Invitation To Classical Analysis Compress

Pure and Applied
Sally
The UNDERGRADUATE TEXTS 17
SERIES
Invitation
to Classical
Analysis
Peter Duren
American Mathematical Society

Invitation
to Classical
Analysis
Pure and Applied
Sally
The UNDERGRADUATE TEXTS • 17
SERIES
Invitation
to Classical
Analysis
Peter Duren
American Mathematical Society

Providence, Rhode Island
EDITORIAL COMMITTEE
Paul J. Sally, Jr. (Chair) Joseph Silverman
Francis Su Susan Tolman
2010 Mathematics Subject Classification. Primary 11–01, 11B68, 26–01, 33–01, 34–01,
40E05, 40–01, 41–01, 42–01.
For additional information and updates on this book, visit

www.ams.org/bookpages/amstext-17
Library of Congress Cataloging-in-Publication Data

Duren, Peter L., 1935–
Invitation to classical analysis / Peter Duren.
p. cm. – (Pure and applied undergraduate texts ; v. 17)
Includes bibliographical references and indexes.
ISBN 978-0-8218-6932-1 (alk. paper)
1. Functional analysis–Textbooks. I. Title.
QA320.D87 2012
515.7–dc23
2011045853
Copying and reprinting. Individual readers of this publication, and nonprofit libraries
acting for them, are permitted to make fair use of the material, such as to copy a chapter for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for such
permission should be addressed to the Acquisitions Department, American Mathematical Society,
201 Charles Street, Providence, Rhode Island 02904-2294 USA. Requests can also be made by
e-mail to reprint-permission@ams.org.

c 2012 by the American Mathematical Society. All rights reserved.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www.ams.org/
10 9 8 7 6 5 4 3 2 1 17 16 15 14 13 12
Jacob Bernoulli
Carl Friedrich Gauss
Augustin-Louis Cauchy
Wikimedia Commons. Public Domain. Courtesy of the Archives of the Mathematisches Forschungsinstitut Ober- Wikimedia Commons. Public Domain.
wolfach. © Universität Göttingen Sammlung Sternwarte.
Leonard Euler
Bernard Bolzano
Joseph-Louis Lagrange
Wikimedia Commons. Public Domain. Wikimedia Commons. Public Domain. Wikimedia Commons. Public Domain.
Georg Cantor
Joseph Fourier
Bernhard Riemann
Archives of the Mathematisches Forschungsinstitut Oberwolfach. Wikimedia Commons. Public Domain. Wikimedia Commons. Public Domain.
Karl Weierstrass
Niels Henrik Abel
Peter Gustav Lejeune Dirichlet

Archives of the Mathematisches Forschungsinstitut Oberwolfach. Photogravure u. Druck v.H. Riffarth. Berlin. Wikimedia Commons. Public Domain.
Contents
Preface xi
Chapter 1. Basic Principles 1

1.1. Mathematical induction 1
1.2. Real numbers 2
1.3. Completeness principles 5
1.4. Numerical sequences 12
1.5. Infinite series 18
1.6. Continuous functions and derivatives 22
1.7. The Riemann integral 27
1.8. Uniform convergence 33
1.9. Historical remarks 39
1.10. Metric spaces 41
1.11. Complex numbers 42
Exercises 46
Chapter 2. Special Sequences 51

2.1. The number e 51
2.2. Irrationality of π 55
2.3. Euler’s constant 56
2.4. Vieta’s product formula 60
2.5. Wallis product formula 61
2.6. Stirling’s formula 63
Exercises 66
v
vi Contents
Chapter 3. Power Series and Related Topics 73

3.1. General properties of power series 73
3.2. Abel’s theorem 76
3.3. Cauchy products and Mertens’ theorem 81
3.4. Taylor’s formula with remainder 83
3.5. Newton’s binomial series 87
3.6. Composition of power series 89
3.7. Euler’s sum 92
3.8. Continuous nowhere differentiable functions 98
Exercises 102
Chapter 4. Inequalities 109

4.1. Elementary inequalities 109
4.2. Cauchy’s inequality 112
4.3. Arithmetic–geometric mean inequality 117
4.4. Integral analogues 118
4.5. Jensen’s inequality 119
4.6. Hilbert’s inequality 122
Exercises 127
Chapter 5. Infinite Products 131

5.1. Basic concepts 131
5.2. Absolute convergence 135
5.3. Logarithmic series 136
Exercises 141
Chapter 6. Approximation by Polynomials 145

6.1. Interpolation 145
6.2. Weierstrass approximation theorem 151
6.3. Landau’s proof 153
6.4. Bernstein polynomials 157
6.5. Best approximation 160
6.6. Stone–Weierstrass theorem 164
6.7. Refinements of Weierstrass theorem 168
Exercises 171
Contents vii
Chapter 7. Tauberian Theorems 179

7.1. Summation of divergent series 179
7.2. Tauber’s theorem 182
7.3. Theorems of Hardy and Littlewood 183
7.4. Karamata’s proof 185
7.5. Hardy’s power series 190
Exercises 193
Chapter 8. Fourier Series 197

8.1. Physical origins 197
8.2. Orthogonality relations 199
8.3. Mean-square approximation 200
8.4. Convergence of Fourier series 203
8.5. Examples 207
8.6. Gibbs’ phenomenon 212
8.7. Arithmetic means of partial sums 215
8.8. Continuous functions with divergent Fourier series 219
8.9. Fourier transforms 221
8.10. Inversion of Fourier transforms 228
8.11. Poisson summation formula 232
Exercises 236
Chapter 9. The Gamma Function 247

9.1. Probability integral 247
9.2. Gamma function 249
9.3. Beta function 251
9.4. Legendre’s duplication formula 252
9.5. Euler’s reflection formula 253
9.6. Infinite product representation 255
9.7. Generalization of Stirling’s formula 257
9.8. Bohr–Mollerup theorem 257
9.9. A special integral 261
Exercises 262
viii Contents
Chapter 10. Two Topics in Number Theory 269

10.1. Equidistributed sequences 269
10.2. Weyl’s criterion 271
10.3. The Riemann zeta function 276
10.4. Connection with the gamma function 280
10.5. Functional equation 282
Exercises 286
Chapter 11. Bernoulli Numbers 291

11.1. Calculation of Bernoulli numbers 291
11.2. Sums of positive powers 294
11.3. Euler’s sums 295
11.4. Bernoulli polynomials 297
11.5. Euler–Maclaurin summation formula 300
11.6. Applications of Euler–Maclaurin formula 302
Exercises 305
Chapter 12. The Cantor Set 309

12.1. Cardinal numbers 309
12.2. Lebesgue measure 313
12.3. The Cantor set 315
12.4. The Cantor–Scheeffer function 317
12.5. Space-filling curves 320
Exercises 323
Chapter 13. Differential Equations 327

13.1. Existence and uniqueness of solutions 327
13.2. Wronskians 333
13.3. Power series solutions 336
13.4. Bessel functions 343
13.5. Hypergeometric functions 348
13.6. Oscillation and comparison theorems 354
13.7. Refinements of Sturm’s theory 358
Exercises 360
Contents ix
Chapter 14. Elliptic Integrals 369

14.1. Standard forms 369
14.2. Fagnano’s duplication formula 371
14.3. The arithmetic–geometric mean 373
14.4. The Legendre relation 381
Exercises 384
Index of Names 387
Subject Index 389

Preface
This is a book for undergraduates. To be more precise, it is designed

for students who have learned the basic principles of analysis, as taught
to undergraduates in advanced calculus courses, and are prepared to ex-
plore substantial topics in classical analysis. And there is much to explore:
Fourier series, orthogonal polynomials, Stirling’s formula, the gamma func-
tion, Bernoulli numbers, elliptic integrals, Bessel functions, Tauberian the-
orems, etc. Yet the modern undergraduate curriculum typically does not
encompass such topics, except perhaps by way of physical applications. In
effect the student struggles to master abstract concepts and general theo-
rems of analysis, then is left wondering what to do with them.
It was not always so. Around 1950 the typical advanced calculus course
in American colleges contained a selection of concrete topics such as those
just mentioned. However, the development could not be entirely rigorous
because the underlying theory of calculus had been deferred to graduate
courses. To remedy this unsatisfactory state of affairs, the theory of calculus
was moved to the undergraduate level. Textbooks by Walter Rudin and
Creighton Buck helped transform advanced calculus to a theoretical study
of basic principles. Certainly much was gained in the process, but also much
was lost. Various concrete topics, natural sequels to the abstract theory,
were crowded out of the curriculum.
The purpose of this book is to recover the lost topics and introduce
others, making them accessible at the undergraduate level by building on the
theoretical foundation provided in modern advanced calculus courses. My
aim has been to develop the mathematics in a rigorous way while holding
the prerequisites to a minimum. The exposition probes rather deeply into
each topic and is at times intellectually demanding, but every effort has
xi
xii Preface
been made to include the background necessary for full comprehension. It is

hoped that undergraduate students (and other readers) will find the material
exciting and will be inspired to make further studies in the realm of classical
analysis.
The book evolved from a collection of notes prepared over the years
for students in my own advanced calculus courses, as supplements to the
text, presenting concrete topics such as the Wallis product formula, Stir-
ling’s formula, the Weierstrass approximation theorem, Euler’s sum, and
the convergence of Fourier series. Similar handouts for students in differen-
tial equations treated the method of Frobenius and the Sturm comparison
theorem, both with applications to Bessel functions. Versions of the chapters
on infinite products and the gamma function served as background material
for students in my complex analysis courses. Feedback from students in all
of those courses helped to improve the presentation.
Although the reader is assumed to have acquired a good command of
basic principles, a preliminary chapter is included by way of review, and to
provide a convenient reference when the principles are invoked in later chap-
ters. The second and third chapters also contain some standard material.
The book is designed for individual study, but may serve also as a text for a
second-semester course in advanced calculus. There is some progression of
material throughout the book, but each chapter is largely self-contained.
The desire to maintain an elementary level, excluding techniques of com-
plex analysis and Lebesgue integration, sometimes led to awkward problems
of exposition. In the end I am especially impressed by the power of Fourier
analysis. The convergence theorem for Fourier series often turns out to be
a viable substitute for the more systematic methods of complex analysis.
On the other hand, some topics do call for more advanced techniques. For
instance, Lebesgue integrals are the natural setting for Fourier transforms,
and the functional equation for the Riemann zeta function can be fully un-
derstood only in the context of complex function theory. Nevertheless, I felt
compelled to discuss both topics in the book.
Classical analysis is a vast area that defies comprehensive treatment.
Any book on the subject must make a selection of topics. With its focus
on functions of a single real variable, this book automatically excludes a
great many attractive topics. Otherwise, the choices partly reflect my own
background and interests.
Many exercises are offered. Mathematics is not a spectator sport; it is
learned by participation. My conviction too is that abstract principles are
best appreciated when applied to good problems. Therefore, while some
exercises are straightforward, I have tried to make others more interesting,
Preface xiii
more challenging, and consequently more rewarding. Hints are often pro-
vided. A few of the problems, not really exercises, come with an asterisk
and an invitation to consult the literature. References for each topic are
grouped at the end of the relevant chapter.
Historical notes are sprinkled throughout the book. To put a human
face on the mathematics, the book includes capsule scientific biographies of
the major players and a gallery of portraits. Some historical notes also shed
light on the origin and evolution of mathematical ideas. A few discussions
of physical applications serve the same purpose.
Many friends and colleagues helped to shape the book. I am especially
indebted to Dick Askey, Martin Chuaqui, Dima Khavinson, Jeff Lagarias,
Hugh Montgomery, Brad Osgood, Bill Ross, and Harold Shapiro for math-
ematical suggestions and encouragement as the writing progressed. Alex
Lapanowski, a student in the undergraduate honors program at Michigan,
read large parts of the manuscript and made valuable suggestions. Dragan
Vukotić and his students Irina Arévalo and Diana Giraldo also read portions
of the manuscript and spotted a number of small errors. I am enormously
grateful to David Ullrich, who read the manuscript carefully, checked many
of the exercises, pointed out errors and inaccuracies, and suggested impor-
tant improvements in the exposition. In particular, he devised the relatively
simple proof of Hilbert’s inequality presented in Chapter 4. Thanks to all of
these readers the book is better, but any remaining faults are the author’s
responsibility.
In the final stages of preparation the AMS production staff made expert
contributions to the book. Special thanks go to the editor Ina Mette for her
continual encouragement and willingness, for better or worse, to accommo-
date my peculiar wishes in regard to content and format. Finally, I must
acknowledge the important role of my wife Gay. Without her support the
book would not have been written.
Peter Duren
Chapter 1
Basic Principles
Classical analysis is rooted in differential and integral calculus. Calculus

began with Newton and Leibniz as a collection of ideas not precisely for-
mulated, but highly effective in scientific work. As the years progressed,
however, mathematicians recognized the limitations of intuitive concepts
and sought to place calculus on a firm theoretical foundation. Their efforts
led to a framework of definitions, theorems, and proofs that has become part
of the standard curriculum for students of advanced calculus.
This preliminary chapter offers a rapid overview of the basic theory of
calculus. Definitions and major theorems are stated for later reference, and
proofs are generally included. It is assumed that the reader is already ac-
quainted with most of this material but may feel a need for review. The
coverage is not complete, as some relevant topics are omitted. More exten-
sive treatments can be found in introductory texts such as Rudin [8], Ross
[7], or Mattuck [6].
1.1. Mathematical induction

Before turning to the calculus, we want to give a brief review of mathe-
matical induction. This is an important method for verifying relations that
depend on positive integers. The positive integers, also known as the natural
numbers, are simply the numbers n = 1, 2, 3, . . . . The symbol N will denote
the set of all positive integers.
Suppose now that Pn is a proposition expressed in terms of an integer
n, which is to be proved for all n ∈ N. The first step is to verify that P1
is true. This is usually quite trivial. Next comes the inductive step. Under
1
2 1. Basic Principles
the inductive hypothesis that Pn is true for some n ∈ N, it is deduced that

Pn+1 must also be true. Since P1 is true, we conclude that P2 is true, which
implies that P3 is true, and so on. In this manner we are able to prove that
Pn is true for every n ∈ N.
As a simple illustration, suppose that Pn is the proposition that
n(n + 1)
1 +2 +3 + ···+ n = .
2
1(1+1)
Then P1 is clearly true, since 2 = 1. Suppose now that Pn is true for
some n ∈ N. Then
n(n + 1) (n + 1)(n + 2)
1 + 2 + 3 + · · · + n + (n + 1) = +n+1= ,
2 2
which says that Pn+1 is true. Thus the formula holds for all n ∈ N.
To give another example, let Pn be the inequality n2 ≤ 2n . The inequal-
ity is easily verified for n = 1 and n = 2, but it is false for n = 3, since
32 = 9 > 8 = 23 . However, it is again true for n = 4, and an inductive
argument shows that it remains true for all n ≥ 4. Indeed, suppose that
n2 ≤ 2n for some n ≥ 4. Then
2
n+1 1 2 n
2
(n + 1) = n ≤ 1+
2
2 ≤ 2n+1 ,
n n
which shows that the truth of Pn implies that of Pn+1 , provided that n ≥ 4.
Therefore, since P4 is true, it follows that Pn is true for all n ≥ 4.
Other applications of induction appear in exercises at the end of this
chapter.
1.2. Real numbers

The mathematician Leopold Kronecker is reputed to have said, “God
created the positive integers. All the rest is the work of man.” We will not
attempt a formal construction of the real number system here but will need
to review some basic principles.
The set of all integers (positive, negative, or zero) is customarily denoted
by Z, presumably because “Zahlen” is the German word for “numbers”. The
rational numbers are the quotients m/n of integers m, n ∈ Z with n = 0.
A rational number m/n is said to be in lowest terms if m and n have no
1.2. Real numbers 3
common divisor; that is, if there is no integer k ≥ 2 such that m/k and n/k
are integers. The set of all rational numbers is denoted by Q.
The symbol R denotes the set of all real numbers. We will take for
granted the existence of the system of real numbers with its familiar prop-
erties of addition and multiplication. It is well known that the rational
numbers are everywhere dense in R. In other words, if a, b ∈ R and a < b,
then a < r < b for some r ∈ Q.
However,
√ not every real number is rational. To confirm this, let us show
that 2 is irrational; more precisely, there is no rational number r such that
r2 = 2. Suppose, for purpose of contradiction, that some rational number
r has the property r2 = 2. We may take r > 0 and express it in lowest
terms by r = m/n, where m, n ∈ N. Then r2 = 2 implies m2 = 2n2 , so
that m2 is even. Therefore, m is even and so m2 = (2k)2 = 4k 2 for some
k ∈ N. Hence n2 = 2k 2 and n2 is even, so that n is also even. Since m and
n are both even, they have a common divisor 2, contrary to our assumption
that r = m/n is in lowest terms. This contradiction
√ shows that no rational
2
number r can have the property r = 2. Thus 2 is irrational.
Like the rationals, the irrational numbers are everywhere dense in R. In
fact,
√ it follows from the density of the rationals that the rational multiples
of 2 are everywhere dense.
A number x0 ∈ R is said to be algebraic if it satisfies an equation
P (x) = a0 + a1 x + a2 x2 + · · · + an xn = 0
for some polynomial P whose coefficients a0 , a1 , . . . , an are integers, not all

zero.
√ It is easily seen that every rational number is algebraic. The number
2 is also algebraic, since it satisfies the equation x2 − 2 = 0. Real numbers
that are not algebraic are called transcendental. It is known, for instance,
that the numbers π and e are transcendental. (See Chapter 2, Sections 2.1
and 2.2 for further discussion.)
It is an important fact that the rational numbers form a countable set.
An infinite set S is said to be countable or denumerable if it can be put into
one-to-one correspondence with the set N of positive integers. This amounts
to saying that the elements of S can be arranged in a sequence x1 , x2 , x3 , . . .
with no repetitions. To show that the set Q has this property, it suffices to
show that the set of all positive rational numbers is countable. To see this,
arrange the numbers r = m/n, where m, n ∈ N, in an infinite rectangular
array with m/n in the mth row and nth column. Then by weaving diagonally
through the array, as indicated in the diagram below, the set of all positive
rationals can be placed in a sequence without repetition. Whenever the path
encounters a rational number already recorded in different form, it passes
over that number without including it in the sequence.

1
1 → 1
2
1
3 → 1
4
1
5 ···

2
1
2
2
2
3
2
4
2
5 ···
↓
3
1
3
2
3
3
3
4
3
5 ···

4
1
4
2
4
3
4
4
4
5 ···
.. .. .. .. ..
. . . . .
Thus the resulting sequence is

1 1 2 3 1 1 2 3 4 5 1 1
1, 2, 1, 1, 3, 4, 3, 2, 1, 1, 5, 6, . . . .
Every positive rational number must appear exactly once in the sequence.
This proves that Q is a countable set.
On the other hand, the set R of all real numbers is not countable. To
see this, it suffices to show that the real numbers in the interval 0 ≤ x < 1
do not form a countable set. The idea of the proof is quite ingenious and
is due to Georg Cantor (1845–1918), the founder of modern set theory.
(More of Cantor’s ideas will be discussed in Chapter 12.) Cantor’s proof
proceeds as follows. Suppose, for purpose of contradiction, that the set
[0, 1) = {x ∈ R : 0 ≤ x < 1} is countable, so that all of its members can be
arranged in a sequence
x1 = 0.d11 d12 d13 . . .

x2 = 0.d21 d22 d23 . . .
x3 = 0.d31 d32 d33 . . .
. . .
Here xn = 0.dn1 dn2 dn3 . . . is the decimal expansion of the number xn , chosen
to end in a sequence of 0’s (not 9’s) in cases of ambiguity. We will reach
a contradiction by displaying a number in [0, 1) that is not on the list.
If 0 ≤ dnn ≤ 4, set bn = 7. If 5 ≤ dnn ≤ 9, set bn = 2. Then the
number y = 0.b1 b2 b3 . . . lies in [0, 1) but does not appear in the purported
enumeration. Indeed, y = xn for each n because the nth digit of its decimal
expansion differs from that of xn by at least 3, so that |y − xn | > 10−n .
This shows that no sequence of numbers in [0, 1) can exhaust the entire
interval, and so [0, 1) is uncountable. Therefore, the set of all real numbers
is uncountable.
The real number system has a very important property, known as com-
pleteness, that manifests itself in various ways and lies at the very founda-
tion of analysis. One form of completeness is the least upper bound principle,
whose statement requires us first to recall some terminology. A nonempty
set S ⊂ R is said to be bounded above if there is a number y ∈ R such that
x ≤ y for all x ∈ S. Such a number y is called an upper bound of S. A
number M is called the least upper bound or supremum of S if M is an upper
bound and M ≤ y for every upper bound y. This is indicated by writing
M = sup S. Similarly, a set S ⊂ R is bounded below if there is a number
y ∈ R, called a lower bound of S, such that y ≤ x for all x ∈ S. A number m
is called the greatest lower bound or infimum of S if y ≤ m for every lower
bound y. For this we write m = inf S. A set S ⊂ R is said to be bounded if
it is bounded both above and below.
Least Upper Bound Principle. If a set of real numbers is bounded above,

then it has a least upper bound.
Equivalently, every set of real numbers that is bounded below has a

greatest lower bound. We will take the least upper bound principle as an
axiom. If the real number system is actually constructed, for instance by the
method of “Dedekind cuts”, the least upper bound principle can be deduced
as a theorem. (See for instance Rudin [8].)
It is instructive to note that the system of rational numbers is not com-
plete, since it fails to have the least upper bound property. Consider, for
instance, the set S = {r ∈ Q : r2 < 2}. Clearly, the set S has a (ratio-
nal) upper bound, √ but it does not have a smallest rational upper bound,
essentially because 2 is irrational.
Finally, we need to set some notation for intervals of real numbers. For
a, b ∈ R with a < b, we define [a, b] = {x ∈ R : a ≤ x ≤ b} and (a, b) =
{x ∈ R : a < x < b}. The intervals a ≤ x a and
x < b are denoted by (a, ∞) and (−∞, b). We will usually write R instead
of (−∞, ∞). The empty set is denoted by ∅.
1.3. Completeness principles

We now develop some further principles, basic to analysis, that can be
regarded as expressing the completeness of the real number system. These
are the monotone boundedness theorem, the Bolzano–Weierstrass theorem,
the Cauchy criterion for convergence, and the Heine–Borel theorem. Each
will be derived ultimately from the least upper bound principle, and all are
in some sense equivalent to it. We begin with sequences of real numbers and
the concept of convergence.
A sequence of real numbers can be defined as a mapping from N into
R, a real-valued function defined on the positive integers. It is customarily
written as {xn } = {x1 , x2 , x3 , . . . }, where xn ∈ R for each n ∈ N. The se-
quence is said to converge to a limit L ∈ R if to each ε > 0 there corresponds
a number N such that |xn − L| < ε for all n ≥ N . This is tantamount to
saying that for each prescribed discrepancy ε the numbers xn are within
distance ε of L for all but a finite number of indices n. It is easy to see
that a sequence can have at most one limit. The range of a sequence is its
set of values {xn : n ∈ N}. A sequence is said to be bounded if its range
is a bounded set. Every convergent sequence is bounded, but the converse
is false. The convergence of a sequence {xn } to a limit L is indicated by
writing L = limn→∞ xn or simply xn → L as n → ∞. A sequence that does
not converge is called divergent.
If xn → L and yn → M , then (xn + yn ) → L + M and xn yn → LM . If
also M = 0, then yn = 0 for all n sufficiently large and xn /yn → L/M .
More generally, the notation xn → +∞ (or simply xn → ∞) as n → ∞
indicates that to each a ∈ R there corresponds a number N such that xn > a
for all n ≥ N . This is expressed by saying that the sequence {xn } tends to
+∞. Thus it diverges in a specific way. Similarly, the notation xn → −∞
means that for each a ∈ R there is a number N such that xn < a for all
n ≥ N.
A sequence is called monotonic if it is either nondecreasing or nonin-
creasing; that is, if either
x1 ≤ x2 ≤ x3 ≤ . . . or x1 ≥ x2 ≥ x3 ≥ . . . .
Monotone Boundedness Theorem. Every monotonic and bounded se-

quence of real numbers is convergent.
Proof. It suffices to suppose that the sequence {xn } is nondecreasing. Then

by hypothesis, its range S = {xn : n ∈ N} is bounded above. Therefore, by
the least upper bound principle, the set S has a least upper bound λ = sup S.
We will show that {xn } converges to λ. Since λ is an upper bound for S,
we have xn ≤ λ for all n ∈ N. On the other hand, for each ε > 0 the
number λ − ε is not an upper bound for S, since λ is the least upper bound.
Consequently, xN > λ − ε for some N ∈ N. Since the sequence {xn } is
nondecreasing, it follows that xn > λ − ε for all n ≥ N . Combining this
with the inequality xn ≤ λ, we see that −ε < 0 ≤ λ − xn ≤ ε, or |xn − λ| < ε
for all n ≥ N . This proves that xn → λ as n → ∞.
With the monotone boundedness theorem in hand, we can now prove

the Bolzano–Weierstrass theorem. To state the latter result, we need a
definition. A number ξ is a cluster point of a set S ⊂ R if each deleted
neighborhood of ξ contains a point of S. A deleted neighborhood of ξ is an
interval (a, b) with the point ξ (a < ξ < b) removed. Thus ξ is a cluster point
of S if for each ε > 0 there exists a point x ∈ S such that 0 < |x − ξ| < ε.
Note that a finite set cannot have any cluster points. Also an unbounded
infinite set need not have any cluster points; consider for example the set
N of positive integers. A cluster point of a set need not belong to the set.
For instance, 0 is a cluster point of the set S = { n1 : n ∈ N} but 0 ∈ / S.
Observe also that if ξ is a cluster point of a set S, then every neighborhood
of ξ contains infinitely many points of S.
Bolzano–Weierstrass Theorem. Every bounded infinite set of real num-

bers has a cluster point.
Proof. Let S ⊂ R be a bounded infinite set. Then S is contained in some

closed bounded interval [a, b]. The strategy of proof is to “divide and con-
quer”. The midpoint of the interval is a+b 2 . At least one of the intervals
a+b a+b
[a, 2 ] and [ 2 , b] must contain infinitely many points of S. If the left-
hand interval qualifies, let a1 = a and b1 = a+b 2 . Otherwise, let a1 = 2
a+b
and b1 = b. In either case the subinterval [a1 , b1 ] contains infinitely many

points of S. Now repeat the process. Divide the interval [a1 , b1 ] in half and
choose a subinterval [a2 , b2 ] of the form [a1 , a1 +b a1 +b1
2 ] or [ 2 , b1 ] that again
1
contains infinitely many points of S. Proceeding inductively, construct a

nested sequence of intervals [an , bn ] with lengths 2−n (b − a), each containing
infinitely many points of S.
Since a1 ≤ a2 ≤ a3 ≤ · · · ≤ b and b1 ≥ b2 ≥ b3 ≥ · · · ≥ a, it follows
from the monotone boundedness theorem that both of the sequences {an }
and {bn } are convergent. Thus an → α ≤ b and bn → β ≥ a as n → ∞.
Since an < bn for all n, it is clear that α ≤ β. But in fact α = β, since for
each ε > 0 we have
|α − β| ≤ |α − an | + |an − bn | + |bn − β|
= |α − an | + 2−n (b − a) + |bn − β| < ε
for all n sufficiently large. Let ξ denote the common value of α and β, so
that
lim an = lim bn = ξ .
n→∞ n→∞
The final step in the proof is to show that ξ is a cluster point of S. But
for each given ε > 0 we see that [an , bn ] ⊂ [ξ − ε, ξ + ε] for all n sufficiently
large. Indeed, since {an } increases to ξ and {bn } decreases to ξ, there exists
a number N such that ξ − ε < an ≤ ξ and ξ ≤ bn < ξ + ε for all n ≥ N . But

by construction each interval [an , bn ] contains an infinite number of points
in S. Therefore, for each ε > 0 the deleted neighborhood
{x ∈ R : 0 < |x − ξ| < ε}
contains a point of S, in fact infinitely many. This says that ξ is a cluster

point of S, which proves the theorem.
It is now a short step to the sequential form of the theorem. A subse-

quence of a sequence {xn } is a sequence of the form {xnk }, where nk ∈ N
and 1 ≤ n1 < n2 < n3 < . . . .
Bolzano–Weierstrass Theorem (sequential form). Every bounded se-

quence of real numbers has a convergent subsequence.
Proof. Let {xn } be a bounded sequence and let S = {xn : n ∈ N} be its

set of values. Then S is a bounded set. If S is a finite set, then some value
is repeated infinitely often and so {xn } has a constant subsequence, which is
certainly convergent. If S is an infinite set, the earlier form of the Bolzano–
Weierstrass theorem says that S has a cluster point ξ. Therefore, each
deleted neighborhood of ξ contains a point xn for some n. Having selected a
point xn1 for which 0 < |ξ − xn1 | < 1, we can choose an index n2 > n1 such
that 0 < |ξ−xn2 | < 12 , and then an index n3 > n2 such that 0 < |ξ−xn3 | < 13 ,
and so forth. Then the subsequence {xnk } converges to ξ.
We are now ready to establish the Cauchy criterion for convergence of

a numerical sequence. First let us record the key definition. A Cauchy
sequence of real numbers is a sequence {xn } with the property that for each
ε > 0 there exists an integer N depending on ε such that |xn − xm | < ε
for all n, m ≥ N . It is easy to see that every convergent sequence has this
property. The converse, however, is not so obvious.
Theorem (Cauchy criterion). Every Cauchy sequence of real numbers

is convergent.
Proof. The proof will consist of three main steps. The first step is to show
that every Cauchy sequence is bounded. To prove this, choose for instance
ε = 1 and refer to the definition of a Cauchy sequence to conclude that for
some integer N the inequality |xn − xN | < 1 holds for all n > N . Therefore,
|xn | ≤ |xN | + |xn − xN | ≤ |xN | + 1
for all n > N . But this implies that the sequence is bounded, since it follows
that
|xn | ≤ max |x1 |, |x2 |, . . . , |xN |, |xN | + 1
for all n ∈ N.
The second step is to apply the sequential form of the Bolzano–
Weierstrass theorem. Since a Cauchy sequence is bounded, it must have
a convergent subsequence {xnk }.
The final step is to observe that if a Cauchy sequence has a convergent
subsequence, then in fact the full sequence is convergent. Indeed, if xnk → L
as k → ∞, then for each ε > 0 we see that
ε
|xn − L| ≤ |xn − xnk | + |xnk − L| < |xn − xnk | +
2
for all indices nk sufficiently large. But because {xn } is a Cauchy sequence,
it then follows that
ε ε
|xn − L| < + = ε
2 2
for all n sufficiently large. This shows that xn → L as n → ∞, which
completes the proof.
Next we apply the Bolzano–Weierstrass theorem to prove the Heine–

Borel theorem. The proof will require two preliminary results, known as
the nested sets theorem and the countable basis theorem, which will now be
discussed.
First recall that a set G of real numbers is said to be open if each point
x0 ∈ G has a neighborhood
Nr (x0 ) = {x ∈ R : |x − x0 | < r}
that lies entirely in G. A set F ⊂ R is said to be closed if its complement

= R \ F = {x ∈ R : x ∈
F / F}
is open. The union of any collection of open sets is open. The intersection of
any finite collection of open sets is open. Similarly, an arbitrary intersection
of closed sets is closed, and every finite union of closed sets is closed. It can
be proved (cf. Exercise 11) that a set is closed if and only if it contains all
of its cluster points.
Nested Sets Theorem. Let F1 , F2 , . . . be a sequence of nonempty closed
nested in the sense that Fn+1 ⊂ Fn
bounded sets of real numbers which are
for every n ∈ N. Then the intersection ∞ n=1 Fn is nonempty.
The theorem says that under the stated hypotheses the sets must con-
tain some common element. This may seem obvious, but if either of the
hypotheses “closed” or “bounded” is relaxed the conclusion is false. For

instance, the open intervals In = (0, 1/n), for n = 1, 2, . . . , are nonempty,

bounded, and nested: I1 ⊃ I2 ⊃ . . . . However, their intersection ∞ n=1 In
is empty. The intervals [n, ∞) are closed and nested but not bounded, and
again their intersection is empty.
Proof of theorem. The theorem is trivial if some set Fm is finite, because

then Fn is finite for every n ≥ m. Since the sets Fn are nonempty and
nested, this clearly implies that some point belongs to all of the sets. Thus
we may suppose that each set Fn is infinite.
An infinite sequence of distinct points xn will now be constructed induc-
tively as follows. Let x1 be any point in F1 . Choose x2 ∈ F2 with x2 = x1 .
At the nth stage of the construction, having chosen distinct points xk ∈ Fk
for k = 1, 2, . . . , n − 1, choose xn ∈ Fn different from all of its predecessors.
Since every set Fn is infinite, it is always possible to make such a choice. Let
S = {xn : n ∈ N} be the set of all points chosen in this process. Note that
S is a bounded set, since S ⊂ F1 and F1 is bounded. Also S is an infinite
set, since all of the numbers xn were chosen to be distinct. Thus by the
Bolzano–Weierstrass theorem, S has a cluster point ξ. Then ξ is a cluster
point of each set Fm , since the points xn belong to Fm for all n ≥ m. But
Fm is closed and therefore contains all of its cluster points. Hence ξ ∈ Fn
for every n ∈ N. In other words, ξ belongs to the intersection of the sets Fn ,
which is therefore not empty. This proves the nested sets theorem.
Countable Basis Theorem. There exists a countable collection of bounded

open intervals I1 , I2 , · · · ⊂ R with the following property. For each x ∈ R
and for each open set G ⊂ R with x ∈ G, there is an interval In such that
x ∈ In ⊂ G.
Proof. Consider the family F of all open neighborhoods
Nr (p) = {x ∈ R : |x − p| < r}
with rational centers p and rational radii r. In other words,
F = {Nr (p) : p ∈ Q , r ∈ Q} .
The family F is in one-to-one correspondence with the Cartesian product

Q × Q and is therefore countable. It can be enumerated as I1 , I2 , . . . . Now
if x belongs to an open set G, then by the definition of open set, Nε (x) ⊂ G
for some ε > 0. But then by the density of the rational numbers, there exist
p, r ∈ Q with x ∈ Nr (p) ⊂ Nε (x). Thus x ∈ In ⊂ G for some In ∈ F, as
required.
In order to state the Heine–Borel theorem, we need to make some def-

initions. We say that a collection C of subsets of R is a covering of a set
S ⊂ R if S is contained in the union of all of the sets in C. Equivalently, C
is a covering of S if each point x ∈ S belongs to some set in the collection C.
An open covering is a covering that consists of open sets. A finite covering
is a covering that consists of finitely many sets. A subcovering of a covering
C is a subcollection of the sets in C (in other words, a subset of C) which
again constitutes a covering of the set S.
Heine–Borel Theorem. Every open covering of a closed bounded set of

real numbers contains a finite subcovering.
The theorem says that whenever a closed bounded set of real numbers
is contained in the union of some collection of open sets, then it is actually
contained in the union of finitely many of those sets. The statement be-
comes false if either of the adjectives “closed” or “bounded” is omitted. For
instance, the collection
C = {( n1 , 1) : n ∈ N}

is an open covering of the open set (0, 1), and in fact ∞ 1
n=1 ( n , 1) = (0, 1), yet
no finite subcollection of C is a covering of (0, 1). To give another example,
the collection
C = {(−n, n) : n ∈ N}
is an open covering of any set S ⊂ R, but no finite subcollection of C can be
a covering of any unbounded set S.
Proof of theorem. We will show first that every countable open covering
of a closed bounded set contains a finite subcovering. Thus we assume that
a closed bounded set S is contained in the union of a countable family of
open sets G1 , G2 , . . . . In other words,
∞

S⊂ Gk .
k=1
n
We want to show that S ⊂ k=1 Gk for some n ∈ N. Let Hn = nk=1 Gk .
Then each set Hn is open and we see that H1 ⊂ H2 ⊂ H3 ⊂ . . . and S ⊂
∞

n=1 Hn . Now define the sets Fn = Hn ∩ S , where Hn is the complement

of Hn . Then each set Fn is closed and bounded, and F1 ⊃ F2 ⊃ F3 . . . . We
want to prove that S ⊂ Hn for some n, which is equivalent to saying that Fn
is empty. Suppose on the contrary that Fn= ∅ for every n ∈ N. Then by the
nested sets theorem there is a point x ∈ ∞ n=1 Fn . In other words, x ∈ Fn
for every n ∈ N. According to the definition of Fn , this says that
x ∈ S and
x∈ / Hn for all n, which contradicts the hypothesis that S ⊂ ∞ n=1 Hn . The
contradiction shows that in fact some set Fn is empty, or that S ⊂ Hn for

some n ∈ N, which was to be shown.
It remains to prove that a countable subcovering can be extracted from
an arbitrary open covering of S. Suppose, then, that

S⊂ Gα , where C = {Gα : α ∈ A}
α∈A
is a collection, not necessarily countable, of open sets Gα indexed by a set

A. By assumption, each point x ∈ S belongs to Gα for some α ∈ A. Choose
one of the sets Gα that contains x and write it as Gα(x) . Now let F = {In }
be the countable collection of bounded open intervals In supplied by the
countable basis theorem. For each x ∈ S, that theorem allows us to choose
an interval In = In(x) such that x ∈ In(x) ⊂ Gα(x) . Note that the collection
of sets {In(x) : x ∈ S} is a subcollection of F and is therefore a countable
covering of S. Now with each interval In in this covering we can associate a
set Gα in the original covering C such that In ⊂ Gα . This gives the required
countable subcovering of C.
To summarize the proof, a countable covering is first extracted from an

arbitrary open covering of S. In this part of the argument it is not necessary
to make any special assumptions about S. However, in the next step, where
a further finite subcovering is extracted from the countable subcovering, it
is essential to assume that S is closed and bounded.
In the general context of a metric space, any set with the Heine–Borel
property is said to be compact. Thus the content of the Heine–Borel theorem
is that a set of real numbers is compact if (and only if) it is closed and
bounded.
1.4. Numerical sequences

At this point it is convenient to discuss some standard examples of se-
quences whose limits can be found with help of the binomial theorem
n
n n(n − 1) 2
(1 + x)n = 1 + xk = 1 + nx + x + · · · + xn ,
k 2!
k=1
where n ∈ N and

n n! n(n − 1) · · · (n − k + 1)
= =
k k!(n − k)! k!
denotes a binomial coefficient. Here are the examples.

Example 1. For each fixed number r ∈ R with |r| < 1 ,
lim rn = 0 .
n→∞
Proof. Since (−r)n = (−1)n rn , it suffices to suppose that 0 < r < 1. Write
1
r = 1+a , where a > 0. Then na ≤ (1 + a)n and it follows that
1 1
0 < rn = n
≤ →0 as n → ∞ .
(1 + a) na
By the “squeeze principle”, this shows that rn → 0.
Example 2. For each constant a > 0 ,

√
n
lim a = 1.
n→∞
√ √
Proof. Since n 1/a = 1/ n a, it suffices to consider a > 1. Then n a > 1,
√
and we can write n a = 1 + hn for some hn > 0. It follows that
a
a = (1 + hn )n ≥ n hn , so that 0 < hn < → 0.
n
√
This shows that hn → 0 as n → ∞, which says that n
a → 1 as n → ∞.
Example 3.
√
n
lim n = 1.
n→∞
√
n
√
Proof. Again n > 1, so we can write n n = 1 + hn for some hn > 0.
Then
n(n − 1) 2 n(n − 1) 2
n = (1 + hn )n ≥ 1 + nhn + hn > hn , for n ≥ 2 ,
2! 2
which shows that

2
0 < hn < →0 as n → ∞ .
n−1
√
This proves that hn → 0, so that n
n → 1 as n → ∞.
Example 4. For each pair of numbers p > 0 and r with |r| < 1,
lim np rn = 0 .
n→∞
Proof. This is a stronger form of Example 1. Again it is enough to take

0 < r < 1. Let an = np rn and observe that the ratios

an+1 n+1 p
= r→r as n → ∞ .
an n
Therefore, for any fixed R in the interval r < R < 1, we can infer that
an+1
0< <R
an
for all n sufficiently large, say for all n ≥ N . Then
aN +1 < R aN , aN +2 < R aN +1 < R2 aN ,
and in general
0 < aN +k < aN Rk , k = 1, 2, . . . .
Thus aN +k → 0 as k → ∞, since 0 < R < 1 and so Rk → 0. In other words,
an → 0 as n → ∞, as claimed.
Example 5. For each number p > 0,

log n
lim = 0.
n→∞ np
Proof. Consider first what happens as n → ∞ over the subsequence {2k }.

For n = 2k , we have
log n log(2k ) k log 2
= = →0 as k → ∞ ,
np (2k )p (2p )k
by Example 4, since r = 1/2p < 1 for p > 0.
Now observe that each integer n ≥ 2 lies in an interval 2k ≤ n < 2k+1 ,
for some uniquely determined integer k = 1, 2, . . . . Thus
0 < log n < log(2k+1 ) = (k + 1) log 2 ,
whereas
np ≥ (2k )p = (2p )k .
Combining the two estimates, we see that
log n (k + 1) log 2
0< ≤ .
np (2p )k
But k → ∞ as n → ∞, and
(k + 1) log 2
lim = 0,
k→∞ (2p )k
by Example 4, again because r = 1/2p < 1 for p > 0. An application of the
“squeeze principle” completes the proof.
For an arbitrary sequence {xn } of real numbers, the upper and lower
limits
lim sup xn and lim inf xn
n→∞ n→∞
will now be defined and characterized in three equivalent ways. We will see
that
lim sup xn = lim inf xn = lim xn
n→∞ n→∞ n→∞
if the sequence {xn } converges. With suitable interpretation, the same will
be true if xn → +∞ or xn → −∞. However, the “lim sup” and “lim inf ”
are more general quantities that help to specify the eventual behavior of
divergent sequences that do not tend to +∞ or −∞.
First some conventions for unbounded sequences. If {xn } is not bounded
above, we write
lim sup xn = +∞ .
n→∞
Similarly, if {xn } is not bounded below, we write
lim inf xn = −∞ .
n→∞
If xn → −∞, we say that lim supn→∞ xn = −∞. If xn → +∞, we say that

lim inf n→∞ xn = +∞.
Suppose now that {xn } is bounded above and does not tend to −∞. Let
S denote the set of subsequential limits of {xn }. This is the set of numbers
λ for which some subsequence {xnk } = {xn1 , xn2 , xn3 , . . . } converges to λ.
We then define lim supn→∞ xn to be sup S. Observe that S is nonempty, in
view of the Bolzano–Weierstrass theorem. Indeed, if {xn } is bounded above
and does not tend to −∞, then it has a bounded subsequence, and so a
further subsequence is convergent. Similarly, if {xn } is bounded below and
does not tend to +∞, we define lim inf n→∞ xn = inf S.
It may be remarked that the number sup S actually belongs to S, so that
lim sup xn is in fact the largest subsequential limit. Similarly, lim inf xn is
the smallest subsequential limit if the sequence is bounded below and does
not tend to +∞. (See Exercise 21.)
With these definitions, we see that
lim inf xn ≤ lim sup xn

n→∞ n→∞
for every sequence {xn }, with the usual interpretation of the symbol “≤” if
either of the quantities is infinite. Equality occurs if and only if the sequence
{xn } converges or tends to +∞ or −∞. For bounded sequences, equality
occurs if and only if the set S contains only one number, which must then
be the limit of the sequence.
Another simple consequence of the definitions is that
lim sup (−xn ) = − lim inf xn ,

n→∞ n→∞
with the standard interpretation if either quantity is infinite. This relation

shows that the upper and lower limits are in some sense reflections of each
other, so that properties of one will transfer automatically to the other.
The preceding definitions of “lim sup” and “lim inf ” were given in terms
of the set S of subsequential limits. Although this description is perhaps the
easiest to comprehend, it is awkward to apply. We now give an equivalent
formulation that is better suited to manipulation.
Theorem. If a sequence {xn } is bounded above and does not tend to −∞,
then the number
β = lim sup xn = sup S
n→∞
is characterized by the following property. For each ε > 0,

(i) xn < β + ε for all but a finite number of indices n
and
(ii) xn > β − ε for infinitely many indices n .
Similarly, if {xn } is bounded below and does not tend to +∞, the number
α = lim inf xn = inf S

n→∞
is characterized by the property that for each ε > 0,

(i) xn > α − ε for all but a finite number of indices n
and
(ii) xn < α + ε for infinitely many indices n .
Proof. We need only discuss the description of “lim sup”. Let us show
first that β = lim supn→∞ xn satisfies the conditions (i) and (ii). If (i)
fails for some ε > 0, then the inequality xn ≥ β + ε holds for infinitely
many indices n. This implies that xnk ≥ β + ε for some subsequence {xnk }.
But by hypothesis {xn } is bounded above, so the subsequence {xnk } is also
bounded above. Hence by the Bolzano–Weierstrass theorem it has a further
subsequence that converges to a limit λ ≥ β + ε. Thus λ belongs to the
set S of subsequential limits of {xn } and λ > β, which contradicts the fact
that β = sup S is an upper bound for S. This shows that β satisfies (i).
If (ii) fails to hold for some ε > 0, then xn ≤ β − ε for all but a finite
number of indices n, and so β − ε is an upper bound for S, which violates
the definition of β as the least upper bound of S. Thus we have proved that
β = lim supn→∞ xn satisfies both (i) and (ii). Conversely, it is clear that at
most one number β can satisfy both (i) and (ii) for each ε > 0, so that a
number with this property must be lim supn→∞ xn .
The proof for “lim inf ” can be carried out in a similar way, or the result
can be deduced from the relation lim inf n→∞ xn = − lim supn→∞ (−xn ) .
The condition (i) for “lim sup” is equivalent to saying that for each
ε > 0 there exists a number N (depending on ε) such that xn < β + ε for
all n ≥ N . A similar remark applies to “lim inf ”.
For a third formulation of upper and lower limits, one can show that

lim sup xn = lim sup xk and lim inf xn = lim inf xk ,
n→∞ n→∞ k≥n n→∞ n→∞ k≥n
with appropriate interpretation for unbounded sequences. To see that the

limits exist for bounded sequences {xn }, let
yn = sup xk and zn = inf xk ,
k≥n k≥n
and observe that yn+1 ≤ yn , whereas zn+1 ≥ zn . Thus both limits exist by
the monotone boundedness theorem. It is left as an exercise to identify the
limits as lim sup xn and lim inf xn , respectively.
The following theorem is useful in the study of infinite series.
Theorem. For any sequence {rn } of positive numbers,
√ rn+1
(i) lim sup n rn ≤ lim sup
n→∞ n→∞ rn
and
rn+1 √
(ii) lim inf ≤ lim inf n rn .
n→∞ rn n→∞
Proof. We will treat only the inequality (i). (The proof of the inequal-
ity (ii) is similar and is left as an exercise.) We may suppose that β =
lim supn→∞ rn+1 /rn is finite. Then for each ε > 0 there is a number N such
that 0 < rn+1 /rn < β + ε for all n ≥ N . For n > N it follows that
rN +1 rN +2 rn
rn = rN · · ··· < rN (β + ε)n−N , and so
rN rN +1 rn−1

√n
rn < n rN (β + ε)−N (β + ε) < β + 2ε
√
for all n sufficiently large, since n a → 1 for each fixed a > 0. It follows that
√
lim sup n rn ≤ β + 2ε for each ε > 0 ,
n→∞
√
which implies that lim supn→∞ n rn ≤ β.
1.5. Infinite series

A formal infinite series ∞ k=1 ak is said
n to be convergent to a sum s if
its sequence {sn } of partial sums sn = k=1 ak converges to s. One very

important example is the geometric series ∞ k
k=0 r , with partial sums

n
1 − rn+1
sn = rk = , r = 1 .
1−r
k=0
The series converges if and only if |r| < 1, in which case its sum is s = 1
1−r .
If a series converges, then an = sn − sn−1 → s − s = 0 as n → ∞. In
other words, a necessary condition for convergence of an infinite series ak
is that its general term ak tend to zero as k → ∞. However, this condition
is far from sufficient.
Infinite series with terms of constant sign are easiest to analyze. If ak ≥ 0
for all k, then sn ≤ sn+1 and the partial sums form a nondecreasing sequence.
By the monotone boundedness theorem, then, the series converges if and
only if its partial sums are bounded above. For an example, let ak = 1/k
and observe that

1 1 1 1 1 1 1
s2m = 1 + + + + + + + + ...
2 3 4 5 6 7 8

1 1
+ + ··· + m
2m−1 + 1 2
1 1 1 1 m
≥ 1 + + 2 · + 4 · + · · · + 2m−1 · m = 1 + →∞
2 4 8 2 2

as m → ∞, which shows that the series ∞ k=1 1/k diverges. On the other
hand, if ak = 1/k 2 , then

1 1 1 1 1 1
s2m = 1 + + + + + + + ...
22 32 42 52 62 72

1 1 1
+ + ··· + m + 2m
2 2(m−1) (2 − 1) 2 2
1 1 1 1
≤ 1 + 2 · 2 + 4 · 2 + · · · + 2m−1 · 2(m−1) + 2m
2 4 2 2
1 1 1 1
= 1 + + + · · · + m−1 + 2m < 2 ,
2 4 2 2
∞
so the sequence {sn } of partial sums is bounded and the series ∞ k=1p 1/k
2
converges. In similar manner it can be seen that the series k=1 1/k con-
verges if p > 1 and diverges if p ≤ 1.
Series with positive decreasing terms are often analyzed most easily by
comparison with an integral. To be more specific, suppose f (x) is a positive
continuous nonincreasing function on the interval [1, ∞). Then by compar-

ison of areas it is apparent that

n n
n−1
f (k) ≤ f (x) dx ≤ f (k) , n = 2, 3, . . . ,
k=2 1 k=1
∞
which shows that the sum ∞ k=1 f (k) and the integral 1 f (x) dx converge
or diverge together.

A series ak is said to be absolutely convergent if |ak | is convergent.
Theorem. Every absolutely convergent series is convergent.

Proof. Suppose that a series ak is absolutely convergent. Let
sn = a1 + a2 + · · · + an and tn = |a1 | + |a2 | + · · · + |an | .
By hypothesis, the sequence {tn } is convergent, and so it is a Cauchy se-

quence. Thus for each ε > 0 there exists a number N such that |tn − tm | < ε
whenever n > m ≥ N . But
|sn − sm | = |am+1 + am+2 + · · · + an | ≤ |am+1 | + |am+2 | + · · · + |an | = tn − tm ,
} is also a Cauchy sequence, hence is convergent. This

so it follows that {sn
says that the series ak is convergent, as claimed.
The theorem just proved is essentially equivalent to the comparison

test
for convergence ofan infinite series. Specifically, if |ak | ≤ bk and bk is
convergent, then ak is (absolutely) convergent.
Comparison with a geometric series is facilitated by two special devices
known as the ratio test and the root test.

Ratio Test. If ak = 0 for all k, the infinite series ak is convergent if

ak+1 ak+1

lim sup <1 and divergent if
lim inf > 1.
k→∞ ak k→∞ ak

Root Test. The infinite series ak is convergent if

lim sup k
|ak | < 1 and divergent if lim sup k
|ak | > 1 .
k→∞ k→∞
Proofs. In view of the theorem at the end of Section 1.4, the validity of
the ratio test follows from that of the root test. More specifically, the con-
dition for convergence in the ratio test implies that in the root test, and the
conditions for divergence are similarly related. Therefore, it will suffice to
prove the validity of the root test.

Let β = lim supk→∞ k |ak | and suppose that β < 1. Then for any
number r with β < r < 1 there is an index N such that k |ak | ≤ r for all k ≥

N . Thus |ak | ≤ rk for all k ≥ N and the series ak is absolutely
kconvergent,
hence convergent, by comparison with the geometricseries r . If β > 1,
choose any number ρ with 1 < ρ < β and note that |ak | > ρ for infinitely
k
many indices k, by the definition of “lim sup”. Consequently, |ak | > ρk → ∞

as k →∞ through some subsequence. In particular, the general term of the
series ak does not tend to zero, so the series is divergent.
Although the root test is more powerful than the ratio test, the latter is
often more convenient to apply, especially when factorials are involved.
The question of convergence becomes more delicate when the series is
not absolutely convergent. The basic theorem on series with alternating
signs goes back to Leibniz and is a useful test for convergence. Gottfried
Wilhelm Leibniz (1646–1716) is known as a founder of calculus, along with
Isaac Newton (1643–1727).
Leibniz’ Alternating Series Theorem. Suppose that a1 ≥ a2 ≥ a3 ≥
· · · ≥ 0 and limk→∞ ak = 0 . Then the infinite series
∞

(−1)k+1 ak = a1 − a2 + a3 − a4 + . . .
k=1
is convergent.
Proof. Let

n
sn = (−1)k+1 ak
k=1
denote the partial sums. The strategy is to apply the monotonicity to show
that each of the subsequences {s2n } and {s2n+1 } is convergent, then to use
the hypothesis that ak → 0 to see that the two limits agree.
Observe first that
s2n+2 = s2n + (a2n+1 − a2n+2 ) ≥ s2n ,
since a2n+1 ≥ a2n+2 by hypothesis. This shows that
s2 ≤ s4 ≤ s6 ≤ s8 ≤ . . . .
Similarly,
s2n+3 = s2n+1 − (a2n+2 − a2n+3 ) ≤ s2n+1 ,
so that
s1 ≥ s3 ≥ s5 ≥ s7 ≥ . . . .
In other words, the sequence {s2n } is nondecreasing, whereas the sequence
{s2n+1 } is nonincreasing. To see that both subsequences are bounded, note
that
s2 ≤ s2n ≤ s2n + a2n+1 = s2n+1 ≤ s1 .
Therefore, by the monotone boundedness theorem, each of the subsequences
{s2n } and {s2n+1 } is convergent. Let
s = lim s2n and t = lim s2n+1 .

n→∞ n→∞
Now invoke the hypothesis that ak → 0 to show that the two limits agree:
t = lim s2n+1 = lim (s2n + a2n+1 ) = s + 0 = s .

n→∞ n→∞
From this it is easy to deduce that the full sequence {sn } converges to the
same limit s.
As a simple application of the theorem, we can infer that the infinite

series
1 − 12 + 13 − 14 + . . .
is convergent. We will see later, in several different ways, that its sum is
log 2. This is an example of a conditionally convergent series, convergent
but not absolutely convergent.
We now record a less familiar result known as Cauchy’s double series
theorem, an important tool that will be applied to power series in Chapter
3. By way of orientation, recall that an area integral of a function of two
real variables can be iterated in two ways, and an interchange of the order
of integration is often productive, although it is not always permissible and
must be justified. Cauchy’s theorem deals with a discrete analogue, a double
series
∞ ∞
anm ,
n=1 m=1
where the anm are real numbers. Such a double ∞series is said to con-
verge to the sum s if for each fixed n the series m=1 anm converges to
asum An and ∞ n=1 A n = s. The double series is absolutely convergent if
∞ ∞
n=1 |a
m=1 nm | is convergent.
Cauchy’s Double Series Theorem. If either of the double series
∞
∞ ∞
∞
anm or anm
n=1 m=1 m=1 n=1
is absolutely convergent, then both are convergent and their sums are equal.
Without the hypothesis of absolute convergence, the statement is false.

Even if both series converge, their sums need not be equal (cf. Exercise 34).
Proof of theorem.
∞ Suppose first that anm ≥ 0 for all n, m = 1, 2, . . . , and
∞
assume
∞ that n=1 m=1 anm is (absolutely)
∞ convergent. In other words,
m=1 anm = An < ∞ for each n, and n=1 An converges. Then anm ≤ An
and so
∞ ∞

Bm = anm ≤ An < ∞ for each m .
n=1 n=1
∞
Also, an1 +an2 +· · ·+anm ≤ An , so that B1 +B2 +· · ·+Bm ≤ n=1 An < ∞.
It follows that
∞
∞ ∞
∞
∞
∞
B= anm = Bm ≤ An = anm = A .
m=1 n=1 m=1 n=1 n=1 m=1
In particular, the left-hand double series is convergent, so by interchanging

the roles of n and m we can deduce by a similar argument that A ≤ B.
Hence A = B and the theorem is proved in the special case where all of the
numbers anm have the same sign.
In the general case where the numbers anm have variable signs, we write
anm = αnm + βnm , where αnm = anm if anm > 0 and αnm = 0 otherwise.
Thus αnm ≥ 0 and βnm ≤ 0, so we infer from the special case already proved
that
∞
∞ ∞
∞ ∞
∞
anm = αnm + βnm
n=1 m=1 n=1 m=1 n=1 m=1
∞ ∞ ∞ ∞ ∞
∞
= αnm + βnm = anm ,
m=1 n=1 m=1 n=1 m=1 n=1
as claimed.
1.6. Continuous functions and derivatives

A real-valued function f defined on a set S ⊂ R is said to be continuous
at a point x0 ∈ S if for each ε > 0 there exists a number δ > 0 such that
|f (x) − f (x0 )| < ε for all x ∈ S with |x − x0 | < δ. It is equivalent to require

that f (xn ) → f (x0 ) for each sequence of points xn ∈ S such that xn → x0 .
A function is said to be continuous on the set S if it is continuous at each
point of S.
It is easy to prove that if f and g are defined on a common set S and
are continuous at a point x0 ∈ S, then their sum f + g and product f g are
continuous at x0 , and the quotient f /g is continuous there if g(x0 ) = 0. This
allows us to conclude that every polynomial is continuous on R, since it is
trivially true that the function f (x) = x is continuous. More generally, every
rational function (the quotient of two polynomials) is continuous except at
zeros of the denominator.
An important property of continuous functions is that they map con-
nected sets to connected sets. To put it in more prosaic terms, if a con-
tinuous function on an interval assumes two different values, then it must
assume every value in between. Although this fact is intuitively obvious,
the proof is not trivial.
Intermediate Value Theorem. If a function f is continuous on an

interval [a, b] and if f (a) < λ < f (b), then f (ξ) = λ for some ξ ∈ (a, b).
Of course, the same conclusion holds if f (b) < λ < f (a).
Proof of theorem. Let S = {x ∈ [a, b] : f (x) < λ} and define ξ = sup S.

Then ξ ∈ (a, b) since f (x) < λ in some interval [a, a + δ) and f (x) > λ
in some interval (b − δ, b]. If f (ξ) < λ, then f (x) < λ throughout some
neighborhood of ξ, which contradicts the definition of ξ as an upper bound
of S. If f (ξ) > λ, then f (x) > λ in some neighborhood of ξ, and so ξ cannot
be the smallest upper bound of S. Hence f (ξ) = λ.
A function f is bounded above on a set S if its range of values f (S) =

{f (x) : x ∈ S} has an upper bound. Similarly, f is bounded below on S if
f (S) has a lower bound. The function is said to be bounded on S if it is
bounded above and below; that is, if |f (x)| ≤ B for some constant B and all
x ∈ S. We say that f has a maximum at a point x0 ∈ S if f (x) ≤ f (x0 ) for
all x ∈ S. Similarly, f has a minimum at x1 if f (x) ≥ f (x1 ) for all x ∈ S.
Theorem. If a function is continuous on a closed bounded set S ⊂ R, then

it attains a maximum and a minimum on S.
Proof. Let f be continuous on S, and let M = sup{f (x) : x ∈ S}, where

M = +∞ if f is not bounded above on S. Let {xn } be a sequence of points
in S for which f (xn ) → M as n → ∞. By the Bolzano–Weierstrass theorem,
some subsequence {xnk } converges to a point x0 ∈ S. Since f is continuous
at x0 , it follows that f (xnk ) → f (x0 ) as k → ∞. But f (xnk ) → M , so this

proves that M = f (x0 ) < ∞. Hence f attains a maximum value at x0 . A
similar argument shows that f attains a minimum on S.
A function f defined on a set S ⊂ R is said to be uniformly continuous

on S if for each ε > 0 there exists a δ > 0 such that |f (x) − f (t)| < ε for
all pairs of points x, t ∈ S with |x − t| < δ. Every uniformly continuous
function is pointwise continuous, but the converse is false. For instance,
the function f (x) = 1/x is continuous at each point of the interval (0, 1]
but is not uniformly continuous there. Similarly, the function f (x) = x2 is
pointwise continuous but not uniformly continuous on [0, ∞). However, the
converse is true if the set S is closed and bounded, as we now proceed to
prove.
Theorem. If a function is continuous at each point of a closed bounded set

S ⊂ R, then it is uniformly continuous on S.
Proof. The result can be obtained with help of the Bolzano–Weierstrass

theorem, but it is more natural to base a proof on the Heine–Borel theorem.
Suppose a function f is pointwise continuous on the closed bounded set S.
Given ε > 0, we want to produce a number δ > 0 such that |f (x) − f (t)| < ε
for all pairs of points x, t ∈ S with |x − t| < δ. The continuity of f at each
point x ∈ S says that there is a number δ(x) > 0 depending on x such that
ε
|f (t) − f (x)| < for all t ∈ S ∩ I(x) ,
2
where I(x) is the interval (x − δ(x), x + δ(x)). Define the smaller interval
J(x) = (x − 12 δ(x), x + 12 δ(x)) ⊂ I(x) .
The collection C = {J(x) : x ∈ S} of all such intervals J(x) is an open

covering of S, since each x ∈ S belongs to its associated set J(x). Thus
by the Heine–Borel theorem, C contains a finite subcovering of S. In other
words, there is a finite set of points x1 , x2 , . . . , xn ∈ S for which
S ⊂ J(x1 ) ∪ J(x2 ) ∪ · · · ∪ J(xn ) .
Now set δ = 12 min{δ(x1 ), δ(x2 ), . . . , δ(xn )}, and observe that δ > 0. Choose
arbitrary points x, t ∈ S with |x − t| < δ. Then x ∈ J(xk ) for some k, since
the sets J(x1 ), . . . , J(xn ) form a covering of S. Hence x ∈ I(xk ). But the
point t also belongs to the interval I(xk ), in view of the inequality
|t − xk | ≤ |t − x| + |x − xk | < δ + 12 δ(xk ) ≤ δ(xk ) .

Because the points x and t belong to the same interval I(xk ), we conclude
that
ε ε
|f (x) − f (t)| ≤ |f (x) − f (xk )| + |f (xk ) − f (t)| < + = ε,
2 2
as desired. This proves that f is uniformly continuous on S.
Now let f (x) be a real-valued function defined on an arbitrary set S ⊂ R,

and suppose that a is a cluster point of S. Then f (x) is said to have a limit
L at a if for each ε > 0 there exists a δ > 0 such that |f (x) − L| < ε for all
x ∈ S with 0 < |x − a| < δ. This is indicated by writing
lim f (x) = L or f (x) → L as x → a .

x→a
It is equivalent to require that f (xn ) → L for every sequence of points

xn ∈ S such that xn → a as n → ∞.
It is sometimes convenient to speak of one-sided limits. If a is a right-
hand cluster point of S, meaning that every neighborhood of a contains a
point x ∈ S with x > a, then the function f (x) has a right-hand limit L if
for each ε > 0 there is a δ > 0 such that |f (x) − L| < ε for all x ∈ S with
0 < x − a < δ. This is indicated by writing
lim f (x) = L or f (x) → L as x → a + .

x→a+
Similarly, if the point a is a left-hand cluster point of S, then f (x) is said

to have a left-hand limit L, written
lim f (x) = L or f (x) → L as x → a − .

x→a−
if to each ε > 0 there corresponds a δ > 0 such that |f (x) − L| < ε for all
x ∈ S with 0 < a − x < δ.
A limit of special importance is the derivative
f (x) − f (x0 )
f (x0 ) = lim ,
x→x0 x − x0
where f is a function defined in an open interval containing the point x0 .
If the derivative exists, the function f is said to be differentiable at the
point x0 . A function is said to be differentiable on an open set S if it has
a derivative at each point of S. One-sided derivatives can be defined as
one-sided limits of the difference quotient.
The familiar rules of differentiation are easily deduced. The derivative
of the sum of two differentiable functions is the sum of their derivatives,
the product f (x)g(x) has derivative f (x)g(x) + f (x)g (x), and the quotient
f (x)/g(x) has derivative
f (x)g(x) − f (x)g (x)
g(x)2
wherever g(x) = 0. The “chain rule” for differentiation of the composition
h(x) = f (g(x)) is more subtle but is found to be h (x) = f (g(x))g (x).
Every student of calculus learns that the maximum and minimum of
a function are likely to occur at a critical point; that is, at a point where
the derivative vanishes. This principle has a more precise statement, and
it applies more generally to local maxima and minima. If a function f is
defined on an open set S containing a point x0 , we say that it has a local
maximum at x0 if f (x) ≤ f (x0 ) for all x in some neighborhood
Nδ (x0 ) = {x ∈ R : |x − x0 | < δ} ⊂ S .
A local minimum is defined similarly. Here is the precise statement.
Theorem. If a function f is defined on an open set S and has a local
maximum or local minimum at a point x0 ∈ S, and if f is differentiable at
x0 , then f (x0 ) = 0.
Proof. It suffices to consider a local minimum. By hypothesis, f (x) ≥
f (x0 ) for all x in some neighborhood Nδ (x0 ). Consequently, the difference
quotient
f (x) − f (x0 )
g(x) = , x = x0 ,
x − x0
defined in Nδ (x0 ) except at x0 , satisfies the inequalities g(x) ≥ 0 for x > x0
and g(x) ≤ 0 for x < x0 . Since f is differentiable at x0 , it follows that
f (x0 ) = lim g(x) ≥ 0 and f (x0 ) = lim g(x) ≤ 0 .
x→x0 + x→x0 −
Hence f (x 0) = 0, as claimed.
It is now a short step to a time-honored result.

Rolle’s Theorem. If a function f is continuous in an interval [a, b] and
differentiable in the open interval (a, b), and if f (a) = f (b), then f (ξ) = 0
for some point ξ ∈ (a, b).
Proof. If f is constant in [a, b], then f (x) ≡ 0 in (a, b) and the conclusion
is trivial. If f is not constant, then f attains a maximum or a minimum
at some point ξ ∈ (a, b), and it follows from the previous theorem that
f (ξ) = 0.
Rolle’s theorem can be generalized as follows.

Mean Value Theorem. If a function f is continuous in an interval [a, b]

and differentiable in the open interval (a, b), then
f (b) − f (a)
f (ξ) = for some ξ ∈ (a, b) .
b−a
Proof. Define the linear function

f (b) − f (a)
L(x) = (x − a) + f (a) ,
b−a
so that L(a) = f (a) and L(b) = f (b). Then apply Rolle’s theorem to the
function g(x) = f (x) − L(x).
The mean value theorem can be applied to show that a function with
zero derivative throughout an interval must be constant. More generally, if
the derivatives of two functions coincide in some interval, the two functions
must differ by a constant. Another corollary of the mean value theorem is
that a function with positive derivative in an interval is strictly increasing.
Similarly, a function with negative derivative is strictly decreasing.
1.7. The Riemann integral

Loosely speaking, the integral of a positive function is the area under its
graph. The problem is to make that notion precise, and to extend it to a
wider class of functions. Such an effort leads naturally to the definition of the
Riemann integral. It is often possible to calculate an integral directly from
the definition, but the most effective method comes via the fundamental
theorem of calculus, which allows calculation of an integral by a process of
antidifferentiation. In fact, the deepest results emerge when both methods
are applied to the same problem, as in the derivation of Stirling’s formula
to be given in the next chapter.
For a formal development of the Riemann integral, consider a function
f defined and bounded on an interval [a, b]. Let P be an arbitrary partition
a = x0 < x1 < x2 < · · · < xn = b of the interval [a, b]. For k = 1, 2, . . . , n,
let
Mk = sup {f (x) : xk−1 ≤ x ≤ xk } and mk = inf {f (x) : xk−1 ≤ x ≤ xk } ,
and form the upper and lower sums

n
n
U (f, P ) = Mk (xk − xk−1 ) and L(f, P ) = mk (xk − xk−1 )
k=1 k=1
of f with respect to the partition P . Clearly, L(f, P ) ≤ U (f, P ) for every

partition P . If the partition P is refined to a partition Q ⊃ P by ad-
joining extra points of subdivision, it can be seen that L(f, P ) ≤ L(f, Q)
and U (f, Q) ≤ U (f, P ). In particular, if Q is a common refinement of two
partitions P1 and P2 , then
L(f, P1 ) ≤ L(f, Q) ≤ U (f, Q) ≤ U (f, P2 ) .
Thus L(f, P1 ) ≤ U (f, P2 ) for every pair of partitions P1 and P2 , and so it

follows that
sup L(f, P ) ≤ inf U (f, P ) .
P P
In other words, the supremum of the lower sums taken over all partitions is
less than or equal to the infimum of the upper sums. If equality occurs, the
function f is said to be Riemann integrable over the interval [a, b], and its
integral is defined to be the common value, written as
b
f (x) dx = sup L(f, P ) = inf U (f, P ) .
a P P
Equivalently, we can say that f is Riemann integrable if and only if for each
ε > 0 there exists a partition P such that U (f, P ) − L(f, P ) < ε.
Not every bounded function is Riemann integrable. Consider, for in-
stance, the function defined by f (x) = 1 if x is rational and f (x) = 0 if x
is irrational. Then L(f, P ) = 0 and U (f, P ) = b − a for every partition P ,
since both the rationals and the irrationals are everywhere dense. Thus f is
not Riemann integrable.
On the other hand, every continuous function on an interval [a, b] is
integrable. To see this, we apply the theorem (proved in the previous section)
that every continuous function on a closed bounded interval is uniformly
continuous there. Thus the pointwise continuity of a function f implies that
for each ε > 0 there is a δ > 0 such that |f (x) − f (t)| < ε/(b − a) for all
pairs of points x, t ∈ [a, b] with |x − t| < δ. Now choose a partition P with n
points and all subintervals of length less than δ. Then Mk − mk ≤ ε/(b − a)
and so
n
U (f, P ) − L(f, P ) = (Mk − mk )(xk − xk−1 ) ≤ ε ,
k=1
which shows that f is integrable over [a, b].

The argument can be modified to show that f is integrable if it is contin-
uous except for a finite number of jump discontinuities. It can also be seen
that every monotonic bounded function is integrable. Suppose, for instance,
that f is nondecreasing and |f (x)| ≤ B on [a, b]. Given ε > 0, choose a

partition P comprised of n subintervals of length less than ε/2B. Then

n
ε
U (f, P )−L(f, P ) = f (xk )−f (xk−1 ) (xk −xk−1 ) < f (b)−f (a) ≤ ε .
2B
k=1
Hence f is integrable over [a, b].

Standard properties of the integral can now be derived. For instance, if
f and g are integrable over [a, b], then so is c1 f + c2 g for any constants c1
and c2 , and
b b b

c1 f (x) + c2 g(x) dx = c1 f (x) dx + c2 g(x) dx .
a a a
The product f g is also integrable. If f and g are integrable and f (x) ≤ g(x)
on [a, b], then
b b
f (x) dx ≤ g(x) dx .
a a
In particular, the inequality
b b

f (x) dx ≤ |f (x)| dx

a a
holds whenever f is integrable. (Here the main difficulty is to show that |f |

is integrable.) If a < c < b, then
b c b
f (x) dx = f (x) dx + f (x) dx .
a a c
Proofs are omitted here, but can be found in introductory texts such as Ross
[7].
It is often possible to calculate an integral directly from the definition.
The technique was known to Archimedes, who essentially applied a limiting
process to calculate the volumes of geometric solids such as the cone and
sphere. As a simple illustration, let us calculate the integral of f (x) = x2
over the interval [0, 1]. Let Pn denote the partition of [0, 1] with n subinter-
vals of equal length 1/n. Then
n
1 k − 1 2 (n − 1)(2n − 1)
n
L(f, Pn ) = f (xk−1 )(xk − xk−1 ) = =
n n 6n2
k=1 k=1
and
2
1
n n
k (n + 1)(2n + 1)
U (f, Pn ) = f (xk )(xk − xk−1 ) = = ,
n n 6n2
k=1 k=1
where the formula for the sum of squares (cf. Exercise 1) has been invoked.
Since both L(f, Pn ) and U (f, Pn ) converge to 1/3 as n → ∞, we conclude
1
that 0 x2 dx = 1/3.
Riemann’s original definition of the integral was based not on upper
and lower sums but on “Riemann sums”, as they are now called. Given a
partition P of [a, b] in the form a = x0 < x1 < x2 < · · · < xn = b, select an
arbitrary point ξk in each subinterval [xk−1 , xk ] and let ξ = (ξ1 , ξ2 , . . . , ξn ).
Then the corresponding Riemann sum is

n
S(f, P, ξ) = f (ξk )(xk − xk−1 ) .
k=1
It is clear that L(f, P ) ≤ S(f, P, ξ) ≤ U (f, P ) for every partition P and

for each choice ξ of intermediate points. With P = maxk (xk − xk−1 )
denoting the norm of the partition P , it can be shown that the Riemann
sums converge to the integral as P → 0. More precisely, we have the
following theorem (proof omitted).
Theorem. A bounded function f is Riemann integrable on an interval [a, b]

if and only if there is a number I with the following property. For each ε > 0
there exists a δ > 0 such that |S(f, P, ξ) − I| < ε for each partition P with
P < δ and all choices ξ = (ξ1 , ξ2 , . . . , ξn ) of intermediate points. If such
b
a number I exists, then I = a f (x)dx.
The intermediate value theorem (cf. Section 1.6) for continuous func-
tions can be applied to establish a corresponding theorem for integrals.
Intermediate Value Theorem for Integrals. If a function f is con-

tinuous on an interval [a, b], then
b
1
f (x) dx = f (x0 ) for some x0 ∈ [a, b] .
b−a a
Proof. As previously shown (cf. Section 1.6), the continuous function f

attains a minimum value m and a maximum value M on the interval [a, b].
Since m ≤ f (x) ≤ M for all x ∈ [a, b], it follows that
b
1
m≤ f (x) dx ≤ M .
b−a a
An appeal to the intermediate value theorem for continuous functions now

gives the desired result.
We now turn to the fundamental theorem of calculus, which says that

integration and differentiation are inverse processes. It is convenient to
adopt the convention that
a b
f (x) dx = − f (x) dx .
b a
Fundamental Theorem of Calculus.

(a) Let f be an integrable function on an interval [a, b], and define
x
F (x) = f (t) dt , a ≤ x ≤ b.
a
Then F is continuous on [a, b]. Furthermore, F is differentiable at each

point x0 ∈ (a, b) where f is continuous, and F (x0 ) = f (x0 ).
(b) Let g be a function continuous on [a, b] and differentiable on (a, b),

and suppose that g is integrable on [a, b]. Then
b
g (x) dx = g(b) − g(a) .
a
Part (b) of the theorem provides the standard device of antidifferentia-

tion to compute an integral. It is also the basis for the well-known formula
of integration by parts:
b b

u(x)v (x) dx = u(b)v(b) − u(a)v(a) − v(x)u (x) dx .
a a
Proof of theorem. (a). By definition, an integrable function is bounded.

Suppose that |f (x)| ≤ B for all x ∈ [a, b]. Then
y

|F (y) − F (x)| = f (t) dt ≤ B|y − x| , x, y ∈ [a, b] ,
x
which shows that F is (uniformly) continuous in [a, b]. For x0 ∈ (a, b), write
x
F (x) − F (x0 ) 1
−f (x0 ) = f (t)−f (x0 ) dt , x ∈ (a, b), x = x0 .
x − x0 x − x0 x0
Let ε > 0 be given. If f is continuous at x0 , then for some δ > 0 the
inequality |f (x) − f (x0 )| < ε holds for all x ∈ [a, b] with |x − x0 | < δ.
Therefore,

F (x) − F (x0 )
− f (x0 ) < ε if 0 < |x − x0 | < δ ,
x − x0
which proves that F is differentiable at x0 and F (x0 ) = f (x0 ).
(b). Since g is integrable, for each ε > 0 there is a partition P such that
U (g , P ) − L(g , P ) < ε .
On the other hand, an application of the mean value theorem to the subin-
terval [xk−1 , xk ] of this partition P gives
g (ξk )(xk − xk−1 ) = g(xk ) − g(xk−1 ) for some ξk ∈ (xk−1 , xk ) .
Addition then produces the Riemann sum

n
n

S(g , P, ξ) = g (ξk )(xk − xk−1 ) = g(xk ) − g(xk−1 ) = g(b) − g(a) .
k=1 k=1
Since L(g , P ) ≤ S(g , P, ξ) ≤ U (g , P ), we see that both of the numbers

b
[g(b) − g(a)] and a g (x)dx lie between L(g , P ) and U (g , P ), so that
b

g
(x) dx − g(b) − g(a) < ε.

a
Because ε > 0 was prescribed arbitrarily, this gives the desired result.
Integrals over unbounded intervals are defined by limiting processes. If

a function f is bounded and Riemann integrable over each bounded interval
[a, b], we define
∞ b b b
f (x) dx = lim f (x) dx and f (x) dx = lim f (x) dx ,
a b→∞ a −∞ a→−∞ a
provided the limits exist. Then we define

∞ ∞ b
f (x) dx = lim f (x) dx = lim f (x) dx ,
−∞ a→−∞ a b→∞ −∞
assuming again that the relevant limits exist.

Integrals are defined in a similar way for integrands with a singularity at
an endpoint of the interval of integration. For instance, if f is defined in a
half-open interval (a, b] and is bounded and integrable over each subinterval
[c, b] for a < c < b, we define
b b
f (x) dx = lim f (x) dx
a c→a+ c
if the limit exists. For example,

1
1 √ 1
√ dx = lim 2 x c = 2 .
0 x c→0+
1.8. Uniform convergence

Suppose that f1 , f2 , . . . are functions defined on a set S ⊂ R. The se-
quence {fn (x)} is said to be pointwise convergent on S to a function f (x) if
for each point x0 ∈ S the numerical sequence {fn (x0 )} converges to f (x0 ).
Simple examples show that desirable properties of the functions fn such as
continuity or integrability need not be preserved under pointwise conver-
gence. However, under a stronger mode of convergence known as uniform
convergence, the situation is more satisfactory.
By definition, a sequence of functions {fn } is uniformly convergent to f
on a set S if for each ε > 0 there exists a number N such that |fn (x)−f (x)| <
ε for all x ∈ S whenever n ≥ N . Uniform convergence implies pointwise
convergence, but it imposes the stronger requirement that the number N
depend only on ε and not on any particular point x ∈ S.
Some examples will illustrate the distinction. The functions fn (x) = x/n
converge pointwise to f (x) ≡ 0 but the convergence is not uniform on R. The
functions fn (x) = xn converge pointwise but not uniformly in the interval
[0, 1] to the limit f (x) = 0 for x ∈ [0, 1) and f (1) = 1. Note that each of
the functions xn is continuous on [0, 1] but their limit function is not. The
functions ⎧ nx ,
⎨ 0 ≤ x ≤ 1/n
gn (x) = 2 − nx , 1/n ≤ x ≤ 2/n
⎩
0, 2/n ≤ x ≤ 1
converge pointwise to g(x) ≡ 0 on [0, 1], but the convergence is not uniform.
On a more positive note, we now show that continuity is preserved under
uniform convergence.
Theorem. Suppose that the functions f1 , f2 , . . . are defined and continuous

on a set S ⊂ R, and that fn (x) → f (x) uniformly on S as n → ∞. Then f
is continuous on S.
Proof. Let ε > 0 be given. For any point x0 ∈ S, the inequality
|f (x) − f (x0 )| ≤ |f (x) − fn (x)| + |fn (x) − fn (x0 )| + |fn (x0 ) − f (x0 )|
holds for all x ∈ S. By the uniform convergence, we infer that there exists
a number N (depending only on ε) such that
ε ε
|f (x) − f (x0 )| < + |fn (x) − fn (x0 )| + for all x ∈ S
3 3
whenever n ≥ N . Choose any n ≥ N and apply the continuity of fn to
conclude that for some δ > 0 the inequality |f (x) − f (x0 )| < ε holds for all
x ∈ S with |x − x0 | < δ. Thus f is continuous at x0 .

An infinite series ∞ n=1 un (x) of functions defined on a set S ⊂ R is
understood to be uniformly
convergent on S to a sum s(x) if its sequence
of partial sums sn (x) = nk=1 uk (x) converges uniformly to s(x) on S. The
previous theorem has an equivalent expression for series.

Corollary. If an infinite series ∞ n=1 un (x) of continuous functions con-
verges uniformly on a set S, then its sum is continuous on S.
When a sequence of integrable functions fn (x) converges pointwise to
f (x) on an interval [a, b], the limit function f need not be integrable. If
b
f is integrable, it may well happen that the integrals a fn (x) dx do not
b
converge to a f (x) dx. Consider, for example, fn (x) = ngn (x), where gn (x)
is the “triangle function” defined above. The sequence {fn (x)} converges
pointwise (but not uniformly) on [0, 1] to f (x) ≡ 0, yet
1 1
fn (x) dx = 1 for every n, whereas f (x) dx = 0 .
0 0
A similar example is given in Exercise 31. The next theorem shows, however,
that integrals behave as expected under uniform convergence.
Theorem. If the functions f1 , f2 , . . . are integrable over an interval [a, b]
and fn (x) → f (x) uniformly on [a, b], then f is integrable and
b b
f (x) dx = lim fn (x) dx .
a n→∞ a
Proof. The main difficulty is to show that f is integrable. Observe that

U (f, P ) − L(f, P )
≤ |U (f, P ) − U (fn , P )| + |U (fn , P ) − L(fn , P )| + |L(fn , P ) − L(f, P )|
for any partition P of [a, b]. Let ε > 0. Then by the uniform convergence
there exists an N such that the inequalities
ε ε
|U (f, P ) − U (fn , P )| < and |L(fn , P ) − L(f, P )| <
3 3
hold for every partition P and all n ≥ N . Now for any fixed n ≥ N the
integrability of fn allows the selection of a partition P for which U (fn , P ) −
L(fn , P ) < ε/3. Hence U (f, P ) − L(f, P ) < ε, and f is integrable. Finally,
another appeal to the uniform convergence shows that
b b b

f (x) dx − f (x) dx ≤ |f (x) − fn (x)| dx < ε
n
a a a
for all n sufficiently large, which gives the desired result.

Corollary. If the functions

∞ u1 , u2 , . . . are integrable over an interval [a, b]
and the infinite series n=1 un (x) converges uniformly to s(x) on [a, b], then
s is integrable and
b ∞ b
s(x) dx = un (x) dx .
a n=1 a
On the other hand, the uniform limit of a sequence of differentiable

functions need not be differentiable, and its sequence of derivatives need not
converge. For example, the functions fn (x) = n1 sin(n2 x) converge to f (x) ≡
0 uniformly on the whole real line, but their derivatives fn (x) = n cos(n2 x)
do not even stay bounded as n → ∞. Nevertheless, we have the following
theorem.
Theorem. Let the functions f1 , f2 , . . . be continuous on an interval [a, b]
and have continuous derivatives on (a, b). Suppose that fn (x) → f (x) for
each x ∈ [a, b] and that the sequence of derivatives fn (x) is uniformly con-
vergent on (a, b). Then f is differentiable on (a, b) and fn (x) → f (x).
Proof. Let g(x) denote the uniform limit of {fn (x)}. Then by the two
preceding theorems, g is continuous on (a, b) and
x x

g(t) dt = lim fn (t) dt = lim fn (x) − fn (a) = f (x) − f (a)
a n→∞ a n→∞
for each x ∈ (a, b). Thus by the fundamental theorem of calculus, f is
differentiable and f (x) = g(x) on (a, b).
Corollary. Let the functions u1 , u2 , . . . be continuous on an interval [a, b]

and have continuous derivatives on (a, b). Suppose that the infinite se-
∞
ries
∞ n=1 un (x) converges to s(x) for each x ∈ [a, b], and that the series
u (x) is uniformly convergent on (a, b). Then s is differentiable on
n=1 n
(a, b) and s (x) = ∞
n=1 un (x).
The corollary provides a condition that allows an infinite series to be

differentiated term by term. The continuous analogue of this process, dif-
ferentiation under the integral sign, can be justified by a similar argument.
In the simplest case the integral extends over a bounded interval.
Theorem. Let f (x, y) be continuous together with its partial derivative
∂f
∂x (x, y) on a rectangle
R = [a, b] × [c, d] = {(x, y) : a ≤ x ≤ b , c ≤ y ≤ d} , a < b, c < d.
Then d d
d ∂f
f (x, y) dy = (x, y) dy .
dx c c ∂x
In particular, the left-hand integral is differentiable.
Proof. The continuity of f and ∂f∂x on the compact set R ⊂ R implies that
2
both functions are uniformly continuous there, so that the integrals

d d
∂f
Φ(x) = f (x, y) dy and Ψ(x) = (x, y) dy
c c ∂x
are continuous on the interval [a, b]. Therefore, for each point x ∈ [a, b] we
have
x x d d x
∂f ∂f
Ψ(t) dt = (t, y) dy dt = (t, y) dt dy
a a c ∂t c a ∂t
d

= f (x, y) − f (a, y) dy = Φ(x) − Φ(a) .
c
Thus by the fundamental theorem of calculus, Φ is differentiable and Φ (x) =

Ψ(x) for each x ∈ (a, b).
Essentially the same proof works for improper integrals

∞ ∞
∂f
Φ(x) = f (x, y) dy and Ψ(x) = (x, y) dy , a ≤ x ≤ b,
c c ∂x
∞
under the additional hypotheses that the integral c f (x, y) dy converges
∞
pointwise and the integral c ∂f ∂x (x, y) dy converges uniformly on [a, b]. Then
Ψ(t) is continuous
r on [a, b], since it is the uniform limit of the continuous
functions c (∂f /∂t)(t, y) dy as r → ∞. The uniform convergence also im-
plies that
x x r r x
∂f ∂f
Ψ(t) dt = lim (t, y) dy dt = lim (t, y) dt dy
a r→∞ a c ∂t r→∞ c a ∂t
= Φ(x) − Φ(a) ,
so that Φ (x) = Ψ(x). Further details and related results can be found, for
instance, in the book by Burkill and Burkill [2].
The remainder of this section is devoted to some important conditions

for uniform convergence. We begin with the observation that the Cauchy
criterion for convergence of a numerical sequence can be adapted to a cri-
terion for uniform convergence. Let f1 , f2 , . . . be functions defined on a set
S ⊂ R. Then the sequence {fn (x)} is called a uniform Cauchy sequence on
S if for each ε > 0 there is a number N such that
|fn (x) − fm (x)| < ε for all x ∈ S
if n, m ≥ N .
Theorem. Every uniform Cauchy sequence is uniformly convergent.
Proof. If {fn (x)} is a uniform Cauchy sequence on S, then clearly {fn (x0 )}
is a numerical Cauchy sequence for each x0 ∈ S. Thus by the Cauchy
criterion, the sequence {fn (x)} converges pointwise on S to a function f (x).
To see that the convergence is uniform, apply the uniform Cauchy property:
|fn (x) − fm (x)| < ε for all x ∈ S
when n, m ≥ N . Hold n ≥ N fixed and let m → ∞ to deduce that
|fn (x) − f (x)| ≤ ε for all x ∈ S
if n ≥ N . This proves the theorem.
As an application of the uniform Cauchy criterion, we can now estab-

lish a very useful condition for uniform convergence of an infinite series,
traditionally known as the Weierstrass M-test.
Theorem (Weierstrass M-test). Let the functions u1 , u2 , . . . be defined

on a set S ⊂ R, and suppose that |un (x)| ≤ Mn for all x ∈ S, for n =
1, 2, . . . . Supposefurther that the numerical series ∞n=1 Mn is convergent.
∞
Then the series n=1 un (x) converges uniformly on S.
Proof. Let ε > 0. Denote the partial sums of the series by

n
sn (x) = uk (x) , n = 1, 2, . . . .
k=1
Then for m < n we see that

n
n

|sn (x) − sm (x)| = uk (x) ≤ |uk (x)|

k=m+1 k=m+1
n
≤ Mk < ε for all x ∈ S
k=m+1

if m is sufficiently large, since the series ∞k=1 Mk converges. Thus {sn (x)}
is a uniform Cauchy sequence, and so it converges uniformly on S.
We conclude this section with a special result, known as Dini’s theorem,

which provides another useful criterion for uniform convergence.
Dini’s Theorem. Suppose that the functions f1 (x), f2 (x), . . . are contin-
uous on a closed bounded interval [a, b], and f1 (x) ≥ f2 (x) ≥ · · · ≥ 0 for
all x ∈ [a, b]. If fn (x) → 0 pointwise for each x ∈ [a, b], then fn (x) → 0
uniformly in [a, b].
Proof. It is convenient to begin by extending each function fn continu-

ously to the entire real line by defining fn (x) = fn (a) for x < a and
fn (x) = fn (b) for x > b. Let ε > 0 be given. Then for each fixed
x0 ∈ [a, b] there is an index N = N (x0 ) such that 0 ≤ fn (x0 ) < ε/2
for all n ≥ N . By continuity it follows that 0 ≤ fN (x) < ε for all x
in the interval I(x0 ) = (x0 − δ, x0 + δ) for some radius δ = δ(x0 ) > 0.
In view of the hypothesis that 0 ≤ fn+1 (x) ≤ fn (x), this implies that
0 ≤ fn (x) < ε for all n ≥ N (x0 ) and all x ∈ I(x0 ). Now observe that
the collection {I(x0 ) : x0 ∈ [a, b]} of all such intervals forms an open cover-
ing of [a, b], and apply the Heine–Borel theorem to extract a finite subcol-
lection {I(x1 ), I(x2 ), . . . , I(xm )} whose union contains [a, b]. Taking N ∗ to
be the maximum of the indices N (x1 ), N (x2 ), . . . , N (xm ), we conclude that
0 ≤ fn (x) < ε for all x ∈ [a, b] whenever n ≥ N ∗ . This proves the uniform
convergence.
Corollary 1. If a sequence of continuous functions fn (x) converges mono-

tonically to f (x) for each x ∈ [a, b] and if f (x) is continuous, then the
convergence is uniform.
Proof. By assumption, either fn (x) ≤ fn+1 (x) ≤ f (x) or fn (x) ≥ fn+1 (x) ≥
f (x). Define gn (x) = f (x) − fn (x) or gn (x) = fn (x) − f (x), respectively, to
obtain a sequence of continuous functions gn (x) that decreases pointwise to
0. The theorem then says that gn (x) → 0 uniformly in [a, b] as n → ∞.
Corollary 2. Let fn (x) and f (x) be nonnegative continuous functions that

are integrable over the real line. Suppose that either
0 ≤ fn (x) ≤ fn+1 (x) ≤ f (x) or 0 ≤ f (x) ≤ fn+1 (x) ≤ fn (x) ,
and that fn (x) → f (x) pointwise in R as n → ∞. Then

∞ ∞
lim fn (x) dx = f (x) dx .
n→∞ −∞ −∞
Proof. Let ε > 0 be given. Suppose first that 0 ≤ fn (x) ≤ f (x). Since f is
integrable, we can choose a number R > 0 for which
−R ∞
ε
0≤ f (x) dx + f (x) dx < ,
−∞ R 2
1.9. Historical remarks 39
which implies that

−R ∞
ε
0≤ [f (x) − fn (x)] dx + [f (x) − fn (x)] dx <
−∞ R 2
for all n. On the other hand, fn (x) → f (x) uniformly on [−R, R] by Dini’s
theorem, so that
R
ε
0≤ [f (x) − fn (x)] dx <
−R 2
for all n sufficiently large. Addition of these inequalities now shows that
∞
0≤ [f (x) − fn (x)] dx < ε
−∞
for sufficiently large n, which is the desired result. The case where 0 ≤
f (x) ≤ fn (x) is treated in similar fashion, with the integrability of f1 allow-
ing the reduction to a finite interval [−R, R].
Corollary 3. If the functions un (x) are

continuous, nonnegative, and inte-
grable over R , and if the infinite series ∞n=1 un (x) converges pointwise in
R to a continuous sum s(x) that is integrable over R , then
∞ ∞
∞
s(x) dx = un (x) dx .
−∞ n=1 −∞
Proof. Since un (x) ≥ 0, the partial sums sn (x) increase to s(x), so the
conclusion follows from Corollary 2.
1.9. Historical remarks

Sometime around the year 1670, Isaac Newton (1642–1727) and Got-
tfried Wilhelm Leibniz (1646–1716) discovered the basic principles of cal-
culus. The techniques were effective in solving a whole host of scientific
problems, notably problems of mechanics, but the logical foundations were
insecure. Concepts such as limits were understood intuitively at best and
were the focus of great controversy in early years. In fact, two centuries
would elapse before those concepts were clearly formulated and the subject
acquired a sound logical structure.
Leonhard Euler (1707–1783) was the greatest mathematician of the 18th
century and the most prolific of all time. His work spanned virtually every
aspect of mathematics and its applications. He wrote a series of three calcu-
lus texts that introduced the now-standard notation and served as models
for later treatments. In those texts Euler made functions, not curves, the
primary objects of study. However, his definition of a function was rather
restrictive and he accepted without question the notion of an “infinitely
small” quantity known as a differential. His proofs and derivations of formu-
las tended to be nothing more than plausible arguments, not at all rigorous
by modern standards. Joseph-Louis Lagrange (1736–1813), best known for
his fundamental work in mechanics, tried to free calculus from differentials
and limits by placing heavy emphasis on power series. His attempt was not
successful, but he did give the first clear analysis of Taylor series expansion.
In 1821, Augustin-Louis Cauchy (1789–1857) published the first rigorous
textbook in calculus, a Cours d’analyse written for his students at the École
Royale Polytechnique in Paris. There he gave clear definitions of limits and
continuity, essentially the definitions that are commonly accepted today. He
formulated the notion of a Cauchy sequence and he introduced the defi-
nition of integral that was later refined and completed by Riemann. The
book is a landmark in the history of mathematics. However, Cauchy failed
to distinguish pointwise continuity from uniform continuity, and pointwise
convergence from uniform convergence, making some of his arguments falla-
cious. Cauchy, like Euler, published a huge volume of work in many areas of
mathematical science. Perhaps his greatest achievement was to develop the
theory of functions of a complex variable, including the calculus of residues.
The dominant figure in the second half of the 19th century was Karl
Weierstrass (1815–1897). His career was truly remarkable. For some years
he taught at a Gymnasium, a German secondary school, carrying out high-
level mathematical research in obscurity. Then at age 39 he published a
groundbreaking paper in Crelle’s Journal. It was an immediate sensation.
Two years later, Weierstrass had moved to a professorship at the University
of Berlin. There his lectures contained many original results and attracted
students from all over the world, many of whom became outstanding math-
ematicians in later years. His lectures on introductory analysis perfected
the program that Cauchy had initiated, bringing new rigor and clarity to
calculus while introducing further results.
Bernard Bolzano (1781–1848), a Bohemian monk in Prague, was a con-
temporary of Cauchy who independently obtained many of the same results
in the foundations of analysis. In 1817 he published a proof of the inter-
mediate value theorem, often called “Bolzano’s theorem”. In the same long
paper, Bolzano introduced the least upper bound principle and “Cauchy’s
criterion” for convergence of a sequence. However, his paper appeared in
an obscure journal published in Prague, far from the scientific centers of
the day, and there is no evidence that Cauchy knew of Bolzano’s ideas be-
fore writing his own Cours d’analyse a few years later. As suggested by
1.10. Metric spaces 41
Bolzano’s preoccupation with the foundations of analysis, he was a philoso-

pher and a free thinker. His sermons in Prague attracted large audiences.
The Habsburg regime regarded him as a threat and issued an order in 1819
that placed Bolzano under house arrest and prohibited him from publishing
even the results of his mathematical research. At his death he left a large
number of manuscripts that were edited and published only after long delay,
starting around 1920. It then came to light that Bolzano had independently
discovered the Bolzano–Weierstrass theorem, had proved that a continuous
function on a closed bounded interval attains a maximum and minimum
value, and had constructed a continuous nowhere differentiable function.
Bernhard Riemann (1826–1866) gave his definition of an integral in his
Habilitationschrift of 1853, using “Riemann sums”. The approach in terms
of upper and lower sums is due to Gaston Darboux (1842–1917) and ap-
peared in 1875.
The history of the Heine–Borel theorem is rather complicated. It is
named for Eduard Heine (1821–1881) and Émile Borel (1871–1956), but
the attribution is considered inappropriate. Heine obtained a variant of the
result in 1872, and in 1895 Borel stated and proved what is now called the
Heine–Borel theorem, but only for countable coverings. Several years later,
Lebesgue and others generalized the result to arbitrary coverings. As early
as 1862, however, Dirichlet had used a form of the theorem implicitly in his
lectures, to prove that a continuous function on a closed bounded interval
is uniformly continuous. The theorem had also been known to Weierstrass
and others.
The books by Bottazzini [1], Edwards [3], Grabiner [4], and Jahnke [5]
are good sources of information about the historical development of analysis.
1.10. Metric spaces

In later chapters of this book, we will make occasional remarks that
refer to the relatively modern concept of a metric space. For the sake of
completeness, we include here a brief discussion of this important topic,
which unifies certain aspects of classical analysis.
A metric space is an ordered pair (S, ρ), where S is an arbitrary set and
ρ is a real-valued function defined on the Cartesian product S × S with the
properties
(i) ρ(x, y) ≥ 0 for all x, y ∈ S, with equality if and only if x = y ,
(ii) ρ(x, y) = ρ(y, x) for all x, y ∈ S , and
(iii) ρ(x, z) ≤ ρ(x, y) + ρ(y, z) for all x, y, z ∈ S .
A function ρ with these three properties is called a metric. Property (iii) is
known as the triangle inequality.
The leading example is (R, d), the set of real numbers equipped with
the usual metric, or distance function, d(x, y) = |x − y|. A more general
example is (Rn , dn ), the n-dimensional Euclidean space with metric
1/2

n
dn (x, y) = (xk − yk )2 , x = (x1 , . . . , xn ) , y = (y1 , . . . , yn ) .
k=1
A more artificial example consists of an arbitrary set endowed with the

metric
1 , if x = y
ρ(x, y) =
0 , if x = y.
Many basic concepts of analysis generalize naturally to the abstract set-
ting of a metric space (S, ρ). For instance, a sequence {xn } of elements
xn ∈ S is said to converge to a limit x ∈ S if for each ε > 0 there is a
number N such that ρ(xn , x) < ε for all n ≥ N . A sequence {xn } is called a
Cauchy sequence if for each ε > 0 there exists an N such that ρ(xn , xm ) < ε
whenever n, m ≥ N . A metric space is said to be complete if every Cauchy
sequence is convergent. Thus the space (R, d) is complete, whereas (Q, d) is
not. (Recall that Q denotes the set of rational numbers.)
For applications to analysis, it is especially useful to consider metric
spaces whose elements are functions. For instance, the set S can be taken
to consist of all continuous functions on a fixed interval [a, b], with metric
ρ(f, g) = sup |f (x) − g(x)| , f, g ∈ S .

x∈[a,b]
Convergence in this metric is equivalent to uniform convergence on [a, b].

Another space is given by the same set of functions with metric
b
ρ(f, g) = |f (x) − g(x)| dx .
a
The first of these two metric spaces is complete, but the second is not.
An extensive theory of metric spaces has been developed, and there are
many examples important in analysis. Further information can be found in
the book by Rudin [8] and in other texts.
1.11. Complex numbers

Although this book deals primarily with real-valued functions of a real
variable and makes no use of complex function theory, it is often conve-
nient to work with complex arithmetic. Hence we conclude this preliminary
chapter with a brief account of the complex number system.
The need for an extension of the real number system is already apparent
in the problem of solving a quadratic equation ax2 + bx + c = 0, where
a, b, c ∈ R and a = 0. According to the quadratic formula, the solutions
have the form √
−b ± b2 − 4ac
x= .
2a
Of course, it may well happen that b2 − 4ac < 0, in which case the square-
root is undefined and the equation has no solutions in the algebra of real
numbers. In this case, however, it is fruitful to interpret the quadratic
formula as providing a pair of solutions
√
b 4ac − b2 √
− ±i , where i = −1 .
2a 2a
In effect, this amounts to extending the algebra of real numbers to the
algebra of complex numbers z = x + iy, where x and y are real numbers.
It should be emphasized that in the process of enlarging the set R of
real numbers to the set C of complex numbers, the algebraic operations of
addition and multiplication are also extended to C. The sum and product
of two complex numbers z = x + iy and w = u + iv are defined by
z + w = (x + u) + i(y + v) and zw = (xu − yv) + i(xv + yu) .
These expressions result from formal manipulations, assuming that the imag-
inary number i is subject to the usual rules of arithmetic. In this way the
algebra of real numbers is extended to the algebra of complex numbers,
preserving the basic properties of the arithmetic operations. Specifically,
addition and multiplication are both commutative and associative, and mul-
tiplication is distributive over addition: z(w1 + w2 ) = zw1 + zw2 .
A complex number z = x + iy is said to be real if y = 0, and purely
imaginary if x = 0. The real part of z is Re{z} = x and the imaginary
part is Im{z} = y. The complex conjugate of z is z = x − iy. Conjugation
commutes with addition and multiplication: z +w = z + w and zw = z w.
The modulus of a complex number z is |z| = x2 + y 2 , a generalization
of the absolute value of a real number. Observe that z z = |z|2 and that
|z| = |z|. Two useful formulas are
1 1
Re{z} = (z + z) and Im{z} = (z − z) .
2 2i
It may be noted that |Re{z}| ≤ |z| and |Im{z}| ≤ |z|. A simple calculation
shows that
|z + w|2 = (z + w)(z + w) = |z|2 + 2 Re{zw} + |w|2 .

Complex numbers have a natural geometric interpretation. Just as real

numbers can be identified with points on the real line, complex numbers
can be identified with points in the complex plane. Specifically, a complex
number z = x + iy corresponds to the point in the plane with Cartesian
coordinates (x, y). The x-axis is then called the real axis, and the y-axis
is called the imaginary axis. The real and imaginary parts of z are its
orthogonal projections onto the real and imaginary axes, respectively. The
complex conjugate of z is its reflection across the real line. The modulus
of z is its distance from the origin. More generally, |z − w| is the distance
between the points z and w, and so |z − α| = R is the equation of a circle
with center α ∈ C and radius R > 0. Representing points in the plane by
polar coordinates (r, θ), we can write x = r cos θ and y = r sin θ, so that
z = x + iy = r(cos θ + i sin θ). The angle θ is called the argument of z and is
denoted θ = arg{z}. Of course, the argument is determined only up to an
additive multiple of 2π.
From the geometric viewpoint of the complex plane, the sum of two
complex numbers is given by the parallelogram law, as in vector addition.
For a geometric interpretation of the product, it is useful to consider polar
representations
z = r(cos θ + i sin θ) and w = ρ(cos ϕ + i sin ϕ) .
Then by the standard addition formulas for sine and cosine,
zw = rρ[(cos θ cos ϕ − sin θ sin ϕ) + i(cos θ sin ϕ + sin θ cos ϕ)]

= rρ[cos(θ + ϕ) + i sin(θ + ϕ)] .
In other words, the product of two complex numbers is the number whose
modulus is the product of their moduli and whose argument is the sum of
their arguments. In symbols,
|zw| = |z||w| and arg{zw} = arg{z} + arg{w} .
One corollary of the relation |zw| = |z||w| is the triangle inequality

|z + w| ≤ |z| + |w|. For a proof, observe that
|z + w|2 = |z|2 + 2 Re{zw} + |w|2 ≤ |z|2 + 2 |zw| + |w|2

= |z|2 + 2 |z||w| + |w|2 = (|z| + |w|)2 .
For a geometric interpretation that justifies the name “triangle inequality”,

rewrite it in the form
|α − β| ≤ |α − γ| + |γ − β| , where α, β, γ ∈ C .
Then the inequality says that the distance from α to β is no greater than
the distance from α to γ plus the distance from γ to β. The three points
α, β, γ can be viewed as vertices of a triangle.
In fact the formula
|α − β|2 = |α|2 + |β|2 − 2 Re{αβ}
is none other than the law of cosines. To see this, view the three points 0,
α, and β as vertices of a triangle. Then the sides have lengths |α|, |β|, and
|α − β|. Let θ = arg{α} and ϕ = arg{β}, where 0 < θ − ϕ < π. Then
the angle at the origin, the vertex opposite the side of length |α − β|, is
ψ = θ − ϕ. But Re{αβ} = |α||β| cos(θ − ϕ), so the formula does reduce to
the law of cosines: c2 = a2 + b2 − 2ab cos ψ.
Another consequence of the product formula is the de Moivre formula
(cos θ + i sin θ)n = cos nθ + i sin nθ ,
which is a compact expression for a family of trigonometric identities. In
view of the binomial theorem it gives, for instance
cos 5θ = cos5 θ − 10 cos3 θ sin2 θ + 5 cos θ sin4 θ .
The product formula appears more natural when the polar form of a
complex number is expressed in exponential form. Euler’s formula eiθ =
cos θ + i sin θ results from a formal comparison of Taylor series expansions.
(See Chapter 3, Exercise 23.) If we write z = reiθ and w = ρeiϕ , the
product formula is zw = rρei(θ+ϕ) . It is possible to define the exponential
of a complex number as a power series, but for the moment we may simply
regard eiθ as a notation for cos θ + i sin θ, with the property eiθ eiϕ = ei(θ+ϕ)
that is characteristic
of an exponential function. Observe also that eiθ = e−iθ
iθ
and e = 1, and that
d iθ
e = − sin θ + i cos θ = ieiθ .
dθ
The quotient of two complex numbers a + ib and c + id = 0 can be
defined as the solution z of the equation (c + id)z = a + ib. To see that
this equation has a unique solution, let z = x + iy and write the complex
equation in real form as a pair of linear equations
cx − dy = a , dx + cy = b .
This system has the unique solution
ac + bd bc − ad
x= 2 2
, y= 2 .
c +d c + d2
The quotient can be calculated by writing
a + ib (a + ib)(c − id) ac + bd bc − ad
= = 2 +i 2 .
c + id (c + id)(c − id) c + d2 c + d2
In more concise form, z/w = z w/|w|2 .
Exercises
1. Use mathematical induction to prove that
n(n + 1)(2n + 1)
(a) 12 + 22 + · · · + n2 = , n = 1, 2, . . . ,
6
(b) 13 + 23 + · · · + n3 = (1 + 2 + · · · + n)2 , n = 1, 2, . . . .
2. (a) Verify the identity of Pascal’s triangle:

n n n+1
+ = , k = 1, 2, . . . , n .
k−1 k k
(b) Use mathematical induction to prove the binomial theorem
n
n n k n−k
(x + y) = x y , n = 1, 2, . . . .
k
k=0
3. (a) Use the binomial theorem to show that
(1 + x)n ≥ 1 + nx for all x ≥ 0 .
(b) Prove that the inequality holds more generally for all x > −1.
4. Prove the inequality | sin nθ| ≤ n| sin θ| , n = 1, 2, . . . .
5. The Fibonacci sequence {xn } is defined by
x0 = 0, x1 = 1, x2 = 1, x3 = 2, x4 = 3, x5 = 5, x6 = 8, . . . .
Here the general rule is that each number in the sequence is the sum of the
two previous numbers: xn+2 = xn+1 + xn for n = 0, 1, 2, . . . .
(a) Show that the recurrence relation xn+2 = xn+1 + xn has solutions
of the form xn = αn for exactly two real constants α. Find this pair of
numbers.
(b) Show that the Fibonacci sequence has the form
xn = A α n + B β n , n = 0, 1, 2, . . . ,
for some real constants A, B, α, and β. Determine these numbers.

(c) Check your formula for xn by induction.
6. If X and Y are arbitrary countable sets, show that their Cartesian
product
X × Y = {(x, y) : x ∈ X , y ∈ Y }
is countable.
Exercises 47
7. Prove that the collection of all infinite sequences of 0’s and 1’s is un-
countable.
8. Prove that the collection of all sets of positive integers is uncountable.
9. Prove that the set of all algebraic numbers is countable and deduce the
existence of transcendental numbers.
Hint. Use the fact that a polynomial of degree n has at most n real roots.
10. (a) Show that the reciprocal of each nonzero algebraic number is
algebraic.
(b) Show that the square root of each positive algebraic number is
algebraic.
11. Show that a set of real numbers is closed if and only if it contains all of
its cluster points.
12. For any set S of real numbers, the derived set S is defined as the set
all cluster points of S. (It may be empty.) Show that for every set S ⊂ R
the set S is closed.
13. Let {sn } be a sequence of real numbers and let
s1 + s2 + · · · + sn
σn =
n
be its sequence of arithmetic means.
(a) If {sn } converges to s, show that {σn } also converges to s.
(b) Give an example to show that the converse is false: the sequence
{σn } may converge although the sequence {sn } does not.
14. Without appeal to Stirling’s formula, calculate
1
lim (n!)1/n .
n→∞ n
Suggestion. Apply the last theorem in Section 1.4.
15. Suppose a1 ≥ a2 ≥ a3 ≥ · · · > 0. Prove that the series
∞
∞
an converges if and only if 2k a2k converges.
n=1 k=0
16. Show that if

an
lim (an+1 − an ) = λ , then lim = λ.
n→∞ n→∞ n
17. If x1 = 1 and
√
xn+1 = 1 + xn , n = 1, 2, . . . ,
prove that the sequence {xn } converges and find its limit.
18. Given numbers a1 and b1 with 0 < a1 < b1 , define the sequences {an }
and {bn } inductively by
an + bn
an+1 = an bn and bn+1 = , n = 1, 2, . . . .
2
Show that {an } and {bn } both converge, and to the same limit.
√
Hint. First observe that ab ≤ a+b2 for all a > 0 and b > 0.
19. Using the results of Section 1.4, verify the unrestricted limits
(a) lim xp rx = 0 for p > 0 and 0 < r < 1 ,
x→∞
(b) lim x−p log x = 0 for p > 0 ,
x→∞
(c) lim xp log x = 0 for p > 0 ,
x→0+
(d) lim xx = 1 .
x→0+
20. Without appeal to Stirling’s formula, verify the asymptotic relation

n 1
∼ nk for each fixed k ∈ N, as n → ∞ .
k k!
In other words, show that the ratio of the two expressions tends to 1 as
n → ∞.
21. For a sequence {xn } that is bounded above and does not tend to −∞,
let S be the set of all subsequential limits. Show that S contains the number
sup S, so that {xn } has a largest subsequential limit.
22. For bounded sequences {an } and {bn } prove that
lim sup (an + bn ) ≤ lim sup an + lim sup bn and
n→∞ n→∞ n→∞
lim inf (an + bn ) ≥ lim inf an + lim inf bn .
n→∞ n→∞ n→∞
Give examples to show that strict inequality may occur.
23. For all bounded sequences {xn } verify that

lim sup xn = lim sup xk and lim inf xn = lim inf xk .
n→∞ n→∞ k≥n n→∞ n→∞ k≥n
24. Prove the second inequality in the theorem at the end of Section 1.4:
rn+1 √
lim inf ≤ lim inf n rn .
n→∞ rn n→∞
25. Suppose a function f (x) is continuous and has the property f (x + y) =

f (x) + f (y) for all x, y ∈ R. Prove that f (x) = Cx , where C = f (1).
Hint. Prove it first for rational numbers x.
Exercises 49
26. Define a function f on the interval (0, 1) as follows. If x is rational and

x = p/q in lowest terms, let f (x) = 1/q. If x is irrational, let f (x) = 0. Show
that f is discontinuous at each rational point of (0, 1) but is continuous at
every irrational point.
27. Generalize the mean value theorem by proving that if f and g are
continuous on [a, b] and differentiable on (a, b), then
[f (b) − f (a)]g (ξ) = [g(b) − g(a)]f (ξ) for some ξ ∈ (a, b) .
28. Apply the generalized mean value theorem of the preceding exercise to
establish l’Hôpital’s rule in the following form. If f and g are differentiable
in an open interval (a, b) and f (x)/g (x) → L as x → a+, and if f (x) → 0
and g(x) → 0 as x → a+. then f (x)/g(x) → L as x → a+.
29. Verify that the function defined by
2
x sin x1 , x = 0
f (x) =
0, x = 0.
is differentiable everywhere on R, but its derivative is not continuous at the
origin.
30. The preceding example shows that a derivative may exist but fail to be
continuous. Show nevertheless that a derivative always has the intermediate
value property. Specifically, prove that if f is differentiable on an interval
[a, b] and f (a) < λ < f (b), then f (ξ) = λ for some ξ ∈ (a, b).
Hint. Show that g(x) = f (x) − λx attains its minimum value at a point
ξ ∈ (a, b).
31. For 0 ≤ x ≤ 1, define the functions fn (x) = n2 xn (1 − x), where
n = 1, 2, . . . .
(a) Show that limn→∞ fn (x) = 0 for each x ∈ [0, 1].
n
(b) Show directly, by consideration of fn ( n+1 ), that the convergence is
not uniform on [0, 1].
1
(c) Show that limn→∞ 0 fn (x)dx = 1, and conclude again that the
convergence of fn (x) to 0 is not uniform on [0, 1].
32. Let the functions f1 (x), f2 (x), . . . be continuous on the interval [0, 1]
and suppose that fn (x) → f (x) uniformly on [0, 1] as n → ∞. Show that
fn (1/n) → f (0).
33. Apply the Bolzano–Weierstrass theorem instead of the Heine–Borel
theorem to show that a function continuous at each point of a closed bounded
set S ⊂ R is uniformly continuous on S.
34. Let the numbers anm be defined by

⎧ 0, for 1 ≤ n < m
⎨
anm = −1 , for n = m
⎩ m−n
2 , for 1 ≤ m < n .
Show that
∞
∞ ∞
∞
anm = −2 but anm = 0 .
n=1 m=1 m=1 n=1
Explain why this example does not contradict Cauchy’s double series theo-
rem (see Section 1.5).
References
[1] Umberto Bottazzini, The Higher Calculus: A History of Real and Complex
Analysis from Euler to Weierstrass, Springer-Verlag, New York, 1986.
[2] J. C. Burkill and H. Burkill, A Second Course in Mathematical Analysis, Cam-

bridge University Press, Cambridge, U.K., 1970.
[3] C. H. Edwards, The Historical Development of the Calculus, Springer-Verlag,

New York, 1979.
[4] Judith V. Grabiner, The Origins of Cauchy’s Rigorous Calculus, MIT Press,
Cambridge, Massachusetts, 1981.
[5] Hans Niels Jahnke (editor), A History of Analysis, American Mathematical
Society, Providence, Rhode Island, and London Mathematical Society, London,
2003.
[6] Arthur Mattuck, Introduction to Analysis, Prentice–Hall, Upper Saddle River,
New Jersey, 1999.
[7] Kenneth Ross, Elementary Analysis: The Theory of Calculus, Springer-Verlag,
New York, 1980.
[8] Walter Rudin, Principles of Mathematical Analysis, Third edition, McGraw–
Hill, New York, 1976.
Chapter 2
Special Sequences
This chapter is devoted to a study of some special sequences that arise often
in analysis. We begin with two familiar sequences which converge to e, the
base of natural logarithms. As a byproduct of the analysis, the number e is
found to be irrational. This prompts an elementary proof that π is irrational
as well. Further topics include Euler’s constant γ and the infinite product
formulas of Vieta and Wallis. The chapter concludes with a crown jewel of
classical analysis, Stirling’s approximation to the factorial function.
2.1. The number e

The number e = 2.71828182845904 . . . , known as the base of natural
logarithms, arises through the limit

1 n
e = lim 1 +
n→∞ n
and is found to have the form
1 1
e=1+1+ + + ... .
2! 3!
Our aim is to reconcile the two expressions and to show in particular that
the first limit exists. We will also show that e is an irrational number.
First of all, let us prove that the infinite series converges, so that we can
legitimately define the number e as its sum. Let
1 1 1
sn = 1 + 1 + + + ···+
2! 3! n!
51
52 2. Special Sequences
be the nth partial sum, and observe that sn < sn+1 , since all terms are
positive. On the other hand, the inequality k! ≥ 2k−1 shows that
1 1 1
sn ≤ 1 + 1 + + 2 + · · · + n−1 < 3
2 2 2
for all n. Thus the sequence {sn } is monotonic and bounded, and is therefore
convergent. We denote its limit, or the sum of the infinite series, by e.
Now consider the expressions

1 n
xn = 1 + , n = 1, 2, 3, . . . .
n
In order to study the behavior of the sequence {xn }, we can use a hand
calculator to compute
x1 = 2 , x2 = 2.25 , x3 = 2.370 . . . , x4 = 2.441 . . . , x5 = 2.488 . . . , etc.
The data suggest that {xn } is an increasing sequence. Our next step will be
to show that this is actually true.
According to the binomial theorem, we can write
n k
1 n n 1
xn = 1 + =
n k n
k=0
2
1 n(n − 1) 1 n(n − 1)(n − 2) 1 3
=1+n + +
n 2! n 3! n
n
n(n − 1)(n − 2) · · · 1 1
+ ··· +
n! n

1 1 1 1 2
=1+1+ 1− + 1− 1−
2! n 3! n n

1 1 2 n−1
+ ··· + 1− 1− ··· 1 − .
n! n n n
A similar expansion of xn+1 gives
n+1
n+1 k
1 n+1 1
xn+1 = 1+ = .
n+1 k n+1
k=0
Observe now that for each fixed k in the range 2 ≤ k ≤ n, the term
k
n+1 1 1 1 2 k−1
= 1− 1− ··· 1 −
k n+1 k! n+1 n+1 n+1
k
1 1 2 k−1 n 1
≥ 1− 1− ··· 1 − = .
k! n n n k n
2.1. The number e 53
n+1
1
Furthermore, the expansion for xn+1 contains an extra term n+1 > 0,
corresponding to k = n + 1. These two inequalities combine to show that
xn+1 > xn for n = 1, 2, 3, . . . .
The above expansion for xn also shows that
1 1 1
xn ≤ 1 + 1 + + + ··· + = sn < 3 .
2! 3! n!
Therefore, x1 < x2 < x3 < · · · < 3, and so the sequence {xn } is monotonic
and bounded, hence convergent. We are going to prove that
lim xn = e ,
n→∞
the limit of the sequence {sn }. Since xn ≤ sn for all n, it follows that
lim xn ≤ lim sn = e .
n→∞ n→∞
On the other hand, for each fixed m ≤ n the sum in the previous calcu-
lation can be truncated to give

1 1 1 1 2
xn ≥ 1 + 1 + 1− + 1− 1−
2! n 3! n n

1 1 2 m−1
+ ··· + 1− 1− ··· 1 − .
m! n n n
Now hold m fixed and let n → ∞ to infer that

1 1 1
lim xn ≥ 1 + 1 + + +··· + = sm
n→∞ 2! 3! m!
for each index m. Letting m → ∞, we see from this that
lim xn ≥ lim sm = e .
n→∞ m→∞
Combining this with the earlier inequality limn→∞ xn ≤ e, we conclude that
1 1
lim xn = lim sn = e = 1 + 1 + + + ... .
n→∞ n→∞ 2! 3!
The infinite series

1 1
e=1+1+ + + ...
2! 3!
converges quite rapidly. For instance s7 = 2.71825396 . . . already gives the

correct value of e to four decimal places. To estimate the rate of convergence,
we can write
1 1 1
e − sn = + + +...
(n + 1)! (n + 2)! (n + 3)!
2
1 1 1 1
< 1+ + + ... = ,
(n + 1)! n+1 n+1 n! n
after summing the geometric series. Therefore,
1
0 < e − sn < for n = 1, 2, 3, . . . .
n! n
It is now easy to show that e is irrational. Suppose on the contrary that

e is rational, so that e = mn for some positive integers m and n. By the
estimate just given,
1
0 < n!(e − sn ) < ≤ 1 .
n
But
1 1 1
n!sn = n! 1 + 1 + + + · · · +
2! 3! n!
is an integer, and n!e is also an integer under the assumption that e = m n.

Hence that assumption has led to the conclusion that n!(e − sn ) is an integer
between 0 and 1, which is impossible. The contradiction shows that e is
irrational.
In fact, it is known that e is a transcendental number. Recall that a real
number x is said to be algebraic if it satisfies some polynomial equation
a0 + a1 x + a2 x2 + · · · + an xn = 0
with integer coefficients a0 , a1 , . . . , an . All rational

√ numbers are algebraic,
and so are many irrational numbers such as 2. A real number that is not
algebraic is said to be transcendental. Both e and π are known to be tran-
scendental, but those assertions are not easy to prove. The transcendence of
e was proved by Charles Hermite [2] in 1873. Then in 1882 Ferdinand von
Lindemann [4] adapted Hermite’s method to establish the transcendence of
π. A simpler version of the Hermite–Lindemann proof can be found in the
book by Ivan Niven [5]. In the next section we present Niven’s proof of the
more elementary fact that π is irrational.
2.2. Irrationality of π 55
2.2. Irrationality of π
We now digress from the theme of this chapter to prove that the number
π is irrational. This fact lies intrinsically deeper than the irrationality of e,
and was proved by more sophisticated methods before Ivan Niven [5] found
the remarkably elementary proof that will be presented here. In fact, the
proof yields the stronger result that π 2 is irrational.
Consider the polynomial
1 n
f (x) = x (1 − x)n ,
n!
where n is a positive integer to be specified later. This function satisfies
1
0 < f (x) < for 0 < x < 1 .
n!
It is easy to see that each of the derivatives f (k) (0) is an integer. Indeed,
f (k) (0) = 0 for 0 ≤ k < n, and for k ≥ n a calculation shows that the kth
derivative of xn (1 − x)n at the origin is an integer divisible by n!. (This
remains true if the factor (1 − x)n is replaced by any other polynomial
with integer coefficients.) By the symmetry relation f (1 − x) = f (x), it
follows that every derivative f (k) (1) is also an integer. Observe finally that
f (k) (x) ≡ 0 for all k > 2n, since f is a polynomial of degree 2n.
Now suppose, for purpose of contradiction, that π 2 is rational, so that
π 2 = p/q for some positive integers p and q. Define the polynomial

g(x) = q n π 2n f (x) − π 2n−2 f (2) (x) + π 2n−4 f (4) (x) − · · · + (−1)n f (2n) (x) ,
and note that both g(0) and g(1) are integers under the supposition that
π 2 = p/q. Because of the structure of g, we see that
g (x) + π 2 g(x) = q n π 2n+2 f (x) .
In view of this relation, a simple calculation gives
d
g (x) sin πx − πg(x) cos πx = g (x) + π 2 g(x) sin πx
dx
= q n π 2n+2 f (x) sin πx = π 2 pn f (x) sin πx .
Consequently,
1 1
1
πp n
f (x) sin πx dx = g (x) sin πx − g(x) cos πx = g(1) + g(0) ,
0 π 0
which is an integer. On the other hand, since 0 < f (x) < 1/n! for 0 < x < 1 ,
we find that 1
n πpn
0 < πp f (x) sin πx dx < <1
0 n!
if n is chosen sufficiently large. But this is impossible, because there is no
integer between 0 and 1. Thus the assumption that π 2 is rational has led to
a contradiction, and so we conclude that π 2 is irrational, which implies that
π is irrational.
The transcendence of π has an interesting application to classical geom-
etry. It settles once and for all the ancient problem of “squaring the circle”
with straight edge and compass. Given an arbitrary circle, the problem is
to construct a square of the same area. Since the circle has area πr2 , this
amounts to starting with a line segment of unit length and constructing a
√
segment of length π. A segment of length π could then be constructed,
since from any segment of length it is possible to construct a segment of
length 2 . But it can be shown that the length of every segment constructible
with straight edge and compass, starting with a segment of unit length, is
an algebraic number. (A good reference for this fact is the book by Courant
and Robbins [1].) Therefore, if it were possible to square the circle with
straight edge and compass, the number π would have to be algebraic. But
π is transcendental, so it is impossible to square the circle.
2.3. Euler’s constant

Euler’s constant is

n
1
γ = lim − log n .
n→∞ k
k=1
It is named for Leonhard Euler, who first discussed it in 1734. The number
γ is an important constant that occurs frequently in mathematical formulas.
The existence of the limit is not obvious. Our aim is to prove that the limit
exists and to determine its approximate numerical value.
Consider the curve y = 1/x for 1 ≤ x ≤ n, where n = 2, 3, . . . . The area
under the curve is given by
n
1
An = dx = log n .
1 x
Now construct rectangular boxes of heights 1/k over the intervals [k, k + 1],
as shown in Figure 1.
x
1 2 3 4 ... n
Figure 1. The curve y = 1/x and rectangular boxes.
Since
1 1 1
≤ ≤ for k ≤ x ≤ k + 1 ,
k+1 x k
it follows that
k+1 k+1 k+1
1 1 1 1 1
= dx ≤ dx ≤ dx =
k+1 k k+1 k x k k k
for k = 1, 2, . . . . Adding these inequalities over k = 1, 2, . . . , n − 1, we have

n−1 n 1 n−1
1 1
≤ dx ≤ .
k+1 1 x k
k=1 k=1
With the notation

n
1
Sn = ,
k
k=1
this says that Sn − 1 ≤ An ≤ Sn−1 . The two inequalities can be rearranged

to give
1
0 ≤ Sn−1 − An ≤ 1 − Sn + Sn−1 = 1 − .
n
This shows that the sequence {Sn−1 − An } is positive and is bounded above
by 1.
Geometrically, the quantity Sn−1 − An is the sum of areas of those por-
tions of the boxes that lie above the curve y = 1/x from x = 1 to n. In order
to estimate this total area, imagine that all of these boxes are slid to the left
until they lie inside the first box, as shown in Figure 2, where the shaded
regions have total area Sn−1 − An . Since the regions are nonoverlapping
and lie inside a square of area 1, this conceptual exercise gives a geometric
interpretation of the inequality Sn−1 − An ≤ 1.
Figure 2. Geometric estimate of Sn−1 − An .
Next observe that

n+1
1 1
An+1 − An = dx ≤ = Sn − Sn−1 ,
n x n
or Sn−1 − An ≤ Sn − An+1 , which says that the sequence {Sn−1 − An } is

nondecreasing. An appeal to the monotone boundedness theorem now shows
that the sequence converges. Denoting its limit by γ, we have

γ = lim (Sn−1 − An ) = lim Sn − 1
n − An = lim (Sn − An ) .
n→∞ n→∞ n→∞
The existence of Euler’s constant has now been established, and the proof
shows that 0 ≤ γ ≤ 1.
In fact, it is visually clear from Figure 2 that γ is slightly larger than 12 .
Our next task is to derive quantitative bounds on γ by estimating the area
k+1
1 1
αk = − dx
k k x
of the region in the kth box that lies above the curve y = 1/x. Since the
curve is convex, that region contains a triangle of area

1 1 1
−
2 k k+1
and is contained in a trapezoid of area
1 1
− 1 ,
k k+ 2
1
k k+2 k+1
Figure 3. Estimation of the area αk .
constructed by drawing the tangent line to the curve at the point where
x = k + 12 . (See Figure 3 and note that the trapezoid above the tangent line
is obtained by removing the lower trapezoid from the entire rectangle.)
Thus a comparison of areas shows that

1 1 1 1 1 1 1
− ≤ αk ≤ − 1 =2 − .
2 k k+1 k k+ 2
2k 2k + 1
Summing these inequalities, we find
n−1 n−1 1
1 1 n−1
1 1
− ≤ αk ≤ 2 − ,
2 k k+1 2k 2k + 1
k=1 k=1 k=1
or

1 1 1 1 1 1 1
1− ≤ Sn−1 − An ≤ 2 − + − +··· − .
2 n 2 3 4 5 2n − 1
Letting n → ∞, we infer that

1 1 1 1
0.5 ≤ γ ≤ 2 − + − + . . . = 2(1 − log 2) = 0.6137 . . . ,
2 3 4 5
since (see Exercise 7)
1 1 1
1− + − + · · · = log 2 .
2 3 4
The actual value of Euler’s constant is
γ = 0.577215664901532 . . . .
It has been computed to thousands of decimal places and no periodicities
have been detected, so it is strongly suspected to be an irrational number.
In fact, it is generally conjectured to be a transcendental (or nonalgebraic)
number, like the constants π and e. However, no one has ever been able to
prove that γ is irrational!
2.4. Vieta’s product formula

The infinite product formula

2
1
2
1
2 + 1
2
1
2
1
2 + 1
2
1
2 + 1
2
1
2 ··· =
π
was discovered in 1593 by the French lawyer and mathematician François

Viète (1540–1603), whose name was rendered into Latin as Franciscus Vieta.
The formula derives special interest from the unexpected appearance of the
number π. It can be written more explicitly as
2
lim (b1 b2 · · · bn ) = ,
n→∞ π
where

1 1
b1 = 2 and bn+1 = 2 + 12 bn , n = 1, 2, . . . .
The proof is based on a simple trigonometric identity. We begin with

the double-angle formula
x x
sin x = 2 cos sin , x > 0.
2 2
Introducing the same identity in the form
x x x
sin = 2 cos sin ,
2 4 4
we see that
x x x
cos sin .
sin x = 4 cos
2 4 4
After n iterations, the process leads to the formula
x x x x x
sin x = 2n cos cos cos · · · cos n sin n ,
2 4 8 2 2
or
x x x x sin x
cos cos cos · · · cos n = n ,
2 4 8 2 2 sin(x/2n )
provided that x/2n is not an integer multiple of π. In particular, the formula
holds for any fixed x > 0 whenever n is sufficiently large. Observe now that
x sin θ
lim 2n sin = x, since lim = 1.
n→∞ 2n θ→0 θ
2.5. Wallis product formula 61
Consequently, we have derived the infinite product formula

x x x x sin x
lim cos cos cos · · · cos n = , 0 < x < π,
n→∞ 2 4 8 2 x
written in standard notation as
∞
x sin x
cos n = .
2 x
n=1
With the choice x = π/2, the formula reduces to

∞
π 2
cos n
= .
2 π
n=2

But cos π4 = 1
2, and the half-angle formula

cos 2 = 12 + 12 cos θ ,
θ
0 < θ < π,
shows that

π 1 π
cos = 2 + 12 cos , n = 1, 2, 3, . . . .
2n+1 2n
Therefore, bn = cos(π/2n+1 ) and the last infinite product is simply an equiv-
alent statement of Vieta’s formula.
2.5. Wallis product formula

The remarkable Wallis product formula is
2·2 4·4 6·6 (2n)(2n) π

(1) lim ··· = .
n→∞ 1·3 3·5 5·7 (2n − 1)(2n + 1) 2
It was first recorded by John Wallis (1616–1703) as early as 1656. Equivalent
formulations are

1 1 1 2
(2) lim 1− 2 1− 2 ··· 1 − =
n→∞ 2 4 (2n)2 π
and
22n (n!)2 √
(3) lim √ = π.
n→∞ (2n)! n
The factors in (2) are simply the reciprocals of those in (1). To see that (1)
2n → 1, the formula (1)
and (3) are equivalent, observe first that because 2n+1
implies
22 42 62 · · · (2n − 2)2 π
lim (2n) = .
n→∞ 3 5 7 · · · (2n − 1)
2 2 2 2 2
Taking square roots, we deduce that

π 2 · 4 · 6 · · · (2n − 2) √
= lim 2n
2 n→∞ 3 · 5 · 7 · · · (2n − 1)
22 42 62 · · · (2n − 2)2 (2n)2
= lim √
n→∞ (2n)! 2n
1 22n (n!)2
= √ lim √ ,
2 n→∞ (2n)! n
which shows that (1) implies (3). Since all steps are reversible, a similar
argument shows that (3) implies (1).
Taken in the version (3), the Wallis product√formula may be regarded
as a weak form of Stirling’s formula n! ∼ nn e−n 2πn , or
n!
lim √ = 1,
n→∞ nn e−n 2πn
which will be proved in the next section.
For a proof of the product formula (1), consider the integrals
π/2
In = sinn x dx , n = 0, 1, 2, . . . .
0
Observe that I0 = π
2 and I1 = 1. To compute In for n ≥ 2, write
sinn x = sinn−1 x sin x
and integrate by parts to obtain In = (n − 1)(In−2 − In ), which gives the

reduction formula
n−1
In = In−2 , n = 2, 3, . . . .
n
Iterating the reduction formula and using the values of I0 and I1 , we find
2n − 1 2n − 3 1 π
I2n = ··· ·
2n 2n − 2 2 2
and
2n 2n − 2 2
I2n+1 = ··· · 1.
2n + 1 2n − 1 3
Dividing one equation by the other, we arrive at the expression

π 2·2 4·4 6·6 (2n)(2n) I2n
= ··· .
2 1·3 3·5 5·7 (2n − 1)(2n + 1) I2n+1
It is now clear that a proof of the Wallis product formula reduces to

showing that I2n /I2n+1 → 1 as n → ∞. But the inequalities
0 < sin2n+1 x < sin2n x < sin2n−1 x

π
hold for 0 < x < 2 , so that 0 < I2n+1 < I2n < I2n−1 . It follows that
I2n I2n−1 2n + 1
1< < = → 1,
I2n+1 I2n+1 2n
which shows that I2n /I2n+2 → 1 as n → ∞. This completes the proof of the
Wallis product formula (1).
The formula is actually a special case of a much more general relation.
The sine function has the infinite product representation
∞

x2
sin πx = πx 1− 2 , x ∈ R,
n
n=1
1
as will be proved in Chapter 8. Set x = 2 to obtain the Wallis product
formula (2).
2.6. Stirling’s formula

The asymptotic formula
√
n! ∼ nn e−n 2πn , n → ∞,
is known as Stirling’s formula. It is of basic importance for instance in

probability theory and combinatorics, because it gives precise information
about the growth of the factorial function. The symbol “∼” means that
n!
lim √ = 1.
n→∞ nn e−n 2πn
The formula was discovered in 1730, and the Scottish mathematician

James Stirling (1692–1770) included it in his book Methodus Differentialis
(London, 1730). However, most of the credit actually belongs to Abraham
de Moivre (1667–1754), a mathematician born in France who spent his adult
life in England, where he became an intellectual companion to Isaac Newton.
Although de Moivre was never able to obtain a suitable position in England,
he made fundamental discoveries in the early theory of probability, including

a preliminary version of the central limit theorem. In 1730, he found that
1
n! ∼ C nn+ 2 e−n for some constant C, and Stirling determined that C =
√
2π. In fact, as we shall see, the value of the constant is implicit in the
Wallis product formula.
We propose to prove Stirling’s formula by showing that
n! 1
1< √ <1+ , n = 1, 2, . . . .
nn e−n 2πn 4n
The error estimates are important in quantitative applications of the for-
mula.
With the observation that
log(n!) = log 1 + log 2 + · · · + log n ,
it is natural to base a proof on a careful study of the area under the loga-
rithmic curve y = log x from x = 1 to n. (See Figure 4.) An integration by
parts calculates this area as
n
n
An = log x dx = x log x − x = n log n − n + 1 .
1 1
On the other hand, the area can be estimated geometrically. Since the
logarithm is a concave function, the curve y = log x lies above each of its
chords connecting successive points (k, log k), for k = 1, 2, . . . , n. Thus An
is larger than the sum of areas of the trapezoids under those line segments.
The total area of the trapezoids is
Tn = 1
2 log 2 + 12 (log 2 + log 3) + · · · + 12 (log(n − 1) + log n)
= log 2 + log 3 + · · · + log(n − 1) + 1
2 log n
= log(n!) − 1
2 log n .
x
1 2 3 4 ... n
Figure 4. The logarithmic curve and inscribed trapezoids.

Now let αk denote the area of the small region bounded by the curve
y = log x and the line segment joining the two points (k, log k) and (k +
1, log(k + 1)), for k = 1, 2, . . . , n − 1. Then the total area under the curve is
An = Tn + En , where En = α1 + α2 + · · · + αn−1 .
Inserting the expressions for An and Tn , we can write this relation in the
form
log(n!) = n + 12 log n − n + 1 − En ,
or
n! = Cn nn+ 2 e−n ,
1
where Cn = e1−En .
The sequence {En } is increasing, since each term αk is positive. We
now show that the sequence {En } has an upper bound and is therefore
convergent. In order to estimate αk , we construct the tangent line to the
curve y = log x at the point where x = k + 12 (see Figure 5) and compare
areas:

1 1
αk < log k + − log k + log(k + 1)
2 2

1 k + 12 1 k+1
= log − log
2 k 2 k + 12

1 1 1 1
= log 1 + − log 1 +
2 2k 2 2k + 1

1 1 1 1
< log 1 + − log 1 + .
2 2k 2 2(k + 1)
k k + 1/2 k
Figure 5. Estimation of the area αk .

Adding these inequalities, we find that

n−1
n−1
1

1

1

1

En = αk < log 1 + − log 1 +
2 2k 2 2(k + 1)
k=1 k=1

1 3 1 1 1 3
= log − log 1 + < log ,
2 2 2 2n 2 2
since
the dominant series telescopes. Thus En increases to a finite limit
E = ∞ k=1 αk , and so Cn = e
1−En decreases to a limit C = e1−E > 0. In
particular, Cn > C, and so 1 < Cn /C = eE−En . But

∞
∞

1 1 1 1
E − En = αk < log 1 + − log 1 +
2 2k 2 2(k + 1)
k=n k=n

1 1
= log 1 + ,
2 2n
again because the dominant series telescopes. Therefore,

1 1
1 < Cn /C = eE−En < 1 + <1+ .
2n 4n
In summary, we have shown that

n! 1
0 < C < Cn = 1 <C 1+ .
nn+ 2 e−n 4n
In order to finish the proof of Stirling’s

√ formula with error estimates,
it now remains only to show that C = 2π. This is where we invoke the
Wallis product formula. It gives
1 2
√ 22n (n!)2 22n Cn nn+ 2 e−n
π = lim √ = lim √
n→∞ (2n)! n n→∞ C (2n)2n+ 21 e−2n n
2n
1 C2 C2 C
= √ lim n = √ =√ .
2 n→∞ C2n 2C 2
√
Thus C = 2π and the proof is complete.
Exercises
1. (a) Verify the inequality
x2
x− < log(1 + x) < x , x > 0,
2
Exercises 67
and use it to show that

n
1 n 1
lim 1 + 2 =1 and lim 1− 2 = 1.
n→∞ n n→∞ n
Conclude that n
1 1
lim 1− = .
n→∞ n e
(b) Find

1 2 n
lim 1+ 2 1+ 2 ··· 1 + 2 .
n→∞ n n n
2. Sharpen the inequality of the previous problem to
x2 x2 x3
x− < log(1 + x) < x − + , x > 0.
2 2 3
(This is an improvement for small x.) Apply the result to calculate
n 2
1
lim 1+ e−n .
n→∞ n
3. (a) Evaluate
1/n
1 (2n)!
lim
n→∞ n n!
by showing, directly from the definition of an integral, that the logarithm
2
of the expression tends to 1 log x dx. Calculate the integral.
(b) Check your result with the help of Stirling’s formula.
4. Prove that the infinite series

∞
1
(n!)2
n=1
converges, and that its sum is an irrational number.
5. Write e = sn + rn , where sn = 1 + 1
1! + 1
2! +··· + 1
n! .
1 1
(a) Show that < n!rn < .
n+1 n
(b) Calculate lim n sin(2πn!e) .
n→∞
(c) Conclude that e is irrational.

6. Test the following infinite series for convergence and absolute conver-
gence. Explain your reasoning.
∞
n!en
(a) , p ∈ R.
nn+p
n=1
∞
a
(b) (−1)n sin , a > 0.
n
n=1
∞
1
(c) 1 − cos .
n
n=1
Note. These facts about the sine function may be useful:
2 π sin x
x ≤ sin x ≤ x for 0 ≤ x ≤ ; lim = 1.
π 2 x→0 x
7. Use the existence of Euler’s constant γ to verify that
1 1 1
1− + − + · · · = log 2 .
2 3 4
8. Show by a similar method that
1 1 1 1 1 1 1 1 3
1+ − + + − + + − + · · · = log 2 .
3 2 5 7 4 9 11 6 2
9. Show that

n
1 1 γ
lim − log n = + log 2 .
n→∞ (2k + 1) 2 2
k=0
10. Prove the existence of the limit

1 1 1
lim + +··· + ,
n→∞ n + 1 n+2 2n
and calculate its value.
11. Let

n
1 1
xn = − log n and yn = xn − .
k 2n
k=1
(a) Show that the sequence {xn } is decreasing and {yn } is increasing,
so that yn < γ < xn for n = 1, 2, . . . .
Exercises 69
(b) Use a hand calculator to compute x10 = 0.62638 . . . . Conclude that

0.576 < γ < 0.627.
(c) Use Mathematica
R or a similar computer software system to cal-
culate xn for n = 10, 000. Use the result to give the bounds 0.57721 < γ <
0.57726 and thus to calculate γ = 0.5772 . . . to 4 decimal places.
*12. With xn defined as in the preceding exercise, it can be shown by an

elementary geometric (or graphical) argument due to Robert Young [10]
that
1 1
< xn − γ < , n = 1, 2, . . . .
2(n + 1) 2n
This indicates that the convergence of {xn } to Euler’s constant is very slow.
13. Prove that

∞
log k 1
(−1)k+1 = (log 2)2 − γ log 2 ,
k 2
k=1
where γ is Euler’s constant. Note that Leibniz’ alternating series theorem

implies that the series converges. Evaluate its sum in the following two
steps.
(a) Integrate by parts to calculate
x
log t 1
dt = (log x)2 , x ≥ 1,
1 t 2
and apply the result to show that the sequence {dn } given by

n
log k 1
dn = − (log n)2
k 2
k=1
is convergent. (Imitate the argument used to prove existence of Euler’s

constant.)
(b) Write the partial sums in the form

2n
2n
n
k+1 log k log k log 2k
(−1) = −2 ,
k k 2k
k=1 k=1 k=1
express the right-hand side in terms of d2n and dn , and let n → ∞.

Note. See also Chapter 9, Exercise 18. The proof is due to H. F. Sandham
[8].

R Mathematica is a registered trademark of Wolfram Research, Inc.
14. Verify the formula

√
3 3
1
2
1
2 + 1
2 · 1
2
1
2 + 1
2
1
2 + 1
2 · 1
2
1
2 + 1
2
1
2 + 1
2
1
2 + 1
2 · ··· =
1
2 .
4π
15. Show that the integrals
π/2
In = sinn x dx , n = 1, 2, . . . ,
0
are evaluated by the formulas
(2n)! π 22n (n!)2

I2n = 2n+1
and I2n+1 = .
2 (n!)2 (2n + 1)!
16. Derive the Wallis product formula from Stirling’s formula.
17. Use Stirling’s formula to verify the asymptotic relation

2n 22n
∼√ as n → ∞ .
n πn
18. Integrate the inequalities
π
sin2n+1 x < sin2n x < sin2n−1 x , 0<x< ,
2
from 0 to π/2 to prove that
2 2
2 (2n)!! 1 (2n)!!
<π< , n = 1, 2, . . . ,
2n + 1 (2n − 1)!! n (2n − 1)!!
where
(2n)!! = 2 · 4 · 6 · · · (2n) and (2n − 1)!! = 1 · 3 · 5 · · · (2n − 1) .
Obtain the Wallis product formula as a corollary.

References 71
References
[1] Richard Courant and Herbert Robbins, What is Mathematics?, Oxford Univer-
sity Press, London and New York, 1941.
[2] Charles Hermite, “Sur la fonction exponentielle”, Comptes Rendus Acad. Sci.
Paris 77 (1873), 18–24, 74–79, 226–233, 285–293.
[3] W. J. Kaczor and M. T. Nowak, Problems in Mathematical Analysis I : Real

Numbers, Sequences and Series, Student Mathematical Library, Volume 4, Ameri-
can Mathematical Society, Providence, RI, 2000.
[4] Ferdinand von Lindemann, “Über die Zahl π”, Math. Annalen 20 (1882), 213–
225.
[5] Ivan Niven, Irrational Numbers, Mathematical Association of America, Wash-

ington, DC, 1956.
[6] Ivan Niven, Numbers: Rational and Irrational , New Mathematical Library,
Volume 1, Random House, New York-Toronto, 1961.
[7] Walter Rudin. Principles of Mathematical Analysis, Third edition, McGraw–

[8] H. F. Sandham, Solution to Advanced Problem 4592 (proposed by M. S. Klamkin),

Amer. Math. Monthly 62 (1955), 588–589.
[9] David V. Widder, Advanced Calculus, Second edition, Prentice–Hall, Englewood

Cliffs, NJ, 1961; reprinted by Dover Publications, Mineola, NY, 1989.
[10] Robert M. Young, “Euler’s constant”, Math. Gazette 75 (1991), 187–190.
[11] Robert M. Young, Excursions in Calculus: An Interplay of the Continuous

and the Discrete, Mathematical Association of America, Washington, DC, 1992.
Chapter 3
Power Series and
Related Topics
This chapter begins with a review of the basic properties of power series. The
more delicate problem of boundary behavior is then addressed with a dis-
cussion of Abel’s theorem. In the converse direction, Taylor’s formula with
remainder is developed and applied to the expansion of certain functions in
power
series. The results then allow an elementary evaluation of Euler’s sum
1/k = π 2 /6. The chapter concludes with Weierstrass’s famous example
2
of a continuous function that is not differentiable at any point.
3.1. General properties of power series

A power series is an infinite series of the form
∞

f (x) = cn (x − a)n = c0 + c1 (x − a) + c2 (x − a)2 + . . . .
n=0
The number a is called the center of the power series, and the numbers cn
are its coefficients. The first problem is to find the set of numbers x ∈ R for
which the series converges.
Obviously, the series converges for x = a since every term then vanishes
except possibly for c0 = f (a). It need not converge anywhere else, but if
it does converge at some point x0 = a, then it will converge in the interval
|x − a| < |x0 − a|, uniformly in each closed subinterval |x − a| ≤ ρ < |x0 − a|.
73
74 3. Power Series and Related Topics
To see this, observe first that the convergence at x0 implies that the terms
cn (x0 − a)n tend to zero as n → ∞, so that
|cn (x0 − a)n | ≤ A , n = 0, 1, 2, . . . ,
for some constant A. Therefore,

n
x − a n ρ
n n
|cn (x − a) | = |cn (x0 − a) | ≤A , |x − a| ≤ ρ .
x0 − a |x0 − a|

But the geometric series (ρ/|x0 −a|)n converges, so the series cn (x−a)n
converges uniformly in the interval |x − a| ≤ ρ by the Weierstrass M-test.
As a consequence, we see that if the series diverges at some point x1 , then it
must diverge at every point x with |x − a| > |x1 − a|. The conclusion is that
the series converges throughout some symmetric interval |x − a| < R and it
diverges everywhere outside that interval, except possibly at the endpoints.
The number R is known as the radius of convergence of the power series. It
can happen that R = ∞, meaning that the series converges for all x ∈ R ,
or that R = 0 if it converges only for x = a.
These conclusions are summarized as follows.

Theorem. A power series ∞ n=0 cn (x − a) has a radius of convergence R ,
n
where 0 ≤ R ≤ ∞. If R = ∞, the series converges for all x and converges

uniformly in each interval |x − a| ≤ ρ < ∞ . If R = 0, it converges only at
x = a. If 0 < R < ∞,the series converges in the open interval |x − a| < R
and converges uniformly in each closed subinterval |x − a| ≤ ρ < R, whereas
it diverges everywhere outside the closed interval |x − a| ≤ R.

Here are some examples. The geometric series ∞ n
n=0 x has radius of
convergence R = 1. It converges in the interval (−1, 1) but diverges at both
endpoints. The series ∞ 1 n
n=1 n x again has R = 1, and it diverges at x = 1
∞ 1 n
but converges at x = −1. The
series n n=1 n2 x has R = 1 and converges at
both endpoints. The series ∞ n=1 n!x converges only for x = 0; its radius of
∞ 1 n
convergence is R = 0. At the other extreme, the series n=1 n! x converges
for all x, so R = ∞.

The radius of convergence of an arbitrary power series cn (x − a)n is
given by the formula
1
R= , where β = lim sup n
|cn | .
β n→∞
Here it is understood that R = ∞ if β = 0 and R = 0 if β = ∞. For a proof,

we need only appeal to the root test for convergence of an infinite series.
3.1. General properties of power series 75
In view of the uniform convergence within the interval of convergence,

the sum function
∞

f (t) = cn (t − a)n
n=0
is continuous in the interval a − R < t < a + R and the series can be
integrated term by term to give
x ∞
x ∞
1
f (t) dt = cn (t − a) dt =
n
cn (x − a)n+1
a a n+1
n=0 n=0
for x ∈ (a − R, a + R). As an illustration, the geometric series

∞
1
= tn , |t| < 1 ,
1−t
n=0
is integrated to produce the series expansion

x ∞

1 1 1
log = dt = xn+1 = x+ 12 x2 + 13 x3 +. . . , |x| < 1 .
1−x 0 1 − t n + 1
n=0
For another important example, start with the series

∞
1
2
= (−1)n t2n , |t| < 1 ,
1+t
n=0
to obtain the expansion

x ∞
x
−1 1 n
tan x = 2
dt = (−1) t2n dt
0 1+t 0 n=0
∞
(−1)n 2n+1
= x = x − 13 x3 + 15 x5 − 17 x7 + . . . , |x| < 1 .
2n + 1
n=0
A power series can also be differentiated term by term within its interval
of convergence. We have the following theorem.
Theorem. The sum of a power series
∞

f (x) = cn (x − a)n , |x − a| < R ,
n=0
is differentiable and its derivative is

∞

f (x) = ncn (x − a)n−1 , |x − a| < R .
n=1
Proof. Observe that the formally differentiated power series

∞
g(x) = ncn (x − a)n−1
n=1
√
has the same radius of convergence R, since n n → 1. Term-by-term inte-
gration gives
x ∞

G(x) = g(t) dt = cn (x − a)n = f (x) − c0 , |x − a| < R .
a n=1
But the integrand is continuous, so the fundamental theorem of calculus
tells us that G(x) is differentiable for |x − a| < R and G (x) = g(x). Since
G(x) = f (x) − c0 , this proves the theorem.
Because the derivative f (x) is given by another power series with the
same radius of convergence, the theorem shows that f (x) exists and is
expressed by the twice-differentiated power series. Continuing in this way,
we see that the sum of a power series has derivatives of all orders, calculated
by successive term-by-term differentiation.
3.2. Abel’s theorem

The short life of Niels Henrik Abel (1802–1829) was filled with remark-
able discoveries in a wide variety of mathematical areas, ranging from the
insolubility of quintic equations by radicals to the theory of elliptic func-
tions. From an obscure village in southern Norway he rose to international
prominence but was never able to obtain an appropriate scientific position.
His posthumous statue (1908) by the famous Norwegian sculptor Gustav
Vigeland now stands in a small park near the center of Oslo, adjacent to the
Royal Palace.
The theorem of Abel to be presented here is relatively elementary but
quite useful in evaluating the sums of certain infinite series. It is the pro-
totype for a class of theorems now known as “Abelian theorems”, and its
proof involves a basic device now called “Abel summation”.

Abel’s Theorem. If an infinite series ∞ n=0 an converges to the sum s,
then
∞

lim an xn = s .
x→1−
n=0

Note that the convergence of the series ∞ n=0 an says that the power
series
∞

f (x) = an xn
n=0
converges at the point x = 1. It then follows that the power series converges
for all x in the interval (−1, 1], and Abel’s theorem asserts that the sum
function f is continuous at the point 1. Since the sum of a power series
is known to be continuous within its interval of convergence (−R, R), the
theorem as stated is of interest only when R = 1.
Here are some examples. The function f (x) = log(1 + x) has the power
series expansion
log(1 + x) = x − 12 x2 + 13 x3 − 14 x4 + . . . , |x| < 1 .
For x = 1, the series reduces to 1 − 12 + 13 − 14 + . . . , which converges by

Leibniz’ alternating series theorem. Thus Abel’s theorem tells us that
1− 1
2 + 1
3 − 1
4 + · · · = lim log(1 + x) = log 2 .
x→1−
Another example is provided by the power series expansion
tan−1 x = x − 13 x3 + 15 x5 − 17 x7 + . . . , |x| < 1 .
Again the series converges for x = 1, so Abel’s theorem says that
1− 1
3 + 1
5 − 1
7 + · · · = lim tan−1 x = tan−1 1 = π
4 .
x→1−
On the other hand, the converse of Abel’s theorem is false in the absence
of additional hypotheses. In other words, the existence of the limit of f (x)
as x → 1− does not imply the convergence of the power series at x = 1. For
instance,
1
1 − x + x2 − x3 + · · · = → 12
1+x
as x → 1−, but the series 1 − 1 + 1 − 1 + . . . is divergent.
The proof of Abel’s theorem makes use of a device known as Abel sum-
mation or summation by parts, the discrete analogue of integration by parts.
It can be expressed as follows.
Let {ak } and {bk } be arbitrary sequences of real numbers, and let
Lemma.
sn = nk=0 ak . Then for 0 ≤ m < n,

n
n−1
ak bk = sk (bk − bk+1 ) + sn bn − sm−1 bm ,
k=m k=m
where s−1 = 0.
Proof.

n
n
n
n−1
ak bk = (sk − sk−1 )bk = sk bk − sk bk+1
k=m k=m k=m k=m−1

n−1
= sk (bk − bk+1 ) + sn bn − sm−1 bm .
k=m
n
Proof of Abel’s theorem. Again let sn = k=0 ak denote the partial
sums of the series, so that sn → s as n → ∞. For |x| < 1, let
∞

n
k
f (x) = ak x , and let fn (x) = ak x k
k=0 k=0
denote the partial sums of the power series. Then by the lemma,

n−1
fn (x) = sk (1 − x)xk + sn xn .
k=0
Fixing x in the interval (0, 1) and letting n → ∞, we arrive at the expression
∞

f (x) = (1 − x) sk xk , 0 < x < 1,
k=0
since sn → s and xn → 0. But
∞

(1 − x) xk = 1 , |x| < 1 ,
k=0
so we can write
∞

f (x) − s = (1 − x) (sn − s)xn , 0 < x < 1.
n=0
Given ε > 0, choose N such that |sn − s| < 2ε for all n ≥ N . Then
N −1 ∞

|f (x) − s| ≤ (1 − x) |sn − s|x +
n
|sn − s|x n
n=0 n=N

N −1
< (1 − x) |sn − s| + ε
2 < ε
2 + ε
2 =ε
n=0
if x is sufficiently close to 1. In other words, there is a number δ > 0 such
that
|f (x) − s| < ε when 0 < 1−x < δ.
This proves the theorem.
It is now easy to deduce a more general version of Abel’s theorem.

Corollary. If a power series

∞

f (x) = an x n
n=0
has radius of convergence R, where 0 < R < ∞, and if the series converges
for x = R, then limx→R− f (x) = f (R). In other words, the function f (x) is
left-continuous at x = R.
For a proof, we need only consider the function g(x) = f (Rx), whose
power series development has radius of convergence 1.
Abel’s theorem is much easier to prove in the special case where all of
the coefficients an are of one sign. Then the result is a direct consequence
of the Weierstrass M-test for uniform convergence. To see this, observe
that
|an xn | ≤ an for |x| ≤ 1 if an ≥ 0, so the convergence of the series ∞
n=0 an
implies the uniform convergence of the power series
∞

f (x) = an x n
n=0
in the closed interval [−1, 1]. Therefore, the sum f (x) is continuous in [−1, 1]
and is in particular continuous at the point 1, as Abel’s theorem asserts.
The technique of Abel summation has many applications. For instance,
it can be used to obtain the following result.

Theorem. If an infinite series ∞ n=1 an has bounded partial sums, and if
n∞} is a sequence of positive numbers that decrease to zero, then the series
{b
n=1 an bn is convergent.

Proof. By hypothesis, the partial sums sn = nk=1 ak have the property
|sn | ≤ M for some constant M and all n. An Abel summation gives

n
n−1
ak bk = sk (bk − bk+1 ) + sn bn .
k=1 k=1
Since sn bn → 0 as n → ∞ and

n−1
n−1
|sk (bk − bk+1 )| ≤ M (bk − bk+1 ) = M (b1 − bn ) ≤ M b1 ,
k=1 k=1
∞
it follows that the series k=1 ak bk converges.
Taking for instance an = (−1)n+1 , we see that |sn | ≤ 1, and so we have

a new proof of Leibniz’ alternating series theorem.
There is an integral analogue of Abel’s theorem that involves Laplace

transforms instead of power series. A bounded function is said to be inte-
grable over the positive half-line [0, ∞) if its Riemann integral exists over
each bounded interval [0, R] and tends to a finite limit as R → ∞. Such
a function need not be absolutely integrable over [0, ∞). For instance, the
function g(t) = sint t is integrable, by Leibniz’ alternating series theorem, but
|g(t)| is not integrable over [0, ∞).
On the other hand, if g(t) is any bounded locally integrable function,
then e−xt g(t) is absolutely integrable over [0, ∞) for each x > 0, so the
Laplace transform
∞
f (x) = e−xt g(t) dt , x > 0,
0
is well defined. Here is the integral analogue of Abel’s theorem.
Theorem. If g(t) is bounded and integrable over [0, ∞), then
∞ ∞
−xt
lim e g(t) dt = g(t) dt .
x→0+ 0 0
Proof. The strategy is to apply integration by parts essentially as in the

proof of Abel’s theorem. Let
t
G(t) = g(s) ds , 0 ≤ t < ∞,
0
denote the indefinite integral of g, so that
∞
lim G(t) = I = g(s) ds .
t→∞ 0
We want to show that the Laplace transform f (x) converges to I as x → 0+.
For this purpose we integrate by parts to obtain
∞ ∞
−xt
f (x) = e g(t) dt = x e−xt G(t) dt , x > 0.
0 0
Hence ∞
f (x) − I = x e−xt G(t) − I dt .
0
Given ε > 0, choose R large enough that |G(t) − I| < ε/2 for all t ≥ R, and
observe that
R ∞
−xt
|f (x) − I| ≤ x e |G(t) − I| dt + x e−xt |G(t) − I| dt
0 R
ε
≤ xR max |G(t) − I| + < ε
0≤t≤R 2
for all x > 0 sufficiently small. This proves that f (x) → I as x → 0+.
3.3. Cauchy products and Mertens’ theorem 81
3.3. Cauchy products and Mertens’ theorem

The Cauchy product of two infinite series
∞
∞

an and bn
n=0 n=0
∞
is the series n=0 cn , where

n
cn = an−k bk = an b0 + an−1 b1 + · · · + a0 bn , n = 0, 1, 2, . . . .
k=0
This definition is suggested by power series considerations. Multiplication

of two power series
∞
∞

an xn and bn x n
n=0 n=0
gives
∞
∞

an xn bn x n
n=0 n=0

= a0 + a1 x + a2 x2 + . . . b0 + b1 x + b2 x2 + . . .
= a0 b0 + (a1 b0 + a0 b1 )x + (a2 b0 + a1 b1 + a0 b2 )x2 + . . .
= c 0 + c 1 x + c 2 x2 + . . . ,
where the coefficients cn are those in the Cauchy product.

Such a calculation requires justification, which will be given presently
in the form of Mertens’ theorem (cf. Exercise 14). Meanwhile, the result of
formal manipulation will serve to motivate the notion of a Cauchy product.
Observe that this operation of multiplication is commutative in the sense
that
n n
cn = an−k bk = bn−k ak .
k=0 k=0
The Cauchy product of two convergent series need not converge. Con-
sider for example the series
∞
∞
(−1)n 1 1 1
an = √ = 1 − √ + √ − √ +... ,
n=0 n=0
n+1 2 3 4
which converges by Leibniz’ alternating series theorem. We will show that

the Cauchy product of this series with itself is divergent. Here we have

n
n
1
cn = an−k ak = (−1) n
√ √ .
k=0 k=0
n−k+1 k+1
But n 2 n 2 n 2
(n − k + 1)(k + 1) = +1 − −k ≤ +1 ,
2 2 2
so that

n
2 2(n + 1)
|cn | ≥ = ≥ 1, n = 0, 1, 2, . . . .
n+2 n+2
k=0

Hence the series cn diverges, since its terms do not tend to 0.

The preceding example involves a series an that is convergent but not
absolutely convergent. If an and bn are both absolutely convergent,
it is permissible to rearrange terms to show that their Cauchy product is
convergent. A theorem of Mertens says that absolute convergence of only
one of the series allows the same conclusion.

Mertens’ Theorem. Suppose the series an and bn converge to sums
A
and B, respectively. Let c n be the
Cauchy product of the two series. If
an is absolutely convergent, then cn converges to sum C = AB.
Proof. Let

n
n
n
An = ak , Bn = bk , and Cn = ck
k=0 k=0 k=0
denote the partial sums, and let Rn = B − Bn . Then we can write

Cn = a0 b0 + (a0 b1 + a1 b0 ) + · · · + (a0 bn + a1 bn−1 + · · · + an b0 )
= a0 Bn + a1 Bn−1 + · · · + an B0
= An B − a0 Rn − a1 Rn−1 − · · · − an R0 .
Since An → A, it will suffice to show that
σn = a0 Rn + a1 Rn−1 + · · · + an R0 → 0 as n → ∞ .
But Rn → 0, so for each ε > 0 there is a number N such that |Rn | < ε for
all n ≥ N . Then for n > N we have
|σn | ≤ |a0 Rn + a1 Rn−1 + · · · + an−N RN | + |an−N +1 RN −1 + · · · + an R0 |
n
≤ε |ak | + |an−N +1 RN −1 + · · · + an R0 | .
k=0
Since an → 0 as n → ∞, we conclude that

∞

|σn | ≤ ε 1 + |ak |
k=0
for all n sufficiently large. This proves that σn → 0, and it follows that
Cn → AB as n → ∞, which proves the theorem.
Franz Josef Mertens (1840–1927) is best known for his work in num-
ber theory. As a student at the University of Berlin, he attended lectures
by Kummer, Kronecker, and Weierstrass. Later he held professorships at
universities in Krakow, Graz, and Vienna.
3.4. Taylor’s formula with remainder

Suppose a function f (x) is given as the sum of a power series with
positive radius of convergence:
∞

f (x) = cn (x − a)n , |x − a| < R ,
n=0
where a is some real number. Then, as we have seen, f ∈ C ∞ . In other

words, f has derivatives of all orders in the interval (a−R, a+R). Moreover,
the derivatives of f can be calculated by successive term-by-term differenti-
ation. Thus
∞
∞

f (x) = ncn (x − a) n−1
, f (x) = n(n − 1)cn (x − a)n−2 , etc.
n=1 n=2
Proceeding in this manner, we see that f (a) = c1 , f (a) = 2c2 , and in

general the kth derivative of f at the point a is f (k) (a) = k!ck . Thus the
coefficients cn can be expressed in terms of f by the simple formula
1 (n)
(1) cn = f (a) , n = 0, 1, 2, . . . .
n!
The problem now arises to describe the class of functions f that can be
expressed as the sum of a power series in some open interval (a − R, a + R).
We know that any such function must be of class C ∞ , and if it has a power
series expansion at the point a, the coefficients must have the form (1). But
is the converse true? In other words, if f ∈ C ∞ in some interval (a−R, a+R),
must it have a power series expansion
∞
1 (n)
(2) f (x) = f (a)(x − a)n
n!
n=0
in that interval, or at least in some smaller open interval about the point a?
This is a much more delicate question, and the answer in general is no. A
counterexample is outlined in Exercise 6.
The formal expansion (2) of a function f into power series is known as the
Taylor series of f at the point a. It is so named after Brooke Taylor (1685–
1731), an English mathematician who recorded it as early as 1715, without
considering the question of convergence. The problem of convergence of
the Taylor series to a given function was first addressed by Colin Maclaurin
(1698–1746) in 1742 and was successfully analyzed by Joseph-Louis Lagrange
(1736–1813) in his book Théorie des fonctions analytiques, published in
1797.
Although the problem of Taylor series representation is fully understood
only in the context of functions of a complex variable, sufficient conditions
for validity of the expansion are available through an explicit formula for the
difference between f (x) and the nth partial sum of its Taylor series. This
expression is commonly called Taylor’s formula with remainder, but its first
clear statement appears in the work of Lagrange.
Here is a derivation of Taylor’s formula. Let f (x) have continuous deriva-
tives of all orders on some open interval containing a given point a ∈ R. By
the fundamental theorem of calculus,
x
f (x) = f (a) + f (t) dt for x near a.
a
Now hold x fixed and integrate by parts, letting u(t) = f (t) and v(t) =
−(x − t) in the usual notation. This leads to
x x
f (x) = f (a) + f (t)v (t) dt = f (a) + f (a)(x − a) + f (t)(x − t) dt .
a a
Another integration by parts produces

x
1
f (x) = f (a) + f (a)(x − a) + 2 f (a)(x −
2
a) + 1
2 f (t)(x − t)2 dt .
a
Continuing in this manner, we obtain
f (x) = Sn (x) + Rn (x) , n = 1, 2, . . . ,
where

n
1 (k)
Sn (x) = f (a)(x − a)k
k!
k=0
and x
1
Rn (x) = f (n+1) (t)(x − t)n dt .
n! a
This is Taylor’s formula. The polynomial Sn (x) is the nth partial sum of
the Taylor series of f at the point a, and Rn (x) is called the remainder.
In order to prove that Sn (x) converges to f (x) as n → ∞, or that f has
the Taylor series expansion
∞
1 (k)
f (x) = f (a)(x − a)k ,
k!
k=0
we have to show that Rn (x) tends to 0 as n → ∞. This is equivalent to

proving that the formal Taylor series of f actually converges to the sum
f (x).
For this purpose, the Lagrange form of the remainder is often convenient.
It is
(x − a)n+1
Rn (x) = f (n+1) (ξ) ,
(n + 1)!
where ξ is some point between a and x. (We are allowing either x > a or
x < a.) A derivation of the Lagrange formula makes use of the following
lemma, which can be proved by appeal to the intermediate value theorem
for continuous functions.
Lemma. Let g(t) and h(t) be continuous functions on an interval [a, b], and
suppose that h(t) ≥ 0. Then
b b
g(t)h(t) dt = g(ξ) h(t) dt
a a
for some ξ ∈ (a, b).
To derive the Lagrange form of the remainder from the integral repre-
sentation, we take g(t) = f (n+1) (t) and h(t) = (x − t)n in the lemma to
conclude that
x
1 (n+1) (x − a)n+1
Rn (x) = f (ξ) (x − t)n dt = f (n+1) (ξ)
n! a (n + 1)!
for some ξ between a and x. If instead we take h(t) = 1, an application of

the lemma leads to the Cauchy form of the remainder:
(x − ξ)n
Rn (x) = f (n+1) (ξ) (x − a)
n!
for some ξ between a and x. Specifically, this means that a < ξ < x or
x < ξ < a.
Here are two examples.
Example 1. Let f (x) = sin x and let a = 0. We are going to show that the
familiar Taylor series of the sine function actually converges to sin x for every
x ∈ R. Since it is a power series, we can then conclude that the convergence
is uniform on each bounded subset of R, but this will also follow directly
from our estimate of the remainder in Taylor’s formula. We find that
f (0) = 0 , f (0) = 1 , f (0) = 0 , f (0) = −1 , f (4) (0) = 0 ,

f (5) (0) = 1 , . . . , so the formal Taylor series of sin x is
x3 x5 x7
x− + − + ... .
3! 5! 7!
It is easy to see (by the ratio test, for instance) that this infinite series
converges for all x ∈ R, but convergence alone does not imply that its sum
is sin x. To prove this, we have to verify that Rn (x) → 0 as n → ∞ for each
fixed x ∈ R. Referring to the Lagrange form of the remainder and noting
that f (n+1) (ξ) is equal to ± sin ξ or ± cos ξ, we see that |f (n+1) (ξ)| ≤ 1, so
that
|x|n+1
|Rn (x)| ≤ →0 as n → ∞ .
(n + 1)!
A similar analysis verifies the expansions
x2 x4 x6 x2 x3
cos x = 1 − + − + ... and ex = 1 + x + + +...
2! 4! 6! 2! 3!
for all x ∈ R.
Example 2. Next let f (x) = log(1 + x) for x > −1, and again take a = 0.
A calculation gives
f (n) (x) = (−1)n+1 (n − 1)! (1 + x)−n , n = 1, 2, . . . .
Thus the formal Taylor series of log(1 + x) about the origin is found to be
x − 12 x2 + 13 x3 − 14 x4 + . . . .
We invoke Taylor’s formula to write f (x) = Sn (x) + Rn (x), where
Sn (x) = x − 12 x2 + 13 x3 − · · · + (−1)n+1 n1 xn
and
x n
1 x
x−t 1
Rn (x) = f (n+1)
(t) (x − t) dt = (−1)
n n
dt .
n! 0 0 1+t 1+t
3.5. Newton’s binomial series 87

We now apply the inequality x−t 1+t ≤ |x|, which can be seen to hold for all
t between 0 and x. (See Exercise 4.) It provides the estimate
x
1
|Rn (x)| ≤ |x|n dt = |x|n log(1 + |x|) ,
0 1+t
which tends to 0 as n → ∞ if −1 < x < 1. This confirms that
log(1 + x) = x − 12 x2 + 13 x3 − 14 x4 + . . . , |x| < 1 ,
a result previously obtained by direct integration of a geometric series.

For x = 1 the Lagrange form of the remainder is more effective. It yields
1 (1 + ξ)−n+1
Rn (1) = f (n+1) (ξ) = (−1)n
(n + 1)! n+1
for some ξ in the interval 0 < ξ < 1. Thus |Rn (1)| ≤ n+1
1
→ 0, which shows
that the power series converges to f (1) when x = 1. Observe that we have
given another derivation, without appeal to Abel’s theorem, of the formula
1− 1
2 + 1
3 − 1
4 + · · · = log 2 .
3.5. Newton’s binomial series
This important generalization of the binomial theorem was formulated

by Isaac Newton (1642–1727) as early as 1665. Recall that the binomial
theorem gives the formula
n
n n k
(1 + x) = 1 + x ,
k
k=1
where n is a positive integer and

n n! n(n − 1) · · · (n − k + 1)
= =
k k!(n − k)! k!
is a binomial coefficient. Newton’s generalization is

∞
α
α
(1 + x) = 1 + xk , |x| < 1 ,
k
k=1
where α is an arbitrary real number and

α α(α − 1) · · · (α − k + 1)
=
k k!
is a generalized
α binomial coefficient. Observe that if α is a positive integer,
then k = 0 for all k ≥ α + 1 and Newton’s binomial series reduces to the
standard binomial theorem.
We can derive Newton’s formula as a Taylor series expansion about the
origin of the function f (x) = (1 + x)α , where α = 0, 1, 2, . . . and x > −1.
Our assumption that α is not a positive integer or zero implies that αk = 0
for all k. Successive differentiations produce the formula

1 (k) α
f (x) = (1 + x)α−k , k = 1, 2, . . . .
k! k
Thus
n
α
Sn (x) = 1 + xk
k
k=1
and
x n
x−t α(α − 1) · · · (α − n)
Rn (x) = rn (1 + t)α−1 dt , where rn = .
0 1+t n!
To estimate the factors rn , observe that

α − n + 1
|rn+1 /rn | = →1 as n → ∞ ,
n+1
which implies that for each fixed number ρ > 1, the inequality |rn+1 /rn | ≤ ρ
holds for all n ≥ N . Thus |rn | ≤ Cρn for all n, where C is a constant.
x−t
Appealing to the elementary inequality 1+t ≤ |x| used in the previous
section, we can estimate Rn (x) by
x

n α−1

|Rn (x)| ≤ |rn | |x| (1 + t) dt .
0
Given x ∈ (−1, 1), choose ρ > 1 such that ρ|x| < 1. Then
|rn | |xn | ≤ C(ρ|x|)n → 0 as n → ∞ .
It follows that Rn (x) → 0 in the interval −1 < x < 1 as n → ∞ , which

proves Newton’s binomial series expansion
∞
α k
α
(1 + x) = 1 + x , −1 < x < 1 .
k
k=1
In fact, the proof shows that for α > 0 the series converges uniformly
to (1 + x)α in the closed interval −1 ≤ x ≤ 1. To see this, note that
r → 0 for every α ∈ R (cf. Chapter 5, Exercise 16), whereas the integral

n1 α−1 dt converges for α > 0. Thus R (x) → 0 uniformly in [−1, 1].
−1 (1 + t) n
To give an explicit example of Newton’s series, let α = − 12 . Then

1 1 3 2k−1
−2 −2 −2 · · · − 2
=
k k!
1 · 3 · 5 · · · (2k − 1) (−1)k (2k)!
= (−1)k k
= ,
2 k! 22k (k!)2
and the series reduces to
∞

−1/2 (−1)k (2k)!
(1 + x) =1+ xk , |x| < 1 .
22k (k!)2
k=1
3.6. Composition of power series

Taylor’s formula offers a systematic way to verify that a given func-
tion has a power series expansion and to find the expansion. However, the
calculation of successive derivatives is often a daunting prospect and other
methods are likely to be more practical. In particular, if two functions f
and g have power series expansions about the origin, it is useful to know
that the same is true for their product f g and quotient f /g (if g(0) = 0),
and to be able to calculate at least the first few terms of those expansions
conveniently from the power series for f and g.
Suppose now that
∞

f (x) = an x n for |x| < R1 and
n=0
∞
g(x) = bn xn for |x| < R2 .
n=0
Then, as we saw in Section 3.3, Mertens’ theorem can be used to prove that
the product has the power series expansion
∞

n
f (x)g(x) = c n xn , where cn = an−k bk ,
n=0 k=0
convergent at least in the interval |x| < R = min{R1 , R2 }. In fact, the radius
of convergence R may well be larger than both R1 and R2 . For instance,
consider the pair of functions
1+x 1−x
f (x) = and g(x) = .
1−x 1+x
Although f and g have power series expansions of radii R1 = R2 = 1, their

product f (x)g(x) ≡ 1 has a (trivial) expansion with radius of convergence
R = ∞.
Similarly, we will be able to show that the quotient f /g has a power
series development in some interval |x| < R where g(x) = 0, but the radius
of convergence R may be smaller than both R1 and R2 . To give a simple
example, let f (x) ≡ 1 and g(x) = 1 + x2 . Then R1 = R2 = ∞ and g(x) > 0
for all x ∈ R, yet f /g has the expansion
1
= 1 − x2 + x4 − x6 + . . . ,
1 + x2
with radius of convergence R = 1.
In order to show in general that f /g can be represented by a power
series, it will suffice to give a proof for the reciprocal 1/g, since f /g =
f · (1/g). Supposing for convenience that g(0) = 1, we can view 1/g as the
postcomposition of g − 1 with the function
1
ϕ(y) = = 1 − y + y2 − y3 + . . . , |y| < 1 .
1+y
Thus, if |x| is sufficiently small we find
1 1
= = 1 − (b1 x + b2 x2 + . . . )
g(x) 1 + (b1 x + b2 x2 + . . . )
+ (b1 x + b2 x2 + . . . )2 − . . .
= 1 − b1 x + (b21 − b2 )x2 + (−b31 + 2b1 b2 − b3 )x3 + . . . .
In this way we can calculate, for instance,
x 1
= = 1 + 16 x2 + 7 4
360 x + 31
15120 x
6
+ ... .
sin x 1 − 6 x + 120
1 2 1 4
x − ...
Of course, these formal manipulations need to be justified. In doing so,

we will consider composition of arbitrary power series, since the extra gen-
erality does not complicate the argument. For convenience we will assume
that both functions vanish at the origin. Suppose, then, that
∞
∞

f (x) = an x n and g(x) = bn xn
n=1 n=1
are power series developments in some interval |x| < r. Then |g(x)| < r if x
is sufficiently small, say |x| < ρ ≤ r, and the composition
∞

h(x) = f (g(x)) = an [g(x)]n
n=1
is defined. Each of the functions [g(x)]n is a product of power series and so

has its own expansion
∞

[g(x)]n = bnm xm , n = 1, 2, . . . ,
m=1
where bnm = 0 for all m < n. Therefore, the composite function has the
form
∞
∞
(3) h(x) = an bnm xm .
n=1 m=1
Interchanging the order of summation, we find

∞ ∞ ∞
m

m
h(x) = an bnm x = an bnm xm ,
m=1 n=1 m=1 n=1
since bnm = 0 for all n > m. This is the desired representation of h as

a power series. Moreover, it validates the formal manipulation previously
used to calculate the coefficients in the expansion.
However, we have to justify the interchange of summation. For this pur-
pose we can apply Cauchy’s double series theorem (see Chapter 1, Section
1.5), provided we are able to show that the double series (3) is absolutely
convergent. This is not clear, but the proof can be completed by the follow-
ing argument. Recall that a power series is absolutely convergent within its
interval of convergence. Let
∞
∞

F (x) = |an |xn and G(x) = |bm ||x|m ,
n=1 m=1
and choose a positive number ρ1 ≤ ρ small enough that |G(x)| < r whenever
|x| < ρ1 . Then the composition
∞

H(x) = F (G(x)) = |an |[G(x)]n
n=1
is defined for |x| < ρ1 . Writing

∞

[G(x)] = n
Bnm |x|m ,
m=1
we infer in particular that the double series

∞
∞
|an |Bnm |x|m
n=1 m=1
is convergent. But since |bnm | ≤ Bnm , this implies that the double series
(3) is absolutely convergent for |x| < ρ1 , which completes the proof that the
composite function h(x) = f (g(x)) can be expanded in power series in some
neighborhood of the origin, and that the formal method for calculating the
coefficients is valid.
Finally, it must be acknowledged that from the viewpoint of complex
function theory, some paradoxes encountered in the real domain are readily
explained, and the theory becomes much more transparent. In particular,
the tools of complex analysis provide easy proofs of the existence of power
series developments for the product, quotient, and composition of two func-
tions known to have power series expansions. However, the actual calcula-
tion of coefficients in those developments involves essentially the same tricks
discussed here for real power series. Thus the argument presented in this
section, justifying the formal calculation of coefficients in the composition
of power series on the basis of Cauchy’s double series theorem, is also of
interest for complex power series and is readily adapted to that setting.
3.7. Euler’s sum

Our purpose is to evaluate the sum of the infinite series
1 1 1 π2
1+ + + + · · · = .
22 32 42 6
It is easy to see that the series converges, but more difficult to show that
its sum is π 2 /6. The problem goes back at least to the year 1690, when
Jacob Bernoulli (1654–1705) determined that the series converges to a sum
less than 2 but was unable to evaluate it precisely. He and his younger
brother Johann Bernoulli (1667–1748), his eventual successor to a chair at
the University of Basel (Switzerland), drew attention to the problem, which
became known as the Basel problem. Johann Bernoulli was the principal
teacher of Leonhard Euler (1707–1783), who entered the University of Basel
at the age of 14.
Euler became interested in the Basel problem and sought to calculate
the sum numerically. The series converges too slowly for accurate numerical
evaluation, but in 1731 Euler found a clever transformation (see Exercise
17) to express it in the form
∞
∞

1 2 1
2
= (log 2) + 2 ,
k k 2k
2
k=1 k=1
thereby gaining a series that converges much more rapidly. To calculate
log 2 he took the Taylor series expansion
x2 x3
− log(1 − x) = x + + + ...
2 3
1
and set x = 2 to obtain
1 ∞
1 1 1
log 2 = + + + ··· = ,
2 8 24 k2k
k=1
a series that converges rapidly enough for effective calculation. As a result,

Euler was able to compute
∞
1
= 1.644934 . . . ,
k2
k=1
accurate to 6 decimal places. Later he used what is now called the “Euler-
Maclaurin summation formula” (developed in Chapter 11 of this book) to
compute the sum to 20 decimal places. Then in 1735 Euler solved the Basel
problem, making the sensational discovery that the sum of the series is π 2 /6.
Euler’s first derivation of the sum π 2 /6 was not at all rigorous. Other
mathematicians criticized his method, but they could hardly question the
result because π 2 /6 matched the true sum numerically to 20 decimal places!
(Euler’s “proof ” is described at the end of Section 8.5.)
Soon after his initial triumph, Euler evaluated the more general sums
∞
2n−1 π 2n
1 n+1 2
= (−1) B2n , n = 1, 2, 3, . . . ,
k 2n (2n)!
k=1
in terms of the Bernoulli numbers Bn , defined by the power-series expansion
1 ∞
x
= Bn xn , |x| < 1 .
e −1
x n!
n=0
A calculation gives the values
B2 = 1
6 , B4 = − 30
1
, B6 = 1
42 , B8 = − 30
1
, B10 = 5
66 .
Thus
∞
∞
∞

1 π2 1 π4 1 π6
= , = , = , etc.
k2 6 k4 90 k6 945
k=1 k=1 k=1
We will evaluate these more general sums in Chapter 11, which is devoted
to Bernoulli numbers and their applications.

Many proofs of Euler’s basic relation 1/k 2 = π 2 /6 have been found,
but none is really simple. The standard proofs use advanced methods such
as Fourier series or complex function theory. Here we present two relatively

elementary derivations. The first is a simplified version of a proof given by
Boo Rim Choe [4], which appeals to Newton’s binomial series and the Wallis
product formula. The second method is an ingenious argument based on an
elementary trigonometric sum. It was found by the Russian twins Yaglom
and Yaglom ([18]; see [19], Problem 145), and rediscovered independently by
Holme [8] and Papadimitriou [11]. The paper by Stark [14] gives references
to other elementary proofs.
First Proof. We begin with the observation that

∞
∞
∞

1 1 1
2
= 2
+ ,
n (2k) (2k + 1)2
n=1 k=1 k=0
or
∞ ∞
3 1 1
= ,
4 n2 (2k + 1)2
n=1 k=0
so it will be equivalent to show that

∞
1 π2
= .
(2k + 1)2 8
k=0
The proof will use Newton’s binomial series

∞
α
(1 + x) = 1 + α
xn , |x| < 1 ,
n
n=1
in the special case where α = −1/2. Our previous calculation (cf. Section
3.5) led to the expression

− 12 (−1)n (2n)!
= , n = 1, 2, . . . .
n 22n (n!)2
Therefore, after replacing x by −t2 we arrive at the expansion

∞
(2n)!
(1 − t2 )− 2 = 1 +
1
cn t2n , |t| < 1 , with cn = .
22n (n!)2
n=1
Now recall the Wallis product formula in the form
22n (n!)2 √
lim √ = π.
n→∞ (2n)! n
As a consequence, we have the asymptotic formula

1
cn ∼ √ , n → ∞.
πn
The next step is to integrate the power series term by term to obtain
x ∞
−1 2 − 12 1
sin x = (1 − t ) dt = x + cn x2n+1 , |x| < 1 .
0 2n + 1
n=1
√
Since cn ∼ 1/ πn, this last power series converges at x = 1. In other words,
∞
1
cn < ∞ .
2n + 1
n=1
Thus by the Weierstrass M-test, the series converges uniformly in the interval
[−1, 1], and so the sum is continuous there. We may therefore infer that
∞
−1 1
sin x = x + cn x2n+1
2n + 1
n=1
throughout the closed interval [−1, 1].
Now substitute x = sin θ to arrive at the identity
∞
1
θ = sin θ + cn sin2n+1 θ , |θ| ≤ π
2 .
2n + 1
n=1
The last series converges uniformly in the interval [−π/2, π/2], so we may
integrate term by term:
π/2 π/2 ∞
π/2
1
θ dθ = sin θ dθ + cn sin2n+1 θ dθ .
0 0 2n + 1 0
n=1
Now recall the formula
π/2
2 · 4 · 6 · · · (2n) 22n (n!)2
I2n+1 = sin2n+1 θ dθ = = ,
0 1 · 3 · 5 · · · (2n + 1) (2n + 1)!
obtained from the reduction formula and used in the proof of the Wallis
product formula (see Chapter 2, Section 2.5). Comparing this with the
formula for cn , we see that
1
cn I2n+1 = ,
2n + 1
so that the result of the term by term integration reduces to
∞
π2 1
=1+ .
8 (2k + 1)2
k=1
But as we noted at the outset, this relation is equivalent to
∞
1 π2
= .
n2 6
n=1
Second Proof. Here is a totally different approach, based upon the curious
identity

n
2 kπ n(2n − 1)
(4) cot = , n = 1, 2, . . . .
2n + 1 3
k=1
Deferring a proof of (4), let us show first how it leads to an evaluation of

Euler’s sum. Start with the inequality
π
sin x < x < tan x , 0<x< 2 .
(See Chapter 4, Section 4.1.) Take reciprocals and squares to infer that
cot2 x < 1/x2 < csc2 x = 1 + cot2 x , 0<x< π

2 .
Insert the numbers x = kπ/(2n + 1), for k = 1, 2, . . . , n, and take sums to

arrive at the inequality

(2n + 1)2 1
n n n
2 kπ 2 kπ
cot < <n+ cot .
2n + 1 π2 k2 2n + 1
k=1 k=1 k=1
In view of (4), this reduces to
(2n + 1)2 1
n
n(2n − 1) n(2n − 1)
< 2 2
<n+ .
3 π k 3
k=1
Now multiply the inequality by π 2 /(4n2 ) and let n → ∞ to conclude that

∞ ∞
π2 1 π2 1 π2
≤ ≤ , or = .
6 k2 6 k2 6
k=1 k=1
A proof of the identity (4) can be based on the de Moivre formula
(cos θ + i sin θ)n = cos nθ + i sin nθ ,

√
where i = −1. The de Moivre formula can be recast in the form
n

n n n n k
cos nθ + i sin nθ = sin θ(cot θ + i) = sin θ i cotn−k θ .
k
k=0
Equating imaginary parts, we find that

n n n
n
sin nθ = sin θ cot n−1
θ− cot n−3
θ+ cot n−5
θ − ... .
1 3 5
If n = 2m + 1 is an odd integer, this relation takes the more specific form

(5) sin(2m + 1)θ = sin2m+1 θ Pm (cot2 θ) , 0<θ< π
2 ,
where Pm is the polynomial of degree m defined by

2m + 1 m 2m + 1 m−1 2m + 1 m−2
Pm (x) = x − x + x − · · · + (−1)m .
1 3 5
Now observe that sin2m+1 θ = 0 for 0 < θ < π/2, whereas sin(2m + 1)θ
vanishes at the m distinct points
kπ
θ= , k = 1, 2, . . . , m .
2m + 1
In view of the relation (5), we conclude that Pm (x) = 0 at the m distinct
points
2 kπ
xk = cot , k = 1, 2, . . . , m .
2m + 1
But from the coefficients of the polynomial Pm (x), we see that
2m+1

m
kπ
m
m(2m − 1)
cot2
= xk = 2m+1
3
= ,
2m + 1 1
3
k=1 k=1
which proves (4).
The last step of the argument used the fact that a monic polynomial of
degree m with roots x1 , x2 , . . . , xm has the factorization
xm + am−1 xm−1 + · · · + a1 x + a0 = (x − x1 )(x − x2 ) · · · (x − xm ) .
By expanding the product and equating coefficients of xm−1 , one sees that
the sum of the roots is
m
xk = −am−1 .
k=1
Apostol [1] generalized this proof of (4) to show that

m
kπ 24n−1
cot 2n
= (−1)n+1 B2n m2n + O m2n−1 , m → ∞.
2m + 1 (2n)!
k=1
In the same way as before, this formula then leads to the general result
∞
2n−1 π 2n
1 n+1 2
= (−1) B2n , n = 1, 2, 3, . . . .
k 2n (2n)!
k=1
Yaglom and Yaglom [19] indicate a similar generalization of the argument.

Leonhard Euler (1707–1783) was the greatest mathematician of the 18th

century and one of the greatest ever. His work spanned all branches of math-
ematics, pure and applied. The sheer quantity of his output is astounding.
A project to publish his collected works (Opera Omnia) was started in 1911
and has run to over 70 volumes but is still far from complete. Euler made
lasting contributions to many areas of mathematics, including differential
equations, elliptic integrals, infinite series and products, number theory, al-
gebra, hydrodynamics, optics, and classical geometry. His work initiated
major fields of investigation such as calculus of variations, graph theory,
and combinatorial topology. Born and educated in Basel, he worked for
most of his life in St. Petersburg (where he is buried) and Berlin. At age 31
he suffered a severe illness that left him blind in one eye, and eventually he
became totally blind. His solution of the Basel problem was his first work
to gain international recognition.
The recent book by Varadarajan [15] gives a full account of Euler’s life
and mathematical work. Euler’s studies of the Riemann zeta function
∞
1
ζ(x) = , x > 1,
nx
n=1
as it came to be called, are described in an article by Ayoub [2]. We will

develop some of this material in Chapter 10.
3.8. Continuous nowhere differentiable functions

In the early days of calculus, the prevailing wisdom was that a contin-
uous function must have a derivative at “almost all” points. It was hard
to imagine otherwise. Thus it came as a rude shock when Karl Weierstrass
(1815–1897) produced an example of a continuous function that is not differ-
entiable at any point! Writing in 1872, Weierstrass [17] reported having met
a former student of Bernhard Riemann(1826–1866) who told him Riemann
had claimed as early as 1861 that the continuous function
∞
sin(k 2 x)
(6) f (x) =
k2
k=1
is nowhere differentiable. However, Riemann had not published a proof and

Weierstrass was unable to verify it, so he produced his own example. Perhaps
inspired by Riemann’s claim, Weierstrass [17] constructed the function
∞

(7) f (x) = bk cos(ak πx) ,
k=0
where a is an odd integer and b is a number in the interval 0 1 + 3π/2. Since 0 1 + 3π/2 is needed for the more technical proof that the function is
nowhere differentiable. We will come to the details presently.
After Weierstrass published his construction, many other examples were
devised. One idea was to look at a general function of the form
∞

f (x) = bk ϕ(ak x) ,
k=0
where ϕ is a suitable periodic function. In a scholarly paper written in 1918,
Knopp [9] discussed this approach among others, and gave many references
to earlier literature. A few years later, van der Waerden [16] discovered a
relatively simple example of this type. In modified form presented by Rudin
[12], van der Waerden’s example is
∞

(8) f (x) = (3/4)k ϕ(4k x) ,
k=0
where the function ϕ is defined by ϕ(x) = |x| for −1 ≤ x ≤ 1 and is extended

periodically to the whole line so that ϕ(x + 2) = ϕ(x).
It would come to light, however, that Bernard Bolzano (1781–1848) had
already constructed an example in 1834, although his discussion was not
entirely precise and he claimed the nonexistence of a derivative only at a
countable dense set of points. Bolzano’s work remained obscure and was
not properly published until 1921, at which point his function was found to
be nowhere differentiable. His construction is geometric. A curve undergoes
successive modifications to produce in the limit a curve that does not have
a well determined tangent direction at any point. A more precise exposition
was given by Kowalewski [10].
Presumably unaware of Bolzano’s construction, Riemann and Weier-
strass used much the same idea. Each new term of the series has smaller am-
plitude but much higher frequency than the preceding terms, so the graphs
of the partial sums become increasingly “wiggly” and it seems plausible that
the limiting function will have an “infinitely wiggly” graph at each point.
The problem is to make these vague notions precise.
To prove that the Weierstrass function (7) is nowhere differentiable,
write f (x) = sn (x) + rn (x), where

n−1
sn (x) = bk cos(ak πx)
k=0
is the nth partial sum of the series and rn (x) is the remainder. For an
arbitrarily chosen point x, let
sn (x + h) − sn (x) rn (x + h) − rn (x)
φn (h) = and ψn (h) =
h h
denote the respective difference quotients. By the mean value theorem,

n−1
φn (h) = sn (t) = − ak πbk sin(ak πt)
k=0
for some point t between x and x + h. It follows that

n−1
(ab)n − 1 π(ab)n
(9) |φn (h)| ≤ π (ab)k = π < .
ab − 1 ab − 1
k=0
Now hold n fixed and let m be the integer closest to an x, so that |an x −
m| ≤ 12 . (In case of ambiguity, when an x is a half-integer, take for instance
m = an x + 12 .) Let λn = an x − m and define the increment
1 − λn m+1
hn = n
= − x,
a an
so that 0 < hn ≤ 32 a−n . Since a is an odd integer, we see that

cos ak π(x + hn ) = cos ak−n (m + 1)π = (−1)m+1 for k ≥ n .
Also, we can write

cos ak πx = cos ak−n (m + λn )π

= cos ak−n mπ cos ak−n λn π

= (−1)m cos ak−n λn π , k ≥ n.
Combining the last two formulas, we have
1 k
∞
ψn (hn ) = b cos ak π(x + hn ) − cos ak πx
hn
k=n
(−1)m+1 k
∞
= b 1 + cos(ak−n λn π) .
hn
k=n
Every term in this infinite series is nonnegative, so we can discard all but
the first term to obtain the estimate
bn bn
|ψn (hn )| ≥ [1 + cos(λn π)] ≥ ≥ 23 (ab)n ,
hn hn
since |λn | ≤ 1
2 and 0 < hn ≤ 32 a−n .
The final step is to invoke the estimate (9) and conclude that

f (x + hn ) − f (x)
= |φn (hn ) + ψn (hn )|
hn
≥ |ψn (hn )| − |φn (hn )|

n 2 π
≥ (ab) − .
3 ab − 1
Because ab > 1 + 3π/2, the lower bound is positive and we see that

f (x + hn ) − f (x)
→∞ as n → ∞ .
hn
But hn > 0 and hn → 0 as n → ∞, so this shows that the Weierstrass

function f has unbounded right-hand difference quotients at the point x.
Since x was chosen arbitrarily, we conclude in particular that Weierstrass’s
function is not differentiable at any point.
G. H. Hardy [7] showed that the requirement ab > 1 + 3π/2 can be
replaced by ab ≥ 1, but then the proof of nondifferentiability is more difficult.
A similar argument applies to van der Waerden’s function (8), but easier
estimates suffice to show that its difference quotients are unbounded at each
point. The details are outlined in Exercise 20.
And what about Riemann’s function? Is it nowhere differentiable as
Riemann claimed? In 1916, Hardy [7] showed that Riemann’s function (6)
is not differentiable at any irrational multiple of π or at any rational multiple
ξπ with ξ of the form 2p/(4q + 1) or (2p + 1)/(4q + 2), where p and q are
positive integers. There the problem lay until 1970, when Joseph Gerver
[5] made the unexpected discovery that Riemann’s function actually has
a derivative, whose value is always −1/2, at every point ξπ where ξ =
(2p+1)/(2q +1); in other words, at every multiple of π by a rational number
whose numerator and denominator are both odd. Gerver [6] then went on
to show that Riemann’s function is not differentiable at any point ξπ with
ξ = 2p/(2q + 1) or (2p + 1)/(2q), thereby arriving at a full solution of the
problem.
A. Smith [13] gave a simpler proof of Gerver’s result that Riemann’s
function has a derivative at the specified points.
Finally, it should be mentioned that the existence of continuous nowhere
differentiable functions can be proved without actually constructing an ex-
ample. On the basis of a result from metric space theory known as the Baire
category theorem, it can be shown that many such functions exist. In fact,
it turns out to be very unusual, in a certain precise sense, for a continu-

ous function to have bounded one-sided difference quotients at any point!
Details may be found for instance in the book by Boas [3].
Exercises

1. Find the radius of convergence of the power series ∞ n! n
n=1 nn x .

2. Directly from the geometric series ∞ n=0 x = 1/(1 − x) , calculate the
n
sums
∞
x
(a) nxn = ,
(1 − x)2
n=1
∞
x(1 + x)
(b) n2 xn = ,
(1 − x)3
n=1
for all |x| < 1.

∞
3. (a) If the power series ∞ n
n=0 an x and
n
n=0 bn x have∞ respective radii
of convergence R1 and R2 , show that the power series n
n=0 an bn x has
radius of convergence R ≥ R1 R2 . Give an example where R1 = R2 = 0 but
R = ∞.
∞ n
(b) If a power series n=0 an x has only a finite number of distinct
coefficients an and is not a polynomial, show that it has radius of convergence
R = 1.
4. For |x| ≤ 1, verify the inequality

x − t

1 + t ≤ |x| for all t between 0 and x.
5. Show that
1+x
1
log = x + 13 x3 + 15 x5 + . . . , |x| < 1 .
2 1−x
Use the result to calculate log 11
9 = 0.20067 . . . by choosing only the first
3 nonzero terms. Estimate the remainder in the Taylor series expansion to
show that this numerical value is correct to 5 decimal places. Finally, check
the number on a calculator or computer and record it to 8 or 10 decimal
places.
6. Consider the function f (x) = e−1/x for x > 0 and f (x) = 0 for x ≤ 0.
Show that f is of class C ∞ on R , and that f (n) (0) = 0 for n = 0, 1, 2, . . . .
Conclude that f cannot be expanded into Taylor series at the origin.
Exercises 103
7. Here is another way to prove the validity of Newton’s binomial series

expansion
∞
α k
α
(1 + x) = 1 + x , |x| < 1 .
k
k=1
First prove that the series converges for |x| < 1, and let f (x) denote its sum.
Verify the identity

α−1 α−1 α
+ =
k k−1 k
and use it to show that (1 + x)f (x) = αf (x). Then solve the differential
equation to conclude that f (x) = (1 + x)α .
8. Specialize Newton’s binomial series to verify the expansion
∞

√ (−1)k+1 (2k)! k
1+x=1+ x , |x| < 1 .
22k (k!)2 (2k − 1)
k=1
9. Apply Abel’s theorem to show that
∞
(−1)k (2k)! √
1+ = 1/ 2 .
22k (k!)2
k=1
10. Show that

∞
(2k)! π
(a) 1+ = and
22k (k!)2 (2k + 1) 2
k=1
∞
(−1)k+1 (2k)! √
(b) 1+ = 2.
2 (k!) (2k − 1)
2k 2
k=1
11. Suppose that the functions u1 (x), u2 (x), . . . are continuous on a set
E ⊂ R, and there is a constant M such that

n

uk (x) ≤ M , n = 1, 2, . . . ,

k=1
for all x ∈ E. Let {cn } be a sequence of positive numbers that decrease to

zero: c1 ≥ c2 ≥ · · · ≥ cn > 0 and cn → 0 as n → ∞. Prove that the series
∞
n=1 cn un (x) converges uniformly on E.
12. (a) Use the identity
cos(k − 12 )x − cos(k + 12 )x = 2 sin 12 x sin kx
to calculate the sum

n
cos 12 x − cos(n + 12 )x
sin kx = .
k=1
2 sin 12 x
(b) If c1 ≥ c2 ≥ · · · ≥ cn > 0 and cn → 0 as n → ∞, show that the series

∞

cn sin nx
n=1
is uniformly convergent in the interval [δ, 2π − δ] for each δ > 0.

13. If the infinite series ∞k=1 ak convergesand the sequence {bk } is mono-
tonic and bounded, prove that the series ∞ k=1 ak bk converges.
14. If two functions f and g have power series expansions

∞
∞

f (x) = an xn and g(x) = bn xn , |x| < R ,
n=0 n=0
apply Mertens’ theorem to prove that their product h = f g is given by the

power series
∞

n
n
h(x) = cn x , where cn = an−k bk ,
n=0 k=0
convergent in the interval |x| < R .

∞ ∞
15. (a) Suppose the series n n=0 a n and n=0 bn converge to sums A and
∞
B, respectively. Let cn = k=0 an−k bk . If n=0 cn converges to sum C, use
Abel’s theorem to prove that C = AB.
∞
(b) If the series ∞
n=0 an and n=0 bn are absolutely convergent, prove
that ∞ c
n=0 n is absolutely convergent and
∞
∞ ∞

|cn | ≤ |an | |bn | .
n=0 n=0 n=0
16. The Catalan numbers Cn are defined recursively by C0 = 1 and

n−1
Cn = Cn−1−k Ck , n = 1, 2, . . . .
k=0
Exercises 105
Thus C1 = C02 = 1, C2 = 2C0 C1 = 2, C3 = 2C0 C2 + C12 = 5, C4 = 14,

C5 = 42, etc. Derive the formula

1 2n (2n)!
Cn = = , n = 1, 2, . . . ,
n+1 n (n + 1)(n!)2
and check it by induction.

Suggestion. Consider the generating function G(x) = ∞ n
n=0 Cn x and show
that
√
1 − 1 − 4x 1
xG(x) = G(x) − 1 ,
2
so that xG(x) = , |x| < .
2 4
17. Carry out the following steps
to derive the formula that Euler used for
numerical calculation of the sum 1/k 2 .
(a) Observe that
1 1
∞

∞

log(1 − x) xk−1 1
− dx = dx = .
0 x 0 k k2
k=1 k=1
Justify the term-by-term integration.

(b) Split the integral and change variables to write
∞
1/2 1/2
1 log(1 − x) log t
=− dx − dt .
k 2
0 x 0 1−t
k=1
(c) Integrate by parts to obtain

∞
1/2
1 log(1 − x)
= (log 2) − 2
2
dx
k2 0 x
k=1
∞

2 1
= (log 2) + 2 .
k 2 2k
k=1
18. The function

∞
xn
f (x) = , |x| ≤ 1 ,
n2
n=1
is known as the dilogarithm. Generalize the formula of Exercise 17 by show-
ing that
∞
1
= (log x) log(1 − x) + f (x) + f (1 − x) , 0 < x < 1,
k2
k=1
a result also known to Euler.

Hint. Compute the derivative of the right-hand side.
19. Calculate the integral

∞
x π2
dx = .
0 ex − 1 6
Hint. Expand the integrand into a geometric series.
*20. (a) Show that the van der Waerden function f defined by (8) is
continuous on the whole real line.
(b) Show that the function ϕ in (8) satisfies the Lipschitz condition
|ϕ(s) − ϕ(t)| ≤ |s − t|.
(c) For any fixed number x ∈ R and any integer n > 0, define the
number δn = ± 12 4−n , where the sign is chosen so that no integer lies strictly
between 4n x and 4n (x + δn ). Next define the difference quotients
ϕ(4k (x + δn )) − ϕ(4k x)
Qk = , k = 0, 1, 2, . . . ,
δn
and show that Qk = 0 for every k > n.
(d) Show that |Qk | ≤ 4k for 0 ≤ k ≤ n and that |Qn | = 4n .
(e) Show that the difference quotients of f satisfy

f (x + δn ) − f (x) n−1

≥ 3n 4−n Qn −
3k 4−k Qk ≥ 12 (3n + 1) ,
δn
k=0
and conclude that f is nowhere differentiable.

Reference. Additional details can be found in Rudin [12].
21. Use the identity (1 + x)n (1 + x)n = (1 + x)2n to prove that

n 2

n 2n
= , n = 1, 2, . . . .
k n
k=0
22. Compare power series to prove that

1+x
e2x ≤ , 0 ≤ x < 1.
1−x
23. The theory of power series can be extendedto the complex plane.
For any complex coefficients cn the power series ∞ n
n=0 cn z in a complex
variable z can be shown to converge for every point in some disk |z| < R
References 107
and to diverge whenever |z| > R. (Here, as for real power series, we adopt
the convention that 0 ≤ R ≤ ∞.) For instance, this analysis allows us to
define the exponential function
z2 z3
ez = 1 + z + + + ...
2! 3!
for all complex numbers z. Use this definition to derive Euler’s formula
eix = cos x + i sin x , x ∈ R.
Note in particular that eiπ = −1.
References
[1] T. M. Apostol, “Another elementary proof of Euler’s formula for ζ(2n) ”, Amer.
Math. Monthly 80 (1973), 425–431.
[2] Raymond Ayoub, “Euler and the zeta function”, Amer. Math. Monthly 81
(1974), 1067–1086.
[3] R. P. Boas, A Primer of Real Functions, Third edition, Mathematical Associa-
tion of America, Washington, D.C., 1981.

[4] B. R. Choe, “An elementary proof of ∞ 2 2
n=1 1/n = π /6 ”, Amer. Math. Monthly
94 (1987), 662–663.
[5] Joseph Gerver, “The differentiability of the Riemann function at certain rational
multiples of π”, Amer. J. Math. 92 (1970), 33–55.
[6] Joseph Gerver, “More on the differentiability of the Riemann function”, Amer.
J. Math. 93 (1971), 33–41.
[7] G. H. Hardy, “Weierstrass’s non-differentiable function” Trans. Amer. Math.
Soc. 17 (1916), 301–325.
∞
[8] F. Holme, “En enkel beregning av k=1 1/k2 ”, Nordisk Mat. Tidskr. 18 (1970),
91–92; 120. [Norwegian, English summary]
[9] Konrad Knopp, “Ein einfaches Verfahren zur Bildung stetiger nirgends differen-
zierbarer Funktionen”, Math. Zeitschrift 2 (1918), 1–26.
[10] Gerhard Kowalewski, “Über Bolzanos nichtdifferenzierbare stetige Funktion”,
Acta Math. 44 (1923), 315–319.
∞
[11] I. Papadimitriou, “A simple proof of the formula k=1 k−2 = π 2 /6 ”, Amer.
Math. Monthly 80 (1973), 424–425.
[12] Walter Rudin, Principles of Mathematical Analysis, Third edition, McGraw–
[13] A. Smith, “The differentiability of Riemann’s functions”, Proc. Amer. Math.

Soc. 34 (1972), 463–468.

[14] E. L. Stark, “The series ∞
k=1 k
−s
, s = 2, 3, 4, . . . once more”, Math. Magazine
47 (1974), 197–202.
[15] V. S. Varadarajan, Euler Through Time: A New Look at Old Themes, American
Mathematical Society, Providence, R. I., 2006.
[16] B. L. van der Waerden, “Ein einfaches Beispiel einer nichtdifferenzierbaren
stetigen Funktion”, Math. Zeitschrift 32 (1930), 474–475.
[17] Karl Weierstrass, “Über continuirliche Functionen eines reellen Arguments, die
für keinen Werth des letzteren einen bestimmten Differentialquotienten besitzen”,
Königlich Preussischen Akademie der Wissenschaften, 1872; Mathematische Werke,
Band II, Preussischen Akademie der Wissenschaften, Berlin, 1895, 71–74.
[18] A. M. Yaglom and I. M. Yaglom, “An elementary derivation of the formulas of
Wallis, Leibniz, and Euler for the number π ”, Uspekhi Mat. Nauk 8 (1953), no. 5
(57), 181–187. [in Russian]
[19] A. M. Yaglom and I. M. Yaglom, Challenging Mathematical Problems with
Elementary Solutions, Gos. Izdat. Tekhn.-Teor. Lit. (Moscow, 1954) [in Rus-
sian]; English translation, Holden–Day (San Francisco, 1967); reprinted by Dover
Publications, Mineola, NY, 1987.
[20] Robert M. Young, Excursions in Calculus: An Interplay of the Continuous and
the Discrete, Mathematical Association of America, Washington, D.C., 1992.
Chapter 4
Inequalities
Inequalities are essential tools for analysis. Proofs of theorems or derivations

of formulas often depend on estimates that are conveniently performed by
appeal to general inequalities. The primary goal of this chapter is to develop
some basic inequalities, involving sums and integrals, that are commonly
applied in analysis.
4.1. Elementary inequalities

We begin by recalling three elementary but extremely useful inequalities.
The first is
1 + x ≤ ex , −∞ < x < ∞ ,
or equivalently
log x ≤ x − 1 , 0 < x < ∞.
The graphs in Figure 1 illustrate the two forms of the inequality.
The proof is quite easy. The function f (x) = ex − x − 1 has the value
f (0) = 0 and derivative f (x) = ex − 1, with f (0) = 0. But ex is an
increasing function, so f (x) > 0 for x > 0 and f (0) < 0 for x < 0. Thus
f (x) decreases to zero for x < 0 and increases from zero for x > 0, so that
f (x) > 0 for all x = 0. In other words, 1 + x < ex for all x = 0.
The next inequality is
2 π
x < sin x < x , 0<x< ,
π 2
as illustrated by Figure 2.
109
110 4. Inequalities
Figure 1. The inequalities 1 + x ≤ ex and log x ≤ x − 1.
0 π/2
Figure 2. The inequality 2x/π < sin x < x.
We will give a proof based directly on the geometric definitions of sine

d
and cosine, not assuming the differentiation formula {sin x} = cos x but
dx
deriving it along the way. The proof proceeds by comparing areas of the
two right triangles and the circular sector in Figure 3.
This shows that

1
2 cos θ sin θ < 12 θ < 1
2 tan θ ,
4.1. Elementary inequalities 111
θ
0 1
Figure 3. The inequality cos θ sin θ < θ < tan θ.
so that θ < tan θ and

sin θ 1 π
cos θ < < , 0<θ< .
θ cos θ 2
It now follows from the squeeze principle that
sin x
lim = 1,
x→0 x
which implies, in view of the half-angle formula 1 − cos x = 2 sin2 (x/2), that
1 − cos x
lim = 0.
x→0 x
Hence the addition formula for the sine function gives

sin(x + h) − sin x cos h − 1 sin h
= sin x + cos x → cos x
h h h
d
as h → 0. In other words, {sin x} = cos x. Now observe that
dx

d sin x x cos x − sin x π
= < 0, 0<x< ,
dx x x2 2
sin x
since x < tan x. Thus the function g(x) = decreases from g(0) = 1 to
x
g(π/2) = 2/π , so that
2 sin x π
< < 1, 0<x< ,
π x 2
which was to be proved.
112 4. Inequalities
Our third inequality is
(1 + x)p ≥ 1 + px , p ≥ 1 , x > −1 .
If p is a positive integer and x ≥ 0, the inequality is a simple consequence

of the binomial theorem. An inductive proof (see Exercise 1) extends its
validity to x > −1 when p = 1, 2, . . . . In the general case where p need not
be an integer, the inequality can be proved by calculus. Take p > 1 and
define the function
h(x) = (1 + x)p − px − 1 ,
with h(0) = 0 and derivative

h (x) = p (1 + x)p−1 − 1 .
Then h (x) < 0 for −1 < x < 0, whereas h (x) > 0 for x > 0. Since h(0) = 0,
it follows that h(x) ≥ 0 for all x > −1, which is the desired result.
4.2. Cauchy’s inequality

Cauchy’s inequality is well known and is one of the most widely used of
all inequalities. It can be stated as follows.
Cauchy’s Inequality. For arbitrary real numbers a1 , a2 , . . . , an and b1 ,

b2 , . . . , bn ,
n n
1/2 n 1/2

ak bk ≤ 2
ak 2
bk .

k=1 k=1 k=1
Unless ak ≡ 0 or bk ≡ 0, equality occurs if and only if ak = λbk for some

real number λ and all k = 1, 2, . . . , n.
This famous result was recorded by Augustin-Louis Cauchy (1789–1857)

in his text Cours d’analyse for students at the École Royale Polytechnique in
Paris. The book, published in 1821, was a landmark in the history of math-
ematics because it was the first serious attempt at a rigorous development
of calculus.
Cauchy’s inequality can be recast in vector form. With the notation
a = (a1 , a2 , . . . , an ) and b = (b1 , b2 , . . . , bn ) for vectors in Rn , one defines
their dot product (or scalar product) by
a · b = a1 b1 + a2 b2 + · · · + an bn .

1/2
The norm of a is defined to be a = a21 + a22 + · · · + a2n , so that a·· a =
a2 . Cauchy’s inequality then says that |a · b| ≤ ab.
For a proof, let t be a real parameter and write
0 ≤ a − tb2 = (a − tb) · (a − tb) = a2 − 2t a · b + t2 b2 .
But if a quadratic polynomial At2 + Bt + C is nonnegative for all values of

the variable t, then B 2 − 4AC ≤ 0. Thus we conclude that |a · b| ≤ a b,
which is Cauchy’s inequality. If b = 0, equality occurs if and only if a−tb =
0, or a = tb for some real number t.
A more illuminating proof comes from Lagrange’s identity
2
1
n n n n n
ak bk = a2k b2k − (aj bk − ak bj )2 ,
2
k=1 k=1 k=1 j=1 k=1
which can be viewed as a quantitative form of Cauchy’s inequality since

it gives an explicit formula, obviously nonnegative, for the discrepancy
a2 b2 − (a · b)2 . Once discovered, Lagrange’s identity can be verified
by straightforward algebra.
Lagrange’s identity plays an important role in vector analysis. For vec-
tors a and b in R3 , the cross product (or vector product) is defined by
a × b = (a2 b3 − a3 b2 , a3 b1 − a1 b3 , a1 b2 − a2 b1 ) ,
so that
1
3 3
a × b = 2
(aj bk − ak bj ) =
2
(aj bk − ak bj )2 .
2
j<k j=1 k=1
Hence Lagrange’s identity reduces to (a · b)2 = a2 b2 − a × b2 when

n = 3. But it follows from the law of cosines that a·· b = ab cos θ, where
θ is the angle between the vectors a and b. Therefore, Lagrange’s identity
gives the formula a × b = ab sin θ.
An important generalization of Cauchy’s inequality was discovered in
1884 by Otto Hölder (1859–1937), who is known primarily for his work in
abstract algebra. Hölder stated his inequality in different form, but here is
the modern formulation.
1 1
Hölder’s Inequality. Let p and q be positive numbers such that p+q = 1.
Then for arbitrary real numbers a1 , a2 ,. . . , an and b1 , b2 , . . . , bn ,
n n 1/p 1/q
n

ak bk ≤ |ak | p
|bk | q
.

k=1 k=1 k=1
114 4. Inequalities
Unless ak ≡ 0 or bk ≡ 0, equality occurs if and only if ak bk has constant

sign and |ak |p = λ|bk |q for some number λ > 0 and all k = 1, 2, . . . , n.
Two positive numbers p and q such that 1p + 1q = 1 are called conjugate

indices. The standard proof of Hölder’s inequality makes use of Young’s
inequality
1 1
ab ≤ ap + bq , a > 0, b > 0,
p q
where p and q are conjugate indices. For p = q = 2 it is simply the inequality
(a − b)2 ≥ 0. Equality occurs in Young’s inequality if and only if ap = bq .
Here is a geometric proof of Young’s inequality. The area under the
curve y = xp−1 , for 0 ≤ x ≤ a, is given by the integral
a
1 p
xp−1 dx = a .
0 p
1 1
The inverse function is found to be x = y 1/(p−1) = y q−1 , since p + q = 1.
The area under the curve x = y q−1 , for 0 ≤ y ≤ b, is
b
1 q
y q−1 dy = b .
0 q
But it is clear geometrically (see Figure 4) that the sum of these two areas
is greater than or equal to the area of the rectangle 0 ≤ x ≤ a, 0 ≤ y ≤ b.
Hence
1 1
ab ≤ ap + bq ,
p q
with equality if and only if ap−1 = b, so that the curve y = xp−1 passes
through the vertex (a, b) of the rectangle. The condition for equality reduces
to ap = bq , as claimed.
y y
b
b
x x
a a
Figure 4. Young’s inequality: ap−1 b.

For a proof of Hölder’s inequality, it is enough to consider the case where

n
n
|ak |p = |bk |q = 1 ,
k=1 k=1
because the general case can then be deduced by normalization. After this
reduction, Young’s inequality gives
n
1 n
1
n
1 1

ak bk ≤ |ak | +
p
|bk |q = + = 1
p q p q
k=1 k=1 k=1
n
1/p n 1/q
= |ak |p |bk |q ,
k=1 k=1
which is Hölder’s inequality. A close examination of the proof shows that

equality occurs if and only if ak bk has constant sign and |ak |p = λ|bk |q for
all k. The details are left as an exercise.
In terms of a parameter p ≥ 1, the notion of norm of a vector a =
(a1 , a2 , . . . an ) can be generalized to

n 1/p
ap = |ak |p .
k=1
Previously we wrote a for a2 . Hölder’s inequality is compactly stated as

|a·· b| ≤ ap bq , where p and q are conjugate indices. With the definition
b∞ = max |bk | ,

1≤k≤n
Hölder’s inequality extends to the case p = 1, q = ∞, where the proof is

trivial. (For motivation of the notation b∞ , see Exercise 15.)
The term “norm” is justified by the relation
a + bp ≤ ap + bp , p ≥ 1,
known as the triangle inequality because of its geometric interpretation when

p = 2 and n = 2 or 3. For p = 2 the triangle inequality is a simple
consequence of Cauchy’s inequality:
a + b2 = (a + b) · (a + b) = a2 + 2a · b + b2

≤ a2 + 2 ab + b2 = (a + b)2 .
For p = 2 the result is more subtle and is known as Minkowski’s inequality.

116 4. Inequalities
Minkowski’s Inequality. For 1 ≤ p < ∞ and for arbitrary real numbers

a1 , a2 , . . . , an and b1 , b2 , . . . , bn ,
n 1/p n 1/p
n 1/p
|ak + bk | p
≤ |ak | p
+ |bk | p
.
k=1 k=1 k=1
Hermann Minkowski (1864–1909) is famous for his 4-dimensional “space-

time” non-Euclidean model of the universe, which served as a foundation for
the mathematical theory of relativity. He also did important work in number
theory, where he developed a well-known theory of “geometry of numbers”.
Minkowski’s inequality appeared in his book Geometrie der Zahlen, first
published in 1896.
The inequality is trivial for p = 1, since |ak + bk | ≤ |ak | + |bk | for all k.
To prove it for 1 < p < ∞, observe that

n
n
n
|ak + bk | ≤
p
|ak ||ak + bk | p−1
+ |bk ||ak + bk |p−1 ,
k=1 k=1 k=1
and apply Hölder’s inequality to each of the sums on the right-hand side.
Since the conjugate index of p is q = p/(p − 1), we find
n
n 1/p
n (p−1)/p
|ak ||ak + bk | p−1
≤ |ak | p
|ak + bk | p
,
k=1 k=1 k=1
and similarly

n
n 1/p
n (p−1)/p
|bk ||ak + bk | p−1
≤ |bk | p
|ak + bk | p
.
k=1 k=1 k=1
Thus we have arrived at the inequality

a + bpp ≤ (ap + bp ) a + bpp−1 .
Assuming that a + bp = 0 (otherwise there is nothing to prove), we now
divide both sides by a + bpp−1 to obtain Minkowski’s inequality.
All of the foregoing inequalities have immediate extensions to infinite
sequences, or to Euclidean spaces of infinite dimension. For 0 < p < ∞, a
sequence a = (a1 , a2 , . . . ) is said to belong to the space p if
∞

app = |ak |p < ∞ .
k=1
In this terminology, for instance, the extended form of Hölder’s inequality

says that if a ∈ p and b ∈ q , where p and q are conjugate indices, then the
sequence c = (c1 , c2 , . . . ) with ck = ak bk belongs to 1 and c1 ≤ ap bq .
4.3. Arithmetic–geometric mean inequality 117
4.3. Arithmetic–geometric mean inequality
For any set of positive numbers a1 , a2 , . . . , an , the arithmetic mean An

and the geometric mean Gn are defined by
a1 + a2 + · · · + an √
An = and Gn = n a1 a2 · · · an .
n
The arithmetic–geometric mean inequality says that the arithmetic mean is
larger unless all of the numbers are equal.
Theorem. For any positive numbers a1 , a2 , . . . , an , the inequality Gn ≤ An
holds, with equality if and only if a1 = a2 = · · · = an .
This fact has been known for a long time, and many clever proofs have
been devised. We will give Cauchy’s proof, which is particularly ingenious.
The first step is to apply an inductive argument to establish the inequality
in the special case where n is a power of 2. The result for intervening values
of n is then deduced by a method of descent.
For n = 1 there is nothing to prove, since G1 = A1 = a1 . For n = 2 the
inequality G2 ≤ A2 takes the form

a1 + a2 2
a1 a2 ≤
2
√ √
and is equivalent to the trivial relation ( a1 − a2 )2 ≥ 0, where equality
occurs only for a1 = a2 . The proof for n = 4 now follows by writing
a1 a2 a3 a4 = (a1 a2 )(a3 a4 ) and applying the result for n = 2:
2
a1 + a2 2 a3 + a4 2 a1 + a2 a3 + a4
(a1 a2 )(a3 a4 ) ≤ =
2 2 2 2
4
1 a1 + a2 a3 + a4 a1 + a2 + a3 + a4 4
≤ + = .
2 2 2 4
This last argument can be generalized to prove the inequality for all
indices of the form n = 2k , k = 1, 2, . . . . Proceeding inductively, suppose it
has been established for some index n = 2k , so that
k
a1 + a2 + · · · + a2k 2
a1 a2 · · · a2k ≤ .
2k
Then the inequality for n = 2k+1 is deduced by writing

a1 a2 · · · a2k+1 = (a1 a2 · · · a2k ) a2k +1 a2k +2 · · · a2k+1
k k
a1 + a2 + · · · + a2k 2 a2k +1 + a2k +2 + · · · + a2k+1 2
≤
2k 2k
k+1
a1 + a2 + · · · + a2k+1 2
≤ ,
2k+1
118 4. Inequalities
where the last step again used the result for n = 2. This proves the inequality
for all indices n = 2k .
If n is not a power of 2, then 2k−1 < n < 2k for some integer k. Apply
the inequality for index 2k to the numbers
a1 , a2 , . . . , an , An , An , . . . , An ,
where the arithmetic mean An appears m = 2k − n times. Then

2k
a1 + a2 + · · · + an + mAn
a1 a2 · · · an An m
≤
2k
2 k
(n + m)An k
= k
= An2 ,
2
k
or Gn n An m ≤ An 2 , which reduces to the desired inequality Gn ≤ An . An
inspection of the proof shows that equality occurs at each step precisely
when all of the numbers aj are equal.
4.4. Integral analogues

The inequalities of Hölder and Minkowski have clear analogues for inte-
grals. We will operate under the standing assumption that our functions are
Riemann integrable over the real line. Functions defined on a fixed subin-
terval can then be regarded as a special case. We will say that a function f
belongs to the class Lp , where 0 < p < ∞, if it has finite Lp norm:
∞ 1/p
f p = |f (x)| dx
p
< ∞.
−∞
If f is bounded, we say that f ∈ L∞ and define the norm
f ∞ = sup |f (x)| .
x∈R
In reality, the space Lp is customarily defined to consist of the larger class

of Lebesgue measurable functions with finite Lp norm. The Hölder and
Minkowski inequalities remain valid in that more general context, but the
theory of Lebesgue integrals is beyond the scope of this book and we will
confine ourselves to Riemann integrals.
In the terminology of Lp spaces, the continuous forms of Hölder’s and
Minkowski’s inequalities can be stated as follows.
Hölder’s Inequality (continuous form). Let p and q be conjugate in-

dices: 1p + 1q = 1. If f ∈ Lp and g ∈ Lq , then f g ∈ L1 and f g1 ≤ f p gq .
Minkowski’s Inequality (continuous form). Let 1 ≤ p ≤ ∞. If f ∈ Lp

and g ∈ Lp , then (f + g) ∈ Lp and f + gp ≤ f p + gp .
The proofs are essentially the same as for sums and are left as exercises.
Equality occurs in Hölder’s inequality if and only if f (x)g(x) has constant
sign and |f (x)|p = λ|g(x)|q .
The integral analogue of Cauchy’s inequality, or Hölder’s inequality for
p = q = 2, is
1/2 1/2
∞ ∞ ∞
f (x)g(x) dx ≤ |f (x)| dx
2
|g(x)| dx
2
.

−∞ −∞ −∞
It is known as the Schwarz inequality, or the Cauchy–Schwarz inequality,

after Hermann Amandus Schwarz (1843–1921), who recorded it in a paper
on minimal surfaces published in 1885. However, the result had already been
formulated as early as 1859 by a Russian mathematician, Victor Yacovlevich
Bunyakovski (1804–1889), and so it often bears his name as well. In any case,
as Bunyakovski himself observed, the integral form is a simple consequence
of Cauchy’s inequality, by passage from Riemann sums to integrals.
There is also an integral analogue of the arithmetic–geometric mean
inequality, which becomes apparent if the geometric mean is written in the
form n
1
(a1 a2 · · · an )1/n = exp log ak .
n
k=1
This suggests the inequality

b b
1 1
exp log f (x) dx ≤ f (x) dx
b−a a b−a a
for positive functions f (x) on the interval [a, b]. Under the assumption that
f and log f are Riemann integrable, a proof can be based on the classical
arithmetic–geometric mean inequality for sums, approximating the integrals
by their corresponding Riemann sums. However, as we will see presently,
the result depends only on the convexity of the exponential function and is a
special case of a much more general inequality, known as Jensen’s inequality.
4.5. Jensen’s inequality

The statement of Jensen’s inequality will involve the notion of a convex
function. A function ϕ(x) is said to be convex on an interval I ⊂ R if for
each pair of points x1 , x2 ∈ I it satisfies the inequality
ϕ(tx1 + (1 − t)x2 ) ≤ tϕ(x1 ) + (1 − t)ϕ(x2 ) , 0 ≤ t ≤ 1.

120 4. Inequalities
The function is strictly convex if for each pair of distinct points x1 , x2 ∈ I ,
ϕ(tx1 + (1 − t)x2 ) < tϕ(x1 ) + (1 − t)ϕ(x2 ) , 0 < t < 1.
Geometrically, these inequalities say that the graph of the function lies on
or below every chord. Jensen’s inequality, in most primitive form, asserts
that a similar inequality then holds for more general means. Here, then, is
the discrete form of the inequality.
Jensen’s Inequality. Let ϕ(x) be a convex function on an interval I ⊂ R ,

and let t1 , t2 , . . . , tn be any collection of positive numbers with sum t1 + t2 +
· · · + tn = 1. Then for each choice of points x1 , x2 , . . . , xn ∈ I the inequality
ϕ(t1 x1 + t2 x2 + · · · + tn xn ) ≤ t1 ϕ(x1 ) + t2 ϕ(x2 ) + · · · + tn ϕ(xn )
holds. If ϕ is strictly convex, equality occurs only for x1 = x2 = · · · = xn .
The inequality is named for the Danish mathematician J. L. W. V.

Jensen (1859–1925), who recognized the importance of convexity and pub-
lished a more restrictive result in 1906. Jensen was an amateur mathe-
matician who received no formal training in research-level mathematics and
never held an academic position, supporting himself as an employee of the
telephone company in Copenhagen.
Before turning to a proof of Jensen’s inequality, let us observe that the
arithmetic–geometric mean inequality is an immediate corollary. For this we
need only choose the function ϕ(x) = ex and take t1 = t2 = · · · = tn = n1 .
Writing ak = exk , we observe that
√
t 1 x1 + t 2 x2 + · · · + t n xn = 1
n log(a1 a2 · · · an ) = log ( n a1 a2 · · · an ) .
Since the exponential function is convex, Jensen’s theorem gives

√
n
a1 a2 · · · an ≤ n1 (a1 + a2 + · · · + an ) ,
which is the arithmetic–geometric mean inequality. Moreover, because the

exponential function is strictly convex, equality occurs only when all of the
numbers ak are equal.
Jensen’s inequality will now be proved by induction. For n = 2 it is
simply a restatement of the definition of convexity. Suppose the inequality
holds for some n ≥ 2 and write

n+1
n
tk
tk xk = tn+1 xn+1 + (1 − tn+1 ) xk
1 − tn+1
k=1 k=1
for arbitrary points x1 , . . . , xn+1 ∈ I and positive weights tk with sum t1 +

t2 +· · ·+tn+1 = 1. Then by the convexity of ϕ and the inductive hypothesis,
n+1
n
tk
ϕ tk xk ≤ tn+1 ϕ(xn+1 ) + (1 − tn+1 ) ϕ xk
1 − tn+1
k=1 k=1

n
tk
n+1
≤ tn+1 ϕ(xn+1 ) + (1 − tn+1 ) ϕ(xk ) = tk ϕ(xk ) ,
1 − tn+1
k=1 k=1
which is the desired result for n + 1. This proves Jensen’s inequality. The
case of equality is left as an exercise.
It is not difficult to formulate an appropriate analogue for integrals. In
continuous form, Jensen’s inequality may be expressed as follows.
Jensen’s Inequality (continuous form). Let ϕ(y) be a convex function
on an interval J ⊂ R . Suppose the function y = f (x) is integrable over an
interval I ⊂ R , with f (I) ⊂ J. Let w(x) ≥ 0 be a weight function on I with
integral I w(x) dx = 1. Then

ϕ f (x)w(x) dx ≤ ϕ(f (x))w(x) dx .
I I
If ϕ is strictly convex, w(x) > 0, and f is continuous, then equality holds

only when f is constant.
The continuous form can be deduced from the discrete form by the usual
process of approximating integrals by their Riemann sums, but it is more
interesting to give an independent proof. For simplicity we will make the
assumption that ϕ has a nondecreasing derivative on J. It then follows that
the graph of ϕ lies above each of its tangent lines. More precisely, for each
point y0 ∈ J the inequality
ϕ(y0 ) + ϕ (y0 )(y − y0 ) ≤ ϕ(y)
holds for all y ∈ J, with strict inequality for all y = y0 if ϕ is strictly convex.
Setting y = f (x), we have
ϕ(y0 ) + ϕ (y0 )(f (x) − y0 ) ≤ ϕ(f (x)) , x∈I.
Now multiply the inequality by w(x) and integrate over I to obtain

ϕ(y0 ) + ϕ (y0 ) f (x)w(x) dx − y0 ≤ ϕ(f (x))w(x) dx ,
I I

since w is a unit weight. With the choice y0 = I f (x)w(x)dx , this gives the
continuous form of Jensen’s inequality. The case of equality is again left as
an exercise.
122 4. Inequalities
The proof is essentially the same without the simplifying assumption

that ϕ is differentiable and has a nondecreasing derivative, because it is
possible to find a suitable substitute for the tangent line to the graph of
a general convex function ϕ. Details may be found in the classic book of
Hardy, Littlewood, and Pólya [2] or in the recent book by Steele [6].
Specific applications of Jensen’s inequality, in either discrete or continu-
ous form, will depend upon knowing that a given function ϕ is convex. The
most useful criterion for convexity is that ϕ have a second derivative and
ϕ (x) ≥ 0 on an interval (a, b). Similarly, ϕ is strictly convex if ϕ (x) > 0
on (a, b). Although it is intuitively clear from graphical considerations that
the property ϕ (x) ≥ 0 implies convexity, a proof is more elusive.
Here is one approach. Given that ϕ (x) ≥ 0, or that ϕ (x) is nonde-
creasing on (a, b), it is to be shown that
ϕ(tx + (1 − t)y) ≤ tϕ(x) + (1 − t)ϕ(y)
whenever a < x < y < b and 0 < t < 1. To see this, let s = 1 − t and write
tx+sy y

ϕ(tx + sy) − tϕ(x) − sϕ(y) = t ϕ (u) du − s ϕ (u) du
x tx+sy
≤ ts(y − x)ϕ (tx + sy) − st(y − x)ϕ (tx + sy) = 0 ,
by the monotonicity of ϕ . The same argument shows that ϕ is strictly

convex if ϕ is strictly increasing.
4.6. Hilbert’s inequality

This chapter concludes with a special inequality, due to David Hilbert
(1862–1943), sometimes called Hilbert’s double series theorem.
Hilbert’s Inequality. For arbitrary square-summable sequences {aj } and

{bk }, the inequality
∞ ∞
∞ 1/2
∞ 1/2
aj bk
(1) ≤π a 2
b2
j + k − 1 j k
j=1 k=1 j=1 k=1
holds, with strict inequality unless one of the sequences {aj } or {bk } is iden-
tically zero. The constant π cannot be replaced by any smaller number.
In other words, the bound is best possible but is attained only in trivial
cases.
Hilbert’s inequality has found applications to real and complex analysis
and to analytic number theory. Hilbert originally included the result in
lectures at Göttingen on integral equations, but his proof was first published
in the 1908 dissertation of his student Hermann Weyl. Actually, Hilbert
obtained a weaker inequality with constant 2π, and several years later Issai
Schur constructed another proof that led to the sharp constant π. Since that
time many proofs have been found, but none is entirely simple. (See, for
instance, Hardy–Littlewood–Pólya [4], Oleszkiewicz [5], and Steele [6].) The
proof presented here was discovered by David Ullrich [7] and seems relatively
simple and straightforward. The strategy is to begin with a continuous form
of the theorem and to deduce the discrete form from it.
Hilbert’s Inequality (continuous form). For any functions f and g in

the space L2 (0, ∞), the inequality
∞ ∞
f (x)g(y)
(2) dxdy ≤ π f 2 g2
x+y
0 0
holds, with strict inequality unless f or g is the zero-function. The constant

π cannot be replaced by any smaller number.
Proof of continuous form. The change of variables x = u2 , y = v 2 gives

∞ ∞ ∞ ∞
f (x)g(y) f (u2 )g(v 2 )
dxdy = 4 uv dudv .
0 0 x+y 0 0 u2 + v 2
Now introduce polar coordinates u = r cos θ, v = r sin θ and apply the

Cauchy–Schwarz inequality to obtain
∞ ∞
f (x)g(y)
dxdy
x+y
0 0
π/2 ∞
2
≤4 cos θ sin θ f (r cos2 θ) g(r2 sin2 θ) r drdθ
0 0
π/2 ∞ 1/2 ∞ 1/2
≤4 cos θ sin θ 2 2
f (r cos θ) r dr2 2 2 2
g(r sin θ) r dr dθ
0 0 0
π/2 ∞ 1/2 ∞ 1/2
=2 2
f (s) ds 2
g(t) dt dθ = π f 2 g2 ,
0 0 0
with the substitution s = r2 cos2 θ, t = r2 sin2 θ. Thus we have arrived at

the inequality (2). We defer the proofs that the constant π is best possible
and is never attained, passing first to the discrete form of the inequality.
Proof of discrete form. The discrete form (1) of Hilbert’s inequality is

essentially a special case of the continuous form (2). In order to deduce (1)
from (2), we define the functions ψj on (0, ∞) by ψj (x) = 1 for x ∈ (j − 1, j]
124 4. Inequalities
and ψj (x) = 0 elsewhere, j = 1, 2, . . . . Given a pair of sequences {aj } and

{bk } in 2 , let
∞
∞

f (x) = aj ψj (x) and g(y) = bk ψk (y) .
j=1 k=1
Then
∞
∞

f 22 = a2j , g22 = b2k , and
j=1 k=1
∞ ∞ ∞
∞
f (x)g(y)
dxdy = wjk aj bk , where
0 0 x+y
j=1 k=1
k j
1
wjk = dxdy .
k−1 j−1 x + y
It is obvious that wjk > 1/(j + k), and in fact the convexity of the integrand
implies that wjk > 1/(j + k − 1). Specifically, for each fixed y > 0, we see
that
j
1 1
dx > , j = 1, 2, . . . ,
j−1 x + y y + j − 12
by comparing the area under the curve with the area under the tangent line
at the midpoint x = j − 12 . A similar estimate then shows that
k
1 1 1
(3) wjk > dy > 1 = j+k−1.
k−1 y+j− 1
2 (k − 1
2) + (j − 2 )
Therefore, if neither {aj } nor {bk } is the zero-sequence, it follows from (2)
that
∞ ∞ ∞
∞ ∞ ∞
aj bk |f (x)g(y)|
< wjk |aj bk | = dxdy
j + k − 1 x+y
j=1 k=1 j=1 k=1 0 0

∞ 1/2 ∞ 1/2
≤ π f 2 g2 = π aj2 2
bk ,
j=1 k=1
which proves (1) with strict inequality. Note that (3) gives strict inequality
because |aj bk | > 0 for some pair of indices j and k.
Proof of sharpness. The next step is to show that the constant π is best
possible in the continuous form of Hilbert’s inequality, then to adapt the
calculations to deduce that π is best possible in the discrete form as well.
For the first purpose it is useful to choose the functions
√
1/ x for 1 ≤ x ≤ R
(4) f (x) = g(x) = hR (x) =
0 elsewhere in (0, ∞) ,
where R is a large parameter. Then f 2 g2 = log R and

∞ ∞ ∞ ∞
f (x)g(y) f (u2 )g(v 2 )
dxdy = 4 uv dudv
0 0 x+y 0 0 u2 + v 2
√ √
R R
4
= dudv .
1 1 u2 + v 2
The problem is then to obtain an effective lower bound for the last integral
in terms of R. We propose to show that
√ √
R R
4
(5) dudv > π log R − π log 2 − 8 .
1 1 u2 + v 2
Because f 2 g2 = log R , it will then follow, by letting R tend to infinity,

that the constant π in the inequality (2) cannot be replaced by any smaller
number.
To derive the estimate (5), it is convenient to define the disjoint rectan-
gular regions
√ √
QR = (u, v) : 1 ≤ u ≤ R , 1 ≤ v ≤ R ,
√
ER = (u, v) : 0 ≤ u < 1 , 1 ≤ v ≤ R ,
√
FR = (u, v) : 1 ≤ u ≤ R , 0 ≤ v < 1
and the quarter-annulus

√ √
AR = (u, v) : 2 ≤ (u2 + v 2 )1/2 ≤ R , u ≥ 0 , v ≥ 0 .
Then AR ⊂ QR ∪ ER ∪ FR and we have

√ √
R R
4
(6) dudv = > − − .
1 1 u2 + v 2
QR AR ER FR
126 4. Inequalities
But a simple calculation yields

4 4
dudv = dudv
u + v2
2 u2 + v2
FR ER
√
R 1 ∞
4 4
= dudv < dv = 4 ,
1 0 u + v2
2
1 v2
and a change to polar coordinates gives

√
π/2 R
4 4
dudv = √ dr = π(log R − log 2) .
u2 + v 2 0 2 r
AR
Putting these results into (6), we arrive at the inequality (5), which shows
that the constant π is sharp in the continuous form (2) of Hilbert’s inequality.
In order to show that π is also best possible in the discrete form (1), we
choose √
1/ j for j = 1, 2, . . . , N
aj = bj =
0 for j > N ,
where N is a large integer, and let f (x) = g(x) = hN +1 (x) as defined in (4).
Then
∞
∞

f (x) ≤ aj ψj+1 (x) and g(y) ≤ bk ψk+1 (y)
j=1 k=1
for all x, y ∈ (0, ∞), so that

∞ ∞ ∞ ∞
f (x)g(y)
dxdy ≤ wj+1,k+1 aj bk ,
0 0 x+y
j=1 k=1
where
k+1 j+1
1 1 1
wj+1,k+1 = dxdy < < .
k j x+y j+k j+k−1
Consequently, in view of the lower bound (5), we have shown that

∞
∞ ∞ ∞
aj bk
> wj+1,k+1 aj bk
j+k−1
j=1 k=1 j=1 k=1
∞ ∞
f (x)g(y)
≥ dxdy > π log(N + 1) − π log 2 − 8 ∼ π log N
0 0 x+y
Exercises 127
as N → ∞. On the other hand,

∞ 1/2
∞ 1/2
N 1/2
N 1/2
a2j b2k = 1/j 1/k ∼ log N .
j=1 k=1 j=1 k=1
Letting N tend to infinity, we conclude that the constant π in (1) is best

possible.
In summary, the three steps of proof combine to establish the discrete

form of Hilbert’s inequality in its entirety and the continuous form except
for the assertion of strict inequality. Reviewing the proof of the continuous
form, we see that the key to strict inequality is to examine the conditions for
equality in our application of the Cauchy–Schwarz inequality. The details
are technical, however, and will be omitted.
Exercises
1. (a) Use mathematical induction to prove that
(1 + x)n ≥ 1 + nx , x > −1 ,
for n = 1, 2, . . . .
(b) Generalize the result by proving that

n
(1 + a1 )(1 + a2 ) · · · (1 + an ) ≥ 1 + ak
k=1
whenever ak > −1 for k = 1, 2, . . . , n and all ak have the same sign.
2. For ak > 0 and a1 + a2 + · · · + an < 1, prove:

n
1
(a) 1 + ak < (1 + a1 )(1 + a2 ) · · · (1 + an ) < n ,
1− k=1 ak
k=1
n
1
(b) 1 − ak < (1 − a1 )(1 − a2 ) · · · (1 − an ) < n .
1+ k=1 ak
k=1
Hint: Observe that 1 + ak < 1/(1 − ak ).
3. For positive numbers a1 , a2 , . . . , an , the harmonic mean Hn is defined by

n
Hn = .
1
a1 + 1
a2 + ···+ 1
an
Prove that Hn ≤ Gn , with equality if and only if a1 = a2 = · · · = an .

128 4. Inequalities
4. Let a1 , a2 , . . . , an be positive numbers with sum a1 + a2 + · · · + an = n .

Show that a1 a2 · · · an ≤ 1, and that the maximum occurs only when all of
the numbers are equal to 1.
5. Prove without calculus that among all 3-dimensional rectangular boxes
of fixed volume, the cube has the smallest surface area.
6. Show that
√ √ n+1
n ≤ n! ≤
n
, n = 1, 2, . . . .
2
7. Let xn = (1 + n1 )n and yn = (1 + n1 )n+1 . Observe that xn < yn .
Use the arithmetic–geometric mean inequality to show that xn < xn+1 and
yn+1 < yn , and deduce that the sequences {xn } and {yn } both converge
to a common limit. Recall that the common limit is e, the base of natural
logarithms. Conclude that

1 n 1 n+1
1+ <e< 1+ , n = 1, 2, . . . .
n n
Taking n = 10, 000, use a calculator to show that 2.7181 < e < 2.7185.
Hint: Consider the numbers a1 = a2 = · · · = an = n+1 n , an+1 = 1. Then
consider the numbers a1 = a2 = · · · = an+1 = n+1
n
, an+2 = 1.
8. Show that
n 1/2

n
√ √ n
|ak | ≤ n 2
ak ≤ n |ak | , n = 1, 2, . . . .
k=1 k=1 k=1
9. Verify the condition for equality in Hölder’s inequality.

10. Prove the continuous forms of the Hölder and Minkowski inequalities.
11. Verify the conditions for equality in both forms of Jensen’s inequality,
discrete and continuous.
12. Use Jensen’s inequality to prove Minkowski’s inequality.
13. If f ∈ Lr ∩ Ls and 1
p = λ
r + 1−λ
s for 0 < λ < 1, show that f p ≤
λ 1−λ
f r f s . As a special case, show that if f ∈ L1 ∩ L2 . then f ∈ Lp for
1 < p < 2.
1 1 1
14. If p, q, and r are positive numbers satisfying + + = 1, show that
p q r
f gh1 ≤ f p gq hr .
15. (a) Show that if a sequence a = (a1 , a2 , . . . ) belongs to the space p for
some p < ∞, then a ∈ q for all q with p < q < ∞.
Exercises 129
(b) If a ∈ p for some p < ∞, show that ap → a∞ as p → ∞.

√
Hint. If 1 ≤ xn ≤ C, then n xn → 1 as n → ∞.
16. Prove the following generalization of Hilbert’s inequality. Let p and q

lie in the interval (1, ∞) and be conjugate indices, so that 1p + 1q = 1. Then
for arbitrary real sequences {aj } ∈ p and {bk } ∈ q , the inequality
∞ ∞
∞ 1/p
∞ 1/q
aj bk π
≤ |aj | p
|bk | q
j + k − 1 sin(π/p)
j=1 k=1 j=1 k=1
holds, with strict inequality unless one of the sequences {aj } or {bk } is
identically zero. The constant π/ sin(π/p) cannot be replaced by any smaller
number.
17. Let a1 , a2 , . . . , an be arbitrary real numbers, and suppose that M and

m are the maximum and minimum, respectively, of their partial sums sk =
a1 + a2 + · · · + ak for k = 1, 2, . . . , n. If b1 ≥ b2 ≥ · · · ≥ bn ≥ 0, prove that
mb1 ≤ a1 b1 + a2 b2 + · · · + an bn ≤ M b1 .
Hint. Apply an Abel summation.
18. By successive integrations of the inequality cos x ≤ 1, show that for all
x ≥ 0,
x2 x3 x2 x4
sin x ≤ x , cos x ≥ 1 − , sin x ≥ x − , cos x ≤ 1 − + ,
2! 3! 2! 4!
and in general that
x3 x4k+1 x2 x4k+2
sin x ≤ x − +··· + , cos x ≥ 1 − + ···− ,
3! (4k + 1)! 2! (4k + 2)!
x3 x4k+3 x2 x4k+4
sin x ≥ x − +··· − , cos x ≤ 1 − + ···+
3! (4k + 3)! 2! (4k + 4)!
for each k = 0, 1, 2, . . . and all x ≥ 0. Prove by induction.
19. Let f (x) ≥ 0 on the interval [0, 1], and suppose it has a finite integral
1
A = 0 f (x) dx. Prove that
1
1+ A2 ≤ 1 + f (x)2 dx ≤ 1 + A .
0
Interpret the inequalities geometrically by taking f = g , where g is a non-

decreasing continuously differentiable function on [0, 1].
130 4. Inequalities
References
[1] E. F. Beckenbach and R. Bellman, Inequalities, Springer–Verlag, New York,
1965.
[2] G. H. Hardy, J. E. Littlewood, and G. Pólya, Inequalities, Second edition, Cam-
bridge University Press, Cambridge, U.K., 1952.
[3] Nicholas D. Kazarinoff, Analytic Inequalities, Holt, Rinehart and Winston, New
York, 1961; reprinted by Dover Publications, Mineola, N.Y., 2003.
[4] D. S. Mitrinović, Elementary Inequalities, P. Noordhoff, Groningen, The Nether-
lands, 1964.
[5] Krzysztof Oleszkiewicz, “An elementary proof of Hilbert’s inequality”, Amer.
Math. Monthly 100 (1993), 276–280.
[6] J. Michael Steele, The Cauchy–Schwarz Master Class: An Introduction to the
Art of Mathematical Inequalities, Cambridge University Press, Cambridge, 2004.
[7] David Ullrich, “A simple elementary proof of Hilbert’s inequality”, Amer. Math.
Monthly, to appear.
Chapter 5
Infinite Products
Infinite products, like infinite series, arise often and are natural objects
in mathematical analysis. The convergence theory of infinite products is
closely related to that of infinite series but is more subtle. In this chapter
we develop standard criteria for convergence and uniform convergence of
infinite products.
5.1. Basic concepts

Infinite products are analogous to infinite series, and the two are closely
related. The convergence of an infinite product
∞

(1 + ak ) = (1 + a1 )(1 + a2 ) · · · ,
k=1
where ak are real numbers, comes down to the behavior of the sequence of
partial products

n
pn = (1 + ak ) = (1 + a1 )(1 + a2 ) · · · (1 + an )
k=1
as n tends to infinity. The reason for writing the factors in the form (1 + ak )
will become apparent shortly.
It turns out that the “right” notion of convergence is not
the obvious
one. It would seem natural to declare an infinite product ∞ k=1 (1 + ak )
convergent if its partial products pn form a convergent sequence, but that
131
132 5. Infinite Products
convention has several drawbacks. First of all, any reasonable definition of

convergence should depend only on the “eventual behavior” of the factors,
not on the values of finitely many factors. But if a single factor 1 + am = 0,
then pn = 0 for all n ≥ m, and so pn → 0 as n → ∞. To avoid this trivial
complication, we make the standing assumption (to be modified later) that
the product contains no zero factors. In other words, 1 + ak = 0 for all k.
An infinite product is then said to be convergent if its sequence of partial
products converges to a finite, nonzero limit. Thus in the above notation
the requirement is that
lim pn = p = 0 ,
n→∞
and the infinite product is then assigned the value p. The infinite product is
declared to be divergent if it is not convergent; that is, if the sequence {pn }
either diverges or tends to zero. If pn → 0, the product is said to diverge to
zero.
One advantage of insisting on a nonzero limit is that convergence of
an infinite product implies convergence of the corresponding product of re-
ciprocal factors, since the sequence {1/pn } then also converges to a finite,
nonzero limit. Another important consequence is that the factors of a con-
vergent product tend to 1, just as the terms of a convergent series tend to
0. Specifically, if an infinite product converges to p , then
1 + an = pn /pn−1 → p/p = 1 as n → ∞ ,
or equivalently, an → 0. However, this is only a necessary condition for

convergence; it is far from sufficient.
Here are two examples.
Example 1. If ak = k1 , then 1 + ak = k+1 k , and the partial product “tele-
scopes” to give
2 3 4 n+1
pn = · · · · · = n + 1.
1 2 3 n

Thus the infinite product ∞ 1
k=1 (1 + k ) is divergent.
Example 2. Suppose ak = − k12 for k = 2, 3, . . . . Then
1 k2 − 1 (k − 1)(k + 1)
1 + ak = 1 − 2
= 2
= ,
k k k2
and again the partial product telescopes to give
n

1 1·3 2·4 3·5 (n − 1)(n + 1) 1 n+1 1
pn = 1− 2 = · 2 · 2 ··· = · →
k 22 3 4 n2 2 n 2
k=2
5.1. Basic concepts 133
as n → ∞. Thus the infinite product converges, and

n
1 1
1− 2 = .
k 2
k=2
Our two
examples suggest the possibility that convergence of an infinite
product (1 + ak ) is equivalent to convergence of the corresponding infinite
series ak . This is actually true if the numbers ak are all of the same sign,
as we now proceed to show.

Theorem 1. If ak ≥ 0 for all k, then the product ∞ k=1 (1 + ak ) converges
∞
if and only if the series k=1 ak converges. The same is true if ak ≤ 0 for
all k.
n
Proof. nAgain let pn = k=1 (1 + ak ) denote the partial products, and let
sn = k=1 ak denote the partial sums of the related series. Suppose first
that ak ≥ 0 for all k. Then 1 ≤ pn ≤ pn+1 and sn ≤ sn+1 , so by the
monotone boundedness theorem, either of the sequences {pn } or {sn } is
convergent if and only if it is bounded above. But in view of the inequality
1 + x ≤ ex ,
it is easy to see that
1 + sn ≤ pn ≤ esn , n = 1, 2, . . . .
Therefore, the boundedness of {pn } is equivalent to that of {sn }, and the

theorem is proved for the case where ak ≥ 0.
Suppose next that ak ≤ 0, and assume without loss of generality that
ak > −1 for all k. Now 0 < pn+1 ≤ pn , and so the infinite product converges
if and only if pn ≥ b for some constant b > 0. Observe first that
1
1 ≤ 1 − ak ≤ ,
1 + ak

so the convergence of (1 + ak ) implies that of (1 − ak ), which implies,
by what we have just
proved, that the series ak converges. Conversely,
suppose the series ak is convergent. This implies that ak → 0, so after
removing a finite number of factors we may assume that − 12 ≤ ak ≤ 0. Then
an application of the elementary inequality
1 ≤ (1 + x)(1 − 2x) , − 12 ≤ x ≤ 0 ,
shows that
1
≤ 1 + ak .
1 − 2ak

However,
the infinite product (1 − 2ak ) converges, since −2ak ≥ 0 and the
series
ak converges. Thus the above inequality shows that the product
(1 + ak ) is also convergent. This completes the proof.
The question now arises whether the theorem

remains true under more
general
conditions. Do the infinite product (1 + ak ) and the infinite series
ak always converge and diverge together? The answer is no. Consider for
instance the following example.
Example 3. The infinite product
∞

1
1−
k
k=2
diverges to zero, as does the product

1 1 1 1 1 1
1+ √ 1− √ 1+ √ 1− √ 1+ √ 1− √ ··· .
2 2 3 3 4 4
However, the first product corresponds to a divergent series, whereas the

second corresponds to a convergent series.
It is more difficult to find a convergent product whose associated series
diverges. Here is one example.
Example 4. Consider the infinite product

1 1 1 1 1 1
1− √ 1+ √ + 1− √ 1+ √ +
2 2 2 3 3 3

1 1 1
· 1− √ 1+ √ + ··· .
4 4 4
Note that

1 1 1 1 1
1− √ 1+ √ + = 1 − 3/2 and 1− √ → 1,
k k k k k
so the product converges although its associated series

1 1 1 1 1 1 1 1
−√ + √ + −√ + √ + −··· = + + ...
2 2 2 3 3 3 2 3
is divergent. More precisely, the series has a subsequence of partial sums

s2n = 1 + 12 + · · · + n1 , and so it diverges.
5.2. Absolute convergence 135
5.2. Absolute convergence

A useful necessary and sufficient condition for convergence of infinite
products can be adapted from the Cauchy criterion for convergence of a
sequence. Here is the result.

Theorem 2. An infinite product (1 + ak ) is convergent if and only if its
factors have the following property. For each ε > 0 there is a number N
such that
(1) |(1 + am )(1 + am+1 ) · · · (1 + an ) − 1| < ε
for all indices m and n with n > m ≥ N .
Proof. Suppose first that the product converges. Then the partial products
pn → p = 0 as n → ∞. In particular, the partial products are bounded away
from zero: |pn | ≥ b for some b > 0 and all n. For each ε > 0, the Cauchy
condition says that
|pn − pm | < εb , n>m≥N,
for some N sufficiently large. But |pm | ≥ b, so this implies that
|pn /pm − 1| < ε , n>m≥N,
which is essentially the condition (1).

Conversely, if the factors have the property (1), then for each ε > 0 the
inequality |pn /pm − 1| < ε holds whenever n > m ≥ N . Choose ε = 12 to
infer that
|pn − pM | < 12 |pM | , n>M,
for some specific index M . This implies that
1
2 |pM | < |pn | < 3
2 |pM | , n>M.
In other words, the partial products are bounded away from both zero and
infinity. Therefore, for each ε > 0 it follows from (1) that
|pn − pm | < ε |pm | < 3

2 ε |pM |
for all m and n greater than some N ≥ M . This shows that {pn } is a Cauchy
sequence, and so pn → p as n → ∞. Finally, p = 0 since the partial products
pn are bounded away from zero. Thus the infinite product is convergent.
We can now
discuss the question of absolute convergence. An infi-
nite product (1 + ak ) is said to be absolutely convergent if the product

(1 + |ak |) is convergent. In view of Theorem 1, a product is absolutely
convergent if and only if its associated series ak is absolutely convergent.
A convergent product need not be absolutely convergent, as simple examples
show. For instance, the product

1 1 1 1 1 1
1+ 1− 1+ 1− 1+ 1− ···
2 2 3 3 4 4
is not absolutely convergent but it is convergent, because its partial products

tend to the same limit as do those of the convergent product

1 1 1
1− 2 1− 2 1 − 2 ··· .
2 3 4
On the other hand, absolute convergence implies convergence.
Theorem 3. Every absolutely convergent product is convergent.
Proof. The inequality
|(1 + am )(1 + am+1 ) · · · (1 + an ) − 1| ≤ (1+|am |)(1+|am+1 |) · · · (1+|an |)−1
can be verified by multiplying out the left-hand product, subtracting 1, ap-

plying the triangle inequality, and regrouping the factors. Thus the desired
result is an immediate consequence of Theorem 2.
5.3. Logarithmic series

We now come to a very natural necessary and sufficient condition for
convergence of an infinite product. The formal identity
∞ ∞

log (1 + ak ) = log(1 + ak )
k=1 k=1
suggests that the product and its associated series of logarithms will converge
or diverge together. Under an assumption to guarantee that the logarithms
are well-defined, this turns out to be correct.
∞
Theorem 4. If ak > −1 for all k, then the ∞ infinite product k=1 (1 + ak )
converges if and only if the infinite series k=1 log(1 + ak ) converges.
Proof. If the product converges, so that pn → p = 0, then it follows from

the continuity of the logarithm function that log pn → log p. But log pn is
simply the nth partial sum of the infinite series, so this shows that the series
converges. Conversely, the convergence of the series says that log pn → L, so
5.3. Logarithmic series 137
pn → eL = 0, and the infinite product converges. The proof shows that the
value of the product is the exponentiated sum of the series, as expected.
From Theorem 4 it is a short step to a condition under which convergence

∞
of the product ∞ k=1 (1 + ak ) is equivalent to that of the sum k=1 ak .
Cauchy attributes the following theorem to Gaspard-Gustave Coriolis (1792–
1843), who is better known for discovering the Coriolis force on the surface
of a rotating body.
∞ 2 ∞
Theorem 5. If
k=1 a k < ∞, then the product k=1 (1 + ak ) converges if
and only if the series ∞ k=1 ak converges.
Proof. The deduction from Theorem 4 will use the inequality
(2) 1 2
3x < x − log(1 + x) < x2 , |x| < δ ,
for some positive constant δ < 1 sufficiently small. This is a simple conse-
quence of the limit
x − log(1 + x) 1
lim = .
x→0 x2 2
2
If the series ak converges, then ak → 0 and so |ak | < δ for all indices k
larger than some number N . The inequality (2) then gives
|ak − log(1 + ak )| ≤ a2k , k>N.
Thus the series

∞

ak − log(1 + ak )
k=1
converges (absolutely). It follows that if either of the series

∞
∞

ak or log(1 + ak )
k=1 k=1
converges, then so does the other. In view of Theorem 4, this completes the
proof.
If the divergent
product of Example 3 is written in the first
form, with
a
k = −1/k, then a 2 < ∞, so the simultaneous divergence of
k (1+ak ) and
ak is in accordance with Theorem 5. On the other hand, if the product
is
written in the second form, then a2 = ∞ and so the convergence of ak
k
and divergence of (1 + ak ) is beyond the jurisdiction
of Theorem 5. For
Example 4, where 2 the (1 + ak ) converges but the ak diverges, we see
similarly that ak = ∞.
2
If the product (1 + ak ) and the sum ak both converge,
2 then ak
also converges (cf. Exercise 5). However, the convergence of ak is by no
means necessary for √ the simultaneous divergence
of product and sum. For
example, let ak = 1/ k. Then the product (1 + ak ), the sum ak , and
the sum 2
ak all diverge.
5.4. Uniform convergence

Up to this point, we have maintained the standing assumption that an
infinite product has no zero-factors, or equivalently that ak = −1 for all
k. In various applications, however, especially those dealing with infinite
products of functions, it is essential to allow zero-factors and to modify
the definition of convergence accordingly. An infinite product with finitely
many zero-factors is said to be convergent if after removal of those factors
the product of remaining factors is convergent. In this case the product is
said to converge and have value 0, or to converge to zero. A product with
infinitely many zero-factors is regarded as divergent.
One of the most important uses of infinite products is to give explicit
representations of special functions. For instance, it will be shown in Chap-
ter 8 that the sine function has the expression
∞
x2
(3) sin πx = πx 1− 2
k
k=1
for all real numbers x. Note that the infinite product converges for each fixed
x, by Theorem 1. It vanishes precisely for x = 0, ±1, ±2, . . . , which we know
are the zeros of the function sin πx. An infinite product representation of this
type can be viewed as a generalization of the factorization of a polynomial
according to its (real) zeros.
Consider now a general infinite product of the form
∞

(4) 1 + gk (x) ,
k=1
where the functions gk are defined on a common set E ⊂ R. The product

is said to converge pointwise on E if it converges for each point x ∈ E. We
can then write
∞

f (x) = 1 + gk (x) , x∈E.
k=1
The product is said to be uniformly convergent on E if its sequence of partial
products

n

fn (x) = 1 + gk (x)
k=1
converges to f (x) uniformly for x ∈ E. Note that f (x) = 0 for all x ∈ E if

each gk (x) = −1 in E.
The tests for convergence of numerical products, as given in Theorems
1, 2, and 4, can be adapted to provide criteria for uniform convergence
of products of functions. We begin with a uniform version of the Cauchy
criterion.
Theorem 6. Let g1 (x), g2 (x), . . . be bounded functions on a set E ⊂ R, with
gk (x) = −1 for x ∈ E. Then the infinite product (4) converges uniformly on
E if its partial products fn (x) have the following property. For each ε > 0
there is a number N such that
(5) |fn (x)/fm (x) − 1| < ε for all x ∈ E
and for all indices m and n with n > m ≥ N . Conversely, the partial
products have this property if the infinite product (4) converges uniformly on
E to a function f (x) that is bounded below by a positive constant.
Proof. Suppose first that the partial products have the property (5). Then
corresponding to ε = 12 there is an index M for which the uniform inequality
|fn (x) − fM (x)| < 1
2 |fM (x)| , x∈E,
holds for all indices n > M . Consequently,
(6) 1
2 |fM (x)| < |fn (x)| < 3
2 |fM (x)| , x∈E, n>M.
Then since each of the factors 1 + gk (x) is bounded, the condition (5) and
its consequence (6) imply that
|fn (x) − fm (x)| < ε|fm (x)| < 3
2 ε|fM (x)| ≤ C ε , x∈E,
for all m and n greater than some index N , where C is an absolute constant.
This says that {fn (x)} is a uniform Cauchy sequence, so that fn (x) →
f (x) uniformly on E as n → ∞. The inequality (6) shows that |f (x)| ≥
2 |fM (x)| > 0 for each x ∈ E, so we can conclude that the infinite product
1
(4) is uniformly convergent on E.

Next suppose that the product (4) converges uniformly on E to a func-
tion f (x) with the property |f (x)| ≥ δ for some δ > 0 and all x ∈ E. Then
{fn (x)} is a uniform Cauchy sequence, so for each ε > 0 there is an index
N for which
|fn (x) − fm (x)| < 12 εδ , x∈E,
whenever n > m ≥ N . But the uniform convergence implies that |fm (x)| >
2 |f (x)| ≥ 2 δ for all x ∈ E and all m sufficiently large. This shows that (5)
1 1
holds if N is sufficiently large, so the proof is complete.

∞
Theorem 7. If the
∞series

k=1 |g
k (x)| converges uniformly on a set E ⊂ R,
then the product k=1 1 + gk (x) converges uniformly on E.
Proof. Let

n
n

sn (x) = |gk (x)| and Fn (x) = 1 + |gk (x)| .
k=1 k=1
Then for n > m,
1 ≤ Fn (x)/Fm (x) ≤ esn (x)−sm (x) , x∈E.
Therefore, because {sn (x)} is a uniform Cauchy sequence on E, there cor-

responds to each ε > 0 a number N such that
0 ≤ Fn (x)/Fm (x) − 1 < ε , x∈E,
whenever n > m ≥ N . As in the proof of Theorem 3, it follows that
|fn (x)/fm (x) − 1| < ε , x∈E,
for n > m ≥ N . The desired result

can now be inferred from Theorem 6,
since the uniform convergence of |gk (x)| implies that gk (x) → 0 uniformly
on E.

Theorem 8. If gk (x) > −1 and the series ∞ log
1 + gk (x) converges

k=1
uniformly on a set E ⊂ R, then the product ∞ k=1 1 + gk (x) converges
uniformly on E.
Proof. The partial sums

n

tn (x) = log 1 + gk (x)
k=1
form a uniform Cauchy sequence on E. Thus to each ε > 0 there corresponds

a number N such that
1 − ε < 1 + tn (x) − tm (x) ≤ etn (x)−tm (x) < 1 + ε , x∈E,
whenever n > m ≥ N . In other words, the partial products

n

fn (x) = 1 + gk (x)
k=1
have the property (5), so Theorem 6 ensures that the infinite product con-
verges uniformly on E.
Exercises 141
Exercises
1. (a) Calculate partial products and show directly that both infinite
products in Example 3 diverge to zero.
(b) Carry out the details to show that the infinite product in Example
4 is convergent.
2. Calculate partial products and show that
∞

2 1
1− = .
k(k + 1) 3
k=2
3. Show that
∞

(−1)k+1
1+ = 1.
k
k=1
∞
4. Let {xk } be a
sequence of real numbers for which 2
k=1 xk < ∞. Prove
that the product ∞ k=1 cos xk converges.
2
5. Adapt the proof
of Theorem 5 to show
that the sum ak converges
whenever the sum ak and the product (1 + ak ) both converge.
6. Prove that
∞
∞

4 1 9 1
1− 2 = and 1− 2 = .
k 6 k 20
k=3 k=4
7. Generalize the results of Exercise 6 by showing that

∞

m2 (m!)2
1− 2 = , m = 1, 2, 3, . . . .
k (2m)!
k=m+1
8. Conclude from Exercise 7 that

∞

k2 2m
= , m = 1, 2, 3, . . . ,
k −m
2 2 m
k=m+1
and deduce that

∞
k2
= (−1)m+1 2 , m = 2, 3, 4, . . . .
k 2 − m2
k=1
k=m
9. (a) Suppose a regular n-gon is circumscribed about a circle of radius r

and another circle is circumscribed about the n-gon. Show that the second
circle has radius R = r sec πn .
(b) Suppose now that an equilateral triangle is circumscribed about a

circle of radius 1, then a circle C3 is circumscribed about the triangle, then a
square is circumscribed about C3 , another circle C4 is circumscribed about
the square, a regular pentagon is circumscribed about C4 , and so on. If Cn
has radius Rn , show that
∞
π
lim Rn = sec ,
n→∞ n
n=3
and prove that the infinite product converges. (Its numerical value is known
to be approximately 8.700 .)
10. Show that the infinite product

∞

x2
1− 2
k
k=1
converges uniformly in every bounded subset E ⊂ R. Conclude that

∞

x2
s(x) = πx 1− 2
k
k=1
is continuous on the whole real line, without using the fact that s(x) =
sin πx.

ix
11. Recall that sinh x = 12 ex − e−x and sin x = 2i 1
e − e−ix . Use the
formula (3) for sin πx to conclude, at least formally, that
∞

x2
sinh πx = πx 1+ 2 .
k
k=1
12. Show that

∞
∞

1 sinh π k2 − 1 π
1+ 2 = and 2
= .
k π k +1 sinh π
k=1 k=2
13. Find a general formula for the partial products and show that
∞
k3 − 1 2
3
= .
k +1 3
k=2
Hint. Factor the polynomials k 3 − 1 and k 3 + 1.

References 143
1
(1 + x)(1 + x2 )(1 + x4 )(1 + x8 ) · · · = , |x| < 1 .
1−x
15. Use the identity sin 2x = 2 sin x cos x to show that

∞
x sin x
cos = , x = 0 .
2k+1 x
k=1
This is a generalization of Vieta’s product formula (see Chapter 2, Section

2.4).
16. For each real number α > 0, show that the generalized binomial coeffi-
cients
α α(α − 1) · · · (α − n + 1)
=
n n!
tend to zero as n → ∞.
Reference
[1] Konrad Knopp, Theory and Application of Infinite Series, Second English edi-
tion, Blackie & Son, London and Glasgow, 1951; reprinted by Dover Publications,
Mineola, NY, 1990.
Chapter 6
Approximation by
Polynomials
The Weierstrass approximation theorem is one of the gems of classical anal-

ysis. In this chapter we give three different proofs of the basic theorem, plus
a proof of Stone’s far-reaching generalization. The discussion begins with
interpolation theory and leads naturally to Chebyshev polynomials, which
solve a particular problem of best approximation. Later in the chapter we
consider the general problem (still unsolved) of finding the best uniform
approximation to an arbitrary continuous function by polynomials of given
degree.
6.1. Interpolation
The main focus of this chapter is a theorem of Weierstrass [15] that
every continuous function on a closed bounded interval can be approximated
uniformly by algebraic polynomials; that is, by functions of the form
P (x) = a0 + a1 x + a2 x2 + · · · + an xn ,
where n is a positive integer and the coefficients ak are real numbers. The
result is important because it often reduces a problem about continuous
functions to the corresponding problem for polynomials. Broadly speaking,
the theorem is a prototype for a large collection of results in approximation
theory.
To put the approximation problem in better perspective, we begin with
145
146 6. Approximation by Polynomials
a discussion of interpolation. Suppose a set of points
(x0 , y0 ), (x1 , y1 ), . . . , (xn , yn )
is specified with distinct abcissas xk . Then it is possible to construct a

polynomial of degree at most n whose graph passes through all of the points,
and this can be done in exactly one way. In other words, there is a unique
polynomial P of degree ≤ n such that P (xk ) = yk for k = 0, 1, . . . , n. To
see this, observe that the problem amounts to solving the system of linear
equations
a0 + a1 x0 + a2 x20 + . . . + an xn0 = y0
a0 + a1 x1 + a2 x21 + . . . + an xn1 = y1
···
a0 + a1 xn + a2 x2n + . . . + an xnn = yn
for the variables a0 , a1 , . . . , an . The problem has a unique solution because

the matrix of coefficients is nonsingular; that is, its determinant does not
vanish. The determinant is known as a Vandermonde determinant and it is
calculated by the formula

1 x0 x20 . . . xn0

1 x1 x21 . . . xn1
. . = (xk − xj ) .
. . .
.. ..
.
. .
j<k
1 xn x2n . . . xnn
(See Exercise 1 for a derivation.) In particular, a polynomial of degree ≤ n
cannot vanish at n + 1 points unless it vanishes identically.
In fact, the required polynomial is given by the Lagrange interpolation
formula

n
yj (x)
P (x) =
,
(xj )(x − xj )
j=0
where (x) = (x − x0 )(x − x1 ) · · · (x − xn ). More explicitly,

n
(x − x0 ) · · · (x − xj−1 )(x − xj+1 ) · · · (x − xn )
P (x) = yj .
(xj − x0 ) · · · (xj − xj−1 )(xj − xj+1 ) · · · (xj − xn )
j=0
It is readily verified that P is a polynomial of degree less than n + 1, and

that P (xk ) = yk for k = 0, 1, . . . , n. Thus by uniqueness it is the solution
to the interpolation problem.
The Lagrange interpolation formula is customarily used as a practical
method of extrapolation, or approximate reconstruction of a function whose
values are known (or have been measured) at a finite set of points. If a
function f is continuous over an interval [a, b], and x0 , x1 , . . . , xn are distinct
points in that interval, the Lagrange interpolation

n
f (xj ) (x)
P (x) = ,
(xj )(x − xj )
j=0
can be viewed as an approximation to f (x) over the interval [a, b]. But how
good is the approximation? Certainly it is exact if f is itself a polynomial
of degree ≤ n, but in the absence of further information there will be no
control away from the nodes xk , and the error
f − P ∞ = max |f (x) − P (x)|

a≤x≤b
as measured in the uniform norm may be quite large. The error can be
controlled, however, if f is known to be sufficiently smooth.
Theorem 1. Suppose a function f has n + 1 continuous derivatives on the
interval [a, b], and let P be its Lagrange interpolation polynomial with nodes
x0 , x1 , . . . , xn . Then
1
|f (x) − P (x)| ≤ f (n+1) ∞ |(x)| ,
(n + 1)!
where (x) = (x − x0 )(x − x1 ) · · · (x − xn ).
Proof. Fix an arbitrary point x in [a, b], not equal to any of the nodes xk .
Define the function
g(t) = f (t) − P (t) − c (t) ,
where c = [f (x) − P (x)]/(x), so that g(x) = 0. It is clear that g(xk ) = 0
for k = 0, 1, . . . , n. Thus g has at least n + 2 distinct zeros in [a, b]. By
Rolle’s theorem, the derivative g (t) vanishes at least once between each pair
of zeros, so g has at least n + 1 zeros. Continuing the argument by taking
successive derivatives, we find ultimately that g (n+1) (ξ) = 0 at some point
ξ ∈ (a, b). But P is a polynomial of degree ≤ n and is a monic polynomial
of degree n + 1, so P (n+1) (t) ≡ 0 and (n+1) (t) ≡ (n + 1)! . Therefore, the
conclusion is that f (n+1) (ξ) = c(n + 1)! , or
1
f (x) − P (x) = f (n+1) (ξ) (x) ,
(n + 1)!
which yields the desired result.
The question now arises as to how the nodes should be situated to min-
imize the error of the Lagrange approximation, as measured by the uniform
norm F − P ∞ . Theorem 1 suggests that the key is to minimize ∞ .

A more general formulation is to find the monic polynomial of fixed degree
that has smallest norm over a given interval. It will be convenient to work
with the interval [−1, 1] and to consider monic polynomials
Q(x) = xn + an−1 xn−1 + . . . + a1 x + a0
of fixed degree n (instead of n + 1).
We begin by recalling that cos nθ can be expressed as a polynomial of
degree n in cos θ. This follows most easily from the de Moivre formula
cos nθ + i sin nθ = (cos θ + i sin θ)n .
Expanding the right-hand side by the binomial theorem and taking real
parts, one finds

n n
cos nθ = cos θ −
n
cos n−2 2
θ sin θ + cosn−4 θ sin4 θ − . . . .
2 4
But sin2 θ = 1 − cos2 θ, so we can write cos nθ = Tn (cos θ) for a uniquely
determined polynomial
Tn (x) = cn xn + cn−1 xn−1 + . . . + c1 x + c0
of degree n, known as the nth Chebyshev polynomial. The “T” is for
“Tchebicheff”, a French transliteration of “Chebyshev”.
P. L. Chebyshev (1821–1894) was a primary founder of the Russian
school of analysis. In addition to his work in approximation theory, he made
seminal contributions to number theory, probability theory, and mechanics.
He introduced the Chebyshev polynomials in 1854.
The first few Chebyshev polynomials can be calculated explicitly from
the preceding formulas. They are
T0 (x) = 1 , T1 (x) = x , T2 (x) = 2x2 − 1 , T3 (x) = 4x3 − 3x ,
T4 (x) = 8x4 − 8x2 + 1 , T5 (x) = 16x5 − 20x3 + 5x , ... .
It is easier to calculate the Chebyshev polynomials recursively by the relation
Tn+1 (x) = 2xTn (x) − Tn−1 (x) , n = 2, 3, . . . .
To derive the recurrence relation, introduce the definition cos nθ = Tn (cos θ)
into the identity
cos(n + 1)θ + cos(n − 1)θ = 2 cos θ cos nθ
and set x = cos θ. An inductive argument now shows that Tn (x) has leading
coefficient cn = 2n−1 for n > 0. Additional properties are collected in the
following lemma.
Lemma 1. For n > 0 the Chebyshev polynomial Tn (x) has simple zeros at
the points
2k − 1
xk = cos π , k = 1, 2, . . . , n .
2n
In the interval [−1, 1], the local maxima of |Tn (x)| occur at the points
ξk = cos(kπ/n) , k = 0, 1, . . . , n ,
and Tn (ξk ) = (−1)k .
Proof. Note that Tn (x) = cos nθ = 0 for

2k − 1 2k − 1
θ= π, or x = xk = cos π , k = 1, 2, . . . , n .
2n 2n
A simple calculation shows that
d
n
Tn (x) = cos n cos−1 x = √ sin n cos−1 x = 0
dx 1 − x2
for
cos−1 x = kπ/n , or x = ξk = cos(kπ/n) , k = 0, 1, . . . , n .
In particular, Tn (xk ) = 0, so all of the zeros are simple. Finally,

Tn (ξk ) = cos n cos−1 ξk = cos(kπ) = (−1)k
Figure 1 shows graphs of the Chebyshev polynomials of degrees 10 and

15 over the interval [−1, 1].
Figure 1. Chebyshev polynomials Tn (x) for n = 10, 15.

We can now show that among all monic polynomials of fixed degree, the
normalized Chebyshev polynomials have smallest uniform norm.
Theorem 2. Every monic polynomial
Q(x) = xn + an−1 xn−1 + . . . + a1 x + a0
of degree n > 0 satisfies the inequality
Q∞ = max |Q(x)| ≥ 21−n = 21−n Tn ∞ ,

−1≤x≤1
where Tn is the nth Chebyshev polynomial.
The theorem says in particular that for all choices of points x1 , x2 , . . . , xn

on the real line, the function
(x) = (x − x1 )(x − x2 ) · · · (x − xn )
has smallest norm for

2k − 1
xk = cos π , k = 1, 2, . . . , n .
2n
It is clear a priori that for minimum norm all of the points xk must lie in
the interval [−1, 1], since |(x)| is a product of distances. Another interpre-
tation of the theorem is that among all polynomials of degree ≤ n − 1, the
polynomial
Pn (x) = xn − 21−n Tn (x)
gives the best uniform approximation to the function xn over the interval
[−1, 1].
Proof of Theorem 2. Let T n = 21−n Tn and suppose that Q∞ < T n ∞

for some monic polynomial Q of degree n. According to Lemma 2, the
maximum of |T n (x)| occurs at the points ξk = cos(kπ/n), with alternating
signs. Specifically,
n (ξk ) = 21−n = T
(−1)k T n ∞ , k = 0, 1, . . . , n , .
Therefore, Q(ξk ) < Tn (ξk ) when k is even, and Q(ξk ) > Tn (ξk ) when k is
odd. This implies that the difference T n (x) − Q(x) vanishes at some point in
each of the n intervals (ξ1 , ξ0 ), (ξ2 , ξ1 ), . . . , (ξn , ξn−1 ). But this is impossible,
because Tn and Q are both monic polynomials of degree n, so T n − Q is a
polynomial of degree ≤ n − 1 and can have at most n − 1 zeros. Thus we
have arrived at a contradiction, which proves the theorem.
6.2. Weierstrass approximation theorem 151
6.2. Weierstrass approximation theorem
It can be truly said that the theory of approximation was launched in

1885 with Weierstrass’s discovery that every continuous function can be
approximated uniformly by polynomials. Here is a precise statement.
Theorem 3 (Weierstrass Approximation Theorem). If a function

f (x) is continuous on a closed bounded interval [a, b], then for each ε > 0
there exists a polynomial P (x) such that |f (x) − P (x)| < ε for all x in [a, b].
Karl Weierstrass (1815–1897) was 70 years old at the time of his discov-
ery. In fact, the theorem was discovered independently at almost the same
time by Carl Runge (1856–1927), who had been a student of Weierstrass
five years earlier. Runge obtained the result as part of his groundbreaking
work on approximation theory in the complex domain.
The theorem of Weierstrass is remarkable because continuous functions
can have a much more complicated behavior than polynomials. For instance,
in 1872 Weierstrass had constructed examples of continuous functions that
are not differentiable at any point. If the function f is sufficiently smooth,
it can be approximated locally by partial sums of its Taylor series, but
in general these polynomials will not provide the global approximation of
the Weierstrass theorem. A more promising idea is to construct from f its
Lagrange interpolation polynomials at n equally spaced points in the interval
[a, b]. As n → ∞, it is reasonable to expect this sequence of polynomials
to converge uniformly to f . However, these interpolation polynomials need
not converge pointwise to f , even if f has derivatives of all orders. In
fact, the interpolation polynomials can fail to be uniformly bounded. This
surprising state of affairs makes the Weierstrass theorem appear all the more
remarkable.
Runge gave a striking example in 1901 that illustrates the failure of
approximation by Lagrange interpolation polynomials. For the function
f (x) = 1/(1+x2 ), Runge considered the interpolation polynomials at equally
spaced nodes over the interval [−5, 5] and showed that they are unbounded
for c < |x| < 5, where c = 3.63 . . . . Sergei Bernstein showed in 1912 that for
the function f (x) = |x| in the interval [−1, 1] the interpolation polynomials
at equally spaced nodes converge only at the points −1, 0, and 1. One might
suspect, in light of the results discussed in the preceding section, that the
bad behavior in these examples is the fault of equally spaced nodes, and that
nodes at the scaled roots of Chebyshev polynomials would produce better
results. Indeed, this is true when the target function f is sufficiently smooth;
for instance, when it has a continuous derivative. However, it is a theorem
of Georg Faber, proved in 1914, that for any preassigned sequence of nodes
in a given interval, there is some continuous function whose interpolation

polynomials are unbounded in that interval.
Many proofs of the Weierstrass theorem are known. We shall present
three of them, all quite elegant, which shed light on various aspects of the
problem. The first of these proofs, due to Lebesgue [5], is ingenious but quite
simple. The second proof, due to Landau [4], proceeds by convolution of the
given continuous function f with a polynomial “peaking kernel”. Landau’s
proof closely resembles the original argument of Weierstrass (details to be
discussed later), but is more direct. The third proof, due to Sergei Bernstein
[1], is motivated by the theory of probability. It uses the values of f at
equally spaced points to construct a sequence of polynomials, known as the
Bernstein polynomials, that are shown to converge uniformly to f . The
proofs by Landau and Bernstein are constructive in the sense that they
produce explicit sequences of polynomials which converge uniformly to the
given continuous function.
Henri Lebesgue (1875–1941) is famous for developing what is now called
the Lebesgue integral. It extends the Riemann integral to a wider class of
functions and has become an essential tool in mathematical analysis. Here
is Lebesgue’s proof of the Weierstrass theorem.
Lebesgue’s proof. This proof has a striking feature. It reduces the ap-
proximation of an arbitrary continuous function f to the approximation of
a single function, namely f (x) = |x|. The reduction proceeds as follows. If a
function f is continuous at each point of the closed bounded interval [a, b], it
is uniformly continuous there. Therefore, f can be approximated uniformly
by a continuous piecewise linear function g whose graph is a polygonal path
connecting finitely many points (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) on the curve
y = f (x), where
a = x1 < x2 < · · · < xn = b .
Observe that the function
ϕk (x) = |x − xk | + (x − xk )
is equal to 0 for x ≤ xk and is equal to 2(x − xk ) for x ≥ xk . Consider the

function

n−1
h(x) = c0 + cj ϕj (x) ,
j=1
where the constants cj are to be determined. The function h varies linearly

in each of the intervals [xk , xk+1 ], so it will coincide with the function g if
the constants cj can be chosen such that h(xk ) = yk for k = 1, 2, . . . , n.
These conditions can be written in the form

k−1
c0 + 2 cj (xk − xj ) = yk , k = 1, 2, . . . , n ,
j=1
and so a unique solution (c0 , c1 , . . . , cn ) can be calculated recursively. With

this choice of constants, h(x) = g(x) in [a, b]. Thus the uniform approx-
imation of g(x) by polynomials reduces to the uniform approximation of
each function ϕk (x), hence to that of |x − xk |, and finally to the uniform
approximation of |x| by polynomials in [a, b]. This last problem is trivial
unless a < 0 < b, and then it suffices to consider a symmetric interval,
say [−b, b]. After the substitution x = bt, the problem reduces to uniform
approximation of |t| by polynomials in the interval [−1, 1].
To perform the last approximation, begin with the binomial expansion
√
1 − x = 1 − 12 x − 18 x2 − . . . ,
recalling that the series converges uniformly in the interval [−1, 1]. (See
Chapter 3, Section 3.5.) Now replace x by 1 − t2 to see that
|t| = 1 − 12 (1 − t2 ) − 18 (1 − t2 )2 − . . . ,
and the convergence is again uniform in [−1, 1]. This shows that |t| can be
approximated uniformly by partial sums of the series, which are polynomials
in t, and the proof is complete.
6.3. Landau’s proof

Edmund Landau (1877–1938) was primarily an analytic number theo-
rist who discovered many new results on the distribution of prime num-
bers. His two-volume work Handbuch der Lehre von der Verteilung der
Primzahlen (1909) was the first systematic presentation of analytic number
theory. Landau was a perfectionist known for the beauty and clarity of his
writing. His little books Grundlagen der Analysis (1930) and Darstellung
und Begründung einiger neuerer Ergebnisse der Funktionentheorie (1916)
are classics. In addition to his original discoveries, he was noted for finding
simpler and more elegant proofs of known results. His proof of the Weier-
strass approximation theorem is no exception.
Landau’s proof. The following streamlined version of Landau’s proof is

adapted from the book of Rudin [11]. It is sufficient to take the interval
[a, b] to be [0, 1], since the general result then follows by translation and
change of scale. It is convenient to assume also that f (0) = f (1) = 0. This
assumption entails no loss of generality, because the general result is then

deduced by first subtracting a linear polynomial to produce a function
g(x) = f (x) − f (0) − [f (1) − f (0)] x
continuous on [0, 1] with the property g(0) = g(1) = 0.

After these reductions, consider the functions
Kn (x) = cn (1 − x2 )n , n = 1, 2, 3, . . . ,
where the positive constants cn are chosen so that

1
Kn (x) dx = 1 .
−1
Note that Kn (x) is a polynomial of degree 2n and is an even function, mean-

ing that Kn (−x) = Kn (x). It is also evident that 0 < Kn (x) ≤ Kn (0) = cn
for −1 < x < 1. The functions Kn (x) peak more and more sharply at the
origin as n increases. Some sample graphs are shown in Figure 2.
One further property of the functions Kn (x), suggested by Figure 2, is
that Kn (x) → 0 uniformly in the set δ ≤ |x| ≤ 1 for each fixed δ > 0. To see
this, we will need an estimate on the rate of growth of the constant factors
cn . Direct calculation of the integral and an appeal to Stirling’s formula (cf.
Chapter 9, Exercise 24) shows in fact that
(2n + 1)!
cn = 2n+1 2
∼ n/π , n → ∞,
2 (n!)
x
-1 1
Figure 2. The peaking kernel Kn (x) for n = 5, 10, 20.

√
but the more elementary estimate cn < n will suffice for our purpose.
Recall first that the inequality (1 + t)n ≥ 1 + nt holds for all t > −1, a fact
easily verified by induction. Hence
(1 − x2 )n ≥ 1 − nx2 , |x| < 1 .
Using this inequality, we find that

1 1 1/√n
1
= (1 − x ) dx = 2
2 n
(1 − x ) dx ≥ 2
2 n
(1 − x2 )n dx
cn −1 0 0
1/√n
4 1
≥2 (1 − nx2 ) dx = √ > √ ,
0 3 n n
√
which shows that cn < n for n = 1, 2, 3, . . . . It follows that for 0 < δ < 1,
√
Kn (x) = cn (1 − x2 )n ≤ n(1 − δ 2 )n , δ ≤ |x| ≤ 1 .
Thus Kn (x) → 0 as n → ∞, uniformly in each set δ ≤ |x| ≤ 1.

Now define the convolution
1
Pn (x) = Kn (x − t)f (t) dt , 0 ≤ x ≤ 1.
0
For each fixed x the number Pn (x) may be viewed as a “weighted average”
of the numbers f (t) as t ranges over the interval [0, 1], with greatest weight
attached (when n is large) to the values of t near x. Intuitively, then, it is
to be expected that Pn (x) → f (x) as n → ∞. Because Kn is a polynomial,
it is easy to see that Pn is also a polynomial.
In order to prove that Pn (x) → f (x) uniformly in [0, 1], it is convenient
first to extend the given function by setting f (x) = 0 for all x outside the
interval [0, 1]. The extended function is continuous on R because of our
initial assumption that f (0) = f (1) = 0. Since Kn is an even function, the
interval of integration can be shifted to give
1 1−x
Pn (x) = Kn (t − x)f (t) dt = Kn (t)f (x + t) dt .
0 −x
But f (x + t) = 0 for t ≤ −x and for t ≥ 1 − x, so we can write

1
Pn (x) = Kn (t)f (x + t) dt .
−1
This shows that

1
Pn (x) − f (x) = Kn (t)[f (x + t) − f (x)] dt ,
−1
1
since −1 Kn (t) dt = 1. Now write |f (x)| ≤ M in [0, 1], and invoke the
basic theorem that a function continuous at each point of a closed bounded
interval is uniformly continuous there. Thus to each ε > 0 there corresponds
a number δ > 0 such that
|f (x + t) − f (x)| < ε whenever x ∈ [0, 1] and |t| < δ .
Therefore, recalling the estimate

√
Kn (t) ≤ n(1 − δ 2 )n , δ ≤ |t| ≤ 1 ,
we find
1

|Pn (x) − f (x)| = Kn (t)[f (x + t) − f (x)] dt
−1
1
≤ Kn (t)|f (x + t) − f (x)| dt
−1
−δ δ
≤ 2M Kn (t) dt + Kn (t)|f (x + t) − f (x)| dt
−1 −δ
1
+ 2M Kn (t) dt
δ
δ
√
≤ 4M n(1 − δ ) + ε
2 n
Kn (t) dt
−δ
√
< 4M n(1 − δ 2 )n + ε < 2ε
for all n sufficiently large. Thus Pn (x) → f (x) as n → ∞, uniformly in

[0, 1]. This completes Landau’s proof of the theorem.
The main idea of Landau’s proof, convolution with a peaking kernel, was
already present (as Landau [4] acknowledged) in the original proof given by
Weierstrass [15]. But instead of a polynomial peaking kernel, Weierstrass
used the kernel e−(x/a) with a > 0, for which
2
∞
√
e−(x/a) dx = a π .
2
−∞
By a similar argument he could show that the integrals

∞
1 x−t 2
√ e−( a ) f (t) dt
a π −∞
converge uniformly to f (x) as a → 0. He then approximated the integrals
uniformly by taking partial sums of the Taylor series of the kernel e−(x/a) ,
2
and so approximated f (x) uniformly by polynomials.

6.4. Bernstein polynomials

Sergei Bernstein (1880–1968) was born in the Ukraine but got his early
mathematical training in Paris. He then returned to the Ukraine and even-
tually held positions in Leningrad and Moscow. He did basic work in the
area of interpolation and approximation and made important contributions
to the theory of probability. In his proof of the Weierstrass theorem he in-
troduced the Bernstein polynomials, as they are now called, which turned
out to have many interesting properties.
Bernstein’s proof. Again we may assume that the underlying interval is
[0, 1]. Given a function f (x) continuous on [0, 1], we propose to show that
the Bernstein polynomials
n
n
Bn (x) = f (k/n) xk (1 − x)n−k
k
k=0
converge uniformly to f (x). Here

n n!
=
k k!(n − k)!
is a binomial coefficient and 0! = 1.
We now digress to discuss the heuristic source of the Bernstein polyno-
mials. As indicated by the title of Bernstein’s paper [1], the construction is
motivated by probability theory, although the proof of uniform convergence
makes no essential reference to these concepts. Let x be a fixed number with
0 < x < 1. If a point is chosen at random from the interval [0, 1], it will lie
in the interval [0, x] with probability x and in the complementary subinter-
val [x, 1] with probability 1 − x. If n points are chosen independently, the
probability that exactly k of them lie in the interval [0, x] is given by the
binomial distribution

n k
b(k; n, x) = x (1 − x)n−k , k = 0, 1, . . . , n .
k
Consider now a random variable Yn that has the value f (k/n) if k of the
chosen points are in [0, x]. The expectation of Yn is

n
E(Yn ) = f (k/n)b(k; n, x) = Bn (x) .
k=0
It is plausible (and can be proved on probabilistic grounds, by appeal to the

“strong law of large numbers”) that E(Yn ) → f (x) as n → ∞.
For an analytic proof, we will need some standard facts about the bino-
mial distribution, as stated in the following lemma.
Lemma 2.

n
(a) b(k; n, x) = 1 ;
k=0

n
(b) (k − nx)2 b(k; n, x) = nx(1 − x) .
k=0
The proof of (a) is the simple observation that by the binomial theorem,
the sum is equal to [x + (1 − x)]n = 1. The sum in (b) is a calculation of the
variance of the binomial distribution; the proof is deferred for the moment.
In view of (a), we have

n
Bn (x) − f (x) = [f (k/n) − f (x)] b(k; n, x) .
k=0
The proof that Bn (x) tends uniformly to f (x) appeals again to the uni-
form continuity of f on [0, 1]. For each ε > 0, there is a δ > 0 such that
|f (x) − f (y)| < ε for all pairs of points x, y ∈ [0, 1] with |x − y| < δ. Again
suppose |f (x)| ≤ M for all x in the interval [0, 1], and write

n
|Bn (x) − f (x)| ≤ |f (k/n) − f (x)| b(k; n, x) = S1 + S2 ,
k=0
k
where the sum S1 extends over all integers k (0 ≤ k ≤ n) with − x < δ
n
and the sum S2 is taken over those integers k with nk − x ≥ δ. By the
uniform continuity of f and part (a) of the lemma, we see that S1 < ε. On
the other hand,

n
(k/n − x)2
S2 ≤ 2M b(k; n, x) ≤ 2M b(k; n, x)
δ2
|n
k
−x|≥δ k=0
2M
n
2M x(1 − x) 2M
= 2 2
(k − nx)2 b(k; n, x) = 2
≤ 2 <ε
δ n δ n δ n
k=0
for all n ≥ N > 2M/(δ 2 ε), where part (b) of the lemma has been used.
Observe that N depends only on ε, so we have shown that
|Bn (x) − f (x)| ≤ 2ε for all x ∈ [0, 1]
whenever n ≥ N . This shows that Bn (x) → f (x) uniformly in [0, 1], which
completes Bernstein’s proof of the Weierstrass approximation theorem.
Proof of Lemma 2. It was already observed that
n

n
n
1 = [x + (1 − x)] = n
x (1 − x)
k n−k
= b(k; n, x) ,
k
k=0 k=0
which proves (a). To prove (b), we first calculate

n n

n−1
μ= k b(k; n, x) = nx xk−1 (1 − x)n−k
k−1
k=0 k=1

n−1
= nx b(j; n − 1, x) = nx ,
j=0
by (a), after the shift of index j = k − 1. Next observe that with μ = nx

n
n
(k − nx) b(k; n, x) =
2
(k 2 − 2kμ + μ2 )b(k; n, x)
k=0 k=0

n
= k 2 b(k; n, x) − μ2 .
k=0
But

n
n
2
k b(k; n, x) = μ + k(k − 1)b(k; n, x)
k=0 k=2
n

n−2
= μ + n(n − 1)x 2
xk−2 (1 − x)n−k
k−2
k=2

n−2
= μ + n(n − 1)x2 b(j; n − 2, x)
j=0
= μ + n(n − 1)x = nx(1 − x) + μ2 ,

2
which combines with the previous formula to yield the desired result.
The Bernstein polynomials have the remarkable property that for func-
(ν)
tions f with continuous νth derivative, Bn (x) → f (ν) (x) uniformly in [0, 1].
(See Exercise 14.) The books by Cheney [2], Davis [3], and Lorentz [6] may
be consulted for this and other facts about Bernstein polynomials.
6.5. Best approximation

The theorem of Weierstrass is a qualitative statement that uniform ap-
proximation by polynomials is possible. For any given function f continuous
over an interval [a, b], it says that the uniform distance
f − P ∞ = max |f (x) − P (x)|
a≤x≤b
can be made arbitrarily small with some choice of polynomial P . The theo-
rem suggests the problem of finding the best approximation by polynomials
of fixed degree. More precisely, it is of interest to determine the quantity
En (f ) = inf f − P ∞ ,
P
where the infimum is taken over all polynomials of degree ≤ n. Is the

infimum attained, and if so what are the polynomials of best approximation?
We will see presently that an extremal polynomial of degree ≤ n always
exists and is unique. On the other hand, no explicit formula is known for
this polynomial of best approximation to an arbitrarily prescribed function
f that is continuous on a specified interval [a, b]. To find such a formula is an
open problem of long standing. It turns out, however, that the polynomial
of best approximation can be characterized by the alternation property al-
ready encountered in the special case of Chebyshev polynomials. Deferring
a precise statement, we begin with some general remarks about polynomials.
According to the Lagrange interpolation formula, any polynomial P of
degree ≤ n can be written as

n
P (xk ) (x)
n
P (x) =
, (x) = (x − xj ) ,
(xk )(x − xk )
k=0 j=0
for any system {x0 , x1 , . . . , xn } of distinct points. The representation has

the form
n
P (x) = P (xk ) qk (x) ,
k=0
where the functions qk (x) are fixed polynomials of degree ≤ n, depending
only on the points xj . This formula makes clear that if a sequence of polyno-
mials of degree ≤ n is bounded at n + 1 specified points (in particular, if the
polynomials are uniformly bounded on an interval), then the corresponding
sequences of coefficients are also bounded.
Theorem 4. For any function f continuous over an interval [a, b], and
for any positive integer n, there exists a polynomial P of degree ≤ n for
which the distance f − P ∞ attains a minimum value. Furthermore, this
polynomial of best approximation is unique.
Proof of existence. Since En (f ) is the infimum of f − P ∞ , there is a

sequence of polynomials

n
Pj (x) = ajk xk , j = 1, 2, . . . ,
k=0
of degree ≤ n such that f − Pj ∞ → En (f ) as j → ∞. Then the norms

Pj ∞ are bounded, so by the preceding remarks the coefficients ajk are
also bounded. Thus by the Bolzano–Weierstrass theorem, some subsequence
{ajν k } of the coefficients converges, say to ak , for each k. Then the corre-
sponding polynomials Pjν (x) converge to

n
P (x) = ak xk ,
k=0
uniformly in the interval [a, b]. Therefore,

En (f ) = lim f − Pjν ∞ = f − P ∞ ,
ν→∞
which shows that P is a polynomial of best approximation.
The uniqueness property will be derived from the following theorem.

Theorem 5. Let f be a continuous function on an interval [a, b], and sup-
pose P is a polynomial of degree ≤ n that gives the best uniform approxima-
tion to f on [a, b], so that f − P ∞ = En (f ). Then there are n + 2 points
xk in [a, b],
a ≤ x1 < x2 < · · · < xn+2 ≤ b ,
at which f (xk ) − P (xk ) = ± En (f ) with alternating signs; that is,
f (xk+1 ) − P (xk+1 ) = −[f (xk ) − P (xk )] , k = 1, 2, . . . , n + 1 .
Before embarking on the proof, let us remark that the Chebyshev poly-
nomials provide a special case of this alternating property. As previously
noted, Theorem 2 can be interpreted as saying that among all polynomials
of degree ≤ n − 1,
P (x) = xn − 21−n Tn (x)
gives the best uniform approximation to the function f (x) = xn over the
interval [−1, 1]. Lemma 1 then confirms that
f (x) − P (x) = 21−n Tn (x)
attains its maximum absolute value in the interval [−1, 1], with alternating
signs at n + 1 points ξk = cos(kπ/n) , for k = 0, 1, . . . , n. Note that n is
replaced here by n − 1.
Proof of Theorem 5. For convenience, set ϕ(x) = f (x) − P (x) and δ =

En (f ). Suppose that δ > 0, excluding the trivial case where f is itself a
polynomial of degree ≤ n. If ϕ(x) has the same sign wherever |ϕ(x)| = δ,
then the approximation can be improved by adding to P a small constant of
that sign. But this is impossible if P is a polynomial of best approximation.
Thus there must be some variation of signs at points where |ϕ(x)| = δ.
Because the function ϕ is uniformly continuous, the interval [a, b] can
be divided into a finite number of subintervals in such a way that ϕ(x) does
not vanish in any interval where |ϕ| attains its maximum value δ. Let
I1 = [a1 , b1 ] , I2 = [a2 , b2 ] , . . . , Im = [am , bm ]
denote the subintervals in which |ϕ| attains the value δ , with numbering
chosen so that
a ≤ a1 < b1 ≤ a2 < b2 ≤ · · · ≤ am < bm ≤ b .
Let σk denote the sign of ϕ(x) in Ik ; in other words, σk = +1 if ϕ(x) > 0

in Ik , and σk = −1 if ϕ(x) < 0 in Ik . As previously noted, the sequence
{σ1 , σ2 , . . . , σm } must display some variation of sign. If σk+1 = −σk , then
bk < ak+1 and |ϕ(x)| < δ in the interval (bk , ak+1 ). Thus after slight en-
largement of the intervals Ik and Ik+1 if necessary, we may assume that
|ϕ(bk )| < δ and |ϕ(ak+1 )| < δ while maintaining the inequality bk < ak+1
together with the property that ϕ(x) = 0 in Ik and Ik+1 . After these ad-
justments have been made, we can assert that |ϕ(x)| ≤ δ < δ for all x in
the complementary set
S = [a, b] \ (I1 ∪ I2 ∪ · · · ∪ Im ) .
We are to prove that the sequence {σ1 , σ2 , . . . , σm } exhibits at least

n + 1 changes of sign from an entry σk to its successor σk+1 . Suppose, on
the contrary, that there are only ν ≤ n changes of sign, and let k1 , k2 , . . . , kν
denote the indices where these changes occur. Thus
σkj +1 = −σkj , j = 1, 2, . . . , ν .
Then we can choose points ξj ∈ S so that
bkj < ξj < akj +1 , j = 1, 2, . . . , ν ,
and form the function
q(x) = σk1 (ξ1 − x)(ξ2 − x) · · · (ξν − x) ,

which is a polynomial of degree ν ≤ n. By construction, q(x) has the same

sign as ϕ(x) in each of the intervals I1 , I2 , . . . , Im .
We now claim that for sufficiently small ε > 0, the polynomial Q = P +εq
provides a better approximation:
f − Q∞ < f − P ∞ = δ .
This will contradict the extremal property of P , thus proving the theorem.
To establish the claim, choose ε > 0 so small that
ε|q(x)| < δ − δ for all x ∈ [a, b] .
Then
|f (x) − Q(x)| = |ϕ(x) − εq(x)| < δ
for all x ∈ (I1 ∪ I2 ∪ · · · ∪ Im ), since ϕ(x) and q(x) have the same sign there.
On the other hand, if x ∈ S, then |ϕ(x)| ≤ δ and so
|f (x) − Q(x)| = |ϕ(x) − εq(x)| ≤ |ϕ(x)| + ε|q(x)|
< δ + (δ − δ ) = δ .
Combining the two estimates, we see that
f − Q∞ < f − P ∞ ,
which contradicts the hypothesis that P is a polynomial of degree ≤ n of
best approximation to f . This shows that the sequence {σ1 , σ2 , . . . , σm }
must have at least n + 1 changes of sign, as the theorem asserts.
As a consequence of Theorem 5, we can now infer that the polynomial

of best approximation is unique. Suppose that two polynomials P and Q of
degree ≤ n both give the best approximation to f :
f − P ∞ = f − Q∞ = En (f ) .
Then the polynomial R = 12 (P + Q) also gives the best approximation, since
f − R∞ ≤ 12 f − P ∞ + 12 f − Q∞ = En (f ) .
In view of Theorem 5, it follows that |f (x) − R(x)| attains its maximum
value En (f ) at n + 2 distinct points of the interval [a, b]. But since

f (x) − R(x) = 12 f (x) − P (x) + 12 f (x) − Q(x) ,
an application of the triangle inequality shows that
f (x) − P (x) = f (x) − Q(x) = f (x) − R(x)
at any point where |f (x) − R(x)| = En (f ). Thus P (x) = Q(x) at n + 2
distinct points, which implies that P (x) ≡ Q(x), since both are polynomials
of degree ≤ n. This proves the uniqueness.
Finally, we can show that the alternation property of Theorem 5 is not

only necessary for P to be the polynomial of best approximation, but suffi-
cient as well.
Theorem 6. The alternation property of Theorem 5 characterizes the poly-

nomial P of degree ≤ n of best approximation. More precisely, if f is con-
tinuous on [a, b] and Q is a polynomial of degree ≤ n such that f (x) − Q(x)
attains its maximum absolute value at n + 2 distinct points in [a, b] with
alternating signs, then Q = P .
Proof. Let d = f − Q∞ and suppose that
|f (xk ) − Q(xk )| = d for a ≤ x1 < x2 < . . . xn+2 ≤ b .
Suppose also that the signs alternate, so that
f (xk+1 ) − Q(xk+1 ) = −[f (xk ) − Q(xk )] , k = 1, 2, . . . , n + 1 .
Since P is the polynomial of best approximation, it is clear that d ≥ δ =

f − P ∞ . If d > δ, then the polynomial
P − Q = (f − Q) − (f − P )
has the sign of f − Q at the points xk . This implies that P − Q changes

sign at least n + 1 times, and so it has at least n + 1 distinct zeros. But
P − Q is a polynomial of degree ≤ n, so it follows that P (x) − Q(x) ≡ 0, or
Q = P . This is not possible, however, if d > δ. Thus d = δ and Q = P by
uniqueness of the polynomial of best approximation.
The alternation results of Theorems 5 and 6, which together character-

ize the polynomial of best approximation, go back at least to Émile Borel
(1905) and Charles de la Vallée Poussin (1910). The exposition of this sec-
tion is adapted from the book by de la Vallée Poussin [14], which contains
much more information. Cheney [2] presents the same material in terms of
“Haar systems”, which are more general than polynomials. The alternation
criterion is the basis for an algorithm, developed by Evgeny Remez in 1934,
for the numerical calculation of the polynomial of best approximation. For
a full description the reader is referred to the book by Cheney [2].
6.6. Stone–Weierstrass theorem

The Weierstrass approximation theorem can be generalized to several
variables. Weierstrass showed that a continuous function f (x1 , x2 , . . . , xn )
can be approximated uniformly on any closed bounded subset of Rn by linear
combinations of monomials xk11 xk22 · · · xknn , where the exponents k1 , k2 , . . . , kn

are nonnegative integers. Landau also extended the theorem to several vari-
ables, by a simple inductive argument.
The extension to several variables is actually a special case of a far-
reaching generalization of Weierstrass’s theorem discovered in 1937 by the
American mathematician Marshall Stone (1903–1989). Ten years later,
Stone wrote an expository article [12] giving a simpler and more elemen-
tary proof. The theorem in generalized form, now known as the Stone–
Weierstrass theorem, has a wide range of applications. In order to state it,
we need to define a few terms. A family A of real-valued functions defined
on an abstract set E is said to be an algebra if f, g ∈ A and c ∈ R imply
that f + g ∈ A, f g ∈ A, and cf ∈ A. A family A of functions defined on E
is said to separate points if for each pair of distinct points x, y ∈ E there is
a function f ∈ A for which f (x) = f (y). The family A is said to vanish at
a point x ∈ E if f (x) = 0 for every f ∈ A.
Theorem 7 (Stone–Weierstrass Theorem). Let A be an algebra of real-

valued continuous functions on a compact subset E of a metric space. If
A separates points and vanishes at no point of E, then every continuous
function can be approximated uniformly on E by functions in A.
It is clear that the set of all polynomials P (x) in one real variable is an
algebra that separates points in any interval [a, b] and does not vanish at
any point, so the classical approximation theorem of Weierstrass is a special
case of the Stone–Weierstrass theorem. The set of polynomials P (F (x)) of
a fixed function F is a more general example of an algebra that does not
vanish at any point, but it separates points only if F is univalent on E. As
a special case, the set of all even polynomials is an algebra that does not
separate points in any symmetric interval [−a, a].
Recall that a subset of a metric space is said to be compact if it has the
Heine–Borel property: each open covering contains a finite subcovering. For
the sake of simplicity, we will prove the Stone–Weierstrass theorem under
the special assumption that E is a closed bounded subset of the real line,
but the argument generalizes with little change to arbitrary metric spaces.
A simple algebraic lemma will be needed.
Lemma 3. Let A be an algebra of real-valued functions on a set E that

separates points and does not vanish at any point. Then for each pair of
distinct points x1 , x2 ∈ E and each pair of numbers y1 , y2 ∈ R, there is a
function f ∈ A for which f (x1 ) = y1 and f (x2 ) = y2 .
Proof of lemma. The algebra A separates points, so it contains a function

g for which g(x1 ) = g(x2 ). Because A vanishes at no point of E, there are
functions h1 and h2 in A with h1 (x1 ) = 0 and h2 (x2 ) = 0. Define
ϕ = [g − g(x1 )] h2 , ψ = [g − g(x2 )] h1 .
Then ϕ, ψ ∈ A and ϕ(x1 ) = ψ(x2 ) = 0, whereas ϕ(x2 ) = 0 and ψ(x1 ) = 0.

Thus the function
y1 y2
f= ψ+ ϕ
ψ(x1 ) ϕ(x2 )
belongs to A and has the properties f (x1 ) = y1 , f (x2 ) = y2 .
Proof of theorem. Let B denote the uniform closure of A ; that is, the set
of all functions uniformly approximable by functions in A. More precisely,
g ∈ B if for each ε > 0 there exists f ∈ A such that |g(x) − f (x)| < ε for
all x ∈ E. It is easy to see that B is an algebra of continuous functions. We
are to prove that B is the algebra of all continuous functions on E.
The first step is to show that |f | ∈ B whenever f ∈ B. To see this, let
M = sup{|f (x)| : f ∈ E} and let ε > 0. By the Weierstrass approximation
theorem, there is a polynomial P such that
|P (y) − |y|| < ε , y ∈ [−M, M ] .
This implies that
|P (f (x)) − |f (x)|| < ε , x∈E.
But B is an algebra, so P ◦f ∈ B if f ∈ B. Thus |f | ∈ B, since B is uniformly

closed.
It now follows that
max(f, g) = 12 (f + g) + 12 |f − g| and
min(f, g) = 1
2 (f + g) − 2 |f
1
− g|
belong to B whenever f, g ∈ B. Iteration gives the more general result that

max{f1 , f2 , . . . , fn } and min{f1 , f2 , . . . , fn } are in B if f1 , f2 , . . . , fn ∈ B.
Next we claim that for each function f continuous on E, for each point
s ∈ E, and for each ε > 0, there exists a function gs ∈ B such that gs (s) =
f (s) and gs (x) > f (x) − ε for all x ∈ E. To see this, observe that because
A is a subalgebra of B that separates points and does not vanish at any
point, B has the same properties. For an arbitrarily specified point t ∈ E,
the lemma then shows that some function ht ∈ B satisfies
ht (s) = f (s) and ht (t) = f (t) .

By the continuity of ht and f , it follows that ht (x) > f (x) − ε for all x in
some open set Ut containing t. But the collection of sets {Ut : t ∈ E} forms
an open covering of the compact set E, so the Heine–Borel theorem says
that
E ⊂ Ut1 ∪ Ut2 ∪ · · · ∪ Utn
for some finite set of points t1 , t2 , . . . , tn ∈ E. Now define
h = max{ht1 , ht2 , . . . , htn } .
Then h ∈ B and it has the properties h(s) = f (s) and h(x) > f (x) − ε for
all x ∈ E, by construction of the functions ht and the neighborhoods Ut .
Thus with gs = h our claim is established.
The final step is to show that every continuous function is uniformly
approximable by functions in B, or equivalently that f ∈ B. The proof
involves another appeal to the Heine–Borel theorem. Given a function f
continuous on E, a point s ∈ E, and a number ε > 0, we have constructed
a function gs ∈ B for which gs (s) = f (s) and gs (x) > f (x) − ε for all x ∈ E.
By continuity, the inequality gs (x) < f (x) + ε persists in some open set Vs
containing s. The sets Vs form an open covering of E, so there is a finite
subcovering:
E ⊂ Vs1 ∪ Vs2 ∪ · · · ∪ Vsm
for some finite collection of points s1 , s2 , . . . , sm . Now define
g = min{gs1 , gs2 , . . . , gsm } .
Then g ∈ B and g(x) < f (x) + ε for all x ∈ E, by the construction of g and
the sets Vs . On the other hand, g(x) > f (x) − ε for all x ∈ E because each
function gs has this property. Combination of the two inequalities gives
|f (x) − g(x)| < ε for all x ∈ E ,
which completes the proof of the Stone–Weierstrass theorem.
The theorem is not true as stated for algebras of complex-valued func-

tions, but it does hold for self-conjugate algebras. In other words, the the-
orem remains true for complex algebras under the additional hypothesis
that f ∈ A whenever f ∈ A, where f (x) = f (x) is the complex conjugate
function. For further details the reader may consult Davis [3] or Rudin [11].
References to additional proofs of the Weierstrass approximation theo-
rem and further discussion may be found in Cheney [2] and Davis [3]. The
article by Pincus [10] discusses the historical evolution of ideas leading to
the discovery of the theorem.
6.7. Refinements of Weierstrass theorem

We conclude this chapter with some sharpened versions of the Weier-
strass approximation theorem. The primary result, found in 1914 by Julius
Pál [8], asserts that the approximation can be achieved by polynomials with
prescribed initial coefficients. As Pál [8, 9] observed, this comparatively
simple result has some astonishing implications.
Theorem 8 (Pál’s Theorem). Let f be a continuous function on an
interval [−r, r] for some r > 0, and let a0 = f (0). Then for each ε > 0
and arbitrarily prescribed numbers a1 , a2 , . . . , an , there exists a polynomial
P (x) with initial coefficients a0 , a1 , . . . , an such that |f (x) − P (x)| < ε for
all x ∈ [−r, r].
Proof. Without loss of generality, we may take r = 1. For convenience, set
a0 + a1 x + a2 x2 + · · · + an xn = g(x) .
Then g(0) = a0 = f (0), and by continuity we see that |f (x) − g(x)| < ε/3
throughout an interval [−δ, δ] for some δ > 0. Define the function
f (x) − g(x)
ϕ(x) = for δ ≤ |x| ≤ 1 .
xn+1
Set ϕ(0) = 0 and define ϕ to be linear in each of the intervals [0, δ] and
[−δ, 0]. This function ϕ is continuous on [−1, 1], so by the Weierstrass
approximation theorem there is a polynomial Q for which
ε
|ϕ(x) − Q(x)| < for all x ∈ [−1, 1] .
3
Now define the polynomial P (x) = g(x) + xn+1 Q(x) and conclude that
ε ε
|f (x) − P (x)| < |x|n+1 ≤ for δ ≤ |x| ≤ 1 .
3 3
For 0 ≤ x ≤ δ we have |ϕ(x) − Q(x)| < ε/3 and |ϕ(x)| ≤ |ϕ(δ)|, so that
|Q(x)| ≤ ε/3 + |ϕ(δ)|. Therefore,
|f (x) − P (x)| < |f (x) − g(x)| + |x|n+1 |Q(x)|
ε ε ε
< + δ n+1 |Q(x)| ≤ + + δ n+1 |ϕ(δ)|
3 3 3
2ε 2ε ε
= + |f (δ) − g(δ)| < + =ε
3 3 3
for 0 ≤ x ≤ δ. A similar estimate applies to the interval where −δ ≤
x ≤ 0. Together with the earlier estimate for δ ≤ |x| ≤ 1, this shows that
|f (x) − P (x)| < ε for all x ∈ [−1, 1]. By construction, the polynomial P (x)
has initial coefficients a0 , a1 , . . . , an , so the proof is complete.
There is an interesting application to power-series “expansion” of an

arbitrary continuous function. As Pál [8] pointed out, Theorem 8 has the
following immediate corollary.
6.7. Refinements of Weierstrass theorem 169
Corollary 1. Corresponding to any ∞continuous function f on an inter-

k
val [−r, r] there is a power series k=0 ak x whose partial sums sn (x) =
n
k=0 ak x have a subsequence {snj } that converges uniformly to f in [−r, r].
k
Of course, the power series need not converge at any point x0 = 0, since
that would make its sum f infinitely differentiable in the interval (−x0 , x0 ),
whereas f is not required to have a derivative at any point.
One particular case of Theorem 8 is worthy of note. Taking a0 = a1 =
· · · = an = 0, we infer that each continuous function f on an interval [−r, r]
with f (0) = 0 can be approximated uniformly by polynomials of the form
P (x) = an+1 xn+1 + an+2 xn+2 + · · · + an+k xn+k ,
where n is any prescribed positive integer. This special case leads to a

striking refinement of the Weierstrass approximation theorem, also due to
Pál [9].
Corollary 2. If 0 < r < 1, then each function f continuous in the in-

terval [−r, r] with f (0) = 0 can be approximated uniformly in [−r, r] by a
polynomial all of whose coefficients are integers.
Proof. Given ε > 0, choose n so large that
rn+1 ε
rn+1 + rn+2 + · · · = < .
1−r 2
Next invoke the special case of Theorem 8 to produce a polynomial of the

form
Q(x) = an+1 xn+1 + an+2 xn+2 + · · · + an+k xn+k
such that |f (x) − Q(x)| < ε/2 for all x ∈ [−r, r]. Finally, define the polyno-
mial
P (x) = cn+1 xn+1 + cn+2 xn+2 + · · · + cn+k xn+k ,
where the coefficients cn+j are integers with |cn+j − an+j | < 1. Then
|f (x) − P (x)| ≤ |f (x) − Q(x)| + |Q(x) − P (x)|

ε ε ε
< + rn+1 + rn+2 + · · · + rn+k < + = ε
2 2 2
for all x ∈ [−r, r], as desired.
Note that Corollary 2 fails unless r < 1. Consider for instance the
function f (x) = x/2 on the interval [−1, 1].
Finally, Corollary 1 can be sharpened to produce a single power series

with the property that some subsequence of its partial sums converges uni-
formly to any given continuous function that vanishes at the origin. This
astounding result, the existence of a universal power series that generates
every continuous function, is the prototype for what has become a large body
of similar results involving infinite series of functions. It was first presented
by Pál [9] but was attributed to his Hungarian colleague Mihály Fekete.
Corollary
∞ 3 (Fekete’s theorem). There exists a single power series
k whose partial sums s (x) = n k
a
k=1 k x n k=1 ak x approximate every func-
tion f continuous on an arbitrary interval [−r, r], provided only that f (0) =
0. More precisely, to each such function f there corresponds an increas-
ing sequence {n1 , n2 , . . . } of positive integers such that snj (x) → f (x) as
j → ∞, uniformly in −r ≤ x ≤ r.
Proof. It follows directly from the Weierstrass theorem that each func-
tion f continuous on an interval [a, b] can be approximated uniformly with
any prescribed accuracy by polynomials with rational coefficients (cf. Ex-
ercise 19). The set of polynomials with rational coefficients is countable.
Let {Q1 (x), Q2 (x), . . . } be an enumeration of all polynomials with rational
coefficients that vanish at the origin: Q1 (0) = Q2 (0) = · · · = 0. Choose
a polynomial P1 (x) with P1 (0) = 0 such that |P1 (x) − Q1 (x)| < 1 for all
x ∈ [−1, 1]. Appealing to Pál’s theorem (Theorem 8), next choose a poly-
nomial P2 (x) that begins with P1 (x) and has the approximation property
1
|P2 (x) − Q2 (x)| < for all x ∈ [−2, 2] .
2
Now proceed inductively. Having chosen P1 , P2 , . . . , Pn−1 , choose a polyno-
mial Pn (x) that begins with Pn−1 (x) and satisfies
1
|Pn (x) − Qn (x)| < for all x ∈ [−n, n] .
n
Then P1 (x), P2 (x), . . . are partial sums of a power series with the desired
property. To see this, suppose f (x) is continuous on an interval [−r, r]
and f (0) = 0. Choose an increasing sequence {nk } of indices such that
Qnk (x) → f (x) uniformly on [−r, r] as k → ∞. Then by construction,
[Pnk (x) − Qnk (x)] → 0 and so Pnk (x) → f (x) uniformly on [−r, r].
Needless to say, a universal power series, as described in Corollary 3,

cannot converge at any point x = 0.
Pál’s theorem strengthens the Weierstrass theorem by showing that fi-
nite linear combinations of the monomials 1, xn , xn+1 , . . . are sufficient for
Exercises 171
uniform approximation of any continuous function on a closed bounded in-

terval. His result suggests the problem of characterizing the sets of mono-
mials 1, xn1 , xn2 , . . . whose finite linear combinations are uniformly dense in
the space of all continuous functions. Sergei Bernstein studied the problem
in 1912 and found a partial solution. The full solution, at least on the in-
terval [0, 1], is quite elegant and was found around 1915 by Chaim Müntz
[7] and Otto Szász [13].
Müntz–Szász Theorem. Let {n1 , n2 , . . . } be an increasing sequence of
positive integers. In order that every function continuous on the interval
[0, 1] be uniformly approximable by polynomials of the form
P (x) = c0 + c1 xn1 + · · · + ck xnk ,

it is necessary and sufficient that the series ∞ k=1 1/nk diverge.
In fact, the exponents nk need not be integers. The theorem remains

valid for any increasing sequence of positive exponents that tend to infinity.
A similar theorem holds for mean-square approximation of functions that are
square-integrable over the interval [0, 1]. There the natural method of proof
(introduced by Szász) involves Gram determinants and a classical formula
of Cauchy to evaluate the particular determinants that arise. The uniform
version of the theorem (as stated here) then follows from the mean-square
version by simple estimates. The details will not be pursued here, but may
be found in books on approximation theory, for instance in Cheney [2] or
Davis [3].
It may be remarked that the Müntz–Szász theorem extends readily to
any interval [0, r], but not to an interval [−r, r] that contains the origin in its
interior. For example, the exponents nk = 2k satisfy the divergence criterion
but linear combinations of the monomials x2k will approximate only the even
functions.
Exercises
1. Use induction to calculate the Vandermonde determinant

1 x0 x20 . . . xn0

1 x1 x21 . . . xn1
. . = (xk − xj ) .
. . .
.. ..
.
. .
j<k
1 xn x2n . . . xnn
Suggestion. Moving from right to left, subtract xn times the kth column
from the (k + 1)st column, then remove common factors and reduce the
problem to the calculation of a Vandermonde determinant with n replaced
by n − 1.
2. (a) For m, n = 1, 2, . . . , derive the trigonometric identity

cos(n + m)θ + cos(n − m)θ = 2 cos mθ cos nθ .
(b) Show that

π
0 if m = n
cos mθ cos nθ dθ =
−π π if m = n > 0 .
(c) Show that the Chebyshev polynomials have the orthogonality prop-
erty 1
dx 0 if m = n
Tm (x) Tn (x) √ =
−1 1−x 2 π/2 if m = n .
3. (a) Show that the Chebyshev polynomials satisfy the differential equa-
tion
(1 − x2 )Tn (x) − xTn (x) + n2 Tn (x) = 0 .
(b) Set W (x) = Tn (x)Tm (x) − Tm
(x)T (x) and show that
n
(1 − x2 )W (x) − xW (x) + (n2 − m2 )Tn (x)Tm (x) = 0 .
(c) Conclude that

1
dx
Tn (x) Tm (x) √ =0 if n = m .
−1 1 − x2
4. Verify the explicit formula
1 n n
Tn (x) = x + i 1 − x2 + x − i 1 − x2
2
for −1 ≤ x ≤ 1 and n = 0, 1, 2, . . . .
5. (a) Show that sin(n + 1)θ/ sin θ is a polynomial of degree n in cos θ.
Denoted by Un (x), it is called a Chebyshev polynomial of second kind. Thus
sin(n + 1)θ
= Un (cos θ) .
sin θ
Suggestion. Base an inductive argument on the identity
sin(n + 1)θ = sin nθ cos θ + cos nθ sin θ .
(b) Calculate U0 (x) = 1 , U1 (x) = 2x , U2 (x) = 4x2 − 1 . Verify the

identity
sin(n + 1)θ + sin(n − 1)θ = 2 cos θ sin nθ
and derive the recurrence relation
Un+1 (x) = 2xUn (x) − Un−1 (x) , n = 1, 2, . . . ,
which is the same as for the Chebyshev polynomials Tn (x).
Exercises 173
(c) Verify the trigonometric identity
cos(n − m)θ − cos(n + m)θ = 2 sin mθ sin nθ
and use it to show that

π
0 if m = n
sin mθ sin nθ dθ =
−π π if m = n > 0.
(d) Derive the orthogonality relation

1 0 if m = n
Um (x)Un (x) 1 − x dx =
2
−1 π/2 if m = n > 0.
6. (a) Show that the Chebyshev polynomials of second kind satisfy the
differential equation
(1 − x2 )Un (x) − 3xUn (x) + n(n + 2)Un (x) = 0 .
(b) Use the differential equation to derive the orthogonality relation

1
Um (x)Un (x) 1 − x2 dx = 0 , m = n .
−1
7. (a) Show that Un (1) = n + 1, n = 0, 1, 2, . . . .

(b) Show that Tn (x) = nUn−1 (x), n = 1, 2, . . . .
(c) For any polynomial P of degree ≤ n, Markov’s inequality asserts
that
max |P (x)| ≤ n2 max |P (x)| .
−1≤x≤1 −1≤x≤1
Give an example to show that the constant n2 cannot be replaced by any

smaller constant.
8. Show that the polynomials Tn (x) and Un (x) are even functions when n
is even and odd functions when n is odd.
9. Apply the Weierstrass approximation theorem to show that if a function

f has a continuous derivative of order n on an interval [a, b], then for each
ε > 0 there is a polynomial P such that
|f (x) − P (x)| < ε , |f (x) − P (x)| < ε , . . . , |f (n) (x) − P (n) (x)| < ε
for all x ∈ [a, b].

10. Suppose f is continuous on [a, b] and has the property

b
xn f (x) dx = 0 , n = 0, 1, 2, . . . .
a
Prove that f (x) ≡ 0 on [a, b].

b
Hint. Show by the Weierstrass approximation theorem that a f (x)2 dx = 0 .
11. Let f (x) be continuous on the interval [0, 1]. Show that
1
√ f (x) π
lim n 2
dx = f (0) .
n→∞ 0 1 + nx 2
Suggestion: First let f (x) = xk , k = 0, 1, 2, . . . .

12. If f is continuous in the interval [0, 1], show that its Bernstein polyno-
mials satisfy
max |Bn (x)| ≤ max |f (x)| , n = 1, 2, . . . .

0≤x≤1 0≤x≤1
13. For m = 0, 1, 2, . . . let

n m
n k
Bn (x; m) = xk (1 − x)n−k
k n
k=0
denote the nth Bernstein polynomial of the function f (x) = xm .

(a) Show that
1
Bn (x; 0) = 1 , Bn (x; 1) = x , Bn (x; 2) = x2 + x(1 − x) .
n
In particular, B1 (x; 2) = x.
(b) Show that
1
Bn (x; 3) = 2
(n − 1)(n − 2)x3 + 3(n − 1)x2 + x .
n
In particular, B1 (x; 3) = x and B2 (x; 3) = 14 (3x2 + x).
Hint. Write k 3 = k(k − 1)(k − 2) + 3k 2 − 2k.
(c) Show by induction that Bn (x; m) has degree n for n ≤ m and degree
m for n > m.
(d) Generalize (c) by showing that if f is any polynomial of degree m,
then its Bernstein polynomial Bn (x) has degree n for n ≤ m and degree
≤ m for n > m.
Exercises 175
14. Let Bn (x; f ) be the nth Bernstein polynomial of a function f defined

on [0, 1].
(a) Perform a calculation to derive the formula

n−1
n−1

k+1

k
Bn (x; f ) =n f −f xk (1 − x)n−1−k ,
k n n
k=0
so that Bn (x; f ) = Bn−1 (x; gn ), where gn is the function defined by

n−1 1 n−1
gn (x) = n f x+ −f x .
n n n
(b) If f has a continuous derivative in [0, 1], show that gn (x) converges
uniformly to f (x) as n → ∞.
(c) Deduce that |Bn−1 (x; gn ) − Bn−1 (x; f )| → 0 uniformly in [0, 1].
Hint. Apply the result of Exercise 12.
(d) Conclude that if f is continuously differentiable, then Bn (x; f ) →
f (x) uniformly in [0, 1].
*(e) By a similar method, prove more generally that if f has a continuous
(ν)
derivative of order ν, then Bn (x; f ) → f (ν) (x) as n → ∞, uniformly in
[0, 1].
Reference: Lorentz [6].
15. Let I n denote the closed unit cube in Rn , consisting of all points
(x1 , x2 , . . . , xn ) with 0 ≤ xj ≤ 1 for j = 1, 2, . . . , n. If a function
f (x1 , x2 , . . . , xn ) is continuous on I n , show that it can be uniformly ap-
proximated on I n by polynomials; that is, by finite linear combinations of
the monomials xk11 xk22 · · · xknn .
16. Apply the Stone–Weierstrass theorem to prove the trigonometric form

of the Weierstrass theorem. Specifically, prove that if f (x) is continuous
on R and f (x + 2π) = f (x) for all x ∈ R, then for each ε > 0 there is a
trigonometric polynomial

n
P (x) = a0 + (ak cos kx + bk sin kx)
k=1
such that |f (x) − P (x)| < ε for all x ∈ R.

17. Let f be continuous on the interval [0, 2π] with f (0) = f (2π), and form
the Poisson integral
2π
1 1 − r2
u(r, θ) = f (t) dt , 0 ≤ r < 1.
2π 0 1 − 2r cos(θ − t) + r2
Show that u(r, θ) → f (θ) as r → 1, uniformly for θ ∈ [0, 2π]. Give two
proofs, one patterned after Landau’s proof of Weierstrass’s theorem, the
other invoking the trigonometric form of Weierstrass’s theorem.
18. Carry out the details in the deduction of Corollary 1 from Pál’s theorem
(Theorem 8).
19. For each function f continuous on a closed bounded interval [a, b]

and for each ε > 0, show that there exists a polynomial Q with rational
coefficients such that |Q(x) − f (x)| < ε for all x ∈ [a, b].
20. For each function f continuous on a closed bounded interval [a, b]

and for each ε > 0, show that there exist polynomials P and Q such that
P (x) < f (x) < Q(x) and Q(x) − P (x) < ε for all x ∈ [a, b].
21. Show directly, without appeal to the Müntz–Szász theorem, that for
each integer m > 0, every continuous function can be approximated uni-
formly on the interval [0, 1] by polynomials of the form
P (x) = c0 + c1 xn1 + · · · + ck xnk , where nk = km for k = 1, 2, . . . .
Hint. If f (x) is continuous on [0, 1], so is g(t) = f (t1/m ).
References
[1] S. N. Bernstein, “Démonstration du théorème de Weierstrass fondée sur le calcul
des probabilités”, Comm. Soc. Math. Kharkow 13 (1912), 1–2.
[2] E. W. Cheney, Introduction to Approximation Theory, McGraw–Hill, New York,
1966.
[3] Philip J. Davis, Interpolation and Approximation, Blaisdell, New York, 1963.
[4] Edmund Landau, “Über die Approximation einer stetigen Funktion durch eine
ganze rationale Funktion”, Rend. Circ. Mat. Palermo 25 (1908), 337–345.
[5] Henri Lebesgue, “Sur l’approximation des fonctions”, Bull. Sci. Math. 22 (1898),
278–287.
[6] G. G. Lorentz, Bernstein Polynomials, University of Toronto Press, Toronto,
1953.
References 177
[7] Chaim Müntz, “Über den Approximationssatz von Weierstrass”, in H.A. Schwarz
Festschrift, Mathematische Abhandlungen, Berlin, 1914, pp. 303–312.
[8] Julius Pál, “Über eine Anwendung des Weierstrass-schen Satzes von der Annä-
herung stetiger Funktionen durch Polynome”, Tôhoku Math. J. 5 (1914), 8–9.
[9] Julius Pál, “Zwei kleine Bemerkungen”, Tôhoku Math. J. 6 (1914), 42–43.
[10] A. Pincus, “Weierstrass and approximation theory”, J. Approx. Theory 107
(2000), 1–66.
[11] Walter Rudin, Principles of Mathematical Analysis, Third edition,
McGraw–Hill, New York, 1976.
[12] M. H. Stone, “The generalized Weierstrass approximation theorem”, Math.
Magazine 21 (1948), 167–184; 237–254.
[13] Otto Szász, “Über die Approximation stetiger Funktionen durch lineare
Aggregate von Potenzen”, Math. Annalen 77 (1916), 482–496.
[14] Charles de la Vallée-Poussin, Leçons sur l’approximation des fonctions d’une
variable réelle, Gauthier–Villars, Paris, 1919; reprinted 1952.
[15] Karl Weierstrass, “Über die analytische Darstellbarkeit sogenannter willkür-
licher Functionen einer reellen Veränderlichen”, Sitzungsberichte der Königlich
Preussischen Akademie der Wissenschaften zu Berlin, 1885, pp. 633–639, 789–805.
Chapter 7
Tauberian Theorems
It makes no sense to speak of the sum of a divergent infinite series. Neverthe-

less, a series that is not “too badly divergent” can be assigned a generalized
sum in a variety of natural ways. We have already encountered this notion
of summability in connection with Abel’s theorem, which asserts that every
convergent series is Abel summable to its ordinary sum. More generally, an
Abelian theorem is any statement to the effect that a method of summabil-
ity assigns to each convergent series its ordinary sum. A Tauberian theorem
goes in the opposite direction and asserts that every summable series which
is not too badly divergent is actually convergent. In this chapter we develop
some elegant and important Tauberian theorems that will find application
to other topics later in the book.
7.1. Summation of divergent series

∞
According to Abel’s theorem, if an infinite series n=0 an converges and
has sum s, then the power series
∞
f (x) = an x n
n=0
converges for |x| < 1 and f (x) → s as x → 1−. The theorem was originally
viewed as a device for evaluating the sums of convergent series such as
1 1 1
1 − + − + · · · = log 2 and
2 3 4
1 1 1 π
1 − + − + ··· = .
3 5 7 4
179
180 7. Tauberian Theorems
However, the possible existence of the Abel limit when a series diverges
suggests
a natural way to assign it a generalized sum. For example, the
series ∞ n=0 (−1) n is divergent, but
∞
1 1
f (x) = (−1)n xn = → as x → 1− ,
1+x 2
n=0
so we may say that the series is Abel summable to the sum 12 . Observe
that
the extended notion of sum retains its linearity. In other words,
if an is
Abel summable to A and bn is Abel summable to B, then (an + bn ) is
summable to A + B and can is summable to cA for any constant c. Also,
Abel’s theorem guarantees that the Abel sum of a convergent series exists
and is equal to the ordinary sum.
A similar technique for summation of divergent series can be based on
the averages of partial sums. Let
sn = a0 + a1 + · · · + an
denote the partial sums of the series, and let
s0 + s1 + · · · + sn
σn =
n+1
denote their arithmetic means, also known as the Cesàro means. The series

an is said to be Cesàro summable to the sum σ if if σn → σ as n → ∞.
Recall (cf. Chapter 1, Exercise 13) that σn → s whenever sn → s, but
the sequence {σn } may converge when {sn } does not. For instance, the
divergent series ∞ n
n=0 (−1) has partial sums sn = 1 when n is even and
sn = 0 when n is odd, so it is Cesàro summable to the sum 12 , which is the

same as its Abel sum. The series ∞ n=1 (−1)
n+1 n is not Cesàro summable,
but is Abel summable to 14 . (See Exercise 1.)

With this meager evidence in hand, one may suspect that Abel summa-
tion is more powerful than Cesàro summation, which is indeed the case. In
1880, Georg Frobenius [2] strengthened Abel’s theorem by showing that a
series is Abel summable if it is Cesàro summable. Here is a more precise
statement, in notation introduced above.
Frobenius’
∞ Theorem. If σn → σ as n → ∞, then the power series
f (x) = n=0 an x converges for |x| < 1 and f (x) → σ as x → 1−.
n
Proof. Two summations by parts give

∞
∞
∞

n
an x = a0 + (sn − sn−1 )x = (1 − x)
n
sn xn
n=0 n=1 n=0
∞
∞

= (1 − x) 2
(s0 + s1 + · · · + sn )x = (1 − x)
n 2
(n + 1)σn xn .
n=0 n=0
7.1. Summation of divergent series 181
In view of the identity

∞
1
(n + 1)xn = ,
(1 − x)2
n=0
this may be written as

∞
∞

an x − σ = (1 − x)
n 2
(n + 1)(σn − σ)xn .
n=0 n=0
But σn → σ, so for each ε > 0 there is an integer N such that
|σn − σ| < ε for all n > N .
Then for 0 < x < 1,

∞ N ∞

n
(n + 1)(σn − σ)x ≤ (n + 1)|σn − σ|x +
n
(n + 1)|σn − σ|xn

n=0 n=0 n=N +1

N
ε
≤ (n + 1)|σn − σ| + ,
(1 − x)2
n=0
and so

∞
N

an xn − σ ≤ (1 − x)2 (n + 1)|σn − σ| + ε < 2ε

n=0 n=0
when x is sufficiently close to 1. In other words, f (x) → σ as x → 1−.
Many other summation procedures have been introduced, and the cor-
responding “Abelian theorems” proved, asserting that whenever a series is
convergent or summable by some method, it must be summable to the same
sum by another more powerful method. For instance, a series is said to be
Borel summable if the limit
∞

−x 1
lim e sn xn
x→∞ n!
n=0
exists. The method is named for Émile Borel, who introduced it in 1899 and
pointed out the corresponding Abelian theorem, that a convergent series is
Borel summable to its ordinary sum (cf. Exercise 2). Borel summability
arises naturally in complex function theory, especially in problems of analytic
continuation.
To give one more example, a series is said to be Lambert summable if

the limit
∞
lim an L(xn )
x→1−
n=0
exists, where
x log(1/x)
L(x) = for 0 < x < 1, and L(1) = 1 ,
1−x
is known as the Lambert kernel. Note that L(x) → 1 as x → 1−. Lambert

summability has important applications to number theory and is involved
in proofs of the prime number theorem.
7.2. Tauber’s theorem

An infinite series may be summable by various methods and yet fail to
converge. We have seen elementary examples of divergent series that are
Cesàro or Abel summable. In the converse direction, however, there is a
general principle that if a series is summable by some method and is not
“too badly divergent”, then it is actually convergent. The first result of
this type was found in 1897 by Alfred Tauber (1866–1942), an Austrian
mathematician who later specialized in actuarial mathematics. Tauber [12]
showed that the converse of Abel’s theorem is valid under the additional
hypothesis that nan → 0.
Tauber’s Theorem. Suppose nan → 0 as n → ∞, so that the power series

∞

f (x) = an xn
n=0
in the interval (−1, 1). If f (x) → A as x → 1−, then the infinite

converges
series ∞n=0 an converges to the sum A.
Briefly, Tauber’s theorem asserts that if a series is Abel summable and its
terms satisfy the additional condition nan → 0, then the series is convergent.
The theorem uncovers a remarkable phenomenon. Not only does the Abel
method fail to sum series that diverge too rapidly, but it also fails to sum
series whose divergence is too slow. For example, if |an | ≤ 1/(n log n) ,
Tauber’s theorem tells us that the series an is not Abel summable unless
it is convergent. The same is true if |an | ≤ 1/n , but the proof is based on a
stronger form of Tauber’s theorem requiring only that the sequence {nan }
be bounded. More about this later.
7.3. Theorems of Hardy and Littlewood 183
n
Proof of Tauber’s theorem. With the notation sn = k=0 ak for the
partial sums, we can write

n ∞

sn − f (x) = ak (1 − x ) −
k
ak x k .
k=1 k=n+1
Taking 0 < x < 1, we apply the inequality
1 − xk = (1 − x)(1 + x + · · · + xk−1 ) ≤ k(1 − x)
to conclude that

n ∞

|sn − f (x)| ≤ (1 − x) k|ak | + |ak |xk .
k=1 k=n+1
Now let ε > 0 be given. By hypothesis, we may choose n large enough that
k|ak | < ε for all k > n. Then
∞
∞
∞
1 k ε k ε
|ak |xk < ε x < x = .
k n n(1 − x)
k=n+1 k=n+1 0
Now choose x = xn = 1 − n1 , so that 1 − xn = 1

n and xn → 1 as n → ∞.
This gives the estimate
1
n
|sn − f (xn )| ≤ k|ak | + ε
n
k=1
for n sufficiently large. But since k|ak | → 0, it follows that the arithmetic
means
1
n
k|ak |
n
k=1
also tend to 0 as n → ∞. Therefore, |sn − f (xn )| < 2ε for all n sufficiently

large. In view of the hypothesis that f (x) → A as x → 1−, this proves that
sn → A.
7.3. Theorems of Hardy and Littlewood

Beginning around 1910, the two British analysts G. H. Hardy (1877–
1947) and J. E. Littlewood (1885–1977) began a series of investigations
initially inspired by Tauber’s theorem. First Hardy [4] proved an analogue
of Tauber’s theorem by postulating the more restrictive Cesàro summability
instead of Abel summability but requiring only that the sequence {nan } be
bounded. His result can be stated as follows.

Hardy’s Theorem. If the infinite series ∞ n=0 an is Cesàro summable
and {nan } is bounded, then the series is convergent.
Hardy conjectured that Tauber’s theorem remains valid if the condition

nan → 0 is weakened to the boundedness of {nan }. His Cambridge col-
league Littlewood [11] then verified the conjecture by proving the following
theorem, which strengthens the theorems of both Tauber and Hardy.
Littlewood’s Theorem. Suppose the sequence {nan } is bounded, so that

the power series
∞

f (x) = an xn
n=0
in the interval (−1, 1). If f (x) → A as x → 1−, then the infinite

converges
series ∞n=0 an converges to the sum A.
This work marked the beginning of the famous Hardy–Littlewood col-

laboration, which extended over a period of 35 years and was one of the
most fruitful collaborations in the history of mathematics. Together these
two mathematicians made seminal contributions to a variety of subjects in-
cluding number theory, Fourier series, and functions of a complex variable.
They also developed a series of results analogous to Tauber’s theorem, which
they called “Tauberian theorems”. A Tauberian theorem is the converse of
an Abelian theorem under a supplementary growth condition called a Taube-
rian condition. For instance, Littlewood’s Tauberian theorem asserts that
the converse of Abel’s theorem is true under the Tauberian condition that
{nan } is bounded.
In 1914, Hardy and Littlewood together found a Tauberian theorem
that goes from Abel to Cesàro summability and therefore provides a partial
converse of Frobenius’ theorem.

Hardy–Littlewood Theorem. If f (x) = ∞ n=0 an x → A as x → 1−
n
and sn ≥ 0 for all n, then σn → A as n → ∞.
It is easy to see that the Hardy–Littlewood theorem remains valid more

generally if sn ≥ −C for some constant C, which is certainly true if the
partial sums sn are bounded. Hardy and Littlewood also found that the
Tauberian condition in Littlewood’s theorem, the boundedness of {nan },
can be relaxed to a one-sided condition an ≥ −C/n.
The first proofs of Littlewood’s theorem and of the Hardy–Littlewood
theorem were similar and were much more difficult than those of Tauber’s
or Hardy’s theorem. For a long time no simplifications were found and
these results were considered to be quite deep. Then in 1930, Karamata [8]
found a clever and surprisingly simple approach based on the Weierstrass
approximation theorem. Karamata’s elegant proof will be discussed in the

next section. His strategy was to prove the Hardy–Littlewood theorem,
then to deduce Littlewood’s theorem from it. Littlewood’s original approach
was to prove Cesàro summability and then appeal to Hardy’s theorem to
conclude that the series converges. Some years later, Wielandt [14] found
a simple refinement of Karamata’s method that avoids the detour through
Cesàro summability and proves Littlewood’s theorem directly. The details
are described in the next section.
7.4. Karamata’s proof

We now turn to Karamata’s proof of the Hardy–Littlewood and Little-
wood Tauberian theorems. Our treatment is adapted from the books of
Titchmarsh [13] and Korevaar [10]. Karamata’s method is based on the
Weierstrass approximation theorem, which can be invoked to establish the
following lemma.
Lemma. Let g(x) be continuous on the interval [0, 1] except for a possible
jump-discontinuity at a point c ∈ (0, 1). Then for each ε > 0 there exist
polynomials P (x) and Q(x) such that P (x) < g(x) < Q(x) and
1 1

g(x) − P (x) dx < ε Q(x) − g(x) dx < ε .
0 0
Proof. Suppose that g(x) has a jump-discontinuity at c ∈ (0, 1), and let
g(c+) and g(c−) denote the right- and left-hand limits. Suppose without
loss of generality that g(c−) ≤ g(c+). For fixed δ ∈ (0, c), let (x) be the
linear function such that
(c − δ) = g(c − δ) + ε/2 , (c) = g(c+) + ε/2 .
Define

g(x) + ε/2 if 0 ≤ x < c − δ or c < x ≤ 1
φ(x) =
max{(x), g(x) + ε/2} if c − δ ≤ x ≤ c .
Then φ(x) is continuous in [0, 1] and φ(x) ≥ g(x)+ε/2. Choose a polynomial
Q(x) such that |Q(x) − φ(x)| < ε/2 for all x ∈ [0, 1]. Then Q(x) > g(x) and
1

Q(x) − g(x) dx < ε
0
if δ is sufficiently small. A similar construction produces a polynomial
1
P (x) < g(x) with 0 [g(x) − P (x)] dx < ε.
With the lemma in hand, we first prove the Hardy–Littlewood theorem,

which is a somewhat easier than Littlewood’s theorem.
Proof of Hardy–Littlewood theorem. By hypothesis,

∞
∞

f (x) = an x = (1 − x)
n
sn xn → A as x → 1 − .
n=0 n=0
We claim that this implies, more generally, that

∞
1
(1) lim (1 − x) sn xn p(xn ) = A p(t) dt
x→1− 0
n=0
for every polynomial p(x). To see this, it suffices to consider monomials

p(x) = xk . Then the left-hand side is
∞

∞

1−x
(1 − x) sn x n+kn
= (1 − x )
k+1 k+1 n
sn (x )
1 − xk+1
n=0 n=0
1
A
→ =A p(t) dt as x → 1− ,
k+1 0
as claimed. Invoking the lemma, we can now infer from (1) that
∞
1
(2) lim (1 − x) sn xn g(xn ) = A g(t) dt
x→1− 0
n=0
if g is continuous in [0, 1] except for a possible jump-discontinuity. It is here

that the hypothesis sn ≥ 0 is needed. To carry out the details, let ε > 0 be
given and choose polynomials P and Q as in the lemma. Then
∞
∞
∞

sn xn P (xn ) ≤ sn xn g(xn ) ≤ sn xn Q(xn ) , 0 < x < 1.
n=0 n=0 n=0
Since sn ≥ 0 implies A ≥ 0, it follows that

∞
∞

lim sup (1 − x) sn xn g(xn ) ≤ lim (1 − x) sn xn Q(xn )
x→1− x→1−
n=0 n=0
1 1
=A Q(t) dt ≤ A g(t) dt + Aε .
0 0
Letting ε → 0, we conclude that

∞
1
lim sup (1 − x) sn x g(x ) ≤ A
n n
g(t) dt .
x→1− n=0 0
In a similar way we find that

∞
1
lim inf (1 − x) sn xn g(xn ) ≥ A g(t) dt .
x→1− 0
n=0
Combining the last two inequalities, we arrive at a proof of (2).

Now choose
0, 0 ≤ t < 1/e
g(t) =
1/t , 1/e ≤ t ≤ 1 .
Then
1 1
g(t) dt = 1/t dt = 1 .
0 1/e
Let xN = e−1/N and observe that xnN ≥ 1/e if and only if n ≤ N , so that
∞

N
sn xnN g(xnN ) = sn = (N + 1)σN .
n=0 n=0
The relation (2) now shows that

∞
1
(N + 1)(1 − xN )σN = (1 − xN ) sn xnN g(xnN ) →A g(t) dt = A
n=0 0
as N → ∞, since xN → 1. But
(N + 1)(1 − xN ) = (N + 1)(1 − e−1/N ) → 1 as N → ∞ ,
so this proves that σN → A.
Proof of Littlewood’s theorem. For each polynomial

m
P (x) = bk x k with P (0) = 0 ,
k=1
the hypothesis that f (x) → A implies

∞

m ∞

m
(3) n
an P (x ) = bk an x kn
→A bk = P (1)A
n=0 k=1 n=0 k=1
as x → 1−. Now choose the “cutoff function”

0 , 0 ≤ t < 1/e
g(t) =
1 , 1/e ≤ t ≤ 1 ,
so that

N ∞

sN = an = an g(xnN ) , where xN = e−1/N .
n=0 n=0
To prove that sN → A as N → ∞, it will suffice to show that

∞

an g(xn ) → g(1)A = A as x → 1 − .
n=0
In order to show that

∞

(4) lim sup an g(xn ) ≤ A ,
x→1− n=0
we will apply the relation (3) to a polynomial P with the properties P (0) =
0, P (1) = 1, and P (t) ≥ g(t) for 0 ≤ t ≤ 1. For this purpose we choose a
polynomial Q with
g(t) − t
Q(t) ≥ = h(t) , 0 < t < 1,
t(1 − t)
which gives a polynomial P (t) = t+t(1−t)Q(t) with the required properties.

The Tauberian condition |nan | ≤ C implies that nan ≥ −C, so that
∞
∞
∞

an g(xn ) − an P (xn ) = − an P (xn ) − g(xn )
n=0 n=0 n=1
∞
∞

1 1−x
≤C P (xn ) − g(xn ) ≤ C P (xn ) − g(xn )
n 1−x n
n=1 n=1
∞
P (t) − g(t)
= C(1 − x) φ(xn ) , where φ(t) = .
1−t
n=1
Here we have used the elementary inequality
1 − xn
= 1 + x + · · · + xn−1 ≤ n for 0 ≤ x < 1.
1−x
We claim now that

∞
1
φ(t)
lim (1 − x) n
φ(x ) = dt ,
x→1− 0 t
n=1
since φ(t) has a continuous extension to the interval [0, 1] except for the
jump-discontinuity at t = 1/e. Indeed, for any fixed x ∈ (0, 1), the integral
is approximated by its Riemann sum
∞
∞

φ(xn )
(x − x
n n+1
) = (1 − x) φ(xn ) ,
xn
n=1 n=1
which tends to the integral as x → 1−. But

1 1 1
φ(t) P (t) − t − g(t) − t
dt = dt = Q(t) − h(t) dt .
0 t 0 t(1 − t) 0
Since Q(t) ≥ h(t), the last integral is positive, and in view of the lemma
it can be made arbitrarily small by suitable choice of the polynomial Q.
Putting everything together and recalling from (3) that
∞

lim an P (xn ) = A ,
x→1−
n=0
we conclude that (4) holds.

A similar argument shows that
∞

(5) lim inf an g(xn ) ≥ A .
x→1−
n=0
For this we choose P (t) = t + t(1 − t)Q(t), where Q is a polynomial with

g(t) − t
Q(t) ≤ h(t) = for 0 < t < 1.
t(1 − t)
Then P (0) = 0, P (1) = 1, and P (t) ≤ g(t) for 0 ≤ t ≤ 1. Hence the
Tauberian condition nan ≥ −C gives
∞
∞
∞
∞

1 n
n
an g(x )− an P (x ) ≥ −C
n
g(x )−P (x ) ≥ C(1−x)
n
φ(xn ) ,
n
n=0 n=0 n=1 n=1
since 1 − xn ≤ n(1 − x) and now

P (t) − g(t)
φ(t) = ≤0 for 0 < t < 1 .
1−t
Consequently, we may conclude as before that
∞
1

lim inf an g(x ) ≥ A + C
n
Q(t) − h(t) dt .
x→1− 0
n=0
Another appeal to the lemma shows that the last integral can be made
arbitrarily close to zero with suitable choice of the polynomial Q, which
gives (5). Combining (4) and (5), we see that
∞

lim an g(xn ) = A .
x→1−
n=0
As previously noted, this implies that sn → A, and so the theorem is proved.

It should be observed that the proof used only the weaker Tauberian condi-
tion nan ≥ −C.
Jovan Karamata (1902–1967) was a Yugoslavian, born in Zagreb (Croa-

tia) and trained in mathematics at the University of Novi Sad (Serbia) and at
Belgrade University, where he received a doctoral degree in 1926. Four years
later his two-page paper [8] appeared, outlining his totally new, relatively
simple approach to Tauberian theorems. The paper created an immediate
sensation and brought him international recognition. Later he made other
significant contributions to Tauberian theory (cf. Korevaar [10]). He was a
professor at Belgrade University until 1951, then spent the rest of his career
in Switzerland at the University of Geneva.
7.5. Hardy’s power series

The following problem appeared in the puzzles column of Emissary, a
newsletter of the Mathematical Sciences Research Institute [1].
Problem. For positive real x less than 1, define
f (x) = x − x2 + x4 − x8 + x16 − . . . .
Does f (x) have a limit as x approaches 1 from below? If so, what is the
limit?
Before considering the problem in earnest, let us make a few simple

observations. First of all, by pairing terms one sees that
f (x) = (x − x2 ) + (x4 − x8 ) + · · · > 0
and
f (x) = x − (x2 − x4 ) − (x8 − x16 ) − · · · < x
for 0 < x < 1. In particular, f (x) is bounded in the interval [0, 1]. The
identity f (x) = x − f (x2 ) shows that if f (x) has a limit as x → 1−, the
limit must be 12 . Iteration of the identity shows that f (x4 ) < f (x), which
suggests that f (x) is an increasing function in the interval 0 < x < 1. If
7.5. Hardy’s power series 191
so, then by the monotone boundedness theorem f (x) does have a limit, and
f (x) → 12 as x → 1−. Figure 1 displays the graph of

30
k
(−1)k x2 , 0 < x < 1,
k=0
as generated by Mathematica. The figure appears to confirm that f (x)

increases to 12 as x → 1−.
However, the limit does not exist! Hardy [3] studied the series in 1907
and found by intricate analysis that as x tends to 1, the function f (x)
undergoes miniscule oscillations about the value 12 and does not converge. A
closer inspection of Figure 1 actually reveals the small oscillations. Figure 2
displays a graph of the same partial sum (up to k = 30), magnified to clarify
the oscillations near the point (1, 12 ).
In fact, the nonexistence of the limit is a direct consequence of a result of
Hardy and Littlewood [7] known as the “high-indices” theorem, a remarkable
Tauberian theorem for lacunary power series. Here is the statement.

High-Indices Theorem. If f (x) = ∞ k=1 ak x
nk for 0 < x < 1, where
the exponents nk are positive integers satisfying a condition of the form

nk+1 /nk ≥ q > 1, and if f (x) → A as x → 1−, then the infinite series
∞
k=1 ak converges to A.
Hardy and Littlewood obtained this result in 1926, relatively late in

the development of basic Tauberian theory. Their discovery appears to
have been inspired in part by the special power series (discussed above)
0.5
0.4
0.3
0.2
0.1
0.2 0.4 0.6 0.8 1
Figure 1. Hardy’s sum for 0 ≤ x ≤ 1.

that Hardy had studied in 1907. The most astounding feature of the high-
indices theorem is that lacunarity alone serves as a Tauberian condition.
For Hardy’s power series we see k
that nk = 2 is sufficiently lacunary, and
the partial sums of the series ak are alternately equal to 0 and 1, so the
high-indices theorem shows that f (x) cannot tend to a limit as x → 1−.
A proof of the high-indices theorem is beyond the scope of this book,
but can be found for instance in the book by Korevaar [10], p. 50 ff. Al-
ternatively, we can appeal to the more elementary theorem of Hardy and
Littlewood, as stated at the end of Section 7.3, to show that Hardy’s sum
does not tend to a limit as x → 1−. For the function
∞
∞
k
f (x) = an xn = (−1)k x2 ,
n=1 k=0
it is seen that all partial sums sn = a1 + · · · + an are either 0 or 1, so

the hypothesis sn ≥ 0 of the Hardy-Littlewood theorem is clearly satisfied.
Therefore, if we can show that the sequence {σn } of Cesàro means does not
converge, it will follow that f (x) does not converge as x → 1−. In fact, a
direct calculation (cf. Exercise 3) reveals that
1 2
lim inf σn = 3 and lim sup σn = 3 ,
n→∞ n→∞

so the series an is not Cesàro summable. Thus by the Hardy-Littlewood
theorem, it is not Abel summable. In other words, f (x) does not tend to a
limit as x → 1−.
A recent paper by Keating and Reade [9] offers further analysis of
Hardy’s power series and similar examples. There the authors apply the
0.52
0.51
0.95 0.96 0.97 0.98 0.99
0.49
Figure 2. Hardy’s sum for 0.95 ≤ x ≤ 1.

Exercises 193
Poisson summation formula of Fourier analysis (see Chapter 8) to prove

that as x → 1−, Hardy’s function undergoes persistent oscillations about
the value 12 with approximate amplitude 0.00275 , which is consistent with
Figure 2.
Hardy’s classic book [5] contains a wealth of information about Abelian
and Tauberian theorems. Korevaar’s more recent book [10] is a comprehen-
sive reference for Tauberian theorems and their applications.
Exercises
1. Show that the series
∞

(−1)n+1 n = 1 − 2 + 3 − 4 + . . .
n=1
is not Cesàro summable, but is Abel summable to 14 .

2. Show that every convergent series is Borel summable to its ordinary
sum. Specifically, show that if sn → s, then
∞
1
e−x sn xn → s as x → ∞ .
n!
n=0
3. Consider Hardy’s power series

∞
∞

n k
an x = (−1)k x2 .
n=1 k=0
Thus a k k
n = (−1) if n = 2 and an = 0 otherwise. Show directly that the
series an is not Cesàro summable by computing
1 2
lim inf σn = 3 and lim sup σn = 3 .
n→∞ n→∞
4. Show that
∞
2 1
(−1)k xk → as x → 1 − .
2
k=0

Suggestion. Show that the series an is Cesàro summable, where an =
k 2
(−1) if n = k and an = 0 otherwise.
5. (a) Show directly, without appeal to a Tauberian theorem, that the
infinite series ∞ 1
n=1 n is not Cesàro summable.
(b) Show directly that the same series is not Abel summable.

6. Show that for an ≥ 0, the series an is Abel summable only if it is
convergent.
7. (a) Prove the Abelian theorem for Lambert summability. In other words,
show that if
∞
∞

an = s , then lim an L(xn ) = s ,
x→1−
n=0 n=0
where L(x) is the Lambert kernel (cf. Section 7.1).

Hint. Apply a summation by parts and observe that L(x) increases from 0
to 1 in the interval [0, 1].
(b) Make the change of variables x = e−t to recast the conclusion in the
form
∞
nan t
a0 + lim = s.
t→0+ e −1
nt
n=1
8. Show that the series

∞
∞

cos nθ and sin nθ , 0 < θ < 2π ,
n=1 n=1
are Abel summable to − 12 and 1

2 cot θ2 , respectively.
∞ n,
Hint. Consider real and imaginary parts of the geometric series n=1 z
where z = reiθ .
References
[1] E. R. Berlekamp and J. P. Buhler (editors), Puzzles Column, Emissary, Math-
ematical Sciences Research Institute, Berkeley, Fall 2004 and Spring 2005.
[2] G. Frobenius, “Ueber die Leibnitzsche Reihe”, J. Reine Angew. Math. 89 (1880),
262–264.
[3] G. H. Hardy, “On certain oscillating series”, Quart. J. Math. 38 (1907), 269–288.
[4] G. H. Hardy, “Theorems relating to the summability and convergence of slowly
oscillating series”, Proc. London Math. Soc. 8 (1910), 301–320.
[5] G. H. Hardy, Divergent Series, Clarendon Press, Oxford, 1949.
[6] G. H. Hardy and J. E. Littlewood, “Tauberian theorems concerning power series
and Dirichlet’s series whose coefficients are positive”, Proc. London Math. Soc. 13
(1914), 174–191.
[7] G. H. Hardy and J. E. Littlewood, “A further note on the converse of Abel’s
theorem”, Proc. London Math. Soc. 25 (1926), 219–236.
References 195
[8] J. Karamata, “Über die Hardy–Littlewoodschen Umkehrung des Abelschen

Stetigkeitssatzes”, Math. Zeitschrift 32 (1930), 319–320.
[9] J. P. Keating and J. B. Reade, “Summability of alternating gap series”, Proc.
Edinburgh Math. Soc. 43 (2000), 95–101.
[10] J. Korevaar, Tauberian Theory: A Century of Developments, Springer-Verlag,
Berlin–Heidelberg, 2004.
[11] J. E. Littlewood, “The converse of Abel’s theorem on power series”, Proc.
London Math. Soc. 9 (1911), 434–448.
[12] A. Tauber, “Ein Satz aus der Theorie der unendlichen Reihen”, Monatsh. Math.
u. Phys. 8 (1897), 273–277.
[13] E. C. Titchmarsh, The Theory of Functions, Second edition, Oxford University
Press, London, 1939.
[14] H. Wielandt, “Zur Umkehrung des Abelschen Stetigkeitssatzes”, Math. Zeit-
schrift 56 (1952), 206–207.
Chapter 8
Fourier Series
Power series expansions of functions are known to play a vital role in anal-
ysis. In this chapter we turn to another kind of expansion, a sum of sines
and cosines known as a Fourier series. It is fundamentally different from a
Taylor series expansion in that the coefficients are determined not by dif-
ferentiation but by integration, exploiting a property of orthogonality; and
the function to be represented need not be differentiable and may even have
points of discontinuity. Here we develop criteria for convergence and Cesàro
summability of Fourier series, with applications to specific examples. The
discussion then shifts to a continuous analogue, the Fourier transform and its
inversion, leading ultimately to the remarkable Poisson summation formula.
8.1. Physical origins

The theory of Fourier series has its roots in problems of mathematical
physics, classical problems of wave propagation, heat conduction, electro-
static potentials, and the like. Consider for example the problem of calcu-
lating the steady-state distribution of temperature u(r, θ) in a circular metal
plate of unit radius, given a boundary temperature u(1, θ) = f (θ), where
r and θ are standard polar coordinates. Here f is an “arbitrary” function,
not necessarily continuous, with the property f (θ + 2π) = f (θ). According
to the laws of physics, the solution u(r, θ) will be a harmonic function; that
is, a solution of the Laplace equation, which takes the form
∂ 2 u 1 ∂u 1 ∂2u
+ + =0
∂r2 r ∂r r2 ∂θ2
197
198 8. Fourier Series
in polar coordinates.
Now if f (θ) = sin nθ or cos nθ for n = 1, 2, . . . , this boundary-value
problem can be solved by inspection. The solutions are u(r, θ) = rn sin nθ
and rn cos nθ, respectively. But the Laplace equation is linear, so a principle
of superposition applies. This means that for any constant c, the solution
with boundary function cf (θ) is cu(r, θ), and if v(r, θ) is the solution cor-
responding to another boundary function g(θ), then the u + v solves the
problem for boundary function f + g.
All of this suggests that the boundary-value problem will be solved for
an arbitrary periodic function f (θ) if that function can be expanded into an
infinite series of the form
∞

1
(1) f (θ) = 2 a0 + (an cos nθ + bn sin nθ) ,
n=1
for some coefficients an and bn . Setting aside questions of convergence, we

will then arrive at the solution
∞

u(r, θ) = 12 a0 + (an rn cos nθ + bn rn sin nθ) .
n=1
An expansion of the form (1) is called a Fourier series, after the French
mathematician and physicist Joseph Fourier (1768–1830), who conceived the
method and developed it in Théorie analytique de la chaleur, a revolution-
ary monograph on heat conduction published in 1822. Fourier’s arguments
were not rigorous, and his ideas were not readily accepted. In particular,
the notion that a discontinuous function could be represented as a sum of
sines and cosines seemed inconceivable and was greeted with skepticism. Ul-
timately the paradoxes inherent in Fourier series played an important role in
leading Cauchy and others to develop more precise concepts in mathematical
analysis, thereby placing the entire apparatus on a more secure foundation.
Fourier series are the prototype for a variety of orthogonal expansions
that arise in similar manner from problems of mathematical physics. For ex-
ample, expansions into series of Bessel functions and Legendre polynomials
will be encountered in Chapter 13 of this book.
Joseph Fourier had a colorful career that combined science with public
service. His early work on the theory of algebraic equations was interrupted
by the French Revolution of 1789. He then became embroiled in politics and
was actually condemned to the guillotine at one time. After the Revolution
he taught for a while at the new École Polytechnique in Paris, then was
appointed as a scientific advisor in Napoleon’s expedition to Egypt. Upon
return to France, he became a civil administrator in Grenoble, during which
8.2. Orthogonality relations 199
time (around 1805) he began the work on heat diffusion that would culminate
in his famous book of 1822.
Since the time of Fourier, an extensive mathematical theory of Fourier se-
ries has emerged, with broad applications not only to mathematical physics,
but to such fields as coding theory, signal transmission, data storage and re-
trieval, crystallography, medical imaging, and pure mathematics itself. The
mathematical theory will be our focus in this chapter.
8.2. Orthogonality relations

The theory of Fourier series is based on orthogonality relations for sines
and cosines. By analogy with vectors in Euclidean spaces, two functions f
and g are said to be orthogonal over an interval [a, b] if
b
f (x)g(x) dx = 0 .
a
The fundamental relations are

π π
cos nx dx = sin nx dx = 0 , n = 1, 2, . . . ,
−π −π
π
cos nx sin mx dx = 0 , n, m = 1, 2, . . . ,
−π
(2) π π
cos nx cos mx dx = sin nx sin mx dx = 0 , n = m ,
−π −π
π π
2
cos nx dx = sin2 nx dx = π , n = 1, 2, . . . .
−π −π
They follow easily from the trigonometric identities

2 cos nx sin mx = sin(n + m)x − sin(n − m)x ,
2 cos nx cos mx = cos(n + m)x + cos(n − m)x ,
2 sin nx sin mx = cos(n − m)x − cos(n + m)x .
Suppose now that for some coefficients ak and bk the trigonometric series
∞

(3) f (x) = 12 a0 + (ak cos kx + bk sin kx)
k=1
converges uniformly in the interval [−π, π] to a sum f (x). (The reason for
the 12 will become apparent shortly.) Because of the uniform convergence,
the function f is continuous and we can integrate term by term to see that
π π
1
f (x) dx = 2 a0 = πa0 .
−π −π
If we multiply by cos nx for some index n = 1, 2, . . . and integrate term by

term, we find
π ∞
π ∞ π
f (x) cos nx dx = ak cos kx cos nx dx + bk sin kx cos nx dx
−π k=1 −π k=1 −π
= πan + 0 = πan ,
in view of the orthogonality relations (2). Similarly, multiplication by sin nx
and integration gives
π
f (x) sin nx dx = πbn .
−π
The numbers

1 π
an = f (x) cos nx dx , n = 0, 1, 2, . . . ,
π −π
(4) π
1
bn = f (x) sin nx dx , n = 1, 2, . . .
π −π
are called the Fourier coefficients of the function f , and the trigonometric
series of the form (3) with these coefficients is known as the Fourier series
of f . As our calculations show, the Fourier series of a continuous function f
is the only trigonometric series of the form (3) that can converge uniformly
to f in the interval [−π, π]. In particular, no function can have more than
one uniformly convergent trigonometric series expansion.
A much more delicate question is whether the Fourier series of a given
continuous periodic function will actually converge to the function, uni-
formly or even pointwise. Before turning to this important question, we
shall consider a problem of best “mean-square” approximation.
8.3. Mean-square approximation

A function f is said to be square-integrable over an interval [a, b] if it
b
is Riemann integrable over [a, b] and the integral a f (x)2 dx is finite. Sup-
pose f is square-integrable over [−π, π], and consider the following question.
Among all trigonometric polynomials

n
1
T (x) = 2 c0 + (ck cos kx + dk sin kx)
k=1
of degree at most n, which one gives the best mean-square approximation

to f ? In other words, how should the coefficients ck and dk be chosen to
minimize the integral π
[f (x) − T (x)]2 dx ?
−π
8.3. Mean-square approximation 201
To answer this question, we let ak and bk be the Fourier coefficients of f , as

defined by (4), and write
(5)
π π π π
[f (x) − T (x)] dx =
2
f (x) dx − 2
2
f (x)T (x) dx + T (x)2 dx
−π −π −π −π
π
n
n
= f (x)2 dx − πa0 c0 − 2π (ck ak + dk bk ) + 12 πc20 + π (c2k + d2k ) ,
−π k=1 k=1
where
π the orthogonality relations (2) have been used to calculate the integral
2
−π T (x) dx. Now let

n
(6) sn (x) = 12 a0 + (ak cos kx + bk sin kx)
k=1
denote the nth partial sum of the Fourier series of f . With the choice of
trigonometric polynomial T (x) = sn (x), the expression (5) reduces to
π π
n
(7) [f (x) − sn (x)]2 dx = f (x)2 dx − 12 πa20 − π (a2k + b2k ) .
−π −π k=1
A straightforward calculation now leads to the identity

π π
[f (x) − T (x)]2 dx = [f (x) − sn (x)]2 dx + 12 π(a0 − c0 )2
−π −π

n

+π (ak − ck )2 + (bk − dk )2 .
k=1
From this it follows that

π π
[f (x) − T (x)]2 dx ≥ [f (x) − sn (x)]2 dx ,
−π −π
with equality if and only if c0 = a0 and ck = ak , dk = bk for k = 1, 2, . . . , n.

In other words, the Fourier polynomial sn gives the best mean-square ap-
proximation to f among all trigonometric polynomials of degree less than
or equal to n.
The identity (7) also shows that

n π
1 2
2 πa0 +π (ak + bk ) ≤
2 2
f (x)2 dx ,
k=1 −π
π
since −π [f (x) − sn (x)]2 dx ≥ 0. Letting n → ∞, one sees that the infinite
series converges and that
∞
π
(8) 1 2
2 πa0 + π (ak + bk ) ≤
2 2
f (x)2 dx .
k=1 −π
The relation (8) is known as Bessel’s inequality. An immediate corollary is

that the Fourier coefficients ak and bk of a square-integrable function tend
to zero as n → ∞.
Actually, Bessel’s inequality is valid for more general systems of orthog-
onal functions; no special properties of sines and cosines were used in its
derivation. However, for the trigonometric system considered here, we can
actually assert that
∞
π
2 2 2
1
2 πa 0 + π (ak + bk ) = f (x)2 dx ,
k=1 −π
an identity known as Parseval’s relation. This will follow from the equation
(7) if we can show that
π
(9) lim [f (x) − sn (x)]2 dx = 0 .
n→∞ −π
For this purpose we will appeal to the trigonometric form of the Weierstrass
approximation theorem, stated here for easy reference.
Weierstrass Approximation Theorem. Let f (x) be continuous on the
interval [−π, π], with f (−π) = f (π). Then for each ε > 0 there is a trigono-
metric polynomial T (x) such that |f (x) − T (x)| < ε for −π ≤ x ≤ π.
The trigonometric form of the Weierstrass approximation theorem is
closely related to the algebraic form, as presented in Chapter 6. A proof is
outlined in Exercise 6. Another proof will be given later in this chapter as
a corollary of Fejér’s theorem. To show that the relation (9) holds for every
function f that is square-integrable over [−π, π], we note first that such a
function can be approximated in mean by a function g that is continuous
on [−π, π] and has the property g(−π) = g(π). Combining this fact with
the Weierstrass approximation theorem, we see that for each ε > 0 there is
a trigonometric polynomial T such that
π
[f (x) − T (x)]2 dx < ε .
−π
But if T has degree m, it then follows that

π
[f (x) − sn (x)]2 dx < ε for all n ≥ m ,
−π
since we have shown that the Fourier polynomial sn gives the best mean-
square approximation to f among all trigonometric polynomials of degree n
or lower. This verifies (9), which proves the Parseval relation.
8.4. Convergence of Fourier series

Our goal is now to develop conditions that ensure the convergence of
a Fourier series to its associated function. In order to obtain a suitable
expression for the partial sums, we begin with an elementary calculation.
Lemma 1.

n
sin(n + 12 )x
1
2 + cos kx = , n = 1, 2, . . . .
k=1
2 sin 12 x
Proof. Multiply the sum by 2 sin 12 x and use the identity

2 sin 12 x cos kx = sin(k + 12 )x − sin(k − 12 )x .
The resulting sum then telescopes and gives the stated result.
The function
sin(n + 12 )x
n
1
Dn (x) = = 2 + cos kx
2 sin 12 x k=1
is known as the Dirichlet kernel. It has the property

1 π
(10) Dn (x) dx = 1 , n = 1, 2, . . . ,
π −π
π
since −π cos kx dx = 0 for k = 1, 2, . . . . The Dirichlet kernel is an even
function, periodic with period 2π. Thus Dn (−x) = Dn (x) and Dn (x+2π) =
Dn (x). A graph of the Dirichlet kernel for n = 10 is shown in Figure 1.
x
−π π
Figure 1. The Dirichlet kernel D10 (x).

With the help of Lemma 1, we may now derive a useful formula for the
nth partial sum sn (x) of the Fourier series of a function f , as defined by
(6). We take f to be integrable over the interval [−π, π], with the property
f (−π) = f (π), and we extend it to the whole real line as a periodic function
with period 2π. Introducing the formulas (4) for the Fourier coefficients of
f , we can write

1 π
1

n

sn (x) = 2 + cos kx cos kt + sin kx sin kt f (t) dt

π −π k=1
π n
1
= 1
2 + cos k(x − t) f (t) dt
π −π k=1
π π
1 1
= Dn (x − t)f (t) dt = Dn (t)f (x + t) dt ,
π −π π −π
since the Dirichlet kernel Dn is even and both Dn and f have period 2π.
This relation

1 π
(11) sn (x) = Dn (t)f (x + t) dt
π −π
is known as the Dirichlet formula for the nth partial sum of a Fourier series.
In view of the property (10), we have deduced that

1 π
(12) sn (x) − f (x) = Dn (t)[f (x + t) − f (x)] dt .
π −π
In order to prove that sn (x) → f (x), as n → ∞, we have only to show that

the integral tends to zero. Were it true that Dn (x) ≥ 0, an argument similar
to Landau’s proof of the Weierstrass approximation theorem (see Chapter
6) could be given for any point x where f is continuous. However, it is
apparent that Dn (t) alternates in sign in the interval [−π, π]. In fact, it can
be shown (see Exercise 7) that
π
|Dn (t)| dt ∼ C log n , n→∞
−π
for some constant C > 0, so this method of proof cannot succeed. What is
worse, it is actually possible for the Fourier series of a continuous periodic
function to diverge at some points. Thus in order to prove that the Fourier
series converges to the function at a given point, we shall have to impose
hypotheses stronger than continuity. We will assume that the function sat-
isfies a smoothness condition near the point where convergence is to occur.
The precise statement is as follows.
Convergence Theorem. Let a function f be square-integrable over the

interval [−π, π], and let sn be the nth partial sum of its Fourier series. Let
f be extended to the real line by the periodicity condition f (x + 2π) = f (x).
Suppose that at some point x ∈ R the extended function satisfies a Lipschitz
condition:
|f (x + t) − f (x)| ≤ C |t| , |t| < δ ,
for some positive constants C and δ. Then sn (x) → f (x) as n → ∞.
Before passing to the proof, let us observe one very remarkable feature
of the theorem. The Fourier coefficients ak and bk are determined by the
values of f throughout an interval of length 2π, yet the convergence of the
resulting Fourier series at any particular point x is governed only by the
behavior of f in an arbitrarily small neighborhood of x. This phenomenon
is known as Riemann’s localization principle.
Proof of theorem. Begin by recalling the corollary of Bessel’s inequality
(8), that the Fourier coefficients of any square-integrable function g tend to
zero: π π
lim g(t) cos nt dt = lim g(t) sin nt dt = 0 .
n→∞ −π n→∞ −π
Now write the formula (12) as

1 π
sn (x) − f (x) = Dn (t)[f (x + t) − f (x)] dt
π −π
(13)
1 π
= ϕx (t) sin(n + 12 )t dt ,
π −π
where
f (x + t) − f (x)
ϕx (t) = .
2 sin 12 t
Then by the smoothness hypothesis on f ,
|f (x + t) − f (x)| C |t|
|ϕx (t)| = ≤ ≤C
2 | sin 12 t| 2 | sin 12 t|
for |t| ≤ δ. On the other hand, | sin 12 t| ≥ | sin(δ/2)| > 0 elsewhere in

the interval [−π, π], and f is a square-integrable function. Therefore, the
function ϕx (t) is square-integrable over the interval [−π, π]. But
sin(n + 12 )t = cos 12 t sin nt + sin 12 t cos nt ,
so the integral
π
1
(14) ϕx (t) sin(n + 12 )t dt
π −π
can be regarded as a sum of Fourier sine and cosine coefficients of the square-
integrable functions ϕx (t) cos 12 t and ϕx (t) sin 12 t, respectively. Thus by the
corollary to Bessel’s inequality, the integral (14) tends to zero as n → ∞. It
therefore follows from (13) that sn (x) → f (x) as n → ∞, which completes
the proof.
It may be remarked that the proof remains valid under the weaker
smoothness hypothesis
|f (x + t) − f (x)| ≤ C |t|α , |t| < δ ,
for some exponent α > 12 . That still ensures the square-integrablity of ϕx (t),
so its Fourier coefficients tend to zero.
There is a more general form of the convergence theorem that applies to
functions with jump discontinuities. If the function f has one-sided limits
f (x+) and f (x−) at a point x, and it is sufficiently smooth in both of the one-
sided neighborhoods of x, then its Fourier series converges to the average
of the two limits. This may be viewed as a corollary of the convergence
theorem. Here is a more precise statement.
Corollary. Let f be periodic with period 2π and square-integrable over
[−π, π]. At some point x ∈ R, suppose f has one-sided limits
f (x+) = lim f (x + t) , f (x−) = lim f (x − t) .
t→0, t>0 t→0, t>0
Suppose further that

|f (x + t) − f (x+)| ≤ Ct and |f (x − t) − f (x−)| ≤ Ct , 0 < t < δ,
for some positive constants C and δ. Then sn (x) → 12 [f (x+) + f (x−)] as
n → ∞.
Proof. The result can be deduced from the more special form of the con-
vergence theorem by subtracting from f the function g that has the same
jump as f and is constant on both sides of the jump-point. Then f − g has
the properties required in the convergence theorem, and a direct calculation
shows that the partial sums of g are all equal to 12 [g(x+) + g(x−)] at the
point x. Rather than pursue this line of proof, however, we will simply adapt
the earlier proof of the convergence theorem to the more general situation.
Dirichlet’s formula (11) allows us to write

1 0
sn (x) − 12 [f (x+) + f (x−)] = Dn (t)[f (x + t) − f (x−)] dt
π −π

1 π
+ Dn (t)[f (x + t) − f (x+)] dt .
π 0
8.5. Examples 207
By considering now the two square-integrable functions

⎧
⎨ f (x + t) − f (x−) , −π ≤ t < 0
ϕx (t) = 2 sin 12 t
⎩
0, 0≤t≤π
and ⎧
⎨ 0, −π ≤ t ≤ 0
ψx (t) = f (x + t) − f (x+)
⎩ , 0 < t ≤ π,
2 sin 12 t
we conclude as before that sn (x) → 12 [f (x+) + f (x−)] as n → ∞.
Most of the preceding analysis is due to Peter Gustav Lejeune Dirichlet

(1805–1859), who gave the first rigorous proof of convergence of Fourier
series in 1829. Dirichlet was a major analyst and number theorist, regarded
as the founder of analytic number theory. Born in the town of Düren (now in
Germany) to a family of Belgian origin, he entered the Collège de France in
Paris at age 17. Among his teachers there were Fourier, Laplace, Legendre,
and Poisson. By 1828 he held a professorship in Berlin, and later moved to
Göttingen as successor to Gauss.
The convergence theorem as presented here is not the most general, but
it is general enough to cover most functions encountered in applications.
We will see that it is also a very effective mathematical tool. Refinements
and further information can be found in many books on Fourier series; for
instance, in those of Jackson [7], Körner [8], Rogosinski [9], Widder [10],
and Zygmund [11].
8.5. Examples
In calculating the Fourier coefficients of a given function f it is often
useful to take advantage of special symmetries. For instance, if f is an even
function, then f (x) sin nx is odd, and it is clear without calculation that the
Fourier sine coefficients of f must vanish: bn = 0 for all n. Also, the Fourier
cosine coefficients take the form

2 π
an = f (x) cos nx dx n = 0, 1, 2, . . . .
π 0
Similarly, if f is an odd function, or if it differs by a constant from an

odd function, then an = 0 for all n ≥ 1 and the formula for bn may be
correspondingly simplified.
Example 1. For a first example, consider the function f (x) = |x| for
−π ≤ x ≤ π, and let it be extended periodically to the whole line so that
f (x + 2π) ≡ f (x). The result is the “sawtooth function” whose graph is

shown in Figure 2. Here f is an even function in the interval [−π, π], so
bn = 0 for all n; while
2 π
a0 = x dx = π
π 0
and an integration by parts gives
4
2 π 2 − 2 , n odd
an = x cos nx dx = 2 (−1) − 1 =n
n π
π 0 n π 0, n even.
The function f satisfies a uniform Lipschitz condition (with C = 1), so the
convergence theorem applies and gives
∞
π 4 1
|x| = − cos(2k + 1)x
2 π (2k + 1)2
k=0

π 4 1 1
= − cos x + cos 3x + cos 5x + . . .
2 π 9 25
for −π ≤ x ≤ π. By the Weierstrass M-test, the convergence is uniform.
With the choice x = 0, the identity reduces to
∞
1 π2
= ,
(2k + 1)2 8
k=0
from which it is easy to deduce Euler’s sum
∞
1 π2
= .
n2 6
n=1
The uniform convergence of the cosine series to the sum |x|, periodically
extended to the sawtooth function f (x), is shown graphically in Figure 2,
where Mathematica plots compare the function with the partial sum
4
2
π 1
s5 (x) = − cos(2k + 1)x
2 π (2k + 1)2
k=0
of its Fourier series. Evidently the convergence is quite rapid, since the two
graphs are barely distinguishable.
Figure 2. Fourier series of |x| extended periodically.

8.5. Examples 209
Example 2. Another example worth exploring is the “square wave” func-

tion defined as the periodic extension of

−1 , −π < x < 0
f (x) =
1, 0 < x < π.
This is an odd function whose Fourier coefficients are an = 0 and
4
2 π 2 , n odd
bn = sin nx dx = 1 − (−1)n = nπ
π 0 nπ 0, n even.
Except at its points of discontinuity, the function f clearly satisfies the
requirements of the convergence theorem, so we infer that
∞
4 1 −1 , −π < x < 0
sin(2k + 1)x =
π 2k + 1 1, 0 < x < π.
k=0
Figure 3 illustrates the convergence by displaying plots of the partial sums

of order 10 and 20, containing 5 and 10 terms, respectively.
Example 3. For a third example, consider the periodic extension of the
linear function
f (x) = 12 (π − x) , 0 < x < 2π .
This is a sawtooth function with discontinuities at 0, ±2π, ±4π, . . . . Be-
cause it is an odd function, it has Fourier coefficients an = 0 and

1 π 1
bn = (π − x) sin nx dx = , n = 1, 2, . . . ,
π 0 n
as an integration by parts shows. Another appeal to the convergence theo-
rem now verifies that
∞ 1
sin kx − (π + x) , −π ≤ x < 0
= 12
2 (π − x) , 0 < x ≤ π.
k
k=1
Figure 3. Fourier series of square wave, n = 10 and 20.

Figure 4. Fourier series of discontinuous sawtooth, n = 10 and 20.
Figure 4 shows graphs of the partial sums of order n = 10 and 20.
Example 4. We turn now to a special example, the Fourier series expansion

of the function cos cx, where the constant c is not an integer. The resulting
identity will be seen to have surprising implications.
For c = 0, ±1, ±2, . . . , the function cos cx does not have period 2π, but
it has a continuous 2π-periodic extension f (x) beyond the interval [−π, π],
since it is an even function and so cos(−cπ) = cos cπ. The periodic extension
f (x) need not be differentiable at the points x = ±π, but it satisfies a
Lipschitz condition there. The basic convergence theorem therefore applies
and guarantees that the Fourier series converges to f (x) at every point
x ∈ R. In particular, it guarantees the validity of the formula
∞

2c 1 (−1)n
(15) cos cx = sin cπ + cos nx , −π ≤ x ≤ π .
π 2c2 c2 − n2
n=1
In order to verify (15), we have only to check that the right-hand side is
indeed the formal Fourier series of the function f . We may use the trigono-
metric identity
2 cos cx cos nx = cos(c + n)x + cos(c − n)x
to calculate the Fourier coefficients

1 π 1 π 2c (−1)n
an = f (x) cos nx dx = cos cx cos nx dx = sin cπ 2
π −π π −π π c − n2
for n = 1, 2, . . . , and

1 π 1 π 2
a0 = f (x) dx = cos cx dx = sin cπ .
π −π π −π cπ
8.5. Examples 211
All of the Fourier sine-coefficients bn vanish, since cos cx is an even function.

Thus the infinite series in the formula (15) is the formal Fourier series of
f (x), and the basic convergence theorem shows that it converges to the sum
cos cx throughout the interval [−π, π].
Now comes an audacious trick. The equation (15) holds for every non-
integer constant c and for all x in the interval [−π, π]. Let us fix x = π and
regard c as the variable. Changing notation from c to t, we deduce from
(15) that
∞
1 2t
(16) π cot πt − = , t = 0, ±1, ±2, . . . .
t t2 − n2
n=1
Since the series in (16) converges uniformly in each interval [−a, a] with
0 < a < 1, we may integrate term by term over an interval [ε, x], where
0 < ε < x < 1, to obtain
∞ 2
sin πx sin πε x − n2
log − log = log .
πx πε ε2 − n2
n=1
Letting ε tend to zero, we conclude that

∞
sin πx x2
log = log 1 − 2 , 0 < x < 1.
πx n
n=1
Exponentiating, we arrive at the infinite product representation

x2 x2
(17) sin πx = πx 1 − x2 1 − 2 1 − 2 ··· , −1 < x < 1 .
2 3
In fact, the formula (17) holds for all x ∈ R. To see this, we need only
show that the right-hand side is periodic with period 2. In other words,
if p(x) denotes the infinite product (which converges for every x ∈ R), we
want to show that p(x + 2) = p(x). But the partial product

x2 x2
pn (x) = πx 1 − x 2
1 − 2 ··· 1 − 2
2 n
can be rewritten as
(−1)n π
pn (x) = (x − n) · · · (x − 1)x(x + 1) · · · (x + n) ,
(n!)2
which gives the relation
(x + n + 1)(x + n + 2)
pn (x + 2) = pn (x) , n = 1, 2, . . . .
(x − n + 1)(x − n)
Letting n tend to infinity, we conclude that p(x + 2) = p(x).
Setting x = 12 in (17), we obtain another proof of the Wallis product

formula
1 1 1 2
1− 2 1− 2 1 − 2 ··· = .
2 4 6 π
By comparing the coefficient of x3 in the product (17) with that of the

Taylor series expansion of sin πx, we are led to Euler’s sum 1/n = π 2 /6.
2
This was Euler’s original derivation, although his discovery of the product
formula for the sine function amounted only to a plausible argument.
8.6. Gibbs’ phenomenon

For Examples 2 and 3, the graphs of partial sums converge nicely to the
function away from the points of discontinuity, but they exhibit a persistent
“overshoot” near the jump points. This strange behavior is known as the
Gibbs phenomenon. Its discovery is generally attributed to the American
mathematical physicist J. W. Gibbs in 1898, but it goes back at least to
the English mathematician Henry Wilbraham in 1848. For this reason it
is sometimes called the Gibbs–Wilbraham phenomenon. Wilbraham made
painstaking hand calculations of partial sums for Example 2 and plotted
graphs much like those in Figure 3.
In order to give a precise formulation of the Gibbs phenomenon, we shall
focus on Example 3. Let

n
sin kx
sn (x) =
k
k=1
denote the partial sums of that Fourier series. Let ξn denote the first positive
point where the function sn (x) attains a local maximum. An inspection of
Figure 4 suggests the conjecture that the sequence {sn (ξn )} tends to a limit
larger than π/2 as n → ∞. In other words, the overshoot persists in the
limit.
To prove the conjecture, first calculate the derivative

n
sn (x) = cos kx = Dn (x) − 12 ,
k=1
where
sin(n + 12 )x
Dn (x) =
2 sin 12 x
is the Dirichlet kernel, as found in Lemma 1. Thus the equation sn (x) = 0
is equivalent to
sin(n + 12 )x = sin 12 x .
8.6. Gibbs’ phenomenon 213
This last equation implies that either

1
2x + 2kπ = (n + 12 )x or π − 12 x + 2kπ = (n + 12 )x
for k = 0, ±1, ±2, . . . . Hence the positive critical points of sn (x) in the
interval (0, π] have the form
2kπ
x= , k = 1, 2, . . . , [n/2] ,
n
and
(2k + 1)π
x= , k = 0, 1, 2, . . . , [n/2] ,
n+1
where [x] denotes the integer part of x. The critical points in the first
group are found to be local minima and those in the second group are local
maxima. Thus the first positive local maximum of sn (x) occurs at the point
π
x = ξn = .
n+1
Let us now investigate the behavior of

n
1 kπ
sn (ξn ) = sin
k n+1
k=1
as n → ∞. A little manipulation brings the sum to the form

n+1 n+1
n+1 kπ π
sn (ξn ) = sin = f (xk ) Δxk ,
kπ n+1 n+1
k=1 k=1
where
sin x kπ π
f (x) = , xk = , and Δxk = .
x n+1 n+1
The last sum is none other than a Riemann sum for the integral
π π
sin x
f (x) dx = dx
0 0 x
with respect to a partition of the interval [0, π] by n + 1 equally spaced

points. With the understanding that f (0) = 1, the function f is continuous
on the interval [0, π]. Therefore, the Riemann sums tend to the integral as
n → ∞, and we have proved that
π
sin x
lim sn (ξn ) = dx .
n→∞ 0 x
π 2π 3π 4π
sin x
Figure 5. Graph of x .
A graph of the function (sin x)/x is displayed in Figure 5. The graph

suggests that
sin x x
>1− , 0 < x < π,
x π
and this is not difficult to confirm (see Exercise 8). It follows that
π π
sin x x π
dx > 1− dx = .
0 x 0 π 2
Numerical evaluation of the (nonelementary) integral actually gives

π
sin x π
dx = 1.8519 · · · = (1.1789 . . . ) .
0 x 2
In the limit, then, the overshoot amounts to about 9% of the total jump π.
Although a particular example has served to illustrate the Gibbs phe-
nomenon, it is important to emphasize that the phenomenon is quite general.
It is true that our analysis has taken advantage of special features of the
sawtooth function, and a similar analysis applies to the square wave of Ex-
ample 2 (see Exercise 10). However, the phenomenon is not peculiar to
special examples and will manifest itself at any jump-discontinuity. To be
more precise, suppose that a periodic function g has a jump-discontinuity
at a point x0 , so that g(x0 +) − g(x0 −) = c > 0, where g(x0 +) and g(x0 −)
denote the right-hand and left-hand limits of g(x), respectively, at x0 . Then,
if f is the sawtooth function of Example 3, the function
c
h(x) = g(x) − f (x − x0 )
π
has a continuous extension to x0 and the partial sums of its Fourier series
are
c
un (x) = tn (x) − sn (x − x0 ) ,
π
where sn and tn are respective partial sums for f and g. If g is suffi-

ciently well behaved elsewhere, it can be shown that un (x) → h(x) pointwise
throughout some open neighborhood of x0 , and the convergence is uniform
at each such point. In particular, the convergence is uniform at x0 , which
is to say that for each ε > 0 there exist a δ > 0 and an integer N such that
|un (x) −h(x)| < ε for |x −x0 | < δ and n > N . Equivalently, un (xn ) → h(x0 )
whenever xn → x0 . Therefore, since the sums sn (x−x0 ) have been shown to
exhibit Gibbs’ phenomenon at x0 , the sums tn (x) must also exhibit Gibbs’
phenomenon there, with the same relative overshoot (0.1789 . . . )c/2, again
about 9% of the total jump c. Further details will not be pursued here, but
can be found for instance in Zygmund’s book [11].
There is a very interesting account of the Gibbs–Wilbraham phenomenon
and its history in the paper by Hewitt and Hewitt [4].
8.7. Arithmetic means of partial sums

The convergence theorem of Section 3 asserts that the Fourier series of a
continuous periodic function converges wherever the function is sufficiently
smooth. In Section 8.8 an example will be constructed to show that conti-
nuity alone does not guarantee the convergence. More precisely, the partial
sums sn (x) of a continuous periodic function f (x) need not converge and can
even be unbounded. However, we now proceed to show that the sequence
of arithmetic means
1
n
(18) σn (x) = sk (x) , n = 1, 2, . . . ,
n+1
k=0
always converges uniformly to the given function. This important result is

known as Fejér’s theorem, after the Hungarian mathematician Leopold Fejér
(1880–1959), who discovered it in 1904. In the language of summability the-
ory, the theorem says that the Fourier series of every continuous function is
uniformly Cesàro summable to the function. The trigonometric polynomials
σn (x) are often called the Fejér means of f .
Fejér’s Theorem. Let f (x) be continuous on R and periodic with period

2π. Let σn (x) be the arithmetic means of the partial sums of the Fourier
series of f , as defined in (18). Then σn (x) → f (x) uniformly on R as
n → ∞.
Corollary. Every continuous periodic function f (x) can be approximated

uniformly on R by trigonometric polynomials. In other words, for each ε > 0
there exists a trigonometric polynomial T (x) such that |f (x) − T (x)| < ε for
all x ∈ R.
The corollary is the trigonometric form of the Weierstrass approximation

theorem. The proof of the theorem will require the calculation of another
trigonometric sum.
Lemma 2.

n
sin2 12 (n + 1)x
sin(k + 12 )x = , n = 1, 2, . . . .
k=0
sin 12 x
Proof of lemma. In view of the identity
2 sin 12 x sin(k + 12 )x = cos kx − cos(k + 1)x ,
the sum telescopes to give

n
1
2 sin 12 x sin(k + 12 )x = 1 − cos(n + 1)x = 2 sin2 2 (n + 1)x .
k=0
Preparatory to a proof of the theorem, we will use Lemma 2 to derive

a representation formula for the arithmetic means σn (x) that is analogous
to Dirichlet’s formula (11) for the partial sums sn (x). Recalling that the
Dirichlet kernel is
sin(k + 12 )t
Dk (t) = , k = 0, 1, 2, . . . ,
2 sin 12 t
and appealing to the Dirichlet formula and Lemma 2, we find

1
n π n
1 1
σn (x) = sk (x) = Dk (t) f (x + t) dt
n+1 n+1 π −π
(19) k=0 k=0
π
1
= Kn (t)f (x + t) dt ,
π −π
where
1 sin2 12 (n + 1)x
Kn (x) = .
n + 1 2 sin2 12 x
The function Kn (x) is called the Fejér kernel. Figure 6 shows a plot for
n = 10.
x
Π Π
Figure 6. The Fejér kernel K10 (x).
Because the Fejér kernel has the form
1
n
Kn (x) = Dk (x) ,
n+1
k=0
the property (10) of the Dirichlet kernel implies that

π
1
(20) Kn (x) dx = 1 , n = 1, 2, . . . .
π −π
Like the Dirichlet kernel, the Fejér kernel peaks at the origin as n increases,
but it has the crucial advantage that Kn (x) ≥ 0 for all x. Since sin2 12 x ≥
(x/π)2 , we see that
π2
(21) 0 ≤ Kn (x) ≤ , δ ≤ |x| ≤ π ,
2(n + 1)δ 2
for each δ in the interval 0 < δ < π.
Proof of Fejér’s theorem. Since f is continuous and periodic, it is uni-

formly bounded: |f (x)| ≤ M for all x ∈ R. Moreover, f is uniformly contin-
uous. Thus for each ε > 0 there is a number δ > 0 such that |f (x)−f (y)| < ε
for each pair of points x, y ∈ R with |x − y| < δ. In view of (19) and (20),
we can write
π
1
σn (x) − f (x) = Kn (t)[f (x + t) − f (x)] dt .
π −π
Given ε > 0, choose δ as above. Then because Kn (x) ≥ 0, we can appeal to

the properties (20) and (21) of the Fejér kernel to infer that
π −δ
1 2M
|σn (x) − f (x)| ≤ Kn (t)|f (x + t) − f (x)| dt ≤ Kn (t) dt
π −π π −π
δ π
1 2M
+ Kn (t)|f (x + t) − f (x)| dt + Kn (t) dt
π −δ π δ
2M π 2
≤ + ε < 2ε for all x ∈ R ,
(n + 1)δ 2
for every n sufficiently large. This proves that σn (x) → f (x) uniformly on
R as n → ∞.
The representation formula (19) for the Fejér means, together with the
property (20), shows at once that
(22) |σn (x)| ≤ max |f (x)| for all x ∈ R .

x∈[−π,π]
Hence there can be no Gibbs phenomenon for the Fejér means, no over-
shooting near points of discontinuity. The dramatic difference in mode of
convergence is illustrated in Figure 7, where the Fejér means σn (x) of the
discontinuous sawtooth function

− 12 (π + x) , −π ≤ x < 0
(23) f (x) = 1
2 (π − x) , 0<x≤π
are shown for n = 10 and 20. This should be compared with the Gibbs
phenomenon displayed by the partial sums sn (x), as shown in Figure 4.
Figure 7. Fejér means of discontinuous sawtooth, n = 10 and 20.

8.8. Continuous functions with divergent Fourier series 219
With the help of the property (22) of the Fejér means, we can now prove
that the partial sums sn (x) of the function (23) are uniformly bounded. The
estimate (22) gives |σn (x)| ≤ π/2. On the other hand, a calculation shows
that
1
n

|sn (x) − σn (x)| = sin kx < 1 .
n+1
k=1
Thus |sn (x)| ≤ 1 + π/2 for all x ∈ R. The details are left as an exercise.
8.8. Continuous functions with divergent Fourier series

We have seen that a Fourier series must converge to its associated func-
tion at every point where that function is sufficiently smooth. The question
remains whether continuity alone will ensure the convergence. The answer is
negative. The first example of a continuous function with divergent Fourier
series was given by Paul du Bois-Reymond in 1873, and was rather compli-
cated. Much later, in 1911, Fejér found a remarkably simple example, which
will now be presented. The construction is reminiscent of Weierstrass’s ex-
ample of a continuous nowhere differentiable function.
First define the trigonometric polynomials
1 1 1
gm (x) = + cos x + cos 2x + · · · + cos(m − 1)x
m m−1 m−2
1 1
− cos(m + 1)x − cos(m + 2)x − · · · − cos 2mx
2 m

m
1 m
sin kx
= cos(m − k)x − cos(m + k)x = 2 sin mx
k k
k=1 k=1
for m = 1, 2, . . . . Now let {mk } be a rapidly increasing sequence of positive

integers, to be specified later. Define the function
∞
1
(24) F (x) = gm (x) .
k2 k
k=1
1
Recall that the partial sums of the series k sin kx are uniformly bounded,
as we saw at the end of the preceding section. Thus the functions gm (x)
are also uniformly bounded, and the Weierstrass M-test shows that the
infinite series (24) converges uniformly on R. Therefore, the sum F (x) is a
continuous function on the real line, periodic with period 2π. With suitable
choice of the sequence {mk }, we will see that the Fourier series of F is
divergent at the origin; in fact, its partial sums are unbounded there.
Let smn (x) denote the nth partial sum of the Fourier series of gm (x).
Then smn (x) = gm (x) for all n ≥ 2m, since gm is a trigonometric polynomial
of degree 2m. In view of the uniform convergence, the Fourier coefficients of
F can be calculated by integrating the relevant infinite series term by term.
This says that each Fourier coefficient of F is the sum of corresponding
Fourier coefficients of the terms of the series (24). It follows that the nth
partial sum Sn (x) of the Fourier series of F is the sum of the nth partial
sums of the terms. In other words,
∞
1
(25) Sn (x) = sm n (x) .
k2 k
k=1
In particular,
∞
1
Sn (0) = sm n (0) .
k2 k
k=1
Observe now that smn (0) ≥ 0 for every pair of indices m and n, and that
m
1 1 dx
smm (0) = 1 + + · · · + > = log m .
2 m 1 x
Consequently, the relation (25) shows that for each index k

∞
1 1 1
Smk (0) = sm m (0) ≥ 2 smk mk (0) > 2 log(mk ) .
j2 j k k k
j=1
3
Now choose mk = 2k to infer that
1 3
Smk (0) > k log 2 = k log 2 ,
k2
so that the partial sums of F are unbounded at the origin.
The function F (x) can be translated to obtain a function F (x−r) whose
Fourier series diverges at a specified point r. Then by forming a suitable
weighted sum of the functions F (x − rk ) corresponding to an arbitrary se-
quence rk , we can construct a continuous function whose Fourier series di-
verges at every point rk . For instance, the sequence rk can be taken to be
an enumeration of the rational numbers, and we then obtain a continuous
function whose Fourier series diverges at a countable dense set of points.
On the other hand, Lennart Carleson proved in 1966 that the Fourier
series of every continuous function converges almost everywhere. In other
words, the points of divergence (if any) constitute a set of measure zero.
8.9. Fourier transforms

We now turn to the continuous analogue of a Fourier series, the Fourier
transform and its inversion. It is customary and convenient to express the
Fourier transform in the notation of complex exponentials eit = cos t+i sin t.
By way of motivation, we begin by recasting Fourier series expansions in
complex form. To fix ideas, suppose that f (t) is a real-valued Lipschitz-
smooth function defined on the real line, periodic with period 2π. We claim
that its Fourier series expansion can be expressed in the form
∞

(26) f (t) = cn eint ,
n=−∞
where the coefficients cn are defined by

2π
1
(27) cn = f (t)e−int dt , n = 0, ±1, ±2, . . . .
2π 0
To see this, observe that

2π
1 1
cn = f (t)(cos nt − i sin nt) dt = (an − ibn ) ,
2π 0 2
where
2π 2π
1 1
an = f (t) cos nt dt and bn = f (t) sin nt dt
π 0 π 0
are the Fourier coefficients as previously defined for n ≥ 0. Then c0 = 12 a0

and c−n = cn = 12 (an + ibn ), which shows that
cn eint + c−n e−int = an cos nt + bn sin nt , n = 1, 2, . . . .
Therefore,

n
1 n
ikt
ck e = a0 + (ak cos kt + bk sin kt) = sn (t)
2
k=−n k=1
is the nth partial sum of the Fourier series of f , expressed in complex nota-
tion.
Recall now that ei(s+t) = eis eit , by the addition formulas for sine and
cosine. Differentiation gives
d iat
e = −a sin at + ia cos at = iaeiat .
dt
The complex form of a Fourier series can be obtained more directly through
the orthogonality relations
2π 2π
int −imt i(n−m)t
0 for m = n
e e dt = e dt =
0 0 2π for m = n .
Formally, if a complex-valued function has an expansion of the form (26),
multiplication by e−imt and integration shows that the coefficients cn are
given by (27).
In what follows we will be dealing with complex-valued functions f (t)
defined on the real line R and Riemann integrable over each bounded inter-
val. For p > 0 we will write f ∈ Lp to indicate that |f (t)|p is integrable over
R . We will denote by L∞ the set of all bounded locally integrable functions
on R .
The Fourier transform
∞

f (x) = e−ixt f (t) dt , −∞ < x < ∞ ,
−∞
of a function f ∈ L1 plays the role of a sequence of Fourier coefficients. In

some sense, to be made precise later, the function can be recovered from its
Fourier transform by the formula
∞
1
f (t) = eixt f(x) dx ,
2π −∞
the continuous analogue of a Fourier series expansion. It is clear that
∞

|f (x)| ≤ |f (t)| dt < ∞ if f ∈ L1 ,
−∞
so that the Fourier transform is always a bounded function. Simple examples

will show, however, that the transform need not be integrable.
It is important to note that Fourier transformation is a linear operator.
In other words, if f, g ∈ L1 and α, β ∈ C , then αf + βg = αf + β
g. It may
also be observed that if g(t) = f (at + b) for some real constants a = 0 and
b, then
1 ibx/a

g (x) = e f (x/a) .
|a|
Here are some examples.
Example 1. For some constant R > 0, consider the function

1 for |t| ≤ R
f (t) =
0 for |t| > R .
Then for x = 0 we have

R R sin xt R
−ixt 2 sin Rx
f (x) = e dt = 2 cos xt dt = 2 = .
−R 0 x 0 x
For x = 0 the result is f(0) = 2R = limx→0 f(x). Thus f is continuous and

f(x) → 0 as |x| → ∞, but f ∈ / L1 .
Example 2. Suppose now that

|t|
1− R for |t| ≤ R
f (t) =
0 for |t| > R .
Then for x = 0 an integration by parts gives
R R
−ixt |t| t
f (x) = e 1− dt = 2 1− cos xt dt
−R R 0 R

2 sin(Rx/2) 2
= (1 − cos Rx) = R .
Rx2 Rx/2
For x = 0 we see that f(0) = R = limx→0 f(x).
Example 3. Next let f (t) = e−a|t| for some constant a > 0. Then since f
is an even function, we see that
∞ ∞ ∞

f (x) = e −ixt −a|t|
e dt = e −a|t|
cos xt dt = 2 e−at cos xt dt .
−∞ −∞ 0
The last expression is a Laplace transform that can be calculated through

two integrations by parts. The result is
2a
f(x) = .
a2 + x2
Example 4. Now, for some constant a > 0, consider the function

−at
e for t ≥ 0
f (t) =
0 for t < 0.
Then
∞ ∞
a − ix
f(x) = e −ixt −at
e dt = e−at (cos xt − i sin xt) dt = .
0 0 a2 + x 2
Observe again that f ∈

/ L1 .
Two more examples can be deduced from Examples 2 and 3 by applying
the inversion formula for Fourier transforms, to be proved in the next section.
Example 5. For some constant R > 0, let

2
2 sin(Rt/2)
f (t) = (1 − cos Rt) = R .
Rt2 Rt/2
Then
|x|
2π 1 − for |x| ≤ R
f(x) = R
0 for |x| > R .
Example 6. If f (t) = a
a2 +t2
for some constant a > 0, then f(x) = πe−a|x| .
Finally, it is interesting to note that the function f (t) = e−t /2 , which

2
plays a central role in probability theory, is essentially its own Fourier trans-
form. Here is a slight generalization.
Example 7. If f (t) = e−at
2 /2
for some constant a > 0, then

f(x) = 2π/a e−x
2 /2a
.
We will be able to calculate this transform in an elementary way after

having established some general properties of Fourier transforms, collected
in the following theorem.
Theorem. Let f ∈ L1 and let f(x) denote its Fourier transform. Then
(a) f(x) is uniformly continuous on R .
(b) f(x) → 0 as |x| → ∞ .
(c) If g(t) = tf (t) and g ∈ L1 , then f is differentiable and f (x) =
− i
g (x) .
(d) If f is differentiable and f ∈ L1 , then f has Fourier transform
f (x) = ixf(x) .
Preparatory to the proof, we record the simple inequality

iθ
(28) e − 1 ≤ |θ| , θ ∈ R.
To verify this, write

eiθ/2 − e−iθ/2
iθ iθ/2 −iθ/2
e − 1 = e −e = 2 = 2 |sin(θ/2)| ≤ |θ| .
2i
Proof of theorem. (a). Given ε > 0, choose a constant R > 0 so large

that −R ∞
ε
|f (t)| dt + |f (t)| dt < .
−∞ R 3
Then for arbitrary real numbers x and y, we can estimate

∞ −ixt

|f (x) − f (y)| = e −e −iyt
f (t) dt
−∞

−R R −ixt ∞
≤2 |f (t)| dt + e − e−iyt |f (t)| dt + 2 |f (t)| dt
−∞ −R R
R
2ε −i(x−y)t
< + e − 1 |f (t)| dt.
3 −R
But by (28) we see that

−i(x−y)t
e − 1 ≤ |(x − y)t| ≤ R |x − y| for |t| ≤ R ,
which shows that for some δ > 0, the inequality

R
2ε
|f(x) − f(y)| ≤ + R |x − y| |f (t)| dt < ε
3 −R
holds whenever |x − y| < δ. This proves the uniform continuity of f.

(b). Consider first the special case where f is the characteristic function
of a bounded interval [a, b]. In other words,

1 for a ≤ t ≤ b
f (t) =
0 for t < a or t > b.
Then we see from Example 1, by translation and change of scale, that

f(x) → 0 as |x| → ∞ . From this it follows more generally that f(x) → 0 for
any step-function f , by which we mean a finite linear combination of charac-
teristic functions of bounded intervals. Now consider an arbitrary function
f ∈ L1 . Given ε > 0, let R > 0 be chosen as in part (a). Then observe
that since f is Riemann integrable over [−R, R], it can be approximated in
L1 norm by step-functions. In other words, there exists a step-function ϕ(t)
supported on the interval [−R, R] for which
R
ε
|f (t) − ϕ(t)| dt < .
−R 3
Combining these estimates, we see that

∞

|f (x)| ≤ |f (x) − ϕ
(x)| + |
ϕ(x)| ≤ |f (t) − ϕ(t)| dt + |
ϕ(x)|
−∞
−R R ∞
= |f (t)| dt + |f (t) − ϕ(t)| dt + |f (t)| dt + |
ϕ(x)|
−∞ −R R
ε ε
< + + |
ϕ(x)| < ε
3 3
for |x| sufficiently large, since ϕ
(x) → 0 as |x| → ∞ . This proves that
f(x) → 0 as |x| → ∞ .
(c). To show that f has a derivative f (x) = −i
g (x), write
∞ −iht
f(x + h) − f(x) e −1
+ i
g (x) = + i e−ixt tf (t) dt .
h −∞ ht
Given ε > 0, choose R > 0 such that
−R ∞
ε
|tf (t)| dt + |tf (t)| dt < .
−∞ R 3
Then apply the inequality (28) to obtain

f(x + h) − f(x) ∞ −iht
−
e 1
+ ig (x) ≤ + i |tf (t)| dt
h −∞ ht
R −iht
e −1 2ε
< + i |tf (t)| dt + .
ht 3
−R
−iθ
d
But dθ e = −i at θ = 0, which says that for some δ > 0 the inequality
−iθ R
e −1 ε
+ i |tf (t)| dt <
θ 3
−R
holds for all θ with |θ| < δ. It follows that

f(x + h) − f(x) 2ε ε

g (x) <
+ i + =ε
h 3 3
when |h| < δ/R, which proves that f (x) = −i

g (x).
(d). The proof that f has Fourier transform ixf(x) will use integration
by parts. The hypothesis f ∈ L1 alone does not guarantee that f (t) → 0 as
|t| → ∞, but with the additional assumption that f ∈ L1 we can write
t ∞

f (t) = f (0) + f (s) ds → f (0) + f (s) ds as t → ∞ ,
0 0
by the fundamental theorem of calculus. Thus f (t) tends to a limit as

t → ∞, and that limit must be zero since f ∈ L1 . The same argument
shows that f (t) → 0 as t → −∞. Consequently, integration by parts gives
∞ ∞ ∞

f (x) = e −ixt
f (t) dt = e −ixt
f (t) + ix e−ixt f (t) dt = ixf(x) ,
−∞ −∞ −∞
as desired.
We are now prepared to calculate the Fourier transform of f (t) = e−at /2 ,

2
as given in Example 7. Since f (t) = −at e−at /2 is also integrable, Parts (c)
2
and (d) of the preceding theorem both apply and show that
ixf(x) = f (x) = −iaf (x) .
Thus f satisfies the differential equation af (x) + xf(x) = 0, which implies
that
d x2 /2a
ex /2a f(x) = C
2
e f (x) = 0 , and so
dx
for some constant C. But
∞ ∞

C = f (0) = e −at2 /2
dt = 2/a e−s ds = 2π/a ,
2
−∞ −∞

which shows that f(x) = 2π/a e −x2 /2a , as claimed. (For calculation of
the last integral, see Section 9.1.)
The convolution f ∗ g of two functions f, g ∈ L1 is defined by the integral
∞
(f ∗ g)(t) = f (s)g(t − s) ds .
−∞
It is easily seen that the integral converges absolutely if one of the functions
f or g is bounded. The same is true, by the Cauchy-Schwarz inequality, if
f and g are both in the space L2 . The operation of convolution behaves
much like multiplication of numbers. It is commutative and associative, and
distributive over addition:
f ∗g =g∗f, f ∗ (g ∗ h) = (f ∗ g) ∗ h , f ∗ (g + h) = f ∗ g + f ∗ h .
Perhaps the most important property of the Fourier transform is that it con-
verts convolution to ordinary pointwise multiplication: f ∗ g(x) = f(x) g(x).
For a proof, we can interchange the order of integration to write
∞ ∞

f ∗ g(x) = e −ixt
f (s)g(t − s) ds dt
−∞ −∞
∞ ∞
= e −ixs
f (s) e−ix(t−s) g(t − s) dt ds = f(x)
g (x) .
−∞ −∞
On the other hand, the operation of convolution does not admit a multi-
plicative identity element. In other words, there is no function g ∈ L1 with
the property that f ∗ g = f for every function f ∈ L1 . If there were such a
function, Fourier transformation would give in particular f(x) g (x) = f(x)
√
for the function f (t) = e −t 2 /2

, with transform f (x) = 2π e −x2 /2
= 0. But
that implies g (x) ≡ 1, contradicting the fact that g (x) → 0 as |x| → ∞,
as shown in the preceding theorem. To circumvent the difficulty, it is of-
ten useful to implement
∞ an “approximate identity”, a sequence of functions
gn ∈ L such that −∞ gn (t)dt = 1 and f ∗ gn → f in some sense for every
1
f ∈ L1 ∩ L∞ . Indeed, this will be our underlying strategy for inversion of a

Fourier transform.
8.10. Inversion of Fourier transforms

We propose to interpret and justify the formula
∞
1
f (t) = eixt f(x) dx ,
2π −∞
which recovers a function from its Fourier transform. This is the continuous
analogue of a Fourier series. As for Fourier series, however, the inversion
problem is by no means straightforward. First of all, the transform f need
not be integrable, as we have seen by examples, so in general the formula does
not make sense as stated. Even in cases where f is integrable, the formula
cannot be proved simply by interchanging the order of integration, since
a complex exponential is not integrable over the entire real line. In order
to deal with these problems, it will be necessary to introduce a truncating
factor to allow the interchange of integration, then to remove the effect of
truncation by passage to a limit. The process will involve the continuous
analogue of Cesàro summation.
For motivation, we begin by recalling the Cesàro means of an infi-
nite
∞ series and their relation to convergence. Given a formal infinite n series
k=−∞ αk of complex numbers, we form the partial sums sn = k=−n αk
and their Cesàro means
n
1 |k|
σn = (s0 + s1 + · · · + sn ) = 1− αk .
n+1 n+1
k=−n

If the series αk converges to a sum s, or equivalently if sn → s as n → ∞,
then the Cesàro means σn also converge to s as n → ∞. We will require a
continuous version of this fact, now stated as a lemma.
Lemma 1. Let α(t) be a complex-valued function that is Riemann integrable
over the real line. Then
R ∞
|t|
lim 1− α(t) dt = α(t) dt .
R→∞ −R R −∞
Proof. By symmetry, it suffices to show that

R ∞
t
lim 1− α(t) dt = α(t) dt .
R→∞ 0 R 0
Integrate by parts to obtain

R
t 1 R t
1− α(t) dt = A(t) dt , where A(t) = α(s) ds .
0 R R 0 0
Let I = limt→∞ A(t). Given ε > 0, choose M large enough that |A(t) − I| <
ε/2 for all t > M . Then for every R > M sufficiently large, we have
R M R
1 1 1
A(t) dt − I ≤ |A(t) − I| dt + |A(t) − I| dt < ε ,
R R 0 R M
0
which was to be proved.
Another lemma, a result on approximate identities under convolution,

will be the key to our proof of the inversion theorem for Fourier transforms.
It is stated here in general form.
∞
Lemma 2. Let g ∈ L1 with integral −∞ g(s)ds = 1, and define gR (s) =
R g(Rs) for R > 0. Let f ∈ L1 ∩ L∞ , and suppose that f is continuous
at some point t ∈ R. Then the convolution (f ∗ gR )(t) converges to f (t) as
R → ∞. If f is uniformly continuous on R, then (f ∗ gR )(t) converges to
f (t) uniformly for all t ∈ R.
∞ ∞
Proof. Note first that −∞ gR (s)ds = −∞ g(s)ds = 1. By hypothesis,
|f (s)| ≤ M for some constant M > 0 and all s ∈ R. Given ε > 0, choose
δ > 0 small enough that |f (t + s) − f (t)| < ε/2 if |s| < δ. Then write
∞ −δ δ ∞
(f ∗ gR )(t) − f (t) = [f (t − s) − f (t)] gR (s)ds = + +
−∞ −∞ −δ δ
and make the estimate

−δ δ ∞
ε
|(f ∗gR )(t)−f (t)| ≤ 2M |gR (s)| ds+ |gR (s)| ds+2M |gR (s)| ds .
−∞ 2 −δ δ
But
δ δ ∞
|gR (s)| ds = R |g(Rs)| ds ≤ |g(s)| ds < ∞ , and
−δ −δ −∞
∞ ∞
|gR (s)| ds = |g(s)| ds → 0 as R → ∞ .
δ δR
−δ
Similarly, −∞ |gR (s)| ds → 0 as R → ∞. It follows that |(f ∗gR )(t)−f (t)| ≤
ε for all R sufficiently large, and so (f ∗ gR )(t) → f (t) as R → ∞. If f is
uniformly continuous on the whole real line, the same argument shows that
the convergence is uniform.
Remark. If g is an even function and f is only piecewise continuous,

essentially the same argument shows that (f ∗ gR )(t) → 12 [f (t+) + f (t−)] for
every t ∈ R . A function f is said to be piecewise continuous on a bounded
interval if it is continuous there except perhaps for a finite number of jump-
discontinuities, where the right-hand and left-hand limits f (t+) and f (t−)
exist but are not equal. To say that a function is piecewise continuous on R
means that it is piecewise continuous on each bounded subinterval.
We turn now to the inversion formula for Fourier transforms.
Inversion Theorem. Let f ∈ L1 ∩ L∞ . Then

1 R
|x|
lim e ixt
1− f(x) dx = f (t)
R→∞ 2π −R R
for each point t ∈ R where f is continuous. If f is uniformly continuous on

R , the integrals converge to f (t) uniformly on R .
Proof. By the definition of the Fourier transform f(x), we have

R ∞
R
|x| |x|
e 1−
ixt
f (x) dx = e ixt
1− e−ixs f (s) ds dx
−R R −R R −∞
∞ R ∞
|x|
= f (s) e−ix(s−t) 1 − dx ds = f (s)gR (t − s) ds ,
−∞ −R R −∞
where
2
g(s) = (1 − cos s) and gR (s) = R g(Rs) .
s2
Here the interchange of the order of integration is justified by the absolute
integrability of the integrand, and Example 2 in the last section has been
applied to calculate the Fourier transform of the function 1 − |x|
R .
We are now in position to apply Lemma 2. Integration by parts gives
∞ ∞ ∞
1 1
g(s) ds = 2 2
(1 − cos s) ds = 4 (1 − cos s) ds
−∞ −∞ s 0 s2
∞
sin s
=4 ds = 2π .
0 s
(See Exercise 12 for calculation of the last integral.) Thus Lemma 2, ad-
justed by the factor 2π, shows that
∞
f (s)gR (t − s) ds = (f ∗ gR )(t) → 2πf (t) as R → ∞ ,
−∞
wherever f is continuous, and the convergence is uniform if f is uniformly

continuous on R .
The preceding statement about uniform convergence may be taken as

a continuous analogue of Fejér’s theorem for Fourier series, since it says
that the Cesàro means of the Fourier integral of f(x) converge uniformly to
f (t) when f is uniformly continuous. On the other hand, when the Fourier
transform is actually integrable, the inversion formula requires no special
interpretation.
Corollary. Let f ∈ L1 ∩L∞ and suppose that f ∈ L1 . Then if f is piecewise
continuous on R , it is uniformly continuous on R and
∞
1
(29) f (t) = eixt f(x) dx for every t ∈ R .
2π −∞
Proof. Since f ∈ L1 , the integral converges. By a slight generalization of

the inversion theorem (see the remark following the proof of Lemma 2), the
integral is Cesàro summable to 12 [f (t+) + f (t−)] for every t ∈ R . Hence by
Lemma 1 the integral converges to 12 [f (t+) + f (t−)] for every t ∈ R . But
the integral is the Fourier transform of a function f ∈ L1 , so as a function
of t it is uniformly continuous on R . Thus f (t+) = f (t−) = f (t) and f has
the representation (29).
The corollary says in particular that a piecewise continuous function

f ∈ L1 ∩ L∞ cannot have a transform f in L1 unless f is continuous.
Observe that in Examples 1 and 4 of the last section, the functions f are
piecewise continuous but not continuous, and indeed f ∈
/ L1 in both cases.
By appeal to the inversion theorem we can easily calculate the Fourier
transforms of Examples 5 and 6. Because in Examples 2 and 3 the functions
f are bounded and continuous, with transforms f ∈ L1 , the formulas of
Examples 5 and 6 follow at once from the corollary.
We conclude this section with a version of Parseval’s relation for Fourier
transforms. In general form this says that for functions f, g ∈ L1 ∩ L∞ and
continuous on R , with transforms f, g ∈ L1 , the formula
∞ ∞
2π f (t) g(t) dt = f(x)
g (x) dx
−∞ −∞
holds. Observe that L1 ∩ L∞ ⊂ L2 , and so the Cauchy–Schwarz inequality

ensures that both integrals converge. To verify the formula, we can apply
(29) to write
∞ ∞ ∞
2π f (t) g(t) dt = f (t) eixt
g (x) dx dt
−∞ −∞ −∞
∞ ∞ ∞
=
g (x) e −ixt
f (t) dt dx = f(x) g (x) dx .
−∞ −∞ −∞
In particular, if f is continuous on R and f ∈ L1 ∩ L∞ , and if f ∈ L1 , then

∞ ∞
2π |f (t)| dt =
2
|f(x)|2 dx .
−∞ −∞
It may be noted that the more general form of Parseval’s relation follows
easily from the particular form just stated, by means of the polarization
identity (cf. Exercise 28).
8.11. Poisson summation formula

This chapter concludes with an important relation between a function
and its Fourier transform. The Poisson summation formula resembles a
discrete version of Parseval’s relation and is very easy to prove, yet it has
profound applications. The leading idea is to form a “periodization” of a
given function on the real line and thus to introduce the discrete mechanism
of Fourier series. Here is a statement of the result.
Poisson Summation Formula. Let f ∈ L1 . Suppose that f is continuous

on R and that the infinite series
∞

g(t) = f (t + k)
k=−∞
converges uniformly in each bounded interval. Suppose also that the series
∞

f(2πn)
n=−∞
is absolutely convergent. Then

∞
∞

(30) f (n) = f(2πn) .
n=−∞ n=−∞
Proof. Observe first that under the hypothesis of uniform convergence the
sum g(t) is continuous and is periodic with period 1 ; that is, g(t + 1) = g(t)
for all t ∈ R . The Fourier coefficients of g, adapted to the change of scale,
are
1 ∞
1
−2πint
cn = g(t) e dt = f (t + k) e−2πint dt
0 k=−∞ 0
∞
k+1 ∞
= f (t) e−2πint dt = f (t) e−2πint dt = f(2πn) ,
k=−∞ k −∞
where the term-by-term integration is justified by uniform convergence of

the series f (t + k). But the hypothesis |f (2πn)| < ∞ implies that the
Fourier series
∞

f(2πn) e2πint .
n=−∞
converges absolutely and uniformly to a continuous sum h(t) for which h(t +
1) ≡ h(t). Since both h and g are continuous periodic functions with the
same Fourier coefficients, it follows that h(t) ≡ g(t). (See Exercise 21.)
Therefore,
∞
∞

(31) g(t) = f (t + n) = f(2πn) e2πint .
n=−∞ n=−∞
Now set t = 0 to obtain the Poisson summation formula (30). It may be

remarked that the formula (31) is really no more general than (30), since it
follows from (30) by a simple translation of the function f .
As a corollary, we see that if f(x) = 0 for all x outside the open interval
(−2π, 2π), then
∞ ∞
f (t) dt = f(0) = f (n) .
−∞ n=−∞
For a first application, consider Jacobi’s theta function

∞

e−πn t ,
2
ϑ(t) = t > 0.
n=−∞
For fixed t > 0 we can write

∞

where f (s) = e−πts .
2
ϑ(t) = f (n) ,
n=−∞
But by Example 7 in Section 8.9, the Fourier transform of f is

√
f(x) = (1/ t)e−x /4πt ,
2
and so
∞
∞
√ √

f (2πn) = (1/ t) e−πn /t = (1/ t) ϑ(1/t) .
2
n=−∞ n=−∞
Consequently, the remarkable identity
√
ϑ(1/t) = t ϑ(t) ,
known as Jacobi’s inversion formula, follows from the Poisson summation
formula.
For another application, let

sin πt 2
f (t) = .
πt
According to Example 5, this function has Fourier transform
|x|
1 − 2π for |x| ≤ 2π
f(x) =
0 for |x| > 2π .
In particular, f(2πn) = 0 for all n except n = 0, and f(0) = 1. Hence the
Poisson summation formula in the form (31) evaluates the sum
∞
sin π(n + t) 2
= 1, t ∈ R.
n=−∞
π(n + t)
But sin2 π(n + t) = sin2 πt for all n, so the expansion reduces to
∞
sin πt 2 1
= 1,
π n=−∞
(n + t)2
which implies that
∞
1 π 2
=
n=−∞
(n + t)2 sin πt
if t is not an integer. For t = 12 this says that
∞
1 π2
= ,
(2n + 1)2 8
n=0

which is equivalent to Euler’s sum ∞ 2 2
n=1 1/n = π /6.
The Poisson summation formula can also be applied to prove the sam-
pling theorem, a result of great importance in signal analysis. Here a signal
is understood to be a continuous function f (t) of time t, where −∞ < t < ∞.
It is said to be band-limited if its Fourier transform vanishes outside some
bounded interval. Band-limited signals are analogous to trigonometric poly-
nomials and have similar structural properties.
Sampling Theorem. Let f ∈ L1 ∩ L∞ be a continuous function with

Fourier transform f(x) = 0 for |x| ≥ π, and suppose that ∞ n=−∞ |f (n)| <
∞. Then f is determined by its values at the integers, and
∞
sin π(t − n)
(32) f (t) = f (n) , t ∈ R.
n=−∞
π(t − n)
The formula (32) is known to mathematicians as the cardinal series. A

more general form can be deduced by change of scale when f(x) = 0 for
|x| ≥ a, where a is an arbitrary positive number. When t is an integer, the
formula is to be interpreted in a limiting sense. Note that for any integer
m,
sin π(t − n) 0 for m = n
lim =
t→m π(t − n) 1 for m = n .
Although the sampling theorem can be obtained as a corollary of the
Poisson summation formula, it is simpler to give a direct proof. The fol-
lowing proof was found by Boas [2], who also derived the theorem from the
Poisson summation formula.
Proof of sampling theorem. Since f(x) = 0 for |x| ≥ π, and f(x) is
continuous, it is clear that f ∈ L1 . Thus the Fourier transform can be
inverted to write
π
1
f (t) = eitx f(x) dx , t ∈ R.
2π −π
In particular, the Fourier coefficients of the periodic extension of f are

π
1
cn = e−inx f(x) dx = f (−n) .
2π −π
Thus by hypothesis, f has an absolutely convergent Fourier series, which

therefore converges uniformly to f(x):
∞
∞

f(x) = cn einx = f (n)e−inx , −π ≤ x ≤ π .
n=−∞ n=−∞
In view of the uniform convergence, the series can be multiplied by the

exponential eitx and integrated term by term to obtain
π ∞ π
1 1
f (t) = eitx f(x) dx = f (n) ei(t−n)x dx
2π −π n=−∞
2π −π
∞ ∞
1 π sin π(t − n)
= f (n) cos(t − n)x dx = f (n) .
n=−∞
π 0 n=−∞
π(t − n)
The article by Higgins [5] may be consulted for further information about
the cardinal series. The book by Folland [3] offers a broad discussion of
Fourier transforms and their applications to problems of physics and engi-
neering.
Siméon-Denis Poisson (1781–1840) was a student at École Polytechnique
in Paris, where Laplace and Lagrange recognized his talent and promoted
his cause. At age 25 he became a professor there, the successor to Fourier.
He made important contributions to mathematical physics (electricity and
magnetism), celestial mechanics, and probability theory (the Poisson dis-
tribution). The Poisson summation formula was known to Poisson, among
others, as early as 1823.
Exercises

x = 2 sin x − 12 sin 2x + 13 sin 3x − . . . , −π < x < π ,
and conclude that 1 − 1

3 + 1
5 − 1
7 + ··· = π
4 .
2. Calculate the Fourier series

∞
π − x 2 π 2 cos nx
= + , 0 ≤ x ≤ 2π .
2 12 n2
n=1
Note the special cases
1 1 1 π2 1 1 1 π2
1+ + + + · · · = and 1− + − + · · · = .
22 32 42 6 22 32 42 12
3. Show that
∞
(−1)n
πt csc πt = 1 + 2t2 , t = 0, ±1, ±2, . . . .
t2 − n2
n=1
Suggestion. Set x = 0 in the Fourier expansion (15) of cos cx.
4. Show that
∞ ∞
2 4 cos 2nx 8 sin2 nx
| sin x| = − = , −∞ < x < ∞ .
π π 4n2 − 1 π 4n2 − 1
n=1 n=1
Hint. 1
2n−1 − 1
2n+1 = 2
4n2 −1
.
Exercises 237
5. (a) If a function f (x) is periodic with period 2π, and if f has a continuous
derivative of order k, show that the Fourier coefficients of f satisfy
|an | ≤ C/nk and |bn | ≤ C/nk
for some constant C > 0.

(b) If a periodic function f (x) has a continuous second derivative, show
that its Fourier series converges uniformly to f (x).
6. (a) Show that the Fourier coefficients of a continuous periodic piecewise

linear function f satisfy
|an | ≤ C/n2 and |bn | ≤ C/n2
for some constant C > 0. Conclude that the Fourier series converges uni-
formly to f (x).
(b) Show that every continuous periodic function can be approximated
uniformly by a continuous piecewise linear function with the same period.
(c) Deduce the trigonometric form of the Weierstrass approximation the-
orem.
7. Prove that π
4
|Dn (t)| dt ∼ log n , n → ∞,
−π π
where Dn (t) is the Dirichlet kernel.
8. Show that h(x) = sin x − x + x2 /π has the properties h(0) = h(π) = 0,

h (x) > 0 for 0 < x < π/2, and h (x) < 0 for π/2 < x < π. Conclude that
sin x x
>1− , 0 < x < π.
x π
9. (a) Show that the partial sums

n
sin kx
sn (x) =
k
k=1
of Example 3 in Section 8.5 have the property sn (x) > 0 for 0 < x < π.
Hint. Proceed by induction. Assuming that sn−1 (x) > 0 for 0 < x < π,
investigate the sum sn (x) at a local minimum.
(b) Show that |sn (x)| ≤ 1 + π/2 for all x ∈ R.
Suggestion. Use the Fejér means σn (x) as indicated at the end of Section
8.7.
10. (a) Derive the formula

n
sin 2(n + 1)x
cos(2k + 1)x = , n = 1, 2, . . . .
2 sin x
k=0
(b) For the Fourier series of the square wave function of Example 2,
show that the partial sums
4 1
n
s2n+1 (x) = sin(2k + 1)x
π 2k + 1
k=0
have the critical points
kπ
x= , k = 1, 2, . . . , 2n + 1 ,
2(n + 1)
in the interval (0, π), which are alternately local maxima and local minima.
(c) By a direct method similar to that of Section 8.6, show that

π 2 π sin x
lim s2n+1 = dx = 1.1789 . . . ,
n→∞ 2(n + 1) π 0 x
so that the partial sums exhibit Gibbs’ phenomenon.
11. For the Fourier series of the discontinuous sawtooth function of Exam-
ple 3, show that the partial sums exhibit an “undershoot” as well as the
“overshoot” of the Gibbs phenomenon. Specifically, for the first positive
local minimum λn = 2π/n, show that
2π
sin x π
lim sn (λn ) = dx = 1.4181 . . . = (0.9028 . . . ) .
n→∞ 0 x 2
12. Use the Dirichlet kernel to calculate the integral

∞
sin x π
dx = .
0 x 2
Writing

π π
1 1 π sin(n + 12 )x
1 − x
1
Dn (x) dx = sin(n + 2 )x dx + dx ,
0 0 2 sin 2 x 0 x
show that the first integral on the right-hand side tends to 0 as n → ∞.

Exercises 239
13. If f (x) is a continuous function of period 2π, except for a jump discon-
tinuity at a point x0 where it has one-sided limits f (x0 +) and f (x0 −), show
that its Fejér means have the property
lim σn (x) = 12 [f (x0 +) + f (x0 −)] .
n→∞
Here no smoothness hypothesis is imposed.

14. The function

n
! n (x) =
D sin kx
k=1
is known as the conjugate Dirichlet kernel. Verify the trigonometric identity
cos(k − 12 )x − cos(k + 12 )x = 2 sin 12 x sin kx
and use it to calculate
! n (x) = cos 2 x − cos(n + 2 )x .

1 1
D
2 sin 12 x
! n (x)| ≤ π/x for 0 < x ≤ π.
Conclude that |D
15. If bn+1 ≤ bn and bn → 0 as n → ∞, and the numbers nbn are bounded,
prove that the partial sums

n
sn (x) = bk sin kx
k=1
are uniformly bounded on 0 ≤ x ≤ 2π, hence on the entire real line.

Outline of proof. Suppose nbn ≤ M , n = 1, 2, . . . . Let N be the integer
defined by N ≤ π/x < N + 1. For n ≤ N , show that |sn (x)| ≤ πM . For
n > N , split the sum as

N
n
sn (x) = bk sin kx + bk sin kx
k=1 k=N +1
and apply Abel summation to the second sum to conclude that |sn (x)| ≤
(π + 2)M .
16. (a) If an+1 ≤ an and an → 0 as n → ∞, show that the series
∞

1
2 a0 + an cos nx
n=1
converges in the interval 0 < x < 2π and converges uniformly in δ ≤ x ≤

2π − δ for each δ > 0.
(b) Similarly, if bn+1 ≤ bn and bn → 0 as n → ∞, show that the series

∞

bn sin nx
n=1
converges for all x ∈ R and converges uniformly in δ ≤ x ≤ 2π − δ for each

δ > 0.
Hint. Use Abel summation.
(c) Integrate the Fourier series
∞
sin nx π−x
= , 0 < x < 2π ,
n 2
n=1
of Example 3 in Section 8.5 to calculate (cf. Exercise 2)

∞
2
cos nx x−π π2
= − .
n2 2 12
n=1
17. (a) By expanding the given function into Fourier series, show that
∞ 1
1 (π − 1)x , 0 ≤ x ≤ 1
sin n sin nx = 21
2 (π − x) , 1 ≤ x ≤ π.
n2
n=1
Conclude that
∞
∞

sin n sin n 2 π−1
= = .
n n 2
n=1 n=1
This observation is due to Robert Baillie [1].

(b) Apply the result of Exercise 2 (see also Exercise 16) to show that
∞
sin2 nx πx − x2
= , 0 ≤ x ≤ π,
n2 2
n=1
thus generalizing the result in Part (a), where x = 1.
18. Let f (θ) be a continuous function of period 2π, and let

∞

1
2 a0 + (an cos nθ + bn sin nθ)
n=1
be its formal Fourier series.

Exercises 241
(a) Show that the series

∞

u(r, θ) = 1
2 a0 + (an cos nθ + bn sin nθ)rn
n=1
converges uniformly in θ for each fixed r ∈ [0, 1).

(b) For the Poisson kernel
1 − r2
P (r, θ) = ,
1 − 2r cos θ + r2
verify the infinite series expansion
∞

P (r, θ) = 1 + 2 rn cos nθ , 0 ≤ r < 1.
n=1
Note that this formula can be viewed either as a Fourier expansion of P (r, θ)
for fixed r or as a Taylor series expansion for fixed θ. Suggestion: Show that

1 + reiθ
P (r, θ) = Re .
1 − reiθ
(c) Use integral expressions for the Fourier coefficients an and bn to es-
tablish the Poisson formula
2π
1
u(r, θ) = P (r, θ − t)f (t) dt , 0 ≤ r < 1.
2π 0
(d) Deduce from Fejér’s theorem that u(r, θ) → f (θ) uniformly in θ as

r → 1. (Compare with Exercise 17 in Chapter 6.) Interpret the result as a
statement that the Fourier series of a continuous periodic function f (θ) is
uniformly Abel summable to f (θ). (Apply a uniform version of Frobenius’
theorem in Chapter 7.)
19. (a) If ϕ(x) is a function of bounded variation over the interval [−π, π],
integrate by parts to show that its Fourier coefficients

1 π 1 π
an = cos nx ϕ(x) dx and bn = sin nx ϕ(x) dx
π −π π −π
satisfy an = O(1/n) and bn = O(1/n) as n → ∞.

(b) Invoke Littlewood’s Tauberian theorem to conclude that the Fourier
series of a function of bounded variation is convergent wherever it is Abel
summable.
20. (a) Suppose that a function f (x) is continuous and periodic with period
2π on the real line, and its Fourier coefficients satisfy an = O(1/n) and
bn = O(1/n) as n → ∞. Invoke Hardy’s Tauberian theorem to show that
the Fourier series converges to f (x) uniformly on R.
(b) Conclude that if f (x) is a continuous function of bounded variation
over [−π, π] with f (−π) = f (π), then the Fourier series converges to f (x)
uniformly in [−π, π].
21. Let f (x) be a continuous function with period 2π, and suppose that all
of its Fourier coefficients vanish:
π π π
f (x) dx = 0 and f (x) cos nx dx = f (x) sin nx dx = 0
−π −π −π
for n = 1, 2, 3, . . . . Show that f (x) ≡ 0.

Suggestion. Apply the trigonometric form of the Weierstrass approximation
theorem. Compare with Exercise 10 in Chapter 6.
22. Let f (x) be a continuous function with period 2π, and

∞suppose
that its
Fourier series is absolutely convergent in the sense that n=1 |an | + |bn | <
∞. Show that the Fourier series converges uniformly to the sum f (x). In
other words, show that sn (x) → f (x) uniformly on the real line as n → ∞.
23. Suppose that a function f (x) has a continuous derivative on the interval
2π
[0, 2π] and that f (0) = f (2π) and 0 f (x)dx = 0. Use Parseval’s relation
to show that
2π 2π
[f (x)]2 dx ≤ [f (x)]2 dx ,
0 0
with strict inequality unless f (x) = A cos x + B sin x for some constants A
and B. The result is known as Wirtinger’s inequality.
24. Apply Parseval’s relation to prove the isoperimetric inequality: Among

all simple closed curves of given length, the circle encloses the region of
largest area. Let C be a smooth simple closed curve of length 2π, represented
parametrically by (x, y) = (f (s), g(s)), 0 ≤ s ≤ 2π, where s is arclength.
Then the coordinate functions have Fourier series expansions
∞
a0
x = f (s) = + (an cos ns + bn sin ns) ,
2
n=1
∞

c0
y = g(s) = + (cn cos ns + dn sin ns) .
2
n=1
Exercises 243
(a) Apply Green’s theorem to show that the area of the region enclosed
by C is
∞

A= x dy = π n(an dn − bn cn ) .
C n=1
(b) Noting that (dx/ds)2 + (dy/ds)2 = 1, show by integration that
∞

π n2 (a2n + b2n + c2n + d2n ) = 2π .
n=1
(c) Combine the results of parts (a) and (b) to deduce that
∞
∞

2(π − A) = π (n2 − n)(a2n + b2n + c2n + d2n ) + π n[(an − dn )2 + (bn + cn )2 ] .
n=1 n=1
(d) Conclude that A < π unless a1 = d1 , b1 = −c1 , and an = bn = cn =

dn = 0 for every n > 1, so that C is a circle.
Note. This argument is due to Hurwitz [6]. It is closely related to the proof
of Wirtinger’s inequality suggested in Exercise 23.
25. Show that

√
1 1 1 3 3
1− 2 1− 1 − 2 ··· = and
3 62 9 2π

1 1 1 3
1− 2 1− 2
1 − 2 ··· = .
6 12 18 π
26. Refer to Example 3 in Section 8.9 for the Fourier transform of f (t) =
e−a|t| , where a > 0.
(a) Calculate the Fourier transform of f (t).
(b) Use the Parseval relation to calculate the integral
∞
x2 π
2 2 2 2
dx = , a > 0, b > 0.
−∞ (x + a )(x + b ) a+b
27. Calculate the Fourier transform of the function f (t) = sint t , and explain
your reasoning. Note that f is Riemann integrable but |f | is not. In other
words, f ∈
/ L1 .
28. For functions f, g ∈ L2 , let

∞
f, g = f (t)g(t) dt ,
−∞
and let f 2 = f, f . Verify the polarization identity
4f, g = f + g2 − f − g2 + if + ig2 − if − ig2 ,
and apply it to derive the general version 2πf, g = f,

g of Parseval’s
relation from the particular form where f = g.
29. Apply the Poisson summation formula to calculate the sum

∞
1 π 1 + e−2πt
= , t > 0.
n=−∞
n2 + t2 t 1 − e−2πt
∞ 2
Use the result to show that n=1 1/n = π 2 /6 .
Hint. Subtract 1/t2 from both sides of the equation and find the limit as
t → 0. Use the Taylor expansion of the exponential function.
30. Use Fourier transforms to find a solution of the ordinary differential

equation y + y = g(t) , where g is an arbitrary continuous function in L1 .
31. Derive the sampling theorem from the Poisson summation formula.
32. Generalize the sampling theorem by showing that if f(x) = 0 for |x| ≥ a,
then
∞
sin(at − nπ)
f (t) = f (nπ/a) .
n=−∞
at − nπ
References
[1] Robert Baillie, Solution to Advanced Problem 6241 (also proposed by R. Baillie),
[2] R. P. Boas, Jr., “Summation formulas and band-limited signals”, Tôhoku Math.
J. 24 (1972), 121–125.
[3] Gerald B. Folland, Fourier Analysis and Its Applications, American Mathemat-
ical Society, Providence, Rhode Island, 1992.
[4] Edwin Hewitt and Robert E. Hewitt, “The Gibbs–Wilbraham phenomenon:
An episode in Fourier analysis”, Archive for History of Exact Sciences 21 (1979),
129–160.
[5] J. R. Higgins, “Five short stories about the cardinal series”, Bull. Amer. Math.
Soc. 12 (1985), 45–89.
References 245
[6] A. Hurwitz, “Sur le problème des isopérimètres”, Comptes Rendus Acad. Sci.
Paris 132 (1901), 401–403.
[7] Dunham Jackson, Fourier Series and Orthogonal Polynomials, Mathematical
Association of America, Washington, D.C., Carus Monograph No. 6, 1941; reprinted
by Dover Publications, Mineola, NY, 2004.
[8 ] T.W. Körner, Fourier Analysis, Cambridge University Press, Cambridge, U.K.,
1988.
[9] W. Rogosinski, Fourier Series, Chelsea Publishing Co., New York, 1950. [Eng-
lish translation of Fouriersche Reihen, Sammlung Göschen, W. de Gruyter, Berlin,
1930.]
[10] David V. Widder, Advanced Calculus, Second edition, Prentice–Hall, Engle-
wood Cliffs, N.J., 1961; reprinted by Dover Publications, Mineola, NY, 1989.
[11] A. Zygmund, Trigonometric Series, Second edition, Cambridge University
Press, Cambridge, U.K., 1959.
Chapter 9
The Gamma Function
The gamma function is a continuous extension of the factorial function.

It is an important function with many remarkable properties. It can be
represented as an integral and as an infinite product. In this chapter we
develop the basic theory of the gamma function and its relation to the beta
function. By way of background we begin with an integral that plays a
central role in probability theory.
9.1. Probability integral

The integral ∞ √
e−x dx =
2
π
−∞
is important in probability theory because of its connection with the Gauss-
ian or normal distribution. It is evaluated by a well known device that makes
use of area integrals. Since the integrand is an even function, it is equivalent
to show that
∞ √
π
e−x dx =
2
(1) .
0 2
To evaluate the integral in (1), we consider the function
f (x, y) = e−x e−y

2 2
and the three regions of integration shown in Figure 1. With the√notation

D1 and D2 for for the quarter-disks with respective radii R and 2R, and
247
248 9. The Gamma Function
x
R
2R
Figure 1. Three regions of integration.
S for the square with side-length R, we observe that D1 ⊂ S ⊂ D2 . Since

f (x, y) > 0, it is therefore clear that

f (x, y) dA ≤ f (x, y) dA ≤ f (x, y) dA .
D1 S D2
Now introduce polar coordinates in the regions D1 and D2 through the

substitutions x = r cos θ and y = r sin θ, and iterate all three integrals to
conclude that
√
π/2 R R R π/2 2R
−r 2 −x2 −y 2
e−r r dr dθ .
2
e r dr dθ ≤ e e dx dy ≤
0 0 0 0 0 0
The advantage of the area-integral approach is that the polar-coordinate

integrals can be calculated explicitly. These calculations reduce the inequal-
ities to
π R 2
π
−R2 −x2 −2R2
1−e ≤ e dx ≤ 1−e .
4 0 4
Now let R → ∞ to obtain

∞ 2
π
e−x dx
2
= ,
0 4
which is equivalent to (1).

9.2. Gamma function 249
9.2. Gamma function

In the early development of calculus, there were various efforts to analyze
the factorial function n! = 1 · 2 · 3 · · · n and the related binomial coefficients.
Stirling’s formula √
n! ∼ nn e−n 2πn , n → ∞,
was published in 1730. Around the same time, Euler had been searching
for a natural extension of the factorial function. The problem was to find
a function f (x), defined by an analytic formula for all x > 0, to interpolate
the values f (n) = n! at the positive integers. In 1729, Euler sketched a
suitable construction in a letter to Christian Goldbach (1690–1764), who
had proposed the problem along with Daniel Bernoulli (1700–1784), the son
of Euler’s teacher Johann Bernoulli. In fact, this letter was the first in a
long series of mathematical correspondence between Euler and Goldbach,
a very talented amateur mathematician who asked thoughtful questions.
(“Goldbach’s conjecture”, that every even number larger than 2 is the sum
of two primes, remains unsettled.)
Euler published the details of his construction, but soon found that his
solution could be expressed in the equivalent integral form
1
f (x) = (− log t)x dt .
0
Focusing on the integral representation, he then studied the function and
developed its basic properties. Only later did Adrien-Marie Legendre (1752–
1833) propose the notation Γ(x) and the now standard definition of the
gamma function:
∞
(2) Γ(x) = e−t tx−1 dt , x > 0.
0
A change of variables shows that Γ(x + 1) is the same as Euler’s function
f (x).
One may well wonder what is so special about the gamma function.
There are many functions g(x), infinitely differentiable for x > 0, with the
property g(n) = n! for all integers n = 1, 2, . . . . Why does the gamma
function provide the “right” extension of the factorial? One response is that
the gamma function not only has the values Γ(n + 1) = n! , but as we shall
see presently, it also satisfies the functional equation Γ(x + 1) = xΓ(x) for
all x > 0. Further justification comes from the Bohr–Mollerup theorem, to
be discussed later, but the real case for the gamma function is its frequent
appearance in formulas of mathematical analysis.
Integration by parts shows that
∞ ∞ ∞
−t x −t x
Γ(x + 1) = e t dt = −e t +x e−t tx−1 dt = xΓ(x) .
0 0 0
x
1 2 3
Figure 2. Graph of y = Γ(x).
Since Γ(1) = 1, the formula Γ(x + 1) = xΓ(x) can be iterated to give

Γ(2) = 1, Γ(3) = 2, and in general Γ(n + 1) = n! for n = 1, 2, . . . . It is clear
that Γ(x) > 0 for all x > 0, because the integrand is positive. A graph of
the gamma function is shown in Figure 2.
The graph suggests that Γ(x) descends to a minimum value at a point
x0 in the interval 1 < x0 < 2, and is increasing for x > x0 . It is not
difficult to verify this (see Exercise 3). Gauss was interested in the number
x0 = 1.4616 . . . and calculated it to many decimal places. He also calculated
Γ(x0 ) = 0.8856 . . . . Another feature suggested by the graph is that Γ(x) is a
convex function. In fact, we will see later that log Γ(x) is convex, a stronger
property. The graph of log Γ(x) is displayed in Figure 3.
y
2
x
1 2 3
Figure 3. Graph of y = log Γ(x).

9.3. Beta function 251
With the substitution t = u2 , the gamma function takes the form

∞
e−u u2x−1 du ,
2
(3) Γ(x) = 2 x > 0.
0
√
In view of the formula (1), this shows that Γ( 12 ) = π.
The functional equation Γ(x + 1) = xΓ(x) can be iterated to show more
generally that
Γ(x + n) = (x + n − 1)(x + n − 2) · · · (x + 1)xΓ(x) , x > 0,
for n = 1, 2, . . . , or
Γ(x + n)
Γ(x) = , x > 0.
x(x + 1) · · · (x + n − 1)
The last formula allows the definition of Γ(x) to be extended to the interval
x > −n, provided that x is not a negative integer or 0. Since n is an
arbitrary positive integer, this extends the definition of Γ(x) to the entire
real line, excluding the singular points x = 0, −1, −2, . . . .
9.3. Beta function

The beta function B(x, y) is defined by
1
(4) B(x, y) = tx−1 (1 − t)y−1 dt , x > 0, y > 0.
0
The change of variables t = 1 − s reveals the symmetry property B(x, y) =

B(y, x). However, the most important property of the beta function is its
expression in terms of the gamma function:
Γ(x)Γ(y)
(5) B(x, y) = .
Γ(x + y)
The proof of (5) makes use of the alternate expression

π/2
(6) B(x, y) = 2 (cos θ)2x−1 (sin θ)2y−1 dθ ,
0
which results from the substitution t = cos2 θ. The argument is very similar
to the calculation of the probability integral (1). The proof begins with a
comparison of integrals of the positive function
g(u, v) = u2x−1 v 2y−1 e−(u

2 +v 2 )
over the same three regions as before, as displayed in Figure 1. Here, how-
ever, x and y are fixed positive parameters, whereas u and v are the variables
of integration. After passing to polar coordinates with the substitutions
u = r cos θ and v = r sin θ, we arrive at the formulas
π/2 R
e−r r2x+2y−1 dr
2x−1 2y−1 2
(cos θ) (sin θ) dθ
0 0
R R
e−u u2x−1 du e−v v 2y−1 dv
2 2
≤
0 0
√
π/2 2R
e−r r2x+2y−1 dr .
2
≤ (cos θ)2x−1 (sin θ)2y−1 dθ
0 0
Letting R → ∞ and referring to (3) and (6), we conclude that
B(x, y)Γ(x + y) ≤ Γ(x)Γ(y) ≤ B(x, y)Γ(x + y) ,
and (5) follows. Another proof is outlined in the exercises.
9.4. Legendre’s duplication formula

Legendre’s duplication formula for the gamma function expresses Γ(2x)
in terms of Γ(x) and Γ(x + 12 ). It is
√
(7) π Γ(2x) = 22x−1 Γ(x) Γ(x + 12 ) , x > 0.
Many proofs have been found. We will give a particularly elementary proof
that uses the basic connection (5) between the beta and gamma functions,
plus the special formula
(8) B(x, x) = 21−2x B(x, 12 ) , x > 0.
To prove (8), observe first that because of the symmetry of the integrand
in (4) when x = y, we can write
1/2

x−1
B(x, x) = 2 t(1 − t) dt , x > 0.
0
Next observe that u = 4t(1 − t) increases from 0 to 1 as t increases from 0

to 12 . Make this substitution in the integral, noting that
du = 4(1 − 2t) dt and 1 − u = (1 − 2t)2 ,
to arrive at the formula

1
B(x, x) = 2 1−2x
ux−1 (1 − u)−1/2 du = 21−2x B(x, 12 ) ,
0
9.5. Euler’s reflection formula 253
which proves (8).

Legendre’s formula (7) is now deduced from (8) through the relation (5).
When (5) is substituted into (8), the formula reduces to
Γ(x)2 Γ(x) Γ( 12 )
= 21−2x .
Γ(2x) Γ(x + 12 )
√
Since Γ( 12 ) = π, this is the Legendre duplication formula.
It may be observed that the special formula (8) is actually a disguised
expression of Legendre’s formula.
The name of Adrien-Marie Legendre (1752–1833) occurs frequently in
topics of classical analysis. In Chapter 13 of this book we will encounter
the Legendre differential equation and the associated Legendre polynomi-
als, which are orthogonal over the interval [−1, 1]. Chapter 14 contains a
discussion of the Legendre relation for elliptic integrals, a field that Legendre
cultivated for many years. Yet some of Legendre’s greatest contributions lay
in celestial mechanics and in number theory. For instance, he developed and
applied (before Gauss) the method of least squares for determination of as-
tronomical orbits. He formulated the prime number theorem (cf. Chapter
10), the theorem that every arithmetic progression without common factor
contains infinitely many primes (proved later by Dirichlet), and the law of
quadratic reciprocity, verified by Gauss after Legendre had found a partial
proof. Little is known about Legendre’s personal life. A portrait tradition-
ally believed to be that of Adrien-Marie Legendre has now been discredited
and replaced by a recently discovered caricature portrait, the only true im-
age of Legendre known to exist. An article by Duren [5] tells the story of
the two portraits.
9.5. Euler’s reflection formula

The important relation
π
(9) Γ(x)Γ(1 − x) = , 0 < x < 1,
sin πx
known as Euler’s reflection formula, links the gamma function with the sine
function. It lies deeper than other properties of the gamma function, and
there is no simple derivation. We will base a proof on the formula
∞

2c 1 (−1)n
(10) cos cx = sin cπ + cos nx , −π ≤ x ≤ π ,
π 2c2 c2 − n2
n=1
which we derived in Section 8.5 from the convergence theorem for Fourier
series. Set x = 0 in (10) to obtain the expression
∞

π 1 (−1)n
(11) = 2c + , c = 0, ±1, ±2, . . . .
sin cπ 2c2 c2 − n2
n=1
On the other hand, since Γ(1) = 1, the formula

Γ(x)Γ(y)
B(x, y) = , x > 0, y > 0
Γ(x + y)
for the beta function gives
1 ∞
c−1 −c xc−1
Γ(c)Γ(1−c) = B(c, 1−c) = t (1−t) dt = dx , 0 < c < 1,
0 0 1+x
t
with the substitution x = 1−t . But the substitution u = 1/x reduces part
of the last integral to
∞ c−1 1 −c
x u
dx = du ,
1 1+x 0 1+u
so we arrive at the expression

1 c−1 1 −c
x x
Γ(c)Γ(1 − c) = dx + dx , 0 < c < 1.
0 1 + x 0 1 +x
Now introduce the identity
1 x
=1−
1+x 1+x
into the first integral to obtain
1 1 −c
x − xc
Γ(c)Γ(1 − c) = xc−1 dx + dx
0 0 1+x
1
1
= + (x−c − xc )(1 − x + x2 − x3 + . . . ) dx
c 0
∞ 1
1
= + (−1) n
(x−c − xc )xn dx
c 0
n=0
∞
1 2c π
= + (−1)n = , 0 < c < 1,
c c2 −n 2 sin cπ
n=1
in view of (11). The term-by-term integration is justified by showing that

the tail of the integrated series tends to zero:
1 n+1
x
(12) lim (x−c − xc ) dx = 0 .
n→∞ 0 1 + x
Indeed, the factor

x1−c − x1+c
1+x
9.6. Infinite product representation 255
is bounded in the interval [0, 1], while

1
1
xn dx = →0 as n → ∞ .
0 n+1
This proves (12) and completes the completes the proof of the Euler reflec-
tion formula (9).
√
With the choice x = 12 , the reflection formula says that Γ( 12 ) = π.
Another proof of the reflection formula can be based on the infinite product
representation of the gamma function, as developed in the next section (see
Exercise 16).
9.6. Infinite product representation

Before turning to a representation of the gamma function as an infinite
product, we will derive the formula
n! nx
(13) Γ(x) = lim , x > 0.
n→∞ x(x + 1) · · · (x + n)
The existence of the limit is not obvious and is part of the assertion. The
result is called the Gauss product formula after Carl Friedrich Gauss (1777–
1855), although it was known to Euler. In fact, this was Euler’s original
construction in 1729, but Gauss rediscovered the formula and recognized its
importance.
One approach to (13) is to observe that
1
Γ(x)Γ(n + 1)
tx−1 (1 − t)n dt = B(x, n + 1) =
0 Γ(x + n + 1)
n! Γ(x) n!
= = .
(x + n)(x + n − 1) · · · xΓ(x) x(x + 1) . . . (x + n)
The substitution of t/n for t then leads to the representation
n ∞
n! nx t n
= t x−1
1− dt = gn (t)tx−1 dt ,
x(x + 1) · · · (x + n) 0 n 0
where n
1 − nt , for 0 ≤ t ≤ n
gn (t) =
0, for n < t < ∞ .
−t
Because gn (t) increases to the limit e as n → ∞, Dini’s theorem (see
Section 1.8) ensures that the integrals converge and
∞
n! nx
lim = e−t tx−1 dt = Γ(x) .
n→∞ x(x + 1) · · · (x + n) 0
This concludes the proof of the Gauss product formula.
It is now a short step to the infinite product representation of the gamma
function.
Theorem. The formula
∞
1 γx
x −x/n
= xe 1+ e
Γ(x) n
n=1
holds for all x ∈ R with x = 0, −1, −2, . . . . Here

n
1
γ = lim − log n
n→∞ k
k=1
is Euler’s constant.
The infinite product converges for each x ∈ R because

x −x/n x x x2
1+ e = 1+ 1 − + 2 + ···
n n n 2n

x2 1
=1− 2 +O
2n n3

and the series 1/n2 converges. In fact, the product can be shown to
converge uniformly on each bounded subset of R, so that it represents a
continuous function on R. Note that the product is equal to zero at the
points x = 0, −1, −2, . . . where the gamma function is infinite.
Proof of theorem. The Gauss product formula (13) allows us to write
1 x(x + 1) · · · (x + n)
= lim
Γ(x) n→∞ n! nx
x x x −x log n
= lim x 1 + 1+ ··· 1 + e
n→∞ 1 2 n
n
x −x/k
= lim xex(1+ 2 +···+ n −log n)
1 1
1+ e
n→∞ k
k=1
∞
x
= xeγx 1+ e−x/k .
k
k=1
The proof shows that the Gauss product formula (13) is essentially the
same as the infinite product representation of 1/Γ(x).
9.7. Generalization of Stirling’s formula

Implicit in the preceding calculations is an important generalization of
Stirling’s formula
√
2π nn+ 2 e−n ,
1
n! ∼ n → ∞,
n
to the gamma function. Because Γ(n) = (n − 1)! and 1 − n1 → 1/e , an
equivalent form is
√
2π nn− 2 e−n ,
1
Γ(n) ∼ n → ∞.
But the result can be strengthened to

√
2π xx− 2 e−x ,
1
(14) Γ(x) ∼ x → ∞,
where x → ∞ unrestrictedly. This is Stirling’s formula for the gamma

function.
Stirling’s formula (14) is easily deduced from the relation (13), which
can be rewritten as
Γ(n + a + 1)
lim = 1, a > 0,
n→∞ n! na
where the convergence is uniform for 0 < a ≤ 1, as is evident from the proof
of (13). In view of Stirling’s formula for n!, the last limit implies that
Γ(n + a) √
lim = 2π ,
n+a− 21
n→∞ n e−n
or
Γ(n + a) √
lim = 2π , 0 < a ≤ 1,
n+a− 21
n→∞ (n + a) e−(n+a)
n
since 1 + na → ea uniformly in the interval 0 < a ≤ 1. This proves
Stirling’s formula for the gamma function.
9.8. Bohr–Mollerup theorem

Finally, we come to the remarkable fact that the gamma function is
completely determined by the relation Γ(x + 1) = xΓ(x), the normalization
Γ(1) = 1, and the fact that log Γ(x) is a convex function. Recall that a
function f defined in an interval I ⊂ R is convex if for each pair of points
a, b ∈ I it satisfies the inequality
f (ra + (1 − r)b) ≤ rf (a) + (1 − r)f (b) , 0 < r < 1.

Geometrically, this says that the graph of f lies beneath every chord.
We begin by proving that log Γ(x) is convex, a property evident from its
graph in Figure 3. Suppose 0 < a < b and 0 < r < 1. Then by definition
∞
Γ(ra + (1 − r)b) = e−t tra+(1−r)b−1 dt
∞
0
−t a−1 r −t b−1 1−r
= e t e t dt
0
∞ r ∞ 1−r
−t a−1 −t b−1
≤ e t dt e t dt = Γ(a)r Γ(b)1−r ,
0 0
where Hölder’s inequality has been applied with the conjugate indices p =
1/r and q = 1/(1 − r). Taking logarithms, we conclude that
log Γ(ra + (1 − r)b) ≤ r log Γ(a) + (1 − r) log Γ(b) , 0 < r < 1,
which shows that log Γ(x) is a convex function. (See also Exercise 26.)
Harald Bohr and Johannes Mollerup [3] discovered that the gamma func-
tion is actually characterized by its logarithmic convexity. Emil Artin [2]
gave an elegant presentation of their argument and clarified the role of log-
arithmic convexity.
Bohr–Mollerup Theorem. Let G(x) be a positive function on the pos-

itive real axis x > 0 with the properties G(x + 1) = xG(x) and G(1) = 1.
Suppose further that log G(x) is a convex function. Then G(x) ≡ Γ(x).
Proof. The hypotheses G(x + 1) = xG(x) and G(1) = 1 imply that G(n +
1) = n! for n = 1, 2, . . . . For any positive integer n and for 0 < x ≤ 1, we
express
n + x = (1 − x)n + x(n + 1)
as a convex combination of n and n + 1. Then by the convexity hypothesis,
log G(n + x) ≤ (1 − x) log G(n) + x log G(n + 1) ,
or
G(n + x) ≤ G(n)1−x G(n + 1)x = n! nx−1 .
In a similar way, the convex combination
n + 1 = x(n + x) + (1 − x)(n + x + 1)
produces the inequality
n! = G(n + 1) ≤ G(n + x)x G(n + x + 1)1−x = G(n + x)(n + x)1−x ,

since G(n + x + 1) = (n + x)G(n + x). When the two inequalities are

combined, we find that
(15) n!(n + x)x−1 ≤ G(n + x) ≤ n! nx−1 .
But the property G(x + 1) = xG(x) can be iterated to give
G(x + n) = x(x + 1) · · · (x + n − 1)G(x) ,
so the inequality (15) takes the form

n!(n + x)x n! nx
(16) ≤ G(x) ≤ .
x(x + 1) · · · (x + n) x(x + 1) · · · (x + n − 1)n
We now let n → ∞ and appeal to the Gauss product formula (13) to conclude
from (16) that
Γ(x) ≤ G(x) ≤ Γ(x) ,
or G(x) = Γ(x) for 0 < x ≤ 1. The extension to all x > 0 now results from
the two identities G(x + 1) = xG(x) and Γ(x + 1) = xΓ(x).
Alternatively, the inequality (16) can be used to establish the existence

of the limit in (13) and to show that it is equal to G(x). Since the reasoning
applies equally well to Γ(x), it then follows that G(x) = Γ(x), which gives
an independent proof of the Gauss product formula.
Aside from its aesthetic appeal, the Bohr–Mollerup theorem offers an
effective way to verify formulas involving the gamma function. As an illus-
tration, we now apply it to establish the Gauss multiplication formula
(17)

mmx−1/2 1 2 m−1
Γ(mx) = Γ(x) Γ x + Γ x + · · · Γ x +
(2π)(m−1)/2 m m m
for x > 0 and m = 2, 3, . . . . This is a generalization of Legendre’s duplication
formula (7), which corresponds to the special case m = 2. Replacing mx by
x in (17), we want to show that the function
mx−1/2 x x + 1
x+m−1

G(x) = Γ Γ ···Γ
(2π)(m−1)/2 m m m
is equal to Γ(x). In view of the Bohr–Mollerup theorem, it will suffice to
show that log G(x) is convex, G(x + 1) = xG(x), and G(1) = 1. The
logarithmic convexity follows at once from that of the gamma function. To
see that G(x+1) = xG(x), it is convenient to write G(x) = αm gm (x), where
m−1/2
αm =
(2π)(m−1)/2
and x x + 1
x+m−1

gm (x) = mx Γ Γ ···Γ .
m m m
Now observe that

x+m x
gm (x + 1) = m Γ Γ gm (x) = xgm (x) ,
m m
x x
since Γ x+m
m = Γ m + 1 = m
x
Γ m . Thus G(x + 1) = xG(x).
The difficulty is to show that G(1) = 1. In other words, it is to be shown

that
m
1 2
(18) αm m Γ Γ ···Γ = 1.
m m m
For this purpose we apply the Gauss product formula (13) to write

k n! nk/m mn+1
Γ = lim , k = 1, 2, . . . , m .
m n→∞ k(k + m)(k + 2m) · · · (k + nm)
The product of these limits is

m
1 2 (n!)m n(m+1)/2 mmn+m
(19) Γ Γ ···Γ = lim .
m m m n→∞ (m + nm)!
But
(m + nm)! 1 2 m
= 1+ 1+ ··· 1 + →1
(nm)!(nm)m nm nm nm
as n → ∞, and so the expression (19) reduces to
m
1 2 (n!)m mmn
Γ Γ ···Γ = lim .
m m m n→∞ (nm)! n(m−1)/2
√
Now Stirling’s formula n! ∼ nn e−n 2πn can be applied to give
(n!)m (2π)(m−1)/2 n(m−1)/2

∼ as n → ∞ ,
(nm)! mmn+1/2
so that
1 2 m (2π)(m−1)/2 1
Γ Γ ···Γ = 1/2
= .
m m m m mαm
This proves (18), which shows that G(1) = 1. Hence it follows from the
Bohr–Mollerup theorem that G(x) = Γ(x), which completes the proof of the
Gauss multiplication formula.
9.9. A special integral 261
9.9. A special integral

We close this chapter with a calculation of the integral
∞
π
(20) x−a sin x dx = csc(πa/2) , 0 < a < 2.
0 2 Γ(a)
This result will be applied in the next chapter to derive a functional equation
for the Riemann zeta function. Although the integral can be calculated by
contour integration in the complex plane, the method used here is entirely
elementary and is based only on standard properties of the gamma function.
Recall first the expression for the Laplace transform
∞
1
(21) F (s) = e−st sin t dt = , s > 0,
0 1 + s2
which can be derived via two integrations by parts. Specifically,

1 ∞ −st 1 1 1 ∞ −st
F (s) = e cos t dt = − e sin t dt ,
s 0 s s s 0
so that (1 + s2 )F (s) = 1.
Next make a change of variables to obtain
∞ ∞
Γ(a) = ta−1 e−t dt = xa ta−1 e−xt dt , x > 0, a > 0.
0 0
Take ε > 0 and use this form of the gamma function to write
∞ ∞ ∞
1
e−εx x−a sin x dx = e−εx sin x ta−1 e−xt dt dx
0 Γ(a)
0 ∞ ∞ 0
1
= ta−1 e−(ε+t)x sin x dx dt
Γ(a) 0 0
∞
1 1
= ta−1 dt ,
Γ(a) 0 1 + (ε + t)2
where the formula (21) has been invoked. To justify the interchange in order
of integration, observe that the exponential factor e−εx makes the integrand
absolutely integrable over the first quadrant of the (x, t) plane.
Now appeal to the integral analogue of Abel’s theorem (Section 3.2) to
conclude that
∞ ∞
−a
x sin x dx = lim e−εx x−a sin x dx
0 ε→0 0
∞ a−1
1 t
= dt .
Γ(a) 0 1 + t2
But the last integral is a beta function in disguise. Let u = t2 and refer to
Exercise 13 to write
∞ a−1
1 ∞ u 2 −1
a
t 1
2
dt = du = B( a2 , 1 − a2 ) .
0 1+t 2 0 1+u 2
Then apply the Euler reflection formula to obtain
B( a2 , 1 − a2 ) = Γ( a2 ) Γ(1 − a2 ) = π csc(πa/2) ,
which verifies the formula (20).

The expression (20) can be recast in other forms. For 0 < a < 1 it
combines with the Euler reflection formula to give
∞
π
xa−1 sin x dx = csc(π(1 − a)/2)
0 2 Γ(1 − a)
1 sin πa
= Γ(a) = Γ(a) sin(πa/2) .
2 cos(πa/2)
More generally,
∞
xa−1 sin bx dx = b−a Γ(a) sin(πa/2) , 0 < a < 1, b > 0.
0
Exercises
1. Prove that Γ(x) is continuous for x > 0.
2. Show that limx→0+ xΓ(x) = 1 .
3. (a) Prove that the function Γ(x) has derivatives of all orders at every
point x > 0, and that its nth derivative is given by
∞
Γ(n) (x) = e−t tx−1 (log t)n dt , n = 1, 2, . . . .
0
(b) Show that Γ (x) > 0 for x > 0, so that the curve y = Γ(x) is convex.
(c) Show that Γ(x) attains a minimum value for x > 0 at a point x0 in
the interval 1 < x0 < 2. Show further that Γ(x) is decreasing in the interval
(0, x0 ) and increasing in (x0 , ∞).
4. Show that
√
(2n)! π
Γ n + 12 = , n = 1, 2, . . . .
4n n!
Exercises 263
5. Verify that Euler’s definition

1 x−1
Γ(x) = log(1/t) dt
0
is equivalent to the definition (2). Show more generally that
1
y−1
tx−1 log(1/t) dt = x−y Γ(y) , x > 0, y > 0.
0
6. Calculate the integrals
1 √
1/2
π
(a) log(1/t) dt =
0 2
1
−1/2 √
(b) log(1/t) dt = π
0
1 √
1/2 2π
(c) t log(1/t) dt = √
0 3 3
1 √
−1/2
(d) t log(1/t) dt = 2π .
0
1
π
t(1 − t) dt = .
0 8
8. Express the function
π/2
f (x) = (sin 2θ)x dθ , x > −1 ,
0
in terms of the gamma function.
9. Show that √
3
Γ( 16 ) = √
3 √ Γ( 13 )2 .
2 π
10. Calculate the integrals
1 √
3
(a) 1 − t3 dt = √3
Γ( 13 )3
0 10π 2
1
1
(b) 1 − t4 dt = √ Γ( 14 )2
0 6 2π
1
1 1
(c) √ dt = √ √ Γ( 13 )3
0 1 − t 3 2π 3 3
2
1
1 1
(d) √ dt = √ Γ( 14 )2
0 1−t 4 4 2π
1 2
t 1
(e) √ dt = √ Γ( 34 )2 .
0 1−t 4 2π
11. Use the beta function to calculate

π/2 √ π/2 √ √
2
sin θ dθ = cos θ dθ = √ Γ( 34 )2 .
0 0 π
Compare the result with that of Exercise 10(e).
12. Find the area of the region bounded by the hypocycloid x2/3 + y 2/3 = 1.
u
13. Use the substitution t = 1+uto derive the formula
∞
ux−1
B(x, y) = du .
0 (1 + u)x+y
14. Apply the formula of the preceding exercise to calculate the integral
∞
x3 1
7
dx = .
0 (1 + x) 60
∞ −a
x π
dx = , 0 < a < 1.
0 1+x sin πa
More generally, calculate
∞ m−1
x π
n
dx = , 0 < m < n,
0 1+x n sin(mπ/n)
a result known to Euler. (Here m and n need not be integers.)
16. Compare the infinite product representation of Γ(x) with that of the
sine function to obtain another proof of the Euler reflection formula.
17. Prove that Γ (1) = −γ, where γ is Euler’s constant. Conclude that
∞
γ=− e−t log t dt .
0
Hint. Take logarithmic derivatives in the infinite product formula for Γ(x).
Justify the term-by-term differentiation.
∞ ∞

log x log n 1
x
dx = −γ log 2 + (−1)n = − (log 2)2 .
0 1+e n 2
n=1
Hint. Expand 1/(1 + ex ) = e−x /(1 + e−x ) into geometric series and integrate
term by term (justify), then refer to Exercise 13 of Chapter 2 for evaluation
of the infinite series.
Exercises 265
19. The Laplace transform of a function f is the function F = L(f ) defined

by the integral ∞
F (s) = f (t)e−st dt , s > s0 ,
0
which is assumed to converge for all real numbers s larger than some number
s0 . For each exponent a > −1, show that the Laplace transform of f (t) = ta
is
Γ(a + 1)
F (s) = , s > 0.
sa+1
20. The convolution of two functions f and g defined on (0, ∞) is the

function h = f ∗ g given by
t
h(t) = f (u)g(t − u) du , 0 < t < ∞.
0
Show that L(f ∗ g) = L(f )L(g). In other words, show that the Laplace
transform of a convolution is the product of transforms.
21. If f (t) = tx−1 and g(t) = ty−1 for some x > 0 and y > 0, show that the
convolution h = f ∗ g has the form h(t) = tx+y−1 B(x, y). By taking Laplace
transforms, conclude that
Γ(x) Γ(y)
B(x, y) = .
Γ(x + y)
22. Use Stirling’s formula for the gamma function to show that
Γ(n + c − a) Γ(n + c − b)
lim =1
n→∞ Γ(n + c) Γ(n + c − a − b)
for arbitrary real numbers a, b, and c.

∞
π
x−a cos x dx = sec(πa/2) , 0 < a < 1.
0 2 Γ(a)
24. (a) Calculate the integral

1
22n+1 (n!)2
(1 − x2 )n dx = , n = 1, 2, . . . ,
−1 (2n + 1)!
by making the substitution t = 1 − x2 to transform it into a beta function.

(b) Deduce that

π/2
22n (n!)2
sin2n+1 θ dθ = .
0 (2n + 1)!
(c) Apply the formula (6) for the beta function to reprove (b) and to
calculate π/2
(2n)! π
sin2n θ dθ = 2n+1 .
0 2 (n!)2
(Compare with Exercise 15 in Chapter 2.)
25. Let vn (r) denote the volume of the n-dimensional sphere

Bn (r) = (x1 , x2 , . . . , xn ) ∈ Rn : x21 + x22 + · · · + x2n ≤ r2
of radius r, where n = 1, 2, . . . . Then v1 (r) = 2r, v2 (r) = πr2 , and v3 (r) =
4 3 n
3 πr . In general, it is easy to see that vn (r) = Vn r , where Vn is the volume
of the unit sphere Bn (1). The numbers Vn can be calculated recursively by
slicing:
1 1

Vn+1 = vn 1 − x dx = Vn
2 (1 − x2 )n/2 dx .
−1 −1
Derive the formula
π n/2
Vn = , n = 1, 2, . . . .
Γ( n2 + 1)
Show in particular that
v4 (r) = 12 π 2 r4 , v5 (r) = 8 2 5
15 π r , v6 (r) = 16 π 3 r6 , and v7 (r) = 16 3 7
105 π r .
26. The digamma function is ψ(x) = dx {log Γ(x)}

d
= Γ (x)/Γ(x).
(a) Apply the functional equation Γ(x + 1) = xΓ(x) to show that ψ(x +
1) = ψ(x) + x1 , x > 0.
(b) Generalize the relation Γ (1) = − γ by proving that

ψ(n + 1) = 1 + 12 + · · · + n1 − γ , n = 1, 2, . . . .
(c) Use the infinite product formula for 1/Γ(x) to show that
∞

1 1
ψ(x) = − −γ, x > 0.
n x+n−1
n=1
(d) Deduce that

∞
1
ψ (x) = , x > 0,
(x + n)2
n=0
and conclude that the function log Γ(x) is convex.
References 267
27. Apply Stirling’s formula to verify the Gauss product formula (13) in
the special case where x is a positive integer.
References
[1] G. E. Andrews, R. Askey, and R. Roy, Special Functions, Cambridge University

[2] Emil Artin, The Gamma Function, English translation, Holt, Rinehart and
Winston, New York, 1964.
[3] H. Bohr and J. Mollerup, Lærebog i Matematisk Analyse, Volume 3, J. Gjellerup,
Copenhagen, 1922.
[4] Philip J. Davis, “Leonhard Euler’s integral: A historical profile of the gamma
function”, Amer. Math. Monthly 66 (1959), 849–869.
[5] P. Duren, “Changing faces: The mistaken portrait of Legendre”, Notices Amer.
Math. Soc. 56 (2009), 1440–1443; Ibid. 57 (2010), 703.
[6] D. V. Widder, Advanced Calculus, Second edition, Prentice–Hall, Englewood
Cliffs, N. J., 1961; reprinted by Dover Publications, Mineola, NY, 1989.
Chapter 10
Two Topics in
Number Theory
In this chapter we discuss two unrelated topics that lie at the intersection
of analysis with number theory. First we consider the notion of equidistri-
bution, a classical topic that has found applications beyond number theory
to such areas as probability, functional analysis, and topological algebra.
(The German and French terms are Gleichverteilung and équirépartition,
respectively. Another English term is uniform distribution.) Our discussion
will focus on a beautiful criterion for equidistribution (modulo 1) due to
Hermann Weyl.
Our second topic is the Riemann zeta function, which has intimate con-
nections with the prime numbers. We develop Euler’s product formula for
the zeta function, then give two relatively elementary derivations of the
functional equation, which involves the gamma function. In truth a full
appreciation of the functional equation requires some familiarity with com-
plex analysis, which guarantees the uniqueness of an analytic continuation.
Otherwise, however, the derivations are entirely self-contained.
10.1. Equidistributed sequences
Loosely speaking, a sequence of points is said to be equidistributed over
a given set if each subset receives its proper share of points. To make the
notion precise, let us consider the special case of a numerical sequence {αn }
in the interval [0, 1). The reason for choosing a half-open interval will become
apparent later. For any subinterval I ⊂ [0, 1), let ν(n, I) denote the number
269
270 10. Two Topics in Number Theory
of points among α1 , α2 , . . . , αn that lie in I. Then the sequence {αn } is

defined to be equidistributed over [0, 1) if
ν(n, I)
lim = |I|
n→∞ n
for every interval I ⊂ [0, 1), where |I| denotes the length of I. The existence
of each such limit is part of the requirement.
For example, it is intuitively clear that the sequence
0 0 1 0 1 2 0 1 2 3

1, 2, 2, 3, 3, 3, 4, 4, 4, 4, . . .
is equidistributed over [0, 1), although the proof requires a bit of effort. At
the other extreme, a convergent sequence can never be equidistributed. In
fact, it is easy to see that an equidistributed sequence must be everywhere
dense; that is, it must have a subsequence that converges to each point of the
interval [0, 1]. If not, then some interval I ⊂ [0, 1) of positive length is free
of all but a finite number of points αn in the sequence, and so ν(n, I)/n → 0
as n → ∞. Since |I| = 0, this violates the requirement for equidistribution.
In general it is not easy to determine whether a given sequence is equidis-
tributed. Our main purpose is to develop an effective criterion. Actually,
our discussion will focus on a modular notion of equidistribution. Two num-
bers x and y are said to be congruent modulo m, where m > 0 is a prescribed
number, if their difference x − y is an integer multiple of m. Such a con-
gruence is indicated by writing x ≡ y(mod m). The integer part of any real
number x is defined to be the greatest integer less than or equal to x, and is
denoted by [x]. The number x = x − [x] is called the fractional part of x.
In particular, 0 ≤ x < 1 and x ≡ x(mod 1) for every x ∈ R. Hence the
fractional part of a number x may be viewed as the unique representative
in the interval [0, 1) of the equivalence class determined by x.
An arbitrary numerical sequence {αn } is said to be equidistributed mod-
ulo 1 if its sequence {αn } of fractional parts is uniformly distributed in
[0, 1). If ξ is a fixed rational number, the sequence defined by αn = nξ
has periodic fractional parts and therefore is certainly not equidistributed
modulo 1. On the other hand, it is a remarkable fact that for any irrational
number ξ, the sequence {ξ, 2ξ, 3ξ, . . . } is equidistributed modulo 1. This
generalizes a classical result of Kronecker, who showed that the fractional
parts of each such sequence are everywhere dense in [0, 1]. Many elementary
proofs are now known. We will obtain a proof later as a corollary of Weyl’s
criterion.
Meanwhile, however, we digress to give a direct proof of Kronecker’s
theorem. Since ξ is irrational, the fractional parts nξ are all distinct. In
other words, nξ = mξ if n = m. By the Bolzano–Weierstrass theorem,

some subsequence of the sequence {nξ} converges to a point in the interval
[0, 1]. In particular, for each ε > 0 there exist positive integers m and k such
that
0 < |(m + k)ξ − mξ| < ε .
This says that the point kξ is within ε of either 0 or 1. Consequently, the
sequence of points
kξ, 2kξ, 3kξ, . . . ,
suitably terminated, forms a chain that moves across the interval (0, 1) in one
direction (modulo 1) or the other, with distance between successive points
always less than ε except for a possible jump near one of the endpoints 0 or
1. As a result, each point of the interval [0, 1] is within distance ε of a point
nξ, where n is some multiple of k. Thus the points nξ are everywhere
dense in the interval [0, 1], as Kronecker’s theorem asserts. (It should be
noted that Kronecker actually obtained a stronger quantitative form of the
theorem and gave a generalization to higher dimensions; cf. Hardy and
Wright [5].)
10.2. Weyl’s criterion

We now turn to the elegant criterion due to Hermann Weyl [11] for
equidistribution modulo 1. Weyl’s key idea is to express the definition in
terms of integrals. For any given interval I ⊂ [0, 1), let f be the periodic
extension of its characteristic function. This is the function defined on the
interval [0, 1) by f (x) = 1 for x ∈ I and f (x) = 0 for x ∈/ I, then extended
to the whole real line by f (x + 1) = f (x), so that it is periodic with period
1. Then the sequence {αn } is equidistributed modulo 1 if and only if
1
1
n
(1) lim f (αk ) = f (x) dx
n→∞ n 0
k=1
for every interval I ⊂ [0, 1). If the sequence {αn } is equidistributed modulo
1, the relation (1) can be generalized by linearity to periodic extensions of
arbitrary step functions on [0, 1). Here a step function is understood to be
a finite linear combination of characteristic functions of intervals in [0, 1).
This in turn implies the validity of (1) for each bounded Riemann integrable
function f that is periodic with period 1. To see this, recall that by the
definition of Riemann integrability, any such function f can be approximated
below and above by periodic extensions g and h of step functions on [0, 1).
More precisely, for each prescribed ε > 0, there are periodic extensions g
and h of step functions on [0, 1) for which g(x) ≤ f (x) ≤ h(x) and 0 ≤
h(x) − g(x) < ε on R , so that

1 1
0≤ h(x) dx − g(x) dx < ε .
0 0
Since the equation (1) holds for both g and h, we see that
1 1
1
n
lim g(αk ) = g(x) dx > f (x) dx − ε ,
n→∞ n 0 0
k=1
so that 1
1 1
n n
f (αk ) ≥ g(αk ) > f (x) dx − ε
n n 0
k=1 k=1
for all n sufficiently large. In a similar way, we find that

1
1 1
n n
f (αk ) ≤ h(αk ) < f (x) dx + ε
n n 0
k=1 k=1
for all n sufficiently large. But ε was chosen arbitrarily, so the last two
inequalities combine to show that (1) holds for f as well. In summary, we
have shown that whenever a sequence {αn } is equidistributed modulo 1, it
has the property (1) for every periodic function f that is Riemann integrable
over [0, 1). In particular, the limit in (1) exists.
The most familiar examples of nonconstant Riemann integrable func-
tions of period 1 are sin 2πmx and cos 2πmx for integers m > 0. Apply-
ing the formula (1) to these functions, and using Euler’s formula eiθ =
cos θ + i sin θ to simplify the writing, we conclude that if a sequence {αn } is
equidistributed modulo 1, then
1 2πimαk
n
lim e = 0, m = 1, 2, . . . .
n→∞ n
k=1
Weyl’s theorem asserts that the converse is also true.
Weyl’s Theorem. A sequence {αk } of real numbers is equidistributed mod-

ulo 1 if and only if
1 2πimαk
n
(2) lim e =0
n→∞ n
k=1
for each integer m = 1, 2, . . . .

Proof. We have already seen that (2) holds whenever the sequence {αk } is
equidistributed modulo 1. For the converse, suppose the sequence {αk } has
the property (2). It then follows by linearity that the relation (1) holds for
every trigonometric polynomial

N
f (x) = a0 + (am cos 2πmx + bm sin 2πmx) ,
m=1
since it holds trivially for constant functions. But according to the trigono-
metric version of the Weierstrass approximation theorem, every continuous
function that is periodic with period 1 can be approximated uniformly by
such polynomials. In fact, given any ε > 0, we can approximate such a
function f above and below by trigonometric polynomials t and u so that
t(x) ≤ f (x) ≤ u(x) , f (x) < t(x) + ε , and f (x) > u(x) − ε .
Then we see that

1 1
1 1
n n
lim sup f (αk ) ≤ lim u(αk ) = u(x) dx < f (x) dx + ε .
n→∞ n n→∞ n 0 0
k=1 k=1
Similarly,
1 1
1 1
n n
lim inf f (αk ) ≥ lim t(αk ) = t(x) dx > f (x) dx − ε .
n→∞ n n→∞ n 0 0
k=1 k=1
Since ε was chosen arbitrarily, it follows that the limit exists and
1
1
n
lim f (αk ) = f (x) dx .
n→∞ n 0
k=1
Thus the relation (1) holds for every continuous function f that is periodic
with period 1.
Finally, for any given interval I ⊂ [0, 1), let f be the periodic extension
of the characteristic function of I. It is to be shown that (1) holds for
each such function f . But f can be approximated in integral norm above
and below by continuous periodic functions, and since we have shown that
such functions satisfy (1), an argument similar to the above leads to the
conclusion that f satisfies (1) as well. This completes the proof that the
sequence {αk } is equidistributed modulo 1.
As an application, we can now prove that the sequence {kξ} is equidis-

tributed modulo 1 when ξ is irrational. We need to show that the condition
(2) of Weyl’s theorem is satisfied for αk = kξ. But for αk = kξ the sum is a
geometric series, which can be evaluated as

n
e2πim(n+1)ξ − e2πimξ
e2πimkξ = .
e2πimξ − 1
k=1
Therefore, the sum has modulus

e2πim(n+1)ξ − e2πimξ
2 1
≤ 2πimξ = .
e 2πimξ −1 |e − 1| | sin πmξ|
Because ξ is irrational, the number mξ is never an integer, so we have shown

that each of the sums in (2) remains bounded as n → ∞. In particular, the
condition (2) is easily satisfied for each m ∈ N, so by Weyl’s theorem the
sequence {nξ} is equidistributed modulo 1.
As a further application of Weyl’s criterion, we now establish a theorem
that will provide many more examples of equidistributed sequences. Here is
a discrete version of a result known as Fejér’s theorem.
Theorem. Let {αn } be a sequence of real numbers whose differences Δαn =

αn+1 − αn are monotonic, and suppose that Δαn → 0 and n|Δαn | → ∞ as
n → ∞. Then the sequence {αn } is equidistributed modulo 1.
To give a proof it suffices to show, according to Weyl’s theorem, that

the sequence {αn } has the property (2) for each m ∈ N. For this purpose a
simple inequality will be useful.

Lemma. eiθ − 1 − iθ ≤ 12 θ2 , θ ∈ R.
Proof of lemma. It is enough to take θ > 0. An integration by parts

verifies the representation
θ
e − 1 − iθ =
iθ
(t − θ)eit dt ,
0
which is actually a version of Taylor’s formula with remainder. Hence

θ
θ
iθ (t − θ)eit dt =
e − 1 − iθ ≤ (θ − t) dt = 1
2 θ2 .
0 0
Proof of theorem. Since {−αn } is equidistributed modulo 1 if {αn } is,

there is no loss of generality in assuming that Δαn > 0, so that {Δαn }
decreases to 0. Fixing a positive integer m, choose θ = 2πmΔαk and apply

the lemma to obtain after minor rearrangement
2πimα
e k+1 e2πimαk
− − 2πimαk
≤ 2π m Δαk ,
2 2
Δα 2πim e k = 1, 2, . . . .
k Δα k
By the triangle inequality, this implies that

2πimα
e k+1 e2πimαk 1 1
− − 2πim e 2πimαk
≤ − + 2π 2 m2 Δαk .
Δα Δαk Δαk+1 Δαk
k+1
Another application of the triangle inequality now yields the estimate

n n 2πimαk+1 2πimαk
e e
2πim e2πimαk ≤ −
Δαk+1 Δαk
k=1 k=1
n
1 1
+ − + 2π 2 m2 Δαk
Δαk+1 Δαk
k=1
2πimα
e n+1 e 2πimα1 1 1
= − +
−
Δαn+1 Δα1 Δαn+1 Δα1
+ 2π 2 m2 (αn+1 − α1 ) ,
since the sums telescope. Dividing the last inequality by n and using the
hypothesis that n Δαn → ∞, we conclude that (2) holds. To see that
(αn+1 − α1 ) /n → 0, observe that those expressions are arithmetic means
of the sequence {Δαn }, which tends to 0. Because (2) holds for each m =
1, 2, . . . , Weyl’s theorem shows that the sequence {αn } is equidistributed
modulo 1.
It is now a short step to a special case which Pólya and Szegő [7] attribute
to Fejér.
Corollary (Fejér’s Theorem). Let g(x) be continuous on the interval

[1, ∞) and differentiable on (1, ∞). Suppose that g (x) tends monotonically
to 0 and x|g (x)| → ∞ as x → ∞. Then the sequence {αn } = {g(n)} is
equidistributed modulo 1.
Proof. By the mean value theorem, Δαn = g (xn ) for some point xn ∈
(n, n+1), and so the sequence {αn } satisfies the hypotheses of the preceding
theorem.
√
Fejér’s theorem shows for instance that the sequences { n log n},
{(log n)2 }, and {n/ log n} are equidistributed modulo 1. The discrete form of
the theorem is more flexible. For example, it shows that the partial sums of
√
the divergent series ∞n=1 1/ n constitute a sequence that is equidistributed
modulo 1.
On the other hand, the sequence {log n} is not equidistributed modulo
1, as can be seen from the following partial converse of Fejér’s theorem.
Theorem. If a sequence {αn } is equidistributed modulo 1, then the differ-

ences Δαn = αn+1 − αn have the property
lim sup n|Δαn | = ∞ .

n→∞
Proof. The proof will invoke Hardy’s Tauberian theorem, as discussed in

Chapter 7. With the notation sn = a1 +a2 +· · ·+an and σn = (s1 +s2 +· · ·+
sn )/n , Hardy’s theorem states that if an = O(1/n) and σn → s as n → ∞,
then sn → s. Suppose now that the sequence {αn } is equidistributed modulo
1. Then by Weyl’s theorem, the arithmetic means
1 2πiαk
n
σn = e →0 as n → ∞ .
n
k=1
But in view of the inequality |eiθ − 1| ≤ |θ|, the assumption that
lim sup n|Δαn | < ∞

n→∞
would imply 2πiα

e n+1
− e2πiαn ≤ |Δαn | = O(1/n) ,
and so it would follow from Hardy’s theorem that e2πiαn → 0, which is

clearly impossible. Therefore, it must be true that lim supn→∞ n|Δαn | = ∞
whenever the sequence {αn } is equidistributed modulo 1.
Further information about equidistribution can be found in the book

by Pólya and Szegő [7]. For a broad discussion of the topic, the reader is
referred to the book by Kuipers and Niederreiter [6].
10.3. The Riemann zeta function

The Riemann zeta function is defined by
∞
1
ζ(x) = , x > 1.
nx
n=1
Euler showed that ζ(2) = π 2 /6 and he found a general expression for ζ(2m),
where 2m is any even integer, in terms of the Bernoulli numbers. (Details
will be given in Chapter 11.) Some closely related series are
∞
∞ ∞
1 1 1 1
= − = 1 − ζ(x)
(2k + 1)x nx (2k)x 2x
k=0 n=1 k=1
and
∞
∞
∞

(−1)n+1 1 1 1
= −2 = 1 − x−1 ζ(x).
nx nx (2k)x 2
n=1 n=1 k=1
The zeta function has intrinsic connections with number theory. For
instance, it is not hard to see that
∞

ζ(x)2 = d(n) n−x , x > 1,
n=1
where d(n) is the number of divisors of n. For instance, d(5) = 2, d(9) = 3,

and d(12) = 6. Similarly, it can be shown that
∞

ζ(x)ζ(x − 1) = σ(n) n−x , x > 2,
n=1
where σ(n) denotes the sum of the divisors of n. For example,

σ(12) = 1 + 2 + 3 + 4 + 6 + 12 = 28 .
At a more basic level, the zeta function is an essential tool in the study
of prime numbers. An integer k is a divisor of n if n = km for some integer
m. An integer p > 1 is said to be prime if it has no positive divisors except
1 and p. The first few primes are
2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, . . . .
An integer larger than 1 that is not prime is called composite. It is not
difficult to see that every composite number has a prime divisor.
It was known to the ancient Greeks that there are infinitely many prime
numbers. Here is Euclid’s elegant proof. Suppose, on the contrary, that
there were only a finite number of primes: p1 , p2 , . . . , pn . Then since the
number N = p1 p2 · · · pn + 1 is greater than every prime pk , it is composite
and therefore has a prime divisor. But clearly N is not divisible by any of
the primes pk . This contradiction shows that the number of primes cannot
be finite.
Each positive integer has a unique prime factorization. For instance,
60 = 22 · 3 · 5.
Fundamental Theorem of Arithmetic. Each integer n ≥ 2 can be

represented in one and only one way as a product of primes.
The existence of a prime factorization is fairly obvious. A formal proof
proceeds by induction. Suppose that every integer less than n has a prime
factorization. If n is prime, there is nothing to prove. If n is composite, it is
the product of two smaller integers, each of which has a prime factorization,
by the inductive hypothesis. Hence n has a prime factorization.
The uniqueness of a prime factorization, however, is not at all obvious.
Suppose the assertion is false, and let n be the smallest positive integer that
has two different representations
n = p1 p2 · · · pk = q1 q2 · · · q
as a product of primes, not necessarily distinct. The same prime factor
cannot occur in both representations, because any common factor could be
divided out to produce a smaller number with two different prime factor-
izations, contrary to our assumption that n is the smallest such number.
Suppose without loss of generality that p1 and q1 are the smallest primes
in their respective factorizations. Then p21 ≤ n and q12 ≤ n. Since p1 = q1 ,
one of these two inequalities must be strict, so that p1 q1 < n. In fact,
p1 q1 < n − 1, because p1 ≥ 2 and q1 ≥ 2. Thus the number N = n − p1 q1
lies in the range 2 ≤ N < n, so its prime factorization is unique. But p1
divides N , so it must occur in that factorization. Similarly, q1 also occurs in
the prime factorization of N . Thus p1 q1 is a divisor of N , so it is also a divi-
sor of n. Consequently, q1 divides the number n/p1 = p2 p3 · · · pk . But this
is not possible, because n/p1 is less than n and therefore has a unique prime
factorization, and q1 does not occur in that factorization. This contradiction
proves that the prime factorization of every number is unique.
For further study of elementary number theory, the book by Davenport
[2] is especially recommended.
The connection of the zeta function with the prime numbers is apparent
from its infinite product representation, which goes back to Euler.
Euler’s Product Formula.

1
ζ(x) = , x > 1,
p
1 − p−x
where the product extends over all primes p.
Implicit in the statement is the assertion that the infinite product con-
verges for every x > 1. The keys to the proof of the product formula are
the geometric series expansion
1 1 1
−x
= 1 + x + 2x + . . .
1−p p p
and the fundamental theorem of arithmetic. For each fixed x > 1 and
integers M > N ≥ 2 we have

1 1 1
∞
1
1 + x + 2x + · · · + M x ≤ = ζ(x) .
p p p nx
p≤N n=1
Let M → ∞ to infer that

1
≤ ζ(x) ,
1 − p−x
p≤N
so the infinite product converges and

1
≤ ζ(x) .
p
1 − p−x
For the reverse inequality, again choose M > N ≥ 2 and write

N
1
1 1 1

≤ 1 + x + 2x + · · · + M x
nx p p p
n=1 p≤N

1
1
≤ −x
≤ .
1−p p
1 − p−x
p≤N
Let N → ∞ to deduce that

1
ζ(x) ≤ .
p
1 − p−x
The two inequalities now combine to prove Euler’s product formula.

The product formula is valid only for x > 1, but Euler did not hesitate
∞ to 1
take x = 1 and “deduce” from
the divergence of the infinite series n=1 n

that the product p 1 − 1p diverges to zero. This in turn implies (see
Chapter 5, Theorem 1) that the sum of reciprocals of the primes diverges,
which says that the sequence of prime numbers tends to infinity rather
slowly.

For a more convincing argument that p p1 = ∞, we need only examine
the above proof of Euler’s product formula to write
N
1
1
≤
n=1
n
p≤N
1− 1
p
for each integer N ≥ 2. Letting N → ∞, we conclude that the product

p 1 − p diverges to zero.
1
The problem now arises to determine the density of the primes. If pn

denotes the nth prime, what can be said about the order of magnitude of
pn ? More precisely, what is the asymptotic behavior of pn as n → ∞? To
put the question another way, if π(x) denotes the number of primes not
exceeding x, what is the asymptotic behavior of π(x) as x → ∞? In 1792,
Gauss conjectured on the basis of extensive numerical data that
x
(3) π(x) ∼ as x → ∞ .
log x
Legendre made the same conjecture independently in 1798. Half a century

later, Chebyshev proved that
x x
A ≤ π(x) ≤ B
log x log x
for some positive constants A and B. Then in 1896, Jacques Hadamard and
Charles de la Vallée Poussin used results of Riemann on the zeta function
to give independent proofs of the asymptotic relation (3), now known as the
prime number theorem.
An equivalent form of the prime number theorem is that pn ∼ n log n
as n → ∞. To see this, note that there are exactly n primes less than or
equal to pn , which says that π(pn ) = n. Since pn → ∞ as n → ∞, the
prime number theorem in the form (3) implies that π(pn ) ∼ pn / log pn , so
that pn ∼ n log pn . In other words, pn /(n log pn ) → 1, and it follows that
log pn − log n − log log pn → 0 as n → ∞ .
But log log pn / log pn → 0, so this shows that log pn ∼ log n, which allows us
to conclude that pn ∼ n log n. A similar argument shows that the asymptotic
relation pn ∼ n log n implies (3). The details are left as an exercise.
Further discussion of prime numbers can be found for instance in the
book by Tenenbaum and Mendès France [8]. The article by Bateman and
Diamond [1] gives a nice historical account of the prime number theorem.
10.4. Connection with the gamma function

We now derive the relation
∞
tx−1
(4) Γ(x)ζ(x) = dt , x > 1,
0 et − 1
10.4. Connection with the gamma function 281
between the zeta function and the gamma function

∞
Γ(x) = e−u ux−1 du , x > 0.
0
The change of variables u = nt converts the latter integral to

∞
Γ(x) = nx e−nt tx−1 dt , x > 0 , n = 1, 2, . . . .
0
Now take x > 1 and sum over n to obtain (formally)

∞
∞
∞
Γ(x)ζ(x) = Γ(x) n−x = e−nt tx−1 dt
n=1 n=1 0
∞

∞ ∞
−nt x−1 tx−1
= e t dt = dt ,
0 0 et − 1
n=1
as claimed. To justify the interchange of summation and integration, we can

invoke Dini’s theorem (cf. Section 1.8, Corollary 3).
Although the relation (4) has been derived under the assumption that
x > 1, it can be used to extend the definition of ζ(x) to 0 < x < 1. Since
1
1
tx−2 dt = for x > 1 ,
0 x−1
we can rewrite (4) in the form

1 ∞
1 1 1 tx−1
Γ(x)ζ(x) = − t x−1
dt + + dt .
0 et − 1 t x−1 1 et − 1
Observe now that each of the terms on the right-hand side is defined and
continuous for all x > 0, except for the singularity of 1/(x − 1) at x = 1.
Consequently, since Γ(x) is defined and Γ(x) > 0 for x > 0, we have extended
the definition of ζ(x) to all x > 0, with a singularity at x = 1. Moreover,
we see that ζ(x) − 1/(x − 1) is continuous in the interval (0, ∞) because
Γ(1) = 1.
This extension of the zeta function into the interval (0, 1) may appear
rather arbitrary, but in the context of functions of a complex variable the
extension is found to be uniquely determined. If z is a complex number, the
formulas
∞ ∞
−z
ζ(z) = n and Γ(z) = e−t tz−1 dt
n=1 0
define ζ(z) and Γ(z) as analytic functions in the half-planes Re{z} > 1 and
Re{z} > 0, respectively. A generalization of the process just described then
extends ζ(z) to an analytic function in the half-plane Re{z} > 0, except
for a pole at z = 1. But a basic principle of complex analysis says that an
analytic extension can be given in at most one way.
In the language of analytic number theory, the vertical strip 0 < Re{z} <
1 is called the critical strip for the zeta function. The Riemann hypothesis,
perhaps the most famous unsolved problem in mathematics, is the conjecture
that all zeros of ζ(z) in the critical strip are situated on the line Re{z} = 12 ,
known as the critical line. The Riemann hypothesis, if true, would have
important implications in number theory. In particular, it would lead to a
sharper form of the prime number theorem.
10.5. Functional equation

The functional equation for the Riemann zeta function is
x
1−x
(5) π −x/2 Γ ζ(x) = π −(1−x)/2 Γ ζ(1 − x) .
2 2
For 0 < x < 1 this equation exhibits a certain symmetry of the extended
function ζ(x) about the point x = 12 . For x > 1 it defines a natural ex-
tension of the zeta function to the negative real axis, the restriction of its
analytic continuation to the left half-plane Re{z} < 0. Since the left-hand
side is continuous and positive for x > 1, the functional equation (5) shows
that the same is true for Γ(x/2)ζ(x) when x < 0. But the gamma func-
tion is continuous and nonvanishing on the negative real line except for its
singularities at the negative integers, so the extended zeta function has the
same properties except that the singularities of Γ(x/2) at the negative even
integers must be canceled by zeros of ζ(x). In other words, the functional
equation shows that the zeta function is continuous and nonvanishing on the
negative real line except that ζ(x) = 0 at the points x = −2, −4, −6, . . . .
These points are sometimes called the “trivial zeros” of the zeta function.
It is known that infinitely many zeros lie on the critical line, and that only
the trivial zeros lie outside the critical strip. Computer calculations have
located millions of zeros on the critical line and none elsewhere in the critical
strip, providing overwhelming numerical evidence in favor of the Riemann
hypothesis, but no proof has been found.
We will give two distinct proofs of the functional equation. Both are
elementary, but a full appreciation of either argument requires a little back-
ground in complex analysis; in particular, some familiarity with analytic
functions and the uniqueness of an analytic continuation.
The first proof is due to G. H. Hardy [3, 4]. It begins with the Fourier
series
∞
sin(2n + 1)t π
(6) F (t) = = (−1)m , mπ < t < (m + 1)π ,
2n + 1 4
n=0
for m = 0, 1, 2, . . . , which was derived and discussed in Chapter 8. The

strategy is to multiply the series (6) by tx−1 and integrate term by term,
using the general formula
∞
(7) ta−1 sin bt dt = b−a Γ(a) sin(πa/2) , 0 < a < 1, b > 0,
0
derived at the end of the last chapter (Section 9.9). Deferring a justification
of the term-by-term integration, we obtain
∞ ∞
∞
x−1 1
t F (t) dt = tx−1 sin(2n + 1)t dt
0 2n + 1 0
n=0
∞

1
= Γ(x) sin(πx/2)
(2n + 1)x+1
n=0

= Γ(x) sin(πx/2) 1 − 2−x−1 ζ(x + 1) , 0 < x < 1.
On the other hand, taking account of the sum of the Fourier series (6), we
find that
∞ ∞ (m+1)π
x−1 π m
t F (t) dt = (−1) tx−1 dt
0 4 mπ
m=0
∞

π x+1
= 1+ (−1) (m + 1) − m
m x x
.
4x
m=1
The last series converges for 0 < x < 1, by Leibniz’ alternating series theo-
rem, since the sequence {(m + 1)x − mx } decreases to zero as m → ∞. But
for x < 0 the same sequence increases to zero, so the series again converges.
In fact, the convergence is uniform in each interval −∞ < x ≤ b < 1, so the
sum
∞

S(x) = 1 + (−1)m (m + 1)x − mx
m=1
is well defined and continuous in the interval (−∞, 1). Trivially, S(0) = 1.
Observe now that for x < −1 the sum is

S(x) = 2 (1x − 2x + 3x − . . . ) = 2 1 − 2x+1 ζ(−x) .

By analytic continuation, this relation S(x) = 2 1 − 2x+1 ζ(−x) extends
to all x < 1 and effectively defines ζ(−x) for 0 ≤ x < 1. In particular, for
0 < x < 1 the two calculations of the integral can be equated to obtain
π x+1
1 − 2x+1 ζ(−x) = Γ(x) sin(πx/2) 1 − 2−x−1 ζ(x + 1) ,
2x
which reduces to
2x π x+1 ζ(−x) = Γ(x + 1) sin(πx/2) ζ(x + 1) .
Replacing x by x − 1, we arrive at the functional equation
(8) ζ(1 − x) = 21−x π −x cos(πx/2)Γ(x)ζ(x)
for the zeta function, an equivalent form of (5). The details of this equiva-
lence are left as an exercise.
It remains to justify the interchange of summation and integration:
∞ ∞ ∞
sin(2n + 1)t
(9) tx−1 F (t) dt = tx−1 dt , 0 < x < 1.
0 0 2n + 1
n=0
Observe first that the partial sums of the series
∞
∞ ∞
sin(2n + 1)t sin nt sin 2nt
= −
2n + 1 n 2n
n=0 n=1 n=1
are uniformly bounded. (See Chapter 8, Exercise 15.) On the other hand, it
can be shown by the technique of Abel summation (cf. Chapter 3, Exercise
12) that this series converges uniformly in each closed set that contains no
multiple of π (cf. Chapter 8, Exercise 16). These two facts allow us to
conclude that
R ∞ R
x−1 sin(2n + 1)t
t F (t) dt = tx−1 dt , 0 < x < 1,
0 0 2n + 1
n=0
for arbitrary R < ∞. Hence the proof of (9) reduces to showing that
∞
∞
1
(10) lim tx−1 sin(2n + 1)t dt = 0 .
R→∞ 2n + 1 R
n=0
But an integration by parts gives
∞ x−1
R cos(2n + 1)R
t x−1
sin(2n + 1)t dt =
2n + 1
R
∞
x−1
+ t x−2
cos(2n + 1)t dt
2n + 1 R
∞
Rx−1 1−x 2 Rx−1
≤ + tx−2 dt = ,
2n + 1 2n + 1 R 2n + 1
and (10) follows. This completes the proof of (9) and justifies the term-by-
term integration used to derive the functional equation for the zeta function.
Our second proof of the functional equation is quite different √

and is due
to Riemann himself. It relies on the inversion formula ϑ(1/t) = t ϑ(t) for
Jacobi’s theta function
∞

e−n
2 πt
ϑ(t) = , t > 0,
n=−∞
which we derived in Section 8.11 as an application of the Poisson summation

formula. The first step is to calculate the integral
∞ x
x/2−1 −n2 πt −x/2
t e dt = π Γ n−x , x > 0,
0 2
directly from the definition of the gamma function. Taking x > 1 and
summing over n, we then deduce that
x ∞ ∞
∞
−x/2 x/2−1 −n2 πt
(11) π Γ ζ(x) = t e dt = tx/2−1 ψ(t) dt ,
2 0 0
n=1
where
∞

e−n
2 πt
(12) ψ(t) = .
n=1
The interchange of summation and integration is justified by the uniform

√ [0, ∞). Observe now that ϑ(t) =
convergence of the sum (12) in the interval
2ψ(t) + 1, so that the formula ϑ(1/t) = t ϑ(t) implies

(13) ψ(t) = 12 t−1/2 − 1 + t−1/2 ψ(1/t) .
Using the relation (13), we can write

∞ 1 ∞
x/2−1 x/2−1
t ψ(t) dt = t ψ(t) dt + tx/2−1 ψ(t) dt
0 0 1
1
1 1 x/2−3/2
= t −t x/2−1
dt + tx/2−3/2 ψ(1/t) dt
2 0
∞ 0
+ tx/2−1 ψ(t) dt
1
∞ ∞
1
= + s−x/2−1/2 ψ(s) ds + tx/2−1 ψ(t) dt ,
x(x − 1) 1 1
with the substitution s = 1/t in the second integral. Consequently, the

equation (11) reduces to
(14) ∞
x 1
−x/2
π Γ ζ(x) = + t−x/2−1/2 + tx/2−1 ψ(t) dt , x > 1 .
2 x(x − 1) 1
But it is easy to see that ψ(t) ≤ Ce−πt for some constant C > 0 and all
t ≥ 1, and so the last integral in (14) converges for every x ∈ R. Thus the
relation (14) defines a natural extension of the zeta function to the whole
real line, except for the singularity at x = 1. A simple calculation reveals
that the right-hand side of (14) is unchanged if x is replaced by 1 − x, and
so the functional equation (5) follows.
The essence of the proof is to derive the functional equation for the zeta
function from Jacobi’s inversion formula for the theta function. On the other
hand, it is possible to reverse the argument and derive Jacobi’s formula
from the functional equation. Thus the two relations can be regarded as
equivalent.
Readers familiar with complex function theory may wish to consult the
book by Titchmarsh [10], where both of the preceding proofs of the func-
tional equation can be found, together with several others, in the more
natural setting of the complex plane. Further details of Hardy’s proof are
given in Titchmarsh [9], Section 4.45.
Bernhard Riemann had an unusually fertile imagination. Although he
published only a few papers, each one introduced new ideas of fundamental
importance. Born in Hanover (now part of Germany) in 1826, he entered
the University of Göttingen at age 19 but soon transferred to Berlin, where
he studied under Dirichlet and Jacobi. Returning to Göttingen in 1849, he
became a student of Gauss and completed an inaugural dissertation on geo-
metric aspects of complex function theory, including the Riemann mapping
theorem and the concept of a Riemann surface. His Habilitationschrift in
1853 focused on functions representable as the sum of a trigonometric se-
ries and led to him to formalize the notion of a Riemann integral. In his
famous Habilitationsvortrag of 1854 he presented the fundamental ideas of
Riemannian geometry. When Gauss died in 1855, Dirichlet was appointed
his successor, then when Dirichlet died four years later the position passed
to Riemann. At that point Riemann produced his celebrated paper on the
zeta function, viewing it not as a function of a real variable as Euler had
done, but of a complex variable. There he derived the functional equation,
stated the Riemann hypothesis, and showed its relevance to the distribution
of prime numbers. After several years of ill health, Riemann succumbed to
tuberculosis in 1866, two months short of his fortieth birthday.
Exercises
1. Show that the sequence
0 0 1 0 1 2 0 1 2 3
1, 2, 2, 3, 3, 3, 4, 4, 4, 4, . . .
is equidistributed over [0, 1).

Exercises 287
2. If each of the sequences {αn } and {βn } is equidistributed modulo 1, show

that the sequence {α1 , β1 , α2 , β2 , . . . } is also equidistributed modulo 1.
3. Give an example of a sequence that is everywhere dense in the interval

[0, 1] but is not equidistributed.
4. Show that the sequence {cos n} is everywhere dense in the interval [−1, 1].
In particular,
lim sup cos n = 1 and lim inf cos n = −1 .

n→∞ n→∞
5. Since the number e is irrational, the sequence {ne} is equidistributed

modulo 1. Show, however, that the sequence {n!e} is not equidistributed
modulo 1.
Hint. Apply
n an estimate for the difference between e and the partial sums
sn = 1 + k=1 1/k! .
√ n
1+ 5
6. Prove that the sequence 2 is not equidistributed modulo 1.
Can you generalize the result?
√ n √ n
Hint. Show that the numbers xn = 1+2 5 + 1−2 5 satisfy the relation
xn+2 = xn + xn+1 with x1 = 1 and x2 = 3.
7. Show that the sequence {αn } is equidistributed modulo 1 if
(a) αn = np , 0 1,
p
(c) αn = n(log n) , p < 0.
8. Show that the sequence {(log n)p } is not equidistributed modulo 1 when
p ≤ 1.
9. Prove that a sequence {αn } is equidistributed in the interval [0, 1] if and

only if
αk + α2k + · · · + αnk 1
lim 1 = , k = 1, 2, . . . .
n→∞ n k+1
10. Show that
∞

ζ(x)2 = d(n) n−x , x > 1,
n=1
where d(n) is the number of divisors of n.

11. Show that

∞

ζ(x)ζ(x − 1) = σ(n) n−x , x > 2,
n=1
where σ(n) denotes the sum of the divisors of n.
*12. Show that
∞
ζ(x − 1)
= ϕ(n) n−x , x > 2,
ζ(x)
n=1
where ϕ(n) denotes the number of integers less than or equal to n that are
relatively prime to n. (Two positive integers are said to be relatively prime
if their greatest common divisor is 1.)
Discussion. The function ϕ, introduced by Euler, has remarkable prop-
erties. It is multiplicative in the sense that ϕ(nm) = ϕ(n)ϕ(m)
whenever n
and m are relatively prime. A theorem of Gauss asserts that k|m ϕ(k) = m.
In other words, the sum of the numbers ϕ(k), extended over all divisors k
of m, is equal to m. See for instance Davenport [2] for further details.
13. As shown in Section 10.3, the prime number theorem π(x) ∼ x/ log x
implies that the nth prime pn satisfies the asymptotic relation pn ∼ n log n
as n → ∞. Prove that conversely, pn ∼ n log n implies π(x) ∼ x/ log x as
x → ∞.
14. The logarithmic integral function
x
1
Li(x) = dt
2 log t
is known to give a better approximation to π(x) than does x/ log x. Show
that Li(x) ∼ x/ log x by proving

x x
Li(x) = +O .
log x (log x)2
Hint. An integration by parts reduces the problem to showing
x
1 x
2
dt = O ,
2 (log t) (log x)2
√
and the estimate is completed by splitting the integral at x.
15. Show directly that the two forms (5) and (8) of the functional equation
for the zeta function are equivalent.
16. Conclude from the functional equation that ζ(0) = − 12 . Show further
that ζ(−1) = − 12 1 1
, ζ(−3) = 120 , and in general that ζ(1 − 2n) = − 2n
1
B2n
for n = 1, 2, . . . . (Apply Euler’s formula for ζ(2n) in terms of the Bernoulli
number B2n , quoted in Section 3.7 and proved in Section 11.3.)
References 289
References
[1] P. T. Bateman and H. G. Diamond, “A hundred years of prime numbers”, Amer.
Math. Monthly 103 (1996), 729–741.
[2] H. Davenport, The Higher Arithmetic: An Introduction to the Theory of Num-
bers, Eighth edition, Cambridge University Press, Cambridge, U.K., 2008.
[3] G. H. Hardy, “A new proof of the functional equation for the zeta-function”,
Matematisk Tidsskrift B, 1922, 71–73.
[4] G. H. Hardy, “On the integration of Fourier series”, Messenger of Math. 51
(1922), 186–192.
[5] G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers,
Third edition, Oxford University Press, Oxford, 1954.
[6] L. Kuipers and H. Niederreiter, Uniform Distribution of Sequences, New York,
Wiley, 1974; reprinted by Dover Publications, Mineola, NY, 2006.
[7] G. Pólya and G. Szegő, Aufgaben und Lehrsätze aus der Analysis, Band 1.
Vierte Auflage, Springer-Verlag, Heidelberg, 1970; English edition: Problems and
Theorems in Analysis, Volume 1, Springer-Verlag, New York, 1972.
[8] G. Tenenbaum and M. Mendès France, Les Nombres Premiers, Presses Univer-
sitaires de France, Paris, 1997; English translation: The Prime Numbers and Their
Distribution, American Mathematical Society, Providence, RI, 2000.
[9] E. C. Titchmarsh, The Theory of Functions, Second edition, Oxford University
Press, London, 1939.
[10] E. C. Titchmarsh, The Theory of the Riemann Zeta-Function, Oxford Univer-
sity Press, London, 1951.
[11] Hermann Weyl, “Über die Gleichverteilung von Zahlen mod. Eins”, Math.
Annalen 77 (1916), 313–352.
Chapter 11
Bernoulli Numbers
The Bernoulli numbers were introduced by Jacob Bernoulli (1654–1705)

in his posthumously published treatise Ars Conjectandi (1713) to give a
general formula for sums of the form 1p + 2p + · · · + np , where
p is a positive
integer. Euler used them later to evaluate the infinite series ∞ n
k=1 1/k for
all even integers n. They have become important in number theory and
combinatorics, and they occur in algorithms such as the Euler–Maclaurin
summation formula for numerical computation of definite integrals. We will
discuss some of these connections later, but our first priority is to define the
Bernoulli numbers and to develop an efficient scheme for their calculation.
11.1. Calculation of Bernoulli numbers

The Bernoulli numbers Bn are defined by the generating relation
∞

x Bn n
(1) = x .
ex − 1 n!
n=0
Calculations with the power series
x2 x3
ex − 1 = x + + + ...
2! 3!
(cf. Section 3.6 for justification) show that
x 1
=
e −1
x 1 + 2 x + 16 x2 + . . .
1
2
= 1 − 12 x + 16 x2 + . . . + 12 x + 16 x2 + . . . − . . .
= 1 − 12 x + 1 2
12 x +...
291
292 11. Bernoulli Numbers
for sufficiently small x, so that B0 = 1, B1 = − 12 , and B2 = 16 . In particular,
Bn ∞
x
= 1 − 1
x + xn .
ex − 1 2 n!
n=2
Observe now that

x Bn ∞
x 1 e +1
(2) f (x) = + 1
x = x = 1 + xn
ex − 1 2 2 ex − 1 n!
n=2
is an even function; it has the property f (−x) = f (x). This implies that
its power series contains only even powers of x. In other words, B3 = B5 =
B7 = · · · = 0.
The Bernoulli numbers of even index are best calculated with the help
of a recurrence relation. Multiplying both sides of the generating relation
(1) by the function
∞
1 n
ex − 1 = x ,
n!
n=1
we find that
∞
∞
∞
Bn 1 n
n
x= x x = c n xn ,
n! n!
n=0 n=1 n=1
where
B0 B1 B2 Bn−1
cn = + + +··· + .
0! n! 1! (n − 1)! 2! (n − 2)! (n − 1)! 1!
A comparison of coefficients now shows that c1 = 1 and cn = 0 for all n ≥ 2.

When binomial coefficients are introduced, the relation cn = 0 becomes

n−1
n

(3) Bk = 0 , n = 2, 3, . . . .
k
k=0
The equations c1 = 1, c2 = 0, c3 = 0 reduce respectively to

1 1
B0 = 1 , 2 B0 + B1 = 0 , 3 B0 + B1 + B2 = 0 ,
so that B1 = − 12 and B2 = 16 . The equation c4 = 0 then reduces to
1
1
24 B0 + 16 B1 + 14 B2 + B3 = 1
24 − 1
12 + 1
24 + 16 B3 = 0 ,
6
11.1 Calculation of Bernoulli numbers 293
or B3 = 0. The equation c5 = 0 gives
1
5 B0 + B1 + 2B2 + 2B3 + B4 = 0 ,
so that
B4 = − 15 + 1
2 − 1
3 = − 30
1
.
More generally, the equation c2n+1 = 0 takes the form
1
n
1 (2n)!
(4) − + B2k = 0 ,
2n + 1 2 (2n − 2k + 1)! (2k)!
k=1
since B2k+1 = 0 for k ≥ 1. The expression (4) is a recurrence relation

that allows the successive calculation of the Bernoulli numbers B2n from
previously found values of B2 , B4 , . . . B2n−2 . For n = 3 it reduces to
1
7 − 1
2 + 3B2 + 5B4 + B6 = 0 ,
which gives
B6 = − 17 + 1
2 − 1
2 + 1
6 = 1
42 .
For n = 4 the relation (4) is
1
9 − 1
2 + 4B2 + 14B4 + 28
3 B6 + B8 = 0 ,
and so
B8 = − 19 + 1
2 − 2
3 + 7
15 − 2
9 = − 30
1
.
For convenient reference, here is a summary of results:
B0 = 1 , B1 = − 12 , B2 = 1
6 , B4 = − 30
1
, B6 = 1
42 , B8 = − 30
1
.
Similar calculations, increasingly laborious, produce the values
B10 = 5
66 , B12 = − 2730
691
, B14 = 7
6 , B16 = − 3617
510 , etc.
The Bernoulli numbers can be found in books of mathematical tables and in

some computer software packages. For instance, Mathematica will display
Bn as a numerical fraction if one enters the code BernoulliB[n] with a
specific integer n.
The values of B2n so far calculated suggest that the signs always alter-
nate: B2 > 0, B4 < 0, B6 > 0, B8 < 0, etc. This is true and will be verified
by inspection of the formulas for Euler’s sums.
11.2. Sums of positive powers

The familiar formula
n(n + 1)
1 + 2 + 3 + ··· +n =
2
for the sum of the first n integers can be discovered in many ways. Its proof
is a popular exercise in mathematical induction. The formulas
n(n + 1)(2n + 1)
12 + 22 + 32 + · · · + n2 = ,
6
2
n(n + 1)
13 + 23 + 33 + · · · + n3 = , ...
2
are also easy to verify by induction, but are perhaps more difficult to dis-
cover. One method of discovery is to guess by analogy with the integral
of xp that the sum 1p + 2p + · · · + np will be a polynomial in n of degree
p + 1 and to determine the coefficients by interpolation of numerical data.
However, this method does not lead to a general formula for the sum. Jacob
Bernoulli found such a formula in terms of Bernoulli numbers.
Bernoulli’s formula is
p+1
1 p+1
n
p
(5) k = Bp−j+1 (n + 1)j , p = 1, 2, . . .
p+1 j
k=1 j=1
where
p+1 (p + 1)!
=
j (p − j + 1)! j!
is a binomial coefficient. If p = 2, for instance, the formula (5) reduces to

n
n(n + 1)(2n + 1)
k 2 = B2 (n + 1) + B1 (n + 1)2 + 13 B0 (n + 1)3 = .
6
k=1
For a proof of (5), observe first that

∞ ∞
n

n
n
1 xp
ekx = (kx)p = kp , x = 0 .
p! p!
k=0 k=0 p=0 p=0 k=0
On the other hand, the geometric progression has the sum

n
n
e(n+1)x − 1
e kx
= (ex )k = , x = 0 .
ex − 1
k=0 k=0
11.3. Euler’s sums 295
Comparing the two expressions, and recalling the generating relation (1),
we see that
n
x
(n+1)x
∞ xp+1
kp = x e −1
p! e −1
p=0 k=0
∞ ∞⎛ ⎞
Bk (n + 1)j ∞

= xk ⎝ xj ⎠ = ap xp+1 ,
k! j!
k=0 j=1 p=0
where

p+1
1
ap = Bp−j+1 (n + 1)j .
(p − j + 1)! j!
j=1
Equating the coefficients of xp+1 , we arrive at the formula

n
p+1
p!
k p = p! ap = Bp−j+1 (n + 1)j , p = 1, 2, . . . .
(p − j + 1)! j!
k=0 j=1
which is equivalent to (5).
11.3. Euler’s sums

In Chapter 3 we discussed Euler’s solution of the “Basel problem”, first
proposed by Jacob Bernoulli in the year 1690, to find the sum of the infinite
series
∞
1
.
n2
n=1
(See Section 3.7.) Euler solved the problem in 1735 with the sensational
discovery that the sum of the series is π 2 /6. Shortly thereafter, Euler also
found that
∞
∞

1 π4 1 π6
(6) = , = ,
n4 90 n6 945
n=1 n=1
and in general that

∞
2k
1 k+1 (2π)
(7) = (−1) B2k , k = 1, 2, . . . ,
n2k 2(2k)!
n=1
B2k2 is a Bernoulli
where number. Observe that for k = 1 the sum (7) reduces
to 1/n = π 2 /6, since B2 = 1/6.

Euler’s formula 1/n2 = π 2 /6 has been verified by many methods,
some quite elementary, but no proof is really simple. Two elementary proofs
were presented in Section 3.7 of this book. Euler based a proof on the infinite
product formula for the sine function (cf. Section 8.5). His result often
appears as a special case of a Fourier expansion, but then it depends on a
theorem ensuring that the Fourier series actually converges to the function
at the point in question. In fact, Fourier series provide a key to the proof of
Euler’s general formula (7), since Fourier expansion of the function cos cx
leads to the formula
∞
2t2
πt cot πt = 1 + ,
t2 − n2
n=1
as shown in Section 8.5. To understand the relevance of this formula to

Euler’s sums, expand the summands in geometric series to infer that
∞
∞ ∞
(t/n)2
πt cot πt = 1 − 2 = 1 − 2 (t/n)2k
1 − (t/n)2
n=1 n=1 k=1
(8) ∞ ∞ ∞
∞
1
=1−2 (t/n)2k = 1 − 2 t2k .
n2k
k=1 n=1 k=1 n=1
On the other hand, the function πt cot πt can be expanded in another

way that introduces Bernoulli numbers. Letting x = 2t in the generating
relation (2), we obtain the formula
∞
(2t)2k 2t t(e2t + 1) t(et + e−t )
(9) 1+ B2k = 2t + 12 (2t) = 2t = t .
(2k)! e −1 e −1 e − e−t
k=1
Now recall Euler’s formula eit = cos t + i sin t and its consequences:
eit + e−it eit − e−it

cos t = , sin t = .
2 2i
Replacing t by iπt in (9), we deduce that
∞
(2π)2k 2k
(10) πt cot πt = 1 + (−1)k B2k t .
(2k)!
k=1
Equating the coefficients of t2k in the two expansions (8) and (10) for
πt cot πt, we arrive at Euler’s formula (7).
A more direct version of the proof can be based instead on the expression
for πt coth πt obtained in Exercise 29 of Chapter 8 as an application of the
Poisson summation formula. Expand this expression as in (8) to introduce

Euler’s sums, then replace t by πt in (9) and compare coefficients of t2k to
derive the formula (7). A different proof will appear in Section 11.4.
It is clear from the equation (7) that (−1)k+1 B2k > 0 for all integers
k ≥ 1. This confirms the pattern of alternating signs in the sequence of
Bernoulli numbers B2k .
Euler’s formula (7) has two simple corollaries:
∞

k+1 (2 − 1)π
1 2k 2k
(11) = (−1) B2k
(2m + 1)2k 2(2k)!
m=0
∞ 2k−1 − 1)π 2k
1 k+1 (2
(12) (−1)n+1 = (−1) B2k
n2k (2k)!
n=1
The proofs are left as exercises.

The Riemann zeta function
∞
1
ζ(s) = , Re{s} > 1 ,
ns
n=1
a function of a complex variable, plays an important role in number theory

and was discussed in Chapter 10. Euler’s formula evaluates ζ(2k) for all
even integers 2k. However, the value of the zeta function is not known for
any odd integer. In 1973, Roger Apéry proved that ζ(3) is irrational, but it
is still not known whether ζ(3) is a rational multiple of π 3 .
11.4. Bernoulli polynomials

The Bernoulli polynomials bn (t) are defined by the generating relation
∞
bn (t)
xetx
(13) = xn .
e −1
x n!
n=0
For t = 0 this reduces to the generating relation (1) for the Bernoulli num-
bers, and so bn (0) = Bn . To see that bn (t) is a polynomial of degree n,
multiply the power series
∞ k
t
etx = 1 + xk
k!
k=1
by that of (1) and compare with (13) to calculate the coefficients b0 (t) ≡ 1
and
bn (t) Bn−k
n
= tk , n = 1, 2, . . . ,
n! (n − k)! k!
k=0
or
n
n
(14) bn (t) = Bn−k tk .
k
k=0
This shows that bn (t) is a monic polynomial of degree n, since B0 = 1.

The values
B0 = 1, B1 = − 12 , B2 = 16 , B3 = 0, B4 = − 30
1
, B5 = 0, B6 = 1
42
can be inserted to calculate

b1 (t) = t − 12 , b2 (t) = t2 − t + 16 ,
b3 (t) = t3 − 32 t2 + 12 t , b4 (t) = t4 − 2t3 + t2 − 1
30 ,
b5 (t) = t5 − 52 t4 + 53 t3 − 16 t , b6 (t) = t6 − 3t5 + 52 t4 − 12 t2 + 1
42 .
We will be interested in the values of bn (1). Calculations show that

b1 (1) = −B1 and bn (1) = Bn for n = 2, 3, 4, 5, 6, which suggests that bn (1) =
Bn for all n ≥ 2. This is most easily proved by inserting t = 1 into the
generating relation (13) to obtain
∞
Bn ∞
bn (1) xex −x
xn = = −x = (−x)n .
n! e −1
x e −1 n!
n=0 n=0
Comparison of coefficients then gives bn (1) = (−1)n Bn . Since Bn = 0 for

all odd n ≥ 3, this confirms that bn (1) = Bn for n ≥ 2. For another proof,
one can apply the recurrence relation (3) directly to the formula (14).
The most important property of the Bernoulli polynomials is the identity
(15) bn (t) = nbn−1 (t) , n = 1, 2, . . . .
This again is most easily deduced from the generating relation (13). Term-
by-term differentiation of (13) gives
∞
∞

bn (t) n x2 etx bn (t) n+1
x = x = x ,
n! e −1 n!
n=1 n=0
and a comparison of coefficients verifies (15). However, the term-by-term

differentiation needs to be justified, and this is not so easy. A more pedes-
trian proof of (15) can be based directly on the representation (14).
In terms of Bernoulli polynomials, the formula (5) for the sum of positive
powers can be recast as

n
1
(16) kp = bp+1 (n + 1) − Bp+1 .
p+1
k=1
The proof is left as an exercise.

The Bernoulli polynomials have an elegant Fourier series expansion,
most conveniently expressed in complex form as
n! e2πikt
(17) bn (t) = − , 0≤t≤1
(2πi)n kn
k=0
for n = 2, 3, . . . , where the sum extends over all integers k, positive and
negative, except for k = 0. The same formula holds for n = 1 but is restricted
to the open interval 0 < t < 1, since the periodic extension of b1 (t) = t − 12
is discontinuous at the endpoints. Recall that bn (0) = bn (1) = Bn for every
n ≥ 2.
To verify the expansion (17), write
∞

bn (t) = cnk e2πikt , 0 < t < 1,
k=−∞
where the Fourier coefficients are given by

1
cnk = bn (t)e−2πikt dt .
0
Observe first that

1 1
1
cn0 = bn (t) dt = bn+1 (t) dt
0 n+1 0
1 1
= bn+1 (1) − bn+1 (0) = Bn+1 − Bn+1 = 0
n+1 n+1
for n = 1, 2, . . . . For k = 0, an integration by parts gives
1 1
−2πikt
c1k = b1 (t)e dt = (t − 12 )e−2πikt dt
0 0
1
1 1 −2πikt
1 1 1
=− (t − 2 )e + e−2πikt dt = − .
2πik 0 2πik 0 2πik
For fixed k = 0, we now proceed by induction on n. Suppose that the
formula cnk = −n!/(2πik)n holds for some n ≥ 1. Then by (15)
1
cn+1,k = bn+1 (t)e−2πikt dt
0
1
1 −2πikt
1 1
=− bn+1 (t)e + b (t)e−2πikt dt
2πik 0 2πik 0 n+1

n+1 1 n+1 (n + 1)!
= bn (t)e−2πikt dt = cnk = − .
2πik 0 2πik (2πik)n+1
This completes the inductive argument and shows that the formula holds
for each k = 0 and all n ≥ 1. Hence the polynomial bn (t) has the Fourier
expansion (17).
For n = 2m and t = 0, the expression (17) reduces to
∞
2(2m)! 1
B2m = b2m (0) = − ,
(2πi)2m k 2m
k=1
so that
∞
2m
1 m+1 (2π)
= (−1) B2m .
k 2m 2(2m)!
k=1
Thus we have found another proof of Euler’s formula (7) for the value of the
zeta function at an even integer.
In general, the formula (17) takes the form
∞
2(2m)! cos 2πkt
(18) b2m (t) = (−1) m+1
, 0 ≤ t ≤ 1,
(2π)2m k 2m
k=1
for even indices n = 2m, and

∞
2(2m + 1)! sin 2πkt
(19) b2m+1 (t) = (−1)m+1 , 0 ≤ t ≤ 1,
(2π)2m+1 k 2m+1
k=1
for odd indices n = 2m+1, with the proviso that (19) holds only for 0 < t < 1
when m = 0.
11.5. Euler–Maclaurin summation formula

Bernoulli polynomials provide the key to a basic method for numeri-
cal calculation and asymptotic analysis of integrals, known as the Euler–
Maclaurin summation formula. Euler developed the technique in 1732 and
applied it in 1736 to calculate Euler’s constant γ accurately to 16 decimal
places, a stunning feat in that era of hand calculation. Earlier, before dis-
covering the exact values of ζ(2) and ζ(4), he had applied the formula to
calculate ζ(2), ζ(3), and ζ(4) to over 15 decimal places. The Scottish mathe-
matician Colin Maclaurin (1698–1746) discovered the method independently
and discussed it in his book A Treatise of Fluxions (1742).
The basic idea is to improve upon the Riemann sum approximation of
a definite integral by taking into account the derivatives of the integrand.
This is accomplished through successive integrations by parts. To fix ideas,
11.5. Euler–Maclaurin summation formula 301
suppose a function f (t) is defined on the whole real line and has derivatives
of all orders. Introduce the Bernoulli polynomial
b1 (t) = t − 1
2 = 12 b2 (t)
and integrate by parts to obtain

1 1 1 1
f (t) dt = b1 (t)f (t) dt
= b1 (t)f (t) − b1 (t)f (t) dt
0 0 0 0

1 1 1
= f (0) + f (1) − b (t)f (t) dt .
2 2 0 2
Now integrate by parts again, invoking the relations b2 (0) = b2 (1) = B2 and
b2 (t) = 13 b3 (t), to write

1
1 B2 1 1
f (t) dt = f (0) + f (1) + f (0) − f (1) + b3 (t)f (t) dt .
0 2 2! 3! 0
The last integral is now transformed in the same way, making repeated use
of the relations
1
bk (t) = b (t) and bk (0) = bk (1) = Bk ,
k + 1 k+1
and recalling that Bk = 0 for odd indices k ≥ 3. The result is

1 1
1 1 1 1
b3 (t)f (t) dt = b3 (t)f (t) − b (t)f (t) dt
3! 0 3! 0 4! 0 4

1 1 1 1
= − b4 (t)f (t) + b4 (t)f (4) (t) dt
4! 0 4! 0

B4
1 1
= f (0) − f (1) − b5 (t)f (5) (t) dt .
4! 5! 0
Continuing in this manner, we arrive at the general formula

1
1
B2k (2k−1)
f (t) dt = f (0) + f (1) + f (0) − f (2k−1) (1)
0 2 (2k)!
(20) k=1
1
1
− b2+1 (t)f (2+1)(t) dt , = 1, 2, . . . .
(2 + 1)! 0
The formula (20) is easily generalized to an arbitrary interval [j, j + 1],

for j = 0, ±1, ±2, . . . . The only difference is that the Bernoulli polynomial
bn (t) must be replaced by its periodic extension Pn (t) to the real line, defined
so that Pn (t) = bn (t) for 0 ≤ t ≤ 1 and Pn (t+1) = Pn (t) for all t ∈ R. These
functions Pn (t) are called the Bernoulli periodic functions. They retain the
properties Pn (t) = nPn−1 (t) and Pn (j) = Pn (j +1) = Bn , and so the formula
(20) is generalized to
(21)
j+1
1
B2k (2k−1)
f (t) dt = f (j) + f (j + 1) + f (j) − f (2k−1) (j + 1)
j 2 (2k)!
k=1
j+1
1
− P2+1 (t)f (2+1)(t) dt , = 1, 2, . . . .
(2 + 1)! j
Finally, the formulas (21) for intervals [j, j + 1] can be added to obtain
a general formula for any interval [m, n] where m and n are integers with
m < n. It is
n
n−1
f (t) dt = 12 f (m) + f (j) + 12 f (n)
m j=m+1
(22)
B2k (2k−1)
+ f (m) − f (2k−1) (n)
(2k)!
k=1
n
1
− P2+1 (t)f (2+1)(t) dt , = 1, 2, . . . .
(2 + 1)! m
This is the Euler–Maclaurin summation formula. It is valid for any function
f that has continuous derivatives up to order 2 + 1 on the interval [m, n].
Integration by parts gives the alternate form
n n
1 1
(23) P2 (t)f (t) dt = −
(2)
P2+1 (t)f (2+1)(t) dt
(2)! m (2 + 1)! m
for the remainder. It is often more convenient for estimation because the
inequality |Pn (t)| ≤ |Bn | holds for even indices n (cf. Exercise 15).
11.6. Applications of Euler–Maclaurin formula

As a first illustration, consider the problem of calculating Euler’s con-
stant n
1
γ = lim − log n = 0.5772156649 . . . .
n→∞ k
k=1
Apply the Euler–Maclaurin formula (22) with m = 1 and f (t) = 1/t. The
derivatives are
k!
f (k) (t) = (−1)k , k = 1, 2, . . . ,
tk+1
11.6. Applications of Euler–Maclaurin formula 303
and so the formula (22) becomes

n n
1 1 1 B2k 1
− log n = 1+ + 1 − 2k − P2 (t) t−(2+1) dt ,
k 2 n 2k n 1
k=1 k=1
with the alternate form (23) of the remainder. Now hold fixed and let
n → ∞ to arrive at the exact formula
∞
1 B2k

γ= + − P2 (t) t−(2+1) dt ,
2 2k 1
k=1
valid for an arbitrary positive integer . If we choose = 3, the formula is
∞
γ = 2 + 2 B2 + 4 B4 + 6 B6 −
1 1 1 1
P6 (t) t−7 dt
∞ 1
= 2 + 12 − 120 + 252 −
1 1 1 1
P6 (t) t−7 dt
∞1
= 0.578968253 . . . − P6 (t) t−7 dt .
1
Now recall that P6 (t) is the periodic extension of the Bernoulli polynomial
b6 (t) = t6 − 3t5 + 52 t4 − 12 t2 + 1
42 ,
and |P6 (t)| ≤ |B6 | = 42 1
. If we calculate the integral over the interval
1 ≤ t ≤ 5, we will have the error estimate
∞ ∞

P6 (t) t dt ≤ B6
−7
t−7 dt = 252
1
5−6 < 10−6 .

5 5
Thus the problem reduces to the straightforward but laborious calculation
of the integral
5 4 1

−7
P6 (t) t dt = b6 (t)(t + k)−7 dt = 0.001752579 . . . .
1 k=1 0
Subtracting the two numbers, we calculate

∞
γ = 0.578968253 . . . − 0.001752579 . . . − P6 (t) t−7 dt = 0.57721 . . . ,
5
accurate to 5 decimal places since the error is less than 10−6 . (In fact, the
error is much smaller, since the calculation gives γ ≈ 0.577215674, correct
to 7 decimal places.)
It must be emphasized that γ is not equal to
∞
1 B2k
+ .
2 2k
k=1
In fact, that infinite series diverges rapidly (cf. Exercise 4).
For a second illustration of the Euler–Maclaurin summation formula, we

now develop what is called Stirling’s series, a sharpened form of Stirling’s
formula √
n! ∼ nn e−n 2πn , n → ∞,
or equivalently,
√
lim log n! − n + 12 log n + n = log 2π .
n→∞
Now the integrand is chosen to be f (t) = log t, with derivatives

(k − 1)!
f (k) (t) = (−1)k+1 , k = 1, 2, . . . .
tk
The formula (22), with m = 1, becomes
n
n
1 B2k 1
log t dt = log k − log n + 1−
1 2 (2k − 1)(2k) n2k−1
k=1 k=1
n
1
− P2+1 (t) t−(2+1) dt .
2 + 1 1
Since n
log t dt = n log n − n + 1 ,
1
this reduces to

B2k 1
log n! − n + 2 log n + n = 1 +
1
−1
(2k − 1)(2k) n2k−1
(24) k=1
n
1
+ P2+1 (t) t−(2+1) dt .
2 + 1 1
√
But according to Stirling’s formula, the left-hand side tends to log 2π as
n → ∞, which implies that

∞ √
B2k 1
(25) 1− + P2+1 (t) t−(2+1) dt = log 2π
(2k − 1)(2k) 2 + 1 1
k=1
for arbitrary integers = 1, 2, . . . . Substituting (25) into (24), we conclude

that
√

B2k 1
log n! = n + 12 log n − n + log 2π +
(2k − 1)(2k) n2k−1
k=1
∞
1
− P2+1 (t) t−(2+1) dt .
2 + 1 n
Exercises 305
Now observe that the error term is estimated by

∞ ∞

− 1 P2+1 (t) t −(2+1)
dt ≤ C
1
t−(2+1) dt ≤ C 2
2 + 1 n
n n
for some constant C depending only on . Thus we have arrived at the
asymptotic formula
√ B2 1 B4 1
log n! = n + 12 log n − n + log 2π + + + ...
1 · 2n 3 · 4 n3
(26)
B2k 1 1
+ +O , n → ∞,
(2k − 1)(2k) n2k−1 n2k
valid for any positive integer k. (Note the switch of notation, changing to
k.) If k = 3, for instance, this says that

√ 1 1 1 1
log n! − n + 2 log n + n = log 2π +
1
− 3
+ 5
+O .
12 n 360 n 1260 n n6
In the notation of asymptotic series, the formula (24) can be written

∞

√ B2k 1
(27) log n! − n + 1
log n + n log 2π + .
2 (2k − 1)(2k) n2k−1
k=1
The infinite series in (27) diverges for every fixed value of n, but any partial
sum approximates the left-hand side as n → ∞ with error tending to zero
at the same rate as the first neglected term.
The formal infinite series in the asymptotic formula (27) is known as
Stirling’s series. The exponential form is
∞
n! B2k 1
√ exp
n −n 2πn (2k − 1)(2k) n2k−1
(28) n e k=1
1 1 139 571
1+ + 2
− 3
− + ... .
12 n 288 n 51840 n 2488320 n4
Exercises
1. Use Bernoulli’s formula (5) to derive the relation
n
3 n(n + 1) 2
k = , n = 1, 2, . . . .
2
k=1
2. Use Bernoulli’s formula to work out the expression

n
1
k4 = n(n + 1)(2n + 1)(3n2 + 3n − 1) , n = 1, 2, . . . ,
30
k=1
and check it by induction.

3. Use Euler’s formula to check the sums (6) for ζ(2), ζ(4), and ζ(6), and
to calculate the sum
∞
1 π8
ζ(8) = = .
n8 9450
n=1
4. Prove the asymptotic formula

n 2n √
|B2n | ∼ 4 πn , n → ∞.
πe
Conclude that the generating relation (1) for Bernoulli numbers has radius
of convergence 2π, and that |B2n |(2π)2n/(2n)! → 2 as n → ∞.
Suggestion. Use Euler’s formula (7) for ζ(2n).
5. Use the formula (14) to give a direct proof of (15).
6. Prove that
bp+1 (t + 1) − bp+1 (t)
= tp , p = 0, 1, 2, . . . .
p+1
Suggestion. Prove by induction. Suppose that bp (t + 1) − bp (t) = ptp−1 for
some p ≥ 1. Then observe that
bp+1 (t + 1) − bp+1 (t)
= bp (t + 1) − bp (t) ,
p+1
and note that bp+1 (1) − bp+1 (0) = 0.
7. Verify the formula (16) for the sum of positive powers.
Suggestion. Apply the formula of Exercise 6.
8. Show that x+1
bn (t) dt = xn , n = 1, 2, . . . ,
x
for all x ∈ R . Conclude in particular that
1
bn (t) dt = 0 , n = 1, 2, . . . .
0
9. Deduce the formulas (11) and (12) from Euler’s formula (7).
10. Show that the tangent function has the Taylor series expansion
∞
4n (4n − 1)
tan x = (−1)n+1 B2n x2n−1
(2n)!
n=1
π
= x + 13 x3 + 2 5
15 x + 17 7
315 x + ... , |x| < .
2
Hint. tan x = cot x − 2 cot 2x. For radius of convergence, use Exercise 4.
Exercises 307
11. Verify the Taylor series expansion

∞
x 4n − 2
=1+ (−1)n+1 B2n x2n
sin x (2n)!
n=1
=1+ 1 2
6x + 7 4
360 x + 31
15120 x
6
+ ... , |x| < π .
Hint. csc x = cot x + tan x2 .
12. Verify the symmetry property bn (1 − t) = (−1)n bn (t) for Bernoulli

polynomials. Conclude that bn ( 12 ) = 0 if n is odd.
13. Apply the Fourier expansion (18) of the Bernoulli polynomial b2n (t)
and Euler’s formula (12) to show that

b2n 12 = (21−2n − 1)B2n , n = 1, 2, . . . ,
a formula that holds trivially for odd indices (cf. Exercise 12). Conclude
that |b2n ( 12 )| < |B2n | for n = 1, 2, . . . .
1
14 . Show that b4n+1 (t) < 0 and b4n+3 (t) > 0 for 0 < t < 2, where
n = 0, 1, 2, . . . .
Suggestion. Proceed by induction and consider bn (t).
15. Prove that |b2n (t)| < |B2n | for 0 < t < 1 , n = 1, 2, . . . .
Suggestion. Use a computer to plot the polynomials and see what needs to
be proved. Mathematica will plot the nth Bernoulli polynomial bn (t) over
the interval 0 ≤ t ≤ 1 from the code Plot[BernoulliB[n,t], {t,0,1}].

16. Derive the formula bn 12 = (21−n − 1)Bn directly from the generating
relation (13), by appeal to the expansion (9) for t coth t.
Hint. coth x − coth 2x = csch 2x .
17. Calculate the sum

∞
(−1)k π3
= .
(2k + 1)3 32
k=0
Hint. Think Bernoulli polynomials.
18. Show that Bernoulli polynomials satisfy the relation
bn (t) + (−1)n+1 bn (−t) = −ntn−1 , n = 1, 2, . . . .
19. Show that Bernoulli’s formula (5) for the sum of positive powers is a
special case of the Euler–Maclaurin summation formula.
20. Carry out the calculations to derive the numerical formula (28) from
Stirling’s series (27).
21. Use the Euler–Maclaurin summation formula and a hand calculator to

compute
∞
1
= 1.6449340 . . . ,
n2
n=1
and prove that the result is accurate to 7 decimal places.

Suggestion. Take m = 5 and = 3 in the formula (22), and use the alternate
form (23) of the remainder.
* 22. There is an explicit formula for the Bernoulli numbers, discovered in

1883 by J. Worpitzky. It is

(−1)n n 1
n k−1
j k−1
Bn = n (−1) (j + 1)n−1 .
2 −1 2k j
k=1 j=0
For a reference, the paper by Rza̧dkowski [3] is recommended.
References
[1] Konrad Knopp, Theory and Application of Infinite Series, Second English edi-
tion, Blackie & Son, London and Glasgow, 1951; reprinted by Dover Publications,
Mineola, NY, 1990.
[2] H. L. Montgomery and R. C. Vaughn, Multiplicative Number Theory I. Classical
Theory, Cambridge University Press, Cambridge, U.K., 2007.
[3] G. Rza̧dkowski, “A short proof of the explicit formula for Bernoulli numbers”,
[4] V. S. Varadarajan, Euler Through Time: A New Look at Old Themes, American
Mathematical Society, Providence, RI, 2006.
Chapter 12
The Cantor Set
In this chapter we explore some of Georg Cantor’s ideas about abstract sets,
including the notion of cardinality. The centerpiece of our discussion is Can-
tor’s famous “middle-thirds” set, an uncountable set of Lebesgue measure
zero that is a valuable source of counterexamples in analysis. For instance,
the Cantor set is the basis for construction of the “devil’s staircase”, de-
scribed by a continuous nondecreasing function that is not constant, yet
has zero derivative at almost every point. The Cantor set also provides the
underlying concept for construction of a space-filling curve, as presented at
the end of the chapter. We begin, however, with an extended discussion of
cardinal numbers.
12.1. Cardinal numbers

Georg Cantor (1845–1918) was the founder of modern set theory. He
received his mathematical training in Berlin, where Ernst Kummer, Karl
Weierstrass, and Leopold Kronecker were among his teachers. After writing
a dissertation in number theory in 1867, he turned to problems of uniqueness
of trigonometric series representations. That led to questions about the size
of exceptional sets of real numbers, and ultimately to a precise definition of
a real number as the limit of a sequence of rational numbers. About the
same time (1872), Richard Dedekind published a book giving his definition
of real number in terms of what are now called Dedekind cuts. Dedekind
and Cantor both realized that the set of real numbers is in some sense “more
infinite” than the set of rationals. Cantor transformed that insight into a
precise mathematical concept, developing the notion of cardinal number. His
309
310 12. The Cantor Set
ideas were revolutionary, and they met with considerable resistance at the
time, notably from Cantor’s former teacher Kronecker, who firmly believed
that mathematical proofs should be constructive. The opposition was so
intense that Cantor was unable to obtain a proper academic position until
1879, when he became a professor at the University of Halle. He remained
in Halle for the rest of his career.
Two abstract sets (not necessarily sets of real numbers) are said to have
the same cardinality, or the same cardinal number, if they are bijectively
equivalent; that is, if they can be put in one-to-one correspondence. Thus
two finite sets have the same cardinality n if both sets contain exactly n
elements. Each set can then be put in one-to-one correspondence with the
set {1, 2, . . . , n} of the first n positive integers. A set is said to be countable
or denumerable if either it is finite or it can be put in one-to-one correspon-
dence with the set N = {1, 2, 3, . . . } of all positive integers. In the latter
case it is called countably infinite. Clearly, any subset of a countable set
is countable. According to Cantor’s notation, a countably infinite set has
cardinality ℵ0 , usually pronounced “aleph-naught” or “aleph-null”. (Aleph
is the first letter of the Hebrew alphabet.) Thus the set Q of all rational
numbers has cardinality ℵ0 . Cantor also proved that the set of all algebraic
numbers is countable; it too has cardinality ℵ0 . Cardinal numbers of infinite
sets are sometimes called transfinite numbers.
Observe that two infinite sets may have the same cardinality even though
one is a proper subset of the other. For instance, the set {2, 4, 6, . . . } of all
even integers is in obvious one-to-one correspondence with the set N through
the mapping n ↔ 2n. Cantor introduced the diagonal method to prove that
the set R of all real numbers is not countable. (See Section 1.2.) In other
words, its cardinality is larger than ℵ0 , and it is in this sense “more infinite”
than its subset Q. The cardinality of R is customarily denoted by c and is
called the cardinality of the continuum.
The notion that c is larger than ℵ0 can be formulated in greater gener-
ality. The cardinal number of an abstract set A is denoted by |A|. If B is
another set and there exists an injective mapping f : A → B, then we say
that |A| ≤ |B|. If in fact there exists a injection of A onto B, or a bijection,
then |A| = |B|. If |A| ≤ |B| but no such bijection exists, then we say that
|A| < |B|. In this sense we know that ℵ0 < c.
It is worth remarking that the existence of an injective mapping f : A →
B is equivalent to the existence of a surjective mapping g : B → A. In other
words, there is a one-to-one mapping of A into B if and only if there is a
mapping of B onto A.
For finite cardinal numbers, or ordinary positive integers n and m, the
two relations n ≤ m and m ≤ n imply n = m. This implication can be shown
12.1. Cardinal numbers 311
to extend to transfinite cardinals. Specifically, if A and B are arbitrary sets

with the properties |A| ≤ |B| and |B| ≤ |A|, then |A| = |B|. This is the
content of the Schröder–Bernstein theorem, which can be stated as follows.
Schröder–Bernstein Theorem. Let A and B be arbitrary sets. If there

exist injective mappings f : A → B and g : B → A, then there is a bijection
h : A → B.
The result is very useful because it allows us to conclude that two sets
have the same cardinality without explicitly constructing a bijection. The
theorem is sometimes called the Cantor–Schröder–Bernstein theorem or the
Cantor–Bernstein theorem, or simply Bernstein’s theorem. After Cantor
made the conjecture, Ernst Schröder (1841–1902) and Felix Bernstein (1878–
1956) gave independent proofs in 1896 and 1897, respectively. Bernstein
presented his proof, at age 19, in Cantor’s seminar in Halle.
Proof of Schröder–Bernstein theorem. If g maps B onto A, the theo-

rem is proved. Otherwise, the complementary set E0 = A\g(B) is nonempty.
Let
En+1 = g f (En ) , n = 0, 1, 2, . . . ,
and let E = ∪∞
n=0 En . For x ∈ A, define

f (x) , for x ∈ E
h(x) =
g −1 (x) , for x ∈
/ E.
Note that x ∈
/ E implies x ∈ / E0 , so x ∈ g(B) and g −1 (x) is well defined.
We claim that h is the required bijection.
To show that h is a one-to-one mapping, suppose on the contrary that
f (x1 ) = g −1 (x2 ) for some elements x1 ∈ E and x2 ∈
/ E. But x1 ∈ E implies
x1 ∈ En for some index n, so
x2 = g(f (x1 )) ∈ En+1 ⊂ E ,
which is a contradiction. Thus h is an injection.

To show that h(A) = B, suppose on the contrary that h(x) = y for some
y ∈ B and all x ∈ A. By the definition of h, this is equivalent to the two
properties
(i) y = f (x) for all x ∈ E , and

(ii) y = g −1 (x) for all x ∈
/ E.
The property (i) implies that g(y) = g(f (x)) for all x ∈ E. In particular,
g(y) = g(f (x)) for all x ∈ En , where n = 0, 1, 2, . . . . Therefore, g(y) ∈
/ En+1
for each n = 0, 1, 2, . . . . But obviously g(y) ∈ g(B), so it is also clear that

g(y) ∈/ E0 . Thus g(y) ∈ / E. On the other hand, the property (ii) implies
that g(y) = x for all x ∈ / E, and so g(y) ∈ E. This contradicts our earlier
conclusion that g(y) ∈ / E, so h(A) = B and the proof is complete.
As an application of the Schröder–Bernstein theorem, it is easy to show

that every interval on the real axis, bounded or unbounded, has the cardi-
nality of the continuum. It is rather awkward to construct a specific bijec-
tion between the intervals (0, 1) and (0, 1], for instance, but the Schröder–
Bernstein theorem circumvents the problem.
It is also possible to show that any rectangle in the plane, or even the
whole plane R2 , has cardinality c. The same is true for Euclidean space Rn
of any dimension. On the other hand, it is not difficult to find sets with
cardinality greater than c, as we shall now see.
For any set A, the power set P(A) is defined as the set of all subsets of
A, including the empty set and the full set A. The power set is essentially
the same as the set of all functions f defined on A with values 0 or 1. Such
a function corresponds to a particular subset B ⊂ A, where f (x) = 0 or 1
is taken to indicate that x ∈/ B or x ∈ B, respectively. Cantor proved that
the power set of any set A has cardinality larger than that of A.
Cantor’s Theorem. The power set of every nonempty set A has cardinal-
ity |P(A)| > |A|.
If A is a finite set with n elements, it is easy to see that P(A) has 2n

elements, so the main interest of the theorem is for infinite sets. The theorem
shows in particular that there exist sets with cardinality greater than that
of the continuum, and in fact that there is no largest cardinal number.
Proof of theorem. First note that |A| ≤ |P(A)|, since we can construct
an injection of A into P(A) by assigning the singleton subset {x} to any
element x ∈ A. To show that the cardinality of P(A) is strictly larger than
that of A, let f : A → P(A) be an arbitrary injection, and define the set
B ∈ P(A) by
B = {x ∈ A : x ∈
/ f (x)} .
Suppose f (x0 ) = B for some x0 ∈ A. If x0 ∈ B, then by definition x0 ∈ /

f (x0 ) = B, a contradiction. If x0 ∈
/ B, then x0 ∈ f (x0 ) = B, which is again
a contradiction. Thus there cannot be any point x ∈ A for which f (x) = B.
This shows that no injection f of A into P(A) can have range f (A) = P(A).
Therefore, there is no bijection of A onto P(A), and the cardinality of P(A)
is larger than that of A.
12.2. Lebesgue measure 313
As an illustration of Cantor’s theorem, it will now be shown that the

collection of all subsets of positive integers has the cardinality of the con-
tinuum. In symbols, |P(N)| = c. To see that c ≤ |P(N)|, define the mapping
f : P(N) → [0, 1] by f (∅) = 0 and
1
f (A) = for A ⊂ N , A = ∅ .
2n
n∈A
Then f (A) is the number x with binary expansion x = 0.b1 b2 . . . , where

bn = 1 if n ∈ A and bn = 0 if n ∈ / A. Because every number x in [0, 1] has a
binary expansion, it is clear that f is a surjection, and so c = |[0, 1]| ≤ |P(N)|.
On the other hand, f is not an injection because every dyadic rational
number has two binary expansions, one ending in a string of zeros, the other
in a string of ones. To remedy the situation, we modify f and define a new
mapping g : P(N) → R by
1
g(A) = if N \ A is infinite, and
2n
n∈A
1
g(A) = 1 + if N \ A is finite.
2n
n∈A
Then g(A) has a binary expansion ending is a string of ones precisely when
N \ A is finite. The corresponding binary expansion of the same number
g(A) − 1 that ends in a string of zeros will occur when A is finite, which is
part of the case where N \ A is infinite. These considerations show that g
is an injection of P(N) into the interval [0, 2], so that P(N) has cardinality
|P(N)| ≤ |[0, 2]| = c. Thus |P(N)| = c, by the Schröder–Bernstein theorem.
12.2. Lebesgue measure

The Lebesgue measure of a set of real numbers is a generalization of the
length of an interval. An open interval I = (a, b) has length m(I) = b − a.
A set E ⊂ R is said to have measure zero, written m(E) = 0, if for each
ε > 0there is a finite or countable collection ofopen intervals Ik such that
E ⊂ k Ik and the intervals have total length k m(Ik ) < ε.
Clearly, every finite set has measure zero. In fact, it is not difficult to
show that every countable set has measure zero. To see this, let E ⊂ R be a
countable set consisting of points x1 , x2 , . . . . Given ε > 0, let Ik be the open
interval centered at xk with length m(Ik ) < ε/2k , for k = 1, 2, . . . . Then
∞
∞
∞
ε
E⊂ Ik and m(Ik ) < = ε.
2k
k=1 k=1 k=1
Thus E has measure zero.

Our subsequent discussion will refer only to the notion of a set of mea-
sure zero, but for the sake of completeness we now sketch the more general
definition of a measurable set and its Lebesgue measure. For simplicity we
will assume that our sets are bounded. Everything depends on the following
lemma.
Lemma. Every bounded open set E ⊂ R has a unique representation E =

k Ik as the union of a countable collection of disjoint open intervals Ik .
The proof is left as an exercise. On the basis of the lemma, the measure
of a bounded open set E ⊂ R can be defined by

m(E) = m(Ik ) , where E= Ik
k k
is its canonical representation as a disjoint union of open intervals.

Next let F ⊂ R be a bounded closed set. Since F is bounded, it is
contained in some bounded open interval J. The complementary set E =
J \ F is open and bounded, so the measure of F can be defined by
m(F ) = m(J) − m(E) .
It has to be checked that the definition is independent of the choice of J.

Finally, let A ⊂ R be an arbitrary bounded set of real numbers. The set
A is said to be measurable if
sup m(F ) = inf m(E) ,

F ⊂A E⊃A
where the supremum is taken over all closed sets F contained in A and
the infimum extends over all bounded open sets E that contain A. If A is
measurable, its Lebesgue measure m(A) is defined as the common value of
the supremum and infimum.
Lebesgue measure has many desirable properties. If A and B are mea-
surable sets with A ⊂ B, it is easily seen that m(A) ≤ m(B). For arbitrary
measurable sets A and B, it is possible to prove that A ∪ B is measurable
and
m(A ∪ B) ≤ m(A) + m(B) ,
with equality if A ∩ B = ∅. More generally, the union of any collection of
measurable sets A1 , A2 , . . . is measurable and

m Ak ≤ m(Ak ) ,
k k
12.3. The Cantor set 315
with equality if the sets Ak are pairwise disjoint.

Henri Lebesgue (1875–1941) developed this concept of measure in 1901
and used it as a basis for what is now called the Lebesgue integral. Every
Riemann integrable function is Lebesgue integrable, and the two integrals
are equal. However, the Lebesgue integral applies to a wider class of func-
tions and has significant technical advantages. It has become a standard
tool in modern analysis.
12.3. The Cantor set

We are now ready to construct Cantor’s famous “middle-thirds” set.
Begin with the closed unit interval [0, 1], and remove the open middle third
( 13 , 23 ) to form the set
F1 = [0, 13 ] ∪ [ 23 , 1] .
From each remaining interval, again extract the open middle third, leaving
the set
F2 = [0, 19 ] ∪ [ 29 , 13 ] ∪ [ 23 , 79 ] ∪ [ 89 , 1] .
Continue in this manner, removing the open middle third from each remain-
ing interval at each stage of the dissection, thus obtaining a nested sequence
of closed sets
F1 ⊃ F2 ⊃ F3 ⊃ · · · .
After n stages of dissection, the remaining set Fn consists of 2n disjoint
closed intervals, each of length 3−n . Figure 1 depicts the first 3 sets F1 , F2 ,
and F3 .
The Cantor set C is defined as the intersection
∞

C= Fn .
n=1
0 1 2 1 2 7 8 1
9 9 3 3 9 9
Figure 1. Construction of Cantor set: F1 , F2 , F3

The nested sets theorem guarantees that C is not the empty set, but this
is also evident by inspection, since each set Fn contains all of the endpoints
0, 1, 13 , 23 , 19 , 29 , 79 , 89 , 1 2 7 8
27 , 27 , 27 , 27 , ...
of remaining intervals, and so C must also contain these points. We shall

see that the Cantor set contains many other points as well.
It is easy to show that the Cantor set has measure zero. Indeed, Fn has
total length (2/3)n , so we can assert that

n
C ⊂ Fn ⊂ Ik
k=1
for some collection of open intervals Ik with total length

n
2 n
m(Ik ) < 2 3 .
k=1
But for each ε > 0 we can choose n large enough that 2(2/3)n < ε. This
shows that C has measure zero.
Our next aim is to show that the Cantor set is uncountable. By virtue
of its construction, Fn consists precisely of those points x that have some
ternary expansion
∞

x = 0.a1 a2 a3 · · · = ak 3−k
k=1
with ak = 0 or 2 for k = 1, 2, . . . , n. Thus a point x belongs to the Cantor
set if and only if it has a ternary expansion of the form x = 0.a1 a2 a3 · · ·
with ak = 0 or 2 for all k. The endpoints of removed intervals are triadic
rationals, numbers of the form m/3n . These points have two distinct ternary
expansions, exactly one of which has all digits ak = 0 or 2. For example,
7
9 = 0.21000 · · · = 0.20222 . . . .
Given a point x = 0.a1 a2 a3 · · · ∈ C with all ak = 0 or 2, define bk = 12 ak and

consider the binary expansion
∞

y = 0.b1 b2 b3 · · · = bk 2−k .
k=1
This operation defines a mapping x → y of the Cantor set onto the interval
[0, 1]. However, the mapping is not injective. For instance,
x= 7
9 = 0.20222 · · · → y = 0.10111 · · · = 3
4 and
x= 8
9 = 0.22000 · · · → y = 0.11000 · · · = 3
4 .
In any case, at most two points x ∈ C correspond in this manner to the same
point y ∈ [0, 1]. This shows that a proper subset of C is bijectively equiv-
alent to the interval [0, 1], and so has the cardinality of the continuum. In
particular, the Cantor set is uncountable, a result that can also be obtained
by a diagonal argument similar to the proof that the set of real numbers is
uncountable.
To show that the full Cantor set has the cardinality of the continuum,
we may apply the Schröder–Bernstein theorem. Since C ⊂ [0, 1], there is
an obvious injection of C into [0, 1]. To construct an injective mapping of
[0, 1] into C, represent an arbitrary point y ∈ [0, 1] by its binary expansion
y = 0.b1 b2 b3 . . . , adopting the convention that in cases of ambiguity the
expansion shall end in an infinite string of zeros. Let ak = 2bk and define
∞

x = 0.a1 a2 a3 · · · = ak 3−k
k=1
by a ternary expansion with digits ak . Then x ∈ C, since all ak = 0 or 2.

Thus we have constructed an injective mapping of the interval [0, 1] into C,
which is a subset of [0, 1]. By the Schröder–Bernstein theorem, the two sets
C and [0, 1] have the same cardinality. In other words the Cantor set has
the cardinality of the continuum.
12.4. The Cantor–Scheeffer function

We turn now to the construction of a remarkable function, known as
the Cantor function although it was discovered in 1883 by Ludwig Scheeffer
(1859–1885), a talented young mathematician who was in communication
with Cantor, and whose work Cantor promoted. We prefer to call it the
Cantor–Scheeffer function.
This function f (x) is continuous and nondecreasing on the interval [0, 1]
and it has the values f (0) = 0 and f (1) = 1, yet its derivative vanishes
almost everywhere; that is, outside a set of measure zero. As we shall see,
the exceptional set is the Cantor set.
For a point x ∈ C with ternary expansion
x = 0.a1 a2 a3 . . .
and all digits ak = 0 or 2, the Cantor function f can be defined by the

binary expansion
f (x) = 0.b1 b2 b3 . . . ,
where bk = 12 ak . For points at opposite ends of an interval
(x1 , x2 ) = (m/3n , (m + 1)/3n )

removed from [0, 1] in the construction of the Cantor set, it is not difficult
to see that f (x1 ) = f (x2 ). The definition of f is then extended to the whole
interval [0, 1] by setting f (x) ≡ f (x1 ) in each removed interval x1 ≤ x ≤ x2 .
The resulting function can then be shown to be continuous and to have the
other stated properties.
However, there is an alternate approach that is more geometric and
perhaps more transparent. We will produce the function f (x) as the uniform
limit of a sequence of continuous nondecreasing functions fn (x) with values
fn (0) = 0 and fn (1) = 1. First define
⎧ 3
⎪
⎨ 2x , 0 ≤ x ≤ 13
3 ≤x≤ 3
1 1 2
f1 (x) =
⎪ 2,
⎩ 3
2x − 2 , 3 ≤ x ≤ 1.
1 2
Note that f1 (x) is constant on the removed interval ( 13 , 23 ) and it increases

linearly with derivative f1 (x) = 32 on each interval of the set F1 . Next
define f2 (x) to have the constant values 14 on ( 19 , 29 ) and 12 on ( 13 , 23 ) and
3 7 8 3 2
4 on ( 9 , 9 ), and to increase linearly with derivative f2 (x) = ( 2 ) on each
interval of the set F2 . Continue in this manner. Thus define f3 (x) by
replacing the linear function on each interval of F2 with a scaled copy of
f1 (x). The resulting function f3 (x) will then increase linearly with derivative
f3 (x) = ( 32 )3 on each interval of F3 and will be constant on each of the
intervening intervals. In general, the function fn (x) is linear with derivative
fn (x) = ( 32 )n on each interval of the set Fn and has a constant value j/2n
(for some j = 1, 2, . . . , 2n − 1) on each of the intervening intervals. It is
continuous and nondecreasing on the interval [0, 1], with the values fn (0) = 0
and fn (1) = 1. Graphs of the functions f1 (x), f2 (x), and f3 (x) are displayed
in Figures 2a, 2b, and 2c.
We will now show that the functions fn (x) constitute a uniform Cauchy
sequence on [0, 1]. For arbitrary n > m ≥ 1, it is apparent from the con-
struction that
1
|fn (x) − fm (x)| ≤ , 0 ≤ x ≤ 1.
2m
This makes clear that {fn (x)} is a uniform Cauchy sequence, so it converges
uniformly on [0, 1] to a function f (x). Because each function fn (x) is con-
tinuous and nondecreasing, with values fn (0) = 0 and fn (1) = 1, the same
properties are bestowed upon the limit function.
Finally, the function fm (x) is constant on each open interval of the
complementary set [0, 1] \ Fm , which has measure 1 − ( 23 )m . But for each
n > m the function fn (x) = fm (x) on that set, and so f (x) is constant on
y
1
1

2
x
1 2 1

3 3
Figure 2a. Graph of y = f1 (x)
y
1
3

4
1

2
1

4
x
1 2 1 2 7 8 1

9 9 3 3 9 9
Figure 2b. Graph of y = f2 (x)

y
1
7

8
3

4
5

8
1

2
3

8
1

4
1

8
x
1 2 1 2 7 8 1

9 9 3 3 9 9
Figure 2c. Graph of y = f3 (x)
[0, 1] \ Fm for each m. Therefore, f (x) = 0 on

∞
∞

([0, 1] \ Fm ) = [0, 1] \ Fm = [0, 1] \ C .
m=1 m=1
Since the Cantor set has measure zero, this shows that f (x) = 0 almost
everywhere on the interval [0, 1].
The curve in the graph of the Cantor–Scheeffer function is sometimes
called the devil’s staircase. The function was popularized by Lebesgue, who
adopted it as an example of a continuous nonconstant monotonic function
that is singular in the sense that f (x) = 0 almost everywhere. The con-
struction can be modified to produce a continuous singular function that
is strictly monotonic; in other words, it is strictly increasing on every open
subinterval of [0, 1].
12.5. Space-filling curves

A curve in the plane is defined by its parametric representation
x = ϕ(t) , y = ψ(t) , a ≤ t ≤ b,
where ϕ and ψ are continuous functions. More precisely, the curve is defined
as an equivalence class of such representations, since the actual choice of pa-
rameter is unimportant. The reason for such a pedantic definition is that the
12.5. Space-filling curves 321
more intuitive notion is not entirely adequate. It would seem more natural
to view a curve as a continuous image of a line segment, a connected path
of points in the plane. However, such an image can have totally unexpected
form. In 1890, Giuseppe Peano (1858–1932) astounded the mathematical
world by constructing a continuous curve that passes through every point
of a square!
Soon after Peano [4] produced his example, David Hilbert and E. H.
Moore found different constructions. Today such curves are known as space-
filling curves or Peano curves. Lebesgue based another construction on the
Cantor set, and I. J. Schoenberg [6] modified Lebesgue’s construction to
produce a very elegant example, which we now proceed to describe.
First define the function
⎧
⎪
⎨ 0, 0≤t≤ 1
3
g(t) = 3t − 1 , 1
≤t≤ 2
⎪
⎩
3 3
1, 2
3 ≤ t ≤ 1.
Let g(−t) = g(t) and g(t + 2) = g(t), so that g is a continuous even function
of period 2. Its graph is shown in Figure 3.
The curve is now defined by the parametric representation
1 1 1
x = ϕ(t) = g(t) + 2 g(32 t) + 3 g(34 t) + . . .
2 2 2
1 1 1
y = ψ(t) = g(3t) + 2 g(3 t) + 3 g(35 t) + . . . ,
3
2 2 2
where 0 ≤ t ≤ 1. Since 0 ≤ g(t) ≤ 1, the convergence of both series is

uniform, and so the functions ϕ(t) and ψ(t) are continuous on the interval
[0, 1]. Note also that 0 ≤ ϕ(t) ≤ 1 and 0 ≤ ψ(t) ≤ 1, so that the point
(x, y) = (ϕ(t), ψ(t)) lies in the closed unit square. It is to be shown that the
curve meets each point of the square for some value of t ∈ [0, 1]. Suppose,
y
1
t
1 2 3 4 5
Figure 3. Graph of y = g(t)

then, that an arbitrary point (x0 , y0 ) is prescribed, where 0 ≤ x0 ≤ 1 and

0 ≤ y0 ≤ 1. Represent x0 and y0 by their binary expansions
1 1 1
x0 = b1 + 2 b3 + 3 b5 + . . .
2 2 2
1 1 1
y0 = b2 + 2 b4 + 3 b6 + . . . ,
2 2 2
where bk = 0 or 1. Let ak = 2bk and define the number t0 by the ternary
expansion
∞

t0 = ak 3−k .
k=1
It is clear that 0 ≤ t0 ≤ 1. If b1 = 0, then 0 ≤ t0 ≤ 13 and so g(t0 ) = 0. If

b1 = 1, then 23 ≤ t0 ≤ 1 and so g(t0 ) = 1. In either case, g(t0 ) = b1 . More
generally, the same argument shows that
g(3n t0 ) = bn+1 , n = 0, 1, 2, . . . .

To see this, observe that 3n t0 differs from ∞ k=1 an+k 3
−k by an even integer,
because all of the coefficients ak are even. Since g(t + 2) = g(t), this shows
that ⎛ ⎞
∞
g(3n t0 ) = g ⎝ an+j 3−j ⎠ = bn+1 , n = 0, 1, 2, . . . .
j=1
Consequently, the definition of the curve gives

1 1 1
ϕ(t0 ) = g(t0 ) + 2 g(32 t0 ) + 3 g(34 t0 ) + . . .
2 2 2
1 1 1
= b1 + 2 b3 + 3 b5 + · · · = x0
2 2 2
and
1 1 1
ψ(t0 ) = g(3t0 ) + 2 g(33 t0 ) + 3 g(35 t0 ) + . . .
2 2 2
1 1 1
= b2 + 2 b4 + 3 b6 + · · · = y0 .
2 2 2
This shows that the curve passes through the arbitrarily prescribed point
(x0 , y0 ). Therefore, it passes through every point of the unit square.
The coordinate functions ϕ(t) and ψ(t) of Schoenberg’s space-filling
curve are nowhere differentiable. For an elementary proof, see Sagan [5].
Other examples of space-filling curves, but not all, share this property. For
instance, the coordinate functions of Lebesgue’s example are known to be
differentiable almost everywhere.
Exercises 323
Exercises
1. (a) The Cartesian product A × B of two sets A and B is the set of all
ordered pairs (a, b) with a ∈ A and b ∈ B. If A and B are countable sets,
prove that A × B is countable.
(b) Prove that every countable union of countable sets is countable. More
precisely,
∞ if each of the sets E1 , E2 , . . . is countable, then their union E =
E
n=1 n is also countable.
2. (a) Apply the Schröder–Bernstein theorem to show that the intervals

(0, 1) and (0, 1] have the same cardinality.
(b) Find a bijective mapping of the interval (0, 1) to (0, ∞).
(c) Find a bijective mapping of the interval (0, 1) to R.
3. A number x is said to be algebraic if it is a root of a polynomial

with integer coefficients: a0 + a1 x + · · · + an xn = 0 for some n ∈ N and
a0 , a1 , . . . , an ∈ Z with an = 0. √Every rational number is algebraic, but
many algebraic numbers such as 2 are irrational. Show that the set of all
algebraic numbers is countable.
4. Let A be an arbitrary set and let P(A) denote its power set, the set of
all subsets of A. Show that there is a bijection between P(A) and the set of
all mappings f : A → {0, 1}, the set consisting of the two numbers 0 and 1.

5. Prove that every open set E ⊂ R has a unique representation E = In
as a union of a countable collection of open intervals In .
6. Use a diagonal argument to prove that the Cantor set is uncountable.
7. For any bounded set E ⊂ R, the derived set E consists of all cluster
points of E. We know that E is always a closed set, although it may be
empty. A bounded set E ⊂ R is said to be perfect if E = E . Thus every
closed interval [a, b] is a perfect set. Prove that the Cantor set C is a perfect
set.
8. The Cantor set construction can be generalized to any constant ratio of

dissection. Fix a number ξ with 0 < ξ < 12 . Begin with the unit interval
[0, 1] and remove the central open interval (ξ, 1 − ξ) to form the set F1 =
[0, ξ] ∪ [1 − ξ, 1]. From each of the two intervals comprising F1 remove the
central open interval in the same ratio ξ : 1 − 2ξ : ξ to form the set F2 .
Continue in this manner to generate a sequence of sets F1 , F2 , F3 , . . . .
(a) Show that Fn is composed of 2n closed intervals, each of length ξ n ,
for n = 1, 2, . . . .
∞
(b) Show that the intersection F = n=1 Fn is an uncountable closed set
of measure zero.
9. The Cantor set construction can be further generalized by considering

variable ratios of dissection. Fix a sequence of numbers ξk with 0 < ξk < 12 .
Begin by removing the interval (ξ1 , 1−ξ1 ) from the unit interval [0, 1] to form
the set F1 . From each interval of F1 remove the central open interval in the
ratio ξ2 : 1 − 2ξ2 : ξ2 to form the set F2 . Proceeding inductively, construct
a nested sequence of closed sets F1 ⊃ F2 ⊃ F3 ⊃ · · · by removing from each
interval of Fn−1 the central open interval in the ratio ξn : 1 − 2ξn : ξn to
form the set Fn .
(a) Show that Fn consists of 2n closed intervals, each of length ξ1 ξ2 · · · ξn ,
for n = 1, 2, . . . .

(b) Show that the intersection F = ∞ n=1 Fn is a closed set with empty
interior; it contains no open intervals.
∞(c) Show that F has measure zero if and only if the infinite product
n=1 (2ξn ) diverges to 0.
(d) Exhibit
∞ a sequence {ξk } with 0 < ξk < 12 for which the infinite
product n=1 (2ξn ) converges. Thus obtain a closed set of positive measure
with empty interior.
10. Show that the Cantor–Scheeffer function satisfies a Lipschitz condition
|f (x) − f (t)| ≤ C |x − t|α , α = log 2/ log 3 ,
for x, t ∈ [0, 1], where C is a positive constant.
11. Show that the “devil’s staircase” curve given by the graph of the Cantor–
Scheeffer function, y = f (t) for 0 ≤ t ≤ 1, has arclength 2. Note the failure
of the usual formula
1
1 + f (t)2 dt
0
for arclength. Note also the failure of the fundamental theorem of calculus,
since
1
f (t) dt = f (1) − f (0) .
0
Here both integrals are taken in the Lebesgue sense. Two functions that
differ only on a set of measure zero have equal Lebesgue integrals.
References 325
References
[1] Ralph P. Boas, A Primer of Real Functions, Third edition, Mathematical
Association of America, Washington, DC, 1981.
[2] E. Hille and J. D. Tamarkin, “Remarks on a known example of a monotone
continuous function”, Amer. Math. Monthly 36 (1929), 255–264.
[3] E. Kamke, Theory of Sets, English translation, Dover Publications, New York,
1950.
[4] G. Peano, “Sur une courbe qui remplit toute une aire plane”, Math Annalen 36
(1890), 157–160.
[5] Hans Sagan, “An elementary proof that Schoenberg’s space-filling curve is
nowhere differentiable”, Math. Magazine 65 (1992), 125–128.
[6] I. J. Schoenberg, “On the Peano curve of Lebesgue”, Bull. Amer. Math. Soc. 44
(1938), 519.
Chapter 13
Differential Equations
Differential equations occur in all branches of science. They provide impor-

tant mathematical models for classical problems of physics. Many of those
models involve linear differential equations of second order, which give rise
to special functions such as Bessel functions, Legendre polynomials, and
hypergeometric functions. Traditional studies of differential equations em-
phasize techniques for finding explicit solutions either in closed form or as
infinite series. In this chapter we devote some attention, for instance, to the
construction of power series solutions at a regular singular point, and we
develop formulas for Bessel functions and hypergeometric functions. How-
ever, the most important properties of solutions are typically deduced not
from explicit formulas but directly from the fact that the functions satisfy
the differential equation. This approach comes to full fruition in the Sturm
comparison theorem and related results, discussed at the end of the chapter.
13.1. Existence and uniqueness of solutions

Our study of differential equations will begin at the beginning. The
questions of existence and uniqueness of solutions have direct importance
in applications, and are central to the entire mathematical theory. Here we
propose to discuss the existence and uniqueness of solutions to an initial-
value problem
(1) y = f (x, y) , y(x0 ) = y0 .
Is there a function y = ϕ(x) that satisfies the differential equation and the
initial condition, and is there only one such function?
327
328 13. Differential Equations
To be more precise, we will assume that the function f (x, y) is continuous

in a rectangle
R = {(x, y) : |x − x0 | ≤ a , |y − y0 | ≤ b} ,
where it satisfies a Lipschitz condition
|f (x, y) − f (x, z)| ≤ C |y − z|
for some constant C > 0. Then we can prove the following theorem.
Existence and Uniqueness Theorem. Let f (x, y) be continuous in the
rectangle R and satisfy a Lipschitz condition there. Let M be the maximum
value of |f (x, y)| on R, and let δ = min{a, b/M }. Then there exists a
unique continuously differentiable function ϕ(x) that satisfies the differential
equation ϕ (x) = f (x, ϕ(x)) in the interval |x − x0 | < δ and has a prescribed
value ϕ(x0 ) = y0 .
Intuitively, the theorem is quite plausible. The differential equation
specifies a direction field for the solution curves. In other words, it pre-
scribes the slope of the tangent line as the curve passes through any point
of the region. If the direction field is continuous, there would seem to be
exactly one way to start at a given point and trace a curve with the specified
tangent directions. Nevertheless, the theorem is false without the Lipschitz
condition, as a simple example shows. Consider the initial-value problem
y = 3 y 2/3 , y(0) = 0 ,
in some neighborhood of the origin. This problem is easily seen to have two
distinct solutions, y = x3 and y = 0. The solution is not unique, but the
function f (x, y) = 3 y 2/3 does not satisfy a Lipschitz condition as required
by the theorem, since
|f (x, y) − f (x, 0)|/|y − 0| = 3 |y|−1/3 → ∞ as y → 0 .
We propose to prove the theorem by the method of successive approxi-

mations, a technique introduced in 1890 by the French mathematician Émile
Picard (1856–1941). His method is constructive in the sense that it provides
a mechanism for calculating the solution to any desired accuracy. Another
proof, essentially due to Cauchy, was developed and published by Giuseppe
Peano in the same year 1890. Peano’s proof arrives at the solution through
a sequence of polygonal approximations. The Runge–Kutta method, com-
monly used today for numerical solution of differential equations, is a re-
finement of the Cauchy–Peano scheme. As a theoretical device, however,
Picard’s method of iteration has the advantage of elegance and relative sim-
plicity.
Proof of theorem. The first step is to observe that the initial-value prob-
lem (1) is equivalent to the integral equation
x
(2) ϕ(x) = y0 + f (t, ϕ(t)) dt .
x0
In other words, any continuously differentiable function ϕ that solves the
initial-value problem (1) must satisfy the integral equation (2), and con-
versely any continuous solution of the integral equation must be continuously
differentiable and satisfy (1). Thus our problem reduces to showing that the
integral equation (2) has a continuous solution, and that this solution is
unique.
Turning first to the question of existence, we will show that the equation
(2) has a continuous solution ϕ(x) in the interval |x − x0 | < δ. The strategy
is to obtain the solution as a uniform limit of continuous functions defined
inductively by ϕ0 (x) ≡ y0 and
x
(3) ϕn+1 (x) = y0 + f (t, ϕn (t)) dt . |x − x0 | < δ ,
x0
for n = 0, 1, 2, . . . . For convenience we will restrict attention to the half-
interval [x0 , x0 + δ). Similar arguments apply to the interval (x0 − δ, x0 ].
The first step is to show that each of the functions ϕn (x) is well defined
and continuous on [x0 , x0 + δ), and that
(4) |ϕn (x) − y0 | ≤ M (x − x0 ) , x0 ≤ x < x0 + δ .
Obviously the constant function ϕ0 (x) has these properties. Proceeding
inductively, suppose that some function ϕn (x) has the stated properties.
Then since δ ≤ b/M it follows that (x, ϕn (x)) ∈ R for all x ∈ [x0 , x0 + δ),
so that ϕn+1 (x) is well defined and continuous on this interval and
x
|ϕn+1 (x) − y0 | ≤ |f (t, ϕn (t))| dt ≤ M (x − x0 ) .
x0
Thus every function ϕn (x) has the stated properties.
The proof of uniform convergence will rely on the inequality
M Cn
(5) |ϕn+1 (x) − ϕn (x)| ≤ (x − x0 )n+1 , x0 ≤ x < x0 + δ ,
(n + 1)!
for n = 0, 1, 2, . . . , where C is the Lipschitz constant of the function f (x, y).
For n = 0 the inequality (5) is the same as (4). To prove it for general n,
we apply the iterative definition (3) to see that
x
|ϕn+2 (x) − ϕn+1 (x)| ≤ |f (t, ϕn+1 (t)) − f (t, ϕn (t))| dt
x0
x
≤C |ϕn+1 (t) − ϕn (t))| dt .
x0
Therefore, if (5) holds for some integer n, then

x
M C n+1
|ϕn+2 (x) − ϕn+1 (x)| ≤ (t − t0 )n+1 dt
(n + 1)! x0
M C n+1
= (x − x0 )n+2 , x0 ≤ x < x0 + δ) ,
(n + 2)!
which shows that (5) holds for the next integer n + 1 as well. This proves
(5) for all n.
In particular, the inequality (5) gives the uniform estimate
M C n δ n+1
|ϕn+1 (x) − ϕn (x)| ≤ , n = 0, 1, 2, . . .
(n + 1)!
in the interval [x0 , x0 + δ) . From this it follows that {ϕn (x)} is a uniform
Cauchy sequence, hence it converges uniformly on [x0 , x0 + δ) to some con-
tinuous function ϕ(x). Alternatively, the Weierstrass M-test shows that

n

ϕn+1 (x) = y0 + ϕk+1 (x) − ϕk (x)
k=0
converges to ϕ(x), uniformly on [x0 , x0 + δ), as n → ∞. In view of the

Lipschitz condition, we can infer that the sequence {f (t, ϕn (t))} converges
uniformly to f (t, ϕ(t)), so that
x x
lim f (t, ϕn (t)) dt = f (t, ϕ(t)) dt .
n→∞ x x0
0
Consequently, we can pass to the limit in the iterative relation (3) and con-
clude that ϕ(x) satisfies the integral equation (2). This proves the existence
of a solution of (2), hence of the initial-value problem (1).
Uniqueness is proved in a similar way. If ϕ(x) and ψ(x) are solutions of
(2), then
x
|ϕ(x) − ψ(x)| ≤ |f (t, ϕ(t)) − f (t, ψ(t))| dt
x0
x
≤C |ϕ(t) − ψ(t)| dt , x0 ≤ x < x0 + δ .
x0
Since |ϕ(x) − x0 | < b and |ψ(x) − x0 | < b, so that |ϕ(x) − ψ(x)| < 2b, it
follows that
|ϕ(x) − ψ(x)| ≤ 2b C(x − x0 ) , x0 ≤ x < x0 + δ .

Iteration gives
x
2b C 2
|ϕ(x) − ψ(x)| ≤ 2b C 2
(t − x0 ) dt = (x − x0 )2
x0 2!
and in general
2b C n 2b C n δ n
|ϕ(x) − ψ(x)| ≤ (x − x0 )n ≤ , x0 ≤ x < x0 + δ ,
n! n!
for n = 1, 2, . . . . Letting n → ∞, we conclude that ϕ(x) ≡ ψ(x) in the
interval [x0 , x0 + δ). This proves the uniqueness of a solution of (2), hence
of (1), and completes the proof of the theorem.
It is not difficult to estimate the error committed in approximating the

true solution ϕ by the function ϕn . It can be shown that
M eCδ C n δ n+1
|ϕ(x) − ϕn (x)| ≤ , |x − x0 | < δ , n = 0, 1, 2, . . . .
(n + 1)!
The proof is left as an exercise.

The existence and uniqueness theorem and its proof are readily gener-
alized to initial-value problems for systems of differential equations
(6) y = f (x, y) , y(x0 ) = b ,
where y(x) = (y1 (x), . . . , yn (x)) is a vector-valued function of x, f =

(f1 , . . . , fn ) is a vector-valued function of x and y, and b = (b1 , . . . , bn )
is a prescribed point. It is convenient to work with the norm
y = |y1 | + |y2 | + · · · + |yn | .
The function f is assumed to be defined and continuous in some rectangular

parallelepiped centered at the point (x0 , b) = (x0 , b1 , . . . , bn ), and to satisfy
a Lipschitz condition
f (x, y) − f (x, z) ≤ C y − z
there. Then the theorem asserts the existence of a unique solution y = ϕ(x)
to the initial-value problem (6), valid in some interval |x − x0 | < δ. Picard’s
iterative proof can be adapted to this more general situation. The problem
(6) is converted to an equivalent integral equation
x
y(x) = b + f (t, y(t)) dt ,
x0
and a solution y = ϕ(x) is constructed as a uniform limit of the vector-

valued functions ϕk (x) defined inductively by ϕ0 (x) = b and
x
ϕk+1 (x) = b + f (t, ϕk (t)) dt , k = 0, 1, 2, . . . .
x0
The details are similar to the scalar case and are left as an exercise.
The last theorem applies in particular to linear systems
y1 = a11 y1 + a12 y2 + · · · + a1n yn

y2 = a21 y1 + a22 y2 + · · · + a2n yn
· · ·
yn = an1 y1 + an2 y2 + · · · + ann yn ,
where the functions ajk = ajk (x) are continuous in some neighborhood of
x0 .
A single linear differential equation of higher order can be viewed as a
special case of a first-order linear system. Consider a differential equation
of the form
(7) y (n) + a1 (x)y (n−1) + · · · + an−1 (x)y + an (x)y = 0 ,
where each of the functions ak (x) is continuous in some neighborhood of a

point x0 . The initial-value problem is to find a solution of (7) that satisfies
a set of initial conditions
(8) y(x0 ) = b1 , y (x0 ) = b2 , . . . , y (n−1) (x0 ) = bn
for some specified numbers b1 , b2 , . . . , bn . To view this problem from the

perspective of a first-order linear system, set
z1 = y , z2 = y , . . . , zn = y (n−1) .
Then the differential equation (7) is expressed by the linear system
z1 = z2 , z2 = z3 , . . . , zn−1

= zn ,
zn = −an z1 − an−1 z2 − · · · − a1 zn ,
and the initial conditions (8) become z(x0 ) = b, where z = (z1 , . . . , zn ) and
b = (b1 , . . . , bn ). Consequently, the existence and uniqueness theorem for
first-order linear systems implies the existence and uniqueness of a solution
to the linear differential equation (7) under the initial conditions (8).
13.2. Wronskians
Recall that a set of functions f1 (x), f2 (x), . . . , fn (x) is said to be linearly
independent over an open interval I ⊂ R if no linear combination
c1 f1 (x) + c2 f2 (x) + · · · + cn fn (x)
vanishes throughout I, except for the trivial combination with all coefficients
ck = 0. It is equivalent to require that no function in the set be expressible
as a linear combination of the others. The functions are said to be linearly
dependent if some nontrivial combination vanishes identically in I.
If all of the functions fn (x) have derivatives up to order n − 1 on I, their
linear dependence implies that the system of linear equations
c1 f1 (x) + c2 f2 (x) + · · · + cn fn (x) = 0

c1 f1 (x) + c2 f2 (x) + · · · + cn fn (x) = 0
· · ·
(n−1) (n−1)
c 1 f1 (x) + c 2 f2 (x) + · · · + cn fn(n−1) (x) = 0
has a nontrivial solution (c1 , c2 , . . . , cn ) for all x ∈ I. Thus the determinant

f1 (x) f2 (x) ... fn (x)

f1 (x) f2 (x) ... fn (x)

W (x) = W (x; f1 , f2 , . . . , fn ) =
.. .. .. ..
. . . .
(n−1) (n−1) (n−1)
f1 (x) f2 (x) . . . fn (x)
must vanish identically on I. An equivalent statement is that the functions
f1 , . . . , fn are linearly independent on I if W (x0 ) = 0 for some point x0 ∈ I.
The determinant W (x) is called the Wronskian of the functions f1 , . . . , fn ,
in honor of the mathematician Josef Wronski (1778–1853).
It may well happen that a set of functions is linearly independent and yet
their Wronskian vanishes at some points of the interval. (See Exercise 3 for
an example.) This cannot happen, however, if the functions are solutions of
the same linear differential equation. Indeed, suppose that y1 , y2 , . . . , yn are
solutions of the differential equation (7) whose coefficients ak (x) are contin-
uous on an interval I. Let W (x) = W (x; y1 , . . . , yn ) denote the Wronskian.
If W (x0 ) = 0 for some point x0 ∈ I, then the corresponding system of linear
equations
c1 y1 (x0 ) + c2 y2 (x0 ) + · · · + cn yn (x0 ) = 0

c1 y1 (x0 ) + c2 y2 (x0 ) + · · · + cn yn (x0 ) = 0
· · ·
(n−1) (n−1)
c1 y1 (x0 ) + c2 y2 (x0 ) + · · · + cn yn(n−1) (x0 ) = 0
has a nontrivial solution (c1 , c2 , . . . , cn ). This says that the function
y = c1 y1 + c2 y2 + · · · + cn yn
has the properties
y(x0 ) = y (x0 ) = · · · = y (n−1) (x0 ) = 0 .
But the differential equation is linear, so y is a solution with the same data
at x0 as the trivial solution z(x) ≡ 0. By the uniqueness of the solution
to an initial-value problem, it follows that y(x) = z(x) = 0 for all x ∈ I.
In other words, the functions y1 , . . . , yn are linearly dependent on I if their
Wronskian vanishes at any point of the interval.
In summary, these results demonstrate a sharp dichotomy for sets of
solutions y1 , . . . , yn of the same differential equation (7). Either these func-
tions are linearly dependent and their Wronskian vanishes everywhere on
the interval I, or they are linearly independent and their Wronskian van-
ishes nowhere on I. There is no middle ground. The same dichotomy will be
apparent from an explicit formula for the Wronskian that we now proceed
to derive.
The case n = 2 is the most important, and will be considered first. Let
y1 and y2 be solutions of a differential equation
(9) y + p(x)y + q(x)y = 0 , x∈I,
where the coefficient functions p and q are continuous on I. Let

y1 (x) y2 (x)
W (x) = = y1 (x)y2 (x) − y1 (x)y2 (x)
y1 (x) y2 (x)
be the Wronskian of y1 and y2 . Differentiation gives
W = y1 y2 − y1 y2
= (py1 + qy1 )y2 − y1 (py2 + qy2 )
= p(y1 y2 − y1 y2 ) = −pW .
This is a linear first-order differential equation for W, namely W

x + pW =
0. After multiplication by the integrating factor exp x0 p(t)dt with any
conveniently chosen point x0 ∈ I, the equation reduces to

x
d
W (x) exp p(t)dt = 0,
dx x0
so that
x
(10) W (x) = W (x0 ) exp − p(t)dt .
x0
The formula (10) is known as the Wronskian relation for the differential
equation (9). Its most remarkable feature is its dependence only on the
coefficient function p, not on the solutions y1 and y2 , except for the value
of their Wronskian at a single point. The formula shows again that either
W (x) = 0 for all x ∈ I or W (x) ≡ 0 on I.
For a set of solutions y1 , y2 , . . . , yn of the general differential equation
(7) of order n, a similar calculation shows that the Wronskian W (x) =
W (x; y1 , . . . , yn ) satisfies the equation W = −a1 W , and so
x
(11) W (x) = W (x0 ) exp − a1 (t)dt .
x0
Here the calculation involves some general properties of determinants and

is left as an exercise.
Finally, it should be noted that the existence and uniqueness theorem
leads to an important structural formula for the solutions of a (homoge-
neous) linear differential equation. Suppose that y1 , y2 , . . . , yn is a linearly
independent set of solutions to the differential equation (7). Then any lin-
ear combination y = c1 y1 + c2 y2 + · · · + cn yn is again a solution, and linear
independence implies the non-vanishing of the Wronskian, so that the con-
stant coefficients c1 , . . . , cn can be chosen to make the function y satisfy any
prescribed initial conditions
y(x0 ) = b1 , y (x0 ) = b2 , . . . , y (n−1) (x0 ) = bn .
In particular, if z(x) is an arbitrary solution of (7), the coefficients c1 , . . . , cn

can be chosen so that
y(x0 ) = z(x0 ) , y (x0 ) = z (x0 ) , . . . , y (n−1) (x0 ) = z (n−1) (x0 ) .
It then follows from the uniqueness property that z(x) ≡ y(x) for this choice
of coefficients. In other words, every solution of (7) is a linear combination
of the set of linearly independent solutions y1 , . . . , yn .
In the special case where the equation (7) has constant coefficients, there
are well known procedures for explicit construction of a complete set of lin-
early independent solutions. The general solution is then expressible as some
linear combination of these functions. In the case of variable coefficients, the
calculation of solutions is generally not so easy. The problem is even more
difficult near points where the coefficients have singularities. In the next
section we discuss methods for finding solutions in the form of power series.
13.3. Power series solutions

In discussing solutions by power series, we restrict attention to linear
differential equations of second order,
(12) y + p(x)y + q(x)y = 0 .
A point x0 ∈ R is said to be a regular point if in some neighborhood of

x0 both of the functions p(x) and q(x) can be represented as power series
centered at x0 . At a regular point it is always possible to find a pair of
independent solutions, also represented by power series. To illustrate the
procedure, consider the simple example
y + 2xy + 2y = 0 .

We look for a solution in the form of a power series y = ∞ k
k=0 ck x with
positive radius of convergence. Term-by-term differentiation gives
∞
∞

y = kck xk−1 , y = k(k − 1)ck xk−2 .
k=1 k=2
Putting these expressions into the differential equation, we are led to the
requirement that
∞
∞
∞

(k + 2)(k + 1)ck+2 xk + 2 kck xk + 2 c k xk = 0 .
k=0 k=0 k=0
This imposes the equivalent condition that the coefficients of all powers xk
must vanish, so that
(k + 2)(k + 1)ck+2 + 2(k + 1)ck = 0 ,
or
2
(13) ck+2 = − ck , k = 0, 1, 2, . . . .
k+2
It is now clear that the first two coefficients c0 and c1 can be chosen ar-
bitrarily, and then the recurrence relation (13) will determine the others.
Specifically, we find from (13) that
1 1 1 1
c2 = −c0 , c4 = − c2 = c0 , c6 = − c4 = − c0 , . . . ;
2 2 3 3!
2 2 2 2
c3 = − c1 , c5 = − c3 = c1 , . . . ,
3 5 1·3·5
and in general
(−1)k (−1)k 2k
c2k = c0 and c2k+1 = c1 , k = 1, 2, 3, . . . .
k! 1 · 3 · 5 · · · (2k + 1)
Thus the general solution to the differential equation is

∞
∞

(−1)k (−1)k 2k
y = c0 x2k + c1 x2k+1 .
k! 1 · 3 · 5 · · · (2k + 1)
k=0 k=0
Both series are easily seen to converge on the whole real line, so that the
term-by-term differentiation is justified and the series represent actual solu-
tions to the differential equation. In fact, the first series sums to the function
e−x .
2
In a similar way, near any regular point x0 of a differential equation (12)

two linearly independent power series solutions can always be constructed.
It can be proved that the formal power series so constructed necessarily
converge in some neighborhood of x0 , hence represent bonafide solutions.
The method fails if x0 is a singular point ; that is, if it is not a regular
point of the differential equation (12). However, there may still exist a power
series solution at x0 , which can be found by the same device. Consider, for
example, Bessel’s equation
1
xy + y + xy = 0 , or y + y + y = 0.
x
Here x0 = 0 is a singular point. Nevertheless, we look for a power series
solution of the form
∞

y(x) = c k xk .
k=0
Inserting this series and its derivatives into Bessel’s equation, we find that
∞
∞
∞

k(k − 1)ck xk−1 + kck xk−1 + ck xk+1 = 0 ,
k=2 k=1 k=0
or
∞

c1 + (k + 2)2 ck+2 + ck xk+1 = 0 .
k=0
Thus the requirements are that c1 = 0 and
1
ck+2 = − ck , k = 0, 1, 2, . . . .
(k + 2)2
Since c1 = 0, the recurrence relation implies that ck = 0 for all odd indices
k. With the choice c0 = 1, the relation gives
1 1 1
c2 = − , c4 = , c6 = − ,... ,
22 22 42 22 42 62
so that, in standard notation,
x2 x4 x6
y(x) = J0 (x) = 1 − + − + ...
22 22 42 22 42 62
is a solution of Bessel’s equation. This power series converges on the whole
real line and is called a Bessel function of first kind. Its graph is shown in
Figure 1.
Because of the requirement that c1 = 0, the only power series solutions
are constant multiples of J0 (x). The search for a second solution leads us to
a general discussion of regular singular points and the method of Frobenius.
We will now describe this method in detail, applying it to specific equations
but omitting a proof of its general validity. For a more complete theoretical
discussion, the reader is referred to the book by Coddington and Levinson
[4].
A singular point x0 of the differential equation (12) is called a regular
singular point if (x − x0 )p(x) and (x − x0 )2 q(x) have power series expansions
about x0 :
∞
∞

(x − x0 )p(x) = ak (x − x0 )k and (x − x0 )2 q(x) = bk (x − x0 )k ,
k=0 k=0
0.8
0.6
0.4
0.2
5 10 15 20 25 30
-0.2
-0.4
Figure 1. Graph of Bessel’s function J0 (x).

where both series converge in some interval |x − x0 | < δ.

The most commonly occurring differential equations of mathematical
physics have at worst regular singular points. For instance, the Legendre
equation
(1 − x2 )y − 2xy + λy = 0 , λ a real constant,
has regular singular points at x = 1 and x = −1, and is regular elsewhere

on the real axis. Bessel’s equation
xy + y + xy = 0
has a regular singular point at the origin.

For ease of notation, we will suppose that our differential equation has a
regular singular point at x0 = 0. The strategy is then to seek a generalized
power series solution of the form
∞

y = xα c k xk , c0 = 1 ,
k=0
where the real number α and the coefficients ck are to be determined. Formal
calculations give
∞
∞

y = (k + α)ck xk+α−1 and y = (k + α)(k + α − 1)ck xk+α−2 ,
k=0 k=0
so that
∞
∞
∞

j−1 k+α−1
py = aj x (k + α)ck x = aj (k + α)ck xn+α−2
j=0 k=0 n=0 j+k=n
and
∞
∞
∞

j−2 k+α
qy = bj x ck x = bj ck xn+α−2 .
j=0 k=0 n=0 j+k=n
Addition of these expressions gives, after some rearrangement of terms,
y + py + qy =F (α)xα−2
∞
n−1

+ F (n + α)cn + (k + α)an−k + bn−k ck xn+α−2 ,
n=1 k=0
where F (α) = α(α − 1) + a0 α + b0 . Therefore, in order that y be a solution

to the differential equation, we must require that the coefficient of xn+α−2
vanish for every integer n = 0, 1, 2, . . . . Thus the requirements are that

F (α) = 0 and

n−1

(14) F (n + α)cn = − (k + α)an−k + bn−k ck , n = 1, 2, . . . .
k=0
The quadratic equation F (α) = 0 is known as the indicial equation

associated with the given differential equation. Let α1 and α2 denote its
solutions, real or complex. There are then 4 cases to be treated separately.
Case 1. α1 and α2 are real and distinct, and do not differ by an integer.
Then set α = α1 in (14). Starting with c0 = 1, this equation determines
recursively the coefficients cn = cn (α1 ), because in this case F (n+α1 ) = 0 for
all n = 1, 2, . . . . The same procedure serves to determine coefficients cn =
cn (α2 ). In this manner we have constructed a pair of linearly independent
formal (generalized) power series solutions
∞

y = xα j cn (αj ) xn , j = 1, 2, .
n=0
It can be proved that each of these formal power series actually converges
in some neighborhood of the regular singular point x0 = 0, and therefore
represents an actual solution to the differential equation.
Case 2. α1 and α2 are real and α1 = α2 . Then only one solution can
be obtained by the process just described. However, F (α1 ) = F (α1 ) = 0,
since α1 is a double root of the polynomial F (α). This observation allows
us to obtain a second (independent) solution by a process of differentiation
with respect to α. For arbitrary real α near α1 , let the functions c0 (α) = 1,
c1 (α), c2 (α), . . . be defined recursively through the equation (14). Then each
coefficient cn (α) is a rational function of α, the quotient of two polynomials.
Consider the function
∞

y = y(x, α) = xα cn (α) xn .
n=0
Because of the way the coefficients cn (α) were constructed, we see that
yxx (x, α) + p(x)yx (x, α) + q(x)y(x, α) = F (α) xα−2 ,
where the subscripts denote partial derivatives. Once more it is clear that
y(x, α1 ) satisfies the differential equation, but we can say more. Differenti-
ation with respect to α gives
(yα )xx + p(yα )x + q yα = [F (α) log x + F (α)] xα−2 .

Because F (α1 ) = F (α1 ) = 0, this shows that

∞

yα (x, α1 ) = y(x, α1 ) log x + xα1 cn (α1 )xn
n=0
is a formal solution to the differential equation, and it can be shown to be

an actual solution; in other words, the power series converges in some open
interval about the origin.
Case 3. α1 and α2 are real and distinct, and differ by an integer. Number
the two roots so that α2 = α1 + m, where m is a positive integer. Then
a solution y(x, α2 ) can be constructed by the method described in Case 1.
However, if we make a similar attempt to use the relation (14) for recursive
determination of the coefficients cn (α1 ), we encounter a major obstacle when
n = m, since F (m + α1 ) = F (α2 ) = 0. Thus a direct construction fails, and
we are in search of a second solution.
It is here that the method of Frobenius comes to the rescue. Consider
an expansion of the form
∞

y = ϕ(x, α) = (α − α1 )x + x
α α
c n xn .
n=1
This amounts to replacing c0 = 1 by (α − α1 ). With this expression for y, a

formal calculation gives, as before,
y + py + qy = (α − α1 )F (α) xα−2
∞
+ F (n + α)cn + (αan + bn )(α − α1 )
n=1

n−1

+ (k + α)an−k + bn−k ck xn+α−2 ,
k=1
where the last sum is absent for n = 1. For α near α1 , we now use the
relation

n−1

(15) F (n + α)cn = −(αan + bn )(α − α1 ) − (k + α)an−k + bn−k ck
k=1
to generate recursively the coefficients cn = cn (α) for n = 1, 2, . . . . It is

easily seen that for n = 1, 2, . . . , m − 1, the functions cn (α) all carry (α − α1 )
as a factor. Since the polynomial F (α) factors as F (α) = (α − α2 )(α − α1 ),
and α2 = α1 + m, we see that
F (m + α) = (α − α1 )(α + m − α1 ) ,
and so it follows that cm (α) is well determined at α = α1 . Indeed, we have

only to divide the equation (15) by (α − α1 ) and set α = α1 . Consequently,
cn (α1 ) is defined by (15) for all n, and
ϕxx (x, α) + p(x)ϕx (x, α) + q(x)ϕ(x, α) = (α − α1 )F (α) xα−2 .
In particular, y = ϕ(x, α1 ) is a solution of the differential equation, but it is
not in general independent of y(x, α2 ). To obtain an independent solution,
differentiate the last equation with respect to α. This shows that y =
ϕα (x, α1 ) is a solution.
Case 4. α1 and α2 are (nonreal) complex conjugates: α2 = α1 . Then the
procedure of Case 1 can be used to generate two complex-valued solutions
y(x, α1 ) and y(x, α2 ) = y(x, α1 ) .
For a pair of independent real-valued solutions, take real and imaginary
parts.
Bessel’s equation xy + y + xy = 0 offers an important illustration of
the method described in Case 2, the easier case for the method of Frobenius.
Here xp(x) = 1 and x2 q(x) = x2 . Thus a0 = 1 and an = 0 for all n ≥ 1;
whereas b2 = 1 and bn = 0 for all n = 2. Therefore, the indicial polynomial
is
F (α) = α(α − 1) + α = α2 ,
with roots α1 = α2 = 0. The recurrence relation (14) reduces to c1 (α) = 0
and
cn (α) = −(n + α)−2 cn−2 (α) , n = 2, 3, . . . .
It follows recursively that cn (α) = 0 for all odd indices n. Also, since
c0 (α) = 1, recursion leads to the formula
c2k (α) = (−1)k (2 + α)−2 (4 + α)−2 · · · (2k + α)−2

for the coefficients of even index n = 2k. Differentiation gives the expression

c2k (0) = (−1)k+1 2−2 4−2 · · · (2k)−2 1 + 12 + · · · + k1 .
These calculations show that the two solutions y(x, 0) and yα (x, 0) are
∞

J0 (x) = 1 + (−1)k 2−2 4−2 · · · (2k)−2 x2k ,
k=1
as calculated previously, and

∞

Y0 (x) = J0 (x) log x + (−1)k+1 2−2 4−2 · · · (2k)−2 1 + 1
2 + ···+ 1
k x2k .
k=1
0.5
5 10 15 20 25 30
-0.5
-1
-1.5
Figure 2. Graph of Bessel’s function Y0 (x).
Both power series converge for all x ∈ R. The functions J0 (x) and Y0 (x) are
known as Bessel functions of first and second kinds, respectively. Figure 2
shows the graph of Y0 (x). It should be remarked that conventions differ and
the definition of Y0 (x) is not entirely standard in the literature.
Georg Frobenius (1849–1917) was a student of Weierstrass in Berlin.

After serving for many years as a professor in Zürich, he returned to the
University of Berlin as one of the leading mathematicians of his day. His
work spanned several areas of mathematics but he is perhaps best known
for work in algebra, specifically in the theory of groups.
13.4. Bessel functions

Bessel’s equation xy +y +xy = 0 arises in the study of physical problems
that involve circular disks or cylinders. As an illustration we now discuss the
vibrations of a circular drumhead. Suppose an elastic membrane is clamped
along a circular boundary of unit radius as it undergoes vibrations. Denote
by u(r, θ, t) the vertical displacement of the membrane at time t above the
point (r, θ), described in polar coordinates. The displacement is governed
by the wave equation, which takes the form
∂ 2u ∂ 2 u 1 ∂u 1 ∂ 2u
= + +
∂t2 ∂r2 r ∂r r2 ∂θ2
if the units of time and length are chosen appropriately. The boundary
condition u(1, θ, t) = 0 applies, since the membrane is clamped at its circular
edge. The motion is then determined by initial position and velocity. If the
initial conditions are circularly symmetric (i.e., independent of θ), so that
∂u
u(r, θ, 0) = f (r) and (r, θ, 0) = g(r) ,
∂t
then the subsequent motion will also be circularly symmetric. In other

words, the vertical displacement is independent of θ and can be written as
u(r, t). Thus the wave equation reduces to
∂2u ∂ 2 u 1 ∂u
2
= + , u(1, t) = 0 .
∂t ∂r2 r ∂r
We look for a particular solution of the form u(r, t) = R(r)T (t). Then
the wave equation gives R T = [R + (1/r)R ]T . Separating the variables,
we can write
T R + (1/r)R
= .
T R
The last equation says that a function of t is identically equal to a function
of r, so both functions must be constant. The constant is negative, because
otherwise the differential equation for T would have unbounded solutions.
Writing the constant as −λ2 for λ > 0, we see that R and T satisfy the
differential equations
rR + R + λ2 rR = 0 , T + λ2 T = 0 .
It is easily verified that the solutions of the first equation have the form
R(r) = y(λr) where y(x) is a solution of Bessel’s equation xy + y + xy = 0.
We have seen that every solution of Bessel’s equation is a linear combination
of J0 (x) and Y0 (x), the Bessel functions of first and second kind. But Y0 (x)
is unbounded near the origin, where it has a logarithmic singularity. Conse-
quently, the only bounded solutions, hence the only physically relevant ones,
are constant multiples of R(r) = J0 (λr). On the other hand, the solutions
of the differential equation T + λ2 T = 0 are linear combinations of cos λt
and sin λt. We conclude that the relevant solutions u(r, t) = R(r)T (t) have
the form
u(r, t) = J0 (λr) a cos λt + b sin λt
for arbitrary constants a and b. The boundary condition u(1, t) = 0 requires
that J0 (λ) = 0. In other words, λ must be a zero of the Bessel function
J0 (x).
As can be surmised from the graph in Figure 1, and will be proved later,
the Bessel function J0 (x) has infinitely many positive zeros λn , ordered so
that 0 < λ1 < λ2 < . . . . With the boundary condition taken into account,
we have arrived at the particular solutions

un (r, t) = J0 (λn r) an cos λn t + bn sin λn t , n = 1, 2, . . . .
By the superposition principle, any sum of such functions is again a solution.
We want to choose the coefficients an and bn so that the series
∞

(16) u(r, t) = J0 (λn r) an cos λn t + bn sin λn t
n=1
converges and satisfies the prescribed initial conditions u(r, 0) = f (r) and
ut (r, 0) = g(r). Formally, these requirements reduce to
∞
∞

an J0 (λn r) = f (r) and λn bn J0 (λn r) = g(r) .
n=1 n=1
Here we encounter an expansion problem similar to that in the theory of

Fourier series. The coefficients an and bn are calculated by means of the
orthogonality relation
1
(17) rJ0 (λn r)J0 (λm r) dr = 0 , n = m .
0
For a proof of (17), recall that yn (x) = J0 (λn x) satisfies the differential
equation
d
xyn + λ2n xyn = 0 .
dx
Multiply the equation by ym and integrate by parts to see that
1 1
d
2
λn xyn (x)ym (x) dx = − ym (x) xyn
0 0 dx
1 1

= − xyn (x)ym (x) + xyn (x)ym

(x) dx
0 0
1
= xyn (x)ym

(x) dx ,
0
because ym (1) = J0 (λm ) = 0. Now interchange n and m and subtract one

equation from the other to conclude that
1
(λn − λm )
2 2
xyn (x)ym (x) dx = 0 ,
0
which implies (17) since λ2n = λ2m .

A similar calculation shows that
1
xJ0 (λn x)2 dx = 1
2 J0 (λn )2 .
0
(See Exercise 8 for further details.) Consequently, the equations

1
1 2
2 J0 (λn ) an = rJ0 (λn r)f (r) dr and
0
1
1 2
2 J0 (λn ) λn bn = rJ0 (λn r)g(r) dr
0
determine the “Fourier–Bessel” coefficients an and bn that should be inserted

into the series (16) to solve the problem of the vibrating drumhead. A
rigorous treatment of the convergence problem for Fourier–Bessel series is
beyond the scope of this book, but can be found for instance in Watson [13].
A more general form of Bessel’s differential equation is
x2 y + xy + (x2 − n2 )y = 0 ,
where n is a real parameter. If n is a positive integer, one solution is
(−1)k x n+2k
∞

Jn (x) =
k!(n + k)! 2
(18) k=0

xn x2 x4
= n 1− + − ...
2 n! 2(2n + 2) 2 · 4(2n + 2)(2n + 4)
The graph of J1 (x) is shown in Figure 3. It is an odd function with infin-

itely many zeros. Bessel functions of negative integer order are defined by
J−n (x) = (−1)n Jn (x).
The integral representations
π
1 2n n! n
(19) Jn (x) = x cos(x cos θ) sin2n θ dθ
π (2n)! 0
and
π
1
(20) Jn (x) = cos(x sin θ − nθ) dθ
π 0
0.6
0.4
0.2
5 10 15 20 25 30
-0.2
Figure 3. Graph of Bessel’s function J1 (x).

are valid for n = 0, 1, 2, . . . , with the usual convention that 0! = 1. For

n = 0 the two representations reduce to virtually the same formula
π π
1 1
J0 (x) = cos(x cos θ) dθ = cos(x sin θ) dθ .
π 0 π 0
Each of the representations (19) and (20) for the Bessel function Jn (x) can
be verified by showing that the right-hand side satisfies Bessel’s differential
equation and is finite and properly normalized at the origin (see Exercises
10, 13, and 14). The representation (20) is often called Bessel’s formula. A
standard calculation in the theory of functions of a complex variable (contour
integration to find the coefficients in a Laurent expansion) produces Bessel’s
formula as a consequence of the remarkable generating relation
∞

x 1
(21) e 2 (t− t ) = Jn (x) tn .
n=−∞
Conversely, the generating relation can be established as a consequence of

Bessel’s formula. However, the generating relation (21) can also be derived
in a more elementary way, without resort to methods of complex analysis.
One such derivation proceeds as follows.
Recalling that J−n (x) = (−1)n Jn (x) and introducing the series (18), we
can write
∞
∞
∞

Jn (x) t =n n
Jn (x) t + (−1)n Jn (x) t−n
n=−∞ n=0 n=1
∞
∞
(−1)k x n+2k
= tn
k!(n + k)! 2
n=0 k=0
(−1)n+k x n+2k −n
∞
∞
+ t = S1 + S2 .
k!(n + k)! 2
n=1 k=0
Now set m = n + 2k and rewrite the sums as
∞ [m/2]
(−1)k m
m−2k x
S1 = 1 + t and
k!(m − k)! 2
m=1 k=0
(−1)m−k 2k−m x m
∞ [(m−1)/2]

S2 = t ,
k!(m − k)! 2
m=1 k=0
where [α] denotes the largest integer less than or equal to α. On the other
hand, the binomial theorem gives

m
m
1 m!
t− = (−1)k tm−2k
t k!(m − k)!
k=0

[m/2]
m!
[(m−1)/2]
m!
= (−1)k tm−2k + (−1)m−j t2j−m ,
k!(m − k)! (m − j)!j!
k=0 j=0
with the substitution k = m − j in the second part of the sum. Therefore,

1 m x m
∞
1 x 1
S1 + S2 = 1 + t− = e 2 (t− t ) ,
m! t 2
m=1
which is the desired result (21). (For another proof, see Exercise 15.)
It was essentially through the generating relation that the functions now
called Bessel functions originally arose in 1817, when Bessel was studying
perturbations of astronomical orbits. Other mathematicians had previously
encountered special cases of Bessel functions, but Bessel developed their
properties in a systematic treatise published in 1824.
Friedrich Wilhelm Bessel (1784–1846) had an unusual life history. He
left school at age 14 to become apprentice to a shipping merchant in Bre-
men. Thereafter he was mainly self-educated. Attracted by the problem
of determining longitude at sea, he developed a keen interest in astronomy.
His work drew the attention of professional astronomers, and at age 26 he
became director of the observatory in Königsberg, where he remained for
the rest of his career. Using sophisticated mathematical methods, he made
important contributions to astronomy, including improved data on the po-
sitions of stars.
13.5. Hypergeometric functions

The hypergeometric differential equation is
(22) x(1 − x)y + [c − (a + b + 1)x]y − ab y = 0 ,
where a, b, and c are real parameters. There are many important special
cases. It has regular singular points at 0 and 1, and (with suitable interpre-
tation) at ∞, and it is in a sense the most general linear differential equation
of second order with three regular singular points. In the notation of Section
13.3,
c − (a + b + 1)x ab x
xp(x) = and x2 q(x) = − ,
1−x 1−x
so the indicial equation at the origin is
F (α) = α(α − 1) + a0 α + b0 = α2 + (c − 1)α = 0 ,
with roots α1 = 0 and α2 = 1−c. In particular, the hypergeometric equation

admits a power series solution of the form
∞

y =1+ γk xk , |x| < 1 ,
k=1
provided that α2 = 1 − c is not a positive integer. The coefficients γk are

determined by inserting the series into the equation (22) to obtain after
some manipulation
∞

(k + 1)(k + c)γk+1 − (k + a)(k + b)γk xk = 0 .
k=0
Since the coefficient of every power xk must vanish, this gives the recurrence
relation
(k + 1)(k + c)γk+1 = (k + a)(k + b)γk , k = 0, 1, 2, . . . .
Under the assumption that 1 − c is not a positive integer, or that c =

0, −1, −2, . . . , it is clear that k + c = 0 for all k = 0, 1, 2, . . . . Thus with
γ0 = 1 the recurrence relation leads to the formula
a(a + 1) · · · (a + k − 1) b(b + 1) · · · (b + k − 1)
γk = , k = 1, 2, . . . .
k! c(c + 1) · · · (c + k − 1)
It is convenient to adopt the notation
(a)k = a(a + 1) · · · (a + k − 1) , with (a)0 = 1 ,
for a shifted factorial. Then the solution, known as a hypergeometric func-

tion, can be written more compactly as
∞
(a)k (b)k
(23) y = F (x) = F (a, b; c; x) = 1 + xk .
k! (c)k
k=1
In terms of the gamma function, the shifted factorial is (a)k = Γ(a + k)/Γ(a) .
An application of the ratio test confirms that the power series has radius of
convergence equal to 1 unless either a or b is a negative integer, in which
case the series terminates and F (x) is a hypergeometric polynomial.
Many important functions of mathematical analysis can be expressed as

hypergeometric series. For instance, the geometric series is
∞
1
=1+ xk = F (1, 1; 1; x) , whereas
1−x
k=1
log(1 + x) = x F (1, 1; 2; −x) ,
sin−1 x = x F ( 12 , 12 ; 32 ; x2 ) ,
tan−1 x = x F ( 12 , 1; 32 ; −x2 ) .
Other examples are complete elliptic integrals and Jacobi polynomials, which
include the Legendre polynomials and Chebyshev polynomials as special
cases.
Gauss made the first systematic study of hypergeometric functions in a
famous paper of 1812, but Euler had found a basic integral representation
in 1769.
Euler integral representation. If c > b > 0, then

1
Γ(c)
F (a, b; c; x) = tb−1 (1 − t)c−b−1 (1 − xt)−a dt , |x| < 1 .
Γ(b)Γ(c − b) 0
Proof. The restriction of parameters to c > b > 0 ensures that the integral
converges. For fixed x with |x| < 1, expand the last factor into binomial
series:
∞

−a −a
(1 − xt) = (−xt)k ,
k
k=0
where

−a (−a)(−a − 1) · · · (−a − k + 1) (−1)k (a)k
= = .
k k! k!
Since the binomial series converges uniformly in the interval 0 ≤ t ≤ 1, the

integration can be performed term by term, so that
∞

1
−a
1
−a
t b−1
(1 − t) c−b−1
(1 − xt) dt = t b+k−1
(1 − t) c−b−1
dt (−x)k
0 0 k
k=0
∞

∞

−a Γ(b + k) Γ(c−b) (−1)k (a)k
= B(b + k, c−b) (−x)k = (−x)k
k Γ(c + k) k!
k=0 k=0
∞
Γ(b) Γ(c − b) (a)k (b)k k Γ(b) Γ(c − b)
= x = F (a, b; c; x) ,
Γ(c) k!(c)k Γ(c)
k=0
which verifies Euler’s integral formula.
Hypergeometric functions satisfy a wide variety of interesting identities.

We will give only a few examples here. Much more comprehensive treat-
ments can be found in books on special functions such as Rainville [11] or
Andrews, Askey, and Roy [1].
It is obvious from the definition (23) that F (b, a; c; x) = F (a, b; c; x), but
strangely this symmetry is not apparent in Euler’s integral representation.
The change of variable t = 1 − s in Euler’s integral gives
1
Γ(c)
F (a, b; c; x) = sc−b−1 (1 − s)b−1 (1 − x + xs)−a ds
Γ(b) Γ(c − b) 0

−a
Γ(c) (1 − x)−a 1 c−b−1 x
= s (1 − s) b−1
1− s ds
Γ(b) Γ(c − b) 0 x−1

−a x
= (1 − x) F a, c − b; c; ,
x−1
provided that c > b > 0, |x| < 1, and |x/(x − 1)| < 1. This identity is
called Pfaff ’s transformation. After interchange of the first two parameters,
another application of Pfaff’s transformation gives

x x
F a, c − b; c; = F c − b, a; c; = (1−x)c−b F (c−b, c−a; c; x)
x−1 x−1
if c > a > 0, with the conclusion that
F (a, b; c; x) = (1 − x)c−a−b F (c − a, c − b; c; x) .
The last identity was known to Euler.
The hypergeometric series (23) need not converge at the point x = 1.
However, it can be shown to converge there if c > a + b. For a proof, recall
the Gauss product formula (Section 9.6)
k! k a
Γ(a) = lim , a > 0,
k→∞ a(a + 1) · · · (a + k)
which implies that
k! k a−1
(a)k ∼ , k → ∞.
Γ(a)
This asymptotic relation fails if a < 0, but it can be applied to show that
(a)k ∼ M k! k a−1 for some constant M depending only on a. Therefore,
(a)k (b)k
(24) ∼ M k a+b−c−1 , k → ∞,
k!(c)k
where the constant M depends only on a, b, and c. In particular, the series
(23) converges for x = 1 if c > a + b.
The actual value of F (a, b; c; 1), when the series converges, turns out to
be important. Gauss evaluated the sum in his paper of 1812.
Gauss summation formula. If c > a + b, then

∞
(a)k (b)k Γ(c) Γ(c − a − b)
F (a, b; c; 1) = = .
k! (c)k Γ(c − a) Γ(c − b)
k=0
We will give two proofs. The first uses Euler’s integral representation and
requires the additional assumption that c > b > 0. Euler’s representation is
easily seen to give
1
Γ(c)
lim F (a, b; c; x) = tb−1 (1 − t)c−a−b−1 dt
x→1− Γ(b) Γ(c − b) 0
Γ(c) Γ(c) Γ(c − a − b)
= B(b, c − a − b) = .
Γ(b) Γ(c − b) Γ(c − a) Γ(c − b)
Because the condition c > a + b ensures that the series converges at x = 1,
we can infer from Abel’s theorem that its sum agrees with its Abel limit, so
the Gauss summation formula is proved for c > b > 0.
The second proof shows that the condition c > b > 0 is superfluous.
Assuming only that c > a + b, we begin by establishing the identity
(c − a)(c − b)
(25) F (a, b; c; 1) = F (a, b; c + 1; 1) .
c(c − a − b)
For a proof of (25), denote the terms of the two series by
(a)k (b)k (a)k (b)k
Ak = and Bk = ,
k! (c)k k! (c + 1)k
respectively. Then
(a + k)(b + k) c
Ak+1 = Ak and Bk = Ak ,
(k + 1)(c + k) c+k
and a straightforward calculation confirms that

c(c − a − b)Ak = (c − a)(c − b)Bk + c kAk − (k + 1)Ak+1 .
Consequently,

n
n
c(c − a − b) Ak = (c − a)(c − b) Bk − c(n + 1)An+1 ,
k=0 k=0
since the last sum telescopes. But because c > a+b, the asymptotic relation
(24) shows that
(n + 1)An+1 ∼ M na+b−c → 0 as n → ∞ ,
and so (25) follows. Iterate the relation (25) to obtain
(c − a)n (c − b)n
F (a, b; c; 1) = F (a, b; c + n; 1) ,
(c)n (c − a − b)n
or
(26)
Γ(c − a) Γ(c − b) Γ(c − a + n) Γ(c − b + n)
F (a, b; c; 1) = F (a, b; c + n; 1)
Γ(c) Γ(c − a − b) Γ(c + n) Γ(c − a − b + n)
for n = 1, 2, . . . . Now pass to the limit as n → ∞. An application of
Stirling’s formula for the gamma function (cf. Chapter 9, Exercise 22) shows
that
Γ(c − a + n) Γ(c − b + n)
lim = 1.
n→∞ Γ(c + n) Γ(c − a − b + n)
We leave it as an exercise to show that also F (a, b; c + n; 1) → 1 as n → ∞.

Then the right-hand side of (26) tends to 1 as n → ∞, and the Gauss
summation formula is proved.
As a final example, we derive the identity
√
F (a, b; a + b + 12 ; 4x(1 − x)) = F (2a, 2b; a + b + 12 ; x) , 2 (1 −
1
2) < x < 1 ,
1
where a + b + 2 is not a negative integer or zero. The proof uses the fact
that
y(t) = F (a, b; a + b + 12 ; t)
satisfies the hypergeometric differential equation
t(1 − t)y + [a + b + 1
2 − (a + b + 1)t]y − ab y = 0 .
With the substitution t = 4x(1 − x), a calculation then shows, in view of

the relation (1 − 2x)2 = 1 − t, that the function
Y (x) = y(4x(1 − x)) = F (a, b; a + b + 12 ; 4x(1 − x))
satisfies the hypergeometric equation
x(1 − x)Y + [a + b + 1
2 − (2a + 2b + 1)x]Y − 4ab Y = 0 .
But since c = a + b + 12 is not a negative integer or zero, the only power

series solution, up to a constant factor, is Y (x) = F (2a, 2b; a+b+ 12 ; x). This
√
gives the desired result. The restriction to the interval 12 (1 − 2) < x < 1
guarantees that |4x(1−x)| < 1 except at x = 12 , where the identity is proved
by a continuity argument. Note that a+b+ 12 > a+b, so the hypergeometric
series F (a, b; a + b + 12 ; t) actually converges for t = 1.
13.6. Oscillation and comparison theorems

When confronted with a differential equation, it is customary to search
for explicit formulas for the solutions. But often the main interest lies in
qualitative properties, and explicit formulas may be too complicated to shed
light on the behavior of solutions. A case in point is Bessel’s function
x2 x4 x6
J0 (x) = 1 − + − + ... ,
22 22 42 22 42 62
which is a solution to Bessel’s equation xy + y + xy = 0 . Numerical
calculations lead to the graph shown in Figure 1. The graph suggests that
J0 (x) has infinitely many zeros, rather evenly spaced, and it oscillates with
decreasing amplitude, tending to zero as x → ∞. However, the power series
representation is of little use in confirming such conjectures. On the other
hand, we will see that the Sturm comparison theorem and related results
allow us to verify all of this information directly from the fact that J0 (x)
satisfies Bessel’s differential equation.
In fact, the method has broad applications. Many classical problems of
mathematical physics lead to linear differential equations of second order. It
was Sturm’s great insight that such an equation need not be solved explicitly
in order to determine the qualitative behavior of its solutions.
Sturm’s analysis is clever but entirely elementary. He developed the core
of the theory in a famous paper [12] of 1836, but his ideas have continued to
be refined and extended. Although we propose to discuss only the most basic
of Sturm’s results, they will reveal some remarkably precise information
about Bessel functions.
Consider a linear differential equation with variable coefficients, ex-
pressed in “self-adjoint” form
(27) (py ) + qy = 0 ,
where q(x) is continuous, p(x) is continuously differentiable, and p(x) > 0
over an interval [a, b]. Observe first that each solution of (27) that is not
identically zero can have at most a finite number of zeros in the closed
bounded interval [a, b]. Indeed, if there were infinitely many zeros they
would have to cluster at some point in [a, b], and so by Rolle’s theorem, the
zeros of the derivative would also cluster at the same point. The function
would then vanish together with its derivative at this point, and so would
vanish identically in [a, b], since it is a solution of a differential equation (27)
of second order. Thus each solution of (27) has at most a finite number of
zeros in the interval, and all zeros are simple.
It is a familiar fact that the zeros of sin x and cos x are interlaced; they
occur alternately as x increases. But these two functions are independent
solutions of the differential equation y + y = 0, a very special form of the

equation (27). Sturm discovered that the interlacing phenomenon is actually
quite general.
Sturm Separation Theorem. Let y1 and y2 be independent solutions of

the differential equation (27), and let x1 and x2 be consecutive zeros of y1
in [a, b]. Then y2 vanishes exactly once in the interval (x1 , x2 ).
To say that x1 and x2 are consecutive zeros of y1 means that y1 (x1 ) =

y1 (x2 ) = 0 but y1 (x) = 0 for x1 < x < x2 .
Proof of theorem. Because the two solutions y1 and y2 are independent,

they cannot vanish simultaneously. Indeed, if both functions were to vanish
at some point in [a, b], then every solution would vanish at that point, which
is certainly not true. Hence, y2 (x1 ) = 0 and y2 (x2 ) = 0. Suppose now
that y2 (x) = 0 throughout the interval (x1 , x2 ). Then the quotient v =
y1 /y2 is continuously differentiable in [x1 , x2 ] and v(x1 ) = v(x2 ) = 0. By
Rolle’s theorem, it follows that v (ξ) = 0 at some intermediate point ξ ∈
(x1 , x2 ). But v = (y2 y1 − y1 y2 )/y22 , so this says that the Wronskian of y1
and y2 vanishes at the point ξ. This is impossible, however, since the two
solutions are independent. Thus the supposition that y2 (x) = 0 has led to a
contradiction, and we conclude that y2 (x) = 0 at some point in the interval
(x1 , x2 ). If y2 had two zeros in that interval, we could apply the result
just established, with the roles of y1 and y2 reversed, to see that y1 (x) = 0
somewhere between x1 and x2 , contrary to the hypothesis that these two
points are consecutive zeros of y1 . This concludes the proof.
The next theorem compares zeros of solutions of different but closely

related differential equations of the form (27).
Sturm Comparison Theorem. Let y1 and y2 be solutions of the respec-

tive differential equations (py ) + q1 y = 0 and (py ) + q2 y = 0. Let x1 and
x2 be consecutive zeros of y1 , and suppose that q1 (x) ≤ q2 (x) in [x1 , x2 ],
with strict inequality throughout some subinterval. Then y2 (x) = 0 for some
point x in the open interval (x1 , x2 ).
In particular, if y2 (x1 ) = 0, then y2 vanishes again before y1 does; it

“oscillates faster”. The theorem is best illustrated by the two equations
y + n2 y = 0 and y + m2 y = 0 with m > n > 0. Then the solutions are
sinusoids with regularly spaced zeros at respective distances π/n and π/m
apart.
Proof of theorem. If y2 (x) = 0 in (x1 , x2 ), we may assume without loss

of generality that y1 (x) > 0 and y2 (x) > 0 in (x1 , x2 ). Then since y1 (x1 ) =
y1 (x2 ) = 0, it follows that y1 (x1 ) > 0 and y1 (x2 ) < 0; while y2 (x1 ) ≥ 0
and y2 (x2 ) ≥ 0. Observe now that because y1 and y2 satisfy their respective
differential equations, we have the identity

p(y2 y1 − y1 y2 ) = y2 (py1 ) − y1 (py2 ) = (q2 − q1 )y1 y2 .
Integration gives
x2 x2
p(y2 y1 − y1 y2 ) = q2 (x) − q1 (x) y1 (x)y2 (x) dx > 0 ,
x1 x1
in view of the hypothesis that q1 (x) ≤ q2 (x) with strict inequality in some
subinterval, and the assumptions that y1 (x) > 0 and y2 (x) > 0 in (x1 , x2 ).
On the other hand, since y1 vanishes at both x1 and x2 , the left-hand side
reduces to
x2
p y2 y1 = p(x2 )y2 (x2 )y1 (x2 ) − p(x1 )y2 (x1 )y1 (x1 ) ≤ 0 ,
x1
by virtue of the standing assumption that p(x) > 0 and the supposed in-
equalities y2 (x1 ) ≥ 0, y2 (x2 ) ≥ 0, y1 (x1 ) > 0, and y1 (x2 ) < 0. This con-
tradicts the previous conclusion that the right-hand side is positive, thereby
invalidating our initial supposition that y2 (x) = 0 throughout the interval
(x1 , x2 ). Consequently, y2 (x) = 0 for some point x in (x1 , x2 ), as the theorem
asserts.
We now apply the Sturm comparison theorem to study the behavior

of the Bessel function J0 , which is a solution of the differential equation
(xy ) + xy = 0. Since J0 is an even function, it is enough to consider x > 0.
We are going to prove that J0 has an infinite number of zeros, as suggested
by its graph in Figure 1.
In order to study the zeros of J0 , and in fact to prove their existence, we
first transform the Bessel equation through the substitution y(x) = xα u(x),
where α is a real parameter to be specified shortly. A simple calculation
shows that

xy + y + xy = xα+1 u + (2α + 1)xα u + α2 xα−1 + xα+1 u = 0 .
Choosing α = − 12 to eliminate the term in u , we find that the differential

equation reduces to

1
(28) u + 1 + 2 u = 0.
4x
As x → ∞, the equation (28) approaches in some sense the familiar equation
u + u = 0 , whose solutions are sinusoids. Thus we can expect the zeros of
u , which are the same as those of J0 (x) = x−1/2 u(x) , to behave for large x
like those of sin x and cos x.
These notions are made precise by appeal to the Sturm comparison
theorem. Since u(x) satisfies the differential equation u + qu = 0 with
q(x) = 1 + 1/(4x2 ) > 1 for all x > 0 ,
we can use Sturm’s theorem to compare the zeros of u(x) with those of
cos x, for instance, which satisfies the differential equation y + y = 0. The
theorem guarantees that u(x) has at least one zero between each pair of
zeros of cos x, which lie apart at distance π. We conclude in particular
that J0 (x) has infinitely many positive zeros λn , which can be ordered by
0 < λ1 < λ2 < . . . with λn → ∞ as n → ∞, since the zeros can have no
finite cluster point.
In fact, we can say more. Let y1 = cos(x − δ) be a solution to the
differential equation y + y = 0 which vanishes together with J0 at some
point λn . Then the next zero of y1 will occur at λn + π. But by the
√
comparison theorem, the function y2 (x) = u(x) = xJ0 (x) must vanish
between successive zeros of y1 . This shows that λn < λn+1 < λn + π, or
0 < λn+1 − λn < π for n = 1, 2, . . . .
But we can say even more. By reversing the roles of the two differential
equations, we can apply the Sturm comparison theorem to prove that λn+1 −
λn → π as n → ∞. For this it will suffice to show that for each ε > 0 there
is a number N such that λn+1 − λn > π/(1 + ε) for all n ≥ N . Given ε > 0,
choose N so large that
1 + 1/(4λ2n ) < (1 + ε)2 for all n ≥ N .
This is possible because λn → ∞, and it follows that
1 + 1/(4x2 ) < (1 + ε)2 for all x ≥ λn if n ≥ N .
Now define y2 (x) = cos((1 + ε)x − δ), choosing δ so that y2 (λn ) = 0. Then
since y2 satisfies the equation y + (1 + ε)2 y = 0, the Sturm comparison
√
theorem ensures that y2 (x) will vanish again before u(x) = xJ0 (x) does.
But the zeros of y2 are spaced at distance π/(1 + ε) apart, so this says that
λn < λn + π/(1 + ε) < λn+1 , or π/(1 + ε) < λn+1 − λn
for all n ≥ N . Because ε > 0 was chosen arbitrarily, and we have already
found that λn+1 − λn < π for all n, this implies that λn+1 − λn → π as
n → ∞.
Listed below are some approximate numerical values of the zeros λn .
λ1 = 2.4048256 λ2 = 5.5200781 λ3 = 8.6537279

λ4 = 11.7915344 λ5 = 14.9309177 λ6 = 18.0710640
λ7 = 21.2116366 λ8 = 24.3524715 λ9 = 27.4934791
λ10 = 30.6346065 λ11 = 33.7758202 λ12 = 36.9170984
Note that λ12 − λ11 ≈ 3.14128 , which is already very close to π ≈ 3.14159 .
The data suggest that λn+1 − λn increases to π, and this can be verified
analytically. (See Exercises 28 and 29.)
Charles-François Sturm (1803–1855) was born in Geneva but worked

mainly in Paris, where eventually he secured a professorship. He made
important contributions to several areas of science. In 1826 he produced
a prize essay on compressible fluids. In 1829 he found a basic algorithm
for locating the real roots of a polynomial, a method which Cauchy then
extended to complex roots. His seminal work on differential equations led
to a fruitful collaboration with Joseph Liouville (1809–1882). Together they
produced the celebrated Sturm–Liouville theory of eigenfunction expansions
associated with boundary-value problems for linear differential equations of
second order. The article by Loria [10] gives a more detailed account of
Sturm’s wide-ranging scientific career.
13.7. Refinements of Sturm’s theory

The proof of the Sturm comparison theorem, with slight modification,
actually leads to a stronger result. Suppose again that the functions q1 (x)
and q2 (x) are continuous on the interval (a, b), and that p(x) is continuously
differentiable there with p(x) > 0. Under these assumptions, we can state a
simple variant of the Sturm comparison theorem.
Comparison Theorem. Let y1 and y2 be solutions of the respective dif-

ferential equations (py ) + q1 y = 0 and (py ) + q2 y = 0, where q1 (x) < q2 (x)
in (a, b). Suppose that y1 (x1 ) = y2 (x1 ) ≥ 0 and y1 (x1 ) = y2 (x1 ) > 0
at some point x1 ∈ (a, b). Then y2 (x) < y1 (x) for all x in each interval
(x1 , x2 ) ⊂ (x1 , b) where y2 (x) = 0.
Proof. As in the proof of the Sturm comparison theorem, we can write

x x

(29) p(y2 y1 − y1 y2 ) = q2 (t) − q1 (t) y1 (t)y2 (t) dt .
x1 x1
The hypotheses guarantee that y2 (t) > 0 in (x1 , x2 ), and so it follows from
the Sturm comparison theorem that y1 (t) > 0 as well. Thus, in view of the
13.7. Refinements of Sturm’s theory 359
assumption that p(x) > 0, the relation (29) implies that
y2 (x)y1 (x) − y1 (x)y2 (x) > 0 for all x ∈ (x1 , x2 ) .
Therefore, the function y1 /y2 has positive derivative in this interval. But
y1 (x)/y2 (x) → 1 as x → x1 + (apply l’Hôpital’s rule if y1 (x1 ) = y2 (x1 ) = 0),
so we conclude that y1 (x)/y2 (x) > 1 in (x1 , x2 ), as claimed.
We turn next to a different refinement of Sturm’s basic theory that will

provide further information about Bessel functions. Using the Sturm com-
parison theorem, we have already confirmed that the Bessel function J0 (x)
has infinitely many zeros with essentially constant spacing, as suggested by
its graph in Figure 1. Other features of the graph suggest that the Bessel
function oscillates with decreasing amplitude, and that J0 (x) → 0 as x → ∞.
To verify these latter properties, we now develop an elementary result known
as the Sonin–Pólya oscillation theorem.
Sonin–Pólya Oscillation Theorem. Let y = y(x) be a solution of the

differential equation (py ) + qy = 0, where p(x) and q(x) are continuously
differentiable on an interval [a, b]. Suppose that p(x) > 0, q(x) = 0, and
(pq) (x) > 0 on (a, b). Then the successive local maxima (if any) of |y(x)|
form a decreasing sequence.
Proof. Consider the function

p(x) y (x)2
φ(x) = y(x)2 + .
q(x)
Writing φ = y 2 + (py )2 (pq)−1 and noting that (py ) = −qy, we find

2
2py
(pq) y
φ = 2yy + (py ) − (py )2 = − (pq) .
pq (pq)2 q
Since (pq) (x) > 0 on (a, b) by hypothesis, this shows that φ(x) is decreasing
on the interval [a, b]. In particular, if ξn are successive local maxima of
|y(x)|, so that a < ξ1 < ξ2 < · · · < b, then φ(ξ1 ) > φ(ξ2 ) > . . . . But
φ(ξn ) = y(ξn )2 , because y (ξn ) = 0 at each local maximum of |y(x)|. Thus
|y(ξ1 )| > |y(ξ2 )| > . . . , as the theorem asserts. Note that the inequalities
extend to the endpoints a or b if y (a) = 0 or y (b) = 0, and if p(x)/q(x)
tends to a finite limit as x → a+ or x → b−, respectively.
Applying the Sonin–Pólya theorem to Bessel’s equation (xy ) + xy = 0,

where p(x) = q(x) = x, we see at once that J0 (x) oscillates with strictly
decreasing amplitude. More precisely, if 0 < ξ1 < ξ2 < . . . are the successive
local maxima of |J0 (x)| in (0, ∞), then |J0 (ξ1 )| > |J0 (ξ2 )| > . . . . In fact,
it is easy to see from the differential equation that the points ξn interlace
with the zeros λn of J0 , since J0 (ξn ) = −J0 (ξn ) = 0 at each of these critical
points, so that J0 has a local maximum at ξn if J0 (ξn ) > 0 and a local
minimum if J0 (ξn ) < 0, as the graph in Figure 1 suggests. Recall that J0
cannot vanish at a critical point because it is a solution of a linear differential
equation of second order. Finally, we note that J0 has a local maximum the
origin, since J0 (0) = 0 and J0 (0) = − 12 < 0. Also J0 (0) = 1 > 0, and the
Sonin–Pólya theorem extends to show that J0 (0) > |J0 (ξ1 )| despite the fact
that the function p(x) = x vanishes at the origin, because p(x)/q(x) ≡ 1. In
particular, it follows that |J0 (x)| ≤ 1 for all x ≥ 0.
A similar elementary argument shows that J0 (x) → 0 as x → ∞. In
√
fact, it may be expected that J0 (x) decays like 1/ x, because as we have
√
seen, the substitution u(x) = x y(x) transforms Bessel’s equation to u +
1 + 4x1 2 u = 0, whose solutions behave like sinusoids for large x. To make
these notions precise, write the transformed equation as
1
u + Qu = 0 , where Q(x) = 1 + 2 ,
4x
and consider the function
ψ(x) = u (x)2 + Q(x)u(x)2 .
A simple calculation shows that ψ (x) = Q (x)u(x)2 ≤ 0, which implies that
ψ(x) is nonincreasing over the interval (0, ∞). In particular, if νn are the
successive local maxima of |u(x)|, with 0 < ν1 < ν2 < . . . , then
Q(νn )u(νn )2 = ψ(νn ) ≤ ψ(ν1 ) = Q(ν1 )u(ν1 )2 , n = 1, 2, . . . ,
since u (ν
n ) =√0. It follows that u(x) remains bounded as x → ∞, so that
J0 (x) = O(1/ x) as x → ∞. In particular, J0 (x) → 0 as x → ∞.
A more general form of the preceding argument appears in an article
by Duren and Muldoon [5]. Our presentation of the Sonin-Pólya theorem
is adapted from the book of Kreider, Kuller, and Ostberg [8]. For further
refinements of the Sturm theory, the reader is referred to Birkhoff and Rota
[2] and the classic book by Ince [6], and to the article by Laforgia and
Muldoon [9], which gives additional references.
Exercises
1. For the initial-value problem y = f (x, y), y(x0 ) = y0 , verify the error
estimate
M eCδ C n δ n+1
|ϕ(x) − ϕn (x)| ≤ , |x − x0 | < δ ,
(n + 1)!
as stated after the proof of the existence and uniqueness theorem in Section
13.1.
Exercises 361
2. Carry out the details in the proof of existence and uniqueness of a solution
to the initial-value problem (6) for first-order linear systems of differential
equations.
3. Show that the functions f (x) = x and g(x) = x2 are linearly independent
over the interval (−1, 1), yet their Wronskian vanishes at the origin.
4. Verify the Wronskian relation (11) for solutions to a differential equation
(7) of order n.
Hint. The derivative of an n × n determinant is equal to the sum of n
determinants, each involving the derivatives of one row. Loosely speaking,
the determinant can be differentiated row by row.
5. Use the method of reduction of order to derive a second solution, in-
dependent of J0 (x), for Bessel’s equation xy + y + xy = 0. Set y = J0 v
and derive a linear differential equation of first order for v. Solve the latter
equation to obtain

1
Y0 (x) = J0 (x) dx = J0 (x) log x + 14 x2 − 128
3 4
x + ... .
xJ0 (x)2

6. (a) Find the general power series solution y = ∞ k
k=0 ck x of Legendre’s
equation
(1 − x2 )y − 2xy + λy = 0 .
Show that the series is convergent for |x| < 1.
(b) Show that the Legendre equation has a (nontrivial) polynomial so-
lution if and only if λ = n(n + 1) for some integer n = 0, 1, 2, . . . . When
λ = n(n + 1), show that the polynomial solution is unique up to a constant
factor and has degree n. With standard normalization (choice of constant
factor), this solution is called the nth Legendre polynomial and is denoted
by Pn (x). The standard normalization is to require that Pn (1) = 1. Show
that Pn (−x) = (−1)n Pn (x), or in other words that the polynomial is an
even function if n is even, odd if n is odd.
(c) Verify that
d
(1 − x2 )Pn (x) = −n(n + 1)Pn (x) ,
dx
and use this relation to establish the orthogonality property
1
Pn (x) Pm (x) dx = 0 , n = m .
−1
Note. Legendre’s differential equation and Legendre polynomial expansions
(orthogonal series expansions analogous to Fourier series) occur naturally
when solving Laplace’s equation in a spherical region. The book by Jackson
[7] is a good reference.
7. (a) Show that the Legendre polynomial Pn (x) is orthogonal to every

polynomial Q(x) of degree lower than n :
1
Pn (x) Q(x) dx = 0 .
−1
(b) Show that every zero of Pn (x) in the interval (−1, 1) is simple.
Hint. Remember that Pn (x) satisfies a differential equation.
(c) Prove that the Legendre polynomial Pn (x) has exactly n (distinct)
zeros in the interval (−1, 1).
Hint. Suppose that Pn (x) = 0 in (−1, 1) only at points x1 , x2 , . . . , xk for
k < n, and apply the result of (a) to the polynomial Q(x) = (x − x1 )(x −
x2 ) · · · (x − xk ) .
8. (a) For any constant λ > 0, show that
1

x J0 (λx)2 dx = 12 J0 (λ)2 + J0 (λ)2 .
0
In particular, if λn > 0 is a zero of J0 (x), conclude that

1
x J0 (λn x)2 dx = 12 J0 (λn )2 .
0
Hint. Transform the first integral by substituting t = λx. Then show that
y = J0 (t) satisfies
d2 2 2
2ty 2 = t (y + y ) .
dt
(b) Generalize the result by proving that
1

x Jn (λx)2 dx = 12 Jn (λ)2 + 1 − n2
λ2
Jn (λ)2
0
for λ > 0 and n = 1, 2, . . . .

9. If λ and μ are any distinct positive zeros of Jn (x), prove the orthogonality
relation 1
x Jn (λx)Jn (μx) dx = 0 , n = 1, 2, . . . .
0
10. (a) For n = 1, 2, . . . , show that Bessel’s equation x2 y +xy +(x2 −n2 )y =
0 has a regular singular point at the origin with indicial equation α2 −n2 = 0.
(b) Derive the series expansion (18) for Jn (x) as a solution to Bessel’s
equation, and show that it is the only solution, up to constant multiples,
that is bounded near the origin.
Exercises 363
11. (a) Show that J0 (x) = −J1 (x) .

(b) For n = 1, 2, . . . , show that d
dx {x−n Jn (x)} = −x−n Jn+1 (x).
(c) For n = 1, 2, . . . , show that Jn (x) = 12 {Jn−1 (x) − Jn+1 (x)}.
12. Prove the recurrence relaton

2n
Jn+1 (x) = Jn (x) − Jn−1 (x) , n = 1, 2, . . . .
x
13. (a) Verify the integral representation

1 π
J0 (x) = cos(x cos θ) dθ .
π 0
by showing that the integral satisfies Bessel’s differential equation. Justify
the process of differentiation under the integral sign.
(b) Verify the same integral representation for J0 (x) by introducing the
Taylor expansion
(x cos θ)2 (x cos θ)4

cos(x cos θ) = 1 − + − ...
2! 4!
and integrating term by term. Justify the term-by-term integration and
apply the formula in Chapter 2, Exercise 15. (See also Section 2.5.)
(c) Verify the integral representation (19) for Jn (x), where n = 1, 2, . . . .
14. Verify Bessel’s formula (20) for Jn (x), where n = 1, 2, . . . .

Suggestion. If In (x) denotes the integral, show by differentiation under the
integral sign that
π
2 d
x In (x)+xIn (x)+(x −n )In (x) = −
2 2
(n+x cos θ) sin(nθ−x sin θ) dθ .
0 dθ
15. Write
x 1 xt x
e 2 (t− t ) = e 2 e− 2t
and multiply the power series expansions of the two exponential functions
to recover the generating relation (21) for Bessel functions.
16. Differentiate the generating relation (21) term by term to obtain proofs
of the identities
x
Jn (x) = 12 Jn−1 (x) − Jn+1 (x) and n Jn (x) = Jn−1 (x) + Jn+1 (x) ,
2
where n = 1, 2, . . . . Justify the term-by-term differentiation.
17. Show that

2 2
y = J1/2 (x) = sin x and y = J−1/2 (x) = cos x
πx πx
are solutions of Bessel’s equation x2 y + xy + (x2 − 14 )y = 0 for x > 0.

Suggestion. First transform the differential equation by the substitution
√
u = x y.
18. The hypergeometric function F (a, b; c; x) is a solution of the hypergeo-

metric differential equation (22) if c is not a negative integer or zero. If c is
not an integer, show that
x1−c F (a − c + 1, b − c + 1; 2 − c; x)
is a second solution, independent of F (a, b; c; x), on the interval 0 < x < 1.
19. Show that the Legendre polynomials are hypergeometric functions

Pn (x) = F −n, n + 1; 1; 1−x
2 .
Recall that Pn (x) is the polynomial solution of Legendre’s differential equa-

tion with parameter λ = n(n + 1), normalized by Pn (1) = 1. (See Exercise
6.)
20. Verify the relations
1+x
(a) log = 2x F 12 , 1; 32 ; x2 ,
1−x

(b) (1 + x)−2a + (1 − x)−2a = 2 F a, a + 12 ; 12 ; x2 ,
√ −2a
(c) 1+ 1−x = 2−2a F a, a + 12 ; 2a + 1; x .
21. Show that

d ab
(a) F (a, b; c; x) = F (a + 1, b + 1; c + 1; x) ,
dx c
d a
(b) x F (a, b; c; x) = axa−1 F (a + 1, b; c; x) .
dx
22. Suppose that a + b + 1 = 2c and let y(x) be a solution of the hyperge-

ometric differential equation (22) on the interval 0 < x < 1. Show that the
function u(x) = y(1 − x) is also a solution.
23. Prove that F (a, b; c + n; 1) → 1 as n → ∞.

Exercises 365
24. Obtain the sums

(a) F (2a, 2b; a + b + 12 ; 12 )
Γ(a + b + 12 ) Γ( 12 )
= , a + b + 12 = 0, −1, −2, . . . ,
Γ(a + 12 ) Γ(b + 12 )
(c − b)n
(b) F (−n, b; c; 1) = , n = 1, 2, . . . ; c = 0, −1, −2, . . . .
(c)n
25. Show that J1 (x), the Bessel function given by the series (18) with n = 1,
has infinitely many positive zeros μk , arranged so that 0 < μ1 < μ2 < . . . ,
and μk+1 − μk → π as k → ∞. (This property is suggested by the graph in
Figure 3.)
26. Show that every solution of Airy’s equation y + xy = 0 has infinitely
many zeros in the interval (0, ∞).
27. (a) Let q(x) be continuous and q(x) < 0 in an interval (a, b). Use
the Sturm comparison theorem to show that no nontrivial solution of the
differential equation y + qy = 0 can have more than one zero in (a, b).
(b) Let y = y(x) be a nontrivial solution of the differential equation
y + p(x)y + q(x)y = 0 on an interval (a, b) where the coefficient functions
p(x) and q(x) are continuous and q(x) < 0. Show that y(x) has at most one
zero on (a, b).
Suggestion. Apply Rolle’s theorem.
28. Let q(x) be continuous and q(x) > 0 in an interval (a, b). Suppose that
some nontrivial solution of the differential equation y + qy = 0 has zeros
at the points xk , arranged so that a < x1 < x2 < · · · < b. If q(x) is strictly
decreasing, use the Sturm comparison theorem to show that
xk − xk−1 < xk+1 − xk , k = 2, 3, . . . .
If q(x) is increasing, show that xk − xk−1 > xk+1 − xk . This result is due to
Sturm and is sometimes called the Sturm convexity theorem.
Suggestion. Consider y(x − xk + xk−1 ) in the interval (xk , xk+1 ).
29. (a) Let y = y(x) be a nontrivial solution of Bessel’s equation x2 y +
√
xy + (x2 − n2 )y = 0. Show that u(x) = x y(x) satisfies the differential
equation

1 − 4n2
u + 1+ u = 0.
4x2
(b) Conclude that y has infinitely many positive zeros, whose spacing
tends to π. More precisely, if 0 < x1 < x2 < . . . are the consecutive positive
zeros of y, then xk+1 − xk → π as k → ∞.
(c) Show further that xk+1 − xk increases to π when 0 ≤ n < 1

2 and
decreases to π when 12 < n < ∞.
30. Generalizing the device used to study of zeros of the Bessel function J0 ,
show that for some choice of function ϕ(x) > 0 the transformation y = ϕu
converts a differential equation y + py + qy = 0 to an equation of the form
u + Qu = 0. This is known as Liouville’s normal form. Note that the zeros
of y(x) are the same as those of u(x).
31. Show that all real zeros of a hypergeometric polynomial F (−n, b; c; x)

are simple, except possibly for zeros at x = 0 or 1. Assume, as usual, that
the parameter c is not a negative integer or zero.
References
[1] G. E. Andrews, R. Askey, and R. Roy, Special Functions, Cambridge University

[2] Garrett Birkhoff and G.-C. Rota, Ordinary Differential Equations, 4th edition,
Wiley, New York, 1989.
[3] F. Bowman, Introduction to Bessel Functions, Dover Publications, Mineola, NY,

2010.
[4] E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equa-

tions , McGraw–Hill, New York, 1955.
[5] P. Duren and M. Muldoon, “Asymptotic behavior of Bessel functions”, to

appear.
[6] E. L. Ince, Ordinary Differential Equations, Longmans, Green; London, 1926;

reprinted by Dover Publications, Mineola, NY, 1956.
[7] Dunham Jackson, Fourier Series and Orthogonal Polynomials, Mathematical

Association of America, Washington, D.C., Carus Monograph No. 6, 1941; reprinted
by Dover Publications, Mineola, NY, 2004.
[8] D. L. Kreider, R. G. Kuller, and D. R. Ostberg, Elementary Differential Equa-

tions, Addison–Wesley, Reading, Massachusetts, 1968.
[9] A. Laforgia and M. E. Muldoon, “Some consequences of the Sturm comparison

theorem”, Amer. Math. Monthly 93 (1986), 89–94.
[10] G. Loria, “Charles Sturm et son œuvre mathématique”, Enseignement

Mathématique 37 (1938), 249–274.
[11] E. D. Rainville, Special Functions, Macmillan, New York, 1960.

References 367
[12] C. Sturm, “Mémoire sur les équations différentielles linéaires du second ordre”,
J. Math. Pures Appl. 1 (1836), 106–186.
[13] G. N. Watson, A Treatise on the Theory of Bessel Functions, Second edition,
Cambridge University Press, London, 1966.
Chapter 14
Elliptic Integrals
The study of elliptic integrals dates back at least to the early 18th century,
when they arose in connection with problems of mechanics and classical ge-
ometry. They comprise a large class of nonelementary integrals whose calcu-
lation reduces to standard forms that were tabulated at an early stage. The
standard elliptic integrals enjoy some nice transformation properties and
have a close, totally unexpected connection with the arithmetic–geometric
mean. They also satisfy the Legendre relation, an elegant identity that pro-
vides an effective method for numerical calculation of π. These topics will
be the focus of this final chapter.
14.1. Standard forms
In the early days of calculus, Leibniz and others embarked on a program

to calculate the integrals of all “elementary” functions, those functions ex-
pressible as compositions of rational functions, logarithms and exponentials,
trigonometric functions, etc. It soon became apparent, however, that an el-
ementary function need not have an elementary integral. One such example
arises in calculating the arclength of an ellipse. An example of greater histor-
ical importance appears in connection with Bernoulli’s lemniscate, described
in polar coordinates by the equation r2 = cos 2θ. This familiar “figure-eight”
curve, displayed in Figure 1, was known to ancient Greek geometers and was
discussed in detail by Jacob Bernoulli in 1694. Its equation in rectangular
coordinates is
(x2 + y 2 )2 = x2 − y 2 .
369
370 14. Elliptic Integrals
Figure 1. Bernoulli’s lemniscate.
The curve can be described geometrically as the √ locus of points the

product of whose distances from the two points (±1/ 2, 0) has constant
value 12 . The arclength of Bernoulli’s lemniscate is found to be
1
dx
4 √ .
0 1 − x4
The formula is most easily derived from the polar-coordinate form
sin2 2θ 2
ds2 = dr2 + r2 dθ2 = 2
dθ + r2 dθ2
r
1 − r4 1
= 2
+ r dθ2 = 2 dθ2 ,
2
r r
which gives the total arclength
π/4 1
dθ dr
4 √ =4 √ .
0 cos 2θ 0 1 − r4
The arclength integrals for the ellipse and lemniscate are examples of
what are now called elliptic integrals. This term refers to integrals of the
form
R(t, p(t)) dt ,
where R(x, y) is a rational function and p(t) is a polynomial of degree 3

or 4 with no repeated factors. Elliptic integrals occur in many problems of
mathematics and physics. Every elliptic integral can be reduced to a linear
combination of three basic forms. In standard notation, the complete elliptic
integrals of first and second kind are, respectively,
π/2 1
dθ dt
K = K(k) = = √ √
0 1 − k sin θ
2 2 0 1 − t 1 − k 2 t2
2
and π/2 √
1
1 − k 2 t2
E = E(k) = 1 − k 2 sin2 θ dθ = √ dt ,
0 0 1 − t2
14.2. Fagnano’s duplication formula 371
√
for 0 < k < 1. The parameter k is called the modulus, and k = 1 − k 2 is
known as the complementary modulus. Symmetrically related to K and E
are K and E , defined respectively by K (k) = K(k ) and E (k) = E(k ).
It is customary to use the notation K (k) and E (k) despite possible misin-
terpretation as derivatives. The corresponding indefinite integrals are called
incomplete elliptic integrals. The complete elliptic integral of third kind is
π/2
dθ
Π(ν, k) =
0 (1 + ν sin2 θ) 1 − k 2 sin2 θ
1
dt
= √ √ .
0 (1 + νt2 ) 1 − t2 1 − k 2 t2
It is less common and will not concern us here.

A simple calculation identifies the lemniscate integral as a complete el-
liptic integral of first kind. Specifically,
√ √ π/2
dθ √ 1
dx
K(1/ 2) = 2 = 2 √ ,
0 2 − sin2 θ 0 1 − x4
with the substitution x = cos θ. The lemniscate integral can also be ex-
pressed in terms of the gamma function (see Exercise 2). The arclength of
an ellipse is given by a complete elliptic integral of second kind (see Exercise
1).
14.2. Fagnano’s duplication formula
The arclength function

x
dt
L(x) = √ , 0 ≤ x ≤ 1,
0 1 − t4
for the lemniscate is not elementary, but it has some remarkable properties.
The duplication formula
√
2x 1 − x4
2 L(x) = L
1 + x4
was discovered in 1718 by Guilio Fagnano (1682–1766), an Italian magistrate

(later a marquis) and amateur mathematician. The formula as stated is valid
only for x sufficiently small. Analysis of the function
√
2x 1 − x4
g(x) =
1 + x4
0.8
0.6
0.4
0.2
0.2 0.4 0.6 0.8 1
Figure 2. Fagnano’s function y = g(x).
shows that it increases to a maximum value 1 at

√
x = x0 = (3 − 2 2)1/4 = 0.64359 . . . ,
then decreases to 0 at x = 1. Its graph is shown in Figure 2.
Fagnano’s derivation of the duplication formula begins with the substi-

tution t2 = 2u2 /(1 + u4 ), which gives
1 − u4 dt √ du
1 − t4 = and √ = 2√ .
1 + u4 1−t 4 1 + u4
Next the substitution u2 = 2v 2 /(1 − v 4 ) gives
1 + v4 du √ dv
1 + u4 = and √ = 2√ .
1 − v4 1+u 4 1 − v4
A combination of the two transformations produces the relation t = g(v)
and leads to the duplication formula
x g(x)
dv dt
2 √ = √ , 0 ≤ x ≤ x0 .
0 1−v 4
0 1 − t4
Fagnano’s discovery languished in obscurity until 1751, when Euler was

asked to examine a copy of Fagnano’s collected works. Euler quickly saw that
Fagnano’s substitution device could be generalized to obtain the addition
formula √
x 1 − y 4 + y 1 − x4
L(x) + L(y) = L
1 + x2 y 2
for small x and y. He then generalized the addition formula to a large class
of elliptic integrals, including all incomplete elliptic integrals of first kind.
To understand the significance of the addition formula, it is instructive

to consider an analogy with the arcsine integral
x
−1 dt
sin x = √ , 0 ≤ x ≤ 1.
0 1 − t2
If x = sin θ and y = sin ϕ, the addition formula
sin(θ + ϕ) = sin θ cos ϕ + sin ϕ cos θ
for the sine function can be written in the form

sin−1 x + sin−1 y = sin−1 x 1 − y 2 + y 1 − x2
for small x and y, which is similar to Euler’s formula for the lemniscate
integral.
Clearly, it would be awkward to base a study of trigonometry on the
inverse trigonometric functions. It is the sine and cosine that have the
elegant addition formulas, plus another key property: periodicity. Yet after
Euler’s work another half-century elapsed before anyone thought seriously
of inverting the incomplete elliptic integrals and basing the theory on their
inverse functions, now known as elliptic functions. Legendre took some
preliminary steps in this direction. Gauss recorded the idea in his diary but
did not publish it. In 1827, Abel published the first account of the theory
of elliptic functions. He viewed them as functions of a complex variable and
found them to be doubly periodic, with two independent complex periods.
Around the same time, Carl Jacobi (1804–1851) made similar innovations,
which he first published also in 1827. Abel died in 1829 at age 26, but
Jacobi gradually developed the theory that revolutionized the subject. For
many years, Legendre had labored to produce what was to have been a
two-volume treatise [7] on the theory of elliptic integrals, complete with
extensive numerical tables. The first volume was finally published in 1825
and the second in 1826, but Legendre soon realized that much of his work had
been rendered obsolete by the brilliant insights of the young mathematicians
Abel and Jacobi. He might well have reacted with dismay, but to his credit
Legendre embraced the new ideas with great enthusiasm and presented them
in coordination with his earlier work as extensive supplements comprising a
third volume of the treatise.
14.3. The arithmetic–geometric mean

We now digress to discuss a classical algorithm for computing a quantity
called the arithmetic–geometric mean of two positive numbers. This algo-
rithm would appear to have little to do with elliptic integrals, but it turns
out to be very relevant.
Given an arbitrary pair of real numbers a and b with 0 < b < a, calculate
their arithmetic and geometric means
√
a1 = 12 (a + b) and b1 = ab .
Then 0 < b < b1 < a1 < a. Now iterate the process, inductively defining

an+1 = 12 (an + bn ) and bn+1 = an bn , n = 1, 2, . . . .
Then
0 < b < b1 < b2 < · · · < bn < an < · · · < a2 < a1 < a .
Each of the sequences {an } and {bn } is monotonic and bounded, hence
convergent. Let
α = lim an , β = lim bn .
n→∞ n→∞
Observe that
1
α = lim an+1 = lim (an + bn ) = 12 (α + β) ,
n→∞ n→∞ 2
which shows that α = β. The common limit is called the arithmetic–

geometric mean (AGM) of a and b and is denoted by M (a, b). It is easy
to see that the AGM has the properties
√
M (a, b) = M 12 (a + b), ab , M (a, b) = a M (1, b/a) .
The AGM algorithm was discovered by Lagrange around the year 1784.
Gauss rediscovered it in 1791, at the age of 14. The algorithm converges
rapidly, allowing easy numerical calculation of the AGM.
On May 30, 1799, Gauss made a startling discovery. On that day he
recorded a note in his diary to the effect that the two numbers
1
1 2 dx
√ and √
M ( 2, 1) π 0 1 − x4
agree to 11 decimal places. He predicted that “the demonstration of this fact

will surely open up an entirely new field of analysis.” He soon proved that
the two numbers are indeed equal, and he generalized the result to other
elliptic integrals. In fact, Lagrange had already discovered and published
the result in 1785, but Gauss developed other applications of the AGM over
a long period of time, especially in the realm of complex function theory.
It is no great exaggeration to say that Gauss single-handedly fulfilled the
prophecy he made in 1799.
Here is the unexpected connection between the AGM and elliptic inte-
grals of first kind.
√
Theorem 1. For 0 < k < 1 and k = 1 − k 2 ,
1 2
= K(k ) .
M (1, k) π
The theorem provides an explicit formula for the AGM of any two posi-
tive numbers. More importantly, it offers an efficient method for the numer-
ical calculation of elliptic integrals of first kind. Elliptic integrals of second
kind can be computed by a similar process. The extensive tables of elliptic
integrals in Legendre’s treatise [7] were assembled in this manner.
The proof of the theorem will depend on the fact that the elliptic inte-
gral and the AGM have essentially the same transformation property. It is
convenient to express this relation in terms of the function
π/2
dθ 1 b
I(a, b) = = K , 0 < b < a.
0 a2 cos2 θ + b2 sin2 θ a a
Lemma 1. For 0 < b < a,

√
I 12 (a + b), ab = I(a, b) .
Deferring the proof of the lemma, let us observe first that the theorem
is an immediate consequence.
Proof of Theorem 1. Recall the notation {an } and {bn } for the sequences
of arithmetic and geometric means converging to the common limit M (a, b).
In this notation, Lemma 1 says that I(a, b) = I(a1 , b1 ). Iterating this iden-
tity, we infer that I(a, b) = I(an , bn ) for n = 1, 2, . . . . Because the function
I(a, b) is continuous, it follows that
I(a, b) = I(an , bn ) → I(M (a, b), M (a, b))

π/2
1 dθ π 1
= = .
M (a, b) 0 2 2
cos θ + sin θ 2 M (a, b)
Since I(1, k) = K (k) = K(k ) for 0 < k < 1, this is the desired result.
The identity in Lemma 1 is a version of what is called the Landen trans-

formation, so named for the English mathematician John Landen (1719–
1790), who introduced it in 1771. It is usually stated, in less symmetric
form, as follows.
Lemma 2. For 0 < k < 1, the complete elliptic integral of first kind satisfies
√
2 k
(i) K = (1 + k)K(k)
1+k

1−k 1+k
(ii) K = K(k ) .
1+k 2
The three identities in Lemmas 1 and 2 are all equivalent. To see this,
observe first that the conjugate modulus of
√
1−k 2 k
= is = .
1+k 1+k
Inversion gives
√
1− 1+k 1 2

k= , = , k = .
1+ 2 1+ 1+
Therefore, the identity (ii) of Lemma 2 reduces to
√
1 2
K() = K ,
1+ 1+
which is (i) with k replaced by . Thus (i) and (ii) are equivalent.
To see that Lemma 1 is another version of the same identity, let k = b/a
and note that the relation

1 b
I(a, b) = K
a a
implies
√ √
√ 2 2 ab 1 2 2 k
I 12 (a + b), ab = K = K .
a+b a+b a 1+k 1+k
But by (ii) in Lemma 2, the last expression is equal to

1 2 1−k 1 1
K = K(k ) = K (k) = I(a, b) .
a 1+k 1+k a a
This shows that Lemma 2 implies Lemma 1. The reverse implication is
similar.
The validity of Landen’s transformation (Lemma 2) can be verified in
many ways. One of the most elegant proofs was found in 1796 by the English
mathematician and astronomer James Ivory (1765–1842). We give a version
of Ivory’s proof.
Proof of Lemma 2. Recall the definition of the complete elliptic integral

K(k) and expand the integrand in binomial series:
∞

−1/2 −1/2
1 − k sin θ
2 2
=1+ (−1)n k 2n sin2n θ .
n
n=1
Integrate term by term, using the formula

π/2
2n (2n)! π n −1/2 π
sin θ dθ = 2n = (−1) ,
0 2 (n!)2 2 n 2
(cf. Sections 2.5 and 3.5, and Exercise 15 in Chapter 2) to arrive at the
expression
∞
π −1/2 2 2n
K(k) = 1+ k .
2 n
n=1
Now write
√ −1/2
π
2 k 4k
2K = 1− 2
sin θ dθ
1+k 0 (1 + k)2
π −1/2
2k
= 1− (1 − cos 2θ) dθ
0 (1 + k)2
π

−1/2
= (1 + k) 1 + k 2 + 2k cos 2θ dθ .
0

Since 1 + k 2 + 2k cos 2θ = 1 + ke2iθ 1 + ke−2iθ , the last expression takes
the form
π −1/2 −1/2
(1 + k) 1 + ke2iθ 1 + ke−2iθ dθ
0
∞
∞

−1/2 n −1/2 m π 2i(n−m)θ
= (1 + k) k k e dθ ,
n m 0
n=0 m=0

π
with the convention that α0 = 1. But 0 e2i(n−m)θ dθ = 0 if n = m, and so
a calculation leads to the relation
√ ∞
2 k −1/2 2 2n
2K = π(1 + k) 1 + k = 2(1 + k)K(k) ,
1+k n
n=1
which is the identity (i) of Lemma 2.
As a consequence of Lemma 2, one can derive corresponding identities for

elliptic integrals of second kind. These relations are also known as Landen
transformations.
Corollary. For 0 < k < 1, the complete elliptic integral of second kind
satisfies
√
2 k 2
(iii) E = E(k) − (1 − k) K(k)
1+k 1+k

1−k 1 k
(iv) E = E(k ) + K(k ) .
1+k 1+k 1+k
Derivation of corollary. To derive the relation (iii), the strategy is to

differentiate the identity (i) of Lemma 2, using the formula
dK E − k2 K
= ,
dk kk 2
which comes from
√ the power series for K and E (see Exercise 9). Define the
modulus = 2 k/(1 + k) and apply the formulas
1−k d 1−k
= , =√
1+k dk k(1 + k)2
√
to differentiate the composite function K(2 k/(1 + k)). Specifically,
dK d d d dK
= K() = (1 + k)K(k) = K(k) + (1 + k) ,
d dk dk dk dk
or
E() − 2 K() 1−k E(k) − k 2 K(k)
√ = K(k) + (1 + k) ,
2 k(1 + k)2 kk 2
which reduces to
1+k 1−k 1 1
E() − K(k) = E(k) − K(k) ,
2k(1 − k) 2k k(1 − k) k
since K() = (1 + k)K(k) by Lemma 2. Further simplification leads to the
relation (iii). The identity (iv) can be verified either by algebraic manip-
ulations from (iii) or by differentiating (ii). In fact, as with (i) and (ii),
the relations (iii) and (iv) are equivalent forms of the same identity. The
details are left as an exercise.
We are now in position to develop a method for computing complete

elliptic integrals of second kind via the AGM algorithm. The integral
π/2
J(a, b) = a2 cos2 θ + b2 sin2 θ dθ
0
corresponds to the expression I(a, b) introduced earlier. For 0 < b < a and
k = b/a, we find that J(a, b) = a E(k ) The substitution θ = π2 − t shows
that J(b, a) = J(a, b). Note that the arclength of the ellipse
x2 y 2
+ 2 =1
a2 b
is given by 4J(a, b) (see Exercise 1). It will be convenient to use the following
analogue of Lemma 1, which is closely related to Landen’s transformation
for E.
Lemma 3. For 0 < b < a,

√
1
2J 2 (a + b), ab = J(a, b) + ab I(a, b) .
Proof. Let k = b/a and recall the relations I(a, b) = a1 K(k ) and J(a, b) =
a E(k ). Then by Landen’s transformation for E, as given by identity (iv)
in the corollary to Lemma 2, we have
√
√
2 ab
2 J 12 (a + b), ab = 2 12 (a + b) E
a+b
√
2 k
= a(1 + k)E
1+k

1−k

= a(1 + k)E = a E(k ) + kK(k )
1+k
= J(a, b) + ab I(a, b) .

In the notation of the AGM algorithm,
∞ n 2 let c n = a2n − b2n . It can be
shown that the infinite series n=1 2 cn converges (see Exercise 7). Here is
a formula for computing elliptic integrals of second kind.
Theorem 2. For 0 < k < 1, let {an } and {bn } be the sequences generated
by theAGM algorithm, beginning with a = a0 = 1 and b = b0 = k . Let
cn = a2n − b2n , so that c0 = k. Then
∞

E(k) = 1− 2n−1 c2n K(k) .
n=0
Proof. By Lemma 3 and Lemma 1, we have
2 J(an+1 , bn+1 ) − J(an , bn ) = an bn I(an , bn ) = an bn I(1, k ) = an bn K(k)

for n = 0, 1, 2, . . . . This can be rewritten in the form

2n+1 J(an+1 , bn+1 ) − a2n+1 K(k) − 2n J(an , bn ) − a2n K(k) = 2n−1 c2n K(k) ,
since
4a2n+1 − 2a2n + c2n = (an + bn )2 − 2a2n + (a2n − b2n ) = 2an bn .

Now take the sum of these identities as n ranges from 0 to N . Because the
left-hand side telescopes, this gives

2N +1 J(aN +1 , bN +1 ) − a2N +1 K(k) − J(a0 , b0 ) − a20 K(k)
N

= 2n−1 c2n K(k) ,
n=0
or
E(k) = J(1, k )

N

= 1− 2 cn K(k) + 2N +1 J(aN +1 , bN +1 ) − a2N +1 K(k) .
n−1 2
n=0
It remains to show that the last term tends to 0 as N → ∞. Thus it is to

be shown that

Rn (k) = 2n a2n K(k) − J(an , bn ) → 0 as n → ∞ .
But K(k) = I(1, k ) = I(an , bn ) by Lemma 1, and so
π/2
sin2 θ
Rn (k) = 2 cnn 2
dθ .
0 a2n cos2 θ + b2n sin2 θ
n0 2≤ Rn (k) ≤ 2 cn I(an , bn ) = 2 cn K(k) → 0 as n → ∞, since the

Therefore, n 2 n 2
series 2 cn converges, and the theorem is proved.
Much more information about the AGM and its applications can be
found in the book by Borwein and Borwein [2]. The articles by Cox [3, 4]
focus on the remarkable connections that Gauss found between the AGM
and functions of a complex variable. Almqvist and Berndt [1] give an en-
tertaining account of historical developments related to elliptic integrals.
Watson [9] discusses the early history, with emphasis on the contributions
of Fagnano and Landen. The book by Stillwell [8] presents similar material
in broader historical perspective.
Carl Friedrich Gauss (1777–1855) has been called the prince of math-
ematicians. He made groundbreaking advances in many scientific fields,
including number theory, probability and statistics, celestial mechanics, spe-
cial functions, differential geometry, electricity and magnetism, and poten-
tial theory. Born in Braunschweig to parents of humble means and little
education, Gauss was a true prodigy whose conspicuous talent gained him
financial support for higher study in Braunschweig and at the University of
Göttingen. He spent most of his career in Göttingen, where he was direc-
tor of the Sternwarte (astronomical observatory). For a concise account of
Gauss’s life and work, the biography by Hall [6] is recommended.
14.4. The Legendre relation
In his integral calculus text of 1811, Legendre gave some illustrations of

his technique for the reducing elliptic integrals to normal form. His calcula-
tions led him to the observe that for the special value
√
π 3−1
k = sin = √
12 2 2
of the modulus, the combination P = K E +E K −K K of complete elliptic

integrals of first and second kind
√ has the value π/2. Remarking that P also
has the value π/2 for k = 1/ 2 and as k → 0, Legendre proceeded to
compute the derivative. A straightforward calculation, using the formulas
dK E − k2 K dE E−K
= and = ,
dk kk 2 dk k
shows that dP/dk = 0, so P is constant. In other words,
π
K E + EK − K K = , 0 < k < 1.
2
This remarkable identity has become known as the Legendre relation.
Legendre seems to have discovered his relation by accident, and his proof
is not illuminating. It verifies the Legendre relation but does not explain
it. In his treatise [7] of 1825, Legendre reproduced the same calculations
without adding any new insights. Later, when Abel and Jacobi introduced
their theory of elliptic functions, and when Weierstrass developed another
version of the theory, the Legendre relation emerged in a natural way. There
is, however, a more elementary approach which exploits the fact that K and
K are solutions of the same hypergeometric differential equation and so
satisfy a Wronskian relation. It turns out that the Wronskian relation is
none other than the Legendre relation.
Here are the details. Recall that the Gauss hypergeometric function is
defined by
ab 1 a(a + 1)b(b + 1) 2
F (a, b; c; x) = 1 + x+ x +... ,
c 2! c(c + 1)
where a, b, c are real parameters with c = 0, −1, −2, . . . , and |x| < 1. The
function y = F (a, b; c; x) satisfies the hypergeometric equation
x(1 − x)y + [c − (a + b + 1)x]y − ab y = 0 .
Under the special condition a + b + 1 = 2c, it is easily verified that if some
function y(x) is a solution, then so is y(1 − x). (See Chapter 13, Exercise
22.)
The complete elliptic integrals K and E are hypergeometric functions
of x = k 2 . Specifically,
π
π

K(k) = F 12 , 12 ; 1; k 2 , E(k) = F − 12 , 12 ; 1; k 2 .
2 2
These formulas can be derived by expanding the integrands in binomial
series and integrating term by term. As in the proof of Lemma 2, this
process leads to the representation
∞
π −1/2 2 2n
K(k) = 1+ k .
2 n
n=1
On the other hand, the coefficient of xn in the power series for F ( 12 , 12 ; 1; x)

is
1 3 2 2
( 2 )( 2 ) · · · (n − 12 ) (2n)! −1/2 2
= = , n = 1, 2, . . . ,
(n!)2 22n (n!)2 n
π
which shows that K(k) = 2 F ( 12 , 12 ; 1; k 2 ). The formula for E(k) can be
derived in a similar way.
In fact, both formulas are special cases of the Euler integral representa-
tion (cf. Section 13.5)
1
Γ(c)
F (a, b; c; x) = tb−1 (1 − t)c−b−1 (1 − xt)−a dt ,
Γ(b)Γ(c − b) 0
√
valid for c > b > 0. Since Γ( 12 ) = π, the Euler representation reduces to

1 1 1 1 −1/2
2
F 2 , 2 ; 1; k = t (1 − t)−1/2 (1 − k 2 t)−1/2 dt
π 0

2 1 2
= (1 − u2 )−1/2 (1 − k 2 u2 )−1/2 du = K(k)
π 0 π
and

1 1 −1/2
F − 12 , 12 ; 1; k 2 = t (1 − t)−1/2 (1 − k 2 t)1/2 dt
π 0

2 1 2
= (1 − u2 )−1/2 (1 − k 2 u2 )1/2 du = E(k) ,
π 0 π
with the substitution t = u2 .

Recall now that any pair of solutions y1 and y2 of a linear differential
equation y + py + qy = 0 have a Wronskian W = y1 y2 − y1 y2 with the
property x
W (x) = W (x0 ) exp − p(t) dt ,
x0
where x0 is any convenient base-point. The Wronskian relation was derived

in the preceding chapter (Section 13.2). For the hypergeometric equation
with a = b = 12 and c = 1, it takes the form

x
1 − 2t
W (x) = W ( 12 ) exp − dt ,
1/2 t(1 − t)
and an easy calculation of the integral gives
1 1
W (x) = W ( 12 ) , 0 < x < 1.
4 x(1 − x)
Now choose the particular solutions

π
y1 (x) = K(k) = F ( 12 , 12 ; 1; x) ,
2
π
y2 (x) = K (k) = F ( 12 , 12 ; 1; 1 − x) , x = k2 ,
2
noting that K (k) = K(k ) = y1 (1 − x) is also a solution of the hypergeo-
metric equation because the condition a + b + 1 = 2c is satisfied in this case.
The differentiation formula for K(k) yields
dy1 E − (1 − x)K dy2 E − xK

= , =− ,
dx 2x(1 − x) dx 2x(1 − x)
and so the Wronskian W of K and K is given by
−2x(1 − x)W (x) = (E − (1 − x)K)K + K(E − xK )

= K E + EK − K K .
But the Wronskian relation says that x(1 − x)W (x) is constant, so it follows
that Legendre’s expression K E + E K − K K is constant.
To show that the constant is π/2, it is simplest to compute the limit √ of
Legendre’s
√ expression
√ as k → 0. An
√ alternative is to calculate K(1/ 2) =

K (1/ 2) and E(1/ 2) = E (1/ 2) in terms of the gamma function and
then show that
√ √ √ π
2K(1/ 2)E(1/ 2) − K(1/ 2)2 = .
2
The details are left as an exercise.
√ √
Because the elliptic integrals K(1/ 2) and E(1/ 2) can be computed
by the rapidly converging AGM algorithm, the Legendre relation offers an
effective method for numerical calculation of π. For further information, see
the book by Borwein and Borwein [2]. For an exposition of the Legendre
relation as it occurs in the theory of elliptic functions, see also the paper by
Duren [5].
Exercises
1. Show that for 0 < b < a the arclength of the ellipse
x2 y 2
+ 2 =1 is 4a E (b/a) ,
a2 b
where E (k) = E(k ) is a complete elliptic integral of second kind. Use the
parametric representation x = a cos t, y = b sin t.
2. (a) Evaluate the lemniscate integral
1 √
dt 1
√ to show that K(1/ 2) = √ Γ( 14 )2 .
0 1 − t4 4 π
(b) Confirm Gauss’s conjecture (made on the basis of numerical evidence)
that
1 2 1 dx
√ = √ .
M ( 2, 1) π 0 1 − x4
√
3. Use the substitution u = 1 − t2 to show that
√ √ 1
1 + u2
2 E(1/ 2) = √ du .
0 1 − u4
Deduce that
√ 1
E(1/ 2) = √ Γ( 14 )2 + 4 Γ( 34 )2 .
8 π
√ √
4. Show that values of K(1/ 2) and E(1/ 2) found in the preceding exer-
cises are consistent with the Legendre relation.
Hint. Use the Euler reflection formula for the gamma function.
Exercises 385
5. Use a hand calculator to compute the arithmetic–geometric mean M (2, 1)

to 6 decimal places.
6. In the notation of the AGM algorithm, prove that a1 − b1 < 12 (a − b)
and conclude that an − bn < ( 12 )n (a − b). This shows that the algorithm
converges rapidly.

7. In the notation of the AGM algorithm, let cn = a2n − b2n . Show that

c1 = 12 (a − b) ≤ 12 a2 − b2
and in general that cn+1 ≤ 12 cn . Conclude that c2n ≤ ( 14 )n (a2 − b2 ), so that

the series ∞ n 2
n=1 2 cn is convergent.
8. Directly from the definitions of K(k) and E(k), show that

dE E−K
= .
dk k
Justify the process of differentiation under the integral sign.
9. (a) Derive the power series expansion
∞
π 1/2 2 2n
E(k) = 1− (2n − 1) k
2 n
n=1
∞
π 1 (2n)! 2 2n
= 1− k .
2 2n − 1 22n (n!)2
n=1
(b) Use the power series expansions of K(k) and E(k) to verify the formula
dK E − k2 K
= .
dk kk 2
Discussion. In contrast with dE/dk in Exercise 8, the expression for dK/dk
is not readily calculated by differentiation under the integral sign. The power
series calculation is somewhat laborious, but a remarkable combination of
terms gives the result.
10. Verify the Landen transformation for complete elliptic integrals of sec-
ond kind as given by the formula (iv) in the corollary to Lemma 2.
11. As early as 1738, Euler discovered the formula
1 1
dx x2 dx π
√ √ = .
0 1 − x4 0 1 − x4 4
(a) Show
√ that this is a special case of the Legendre relation, with k =
k = 1/ 2.
(b) Without appeal to elliptic integrals, verify Euler’s formula by calcu-

lating each integral in terms of the gamma function and applying Euler’s
reflection formula. (See Chapter 9, Exercise 10.)
√
12. Use
√ the AGM methods of Theorems 1 and 2 to calculate K(1/ 2) and
E(1/ 2) to at least 6 decimal places. Then use the Legendre relation to
calculate π.
√
13. For 0 < k < 1 and k = 1 − k 2 , show that
1/k 1
ds dt
√ √ = √ = K (k) .
1 s2 − 1 1 − k 2 s2 0 2
1−t 1−k t 2 2
Suggestion. Let s = (1 − k 2 t2 )−1/2 .
References
[1] G. Almqvist and B. Berndt, “Gauss, Landen, Ramanujan, the arithmetic–

geometric mean, ellipses, π, and the Ladies Diary ”, Amer. Math. Monthly 95
(1988), 585–608.
[2] J. M. Borwein and P. B. Borwein, Pi and the AGM: A study in Analytic Number
Theory and Computational Complexity, John Wiley & Sons, New York, 1987.
[3] D. A. Cox, “The arithmetic–geometric mean of Gauss”, Enseignement Math. 30
(1984), 275–330.
[4] D. A. Cox, “Gauss and the arithmetic–geometric mean”, Notices Amer. Math.
Soc. 32 (1985), 147–151.
[5] P. Duren, “The Legendre relation for elliptic integrals”, in Paul Halmos: Cel-
ebrating 50 Years of Mathematics (J. H. Ewing and F. W. Gehring, editors),
Springer-Verlag, New York, 1991; pp. 305–315.
[6] Tord Hall, Carl Friedrich Gauss: A Biography, English translation, MIT Press,
Cambridge, Massachusetts, 1970.
[7] A. M. Legendre, Traité des Fonctions Elliptiques, Paris, 1825 (Tome I), 1826
(Tome II), 1828 (Tome III).
[8] J. Stillwell, Mathematics and Its History, Springer-Verlag, New York, 1989.
[9] G. N. Watson, “The marquis and the land agent”, Math. Gazette 17 (1933),
5–17.
Index of Names
Abel, Niels Henrik, 76, 373 Fourier, Joseph, 198

Frobenius, Georg, 180, 343
Bernoulli, Daniel, 249
Bernoulli, Jacob, 92, 291, 294, 369 Gauss, Carl Friedrich, 250, 253, 255,
Bernoulli, Johann, 92, 249 280, 350, 374, 381
Bernstein, Felix, 311 Gerver, Joseph, 101
Bernstein, Sergei, 151, 157 Gibbs, J. W., 212
Bessel, Friedrich Wilhelm, 348 Goldbach, Christian, 249
Bolzano, Bernard, 40, 99
Borel, Émile, 41, 164, 181 Hadamard, Jacques, 280
Bunyakovski, Victor Yacovlevich, 119 Hardy, G. H., 101, 183, 184, 191, 283
Heine, Eduard, 41
Cantor, Georg, 4, 309 Hilbert, David, 122, 321
Carleson, Lennart, 220 Hölder, Otto, 113
Cauchy, Augustin-Louis, 40, 112, 117,
Ivory, James, 376
328
Chebyshev, P. L., 148, 280 Jacobi, Carl, 373
Coriolis, Gaspard-Gustave, 137 Jensen, J. L. W. V., 120
Darboux, Gaston, 41 Karamata, Jovan, 184, 190
de la Vallée Poussin, Charles, 164, 280 Kronecker, Leopold, 2, 270, 310
de Moivre, Abraham, 63
Dedekind, Richard, 309 Lagrange, Joseph-Louis, 40, 84, 374
Dirichlet, Peter Gustav Lejeune, 41, Landau, Edmund, 153
207, 253 Landen, John, 375
du Bois-Reymond, Paul, 219 Lebesgue, Henri, 152, 315, 321
Legendre, Adrien-Marie, 249, 253, 280,
Euler, Leonhard, 39, 56, 92, 98, 212, 373, 381
249, 279, 295, 300, 350, 372 Leibniz, Gottfried Wilhelm, 20, 39
Liouville, Joseph, 358
Faber, Georg, 151 Littlewood, J. E., 183, 184
Fagnano, Guilio, 371
Fejér, Leopold, 215, 219, 275 Maclaurin, Colin, 84, 300
Fekete, Mihály, 170 Mertens, Franz Josef, 83
387
388 Index of Names
Minkowski, Hermann, 116

Moore, E. H., 321
Müntz, Chaim, 171
Newton, Isaac, 20, 39, 87
Pál, Julius, 168

Peano, Giuseppe, 321, 328
Picard, Émile, 328
Poisson, Siméon-Denis, 236
Remez, Evgeny, 164

Riemann, Bernhard, 41, 98, 286
Runge, Carl, 151
Scheeffer, Ludwig, 317

Schröder, Ernst, 311
Schur, Issai, 123
Schwarz, Hermann Amandus, 119
Stirling, James, 63
Stone, Marshall, 165
Sturm, Charles-François, 354, 358
Szász, Otto, 171
Tauber, Alfred, 182

Taylor, Brooke, 84
Viète, François, 60
Wallis, John, 61
Weierstrass, Karl, 40, 98, 151, 156, 381
Weyl, Hermann, 123, 271
Wilbraham, Henry, 212
Wronski, Josef, 333
Subject Index
Abel summability, 180 bounded set, 5

Abel summation, 77
Abel’s theorem, 76 Cantor set, 315, 317, 321, 323, 324
integral form, 80, 261 Cantor–Scheeffer function, 317, 324
Abelian theorem, 179 cardinal number, 310
absolute convergence of infinite cardinal series, 235
product, 135 Catalan numbers, 104
absolutely convergent series, 19 Cauchy criterion, 8
algebraic number, 3 Cauchy product, 81
alternating series theorem, 20 Cauchy sequence, 8
arithmetic mean, 117 Cauchy’s double series theorem, 22, 91
arithmetic–geometric mean, 374 Cauchy’s inequality, 112
arithmetic–geometric mean inequality, Cauchy–Schwarz inequality, 119
117 Cesàro means, 180, 215, 228
Cesàro summability, 180
band-limited signals, 234 Cesàro summation, 228
Basel problem, 92, 295 Chebyshev polynomials, 148, 150, 172,
Bernoulli numbers, 93, 291 173
Bernoulli periodic functions, 302 circular drumhead, 343
Bernoulli polynomials, 297 closed set, 9
Bernoulli’s lemniscate, 369 cluster point, 7
Bernstein polynomials, 157, 174, 175 compact set, 12
Bessel functions, 338, 343, 346, 356, 362 comparison test, 19
Bessel’s equation, 337, 346 complement, 9
Bessel’s inequality, 202 complementary modulus, 371
beta function, 251 complete elliptic integrals, 370
binary expansion, 316 complete metric space, 42
binomial coefficient, 12 completeness, 5
binomial theorem, 12 complex conjugate, 43
Bohr–Mollerup theorem, 258 complex numbers, 42
Bolzano–Weierstrass theorem, 7 composite number, 277
Borel summability, 181 composition of power series, 89
bounded sequence, 6 conditionally convergent series, 21
389
390 Subject Index
conjugate indices, 114 Fejér means, 215

continuous function, 22 Fejér’s theorem on equidistribution,
continuous nowhere differentiable 274, 275
functions, 98 Fejér’s theorem on Fourier series, 215
Bolzano’s example, 99 Fekete’s theorem, 170
Riemann’s example, 98, 101 Fibonacci sequence, 46
van der Waerden’s example, 99, 106 Fourier coefficients, 200
Weierstrass’s example, 98, 99 Fourier series, 197, 198, 200, 299
convergence of infinite product, 131 convergence theorem, 205
convergent sequence, 6 examples, 207
convex function, 119 partial sums, 204
convolution, 227, 265 Fourier transform, 222
countable basis theorem, 10 examples, 222
countable set, 3, 310 inversion, 228, 230
covering, 11 properties, 224
critical point, 26 Frobenius’ theorem, 180
cross product, 113 fundamental theorem of arithmetic, 278
fundamental theorem of calculus, 31
de Moivre formula, 45, 96
deleted neighborhood, 7 gamma function, 247, 280
derivative, 25 duplication formula, 252
devil’s staircase, 320 multiplication formula, 259
differentiable function, 25 product representation, 256
differential equations reflection formula, 251
Bessel’s equation, 337, 346 Gauss multiplication formula, 259
hypergeometric equation, 348 Gauss product formula, 255
Legendre’s equation, 339, 361 Gauss summation formula, 352
power series solutions, 336 geometric mean, 117
differentiation term by term, 35 geometric series, 18
differentiation under integral sign, 35 Gibbs phenomenon, 212
digamma function, 266
Dini’s theorem, 38 Hardy’s power series, 190
Dirichlet formula, 204 Hardy’s Tauberian theorem, 184, 242,
Dirichlet kernel, 203 276
divergent sequence, 6 Hardy–Littlewood Tauberian theorem,
dot product, 112 184, 192
harmonic function, 197
e, 51 harmonic mean, 127
elliptic functions, 373 heat conduction, 197
elliptic integrals, 369, 370 Heine–Borel theorem, 11
empty set, 5 high-indices theorem, 191
equidistribution, 269 Hilbert’s inequality, 122
Euler integral representation, 350 Hölder’s inequality, 113
Euler’s constant, 56, 68, 256, 264, 302 hypergeometric equation, 348, 349
Euler’s product formula, 278 hypergeometric functions, 349
Euler’s reflection formula, 253 examples, 350, 364
Euler’s sums, 92, 208, 212, 295 integral representation, 350
Euler–Maclaurin summation formula, summation formula, 352
93, 300, 302 transformations, 351, 353
Fagnano’s duplication formula, 371 indicial equation, 340

Fejér kernel, 216 induction, 1
Subject Index 391
infimum, 5 measure zero, 313

infinite products, 131 Mertens’ theorem, 82, 89
infinite series, 18 method of Frobenius, 338, 341
intermediate value theorem, 23 metric space, 41
for integrals, 30 minimum, 23
interpolation, 145 Minkowski’s inequality, 116
intervals, 5 modulus of complex number, 43
inversion of Fourier transforms, 228, 230 of elliptic integral, 371
irrational number, 3 monotone boundedness theorem, 6
irrationality of e, 54 monotonic sequence, 6
irrationality of π, 55 Müntz–Szász theorem, 171
isoperimetric inequality, 242
n-dimensional sphere, 266
Jacobi’s inversion formula, 234, 285 nested sets theorem, 9
Jacobi’s theta function, 233, 285 Newton’s binomial series, 87
Jensen’s inequality, 120
open set, 9
Karamata’s method, 185 orthogonal functions, 199
Kronecker’s theorem, 270
Pál’s theorem, 168
Lagrange interpolation formula, 146 Parseval’s relation, 202, 231
Lagrange’s identity, 113 partial sum, 18
Lambert summability, 182 Peano curves, 321
Landen transformation, 375, 377 Pfaff’s transformation, 351
Laplace equation, 197 pointwise convergence, 33
Laplace transform, 80, 223, 261, 265 Poisson formula, 241
least upper bound principle, 5 Poisson kernel, 241
Lebesgue measure, 314 Poisson summation formula, 193, 232
left-hand limit, 25 polynomial of best approximation, 160
Legendre polynomials, 361 power series, 73
Legendre relation, 381 power series solutions, 336
Legendre’s duplication formula, 252 power set, 312
Legendre’s equation, 339, 361 prime number, 277
Leibniz’ theorem, 20 prime number theorem, 280
l’Hôpital’s rule, 49
lim inf, 15 radius of convergence, 74
lim sup, 15 ratio test, 19
limit of a function, 25 rational number, 2
limit of a sequence, 6 real numbers, 3
Liouville’s normal form, 366 reduction formula, 62
Lipschitz condition, 205, 328 regular point, 336
Littlewood’s Tauberian theorem, 184, regular singular point, 338
241 remainder, 85
local maximum, 26 Cauchy form, 85
local minimum, 26 Lagrange form, 85
lower bound, 5 Riemann hypothesis, 282
lower limit, 15 Riemann integrable function, 28
lowest terms, 2 Riemann integral, 27
Riemann sum, 30
Markov’s inequality, 173 Riemann zeta function, 98, 276, 297
maximum, 23 functional equation, 282
mean value theorem, 27 product formula, 278
mean-square approximation, 200 zeros, 282
392 Subject Index
Riemann’s localization principle, 205 Weierstrass approximation theorem,

right-hand limit, 25 145, 151, 185, 202, 273
Rolle’s theorem, 26 Weierstrass M-test, 37
root test, 19 Weyl’s theorem, 272
Wirtinger’s inequality, 242
sampling theorem, 235 Wronskian, 333
Schröder–Bernstein theorem, 311 Wronskian relation, 335, 381
Schwarz inequality, 119
sequence of real numbers, 6 Young’s inequality, 114
shifted factorial, 349
sine function
product formula, 211
singular point, 337
Sonin–Pólya oscillation theorem, 359
space-filling curves, 320, 321
squaring the circle, 56
Stirling’s formula, 63
generalization, 257
Stirling’s series, 304, 305
Stone–Weierstrass theorem, 165
Sturm comparison theorem, 355, 358
Sturm convexity theorem, 365
Sturm separation theorem, 355
subsequence, 8
summability, 179
summation by parts, 77
supremum, 5
Tauber’s theorem, 182

Tauberian condition, 184
Tauberian theorem, 179, 184
Taylor series, 85
Taylor’s formula, 84, 85
ternary expansion, 316
transcendental number, 3
transfinite numbers, 310
uniform Cauchy sequence, 36

uniform convergence, 33
uniform convergence of infinite product,
138
uniform distribution, 269
uniformly continuous function, 24
universal power series, 170
upper bound, 5
upper limit , 15
Vandermonde determinant, 146, 171

Vieta’s product formula, 60
Wallis product formula, 61, 212

wave equation, 343
Titles in This Series
Volume
17 Peter Duren
Invitation to Classical Analysis
2012
16 Joseph L. Taylor
Complex Variables
2011
15 Mark A. Pinsky
Partial Differential Equations and Boundary-Value Problems with Applications,
Third Edition
1998
14 Michael E. Taylor
Introduction to Differential Equations
2011
13 Randall Pruim
Foundations and Applications of Statistics: An Introduction Using R
2011
12 John P. D’Angelo
An Introduction to Complex Analysis and Geometry
2010
11 Mark R. Sepanski
Algebra
2010
10 Sue E. Goodman
Beginning Topology
2005
9 Ronald Solomon
Abstract Algebra
2003
8 I. Martin Isaacs
Geometry for College Students
2001
7 Victor Goodman and Joseph Stampfli
The Mathematics of Finance: Modeling and Hedging
2001
6 Michael A. Bean
Probability: The Science of Uncertainty with Applications to Investments,
Insurance, and Engineering
2001
TITLES IN THIS SERIES
5 Patrick M. Fitzpatrick
Advanced Calculus, Second Edition
2006
4 Gerald B. Folland
Fourier Analysis and Its Applications
1992
3 Bettina Richmond and Thomas Richmond
A Discrete Transition to Advanced Mathematics
2004
2 David Kincaid and Ward Cheney
Numerical Analysis: Mathematics of Scientific Computing, Third Edition
2002
1 Edward D. Gaughan
Introduction to Analysis, Fifth Edition
1998
This book gives a rigorous treatment of selected topics in
classical analysis, with many applications and examples.
UM Photo Services, Martin Vloet.

The exposition is at the undergraduate level, building on
basic principles of advanced calculus without appeal to more
sophisticated techniques of complex analysis and Lebesgue
integration.
Among the topics covered are Fourier series and integrals,
approximation theory, Stirling’s formula, the gamma function,
Bernoulli numbers and polynomials, the Riemann zeta function, Tauberian theo-
rems, elliptic integrals, ramifications of the Cantor set, and a theoretical discussion
of differential equations including power series solutions at regular singular
points, Bessel functions, hypergeometric functions, and Sturm comparison
theory. Preliminary chapters offer rapid reviews of basic principles and further
background material such as infinite products and commonly applied inequalities.
This book is designed for individual study but can also serve as a text for second-
semester courses in advanced calculus. Each chapter concludes with an abundance
of exercises. Historical notes discuss the evolution of mathematical ideas and their
relevance to physical applications. Special features are capsule scientific biogra-
phies of the major players and a gallery of portraits.
Although this book is designed for undergraduate students, others may find it an
accessible source of information on classical topics that underlie modern develop-
ments in pure and applied mathematics.
For additional information

and updates on this book, visit
www.ams.org/bookpages/amstext-17
AMS on the Web

AMSTEXT/17 www.ams.org
Sally
The This series was founded by the highly respected
mathematician and educator, Paul J. Sally, Jr.

Invitation To Classical Analysis Compress

Uploaded by

Copyright:

Available Formats

Invitation To Classical Analysis Compress

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Invitation To Classical Analysis Compress

Uploaded by

Copyright:

Available Formats

Pure and Applied

American Mathematical Society

American Mathematical Society

For additional information and updates on this book, visit

Library of Congress Cataloging-in-Publication Data

Carl Friedrich Gauss

Peter Gustav Lejeune Dirichlet

Chapter 1. Basic Principles 1

Chapter 2. Special Sequences 51

Chapter 3. Power Series and Related Topics 73

Chapter 4. Inequalities 109

Chapter 5. Inﬁnite Products 131

Chapter 6. Approximation by Polynomials 145

Chapter 7. Tauberian Theorems 179

Chapter 8. Fourier Series 197

Chapter 9. The Gamma Function 247

Chapter 10. Two Topics in Number Theory 269

Chapter 11. Bernoulli Numbers 291

Chapter 12. The Cantor Set 309

Chapter 13. Diﬀerential Equations 327

Chapter 14. Elliptic Integrals 369

Index of Names 387

Subject Index 389

This is a book for undergraduates. To be more precise, it is designed

been made to include the background necessary for full comprehension. It is

Classical analysis is rooted in diﬀerential and integral calculus. Calculus

1.1. Mathematical induction

the inductive hypothesis that Pn is true for some n ∈ N, it is deduced that

1.2. Real numbers

for some polynomial P whose coeﬃcients a0 , a1 , . . . , an are integers, not all

over that number without including it in the sequence.

Thus the resulting sequence is

x1 = 0.d11 d12 d13 . . .

Least Upper Bound Principle. If a set of real numbers is bounded above,

Equivalently, every set of real numbers that is bounded below has a

1.3. Completeness principles

Monotone Boundedness Theorem. Every monotonic and bounded se-

Proof. It suﬃces to suppose that the sequence {xn } is nondecreasing. Then

With the monotone boundedness theorem in hand, we can now prove

Bolzano–Weierstrass Theorem. Every bounded inﬁnite set of real num-

Proof. Let S ⊂ R be a bounded inﬁnite set. Then S is contained in some

and b1 = b. In either case the subinterval [a1 , b1 ] contains inﬁnitely many

contains inﬁnitely many points of S. Proceeding inductively, construct a

a number N such that ξ − ε < an ≤ ξ and ξ ≤ bn < ξ + ε for all n ≥ N . But

contains a point of S, in fact inﬁnitely many. This says that ξ is a cluster

It is now a short step to the sequential form of the theorem. A subse-

Bolzano–Weierstrass Theorem (sequential form). Every bounded se-

Proof. Let {xn } be a bounded sequence and let S = {xn : n ∈ N} be its

We are now ready to establish the Cauchy criterion for convergence of

Theorem (Cauchy criterion). Every Cauchy sequence of real numbers

|xn | ≤ |xN | + |xn − xN | ≤ |xN | + 1

Next we apply the Bolzano–Weierstrass theorem to prove the Heine–

that lies entirely in G. A set F ⊂ R is said to be closed if its complement

hypotheses “closed” or “bounded” is relaxed the conclusion is false. For

Proof of theorem. The theorem is trivial if some set Fm is ﬁnite, because

Countable Basis Theorem. There exists a countable collection of bounded

Proof. Consider the family F of all open neighborhoods

with rational centers p and rational radii r. In other words,

The family F is in one-to-one correspondence with the Cartesian product