terms on the right hand side
of (50) are each less than 2 . Using the Refinement Lemma (Lemma 8.3), we have
Z b
Z b

f L(f, P 0 )
f L(f, P ) <
0
2
a
a
and
Z


.
2
a
a
This gives two of the four inequalities. As for the other two: since P 0 is a refinement
of P = {a = x0 < . . . < . . . < xN 1 < xN = b}, for 0 i N 1, Pi0 :=
P [xi , xi+1 ] is a partition of [xi , xi+1 ]. By the Refinement Lemma,
0 U (f, P 0 )

0 L(f, P 0 ) L(f, P) =

f U (f, P )

n1
X

f<

(L(f, Pi0 ) inf(f, [xi , xi+1 ])) ,

i=0

0 U (f, P) U (f, P 0 ) =

n1
X

(sup(f, [xi , xi+1 ]) U (f, Pi0 )) .

i=0

Because P 0 has at most N 1 elements which are not contained in P, there are at
most N 1 indices i such that (xi , xi+1 ) contains at least one point of Pi0 . For all
other indices the terms are zero. Further, each nonzero term in either sum satisfies
0 L(f, Pi0 ) inf(f, [xi , xi+1 ]) 2M (xi+1 xi ) < 2M ,
0 sup(f, [xi , xi+1 ]) U (f, Pi0 ) 2M (xi+1 xi ) < 2M .
Because there are at most N 1 nonzero terms, we get

0 L(f, P 0 ) L(f, P) < 2M N < ,
2

0
0 U (f, P) U (f, P ) < 2M N < .
2
So the last terms on the right hand sides of (49) and (50) are each less than 2 . 
We can now deduce the main result of this section.
Theorem 8.26. a) For a function f : [a, b] R, the following are equivalent:
(i) f is Darboux integrable.
(ii) There exists a number I such that for all  > 0, there exists > 0 such that for
all partitions P of [a, b] of mesh at most and all taggings of P,
|R(f, P, ) I| < .
(iii) For every sequence (Pn , n ) of tagged partitions of [a, b] such that |Pn | 0,
the sequence of Riemann sums R(f, Pn , n ) is convergent.
Rb
b) If condition (ii) holds for some real number I, then necessarily I = a f .
c) If condition (iii) holds, then for every sequence (Pn , n ) of tagged partitions with
Rb
|Pn | 0, R(f, Pn , n ) a f .
Rb
Rb
Proof. a) (i) = (ii): if f is Darboux integrable, then a f = a f , and
property (ii) follows immediately from the Dicing Lemma (Lemma 8.25).
(ii) = (iii): Indeed, if (ii) holds then for any sequence of tagged partitions
(Pn , n ) with |Pn | 0, we have R(f, Pn , n ) I.

174

8. INTEGRATION

(iii) = (i): We will show the contrapositive: if f is not Darboux integrable,

then there is a sequence (Pn , n ) of tagged partitions with |Pn | 0 such that the
sequence of Riemann sums R(f, Pn , n ) is not convergent.
Case 1: Suppose f is unbounded. Then for any partition P of [a, b] and any M > 0,
there exists a tagging such that |R(f, P, )| > M . Thus we can build a sequence
of tagged partitions (Pn , n ) with |Pn | 0 and |R(f, Pn , n )| .
Case 2: Suppose f is bounded but not Darboux integrable, i.e.,
Z b
Z b
f < .
f<
<
a

For n Z , let Pn be the partition into n subintervals each of length ba

n . Since
U (f, P) = sup R(f, P, ) and L(f, P) = inf L(f, P, ), for all n Z+ there is one
tagging tn of Pn with L(f, Pn ) R(f, Pn , tn ) < L(f, Pn ) + n1 and another tagging
Tn of Pn with U (f, Pn ) R(f, Pn , Tn ) > U (f, Pn ) n1 . By the Dicing Lemma,
Z b
Z b
lim L(f, Pn ) =
f, lim U (f, Pn ) =
f,
+

and it follows, for instance by a squeezing argument, that

Z b
f,
lim R(f, Pn , tn ) =
n

a
b

Z
lim R(f, Pn , Tn ) =

f.
a

Now let n be tn if n is odd and Tn is n is even. Then we get a sequence of tagged

partitions (Pn , n ) with |Pn | 0 such that
Z b
f,
lim R(f, P2n+1 , 2n+1 ) =
n

Z
lim R(f, P2n , 2n ) =

f.
a

Rb
Rb
Since a f 6= a f , the sequence {R(f, Pn , n )}
n=1 does not converge.
b) This follows from Theorem 8.23: therein, the number I satisfying (ii) was unique.
Our condition (ii) is more stringent, so there can be at most one I satisfying it.
c) This is almost immediate from the equivalence (ii) (iii) and part b): we
leave the details to the reader.

4.3. The Riemann Integral.
By definition, a function f : [a, b] R satisfying condition (ii) of Theorem 8.26 is
Riemann integrable, and the number I associated to f is called the Riemann
integral of f . In this language, what we have shown is that a function is Riemann
integrable iff it is Darboux integrable, and the associated integrals are the same.
As mentioned above, Riemann set up his integration theory using the Riemann integral. Some contemporary texts take this approach as well. It really is a bit messier
though: on the one hand, the business about the taggings creates another level
of notation and another (minor, but nevertheless present) thing to worry about.

4. RIEMANN SUMS, DICING, AND THE RIEMANN INTEGRAL

175

But more significantly, the more stringent notion of convergence in the definition
of the Riemann integral can be hard to work with: directly showing that the composition of a continuous function with a Riemann integrable function is Riemann
integrable seems troublesome. On the other hand, there are one or two instances
where Riemann sums are more convenient to work with than upper and lower sums.
Example 4.2: Suppose f, g : [a, b] R are both Darboux integrable. We wanted
to show that f + g is also Darboux integrable...and we did, but the argument was
slightly complicated by the fact that we had only inequalities
L(f, P) + L(g, P) L(f + g, P), U (f, P) + U (g, P) U (f + g, P).
However, for any tagging of P, the Riemann sum is truly additive:
R(f + g, P, ) = R(f, P, ) + R(g, P, ).
Using this equality and Theorem 8.23 leads to a more graceful proof that f + g is
Rb
Rb
Rg
integrable and a f + g = a f + a . I encourage you to work out the details.
Example 4.3: Let f : [a, b] R be differentiable such that f 0 is Darboux integrable. Choose a partition P = {a = x0 < x1 < . . . < xn1 < xn = b} of [a, b].
Apply the Mean Value Theorem to f on [xi , xi+1 ]: there is xi (xi , xi+1 ) with
f (xi+1 ) f (xi ) = f 0 (xi )(xi+1 xi ).
Now {xi }n1
i=0 gives a tagging of P. The corresponding Riemann sum is
R(f, P, ) = (f (x1 ) f (x0 )) + . . . + (f (xn ) f (xn1 )) = f (b) f (a).
Thus, no matter what partition of [a, b] we choose, there is some tagging such that
Rb
the corresponding Riemann sum for a f 0 is exactly f (b) f (a)! Since the integral
of an integrable function can be evaluated as the limit of any sequence of Riemann
Rb
sums over tagged partitions of mesh approaching zero, we find that a f 0 is the limit
of a sequence each of whose terms has value exactly f (b) f (a), and thus the limit
is surely f (b) f (a). This is not really so different from the proof of the supplement to the Fundamental Theorem of Calculus we gave using upper and lower sums
(and certainly, no shorter), but I confess I find it to be a more interesting argument.
Remark: By distinguishing between Darboux integrable functions and Riemann
integrable functions, we are exhibiting a fastidiousness which is largely absent in
the mathematical literature. It is more common to refer to the Riemann integral to mean either the integral defined using either upper and lower sums and
upper and loewr integrals or using convergence of Riemann sums as the mesh of a
partition tends to zero. However, this ambiguity leads to things which are not completely kosher: in the renowned text [R], W. Rudin gives the Darboux version of
the Riemann integral, but then gives an exercise involving recognizing a certain
limit as the limit of a sequence of Riemann sums and equating it with the integral
of a certain function: hes cheating! Let us illustrate with an example.
Example 4.4: Compute limn

Pn

n
k=1 k2 +n2 .

Solution: First observe that as a consequence of Theorem 8.26, for any Darboux

176

8. INTEGRATION

integrable function f : [0, 1] R, we have

Z 1
n
1X k
lim
f( ) =
f.
n n
n
0
k=1

Now observe that our limit can be recognized as a special case of this:
n
n
n
n
X
1 X n2
1X
1
n
1X k
=
lim
=
lim
f ( ),
lim
=
lim
n n
n n
n
k 2 + n2
k 2 + n2
n
( nk )2 + 1 n n
k=1

k=1

where f (x) =

1
x2 +1 .

k=1

k=1

Thus the limit is

Z 1
dx

= arctan 1 arctan 0 = .
2
2
0 x +1

Anyway, we have done our homework: by establishing Theorem 8.26 we have earned
the right to use Darboux integral and Riemann integral interchangeably. In
fact however we will generally simply drop both names and simply speak of integrable functions and the integral. For us, this is completely safe. However,
as mentioned before, you should be aware that in more advanced mathematical
analysis one studies other kinds of integrals, especially the Lebesgue integral.
5. Lesbesgues Theorem
5.1. Statement of Lebesgues Theorem.
In this section we give a characterization of the Riemann-Darboux integrable functions f : [a, b] R due to H. Lebesgue. Lebesgues Theorem is a powerful, definitive
result: many of our previous results on Riemann-Darboux integrable functions are
immediate corollaries. Here we give an unusually elementary proof of Lebesgues
Theorem following lecture notes of A.R. Schep.
For an interval I, we denote by `(I) its length: for all < a b < ,
`((a, b)) = `([a, b)) = `((a, b]) = `([a, b]) = b a,
`((a, )) = `([a, )) = `((, b)) = `((, b]) = `((, )) = .
We define a subset S R to have measure zero if for all  > 0, there is a sequence
PN
{In }Sof open intervals in R such that (i) for all N 1, n=1 `(In ) , and (ii)

S i=1 In , i.e., every point of S lies in at least one of the intervals I.

Proposition 8.27. Let {Sn }
of subsets of R. If each Sn has
n=1 be a sequence
S
measure zero, then so does their union S = n=1 Sn .
Proof. Fix  > 0 and n Z+P
. Since Sn has measure zero, it admits a covering

by open intervals {In,k }
k=1 with
k `(In,k ) < 2k+1 . Then if we collect together all
the open intervals into one big set {In,k }n,k and use the fact that sums of series
with positive terms do not depends on the ordering of the terms, we see that
X
XX
X 
In,k =
`(In,k ) <
= .
2k+1
n
n
n,k
k
S
Further, {In,k } is a covering of S = n=1 Sn by open intervals. There is one nagging
technical problem: the covering is not given by a sequence i.e., not indexed by
the positive integers but rather by a double sequence i.e., indexed by ordered

5. LESBESGUES THEOREM

177

pairs (n, k) of positive integers. But this is easily solved, simply by reindexing the
terms of a double sequence an,k via a single sequence bn , e.g. as follows:
a1,1 , a1,2 , a2,1 , a1,3 , a2,2 , a3,1 , a1,4 , a2,3 , a3,2 , a4,1 , a1,5 . . . .

Theorem 8.28. (Lebesgue) For a function f : [a, b] R, let D(f ) be the set
of x [a, b] at which f is discontinuous. The following are equivalent:
(i) f is Riemann-Darboux integrable on [a, b].
(ii) f is bounded and D(f ) has measure zero.
Thus a function is Riemann-Darboux integrable on [a, b] iff it is bounded (which
we already know is necessary) and its set of discontinuities is small in the sense
of having measure zero.
Exercise: Let f, g : [a, b] R be Riemann-Darboux integrable functions. Use
Theorem 8.28 to give a quick proof that f + g and f g are both Riemann-Darboux
integrable. (Suggestion: also use Proposition 8.27.)
5.2. Preliminaries on Content Zero.
The notion of sets of measure zero is a vitally important one in the further study
of functions of a real variable. In fact one goes further, by assigning a measure
(to be entirely technicallyPprecise, an outer measure) to any subset S R as the

infimum of all quantities

L n=1 `(In ) as {In }n=1 ranges over all sequences of open
intervals with S
n=1 In . This is the beginning of a branch of real analysis
called measure theory. In contrast, the following definition ought to be viewed
as merely a technical tool used in the proof of Lebesgues Theorem.
A subset S R has content zero if for every  > 0, there exists N Z+ and
PN
SN
open intervals I1 , . . . , IN such that i=1 `(Ii )  and S i=1 Ii .
Exercise: Show that if in the definitions of measure zero and content zero we replace
open interval by interval, we do not change the classes of sets of measure zero
or content zero.
Exercise: a) Show that Q [0, 1] has measure zero.
b) Show that Q [0, 1] does not have content zero. (Suggestion: induct on N , the
number of intervals.)
Lemma 8.29. Let f : [a, b] R be non-negative and Riemann-Darboux inteRb
grable, with a f = 0. Then:
a) For all C > 0, the set
SC = {x [a, b] | f (x) C}
has content zero.
b) The set S = {x [a, b] | f (x) > 0} has measure zero.
Proof. Fix  > 0, and choose a partition P of [a, b] such that U(f, P)) <  C.
Let i be the set of indices such that SC [xi , xi+1 ] is nonempty. Thus, i I

178

8. INTEGRATION

Mi = sup(f, [xi , xi+1 ]) C. It follows that

X
X
 C > U(f, P)
Mi (xi+1 xi ) C
`([xi , xi+1 ]).
iI

iI

Thus SC iI [xi , xi+1 ] and iI `([xi , xi+1 ]) < . Thus SC has content zero.
S
b) We have S = n=1 S n1 . Each S n1 has content zero, hence certainly has measure
zero. Now apply Proposition 8.27.

Exercise: Let E be a subset of R, and let E be the set of x R such that for all
> 0, there is y E with |x y| . Show that E has content zero iff E has
measure zero.
5.3. Proof of Lebesgues Theorem.
The proof of Lebesgues Theorem uses Heine-Borel (Theorem 6.17) and also the
fact that a bounded monotone sequence is convergent (Theorem 10.11), which we
will not discuss until Chapter 9. All in all this section should probably be omitted
on a first reading, and only the extremely interested student need proceed.
Step 0: Let P = {a = x0 < x1 < . . . < xn1 < xn = b} be a partition of
[a, b]. As usual, for 0 i n 1, put
mi = inf(f, [xi , xi+1 ]), Mi = sup(f, [xi , xi+1 ],
so that
L(f, P) =

n1
X

mi `([xi , xi+1 ]), U (f, P) =

n1
X

i=0

Mi `([xi , xi+1 ]).

i=0

Let , : [a, b] R be the lower and upper step functions, i.e., takes the value mi
on [xi , xi+1 ) and takes the value Mi on [xi , xi+1 ). The functions , are bounded
with only finitely many discontinuities, so are Riemann-Darboux integrable by X.X;
moreover
Z
Z
b

= L(f, P),
a

= U (f, P).
a

Step 1: Suppose that f : [a, b] R is Riemann-Darboux integrable. By X.X,

f is bounded, so it suffices to show that the set D(f ) of discontinuities of f has
measure zero. Let {Pk } be a sequence of partitions of [a, b] with Pk Pk+1 such
that |Pk | 0 (e.g. let Pk be the uniform subdivision of [a, b] into 2k subintervals).
Let k and k be the lower and upper step functions for Pk , so that for all k and
all x [a, b) and
Z b
Z b
k = L(f, Pk )
f,
Z

a
b

Z
k = U(f, Pk )

f.
a

Further, since Pk Pk+1 for all k, for all x [a, b), the sequence k (x) is increasing
and bounded above, hence convergent, say to (x); and similarly the sequence k (x)
is decreasing and bounded below, hence convergent, say to (x). For all x [a, b),
we have
k (x) (x) f (x) (x) k (x),

6. IMPROPER INTEGRALS

179

and thus
Z

Z
k

Z
f

k .
a

Rb
These inequalities show that and are Riemann-Darboux integrable and a =
Rb
Rb
= a f . Applying Lemma 8.29 to , we get that = except on a set of
a
measure zero. Let
[
E = {x [a, b] | (x) 6= (x)} Pk .
k

Since E is the union of a set of measure zero with a the union of a sequence of
finite sets, it has measure zero. We claim that f is continuous at every point of
[a, b] \ E, which will be enough to complete this direction of the proof.
proof of claim Fix x0 [a, b] \ E. Since (x) = (x), there is k Z+ such
that k (x0 ) k (x0 ) < . Further, since x0
/ Pk , there is > 0 such that k k
is constant on the interval (x0 , x0 + ), and for x in this interval we have
 < k (x0 ) k (x0 ) f (x) f (x0 ) k (x0 ) k (x0 ) < ,
so f is continuous at x0 .
Step 2: Suppose now that f is bounded on [a, b] and continuous on [a, b] \ E for a
subset E of measure zero. We must show that f is Riemann-Darboux integrable
on [a, b]. Let M be such that |f (x)| M for all x [a, b]. Fix  > 0. Since E has
measure
zero, there is a sequence {In }
n=1 of open intervals covering E such that
P

.
For
x

[a,
b]
\
E,
by
continuity of f there is an open interval
`(I
)
<
n
n=1
4M

Jx containing x such that for all y, z Jx , |f (y) f (z)| 2(ba)
. Applying the
Heine-Borel Theorem (Theorem 6.17) to the open covering {Ik } {Jx }x[a,b]\E of
[a, b], there is a finite subcovering, say {Ik }N
k=1 {Jxi }. Let P = {a = t0 < t1 <
. . . < tN = b} be the partition of [a, b] whose points are the endpoints of I1 , . . . , IN
and the endpoints of the intervals Jxi which lie in [a, b]. or 0 j N 1, the
subinterval (tj , tj+1 ) is contained in some Ik or some Jxi . Let I be the set of i such
that (ti , ti+1 ) is contained in Ik for some k. Then
U(f, P) L(f, P) =

N
1
X

`([ti , ti+1 ]) sup(f (x) f (y) : x, y [ti , ti+1 ])

i=0

`([ti , ti+1 ]) 2M +

iI

`([ti , ti+1 ])

iI
/

<


2(b a)

  

2M + (b a)
= .
4M
2(b a)

Since  > 0 was arbitrary, f is Riemann-Darboux integrable on [a, b].

6. Improper Integrals
6.1. Basic definitions and first examples.

180

8. INTEGRATION

6.2. Non-negative functions.

Things become much simpler if we restrict to functions f : [a, ) [0, ): in
words, we assume that f is defined and non-negative for all sufficiently large x.
As usual we Rsuppose that f is integrable on [a, b] for all bR a, so we may dex

fine F (x) = a f for x a. Then the improper integral a f is convergent iff

limx F (x) exists. But here is the point: since Rf is non-negative,
R x F is weakly
Rx
x
increasing: indeed, for x1 x2 , F (x2 ) F (x1 ) = a 2 f a 1 f = x12 0. Now
for any weakly increasing function F : [a, ) R we have
lim F (x) = sup(f, [a, )).

In other words, the limit exists as a real number iff F is bounded; otherwise, the
limit is : there is no oscillation! We deduce:
Proposition 8.30. Let f : [a,
R ) [0, ) be integrable
R on every finite interval
[a, N ] with N a. Then either a f is convergent or a f = .
InRview of Proposition 8.30, we may write the two alternatives as

a f < (convergent)
R
a f = (divergent).
R
2
Example: Suppose we wish to compute ex . Well, we are out of luck: this
integral cannot be destroyed ahem, I mean computed by any craft that we
here possess.4 The problem is that we do not know any useful expression for the
2
antiderivative of ex (and in fact it can be shown that this antiderivative is not an
elementary function). But because we are integrating a non-negative function,
we know that the integral is either convergent or infinite. Can we at least decide
which of these alternatives is the case?
Yes we can. First, since we are integrating an even function, we have
Z
Z
2
2
ex = 2
ex .

x2

Now the function e

= ex2 is approaching 0 very rapidly; in fact a function like
1
x
e = ex exhibits exponential decay, and our function is even smaller than that,
R
2
at least for sufficiently large x. So it seems like a good guess that 0 ex < .
Can we formalize this reasoning?
Yes we can. First, for all x 1, x x2 , and since (ex )0 = ex > 0, ex is in2
2
creasing, so for all x 1 ex ex , and finally, for all x 1, ex ex . By the
familiar (I2) property of integrals, this gives that for all N 1,
Z N
Z N
2
ex
ex ,
1

and taking limits as N we get

Z
Z
2
ex
1

ex .

4Elrond: The Ring cannot be destroyed, Gimli, son of Gl

oin, by any craft that we here
possess. The Ring was made in the fires of Mount Doom. Only there can it be unmade. It must
be taken deep into Mordor and cast back into the fiery chasm from whence it came.

6. IMPROPER INTEGRALS

181

This integral is much less scary, as we know an antiderivative for ex : ex . Thus

Z
Z
1
x2

ex = ex |
e1 ) = .
1 = (e
e
1
1
R x2
R x2
Note that we replaced 0 e
with 1 e
: does that make a difference? Well,
R1
2
yes, the difference between the two quantities is precisely 0 ex , but this is a
proper integral, hence finite, so removing it changes the value of the integral
which we dont know anyway! but not whether it converges. However we can be
2
slightly more quantitative: for all x R, x2 0 so ex 1, and thus
Z 1
Z 1
2
ex
1 = 1,
0

and putting it all together,

Z
Z
Z
x2
x2
e
=2
e
=2

x2

Z
+

e
1

x2



1
= 2.735 . . . .
2 1+
e

The exact value of the integral is known we just dont possess the craft to find it:
Z

2
ex = = 1.772 . . . .

This is indeed smaller than our estimate, but it is a bit disappointingly

R far 2off.
Later in the course we will develop methods suitable for approximating 0 ex to
any desired finite degree of
accuracy, and we will be able to check for instance that
this integral agrees with to, say, 20 decimal places.
The following simple theorem formalizes the argument we used above.
Theorem 8.31. (Comparison Test For Improper Integrals) Let f, g : [a, )
[0, )R be integrable onR[a, N ] for all N a. Suppose f (x) g(x) for all x a.

a) If a g < , then a f < .

R
R
b) If a f = , then a g = .
Proof. By property (I2) of integrals, for all N a since f (x) g(x) on [a, N ]
RN
RN
we have a f a g. Taking limits of both sides as N gives
Z
Z
(51)
f
g;
a

here each Rside is either a Rnon-negative real number

(51) both parts
R or . FromR


follow: if a g < then a f < , whereas if a f = then a g = .5
Theorem 8.32. (Limit Comparison Test For Improper Integrals)
Let f, g : [a, ) [0, ) be integrable on [a, N ] for all N a. Consider condition
(S): there exists b Ra and M > 0 such
R that f (x) M g(x) for all x b.

a) If (S) holds and a g < , then a f < .

R
R
b) If (S) holds and a f = , then a g = .
(x)
=
c) Suppose there exists b a such that g(x) > 0 for all x b and that limx fg(x)
L < . Then (S) holds.
5In fact parts a) and b) are contrapositives of each other, hence logically equivalent.

182

8. INTEGRATION

d) Suppose there exists b a such that g(x) > 0 for all x b and that limx
L with 0 < L < . Then
Z
Z
g < .
f <

f (x)
g(x)

Proof. For any b a, since f and g are integrable on [a, b], we have
Z
Z
Z
Z
f <
f < ,
g <
g < .
a

R
R
a) If f (x) M g(x) for all x b, then b f b M g = M b g. Thus if
R
R
R
R
g < , b g < , so b f < and thus finally a f < .
a
b) Note that this is precisely the contrapositive of part a)! Or to put it in a
slightly different way: suppose (S) holds. Seeking a contradiction, we also suppose
R
Rf
g < . Then by part a), a f < , contradiction.
a
(x)
(x)
c) Since limx fg(x)
= L < , there is b a such that for all x b, fg(x)
L + 1.
Thus for all x b we have f (x) (L + 1)g(x), so (S) holds.
(x)
d) Suppose limx fg(x)
= L (0, ). By part c), (S) holds, so by part a), if
R
R
1
g < , then a f < . Moreover, since L 6= 0, limx fg(x)
(x) = RL (0, ). So
a
R

part c) applies with the roles of f and g reversed: if a f < then a g < . 
Although from a strictly logical perspective part d) of Theorem 8.32 is the weakest,
it is the most useful in practice.
7. Some Complements
Riemann was the first to give a complete, careful treatment of integration as a process that applies to a class of functions containing continuous functions and yields
the Fundamental Theorem of Calculus: he was certainly not the last.
In more advanced analysis courses one studies the Lebesgue integral. This is
a generalization of the Riemann integral that, extremely roughly, is modelled on
the idea of approximations in small ranges of y-values rather than small ranges of
x-values. Suppose for instance that f : [0, 1] R has a finite range y1 , . . . , yN .
Rb
Then the idea is that a f is determined by the size of each of the sets f 1 (yi ) on
which f takes the value yi . For instance this is the perspective of the expected value
in probability theory: given a function defined on a probability space, its expected
value is equal to the sum of each of the values at each possible outcome multipled
by the probability that that outcome occurs. Notice that when f is a step function,
the sets f 1 (yi ) are just intervals, and the size of an interval (a, b), [a, b), (a, b] or
[a, b] is just its length b a. However, a general function f : [0, 1] R with finite
image need not have the sets f 1 (yi ) be intervals, so the first and hardest step in
Lebesgues theory is a rigorous definition of the measure of a much more general
class of subsets of [a, b]. This has grown into a branch of mathematics in its own
right, measure theory, which is probably at least as important as the integration theory it yields. However, Lebesgues integral has many technical advantages
over the Riemann integral: the supply of Lebesgue integrable functions is strictly
larger, and in particular is much more stable under limiting operations. Later we

7. SOME COMPLEMENTS

183

will study sequences of functions fn : [a, b] R and see that a crucial issue is the
permissibility of the interchange of limit and integral: i.e., is
Z
n

lim fn ?

fn =

lim

For the Riemann integral it is difficult to prove general theorems asserting that the
left hand side exists, a phenomenon which evidently makes the identity hard to
prove! Probably the real showpiece of Lebesgues theory is that useful versions of
such convergence theorems become available without too much pain.
In classical probability theory one encounters both discrete and continuous
probability distributions. The idea of a continuous probability distribution on [a, b]
can be modelled by a non-negative integrable function f : [a, b] R such that
Rb
f = 1. This is studied briefly in the next chapter. Lebesgues theory allows for
a
distributions like the Dirac delta function, a unit mass concentrated at a single
point. It turns out that one can model such discrete distributions using a simpler
theory than Lebesgues theory due to the Dutch mathematician T.J. Stieltjes (born
in 1856; Riemann was born in 1826, Lebesgue in 1875). This is a relatively mild
souping up of Riemanns theory which views the dx as referring to the function
f (x) = x and replaces it by dg for a more general function g : [a, b] R. Some
undergraduate texts develop this Riemann-Stieltjes integral, most notably [R].
But this integral adds one extra layer of technical complication to what is already
the most technical part of the course, so I feel like in a first course like this one
should stick to the Riemann integral.
In the twentieth century various generalizations of the Riemann integral were developed which are much more closely related to the Riemann integral than the
Lebesgue integral avoiding any need to develop measure theory but are yet
more powerful than the Lebesgue integral: i.e., with a larger class of integral functions, good convergence theorems, and a simple general form of the Fundamental
Theorem of calculus. This integral is now called the Kurzweil-Henstock integral
although it was also developed by Denjoy, Perron and others. This theory however
is not so widely known, whereas every serious student of mathematics studies the
Lebesgue integral. As alluded to above, the reason is probably that the measure
theory, while a lot to swallow the first time around, is in fact a highly important
topic (in particular making a profound connection between analysis and geometry)
in its own right. It is also extremely general, making sense in contexts far beyond
intervals on the real line, whereas the Kurzweil-Henstock integral is more limited
in scope.
Aside from developing his new and more sophisticated theory of integration, Lebesgue
did much insightful work in the more classical context, as for instance his remarkable characterization of Riemann integrable functions presented above. I want to
mention a further achievement of his which has been almost forgotten until recently: from the perspective of differential calculus, a natural but very difficult
question is Which functions are derivatives? We now know that in particular
every continuous function f : [a, b] R is the derivative of some other function
F , but in order to construct the antiderivative we had to develop the theory of

184

8. INTEGRATION

the Riemann integral. Is it possible to construct an antiderivative of an arbitrary

Rb
continuous function f in some other way than as a f ? (I dont mean is there some
Rb
other function which is antiderivative: as we know, a f + C is the most general
antiderivative of f . But for instance we certainly dont need an integration theory
to know that polynomials have antiderivatives: we can just write them down. The
question is whether one can find an antiderivative without expressing it in terms of
an integration process.)
Lebesgue gave a considerably more elementary construction of antiderivatives of
every continuous function. The basic idea is shockingly simple: on the one hand it
is easy to write down an antiderivative of a piecewise linear function: it is simply
an appropriate piecewise quadratic function. On the other hand every continuous function can be well approximated by piecewise linear functions, so that one
can write down the antiderivative of any continuous function as a suitable limit of
piecewise quadratic functions. A nice modern exposition is given in [Be13].

CHAPTER 9

Integral Miscellany
1. The Mean Value Therem for Integrals
Theorem 9.1. Let f, g : [a, b] R. Suppose that f is continuous and g is
integrable and non-negative. Then there is c [a, b] such that
Z b
Z b
(52)
f g = f (c)
g.
a

Proof. By the Extreme Value Theorem, f assumes a minimum value m and

Rb
a maximum value M on [a, b]. Put I = a g; since g is non-negative, so is I. We
have
Z b
Z b
Z b
(53)
mI =
mg
fg
M g = M I.
a

Case 1: Suppose I = 0. Then by (53), for all c [a, b] we have

Z b
Z b
f g = 0 = f (c)
g.
a

Case 2: Suppose I 6= 0. Then dividing (53) through by I gives

Rb
fg
M.
m a
I
Rb

fg

Thus aI lies between the minimum and maximum values of the continuous function f : [a, b]R R, so by the Intermediate Value Theorem there is c [a, b] such
b
Rb
fg
that f (c) = Rab g . Multiplying through by a g, we get the desired result.

a

Exercise 9.1.1: Show that in the setting of Theorem 9.1, we may take c (a, b).
Exercise 9.1.2: Show by example that the conclusion of Theorem 9.1 becomes false
even when g 1 if the hypothesis on continuity of f is dropped.
Exercise 9.1.3: Suppose that f : [a, b] R is differentiable and f 0 is continuous on [a, b]. Deduce the Mean Value Theorem for f from the Mean Value Theoem
for Integrals and the Fundamental Theorem of Calculus.1
2. Some Antidifferentiation Techniques
2.1. Change of Variables.
1Since the full Mean Value Theorem does not require continuity of f 0 , this is not so exciting.
But we want to keep track of the logical relations among these important theorems.
185

186

9. INTEGRAL MISCELLANY

2.2. Integration By Parts.

Recall the product rule: if f, g : I R are differentiable, then so is f g and
(f g)0 = f 0 g + f g 0 .
Equivalently,
f g 0 = (f g)0 f 0 g.
We have already (!) essentially derived Integration by Parts.
Theorem 9.2. (Integration by Parts) Let f, g : I R be continuously differ0
0
entiable functions:
R 0 i.e., f and
R 0 g are defined and continuous on I.
a) We have f g = f g f g, in the sense that subtracting from f g any antiderivative of f 0 g gives an antiderivative of f g 0 .
b) If [a, b] I then
Z b
Z b
f g 0 = f (b)g(b) f (a)g(a)
f 0 g.
a

Exercise 9.2.1: a) Prove Theorem 9.2. (Yes, the point is that its easy.)
b) Can the hypothesis on continuous differentiability of f and g be weakened?
Exercise 9.2.2: Use Theorem 9.2a) to find antiderivatives for the following functions.
a) xn ex for 1 n 6.
b) log x.
c) arctan x.
d) x cos x.
e) ex sin x.
f) sin6 x.
g) sec3 x, given that log(sec x + tan x) is antiderivative of sec x.
Exercise 9.2.3: a) Show that for each n Z+ , there is a unique monic (= leading
d
(Pn (x)ex ) = xn ex .
coefficient 1) polynomial Pn (x) such that dx
b) Can you observe/prove anything about the other coefficients of Pn (x)? (If you
did part a) of the preceding exercise, you should be able to find patterns in at least
three of the non-leading coefficients.)
Exercise 9.2.4: Use Theorem 9.2a) to derive the following reduction formulas:
here m, n Z+ and a, b, c R.
a)
Z
Z
n1
1
cosn2 x.
cosn x = cosn1 x sin x +
n
n
b)
Z
Z
1
1
x
2n 3
=
+
.
2
2
n
2
2
2
n1
2
2
(x + a )
2a (n 1)(x + a )
2a (n 1)
(x + a2 )n1
c)
Z

m1
1
sinm1 ax cosn+1 ax +
sin ax cos ax =
a(m + n)
m+n
m

sinm2 ax cosn ax

3. APPROXIMATE INTEGRATION

1
n1
sinm+1 ax cosn1 ax +
a(m + n)
m+n

187

sinm ax cosn2 ax.

In this text we are not much interested in antidifferentiation techniques per se we

regard this as being somewhat forward-looking, since computer algebra packages
are by now better at this than any human ever could be so most of our interest in
integration by parts comes from part b) of Theorem 9.2. In fact this result, simple
though it is, is a remarkably powerful tool throughout honors calculus and analysis.
R
Proposition 9.3. For all n N, 0 xn ex dx = n!.
By induction on n. Base case (n = 0):
R Proof.
x
x
e
=
e
|0 = e (e0 ) = 0 (1) = 1 = 0!
0
R
Induction step: let n N and assume 0 xn ex dx = n!. Now to make progress in
R n+1 x
e dx, we integrate by parts, taking u = xn + 1, dv = ex dx.
evaluating 0 x
Then du = (n + 1)xn dx, v = ex , and
Z
Z
n+1 x
n x oo
x
e dx = (n + 1)x e |0
(ex (n + 1)xn )dx
0

Z
= (0 0) + (n + 1)

IH

xn ex dx = (n + 1)n! = (n + 1)!


2.3. Integration of Rational Functions.
3. Approximate Integration
Despite the emphasis on integration (more precisely, antidifferentiation!) techniques
in a typical freshman calculus class, it is a dirty secret of the trade that in practice
many functions you wish to integrate do not have an elementary antiderivative,
i.e., one that can be written in (finitely many) terms of the elementary functions one
learns about in precalculus mathematics. Thus one wants methods for evaluating
definite integrals other than the Fundamental Theorem of Calculus. In practice it
Rb
would often be sufficient to approximate a f rather than know it exactly.2
Theorem 9.4. (Endpoint Approximation Theorem) Let f : [a, b] R be differentiable with bounded derivative: there is M 0 such that |f 0 (x)| M for
all x [a, b]. For n Z+ , let Ln (f ) be the left endpoint Riemann sum obtained by dividing
spaced subintervals: thus for 0 i n 1,

[a, b] into n equally
Pn1
ba
and
L
(f
)
=
f
xi = a + i ba
n
i=0 (xi ) n . Then
n
Z
 

 b

(b a)2 M 1


f Ln (f )

 a

2
n
Proof. Step 1: We establish the result for n = 1. Note that L1 (f ) = (b
a)f (a). By the Racetrack Principle, for all x [a, b] we have
M (x a) + f (a) f (x) M (x a) + f (a)
2It would be reasonable to argue that if one can approximate the real number R b f to any
a

degree of accuracy, then in some sense one does know it exactly.

188

9. INTEGRAL MISCELLANY

and thus
Z

Z
(M (x a) + f (a)))

Z
f

(M (x a) + f (a)).
a

Thus
Z b
M
M
2
(b a) + (b a)f (a)
f
(b a)2 + (b a)f (a),
2
2
a
which is equivalent to

Z
 M
 b


f L1 (f )
(b a)2 .


 a
2
Step 2: Let n Z+ . Then

Z

!
n1
n1

 b
X Z xi+1
X Z xi+1
ba


ba
f Ln (f ) = |
f f (xi )(
) |
f f (xi )(
)|



 a

n
n

xi
xi
i=0

i=0

Step 1 applies to each term in the latter sum to give

Z
 n1


 b
 X M ba
(b a)2 M 1


2
(
)
=
.
f

L
(f
)



n
 a

2
n
2
n
n=0


Exercise 9.3.1: Show that Theorem 9.4 holds verbatim for with left endpoint sums
replaced by right endpoint sums Rn (f ).
Exercise 9.3.2: a) Suppose f : [a, b] R is increasing. Show that
Z b
1
0
f Ln (f ) (f (b) f (a))(b a) .
n
a
b) Derive a similar result to part a) for the right endpoint sum Rn (f ).
c) Derive similar results to parts a) and b) if f is decreasing.
In view of the preceding exercise there is certainly no reason to prefer left endpoint sums over right endpoint sums. In fact, if you stare at pictures of left and
right endpoint sums for a while, eventually it will occur to you to take the average
of the two of them: with = ba
n , we get
(54)
1
1
Tn (f ) = Ln (f )+Rn (f ) = f (a)+f (a+)+f (a+2) . . .+f (a+(n1))+ f (b).
2
2
On each subinterval [xi , xi+1 ] the
 average of the left and right endpoint sums is
f (xi )(xi+1 xi )+f (xi+1 )(xi+1 xi )
2

f (xi )+f (xi+1 )

2

(xi+1 xi ). Notice that this is the

area of the trapezoid whose upper edge is the line segment which joins (xi , f (xi ))
Rb
to (xi+1 , f (xi+1 ). For this reason the approximation Tn (f ) to a f is often called
the trapezoidal rule. Thats cute, but a more useful way of thinking about Tn (f )
is that it is obtained by linearly interpolating f on each subinterval [xi , xi+1 ]
via the line which matches f at the two endpoints of the interval.
Thinking of the trapezoidal rule this way suggests that Tn (f ) should, at least for
Rb
nice f and large n, a better approximation to a f than either the left or right

3. APPROXIMATE INTEGRATION

189

endpoint sums. For instance, if f is linear then on each subinterval we are approxRb
imating f by itself and thus Tn (f ) = a f . This was certainly not the case for
the endpoint rules: in fact, our proof of Theorem 9.4 showed that for each fixed
M , taking f to be linear with slope M (or M ) was the worst case scenario: if
we approximate a linear function `(x) with slope M on the interval [a, b] by the
horizontal line given by the left endpoint (say), then we may as well assume that
Rb
`(a) = 0 and then we are approximating ` by the zero function, whereas a ` is
2
the area of the triangle with base (b a) and height M (b a), hence M
2 (b a) .
Another motivation is that if f is increasing, then Ln (f ) is a lower estimate and
Rn (f ) is an upper estimate, so averaging the two of them gives something which is
closer to the true value.
The following result confirms our intuitions.
Theorem 9.5. (Trapezoidal Approximation Theorem) Let f : [a, b] R be
twice differentiable with bounded second derivative: there is M 0 such that
|f 00 (x)| M for all x [a, b]. Then for all n Z+ we have
Z
 

 b

(b a)3 M
1


f Tn (f )
.

 a

12
n2
ba
n ,

and for i = 0, . . . , n 1, let xi = a + i. Also put

Z xi+1
t
i : [0, ] R, i (t) = (f (xi ) + f (xi + t))
f.
2
xi

Proof. Put =

As motivation for this definition we observe that

n1
X

(55)

Z
i () = Tn (f )

i=0

f.
a

We have i (0) = 0 and, using the Fundamental Theorem of Calculus,

1
t
1
t
(f (xi ) + f (xi + t))+ f 0 (xi +t)f (xi +t) = (f (xi )f (xi +t))+ f 0 (xi +t),
2
2
2
2
so 0i (0) = 0 and
0i =

00i =

1 0
1
1
1
f (xi + t) + f 0 (xi + t) + tf 00 (xi + t) = tf 00 (xi + t).
2
2
2
2

Put
A = inf(f 00 , [a, b]) B = sup(f 00 , [a, b]).
Then for all 0 i n 1 and t [0, ] we have
A
B
t 00i (t) t.
2
2
0
Since i (0) = 0, integrating and applying the Racetrack Principle gives
A 2
B
t 0i (t) t2 ;
4
4
doing it again (since i (0) = 0) and plugging in t = gives
A 3
B 3
i ()
.
12
12

190

9. INTEGRAL MISCELLANY

Summing these inequalities from i = 0 to n 1 and using (55) we get

Z b
B 3
A 3
f
n Tn (f )
n.
12
12
a
Substituting in = ba
n gives




Z b
B(b a)3
A(b a)3 1
Tn (f )
f
.
12
n2
12
a
Since |A|, |B| M , we get the desired result:
 
Z


 b
(b a)3 M
1


.
f Tn (f )


 a
12
n2

00

Exercise 9.3.3: In the setting of Theorem 9.5, suppose

on [a, b].

 f is3 continuous
Rb
(ba) f 00 (c)
1
a) Show: there is c [a, b] such that Tn (f ) a f =
12
n2 .
Rb
b) Show: if f is linear, the trapezoidal rule is exact: a f = Tn (f ) for all n Z+ .
Rb
c) Show: if f = x2 + bx + c, the trapezoidal rule is never exact: a f 6= Tn (f ).
Exercise 9.3.4: Show: if f : [a, b] R is convex, then Tn (f )

Rb
a

f.

Looking back at the formula (54) for the trapezoidal rule we notice that we have
given equal weights to all the interior sample points but only half as much weight
to the two endpoints. This is an instance of a heuristic in statistical reasoning:
extremal sample points are not as reliable as interior points. This suggests a different kind of approximation scheme: rather than averaging endpoint Riemann sums,
lets consider the Riemann sums in which each sample point xi is the midpoint
of the subinterval [xi , xi+1 ]. Dividing the interval [a, b] into n subintervals of equal
width = ba
n as usual, this leads us to the midpoint rule
Mn (f ) =

n1
X

1
f (a + (i + )).
2
i=0

Exercise 9.3.5: a) Show: if f is linear, the midpoint rule is exact:

Rb
b) Show: if f : [a, b] R is convex, then Mn (f ) a f .

Rb
a

f = Mn (f ).

The following result concerning the midpoint rule is somewhat surprising: it gives
a sense in which the midpoint rule is twice as good as the trapezoidal rule.
Theorem 9.6. (Midpoint Approximation Theorem) Let f : [a, b] R be twice
differentiable with bounded second derivative: there is M 0 such that |f 00 (x)| M
for all x [a, b]. Then for all n Z+ we have
Z
 

 b

1
(b a)3 M


.
f Mn (f )

 a

24
n2
for i = 0, . . . , n 1, let xi = a + (i + 21 ). Also put
Z xi +t
i : [0, /2] R, i (t) =
f 2tf (xi ).

Proof. Put =

ba
n ;

xi t

3. APPROXIMATE INTEGRATION

191

As motivation for this definition we observe that

Z b
n1
X
(56)
(/2) =
f Mn (f ).
a

i=0

We have i (0) = 0, and using the Fundamental Theorem of Calculus,

0i (t) = f (xi + t) f (xi t)(1) 2f (xi ) = (f (xi + t) + f (xi t)) 2f (xi ).
So 0i (0) = 0,
00 (t) = f 0 (xi + t) f 0 (xi t).
Applying the Mean Value Theorem to f 0 on [xi t, xi + t] we get a point xi,t
(xi t, xi + t) such that
f 00 (xi,t ) =

f 0 (xi + t) f 0 (xi t)
00 (t)
= i ,
2t
2t

or
00i (t) = 2tf 00 (xi,t ).
Put
A = inf(f 00 , [a, b]) B = sup(f 00 , [a, b]).
Then for all 0 i n 1 and t [0,
2 ] we have
2tA 00i (t) 2tB.
Integrate and apply the Racetrack Principle twice as in the proof of Theorem 9.5,
and plug in t =
2 to get
(57)

B 3
A 3
i (/2)
.
24
24

Summing (57) from i = 0 to n 1, using (56) and substituting = ba

n gives




Z
b
(b a)3 A 1
(b a)3 B 1

.
f Mn (f )
2
24
n
24
n2
a
Since |A|, |B| M , we get the desired result:
Z
 

 b

(b a)3 M
1


f Mn (f )
.

 a

24
n2

For integrable f : [a, b] R and n Z+ , we define Simpsons Rule
2
1
Mn (f ) + Tn (f ).
3
3
Thus S2n (f ) is a weighted average of the midpoint rule and the trapezoidal rule.
Since our previous results suggest that the midpoint rule is twice as good as the
trapezoidal rule, it makes some vague sense to weight in this way.
S2n (f ) =

Exercise: a) Show that for any a, b R such that a+b = 1, aMn (f )+bTn (f )
Rb
b) Deduce that S2n (f ) a f .

Rb
a

f.

Let us better justify Simpsons Rule. As with the other approximation rules, it

192

9. INTEGRAL MISCELLANY

is really a composite rule, i.e., it is based on iterating one simple approximation on

many equally spaced subintervals. So lets first concentrate on S2 (f ). We have
2
1
2
a+b
1 f (a) + f (b)
M2 (f ) + T2 (f ) = ((b a)f (
)) + (
(b a))
3
3
3
2
3
2


ba
a+b
=
f (a) + 4f (
) + f (b) .
6
2
Since we are dividing [a, b] into two subintervals, the width of each is = ba
2 , so
with x0 = a, x1 = a + , x2 = a + 2 = b, we have

S2 (f ) = (f (x0 ) + 4f (x1 ) + f (x2 )).

3
ba
+
For n Z , putting = 2n , x0 = a, x1 = a + , . . . , xn1 a + (2n 1), xn = b,
we get

S2n (f ) = (f (x0 ) + 4f (x1 ) + f (x2 ) + f (x2 ) + 4f (x3 ) + f (x4 ) + f (x4 ) + . . . + f (b)

3

= (f (x0 ) + 4f (x1 ) + 2f (x2 ) + . . . + 4f (xn1 ) + f (xn ).

3
S2 (f ) =

Lemma 9.7. For any quadratic function f (x) = Ax2 + Bx + C, Simpsons Rule
Rb
is exact: a f = S2n (f ) for all n Z+ .
Proof. Let = ba
2n . By splitting up [a, b] into n pairs of subintervals, it sufRb
fices to show a f = S2 (f ). This is done by a direct, if unenlightening, calculation:
Z b
B
A
(Ax2 + Bx + C) = (b3 a3 ) + (b2 a2 ) + C(b a)
3
2
a


A 2
B
= (b a)
(a + ab + b2 ) + (a + b) + C) ,
3
2
whereas




ba
a+b
f (a) + 4f
+ f (b)
6
2
!

2


ba
a+b
a+b
2
2
=
Aa + Ba + C + 4(A
+B
+ C) + (Ab + Bb + C)
6
2
2


ba
A(a2 + (a2 + 2ab + b2 ) + b2 ) + B(a + 2a + 2b + b) + 6C
=
6


A 2
B
= (b a)
(a + ab + b2 ) + (a + b) + C .
3
2

Thus whereas the endpoint rule is an approximation by constant functions and
the trapezoidal rule is an approximation by linear functions, Simpsons rule is an
approximation by quadratic functions. We may therefore expect it to be more accurate than either the Trapezoidal or Midpoint Rules. The following exercise (which
again, can be solved by a direct, if unenlightening, calculation), shows that it is
even a little better than we might have expected.
Exercise 9.3.6: Suppose f is a cubic polynomial. Show that Simpsons Rule is
Rb
exact: S2n (f ) = a f .

3. APPROXIMATE INTEGRATION

193

Since Simpsons Rule is exact for polynomials of degree at most 3 and a function f is a polynomial of degree at most three if and only if f (4) 0, it stands to
reason that Simpsons Rule will be a better or worse approximation according to
the magnitude of f (4) on [a, b]. The following result confirms this.
Theorem 9.8. (Simpson Approximation Theorem) Let f : [a, b] R be four
times differentiable with bounded fourth derivative: there is M 0 such that
|f (4) (x)| M for all x [a, b]. Then for all even n Z+ we have
 
Z


 b
(b a)4 M
1


f Sn (f )
.


 a
180
n4
Proof. For 0 i

n
2

1, let xi = a + (2i + 1) and put

i : [0, ] R, i (t) =

1
t(f (xi t) + 4(xi ) + f (xi + t))
3

As motivation for this definition we observe that

Z
n1
X
(58)
i (t) = Sn (f )
i=0

xi +t

f.
xi t

f.

We have i (0) = 0 and using the Fundamental Theorem of Calculus,

1
1
t(f 0 (xi t)+f 0 (xi +t))+ (f (xi t)+4f (xi )+f (xi +t))f (xi +t)f (xi t))
3
3
1
2
= t(f 0 (xi + t) f 0 (xi t)) (f (xi t) 2f (xi ) + f (xi + t)).
3
3
So 0i (0) = 0 and
0i (t) =

1
2
2
1 00
t(f (xi t)+f 00 (xi +t))+ (f 0 (xi +t)f 0 (xi t))+ f 0 (xi t) f 0 (xi +t)
3
3
3
3
1
1
= t(f 00 (xi + t) + f 00 (xi t)) (f 0 (xi + t) f 0 (xi t)).
3
3
So 00i (0) = 0 and
00i (t) =

1 000
1
1
1
t(f (xi +t)f 000 (xi t))+ (f 00 (xi +t)+f 00 (xi t)) f 00 (xi +t) f 00 (xi t)
3
3
3
3
1
= t(f 000 (xi + t) f 000 (xi t)).
3
000
So 000
on [xi t, xi + t], there
i (0) = 0. Applying the Mean Value Theorem to f
is (xi t, xi + t) such that

000
i (t) =

3 00
f 000 (xi + t) f 000 (xi t)
i (t)/t2 =
= f (4) ().
2
2t
Let A = inf(f (4) , [a, b]), B = sup(f (4) , [a, b]), so for all i and all t [0, ] we have
2A 2
2B 2
t 000
t .
i (t)
3
3
Integrate and apply the Racetrack Principle three times and plug in t = to get
(59)

A 3
B 3
i ()
.
90
90

194

9. INTEGRAL MISCELLANY

Summing (59) from i = 0 to n2 1,3 using (58) and substituting =





Z b
B(b a)5
A(b a)5 1

S
(f
)

.
n
180
n4
180
a

ba
n

gives


Exercise 9.3.7: a) In the setting of the Trapezoidal Rule, suppose moreover that
f 00 is continuous on [a, b]. Adapt the proof of Theorem 9.5 to show that there is
[a, b] such that


Z b
(b a)3 f 00 () 1
.
Tn (f )
f=
12
n2
a
b) Derive similar error equalities for the Endpoint, Midpoint and Simpson rules.
The proofs of Theorems 9.5, 9.6 and 9.8 are taken from [BS, Appendix D]. They
are admirably down-to-earth, using only the Mean Value Theorem and the Fundamental Theorem of Calculus. However it must be admitted that they are rather
mysterious.
Our motivations for considering the various rules were also a little light, but
I hope the reader can see that they have something to do with polynomial interpolation. This will be taken up in more detail in Chapter 12. For now we
just mention the sobering fact that all these rules and higher-degree analogues
of them were already known to Newton, developed in jointly with his younger
collaborator Roger Cotes in the early years of the 18th century.
4. Integral Inequalities
Theorem 9.9. (The Hermite-Hadamard Inequality) Let f : [a, b] R be convex
and continuous. Then
Rb


f
a+b
f (a) + f (b)
f
.
a
2
ba
2
a+b
Proof. Let s(x) = f ( a+b
2 ) +m(x 2 ) be a supporting line for f at x =

and let S(x) = f (a) +

f (b)f (a)
ba

a+b
2 ,

(x a) be the secant line, so

x [a, b], s(x) f (x) S(x).

(60)

Integrating these inequalities and dividing by b a we get

Rb
Rb
Rb
s
f
S
a
a

a .
ba
ba
ba
Now we have
Rb
R b a+b
Z b
s(x)
f ( 2 ) + m(x a+b
a+b
m
a+b
2 )
a
= a
= f(
)+
(x
)
ba
ba
2
ba a
2
 2

a+b
m
b
a2
a+b
a+b
m
a+b
= f(
)+

(b a) = f (
)+
0=
,
2
ba 2
2
2
2
ba
2
3Note that there are n terms here half as many as in the previous results. This gives rise
2

to an extra factor of

1
.
2

4. INTEGRAL INEQUALITIES

Rb

S(x)dx
a
=
ba

Rb
a

(f (a) +

f (b)f (a)
ba

(x a))dx
=

ba
Rb

Substituting these evaluations of

Rb

s
ba

195

f (a) + f (b)
.
2

a
and ba
into (60) gives
Rb
f
a+b
f (a) + f (b)
f(
) a
.
2
ba
2
a


Exercise: Show that the hypothesis of continuity in Theorem 9.9 is not necessary:
the inequality holds for any convex f : [a, b] R.
Let [a, b] be a closed interval, and let P : [a, b] [0, ) be a probability density:
Rb
i.e., P is integrable on [a, b] and a P = 1. For an integrable function f : [a, b] R,
we define the expected value
Z b
f (x)P (x)dx.
E(f ) =
a

Theorem 9.10. (Jensens Integral Inequality) Let P : [a, b] [0, ) be a

probability density function, : [c, d] R be convex, and let f : [a, b] [c, d] be
integrable. Then E(f ) [c, d] and
(E(f )) E((f )).
Proof. Since f : [a, b] [c, d], we have
Z b
Z b
Z
c=
cP (x)dx
f (x)P (x)
a

dP (x) = d,

Rb
so a f (x)P (x)dx = E(f ) [c, d] and thus (E(f )) is defined. Now put x0 = E(f )
and let s(x) = mx + B be a supporting line for the convex function at x0 , so
s(x) (x) for all x [c, d] and s(x0 ) = (x0 ). Now
Z b
Z b
E((f )) =
(f (x))P (x)dx
(mf (x) + B)P (x)dx
a

Z
=m

f (x)P (x)dx + B = mE(f ) + B = mx0 + b = (x0 ) = (E(f )).

a


Theorem 9.11 (H
olders Integral Inequality). Let p, q R>0 be such that
1
+ q = 1. Let f, g : [a, b] R be Riemann integrable functions. Then
Z

! q1
! p1 Z
Z b
 b 
b


(61)
f g
|f |p
|g|q
.

 a

a
a
1
p

Rb
Proof. Step 1: Suppose a |f |p = 0. Let Mf and Mg be upper bound for f
and g on [a, b]. By Lemma 8.29, the set of x [a, b] such that |f |p >  has content
zero: thus, for every > 0 there is a finite collection of subintervals of [a, b], of total
length at most , such that on the complement of those subintervals |f |p . On
1
this complement, |f g|  p Mg and thus the sum of the integrals of |f g| over the

196

9. INTEGRAL MISCELLANY
1

complement is at most (b a) p Mg . The sum of the Riemann integrals over the
subintervals of total length of |f g| is at most Mf Mg , so altogether we get
Z b
Z b
1
|
f g|
|f g| (b a) p Mg + Mf Mg .
a

Rb
Since all the other quantities are fixed and , are arbitrary, we deduce | a f g| = 0.
Thus both sides of (61) are zero in this case and the inequality holds. A similar
Rb
argument works if a |g|q = 0. Henceforth we may suppose
Z b
Z b
If =
|f |p > 0, Ig =
|g|q > 0.
a

1
p

1
q

Step 2: Put f = f /If and g = g/Ig . Then by Youngs Inequality,

!
Z b
Z b
Z b
q
|p
|
g
|
1 1
|
f
+
= + = 1.
|
fg|
|f||
g|
p
q
p
q
a
a
a
1

Multiplying through by Ifp Igq =

R

b
a

|f |p

 p1 R
b
a

|g|q

 q1

gives the desired result.

Remark: Admittedly Step 1 of the above proof is rather technical. To an extent

this points to the finickiness of the class of Riemann integrable functions. The student who continues on will learn the corresponding result for the Lebesgue integral,
in which the proof of this step is immediate. Alternately, the student may wish
to restrict to continuous f and g, since a non-negative continuous function f with
Rb
f = 0 must be identically zero.
a
The case p = q = 2 of Theorem 9.11 is important enough to be restated as a
result in its own right.
Corollary 9.12 (Cauchy-Schwarz Integral Inequality). Let f, g : [a, b] R
be Riemann integrable functions. Then
Z b !2 Z b Z b
fg

f2
g2 .
a

5. The Riemann-Lebesgue Lemma

Theorem 9.13. (Riemann-Lebesgue Lemma) Let f : [a, b] R be Riemann
integrable. Then:
Rb
a) lim a f (x) cos(x)dx = 0.
Rb
b) lim a f (x) sin(x)dx = 0.
Proof. As the reader surely expects, the arguments for parts a) and b) are
virtually identical, so we will establish part a) and leave part b) as an exercise.
Rb
Here is the idea of the proof: if f (x) C is a constant function, then a C cos(x)
oscillates increasingly rapidly as increases. If the length of the interval a b were
a precise multiple of the period 2
then we get several complete sine waves and
the positive and negative area cancel out exactly, making the integral exactly zero.
As increases the remaining incomplete sine wave is a bounded function living
on a smaller and smaller subinterval, so its contribution to the integral goes to zero.

5. THE RIEMANN-LEBESGUE LEMMA

197

A similar argument holds for step functions, which become constant functions
when split up into appropriate subintervals. And now the key: the integrable functions are precisely those which can be well approximated by step functions in the
sense that the integral of the difference can be made arbitrarily small.
Okay, lets see the details: fix  > 0. By Darbouxs Criterion, there is a partition
P = {a = x0 < x1 < . . . < xn1 < xn = b} of [a, b] such that
Z b

0 L(f, P)
f< .
2
a
Let g be the step function which is constantly equal to mi = inf(f, [xi , xi+1 ]) on
Rb
the subinterval [xi , xi+1 ) of [a, b], so g f and a g = L(f, P), so
Z b

0
(f g) .
2
a
Now
Z
 Z
Z

 b

 b

b




f (x) cos(x)dx
|f (x) g(x)|| cos(x)|dx + 
g(x) cos(x)dx

 a



a
a
(62)





n1 Z
n1


  X xi+1
  X mi


+
mi cos(x)dx + 
(sin(xi+1 ) sin(xi )) .
 2 

2  i=0 xi

i=0

Here we have a lot of expressions of the form | sin(A) sin(B)| for which an obvious
upper bound is 2. Using this, the last expression in (62) is at most
Pn1
2 i=0 |mi |

+
.
2

But this inequality holds for any > 0, so taking sufficiently large we can make
Rb
the last term at most 2 and thus | a f (x) cos(x)dx| < .


CHAPTER 10

Infinite Sequences
Let X be a set. An infinite sequence in X is given by a function x : Z+ X.
Less formally but entirely equivalently, we are getting an ordered infinite list of
elements of X: x1 , x2 , x3 , . . . , xn , . . .. Note that the function is not required to be
injective: i.e., we may have xi = xj for i 6= j. In fact, a simple but important example of a sequence is a constant sequence, in which we fix some element x X
and take xn = x for all n.
The notion of an infinite sequence in a general set X really is natural and important
throughout mathematics: for instance, if X = {0, 1} then we are considering infinite sequences of binary numbers, a concept which comes up naturally in computer
science and probability theory, among other places. But here we will focus on real
infinite sequences xbullet : Z+ R. In place of x , we will write {xn }
n=1 or
even, by a slight abuse of notation, xn .
We say that an infinite sequence an converges to a real number L if for all  > 0
there exists a positive integer N such that for all n N , we have |an L| < .
A sequence is said to be convergent if it converges to some L R and otherwise divergent. Further, we say a sequence an diverges to infinity and write
limn an = or an if for all M R there exists N Z+ such that
n N = an > M . Finally, we define divergence to negative infinity: I leave it
to you to write out the definition.
This concept is strongly reminiscent of that of the limit of a function f : [1, ) R
as x approaches infinity. In fact, it is more than reminiscent: there is a direct connection. If limx f (x) = L, then if we form the sequence xn = f (n), then it
follows that limn xn = L. If xn = f (x) for a function f which is continuous or
better, differentiable then the methods of calculus can often be brought to bear
to analyze the limiting behavior of xn .
Given a sequence {xn }, we say that a function f : [1, ) R interpolates f
if f (n) = xn for all n Z+ .
Example: Supose xn =

log n
n .

Then f (x) =

log x
x

interpolates the sequence, and

1
log x
LH
1
=
= lim x =
= 0.
x x
x 1

lim

It follows that xn 0.
Exercise: Let {an }
n=1 be a real sequence. Define f : [1, ) R as follows:
for x [n, n + 1), f (x) = (n + 1 x)an + (x n)an+1 .
199

200

10. INFINITE SEQUENCES

a) Take, for example, an = n2 , and sketch the graph of f .

b) Show that the sequence limx f (x) exists iff the sequence {an } is convergent,
and if so the limits are the same.
The previous exercise shows that in principle every infinite sequence {an } can be
interpolated by a continuous function. The given f is piecewise linear but generally not differentiable at integer values. However, with only a little more trouble we
could round off the corners and find a differentiable function f which interpolates
{an }. But in practice this is only useful if the interpolating function f is simple
enough to have a known limiting behavior at infinity. Many sequences which come
up in practice cannot be interpolated in a useful way.
Example (Fibonacci Sequence)
Example (Newtons method sequences).
In fact we will be most interested in sequences of finite sums. For instance, let
n
X
1
1
1
= 1 + + ... + ,
an =
i
2
n
i=1
and let

n
X
1
1
1
= 1 + 2 + ... + 2.
bn =
2
i
2
n
i=1

What is the limiting behavior of an and bn ? In fact it turns out that an and
2
bn 6 : whatever is happening here is rather clearly beyond the tools we have at
the moment! So we will need to develop new tools.
1. Summation by Parts
Lemma 10.1. (Summation by Parts) Let {an } and {bn } be two sequences. Then
for all m n we have
n
n
X
X
ak (bk+1 bk ) = (an+1 bn+1 am bm ) +
(ak+1 ak )bk+1 .
k=m

k=m

Proof.
n
X
ak (bk+1 bk ) = am bm+1 + . . . + an bn+1 (am bm + . . . + an bn )
k=m

= an bn+1 am bm ((am+1 am )bm+1 + . . . + (an an1 )bn )

n1
X
= an bn+1 am bn
(ak+1 ak )bk+1
k=m

= an bn+1 am bn + (an+1 an )bn+1

n
X

(ak+1 ak )bk+1

k=m

= an+1 bn+1 am bm

n
X

(ak+1 ak )bk+1 .

k=m

2. EASY FACTS

201

The proof of Lemma 10.1 could hardly be more elementary or straightforward: literally all we are doing is regrouping some finite sums. Nevertheless the human
mind strives for understanding and rebels against its absence: without any further
explanation, summation by parts would seem (to me at least) very mysterious.
The point is that Lemma 10.1 is a discrete analogue of integration by parts:
Z b
Z b
f 0 g.
f g 0 = f (b)g(b) f (a)g(a)
a

Just as integration by parts is a fundamental tool in the study of integrals and,

after taking limits, improper integrals summation by parts is a fundamental tool
in the study of finite sums and, after taking limits, infinite series.
Whereas for integration by parts there really is just the one identity, recalled
above, for summation by parts, there are several alternate formulas and special
cases, one of which is given in the following exercise.
So really one should rememP
ber the idea starting with a sum of the form n an bn , there are several identities
involving expressions in which we have discretely differentiated {an } and discretely
integrated {bn }, or vice versa rather than any one specific formula.

+
Exercise:
Let {a
Pn
Pnn}n=1 and {bn }n=1 be sequences, and for N Z , put An =
i=1 an , Bn =
i=1 bn , A0 = B0 = 0. Show that
N
X

(63)

an bn =

N
1
X

An (bn bn+1 ) + AN bN .

n=1

n=1

Proposition 10.2. (Abels Lemma) Let {an }

n=1 be a real sequence with the
following property: there is an M > 0 such that |a1 + . . . + aN | M for all positive
integers N . Let {bn }
n=1Pbe a real sequence such that b1 b2 . . . bn . . . 0.
N
Then, for all N Z+ , | n=1 an bn | b1 M .
Proof. Using (63), we have
|

N
X
n=1

ab bn | = |

N
1
X

An (bn bn+1 ) + AN bN |

n=1

N
1
X

(bn bn+1 )|AN | + bN |AN | = b1 |AN | b1 M.

n=1


We will put Abels Lemma to use...but only much later on. For now though you
might try to prove it without using the summation by parts variant (63): by doing
so, youll probably gain some appreciation that these formulas, though in some
sense trivial, can be used in distinctly nontrivial ways.
2. Easy Facts
The following result collects some easy facts about convergence of infinite sequences.
Theorem 10.3. Let {an }, {bn }, {cn } be real infinite sequences.
a) If an = C for all n a constant sequence then an C.
b) The limit of a convergent sequence is unique: if for L1 , L2 R we have an L1

202

10. INFINITE SEQUENCES

and an L2 , then L1 = L2 .
c) If an L and bn M then:
(i) For all C R, Can CL.
(ii) an + bn L + M .
(iii) an bn LM .
L
(iv) If M 6= 0, abnn M
.
d) If an bn for all n, an L and bn M , then L M .
e) If a b are such that an [a, b] for all n and an L, then L [a, b].
f ) (Three Sequence Principle) Suppose cn = an + bn . Then it is not possible for
exactly two of the three sequences {an }, {bn }, {cn } to be convergent: if any two are
convergent, then so is the third.
g) Suppose there exists N Z+ such that bn = an for all n N . Then for any
L [, ], an L bn L.
Most of these facts are qiute familiar, and the ones that may not be are routine.
In fact, every part of Theorem 10.3 holds verbatim for functions of a continuous
variable approaching infinity. Hence one method of proof would be to establish
these for functions or maintain that we have known these facts for a long time
and then apply the Sequence Interpolation Theorem. But be honest with yourself:
for each part of Theorem 10.3 for which you have an iota of doubt as to how to
prove, please take some time right now to write out a careful proof.
We say that a sequence {an } is eventually constant if there is C R and N Z+
such that an = C for all n N . It is easy to see that if such a C exists then it
is unique, and we call such a C the eventual value of the sequence. Of course
an eventually constant sequence converges to its eventual value e.g. by applying
parts a) and g) of Theorem 10.3, but really this is almost obvous in any event.
Proposition 10.4. Let {an } be an infinite sequence with values in the integers
Z. Then an is convergent iff it is eventually constant.
Proof. As above, it is clear that an eventually constant sequence is convergent.
Conversely, suppose an L R. First we claim that L Z. If not, the distance
from L to the nearest integer is a positive number, say . But since an L, there
exists N Z+ such that for all n N , |an L| < . But the interval (L , L + )
contains no integers: contradiction.
Now take  = 1 in the definition of convergence: there is N Z+ such that for
n N , we have |an L| < 1, and since an and L are both integers this implies
an = L. Thus the sequence is eventually constant with eventual value L.

Proposition 10.4 goes a long way towards explaining why we have described a function from Z+ to R as semi -discrete. A function from Z+ to Z+ is fully discrete,
and thus the limiting behavior of such functions is very limited.
Exercise: A subset S R is discrete if for all x S, there is  > 0 such that the
only element of S which lies in (x , x + ) is x.
a) Which of the following subsets of R are discrete?
(i) A finite set.
(ii) The integers Z.
(iii) The rational numbers Q.
(iv) The set { n1 | n Z+ } of reciprocals of positive integers.

4. MONOTONE SEQUENCES

203

S
(v) The set { n1 | n Z+ } {0}.
b) For a subset S R, show that the following are equivalent:
(i) S is discrete.
(ii) Every convergent sequence {an } with an S for all n is eventually constant.
It is often convenient to consider sequences whose initial term is something other
than 1. It is certainly no problem to entertain sequences starting at any integer
N0 : informally, they look like
aN0 , aN0 +1 , . . . .
Formally, instead of a function from Z+ to R, we have a function from {n Z | n
N0 } to R. This mild generalization changes nothing we have said so far or will say
later. We leave it to the reader to make her own peace with this.
3. Characterizing Continuity
Theorem 10.5. Let I be an interval, let c be an interior point of I, and let
f : I R be a function. The following are equivalent:
(i) f is continuous at c.
(ii) For every infinite sequence an c, we have f (an ) f (c).
Proof. (i) = (ii): This argument is very similar to the (easy!) proof that
a composition of continuous functions is continuous. Namely, fix  > 0. Since f
is continuous at c, there is > 0 such that |x c| < = |f (x) f (c)| < .
Moreover, since an c, there exists N Z+ such that n N = |an c| < .
So if n N we have |an c| < and thus |f (an ) f (c)| < .
(ii) = (i): We prove the contrapositive: that is, we will suppose that f is not
continuous at c and find a sequence an c but such that f (an ) does not converge
to f (c). If f is not continuous at c then there exists  > 0 such that for all > 0,
there is a with |a c| < and |f (a ) f (c)| . In particular, for n Z+ we
may take = n1 and choose an with |an c| < n1 and |f (an ) f (c)| , and then
indeed we have an c but f (an ) does not approach f (c).

Spivak gives a slightly more general result, the proof of which (essentially the same
as the one given above) we leave as an exercise.
Theorem 10.6. Let I be an interval, c an interior point of I, and let f :
I \ {c} R be a function such that limxc f (x) = L. Let {an }
n=1 be a real
sequence such that for all n Z+ , an I \ {c} and limn an = c. Then
lim f (an ) = L.

Exercise: Prove Theorem 10.6.

4. Monotone Sequences
One commonality between sequences a : Z+ R and non-discrete functions
f : I R is that since indeed both the domain and codomain are subsets of
R we have a natural order structure and it makes sense to consider functions
which systematically preserve or systematically reverse! this order structure.
Thus the following definitions will be quite familiar.
A sequence {an } is weakly increasing if for all m, n Z+ , m n = am an .

204

10. INFINITE SEQUENCES

A sequence {an } is increasing if for all m, n Z+ , m < n = am < an .

A sequence {an } is weakly decreasing if for all m, n Z+ , m n = am an .
A sequence {an } is decreasing if for all m, n Z+ , m < n = am > an .
A sequence is monotone if it is weakly increasing or weakly decreasing.
Remark: As before, we need to counsel the reader that the terminology here is
not agreed upon: for instance, many use the term strictly increasing for sequences with m < n = am < an . If you are ever in doubt whether someone
means m n = am an or m < n = am < an , there is a simple remedy: ask!
Exercise: Let {an } be a sequence.
a) Show that {an } is increasing iff an < an+1 for all n Z+ .
b) Formulate (and, if you feel it is worth your time, prove) analogues of part a)
with increasing replaced by each of weakly increasing, decreasing, weakly decreasing.
Exercise: For a real sequence {an }, show that the following are equivalent:
(i) There is n Z+ such that either an < an+1 > an+2 (-configuration) or
an > an+1 < an+2 (V-configuration).
(ii) The sequence {an } is not monotone.
Lemma 10.7. For a real sequence {an }, the following are equivalent:
(i) {an } is both weakly increasing and weakly decreasing.
(ii) {an } is constant.
Exercise: Prove it.
Lemma 10.8. (Reflection Principle) Let {an } be a real sequence. Then:
a) {an } is decreasing iff {an } is increasing.
b) {an } is weakly increasing iff {an } is weakly increasing.
Exercise: Prove it.
Just as for functions of a continuous variable, Lemma 10.8 implies that whatever
we can establish for increasing / weakly increasing sequences will carry over immediately to decreasing / weakly decreasing sequences (and conversely, of course).
What a monotone sequence lacks is oscillation, and as for functions of a continuous
variable, this implies a much simpler limiting behavior. Especially, if a : Z+ R
is weakly increasing, there is really only one way for limn an to fail to exist.
To speak of this it is convenient to make another definition: since a sequence an
is really a function Z+ R, it makes sense to consider its image, i.e., the set of
all real numbers of the form an for some n Z+ . Strangely, there is not really a
standard name for this: I vaguely recall it being called the trace of the sequence,
but the term trace has many other, unrelated meanings in mathematics so this
is really not optimal. We avoid the problem by giving the concept a very clunky
name: we define the term set of {an } to be A = {an R | n Z+ }.

4. MONOTONE SEQUENCES

205

We are therefore able to speak of bounded sequences just as for bounded functions: i.e., in terms of the image...um, I mean the term set.
A sequence a : Z+ R is bounded above if its term set is bounded above:
that is, if there exists M R such that an M for all n Z+ . Otherwise we say
the sequence is unbounded above. Similarly, we say a is bounded below if its
term set is bounded below: that is, if there exists m R such that m an for all
n Z+ . Otherewise we say the sequence is unbounded below. Finally, a sequenc
is bounded if it is both bounded above and bounded below, and a sequence is
unbounded if it is not bounded.
Proposition 10.9. Let {an }
n=1 be a weakly increasing sequence.
a) If the sequence converges to L R, then L is the least upper bound of the term
set A = {an | n Z+ }.
b) Conversely, if the term set A has an upper bound L < , then an L.
Proof. a) First we claim L = limn an is an upper bound for the term
set A. Indeed, suppose not: then there is N Z+ with L < aN . But since the
sequence is weakly increasing, this implies that for all n N , L < aN an . Thus
if we take  = aN L, then for no n N do we have |an L| < , contradicting
our assumption that an L. Second we claim L is the least upper bound. Indeed,
suppose not: then there is L0 such that for all n Z+ , an L0 < L. Let  = L L0 .
For no n do we have |an L| < , contradicting our asumption that an L.
b) Let  > 0. We need to show that for all but finitely many n Z+ we have
 < L an < . Since L is the least upper bound of A, in particular L an for
all n Z+ , so L an 0 > . Next suppose that there are infinitely many terms
an with L an , or L an + . But if this inequality holds for ifninitely many
terms of the sequence, then because an is increasing, it holds for all terms of the
sequence, and this implies that L  an for all n, so that L  is a smaller upper
bound for A than L, contradiction.

Remark: In the previous result we have not used the completeness property of
R, and thus it holds for sequences with values in the rationals Q (and where by
converges we mean converges to a rational number !) or really in any ordered field.
By combining this with the least upper bound axiom, we get a much stronger result.
Theorem 10.10. Let {an }
n=1 be a weakly increasing real sequence. Let L
(, ] be the least upper bound of the term set of A. Then an L.
This is so important as to be worth spelling out very carefully. We get:
Theorem 10.11. (Monotone Sequence Theorem) a) Every bounded monotone
real sequence is convergent. More precisely:
b) Let {an } be weakly increasing. Then if {an } is bounded above, it converges to its
least upper bound, and if it is unbounded above it diverges to .
c) Let {an } be weakly decreasing. Then if {an } is bounded below, it converges to its
greatest lower bound, and it is unbounded below it diverges to .
In fact, in proving the Monotone Sequence Theorem we did not just invoke the
completeness of the real field: we used its full force, in the following sense.

206

10. INFINITE SEQUENCES

Theorem 10.12. Let F be an ordered field in which every bounded monotone

sequence converges. Then F is Dedekind complete: every nonempty bounded above
subset has a least upper bound.
Proof. Step 1: We claim that F is Archimedean.
proof of claim: Suppose not: then the sequence xn = n is increasing and
bounded above. Suppose that it were convergent, say to L F . By Proposition
10.9, L is the least upper bound of Z+ . But this is absurd: if n L for all n Z+
then n + 1 L for all n Z+ and thus n L 1 for all n Z+ , so L 1 is a
smaller upper bound for Z+ .
Step 2: Let S R be nonempty and bounded above by M0 .
claim For all n Z+ , there exists yn S such that for any x S, x yn + n1 .
Proof of claim: Indeed, first choose any element z1 of S. If for all x S, x z1 + n1 ,
then we may put yn = z1 . Otherwise there exists z2 S with z2 > z1 + n1 . If for
all x S, x z2 + n1 , then we may put yn = z2 . Otherwise, there exists z3 S
with z3 > z2 + n1 . If this process continues infinitely, we get a sequence with
zk z1 + k1
n . But by Step 1, F is Archimedean, so that for sufficiently large k,
zk > M , contradiction. Therefore the process musts terminate and we may take
yn = zk for sufficiently large k.
+
Now we define a sequence of upper bounds {Mn }
n=1 of S as follows: for all n Z ,
1
Mn = min(Mn1 , yn + n ). {Mn } is decreasing and bounded below by any element
of S, so by hypothesis it converges, say to M , and by the reflected version of
Proposition 10.9, M is the infimum of the set {Mn }. Moreover M must be the
supremum of S, since again by the Archimedean nature of the order, for any m < M ,
for sufficiently large n we have m + n1 < M Mn yn + n1 and thus m < yn . 
5. Subsequences
A real infinite sequence {an } is, informally, an orderest list of real numbers:
a1 , a2 , . . . , an , . . .
A subsequence of {an }, is again informally, for a little while; a precise definition
will be coming soon obtained by selecting some infinitely many terms of the
sequence, e.g
a1 , a3 , a6 , a100 , a703 , . . . .
(There is not meant to be a clearly evident pattern as to which terms of the sequence we are choosing: the point is that there does not need to be.) To learn more
about subsequences and what they are good for, let us look at some key examples.
Example: Let an = (1)n , so the sequence is
(64)

1, 1, 1, 1, . . .

By selecting all the odd-numbered terms we get one subsequence:

(65)

1, 1, 1, 1, . . .

Similarly, by selecting all the even-numberd terms we get another subsequence:

(66)

1, 1, 1, 1, . . . .

There are other choices many other choices. In fact, a real sequence can be obtained as a subsequence of {an } iff it takes values in {1}.

5. SUBSEQUENCES

207

But note that something very interesting happened in the passage from our original
sequence to each of the first two subsequences. The original sequence (64) is not
convergent, due to oscillation. However, the subsequence (65) is constant, hence
convergent to it constant value 1. Similarly, the subsequence (66) converges to
its constant value 1.
Lets recap: we began with a sequence (66) which did not converge due to oscillation. However, by choosing appropriate subsequences we were able to remove the
oscillation, resulting in this case, at least in a convergent subsequence. (There
are also lots of subsequences which are inappropriate for this purpose.)
Example: Let an = n, so the sequence is
(67)

1, 2, 3, 4, . . .

Here are some subsequences we can get:

(68)

1, 3, 5, 7, . . .

(69)

2, 4, 6, 8, . . .

(70)

1, 4, 9, 16, . . .

(71)

1, 2, 4, 8, . . .

And so forth. Indeed, the subsequences of (67) are precisely the increasing sequences with values in the positive integers. Note moreover that our sequence (67)
fails to converge, but not due to oscillation. It is an increasing sequence which is
not unbounded above, and thus it diverges to infinity. For that matter, so do the
subsequences (68), (69), (70), (71), and a little thought suggests that every subsequence will have this property. Thus, passing to a subsequence can cure divergence
due to oscillation, but not divergence to infinity.
Example (subsequences of a convergent sequence):
We are now well-prepared for the formal definition. In fact, we practically saw
it in the example above. Given a real sequence {an }, we view it as a function
a : Z R, n 7 an . To obtain a subsequence we choose an increasing sequence
n : Z+ Z+ , k 7 nk and form the composite function
a n : Z+ R, k 7 ank .
Less formally, we choose an increasing list n1 < n2 < . . . < nk of positive integers
and use this to tell us which terms of the sequence to take, getting
an1 , an2 , an3 , . . . , ank , . . . .
Lets formalize these observations about what passing to subsequences does for us.
Exercise: Let n : Z+ Z+ be increasing. Show that for all k Z+ , nk k.
Proposition 10.13. Let {an } be a real sequence, L [, ], and suppose
that an L. Then for any subsequence {ank }
k=1 , we have ank L.

208

10. INFINITE SEQUENCES

Proof. Case 1: Suppose that L R. Since an L, for each  > 0 there

exists N = N () Z+ such that for all n N , |an L| < . Then, by Exercise
X.X, for all k N we have nk N and thus |ank L| < .
Case 2: Suppose L = . Since an , for each M R there exists N =
N (M ) Z+ such that for all n N , an > M . If k N , then nk k N and
thus ank > M .
Case 3: We leave the case of L = to the reader as a (not very challenging)
exercise in modifying the argument of Case 2.

Proposition 10.14. Every subsequence of a monotone sequence is monotone.
More precisely:
a) If {an } is weakly increasing, then every subsequence ank is weakly increasing.
b) If {an } is increasing, then every subsequence ank is increasing.
c) If {an } is weakly decreasing, then every subsequence ank is weakly decreasing.
d) If {an } is decreasing, then every subsequence ank is decreasing.
e) If {an }, is constant, then every subsequence ank is constant.
Proof. a) If k1 k2 , then by definition of a subsequence nk1 nk2 , and then
by definition of weakly increasing, ank1 ank2 .
b),c),d),e) These may safely be left to the reader.

Corollary 10.15. For a montone sequence {an }, the following are equivalent:
(i) {an } is convergent.
(ii) {an } is bounded.
(iii) Every subsequence of {an } is convergent.
(iv) At least one subsequence of {an } is convergent.
6. The Bolzano-Weierstrass Theorem For Sequences
6.1. The Rising Sun Lemma.
I learned of the following result from Mr. Evangelos Kobotis in late 1994, in my
first quarter of college at the University of Chicago. Because of its elegance, generality and usefulness, it has stayed with me through my entire adult mathematical
career. It seems that this argument goes back to a short note of Newman and
Parsons [NP88].
Lemma 10.16. Every infinite sequence has a monotone subsequence.
Proof. Let us say that m N is a peak of the sequence {an } if for all n > m,
an < am . Suppose first that there are infinitely many peaks. Then any sequence
of peaks forms a strictly decreasing subsequence, hence we have found a monotone
subsequence. So suppose on the contrary that there are only finitely many peaks,
and let N N be such that there are no peaks n N . Since n1 = N is not a
peak, there exists n2 > n1 with an2 an1 . Similarly, since n2 is not a peak, there
exists n3 > n2 with an3 an2 . Continuing in this way we construct an infinite (not
necessarily strictly) increasing subsequence an1 , an2 , . . . , ank , . . .. Done!

Exercise: Show that every infinite sequence admits a subsequence which is (i) increasing, (ii) decreasing, or (iii) constant.
Remark: Although it is nothing to do with our development of infinite sequences, we
cannot resist mentioning the following finite analogue of the Rising Sun Lemma.

6. THE BOLZANO-WEIERSTRASS THEOREM FOR SEQUENCES

209

Theorem 10.17. (Erd

os-Szekeres [ES35]) Let r, s Z+ .
a) Consider a finite real sequence x1 , . . . , x(r1)(s1)+1 of length (r 1)(s 1) + 1.
(i) There is a weakly increasing subsequence of length r, or
(ii) There is a weakly decreasing subsequence of length s.
b) The number (r 1)(s 1) + 1 is best possible, in the sense that there are real
sequences of length (r 1)(s 1) without either a weakly increasing subsequence of
length r or a weakly decreasing subsequence of length s.
Proof. a) ([Se59], [Er94]) Seeking a contradiction, we suppose every weakly
increasing subsequence has length at most r 1 and every weakly decreasing subsequence has length at most s1. We define a function F : {1, . . . , (r1)(s1)+1}
{1, . . . , r 1} {1, . . . , s 1} by F (j) = (ij , dj ), where ij is the length of the
longest weakly increasing subsequence beginning with xj and dj is the length of
the longest weakly decreasing subsequence beginning with xj . Since the domain of
F has (r 1)(s1)+1 elements and the codomain has (r 1)(s1) elements, there
must be j < k such that F (j) = F (k): thus ij = ik and dj = dk . But if xj xk
then adding xj to the beginning of the longest weakly increasing sequence starting
with xk gives a longer weakly increasing sequence starting at xj , a contradiction;
similarly, if xj xk , then adding xj to the beginning of the longest weakly decreasing sequence starting with xk gives a longer weakly decreasing sequence starting at
xj , a contradiction. b) Consider the length (r 1)(s 1) sequence
r, r 1, . . . , 1, 2r, 2r 1, . . . , r +1, 3r, 3r 1, . . . , 2r +1, . . . , sr, sr 1, . . . , (s1)r +1.

6.2. Bolzano-Weierstrass for Sequences.
Theorem 10.18. (Bolzano-Weierstrass) Every bounded real sequence admits
a convergent subsequence.
Proof. Let {an } be a sequence with |an | M for all n Z+ . By the Rising
Sun Lemma, there is a monotone subsequence ank , which of course is also bounded:
|ank | M for all k Z+ . By Corollary 10.15, the subsequence ank is convergent.

Exercise: Show that for an ordered field F , the following are equivalent:
(i) Every bounded monotone sequence converges.1
(ii) Every bounded sequence admits a convergent subsequence.
6.3. Supplements to Bolzano-Weierstrass.
Theorem 10.19. a) A real sequence which is unbounded above admits a subsequence diverging to .
b) A real sequence which is unbounded below admits a subsequence diverging to .
Proof. We will prove part a) and leave the task of adapting the argument to
prove part b) to the reader. Let {xn } be a real sequence which is unbounded above.
Then for every M R, there exists at least one n such that xn M . Let n1 be the
least positive integer such that xn1 > 1. Let n2 be the least positive integer such
that xn2 > max(xn1 , 2). And so forth: having defined nk , let nk+1 be the least
positive integer such that xnk+1 > max(xnk , k + 1). Then limk xnk = .

1Recall that we have already shown that this is equivalent to Dedekind completeness.

210

10. INFINITE SEQUENCES

6.4. Applications of Bolzano-Weierstrass.

We will now give two quite different proofs of two of the three Interval Theorems.
Theorem 10.20. (Extreme Value Theorem Again) Let f : [a, b] R be continuous. Then f is bounded and attains its minimum and maximum values.
Proof. Seeking a contradiction we suppose f is unbounded : then, for each
n Z+ , there exists xn [a, b] with |f (xn )| > n. The sequence {xn } takes values in [a, b] so is bounded, so by Bolzano-Weierstrass there is a subsequence xnk
such that xnk L R. Now, on the one hand, since f is continuous we have by
Theorem 10.5 that f (xnk ) f (L) R. On the other hand, we have for all k that
|f (xnk )| > k, so the sequence f (xnk ) is divergent: contradiction!
We do not have a new argument for the attainment of the minimum and maximum values, but the original argument was sufficiently short and sweet that we
dont mind repeating it here: let M be the supremum of the set of f (x) for x [a, b],
so by what we just proved, M < . If M is not attained then the function f (x)M
1
is continuous and nowhere zero, so the function g : [a, b] R by g(x) = f (x)M
is continuous on [a, b], hence bounded. But this is impossible: by definition of M ,
f (x) takes values arbitrarily close to M , so f (x) M takes values arbitrarily close
to zero and thus |g(x)| take arbitrarily large values. By reflection, f attains its
minimum as well.

Theorem 10.21. (Uniform Continuity Theorem Again) Let f : [a, b] R be
continuous. Then f is uniformly continuous.
Proof. Suppose not: then there exists  > 0 such that for all n Z+ we have
xn , yn [a, b] with |xn yn | < n1 but |f (xn ) f (yn )| . By Bolzano-Weierstrass,
after passing to a subsequence we may assume xnk is convergent, say to L, Since
xnk ynk 0, by the Three Sequence Principle ynk is also convergent to L. But
then f (xnk ) f (L) and f (ynk ) f (L), and since the sequences f (xnk ) and f (ynk )
have the same limit, we can make |f (xnk )f (ynk )| |f (xnk )L(f (ynk ))L)|
|f (xnk ) L| + |f (ynk ) L| arbitrarily small by taking n sufficiently large. In
particular, for sufficiently large n we have |f (xnk ) f (ynk )| < : contradiction! 
Remark: Although I am very fond of the Real Induction proof of the Uniform Continuity Theorem, I must admit this argument seems shorter and easier. Very often
in mathematics if one takes the time to develop additional, seemingly unrelated
technology in this case, the theory of infinite sequences one is richly rewarded
with the ability to come up with shorter, simpler (but more high tech) proofs.
6.5. Bolzano-Weierstrass for Subsets Revisited.
Recall that earlier we proved a different result which we also called the BolzanoWeierstrass Theorem: it concerned the notion of a limit point of a subset X R: a
point a R is a limit point for X if for all  > 0, there is x R with 0 < |ax| < .
It was an easy observation that any finite subset of R has no limit points. We proved
that the converse holds for bounded sets:
Theorem 10.22. (Bolzano-Weierstrass for Subsets) An infinite subset X of a
closed, bounded interval [a, b] has a limit point.

7. PARTIAL LIMITS; LIMITS SUPERIOR AND INFERIOR

211

Our task now is to explain why we have two different results called the BolzanoWeierstrass Theorem. In fact they are really equivalent results, which means
roughly that is much easier to deduce each from the other than it is to prove
either one. Indeed:
Assume the Bolzano-Weierstrass Theorem for sequences, and let X be an infinite subset of [a, b]. Since X is infinite, we can choose a sequence {xn }
n=1 of
distinct elements of X. Since xn X [a, b], we have a xn b for all n;
in particular, {xn } is bounded, so by Bolzano-Weierstrass for sequences there is a
subsequence xnk converging to some L [a, b]. We claim that L is a limit point for
X: indeed, for any  > 0, there is K Z+ such that for all k K, |xnk L| < :
since the terms are distinct, at most one of them is equal to L and thus the interval
(L , L + ) contains infinitely many elements of X.
Assume the Bolzano-Weierstrass Theorem for subsets, and let {xn } be a bounded
sequence: thus there are a, b R with a xn b for all n Z+ . Let X = {xn | n
Z+ } be the term set of the sequence. If X is finite, then the sequence has a constant (hence convergent) subsequence. Otherwise X is infinite and we may apply
Bolzano-Weierstrass for subsets to get a limit point L of X. This implies: there is
n1 Z+ such that |xn1 L| < 1; having chosen such an n1 , there is n2 Z+ such
that n2 > n1 and |xn2 L| < 21 : continuing in this way we build a subsequence
{xnk } such that for all k Z+ , |xnk L| < k1 , and thus xnk L.
7. Partial Limits; Limits Superior and Inferior
7.1. Partial Limits.
For a real sequence {an }, we say that an extended real number L [, ]
is a partial limit of {an } if there exists a subsequence ank such that ank L.
Lemma 10.23. Let {an } be a real sequence. Suppose that L is a partial limit of
some subsequence of {an }. Then L is also a partial limit of {an }.
Exercise: Prove Lemma 10.23. (Hint: this comes down to the fact that a subsequence of a subsequence is itself a subsequence.)
Theorem 10.24. Let {an } be a real sequence.
a) {an } has at least one partial limit L [, ].
b) The sequence {an } is convergent iff it has exactly one partial limit L and L is
finite, i.e., L 6= .
c) an iff is the only partial limit.
d) an iff is the only partial limit.
Proof. a) If {an } is bounded, then by Bolzano-Weierstrass there is a finite
partial limit L. If {an } is unbounded above, then by Theorem 10.19a), is a
partial limit. It {an } is unbounded below, then by Theorem 10.19b) is a
partial limit. Every sequence is either bounded, unbounded above or unbounded
below (and the last two are not mutually exclusive), so there is always at least one
partial limit.
b) Suppose that L R is the unique partial limit of {an }. We wish to show that
an L. First observe that by Theorem 10.19, {an } must be bounded above and
below, for otherwise it would have an infinite partial limit. So choose M R such

212

10. INFINITE SEQUENCES

that |an | M for all n.

Now suppose that an does not converge to L: then there exists  > 0 such that
it is not the case that there exists N N such that |an L| <  for all n N .
What this means is that there are infinitely many values of n such that |an L| .
Moreover, since |an L|  means either M an L  or L +  an M ,
there must in fact be in infinite subset S N such that either for all n S we have
an [M, L ] or for all n S we have an [L + , M ].
Let us treat the former case. The reader who understands the argument will
have no trouble adapting to the latter case. Writing the elements of S in increasing
order as n1 , n2 , . . . , nk , we have shown that there exists a subsequence {ank } all of
whose terms lie in the closed interval [M, L ]. Applying Bolzano-Weierstrass
to this subsequence, we get a subsubsequence (!) ank` which converges to some L0 .
We note right away that a subsubsequence of an is also a subsequence of an : we
still have an infinite subset of N whose elements are being taken in increasing order.
Moreover, since every term of ank` is bounded above by L , its limit L0 must
satisfy L0 L . But then L0 6= L so the sequence has a second partial limit L0 :
contradiction.
c) Suppose an . Then also every subsequence diverges to +, so + is
a partial limit and there are no other partial limits. We will prove the converse
via its contrapositive (the inverse): suppose that an does not diverge to . Then
there exists M R and infinitely many n Z+ such that an M , and from this
it follows that there is a subsequence {ank } which is bounded above by M . This
subsequence does not have + as a partial limit, hence by part a) it has some
partial limit L < . By Lemma 10.23, L is also a partial limit of the original
sequence, so it is not the case that + is the only partial limit of {an }.
d) Left to the reader to prove, by adapting the argument of part c) or otherwise. 
Exercise: Let {xn } be a real sequence. Suppose that:
(i) Any two convergent subsequences converge to the same limit.
(ii) {xn } is bounded.
Show that {xn } is convergent. (Suggestion: Combine Theorem 10.24b) with the
Bolzano-Weierstrass Theorem.)
Exercise: Let {xn } be a real sequence, and let a b be extended real numbers.
Suppose that there exists N Z+ such that for all n N , a xn b. Show that
{an } has a partial limit L with a L b.
7.2. The Limit Supremum and Limit Infimum.
For a real sequence {an }, let L be the set of all partial limits of {an }.
We define the limit supremum L of a real sequence to be the least upper bound
of the set of all partial limits of the sequence.
Theorem 10.25. For any real sequence {an } , L is a partial limit of the sequence and is thus the largest partial limit.
Proof. Case 1: The sequence is unbounded above. Then + is a partial
limit, so L = + is a partial limit.

7. PARTIAL LIMITS; LIMITS SUPERIOR AND INFERIOR

213

Case 2: The sequence diverges to . Then is the only partial limit and
thus L = is the largest partial limit.
Case 3: The sequence is bounded above and does not diverge to . Then it
has a finite partial L (it may or may not also have as a partial limit), so
L (, ). We need to find a subsequence converging to L.
For each k Z+ , L k1 < L, so there exists a subsequence converging to some
L0 > L k1 . In particular, there exists nk such that ank > L k1 . It follows from
these inequalities that the subsequence ank cannot have any partial limit which
is less than L; moreover, by the definition of L = sup L the subsequence cannot
have any partial limit which is strictly greater than L: therefore by the process of
elimination we must have ank L.

Similarly we define the limit infimum L of a real sequence to be the infimum of the
set of all partial limits. By reflection, the proof of Theorem 10.25 shows that L is a
partial limit of the sequence, i.e., there exists a subsequence ank such that ank L.
Here is a very useful characterization of the limit supremum of a sequence {an } it is
the unique extended real number L such that for any M > L, {n Z+ | an M }
is finite, and such that for any m < L, {n Z+ | an m} is infinite.
Exercise:
a) Prove the above characterization of the limit supremum.
b) State and prove an analogous characterization of the limit infimum.
Proposition 10.26. For any real sequence an , we have
(72)

L = lim sup ak
n kn

and
(73)

L = lim inf ak .
n kn

Because of these identities it is traditional to write lim sup an in place of L and

lim inf an in place of L.
Proof. As usual, we will prove the statements involving the limit supremum
and leave the analogous case of the limit infimum to the reader.
Our first order of business is to show that limn supkn ak exists as an extended real number. To see this, define bn = supkn ak . The key observation is
that {bn } is decreasing. Indeed, when we pass from a set of extended real numbers
to a subset, its supremum either stays the same or decreases. By Theorem 10.21,
bn L0 [, ].
Now we will show that L = L0 using the characterization of the limit supremum stated above. First suppose M > L0 . Then there exists n Z+ such that
supkn ak < M . Thus there are only finitely many terms of the sequence which are
at least M , so M L. It follows that L0 L.
On the other hand, suppose m < L0 . Then there are infinitely many n Z+
such that m < an and hence m L. It follows that L L0 , and thus L = L0 . 
The merit of these considerations is the following: if a sequence converges, we have a
number to describe its limiting behavior, namely its limit L. If a sequence diverges

214

10. INFINITE SEQUENCES

to , again we have an extended real number that we can use to describe

its limiting behavior. But a sequence can be more complicated than this: it may
be highly oscillatory and therefore its limiting behavior may be hard to describe.
However, to every sequence we have now associated two numbers: the limit infimum
L and the limit supremum L, such that
L L +.
For many purposes e.g. for making upper estimates we can use the limit
supremum L in the same way that we would use the limit L if the sequence were
convergent (or divergent to ). Since L exists for any sequence, this is very
powerful and useful. Similarly for L.
Corollary 10.27. A real sequence {an } is convergent iff L = L (, ).
Exercise: Prove Corollary 10.27.
8. Cauchy Sequences
8.1. Motivating Cauchy sequences.
Let us look back at our definition of a convergent sequence: a real sequence {an }
is convergent if there is some L R such that {an } converges to L.
There is something slightly curious about this definition: do you see it? It is
this: we are defining the notion convergent in terms of the notion convergent
to L. One might think that things should go the other way around: shouldnt the
concept of convergene should be somehow logically prior than convergence to L?
Anyway, this is the way we tend to think of things in practice: if one is given an
infinite sequence, first we want to know whether it is convergent or divergent, and
then if we know it is convergent we can ask the more refined question: To which
real number L does our convergent sequence converge?
This is a good opportunity to think back on our other notions of convergence.
If you think about it, our definition of the limit of f (x) as x c also has this
feature: what we actually defined was limxc f (x) = L. For continuity the issue is
not pressing because we know that we want the limit to be f (c).2 The concept of
uniform continuity really does not mention a specific limiting value, i.e., a function
is not uniformly continuous to L at c.
Since derivatives are defined in terms of limits (and not in terms of continuity!)
they have the same problem: to show that f is differentiable at c we have to show
that the limiting slope of the secant line is some specific real number.
For integrals, the plot thickens. Our first, awkward, definition involved a fixed
Rb
real number I. Then we gave a modified definition in terms of two quantities a f
Rb
and a f which exist for any function, and our definition of integrability was that
these two quantities are finite and equal. This is sort of in between in that the definition includes a particular limiting value but we dont have to find it: we have to
show that an underestimate and an overestimate are actually the same. Moreover
we have Darbouxs criterion, which remedies the issue completely: this gives us a
procedure to show that a function is integrable without requiring any knowledge on
2I think this partially explains why the concept of continuity is simpler than that of a limit.

8. CAUCHY SEQUENCES

215

what number the integral should be. And thats what made Darbouxs criterion
so useful: we used it to show that every continuous function and every monotone
function is integrable, but of course without having to find in any sense the value
of the integral. (This inexplicitness is not entirely a good thing, and the main point
of our discussion of Riemann sums was to make the convergence more explicit.)
Upshot: it would be nice to have some way of expressing/proving that a sequence
is convergent which doesnt have the limit of the sequence built into it. This is
exactly what Cauchy sequences are for.
8.2. Cauchy sequences.
+
A sequence {an }
such
n=1 in is Cauchy if for all  > 0, there exists N Z
that for all m, n N , |am an | < .

Here are some elementary properties of Cauchy sequences.3

Lemma 10.28. A subsequence of a Cauchy sequence is Cauchy.
Proof. Left to the reader as an exercise.

Proposition 10.29. A convergent sequence is Cauchy.

Proof. Suppose an L. Then there exists N Z+ such that for all n N ,
|an L| < 2 . Thus for all m, n N we have


|an am | = |(an L) (am L)| |an L| + |am L| < + = .
2 2

Proposition 10.30. A Cauchy sequence is bounded.
Proof. Let {an } be a Cauchy sequence. There exists N Z+ such that for
all m, n N , |am an | < 1. Therefore, taking m = N we get that for all n N ,
|an aN | < 1, so |an | |aN | + 1 = M1 , say. Moreover put M2 = max1nN |an |
and M = max(M1 , M2 ). Then for all n Z+ , |an | M .

Proposition 10.31. A Cauchy sequence which admits a convergent subsequence is itself convergent.
Proof. Let {an } be a Cauchy sequence and suppose there is a subsequence
ank converging to L F . We claim that an converges to L. Fix any  > 0.
Choose N1 Z+ such that for all m, n N1 we have |an am | = |am an | < 2 .
Further, choose N2 Z+ such that for all k N2 we have |ank L| < 2 , and put
N = max(N1 , N2 ). Then nN N and N N2 , so


|an L| = |(an anN ) (anN L)| |an anN | + |anN L| < + = .
2 2

3You might notice that they are reminiscent of the elementary properties of monotone sequences we proved above. But I myself noticed this only quite recently more than 15 years
after I first learned about monotone sequences and Cauchy sequences. I wish I had a deeper
understanding of the underlying source if any! of these commonalities.

216

10. INFINITE SEQUENCES

The above proofs did not use completeness, and thus the results hold in any ordered
field. In contrast, the next result does crucially use the Dedekind completeness of
the real numbers, in the form of the Bolzano-Weierstrass Theorem.
Theorem 10.32. (Cauchy Criterion) Any real Cauchy sequence is convergent.
Proof. Let {an } be a real Cauchy sequence. By Proposition 10.30, {an } is
bounded. By Bolzano-Weierstrass there exists a convergent subequence. Finally,
by Proposition 10.31, this implies that {an } is convergent.

It can be further shown that an Archimedean ordered field F in which every
Cauchy sequence is convergent must be Dedekind complete. However, there are
non-Archimedean and thus necessarily not Dedekind complete ordered fields
in which every Cauchy sequence converges. (In fact there are non-Archimedean
ordered fields in which the only Cauchy sequences are the eventually constant sequences!) But we had better not get into such matters here.
Exercise: Is there a Cauchy criterion for a function to be differentiable at a
point? (Hint: yes. See [Ma56].)
9. Geometric Sequences and Series
A geometric sequence is a sequence {xn }
n=0 of real numbers such that there is
a fixed real number r with xn+1 = rxn . We call r the geometric ratio since, if
= r.
for all n Z+ xn 6= 0 we have xxn+1
n
Theorem 10.33. (Geometric Sequences) Let {xn } be a geometric sequence with
geometric ratio r and x0 6= 0.
a) We have xn = x0 rn .
b) If |r| < 1, then xn 0.
c) If r = 1, then xn x0 .
d) If r = 1, then the sequence is x0 , x0 , x0 , x0 , . . ., which diverges.
e) If |r| > 1, then |xn | .
Proof. a) A simple induction argument which is left to the reader.
Suppose xn L [, ]. Since xn+1 = rxn , L = limn xn+1 =
limn rxn = rL. The solutions to L = rL for r R, L [, ] are: (1, L) for
any L, (r, 0) for any r, (r, ) for positive r and (r, ) for negative r. Now:
b) If |r| < 1, then |xn+1 | = |r||xn | < |xn |, so {|xn |} is decreasing and bounded
below by 0. By the Montone Sequence Lemma, |xn | converges to a finite, nonnegative number L, and by the above analysis L = 0. Since |xn | 0, xn 0.
c), d) These are immediate and are just recorded for easy reference.
e) If |r| > 1, then |xn+1 | = |r||xn | > |xn |, so the sequence {|xn |} is increasing and
bounded below by |x0 |. By the Monotone Sequence Lemma, |xn | L (|x0 |, ],
and by the above analysis we must have L = .

For x0 , r R, we define the finite geometric series
(74)

Sn = x0 + x0 r + . . . + x0 rn .

We claim that quite luckily! we are able to obtain a simple closed-form expression for Sn . We may dispose of the case r = 1: then Sn = (n + 1). When r 6= 1, we
use a very standard trick : multiplying (75) by r gives
(75)

rSn = x0 r + . . . + x0 rn + x0 rn+1 ,

10. CONTRACTION MAPPINGS REVISITED

217

and subtracting (75) from (74) gives

(1 r)Sn = x0 (1 rn+1 .
Since r 6= 1, this yields
(76)

Sn =

n
X

x0 r k =

k=0

1 rn+1
.
1r

Having this closed form enables us to determine easily whether the sequence {Sn }
converges, and if so, to what.
Theorem 10.34. (The Geometric Series) For x0 , r R, let
Sn =

n
X

x0 r n = x0 + x0 r + . . . + x0 r n .

k=0
1
.
a) If |r| < 1, then limn Sn = 1r
b) If |r| 1, then {Sn } diverges.

Exercise: Use Theorem 10.33 to prove Theorem 10.34.

10. Contraction Mappings Revisited
The material of this section owes a debt to a paper of K. Conrad [CdC].
10.1. Preliminaries.
Recall that we last studied contractive mappings, fixed points and sequences of
iterates in the context of Newtons method. Now, towards the end of a long chapter on infinite sequences, we have more tools at our disposal, so it is reasonable to
take another look at this important circle of ideas. Lets take it again from the top:
A function f : I R is contractive, or a contraction mapping, if there is
a constant 0 < C < 1 such that:
(77)

x, y I, |f (x) f (y)| C|x y|.

Recall that f : I R is Lipschitz if (77) holds for some C > 0. If so, for
any  > 0, we may independently of x choose < C , and then |x y|
= |f (x) f (y)| C < . So a Lipschitz function is uniformly continuous. A
contraction mapping is a Lipchitz mapping with Lipschitz constant C < 1.
A function f : I R is weakly contractive if:
x 6= y I, |f (x) f (y)| < |x y|.
A function f : I R is a short map if 1 is a Lipschitz constant for f , i.e.,
x, y I, |f (x) f (y)| |x y|.
Thus contractive implies weakly contractive implies short map.
Exercise: Let f : I I be differentiable.

218

10. INFINITE SEQUENCES

a) Show that if |f 0 (x)| C for all x I, then C is a Lipschitz constant for f .

b) Deduce: if there is C (0, 1) with |f 0 (x)| C for all x I, f is contractive.
c) State and prove analogous sufficient conditions in terms of bounds on f 0 for f
to be weakly contractive (resp. a short map).
Let X and Y be sets, and let f : X Y be a function (or, synonymously, a
map or mapping). An element L X is a fixed point for f if f (L) = L. The
study of fixed points of mappings is highly important throughout mathematics.
A fixed point L of f is, in particular, an element of both X and Y . Perhaps
the sets X and Y have no elements in common: then of course no function between
them can have a fixed point: e.g., looking for fixed points for mappings from the set
X of real numbers to the set Y = {Barack Obama, Lady Gaga, William Faulkner}
would be an exercise in futility.
To alleviate this problem, we often restrict attention to mappings f : X X.
Further, such a mapping can be iterated: that is, we can take the output and
plug it back in as the input: x X 7 f (x) 7 f (f (x)). And we can keep going.
Formally, starting from any initial value x0 X, we get a sequence of iterates
{xn }
n=0 where for all n N, xn+1 = f (xn ). (Recall that Newtons method involved consideration of sequences of iterates of the function T (x) = x ff0(x)
(x) .)
And now, back to earth: the methods of calculus allow us to analyze functions
defined on subsets, especially intervals, of the real numbers. So suppose we have
an interval I R and a function f : I I. Here is what we want to know.
Question 10.35. Let f : I I be a function.
a) Does f have at least one fixed point?
b) Does f have exactly one fixed point?
c) If f has fixed points, how can we explicitly find them, at least approximately?
10.2. Some Easy Results.
Our first result is a very modest generalization of Lemma 7.7.
Theorem 10.36. A weakly contractive function has at most one fixed point.
Proof. Suppose there are distinct L1 , L2 I such that f (L1 ) = L1 and
f (L2 ) = L2 . Then |f (L1 ) f (L2 )| = |L1 L2 |, contradiction.

Exercise: Show that a short map f : I I can have more than one fixed point.
Lemma 10.37. Let f : I I be a continuous function. Let x0 I and let
x0 , x1 = f (x1 ), x2 = f (x2 ), . . . be the sequence of iterates of x0 under f . Suppose
that xn converges to some L I. Then L is a fixed point of f .
Proof. Since f is continuous and xn L, f (xn ) f (L). But since f (xn ) =
xn+1 , the sequence {f (xn )} = {xn+1 } is just a reindexed version of the sequence
{xn }, so it must have the same limit: f (L) = L.

Exercise:4 Consider the function f : (0, 1] (0, 1] given by f (x) = x2 .
a) Show that f is contractive.
4This exercise and the next are easy bordering on insulting, but to properly appreciate later
results it will be useful to have these examples on record.

10. CONTRACTION MAPPINGS REVISITED

219

b) Show that f has no fixed point on (0, 1].

Exercise: Consider the function f : R R given by f (x) = x + 1.
a) Show that f is a short map, hence continuous.
b) Show that f has no fixed point on R.
Exercise: Consider the function f : R R given by f (x) =
a) Show that f is weakly contractive.
b) Show that f has no fixed point on R.
c) Show that f is not contractive.

x2 + 1.

When f : I I has a fixed point L, we would like to be able to find L as the

limit of a sequence of iterates. And we can: if x0 = L, then f (x0 ) = L, so the
sequence of iterates is constantly L. But this observation is not very useful in finding L! The next, slightly less trivial, observation is that a convegent sequence of
iterates of a continuous function yields a fixed point. To state this more precisely
we introduce some further terminology.
We say L I is an attracting point for f if for every x0 I, the sequence
of iterates x0 , f (x0 ), f (f (x0 )), . . . converges to L. We say f is attractive if it has
an attracting point.
Lemma 10.38. Suppose f : I I is an attractive function.
a) The attracting point L is unique.
b) f has at most one fixed point.
c) If f is continuous, then the unique attracting point is the unique fixed point.
Proof. a) Suppose L1 6= L2 are attractive points. Starting at any initial point
x0 L, the sequence of iterates of x0 under f must converge to both L1 and L2 .
Since a sequence can have at most one limit, this is a contradiction.
b) If L is a fixed point for f , then the sequence of iterates of L under f is constantly
L and thus converges to L. So if we had at least two fixed points, different sequences
of iterates would have at least two different limits, contradicting attractivity.
c) This follows immediately from the first two parts and from Lemma 10.37.

Thus attractive maps are precisely the maps with a unique fixed point which can
be found as the limit of a sequence of iterates starting from any point x0 in the
domain. So the fixed point of an attractive map is easy to find: we just start anywhere, iterate, and we get sucked into it.
Exercise: Consider the function f : [0, 1] [0, 1] such that

f (x) =

x
2,

0<x1
1, x = 0

a) Show that 0 is an attractive point of f but not a fixed point.

b) Why does this not contradict Lemma 10.38?
Remark: Note that, although our terminology suppresses this, whether L is an
attracting fixed point for f depends very much on the interval of definition I. In

220

10. INFINITE SEQUENCES

the literature, the term attracting point is often used for a weaker, local property which is studied in the following exercise.
Exercise: Let f : I I be a function. A point L I is a locally attracting
point if there exists > 0 such that f maps [L , L + ] into itself and the
restriction of f to [L , L + ] has L as an attracting point.
Now let f : I I be C 1 , and let L I be a fixed point of f .
a) Show that if |f 0 (L)| < 1, then L is a locally attracting point for f .
b) Show that if |f 0 (L)| > 1, then L is not a locally attracting point for f .
c) Exhibit f : [1, 1] R such that L = 0 is a fixed point, f 0 (0) = 1, and 0 is a
locally attracting point for f .
d) Exhibit f : [1, 1] R such that L = 0 is a fixed point, f 0 (0) = 1 and 0 is not
a locally attracting point for f .
10.3. The Contraction Mapping Theorem.
Theorem 10.39. (Contraction Mapping Theorem) Let I R be a closed interval, and let f : I I be contractive with constant C (0, 1).
a) Then f is attractive: there is a unique fixed point L, and for all x0 I, the
sequence {xn } of iterates of f under x0 converges to L.
b) Explicitly, for all x0 I and all n N,


|x0 f (x0 )|
C n.
(78)
|xn L|
1C
Proof. Step 0: By Theorem 10.36, f has at most one fixed point.
Step 1: Let x0 I, fix  > 0, let N be a large positive integer to be chosen (rather
sooner than) later, let n N and let k 0. Then
|xn+k xn | |xn+k xn+k1 | + |xn+k1 xn+k2 | + . . . + |xn+1 xn |
C n+k1 |x1 x0 | + C n+k2 |x1 x0 | + . . . + C n |x1 x0 |

 



1 Ck
|x1 x0 |
= |x1 x0 |C n 1 + C + . . . + C k1 = |x1 x0 |C n
<
C n.
1C
1C
Since |C|< 1, C n 0, so we may choose N such that for all n N and all k N,

|x1 x0 |
1C

C n < . So {xn } is Cauchy and hence, by Theorem 10.32, convergent.

Step 2: By Step 1 and Lemma 10.37, for each x0 I the sequence of iterates of x0
under f converges to a fixed point. By Step 0, f has at most one fixed point. So
there must be a unique fixed point L I, which is the limit of every sequence of
iterates: i.e., L is an attracting point for f . This completes the proof of part a).
Step 3: We leave the derivation of (78) as an exercise for the reader.

Exercise: Complete the proof of Theorem 10.39 by establishing (78). (Suggestion:
adapt the argument of Step 1 of the proof.)
Thus the Contraction Mapping Theorem asserts in particular that a contractive
mapping on a closed interval is attractive. Thats a fine result in and of itself.
But we have more, namely the explicit estimate (78). It guarantees that starting
at any point x0 I the sequence of iterates converges to the attracting point L
exponentially fast, and is so explicit that it immediately gives an efficient algorithm
for approximating L to any desired accuracy. Now thats a theorem!

10. CONTRACTION MAPPINGS REVISITED

221

Exercise: a) Show that there is a unique real number x such that cos x = x.
b) Explain how to use (say) a handheld calculator to approximate x to as many
decimal places of accuracy as your calculator carries.
Exercise: Let f : [a, b) [a, b) be a contractive map with constant C (0, 1).
a) Show that limxb f (x) exists (as a real number!), and that by defining f at
b to be this limit, the extended function f : [a, b] R remains contractive with
constant C.5
b) Use part a) to extend the Contraction Mapping Principle to contractions f :
[a, b) [a, b), with the proviso that the unique fixed point may be b.
c) Give an example of a contraction mapping on [a, b) with fixed point b.
d) State and prove a version of the Contraction Mapping Principle valid for an
arbitrary interval.
Exercise: Let f : I I be a function. For n Z+ , let f n = f . . . f be
the nth iterate of f . Suppose that for some N Z+ , f N is contractive.
a) Show that any fixed point of f is a fixed point of f N .
b) Show that f N has a unique fixed point L I, and deduce that f has at most
one fixed point in I.
c) Show that f (L) is also a fixed point for f N , and deduce that f (L) = L: thus f
has a unique fixed point in I.
d) Show that in fact L is an attracting point for f .
e) Consider the function f : [0, 1] [0, 1] by

0, 0 x [0, 1]
f (x) =
1, x (1, 2].
Show that f is not a contraction but f f is a contraction.
10.4. Further Attraction Theorems.
Let I be an interval, and let f : I I be continuous. Recall that L I is
attracting for f if for all x0 I, the sequence of iterates of x0 under f converges to
L; above, we showed that if L is attracting for f then L is the unique fixed point
of f . Let us say sup I is an attracting point for f if for all x0 I, the sequence
of iterates of x0 under f approaches sup I (sup I = iff I is unbounded above).
Similarly, we say inf I is an attracting point for f if for all x0 I, the sequence of
iterates of x0 under f approaches inf I (inf I = iff I is unbounded below).
Theorem 10.40. Let I R be an interval, and let f : I I be continuous.
a) At least one of the following holds:
(i) f has a fixed point.
(ii) sup I is an attracting point for f .
(iii) inf I is an attracting point for f .
b) If I = [a, b] then f has a fixed point.
Proof. a) Define g : I R by g(x) = f (x) x. A fixed point of f is precisely
a root of g, so to prove part a) we may assume that g has no roots on I and show
5This is the hardest part of the problem. The result on extension will be much easier after
you have read the next section, which treats similar but more general problems. You may wish to
assume this part for now and come back to it later.

222

10. INFINITE SEQUENCES

that either (ii) or (iii) holds. If the continuous function g has no roots then either
(I) f (x) > x for all x I or (II) f (x) < x for all x I. We will show (I) = (ii);
the very similar proof that (II) = (iii) is left to the reader.
Suppose f (x) > x for all x I. Then, for any x0 I, the sequence of iterates
is increasing. This sequence cannot converge to any element L of I, for by Lemma
10.37, L would then be a fixed point of f , contradiction. So xn must approach
sup I.
b) Let I = [a, b] and let x0 I. If there were no fixed point, then by part a) the
sequence of iterates of x0 under f would approach either sup I = b or inf I = a.
But both are elements of I, so by Lemma 10.37 either a or b is a fixed point.

Theorem 10.41. Let I R be an interval, and let f : I I be weakly
contractive. Then f has an attracting point in [inf I, sup I].
Proof. Step 0: If f has no fixed point in I, then by Theorem 10.40 either
sup I or inf I is an attracting point for f . Thus we may assume that f has a fixed
point L I, and our task is to show that L is attracting for f .
Step 1: Let x0 I, and for n N, put dn = |xn L|. If for some N we have
xN = L, then xn = L for all n N , so the sequence of iterates converges to L.
So we may assume dn > 0 for all n N, and then by weak contractivity {dn } is
decreasing. Since 0 is a lower bound, there is d 0 such that dn d. Observe
that the desired conclusion that xn L is equivalent to d = 0.
Step 2: For all n N, xn [L d0 , L + d0 ], so {xn } is bounded; by BolzanoWeierstrass, there is a convergent subsequence, say xnk y. As k we have:
|xnk L| |y L|, |xnk L| = dnk d,
|f (xnk ) L| = |xnk +1 L| = dnk +1 d, |f (xnk ) L| |f (y) L|,
so |y L| = d = |f (y) L| = |f (y) f (L)|. By weak contractivity, d = 0.

Remark: The case I = R of Theorem 10.41 is due to A. Beardon [Be06]. The case
I = [a, b] is an instance of a general result of M. Edelstein [Ed62] which is described
in the next section. Our proof of Theorem 10.41 draws ideas from Beardons proof
and also from K. Conrads treatment of Edelsteins Theorem in [CdC].
Exercise: Show that the function f : [0, 1] [0, 1] defined by f (x) =
weakly contractive but not contractive.

1
1+x

is

10.5. Fixed Point Theorems in Metric Spaces.

Theorem 10.39 is a special case of a celebrated theorem of Banach,6 itself called
either the Contraction Mapping Principle or Banachs Fixed Point Theorem. The flavor of the generalization can perhaps be appreciated by looking back
at what tools we used to prove Theorem 10.39. The general strategy of the proof
was to show that any sequence of iterates of x0 under f was a Cauchy sequence.
To show this, we used nothing other than the definition of a contractive mapping
and (repeatedly) the triangle inequality.
Now there is a general mathematical structure called a metric space: this is
a set X equipped with a metric function d : X X R. The intuition is that
for x, y X, d(x, y) measures the distance between x and y. When X = R, this
6Stefan Banach (1892-1945)

10. CONTRACTION MAPPINGS REVISITED

223

function is nothing else than d(x, y) = |x y|. In order for a function d to be a

metric, it must satisfy three very familiar properties:
(M1) (Definiteness) For all x, y X, d(x, y) 0, with equality iff x = y.
(M2) (Symmetry) For all x, y X, d(x, y) = d(y, x).
(M3) (Triangle Inequality) For all x, y, z X, d(x, z) d(x, y) + d(y, z).
Thus d(x, y) = |x y| is certainly a metric on R. It turns out that these three
simple properties enable us to carry over a large portion of the theory of real infinite sequences to study sequences with values in a metric space X, i.e., functions
x : Z+ X. Essentially, wherever in any of our definitions we see |x y|, we
replace it with d(x, y). For instance, we say that a sequence {xn } in a metric space
X converges to an element L X if for all  > 0, there is N Z+ such that for
all n N , d(xn , L) < . Similarly, one can directly carry over the definition of a
Cauchy sequence: try it!
Similarly, if (X, dX ) and (Y, dY ) are metric spaces, then a function f : X Y
is continuous at x X if for all  > 0, there exists > 0 such that for all x0 X
with dX (x, x0 ) < , dY (f (x), f (x0 )) < . A function f is continuous if it is continuous at every point of X. (The reader should recognize this as a rather straightforward adaptation of our cherished - definition of continuity to this new context.)
One important class of examples of metric spaces are obtained simply by taking a
subset X R and definining d(x, y) : X X R as before: d(x, y) = |x y|.
The point here is that for some subspaces of X, a Cauchy sequence with values
in X need not converge to an element of X. This is not really new or surprising:
for instance, consider the sequence xn = n1 as a sequence in (0, ): it is a Cauchy
sequence there, but it does not converge to an element of (0, ); rather it converges
to 0. Similarly, in a general metric space X a Cauchy sequence need not converge
to an element of X: we say that a metric space X is complete if every Cauchy
sequence in X is convergent in X. Now we can state the theorem.
Theorem 10.42. (Banach Fixed Point Theorem [Ba22]) Let X be a complete
metric space, and let f : X X be a contraction mapping: that is, there is
C (0, 1) such that for all x, y X,
d(f (x), f (y)) Cd(x, y).
Then:
a) There is a unique L X with f (L) = L, i.e., a fixed point of f .
b) For every x0 X, define the sequence of iterates of x0 under f by xn+1 = f (xn )
for n N. Then for all n N,
d(xn , L) C n d(x0 , L).
The proof of Theorem 10.42 is exactly the same as that of Theorem 10.39.
For the last century, Banachs Fixed Point Theorem has been one of the most
important and useful results in mathematical analysis: it gives a very general condition for the existence of fixed points, and a remarkable number of existence
theorems can be reduced to the existence of a fixed point of some function on
some metric space. For instance, if you continue on in your study of mathematics

224

10. INFINITE SEQUENCES

you will surely learn about systems of differential equations, and the most important result in this area is that with suitable hypotheses and precisions, of course
every system of differential equations has a unique solution. The now standard
proof of this seminal result uses Banachs Fixed Point Theorem!7
Let X be a metric space. Then the statement Every sequence with values in
X admits a convergent subsequence is certainly meaningful, but as is already
the case with intervals on the real line! whether it is true or false certainly depends on X. We say that a metric space is compact if every sequence with values
in X admits a convergent subsequence.
In fact every compact metric space is complete, and the proof again requires no
ideas other than the ones we have already developed: indeed, if {xn } is a Cauchy
sequence in a compact metric space, then by definition of compactness it admits
a subsequence xnk converging to some L X, and then we prove exactly as we
did before that if a subsequence of a Cauchy sequence converges to L then the
Cauchy sequence itself must converge to L.
Above we showed that a weakly contractive map on a closed, bounded (and thus
compact) interval was attractive and attributed this to M. Edelstein. What Edelstein actually showed was the following result.
Theorem 10.43. (Edelstein [Ed62]) Let X be a compact metric space, and let
f : X X be a weakly contractive mapping. Then f is attractive:
a) There is a unique fixed point L of f .
b) For all x0 X, the sequence of iterates of x0 under f converges to f .
Proof. We follow [CdC].
Step 0: The Extreme Value Theorem has the following generalization to compact
metric spaces: if X is a compact metric space, then any continuous function f :
X R is bounded and attains its maximum and minimum values. Recall that
we gave two proofs of the Extreme Value Theorem for X = [a, b]: one using Real
Induction and one using the fact that every sequence in [a, b] admits a convergent
subsequence. Since by definition this latter property holds in a compact metric
space, it is the second proof that we wish to carry over here, and we ask the
interested reader to check that it does carry over with no new difficulties.
Step 1: We claim f has a fixed point. Here we need a new argument: the one we
gave for X = [a, b] used the Intermediate Value Theorem, which is not available in
our present context. So here goes: let g : X R by g(x) = d(x, f (x)). Since f is
continuous, so is g and thus the Extreme Value Theorem applies and in particular
g attains a minimum value: there is L X such that for all y X, d(L, f (L))
d(y, f (y)). But if f (L) 6= L, then by weak contractivity we have d(f (L), f (f (L))) <
d(L, f (L)), i.e., g(f (L)) < g(L), contradiction. So L is a fixed point for f .
Step 2: The argument of Step 2 of the proof of Theorem 10.41 carries over directly
to show that L is an attracting point for f .

7In fact the title of [Ba22] indicates that applications to integral equations are being explicitly
considered. An integral equation is very similar in spirit to a differential equation: it is an
equating relating an unknown function to its integral(s).

11. EXTENDING CONTINUOUS FUNCTIONS

225

11. Extending Continuous Functions

Let I R be an interval.8 A subset S of I is dense in I if for all x I and all
 > 0, there exists at least one element of s with s (x , x + ).
For example, for every interval I the subset I Q of rational numbers in I is
dense in I, as is the subset I (R \ Q) of irrational numbers in I.
Here we wish to consider the following problem: suppose we are given a dense
subset S of I and a function f : S R. Under which circumstances does f extend to a function f : I R?
Stated this way the answer is easy: always. For instance, we may set f(x) = 0
for all x I \ S. So we were not careful enough. Heres what we really mean:
Question 10.44. Let S be a dense subset of the interval I, and suppose we
are given a function f : S R. Under what circumstances is there a continuous
function f : I R extending f , i.e., such that f(x) = f (x) for all x S?
Here are some preliminary observations. First, the uniqueness problem is easier to
solve than the existence problem, and it is helpful to nail that down first.
Proposition 10.45. (Uniqueness of Extension) Let S be a dense subset of an
interval I, and let f : S R be a function. There is at most one continuus
extension of f to I: that is, if f1 , f2 : I R are two continuous functions such
that f1 (x) = f2 (x) = f (x) for all x S, then f1 (x) = f2 (x) for all x I.
Proof. Step 1: First suppose f 0 on S. Since S is dense in I, for any x I
and n Z+ , there is xn S with |xn x| < n1 . Thus we get a sequence {xn }
n=1
taking values in S such that xn x. Let f : I R be any continuous extension
of f . Then since xn x, f (xn ) f (x). But {f (xn )} = 0, 0, 0, . . ., so it converges
(uniquely!) to 0 and thus f (x) = 0.
Step 2: In general, suppose f1 , f2 : I R are two continuous extensions of f ,
and put g = f1 f2 : I R. Then g is continuous and for all x S, g(x) =
f1 (x) f2 (x) = 0. By Step 1, for all x I,
0 = g(x) = f1 (x) f2 (x),
so f1 (x) = f2 (x) for all x I.

To go further, we need the definition of a continuous function f : S R for an

arbitrary subset S of R. There is nothing novel here: f is continuous at s0 S if for
all  > 0, there is > 0 such that for all s S with |s s0 | < , |f (s) f (s0 )| < ,
f : S R is continuous if it is continuous at s for every s S.
Proposition 10.46. If f admits a continuous extension f to I, then f is
continuous on S.
Proof. Let s0 S, and fix  > 0. Since S I, f is continuous at s0 , so there
is > 0 such that for all x I with |x s0 | < , |f(x) f(s0 )| < . In particular
then, for all s S with |s s0 | < , |f (s) f (s0 )| = |f(s) f(s0 )| < .

8To evade trivialities, we assume that all of our intervals contain more than one point.

226

10. INFINITE SEQUENCES

Example: I = R, S = R \ { 2}, f (x) = 0 for x < 2, f (x) = 1 for x > 2. Then

f is continuous on S but admits no continuous extension to I.
Here is a key observation: our counterexample function f is continuous on S but
not uniformly continuous there: there is no > 0 such that for all x1 , x2 S,
|x1 x2 | < = |f (x1 ) f (x2 )| < 1.
Theorem 10.47. (Extension Theorem) Let S be a dense subset of an interval
I, and let f : S R be a function. If f is uniformly continuous on S, then it
extends uniquely to a uniformly continuous function f : I R.
Proof. As above, since S is dense in I, for each x I there is a sequence
{an } in S with an x. This suggests that we define f(x) = limn f (an ). And
we will, but there are several things we need to check:
(i) The sequence f (an ) is convergent.
(ii) For any other sequence {bn } in S with bn x, limn f (an ) = limn f (bn ).
(iii) The (now well-defined) extension function f is continuous on I.
So lets do it:
(i) By Theorem 10.32, it suffices to show that {f (an )} is Cauchy. Fix  > 0.
Since f is uniformly continuous on S, there exists > 0 such that |s1 s2 | <
= |f (s1 ) f (s2 )| < . Since an x, there is N Z+ such that for n N ,
|an x| < 2 . Then for m, n N , |am an | < , so |f (am ) f (an )| < .
(ii) Suppose that an x and bn x. By (i), we know that f (an ) L and
f (bn ) M , say, and we must show that L = M . Fix  > 0. By uniform continuity
there exists > 0 such that |s1 s2 | < = |f (s1 ) f (s2 )| < . Moreover for all
sufficiently large n we have |an x| < 2 , |bn x| < 2 , so |an bn | < and hence
|f (an ) f (bn )| < . It follows that |L M | . Since  was arbitrary, L = M .
(iii) Let x I, and suppose that f is not uniformly continuous on I: then there
exists some  > 0 such that for all > 0, there are x1 , x2 I with |x1 x2 | < 3
but |f (x1 ) f (x2 )| . By (ii), there are s1 , s2 S such that


|s1 x1 |, |s2 x2 | < , |f (s1 ) f (x1 )|, |f (s2 ) f (x2 )| < .
3
3
Then
|s1 s2 | |s1 x1 | + |x1 x2 | + |x2 s2 | < ,
and
2
|f (x1 )f (x2 )| |f (x1 )f (s1 )|+|f (s1 )f (s2 )|+|f (s2 )f (x2 )| +|f (s1 )f (s2 )|.
3
But since f : S R is uniformly continuous, we may take to be sufficiently small
so that |s1 s2 | < = |f (s1 ) f (s2 )| < 3 . For such a we get
|f (x1 ) f (x2 )| <

2

+ = ,
3
3

contradiction.

Theorem 10.47 gives necessary and sufficient conditions for a function f : S R to
admit a uniformly continuous extension to I. When I is closed and bounded, this
solves our extension problem because uniform continuity is equivalent to continuity.
However, for an interval which is not closed and bounded, being uniformly continuous is much stronger than being continuous. For instance, the only polynomial

11. EXTENDING CONTINUOUS FUNCTIONS

227

functions which are uniformly continuous on all of R are the linear polynomials.
However there is an easy, but sneaky, way to soup up the Extension Theorem: if I is not closed and bounded, we dont have to extend f to I all at once;
we can do it via an increasing sequence of closed, bounded subintervals of I.
Corollary 10.48. Let S be a dense subset of R, and let f : S R. The
following are equivalent:
(i) For all M > 0, the restriction of f to S [M, M ] is uniformly continuous.
(ii) There is a unique extension of f to a continuous function f : R R.
Proof. (i) = (ii): Applying the Extension Theorem to f on S [M, M ],
we get a unique continuous extension fM : [M, M ] R. Since the extensions
are unique, for any x R, we may choose any M with M |x| and define f(x) =
fM (x): this does not depend on which M we choose. Moreover, since continuity at
a point depends only on the behavior of the function in small intervals around the
point, it is immediate that any function constructed from an expanding family of
continuous functions in this way is continuous on all of R.


CHAPTER 11

Infinite Series
1. Introduction
1.1. Zeno Comes Alive: a historico-philosophical introduction.
Humankind has had a fascination with, but also a suspicion of, infinite processes
for well over two thousand years. Historically, the first kind of infinite process that
received detailed infomation was the idea of adding together infinitely many quantitties; or, to put a slightly different emphasis on the same idea, to divide a whole
into infinitely many parts.
The idea that any sort of infinite process can lead to a finite answer has been
deeply unsettling to philosophers going back at least to Zeno,1 who believed that a
convergent infinite process was absurd. Since he had a sufficiently penetrating eye
to see convergent infinite processes all around him, he ended up at the lively conclusion that many everyday phenomena are in fact absurd (so, in his view, illusory).
We will get the flavor of his ideas by considering just one paradox, Zenos arrow paradox. Suppose that an arrow is fired at a target one stadium away. Can
the arrow possibly hit the target? Call this event E. Before E can take place,
the arrow must arrive at the halfway point to its target: call this event E1 . But
before it does that it must arrive halfway to the halfway point: call this event E2 .
We may continue in this way, getting infinitely many events E1 , E2 , . . . all of which
must happen before the event E. That infinitely many things can happen before
some predetermined thing Zeno regarded as absurd, and he concluded that the arrow never hits its target. Similarly he deduced that all motion is impossible.2
Nowadays we have the mathematical tools to retain Zenos basic insight (that a
single interval of finite length can be divided into infinitely many subintervals)
without regarding it as distressing or paradoxical. Indeed, assuming the arrow
takes one second to hit its target and (rather unrealistically) travels at uniform
velocity, we know exactly when these events Ei take place: E1 takes place after
1
1
1
2 seconds, E2 takes place after 4 seconds, and so forth: En takes place after 2n
seconds. Nevertheless there is something interesting here: we have divided the total
time of the trip into infinitely many parts, and the conclusion seems to be that
(79)

1 1
1
+ + . . . + n + . . . = 1.
2 4
2
1Zeno of Elea, ca. 490 BC - ca. 430 BC.
2One has to wonder whether he got out much.
229

230

11. INFINITE SERIES

So now we have not a problem not in the philosophical sense but in the mathematical
one: what meaning can be given to the left hand side of (79)? Certainly we ought
to proceed with some caution in our desire to add infinitely many things together
and get a finite number: the expression
1 + 1 + ... + 1 + ...
represents an infinite sequence of events, each lasting one second. Surely the aggregate of these events takes forever.
We see then that we dearly need a mathematical definition of an infinite series
of numbers and also of its sum. Precisely, if a1 , a2 , . . . is a sequence of real numbers
and S is a real number, we need to give a precise meaning to the equation
a1 + . . . + an + . . . = S.
So here it is. We do not try to add everything together all at once. Instead, we
form from our sequence {an } an auxiliary sequence {Sn } whose terms represent
adding up the first n numbers. Precisely, for n Z+ , we define
Sn = a1 + . . . + an .
The associated sequence {Sn } is said to be the sequence of partial sums of the
sequence {an }; when necessary we call {an } the sequence
of terms. Finally, we
P
say that the infinite series a1 + . . . + an + . . . = n=1 an converges to S or
has sum S if limn Sn = S in the familiar sense of limits of seqeunces. If
the sequence of partial sums {Sn } converges to some number S we say thePinfinite

series is convergent; if the sequence {Sn } diverges then the infinite series n=1 an
is divergent.
P
Thus the trick of defining the infinite sum n=1 an is to do everything in terms
of the associated sequence
P of partial sums Sn = a1 + . . . + an .
In particular
by
n=1 an = we mean the sequence of partial sums diverges
P
to , and by n=1 an = we mean the sequence
Pof partial sums diverges to
. So to spell out the first definition completely, n=1 an = means: for every
M R there exists N Z+ such that for all n N , a1 + . . . + an M .
Let us revisit the examples above using the formal definition of convergence.
Example 1: Consider the infinite series 1 + 1 + . . . + 1 + . . ., in which an = 1
for all n. Then Sn = a1 + . . . + an = 1 + . . . + 1 = n, and we conclude

1 = lim n = .
n

n=1

Thus this infinite series indeed diverges to infinity.

Example 2: Consider
(80)

1
2

1
4

+ ... +

1
2n

Sn =

+ . . ., in which an =

1
2n

for all n, so

1
1
+ ... + n.
2
2

There is a standard trick for evaluating such finite sums. Namely, multiplying (80)
by 12 and subtracting it from (80) all but the first and last terms cancel, and we get

1. INTRODUCTION

231

1
1
1
1
Sn = Sn Sn = n+1 ,
2
2
2 2
and thus
Sn = 1
It follows that

1
.
2n

X
1
1
= lim (1 n ) = 1.
n
n
2
2
n=1

So Zeno was right!

Remark: It is not necessary for the sequence of terms {an } of an infinite series
to start with a1 . In our applications it will be almost as common
consider series
Pto

starting with a0 . More generally, if N is any integer, then by n=N an we mean

the sequence of partial sums aN , aN + aN +1 , aN + aN +1 + aN +2 , . . ..
1.2. Telescoping Series.
Example: Consider the series

1
n=1 n2 +n .

We have

1
,
2
1 1
2
S2 = S1 + a2 = + = ,
2 6
3
1
3
2
= ,
S3 = S2 + a3 = +
3 12
4
3
1
4
S4 = S3 + a4 = +
= .
4 20
5
1
n
It certainly seems as though we have Sn = 1 n+1
= n+1
for all n Z+ . If this is
the case, then we have

X
n
an = lim
= 1.
n n + 1
n=1
S1 =

How to prove it?

First Proof: As ever, induction is a powerful tool to prove that an identity holds
for all positive integers, even if we dont really understand why the identity should
hold! Indeed, we dont even have to fully wake up to give an induction proof: we
wish to show that for all positive integers n,
n
X
n
1
=
.
(81)
Sn =
2
k +k
n+1
k=1

When n = 1, both sides equal

Sn+1 = Sn +

(n +

1)2
=

1
2.

Now suppose (81) holds for some n Z+ . Then

1
n
1
n
1
IH
=
+ 2
=
+
+ (n + 1)
n + 1 n + 3n + 2
n + 1 (n + 1)(n + 2)

(n + 2)n + 1
(n + 1)2
n+1
=
=
.
(n + 1)(n + 2)
(n + 1)(n + 2)
n+2

This approach will work whenever we have some reason to look for and successfully
guess a simple closed form identity for Sn . But in fact, as we will see in the coming

232

11. INFINITE SERIES

sections, in practice it is exceedingly rare that we are able to express the partial
sums Sn in a simple closed form. Trying to do this for each given series would turn
out
Pto be1 a discouraging waste of time. We need some insight into why the series
n=1 n2 +n happens to work out so nicely.
Well, if we stare at the induction proof long enough we will eventually notice how
convenient it was that the denominator of (n+1)21+(n+1) factors into (n + 1)(n + 2).
1
Equivalently, we may look at the factorization n21+n = (n+1)(n)
. Does this remind
us of anything? I certainly hope so: this is a partial fractions decomposition. In
this case, we know there are constants A and B such that
1
A
B
= +
.
n(n + 1)
n
n+1
I leave it to you to confirm in whatever manner seems best to you that we have
1
1
1
=
.
n(n + 1)
n n+1
This makes the behavior of the partial sums much more clear! Indeed we have
1
S1 = 1 .
2
1
1 1
1
S2 = S1 + a2 = (1 ) + ( ) = 1 .
2
2 3
3
1
1 1
1
S3 = S2 + a3 = (1 ) + ( ) = 1 ,
3
3 4
4
1
and so on. This much simplifies the inductive proof that Sn = 1 n+1
. In fact
induction is not needed: we have that
1
1 1
1
1
1
Sn = a1 + . . . + an = (1 ) + ( ) + . . . + (
)=1
,
2
2 3
n n+1
n+1
the point being that every term
cancelled out by some
Pexcept the first and last is
1
other term. Thus once again n=1 n21+n = limn 1 n+1
= 1.
Finite sums which cancel in this way are often called telescoping sums, I believe P
after those old-timey collapsible nautical telescopes. In general an infinite

sum n=1 an is telescoping when we can find an auxiliary sequence {bn }

n=1 such
that a1 = b1 and for all n 2, an = bn bn1 , for then for all n 1 we have
Sn = a1 + a2 + . . . + an = b1 + (b2 b1 ) + . . . + (bn bn1 ) = bn .
But looking at these formulas shows something curious: every infinite series is
telescoping: we need only take bn = Sn for all n! Another, less confusing, way to
say this is that if we start with any infinite sequence {Sn }
n=1 , then there is a unique
such
that
S
is
the
sequence
of
partial
sums Sn = a1 + . . . + an .
sequence {an }
n
n=1
Indeed, the key equations here are simply
S1 = a1 ,
n 2, Sn Sn1 = an ,
which tells us how to define the an s in terms of the Sn s.
In practice all this seems to amount to the following: if you can find a simple
closed form expression for the nth partial sum Sn (in which case you are very

2. BASIC OPERATIONS ON SERIES

233

lucky), then in order to prove it you do not need to do anything so fancy as mathematical induction (or fancier!). Rather, it will suffice to just compute that S1 = a1
and for all n 2, Sn RSn1 = an . This is the discrete analogue of the fact that if
you want to show that f dx = F i.e., you already have a function F which you
believe is an antiderivative of f then you need not use any integration techniques
whatsoever but may simply check that F 0 = f .
Pn
Exercise: Let n Z+ . We define the nth harmonic number Hn = k=1 k1 =
1
1
1
1 + 2 + . . . + n . Show that for all n 2, Hn Q \ Z. (Suggestion: more specifically,
show that for all n 2, when written as a fraction ab in lowest terms, then the
denominator b is divisible by 2.)3
+
Exercise:
P Let1 k Z . Use the method of telescoping sums to give an exact formula
for n=1 n(n+k) in terms of the harmonic number Hk of the previous exercise.

2. Basic Operations on Series

P
Given an infinite series n=1 an there are two basic questions to ask:
P
Question 11.1. For an infinite series n=1 an :
a) Is the series convergent or divergent?
b) If the series is convergent, what is its sum?
It may seem that this is only one and a half questions because if the series diverges
we cannot ask about its sum (other than to ask whether it diverges to or due
to oscillation). However, later on we will revisit this missing half a question: if
a series diveges we may ask how rapidly it diverges, or in more sophisticated language we may ask for an asymptotic estimate for the sequence of partial sums
PN
n=1 an as a function of N as N .
Note that we have just seen an instanceP
in which we asked and answered both

of these questions: for a geometric series n=N crn , we know that the series concr N
. We should keep this success
verges iff |r| < 1 and in that case its sum is 1r
story in mind, both because geometric series are ubiquitous and turn out to play
a distinguished role in the theory in many ways, but also because other examples
of series in which we can answer Question 11.1b) i.e., determine the sum of a
convergent series are much harder to come by. Frankly, in a standard course
on infinite series one all but forgets about Question 11.1b) and the game becomes
simply to decide whether a given series is convergent or not. In these notes we try
to give a little more attention to the second question in some of the optional sections.
In any case, there is a certain philosophy at work when one is, for the P
moment,

interested in determining the convergence / divergence of a given series n=1 an

rather than the sum. Namely, there are certain operations that one can perform
on an infinite series that will preserve the convergence / divergence of the series
i.e., when applied to a convergent series yields another convergent series and when
applied to a divergent series yields another divergent series but will in general
3This is a number theory exercise which has, so far as I know, nothing to do with infinite
series. But I am a number theorist...

234

11. INFINITE SERIES

change the sum.

First, we may add or remove any finite number of terms from an infinite series
without
P affecting its convergence. In other words, suppose we start
Pwith a series
a
.
Then,
for
any
integer
N
>
1,
consider
the
series
n
n=1
n=N +1 an =
aN +1 + aN +2 + . . .. Then the first series converges iff the second series converges.
Here is one (among many) ways to show this formally: write Sn = a1 + . . . + an
and Tn = aN +1 + aN +2 + . . . + aN +n . Then for all n Z+ ,
!
N
X
ak + Tn = a1 + . . . + aN + aN +1 + . . . + aN +n = SN +n .
k=1

P
Thus if limn Tn = n=N +1 an exists, so does limn SN +n = limn Sn =
P
P
PN
an . Conversely if
n=1 an exists, then
Pn=1
Pso does limn k=1 ak + Tn =
n
k=1 ak + limn Tn , hence limn Tn =
n=N +1 an exists.
Similarly, if we are so inclined (and we will be, on occasion), we could add finitely
many terms to the series, or for that matter change finitely many terms of the
series, without affecting the convergence. We record this as follows.
Proposition 11.2. The addition, removal or altering of any finite number of
terms in an infinite series does not affect the convergence or divergence of the series
(though of course it may change the sum of a convergent series).
As the reader has probably already seen for herself, reading someone elses formal
proof of this result can be more tedious than enlightening, so we leave it to the
reader to construct a proof that she finds satisfactory.
P
Because the convergence or divergence of a series n=1 an is not affected by changing
P the lower limit 1 to any other integer, we often employ a simplified notation
n an when discussing series only up to convergence.
P
P
Proposition 11.3. Let n=1 an , n=1 bn be two infinite series, and let be
any real
Pnumber.
P
P
a) If n=1 an = A and n=1 bn = B are both convergent, then the series n=1 an +
bn is P
also convergent, with sum A + B.
P

b) If n=1 an = A is convergent, then so also is n=1 an , with sum A.

Proof. a) Let An = a1 + . . . + an , Bn = b1 + . . . + bn and Cn = a1 + b1 + . . . +
an + bn . Then An A and Bn B, so for any  > 0, there is N Z+ such that
for all n N , |An A| < 2 and |Bn B| < 2 . It follows that for all n N ,


|Cn (A + B)| = |An + Bn A B| |An A| + |Bn B| + = .
2 2
b) Left to the reader.

P
Exercise: Let n an bePan infinite series and R.
a) If = 0, show that nPan = 0.
P
b) Suppose 6= 0. Show n an converges iff n an converges.
P
P
P
Exercise: Prove the Three Series Principle: let n an , n bn , n cn be
Pthree
infinite
series
with
c
=
a
+
b
for
all
n.
If
any
two
of
the
three
series
n
n
n
n an ,
P
P
b
,
c
converge,
then
so
does
the
third.
n
n
n
n

2. BASIC OPERATIONS ON SERIES

235

2.1. The Nth Term Test.

P
P
Theorem 11.4. (Nth Term Test) Let n an be an infinite series. If n an
converges, then an 0.
P
Proof. Let S = n=1 an . Then for all n 2, an = Sn Sn1 . Therefore
lim an = lim Sn Sn1 = lim Sn lim Sn1 = S S = 0.


We will often apply theP
contrapositive form of Theorem 11.4: if for a series
we have an 9 0,4 then n an diverges.

an

Warning: The P
converse of Theorem 11.4 is not valid! It may well be the case
that an 0 but n an diverges. Later we will see many examples. Still, when put
under duress (e.g. while taking an exam) many students can will themselves into
believing that the converse might be true. Dont do it!
P (x)
be a rational function. The polynomial Q(x) has only finitely
Exercise: Let Q(x)
many roots, so we may choose N Z+ such that for all n N , Q(x) 6= 0. Show
P P (n)
that if deg P deg Q, then n=N Q(n)
is divergent.

2.2. The Cauchy Criterion.

Recall that we proved that a sequence {xn } of real numbers is convergent if and
only if it is Cauchy: that is, for all  > 0, there exists N Z+ such that for all
m, n N we have |xn xm | < .
Applying the Cauchy
to the sequence of partial sums {Sn = a1 +. . .+an }
Pcondition

of an infinite series n=1 an , we get the following result.

Proposition
11.5. (Cauchy criterion for convergence of series) An infinite
P
series n=1 an converges if and only if: for every  > 0, there exists N0 Z+ such
PN +k
that for all N N0 and all k N, | n=N an | < .
Note that taking k = 0 in the Cauchy criterion, we recover the Nth Term Test for
convergence (Theorem 11.4). It is important to compare these two results: the Nth
Term Test gives a very weak necessary condition for the convergence of the series.
In order to turn this into a necessary and sufficient condition we must require not
only that an 0 but also an + an+1 0 and indeed that an + . . . + an+k 0 for
a k which is allowed to be (in a certain precise sense) arbitrarily large.
PN +k
Let us call
a sum of the form n=N = aN + aN +1 + . . . + aN +k a finite tail
P
of the series n=1 an . As a matter of notation, if for a fixed N Z+ and all k N
PN +k
we have | n=N an | , let us abbreviate this by
|

an | .

n=N

PN +k
In other words the supremum of the absolute values of the finite tails | n=N an |
is at most . This gives a nice way of thinking about the Cauchy criterion.
4This means: a L 6= 0, or a diverges.
n
n

236

11. INFINITE SERIES

P
Proposition 11.6. An infinite series n=1 an converges
P if and only if: for all
 > 0, there exists N0 Z+ such that for all N N0 , | n=N an | < .
In other (less precise) words, an infinite series converges iff by removing sufficiently
many of the initial terms, we can make what remains arbitrarily small.
3. Series With Non-Negative Terms I: Comparison
3.1. The sum is the supremum.
Starting
in this section we get down to business by restricting our attention to
P
series n=1 an with an 0 for all n Z+ . This simplifies matters considerably
and places an array of powerful tests at our disposal.
Why? Assume an 0 for all n Z+ and consider the sequence of partial sums:
S1 = a1 a1 + a2 = S2 a1 + a2 + a3 = S3 ,
and so forth. In general, Sn+1 Sn = an+1 0, so the sequence of partial sums
{Sn } is increasing. Applying the Monotone Sequence Lemma we get:
P
Proposition 11.7. Let n an be an infinite series with an 0 for all n. Then
the series converges if and only if the partial sums are bounded above, i.e., if and
only if there exists M R such that for all n, a1 + . . . + an M . Moroever if the
series converges, its sum is precisely the least upper bound of the sequence of partial
sums. If the partial sums are unbounded, the series diverges to .
Because of this, when dealing
with series with non-negative terms
P
P we may express
convergence by writing n an < and divergence by writing n an = .
3.2. The Comparison Test.
P
Example: Consider the series n=1 n21n . Its sequence of partial sums is
 
 
 
1
1
1
1
1
Tn = 1
+
+ ... +
.
2
2
4
n
2n
Unfortunately we do not (yet!) know a closed form expression for Tn , so it is
not possible for us to compute limn Tn directly. But if we just want P
to decide

whether the series converges, we can compare it with the geometric series n=1 21n :
Sn =

1 1
1
+ + ... + n.
2 4
2

Since n1 1 for all n Z+ , we have that for all n Z+ , n21n 21n . Summing these
inequalities from k = 1 to n gives Tn Sn for all n. By our work with geometric
series we know that Sn 1 for all n and thus also TnP
1 for all n. Therefore our

given series has partial sums bounded above by 1, so n=1 n21n 1. In particular,
the series converges.
P
Example:
conside the series n=1 n. Again, a closed form expression for Tn =

1 + . . . + n is not easy to come by. But we dont need it: P

certainly
Tn

1 + . . . + 1 = n. Thus the sequence of partial sums is unbounded, so n=1 n = .

3. SERIES WITH NON-NEGATIVE TERMS I: COMPARISON

237

P
P
Theorem 11.8. (Comparison Test) Let n=1 an , n=1 bn be two series with
non-negative terms, and suppose that an bn for all n Z+ . Then

In particular: if

n bn

< then

an

bn .

n=1

n=1

an < , and if

an = then

n bn

= .

Proof. There is really nothing new to say here, but just to be sure: write
Sn = a1 + . . . + an , Tn = b1 + . . . + bn .
Since ak bk for all k we have Sn Tn for all n and thus

an = sup Sn sup Tn =

n=1

bn .

n=1

The assertions about convergence and divergence follow immediately.

3.3. The Delayed Comparison Test.

The Comparison Test
P is beautifully simple when it works. It has two weaknesses:
first, given a series n an we need to find some other series to compare it to. Thus
the test will be more or less effective according to the size of our repertoire of known
convergent/divergent series. (At the moment, we dont know much, but that will
soon change.) Second, the requirement that an bn for all n Z+ is rather inconveniently strong. Happily, it can be weakened in several ways, resulting in minor
variants of the Comparison Test with a much wider range of applicability.
Example: Consider the series

X
1
1
1
1
1
=1+1+ +
+
+ ... +
+ ....
n!
2
2

3
2

4
2

... n
n=0

We would like to show that the series converges by comparison, but what to compare
it to? Well, there is always the geometric series! Observe that the sequence n!
grows faster than any geometric rn in the sense that limn rn!n = . Taking
1
reciprocals, it follows that for any 0 < r < 1 we will have n!
< r1n not necessarily
+
for all n Z , but at least for all sufficiently large n. For instance, one easily
1
1
establishes by induction that n!
< 21n if and only if n 4. Putting an = n!
and
1
bn = 2n we cannot apply the Comparison Test because we have an bn for all
n 4 rather than for all n 0. But this objection is more worthy of a bureaucrat
than a mathematician: certainly the idea of the Comparison Test is applicable here:

X
X
X
X
1
1
1
1
8 1
67
=
+
8/3 +
= + =
< .
n
n!
n!
n!
2
3
8
24
n=0
n=0
n=4
n=4

So the series converges. More than that, we still retain a quantitative estimate on
the sum: it is at most (in fact strictly less than, as a moments thought will show)
67
24 = 2.79166666 . . .. (Perhaps this reminds you of e = 2.7182818284590452353602874714 . . .,
which also happens to be a bit less than 67
24 . It should! More on this later...)
We record the technique of the preceding example as a theorem.

238

11. INFINITE SERIES

P P
Theorem 11.9. (Delayed Comparison Test) Let n=1 , n=1 bn be two series
with non-negative terms. Suppose that there exists N Z+ such that for all n > N ,
an bn . Then
!

X
X
X
an
an bn +
bn .
n=1

In particular: if

n bn < then

n=1
Pn=1
P
P
a
<
,
and
if
n
n
n an = then
n bn = .

Exercise: Prove Theorem 11.9.

Thus the Delayed Comparison Test assures us that we do not need an bn for
all n but only for all sufficiently large n. A different issue occurs when we wish to
apply the Comparison Test and the inequalities do not go our way.
3.4. The Limit Comparison Test.
P
P
Theorem 11.10. (Limit Comparison Test) Let n an , n bn two series. Sup+
pose thatP
there exists N ZP
and M R0 such that for all n N , 0 an M bn .
Then if n bn converges, n an converges.
Exercise: Prove Theorem 11.10.
P
Corollary
11.11. (Calculus Students Limit Comparison Test) Let
n an
P
and n bn be two series. Suppose that for all sufficiently large n both an and bn
are positive and limn abnn = L [0, ].
P
P
a) If 0 < L < , the series n an and n bn converge or diverge together (i.e.,
either both convege
Por both diverge).
P
b) If L = andP n an converges, thenP n bn converges.
c) If L = 0 and n bn converges, then n an converges.
Proof. In all cases we deduce the result from the Limit Comparison Test.
an (2L)bn .
a) If 0 < L < , then there exists N Z+ such that 0 < L2 bnP
Applying
Theorem
11.10
to
the
second
inequality,
we
get
that
if
n bn converges,
P
then n an converges. The first inequality is equivalent toP0 < bn L2 an for all
n
we get that if n an converges, then
P N , and applying Theorem 11.10
Pto thisP
b
converges.
So
the
two
series
a
,
n n
n n
n bn converge or diverge together.
b) If L = , then there exists N Z+ such
Pthat for all n N
P, an bn 0.
Applying Theorem 11.10 to this we get that if n converges, then n bn converges.
c) This case is left to the reader as an exercise.

Exercise: Prove Corollary 11.11c).
P
Example: We will show that for all p 2, the p-series n=1 n1p converges. In
fact it is enough to show this P
for p = 2, P
since for p > 2 we have for all n Z+ that
1
1
1
2
p
n < n and thus np < n2 so n np n n12 . For p = 2, we happen to know that



X
X
1
1
1
=

= 1,
n2 + n n=1 n n + 1
n=1
P
and in particular that n n21+n converges. For large n, n21+n is close to n12 . More
precisely, putting an = n21+n and bn = n12 we have an bn , i.e.,
an
n2
1
= lim 2
= lim
n bn
n n + n
n 1 +
lim

1
n

= 1.

3. SERIES WITH NON-NEGATIVE TERMS I: COMPARISON

239

P
P
Applying Theorem 11.11, we find that n n21+n and n n12 converge or diverge
P
together. Since the former series converges, we deduce that n n12 converges, even
though the Comparison Test does not directly apply.
Exercise: Let

P (x)
Q(x)

be a rational function such that the degree of the denominator

P P (n)
converges.
minus the degree of the numerator is at least 2. Show n=N Q(n)
Exercise:
P Determine whether each of the following series converges or diverges:
a) n=1 sin n12 .
P
b) n=1 cos n12 .
3.5. Cauchy products I: non-negative terms.
Let

n=0

an and

n=0 bn

be infinite series. Can we, in some sense, multiply them?

In order to forestall possible confusion, let us point out that many students are
tempted to consider the following product operation on series:
!
!

X
X
X
??
an
bn =
an bn .
n=0

n=0

n=0

In other words, given two sequences of terms {an }, {bn }, we form a new sequence
of terms {an bn } and then theP
associated series.
P In fact this is not a very useful
candidate for the product. If n an = A and n bn = B, we want our product
series to converge to AB. But for instance, take {an } = {bn } = 21n . Then

an =

n=0

bn =

n=0

1
1

1
2

= 2,

so AB = 4, whereas

X
n=0

Unfortunately

4
3

an bn =

X
1
1
=
n
4
1

n=0

1
4

4
.
3

6= 4. What went wrong?

Plenty! We have ignored the laws of algebra for finite sums: e.g.
(a0 + a1 + a2 )(b0 + b1 + b2 ) = a0 b0 + a1 b1 + a2 b2 + a0 b1 + a1 b0 + a0 b2 + a1 b1 + a2 b0 .
The product is different and more complicated and indeed, if all the terms are
positive, strictly larger than just a0 b0 + a1 b1 + a2 b2 . We have forgotten about
the cross-terms which show up when we multiply one expression involving several
terms by another expression involving several terms.5
Let us try again at formally multiplying out a product of infinite series:
(a0 + a1 + . . . + an + . . .)(b0 + b1 + . . . + bn + . . .)
= a0 b0 + a0 b1 + a1 b0 + a0 b2 + a1 b1 + a2 b0 + . . . + a0 bn + a1 bn1 + . . . + an b0 + . . . .
5To the readers who did not forget about the cross-terms: my apologies. But it is a common
enough misconception that it had to be addressed.

240

11. INFINITE SERIES

The notation is getting complicated. In order to shoehorn the right hand side into
a single infinite series, we need to either (i) choose some particular ordering of the
terms ak b` or (ii) collect some terms together into an nth term.
For the moment we choose the latter: we define for any n N
n
X
cn =
ak bnk = a0 bn + a1 bn1 + . . . + an bn
k=0

and then we define the Cauchy product of

X
n=0

cn =

n
X

n=0

k=0

n=0

an and
!

ak bnk

n=0 bn

to be the series

Theorem
11.12. LetP{an }
n=0 , {bn }n=0 be two series
P
Pnwith non-negative terms.

Let
a
=
A
and
b
=
B.
Putting
c
=
n
n
n
n=0
k=0 ak bnk we have that
P n=0
c
=
AB.
In
particular,
the
Cauchy
product
series
converges iff the two
n
n=0
P
P
factor series n an and n bn both converge.

Proof. We define another sequence, the box product, as follows: for N N,

X
N =
ai bj = (a0 + . . . + aN )(b0 + . . . + bN ) = AN BN .
0i,jN

Thus by the usual product rule for sequences, we have

lim N = lim AN BN = AB.

So the box product clearly converges to the product of the sums of the two series.
This suggests that we compare the Cauchy product to the box product. The entries
of the box product can be arranged to form a square, viz:
N = a0 b0 + a0 b1 + . . . + a0 bN
+a1 b0 + a1 b1 + . . . + a1 bN
..
.
+aN b0 + aN b1 + . . . + aN bN .
On the other hand, the terms of the N th partial sum of the Cauchy product can
naturally be arranged in a triangle:
CN =

a0 b0
+a0 b1 + a1 b0
+ a0 b2 + a1 b1 + a2 b0
+a0 b3 + a1 b2 + a2 b1 + a3 b0
..
.

+a0 bN + a1 bN 1 + a2 bN 2 + . . . + aN b0 .
Thus while N is a sum of (N + 1)2 terms, CN is a sum of 1 + 2 + . . . + N + 1 =
(N +1)(N +2)
terms: those lying on our below the diagonal of the square. Thus in
2
considerations involving the Cauchy product, the question is to what extent one
can neglect the terms in the upper half of the square i.e., those with ai bj with
i + j > N as N gets large.

4. SERIES WITH NON-NEGATIVE TERMS II: CONDENSATION AND INTEGRATION 241

Here, since all the ai s and bj s are non-negative and N contains all the terms
of CN and others as well, we certainly have
CN N = AN BN AB.
Thus C = limN CN AB. For the converse, the key observation is that if we
make the sides of the triangle twice as long, it will contain the box: that is, every
term of N is of the form ai bj with 0 i, j N ; thus i + j 2N so ai bj appears
as a term in C2N . It follows that C2N N and thus
C = lim CN = lim C2N lim N = lim AN BN = AB.
N

Having shown both that C AB and C AB, we conclude

! !

X
X
X
C=
an = AB =
an
bn .
n=0

n=0

n=0


4. Series With Non-Negative Terms II: Condensation and Integration
P
We have recently been studying criteria for convergence of an infinite series n an
which are valid under the assumption that an 0 for all n. In this section we place
ourselves under more restrictive hypotheses: that for all n N, an+1 an 0,
i.e., that the sequence of terms is non-negative and decreasing.
Remark: Its no loss of generality to assume an > 0 for all n. If not, aN = 0 for
some N and, since the terms are assumed decreasing we have 0 = aN = aN +1 = . . .
PN 1
and our infinite series reduces to the finite series n=1 an : this converges!
4.1. The Harmonic Series.
P 1
Consider
n=1 n , the harmonic series. Does it converge? None of the tests
we have developed so far are up to the job: especially, an 0 so the Nth Term
Test is inconclusive.
Let us take a computational approach by looking at various partial sums. S100
is approximately 5.187. Is this close to a familiar real number? Not really. Next
we compute S150 5.591 and S200 5.878. Perhaps the partial sums never exceed
6? (If so, the series would converge.) Lets try a significantly larger partial sums:
S1000 7.485, so the above guess is incorrect. Since S1050 7.584, we are getting
the idea that whatever the series is doing, its doing it rather slowly, so lets instead
start stepping up the partial sums multiplicatively:
S100 5.878.
S103 7.4854.
S104 9.788.
S105 12.090.
Now there is a pattern for the perceptive eye to see: the difference S10k+1 S10k
appears to be approaching 2.30 . . . = log 10. This points to Sn log n. If this is so,
then since log n the series would diverge. I hope you notice that the relation
between n1 and log n is one of a function and its antiderivative. We ask the reader
to hold this thought until we discuss the integral test a bit late on.

242

11. INFINITE SERIES

For now, we give the following brilliant and elementary argument due to Cauchy.
Consider the terms arranged as follows:
  
 

1
1 1
1 1 1 1
+
+
+
+ + +
+ ...,
1
2 3
4 5 6 7
i.e., we group the terms in blocks of length 2k . The power of 21 which begins each
block is larger than every term in the preceding block, so if we replaced every term
in the current block the the first term in the next block, we would only decrease
the sum of the series. But this latter sum is much easier to deal with:
  
 


X
1
1 1
1 1 1 1
1 1 1
1

+
+
+
+ + +
+ . . . = + + + . . . = .
n
2
4
4
8
8
8
8
2
2 2
n=1
P
Therefore the harmonic series n=1 diverges.
Exercise: Determine the convergence of

1
1

n1+ n

P (x)
Exercise: Let Q(x)
be a rational function with deg Q deg P = 1. Show that
P P (n)
6
n=N Q(n) diverges.

We deduce the following result.

Proposition 11.13. For a rational function
verges iff deg Q deg P 2.

P (x)
Q(x) ,

the series

P (n)
n=N Q(n)

Proof. Left to the reader.

con

4.2. The Condensation Test.

The apparently ad hoc argument used to prove the divergence of the harmonic
series can be adapted to give the following useful test, due to A.L. Cauchy.
P
Theorem 11.14. (Condensation Test) Let n=1 an be an infinite series such
that an an+1
n N. Then: P
P 0 for all
P

a) We have n=1 P
an n=0 2n a2n 2 n=1 an .
P
b) Thus the series n an converges iff the condensed series n 2n a2n converges.
Proof. We have

X
an = a1 + a2 + a3 + a4 + a5 + a6 + a7 + a8 + . . .
n=1

a1 + a2 + a2 + a4 + a4 + a4 + a4 + 8a8 + . . . =

2n a2n

n=0

= (a1 + a2 ) + (a2 + a4 + a4 + a4 ) + (a4 + a8 + a8 + a8 + a8 + a8 + a8 + a8 ) + (a8 + . . .)

(a1 + a1 ) + (a2 + a2 + a3 + a3 ) + (a4 + a4 + a5 + a5 + a6 + a6 + a7 + a7 ) + (a8 + . . .)

X
=2
an .
n=1
6Take N larger than any of the roots of Q(x), so that every term in the sum is well defined.

4. SERIES WITH NON-NEGATIVE TERMS II: CONDENSATION AND INTEGRATION 243

This establishes part a), and part b) follows immediately.

The Cauchy Condensation Test is, I think, an a priori interesting result: it says
that, under the given hypotheses, in order to determine whether a series converges
we need to know only a very sparse set of the terms of the series whatever is
happening in between a2n and a2n+1 is immaterial, so long as the sequence remains
decreasing. This is a very curious phenomenon, and of couse without the hypothesis
that the terms are decreasing, nothing like this could hold.
On the other hand, it may be less clear thatP
the Condensation Test is of any
practical use: afterP
all, isnt the condensed series n 2n a2n more complicated than
the original series n an ? In fact the opposite is often the case: passing from the
given series to the condensed series preserves the convergence or divergence but
tends to exchange subtly convergent/divergent series for more obviously (or better:
more rapidly) converging/diverging series.
Example: Fix a real number p and consider the p-series7
is to find all values of p for which the series converges.

1
n=1 np .

Our task

Step 1: The sequence an = n1p has positive terms. The terms are decreasing iff
the sequence np is increasing iff p > 0. So we had better treat the cases p 0 separately. First, if p < 0, then limn n1p = limn n|p| = , so the p-series diverges
P
P
by the nth term test. Second, if p = 0 then our series is simply n n10 = n 1 = .
So the p-series obviously diverges when p 0.
Step 2: Henceforth we assume P
p > 0, so that the hypotheses
Cauchys
P n nofp
P nCondenp
np
sation
Test
apply.
We
get
that
n
converges
iff
2
(2
)
=
=
n
n
n2 2
P 1p n
(2
)
converges.
But
the
latter
series
is
a
geometric
series
with
geometric
n
ratio r = 21p , so it converges iff |r| < 1 iff 2p1 > 1 iff p > 1.
Thus we have proved the following important result.
P
Theorem 11.15. For p R, the p-series n n1p converges iff p > 1.
Example (p-series continued): Let
Pp > 1. By applying part b) of Cauchys Condensation Test we showed that n=1 n1p < . What about part a)? It gives an
explicit upper bound on the sum of the series, namely

X
X
X
1
1
n n p

2
(2
)
=
(21p )n =
.
p
n
1 21p
n=0
n=0
n=1

For instance, taking p = 2 we get

X
1
1

= 2.
2
n
1

212
n=1

Using a computer algebra package I get

1

1024
X
n=1

1
= 1.643957981030164240100762569 . . . .
n2

7Or sometimes: hyperharmonic series.

244

11. INFINITE SERIES

P
So it seems like n=1 n12 1.64, whereas the Condensation Test tells us that it
is at most 2. (Note that since the terms are positive, simply adding up any finite
number of terms gives a lower bound.)
The following
exercise gives a technique for using the Condensation Test to esP
timate n=1 n1p to arbitrary accuracy.
Exercise: Let N be a non-negative integer.
a) Show that under the hypotheses of the Condensation Test we have

an

2n a2n+N .

n=0

n=2N +1

b) Apply part a) to show that for any p > 1,

X 1
1
Np
.
p
n
2
(1
21p )
N
n=2 +1

1
1
Example:
n=2 n log n . an = n log n is positive and decreasing (since its reciprocal
is positive and increasing) so the Condensation Test applies. We get that the
convergence of the series is equivalent to the convergence of
X
1 X1
2n
=
= ,
n
n
2 log 2
log 2 n n
n
P 1
so the series diverges. This is rather subtle: we know that for any  > 0, n nn

converges, since it is a p-series with p = 1 + . But log n grows more slowly than
n for any  > 0, indeed slowly enough so that replacing n with log n converts a
convergent series to a divergent one.

Exercise: Determine whether the series

1
n log(n!)

converges.

Exercise: Let P
p, q, r be positive real numbers.
a) Show that n n(log1 n)q converges iff q > 1.
P
1
b) Show that n np (log
n)q converges iff p > 1 or (p = 1 and q > 1).
P
c) Find all values of p, q, r such that n np (log n)q1(log log n)r converges.
The pattern of Exercise X.X can be continued indefinitely, giving series which converge or diverge excruciatingly slowly, and showing that the difference between
convergence and divergence can be arbitarily subtle.
4.3. The Integral Test.
Theorem 11.16. (Integral Test) Let f : [1, ) R be a positive decreasing
function, and for n Z+ put an = f (n). Then
Z

X
X
an
f (x)dx
an .
n=2

Thus the series

n=1

an converges iff the improper integral

R
1

f (x)dx converges.

4. SERIES WITH NON-NEGATIVE TERMS II: CONDENSATION AND INTEGRATION 245

Proof. This is a rare opportunity in analysis in which a picture supplies a

perfectly rigorous proof. Namely, we divide the interval [1, ) into subintervals
RN
[n, n + 1] for all n N and for any N N we compare the integral 1 f (x)dx with
the upper and lower Riemann sums associated to the partition {1, 2, . . . , N }. From
the picture one sees immediately that since f is decreasing the lower sum is
PN +1
PN
n=2 an and the upper sum is
n=1 an , so that
Z N
N
+1
N
X
X
an
f (x)dx
an .
n=2

n=1

Taking limits as N , the result follows.

Remark: The Integral Test is due to Maclaurin8 [Ma42] and later in more modern
form to Cauchy [Ca89].
Among series which arise naturally in undergraduate analysis, it usually holds that
the Condensation Test can be successfully applied to determine convergence / divergence of a series if and only if the Integral Test can be successfully applied.9
Example:
Let us use the Integral Test to determine the set of p > 0 such
that
R dx
P 1
converges.
Indeed
the
series
converges
iff
the
improper
integral
is
p
n n
1 xp
finite. If p 6= 1, then we have
Z
x1p x=
dx
=
|
.
p
x
1 p x=1
1
The upper limit is 0 if p 1 < 0 p > 1 and is if p < 1. Finally,
Z
dx
= log x|
x=1 = .
x
1
So, once again, the p-series diverges iff p > 1.
Exercise: Verify that all of the above examples involving the Condensation Test
can also be done using the Integral Test.
Given the similar applicability of the Condensation and Integral Tests, it is perhaps
not so surprising that many texts content themselves to give one or the other. In
calculus texts, one almost always finds the Integral Test, which is logical since often
integration and then improper integation are covered earlier in the same course in
which one studies infinite series. In elementary analysis courses one often develops
sequences and series before the study of functions of a real variable, which is logical because a formal treatment of the Riemann integral is necessarily somewhat
involved and technical. Thus many of these texts give the Condensation Test.
The Condensation Test and the Integral Test have a similar range of applicability: in most textbook examples, where one test succeeeds, so will the other.
From an aesthetic standpoint, the Condensation Test is more appealing (to me).
On the other hand, under a mild additional hypothesis the Integral Test can be
used to give asymptotic expansions for divergent series. Our treatment of the
8Colin Maclaurin, 1698-1746
9Why this should be so is not clear to me: the observation is purely empirical.

246

11. INFINITE SERIES

next two results owes a debt to [CdD].

Let us first establish some notation: Suppose that f, g : [a, ) R and that
g(x) 6= 0 for all sufficiently large x. We write f g and say f is asymptotic to
(x)
g if limx fg(x)
= 1. Similarly, given sequences {an }, {bn } with bn 6= 0 for all
sufficiently large n, we write an bn if limn abnn = 1.
Exercise: Let f and g be nonzero polynomial functions. Show that f g iff f
and g have the same degree and the same leading coefficient.
Lemma 11.17. Let {an } and {bn } be two sequences of positive real numbers
P
P
PN
PN
with an bn and n an = . Then n bn = and n=1 an n=1 bn .
P
Proof. By the Limit Comparison Test, n an = . Now fix  > 0 and choose
K Z+ such that for all n K we have an (1 + )bn . Then for N K,
N
X

an =

K1
X

n=1

K1
X
n=1

an

an +

n=1
K1
X

N
X

an

(1 + )bn

n=1

N
X

an +

n=1

n=K

K1
X

N
X

(1 + )bn

n=K

(1 + )bn = C,K + (1 + )

n=1

N
X

bn ,

n=1

PN
say, where C,K does not depend on N . Dividing both sides by n=1 bn and using
PN
PN
an
limN n=1 bn = , we find that Pn=1
is at most 1 + 2 for all sufficiently
N
n=1 bn
P
P
large N . Because our hypotheses are symmetric in n an and n bn , we also have
that

PN
n=1 bn
PN
n=1 an

is at most 1 + 2 for all suffiicently large N . It follows that

PN
an
lim Pn=1
= 1.
N
N
n=1 bn


Theorem P
11.18. Let f : [1, ) (0, ) be continuous and monotone. Suppose the series n f (n) diverges and that as x , f (x) f (x + 1). Then
Z N
N
X
f (n)
f (x)dx.
n=1

Proof.
R n+1 Case 1: Suppose f is increasing. Then, for n x n + 1, we have
f (n) n f (x)dx f (n + 1), or
R n+1
f (x)dx
f (n + 1)

.
1 n
f (n)
f (n)
By assumption we have
lim

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

so by the Squeeze Principle we have

Z n+1
(82)
f (x)dx f (n).
n

5. SERIES WITH NON-NEGATIVE TERMS III: RATIOS AND ROOTS

247

R n+1
Applying Lemma 11.17 with an = f (n) and bn = n f (x)dx, we conclude
Z N +1
N Z k+1
N
X
X
f (x)dx =
f (x)dx
f (n).
1

k=1

n=1

Further, we have
R N +1
1

lim

RN

f (x)dx

f (x)dx

f (N + 1)
= lim
= 1,
N f (N )

where in the starred equality we have applied LHopitals Rule and then the Fundamental Theorem of Calculus. We conclude
Z N
Z N +1
N
X
f (x)dx
f (x)dx
f (n).
1

n=1

Case 2: Suppose f is decreasing. Then for n x n + 1, we have

Z n+1
f (n + 1)
f (x)dx f (n),
n

or

R n+1
f (x)dx
f (n + 1)
n
1.
f (n)
f (n)
Once again, by our assumption that f (n) f (n + 1) and the Squeeze Principle we
get (82); the remainder of the proof proceeds as in the previous case.

5. Series With Non-Negative Terms III: Ratios and Roots
P
We continue our analysis of series n an with an 0 for all n. In this section we
introduce two important tests based on a very simple yet powerful idea: if for
sufficiently large n an is bounded above by a non-negative constant M times rn for
0 rP
< 1, then the series converges by comparison to the convergent geometric
series n M rn . Conversely, if for sufficiently large n an is bounded below by a
positive constant M times rn for
P r 1, then the series diverges by comparison to
the divergent geometric series n M rn .
5.1. The Ratio Test.
P
Theorem 11.19. (Ratio Test) Let n an be a series with an > 0 for all n.
a) Suppose there exists N Z+ and 0 < r < 1 such that for all n N , aan+1
r.
n
P
Then the series n an converges.
b) Suppose there exists N Z+ and r 1 such that for all n N , aan+1
r. Then
n
P
the series n an diverges.
c) The hypothesis of part a) holds if = limn aan+1
exists and is less than 1.
n
exists
and is greater than 1.
d) The hypothesis of part b) holds if = limn aan+1
n
Proof. a) Our assumption is that for all n N , aan+1
r < 1. Then
n
an+2 an+1
2
= an+1 an r . An easy induction argument shows that for all k N,

an+2
an

aN +k
rk ,
aN
so
aN +k aN rk .

248

11. INFINITE SERIES

Summing these inequalities gives

X
X
X
ak =
aN +k
aN rk < ,
k=N

k=0

k=0

so the series n an converges by comparison.

r 1. As above, it
b) Similarly, our assumption is that for all n N , aan+1
n
follows that for all k N,
aN +k
rk ,
aN
so
aN +k aN rk aN > 0.
It follows that an 9 0, so the series diverges by the Nth Term Test.
We leave the proofs of parts c) and d) as exercises.

Exercise: Prove parts c) and d) of Theorem 11.19.
P n
Example: Let x > 0. We will show that the series n=0 xn! converges. (Recall
we showed this earlier when x = 1.) We consider the quantity
an+1
=
an
It follows that limn

an+1
an

xn+1
(n+1)!
xn
n!

x
.
n+1

= 0. Thus the series converges for any x > 0.

5.2. The Root Test.

In this section we give a variant P
of the Ratio Test. Instead of focusing on the
property that the geometric series n rn has constant ratios of consecutive terms,
we observe that the nth root of the nth term is equal to r. Suppose now that
1
P
n
n an is a series with non-negative terms such that an r for some r < 1. Raising both sides to the nth power gives an rn , and once again we find that the
series converges by comparison to a geometric series.
P
Theorem 11.20. (Root Test) Let n an be a series with an 0 for all n.
1

a) Suppose therePexists N Z+ and 0 < r < 1 such that for all n N , ann r.
Then the series n an converges.
1

n
b) Suppose
P that for infinitely many positive integers n we have an 1. Then the
series n an diverges.
1

c) The hypothesis of part a) holds if = limn ann exists and is less than 1.
1

d) The hypothesis of part b) holds if = limn ann exists and is greater than 1.
Exercise: Prove Theorem 11.20.
5.3. Ratios versus Roots.
It is a fact a piece of calculus folklore that the Root Test is stronger than
the Ratio Test. That is, whenever the ratio test succeeds in determining the convergence or divergence of a series, the root test will also succeed.
In order to explain this result we need to make use of the limit infimum and limit

5. SERIES WITH NON-NEGATIVE TERMS III: RATIOS AND ROOTS

249

supremum. First we recast the ratio and root tests in those terms.
Exercise: Let

Exercise: Let

an be a series with positive terms. Put

an+1
an+1
= lim inf
, = lim sup
.
n
an
an
n
P
a) Show that if < 1, the series P n an converges.
b) Show that if > 1 the series n an diverges.
n

an be a series with non-negative terms. Put

1

= lim sup ann .

n

P
a) Show that if < 1, the series n an converges.
P
b) Show that if > 1, the series n an diverges.10
Exercise: Consider the following conditions on a real sequence {xn }
n=1 :
(i) lim supn xn > 1.
(ii) For infinitely many n, xn 1.
(iii) lim supn xn 1.
a) Show that (i) = (ii) = (iii) and that neither implication can be reversed.
b) Explain why the result of part b) of the previous Exercise is weaker than part
b) of Theorem 11.20.
1
P
c) Give anPexample of a non-negative series n an with = lim supn ann = 1
such that n an = .
P
Proposition 11.21. For any series n an with positive terms, we have
1
1
an+1
an+1
= lim inf
= lim inf ann lim sup ann = lim sup
.
n
n
an
an
n
n
Exercise: Let A and B be real numbers with the following property: for any real
number r, if A < r then B r. Show that B A.
Proof. Step 1: For any sequence {xn }, lim inf xn lim sup xn , hence .
Step 2: We show that . For this, suppose r > , so that for all sufficiently
large n, aan+1
r. As in the proof of the Ratio Test, we have an+k < rk an for all
n
k N. We may rewrite this as
a 
n
an+k < rn+k n ,
r
or
1
 a  n+k
1
n
n+k
an+k
<r n
.
r
Now
1
 a  n+k
1
1
n
k
= lim sup ann = lim sup an+k
lim sup r n
= r.
r
n
k
k
By the preceding exercise, we conclude .
Step 3: We must show that . This is very similar to the argument of Step 2,
and we leave it as an exercise.

10This is not a typo: we really mean the limsup both times, unlike in the previous exercise.

250

11. INFINITE SERIES

Exercise: Give the details of Step 3 in the proof of Proposition 11.21.

P
Now let
n an be a series which the Ratio Test succeeds in showing is convergent: that is, < 1. Then by Proposition 11.21, we have 1, so the Root
Test also shows that the series is convegent. Now suppose that the Ratio Test
succeeds in showing that the series is divergent: that is > 1. Then > 1,
so the Root Test also shows that the series is divergent.
P
n
Exercise: Consider the series n 2n+(1) .
a) Show that = 18 and = 2, so the Ratio Test fails.
b) Show that = = 12 , so the Root Test shows that the series converges.
Exercise: Construct further examples of series for which the Ratio Test fails but
the Root Test succeeds to show either convergence or divergence.
Warning: The sense in which the Root Test is stronger than the Ratio Test is
a theoretical one. For a given relatively benign series, it may well be the case that
the Ratio Test is easier to apply than the Root Test, even though in theory whenever the Ratio Test works the Root Test must also work.
P 1
Example: Consider again the series
n=0 n! . In the presence of factorials one
should always attempt the Ratio Test first. Indeed
an+1
n!
1
1/(n + 1)!
= lim
= lim
= 0.
= lim
n an
n (n + 1)n!
n n + 1
n
1/n!
lim

Thus the Ratio Test limit exists (no need for liminfs or limsups) and is equal to 0,
so the series converges. If instead we tried the Root Test we would have to evaluate
1
1 n
limn ( n!
) . This is not so bad if we keep our head e.g. one can show that for
1
1
1 n
any fixed R > 0 and sufficiently large n, n! > Rn and thus ( n!
) ( R1n ) n = R1 .
Thus the root test limit is at most R1 for any positive R, so it is 0. But this is
elaborate compared to the Ratio Test computation, which was immediate.
Turning these ideas around, Proposition 11.21 can be put to the following
sneaky use.
Corollary 11.22. Let {an }
n=1 be a sequence of positive real numbers. Assume
that limn

an+1
an

L [0, ]. Then also limn ann = L.

Proof. Indeed, the hypothesis gives that for the infinite series
= L, so by Proposition 11.21 we must also have = L.

an we have


Exercise: Use Corollary 11.22 to evaluate the following limits:

1
a) limn n n .

b) For R, limn n n .
1
c) limn (n!) n .
6. Absolute Convergence
6.1. Introduction to absolute convergence.
We turn now to the serious study of series with both positive and negative terms.

6. ABSOLUTE CONVERGENCE

251

It turns out that under one relatively mild additional hypothesis, virtually all of
our work on series with non-negative terms can be usefully applied in this case. In
this section we study this wonderful hypothesis: absolute convergence. (In the next
section we get really serious by studying series in the absence of absolute convergence. This will lead to surprisingly delicate and intricate considerations.)
P
P
A
P real series n an is absolutely convergent if n |an | converges. Note that
it is convergent
n |an | is a series with non-negative terms, so to decide whether
P
we may use allP
the tools of the last three sections. A series n an which converges
but for which n |an | diverges is said to be nonabsolutely convergent.11
The P
terminology absolutely convergent suggests that the convergence
P of the series n |an | is somehow better than the convergence of the series n an . This is
indeed the case, although it is not obvious. But the following result already clarifies
matters a great deal.
Proposition 11.23. Every absolutely convergent real series is convergent.
Proof. We shall give two proofs of this
result. P
P important
P
First Proof : Consider
the
three
series
a
,
|a
n |. Our
n n
n n | and
n an + |aP
P
hypothesis is that n |an | converges. But we claim that this implies that n an +
|an | converges as well. Indeed, consider the expression an + |an |: it is equal to
2an =P
2|an | when an is non-negative and 0 when
Pan is negative.PIn particular the
series
a
+
|a
|
has
non-negative
terms
and
n
n n
n an +
P
P|an | n 2|an | < . So
a
+
|a
|
converges.
By
the
Three
Series
Principle,
n
n n
n an converges.
Second Proof : The above argument
is
clever

maybe
too clever! Lets try
P
something more fundamental: since
|a
|
converges,
for
every
 > 0 there exists
P n n
N Z+ such that for all n N , n=N |an | < . Therefore for all k 0,
|

N
+k
X
n=N

and

an |

N
+k
X
n=N

|an |

|an | ,

n=N

an converges by the Cauchy criterion.


P
P
P
Exercise: If n=0 an is absolutely convergent, show that | n=0 an | n=0 |an |.
n

Exercise: Find a sequence

P {an }n=1 of rational numbers such that
a rational number but n=1 an is an irrational number.

n=1

|an | is

The main idea is that when we try to extend our rich array of convergence tests for
non-negative series to series with both positive and negative terms, a sufficient (and
often necessary) hypothesis is that the series be not just convergent but absolutely
convergent.

11We warn the reader that the more standard terminology is conditionally convergent.
We will later on give a separate definition for conditionally convergent and then it will be a
theorem that a real series is conditionally convergent if and only if it is nonabsolutely convergent.
The reasoning for this which we admit will seem abstruse at best to our target audience is
that in functional analysis one studies convergence and absolute convergence of series in a more
general context, such that nonabsolute converge and conditional convergence may indeed differ.

252

11. INFINITE SERIES

Exercise: State and prove a Comparison Test and Limit Comparison Test for absolute convergence.

Theorem 11.24. (Ash [As12]) Let {an }

n=1 and {bn }n=1 be real sequences
+
such that for all n Z we have bn 6= 0. Suppose:
(i) limn abnn = 1, and
P
P
(ii) n=1 an diverges and n=1 bn converges.
P
an bn
= 0.
Then:
n=1 |bn | = and limn
bn
P
Proof.
If n=1 |bn | <
then applying the P
limit comparison test to the absoP,
P

lute series n=1 |an | and n=1 |bn |, we get that n=1 an is absolutely convergent
and thus convergent, contradicting (ii). Moreover

lim

an bn
an
= lim
1 = 0.
n bn
bn

limn | abnn |

= L (0, ), then still

Exercise: Show that if we weaken (i) to
P
n=1 bn is not absolutely convergent.
As an example of how Theorem 11.23 may be combined with the previous tests
to give tests for absolute convergence, we record the following result.
P
Theorem 11.25. (Ratio & Root Tests for Absolute Convergence) Let n an
be a real series.
a) Assume an 6= 0 for all n. If there
Pexists 0 r < 1 such that for all sufficiently
|

r,
then
the
series
large n, | aan+1
n an is absolutely convergent.
n
b) Assume an 6= 0 for allP
n. If there exists r > 1 such that for all sufficiently large
n, | aan+1
| r, the series n an is divergent.
n
1
c) If there exists r < 1 such that for all sufficiently large n, |an | n r, the series
P
n an is absolutely convergent.
1
d) If there are infinitely many n for which |an | n 1, then the series diverges.
Proof. Parts a) and
Pc) are immediate: applying Theorem 11.19 (resp. Theorem 11.20) we find that P n |an | is convergent and the point is that by Theorem
11.23, this implies that n an is convergent.
P There is something to say in parts
P b) and d), because in general just because
|a
|
=

does
not
imply
that
n
n
n an diverges. (We will study this subtlety
later on in detail.) But recall that whenever
the Ratio or Root tests establish the
P
divergence of a non-negative series n bn , they do so by showing that bn 9 0.
Thus
P under the hypotheses of parts b) and d) we have |an | 9 0, hence also an 9 0
so n an diverges by the Nth Term Test (Theorem 11.4).

P
In particular, for a real series n an define the following quantities:
an+1
| when it exists,
an
an+1
= lim inf |
|,
n
an
an+1
= lim sup |
|,
an
n

= lim |
n

= lim |an | n when it exists,

n

6. ABSOLUTE CONVERGENCE

253

= lim sup |an | n ,

n

and then all previous material on Ratio and Root Tests applies to all real series.
P
P
Theorem 11.26. Let Pn=0 an = A and n=0 bn = B be two absolutely
P conn
vergent series, and let cn = k=0 ak bnk . Then the Cauchy product series n=0 cn
is absolutely convergent, with sum AB.
Proof. We have proved this result already when an , bn 0 for all n. We wish,
of course, to reduce to that case. As far as the convergence of the Cauchy product,
this is completely straightforward: we have

|cn | =

n=0

X
n
X
n
X
X
|
ak bnk |
|ak ||bnk | < ,
n=0 k=0

n=0 k=0

P Pn
the last inequality following from the fact
|ak ||bnk | is the Cauchy
Pthat n=0 Pk=0

product of P
the two non-negative series n=0 |an | and n=0 |bn |, hence it converges.
P
Therefore n |cn | converges by comparison, so the Cauchy product series n cn
converges.
P
We now wish to show that limN CN = n=0 cn = AB. Recall the notation
X
N =
ai bj = (a0 + . . . + aN )(b0 + . . . + bN ) = AN BN .
0i,jN

We have
|CN AB| |N AB| + |N CN |
= |AN BN AB| + |a1 bN | + |a2 bN 1 | + |a2 bN | + . . . + |aN b1 | + . . . + |aN bN |

!
!

X
X
X
X
|AN BN AB| +
|an |
|bn | +
|bn |
|an | .
n=0

n=0

nN

nN

Fix  > 0; since AN BN AB, for sufficiently large N |AN BN AB| < 3 . Put
A=

|an |, B =

n=0

|bn |.

n=0

By the Cauchy criterion, for sufficiently large N we have

P

nN |an | < 3B and thus |CN AB| < .

nN

|bn | <


3A

and


6.2. Cauchy products II: Mertenss Theorem.

While the proof of Theorem 11.26 may seem rather long, it is in fact a rather
straightforward argument: one shows that the difference between the box product and the partial sums of the Cauchy product becomes negligible as N tends
to infinity. In less space but with a bit more finesse, one can prove the following
stronger result, a theorem of F. Mertens [Me72].12
P
Theorem 11.27.
P(Mertens Theorem) Let n=0 an = A be an absolutely convegent P
series and n=0 bn = B be a convergent series. Then the Cauchy product

series n=0 cn converges to AB.

12Franz Carl Joseph Mertens, 1840-1927

254

11. INFINITE SERIES

Proof. (Rudin [R, Thm. 3.50]): define (as usual)

AN =

N
X

an , BN =

n=0

N
X

bn , CN =

n=0

N
X

cn

n=0

and also (for the first time)

n = Bn B.
Then for all N N,
CN = a0 b0 + (a0 b1 + a1 b0 ) + . . . + (a0 bN + . . . + aN b0 )
= a0 BN + a1 BN 1 + . . . + aN B0
= a0 (B + N ) + a1 (B + N 1 ) + . . . + aN (B + 0 )
= AN B + a0 N + a1 N 1 + . . . + aN 0 = AN B + N ,
say, where N = a0 N + a1 N 1 + . . . + an 0 . Since our goal is to show that
CN AB
and we know that AN B AB, it suffices to show that N 0. Now,
P
put = n=0 |an |. Since BN B, N 0, and thus for any  > 0 we may choose

. Put
N0 N such that for all n N0 we have |n | 2
M = max |n |.
0nN0

By the Cauchy criterion, for all sufficiently large N , M

nN N0

|an | /2. Then

|N | |0 aN + . . . + N0 aN N0 | + |N0 +1 aN N0 1 + . . . + N a0 |

|0 aN + . . . + N0 aN N0 | +


M
2

nN N0

So N 0.

|an | +




+ = .
2
2 2


7. Non-Absolute Convergence
P
We say that
Pa real series n an is nonabsolutely convergent if the series converges but n |an | diverges, thus if it is convergent but not absolutely convergent.13
A series which is nonabsolutely convergent is a more delicate creature than any
we have studied thus far. A test which can show that a series is convergent but
nonabsolutely convergent is necessarily subtler than those of the previous section.
In fact the typical undergraduate student of calculus / analysis learns exactly one
such test, which we give in the next section.
13One therefore has to distinguish between the phrases not absolutely convergent and
nonabsolutely convergent: the former allows the possibility that the series is divergent, whereas
the latter does not. In fact our terminology here is not completely standard. We defend ourselves
grammatically: nonabsolutely is an adverb, so it must modify convergent, i.e., it describes
how the series converges.

7. NON-ABSOLUTE CONVERGENCE

255

7.1. The Alternating Series Test.

Consider the alternating harmonic series

X
(1)n+1
1 1
= 1 + ....
n
2 3
n=1

Upon taking the absolute value of every term we get the usual harmonic series,
which diverges, so the alternating harmonic series is not absolutely convergent.
However, some computations with partial sums suggests that the alternating harmonic series is convergent, with sum log 2. By looking more carefully at the partial
sums, we can find a pattern that allows us to show that the series does indeed
converge. (Whether it converges to log 2 is a different matter, of course, which we
will revisit much later on.)
It will be convenient to write an = n1 , so that the alternating harmonic series
n+1
P
is n (1)
n+1 . We draw the readers attention to three properties of this series:
(AST1) The terms alternate in sign.
(AST2) The nth term approaches 0.
(AST3) The sequence of absolute values of the terms is weakly decreasing:
a1 a2 . . . an . . . .
These are the clues from which we will make our case for convergence. Here it is:
consider the process of passing from the first partial sum S1 = 1 to S3 = 1 21 + 13 =
5
6 . We have S3 1, and this is no accident: since a2 a3 , subtracting a2 and then
adding a3 leaves us no larger than where we started. But indeed this argument is
valid in passing from any S2n1 to S2n+1 :
S2n+1 = S2n1 (a2n a2n+1 ) S2n1 .
Thus the sequence of odd-numbered partial sums {S2n1 } is decreasing. Moreover,
S2n+1 = (a1 a2 ) + (a3 a4 ) + . . . + (a2n1 | |a2n ) + a2n1 0.
Therefore all the odd-numbered terms are bounded below by 0. By the Monotone
Sequence Lemma, the sequence {S2n+1 } converges to its greatest lower bound, say
Sodd . On the other hand, just the opposite sort of thing happens for the evennumbered partial sums:
S2n+2 = S2n + a2n+1 a2n+2 S2n
and
S2n+2 = a1 (a2 a3 ) (a4 a5 ) . . . (a2n a2n+1 |) a2n+2 a1 .
Therfore the sequence of even-numbered partial sums {S2n } is increasing and
bounded above by a1 , so it converges to its least upper bound, say Seven . Thus we
have split up our sequence of partial sums into two complementary subsequences
and found that each of these series converges. Now the sequence {Sn } converges iff
Sodd = Seven , and the inequalities
S2 S4 . . . S2n S2n+1 S2n1 . . . S3 S1

256

11. INFINITE SERIES

show that Seven Sodd . Moreover, for any n Z+ we have

Sodd Seven S2n+1 S2n = a2n+1 .
Since a2n+1 0, we conclude Sodd = Seven = S, i.e., the series converges.
Further, since for all n we have S2n S2n+2 S S2n+1 , it follows that
|S S2n | = S S2n S2n+1 S2n = a2n+1
and similarly
|S S2n+1 | = S2n+1 S S2n+1 S2n+2 = a2n+2 .
P
Thus the error in cutting off the infinite sum n=1 (1)n+1 |an | after N terms is in
absolute value at most the absolute value of the next term: aN +1 .
Of course in all this we never used that an = n1 but only that we had a series
satisfying (AST1) (i.e., an alternating series), (AST2) and (AST3). Therefore the
preceding arguments have in fact proved the following more general result, due
originally due to Leibniz.
Theorem 11.28. Let {an }
n=1 be a sequence of non-negative real numbers which
is weakly decreasing and such that limn
P an = 0. Then:
a) The associated alternating series n (1)n+1 an converges.
b) For N Z+ , put
(83)

EN = |(

n+1

(1)

an ) (

n=1

N
X

(1)n+1 an )|.

n=1

Then we have the error estimate

EN aN +1 .
Exercise: Let p R: Show that the alternating p-series
(i) divergent if p 0,
(ii) nonabsolutely convergent if 0 < p 1, and
(iii) absolutely convergent if p > 1.

n=1

(1)n+1
np

is:

P (x)
be a rational function. Give necessary and sufficient condiExercise: Let Q(x)
P
n P (x)
tions for n (1) Q(x) to be nonabsolutely convergent.

For any convergent series

n=1

an = S, we may define EN as in (83) above:

EN = |S

N
X

an |.

n=1

Then because the series converges to S, limN EN = 0, and conversely: in other

words, to say that the error goes to 0 is a rephrasing of the fact that the partial
sums of the series converge to S. Each of these statements is (in the jargon of
mathematicians working in this area) soft: we assert that a quantity approaches 0
and N , so that in theory, given any  > 0, we have EN <  for all suffuciently
large N . But as we have by now seen many times, it is often possible to show that
EN 0 without coming up with an explicit expression for N in terms of . But
this stronger statement is exactly what we have given in Theorem 11.28b): we have

7. NON-ABSOLUTE CONVERGENCE

257

given an explicit upper bound on EN as a function of N . This type of statement

is called a hard statement or an explicit error estimate: such statements tend
to be more difficult to come by than soft statements, but also more useful to have.
Here, as long as we can similarly make explicit how large N has to be in order for
aN to be less than a given  > 0, we get a completely explicit error estimate and
can use this to actually compute the sum S to arbitrary accuracy.
n+1
P
to three decimal place accuracy. By TheExample: We compute n=1 (1)n
1
orem 11.28b), it is enough to find N Z+ such that aN +1 = N1+1 < 1000
. We may
take N = 1000. Therefore
1000

X
1
(1)n+1 X (1)n+1

|
.
|
n
n
1001
n=1
n=1

Using a software package, we find that

1000
X
n=1

(1)n+1
= 0.6926474305598203096672310589 . . . .
n

Again, later we will show that the exact value of the sum is log 2, which my software
package tells me is14
log 2 = 0.6931471805599453094172321214.
Thus the actual error in cutting off the sum after 1000 terms is
E1000 = 0.0004997500001249997500010625033.
It is important to remember that this and other error estimates only give upper
bounds on the error: the true error could well be much smaller. In this case we
1
were guaranteed to have an error at most 1001
and we see that the true error is
about half of that. Thus the estimate for the error is reasonably accurate.
Note well that although the error estimate of Theorem 11.28b) is very easy
to apply, if an tends to zero rather slowly (as in this example), it is not especially
efficient for computations. For instance, in order to compute the true sum of the alternating harmonic series to six decimal place accuracy using this method, we would
need to add up the first million terms: thats a lot of calculation. (Thus please be
assured that this is not the way that a calculator or computer would compute log 2.)
n
P
Example: We compute n=0 (1)
to six decimal place accuracy. Thus we need to
n!
1
choose N such that aN +1 = (N +1)! < 106 , or equivalently such that (N +1)! > 106 .
A little calculation shows 9! = 362, 880 and 10! = 3, 628, 800, so that we may take
N = 9 (but not N = 8). Therefore

9
X
(1)n X (1)n
1
|

|<
< 106 .
n!
n!
10!
n=0
n=0

Using a software package, we find

9
X
(1)n
= 0.3678791887125220458553791887.
n!
n=0
14Yes, you should be wondering how it is computing this! More on this later.

258

11. INFINITE SERIES

In this case the exact value of the series is

1
== 0.3678794411714423215955237701
e
so the true error is
E9 = 0.0000002524589202757401445814516374,
which this time is only very slightly less than the guaranteed
1
= 0.0000002755731922398589065255731922.
10!
7.2. Dirichlets Test.
What lies beyond the Alternating Series Test? We present one more result, an
elegant (and useful) test due originally to Dirichlet.15
P
P
Theorem 11.29. (Dirichlets Test) Let n=1 an , n=1 bn be two infinite series. Suppose that:
(i) The partial sums Bn = b1 + . . . + bn are bounded.
P
(ii) The sequence an is decreasing with limn an = 0. Then n=1 an bn is convergent.
Proof. Let M R be such that Bn = b1 + . . . + bn M for all n Z+ . For

. Then for n > m N ,
 > 0, choose N > 1 such that aN 2M
|

n
X

n
X

ak bk | = |

k=m

ak (Bk Bk1 )| = |

k=m

=|

n1
X

k=m
n1
X

n
X

ak Bk

k=m

n1
X

ak+1 Bk |

k=m1

(ak ak+1 )Bk + an Bn am Bm1 |

!

|ak ak+1 ||Bk |

+ |an ||Bn | + |am ||Bm1 |

k=m

M(

n1
X
k=m

Therefore

n1
X

|ak ak+1 |) + |an | + |am | = M

!
(ak ak+1 ) + an + am

k=m

= M (am an + an + am ) = 2M am 2M aN < .
a
bn converges by the Cauchy criterion.
n
n

In the preceding proof, without saying what we were doing, we used the technique
of summation by parts.
If we take bn = (1)n+1 , then B2n+1 = 1 for all n and B2n = 0 for all n,
so {bn } has bounded partial sums. Applying Dirichlets Test with a sequence
an which decreases
to 0 and with this sequence {bn }, we find that the series
P
P
n+1
an converges. We have recovered the Alternating Series
n an bn =
n (1)
Test!
In fact Dirichlets Test yields the following Almost Alternating Series Test:
let {an } be a sequence decreasing to 0, and for all n let bn {1} be a sign sequence which is almost alternating in the sense P
that the sequence of partial
sums Bn = b1 + . . . + bn is bounded. Then the series n bn an converges.
15Johann Peter Gustav Lejeune Dirichlet, 1805-1859

7. NON-ABSOLUTE CONVERGENCE

259

Exercise: Show that Dirichlets generalization of the Alternating Series Test is

as strong as possible in the following sense: if {bn } is a sequence of elements,
each 1, such that the sequence of partial sums Bn = b1P
+ . . . + bn is unbounded,
then there is a sequence an decreasing to zero such that n an bn diverges.
Exercise:
a) Use the trigonometric identity
cos n =

sin(n + 12 ) sin(n 21 )
2 sin( 21 )

to show that the sequence Bn = cos 1 + . . . + cos n

Pis bounded.
cos n
b) Apply Dirichlets
Test
to
show
that
the
series
converges.
n=1 n
P cos n
c) Show that n=1 n is not absolutely convergent.
Exercise: Show that

n=1

sin n
n

is nonabsolutely convergent.

Remark: Once we know about series of complex numbers and Eulers formula
eix = cos x + i sin x, we will be able to give a trigonometry-free proof of the preceding two exercises.
Dirichlet himself applied his test to establish the convergence of a certain class
of series of a mixed algebraic and number-theoretic nature. The analytic properties of these series were used to prove his celebrated theorem on prime numbers in
arithmetic progressions. To give a sense of how influential this work has become, in
modern terminology Dirichlet studied the analytic properties of Dirichlet series
associated to nontrivial Dirichlet characters. For more information on this work,
the reader may consult (for instance) [DS].
7.3. Cauchy Products III: A Divergent Cauchy Product.
Let us give an example due to Cauchy! of a Cauchy product of two nonabsolutely
P
P
P (1)n
convergent series which fails to converge. Take n=0 an = n=0 bn = n=0
.
n+1
The nth term in the Cauchy product is
X
1
1

cn =
(1)i (1)j
.
i+1 j+1
i+j=n
Now (1)i (1)j = (1)i+j = (1)n , so cn is equal to (1)n times a sum of positive
1
1
1
n+1
terms. Since i, j n, i+1
, j+1
, and thus each term in cn has absolute
1
1
2
value at least ( n+1 ) = n+1 . Since we are summing from i = 0 to n there are
n +P1 terms, all of the same size, we find |cn | 1 for all n. Thus the general term
of n cn does not converge to 0, so the series diverges.
7.4. Decomposition into positive and negative parts.
For a real number r, we define its positive part
r+ = max(r, 0)

260

11. INFINITE SERIES

and its negative part

r = min(r, 0).
Exercise: Let r be a real number. Show:
a) r = r+ r .
b) |r| = r+ + r .
For any real series

an we have a decomposition
X
X
X
an =
a+
a
n
n,
n

P
P
at least if all three series converge.
Let us call n a+
n and P n an the positive
P
part and negative part of n an . Let us now suppose that n an converges. By
the Three Series Principle there are two cases:
P +
P
P
P +

Case 1: Both
P n an and n an converge. Hence n |an | = n (an + an ) converges: i.e., n an is absolutely convergent.
P
P
P
P

Case 2: Both n a+
| = n a+
n and
n + an diverges:
n an diverge. Hence
n |an
P
indeed,
if it converged,
n an we would get that
P
P then adding and subtracting
2 n a+
and
2
converge,
contradiction.
Thus:
a
n
n n
P
Proposition 11.30. If a series nPis absolutely convergent, both its positive
and negative parts converge. If a series n is nonabsolutely convergent, then both
its positive and negative parts diverge.
P
Exercise: Let P
n an be a real series. P
P

a) Show that if Pn a+
n converges andP n an diverges then Pn an = .

b) Show that if n a+
n diverges and
n an converges then
n an = .
8. Power Series I: Power Series as Series
8.1. Convergence of Power Series.
P
n
Let {an }
n=0 be a sequence of real numbers. Then a series of the form
n=0 an x
is called a power series.
Thus, for instance, if we had an = 1 for all n we would
P
1
.
get the geometric series n=0 xn which converges iff x (1, 1) and has sum 1x
Pn
k
The nth partial sum of a power series is k=0 ak x , a polynomial in x. One
of the major themes of Chapter three will be to try to view power series as infinite
polynomials: in particular, we will regard x as a variable and be interested in the
propeties
continuity, differentiability, integrability, and so on of the function
P
f (x) = n=0 an xn defined by a power series.
P
However, if we want to regard the series n=0 an xn as a function of x, what is
its domain? The natural domain of a power series is the set of all values of x for
which the series converges. Thus the basic question about power series that we will
answer in this section is the following.
Question 11.31. For a P
sequence {an }
n=0 of real numbers, for which values of

x R does the power series n=0 an xn converge?

8. POWER SERIES I: POWER SERIES AS SERIES

261

There is one value of x for which the answer is trivial. Namely, if we plug in x = 0
to our general power series, we get

an 0n = a0 + a1 0 + a2 02 = a0 .

n=0

So every power series converges at least at x = 0.

Example 1: Consider the power series

n=0

n!xn . We apply the Ratio Test:

(n + 1)!xn+1
= lim (n + 1)|x|.
n
n
n!xn
The last limit is 0 if x = 0 and otherwise is +. Therefore the Ratio Test shows
that (as we already knew!) the series converges absolutely at x = 0 and diverges
at every nonzero x. So it is indeed possible for a power series
P to converge only at
x = 0. This is disappointing if we are interested in f (x) = n an xn as a function
of x, since in this case it is just the function from {0} to R which sends 0 to a0 .
There is nothing interesting going on here.
lim

Example 2: Consider

xn
n=0 n! .

lim |

We apply the Ratio Test:

xn+1
n!
|x|
|| | = lim
= 0.
n n + 1
(n + 1)! xn

So the power series converges for all x R and defines a function f : R R.

P 1 n
Example 3: Fix R (0, ); consider
n=0 Rn x . This is a geometric series
x
x
with geometric ratio = R
, so it converges iff || = | R
| < 1, i.e., iff x (R, R).
Example 4: Fix R (0, ); consider

1
n
n=1 nRn x .

We apply the Ratio Test:

nRn
|x|n+1
n+1 1
|x|
= |x| lim

=
.
n+1
n
n (n + 1)R
n
|x|
n
R
R
lim

Therefore the series converges absolutely when |x| < R and diverges when |x| > R.
We must look separately at the
P case |x| = R i.e., when x = R. When x = R, the
series is the harmonic series n n1 , hence divergent. But when x = R, the series
n
P
is the alternating harmonic series n (1)
n , hence (nonabsolutely) convergent. So
the power series converges for x [R, R).
P (1)n n
Example 5: Fix R (0, ); consider
n=1 nRn x . We may rewrite this seP
1
n
ries as n=1 nR
n (x) , i.e., the same as in Example 4 but with x replaced by x
throughout. Thus the series converges iff x [R, R), i.e., iff x (R, R].
Example 6: Fix R (0, ); consider

1
n
n=1 n2 Rn x .

n2 R n
|x|n+1
= |x| lim
2
n+1
n (n + 1) R
n
|x|n
lim

We apply the Ratio Test:

2
1
|x|
n+1

=
.
n
R
R

So once again the series converges absolutely when |x| < R, diverges when |x|
P> R,
and we must look separately at x = R. This time plugging in x = R gives n n12
n
P
which is a convergent p-series, whereas plugging in x = R gives n (1)
n2 : since

262

11. INFINITE SERIES

the p-series with p = 2 is convergent, the alternating p-series with p = 2 is absolutely convergent. Therefore the series converges (absolutely, in fact) on [R, R].
Thus the convergence set of a power series can take any of the following forms:

the single point {0} = [0, 0].

the entire real line R = (, ).
for any R (0, ), an open interval (R, R).
for any R (0, ), a half-open interval [R, R) or (R, R]
for any R (0, ), a closed interval [R, R].

In each case the set of values is an interval containing 0 and with a certain radius,
i.e., an extended real number R [0, ) such that the series definitely converges
for all x (R, R) and definitely diverges for all x outside of [R, R]. Our goal is
to show that this is the case for any power series.
This goal can be approached at various degrees of sophistication. At the calculus level, we have already said what is needed: we use the Ratio Test to see that
the convergence set isPan interval around 0 of a certain radius R. Namely, taking a
general power series n an xn and applying the Ratio Test, we find
an+1
|an+1 xn+1 |
= |x| lim
.
n an
n
|an xn |
lim

So if = limn aan+1
, the Ratio Test tells us that the series converges when
n
|x| < 1 i.e., iff |x| < 1 and diverges when |x| > 1 i.e., iff |x| > 1 . That
is, the radius of convergence R is precisely the reciprocal of the Ratio Test limit ,
1
with suitable conventions in the extreme cases, i.e., 01 = ,
= 0.
So what more is there to say or do? The issue here is that we have assumed
exists. Although this is usually the case in simple examples of
that limn aan+1
n
interest, it is certainly does not happen in general (we ask the reader to revisit X.X
for examples of this). This we need to take a different approach in the general case.
P
P
n
n
Lemma 11.32.
A > 0 and let
n an x be a power series. If
n an A
P Let
n
converges, then n x converges absolutely for all x (A, A).
P
Proof. Let 0 < BP< A. It is enough to
show n an B n is absolutely conP
vergent, for then so is n an (B)n . Since n an An converges, an An 0: by
omitting finitely many terms, we may assume |an An | 1 for all n. Since 0 < B
A < 1,




n
n
X
X
X B
B
|an B n | =
|an An |

< .
A
A
n
n
n

P

Theorem 11.33. Let n=0 an xn be a power series.

a) There exists R [0, ] such
P that:
(i) For all x with |x| < R, Pn an xn converges absolutely and
(ii) For all x with |x| > R, n an xn diverges.
b) If R = 0, then the power series converges only at x = 0.
c) If R = , the power series converges for all x R.

8. POWER SERIES I: POWER SERIES AS SERIES

263

d) If 0 < R < , the convergence set of the power series is either (R, R), [R, R),
(R, R] or [R, R].
a) Let R be the least upper bound of the set of x 0 such that
P Proof.
n
a
x
converges.
If y is such that |y| < R, then there exists A with |y| < A <
n n
P
R such that n an An converges, so by Lemma 11.32 the power series converges
absolutely on (A, A), so in particular it converges absolutely at y. Thus RPsatisfies
property (i). Similarly, suppose there exists y with |y| > R such that n an y n
converges. Then there exists A with R < A < |y| such that the power series
converges on (A, A), contradicting the definition of R.
We leave the proof of parts b) through d) to the reader as a straightforward exercise.

Exercise: Prove parts) b), c) and d) of Theorem 11.33.
P
P
Exercise: Let n=0 an xn and n=0 bn xn be two power seriesPwith positive radii
n
of convergence Ra and Rb . Let R = min(Ra , Rb ). Put cn = k=0 ak bnk . Show
that the formal identity
!
!

X
X
X
n
n
an x
bn x
=
cn xn
n=0

n=0

n=0

is valid for all x (R, R). (Suggestion: use past results on Cauchy products.)
The drawback of Theorem 11.33 is that it does not give an explicit description
of the radius of convergence R in terms of the coefficients of the power series, as is
|
exists. In order to achieve this
the case when the ratio test limit = limn |a|an+1
n|
in general, we need to appeal instead to the Root Test and make use of the limit
supremum. The following elegant result is generally attributed to J.S. Hadamard,16
who published it in 1888 [Ha88] and included it in his 1892 PhD thesis. This seems
remarkably late in the day for a result which is so closely linked to (Cauchys) Root
Test. It turns out that the result was established by our most usual suspect: it was
first proven by Cauchy in 1821 [Ca21] but apparently had been nearly forgotten.
P
Theorem 11.34. (Cauchy-Hadamard) Let n an xn be a power series, and put
1

= lim sup |an | n .

n

Then the radius of convergence of the power series is R =

converges absolutely for |x| < R and diverges for |x| > R.
1

1
:

that is, the series

1

Proof. We have lim supn |an xn | n = |x| lim supn |an | n = |x|. Put
R = 1 . If |x| < R, choose A such that |x| < A < R and then A0 such that
1
1
< A0 < .
R
A
1
Then for all sufficiently large n, |an xn | n A0 A < 1, so the series converges absolutely by the Root Test. Similarly, if |x| > R, choose A such that R < |x| < A and
then A0 such that
1
1
< A0 <
= .
A
R
=

16Jacques Salomon Hadamard (1865-1963)

264

11. INFINITE SERIES

1

Then there are infinitely

many non-negative integers n such that |an xn | n A0 A >
P
n
1, so the series n an x diverges: indeed an xn 9 0.

Here is a useful criterion for the radius of convergence of a power series to be 1.
Corollary 11.35. Let {an }
n=0 be a sequence
P of real numbers, and let R be
the radius of convergence of the power series n=0 an xn .
a) If {an } is bounded, then R 1.
b) If an 9 0, then R 1.
c) Thus if {an } is bounded but not convergent to zero, R = 1.
Exercise: Prove Corollary 11.35.
P
Theorem 11.36. Let n an xn be a power series
P with radius of convergence R.
Then, for any k Z, the radius of convergence of n nk an xn is also R.

k
k
= 1, by Corollary 11.22,
= limn n+1
Proof. Since limn (n+1)
n
nk
lim (nk )1/n = lim nk/n = 1.

(Alternately, take logarithms and apply LHopitals Rule.) Therefore




1
1
1
1
lim sup |an | n = lim sup |an | n .
lim sup(nk |an |) n = lim (nk ) n
n

The result now follows from the Cauchy-Hadamard Formula.

Remark: For the reader who is less than comfortable with limits infimum and supre|
mum, we recommend simply assuming that the Ratio Test limit = limn | aan+1
n
exists and proving Theorem 11.36 under that additional assumption using the Ratio
Test. This will be good enough for most of the power series encountered in practice.
P
Exercise: Let n an xn be a power
P series with interval of convergence I. Let J
be the interval of convergence of n nan xn1 .
a) Show that J I.
b) Give an example in which J 6= I.

CHAPTER 12

Taylor Taylor Taylor Taylor

1. Taylor Polynomials
Recall that early on we made a distinction between polynomial expressions and
polynomial functions. A polynomial expression is something of the form P (t) =
PN
n
expression we asn=0 an t with N N and a0 , . . . , aN R. To every polynomialP
N
sociated a polynomial function P (x) from R to R, given by x 7 n=0 an xn . Then
we showed that the process is reversible in the following sense: if two expressions
PN
PN
n
n
n=0 an t and
n=0 bn t define the same polynomial function, then an = bn for
all
P0 n n N . (This came down to showing that that if a polynomial expression
n=0 an t defines the zero function, then an = 0 for all n.) And after that
thankfully! we have spoken only of polynomials.
We wish briefly! to revisit this formalism in a slightly generalized context.
Namely we want to consider polynomial c-expansions and their associated funcPN
tions. A polynomial c-expansion is a formal expression n=0 an (t c)n with
N N and c, an R. A polynomial c-expansion induces a function in an evident
PN
manner: we map x to n=0 an (x c)n . Now we wish to establish the following
simple indeed, almost trivial but nevertheless important result.
Theorem 12.1. Let P be a polynomial of degree at most N , and let c R.
PN
a) There is a unique polynomial c-expansion n=0 bn (tc)n such that for all x R,
PN
P (x) = n=0 bn (x c)n .
b) Explicitly, we have bn =

P (n) (c)
n!

for all 0 n N .

Proof. a) We first discuss the existence of c-expansions. Perhaps this is most

cleanly handled by induction on the degree of P . If P is the zero polynomial or is
a constantPpolynomial, then there is nothing to show: a constant a is certainly of
the form n bn (x c)n : take b0 = a and bn = 0 for all n > 0. Suppose now that
N > 0 and we have established that every polynomial function of degree less than
N has a polynomial c-expansion, and consider a degree d polynomial expression
P (t) = a0 + a1 t + . . . + aN 1 tN 1 + aN tN . Then we may write
P (t) = aN (t c)N + Q(t),
where, since the leading term of aN (t c)N is aN tn , Q(t) is a polynomial of degree
less than N and thus by induction can be written as
Q(t) =

N
1
X

bn (t c)n .

n=0

Then
P (t) = b0 + b1 (t c) + . . . + bN 1 (t c)N 1 + aN (t c)N .
265

266

12. TAYLOR TAYLOR TAYLOR TAYLOR

Alternately and perhaps more directly, simply write

a0 +a1 t+a2 t2 +. . .+aN tN = a0 +a1 ((tc)+c)+a2 ((tc)+c)2 +. . .+aN ((tc)+c)N
and use the binomial theorem to expand each ((t c) + c)n in powers of t c.
PN
As for the uniqueness, suppose that we have two c-expansions n=0 an (t c)n ,
PN
n
same function. By an argument almost identical
n=0 bn (t c) which induce the
PN
PN
to that just given, we may write n=0 (an bn )(t c)n as n=0 dn tn , with highest
PN
degree term dN = aN bN . Since the corresponding function x 7 n=0 dn xn is
identically zero, by what we showed about polynomial functions we have dn = 0
PN
n
for all n. In particular 0 = dN = aN bN , so that
n=0 (an bn )(t c) =
P
PN 1
N 1
n
n
n=0 dn t . Reasoning as above we find that dN 1 =
n=0 (an bn )(t c) =
aN 1 bN 1 = 0, and so forth: continuing in this way we find that an = bn for all
0 6= n N , and thus the c-expansion is unique.
b) Consider the identity
(84)

P (x) =

N
X

bn (x c)n = b0 + b1 (x c) + b2 (x c)2 + . . . + bN (x c)N .

n=0

Evaluating (84) at x = c gives P (c) = b0 . Differentiating (84) gives

P 0 (x) = b1 + 2b2 (x c) + 3b3 (x c)2 + . . . + N bN (x c)N 1 ,
and evaluating at x = c gives P 0 (c) = b1 . Differentiating (84) again and evaluating
00
at x = c gives P 00 (c) = 2b2 , so b2 = P 2(c) . Continuing in this manner one finds
that for 0 k N , P (k) (c) = k!bk , so bk =

P (k) (c)
.
k!

Corollary 12.2. Let I be an interval and c an interior point of I, and let

f : I R be a function which is N times differentiable at c. Let
TN (x) =

N
X
f (k) (c)
k=0

k!

(x c)k .

Then TN (x) is the unique polynomial function of degree at most N such that
(k)
TN (c) = f (k) (c) for all 0 k N .
Proof. Applying Theorem 0 to TN (x) =
(k)

TN (c)
,
k!

(k)
TN (c)

PN

k=0

f (k) (c)
k! (x

c)k gives

f (k) (c)
k!

(k)

hence
= f (c) for all 0 k N . As for the uniqueness: let P (x) be
any polynomial of degree at most N such that P (k) (c) = f (k) (c) for 0 k N , and
let Q = TN P . Then Q is a polynomial of degree at most N such that Q(k) (c) = 0
PN Q(k) (c)
for 0 k N ; applying Theorem 12.1 we get Q(x) =
(x c)k =
k=0
k!
PN
k

k=0 0 (x c) = 0, i.e., Q is the zero polynomial and thus P = TN .
By definition, TN (x) is the degree N Taylor polynomial of f at c.
2. Taylors Theorem Without Remainder
For n N and c I , we say two functions f, g : I R agree to order n at c if
lim

xc

f (x) g(x)
= 0.
(x c)n

2. TAYLORS THEOREM WITHOUT REMAINDER

267

Exercise: Show: if m n and f and g agree to order n at c, then f and g agree to

order m at c.
Example 0: We claim that two continuous functions f and g agree to order 0
at c if and only if f (c) = g(c). Indeed, suppose that f and g agree to order 0 at c.
Since f and g are continuous, we have
0 = lim

xc

f (x) g(x)
= lim f (x) g(x) = f (c) g(c).
xc
(x c)0

The converse, that if f (c) = g(c) then limxc f (x) g(x) = 0, is equally clear.
Example 1: We claim that two differentiable functions f and g agree to order
1 at c if and only if f (c) = g(c) and f 0 (c) = g 0 (c). Indeed, by the exercise above
both hypotheses imply f (c) = g(c), so we may assume that, and then we find
f (x) g(x)
f (x) f (c) g(x) g(c)
= lim

= f 0 (c) g 0 (c).
xc
xc
xc
xc
Thus assuming f (c) = g(c), f and g agree to order 1 at c if and only f 0 (c) = g 0 (c).
lim

xc

The following result gives the expected generalization of these two examples. It
is generally attributed to Taylor,1 probably correctly, although special cases were
known to earlier mathematicians. Note that Taylors Theorem often refers to a
later result (Theorem 12.4) that we call Taylors Theorem With Remainder, even
though it is Theorem 12.3 (only) that was proved by Brook Taylor.
Theorem 12.3. (Taylor) Let n N and f, g : I R be two n times differentiable functions. Let c be an interior point of I. The following are equivalent:
(i) We have f (c) = g(c), f 0 (c) = g 0 (c), . . . , f (n) (c) = g (n) (c).
(ii) f and g agree to order n at c.
Proof. Set h(x) = f (x) g(x). Then (i) holds iff h(c) = h0 (c) = . . . =
h(x)
h (c) = 0 and (ii) holds iff limxc (xc)
n = 0. So we may work with h instead of
f and g. Since we dealt with n = 0 and n = 1 above, we may assume n 2.
h(x)
0
(i) = (ii): L = limxc (xc)
opitals Rule gives
n is of the form 0 , so LH
(n)

h0 (x)
,
xc n(x c)n1

L = lim

provided the latter limit exists. By our assumptions, this latter limit is still of the
form 00 , so we may apply LHopitals Rule again. We do so iff n > 2. In general,
we apply LH
opitals Rule n 1 times, getting


h(n1) (x)
1
h(n1) (x) h(n1) (c)
L = lim
=
lim
,
xc n!(x c)
n! xc
xc
provided the latter limit exists. But the expression in parentheses is nothing else
than the derivative of the function h(n1) (x) at x = c i.e., it is h(n) (c) = 0 (and,
in particular the limit exists; only now have the n 1 applications of LHopitals
Rule been unconditionally justified), so L = 0. Thus (ii) holds.
(ii) = (i): Let Tn (x) be the degree N Taylor polynomial to h at c. By Corollary
1Brook Taylor, 1685 - 1731

268

12. TAYLOR TAYLOR TAYLOR TAYLOR

12.2, f and Tn agree to order n at c, so by the just proved implication (i) = (ii),
h(x) and Tn (x) agree to order n at x = c:
lim

xc

h(x) Tn (x)
= 0.
(x c)n

Moreover, by assumption h(x) agrees to order n with the zero function:

h(x)
= 0.
xc (x c)n
lim

Subtracting these limits gives

(85)
h(c) + h0 (c)(x c) +
Tn (x)
lim
=
lim
xc (x c)n
xc

h00 (c)
2 (x

c)2 + . . . +
(x c)n

Clearly limxc Tn (x) = Tn (c), so if Tn (c) 6= 0, then limxc

so we must have Tn (c) = h(c) = 0. Therefore
h0 (c) +
Tn (x)
=
lim
lim
xc
xc (x c)n

h00 (c)
2 (x

c) + . . . + h
(x c)n1

(n)

h(n) (c)
n! (x

Tn (x)
(xc)n

(c)
n! (x

c)n

= 0.

would not exist,

c)n1

= 0.

As above, we have the limit of a quotient of continuous functions which we know

exists such that the denominator approaches 0, so the numerator must also approach
zero (otherwise the limit would be infinite): evaluating the numerator at c gives
h0 (c) = 0. And so forth: continuing in this way we find that the existence of the
limit in (85) implies that h(c) = h0 (c) = . . . = h(n1) (c) = 0, so (85) simplifies to
0 = lim

xc

(n)

so h

h(n) (c)
n! (x

c)n
h(n) (c)
=
,
n
(x c)
n!

(c) = 0.

Remark: Above we avoided a subtle pitfall: we applied LHopitals Rule n 1 times

h(x)
0
to limxc (xc)
n , but the final limit we got was still of the form 0 so why not
apply LH
opital one more time? The answer is if we do we get that
h(n) (x)
,
xc
n!

L = lim

assuming this limit exists. But to assume this last limit exists and is equal to h(n) (0)
is to assume that nth derivative of h is continuous at zero, which is slightly more
than we want (or need) to assume.
f (x)
For n N, a function f : I R vanishes to order n at c if limxc (xc)
n = 0.
Note that this concept came up prominently in the proof of Theorem 12.3 in the
form: f and g agree to order n at c iff f g vanishes to order n at c.

Exercise: Let f be a function which is n times differentiable at x = c, and let

Tn be its degree n Taylor polynomial at x = c. Show that f Tn vanishes to order
n at x = c. (This is just driving home a key point of the proof of Theorem 12.3 in
our new terminology.)
Exercise:

3. TAYLORS THEOREM WITH REMAINDER

269

a) Show that for a function f : I R, the following are equivalent:

(i) f is differentiable at c.
(ii) We may write f (x) = a0 + a1 (x c) + g(x) for a function g(x) vanishing to
order 1 at c.
b) Show that if the equivalent conditions of part a) are satisfied, then we must
have a0 = f (c), a1 = f 0 (c) and thus the expression of a function differentiable at
c as the sum of a linear function and a function vanishing to first order at c is unique.
Exercise:2 Let a, b Z+ , and consider the following function fa,b : R R:


 a
if x 6= 0
x sin x1b
fa,b (x) =
0
if x = 0
a) Show that fa,b vanishes to order a 1 at 0 but does not vanish to order a at 0.
0
b) Show that fa,b is differentiable at x = 0 iff a 2, in which case fa,b
(0) = 0.
c) Show that fa,b is twice differentiable at x = 0 iff a b + 3.
d) Deduce that for any n 2, fn,n vanishes to order n at x = 0 but is not twice
differentiable at x = 0.
3. Taylors Theorem With Remainder
To state the following theorem, it will be convenient to make a convention: real
numbers a, b, by |[a, b]| we will mean the interval [a, b] if a b and the interval [b, a]
if b < a. So |[a, b]| is the set of real numbers lying between a and b.
For a function f : |[a, b]| R, we put ||f || = supx[a,b] |f (x)|.
Theorem 12.4. (Taylors Theorem With Remainder) Let n N, let I be an
interval, and let f : I R be defined and n + 1 times differentiable. Let c I and
x I. Moreover, let
n
X
f (k) (c)(x c)k
Tn (x) =
k!
k=0

be the degree n Taylor polynomial of f at c, and let

Rn (x) = f (x) Tn (x)
be the remainder. Then:
a) There is z |[c, x]| such that
(86)

Rn (x) =

f (n+1) (z)
(x z)n (x c).
(n + 1)!

b) There exists z |[c, x]| such that

(87)

Rn (x) =

f (n+1) (z)
(x c)n+1 .
(n + 1)!

c) If f (n+1) is integrable on |[c, x]|, then

Z x n+1
f
(t)(x t)n dt
Rn (x) =
.
n!
c
2Thanks to Didier Piau for a correction that led to this exercise.

270

12. TAYLOR TAYLOR TAYLOR TAYLOR

d) We have
|Rn (x)| = |f (x) Tn (x)|

||f (n+1) ||
|x c|n+1 .
(n + 1)!

Proof. We closely follow [S, Thm. 20.4].

a) First note that, notwithstanding the notation chosen, the expression Rn (x) =
f (x) Tn (x) certainly depends upon the expansion point c I : thus it would
be more accurate to write it as Rn,c (x). In fact this opens the door to a key idea:
for any t I , we may consider the Taylor polynomial Tn,t to f centered at t and
thus the remainder function
n
X
f (k) (t)
(x t)k .
(88)
Rn,t (x) = f (x) Tn,t (x) = f (x)
k!
k=0

To emphasize the dependence on t, we define a new function S : |[c, x]| R by

S(t) = Rn,t (x). At first thought (my first thought, at least) consideration of S(t)
just seems to make things that much more complicated. But in fact the assumed
differentiability properties of f easily translate into differentiability properties of
S(t), and these can used to our advantage in a rather beautiful way. Indeed, since
f is n + 1 times differentiable on I, S is differentiable on |[c, x]|. Differentiating
both sides of (88) with respect to t gives

n  (k+1)
X
f (k) (t)
f (n+1) (t)
f
(t)
k
k1
0
0
(x t) +
(x t)
=
(xt)n .
S (t) = f (t)
k!
(k 1)!
n!
k=1

We apply the Mean Value Theorem to S on |[c, x]|: there is z |(c, x)| such that
f (n+1) (z)
S(x) S(c)
= S 0 (z) =
(x z)n .
xc
n!
Noting that S(x) = Rn,x (x) = 0 and S(c) = Rn,c (x) = Rn (x), this gives
f (n+1) (z)
(x z)n (x c).
n!
b) Apply the Cauchy Mean Value Theorem to S(t) and g(t) = (x t)n+1 on |[c, x]:
there is z |(c, x)| such that
Rn (x) = S(c) S(x) =

(n+1)

f
(z)
(x z)n
S(x) S(c)
S 0 (z)
f (n+1) (z)
Rn (x)
n!
=
=
=
=
,
(x c)n+1
g(x) g(c)
g 0 (z)
(n + 1)(x z)n
(n + 1)!

so

f (n+1) (z)
(x c)n+1 .
(n + 1)!
c) If f (n+1) is integrable on |[c, x]|, then
Z x
Z x (n+1)
f
(t)(x t)n dt
.
Rn (x) = (S(x) S(c)) =
S 0 (t) =
n!
c
c
Rn (x) =

d) This follows almost immediately from part b); the proof is left to the reader.

Exercise: Show that Theorem 12.4 (Taylors Theorem With Remainder) implies
Theorem 12.3 (Taylors Theorem) under the additional hypothesis that f (n+1) exists
and is continuous3 on the interval |[c, x]|.
3Thanks to Nick Fink for pointing out the hypothesis of continuity seems to be needed here.

4. TAYLOR SERIES

271

4. Taylor Series
4.1. The Taylor Series. Let f : I R be an infinitely differentiable function, and let c I. We define the Taylor series of f at c to be

X
f (n) (c)(x c)n
T (x) =
.
n!
n=0
Thus T (x) = limn Tn (x), where Tn is the degree n Taylor polynomial at c. In
particular T (x) is a power series, so all of our prior work on power series applies.
Just as with power series, it is no real loss of generality to assume that c = 0,
in which case our series takes the simpler form

X
f (n) (0)xn
;
T (x) =
n!
n=0
indeed, to get from this to the general case one merely has to make the change of
variables x 7 x c. It is traditional to call Taylor series centered around c = 0
Maclaurin series. But I know no good reason for this Taylor series were introduced by Taylor in 1721, whereas Colin Maclaurins Theory of fluxions was not
published until 1742 and makes explicit attribution is made to Taylors work.4 Using separate names for Taylor series centered at 0 and Taylor series centered at c
often suggests misleadingly! to students that there is some conceptual difference
between the two cases. So we will not use the term Maclaurin series here.
1

Exercise: Define a function f : R R by f (x) = e x2 for x 6= 0 and f (0) = 0. Show

that f is infinitely differentiable and in fact f (n) (0) = 0 for all n N.
When dealing with Taylor series there are two main issues.
Question 12.5. Let f : I R be an infinitely differentiable function and
(n)
P
(0)xn
be its Taylor series.
T (x) = n=0 f n!
a) For which values of x does T (x) converge?
b) If for x I, T (x) converges, do we have T (x) = f (x)?
Notice that Question 12.5a) is simply asking for which values of x R a power series is convergent, a question to which we worked out a very satisfactory answer in
X.X. Namely, the set of values x on which a power series converges is an interval
of radius R [0, ] centered at 0. More precisely, in theory the value of R is given
1
by Hadamards Formula R1 = lim supn |an | n , and in practice we expect to be
able to apply the Ratio Test (or, if necessary, the Root Test) to compute R.
If R = 0 then T (x) only converges at x = 0 and we have T (0) = f (0): this is a
trivial case. Henceforth we assume that R (0, ] so that f converges (at least)
on (R, R). Fix a number A, 0 < A R such that (A, A) I. We may then
move on to Question 12.5b): must f (x) = T (x) for all x (A, A)?
The answer is no: consider the function f (x) of Exercise X.X. f (x) is infinitely differentiable and has f (n) (0) = 0 for all n N, so its Taylor series is
4Special cases of the Taylor series concept were well known to Newton and Gregory in the
17th century and to the Indian mathematician Madhava of Sangamagrama in the 14th century.

272

12. TAYLOR TAYLOR TAYLOR TAYLOR

T (x) =

0xn
n=0 n!

n=0

0 = 0, i.e., it converges for all x R to the zero func1

tion. Of course f (0) = 0, but for x 6= 0, f (x) = e x2 6= 0. Therefore f (x) 6= T (x) in

any open interval around x = 0.
There are plenty of other examples. Indeed, in a sense that we will not try to make
precise here, most infinitely differentiable functions f : R R are not equal to
their Taylor series expansions in any open interval about any point. Thats the bad
news. However, one could interpret this to mean that we are not really interested
in most infinitely differentiable functions: the special functions one meets in
calculus, advanced calculus, physics, engineering and analytic number theory are
almost invariably equal to their Taylor series expansions, at least in some small
interval around any given point x = c in the domain.
4.2. Easy Examples.
In any case, if we wish to try to show that a T (x) = f (x) on some interval
(A, A), we have a tool for this: Taylors Theorem With Remainder. Indeed,
since Rn (x) = |f (x) Tn (x)|, we have
f (x) = T (x) f (x) = lim Tn (x)
n

lim |f (x) Tn (x)| = 0 lim Rn (x) = 0.

n

So it comes down to being able to give upper bounds on Rn (x) which tend to zero as
n . According to Taylors Theorem with Remainder, this will hold whenever
we can show that the norm of the nth derivative ||f (n) || does not grow too rapidly.
Example: We claim that for all x R, the function f (x) = ex is equal to its
Taylor series expansion at x = 0:

X
xn
x
.
e =
n!
n=0
First we compute the Taylor series expansion: f (0) (0) = f (0) = e0 = 1, and f 0 (x) =
ex , hence every derivative of ex is just ex again. We conclude that f (n) (0) = 1 for
P n
all n and thus the Taylor series is n=0 xn! , as claimed. Next note that this power
series converges for all real x, as we have already seen: just apply the Ratio Test.
Finally, we use Taylors Theorem with Remainder to show that Rn (x) 0 for each
fixed x R. Indeed, Theorem 12.4 gives us
Rn (x)

||f (n+1) ||
|x c|n+1 ,
(n + 1)!

where ||f (n+1) || is the supremum of the the absolute value of the (n+1)st derivative
on the interval |[0, x]|. But lucky us in this case f (n+1) (x) = ex for all n and
the maximum value of ex on this interval is ex if x 0 and 1 otherwise, so in either
way ||f (n+1) || e|x| . So
 n+1 
x
.
Rn (x) e|x|
(n + 1)!
And now we win: the factor inside the parentheses approaches zero with n and is
being multiplied by a quantity which is independent of n, so Rn (x) 0. In fact a

4. TAYLOR SERIES

273

moments thought shows that Rn (x) 0 uniformly on any bounded interval, say
on [A, A], and thus our work on the general properties of uniform convergence of
power series (in particular the M -test) is not needed here: everything comes from
Taylors Theorem With Remainder.
Example continued: we use Taylors Theorem With Remainder to compute e = e1
accurate to 10 decimal places.
A little thought shows that the work we did for f (x) = ex carries over verbatim
under somewhat more general hypotheses.
Theorem 12.6. Let f (x) : R R be a smooth function. Suppose that for all
A [0, ) there exists a number MA such that for all x [A, A] and all n N,
|f (n) (x)| MA .
(n)
P
(0)xn
converges absolutely for all x R.
a) The Taylor series T (x) = n=0 f n!
b) For all x R we have f (x) = T (x): that is, f is equal to its Taylor series
expansion at 0.

Exercise: Prove Theorem 12.6.

Exercise: Suppose f : R R is a smooth function with periodic derivatives:
there exists some k Z+ such that f = f (k) . Show that f satisfies the hypothesis
of Theorem 12.6 and therefore is equal to its Taylor series expansion at x = c.
4.3. The Binomial Series.
Even for familiar, elementary functions, using Theorem 12.4 to show Rn (x) 0
may require nonroutine work. We give a case study: the binomial series.
Let R. For x (1, 1), we define
f (x) = (1 + x) .
Case 1: Suppose N. Then f is just a polynomial; in particular f is defined and
infinitely differentiable for all real numbers.
Case 2: Suppose is positive but not an integer. Depending on the value of , f
may or may not be defined for x < 1 (e.g. it is for = 32 and it is not for = 32 ),
but in any case f is only hi times differentiable at x = 1.
Case 3: Suppose < 0. Then limx1+ f (x) = .
The upshot of this discussion is that if is not a positive integer, then f is defined
and infinitely differentiable on (1, ) and on no larger interval than this.
For n Z+ , f (n) (x) = ()( 1) ( (n 1))(1 + x)n , so f (n) (0) =
()( 1) ( (n 1)). Of course we have f (0) (0) = f (0) = 1, so the Taylor
series to f at c = 0 is
T (x) = 1 +

X
()( 1) ( (n 1)) n
x .
n!
n=1

274

12. TAYLOR TAYLOR TAYLOR TAYLOR

If
N, we recognize the nth Taylor series coefficient as the binomial coefficient

n , and this ought not to be surprising because for N, expanding out T (x)
simply gives the binomial theorem:
 
X
n

N, (1 + x) =
x .
n
n=0
So lets extend our definition of binomial coefficients: for any R, put
 

= 1,
0
 

()( 1) ( (n 1))
+
.
n Z ,
=
n!
n
Exercise: For any R, n Z+ , show
  
 


1
1
(89)
=
+
.
n
n1
n
Finally, we rename the Taylor series to f (x) as the binomial series
 
X
n
x .
B(, x) =
n
n=0
The binomial series is as old as calculus itself, having been studied by Newton in the
17th century.5 It remains one of the most important and useful of all power series.
For us, our order of business is the usual one when given a Taylor series: first, for
each fixed we wish to find the interval I on which the series B(, x) converges.
Second, we would like to show if possible! that for all x I, B(, x) = (1 + x) .
Theorem 12.7. Let R \ N, and consider the binomial series
 
 
X
X
n
n
B(, x) =
x =1+
x .
n
n
n=0
n=1
a)
b)
c)
d)

For all such , the radius of convergence of B(, x) = 1.

For all > 0, the series B(, 1) and B(, 1) are absolutely convergent.
If (1, 0), the series B(, 1) is nonabsolutely convergent.
If 1, then B(, 1) and B(, 1) are divergent.
Proof. a) We apply the Ratio Test:






 n


 = 1,



= lim  n+1
=
lim


n 
 n  n + 1 
n

so the radius of convergence is 1 = 1.

b) Step 0: Let a > 1 be a real number and m Z+ . Then we have

m 
m
a
1
m
m
a+m
= 1+
1+
>1+
=
> 0,
a1
a1
a1
a
a
where in the first inequality we have just taken the first two terms of the usual
(finite!) binomial expansion. Taking reciprocals, we get

m
a1
a
<
.
a
a+m
5In fact it is older. For an account of the early history of the binomial series, see [Co49].

4. TAYLOR SERIES

275

1
Step 1: Suppose (0, 1). Choose an integer m 2 such that m
< . Then
 

(1 ) (n 1 )
1
1 1
|
|=
< 1(1 ) (n 1 )
n
n!
m
m n!

m 1 2m 1
(n 1)m 1 1
1

= an ,
m
2m
(n 1)m n
n
say. Using Step 0, we get
=

am1
<
n

m
2m
(n 1)m

2m 1 3m 1
nm 1

m
2m
(n 1)m m 1
1 1

m1
2m 1
(n 1)m 1 nm 1
an n

1
1
It follows that an < 1 , so | n | < 1+ 1 , so
nm
n m
X 
X 1
|
|
1 < .
1+ m
n
n
n n

n
This shows that B(, 1) is absolutely convergent; since |
n (1) | = | n |, it also
shows that B(, 1) is absolutely convergent.
Step 2: Using the identity (89), we find
 

 

X
X
n
1
1
S(, x) = 1 +
x = 1+
+
xn = (1 + x)S( 1, x).
n
n

1
n
n=1
n=1
=

Thus for any fixed x, if S( 1, x) (absolutely) converges, so does S(, x). By an

evident induction argument, if S(, x) (absolutely) converges, so does S( + n, x)
for all n N. Since S(, 1) and S(, 1) are absolutely convergent for all (0, 1),
they are thus absolutely convergent for all non-integers > 0.
c) If (1, 0) and n N, then

  

n
(1, 0);
/
=
n+1
n+1
n
P

this shows simultaneously that the sequence of terms of B(, 1) =
n=0 n is
decreasing in absolute value and alternating in sign. Further, write = 1, so
1
that (0, 1). Choose an integer m 2 such that m
< . Then
 

(1 )(2 ) (n 1 ) n
n
|
|=
= bn
.
(n 1)!
n
n
n
Arguing as in Step 1 of part b) shows that bn < 11 , and hence
nm
 

n
= 0 1 = 0.
lim |
| = lim bn lim
n
n
n
n
n
Therefore the Alternating Series Test applies to show that S(, 1) converges.

d) The absolute value of the nth term of both B(, 1) and B(, 1) is |
n |. If
1, then | n| n + 1 and thus

  

n
|
/
|=|
| 1,
n+1
n
n+1



and thus
n 6 0. By the N th term test, S(, 1) and S(, 1) diverge.

276

12. TAYLOR TAYLOR TAYLOR TAYLOR

Exercise*: Show that for (1, 0), the binomial series B(, 1) diverges.
Remark: As the reader has surely noted, the convergence of the binomial series
S(, x) at x = 1 is a rather delicate and tricky enterprise. In fact most texts at
this level even [S] do not treat it. We have taken Step 1 of part b) from [Ho66].
Remark: There is an extension of the Ratio Test due to J.L. Raabe which simplifies much the of the above analysis, including the preceding exercise.
Theorem 12.8. Let R \ N; let f (x) = (1 + x) , and consider its Taylor
series at zero, the binomial series
 
X
n
B(, x) =
x .
n
n=0
a) For all x (1, 1), f (x) = B(, x).
b) If > 1, f (1) = B(, 1).
c) If > 0, f (1) = B(, 1).
Proof. [La] Let Tn1 (x) be the (n 1)st Taylor polynomial for f at 0, so
B(, x) = lim Tn1 (x)
n

is the Taylor series expansion of f at zero. As usual, put Rn1 (x) = f (x)Tn1 (x).
a) By Theorem 12.4b),
Z x
Z x n
1
f (t)(x t)n1 dt
=
(1) (n+1)(1+t)n (xt)n1 dt.
Rn1 (x) =
(n 1)!
(n 1)! 0
0
By the Mean Value Theorem for Integrals, there is (0, 1) such that
Rn1 (x) =

( 1) ( n + 1)
(1 + x)n (x x)n1 (x 0).
(n 1)!

Put
1
t=
, cn (s) =
1 + x


1 n1
s
.
n1

Then
(1 + s)1 =

cn (s)

n=1

and
Rn1 (x) = cn (xt)x(1 + x)1 .
P
Since x (1, 1), we have t (0, 1), so |xt| < 1. It follows that
n=1 cn (xt)
converges, so by the nth term test cn (xt) 0 as n and thus Rn1 (x) 0.
b) The above argument works verbatim if x = 1 and > 1.
c) IfP > 0, then by Theorem 12.7b), S(, 1) is convergent. Moreover, 1 > 1,

so
n=1 cn (1) converges and thus cn (1) 0. But |cn (1)| = |cn (1)|, so also
cn (1) 0 and thus Rn1 (1) 0.


5. HERMITE INTERPOLATION

277

5. Hermite Interpolation
It is a well-known fact that two points determine a line. One version of this
is: given real numbers x1 < x2 and real numbers f1 , f2 , there is a unique linear
function f : R R such that f (x1 ) = f1 and f (x2 ) = f2 . In a similar way, three
points determine a parabola: given real numbers x1 < x2 < x3 and real numbers
f1 , f2 , f3 , there is a unique quadratic polynomial P (x) = ax2 + bx + c such that
P (xi ) = fi for i = 1, 2, 3. The following is a generalization of this.
Theorem 12.9. (Polynomial Interpolation) Let n Z+ , let x0 < . . . < xn be
real numbers, and let y0 , . . . , yn be real numbers. Then there is a unique polynomial
P of degree at most n such that P (xi ) = yi for 0 i n. Indeed, there are real
numbers A0 , . . . , An such that
(90) P (x) = A0 + A1 (x x0 ) + A2 (x x0 )(x x1 ) + . . . + An (x x0 ) (x xn1 ).
Proof. Uniqueness: Suppose that f and g are two such polynomials. Then
the polynomial h = f g has degree at most n and vanishes at the n + 1 points
x0 , . . . , xn . Since a nonzero polynomial cannot have more roots than its degree, we
must have h 0 and thus f = g.
Existence: Often the way to find something is to postulate (i.e., guess) a certain
form that the solution should take, while leaving certain parameters undetermined,
then algebraically solve for these parameters. (This is sometimes called the method
of undetermined coefficients.) The present result is an instance of this: (90)
gives the general form of the solution. If we plug in x = x0 we find
y0 = P (x0 ) = A0 .
Plugging in x = x1 we find
y1 = P (x1 ) = A0 + A1 (x1 x0 ),
which we can uniquely solve for A1 : explicitly
A1 =

y1 y0
y1 A0
=
.
x1 x0
x1 x0

Inductively, having solved uniquely for A0 , . . . , Ak1 , we plug in x = xk , getting

yk = P (xk ) = A0 +A1 (xk x0 )+A2 (xk x0 )(xk x1 )+. . .+Ak (xk x0 ) (xk xk1 ),
which we may solve uniquely for Ak .

In fact, following Lagrange, we can give a more explicit formula for the interpolating
polynomial of Theorem 12.9. (Why didnt we start with this? Because often it is
useful to make a distinction between proving the existence of something and giving
an explicit construction of it. The former is often easier, whereas the latter is often
more useful. In fact we will not need the explicit formula for most of our work here,
but in our study of quadrature we will want to have it.)
Theorem 12.10. (Lagrange Interpolation Formula) Let n Z+ , let x0 < . . . <
xn be real numbers, and let y0 , . . . , yn be real numbers. For 0 j n, define
Y
x x0
x xj1 x xj+1
x xn
x xn
=

.
`j (x) =
xj xn
xj x0
xj xj1 xj xj+1
xj xn
0in, i6=j

278

12. TAYLOR TAYLOR TAYLOR TAYLOR

Then an explicit formula for the unique polynomial P (x) of degree at most n such
that P (xi ) = yi for all 0 i n of Theorem 12.9 is
(91)

P (x) =

n
X

yj `j (x).

j=0

Proof. Each `j (x) is a product of n + 1 1 = n linear polynomials, so is a

polynomial of degree n. Therefore P (x) is a polynomial of degree at most n. To
establish (91) the key observation is that for all 1 i, j n, `j (xi ) is equal to 1
if i = j and 0 if i 6= j. We ask the reader to just stop and think about this: the
formula looks a little complicated at first, so it is natural to worry that it may not
be so easy to see this...but in fact it is immediate. The equation (91) follows, also
immediately.

The polynomials `j (x) are called Lagrange basis polynomials. One of the merits
of the formula is that they do not depend on the values y0 , . . . , yn or, if you like,
the function y = f (x) that we are interpolating but only on the values x0 , . . . , xn
which in the lingo of this field are often called the interpolation nodes.
Now look back at (90): it is reminiscent of the degree n Taylor polynomial, except instead of taking powers of x x0 for a fixed center x0 we are taking different
points xi . Just brainstorming then, might it be possible to generalize Theorem 12.9
by allowing some of the xi s to coincide, i.e., replacing the condition x0 < . . . < xn
with x0 x1 . . . xn ? At first glance no: if for instance x1 = x2 then the
conditions f (x1 ) = f1 and f (x2 ) = f2 are inconsistent unless f1 = f2 , in which case
they are simply redundant. But there is a way of salvaging the situation. The key
observation is that multiple roots of a polynomial function show up as roots of the
derivative. More precisely, if a polynomial f can be written as (x c)m g(x) with
g(c) 6= 0, then
f 0 = m(x c)m1 g(x) + (x c)m g 0 (x) = (x c)m1 (mg(x) + (x c)g 0 (x)) .
Observe that when we plug in x = c the expression mg(c)+(cc)g 0 (c) = mg(c) 6= 0.
This shows that if a polynomial f has a root of multiplicity m 1 at a point c,
then the polynomial f 0 has a root of multiplicity m 1 at c.
A reasonable way of repairing f (x1 ) = f1 , f (x2 ) = f2 when x1 = x2 then
is to take the second condition as a condition on the derivative of f at x2 : i.e.,
f 0 (x2 ) = f2 . But it is probably clearer to switch to a different notation: suppose
that we have r distinct real numbers x1 < . . . < xr , and each occurs with a
certain multiplicity mi . Thus our list of numbers contains m1 instances of x1 , m2
instances of x2 , up to mr instances of xr , thus m1 + . . . + mr entries altogether.
Lets put n + 1 = m1 + . . . + mr . Here is a more general interpolation theorem.
Theorem 12.11. (Generalized Polynomial Interpolation) Let r, m1 , . . . , mr
Z+ , and put n + 1 = m1 + . . . + mr . Let x1 < . . . , xr be real numbers, and for
each 1 i r and 1 j mr , let fij be a real number. Then there is a unique
polynomial P of degree at most n such that for all 1 i r and 0 j mi 1,
P (j) (xi ) = fij .
Proof. . . .

5. HERMITE INTERPOLATION

279

Let us now try to switch back to the old notation: we give ourselves n + 1 real
numbers with multiplicity: x0 x1 . . . xn , a and n + 1 real numbers
f0 , . . . , fn . We write the interpolation problem as above as f (xi ) = fi , but with the
understanding that when a root is repeated more than once, the further conditions
are conditions on the derivatives of f . In this case we claim that the interpolation
polynomial can be taken in the same form as above: namely, there are unique real
numbers A0 , . . . , An such that
f = f (x) = A0 + A1 (x x0 ) + A2 (x x0 )(x x1 ) + . . . + An (x x0 ) (x xn1 ).
At the moment we will prove this by linear algebraic considerations (which is cheating: we are not supposed to be assuming any knowledge of linear algebra in this
text!). Namely, since we have already shown the existence of an interpolating polynomial f of degree at most n, it suffices to show that the set of polynomials
S = {1, x x0 , (x x0 )(x x1 ), . . . , (x x0 ) (x xn1 )}
spans the R-vector space Pn of all polynomials of degree at most n. The set S is
linearly independent: indeed, the polynomials have distinct degrees, so a nontrivial
linear indeendence relationship would allow us to write a nonzero polynomial as
a linear combination of polynomials of smaller degree, which is absurd. Further,
#S = n + 1. But Pn has dimension n, so S must be a basis for Pn : in paticular S
spans Pn .
Having billed Theorem 12.11 as generalizing Theorem 12.9, let us now call attention to the other extreme: suppose x0 = . . . = xn = c, say. Then the interpolating
polynomial P is precisely the degree n Taylor polynomial at c to any n times differentiable function f with f (j) (c) = fj for 0 j n. This brings up a key
idea in general: let I be an interval containing the points x0 . . . xn , and let
f : I R be an n times differentiable function. We define the Hermite interpolation polynomial P (x) to be the unique polynomial of degree at most n such
that for all 0 i n, P (xi ) = f (xi ): here we are using the above slightly shady
convention that when the xi s occur with multiplicity greater than 1, the conditions
P (xi ) = f (xi ) are actually conditions on the derivatives of P and f at xi .
Let us define the remainder function: for x I,
R(x) = f (x) P (x).
Following [CJ], we will now give an expression for R which generalizes one form of
Taylors Theorem With Remainder. We begin with one preliminary result.
Theorem 12.12. (Generalized Rolles Theorem) Let f : I R be n times differentiable, and assume that f has at least n+1 roots on I, counted with multiplicity.
Then there is I with f (n) () = 0.
Exercise: Prove Theorem 12.12.
Theorem 12.13. (Hermite With Remainder) Let x0 . . . xn I, and let
f : I R be (n + 1) times differentiable. Let P be the Hermite Interpolation
Polynomial for f . Then, for all x I, there is I in fact, lying in any closed
interval containing x, x0 , . . . , xn such that
(x x0 ) (x xn ) (n+1)
f
().
(92)
R(x) = f (x) P (x) =
(n + 1)!

280

12. TAYLOR TAYLOR TAYLOR TAYLOR

Proof. If x = xi for some i, then both sides of (92) are 0, so equality holds.
We may thus assume x 6= xi for any i. Let c R, and consider
K(x) = R(x) c(x x0 ) (x xn ).
There is a unique value of c such that K(x) = 0: namely,
R(x)
.
(x x0 ) (x xn )
The function K : I R thus vanishes at least n + 2 times on I with multiplicity,
so by the Generalized Rolles Theorem there is I such that
c=

0 = K (n+1) () = f (n+1) () P (n+1) () (n + 1)!c = f (n+1) () (n + 1)!c,

so
c=
It follows that
R(x) =

f (n+1) ()
.
(n + 1)!

(x x0 ) (x xn ) (n+1)
f
().
(n + 1)!


Remark: Restricting to Taylor polynomials, our earlier argument for the existence
of the interpolating polynomial is certainly easier: recall this consisted of simply
writing down the answer and checking that it was correct. However this proof of
part b) of Taylors Theorem with Remainder seems easier.
Exercise: a) Let x0 . . . xn , and let m n. Let
Pn (x) = A0 + A1 (x x0 ) + . . . + An (x x0 ) (x xn1 )
be the Hermite interpolation polynomial for a function f . Show that the Hermite
interpolation polynomial for f with respect to the approximation points x0 . . .
xk is
Pm (x) = A0 + A1 (x x0 ) + . . . + Am (x x0 ) (x xm1 ).
b) Suppose xn1 6= xn . Show that there is [x0 , xn ] such that
f (n) ()
.
n!
c) Show that there is a sequence {k } taking values in I such that
An =

f (n) (k )
.
k
n!
d) Suppose that f has a continuous (n + 1)st derivative. Use part c) to recover the
formula or the nth Taylor series coefficient.
An = lim

CHAPTER 13

Sequences and Series of Functions

1. Pointwise Convergence
All we have to do now is take these lies and make them true somehow. G. Michael1
1.1. Pointwise convergence: cautionary tales.
Let I be an interval in the real numbers. A sequence of real functions is a
sequence f0 , f1 , . . . , fn , . . ., with each fn a function from I to R.
P
For us the following example is all-important: let f (x) = n=0 an xn be a power
series with radius of convergence
Pn R > k0. So f may be viewed as a function
f : (R, R) R. Put fn =
k=0 ak x , so each fn is a polynomial of degree
at most n; therefore fn makes sense as a function from R to R, but let us restrict
its domain to (R, R). Then we get a sequence of functions f0 , f1 , . . . , fn , . . ..
As above, our stated goal is to show that the function f has many desirable properties: it is continuous and indeed infinitely differentiable, and its derivatives and
antiderivatives can be computed term-by-term. Since the functions fn have all these
properties (and more each fn is a polynomial), it seems like a reasonable strategy
to define some sense in which the sequence {fn } converges to the function f , in
such a way that this converges process preserves the favorable properties of the fn s.
The previous description perhaps sounds overly complicated and mysterious, since
in fact there is an evident sense in which the sequence of functions fn converges to
f . Indeed, to say that x lies
the open interval (R, R) of convergence is to say
Pin
n
that the sequence fn (x) = k=0 ak xk converges to f (x).
This leads to the following definition: if {fn }
n=1 is a sequence of real functions
defined on some interval I and f : I R is another function, we say fn converges
to f pointwise if for all x I, fn (x) f (x). (We also say f is the pointwise
limit of the sequence {fn }.) In particular the sequence of partial sums of a power
series converges pointwise to the power series on the interval I of convergence.
P
Remark: There is similarly a notion of an infinite series of functions
n=0 fn
and of pointwise convergence of this series to some limit function f . Indeed,Pas in
the case of just one series, we just define Sn = f0 + . . . + fn and say that n fn
converges pointwise to f if the sequence Sn converges pointwise to f .

1George Michael, 1963

281

282

13. SEQUENCES AND SERIES OF FUNCTIONS

The great mathematicians of the 17th, 18th and early 19th centuries encountered
many sequences and series of functions (again, especially power series and Taylor
series) and often did not hesitate to assert that the pointwise limit of a sequence of
functions having a certain nice property itself had that nice property.2 The problem
is that statements like this unfortunately need not be true!
Example 1: Define fn = xn : [0, 1] R. Clearly fn (0) = 0n = 0, so fn (0) 0. For
any 0 < x 1, the sequence fn (x) = xn is a geometric sequence with geometric
ratio x, so that fn (x) 0 for 0 < x < 1 and fn (1) 1. It follows that the
sequence of functions {fn } has a pointwise limit f : [0, 1] R, the function which
is 0 for 0 x < 1 and 1 at x = 1. Unfortunately the limit function is discontinuous
at x = 1, despite the fact that each of the functions fn are continuous (and are
polynomials, so really as nice as a function can be). Therefore the pointwise
limit of a sequence of continuous functions need not be continuous.
Remark: Example 1 was chosen for its simplicity, not to exhibit maximum pathology. It is possible to construct a sequence {fn }
n=1 of polynomial functions converging pointwise to a function f : [0, 1] R that has infinitely many discontinuities!
(On the other hand, it turns out that it is not possible for a pointwise limit of
continuous functions to be discontinuous at every point. This is a theorem of R.
Baire. But we had better not talk about this, or well get distracted from our stated
goal of establishing the wonderful properties of power series.)
One can also find assertions in the math papers of old that if fn converges to
Rb
Rb
f pointwise on an interval [a, b], then a fn dx a f dx. To a modern eye, there
are in fact two things to establish here: first that if each fn is Riemann integrable,
then the pointwise limit f must be Riemann integrable. And second, that if f is
Riemann integrable, its integral is the limit of the sequence of integrals of the fn s.
In fact both of these are false!
Example 2: Define a sequence {fn }
n=0 with common domain [0, 1] as follows. Let
f0 be the constant function 1. Let f1 be the function which is constantly 1 except
f (0) = f (1) = 0. Let f2 be the function which is equal to f1 except f (1/2) = 0.
Let f3 be the function which is equal to f2 except f (1/3) = f (2/3) = 0. And
so forth. To get from fn to fn+1 we change the value of fn at the finitely many
rational numbers na in [0, 1] from 1 to 0. Thus each fn is equal to 1 except at a finite
set of points: in particular it is bounded with only finitely many discontinuities,
so it is Riemann integrable. The functions fn converges pointwise to a function f
which is 1 on every irrational point of [0, 1] and 0 on every rational point of [0, 1].
Since every open interval (a, b) contains both rational and irrational numbers, the
function f is not Riemann integrable: for any partition of [0, 1] its upper sum is
1 and its lower sum is 0. Thus a pointwise limit of Riemann integrable functions
need not be Riemann integrable.
2This is an exaggeration. The precise definition of convergence of real sequences did not come
until the work of Weierstrass in the latter half of the 19th century. Thus mathematicians spoke of
functions fn approaching or getting infinitely close to a fixed function f . Exactly what they
meant by this and indeed, whether even they knew exactly what they meant (presumably some
did better than others) is a matter of serious debate among historians of mathematics.

2. UNIFORM CONVERGENCE

283

Example 3: We define a sequence of functions fn : [0, 1] R as follows: fn (0) = 0,

and fn (x) = 0 for x n1 . On the interval [0, n1 ] the function forms a spike:
1
1
) = 2n and the graph of f from (0, 0) to ( 2n
, 2n) is a straight line, as is the
f ( 2n
1
1
graph of f from ( 2n , 2n) to ( n , 0). In particular fn is piecewise linear hence continuous, hence RiemannR integable, and its integral is the area of a triangle with
1
base n1 and height 2n: 0 fn dx = 1. On the other hand this sequence converges
pointwise to the zero function f . So
Z 1
Z 1
lim fn .
fn = 1 6= 0 =
lim
n

0 n

Example 4: Let g : R R be a bounded differentiable function such that limn g(n)

+
does not exist. (For instance, we may take g(x) = sin( x
2 ).) For n Z , define
g(nx)
fn (x) = n . Let M be such that |g(x)| M for all x R. Then for all
x R, |fn (x)| M
n , so fn conveges pointwise to the function f (x) 0 and thus
f 0 (x) 0. In particular f 0 (1) = 0. On the other hand, for any fixed nonzero x,
0
fn0 (x) = ng n(nx) = g 0 (nx), so
lim fn0 (1) = lim g 0 (n) does not exist.

Thus
lim fn0 (1) 6= ( lim fn )0 (1).

A common theme in all these examples is the interchange of limit operations:

that is, we have some other limiting process corresponding to the condition of
continuity, integrability, differentiability, integration or differentiation, and we are
wondering whether it changes things to perform the limiting process on each fn
individually and then take the limit versus taking the limit first and then perform
the limiting process on f . As we can see: in general it does matter! This is not
to say that the interchange of limit operations is something to be systematically
avoided. On the contrary, it is an essential part of the subject, and in natural
circumstances the interchange of limit operations is probably valid. But we need
to develop theorems to this effect: i.e., under some specific additional hypotheses,
interchange of limit operations is justified.
2. Uniform Convergence
Most of the above pathologies vanish if we consider a stronger notion of convergence.
Let {fn } be a sequence of functions with domain I. We say fn converges uniu
formly to f and write fn f if for all  > 0, there exists N Z+ such that for
all n N and all x I, |fn (x) f (x)| < .
How does this definition differ from that of pointwise convergence? Lets compare: fn f pointwise if for all x I and all  > 0, there exists n Z+ such that
for all n N , |fn (x) f (x)|. The only difference is in the order of the quantifiers:
in pointwise convergence we are first given  and x and then must find an N Z+ :
that is, the N is allowed to depend both on  and the point x I. In the definition of uniform convergence, we are given  > 0 and must find an N Z+ which

284

13. SEQUENCES AND SERIES OF FUNCTIONS

works simultaneously (or uniformly) for all x I. Thus uniform convergence is

a stronger condition than pointwise convergence, and in particular if fn converges
to f uniformly, then certainly fn converges to f pointwise.
Lemma 13.1. (Cauchy Criterion For Uniform Convergnece) Let fn : I R be
a sequence of functions. The following are equivalent:
(i) For all  > 0, there is N Z+ such that for all m, n N and all x I,
|fm (x) fn (x)| < .
(ii) For all  > 0, there is N Z+ such that for all n N , all k 0 and all x I,
|fn+k (x) fn (x)| < .
u
(iii) fn f .
Exercise: Prove Lemma 13.1.
2.1. Uniform Convergence and Inherited Properties.
The following result is the most basic one fitting under the general heading uniform
convegence justifies the exchange of limiting operations.
Theorem 13.2. Let {fn } be a sequence of functions with common domain I,
and let c be a point of I. Suppose that for all n Z+ , limxc fn (x) = Ln . Suppose
u
moreover that fn f . Then the sequence {Ln } is convergent, limxc f (x) exists
and we have equality:
lim Ln = lim lim fn (x) = lim f (x) = lim lim fn (x).

n xc

xc

xc n

Proof. Step 1: We show that the sequence {Ln } is convergent. Since we dont
yet have a real number to show that it converges to, it is natural to try to use the
Cauchy criterion, hence to try to bound |Lm Ln |. Now comes the trick: for all
x I we have
|Lm Ln | |Lm fm (x)| + |fm (x) fn (x)| + |fn (x) Ln |.
By the Cauchy criterion for uniform convergence, for any  > 0 there exists N Z+
such that for all m, n N and all x I we have |fm (x) fn (x)| < 3 . Moreover,
the fact that fm (x) Lm and fn (x) Ln give us bounds on the first and last
terms: there exists > 0 such that if 0 < |x c| < then |Ln fn (x)| < 3
and |Lm fm (x)| < 3 . Combining these three estimates, we find that by taking
x (c , c + ), x 6= c and m, n N , we have



|Lm Ln | + + = .
3 3 3
So the sequence {Ln } is Cauchy and hence convergent, say to the real number L.
Step 2: We show that limxc f (x) = L (so in particular the limit exists!). Actually
the argument for this is very similar to that of Step 1:
|f (x) L| |f (x) fn (x)| + |fn (x) Ln | + |Ln L|.
Since Ln L and fn (x) f (x), the first and last term will each be less than 3
for sufficiently large n. Since fn (x) Ln , the middle term will be less than 3 for
x sufficiently close to c. Overall we find that by taking x sufficiently close to (but
not equal to) c, we get |f (x) L| <  and thus limxc f (x) = L.

Corollary 13.3. Let fn be a sequence of continuous functions with common
u
domain I and suppose that fn f on I. Then f is continuous on I.

2. UNIFORM CONVERGENCE

285

Since Corollary 13.3 is easier than Theorem 13.2, we include a separate proof.
Proof. Let x I. We need to show that limxc f (x) = f (c), thus we need to
show that for any  > 0 there exists > 0 such that for all x with |x c| < we
have |f (x) f (c)| < . The idea again! is to trade this one quantity for three
quantities that we have an immediate handle on by writing
|f (x) f (c)| |f (x) fn (x)| + |fn (x) fn (c)| + |fn (c) f (c)|.
By uniform convergence, there exists n Z+ such that |f (x) fn (x)| < 3 for
all x I: in particular |fn (c) f (c)| = |f (c) fn (c)| < 3 . Further, since fn (x)
is continuous, there exists > 0 such that for all x with |x c| < we have
|fn (x) fn (c)| < 3 . Consolidating these estimates, we get



|f (x) f (c)| < + + = .

3 3 3
Exercise: Consider again fn (x) = xn on the interval [0, 1]. We saw in Example 1
above that fn converges pointwise to the discontinuous function f which is 0 on
[0, 1) and 1 at x = 1.
a) Show directly from the definition that the convergence of fn to f is not uniform.
b) Try to pinpoint exactly where the proof of Theorem 13.2 breaks down when
applied to this non-uniformly convergent sequence.
Exercise: Let fn : [a, b] R be a sequence of functions. Show TFAE:
u
(i) fn f on [a, b].
u
(ii) fn f on [a, b) and fn (b) f (b).
Theorem 13.4. Let {fn } be a sequence of (Riemann-Darboux) integrable funcu
tions with common domain [a, b]. Suppose that fn f . Then f is integrable and
Z b
Z b
Z b
lim
fn =
lim fn =
f.
n

a n

a
u

Proof. Step 1: We prove the integrability of f . Fix  > 0; since f f , there

is N Z+ such that for all n N and all x [a, b], |fn (x) f (x)| < ; it follows
that for any subinterval [c, d] [a, b],
| sup(fn , [c, d]) sup(f, [c, d])| , | inf(fn , [c, d]) inf(f, [c, d])| .
So for any partition P = {a = x0 < x1 < . . . < xn1 < xn = b} of [a, b] and n N ,
|U (fn , P) U (f, P)|

n1
X

| sup(fn , [xn , xn+1 ]) sup(f, [xn , xn+1 ]|(xi+1 xi )|

i=0

n1
X

(xi+1 xi ) = (b a),

i=0

and similarly,
|L(fn , P) L(f, P)| (b a).
Since fN is integrable, by Darbouxs Criterion there is a partition P of [a, b] such
that U (fN , P) L(fN , P) < . Thus
|U (f, P)L(f, P)| |U (f, P)U (fn , P)|+|U (fn , P)L(fn , P)|+|L(fn , P)L(f, P)|
(b a) +  + (b a) = (2(b a) + 1).

286

13. SEQUENCES AND SERIES OF FUNCTIONS

Since  > 0 was arbitrary, Darbouxs Criterion shows f is integrable on [a, b].
Step 2: If f, g : [a, b] R are integrable and |f (x) g(x)|  for all x [a, b], then
Z b
Z b
Z b
Z b
|
f
g| = |
f g|
|f g| (b a).
a

a
u

From this simple observation and Step 1 the fact that fn f implies
is almost immediate. The details are left to you.

Rb
a

fn

Rb
a

f


Exercise: It follows from Theorem 13.4 that the sequences in Examples 2 and 3
above are not uniformly convergent. Verify this directly.
Corollary 13.5. Let {fn } be a sequence of continuous functions defined on
P
u
the interval [a, b] such that n=0 fn f . For each n, let Fn : [a, b] R be the
0
unique function with Fn = fn and Fn (a) = 0, and similarly let F : [a, b] R be the
P
u
unique function with F 0 = f and F (a) = 0. Then n=0 Fn F .
Exercise: Prove Corollary 13.5.
Our next order of business is to discuss differentiation of sequences of functions.
For this we should reconsider Example 4: let g : R R be a bounded differentiable
function such that limn g(n) does not exist, and let fn (x) = g(nx)
n . Let M be
u
M
such that |g(x)| M for all R. Then for all x R, |fn (x)| n , so fn 0. But
as we saw above, limn fn0 (1) does not exist.
Thus we have shown the following somewhat distressing fact: uniform convergence of fn to f does not imply that fn0 converges.
Well, dont panic. What we want is true in practice; we just need suitable hypotheses. We will give a relatively simple result sufficient for our coming applications.
Theorem 13.6. Let {fn }
n=1 be a sequence of functions on [a, b]. We suppose:
(i) Each fn is continuously differentiable on [a, b],
(ii) The functions fn converge pointwise on [a, b] to some function f , and
(iii) The functions fn0 converge uniformly on [a, b] to some function g.
Then f is differentiable and f 0 = g, or in other words
( lim fn )0 = lim fn0 .
n

u
fn0

Proof. Let x [a, b]. Since

g on [a, b], certainly fn0 g on [a, x]. Since
0
each fn is continuous, by Corollary 13.3 g is continuous. Now applying Theorem
13.4 and the Fundamental Theorem of Calculus we have
Z x
Z x
Z x
0
g=
lim fn = lim
fn0 = lim fn (x) fn (a) = f (x) f (a).
a

a n

Differentiating and applying the Fundamental Theorem of Calculus, we get

g = (f (x) f (a))0 = f 0 .
P

Corollary 13.7. Let n=0 fn (x) be a series of functions converging pointwise

P
u
to f (x). Suppose that each fn0 is continuously differentiable and n=0 fn0 (x) g.
0
Then f is differentiable and f = g:

X
X
(93)
(
fn )0 =
fn0 .
n=0

n=0

2. UNIFORM CONVERGENCE

287

Exercise: Prove Corollary 13.7.

P
P
P 0
0
When for a series
n fn it holds that (
n fn ) =
n fn , we say that the series can be differentiated termwise or term-by-term. Thus Corollary 13.7 gives
a condition under which a series of functions can be differentiated termwise.
Although Theorem 13.6 (or more precisely, Corollary 13.7) will be sufficient for
our needs, we cannot help but record the following stronger version.
Theorem 13.8. Let {fn } be differentiable functions on the interval [a, b] such
that {fn (x0 )} is convergent for some x0 [a, b]. If there is g : [a, b] R such that
u
u
fn0 g on [a, b], then there is f : [a, b] R such that fn f on [a, b] and f 0 = g.
Proof. [R, pp.152-153]
Step 1: Fix  > 0, and choose N Z+ such that m, n N implies |fm (x0 )

0
(t) fn0 (t)| < 2(ba)
for all t [a, b]. The latter inequality is
fn (x0 )| 2 and |fm
telling us that the derivative of g := fm fn is small on the entire interval [a, b].
Applying the Mean Value Theorem to g, we get a c (a, b) such that for all
x, t [a, b] and all m, n N ,




(94)
|g(x) g(t)| = |x t||g 0 (c)| |x t|
.
2(b a)
2
It follows that for all x [a, b],


+ = .
2 2
By the Cauchy criterion, fn is uniformly convergent on [a, b] to some function f .
Step 2: Now fix x [a, b] and define
|fm (x) fn (x)| = |g(x)| |g(x) g(x0 )| + |g(x0 )| <

n (t) =

fn (t) fn (x)
tx

and
f (t) f (x)
,
tx
so that for all n Z+ , limxt n (t) = fn0 (x). Now by (94) we have

|m (t) n (t)|
2(b a)
(t) =

for all m, n N , so once again by the Cauchy criterion n converges uniformly for
u
u
all t 6= x. Since fn f , we get n for all t 6= x. Finally we apply Theorem
13.2 on the interchange of limit operations:
f 0 (x) = lim (t) = lim lim n (t) = lim lim n (t) = lim fn0 (x).
tx

tx n

n tx


2.2. The Weierstrass M-test.
We have just seen that uniform convergence of a sequence of functions (and possibly,
of its derivatives) has many pleasant consequences. The next order of business is to
give a useful general criterion for a sequence of functions to be uniformly convergent.

288

13. SEQUENCES AND SERIES OF FUNCTIONS

For a function f : I R, we define

||f || = sup |f (x)|.
xI

In (more) words, ||f || is the least M [0, ] such that |f (x)| M for all x I.
Theorem 13.9. (Weierstrass M-Test) Let {fn }
n=0 be a sequence of functions
defined on an interval I.PLet {Mn }
be
a
non-negative
sequence such that ||fn ||
n=0
P

Mn for all n and M = n=0 Mn < . Then n=0 fn is uniformly convergent.

PN
P
Proof. Let SN (x) = n=0 fn (x).
Since n Mn < , for each  > 0 there is
P
N0 Z+ such that for all N N0 , n>N Mn < . For x I, N N0 and k N,


+k
 NX

X
X


|fn (x)|
Mn < .
|SN +k (x) SN (x)| = 
fn (x)


n>N

n=N +1

n>N

Therefore the series is uniformly convergent by the Cauchy criterion.

3. Power Series II: Power Series as (Wonderful) Functions

P
Theorem 13.10. (Wonderful Properties of Power Series) Let P n=0 an xn be a

power series with radius of convergence R > 0. Consider f (x) = n=0 an xn as a

function f : (R, R) R. Then:
a) f is continuous.
b) f is differentiable. Morever, its derivative may be computed termwise:

X
f 0 (x) =
nan xn1 .
n=1
0

c) Since the power series f has the same radius of convergence R > 0 as f , f is
in fact infinitely differentiable.
d) For all n N, f (n) (0) = (n!)an .
Proof.
a) Let P
0 < A < R, so f defines a function from [A, A] to R. We claim that
the series n an xn converges to f uniformly on [A,
as a function
P A]. Indeed,
P
n
n
on [A, A], we have ||an xn || = |an |An , and thus
||a
x
||
=
<
n
n
n |an |A
, because power series converge absolutely on the interior of their interval of
convergence.
Pn Thus by the Weierstrass M -test f is the uniform limit of the sequence
Sn (x) = k=0 ak xk . But each Sn is a polynomial function, hence continuous and
infinitely differentiable. So by Theorem 13.2 f is continuous on [A, A]. Since any
x (R, R) lies in [A, A] for some 0 < A < R, f is continuous
R).
P on (R, P
b) According to Corollary 13.7, in order to show that f = n an xn = n fn is
differentiable and the derivative may be compuited termwise,
it is enough to check
P
that (i) each fn is continuously differentiable and (ii) n fn0 is uniformly convergent.
But (i) is trivial, since fn = an xn of course
Pmonomial
P functions are
P continuously
differentiable. As for (ii), we compute that n fn0 = n (an xn ) = n nan1 xn1 .
By X.X, this power series also has radius of convergence R, hence by the result of
part a) it is uniformly
convergent on [A, A]. Therefore Corollary 13.7 applies to
P
show f 0 (x) = n=0 nan xn1 .
c) We have just seen that for a power series f convergent on (R, R), its derivative
f 0 is also given by a power series convergent on (R, R). So we may continue in

3. POWER SERIES II: POWER SERIES AS (WONDERFUL) FUNCTIONS

289

this way: by induction, derivatives of all orders exist.

d) The formula f (n) (0) = (n!)an is simply what one obtains by repeated termwise
differentiation. We leave this as an exercise to the reader.

Exercise: Prove Theorem 13.10d).
P
Exercise:PShow that if f (x) = n=0 an xn has radius of convergence R > 0, then

an n+1
F (x) = n=0 n+1 x
is an anti-derivative of f .
The following exercise drives home that uniform convergence of a sequence or series
of functions on all of R is a very strong condition, often too much to hope for.
P
Exercise: Let n an xn be a power series with infinite radius of convergence, hence
defining a function
P f : R R. Show that the following are equivalent:
(i) The series n an xn is uniformly convergent on R.
(ii) We have an = 0 for all sufficiently large n.
P
Exercise: Let f (x) = n=0 an xn be a power series with an 0 for all n. Suppose
that the radius of convergence is 1, so that f defines a function on (1, 1). Show
that the following
are equivalent:
P
(i) f (1) = n an converges.
(ii) The power series converges uniformly on [0, 1].
(iii) f is bounded on [0, 1).
The fact that for any power series f (x) =
f

(n)

an xn with positive radius of conver-

(0)
n!

yields the following important result.

P
13.11. (Uniqueness Theorem) Let f (x) = n an xn and g(x) =
P Corollary
n
n bn x be two power series with radii of convergence Ra and Rb with 0 < Ra Rb ,
so that both f and g are infinitely differentiable functions on (Ra , Ra ). Suppose
that for some with 0 < Ra we have f (x) = g(x) for all x (, ). Then
an = bn for all n.
P
P
Exercise: Suppose f (x) = n an xn and g(x) = n bn xn are two power series each
converging on some open interval (A, A). Let {xn }
n=1 be a sequence of elements
of (A, A) \ {0} such that limn xn = 0. Suppose that f (xn ) = g(xn ) for all
n Z+ . Show that an = bn for all n.
gence we have an =

The upshot of Corollary 13.11 is that the only way that two power series can be
equal as functions even in some very small interval around zero
P is if allPoftheir
coefficients are equal. This is not obvious, since in general n=0 an = n=0 bn
does not imply an = bn for all n. Another way of saying this is that the only power
series a function can be equal to on a small interval around zero is its Taylor series.

CHAPTER 14

Serial Miscellany
1.

2
1
n=1 n2 = 6
P 1
n=1 np converges

Let p R. Recall that the p-series

iff p > 1. It is another
matter entirely to determine the sum of the series exactly. In this section we
devote our attention to the p = 2 case.
P
2
Theorem 14.1. (Euler) n=1 n12 = 6 .
Eulers original argument is brilliant but not fully rigorous by modern standards.
Since then several branches of mathematical analysis have been founded which give
systematic tools for finding sums of this and similar series. In particular if one
learns about Fourier series or complex analysis then very natural proofs can
be given, but both of these topics are beyond the scope of an honors calculus course.
On the other hand, in the intervening centuries literally hundreds of proofs of
Theorem 14.1 have been given, some of which use only tools we have developed
(or indeed, no tools beyond standard freshman calculus). Among these we give
here a particularly nice argument due to D. Daners [Da12] following Y. Matsuoka
[Ma61]. In fact this argument barely uses notions from infinite series! Rather,
PN
2
it gives an upper bound on 6 n=1 n12 in terms of N which approaches 0 as
N , and this certainly suffices. Precisely, we will show the following result.
Theorem 14.2. For all positive integers N ,
N

2
2 X 1

.
6
n2
4(N + 1)
n=1

The proof will exploit a family of trigonometric integrals. For n Z+ , we define

Z 2
Z 2
An =
cos2n xdx, Bn =
x2 cos2n xdx.
0

Exercise: Show that An , Bn > 0 for all n 1.

Lemma 14.3. a) For all positive integers n, we have:
Z
(95)

sin2 x cos2(n1) xdx =

An
An1
=
.
2n 1
2n

b) For all positive integers n, we have

(96)

An = (2n 1)nBn1 2n2 Bn .

Proof. a) Integrating by parts gives
Z 2
An =
cos x cos2n1 xdx
0
291

292

14. SERIAL MISCELLANY

= (sin x)(cos2n1 x) |02

(sin x)((2n 1) cos2(n1) x)( sin x)dx

sin2 x cos2(n1) xdx = (2n 1)

= (2n 1)
0

(1 cos2 x) cos2(n1) xdx

= (2n 1)(An1 An ).
Thus

sin2 x cos2(n1) xdx =

An
= An1 An ,
2n 1

and
1
An1
=
2n
2n


An +

An
2n 1


=

1
2n

(2n 1)An + An
2n 1

b) Integrating by parts twice gives1

Z
Z 2
2n
1 cos xdx = 2n
An =
0

Z
=n


=

An
.
2n 1

x sin x cos2n1 xdx



x2 cos x cos2n1 x (2n 1) sin2 x cos2(n1) x dx

Z
= nBn + n(2n 1)

x2 (1 cos2 x) cos2(n1) xdx

= (2n 1)nBn1 2n2 Bn .

Dividing (96) by n2 An and using (95), we get

(2n 1)Bn1
2Bn
2Bn1
2Bn
1
=

.
2
n
nAn
An
An1
An
Thus for all n Z+ we have

N
N 
X
X
2Bn1
1
2Bn
2BN
2B0
=

.
=
2
n
An1
An
A0
AN
n=1
n=1
Since
Z
A0 =

, B0 =
2

dx =
0

we have

x2 dx =

3
,
24

2B0
2
=
.
A0
6

and thus for all N Z+ ,

N
X
1
2
2BN
=

.
2
n
6
AN
n=1

Equivalently
N

(97)

2 X 1
2BN

=
> 0.
2
6
n
AN
n=1

1This time we leave it to the reader to check that the boundary terms uv | 2 evaluate to 0.
0

2. REARRANGEMENTS AND UNORDERED SUMMATION

293

Lemma 14.4. For all x [0, 2 ], we have:

 
2
x sin x

and thus
 2
sin2 x.
x2
2
Exercise: Prove Lemma 14.4. (Hint: use convexity!)
Using Lemma 14.4 and Lemma 14.3a) with N = n 1 we get
Z 2
N
2 X 1
2BN
2
0<
=

=
x2 cos2N xdx
2
6
n
A
A
N
N
0
n=1
Z
2  2
AN
2
2  2 2
=
,
sin2 x cos2N xdx =
AN 2
AN 2
2(N + 1)
4(N + 1)
0
which proves Theorem 14.2.

2. Rearrangements and Unordered Summation

2.1. The Prospect of Rearrangement.
In this section we systematically investigate the validity of the commutative law
for infinite sums. Namely, the definition we gave for convergence of an infinite series
a1 + a2 + . . . + an + . . .
in terms of the limit of the sequence of partial sums An = a1 + . . . + an makes
at least apparent use of the ordering of the terms of the series. Note that this is
somewhat surprising even from the perspective of infinite sequences: the statement
an L can be expressed as: for all  > 0, there are only finitely many terms of the
sequence lying outside the interval (L , L + ), a description which makes clear
that convergence to L will not be affected by any reordering of theP
terms of the

sequence. However, if we reorder the terms {an } of an infinite series n=1 an , the
corresponding change in the sequence An of partial sums is not simply a reordering,
as one can see by looking at very simple examples. For instance, if we reorder
1
1 1 1
+ + + ... + n + ...
2 4 8
2
as
1 1 1
1
+ + + ... + n + ...
4 2 8
2
Then the first partial sum of the new series is 14 , whereas every nonzero partial sum
of the original series is at least 12 .
Thus there is some evidence to fuel suspicion that reordering the terms of an infinite series may not be so innocuous an operation as for that of an infinite seuqence.
All of this discussion is mainly justification for our setting up the rearrangement
problem carefully, with a precision that might otherwise look merely pedantic.
Namely, the formal notion of rearrangement of a series

n=0

an begins with a

294

14. SERIAL MISCELLANY

permuation
of N, i.e., a bijective function
P : N N. We define the rearrangeP
ment of n=0 an by to be the series n=0 a(n) .
2.2. The Rearrangement Theorems of Weierstrass and Riemann.
The most basic questions on rearrangements of series are as follows.
P
Question 14.5. Let n=0 an = S is a convergent infinite series, and let be
a permutation of N. Then: P

a) Does the rearranged series n=0 a(n) converge?

b) If it does converge, does it converge to S?
As usual, the special case in which all terms are non-negative is easiest, the case
of absolute convergence is not much harder than that, and the case of nonabsolute
convergence is where all the real richness and subtlety lies.
P
Indeed, suppose that an 0 for all n. In this case the sum A = n=0 an [0,
Pk
is simply the supremum of the set An =
n=0 ak of finite sums. More generally, let S = {n1 , . . . , nk } be any finite subset of the natural numbers, and put
AS = an1 + . . . + ank . Now every finite subset S N is contained in {0, . . . , N } for
some N N, so for all S, AS AN for some (indeed, for all sufficiently large) N .
This shows that if we define
A0 = sup AS
S

as S ranges over all finite subsets of N, then A0 A. On the other hand, for all
N N, AN = a0 + . . . + aN = A{0,...,N } : in other words, each partial sum AN
arises as AS for a suitable finite subset S. Therefore A A0 and thus A = A0 .
P
The point here is that the description n=0 an = supS AS is manifestly unchanged
by rearranging the terms of the series by any permutation : taking S 7 (S)
gives a bijection on the set of all finite subsets of N, and thus

X
X
a(n) .
an = sup AS = sup A(S) =
n=0

n=0

The case of absolutely convergent series follows rather easily from this.
P
Lemma 14.6. Let n an be a real series with an 0 for all n P
and sum A
[0, ]. Then A is the supremum of the set of all finite sums AS = nS an as S
ranges over all nonempty finite subsets of N.
Proof. Let A be the supremum of the finite sums AS . For N N, let
PN
AN =
n=0 an . Since an 0 for all n, the sequence AN is increasing, so
A = limN AN = supN AN . Since AN = A{0,...,N } , we have A = supN AN
supS AS = A. On the other hand, for any nonempty finite subset S of N, let N be
P
PN
the largest element of S. Then S {0, . . . , N } so AS = nS an n=0 aN =
AN A, so A = sup AS A. Thus A = A.

The point of Lemma 14.6 is that we have expressed the sum of a series with nonnegative terms in a way which is manifestly independent of the ordering of the
terms:2 for any bijection of N, as S = {n1 , . . . , nk } ranges over all finite subsets
2This is a small preview of unordered summation, the subject of the following section.

2. REARRANGEMENTS AND UNORDERED SUMMATION

295

of N, so does (S) = {(n1 ), . . . , (nk )}. It follows that

an =

n=0

a(n) [0, ],

n=0

i.e., rearrangement of a series with non-negative terms does not disturb the convergence/divergence or the sum.
P
Theorem 14.7. (Weierstrass) Let n=0 an be an absolutely convergent
P series
with sum A. Then for every permutation of N, the rearranged series n=0 a(n)
converges to A.
P
P
Proof. Put A = n=0 an . Fix  > 0 and let N0 N be such that n=N0 |an | <
. Let M0 N be sufficiently large so that the terms a(0) , . . . , a(M0 ) include all
the terms a0 , . . . , aN0 1 (and possibly others). Then for all M M0 ,
|

M
X
n=0

a(n) A| = |

M
X
n=0

a(n)

X
n=0

an |

|an | < .

n=N0

Indeed: by our choice of M we know that the terms a0 , . . . , aN0 1 appear in both
PM
P
n=0 a(n) and
n=0 an and thus get cancelled; some further terms may or may
not be cancelled, but by applying the triangle inequality and summing the absolute
values we get an upper bound by assuming no further cancellation. This shows
P
PM

n=0 a(n) = limM
n=0 a(n) = A.
Exercise: a) Give a proof of Step 1 of Theorem 14.7 that bypasses Lemma 14.6.
(Suggestion:
Pby reasoning as in Step 2, argue that for each  > 0 and all sufficiently
large N , n=N |a(n) | < .
P
P +
b) Use the decomposition
of
n an into its series of positive parts
n an and
P
negative parts n an to give a second proof of Step 2 of Theorem 14.7.
P
Theorem 14.8. (Riemann Rearrangement Theorem) Let n=0 an be a nonabsolutely convergent
P series. For any B [, ], there exists a permutation
of N such that n=0 a(n) = B.
Proof. P
Step 1: Since n an is convergent, we have an 0 and thus that {an } is bounded,
so we may choose M such that |an | M for all n. We are not going to give an
explicit formula for ; rather, we are going to describe by a certain process.
For this it is convenient to imagine that the sequence {an } has been sifted into a
disjoint union of two subsequences, one consisting of the positive terms and one
consisting of the negative terms (we may assume without loss of generality that
there an 6= 0 for all n). If we like, we may even imagine both of these subseqeunce
ordered so that they are decreasing in absolute value. Thus we have two sequences
p1 p2 . . . pn . . . 0,
n1 n2 . . . nn . . . 0
so that together {pn , nn } comprise the terms of the series. The key point
Phere is
Proposition
11.30
which
tells
us
that
since
the
convergence
is
nonabsolute,
n pn =
P
, n nn = . So we may specify a rearangement as follows: we specify a choice
of a certain number of positive terms taken in decreasing order and then a choice
of a certain number of negative terms taken in order of decreasing absolute value

296

14. SERIAL MISCELLANY

and then a certain number of positive terms, and so on. As long as we include a
finite, positive number of terms at each step, then in the end we will have included
every term pn and nn eventually, hence we will get a rearrangement.
Step 2 (diverging to ): to get a rearrangement diverging to , we proceed as
follows: we take positive terms p1 , p2 , . . . in order until we arrive at a partial sum
which is at least M + 1; then we take the first negative term n1 . Since |n1 | M ,
the partial sum p1 + . . . + pN1 + n1 is still at least 1. Then we take at least one
more positive term pN1 +1 and possibly further terms until we arrive at a partial
sum which is at least M + 2. Then we take one more negative term n2 , and note
that the partial sum is still at least 2. And we continue in this manner: after the
kth step we have used at least k positive terms, at least k negative terms, and all
the partial sums from that point on will be at least k. Therefore every term gets
included eventually and the sequence of partial sums diverges to +.
Step 3 (diverging to ): P
An easy adaptation of the argument of Step 2 leads to

a permutation such that n=0 a(n) = . We leave this case to the reader.
Step 4 (converging to B R): if anything, the argument is simpler in this case. We
first take positive terms p1 , . . . , pN1 , stopping when the partial sum p1 + . . . + pN1
is greater than B. (To be sure, we take at least one positive term, even if 0 >
B.) Then we take negative terms n1 , . . . , nN2 , stopping when the partial sum
p1 + . . . + pN1 + n1 + . . . + nN2 is less than B. Then we repeat the process, taking
enough positive terms to get a sum strictly larger than B then enough negative
terms to get a sum strictly less than B, and so forth. Because both the positive
and negative parts diverge, this construction can be completed. Because the general
term an 0, a little thought shows that the absolute value of the difference between
the partial sums of the series and B approaches zero.

The conclusion of Theorem 14.8 holds under somewhat milder hypotheses.
P
P +
P
Exercise: Let n an be a real series such that an 0,
n an = and
n an =
. Show that the conclusion of Theorem
14.8
holds:
for
any
A

[,
],
there
P
exists a permutation of N such that n=0 a(n) = A.
P
P
Exercise: Let n an be a real series such that n a+
n = .
a) Suppose that the sequence
{a
}
is
bounded.
Show
that there exists a permutan
P
tion of N such that n a(n) = .
b) Does the conclusion of part a) hold without the assumption that the sequnece
of terms is bounded?
Theorem 14.8 exposes the dark side of nonabsolutely convergent series: just by
changing the order of the terms, we can make the series diverge to or converge
to any given real number! Thus nonabsolute convergence is necessarily of a more
delicate and less satisfactory
P nature than absolute convergence. With these issues
in mind, we define a series n an to be unconditionally convergent if it is convergent and every rearrangement converges to the same sum, and a series to be
conditionally convergent if it is convergent but not unconditionally convergent.
Then much of our last two theorems may be summarized as follows.
Theorem 14.9. (Main Rearrangement Theorem) A convergent real series is
unconditionally convergent if and only if it is absolutely convergent.

2. REARRANGEMENTS AND UNORDERED SUMMATION

297

Many texts do not use the term nonabsolutely convergent and instead define a series to be conditionally convergent if it is convergent but not absolutely convergent.
Aside from the fact that this terminology can be confusing to students to whom
this rather intricate story of rearrangements has not been told, it seems correct to
make a distinction between the following two a priori different phenomena:
P
P
Pn an converges but n |an | does not, versus P
n an converges to A but some rearrangement n a(n) does not.
As we have seen, these two phenomena
are equivalent for real series. However
P
the notion of an infinite series
a
,
absolute
and unconditional convergence
n
n
makes sense in other contexts, for instance3 for series with values in an infinitedimensional Banach space or with values in a p-adic field. In the former case
it is a celebrated theorem of Dvoretzky-Rogers [DR50] that there exists a series
which is unconditionally convergent but not absolutely convergent, whereas in the
latter case one can show that every convergent series is unconditionally convergent
whereas there exist nonabsolutely convergent series.
P
Exercise: Let n=0 an be any nonabsolutely convergent real series, and let
a A . Show
P that there exists a permutation of N such that the set of
partial limits of n=0 a(n) is the closed interval [a, A].
2.3. Unordered summation.
It is very surprising that the ordering of the terms of a nonabsolutely convergent
series affects both its convergence and its sum it seems fair to say that this phenomenon was undreamt of by the founders of the theory of infinite series.
ArmedP
now, as we are, with the full understanding of the implications of our defini
ion of n=0 an as the limit of a sequence of partial sums, it seems reasonable to
ask: is there an alternative definition for the sum of an infinite series, one in which
the ordering of the terms is a priori immaterial?
The answer to this question is yes and is given by the theory of unordered
summation.
To be sure to get a definition of the sum of a series which does not depend on
the ordering of the terms, it is helpful to work in a context in which no ordering is
present. Namely, let S be a nonempty set, and define an S-indexed sequence of
real numbers to be a function a : S R. The point here is that we recover the
usual definition of a sequence by taking S = N (or S = Z+ ) but whereas N and Z+
come equipped with a natural ordering, the naked set S does not.
P
We wish to define
sS as , i.e., the unordered sum of the numbers as as s
ranges over all elements
P of S. Here it is: for every finite subset T = {s1 , . . . , sN }
of S, we define aT = sT as = as1 + . . . + aP
sN . (We also define a = 0.) Finally,
for A R, we say that the unordered sum sS as converges to A if: for all
3Both of these are well beyond the scope of these notes, i.e., you are certainly not expected
to know what I am talking about here.

298

14. SERIAL MISCELLANY

 > 0, there exists a finite subset T0 S such that for allPfinite subsets T0 T S
we
P have |aT A| < . If there exists A R such that sS as = A, we say that
sS as is convergent or that the S-indexed sequence a is summable. (When
S = ZN we already have a notion of summability, so when we need to make the
distinction we will say unordered summable.)
Notation: because we are often going to be considering various finite subsets T
of a set S, we allow ourselves the following time-saving notation: for two sets A
and B, we denote the fact that A is a finite subset of B by A f B.
Exercise: Suppose S is finite.PShow that every S-indexed sequence a : S R
is summable, with sum aS = sS as .
Exercise: If S = , there is a unique function a : R, the
P empty function. Convince yourself that the most reasonable value to assign s as is 0.
Exercise: Give reasonable definitions for

sS

as = and

sS

as = .

Confusing Remark: We say we are doing unordered summation, but our sequences take values in R, in which the absolute value is derived from the order
structure. One could also consider unordered summation of S-indexed sequences
with values in an arbtirary normed abelian group (G, | |).4 A key feature of R is
the positive-negative part decomposition, or equivalently the fact that for M R,
|M | A implies M A or M A. In other words, there are exactly two
ways for a real number to be large in absolute value: it can be very positive or very
negative. At a certain point in the theory considerations like this must be used in order to prove the desired results, but we delay such arguments for as long as possible.
The following result holds without using the positive-negative decomposition.
Theorem 14.10. (Cauchy Criteria for Unordered Summation) Let a : S R
be an S-index sequence, and consider the following assertions:
(i) The S-indexed sequence a is summable.
(ii) For all  > 0, there exists a finite subset T S such that for all finite subsets
T, T 0 of S containing T ,
X
X
|
as
as | < .
sT

sT 0

(iii) For all  > 0,

Pthere exists T f S such that: for all T f S with T T = ,
we have |aT | = | sT as | < .
(iv) There exists M R such that for all T f S, |aT | M .
Proof. (i) = (ii) is immediate from the definition.
(ii) = (i): We may choose, for each n Z+ , a finite subset Tn of S such that
Tn Tn+1 for all n and such that for all finite subsets T, T 0 of S containing Tn ,
|aT aT 0 | < 2 . It follows that the real sequence {aTn } is Cauchy, hence convergent,
say to A. We claim that a is summable to A: indeed, for  > 0, choose n > 2 .
4And its not totally insane to do so: these arise in functional analysis and number theory.

2. REARRANGEMENTS AND UNORDERED SUMMATION

299

Then, for any finite subset T containing Tn we have



+ = .
2 2
(ii) = (iii): Fix  > 0, and choose T0 f as in the statement of (ii). Now let
T f S with T T0 = , and put T 0 = T T0 . Then T 0 is a finite subset of S
containing T0 and we may apply (ii):
X
X
X
|
as | = |
as
as | < .
|aT A| |aT aTn | + |aTn A| <

sT 0

sT

sT0

(iii) = (ii): Fix  > 0, and let T f S be such that for all finite subsets T of S
with T T = , |aT | < 2 . Then, for any finite subset T 0 of S containing T ,

|aT 0 aT | = |aT 0 \T | < .
2
From this and the triangle inequality it follows that if T and T 0 are two finite
subsets containing T ,
|aT aT 0 | < .
(iii) = (iv): Using (iii), choose T1 f S such that for all T 0 f S with T1 T 0 =
, |aT 0 | 1. Then for any T f S, write T = (T \ T1 ) (T T1 ), so
X
X
X
as | + |
as | 1 +
|as |,
|aT | |
sT T1

sT \T1

so we may take M = 1 +

sT1

|as |.

sT1

Confusing Example: Let G = Zp with its standard norm. Define an : Z+ G

by an = 1 for all n. Because of the non-Archimedean nature of the norm, we have
for any T f S |aT | = |#T | 1. Therefore a satisfies condition (iv) of Theorem
14.10 above but not condition (iii): given any finite subset T Z+ , there exists a
finite subset T 0 , disjoint from T , such that |aT 0 | = 1: indeed, we may take T 0 = {n},
where n is larger than any element of T and prime to p.
Although we have no reasonable expectation that the reader will be able to
make any sense of the previous example, we offer it as motivation for delaying
the proof of the implication (iv) = (i) above, which uses the positive-negative
decomposition in R in an essential way.
Theorem 14.11. An S-indexed sequence a : S R is summable iff the finite
sums ae uniformly bounded: i.e., there exists M R such that for all T f S,
|aT | M .
Proof. In Theorem 14.10 above we showed that if a is summable, the finite
sums are uniformly bounded. Now suppose a is not summable, so by Theorem
14.10 there is  > 0 with the following property: for any T f S there is T 0 f S
with T T 0 = and |aT 0 | . Of course, if we can find such a T 0 , we can also find
a T 00 disjoint from T T 0 with |aT 00 | , and so forth: there will be a sequence
{Tn }
n=1 of pairwise disjoint finite subsets of S such that for all n, |aTn | . But
now decompose Tn = Tn+ Tn , where Tn+ consists of the elements s such that
as 0 and Tn consists of the elements s such that as < 0. It follows that
aTn = |aTn+ | |aTn |
hence
 |aTn | |aTn+ | + |aTn |,

300

14. SERIAL MISCELLANY

from which it follows that max |aTn+ , aTn | 2 , so we may define for all n a subset
Tn0 Tn such that |aTn0 | 2 and the sum aTn0 consists either entirely of non-negative
0
, then by
elements of entirely of negative elements. If we now consider T10 , . . . , T2n1
the Pigeonhole Principle there must exist 1 i1 < . . . < in 2n 1 such that all
the terms
Ti0 are non-negative or all the terms in each Ti0 are negative. Let
Sn in each
0
Tn = j=1 Tij . Then we have a disjoint union and no cancellation, so |Tn | n
2 :
the finite sums aT are not uniformly bounded.

Proposition 14.12. Let a : S R be an S-indexed sequence with as 0 for
all s S. Then
X
as = sup aT .
sS

T f S

Proof. Let A = supT f S aT .

We first suppose that A < . By definition of the supremum,, for any  > 0,
there exists a finite subset T S such that A  < aT A. Moreover, for any
finite subset T 0 T , we have A aT aT 0 A, so a A.
Next suppose A = . We must show that for any M R, there exists a subset
TM f S such that for every finite subset T TM , aT M . But the assumption
A = implies there exists T f S such that aT M , and then non-negativity
gives aT 0 aT M for all finite subsets T 0 T .

Theorem 14.13. (Absolute Nature of Unordered Summation) Let S be any set
and a : S R be an S-indexed sequence. Let |a | be the S-indexed sequence
s 7 |as |. Then a is summable iff |a | is summable.
Proof. Suppose |a | is summable. Then for any  > 0, there exists T such
that for all finite subsets T of S disjoint from T , we have ||a|T | < , and thus
X
X
|aT | = |
as | |
|as || = ||a|T | < .
sT

sT

Suppose |a | is not summable. Then by Proposition 14.12, for every M > 0, there
exists T f S such that |a|T 2M . But as in the proof of Theorem 14.11, there
must exist a subset T 0 T such that (i) aT 0 consists entirely of non-negative terms
or entirely of negative terms and (ii) |aT 0 | M . Thus the partial sums of a are
not uniformly bounded, and by Theorem 14.11 a is not summable.

Theorem 14.14. For
P a : N R an ordinary sequence and A R, TFAE:
(i) The unordered sum nZ+ an is convergent, with sum A.
P
(ii) The series n=0 an is unconditionally convergent, with sum A.
Proof. (i) = (ii): Fix  > 0. Then there exists T f S such that for every
finite subset T of N containing T weP
have |aT A| < . Put N = maxnT n. Then
n
for all P
n N , {0, . . . , n} T so | k=0 ak A| < . It follows that the infinite

series n=0 an converges

P to A in the usual sense. Now for any permutation of
N, the unordered sum nZ+ a(n) is manifestly the same as the unordered sum
P
P
nZ+ an , so the rearranged series
n=0 a(n) also converges to A.
P
(ii) = (i): We will prove the contrapositive: suppose the unordered sum nN an
is divergent.
Then by Theorem 14.11 for every M 0, there exists T S with
P
|aT = sT as | M . Indeed, as the proof of that result shows, we can choose T
to be disjoint from any given finite subset. We leave it to you to check that we can
therefore build a rearrangement of the series with unbounded partial sums.


2. REARRANGEMENTS AND UNORDERED SUMMATION

301

Exercise: Fill in the missing details of (ii) = (i) in the proof of Theorem 14.14.
Exercise*: Can one prove Theorem 14.14 without appealing to the fact that |x| M
implies x M or x M ? For instance, does Theorem 14.14 holds for S-indexed
sequences with values in any Banach space? Any complete normed abelian group?
Comparing Theorems 14.12 and 14.13 we get a second proof of the portion of
the Main Rearrangement Theorem that says that a real series is unconditionally
convergent iff it is absolutely convergent. Recall that our first proof of this depended on the Riemann Rearrangement Theorem, a more complicated result.
On the other hand, if we allow ourselves to use the previously derived result
that unconditional convergence and absolute convergence
coincide, then we can get
P
an easier
proof
of
(ii)
=
(i):
if
the
series
a
is
unconditionally
convergent,
n
n
P
then n |an | < , so by Proposition 14.10 the unordered sequence |a | is summable, hence by Theorem 14.12 the unordered sequence a is summable.
To sum up (!), when we apply the very general definition of unordered summability to the classical case of S = N, we recover precisely the theory of absolute
(= unconditional) convergence. This gives us a clearer perspective on exactly what
the usual, order-dependent notion of convergence is buying us: namely, the theory
of conditionally convergent series. It may perhaps be disappointing that such an
elegant theory did not gain us anything new.
However when we try to generalize the notion of an infinite series in various
ways, the results on unordered summability become very helpful. For instance,
often in nature one encounters biseries

X
an
n=

and double series

X

am,n .

m,nN

We may treat the first case as the unordered sum associated to the Z-indexed sequence n 7 an and the second as the unordered sum associated to the NN-indexed
sequence (m, n) 7 am,n and we are done: there is no need to set up separate theories of convergence. Or, if we prefer, we may shoehorn these more ambitiously
indexed series into conventional N-indexed series: this involves choosing a bijection
b from Z (respectively N N) to N. In both cases such bijections exist, in fact
in great multitude: if S is any countably infinite set, then for any two bijections
b1 , b2 : S N, b2 b1
: N N is a permutation of N. Thus the discrepancy
1
between two chosen bijections corresponds precisely to a rearrangement of the series. By Theorem 14.13, if the unordered sequence is summable, then the choice of
bijection b is immaterial, as we are getting an unconditionally convergent series.
The theory of products of infinite series comes out especially cleanly in this unordered setting (which is not surprising, since it corresponds to the case of absolute
convergence, where Cauchy products are easy to deal with).
Exercise: Let S1 and S2 be two sets, and let a : S1 R, b : S2 R. We

302

14. SERIAL MISCELLANY

assume the following nontriviality condition: there exists s1 S1 and s2 S2 such

that as1 6= 0 and as2 6= 0. We define (a, b) : S1 S2 R by
(a, b)s = (a, b)(s1 ,s2 ) = as1 bs2 .
a) Show that a and b are both summable iff (a, b) is summable.
b) Assuming the equivalent conditions of part a) hold, show
!
!
X
X
X
(a, b)s =
as1
bs2 .
sS1 S2

s1 S1

s2 S2

c) When S1 = S2 = N, compare this result with the theory of Cauchy products we

have already developed.
Exercise: Let S be an uncountable set,5 and let a : S R be an S-indexed
sequence. Show that if a is summable, then {s S | as 6= 0} is countable.
3. Abels Theorem
3.1. Statement and Proof.
Theorem 14.15. (Abels Theorem)
P

Let n=0 an be
series.
Pa convergent
n
a) The series n=0 aP
n x is uniformly
Pconvergent on [0, 1].

b) We have limx1 n=0 an xn = n=0 an .

P
P
n
Proof. a) ([C, p. 47]) Since
n=0 an 1 =
n=0 an , convergence at x =
1 is our hypothesis. By our work on power series specifically
P Lemma 11.32
convergence at 1 implies convergence on (1, 1), and thus n=0 an xn converges
pointwise on [0, 1]. Since we have convergence
P at x = 1, it suffices to show uniform
+
convergence
on
[0,
1).
Fix

>
0;
because
n an converges, there is N Z such
P
that | n=N an | < . Now we apply Abels Lemma (Proposition 10.2) with an
sequence aN , aN +1 , . . ., withPbn sequence xN , xN +1 , . . . (note this is positive and

decreasing) and with M = | n=N an |. The conclusion is that for any k N,

|

N
+k
X

an xn | xN M xN  < .

n=N

PN
u
By the Cauchy P
Criterion (Lemma 13.1), n=0 an xn f on [0, 1).

b) By part a), n=0 an xn is a uniform limit of continuous functions (indeed, of

polynomials) on [0, 1], so by Theorem 13.3, itPis continuousP
on [0, 1]. In particular
P

n
n

a
x
is
continuous
at
x
=
1:
lim
a
x
=

n
n
x1
n=0
n=0
n=0 an .
P
Remark: Usually Abels Theorem means part b) of the above result: n=0 an =
limx1 an xn . But this is an immediate consequence of the uniformity of the
convergence on [0, 1], so having this statement be part of Abels Theorem gives a
stronger and also more conceptually transparent result.

5Here we are following our usual convention of allowing individual exercises to assume knowledge that we do not want to assume in the text itself. Needless to say, there is no need to attempt
this exercise if you do not already know and care about uncountable sts.

3. ABELS THEOREM

303

Above we followed an exercise in the text [C] of Cartan.6 For comparison, here
is a different proof of Theorem 14.15b) from a famous text of Rudin.
P
Proof. [R, Thm. 8.2] Since n an converges,
{an } is bounded, so by CorolP
lary 11.35 the radius of convergence of f (x) = n an xn is at least
P1. Put A1 = 0;
for n 0, put An = a0 + . . . + an ; and put A = limn An = n=0 an . Then
N
X

an xn =

n=0

N
X

(An An1 )xn = (1 x)

n=0

N
1
X

An x n + AN x N .

n=0

For each fixed x [0, 1), we let N to get

f (x) =

an xn = (1 x)

n=0

An x n .

n=0

Now fix  > 0, and choose N such that n N implies |A An | < . Then, since
(1 x)

(98)

xn = 1

n=0

for all x [0, 1), we get

|f (x) A| = |(1 x)

An x A(1 x)

n=0

(1 x)

N
X
n=0

|An A|xn +

x | = |(1 x)

n=0


2

(1 x)

X
n=N +1

xn (1 x)

(An A)xn |

n=0
N
X

|An A|xn + .

n=0

The last quantity above approaches  as x approaches 1 from the left. Since  was
arbitrary, this shows limx1 f (x) = A.

As you can see, Rudins proof uses much less than Cartans: rather than relying on
Abels Lemma, a bit of partial summation is done on the fly. Moreover, most of the
appeals to the theory of power series and uniform convergence are replaced by a
clever introduction of the geometric series! Nevertheless I must say that, although
Rudins argument is easy enough to follow line by line, in terms of whats going
on in the proof I find it absolutely impenetrable.
The rest of this section is an extended exercise in Abels Theorem appreciation.
First of all, it may help to restate the result in a form which is slightly more general
and moreover makes more clear exactly what has been established.
P
Theorem 14.16. (Abels Theorem Mark II) Let f (x) = n an (x c)n be a
power series with radius of convergence R > 0, hence convergent at least for all
x (c R, c + R).
a) Suppose that the power series converges at x = c + R. Then the function f :
(c R, c + R] R is continuous at x = c + R: limx(c+R) f (x) = f (c + R).
b) Suppose that the power series converges at x = c R. Then the function f :
[c R, c + R) R is continuous at x = c R: limx(cR)+ f (x) = f (c R).
6Henri Cartan (1904-2008) was one of the leading mathematicians of the 20th century.

304

14. SERIAL MISCELLANY

Exercise: Prove Theorem 14.16.

P
1
Exercise: Consider f (x) = 1x
= n=0 xn , which converges for all x (1, 1).
Show limx1+ f (x) exists, so f extends to a continuous function on [1, 1). Nevertheless f (1) 6= limx1+ f (x). Why doesnt this contradict Abels Theorem?
3.2. An Application to the Cauchy Product.
As our first application, we round out our treatment of Cauchy products by showing
that the Cauchy product never wrongly converges.
P
P
Theorem 14.17. Let n=0 an be a series converging
Pn to A and n=0 bn be a
series converging
define cn = k=0 ak bnk and the Cauchy
Pto B. As usual, weP

product series n=0 cn . Suppose that n=0 cn converges to C. Then C = AB.

P
P
P
Proof. Put f (x) = n=0 an xn , g(x) = n=0 bn xn and h(x) = n=0 cn xn .
By assumption, f (x), g(x)
P and h(x)
P all convergePat x = 1, so by Lemma 11.32 the
radii of convergence of n an xn , n bn xn and n cn xn are all at least one. Now
all we need to do is apply Abels Theorem:
AT

C = h(1) = lim h(x) = lim f (x)g(x)

x1
x1



AT
lim g(x) = f (1)g(1) = AB.
= lim f (x)
x1

x1

3.3. Two Identities Justified By Abels Theorem.

Here are two surprising and beautiful identities.

X
1 1 1
(1)n+1
= 1 + + . . . = log 2.
n
2 3 4
n=1

(99)

X
(1)n
1 1 1

= 1 + + ... = .
2n + 1
3 5 7
4
n=0

(100)

The identity (100) was known to Leibniz (as would be logical, given that the
convergence of both series follows from Leibnizs Alternating Series Test), where
the quotation marks are meant to suggest that Leibniz probably did not have an
argument we would accept as a rigorous proof. Suppose someone shows you such
identities, as I now have: what would you make of them?
A good first reaction is to attempt numerical verification. Even this is not as
easy as one might expect, because the convergence of both series is rather slow.
(In particular, among all ways one might try to numerically compute , (100) is
one of the worst I know.) Like any series which is shown to be convergent by the
Alternating Series Test, there is a built in error estimate for the sum: if we cut off
after N terms the error is in absolute value at most |aN +1 |. The problem with this
is that the N th terms of these series tend to zero quite slowly! So for instance,
4

S104

10
X
(1)n+1
= 0.6930971830599452969172323714 . . .
=
n
n=1

3. ABELS THEOREM

305

(Using a software package I asked for the exact sum of the series, which is of course
a rational number, but a very complicated one: it occupies more than one full screen
on my computer. The amount of time spent to compute this rational number was
small but not instantaneous. In fact the reason I chose 104 is that the software managed this but had trouble with S105 . Then I converted the fraction to a decimal.
Of course a much better way to do this would be to convert the fractionals to decimals as we go along, but this needs to be done carefully to prevent rounding errors:
to be serious about this sort of thing one needs to know some numerical analysis.)
By the Alternating Series Test, the difference between the infinite sum and the
1
, so we are guaranteed (roughly) four decimal places
finite sum S104 is at most 1001
of accuracy. For comparison, computing log 2 e.g. by writing it as log( 12 ) and
using the Taylor series for log(1 + x) we get
log 2 = 0.6931471805599453094172321214 . . . .
So indeed the identity (99) holds true at least up to four decimal places. If we
wanted to do much more numerical verification than this, we would probably have
to do something a little more clever.
Similarly, we have
4

10
X
(1)n
= 0.78542316 . . . ,
2n + 1
n=0

and by the Alternating Series test this approximates the infinite sum
to at least four decimal places of accuracy, whereas

= 0.7853981633974483096156608458 . . . ,
4
which shows that (100) holds true at least up to four decimal places.

(1)n
n=0 2n+1

Okay, so these identities are probably true: how do we prove them?!?

We exploit our knowledge of power series. First, take f (x) = log(1 + x). Then
(101)

f 0 (x) =

X
X
1
1
(1)n xn .
(x)n =
=
=
1+x
1 (x) n=0
n=0

Integrating termwise gives

(102)

log(1 + x) = f (x) = f (0) +

X
X
(1)n xn+1
(1)n xn+1
=
.
n+1
n+1
n=1
n=1

(This was known to Newton and Leibniz.) So aha! we plug in x = 1 to get

log(2) = log(1 + 1) =

X
(1)n
.
n+1
n=1

But not so fast. Although you will find this explanation in many freshman calculus
books, it is not yet justified. We were being sloppy in our above work by writing
down power series expansions and not keeping track of the interval of convergence.
The identity (101) holds for x (1, 1), and thats really the best we can do, since
the power series on the right hand side does not converge for any other values of x.

306

14. SERIAL MISCELLANY

Integrating this term by term, we find that (102) holds for x (1, 1). Above, in
our excitement, we plugged in x = 1: so close, but out of bounds. Too bad!
But dont despair: its Abels Theorem for the win! Indeed, because the series
converges at x = 1 and the function log x is defined and continuous at x = 1,

X
(1)n xn+1 AT X (1)n
log(2) = log(1 + 1) = lim log x = lim
=
.
n+1
n+1
x1
x1
n=1
n=1
Exercise: Establish (100) similarly, starting with the function f (x) = arctan x.
3.4. Abel Summability.
Abels Theorem gives rise to a summability
method: a way to extract numerical
P
values out of certain divergent series n an as though they converged. Instead
of forming the sequence of partial sums
P An = a0 + . . . + an and taking
P the limit,
suppose instead we look at limx1 n=0 ax xn . We say the series
P n an is Abel
summable if this limit exists, in which case we write it as A n=0 an , the
PAbel
sum of the series. The point of this is that by Abels theorem, if a series n=0 an
is actually convergent, say to A, then it is also Abel summable and its Abel sum is
also equal to A. However, there are series which are divergent yet Abel summable.
P
Example 14.1. Consider the series n=0 (1)n . As we saw, the partial sums
alternate between 0 and 1 so the series does not diverge. We mentioned earlier
that (the
great) L. Euler believed that nevertheless the right number to attach to the
P
series n=0 (1)n is 21 . Since the two partial limits of the sequence of partial sums
are 0 and 1, it seems vaguely plausible to split the difference.
Theorem provides a much more convincing argument. The power series
P Abels
n n
(1)
x
converges for all x with |x| < 1, and moreover for all such x we have
n

(1)n xn =

lim

x1

That is, the series

(x)n =

n=0

n=0

and thus

n (1)

1
1
=
,
1 (x)
1+x

(1)n xn = lim

n=0
n

x1

1
1
= .
1+x
2

is divergent but Abel summable, with Abel sum 21 .

So Eulers idea was better than we gave him credit for.

P
Exercise: Suppose that
series with an 0 for
n. Show the
n=0 an is a P
Pall

converse of Abels Theorem: if limx1 n=0 an xn = L, then n=0 an = L.

4. The Peano-Borel Theorem
4.1. Statement.
The Taylor series of a smooth (i.e., infinitely differentiable) function is surely one
of the most useful concepts of basic calculus. It is only reasonable to concentrate
first on the good news here: most naturally occurring functions, including all of
the elementary functions one meets in precalculus mathematics, are analytic: at
every point c in their domain I there is a c > 0 such that the Taylor series Tf,c

4. THE PEANO-BOREL THEOREM

307

converges to f on (c c , c + c ).
This good news builds up a strong intuition that smooth functions should have
well-behaved Taylor series. In fact this is far from the case. To fix ideas, consider
smooth functions f : R R. Then there are two ways for f to fail to be analytic:
(i) At some c R, Tf,c converges only at 0.
(ii) At some c R, Tf,c has positive radius of convergence, but there is no interval
(c c , c + c ) about c on which Tf,c = f .
Both of these pathologies occur. In the second case, an example was constructed
by Cauchy in 1823 and remains well known to this day.
(
2
ex x 6= 0
Proposition 14.18. Let f (x) =
. Then:
0
x=0
a) f : R R is smooth.
b) For all n N, f (n) (0) = 0.
c) Thus the Taylor series expansion of f at 0 is the identically zero function. It
has infinite radius of convergence and is equal to f only at the central point x = 0.
Exercise: Prove Proposition 14.18.
That a smooth function can have a Taylor series which converges only at the central
point which we may certainly assume to be 0 is more subtle. The first example
of such a function was given by du Bois-Reymond in 1876 [dBR76], [dBR83].
P
It is no problem to construct a power series
an xn which converges only at 0. By
1
n
the Cauchy-Hadamard formula we need only select a sequence with lim sup |a
Pn | =
an+1
, which as we have seen is implied by = limn an = , so e.g.
n!xn
works. But is this power series a Taylor series? It is if and only if we can find a
(n)
smooth function f : R R with f n!(0) = n!, i.e., that
n N, f (n) (0) = (n!)2 .
Thus the essential problem is to construct a smooth function f : R R whose
sequence of derivatives at 0 grows (at least along a subsequence) sufficiently rapidly,
and in particular much more rapidly than n!.
It is not so auspicious to search in nature for a function with such rapidly
growing derivatives. It turns out to be simpler solve a more general problem: show
that any real sequence whatsoever can serve as the sequence of derivatives at 0 of a
smooth function f : R R. This is the content of the following remarkable result.
Theorem 14.19. (Peano-Borel) Let {cn }
n=0 be a real sequence. Then there
is an infinitely differentiable function f : R R such that for all n N we have
f (n) (0) = cn .
Exercise: Show that the Peano-Borel Theorem is equivalent to the statement that
every power series is a Taylor series. More precisely, for any c R and any real
sequence {an }
differentiable function f : R R such
n=0 , show there is an infinitely P

that the Taylor series expansion of f at c is n=0 an (x c)n .

308

14. SERIAL MISCELLANY

Exercise: Prove du Bois-Reymonds Theorem that there is an infinitely differentiable function f : R R whose Taylor series expansion at 0 converges only at
x = 0.
Exercise: Let {an }
n=0 be a real sequence.
a) Show that
S = {infinitely differentiable f : R R | the Taylor series of f at 0 is

an xn }

n=0

is infinite.
b) (This part is for those who know some linear algebra.) Let f0 S. Show:
{f f0 | f S} is a real vector space of infinite dimension.
Theorem 14.19 is generally known as Borels Theorem. It was proved along
Borel [Bo95]. But acwith the Heine-Borel Theorem in the doctoral thesis of E.

cording to a recent article of A. Besenyei [Be14] the result was in fact first proved
by G. Peano in 1884, so we speak of the Peano-Borel Theorem.
In fact Besenyeis article is the source of this entire section, both the historical
material and the account of Peanos proof that we will now give.
4.2. Proof.
The following argument uses a bit of complex numbers, which are presented in
more detail in a later chapter.

Step 1: Let {an }

n=0 , {bn }n=0 be real sequences with bn 0 for all n N. Put

f (x) =

X
k=0

ak xk
.
1 + bk x 2

Suppose f converges on all of R and that for all n N and x R we have

(n)

X
ak xk
(n)
f (x) =
.
1 + bk x 2
k=0

Later we will show how this can be achieved with suitable choices of the sequences
{an } and {bn }. Now we have
|bk x2 | < 1 =

X
X
ak xk
k
j j 2j
=
a
x
(1)
b
x
=
(1)j ak bjk x2j+k .
k
k
1 + bk x 2
j=0
j=0

Thus


ak xk
1 + bk x2

(n)

(
n!(1)j an2j bjn2j
(0) =
0

if k = n 2j for some j
otherwise.

Thus
(103)

f (0) = a0 , f 0 (0) = a1 , and

4. THE PEANO-BOREL THEOREM

309

b2c
X
f (n) (0)
n 2,
= an +
(1)j an2j bjn2j .
n!
j=1

(104)

From (103) and (104) it follows that given {bn } and {cn }, there is a uniquely
determined sequence {an } such that f (n) (0) = cn for all n N.
Step 2: Let b > 0 and consider


xk1
1
1
xk
=
+
.
b2 + x 2
2
x + bi x bi
By the Generalized Leibniz Rule, we have


xk
b2 + x2

(n)

n  
1X n
=
(k 1) (k 1 n + j)xk1n+j
2 j=0 j

(1)j j!
(1)j j!
+
(x + bi)j+1
(x bi)j+1



n
n! kn2 X
xj+1
xj+1
j (k 1) (k 1 n + j)
= x
(1)
+
.
2
(n j)!
(x + bi)j+1
(x bi)j+1
j=0
We have (see the following exercise)




xj+1
xj+1


(105)
 (x + bi)j+1 + (x bi)j+1  2.
It follows that
(106)


(n) 

ak xk

 (n + 1)!k!|ak | kn2
k n + 2, 
|x|
.

 1 + bk x2

bk

Taking bk = (k!)2 |ak |, we get



X  ak xk (n) 
X |x|kn2
(107)

(n
+
1)!
.


 1 + bk x2

k!
kn+2

kn+2

The right hand side of (107) is uniformly convergent on every bounded interval,
hence for every n N the convergence of
(n)

X
ak xk
1 + bk x 2

k=0

follows from the Weierstrass M-Test. Applying Theorem 13.6 establishes the fact
that the given series of nth derivatives converges to the nth derivative of f . This
completes the proof.
Exercise: a) Let w C. Show that if |w| 1, then for all n Z+ , |wn + wn | 2.
b) Verify equation (105).

310

14. SERIAL MISCELLANY

5. The Weierstrass Approximation Theorem

5.1. Statement of Weierstrass Approximation.
Theorem 14.20. (Weierstrass Approximation Theorem) Let f : [a, b] R be
a continuous function and  any positive number. Then there exists a polynomial
function P such that for all x [a, b], |f (x) P (x)| < . In other words, any
continuous function defined on a closed, bounded interval is the uniform limit of a
sequence of polynomials.
Exercise: For each n Z+ , let Pn : R R be a polynomial function. Suppose
u
that there is f : R R such that Pn f on all of R. Show that the sequence
of functions {Pn } is eventually constant: there exists N Z+ such that for all
m, n N , Pn (x) = Pm (x) for all x R.
It is interesting to compare Theorem 14.20 with Taylors theorem, which gives
conditions for a function to be equal to its Taylor series. Note that any such
function must be C (i.e., it must have derivatives of all orders), whereas in the
Weierstrass Approximation Theorem we can get any continuous function. An important difference is that the Taylor polynomials TN (x) have the property that
TN +1 (x) = TN (x) + aN +1 xN , so that in passing from one Taylor polynomial to the
next, we are not changing any of the coefficients from 0 to N but only adding a
higher order term. In contrast, for the sequence of polynomials Pn (x) uniformly
converging to f in Theorem 1, Pn+1 (x) is not required to have any simple algebraic
relationship to Pn (x).
Theorem 14.20 was first established by Weierstrass in 1885. To this day it is one
of the most central and celebrated results of mathematical analysis. Many mathematicians have contributed novel proofs and generalizations, notably S.J. Bernstein
[Be12] and M.H. Stone [St37], [St48]. But more than any result of undergraduate mathematics I can think of except the quadratic reciprocity law the
passage of time and the advancement of mathematical thought have failed to single
out any one preferred proof. We have decided to follow an argument given by Noam
Elkies.7 This argument is reasonably short and reasonably elementary, although as
above, not definitively more so than certain other proofs. However it unfolds in a
logical way, and every step is of some intrinsic interest. Best of all, at a key stage
we get to apply our knowledge of Newtons binomial series!
5.2. Piecewise Linear Approximation.
A function f : [a, b] R is piecewise linear if it is a continuous function made up
out of finitely many straight line segments. More formally, there exists a partition
P = {a = x0 < x1 . . . < xn = b} such that for 1 i n, the restriction of f to
[xi1 , xi ] is a linear function. For instance, the absolute value function is piecewise
linear. In fact, the general piecewise can be expressed in terms of absolute values
of linear functions, as follows.

7http://www.math.harvard.edu/elkies/M55b.10/index.html

5. THE WEIERSTRASS APPROXIMATION THEOREM

311

Lemma 14.21. Let f : [a, b] R be a piecewise linear function. Then there is

n Z+ and a1 , . . . , an , m1 , . . . , mn , b R such that
n
X
f (x) = b +
|mi xi + bi |.
i=1

Proof. We leave this as an elementary exercise. Some hints:

(i) If aj a, then as functions on [a, b], |x aj | = x aj .
(ii) The following identities may be useful:
|f + g| |f g|
+
2
2
|f + g| |f g|
min(f, g) =

.
2
2
(iii) One may, for instance, go by induction on the number of corners of f .
max(f, g) =

Now every continuous function f : [a, b] R may be uniformly approximated

by piecewise linear functions, and moreover this is very easy to prove.
Proposition 14.22. (Piecewise Linear Approximation) Let f : [a, b] R be a
continuous function and  > 0 be any positive number. Then there exists a piecewise
linear function P such that for all x [a, b], |f (x) P (x)| < .
Proof. Step 1: Let f : [c, d] R be a continuous function, and put M =
(f, [c, d]) = max(f, [c, d]) min(f, [c, d]). Let L : [c, d] R be the unique linear
function with L(c) = f (c) and L(d) = f (d). We claim that for all x [c, d],
|f (x) L(x)| M . To establish the claim we may argue (for instance) as follows:
for all x, either f (x) L(x) or f (x) L(x). In the former case,
|f (x) L(x)| max(f, [c, d]) min(L, [c, d]) = max(f, [c, d]) min(L(c), L(d))
= max(f, [c, d]) min(f (c), f (d)) max(f, [c, d]) min(f, [c, d]) = M,
and the argument in the latter case is very similar.
Step 2: By the Uniform Continuity Theorem, there is > 0 such that whenever
|x y| < , |f (x) f (y)| < . Choose n large enough so that ba
n < , and consider
the partition Pn = {a = x0 < x1 < . . . < xn1 < xn } of [a, b] into n subintervals of
equal length ba
n . Let P : [a, b] R be piecewise linear, linear on [xi , xi+1 ] and such
that P (xi ) = f (xi ) for all 0 i n. Then for all 0 i n, (f, [xi , xi+1 ]) , so
by Step 1 |f (x) P (x)|  for all x [xi , xi+1 ] and thus for all x [a, b].

5.3. A Very Special Case.
P
Lemma 14.23. (Elkies) Let f (x) = n=0 an xn be a power series. We suppose:
(i) The sequence of signs of the coefficients an is eventually constant.
(ii) The radius of convergence is 1.
f (x) = L exists.
(iii) lim
Px1

Then
n=0 an = L, and the convergence of the series to the limit function is
uniform on [0, 1].
Exercise: Prove Lemma 14.23. Two suggestions:
(i) Reduce to the case in which an 0 for all n N.
(ii) Use the Weierstrass M-Test.
Proposition 14.24. For any > 0, the function f (x) = |x| on [, ] can be
uniformly approximated by polynomials.

312

14. SERIAL MISCELLANY

Proof. Step 1: Suppose that for all  > 0, there is a polynomial function
P : [1, 1] R such that |P (x) |x|| <  for all x [1, 1]. Put x = y . Then for
all y [, ] we have
y
y
|P ( ) | || = |Q(y) |y|| < ,

where Q(y) = P ( y ) is a polynomial function of y. Since  > 0 was arbitrary: if

x 7 |x| can be uniformly approximated by polynomials on [1, 1], then it can be
uniformly approximated by polynomials on [, ]. So we are reduced to = 1.
Step 2: Let TN (y) be the degree N Taylor polynomial at zero for the function
f (y) = 1 y. By the Binomial Theorem,
 
X
n

(1 + y) =
y
n
n=0
valid for all R and all y (1, 1). Taking = 21 and substituting y for y, we

see that the degree N Taylor polynomial for 1 y at zero is

1
N
X
(1)n 2 y n ,
TN (y) =
n
n=0

1

and limN TN (y) = 1 y for y [0, 1). Further, (1)n n2 < 0 for n 1, and
p
lim f (y) = lim 1 y = 0.
y1

y1

u
Thus we may apply Elkies Lemma to get TN (y) 1 y on [0, 1]. For x [1, 1],
y = 1 x2 [0, 1], so making this substitution we find that on [1, 1],
1
N
X

u p
2
n 2
TN (1 x ) =
(1)
(1 x2 )n 1 (1 x2 ) = x2 = |x|.
n
n=0


5.4. Proof of the Weierstrass Approximation Theorem.
For a < b R, let C[a, b] be the set of all continuous functions f : [a, b] R,
and let P be the set of all polynomial functions f : [a, b] R. Let PL([a, b]) denote
the set of piecewise linear functions f : [a, b] R.
For a subset S C[a, b], we define the uniform closure of S to be the set S
of all f C[a, b] which are uniform limits of sequences in S: precisely, for which
u
there is a sequence of functions fn : [a, b] R with each fn S and fn f .
Lemma 14.25. For any subset S C[a, b], we have S = S.
Proof. Simply unpacking the notation is at least half of the battle here. Let
u
f S, so that there is a sequence of functions gi S with gi f . Similarly, since
u
each gi S, there is a sequence of continuous functions fij gi . Fix k Z+ :
1
1
choose n such that ||gn f || < 2k
and then j such that ||fnj gn || < 2k
; then
||fnj f || ||fnj gn || + ||gn f || <

1
1
1
+
= .
2k 2k
k

Thus if we put fk = fnj , then for all k Z+ , ||fk f || <

1
k

and thus fk f .

6. A CONTINUOUS, NOWHERE DIFFERENTIABLE FUNCTION

313

Observe that the Piecewise Linear Approximation Theorem is

PL[a, b] = C[a, b],
whereas the Weierstrass Approximation Theorem is
P = C[a, b].
Finally, the point: its enough to show that every piecewise linear function can be
uniformly approximated by polynomial functions, for then P PL[a, b], so
P = P PL[a, b] = C[a, b].
Thus the following result completes the proof of Theorem 14.20.
Proposition 14.26. We have P PL[a, b].
Proof. Let f PL[a, b]. By Lemma 14.21, we may write
n
X
f (x) = b +
|mi x + bi |.
i=1

Choose > 0 such that for all 1 i n, if x [a, b], then mi x + bi [, ]. For
each 1 i n, by Lemma 14.24 there is a polynomial Pi such that
Pn for all x [a, b],
|Pi (mi x+bi )|mi x+bi || < n . Let P : [a, b] R by P (x) = b+ i=1 Pi (mi x+bi ).
Then P P and for all x [a, b],


n
n
X
 X



= .
|P (x) f (x)| = 
(Pi (mi x + bi ) |mi x + bi |) <


n
i=1
i=1

6. A Continuous, Nowhere Differentiable Function
We are going construct a function f : R R with the following striking property:
for all x0 R, f is continuous at x0 but f is not differentiable at x0 . In short, we
say that f is continuous but nowhere differentiable.
The first such construction (accompanied by a complete, correct proof) was given
in a seminal 1872 paper of Weierstrass. Weierstrasss example was as follows: let
(0, 1), and let b be a positive odd integer such that b > 1 + 3
2 . Then the
function f : R R given by

X
(108)
f (x) =
n cos(bn x)
n=0

is continuous on R but not differentiable at any x R.

Exercise: Show that the function defined by (108) above is continuous.
Unfortunately the proof that f is nowhere differentiable is not so easy, as indicated
by the rather specific conditions given on the parameters , b. (For less carefully
chosen , b the function f can have a small set of points of differentiability.)
Thus, as with most other contemporary treatments, we will switch to a different
function for which the nowhere differentiability calculation is more straightforward.
More specifically, we will switch from trigonometric functions to our new friends
the piecewise linear functions, so first we interpose the following exercise nailing

314

14. SERIAL MISCELLANY

down some further (simple) properties of these functions.

Exercise:
a) Let f : [a, b] R be a piecewise linear function with slopes m1 , . . . , mn . Show
that f is Lipschitz, and the smallest possible Lipschitz constant is C = maxi |mi |.
b) Let f : R R be a function. Suppose that there is C > 0 such that for every
closed subinterval [a, b] of R, C is a Lipschitz constant for the restriction of f to
[a, b]. Show that C is a Lipschitz constant for f .
c) Let f : R R be a piecewise linear function with corners at the integers
i.e., f is differentiable on (n, n + 1) for all n Z+ and is not differentiable at
any integer n. For n Z, let mn be the slope of f on the interval (n, n + 1). Let
C = supnZ mn . Show that f is Lipschitz iff C < , in which case C is the smallest
Lipschitz contant for f .
Now we begin our construction with the sawtooth function S : R R: the
unique piecewise linear function with corners at the integers and such that S(n) = 0
for every even integer n and S(n) = 1 for every odd integer n. The slopes of S are
all 1, so by the preceding exercise S is Lipschitz (hence continuous):
x, y R, |S(x) S(y)| |x y|.
Also S is 2-periodic: for all x R, S(x + 2) = S(x). For k N, define
 k
3
S(4k x).
fk : R R, fk (x) =
4
We suggest that the reader sketch the graphs of the functions fk : roughly speaking
they are sawtooth functions which, as k increases, oscillate more and more rapidly

k
but with smaller amplitude: indeed ||fk (x)|| = 43 . We define f : R R by
 k

X
X
3
fk (x) =
f (x) =
S(4k x).
4
k=0

k=0

3 k

Since k=0 ||fk || = k=0 4 < , the series defining f converges uniformly by
the Weierstrass M-Test. This also gives that f is continuous, since f is a uniform
limit of a sequence of continuous functions. We claim however that f is nowhere
differentiable. To see this, fix x0 R. We will define a sequence {n } of nonzero
real numbers such that n 0 and the sequence
Dn =

f (x0 + n ) f (x0 )
n

is divergent. This implies that f is not differentiable at x0 .

Lets do it. First suppose that the fractional part of x0 lies in [0, 21 ), so that
the interval (x0 , x0 + 12 ) contains no integers. In this case we put
n =

4n
,
2

and the reason for our choice is that the interval (4n x0 , 4n (x0 + n )) contains no
integers. Let k, n N. We claim the following inequalities:
(109)

k > n, |S(4k x0 + 4k n ) S(4k x0 )| = 0.

7. THE GAMMA FUNCTION

(110)
(111)

k = n, |S(4k x0 + 4k n ) S(4k x0 )| =

315

1
.
2

]k < n, |S(4k x0 + 4k n ) S(4k x0 )| |4k n |.

kn

Indeed: (109) holds because if k > n, 4k x0 +4k n 4k x0 = 4k n = 4 2 is a multiple

of 2 and S is a 2-periodic function; (110) holds because if k = n, 4k x0 + 4k n =
4k x0 + 12 , so by our choice of n , the function S is linear on [4k x0 , 4k x0 + 12 ] of slope
1, hence the difference between its values at the endpoints is 1
2 . Finally, (111)
holds because 1 is a Lipschitz constant for S. Using these results and the Reverse
Triangle Inequality gives

  n  k


k
k
k
 f (x0 + n ) f (x0 )  X
3
S(4
x
+
4

S(4
x
)

0
n
0
=


 


n
4
n
k=0

 n
n1
X  3 k  S(4k x0 + 4k n ) S(4k x0 ) 
3
n


4


4
4
n
k=0

3n

n1
X
k=0

3k = 3n

3n 1
3n

.
2
2

Thus Dn , so f is not differentiable at x0 .

Were not quite done: recall that we assumed that the fractional part of x0 lay
in [0, 12 ), with the consequence that S was linear on the interval [(4n x0 , 4n (x0 +n )].
n
What to do if the fractional part of x0 lies in [ 12 , 1)? In this case we take n = 42
so that the interval (4n (x0 + n ), 4n x0 ) contains no integers so S is linear on the
interval [4n (x0 + n ), 4n x0 ], and the rest of the proof goes through as above.
So, albeit with a different function, we have proved Weierstrasss Theorem.
Theorem 14.27. (Weierstrass, 1872) There is a function f : R R which is
continuous at every point of R but differentiable at no point of R.
Notice that if we restrict f to some closed interval, say [0, 2], then by the Weierstrass
Approximation Theorem f is like any continuous function on [0, 2] a uniform
limit of polynomials. Thus even a uniform limit of polynomials on a closed, bounded
interval need not have any good differentiability properties whatsoever!
7. The Gamma Function
There are many real functions. Some are elementary: i.e., built up from power
functions, trigonometric functions, exponentials
and logarithms. Most are not, like
Rx
2
the Gaussian error function E(x) = et dt. As mathematical analysis developed from Newton and Leibniz in the 17th century on down, eventually it became
clear that certain non-elementary functions arise again and again in a variety of
contexts, and satisfy some remarkable, beautiful identities.
Such functions are called special functions. There is no precise mathematical
definition of a special function. Rather, the appellation is historical and cultural.
Unfortunately the days when the majority of students and practitioners of mathematics naturally learned about special functions in the context of their work are
past, since indeed it is no longer the case that that the majority of students and
practitioners of mathematics are deeply concerned with real function theory. In
our day your smart friend the physicist or the engineer probably knows more about

316

14. SERIAL MISCELLANY

special functions than you do...and this may not be a good thing.
In this section we will study one of the most ubiquitous special functions of
them all, the Gamma function.
7.1. Definition and Basic Properties.
For x (0, ) we define
Z

(x) =

tx1 et dt.

As this is an improper integral, thre is of course something to check.

Proposition 14.28. For x (0, ), the integral defining (x) is convergent.
Exercise: Prove Proposition 14.28. Some suggestions:
(i) First deal with the case x 1, since in this case the function t 7 tx1 is continuous on [0, ) and the integral is improper only at . To show the convergence
use the fact that exponentials grow faster than polynomials.
(ii) When 0 < x < 1, the integral is also improper at 0, so it should be split into
R1
R
two pieces, say 0 and 1 . The argument of part a) will handle the latter integral.
R1
Reduce the former integral to 0 xdxp .
R
Exercise: a) Show that the improper integral 0 tx1 et dt is divergent for x 0.
b) Show that (x) > 0 for all x (0, ).
b) Show that limx0+ (x) limx (x) = .
c) Show that (1) = (2) = 1.
Exercise: a) By making the change of variables t = s2 , show that for all x (0, ),
Z
2
(112)
(x) = 2
s2x1 es ds.
0

b) Deduce that
(113)

1
( ) = 2
2

x2

e
0

dx =

ex dx.

Theorem 14.29. a) For all x (0, ) we have

(114)

(x + 1) = x(x).

b) For all n N, (n + 1) = n!.

Exercise: Prove Theorem 14.29. Suggestions:
For part a), integrate by parts, much as in XXXX.
For part b), use part a), the previous exercise, and induction.
Thus the Gamma function is a continuous interpolation of the factorial function
(with a slight shift in the argument that everyone finds a little distressing at first
but eventually learns to live with). Of course there are infinitely many continuous
(even infinitely differentiable) functions which interpolate any real sequence. Is
there a sense in which the Gamma function is the right interpolation?
Yes, although it is rather curious.

7. THE GAMMA FUNCTION

317

Theorem 14.30. (Bohr-Mollerup [BM22])

a) The function log : (0, ) R is convex.
b) Conversely, let f : (0, ) (0, ) satisfy:
(i) f (x + 1) = xf (x) for all x (0, );
(ii) f (1) = 1; and
(iii) log f is convex.
Then f = .
Proof. a) Let f = log . We will verify the secant-graph inequality: for all
0 < x < y < and (0, 1),
f ((1 )x + y) (1 )f (x) + f (y),
and for this we will use H
olders Integral Inequality with p = 1 and q =
Z
f ((1 )a + b) = log
tx+(1)y1 et dt

1
1 :

Z

x1 t

= log

(t

y1 t 1

e ) (t

e )


dt

Z
log

 Z
e dt

x1 t

1 !
e dt

y1 t

= log (x) + (1 ) log (y).

b) . . .

So it seems interesting to investigate non-integer values of the Gamma function.

In some sense the first order of business is to evaluate ( 12 ). Then using (114) we
can evaluate ( n2 ) for all integers n. Moreover, by (113) we know that ( 21 ) =
R x2
e
dx, an improper integral that is easily seen to be convergent but whose

precise value is not so easy to determine.

To evaluate ( 12 ) we introduce a second special function and relate it back to the
Gamma function using the Bohr-Mollerup Theorem.
7.2. The Beta Function.
For x, y (0, ), we define the Beta function
Z 1
B(x, y) =
tx1 (1 t)y1 dt.
0

When x, y 1, this is a proper integral, and there are no convergence issues.

However, when one of x, y is less than 1, there is something to check.
Exercise: Show that the integral defining B(x, y) is convergent for all x, y > 0.
Theorem 14.31. For all x, y > 0, we have
B(x, y) =
Proof. . . .

(x)(y)
.
(x + y)


318

14. SERIAL MISCELLANY

Making the substitution t = sin2 in the integral defining B(x, y) and applying
Theorem 14.31 we get
Z 2
(x)(y)
2x1
2y1
(115)
2
(sin )
(cos )
d =
.
(x + y)
0
Taking x = y =

1
2

in (115) we deduce:

Theorem 14.32.
(116)

  Z

2
1
=
ex dx = .
2

7.3. Interlude: A Dominated Convergence Theorem.

Theorem 14.33. Let fn , f, g : (0, ) R. We suppose:
(i) fn and g are Riemann integrable on every closed, bounded subinterval of (0, );
(ii) fn f pointwise, and the convergence is uniform on each closed, bounded
subinterval of [0, ).
(iii) R|fn (x)| g(x) for all x (0, ); and

(iv) 0 g(x)dx < .

R
R
Then limn 0 fn (x)dx = 0 f (x)dx.
Proof. Since |fn (x)| g(x) for all x and
R fn f , also |f (x)| g(x) for all x.
Fix  > 0. Since g is non-negative and 0 g(x)dx < , there is N Z+ with
Z
Z N1
g(x)dx +
g(x)dx < ,
0

Thus for all n Z+ ,

Z 1
 Z
 Z 1
Z
 N
 
N



fn (x)dx
fn (x)dx + 
|fn (x)|dx +
|fn (x)|dx < ,

 0

N
0
N
and similarly
Z 1
 Z
 Z 1
Z
 N
 
N


 

f (x)dx + 
f (x)dx
|f (x)|dx +
|f (x)|dx < .

 0

N
0
N
Since fn f uniformly on [ N1 , N ],
Z
Z N
lim
fn (x)dx =
n

1
N

1
N

Z
|fn (x)|dx +

+
0

f (x)dx.
1
N

It follows that for all sufficiently large n,

Z
Z


f
(x)dx

n

!

1
N

+
0



f (x)dx

Z

 N



|f (x)|dx + 
(fn (x) f (x))dx
 1

N

 +  +  = 3.
Since  was arbitrary, the proof is complete.

7. THE GAMMA FUNCTION

319

Remark: Theorem 14.33 is given as an exercise in W. Rudins text [R, p. 167].

It is a weak form of what is perhaps the most important and useful single result
of graduate level real variable theory, the Lebesgue Dominated Convergence
Theorem. In Lebesgues version, the integrals are taken in his more permissive
sense (which we certainly have not discussed). More significantly, the hypothesis of
uniform convergence on closed bounded subintervals is dropped entirely: all that is
needed is that fn f pointwise.
Since uniform convergence on bounded subintervals is a much stronger hypothesis than pointwise convergence, Rudins version of the Dominated Convergence
Theorem is significantly weaker than Lebesgues. (For a version of the Dominated
Convergence Theorem which uses only pointwise convergence but stays in the context of the Riemann integral, see [Ke70]. But I must warn you that even this paper
requires somewhat more background than we have developed here.) Nevertheless
Theorem 14.33 has some useful applications.
7.4. Stirlings Formula.
Theorem 14.34.
(117)
Proof. . . .

lim

(x + 1)

= 1.
(x/e)x 2x


CHAPTER 15

Several Real Variables and Complex Numbers

1. A Crash Course in the Honors Calculus of Several Variables
Let M, N Z+ . In multivariable calculus one studies functions f : RN RM ,
that is vector valued functions of several variables. Here we will be briefly interested in several aspects of the case M = 1, i.e., of functions f : RN R and also
of sequences and series with values in RN .
First, recall that by RN we mean the set of all ordered N -tuples x = (x1 , . . . , xN )
of real numbers. To define things like continuity and convergence, need a way to
measure the distance between x, y RN . For this, we define
q
|x| = |(x1 , . . . , xN )| = x21 + . . . + x2N .
Notice that, at least when 1 N 3, this is the distance from the origin 0 =
(0, . . . , 0) to the vector x that we learn about in high school geometry. The basic
fact that allows estimates to be made is the triangle inequality:
x, y RN , |x + y| |x| + |y|.
We proved this at the end of Chapter 8 as an application of Jensens Inequality.
N
Here are the two key definitions: a sequence {xn }
is given by a function
n=1 in R
n 7 xn from Z+ to RN . We say that the sequence xn converges to x RN if for
all  > 0, there exists N0 Z+ such that for all n N0 , |xn x| < . A sequence
is convergent if it converges to some element of RN and otherwise divergent.

Let D RN , and let x D. A function f : D R is continuous at x if for all

 > 0, there exists > 0 such that for all y D with |x y| < , |f (x) f (y)| < .
A function f : D R is continuous if it is continuous at every point of D.
Exercise 15.1: Let x = (x1 , . . . , xN ), y = (y1 , . . . , yN ) RN . Show that for all
1 i N , |xi yi | |x y|.
Example 15.2: For 1 i N , let i : RN R be the function which maps
(x1 , . . . , xN ) to xi . Then i is continuous: indeed, for any  > 0, we may take
= , since then if |x y| < , |i (x) i (y)| = |xi yi | |x y| < .
Exercise 15.3: Fix x0 RN , and consider the function dx0 : RN R, dx0 (x) =
|x x0 |. Show that dx0 is continuous.
Proposition 15.1. Let f1 , f2 : D RN R be two functions, and let ,
R.
321

322

15. SEVERAL REAL VARIABLES AND COMPLEX NUMBERS

a) If f1 and f2 are continuous, so is f1 + f2 .

b) If f1 and f2 are continuous, so is f1 f2 .
c) If f1 and f2 are continuous and for all x D, f2 (x) 6= 0, then

f1
f2

is continuous.

Exercise 15.4: Prove Proposition 15.1. (Suggestion: adapt the N = 1 case.)

Exercise 15.5: a) Give a reasonably careful definition of a polynomial function
f : RN R. (Hint: the functions i defined above are relevant.)
b) Show that every polynomial function is continuous.
Proposition 15.2. Let {xn } be a sequence in RN , and let x RN . TFAE:
(i) xn x.
(ii) For all 1 i n, i (xn ) i (x).
Proof. (i) = (ii): Fix  > 0. By assumption, there is N0 Z+ such
that for all n N0 , |xn x| < . Hence by Exercise 15.1, for all 1 i N ,
|i (xn ) i (x)| < .
(ii) = (i): Fix  > 0. For each 1 i N , there exists Ni Z+ such that
for all n pNi , |i (xn ) i (x)| < . Take N0 =max(N1, . . . , NN ); for n N0 ,
|xn x| = (xn,1 x1 )2 + . . . + (xn,N xN )2 < N 2 = N . Good enough. 
Proposition 15.3. Let f : D RN R be a continuous function, and let
{xn } be a sequence in D. if xn x, then f (xn ) f (x).
Exercise 15.6: Prove Proposition 15.3.
We now press on to give analogues of the fundamental Interval Theorems of
honors calculus. Our first order of business is to find a suitable analogue in RN
of the notion of a closed, bounded interval in R. It turns out that we have some
leeway here. Let us consider two different kinds of sets.
For 1 i N , choose real numbers ai bi . We define the closed coordinate box [a, b] to be the set of all x = (x1 , . . . , xN ) RN such that ai xi bi
for all 1 i N . Further, for x RN and r 0, we define the open ball Br (x)
to be the set of all y RN such that |x y| < r and the closed ball B r (x) to be
the set of all y RN such that |x y| r.
Exercise 15.7: For a subset D of RN , show that the following are equivalent:
(i) There is some closed coordinate box [a, b] with D [a, b].
(ii) There are x RN and r 0 with D Br (x).
(iii) There is r 0 with D B0 (x).
A subset D satisfying these equivalent properties is bounded.
Exercise 15.8: Let D RN be a bounded subset. For all 1 i N , show
thaqt i (D) is a bounded subset of R.
Exercise 15.9: Show that all closed coordinate boxes and all open and closed balls
are convex subsets of RN .
Theorem 15.4. Let D RN be a nonempty convex subset, and let f : D R
be a continuous function. Then f (D) = {f (x) | x D} is an interval in R.

2. COMPLEX NUMBERS AND COMPLEX SERIES

323

Proof. By Theorem 8.1, it is enough to show f (D) is convex, i.e., if y1 , y2

f (D), then the entire interval from y1 to y2 is in f (D). Suppose y1 = f (x1 ) and
y2 = f (x2 ). Define a function g : [0, 1] R by g : t 7 f ((1 t)x1 + tx2 ). The
function g is well-defined since D is convex; it is continuous since f is continuous,
and g([0, 1]) f (D). By the usual Intermediate Value Theorem, g([0, 1]) is an
interval in R. Since g(0) = f (x1 ) = y1 and g(1) = f (x2 ) = y2 , this implies that
every point between y1 and y2 lies in f (D).

Theorem 15.5. (Multivariate Bolzano-Weierstrass Theorem) Every bounded
sequence in RN admits a convergent subsequence.
Proof. When N = 1 this is the usual Bolzano-Weierstrass Theorem. We will
reduce to this case by repeated passage to subsequences. Let {xn } be a bounded
sequence in RN . By Exercise 15.8, 1 (xn ) is a bounded sequence of real numbers.
By the usual Bolzano-Weierstrass Theorem, there is a subsequence nk such that
1 (xnk ) converges. Since {xn } is bounded, so is its subequence xnk , and hence
so is 2 (xnk ). Applying Bolzano-Weierstrass again, we get a subsubsequence xnkl
such that 2 (xnkl ) converges. Continuing in this manner i.e., passing again to a
subsequence for each coordinate in turn we eventually get a subsub...subsequence
(N subs in all!) all of whose coordinates converge. This is a subsequence of the
original sequence which converges by Proposition 15.2.

A subset D RN is closed if for all sequences {xn } with values in D, if xn x
RN , then x D.
Exercise 15.10: Show that a closed interval I R is a closed subset of R.
Proposition 15.6. Closed coordinate boxes [a, b] and closed balls B r (x) are
closed subsets of RN .
QN
Proof. Let D = [a, b] = i=1 [ai , bi ] be a closed coordinate box, and let xn be
a sequence in D converging to x RN . Then for 1 i N , i (xn ) i (x). Since
i (xn ) [ai , bi ], i (x) [ai , bi ]. It follows that x [a, b] and thus [a, b] is closed.
Let D = B r (x0 ) be a closed ball and let xn be a sequence in D converging to
x RN . By Exercise 15.3, y 7 |x0 y| is continuous, and thus |x0 xn | |x0 x|.
Since |x0 xn | r for all n, |x0 x| r, i.e., x B r (x0 ): B r (x0 ) is closed.

Theorem 15.7. (Multivariable Extreme Value and Uniform Continuity) Let
D RN be closed and bounded.
a) Every sequence in D admits a subsequence converging to an element of D.
b) Every continuous function f : D R is bounded and attains its maximum and
minimum values.
c) Eery continuous function f : D R is uniformly continuous.
Proof. a) Since D is bounded, by Theorem 15.5, every sequence in D admits
a subsequence converging to x RN . Since D is closed, x must lie in D.
b),c) The proofs of Theorems 10.20 and 10.21 which treat the case D = [a, b] adapt
easily to the general case. Details are left to the reader.

2. Complex Numbers and Complex Series
It turns out that the sine, cosine and exponential functions are all very closely
related to each other, provided we are willing to work with complex values. In

324

15. SEVERAL REAL VARIABLES AND COMPLEX NUMBERS

this section we provide a review of complex numbers and the rudiments of complex
power series. This theory can be developed to an amazing extent further than the
theory of real power series, in fact! but such qualitatively different developments
are the subject of another course. Here we just want to develop the theory enough
so that we can make sense of plugging complex numbers into power series.
Recall that a complex number is an expression of the form z = a + bi. Here a
and b are real numbers and i is a formal symbol having the property that i2 = 1.
For many years people had philosophical difficulties with complex numbers; indeed, numbers of the form ib were called imaginary, and the prevailing view was
that although they did not exist, they were nevertheless very useful.
From a modern point of view this is neither acceptable (we cannot work with
things that dont exist, no matter how useful they may be!) nor necessary: we can
define the complex numbers entirely in terms of the real numbers. Namely, we may
identify a complex number a + bi with the ordered pair (a, b) of real numbers, and
we will define addition and multiplication. Since we would want (a + bi) + (c + di) =
(a + c) + (b + d)i, in terms of ordered pairs this is just (a, b) + (c, d) = (a + c, b + d).
In other words, this is the usual addition of vectors in the plane. The multiplication
operation is more interesting but still easy enough to write down in terms of only
real numbers: to compute (a + bi)(c + di), we would want to use the distributive
law of multiplication over addition and the relation i2 = 1. In other words, we
would like (a + bi) (c + di) = ac + bci + adi + bdi2 = (ac bd) + (ad + bc)i. Thus
in terms of ordered pairs we define a multiplication operation
(a, b) (c, d) = (ac bd, ad + bc).
Note that with this convention, we may identify real numbers a (i.e., those with
b = 0) with pairs of the form (a, 0); moreover, what we were calling i corresponds
to (0, 1), and now any ordered pair a + bi can be expressed as (a, 0) + (b, 0) (0, 1).
Exercise: Show that the above operations of addition and multiplication on ordered pairs satisfy all the field axioms (P0) through (P9). The resulting structure
is called the complex numbers and denoted C.
Exercise: Show that because of the relation i2 = 1, C cannot be endowed with
the structure of an ordered field.
Two other important operations on the complex numbers are conjugation and taking the modulus. For any complex number z = a + bi, we define its complex
conjugate to be z = a bi. Conjugation fits in nicely with the rest of the algebraic
structure: one has z1 + z2 = z1 + z2 and (z1 z2 ) = z1 z2 .
For any complex number
z = a + bi, we define its modulus (or norm, or abso
lute value) to be |z| = a2 + b2 . This is just the usual norm of an element of RN
specialized to the case N = 2. In particular, we have the triangle inequality
z1 , z2 C, |z1 + z2 | |z1 | + |z2 |.
However, the norm also behaves nicely with respect to the multiplicative structure.

3. ELEMENTARY FUNCTIONS OVER THE COMPLEX NUMBERS

325

Proposition 15.8. a) For all z1 , z2 C, |z1 z2 | = |z1 ||z2 |.

b) For all z C, zz = |z|2 .
Exercise: Prove Proposition 15.8.
Finally, we mention that the theory of complex series, and especially, of complex power series, works (at least) as well as the theory of real series. Namely, if
P
1
n
n
n=0 an z is a power series with complex coefficients, then defining = lim sup |an | ,
we find that is the radius of convergence of the complex power series in the sense
that the seriesPconverges for all z with |z| < and diverges for all z with |z| > .
Especially, if n an xn is a power series with real coefficients and infinite radius of
convergence, then because for a real number x, its absolute value |x|
P is the same
as the modulus of the complex number x + 0i, then the power seres n an z n must
converge for all complex numbers z.
3. Elementary Functions Over the Complex Numbers
3.1. The complex exponential function.
Consider the following complex power series:

X
zn
.
E(z) :=
n!
n=0
1

1
Because the ratio test limit is limn (n+1)!
= limn n+1
= 0, the radius of
1
n!
convergence is infinite: the series converges for all complex numbers z.

Proposition 15.9. For all complex numbers z and w, E(z + w) = E(z)E(w).

Proof. Since the series representations of E(z) and E(w) are absolutely convergent, we know that E(z)E(w) is given by the Cauchy product, namely

n  

X
n
X
X
1 X n k nk X (z + w)n
z k wnk
=
z w
=
= E(z+w).
E(z)E(w) =
k!(n k)! n=0 n!
k
n!
n=0
n=0
k=0

k=0


Since E(0) = 1, we have for all z that
E(z)E(z) = E(z z) = E(0) = 1,
1
E(z) .

Note in particular that E(z) is never zero. Restricting attention

or E(z) =
to real values, since E : x 7 E(x) is a continuous function which is never zero and
such that E(0) = 1, we conclude E(x) > 0 for all real x.
3.2. The trigonometric functions.
Let us now turn to the functions sin x and cos x. Recall that we have already shown
that any pair of differentiable functions S(x) and C(x) such that S 0 (x) = C(x),
C 0 (x) = S(x), S(0) = 0 and C(0) = 1 must be equal to their Taylor series and
given by the following expansions:

X
(1)n x2n+1
.
S(x) :=
(2n + 1)!
n=0

326

15. SEVERAL REAL VARIABLES AND COMPLEX NUMBERS

C(x) :=

X
(1)n x2n
.
(2n)!
n=0

Of course we would like to say S(x) = sin x and C(x) = cos x, but we do not
want to have to resort to discussions involving angles, lengths of arcs and other
such things. We want to see how much can be derived directly from the power series expansions themselves. For instance, we would like the show that C 2 + S 2 = 1.
Unfortunately, although this identity does hold, showing it directly from the power
series expansions involves some rather unpleasant algebra (try it and see).
This is where complex numbers come in to save the day:
Proposition 15.10. For all real x, we have the following identities:
1
C(x) = (E(ix) + E(ix)),
2
1
S(x) = (E(ix) E(ix)),
2i
E(ix) = C(x) + iS(x).
Exercise: Prove Proposition 15.10.
Now we are in business: since the coefficients of E(z) are real, we have E(ix) =
E(ix) = E(ix) for all real x, hence
C(x)2 + S(x)2 = |E(ix)|2 = E(ix)E(ix) = E(ix)E(ix) = E(ix ix) = E(0) = 1.
Were not done yet: wed like to prove that S(x) and C(x) are periodic functions,
whose period is a mysterious number approximately equal to 2 3.141592653 . . ..
This can also be worked out from the power series expansions, with some cleverness:
We first claim that there exists x0 > 0 such that C(x0 ) = 0. Otherwise, since
C(0) = 1 > 0, wed have C(x) > 0 for all x, hence S 0 (x) = C(x) > 0 for all x, hence
S would be strictly increasing on the entire real line. Since S(0) = 0, it follows that
S(x) > 0 for all x > 0. Now, if 0 < x < y, we have
Z y
S(x)(y x) <
S(t)dt = C(x) C(y) 2.
But now for fixed x and y > x +

x
2
S(x) ,

this gives a contradiction.

Lemma 15.11. Let f : [0, ) R be continuous such that f (0) > 0 and
f (x) = 0 for some x > 0. Then there is a least positive number x0 such that
f (x0 ) = 0.
Proof. Left to the reader.

Now we define the number by = 2x0 . where x0 is the lesat positive number x
such that C(x) = 0. The relation C(x)2 + S(x)2 = 1 together with C( 2 ) = 0 shows
that S( 2 ) = 1. On the other hand, since C(x) = S 0 (x) is non-negative on [0, 2 ],
S(x) is increasing on this interval, so it must be that S( 2 ) = 1. Thus E( i
2 ) = i.
Using the addition formula for E(z) we recover Eulers amazing identity
 i 2
ei = e 2
= 1,

4. THE FUNDAMENTAL THEOREM OF ALGEBRA

327

and also e2i = 1. In general, ez+2i = ez e2i = ez , so E is periodic with period 2i.
Using the periodicity of E and the formula of Proposition 2, we get that for all
x C(x + 2) = C(x) and S(x + 2) = S(x).
Since for all real t, |eit | = 1, the parameterized curve
r(t) = eit = C(t) + iS(t) (C(t), S(t)) = (x(t), y(t))
has image contained in the unit circle. We claim that every point on the unit circle
is of the form eit for a unique t [0, 2). To see this, start at the point 1 = ei0 ,
and consider t [0, 2 ]. The function C : [0, 2 ] R is continuous and decreasing,
hence injective, with C(0) = 1 and C( 2 ) = 0. By the Intermediate Value Theorem, all values in [0, 1] are assumed for a (necessarily unique, by the injectivity)
t [0, pi
2 ], and every point in the first quadrant of the unit circle is of the form
(x, y) for a unique x [0, 1]. By making similar arguments in the intervals [ 2 , ],
3
[, 3
2 ], [ 2 , 2] we establish the claim.
Finally, if we grant that by the arclength of the parameterized curve r(t) = (x(t), y(t))
from t = a to t = b we mean the integral
Z b r
dx
dy
| |2 + | |2 dt
dt
dt
t=a
it is easy to show that C(x) = cos x and S(x) = sin x.
(C(t), S(t)), the arclength integral is
Z
S 2 (t) + C 2 (t)dt = ,

Indeed, for r(t) =

t=0

so the point r() = (C(), S()) really is the point that we arrive at by starting at
the point (1, 0) on the unit circle and traversing units of arc.
S(x)
Exercise: Show that for all x (0, /2), S(x) < x < C(x)
.
(Show this directly from the power series definitions: no pictures involving the unit
circle! Hint: use the power series representation for arctan x.)

Lemma 15.12. (DeMoivre) Let k Z+ and z C. Then there is w C such

that wk = z.
Proof. If z = 0 we may take w = 0. Otherwise, r := |z| > 0, so zr lies on
the unit circle and thus zr = ei for a unique [0, 2). We well know (as a
consequence of the Intermediate Value Theorem) that every positive real number

1
1
has a positive kth root, denoted r k . Thus if w = r k ei / k ,
1

wk = (r k ei / k )k = r(ei k )k = rei = z.

Exercise: Let z be a nonzero complex number and k a positive integer. Show that
there are preciesly k complex numbers w such that wk = z.
4. The Fundamental Theorem of Algebra
4.1. The Statement and Some Consequences.

328

15. SEVERAL REAL VARIABLES AND COMPLEX NUMBERS

Theorem 15.13. (Fundamental Theorem of Algebra) Let

P (z) = an z n + . . . + a1 z + z0
be a polynomial with complex coefficients and positive degree. Then P has a root in
the complex numbers: there is z0 C such that P (z0 ) = 0.
Theorem 15.13 is not easy to prove, and we defer the proof until the next section.
For now we give some important consequences of this seminal result.
Corollary 15.14. Every nonconstant polynomial with complex coefficients
factors as a product of linear polynomials. More precisely, let
P (z) = an z n + . . . + a1 z + a0 , with an 6= 0.
Then there are 1 , . . . , n C (not necessarily distinct) such that
(118)

P (z) = an (z 1 )(z 2 ) (z n ).

Proof. First observe that if the result holds for P (z) then it holds for P (z)
for any C \ {0}. It is therefore no loss of generality to assume that the leading
coefficient an of P (z) is equal to 1. Let us do so.
We now prove the result by induction on n, the degree of the polynomial P .
Base Case (n = 1): A degree one polynomial with leading coefficient 1 is of the
form z + a0 = z 1 , with 1 = a0 .
Induction Step: Let n Z+ , suppose the result holds for all polynomials of degree
n, and let P (z) be a polynomial of degree n + 1. By Theorem 15.13, there is
z0 C such that P (z0 ) = 0. By the Root-Factor Theorem, we may write P (z) =
(z z0 )Q(z), with Q(z) a polynomial of degree n and leading coefficient 1. By
induction, Q(z) = (z 1 ) (z n ), so putting n+1 = z0 we get
P (z) = (z 1 ) (z n )(z n+1 ).

More generally, let F be a field: that is, a set endowed with two binary operations
denoted + and and satisfying the field axioms (P0) through (P9) from Chapter
1. We say that F is algebraically closed if every nonconsant polynomial P (x)
with coefficients in F has a root in F , i.e., there is x0 F such that P (x0 ) = 0. In
this terminology, the Fundamental Theorem of Algebra asserts precisely that the
complex field C is algebraically closed.
Exercise: Let F be an algebraically closed field. Show that the conclusion of Corollary 15.14 holds for F : that is, every nonconstant polynomial factors as a product
of linear polynomials.
Since every real number is, in particular, a complex number, Corollary 15.14 applies
in particular to polynomials over R: if P (x) = an xn + . . . + a1 x + a0 with an 6= 0,
then there are complex numbers 1 , . . . , n such that
P (x) = an (x 1 )(x 2 ) (x n ).
But since the coefficients of P are real, it is natural to ask whether or to what
extent some or all of the roots i must also be real. Recall that we need not have
any real roots (that is, the field R is not algebraically closed), for any n Z+ , the
polynomial Pn (x) = (x2 + 1)n is positive for all real x, so has no real roots. And
indeed, its factorization over C is (x2 + 1)n = (x + i)n (x i)n .
However, the polynomials Pn all had even degree. Recall that, as a consequence

4. THE FUNDAMENTAL THEOREM OF ALGEBRA

329

of the Intermediate Value Theorem, every polynomial of odd degree has at least
one real root (and need not have more than one, as the family of examples xPn (x)
shows). So there is some relation between the parity (i.e., the evenness or oddness)
of the degree of a real polynomial and its real and complex roots. This observation
can be clarified and sharpened in terms of the operation z = a + bi 7 a bi = z of
complex conjugation.
Lemma 15.15. a) For z, a0 , . . . , an C, we have
an z n + . . . + a1 z + a0 = an z n + . . . + a1 z + a0 .
b) Thus, if a0 , . . . , an R, we have
an z n + . . . + a1 z + a0 = an z n + . . . + a1 z + a0 .
c) Let a0 , . . . , an R and put P (z) = an z n + . . . + a1 z + a0 . If z0 C is such that
P (z0 ) = 0, then also P (z0 ) = 0.
Proof. We have already observed that for any z1 , z2 C, z1 + z2 = z1 +z2 and
z1 z2 = z1 z2 . Keeping these identities in mind, the proof becomes a straightforward
exercise which we leave to the reader.

It is part c) of Lemma 15.15 that is important for us: this well-known result often
goes by the description The complex roots of a real polynomial occur in conjugate
pairs. To see why this is relevant, consider the following extremely simple but
extremely important result.
Lemma 15.16. For a complex number z, z R z = z.
Proof. If z R, then z = z + 0i, so z = z 0i = z. Conversely, let z = a + bi.
IF a bi = z = z = a + bi, then (2i)b = 0. Multiplying through by (2i)1 = i
2 ,
we get b = 0, so z = a + 0b R.

Let us say that C is properly complex if it is not a real number. If P (x)
is a polynomial with real coefficients and has the properly complex number as a
root, then by Lemma 15.16 it also has the (distinct!) properly complex number
as a root. By the Root-Factor Theorem, we may write P (z) = (z )(z )P2 (z).
Now wake up! something interesting is about to happen. Namely, if we write
P1 (z) = (z )(z ),
then we claim both P1 (z) and P2 (z) have real coefficients, so we have obtained a
factorization of real polynomials
P (x) = P1 (x)P2 (x).
For P1 (z) we need only write = a + bi and multiply it out:
(z )(z ) = z 2 ( + )z + = z 2 (2a)z + (a2 + b2 ).
For P2 (z) we have to argue a bit more abstractly. Namely, for polynomials over R
(and really, over any field) we can always perform division with remainder: there
are unique real polynomials Q(x), R(x) such that
(119)

P (x) = Q(x)P1 (x) + R(x), deg R(x) < deg P1 (x).

We claim R(x) 0 (i.e., it is the zero polynomial). To see this, we use the uniqueness part of the division algorithm in a slightly sneaky way: namely, consider P (x)

330

15. SEVERAL REAL VARIABLES AND COMPLEX NUMBERS

and P1 (x) as polynomials with complex coefficients and perform the division algorithm there: there are unique complex polynomials, say QC (x) and RC (x), such
that
(120)

P (x) = QC (x)P1 (x) + RC (x), deg RC < deg QC .

Heres the point: on the one hand, real polynomials are complex polynomials, so
comparing (119) and (120) we deduce Q(x) = QC (x) and R(x) = RC (x). On
the other hand, the identity P (x) = P1 (x)P2 (x) of complex polynomials shows
that we may take QC (x) = P2 (x) and RC (x) 0. Putting these together, we get
Q(x) = P2 (x) and R(x) 0, so indeed P (x) = P1 (x)P2 (x) is a factorization of real
polynomials.
A positive degree polynomial P (x) over a field F is called irrreducible if it cannot
be factored as P (x) = P1 (x)P2 (x) with deg P1 , deg P2 < deg P . (This last condition
is there to prevent trivial factorizations like x2 + 1 = (2) ( 21 x2 + 12 .) A polynomial
of positive degree which is not irreducible is called reducible.
Exercise: Let F be a field, and let P (x) be a polynomial with coefficients in F .
a) Show that if P is irreducible, it has no roots in F .
b) Suppose P has degree 2 and no roots in F . Show that P is irreducible.
c) Suppose P has degree 3 and no roots in F . Show that P is irreducible.
d) For each n 4, exhibit a degree n polynomial with coefficients in R which is
reducible, but has no real roots.
Theorem 15.17. Let P (x) be a real polynomial of degree n 1.
a) There are natural numbers r, s N such that r + 2s = n, linear polynomials
L1 (x), . . . , Lr (x) and irreducible quadratic polynomials Q1 (x), . . . , Qs (x) such that
P (x) = L1 (x) Lr (x)Q1 (x) Qs (x).
b) If n is odd, then P has at least one real root.
Proof. Its harmless to assume that P has leading coefficient 1, and we do so.
a) We go by strong induction on n. When n = 1, P (x) is a linear polynomial, so
we may take r = n = 1, s = 0 and P (x) = L(x). Suppose n 2 and the result
holds for all polynomials of degree less than n, and let P (x) be a real polynomial
of degree n. By the Fundamental Theorem of Algebra, P (x) has a complex root .
If R, then by the Root-Factor Theorem P (x) = (x )P2 (x) and we are done
by induction. If is properly complex, Q(x) = (x )(x ) is a real, irreducible
quadratic polynomial and P (x) = Q(x)P2 (x), and again we are done by induction.
b) Since r + 2s = n, if n is odd we cannot have r = 0. Thus P has at least one
linear factor and thus at least one real root.

The partial fractions decomposition rests on the foundation of Theorem 15.17.
Exercise:1 Let f (z) = an z n + . . . + a1 z + a0 be a complex polynomial with positive
degree. Suppose that for all z C, we have z R f (z) R.
a) Show: a0 , . . . , an R.
b) Show: n = 1.
1Taken from http://math.stackexchange.com/questions/1332641/

4. THE FUNDAMENTAL THEOREM OF ALGEBRA

331

4.2. Proof of the Fundamental Theorem of Algebra.

We now give a proof of Theorem 15.13, closely following W. Rudin [R, Thm. 8.8].
Let P (z) be a polynomial with complex coefficients, degree n 1, and leading
coefficient an . We want to show that P (z) has a complex root; certainly this holds
iff a1n P (z) has a complex root, so it is no loss of generality to assume that the
leading coefficient is 1 and thus
P (z) = 1 + an1 z n1 + . . . + a1 z + a0 , ai C.
Let
= inf |P (z)|.
zC

Thus is a non-negative real number and our job is to show (i) that is actually
attained as a value of M and (ii) = 0.
Step 1: Since for z 6= 0,


|a0 |
|an1 |
+ ... + n ,
|P (z)| = |z|n 1 +
|z|
|z|
it follows that
lim |P (z)| = 1 = .
z

Thus there is R > 0 such that for all z C with |z| > R, |P (z)| > |P (0)|. By
Theorem 15.7, the continuous function |P (z)| assumes a minimum value on the
closed, bounded set {z C | |z| R}, say at z0 . But R was chosen so that
|P (z)| > |P (0)| |P (z0 )| for all z with |z| > R, so altogether |P (z)| |P (z0 )| for
all z C and thus = inf zC |P (z)| = |P (z0 )|.
Step 2: Seeking a contradiction, we suppose > 0. Define Q : C C by
0)
Q(z) = PP(z+z
(z0 ) . Thus Q is also a degree n polynomial function, Q(z) = 1, and by
minimality of z0 , |Q(z)| 1 for all z C. We may write Q(z) = 1+bk z k +. . .+bn z n
with bk 6= 0 for some 1 k n. Let w C be such that
|bk |
wk =
;
bk
the existence of such a w is guaranteed by Lemma 15.12. Then for r (0, ),
Q(rw) = 1 + bk rk wk + bk+1 rk+1 wk+1 + . . . + bn rn wn


= 1 rk |bk | rwk+1 bk+1 . . . rnk wn bn = 1 rk (|bk | + C(r)) ,
where we have set C(r) = rwk+1 bk+1 . . . rnk wb bn . Thus
|Q(rw)| = |1 rk (|bk | + C(r))| |1 rk |bk || + |rk C(r)|.
As r approaches 0 from the right, rk |bk | and C(r) both approach 0. Thus for
sufficiently small r, we have rk |bk | < 1 and |C(r)| < |bk | and then
|Q(rw)| |1 rk |bk || + |rk C(r)| = 1 rk (|bk | |C(r)|) < 1.
This contradicts the fact that minzC |Q(z)| = 1 and completes the proof.

CHAPTER 16

Foundations Revisited
The reader should picture a street mime juggling non-existent balls. As the mime
continues, the action of juggling slowly brings the balls into existence, at first in
dim outline and then into solid reality. T.W. Korner1
An ordered field F is Dedekind complete if every nonempty subset which is
bounded above has a least upper bound (or supremum).
Exercise 16.0: Show that an ordered field is Dedekind complete iff every nonempty
subset which is bounded below has a greatest upper bound (or infimum).
Our initial definition of R was precisely that it was a Dedekind complete ordered
field. Practically speaking, this is a great foundation for honors calculus and real
analysis, because it contains all the information we need to know about R.
In other words, we have put a neat little black box around our foundational problems. Real analysis works perfectly well without ever having to look in the box.
But curiosity is a fundamental part of mathematics, and at some point most of us
will want to look in the box. This chapter is for those who have reached that point,
i.e., who want to understand a proof of the following theorem.
Theorem 16.1. (Black Box Theorem)
a) There is a Dedekind complete ordered field.
b) If F1 and F2 are Dedekind complete ordered fields, they are isomorphic: that is,
there is a bijection : F1 F2 such that:
(i) For all x, y F1 , (x + y) = (x) + (y).
(ii) For all x, y F1 , (xy) = (x)(y).
(iii) For all x, y F1 , x y (x) (y).
c) The isomorphism of part b) is unique: there is exactly one such map between
any two Dedekind complete ordered fields.
The Black Box Theorem explains why we never needed any further axioms of R
beyond the fact that it is a Dedekind complete ordered field: there is exactly one
such structure, up to isomorphism.2
1Thomas William K
orner, 1946
2The student unfamiliar with the notion of isomorphism should think of it as nothing else

than a relabelling of the points of R. For instance consider the x-axis Rx = {(a, 0) | a R}
and the y-axis Ry = {(0, a) | a R} in the plane. These are two copies of R. Are they the
same? Not in a hard-nosed set-theoretic sense: they are different subsets of the plane. But
they are essentially the same: the bijection which carries (a, 0) 7 (0, a) preserves the addition,
multiplication and order relation. So really we have two slightly different presentations of the same
333

334

16. FOUNDATIONS REVISITED

We will prove the Black Box Theorem...eventually. But rather than taking the
most direct possible route we broaden our focus to a study of the structure of
ordered fields, not just Q and R.
1. Ordered Fields
1.1. Basic Definitions.
In this section we revisit the considerations of 1.2 from a somewhat different perspective. Before we listed certain ordered field axioms, but the perspective there
was that we were collecting true, and basic, facts about the real numbers for use in
our work with them. This time our perspective is to study and understand the collection of all ordered fields. One of our main goals is to construct the real numbers
R in terms of the rational numbers Q and to understand this in terms of a more
general process, completion, which can be applied in any ordered field.
A field is a set F endowed with two binary operations + and which satisfy
all of the field axioms (P0) through (P9). To a first approximation, these axioms
simply encode the usual rules of addition, subtraction, multiplication and division
of numbers, so any field can be thought of as a kind of generalized number system. The most important basic examples are the rational numbers Q, the real
numbers R, and the complex numbers C. But there are other examples which seem
farther removed from the usual numbers: e.g. finite fields like F2 = {0, 1} are
smaller than what we normally think of as a number system, whereas the set R(t) of
all rational functions (with real coefficients) is a field whose elements are naturally
regarded as functions, not as numbers.
Field theory is an active branch of mathematical research, with several texts and
thousands of papers devoted to it (e.g. [FT]). Nevertheless the very simple properties of fields established in 1.2.1 will be sufficient for our needs here, in part
because we are not interested in fields per se but rather ordered fields. An ordered
field is a field equipped with the additional structure of a total order relation,
namely a binary relation which satisfies:
(TO1)
(TO2)
(TO3)
(TO4)

Reflexivity: for all x F , x x.

Antisymmetry: for all x, y F , if x y and y x, then x = y.
Transitivity: for all x, y, z F , if x y and y z then x z.
Totality: for all x, y F , either x y or y x.

Given a total order, we may define a strict ordering x < y by x y but x 6= y.

This is (as we well know!) natural and useful. From an abstract perspective the
key remark is that it is essentially equivalent to a total ordering. Namely, any strict
ordering coming from a total ordering satisfies the following modified versions of
the above axioms:
(TR) Trichotomy: for all x, y F , exactly one of the following holds:
x < y; x = y; or x > y.
essential structure. An arbitrary isomorphism is no more than this, except that the relabelling
map might be more complicated.

1. ORDERED FIELDS

335

(TO3): Transitivity: for all x, y, z F , if x < y and y < z then x < z.

Exercise 16.1: Let (F, ) be a total order relation on the set F .
a) Show that the associated strict ordering satisfies (TR) and (TO3).
b) Let < be a binary relation on F satisfying (TR) and (TO3). Define x y by
x < y or x = y. Show that this gives a total ordering on F .
c) Show that the processes of passing from a total ordering to its strict ordering and
from a strict ordering to its total ordering are mutually inverse: doing one followed
by the other, in either order, brings us back where we started.
In light of the preceding exercise we may (and shall) in fact pass from a total
ordering to its strict ordering and back without explicit comment.
An ordered field is a field (F, +, ) equipped with a total ordering which is
compatible with the field structure in the sense of satisfying the following two
familiar axioms:
(P11) If x, y > 0, then x + y > 0.
(P12) If x, y > 0, then xy > 0.
We refer the reader back to to 1.2.2 for the most elementary consequences of
these axioms, e.g. Proposition 1.11.
Now we introduce a new idea. Let F and F 0 be ordered fields. (To be formally
corrrect, we should speak of the ordered field (F, +, , <), and similarly for F 0 .
In practice this extra notation weighs us down without any advantages in clarity
or precision.) An ordered field homomorphism is a map f : F F 0 satisfying
all of the following:
(FH1) For all x, y F , f (x + y) = f (x) + f (y).
(FH2) For all x, y F , f (xy) = f (x)f (y).
(FH3) f (1) = 1.
(OFH) For all x F , if x > 0, then f (x) > 0.
Exercise 16.2: Show that for any homomorphism of ordered fields, f (0) = 0.
Because of Exercise 16.2 it is natural to wonder whether we really need (FH3)
or whether in fact f (1) = 1 follows automatically from the other axioms. The
answer is a resounding no. Indeed, in the absence of (FH3), for any ordered fields
F and F 0 we could get a homomorphism betweeen them simply by mapping every
x F to 0. This is not an interesting map,3 and as we shall see the theory is
cleaner for not allowing it.
If F and F 0 are fields, then a field homomorphism is a map f : F F 0 sayisfying
(FH1) through (FH3).
Proposition 16.2. Every field homomorphism f : F F 0 is injective.
3Nothing can come of nothing. King Lear

336

16. FOUNDATIONS REVISITED

Proof. Assume to the contrary that there are x 6= y F with f (x) = f (y).
Then f (x y) = f (x) f (y) = 0. Since x 6= y, x y 6= 0, and thus we have a
1
multiplicative inverse xy
. Then
1 = f (1) = f ((x y)

1
1
1
) = f (x y)f (
) = 0 f(
) = 0,
xy
xy
xy

contradicting axiom (P0). Note the use of (FH3)!

A field isomorphism is a field homomorphism f : F F 0 which has an inverse:

i.e., there is a field homomorphism f 0 : F 0 F such that f 0 f = 1F , f f 0 = 1F 0 .
Exercise 16.3: Show that for a field homomorphism f : F F 0 , TFAE:
(i) f is a field isomorphism.
(ii) f is bijective.4
(iii) f is surjective.
Exercise 16.4: Let f : F F 0 be a field homomorphism. Show that f induces
a field isomorphism F f (F ).
A subfield of a field K is a subset of K which is a field under the binary operations + and on K. For example, Q is a subfield of R and R is a subfield of C.
Exercise 16.5: Let K be a field and F a subset of K. Show that F is a subfield iff all of the following hold:
(SF1) 0, 1 F . (In particular F is nonempty.)
(SF2) For all x, y F , x + y, x y, xy F .
(SF3) If x F and x 6= 0, then x1 F .
Thus whenever we have a field homomorphism f : F F 0 , we get an isomorphic copy of F as a subfield of F 0 , namely f (F ). Because of this it is safe to think
of of F itself as a subfield of F 0 , and this perspective is often useful.
Exercise 16.6: Let K be an ordered field, and let F K be a subfield. Show
that restricting the total ordering on K to F endows F with the structure of an
ordered field in such a way that the inclusion map : F K is a homomorphism
of ordered fields.
We say that an element a of a field F is a sum of squares if there is n Z+ and
x1 , . . . , xn F such that a = x21 + . . . + x2n . Notice that in any field 0 and 1 are
sums of squares.
Exercise 16.7: Let F be a field.
a) Let x F be a sum of squares. Show that for any compatible ordering on F ,
4Some philistines take this as the definition of a field isomorphism. Without getting into
too much detail about it, this is really the wrong definition. It happens to be equivalent to the
right definition for fields, but it has an analogue for other types of isomorphisms of mathematical
structure which is not always true. E.g. this definition suggests that an isomorphism of topological
spaces should be a continuous bijection, and this is truly weaker than the correct definition, namely
a continuous map between topological spaces which admits a continuous inverse.

1. ORDERED FIELDS

337

x 0. b) Suppose that there is a nonzero element x F such that both x and x

are sums of squares. Show that there does not exist an ordering on F compatible
with the field structure.
c) Deduce that the field C of complex numbers admits no compatible ordering.
Exercise 16.8: Let F be a field.
a) Suppose that F has the following property: for each nonzero x F , exactly one
of x and x is a sum of squares. Show that F admits a unique compatible ordering:
namely x 0 iff x is a sum of squares.
b) Show that the standard ordering on R is the only compatible ordering.
c) Show that the standard ordering on Q is the only compatible ordering.
Exercise 16.9: A field F is formally real if 1 is not a sum of squares in F .
a) Suppose that F admits a compatible ordering. Show that F is formally real.
(Hint: show the contrapositive.)
b) We suppose that 1 + 1 6= 0 F . (If not, then 1 = 1 is a sum of squares, so by
part a) F admits no compatible ordering.) remarkable theorem of Artin-Schreier
asserts for x F , the following are equivalent:
(i) For every compatible ordering on F , x 0.
(ii) x is a sum of squares.
Deduce from this theorem that if F is formally real then it admits at least one
ordering. (Hint: (i) holds vacuously if there are no compatible orderings!)
Proposition 16.3. For every ordered field F , there is a unique ordered field
homomorphism : Q F .
Proof. Step 1: Since F can be ordered, the sum of positive elements is positive: in particular the sum of positive elements is not zero. Thus for any positive
integer n, 1 + . . . + 1 (n times) is not zero. The map which sends 0 to 0 and
each positive integer to (n) = 1 + . . . + 1 (n times) therefore gives an injective
map N F . We can extend this map to negative integers by mapping n to the
additive inverse of 1 + . . . + 1 (n times). Further, for each positive integer n, since
1
1
, and we may map m
(n) 6= 0, it has a multiplicative inverse (n)
n 7 (m) (n) . It is
now straightforward (but not completely trivial) to check that this map : Q F
is a field homomorphism: we leave this verification to the reader.
Step 2: In fact we had no choice in the matter: the map is the unique field homomorphism from Q to F . We leave this to the reader.
1
Step 3: We must check that if x = m
n Q is positive, then so is (a) = (m) (n) .
But this is easy: (m) is 1 + . . . + 1, so it is a sum of positive elements and is
1
thus positive. Similarly (n) is positive, and thus so is its reciprocal (n)
. Finally,
m
1
( n ) = (m) (n) is a product of two positive elements, hence positive.

In light of this result, for any ordered field F , we may view Q as a subfield.
1.2. Some Topology of Ordered Fields.
Let (K, ) be a linearly ordered set, and let F K be a subset. We say that
F is cofinal in K if for all x K, there is y F with y > x. In fact this is precisely the concept that in the case of subsets of R we call unbounded above. We
use this terminology as an indication that this property can behave a bit differently

338

16. FOUNDATIONS REVISITED

in a non-Archimedean field.
By definition, an ordered field K is Archimedean if Z is cofinal in K. Equivalently, K is Archimedean iff Q is cofinal in K.
Let F be a subfield of the ordered field F . We say that F is dense in K if
for all x < y K, there is z F such that x < z < y.
Lemma 16.4. Let K be an ordered field, and let F be a subfield of K. If F is
dense in K, then F is cofinal in K.
Proof. We show the contrapositive: suppose F is not cofinal in K: there is
x K such that for all y F , y x. Then the interval (x, x + 1) contains no
points of F , so F is not dense in K.

More generally, let f : F F 0 be a homomorphism of ordered fields. We say that
f is cofinal if the image f (F ) is a cofinal subfield of F 0 . We say that f is dense
if the image f (F ) is a dense subfield of F 0 .
Exercise 16.10: Let K be a subfield of F .
a) Suppose that for every F , there is a sequence {xn } of elements of K such
that xn . Show that K is a dense subfield of F .
b) Does the converse of part a) hold? (Hint: no, but counterexamples are not so
easy to come by.)
Lemma 16.5. For a homomorphism f : F F 0 of ordered fields, TFAE:
(i) f is cofinal.
(ii) For every positive 0 F 0 , there is a positive F such that f () < 0 .
Exercise 16.11: Prove Lemma 16.5. (Hint: take reciprocals!)
Lemma 16.6. For an ordered field F , the following are equivalent:
(i) F is Archimedean.
(ii) Q is a dense subfield of F .
Proof. (i) = (ii): Suppose F is Archimedean and let x < y F . Let
1
n Z+ be such that yx
< n; then 0 < n1 < y x, so
1
< y.
n
(ii) = (i): If Q is dense in F then by Lemma 16.4, Q is cofinal in F .
x<x+

Let S be a subset of an ordered field F . For s S, we say that S is discrete

at s if there is a positive  F such that (s , s + ) S = {s}. We say that
S is discrete if it is discrete at s for all s S. An ordered field homomorphism
f : F F 0 is discrete if f (F ) is a discrete subset of F 0 .
Proposition 16.7. (Doraiss Dichotomy) Let f : F F 0 be an ordered field
homomorphism. Then exactly one of the following holds:
(i) f is cofinal.
(ii) f is discrete.
Proof. We will use the characterization of cofinality from Lemma 16.5.
Suppose f is cofinal. Then 0 f (F ) and for every positive 0 F 0 , there is a

1. ORDERED FIELDS

339

positive F such that f () (, ). It follows that f (F ) is not discrete.

Suppose f is not cofinal: there is 0 F 0 such that (0 , 0 ) f (F ) = {0}: f (F )
is discrete at 0. Now let x F . Suppose there is y F such that x  < y < x + .
Then y x (, ). By what weve just seen we must have y x = 0, i.e., y = x.
Thus f (F ) is discrete at each of its elements, so it is discrete.

For any element x in an ordered field, we can define |x| in the usual way, i.e., x if
x 0 and x if x 0.
Exercise 16.12: Let f : F F 0 be an ordered field homomorphism. Show that
for all x F , |f (x)| = f (|x|). (The absolute value on the left hand side is taking
place in F 0 ; the one on the right hand side is taking place in F .)
It is now possible to carry over our definitions of convergent sequencess, Cauchy
sequences and continuous functions to the context of ordered fields and homomorphisms between them. We repeat the basic definitions. In many cases the proofs
are exactly the same as in the case F = R treated in loving detail in this text; when
this is so we leave the proofs to the reader. However, there are some things which do
not carry over to the context of all ordered fields, and we treat these in some detail.
First recall that for any set S, a sequence {xn } in S is given by a function f :
Z+ S; we write xn in place of f (n).
Let {xn } be a sequence in an ordered field F and let x F . We say that xn
converges to x and write xn x if: for all positive  F , there is N Z+ such
that for all n N , |xn x| < . We say that a sequence {xn } is convergent if
xn x for some x F .
We observe that this is verbatim the same as the definition in R. At the same
itme, the fact that our small  is now an element of the ordered field F rather
than (necessarily) a real number has certain surprising implications. The following
exercises exhibits one.
Exercise 16.13: For an ordered field F , show that the following are equivalent:
(i) The sequence { n1 }
n=1 is convergent in F .
(ii) F is Archimedean.
Lemma 16.8. Let {xn } be a sequence in an ordered field F , and let x, y F .
If xn x and xn y, then x = y.
Proof. Left to the reader.

Thus if xn x we may call x the limit of the sequence: it is unique.

A sequence {xn } in an ordered field F is Cauchy if for all positive  F there is
N Z+ such that for all m, n N , |xm xn | < .
Proposition 16.9. A Cauchy sequence in an ordered field which admits a
convergent subsequence is itself convergent.
Proof. Let F be an ordered field, let {xn } be a Cauchy sequence in F , and
let {xnk } be a subsequence converging to x F . Let  > 0. Since {xn } is Cauchy,

340

16. FOUNDATIONS REVISITED

there is N Z+ such that for all m, n N , |xm xn | < 2 . Also there is K N

such that for all k K, |xnk x| < 2 . If n K, then since nK K we have


|xn x| |xn xnK | + |xnK x| < + = ,
2 2
so xn x.

Lemma 16.10. A convergent sequence in an ordered field is Cauchy.
Proof. Left to the reader.

On the other hand, in an arbitrary ordered field a Cauchy sequence need not be
convergent. For instance the Babylonian
sequence
{xn } of VII.2.2 is a sequence

/ Q and limits in R are unique,

of rational numbers converging to 2 R. Since 2
viewed as a sequence in the ordered field Q, the sequence {xn } is not convergent.
It is however Cauchy. More generally we have the following result.
Proposition 16.11. Let f : F F 0 be an ordered field homomorphism, and
let {xn } be a sequence in F . If the sequence {f (xn )} is Cauchy in F 0 , then {xn }
is Cauchy in F .
Proof. We show the contrapositive: suppose {xn } is not Cauchy in F . Then
there is some positive  F such that for all N Z+ there are m, n N such that
|xm xn | . Then
f () f (|xm xn |) = |f (xm xn )| = |f (xm ) f (xn )|.
0

Since  = f () is a positive element of F 0 , this shows that {f (xn )} is not Cauchy.

Thus we get a method for producing nonconvergent Cauchy sequences in an ordered
field F : find a sequence {xn } in F and a homomorphism f : F F 0 such that
f (xn ) converges to an element of F 0 \ f (F ). In fact every nonconvergent Cauchy
sequence arises this way, as we will see later on.
Let F and F 0 be ordered fields, and let f : F F 0 be a map between them.
We say that f is continuous at c F if for all positive 0 F 0 , there is a positive F such that for all x F with |x c| < , |f (x) f (c)| < 0 . We say
that f : F F 0 is continuous if it is continuous at every c F . We say that
f : F F 0 is uniformly continuous if for all positive 0 F 0 there is a positive
F such that for all x, x0 F , if |x x0 | < then |f (x) f (x0 )| < 0 .
Lemma 16.12. Let f : F F 0 be a function.
a) If f is continuous and xn is a sequence in F which converges to x F , then
f (xn ) f (x) in F 0 .
b) If f is uniformly continuous and {xn } is a Cauchy sequence in F , then {f (xn )}
is a Cauchy sequence in F 0 .
Proof. Left to the reader.

Warning: When F = F 0 = R, the converse of Lemma 16.6a) holds (Theorem 10.5):

a map which preserves limits of convergent sequences is necessarily continuous. This
does not hold in general; but unfortunately counterexamples lie beyond the scope
of this text.

1. ORDERED FIELDS

341

Theorem 16.13. (F. Dorais) Let f : F F 0 be a homomorphism of ordered

fields. The following are equivalent:
(i) f is uniformly continuous.
(ii) f is continuous.
(iii) f is continuous at 0.
(iv) f is cofinal.
Proof. (i) = (ii) = (iii) is immediate.
(iii) (iv): By Lemma 16.5, f (F ) is cofinal in F 0 iff for every positive 0 in
F 0 , there is a positive in F such that f () < 0 . Suppose this holds, and let
0 be positive in F 0 , and choose a positive F as above. Then for x F , if
|x| = |x 0| < , then
|f (x)| = f (|x|) < f () = |f () f (0)| < 0 .
This shows f is continuous at 0. Conversely, if f is continuous at 0, then for each
positive 0 F there is a positive F such that if |x| < , |f (x)| < 0 . Thus 2
0
0
is a positive element of F such that f ()
2 <  , so f (F ) is cofinal in F by Lemma
16.5.
(iii) = (i): Let 0 be a positive element of F 0 ; since f is continuous at 0 there is
a positive F such that for all x F with |x| < , |f (x)| < 0 . So, if x, y F
are such that |x x0 | < , then
|f (x) f (y)| = |f (x y)| < 0 .

1.3. A Non-Archimedean Ordered Field.
Let K = R((t)) be the field of P
formal Laurent series with R-coefficients: an
n
element of K is a formal sum
nZ an t where there exists N Z such that
an = 0 for all n < N . We add such series term by term and multiply them in the
same way that we multiply polynomials.
Exercise 16.14: Show that K = R((t)) is a field.
We need to equip K with an ordering; equivalently, we need to specify a set of
positive elements. For every nonzero element x K, we take v(x) to be the smallest n Z such that an 6= 0. Then we say that x is positive if the coefficient av(x )
of the smallest nonzero term is a positive real number. It is straightforward to see
that the sum and product of positive elements is positive and that for each nonzero
x K, exactly one of x and x is positive, so this gives an ordering on K in the
usual way: we decree that x < y iff y x is positive.
We observe that this ordering is non-Archimedean. Indeed, the element 1t is
positive its one nonzero coefficient is 1, which is a positive real number and
infinitely large: for any n Z, 1t n is still positive recall that we look to the
smallest degree coefficient to check positivity so 1t > n for all n.
Next we observe that the set { t1n } is unbounded in K. Taking reciprocals, it
follows that the sequence {tn } converges to 0 in K: explicitly, given any  > 0 here
 is not necessarily a real number but any positive element of F ! for all sufficiently
large n we have that t1n > 1 , so |tn | = tn < . We will use this fact to give a simple
explicit description of all convergent sequences in F . First, realize that a sequence

342

16. FOUNDATIONS REVISITED

P
in K consists of, for each m Z+ a formal Laurent series xm = nZ am,n tn , so in
fact for each n Z we have a real sequence {am,n }
m=1 . Now consider the following
conditions on a sequence {xm } in K:
(i) There is an integer N such that for all m Z+ and n < N , am,n = 0, and
(ii) For each n Z the sequence am,n is eventually constant: i.e., for all sufficiently
large m, am,n = Cn R. (Because of (i) we must have Cn = 0 for all n < N .)
Then condition (i) is equivalent to boundedness of the sequence.
P
I claim that if the sequence converges say xm x = n=N an tn F then (i)
and (ii) both hold. Indeed convergent sequences are bounded, so (i) holds. Then for
all n N , am,n is eventually constant in m iff am,n an is eventually constant in m,
so we may consider xm x instead of xm and thus we may assume that xm 0 and
try to show that for each fixed n, am,n is eventually equal to 0. As above, this holds
iff for all k 0, there exists Mk such that for all m Mk , |xm | tk . This latter
condition holds iff the coefficient am,n of tn in xn is zero for all N < k. Thus, for all
m Mk , am,N = am,N +1 = . . . = am,k1 = 0, which is what we wanted to show.
Conversely, suppose (i) and (ii) hold. Then since for all n N the sequence
am,n is eventually constant, we may define an to be this P
eventual value, and an
argument very similar to the above shows that xm x = nN an tn .
Next I claim that if a sequence {xn } is Cauchy, then it satisfies (i) and (ii) above,
hence is convergent. Again (i) is immediate because every Cauchy sequence is
bounded. The Cauchy condition here says: for all k 0, there exists Mk such
that for all m, m0 Mk we have |xm x0m | tk , or equivalently, for all n < k,
am,n am0 ,n = 0. In other words this shows that for each fixed n < k and all
m Mk , the sequence am,n is constant, so in particular for all n N the sequence
am,n is eventually constant in m, so the sequence xm converges.
Exercise 16.15: Show that the subfield R((t2 )) of R((t)) is cofinal but not dense.
2. The Sequential Completion
2.1. Sequentially Complete Fields.
An ordered field F is sequentially complete if every Cauchy sequence in F converges to an element of F . We have seen that a Dedekind complete ordered field is
sequentially complete. Here we wish to examine the converse.
Theorem 16.14. For an Archimedean ordered field F , TFAE:
(i) F is Dedekind complete.
(ii) F is sequentially complete: every Cauchy sequence converges.
Proof. The implication (i) = (ii) is the content of Theorem 10.32, since
the Bolzano-Weierstrass Theorem holds in any ordered field satisfying (LUB).
(ii) = (i): Let S F be nonempty and bounded above, and write U(S) for
the set of least upper bounds of S. Our strategy will be to construct a decreasing
Cauchy sequence in U(S) and show that its limit is sup S.

2. THE SEQUENTIAL COMPLETION

343

Let a S and b U(S). Using the Archimedean property, we choose a negative

integer m < a and a positive integer M > b, so
m < a b M.
+

For each n Z , we define

k
U(A) and k 2n M }.
2n
Every element of Sn lies in the interval [2n m, 2n M ] and 2n M Sn , so each Sn is
kn
n
finite and nonempty. Put kn = min Sn and an = k2nn , so 22k
n+1 = 2n U(S) wihle
2kn 2
kn 1
/ U(S). It follows that we have either kn+1 = 2kn or kn+1 = 2kn 1
2n+1 = 2n
1
and thus either an+1 = an or an+1 = an 2n+1
. In particular {an } is decreasing.
For all 1 m < n we have
Sn = {k Z |

0 am an = (am am+1 ) + (am+1 am+2 ) + . . . + (an1 an )

2(m+1) + . . . + 2n = 2m .
Thus {an } is Cauchy, hence by our assumption on F an L F .
We claim L = sup(S). Seeking a contradiction we suppose that L
/ U(S).
Then there exists x S such that L < x, and thus there exists n Z+ such that
an L = |an L| < x L.
It follows that an < x, contradicting an U(S). So L U(S). Finally, if there
exists L0 U(S) with L0 < L, then (using the Archimedean property) choose
n Z+ with 21n < L L0 , and then
1
1
L n > L0 ,
2n
2
U(S), contradicting the minimality of kn .
an

so an

1
2n

kn 1
2n

The proof of (ii) = (i) in Theorem 16.14 above is taken from [HS] by way of
[Ha11]. It is rather unexpectedly complicated, but I do not know a simpler proof
at this level. However, if one is willing to introduce the notion of convergent and
Cauchy nets, then one can show first that in an Archimedean ordered field, the
convergence of all Cauchy sequences implies the convergence of all Cauchy nets,
and second use the hypothesis that all Cauchy nets converge to give a proof which
is (in my opinion of course) more conceptually transparent. This is the approach
taken in my (more advanced) notes on Field Theory [FT].
In fact there are (many!) non-Archimedean sequentially complete ordered fields.
We will attempt to describe two very different examples of such fields here. We
hasten to add that this is material that the majority of working research mathematicians are happily unfamiliar with, and which is thus extremely rarely covered
in undergraduate courses. Only the exceptionally curious need the next section.
2.2. Sequential Completion I: Statement and Applications.
We will now establish one of our main results: for every ordered field F , there
is a sequentially complete ordered field R and a homomorphism f : F R.
In fact we can, and will prove, even more than this. The point is that there will
be many (nonisomorphic) sequentially complete fields into which any given ordered

344

16. FOUNDATIONS REVISITED

field embeds. For example, when we construct the real numbers we will have an
embedding Q , R. But we also have an embedding R , R((t)), so taking the
composite gives an embedding Q R((t)). (There is no way that R and R((t)) are
isomorphic, since the former is Archimedean and the latter is not.)
We would like a general definition which allows us to prefer the embedding
Q , R to the embedding Q , R((t)). The key observation is that, since R is
Archimedean, the embedding of Q into R is dense, whereas since R((t)) is not
Archimedean, the embedding of Q into R is not dense. This leads to the following
important definition.
A sequential completion of an ordered field F is a dense embedding F , R
into a sequentially complete ordered field.
Lemma 16.15. For an ordered field F , the following are equivalent.
(i) F is Dedekind complete.
(ii) The inclusion : Q , F makes F into a sequential completion of Q.
Proof. (i) = (ii): By Theorem 16.14, F is sequentially complete and
Archimedean. By Lemma 16.6, Q = (Q) is a dense subfield of F , and it follows
that F is a sequential completion of Q.

We will prove that every ordered field admits a sequential completion. And again,
we will in fact prove a bit more.
Theorem 16.16. Let F be an ordered field.
a) F admits a sequential completion : F R.
b) If L is any sequentially complete ordered field and f : F L is a cofinal ordered
field homomorphism, then there is a unique ordered field homomorphism g : R L
such that f = g .
Corollary 16.17. Two sequential completions of the same ordered field are
isomorphic.
Proof. Let 1 : F R1 and 2 : F R2 be two sequential completions.
Applying Theorem 16.16 with R = R1 and f = 2 : F R2 , we get a unique
homomorphism g : R1 R2 such that 2 = g 1 . Interchanging the roles of R1
and R2 we also get a unique homomorphism g 0 : R2 R1 such that 1 = g 0 2 .
Now consider g 0 g : R1 R . We have
(g 0 g) 1 = g 0 (g 1 ) = g 0 2 = 1 .
Applying Theorem 16.16 with L = R = R1 we get that there is a unique homomorphism G : R R1 such that G1 = 1 , but clearly the identity map 1R1 has this
property. Thus we must have g 0 g = 1R1 . Similarly considering g g 0 : R2 R2 ,
then in view of
(g g 0 ) 2 = g (g 0 2 ) = g 1 = 2 ,
we deduce that g g 0 = 1R2 . In other words, g and g 0 are mutually inverse isomorphisms...so R1 and R2 are isomorphic.

Applying Theorem 16.16 to the ordered field Q, we get a sequential completion R
of Q. Since R is Archimedean and sequentially complete, by Theorem 16.14, R is
Dedekind complete. Conversely, by Lemma 16.15any Dedekind complete ordered

2. THE SEQUENTIAL COMPLETION

345

field R0 is isomorphic to R. Thus the existence and uniqueness statements of Theorem 16.16 imply the existence and uniqueness up to isomorphism of a Dedekind
complete ordered field.
The uniqueness statement can be strengthened: let R1 and R2 be two Dedekind
complete ordered fields. We claim that not only are they isomorphic, but that the
isomorphism between them is unique. Indeed, for i = 1, 2 let i : Q Ri be the
inclusion maps. We saw above that there is a unique map g : R1 R2 such that
g 1 = 2 and this g is an isomorphism. But any isomorphism h : R1 R2
will satisfy h 1 = 2 , since in fact there is exactly one embedding from Q into
any ordered field. Thus whereas in general there is an isomorphim g between two
sequential completions of a given ordered field F which is unique such that blah
blah blah (more precisely, such that g 1 = 2 ), in this case the blah blah blah
is vacuous and the isomorphism is unique full stop.
In abstract mathematics, uniqueness up to a unique isomorphism is as close to
identical as we can reasonably ask for two structures to be. (Even the horizontal
copy of R and the vertical copy of R are different sets, but the obvious isomorphism between them is the only isomorphism, so no trouble can arise by identifying
the two.) We denote this unique field by R and call it the real numbers...of course.
2.3. Sequential Completion II: The Proof.
Now we are properly motivated to roll up our sleeves and endure the rather lengthy,
technical proof of Theorem 16.16. The essential idea (which is indeed due to A.L.
Cauchy) is to build the sequential completion directly from the set C of all Cauchy
sequences in F .
We can observe that C itself has some structure reminiscent of an ordered field
but that things do not quite work out: it is somehow too large to itself be an ordered field. Namely, it makes perfectly good sense to add, subtract and multiply
Cauchy sequences in F . For that matter, it makes perfectly good sense to add,
subtract and multiply arbitrary sequences in F : we simply put
{xn } + {yn } = {xn + yn },
{xn } {yn } = {xn yn },
{xn } {yn } = {xn yn }.
It remains to check that these operations take Cauchy sequences to Cauchy sequences. At the very beginning of our study of sequences we showed this for
convergent sequences (in R, but the proofs certainly did not use any form of the
completeness axiom). It is no more difficult to establish the analogue for Cauchy
sequences in F .
Lemma 16.18. Let F be any ordered field, and let a , b be Cauchy sequences.
Then a + b and a b are both Cauchy.
Proof. Since a and b are both Cauchy, for  > 0 there is N Z+ such that
for m, n N , |am an | <  and and |bm bn | < . Then
|(am + bm ) (an + bn )| |am an | + |bm bn | < 2.

346

16. FOUNDATIONS REVISITED

Further, since the sequences are Cauchy, they are bounded: there are Ma , Mb F
such that |an | Ma and |bn | Mb for all n Z+ . Then for m, n N ,
|am bm an bn | |am an ||bm | + |an ||bm bn | (Ma + Mb ).

So does this addition and multiplication endow C with the structure of a field?
There is an additive identity, namely the sequence with xn = 0 for all n. There
is also a multiplicative identity, namely the sequence with xn = 1 for all n. It all
works well until we get to multiplicative inverses.
Exercise 16.16: Let {xn } be a sequence in the ordered field F .
a) Show that there is a sequence {yn } with {xn } {yn } = {1} if and only if for all
n Z+ , xn 6= 0.
b) Show that if {xn } is Cauchy and xn 6= 0 for all n, then its inverse { x1n } is again
a Cauchy sequence.
Thus there are plenty of Cauchy sequences other than the constantly zero sequence
which do not have multiplicative inverses: e.g. (0, 1, 1, 1, . . .), or indeed any constant
sequence which takes the value 0 at least once. Thus C has many good algebraic
properties, but it is not the case that every nonzero element has a multiplicative
inverse, so it is not a field.5
We also have some order structure on C. For instance, it is tempting to define
{xn } > {yn } if xn > yn for all n. This turns out not to be a good definition in the
sense that it does not lead to a trichotomy: there will be unequal Cauchy sequences
{xn } and {yn } for which neither is less than the other, e.g.
{xn } = {0, 1, 1, . . .}, {yn } = {1, 0, 0, . . .}.
As in the definition of convergence, it is more fruitful to pay attention to what
a Cauchy sequence is doing eventually. Exploiting this idea we can get a sort of
trichotomy result.
Lemma 16.19. (Cauchy Trichotomy) For a Cauchy sequence {xn } in an ordered
field F , exactly one of the following holds:
(i) There is a positive element  F and N Z+ such that xn  for all n N .
(ii) There is a positive element  F and N Z+ such that xn  for all n N .
(iii) xn converges to 0.
Proof. It is easy to see that the conditions are mutually exclusive. Let us
suppose that (iii) does not hold: thus there is  > 0 and a subsequence {xnk }
such that |xnk |  for all k Z+ . By passing to a further subsequence we may
assume either that xnk  for all k or xnk  for all k. Let us suppose that
the former holds and show that this implies (i): if so, replacing x by x shows
that the latter alternative implies (ii). Since {xn } is Cauchy, there is N Z+ such
that |xm xn | 2 for all m, n N . Putting these two conditions together we get
xn 2 for all n N .

5for those who know some abstract algebra: what weve shown is that (C, +, ) is a commutative ring. There is a very general algebraic method which, when given a commutative ring,
will yield a collection of fields defined in terms of that ring. The present construction is indeed
an instance of this.

2. THE SEQUENTIAL COMPLETION

347

Unfortunately this is not quite the kind of trichotomy which defines a total order
relation: we have some elements that we regard as positive case (i) above, some
elements that we regard as negative case (ii) above but for an order relation
only the zero element should be neither positive nor negative, whereas case (iii)
above includes the much larger collection of elements converging to zero.
Lemma 16.19 suggests that if we could somehow squash down the subset of
Cauchy sequences which converge to 0 to a single point, then we would actually get
a total order relation. This business of squashing subsets to a point is formalized
in mathematics (more so in algebra and topology than the part of mathematics
weve been studying for most of this text!) by an equivalence relation. Rather
than providing a logically complete but pedagogically useless whirlwind tour of
equivalence relations, we will simply assume that the reader is familiar with them.6
Namely, we regard any two Cauchy sequences which converge to 0 as equivalent.
We are left with the question of when to regard two Cauchy sequences which do
not converge to zero as equivalent. We could simply not squash them, i.e., declare two such sequences to be equivalent exactly when they are equal. But a
little exploration shows that this wont work: well get a total order relation but it
wont interact well with the algebraic structure. For instance, consider the Cauchy
sequences
{xn } = (0, 0, 1, 1, 1, . . .), {yn } = (1, 1, 1, 1, . . .).
The difference {xn } {yn } converges to 0 so is getting identified with 0. Thus we
should identify {xn } and {yn } as well. This leads to the following key definitions.
Let Z be the set of all sequences in F which converge to 0; convergent sequences
are Cauchy, so certainly Z C. For two Cauchy sequences a , b C, we put
a b a b Z.
In words, two Cauchy sequences are equivalent iff their difference converges to zero.
Exercise 16.17 (if you know abstract algebra): Show that Z is a maximal ideal
in the commutative ring C. Why is this an exciting sign that were on the right
track?
Exercise 16.18: Let {xn } be a Cauchy sequence in F , and let {xnk } be any subsequence. Show that {xn } {xnk }.
Now we define R as C/ , that is, the set of equivalence classes of Cauchy sequences. This will be the underlying set of our sequential completion. It remains
to endow it with all the rest of the structure. The idea here is that when we
pass to a quotient by an equivalence relation we can try to simply carry over the
structure we already had, but at every step we must check that the operations are
well-defined, meaning they are independent of the chosen equivalence class. At no
point are these verifications difficult, but we admit they can be somewhat tedious.
Let us check the addition and multiplication induced well-defined operations on
6At UGA they are covered in the transitional Math 3200 course. The reader who has made
it through most of this text will have no problem learning this concept.

348

16. FOUNDATIONS REVISITED

the set R of equivalence classes. This means: if we have four Cauchy sequences
a , b , c , d and a c , b d , then
a + b c + d , a b c d .
All right: since a c and b d , a c 0 and b d 0, so
(a + b (c + d ) = (a c ) + (b d ) 0 + 0 = 0,
so a + b c + d . Similarly,
a b c d = (a c )b + (b d )c ,
and this converges to 0 because a c , b d 0 and b , c are bounded. Thus
we have equipped our set R with two binary operations + and .
Proposition 16.20. (R, +, ) is a field.
Proof. Most of these axioms are completely straightforward (but, yes, somewhat tedious) to verify and are left to the reader as exercises. Let us single out:
(P3) The additive identity is [0 ], the class of the constant sequence 0.
(P7) The multiplicative identity is [1 ], the class of the constant sequence 1.
(P8) Suppose that x R \ {[0 ]}, and let x be any Cauchy sequence representing
x. Then we must have xn 6= 0 for all sufficiently large n: indeed, otherwise we
would have 0 = (0, 0, 0, . . .) as a subsequence, and if a subsequence of a Cauchy
sequence converges to 0, then the Cauchy sequence itself converges to 0, contradiction. Suppose xn 6= 0 for all n > N . Then define y by yn = 0 for all 1 n N
(or whatever you want: it doesnt matter) and yn = x1n for n > N . Then xn yn = 1
for all n > N , so x y differs from 1 by a sequence which is convergent to zero:
[x ][y = [1 ] = 1, so y = [y ] is the multiplicative inverse of x = [x ].

We now equip R with an ordering. For a , b C, we put [a ] > [b ] if there is a
positive element  in F such that an bn  for all sufficiently large n. We claim
that this is well-defined independent of the representatives a and b chosen. Indeed,
if x and y converge to zero, then for sufficiently large n we have |xn yn | < 2
and then

an + xn ) (bn + yn ) = (an bn ) + (xn yn ) .
2
Theorem 16.21. The strict ordering < on R endows it with the structure of
an ordered field.
Proof. Step 1: The trichotomy property follows from Lemma 16.19. The
transitive property is easy: if [a ] < [b ] and [b ] < [c ], then there is N Z+
and 1 , 2 > 0 such that for all n N , bn an 1 and cn bn 2 , hence
cn an = (cn bn ) + (bn an ) 1 + 2 . Thus < is a strict ordering on R.
Step 2: If [a ], [b ] > 0, then as above there is N Z+ and 1 , 2 > 0 such that
for all n N , an 1 and bn 2 . Then for all n N , an + bn 1 + 2 and
an bn 1 2 .

Proposition 16.22. For an ordered field F , let : F R by sending x F
to the class of the constant sequence (x, x, . . .). Then is a dense homomorphism
of ordered fields.
Proof. Step 1: It is immediate that is a field homomorphism. Further, if
x > 0, then (x, x, . . .) > 0: indeed, we can take  = x. So is a homomorphism of
ordered fields.

2. THE SEQUENTIAL COMPLETION

349

Step 2: Let x = [a ] in R. We claim that (an ) x. Indeed, if  = [ ] is a positive

element of R, then there is e > 0 in F such that n  for all sufficiently large n.
Since {an } is Cauchy in F , there is N Z+ such that for all m, n N ,
e
|an am | < .
2
Then for sufficiently large m,
e
m |an am | e |an am | ,
2
which shows that for all n N , |(an ) x| < . Step 3: It remains to show that
ba
is dense, so let a = [a ] < b = [b ] R. Let m = a+b
2 and let  = 2 ; thus the
interval of width  centered at m has as its endpoints a and b. By Step 2, there is
a sequence {xn } F such that (xn ) m. By definition then, for all sufficiently
large n we have|(xn ) m| < , and thus (xn ) (a, b).

Finally, we prove Theorem 16.16 in the following explicit form.
Theorem 16.23. For any ordered field F , the homomorphism : F R =
C/Z is a sequential completion. Moreover, if L is a sequentially complete field and
f : F L is a cofinal ordered field homomorphism, then there is a unique ordered
field homomorphism g : R L such that f = g .
Proof. We already know that : F R is a dense ordered field homomorphism. It remains to check first that R is sequentially complete and second that
satisfies the universal property of Theorem 16.16b).
Step 1: Let {xn } be a Cauchy sequence in R; we must show that it is convergent. By Proposition 16.9 it suffices to check this after passing to a subsequence.
If {xn } has only finitely many distinct points, we have a constant subsequence,
which is certainly convergent. Otherwise {xn } has infinitely many distinct points,
so that after passage to a subsequence we may assume they are all distinct. Since
by Proposition 16.22 is a dense homomorphism, for all n Z+ there is yn F
such that yn lies strictly between xn and xn+1 . It follows that for any n, k Z+ ,
yn+k lies in any interval containing xn , xn+1 , . . . , xn+k+1 . Since {xn } is Cauchy,
this forces {yn } be to Cauchy, and thus y = [{yn }] R. We also have xn yn 0,
so limn xn = limn yn = y.
Step 2: Let L be a sequentially complete ordered field, and let f : F L be a
cofinal ordered field homomorphism. We must show that there is a unique ordered
field homomorphism g : R L such that f = g .
Uniqueness: Let g : R L be such that f = g . We will see explicitly what
g must be on each x R. Write x = [a ]. Since {an } is Cauchy in F and f is
cofinal, f is uniformly continuous, so {f (an )} is Cauchy in L, hence convergent, say
to y L. Since f is cofinal, is g, hence g is uniformly continuous. Since (an ) x,
f (an ) = g((an )) g(x). Since f (an ) y and limits are unique, we must have
g(x) = y = lim f (an ).
n

Existence: We must show that putting g(x) = y as above defines an ordered field
homomorphism from R to L. If x1 = [an ] and x2 = [bn ], let y1 = limn f (an )
and y2 = limn f (bn ). Then an + bn x1 + x2 and f (an + bn ) = f (an ) +
f (bn ) y1 + y2 , so g(x1 + x2 ) = y1 + y2 = g(x1 ) + g(x2 ). Similarly, an bn x1 x2
and f (an bn ) = f (an )f (bn ) y1 y2 , so g(x1 x2 ) = y1 y2 = g(x1 )g(x2 ). Finally, if

350

16. FOUNDATIONS REVISITED

x = [a ] > 0, then there is a positive  F such that an  for all sufficiently large
n, so
g(x) = lim f (an ) lim f () = f () > 0.
n

Bibliography
[A]
T.M. Apostol, Calculus. Volume One.
[Ab84] Y.S. Abu-Mostafa, A Differentiation Test for Absolute Convergence. Math. Mag. 57
(1984), 228-231.
[Ac00] F. Acerbi, Plato: Parmenides 149a7-c3. A Proof by Complete Induction? Archive for
History of the Exact Sciences 55 (2000), 5776.
[ADC12] S. Ali and M. Deutsche Cohen, The phi-ratio tests. Elem. Math. 67 (2012), 164-168.
[Al08] S.A. Ali, The mth ratio test: new convergence tests for series. Amer. Math. Monthly 115
(2008), 514-524.
[An95] P. Andrews, Where not to find the critical points of a polynomialvariation on a Putnam
theme. Amer. Math. Monthly 102 (1995), 155-158.
[Ap73] T.M. Apostol, Another Elementary Proof of Eulers Formula for (2n). Amer. Math.
Monthly 80 (1973), 425431.
[As12] J.M. Ash, The Limit Comparison Test Needs Positivity. Math. Magazine 85 (2012), 374
375.
[Ba96] R.G. Bartle, Return to the Riemann integral. Amer. Math. Monthly 103 (1996), 625-632.
[B]
J.C. Bowman, Honours Calculus, online lecture notes.
[Ba70] R. Baer, Dichte, Archimedizit
at und Starrheit geordneter K
orper. Math. Ann. 188 (1970),
165-205.
[Ba22] S. Banach, Sur les op
erations dans les ensembles abstraits et leur applications aux

equations int
egrales. Fund. Math. 3 (1922), 133181.
[Ba98] B. Banaschewski, On proving the existence of complete ordered fields. Amer. Math.
Monthly 105 (1998), 548551.
[Be06] A.F. Beardon, Contractions of the Real Line. Amer. Math. Monthly 113 (2006), 557558.
[Be12] S.J. Bernstein, D
emonstration du th
eor`
eme de Weierstrass fond
ee sur le calcul des probabilit
es. Communications of the Kharkov Mathematical Society 13 (1912), 12.
Besenyei, Lebesgues Road to Antiderivatives. Math. Mag. 86 (2013), 255260.
[Be13] A.
Besenyei, Peanos unnoticed proof of Borels theorem. Amer. Math. Monthly 121
[Be14] A.
(2014), 69-72.
[BM22] H. Bohr and J. Mollerup, Lrebog i matematisk Analyse, vol. 3, Jul. Gjellerups Forlag,
Copenhagen, 1922.
[Bo71] R.P. Boas, Jr., Signs of Derivatives and Analytic Behavior. Amer. Math. Monthly 78
(1971), 10851093.
Borel, Sur quelque points de la th

[Bo95] E.
eorie des fonctions. Ann. Sci. lEcole
Norm. Sup. 12
(1895), 955.
[BS]
R.G. Bartle and D.R. Sherbert, Introduction to real analysis. Second edition. John Wiley
& Sons, Inc., New York, 1992.
[C]
H. Cartan, Elementary theory of analytic functions of one or several complex variables.
Dover Publications, Inc., New York, 1995.
[Ca21] A.L. Cauchy, Analyse alg
ebrique, 1821.
[Ca89] A.L. Cauchy, Sur la convergence des s
eries, in Oeuvres compl`
etes S
er. 2, Vol. 7, GauthierVillars (1889), 267279.
[Ch01] D.R. Chalice, How to Differentiate and Integrate Sequences. Amer. Math. Monthly 108
(2001), 911921.
[CJ]
R. Courant and F. John, Introduction to Calculus and Analysis.
[Cl10] P.L. Clark, Real induction. http://math.uga.edu/pete/realinduction.pdf
[Cl11] P.L.
Clark,
Induction
and
completeness
in
ordered
sets.
http://math.uga.edu/pete/induction completeness brief.pdf
351

352

[Co77]

BIBLIOGRAPHY

G.L. Cohen, Is Every Absolutely Convergent Series Convergent? The Mathematical

Gazette 61 (1977), 204213.
[CdC] K. Conrad, The contraction mapping theorem. http://www.math.uconn.edu/
kconrad/blurbs/analysis/contractionshort.pdf
[CdD] K.
Conrad,
Estimating
the
size
of
a
divergent
sum.
http://www.math.uconn.edu/kconrad/blurbs/analysis/sumest.pdf
[Co49] J.L. Coolidge, The story of the binomial theorem. Amer. Math. Monthly 56 (1949), 147157.
[Cu65] F. Cunningham, Jr., Classroom Notes: The Two Fundamental Theorems of Calculus.
Amer. Math. Monthly 72 (1965), 406-407.
P 1
2
[Da12] D. Daners, A Short Elementary Proof of
= 6 . Math. Mag. 85 (2012), 361364.
k2

[dBR76] P. du Bois-Reymond, Uber

den G
ultigkeitschbereich der Taylorschen Reihenentwickelung. Sitzungb. k. Bayer. Akad. Wiss., math.-phys. Klasse (1876), 225237.

[dBR83] P. du Bois-Reymond, Uber

den G
ultigkeitschbereich der Taylorschen Reihenentwickelung. Math. Ann. 21 (1883), 107119.
[DC]
P.L. Clark, Discrete calculus. In preparation. Draft available on request.
[DR50] A. Dvoretzky and C.A. Rogers, Absolute and unconditional convergence in normed linear
spaces. Proc. Nat. Acad. Sci. USA 36 (1950), 192197.
[DS]
Dirichlet
series,
notes
by
P.L.
Clark,
available
at
http://math.uga.edu/pete/4400dirichlet.pdf
[Ed62] M. Edelstein, On fixed and periodic points under contractive mappings. J. London Math.
Soc. 37 (1962), 74-79.
[Eh94] P. Ehrlich, Dedekind cuts of Archimedean complete ordered abelian groups. Algebra Universalis 37 (1997), 223-234.
[Er94] M. Erickson, An introduction to combinatorial existence theorems. Math. Mag. 67 (1994),
118-123.
[ES35] P. Erd
os and G. Szekeres, A combinatorial problem in geometry. Compositio Math. 2
(1935), 463-470.
[FT]
Field
Theory,
notes
by
P.L.
Clark,
available
at
http://www.math.uga.edu/pete/FieldTheory.pdf
[Go]
R. Gordon, Real Analysis: A First Course. Second Edition, Addison-Wesley, 2001.
[Gr]
P.M. Gruber, Convex and discrete geometry. Grundlehren der Mathematischen Wissenschaften 336. Springer, Berlin, 2007.
[Ha88] J. Hadamard, Sur le rayon de convergence des s
eries ordonn
ees suivant les puissances
dune variable. C. R. Acad. Sci. Paris 106 (1888), 259262.
[Ha11] J.F. Hall, Completeness of Ordered Fields. 2011 arxiv preprint.
[Ha50] H.J. Hamilton, A type of variation on Newtons method. Amer. Math. Monthly 57 (1950),
517-522.
[H]
G.H. Hardy, A course of pure mathematics. Centenary edition. Reprint of the tenth (1952)
edition with a foreword by T. W. K
orner. Cambridge University Press, Cambridge, 2008.
[Ha02] F. Hartmann, Investigating Possible Boundaries Between Convergence and Divergence.
College Math. Journal 33 (2002), 405406.
[Ha11] D. Hathaway, Using Continuity Induction. College Math. Journal 42 (2011), 229231.
[HS]
E. Hewitt and K. Stromberg, Real and abstract analysis. A modern treatment of the
theory of functions of a real variable. Third printing. Graduate Texts in Mathematics,
No. 25. Springer-Verlag, New York-Heidelberg, 1975.
[Ho66] A. Howard, Classroom Notes: On the Convergence of the Binomial Series. Amer. Math.
Monthly 73 (1966), 760-761.
[Ho95] M.E. Hoffman, Derivative polynomials for tangent and secant. Amer. Math. Monthly 102
(1995), 23-30.
[Ka07] I. Kalantari, Induction over the continuum. Induction, algorithmic learning theory, and
philosophy, 145154, Log. Epistemol. Unity Sci., 9, Springer, Dordrecht, 2007.
[Ke70] H. Kestelman, Riemann Integration of Limit Functions. Amer. Math. Monthly 77 (1970),
182187.
[Kn80] W.J. Knight, Functions with zero right derivatives are constant. Amer. Math. Monthly
87 (1980), 657-658.

BIBLIOGRAPHY

353

[K
o91] T.W. K
orner, Differentiable functions on the rationals. Bull. London Math. Soc. 23
(1991), 557-562.
[La]
J. Labute, Math 255, Lecture 22:
Power Series:
The Binomial Series.
http://www.math.mcgill.ca/labute/courses/255w03/L22.pdf
[L]
S. Lang, Undergraduate analysis. Second edition. Undergraduate Texts in Mathematics.
Springer-Verlag, New York, 1997.
[Li33] F.A Lindemann, The Unique Factorization of a Positive Integer. Quart. J. Math. 4,
319320, 1933.
[LV06] M. Longo and V. Valori, The Comparison Test Not Just for Nonnegative Series. Math.
Magazine 79 (2006), 205210.
[Lu99] J. Lu, Is the Composite Function Integrable? Amer. Math. Monthly 106 (1999), 763766.
[Ma56] A.M. Macbeath, A criterion for differentiability. Edinburgh Math. Notes (1956), 8-11.
[Ma42] C. Maclaurin, Treatise of fluxions, 1. Edinburgh (1742), 289290.
P
2
1
[Ma61] Y. Matsuoka, An elementary proof of the formula
k=1 k2 = 6 . Amer. Math. Monthly
68 (1961), 485487.
[MK09] M.M.
Marjanovi
c
and
Z.
Kadelburg,
Limits
of
composite
functions.
The
Teaching
of
Mathematics,
XII
(2009),
16.
http://elib.mi.sanu.ac.rs/files/journals/tm/22/tm1211.pdf
[Me08] A. Melman, Bounds on the zeros of the derivative of a polynomial with all real zeros.
Amer. Math. Monthly 115 (2008), 145-147.
[Me72] F. Mertens, Ueber die Multiplicationsregel f
ur zwei unendliche Reihen. Journal f
ur die
Reine und Angewandte Mathematik 79 (1874), 182184.
[MS22] E.H. Moore and H.L. Smith, A General Theory of Limits. Amer. J. of Math. 44 (1922),
102121.
[Mo50] R.K. Morley, Classroom Notes: The Remainder in Computing by Series. Amer. Math.
Monthly 57 (1950), 550551.
[Mo51] R.K. Morley, Classroom Notes: Further Note on the Remainder in Computing by Series.
Amer. Math. Monthly 58 (1951), 410412.
[Mo57] T.E. Mott, Newtons method and multiple roots. Amer. Math. Monthly 64 (1957), 635638.
[Mu63] A.A. Mullin, Recursive function theory (A modern look at a Euclidean idea). Bulletin of
the American Mathematical Society 69 (1963), 737.
[Ne03] Nelsen, R.B. An Impoved Remainder Estimate for Use with the Integral Test. College
Math. Journal 34 (2003), 397399.
[Ne81] D.J. Newman, Differentiation of asymptotic formulas. Amer. Math. Monthly 88 (1981),
526-527.
[NP88] D.J. Newman and T.D. Parsons, On monotone subsequences. Amer. Math. Monthly 95
(1988), 44-45.
[No52] M.J. Norris, Integrability of continuous functions. Amer. Math. Monthly 59 (1952), 244245.
[Ol27] L. Olivier, Journal f
ur die Reine und Angewandte Mathematik 2 (1827), 34.
[PFD] P.L.
Clark,
Partial
Fractions
Via
Linear
Algebra,
http://www.math.uga.edu/pete/partialfractions.pdf.
[Ro63] K. Rogers, Classroom Notes: Unique Factorization. Amer. Math. Monthly 70 (1963),
547548.
[R]
W. Rudin, Principles of mathematical analysis. Third edition. International Series in
Pure and Applied Mathematics. McGraw-Hill, New York-Auckland-D
usseldorf, 1976.
[Sa95] H. Samelson, More on Kummers test. Amer. Math. Monthly 102 (1995), 817-818.
[Sc]
M. Schramm, Introduction to Real Analysis. Dover edition, 2008.
[SP88] D. Scott and D.R. Peeples, The Teaching of Mathematics: A Constructive Proof of the
Partial Fraction Decomposition. Amer. Math. Monthly 95 (1988), 651-653.
[Se59] A. Seidenberg, A simple proof of a theorem of Erd
os and Szekeres. J. London Math. Soc.
34 (1959), 352.
[Sm86] R.S. Smith, Rolle over Lagrange Another Shot at the Mean Value Theorem. College
Math. J. 17 (1986), 403406.
[S]
M. Spivak, Calculus. Fourth edition.
[St95] S.K. Stein, Do Estimates of an Integral Really Improve as n Increases? Math. Mag. 68
(1995), 16-26.

354

[St37]
[St48]
[Str90]
[Ta55]
[To94]
[T]
[Ve02]
[Wa48]
[Wa95]
[Wa36]
[Ze34]

BIBLIOGRAPHY

M.H. Stone, Applications of the theory of Boolean rings to general topology. Trans. Amer.
Math. Soc. 41 (1937), 375-481.
M.H. Stone, The generalized Weierstrass approximation theorem. Math. Mag. 21 (1948),
167-184.
G. Strang, Sums and Differences vs. Integrals and Derivatives. College Math. Journal 21
(1990), 2027.
A. Tarski, A lattice-theoretical fixpoint theorem and its applications. Pacific J. Math. 5
(1955), 285-309.
J.C. Tong, Kummers test gives characterizations for convergence or divergence of all
positive series. Amer. Math. Monthly 101 (1994), 450-452.
J.W. Tukey, Convergence and Uniformity in Topology. Annals of Mathematics Studies,
no. 2. Princeton University Press, 1940.
D.J. Velleman, Partial fractions, binomial coefficients, and the integral of an odd power
of sec . Amer. Math. Monthly 109 (2002), 746-749.
H.S. Wall, A modification of Newtons method. Amer. Math. Monthly 55 (1948), 90-94.
J.A. Walsh, The Dynamics of Newtons Method for Cubic Polynomials. The College
Mathematics Journal 26 (1995), 2228.
M. Ward, A Calculus of Sequences. Amer. J. Math. 58 (1936), 255266.
E. Zermelo, Elementare Betrachtungen zur Theorie der Primzahlen. Nachr. Gesellsch.
Wissensch. G
ottingen 1, 4346, 1934.