This action might not be possible to undo. Are you sure you want to continue?
FOR ENGINEERS I
(DIFFERENTIAL CALCULUS)
SEVER ANGEL POPESCU
Contents
Preface v
Chapter 1. The real line. 1
1. The real line. Sequences of real numbers 1
2. Sequences of complex numbers 26
3. Problems 28
Chapter 2. Series of numbers 31
1. Series of nonnegative real numbers 31
2. Series with arbitrary terms 45
3. Approximate computations 50
4. Problems 52
Chapter 3. Sequences and series of functions 55
1. Continuous and di¤erentiable functions 55
2. Sequences and series of functions 65
3. Problems 76
Chapter 4. Taylor series 79
1. Taylor formula 79
2. Taylor series 89
3. Problems 92
Chapter 5. Power series 95
1. Power series on the real line 95
2. Complex power series and Euler formulas 102
3. Problems 107
Chapter 6. The normed space R
n
. 109
1. Distance properties in R
n
109
2. Continuous functions of several variables 120
3. Continuous functions on compact sets 126
4. Continuous functions on connected sets 133
5. The Riemann’s sphere 135
6. Problems 137
7. Partial derivatives. Di¤erentiability. 140
iii
iv CONTENTS
8. Chain rules 152
9. Problems 162
Chapter 7. Taylor’s formula for several variables. 165
1. Higher partial derivatives. Di¤erentials of order /. 165
2. Chain rules in two variables 175
3. Taylor’s formula for several variables 178
4. Problems 183
Chapter 8. Contractions and …xed points 185
1. Banach’s …xed point theorem 185
2. Problems 189
Chapter 9. Local extremum points 191
1. Local extremum points for many variables 191
2. Problems 197
Chapter 10. Implicitly de…ned functions 199
1. Local Inversion Theorem 199
2. Implicit functions 202
3. Functional dependence 208
4. Conditional extremum points 211
5. Change of variables 215
6. The Laplacian in polar coordinates 218
7. A proof for the Local Inversion Theorem 219
8. The derivative of a function of a complex variable 223
9. Problems 228
Bibliography 231
Preface
I start this preface with some ideas of my former Teacher and
Master, senior researcher I, corresponding member of the Romanian
Academy, Dr. Doc. Nicolae Popescu (Institute of Mathematics of the
Romanian Academy).
Question: What is Mathematics?
Answer: It is the art of reasoning, thinking or making judgements.
It is di¢cult to say more, because we are not able to exactly de…ne the
notion of a "table", not to say Math! In the greek language "mathema"
means "knowledge". Do you think that there is somebody who is able
to de…ne this last notion? And so on... Let us do Math, let us apply
or teach it and let us stop to search for a de…nition of it!
Q: Is Math like Music?
A: Since any human activity involves more or less need of reasoning,
Mathematics is more connected with our everyday life then all the other
arts. Moreover, any description of the natural or social phenomena use
mathematical tools.
Q: What kind of Mathematics is useful for an engineer?
A: Firstly, the basic Analysis, because this one is the best tool
for strengthening the ability of making correct judgements and of tak
ing appropriate decisions. Formulas and notions of Analysis are at
the basis of the particular language used by the engineering topics
like Mechanics, Material Sciences, Elasticity, Concrete Sciences, etc.
Secondly, Linear Algebra and Geometry develop the ability to work
with vectors, with geometrical object, to understand some speci…c alge
braic structures and to use them for applying some numerical methods.
Di¤erential Equations, Calculus of Variations and Probability Theory
have a direct impact in the scienti…c presentation of all the engineering
applications. Computer Science cannot be taught without the basic
knowledge of the above mathematical topics. Mathematics comes from
reality and returns to it.
Q: How can we learn Math such that this one not becomes abstract,
annoying, di¢cult, etc.?
v
vi PREFACE
A: There is only one way. Try to clarify and understand everything,
step by step, from the simplest notions up to the more complicated
ones. Without gaps! Try to work with all the new notions, de…nitions,
theorems, by looking at appropriate simple examples and by doing
appropriate exercises. Do not learn by heart! This is the most useless
thing you can do in trying to become a scientist, an engineer or an
economist! Or anything else!
Math becomes nice and easy to you if it is presented in a LIVELY
way and if you make some e¤orts to come closer and closer to it. If
you hate it from the beginning, don’t say that it is di¢cult!
The present course of Mathematical Analysis covers the Di¤erential
Calculus part only.
It is assumed that students have the basic skills to compute simple
limits, di¤erentials and the integrals of some elementary functions. My
teaching experience of almost 30 years at the Technical University of
Civil Engineering Bucharest made me clear that the Math syllabus
for engineering courses is not only a "part" from the syllabus of the
faculties of mathematics. The engineering teaching should have at its
basis very "concrete" facts. The mathematics for engineers should be
very LIVE. The student should realize that such type of Math came
from "practice", returns to it and, what is most important, it helps a
lot to make rational "models" for some speci…c phenomena. Besides
this point of view, we have not to forget that the most important tool
of an engineer, economist, etc. is his (her) power of reasoning. And
this power of reasoning can be strengthened by mathematical training.
My opinion is that some motivations and drawings are always very
useful in the complicated process of making "easy" and "nice" the
mathematical teaching.
I consider that it is better to start with the notion of a real num
ber, which re‡ects a measurement. Then to consider sequences, series,
functions, etc.
In a chapter I tried to put together some notions and ideas which
have more features in common. We end every chapter with some prob
lems and exercises. In some places you will …nd more detailed examples
and worked problems, in others you will …nd fewer. At any moment I
have in my mind a beginner student and not a moment a professional
in Math. My last goal in this was "the art of teaching Math for engi
neers" and not "the art of solving sophisticated Math problems". We
should be very careful that a good Math teaching means "not multa,
sed multum" (C. F. Gauss, in Latin). Gauss wanted to say that the
PREFACE vii
quality is more important then the quantity, "not much and super…
cial, but fewer and deep". We have computers which are able to supply
us with formulas, with complicated and long computations but, up to
now, they are not able to learn us the deep and the original creative
work. They are useful for us, but the last decision is better to be ours.
The deep "feeling" of an experienced engineer is as important as some
long computations of a computer. If we consider a computer to be
only a "tool" is OK. But, how to obtain this "feeling"? The answer
is: a GOOD background (including Math training) + practice + the
capacity of doing things better and better.
I tried to use as proofs for theorems, propositions, lemmas, etc. the
most direct, simple and NATURAL proofs that I know, such that the
student be able to REALLY understand what the statement wants to
say. The mathematical "tricks" and the simpli…cations by using more
abstract mathematical machinery are not so appropriate in teaching
Math at least for the non mathematical community. This is why we
(teachers) should think twice before accepting a new "shorter" way.
My opinion is that student should begin with a particular case, with
an example, in order to understand a more general situation. Even in
the case of a de…nition you should search for examples and "counterex
amples", you should work with them to become "a friend" of them...
.
I am grateful to many people who helped me directly or indirectly.
The long discussions with some of my colleagues from the Department
of Mathematics and Computer Sciences of the Technical University of
Civil Engineering Bucharest enlightened me a lot. In particular, the
teaching skill, the knowledge and the enthusiasm of Prof. Dr. Gavriil
P¼ altineanu impressed and encouraged me in writing this course. He
was always trying to really improve the way of Math Analysis teaching
in our university.
Many thanks go to Prof. Dr. Octav Olteanu (University Politehnica
Bucharest) for many useful remarks on a previous version of this course.
To be clear and to try to prove "everything" I learned from Prof.
Dr. Mihai Voicu, who was previously teaching this course for many
years.
The friendly climate created around us by our departmental chiefs
(Prof. Dr. ing. Nicoleta R¼ adulescu, Prof. Dr. Gavriil P¼ altineanu,
Prof. Dr. Romic¼ a Tranda…r, etc.) had a great contribution to the
natural development of this project.
I thank to my assistant professor Marilena Jianu for many correc
tions made during the reading of this material.
viii PREFACE
A special thought goes to the late Dr. Ion Petric¼ a who (many years
ago) had the "feeling" that I could write a "popular" book of Math
Analysis with the title "Analysis is easy, doesn’t it?".
The last, but not the least, I express my gratitude to my wife for
helping me with drawings and for a lot of patience she had during my
writing of this book.
I will be very grateful to all the readers who will send me their re
marks on this course to the email address: angel.popescu@gmail.com,
in order to improve everything in future editions.
Prof. Dr. Sever Angel Popescu
Bucharest, January, 2009.
CHAPTER 1
The real line.
1. The real line. Sequences of real numbers
To measure is a basic human activity. To measure time, tempera
ture, velocity, etc., reduces to measure lengths of segments on a line.
For this, we need a …xed point ( on a straight line (d) and a "wit
ness" oriented segment [(¹
1
] (¹
1
= (), i.e. a unitary vector
÷÷
(¹
1
(see
Fig.1.1). Here, unitary means that always in our considerations the
length of the segment [(¹
1
] will be considered to have 1 meter. The
pair ((.
÷÷
i ). where
÷÷
i =
÷÷
(¹
1
is called a Cartesian (from the French
mathematician R. Descartes, the father of the Analytical Geometry,
what shortly means to study …gures by means of numbers) coordinate
system (or a frame of reference). We assume that the reader has a
practical knowledge of the digits 0. 1. 2. 3. 4. 5. 6. 7. 8. 9 which represent
(in Fig.1.1) the points (. ¹
1
. ¹
2
. .... ¹
9
. Let us now consider the point
1 on the line (d) such that the length
÷÷÷
¹
9
1
of the vector
÷÷÷
¹
9
1 is 1
meter and 1 = ¹
8
. i.e.
÷÷÷
¹
9
1 =
÷÷
(¹
1
as FREE vectors.
right orientation
Fig. 1.1
O A
1
A
2
A
3
A
4
A
n
A
n+1
(d) +
_
A
n
A
4
A
3
A
2
A
1
A
[11]
A
[12]
OA
3
= 3 OA
1
inverse
orientation
Our intention is to associate a sequence of digits to the point 1.
Here appears a …rst GREAT IDEA of an anonymous inventor who
denoted 1 by ¹
10
. this means one group of ten units (a unit is one
÷÷
(¹
1
) and 0 (nothing) from the next similar group. For instance, ¹
64
is
the point on (d) which is between the points ¹
60
and ¹
70
such that it
marks 6 groups of ten units + 4 units from the 7th group. Now ¹
269
marks 2 groups of hundreds + 6 groups of tens + 9 units, ... and so on.
In this way we can represent on the real line (d) any quantity which is
a multiple of a unity (for instance 130 km/h if the unity is 1 km/h).
The idea of grouping in units, tens, hundreds, thousands, etc. supply
1
2 1. THE REAL LINE.
us with an addition law for the set of the so called "natural numbers":
0. 1. 2. ...9. 10. 11. ...99. 100. 101. .... We denote this last set by N.
For instance, let us explain what happens in the following addition:
(1.1)
3 6 8 +
9 7
4 6 5
First of all let us see what do we mean by 368. Here one has 3 groups
of one hundred each + 6 groups of one ten each + 8 units (i.e. 8 times
÷÷
(¹
1
). We explain now the result 465 (= 368 + 97) : 8 units + 7 units
is equal to 15 units. This means 5 units and 1 group of ten units. This
last 1 must be added to 6 + 9 and we get 16 groups of ten units each.
Since 10 groups of 10 units means a group of 1 hundred, we must write
6 for tens and add to 3 this last 1. So one gets 4 for hundreds. We say
that a point ¹ on the line (d) is "less" than the point 1 on the same
line if the point 1 is on the right of ¹ and not equal to it. Assume now
that ¹ is represented by the sequence of digits c
a
c
a÷1
...c
0
(c
0
units, c
1
tens, etc.) and 1 by the sequence /
n
/
n÷1
.../
0
. Here we suppose that c
a
and /
n
are distinct of 0 and that : _ :. Otherwise, we change ¹ and
1 between them. Think now at the way we de…ned these sequences!
If : _ :. ¹ must be on the right of 1 or identical to it. If : :
then ¹ is greater than 1. If : = :. but c
a
/
a
. again ¹ is greater
than 1. If : = :. c
a
= /
a
. but c
a÷1
< /
a÷1
. then 1 is greater than
¹. If : = :. c
a
= /
a
. c
a÷1
= /
a÷1
. we compare c
a÷2
with /
a÷2
and
so on. If all the corresponding terms of the above sequences are equal
one to each other (and : = :) we have that ¹ is identical with 1. If
for instance, : = :. c
a
= /
a
. c
a÷1
= /
a÷1
. ..... c
I
= /
I
. but c
I÷1
/
I÷1
we must have ¹ 1 (¹ is greater than 1). Here in fact we described
what is called the "lexicographic order" in the set of …nite sequences
(de…ne it!). If ¹ _ 1 one can subtract 1 from ¹ as it follows in this
example:
(1.2)
3 6 8 ÷
9 7
2 7 1
This operation is as natural as the addition. Namely, 8 units minus
7 units is 1 unit. Since we cannot subtract 9 tens from 6 tens, we
"borrow" 1 hundred = 10 tents from 3. So, now 10 tens + 6 tens
= 16 tens minus 9 tens is equal to 7 tens. It remains 2 hundreds from
which we subtract 0 hundreds and obtain 2 hundreds. Instead of 10
tens we write 10 10 = 10
2
units, etc. Thus, any natural number
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 3
¹ = c
a
c
a÷1
...c
0
(we identi…ed here the name of the point with its
corresponding sequence of digits) can be uniquely written as:
(1.3) ¹ = c
0
+ 10c
1
+ 10
2
c
2
+... + 10
a
c
a
This is also called the representation of ¹ in the base (of numeration)
10. If instead of grouping units, tens, hundreds, etc., in groups of 10.
we group them in groups of 2 for instance, we obtain the writing of
same point ¹ in base 2. etc. Why our ancestors chose 10. ... we do not
know! Maybe because we have 10 …ngers...!!
Hence, the subtraction is not de…ned for any pair ¹. 1. This means
that ¹÷1 does not belong to N for any pair ¹. 1. For instance, 3 ÷4
is not in N, but it is in Z! The algebraists say that N is a monoid
and Z is a group (see any advanced Algebra course), relative to the
addition. We can also introduce a multiplication in Z. First of all, if
:. : are in N and both are not zero (otherwise we put : : = 0), we
de…ne : :
act
= :: by : + : + ... + :. : times. For extending this
operation to Z, we put by de…nition (÷:): = :(÷:) = ÷(::). for
any pair :. : of N. The algebraists say that Z is a ring relative to the
addition and this last de…ned multiplication (see the Algebra course).
We use here freely the elementary basic properties of the addition and
multiplication. For instance, 5 (7 ÷ 9) = 5 7 ÷ 5 9. because of the
distributive property.
We also have a dynamic interpretation of the set N. 0 is for (. 1
is for the extremity ¹
1
of the vector
÷÷
(¹
1
. 2 is for the extremity of the
vector
÷÷
(¹
2
which is twice the vector
÷÷
(¹
1
. etc. We must remark that
we just have chosen "an orientation" on the line (d). namely, we started
our above construction "from ( to the right", not "to the left". So,
on (d) one has two orientations: the direct one, "to the right" and the
inverse one, "to the left". If we construct everything again, "on the
left" (by symmetry) we get the set of negative integers: ÷1. ÷2. ÷3....
. The whole set Z = ¦.... ÷3. ÷2. ÷1. 0. 1. 2. 3. ...¦ is called the set of
integers.
By "Arithmetic" we mean all the properties of N (or Z) derived from
the "algebraic" operations of addition and multiplication. A prime
number j is a natural number distinct of 1. which cannot be written as
a product j = ::. where : and : are natural numbers, both distinct of
1 (or of j). For instance, 2. 3. 5. 7. 11. 13. 17. ... are prime numbers. Any
natural number : greater than 1 is either a prime number or it can be
decomposed into a …nite product of prime numbers (Euclid). Indeed,
if : is not a prime number, there are :
1
. :
2.
natural numbers such that
: = :
1
:
2
. where :
1
. :
2
< :. We go on with the same procedure for :
1
4 1. THE REAL LINE.
and :
2
instead of :. etc., up to the moment when : = j
1
j
2
j
3
...j
I
. where
all j
1
. j
2
. .... j
I
are prime numbers. Maybe some of them are equal one
to the other so, we can write : = ¡
n
1
1
¡
n
2
2
...¡
n
h
I
. where ¡
1
. ¡
2
. .... ¡
I
are
distinct primes. All the scientists up to C. F. Gauss considered the
uniqueness of ¡
1
. ¡
2
. .... ¡
I
to be "evident". Gauss was the …rst who
proved what we call now "The Fundamental Theorem of Arithmetic":
Theorem 1. Any natural number : greater than 1 is either a
prime number or it can be uniquely written as : = ¡
n
1
1
¡
n
2
2
...¡
n
h
I
. where
¡
1
. ¡
2
. .... ¡
I
are distinct prime numbers.
All the other basic results in number theory are directly or indi
rectly connected with this main result. For instance, Euclid proved
that the set of all prime numbers is in…nite. Indeed, if it was not so,
let ¡
1
. ¡
2
. .... ¡
.
be all the distinct primes. Then, let us consider the
natural number : = ¡
1
¡
2
...¡
.
+ 1. It is either a prime number or it
is divisible by a prime number j. Since ¡
1
. ¡
2
. .... ¡
.
are all the prime
numbers, this j must be equal to a ¡
;
for a , ÷ ¦1. 2. .... `¦. Then 1 is
divisible by ¡
;
. a contradiction (Why?). Thus, our assumption is false,
i. e. the set of prime numbers is in…nite. The most delicate hypotheses
and results in Mathematics are connected with THIS SET.
Recall that a function 1 : A ÷ ). where A and ) are arbitrary
sets, is said to be injective (or onetoone) if for any pair of distinct
elements c and / from A. their images 1(c) and 1(/) are distinct in ).
1 is surjective (or onto... ) ) if any element n of ) is the image of an
element r of A. i. e. n = 1(r). Injective + surjective means bijective.
If 1 is bijective we simply say that it is "a bijection" between the sets
A and ). Or that they have "the same cardinal ". For instance, N and
Z have the same cardinal because 1 : N ÷ Z, 1(0) = 0. 1(2:) = ÷:
and 1(2: ÷1) = :. for : = 1. 2. ... is a bijection (Why?).
Generally, if a set ¹ has the same cardinal with N we say that it
is countable. If a set 1 has the same cardinal with a set of the form
¦1. 2. .... :¦ we say that it is …nite and that it has : elements, or that
its cardinal is :. Why a set ¹ cannot be …nite and countable at the
same time?
Any countable set ¹ can be represented like a sequence: c
0
= 1(0).
c
1
= 1(1). c
2
= 1(2). ... where 1 : N ÷¹ is a bijection between N and
¹ (see the de…nition of countability!). Conversely, any set ¹ which
can be represented like a sequence is countable, i.e. it is the image
of the natural number set N through a bijection 1 (prove this!). An
in…nite subset ¦/
n
¦ of a sequence ¦c
a
¦ is called a "subsequence" of
the sequence ¦c
a
¦ if there is a sequence /
1
< /
2
< ... < /
a
< ... of
natural numbers such that for any : ÷ N, /
n
is equal to c
Im
. Here, by
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 5
¦c
a
¦ we mean the set ¦c
0
. c
1
. c
2
. .... c
a
. ...¦ and c
a
is called the general
term of the sequence ¦c
a
¦. For instance ¦/
I
= 2/¦. / = 0. 1. 2. ... is
a subsequence of N = ¦0. 1. 2. ...¦. But the sequence ¦0. 1. 2. 2. 2. ...¦ is
NOT a subsequence of N (Why?). Yes, the set ¦0. 1. 2¦ IS a subset of
N, but not ...a subsequence! Can N be a subsequence of Z?
Now our question is: "How do we represent 2 kg and a quarter
on the line (d)?" More exactly, to the point ( on (d) which is the
extremity of a vector
÷÷
((, obtained by taking
÷÷
(¹
1
twice + a quarter
from the same vector
÷÷
(¹
1
. what kind of sequence of digits 0. 1. 2. .... 9
could we associate? Let us divide the segment [(¹
1
] into 10 equal
parts and let us associate the symbol 0.1 to the extremity ¹
[11]
of
the vector
÷÷
(¹
[11]
which is the 10th part of
÷÷
(¹
1
. In the same way
we construct ¹
[12]
. ¹
[13]
. .... ¹
[19]
and their corresponding symbols 0.2.
0.3. .... 0.9. We continue by dividing the segment [(¹
[11]
] into 10 equal
parts and obtain the new symbols 0.01. 0.02. ..... 0.09, etc. We say
that 0.1 =
1
10
. 0.01 =
1
100
. and so on. For instance, the sequence (or the
number) 23.0145 represents the point 1 on (d) obtained in the following
way. To the vector
÷÷
(¹
23
we add:
1
100
÷÷
(¹
1
+
4
1000
÷÷
(¹
1
+
5
10000
÷÷
(¹
1
. The
resultant vector is
÷÷÷
(1. etc. If one works (by symmetry) on the left of
(. one gets the "negative" numbers of the form: ÷c
a
c
a÷1
...c
0
./
1
/
2
.../
n
,
where c
j
and /
;
are digits from the set ¦0. 1. 2. ....9¦. This last number
can be written as:
÷(10
a
c
a
+ 10
a÷1
c
a÷1
+... +c
0
+
/
1
10
+
/
2
10
2
+... +
/
n
10
n
)
(1.4) = ÷
c
a
c
a÷1
...c
0
/
1
/
2
.../
n
10
n
Here appeared fractions like
o
b
. where c and / are natural numbers
and / = 0. We suppose that the reader is familiar with the operations of
addition, subtraction, multiplication and division with such fractions.
If c ÷ Z and / = 10
n
. from this discussion, we have the geometrical
meaning of the fraction
o
b
. We also call any fraction, a number. What
is the geometrical meaning of
4
7
? Take again the vector
÷÷
(¹
1
and di
vide it into 7 equal parts. Let
÷÷
(G be the 7th part of
÷÷
(¹
1
. Then
4
÷÷
(G =
÷÷÷
(H and H will be the point which corresponds to the number
4
7
. The Greeks said that the number
4
7
is obtained when we want to
measure a segment [(`] with another segment [(`] and if we can …nd
a third segment [(1] such that [(`] = 4[(1] and [(`] = 7[(1]. i.e.
6 1. THE REAL LINE.
[O.]
[OA]
=
4
7
. A representation of a number (for instance a fraction) as
± c
a
c
a÷1
...c
0
./
1
/
2
.../
n
is called a decimal representation (or a decimal
fraction). Let us try to …nd a decimal representation for the fraction
4
7
.
The idea is to write
4
7
as
1
10
40
7
. Then, 40 = 5 7 +5 implies
40
7
= 5 +
5
7
.
where
5
7
< 1. Hence
4
7
=
5
10
+
1
10
5
7
. Now we do the same for
5
7
. Namely,
5
7
=
1
10
50
7
=
1
10
(7 +
1
7
). so
4
7
=
1
10
[5 +
1
10
(7 +
1
7
)] =
5
10
+
7
10
2
+
1
10
2
1
7
.
Write now
1
7
=
1
10
10
7
=
1
10
(1 +
3
7
).
So
4
7
=
5
10
+
7
10
2
+
1
10
3
(1 +
3
7
) =
5
10
+
7
10
2
+
1
10
3
+
1
10
3
3
7
.
Since the remainders obtained by dividing natural numbers by 7 can
be 0. 1. 2. 3. 4. 5. or 6. in the sequence
4
7
.
5
7
.
1
7
.
3
7
. ..., at least one of the
fraction must appear again after at most 7 steps. Thus, let us go on!
Write
3
7
=
1
10
30
7
=
1
10
(4 +
2
7
).
So
4
7
=
5
10
+
7
10
2
+
1
10
3
+
4
10
4
+
1
10
4
2
7
.
But
2
7
=
1
10
20
7
=
1
10
(2 +
6
7
) =
2
10
+
1
10
2
60
7
=
2
10
+
1
10
2
(8 +
4
7
).
So
4
7
=
5
10
+
7
10
2
+
1
10
3
+
4
10
4
+
2
10
5
+
8
10
6
+
1
10
6
4
7
.
But
4
7
=
1
10
40
7
=
1
10
(5 +
5
7
).
Hence
(1.5)
4
7
=
5
10
+
7
10
2
+
1
10
3
+
4
10
4
+
2
10
5
+
8
10
6
+
5
10
7
+...
Since the digit 5 appears again, we must have:
4
7
= 0.5714285714285...
act
= 0.(571428).
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 7
We sat that
4
7
is a simple periodical decimal fraction. Here we meet
with an "in…nite" sum, i.e. with a series:
0.(571428) =
5
10
(1 +
1
10
6
+...) +
7
10
2
(1 +
1
10
6
+...) +...
= (
5
10
+
7
10
2
+
1
10
3
+
4
10
4
+
2
10
5
+
8
10
6
)(1 +
1
10
6
+
1
10
12
+...).
But 1 +
1
10
6
+
1
10
12
+ ... is an in…nite geometrical progression with the
…rst term 1 and the ratio
1
10
6
. The actual mathematical meaning of this
in…nite sum will be explained later.
The next question is if always one can measure a segment c by
another segment / and obtain as a result a fraction
n
a
. Even Greeks
discovered in Antiquity that this operation is not always possible. For
instance, if one wants to measure the diagonal d of a square with the
side c of the same square we obtain a new number
o
o
such that
o
o
2
= 2
(apply Pythagoras’ Theorem). If
o
o
was a fraction
n
a
. where :. : ÷ N,
: = 0 and :. : have no common divisor except 1. then :
2
= 2:
2
and 2 would be a divisor of :. i.e. : = 2:
t
. Thus, 2:
t2
= :
2
and
then : would also have 2 as a divisor, a contradiction. Usually such
a number
o
o
is denoted by
2 because its square is 2. Such numbers
were not accepted by Greeks as being "real" numbers ! But
2 can
be represented on the real line (d). It is the point l which denotes the
extremity of a vector
÷÷
(l such that its length is equal to the length of
the diagonal of a square of side 1 (= the length of
÷÷
(¹
1
). Any fraction
is called a rational number and any other number (like
2) is called
an irrational number.
2 is an algebraic number because it is a root
of an equation with rational coe¢cients (A
2
÷ 2 = 0). We say that a
number is a real number if it is the result of a measurement, i.e. it can
be associated with a point of the real line (d). Up to now we know that
NOT all real numbers can be represented by ordinary fractions (like
2). We shall indicate below a natural way to associate to any point of
the line (d) a decimal fraction, usually in…nite. Recall that to the point
¹
a
(
÷÷
(¹
a
= :
÷÷
(¹
1
) we associated a natural number : (given as a …nite
sequence of digits). The symmetric point of ¹
a
relative to the origin
( was denoted by ¹
÷a
(see Fig.1.1). Our intuition says that any point
` belongs to a segment of the type [¹
a
. ¹
a+1
). where : here can be
positive or nonpositive (i.e. : ÷ Z). We want to associate to the point
` its coordinate r
A
i.e. a decimal number in the interval [:. : + 1) =
the set of all the real numbers (known or unknown up to now!) which
are greater or equal to : and less than : + 1 (relative to the above
lexicographic order). So '
a÷Z
[¹
a
. ¹
a+1
) = all the points of (d). But this
8 1. THE REAL LINE.
last assertion CANNOT be mathematically proved using only previous
simpler results! It is called the Archimedes’ Axiom. In the language of
the real numbers it says that any such number : belongs to an interval
of the type [:. : + 1). This : is called the integral part of : and it is
denoted by [:]. For instance, [3.445] = 3. but [÷3.445] = ÷4. because
÷3.445 ÷ [÷4. ÷3). So, our point ` belongs to an interval of the type
[¹
a
. ¹
a+1
) for ONLY one : = ±c
I
c
I÷1
...c
0
. where c
j
are digits. Let
us divide the segment [¹
a
. ¹
a+1
) into 10 equal parts by 9 points 1
1
.
1
2
. .... 1
9
. such that:
[¹
a
. ¹
a+1
) = [¹
a
act
= 1
0
. 1
1
) ' [1
1
. 1
2
) ' ... ' [1
9
. ¹
a+1
act
= 1
10
).
To these points we obviously associate the following rational numbers:
1
1
÷: + 0.1.
1
2
÷: + 0.2. .... 1
9
÷: + 0.9.
Since ` ÷ [¹
a
. ¹
a+1
). ` belongs to one and only to one subsegment
[1
j
. 1
j+1
). where i ÷ ¦0. 1. .... 9¦. By de…nition we take as the …rst
decimal of r
A
to be this last digit /
1
= i. If ` is just 1
j
we have
r
A
= ±c
I
c
I÷1
...c
0
./
1
. If ` is on the right of 1
j
the actual r
A
will
be greater then the rational number c
I
c
I÷1
...c
0
./
1
and we continue our
above division process. Namely, instead of [¹
a
. ¹
a+1
) we take [1
j
. 1
j+1
)
that ` belongs to and divide this last interval into 10 equal parts by
the points (
0
= 1
j
. (
1
. .... (
9
and (
10
= 1
j+1
. There is only one , such
that ` ÷ [(
;
. (
;+1
). By de…nition, the second decimal of r
A
is /
2
= ,.
If ` = (
;
. then r
A
= ±c
I
c
I÷1
...c
0
./
1
/
2
and r
A
would be a rational
number. If NOT, then we go on with the segment [(
;
. (
;+1
) instead of
[1
j
. 1
j+1
). etc. If at a moment ` will be the left edge of an interval
obtained like above, then r
A
will have a …nite decimal representation,
i.e. it will be a rational number. If ` will never be in this situation,
then r
A
can or cannot be a rational number. For instance, the point
1 which corresponds to the fraction
4
7
is in this last position but, ... it
is represented by a fraction, so r
1
is a rational number. The point \
which corresponds to
2 is in the same position as 1. but r
\
is not a
rational number as we proved above. The segments constructed above,
are contained one into the other:
[¹
a
. ¹
a+1
) [1
j
. 1
j+1
) [(
;
. (
;+1
) ....
If ` is not the left edge of no one of these segments, then their inter
section is exactly ` (Why?).
In general, the following question arises. If one has a tower of closed
segments
[1
1
. l
1
] [1
2
. l
2
] ... [1
a
. l
a
] ...
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 9
on the real line (d). their intersection is empty or not? Our intuition
says that it could not be empty for ever! But,... there is no mathemat
ical proof for this! This is way this last assertion is an axiom, called
the Cantor’s Axiom. Now we can call a real number : any decimal
fraction (…nite or not) of the type:
(1.6) : = ±c
I
c
I÷1
...c
0
./
1
/
2
.../
n
...
We can write this "number" as a sum of some special type of fractions
(1.7) : = ±
10
I
c
I
+... + 10c
1
+c
0
+
/
1
10
+
/
2
10
2
+... +
/
n
10
n
+...
Using this last representation, it is not di¢cult to de…ne the usual
elementary operations of addition, subtraction, multiplication, and di
vision for the set R of all the real numbers (do it and …nd a natural
explanation for the rules you learned in the high school!You must also
use the fact that : = lim
n÷o
:
n
. where
:
n
= ±
10
I
c
I
+... + 10c
1
+c
0
+
/
1
10
+
/
2
10
2
+... +
/
n
10
n
and the usual operations with convergent sequences). The algebraists
say that R together with the addition and multiplication is a …eld (see
the exact de…nition of a …eld in any Algebra course and verify this last
assertion!). Because of the fact that the real numbers are nothing else
than a representation of the points of the real line (together with a
Cartesian reference frame on it!), the Archimedes’s and the Cantor’s
axioms work on R. They can be expressed in the following way (in
language of numbers...):
Axiom 1. (Archimedes’s Axiom) For any real number : there is
one and only one integer number : such that : _ : < : + 1.
Axiom 2. (Cantor’s Axiom) Let c
1
_ c
2
_. .... _ c
a
_. ... and
/
1
_ /
2
_. .... _ /
a
_. ... be two sequences of real numbers such that for
any : one has that c
a
_ /
a
. Then there is at least one real number :
between c
a
and /
a
for any : ÷ N. If in addition, the di¤erence /
a
÷c
a
becomes smaller and smaller to zero, whenever : becomes larger and
larger, then this real number : is unique (in fact, this last assertion is
not an axiom !).
Hence, the real numbers can always be seen like points on a real
line (d). If we change the line and (or) the Cartesian reference frame we
clearly obtain di¤erent sets of real numbers. But,...all these …elds of real
numbers are isomorphic like ordered …elds. This means that for any
two such …elds 1
1
and 1
2
there is at least one bijection 1 : 1
1
÷ 1
2
10 1. THE REAL LINE.
such that 1(r + n) = 1(r) + 1(n). 1(rn) = 1(r)1(n) (1 preserves
the algebraic structure of …elds) and 1(r) _ 1(n). whenever r _ n
(1 preserves the order introduced above). Here r. n ÷ 1
1
. In fact, it
is not di¢cult to construct such a bijection. If we take r ÷ 1
1
. it
is the decimal representation of a point A on the …rst real line (d
1
).
But always one can construct a natural bijection o between the points
of (d
1
) and the points of (d
2
) which carries the Cartesian coordinate
system of the …rst line into the coordinate system of the second line.
Now we take for 1(r) the real number which corresponds to the point
o(A) of the second line (prove that this construction works).
From now on we …x a …eld R of real numbers and we assume that
the reader knows the usual elementary rules of operating in this R. It is
of a great bene…t if one always think of a real number as being a point
on a …xed real line (d). So, ...draw everything or almost everything!
This is why we say a point instead of a number and a number instead
of a point!
We must remark that the correspondence between the points of the
real line (d) and the decimal representations is NOT a bijection. For
instance, 0.999... = 1. But,...the correspondence between the points
of the real line (d) and the real numbers IS a bijection! (Descartes’
bijection).
Let us come back and recall that the set of natural numbers
N = ¦0. 1. .... 9. 10. 11. .... 20. 21. .... :. ...¦
can be naturally embedded in the ring of integers
Z = ¦0. 1. ÷1. 2. ÷2. .... :. ÷:. ...¦.
where : is a natural number. This embedding preserves the usual
operations of addition and multiplication. Both sets N and Z are clearly
countable because they are naturally represented like sequences. What
is the di¤erence between N and Z? The equation A ÷ 3 = 0 has a
solution in N, r = 3. whereas the equation A + 3 = 0 has NO solution
in N, but it has the solution r = ÷3 in Z. The next step is to see that
the general linear equation of the form cA + / = 0. where c. / ÷ Z,
may have no solution in Z. For instance, 2A+1 = 0 has no solution in
Z, but its solution is the fraction
÷1
2
= ÷
1
2
which is a rational number.
Let us denote by Q the …eld of rational numbers and see that any
integer number : can be represented as a rational number: : =
n
1
.
So, N · Z · Q · R, since any rational number is a particular real
number by the de…nition of a real number.
Theorem 2. The rational number …eld Q is also a countable set.
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 11
Proof. It will be enough to represent the positive elements of Q as
a subsequence of a sequence (Why?Use the same trick like in the case
of the countability of Z). Look now carefully to the following in…nite
table
1
1
÷
1
2
1
3
÷
1
4
1
5
÷
1
6
1
7
÷
1
8
. . . . .
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
. . . . .

3
1
3
2
3
3
3
4
3
5
3
6
3
7
3
8
. . . . .
4
1
4
2
4
3
4
4
4
5
4
6
4
7
4
8
. . . . .

5
1
5
2
5
3
5
4
5
5
5
6
5
7
5
8
. . . . .
6
1
6
2
6
3
6
4
6
5
6
6
6
7
6
8
. . . . .

7
1
7
2
7
3
7
4
7
5
7
6
7
7
7
8
. . . . .
8
1
8
2
8
3
8
4
8
5
8
6
8
7
8
8
. . . . .

. . . . . . . . . . . . . . . . . . . .
and to the arrows which indicate "the next term" in the sequence.
This sequence covers ALL the entries of this table and any positive
rational number is an element of this sequence, i.e. Q can be viewed
as a subsequence of this last sequence. Thus Q is countable.
Recall that a real number : is a "union" of two sequences of digits
with + or ÷ in front of it:
(1.8) : = ±c
I
c
I÷1
...c
0
./
1
/
2
.../
a
...
The …rst sequence is always …nite: c
I
. c
I÷1
. .... c
0
. After its last digit
c
0
(the units digit) we put a point "." . Then we continue with the
digits of the second sequence: /
1
. /
2
. .... /
a
. ... . As we saw above, this
last sequence can be in…nite. If this last sequence is …nite, i.e. if from
a moment on /
a+1
= /
a+2
= ... = 0. we say that : is a simple rational
number. Any simple rational number is a fraction of the form
o
10
n
where c ÷ Z and : ÷ N. If : is not a simple rational number, it can be
canonically approximated by the simple rational numbers
:
a
= ±c
I
c
I÷1
...c
0
./
1
/
2
.../
a
.
12 1. THE REAL LINE.
for : = 1. 2. .... This means that when : becomes larger and larger, the
absolute value
(1.9)
c::o:
a
= [: ÷:
a
[ = 0. 00...0
. .. .
a÷tjncc
/
a+1
/
a+2
... =
1
10
a+1
(/
a+1
+
/
a+2
10
+
/
a+3
10
2
+...)
becomes closer and closer to 0. Indeed,
1
10
a+1
(/
a+1
+
/
a+2
10
+
/
a+3
10
2
+...) _
1
10
a+1
(9 +
9
10
+
9
10
2
+...) =
1
10
a
and, since
1
10
n
<
1
a
(prove it!), one gets that [: ÷:
a
[ ÷0 (tends to 0).
when : ÷· (the values of : become larger and larger).
Remark 1. Hence, in any interval (c. /). c = /. c. / real numbers,
one can …nd an in…nite numbers of simple rational numbers (prove it!).
But, what is the mathematical model for the fact that a sequence
. ¦r
a
¦. : = 0. 1. ... tends to 0 (i.e. [r
a
[ becomes closer and closer to 0,
when : becomes larger and larger (: ÷·)?
Definition 1. We say that a sequence ¦r
a
¦. : = 0. 1. ... is conver
gent to 0 (or tends to 0). when : tends to · (: ÷·). if for any posi
tive (small) real number 0. there is a natural number `
.
(depending
on ) such that [r
a
[ < for any : _ `
.
. We simply write this: r
a
÷0.
or, more formally: lim
a÷o
r
a
= 0. or, less formally: limr
a
= 0. We also
say that a sequence ¦r
a
¦. : = 1. 2. ... is convergent to a real number
r (or that r is the limit of ¦r
a
¦; write lim
a÷o
r
a
= r) if the di¤erence
sequence ¦r
a
÷r¦. : = 1. 2. ... is convergent to 0. or, if the "distance"
[r
a
÷r[ between r
a
and r becomes smaller and smaller as : ÷ ·.
This is equivalent to saying that for any positive (small) real number .
all the terms of the sequence ¦r
a
¦. : = 0. 1. .... except a …nite number
of them, belong to the open interval (r ÷ . r + ). Such an interval,
centered at r and of "radius ", is called an neighborhood of r.
Theorem 3. Let ¦r
a
¦ be a convergent sequence. Then its limit is
a unique real number.
Proof. Let us assume that r and r
t
are two distinct limits of the
sequence ¦r
a
¦ and let be a positive small real number such that
< [r ÷r
t
[ . Since both r and r
t
are limits of the sequence ¦r
a
¦. for
: large enough, one must have [r
a
÷r[ <
.
4
and [r
t
÷r
a
[ <
.
4
. Now
< [r
t
÷r[ = [r
t
÷r
a
+r
a
÷r[ _ [r
t
÷r
a
[ +[r
a
÷r[ <
4
+
4
=
2
.
or <
.
2
. a contradiction! So, any two limits of the sequence ¦r
a
¦ must
be equal!
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 13
In (1.9) we have in fact that any real number : can be approximated
by its simple rational number components (or approximates) :
a
. i.e.
lim:
a
= :. We say that the set of simple rational numbers is dense in
R. In particular, Q is dense in R. Let : be a …xed nonzero natural
number and let (
n
be the set of fractions of the form
o
n
n
, where c runs
in Z and : runs in N. Then any real number : is a limit of elements
from (
n
. i.e. (
n
is dense in R (prove it!write : in the basis :. instead
of 10).
We just used above that the sequence ¦
1
a
¦. : = 1. 2. ... is convergent
to 0. Our intuition says that if we divide the unity vector
÷÷
(¹
1
(see
Fig.1.1) into : equal parts, the length
1
a
of one of them becomes smaller
and smaller. But,...why? What is the mathematical explanation for
this?
Theorem 4. The sequence ¦
1
a
¦ is convergent to 0.
Proof. We apply De…nition 1. Let 0 be a small positive real
number and, by using the Archimedes’s Axiom, let `
.
be the unique
natural number such that
1
.
÷ [`
.
÷ 1. `
.
). So, for any : _ `
.
. one
has that
1
.
< `
.
_ :. i.e.
1
a
< .
Remark 2. The absolute value or the modulus [:[ of the real number
: from (1.8) is simply
c
I
c
I÷1
...c
0
./
1
/
2
.../
a
....
i.e. : without minus if it has one. For instance, [÷3.14[ = 3.14 =
[3.14[ . Since the function di:t. which associates to any pair of real
number (r. n) the nonnegative real number [r ÷n[ . i.e. di:t(r. n) =
[r ÷n[ . has the following basic properties (prove them!):
i) di:t(r. n) = 0. if and only if r = n.
ii) di:t(r. n) = di:t(n. r).
iii) di:t(r. n) _ di:t(r. .) +di:t(.. n) (the triangle inequality),
for any r. n. . in R, we say that di:t(r. n) = [r ÷n[ is the distance
between r and n and that R together with this distance function di:t is
a metric space.
Another example of a metric space is the Cartesian plane r(n
with the distance function between two points `
1
(r
1
. n
1
) and `
2
(r
2
. n
2
)
given by the formula:
di:t(`
1
. `
2
) =
÷÷÷÷
`
1
`
2
=
(r
2
÷r
1
)
2
+ (n
2
÷n
1
)
2
.
i.e. the length of the segment [`
1
`
2
]. Here we can see why the property
iii) was called "the triangle property" (be conscious of this by drawing
a triangle in plane...!).
14 1. THE REAL LINE.
Now, what is the di¤erence between the rational number …eld Q
and the real number …eld R? The …rst one is that Q is countable
and, as the following result says, R is NOT countable, so the subset of
irrational numbers is "greater" than the subset of rational numbers.
Theorem 5. (Cantor’s Theorem). The set R is NOT countable,
i.e. one can NEVER represent the whole set of the real numbers as a
sequence.
Proof. Let : be like in (1.8). It is enough to prove that the set o
of all the sequences ¦/
1
. /
2
. .... /
a
. ...¦. where /
a
is a digit, is not count
able. Suppose on the contrary, namely that o can be represented like
a sequence of ... sequences: o = ¦1
1
. 1
2
. .... 1
a
. ...¦. where
1
a
= ¦/
a1
. /
a2
. /
a3
. .... /
aa
. ...¦.
and /
a;
are digits. In order to obtain a contradiction, it is enough
to construct a new sequence of digits, which is distinct of any 1
j
for
i = 1. 2. ... . Let ( = ¦c
1
. c
2
. .... c
a
. ...¦ with the following property:
c
a
= /
aa
+1. if /
aa
= 9 and c
a
= 0. if /
aa
= 9. Now, let us see that ( is
not in o. Assume that ( = 1
I
for a / ÷ ¦1. 2. ...¦. By the de…nition of
c
I
, this last one cannot be equal to /
II
. thus the /th term of ( is not
equal to the /th term of 1
I
and so, ( = 1
I
. a contradiction! Hence
( ÷ o. So o cannot be represented like a sequence.
It is not di¢cult to prove that the subset of R which consists of all
the algebraic elements over Q (roots of polynomials with coe¢cients in
Q) is countable. So, R contains an uncountable subset of transcendental
numbers (numbers which are not algebraic). In fact we know very
few of them, c. :. c
2
. etc. A real number which is not rational is
called an irrational number. Since any interval (c. /) is in a oneto
one correspondence onto the interval (0. 1) (1 : (0. 1) ÷ (c. /). 1(t) =
c + (/ ÷ c)t is a bijection between (0. 1) and (c. /)) and since tan :
(÷
¬
2
.
¬
2
) ÷R is a bijection between (÷
¬
2
.
¬
2
) and R, there is a bijection
between R and any nontrivial interval (c. /). does not matter as small
as this last interval is.
Remark 3. Hence, (c. /) with c = / is not countable. Thus, in
(c. /) one can …nd an in…nite number of irrational numbers and even
an in…nite number of transcendental numbers (why?explain step by
step!).
Can we solve any equation in R ? The answer is NO! Even the
simple equation A
2
+ 1 = 0. with the coe¢cients in Z has no real
solution. Why? Because r = 0 is not a solution and, if r = 0. then r
2
is positive (see the multiplication rule of signs!). So, r
2
+ 1 is greater
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 15
than 1. thus it cannot be zero. In order to solve this last equation we
need to enlarge R up to another …eld C, the complex number …eld.
Its algebraic structure is the following. Take the 2dimensional real
vector space \ = R R with the componentwise addition and the
componentwise scalar multiplication. Then we introduce a "strange"
multiplication:
(1.10) (c. /)(c. d)
oc;
= (cc ÷/d. cd +/c).
It is not di¢cult to prove that \ together with this multiplication
becomes a …eld in which (0. 1)
2
= (÷1. 0). identi…ed with the real
number ÷1. because c ÷ (c. 0) is a canonical embedding of R into
\. This new …eld is usually denoted by C. It is clear that ±(0. 1) are
the solutions of the equation A
2
+ 1 = 0. What is amazing is that C.
F. Gauss proved that any polynomial with coe¢cients in C has all its
roots in C. The algebraists say that C is algebraically closed (it cannot
be enlarged by adding to it new roots of polynomials with coe¢cients
in it). Later, Frobenius proved that there is no other super…eld of R,
which has a …nite dimension over it, but C (which has dimension 2
over R). Here dimension means the dimension of C as a vector space
over R. Since any . = c + i/. where i = (0. 1) and c. / are unique real
numbers, ¦(1. 0). (0. 1)¦ is a basis in C. So the dimension of C over R
is 2.
Let us now come back to our problem relative to the di¤erences
between Q and R. Since Q is a sub…eld of R, the Archimedes Axiom
also works on Q. But, what about Cantor’s Axiom? We know that
2 is not in Q. Let us consider the (in…nite) decimal representation of
2 :
(1.11)
2 = 1.41/
3
/
4
.../
a
...
and let us denote by r
a
= 1.41/
3
/
4
.../
a
. the corresponding :th simple
rational number of
2. It is clear that the sequence ¦r
a
¦ is an increas
ing sequence which converges to
2. Let us also consider the following
decreasing sequence ¦n
a
¦ of simple rational numbers, convergent to
the same
2. n
1
= 1.5. n
2
= 1.42. .... n
a
= 1.41/
3
/
4
.../
a÷1
c
a
/
a+1
/
a+2
....
where c
a
= /
a
+ 1. if /
a
= 9 and c
a
= /
a
= 9. if /
a
= 9. It is easy to
see that the intersection of all the closed intervals [r
a
. n
a
]. : = 1. 2. ....
in Q, is empty in Q (since the intersection in R is exactly
2. which is
not in Q). Hence the Cantor axiom does not work for the ordered …eld
Q.
In this last counterexample we needed some tricks, so it will be
desirable to have an equivalent statement to the Cantor’s Axiom. For
this we introduce two important new notions, namely the notion of the
16 1. THE REAL LINE.
least upper bound (LUB) and the notion of the greatest lower bound
(GLB) of a given subset of R. We do everything for the LUB and we
leave to the reader to translate all of these in the case of the GLB.
Let ¹ be a nonempty subset in R. A real number . is called an
upper bound for ¹ if any element c of ¹ is less or equal to .. A least
upper bound (LUB) for ¹ is (if it does exist!) the least possible . which
is an upper bound for ¹. For instance, the LUB of ¹ = [0. 7) is 7 and
the GLB of ¹ is 0. We cannot have two distinct LUB for the same
subset ¹ (Why?). If ¹ is (upper) unbounded (i.e. if for any natural
number : there is at least one element / of ¹ such that / :), then ¹
has no upper bound in R and as a logical consequence it has no LUB
in R. For instance, ¹ = [0. ·) has no upper bound in R, but 0 is the
GLB of ¹. R and Z have neither an LUB nor a GLB in R.
Usually, the LUB of a subset ¹ is denoted by sup ¹ (the supremum
of ¹) and the GLB of a subset 1 is denoted by inf 1 (in…mum of 1).
Theorem 6. (LUB test) Let ¹ be a subset of R. Then c is the LUB
of ¹ if and only if for any small positive real number 0. there are
an element c of ¹ such that c ÷ < c _ c and an upper bound . of ¹
with c _ . < c+. This is equivalent to saying that any neighborhood
of c must simultaneously contain an element c of ¹ and an upper bound
. of ¹ (Why?).
Proof. Let us suppose that c = sup ¹. Assume that we found an
0 such that all the elements of ¹ are less or equal to c ÷ . So
c ÷ is an upper bound of ¹ less than c. a contradiction, because, by
de…nition, c is the least upper bound of ¹. Hence, there is at least one
c ÷ ¹ in the interval (c÷. c]. If all the upper bounds of ¹ were greater
or equal to c +. then c would not be the least upper bound of ¹ and
we would obtain again a contradiction.
Conversely, let us assume that c is a real number with the property
described in the statement of the above theorem. If c were not sup ¹.
we have two options: 1) c is not an upper bound of ¹. i.e. there is
at least one c greater than c. Taking now = c ÷ c and using our
hypothesis for this particular 0. we get an upper bound . of ¹ in
the interval [c. c + = c). i.e. . is less than c. This is in contradiction
with the fact that . is an upper bound of ¹. Hence 1) cannot appear.
It remains only the second option: 2) c is an upper bound of ¹, but
it is not the least, namely there is another upper bound n which is
less than c. Take now = c ÷ n 0 and use again the hypothesis of
the theorem for this new . So, one can …nd an element / of ¹ in the
interval (c÷ = n. c]. Thus, / is greater than n. which was considered to
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 17
be an upper bound of ¹. Again a contradiction! Therefore, the second
option is also impossible and the proof is complete.
The LUB test is very useful because it supply us with some impor
tant results.
Theorem 7. The following statements are logically equivalent: i)
The Cantor Axiom (see Axiom 2) works in R, ii) Any upper bounded
subset ¹ of R has a LUB in R and, iii) Any lower bounded subset 1
of R has a GLB in R.
Proof. First of all let us see that ii) and iii) are equivalent. Let
us prove for instance that ii)= iii). For the lower bounded subset 1
of R let us put ÷1 = ¦r ÷ R : ÷r ÷ 1¦. the symmetric subset of 1
with respect to the origin ( (on the real line (d)). It is not di¢cult to
see that the new subset ÷1 is upper bounded in R and so, from ii) it
has a LUB / in R. We leave the reader (eventually using Theorem 6)
to prove that ÷/ is the GLB of 1 in R.
We leave as an exercise for the reader to prove that iii)== i).
Now we prove that i)== ii). Let /
0
be an upper bound of ¹ and
let c
0
be an element of ¹. It is clear that c
0
_ /
0
. If c
0
= /
0
we have
nothing more to prove because the LUB of ¹ will be this common value
c = c
0
= /
0
. Assume that c
0
is less than /
0
an let us divide the closed
interval [c
0
. /
0
] into two equal closed subintervals by the mid point c
0
.
By the "essential choice" we mean to choose the subinterval [c
0
. c
0
] if c
0
is an upper bound for ¹. or to choose the subinterval [c
0
. /
0
] if there is
at least one element c
t
1
÷ ¹ in the second subinterval, [c
0
. /
0
]. After we
have performed "the essential choice", let us denote by [c
1
. /
1
] either
the subinterval [c
0
. c
0
] in the …rst choice, or the subinterval [c
0
. /
0
] in
the case of the second choice. In both situations c
1
÷ ¹. /
1
is an upper
bound of ¹ and c
0
_ c
1
_ /
1
_ /
0
. Now we take the interval [c
1
. /
1
].
divide it into two equal parts and repeat the "essential choice" for this
new interval [c
1
. /
1
], …nd c
2
÷ ¹ and /
2
an upper bound of ¹ with
c
0
_ c
1
_ c
2
_ /
2
_ /
1
_ /
0
and so on. We obtain two sequences: an increasing one and a decreasing
one in the following position:
c
0
_ c
1
_ ... _ c
a
_ ... _ /
a
_ ... _ /
1
_ /
0
.
such that the distance di:t(c
a
. /
a
) =
ojct(o
0
.b
0
)
2
n
. In particular, di:t(c
a
. /
a
) ÷
0. whenever : ÷ ·. Now we can apply the Cantor Axiom and …nd a
unique point c belonging to all the intervals [c
a
. /
a
] for any : = 1. 2. ....
i. e. limc
a
= lim/
a
= c (Why?). We prove now that this c is exactly
sup ¹. Let us now apply the LUB test (see Theorem 6). Take an and
18 1. THE REAL LINE.
let us consider the neighborhood (c÷. c+). Since limc
a
= lim/
a
=
c. there is an : ÷ ¦1. 2. ...¦ such that [c
a
. /
a
] · (c ÷ . c + ). But, by
the above construction, c
a
÷ ¹ and /
a
is an upper bound of ¹. So, by
the criterion of Theorem 6, we get that c = sup ¹.
ii)== i) Let ¦c
a
¦ and ¦/
a
¦ be two sequences of real numbers such
that
c
0
_ c
1
_ ... _ c
a
_ ... _ /
a
_ ... _ /
1
_ /
0
.
The subset ¹ = ¦c
0
. c
1
. .... c
a
. ...¦ is upper bounded in R by any term of
the second sequence ¦/
a
¦. From ii) we have that ¹ has a LUB c = sup ¹
and c _ /
a
for any : = 0. 1. ... . Since c is in particular an upper bound
of ¹. one also has that c
a
_ c _ /
a
for any : = 0. 1. ... . Hence the
Cantor Axiom works on R.
A sequence is said to be monotonous if it is either an increasing or
a decreasing sequence. For instance, r
a
=
1
a
2
+1
and n
a
= ÷
1
a
2
+1
are
monotonous sequences.
Remark 4. Let us now introduce two symbols: 1) ·, which is
considered to be greater than any real number :, : +·= ·. ·+·=
·. and 2) ÷·. which is considered to be less then any real number :,
: +(÷·) = ÷·. ÷·÷(·) = ÷·, : ·= ·. if : 0. : ·= ÷·.
if : < 0. Moreover, : (÷·) = ÷· if : 0 and : (÷·) = ·. if : is
negative. In the same logic,
··= (÷·)(÷·) = ·. (÷·)·= ÷·= ·(÷·).
:
±·
= 0. ctc.
The operations 0(±·). ·÷·.
0
0
and
o
o
are not permitted. We denote
by R = ¦÷·¦ ' R ' ¦·¦ and call it the accomplished (or completed)
real line. By de…nition, a neighborhood of · is an open interval of
the form (`. ·) and a neighborhood of ÷· is an interval of the form
(÷·. 1). where `. 1 are real numbers. For instance, in R any subset
of real numbers is bounded (upper or lower) and an unbounded (in R)
increasing sequence is convergent to · (for example, r
a
= :
3
÷ ·).
But the sequence n
a
= (÷1)
a
: is bounded in R but it is not convergent
there (Why?). If ¹ = ¦r
a
¦ is a sequence in R. we denote by limsup r
a
.
the sup ¹ and by liminf r
a
. the inf ¹. For instance, for the sequence
r
a
= sin(
2a+1
2
:) = (÷1)
a
. limsup r
a
= 1 and liminf r
a
= ÷1 (prove
this!).
It is easy to see that limsup r
a
is the greatest limit of all the
convergent subsequences of ¦r
a
¦. What is liminf r
a
?
Theorem 8. a) Let ¦r
a
¦ be an increasing sequence in R. Then
limsup r
a
exist in R and the sequence is convergent to limsup r
a
in R.
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 19
If ¦r
a
¦ is also upper bounded in R, then limsup r
a
is its limit in R
too, i.e. limr
a
= limsup r
a
. b) Let ¦n
a
¦ be a decreasing sequence in
R. Then liminf r
a
always exist in R and the sequence is convergent to
liminf r
a
in R. If ¦r
a
¦ is also lower bounded in R, then liminf r
a
is
also in R and so limr
a
= limsup r
a
.
Proof. We prove only a) and we think that b) is a good exercise
for the reader. If ¦r
a
¦ is upper unbounded then, for any real number
`. there is at least one : with r
a
_ `. Since ¦r
a
¦ is an increasing
sequence, r
a+j
_ r
a
for any j = 1. 2... . So, outside the neighborhood
(`. ·) of · we have only a …nite number of terms of our sequence,
i.e. r
a
÷ ·. which is at the same time limsup r
a
(Why?). If ¦r
a
¦ is
upper bounded, then, using Theorem 7, we get that c = limsup r
a
is a
real number. Take now an neighborhood (c ÷. c +) of c. Since c is
the LUB of the set ¦r
a
¦. we can apply Theorem 6 and …nd an r
n
in the
interval (c ÷ . c]. Since the sequence is increasing, r
n+1
. r
n+2
. ... are
in the same interval (Why?). So, outside this interval one has at most
a …nite number of terms of our sequence, i.e. r
a
÷ c (see De…nition
1).
Let us come back to the approximation of
2 = 1.41/
3
/
4
.../
a
...
(see (1.11)) by the increasing sequence r
a
= 1.41/
3
/
4
.../
a
. : = 1. 2. ...
of simple rational numbers. This last sequence ¦r
a
¦ is a sequence
in Q but its limit
2 is not in Q. However, this sequence has an
interesting property. If we …x an : ÷ N, and if we consider the terms
r
a
. r
a+1
. r
a+2
. ...r
a+j
. we see that the distance between r
a
and r
a+j
goes to 0 independently of j ÷ N, but dependently of :. This means
that from a rank ` on the distance di:t(r

. r
n
) becomes smaller and
smaller (. : _ `). Indeed,
di:t(r
a
. r
a+j
) = 0. 00...0
. .. .
a÷tjncc
/
a+1
/
a+2
.../
a+j
_ 0. 00...0
. .. .
a÷tjncc
999... =
1
10
a
÷0
independently on j. i.e. for any small real number 0. there is a
rank `
.
such that whenever : _ `
.
one has that di:t(r
a
. r
a+j
) < .
for any j = 1. 2. ... .
Definition 2. Let ¦r
a
¦ be a sequence of real numbers. We say
that ¦r
a
¦ is a Cauchy sequence or a fundamental sequence if for any
small positive real number 0. there is a rank `
.
(depending on )
such that [r
a+j
÷r
a
[ < for any : _ `
.
and for any j = 1. 2. .... This
means that [r
a+j
÷r
a
[ ÷0. when : ÷·. independently on j.
For instance, the above sequence r
a
= 1.41/
3
/
4
.../
a
. : = 1. 2. ... is a
Cauchy sequence of rational numbers which is NOT convergent in Q,
20 1. THE REAL LINE.
but which is convergent in R, its limit being the real number
2. This
is why we say that Q is not "complete".
Definition 3. In general, a metric space A with its distance di:t
(see Remark 2) is said to be complete if any Cauchy sequence ¦r
a
¦ with
terms in A is convergent to a limit r of A.
Let us consider the following sequence
r
a
=
cos 1
2
+
cos 2
2
2
+
cos 3
2
3
+... +
cos :
2
a
.
where the arcs are measured in radians. Let us prove that this last
sequence is a Cauchy sequence. For this, let us evaluate the distance
di:t(r
a
. r
a+j
) = [r
a+j
÷r
a
[ =
=
cos(: + 1)
2
a+1
+
cos(: + 2)
2
a+2
+... +
cos(: +j)
2
a+j
<
<
1
2
a+1
(1 +
1
2
+
1
2
2
+...) =
1
2
a
.
This last equality comes from the de…nition of the in…nite geomet
rical progression
1 +
1
2
+
1
2
2
+...
oc;
= lim
a÷o
1 +
1
2
+
1
2
2
+... +
1
2
a
= lim
a÷o
1 ÷
1
2
n+1
1 ÷
1
2
= 2
So di:t(r
a
. r
a+j
) tends to 0 independently of j. because
1
2
n
goes to
0. whenever : ÷·. independently of j. Indeed, for a small 0. let
us …nd the …rst natural number `
.
such that
1
2
N"
< . Applying log
2
we get `
.
÷log
2
. so `
.
= [÷log
2
] + 1. Now, if : _ `
.
.
di:t(r
a
. r
a+j
) <
1
2
a
_
1
2
."
< .
independently on j.
Theorem 9. Any convergent sequence ¦r
a
¦ to r is also a Cauchy
sequence. Thus, the class of Cauchy sequences "appears" to be larger
then the class of convergent sequences.
Proof. We simply verify De…nition 2. Let be a positive small real
number and let `
.
be a rank (dependent on ) such that [r
a
÷r[ <
.
2
for any : _ `
.
(see De…nition 1 with
.
2
instead of ). So,
[r
a+j
÷r
a
[ = [r
a+j
÷r +r ÷r
a
[ _ [r
a+j
÷r[ +[r
a
÷r[ _
2
+
2
=
for any : _ `
.
. Hence our convergent sequence is also a Cauchy se
quence.
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 21
A basic result in Mathematics was discovered by Cauchy: "Any
fundamental sequence of real numbers is convergent to a real number,
i.e. R is a "complete metric space".
To prove this important result we need some speci…c properties of
the Cauchy sequences.
Theorem 10. Any Cauchy sequence ¦r
a
¦ is bounded, i.e. there is
a positive real number ` such that [r
a
[ _ ` for any : = 0. 1. ... or,
equivalently, if there is an interval [¹. 1] in R such that all the terms
of the sequence ¦r
a
¦ belong to this interval, i.e. r
a
÷ [¹. 1] for any
: = 0. 1. ... (Why this equivalence?).
Proof. Take an arbitrary positive real number, for instance 2.
Since ¦r
a
¦ is a Cauchy sequence, there is a rank ` such that whenever
: _ `. [r
a+j
÷r
a
[ < 2 for any j = 1. 2... (see De…nition 2). In
particular, [r
.+j
÷r
.
[ < 2. or r
.+j
÷ (r
.
÷2. r
.
+2) for any j ÷ N.
So, outside this last interval one may have at most r
0
. r
1
. .... r
.÷1
as
terms of our sequence. Take now ¹ = min¦r
0
. r
1
. .... r
.÷1
. r
.
÷ 2¦
and 1 = max¦r
0
. r
1
. .... r
.÷1
. r
.
+ 2¦. It is easy to see that all the
terms of the sequence ¦r
a
¦ belong to the interval [¹. 1]. If one takes
now ` = max¦[¹[ . [1[¦. then r
a
÷ [÷`. `]. or [r
a
[ _ ` for any
: = 0. 1. ... .
Here is a strange property of the Cauchy sequences.
Theorem 11. If a Cauchy sequence ¦r
a
¦ contains at least one sub
sequence ¦r
In
¦. (/
0
< /
1
< /
2
< ... < /
a
< ... ) which is convergent to
r. then the whole sequence ¦r
a
¦ is convergent to the same r. Therefore,
all the other subsequences of ¦r
a
¦ are convergent to r.
Proof. Let be a small positive real number. Since ¦r
In
¦ is con
vergent to to r whenever : ÷ ·. for : large enough, let us assume
that for : _ `
t
. one has
(1.12) [r
In
÷r[ <
2
.
Since ¦r
a
¦ is a Cauchy sequence, for : large enough, suppose : _ `
tt
.
one has that
(1.13) [r
a+j
÷r
a
[ <
2
.
for any j = 1. 2. ... . Let now ` be a natural number greater than
`
t
and than `
tt
. at the same time. Let : be a …xed natural number
greater than ` and let us choose /
n
such that it is greater than this
…xed : and : itself is greater than `. So, /
n
= : + j. for a natural
22 1. THE REAL LINE.
number j (= /
n
÷:). From (1.13) we get that
(1.14) [r
Im
÷r
a
[ <
2
.
because : ` `
tt
. From (1.12) one has that
(1.15) [r
Im
÷r[ <
2
.
because : ` `
t
. Now,
[r
a
÷r[ = [r
a
÷r
Im
+r
Im
÷r[ _ [r
Im
÷r
a
[ +[r
Im
÷r[ <
2
+
2
= .
And this is true for any : `. Hence, the sequence ¦r
a
¦ is convergent
to r. We leave to the reader to convince himself (or herself) that if a
sequence ¦r
a
¦ is convergent to a real number r. then any subsequence
of it is also convergent to the same r.
We prove now a basic property of a bounded in…nite subset ¹ of
real numbers. For this we give a de…nition.
Definition 4. We say that a subset ¹ of real numbers has the point
(real number) r as a limit point if there is a nonconstant sequence ¦c
a
¦.
with terms c
a
from ¹. which is convergent to r.
For instance, 0 is a limit point of
¹ = ¦1.
1
2
.
1
3
. ....
1
:
. ...¦
and of the interval [0. 1]. But 0 is NOT a limit point of the set 1 =
¦0. 1. 2¦ (Why?). N and Z have NO limit points in R! (Why?). Find
all the limit points of Q in R! (Hint: the whole R is the set of all the
limit points of Q, why?)
Theorem 12. (CesaroBolzanoWeierstrass Theorem). Any in…
nite and bounded subset ¹ of R has at least one limit point in R, i.e.
there is an r ÷ R and a nonconstant sequence ¦c
a
¦ with c
a
÷ ¹ for
any : = 0. 1. ... , such that c
a
÷r.
Proof. Since ¹is bounded, there is a closed interval [c
0
. /
0
] (c
0
. /
0
÷
R) which contains ¹. Let us divide this last interval into two equal
closed subintervals and let denote by [c
1
. /
1
] that subinterval which con
tains an in…nite number of elements of ¹. Let r
1
be in [c
1
. /
1
] and in ¹,
i.e. r
1
÷ [c
1
. /
1
]¨¹. Let us divide now the interval [c
1
. /
1
] into two equal
closed subintervals and let us choose that one [c
2
. /
2
] which contains an
in…nite number of elements from¹. Let r
2
be in ¹¨[c
2
. /
2
] and r
2
= r
1
.
We continue to construct subintervals [c
3
. /
3
]. [c
4
. /
4
]. .... [c
a
. /
a
]. ... and
elements r
a
of ¹ ¨ [c
a
. /
a
]. such that r
a
÷ ¦r
1
. r
2
. .... r
a÷1
¦ for any
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 23
: = 3. 4. .... :. ... . Since the length of the interval [c
a
. /
a
] is

2
n
. where
 is /
0
÷c
0
. the length of the initial interval, we can use Cantor Axiom
(Axiom 2) and …nd a unique real number r in the common intersection
o
¨
a=0
[c
a
. /
a
] of all the intervals [c
a
. /
a
]. Since r
a
and r are in [c
a
. /
a
].
di:t(r
a
. r) _

2
n
so, r
a
÷r (see De…nition 1). Because r
a
. : = 1. 2. ...
are distinct elements of ¹. one has that r is a limit point of ¹ and the
theorem is completely proved.
Theorem 13. (Cauchy test 1). Any fundamental (Cauchy) se
quence in R is convergent in R, i.e. R is a complete metric space.
This means that in R there is no di¤erence between the set of conver
gent sequences and the set of Cauchy sequences (In Q there is!Why?)
Proof. Let ¦n
a
¦ be a fundamental sequence in R. If ¦n
a
¦ has
only a …nite distinct terms then, from a rank on, the sequence becomes
a constant sequence, so it would be convergent to the value of the
constant terms. Let us assume that ¦n
a
¦ has an in…nite number of
distinct terms, i.e. that the set ¹ = ¦n
a
¦ is in…nite. Since ¹ is bounded
(see Theorem 10) and in…nite, it has a limit point n (see Theorem
12), i.e. there is a nonconstant subsequence ¦n
In
¦. : = 1. 2. ... of the
sequence ¦n
a
¦. which is convergent to n. We apply now Theorem 11
and …nd that the whole sequence ¦n
a
¦ is convergent to n.
This theorem has not only a great theoretical importance, but a
practical one too. For instance, take again the sequence
r
a
=
cos 1
2
+
cos 2
2
2
+
cos 3
2
3
+... +
cos :
2
a
.
We proved that ¦r
a
¦ is a Cauchy sequence. Now, we know (see The
orem 13) that it is also a convergent sequence to an unknown limit
(we cannot express this limit as a decimal fraction!) r. Knowing that
r
a
÷ r is a very good situation! For a large : we can approximate r
with r
a
. But this last one can be easily computed with an usual com
puter. So, we have a good idea about the limit. Moreover, the Cauchy
test 1 is useful to check if a sequence is convergent or not. For instance,
the sequence ¦c
a
¦ is recurrently de…ned: c
0
= 0. c
a
=
2 +c
a÷1
for
: = 1. 2. ... . Let us prove that it is a Cauchy sequence. Indeed,
(1.16) c
a
÷c
a÷1
=
2 +c
a÷1
÷
2 +c
a÷2
=
c
a÷1
÷c
a÷2
2 +c
a÷1
+
2 +c
a÷2
<
1
2
(c
a÷1
÷c
a÷2
).
24 1. THE REAL LINE.
We can apply (1.16) (: ÷1)times and …nd
c
a
÷c
a÷1
<
1
2
(c
a÷1
÷c
a÷2
) <
1
2
2
(c
a÷2
÷c
a÷3
) < ... <
1
2
a÷1
(c
1
÷c
0
).
So,
c
a+j
÷c
a
= c
a+j
÷c
a+j÷1
+c
a+j÷1
÷c
a+j÷2
+... +c
a+1
÷c
a
<
< (
1
2
a+j÷1
+
1
2
a+j÷2
+... +
1
2
a
)(c
1
÷c
0
) <
<
1
2
a
(1 +
1
2
+
1
2
2
+...)(c
1
÷c
0
) =
1
2
a÷1
(c
1
÷c
0
).
Here we just used that
1 +
1
2
+
1
2
2
+...
oc;
= lim
a÷o
(1 +
1
2
+... +
1
2
a
) = lim
1 ÷
1
2
n+1
1 ÷
1
2
= 2.
Since ¦c
a
¦ is an increasing sequence (Why?), one has that
[c
a+j
÷c
a
[ <
1
2
a÷1
(c
1
÷c
0
).
so, [c
a+j
÷c
a
[ can be made as small as we want when : ÷ ·. inde
pendently on j. Thus, ¦c
a
¦ is a Cauchy sequence (see De…nition 2).
Hence ¦c
a
¦ is convergent to a limit  (see Cauchy test 1). As we shall
see in the following theorem (Theorem 14), we can apply the "oper
ation" lim to the equality: c
a
=
2 +c
a÷1
and …nd:  =
2 +. or
 = 2. Therefore, lim
a÷o
c
a
= 2.
Now, we describe some compatibilities of the "operation" lim (which
associates to a convergent sequence its limit), with the algebraic op
erations "+". " ÷ ". " ". " ÷ ". with the order relation "_ ". with the
functions r
n
.
m
r. exp r. ln r. c
a
. log
o
. c 0. sin r. cos r. tan r. cot r
and with their compositions. This means, ... with all the elementary
functions. We recall a basic de…nition:
Definition 5. Let (A. d
1
) and (). d
2
) be two metric spaces and let
1 : A ÷) be a mapping de…ned on A with values in ). We say that 1
is continuous at r ÷ A (with respect to these metric space structures) if
for any convergent sequence ¦r
a
¦ in A. ¦r
a
¦ ÷ r. i.e. d
1
(r
a
. r) ÷ 0
as : ÷ ·. one has that the corresponding sequence of the images,
¦1(r
a
)¦ is convergent to 1(r) in ). i.e. d
2
(1(r
a
). 1(r)) ÷ 0. when
: ÷·. If 1 is continuous at any r of A. we say that 1 is continuous
in A.
1. THE REAL LINE. SEQUENCES OF REAL NUMBERS 25
All the elementary functions (polynomials, rational functions, power
functions, exponential and logarithmic functions, trigonometric func
tions and their compositions) are continuous on their de…nition do
mains. To prove this, it is not always so easy. For instance, what
do we mean by 3
2
? First of all, we de…ne 3
1
m
, : = 1. 2. .... by the
unique positive real root of the equation A
n
÷3 = 0. Then we de…ne
3
n
m
oc;
=
3
1
m
a
. By 3
÷
5
7
we understand
1
3
5
7
. Then, we approximate
2
with an increasing sequence ¦:
a
¦ of rational numbers, i.e. :
a
÷
2
and :
a
< :
a+1
for any : = 1. 2. .... As we know, we simply take for :
a
the rational number 1./
1
/
2
.../
a
. i.e. we get out all the decimals of
2
from the (: + 1)th decimal on. Now, by de…nition, 3
2
= lim
a÷o
3
vn
. To
prove the existence of this limit is not an easy task. It is su¢cient to
prove that the sequence ¦3
vn
¦ is a Cauchy sequence. But,... even this
one is di¢cult! So, the proof of the continuity of the power function
r ÷3
a
is not so easy at all! This is why we tacitly assume that all the
elementary functions are continuous.
Theorem 14. Let ¦r
a
¦ and ¦n
a
¦ be two convergent sequences to r
and to n respectively. Then:
a) ¦r
a
±n
a
¦ ÷r ±n.
b) ¦r
a
n
a
¦ ÷rn.
c) If n
a
and n are not zero for any : = 0. 1. ... , then ¦
an
&n
¦ ÷¦
a
&
¦.
d) If r
a
_ n
a
for any : = 0. 1. ... , then r _ n.
e) ¦(r
a
)
n
¦ ÷ r
n
for any …xed natural number :.
f )
m
r
a
÷
m
r if : is odd and, for r
a
_ 0.
m
r
a
÷
m
r for any
natural number :.
g) ¦exp r
a
¦ ÷exp r and, if r
a
0. then ¦ln r
a
¦ ÷ln r.
h) ¦c
an
¦ ÷ c
a
and, if r
a
0. ¦log
o
r
a
¦ ÷ log
o
r for any …xed
c 0.
i) sin r
a
÷sin r. cos r
a
÷cos r. tan r
a
÷tan r. cot r
a
÷cot r.
j) Let 1 : ¹ ÷ 1 and o : 1 ÷ ( (¹. 1. ( are subsets in R)
be two functions with the following property: If 1(r
a
) ÷ 1(r) and
o(n
a
) ÷o(n) for ANY convergent sequences ¦r
a
¦ to r and ¦n
a
¦ to n,
then (o·1)(r
a
) ÷(o·1)(r). The functions 1 and o considered here are
continuous on their de…nition domains in the sense of De…nition 5. So,
the composition between two continuous functions is also a continuous
function. The sum, the di¤erence, the product and the quotient of two
continuous functions is also a continuous function.
26 1. THE REAL LINE.
Proof. (Partially) a) Let us prove for instance that ¦r
a
+ n
a
¦ ÷
r +n. For this, let us evaluate the di¤erence:
[r
a
+n
a
÷(r +n)[ = [(r
a
÷r) + (n
a
÷n)[ _ [r
a
÷r[ +[n
a
÷n[ .
But [r
a
÷r[ ÷0 and [n
a
÷n[ ÷0. so their sum tends to 0 too (Why?).
Thus, [r
a
+n
a
÷(r +n)[ also goes to 0.
d) Assume that r n and take c =
a÷&
2
. Let us consider the open
intervals: 1 = (n ÷ c. n + c) and J = (r ÷ c. r + c). Since r
a
÷ r and
n
a
÷n. for a large : one can …nd r
a
÷ J and n
a
÷ 1. But any element
of 1 is less than any element of J. Hence n
a
< r
a
and we obtain a
contradiction, because, for any :. one has in the hypothesis of d) that
r
a
_ n
a
.
i) Let us prove for instance that sin r
a
÷ sin r. whenever r
a
÷ r.
First of all we remark that [sin c[ = sin [c[ for any c ÷ (÷
¬
2
.
¬
2
). Since
r
a
÷ r. one can take : large enough such that r
a
÷ r ÷ (÷
¬
2
.
¬
2
). If
c is measured in radians and c ÷ (÷
¬
2
.
¬
2
) then, an easy geometrical
construction (see Fig.1.2) tell us that sin [c[ _ [c[ .
Let us use now some trigonometry:
[sin r
a
÷sin r[ = 2
sin
r
a
÷r
2
cos
r
a
+r
2
_ 2
r
a
÷r
2
= [r
a
÷r[ .
so [sin r
a
÷sin r[ ÷0. whenever r
a
÷r.
j) Since 1 and o are continuous (see the de…nition in the statement
of the theorem) then, r
a
÷r implies 1(r
a
) ÷1(r) (continuity of 1).
Since o is continuous, o(1(r
a
)) ÷o(1(r)). i.e. (o·1)(r
a
) ÷(o · 1) (r).
Thus o · 1 is also continuous. The other statements in j) are easy
consequences of some of the previous statements of the theorem (prove
them!).
O A
B
C
1
1
α
BC
=
sin α <
BA < lenght (arcBA) = α
Fig. 1.2
2. Sequences of complex numbers
Let C be the complex number …eld. Since any element . of C is a
pair . = (r. n) of two real numbers and since the element i = (0. 1) has
the property that i(n. 0) = (0. n) (see the multiplication rule de…ned in
(1.10)), we can write . = r+in. where we identify (r. 0) and (n. 0) with
2. SEQUENCES OF COMPLEX NUMBERS 27
r and n respectively. Let us …x a Cartesian coordinate system ¦(; i. j¦
in a plane (1). Here i and j are orthogonal versors and they give the
directions and the orientations of the (raxis and (naxis respectively.
Since any vector
÷÷÷
(`. where ` is an arbitrary point in the plane (1).
can be uniquely written as:
÷÷÷
(` = ri + nj. where r. n ÷ R, we call r
and n the coordinates of the point `. Write `(r. n). The association
. = r+in ÷÷`(r. n) give rise to a geometrical representation of the
complex number …eld C. This is way we always call C, the complex
plane. The distance d between two complex numbers .
1
= r
1
+in
1
and
.
2
= r
2
+in
2
is simply the distance between their corresponding points
`
1
(r
1
. n
1
) and `
2
(r
2
. n
2
) respectively, i.e.
d(.
1
. .
2
)
oc;
=
(r
2
÷r
1
)
2
+ (n
2
÷n
1
)
2
It is not di¢cult to check the three properties of a distance function
for this d.
A sequence ¦.
a
¦ of complex numbers is said to be convergent to .
if the numerical sequence of real numbers ¦d(.
a
÷.)¦ is convergent to
0. For instance, .
a
=
1
a
+ (1 +
1
a
)
a
i is convergent to ci because
d(.
a
. ci) =
(
1
:
÷0)
2
+ [(1 +
1
:
)
a
÷c]
2
÷0.
The sequence ¦.
a
¦ is said to be fundamental (or Cauchy) if for any
0. there is a natural number `
.
(depending of ) such that d(.
a+j
. .
a
) <
for any j = 1. 2. ... .
The following result reduces the study of the convergence of a se
quence .
a
= r
a
+ in
a
in C to the study of the convergence of the real
and imaginary part ¦r
a
¦ and ¦n
a
¦ respectively.
Theorem 15. Let ¦.
a
= r
a
+ n
a
i¦ be a sequence of complex num
bers (here r
a
and n
a
are real numbers). Then the sequence ¦.
a
¦ is
convergent to the complex number . = r+ni if and only if r
a
÷r and
n
a
÷n as sequences of real numbers.
Proof. One has the following double implications:
.
a
÷. =d(.
a
. .) =
(r
a
÷r)
2
+ (n
a
÷n)
2
÷0 =r
a
÷r ÷0
and n
a
÷ n ÷ 0 (simultaneously), i.e. if and only if r
a
÷ r and
n
a
÷n.
The sequence .
a
= 3 + (2:sin
1
a
)i tends to 3 + 2i because 3 ÷ 3
and 2:sin
1
a
= 2
sin
1
n
1
n
÷2.
28 1. THE REAL LINE.
Theorem 16. Relative to the distance d. the complex number …eld
C is complete, i.e. any Cauchy sequence ¦.
a
¦ of C is convergent to a
complex number ..
Proof. Let .
a
= r
a
+n
a
i. where r
a
and n
a
are real numbers. Since
¦.
a
¦ is a Cauchy sequence if and only if d(.
a+j
. .
a
) is as small as we
want when : is large enough, independent on j = 1. 2. ... and since
d(.
a+j
. .
a
) =
(r
a+j
÷r
a
)
2
+ (n
a+j
÷n
a
)
2
.
one sees that [r
a+j
÷r
a
[ and [n
a+j
÷n
a
[ are simultaneously small enough
whenever : is large enough, independent on j. But this is equivalent
to saying that ¦r
a
¦ and ¦n
a
¦ are both Cauchy sequences. Since R is
complete (see Theorem 13), ¦r
a
¦ is convergent to a real number r and
¦n
a
¦ is convergent to another real number n. Let us put . = r + ni.
Applying now Theorem 15 we get that .
a
is convergent to ..
We say that a subset ¹ of C is bounded if there is a su¢ciently
large ball 1(0. :) = ¦. ÷ C [ [.[ = d(0. .) < :¦. with centre at 0 and
of radius : 0. such that ¹ · 1(0. :). We also have for C a Bolzano
Weierstrass type theorem. Namely, any in…nite bounded sequence ¦.
a
¦
of complex numbers has a convergent subsequence. If we add a symbol
· to C with similar properties like the in…nite · for R, we get C =
C'¦·¦. the Riemann sphere. It is easy to see that in C any sequence
has a convergent subsequence. Because of this last property, we say
that C and R are the "compacti…cations" of C and of R respectively.
Generally, in a metric space (¹. d) a subset ` is said to be compact
if any sequence of ` has at least a convergent subsequence with its
limit in `. For instance, any closed interval [c. /] is a compact subset
of R (because of BolzanoWeierstrass Theorem). A subset ( of C is
said to be closed if for any sequence ¦.
a
¦ of elements in (. which is
convergent to . in C, its limit . is also in (. Then, the compact subsets
of C are exactly the closed and bounded subsets of C (have you any
idea to prove this?try a similar idea like that one from the real line
situation!)
3. Problems
1. Prove that the following subsets of R have the same cardinal:
a) ¹ = (0. 1) and 1 = R, b) ¹ = (0. 1] and 1 = R, c) ¹ = (÷·. c)
and 1 = R, d) ¹ = (0. 1) and 1 = (c. /). e) ¹ = (c. ·) and 1 = (0. 1].
f) ¹ = Q¨ [0. 3] and 1 = Q¨ [÷7. 3].
2. Prove that sup(¹+1) = sup ¹+sup 1. sup(¹1) = sup ¹sup 1.
where ¹ + 1 = ¦r + n [ r ÷ ¹. n ÷ 1¦ and ¹ 1 = ¦rn [ r ÷ ¹.
3. PROBLEMS 29
n ÷ 1¦. De…ne inf ¹ and prove the same equalities for inf instead of
sup .
3. Construct R = R ' ¦÷·. ·¦ and prove that any sequence
of elements in R has a convergent subsequence in R. Prove that if
a sequence ¦r
a
¦ is convergent in R. then it has only one limit point,
namely the limit of the sequence. Find the limit points for the sequence
c
a
= cos
a¬
3
. : = 0. 1. 2. ... . Recall that r ÷ ` is a limit point of a
subset ¹ of a metric space (`. d) if there is a nonconstant sequence
¦r
a
¦ of elements from ¹. which is convergent to r.
4. Prove that if
o
n+1
on
÷ . where c
a
0 for any :. then
n
c
a
÷ .
Apply this result to compute the limit: lim
n
(2a)!
1·3·5·...·(4a+1)
. whenever
: ÷·.
5. Prove that the set R ` Q of irrational numbers is not countable.
Prove that it has the same cardinal as the cardinal of R (i.e. there is a
bijection between R ` Q and R).
6. Prove that the length of the diagonal of a square which has the
side a rational number, is not a rational number.
7. Are
3
5 and
7
3 rational numbers? Are they algebraic numbers?
8. Prove that the metric space ([0. 1). d). where d(r. n) = [r ÷n[ .
is not a complete metric space, i.e. there is at least a Cauchy sequence
¦r
a
¦. r
a
÷ [0. 1). which has no limit in [0. 1). Prove that this limit must
be 1.
9. De…ne the notion of "boundedness" in a general metric space.
Is Cesaro’s Lemma (any in…nite bounded sequence has at least a con
vergent subsequence) true in a general metric space? Find a simple
counterexample.
10. Why a decreasing sequence always has a limit in R? If instead
of R you put Q = Q' ¦÷·. ·¦. is the last statement also true?
11. Prove that the Archimedes’ Axiom is equivalent to the fact that
lim
a÷o
1
a
= 0. If instead of this last limit we put lim
a÷o
2a+3
3a÷2
=
2
3
. does our
statement work too?
CHAPTER 2
Series of numbers
1. Series of nonnegative real numbers
We know to add a …nite number of real numbers c
1
. c
2
. .... c
a
:
:
a
= (... ((c
1
+c
2
) +c
3
) +...) +c
a÷1
) +c
a
)
For instance,
:
4
= 7 + 3 + (÷4) + 5 = 10 + (÷4) + 5 = 6 + 5 = 11.
However, we have just met in…nite sums when we discussed about
the representation of a real number as a decimal fraction. For instance,
: = 3.3444... = 3.3(4) = 3 +
3
10
+
4
10
2
+
4
10
3
+... =
= lim
a÷o
(3 +
3
10
+
4
10
2
+
4
10
3
+... +
4
10
a
) =
=
33
10
+
4
10
2
lim
a÷o
1 ÷
1
10
n1
1 ÷
1
10
=
301
90
.
Since such in…nite sums (called series) appear in many applications
of Mathematics, we start here a systematic study of them.
Definition 6. Let ¦c
a
¦ be a sequence of real numbers. The in…nite
sum
(1.1)
o
¸
a=0
c
a
= c
0
+c
1
+... +c
a
+...
is by de…nition the value (if this one exists) of the limit : = lim
a÷o
:
a
.
where :
a
= c
0
+c
1
+...+c
a
is called the partial sum of order :. The new
mathematical object de…ned in (1.1) is said to be the series of general
term c
a
and of sum : (if the limit exists). If : exists we say that the
series (1.1) is convergent. If the limit does not exist we say that the
series (1.1) is divergent.
31
32 2. SERIES OF NUMBERS
For instance, the series
o
¸
a=0
1
2
a
= lim
a÷o
(1 +
1
2
+
1
2
2
+... +
1
2
a
) = 2
is convergent to 2. or its sum is 2. whereas the series
o
¸
a=0
: = ·. or
o
¸
a=0
(÷1)
a
are divergent. The last divergent series is said to be oscillatory
because its partial sums have the values 0 or 1. i.e. it oscillates between
the distinct values ¦0. 1¦.
Theorem 17. Let r be a real number. The geometrical series
o
¸
a=0
r
a
is convergent (and its sum is
1
1÷a
) if and only if [r[ is less then 1.
Proof. By De…nition 6,
o
¸
a=0
r
a
= lim
a÷o
(1 +r +r
2
+... +r
a
) = lim
a÷o
1 ÷r
a+1
1 ÷r
.
Since lim
a÷o
r
a+1
exists and is …nite if and only if [r[ < 1 (when the limit
is 0), the series
o
¸
a=0
r
a
is convergent if and only if [r[ < 1. In this last
case, its sum is : = lim
a÷o
1÷a
n+1
1÷a
=
1
1÷a
.
Theorem 18. (The Cauchy general test) A series
o
¸
a=0
c
a
is con
vergent if and only if the sequence of partial sums ¦:
a
¦ is a Cauchy
sequence.
Proof. We only use the fact that R is complete, i.e. that the
sequence ¦:
a
¦ is convergent if and only if it is a Cauchy sequence.
Corollary 1. (The zero test) If the sequence ¦c
a
¦ does not tend
to zero, then the series
o
¸
a=0
c
a
is divergent. Or, if the series
o
¸
a=0
c
a
is
convergent, then c
a
÷0.
Proof. If the series
o
¸
a=0
c
a
was convergent, then the sequence of
partial sums ¦:
a
¦ would be a Cauchy sequence (see Theorem18). Thus,
for : large enough, c
a
= :
a
÷ :
a÷1
becomes smaller and smaller, i.e.
c
a
÷ 0. In fact, we do not need the previous theorem. Indeed, let
: =
o
¸
a=0
c
a
and write c
a
= :
a
÷:
a÷1
. Then, limc
a
= : ÷: = 0.
1. SERIES OF NONNEGATIVE REAL NUMBERS 33
For instance,
o
¸
a=0
a+1
a
a
is divergent, because c
a
=
a+1
a
a
÷c = 0.
Theorem 19. (The renouncement test) Let us consider the series:
: =
o
¸
a=0
c
a
and :
t
=
o
¸
a=.
c
a
= c
.
+c
.+1
+... (we just got out the terms
c
0
. c
1
. .... c
.÷1
in the previous series). Then these two series have the
same nature (i.e. they are convergent or divergent) at the same time.
Moreover, if they are convergent, then : = :
t
+c
0
+c
1
+... +c
.÷1
.
Proof. Let : be large enough (: _ `) and let :
a
= c
0
+c
1
+... +
c
.÷1
+c
.
+...+c
a
. If we denote :
t
a
= c
.
+...+c
a
. then :
t
a
is the partial
sum of order : of the series :
t
. It is clear that :
a
= :
t
a
+c
0
+c
1
+...+c
.÷1
and that the sequences ¦:
a
¦ and ¦:
t
a
¦ are convergent or divergent at the
same time (prove it!). Now, in the last equality, let us make : ÷ ·.
We get: : = :
t
+c
0
+c
1
+... +c
.÷1
and the proof is completed.
Let
o
¸
a=0
c
a
be a series with
c
a
= :. if : _ 100 and c
a
=
1
3
a
. if : 100.
The question is:"What is the nature of this series?" So we must decide if
our series is convergent or not. Let us renounce the terms c
0
. c
1
. .... c
100
in the initial series. We get a new series
o
¸
a=101
1
3
a
=
1
3
101
(1 +
1
3
+
1
3
2
+...).
Let us use now Theorem 17 and …nd that
o
¸
a=0
c
a
= 0 + 1 +... + 100 +
1
3
101
1
1 ÷
1
3
=
100 101
2
+
1
2 3
100
.
Theorem 20. (The boundedness test) Let : =
o
¸
a=0
c
a
be a series
with nonnegative terms (c
a
_ 0). Then the series : is convergent if
and only if the partial sums sequence ¦:
a
¦. :
a
= c
0
+ c
1
+ ... + c
a
. is
bounded.
Proof. Let us assume that the series : =
o
¸
a=0
c
a
is convergent,
i.e. the sequence ¦:
a
¦ is convergent. Since any convergent sequence is
bounded (see also Theorem 10), one has that ¦:
a
¦ is bounded.
Conversely, we suppose that ¦:
a
¦ is bounded. Since c
a
_ 0. :
a
_
:
a+1
. i.e. the sequence ¦:
a
¦ is increasing. But Theorem 8 says that an
increasing and bounded sequence ¦:
a
¦ is convergent to its superior limit
34 2. SERIES OF NUMBERS
limsup :
a
. Thus the series : =
o
¸
a=0
c
a
is convergent to this limsup :
a
.
i.e. : = limsup :
a
.
Theorem 21. (The integral test) Let c be a …xed real number and let
1 : [c. ·) ÷[0. ·) be a decreasing continuous function (see De…nition
5). Let :
0
be a natural number greater or equal to c. For any : _ :
0
let c
a
= 1(:) and let ¹
a
=
a
a
0
1(r)dr for : _ :
0
. Then the series : =
o
¸
a=a
0
c
a
is convergent if and only if the sequence ¦¹
a
¦ is convergent (it
is su¢cient to be boundedwhy?).
Proof. Suppose that the series : =
o
¸
a=a
0
c
a
=
o
¸
a=a
0
1(:) is con
vergent. Since in Fig.2.1 :
a
= 1(:
0
) + ... + 1(:) is exactly the sum
of the hatched and of the double hatched areas and since the integral
¹
a
=
a
a
0
1(r) dr is equal to the area under the graphic of n = 1(r)
which corresponds to the interval [:
0
. :]. then ¹
a
_ :
a
. Since : =
o
¸
a=a
0
c
a
is convergent, the sequence ¦:
a
¦ is bounded, thus the sequence
¦¹
a
¦ is bounded.
Conversely, let us assume that the sequence ¦¹
a
¦ is bounded. Look
again at Fig.2.1! We see that the double hatched area is just equal to
c
ao+1
+c
a
0
+2
+... +c
a+1
= :
a+1
÷c
a
0
. Since this double hatched area
is less then the area ¹
a+1
=
a+1
a
0
1(r) dr. one has that the sequence
¦:
a+1
÷ c
a
0
¦ is bounded. Hence the sequence ¦:
a
¦ is also bounded
(why?). Now, the Theorem 20 tells us that the series : =
o
¸
a=a
0
c
a
is
convergent.
Why we say that if lim
a÷o
1(r) = 0. then the above series is divergent?
O 1 2 c n
0
n
0
+1 n
0
+2 ................. n1 n n+1 x
y = f(x)
y
Fig. 2.1
1. SERIES OF NONNEGATIVE REAL NUMBERS 35
The integral test is very useful in practice. Suppose that somebody
is interested in the nature of the series
o
¸
a=2
1
aln(a)
. Let us apply the
integral test and consider the associated decreasing continuous function
1 : [2. ·) ÷[0. ·). 1(r) =
1
r ln r
(we simply put r instead of : in c
a
=
1
aln(a)
for : _ 2). Since
¹
a
=
a
2
1
r ln r
dr = ln(ln(r))[
a
2
= ln(ln :) ÷ln(ln(2)) ÷·.
¹
a
is unbounded, thus our series is divergent (see Theorem 21).
In the last 150 years one of the most interesting function in Mathe
matics, which was highly considered, is the Zeta function of Riemann.
"Zeta" comes from the Greek letter .. The notation of this function
was …rstly used by the great German mathematician B. Riemann. Its
analytic expression is:
(1.2) 2(c) =
o
¸
a=1
1
:
c
. c ÷ R
This famous function is usually de…ned by a series. Thus, the maximal
domain of de…nition for this function is exactly the set of all c ÷ R
with the property that the numerical series
o
¸
a=1
1
a
is convergent. We
call this last set, the convergence set of our series. In the following,
using the integral test, we …nd the convergence set for the Riemann
(zeta) series
o
¸
a=1
1
a
.
Theorem 22. (Riemann zeta series) The Riemann zeta series is
convergent if and only if c 1. This means that the real de…nition
domain of the function 2 is the interval (1. ·).
Proof. Let us take in Theorem 21 1(r) =
1
a
for r _ 1. Since
¹
a
=
a
1
1
r
c
dr =
1
1 ÷c
[:
÷c+1
÷1] if c = 1
and ¹
a
= ln :. if c = 1. then ¹
a
is bounded if and only if c 1(why?).
Now, Theorem 21 says that the Riemann series
o
¸
a=1
1
a
is convergent if
and only if c 1.
36 2. SERIES OF NUMBERS
The sum
: = 1 +
1
2
+
1
3
+... =
o
¸
a=1
1
:
= 2(1) = ·.
because the series
o
¸
a=1
1
a
is divergent for c = 1. thus the sequence of
partial sums
:
a
= 1 +
1
2
+
1
3
+... +
1
:
is strictly increasing and unbounded. Hence : = lim:
a
= ·. The
Theorem 22 says that the series
2(2) = 1 +
1
2
2
+
1
3
2
+...
is convergent. So it can be approximated by
:
.
= 1 +
1
2
2
+
1
3
2
+... +
1
`
2
for ` large enough. We call the series
o
¸
a=1
1
a
the harmonic series. It is
very important in Analysis. Sometimes the following test is useful.
Theorem 23. (The Cauchy’s compression test) Let ¦c
a
¦ be a de
creasing sequence of nonnegative real numbers. Then the series : =
o
¸
a=0
c
a
and o =
o
¸
a=0
2
a
c
2
n have one and the same nature, i.e. they are
simultaneous convergent or divergent.
Proof. Let :
I
=
I
¸
a=0
c
a
and o
n
=
n
¸
a=0
2
a
c
2
n be the /th and the
:th partial sums of : and o respectively.
Let us …x / and let us take a : such that / _ 2
n
÷1. Then,
:
I
= c
0
+c
1
+... +c
I
_ c
0
+c
1
+... +c
2
m
÷1
= c
0
+c
1
+ (c
2
+c
3
)+
+(c
4
+c
5
+c
6
+c
7
) +... +(c
2
m1 +c
2
m1
+1
+c
2
m1
+2
+... +c
2
m
÷1
) _
_ c
0
+c
1
+ 2c
2
+ 2
2
c
2
2 +... + 2
n÷1
c
2
m1 = c
0
+o
n÷1
.
So
(1.3) :
I
_ c
0
+o
n÷1
Now, if the series o =
o
¸
a=0
2
a
c
2
n is convergent, then the increasing se
quence ¦o
n
¦ is bounded. The inequality (1.3) says that the sequence
1. SERIES OF NONNEGATIVE REAL NUMBERS 37
¦:
I
¦ is also bounded, thus the series : =
o
¸
a=0
c
a
is convergent (see
Theorem 20). If : =
o
¸
a=0
c
a
is divergent, then the sequence ¦:
I
¦ is un
bounded. From (1.3) we see that the sequence ¦o
n
¦ is also unbounded,
so the series o =
o
¸
a=0
2
a
c
2
n is divergent.
Assume now that : is …xed and let us take / such that / _ 2
n
.
Then
:
I
= c
0
+c
1
+... +c
I
_ c
0
+c
1
+... +c
2
m =
= c
0
+c
1
+c
2
+ (c
3
+c
4
) + (c
5
+c
6
+c
7
+c
8
)+
...+(c
2
m1 +c
2
m1
+1
+...+c
2
m) _ c
0
+
1
2
c
1
+c
2
+2c
4
+2
2
c
8
+...+2
n÷1
c
2
m
_
1
2
(c
1
+ 2c
2
+ 2
2
c
2
2 +... + 2
n
c
2
m) =
1
2
o
n
.
thus,
(1.4) :
I
_
1
2
o
n
If the series : =
o
¸
a=0
c
a
is convergent, then the sequence ¦:
I
¦ is bounded
and, using (1.4), we get that the sequence ¦o
n
¦ is also bounded (why?).
Hence, the series o =
o
¸
a=0
2
a
c
2
n is convergent (why?). If o =
o
¸
a=0
2
a
c
2
n
is divergent, then the sequence ¦o
n
¦ tends to ·(why?), so, from (1.4),
we get that the sequence ¦:
I
¦ also goes to · and thus, the series : =
o
¸
a=0
c
a
is also divergent. Now the theorem is completely proved.
We can use this test to …nd again the result on the Riemann zeta
function 2(c) =
o
¸
a=0
1
a
(see Theorem 22). Indeed, here c
a
=
1
a
and
c
2
n =
1
2
n
=
1
2
a
. The series
o
¸
a=0
2
a
1
2
c
a
=
o
¸
a=0
1
2
c÷1
a
is obviously convergent if and only if c 1 (see Theorem 17). Thus,
from the Cauchy compression test, we get that the Riemann series is
convergent if and only if c 1.
38 2. SERIES OF NUMBERS
Now, let us …nd all the values of c ÷ R such that the series
o
¸
a=2
1
a(log
7
a)
is convergent. If in
1
a(log
7
a)
we put instead of :. 2
a
and
if we multiply the result by 2
a
. we get the series
o
¸
a=2
2
a
1
2
a
(log
7
2
a
)
c
=
1
(log
7
2)
c
o
¸
a=2
1
:
c
.
Thus, the nature of our series is the same like the nature of the Riemann
series. Therefore, our series is convergent if and only if c 1.
Another useful convergence test is the following:
Theorem 24. (The comparison test) Let : =
o
¸
a=0
c
a
and n =
o
¸
a=0
/
a
be two series with c
a
_ 0. /
a
_ 0 and c
a
_ /
a
for : = 0. 1. 2. ... . a)
If the series n =
o
¸
a=0
/
a
is convergent, then the series : =
o
¸
a=0
c
a
is
also convergent. b) If the series : =
o
¸
a=0
c
a
is divergent, then the series
n =
o
¸
a=0
/
a
is also divergent.
Proof. Since c
a
_ /
a
for : = 0. 1. 2. .... then
:
a
= c
0
+c
1
+... +c
a
_ /
0
+/
1
+... +/
a
oc;
= n
a
.
the partial :th sum of the series n. a) If the series n =
o
¸
a=0
/
a
is
convergent, the sequence ¦n
a
¦ is bounded. Hence the sequence ¦:
a
¦ is
also bounded, and so the series : =
o
¸
a=0
c
a
is convergent (see Theorem
20). b) If the series : =
o
¸
a=0
c
a
is divergent, then the sequence ¦:
a
¦ is
unbounded (see Theorem 20). Hence the sequence ¦n
a
¦ is unbounded
(why?), so the series n =
o
¸
a=0
/
a
is divergent.
For instance, the series
o
¸
a=0
1
a
2
+7
is convergent because
1
a
2
+7
<
1
a
2
and because the series
o
¸
a=0
1
a
2
= 2(2) is convergent (see Theorem 22).
The comparison test is also useful in proving the following basic
convergence test (see Theorem 25).
1. SERIES OF NONNEGATIVE REAL NUMBERS 39
First of all we remark that the natural way to add two series is the
following
(1.5)
o
¸
a=0
c
a
+
o
¸
a=0
/
a
=
o
¸
a=0
(c
a
+/
a
).
It is easy to see that if the both series are convergent, then the
resulting series on the right is also convergent (prove it!). If c
a
. /
a
are
nonnegative then, if at least one series is divergent, the series on the
right in (1.5) is also divergent (prove it!). In general this is not true.
For instance,
o
¸
a=0
: +
o
¸
a=0
(÷:) = 0!
Now, if \ is a real number, by de…nition,
\
o
¸
a=0
c
a
=
o
¸
a=0
\c
a
If \ = ÷1. we can de…ne the subtraction:
o
¸
a=0
c
a
÷
o
¸
a=0
/
a
=
o
¸
a=0
c
a
+
o
¸
a=0
(÷/
a
).
For \ = 0. the series
o
¸
a=0
c
a
and \
o
¸
a=0
c
a
have the same nature (prove
it!). Pay attention to the following wrong calculation:
o
¸
a=2
1
: + 1
÷
o
¸
a=2
1
: ÷1
= ÷2
o
¸
a=0
1
:
2
÷1
The series on the right side is convergent, but on the left side we have
·÷·. an undetermined operation, so it cannot be equal to a deter
mined one!
Theorem 25. (The limit comparison test) Let
o
¸
a=0
c
a
and
o
¸
a=0
/
a
be two numerical series of real numbers such that c
a
_ 0 and /
a
0
for any : = 0. 1. 2. .... Suppose that the sequence
on
bn
¸
is convergent
to  ÷ R ' ¦·¦. Then, a) if  = 0. ·. both series have the same
nature (they are convergent or not) at the same time, b) if  = 0.
o
¸
a=0
/
a
convergent implies
o
¸
a=0
c
a
convergent and, c) if  = ·.
o
¸
a=0
/
a
divergent
implies
o
¸
a=0
c
a
divergent. This is why the series
o
¸
a=0
/
a
is called a witness
series.
40 2. SERIES OF NUMBERS
Proof. a) Since  = 0. ·.  0. so there is an 0 such that
 ÷ 0. Since lim
a÷o
on
bn
= . there is a natural number ` (depending
on ) with  ÷ <
on
bn
<  + for any : _ `. Because of the last double
inequality and since /
a
0. one can write
(1.6) ( ÷)/
a
< c
a
< ( +)/
a
.
for any : _ `. Now, if for instance,
o
¸
a=0
c
a
is convergent (this means
that the series
o
¸
a=.
c
a
is also convergent from Theorem 19) then, using
the inequality ( ÷ )/
a
< c
a
and the comparison test (Theorem 24)
we get that the series ( ÷ )
o
¸
a=.
/
a
is convergent. Since  ÷ = 0
we …nally obtain that the series
o
¸
a=.
/
a
is convergent, i.e. the series
o
¸
a=0
/
a
is convergent (see the renouncement test). If this last series is
convergent, using the second inequality, c
a
< ( +)/
a
. from (1.6), one
gets that the …rst series
o
¸
a=0
c
a
is convergent (complete the reasoning!).
b) If  = 0. take an 0 and take a natural number `
1
(depending
on ) such that for any : _ `
1
we have 0 _
on
bn
< or c
a
< /
a
. If the
series
o
¸
a=0
/
a
is convergent, then the series
o
¸
a=.
1
/
a
is also convergent, so
the series
o
¸
a=.
1
c
a
is convergent (see the comparison test). Using again
the renouncement test we get that the series
o
¸
a=0
c
a
is convergent. c)
If  = ·. take a positive real number ` 0 and take a natural
number `
2
(depending on `) such that for : _ `
2
.
on
bn
`. or
c
a
`/
a
. Now, if the series
o
¸
a=0
/
a
is divergent, then the series
o
¸
a=.
2
/
a
is also divergent (see Theorem 19). Use the inequality c
a
`/
a
to
obtain that the series
o
¸
a=.
2
c
a
is divergent (see the comparison test).
Using again the renouncement test we get that the series
o
¸
a=0
c
a
is
divergent.
1. SERIES OF NONNEGATIVE REAL NUMBERS 41
Let us decide if the series
o
¸
a=0
3
a
a
2
+4
is convergent or not. We intend
to use the limit comparison test with c
a
=
3
a
a
2
+4
and /
a
=
1
a
. We try
to …nd an c such that the limit  = lim
a÷o
on
bn
be …nite and nonzero. If we
can do this, such an c is unique. Its value is called the "Abel degree"
of the function 1(r) =
3
a
a
2
+4
. So,
 = lim
a÷o
c
a
/
a
= lim
a÷o
:
c+
1
3
:
2
(1 +
4
a
2
)
= 0. ·
(= 1) if and only if c+
1
3
= 2. i.e.
5
3
1. Since the series
o
¸
a=1
1
a
5
3
= 2(
5
3
)
is convergent (see the Riemann Zeta series), from the limit comparison
test one has that the series
o
¸
a=1
3
a
a
2
+4
is convergent. Applying again the
renouncement test we get that our initial series
o
¸
a=0
3
a
a
2
+4
is convergent.
Let us put in a systematic manner all the reasonings in this last
example.
Theorem 26. (The ccomparison test) Let
o
¸
a=0
c
a
be a series with
nonnegative terms (c
a
_ 0). We assume that there is a real number c.
such that the following limit does exist: lim
a÷o
:
c
c
a
=  ÷ R'¦·¦. a) If
 = 0. · then, the series
o
¸
a=0
c
a
is convergent if and only if c 1. b)
If  = 0 and c 1. then our series
o
¸
a=0
c
a
is convergent. c) If  = ·
and c _ 1. then the series
o
¸
a=0
c
a
is divergent and equal to ·.
Proof. It is enough to take /
a
=
1
a
in the Theorem 25 (do every
thing slowly, step by step!).
Let us apply this last test to the following situation. For a large `
( 100. for instance), can we use the approximation
o
¸
a=0
:
3
+ 7: + 1
:
9
+ 2: + 2

.
¸
a=0
:
3
+ 7: + 1
:
9
+ 2: + 2
?.
42 2. SERIES OF NUMBERS
We can do this if and only if our series is convergent (why?). In order
to see if our series is convergent or not, let us consider the limit:
lim
a÷o
:
c
:
3
+ 7: + 1
:
9
+ 2: + 2
= lim
a÷o
:
c+3
(1 +
7
a
2
+
1
a
3
)
:
9
2
1 +
2
a
8
+
2
a
9
= lim
a÷o
:
c+3
:
9
2
.
But, this last limit is neither 0 nor ·. if and only if c + 3 =
9
2
. or
c =
3
2
(why?). Since in this case c 1 and the limit  is 1. we apply
the ccomparison test (Theorem 26) and …nd that our initial series is
convergent. Hence the above approximation works!
A very useful test is the ratio test or D’Alembert test.
Theorem 27. (the ratio test) Let
o
¸
a=0
c
a
be a series with positive
terms.
a) If there is a real number \ such that 0 < \ < 1 and
o
n+1
on
_ \
for any : _ `. where ` is a …xed natural number, then the series is
convergent. This is equivalent to say that limsup
o
n+1
on
< 1.
b) If
o
n+1
on
_ 1 for any : _ `. where ` is a …xed natural number,
then the series is divergent.
c) If limsup
o
n+1
on
= 1. and if
o
n+1
on
is not equal to 1 from a rank on,
then, in general, we cannot decide if the series is convergent or not (in
this situation use more powerful tests, for instance the "RaabeDuhamel
Test").
Proof. a) Let us put : = `. ` + 1. ` + 2. ... in the inequality
o
n+1
on
_ \. We …nd:
c
.+1
_ \c
.
. c
.+2
_ \c
.+1
_ \
2
c
.
. .... c
.+n
_ \
n
c
.
. ....
Hence,
c
.
+c
.+1
+c
.+2
+... +c
.+n
+... _
_ c
.
(1 +\ +\
2
+... +\
n
+...) = c
.
1
1 ÷\
.
So any partial sum of the series
o
¸
a=.
c
a
is bounded. Since c
a
_ 0.
the series
o
¸
a=.
c
a
is convergent (Theorem 20). The renouncement test
says that the whole series
o
¸
a=0
c
a
is also convergent.
b) If
o
n+1
on
_ 1 for any : _ `. then
c
A
+c
A+1
+... +c
A+n
+... _ c
A
+c
A
+... +c
A
+... = ·.
1. SERIES OF NONNEGATIVE REAL NUMBERS 43
so the series
o
¸
a=0
c
a
is divergent (explain everything slowly, step by
step!).
c) For instance, the harmonic series
o
¸
a=1
1
a
is divergent, but limsup
a÷o
1
n+1
1
n
=
1. This last property is also true for the series
o
¸
a=1
1
a
2
. but this last series
is convergent! This is why we cannot say anything in general if one can
…nd numbers of the form
o
n+1
on
< 1 as close as we want to 1.
Remark 5. The condition from a) of Theorem 27 is equivalent to
saying that limsup
o
n+1
on
< 1 (why?). If the sequence
o
n+1
on
¸
is conver
gent to . then the Theorem 27 is more exactly. Namely, in this last
case, the series
o
¸
a=0
c
a
is convergent if  < 1. it is divergent if  1 and
if  = 1 we cannot say anything (prove it!).
For instance, the series
o
¸
a=0
1
2
n
a!
is convergent because lim
a÷o
o
n+1
on
=
0 < 1 (see Remark 5).
Usually, if lim
a÷o
o
n+1
on
= 1. we try to apply the following "more pow
erful" test.
Theorem 28. (The RaabeDuhamel test) Let
o
¸
a=0
c
a
be a series with
positive terms.
a) If there is a real number \ ÷ (1. ·) and a natural number ` such
that :
on
o
n+1
÷1
_ \ for any : _ `. then the series is convergent.
b) If :
on
o
n+1
÷1
< 1 for : _ `. where ` is a …xed natural
number, then the series is divergent.
c) Assume that the following limit exists, lim
a÷o
:
on
o
n+1
÷1
=  ÷
R'¦·¦. Then, if  1. the series is convergent, if  < 1. the series is
divergent and if  = 1. we cannot decide on the nature of this series.
One can …nd a proof of this result in [Nik], or in [Pal]. See also
Problem 11 of this chapter.
Let us …nd the nature of the series
o
¸
a=1
1 3 5 ... (2: + 1)
2 4 6 ... 2:
1
2: + 3
.
Since
c
a+1
c
a
=
(2: + 3)
2
(2: + 2)(2: + 5)
÷1.
44 2. SERIES OF NUMBERS
let us apply RaabeDuhamel test. Since
:
c
a
c
a+1
÷1
=
2:
2
+:
(2: + 3)
2
÷
1
2
< 1.
the series is divergent.
Theorem 29. (The Cauchy root test) Let
o
¸
a=0
c
a
be a series with
nonnegative terms.
a) If there is a real number \ ÷ (0. 1) such that
n
c
a
_ \ for : _ `.
where ` is a …xed natural number, then the series is convergent.
b) If
n
c
a
_ 1 for all : _ `. where ` is a …xed natural number,
then the series is divergent.
c) Assume that the following limit exists, lim
a÷o
n
c
a
=  ÷ R '
¦·¦.Then, if  < 1. the series is convergent, if  1. the series is
divergent and if  = 1. we cannot decide on the nature of this series.
Proof. a) The condition
n
c
a
_ \ for : _ ` implies
c
.
+c
.+1
+... +c
.+n
+... _ c
.
\
.
(1 +\ +... +\
n
+...) =
= c
.
\
.
1 ÷\
<
c
.
1 ÷\
.
so, the partial sums of the series
o
¸
a=.
c
a
are bounded. Hence the
series
o
¸
a=.
c
a
is convergent (see Theorem 20). From the renouncement
test we derive that the series
o
¸
a=0
c
a
is convergent.
b) The condition
n
c
a
_ 1 for : _ `. implies c
a
_ 1 for an in…nite
number of terms, so ¦c
a
¦ does not tend to zero. Hence the series is
divergent (see Corollary 1).
c) Take 0 such that  + < 1. Since
n
c
a
÷. there is a natural
number ` such that if : _ `.
n
c
a
<  + . Apply now a) and …nd
that the series is convergent. If  1. there is a rank ` from which
on
n
c
a
_ 1 for : _ ` and so, the series is divergent (see b)). If
 = 1. there are some cases in which the series is convergent and there
are other cases in which the series is divergent. For instance, the series
o
¸
a=1
1
a
2
is convergent and  = lim
a÷o
n
1
a
2
= 1 (since
n
: ÷ 1; prove this!
Hint:
c
a
=
n
: ÷1 ==: = (1 +c
a
)
a
= 1 +:c
a
+
:(: ÷1)
2
c
2
a
+...
2. SERIES WITH ARBITRARY TERMS 45
:(: ÷1)
2
c
2
a
==c
a
<
2
: ÷1
.
so, c
a
÷0. But the series
o
¸
a=1
1
a
is divergent and  = lim
a÷o
n
1
a
= 1.
The series
o
¸
a=0
1
(2+a)
n
is convergent because
n
c
a
=
1
2+a
_
1
2
for any
: = 0. 1. ... (we just applied the Cauchy Root Test, a)). We can also
apply the Comparison Test:
1
(2+a)
n
<
1
a
2
for any : = 1. 2. ... , etc.
Remark 6. A natural question arises: what is the connection (if
there is one!) between the ratio test and the root test? To explain
this we need a powerful result from the calculus of the limits of se
quences. This is the famous CesaroStolz Theorem: Let ¦c
a
¦ be an ar
bitrary sequence and let ¦/
a
¦ be an increasing and unbounded sequence
of positive numbers such that the sequence
o
n+1
÷on
b
n+1
÷bn
¸
is convergent to
 ÷ R = R'¦÷·. ·¦. Then
on
bn
÷. A direct consequence of this result
is the Cesaro Theorem: Let ¦c
a
¦ be a convergent to  sequence. Then
the "means" sequence
¸
c
0
+c
1
+...c
n1
a
is also convergent to  (prove it as
an application of the CesaroStolz Theorem). We prove now that for a
sequence ¦c
a
¦ of positive numbers, such that the limit of the sequence
o
n+1
on
¸
does exist in R. then
o
n+1
on
¸
÷ if and only if ¦
n
c
a
¦ ÷. Sup
pose that
o
n+1
on
¸
÷. then ln c
a+1
÷ln c
a
÷ln . or
ln o
n+1
÷ln on
(a+1)÷a
÷ln .
From the CesaroStolz Theorem we get that
ln on
a
= ln
n
c
a
÷ ln . or
n
c
a
÷ . Conversely, assume that ¦
n
c
a
¦ ÷  and that
o
n+1
on
¸
÷ 
t
.
From the …rst implication, one has that  = 
t
and the statement is
completely proved.
Suppose we have a series
o
¸
a=0
c
a
with c
a
0 for any : `. such
that
o
n+1
on
¸
÷ 1. We cannot decide on the nature of this series. Re
mark 6 says that it is not a good idea to try to apply the Cauchy Root
Test because this one also cannot decide if the series is convergent or
not.
2. Series with arbitrary terms
Up to now we just considered (in principal) series with nonnegative
terms. If the number of positive or negative terms in a series are …nite,
to decide the nature of this series, it is su¢cient to get out those terms
46 2. SERIES OF NUMBERS
and thus to obtain a new series with all its term positive or negative
(see the renouncement test). If c
a
_ 0 in a series
o
¸
a=0
c
a
. we consider
the new series
o
¸
a=0
(÷c
a
) = ÷
o
¸
a=0
c
a
and apply the results obtained in
the previous section. For instance,
o
¸
a=0
÷
1
a
3
= ÷
o
¸
a=0
1
a
3
is convergent,
because
o
¸
a=0
1
a
3
is convergent (it is the value of the Riemann series for
c = 3 1). A numerical series
o
¸
a=0
c
a
is said to have arbitrary terms if
the sign of its terms c
a
may be positive, negative or zero, but not all
(or a …nite number of them) are of the same sign. We also call such a
series a general series. The Cauchy general test (see Theorem 18) and
the zero test are the only tests we know (up to now) on general series.
Here is another important one.
Theorem 30. (The AbelDirichlet test) Let ¦c
a
¦ be a decreasing
to zero (c
a
÷ 0) sequence of nonnegative (c
a
_ 0) real numbers. Let
o
¸
a=0
/
a
be a series with bounded partial sums (i.e. there is a real number
` 0 such that for :
a
= /
0
+ /
1
+ ... + /
a
. one has [:
a
[ < `. where
: = 0. 1. ...). Then the series
o
¸
a=0
c
a
/
a
is convergent.
Proof. We intend to apply the Cauchy general test (Theorem 18).
Let us denote o
a
= c
0
/
0
+c
1
/
1
+... +c
a
/
a
the :th partial sum of the
series
o
¸
a=0
c
a
/
a
and let us evaluate
[o
a+j
÷o
a
[ = [c
a+1
/
a+1
+... +c
a+j
/
a+j
[ =
= [c
a+1
(:
a+1
÷:
a
) +c
a+2
(:
a+2
÷:
a+1
) +... +c
a+j
(:
a+j
÷:
a+j÷1
)[ =
[÷c
a+1
:
a
+ (c
a+1
÷c
a+2
):
a+1
+... + (c
a+j÷1
÷c
a+j
):
a+j÷1
+c
a+j
:
a+j
[
(2.1)
_ c
a+1
[:
a
[+(c
a+1
÷c
a+2
) [:
a+1
[+...+(c
a+j÷1
÷c
a+j
) [:
a+j÷1
[+c
a+j
[:
a+j
[ .
Let 0 be a small positive real number. In the last row of (2.1) we
put instead [:
;
[ . , = :. : + 1. .... : + j. the greater number `. So we
get
(2.2)
[o
a+j
÷o
a
[ _ `(c
a+1
+c
a+1
÷c
a+2
+c
a+2
÷c
a+3
+...+c
a+j÷1
÷c
a+j
+c
a+j
)
2. SERIES WITH ARBITRARY TERMS 47
= 2`c
a+1
Since ¦c
a
¦ tends to 0 as : ÷·. there is a natural number ` (which
depend on ) such that for any : _ `. on has that 2`c
a+1
< . Since
[o
a+j
÷o
a
[ _ 2`c
a+1
(see (2.2)), we get that [o
a+j
÷o
a
[ < for any
: _ `. This means that the sequence ¦o
a
¦ is a Cauchy sequence, i.e.
the series
o
¸
a=0
c
a
/
a
is convergent (see Theorem 18) and our theorem is
completely proved.
A famous example is the standard alternate series
(2.3)
o
¸
a=1
(÷1)
a÷1
1
:
= 1 ÷
1
2
+
1
3
÷
1
4
+....
This series is a general series (why?) and it is convergent. Indeed,
¸
c
a
=
1
a
is a decreasing to zero sequence with nonnegative terms and
the series
o
¸
a=1
(÷1)
a÷1
has bounded partial sums: :
1
= 1. :
2
= 0. :
3
=
1. .... i.e. [:
a
[ _ 1. We apply now the AbelDirichlet test.
The following test is a direct consequence of the same AbelDirichlet
test.
Corollary 2. (The Leibniz test) Let ¦c
a
¦ be a decreasing to zero
(c
a
÷ 0) sequence of nonnegative (c
a
_ 0) real numbers. Then the
series
o
¸
a=1
(÷1)
a÷1
c
a
= c
1
÷c
2
+c
3
÷...
is convergent.
For instance, applying this test, we get that the series
o
¸
a=1
(÷1)
a a+1
a
2
+3
=
÷
o
¸
a=1
(÷1)
a÷1 a+1
a
2
+3
is convergent (do it!).
Definition 7. (absolute convergence) A series
o
¸
a=0
c
a
is said to be
absolutely convergent if the series of moduli
o
¸
a=0
[c
a
[ is convergent.
For instance, the series
o
¸
a=1
(÷1)
a 1
a
2
is convergent (why?) and ab
solutely convergent, but the series
o
¸
a=0
(÷1)
a 1
a
is convergent (why?) and
it is not absolutely convergent, because the harmonic series
o
¸
a=1
1
a
=
48 2. SERIES OF NUMBERS
2(1) = · (see the Riemann series). A series which is convergent, but
not absolutely convergent, is called semiconvergent.
The following result says that the notion of absolutely convergence
is stronger then the notion of (simple) convergence.
Theorem 31. Any absolute convergence series
o
¸
a=0
c
a
is also (sim
ple) convergent.
Proof. We use again the Cauchy General Test (see Theorem 18).
Let :
a
= c
0
+ c
1
+ ... + c
a
be the :th partial sum of the initial series
o
¸
a=0
c
a
and let o
a
= [c
0
[ +[c
1
[ +... +[c
a
[ be the :th partial sum of the
series
o
¸
a=0
[c
a
[ . Let us evaluate
(2.4) [:
a+j
÷:
a
[ = [c
a+1
+c
a+2
+... +c
a+j
[ _
[c
a+1
[ +[c
a+2
[ +... +[c
a+j
[ = [o
a+j
÷o
a
[ .
Let 0 be a small positive real number and let ` be a su¢ciently
large natural number such that for any : _ ` one has [o
a+j
÷o
a
[ <
for any j = 1. 2. ... (since ¦o
a
¦ is a Cauchy sequence). From (2.4) we
have that [:
a+j
÷:
a
[ _ [o
a+j
÷o
a
[ . so [:
a+j
÷:
a
[ _ for any : _ `
and for any j = 1. 2. ... . But this means that the sequence ¦:
a
¦ is a
Cauchy sequence. Hence the series
o
¸
a=0
c
a
is convergent (see Theorem
18).
For instance, the series
o
¸
a=1
sin(5a)
a
2
is convergent because it is ab
solutely convergent. Indeed, since
sin(5a)
a
2
_
1
a
2
and since the series
o
¸
a=1
1
a
2
= 2(2) is convergent (see the Riemann series), the Comparison
Test says that the series of moduli
o
¸
a=1
[sin(5a)[
a
2
is convergent, i.e. the
initial series
o
¸
a=1
sin(5a)
a
2
is convergent.
Remark 7. (see [Nik] or [Pal]) We saw above that any absolutely
convergent series is convergent, but the converse is not true. Cauchy
proved that in any absolutely convergent series one can change the order
of the terms in the in…nite sum (by any permutation) and the sum of
the series remains the same. On the contrary, Riemann proved that
2. SERIES WITH ARBITRARY TERMS 49
for a semiconvergent series
o
¸
a=0
c
a
and for any number ¹ ÷ R = R '
¦÷·. ·¦. one can …nd a permutation of the terms of the series
o
¸
a=0
c
a
such that its sum becomes exactly ¹. Two absolutely convergent series
can be multiplied by the usual polynomial multiplication rule
o
¸
a=0
c
a
o
¸
a=0
/
a
=
o
¸
a=0
c
a
. where c
a
= c
0
/
a
+c
1
/
a÷1
+... +c
a
/
0
.
and the resulting product series is again absolutely convergent (Mer
taens).
Remark 8. If instead of series with real numbers we consider a
series with complex numbers
o
¸
a=0
.
a
. where .
a
= r
a
+in
a
. r
a
. n
a
÷ R for
any : = 0. 1. 2. ..., we say that such a series is convergent to its sum
: = n +i·. n. · ÷ R if the sequence of partial sums
:
a
= .
0
+.
1
+... +.
a
= (r
0
+r
1
+... +r
a
) +i(n
0
+n
1
+... +n
a
)
is convergent to :. i.e.
[: ÷:
a
[ =
[n ÷(r
0
+r
1
+... +r
a
)]
2
+ [· ÷(n
0
+n
1
+... +n
a
)]
2
÷0.
when : ÷ ·. This is equivalent to saying that both series with real
numbers,
o
¸
a=0
r
a
(the real part) and
o
¸
a=0
n
a
(the imaginary part) are con
vergent to n and · respectively. Hence,
o
¸
a=0
.
a
=
o
¸
a=0
r
a
+ i
o
¸
a=0
n
a
and
the calculus with complex series reduces to the calculus with real series.
Practically, in general, it is di¢cult to decide if both the "real part"
and the "imaginary part" are convergent. For instance, let us consider
the series
: =
o
¸
a=0
(1 +i)
a
:!
=
o
¸
a=0
2
a
1
2
+i
1
2
a
:!
=
o
¸
a=0
2
a
cos
¬
4
+i sin
¬
4
a
:!
Let us use now the Moivre formula and …nd:
: =
o
¸
a=0
2
a
cos :
¬
4
:!
+i
o
¸
a=0
2
a
sin :
¬
4
:!
.
Since
2
a
cos :
¬
4
:!
_
2
a
:!
50 2. SERIES OF NUMBERS
and since
lim
a÷o
2
n+1
(a+1)!
2
n
a!
= 0.
the series
o
¸
a=0
2
n
cos a
4
a!
is absolutely convergent, so it is convergent
(why?precise the theorems that we used!). In the same way we prove
that the imaginary part series
o
¸
a=0
2
n
sin a
4
a!
is also convergent. An eas
ier way to prove the convergence of the complex series : =
o
¸
a=0
(1+j)
n
a!
is the following. It is not di¢cult to prove that an absolutely conver
gent series
o
¸
a=0
.
a
(i.e.
o
¸
a=0
[.
a
[ is convergent) is also convergent (see
the proof of Theorem 31). In our case,
(1 +i)
a
:!
=
([1 +i[)
a
:!
=
2
a
:!
.
So, the series
o
¸
a=0
[.
a
[ =
o
¸
a=0
2
n
a!
is convergent (use the ratio test),
i.e. the series : =
o
¸
a=0
(1+j)
n
a!
is absolutely convergent. Hence, it is
convergent. If a series
o
¸
a=0
.
a
is not absolutely convergent, the general
way to study it is to write it as:
o
¸
a=0
.
a
=
o
¸
a=0
r
a
+i
o
¸
a=0
n
a
and to study separately the real series
o
¸
a=0
r
a
and
o
¸
a=0
n
a
. If both of them
are convergent, the initial series is also convergent. If at least one of
them is divergent, the series
o
¸
a=0
.
a
is divergent (why?).
3. Approximate computations
Usually, whenever one cannot exactly compute the sum of a con
vergent series : =
o
¸
a=0
c
a
. one approximate : by its :th partial sum
:
a
= c
0
+c
1
+... +c
a
. for su¢ciently large :. For instance,
: =
o
¸
a=0
1
:
2
 :
1000
=
1
1
2
+
1
2
2
+... +
1
1000
2
.
3. APPROXIMATE COMPUTATIONS 51
The di¤erence
a
= [: ÷:
a
[ is called the (absolute) error of order : in
our process of approximation. It is clear enough why we are interested
in the evaluation of this error. Since the series is convergent,
a
÷ 0.
when : becomes large enough. Given a small positive real number
0. the problem is to …nd an : (very small if it is possible!) which
depend on . such that the error
a
< . For instance, if =
1
10
3
. we
say that ": is approximated by :
a
with 3 exact decimals".
We study this problem in two cases.
Case 1 Let : =
¸
o
a=0
c
a
be a series with positive terms (c
a
0.
: = 0. 1. ...) and let c ÷ (0. 1) such that
o
n+1
on
_ c for : _ ` (remember
yourself the Ratio Test). The series is convergent (see Theorem 27).
Let now / be a natural number greater or equal to `. Let us evaluate
the error
I
= : ÷:
I
:
(3.1)
I
= c
I+1
+c
I+2
+... _ cc
I
+c
2
c
I
+... =
c
1 ÷c
c
I
We see that if 0 is an arbitrary small positive real number, always
one can …nd a least / ÷ N such that
c
1÷c
c
I
< . Since
I
_
c
1÷c
c
I
. for
this / one also has:
I
< . If we want a small /. we must …nd a small
c ÷ (0. 1) such that for a small ` (0 if it is possible), we have
o
n+1
on
_ c
for : _ `.
Let us compute the value of
o
¸
a=0
1
a!
(we shall see later that it is
exactly c. the base of the Neperian logarithm) with 2 exact decimals.
Since
o
n+1
on
=
1
a+1
_
1
2
for : _ 1.
I
= : ÷:
I
_
1
2
1 ÷
1
2
1
/!
=
1
/!
.
Let us …nd the least / such that
1
I!
< =
1
10
2
. By trials, / = 1. 2. ....
we …nd / = 5. So
:  :
5
= 1 +
1
1!
+
1
2!
+
1
3!
+
1
4!
+
1
5!
= 2.71666....
i.e. we obtained the value of c with 2 exact decimals, c  2.71.
Let : =
¸
c
a
be a series with nonnegative terms (c
a
_ 0. : =
0. 1. ...) and let c ÷ (0. 1) such that
n
c
a
_ c for : _ ` (remember
yourself the Cauchy Root Test). The series is convergent (see Theorem
29). Let now / be a natural number greater or equal to `. Let us
evaluate the error
I
= :÷:
I
. Prove that
I
_
c
k+1
1÷c
. Use this estimation
to …nd the value of : =
o
¸
a=1
1
a
n
2
with 3 exact decimals.
52 2. SERIES OF NUMBERS
Case 2 Suppose now that we want to approximate the value of an
alternate series, : =
o
¸
a=1
(÷1)
a÷1
c
a
. where ¦c
a
¦ is a decreasing sequence
with nonnegative terms and c
a
÷ 0. The Leibniz test (see Corollary
2) says that our series is convergent. Since
:
2a
= :
2a÷2
+ (c
2a÷1
÷c
2a
) _ :
2a÷2
and since
:
2a+1
= :
2a÷1
÷(c
2a
÷c
2a+1
) _ :
2a÷1
.
one has:
(3.2) :
2
_ :
4
_ :
6
_ ... _ :
2a
_ ... _ : _ ... _ :
2a+1
_ ... _ :
3
_ :
1
.
So,
0 _ : ÷:
2a
_ :
2a+1
÷:
2a
= c
2a+1
and
0 _ :
2a+1
÷: _ :
2a+1
÷:
2a+2
= c
2a+2
.
Hence
(3.3)
a
= [: ÷:
a
[ _ c
a+1
i.e. the absolute error is less or equal to the modulus of the …rst ne
glected term. Here, in fact we have another proof of the Leibniz Test
(see Theorem 2). This one is independent of the AbelDirichlet Test
(Theorem 30). It uses only Cantor Axiom (Axiom 2) (where?).
Let us compute : =
o
¸
a=1
(÷1)
a÷1 1
(a!)
2
with 2 exact decimals. We use
the estimation (3.3) and force with
c
a+1
=
1
[(: + 1)!]
2
<
1
10
2
for : _ 3. so
:  :
3
=
1
1
÷
1
4
+
1
36
= 0.777... = 0.(7)
4. Problems
1. Compute the sum of the following series:
a)
o
¸
a=2
ln
1 ÷
1
a
2
; b)
o
¸
a=1
2
n1
+3
n
5
n+1
; c)
o
¸
a=1
1
a(a+2)
; d)
o
¸
a=1
1
a(a+1)(a+2)
;
e)
o
¸
a=1
1
(a+2)(a+4)
; f)
o
¸
a=1
(÷1)
a 1+2
n1
3
n2
;
2. Decide if the following series are convergent or not:
4. PROBLEMS 53
a)
o
¸
a=0
2
n
a!
; b)
o
¸
a=1
1·4·7·...·(1+3a)
1·5·9·...·(1+4a)
1
a
; c)
o
¸
a=0
(÷1)
a 1
a!
; d)
o
¸
a=0
2
n
+1
2
n+1
+1
c
a
. c _ 0
(discussion on c); e)
o
¸
a=1
:
2c÷1
2
a
(discussion on c ÷ R); f)
o
¸
a=1
(÷1)
n
10
n
a!
;
g)
o
¸
a=1
2·7·12·...·[2+5(a÷1)]
3·8·13·...·[3+5(a÷1)]
; h)
o
¸
a=0
(c+2)
n
2
n
+3
n
. (discussion on c _ 0); i)
o
¸
a=1
1
a
(2\ ÷
1)
a
. (discussion on \ ÷ R); j)
o
¸
a=1
(4c÷5)
n
a·5
n
. c _ 2 (discussion on c);
k)
o
¸
a=0
1
3
a
+2
(discussion on c); l)
o
¸
a=1
2
n
1·3·5·...·(2a÷1)
(2c÷1)
a
. (discussion on
c _ 1); m)
o
¸
a=1
1
3
4a+1÷
3
4a÷1
; n)
o
¸
a=1
3
ln a
; o)
o
¸
a=1
2(a!)
(2a)!
; p)
o
¸
a=0
(÷1)
n
a!
(1 + 3
a
);
r)
o
¸
a=0
2
n2
3
n+1
+1
; s)
o
¸
a=1
5a+1
6a÷2
c
a
(discussion on c _ 0).
3. Find the Abel’s degree of the expression 1 =
3
a
5
+2
5
a
3
+a+3
a+2÷
a
.
: ÷ N.
4. Use the cComparison Test to decide if the series
o
¸
a=1
sin
1
3
a+1
is convergent or not.
5. Find all r ÷ R such that the series
o
¸
a=0
a
2
+1
a+1
r
a
be convergent.
What about all r ÷ C such that the same series be convergent?
6. Find all . in C such that the following series be absolute conver
gent.
a)
o
¸
a=0
:
n
a!
; b)
o
¸
a=1
(:÷j)
n
a
; c)
o
¸
a=0
:.
a
; d)
o
¸
a=0
(. ÷3i + 2)
a
;
7. Draw the set ` =
r ÷ R [
o
¸
a=1
(÷1)
a a
n
a3
n
is convergent
¸
on the
real line.
8. Draw the set l =
. ÷ C [
o
¸
a=1
(÷1)
a :
n
a3
n
is convergent
¸
in the
complex plane.
9. Compute
o
¸
a=1
(÷1)
a 1
a
2
with 2 exact decimals.
10. Compute
o
¸
a=1
2
n
a!
with one exact decimal.
11. Prove the RaabeDuhamel test. Hint:
a) Write:
`c
.
÷(` + 1)c
.+1
_ (\ ÷1)c
.+1
(` + 1)c
.+1
÷(` + 2)c
.+2
_ (\ ÷1)c
.+2
............................................
54 2. SERIES OF NUMBERS
(` +j)c
.+j
÷(` +j + 1)c
.+j+1
_ (\ ÷1)c
.+j+1
Sum these inequalities on columns and get:
`c
.
÷(`+j+1)c
.+j+1
_ (\÷1) [c
.+1
+c
.+2
+c
.+3
+... +c
.+j+1
]
So
`c
.
\ ÷1
_ c
.+1
+c
.+2
+c
.+3
+... +c
.+j+1
for any j = 1. 2. .... Hence, the partial sums of our initial series are
bounded. Thus the series is convergent.
b) Since :c
a
< (: +1)c
a+1
for : _ `. the limit lim
a÷o
:c
a
is greater
than 0. So, using the ccomparison test for c = 1. we get that our
initial series is divergent (why?).
c) Apply a) and b).
12. Compute
¸
o
a=1
1
a
n
with 3 exact decimals (use the approximate
computation with the Root Test).
CHAPTER 3
Sequences and series of functions
1. Continuous and di¤erentiable functions
Recall that a metric space is a set A with a distance d on it. A
distance d on A is a function which associates to any pair (r. n) of A
a nonnegative real number d(r. n) with the following properties:
d1. d(r. n) = 0 if and only if r = n.
d2. d(r. n) = d(n. r) for any r and n in A.
d3. d(r. n) _ d(r. .) +d(.. n) for any r. n and . in A.
See also the Remark 2. We usually denote by (A. d) a metric space
A with a distance d on it. The standard example of a metric space
is (R, d). where d(r. n) = [r ÷n[ . We say that r
a
÷ r in (A. d) if
the numerical sequence ¦d(r
a
. r)¦ tends to zero, i.e. if the distance
between r
a
and r becomes smaller and smaller to zero as : ÷·. We
de…ne again the basic notion of continuity.
Definition 8. (continuity of a function at a point) Let (A. d).
(A
t
. d
t
) be two metric spaces, let 1 : A ÷ A
t
be a function de…ned
on A with values in A
t
and let r be a …xed element in A. We say
that 1 is continuous at r if for any sequence ¦r
a
¦ which converges to
r. we have that 1(r
a
) ÷ 1(r). For instance, if A = A
t
= R, with
the usual distance, 1 is continuous at a point r if the graphic of 1 is
not "broken (or interrupted)" in r (see Fig.3.1). All the elementary
functions (polynomials, rational functions, power functions, exponen
tial functions, logarithmic functions, trigonometric functions) and their
compositions are continuous on their de…nition domains, i.e. in any
point of their de…nition domains (see also the Theorem 14). Hence, the
continuity is essentially a "local" property, i.e. its de…nition shows the
behavior of the function 1 at a given point r.
55
56 3. SEQUENCES AND SERIES OF FUNCTIONS
Fig. 3.1
For instance, a) 1 : R ÷R, 1(r) =
a
3
+1
a
2
+1
is continuous on the whole
R. Indeed, let c be a …xed point in R and let ¦c
a
¦ be a sequence
convergent to c. Then, using the basic properties of the convergent
sequences relative to the elementary algebraic operations (+. ÷. . :. see
the Theorem 14), we …nd that
1(c
a
) =
c
3
a
+ 1
c
2
a
+ 1
÷
c
3
+ 1
c
2
+ 1
= 1(c).
i.e. the function 1 is continuous at c. for any c ÷ R. Hence 1 is contin
uous on R. Now, if we compose the function ln r (which is continuous
on (0. ·)) with 1(r) we get a new continuous function o(r) = ln
a
3
+1
a
2
+1
on (÷1. ·) (why?).
Remark 9. We need in this chapter another basic "local" notion,
namely the notion of di¤erentiability of a function 1 at a given point
c. Recall that a subset ¹ of R is said to be open if for any point c
of ¹. there is a small positive real number . such that the interval
(c ÷. c +) (the "ball" with centre at c and of radius . usually called
the neighborhood of c) is completely included in ¹ (de…ne the notion
of an open subset in a metric space (A. d); instead of neighborhoods
use open balls 1(c. ) = ¦r ÷ A : d(r. c) < ¦. etc.). A subset 1 of
R is said to be closed if its complementary R` 1 is an open subset (1
is closed in an arbitrary metric space (A. d) if A ` 1 is open in A).
For instance, (÷·. 1) is open and [÷3. 7] is closed. If A = (÷1. 7).
1. CONTINUOUS AND DIFFERENTIABLE FUNCTIONS 57
with the induced distance of R, then [0. 7) is closed in A. but NOT
in R (why?). It is not di¢cult to prove that a subset 1 is closed if
and only if for any sequence ¦/
a
¦ ÷ /. with all /
a
in 1. one has that
/ ÷ 1 (prove it!). For instance, if 1 : A ÷R is a continuous function
de…ned on a metric space (A. d) and if \ is a real number, then the
set 1
A
= ¦r ÷ A : 1(r) _ \ (or _ \. or = \) ¦ is closed in A.
Indeed, let ¦/
a
¦ be a sequence of elements in 1. which is convergent
to an element / in A. Since 1 is continuous, 1(/
a
) ÷ 1(/). Because
/
a
÷ 1. 1(/
a
) _ \ for any : = 0. 1. .... Then 1(/) _ \ (otherwise,
1(/) < \ and, from a rank ` on, 1(/
a
) < \. for : _ ` (why?see the
de…nition of the limit 1(/
a
) ÷1(/)!)), a contradiction i.e. / itself is in
1 and so 1 is a closed subset in A.
Definition 9. Let ¹ be an open subset of R (for instance an open
interval (c. d)), let 1 : ¹ ÷R be a function de…ned on ¹ with values real
numbers and let c be a …xed point in ¹. We say that 1 is di¤erentiable
at c if the following limit exists (and is …nite!):
(1.1) lim
a÷o
1(r) ÷1(c)
r ÷c
act
= 1
t
(c)
The limit of a function o : ¹ ÷R in a limit point / (it is the limit
of at least one sequence of elements from ¹) of ¹ is a unique number
 ÷ R such that for any sequence ¦/
a
¦. /
a
÷ ¹ which is convergent to
/. one has that o(/
a
) ÷. We shortly write lim
a÷b
o(r) = . Not always a
function o has a limit at a given limit point /. For instance, the function
:io: : R ÷¦÷1. 0. 1¦.
(1.2) :io:(r) =
÷1. if r < 0
0. if r = 0
1. if r 0
has the limit  = ÷1 at any point c < 0. has the limit  = 1 at any
point c 0 and at 0 it has no limit at all (prove this!).
We recall that the limit "on the left" of a function 1 : ¹ ÷ R,
¹ · R, ¹ an open subset, at a point c of ¹ is a number 

such that
for any sequence ¦r
a
¦. r
a
_ c. which is convergent to c. one has that

c
= lim1(r
a
). If we take r
a
"on the right" of c. we get the notion of
the limit 
v
"on the right" of 1 at c. A function 1 has the limit  at c
if and only if 

= 
v
=  (prove it!).
It is clear enough that a continuous function 1 at a point c ÷ ¹
has the limit  = 1(c) at c (why?). In fact, a function 1 : ¹ ÷ R is
continuous at a point c ÷ ¹ if and only if it has a limit  at c and if
that one is exactly  = 1(c) (prove it!).
58 3. SEQUENCES AND SERIES OF FUNCTIONS
We call the number 1
t
(c) from (1.1) the derivative of 1 at c. The
linear function d1(c) : R ÷ R, d1(c)(r) = 1
t
(c) r is called the (…rst)
di¤erential of 1 at c. This is simply a dilation (or a homotety) of mod
ulus 1
t
(c) of the real line R. If the function 1 is di¤erentiable at any
point c of ¹. we say that 1 is di¤erentiable (or has a derivative) on ¹. In
this last case, the new function c 1
t
(c). where c runs on ¹. is called
the (…rst) derivative of 1. It is denoted by 1
t
. We know (see any elemen
tary course in Calculus for the di¤erent rules in computing derivatives!)
that almost all the elementary functions (described above) and their
compositions (recall the chain rule: (1 · o)
t
(c) = 1
t
(o(c)) o
t
(c)) are
di¤erentiable on their de…nition domains. "Almost" because of some
exceptions like 1(r) =
r. 1 : [0. ·) ÷ R. Since 1
t
(r) =
1
2
a
. the
derivative of 1 does not exists at c = 0. Indeed, lim
a÷0. a0
a÷0
a
= ·! One
can interpret the derivative of a function 1 at a point c. either as "the
velocity" of 1 at c or as the slope of the tangent line at c to the graphic
of 1 (why?). Not all the continuous functions at a given point c are also
di¤erentiable at c (see Fig.3.2). But a di¤erentiable function 1 at a
given point c is continuous. Indeed, let r
a
÷c. lim
an÷o
;(an)÷;(o)
an÷o
= 1
t
(c)
(see De…nition 9 and what follows) says that ONLY the nondetermin
istic case
0
0
could give a …nite number 1
t
(c). Hence, 1(r
a
) ÷1(c). i.e.
1 is continuous at c.
y
tg α = f'(x
1
)
α
O
x
1
x
2
continuous but
not differentiable
in x
2
differentiable
in x
1
x
y = f(x)
y = g(x)
Fig. 3.2
1. CONTINUOUS AND DIFFERENTIABLE FUNCTIONS 59
Let ( be a set and let 1 : ( ÷R be a function de…ned on ( with
values in R. We say that 1 is bounded if its image 1(() = ¦1(r) : r ÷
(¦ is a bounded subset in R. This means that there is a positive real
number ` 0 such that [1(r)[ < ` (i.e. ÷` < 1(r) < `) for any
r ÷ (. Equivalently, if ( · R. then 1 is bounded if the graphic of it
is contained into the band bounded by the horizontal lines: n = ÷`
and n = `
A fundamental property of continuous functions is the following:
Theorem 32. (Weierstrass boundedness theorem) Let 1 : [c. /] ÷
R be a continuous function de…ned on the closed and bounded inter
val [c. /]. Then 1 is bounded, `
oc;
= sup 1([c. /]) = 1(c) and :
oc;
=
inf 1([c. /]) = 1(d). where c. d ÷ [c. /]. This means that the least up
per bound (sup 1([c. /]) and the greatest lower bound (inf 1([c. /]) of the
bounded set 1([c. /]) are realized at c and at d respectively.
Proof. a) Let us prove that ` = sup 1([c. /]) < ·. Suppose
on the contrary, namely that ` = ·. Then, there is at least one
sequence ¦r
a
¦ of elements from [c. /] such that 1(r
a
) ÷·. Since ¦r
a
¦
is bounded, we can apply the CesaroBolzanoWeierstrass Theorem(see
Theorem 12) and …nd a subsequence ¦r
a
k
¦ of ¦r
a
¦ which is convergent
to an r
+
÷ [c. /] (here we use the fact that [c. /] is closed, how?). Since
1 is continuous, one has that 1(r
a
k
) ÷ 1(r
+
) when / ÷ ·. But
1(r
a
) ÷· and the uniqueness of the limit implies that 1(r
+
) = ·. a
contradiction (why?). Hence 1 is upper bounded. In the same way we
can prove that 1 is lower bounded (do it!).
b) Let us prove now that ` = 1(c) for a c in [c. /]. Since ` is the
least upper bound, for any natural number : we can …nd an element
n
a
÷ [c. /] such that
(1.3) ` ÷
1
:
_ 1(n
a
) _ ` (why?)
The sequence ¦n
a
¦ is bounded and nonconstant (why?). Applying
again the CesaroBolzanoWeierstrass Theorem, one can …nd a sub
sequence ¦n
a
k
¦ of ¦n
a
¦ which is convergent to an element c ÷ [c. /]
(because the interval is closed). Since 1 is continuous, 1(n
a
k
) ÷ 1(c).
when / ÷·. Making / ÷· in the inequality `÷
1
a
k
_ 1(n
a
k
) _ `
and using the de…nition of a subsequence (:
1
< :
2
< ... ), we get that
` = 1(c). To prove that : = 1(d). d ÷ [c. /]. we work in the same
manner (do it!).
Theorem 33. (Darboux) Let 1 : [c. /] ÷ R be a continuous func
tion de…ned on the closed and bounded interval [c. /]. Let ` = sup 1([c. /])
and let : = inf 1([c. /]). Then the image of the interval [c. /] through 1
60 3. SEQUENCES AND SERIES OF FUNCTIONS
is exactly the closed interval [:. `]. More general, a continuous func
tion carries intervals into intervals.
Proof. Let \ be an element in [:. `]. We want to …nd an element
. in [c. /] such that 1(.) = \. If \ is equal to : or to `. we can take
. = d or c (from Theorem 32) respectively. So, we can assume that
\ ÷ (:. `) and that 1 is not a constant function (in this last case
the statement of the theorem is obvious). We de…ne two subsets of the
interval [c. /]:
¹
1
= ¦r ÷ [c. /] : 1(r) _ \¦
and
¹
2
= ¦r ÷ [c. /] : 1(r) _ \¦.
If ¹
1
¨ ¹
2
is not empty, take . in this intersection and the proof is
…nished. Suppose on the contrary, namely that ¹
1
¨ ¹
2
= ?. Since
\ cannot be either : or `. ¹
1
and ¹
2
are not empty (why?). Now,
[c. /] = ¹
1
'¹
2
(why?) and, since 1 is continuous, ¹
1
and ¹
2
are closed
in R (see Remark 9). In order to obtain a contradiction, we shall prove
that it is not possible to decompose (to write as a union, or to cover) an
interval [c. /] into two disjoint closed and nonempty subsets. Indeed,
let c
2
= sup ¹
2
. Since 1 is continuous, 1(c
2
) _ \ (why?remember the
de…nition of the least upper bound and of the continuity!) i.e. c
2
÷ ¹
2
.
If c
2
= /. then the subset o
1
= ¦r ÷ ¹
1
: r c
2
¦ is not empty (why?).
Take now c
1
= inf o
1
. Since ¹
1
is closed, c
1
÷ ¹
1
(why?). If c
1
c
2
.
take / ÷ (c
2
. c
1
). This / ÷ [c. /] and it cannot be either in ¹
1
or in ¹
2
(why?). Since c
1
_ c
2
. the unique possibility for c
1
is to be equal to c
2
.
But then, c = c
1
= c
2
÷ ¹
1
¨¹
2
= ?. a contradiction! Hence, c
2
= sup
¹
2
= /. Take now d
2
= inf ¹
2
. Since ¹
2
is closed, one has that d
2
÷ ¹
2
.
If d
2
= c. then the subset o
2
= ¦r ÷ ¹
1
: r < d
2
¦ is not empty (why?).
Take now d
1
= sup o
2
. Since ¹
1
is closed, d
1
÷ ¹
1
(why?). If d
1
< d
2
.
take again o ÷ (d
1
. d
2
) and this last one cannot be either in ¹
1
or in ¹
2
.
Hence d
1
= d
2
act
= d and this one must be in ¹
1
¨ ¹
2
. a contradiction!
So, d
2
= c. i.e. inf ¹
2
= c and sup ¹
2
= /. thus ¹
2
= [c. /]. Since ¹
1
is not empty and it is included in [c. /]. ¹
1
· ¹
2
. and we get again a
new and the last contradiction! Hence ¹
1
¨ ¹
2
cannot be empty and
the proof of the theorem is over.
We agree with the reader that the proof of this last theorem is too
long! But,...it is so clear and so elementary! Trying to understand and
to reproduce logically the above proof is a good exercise for strengthen
your power of concentration and not only!
Theorem 34. Let 1 be an open interval on the real line and let
1 : 1 ÷ R, be a continuous function de…ned on 1 with real values.
1. CONTINUOUS AND DIFFERENTIABLE FUNCTIONS 61
1) Assume that there are two points / and d in 1 (/ < d) such that
the values 1(/) and 1(d) are nonzero and have distinct signs. Then,
there is a point c in the interval [/. d] at which the value of 1 is zero,
i.e. 1(c) = 0. 2) Now suppose that at c ÷ 1 the value 1(c) 0 (or
1(c) < 0). Then there is an neighborhood (c ÷ . c + ) · 1. such
that 1(r) 0 (or 1(r) < 0) for any r ÷ (c ÷. c +).
Proof. 1) We can simply apply Theorem 33. Indeed, since 1(1)
is an interval (Theorem 33), the segment generated by 1(/) and 1(d)
is completely contained in 1([/. d]). Since 1(/) and 1(d) have distinct
signs, 0 is between them, so, 0 ÷ 1([/. d]). or 0 = 1(c) for a c ÷ [/. d].
2) Suppose that 1(c) 0. Let us assume contrary, i.e. for all small
possible we can …nd in (c ÷ . c + ) at least on number r
.
(an r
which depends on ) such that 1(r
.
) _ 0. Take for such epsilons the
values
1.
1
2
.
1
3
. ....
1
:
. ....
and …nd r1
n
÷ (c ÷
1
a
. c +
1
a
) with 1(r1
n
) _ 0. : = 1. 2. ... . Since
1 is continuous at c and since the sequence ¦r1
n
¦ tends to c (why?),
one has that 1(r1
n
) ÷ 1(c). But 1(r1
n
) are all nonpositive, so 1(c) is
nonpositive, a contradiction! Hence, there is at least one small enough
such that for any r in (c ÷. c + ). 1(r) 0. The case 1(c) < 0 can
be similarly manipulated (do it!).
Definition 10. Let (A. d) be a metric space and let 1 be an interval
on the real line R (a subset 1 of R is said to be an interval if for any
pair of numbers :
1
. :
2
÷ 1 and any real number : with :
1
_ : _ :
2
.
one has that : ÷ 1). Practically, we think of a curve in A as being the
image in A of an interval 1 through a continuous function / : 1 ÷A.
More exactly, we denote the couple (1. /) by a small greek letter and
say that is a curve in A. If ¹ and 1 are two "points" (elements)
in A. we say that a curve = (1. /) connects ¹ and 1 if there are
c. / ÷ A such that ¹ = /(c) and 1 = /(/). By an (closed) arc [¹1]
in A we mean the image in A of a closed interval [c. /] of R through
a continuous function / : [c. /] ÷ A. i.e. [¹. 1] = ¦r ÷ A : there is
c ÷ [c. /] with /(c) = r¦.
Example 1. a) Let ¦(; i. j. k¦ be a Cartesian coordinate system
in the vector space \
3
of all free vectors in our 31 space (identi…ed
with R
3
). Any point ` in R
3
has 3 coordinates: `(r. n. .). where
÷÷÷
(` = ri+nj+.k. r. n. . ÷ R. Let ¹(c
1
. c
2
. c
3
) and 1(/
1
. /
2
. /
3
) be two
62 3. SEQUENCES AND SERIES OF FUNCTIONS
points in R
3
. The usual segment [¹. 1] is a closed arc which connect the
points ¹ and 1. Indeed, let / : [0. 1] ÷R
3
. /(t) = (c
1
+t(/
1
÷c
1
). c
2
+
t(/
2
÷c
2
). c
3
+t(/
3
÷c
3
)). be the usual continuous parameterization of
the segment [¹. 1] :
r = c
1
+t(/
1
÷c
1
)
n = c
2
+t(/
2
÷c
2
)
. = c
3
+t(/
3
÷c
3
)
. t ÷ [0. 1]
Here = ([0. 1]. /) is a curve in R
3
. This function / describes a com
position between the dilation of moduli /
1
÷ c
1
. /
2
÷ c
2
. /
3
÷ c
3
. along
the (r. (n. and (. axes respectively, and the translation x ÷ a + x.
of center a = (c
1
. c
2
. c
3
).
b) Let ( = ¦(r. n) ÷ R
2
: (r÷c)
2
+(n ÷/)
2
= :
2
¦ be the circle with
center at (c. /) and radius :. The parametrization of (
r = c +: cos t
n = / +: sin t
. t ÷ [0. 2:]
give rise to a curve = ([0. 2:]. /). where /(t) = (c+: cos t. /+: sin t).
In fact, / describes the continuous deformation process of the segment
[0. 2:] · R into the circle ( in the metric space R
2
.
Definition 11. A subset ¹ of a metric space (A. d) is said to be
connected if any pair of two points `
1
and `
2
of ¹ can be connected
by a continuous curve = (1. /). / : 1 ÷A.
Corollary 3. The connected subsets in R are exactly the intervals
of R (for proof use the Darboux Theorem 33).
For instance, ¹ = [0. 1] '[5. 8] is not connected because it is not an
interval (4 is between 0 and 8. but it is not in ¹!).
Remark 10. A subset o of R
3
is said to be convex if for any pair
of points ¹. 1 ÷ o. the whole segment [¹. 1] is included in o. For
instance, the parallelepipeds, the spheres, the ellipsoids, etc., are convex
subsets of R
3
. The union between two tangent spheres is connected but
it is not convex! (why?). It is clear that any convex subset of R
3
is also
a connected subset in R
3
(prove it!).
Definition 12. Let 1 : ¹ ÷ R be a function de…ned on an open
subset ¹ of R with values in R. A point c of ¹ is a local maximum
point of 1 if there is an neighborhood of c. (c÷. c+) · ¹. such that
1(r) _ 1(c) for any r ÷ (c÷. c+). The value 1(c) of 1 at c is called
a local extremum (maximum) for 1. A point / of ¹ is said to be a local
minimum point for 1 if there is an :neighborhood of /. (/÷:. /+:) · ¹.
such that 1(r) _ 1(/) for any r ÷ (/ ÷ :. / + :). The value 1(/) of 1
1. CONTINUOUS AND DIFFERENTIABLE FUNCTIONS 63
at / is called a local extremum (minimum) for 1. A local maximum
point or a local minimum point is called a local extremum point. The
local extrema of 1 on ¹ are all the local maxima and the local minima
of 1 in ¹. The (global) maximum of 1 on ¹ is max 1(¹) (÷ R). The
(global) minimum of 1 on ¹ is min 1(¹) (÷ R) (see Fig.3.3).
y
O x x
1 x
2
x
3
x
4
global
max.
global
min.
not local
extremum
local
min.
local
max.
( )
Fig. 3.3
A critical (or stationary) point c ÷ ¹ for a di¤erentiable function
1 : ¹ ÷R on ¹ is a root of the equation 1
t
(r) = 0. i.e. 1
t
(c) = 0. For
instance, c = 2 is a stationary point for 1(r) = (r ÷ 2)
3
. 1 : R ÷ R,
but it is not an extremum point for 1 (why?). The next result clari…es
the converse situation.
Theorem 35. (11 Fermat’s Theorem) Let c be a local extremum
(local maximum or local minimum) point for a function 1 : ¹ ÷ R
(¹ is open). Assume that 1 is di¤erentiable at c. Then 1
t
(c) = 0. i.e.
c is a critical point of 1. Practically, this statement says that for a
di¤erentiable function 1 we must search for local extrema between the
critical points of 1. i.e. between the solutions of the equation 1
t
(r) = 0.
r ÷ ¹.
Proof. Suppose that c is a local maximum point for 1. i.e. there
is a small 0 such that (c ÷ . c + ) · ¹ and 1(r) _ 1(c) for any
64 3. SEQUENCES AND SERIES OF FUNCTIONS
r in (c ÷ . c + ) (if c is a local minimum point, one proceeds in the
same way, do it!). Look now at the formula:
(1.4) lim
a÷o
1(r) ÷1(c)
r ÷c
= 1
t
(c)!
If r ÷ (c ÷ . c + ) and r < c. since 1(r) _ 1(c). one has that
1
t
(c) _ 0 (why?). Now, if r ÷ (c ÷ . c + ). but r c. again since
1(r) _ 1(c). one gets that 1
t
(c) _ 0. Both inequalities give us that
1
t
(c) = 0 and the Fermat’s theorem for a function of one variable is
proved.
However, the Fermat’s Theorem works only at the points at which
our function is di¤erentiable. For instance, 1(r) = [r[ has at r = 0
a local (even a global) minimum (why?), but it is not di¤erentiable
at this point (why?). The moral is that we must consider separately
the points at which a function is not di¤erentiable and see (using the
de…nition only!) if these points are or not local extremum points for
our function.
Theorem 36. (Rolle Theorem) Let 1 : [c. /] ÷ R (c < /) be
a continuous function. Assume that 1 is di¤erentiable on the open
subinterval (c. /) and that 1(c) = 1(/). Then there is at least one point
c ÷ (c. /) such that 1
t
(c) = 0.
Proof. Let us apply the Weierstrass boundedness theorem (The
orem 32) and …nd : = inf 1([c. /]) and ` = sup 1([c. /]) as real num
bers. If : = `. then our function is a constant function and so,
1
t
(r) = 0 for any r in (c. /). Hence we assume that : = `. So the
number 1(c) = 1(/) cannot be simultaneously equal to : and `. Sup
pose for instance that 1(c) = 1(/) = `. Thus, a c with ` = 1(c).
c ÷ [c. /] (see the Weierstrass boundedness theorem) cannot be either
c or /. i.e. c ÷ (c. /). Therefore, this c is a local maximum for 1. Use
now Fermat’s Theorem and …nd that 1
t
(c) = 0.
For instance, if 1(r) = r
4
÷ 16. r ÷ [÷1. 1]. then 1(÷1) = 1(1) =
÷15 and 1
t
(r) = 0 supplies us with a unique solution c = 0. The
continuity at the ends of the interval [c. /] is necessary, as we can see
in the following example. Let us take
1(r) =
r. if r ÷ [0. 1)
0. if r = 1
. r ÷ [0. 1].
This function is de…ned on [0. 1]. it is di¤erentiable on (0. 1) and 1(0) =
1(1), but its derivative 1
t
(r) = 1 has no zero on (0. 1).
2. SEQUENCES AND SERIES OF FUNCTIONS 65
2. Sequences and series of functions
We know to measure the length a =
c
2
1
+c
2
2
+c
2
3
of a vector
a = c
1
i + c
2
j + c
3
k of \
3
. the 3dimensional vector space of all free
vectors (here c
1
. c
2
. c
3
÷ R are the coordinates of a). The function a
a . which associates to a vector a its length a . has the following
basic properties:
:1. a = 0. if and only if a = 0.
:2. a +b _ a +b .
for any a. b ÷\
3
.
(2.1) :3. \a = [\[ a for any \ ÷ R and a ÷\
3
.
If instead of \
3
we take any real vector space \ together with a
mapping like above, r ÷ r ÷ [0. ·). r ÷ \. which ful…ls the analo
gous requirements :1. :2 and :3 from (2.1), we get the general notion
of a normed space (\. .).
Definition 13. Let \ be an arbitrary real vector space and let
1 1 be a mapping which associates to any element 1 of \ a
nonnegative real number 1 . If this mapping satis…es the following
properties:
::1. 1 = 0. if and only if 1 = 0. 1 ÷ \.
::2. 1 +o _ 1 +o .
for any 1. o ÷ \ and,
::3. \1 = [\[ 1 for any \ ÷ R and 1 ÷ \.
we say that the pair (\. .) is a normed space and the mapping
r r (the norm of r) is called a norm application (function) or
simply a norm on \.
For instance, the norm of a matrix ¹ = (c
j;
). i = 1. 2. .... :; , =
1. 2. .... :. is
¹ =
a
¸
j=1
n
¸
;=1
c
2
j;
.
66 3. SEQUENCES AND SERIES OF FUNCTIONS
The mapping ¹ ¹ satis…es the properties of a norm (prove it!)
on the vector space of all : : matrices. In addition, one can prove
(not so easy!) that
(2.2) ::4. ¹1 _ ¹ 1
for any two matrices : : and :j respectively.
Remark 11. It is easy to see that a normed space (\. .) is also
a metric space with the induced distance d. where d(r. n) = r ÷n
(prove this!). For instance, ¦r
a
¦ ÷ r if and only if r
a
÷r ÷ 0 as
: ÷·.
If we consider now a bounded function 1 : ¹ ÷ R de…ned on an
arbitrary set ¹ with real values, we can de…ne the norm ("length") of
1 by the formula: 1 = sup [1(¹)[ . where [1(¹)[ = ¦[1(c)[ : c ÷ ¹¦ is
the absolute value of the image of ¹ through 1. or simply the modulus
of the image of 1. This norm is also called the sup:o::.
Theorem 37. Let E(¹) = ¦1 : ¹ ÷ R, 1 bounded¦ be the vec
tor space of all bounded functions de…ned on a …xed set ¹. Then the
mapping 1 1 is a norm on E(¹) with the additional property:
:4. 1o _ 1 o
for any 1. o ÷ E(¹). Moreover, any Cauchy sequence ¦1
a
¦ with respect
to this norm is a convergent sequence in E(¹).
Proof. Let us prove for instance ::2. Since
[1(c) +o(c)[ _ [1(c)[ +[o(c)[ _
_ sup¦[1(c)[ : c ÷ ¹¦ + sup¦[o(c)[ : c ÷ ¹¦.
taking sup on the left side (it exists, because it is upper bounded by a
constant quantity), we get the property :2. : 1 +o _ 1+o . The
property ::4. can be proved in the same manner (do it!). The other
properties are obvious (prove them with all details!). Let us prove the
last statement. Since
[1
a+j
(r) ÷1
a
(r)[ _ sup¦[1
a+j
(r) ÷1
a
(r)[ : r ÷ ¹¦ = 1
a+j
÷1
a
 .
for a …xed r in ¹. the numerical sequence ¦1
a
(r)¦ is a Cauchy sequence
in R. Since R is complete, i.e. any Cauchy sequence in R has a (unique)
limit in R, let us associate to r the limit lim
a÷o
1
a
(r). denoted by 1(r).
i.e. a real number which depends on r. We shall prove that this new
function 1 : ¹ ÷ R :1) is bounded, i.e. belongs to E(¹) and 2) it is
the limit of the sequence ¦1
a
¦ in E(¹). relative to the supnorm. For
2. SEQUENCES AND SERIES OF FUNCTIONS 67
2) let us take a small 0 and let us …nd a rank ` which depends on
such that
(2.3) 1
a+j
÷1
a
 <
for any : _ ` and for any j = 1. 2. .... Since 1
a
(r) ÷ 1(r) for any
…xed r in ¹ and since
[1
a+j
(r) ÷1
a
(r)[ _ 1
a+j
÷1
a
 <
for any : _ ` and any j. let us make j large enough, i.e. j ÷ · in
the last inequality. We get [1(r) ÷1
a
(r)[ _ (why?) for : _ ` and
for any r in ¹. Take now sup on the left and get:
(2.4) 1 ÷1
a
 _
for any : _ `. Hence 1
a
.
÷1 . We make : = ` in (2.4) and write
[1(r)[ _ [1(r) ÷1
.
(r)[ +[1
.
(r)[ _ 1 ÷1
.
 +1
.
 _ +1
.
 .
Take now sup on the left and we get:
1 _ +1
.
 .
i.e. 1 is bounded and so, 1
a
.
÷1 in E(¹) .
Definition 14. Let ¦1
a
¦ be a sequence of bounded functions on ¹
and let 1 be another bounded function on ¹. We say that the sequence
¦1
a
¦ is uniformly convergent to 1 (write 1
a
&c
÷ 1) if the sequence of
numbers ¦1
a
÷1¦ is convergent to 0. If for any …xed r ÷ ¹ the
sequence of numbers ¦1
a
(r)¦ is convergent to 1(r). we say that the
sequence of functions ¦1
a
¦ is simply (or pointwise) convergent to 1
(1
a
cc
÷ 1). Since [1
a
(r) ÷1(r)[ _ 1
a
÷1 . the uniform convergence
implies the simple convergence (why?give details!).
The notion of uniform convergence is stronger then the notion of
simple convergence. For instance, let
1
a
(r) = r
a
. r ÷ [0. 1].
Here ¹ = [0. 1] and, for r ÷ [0. 1). lim
a÷o
1
a
(r) = 0 (why?). For r = 1.
lim
a÷o
1
a
(1) = 1. So, the pointwise limit function 1(r) = 0. if 0 _ r <
1 and 1(1) = 1. Hence, the sequence of functions ¦1
a
¦ is pointwise
convergent to this 1. Let us evaluate now
1
a
÷1 = sup¦[1
a
(r) ÷1(r)[ : r ÷ [0. 1]¦ = 1.
Hence 1
a
÷1 = 1 does not tend to 0! So, the sequence of functions
is not uniformly convergent.
68 3. SEQUENCES AND SERIES OF FUNCTIONS
Remark 12. (Weierstrass) Not always we must compute exactly
the norm 1
a
÷1 . In fact, for the uniform convergence to 1 of the se
quence ¦1
a
¦. it is su¢cient to …nd a sequence of numbers ¦c
a
¦ such that
[1
a
(r) ÷1(r)[ _ c
a
for any r ÷ ¹ and for any : _ ` (a …xed natural
number) such that ¦c
a
¦ ÷0 (why?). For instance, take 1
a
(r) =
sin aa
a
.
Since for any …xed r ÷ R,
sin aa
a
_
1
a
. we have that 1
a
(r) ÷ 0. when
: ÷ ·. But the right side of this last inequality is independent on r.
So we can take c
a
=
1
a
and apply the above remark of Weierstrass.
Hence 1
a
(r) =
sin aa
a
is uniformly convergent to 0 on R. If instead of
sin :r one takes any other bounded function o(r) on an arbitrary in
terval 1 · R, we get that 1
a
(r) =
j(a)
a
is uniformly convergent to 0 on
1 (prove it!).
In order to test the uniform convergence of a sequence of continuous
functions we can use the following result.
Theorem 38. Let (A. d) be a metric space and let ¦1
a
¦ be a uni
formly convergent sequence of bounded continuous functions de…ned on
A with real or complex values. Let 1 be the limit function of ¦1
a
¦.
Then the function 1 itself is a bounded and continuous function on A.
Proof. Recall that 1
a
 = sup [1
a
(A)[ < · for any : = 1. 2. ...
(1
a
is bounded). Let 0 be a small positive real number and let `
be a rank (a …xed natural number) such that
(2.5) 1 ÷1
a
 < for any : _ `.
1) Let us prove that 1 is bounded on A. Take : = ` in (2.5),
remember the basic property of the norm function (see Theorem 37)
and write
1 = (1 ÷1
.
) +1
.
 _ 1 ÷1
.
 +1
.
 < +1
.
 .
Since 1
.
is bounded (1
.
 < ·), we get that 1 is also bounded.
2) In order to prove the continuity of 1 at a …xed point c of A. let
us take a sequence ¦c
I
¦ which is convergent to c. when / ÷·. Since
¦1
a
¦ is uniformly convergent to 1. there is a large number 1 such that
1 ÷1
1
 <
.
3
. Since this 1
1
is continuous, there is a rank 1 such that
for any / _ 1 one has
[1
1
(c
I
) ÷1
1
(c)[ <
3
.
Now,
(2.6) [1(c
I
) ÷1(c)[ = [1(c
I
) ÷1
1
(c
I
) +1
1
(c
I
) ÷1(c)[ _
_ [1(c
I
) ÷1
1
(c
I
)[ +[1
1
(c
I
) ÷1(c)[ _
2. SEQUENCES AND SERIES OF FUNCTIONS 69
_ sup¦[1(r) ÷1
1
(r)[ : r ÷ A¦ +[1
1
(c
I
) ÷1(c)[ =
= 1 ÷1
1
 +[1
1
(c
I
) ÷1(c)[
But,
(2.7) [1
1
(c
I
) ÷1(c)[ = [1
1
(c
I
) ÷1
1
(c) +1
1
(c) ÷1(c)[ _
_ [1
1
(c
I
) ÷1
1
(c)[+[1
1
(c) ÷1(c)[ _
3
+sup¦[1
1
(r) ÷1(r)[ : r ÷ A¦ =
=
3
+1
1
÷1 .
for any / _ 1 (here we just used the continuity of 1
1
). Combining the
inequalities (2.6) and (2.7), we …nd
[1(c
I
) ÷1(c)[ _ 1 ÷1
1
 +
3
+1
1
÷1 _
3
+
3
+
3
= .
for any / _ 1. Hence 1(c
I
) ÷1(c). so 1 is continuous at c.
This last result is useful whenever we want to prove that a sequence
of continuous functions ¦1
a
¦ is NOT uniformly convergent. Namely,
we construct the limit function 1(r) = lim
a÷o
1
a
(r) for any …xed r. If the
function 1(r) is not continuous, then, because of Theorem 38, we must
conclude that ¦1
a
¦ cannot be uniformly convergent to 1.
For instance, the sequence 1
a
(r) = r
a
. r ÷ [0. 1] is convergent to
1(r) = 0 if r ÷ [0. 1) and 1(1) = 1. Since this last function is not
continuous, our sequence cannot be uniformly convergent to 1. It is
only simply convergent to 1.
Sometimes it is useful to integrate term by term a sequence of func
tions and see what happens with the limit function.
Theorem 39. Let ¦1
a
¦ be a sequence of continuous functions,
which is uniformly convergent to a continuous (see Theorem 38) func
tion 1 on the interval [c. /]. For any …xed r ÷ [c. /] one de…nes 1
a
(r) =
a
o
1
a
(t)dt. : = 0. 1. ... and 1(r) =
a
o
1(t)dt be the canonical primi
tives of 1
a
and of 1 respectively on [c. /]. Then, the sequence ¦1
a
¦ is
uniformly convergent to 1 on [c. /]. In particular, for r = /. we get a
very useful relation:
(2.8) lim
a÷o
b
o
1
a
(t)dt =
b
o
lim
a÷o
1
a
(t)dt.
Proof. Let us evaluate
1
a
÷1 = sup¦[1
a
(r) ÷1(r)[ . r ÷ [c. /]¦ _
_ sup¦
a
o
[1
a
(t) ÷1(t)[ dt : r ÷ [c. /]¦ _
70 3. SEQUENCES AND SERIES OF FUNCTIONS
(2.9) _ 1
a
÷1 sup¦
a
o
dt : r ÷ [c. /]¦ = (/ ÷c) 1
a
÷1 .
Now, since ¦1
a
¦ is uniformly convergent to 1. the numerical se
quence 1
a
÷1 tends to zero. Hence, since 2.9 says that
1
a
÷1 _ 1
a
÷1 (/ ÷c).
we have that 1
a
÷1 ÷0. i.e. ¦1
a
¦ is uniformly convergent to 1 on
[c. /].
In the following we show how to use this result in practice.
Let us take the sequence of functions 1
a
(r) = :rc
÷aa
2
. r ÷ [0. 1].
It is clear that this sequence is simply convergent to the continuous
function 1(r) = 0 for any r in [0. 1]. Since 1 is continuous we cannot
decide if our sequence is uniformly convergent or not, only by using
Theorem 38. If the sequence were uniformly convergent, then, using
the relation (2.8) we would get:
(2.10) lim
a÷o
1
0
:rc
÷aa
2
dr =
1
0
lim
a÷o
:rc
÷aa
2
dr = 0.
But
1
0
:rc
÷aa
2
dr = ÷
1
2
c
÷aa
2
[
1
0
= ÷
1
2
[c
÷a
÷1] ÷
1
2
= 0.
Hence, our assumption cannot be true. So, our sequence is not uni
formly convergent on [0. 1].
Remark 13. In Theorem 39 we saw that a uniformly convergent
sequence of continuous functions can be "termwisely" integrated. But
what about their "termwise" derivatives? Can we "termwisely" di¤er
entiate a uniformly convergent sequence of di¤erentiable functions? In
general, we cannot, as the following example shows. Let 1
a
(r) =
a
n
a
.
r ÷ [0. 1]. Since 1
a
÷0 = sup¦
a
n
a
: r ÷ [0. 1]¦ =
1
a
÷ 0. when
: ÷ ·. we …nd that ¦1
a
¦ is uniformly convergent to 1(r) = 0 on
[0. 1]. But 1
t
a
(r) = r
a÷1
is not uniformly convergent on [0. 1] as we saw
above.
Theorem 40. If we want to di¤erentiate "termwisely" the sequence
¦1
a
¦ of di¤erentiable functions on [c. /]. the following conditions are
su¢cient: 1) ¦1
a
¦ is nc to 1 on [c. /]. 2) ¦1
t
a
¦ is uniformly convergent
to o on [c. /] and 3) 1
a
÷ (
1
[c. /] for any : = 0. 1. ... . Then 1 is also
di¤erentiable and 1
t
= o (=1 is also of class (
1
on [c. /]).
2. SEQUENCES AND SERIES OF FUNCTIONS 71
Proof. Indeed, using Theorem 39 for the sequence 1
t
a
&c
÷ o. one
has that
(2.11) 1
a
(r) =
a
o
1
t
a
(t)dt = 1
a
(r) ÷1
a
(c)
&c
÷
a
o
o(t)dt.
Since 1
a
&c
÷ 1 one has that 1(r) ÷1(c) =
a
o
o(t)dt (why?), i.e. 1 is a
primitive of o (see Newton’s formula). Hence, 1
t
= o on [c. /].
Definition 15. Let ¦1
a
¦ be a sequence of functions de…ned on a
subset ¹ of R. For every : = 0. 1. ... we denote by
:
a
(r) = 1
0
(r) +1
1
(r) +... +1
a
(r).
A series of functions 1
a
is an "in…nite" sum
:(r) = lim
a÷o
:
a
(r)
act
=
o
¸
I=0
1
I
(r).
If the sequence of "partial sums" ¦:
a
¦ is simply convergent to the func
tion : on ¹. we say that the series
o
¸
I=0
1
I
(r) is simply (pointwise) con
vergent to : (its sum) on ¹. If the sequence ¦:
a
¦ is uniformly convergent
to : on ¹. we say that the series
o
¸
I=0
1
I
(r) is uniformly convergent to :
(its sum) on ¹. In this last case, we simply write : =
o
¸
I=0
1
I
.
Let the series of functions
o
¸
I=0
r
I
= lim
a÷o
(1 +r +r
2
+... +r
a
) = lim
a÷o
1 ÷r
a+1
1 ÷r
=
1
1 ÷r
.
for any r ÷ (÷1. 1). So, the (geometric) series
o
¸
I=0
r
I
is simply (point
wise) convergent to
1
1÷a
on (÷1. 1). Let us see if it is uniformly conver
gent on (÷1. 1). For this, let us evaluate
:
a
÷: =
1 ÷r
a+1
1 ÷r
÷
1
1 ÷r
=
=
r
a+1
1 ÷r
= sup¦
r
a+1
1 ÷r
: r ÷ (÷1. 1)¦ = ·.
Hence, our series is NOT uniformly convergent on the whole interval
(÷1. 1) but,...it IS uniformly convergent on every closed subinterval
72 3. SEQUENCES AND SERIES OF FUNCTIONS
[c. /] of (÷1. 1). Indeed, in this case, if we denote by c = max¦[c[ . [/[¦,
we get
:
a
÷: _
c
a+1
1 ÷c
÷0. when : ÷·.
because c ÷ (0. 1). Thus the series is uniformly convergent on [c. /].
Sometimes, it is very di¢cult to evaluate "the error function" :
a
÷
:. This is why we need some other tools for deciding if a series is
uniformly convergent or not. A series of functions
o
¸
I=0
1
I
is said to
be absolutely uniformly convergent if the series of the moduli of these
functions
o
¸
I=0
[1
I
[ is uniformly convergent. Recall that [1[ (r)
oc;
= [1(r)[ .
It is not di¢cult to see that an absolutely uniformly convergent series of
functions
o
¸
I=0
1
I
is also uniformly convergent. Indeed, let o
a
=
a
¸
I=0
[1
I
[
and let o =
o
¸
I=0
[1
I
[ be the sum of the series of moduli. Then
[:(r) ÷:
a
(r)[ = [1
a+1
(r) +1
a+2
(r) +...[ _ [1
a+1
(r)[ +[1
a+2
(r)[ +...
(why?)
= o(r) ÷o
a
(r) _ sup¦[o(r) ÷o
a
(r)[ : r ÷ ¹¦ = o ÷o
a
 .
Hence [:(r) ÷:
a
(r)[ _ o ÷o
a
 for any r ÷ ¹. Taking now sup on
r ÷ ¹ we get that :
a
÷: _ o ÷o
a
 . Since our series is absolutely
uniformly convergent, then o ÷o
a
 ÷ 0. when : ÷ ·. Using now
the last inequality, we get that :
a
÷: ÷ 0. i.e. the initial series
is uniformly convergent. A powerful and useful test for the absolute
uniform convergence is the following test.
Theorem 41. (Weierstrass Test for series of functions) Let ¹ be a
subset of real numbers and let
o
¸
I=0
1
I
be a series of functions de…ned on
¹. Assume that 1
a
 can be upper bounded by c
a
÷ [0. ·) ([1
a
(r)[ _
c
a
where r runs on ¹) for any : = 0. 1. ... and that the numerical
series
o
¸
I=0
c
I
is convergent. Then the series
o
¸
I=0
1
I
is absolutely uniformly
convergent. In particular, it is also uniformly convergent.
Proof. Let us …x a small positive real number 0 and an r ÷ ¹.
Let
o
a
= [1
0
[ +[1
1
[ +... +[1
a
[
2. SEQUENCES AND SERIES OF FUNCTIONS 73
be the :th partial sum of the series
o
¸
I=0
[1
I
[. Since the numerical series
o
¸
I=0
c
I
is convergent, there is a rank ` such that
c
a+1
+c
a+2
+... +c
a+j
<
for any : _ ` and for any natural number j.
Let us evaluate [o
a+j
(r) ÷o
a
(r)[ :
(2.12) [o
a+j
(r) ÷o
a
(r)[ = [1
a+1
(r)[ +[1
a+2
(r)[ +... +[1
a+j
(r)[ _
c
a+1
+c
a+2
+... +c
a+j
< .
From (2.12) we obtain that the sequence ¦o
a
(r)¦ is a Cauchy se
quence of real numbers (see De…nition 2). Since on the real line any
Cauchy sequence is convergent (see Theorem 13) we get that the se
quence ¦o
a
(r)¦ is convergent to a real number o(r) (this means that
this real number depends on r. i.e. it is changing if we change r. so it is
a function of r). Come back now in (2.12) and make j ÷·. We …nd
that [o(r) ÷o
a
(r)[ < for any : _ ` and for ANY r ÷ ¹. If here,
in the last inequality, we take sup on r. we …nally get: o ÷o
a
 <
for any : _ `. Hence, the series
o
¸
I=0
[1
I
[ is uniformly convergent to
o (its sum). Thus, our initial series
o
¸
I=0
1
I
is uniformly and absolutely
convergent.
The series of functions
o
¸
a=1
arctan(aa)
a
2
is absolutely uniformly conver
gent because
arctan(aa)
a
2
_
¬
2
1
a
2
and the numerical series
o
¸
a=1
¬
2
1
a
2
=
¬
2
o
¸
a=1
1
a
2
is convergent (why?) (see the Weierstrass Test, Theorem 41).
Another very useful test is the AbelDirichlet Test for series of func
tions, a generalization of the test with the same name for numerical
series.
Theorem 42. (AbelDirichlet Test for series of functions)
Let ¦c
a
(r)¦. ¦/
a
(r)¦ be two sequences of functions de…ned on the
same interval 1 of R. We assume that c
a
 is a decreasing to zero
sequence and that the partial sums :
a
(r) =
¸
a
I=0
/
a
(r) of the series
of functions
¸
o
I=c
/
a
(r) are uniformly bounded, i.e. there is a positive
real number ` 0 such that :
a
 < ` for any : = 1. 2. ....
Then the series of functions
¸
o
a=0
c
a
(r)/
a
(r) is (absolutely) uni
formly convergent on the interval 1.
74 3. SEQUENCES AND SERIES OF FUNCTIONS
Proof. Let us come back to the AbelDirichlet’s Test for numerical
series and substitute the numbers c
a
. /
a
. :
a
. o
a
with the corresponding
functions c
a
(r). /
a
(r). :
a
(r) and o
a
(r) =
¸
a
I=0
c
I
(r)/
I
(r) respectively.
We obtain (do it step by step!) that the sequence of functions ¦o
a
(r)¦
is uniformly Cauchy, i.e. for any 0. there is a rank `
.
such that if
: _ `
.
one has that
(2.13) o
a+j
÷o
a
 <
for any j = 1. 2. .... In particular,
[o
a+j
(r) ÷o
a
(r)[ <
for any …xed r in 1. So, the numerical sequence ¦o
a
(r)¦ is convergent
to a number o(r) which depend on r. Making j ÷· in (2.13) we get
[o(r) ÷o
a
(r)[ <
for any : _ `
.
and for any r in 1. Take now sup on r and …nd that
o ÷o
a
 <
for any : _ `
.
. This means that ¦o
a
¦ is uniformly convergent to o.
i.e. our series of functions
¸
o
a=0
c
a
(r)/
a
(r) is uniformly convergent on
the interval 1. With some small changes in the proof, we …nd that this
last series is absolutely uniformly convergent on 1 (do them!).
Let us take the series of functions o(r) =
¸
o
a=1
(÷1)
n1
a
r
a
for r ÷
[÷1 + . 1]. where 0 < < 2. Let us apply the AbelDirichlet Test for
series of functions by taking c
a
(r) =
a
n
a
and /
a
(r) = (÷1)
a÷1
. We easily
see that c
a
(r) =
1
a
and that the series
¸
o
a=1
(÷1)
a÷1
has bounded
partial sums. Hence our series o(r) =
¸
o
a=1
(÷1)
n1
a
r
a
. r ÷ [÷1 +. 1].
is absolutely and uniformly convergent.
The following question arises: can we integrate or di¤erentiate term
by term (termwise) a series of function
o
¸
I=0
1
I
? Since everything reduces
to the sequence of partial sums :
a
= 1
0
+ 1
1
+ ... + 1
a
. we can apply
the results from Theorem 39 and Theorem 40 and …nd:
Theorem 43. Let : =
o
¸
a=0
1
a
be a uniformly convergent series of
continuous functions on the interval [c. /] and let 1
a
(r) be their canon
ical primitives on [c. /] : 1
a
(r) =
a
o
1
a
(t)dt. : = 0. 1. ... . Then the
series of functions o =
o
¸
a=0
1
a
is uniformly convergent on [c. /] and
2. SEQUENCES AND SERIES OF FUNCTIONS 75
o(r) =
a
o
:(t)dt. or
(2.14)
a
o
o
¸
a=0
1
a
(t)
dt =
o
¸
a=0
a
o
1
a
(t)dt.
(this means that the integration symbol
commute with the symbol
¸
of a series). In particular, for r = /. we get a very useful formula:
(2.15)
b
o
o
¸
a=0
1
a
(t)
dt =
o
¸
a=0
b
o
1
a
(t)dt.
If in addition, 1
a
are functions of class (
1
on [c. /] (1
a
are dif
ferentiable and their derivatives are continuous on [c. /]. shortly write
1
a
÷ (
1
[c. /]) and if the series of derivatives, n =
o
¸
a=0
1
t
a
is uniformly
convergent on [c. /]. then : is di¤erentiable on [c. /] and :
t
= n. So,
we can di¤erentiate "term by term" (or termwise) the initial series of
functions.
In the …rst statement : is a continuous function on [c. /] because of
the basic Theorem 38. In this last theorem there is a requirement: 1
a
must be bounded. This is true because 1
I
are continuous and de…ned
on a bounded and closed interval (see Theorem 32).
Let us study the following series of functions
o
¸
a=0
(÷1)
a
r
a
on (÷1. 1).
For any …xed r, one has the formula
(2.16) 1 ÷r +r
2
÷... =
1
1 +r
. r ÷ (÷1. 1).
the famous geometric series with ratio ÷r. Hence, our series is simply
convergent on (÷1. 1). It is not uniformly convergent on (÷1. 1) but it is
absolutely and uniformly convergent on any closed subinterval [c. /] of
(÷1. 1) (apply the same reason as in the case of the in…nite geometrical
series). Let us derive an interesting and useful formula from (2.16). Let
us …x an r
0
in (÷1. 1) and take c. / such that r
0
÷ [c. /]. c or / is 0
(if r
0
< 0. take / = 0. if r
0
_ 0. take c = 0) and [c. /] is included in
(÷1. 1). Since all conditions in Theorem 43 are ful…lled, we integrate
term by term formula (2.16) and get
a
0
0
(1 ÷t +t
2
÷... + (÷1)
a
t
a
+...)dt =
= (t ÷
t
2
2
+
t
3
3
÷... + (÷1)
a
t
a+1
: + 1
+...) [
a
0
0
=
76 3. SEQUENCES AND SERIES OF FUNCTIONS
=
o
¸
a=1
(÷1)
a÷1
r
a
0
:
=
a
0
0
1
1 +t
dt = ln(1 +r
0
).
Now, let us put instead of r
0
an arbitrary r in (÷1. 1) and obtain
(2.17) ln(1 +r) =
o
¸
a=1
(÷1)
a÷1
r
a
:
, for any r ÷ (÷1. 1).
The value of the alternate series
¸
o
a=1
(÷1)
a÷1 1
a
is ln 2 but, to prove
this, one needs the continuity of the function on the right in the formula
2.17. And this is not so easy to be proved (see the Abel Theorem,
Theorem 46).
Let us compute the sum of the series of functions
¸
o
a=0
:r
a
on
its maximal domain of de…nition. First of all, let us …x an r on the
real line and try to …nd conditions for the convergence of the series
¸
o
a=0
:r
a
. Let us see where the series (numerical series this time!) is
absolutely convergent. Applying the Ratio Test (Theorem 27) to the
series of moduli
¸
o
a=0
:[r[
a
. we get lim
a÷o
o
n+1
on
= [r[ . We know that if
[r[ < 1. the series is absolutely convergent, in particular it is convergent
on (÷1. 1). If [r[ 1. the series is divergent, because, in this case, the
sequence ¦:r
a
¦ is not bounded (why?) so, it cannot be convergent to
0. For r = 1 or r = ÷1. the series is divergent. Hence, the de…nition
domain of the function :(r) =
¸
o
a=0
:r
a
is exactly (÷1. 1). Let us
compute :(r).
:(r) = 1r+2r
2
+3r
3
+.... +:r
a
+... = r(1+2r+3r
2
+... +:r
a÷1
+...)
= r(r +r
2
+... +r
a
+...)
t
= r
r
1 ÷r
t
=
r
(1 ÷r)
2
.
Here we used Theorem 43 to di¤erentiate term by term the series
r + r
2
+ ... + r
a
+ ... =
a
1÷a
(why the hypotheses of this theorem are
ful…lled?).
3. Problems
1. Find the convergence set and the limit for the following sequences
of functions: a) 1
a
(r) = r
a
; b) 1
a
(r) =
a
a
; c) 1
a
(r) =
a
a+a
. r ÷ (0. ·);
d) 1
a
(r) =
aa
1+a+a
. r ÷ [0. 1]; e) 1
a
(r) =
2aa
1+a
2
a
2
. r ÷ [1. ·); f) 1
a
(r) =
a
2
a
4
+a
2
. r ÷ [1. ·).
3. PROBLEMS 77
2. Say if the convergence of the above sequences (see Problem 1.)
is uniform or not. Study the absolute uniform convergence of the same
sequences.
3. Let 1
a
(r) =
aa
1+a
2
a
2
. r ÷ [0. 1]. Prove that ¦1
a
¦ is not uniformly
convergent but
1
0
1
a
(r)dr ÷
1
0
lim
a÷o
1
a
(r)dr.
4. Prove that 1
a
(r) =
a
1+a
2
a
2
. r ÷ [÷1. 1] is uniformly convergent
to 1(r) (…nd it!) but 1
t
a
is not uniformly convergent to 1
t
. Do the same
for 1
a
(r) =
a
n
a
. r ÷ [0. 1].
5. Prove that the series of functions
¸
o
a=1
(r
a
÷r
a÷1
) is uniformly
convergent on [0. 0.5]. but not on [0. 1].
6. Is the series of functions
¸
o
a=1
sin
a
a+1
÷sin
a
a
uniformly con
vergent on R? But on [0. 1]? But on [c. /]?
7. Prove that the following series of functions are absolutely and
uniformly convergent on the indicated domain: a)
¸
o
a=1
(÷1)
n+1
a
2
+a
a
. r ÷ R;
b)
¸
o
a=1
(÷1)
n
3
nx
a+2
n
. r ÷ [0. ·); c)
¸
o
a=1
sin aa
a
a
. r ÷ R; d)
¸
o
a=1
1
a
2
+a
2
. r ÷
R; e)
¸
o
a=1
sin aa
a
2
+a
4
. r ÷ R.
8. Can we di¤erentiate term by term the following series?
a)
¸
o
a=1
exp(÷:r) sin :r. r ÷ [1. ·); b)
¸
o
a=1
sin(2
p
n
a)
a
2
2
p
n
. r ÷ R;
c)
¸
o
a=1
1
a
2
+a
2
. r ÷ R.
9. Find the image of the following functions:
a) 1(r) = ÷3r + 2. r ÷ [÷3. 12];
b) 1(r) = 2r
2
+r ÷5. r ÷ R;
c) 1(r) = r
3
÷3r + 2. r ÷ [÷120. 120];
d) 1(r) = 3 sin 4r. r ÷ [÷
¬
2
.
¬
2
];
e) 1(r) = [sin r ÷cos 2r[ . r ÷ [0. :];
f) 1(r) = [r
2
+ 2r ÷1[ ÷3. r ÷ (÷·. 9].
10. Find the norm of the following functions: a) 1(r) = 2r ÷ 5.
r ÷ [÷4. 7]; b) 1(r) = 3 cos 5r. r ÷ [:. ·); c) 1(r) = ln(2r
2
+ 3).
r ÷ [÷2. 2]; d) 1 ÷o , where 1(r) = 3r and o(r) = 4r
2
. r ÷ [0. 2].
CHAPTER 4
Taylor series
1. Taylor formula
Always the most elementary functions were considered to be the
polynomial functions. A polynomial function of degree : is a function
de…ned on the whole real line by the formula:
1
a
(r) = c
0
+c
1
r +c
2
r
2
+... +c
a
r
a
.
where c
0
. c
1
. .... c
a
are …xed real numbers and c
a
= 0.
Many mathematicians tried and are trying to reduce the study of
more complicated functions to polynomials.
It is clear enough that not all functions can be represented by a
polynomial. For instance, the exponential function 1(r) = exp(r) = c
a
cannot be represented by a polynomial 1
a
(r). Indeed, if
exp(r) = c
0
+c
1
r +c
2
r
2
+... +c
a
r
a
for r ÷ (c. /). c = /. we di¤erentiate : times and …nd: exp(r) = :!c
a
,
a constant, which is not possible, because the exponential function is
strictly increasing. Here we proved in fact that the exponential function
cannot be represented by a polynomial in any small neighborhood of
any point on the real line. The following problem appears in many
applications. If r is very close to a …xed number c. i.e. if the di¤erence
r ÷c is very small (is very close to zero!), can we represent a function
1 as an "in…nite" polynomial in the variable r ÷c? This means
(1.1) 1(r) = c
0
+c
1
(r ÷c) +c
2
(r ÷c)
2
+...
in a neighborhood (c÷. c+) of c. This would imply that our function
is a function of class (
o
, i.e. it has derivatives of any order. But this
is not true for all functions. So, what can we hope is to "approximate"
a function 1 in a small neighborhood of a point c with a polynomial of
a given degree : in the variable r ÷c :
(1.2) 1(r) = c
0
+c
1
(r ÷c) +c
2
(r ÷c)
2
+... +c
a
(r ÷c)
a
+1
a
(r).
where 1
a
(r) is a remainder which is a function of r (it also depends
on 1 and on c!). This remainder is the error committed when we
79
80 4. TAYLOR SERIES
approximate 1(r) by the polynomial
c
0
+c
1
(r ÷c) +c
2
(r ÷c)
2
+... +c
a
(r ÷c)
a
.
This polynomial is called the Taylor polynomial of order : at c.
If 1(r) is a polynomial of degree :. we can represent 1 as in formula
(1.2) with the remainder zero. Indeed, the set of : + 1 binomials
¦1. r ÷c. (r ÷c)
2
. (r ÷c)
3
. .... (r ÷c)
a
¦
is linear independent in the vector space {
a
of all polynomials of degree
at most :. which has dimension : + 1 over the real …eld (this comes
directly from the de…nition of a polynomialwhy?). Hence,
¦1. r ÷c. (r ÷c)
2
. (r ÷c)
3
. .... (r ÷c)
a
¦
is a basis in {
a
and so, we always can uniquely …nd the constant ele
ments c
0
. c
1
. c
2
. .... c
a
such that
(1.3) 1(r) = c
0
+c
1
(r ÷c) +c
2
(r ÷c)
2
+... +c
a
(r ÷c)
a
.
In this last case we can compute the coe¢cients c
0
. c
1
. .... c
a
by
using the values of 1 and of its derivatives 1
t
. 1
tt
. .... 1
(a)
at c. Indeed,
let us make r = c in the equality (1.3). We get 1(c) = c
0
. If one
di¤erentiates the same equality and makes r = c. one obtains 1
t
(c) =
c
1
. Now, if we di¤erentiate twice this equality (1.3), we get 1
tt
(c) = 2c
2
.
and so on. Take the /th derivative in both sides in (1.3) and …nd
1
(I)
(c) = /!c
I
for any / = 1. 2. .... :. Thus (1.3) becomes:
(1.4)
1(r) = 1(c) +
1
t
(c)
1!
(r ÷c) +
1
tt
(c)
2!
(r ÷c)
2
+... +
1
(a)
(c)
:!
(r ÷c)
a
.
Generally, if the function 1 is not a polynomial of degree :. we
formally can write (it is clear that 1 must be :times di¤erentiable):
(1.5)
1(r) = 1(c)+
1
t
(c)
1!
(r ÷c)+
1
tt
(c)
2!
(r ÷c)
2
+...+
1
(a)
(c)
:!
(r÷c)
a
+1
a
(r).
where
1
a
(r) = 1(r)÷1(c)÷
1
t
(c)
1!
(r ÷c)÷
1
tt
(c)
2!
(r ÷c)
2
÷...÷
1
(a)
(c)
:!
(r÷c)
a
.
The problem is to estimate this remainder. The famous Taylor formula
gives a general estimation for this remainder.
Theorem 44. (Taylor formula) Let ¹ be an open subset of R and
let
1 : ¹ ÷ R be a function de…ned on ¹ with values in R, which
is (: + 1)times di¤erentiable on ¹. Let us …x a point c in ¹ and a
1. TAYLOR FORMULA 81
natural number j = 0. Then, for any r ÷ ¹ such that the segment [c. r]
is included in ¹. there is a point c on this segment with the following
property: the remainder 1
a
(r) from (1.5) has a representation of the
form
(1.6) 1
a
(r) =
r ÷c
r ÷c
j
(r ÷c)
a+1
:!j
1
(a+1)
(c)
This general form of the remainder was discovered by Schömlich. If
j = : + 1. we …nd the Lagrange form of the remainder
(1.7) 1
a
(r) =
1
(a+1)
(c)
(: + 1)!
(r ÷c)
a+1
.
We see that this form is very similar to the general term form in (1.5).
In fact, it is "the next" term after the :th term
;
(n)
(o)
a!
(r÷c)
a
in which
the value of 1
(a+1)
is not computed at c. but at a close point c ÷ [c. r]
(here we do not mean that c is less then r!). Usually, the error made
by approximating 1(r) with its Taylor polynomial 1
a
(r) of order :.
(1.8)
1
a
(r) = 1(c) +
1
t
(c)
1!
(r ÷c) +
1
tt
(c)
2!
(r ÷c)
2
+... +
1
(a)
(c)
:!
(r ÷c)
a
.
is evaluated by the Lagrange form of the remainder 1
a
(r). Since we
have no supplementary information on the number c. we use the fol
lowing upper bounded formula:
(1.9) [1
a
(r)[ _
[r ÷c[
a+1
(: + 1)!
sup¦
1
(a+1)
(.)
: . ÷ [c. r]¦
Since we frequently use Taylor formula with Lagrange remainder, we
write it here in a complete form (together with this last form of the
reminder)
(1.10)
1(r) = 1(c) +
1
t
(c)
1!
(r ÷c) +
1
tt
(c)
2!
(r ÷c)
2
+... +
1
(a)
(c)
:!
(r ÷c)
a
+
1
(a+1)
(c)
(: + 1)!
(r ÷c)
a+1
.
Proof. The proof of this theorem is not so natural. Let us assume
that r c. In this case, the segment [c. r] is exactly the closed interval
[c. r]. Let us denote in (1.5)
(1.11) ((r) =
1
a
(r)
(r ÷c)
j
.
82 4. TAYLOR SERIES
Thus, the formula (1.5) becomes:
(1.12)
1(r) = 1(c) +
1
t
(c)
1!
(r ÷c) +
1
tt
(c)
2!
(r ÷c)
2
+... +
1
(a)
(c)
:!
(r ÷c)
a
+(r ÷c)
j
((r).
In order to obtain a representation for ((r). we consider an auxiliary
function:
(1.13)
o(t) = 1(t)+
1
t
(t)
1!
(r ÷t)+
1
tt
(t)
2!
(r ÷t)
2
+...+
1
(a)
(t)
:!
(r÷t)
a
+(r÷t)
j
((r)
We obtained the expression of o(t) by simply putting t instead of c.
in (1.12). We apply now the Rolle’s Theorem (Theorem 36) on the
interval [c. r]. The function o(t) is continuous and di¤erentiable on
[c. r], o(c) = 1(r) (see 1.12) and o(r) = 1(r) so, o(c) = o(r). Thus,
there is a point c ÷ (c. r) such that o
t
(c) = 0. Let us compute o
t
(t) :
o
t
(t) = 1
t
(t) +
1
tt
(t)
1!
(r ÷t) ÷
1
t
(t)
1!
+
1
ttt
(t)
2!
(r ÷t)
2
÷
1
tt
(t)
1!
(r ÷t) +...
+
1
(a+1)
(t)
:!
(r ÷t)
a
÷
1
(a)
(t)
(: ÷1)!
(r ÷t)
a÷1
÷j(r ÷t)
j÷1
((r).
So we get
(1.14) o
t
(t) =
1
(a+1)
(t)
:!
(r ÷t)
a
÷j(r ÷t)
j÷1
((r).
Make now t = c in (1.14) and …nd
0 = o
t
(c) =
1
(a+1)
(c)
:!
(r ÷c)
a
÷j(r ÷c)
j÷1
((r).
If here, instead of ((r) we put
1n(a)
(a÷o)
p
(see (1.11)), we get
1
(a+1)
(c)
:!
(r ÷c)
a
= j(r ÷c)
j÷1
1
a
(r)
(r ÷c)
j
.
or
1
a
(r) =
(r ÷c)
j
(r ÷c)
j÷1
1
(a+1)
(c)
:!j
(r ÷c)
a
=
(r ÷c)
j
(r ÷c)
j
1
(a+1)
(c)
:!j
(r ÷c)
a+1
.
i.e. formula (1.6). The other statements of the theorem are easily
deduced from this last formula.
Remark 14. A function 1(r) is a zero of another function o(r)
at a point c if lim
a÷o
;(a)
j(a)
= 0. We write this as 1(r) = 0(o(r)) at c.
1. TAYLOR FORMULA 83
For instance, from (1.7) we see that the remainder 1
a
(r) is a zero of
(r ÷c)
a
at r = c. i.e. 1
a
(r) = 0((r ÷c)
a
) at r = c.
If c = 0, the formula (1.5) is called the Mac Laurin formula:
(1.15) 1(r) = 1(0) +
1
t
(0)
1!
r +
1
tt
(0)
2!
r
2
+... +
1
(a)
(0)
:!
r
a
+1
a
(r)
If we use the Lagrange form of the remainder (1.7), we get
(1.16) 1(r) = 1(0)+
1
t
(0)
1!
r+
1
tt
(0)
2!
r
2
+...+
1
(a)
(0)
:!
r
a
+
1
(a+1)
(c)
(: + 1)!
r
a+1
.
where c is a real number between 0 and r. Since it is easier to ma
nipulate Mac Laurin formulas for many functions which are de…ned on
an interval (c. /) with 0 ÷ (c. /) and since the translation r ÷ r ÷ c
makes connections between Taylor formulas and Mac Laurin formu
las, we prefer to deduce these last formulas for the basic elementary
functions.
Example 2. (exp(r)) Let 1(r) = exp(r) = c
a
. r ÷ R. Since the
derivatives of exp(r) is exp(r) itself, the Taylor formula at c = 0 (Mac
Laurin formula) for exp(r) becomes
(1.17) exp(r) = 1 +
r
1!
+
r
2
2!
+... +
r
a
:!
+ exp(c)
r
a+1
(: + 1)!
.
where c ÷ (0. r). if r 0. or to (r. 0). if r < 0.
For instance, let us compute exp(0.03) with 2 exact decimals. Since
c ÷ (0. 0.03). this means that
[1
a
(0.03)[ =
exp(c)
(0.03)
a+1
(: + 1)!
< 3
(0.03)
a+1
(: + 1)!
<
1
100
.
or
3
a+2
100
a+1
(: + 1)!
<
1
100
=3
a+2
< 100
a
(: + 1)!.
It is easy to prove this last inequality by mathematical induction for
: _ 1. So, exp(r)
~
= 1 +
0.03
1!
= 1.03. with 2 exact decimals. This is the
method which computers use to (approximately) calculate exp(:) for a
given real number :. Formula (1.17) can also be written as
(1.18) exp(r) = 1 +
r
1!
+
r
2
2!
+... +
r
a
:!
+ 0(r
a
)
84 4. TAYLOR SERIES
We can use this formula to compute nondeterministic limits. For in
stance, let us compute
lim
a÷0
exp(r
3
) ÷1 ÷r
3
÷
a
6
2
exp(r
2
) ÷1 ÷r
2
÷
a
4
2
=
0
0
.
In formula (1.18) we put instead of r. r
3
and : = 2 :
exp(r
3
) = 1 +r
3
+
r
6
2
+ 0(r
6
).
If we put now in (1.18) instead of r. r
2
and : = 3. we get
exp(r
2
) = 1 +r
2
+
r
4
2
+
r
6
6
+ 0(r
6
).
Hence, our limit becomes
lim
a÷0
0(r
6
)
a
6
6
+ 0(r
6
)
= lim
a÷0
0(a
6
)
a
6
1
6
+
0(a
6
)
a
6
=
lim
a÷0
0(a
6
)
a
6
1
6
+ lim
a÷0
0(a
6
)
a
6
=
0
1
6
+ 0
= 0.
In practice, we do not know in advance how many terms we must con
sider in numerator and in denominator such that the nondeterministic
to be eliminated. So, it is a good idea to consider one or two terms
more than the degree of the polynomial queue which induces the non
deterministic. In our example we write
lim
a÷0
exp(r
3
) ÷1 ÷r
3
÷
a
6
2
exp(r
2
) ÷1 ÷r
2
÷
a
4
2
=
= lim
a÷0
(1 +
a
3
1!
+
a
6
2!
+
a
9
3!
+...) ÷1 ÷r
3
÷
a
6
2
(1 +
a
2
1!
+
a
4
2!
+
a
6
3!
+
a
8
4!
+...) ÷1 ÷r
2
÷
a
4
2
=
= lim
a÷0
a
9
3!
+...
a
6
3!
+
a
8
4!
+...
= lim
a÷0
a
3
3!
+...
1
3!
+
a
2
4!
+...
=
0
1
3!
= 0.
Example 3. (sin(r)) Let 1(r) = sin(r). r ÷ R. Since [sin(r)]
t
=
cos(r). [sin(r)]
tt
= ÷sin(r). [sin(r)]
ttt
= ÷cos(r) and [sin(r)]
(4)
=
sin(r). we obtain that [sin(r)]
(4I+1)
= cos(r). [sin(r)]
(4I+2)
= ÷sin(r).
[sin(r)]
(4I+3)
= ÷cos(r) and [sin(r)]
(4I)
= sin(r) for any / = 0. 1. ... .
Now, sin 0 = 0. cos 0 = 1 and, applying formula (1.16), we get
(1.19) sin(r) =
r
1!
÷
r
3
3!
+
r
5
5!
÷... + (÷1)
a
r
2a+1
(2: + 1)!
+ 0(r
2a+1
).
It is more complicated to express the remainder in this case because the
(: + 1)derivative of sin(r) is either ±sin(r) or ±cos(r). Let us use
the Mac Laurin formula for sin(r) in order to compute sin(0.2) with
1. TAYLOR FORMULA 85
one exact decimal. Here 0.2 means 0.2 radians. Now, the modulus of
the remainder, [1
2a+1
(r)[ is less or equal to
1
(2a+2)!
[r[
2a+2
. So,
[1
2a+1
(0.2)[ _
1
(2: + 2)!
(0.2)
2a+2
.
and this last one must be less then
1
10
. i.e.
1
(2: + 2)!
2
2a+2
< 10
2a+2
or
2
2a+2
< (2: + 2)!10
2a+2
.
But this last one is true for any : _ 0. Hence, sin(0.2) · 0.2 with one
exact decimal.
Example 4. (cos(r)) Let 1(r) = cos(r). r ÷ R. Like in Example
3 we easily deduce the following formula
(1.20) cos(r) = 1 ÷
r
2
2!
+
r
4
4!
÷
r
6
6!
+... + (÷1)
a
r
2a
(2:)!
+ 0(r
2a
).
Example 5. Let
1(r) = ln(1 +r). r ÷ (÷1. ·).
Since
1
t
(r) = (1 +r)
÷1
. 1
tt
(r) = ÷(1 +r)
÷2
. 1
ttt
(r) = 2(1 +r)
÷3
. ...
.... 1
(a)
(r) = (÷1)
a÷1
(: ÷1)!(1 +r)
÷a
. ....
one has that 1(0) = 0. 1
t
(0) = 1. 1
tt
(0) = ÷1. 1
ttt
(0) = 2. .... 1
(a)
(0) =
(÷1)
a÷1
(: ÷1)!. ... . So, the formula (1.16) becomes
(1.21)
ln(1+r) = r÷
r
2
2
+
r
3
3
÷
r
4
4
+... +(÷1)
a÷1
r
a
:
+(÷1)
a
(1 +c)
÷a÷1
: + 1
r
a+1
.
where c is a real number between 0 and r. Hence,
(1.22) ln(1 +r) = r ÷
r
2
2
+
r
3
3
÷
r
4
4
+... + (÷1)
a÷1
r
a
:
+ 0(r
a
).
Let us compute ln(1.02) with 3 exact decimals. Since
ln(1.02) = ln(1 + 0.02) = 0.02 ÷
(0.02)
2
2
+
(0.02)
3
3
+...
+(÷1)
a÷1
(0.02)
a
:
+ (÷1)
a
(1 +c)
÷a÷1
: + 1
(0.02)
a+1
.
86 4. TAYLOR SERIES
where c is between 0 and 0.02. we must evaluate the modulus of the
remainder and force this last upper bound to be less then
1
1000
.
(÷1)
a
(1 +c)
÷a÷1
: + 1
0.02
a+1
<
2
a+1
(: + 1)100
a+1
<
1
1000
.
This last inequality is true for any : _ 1. Thus, ln(1.02) · 0.020 with
3 exact decimals. Pay attention! It is not sure that 020 are the …rst
three decimals of ln(1.02)! What is sure is that [ln(1.02) ÷0.02[ is less
then 0.001 =
1
1000
(this means "with 3 exact decimals!").
Example 6. (Binomial formula) Let 1(r) = (1 + r)
c
. where c is
a …xed real number and r ÷1. Since
1
t
(r) = c(1 +r)
c÷1
. 1
tt
(r) = c(c ÷1)(1 +r)
c÷2
. ...
.... 1
(a)
(r) = c(c ÷1)(c ÷2)...(c ÷: + 1)(1 +r)
c÷a
. ....
one has that
1(0) = 1. 1
t
(0) = c. 1
tt
(0) = c(c ÷1). ...
.... 1
(a)
(0) = c(c ÷1)(c ÷2)...(c ÷: + 1). ....
Now, formula (1.16) becomes
(1 +r)
c
= 1 +
c
1!
r +
c(c ÷1)
2!
r
2
+...
... +
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
r
a
+
(1.23) +
c(c ÷1)(c ÷2)...(c ÷:)(1 + c)
c÷a÷1
(: + 1)!
r
a+1
.
where c is a real number between 0 and r.
Formula (1.23) can also be written as
(1.24)
(1+r)
c
= 1+
c
1!
r+
c(c ÷1)
2!
r
2
+...+
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
r
a
+0(r
a
)
Let us use this formula to approximate the following expression
1 = 1(¡) =
1
o+bo
2
, c. / 0. by a polynomial of degree 2 (it is used
in Physics for ¡ small). In order to apply (1.23) we need to put our
expression in the form (1 +r)
c
. So,
1 = (c +/¡
2
)
÷
1
2
= c
÷
1
2
(1 +
/
c
¡
2
)
÷
1
2
.
1. TAYLOR FORMULA 87
Let us take only (1 +
b
o
¡
2
)
÷
1
2
and use (1.23) up to r
2
. where r =
b
o
¡
2
and c = ÷
1
2
. We get
(1 +
/
c
¡
2
)
÷
1
2
 1 + (÷
1
2
)
/
c
¡
2
+
(÷
1
2
)(÷
3
2
)
2
/
2
c
2
¡
4
.
Hence,
1
c +/¡
2

1
c
÷
/
2c
c
¡
2
+
3/
2
8c
2
c
¡
4
.
If c = :. a natural number, we obtain the famous binomial formula
of Newton:
(1.25) (1 +r)
a
= 1 +
:
1!
r +
:(: ÷1)
2!
r
2
+... +
:(: ÷1)(: ÷2)...1
:!
r
a
.
because the remainder in (1.23) is zero. If instead of r we put
b
o
in
(1.25) we get
(c +/)
a
c
a
= 1 +
:
1
/
c
+
:
2
/
2
c
2
+
:
3
/
3
c
3
+... +
:
:
/
a
c
a
.
Multiplying by c
a
. we get:
(1.26)
(c +/)
a
= c
a
+
:
1
c
a÷1
/ +
:
2
c
a÷2
/
2
+
:
3
c
a÷3
/
3
+... +
:
:
/
a
.
Here,
a
I
=
a(a÷1)(a÷2)...(a÷I+1)
I!
=
a!
I!(a÷I)!
means : objects taken /.
Example 7. The equilibrium position of a homogeneous weighted
string, …xed at the ends, has a form given by the plane curve n =
c c/(
a
b
). where c/(r) =
exp(a)+exp(÷a)
2
and c. / are real numbers (see
Fig.4.1). The function 1(r) = c/(r) is called the hyperbolic cosine of
r.
Fig. 4.1
A B
O
x
y
88 4. TAYLOR SERIES
The derivative of the function c/(r) is :/(r) =
exp(a)÷exp(÷a)
2
. called
the hyperbolic sine of r. Since the derivative of each of them is the other
one, we easily get the formulas
(1.27) :/(r) =
r
1!
+
r
3
3!
+
r
5
5!
+... +
r
2a+1
(2: + 1)!
+ 0(r
2a+1
).
(1.28) c/(r) = 1 +
r
2
2!
+
r
4
4!
+
r
6
6!
+... +
r
2a
(2:)!
+ 0(r
2a
).
For instance, for r small enough, we can approximate c/(r) by the
polynomial 1
4
(r) = 1 +
a
2
2!
+
a
4
4!
. For r = 0.5. c/(0.5)  1 +
0.25
2
+
0.0025
24
.
Taylor’s and Mac Laurin’s formulas have many applications in the
local study of a function (or a curve).
Corollary 4. (Lagrange formula) Let us write Taylor formula
(1.10) for : = 1 : 1(r) = 1(c) + 1
t
(c) (r ÷ c). where c is a number
between c and r. If r = / c. we get the classical Lagrange formula:
1(/) = 1(c) +1
t
(c) (/ ÷c). where c ÷ (c. /).
Remark 15. We can use Taylor formula (1.10) for study the shape
of a function in a neighborhood of a point c. Suppose that
1
t
(c) = 1
tt
(c) = ... = 1
(a÷1)
(c) = 0
and 1
(a)
(c) = 0. We also assume that 1 is of class (
a
on an 
neighborhood (c ÷. c +) of c. Then
(1.29) 1(r) ÷1(c) =
1
(a)
(c)
:!
(r ÷c)
a
.
where c is between c and r. It is clear that the continuity of 1
(a)
(r) at c
implies that the sign of this last function on maybe a smaller subinterval
(c ÷ o. c + o) of (c ÷ . c + ) is constant and it is the same like the
sign of 1
(a)
(c) (see Theorem 34). Suppose that 1
(a)
(r) 0 for any r ÷
(c ÷ o. c + o). Then, in (1.29), c ÷ (c ÷ o. c + o) and so, the sign of
the di¤erence 1(r) ÷1(c) depends exclusively on : and on the sign of
1
(a)
(c). If : is even, and 1
(a)
(c) 0. the di¤erence 1(r) ÷1(c) is 0,
for any r ÷ (c ÷ o. c + o). thus c is a local minimum point for 1. If
: is even, but 1
(a)
(c) < 0. then the di¤erence 1(r) ÷ 1(c) is < 0. for
any r ÷ (c ÷ o. c + o). so c is a local maximum point for 1. If : is
odd, the point c is not an extremum point because the sign of (r ÷c)
a
changes (it is positive if r c and negative otherwise). For instance,
1(r) = (r ÷2)
5
has not an extremum at r = 2.
2. TAYLOR SERIES 89
2. Taylor series
Let us consider a function 1 of class (
o
on an open subset ¹ of
R. This means that 1 has derivatives of any arbitrary order on ¹. It
is clear that all of these derivatives are continuous on ¹. Look at the
formula (1.10) and push the remainder to ·. We obtain the series of
functions on the right side:
(2.1) 1(c) +
1
t
(c)
1!
(r ÷c) +
1
tt
(c)
2!
(r ÷c)
2
+... +
1
(a)
(c)
:!
(r÷c)
a
+...
=
o
¸
a=0
1
(a)
(c)
:!
(r ÷c)
a
.
This series of functions is called the Taylor series associated to
the function 1 at the point c. If this series of functions is uniformly
convergent and its sum is 1(r). we say that
(2.2) 1(r) =
o
¸
a=0
1
(a)
(c)
:!
(r ÷c)
a
is the Taylor’s expansion of 1 around the point c. If the series on the
right side is simple convergent and its sum is 1 on an neighborhood
of c. we say that 1 is analytic at c. If 1 is analytic at any point of
¹ we say that 1 is analytic on ¹. The series on the right in (2.2) is
a particular case of a more general type of series of functions, namely,
the power series. A power series is a series of functions of the form
¸
o
a=0
c
a
(r ÷c)
a
. where ¦c
a
¦ is a sequence of real numbers and c is a
…xed arbitrary number.
Theorem 45. Let 1 : (c. d) ÷ R be an inde…nite di¤erentiable
function on an interval (c. d) (1 ÷ (
o
(c. d)) such that there is a positive
real number ` which veri…es
1
(a)
(r)
_ ` for any r ÷ (c. d) and for
any : = 0. 1. ... (we say that all the derivatives of 1 are uniformly
bounded on (c. d)). Then the series
¸
o
a=0
;
(n)
(o)
a!
(r ÷ c)
a
is absolutely
and uniformly convergent on (c. d) for any …xed c in (c. d). Moreover,
1(r) =
o
¸
a=0
1
(a)
(c)
:!
(r ÷c)
a
for any …xed c in (c. d). The series on the right is absolutely uniformly
convergent to 1.
90 4. TAYLOR SERIES
Proof. Let us denote 1 = d ÷ c. the length of the interval (c. d).
We apply the Weierstrass Test (Theorem 41):
1
(a)
(c)
:!
(r ÷c)
a
_
`
:!
1
a
for any r ÷ (c. d).
and the numerical series
¸
o
a=0
A
a!
1
a
is convergent (use the Ratio Test:
o
n+1
on
=
1
a+1
÷ 0 < 1). Hence, the series
¸
o
a=0
;
(n)
(o)
a!
(r ÷ c)
a
is ab
solutely and uniformly convergent. Let
:
a
(r) =
a
¸
I=0
1
(I)
(c)
/!
(r ÷c)
I
.
Formula (1.10) gives us:
[1(r) ÷:
a
(r)[ =
1
(a+1)
(c)
(: + 1)!
(r ÷c)
a+1
_
`
(: + 1)!
1
a+1
.
Taking sup we obtain 1 ÷:
a
 _
A
(a+1)!
1
a+1
and, since
A
(a+1)!
1
a+1
÷0
as : ÷ · (prove it by using a numerical series!), we get that ¦:
a
¦ is
uniformly convergent to 1. In particular
1(r) =
o
¸
a=0
1
(a)
(c)
:!
(r ÷c)
a
.
Example 8. (Taylor series for the basic elementary functions)
a) We know that
exp(r) = 1 +
r
1!
+
r
2
2!
+... +
r
a
:!
+ exp(c)
r
a+1
(: + 1)!
.
Since all the derivatives of exp(r) are uniformly bounded on any bounded
interval (c. /) (why?) we can apply Theorem 45 and …nd that the se
ries
¸
o
a=0
1
a!
r
a
is absolutely and uniformly convergent on any bounded
interval (c. /). In particular, we have the Taylor expansion
(2.3) exp(r) = 1 +
r
1!
+
r
2
2!
+... +
r
a
:!
+... =
o
¸
a=0
1
:!
r
a
. r ÷ R
b) We leave the reader to deduce the following Taylor expansions:
(2.4) sin(r) =
r
1!
÷
r
3
3!
+
r
5
5!
÷... + (÷1)
a
r
2a+1
(2: + 1)!
+...
2. TAYLOR SERIES 91
=
o
¸
a=0
(÷1)
a
(2: + 1)!
r
2a+1
. r ÷ R
(2.5) cos(r) = 1 +
r
2
2!
÷
r
4
4!
+
r
6
6!
÷... + (÷1)
a
r
2a
(2:)!
+...
=
o
¸
a=0
(÷1)
a
(2:)!
r
2a
. r ÷ R
Since all the derivatives of sin r and cos r are uniformly (inde
pendent of r) bounded (by 1) on R, the series on the right side in
the last two formulas are absolutely and uniformly convergent on any
BOUNDED interval of R (why not on the whole R?).
c)
(2.6) ln(1 +r) = r ÷
r
2
2
+
r
3
3
÷
r
4
4
+... + (÷1)
a÷1
r
a
:
+...
=
o
¸
a=1
(÷1)
a÷1
:
r
a
. r ÷ (÷1. 1).
Since the :th derivative of 1(r) = ln(1 + r) is
1
(a)
(r) = (÷1)
a÷1
(: ÷1)!(1 +r)
÷a
it is not uniformly bounded on the whole interval (÷1. 1) (why? ...
because sup(1+r)
÷a
= ·there!). Even on any other small subinterval
[c. /] of (÷1. 1) the derivatives of ln(1 + r) are not uniformly bounded
(because of :. this time!). Hence, we cannot apply the above Theorem
45. Let us look directly to the absolute value of the remainder in (1.21)
when r ÷ (÷1. 1) :
(÷1)
a
(1 +c)
÷a÷1
: + 1
r
a+1
.
where c belongs to the segment [0. r]
±
. i.e. c ÷ [0. r], or [r. 0] (for
r < 0). It is clear that if r ÷ ÷1. c may become closer and closer to
÷1 and the remainder cannot uniformly go to 0. But, if we take ANY
subinterval [c. /] of (÷1. 1). then
sup
a÷[o.b]
(÷1)
a
(1 +c)
÷a÷1
: + 1
r
a+1
_
1
: + 1
`
a+1
(1 +:)
a+1
.
92 4. TAYLOR SERIES
where ` = max¦[c[ . [/[¦ and : = min¦[c[ . [/[¦. Thus, in this last
case,
ln(1 +r) ÷:
a
 = sup
a÷[o.b]
(÷1)
a
(1 +c)
÷a÷1
: + 1
r
a+1
_
_
1
: + 1
¸
`
1 +:
a+1
÷0.
because
A
1+n
< 1. So, ¦:
a
(r)¦ is uniformly convergent to ln(1 + r).
relative to r, on [c. /] · (÷1. 1).
d)
(1 +r)
c
= 1 +
c
1!
r +
c(c ÷1)
2!
r
2
+...
... +
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
r
a
+...
or
(2.7) (1+r)
c
= 1+
o
¸
a=1
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
r
a
. r ÷ (÷1. 1).
For the series on the right side we shall prove later (Ch.5, Abel Theo
rem, Theorem 46) that this one is absolutely and uniformly convergent
on any closed subinterval [c. /] of (÷1. 1). We leave the reader to try
a direct proof for this last statement. For a …xed r in (÷1. 1) the se
ries in (2.7) is convergent (apply the Ratio Test). Thus, the series of
functions is simple convergent on (÷1. 1).
3. Problems
1. Find the Mac Laurin expansion for the following functions. In
dicate the convergence (or uniformly convergence) domain for each of
them.
a) 1(r) =
1
4
(exp(r) + exp(÷r) + 2 cos r); Hint: Use formula (2.3)
for exp(r) and for exp(÷r) (put ÷r instead r!) and formula (2.5) for
cos(r).
b) 1(r) =
1
2
arctan(r) +
1
4
ln
1+a
1÷a
; Hint: Compute
(arctan(r))
t
=
1
1 +r
2
= 1 ÷r
2
+r
4
÷...
and then integrate term by term; write then
ln
1 +r
1 ÷r
= ln(1 +r) ÷ln(1 ÷r)
and use formula (2.6) twice.
3. PROBLEMS 93
c) 1(r) = r arctan(r) ÷ln
1 +r
2
; Hint: Write
ln
1 +r
2
=
1
2
ln(1 +r
2
) =
1
2
(r
2
÷
r
4
2
+
r
6
3
÷...).
d) 1(r) =
1
a
2
÷3a+2
; Hint: Write
1
a
2
÷3a+2
=
¹
a÷1
+
1
a÷2
. then, for
instance
1
r ÷2
= ÷
1
2
1
1 ÷
a
2
= ÷
1
2
1 +
r
2
+
r
2
2
2
+... +
r
a
2
a
+...
.
e) 1(r) =
5÷2a
6÷5a+a
2
; f) 1(r) = ln(2÷3r+r
2
); Hint: ln(2÷3r+r
2
) =
ln(1 ÷r) + ln(2 ÷r) and
ln(2 ÷r) = ln 2 + ln(1 ÷
r
2
) = ln 2 +
r
2
÷
r
2
2
2
2
+
r
3
2
3
3
÷...
.
g) 1(r) = r exp(÷2r); Hint: in formula (2.3) put instead of r. ÷2r.
etc.
h) 1(r) = sin(3r) + r cos(3r); i) 1(r) = arcsin r; Hint: Compute
1
t
(r) = (1÷r
2
)
÷
1
2
and use the formula (2.7) with ÷r
2
instead of r and
c = ÷
1
2
.
j) 1(r) = sin
3
r; Hint: Write sin
3
r =
3
4
sin r ÷
1
4
sin 3r and use
formula (2.4) twice.
2. Write as a series of the form
¸
o
a=0
c
a
(r + 3)
a
the following
functions (say where this representation is possible):
a) 1(r) = sin(3r +2); Hint: Denote r +3 = . (a new variable) and
write 1(r) as a new function of . :
o(.) = sin(3(. ÷3) + 2) = sin(3. ÷7) = [sin 3.] cos 7 ÷[cos 3.] sin 7 =
= [cos 7]
3. ÷
(3.)
3
3!
+...
÷[sin 7]
1 ÷
(3.)
2
2!
+...
;
now, come back to 1(r) by the substitution . = r + 3. etc.
b) 1(r) =
3
(3 + 2r); c) 1(r) = ln(5 ÷4r); d) 1(r) = exp(2r +5);
e) 1(r) =
1
2÷3a
; f) 1(r) =
1
a
2
+3a+2
.
3. Using Mac Laurin formulas, compute the following limits:
a)lim
a÷0
exp(a
3
)÷1+ln(1+2a
3
)
a
3
; b)lim
a÷0
ln(1+2a)÷sin 2a+2a
2
a
3
; c)lim
a÷0
3
1+3a÷a÷1
1÷4a÷exp(÷4a)
;
d)lim
a÷0
cos a÷exp(÷
x
2
2
)
a
4
;
e) lim
a÷o
r ÷r
2
ln
1 +
1
a
; Hint: Write n =
1
a
; now, r ÷ · if and
only if n 0 and n ÷0; our limit becomes
lim
&÷0
¸
1
n
÷
1
n
2
ln(1 +n)
= lim
&÷0
¸
1
n
÷
1
n
2
n ÷
n
2
2
+
n
3
3
÷...
=
94 4. TAYLOR SERIES
= lim
&÷0
¸
1
2
÷
n
3
+...
=
1
2
.
4. Using Taylor formula approximately compute: a)
1.07 with 2
exact decimal digits; b)exp(0.25) with 3 exact decimals; c)ln(1.2) with
3 exact decimals; d)sin 1
·
with 5 exact decimals; Hint: 1
·
=
¬
180
radians;
so,
sin
:
180

r
1!
÷
r
3
3!
+
r
5
5!
÷... + (÷1)
a
r
2a+1
(2: + 1)!
.
where r =
¬
180
and : is chosen such that [1
2a+1
(r)[ . which is less then
1
(2a+2)!
r
2a+2
. to be less than
1
10
5
. So, we force
1
(2: + 2)!
:
180
2a+2
<
1
10
5
and …nd such a :.
CHAPTER 5
Power series
1. Power series on the real line
We saw that Mac Laurin series are special cases of some particular
series of functions
¸
o
a=0
c
a
r
a
. where ¦c
a
¦ is a …xed numerical sequence.
If one translates r into r÷c. where c is a …xed real number, we obtain a
more general series of functions,
¸
o
a=0
c
a
(r÷c)
a
. These ones are called
power series (with centre at c) on the real line. If we put n = r ÷c in
this last series, we get
¸
o
a=0
c
a
n
a
. i.e. a power series with centre at 0.
but in the variable n. Such translations reduce the study of a general
power series
¸
o
a=0
c
a
(r ÷c)
a
to a power series
¸
o
a=0
c
a
r
a
with centre
at 0. The mapping r ÷
¸
o
a=0
c
a
r
a
give rise to a function o(r) =
¸
o
a=0
c
a
r
a
. The maximal de…nition domain `
c
= ¦r ÷ R :
¸
o
a=0
c
a
r
a
is convergent¦ of this function o is called the convergence set of the
series. At least r = 0 is an element of `
c
(o(0) = c
0
). Sometimes `
c
reduces to the number 0. For instance, o(r) =
¸
o
a=0
:!r
a
is convergent
only at 0. Indeed, let us consider the series
¸
o
a=0
:! [r[
a
of moduli and
apply the Ratio Test: lim
a÷o
o
n+1
on
= lim
a÷o
(:+1) [r[ = ·. except r = 0. In
fact, if r = 0. ¦:!r
a
¦ does not tend to 0 (why?). Sometimes `
c
= R,
as in the case of the series o(r) =
¸
o
a=0
1
a!
r
a
= exp(r).
In the following, we want to describe the general form of the con
vergence set of a power series
¸
o
a=0
c
a
r
a
. The number
1 =
1
limsup¦
o
n+1
on
¦
in [0. ·] (i.e. 1 can be also ·) is called the convergence radius of the
series
¸
o
a=0
c
a
r
a
. Recall that limsup¦r
a
¦ is obtained in the following
way. Take all the convergent subsequences (include the unbounded
subsequences, i.e. subsequences which are convergent to ·) of the
sequence ¦r
a
¦ and the greatest of all these limits of them is called
limsup¦r
a
¦. the superior limit of the sequence ¦r
a
¦.
95
96 5. POWER SERIES
Theorem 46. (Abel Theorem) Let o(r) =
¸
o
a=0
c
a
r
a
be a power
series with real coe¢cients c
0
. c
1
. .... c
a
. ... and let 1 =
1
limsup[
a
n+1
an
[¦
in
[0. ·] be its convergence radius.
i) If 1 = 0. then the series o is absolutely convergent on the inter
val (÷1. 1) and absolutely uniformly convergent on any closed interval
[÷:. :], where 0 < : < 1. Moreover, the series is absolutely and uni
formly convergent on any closed subinterval [c. /] of (÷1. 1). If 1 = ·.
the series o is divergent on (÷·. ÷1) ' (1. ·). so,
(÷1. 1) · `
c
· [÷1. 1].
i.e. the convergence set of the series contains the open interval (÷1. 1).
it is contained in [÷1. 1] and at r = ÷1. or at r = 1 we must decide
in each particular case if the series is convergent or not.
ii) If 1 = 0. then the series o is convergent only at r = 0. i.e.
`
c
= ¦0¦.
iii) If 1 = 0. then the function o : (÷1. 1) ÷ R is of class (
o
on (÷1. 1). o
t
(r) =
¸
o
a=1
:c
a
r
a÷1
(termwise di¤erentiation) and a
primitive of o on (÷1. 1) is l(r) =
¸
o
a=0
on
a+1
r
a+1
(term by term
integration). All these power series l. o. o
t
. o
tt
. o
ttt
. .... o
(a)
. ... and
any other power series obtained from them by a termwise integration
or di¤erentiation process have the same convergence radius. Moreover,
if the series
¸
o
a=0
c
a
r
a
is convergent for r = 1. for instance, then
the function o : (÷1. 1] ÷ R, de…ned by o(r) =
¸
o
a=0
c
a
r
a
if r =
1 and o(1) =
¸
o
a=0
c
a
1
a
is continuous on (÷1. 1]. With this last
hypotheses ful…led, we also have that the series o(r) =
¸
o
a=0
c
a
r
a
is
absolutely and uniformly convergent on each closed subinterval of the
type [÷1+. 1]. where 0 is a small ( < 21) positive real number.
The same is true if we put ÷1 instead of 1 and if the numerical series
o(÷1) =
¸
o
a=0
c
a
(÷1)
a
is convergent.
Proof. i) Let us consider r as a …xed parameter (for the mo
ment) and let us apply the Ratio Test to the series of moduli 1(r) =
¸
o
a=0
[c
a
[ [r[
a
. Let 1 be the limit
1 = limsup
[c
a+1
[ [r[
a+1
[c
a
[ [r[
a
¸
=
¸
limsup
[c
a+1
[
[c
a
[
¸
[r[ =
[r[
1
.
If 1 = ·. then 1 = 0 < 1. so the series is absolutely convergent for
any r ÷ R. If 1 = 0. then 1 = ·. except maybe the case when
r = 0. Hence, if 1 = 0. the series is convergent ONLY for r = 0. i.e.
the statement of ii). Suppose now that 1 = 0. ·. Then, whenever
1 =
[a[
1
< 1. or r ÷ (÷1. 1), the series is absolutely convergent, in
1. POWER SERIES ON THE REAL LINE 97
particular convergent (see Theorem 31). If r ÷ (÷·. ÷1) '(1. ·). or
[r[ 1. then 1 1. Hence,
limsup
[c
a+1
[ [r[
a+1
[c
a
[ [r[
a
¸
1.
This means that there is at least one subsequence
[o
n
k
+1[[a[
n
k
+1
[on
k
[[a[
n
k
¸
of
[o
n+1
[[a[
n+1
[on[[a[
n
¸
such that
[o
n
k
+1[[a[
n
k
+1
[on
k
[[a[
n
k
1. i.e.
[c
a
k
+1
[ [r[
a
k
+1
[c
a
k
[ [r[
a
k
for any / = 0. 1. ... . Thus the sequence ¦c
a
r
a
¦ cannot tend to 0
and so, the series
¸
o
a=0
c
a
r
a
cannot be convergent for such an r. Let
now r ÷ [÷:. :]. where 0 < : < 1. Since for r = : < 1. the series
¸
o
a=0
[c
a
[ :
a
is convergent (: ÷ (÷1. 1). so the series
¸
o
a=0
c
a
r
a
is
absolutely convergent, see i)). But, [c
a
r
a
[ _ [c
a
[ :
a
for any : = 0. 1. ...
implies that the series
¸
o
a=0
c
a
r
a
is absolutely and uniformly conver
gent (we apply here the Weierstrass Test Theorem 41) on [÷:. :]. Since
any interval [c. /] · (÷1. 1) can be embedded in a symmetrical inter
val of the form [÷:. :] · (÷1. 1). we obtain that the series
¸
o
a=0
c
a
r
a
is absolutely and uniformly convergent on ANY closed subinterval [c. /]
of (÷1. 1).
iii) It is easy to see that all the power series l. o
t
. o
tt
. ... have the
same convergent radius 1 as the series o. Applying the Weierstrass
test to each of them on an interval of the form [÷:. :] · (÷1. 1) and
the theorems 39 and 40, we can prove easily the …rst statement of
iii). To prove the absolute and uniform convergence of
¸
o
a=0
c
a
r
a
on
[÷1 +. 1] if the numerical series
¸
o
a=0
c
a
1
a
is convergent, we apply
the AbelDirichlet’s Test for series of functions (Theorem 42) by taking
/
a
(r) = :c
a
1
a
and c
a
(r) =
a
n
a1
n
. Since c
a
(r) =
1
a
on the interval
[÷1 + . 1] and since the series
¸
o
a=0
:c
a
1
a
is convergent (apply the
RaabeDuhamel Test, taking count of the de…nition of 1), i.e. the
partial sums of this last series are bounded (uniformly, because here
we have no r!). the hypothesis of the AbelDirichlet Test are ful…led.
Hence, the series
¸
o
a=0
c
a
r
a
is absolutely and uniformly convergent on
[÷1 +. 1].
Let us consider the power series
o(r) =
o
¸
a=1
(÷1)
a÷1
:
r
a
.
98 5. POWER SERIES
We know that this one is identical with ln(1 + r) on (÷1. 1). Let us
…nd the convergence set `
c
of it. The convergence radius is equal to
1 =
1
limsup¦
o
n+1
on
¦
=
1
limsup¦
1
n+1
1
n
¦
= 1.
At r = ÷1. the series becomes
o(÷1) = ÷
o
¸
a=1
1
:
= ÷·.
so the series is divergent at r = ÷1. Now, o(1) =
¸
o
a=1
(÷1)
n1
a
is
the alternate series, which was proved to be convergent. Since both
functions o(r) and ln(1+r) are continuous at r = 1 (prove it!by using
iii) of the Abel Theorem), one has that o(1) = ln 2. FromAbel Theorem
we see that `
c
is exactly (÷1. 1]. On this interval it is ln(1 + r) but,
the series does not exist outside of (÷1. 1]. while the function ln(1 +r)
does exist, for instance at r = 2!
Let us now look at the binomial series
o(r) = 1 +
o
¸
a=1
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
r
a
.
where c is a …xed real parameter. Let us …nd the convergence radius
of this series:
(1.1) 1 =
1
limsup¦
o
n+1
on
¦
= lim
a÷o.ac
: ÷c
: + 1
= 1
If r = ÷1. the series is not convergent for any c. For instance, if
c = ÷1. then o(÷1) =
¸
o
a=0
(÷1)
a
(÷1)
a
= ·. At r = 1. o(1) =
¸
o
a=0
(÷1)
a
is divergent. If c is a natural number /. then the series
becomes a polynomial, so its convergence set is the whole R. But,...the
formula (1.1) and Abel Theorem say that... `
c
= R ·[÷1. 1] !!! Some
where must be a mistake! Indeed, since c
I+1
= c
I+2
= ... = 0.
limsup¦
o
n+1
on
¦ is nondeterministic, so the computation of 1 in (1.1) is
wrong! We see that the convergence set `
c
(c) of the binomial series
strongly depends on c. We do not give here a complete discussion of
`
c
(c) as a function of c.
Let us …nd the convergence set for the following series of functions
o(r) =
o
¸
a=1
(÷1)
a
:
2
1
2r + 1
a
.
1. POWER SERIES ON THE REAL LINE 99
This is not a power series but, making the substitution n =
1
2a+1
. we
obtain a power series 1(n) =
¸
o
a=1
(÷1)
n
a
2
n
a
in the new variable n. The
convergence radius of this last series is
1 =
1
limsup¦
o
n+1
on
¦
=
1
lim
a÷o
a
2
(a+1)
2
= 1.
For n = ±1. the series is convergent (why?). So, the convergence set
`
c.&
for the power series
1(n) =
o
¸
a=1
(÷1)
a
:
2
n
a
is `
c.&
= [÷1. 1]. Coming back to the variable r. we get that the initial
series of functions
o(r) =
o
¸
a=1
(÷1)
a
:
2
1
2r + 1
a
is convergent if and only of ÷1 _
1
2a+1
_ 1. i. e.
r ÷ (÷·. ÷1] ' [0. ·).
Hence, the set of all r in R such that the series
o(r) =
o
¸
a=1
(÷1)
a
:
2
1
2r + 1
a
is convergent, i.e. the convergence set of this last series, is
(÷·. ÷1] ' [0. ·).
Remark 16. (CauchyHadamard) Another useful formula for com
puting the convergence radius 1 of a power series
¸
o
a=0
c
a
r
a
is the
following CauchyHadamard formula:
(1.2) 1 =
1
limsup
n
[c
a
[
This formula can be used even when an in…nite number of c
a
are zero.
The proof of Abel’s Theorem by using this formula for 1 is completely
analogue to the proof of the same theorem given above. In this case one
must use the Root Test (Theorem 29) instead of the Ratio Test as we
did in proving Abel Theorem. If we start with the de…nition of 1 as
it appears in formula CauchyHadamard (1.2), we get THE SAME in
terval of convergence (÷1. 1) for our series
¸
o
a=0
c
a
r
a
(why?). Thus,
the both formulas give rise to one and the same number.
100 5. POWER SERIES
Let us …nd the convergence set and the sum of the series of functions
o(r) =
o
¸
a=0
1
2: + 1
(3r + 2)
2a+1
.
This one is not a power series but,...we can associate to it a power
series by the following substitution n = 3r + 2. Hence, we must study
the power series in n :
1(n) =
o
¸
a=0
1
2: + 1
n
2a+1
.
Here c
2a+1
=
1
2a+1
and c
2a
= 0 for any : = 0. 1. ... . In our case, it is
not a good idea to apply Abel formula 1 =
1
limsup[
a
n+1
an
[¦
(why?). Let
us apply CachyHadamard formula (1.2):
1 =
1
limsup
n
[c
a
[
= 1.
because the sequence ¦
n
[c
a
[¦ is the union between two convergent
subsequences:
¦
2n+1
[c
2a+1
[¦ = ¦
2n+1
1
2: + 1
¦ ÷1
(why?) and
¦¦
2n
[c
2a
[¦ = ¦0¦ ÷0
and so, limsup
n
[c
a
[ = 1. At n = ÷1 the series
1(n) =
o
¸
a=0
1
2: + 1
n
2a+1
becomes
1(÷1) = ÷
o
¸
a=0
1
2: + 1
= ÷·
(why?). At n = 1 the series is
1(1) =
o
¸
a=0
1
2: + 1
= ·.
Hence, the convergence set for the power series 1(n) is (÷1. 1) (see
Abel Theorem 46). Now,
1
t
(n) =
o
¸
a=0
n
2a
=
1
1 ÷n
2
=
1
2
1
1 ÷n
+
1
2
1
1 +n
.
1. POWER SERIES ON THE REAL LINE 101
Thus,
1(n) =
1
2
ln
1 +n
1 ÷n
+(.
But ( = 0 because 1(0) = 0. Let us come back to the series in r. The
convergence set is
¦r ÷ R : ÷1 < 3r + 2 < 1¦ = (÷1. ÷
1
3
).
Its sum is
o(r) = 1(3r + 2) =
1
2
ln
÷
3r + 3
3r + 1
.
Example 9. (arctan series) Let us …nd the Mac Laurin expansion
for 1(r) = arctan r. For this let us consider
1
t
(r) =
1
1 +r
2
= 1 ÷r
2
+r
4
÷... + (÷1)
a
r
2a
+....
where [r[ < 1 (why?). Apply now Theorem 43 and integrate termwise:
(1.3) arctan r +( = r ÷
r
3
3
+
r
5
5
÷
r
7
7
+... + (÷1)
a
r
2a+1
2: + 1
+....
where [r[ < 1. For r = 0 we get ( = 0. Since for r = 1 the series on
the right is convergent and since the function
o(r) = r ÷
r
3
3
+
r
5
5
÷
r
7
7
+... + (÷1)
a
r
2a+1
2: + 1
+...
is continuous at r = 1 (see Abel’s Theorem, iii)), we get that
(1.4) arctan 1 =
:
4
= 1 ÷
1
3
+
1
5
÷
1
7
+... + (÷1)
a
1
2: + 1
+...
Let us …nd the convergence set and the sum for the power series
o(r) =
o
¸
a=1
:(: + 1)r
a
.
The convergence radius is
1 = lim
a÷o
:(: + 1)
(: + 1)(: + 2)
= 1
(why?). Since at r = ±1 the series is divergent (:(: + 1) 9 0),
the convergence set is `
c
= (÷1. 1). Let us integrate termwise (see
Theorem 43):
o(r)dr =
o
¸
a=1
:r
a+1
=
o
¸
a=1
(: + 2)r
a+1
÷2
o
¸
a=1
r
a+1
.
102 5. POWER SERIES
But the series
o
¸
a=1
r
a+1
= r
2
+r
3
+... =
r
2
1 ÷r
(it is an in…nite geometrical progression). So we get
o(r)dr =
o
¸
a=1
(: + 2)r
a+1
÷
2r
2
1 ÷r
.
Let us integrate again this last equality
¸
o(r)dr
dr =
o
¸
a=1
r
a+2
+r
2
+ 2r + 2 ln(1 ÷r) =
=
r
3
1 ÷r
+r
2
+ 2r + 2 ln(1 ÷r).
Coming back and di¤erentiating twice, we get:
o(r) =
2r
(r ÷1)
3
. for [r[ < 1.
2. Complex power series and Euler formulas
In Chapter 2, Section 2, we introduced the metric space of complex
number …elds C. In fact, C is a normed spaced with the norm given by
the usual complex modulus [.[ =
r
2
+n
2
. where . = r+in. r. n ÷ R
(prove the properties of the norm for this particular norm!). Since a
sequence ¦.
a
= r
a
+ in
a
¦ is convergent to . = r + in in C if and only
if both the real sequences ¦r
a
¦ and ¦n
a
¦ are convergent to r and to
n respectively (see Theorems 1 and 16), the study of the numerical
series with complex terms reduces to the study of the real numerical
series. But this way is not so easy to put in practice. The best way is
to use …rstly the absolute convergence notion like in the case of series
in a general normed space. Namely, let : =
¸
o
a=0
.
a
be a series with
complex numbers terms and let o =
¸
o
a=0
[.
a
[ be the real series of
moduli. The following result is very useful in practice.
Theorem 47. If the series of moduli o =
¸
o
a=0
[.
a
[ is convergent
(like a numerical real series with nonnegative terms), the initial series
with complex terms : =
¸
o
a=0
.
a
is convergent in C.
Proof. Let :
a
=
¸
a
I=0
.
I
be the :th partial sum of the series
: =
¸
o
a=0
.
a
and let o
a
=
¸
a
I=0
[.
I
[ be the :th partial sum of the
series of moduli o =
¸
o
a=0
[.
a
[ . Since
[:
a+j
÷:
a
[ _ [.
a+1
[ +[.
a+2
[ +... +[.
a+j
[ = o
a+j
÷o
a
.
2. COMPLEX POWER SERIES AND EULER FORMULAS 103
and since the series o is convergent (i.e. the sequence ¦o
a
¦ is a Cauchy
sequence), one obtains that the sequences ¦:
a
¦ is a Cauchy sequence.
Thus, it is convergent to a complex number : (the sum of the series
¸
o
a=0
.
a
) in C, because C is a complete metric space (see Theorem
16).
The Cauchy Test and the zero Test also work in the case of a com
plex series (why?Hint: C is a complete metric spacewhy?). Series of
complex functions and power series are de…ned exactly in the same way
like the analogous real case. However, in the complex case, the study
of the convergence set of a series of function is more complicated than
in the real case.
Example 10. (Complex geometrical series). Let us …nd the con
vergence set for the complex geometrical series
:(.) =
o
¸
a=0
.
a
= 1 +. +.
2
+....
Let us consider the series of moduli
o([.[) =
o
¸
a=0
[.[
a
= lim
a÷o
1 ÷[.[
a+1
1 ÷[.[
.
This limit exists if [.[ < 1. Hence, the series is absolutely convergent if
and only if [.[ < 1. In particular, for [.[ < 1. the series is convergent
(see Theorem 47). Is the series convergent for a . with [.[ 1? Let
us see ! If [.[ 1. the sequence ¦.
a
¦ goes to · in C = C ' ¦·¦. the
Riemann sphere (why?), so, the series is divergent (see the zero Test).
What happens if [.[ = 1?. i.e. if . is a complex number on the circle
of radius 1 and with centre at origin. If . = 1. the series is divergent.
If . = 1. but [.[ = 1. the sequence ¦.
a
¦ is NEVER convergent to zero!
(why?). Thus, the convergence set for the series :(.) =
¸
o
a=0
.
a
is
exactly the open disc 1(0; 1) = ¦. ÷ C : [.[ < 1¦ in the complex plane
C.
To de…ne the basic elementary complex functions one uses complex
power series. For instance, the exponential complex function is de…ned
by the formula
(2.1) exp(.) = 1 +
.
1!
+
.
2
2!
+... +
.
a
:!
+... =
o
¸
a=0
.
a
:!
It is easy to prove (do it!) that this series is absolutely convergent
on the whole complex plane C and absolutely uniformly convergent on
any bounded subset of C. The series on the right side is the natural
104 5. POWER SERIES
extension of the Mac Laurin expansion of the real function exp(r) to the
whole complex plane. Using this "trick" we can de…ne other elementary
complex functions:
(2.2) sin(.)
oc;
=
.
1!
÷
.
3
3!
+
.
5
5!
÷... + (÷1)
a
.
2a+1
(2: + 1)!
+...
=
o
¸
a=0
(÷1)
a
(2: + 1)!
.
2a+1
. r ÷ C
(2.3) cos(.)
oc;
= 1 ÷
.
2
2!
+
.
4
4!
÷
.
6
6!
+... + (÷1)
a
.
2a
(2:)!
+...
=
o
¸
a=0
(÷1)
a
(2:)!
.
2a
. . ÷ C
(2.4) ln(1 +.)
oc;
= . ÷
.
2
2
+
.
3
3
÷
.
4
4
+... + (÷1)
a÷1
.
a
:
+... =
=
o
¸
a=1
(÷1)
a÷1
:
.
a
. [.[ < 1.
(1 +.)
c
oc;
= 1 +
c
1!
. +
c(c ÷1)
2!
.
2
+...+
+
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
.
a
+....
so,
(2.5)
(1 +.)
c
= 1 +
o
¸
a=1
c(c ÷1)(c ÷2)...(c ÷: + 1)
:!
.
a
. [.[ < 1. c ÷ C
In the same way we can de…ne any other complex function 1(.) if
we know a Taylor expansion for the real function 1(r) (if this last one
has really real values and if it can be extended beyond the real line!).
For instance, we know that
:/(r) = r +
r
3
3!
+
r
5
5!
+... +
r
2a+1
(2: + 1)!
+.... r ÷ R.
We simply de…ne the complex hyperbolic sine as
(2.6) :/(.)
oc;
= . +
.
3
3!
+
.
5
5!
+... +
.
2a+1
(2: + 1)!
+.... . ÷ C.
2. COMPLEX POWER SERIES AND EULER FORMULAS 105
and
(2.7) c/(.)
oc;
= 1 +
.
2
2!
+
.
4
4!
+... +
.
2a
(2:)!
+.... . ÷ C.
We always have to check if the series on the right side is convergent on
the extrapolated domain (for instance, we extrapolated R to C). The
restrictions of all these functions to their de…nition domains on the real
line give rise to the well known real functions. For instance, ln(1 +.).
[.[ < 1. restricted to R give rise to ln(1 +r). This does not mean that
we de…ned the function ln(.) for any . = 0! To de…ne such a function,
i.e. the inverse of the complex exponential function, is not an easy
task, because it will be not an usual function, i.e. for a . we have more
than one value of ln(.). This is because exp(.) is NOT injective at all.
To see this we need some famous relations, the Euler formulas.
Theorem 48. (Euler relations) For any r a real number and for
i =
÷1 we have
(2.8) exp(ir) = cos(r) +i sin(r).
(2.9) cos(r) =
exp(ir) + exp(÷ir)
2
and
sin(r) =
exp(ir) ÷exp(÷ir)
2i
.
Proof. We simply use formula (2.1) to compute exp(ir) :
exp(ir) = 1 +
ir
1!
÷
r
2
2!
÷
ir
3
3!
... +
(ir)
a
:!
+... = cos(r) +i sin(r).
If now we put instead of r. ÷r in the formula (2.8), we get
(2.10) exp(÷ir) = cos(r) ÷i sin(r).
because cosine is an even function and sine is an odd one. Adding
formulas (2.8) and (2.10), we get the relation exp(ir) + exp(÷ir) =
2 cos(r). Now, subtract formula (2.10) from formula (2.8) and get the
formula exp(ir) ÷exp(÷ir) = 2i sin(r). etc.
Let us justify nowthat the complex function exp(.) is not invertible,
i.e. it cannot have like inverse an usual function. Using Euler formulas
from the theorem we get that
exp(2/:i) = cos(2/:) +i sin(2/:) = 1.
106 5. POWER SERIES
for any integer /. Thus one has an in…nite number of complex numbers
¦2::i¦. : = 0. ±1. ±2. .... in which the exponential function has the
value 1!. This is why the inverse of exp(.) is the multiform function
1:(.) = ln [.[ +i(o + 2/:). / = 0. ±1. ±2. ...
and o is the argument of .. i.e. the unique real number in [0. 2:) such
that . = [.[ [cos(o) + i sin(o)]. the trigonometric representation of .
(prove this last equality by drawing...). It has a double in…nite number
of "branches", i.e. 1:(.) is in fact the set
¦ln
(I)
(.) = ln [.[ +i(o + 2/:)¦. / = 0. ±1. ±2. ...
of usual functions. All of these functions have the same real part ln [.[ .
For / = 0 we get the principal branch, ln(.) = ln [.[ + i arg .. Some
times in books people work with this last expression for the complex
logarithmic function, without mention this. We leave as an exercise for
the reader to de…ne the radical complex multiform function
n
. (it has
only : branches!…nd them!). One can start with the fact that
n
. is
the inverse of the power : function . .
a
and with the equality:
.
a
= [.[
a
[cos(:o) +i sin(:o)].
etc.
Euler’s formulas from the above theorem are very useful in practice.
For instance, the famous de Moivre formula
[cos(r) +i sin(r)]
a
= cos(:r) +i sin(:r)
from trigonometry, can be immediately proved by using the basic prop
erties of the complex exponential function : exp(.) exp(n) = exp(.+n)
(try to prove it!), exp(.)
a
= exp(:.). where .. n ÷ C, and : is an in
teger number. If one extends in a natural way (componentwise!) the
integral calculus from real functions to functions of real variables but
with complex values:
[1(r) +io(r)]dr =
1(r)dr +i
o(r)dr.
one can compute in an easy way more complicated integrals. For in
stance, let us …nd a primitive for a very known family of functions
1(r) = exp(cr) cos(/r). where c. / are two …xed real numbers (para
meters). Let us denote by o(r) = exp(cr) sin(/r) (its partner!) and let
us …nd a primitive for 1(r) +io(r) :
[exp(cr) cos(/r) +i exp(cr) sin(/r)]dr =
exp(cr) exp(i/r)dr =
3. PROBLEMS 107
=
exp(cr +i/r)dr =
exp(cr +i/r)
c +i/
=
=
exp(cr) [cos(/r) +i sin(/r)](c ÷i/)
c
2
+/
2
=
= exp(cr)
c cos(/r) +/ sin(/r)
c
2
+/
2
+i exp(cr)
c sin(/r) ÷/ cos(/r)
c
2
+/
2
.
Hence,
exp(cr) cos(/r)dr = exp(cr)
c cos(/r) +/ sin(/r)
c
2
+/
2
and
exp(cr) sin(/r)dr = exp(cr)
c sin(/r) ÷/ cos(/r)
c
2
+/
2
(why?).
Another example of a nice application of Euler formulas is the fol
lowing. Suppose we forgot the formula for sin 3r and of cos 3r in lan
guage of sin r and cos r respectively. Let us …nd it by writing
cos 3r +i sin 3r = exp(i3r) =
(Euler formula)
= [exp(ir)]
3
= [cos r +i sin r]
3
=
= cos
3
r ÷3 cos r sin
2
r +i[3 cos
2
r sin r ÷sin
3
r].
Since two complex numbers are equal if their real and imaginary
parts are equal, we get the formulas:
cos 3r = cos r[cos
2
r ÷3 sin
2
r] = cos r[4 cos
2
r ÷3].
sin 3r = [3 cos
2
r sin r ÷sin
3
r] = sin r[3 ÷4 sin
2
r].
3. Problems
1. Find the convergence set and the sum for the following series of
functions:
a)
¸
o
a=0
(3r + 5)
a
; b)
¸
o
a=0
(÷1)
a
(4r + 1)
a
; c)
¸
o
a=1
a
n
a
;
d)
¸
o
a=1
(÷1)
a÷1 a
n
a
; e)
¸
o
a=1
:(3r + 5)
a
; f)
¸
o
a=0
a
n
(a+1)2
n
;
2. Find the convergence set for the following series of functions:
a)
¸
o
a=1
1
(1+
1
n
)
n
2
(r ÷3)
a
; b)
¸
o
a=1
a
n
a
2
; c)
¸
o
a=0
:!r
a
; d)
¸
o
a=0
a
n
a!
;
108 5. POWER SERIES
e)
¸
o
a=1
a
n
a
n
; f)
¸
o
a=1
a
5
5
n
r
a
; g)
¸
o
a=0
a
n
2
n
+3
n
; h)
¸
o
a=1
a+1
a
a
2
r
a
;
i)
¸
o
a=0
[1 ÷(÷2)
a
]r
a
; j)
¸
o
a=0
(÷1)
a+1
3
a
r
a
; k)
¸
o
a=1
1
2a+1
1+a
1÷a
a
;
l)
¸
o
a=1
(÷1)
a
2
n
(a÷5)
2n
a
2
; m)
¸
o
a=1
(÷1)
a÷1
(a÷5)
2n
a3
n
(…nd its sum);
3. Use the power series in order to compute the following sums:
a)
¸
o
a=1
(÷1)
a÷1 1
a
; b)
¸
o
a=0
1
(a+1)2
n
; c)
¸
o
a=1
a
2
n
; (Hint: associate the power
series
o(r) =
o
¸
a=1
:r
a
= r(1+2r+3r
2
+...) = r(r+r
2
+r
3
+...)
t
= r
r
1 ÷r
t
;
make then r =
1
2
).
CHAPTER 6
The normed space R
m
.
1. Distance properties in R
n
Motivation Let ¦(; i. j¦ be a Cartesian coordinate system in a
plane ({). To any point ` ÷ ({) we associate the position vector
÷÷÷
(`. We know that there is a unique pair (r. n) of real numbers such
that
÷÷÷
(` = ri + nj. Here i. j are two perpendicular versors with their
origin in (. Usually one calls (r. n) the coordinates of ` relative to the
"basis" ¦i. j¦. But we can view (r. n) as an element in RR
act
= R
2
. If
`
t
is another point in the same plane ({) and if 1 is the unique point
in ({) such that
÷÷÷
(` +
÷÷÷
(`
t
=
÷÷
(1. then the coordinates of 1 are
(r+r
t
. n +n
t
). where (r
t
. n
t
) are the coordinates of `
t
. Let c be a real
number (scalar) and let us denote by
÷÷÷÷
(`
tt
the vector c
÷÷÷
(`. Then, the
coordinates of the point `
tt
are (cr. cn) ÷ R
2
. So, one can endow the
cartesian product R
2
with a natural algebraic structure of a real vector
space with 2 dimensions (the number of the elements in any basis of
it, in particular in the "canonical" basis ¦(1. 0). (0. 1)¦. where (1. 0)
are the coordinates of the versor i and (0. 1) are the coordinates of the
versor j). Hence, one can study the 2dimensional dynamics only in the
"abstract" space R
2
(this is the basic idea of R. Descartes; the word
"cartesian" comes from "Descartes", in Latin "Cartesius"; he invented
a very useful tool for Engineering, namely the Analytic Geometry; here
we work with numbers and equations instead of geometrical objects like
lines, circles, parabolas, etc.). We call R
2
the 2dimensional space (21
space). In the same way we can construct the 31 space R
3
or, more
generally, the :1 space
R
n
= R R ... R
. .. .
a÷tjncc
= ¦x = (r
1
. r
2
. .... r
n
) : r
;
÷ R¦.
We recall that if x = (r
1
. r
2
. .... r
n
) and y = (n
1
. n
2
. .... n
n
) are two
"vectors" in R
n
. then
x +y = (r
1
+n
1
. r
2
+n
2
. .... r
n
+n
n
)
109
110 6. THE NORMED SPACE R
m
:
and
cx = (cr
1
. cr
2
. .... cr
n
)
for any "scalar" c ÷ R (componentwise operations). For instance,
(÷7. 3)+(6. 0) = (÷1. 3) and
2(÷1. 1) = (÷
2.
2). To do analysis in
R
n
means …rstly to introduce a distance in R
n
. R
n
has the "canonical
basis"
¦(1. 0. .... 0). (0. 1. 0. .... 0). ...(0. 0. .... 0. 1)¦
like a real vector space, so it has the dimension :over R. It is more prof
itable to introduce …rst of all a "length" of a vector x = (r
1
. r
2
. .... r
n
)
by the formula
(1.1) x
oc;
=
r
2
1
+r
2
2
+... +r
2
n
.
The nonnegative real number x is called the norm or the length of x.
If : = 1. the norm of a real number r is its absolute value (modulus)
[r[. If : = 2 and if x = (r
1
. r
2
) the norm x =
r
2
1
+r
2
2
is exactly
the length of the diagonal of the rectangle [(¹
1
`¹
2
]. or the length of
the resultant vector
÷÷÷
(` =
÷÷
(¹
1
+
÷÷
(¹
2
(see Fig.6.1).
O
x
y
A
1
A
2
x
1
x
2
M(x
1
,x
2
)
Fig. 6.1
In the 31 space R
3
the norm of x = (r
1
. r
2
. r
3
) is
r
2
1
+r
2
2
+r
2
3
and it is exactly the length of the diagonal of the parallelepiped gener
ated by
÷÷
(¹
1
.
÷÷
(¹
2
and
÷÷
(¹
3
(see Fig.6.2).
1. DISTANCE PROPERTIES IN R
m
111
A
1
A
2
A
3
x
1
x
2
x
3
x
y
z
M(x
1
,x
2
,x
3
)
Fig. 6.2
Example 11. (the spacetime representation) Let us consider the
vector x = (r
1
. r
2
. r
3
. t) ÷ R
4
. where (r
1
. r
2
. r
3
) are the coordinates of
a point `(r
1
. r
2
. r
3
) in the 31 space and t _ 0 is the time when we
"observe" the point `. Then
x =
r
2
1
+r
2
2
+r
2
3
+t
2
.
Example 12. (the space of dynamics) Let us consider a moving
point ` on a trajectory () in the 31 space. The position of ` is
…xed by its coordinates r
1
. r
2
. r
3
. Its velocity v is given by another
3 coordinates
·
r
1
.
·
r
2
.
·
r
3
. the derivatives of the coordinates functions
r
1
(t). r
2
(t). r
3
(t) at `. Thus, the "dynamic" state of ` is described
by the "vectors"
x = (r
1
. r
2
. r
3
.
·
r
1
.
·
r
2
.
·
r
3
) ÷ R
6
and
x =
r
2
1
+r
2
2
+r
2
3
+
·
r
2
1
+
·
r
2
2
+
·
r
2
3
.
Theorem 49. The norm mapping
x x =
r
2
1
+r
2
2
+... +r
2
n
.
from R
n
to R
+
. has the following main properties: 1) x = 0 if
and only if x = 0; 2) cx = [c[ x for any c ÷ R, x ÷R
n
; 3)
x +y _ x +y . for any x, y ÷R
n
.
Proof. 1) and 2) are obvious (prove them!). To be clearer, let us
prove 3) for : = 2. Both sides in 3) are nonnegative, so the inequality
is equivalent to
x +y
2
_ x
2
+y
2
+ 2 x y .
112 6. THE NORMED SPACE R
m
:
If x = (r
1
. r
2
) and y = (n
1
. n
2
). one has
(r
1
+n
1
)
2
+ (r
2
+n
2
)
2
_ r
2
1
+r
2
2
+n
2
1
+n
2
2
+ 2
(r
2
1
+r
2
2
)(n
2
1
+n
2
2
).
or, r
1
n
1
+r
2
n
2
_
(r
2
1
+r
2
2
)(n
2
1
+n
2
2
). By squaring both sides we get
2r
1
r
2
n
1
n
2
_ r
2
2
n
2
1
+r
2
1
n
2
2
.
or 0 _ (r
2
n
1
÷ r
1
n
2
)
2
. This last inequality is obvious. Moreover, from
this last inequality, we can say that in 3) we have equality if and only
if r
2
n
1
÷r
1
n
2
= 0 or, if and only if (r
1
. r
2
) = \(n
1
. n
2
), i.e. x and y are
collinear.
The couple (R
n
. .) is called a normed space. We know that in
general, a normed space is a real vector space A with a norm mapping
. on it, which veri…es the properties 1), 2) and 3) from Theorem 49.
We recall that a normed space (A. .) is also a metric space w.r.t.
a canonically induced distance: d(r. n) = r ÷n for any r. n in A.
In the case of the normed space (R
n
. .) the distance is given by the
formula
(1.2) d(x. y) =r ÷n =
j=n
¸
j=1
(r
j
÷n
j
)
2
This distance is a very special one because it comes from the "scalar
product"
(1.3) < x. y =
n
¸
j=1
r
j
n
j
.
i.e. this last one induces the norm x =< x. x =
¸
n
j=1
r
j
2
on R
n
and this norm gives rise exactly to our distance (1.2). As we know from
the Linear Algebra course, the scalar product (1.3) endows R
n
with a
geometry. The length of a vector x is its norm x =
¸
j=n
j=1
r
j
2
and
the cosine of the angle o between two vectors x and y of R
n
is de…ned
as
cos o =
< x. y
x y
.
The fact that the quantity
<x.y
xy
is always between ÷1 and 1 is exactly
the famous Cauchy SchwarzBuniakowsky inequality
(1.4) [< x. y [ _ x y .
It can be proved only by using the basic properties of a scalar product
(see any course in Linear Algebra).
1. DISTANCE PROPERTIES IN R
m
113
Since R
n
is a metric space relative to the distance d de…ned in (1.2)
we can speak about the convergence of a sequence
¦x
(a)
= (r
(a)
1
. r
(a)
2
. .... r
(a)
n
)¦
from R
n
to a vector x = (r
1
. r
2
. .... r
n
) : we say that x
(a)
÷ x if and
only if d(x
(a)
. x) ÷0. i.e. if and only if
j=n
¸
j=1
(r
(a)
j
÷r
j
)
2
÷0.
But, a sum of squares becomes smaller and smaller if and only if any
square in the sum becomes smaller and smaller. Thus, we just obtained
a part of the following basic result:
Theorem 50. (componentwise convergence). 1) A sequence
¦x
(a)
= (r
(a)
1
. r
(a)
2
. .... r
(a)
n
)¦
of vectors from R
n
is convergent to a vector x = (r
1
. r
2
. .... r
n
) if
and only if for any i = 1. 2. .... :. the numerical sequence ¦r
(a)
j
¦ is
convergent to r
j
, when : ÷·. 2) A sequence
¦x
(a)
= (r
(a)
1
. r
(a)
2
. .... r
(a)
n
)¦
is a Cauchy sequence in R
n
if and only if any "component" "i". ¦r
(a)
j
¦.
is a Cauchy sequence in R for any i = 1. 2. .... :. Since R is a complete
metric space (see Theorem 13), we see that R
n
is also a complete metric
space.
Proof. 1) was just proved before the statement of the theorem.
For 2) let us consider a sequence ¦x
(a)
= (r
(a)
1
. r
(a)
2
. .... r
(a)
n
)¦. It is a
Cauchy sequence if for any 0 we can …nd a rank `
.
such that if
: _ `
.
one has that d(x
(a+j)
. x
(a)
) < for any j = 1. 2. ... . This
means that whenever : is large enough the distance d(x
(a+j)
. x
(a)
) is
small enough, independent on j. But
(1.5) d(x
(a+j)
. x
(a)
) =
j=n
¸
j=1
(r
(a+j)
j
÷r
(a)
j
)
2
.
So,
r
(a+j)
j
÷r
(a)
j
becomes small enough, independent on j whenever
: is large enough. And this is true for any …xed i = 1. 2... . But
this last remark says that the sequence ¦r
(a)
j
¦ is a Cauchy sequence
for any …xed i = 1. 2. ... . Conversely, if all the sequences ¦r
(a)
j
¦ are
Cauchy sequences for i = 1. 2. .... then, in (1.5), all the di¤erences
114 6. THE NORMED SPACE R
m
:
r
(a+j)
j
÷r
(a)
j
become smaller and smaller, independent of j. whenever
: becomes large enough. Hence, the whole sum
¸
j=n
j=1
(r
(a+j)
j
÷ r
(a)
j
)
2
becomes smaller and smaller, independent of j. whenever : ÷·. i.e.
the sequence ¦x
(a)
¦ is a Cauchy sequence in R
n
. The last statement
becomes very easy now (why?).
For instance, the sequence ¦(
1
a
.
a+1
a
)¦ is convergent to (0. 1) in R
2
because the …rst component ¦
1
a
¦ goes to 0 and the second component
a+1
a
goes to 1.
A normed vector space, which is a complete metric space w.r.t. the
distance de…ned by its norm, is called a Banach space. Such spaces are
very useful in many engineering models.
We recall now, in our particular case of the metric space (R
n
. d).
where d is de…ned in (1.2), the following basic notion.
Definition 16. Let a =(c
1
. c
2
. .... c
n
) be a …xed point in R
n
and
let : 0 be a positive real number. The set 1(a.:) = ¦x ÷ R
n
:
x ÷a = d(x. a) < :¦ is called the open ball with centre at a and of
radius :. The set
1[a.:] = ¦x ÷ R
n
: x ÷a = d(x. a) _ :¦
is said to be the closed ball with centre at a and of radius : (_ 0).
For instance, if : = 1. a =c ÷ R then 1(a.:) = (c ÷:. c + :). the
usual open interval with centre at c and of length 2: (prove this!). In
the same case, 1[a.:] = [c ÷ :. c + :]. If : = 2. 1(a.:) is the usual
open (without boundary!) disc, with centre at the point a = (c
1
. c
2
)
and of radius :. If : = 3. 1(a.:) is the common 31 open (without
boundary) ball (a full sphere!) with centre at a = (c
1
. c
2
. c
3
) and of
radius :. The closed ball 1[a.:] is exactly the full sphere of radius :
and with centre at a. which contains its boundary
o = ¦(r. n. .) : (r ÷c
1
)
2
+ (n ÷c
2
)
2
+ (. ÷c
3
)
2
= :
2
¦.
This last surface o is usually called the sphere of centre a and of radius
:.
Let 1 be an arbitrary subset of R
n
. A point d of 1 is said to be
interior in 1. if there is a small ball 1(d. :). : 0 centered at d such
that 1(d. :) · 1. All the interior points of 1 is a subset of 1 denoted
by 1:t1. the interior of 1. It can be empty. For instance, any …nite
set of points has an empty interior.
Definition 17. A subset 1 of R
n
is said to be an open subset if
for any a in 1 there is a small : 0 such that the open ball 1(a.:)
1. DISTANCE PROPERTIES IN R
m
115
with centre at a and of radius : is completely contained in 1. i.e.
1(a.:) · 1. A subset 1 of R
n
is said to be closed if its complementary
1
c
oc;
= R
n
r1
oc;
= ¦x ÷R
n
: x ÷1¦
in R
n
is an open subset of R
n
.
For instance, any point or any …nite set of points are closed subsets
of R
n
. If : = 1. the closed intervals are closed subsets of R. Moreover,
an open ball is an open set and a closed ball is a closed set (prove it!).
It is not di¢cult to prove that a subset 1 of R
n
is open if and only if
it is equal to its interior. The boundary E(1) of a subset 1 of R
n
is by
de…nition the collection of all the points b of R
n
such that ANY ball
1(b. :). centered at b and of radius : 0 has common points with 1
and with the complementary R
n
` 1 of 1. For instance, the boundary
of the disc ¦(r. n) : r
2
+ n
2
_ 1¦ is the circle ¦(r. n) : r
2
+ n
2
= 1¦
(prove it!). It is easy to see that 1 is closed if and only if it contains
its boundary. The set 1' E(1) is called the closure of 1. It is exactly
the union of all the limits of all convergent sequences which have their
terms in 1.
Remark 17. The set O of all the open subsets of R
n
has the fol
lowing basic properties:
1) ?. the empty set, and the whole set R
n
are considered to be in
O.
2) If 1
1
. 1
2
. .... 1
I
are in O, then their intersection
I
¨
j=1
1
j
is also
in O.
3) If ¦1
c
¦ is any family of open subsets in O, then their union
'
c
1
c
is also in O, i.e. it is also open. We propose to the reader
to prove all of these properties and to state and prove the analogous
properties for the set ( of all the closed subsets of R
n
. Mathematicians
say that a collection O of subsets of an arbitrary set `. which ful…l
the properties 1), 2) and 3) from above, gives rise to a topology on `.
For instance, in a metric space (A. d). the collection O of all the open
subsets (the de…nition is the same like that for R
n
!) gives rise to the
natural topology of a metric space of A. A set ` with a topology O on
it (a collection of subsets with the properties 1), 2) and 3)) is called
a topological space and we write it as (`. O). This notion is the most
general notion which can describe a "distance" between two objects in
`. For instance, if (`. O) is a topological space and if c is a "point"
(an element) of `. then an element / is said to be "closer" to c then
the element c. if there are two "open" subsets 1 and 1 of ` such that
116 6. THE NORMED SPACE R
m
:
c. / ÷ 1. c. c ÷ 1 and 1 · 1. Meditate on this fact in a metric space
A. for instance in the usual case A = R.
Now, if (A. d) is a metric space, the de…nition of an open ball 1(c. :)
with centre at an element c of A and of radius : 0 is similar to the
de…nition of the same notion in R
n
. Namely,
1(c. :) = ¦r ÷ A : d(r. c) < :¦.
In the same way, a subset 1 of A is said to be open in A if for any
c ÷ 1 there is an open ball 1(c. :) = ¦r ÷ A : d(r. c) < :¦. with
centre at c and of radius : 0. such that 1(c. :) · 1. A subset 1 of
A is called a closed set if its complementary 1 = A r 1 in A is an
open set of A.
Theorem 51. (a closeness criterion) A subset 1 of a metric space
(A. d) (in particular of A = R
n
) is closed if and only if any sequence
¦r
a
¦ of elements in 1. which is convergent to an element r of A. has
its limit r also in 1.
Proof. Let us assume that 1 is closed and let ¦r
a
¦ be a sequence
of elements in 1 which is convergent to an element r of A. If r were
not in 1 then, since 1 = A r 1 is open, we could …nd a ball 1(r. :)
with : 0. such that 1(r. :) · 1. i.e. 1(r. :)¨1 = ?. the empty set.
But, since r
a
÷ r. i.e. d(r
a
. r) ÷ 0. for : large enough, d(r
a
. r) < :.
or r
a
÷ 1(r. :). Since all the terms r
a
are in 1. we succeeded to …nd
at least one element r
a
÷ 1(r. :) ¨ 1 = ?. which is a contradiction.
So, r itself must be in 1.
Conversely, we suppose now that any sequence of elements of 1
which is convergent to an element r of A has its limit r in 1. If 1
were not closed, 1 = A r 1 were not open. This means that there
is at least one element n of 1 such that any small ball 1(n.
1
a
) cannot
be contained in 1. Hence, for any natural number : 0. one can
…nd at least one element n
a
÷ 1(n.
1
a
) ¨ 1 (why?). This means that
d(n
a
. n) <
1
a
and that n
a
÷ 1 for any : = 1. 2. ... . Since n
a
÷n (why?)
and since 1 has the above property, we see that n must be also in 1.
But,... n was chosen to be in 1 = A r 1. so it cannot be in 1! We
have a new contradiction! So, we cannot suppose that 1 is not open,
i.e. we are forced to say that 1 is closed and the theorem is completely
proved.
Definition 18. Let ¹ be a nonempty subset of R
n
(or of an arbi
trary metric space (A. d)). By the closure ¹ of ¹ in R
n
(or in A) we
mean the set of the limits of all the convergent sequences with terms in
¹.
1. DISTANCE PROPERTIES IN R
m
117
In particular, any element of ¹ is in ¹ (take the constant sequence
c. c. c. ... ,etc.). We can easily see that ¹ is the least closed subset of
A (in particular of R
n
) which contains ¹ (use Theorem 51). We also
see that if ¹ = ¦r
a
¦ is a sequence in A. then sup ¹ = limsup¦r
a
¦ and
inf ¹ = liminf¦r
a
¦.
Remark 18. ¹ is closed if and only if ¹ = ¹. The closure of
the open ball 1(c. :) in a metric space (A. d) is exactly the closed ball
1[c. :]. The operation ¹ ¹ has the following main properties: 1)
¹ ¨ 1 = ¹ ¨ 1, 2) ¹ ' 1 = ¹ ' 1. 3) ¹ ' E(¹) = ¹. where E(¹) =
¦r ÷ A : 1(r. :) ¨ ¹ = ? and 1(r. :) ¨ (A r¹) = ? for any : 0¦
is the boundary of ¹ in A (prove all these statements!).
We naturally extend the de…nition of a limit point for a subset ¹
of R (see De…nition 4) to a subset of an arbitrary metric space (A. d).
Let ¹ be a nonempty subset of a metric space (A. d) (in particular
of R
n
). An element r of A is said to be a limit point for ¹ if there is a
nonconstant sequence ¦r
a
¦ with terms in ¹ which is convergent to r.
For instance, (0. 0) is a limit point for the halfplane ¦(r. n) : n 0¦.
But (0. ÷0.0001) is not a limit point for the same subset in A = R
2
.
The subset ¦(:. :) : :. : ÷ N¦ of R
2
has no limit points. The set of
all the limit points of a subset ¹ of a metric space (A. d) together the
subset ¹ itself is exactly the closure ¹ of ¹ (why?). The set of all the
limit points of the closed cube ( = [0. 1] [0. 1] [0. 1] is the cube (
itself. But,...the set of all the limit points of an arbitrary closed subset
is not always the set itself. For instance, the set of all limit points of
a point c of A is the empty set which is distinct of ¦c¦. A sequence
¦r
a
¦ has exactly only one limit point r. if the sequence has an in…nite
distinct values and it is convergent to r.
Definition 19. A nonempty subset ¹ in a metric space (A. d) is
said to be bounded if there is a "reference" element c ÷ A and a positive
real number ` such that d(c. r) < ` for any element r of ¹.
Remark 19. It appears that the de…nition depends on the choice
of the "reference" element c. i.e. that the boundedness of ¹ is a c
boundedness. In fact, the de…nition does not depend on the element
c. Namely, if a subset ¹ is bounded relative to an element c of A.
it is bounded relative to any other element / of A. Indeed, d(/. r) _
d(/. c) + d(c. r) < d(/. c) + `. which is a …xed positive number w.r.t.
the variable element r of ¹. Hence, ¹ is also /bounded. In a normed
space (see De…nition 13) we take as a "reference" element c the element
c = 0. Thus, ¹ is bounded in a normed space (A. .) if and only if
there is a positive real number ` such that r < ` for any r of ¹.
118 6. THE NORMED SPACE R
m
:
CesaroBolzanoWeierstrass Theorem (see Theorem 12) has an ex
tension to R
n
for any : = 2. 3. ... .
Theorem 52. (BolzanoWeierstrass Theorem). Let ¹ be a bounded
and in…nite subset of R
n
. Then ¹ has at least one limit point in R
n
. In
particular, any bounded sequence in R
n
has a convergent subsequence.
Proof. To understand easier the idea behind the formal proof of
this theorem, we shall take the particular case : = 2 (the case : = 1
was considered in Theorem 12). So, ¹ is an in…nite (contains an in…nite
number of distinct elements) and bounded subset of R
2
. Any element of
¹ is a couple (r. n). where r. n ÷ R. Since ¹ is bounded by a positive
real number `. we can write (r. n) _ `. for any pair (r. n) of
¹. or
r
2
+n
2
_ `. Thus, the projections of ¹ on the coordinates
axes, ¹
1
= ¦c
1
÷ R : there is an c
2
÷ R with (c
1
. c
2
) ÷ ¹¦ and
¹
2
= ¦/
2
÷ R : there is a /
1
÷ R with (/
1
. /
2
) ÷ ¹¦ are bounded in
R (prove it and make a drawing!). Since ¹ is in…nite, at least one of
¹
1
or ¹
2
is in…nite (why?). We suppose that ¹
1
is in…nite. Let us
apply now CesaroBolzanoWeierstrass Theorem (Theorem 12) for the
subset ¹
1
of R. Hence, there is a limit point r
1
for ¹
1
. i.e. there is a
sequence ¦r
(a)
1
¦ of elements in ¹
1
. which is convergent to r
1
. Let us
look now at the de…nition of ¹
1
! For any r
(a)
1
. : = 1. 2. .... we can …nd
an element r
(a)
2
in R such that the couple (r
(a)
1
. r
(a)
2
) is in ¹. In fact,
the sequence ¦r
(a)
2
¦ is bounded and its terms belong to ¹
2
(why?). If
¹
2
is also in…nite, applying again CesaroBolzanoWeierstrass theorem
to the subset ¦r
(a)
2
¦. we get a limit point r
2
of this last sequence. This
means that we can …nd a subsequence ¦r
(In)
2
¦ of ¦r
(a)
2
¦ (/
1
< /
2
< ...
) which is convergent to r
2
. For any /
a
. : = 1. 2. .... we consider the
term r
(In)
1
of the sequence ¦r
(a)
1
¦ just found above. We obtain a new
sequence ¦(r
(In)
1
. r
(In)
2
)¦ of elements from ¹. which is convergent to the
pair (r
1
. r
2
) (why?...because it is componentwise convergent!). Thus
(r
1
. r
2
) is a limit point of ¹. What happens if ¹
2
is …nite? Then, at
least one term r
()
2
repeats itself of an in…nite number of times. We
suppose that for /
1
< /
2
< ... one has that r
(In)
2
= r
()
2
. for any
: = 1. 2. ... . So, the sequence ¦(r
(In)
1
. r
(In)
2
)¦. with terms in ¹. is
convergent to (r
1
. r
()
2
). which becomes in this way a limit point for
¹. A question can arise here: why can we choose all the elements of
the sequence ¦(r
(In)
1
. r
(In)
2
)¦ to be distinct one to each other? Because
the sequence ¦r
(a)
1
¦ can be chosen from the beginning to contain only
distinct elements (¹
1
is in…nite!). Hence, in both cases ¹ has a limit
point and the proof is completed.
1. DISTANCE PROPERTIES IN R
m
119
We shall see in the future the fundamental importance of this the
oretical result. A limit point is also called in the literature an accumu
lation point.
Since the bounded and closed subsets in a space of the form R
n
are very useful in many applications, we shall call them compact sets.
For instance, [c. /]. ¦(r. n) : r
2
+ n
2
_ :
2
¦ and, generally, any closed
balls, are all compact sets in their corresponding arithmetical spaces
of the type R
n
. A …nite union and any intersection of compact sets is
again a compact set (prove it!). An in…nite union of compact sets is not
always a compact set (…nd a counterexample!). For instance 1 = ¦
1
a
¦
is bounded but it is not closed because
1
a
÷ 0 and 0 is not in 1. So,
1 is not a compact set but,...its closure 1 = ¦0¦ ' ¦
1
a
¦ is a compact
subset in R (prove this!). Any …nite set of points in R
n
is a compact
set (why?).
Now we give a useful characterization of compact sets in R
n
.
Theorem 53. A subset ( of R
n
is a compact set if and only if any
sequence of ( contains a convergent subsequence with its limit in (.
Proof. We suppose that ( is a compact set in R
n
and let ¦x
(a)
¦ be
a sequence with terms in (. If ¦x
(a)
¦ has an in…nite number of distinct
elements, ¹ = ¦x
(a)
¦ being bounded (¹ · ( and ( is bounded), we
can apply Theorem 52 and …nd that there is a convergent subsequence
¦x
(In)
¦ of ¦x
(a)
¦. Since ( is closed, the limit of ¦x
(In)
¦ belongs to (
(see Theorem 51). If ¦x
(a)
¦ has only a …nite number of distinct terms,
one of them appears in an in…nite number of places. So, we take the
constant subsequence generated by it.
Conversely, we assume that ( has the property indicated in the
statement of the theorem. Let us prove …rstly that ( is bounded. If
it were not bounded, for any : = 1. 2. ... one can …nd a vector a
a
in (
such that a
a
 :. The hypothesis says that the sequence ¦a
a
¦ has
a convergent subsequence ¦a
In
¦. Let a = lim
a÷o
a
In
be the limit of the
sequence ¦a
In
¦. Then
/
a
< a
In
 _ a
In
÷a +a .
Taking limits in the extreme sides of these inequalities, we get: · _
a . a contradiction. Hence, ( must be bounded. Let us prove now
that ( is closed by using again Theorem 51. For this, let ¦y
a
¦ ÷y be
a convergent to y sequence with elements in ( and its limit y in R
n
.
By the hypothesis on (. the sequence ¦y
a
¦ has a subsequence ¦y
In
¦
which is convergent to an element z of (. Since ¦y
a
¦ is convergent to y,
any subsequence of ¦y
a
¦ is also convergent to y. Indeed, let us prove
120 6. THE NORMED SPACE R
m
:
for instance that z = y. For this, let us evaluate d(z. y), the distance
between z and y :
(1.6) d(z. y) _d(z. y
Im
) +d(y
Im
. y
a
) +d(y
a
. y).
where : and : are arbitrary chosen. If we make :. : ÷ · in this
last inequality, we get that d(z. y) =0. i.e. z = y (why?). Here we just
used the fact that a convergent sequence is also a Cauchy sequence, i.e.
for :. : large enough, the distance d(y
n
. y
a
) goes to zero. Now, since
z is in ( we get that y is also in (. i.e. ( is closed and the theorem is
proved.
The above characterization of compact subsets of R
n
leads us to the
introduction of the notion of a compact subset in an arbitrary metric
space (A. d). We say that a subset ( of A is compact if any sequence of
elements from ( has a subsequence which is convergent to an element
of (.
For instance, any convergent sequence ¦r
a
¦ in a metric space A.
together with its limit r is a compact subset of A (prove it!). Thus,
( = ¦r
a
¦ ' ¦r¦ is a compact subset of A.
2. Continuous functions of several variables
Let ¹be a nonempty subset of R
a
. the "arithmetical" :dimensional
vector space and let 1 : ¹ ÷R. be a function de…ned on ¹ with values
in R. Since the variable x = (r
1
. r
2
. .... r
a
) is a vector determined by
: free scalar quantities, r
1
. r
2
. .... r
a
. we say that our function is a
function of : variables. If : _ 2. we say that 1 is a function of
"several" variables. Since the values of 1 are scalars (real numbers),
we say that 1 is a scalar function of : variables. A map f : ¹ ÷R
n
is
called a vector function of : variables. This time, the values of f are
:dimensional vectors. Hence f (x) = (n
1
. n
2
. .... n
n
) and we see that
the numbers n
1
. n
2
. .... n
n
are themselves functions 1
1
. 1
2
...., 1
n
of x:
n
1
= 1
1
(x). .... n
n
= 1
n
(x). These scalar functions 1
1
. 1
2
. .... 1
n
. de…ned
on ¹ with values in R this time, are called the components of f . We
write this as: f =(1
1
. 1
2
. .... 1
n
) and interpret it as a "vector" of :
components (coordinates) 1
1
. 1
2.
.... 1
n
. In applications f is also called
a vector …eld of : variables. "Field" comes from "…eld of forces". For
instance,
f :R
2
÷R
2
. f (r. n) = (rn. r ÷n)
is a vector …eld in plane (R
2
) of 2 variables. Its components are
1
1
(r. n) = rn and 1
2
(r. n) = r÷n. We can give its image in some points.
For instance, we can translate the vector f (2. 3) = (23. 2÷3) = (6. ÷1)
in the point (2. 3) and so we get "the image" of f at (2. 3). In this way
2. CONTINUOUS FUNCTIONS OF SEVERAL VARIABLES 121
we can …ll the whole plane R
2
with vectors (forces), i.e. we get a
"…eld" of forces on the whole plane. If : = 1, the image of a vec
tor …eld f : ¹ ÷ R
n
(¹ · R) is a "curve" in R
n
. For instance,
f (t) = (1cos t. 1sin t). t ÷ [0. 2:) has as image in the plane R
2
the
usual circle of radius 1 and with centre at the origin (0. 0). We say
that the two components of f , 1
1
(t) = 1cos t and 1
2
(t) = 1sin t are the
parametric equations of this circle. One also write this as: r = 1cos t.
n = 1sin t. t ÷ [0. 2:). We can also interpret the image of a vector …eld
f : [0. 1] ÷R
n
(: = 2 or : = 3) as the trajectory of a moving point
`(1
1
(t). 1
2
(t). .... 1
n
(t))
where t measures the "time" between the starting moment (usually
t = 0) and the ending moment t = 1. For instance, f (t) = (t. t
2
).
t ÷ ¹ = [0. 10]. is a parabolic trajectory, along the arc of the parabola
n = r
2
. r ÷ [0. 10]. The new vector …eld
f
t
(t) = (1
t
1
(t). 1
t
2
(t). .... 1
t
n
(t))
(the componentwise derivative), associated to the vector …eld
f (t) = (1
1
(t). 1
2
(t). .... 1
n
(t)). t ÷ [0. 1].
is called the velocities …eld of the …eld f .
In order to describe the "breaking" phenomena at a given point
a =(c
1
. c
2
. .... c
a
) of R
a
. we need to see what happens with the values
of a vector function (which describes our phenomenon) f : ¹ ÷ R
n
.
whenever we becomes closer and closed to a. For this, a must be a limit
point of the de…nition domain ¹. We have to study the convergence of
the sequence of vectors ¦f (x
(a)
)¦ in R
n
, whenever the sequence ¦x
(a)
¦,
with terms in ¹. converges to a in the metric space R
a
. The most
convenient situation is that when all the values ¦f (x
(a)
)¦. for all the
sequences ¦x
(a)
¦. which are convergent to a. become closer and closer
to one and the same vector L from R
n
. This is why we give now the
following de…nition.
Definition 20. Let ¹ be a subset of R
a
and let a =(c
1
. c
2
. .... c
a
)
be a limit point of ¹. We say that L ÷ R
n
is the limit of a vector
function f : ¹ ÷ R
n
at the point a (write L =lim
x÷a
f (x)), if for every
sequence ¦x
(a)
¦ in ¹. which is convergent to the vector a. one has that
the sequence of images ¦f (x
(a)
)¦ of ¦x
(a)
¦ through f is convergent to L.
If such an L exists, independently on the choice of the sequence ¦x
(a)
¦,
we say that f has limit L at a. This limit L depends only on f and on
a.
122 6. THE NORMED SPACE R
m
:
If there is such a common limit L. this is unique, because, in general,
the limit of a sequence in a metric space is unique (if it exists!).
For instance, let us compute lim
(a.&)÷(÷1.2)
1(r. n). where
1(r. n) = rn +r
2
+ ln(r
2
+n
2
).
Let us take a sequence ¦(r
a
. n
a
)¦ which is convergent to (÷1. 2). This
means that r
a
÷ ÷2 and n
a
÷ 2 (see Theorem 50). But we know
that the "taking limit" operation is compatible with the multiplication,
addition and with the logarithm function (we say that ln is continuous!)
(see also Theorem 14). Hence,
1(r
a
. n
a
) = r
a
n
a
+r
2
a
+ ln(r
2
a
+n
2
a
)
will be convergent to
(÷1) 2 + (÷1)
2
+ ln((÷1)
2
+ 2
2
) = ÷1 + ln 5.
We see that this limit is independent on the starting sequence (r
a
. n
a
)
which tends to (÷1. 2). Thus, for any sequence (r
a
. n
a
) which is con
vergent to (÷1. 2).
lim
(a.&)÷(÷1.2)
1(r
a
. n
a
) = ÷1 + ln 5.
In fact, we see that for any sequence (r
a
. n
a
) which is convergent to
(÷1. 2),
lim
(a.&)÷(÷1.2)
1(r
a
. n
a
) = 1(÷1. 2).
This happens, because any elementary function of several variables is
"continuous" (see the bellow de…nition) on its de…nition domain.
Definition 21. Let ¹ be a subset of R
a
and let a =(c
1
. c
2
. .... c
a
) be
a point of ¹. We say that the vector function f : ¹ ÷R
n
is continuous
at the point a, if for every sequence ¦x
(a)
¦ of ¹. which is convergent
to the vector a. one has that the sequence of the images ¦f (x
(a)
)¦ of
¦x
(a)
¦ through f is convergent to f (a). the value of f at a. We say that
f is continuous on the set ¹ if f is continuous at any point of ¹.
We see that f is continuous at a point a if and only if it has a
limit L at a and this L is equal to f (a). the value of f at the point a.
The above de…nition is in accordance with the engineers perception of
approximation processes. Let us suppose that f describes a physical
phenomenon 1 and we are interested in the variation of this phenome
non around a …xed "point" (vector) a. Let us take a neighboring point
z of a and let us approximate z by a. In this case, can we approximate
f (z) by f (a)? Or, can we consider that 1 is "almost the same" at z
like at a?. We can do this if f is continuous at a. Otherwise, we cannot
2. CONTINUOUS FUNCTIONS OF SEVERAL VARIABLES 123
do such approximations. We must be very careful for instance, in the
case of earthquake models around the so called "singular" points (see
the example bellow). Now we think that the reader is convinced that
the continuity notion is important in the modelling of the physical phe
nomena. It is not di¢cult to prove that all the elementary functions
and their compositions are continuous functions. In the following we
supply with an example in which we shall see that the case of vector
…elds of several variables (for : 1) is more complicated then the case
of one variable. Let us see now if the following nonelementary (why?)
function
1(r. n) =
a&
a
2
+&
2
. if r = 0. or n = 0.
0. if r = 0 and n = 0.
1 : R
2
÷ R, is continuous or not on the whole R. If (c. /) = (0. 0).
then 1(r. n) =
a&
a
2
+&
2
on a small disc (not containing (0. 0)) with centre
at (c. /) (and a small radius). Since the restriction of 1 to this last disc
is an elementary function, 1 is continuous at (c. /). What happens at
(0. 0)? If the function 1 were continuous at (0. 0) then, for any sequence
(r
a
. n
a
) which tends to (0. 0) (i.e. r
a
÷0 and n
a
÷0), we should have
that 1(r
a
. n
a
) ÷1(0. 0) = 0. Let us take a nonzero real number : and
let ¦r
a
¦ be an arbitrary sequence of nonzero real numbers which is
convergent to 0. Take now n
a
= :r
a
for any : = 1. 2. .... This means
that all the pairs (r
a
. n
a
) are on the line n = :r (its slope is :) and
that the sequence ¦(r
a
. n
a
)¦ is convergent to (0. 0). But
1(r
a
. n
a
) =
:r
2
a
r
2
a
+:
2
r
2
a
=
:
1 +:
2
= 0.
So the function 1 is not continuous at (0. 0). Moreover, since the limit
lim
(an.&n)÷(0.0)
1(r
a
. n
a
) =
:
1 +:
2
is dependent on the slope : of the line n = :r. on which we have
chosen our sequence (r
a
. n
a
). we see that the function 1 has no limit
at (0. 0). Hence, we cannot extend 1 "by continuity" at (0. 0) with no
real value. Such a point (0. 0) is called an essential singular point for
1. This means that if we become closer and closer to (0. 0) on di¤erent
sequences ¦(r
a
. n
a
)¦. we obtain an in…nite number of distinct values
for the limit lim
(an.&n)÷(0.0)
1(r
a
. n
a
) (as we just saw above!).
The following criterion reduces the study of the limit or of the
continuity of a vector function f : ¹ ÷R
n
at a point a ÷¹. where ¹ is
an open subset of R
n
and f = (1
1
. 1
2
. .... 1
n
). to the study of the same
properties for the scalar functions 1
1
. 1
2
. .... 1
n
.
124 6. THE NORMED SPACE R
m
:
Theorem 54. With these last notation, 1) f = (1
1
. 1
2
. .... 1
n
) has
the limit L = (1
1
. 1
2
. .... 1
n
) at the point a if and only if every com
ponent function 1
;
has the limit 1
;
at the same point a. for , = 1. 2. ...
and 2) f is continuous at the point a if and only if every component
function 1
;
is continuous at a.
Proof. Everything comes from the fact that the convergence in
the normed spaces R
n
is a componentwise convergence (see Theo
rem 50). Indeed, let us assume that f = (1
1
. 1
2
. .... 1
n
) has the limit
L = (1
1
. 1
2
. .... 1
n
) at a. Hence, for any sequence ¦(x
(a)
)¦ which is
convergent to a. one gets that limf (x
(a)
) = L. i.e. lim1
;
(x
(a)
) = 1
;
for , = 1. 2. ... (we just applied the "componentwise" principle). The
existence is included here! (why?). Conversely, if for any , = 1. 2. ....
the limit lim1
;
(x
(a)
) = 1
;
exists, then the limit limf (x
(a)
) = L exists
and L = (1
1
. 1
2
. .... 1
n
). We add the fact that f = (1
1
. 1
2
. .... 1
n
) is
continuous at a if and only if
L = (1
1
. 1
2
. .... 1
n
) = f (a) = (1
1
(a). 1
2
(a). .... 1
n
(a)).
or if and only if 1
;
(a) =1
;
for any , = 1. 2. ... . But this means exactly
the continuity of every 1
;
at a for , = 1. 2. ... .
Using this last continuity test, we can easily decide if a vector func
tion is continuous or not. For instance,
f (r. n. .) = (r. 2r +n. 2r + 3n ÷2.)
is continuous on R
3
because all the scalar component functions
1
1
(r. n. .) = r. 1
2
(r. n. .) = 2r + 3n
and 1
3
(r. n. .) = 2r +3n ÷2. are polynomial functions so, they are all
continuous on R
3
.
Remark 20. The existence of a limit at a point and the continuity
at a point are "local" properties. They are de…ned "around" a given
point a. If we …x a :1 continuous curve : [c. /] ÷ ¹ · R
a
and
if a =(t
0
) is a point "on " (it is in the image of ), we say that a
vector function f = (1
1
. 1
2
. .... 1
n
). de…ned on ¹ with values in R
n
is
continuous at a along the curve if the composed function f · : [c. /] ÷
R
n
(a new curve in R
n
) is continuous at t
0
. This means that if we take
any sequence of points ¦x
(a)
¦ in ¹ (is considered opened!) on (x
(a)
=
(t
a
)), which becomes closer and closer to a. then limf (x
(a)
) = f (a).
For instance,
1(r. n) =
a&
a
2
+&
2
. if r = 0. or n = 0.
0. if r = 0 and n = 0.
2. CONTINUOUS FUNCTIONS OF SEVERAL VARIABLES 125
1 : R
2
÷ R. is not continuous at a = (0. 0). but it is continuous at
(0. 0) along the both axes of coordinates. It has limits along any other
…xed line n = :r which is passing through (0. 0). but the limits are not
the same! (see the above commentaries on this example). A question
arises: If a scalar function is continuous along any continuous curve
(is it su¢cient to take only lines?) which are passing through a given
point a. is it continuous at a? (Conversely it is clear!why?).
Theorem 55. The composition between two continuous functions
is also a continuous function.
Proof. Let ¹ be an open subset of R
j
. let 1 be another open sub
set of R
a
and let f : ¹ ÷1. g : 1 ÷R
n
be two continuous functions
on their de…nition domains. The theorem says that the composed func
tion h : ¹ ÷R
n
. h = g · f . i.e. h(x) = g(f (x)) for any x ÷ ¹. is also
a continuous function on ¹. For proving this, let us take a point a ÷ ¹
and an arbitrary sequence ¦x
(a)
¦ in ¹ which is convergent to a w.r.t.
the distance of R
j
. Since f is continuous on ¹. in particular, it is also
continuous at a. So, the sequence ¦f (x
(a)
)¦ is convergent to f (a). Now,
since g is continuous on 1. in particular, it is continuous at the point
f (a) of 1. Hence, the sequence ¦g(f (x
(a)
))¦ tends to g(f (a)) = h(a)
and so, h(x
(a)
)= g(f (x
a)
)) is convergent to h(a). This means that the
composed function h is continuous at a. Since a was arbitrary chosen
in ¹, we have that h is continuous on the whole ¹.
This theorem is very useful, because almost all the functions com
monly used in applications are compositions of elementary functions
and these last ones are continuous on their de…nitions domains. For
instance,
1(r. n) = cos
¸
r + sin rn
1 + ln(r
2
+n
2
)
is de…ned on R
2
`. where is the circle: r
2
+n
2
=
1
c
. where c = 2.71... .
Here 1 is the composition between the following continuous functions:
r cos r. (r. n)
r
n
. n = 0. (r. n) r +n. (r. n) rn.
r sin r and r ln r. r 0
(prove everything slowly!). The same theorem is used to prove that the
set of all continuous functions de…ned on the same set ¹ (open, closed,
etc.) is a real in…nite dimensional (contains polynomials!) vector space
(prove it!).
126 6. THE NORMED SPACE R
m
:
3. Continuous functions on compact sets
Let ¹ be an arbitrary nonempty subset of R
a
and let f : ¹ ÷ R
n
be a continuous function (on the whole ¹). Let 1 be an open subset of
R
a
which is contained in ¹. Here is a question: "Is always the image
f (1) of 1 through f open in R
n
? We shall see by simple examples that
the answer is NO! Let us take, for instance, 1 = (0. 1) and 1(r) = 3
for any r in (0. 1). Since the set ¦3¦ is closed in R (why?), 1(1) is not
open. Let now 1 be an open subset of R
n
and f
÷1
(1) = ¦x ÷ ¹ :
f (x) ÷ ¹¦. the preimage of 1 in ¹. We say that a subset 1 of ¹ is
open in ¹ if it is the intersection between ¹ and an open subset 1 of
R
a
. i.e. 1 = ¹ ¨ 1. For instance, 1 = (0. 1] is not open in R (why?),
but it is open in ¹ = [÷1. 1] because, 1 = (0. 3). which is open in R,
intersected with ¹ is exactly 1.
Theorem 56. With the de…nitions and notation given above, f :
¹ ÷R
n
is continuous if and only if f
÷1
(1) is open in ¹ for any open
subset of R
n
. i.e. if f carries back the open subsets of R
n
into open
subsets of ¹.
Proof. a) We assume that f : ¹ ÷ R
n
is continuous and that
1 is an open subset of R
n
. To prove that f
÷1
(1) is open in ¹ it is
equivalent to prove that ( = ¹`f
÷1
(1) is closed in ¹. i.e. for any
convergent sequence ¦x
(a)
¦ of elements in (. convergent to an element
x of ¹ (pay attention!), one has that x is also in (. If it were not in
(. f (x) ÷ 1. Since 1 is open in R
n
. there is a small ball 1(f (x).:).
with center at f (x) and of radius : 0, which is contained in 1. Since
x
(a)
÷x, and since f is continuous, one has that f (x
(a)
) is convergent
to f (x). So, there is at least one x
(a
0
)
with f (x
(a
0
)
) in 1(f (x).:). i.e. in
1. So, x
(a
0
)
is in f
÷1
(1). a contradiction, because we have chosen the
sequence ¦x
(a)
¦ to have all its terms in (. i.e. not in f
÷1
(1).
b) We suppose now that f carries back the open subsets of R
n
into
open subsets of ¹. Let us prove that f is continuous at an arbitrary …xed
point z. For this, let ¦z
(a)
¦ be a sequence in ¹ which is convergent to
z ÷ ¹. We assume that ¦f (z
(a)
)¦ is not convergent to f (z). Then, there
is a small ball 1(f (z).:) in R
n
such that an in…nite number ¦f (z
(In)
)¦
, : = 1. 2. .... of the terms of the sequence ¦f (z
(a)
)¦ are outside of
1(f (z).:). Since 1(f (z).:) is an open subset in R
n
. following the last
hypothesis, we get that the set 1 = f
÷1
(1(f (z).:)) is an open subset
of ¹ which contains z (why?). Let 1(z. :
t
). :
t
0 be a small ball with
centre in z such that G = 1(z. :
t
) ¨¹ · 1 (since 1 is open in ¹). All
the terms of the subsequence ¦z
(In)
¦ are not in G. in particular they
are not in 1(z. :
t
). But this last conclusion contradicts the fact that
3. CONTINUOUS FUNCTIONS ON COMPACT SETS 127
z
(a)
÷z. Thus, our assumption that ¦f (z
(a)
)¦ is not convergent to f (z)
is false and so, f is continuous at z. Since this z was arbitrary chosen,
we get that f is continuous at all the points of ¹.
The following result is very useful in many situations of this course.
It appears as a direct consequence of the above theorem.
Theorem 57. Let ¹ be an open subset of R
a
. let a be a …xed point
of ¹ and let 1 : ¹ ÷ R be a continuous function on ¹ such that
1(a) 0. Then there is an open ball 1(a. :) · ¹. : 0. with the
property that 1(x) 0 for every x in 1(a. :).
Proof. Take 0 such that 1(a)÷ 0 and take the open subset
) = (1(a) ÷ . 1(a) + ) of R. Since 1 is continuous, A = 1
÷1
() ) is
an open subset of ¹ which contains a. So, there is a small ball 1(a. :)
such that 1(a. :) · A. i.e. 1(x) ÷ ) for any x in 1(a. :). But, for
such x we have that 1(x) 1(a) ÷ 0 and the proof is done.
Remark 21. In the same way one can prove that f : ¹ ÷ R
n
is
continuous if and only if f carries back the closed subsets of R
n
into
closed subsets of ¹ (de…ne this notion by analogy!). To prove this, one
can use the last theorem 56.
Not always a continuous function carries a closed set of R
a
in a
closed set of R
n
. For instance, 1 : R ÷ R. 1(r) =
1
1+a
2
. carries the
closed set [0. ·) into (0. 1]. which is not closed more. It is interesting
to see that the closed set [0. ·) in unbounded. If one tries to substitute
it with a closed and bounded interval, for the same function, we shall
not succeed at all to …nd like an image a non closed set! Why? Because
of the following basic result:
Theorem 58. Let ( be a compact (closed and bounded) subset of
R
a
and let f : ( ÷ R
n
be a continuous function. Then, the image
f (() of (. in R
n
. is also a compact subset there (in R
n
). Moreover, if
: = 1. sup f (() = f (z
A
) and inf f (() = f (z
n
). where z
A
, z
n
are in
(.
Proof. We need to prove that: a) f (() is bounded and, b) f (()
is closed. The ideas used for proving this theorem are exactly the same
like those used in the particular case of Theorem 32. We take them
again here.
a) We assume that f (() is not bounded. This means that for every
: = 1. 2. .... one can …nd a point x
(a)
in ( such that
f (x
(a)
)
:
(why?). Since ( is a compact subset in R
a
. we can …nd a conver
gent subsequence ¦x
(In)
¦ to the point x of ( (see Theorem 53). Since
128 6. THE NORMED SPACE R
m
:
f : ( ÷ R
n
is continuous, the sequence ¦f (x
(In)
)¦ is convergent to
f (x). But
f (x
(In)
)
/
a
and /
a
÷ ·. so, the numerical sequence
¦
f (x
(In)
)
¦ is unbounded (goes to ·!). We shall see that this is a
contradiction. Indeed,
f (x
(In)
)
_
f (x
(In)
) ÷f (x)
+f (x) .
If we take limits in this last inequality, we get: ·_ 0 +f (x) . which
is not possible! The contradiction appeared because we supposed that
f (() is unbounded. Hence, it is bounded, i.e. we just proved a).
b) We use now the closeness test (Theorem51) for proving that f (()
is closed. Let us take for this a convergent sequence ¦f (y
(a)
)¦, with
terms in f (() and with its limit c in R
n
. We have to prove that this c
is also in f ((). Since ( is a compact subset of R
a
. there is a subsequence
¦y
(In)
¦ of the sequence ¦y
(a)
¦ such that y
(In)
is convergent to y ÷ (.
Since f is continuous, the sequence ¦f (y
(In)
)¦ is convergent to f (y). But
any subsequence of a convergent sequence is also convergent to the same
limit of the whole sequence. Thus, c = f (y) and so, c ÷ f ((). what we
wanted to prove. The other statements can be proved exactly in the
same manner (see also Theorem 32).
Let us give a nice application to this last result. We can assume
that the surface of the Earth is closed and bounded in the 31 space R
3
(why?you can take it for easy to be o = ¦(r. n. .) : r
2
+n
2
+.
2
= 1
2
¦.
...a sphere of radius 1. etc.; prove that o is closed and bounded!). At a
…xed moment, to any point `(r. n. .) from the Earth we associate its
temperature 1(r. n. .) at that moment. Thus, we obtain a continuous
function 1 de…ned on the compact surface of the Earth, with values in
R. Applying the above theorem, we always can …nd two points on the
Earth in which the temperatures are extreme.
Let ( be a compact (closed and bounded) subset of R
a
and let
f : ( ÷ R
n
be a continuous function. Then, the norm f (() of the
image f (() of (. in R. is also a compact subset there (in R). Moreover,
sup f (() = f (z) and inf f (() = f (y) . where z and y are in
(. Firstly, the function
g : R
n
÷R. g(x) =x .
is a continuous function. Indeed, let ¦x
(a)
¦ be a sequence in R
n
. which
is convergent to x. Since
x
(a)
÷x
_
x
(a)
÷x
. we see that the
sequence ¦g(x
(a)
) = ¦
x
(a)
¦ is convergent to x . i.e. g is continuous.
Secondly, let us consider the composition g · f : ( ÷ R between the
3. CONTINUOUS FUNCTIONS ON COMPACT SETS 129
continuous functions f and g. It is a continuous function (see Theorem
55) and we can apply the last theorem (do it slowly!).
Remark 22. The condition on the closeness of ( in the above
theorem (Theorem 58) is necessary as one can see in the example:
1 : (0. 1] ÷R, 1(r) =
1
a
; this function is continuous (prove it!), the in
terval (0. 1] is bounded, nonclosed and the image 1((0. 1]) = [1. ·)
is not bounded, so not a compact subset of R. If ( is closed but
not bounded, its image through a continuous function 1 may be non
closed and nonbounded at the same time. For instance, ( = [1. ·).
1(r) =
1
a÷1
, so, 1(() = (0. ·). which is neither closed (it is open
in R), nor bounded. This theorem above is not true in general metric
spaces. Because a compact subset ( in a general metric space (A. d) is
de…ned "by sequences". Namely, ( is a compact subset of (A. d) if any
sequence in ( has a convergent subsequence with its limit also in (.
This is not generally equivalent to "bounded and closed". The exam
ples are two "exotic" and we do not give them here. In a metric space
(A. d) we can introduce the "distance" between two compact subsets ¹
and 1 of A. Namely,
di:t(¹. 1) = inf¦d(c. /) : c ÷ ¹. / ÷ 1¦.
Since d is a continuous function this number di:t(¹. 1) is always …
nite and it is realized, i.e. there are c
0
in ¹ and /
0
in 1 such that
di:t(¹. 1) = d(c
0
. /
0
). For instance, the distance between the full square
¹ = [0. 1] [1. 2] and the disc 1 = ¦(r. n) : (r ÷2)
2
+n
2
_ 1 is
2 ÷1
and it is realized at c
0
= (1. 1) ÷ ¹ and at /
0
= (2 ÷
1
2
.
1
2
) (why?).
It is easy to prove that the distance between two compact subsets ¹ and
1 is realized on their boundaries (which are also compact subsets), i.e.
di:t(¹. 1) = di:t(E(¹). E(1)).
Can you organize the set of all compact subsets of A as a metric space
(with the distance function de…ned above)?
In practice, the above Theorem 58 can be applied to optimization
problems. For instance, let us …nd the maximal and the minimal values
of the function 1 : [0. 1] [0. 2] ÷ R, 1(r. n) = r
4
+ n
4
. Since ( =
[0. 1] [0. 2] is a compact subset in R
2
(prove it!), Theorem 58 implies
that its image is a compact subset of R. So, sup 1(() = 1(a) and
inf 1(() = 1(b). It is easy to see that a = (1. 2) and b = (0. 0) (the
function is increasing relative to r and n, separately).
An useful notion in the integral computation (and not only!see the
bellow application) is the notion of "uniform continuity".
130 6. THE NORMED SPACE R
m
:
Definition 22. Let ¹ be a nonempty subset of R
a
and let f : ¹ ÷
R
n
be a function de…ned on ¹ with values in R
n
. We say that f is
uniformly continuous on ¹ if for any small quantity 0. there is
another small quantity o
.
0 (depending on ) such that whenever we
have two points x
t
and x
tt
in ¹ with the distance x
t
÷x
tt
 between
them less then o
.
, the distance
f (x
t
) ÷f (x
tt
)
between their images is
less then .
The word "uniform" reefers to the fact that here the continuity is
not de…ned at a point, but on the whole ¹. Moreover, the variation
f (x
t
) ÷f (x
tt
)
of f (x) is uniform relative to the variation x
t
÷x
tt

of x. Thus, if we want that the variation of f (x) to be less than 0.001
(
f (x
t
) ÷f (x
tt
)
< 0.001) in the case of an uniform continuous func
tion f . we can …nd a constant o = o
0.001
0 such that anywhere
a
t
and a
tt
would be in ¹. with the distance between them less than
this last constant o. we are sure that the corresponding variation of f .
f (a
t
) ÷f (a
tt
)
is less then 0.001.
Remark 23. The notion of uniform continuity is stronger then the
"simple" continuity. Indeed, let f : ¹ ÷R
n
be a uniformly continuous
function on ¹ and let a be a …xed point in ¹. We shall prove that f is
continuous at a. For this, let ¦a
(a)
¦ be a convergent sequence to a in ¹.
We want to prove that the sequence ¦f (a
(a)
)¦ is convergent to f (a) by
using only the de…nition of the convergence. In fact, we want to prove
that the numerical sequence ¦d(f (a
(a)
). f (a))¦ tends to zero. Now we
use the usually De…nition 1. For this, let 0 be a small positive real
number. Since f is uniformly continuous, there is a o
.
0 such that
whenever x
t
÷x
tt
 < o
.
. one has that
f (x
t
) ÷f (x
tt
)
< .
Let us take now x
tt
to be a and x
t
= a
(a)
. with : _ `. this last `
chosen such that
a
(a)
÷a
< o
.
. Thus,
f (a
(a)
) ÷f (a)
< .
whenever : _ ` and so, we have just proved that the sequence ¦f (a
(a)
)¦
is convergent to f (a). i.e. f is continuous at an arbitrary chosen point
a.
But continuity does not always imply uniform continuity. For in
stance, 1(r) = ln r. r ÷ (0. 1]. is a continuous function and not a uni
formly continuous one. Indeed, let the sequences r
t
a
=
1
a
and r
tt
a
=
1
2a
.
It is clear that [r
t
a
÷r
tt
a
[ =
1
2a
÷ 0, but [ln r
t
a
÷ln r
tt
a
[ = ln 2 9 0.
3. CONTINUOUS FUNCTIONS ON COMPACT SETS 131
Thus, if we take < ln 2 in De…nition 22, we can NEVER …nd a small
o
.
0 such that for all pairs (r
t
. r
tt
) with [r
t
÷r
tt
[ < o
.
one has
[ln r
t
÷ln r
tt
[ < < ln 2.
To see this, let us take :
0
large enough such that
r
t
a
0
÷r
tt
a
0
=
1
2:
0
< o
.
.
For the pair (r
t
a
0
. r
tt
a
0
).
ln r
t
a
0
÷ln r
tt
a
0
= ln 2.
which is greater than . so the de…nition of the uniform continuity does
not work for this function.
The next result says that for the functions de…ned on compact sets,
continuity and uniform continuity coincide. Pay attention, in our case
above (0. 1] in not compact! This is way we could prove that 1(r) = ln r
is not uniformly continuous.
Theorem 59. Let ( be a compact subset of R
a
and let f : ( ÷R
n
be a continuous function de…ned on (. Then f is uniformly continuous
on (.
Proof. We suppose on contrary, namely that f is not uniformly
continuous on (. We must carefully negate the statement of De…nition
22. Thus, there is an
0
0 such that for any small enough o 0 there
is at least one pair (x
t
c
. x
tt
c
) with elements in ( such that x
t
c
÷x
tt
c
 < o
and
f (x
t
c
) ÷f (x
tt
c
)
_
0
.
In particular, let us take for these o. o
I
=
1
I
for / = 1. 2. ... . Like
above, for such o
I
. / = 1. 2. .... one can …nd two sequences ¦x
t(I)
¦ and
¦x
tt(I)
¦ with
x
t(I)
÷x
tt(I)
<
1
I
and
f (x
t(I)
) ÷f (x
tt(I)
)
_
0
0.
Since ( is a compact set, we can …nd two subsequences: ¦x
t(It)
¦ of
¦x
t(I)
¦ and ¦x
tt(It)
¦ of ¦x
tt(I)
¦ (why can we take the same /
t
for both
subsequences?) such that these both subsequences are convergent to
the same limit y ÷ ( because
x
t(It)
÷x
tt(It)
<
1
/
t
÷0.
Since f is continuous, one has that the both sequences ¦f (x
t(It)
)¦ and
¦f (x
tt(It)
)¦ are convergent to the same limit f (y). So the distance be
tween the corresponding terms becomes smaller and smaller as : ÷·.
132 6. THE NORMED SPACE R
m
:
i.e.
f (x
t(It)
) ÷f (x
tt(It)
)
÷0.
a contradiction, because
f (x
t(It)
) ÷f (x
tt(It)
)
is always greater or equal
to
0
. Thus, our assumption on the nonuniform continuity of f is false.
Hence, f is uniformly continuous.
This result is very useful in practice. For instance, the function
1(r) = ln r is uniform continuous on any closed interval [c. /] · (0. ·).
Indeed, [c. /] is a compact subset in the de…nition domain (0. ·) of 1.
1 is continuous on [c. /] and so we can apply the above Theorem 59.
Example 13. Let ( be a 31object (( · R
3
), bounded and con
taining its boundary ·(. like usually in practice. We know that ( is
closed if and only if it contains its boundary ·(. Let us assume that in
any point `(r. n. .) of ( we have a density 1(r. n. .). It is commonly to
suppose that the density function 1 : ( ÷R is a continuous function.
The above theorem and our hypotheses on ( say that 1 is uniformly
continuous. We cannot practically work with this function because no
body gives it us in advance. But we can perform some measurements.
How do we perform such measurements 1(r
j
. n
j
. .
j
). i = 1. 2. .... :. such
that if we chose a point `(r. n. .) in (. we can …nd i
0
with
[1(r. n. .) ÷1(r
j
0
. n
j
0
. .
j
0
)[ <
(this is a small positive real number which controls the error, for in
stance = 11000). Since our function is uniformly continuous, there
is a small o 0 such that whenever the distance between two points
x
t
= (r
t
. n
t
. .
t
) and x
tt
= (r
tt
. n
tt
. .
tt
) of ( is less than this o. we have
that
[1(r
t
. n
t
. .
t
) ÷1(r
tt
. n
tt
. .
tt
)[ < .
It remains to us to divide the body ( into subbodies (
j
. i = 1. 2. .... :.
such that ( =
j=a
'
j=1
(
j
and the diameters
.
j
= sup¦x
t
÷x
tt
 : x
t
. x
tt
÷ (
j
¦
of (
j
are less then o. Let us choose now a …xed point `
j
(r
j
. n
j
. .
j
) in
each (
j
for i = 1. 2. .... :. Then the approximation
1(r. n. .) t 1(r
j
. n
j
. .
j
)
is a good one if `(r. n. .) ÷ (
j
. This means that
[1(r. n. .) ÷1(r
j
. n
j
. .
j
)[ < .
Thus, we can perform measurements of the density function values only
in some arbitrarily chosen points `
j
in each (
j
.
4. CONTINUOUS FUNCTIONS ON CONNECTED SETS 133
We give here a very useful result, in a more general setting (de…ne
and prove things slowly!).
Theorem 60. Let A and ) be two compact metric spaces (recall
that a metric space is compact if any sequence of it has at least one
convergent subsequence) and let 1 : A ÷ ) be a continuous bijection
from A on ). Let o : ) ÷A be its inverse. Then o is also continuous.
Proof. Let us prove that o carries back closed subsets of A into
closed subsets of ) (see Remark 21). Let ( be a closed subset of A
and let 1 = o
÷1
(() = 1((). Since ) is compact, ( is also compact
(prove it!). Since 1 is continuous, 1 = 1(() is compact, so 1 itself is
closed in ) (prove it!). Hence, o is continuous.
Corollary 5. Let 1 be a strictly monotone continuous function
which carries the interval [c. /] onto the interval [c. d] (see also the next
section, Darboux’ theorem). Then 1 is inversable and its inverse o is
also continuous.
Proof. Since 1 is strictly monotone it is onetoone (injective).
Since both intervals are compact metric spaces, we simply apply the
previous result. Here, "onto" means surjectivity!.
4. Continuous functions on connected sets
Let ¹ be a subset of R
a
. A continuous curve in ¹ is a vector con
tinuous function : 1 ÷ ¹. de…ned on an interval 1. …nite or not,
opened or not, closed or not. In fact, we think of the image (1) of
the interval 1 through . Let `(r
1
. r
2
. .... r
a
) be a point in ¹. We say
that passes through ` if there is t
0
in 1 such that (t
0
) = `.
Definition 23. We say that the subset ¹ of R
a
is connected if any
two points `
1
and `
2
of ¹ can be connected by a continuous curve,
i.e. if there is a continuous curve : 1 ÷ ¹ and t
1
. t
2
÷ 1 such that
(t
1
) = `
1
and (t
2
) = `
2
. This means that passes through `
1
and
`
2
.
Remark 24. An interval 1 of R is a subset of R with the following
property: if c. / ÷ 1 and r is between c and / (c _ r _ /), then
r is also in 1. In R, the connected subsets are exactly the intervals
of R. Indeed, let 1 be a connected subset of R, let c. / ÷ 1 and let r
with c _ r _ /. Since 1 is connected, let : J ÷ 1 be a continuous
curve which connect c and /. This means that there are t
1
and t
2
in J
such that (t
1
) = c and (t
2
) = /. We can restrict to the interval
[t
1
. t
2
] · J and apply Darboux property for the continuous function
(see Theorem 33). Hence r = (t
3
). where t
3
÷ [t
1
. t
2
]. So r ÷ 1;
134 6. THE NORMED SPACE R
m
:
thus 1 is an interval. Conversely, let 1 be an interval in R and let r
1
.
r
2
÷ 1. Let : [r
1
. r
2
] ÷ 1 be the identity mapping. This is obviously
a continuous curve which connect r
1
and r
2
.
Theorem 61. Let ¹ be a connected subset of R
a
and let f : ¹ ÷R
n
be a continuous mapping de…ned on ¹ with values in R
n
. Then the
image f (¹) of f in R
n
is also a connected subset of R
n
.
Proof. Let f (x) and f (y) be two points in f (¹). x. y ÷ ¹. Since
¹ is connected, there is a continuous curve : 1 ÷ ¹ and two points
c. / ÷ 1 (an interval in R) such that (c) = x and (/) = y. Now, the
composition f · : 1 ÷R
n
is a continuous curve with (f ·)(c) = f (x)
and (f · )(/) = f (y). Thus f (¹) is a connected subset of R
n
.
This is a fundamental result in di¤erent practical exercises. For
instance, let
o = ¦(r. n. .) ÷ R
3
: r
2
+n
2
+.
2
_ 1
2
¦
be the 31ball of radius 1 with centre at origin. Let 1 : o ÷ R be
the functions which associates to any point `(r. n. .) the sum of these
coordinates, namely
1(r. n. .) = r +n +..
Let us …nd the image of o through 1. Since o is connected (in fact o is
a convex subset of R
3
. i.e. for any pair of points 1. 1 of o. the segment
[1. 1] is contained in o) and since 1 is continuous, its image in R is a
connected subset (see Theorem 61), i.e. it is an interval (see Remark
24). In fact, this image is a closed and bounded interval because o
is a compact set (way?) and 1 is continuous. So it is of the form
[:. `] where : = inf 1(o) and ` = sup 1(o). To …nd : and ` is
not an easy task. We only remark that the points where it is realized
the greatest and the smallest values must be on the boundary ·o of
o. namely where r
2
+ n
2
+ .
2
= 1
2
(otherwise, if a point H(c. /. c) of
extremum, say a maximum, was inside the ball, not on the boundary
·o. then we can gently increase (or decrease) one of the values c. /. or
c. such that the new point 1 obtained in this way belongs to the ball
and, in it the function 1 has a greater value then the value of 1 in H).
In a later section (Conditional extremum points) we shall see how to
compute : and `.
The above theorem is helpful in proving the following useful result
(this result provides the basis of the di¤erent algorithms for solving
equations).
Theorem 62. Let 1 : [c. /] ÷R be a continuous function such that
1(c) 1(/) < 0. Then, there is a point c in (c. /) such that 1(c) = 0.
5. THE RIEMANN’S SPHERE 135
This means that the equation 1(r) = 0 has at least one solution in the
interval [c. /].
Proof. The set 1([c. /]) is an interval (see Theorem61 and Remark
24) which contains 1(c) and 1(/). Since 1(c) 1(/) < 0. the numbers
1(c) and 1(/) have distinct signs. Since 1([c. /]) is an interval and since
0 is between 1(c) and 1(/). 0 must be also in 1([c. /]). This means that
there is a c in [c. /] such that 1(c) = 0. Since 1(c) 1(/) < 0. this c
cannot be neither c nor /. so c ÷ (c. /).
We can use Theorem 62 in order to …nd approximative solutions for
an equation 1(r) = 0 in an interval [c. /]. on which the function 1 is
continuous (…nd a counterexample to this theorem in the case when 1 is
not continuous). We also assume that 1(c) 1(/) < 0. Let us divide the
segment [c. /] into two equal parts and chose that one [c
1
. /
1
] for which
1(c
1
) 1(/
1
) < 0 (if 1(c
1
) = 0 or 1(/
1
) = 0. c = c
1
or c = /
1
and we
stop the process). Let us repeat the same with the subinterval [c
1
. /
1
]
instead of [c. /]. and so on. If we cannot …nd c
a
or /
a
, : = 1. 2. .....
such that 1(c
a
) = 0 or 1(/
a
) = 0. the solution c is (the unique point)
in the intersection
o
¨
a=1
[c
a
. /
a
] (why?). So, for a small error indicator
0. if we take :
0
such that
b÷o
2
n
0
< . then the approximation c  c
a
0
(or c  /
a
0
) lead us to an error less then (why?). This is in fact
the description of a very known algorithm in Computer Science for
constructing approximative solutions for a large class of equations.
5. The Riemann’s sphere
In Fig.6.3 we have a sphere o of radius 1 0 and with center at
the origin ((0. 0. 0). Its equation is
(5.1) r
2
+n
2
+.
2
= 1
2
136 6. THE NORMED SPACE R
m
:
M'(a,b)
x
z
y
M(x,y,z)
N(0,0,R)
(C)
(C')
O
Fig. 6.3
We know that the subset
o = ¦(r. n. .) : r
2
+n
2
+.
2
= 1
2
¦
is a compact subset of R
3
(it is closed and bounded, why?). Since B.
Riemann used this model for explaining the "compacti…cation" of the
usual complex plane C (identi…ed here with the coordinate plane r(n),
we call o the Riemann sphere.We call the point `(0. 0. 1). the north
pole of o (see Fig.6.3). Let us associate to any point `(r. n. .) of the
sphere o, the point `
t
(c. /. 0) in the plane r(n (= C). obtained by
intersecting the line `` with the plane r(n (see Fig.6.3). Since for
` we cannot associate in this way a point in r(n. we say that there is
a one to one correspondence between or¦`¦ and C. Let us denote by
1 : o r¦`¦ ÷C, the mapping ` `
t
. or 1(`) = `
t
. It is not so
easy to express c and / as functions of r. n. .. If we think of a sequence
¦`
a
¦ of points on o. which is convergent in R
3
to `. it is easy to see
that the sequence ¦`
t
a
¦ is convergent to `
t
in C. So 1 is a continuous
function on o r ¦`¦. As in the case of the "compacti…cation" of R
by adding of the symbols ¦±·¦ (since in R= R'¦±·¦ any sequence
has at least one convergent subsequencewhy?it is a compact metric
space!)) we take a symbol "·" outside C and consider
´
C = C ' ¦·¦
with some obvious algebraic operations: r +·= ·+r = ·. r ÷ C,
[·[ = · (this is the symbol +· from
´
R), etc. If we extend now the
function 1 to the whole sphere o by putting 1(`) =·, we obtain a
bijection between the Riemann sphere and
´
C. We say that a sequence
¦.
a
¦ of
´
C is convergent to · if [.
a
[ ÷ · ÷
´
R. So this 1 is invertible
and 1
÷1
is also continuous. In particular
´
C is a compact metric space,
6. PROBLEMS 137
the least compact metric space which contains C (why?). This is why
one can also call
´
C the Riemann sphere. For instance, a "ball" with
centre at · is the exterior of an usual closed ball with centre at (
and of radius : 0 : ¦(r. n. .) : r
2
+ n
2
+ .
2
:
2
¦. The notion
of Riemann sphere is very important when we work with functions of
complex variable. Intuitively, · can be realized as the circumference
of a "circle" with center at ( ÷ C and of an in…nite radius. So, the
fundamental "neighborhoods" of ·are of the form¦. ÷ C : [.[ 1¦.
where 1 is any positive (usually large) real number. We …nally remark
that the metric structure on o is that one induced from R
3
.
6. Problems
1. Say if the following sets are open, closed, bounded, compact or
connected. In each case, compute their closure and their boundaries.
Draw them carefully!
a)
¦(r. n) : r
2
+n
2
< 9¦;
b)
¦(r. n) : r
2
+n
2
9¦;
c)
¦(r. n) : r
2
+n
2
= 5¦;
d)
¦(r. n) : r ÷ [0. 1); n ÷ (1. 2]¦;
e)
¦(r. n) : r +n = 3¦;
f)¦(¡. 0) : ¡ ÷ Q¦; g)¦(0.
1
a
) : : = 1. 2. ... ¦; h)¦(r. n) : n
2
= 2r. r ÷
[0. 1)¦; i)
¦(
1
:
.
1
:
) : : = 1. 2. ...¦;
j)
¦(r. n. .) : r +n +. _ 3; r. n. . ÷ [0. ·)¦
k)
¦(r. n. .) : r ÷ [÷1. 1]. n ÷ (0. 4]. . ÷ (÷3. 5]¦
l) ¦. ÷ C : [. ÷2i[ < 3¦; m)¦. ÷ C : [2. + 3[ _ 6¦; n)
¦. ÷ C : [. + 3 ÷2i[ 4¦;
o)
¦. ÷ C : . = r +in. r = 2. n _ 3¦;
138 6. THE NORMED SPACE R
m
:
p)
¦. ÷ C : 2 < [. ÷2[ _ 4¦;
q)
¦. ÷ C : [. ÷3 + 2i[ 2¦;
r)
¦1 ÷ ([0. 2] : 1 < 2¦;
s)
¦1 ÷ ([0. 2:] : 1 _ 3¦;
u)
¦1 ÷ ([0. 2:] : 1 ÷sin r < 0.3¦
v)
¦1 ÷ ([÷3.3] : o ÷
1
10
_ 1 < o +
1
10
.
where o(r) = r. o(r) = ÷r. or o(r) = r
2
¦; w)
¦1 ÷ ([0. 1] : 2 < 1 ÷o < 4¦.
where o(r) = r; y)1 = ¦(r. n) : ln(r
2
+n
2
÷4)(r+2n) is well de…ned¦.
2. Compute the limits of the following sequences:
a)
x
(a)
=
1
2: + 1
.
2: ÷1
3: + 4
. (1 +
4
:
)
2a
;
b)
x
(a)
=
: ÷1
3
: ÷
3
: ÷1
.
:sin
1
a
1 +:
;
c)
.
a
=
3 + 2i:
: + 2i
. i =
÷1;
d) .
a
=
1 +
j+1
a
a
; e) .
a
= exp
i: +
j
a
;
3. Starting with the de…nition of continuity and of uniform con
tinuity, determine what of the following functions are continuous and
what are uniformly continuous.
a) 1(r) = sin r. r ÷ [0. :];
b)
1(r. n) = (r +n.
1
rn
). r ÷ [1. 2]. n ÷ [3. 4];
c) 1(r. n. .) = r ÷n. where r
2
+n
2
+.
2
= 4; d) 1(r) =
1
a
. r ÷ (0. 2].
4. Some of the following limits exist, some do not exist. Say (and
prove!) which of them exist and compute them in the a¢rmative situ
ation.
a) lim
(a.&)÷(0.0)
a
3
+&
3
+1
2a
3
+3&
3
+2
; b) lim
(a.&)÷(0.0)
a&
a&+1÷1
;
6. PROBLEMS 139
c) lim
(a.&)÷(0.0)
a&
2
a
2
+&
2
(Hint:
a&
a
2
+&
2
_
1
2
. etc.);
d)
lim
(a.&)÷(0.0)
r
2
+n
2
[r[ +[n[
(Hint:
a
[a[+[&[
.
&
[a[+[&[
_ 1. etc.); e) lim
(a.&)÷(0.0)
a
3
+&
3
a
2
+&
2
; f) lim
a÷0
[a[
a
; g) lim
a÷0
exp(÷[a[)÷1
a
;
h) lim
(a.&)÷(0.0)
a&
a
2
+&
2
;
i)
lim
(a.&)÷(0.0)
rn
2
r
2
+n
4
(Hint: use (
1
a
. 0) and (
1
a
2
.
1
a
));
5. Compute, if you can, the following directional limits:
a) lim
a÷0.&=na
a&
a
2
+&
2
; b) lim
a÷0.&=na
2a
3
&
a
6
+&
2
;
c)
lim
a÷o.&=na
n
r
exp(÷(r +n));
d)
lim
(a.&)÷(1.0).a
2
+&
2
=1
rn exp(r
2
+n
2
).
6. Compute:
lim
(a.&.:)÷0
1
r
2
+n
2
+ 1
. 1 +rn.. cos(r +n +.)
and explain everything you did, step by step (small steps!).
7. Study the continuity of the following functions:
a)
1 : R ÷R. 1(r) = 1.
if r ÷ Q and 1(r) = 0. if r ÷ Q (Dirichlet’s function);
b)
1 : R ÷R. 1(r) = r.
if r ÷ Q. and 1(r) = ÷r. if r ÷ Q;
c)
1 : R ÷R. 1(r) = exp(÷r).
if r _ 0 and 1(r) = sin r. if r 0;
d)
1 : R
2
÷R
2
. 1(r. n) = (r. 0);
e)
1 : R
2
÷R. 1(r. n) = d((r. n). (0. 0)) =
r
2
+n
2
;
140 6. THE NORMED SPACE R
m
:
f)
1 : R
2
÷R
2
. 1(r. n) =
rn
r
2
+n
2
. rn
.
if (r. n) = (0. 0) and 1(0. 0) = (0. 0);
g)
1 : R
2
÷R. 1(r. n) = rn
r
2
÷n
2
r
2
+n
2
.
if (r. n) = (0. 0) and 1(0. 0) = 0;
h)
1 : R
2
÷R. 1(r. n) =
sin(r
3
+n
3
)
r
2
+n
2
.
if (r. n) = (0. 0) and 1(0. 0) = 0.
8. Prove that 1(r) = r
2
is uniformly continuous on [0. 1]. but
it is not on the whole R (Hint: use r
a
=
:. r
a+1
÷ r
a
÷ 0. but
1(r
a+1
) ÷1(r
a
) = 1 90).
9. Prove that 1(r) =
1
a
2
is uniformly continuous on [1. 2]. but not
on R.
10. Let (A. d) be a metric space. Prove that, for any …xed c in
A. the mapping 1
o
(r) = d(r. c) is a continuous function de…ned on A
with values in R.
11. Let
¹ = ¦(r. n. .) ÷ R
3
: 1 _ r
2
+n
2
+.
2
_ 4¦.
Prove that 1(¹) is an interval in R. Find it.
12. Do the same for
1(r. n) = r +n. r ÷ [1. 2]. n ÷ [1. 2].
Partial derivatives. Di¤erentiability.
7. Partial derivatives. Di¤erentiability.
Let ¹ be an open subset in R, c a …xed point in ¹ and let 1 : ¹ ÷R
be a function de…ned on ¹ with values in R. Let 1(c. :) = (c÷:. c+:),
: 0. be a small ball (an open interval in our particular case) of radius
: and with centre c. which is contained in ¹. Let / be a small quantity
such that c+/ ÷ 1(c. :). We call this / an "increment" of c in 1(c. :)
(or in ¹ if one takes / with c +/ ÷ ¹). The di¤erence 1(c +/) ÷1(c)
is called the increment of 1 at c. corresponding to the increment / of
c. So, here appears a new function .
o.;
(/) = 1(c+/) ÷1(c). This new
function depends on c and on 1. It is de…ned in a small ball, (÷. ).
which contains 0 as its centre and of radius . (at most : (why?)). The
description of this last function is important in the case we want to
7. PARTIAL DERIVATIVES. DIFFERENTIABILITY. 141
evaluate the variation of a phenomenon around a given point c. For
instance, if a worker has his salary c and if his salary increases with /.
what is the increment 1(c+/)÷1(c) of his family educational level? We
say that the increment 1(c +/) ÷1(c) is approximately linear around
c. if
(7.1) 1(c +/) ÷1(c) = \(c. 1) / +/ .
o.;
(/).
where .
o.;
is a function of / de…ned on (÷. ). .
o.;
(0) = 0 and
.
o.;
(/) ÷0. when / ÷0 (i.e. .
o.;
is continuous at 0). Here \(c. 1) is
a real number which depend on 1 and on c.
The birth of di¤erential calculus began with the following result.
Theorem 63. With the above notation and hypotheses, the incre
ment of 1 is approximately linear around c if and only if 1 is di¤eren
tiable at c and, in this case 1
t
(c) = \(c. 1). Thus,
(7.2) 1(c +/) ÷1(c) = 1
t
(c) / +/ .
o.;
(/).
Hence,
1(c +/) ÷1(c)  1
t
(c) /
and the error / .
o.;
(/) is a zero 0(/) of /. i.e.
lim
I÷0
/ .
o.;
(/)
/
= 0.
Proof. Let us divide by / the equality (7.1) and make / ÷0. We
obtain that the limit
lim
I÷0
1(c +/) ÷1(c)
/
= \(c. 1).
So, if the increment 1(c +/) ÷1(c) is approximately linear around c.
1 is di¤erentiable at c and 1
t
(c) = \(c. 1). Conversely, let us assume
that 1 is di¤erentiable at c. Then, if one construct
(7.3) .
o.;
(/) =
1(c +/) ÷1(c)
/
÷1
t
(c).
it is easy to verify that this function .
o.;
is continuous at 0 and it is
zero at / = 0 (do it!). If we take now for \(c. 1) the number 1
t
(c). and
for .
o.;
the function constructed in (7.3), we obtain the formula (7.1),
i.e. the increment of 1 is approximately linear around c.
Let us evaluate the increment of 1(r) = ÷r
2
+ 3r ÷7 at c = 10 if
the increment / of c is 0.5. We simply apply formula (7.2) and …nd
1(10 + 0.5) ÷1(10) = 1
t
(10) 0.5 + 0.5 .
;.10
(0.5)  ÷8.5.
142 6. THE NORMED SPACE R
m
:
Definition 24. With the above notation, the linear mapping d1(c) :
R ÷R, de…ned by
d1(c)(/) = 1
t
(c) /.
is called the …rst di¤erential of 1 at c. This one exists if and only if
the …rst derivative 1
t
(c) of 1 at c exists (why?).
Thus,
d1(c)(/)  1(c +/) ÷1(c).
i.e. the value d1(c)(/) of the …rst di¤erential of 1 at c. computed
in the increment / of c. is approximative equal to the corresponding
increment
1(c +/) ÷1(c)
of 1 at c.
Before extending the notion of a di¤erential to a vector function we
need some other simpler notion.
Let ¹ be an open subset of R
a
. f : ¹ ÷ R
n
, a vector function of
: variables, de…ned on ¹ with values in the normed (or metric) space
R
n
and a = (c
1
. c
2
. .... c
a
) a point in ¹. We write f = (1
1
. 1
2
. .... 1
n
).
where 1
1
. 1
2
. .... 1
n
are the : scalar component functions of f . For the
moment we take : = 1 and write f = 1. like a scalar function (with
values in R). Let us …x a variable r
;
(, = 1. 2. .... :) of the variable
vector
x = (r
1
. r
2
. .... r
;÷1
. r
;
. r
;+1
. .... r
a
).
For this …xed ,. let us de…ne a "partial function" .
;
of 1 at a. For this
we …x all the other variables r
1
. r
2
. .... r
;÷1
. r
;+1
. .... r
a
(except r
;
) by
putting
r
1
= c
1
. r
2
= c
2
. .... r
;÷1
= c
;÷1
. r
;+1
= c
;+1
. .... r
a
= c
a
and let us leave free the variable r
;
in
1(x) =1(r
1
. r
2
. .... r
;÷1
. r
;
. r
;+1
. .... r
a
).
i.e. we de…ne
(7.4) .
;
(t) = 1(c
1
. c
2
. .... c
;÷1
. t. c
;+1
. .... c
a
).
where t runs over the projection j:
;
(¹) of ¹ along the (,axis, where
j:
;
(r
1
. r
2
. .... r
;÷1
. r
;
. r
;+1
. .... r
a
) = r
;
Definition 25. With the above notation, if the function .
;
is dif
ferentiable at t = c
;
. one says that 1 has a partial derivative .
t
;
(a) with
respect to the variable r
;
at a and we denote this last one by
0;
0a
j
(a).
The mapping x
0;
0a
j
(x). x ÷ ¹. is called the partial derivative of 1
with respect to r
;
.
7. PARTIAL DERIVATIVES. DIFFERENTIABILITY. 143
Practically, if we want to compute the partial derivative of a scalar
function 1 of : variables
r
1
. r
2
. .... r
;÷1
. r
;
. r
;+1
. .... r
a
.
with respect to r
;
. we think of the other variables
r
1
. r
2
. .... r
;÷1
. r
;+1
. .... r
a
like being constants (parameters, or "inactivated" variables) and we
perform the usual di¤erential laws on the "active" variable r
;
. If : = 1.
we usually denote r
1
by r. If : = 2. we usually denote r
1
by r and r
2
by n. If : = 3. we usually denote r
1
by r. r
2
by n and r
3
by .. For
instance, let
1(r. n) = sin
2
(r
3
+n
3
)
be de…ned on R
2
and let a = (0.
3
¬
2
) be the …xed point at which we
want to compute the partial derivatives of 1 (with respect to r and
to n respectively). Let us use the de…nition to compute
0;
0a
(a). In our
case,
.
1
(t) = sin
2
(t
3
+
:
2
)
and
.
t
1
(t) = 2 sin(t
3
+
:
2
) cos(t
3
+
:
2
) 3t
2
(we just used the chain rule for computing the derivative of a composed
function of one variable). Now,
·1
·r
((0.
3
:
2
)) = .
t
1
(0) = 0.
Let us compute now
(7.5)
·1
·n
((r. n)) = 2 sin(r
3
+n
3
) cos(r
3
+n
3
) 3n
2
Here, we simply considered the initial function to depend ONLY on
n and we looked at r like to a constant. If we want to compute
0;
0&
((0.
3
¬
2
)). we simply make r = 0 and n =
3
¬
2
in the general expres
sion (7.5) of
0;
0&
((r. n)). Thus,
0;
0&
((0.
3
¬
2
)) is also 0. Since both partial
derivatives of 1 at (0.
3
¬
2
) are zero, we say that this last point is a
stationary (or critical) point.
If 1 is a function de…ned on an open subset ¹ of R
a
which has
partial derivatives with respect to all its variables at a point a. we
de…ne the gradient vector of 1 at a by the formula:
grad 1(a) =
·1
·r
1
(a).
·1
·r
2
(a). ....
·1
·r
a
(a)
.
144 6. THE NORMED SPACE R
m
:
We say that a is a critical (stationary) point for 1 if grad 1(a) = 0.
The gradient is the direct generalization of the notion of "velocity".
We know from any course of "Linear Algebra" that a mapping T :
R
a
÷R
n
is linear if T(x +y) = T(x) +T(y) and T(cx) =cT(x) for
any x. y in R
a
and c in R. It is easy to see that if 1
1
. 1
2
. .... 1
n
are the
component functions of T. then T is a linear mapping if and only if all
the component functions 1
1
. 1
2
. .... 1
n
of T are linear (prove it!).
Theorem 64. Any linear mapping T : R
a
÷ R
n
is a continuous
vector function of : variables.
Proof. It is su¢cient to prove that any component function 1
j
.
i = 1. 2. .... : of T is continuous (see Theorem 54). This means that we
can reduce ourselves to the case of : = 1. i.e. to the case of a scalar
function 1 : R
a
÷R. Let
¦c
1
= (1. 0. 0. .... 0). c
2
= (0. 1. 0. .... 0). .... c
a
= (0. 0. 0. .... 0. 1)¦
be the canonical basis of R
a
. This means that any vector x = (r
1
. r
2
. .... r
a
)
can be uniquely represented as:
x = r
1
e
1
+r
2
e
2
+... +r
a
e
a
.
Let us denote
c
1
= 1(e
1
). c
2
= 1(e
2
). .... c
a
= 1(e
a
).
These are …xed real numbers. Hence,
1(x) = 1((r
1
. r
2
. .... r
a
)) = r
1
c
1
+... +r
a
c
a
.
If
x
(n)
= (r
(n)
1
. r
(n)
2
. .... r
(n)
a
) ÷x = (r
1
. r
2
. .... r
a
).
when : ÷·. then,
r
(n)
1
÷r
1
. r
(n)
2
÷r
2
. .... r
(n)
a
÷r
a
.
when : ÷· (componentwise convergence). Thus,
1(x
(n)
) = r
(n)
1
c
1
+r
(n)
2
c
2
+... +r
(n)
a
c
a
÷r
1
c
1
+... +r
a
c
a
which is just 1(x). Hence, 1 is a continuous mapping.
Remark 25. Let us de…ne the associated matrix of
T = (1
1
. 1
2
. .... 1
n
)
by c
j;
= 1
j
(e
;
) for i = 1. 2. .... : and , = 1. 2. .... :. So the matrix
¹ = (c
j;
) is a :: matrix with entries in R. If we compute now
T(x)
2
= 1
1
(x)
2
+1
2
(x)
2
+... +1
n
(x)
2
=
7. PARTIAL DERIVATIVES. DIFFERENTIABILITY. 145
a
¸
j=1
r
j
c
1j
2
+
a
¸
j=1
r
j
c
2j
2
+... +
a
¸
j=1
r
j
c
nj
2
_
_
a
¸
j=1
r
2
j
a
¸
j=1
c
2
1j
+
a
¸
j=1
r
2
j
a
¸
j=1
c
2
2j
+... +
a
¸
j=1
r
2
j
a
¸
j=1
c
2
nj
= x
2
¹
2
.
where we recall that
¹ =
n
¸
;=1
a
¸
j=1
c
2
;j
.
Thus,
(7.6) T(x) _ ¹ x .
From here we can easily directly prove the continuity of T (do it!).
Now, we come back to the de…nition of the linear approximation of
the increment 1(r + /) ÷ 1(r) of a function 1 around a point c. in a
general situation.
Definition 26. (Frechet) Let 1 be an open subset of R
a
and let
a be a …xed point in 1. Let 1 : 1 ÷ R be a function de…ned on 1
with values in R. We say that 1 is di¤erentiable at a if there is a linear
mapping 1
a
= 1 : R
a
÷ R and a continuous scalar function .(h)
which is continuous at 0 =(0. 0. .... 0
. .. .
a÷tjncc
). de…ned on a small ball 1(0.:) ·
R
a
. : 0. .(0) = 0 with lim
h÷0
,(h)
h
= 0, such that
(7.7) 1(a +h)÷1(a) =1(h) +.(h).
This means that the increment 1(a +h)÷1(a) can be linearly approx
imated by the linear mapping 1 (which depend on a and on 1) around
the point a up to a function .(h) which is a zero of h (0(h)) of order
1 ( lim
h÷0
,(h)
h
= 0). The linear mapping 1 is called the (…rst) di¤erential
of 1 at a. We write it as d1(a). Hence, formula (7.7) becomes
(7.8) 1(a +h)÷1(a) =d1(a)(h) +.(h).
Remark 26. It is clear that 1 is di¤erentiable at a if and only if
there is a linear function 1 : R
a
÷ R such that the following limit
exists and it is zero:
(7.9) lim
h÷0
1(a +h)÷1(a)÷1(h)
h
= 0.
146 6. THE NORMED SPACE R
m
:
Shortly saying, 1 is di¤erentiable at a if it can be "well" approx
imated on a small neighborhood of a by a formula of the following
type:
(7.10) 1(a +h) 1(a)+1(h).
where 1 is a linear mapping and h is a small increment of a. This
last interpretation is very useful in Physics and in Engineering when a
phenomenon is "linearized".
The next big problem is how to compute this 1 in language of 1 and
a. But, …rst of all, let us use only the de…nition and the remark above
to "guess" the di¤erentials for some simple functions. For instance, if
1 has only one variable, we …ned again De…nition 24. If 1 is a constant
function, then d1(a) a is the zero linear mapping (prove this!). The
…rst di¤erential of a linear mapping 1 : R
a
÷R is 1 itself (why?). In
particular, the ith projection j:
j
: R
a
÷R,
j:
j
(/
1
. /
2
. .... /
j
. .... /
a
) = /
j
.
is di¤erentiable and its di¤erential j:
j
is denoted by dr
j
. or dr. dn. d.
in the 31case. So
dn(1. 2. ÷3)(3. 1. ÷7) = 1. d.(c
1
. c
2
. c
3
)(÷2. 3. 5) = 5
for any a = (c
1
. c
2
. c
3
).
Theorem 65. If 1 is di¤erentiable at a ÷ 1. where 1 is an open
subset of R
a
. then 1 is continuous at a. This means that the property
of di¤erentiability is stronger then the property of continuity.
Proof. Let ¦a
(a)
¦ be a sequence of vectors in R
a
which is conver
gent to a and let h
(a)
= a
(a)
÷a (÷0). Then
1(a + h
(a)
) = 1(a) +d1(a)(h
(a)
) +.(h
(a)
)
(see (7.8)). Since d1(a) is a linear mapping, it is continuous (see The
orem 64), so
lim
a÷o
d1(a)(h
(a)
) =0.
Since lim
h÷0
,(h)
h
= 0. one has that lim
a÷o
.(h
(a)
) = 0 (why?). Hence,
1(a +h
(a)
) ÷1(a).
when : ÷·.
Theorem 66. The linear mapping 1 = d1(a) is uniquely deter
mined by 1 and a.
7. PARTIAL DERIVATIVES. DIFFERENTIABILITY. 147
Proof. If there was another one l such that
(7.11) 1(a +h) ÷1(a) =l(h) +.
1
(h).
where .
1
(0) = 0. .
1
is continuous at 0 and lim
h÷0
,
1
(h)
h
= 0. we can write
that
1(h) +.(h) = l(h) +.
1
(h)
for all h in a small ball centered at origin. Moreover,
(7.12) lim
h÷0
(1 ÷l)(h)
h
= lim
h÷0
.
1
(h) ÷.(h)
h
= 0.
We want to prove that for any x in R
a
one has 1(x) = l(x). We assume
contrary, namely that there is a x
0
such that (1 ÷l)(x
0
) = 0. If t 0 is
small, then tx
0
is small, i.e. it is close to 0. because tx
0
 = t x
0
 ÷0.
when t ÷0. t 0. Let us come back to (7.12) and write
lim
t÷0
(1 ÷l)(tx
0
)
tx
0

= lim
t÷0
t (1 ÷l)(x
0
)
t x
0

= 0.
So, (1 ÷l)(x
0
) = 0 and we just obtained a contradiction. Hence, there
is no x
0
with (1 ÷l)(x
0
) = 0 and so 1 = l.
Thus, if we …nd a method to compute 1 = d1(a). this 1 is unique.
It depends only on 1 and on a.
Theorem 67. If 1 is di¤erentiable at a. then all the partial deriv
atives
0;
0a
1
.
0;
0a
2
. ....
0;
0an
exists at a and
(7.13) d1(a)(/
1
. /
2
. .... /
a
) =
·1
·r
1
(a)/
1
+
·1
·r
2
(a)/
2
+... +
·1
·r
a
(a)/
a
.
or, using the projection j:
;
= dr
;
notation (see Remark 26), we get
(7.14) d1(a) =
·1
·r
1
(a)dr
1
+
·1
·r
2
(a)dr
2
+... +
·1
·r
a
(a)dr
a
.
Moreover, if 1 is of class (
1
on a ball 1(a.:). for a small : 0. i.e. if
1 ÷ (
1
(1(a.:)) (this means that 1 has partial derivatives with respect
to all variables r
1
. r
2
. .... r
a
and all of these are continuous on 1(a.:)),
then 1 is di¤erentiable at a and formula (7.14) works.
Proof. We suppose that 1 is di¤erentiable at a and let 1 = d1(a)
be its di¤erential at a. We know from Linear Algebra or from the proof
of Theorem 64 that
1(/
1
. /
2
. .... /
a
) = \
1
/
1
+\
2
/
2
+... +\
a
/
a
.
where \
1
. \
2
. .... \
a
are …xed real numbers (recall that \
j
= 1(e
j
).
where e
j
is the ith vector of the canonical basis of R
a
. etc.). Let us
148 6. THE NORMED SPACE R
m
:
chose now a , in ¦1. 2. .... :¦, let us take 0. close to 0 and let us
also take
h = (0. 0. .... 0.
....
;
. 0. .... 0)
in formula (7.9). We get
lim
~÷0
1(c
1
. c
2
. .... c
;÷1
. c
;
+. c
;+1
. .... c
a
)÷1(a)÷\
;
= 0.
Since this limit exists, the partial derivative with respect to , exists and,
from this last formula we get that
0;
0a
j
(a) = \
;
. for any , ÷ ¦1. 2. .... :¦.
Hence,
1(/
1
. /
2
. .... /
a
) =
·1
·r
1
(a)/
1
+
·1
·r
2
(a)/
2
+... +
·1
·r
a
(a)/
a
and the …rst part of the statement is completely proved.
Let us now assume that 1 is of class (
1
on a ball 1(a. :). : 0.
Let us take the following linear mapping 1 : R
a
÷R:
1(/
1
. /
2
. .... /
a
) =
·1
·r
1
(a)/
1
+
·1
·r
2
(a)/
2
+... +
·1
·r
a
(a)/
a
.
Let us prove that this 1 is indeed the di¤erential of 1 at a. To be easier,
let us also assume that : = 2. Then, we want to prove that
(7.15) lim
I
1
.I
2
÷0
1(c
1
+/
1
. c
2
+/
2
) ÷1(c
1
. c
2
) ÷1(/
1
. /
2
)
h
= 0.
Let us write:
1(c
1
+/
1
. c
2
+/
2
) ÷1(c
1
. c
2
) = 1(c
1
+/
1
. c
2
+/
2
) ÷1(c
1
. c
2
+/
2
)
(7.16) +1(c
1
. c
2
+/
2
) ÷1(c
1
. c
2
).
Now, let us consider the function
.
1
(t) = 1(t. c
2
+/
2
). t ÷ [c
1
. c
1
+/
1
]
±
and let us apply to it Lagrange’s formula:
(7.17) 1(c
1
+/
1
. c
2
+/
2
) ÷1(c
1
. c
2
+/
2
) =
·1
·r
1
(c
1
. c
2
+/
2
) /
1
.
where c
1
÷ [c
1
. c
1
+/
1
]
±
. Let us do the same for 1(c
1
. c
2
+/
2
)÷1(c
1
. c
2
)
by considering the function
.
2
(t) = 1(c
1
. t). t ÷ [c
2
. c
2
+/
2
]
±
.
We get
(7.18) 1(c
1
. c
2
+/
2
) ÷1(c
1
. c
2
) =
·1
·r
2
(c
1
. c
2
) /
2
.
7. PARTIAL DERIVATIVES. DIFFERENTIABILITY. 149
where c
2
÷ [c
2
. c
2
+/
2
]
±
. Let us come back in (7.16) with the expressions
of (7.17) and (7.18). So,
1(c
1
+/
1
. c
2
+/
2
) ÷1(c
1
. c
2
) ÷1(/
1
. /
2
)
(7.19)
=
¸
·1
·r
1
(c
1
. c
2
+/
2
) ÷
·1
·r
1
(c
1
. c
2
)
/
1
+
¸
·1
·r
2
(c
1
. c
2
) ÷
·1
·r
2
(c
1
. c
2
)
/
2
.
Since the function 1 is of class (
1
in a small neighborhood of a =
(c
1
. c
2
). one has that:
·1
·r
1
(c
1
. c
2
+/
2
) ÷
·1
·r
1
(c
1
. c
2
)
÷0.
when h ÷0 i.e. /
1
÷0 and /
2
÷0 and
·1
·r
2
(c
1
. c
2
) ÷
·1
·r
2
(c
1
. c
2
)
÷0.
when h ÷0. Since
[/
1
[
h
.
[/
2
[
h
_ 1.
one has that the limit in (7.15) is zero (do this slowly, step by step!).
Hence, 1 is di¤erentiable at a and its di¤erential has the usual form:
d1(a) =
·1
·r
1
(a)dr
1
+
·1
·r
2
(a)dr
2
.
For an arbitrary : the proof is similar, but the writing is more compli
cated.
This last theorem is very useful in computations. For instance, let
1 : R
3
÷R be de…ned by
1(r. n. .) = ln(1 + r
2
+n
4
+.
6
).
All the partial derivatives
·1
·r
=
2r
1 +r
2
+n
4
+.
6
.
·1
·n
=
4n
3
1 +r
2
+n
4
+.
6
and
·1
·.
=
6.
5
1 +r
2
+n
4
+.
6
exist and are continuous on the whole R
3
. in particular around the
point (1. ÷1. 2). Applying the last theorem (see Theorem 67) we see
that 1 is di¤erentiable at (1. ÷1. 2) and
d1(1. ÷1. 2) =
·1
·r
(1. ÷1. 2)dr +
·1
·n
(1. ÷1. 2)dn +
·1
·.
(1. ÷1. 2)d. =
150 6. THE NORMED SPACE R
m
:
=
2
67
dr ÷
4
67
dn +
192
67
d..
Recall a basic fact: d1(1. ÷1. 2) is NOT a number, but a linear mapping
from R
3
to R. For instance,
d1(1. ÷1. 2)(3. ÷4. 0) =
2
67
dr(3. ÷4. 0)÷
4
67
dn(3. ÷4. 0)+
192
67
d.(3. ÷4. 0) =
=
2
67
3 ÷
4
67
(÷4) +
192
67
0 =
22
67
.
This last one is a real number because d1(1. ÷1. 2) : R
3
÷R is a linear
mapping.
We want now to extend the notion of di¤erentiability from scalar
functions of : variables to vector functions.
Definition 27. Let f : 1 ÷R
n
be a vector function with its com
ponents (1
1
. 1
2
. .... 1
n
). de…ned on an open subset 1 of R
a
with values
in R
n
. We say that 1 is di¤erentiable at a ÷ 1 if all its components
1
1
. 1
2
. .... 1
n
are di¤erentiable at a like scalar functions. Moreover, if
h = (/
1
. /
2
. .... /
a
) is a vector in R
a
and if
d1
j
(a)(h) =c
j1
/
1
+c
j2
/
2
+... +c
ja
/
a
.
where
c
j1
=
·1
j
·r
1
(a). c
j2
=
·1
j
·r
2
(a). ....
·1
j
·r
a
(a).
then the matrix
J
a.f
= (c
j;
=
·1
j
·r
;
(a)).
with : rows and : columns is called the Jacobi (or jacobian) matrix of
1 at a. The linear mapping T : R
a
÷R
n
de…ned by the jacobian matrix
J
a.f
(with respect to the canonical bases of R
a
and R
n
respectively) is
called the di¤erential of f at a. We write T = df (a). The determinant
[J
a.f
[ of J
a.f
. in the particular case : = :. is said to be the jacobian of
f at a.
For instance,
f : 1 ÷R
2
. 1 = ¦(r. n. .) ÷ R
3
: r 0. n 0. . 0¦.
de…ned by
f (r. n. .) =
1
rn.
. rn.
is di¤erentiable at any point a =(c. /. c) of 1 because its components
1
1
(r. n. .) =
1
rn.
7. PARTIAL DERIVATIVES. DIFFERENTIABILITY. 151
and
1
2
(r. n. .) = rn.
have this last property (why?). Since
d1
1
(a) = ÷
1
c
2
/c
dr ÷
1
c/
2
c
dn ÷
1
c/c
2
d.
and
d1
2
(a) = /c dr +cc dn +c/ d..
the jacobian matrix of f at a is the 2 3 matrix
÷
1
o
2
bc
÷
1
ob
2
c
÷
1
obc
2
/c cc c/
.
For instance, if c = 1. / = 1 and c = ÷2. we get the numerical matrix
1
2
1
2
÷
1
4
÷2 ÷2 1
.
Now, if we want to compute the value of d1(1. 1. ÷2) : R
3
÷R
2
at the
point (3. 4. ÷5). from Linear Algebra or from the remark 25, we get
1
2
1
2
÷
1
4
÷2 ÷2 1
¸
3
4
÷5
¸
=
3
2
+
4
2
+
5
4
÷6 ÷8 ÷5
=
19
4
÷19
.
so d1(1. 1. 2)(3. 4. ÷5) = (
19
4
. ÷19).
Remark 27. One can prove that f : 1 ÷R
n
is di¤erentiable at a
point a ÷1 · R
a
if and only if there is a linear mapping T : R
a
÷R
n
which depends on a such that the following limit exists and is equal to
zero:
lim
h÷0
f (a +h) ÷f (a) ÷T(h)
h
= 0.
This is equivalent to say that the increment f (a +h) ÷f (a) of our vec
tor function f at a. corresponding to the increment h of a. can be "well"
approximated by the value of the liner function T at h (do this slowly,
step by step!). The uniqueness of the above T is obvious because its
components are uniquely de…ned, being the di¤erentials of some scalar
functions, the components of f .
Exercise 1. Let f . g : 1 ÷ R
n
. be two di¤erentiable functions on
1 (at any point of 1), where 1 is an open subset in R
a
and let \ be
a real number. Then: f +g. f ÷g. fg (only for : = 1)
f
g
(only for
: = 1 and g(a) = 0), \f . are also di¤erentiable on 1 and
a)
d(f +g)(a) =df (a)+dg(a);
152 6. THE NORMED SPACE R
m
:
b)
d(f ÷g)(a) =df (a)÷dg(a);
c)
d(fg)(a) = g(a)df (a) +f (a)dg(a);
d)
d(
f
g
) =
g(a)df (a) ÷f (a)dg(a)
g(a)
2
;
e) d(\f ) = \ df for \ ÷ R.
8. Chain rules
Let ¹, 1 be two open subsets of R and let c be a point in ¹. Let
1 : ¹ ÷1 be a function de…ned on ¹ with values in 1 such that 1 is
di¤erentiable at c. Let o : 1 ÷R be a di¤erentiable function at 1(c).
Then the composed function o · 1 : ¹ ÷R is di¤erentiable at c and
(o · 1)
t
(c) = o
t
(1(c)) 1
t
(c)
(the simplest chain rule!). Indeed,
lim
a÷o
o(1(r)) ÷o(1(c))
r ÷c
=
= lim
;(a)÷;(o)
o(1(r)) ÷o(1(c))
1(r) ÷1(c)
lim
a÷o
1(r) ÷1(c)
r ÷c
= o
t
(1(c)) 1
t
(c).
So (o·1)
t
(c) exists and is exactly o
t
(1(c))1
t
(c). In particular, if 1 is
invertible and 1
÷1
is di¤erentiable at / = 1(c) then, from 1
÷1
(1(r)) =
r. we get 1
÷1t
(/) 1
t
(c) = 1. i.e. 1
÷1t
(/) =
1
;
0
(o)
. or (1
÷1
)
t
(1(c)) =
1
;
0
(o)
.
We want nowto generalize this simple chain rule to vector functions.
Let us start with a simpler case, namely, let us take a "curve" f : ¹ ÷
1. f = (1
1
. 1
2
. .... 1
a
). where ¹ is an open subset in R and 1 is an open
subset in R
a
. Let o : 1 ÷ R be a di¤erential function at b = f (c)
and let us assume that f is di¤erentiable at c. Let / = o · f : ¹ ÷ R
be the composition between o and f . i.e. the restriction of o to the
:1 "curve" f (to the image of f in the common language!). Then, the
following result is fundamental in applications.
Theorem 68. (di¤erentiation along a curve) With the above nota
tion and hypotheses,
(8.1) (o · f )
t
(c) =
·o
·r
1
(f (c)) 1
t
1
(c) +
·o
·r
2
(f (c)) 1
t
2
(c) +...
... +
·o
·r
a
(f (c)) 1
t
a
(c).
8. CHAIN RULES 153
For : = 1 we …nd again the above formula (o · 1)
t
(c) = o
t
(1(c)
1
t
(c).
Proof. To be easier we take the particular case : = 2 and we
assume that f and o are functions of class (
1
on ¹ and 1 respectively.
Whenever we write limit of something or the derivative of a function,
be sure that we implicitly prove that this limit or this derivative exists
(prove this slowly in what follows!).
In this case, /(r) = o(1
1
(r). 1
2
(r)) for any r ÷ ¹. So,
/
t
(c) = lim
a÷o
/(r) ÷/(c)
r ÷c
= lim
a÷o
o(1
1
(r). 1
2
(r)) ÷o(1
1
(c). 1
2
(c))
r ÷c
=
(8.2) = lim
a÷o
o(1
1
(r). 1
2
(r)) ÷o(1
1
(c). 1
2
(r))
r ÷c
+
lim
a÷o
o(1
1
(c). 1
2
(r)) ÷o(1
1
(c). 1
2
(c))
r ÷c
.
Let us consider the …rst limit in (8.2) and let us apply Lagrange’s
formula (see Corollary 4) for the mapping t ÷ o(1
1
(t). 1
2
(r)) on the
interval [c. r] (or [r. c] if r < c). We get
o(1
1
(r). 1
2
(r)) ÷o(1
1
(c). 1
2
(r)) =
·o
·r
1
(1
1
(c). 1
2
(r)) 1
t
1
(c) (r ÷c).
where c is between c and r. Here we used our chain formula for : = 1
(where?explain!). Coming back to the …rst limit in (8.2) and using the
fact that
0j
0a
1
, 1
t
1
and 1
2
are continuous, we get:
lim
a÷o
o(1
1
(r). 1
2
(r)) ÷o(1
1
(c). 1
2
(r))
r ÷c
= lim
a÷o
·o
·r
1
(1
1
(c). 1
2
(r)) 1
t
1
(c) =
=
·o
·r
1
(1
1
(c). 1
2
(c)) 1
t
1
(c).
We take now the second limit in (8.2) and apply Lagrange’s formula
for the mapping t ÷o(1
1
(c). 1
2
(t)) on the same interval [c. r]. We get
o(1
1
(c). 1
2
(r)) ÷o(1
1
(c). 1
2
(c)) =
·o
·r
2
(1
1
(c). 1
2
(:)) 1
t
2
(:)) (r ÷c).
where : is a number between c and r. Since
0j
0a
2
. 1
2
and 1
t
2
are con
tinuous (by our restrictive hypothesis in the present proof!), we obtain
that
lim
a÷o
o(1
1
(c). 1
2
(r)) ÷o(1
1
(c). 1
2
(c))
r ÷c
= lim
a÷o
·o
·r
2
(1
1
(c). 1
2
(:)) 1
t
2
(:))
=
·o
·r
2
(1
1
(c). 1
2
(c)) 1
t
2
(c)).
154 6. THE NORMED SPACE R
m
:
thus our formula (8.1) is completely proved for : = 2.
The statement of the theorem is true without these restrictions
made here, but the proof is more sophisticated.
If the curve f : R ÷R
3
is a line which passes through the point
`
0
(r
0
. n
0
. .
0
) and having the direction of the versor
u = (cos c. cos . cos )
(these cosines are usually called the directional cosines of the line), i.e.
f (t) = (r
0
+t cos c. n
0
+t cos . .
0
+t cos ). then, the above derivative
(o · f )
t
(0) =
·o
·r
1
(r
0
. n
0
. .
0
) cos c +
·o
·r
2
(r
0
. n
0
. .
0
)) cos +
+
·o
·r
3
(r
0
. n
0
. .
0
) cos = 'grad o(`
0
). u` .
(a scalar product!) is called the directional derivative of o at the
point `
0
along the versor u.
For instance, if u = (1. 0. 0). we get the partial derivative of o at
`
0
with respect to r
1
. etc.
We can now immediately extend the formula (8.1) for the case of
a vector function g : 1 ÷R
n
. g = (o
1
. o
2
. .... o
n
). Thus, for any …xed
, ÷ ¦1. 2. .... :¦. one has
(8.3)
(o
;
·f )
t
(c) =
·o
;
·r
1
(f (c)) 1
t
1
(c)+
·o
;
·r
2
(f (c)) 1
t
2
(c)+... +
·o
;
·r
a
(f (c)) 1
t
a
(c).
If we use now the matrix language, formula (8.3) becomes
(8.4)
¸
¸
¸
¸
¸
¸
(o
1
· f )
t
(c)
(o
2
· f )
t
(c)
.
.
.
(o
n
· f )
t
(c)
¸
=
¸
¸
¸
¸
¸
¸
¸
0j
1
0a
1
(f (c))
0j
1
0a
2
(f (c)) . . .
0j
1
0an
(f (c))
0j
2
0a
1
(f (c))
0j
2
0a
2
(f (c)) . . .
0j
2
0an
(f (c))
. . . . . .
. . . . . .
. . . . . .
0jm
0a
1
(f (c))
0jm
0a
2
(f (c)) . . .
0jm
0an
(f (c))
¸
¸
¸
¸
¸
¸
¸
1
t
1
(c)
1
t
2
(c)
.
.
.
1
t
a
(c)
¸
.
Up to now our function f was a function of one variable t. Let us make
the last generalization and consider a vectorial function f of j variables
t
1
. t
2
. .... t
j
de…ned on an open subset ¹ of R
j
. So we have the following
8. CHAIN RULES 155
composition: ¹
f
÷ 1
g
÷ R
n
. We denote by h = g · f : ¹ ÷ R
n
and
preserve the notation x = (r
1
. r
2
. .... r
a
) for a point (vector!) in R
a
.
Thus,
f (t
1
. t
2
. .... t
j
) = (1
1
(t
1
. t
2
. .... t
j
). 1
2
(t
1
. t
2
. .... t
j
). .... 1
a
(t
1
. t
2
. .... t
j
))
and
g(r
1
. r
2
. .... r
a
) = (o
1
(r
1
. r
2
. .... r
a
). .... o
n
(r
1
. r
2
. .... r
a
)).
Let now a be a …xed point of ¹. a = (c
1
. c
2
. .... c
j
) and b = f (a). We
assume that f and g are di¤erentiable at a and at b respectively.
Theorem 69. (chain rule theorem) With these notation and hy
potheses, the composed function h = g · f is di¤erentiable at a and
one has the following relation between the corresponding jacobian ma
trices :
(8.5) J
a.g·f
= J
b.g
J
a.f
.
This is the most sophisticated chain rule. Moreover, in this case, Linear
Algebra says that
(8.6) d(g · f )(a) =dg(b)·df (a).
this last composition being the composition between the corresponding
linear mappings.
Proof. Formula (8.6) is a direct consequence of formula (8.5) and
the basic result of Linear Algebra which says that there is an isomorphic
bijection between the :: matrices and the linear mapping 1 : R
a
÷
R
n
. This bijection carries the product between two matrices into the
composition of the corresponding linear mappings. Hence, it remains
us to prove formula (8.5). We shall see that this formula is a pure
generalization of formula (8.4). Indeed, let us …x i ÷ ¦1. 2. .... j¦ and
let us consider the mapping
'
(j)
: ¹
j
÷1. '
(j)
= (.
(j)
1
. .
(j)
2
. .... .
(j)
a
)
de…ned by
t f (c
1
. c
2
. .... c
j÷1
. t. c
j+1
. .... c
j
).
It is de…ned on the ith projection ¹
j
= j:
j
(¹) of ¹ (which is again
openwhy?). Let us denote h
(j)
= g · '
(j)
and let us write formula (8.4)
for it:
156 6. THE NORMED SPACE R
m
:
¸
¸
¸
¸
¸
¸
¸
(o
1
· '
(j)
)
t
(c
j
)
(o
2
· '
(j)
)
t
(c
j
)
.
.
.
(o
n
· '
(j)
)
t
(c
j
)
¸
=
¸
¸
¸
¸
¸
¸
¸
0j
1
0a
1
('
(j)
(c
j
))
0j
1
0a
2
('
(j)
(c
j
)) . . .
0j
1
0an
('
(j)
(c
j
))
0j
2
0a
1
('
(j)
(c
j
))
0j
2
0a
2
('
(j)
(c
j
)) . . .
0j
2
0an
('
(j)
(c
j
))
. . . . . .
. . . . . .
. . . . . .
0jm
0a
1
('
(j)
(c
j
))
0jm
0a
2
('
(j)
(c
j
)) . . .
0jm
0an
('
(j)
(c
j
))
¸
(8.7)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
.
(j)
1
t
(c
j
)
.
(j)
2
t
(c
j
)
.
.
.
.
(j)
a
t
(c
j
)
¸
.
We now see that
(o
;
· '
(j)
)
t
(c
j
) =
·/
;
·t
j
(a)
for any , = 1. 2. .... : and i = 1. 2. .... j. Here h = (/
1
. /
2
. .... /
n
) are
the components of the composed function h = g · f .
Another remark is that
·o
;
·r
I
('
(j)
(c
j
)) =
·o
;
·r
I
(f (a))
and
.
(j)
;
t
(c
j
) =
0;
j
0t
i
(a). But, if we substitute all of these in formula
(8.7), we get exactly formula (8.5) from the statement of the theorem.
Remark 28. It is possible to prove the chain rule theorem, namely
the formula (8.6), in a not so long "upgrading" way. But that proof (see
[Nik], or [Pal]) is more abstract, more elaborated and not so natural.
Our proof here is not so general, but it follows the natural historical
way, from a "simpler" to a "more complicated" case.
8. CHAIN RULES 157
Let us take an usual situation and let us apply formula (8.5) to it.
Let ¹and 1 be two open subsets of R
2
and let (r. n) (n(r. n). ·(r. n))
be a di¤erentiable (at any point of ¹) vector function de…ned on ¹
with values in 1. Let 1(n. ·) be a di¤erentiable function de…ned on
1 with values in R. Here we also use n and · for the coordinates of
a free vector in 1 · R
2
. The only connection between n. · and the
functions of two variables n(r. n) and ·(r. n) respectively, is that the
variable n and · are substituted with two functions n(r. n) and ·(r. n)
respectively, in variables r and n. For instance, n = r +n, · = rn and
1(r + n. rn). This is a new function in r and n. Here, n(r. n) = r + n
and ·(r. n) = rn. This abuse of notation is still working for more then
200 years and it did not caused any damage in science. Let /(r. n) =
1(n(r. n). ·(r. n)) be the composition between 1 and the …rst function
(r. n) ÷ (n(r. n). ·(r. n)). This new function is also denoted by 1. i.e.
the notation 1(r. n) = 1(n(r. n). ·(r. n)) produce no confusion for an
working mathematician (another abuse, which is not indicated to be
used by a beginner!). The function / is also di¤erentiable on ¹ and
0I
0a
(c. /)
0I
0&
(c. /)
=
0;
0&
(n(c. /). ·(c. /))
0;
0·
(n(c. /). ·(c. /))
0&
0a
(c. /)
0&
0&
(c. /)
0·
0a
(c. /)
0·
0&
(c. /)
.
Let us normally write this formula:
(8.8)
·/
·r
(c. /) =
·1
·n
(n(c. /). ·(c. /))
·n
·r
(c. /) +
·1
··
(n(c. /). ·(c. /))
··
·r
(c. /).
·/
·n
(c. /) =
·1
·n
(n(c. /). ·(c. /))
·n
·n
(c. /) +
·1
··
(n(c. /). ·(c. /))
··
·n
(c. /).
How do we recall these useful formulas? For this, write again
/(r. n) = 1(n(r. n). ·(r. n)). To …nd
0I
0a
. we look at the variables n
and · of 1 and observe where r is. If r appears in n = n(r. n). we take
the partial derivative of 1 w.r.t. n and multiply it by the partial deriv
ative of n w.r.t. r. Here is a "chain": 1 ÷ n ÷ r. So we get
0;
0&
0&
0a
.
If r also appears in · = ·(r. n), we consider the chain 1 ÷· ÷r and
obtain
0;
0·
0·
0a
. Since r appears both (if it is the case!) in n and in ·.
we must superpose both "e¤ects" (add them!) and …nally obtain:
(8.9)
·/
·r
=
·1
·n
·n
·r
+
·1
··
··
·r
.
The corresponding points at which we compute these partial derivatives
are easy to be …nd. If we change r with n in (8.9) we get the second
essential formula of (8.8):
158 6. THE NORMED SPACE R
m
:
(8.10)
·/
·n
=
·1
·n
·n
·n
+
·1
··
··
·n
.
Example 14. In the Cartesian plane ¦(; i. j¦. we consider a heat
ing source in the origin ((0. 0). The temperature 1(r. n) at the point
`(r. n) veri…es the following equation (a partial di¤erential equation
of order 1÷ a PDE1):
n
·1
·r
÷r
·1
·n
= 0.
It says that at any point `(r. n) the "gradient" vector
o:cd1 =
·1
·r
(r. n).
·1
·n
(r. n)
of the temperature is perpendicular to the normal vector of the position
vector
÷÷÷
(` = ri +nj. at the point `(r. n). Hence, grad1 is colinear to
÷÷÷
(`. Let us change the variables r and n with n = r and · = r
2
+ n
2
.
The new function /(n. ·) is connected to 1 by the rule:
1(r. n) = /(r. r
2
+n
2
).
So,
·1
·r
=
·/
·n
·n
·r
+
·/
··
··
·r
=
·/
·n
+ 2r
·/
··
and
·1
·n
=
·/
·n
·n
·n
+
·/
··
··
·n
= 2n
·/
··
.
Hence,
0 = n
·1
·r
÷r
·1
·n
= n
·/
·n
+ 2rn
·/
··
÷2rn
·/
··
= n
·/
·n
.
Hence, whenever n = 0.
0I
0&
= 0 is the equation in the new function
/. So / is a function of · = r
2
+ n
2
. the square of the distance up to
origin. Thus, the temperature is constant at all the points which are of
the same circle of radius : 0. We say that the level curves (1(r. n) =
constant) of the temperature are all the concentric circles with center
at (.
We must apply the "spirit" of the formulas (8.5) or (8.10), not the
formulas themselves. For instance, let
f (r. n. .) = (sin(r
2
+n
2
). cos(2.
2
). r
2
+n
2
+.
2
).
8. CHAIN RULES 159
Then,
·f
·r
= (2r cos(r
2
+n
2
). 0. 2r).
·f
·n
= (2n cos(r
2
+n
2
). 0. 2n)
and
·f
·.
= (0. ÷4. sin(2.
2
). 2.).
If we want to compute
0f
0a
(1. ÷1. 7) we simply put r = 1. n = ÷1 and
. = 7 in the expression of
0f
0a
. So,
·f
·r
(1. ÷1. 7) = (2 cos 2. 0. 2).
Here cos 2 means the cosinus of two radians.
Example 15. Let `(r(t). n(t). .(t)), t is time, t ÷ (c. /). c _ 0. be
a moving point of mass : = 51o on the curve
: r = r(t). n = n(t). . = .(t).
Let
v(t) = (r
t
(t). n
t
(t). .
t
(t))
and
w(t) = (r
tt
(t). n
tt
(t). .
tt
(t))
be the velocity and the acceleration respectively. We assume that the
kinetic energy
1 =
5
2
[r
t
(t)]
2
+ [n
t
(t)]
2
+ [.
t
(t)]
2
¸
does not depend on time, i.e. 1
t
(t) = 0. Let us use the chain rule to
make the computation in this last equality:
1
t
(t) = 5 ¦[r
t
(t)] [r
tt
(t)] + [n
t
(t)] [n
tt
(t)] + [.
t
(t)] [.
tt
(t)]¦ = 0.
i.e. the scalar (inner) product between v and w is equal to zero. In this
case, the acceleration is perpendicular on the velocity. This restriction
is very useful in physical considerations.
160 6. THE NORMED SPACE R
m
:
Definition 28. A subset 1 of R
a
is said to be a conic subset if
for any x in 1 and any t ÷ R. one has that tx ÷ 1 (see Fig.7.1).
O
K is the whole R if n = 1
K
K
K
K
O
y
n = 2
a conic body, n = 3
x
Fig. 7.1
For instance,
1 = R
a
. 1 = ¦(r. n) ÷ R
2
: n = :r¦.
where : is a …xed parameter (real number)¦.
1 = ¦(r. n. .) ÷ R
3
: r
2
+n
2
= .
2
¦
are conic subsets (prove it!).
Definition 29. Let 1 : 1 ÷ R, be a function de…ned on a conic
subset 1 · R
a
with values in R and let c be a …xed real number. We
say that 1 is homogeneous of degree c if
(8.11) 1(tr
1
. tr
2
. .... tr
a
) = t
c
1(r
1
. r
2
. .... r
a
).
for any x = (r
1
. r
2
. .... r
a
) in 1 and for any t in R
+
.
For instance, the distance to origin function
d(r. n. .) =
r
2
+n
2
+.
2
is a homogeneous function of degree 1. Indeed,
d(tr. tn. t.) =
(tr)
2
+ (tn)
2
+ (t.)
2
= t
r
2
+n
2
+.
2
= td(r. n. .).
L. Euler introduced these functions when he studied the mechanics
of a moving point in plane. For c = 0. we simply call these functions
homogeneous. Euler discovered a very useful property for homogeneous
functions. In the following we consider a generalization of the Euler’s
result.
Theorem 70. (Euler formula for homogeneous functions) Let 1
be a conic open subset in R
a
and let 1 be a function of class (
1
on 1.
which is homogeneous of degree c. Then,
(8.12) r
1
·1
·r
1
(x) +r
2
·1
·r
2
(x) +... +r
a
·1
·r
a
(x) = c 1(x).
8. CHAIN RULES 161
Proof. By the de…nition of a homogeneous function (De…nition
29), we may look at the formula (8.11) and di¤erentiate everything
w.r.t. t (here we use the chain rule...explain slowly this...)
r
1
·1
·r
1
(tx) +r
2
·1
·r
2
(tx) +... +r
a
·1
·r
a
(tx) = ct
c÷1
1(x).
We now make t = 1 in this last formula and obtain Euler formula
(8.12).
If c = 0. i.e. if our function is homogeneous, Euler formula can be
written as
(8.13) 'x. grad 1(x)` = 0.
Here '. ` is the (inner) scalar product in R
a
. This last formula (8.13)
says that at any point x of the trajectory of a moving point in R
a
.
the gradient (a generalization of the velocity for : variables!) of 1 is
perpendicular on the position vector x. For instance, we know that the
temperature 1(r. n) in any point (r. n) of the plane R
2
is the same for
all the points of an arbitrary line n = :r. where : runs freely on R.
This means (in mathematical language) that 1(tr. tn) = 1(r. n) for
any (r. n) ÷ R
2
and any t in R
+
(why?). So, the temperature is a
homogeneous function and we can write the Euler’s formula for c = 0.
i.e. 'x. grad 1(x)` = 0. where x = (r. n) and
grad 1(r. n) =
·1
·r
(r. n).
·1
·n
(r. n)
.
Finally we get the following PDE of order 1 :
r
·1
·r
(r. n) +n
·1
·n
(r. n) = 0.
i.e. in any point the gradient of the temperature is perpendicular on
the position vector (r. n).
In exercises, one usually asks to verify Euler’s formula for a given
homogeneous function 1. For instance, let us verify Euler’s formula for
1(r. n. .) = rn. + 3r
3
+ n
3
. We do not know yet if the function 1
is homogeneous and, if it is so, we also do not know the homogeneity
degree of it. Let us put instead of r. n and .. tr. tn. and t. respectively:
1(tr. tn. t.) = t
3
(rn. + 3r
3
+n
3
) = t
3
1(r. n. .).
Thus, our function is homogeneous of degree 3. So we have to verify
the following formula:
(8.14) r
·1
·r
+n
·1
·n
+.
·1
·.
= 31.
162 6. THE NORMED SPACE R
m
:
Indeed,
0;
0a
= n. + 9r
2
;
0;
0&
= r. + 3n
2
and
0;
0:
= rn. Substituting in
(8.14), we get:
r(n. + 9r
2
) +n(r. + 3n
2
) +.rn = 3(rn. + 3r
3
+n
3
) = 31.
Hence, we just veri…ed Euler’s formula for our particular function.
9. Problems
1. Compute the following partial derivatives:
a)
1(r. n) =
r
2
+n
2
;
·1
·r
(1. 1).
·
2
1
·r·n
(1. 1).
b)
1(r. n) =
sin
2
r + sin
2
n;
·1
·r
(
:
4
. 0).
·1
·n
(
:
4
.
:
4
).
c)
1(r. n) = ln(r +n
2
÷1);
·1
·r
(1. 1).
·
2
1
·n
2
(1. 1).
d)
1(r. n) = r exp(rn);
·
2
1
·r·n
(1. 0).
·
2
1
·r
2
(1. 0).
·
2
1
·n
2
(1. 0).
e)
1(r. n) = r
ln &
(r 0. n 0).
·1
·r
(c. c).
·1
·n
(c. c).
·
2
1
·r·n
(c. c).
f)
1(r. n. .) = r
&
z
(r 0. n 0). grad 1(1. 1. 1).
g)
1(r. n) = arctan rn.
·
3
1
·n·r
2
(1. 1).
·
3
1
·r·n
2
(1. 1).
·
3
1
·r
3
(1. 1).
h)
1(r. n) = arcsin(
r
n
).
·
2
1
·n·r
(1. 2).
2. Prove that the following functions verify the indicated equations:
a)
.(r. n) = rn(r
2
÷n
2
); rn
2
·.
·r
+r
2
n
·.
·n
= (r
2
+n
2
)..
b)
.(r. n) = r(r
2
÷n
2
);
1
r
·.
·r
+
1
n
·.
·n
=
.
n
2
.
9. PROBLEMS 163
c)
n(r. n) = arctan
n
r
; n
oc;
=
·
2
n
·r
2
+
·
2
n
·n
2
= 0.
d)
n(r. t) = (r ÷ct) + (r +ct);
·
2
n
·t
2
÷c
2
·
2
n
·r
2
= 0
(the wave equation).
e)
.(r. n) = r(
n
r
) + (
n
r
); r
2
·
2
.
·r
2
+ 2rn
·
2
.
·r·n
+n
2
·
2
.
·n
2
= 0.
f)
n(r. n. .) =
1
r
2
+n
2
+.
2
; n
oc;
=
·
2
n
·r
2
+
·
2
n
·n
2
+
·
2
n
·.
2
= 0.
Hint: Let us denote : =
r
2
+n
2
+.
2
. Then,
0&
0a
= ÷
1
v
2
0v
0a
. etc.
3. Show that the Euler’s formula is true for the following homoge
neous functions:
a) 1(r. n) =
a+&
a÷&
;
b)
1(r. n. .) =
r +
n +
.;
c)
1(r. n. .) =
r
2
+n
2
+.
2
;
d) 1(r. n. .) =
a
&
exp(
a
:
).
4. Prove that the following function
1(r. n) =
a&
a
2
+&
2
. for (r. n) = (0. 0)
0. if r = 0 and n = 0
is continuous, has partial derivatives, but it is not di¤erentiable at (0. 0)
(Hint:
[a&[
a
2
+&
2
_ [n[ . so
lim
a÷0.&÷0
rn
r
2
+n
2
= 0.
·1
·r
(0. 0) =
·1
·n
(0. 0) = 0.
If it was di¤erentiable at (0. 0) one has that
(9.1) 1(/
1
. /
2
) ÷1(0. 0) =
·1
·r
(0. 0)/
1
+
·1
·n
(0. 0)/
2
+.(/
1
. /
2
).
164 6. THE NORMED SPACE R
m
:
where .(0. 0) = 0. . is continuous at (0. 0) and
lim
a÷0.&÷0
.(r. n)
r
2
+n
2
= 0.
But, from (9.1), one has that .(r. n) =
a&
a
2
+&
2
and so one would have
that
lim
a÷0.&÷0
rn
r
2
+n
2
= 0.
However, this last limit does not exist at all!!).
CHAPTER 7
Taylor’s formula for several variables.
1. Higher partial derivatives. Di¤erentials of order /.
Let
0;
0a
be the partial derivative with respect to r of a function
1 : ¹ ÷ R, where ¹ is an open subset in R
2
. (r. n)
0;
0a
(r. n) is
a new function of two variables r and n. If this new function has a
partial derivative
0
0a
(
0;
0a
)(c. /) w.r.t. r. at a point (c. /). we denote it
by
0
2
;
0a
2
(c. /) and say " d two 1 over d r two at c /". If the same function
(r. n)
0;
0a
(r. n) has a partial derivative
0
0&
(
0;
0a
)(c. /) w.r.t. n. at a point
(c. /). we write it as
0
2
;
0&0a
(c. /) and call it the mixed derivative of 1 at
(c. /). What do we mean by
0
3
;
0a0&
2
(say "d three 1 over d r d n two";
pay attention to the fact that 3 from ·
3
is equal to the sum between
1 and 2. from ·r and ·n
2
respectively). In general, let 1 : ¹ ÷ R,
1(r
1
. r
2
. .... r
a
) be a function of : variables, de…ned on an open subset
¹ of R
a
. such that it is /
a
times di¤erentiable with respect to r
a
. i.e.
0
kn
;
0a
kn
n
exists on ¹. If this new function
x = (r
1
. r
2
. .... r
a
)
·
In
1
·r
In
a
(x)
is /
a÷1
times di¤erentiable with respect to r
a÷1
. the new obtained func
tion
x
·
I
n1
·r
I
n1
a÷1
·
In
1
·r
In
a
(x)
is denoted by
0
kn+k
n1
;
0a
k
n1
n1
0a
kn
n
. And so on. We …nally obtain the function
0
kn+k
n1
+:::+k
1
;
0a
k
1
1
...0a
k
n1
n1
0a
kn
n
. The order of variables r
1
. r
2
. .... r
a
in the denomina
tor can be changed, but then we may obtain another new function.
For instance, if 1(r. n. .) = r
4
n
3
.
5
. then
0
5
;
0&
2
0a
2
0:
can be successively
computed. First of all we compute
o
1
=
·1
·.
= 5r
4
n
3
.
4
.
165
166 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
Then we compute
o
2
=
·o
1
·r
=
·
2
1
·r·.
= 20r
3
n
3
.
4
.
Now we compute
o
3
=
·o
2
·r
=
·
3
1
·r
2
·.
= 60r
2
n
3
.
4
.
Then we consider
o
4
=
·o
3
·n
=
·
4
1
·n·r
2
·.
= 180r
2
n
2
.
4
.
Finally,
o
5
=
·o
4
·n
=
·
5
1
·n
2
·r
2
·.
= 360r
2
n.
4
.
And this last one is our …nal result.
0
kn+k
n1
+:::+k
1
;
0a
k
1
1
...0a
k
n1
n1
0a
kn
n
is said to be the partial / = /
a
+ /
a÷1
+ ... + /
1
derivative of 1. /
a
times w.r.t. r
a
. /
a÷1
times w.r.t. r
a÷1
. .... and /
1

times w.r.t. r
1
. The mapping 1
0;
0a
j
is also denoted by 1
a
j
1. This
1
a
j
is called the partial di¤erential operator w.r.t. the variable r
;
.
So, 1
0
2
;
0a
i
0a
j
is the composition 1
a
i
· 1
a
j
applied to 1. In general, a
mapping de…ned on a set of functions is called not a function more, but
an operator. We also put 1
a
i
a
j
instead of 1
a
i
·1
a
j
. Such an operator is
called a di¤erential operator. In general, the operators 1
a
i
and 1
a
j
do
not commute if i = ,. This means that there are examples of functions
1 and points a for which
0
2
;
0a
i
0a
j
(a) =
0
2
;
0a
j
0a
i
(a). Following [Pal], p. 145,
we consider
(1.1) 1(r. n) =
rn
a
2
÷&
2
a
2
+&
2
. if (r. n) = (0. 0)
0. if r = 0. n = 0.
It is not di¢cult to prove that
0
2
;
0&0a
(0. 0) = ÷1. but
0
2
;
0a0&
(0. 0) = 1 (do
it step by step and explain everything!). Hence, in this case we cannot
commute the order of derivation!
Let ¹ be an open subset of R
a
and let 1 : ¹ ÷R be a function of
: variable de…ned on ¹. We say that 1 is of class (
2
on ¹ if all the
partial derivatives of order two,
0
2
;
0a
i
0a
j
(a). exist and are continuous, at
any point a of ¹. The following theorem gives us a su¢cient condition
under which the change of order of derivation has no in‡uence on the
…nal result.
1. HIGHER PARTIAL DERIVATIVES. DIFFERENTIALS OF ORDER k: 167
Theorem 71. (Schwarz’ Theorem) Let 1 : ¹ ÷ R be a function of
class (
2
on ¹. Then
·
2
1
·r
j
·r
;
(a) =
·
2
1
·r
;
·r
j
(a)
for any point a of ¹ and for any pair (i. ,). This means that for such
a function (of class (
2
on ¹) we can commute the order of derivation.
Proof. One can reduce everything to the two variables case (why?).
Moreover, we can take an open ball (disc) 1(a. :). : 0. a =(c
1
. c
2
). in
cluded in ¹ and consider 1 de…ned on this ball 1(a. :). Let ¦(r
a
. n
a
)¦
be a sequence of points in 1(a. :) which converges to a. For a …xed
natural number : let us consider the segments [c
1
. r
a
] and [c
2
. n
a
] in
1(a. :). Let
(1.2) 1(r
a
. n
a
) = 1(r
a
. n
a
) ÷1(r
a
. c
2
) ÷1(c
1
. n
a
) +1(c
1
. c
2
)
and let o(t) = 1(t. n
a
) ÷ 1(t. c
2
). t ÷ [c
1
. r
a
]. Let us apply Lagrange’s
theorem (see Corollary 4) to function o on [c
1
. r
a
] :
o(r
a
) ÷o(c
1
) = o
t
(c
a
) (r
a
÷c
1
).
where c
a
÷ [c
1
. r
a
]. But
o(r
a
) ÷o(c
1
) = 1(r
a
. n
a
)
and
o
t
(c
a
) =
·1
·r
(c
a
. n
a
) ÷
·1
·r
(c
a
. c
2
).
So,
1(r
a
. n
a
) =
¸
·1
·r
(c
a
. n
a
) ÷
·1
·r
(c
a
. c
2
)
(r
a
÷c
1
).
Now we apply again Lagrange’s theorem to the function
n ÷
·1
·r
(c
a
. n).
where n ÷ [c
2
. n
a
]. Hence,
(1.3) 1(r
a
. n
a
) =
·
2
1
·n·r
(c
a
. d
a
) (r
a
÷c
1
)(n
a
÷c
2
).
where d
a
÷ [c
2
. n
a
]. Now we take a new function
/(t) = 1(r
a
. t) ÷1(c
1
. t).
t ÷ [c
2
. n
a
] and observe that
1(r
a
. n
a
) = /(n
a
) ÷/(c
2
).
Let us apply Lagrange’s theorem to / on [c
2
. n
a
] :
(1.4) 1(r
a
. n
a
) = /
t
(c
a
) (n
a
÷c
2
).
168 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
where c
a
÷ [c
2
. n
a
]. But /
t
(c
a
) =
0;
0&
(r
a
. c
a
) ÷
0;
0&
(c
1
. c
a
) so, applying
again Lagrange’s theorem to the function:
· ÷
·1
·n
(·. c
a
).
where · ÷ [c
1
. r
a
]. we get:
/
t
(c
a
) =
·
2
1
·r·n
(:
a
. c
a
) (r
a
÷c
1
).
where :
a
÷ [c
1
. r
a
]. Hence,
(1.5) 1(r
a
. n
a
) =
·
2
1
·r·n
(:
a
. c
a
) (r
a
÷c
1
)(n
a
÷c
2
).
Comparing the formulas (1.3) and (1.5), we get:
(1.6)
·
2
1
·n·r
(c
a
. d
a
) =
·
2
1
·r·n
(:
a
. c
a
).
Since the functions
0
2
;
0&0a
and
0
2
;
0a0&
are continuous on ¹, since ¦c
a
¦. ¦:
a
¦ ÷
c
1
and since ¦d
a
¦. ¦c
a
¦ ÷c
2
(why?), from formula (1.6), we get:
·
2
1
·n·r
(c
1
. c
2
) =
·
2
1
·r·n
(c
1
. c
2
).
Hence, the proof of the theorem is complete.
In (1.1)
·
2
1
·n·r
(0. 0) = ÷1 =
·
2
1
·r·n
(0. 0) = 1.
because
0
2
;
0&0a
is not continuous at (0. 0). Indeed,
·
2
1
·n·r
(r. n) =
a
6
÷&
6
÷9a
2
&
4
÷15a
4
&
2
(a
2
+&
2
)
3
. if (r. n) = (0. 0)
÷1. if r = 0. n = 0.
.
and this last function has no limit at (0. 0). This is because, if we take
an arbitrary : and consider (r. n) with n = :r. we get that
lim
a÷0.&=na
r
6
÷n
6
÷9r
2
n
4
÷15r
4
n
2
(r
2
+n
2
)
3
=
1 ÷25:
6
(1 +:
2
)
3
.
which is dependent on :. So, the limit at (0. 0) is not a unique number.
It depends on the direction on which we come to (0. 0). All of these
happen because the function
r
6
÷n
6
÷9r
2
n
4
÷15r
4
n
2
(r
2
+n
2
)
3
1. HIGHER PARTIAL DERIVATIVES. DIFFERENTIALS OF ORDER k: 169
is homogeneous of degree 0 (make clear this for yourself!)
In engineering, the case of functions of class (
2
is mostly frequent,
thus we assume in the following that the order of derivation does not
matter. For instance, 1(r. n) = 4r
3
n
2
+ 2r
2
n is of class (
o
on R
2
(why?). In particular, it is of class (
2
because (
o
means that 1 has
partial derivatives of any order (so these derivatives are continuous
why?). Schwarz’ theorem says that
·
2
1
·r·n
(c. /) =
·
2
1
·n·r
(c. /)
for any point (c. /) in R
2
. Indeed,
·
2
1
·r·n
(c. /) =
·
·r
(
·1
·n
)(c. /) =
·
·r
(8r
3
n + 2r
2
) [
(o.b)
=
= 24r
2
n + 4r [
(o.b)
= 24c
2
/ + 4c
and
·
2
1
·n·r
(c. /) =
·
·n
(
·1
·r
)(c. /) =
·
·n
(12r
2
n
2
+ 4rn) [
(o.b)
=
= 24r
2
n + 4r [
(o.b)
= 24c
2
/ + 4c.
Sometimes is more convenient to change the order of derivation.
For instance, 1(r. n) = n ln(r
2
+ n
2
+ 1) is of class (
o
on R
2
(why?).
In order to compute
0
2
;
0a0&
it is easier to compute
0
2
;
0&0a
i.e. to compute
…rstly
0;
0a
=
2a&
a
2
+&
2
+1
. and secondly
·
·n
2rn
r
2
+n
2
+ 1
=
2r(r
2
+n
2
+ 1) ÷2n 2rn
(r
2
+n
2
+ 1)
2
=
2r
3
÷2rn
2
+ 2r
(r
2
+n
2
+ 1)
2
.
then to compute …rstly
·1
·n
= ln(r
2
+n
2
+ 1) +
2n
2
r
2
+n
2
+ 1
and secondly
·
·r
¸
ln(r
2
+n
2
+ 1) +
2n
2
r
2
+n
2
+ 1
(why?count the number of operations and their di¢culties in each
case!).
The following notion will be very helpful in the applications of the
di¤erential calculus.
170 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
Definition 30. Let ¹ be an open subset in R
a
and let a = (c
1
. c
2
. .... c
a
)
be a …xed point (vector) in ¹. Let 1 be a function of class (
2
on ¹.
1 : ¹ ÷R. The symmetric matrix
H
;.a
= (:
j;
) =
·
2
1
·r
j
·r
;
(a)
. i = 1. 2. .... :; , = 1. 2. .... :
is called the Hessian matrix of 1 at a. The quadratic form d
2
1(a) de
…ned on R
a
. relative to its canonical basis
¦e
1
= (1. 0. 0. .... 0). e
2
= (0. 1. 0. .... 0). .... e
a
= (0. 0. 0. ...0. 1)¦
(see a Linear Algebra course!) with values in R,
(1.7) d
2
1(a)(/
1
. /
2
. .... /
a
) =
a
¸
j=1
a
¸
;=1
·
2
1
·r
j
·r
;
(a)/
j
/
;
.
is called the second di¤erential of 1 at a. Its matrix is exactly the
Hessian matrix of 1 at a. For instance, if 1 is a function of 2 vari
ables, r
1
= r. r
2
= n and a = (c. /). then formula (1.7) becomes
(1.8) d
2
1(c. /)(/
1
. /
2
) =
·
2
1
·r
2
(c. /)/
2
1
+2
·
2
1
·r·n
(c. /)/
1
/
2
+
·
2
1
·n
2
(c. /)/
2
2
.
If we introduce the projection functions dr
j
(/
1
. /
2
. .... /
a
) = /
j
for i =
1. 2. .... :. we get a more compact formula for (1.7)
(1.9) d
2
1(a) =
a
¸
j=1
a
¸
;=1
·
2
1
·r
j
·r
;
(a)dr
j
dr
;
.
Here, dr
j
dr
;
is the product between the two linear mappings dr
j
. dr
;
:
R
a
÷R. i.e.
dr
j
dr
;
(h) = dr
j
(h) dr
;
(h) = /
j
/
;
.
where h = (/
1
. /
2
. .... /
a
). For two variables we get
(1.10) d
2
1(c. /) =
·
2
1
·r
2
(c. /)dr
2
+ 2
·
2
1
·r·n
(c. /)drdn +
·
2
1
·n
2
(c. /)dn
2
.
where dr
2
is drdr and NOT d(r
2
) which is equal to 2rdr (why?). The
same for dn
2
... . The analogous formula for a function of 3 variables
1(r. n. .) is
d
2
1(c. /. c) =
·
2
1
·r
2
(c. /. c)dr
2
+
·
2
1
·n
2
(c. /. c)dn
2
+
·
2
1
·.
2
(c. /. c)d.
2
+
(1.11) +2
·
2
1
·r·n
(c. /. c)drdn+2
·
2
1
·r·.
(c. /. c)drd.+2
·
2
1
·n·.
(c. /. c)dnd..
1. HIGHER PARTIAL DERIVATIVES. DIFFERENTIALS OF ORDER k: 171
For instance, let us compute the second di¤erential for
1(r. n. .) = 2r
3
+ 3rn
2
. +.
3
at the point (÷1. 2. 3). First of all we compute
·
2
1
·r
2
(r. n. .) =
·
·r
(
·1
·r
)(r. n. .) =
·
·r
(6r
2
+ 3n
2
.) = 12r.
So,
0
2
;
0a
2
(÷1. 2. 3) = ÷12. It is easy to …nd
·
2
1
·n
2
(÷1. 2. 3) = ÷18.
·
2
1
·.
2
(÷1. 2. 3) = 18.
·
2
1
·r·n
(÷1. 2. 3) = 36.
·
2
1
·r·.
(÷1. 2. 3) = 12.
·
2
1
·n·.
(÷1. 2. 3) = ÷12.
Now we use (1.11) and …nd
(1.12)
d
2
1(÷1. 2. 3) = ÷12dr
2
÷18dn
2
+ 18d.
2
+ 72drdn + 24drd. ÷24dnd..
i.e. we have a quadratic form in 3 variables dr. dn. d.. Clearer, this last
quadratic form is
o(A. ). 2) = ÷12A
2
÷18)
2
+ 182
2
+ 72A) + 24A2 ÷24) 2.
Now, if we substitute A with dr. ) with dn and 2 with d.. we get
(1.12).
Let us compute the value of this last function
d
2
1(÷1. 2. 3) : R
3
÷R
at the point (2. ÷3. ÷4). Since
dr
2
(2. ÷3. ÷4) = 2
2
= 4. dn
2
(2. ÷3. ÷4) = (÷3)
2
= 9.
d.
2
(2. ÷3. ÷4) = (÷4)
2
= 16. drdn(2. ÷3. ÷4) = 2 (÷3) = ÷6.
drd.(2. ÷3. ÷4) = 2 (÷4) = ÷8. dnd.(2. ÷3. ÷4) = (÷3)(÷4) = 12.
we …nally obtain
d
2
1(÷1. 2. 3)(2. ÷3. ÷4) = ÷12 4 ÷18 9 + 18 16 + 72 (÷6)+
+24 (÷8) ÷24 12 = ÷12 4 + 7 18 + 24(÷18 ÷8 ÷12)
= ÷12 4+7 18+24 (÷38) = ÷12(4+76) +7 18 = 6(÷139) = ÷834.
Now, let us look carefully at the formulas (7.13),(1.7) and (1.9).
We introduce some symbolic operations in order to …nd a unitary and
172 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
general formula. We called
0
0a
j
a di¤erential operator. By de…nition,
we multiply two such operators
0
0a
j
and
0
0a
i
by a simple composition:
·
·r
;
·
·r
j
oc;
=
·
2
·r
;
·r
j
=
·
·r
;
·
·
·r
j
.
For instance,
·
·r
·
·n
(3r
2
+ 5rn
3
) =
·
·r
(
·
·n
(3r
2
+ 5rn
3
)) =
·
·r
(15rn
2
) = 15n
2
.
Moreover,
d1(c. /) =
·1
·r
(c. /)dr +
·1
·n
(c. /)dn
can be written as an operator "on 1" at an arbitrary point (which will
not appear)
d =
·
·r
dr +
·
·n
dn.
This is also called a di¤erential operator. How do we multiply two such
operators?
·
·r
dr +
·
·n
dn
·
·.
d. +
·
·n
dn
=
oc;
=
·
2
·r·.
drd. +
·
2
·n·.
dnd. +
·
2
·r·n
drdn +
·
2
·n·n
dndn.
This means that whenever we multiply operators we just compose
them and whenever we multiply linear mappings we just multiply them
as functions. These last are always coe¢cients of di¤erential operators.
For instance
(1.13)
·
·r
dr +
·
·n
dn
2
=
·
2
·r
2
dr
2
+ 2
·
2
·r·n
drdn +
·
2
·n
2
dn
2
.
Hence,
d
2
1(c. /) =
·
·r
dr +
·
·n
dn
2
(1)(c. /).
with this last notation. We observe that in (1.13) one has a binomial
formula of the type (c +/)
2
= c
2
+ 2c/ +/
2
(with the above indicated
multiplication between di¤erential operators). If we multiply again by
0
0a
dr +
0
0&
dn the both sides in (1.13) we easily get
·
·r
dr +
·
·n
dn
3
=
·
3
·r
3
dr
3
+3
·
3
·r
2
·n
dr
2
dn+3
·
3
·r·n
2
drdn
2
+
·
3
·n
3
dn
3
.
i.e. the analogous formula of (c +/)
3
= c
3
+ 3c
2
/ + 3c/
3
+/
3
.
1. HIGHER PARTIAL DERIVATIVES. DIFFERENTIALS OF ORDER k: 173
Definition 31. (the di¤erential of order /) In general, if a func
tion 1 of : variables, 1 : ¹ ÷ R, is of class (
I
on ¹. i.e. it has all
partial di¤erentials of the type
·
I
1
·r
I
1
1
·r
I
2
2
...·r
In
a
(a).
where / = /
1
+ /
2
+ ... + /
a
and 0 _ /
1
. /
2
. .... /
a
_ :. at any point a
of ¹. the /th di¤erential of 1 at a is by de…nition
(1.14) d
I
1(a) =
·
·r
1
dr
1
+
·
·r
2
dr
2
+... +
·
·r
a
dr
a
I
(1)(a).
For instance, if : = 2. r
1
= r, r
2
= n and a =(c. /). then this last
formula becomes
(1.15)
d
I
1(c. /) =
·
·r
dr +
·
·n
dn
I
(1)(c. /) =
I
¸
j=0
/
i
·
I
1
·r
I÷j
·n
j
(c. /)dr
I÷j
dn
j
.
where
I
j
=
I!
j!(I÷j)!
is (the combination) of / objects taken i. The
analogy with the binomial formula
(c +/)
I
=
I
¸
j=0
/
i
c
I÷j
/
j
is now clear.
Let us compute
d
4
1(1. ÷1) =
·
·r
dr +
·
·n
dn
4
(1)(1. ÷1)
for 1(r. n) = r
5
+rn
4
. For / = 4 formula (1.15) becomes
·
·r
dr +
·
·n
dn
4
(1)(1. ÷1) =
4
0
·
4
1
·r
4
(1. ÷1)dr
4
+
4
1
·
4
1
·r
3
·n
(1. ÷1)dr
3
dn +
4
2
·
4
1
·r
2
·n
2
(1. ÷1)dr
2
dn
2
+
4
3
·
4
1
·r·n
3
(1. ÷1)drdn
3
+
4
4
·
4
1
·n
4
(1. ÷1)dn
4
.
Now, everything reduces to the computation of the mixed partial
derivatives.
·
4
1
·r
4
(1. ÷1) = 120.
·
4
1
·r
3
·n
(1. ÷1) = 0.
·
4
1
·r
2
·n
2
(1. ÷1) = 0.
174 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
·
4
1
·r·n
3
(1. ÷1) = ÷24.
·
4
1
·r·n
3
(1. ÷1) = ÷24.
·
4
1
·n
4
(1. ÷1) = 24.
Hence,
·
·r
dr +
·
·n
dn
4
(1)(1. ÷1) = 120dr
4
÷96drdn
3
+ 24dn
4
.
If we want to compute the value of this last di¤erential at (2. 3) for
instance, we obtain
120 2
4
÷96 2 3
3
+ 24 3
4
= ÷1320.
Let us now compute
d
2
1(1. 1. 0) =
·
·r
dr +
·
·n
dn +
·
·.
d.
2
(1)(1. 1. 0)
for 1(r. n. .) = r
2
+n
2
+r.+n.. To be easier, let us recall the elementary
algebraic formula:
(c +/ +c)
2
= c
2
+/
2
+c
2
+ 2c/ + 2cc + 2/c.
Using the above multiplicity between operators, etc., we get
d
2
1(1. 1. 0) =
·
2
1
·r
2
(1. 1. 0)dr
2
+
·
2
1
·n
2
(1. 1. 0)dn
2
+
·
2
1
·.
2
(1. 1. 0)d.
2
+ 2
·
2
1
·r·n
(1. 1. 0)drdn + 2
·
2
1
·r·.
(1. 1. 0)drd.+
2
·
2
1
·n·.
(1. 1. 0)dnd. = 2dr
2
+ 2dn
2
+ 2drd. + 2dnd..
If one wants to compute d
2
1(1. 1. 0)(3. 4. 5) we get
d
2
1(1. 1. 0)(3. 4. 5) = 2 3
2
+ 2 4
2
+ 2 3 5 + 2 4 5 = 120.
Since
(c
1
+c
2
+... +c
a
)
n
=
¸
I
1
+I
2
+...+In=n.I
i
÷N
:!
/
1
!/
2
!.../
a
!
c
I
1
1
c
I
2
2
...c
In
a
.
one has the following de…nition of the :th di¤erential of 1 at a point
a ÷ ¹
d
n
1(a) =
·
·r
1
dr
1
+
·
·r
2
dr
2
+... +
·
·r
a
dr
a
n
=
¸
I
1
+I
2
+...+In=n.I
i
÷N
:!
/
1
!/
2
!.../
a
!
·
n
1
·r
I
1
1
·r
I
2
2
...·r
In
a
dr
I
1
1
r
I
2
2
...dr
In
a
.
2. CHAIN RULES IN TWO VARIABLES 175
2. Chain rules in two variables
During the mathematical modeling process of the physical phenom
ena, usually one must …nd functions . = .(r. n) which verify an equality
of the following form (a partial di¤erential equation of order 2. i.e. a
PDE):
¹(r. n)
·
2
.
·r
2
(r. n) + 21(r. n)
·
2
.
·r·n
(r. n) +((r. n)
·
2
.
·n
2
(r. n)
(2.1) +1
r. n. .(r. n).
·.
·r
(r. n).
·.
·n
(r. n)
= 0.
where ¹. 1. (. 1 are continuous functions of the indicated free vari
ables. Relative to 1 we must add that it is a continuous function
1(A. ). 2. l. \ ) of 5 free variables, where instead of A. ). 2. l. \. we
put r. n. .(r. n).
0:
0a
(r. n) and
0:
0&
(r. n) respectively. In order to …nd all
the functions .(r. n) of class (
2
on a …xed plane domain 1. which ver
i…es (2.1) we change the "old" variables r, n with new ones n = n(r. n)
and · = ·(r. n) respectively (functions of the …rsts) such that some
of the new "coe¢cients" ¹. 1. or ( to become zero. How do we …nd
these new functions n = n(r. n) and · = ·(r. n) is a problem which will
be considered in another course. Our problem here is how to write the
partial derivatives.
·
2
.
·r
2
(r. n).
·
2
.
·r·n
(r. n).
·
2
.
·n
2
(r. n).
·.
·r
(r. n).
·.
·n
(r. n)
as functions of n and ·. Let .(n. ·) be the new searched function of the
new variables n and ·. i.e.
.(n(r. n). ·(r. n)) = .(r. n)
(see also the section "Change of variables"). The chain rules formulas
(8.9) and (8.10) supply us with formulas for
0:
0a
(r. n) and
0:
0&
(r. n) :
(2.2)
·.
·r
(r. n) =
·.
·n
(n(r. n). ·(r. n))
·n
·r
(r. n) +
·.
··
(n(r. n). ·(r. n))
··
·r
(r. n).
and
(2.3)
·.
·n
(r. n) =
·.
·n
(n(r. n). ·(r. n))
·n
·n
(r. n) +
·.
··
(n(r. n). ·(r. n))
··
·n
(r. n).
Let us use these formulas to …nd a similar formula for
0
2
:
0a0&
(r. n). For
this, let us denote by o(r. n) and by /(r. n) the new functions of r and
176 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
n obtained in (2.3)
o(r. n)
oc;
=
·.
·n
(n(r. n). ·(r. n))
and
·.
··
(n(r. n). ·(r. n))
oc;
= /(r. n).
Let us compute
0j
0a
(r. n) and
0I
0a
(r. n) by using the formula (2.2) with o
instead of . and / instead of . respectively:
·o
·r
(r. n) =
·
·n
·.
·n
(n(r. n). ·(r. n))
·n
·r
(r. n)+
(2.4)
·
··
·.
·n
(n(r. n). ·(r. n))
··
·r
(r. n) =
·
2
.
·n
2
(n(r. n). ·(r. n))
·n
·r
(r. n)+
·
2
.
···n
(n(r. n). ·(r. n))
··
·r
(r. n).
and
·/
·r
(r. n) =
·
·n
·.
··
(n(r. n). ·(r. n))
·n
·r
(r. n)+
(2.5)
·
··
·.
··
(n(r. n). ·(r. n))
··
·r
(r. n) =
·
2
.
·n··
(n(r. n). ·(r. n))
·n
·r
(r. n)+
·
2
.
··
2
(n(r. n). ·(r. n))
··
·r
(r. n).
Let us come back to formula (2.3) and let us di¤erentiate it (both sides)
with respect to r. We get:
·
2
.
·r·n
(r. n) =
·o
·r
(r. n)
·n
·n
(r. n) +o
·
2
n
·r·n
(r. n)+
·/
·r
(r. n)
··
·n
(r. n) +/
·
2
·
·r·n
(r. n).
If we take count of the formulas (2.4) and (2.5) we …nally obtain:
(2.6)
·
2
.
·r·n
(r. n) =
·
2
.
·n
2
(n(r. n). ·(r. n))
·n
·r
(r. n)
·n
·n
(r. n)+
+
·
2
.
·n··
(n(r. n). ·(r. n))
¸
·n
·r
(r. n)
··
·n
(r. n) +
·n
·n
(r. n)
··
·r
(r. n)
+
+
·
2
.
··
2
(n(r. n). ·(r. n))
··
·r
(r. n)
··
·n
(r. n)+
2. CHAIN RULES IN TWO VARIABLES 177
+
·.
·n
(n(r. n). ·(r. n))
·
2
n
·r·n
(r. n) +
·.
··
(n(r. n). ·(r. n))
·
2
·
·r·n
(r. n).
We can simply rewrite this formula as:
·
2
.
·r·n
=
·
2
.
·n
2
·n
·r
·n
·n
+
·
2
.
·n··
¸
·n
·r
··
·n
+
·n
·n
··
·r
+
+
·
2
.
··
2
··
·r
··
·n
+
·.
·n
·
2
n
·r·n
+
·.
··
·
2
·
·r·n
.
If in this formula, we formally put r instead of n we get another useful
formula:
(2.7)
·
2
.
·r
2
=
·
2
.
·n
2
·n
·r
2
+ 2
·
2
.
·n··
·n
·r
··
·r
+
·
2
.
··
2
··
·r
2
+
·.
·n
·
2
n
·r
2
+
·.
··
·
2
·
·r
2
.
If here, in this last formula, we put n instead of r. we get the last useful
chain rule formula:
(2.8)
·
2
.
·n
2
=
·
2
.
·n
2
·n
·n
2
+ 2
·
2
.
·n··
·n
·n
··
·n
+
·
2
.
··
2
··
·n
2
+
·.
·n
·
2
n
·n
2
+
·.
··
·
2
·
·n
2
.
Example 16. (vibrating string equation) Let o be a onedimensional
elastic wire (in…nite, homogeneous and perfect elastic) which vibrates
freely, without an exterior perturbing force. It is considered to lay on
the real line (r. Let n _ 0 be time and let .(r. n) be the de‡ection of
the string in the point ` of coordinate r and at the moment n. If one
write the D’Alembert equality, which makes equal the dynamic New
tonian force and the Hook elasticity force, we get a PDE of order 2
(the vibrating string equation):
(2.9)
·
2
.
·n
2
= c
2
·
2
.
·r
2
.
where c 0 is a constant depending on the density and on the elasticity
modulus. In order to …nd all the functions . = .(r. n) which verify the
equality (2.9), i.e. to solve that equation, we must change the variables
r and n with new ones n = r ÷cn and · = r +cn (see the Di¤erential
Equations course). Let us use chain formulas (2.7) and (2.8) in order
to change the variables in the equation (2.9):
·
2
.
·r
2
=
·
2
.
·n
2
+ 2
·
2
.
·n··
+
·
2
.
··
2
.
178 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
and
·
2
.
·n
2
=
·
2
.
·n
2
c
2
÷2
·
2
.
·n··
c
2
+
·
2
.
··
2
c
2
.
If we substitute these expressions in (2.9) we …nally get
(2.10)
·
2
.
·n··
= 0.
But this last PDE of order 2 can easily be solved. From 2.10 we obtain:
0
0&
0:
0·
= 0. i.e.
0:
0·
is only a function /(·). Hence,
.(n. ·) =
/(·)d· = 1(·) +o(n)
(why?), where 1 and o are two arbitrary functions of class (
2
on some
open real subsets. Coming back to r and n we …nally get the "general
solution" of the vibrating string equation:
.(r. n) = 1(r +cn) +o(r ÷cn).
Other examples in which we use higher chain rules (here "higher"
means 2 1!) will appear in the section "Change of variables".
3. Taylor’s formula for several variables
In Theorem 44 we obtained an approximation of a function of one
variable, of class (
n+1
on an neighborhood (c ÷ . c + ) of a …xed
point c. with a polynomial (the Taylor’s polynomial) of degree : (: is
a …xed natural number). We also estimated the error in this approxi
mative process. We write again this classical and fundamental formula
and try to generalize it to the case of a function of : variables.
(3.1) 1(r) = 1(c)+
1
t
(c)
1!
(r ÷c)+
1
tt
(c)
2!
(r ÷c)
2
+...+
1
(a)
(c)
:!
(r÷c)
a
+
1
(a+1)
(c)
(: + 1)!
(r ÷c)
a+1
where c is a number between r and c. Let us write again formula (3.1)
by putting / = r ÷ c. or r = c + / and c = c + t/. where t ÷ (0. 1)
(t =
c÷o
a÷o
. why?):
(3.2)
1(c+/) = 1(c)+
1
t
(c)
1!
/+
1
tt
(c)
2!
/
2
+...+
1
(a)
(c)
:!
/
a
+
1
(a+1)
(c +t/)
(: + 1)!
/
a+1
.
It is enough to generalize this formula for a scalar function of : variables
because, if f = (1
1
. 1
2
. .... 1
I
) is a vector function with / components,
we simply write the Taylor formula for any component, separately, i.e.
we approximate componentwisely.
3. TAYLOR’S FORMULA FOR SEVERAL VARIABLES 179
Let ¹ be an open subset of R
a
and let 1 : ¹ ÷R be a function of
class (
n+1
on ¹. Let a = (c
1
. c
2
. .... c
a
) be a …xed point of ¹ and let
\ = 1(a. :) be an :dimensional open ball (see its de…nition in Chapter
6, Section 1) with centre at a and of radius : 0 which is contained
in ¹ (why such thing is possible?). If a point x = (r
1
. r
2
. .... r
a
) is in
the ball \. the whole segment
[a. x] = ¦z = a+t(x ÷a) : t ÷ [0. 1]¦
is contained in \ (why?in general, a ball is a convex subset...prove
it!). A subset ( of R
a
is said to be convex if whenever a and b are in
(. the whole segment [a. b] is contained in (.
Theorem 72. (Taylor’s formula for : variables) With the above
notation and hypotheses, for any h = (/
1
. /
2
. .... /
a
) small enough, such
that x = a +h ÷ \ (h < :), one has the following Taylor’s formula:
(3.3) 1(a +h) = 1(a)+
1
1!
d1(a)(h)+
1
2!
d
2
1(a)(h)+...+
1
:!
d
n
1(a)(h)
+
1
(:+ 1)!
d
n+1
1(c)(h).
where c ÷ (a. a +h). i.e. c = a+t
+
h for a t
+
÷ (0. 1).
Proof. (: = 2) Let
a = (c
1
. c
2
). x = (r
1
. r
2
). h = (/
1
. /
2
). /
1
= r
1
÷c
1
. /
2
= r
2
÷c
2
.
The segment [a. x] is the usual segment with ends a and x in the plane
r(n (see Fig. 8.1). Let us restrict 1 to the segment [a. x]. This means
that to any point a+th. t ÷ [0. 1] we assign the number 1(a+th). One
obtains a mapping t 1(a+th). denoted here by o : [0. 1] ÷R,
o(t) = 1(a+th) =1(c
1
+t/
1
. c
2
+t/
2
).
Let us denote by n
1
and n
2
the functions n
1
(t) = c
1
+ t/
1
and respec
tively n
2
(t) = c
2
+t/
2
. So, if
u(t) = (c
1
+t/
1
. c
2
+t/
2
).
i.e. if u = (n
1
. n
2
). one has that g = f · u. Here u is a continuous one
toone mapping from [0. 1] onto [a. x]. Since u is of class (
o
on [0. 1]
(why?), we see that o is of class (
n+1
on [0. 1]. Let us apply Mac
Laurin’s formula (1.16) (or the general Taylor formula (3.1) with c = 0
and r = 1) for the function o :
(3.4)
o(1) = o(0) +
1
1!
o
t
(0) +
1
2!
o
tt
(0) +... +
1
:!
o
(n)
(0) +
1
(:+ 1)!
o
(n+1)
(t
+
).
180 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
where t
+
÷ (0. 1). Since o(1) = 1(a +h) and o(0) = 1(a). one has only
to prove that o
(I)
(0) = d
I
1(a)(h) for any / = 1. 2. .... : + 1. We can
use mathematical induction to prove this. Here, we prove only that
o
t
(0) = d1(a)(h) and that o
tt
(0) = d
2
1(a)(h). For this purpose we use
the chain rules formulas and the de…nition of the di¤erential of order
/. Indeed,
(3.5) o
t
(t) =
·1
·r
1
[n
1
(t). n
2
(t)] n
t
1
(t) +
·1
·r
2
[n
1
(t). n
2
(t)] n
t
2
(t).
Hence,
o
t
(0) =
·1
·r
1
(c
1
. c
2
) /
1
+
·1
·r
2
(c
1
. c
2
) /
2
= d1(a)(h).
Let us use the formula (3.5) to compute o
tt
(t) :
o
tt
(t) =
·
2
1
·r
2
1
[n
1
(t). n
2
(t)] [n
t
1
(t)]
2
+
·
2
1
·r
1
·r
2
[n
1
(t). n
2
(t)] n
t
1
(t) n
t
2
(t)+
·1
·r
1
[n
1
(t). n
2
(t)] n
tt
1
(t) +
·
2
1
·r
1
·r
2
[n
1
(t). n
2
(t)] n
t
1
(t) n
t
2
(t)+
·
2
1
·r
2
2
[n
1
(t). n
2
(t)] [n
t
2
(t)]
2
+
·1
·r
2
[n
1
(t). n
2
(t)] n
tt
2
(t).
Since n
tt
1
(t) = 0 and n
tt
2
(t) = 0. one has:
o
tt
(0) =
·
2
1
·r
2
1
(a) /
2
1
+ 2
·
2
1
·r
1
·r
2
(a) /
1
/
2
+
·
2
1
·r
2
2
(a) /
2
2
= d
2
1(a)(h).
If we take c = a+t
+
h. one gets the formula (3.3) for : = 2.
3. TAYLOR’S FORMULA FOR SEVERAL VARIABLES 181
a
c
x
g(t)
f(x)
O
O
x
t 0
1
A
R
x
x
Fig 8.1
y
Let
1(r. n) = 2r
2
n + 3rn
2
+r +n
be a polynomial of two variables r and n. Let us write 1(r. n) as a
polynomial ((r ÷1. n + 2). i.e.
1(r. n) = c
00
+c
10
(r÷1) +c
01
(n +2) +c
20
(r÷1)
2
+c
11
(r÷1)(n +2)+
c
02
(n + 2)
2
+c
30
(r ÷1)
3
+c
21
(r ÷1)
2
(n + 2)+
c
12
(r ÷1)(n + 2)
2
+c
03
(n + 2)
3
.
We stop here because the "total" degree of 1(r. n) is 3 = 2 + 1. We
could …nd the coe¢cients c
j;
by elementary tricks (do it!). However,
let us use Taylor formula (3.3) with
a = (1. ÷2). x = (r. n). /
1
= r ÷1. /
2
= n + 2.
etc. We have only to compute d1(a). d
2
1(a) and d
3
1(a) (why not
d
4
1(a)?). So,
d1(a) =
·1
·r
(a)dr +
·1
·n
(a)dn = (4rn + 3n
2
+ 1) [
(1.÷2)
dr
+(2r
2
+ 6rn + 1) [
(1.÷2)
dn = 5dr ÷9dn
Thus,
d1(a)(h) = 5(r ÷1) ÷9(n + 2).
182 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
Hence,
c
00
= 1(1. ÷2) = 7; c
10
= 5; c
01
= ÷9.
The coe¢cients c
20
. c
11
and c
02
can be computed from the expression
of
1
2!
d
2
1(a)(h). Namely,
·
2
1
·r
2
(a) = (4n) [
(1.÷2)
= ÷8.
·
2
1
·r·n
(a) = (4r + 6n) [
(1.÷2)
= ÷8
and
0
2
1
0&
2
(a) = 6r [
(1.÷2)
= 6. i.e.
1
2!
d
2
1(a)(h) = ÷4(r ÷1)
2
÷8(r ÷1)(n + 2) + 3(n + 2)
2
and so, c
20
= ÷4. c
11
= ÷8 and c
02
= 3. In order to …nd c
30
. c
21
. c
12
and c
03
one must compute
1
3!
d
3
1(a)(h) =
1
6
¸
·
3
1
·r
3
(a)(r ÷1)
3
+ 3
·
3
1
·r
2
·n
(a)(r ÷1)
2
(n + 2)
+3
·
3
1
·r·n
2
(a)(r ÷1)(n + 2)
2
+
·
3
1
·n
3
(a)(n + 2)
3
= 2(r ÷1)
2
(n + 2) + 3(r ÷1)(n + 2)
2
.
Thus, c
30
= 0; c
21
= 2; c
12
= 3 and c
03
= 0. Finally one has:
1(r. n) = 7 + 5(r ÷1) ÷9(n + 2) ÷4(r ÷1)
2
÷8(r ÷1)(n + 2)+
+3(n + 2)
2
+ 2(r ÷1)
2
(n + 2) + 3(r ÷1)(n + 2)
2
.
Theorem 73. (Lagrange’s Theorem for many variables, or the
Mean Value Theorem) Let ¹ · R
a
be an open subset of R
a
. let a
be a point in ¹ and let \ = 1(a. :) · ¹. : 0 be a ball with centre at
a and of radius :. Let 1 : ¹ ÷R. be a function of class (
1
de…ned on
¹. Then, for any x in A. there is a point c in [a. x] such that:
(3.6)
1(x) ÷1(a) =
·1
·r
1
(c)(r
1
÷c
1
) +... +
·1
·r
a
(c)(r
a
÷c
a
) = 'grad 1(c). h` .
i.e. the "increasing" 1(x) ÷1(a) of 1 on the interval [a. x] is equal to
the scalar product between the gradient vector grad 1(c) of 1 at a point
c of the segment [a. x]. and the the vector x ÷a. If x is very close to
a. then we have an "a¢ne" approximation of 1(x) :
(3.7) 1(x)  1(a) +
·1
·r
1
(a)(r
1
÷c
1
) +... +
·1
·r
a
(a)(r
a
÷c
a
).
or a linear approximation of 1(x) ÷1(a) :
(3.8)
1(x)÷1(a) 
·1
·r
1
(a)(r
1
÷c
1
)+... +
·1
·r
a
(a)(r
a
÷c
a
) = 'grad 1(a). h` .
4. PROBLEMS 183
Proof. It is su¢cient to take : = 0 in the formula (3.3).
From formula (3.7) we see that it is su¢cient to know the gradient
vector grad 1(a) of a function 1 at a point a and the value 1(a) of the
same function at a. in order to approximate the values of this functions
in a neighborhood of a. For instance, let us compute approximately
sin 46
·
cos 1
·
. For this, let us consider the function of two variables
1(r. n) = sin r cos n. the point a = (
¬
4
. 0) and the point x = (
¬
4
+
¬
180
.
¬
180
). Then, formula (3.7) says that: sin 46
·
cos 1
·

2
2
+
2
2
¬
180
.
4. Problems
1. Compute d1 and d
2
1 for:
a)
1(r. n) = sin(r
2
+n
2
);
b)
1(r. n. .) =
r
2
+n
2
+.
2
;
c)
1(r. n) = exp(rn)
at (1. 1); …nd also d1(1. 1)(0. 1) and d
2
1(1. 1)(0. 1).
2. Approximate 1 = 1(r. n) ÷ 1(r
0
. n
0
) by d1(r
0
. n
0
)(r. n).
where r = r ÷r
0
. n = n ÷n
0
and then compute:
a)
1(r. n) = r
ln &
at the point ¹(c + 0.1. 1 + 0.2);
b)
1(r. n) =
r
2
+n
2
at ¹(4.001. 3.002);
c)
1(r. n) = r
&
at ¹(1.02; 3.01).
3. Use Taylor’s formula to approximate 1 by the Taylor polynomial
1
a
with Lagrange’s remainder:
a)
1(r. n) = ln(1 +r) + ln(1 + n)
at (0. 0). with 1
4
;
b)
1(r. n) = r
&
at (1. 1). with 1
3
and compute approximately (1.1)
1.2
;
c)
1(r. n) = (exp r) sin n
184 7. TAYLOR’S FORMULA FOR SEVERAL VARIABLES.
at (0. 0) with 1
2
;
d)
1(r. n. .) = r
3
+n
3
+.
3
÷3rn.
at (1. 1. 1). with 1
2
.
4. Write
1(r. n) = 2r
3
÷3r
2
n + 2n
3
+ 9r
2
÷3n + 6r + 3
as ((r + 1. n ÷1).
5. Compute approximately (0.95)
2.01
; Hint: take
o(r. n) = n
a
around ¹(2. 1) and use 1
2
.
6. Compute d
2
1(0. 0. 0) for
1(r. n. .) = r
2
+n
3
+.
4
÷2rn
2
+ 3n. ÷5r
2
.
2
.
7. Compute d
3
1(0. 0)(0. 0) for
1(r. n) = cos(3r + 2n).
8. Prove that
n(r. t) =
1
2c
:t
exp
÷
(r ÷/)
2
4c
2
t
verify the "heat equation":
0&
0t
(r. t) = c
2 0
2
&
0a
2
(r. t).
9. Use Taylor’s formula to justify the following approximations:
a)
cos r
cos n
 1 ÷
r
2
÷n
2
2
around (0. 0);
b)
arctan
r +n
1 +rn
 r +n.
around (0. 0);
c)
ln(1 +r) ln(1 +n)  rn.
around (0. 0).
10. Find d1(1. ÷2)(2. 3); d
2
1(1. ÷2)(2. 3) and d
3
1(1. ÷2)(2. 3) for
1(r. n) = r
3
+ 2r
2
n.
CHAPTER 8
Contractions and …xed points
1. Banach’s …xed point theorem
Let (A. d) be a metric space, i.e. a set A with a distance function
d on it. This function d associates to any pair (r. n) of elements of A
a nonnegative real number d(r. n) with the following properties:
i) d(r. n) = 0 if and only if r = n.
ii) d(r. n) = d(n. r) for any r. n in A and
iii) d(r. .) _ d(r. n) + d(n. .) for any r. n. . in A (the triangle
inequality).
This triangle inequality can be generalized and one obtains the
polygon inequality:
(1.1) d(r
0
. r
a
) _ d(r
0
. r
1
) +d(r
1
. r
2
) +d(r
2
. r
3
) +... +d(r
a÷1
. r
a
).
for any …nite sequence ¦r
0
. r
1
. r
2
. .... r
a
¦ of A. It can be easily proved
if we use mathematical induction on :. For : = 1. or 2. it is clear.
Suppose : 2 and assume that the polygon inequality is true for any
sequence of / _ : elements of A. Let us prove it for a sequence of :+1
elements ¦r
0
. r
1
. r
2
. .... r
a
¦. Thus,
(1.2) d(r
0
. r
a÷1
) _ d(r
0
. r
1
)+d(r
1
. r
2
)+d(r
2
. r
3
)+...+d(r
a÷2
. r
a÷1
).
Now,
d(r
0
. r
a
) _ d(r
0
. r
a÷1
) +d(r
a÷1
. r
a
) _
[d(r
0
. r
1
) +d(r
1
. r
2
) +d(r
2
. r
3
) +... +d(r
a÷2
. r
a÷1
)] +d(r
a÷1
. r
a
).
and the proof of (1.1) is done.
We just met many examples of metric spaces: (R. d(r. n) = [r ÷n[).
(C. d(.. n) = [. ÷n[). (R
a
. d(x. y) =x ÷y). ([c. /] = ¦1 : [c. /] ÷
R. 1 continuous¦ with
d(1. o) = 1 ÷o = sup¦[1(r) ÷o(r)[ : r ÷ [c. /]¦.
etc. All of these metric spaces are complete metric spaces, i.e. metric
spaces (A. d) with the property that any Cauchy sequence has a limit
in A. Not all metric spaces are complete. For instance, A = (0. 1] with
the same distance like that of R is not complete, because the sequence
185
186 8. CONTRACTIONS AND FIXED POINTS
¦
1
a
¦ is a Cauchy sequence in A but it has no limit in A (why?). It is
easy to see that a subset ) of a metric space (A. d) is complete relative
to the same distance like that of A if and only if it is closed in A (prove
it!).
Definition 32. (contraction) Let (A. d) be a metric space. A func
tion 1 : A ÷ A is said to be a contraction on A if there is a number
\ ÷ (0. 1) such that
(1.3) d(1(r). 1(n)) _ \d(r. n)
for any r. n in A. This number \ is called the (contraction) coe¢cient
of 1.
For instance, 1 : [0. 1] ÷ [0. 1]. 1(r) = 0.5r is a contraction of co
e¢cient 0.5 (prove it!). But o : R ÷R, o(r) = 2r. is not a contraction
on R but,...it is a contraction on [0. 0.44] (prove it!).
Any contraction on A is a uniformly continuous function on A
(why?). The same result is true even \ is an arbitrary positive real
number. In this more general case we say that 1 is a Lipschitzian
function on A.
Theorem 74. Let ¹ be a convex subset of R
a
(if a and b are in ¹.
then the whole segment [a. b] is in ¹. Let 1 : ¹ ÷ ¹ be a function of
class (
1
on ¹ such that all the partial derivatives of 1 are bounded by
a number of the form \:. where \ ÷ (0. 1). Then 1 is a contraction of
coe¢cient \ on ¹.
Proof. Let us take a. b in ¹ and let us write Taylor’s formula for
: = 0 (b = a +h):
(1.4)
1(b) ÷f (a) =
·1
·r
1
(c) (/
1
÷c
1
)+
·1
·r
2
(c) (/
2
÷c
2
)+...+
·1
·r
a
(c) (/
a
÷c
a
).
where c is a point on the segment [a. b] and a = (c
1
. c
2
. .... c
a
). b =
(/
1
. /
2
. .... /
a
).
So,
d(1(a).1(b)) =1(b) ÷1(a) _
a
¸
j=1
·1
·r
j
(c)
a ÷b
_
¸
a
¸
j=1
·1
·r
j
(c)
¸
a ÷b _ \d(a. b).
Thus, our function is a contraction.
For instance, 1(r) =
1
5
r
3
is a contraction on [0. 1]. because [1
t
(r)[ =
3
5
[r
2
[ _
3
5
on [0. 1].
1. BANACH’S FIXED POINT THEOREM 187
Theorem 75. (Banach’s …xed point theorem) Let (A. d) be a com
plete metric space and let 1 : A ÷ A be a contraction of coe¢cient
\ ÷ (0. 1). Then there is a unique element r in A such that 1(r) = r
(a …xed point for 1). This unique …xed point r of 1 on A can be ob
tained by the following method (the successive approximates method).
Start with an arbitrary element r
0
of A and recurrently construct:
r
1
= 1(r
0
). r
2
= 1(r
1
). .... r
a
= 1(r
a÷1
). .... Then, the sequence ¦r
a
¦
is convergent to this …xed point r. Moreover, if we approximate r by
r
a
. the error d(r. r
a
) can be evaluated by the following formula
(1.5) d(r. r
a
) _ d(r
1
. r
0
)
\
a
1 ÷\
.
Proof. It is su¢cient to prove that ¦r
a
¦ is a Cauchy sequence
(why?remember that A is complete so, r
a
÷r. then use the continuity
of 1 in the recurrence relationtake limits and …nd r = 1(r)). Let us
evaluate the distance between the terms of the sequence ¦r
a
¦ by using
the contraction formula (1.3).
d(r
2
. r
1
) = d(1(r
1
). 1(r
0
)) _ \d(r
1
. r
0
).
d(r
3
. r
2
) = d(1(r
2
). 1(r
1
)) _ \d(r
2
. r
1
) _ \
2
d(r
1
. r
0
).
and so on, up to a general relation (use mathematical induction if you
want!):
(1.6) d(r
a+1
. r
a
) _ \
a
d(r
1
. r
0
).
Now,
(1.7)
d(r
a+j
. r
a
) _ d(r
a+j
. r
a+j÷1
) +d(r
a+j÷1
. r
a+j÷2
) +... +d(r
a+1
. r
a
)
comes from applying of the polygon inequality (1.1). If in (1.7) we
introduce the formula from (1.6), we get:
d(r
a+j
. r
a
) _ (\
a+j÷1
+\
a+j÷2
+... +\
a
)d(r
1
. r
0
)
(1.8) _ \
a
(1 +\ +\
2
+...)d(r
1
. r
0
) =
\
a
1 ÷\
d(r
1
. r
0
).
Since
A
n
1÷A
÷ 0. independently on j. the sequence ¦r
a
¦ is a Cauchy
sequence. Since (A. d) is complete, this sequence has a limit r = limr
a
.
Making j ÷· in (1.8) we get the desired estimation of the error:
d(r. r
a
) _
\
a
1 ÷\
d(r
1
. r
0
).
188 8. CONTRACTIONS AND FIXED POINTS
(why d(r
a+j
. r
a
) ÷d(r. r
a
) if j ÷·? Prove it!). Since r
a
= 1(r
a÷1
)
and since 1 is continuous, one has that r = 1(r). This …xed point r is
unique. Indeed, if r = 1(r) and n = 1(n). then
d(r. n) = d(1(r). 1(n)) _ \d(r. n).
or
d(r. n) [\ ÷1] _ 0.
Since \ ÷ (0. 1) and since d(r. n) _ 0. the unique possibility is that
d(r. n) = 0. i.e. r = n.
The Banach’s …xed point theorem has many applications. For in
stance, it can be used to …nd approximate solutions for equations and
system of equations (linear or not!).
Take for example the polynomial
1(r) = r
3
÷r
2
+ 2r ÷1
and let us search for a solution of the equation 1(r) = 0 in the interval
A = [0.
1
3
]. The equation r
3
÷r
2
+ 2r ÷1 = 0 can also be written as:
(1.9)
r
2
+ 1
r
2
+ 2
= r.
Let us prove that 1(r) =
a
2
+1
a
2
+2
is a contraction on [0.
1
3
]. Indeed, 1
t
(r) =
2a
(a
2
+2)
2
and
2r
(r
2
+ 2)
2
_
1
6
(why?) on [0.
1
3
]. Applying Theorem 74
we get that 1 is a contraction of coe¢cient \ =
1
6
. So, the equation
(1.9) has a unique solution r in [0.
1
3
]. Let us …nd it approximately with
"two exact decimals". Formula (1.5) says that:
[r ÷r
a
[ _
1
6
a
6
5
[r
1
÷r
0
[ .
Let us take r
0
= 0. Then r
1
= 1(r
0
) =
1
2
. Thus,
[r ÷r
a
[ _
1
10 6
a÷1
.
If we force with
1
10·6
n1
_
1
10
2
. we get : = 3. Hence, the true solution r
is approximately equal to r
3
= 1(1(
1
2
)) = 1(
21
41
). etc. The committed
error is less then 0.01.
2. PROBLEMS 189
2. Problems
1. Using the Banach’s Fixed Point Theorem, …nd approximate
solutions with the error = 10
÷2
for the following equations:
a) r
3
+r ÷5 = 0; b) r
3
÷sin r = 3; c) r =
¬
3
3
cos r.
2. Which of the following mappings are contractions? Study the
…xed points of them.
a) 1 : R ÷R, 1(r) = r; b) 1 : R ÷R, 1(r) = r
7
; c) 1 : C ÷C,
1(.) = .
4
;
d) 1 : C ÷C, 1(.) = .
2
+. + 1; e) 1 : R ÷R, 1(r) =
1
5
r + 3;
f) 1 : R ÷R, 1(r) =
1
5
arctan r; g) 1 : R ÷R, 1(r. n) = (
1
7
r.
1
8
n).
3. Try to …nd approximate solutions with 2 exact decimals for the
following linear system of algebraic equations:
100r + 2n = 1
4r + 200n = 5
.
Hint: Write this system as:
0.01 ÷0.02n = r
0.025 ÷0.02r = n
.
Prove that the vector function f : R
2
÷ R
2
. de…ned by the formula,
f (r. n) = (0.01 ÷ 0.02n. 0.025 ÷ 0.02r) is a contraction of coe¢cient
0.02
2 < 1. Then apply the Banach’s Fixed Point Theorem. At the
end, compare the approximate result with the exact one!
4. What is the particularity of the system from Problem 3? Can
we apply the Banach’s Fixed Point Theorem to all the linear systems?
CHAPTER 9
Local extremum points
1. Local extremum points for many variables
Let ¹ be an open subset of R
a
and let 1 : ¹ ÷R be a scalar func
tion de…ned on ¹. We say that a = (c
1
. c
2
. .... c
a
) is a local maximum
(minimum) point of 1 if there is a small open ball 1(a. :) · ¹. : 0.
such that 1(x) _1(a) (1(x) _1(a)) for any x in 1(a. :). Local maxima
and local minima are referred to as local extrema. A local maximum
point or a local minimum point is called an extremum point.
Remark 29. Let ¹ be an open subset of R
a
and let i be a …xed
natural number in the set ¦1. 2. .... :¦. Then the ith projection j:
j
(¹)
of ¹ is the set of all t ÷ R such that there is an
x = (r
1
. r
2
. .... r
j÷1
. t. r
j+1
. .... r
a
)
in ¹ with t at the ith position. It is also an open subset of R. Indeed,
take t
0
÷ j:
j
(¹) and take a in ¹ such that a = (c
1
. .... c
j÷1
. t
0
. c
j+1
. .... c
a
).
Since ¹ is open, there is a ball 1(a. :) · ¹ with : 0. We prove that
the 1D ball (t
0
÷:. t
0
+:) is contained in j:
j
(¹). It is in fact the ith
projection of 1(a. :). For this, let n ÷ (t
0
÷:. t
0
+ :). i.e. [n ÷t
0
[ < :.
It is easy to see that
v = (c
1
. c
2
. .... c
j÷1
. n. c
j+1
. .... c
a
) ÷ 1(a. :) · ¹.
Thus
n = j:
j
(v) ÷ j:
j
(¹).
So j:
j
(¹) is also open in R.
Theorem 76. (Fermat’s theorem for many variables) Let ¹ be an
open subset of R
a
and let a ÷¹ be an extremum point of a function
1 : ¹ ÷R, de…ned on ¹ with values in R. If 1 has partial derivatives
0;
0a
j
(a). , = 1. 2. .... : at a. then all of these are zero, i.e. any extremum
point a of 1 is a stationary (critical) point for 1. This means that a
is a root of the vector equation: grad 1(x) = 0, i.e. grad 1(a) = 0, or
d1(a) = 0. if this last one exists.
191
192 9. LOCAL EXTREMUM POINTS
Proof. Let us …x an i in ¦1. 2. .... :¦ and let us de…ne a function
of one variable o
j
: (c
j
÷:. c
j
+:) ÷R by the formula:
o
j
(t) = 1(c
1
. .... c
j÷1
. t. c
j+1
. .... c
a
).
Here : 0 is the radius of a small ball 1(a. :) which is contained in
¹ (see the above discussion). Assume that a is a local maximum point
for 1. We can take : to be small enough such that 1(x) _ 1(a) for any
x in the ball 1(a.:) (why?). If n ÷ (c
j
÷:. c
j
+:). then
v = (c
1
. c
2
. .... c
j÷1
. n. c
j+1
. .... c
a
) ÷ 1(a. :)
so,
o
j
(n) = 1(c
1
. .... c
j÷1
. n. c
j+1
. .... c
a
) _ 1(c
1
. .... c
j÷1
. c
j
. c
j+1
. .... c
a
) = o
j
(c
j
).
This means that c
j
is a local maximum for the function o
j
. We use now
Fermat’s theorem 35 for the one variable function o
j
at the point c
j
.
Thus, o
t
j
(c
j
) = 0. But
o
t
j
(t) =
·1
·r
j
(c
1
. .... c
j÷1
. t. c
j+1
. .... c
a
).
Hence, o
t
j
(c
j
) =
0;
0a
i
(a) = 0. for any i = 1. 2. .... : and the proof of the
theorem is complete.
The Fermat’s theorem says that for the class of di¤erential functions
1 de…ned on an open subset ¹ of R
a
. the local extremum points must
be searched between the critical points, i.e. between the points a which
are zeroes for the gradient of 1. For instance, for 1(r. n) = r
4
+n
4
. the
gradient of 1 is grad 1 = (4r
3
. 4n
3
). So, one has only one point (0. 0)
which makes zero this gradient. Since 0 = 1(0. 0) _ r
4
+ n
4
. for any
r. n ÷ R, the point (0. 0) is a "global" minimum point for 1. It is easy
to see that for the function /(r. n) = r
2
÷n
2
. the point (0. 0) is a critical
point, but it is neither a local minimum, nor a local maximum point for
1. because, in any neighborhood of (0. 0) the function /(r. n) has pos
itive and negative values (why?). So we need a criterion to distinguish
the local extremum points between the critical points. We recall that
a quadratic form in : variables A
1
. A
2
. .... A
a
is a homogeneous poly
nomial function o(A
1
. A
2
. .... A
a
) of degree two of these : independent
variables,
o(A
1
. A
2
. .... A
a
) =
a
¸
j=1
a
¸
;=1
c
j;
A
j
A
;
.
where c
j;
= c
;j
for all i. , ÷ ¦1. 2. .... :¦. i.e. if its associated : :
matrix (c
j;
) is symmetric. Here this last matrix is considered with
entries in R. We say that the quadratic form o is positive de…nite if
1. LOCAL EXTREMUM POINTS FOR MANY VARIABLES 193
o(r
1
. r
2
. .... r
a
) _ 0 for any real numbers r
1
. r
2
. .... r
a
and, it is zero if
and only if all of these numbers are zero. For instance,
o(A. ) ) = A
2
+A) +)
2
is positive de…nite. Assume contrary, namely we could …nd (r. n) =
(0. 0). say n = 0. such that
o(r. n) = r
2
+rn +n
2
< 0.
Let us divide by n
2
and put t = rn. We get t
2
+ t + 1 < 0. which is
false because
t
2
+t + 1 = (t + 12)
2
+ 34
cannot be negative for ever (why?). Moreover, if r
2
+rn +n
2
= 0 and
if (r. n) = (0. 0). then we obtain t
2
+ t + 1 = 0 for t = rn or t = nr.
But the equation 2
2
+2 + 1 = 0 has no real root!
We say that the quadratic form o is negative de…nite if
o(r
1
. r
2
. .... r
a
) _ 0
for any real numbers r
1
. r
2
. .... r
a
and, it is zero if and only if all of
these numbers are zero. For instance,
o(A. ) ) = ÷A
2
÷A) ÷)
2
is negative de…nite (prove it!). If a quadratic form is negative de…nite
or positive de…nite, we say that it is de…nite. If it is neither positive
de…nite, nor negative de…nite, we say that it is nonde…nite. For in
stance, o(A. ) ) = A
2
is a quadratic form which is nonde…nite because,
for r = 0 and any n = 0. it is zero! A basic result in the theory of
quadratic forms (see any serious course in Linear Algebra!) gives us a
criterion which says when a quadratic form is positive de…nite, negative
de…nite, or nonde…nite. The point is to consider the principal minors
1
= c
11
.
2
=
c
11
c
12
c
21
c
22
. ....
a
=
c
11
c
12
. . c
1a
c
21
c
22
. . c
2a
. .
. .
c
a1
c
a2
. . c
aa
.
of the matrix (c
j;
).
Theorem 77. (Sylvester’s criterion) A quadratic form
o(A
1
. A
2
. .... A
a
) =
a
¸
j=1
a
¸
;=1
c
j;
A
j
A
;
is positive de…nite if and only if
1
0.
2
0.
3
0. ....
a
0.
194 9. LOCAL EXTREMUM POINTS
It is negative de…nite if and only if
1
< 0.
2
0.
3
< 0.
4
0. .... (÷1)
a
a
0.
If none of these both conditions are ful…lled, the quadratic form o is
nonde…nite.
For instance,
o(r. n. .) = r
2
+n
2
÷.
2
is nonde…nite because
1
= 1 0.
2
= 1 0 and
3
= ÷1 < 0.
Now, we are ready to prove our above announced criterion for dis
tinguishing the local extremum points between all the critical points.
Theorem 78. (The Decision Theorem) Let 1 : ¹ ÷ R be a func
tion of class (
2
(it has continuous partial derivatives of second order
on ¹) de…ned on an open subset ¹ of R
a
. Let a ÷ ¹ be a critical point
of 1 and let
o(/
1
. /
2
. .... /
a
) = d
2
1(a)(/
1
. /
2
. .... /
a
)
be the second di¤erential of 1 at the point a (it is in fact the quadratic
form
o(/
1
. /
2
. .... /
a
) =
a
¸
j=1
a
¸
;=1
·
2
1
·r
j
·r
;
(a)/
j
/
;
.
i) Assume that d
2
1(a) is not identical to zero. Then, a is a local max
imum point for 1 if and only if the quadratic form d
2
1(a) is negative
de…nite.
ii) Assume that d
2
1(a) is not identical to zero. Then, a is a local
minimum point for 1 if and only if the quadratic form d
2
1(a) is positive
de…nite.
Let / be the …rst natural number such that d
I
1(a) exists, 1 is of
class (
I
on ¹ and d
I
1(a) is not identical to zero. Then,
iii) a is a maximum local point of 1 if and only if / is even and
d
I
1(a)(/
1
. /
2
. .... /
a
) < 0
if at least one of the /
1
. /
2
. .... /
a
is not zero and
iv) a is a minimum local point of 1 if and only if / is even and
d
I
1(a)(/
1
. /
2
. .... /
a
) 0
if at least one of the /
1
. /
2
. .... /
a
is not zero.
Proof. Let us denote by h the variable vector (/
1
. /
2
. .... /
a
) and
let us write Taylor’s formula (3.3) for : = 1. We get:
1(a +h)÷1(a) =
1
2
d
2
1(c)(h).
1. LOCAL EXTREMUM POINTS FOR MANY VARIABLES 195
where c is a point on the segment [a. a +h] and h < :. with : 0
a su¢ciently small real number such that 1(a. :) · ¹. Here d1(a) =
0 because a was considered to be a critical point. Since d
2
1(c) is
continuous as a function of c (d
2
1(c)(h) =
¸
a
j=1
¸
a
;=1
0
2
;
0a
i
0a
j
(c)/
j
/
;
)
and the second order derivatives are continuous by our hypothesis!),
eventually in a smaller ball 1
t
(a. :
t
) with centre at a and of radius
:
t
_ :. one has that the sign of d
2
1(c) is the same like the sign of
d
2
1(a) (why?). Hence, the sign of the di¤erence 1(a +h)÷1(a) is the
same with the sign of d
2
1(a) for h < :
t
. Now, all the statements of
the theorem becomes very clear (why?).
Let us apply this theorem to the following problem. Let
1(r. n) = r
4
+n
4
÷4rn. 1 : R
2
÷R.
Let us …nd all the local extrema for 1. First of all we …nd the critical
points:
0;
0a
= 4r
3
÷4n = 0 and
0;
0&
= 4n
3
÷4r = 0 imply r
9
÷r = 0. So we
…nd the following critical points: `
1
(0. 0). `
2
(1. 1) and `
3
(÷1. ÷1).
In order to apply Theorem 78 we need to compute the Hessian matrix
of 1. i.e. the matrix of the quadratic form d
2
1. at every of the three
critical points.
¹ =
0
2
;
0a
2
0
2
;
0a0&
0
2
;
0a0&
0
2
;
0&
2
=
12r
2
÷4
÷4 12n
2
.
At `
1
the matrix is
0 ÷4
÷4 0
.
Since
1
= 0. from Theorem 78 we obtain that `
1
is not a local
extremum for 1. At `
2
and `
3
the Hessian matrix is
12 ÷4
÷4 12
.
So,
1
= 12 0 and
2
= 144 ÷ 16 = 128 0. Thus, both `
2
and
`
3
are local minimum points.
Example 17. (regression line) In the Cartesian r(n plane we con
sider : distinct points `
1
(r
1
. n
1
). `
2
(r
2
. n
2
). .... `
a
(r
a
. n
a
). We search
for the "closest" line n = cr + / (the regression line) with respect to
this set of points. Here, the "distance" from the set ¦`
j
¦ up to the line
n = cr +/ is the "square" distance distance:
(1.1) o1(c. /) =
a
¸
j=1
[n
j
÷(cr
j
+/)]
2
.
196 9. LOCAL EXTREMUM POINTS
The "closest" line n = cr + / is that one for which the nonnegative
function o1(c. /) is minimum. Thus, we must …nd the local minimum
points for the two variable function o1(c. /). Let us …nd the critical
points by solving the 2 2 system:
(1.2)
0S1
0o
= 2
¸
a
j=1
÷r
j
(n
j
÷cr
j
÷/) = 0
0S1
0b
= 2
¸
a
j=1
÷(n
j
÷cr
j
÷/) = 0
.
Let us write this system in the canonical way
(1.3)
(
¸
r
2
j
) c + (
¸
r
j
) / =
¸
r
j
n
j
(
¸
r
j
) c +:/ =
¸
n
j
.
If not all the points ¦`
j
¦ are on the same line (in this last case
the regression line is obvious the line on which these points are!), the
determinant of this system cannot be zero (use the CauchySchwarz
inequality from Linear Algebra, the equality special case!). So we have
a unique solution (c
0
. /
0
) of this system. Let us prove that this point
realize a minimum for the square distance function o1(c. /). Indeed,
the Hessian matrix of 1 is
2
¸
r
2
j
2
¸
r
j
2
¸
r
j
2:
.
In this case,
1
= 2
¸
r
2
j
0 (otherwise all the points `
j
would be
on the (naxis) and
2
= 4
:
¸
r
2
j
÷(
¸
r
j
)
2
. In order to prove that
2
is greater than zero we consider in R
a
the vectors 1 = (1. 1. .... 1),
x = (r
1
. r
2
. .... r
a
) and write the inequality CauchySchwarz for them:
['1. x`[ _ 1 x or (by squaring) (
¸
r
j
)
2
_ :
¸
r
2
j
. We know that
equality appears if and only if the two vectors are collinear, i.e. if and
only if r
1
= r
2
= ... = r
a
. But this last case appears only if the points
¦`
j
¦ are on a vertical line and we just assumed that ¦`
j
¦ are not
collinear. Hence,
2
0 and the point (c
0
. /
0
) is a local (in fact a
globalwhy?) minimum for the square distance function o1.
The method described above is said to be the least squares method
(1o`). It can be generalized to other classes of curves or surfaces.
Let us apply the 1o` for the set of points `
1
(÷1. 1). `
2
(0. 0).
`
3
(1. 2) and `
4
(2. 3). To solve the system (1.3) we must compute
¸
r
2
j
= 6.
¸
r
j
= 2.
¸
r
j
n
j
= 7 and
¸
n
j
= 6. Then the system
becomes:
6c + 2/ = 7
2c + 4/ = 6
.
We get c = 45 and / = 1110. Hence, the regression line is n =
4
5
r+
11
10
.
2. PROBLEMS 197
2. Problems
1. Find the local extrema for:
a)
1(r. n. .) = r
2
+n
2
+.
2
÷rn +r ÷2.;
b)
1(r. n) = r
3
n
2
(6 ÷r ÷n). r 0. n 0;
c)
1(r. n) = (r ÷2)
2
+ (n + 7)
2
(try directly, without the above algorithm!);
d)
1(r. n) = rn(2 ÷r ÷n);
e)
1(r. n) = ln(1 ÷r
2
÷n
2
);
f)
1(r. n) = r
3
+n
3
÷3rn;
g)
1(r. n) = r
4
+n
4
÷2r
2
+ 4rn ÷2n
2
;
h)
1(r. n. .) = rn.(4c ÷r ÷n ÷.).
c. r. n and . are not zero.
2. Find c. . such that
1(r. n) = 2r
2
+ 2n
2
÷3rn +cr +n +
has a minimum equal to zero in ¹(2. ÷1).
3. A price function is of the form
1(r. n) = r
2
+rn +n
2
÷3cr ÷3/n.
where c. / are constant numbers. Find c and / such that the minimum
of 1 be the biggest possible.
4. Study the local extrema for 1(r. n) = r
4
+n
4
÷r
2
.
CHAPTER 10
Implicitly de…ned functions
1. Local Inversion Theorem
Let a be a point in R
a
. By a (open) neighborhood ¹ of a we mean
any open subset ¹ of R
a
which contains the point a. So, if ¹ is a
neighborhood of a, then there is an open ball 1(a.:). centered at a
and of radius : 0 which is contained in ¹.
Definition 33. Let ¹ and 1 be two open subsets of R
a
. A vector
function f : ¹ ÷1 is said to be a di¤eomorphism between ¹ and 1 if:
i) f is a bijection; ii) f is of class (
1
on ¹ and iii) f
÷1
: 1 ÷¹ is of
class (
1
on 1.
For instance, 1
o
: R ÷R, 1
o
(r) = r+c is a di¤eomorphism because
its inverse o(r) = r÷c is of class (
1
on R. But the mapping 1 : R ÷R,
1(r) = r
5
is not a di¤eomorphism because its inverse o(r) =
5
r is not
di¤erentiable at r = 0 (why?).
Remark 30. It is easy to see that the composition between two
di¤eomorphisms is also a di¤eomorphism (prove it!).
Theorem 79. Let f : ¹ ÷ 1 be a di¤eomorphism and let a be a
point in ¹. Then the linear mapping d1(a) : R
a
÷ R
a
is an isomor
phism of real vector spaces. In particular, the Jacobi matrix J
a.f
of f at
a is invertible and its determinant has a constant sign in a neighborhood
of a. This means that there is an open ball 1(a.:). : 0. contained in
¹. such that det J
x.f
0 (or det J
x.f
< 0) for any x ÷ 1(a.:). In fact,
the sign of det J
x.f
is the same with the sign of det J
a.f
for any x in
1(a.:).
Proof. Let g : 1 ÷¹ be the inverse of f and let b = f (a). Then
g · f = 1
¹
. the identity mapping de…ned on ¹. Now, Theorem 69 says
that J
b.g
J
a.f
= 1
aa
, the :: identity matrix. Hence, the Jacobi ma
trix J
a.f
is invertible, i.e. d1(a) is an isomorphism of real vector spaces
(see the connections between the linear mappings and their correspond
ing matrices, w.r.t. a …xed basis in R
a
. Moreover, det J
a.f
cannot be
zero (why?), say positive, for instance. Since f is a function of class (
1
on ¹. all the partial derivatives which appear as entries in the matrix
199
200 10. IMPLICITLY DEFINED FUNCTIONS
of J
x.f
are continuous. Thus, the mapping x det J
x.f
(denoted here
by 1) is a continuous mapping on ¹. particularly at a. Since 1(a) 0.
we state that there is at least one small positive real number : 0
such that for any x in 1(a. :) we have 1(x) 0. Indeed, otherwise, we
could construct a sequence ¦x
n
¦ of elements in ¹ which is convergent
to a and for which 1(x
n
) _ 0, : = 1. 2. .... The continuity of 1 would
imply that 1(a) _ 0. a contradiction! Hence, there is such a small ball
1(a. :). : 0. on which 1(x) is positive and the proof is complete.
Thus, locally, around a …xed point a. the di¤erential d1(x) is in
vertible. We know that the increment f (x) ÷f (a) of the function f at
a can be well approximated by d1(a)(x ÷a) (see Taylor’s formula for
many variables). A natural question arises: " Is f itself invertible in a
neighborhood of a?" If the function f describes a physical phenomenon,
this means that this phenomenon can be reversible whenever we be
come closer and closer to the point a and, this is very important to be
known in the engineering practice. The following result is fundamental
in all pure and applied mathematics. It is a reverse result relative to
the above theorem
Theorem 80. (Local Inversion Theorem) Let ¹ be an open subset
of R
a
and let f : ¹ ÷ R
a
be a function of class (
1
on ¹. Let a be a
point in ¹ such that det J
a.f
= 0. Then there is a neighborhood l of a,
l · ¹. such that the restriction of f to l. f [
l
: l ÷ \ = f (l). is
a di¤eomorphism. In particular, det J
x.f
= 0 on l and if g : \ ÷ l
is the local inverse of f (g =(f [
l
)
÷1
), then det J
f (x).g
=
1
det J
x;f
and
J
f (x).g
= (J
x.f
)
÷1
.
Proof. (only for : = 1. See a complete proof in Section 7 of this
chapter) Let f = 1 and a = c ÷ ¹ · R be the usual notation in this
restricted case. Now det J
a.f
= 1
t
(c) (why?) and the hypotheses says
that 1
t
(c) is not zero, say that 1
t
(c) 0. Since 1
t
is continuous (1 is of
class (
1
on ¹), like in the proof of the above theorem, we can conclude
that there is an open ball l = 1(c. :) = (c ÷:. c +:). : 0. on which
1
t
is positive, i.e. 1
t
(r) 0 for any r in l. This means that on this l
our function 1 is strictly increasing. So, the restriction of 1 to l has an
inverse o : \ = 1(l) ÷l. Since 1 is continuous and strictly increasing,
one can easily prove that 1
÷1
= o is continuous on \ (prove it! or …nd
by yourself a previous result from which this statement immediately
comes!). We now prove that this function o(n) = r. where n = 1(r).
is di¤erentiable on \. Indeed, let / = 1(c) be a point in \ and let
¦n
a
= 1(r
a
)¦ be a convergent sequence to /. Then ¦r
a
= o(n
a
)¦ tends
1. LOCAL INVERSION THEOREM 201
to c (because of the continuity of o) and
lim
&n÷b
o(n
a
) ÷o(/)
n
a
÷/
= lim
an÷o
r
a
÷c
1(r
a
) ÷1(c)
=
1
1
t
(c)
.
Thus, o is di¤erentiable at / and o
t
(/) =
1
;
0
(o)
.
Example 18. (Polar coordinates) Let `(r. n) be a point in the
Cartesian plane ¦(. i. j¦ and let j =
r
2
+n
2
be the distance from
` up to the origin (. Let o be the unique angle in [0. 2:] such that
r = j cos o and n = j sin o (prove that such an angle exists and that
it is unique!see Fig.10.1). Let us consider ¹ = (0. ·) (0. 2:) · R
2
and 1 = R
2
` ¦[0. ·) ¦0¦¦ in the same R
2
. Let f : ¹ ÷1. f (j. o) =
(j cos o. j sin o). It is easy to see that det J
(a.0).f
= j = 0. It it easy
to prove that this f is a di¤eomorphism. The analytical expression of
its inverse f
÷1
is not so simple (why?…nd it!). The new "coordinates"
(j. o) are called the polar coordinates of `. For instance, the Cartesian
equation of the circle r
2
+ n
2
= 1
2
may be simply written in polar
coordinates like j = 1!
y
O
x
y
x
O
M(x,y)
ρ
Fig. 10.1
Definition 34. (regular transformations) Let ¹ be an open subset
of R
a
and let f : ¹ ÷R
a
be a mapping de…ned on ¹ with values in R
a
.
We say that f is a regular transformation at the point a of ¹ if there
is a neighborhood l of a, l · ¹. such that the restriction of f to l
give rise to a di¤eomorphism f [
l
: l ÷ \ = f (l). If f is regular at
any point of ¹. we say that f is a regular transformation on ¹ or that
f is a local di¤eomorphism on ¹.
In particular, for a local di¤eomorphism f . one has that det J
a.f
= 0
on ¹ and, if in addition ¹ is connected, then det J
a.f
has a constant sign
202 10. IMPLICITLY DEFINED FUNCTIONS
on ¹ (why?). For instance, the polar coordinates transformation (see
Example 18) is a regular transformation (prove it!). The composition
between two regular transformations is again a regular transformation.
Such transformations are "good" for engineers. They are locally su¢
ciently "smooth". This means that they do not produce "breaking" or
"noncontinuous (broken) velocities", or "corners".
Remark 31. The local inversion theorem applied to the regular
transformations gives rise to some basic properties of these last ones.
For instance, a regular transformation f : R
a
÷ R
a
carries an open
subset ¹ of R
a
into the open subset f (¹) (why?). If ¹ is a domain,
i.e. if ¹ is an open and a connected subset of R
a
. then f (¹) is also
a domain of R
a
(why?). Moreover, the Jacobian det J
x.f
has the same
sign on ¹. if ¹ is a domain (try to prove it!).
2. Implicit functions
What is the di¤erence between the curves: 1) (
1
= ¦(r. n) ÷ R
2
:
n =
1 ÷r
2
¦ and 2) (
2
= ¦(r. n) : r
2
+n
2
= 1. n _ 0¦? They represent
the same object, the half of the circle of radius 1. with centre at (.
which is above the (raxis, but... the representations are distinct. In
the …rst case we have an "explicit" representation, i.e. we can write
n = 1(r). this means that we can write one variable as a known function
of the other one. In the second case we have to compute n as a function
of r from the "implicit" relation r
2
+ n
2
= 1. In our case this can be
done, but in other cases such an explicit computation cannot be done.
For instance, it is very di¢cult to express n as a function of r if
(+) r
3
+ 2n
3
÷3rn = 0.
But, if we knew that such an expression n = 1(r) exists (theoretically)
in a neighborhood of a point on the curve, say (1. 1). we can compute
the "velocity" 1
t
(1). the "acceleration" 1
tt
(1). 1
ttt
(1). etc. Practically,
we proceed as follows. Let us write again the implicit relation (+) with
1(r) instead of n :
r
3
+ 21(r)
3
÷3r1(r) = 0
and let us di¤erentiate it with respect to r :
(++) 3r
2
+ 61(r)
2
1
t
(r) ÷31(r) ÷3r1
t
(r) = 0.
We see that always (does not matter the implicit relation is!) the …rst
derivative 1
t
(r) appears to power 1. i.e. it can be "linearly" computed
2. IMPLICIT FUNCTIONS 203
from (++) :
(2.1) 1
t
(r) =
1(r) ÷r
2
21(r)
2
÷r
.
If one put r = 1 in (2.1) one obtains 1
t
(1) = 0. If we di¤erentiate
again formula (2.1) with respect to r. we get
1
tt
(r) =
÷21(r)
2
1
t
(r) ÷4r1(r)
2
÷r1
t
(r) + 4r
2
1(r)1
t
(r) +1(r) +r
2
[21(r)
2
÷r]
2
.
If here we substitute 1
t
(r) with its expression from (2.1), we get the
expression of 1
tt
(r) only as an explicit function of r and of 1(r). Let
us put now r = 1 and we obtain 1
tt
(1). etc.
In our above discussion we supposed that our equation can be
uniquely solved with respect to n. But this is not always true. For
instance, if r
2
+ n
2
= 1. then n(r) = ±
1 ÷r
2
. so that in any neigh
borhood of (1. 0) we cannot …nd a UNIQUE function n = n(r) such
that r
2
+ n(r)
2
= 1. Hence, we cannot compute n
t
(1). n
tt
(1). etc. This
is why we need a mathematical result to precisely say when we have or
not such a unique "implicit" function.
Theorem 81. ( (11) Implicit Function Theorem) Let ¹ be an open
subset of R
2
and let 1 : ¹ ÷ R be a function of two variables which
veri…es the following properties at a …xed point (c. /) of ¹ :
i) 1 is a function of class (
1
on ¹.
ii) 1(c. /) = 0. i.e. (c. /) is a solution of the equation 1(r. n) = 0.
iii)
01
0&
(c. /) = 0.
Then there is a neighborhood l of c. a neighborhood \ of / with
l \ · ¹ and a unique function 1 : l ÷\ such that:
1) 1(r. 1(r)) = 0 for all r in l.
2) 1(c) = /.
3) 1 is of class (
1
on l and
1
t
(r) = ÷
01
0a
(r. 1(r))
01
0&
(r. 1(r))
for all r in l.
Proof. We construct an auxiliary function
=(.
1
. .
2
) : ¹ ÷R
2
. (r. n) = (r. 1(r. n))
for all (r. n) in ¹. Thus, .
1
(r. n) = r and .
2
(r. n) = 1(r. n). We are to
apply the Local Inversion Theorem to this function . Let us compute
204 10. IMPLICITLY DEFINED FUNCTIONS
the Jacobi matrix of at (c. /) :
J
(o.b).
=
1 0
01
0a
(c. /)
01
0&
(c. /)
.
Since (c. /) = (c. 0) and since det J
(o.b).
=
01
0&
(c. /) = 0. Local In
version Theorem 80 says that there is an open neighborhood l \ of
(c. /) and an open neighborhood l \ of (c. 0) (why can we take the
same l?) such that the restriction [
l\
: l \ ÷ l \ of to
l \ is a di¤eomorphism. Let = (·
1
. ·
2
) : l \ ÷ l \ the
inverse of this di¤eomorphism. Let us de…ne 1(r) = ·
2
(r. 0) for any r
in l. It is clear that 1 : l ÷ \ is of class (
1
on l. 1(c) = / and for
any r of l we have
(r. 0) = [(r. 0)] = [·
1
(r. 0). ·
2
(r. 0)]
= [r. 1(r)] = (r. 1(r. 1(r))).
i.e. 1(r. 1(r)) = 0. for any r in l. The function 1 : l ÷ \ is of
class (
1
on l because ·
2
(A. ) ) has continuous partial derivative with
respect to A at any point of the form (r. 0) for any r in l. Let us
di¤erentiate totally with respect to r (this means that r is considered
not only like "the …rst" partial free variable of 1(r. n). but even as an
implicit hidden variable in n = 1(r)) the relation 1(r. 1(r)) = 0 :
0 =
·1
·r
(r. 1(r)) +
·1
·n
(r. 1(r)) 1
t
(r).
thus
1
t
(r) = ÷
01
0a
(r. 1(r))
01
0&
(r. 1(r))
.
for any r in l. Since det J
(a.&).
= 0 on l \ (why?) we get from
J
(a.&).
=
1 0
01
0a
(r. n)
01
0&
(r. n)
that
01
0&
(r. 1(r)) = 0 for any r in l.
If o was another function de…ned on an open neighborhood l
1
of c.
which veri…es the conditions 1), 2) and 3) then, on the neighborhood
l
2
= l ¨ l
1
we would have
·
2
(r. 1(r. o(r)) = o(r)
for any r in l
2
. or ·
2
(r. 0) = o(r) = 1(r) for any r in l
2
. Hence, the
uniqueness reefers to ANOTHER smaller neighborhood of l on which
1 and o are equal. In some conditions, this uniqueness can be extended
to the whole initial l or even to the whole j:
a
(¹). the projection of ¹
on the (raxis.
2. IMPLICIT FUNCTIONS 205
Let us consider again the implicit equation
r
3
+ 2n
3
÷3rn = 0
and let us study it around the solution (1. 1). Since
01
0&
(1. 1) = 3 = 0.
the (11) Implicit Function Theorem says that there is a neighborhood
l of r = 1, a neighborhood \ of n = 1 and a function 1 : l ÷ \.
of class (
1
on l. such that the points ¦(r. 1(r)) : r ÷ l¦ are on the
plane curve r
3
+ 2n
3
÷ 3rn = 0. i.e. r
3
+ 21(r)
3
÷ 3r1(r) = 0 for
any r in l. Now, if we are sure on the existence of such a 1. we can
use di¤erent approximation methods to compute it (approximately!).
The worst situation is when the conditions of the Implicit Function
Theorem fail and we try to compute n = 1(r) approximately! Usually,
in this last case one has more then one function n = 1(r) which ver
ify our equation and during our approximate process we "jump" from
one "branch" to another one, the obtained values for "1(r)" having
a chaotic behavior. For instance, around the point (1. 0). the implicit
solution of the equation r
2
+n
2
= 1 with respect to n has two branches:
n =
1 ÷r
2
and n = ÷
1 ÷r
2
. This is because
01
0&
(1. 0) = 0 and the
Implicit Function Theorem fails around the point (1. 0).
There are two directions for generalizations of this basic theorem.
One reefers to increase the number of variables and the other to consider
vector …elds relations, i.e. a system of implicit equations. We do not
prove these generalizations because these proofs do not contain new
ideas and the "many" variables notation are too sophisticated.
Theorem 82. ((: ÷ 1) Implicit Function Theorem) Let ¹ be an
open subset of R
a+1
. let (a. /) = (c
1
. c
2
. .... c
a
. /) be a point of ¹ and let
1 : ¹ ÷ R, 1(r
1
. r
2
. .... r
a
. n ) be a function of : + 1 variables which
veri…es the following conditions:
i) 1 is of class (
1
on ¹. i.e. it has continuous partial derivatives
with respect to each of its : + 1 variable.
ii) 1(a. /) = 0.
iii)
01
0&
(a. /) = 0.
Then there is a neighborhood l of a, a neighborhood \ of / such
that l \ · ¹ and a unique function 1 : l ÷\ such that:
1) 1[x.1(x)] = 0 for all x in l.
2) 1(a) = /.
3) 1 is of class (
1
on l and
·1
·r
j
(x) = ÷
01
0a
i
(x. 1(x))
01
0&
(x. 1(x))
.
for any x in l.
206 10. IMPLICITLY DEFINED FUNCTIONS
For a proof see [FS]. Let us take the following equation:
2r
3
+n
3
+ 2.
3
÷5rn. = 0
and its solution `(1. 1. 1) (prove this!). Since
01
0:
(1. 1. 1) = 1 = 0.
one can apply the last theorem and can write . = .(r. n) around the
point (1. 1). Let us compute
0
2
:
0a0&
(1. 1). The most practical way is to
put . = .(r. n) into our equation:
2r
3
+n
3
+ 2.(r. n)
3
÷5rn.(r. n) = 0
and let us di¤erentiate this with respect to r and to n :
6r
2
+ 6.(r. n)
2
·.
·r
(r. n) ÷5n.(r. n) ÷5rn
·.
·r
(r. n) = 0.
3n
2
+ 6.(r. n)
2
·.
·n
(r. n) ÷5r.(r. n) ÷5rn
·.
·n
(r. n) = 0.
From these equations we compute
(2.2)
·.
·r
(r. n) =
6r
2
÷5n.
5rn ÷6.
2
.
·.
·n
(r. n) =
3n
2
÷5r.
5rn ÷6.
2
.
Now,
(2.3)
·
2
.
·r·n
=
·
·r
3n
2
÷5r.(r. n)
5rn ÷6.(r. n)
2
=
(÷5. ÷5r
0:
0a
)(5rn ÷6.
2
) ÷(3n
2
÷5r.)(5n ÷12.
0:
0a
)
(5rn ÷6.
2
)
2
.
We need to compute
0:
0a
(1. 1). so we must use formula (2.2) and …nd
0:
0a
(1. 1) = ÷1 (because .(1. 1) = 1). Come back to formula (2.3) and
…nd
0
2
:
0a0&
(1. 1) = 34.
We consider now many relations, i.e. instead of the scalar function
1 we take a vector function F = (1
1
. 1
2
. .... 1
n
) : ¹ ÷R
n
. where ¹ is
an open subset in R
a+n
.
Theorem 83. Let ¹ be an open subset of R
a+n
and let
(a. b) =(c
1
. c
2
. .... c
a
; /
1
. /
2
. .... /
n
)
be a point in ¹. Let F = (1
1
. 1
2
. .... 1
n
) : ¹ ÷R
n
be a function which
veri…es the following conditions:
i) F is a function of class (
1
on ¹.
2. IMPLICIT FUNCTIONS 207
ii) F(a. b) = 0. i.e.
1
1
(c
1
. c
2
. ...c
a
; /
1
. /
2
. .... /
n
) = 0
.
.
.
1
n
(c
1
. c
2
. ...c
a
; /
1
. /
2
. .... /
n
) = 0
.
iii) For F(x. y) = F(r
1
. r
2
. .... r
a
; n
1
. n
2
. .... n
n
). we de…ne the Jaco
bian matrix relative to y = (n
1
. n
2
. .... n
n
) only, as follows:
J
y.F
(x. y) =
¸
¸
¸
¸
¸
01
1
0&
1
(x. y) . . .
01
1
0&m
(x. y)
. . . . .
. . . . .
. . . . .
01m
0&
1
(x. y) . . .
01m
0&m
(x. y)
¸
The condition is that det J
y.F
(a. b) =0. This last determinant can be
suggestively denoted by
det J
y.F
(a. b) =
1(1
1
. 1
2
. .... 1
n
)
1(n
1
. n
2
. .... n
n
)
(a. b).
Then there is a neighborhood l = l
1
l
2
... l
a
of a =
(c
1
. c
2
. .... c
a
). a neighborhood \ = \
1
\
2
...\
n
of b = (/
1
. /
2
. .... /
n
),
such that l \ · ¹ and a unique function f = (1
1
. 1
2
. .... 1
n
).
1
j
: l ÷\
j
. i = 1. 2. .... :. with the following properties:
1) F(x. f (x)) = 0 for any x in l.
2) f (a) = b.
3) f is of class (
1
on l and
(2.4)
·1
j
·r
;
(x) = ÷
1(1
1
.1
2
.....1m)
1(&
1
.&
2
.....&
j1
.a
j
.&
j+1;
....&m)
(x. f (x))
1(1
1
.1
2
.....1m)
1(&
1
.&
2
.....&m)
(x. f (x))
.
It is not necessarily to memorize this last cumbersome formula as
we can see in the following example.
Let (() : r
2
+ n
2
÷ .
2
= 0 be a conic surface and let (1) : r
2
+
2n
2
+3.
2
÷4 = 0 be an ellipsoid. Let = (() ¨(1) be the intersection
curve of them. We see that the point `(1. 0. 1) is on this curve. The
question is if we can …nd a parametrization of the form
:
r = r(n)
n
. = .(n)
.
i.e. if we can use n as a parameter for this curve in a neighborhood
of `. This is equivalent to see if the following system of the implicit
208 10. IMPLICITLY DEFINED FUNCTIONS
functions r = r(n) and . = .(n) can be solved around ` :
(2.5)
1
1
(n; r. .) = r
2
+n
2
÷.
2
= 0.
1
2
(n; r. .) = r
2
+ 2n
2
+ 3.
2
÷4 = 0.
Since all our functions are elementary ones, we need only to check the
condition iii) of the theorem:
1(1
1
. 1
2
)
1(r. .)
(1. 0. 1) =
01
1
0a
(1. 0. 1)
01
1
0:
(1. 0. 1)
01
2
0a
(1. 0. 1)
01
2
0:
(1. 0. 1)
= 16 = 0.
So, r and . can be seen like functions of n in a neighborhood of `.
Let us compute the "velocity" and the "acceleration" at `, along the
curve . For this, it is not necessarily to use the formula (2.4). Namely,
let us put in (2.5) instead of r. r(n) and instead of ., .(n) :
r(n)
2
+n
2
÷.(n)
2
= 0.
r(n)
2
+ 2n
2
+ 3.(n)
2
÷4 = 0.
Let us di¤erentiate both equations with respect to the ONLY free vari
able n :
2r(n)r
t
(n) + 2n ÷2.(n).
t
(n) = 0.
2r(n)r
t
(n) + 4n + 6.(n).
t
(n) = 0.
This is an algebraic linear system in the variables r
t
(n) and .
t
(n). Solv
ing it, we get
(2.6) r
t
(n) = ÷
5n
4r(n)
. .
t
(n) = ÷
n
4.(n)
.
To …nd r
tt
(n) and .
tt
(n) we di¤erentiate again in the formulas (2.6) and
get:
(2.7) r
tt
(n) = ÷
5
4
r(n) ÷nr
t
(n)
r(n)
2
. .
tt
(n) = ÷
1
4
.(n) ÷n.
t
(n)
.(n)
2
Now, it is easy to …nd r
t
(0) = 0. .
t
(0) = 0. r
tt
(0) = ÷
5
4
and .
tt
(0) = ÷
1
4
.
Here is an example when the velocity is zero at a point ` but the
acceleration is not zero at the same point. Thus, one has a nonzero
force at a stationary point!
3. Functional dependence
Let ¹ be an open subset of R
a
and let 1
1
. 1
2
. .... 1
n
be : functions
de…ned on ¹ with real values. We assume that each 1
j
is of class (
1
on ¹.
3. FUNCTIONAL DEPENDENCE 209
Definition 35. We say that ¦1
1
. 1
2
. .... 1
n
¦ are functional depen
dent on ¹ if one of them, say 1
n
is "a function" of the others 1
1
. 1
2
. .... 1
n÷1
.
i.e. there is a function c(n
1
. n
2
. .... n
n÷1
) of : ÷ 1 variables, of class
(
1
on R
n÷1
. such that
1
n
(x) = c[1
1
(x). 1
2
(x). .... 1
n÷1
(x)].
for any x in ¹.
For instance,
(3.1) 1
1
(r
1
. r
2
. r
3
) = r
1
+r
2
+r
3
. 1
2
(r
1
. r
2
. r
3
) = r
1
r
2
+r
1
r
3
+r
2
r
3
.
1
3
(r
1
. r
2
. r
3
) = r
2
1
+r
2
2
+r
2
3
are functional dependent because 1
3
= 1
2
1
÷ 21
2
. Thus, c(n
1
. n
2
) =
n
2
1
÷2n
2
.
We know from Linear Algebra that 1
1
. 1
2
. .... 1
n
are linear depen
dent if there are \
1
. \
2
. .... \
n
scalars, not all zero, such that
(3.2) \
1
1
1
+\
2
1
2
+... +\
n
1
n
= 0.
i.e. \
1
1
1
(x) +\
2
1
2
(x) +... +\
n
1
n
(x) = 0 for any x in ¹. Assume that
\
n
= 0, divide the equality (3.2) by \
n
and compute 1
n
:
1
n
= ÷
\
1
\
n
1
1
÷
\
2
\
n
1
2
÷... ÷
\
n÷1
\
n
1
n÷1
.
Hence, 1
1
. 1
2
. .... 1
n
are also functional dependent. Conversely it is not
true. For instance, the functions 1
1
. 1
2
. 1
3
from (3.1) are functional
dependent but they are not linear dependent (prove it!). This shows
that the notion of functional dependence from Analysis is more general
then the notion of linear dependence from Linear Algebra.
Theorem 84. Let ¹ be an open subset of R
a
and let 1
1
. 1
2
. .... 1
n
:
¹ ÷ R be : function of class (
1
on ¹. If ¦1
1
. 1
2
. .... 1
n
¦ are func
tional dependent on ¹. then the rank of the Jacobian matrix of f =
(1
1
. 1
2
. .... 1
n
) : ¹ ÷R
n
is less than :.
Proof. Suppose that 1
n
(x) = c[1
1
(x). 1
2
(x). .... 1
n÷1
(x)] for all x
in ¹. Then,
·1
n
·r
;
=
·c
·n
1
·1
1
·r
;
+
·c
·n
2
·1
2
·r
;
+... +
·c
·n
n÷1
·1
n÷1
·r
;
for all , = 1. 2. .... :. This means that the :th row of the matrix J
x.f
is
a linear combination of the …rst :÷1 rows, so the rank of the Jacobian
matrix J
x.f
is less than : (why?see any Linear Algebra course).
210 10. IMPLICITLY DEFINED FUNCTIONS
We say that 1
1
. 1
2
. .... 1
n
are dependent at a. a point in ¹. if there
is a neighborhood l of a. l · ¹. such that 1
1
. 1
2
. .... 1
n
are dependent
on l. If 1
1
. 1
2
. .... 1
n
are not dependent at a. we say that they are
independent at a. If 1
1
. 1
2
. .... 1
n
are independent at any point of ¹. we
say that 1
1
. 1
2
. .... 1
n
are independent on ¹.
Theorem 85. If the rank of J
x.f
is equal to : for any x in ¹. then
1
1
. 1
2
. .... 1
n
are independent on ¹.
Proof. Suppose contrary, namely that there is a point a in ¹ and
a small neighborhood l of a. such that 1
1
. 1
2
. .... 1
n
are dependent on
l. Applying Theorem 84 we get that the rank of J
a.f
is less than :. A
contradiction! Thus, 1
1
. 1
2
. .... 1
n
are independent on ¹.
We also have a reverse of the last two theorems.
Theorem 86. With the above notation and hypotheses, if : _ :. if
f = (1
1
. 1
2
. .... 1
n
) is of class (
1
on ¹ and if for a …xed point a of ¹ one
has that the rank of J
a.f
is less than :. then there is a neighborhood l of
a. l · ¹. and : functions from ¦1
1
. 1
2
. .... 1
n
¦. say 1
1
. 1
2
. .... 1
c
. which
are independent on l. such that the other functions ¦1
c+1
. 1
c+2
. .... 1
n
¦
are functional dependent on 1
1
. 1
2
. .... 1
c
on l. This means that there
are :÷: functions c
1
. c
2
. .... c
n÷c
of class (
1
on R
c
such that
1
c+1
(x) = c
1
(1
1
(x). .... 1
c
(x)). .... 1
n
(x) = c
n÷c
(1
1
(x). .... 1
c
(x))
for all x in l.
The proof involves some more sophisticated tools and we send the
interested reader to [Pal] or [FS]. Let us apply this last theorem in a
more complicated example. Let
1
1
= r
1
r
3
+r
2
r
4
1
2
= r
1
r
4
÷r
2
r
3
1
3
= r
2
1
+r
2
2
÷r
2
3
÷r
2
4
1
4
= r
2
1
+r
2
2
+r
2
3
+r
2
4
be four functions of variables r
1
. r
2
. r
3
. r
4
. The Jacobian matrix of
f = (1
1
. 1
2
. 1
3
. 1
4
) at a = (1. 1. 0. 0) is
J
a.f
=
¸
¸
¸
0 0 1 1
0 0 ÷1 1
2 2 0 0
2 2 0 0
¸
.
Since the rank of this matrix is 3 and a nonzero 3 3 determinant
involves the …rst 3 rows, one sees that 1
1
. 1
2
. 1
3
are functional inde
pendent at a and 1
4
is a function of the others in a neighborhood of a.
4. CONDITIONAL EXTREMUM POINTS 211
If we look carefully, we see that 1
2
4
= 4(1
2
1
+ 1
2
2
) + 1
2
3
. so 1
1
. 1
2
. 1
3
. 1
4
are functional dependent on the whole R
4
.
4. Conditional extremum points
Sometimes we have to …nd the extremum points for a function 1
de…ned on a compact subset ( of R
a
. For instance, let ( be the closed
ball
1[0. 3] = ¦(r. n. .) : r
2
+n
2
+.
2
_ 9¦.
centered at 0 = (0. 0. 0) and of radius 3. The problem of …nding the
extremum points of the function 1(r. n. .) = r + 2n + 3. de…ned on (
can be divided into two parts. First of all we …nd the local extrema
points of 1 de…ned only on the open set
1(0. 3) = ¦(r. n. .) : r
2
+n
2
+.
2
< 9¦
by using Fermat’s theorem, then we consider only the points on the
sphere r
2
+n
2
+.
2
= 9 and try to …nd the extremum points `(r. n. .)
of 1, which verify this last supplementary condition (a constraint). This
last problem is an example of a conditional extremum points problem.
The general method for solving such problems is the "method of
Lagrange’s multipliers". In the following we shall describe this method.
Let ¹ be an open subset of R
a
and let 1. o
1
. o
2
. .... o
n
(: < :) be
functions of class (
1
on ¹. We assume that o
1
. o
2
. .... o
n
are functional
independent on ¹. particularly, if g = (o
1
. o
2
. .... o
n
). its Jacobian
matrix J
x.g
has the rank : at any point x of ¹. Let o · ¹ be the set
of all solutions (in ¹) of the following system of equations:
(4.1)
o
1
(r
1
. r
2
. .... r
a
) = 0
.
.
.
o
n
(r
1
. r
2
. .... r
a
) = 0
.
These equations are called constraints or supplementary conditions for
the variables r
1
. r
2
. .... r
a
.
Definition 36. We say that a point a = (c
1
. c
2
. .... c
a
) of o is
a local conditional maximum point with the constraints (4.1) if there
is a neighborhood l of a. l · ¹. such that 1(x) _ 1(a) for any x
in l ¨ o. The notion of a local conditional minimum point with the
same constraints, for the same function 1, can be de…ned in the same
manner.
212 10. IMPLICITLY DEFINED FUNCTIONS
For instance, (0. 0) is a local conditional minimum for 1(r. n) =
r
2
+n de…ned on R with the constraint n ÷r
2
= 0. Indeed, 1(r. r
2
) =
2r
2
_ 0 = 1(0. 0) for any r ÷ R.
Let = (\
1
. \
2
. .... \
n
) be a variable vector in R
n
. These new
auxiliary variables \
1
. \
2
. .... \
n
are called Lagrange’s multipliers and
the new auxiliary function
(4.2) (r
1
. r
2
. .... r
a
; \
1
. \
2
. .... \
n
) = (x. ) = 1(x) +
n
¸
;=1
\
;
o
;
(x)
is called Lagrange’s associated function.
Theorem 87. (Lagrange’s Theorem) Let us preserve all the above
notation and hypotheses. Assume that a is a local conditional extremum
point for 1. with the constraints (4.1). Then there is a vector
+
=
(\
+
1
. \
+
2
. .... \
+
n
) in R
n
such that the point
(a.
+
) = (c
1
. c
2
. .... c
a
; \
+
1
. \
+
2
. .... \
+
n
)
is a critical (stationary) point for Lagrange’s function . i.e.
grad(a.
+
) = 0.
Proof. (for : = 2 and : = 1 Suppose that a is a local conditional
maximum point for 1. Since o = o
1
is functional independent, it cannot
be a constant function, say
0j
0a
2
(a) = 0. We can apply the Implicit
Function Theorem and …nd a function / : l
1
÷ l
2
of class (
1
on l
1
.
an appropriate neighborhood of c
1
(l
2
is a neighborhood of c
2
), such
that /(c
1
) = c
2
. o(r
1
. /(r
1
)) = 0 for all r
1
in l
1
and
(4.3) /
t
(r
1
) = ÷
0j
0a
1
(r
1
. /(r
1
))
0j
0a
2
(r
1
. /(r
1
))
for all r
1
in l
1
. We can assume that the neighborhood of a. l = l
1
l
2
is su¢ciently small such that 1(x) _ 1(a) for any x in l. We de…ne
now a new function 1 : l
1
÷ R, 1(r
1
) = 1(r
1
. /(r
1
)) for any r
1
in
l
1
. Since 1(r
1
) _ 1(c
1
). for all r
1
in l
1
, we see that c
1
is a local
maximum point for the function 1. Use now Fermat’s Theorem and
…nd that 1
t
(c
1
) = 0. or that
·1
·r
1
(a) +
·1
·r
2
(a) /
t
(c
1
) = 0.
Thus,
(4.4) /
t
(c
1
) = ÷
0;
0a
1
(a)
0;
0a
2
(a)
.
4. CONDITIONAL EXTREMUM POINTS 213
But the same /
t
(c
1
) can also be computed from the formula (4.3)
/
t
(c
1
) = ÷
0j
0a
1
(c
1
. c
2
)
0j
0a
2
(c
1
. c
2
)
.
If we equals the both expression of /
t
(c
1
) we get
·1
·r
1
(a)
·o
·r
2
(a) ÷
·1
·r
2
(a)
·o
·r
1
(a) = 0.
Let us put
(4.5) \
+
oc;
= ÷
0;
0a
1
(a)
0j
0a
1
(a)
= ÷
0;
0a
2
(a)
0j
0a
2
(a)
and let us write the Lagrange’s auxiliary function for this "multiplier"
\
+
:
(x.\
+
) = 1(x) +\
+
o(x).`
Let us compute the grad(a.\
+
) by taking count of the value of \
+
from
(4.5):
0
0a
1
(a.\
+
) =
0;
0a
1
(a) +\
+ 0j
0a
1
(a) = 0
0
0a
2
(a.\
+
) =
0;
0a
2
(a) +\
+ 0j
0a
2
(a) = 0
0
0A
1
(a.\
+
) = o(a) = 0. because a ÷ o.
Hence grad(a.\
+
) = 0 and the proof is complete.
Look now at the function
(x.
+
) = 1(x) +
n
¸
;=1
\
+
;
o
;
(x).
where
+
= (\
+
1
. \
+
2
. .... \
+
n
) is the vector just constructed in Theorem
87. It is easy to see that a is a local conditional maximum(for instance!)
for 1 if and only if a is an usual local maximum for the function 1(x) =
(x.
+
). Thus, if we want do decide if a stationary point (a.
+
) of the
Lagrange function is a conditional extremum point, we must consider
the second di¤erential of 1 at a. But, in the expression of d
2
1(a) we
must take count of the connections between dr
1
. dr
2
. .... dr
a
. These
connections can be found by di¤erentiating the equations 4.1:
0j
1
0a
1
(a)dr
1
+... +
0j
1
0an
(a)dr
a
= 0
.
.
.
0jm
0a
1
(a)dr
1
+... +
0jm
0an
(a)dr
a
= 0
.
214 10. IMPLICITLY DEFINED FUNCTIONS
Since the rank of the Jacobi matrix J
a.g
is : < :. this linear system
in the unknown quantities dr
1
. dr
2
. .... dr
a
has an in…nite number of
solutions. Namely, say that the last : ÷ : unknowns dr
n+1
. .... dr
a
remain free and the others dr
1
. dr
2
. .... dr
n
can be linearly expressed
as functions of the last :÷:. Thus, the di¤erential d
2
(a.
+
) becomes
a quadratic form in :÷: free variables. The sign of this last one must
be considered in any discussion about the nature of the point a.
Let us …nd the points of the compact r
2
+ n
2
_ 1 in which the
function 1(r. n) = (r÷1)
2
+(n÷2)
2
has the maximumand the minimum
values. Let us …nd …rstly the local extrema inside the disc: r
2
+n
2
< 1.
·1
·r
= 2(r ÷1) = 0.
·1
·n
= 2(n ÷2) = 0.
So the critical point is `(1. 2). The Hessian matrix at (1. 2) is
2 0
0 2
.
thus `(1. 2) is a local minimum point inside the disc. The value of 1
at (1. 2) is 0. Let us consider now the local conditional problem:
max(min)1
with the restriction
o(r. n) = r
2
+n
2
÷1 = 0
The auxiliary Lagrange’s function is
(r. n. \) = 1(r. n) +\(r
2
+n
2
÷1).
Let us …nd its critical points:
0
0a
= 2(r ÷1) + 2\r = 0
0
0&
= 2(n ÷1) + 2\n = 0
0
0A
= r
2
+n
2
÷1 = 0
.
Solve this system and …nd r =
1
A+1
and n =
1
A+1
(why \ cannot be
÷1?). \
1
=
2 ÷ 1. r
1
=
1
2
, n
1
=
1
2
and \
2
= ÷
2 ÷ 1. r
2
= ÷
1
2
, n
1
= ÷
1
2
. Let us denote `
1
(
1
2
.
1
2
) and `
2
(÷
1
2
. ÷
1
2
). In order to
see the nature of these critical points, let us …nd the expression of the
second di¤erential of (r. n. \) for a constant parameter \. We …nd
d
2
(r. n. \) = (2 + 2\)dr
2
+ (2 + 2\)dn
2
.
Since rdr +ndn = 0. then dn = ÷
a
&
dr. so,
d
2
(r. n. \) = (2 + 2\)(1 +
r
2
n
2
)dr
2
.
5. CHANGE OF VARIABLES 215
For \
1
=
2 ÷1. we get that `
1
is a local conditional minimum. For
\
2
= ÷
2 ÷ 1. we obtain that `
2
is a local conditional maximum.
Hence, the global maximum of 1 on the compact subset ¦(r. n) : r
2
+
n
2
_ 1¦ is 1
÷
1
2
. ÷
1
2
= 6 +3
2. Its global minimum is min¦0. 6 ÷
3
2¦ = 0.
Let us consider now a practical problem of conditional extremum.
Let us …nd the distance between the line r ÷ n = 5 and the parabola
n = r
2
. Let 1(r
1
. n
1
) be a running point on the line and let 1(r
2
. n
2
)
be a running point on the parabola. The square 1(r
1
. r
2
. n
1
. n
2
) =
(r
1
÷ r
2
)
2
+ (n
1
÷ n
2
)
2
of the distance between two such points must
be minimum and the constraints are
o
1
(r
1
. r
2
. n
1
. n
2
) = r
1
÷n
1
÷5 = 0
and
o
2
(r
1
. r
2
. n
1
. n
2
) = r
2
2
÷n
2
= 0.
The Lagrange’s function is
(r
1
. r
2
. n
1
. n
2
; \
1
. \
2
) = (r
1
÷r
2
)
2
+ (n
1
÷n
2
)
2
+
+\
1
(r
1
÷n
1
÷5) +\
2
(r
2
2
÷n
2
).
If we solve the 4 4 algebraic system grad = 0. we get r
1
=
23
8
.
n
1
= ÷
17
8
. r
2
=
1
2
. n
2
=
1
4
and the corresponding distance is
19
4
2
.
5. Change of variables
What is the plane curve rn = 2? We know that an equation of the
form
a
2
o
2
÷
&
2
b
2
= 1 is a hyperbola. If we introduce two new variables A
and ) such that r =
1
2
A÷
1
2
) and n =
1
2
A+
1
2
). we introduce in
fact a new cartesian coordinate system A() which is obtained from
r(n by a rotation of 45
·
in the direct sense (see Fig.10.2).
216 10. IMPLICITLY DEFINED FUNCTIONS
2
2
45
o
O
X
Y
x
y
Fig. 10.2
Our initial curve rn = 2 becomes A
2
÷ )
2
= 4. i.e. we have an
usual hyperbola with c = / = 2 relative to the new cartesian coordinate
system A().
The moral is that sometimes is better to change the old cartesian
coordinate system i.e. to change the old variables r
1
. r
2
. .... r
a
with
another new ones n
1
. n
2
. .... n
a
which are functions of the …rst ones:
(5.1)
n
1
= n
1
(r
1
. r
2
. .... r
a
)
.
.
.
n
a
= n
a
(r
1
. r
2
. .... r
a
)
.
Here we forced the notation. The function of : variables which de…nes
the new variable n
1
is also denoted by n
1
. etc.
Definition 37. Let 1. be two open subsets of R
a
and let f : 1 ÷
be a di¤eomorphism of class (
I
on 1. i.e. f is a bijection, it is of
class (
I
on 1 and its inverse f
÷1
is also of class (
I
on . Usually,
/ = 1 or 2. We call such a f a change of variables of class (
I
.
If we write
f (r
1
. r
2
. .... r
a
) = (n
1
(r
1
. r
2
. .... r
a
). .... n
a
(r
1
. r
2
. .... r
a
)).
we have a representation like (5.1) for the vector function f . We also call
such a representation a change of variables. We represent the inverse
5. CHANGE OF VARIABLES 217
of f by:
(5.2)
r
1
= r
1
(n
1
. n
2
. .... n
a
)
.
.
.
r
a
= r
a
(n
1
. n
2
. .... n
a
)
.
In fact, we solved the system (5.1) and we computed r
1
. r
2
. .... r
a
as
functions of n
1
. n
2
. .... n
a
. For instance, if n
1
= r
1
+r
2
and n
2
= 2r
1
÷r
2
.
then r
1
=
1
3
(n
1
+n
2
) and r
2
=
1
3
(2n
1
÷n
2
).
If one considers an expression like
1(r
1
. r
2
. .... r
a
. o(r
1
. r
2
. .... r
a
).
·o
·r
;
.
·
2
o
·r
;
·r
j
. ...).
the problem is to …nd an appropriate change of variables of the form
(5.2) such that the new expression in the new variables n
1
. n
2
. .... n
a
has
a simpler form. Thus, the "old" function o(r
1
. r
2
. .... r
a
) becomes a
"new" function o(n
1
. n
2
. .... n
a
). The relations between these two func
tions are
(5.3) o(n
1
. n
2
. .... n
a
) = o(r
1
(n
1
. n
2
. .... n
a
). .... r
a
(n
1
. n
2
. .... n
a
))
and
(5.4) o(r
1
. r
2
. .... r
a
) = o(n
1
(r
1
. r
2
. .... r
a
). .... n
a
(r
1
. r
2
. .... r
a
)).
Now, the problem is to express the partial derivatives
·o
·r
;
(r
1
. r
2
. .... r
a
).
·
2
o
·r
;
·r
j
(r
1
. r
2
. .... r
a
). ...
only in language of the partial derivatives of the newfunction o(n
1
. n
2
. .... n
a
).
This is an easy job if we know to manipulate the chain rules. For in
stance, if x = (r
1
. r
2
. .... r
a
) and y = (n
1
. n
2
. .... n
a
). from (5.4) one
has:
·o
·r
j
(x) =
·o
·n
1
(y)
·n
1
·r
j
(x) +... +
·o
·n
a
(y)
·n
a
·r
j
(x).
i = 1. 2. .... :. To have "everything" in n
1
. n
2
. .... n
a
we …nally put in
stead of r
1
. r
1
(n
1
. n
2
. .... n
a
). .... instead of r
a
. r
a
(n
1
. n
2
. .... n
a
).
For instance, let us make the substitution (change of variables)
r = exp(t) in the following Euler’s equation:
r
2
d
2
n
dr
2
+r
dn
dr
= 0. r 0.
First of all recall the di¤erential notation: n = n(r). n
t
(r) =
o&
oa
(since
dn = n
t
(r)dr) and n
tt
(r) =
o
2
&
oa
2
(since d
2
n = n
tt
(r)dr
2
see the formula
218 10. IMPLICITLY DEFINED FUNCTIONS
for the second di¤erential!). Let us denote by n(t) = n(exp(t)). Since
n(r) = n(ln r). one has that
dn
dr
=
dn
dt
dt
dr
=
dn
dt
1
r
. i.c.
d
dr
=
d
dt
exp(÷t).
Let us compute
d
2
n
dr
2
=
d
dr
dn
dr
=
d
dr
dn
dt
exp(÷t)
=
d
dt
dn
dt
exp(÷t)
exp(÷t).
Applying the rule of the di¤erential of a product, we get:
d
2
dr
2
=
d
2
dt
2
÷
d
dt
exp(÷2t).
Substituting in the initial equation, we get
o
2
&
oa
2
= 0. i.e. n = (
1
t + (
2
.
where (
1
. (
2
are arbitrary constants. Thus, n(r) = (
1
ln r + (
2
and
we just found the general solution of the initial di¤erential equation.
6. The Laplacian in polar coordinates
The polar coordinates j. o were introduced in Example 18. The
"linear operator" . the Laplacian, carries functions n(r. n) of class
(
2
. de…ned on a …xed domain 1 · R
2
into continuous functions:
n =
·
2
n
·r
2
+
·
2
n
·n
2
. i.c. =
·
2
·r
2
+
·
2
·n
2
.
For instance, in order to solve the famous Laplace equation, n = 0,
which appears in many applications, we sometimes need to write the
operator in polar coordinates j and o. We know that
r = j cos o
n = j sin o
.
where j ÷ (0. ·) and o ÷ [0. 2:). The Jacobian of this transformation
is det J
(a.0).g
= j = 0. where g(j. o) = (j cos o. j sin o). Let us denote
by n(j. o) = n(j cos o. j sin o). the new function in the new variables j
and o. Let us denote by j = j(r. n) and by o = o(r. n) the coordinates
of the inverse function g
÷1
. Thus,
n(r. n) = n(j(r. n). o(r. n)).
Hence,
(6.1)
0&
0a
=
0&
0a
0a
0a
+
0&
00
00
0a
0&
0&
=
0&
0a
0a
0&
+
0&
00
00
0&
7. A PROOF FOR THE LOCAL INVERSION THEOREM 219
These last relations can be represented in a matrix form
(6.2)
0&
0a
0&
0&
=
0a
0a
00
0a
0a
0&
00
0&
0&
0a
0&
00
.
Since g · g
÷1
= the identity mapping, we have that
0a
0a
00
0a
0a
0&
00
0&
tvoac
=
J
(a.0).g
÷1
=
cos o ÷j sin o
sin o j cos o
÷1
=
cos o sin o
÷
sin 0
a
cos 0
a
.
Let us come back to formula 6.2 and …nd:
0&
0a
0&
0&
=
cos o ÷
sin 0
a
sin o
cos 0
a
0&
0a
0&
00
.
Let us write this formula in a nonmatriceal form:
(6.3)
0&
0a
=
0&
0a
cos o ÷
0&
00
sin 0
a
0&
0&
=
0&
0a
sin o +
0&
00
cos 0
a
.
Let us use now these formulas and the chain rules formulas 2.7, 2.8 to
compute n =
0
2
&
0a
2
+
0
2
&
0&
2
:
·
2
n
·r
2
=
·
2
n
·j
2
cos
2
o÷2
·
2
n
·j·o
sin o cos o
j
+
·
2
n
·o
2
sin
2
o
j
2
+
·n
·j
sin
2
o
j
+2
·n
·o
sin o cos o
j
2
.
·
2
n
·n
2
=
·
2
n
·j
2
sin
2
o+2
·
2
n
·j·o
sin o cos o
j
+
·
2
n
·o
2
cos
2
o
j
2
+
·n
·j
cos
2
o
j
÷2
·n
·o
sin o cos o
j
2
.
Hence, the formula for the Laplacian in polar coordinates is:
n =
·
2
n
·j
2
+
1
j
2
·
2
n
·o
2
+
1
j
·n
·j
.
This formula will be used later in the course of partial di¤erential equa
tions with direct applications in Engineering.
7. A proof for the Local Inversion Theorem
Here we present a complete proof for the Local Inversion Theorem
(see Theorem 80). We prefer an elementary longer proof then a shorter
sophisticated one. Let us state again this basic result.
Theorem 88. Let ¹ be an open subset of R
a
and let f : ¹ ÷ R
a
be a function of class (
1
on ¹. Let a be a point in ¹ such that the
Jacobian determinant det J
a.f
= 0. Then there are two open sets A · ¹
and ) · f (¹) and a uniquely determined function g with the following
properties:
i) a ÷ ¹ and f (a) ÷ ).
ii) ) = f (A).
iii) g : ) ÷A. g() ) = A and g(f (x)) = x for any x in A.
220 10. IMPLICITLY DEFINED FUNCTIONS
iv) g is of class (
1
on ) and the restriction of 1 to A. 1 [
A
: A ÷)
is a di¤eomorphism with g = (f [
A
)
÷1
. Particularly,
J
f (x).g
= (J
x.f
)
÷1
and
det J
f (x).g
=
1
det J
x.f
.
Proof. STEP 1. First of all let us remark that if (/
j;
(x)). i. , =
1. 2. .... : are :
2
continuous functions de…ned on ¹. such that det[/
j;
(a)] =
0. then there is a small closed ball 1[a. :] with centre at a and of radius
: 0. 1[a. :] · ¹ with the property that whenever we take :
2
points
¦x
j;
¦ in 1[a. :], one has that det[/
j;
(x
j;
)] = 0. Indeed, let us de…ne a
continuous function of :
2
variables on the product ¹ ¹ ... ¹
. .. .
a
2
÷times
:
1(X
11
. X
12
. .... X
1a
. .... X
a1
. X
a2
. .... X
aa
) = det[/
j;
(X
j;
)].
Since 1(a. a. .... a) = det(/
j;
(a)) is not zero, say 1(a. a. .... a) 0. one
can …nd a small ball 1(a. :
t
) · ¹. :
t
0. on which
1(x
11
. x
12
. .... x
aa
) = det(/
j;
(x
j;
)) 0
for every x
j;
in 1(a. :
t
) (see Theorem57). If one takes any :, 0 < : < :
t
.
then det(/
j;
(x
j;
)) 0 for any arbitrary :
2
elements ¦x
j;
¦ in 1[a. :]. In
our case, det J
a.f
=
det
0;
i
0a
j
(a)
= 0. where f = (1
1
. 1
2
. .... 1
a
). Hence,
we can …nd a small closed ball \ = 1[a. :] · ¹. : 0. on which
det
0;
i
0a
j
(x
j;
)
= 0 for any :
2
elements x
j;
in \.
STEP 2. Let us prove now that the restriction of f to \ is oneto
one. Suppose that x and z are in \ such that f (x) = f (z). This means
that for every i = 1. 2. .... : one has that 1
j
(x) = 1
j
(z). Let us apply
the Lagrange theorem (see Theorem 73) on the segment [x. z] :
(7.1) 0 = 1
j
(x) ÷1
j
(z) =
a
¸
;=1
·1
j
·r
;
(c
(j)
) (r
;
÷.
;
).
where c
(j)
is a point on the segment [x. z] and x =(r
1
. r
2
. .... r
a
). z =
(.
1
. .
2
. .... .
a
). Since the segment [x. z] is contained in \ (why?), all
c
(j)
. i = 1. 2. .... :. are contained in \ and so, det
0;
i
0a
j
(c
(j)
)
= 0.
Hence, the homogeneous linear system
0 =
a
¸
;=1
·1
j
·r
;
(c
(j)
) (r
;
÷.
;
).
7. A PROOF FOR THE LOCAL INVERSION THEOREM 221
i = 1. 2. .... :. in the unknowns r
1
÷.
1
. r
2
÷.
2
. .... r
a
÷.
a
. has only the
trivial solution, i.e. r
1
= .
1
. .... r
a
= .
a
or x = z. Thus, f is onetoone
on \ = 1[a. :].
STEP 3. Let us prove now that the image f (2) of 2 = 1(a. :).
the interior of \. is an open subset of R
a
. Indeed, let us de…ne the
continuous function o : ·2 ÷R (here ·2 = \ r2 is the boundary of
2):
o(x) = f (x) ÷f (a) .
for x ÷ ·2. Since ·2 is a compact subset of R
a
(prove it!) and since
f is onetoone (see STEP 2), the minimum value : of o on ·2 is 0
(why?). Let us denote by 1 = 1(f (a).
n
2
) and let us prove that this
open ball 1 is contained in f (2). For this, let y be a …xed element in
1 and let us de…ne the following continuous function:
/(x) = f (x) ÷y
for any x in \. Let us see that the absolute minimum of / cannot be
attained on the boundary ·2. Indeed, since
/(a) = f (a) ÷y <
:
2
.
one has that min /(x) <
n
2
. But, if x ÷ ·Z. we have
/(x) = f (x) ÷y _ f (x) ÷f (a) ÷f (a) ÷y
o(x) ÷
:
2
_
:
2
.
i.e. /(x)
n
2
for any x in ·2. Hence, let c be in 2 such that
/(c) = min¦/(x) : x ÷ \¦.
This c also realizes the absolute minimum for
/
2
(x) = f (x) ÷y
2
=
a
¸
v=1
[1
v
(x) ÷n
v
]
2
.
Then Fermat’s theorem says that:
·
·r
I
a
¸
v=1
[1
v
(x) ÷n
v
]
2
¸
= 2
a
¸
v=1
[1
v
(x) ÷n
v
]
·1
v
·r
I
(x)
is zero at c. i.e.
a
¸
v=1
·1
v
·r
I
(c) [1
v
(c) ÷n
v
] = 0
for every / = 1. 2. .... :. This is again a homogenous linear system in
the unknowns ¦1
v
(c) ÷ n
v
¦
v
with a nonzero determinant. Hence, we
have only the trivial solution, i.e. 1
v
(c) = n
v
for every : = 1. 2. .... :.
Thus, f (c) = y and so y ÷ f (2). But, the same type of reasoning can
222 10. IMPLICITLY DEFINED FUNCTIONS
be done for any other b = f (e). where e ÷ 2 and b ÷ f (2). Namely,
we take a su¢ciently small open ball 1(e. :
tt
) · 1(a. :) and we repeat
the above reasoning for 1(e. :
tt
) instead of 1(a. :). We …nd that
1
t
= 1(b.
:
t
2
) · f (1(e. :
tt
)) · f (2)
for the minimum :
t
of the function
x ÷f (x) ÷f (e) .
de…ned on ·1(e. :
tt
). Hence, f (2) is open in R
a
. Moreover, f carries an
open subset A of 2 into an open subset f (A) of R
a
(why?).
STEP 4. Let now ) = 1(f (a). :
t
) be an open ball centered at
f (a) such that its closure 1[f (a). :
t
] is included in f (2) and let A =
f
÷1
() ) ¨2. It is clear that the restriction f [
A
: A ÷) is a continuous
bijection between A and ). Let g : ) ÷ A. g(y) = x be its inverse.
Let A and ) be the topological closure of A and ) respectively. They
both are compact subsets of R
a
and f [
A
: A ÷ ) is also a bijection,
because A · \ and f is onetoone on \ (see STEP 1). Its inverse
(f [
A
)
÷1
: ) ÷ A is continuous (because f is continuous and A and
) are compact sets...it reverses closed subsets into closed subsets!).
Since the restriction of (f [
A
)
÷1
to ) is exactly g (why?), g is also a
continuous mapping and g(f (x)) = x for any x in A.
STEP 5. It remains us to prove that g = (o
1
. o
2
. .... o
a
) is of class
(
1
on ). We …x an : = 1. 2. .... : and we shall prove that
0j
j
0&r
exists
at any …xed point y in ) and that they are continuous. Let e
v
=
(0. 0. .... 0. 1. 0. .... 0) be the :th unit vector in R
a
(with 1 at the :th
position!) and let us consider the di¤erence quotient:
(7.2)
o
;
(y +te
v
) ÷o
;
(y)
t
.
where t is a small real number such that y +te
v
÷ ) () is open). Let
x = g(y) and x
t
= g(y+te
v
). Thus,
f (x
t
) ÷f (x) = te
v
implies that
(7.3) 1
j
(x
t
) ÷1
j
(x) =
0. if i = :.
t. if i = :.
Let us apply Lagrange’s theorem(see Theorem73) for 1
j
on the segment
[x. x
t
] · 2. We get:
(7.4) 0 or 1 =
1
j
(x
t
) ÷1
j
(x)
t
=
a
¸
;=1
·1
j
·r
;
(d
(j)
)
r
t
;
÷r
;
t
.
8. THE DERIVATIVE OF A FUNCTION OF A COMPLEX VARIABLE 223
i = 1. 2. .... :. where d
(j)
is a point on the segment [x. x
t
] · 2. Since
det
0;
i
0a
j
(d
(j)
)
= 0. the linear system (7.4), in variables
a
0
j
÷a
j
t
¸
;
has
a unique solution (Cramer’s rule):
r
t
;
÷r
;
t
=
;
.
, = 1. 2. .... :. where and
;
are determinants with entries of the
form
0;
i
0a
j
(d
(j)
). 0. or 1. When t ÷ 0. the determinant ÷ J
x.f
= 0
(why?), so
1
.
2
. ....
a
÷
·o
1
·n
v
(y).
·o
2
·n
v
(y). ....
·o
a
·n
v
(y)
.
i.e all the partial derivatives
0j
j
0&r
(y) exist. Since their expressions in
volve only partial derivatives of the type
0;
i
0a
j
(x) which are continuous,
the function g is of class (
1
on ) and the proof of the Local Inversion
Theorem is now complete.
The proof is long, but elementary and very natural. Trying to
understand this proof one remembers many basic things from previous
chapters. Moreover, the proof itself re‡ects some of the indescribable
Beauty of Mathematical Analysis.
8. The derivative of a function of a complex variable
Let ¹ be an open subset of the complex plane C. If we associate
to any complex number . = r + in of ¹. where r. n are real numbers
and i =
÷1 is a …xed root of the equation r
2
+ 1 = 0. another
complex number n = 1(.). we say that the mapping . ÷ 1(.) is
a function of a complex variable de…ned on ¹. Like in the case of a
function of a real variable, we say that 1 has the limit 1 at the point
.
0
= r
0
+ in
0
of ¹ if for any sequence ¦.
a
¦. : = 1. 2. .... of complex
numbers .
a
= r
a
+ in
a
. r
a
. n
a
÷ R, which tends to c. one has that
1(.
a
) ÷ 1. If 1 = 1(.
0
) we say that 1 is continuous at .
0
. Let us
assume that 1(r + in) = n(r. n) + i·(r. n). where n and · are two
real functions of two variables. One calls n = Re 1. the real part of 1
and · = Im1. the imaginary part of 1. It is not di¢cult to see that
1 is continuous at .
0
= r
0
+ in
0
if and only if n and · are continuous
at (r
0
. n
0
). Let us de…ne the derivative of a function 1 of a complex
variable . at a …xed point .
0
. We say that 1 is di¤erentiable at .
0
if
224 10. IMPLICITLY DEFINED FUNCTIONS
the following limit exists and is …nite:
(8.1) lim
:÷:
0
1(.) ÷1(.
0
)
. ÷.
0
= 1
t
(.
0
).
We denoted its value by 1
t
(.
0
) and we call it the derivative of 1 at .
0
.
For instance, (.
2
)
t
= 2.. because
lim
:÷:
0
.
2
÷.
2
0
. ÷.
0
= lim
:÷:
0
(. +.
0
) = 2.
0
.
Generally speaking, the usual di¤erential rules of the functions of a real
variable also works for functions of a complex variable. For instance,
(1 + o)
t
= 1
t
+ o
t
. (c1)
t
= c1
t
. (1o)
t
= 1
t
o + 1o
t
.
;
j
t
=
;
0
j÷;j
0
j
2
.
(1 · o)
t
(.) = 1
t
(o(.)) o
t
(.). (sin .)
t
= cos .. (exp(.))
t
= exp(.). etc.
Many formulas in complex function theory (the theory of functions
of a complex variable) can be easily proved by using the following
fundamental result.
Theorem 89. (Identity Theorem) Let ¹ be a subset of complex
numbers with at least one limit point and let 1 and o be two di¤er
entiable complex functions de…ned on a complex domain 1 (it is open
and connected) which contains ¹. Assume that 1 and o are equal at any
point of ¹. Then 1 and o are identical, this means that 1(.) = o(.) for
all . of 1.
For a proof of this basic result see any book of complex function
theory (see for instance [ST]). Let us use this result to compute the
derivative of exp(.) =
¸
o
a=0
:
n
a!
. . ÷ C. Let us denote by o(.) the
derivative of exp(.). Since for any real number r one has that exp(r)
t
=
exp(r). we have that o(r) = exp(r) for any r in R. But all the point
of R are limit points so, o(.) = exp(.). Here we tacitly used another
basic result of complex function theory.
Theorem 90. If a complex function 1 : ¹ ÷ C, where ¹ is a
complex domain, is di¤erentiable on ¹. then it has derivatives of any
order on ¹. i.e. it is of class (
o
on ¹.
Following an analogous theory like the Weierstrass theory for the
real series of functions, we can prove that exp(.) is a di¤erential func
tion. Hence, its derivative o(.) is also di¤erentiable on C. This is why
we could apply Theorem 89 for the complex function exp(.).
What can we say about the two variables real functions n = Re 1
and · = Im1 if 1 is di¤erentiable at a point .
0
?
Theorem 91. (CauchyRiemann relations) If the function 1(r +
in) = n(r. n)+i·(r. n) is di¤erentiable at a point .
0
= r
0
+in
0
. then the
8. THE DERIVATIVE OF A FUNCTION OF A COMPLEX VARIABLE 225
two variables real functions n and · have partial derivatives at (r
0
. n
0
)
and between them we have the following relations (the CauchyRiemann
relations):
(8.2)
·n
·r
(r
0
. n
0
) =
··
·n
(r
0
. n
0
).
·n
·n
(r
0
. n
0
) = ÷
··
·r
(r
0
. n
0
)
Moreover, 1
t
(.
0
) =
0&
0a
(r
0
. n
0
) +i
0·
0a
(r
0
. n
0
) =
0·
0&
(r
0
. n
0
) ÷i
0&
0&
(r
0
. n
0
).
Proof. If 1 is di¤erentiable at the point .
0
the following limit
exists:
lim
:÷:
0
1(.) ÷1(.
0
)
. ÷.
0
= 1
t
(.
0
).
This means that for any sequence (r
a
. n
a
) which converges to (r
0
. n
0
)
(in R
2
) one has that
(8.3)
lim
an÷a
0
.&n÷&
0
n(r
a
. n
a
) ÷n(r
0
. n
0
) +i[·(r
a
. n
a
) ÷·(r
0
. n
0
)]
r
a
÷r
0
+i(n
a
÷n
0
)
= 1
t
(.
0
).
Firstly take here n
a
= n
0
for any : = 1. 2. .... We get
(8.4)
·n
·r
(r
0
. n
0
) +i
··
·r
(r
0
. n
0
) = 1
t
(.
0
).
Secondly, let us consider in (8.3) r
a
= r
0
for any : = 1. 2. .... We …nd
(8.5)
1
i
¸
·n
·n
(r
0
. n
0
) +i
··
·n
(r
0
. n
0
)
= 1
t
(.
0
)
Comparing (8.3) and (8.5) we get the CauchyRiemann relations (8.2).
The CauchyRiemann relations imply that the real and the imagi
nary part of a di¤erentiable complex function are harmonic functions,
i.e. they are solutions of the Laplace equation:
(8.6) n =
·
2
n
·r
2
+
·
2
n
·n
2
= 0
and
· =
·
2
·
·r
2
+
·
2
·
·n
2
= 0
(prove it!).
Let 1 = n + i· be a complex function di¤erentiable on a complex
open subset ¹ and let F(r. n) = (·(r. n). n(r. n)) be its associated …eld
of plane forces. By de…nition, the curl (the rotational) of F is the 3D
vector …eld curl F =(0. 0.
0&
0a
÷
0·
0&
). Since
0&
0a
=
0·
0&
on ¹. one sees that
curl F = 0 i.e. the vector …eld F is irrotational. By de…nition, the
226 10. IMPLICITLY DEFINED FUNCTIONS
divergence of F is div F =
0·
0a
+
0&
0&
. But this last one is 0 because of the
second CauchyRiemann relation.
Moreover, if one know one of the two functions n or ·. one can
determine the other up to a complex constant, such that the couple
(n. ·) be the real and the imaginary part respectively of a di¤erentiable
complex function 1. Indeed, suppose we know n and we want to …nd ·
from the CauchyRiemann relations:
(8.7)
··
·r
(r. n) = ÷
·n
·n
(r. n)
and
(8.8)
··
·n
(r. n) =
·n
·r
(r. n)
From (8.7) we can write
·(r. n) = ÷
·n
·n
(r. n)dr +((n).
We prove that we can determine the unknown function ((n) up to a
constant term. Let us come to the relation (8.8) with this last expres
sion of ·. Here we use the famous Leibniz formula on the di¤erential
of an integral with a parameter (see the Integral calculus in any course
of Analysis):
·n
·r
(r. n) = ÷
·
2
n
·n
2
(r. n)dr +(
t
(n).
From (8.6) we …nd
(8.9)
·n
·r
(r. n) =
·
2
n
·r
2
(r. n)dr +(
t
(n) =
·n
·r
(r. n) +1(n) +(
t
(n).
where ((n) and 1(n) are functions of n. From (8.9) we get
(
t
(n) = ÷1(n).
Therefore, always one can …nd the function ((n). and so the function
·(r. n) up to a real constant c. Hence, we can determine the function
1 = n +i· up to a purely imaginary constant ic.
For instance, let us consider n(r. n) = r
2
÷n
2
and let us …nd 1 (if
it is possible! It is, because n is a harmonic function!this is the only
thing we used above!). The CauchyRiemann relations become:
··
·r
(r. n) = 2n
and
··
·n
(r. n) = 2r
8. THE DERIVATIVE OF A FUNCTION OF A COMPLEX VARIABLE 227
Let us integrate the …rst equality with respect to r
·(r. n) = 2rn +((n).
where ((n) is a constant function with respect to r but,...it can depend
on n! Come now to the second relation and …nd
2r = 2r +(
t
(n).
so, (
t
(n) = 0. i.e. ((n) does not depend on n. It is a pure constant c.
Hence, ·(r. n) = 2rn+c and 1(.) = r
2
÷n
2
+i(2rn+c) = (r+in)
2
+ic.
where c is a real arbitrary constant.
Let us now come back to formula (8.1) and consider an arbitrary
smooth curve which passes through .
0
. Let us take . very close to .
0
but on the curve . So, we can approximate:
(8.10)
1(.) ÷1(.
0
)
. ÷.
0
 1
t
(.
0
)
Hence,
[1(.) ÷1(.
0
)[  [. ÷.
0
[ [1
t
(.
0
)[ =
[. ÷.
0
[
¸
·n
·r
(r
0
. n
0
)
2
+
¸
··
·r
(r
0
. n
0
)
2
.
So, the length of the segment [1(.
0
). 1(.)] is proportional to the length
of the segment [.
0
. .]. The "dilation" coe¢cient
\ =
¸
·n
·r
(r
0
. n
0
)
2
+
¸
··
·r
(r
0
. n
0
)
2
does not depend on the curve on which . becomes closer and closer to
.
0
.
Let us recall that any complex number . can be uniquely written as:
. = : exp(ic). where c ÷ [0. 2:). This angle c is called the argument
of .. From the formula (8.10) we get
(8.11) arg [1(.) ÷1(.
0
)]  arg(. ÷.
0
) + arg 1
t
(.
0
).
Here we assume that 1
t
(.
0
) = 0. Formula (8.11) says that in a small
neighborhood of .
0
our di¤erentiable function preserve the angle be
tween two curves which pass through .
0
(why?). So, we can locally
approximate the action of a di¤erentiable function by a rotation of
angle arg 1
t
(.
0
), followed by a "dilation"(or a "contraction") of coe¢
cient [1
t
(.
0
)[. We assume that 1
t
(.
0
) = 0. Otherwise, the transforma
tion . ÷ 1(.) is almost constant around .
0
. A transformation of the
complex plane into itself with this last two properties is called a con
formal transformation. These are very important in some engineering
applications (hydraulics, ‡uid mechanics, electricity, etc.).
228 10. IMPLICITLY DEFINED FUNCTIONS
If we write the plane transformation . ÷1(.) as (r. n) ÷(n(r. n). ·(r. n)).
where 1(.) = n +i·. the Jacobian determinant of this at (r
0
. n
0
) is
0&
0a
(r
0
. n
0
)
0&
0&
(r
0
. n
0
)
0·
0a
(r
0
. n
0
)
0·
0&
(r
0
. n
0
)
=
¸
·n
·r
(r
0
. n
0
)
2
+
¸
··
·r
(r
0
. n
0
)
2
= [1
t
(.
0
)[
2
.
Here we used again the CauchyRiemann relations. If we want that our
transformation . ÷1(.) to be locally invertible around the point .
0
. we
must assume that 1
t
(.
0
) = 0 (see the Local Inversion Theorem). In this
last case, this transformation is locally a conformal transformation, i.e.
it preserves the angles (with their senses) and it changes the lengthens
with the same "velocity" around the point .
0
.
9. Problems
1. Find n
t
(r) if n = 1+n
a
. Why we cannot performthis computation
for the points on the curve rn
a÷1
= 1. n 0?
2. Compute
o&
oa
and
o
2
&
oa
2
. if n = r + ln n. n = 1.
3. If . = .(r. n) and
r
3
+ 2n
3
+.
3
÷3rn. ÷2n + 3 = 0.
…nd d. and d
2
..
4. Find inf 1 and sup 1 for:
a)
1(r. n) = r
3
+ 3rn
2
÷15r ÷12n;
b)
1(r. n) = rn
with r +n ÷1 = 0;
c)
1(r. n. .) = r
2
+n
2
+.
2
with cr +/n +c. ÷1 = 0 (What this means?);
5. Find the distance from `(0. 0. 1) to the curve ¦n = r
2
¦ ¨ ¦. =
r
2
¦.
6. Find the distance between the line 3r+n ÷9 = 0 and the ellipse
a
2
4
+
&
2
9
÷1 = 0.
7. Compute the velocity and the acceleration on the circle
¦r
2
+n
2
+.
2
= c
2
¦ ¨ ¦r +n +. = c¦
by using a parametrization of the type: r = r. n = n(r). . = .(r).
9. PROBLEMS 229
8. Are the functions
n = (r +n +.)
2
. · = 3r ÷n + 3.. n = r
2
+rn +n. +.r
independent at (0. 0. 0)?
9. Change the variables in the following expressions:
a)
(1 ÷r
2
)
d
2
n
dr
2
÷r
dn
dr
+.n = 0.
r = cos t;
b)
r
2
·
2
.
·r
2
÷n
2
·
2
.
·n
2
= 0. n = rn. · =
r
n
;
c)
0&
0a
2
+
0&
0&
2
. r = j cos o. n = j sin o;
10. Find all such that n = (r + n) and · = (r)(n) be
dependent on R
2
.
11. Prove that the following complex functions are di¤erentiable
and …nd their derivatives. Take a point .
0
and study the geometrical
behavior of the transformation . ÷1(.) around this point .
0
.
a) 1(.) = 3. + 2; b) 1(.) = 2i. + 3; c) 1(.) =
1
:
. [.[ 1;
d) 1(.) = exp(i.); e) 1(.) = .
3
+ 2. . = 0; g) 1(.) = . sin .;
Bibliography
[A] T. M. Apostol, Mathematical Analysis, Narosa Publishing House, India,
2002.
[Dem] B. Demidovich, Problems in Mathematical Analysis, Mir Publishers,
Moscow, 1989.
[DOG] C. Dr¼ agu¸sin, O. Olteanu, M. Gavril¼ a, Mathematical Analysis. Theory and
Applications (Romanian), Vol. I and Vol. II, Matrix Rom, Bucharest, 2006,
2007.
[EP] E. Popescu, Mathematical Analysis (Di¤erential Calculus) (Romanian), Ma
trix Rom, Bucharest, 2006.
[FS] P. Flondor, O. St¼ an¼ a¸sil¼ a, Lectures in Mathematical Analysis (Romanian),
All Publishers, Bucharest, 1993.
[GG] G. Groza, Numerical Analysis (Romanian), Matrix Rom, 2005.
[JJ] J. Jost, Postmodern Analysis, Springer, 2002.
[La] S. Lang, Calculus of several variables, Springer Verlag, 1996.
[Nik] S. M. Nikolsky, A course of Mathematical Analysis, Vol. I, II, Mir Publishers,
Moskow, 1981.
[Pal] G. P¼ altineanu, Mathematical Analysis. Di¤erential Calculus (Romanian),
AGIR Publishers, Bucharest, 2002.
[Pro] *** Problems in Mathematical Analysis (Romanian), Department of Math.
and Computer Science, TUCIB, Matrix Rom, Bucharest, 2002.
[ST] A. Sveshnikov, A. Tikhonov, The Theory of Functions of a Complex Vari
able, Mir Publishers, 1978.
[R] W. Rudin, Principles of Mathematical Analysis, McGrawHill, N.Y., 1964.
231
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.