Calculus and applications

Teo Banica
Department of Mathematics, University of Cergy-Pontoise, F-95000
Cergy-Pontoise, France. [Link]@[Link]
 2010 Mathematics Subject Classification. 26A06
Key words and phrases. Calculus, Multivariable calculus

Abstract. This is an introduction to calculus, and its applications to basic questions

from physics. We first discuss the theory of functions f : R → R, with the notion of
Rb
continuity, and the construction of the derivative f ′ (x) and of the integral a f (x)dx.
Then we investigate the case of the complex functions f : C → C, and notably the
holomorphic functions, and harmonic functions. Then, we discuss the multivariable
functions, f : RN → RM or f : RN → CM or f : CN → CM , with general theory,
integration results, maximization questions, and basic applications to physics.
 Preface

Understanding what happens in the real life surrounding us, in phenomena involving
physics, chemistry, biology and so on, is not an easy task. What we can do as humans is
to come up with some machinery, and perform measurements, recording quantities such
as length, volume, temperature, pressure and so on, and then see how these quantities,
called “variables”, and denoted x, y, z, . . . depend on each other, and change in time.

Calculus is the study of the correspondences x → y between such variables. Such

correspondences are called “functions”, and are denoted y = f (x), with f standing for
the abstract machinery, or mathematical formula, producing y out of x.

The basics of calculus were developed by Newton, Leibnitz and others, a long time
ago. The idea is very simple. The simplest functions f : R → R are the linear ones,
f (x) = a + bx with a, b ∈ R, but of course not any function is linear. Miraculously,
however, most functions f : R → R are “locally linear”, in the sense that around any
given point c ∈ R, we have a formula of type f (c + x) ≃ a + bx, for x small. Why?
Obviously, a ∈ R can only be the value of our function at that point, a = f (c). As for the
number b ∈ R, this can be taken to be the rate of change of f around that point, called
derivative of the function at that point, and denoted b = f ′ (c).

So, this was the main idea of calculus, “functions are locally linear”. This idea applies
as well to more complicated functions, such as the “multivariable” ones f : RN → RM ,
relating vector variables x ∈ RN to vector variables y ∈ RM , with the linear approximation
formula f (c + x) ≃ a + bx needing this time as parameters a vector a = f (c) ∈ RM , and
a linear map, or beast called rectangular matrix, b = f ′ (c) ∈ MM ×N (R).

Further ideas of calculus, which are more advanced, include the facts that: (1) the
remainder ε(x) given by f (c + x) = a + bx + ε(x) can studied by using again derivatives,
(2) in several variables, the geometric understanding of the derivatives f ′ (c) ∈ MM ×N (R)
is best done by using complex numbers, (3) in fact, the use of complex numbers is useful
even for one-variable functions f : R → R, and (4) in one variable at least, there is a
magic
Rb relation between derivatives and weighted averages, called integrals and denoted
a
f (x)dx, the idea being that “the derivative of the integral is the function itself”.
3
4 PREFACE

Calculus can be learned from many places, with this being mostly a matter of taste.
Personally as a student I read the books of Rudin [73], [74], and this was a very good
investment, never had any trouble with calculus since, be that for research, or teaching.
And these are still the books that I recommend to my students, although in the present
modern age there are so many alternative resources, for having the basics learned.
The present book is an introduction to calculus, based on lecture notes from various
classes that I taught at Cergy, and previously at Toulouse. The material inside claims of
course no originality, basically going back to Newton, Leibnitz and others. But in what
regards the presentation, there are a few ideas behind it, none of these claiming of course
originality either, but their combination being something original, I hope:
(1) One complex variable comes before several real variables. This is perhaps not that
standard, but as a quantum physicist, I just love complex numbers.
(2) Applications to probability everywhere, scattered throughout the book. Again,
coming from experience with mathematics, physics, and science in general.
(3) Combinatorics, binomials and factorials all over the place, with joy. With this
being a quite popular approach, who in mathematics does not love binomials.
(4) Applications to physics too, including even the hydrogen atom, at the end. In
short, read this book, and you’ll understand how hydrogen 1 H works.
In the hope that you will like this book. High-school or undergraduate students,
wishing to learn calculus in a quick way, graduate students in math and science, wishing
to fine-tune their calculus knowledge, or just math professionals like me, wishing to have
a compact analysis book, so that they can grab the appropriate chapter, before going to
class, no matter what the class is about. Hope you will all find this useful.
As already mentioned, the present book is based on lecture notes from classes at
Toulouse and Cergy, and I would like to thank my students. Many thanks go as well to
my cats, for useful pieces of advice, often complementary to the pieces of advice of my
colleagues, and for some help with the underlying PDE and physics.

Cergy, September 2024

Teo Banica
 Contents

Preface 3

Part I. Basic calculus 9

Chapter 1. Sequences, series 11

1a. Binomials, factorials 11
1b. Real numbers, analysis 16
1c. Sequences, convergence 22
1d. Series, the number e 25
1e. Exercises 32

Chapter 2. Functions, continuity 33

2a. Continuous functions 33
2b. Intermediate values 39
2c. Sequences and series 45
2d. Basic functions 48
2e. Exercises 56

Chapter 3. Derivatives 57
3a. Derivatives, rules 57
3b. Second derivatives 66
3c. The Taylor formula 72
3d. Differential equations 77
3e. Exercises 80

Chapter 4. Integration 81
4a. Integration theory 81
4b. Riemann sums 87
4c. Advanced results 92
4d. Some probability 101
4e. Exercises 104
5
6 CONTENTS

Part II. Complex functions 105

Chapter 5. Complex numbers 107

5a. Complex numbers 107
5b. Exponential writing 115
5c. Equations, roots 121
5d. Roots of unity 126
5e. Exercises 128

Chapter 6. Complex functions 129

6a. Functions, continuity 129
6b. Holomorphic functions 137
6c. Cauchy formula 141
6d. Stieltjes inversion 147
6e. Exercises 152

Chapter 7. Fourier analysis 153

7a. Function spaces 153
7b. Fourier transform 159
7c. Groups, extensions 166
7d. Independence, limits 172
7e. Exercises 176

Chapter 8. Harmonic functions 177

8a. Laplace operator 177
8b. Harmonic functions 183
8c. Light, spectroscopy 191
8d. Atomic spectrum 196
8e. Exercises 200

Part III. Several variables 201

Chapter 9. Linear maps 203

9a. Linear maps 203
9b. Matrix inversion 206
9c. The determinant 213
9d. Diagonalization 220
9e. Exercises 224
 CONTENTS 7

Chapter 10. Partial derivatives 225

10a. Functions, continuity 225
10b. Compact sets 231
10c. Partial derivatives 238
10d. The chain rule 244
10e. Exercises 248

Chapter 11. Some geometry 249

11a. Equations, conics 249
11b. Kepler and Newton 255
11c. Algebraic manifolds 261
11d. Differential manifolds 268
11e. Exercises 272

Chapter 12. Higher derivatives 273

12a. Higher derivatives 273
12b. Matrices, positivity 281
12c. Harmonic functions 290
12d. Lagrange multipliers 292
12e. Exercises 296

Part IV. Integration theory 297

Chapter 13. Multiple integrals 299

13a. Multiple integrals 299
13b. Spherical coordinates 302
13c. Stirling estimates 306
13d. Spherical integrals 313
13e. Exercises 320

Chapter 14. Normal variables 321

14a. Probability basics 321
14b. Normal variables 329
14c. Hyperspherical laws 337
14d. Rotations, reflections 340
14e. Exercises 344

Chapter 15. Partial integration 345

8 CONTENTS

15a. Vector products 345

15b. Einstein addition 351
15c. Charges and flux 359
15d. Gauss, Green, Stokes 365
15e. Exercises 368
Chapter 16. Infinite dimensions 369
16a. Operators, matrices 369
16b. Schrödinger equation 375
16c. Spherical coordinates 377
16d. The hydrogen atom 386
16e. Exercises 392
Bibliography 393
Index 397
 Part I

Basic calculus
I’ve got to stand and fight
In this creation
Vanity I know
Can’t guide I alone
 CHAPTER 1

Sequences, series

1a. Binomials, factorials

We denote by N the set of positive integers, N = {0, 1, 2, 3, . . .}, with N standing
for “natural”. Quite often in computations we will need negative numbers too, and
we denote by Z the set of all integers, Z = {. . . , −2, −1, 0, 1, 2, . . .}, with Z standing
from “zahlen”, which is German for “numbers”. Finally, there are many questions in
mathematics involving fractions, or quotients, which are called rational numbers:
Definition 1.1. The rational numbers are the quotients of type
a
r=
b
with a, b ∈ Z, and b ̸= 0, identified according to the usual rule for quotients, namely:
a c
= ⇐⇒ ad = bc
b d
We denote the set of rational numbers by Q, standing for “quotients”.
Observe that we have inclusions N ⊂ Z ⊂ Q. The integers add and multiply according
to the rules that you know well. As for the rational numbers, these add according to the
usual rule for quotients, which is as follows, and death penalty for forgetting it:
a c ad + bc
+ =
b d bd
Also, the rational numbers multiply according to the usual rule for quotients, namely:
a c ac
· =
b d bd
Beyond rationals,
√ we have the real numbers, whose set is denoted R, and which include
beasts such as 3 = 1.73205 . . . or π = 3.14159 . . . But more on these later. For the
moment, let us see what can be done with integers, and their quotients. As a first
theorem, solving a problem which often appears in real life, we have:
Theorem 1.2. The number of possibilities of choosing k objects among n objects is
 
n n!
=
k k!(n − k)!
called binomial number, where n! = 1 · 2 · 3 . . . (n − 2)(n − 1)n, called “factorial n”.
11
12 1. SEQUENCES, SERIES

Proof. Imagine a set consisting of n objects. We have n possibilities for choosing

our 1st object, then n − 1 possibilities for choosing our 2nd object, out of the n − 1 objects
left, and so on up to n−k +1 possibilities for choosing our k-th object, out of the n−k +1
objects left. Since the possibilities multiply, the total number of choices is:
N = n(n − 1) . . . (n − k + 1)
(n − k)(n − k − 1) . . . 2 · 1
= n(n − 1) . . . (n − k + 1) ·
(n − k)(n − k − 1) . . . 2 · 1
n(n − 1) . . . 2 · 1
=
(n − k)(n − k − 1) . . . 2 · 1
n!
=
(n − k)!
But is this correct. Normally a mathematical theorem coming with mathematical
proof is guaranteed to be 100% correct, and if in addition the proof is truly clever, like
the above proof was, with that fraction trick, the confidence rate jumps up to 200%.
This being said, never knows, so let us doublecheck, by taking for instance n = 3, k = 2.
Here we have to choose 2 objects among 3 objects, and this is something easily done,
because what we have to do is to dismiss one of the objects, and N = 3 choices here, and
keep the 2 objects left. Thus, we have N = 3 choices. On the other hand our genius math
computation gives N = 3!/1! = 6, which is obviously the wrong answer.
So, where is the mistake? Thinking a bit, the number N that we computed is in fact
the number of possibilities of choosing k ordered objects among n objects. Thus, we must
divide everything by the number M of orderings of the k objects that we chose:
 
n N
=
k M
In order to compute now the missing number M , imagine a set consisting of k objects.
There are k choices for the object to be designated #1, then k − 1 choices for the object
to be designated #2, and so on up to 1 choice for the object to be designated #k. We
conclude that we have M = k(k − 1) . . . 2 · 1 = k!, and so:
 
n n!/(n − k)! n!
= =
k k! k!(n − k)!
And this is the correct answer, because, well, that is how things are. In case you
doubt, at n = 3, k = 2 for instance we obtain 3!/2!1! = 3, which is correct. □
All this is quite interesting, and in addition to having some exciting mathematics going
on, and more on this in a moment, we have as well some philosophical conclusions. For-
mulae can be right or wrong, and as the above shows, good-looking, formal mathematical
proofs can be right or wrong too. So, what to do? Here is my advice:
 1A. BINOMIALS, FACTORIALS 13

Advice 1.3. Always doublecheck what you’re doing, regularly, and definitely at the
end, either with an alternative proof, or with some numerics.
This is something very serious. Unless you’re doing something very familiar, that
you’re used to for at least 5-10 years or so, like doing additions and multiplications for
you, or some easy calculus for me, formulae and proofs that you can come upon are by
default wrong. In order to make them correct, and ready to use, you must check and
doublecheck and correct them, helped by alternative methods, or numerics.

Which brings us into the question on whether mathematics is an exact science or not.
Not clear. Chemistry for instance is an exact science, because findings of type “a mixture
of water and salt cannot explode” look rock-solid. Same for biology, with findings of type
“crocodiles eat fish” being rock-solid too. In what regards mathematics however, and
theoretical physics too, things are always prone to human mistake.

And for ending this discussion, you might ask then, what about engineering? After
all, this is mathematics and physics, which is usually 100% correct, because most of the
bridges, buildings and other things built by engineers don’t collapse. Well, this is because
engineers follow, and in a truly maniac way, the above Advice 1.3. You won’t declare a
project for a bridge, building, engine and so on final and correct, ready for production,
until you checked and doublechecked it with 10 different methods or so, won’t you.

Back to work now, as an important adding to Theorem 1.2, we have:

Convention 1.4. By definition, 0! = 1.
This
convention comes, and no surprise here, from Advice 1.3. Indeed, we obviously
have nn = 1, but if we want to recover this formula via Theorem 1.2 we are a bit in
trouble, and so we must declare that 0! = 1, as for the following computation to work:
 
n n! n!
= = =1
n n!0! n! × 1
Going ahead now with more mathematics and less philosophy, with Theorem 1.2
complemented by Convention 1.4 being in final form (trust me), we have:
Theorem 1.5. We have the binomial formula
n  
n
X n k n−k
(a + b) = a b
k=0
k
valid for any two numbers a, b ∈ Q.
Proof. We have to compute the following quantity, with n terms in the product:
(a + b)n = (a + b)(a + b) . . . (a + b)
14 1. SEQUENCES, SERIES

When expanding, we obtain a certain sum of products of a, b variables, with each such
product being a quantity of type ak bn−k . Thus, we have a formula as follows:
Xn
n
(a + b) = Ck ak bn−k
k=0

In order to finish, it remains to compute the coefficients Ck . But, according to our

product formula, Ck is the number of choices for the k needed a variables among the n
available a variables. Thus, according to Theorem 1.2, we have:
 
n
Ck =
k
We are therefore led to the formula in the statement. □
Theorem 1.5 is something quite interesting, so let us doublecheck it with some numer-
ics. At small values of n we obtain the following formulae, which are all correct:
(a + b)0 = 1
(a + b)1 = a + b
(a + b)2 = a2 + 2ab + b2
(a + b)3 = a3 + 3a2 b + 3ab2 + b3
(a + b)4 = a4 + 4a3 b + 6a2 b2 + 4ab3 + b4
(a + b)5 = a5 + 5a4 b + 10a3 b2 + 10a2 b3 + 5a4 b + b5
..
.
Now observe that in these formulae, say for memorization purposes, the powers of the
a, b variables are something very simple, that can be recovered right away. What matters
are the coefficients, which are the binomial coefficients nk , which form a triangle. So, it


is enough to memorize this triangle, and this can be done by using:
Theorem 1.6. The Pascal triangle, formed by the binomial coefficients nk ,

1
1 , 1
1 , 2 , 1
1 , 3 , 3 , 1
1 , 4 , 6 , 4 , 1
1 , 5 , 10 , 10 , 5 , 1
..
.
has the property that each entry is the sum of the two entries above it.
 1A. BINOMIALS, FACTORIALS 15

Proof. In practice, the theorem states that the following formula holds:
     
n n−1 n−1
= +
k k−1 k
There are many ways of proving this formula, all instructive, as follows:
(1) Brute-force computation. We have indeed, as desired:
   
n−1 n−1 (n − 1)! (n − 1)!
+ = +
k−1 k (k − 1)!(n − k)! k!(n − k − 1)!
 
(n − 1)! 1 1
= +
(k − 1)!(n − k − 1)! n − k k
(n − 1)! n
= ·
(k − 1)!(n − k − 1)! k(n − k)
 
n
=
k
(2) Algebraic proof. We have the following formula, to start with:
(a + b)n = (a + b)n−1 (a + b)
By using the binomial formula, this formula becomes:
n  
" n−1   #
X n k n−k X n − 1 r n−1−r
a b = ab (a + b)
k=0
k r=0
r
Now let us perform the multiplication on the right. We obtain a certain sum of terms
k n−k
of type

a b k−1 , and to be more precise, each such ak bn−k

term can either come from the
n−1 n−k n−1 k n−1−k
k−1
terms a b multiplied by a, or from the terms a b multiplied by b.
k n−k n−1

k n−1

Thus, the coefficient of a b on the right is k−1 + k , as desired.
(3) Combinatorics. Let us count k objects among n objects, with one of the n objects
having a hat on top. Obviously, the hat has nothing to do with the count, and we obtain
n


k
. On the other hand, we can say that

there are two possibilities. Either the object
n−1
with hat is counted, and we have k−1 possibilities here, or the object with hat is not
counted, and we have n−1 possibilities here. Thus nk = n−1 + n−1





k k−1 k
, as desired. □
There are many more things that can be said about binomial coefficients, with all
sorts of interesting formulae, but the idea is always the same, namely that in order to find
such formulae you have a choice between algebra and combinatorics, and that when it
comes to proofs, the brute-force computation method is useful too. In practice, the best
is to master all 3 techniques. Among others, because of Advice 1.3. You will have in this
way 3 different methods, for making sure that your formulae are correct indeed.
16 1. SEQUENCES, SERIES

1b. Real numbers, analysis

All the above was very nice, but remember that we are here for doing science and
physics, and more specifically for mathematically understanding the numeric variables
x, y, z, . . . coming from real life. Such variables can be lengths, volumes, pressures and so
on, which vary continuously with time, and common sense dictates that there is little to
no chance for our variables to be rational, x, y, z, . . . ∈
/ Q. In fact, we will even see soon a
theorem, stating that the probability for such a variable to be rational is exactly 0. Or,
to put it in a dramatic way, “rational numbers don’t exist in real life”.

You are certainly familiar with the real numbers, but let us review now their definition,
which
√ is something quite tricky. As a first goal, we would like to construct a number
2
x = 2 having the property x = 2. But how to do this? Let us start with:
Proposition 1.7. There is no number r ∈ Q+ satisfying r2 = 2. In fact, we have
n oGn o
Q+ = p ∈ Q+ p2 < 2 q ∈ Q+ q 2 > 2

with this being a disjoint union.

Proof. In what regards the first assertion, assuming that r = a/b with a, b ∈ N prime
to each other satisfies r2 = 2, we have a2 = 2b2 , so a ∈ 2N. But by using again a2 = 2b2
we obtain b ∈ 2N, contradiction. As for the second assertion, this is obvious. □
√
It looks like we √
are a bit stuck. We can’t really tell who 2 is, and the only piece of
information about 2 that we have comes from the knowledge of the rational numbers
satisfying p2 < 2 or q 2 > 2. To be more precise, the picture that emerges is:
√
Conclusion 1.8. The number 2 is the abstract beast which is bigger than all ratio-
nals satisfying p2 < 2, and smaller than all positive rationals satisfying q 2 > 2.
This does not look very good, but you know what, instead of looking for more clever
solutions to our problem, what about relaxing,√ or being lazy, or coward, or you name it,
and taking Conclusion 1.8 as a definition for 2. This is actually something not that bad,
and leads to the following “lazy” definition for the real numbers:
Definition 1.9. The real numbers x ∈ R are formal cuts in the set of rationals,
Q = Q≤x ⊔ Q>x
with such a cut being by definition subject to the following condition:
p ∈ Q≤x , q ∈ Q>x =⇒ p < q
These numbers add and multiply by adding and multiplying the corresponding cuts.
 1B. REAL NUMBERS, ANALYSIS 17

This might look quite original, but believe me, there is some genius behind this defi-
nition. As a first observation, we have an inclusion Q ⊂ R, obtained by identifying each
rational number r ∈ Q with the obvious cut that it produces, namely:
n o n o
Q≤r = p∈Qp≤r , Q>r = q ∈ Q q > r

As a second observation, the addition and multiplication of real numbers, obtained

by adding and multiplying the corresponding cuts, in the obvious way, is something very
simple. To be more precise, in what regards the addition, the formula is as follows:

Q≤x+y = Q≤x + Q≤y

As for the multiplication, the formula here is similar, namely Q≤xy = Q≤x Q≤y , up to
some mess with positives and negatives, which is quite easy to untangle, and with this
being a good exercise. We can also talk about order between real numbers, as follows:

x ≤ y ⇐⇒ Q≤x ⊂ Q≤y

But let us perhaps leave more abstractions for later, and go back to more concrete
things. As a first success of our theory, we can formulate the following theorem:
2
Theorem 1.10. The equation√ x = 2 has two solutions over the real numbers,
√ namely
the positive solution, denoted 2, and its negative counterpart, which is − 2.
2
√ By using x → −x, it is enough
Proof. √ to prove that x = 2 has exactly one positive
solution 2. But this is clear, because 2 can only come from the following cut:
Gn o n o
Q≤√2 = Q− p ∈ Q+ p2 ≤ 2 , Q>√2 = q ∈ Q+ q 2 > 2

Thus, we are led to the conclusion in the statement. □

√
More generally, the same method works in order to extract the square root r of any
number r ∈ Q+ , or even of any number r ∈ R+ , and we have the following result:

Theorem 1.11. The solutions of ax2 + bx + c = 0 with a, b, c ∈ R are

√
−b ± b2 − 4ac
x1,2 =
2a

provided that b2 − 4ac ≥ 0. In the case b2 − 4ac < 0, there are no solutions.
18 1. SEQUENCES, SERIES

Proof. We can write our equation in the following way:

b c
ax2 + bx + c = 0 x2 + x + = 0
⇐⇒
a a
2
b2

b c
⇐⇒ x+ − 2 + =0
2a 4a a
 2 2
b b − 4ac
⇐⇒ x+ =
2a 4a2
√
b b2 − 4ac
⇐⇒ x + =±
2a 2a
Thus, we are led to the conclusion in the statement. □
Summarizing, we have a nice definition for the real numbers, that we can certainly
do some math with. However, for anything more advanced we are in need of the decimal
writing for the real numbers. The result here is as follows:
Theorem 1.12. The real numbers x ∈ R can be written in decimal form,
x = ±a1 . . . an .b1 b2 b3 . . . . . .
with ai , bi ∈ {0, 1, . . . , 9}, with the convention . . . b999 . . . = . . . (b + 1)000 . . .
Proof. This is something quite non-trivial, assuming that you already have some
familiarity with such things, for the rational numbers. The idea is as follows:
(1) First of all, our precise claim is that any x ∈ R can be written in the form in the
statement, with the integer ±a1 . . . an and then each of the digits b1 , b2 , b3 , . . . providing
the best approximation of x, at that stage of the approximation.
(2) Moreover, we have a second claim as well, namely that any expression of type
x = ±a1 . . . an .b1 b2 b3 . . . . . . corresponds to a real number x ∈ R, and that with the
convention . . . b999 . . . = . . . (b + 1)000 . . . , the correspondence is bijective.
(3) In order to prove now these two assertions, our first claim is that we can restrict
the attention to the case x ∈ [0, 1), and with this meaning of course 0 ≤ x < 1, with
respect to the order relation for the reals discussed in the above.
(4) Getting started now, let x ∈ R, coming from a cut Q = Q≤x ⊔ Q>x . Since the set
Q≤x ∩ Z consists of integers, and is bounded from above by any element q ∈ Q>x of your
choice, this set has a maximal element, that we can denote [x]:
[x] = max (Q≤x ∩ Z)
It follows from definitions that [x] has the usual properties of the integer part, namely:
[x] ≤ x < [x] + 1
 1B. REAL NUMBERS, ANALYSIS 19

Thus we have x = [x] + y with [x] ∈ Z and y ∈ [0, 1), and getting back now to what we
want to prove, namely (1,2) above, it is clear that it is enough to prove these assertions
for the remainder y ∈ [0, 1). Thus, we have proved (3), and we can assume x ∈ [0, 1).
(5) So, assume x ∈ [0, 1). We are first looking for a best approximation from below of
type 0.b1 , with b1 ∈ {0, . . . , 9}, and it is clear that such an approximation exists, simply
by comparing x with the numbers 0.0, 0.1, . . . , 0.9. Thus, we have our first digit b1 , and
then we can construct the second digit b2 as well, by comparing x with the numbers
0.b1 0, 0.b1 1, . . . , 0.b1 9. And so on, which finishes the proof of our claim (1).
(6) In order to prove now the remaining claim (2), let us restrict again the attention,
as explained in (4), to the case x ∈ [0, 1). First, it is clear that any expression of type
x = 0.b1 b2 b3 . . . defines a real number x ∈ [0, 1], simply by declaring that the corresponding
cut Q = Q≤x ⊔ Q>x comes from the following set, and its complement:
[n o
Q≤x = p ∈ Q p ≤ 0.b1 . . . bn
n≥1

(7) Thus, we have our correspondence between real numbers as cuts, and real numbers
as decimal expressions, and we are left with the question of investigating the bijectivity
of this correspondence. But here, the only bug that happens is that numbers of type
x = . . . b999 . . ., which produce reals x ∈ R via (6), do not come from reals x ∈ R via (5).
So, in order to finish our proof, we must investigate such numbers.
(8) So, consider an expression of type . . . b999 . . . Going back to the construction in
(6), we are led to the conclusion that we have the following equality:
Q≤...b999... = Q≤...(b+1)000...
Thus, at the level of the real numbers defined as cuts, we have:
. . . b999 . . . = . . . (b + 1)000 . . .
But this solves our problem, because by identifying . . . b999 . . . = . . . (b + 1)000 . . . the
bijectivity issue of our correspondence is fixed, and we are done. □
The above theorem was of course quite difficult, but this is how things are. You
might perhaps say why bothering with cuts, and not taking x = ±a1 . . . an .b1 b2 b3 . . . . . .
as definition for the real numbers. Well, this is certainly possible, but when it comes to
summing
√ such numbers, or making products, or proving basic things such as the existence
of 2, things become fairly complicated with the decimal writing picture. So, all the above
is not as stupid as it seems. And we will come back anyway to all this later, with a 3rd
picture for the real numbers, involving scary things like ε and δ, and it will be up to you
to decide, at that time, which picture is the one that you prefer.

Moving on, we made the claim in the beginning of this chapter that “in real life, real
numbers are never rational”. Here is a theorem, justifying this claim:
20 1. SEQUENCES, SERIES

Theorem 1.13. The probability for a real number x ∈ R to be rational is 0.

Proof. This is something quite tricky, the idea being as follows:
(1) Before starting, let us point out the fact that probability theory is something quite
tricky, with probability 0 not necessarily meaning that the event cannot happen, but
rather meaning that “better not count on that”. For instance according to my compu-
tations the probability of you winning 1 billion at the lottery is 0, but you are of course
free to disagree, and prove me wrong, by playing every day at the lottery.
(2) With this discussion made, and extrapolating now from finance and lottery to
our question regarding real numbers, your possible argument of type “yes, but if I pick
x ∈ R to be x = 3/2, I have proof that the probability for x ∈ Q is nonzero” is therefore
dismissed. Thus, our claim as stated makes sense, so let us try now to prove it.
(3) By translation, it is enough to prove that the probability for a real number x ∈ [0, 1]
to be rational is 0. For this purpose, let us write the rational numbers r ∈ [0, 1] in the
form of a sequence r1 , r2 , r3 . . . , with this being possible say by ordering our rationals
r = a/b according to the lexicographic order on the pairs (a, b):

Q ∩ [0, 1] = r1 , r2 , r3 , . . .
Let us also pick a number c > 0. Since the probability of having x = r1 is certainly
smaller than c/2, then the probability of having x = r2 is certainly smaller than c/4, then
the probability of having x = r3 is certainly smaller than c/8 and so on, the probability
for x to be rational satisfies the following inequality:
c c c
P ≤ + + + ...
2 4 8 
1 1 1
= c + + + ...
2 4 8
= c
Here we have used the well-known formula 12 + 14 + 18 +. . . = 1, which comes by dividing
[0, 1] into half, and then one of the halves into half again, and so on, and then saying in
the end that the pieces that we have must sum up to 1. Thus, we have indeed P ≤ c, and
since the number c > 0 was arbitrary, we obtain P = 0, as desired. □
As a comment here, all the above is of course quite tricky, and a bit bordeline in respect
to what can be called “rigorous mathematics”. But we will be back to this, namely general
probability theory, and in particular meaning of the mysterious formula P = 0, countable
sets, infinite sums and so on, on several occasions, throughout this book.
Moving
√ ahead now, let us construct now some more real numbers. We already know
about 2 and other numbers of the same type, namely roots of polynomials, and our
knowledge here being quite decent, no hurry with this, we will be back to it later. So, let
us get now into π and trigonometry. To start with, we have the following result:
 1B. REAL NUMBERS, ANALYSIS 21

Theorem 1.14. The following two definitions of π are equivalent:

(1) The length of the unit circle is L = 2π.
(2) The area of the unit disk is A = π.
Proof. In order to prove this theorem let us cut the unit disk as a pizza, into N
slices, and forgetting about gastronomy, leave aside the rounded parts:
◦ ◦

◦ ◦ ◦

◦ ◦
The area to be eaten can be then computed as follows, where H is the height of the
slices, S is the length of their sides, and P = N S is the total length of the sides:
HS
A = N×
2
HP
=
2
1×L
≃
2
Thus, with N → ∞ we obtain that we have A = L/2, as desired. □
In what regards now the precise value of π, the above picture at N = 6 shows that
we have π > 3, but not by much. The precise figure is π = 3.14159 . . . , but we will come
back to this later, once we will have appropriate tools for dealing with such questions. It
is also possible to prove that π is irrational, π ∈
/ Q, but this is not trivial either.

Let us end this discussion about real numbers with some trigonometry. There are
many things that can be said, that you certainly know, the basics being as follows:
Theorem 1.15. The following happen:
(1) We can talk about angles x ∈ R, by using the unit circle, in the usual way, and
in this correspondence, the right angle has a value of π/2.
(2) Associated to any x ∈ R are numbers sin x, cos x ∈ R, constructed in the usual
way, by using a triangle. These numbers satisfy sin2 x + cos2 x = 1.
Proof. There are certainly things that you know, the idea being as follows:
(1) The formula L = 2π from Theorem 1.14 shows that the length of a quarter of the
unit circle is l = π/2, and so the right angle has indeed this value, π/2.
22 1. SEQUENCES, SERIES

(2) As for sin2 x+cos2 x = 1, called Pythagoras’ theorem, this comes from the following
picture, consisting of two squares and four identical triangles, as indicated:
◦ ◦ ◦

◦ 1 sin x

◦ ◦ cos x ◦
Indeed, when computing the area of the outer square, we obtain:
sin x cos x
(sin x + cos x)2 = 1 + 4 ×
2
2 2
Now when expanding we obtain sin x + cos x = 1, as claimed. □
It is possible to say many more things about angles and sin x, cos x, and also talk about
some supplementary quantities, such as tan x = sin x/ cos x. But more on this later, once
we will have some appropriate tools, beyond basic geometry, in order to discuss this.

1c. Sequences, convergence

We already met, on several occasions, infinite sequences or sums, and their limits.
Time now to clarify all this. Let us start with the following definition:
Definition 1.16. We say that a sequence {xn }n∈N ⊂ R converges to x ∈ R when:
∀ε > 0, ∃N ∈ N, ∀n ≥ N, |xn − x| < ε
In this case, we write limn→∞ xn = x, or simply xn → x.
This might look quite scary, at a first glance, but when thinking a bit, there is nothing
scary about it. Indeed, let us try to understand, how shall we translate xn → x into
mathematical language. The condition xn → x tells us that “when n is big, xn is close
to x”, and to be more precise, it tells us that “when n is big enough, xn gets arbitrarily
close to x”. But n big enough means n ≥ N , for some N ∈ N, and xn arbitrarily close to
x means |xn − x| < ε, for some ε > 0. Thus, we are led to the above definition.

As a basic example for all this, we have:

Proposition 1.17. We have 1/n → 0.
Proof. This is obvious, but let us prove it by using Definition 1.16. We have:
1 1 1
− 0 < ε ⇐⇒ < ε ⇐⇒ <n
n n ε
Thus we can take N = [1/ε] + 1 in Definition 1.16, and we are done. □
 1C. SEQUENCES, CONVERGENCE 23

There are many other examples, and more on this in a moment. Going ahead with
more theory, let us complement Definition 1.16 with:
Definition 1.18. We write xn → ∞ when the following condition is satisfied:
∀K > 0, ∃N ∈ N, ∀n ≥ N, xn > K
Similarly, we write xn → −∞ when the same happens, with xn < −K at the end.
Again, this is something very intuitive, coming from the fact that xn → ∞ can only
mean that xn is arbitrarily big, for n big enough. As a basic illustration, we have:
Proposition 1.19. We have n2 → ∞.
Proof. As before, this is obvious, but let us prove it using Definition 1.18. We have:
√
n2 > K ⇐⇒ n > K
√
Thus we can take N = [ K] + 1 in Definition 1.18, and we are done. □
We can unify and generalize Proposition 1.17 and Proposition 1.19, as follows:
Proposition 1.20. We have the following convergence, with n → ∞:

0
 (a < 0)
a
n → 1 (a = 0)

∞ (a > 0)

Proof. This follows indeed by using the same method as in the proof of Proposition
1.17 and Proposition 1.19, first for a rational, and then for a real as well. □
We have some general results about limits, summarized as follows:
Theorem 1.21. The following happen:
(1) The limit limn→∞ xn , if it exists, is unique.
(2) If xn → x, with x ∈ (−∞, ∞), then xn is bounded.
(3) If xn is increasing or descreasing, then it converges.
(4) Assuming xn → x, any subsequence of xn converges to x.
Proof. All this is elementary, coming from definitions:
(1) Assuming xn → x, xn → y we have indeed, for any ε > 0, for n big enough:
|x − y| ≤ |x − xn | + |xn − y| < 2ε
(2) Assuming xn → x, we have |xn − x| < 1 for n ≥ N , and so, for any k ∈ N:
|xk | < 1 + |x| + sup (|x1 |, . . . , |xn−1 |)
24 1. SEQUENCES, SERIES

(3) By using x → −x, it is enough to prove the result for increasing sequences. But
here we can construct the limit x ∈ (−∞, ∞] in the following way:
[
(−∞, xn ) = (−∞, x)
n∈N

(4) This is clear from definitions. □

Here are as well some general rules for computing limits:
Theorem 1.22. The following happen, with the conventions ∞+∞ = ∞, ∞·∞ = ∞,
1/∞ = 0, and with the conventions that ∞ − ∞ and ∞ · 0 are undefined:
(1) xn → x implies λxn → λx.
(2) xn → x, yn → y implies xn + yn → x + y.
(3) xn → x, yn → y implies xn yn → xy.
(4) xn → x with x ̸= 0 implies 1/xn → 1/x.
Proof. All this is again elementary, coming from definitions:
(1) This is something which is obvious from definitions.
(2) This follows indeed from the following estimate:
|xn + yn − x − y| ≤ |xn − x| + |yn − y|
(3) This follows indeed from the following estimate:
|xn yn − xy| = |(xn − x)yn + x(yn − y)|
≤ |xn − x| · |yn | + |x| · |yn − y|
(4) This is again clear, by estimating 1/xn − 1/x, in the obvious way. □
As an application of the above rules, we have the following useful result:
Proposition 1.23. The n → ∞ limits of quotients of polynomials are given by
ap np + ap−1 np−1 + . . . + a0 ap np
lim = lim
n→∞ bq nq + bq−1 nq−1 + . . . + b0 n→∞ bq nq

with the limit on the right being ±∞, 0, ap /bq , depending on the values of p, q.
Proof. The first assertion comes from the following computation:
ap np + ap−1 np−1 + . . . + a0 np ap + ap−1 n−1 + . . . + a0 n−p
lim = lim ·
n→∞ bq nq + bq−1 nq−1 + . . . + b0 n→∞ nq bq + bq−1 n−1 + . . . + b0 n−q
p
ap n
= lim
n→∞ bq nq

As for the second assertion, this comes from Proposition 1.20. □

 1D. SERIES, THE NUMBER E 25

Getting back now to theory, some sequences which obviously do not converge, like for
instance xn = (−1)n , have however “2 limits instead of 1”. So let us formulate:
Definition 1.24. Given a sequence {xn }n∈N ⊂ R, we let
lim inf xn ∈ [−∞, ∞] , lim sup xn ∈ [−∞, ∞]
n→∞ n→∞

to be the smallest and biggest limit of a subsequence of (xn ).

Observe that the above quantities are defined indeed for any sequence xn . For instance,
for xn = (−1)n we obtain −1 and 1. Also, for xn = n we obtain ∞ and ∞. And so on.
Of course, and generalizing the xn = n example, if xn → x we obtain x and x.

Going ahead with more theory, here is a key result:

Theorem 1.25. A sequence xn converges, with finite limit x ∈ R, precisely when
∀ε > 0, ∃N ∈ N, ∀m, n ≥ N, |xm − xn | < ε
called Cauchy condition.
Proof. In one sense, this is clear. In the other sense, we can say for instance that
the Cauchy condition forces the decimal writings of our numbers xn to coincide more and
more, with n → ∞, and so we can construct a limit x = limn→∞ xn , as desired. □
The above result is quite interesting, and as an application, we have:
Theorem 1.26. R is the completion of Q, in the sense that it is the space of Cauchy
sequences over Q, identified when the virtual limit is the same, in the sense that:
xn ∼ yn ⇐⇒ |xn − yn | → 0
Moreover, R is complete, in the sense that it equals its own completion.
Proof. Let us denote the completion operation by X → X̄ = CX / ∼, where CX is
the space of Cauchy sequences over X, and ∼ is the above equivalence relation. Since by
Theorem 1.25 any Cauchy sequence (xn ) ∈ CQ has a limit x ∈ R, we obtain Q̄ = R. As
for the equality R̄ = R, this is clear again by using Theorem 1.25. □

1d. Series, the number e

With the above understood, we are now ready to get into some truly interesting
mathematics. Let us start with the following definition:
Definition 1.27. Given numbers x0 , x1 , x2 , . . . ∈ R, we write
∞
X
xn = x
n=0
Pk
with x ∈ [−∞, ∞] when limk→∞ n=0 xn = x.
26 1. SEQUENCES, SERIES

As before with the sequences, there is some general theory that can be developed for
the series, and more on this in a moment. As a first, basic example, we have:
Theorem 1.28. We have the “geometric series” formula
∞
X 1
xn =
n=0
1−x

valid for any |x| < 1. For |x| ≥ 1, the series diverges.
Proof. Our first claim, which comes by multiplying and simplifying, is that:
k
X 1 − xk+1
xn =
n=0
1−x

But this proves the first assertion, because with k → ∞ we get:

k
X 1
xn →
n=0
1−x

As for the second assertion, this is clear as well from our formula above. □

Less trivial now is the following result, due to Riemann:

Theorem 1.29. We have the following formula:
1 1 1
1+ + + + ... = ∞
2 3 4
1/na converges for a > 1, and diverges for a ≤ 1.
P
In fact, n

Proof. We have to prove several things, the idea being as follows:

(1) The first assertion comes from the following computation:
   
1 1 1 1 1 1 1 1 1 1
1 + + + + ... = 1 + + + + + + + + ...
2 3 4 2 3 4 5 6 7 8
   
1 1 1 1 1 1 1
≥ 1+ + + + + + + + ...
2 4 4 8 8 8 8
1 1 1
= 1 + + + + ...
2 2 2
= ∞
(2) Regarding now the second assertion, we have that at a = 1, and so at any a ≤ 1.
Thus, it remains to prove that at a > 1 the series converges. Let us first discuss the case
 1D. SERIES, THE NUMBER E 27

a = 2, which will prove the convergence at any a ≥ 2. The trick here is as follows:
1 1 1 1 1 1
1+ + + + ... ≤ 1 + + + + ...
4 9 16  3 6 10 
1 1 1 1
= 2 + + + + ...
2 6 12 20
        
1 1 1 1 1 1 1
= 2 1− + − + − + − ...
2 2 3 3 4 4 5
= 2
(3) It remains to prove that the series converges at a ∈ (1, 2), and here it is enough
to deal with the case of the exponents a = 1 + 1/p with p ∈ N. We already know how to
do this at p = 1, and the proof at p ∈ N will be based on a similar trick. We have:
∞
X 1 1
1/p
− =1
n=0
n (n + 1)1/p
Let us compute, or rather estimate, the generic term of this series. By using the
formula ap − bp = (a − b)(ap−1 + ap−2 b + . . . + abp−2 + bp−1 ), we have:
1 1 (n + 1)1/p − n1/p
− =
n1/p (n + 1)1/p n1/p (n + 1)1/p
1
=
n1/p (n + 1)1/p [(n + 1)1−1/p + . . . + n1−1/p ]
1
≥
n1/p (n + 1)1/p · p(n + 1)1−1/p
1
=
pn1/p (n + 1)
1
≥
p(n + 1)1+1/p
We therefore obtain the following estimate for the Riemann sum:
∞ ∞
X 1 X 1
= 1+
n=0
n1+1/p n=0
(n + 1)1+1/p
∞  
X 1 1
≤ 1+p −
n=0
n1/p (n + 1)1/p
= 1+p
Thus, we are done with the case a = 1 + 1/p, which finishes the proof. □
Here is another tricky result, this time about alternating sums:
28 1. SEQUENCES, SERIES

Theorem 1.30. We have the following convergence result:

1 1 1
1 − + − + ... < ∞
2 3 4
However, when rearranging terms, we can obtain any x ∈ [−∞, ∞] as limit.
Proof. Both the assertions follow from Theorem 1.29, as follows:
(1) We have the following computation, using the Riemann criterion at a = 2:
   
1 1 1 1 1 1
1 − + − + ... = 1− + − + ...
2 3 4 2 3 4
1 1 1
= + + + ...
2 12 30
1 1 1
< 2 + 2 + 2 + ...
1 2 3
< ∞
(2) We have the following formulae, coming from the Riemann criterion at a = 1:
 
1 1 1 1 1 1 1 1
+ + + + ... = 1 + + + + ... = ∞
2 4 6 8 2 2 3 4
1 1 1 1 1 1 1
1 + + + + ... ≥ + + + + ... = ∞
3 5 7 2 4 6 8
Thus, both these series diverge. The point now is that, by using this, when rearranging
terms in the alternating series in the statement, we can arrange for the partial sums to
go arbitrarily high, or arbitrarily low, and we can obtain any x ∈ [−∞, ∞] as limit. □
Back now to the general case, we first have the following statement:
Theorem 1.31. The following hold, with the converses of (1) and (2) being wrong,
and with (3) not holding when the assumption xn ≥ 0 is removed:
P
(1) If Pn xn converges then xPn → 0.
(2) If Pn |xn | converges then n xn converges. P
(3) If n xn converges, xn ≥ 0 and xn /yn → 1 then n yn converges.
Proof. This is a mixture of trivial and non-trivial results, as follows:
(1) We know that n xn converges when Sk = kn=0 xn converges. Thus by Cauchy
P P
we have xk = Sk −Sk−1 → 0, and this gives the result. As for the simplest counterexample
for the converse, this is 1 + 12 + 13 + 14 + . . . = ∞, coming from Theorem 1.29.
(2) This follows again from the Cauchy criterion, by using:
|xn + xn+1 + . . . + xn+k | ≤ |xn | + |xn+1 | + . . . + |xn+k |
As for the simplest counterexample for the converse, this is 1 − 12 + 13 − 41 + . . . < ∞,
coming from Theorem 1.30, coupled with 1 + 21 + 13 + 14 + . . . = ∞ from (1).
 1D. SERIES, THE NUMBER E 29

(3) Again, the main assertion here is clear, coming from, for n big:
(1 − ε)xn ≤ yn ≤ (1 + ε)xn
In what regards now the failure of the result, when the assumption xn ≥ 0 is removed,
this is something quite tricky, the simplest counterexample being as follows:
(−1)n 1 (−1)n
xn = √ , yn = + √
n n n
To be more precise, we have yn /xn → 1, soPxn /yn → 1 too, but according
P to the above-
mentioned results from (1,2), modified a bit, n xn converges, while n yn diverges. □
Summarizing, we have some useful positive results about series, which are however
quite trivial, along with various counterexamples to their possible modifications, which
are non-trivial. Staying positive, here are some more positive results:
Theorem 1.32. The following happen, and in all cases, the situtation where c = 1 is
indeterminate, in the sense that the series can converge or diverge:
P
(1) If |x n+1 /xn | → c, the series x converges if c < 1, and diverges if c > 1.
p P n n
(2) If |xn | → c, the series
n
p nx n converges if c < 1, and diverges if c > 1.
P
(3) With c = lim supn→∞ |xn |, n xn converges if c < 1, and diverges if c > 1.
n

Proof. Again, this is a mixture of trivial and non-trivial results, as follows:

(1) Here the main assertions, regarding
P n the cases c < 1 and c > 1, are both clear by
comparing with the geometric series n c . As for the case c = 1, this is what happens
for the Riemann series n 1/na , so we can have both convergent and divergent series.
P

(2) Again, thePmain assertions, where c < 1 or c > 1, are clear by comparing with the
geometric series n cn , and the c = 1 examples come from the Riemann series.
(3) Here the case c < 1 is dealt with as in (2), and the same goes for the examples at
c = 1. As for the case c > 1, this is clear too, because here xn → 0 fails. □
Finally, generalizing the first assertion in Theorem 1.30, we have:
Theorem 1.33. If xn ↘ 0 then n (−1)n xn converges.
P

Proof. We have the n (−1)n xn = k yk , where:

P P

yk = x2k − x2k+1
But, by drawing P for instance the numbers xi on the real line, we see that yk are positive
numbers, and that k yk is the sum P of lengths of certain disjoint intervals, included in
the interval [0, x0 ]. Thus we have k yk ≤ x0 , and this gives the result. □
All this was a bit theoretical, and as something more concrete now, we have:
30 1. SEQUENCES, SERIES

Theorem 1.34. We have the following convergence

 n
1
1+ →e
n
where e = 2.71828 . . . is a certain number.
Proof. This is something quite tricky, as follows:
(1) Our first claim is that the following sequence is increasing:
 n
1
xn = 1 +
n
In order to prove this, we use the following arithmetic-geometric inequality:
v
Pn 1

u n  
1 + i=1 1 + n u Y 1
≥ n+1
t 1· 1+
n+1 i=1
n

In practice, this gives the following inequality:

 n/(n+1)
1 1
1+ ≥ 1+
n+1 n
Now by raising to the power n + 1 we obtain, as desired:
 n+1  n
1 1
1+ ≥ 1+
n+1 n
(2) Normally we are left with proving that xn is bounded from above, but this is
non-trivial, and we have to use a trick. Consider the following sequence:
 n+1
1
yn = 1 +
n
We will prove that this sequence yn is decreasing, and together with the fact that we
have xn /yn → 1, this will give the result. So, this will be our plan.
(3) In order to prove now that yn is decreasing, we use, a bit as before:
v
Pn 1

u n 
1 + i=1 1 − n

u Y 1
≥ n+1
t 1· 1−
n+1 i=1
n

In practice, this gives the following inequality:

 n/(n+1)
1 1
1− ≥ 1−
n+1 n
 1D. SERIES, THE NUMBER E 31

Now by raising to the power n + 1 we obtain from this:

 n+1  n
1 1
1− ≥ 1−
n+1 n
The point now is that we have the following inversion formulae:
 −1  −1
1 n n+1 1
1− = = =1+
n+1 n+1 n n
 −1  −1
1 n−1 n 1
1− = = =1+
n n n−1 n−1
Thus by inverting the inequality that we found, we obtain, as desired:
 n+1  n
1 1
1+ ≤ 1+
n n−1
(4) But with this, we can now finish. Indeed, the sequence xn is increasing, the
sequence yn is decreasing, and we have xn < yn , as well as:
yn 1
=1+ →1
xn n
Thus, both sequences xn , yn converge to a certain number e, as desired.
(5) Finally, regarding the numerics for our limiting number e, we know from the above
that we have xn < e < yn for any n ∈ N, which reads:
 n  n+1
1 1
1+ <e< 1+
n n
Thus e ∈ [2, 3], and with a bit of patience, or a computer, we obtain e = 2.71828 . . .
We will actually come back to this question later, with better methods. □
We should mention that there are many other ways of getting into e. For instance it
is possible to prove that we have the following formula, which is a bit more conceptual
than the formula in Theorem 1.34, and also with the convergence being very quick:
∞
X 1
=e
n=0
n!

Importantly, all this not the end of the story with e. For instance, in relation with
the first formula that we found, from Theorem 1.34, we have, more generally:
 x n
1+ → ex
n
32 1. SEQUENCES, SERIES

Also, in relation with the second formula, from above, we have, more generally:
∞
X xn
= ex
n=0
n!
To be more precise, these latter two formulae are something that we know at x = 1.
The case x = 0 is trivial, the case x = −1 follows from the case x = 1, via some simple
manipulations, and with a bit more work, we can get these formulae for any x ∈ N, and
then for any x ∈ Z. However, the general case x ∈ R is quite tricky, requiring a good
knowledge of the theory of real functions. And, good news, real functions will be what
we will be doing in the remainder of this first part, in chapters 2-4 below.
1e. Exercises
This opening chapter was a bit special, containing a lot of material in need to be
known, and compacted to the maximum. As exercises, again compacted, we have:
Exercise 1.35. Prove that the rational numbers r ∈ Q are exactly the real numbers
whose decimal expansion is periodic.
Exercise 1.36. Find geometric proofs, using triangles in the plane, for the well-known
formulae for sin(x + y) and cos(x + y).
Exercise 1.37. Develop some convergence theory for xn = an with a > 0, notably by
proving that an /nk → ∞ for any a > 1, and any k ∈ N.
x n
Exercise 1.38. Prove that ∞ 1


→ ex , and that
P
n=0 n! = e. Also, prove that 1 + n
P∞ xn x
n=0 n! = e , for x = −1, then for x ∈ Z, then for x ∈ R.

These exercises are probably quite difficult, unless you are already a bit familiar with
all this. If this is not the case, a good idea at this point is to pick a random entry-level
calculus book, and work out a few dozen exercises from there, as a warm-up.
 CHAPTER 2

Functions, continuity

2a. Continuous functions

We are now ready to talk about functions, which are the main topic of this book. A
function f : R → R is a correspondence x → f (x), which to each real number x ∈ R
associates a real number f (x) ∈ R. As examples, we have f (x) = x2 , f (x) = 2x and
so on. This suggests that any function f : R → R should be given by some kind of
“mathematical formula”, but unfortunately this is not correct, because, with suitable
definitions of course, there are more functions than mathematical formulae.

This being said, we will see that under suitable regularity assumptions on f : R → R,
we have indeed a mathematical formula for f (x) in terms of x, at least locally. And with
this being actually the main idea of calculus, that will take some time to be developed.
But more on this later, once we will know more about functions.

Getting started now, let us keep from the above discussion the idea that we should
focus our study on the functions f : R → R having suitable regularity properties. In what
regards these regularity properties, the most basic of them is continuity:
Definition 2.1. A function f : R → R, or more generally f : X → R, with X ⊂ R
being a subset, is called continuous when, for any xn , x ∈ X:
xn → x =⇒ f (xn ) → f (x)
Also, we say that f : X → R is continuous at a given point x ∈ X when the above
condition is satisfied, for that point x.
Observe that a function f : X → R is continuous precisely when it is continuous at
any point x ∈ X. We will see examples in a moment. Still speaking theory, there are
many equivalent formulations of the notion of continuity, with a well-known one, coming
by reminding in the above definition what convergence of a sequence means, twice, for
both the convergences xn → x and f (xn ) → f (x), being as follows:
∀x ∈ X, ∀ε > 0, ∃δ > 0, |x − y| < δ =⇒ |f (x) − f (y)| < ε
At the level of examples, basically all the functions that you know, including powers
x , exponentials ax , and more advanced functions like sin, cos, exp, log, are continuous.
a

However, proving this will take some time. Let us start with:
33
34 2. FUNCTIONS, CONTINUITY

Theorem 2.2. If f, g are continuous, then so are:

(1) f + g.
(2) f g.
(3) f /g.
(4) f ◦ g.
Proof. Before anything, we should mention that the claim is that (1-4) hold indeed,
provided that at the level of domains and ranges, the statement makes sense. For instance
in (1,2,3) we are talking about functions having the same domain, and with g(x) ̸= 0 for
the needs of (3), and there is a similar discussion regarding (4).
(1) The claim here is that if both f, g are continuous at a point x, then so is the sum
f + g. But this is clear from the similar result for sequences, namely:
lim (xn + yn ) = lim xn + lim yn
n→∞ n→∞ n→∞

(2) Again, the statement here is similar, and the result follows from:
lim xn yn = lim xn lim yn
n→∞ n→∞ n→∞

(3) Here the claim is that if both f, g are continuous at x, with g(x) ̸= 0, then f /g is
continuous at x. In order to prove this, observe that by continuity, g(x) ̸= 0 shows that
g(y) ̸= 0 for |x − y| small enough. Thus we can assume g ̸= 0, and with this assumption
made, the result follows from the similar result for sequences, namely:
lim xn /yn = lim xn / lim yn
n→∞ n→∞ n→∞

(4) Here the claim is that if g is continuous at x, and f is continuous at g(x), then
f ◦ g is continuous at x. But this is clear, coming from:
xn → x =⇒ g(xn ) → g(x)
=⇒ f (g(xn )) → f (g(x))
Alternatively, let us prove this as well by using that scary ε, δ condition given after
Definition 2.1. So, let us pick ε > 0. We want in the end to have something of type
|f (g(x)) − f (g(y))| < ε, so we must first use that ε, δ condition for the function f . So, let
us start in this way. Since f is continuous at g(x), we can find δ > 0 such that:
|g(x) − z| < δ =⇒ |f (g(x)) − f (z)| < ε
On the other hand, since g is continuous at x, we can find γ > 0 such that:
|x − y| < γ =⇒ |g(x) − g(y)| < δ
Now by combining the above two inequalities, with z = g(y), we obtain:
|x − y| < γ =⇒ |f (g(x)) − f (g(y))| < ε
Thus, the composition f ◦ g is continuous at x, as desired. □
 2A. CONTINUOUS FUNCTIONS 35

As a first comment, (3) shows in particular that 1/f is continuous, and we will use
this many times, in what follows. As a second comment, more philosophical, the proof of
(4) shows that the ε, δ formulation of continuity can be sometimes more complicated than
the usual formulation, with sequences, which leads us into the question of why bothering
at all with this ε, δ condition. Good question, and in answer:

(1) It is usually said that “for doing advanced math, you must use the ε, δ condition”,
but this is not exactly true, because sometimes what happens is that “for doing advanced
math, you must use open and closed sets”. With these sets, and the formulation of
continuity in terms of them, being something that we will discuss a bit later.

(2) This being said, the point is that the use of open and closed sets, technology that
we will discuss in a moment, requires some prior knowledge of the ε, δ condition. So, you
cannot really run away from this ε, δ condition, and want it or not, in order to do later
some more advanced mathematics, you’ll have to get used to that.

(3) But this should be fine, because you’re here since you love math and science, aren’t
you, and good math and science, including this ε, δ condition, will be what you will learn
from here. So, everything fine, more on this later, and in the meantime, no matter what
we do, always take a few seconds to think at what that means, in ε, δ terms.

Back to work now, at the level of examples, we have:

Theorem 2.3. The following functions are continuous:
(1) xn , with n ∈ Z.
(2) P/Q, with P, Q ∈ R[X].
(3) sin x, cos x, tan x, cot x.
Proof. This is a mixture of trivial and non-trivial results, as follows:
(1) Since f (x) = x is continuous, by using Theorem 2.2 we obtain the result for
exponents n ∈ N, and then for general exponents n ∈ Z too.
(2) The statement here, which generalizes (1), follows exactly as (1), by using the
various findings from Theorem 2.2.
(3) We must first prove here that xn → x implies sin xn → sin x, which in practice
amounts in proving that sin(x + y) ≃ sin x for y small. But this follows from:
sin(x + y) = sin x cos y + cos x sin y
To be more precise, let us first establish this formula. In order to do so, consider the
following picture, consisting of a length 1 line segment, with angles x, y drawn on each
36 2. FUNCTIONS, CONTINUITY

side, and with everything being completed, and lengths computed, as indicated:

1/ cos x
◦
sin x/ cos x
◦ 1
◦
sin y/ cos y
1/ cos y

Now let us compute the area of the big triangle, or rather the double of that area. We
can do this in two ways, either directly, with a formula involving sin(x + y), or by using
the two small triangles, involving functions of x, y. We obtain in this way:
1 1 sin x sin y
· · sin(x + y) = ·1+ ·1
cos x cos y cos x cos y
But this gives the formula for sin(x + y) claimed above.

(4) Now with this formula in hand, we can establish the continuity of sin x, as follows,
with the limits at 0 which are used being both clear on pictures:

lim sin(x + y) = lim (sin x cos y + cos x sin y)

y→0 y→0
= sin x lim cos y + cos x lim sin y
y→0 y→0
= sin x · 1 + cos x · 0
= sin x

(5) Moving ahead now with cos x, here the continuity follows from the continuity of
sin x, by using the following formula, which is obvious from definitions:
π 
cos x = sin −x
2
(6) Alternatively, and let us do this because we will need later the formula, by using
the formula for sin(x + y) we can deduce a formula for cos(x + y), as follows:
π 
cos(x + y) = sin −x−y
h2π  i
= sin − x + (−y)
 π2  π 
= sin − x cos(−y) + cos − x sin(−y)
2 2
= cos x cos y − sin x sin y
 2A. CONTINUOUS FUNCTIONS 37

But with this, we can use the same method as in (4), and we get, as desired:
lim cos(x + y) = lim (cos x cos y − sin x sin y)
y→0 y→0
= cos x lim cos y − sin x lim sin y
y→0 y→0
= cos x · 1 − sin x · 0
= cos x
(7) Finally, the fact that tan x, cot x are continuous is clear from the fact that sin x,
cos x are continuous, by using the result regarding quotients from Theorem 2.2. □
We will be back to more examples later, and in particular to functions of type xa and
x
a with a ∈ R, which are more tricky to define. Also, we will talk as well about inverse
functions f −1 , with as particular cases the basic inverse trigonometric functions, namely
arcsin, arccos, arctan, arccot, once we will have more tools for dealing with them.

Going ahead with more theory, some functions are “obviously” continuous:
Proposition 2.4. If a function f : X → R has the Lipschitz property
|f (x) − f (y)| ≤ K|x − y|
for some K > 0, then it is continuous.
Proof. This is indeed clear from our definition of continuity. □
Along the same lines, we can also argue, based on our intuition, that “some functions
are more continuous than other”. For instance, we have the following definition:
Definition 2.5. A function f : X → R is called uniformly continuous when:
∀ε > 0, ∃δ > 0, |x − y| < δ =⇒ |f (x) − f (y)| < ε
That is, f must be continuous at any x ∈ X, with the continuity being “uniform”.
As basic examples of uniformly continuous functions, we have the Lipschitz ones. Also,
as a basic counterexample, we have the following function:
f :R→R , f (x) = x2
Indeed, it is clear by looking at the graph of f that, the further our point x ∈ R is from
0, the smaller our δ > 0 must be, compared to ε > 0, in our ε, δ definition of continuity.
Thus, given an ε > 0, we have no δ > 0 doing the |x − y| < δ =⇒ |f (x) − f (y)| < ε job
at any x ∈ R, and so our function is indeed not uniformly continuous.

Quite remarkably, we have the following theorem, due to Heine and Cantor:
38 2. FUNCTIONS, CONTINUITY

Theorem 2.6. Any continuous function defined on a closed, bounded interval

f : [a, b] → R

is automatically uniformly continuous.

Proof. This is something quite subtle, and we are punching here a bit above our
weight, but here is the proof, with everything or almost included:

(1) Given ε > 0, for any x ∈ [a, b] we know that we have a δx > 0 such that:
ε
|x − y| < δx =⇒ |f (x) − f (y)| <
2
So, consider the following open intervals, centered at the various points x ∈ [a, b]:
 
δx δx
Ux = x − , x +
2 2
These intervals then obviously cover [a, b], in the sense that we have:
[
[a, b] ⊂ Ux
x∈[a,b]

Now assume that we managed to prove that this cover has a finite subcover. Then we
can most likely choose our δ > 0 to be the smallest of the δx > 0 involved, or perhaps
half of that, and then get our uniform continuity condition, via the triangle inequality.

(2) So, let us prove first that the cover in (1) has a finite subcover. For this purpose,
we proceed by contradiction. So, assume that [a, b] has no finite subcover, and let us cut
this interval in half. Then one of the halves must have no finite subcover either, and we
can repeat the procedure, by cutting this smaller interval in half. And so on. But this
leads to a contradiction, because the limiting point x ∈ [a, b] that we obtain in this way,
as the intersection of these smaller and smaller intervals, must be covered by something,
and so one of these small intervals leading to it must be covered too, contradiction.

(3) With
S this done, we are ready to finish, as announced in (1). Indeed, let us denote
by [a, b] ⊂ i Uxi the finite subcover found in (2), and let us set:
δx i
δ = min
i 2
Now assume |x − y| < δ, and pick i such that x ∈ Uxi . By the triangle inequality we
have then |xi − y| < δxi , which shows that we have y ∈ Uxi as well. But by applying now
f , this gives as desired |f (x) − f (y)| < ε, again via the triangle inequality. □
 2B. INTERMEDIATE VALUES 39

2b. Intermediate values

Moving ahead with more theory, we would like to explain now an alternative formula-
tion of the notion of continuity, which is quite abstract, and a bit difficult to understand
and master when you are a beginner, but which is definitely worth learning, because it is
quite powerful, solving some of the questions that we have left. Let us start with:
Definition 2.7. The open and closed sets are defined as follows:
(1) Open means that there is a small interval around each point.
(2) Closed means that our set is closed under taking limits.
As basic examples, the open intervals (a, b) are open, and the closed intervals [a, b] are
closed. Observe also that R itself is open and closed at the same time. Further examples,
or rather results which are easy to establish, include the fact that the finite unions or
intersections of open or closed sets are open or closed. We will be back to all this later,
with some precise results in this sense. For the moment, we will only need:
Proposition 2.8. A set O ⊂ R is open precisely when its complement C ⊂ R is
closed, and vice versa.
Proof. It is enough to prove the first assertion, since the “vice versa” part will follow
from it, by taking complements. But this can be done as follows:
“ =⇒ ” Assume that O ⊂ R is open, and let C = R − O. In order to prove that C is
closed, assume that {xn }n∈N ⊂ C converges to x ∈ R. We must prove that x ∈ C, and
we will do this by contradiction. So, assume x ∈/ C. Thus x ∈ O, and since O is open we
can find a small interval (x − ε, x + ε) ⊂ O. But since xn → x this shows that xn ∈ O for
n big enough, which contradicts xn ∈ C for all n, and we are done.
“⇐=” Assume that C ⊂ R is open, and let O = R − C. In order to prove that O is
open, let x ∈ O, and consider the intervals (x − 1/n, x + 1/n), with n ∈ N. If one of these
intervals lies in O, we are done. Otherwise, this would mean that for any n ∈ N we have
at least one point xn ∈ (x − 1/n, x + 1/n) satisfying xn ∈
/ O, and so xn ∈ C. But since C
is closed and xn → x, we get x ∈ C, and so x ∈ / O, contradiction, and we are done. □
As basic illustrations for the above result, R − (a, b) = (−∞, a] ∪ [b, ∞) is closed, and
R − [a, b] = (−∞, a) ∪ (b, ∞) is open. Getting now back to functions, we have:
Theorem 2.9. A function is continuous precisely when f −1 (O) is open, for any O
open. Equivalently, f −1 (C) must be closed, for any C closed.
Proof. Here the first assertion follows from definitions, and more specifically from
the ε, δ definition of continuity, which was as follows:
∀x ∈ X, ∀ε > 0, ∃δ > 0, |x − y| < δ =⇒ |f (x) − f (y)| < ε
40 2. FUNCTIONS, CONTINUITY

Indeed, if f satisfies this condition, it is clear that if O is open, then f −1 (O) is open,
and the converse holds too. As for the second assertion, this can be proved either directly,
by using the f (xn ) → f (x) definition of continuity, or by taking complements. □
As a test for the above criterion, let us reprove the fact, that we know from Theorem
2.2, that if f, g are continuous, so is f ◦ g. But this is clear, coming from:
(f ◦ g)−1 (O) = g −1 (f −1 (O))
In short, not bad, because at least in relation with this specific problem, our proof
using open sets is as simple as the simplest proof, namely the one using f (xn ) → f (x),
and is simpler than the other proof that we know, namely the one with ε, δ.

In order to reach to true applications of Theorem 2.9, we will need to know more
about the open and closed sets. Let us begin with a useful result, as follows:
Proposition 2.10. The following happen:
(1) Union of open sets is open.
(2) Intersection of closed sets is closed.
(3) Finite intersection of open sets is open.
(4) Finite union of closed sets is closed.
Proof. Here (1) is clear from definitions, (3) is clear from definitions too, and (2,4)
follow from (1,3) by taking complements E → E c , using the following formulae:
!c !c
[ \ \ [
Ei = Eic , Ei = Eic
i i i i
Thus, we are led to the conclusions in the statement. □
As an important comment, (3,4) above do not hold when removing the finiteness
assumption. Indeed, in what regards (3), the simplest counterexample here is:
\  1 1
− , = {0}
n∈N
n n
As for (4), here the simplest counterexample is as follows:
[ 1

0, 1 − = [0, 1)
n∈N
n
All this is quite interesting, and leads us to the question about what the open and
closed sets really are. And fortunately, this question can be answered, as follows:
Theorem 2.11. The open and closed sets are as follows:
(1) The open sets are the disjoint unions of open intervals.
(2) The closed sets are the complements of these unions.
 2B. INTERMEDIATE VALUES 41

Proof. We have two assertions to be proved, the idea being as follows:

(1) We know that the open intervals are those of type (a, b) with a < b, with the values
a, b = ±∞ allowed, and by Proposition 2.10 a union of such intervals is open.
(2) Conversely, given O ⊂ R open, we can cover each point x ∈ O with an open
interval Ix ⊂ O, and we have O = ∪x Ix , so O is a union of open intervals.
(3) In order to finish the proof of the first assertion, it remains to prove that the union
O = ∪x Ix in (2) can be taken to be disjoint. For this purpose, our first observation is
that, by approximating points x ∈ O by rationals y ∈ Q ∩ O, we can make our union to
be countable. But once our union is countable, we can start merging intervals, whenever
they meet, and we are left in the end with a countable, disjoint union, as desired.
(4) Finally, the second assertion comes from Proposition 2.8. □
The above result is quite interesting, philosophically speaking, because contrary to
what we have been doing so far, it makes the open sets appear quite different from the
closed sets. Indeed, there is no way of having a simple description of the closed sets
C ⊂ R, similar to the above simple description of the open sets O ⊂ R.
Moving towards more concrete things, and applications, let us formulate:
Definition 2.12. The compact and connected sets are defined as follows:
(1) Compact means that any open cover has a finite subcover.
(2) Connected means that it cannot be broken into two parts.
As basic examples, the closed bounded intervals [a, b] are compact, as we know from
the proof of Theorem 2.6, and so are the finite unions of such intervals. As for connected
sets, the basic examples here are the various types of intervals, namely (a, b), (a, b], [a, b),
[a, b], and it looks impossible to come up with more examples. In fact, we have:
Theorem 2.13. The compact and connected sets are as follows:
(1) The compact sets are those which are closed and bounded.
(2) The connected sets are the various types of intervals.
Proof. This is something quite intuitive, the idea being as follows:
(1) The fact that compact implies both closed and bounded is clear from our definition
of compactness, because assuming non-closedness or non-boundedness leads to an open
cover having no finite subcover. As for the converse, we know from the proof of Theorem
2.6 that any closed bounded interval [a, b] is compact, and it follows that any K ⊂ R
closed and bounded is a closed subset of a compact set, which follows to be compact.
(2) This is something which is obvious, and this regardless of what “cannot be bro-
ken into parts” in Definition 2.12 exactly means, mathematically speaking, with several
possible definitions being possible here, all being equivalent. Indeed, E ⊂ R having this
property is equivalent to a, b ∈ E =⇒ [a, b] ⊂ E, and this gives the result. □
42 2. FUNCTIONS, CONTINUITY

We will be back to all this later in this book, when looking at open, closed, compact
and connected sets in RN , or more general spaces, where things are more complicated
than in R. Now with this discussed, let us go back to continuous functions. We have:
Theorem 2.14. Assuming that f is continuous:
(1) If K is compact, then f (K) is compact.
(2) If E is connected, then f (E) is connected.
Proof. These assertions both follow from our definition of compactness and connect-
edness, as formulated in Definition 2.12. To be more precise:
(1) This comes from the fact that if a function f is continuous, then the inverse
function f −1 returns an open cover into an open cover.
(2) This is something clear as well, because if f (E) can be split into two parts, then
by applying f −1 we can split as well E into two parts. □
Let us record as well the following useful generalization of Theorem 2.6:
Theorem 2.15. Any continuous function defined on a compact set
f :X→R
is automatically uniformly continuous.
Proof. We can prove this exactly as Theorem 2.6, by using the compactness of X. □
You might perhaps ask at this point, were Theorems 2.14 and 2.15 worth all this
excursion into open and closed sets. Good point, and here is our answer, a beautiful and
powerful theorem based on the above, which can be used for a wide range of purposes:
Theorem 2.16. The following happen for a continuous function f : [a, b] → R:
(1) f takes all intermediate values between f (a), f (b).
(2) f has a minimum and maximum on [a, b].
(3) If f (a), f (b) have different signs, f (x) = 0 has a solution.
Proof. All these statements are related, and are called altogether “intermediate value
theorem”. Regarding now the proof, one way of viewing things is that since [a, b] is
compact and connected, the set f ([a, b]) is compact and connected too, and so it is a
certain closed bounded interval [c, d], and this gives all the results. However, this is based
on rather advanced technology, and it is possible to prove (1-3) directly as well. □
Along the same lines, we have as well the following result:
Theorem 2.17. Assuming that a function f is continuous and invertible, this function
must be monotone, and its inverse function f −1 must be monotone and continuous too.
Moreover, this statement holds both locally, and globally.
 2B. INTERMEDIATE VALUES 43

Proof. The fact that both f and f −1 are monotone follows from Theorem 2.16.
Regarding now the continuity of f −1 , we want to prove that we have:
xn → x =⇒ f −1 (xn ) → f −1 (x)
But with xn = f (yn ) and x = f (y), this condition becomes:
f (yn ) → f (y) =⇒ yn → y
And this latter condition being true since f is monotone, we are done. □
As a basic application of Theorem 2.17, we have:
Proposition 2.18. The various usual inverse functions, such as the inverse trigono-
metric functions arcsin, arccos, arctan, arccot, are all continuous.
Proof. This follows indeed from Theorem 2.17, with a course the full discussion
needing some explanations on bijectivity and domains. But you surely know all that, and
in what concerns us, our claim is simply that these beasts are all continuous, proved. □
As another basic application of this, we have:
Proposition 2.19. The following happen:
(1) Any polynomial P ∈ R[X] of odd degree
√ has a root.
(2) Given n ∈ 2N + 1, we can extract
√
n
x, for any x ∈ R.
(3) Given n ∈ N, we can extract n x, for any x ∈ [0, ∞).
Proof. All these results come as applications of Theorem 2.16, as follows:
(1) This is clear from Theorem 2.16 (3), applied on [−∞, ∞].
(2) This follows from (1), by using the polynomial P (z) = z n − x.
(3) This follows as well by applying Theorem 2.16 (3) to the polynomial P (z) = z n −x,
but this time on [0, ∞). □
There are many other things that can be said about roots of polyomials, and solutions
of other equations of type f (x) = 0, by using Theorem 2.16. We will be back to this.

As a concrete application, in relation with powers, we have the following result, com-
pleting our series of results regarding the basic mathematical functions:
Theorem 2.20. The function xa is defined and continuous on (0, ∞), for any a ∈ R.
Moreover, when trying to extend it to R, we have 4 cases, as follows,
(1) For a ∈ Qodd , a > 0, the maximal domain is R.
(2) For a ∈ Qodd , a ≤ 0, the maximal domain is R − {0}.
(3) For a ∈ R − Q or a ∈ Qeven , a > 0, the maximal domain is [0, ∞).
(4) For a ∈ R − Q or a ∈ Qeven , a ≤ 0, the maximal domain is (0, ∞).
where Qodd is the set of rationals r = p/q with q odd, and Qeven = Q − Qodd .
44 2. FUNCTIONS, CONTINUITY

Proof. The idea is that we know how to extract roots by using Proposition 2.19, and
all the rest follows by continuity. To be more precise:
(1) Assume a = p/q, with p, q ∈ N, p ̸= 0 and q odd. Given a number x ∈ R, we can
construct the power xa in the following way, by using Proposition 2.19:
√
xa = q xp
Then, it is straightforward to prove that xa is indeed continuous on R.
(2) In the case a = −p/q, with p, q ∈ N and q odd, the same discussion applies, with
the only change coming from the fact that xa cannot be applied to x = 0.
(3) Assume first a ∈ Qeven , a > 0. This√ means a = p/q with p, q ∈ N, p ̸= 0 and q
even, and as before in (1), we can set xa = q xp for x ≥ 0, by using Proposition 2.19. It is
then straightforward to prove that xa is indeed continuous on [0, ∞), and not extendable
either to the negatives. Thus, we are done with the case a ∈ Qeven , a > 0, and the case
left, namely a ∈ R − Q, a > 0, follows as well by continuity.
(4) In the cases a ∈ Qeven , a ≤ 0 and a ∈ R − Q, a ≤ 0, the same discussion applies,
with the only change coming from the fact that xa cannot be applied to x = 0. □
Let us record as well a result about the function ax , as follows:
Theorem 2.21. The function ax is as follows:
(1) For a > 0, this function is defined and continuous on R.
(2) For a = 0, this function is defined and continuous on (0, ∞).
(3) For a < 0, the domain of this function contains no interval.
Proof. This is a sort of reformulation of Theorem 2.20, by exchanging the variables,
x ↔ a. To be more precise, the situation is as follows:
(1) We know from Theorem 2.20 that things fine with xa for x > 0, no matter what
a ∈ R is. But this means that things fine with ax for a > 0, no matter what x ∈ R is.
(2) This is something trivial, and we have of course 0x = 0, for any x > 0. As for the
powers 0x with x ≤ 0, these are impossible to define, for obvious reasons.
(3) Given a < 0, we know from Theorem 2.20 that we cannot define ax for x ∈ Qeven .
But since Qeven is dense in R, this gives the result. □
Summarizing, we have been quite successful with our theory of continuous functions,
having how full results, regarding the definition and continuity property, for all basic
functions from mathematics. All this is of course just a beginning, and we will be back
to these functions on regular occasions, in what follows. In particular, we will discuss the
function ax at the special value a = e, and its inverse log x, at the end of this chapter.
 2C. SEQUENCES AND SERIES 45

2c. Sequences and series

Our goal now is to extend the material from chapter 1 regarding the numeric sequences
and series, to the case of the sequences and series of functions. To start with, we can talk
about the convergence of sequences of functions, fn → f , as follows:
Definition 2.22. We say that fn converges pointwise to f , and write fn → f , if
fn (x) → f (x)
for any x. Equivalently, ∀x, ∀ε > 0, ∃N ∈ N, ∀n ≥ N, |fn (x) − f (x)| < ε.
The question is now, assuming that fn are continuous, does it follow that f is contin-
uous? I am pretty much sure that you think that the answer is “yes”, based on:
lim f (y) = lim lim fn (y)
y→x y→x n→∞

= lim lim fn (y)

n→∞ y→x

= lim fn (x)
n→∞
= f (x)
However, this proof is wrong, because we know well from chapter 1 that we cannot
intervert limits, with this being a common beginner mistake. In fact, the result itself is
wrong in general, because if we consider the functions fn : [0, 1] → R given by fn (x) = xn ,
which are obviously continuous, their limit is discontinuous, given by:
(
0 , x ∈ [0, 1)
lim xn =
n→∞ 1 , x=1
Of course, you might say here that allowing x = 1 in all this might be a bit unnatural,
for whatever reasons, but there is an answer to this too. We can do worse, as follows:
Proposition 2.23. The basic step function, namely the sign function

−1 , x < 0

sgn(x) = 0 , x=0

1 , x>0
can be approximated by suitable modifications of arctan(x). Even worse, there are exam-
ples of fn → f with each fn continuous, and with f totally discontinuous.
Proof. To start with, arctan(x) looks a bit like sgn(x), so to say, but one problem
comes from the fact that its image is [−π/2, π/2], instead of the desired [−1, 1]. Thus,
we must first rescale arctan(x) by π/2. Now with this done, we can further stretch the
variable x, as to get our function closer and closer to sgn(x), as desired. This proves the
first assertion, and the second assertion, which is a bit more technical, and that we will
not really need in what follows, is left as an exercise for you, reader. □
46 2. FUNCTIONS, CONTINUITY

Sumarizing, we are a bit in trouble, because we would like to have in our bag of theo-
rems something saying that fn → f with fn continuous implies f continuous. Fortunately,
this can be done, with a suitable refinement of the notion of convergence, as follows:
Definition 2.24. We say that fn converges uniformly to f , and write fn →u f , if:
∀ε > 0, ∃N ∈ N, ∀n ≥ N, |fn (x) − f (x)| < ε, ∀x
That is, the same condition as for fn → f must be satisfied, but with the ∀x at the end.
And it is this “∀x at the end” which makes the difference, and will make our theory
work. In order to understand this, which is something quite subtle, let us compare
Definition 2.22 and Definition 2.24. As a first observation, we have:
Proposition 2.25. Uniform convergence implies pointwise convergence,
fn →u f =⇒ fn → f
but the converse is not true, in general.
Proof. Here the first assertion is clear from definitions, just by thinking at what
is going on, with no computations needed. As for the second assertion, the simplest
counterexamples here are the functions fn : [0, 1] → R given by fn (x) = xn , that we met
before in Proposition 2.23. Indeed, uniform convergence on [0, 1) would mean:
∀ε > 0, ∃N ∈ N, ∀n ≥ N, xn < ε, ∀x ∈ [0, 1)
But this is wrong, because no matter how big N is, we have limx→1 xN = 1, and so
we can find x ∈ [0, 1) such that xN > ε. Thus, we have our counterexample. □

Moving ahead now, let us state our main theorem on uniform convergence, as follows:
Theorem 2.26. Assuming that fn are continuous, and that
fn →u f
then f is continuous. That is, uniform limit of continuous functions is continuous.
Proof. As previously advertised, it is the “∀x at the end” in Definition 2.24 that will
make this work. Indeed, let us try to prove that the limit f is continuous at some point
x. For this, we pick a number ε > 0. Since fn →u f , we can find N ∈ N such that:
ε
|fN (z) − f (z)| < , ∀z
3
On the other hand, since fN is continuous at x, we can find δ > 0 such that:
ε
|x − y| < δ =⇒ |fN (x) − fN (y)| <
3
 2C. SEQUENCES AND SERIES 47

But with this, we are done. Indeed, for |x − y| < δ we have:

|f (x) − f (y)| ≤ |f (x) − fN (x)| + |fN (x) − fN (y)| + |fN (y) − f (y)|
ε ε ε
≤ + +
3 3 3
= ε
Thus, the limit function f is continuous at x, and we are done. □

Obviously, the notion of uniform convergence in Definition 2.24 is something quite

interesting, worth some more study. As a first result, we have:
Proposition 2.27. The following happen, regarding uniform limits:
(1) fn →u f, gn →u g imply fn + gn →u f + g.
(2) fn →u f, gn →u g imply fn gn →u f g.
(3) fn →u f, f ̸= 0 imply 1/fn →u 1/f .
(4) fn →u f, g continuous imply fn ◦ g →u f ◦ g.
(5) fn →u f, g continuous imply g ◦ fn →u g ◦ f .
Proof. All this is routine, exactly as for the results for numeric sequences from
chapter 1, that we know well, with no difficulties or tricks involved. □

Finally, there is some abstract mathematics to be done as well. Indeed, observe that
the notion of uniform convergence, as formulated in Definition 2.24, means that:
sup fn (x) − f (x) −→n→∞ 0
x

This suggests measuring the distance between functions via a supremum as above,
and in relation with this, we have the following result:
Theorem 2.28. The uniform convergence, fn →u f , means that we have fn → f with
respect to the following distance,
d(f, g) = sup f (x) − g(x)
x

which is indeed a distance function.

Proof. Here the fact that d is indeed a distance, in the sense that it satisfies all the
intuitive properties of a distance, including the triangle inequality, follows from definitions,
and the fact that the uniform convergence can be interpreted as above is clear as well. □

Finally, regarding the series, some general theory can be developed here as well, in
connection with the notion of uniform convergence, and in connection with the notion of
convergence radius. We will see applications of all this, in a moment.
48 2. FUNCTIONS, CONTINUITY

2d. Basic functions

With the above theory in hand, let us get now to interesting things, namely compu-
tations. Among others, because this is what a mathematician’s job is, doing all sorts of
computations. We will be mainly interested in the functions xa and ax , which remain
something quite mysterious. Regarding xa , we first have the following result:
Theorem 2.29. We have the generalized binomial formula
∞  
a
X a k
(1 + x) = x
k=0
k
with the generalized binomial coefficients being given by
 
a a(a − 1) . . . (a − k + 1)
=
k k!
valid for any exponent a ∈ Z, and any |x| < 1.
Proof. This is something quite tricky, the idea being as follows:
(1) For exponents a ∈ N, this is something that we know well from chapter 1, and
which is valid for any x ∈ R, coming from the usual binomial formula, namely:
n  
n
X n k
(1 + x) = x
k=0
k
(2) For the exponent a = −1 this is something that we know from chapter 1 too,
coming from the following formula, valid for any |x| < 1:
1
= 1 − x + x2 − x3 + . . .
1+x
Indeed, this is exactly our generalized binomial formula at a = −1, because:
 
−1 (−1)(−2) . . . (−k)
= = (−1)k
k k!
(3) Let us discuss now the general case a ∈ −N. With a = −n, and n ∈ N, the
generalized binomial coefficients are given by the following formula:
 
−n (−n)(−n − 1) . . . (−n − k + 1)
=
k k!
n(n + 1) . . . (n + k − 1)
= (−1)k
k!
(n + k − 1)!
= (−1)k
(n − 1)!k!
 
k n+k−1
= (−1)
n−1
 2D. BASIC FUNCTIONS 49

Thus, our generalized binomial formula at a = −n, and n ∈ N, reads:

∞  
1 k n+k−1
X
= (−1) tk
(1 + t)n k=0
n − 1
(4) In order to prove this formula, it is convenient to write it with −t instead of t, in
order to get rid of signs. The formula to be proved becomes:
∞  
1 X n+k−1 k
= t
(1 − t)n k=0
n−1
We prove this by recurrence on n. At n = 1 this formula definitely holds, as explained
in (2) above. So, assume that the formula holds at n ∈ N. We have then:
1 1 1
= ·
(1 − t)n+1 1 − t (1 − t)n
∞ ∞  
X
k
X n+l−1 l
= t t
k=0 l=0
n − 1
∞ s  
X
s
X n+l−1
= t
s=0 l=0
n−1
On the other hand, the formula that we want to prove is:
∞  
1 X n+s k
= t
(1 − t)n+1 s=0
n
Thus, in order to finish, we must prove the following formula:
s    
X n+l−1 n+s
=
l=0
n−1 n
(5) In order to prove this latter formula, we proceed by recurrence on s ∈ N. At s = 0
the formula is trivial, 1 = 1. So, assume that the formula holds at s ∈ N. In order to
prove the formula at s + 1, we are in need of the following formula:
     
n+s n+s n+s+1
+ =
n n−1 n
But this is the Pascal formula, that we know from chapter 1, and we are done. □
Quite interestingly, the formula in Theorem 2.29 holds in fact at any a ∈ R, but this
is something non-trivial, whose proof will have to wait until chapter 3 below. However,
in the meantime, let us investigate the case a ∈ Z/2. Indeed, not only the results here
are interesting, and very useful in practice, but also they can be proved with elementary
methods. At a = ±1/2, to start with, we have the following result:
50 2. FUNCTIONS, CONTINUITY

Theorem 2.30. The generalized binomial formula, namely

∞  
a
X a k
(1 + x) = x
k=0
k

holds as well at a = ±1/2. In practice, at a = 1/2 we obtain the formula

∞  k
√ X −t
1+t=1−2 Ck−1
k=1
4
1 2k


with Ck = k+1 k
being the Catalan numbers, and at a = −1/2 we obtain
∞  k
1 X −t
√ = Dk
1 + t k=0 4

with Dk = 2k


k
being the central binomial coefficients.
Proof. This can be done in several steps, as follows:
(1) At a = 1/2, the generalized binomial coefficients are as follows:
 
1/2 1/2(−1/2) . . . (3/2 − k)
=
k k!
1 · 3 · 5 . . . (2k − 3)
= (−1)k−1
2k k!
(2k − 2)!
= (−1)k−1 k−1
2 (k − 1)!2k k!
 k
−1
= −2 Ck−1
4
(2) At a = −1/2, the generalized binomial coefficients are as follows:
 
−1/2 −1/2(−3/2) . . . (1/2 − k)
=
k k!
1 · 3 · 5 . . . (2k − 1)
= (−1)k
2k k!
(2k)!
= (−1)k k k
2 k!2 k!
 k
−1
= Dk
4
(3) Summarizing, we have proved so far that the binomial formula at a = ±1/2 is
equivalent to the explicit formulae in the statement, involving the Catalan numbers Ck ,
 2D. BASIC FUNCTIONS 51

and the central binomial coefficients Dk . It remains now to prove that these two explicit
formulae hold indeed. For this purpose, let us write these formulae as follows:
∞ ∞
√ X
k 1 X
1 − 4t = 1 − 2 Ck−1 t , √ = Dk tk
k=1
1 − 4t k=0

In order to check these latter formulae, we must prove the following identities:
∞
!2 ∞
!2
X
k
X
k 1
1−2 Ck−1 t = 1 − 4t , Dk t =
k=1 k=0
1 − 4t
(4) As a first observation, the formula on the left is equivalent to:
X
Ck Cl = Cn+1
k+l=n

By using the series for 1/(1 − 4t), the formula on the right is equivalent to:
X
Dk Dl = 4n
k+l=n

Finally, observe that if our formulae hold indeed, by multiplying we must have:
X Dn+1
Ck Dl =
k+l=n
2

(5) Summarizing, wePhave to understand 3 formulae, which look quite similar. Let us
first attempt to prove k+l=n Dk Dl = 4n , by recurrence. We have:
     
2k + 2 4k + 2 2k 2
Dk+1 = = = 4− Dk
k+1 k+1 k k+1
Thus, assuming that we have k+l=n Dk Dl = 4n , we obtain:
P

X X  2

Dk Dl = D0 Dn+1 + 4− Dk Dl
k+l=n+1 k+l=n
k+1
X X Dk Dl
= Dn+1 + 4 Dk Dl − 2
k+l=n k+l=n
k+1
X
= Dn+1 + 4n+1 − 2 Ck Dl
k+l=n

Thus, this leads to a sort of half-failure, the conclusion being that for proving by
recurrence the second formula in (4), we need the third formula in (4).
(6) All this suggests a systematic look at the three formulae in (4). According to our
various observations above, these three formulae are equivalent, and so it is enough to
52 2. FUNCTIONS, CONTINUITY

prove one of them. We will chose here to prove the first one, namely:
X
Ck Cl = Cn+1
k+l=n

(7) For this purpose, we will trick. Let us count the Dyck paths in the plane, which
are by definition the paths from (0, 0) to (n, n), marching North-East over the integer
lattice Z2 ⊂ R2 , by staying inside the square [0, n] × [0, n], and staying as well under the
diagonal of this square. As an example, here are the 5 possible Dyck paths at n = 3:
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
In fact, the number Cn′ of these paths is as follows, coinciding with Cn :
1, 1, 2, 5, 14, 42, 132, 429, . . .
(8) So, here is our trick. We will prove on one hand that the numbers Cn′ satisfy the
recurrence for the numbers Cn that we want to prove, from (6), and on the other hand we
will prove that we have Cn′ = Cn . Which is smart, isn’t it. Getting to work now, in what
regards our first task, this is easy, because when looking at where our path last intersects
the diagonal of the square, we obtain the recurrence relation that we want, namely:
X
Cn′ = Ck′ Cl′
k+l=n−1

(9) In what regards now our second task, proving that we have Cn′ = Cn , this is more
tricky. If we ignore
the assumption that our path must stay under the diagonal of the
square, we have 2nn
such paths. And among these, we have the “good” ones, those that
we want to count, and then the “bad” ones, those that we want to ignore.
(10) So, let us count the bad paths, those crossing the diagonal of the square, and
reaching the higher diagonal next to it, the one joining (0, 1) and (n, n + 1). In order to
count these, the trick is to “flip” their bad part over that higher diagonal, as follows:
· · · · · ·
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦
 2D. BASIC FUNCTIONS 53

(11) Now observe that, as it is obvious on the above picture, due to the flipping, the
flipped bad path will no longer end in (n, n), but rather in (n − 1, n + 1). Moreover, more
is true, in the sense that, by thinking a bit, we see that the flipped bad paths are precisely
those ending in (n − 1, n + 1). Thus, we can count these flipped bad paths, and so the
bad paths, and so the good paths too, and so good news, we are done.
(12) To finish now, by putting everything together, we have:
   
′ 2n 2n
Cn = −
n n−1
   
2n n 2n
= −
n n+1 n
 
1 2n
=
n+1 n
′
Thus we have indeed Cn = Cn , and this finishes the proof. □
As already mentioned, the binomial formula holds in fact for any exponent a ∈ Z/2,
after some combinatorial pain, and even for any a ∈ R, but this is non-trivial, and the
elementary study stops with Theorem 2.30. However, we will see in chapter 3 below that
the problem can be solved, and in a very elegant way, by guess whom: calculus.
As another application of our methods, let us get now into the other version of the
exponential function, namely ax . The idea is that some very interesting results appear
with a = e, the number that we know from chapter 1. We first have:
Proposition 2.31. We have the following formula,
 x n
1+ → ex
n
valid for any x ∈ R.
Proof. We already know from chapter 1 that the result holds at x = 1, and this
because the number e was by definition given by the following formula:
 n
1
1+ →e
n
By taking inverses, we obtain as well the result at x = −1, namely:
 n
1 1
1− →
n e
In general now, when ∈ R is arbitrary, the best is to proceed as follows:
 x
 x n x n/x
1+ = 1+ → ex
n n
Thus, we are led to the conclusion in the statement. □
54 2. FUNCTIONS, CONTINUITY

Next, we have the following result, which is something quite far-reaching:

Theorem 2.32. We have the formula

∞
X xk
ex =
k=0
k!

valid for any x ∈ R.

Proof. This can be done in several steps, as follows:

(1) At x = 1, which is the key step, we want to prove that we have the following
equality, between the sum of a series, and a limit of a sequence:

∞  n
X 1 1
= lim 1 +
k=0
k! n→∞ n

(2) For this purpose, the first observation is that we have the following estimate:

∞ ∞
X 1 X 1
2< < k−1
=3
k=0
k! k=0 2

P∞ 1
Thus, the series k=0 k! converges indeed, towards a limit in (2, 3).

(3) In order to prove now that this limit is e, observe that we have:

 n n  
1 X n 1
1+ = · k
n k=0
k n
n
X n(n − 1) . . . (n − k + 1) 1
= ·
k=0
k! nk
n
X 1
≤
k=0
k!

P∞ 1
Thus, with n → ∞, we get that the limit of the series k=0 k! belongs to [e, 3).
 2D. BASIC FUNCTIONS 55

(4) For the reverse inequality, we use the following computation:

n  n n n
X 1 1 X 1 X n(n − 1) . . . (n − k + 1) 1
− 1+ = − · k
k=0
k! n k=0
k! k=0
k! n
n n
X 1 X n(n − 1) . . . (n − k + 1) 1
= − · k
k=2
k! k=2
k! n
n
X nk − n(n − 1) . . . (n − k + 1)
=
k=2
nk k!
n
X nk − (n − k)k
≤
k=2
nk k!
n
k
X 1 − 1 − nk
=
k=2
k!

(5) In order to estimate the above expression that we found, we can use the following
trivial inequality, valid for any number x ∈ (0, 1):

1 − xk = (1 − x)(1 + x + x2 + . . . + xk−1 ) ≤ (1 − x)k

Indeed, we can use this with x = 1 − k/n, and we obtain in this way:
n n n k
·k

X 1 1 X
n
− 1+ ≤
k=0
k! n k=2
k!
n
1X k
=
n k=2 (k − 1)!
n
1X k 1
= ·
n k=2 k − 1 (k − 2)!
n
1X 2
≤
n k=2 2k−2
4
<
n

that the limit of the series ∞ 1

P
Now since with n → ∞ this goes to 0, we obtain k=0 k!
n
is the same as the limit of the sequence 1 + n1 , manely e. Thus, getting back now to


what we wanted to prove, our theorem, we are done in this way with the case x = 1.
56 2. FUNCTIONS, CONTINUITY

(6) In order to deal now with the general case, consider the following function:
∞
X xk
f (x) =
k=0
k!
Observe that, by using our various results above, this function is indeed well-defined.
Moreover, again by using our various results above, f is continuous.
(7) Our next claim, which is the key one, is that we have:
f (x + y) = f (x)f (y)
Indeed, by using the binomial formula, we have the following computation:
∞
X (x + y)k
f (x + y) =
k=0
k!
∞ X k  
X k xs y k−s
= ·
k=0 s=0
s k!
∞ X k
X xs y k−s
=
k=0 s=0
s!(k − s)!
= f (x)f (y)
(8) In order to finish now, we know that our function f is continuous, that it satisfies
f (x + y) = f (x)f (y), and that we have:
f (0) = 1 , f (1) = e
But it is easy to prove that such a function is necessarily unique, and since ex obviously
has all these properties too, we must have f (x) = ex , as desired. □
We will be back to all this, and to the logarithm and trigonometric functions as well,
in chapter 3 below, when talking about derivatives and the Taylor formula.
2e. Exercises
There are many possible exercises on the above, and we have here:
Exercise 2.33. Find some interesting examples of Lipschitz functions.
Exercise 2.34. Find a simple proof for the intermediate value theorem.
Exercise 2.35. Prove the generalized binomial formula, for exponents a ∈ Z/2.
P
Exercise 2.36. Rewrite the theory of e, with e = k 1/k! as definition.
For the rest, business as usual, more exercises are easy to find. Find and solve them.
 CHAPTER 3

Derivatives

3a. Derivatives, rules

Welcome to calculus. In this chapter we go for the real thing, namely development
of modern calculus, following some amazing ideas of Newton, Leibnitz and others. The
material will be quite difficult, mixing geometry and intuition with formal mathematics
and computations, and needing some time to be understood. But we will survive.

The basic idea of calculus is very simple. We are interested in functions f : R → R, and
we already know that when f is continuous at a point x, we can write an approximation
formula as follows, for the values of our function f around that point x:
f (x + t) ≃ f (x)
The problem is now, how to improve this? And a bit of thinking at all this suggests
to look at the slope of f at the point x. Which leads us into the following notion:
Definition 3.1. A function f : R → R is called differentiable at x when
f (x + t) − f (x)
f ′ (x) = lim
t→0 t
called derivative of f at that point x, exists.
As a first remark, in order for f to be differentiable at x, that is to say, in order for
the above limit to converge, the numerator must go to 0, as the denominator t does:
lim [f (x + t) − f (x)] = 0
t→0
Thus, f must be continuous at x. However, the converse is not true, a basic coun-
terexample being f (x) = |x| at x = 0. Let us summarize these findings as follows:
Proposition 3.2. If f is differentiable at x, then f must be continuous at x. However,
the converse is not true, a basic counterexample being f (x) = |x|, at x = 0.
Proof. The first assertion is something that we already know, from the above. As
for the second assertion, regarding f (x) = |x|, this is something quite clear on the picture
of f , but let us prove this mathematically, based on Definition 3.1. We have:
|0 + t| − |0| t−0
lim = lim =1
t↘0 t t↘0 t
57
58 3. DERIVATIVES

On the other hand, we have as well the following computation:

|0 + t| − |0| −t − 0
lim = lim = −1
t↗0 t t↗0 t
Thus, the limit in Definition 3.1 does not converge, so we have our counterexample. □
Generally speaking, the last assertion in Proposition 3.2 should not bother us much,
because most of the basic continuous functions are differentiable, and we will see examples
in a moment. Before that, however, let us recall why we are here, namely improving the
basic estimate f (x + t) ≃ f (x). We can now do this, using the derivative, as follows:
Theorem 3.3. Assuming that f is differentiable at x, we have:
f (x + t) ≃ f (x) + f ′ (x)t
In other words, f is, approximately, locally affine at x.
Proof. Assume indeed that f is differentiable at x, and let us set, as before:
f (x + t) − f (x)
f ′ (x) = lim
t→0 t
By multiplying by t, we obtain that we have, once again in the t → 0 limit:
f (x + t) − f (x) ≃ f ′ (x)t
Thus, we are led to the conclusion in the statement. □
All this is very nice, and before developing more theory, let us work out some examples.
As a first illustration, the derivatives of the power functions are as follows:
Proposition 3.4. We have the differentiation formula
(xp )′ = pxp−1
valid for any exponent p ∈ R.
Proof. We can do this in three steps, as follows:
(1) In the case p ∈ N we can use the binomial formula, which gives, as desired:
n  
p
X p p−k k
(x + t) = x t
k=0
k
= xp + pxp−1 t + . . . + tp
≃ xp + pxp−1 t
(2) Let us discuss now the general case p ∈ Q. We write p = m/n, with m ∈ Z and
n ∈ N. In order to do the computation, we use the following formula:
an − bn = (a − b)(an−1 + an−2 b + . . . + bn−1 )
 3A. DERIVATIVES, RULES 59

We set in this formula a = (x + t)m/n and b = xm/n . We obtain, as desired:

(x + t)m − xm
(x + t)m/n − xm/n =
(x + t)m(n−1)/n + . . . + xm(n−1)/n
(x + t)m − xm
≃
nxm(n−1)/n
mxm−1 t
≃
nxm(n−1)/n
m m−1−m+m/n
= ·x ·t
n
m m/n−1
= ·x ·t
n
(3) In the general case now, where p ∈ R is real, we can use a similar argument.
Indeed, given any integer n ∈ N, we have the following computation:
(x + t)pn − xpn
(x + t)p − xp =
(x + t)p(n−1) + . . . + xp(n−1)
(x + t)pn − xpn
≃
nxp(n−1)
Now observe that we have the following estimate, with [.] being the integer part:
(x + t)[pn] ≤ (x + t)pn ≤ (x + t)[pn]+1
By using the binomial formula on both sides, for the integer exponents [pn] and [pn]+1
there, we deduce that with n >> 0 we have the following estimate:
(x + t)pn ≃ xpn + pnxpn−1 t
Thus, we can finish our computation started above as follows:
pnxpn−1 t
(x + t)p − xp ≃ = pxp−1 t
nxpn−p
But this gives (xp )′ = pxp−1 , which finishes the proof. □
Here are some further computations, for other basic functions that we know:
Proposition 3.5. We have the following results:
(1) (sin x)′ = cos x.
(2) (cos x)′ = − sin x.
(3) (ex )′ = ex .
(4) (log x)′ = x−1 .
Proof. This is quite tricky, as always when computing derivatives, as follows:
60 3. DERIVATIVES

(1) Regarding sin, the computation here goes as follows:

sin(x + t) − sin x
(sin x)′ = lim
t→0 t
sin x cos t + cos x sin t − sin x
= lim
t→0 t
cos t − 1 sin t
= lim sin x · + cos x ·
t→0 t t
= cos x
Here we have used the fact, which is clear on pictures, by drawing the trigonometric
circle, that we have sin t ≃ t for t ≃ 0, plus the fact, which follows from this and from
Pythagoras, sin2 + cos2 = 1, that we have as well cos t ≃ 1 − t2 /2, for t ≃ 0.
(2) The computation for cos is similar, as follows:
cos(x + t) − cos x
(cos x)′ = lim
t→0 t
cos x cos t − sin x sin t − cos x
= lim
t→0 t
cos t − 1 sin t
= lim cos x · − sin x ·
t→0 t t
= − sin x
(3) For the exponential, the derivative can be computed as follows:
∞
!′
k
X x
(ex )′ =
k=0
k!
∞
X kxk−1
=
k=0
k!
x
= e
(4) As for the logarithm, the computation here is as follows, using log(1 + y) ≃ y for
y ≃ 0, which follows from ey ≃ 1 + y that we found in (3), by taking the logarithm:
log(x + t) − log x
(log x)′ = lim
t→0 t
log(1 + t/x)
= lim
t→0 t
1
=
x
Thus, we are led to the formulae in the statement. □
Speaking exponentials, we can now formulate a nice result about them:
 3A. DERIVATIVES, RULES 61

Theorem 3.6. The exponential function, namely

∞
x
X xk
e =
k=0
k!

is the unique power series satisfying f ′ = f and f (0) = 1.

Proof. Consider indeed a power series satisfying f ′ = f and f (0) = 1. Due to

f (0) = 1, the first term must be 1, and so our function must look as follows:
∞
X
f (x) = 1 + ck x k
k=1

According to our differentiation rules, the derivative of this series is given by:
∞
X
f (x) = kck xk−1
k=1

Thus, the equation f ′ = f is equivalent to the following equalities:

c1 = 1 , 2c2 = c1 , 3c3 = c2 , 4c4 = c3 , ...

But this system of equations can be solved by recurrence, as follows:

1 1 1
c1 = 1 , c2 = , c3 = , c4 = , ...
2 2×3 2×3×4
Thus we have ck = 1/k!, leading to the conclusion in the statement. □

Observe that the above result leads to a more conceptual explanation for the number
e itself. To be more precise, e ∈ R is the unique number satisfying:

(ex )′ = ex

Let us work out now some general results. We have here the following statement:

Theorem 3.7. We have the following formulae:

(1) (f + g)′ = f ′ + g ′ .
(2) (f g)′ = f ′ g + f g ′ .
(3) (f ◦ g)′ = (f ′ ◦ g) · g ′ .

Proof. All these formulae are elementary, the idea being as follows:
62 3. DERIVATIVES

(1) This follows indeed from definitions, the computation being as follows:
(f + g)(x + t) − (f + g)(x)
(f + g)′ (x) = lim
t→0
 t 
f (x + t) − f (x) g(x + t) − g(x)
= lim +
t→0 t t
f (x + t) − f (x) g(x + t) − g(x)
= lim + lim
t→0 t t→0 t
= f ′ (x) + g ′ (x)
(2) This follows from definitions too, the computation, by using the more convenient
formula f (x + t) ≃ f (x) + f ′ (x)t as a definition for the derivative, being as follows:
(f g)(x + t) = f (x + t)g(x + t)
≃ (f (x) + f ′ (x)t)(g(x) + g ′ (x)t)
≃ f (x)g(x) + (f ′ (x)g(x) + f (x)g ′ (x))t
Indeed, we obtain from this that the derivative is the coefficient of t, namely:
(f g)′ (x) = f ′ (x)g(x) + f (x)g ′ (x)
(3) Regarding compositions, the computation here is as follows, again by using the
more convenient formula f (x + t) ≃ f (x) + f ′ (x)t as a definition for the derivative:
(f ◦ g)(x + t) = f (g(x + t))
≃ f (g(x) + g ′ (x)t)
≃ f (g(x)) + f ′ (g(x))g ′ (x)t
Indeed, we obtain from this that the derivative is the coefficient of t, namely:
(f ◦ g)′ (x) = f ′ (g(x))g ′ (x)
Thus, we are led to the conclusions in the statement. □
We can of course combine the above formulae, and we obtain for instance:
Proposition 3.8. The derivatives of fractions are given by:
 ′
f f ′g − f g′
=
g g2
In particular, we have the following formula, for the derivative of inverses:
 ′
1 f′
=− 2
f f
In fact, we have (f p )′ = pf p−1 , for any exponent p ∈ R.
 3A. DERIVATIVES, RULES 63

Proof. This statement is written a bit upside down, and for the proof it is better to
proceed backwards. To be more precise, by using (xp )′ = pxp−1 and Theorem 3.7 (3), we
obtain the third formula. Then, with p = −1, we obtain from this the second formula.
And finally, by using this second formula and Theorem 3.7 (2), we obtain:
 ′  ′
f 1
= f·
g g
 ′
′ 1 1
= f · +f
g g
′ ′
f fg
= − 2
g g
f g − f g′
′
=
g2
Thus, we are led to the formulae in the statement. □
All the above might seem to start to be a bit too complex, with too many things to
be memorized and so on, and as a piece of advice here, we have:
Advice 3.9. Memorize and cherish the formula for fractions
 ′
f f ′g − f g′
=
g g2
along with the usual addition formula, that you know well
a c ad + bc
+ =
b d bd
and generally speaking, never mess with fractions.
With this coming from a lifelong calculus teacher and scientist, mathematics can be
difficult, and many things can be pardoned, but not messing with fractions. And with
this going beyond mathematics too, say if you want to make a living by selling apples or
tomatoes at the market, fine, but you’ll need to know well fractions, trust me.

Back to work now, with the above formulae in hand, we can do all sorts of computations
for other basic functions that we know, including tan x, or arctan x:
Proposition 3.10. We have the following formulae,
1 1
(tan x)′ = , (arctan x)′ =
cos2 x 1 + x2
and the derivatives of the remaining trigonometric functions can be computed as well.
64 3. DERIVATIVES

Proof. For tan, we have the following computation:

 ′
′ sin x
(tan x) =
cos x
sin′ x cos x − sin x cos′ x
=
cos2 x
cos2 x + sin2 x
=
cos2 x
1
=
cos2 x
As for arctan, we can use here the following computation:
(tan ◦ arctan)′ (x) = tan′ (arctan x) arctan′ (x)
1
= 2
arctan′ (x)
cos (arctan x)
Indeed, since the term on the left is simply x′ = 1, we obtain from this:
arctan′ (x) = cos2 (arctan x)
On the other hand, with t = arctan x we know that we have tan t = x, and so:
1 1
cos2 (arctan x) = cos2 t = 2
=
1 + tan t 1 + x2
Thus, we are led to the formula in the statement, namely:
1
(arctan x)′ =
1 + x2
As for the last assertion, we will leave this as an exercise. □
At the theoretical level now, further building on Theorem 3.3, we have:
Theorem 3.11. The local minima and maxima of a differentiable function f : R → R
appear at the points x ∈ R where:
f ′ (x) = 0
However, the converse of this fact is not true in general.
Proof. The first assertion follows from the formula in Theorem 3.3, namely:
f (x + t) ≃ f (x) + f ′ (x)t
Indeed, let us rewrite this formula, more conveniently, in the following way:
f (x + t) − f (x) ≃ f ′ (x)t
Now saying that our function f has a local maximum at x ∈ R means that there exists
a number ε > 0 such that the following happens:
f (x + t) ≥ f (x) , ∀t ∈ [−ε, ε]
 3A. DERIVATIVES, RULES 65

We conclude that we must have f ′ (x)t ≥ 0 for sufficiently small t, and since this small
t can be both positive or negative, this gives, as desired:
f ′ (x) = 0
Similarly, saying that our function f has a local minimum at x ∈ R means that there
exists a number ε > 0 such that the following happens:
f (x + t) ≤ f (x) , ∀t ∈ [−ε, ε]
Thus f ′ (x)t ≤ 0 for small t, and this gives, as before, f ′ (x) = 0. Finally, in what
regards the converse, the simplest counterexample here is the following function:
f (x) = x3
Indeed, we have f ′ (x) = 3x2 , and in particular f ′ (0) = 0. But our function being
clearly increasing, x = 0 is not a local maximum, nor a local minimum. □
As an important consequence of Theorem 3.11, we have:
Theorem 3.12. Assuming that f : [a, b] → R is differentiable, we have
f (b) − f (a)
= f ′ (c)
b−a
for some c ∈ (a, b), called mean value property of f .
Proof. In the case f (a) = f (b), the result, called Rolle theorem, states that we have
f ′ (c) = 0 for some c ∈ (a, b), and follows from Theorem 3.11. Now in what regards our
statement, due to Lagrange, this follows from Rolle, applied to the following function:
f (b) − f (a)
g(x) = f (x) − ·x
b−a
Indeed, we have g(a) = g(b), due to our choice of the constant on the right, so we get
g ′ (c) = 0 for some c ∈ (a, b), which translates into the formula in the statement. □
In practice, Theorem 3.11 can be used in order to find the maximum and minimum
of any differentiable function, and this method is best recalled as follows:
Algorithm 3.13. In order to find the minimum and maximum of f : [a, b] → R:
(1) Compute the derivative f ′ .
(2) Solve the equation f ′ (x) = 0.
(3) Add a, b to your set of solutions.
(4) Compute f (x), for all your solutions.
(5) Compute the min/max of all these f (x) values.
(6) Then this is the min/max of your function.
66 3. DERIVATIVES

To be more precise, we are using here Theorem 3.11, or rather the obvious extension
of this result to the case of the functions f : [a, b] → R. This tells us that the local minima
and maxima of our function f , and in particular the global minima and maxima, can be
found among the zeroes of the first derivative f ′ , with the endpoints a, b added. Thus,
what we have to do is to compute these “candidates”, as explained in steps (1-2-3), and
then see what each candidate is exactly worth, as explained in steps (4-5-6).

Needless to say, all this is very interesting, and powerful. The general problem in
any type of applied mathematics is that of finding the minimum or maximum of some
function, and we have now an algorithm for dealing with such questions. Very nice.

3b. Second derivatives

The derivative theory that we have is already quite powerful, and can be used in order
to solve all sorts of interesting questions, but with a bit more effort, we can do better.
Indeed, at a more advanced level, we can come up with the following notion:
Definition 3.14. We say that f : R → R is twice differentiable if it is differentiable,
and its derivative f ′ : R → R is differentiable too. The derivative of f ′ is denoted
f ′′ : R → R
and is called second derivative of f .
You might probably wonder why coming with this definition, which looks a bit abstract
and complicated, instead of further developing the theory of the first derivative, which
looks like something very reasonable and useful. Good point, and answer to this coming
in a moment. But before that, let us get a bit familiar with f ′′ . We first have:
Interpretation 3.15. The second derivative f ′′ (x) ∈ R is the number which:
(1) Expresses the growth rate of the slope f ′ (z) at the point x.
(2) Gives us the acceleration of the function f at the point x.
(3) Computes how much different is f (x), compared to f (z) with z ≃ x.
(4) Tells us how much convex or concave is f , around the point x.
So, this is the truth about the second derivative, making it clear that what we have
here is a very interesting notion. In practice now, (1) follows from the usual interpretation
of the derivative, as both a growth rate, and a slope. Regarding (2), this is some sort
of reformulation of (1), using the intuitive meaning of the word “acceleration”, with the
relevant physics equations, due to Newton, being as follows:
v = ẋ , a = v̇
To be more precise, here x, v, a are the position, speed and acceleration, and the dot
denotes the time derivative, and according to these equations, we have a = ẍ, second
derivative. We will be back to these equations at the end of the present chapter.
 3B. SECOND DERIVATIVES 67

Regarding now (3) in the above, this is something more subtle, of statistical nature,
that we will clarify with some mathematics, in a moment. As for (4), this is something
quite subtle too, that we will again clarify with some mathematics, in a moment.

All in all, what we have above is a mixture of trivial and non-trivial facts, and do not
worry, we will get familiar with all this, in the next few pages.

In practice now, let us first compute the second derivatives of the functions that we
are familiar with, see what we get. The result here, which is perhaps not very enlightening
at this stage of things, but which certainly looks technically useful, is as follows:
Proposition 3.16. The second derivatives of the basic functions are as follows:
(1) (xp )′′ = p(p − 1)xp−2 .
(2) sin′′ = − sin.
(3) cos′′ = − cos.
(4) exp′ = exp.
(5) log′ (x) = −1/x2 .
Also, there are functions which are differentiable, but not twice differentiable.
Proof. We have several assertions here, the idea being as follows:
(1) Regarding the various formulae in the statement, these all follow from the various
formulae for the derivatives established before, as follows:
(xp )′′ = (pxp−1 )′ = p(p − 1)xp−2
(sin x)′′ = (cos x)′ = − sin x
(cos x)′′ = (− sin x)′ = − cos x
(ex )′′ = (ex )′ = ex
(log x)′′ = (−1/x)′ = −1/x2
Of course, this is not the end of the story, because these formulae remain quite opaque,
and must be examined in view of Interpretation 3.15, in order to see what exactly is going
on. Also, we have tan and the inverse trigonometric functions too. In short, plenty of
good exercises here, for you, and the more you solve, the better your calculus will be.
(2) Regarding now the counterexample, recall first that the simplest example of a
function which is continuous, but not differentiable, was f (x) = |x|, the idea behind this
being to use a “piecewise linear function whose branches do not fit well”. In connection
now with our question, piecewise linear will not do, but we can use a similar idea, namely
“piecewise quadratic function whose branches do not fit well”. So, let us set:
(
ax2 (x ≤ 0)
f (x) =
bx2 (x ≥ 0)
68 3. DERIVATIVES

This function is then differentiable, with its derivative being:

(
2ax (x ≤ 0)
f ′ (x) =
2bx (x ≥ 0)
Now for getting our counterexample, we can set a = −1, b = 1, so that f is:
(
−x2 (x ≤ 0)
f (x) =
x2 (x ≥ 0)

Indeed, the derivative is f ′ (x) = 2|x|, which is not differentiable, as desired. □

Getting now to theory, we first have the following key result:
Theorem 3.17. Any twice differentiable function f : R → R is locally quadratic,
f ′′ (x) 2
f (x + t) ≃ f (x) + f ′ (x)t + t
2
with f ′′ (x) being as usual the derivative of the function f ′ : R → R at the point x.
Proof. Assume indeed that f is twice differentiable at x, and let us try to construct
an approximation of f around x by a quadratic function, as follows:
f (x + t) ≃ a + bt + ct2
We must have a = f (x), and we also know from Theorem 3.3 that b = f ′ (x) is the
correct choice for the coefficient of t. Thus, our approximation must be as follows:
f (x + t) ≃ f (x) + f ′ (x)t + ct2
In order to find the correct choice for c ∈ R, observe that the function t → f (x + t)
matches with t → f (x) + f ′ (x)t + ct2 in what regards the value at t = 0, and also in what
regards the value of the derivative at t = 0. Thus, the correct choice of c ∈ R should be
the one making match the second derivatives at t = 0, and this gives:
f ′′ (x) = 2c
We are therefore led to the formula in the statement, namely:
f ′′ (x) 2
f (x + t) ≃ f (x) + f ′ (x)t +t
2
In order to prove now that this formula holds indeed, we will use L’Hôpital’s rule,
which states that the 0/0 type limits can be computed as follows:
f (x) f ′ (x)
≃ ′
g(x) g (x)
 3B. SECOND DERIVATIVES 69

Observe that this formula holds indeed, as an application of Theorem 3.3. Now by
using this, if we denote by φ(t) ≃ P (t) the formula to be proved, we have:
φ(t) − P (t) φ′ (t) − P ′ (t)
≃
t2 2t
φ′′ (t) − P ′′ (t)
≃
2
f ′′ (x) − f ′′ (x)
=
2
= 0
Thus, we are led to the conclusion in the statement. □
The above result substantially improves Theorem 3.3, and there are many applications
of it. As a first such application, justifying Interpretation 3.15 (3), we have the following
statement, which is a bit heuristic, but we will call it however Proposition:
Proposition 3.18. Intuitively speaking, the second derivative f ′′ (x) ∈ R computes
how much different is f (x), compared to the average of f (z), with z ≃ x.
Proof. As already mentioned, this is something a bit heuristic, but which is good to
know. Let us write the formula in Theorem 3.17, as such, and with t → −t too:
f ′′ (x) 2
f (x + t) ≃ f (x) + f ′ (x)t + t
2
f ′′ (x) 2
f (x − t) ≃ f (x) − f ′ (x)t + t
2
By making the average, we obtain the following formula:
f (x + t) + f (x − t) f ′′ (x) 2
= f (x) + t
2 2
Now assume that we have found a way of averaging things over t ∈ [−ε, ε], with the
corresponding averages being denoted I. We obtain from the above:
 2
′′ t
I(f ) = f (x) + f (x)I
2
But this is what our statement says, save for some uncertainties regarding the aver-
aging method, and the precise value of I(t2 /2). We will leave this for later. □
Back to rigorous mathematics now, and of course with apologies for the physics inter-
mezzo, but Proposition 3.18 is really cool isn’t it, and we will be back later to this with
full mathematical details, after developing more theory, that is promised, as a second
application of Theorem 3.17, we can improve as well Theorem 3.11, as follows:
70 3. DERIVATIVES

Theorem 3.19. The local minima and local maxima of a twice differentiable function
f : R → R appear at the points x ∈ R where
f ′ (x) = 0
with the local minima corresponding to the case f ′ (x) ≥ 0, and with the local maxima
corresponding to the case f ′′ (x) ≤ 0.
Proof. The first assertion is something that we already know. As for the second
assertion, we can use the formula in Theorem 3.17, which in the case f ′ (x) = 0 reads:
f ′′ (x) 2
f (x + t) ≃ f (x) + t
2
Indeed, assuming f ′′ (x) ̸= 0, it is clear that the condition f ′′ (x) > 0 will produce a
local minimum, and that the condition f ′′ (x) < 0 will produce a local maximum. □
As before with Theorem 3.11, the above result is not the end of the story with the
mathematics of the local minima and maxima, because things are undetermined when:
f ′ (x) = f ′′ (x) = 0
For instance the functions ±xn with n ∈ N all satisfy this condition at x = 0, which
is a minimum for the functions of type x2m , a maximum for the functions of type −x2m ,
and not a local minimum or local maximum for the functions of type ±x2m+1 .

There are some comments to be made in relation with Algorithm 3.13 as well. Nor-
mally that algorithm stays strong, because Theorem 3.19 can only help in relation with
the final steps, and is it worth it to compute the second derivative f ′′ , just for getting rid
of roughly 1/2 of the f (x) values to be compared. However, in certain cases, this method
proves to be useful, so Theorem 3.19 is good to know, when applying that algorithm.

As a main concrete application now of the second derivative, which is something very
useful in practice, and related to Interpretation 3.15 (4), we have the following result:
Theorem 3.20. Given a convex function f : R → R, we have the following Jensen
inequality, for any x1 , . . . , xN ∈ R, and any λ1 , . . . , λN > 0 summing up to 1,
f (λ1 x1 + . . . + λN xN ) ≤ λ1 f (x1 ) + . . . + λN xN
with equality when x1 = . . . = xN . In particular, by taking the weights λi to be all equal,
we obtain the following Jensen inequality, valid for any x1 , . . . , xN ∈ R,
 
x1 + . . . + xN f (x1 ) + . . . + f (xN )
f ≤
N N
and once again with equality when x1 = . . . = xN . A similar statement holds for the
concave functions, with all the inequalities being reversed.
 3B. SECOND DERIVATIVES 71

Proof. This is indeed something quite routine, the idea being as follows:
(1) First, we can talk about convex functions in a usual, intuitive way, with this
meaning by definition that the following inequality must be satisfied:
 
x+y f (x) + f (y)
f ≤
2 2
(2) But this means, via a simple argument, by approximating numbers t ∈ [0, 1] by
sums of powers 2−k , that for any t ∈ [0, 1] we must have:
f (tx + (1 − t)y) ≤ tf (x) + (1 − t)f (y)
Alternatively, via yet another simple argument, this time by doing some geometry
with triangles, this means that we must have:
 
x1 + . . . + xN f (x1 ) + . . . + f (xN )
f ≤
N N
But then, again alternatively, by combining the above two simple arguments, the
following must happen, for any λ1 , . . . , λN > 0 summing up to 1:
f (λ1 x1 + . . . + λN xN ) ≤ λ1 f (x1 ) + . . . + λN xN
(3) Summarizing, all our Jensen inequalities, at N = 2 and at N ∈ N arbitrary, are
equivalent. The point now is that, if we look at what the first Jensen inequality, that we
took as definition for the convexity, exactly means, this is simply equivalent to:
f ′′ (x) ≥ 0
(4) Thus, we are led to the conclusions in the statement, regarding the convex func-
tions. As for the concave functions, the proof here is similar. Alternatively, we can say
that f is concave precisely when −f is convex, and get the results from what we have. □
As a basic application of the Jensen inequality, which is very classical, we have:
Theorem 3.21. For any p ∈ (1, ∞) we have the following inequality,
p
x 1 + . . . + xN |x1 |p + . . . + |xN |p
≤
N N
and for any p ∈ (0, 1) we have the following inequality,
p
x1 + . . . + xN |x1 |p + . . . + |xN |p
≥
N N
with in both cases equality precisely when |x1 | = . . . = |xN |.
Proof. This follows indeed from Theorem 3.20, because we have:
(xp )′′ = p(p − 1)xp−2
Thus xp is convex for p > 1 and concave for p < 1, which gives the results. □
72 3. DERIVATIVES

Observe that at p = 2 we obtain as particular case of the above inequality the Cauchy-
Schwarz inequality, or rather something equivalent to it, namely:
2
x2 + . . . + x2N

x1 + . . . + xN
≤ 1
N N
We will be back to this later on in this book, when talking scalars products and Hilbert
spaces, with some more conceptual proofs for such inequalities.

Finally, as yet another important application of the Jensen inequality, we have:

Theorem 3.22. We have the Young inequality,
ap b q
ab ≤ +
p q
1 1
valid for any a, b ≥ 0, and any exponents p, q > 1 satisfying p
+ q
= 1.
Proof. We use the logarithm function, which is concave on (0, ∞), due to:
 ′
′′ 1 1
(log x) = − =− 2
x x
Thus we can apply the Jensen inequality, and we obtain in this way:
 p
bq log(ap ) log(bq )

a
log + ≥ +
p q p q
= log(a) + log(b)
= log(ab)
Now by exponentiating, we obtain the Young inequality. □
Observe that for the simplest exponents, namely p = q = 2, the Young inequality
gives something which is trivial, but is very useful and basic, namely:
a2 + b 2
ab ≤
2
In general, the Young inequality is something non-trivial, and the idea with it is that
2 2
“when stuck with a problem, and with ab ≤ a +b 2
not working, try Young”. We will be
back to this general principle, later in this book, with some illustrations.

3c. The Taylor formula

Back now to the general theory of the derivatives, and their theoretical applications,
we can further develop our basic approximation method, at order 3, at order 4, and so
on, the ultimate result on the subject, called Taylor formula, being as follows:
 3C. THE TAYLOR FORMULA 73

Theorem 3.23. Any function f : R → R can be locally approximated as

∞
X f (k) (x) k
f (x + t) = t
k=0
k!
where f (k) (x) are the higher derivatives of f at the point x.
Proof. Consider the function to be approximated, namely:
φ(t) = f (x + t)
Let us try to best approximate this function at a given order n ∈ N. We are therefore
looking for a certain polynomial in t, of the following type:
P (t) = a0 + a1 t + . . . + an tn
The natural conditions to be imposed are those stating that P and φ should match
at t = 0, at the level of the actual value, of the derivative, second derivative, and so on
up the n-th derivative. Thus, we are led to the approximation in the statement:
n
X f (k) (x) k
f (x + t) ≃ t
k=0
k!
In order to prove now that this approximation holds indeed, we can use L’Hôpital’s
rule, applied several times, as in the proof of Theorem 3.17. To be more precise, if we
denote by φ(t) ≃ P (t) the approximation to be proved, we have:
φ(t) − P (t) φ′ (t) − P ′ (t)
≃
tn ntn−1
φ (t) − P ′′ (t)
′′
≃
n(n − 1)tn−2
..
.
φ(n) (t) − P (n) (t)
≃
n!
f (n) (x) − f (n) (x)
=
n!
= 0
Thus, we are led to the conclusion in the statement. □
Here is a related interesting statement, inspired from the above proof:
Proposition 3.24. For a polynomial of degree n, the Taylor approximation
n
X f (k) (x) k
f (x + t) ≃ t
k=0
k!
is an equality. The converse of this statement holds too.
74 3. DERIVATIVES

Proof. By linearity, it is enough to check the equality in question for the monomials
f (x) = xp , with p ≤ n. But here, the formula to be proved is as follows:
p
p
X p(p − 1) . . . (p − k + 1)
(x + t) ≃ xp−k tk
k=0
k!
We recognize the binomial formula, so our result holds indeed. As for the converse,
this is clear, because the Taylor approximation is a polynomial of degree n. □
There are many other things that can be said about the Taylor formula, at the the-
oretical level, notably with a study of the remainder, when truncating this formula at a
given order n ∈ N. We will be back to this later, in chapter 4 below.

As an application of the Taylor formula, we can now improve the binomial formula,
which was actually our main tool so far, in the following way:
Theorem 3.25. We have the following generalized binomial formula, with p ∈ R,
∞  
p
X p p−k k
(x + t) = x t
k=0
k
with the generalized binomial coefficients being given by the formula
 
p p(p − 1) . . . (p − k + 1)
=
k k!
valid for any |t| < |x|. With p ∈ N, we recover the usual binomial formula.
Proof. It is customary to divide everything by x, which is the same as assuming
x = 1. The formula to be proved is then as follows, under the assumption |t| < 1:
∞  
p
X p k
(1 + t) = t
k=0
k
Let us discuss now the validity of this formula, depending on p ∈ R:
(1) Case
p

p ∈ N. According to our definition of the generalized binomial coefficients,
we have k = 0 for k > p, so the series is stationary, and the formula to be proved is:
p  
p
X p k
(1 + t) = t
k=0
k
But this is the usual binomial formula, which holds for any t ∈ R.
(2) Case p = −1. Here we can use the following formula, valid for |t| < 1:
1
= 1 − t + t2 − t3 + . . .
1+t
 3C. THE TAYLOR FORMULA 75

But this is exactly our generalized binomial formula at p = −1, because:

 
−1 (−1)(−2) . . . (−k)
= = (−1)k
k k!
(3) Case p ∈ −N. This is a continuation of our study at p = −1, which will finish the
study at p ∈ Z. With p = −m, the generalized binomial coefficients are:
 
−m (−m)(−m − 1) . . . (−m − k + 1)
=
k k!
m(m + 1) . . . (m + k − 1)
= (−1)k
k!
(m + k − 1)!
= (−1)k
(m − 1)!k!
 
k m+k−1
= (−1)
m−1
Thus, our generalized binomial formula at p = −m reads:
∞  
1 k m+k−1
X
m
= (−1) tk
(1 + t) k=0
m−1
But this is something which holds indeed, as we know from chapter 2.
(4) General case, p ∈ R. As we can see, things escalate quickly, so we will skip the
next step, p ∈ Q, and discuss directly the case p ∈ R. Consider the following function:
f (x) = xp
The derivatives at x = 1 are then given by the following formula:
f (k) (1) = p(p − 1) . . . (p − k + 1)
Thus, the Taylor approximation at x = 1 is as follows:
∞
X p(p − 1) . . . (p − k + 1) k
f (1 + t) = t
k=0
k!
But this is exactly our generalized binomial formula, so we are done with the case
where t is small. With a bit more care, we obtain that this holds for any |t| < 1, and we
will leave this as an instructive exercise, and come back to it, later in this book. □
We can see from the above the power of the Taylor formula, saving us from quite
complicated combinatorics. Remember indeed the mess from chapter 2, when trying to
directly establish particular cases of the generalized binomial formula. Gone all that.

As a main application now of our generalized binomial formula, which is something

very useful in practice, we can extract square roots, as follows:
76 3. DERIVATIVES

Proposition 3.26. We have the following formula,

∞  k
√ X −t
1+t=1−2 Ck−1
k=1
4
1 2k


with Ck = k+1 k
being the Catalan numbers. Also, we have
∞  k
1 X −t
√ = Dk
1 + t k=0 4

with Dk = 2k


k
being the central binomial coefficients.
Proof. This is something that we already know from chapter 2, but time now to
review all this. At p = 1/2, the generalized binomial coefficients are:
 
1/2 1/2(−1/2) . . . (3/2 − k)
=
k k!
(2k − 2)!
= (−1)k−1 k−1
2 (k − 1)!2k k!
 k
−1
= −2 Ck−1
4
Also, at p = −1/2, the generalized binomial coefficients are:
 
−1/2 −1/2(−3/2) . . . (1/2 − k)
=
k k!
(2k)!
= (−1)k k k
2 k!2 k!
 k
−1
= Dk
4
Thus, Theorem 3.25 at p = ±1/2 gives the formulae in the statement. □
As another basic application of the Taylor series, we have:
Theorem 3.27. We have the following formulae,
∞ 2l+1 ∞
X
l x
X x2l
sin x = (−1) , cos x = (−1)l
l=0
(2l + 1)! l=0
(2l)!
as well as the following formulae,
∞ ∞
x
X xk X
k+1 x
k
e = , log(1 + x) = (−1)
k=0
k! k=0
k
as Taylor series, and in general as well, with |x| < 1 needed for log.
 3D. DIFFERENTIAL EQUATIONS 77

Proof. There are several statements here, the proofs being as follows:
(1) Regarding sin and cos, we can use here the following formulae:
(sin x)′ = cos x , (cos x)′ = − sin x
Thus, we can differentiate sin and cos as many times as we want to, so we can compute
the corresponding Taylor series, and we obtain the formulae in the statement.
(2) Regarding exp and log, here the needed formulae, which lead to the formulae in
the statement for the corresponding Taylor series, are as follows:
(ex )′ = ex
(log x)′ = x−1
(xp )′ = pxp−1
(3) Finally, the fact that the formulae in the statement extend beyond the small t
setting, coming from Taylor series, is something standard too. We will leave this as an
instructive exercise, and come back to it later, in chapter 6 below. □

3d. Differential equations

Good news, with the calculus that we know we can do some physics, in 1 dimension.
Let us start with something immensely important, in the history of science:
Fact 3.28. Newton invented calculus for formulating the laws of motion as
v = ẋ , a = v̇
where x, v, a are the position, speed and acceleration, and the dots are time derivatives.
To be more precise, the variable in Newton’s physics is time t ∈ R, playing the role of
the variable x ∈ R that we have used in the above. And we are looking at a particle whose
position is described by a function x = x(t). Then, it is quite clear that the speed of this
particle should be described by the first derivative v = x′ (t), and that the acceleration of
the particle should be described by the second derivative a = v ′ (t) = x′′ (t).

Summarizing, with Newton’s theory of derivatives, as we learned it in this chapter,

we can certainly do some mathematics for the motion of bodies. But, for these bodies to
move, we need them to be acted upon by some forces, right? The simplest such force is
gravity, and in our present, modest 1 dimensional setting, we have:
Theorem 3.29. The equation of a gravitational free fall, in 1 dimension, is
GM
ẍ = − 2
x
with M being the attracting mass, and G ≃ 6.674 × 10−11 being a constant.
78 3. DERIVATIVES

Proof. Assume indeed that we have a free falling object, in 1 dimension:

◦m


•M
In order to reach to calculus as we know it, we must peform a rotation, as to have all
this happening on the Ox axis. By doing this, and assuming that M is fixed at 0, our
picture becomes as follows, with the attached numbers being now the coordinates:
•0 o ◦x
Now comes the physics. The gravitational force exterted by M , which is fixed in our
formalism, on the object m which moves, is subject to the following equations:
Mm
F = −G · , F = ma , a = v̇ , v = ẋ
x2
To be more precise, in the first equation G ≃ 6.674×10−11 is the gravitational constant,
in usual SI units, and the sign is − because F is attractive. The second equation is
something standard and very intuitive, and the last two equations are those from Fact
3.28. Now observe that, with the above data for F , the equation F = ma reads:
Mm
−G · = mẍ
x2
Thus, by simplifying, we are led to the equation in the statement. □
As more phsyics, we can talk as well about waves in 1 dimension, as follows:
Theorem 3.30. The wave equation in 1 dimension is
φ̈ = v 2 φ′′
with the dot denoting time derivatives, and v > 0 being the propagation speed.
Proof. In order to understand the propagation of the waves, let us model the space,
which is R for us, as a network of balls, with springs between them, as follows:
· · · ××× • ××× • ××× • ××× • ××× • ××× · · ·
Now let us send an impulse, and see how balls will be moving. For this purpose, we
zoom on one ball. The situation here is as follows, l being the spring length:
· · · · · · •φ(x−l) ××× •φ(x) ××× •φ(x+l) · · · · · ·
We have two forces acting at x. First is the Newton motion force, mass times accel-
eration, which is as follows, with m being the mass of each ball:
Fn = m · φ̈(x)
 3D. DIFFERENTIAL EQUATIONS 79

And second is the Hooke force, displacement of the spring, times spring constant.
Since we have two springs at x, this is as follows, k being the spring constant:
Fh = Fhr − Fhl
= k(φ(x + l) − φ(x)) − k(φ(x) − φ(x − l))
= k(φ(x + l) − 2φ(x) + φ(x − l))
We conclude that the equation of motion, in our model, is as follows:
m · φ̈(x) = k(φ(x + l) − 2φ(x) + φ(x − l))
Now let us take the limit of our model, as to reach to continuum. For this purpose
we will assume that our system consists of N >> 0 balls, having a total mass M , and
spanning a total distance L. Thus, our previous infinitesimal parameters are as follows,
with K being the spring constant of the total system, which is of course lower than k:
M L
m= , k = KN , l =
N N
With these changes, our equation of motion found in (1) reads:
KN 2
φ̈(x) = (φ(x + l) − 2φ(x) + φ(x − l))
M
Now observe that this equation can be written, more conveniently, as follows:
KL2 φ(x + l) − 2φ(x) + φ(x − l)
φ̈(x) = ·
M l2
With N → ∞, and therefore l → 0, we obtain in this way:
KL2 d2 φ
φ̈(x) = · 2 (x)
M dx
Thus, we are led to the conclusion in the statement. □
Along the same lines, we can talk as well about the heat equation in 1D, as follows:
Theorem 3.31. The heat equation in 1 dimension is
φ̇ = αφ′′
where α > 0 is the thermal diffusivity of the medium.
Proof. As before with the wave equation, this is not exactly a theorem, but rather
what comes out of experiments, but we can justify this mathematically, as follows:
(1) As an intuitive explanation for this equation, since the second derivative φ′′ com-
putes the average value of a function φ around a point, minus the value of φ at that
point, as we know from Proposition 3.18, the heat equation as formulated above tells us
that the rate of change φ̇ of the temperature of the material at any given point must be
80 3. DERIVATIVES

proportional, with proportionality factor α > 0, to the average difference of temperature

between that given point and the surrounding material. Which sounds reasonable.
(2) In practice now, we can use, a bit like before for the wave equation, a lattice model
as follows, with distance l > 0 between the neighbors:
l l
◦x−l ◦x ◦x+l
In order to model now heat diffusion, we have to implement the intuitive mechanism
explained above, and in practice, this leads to a condition as follows, expressing the change
of the temperature φ, over a small period of time δ > 0:
αδ X
φ(x, t + δ) = φ(x, t) + 2 [φ(y, t) − φ(x, t)]
l x∼y
But this leads, via manipulations as before, to φ̇(x, t) = α · φ′′ (x, t), as claimed. □
All this is very nice, so with the calculus that we know, we can certainly talk about
physics. We will see later in this book how to deal with the above equations.
3e. Exercises
Here are some basic exercises, in relation with the above:
Exercise 3.32. Work out, geometrically, estimates for sin, cos, tan at x = 0.
Exercise 3.33. Work out all the details in the proof of the Jensen inequality.
Exercise 3.34. Clarify the use of l’Hôpital’s rule, in the proof of the Taylor formula.
Exercise 3.35. Complete the proof of the generalized binomial formula, at |t| < 1.
As a bonus exercise, try solving the gravity, wave and heat equations in 1D.
 CHAPTER 4

Integration

4a. Integration theory

We have seen so far the foundations of calculus, with lots of interesting results regard-
ing the functions f : R → R, and their derivatives f ′ : R → R. The general idea was
that in order to understand f , we first need to compute its derivative f ′ . The overall
conclusion, coming from the Taylor formula, was that if we are able to compute f ′ , but
then also f ′′ , and f ′′′ and so on, we will have a good understanding of f itself.

However, the story is not over here, and there is one more twist to the plot. Which will
be a major twist, of similar magnitude to that of the Taylor formula. For reasons which
are quite tricky, that will become clear later on, we will be interested in the integration
of the functions f : R → R. With the claim that this is related to calculus.

There are several possible viewpoints on the integral, which are all useful, and good
to know. To start with, we have something very simple, as follows:
Definition 4.1. The integral of a continuous function f : [a, b] → R, denoted
Z b
f (x)dx
a

is the area below the graph of f , signed + where f ≥ 0, and signed − where f ≤ 0.
Here it is of course understood that the area in question can be computed, and with
this being something quite subtle, that we will get into later. For the moment, let us just
trust our intuition, our function f being continuous, the area in question can “obviously”
be computed. More on this later, but for being rigorous, however, let us formulate:
Method 4.2. In practice, the integral of f ≥ 0 can be computed as follows,
(1) Cut the graph of f from 3mm plywood,
(2) Plunge that graph into a square container of water,
(3) Measure the water displacement, as to have the volume of the graph,
(4) Divide by 3 × 10−3 that volume, as to have the area,
and for general f , we can use this plus f = f+ − f− , with f+ , f− ≥ 0.
81
82 4. INTEGRATION

So far, so good, we have a rigorous definition, so let us do now some computations.

In order to compute areas, and so integrals of functions, without wasting precious water,
we can use our geometric knowledge. Here are some basic results of this type:
Proposition 4.3. We have the following results:
(1) When f is linear, we have the following formula:
Z b
f (a) + f (b)
f (x)dx = (b − a) ·
a 2
(2) In fact, when f is piecewise linear on [a = a1 , a2 , . . . , an = b], we have:
Z b n−1
X f (ai ) + f (ai+1 )
f (x)dx = (ai+1 − ai ) ·
a i=1
2
R1 √
(3) We have as well the formula −1 1 − x2 dx = π/2.
Proof. These results all follow from basic geometry, as follows:
(1) Assuming f ≥ 0, we must compute the area of a trapezoid having sides f (a), f (b),
and height b−a. But this is the same as the area of a rectangle having side (f (a)+f (b))/2
and height b − a, and we obtain (b − a)(f (a) + f (b))/2, as claimed.
(2) This is clear indeed from the formula found in (1), by additivity.
(3) The integral in the statement is by definition the area of the upper unit half-disc.
But since the area of the whole unit disc is π, this half-disc area is π/2. □
As an interesting observation, (2) in the above result makes it quite clear that f does
not necessarily need to be continuous, in order to talk about its integral. Indeed, assuming
that f is piecewise linear on [a = a1 , a2 , . . . , an = b], but not necessarily continuous, we
can still talk about its integral, in the obvious way, exactly as in Definition 4.1, and we
have an explicit formula for this integral, generalizing the one found in (2), namely:
n−1 −
f (a+
Z b
i ) + f (ai+1 )
X
f (x)dx = (ai+1 − ai ) ·
a i=1
2
Based on this observation, let us upgrade our formalism, as follows:
Definition 4.4. We say that a function f : [a, b] → R is integrable when the area
below its graph is computable. In this case we denote by
Z b
f (x)dx
a

this area, signed + where f ≥ 0, and signed − where f ≤ 0.

 4A. INTEGRATION THEORY 83

As basic examples of integrable functions, we have the continuous ones, provided

indeed that our intuition, or that Method 4.2, works indeed for any such function. We will
soon see that this is indeed true, coming with mathematical proof. As further examples,
we have the functions which are piecewise linear, or more generally piecewise continuous.
We will also see, later, as another class of examples, that the piecewise monotone functions
are integrable. But more on this later, let us not bother for the moment with all this.

This being said, one more thing regarding theory, that you surely have in mind: is
any function integrable? Not clear. I would say that if the Devil comes with some sort
of nasty, totally discontinuous function f : R → R, then you will have big troubles in
cutting its graph from 3mm plywood, as required by Method 4.2. More on this later.

Back to work now, here are some general results regarding the integrals:
Proposition 4.5. We have the following formulae,
Z b Z b Z b
f (x) + g(x)dx = f (x)dx + g(x)dx
a a a
Z b Z b
λf (x) = λ f (x)
a a
valid for any functions f, g and any scalar λ ∈ R.
Proof. Both these formulae are indeed clear from definitions. □
Moving ahead now, passed the above results, which are of purely algebraic and geo-
metric nature, and perhaps a few more of the same type, which are all quite trivial and
that we we will not get into here, we must do some analysis, in order to compute integrals.
This is something quite tricky, and we have here the following result:
Theorem 4.6. We have the Riemann integration formula,
Z b N  
1 X b−a
f (x)dx = (b − a) × lim f a+ ·k
a N →∞ N N
k=1

which can serve as a definition for the integral.

Proof. This is standard, by drawing rectangles. We have indeed the following for-
mula, which can stand as a definition for the signed area below the graph of f :
Z b N  
1 Xb−a b−a
f (x)dx = lim ·f a+ ·k
a N →∞ N N N
k=1

Thus, we are led to the formula in the statement. □

84 4. INTEGRATION
Rb
Observe that the above formula suggests that a f (x)dx is the length of the interval
[a, b], namely b − a, times the average of f on the interval [a, b]. Thinking a bit, this is
indeed something true, with no need for Riemann sums, coming directly from Definition
4.1, because area means side times average height. Thus, we can formulate:
Theorem 4.7. The integral of a function f : [a, b] → R is given by
Z b
f (x)dx = (b − a) × A(f )
a
where A(f ) is the average of f over the interval [a, b].
Proof. As explained above, this is clear from Definition 4.1, via some geometric
thinking. Alternatively, this is something which certainly comes from Theorem 4.6. □
The point of view in Theorem 4.7 is something quite useful, and as an illustration for
this, let us review the results that we already have, by using this interpretation. First,
we have the formula for linear functions from Proposition 4.3, namely:
Z b
f (a) + f (b)
f (x)dx = (b − a) ·
a 2
But this formula is totally obvious with our new viewpoint, from Theorem 4.7. The
same goes for the results in Proposition 4.5, which become even more obvious with the
viewpoint from Theorem 4.7. However, Rnot√everything trivializes in this way, and the
1
result which is left, namely the formula −1 1 − x2 dx = π/2 from Proposition 4.3 (3),
not only does not trivialize, but becomes quite opaque with our new philosophy.

In short, modesty. Integration is a quite delicate business, and we have several equiv-
alent points of view on what an integral means, and all these points of view are useful,
and must be learned, with none of them being clearly better than the others.

Going ahead with more interpretations of the integral, we have:

Theorem 4.8. We have the Monte Carlo integration formula,
Z b N
1 X
f (x)dx = (b − a) × lim f (xi )
a N →∞ N
k=1

with x1 , . . . , xN ∈ [a, b] being random.

Proof. We recall from Theorem 4.7 that the idea is that we have a formula as follows,
with the points x1 , . . . , xN ∈ [a, b] being uniformly distributed:
Z b N
1 X
f (x)dx = (b − a) × lim f (xi )
a N →∞ N
k=1
 4A. INTEGRATION THEORY 85

But this works as well when the points x1 , . . . , xN ∈ [a, b] are randomly distributed,
for somewhat obvious reasons, and this gives the result. □
Observe that Monte Carlo integration works better than Riemann integration, for
instance when trying to improve the estimate, via N → N + 1. Indeed, in the context of
Riemann integration, assume that we managed to find an estimate as follows, which in
practice requires computing N values of our function f , and making their average:
Z b N  
b−aX b−a
f (x)dx ≃ f a+ ·k
a N k=1 N
In order to improve this estimate, any extra computed value of our function f (y) will
be unuseful. For improving our formula, what we need are N extra values of our function,
f (y1 ), . . . , f (yN ), with the points y1 , . . . , yN being the midpoints of the previous division
of [a, b], so that we can write an improvement of our formula, as follows:
Z b 2N  
b−aX b−a
f (x)dx ≃ f a+ ·k
a 2N k=1 2N
With Monte Carlo, things are far more flexible. Assume indeed that we managed to
find an estimate as follows, which again requires computing N values of our function:
Z b N
b−aX
f (x)dx ≃ f (xi )
a N k=1
Now if we want to improve this, any extra computed value of our function f (y) will
be helpful, because we can set xn+1 = y, and improve our estimate as follows:
Z b N +1
b−a X
f (x)dx ≃ f (xi )
a N + 1 k=1
And isn’t this potentially useful, and powerful, when thinking at practically computing
integrals, either by hand, or by using a computer. Let us record this finding as follows:
Conclusion 4.9. Monte Carlo integration works better than Riemann integration,
when it comes to computing as usual, by estimating, and refining the estimate.
As another interesting feature of Monte Carlo integration, this works better than Rie-
mann integration, for functions having various symmetries, because Riemann integration
can get “fooled” by these symmetries, while Monte Carlo remains strong.

As an example for this phenomeon, chosen to be quite drastic, let us attempt to

integrate, via both Riemann and Monte Carlo, the following function f : [0, π] → R:
f (x) = sin(120x)
86 4. INTEGRATION

The first few Riemann sums for this function are then as follows:
π
I2 (f ) = (| sin 0| + | sin 60π|) = 0
2
π
I3 (f ) = (| sin 0| + | sin 40π| + | sin 80π|) = 0
3
π
I4 (f ) = (| sin 0| + | sin 30π| + | sin 60π| + | sin 90π|) = 0
4
π
I5 (f ) = (| sin 0| + | sin 24π| + | sin 48π| + | sin 72π| + | sin 96π|) = 0
5
π
I6 (f ) = (| sin 0| + | sin 20π| + | sin 40π| + | sin 60π| + | sin 80π| + | sin 100π|) = 0
6
..
.
Based on this evidence, we will conclude, obviously, that we have:
Z π
f (x)dx = 0
0

With Monte Carlo, however, such things cannot happen. Indeed, since there are
finitely many points x ∈ [0, π] having the property sin(120x) = 0, a random point x ∈ [0, π]
will have the property | sin(120x)| > 0, so Monte Carlo will give, at any N ∈ N:
Z π N
b−aX
f (x)dx ≃ f (xi ) > 0
0 N k=1
Again, this is something interesting, when practically computing integrals, either by
hand, or by using a computer. So, let us record, as a complement to Conclusion 4.9:
Conclusion 4.10. Monte Carlo integration is smarter than Riemann integration,
because the symmetries of the function can fool Riemann, but not Monte Carlo.
All this is good to know, when computing integrals in practice, especially with a
computer. Finally, here is one more useful interpretation of the integral:
Theorem 4.11. The integral of a function f : [a, b] → R is given by
Z b
f (x)dx = (b − a) × E(f )
a

where E(f ) is the expectation of f , regarded as random variable.

Proof. This is just some sort of fancy reformulation of Theorem 4.8, the idea being
that what we can “expect” from a random variable is of course its average. We will be
back to this later in this book, when systematically discussing probability theory. □
 4B. RIEMANN SUMS 87

4b. Riemann sums

Our purpose now will be to understand which functions f : R → R are integrable,
and how to compute their integrals. For this purpose, the Riemann formula in Theorem
4.6 will be our favorite tool. Let us begin with some theory. We first have:
Theorem 4.12. The following functions are integrable:
(1) The piecewise continuous functions.
(2) The piecewise monotone functions.
Proof. This is indeed something quite standard, as follows:
(1) It is enough to prove the first assertion for a function f : [a, b] → R which is
continuous, and our claim here is that this follows from the uniform continuity of f . To
be more precise, given ε > 0, let us choose δ > 0 such that the following happens:
|x − y| < δ =⇒ |f (x) − f (y)| < ε
In order to prove the result, let us pick two divisions of [a, b], as follows:
I = [a = a1 < a2 < . . . < an = b]

I ′ = [a = a′1 < a′2 < . . . < a′m = b]

Our claim, which will prove the result, is that if these divisions are sharp enough, of
resolution < δ/2, then the associated Riemann sums ΣI (f ), ΣI ′ (f ) are close within ε:
δ
ai+1 − ai < , a′i+1 − a′i < δ2 =⇒ ΣI (f ) − ΣI ′ (f ) < ε
2
(2) In order to prove this claim, let us denote by l the length of the intervals on the
real line. Our assumption is that the lengths of the divisions I, I ′ satisfy:

δ
δ
l [ai , ai+1 ] < , l [a′i , a′i+1 ] <
2 2
Now let us intersect the intervals of our divisions I, I ′ , and set:
lij = l [ai , ai+1 ] ∩ [a′j , a′j+1 ]

The difference of Riemann sums that we are interested in is then given by:
X X
ΣI (f ) − ΣI ′ (f ) = lij f (ai ) − lij f (a′j )
ij ij

X
= lij (f (ai ) − f (a′j ))
ij
88 4. INTEGRATION

(3) Now let us estimate f (ai ) − f (a′j ). Since in the case lij = 0 we do not need this
estimate, we can assume lij > 0. Now by remembering what the definition of the numbers
lij was, we conclude that we have at least one point x ∈ R satisfying:
x ∈ [ai , ai+1 ] ∩ [a′j , a′j+1 ]
But then, by using this point x and our assumption on I, I ′ involving δ, we get:
|ai − a′j | ≤ |ai − x| + |x − a′j |
δ δ
≤ +
2 2
= δ
Thus, according to our definition of δ from (1), in relation to ε, we get:
|f (ai ) − f (a′j )| < ε
(4) But this is what we need, in order to finish. Indeed, with the estimate that we
found, we can finish the computation started in (2), as follows:
X
ΣI (f ) − ΣI ′ (f ) = lij (f (ai ) − f (a′j ))
ij
X
≤ ε lij
ij

= ε(b − a)
Thus our two Riemann sums are close enough, provided that they are both chosen to
be fine enough, and this finishes the proof of the first assertion.
(5) Regarding now the second assertion, this is something more technical, that we will
not really need in what follows. We will leave the proof here, which uses similar ideas to
those in the proof of (1) above, namely subdivisions and estimates, as an exercise. □
Going ahead with more theory, let us establish some abstract properties of the inte-
gration operation. We already know from Proposition 4.5 that the integrals behave well
with respect to sums and multiplication by scalars. Along the same lines, we have:
Theorem 4.13. The integrals behave well with respect to taking limits,
Z b  Z b
lim fn (x) dx = lim fn (x)dx
a n→∞ n→∞ a
and with respect to taking infinite sums as well,
∞
Z b X ! ∞ Z b
X
fn (x) dx = fn (x)dx
a n=0 n=0 a

with both these formulae being valid, undwer mild assumptions.

 4B. RIEMANN SUMS 89

Proof. This is something quite standard, by using the general theory developed in
chapter 3 for the sequences and series of functions. To be more precise, (1) follows by
using the material there, via Riemann sums, and then (2) follows as a particular case of
(1). We will leave the clarification of all this as an instructive exercise. □
Finally, still at the general level, let us record as well the following result:
Theorem 4.14. Given a continuous function f : [a, b] → R, we have
Z b
∃c ∈ [a, b] , f (x)dx = (b − a)f (c)
a
with this being called mean value property.
Proof. Our claim is that this follows from the following trivial estimate:
min(f ) ≤ f ≤ max(f )
Indeed, by integrating this over [a, b], we obtain the following estimate:
Z b
(b − a) min(f ) ≤ f (x)dx ≤ (b − a) max(f )
a
Now observe that this latter estimate can be written as follows:
Rb
f (x)dx
min(f ) ≤ a ≤ max(f )
b−a
Since f must takes all values on [min(f ), max(f )], we get a c ∈ [a, b] such that:
Rb
a
f (x)dx
= f (c)
b−a
Thus, we are led to the conclusion in the statement. □
At the level of examples now, let us first look at the simplest functions that we know,
namely the power functions f (x) = xp . However, things here are tricky, as follows:
Theorem 4.15. We have the integration formula
Z b
bp+1 − ap+1
xp dx =
a p+1
valid at p = 0, 1, 2, 3.
Proof. This is something quite tricky, the idea being as follows:
(1) By linearity we can assume that our interval [a, b] is of the form [0, c], and the
formula that we want to establish is as follows:
Z c
p cp+1
x dx =
0 p+1
90 4. INTEGRATION

(2) We can further assume c = 1, and by expressing the left term as a Riemann sum,
we are in need of the following estimate, in the N → ∞ limit:
N p+1
1p + 2p + . . . + N p ≃
p+1
(3) So, let us try to prove this. At p = 0, obviously nothing to do, because we have
the following formula, which is exact, and which proves our estimate:
10 + 20 + . . . + N 0 = N
(4) At p = 1 now, we are confronted with a well-known question, namely the compu-
tation of 1 + 2 + . . . + N . But this is simplest done by arguing that the average of the
numbers 1, 2, . . . , N being the number in the middle, we have:
1 + 2 + ... + N N +1
=
N 2
Thus, we obtain the following formula, which again solves our question:
N (N + 1) N2
1 + 2 + ... + N = ≃
2 2
2 2 2
(5) At p = 2 now, go compute 1 + 2 + . . . + N . This is not obvious at all, so as
a preliminary here, let us go back to the case p = 1, and try to find a new proof there,
which might have some chances to extend at p = 2. The trick is to use 2D geometry.
Indeed, consider the following picture, with stacks going from 1 to N :
□
..
.
□ ... □
□ □ ... □
□ □ □ ... □
Now if we take two copies of this, and put them one on the top of the other, with a
twist, in the obvious way, we obtain a rectangle having size N × (N + 1). Thus:
2(1 + 2 + . . . + N ) = N (N + 1)
But this gives the same formula as before, solving our question, namely:
N (N + 1) N2
1 + 2 + ... + N = ≃
2 2
(6) Armed with this new method, let us attack now the case p = 2. Here we obviously
need to do some 3D geometry, namely taking the picture P formed by a succession of
solid squares, having sizes 1 × 1, 2 × 2, 3 × 3, and so on up to N × N . Some quick
thinking suggests that stacking 3 copies of P , with some obvious twists, will lead us to a
 4B. RIEMANN SUMS 91

parallelepiped. But this is not exactly true, and some further thinking shows that what
we have to do is to add 3 more copies of P , leading to the following formula:
N (N + 1)(2N + 1)
12 + 22 + . . . + N 2 =
6
Or at least, that’s how the legend goes. In practice, the above formula holds indeed,
and you can check it for instance by recurrence, and this solves our problem:
2N 3 N3
12 + 22 + . . . + N 2 ≃ =
6 3
(7) At p = 3 now, the legend goes that by deeply thinking in 4D we are led to the
following formula, a bit as in the cases p = 1, 2, explained above:
 2
3 3 3 N (N + 1)
1 + 2 + ... + N =
2
Alternatively, assuming that the gods of combinatorics are with us, we can see right
away the following formula, which coupled with (4) gives the result:
13 + 23 + . . . + N 3 = (1 + 2 + . . . + N )2
In any case, in practice, the above formula holds indeed, and you can check it for
instance by recurrence, and this solves our problem:
N4
13 + 23 + . . . + N 3 ≃
4
(8) Thus, good news, we proved our theorem. Of course, I can hear you screaming,
that what about p = 4 and higher. But the thing is that, by a strange twist of fate, there
is no exact formula for 1p + 2p + . . . + N p , at p = 4 and higher. Thus, game over. □
What happened above, with us unable to integrate xp at p = 4 and higher, not to
mention the exponents p ∈ R − N that we have not even dared to talk about, is quite
annoying. As a conclusion to all this, however, let us formulate:
Conjecture 4.16. We have the following estimate,
N p+1
1p + 2p + . . . + N p ≃
p+1
and so, by Riemann sums, we have the following integration formula,
Z b
bp+1 − ap+1
xp dx =
a p+1
valid for any exponent p ∈ N, and perhaps for some other p ∈ R.
We will see later that this conjecture is indeed true, and with the exact details re-
garding the exponents p ∈ R − N too. Now, instead of struggling with this, let us look at
some other functions, which are not polynomial. And here, as good news, we have:
92 4. INTEGRATION

Theorem 4.17. We have the following integration formula,

Z b
ex dx = eb − ea
a
valid for any two real numbers a < b.
Proof. This follows indeed from the Riemann integration formula, because:
Z b
ea + ea+(b−a)/N + ea+2(b−a)/N + . . . + ea+(N −1)(b−a)/N
ex dx = lim
a N →∞ N
a
e
· 1 + e(b−a)/N + e2(b−a)/N + . . . + e(N −1)(b−a)/N


= lim
N →∞ N
ea eb−a − 1
= lim · (b−a)/N
N →∞ N e −1
1
= (eb − ea ) lim (b−a)/N
N →∞ N (e − 1)
= eb − ea
Thus, we are led to the conclusion in the statement. □

4c. Advanced results

The problem is now, what to do with what we have, namely Conjecture 4.16 and
Theorem 4.17. Not obvious, so stuck, and time to ask the cat. And cat says:
Cat 4.18. Summing the infinitesimals of the rate of change of the function should give
you the global change of the function. Obvious.
Which is quite puzzling, usually my cat is quite helpful. Guess he must be either a
reincarnation of Newton or Leibnitz, these gentlemen used to talk like that, or that I
should take care at some point of my garden, remove catnip and other weeds.

This being said, wait. There is suggestion to connect integrals and derivatives, and
this is in fact what we have, coming from Conjecture 4.16 and Theorem 4.17, due to:
 p+1 ′
x
= xp , (ex )′ = ex
p+1
So, eureka, we have our idea, thanks cat. Moving ahead now, following this idea, we
first have the following result, called fundamental theorem of calculus:
Theorem 4.19. Given a continuous function f : [a, b] → R, if we set
Z x
F (x) = f (s)ds
a
then F ′ = f . That is, the derivative of the integral is the function itself.
 4C. ADVANCED RESULTS 93

Proof. This follows from the Riemann integration picture, and more specifically,
from the mean value property from Theorem 4.14. Indeed, we have:
F (x + t) − F (x) 1 x+t
Z
= f (x)dx
t t x
On the other hand, our function f being continuous, by using the mean value property
from Theorem 4.14, we can find a number c ∈ [x, x + t] such that:
1 x+t
Z
f (x)dx = f (x)
t x
Thus, putting our formulae together, we conclude that we have:
F (x + t) − F (x)
= f (c)
t
Now with t → 0, no matter how the number c ∈ [x, x + t] varies, one thing that we
can be sure about is that we have c → x. Thus, by continuity of f , we obtain:
F (x + t) − F (x)
lim = f (x)
t→0 t
But this means exactly that we have F ′ = f , and we are done. □
We have as well the following result, which is something equivalent, and a hair more
beautiful, also called fundamental theorem of calculus:
Theorem 4.20. Given a function F : R → R, we have
Z b
F ′ (x)dx = F (b) − F (a)
a
for any interval [a, b].
Proof. As already mentioned, this is something which follows from Theorem 4.19,
and is in fact equivalent to it. Indeed, consider the following function:
Z s
G(s) = F ′ (x)dx
a
By using Theorem 4.19 we have G′ = F ′ , and so our functions F, G differ by a constant.
But with s = a we have G(a) = 0, and so the constant is F (a), and we get:
F (s) = G(s) + F (a)
Now with s = b this gives F (b) = G(b) + F (a), which reads:
Z b
F (b) = F ′ (x)dx + F (a)
a
Thus, we are led to the conclusion in the statement. □
As a first illustration for all this, solving our previous problems, we have:
94 4. INTEGRATION

Theorem 4.21. We have the following integration formulae,

Z b Z b
bp+1 − ap+1
 
p 1 b
x dx = , dx = log
a p+1 a x a
Z b Z b
sin x dx = cos a − cos b , cos x dx = sin b − sin a
a a
Z b Z b
x b a
e dx = e − e , log x dx = b log b − a log a − b + a
a a
all obtained, in case you ever forget them, via the fundamental theorem of calculus.
Proof. We already know some of these formulae, but the best is to do everything,
using the fundamental theorem of calculus. The computations go as follows:
(1) With F (x) = xp+1 we have F ′ (x) = pxp , and we get, as desired:
Z b
pxp dx = bp+1 − ap+1
a
(2) Observe first that the formula (1) does not work at p = −1. However, here we can
use F (x) = log x, having as derivative F ′ (x) = 1/x, which gives, as desired:
Z b  
1 b
dx = log b − log a = log
a x a
(3) With F (x) = cos x we have F ′ (x) = − sin x, and we get, as desired:
Z b
− sin x dx = cos b − cos a
a
(4) With F (x) = sin x we have F ′ (x) = cos x, and we get, as desired:
Z b
cos x dx = sin b − sin a
a
(5) With F (x) = e we have F ′ (x) = ex , and we get, as desired:
x

Z b
ex dx = eb − ea
a
(6) This is something more tricky. We are looking for a function satisfying:
F ′ (x) = log x
This does not look doable, but fortunately the answer to such things can be found on
the internet. But, what if the internet connection is down? So, let us think a bit, and try
to solve our problem. Speaking logarithm and derivatives, what we know is:
1
(log x)′ =
x
 4C. ADVANCED RESULTS 95

But then, in order to make appear log on the right, the idea is quite clear, namely
multiplying on the left by x. We obtain in this way the following formula:
1
(x log x)′ = 1 · log x + x · = log x + 1
x
We are almost there, all we have to do now is to substract x from the left, as to get:
(x log x − x)′ = log x
But this this formula in hand, we can go back to our problem, and we get the result. □
Getting back now to theory, inspired by the above, let us formulate:
Definition 4.22. Given a function f , we call primitive of f any function F satisfying:
F′ = f
R
We denote such primitives by f , and also call them indefinite integrals.
Observe that the primitives are unique up to an additive constant, in the sense that if
F is a primitive, then so is F + c, for any c ∈ R, and conversely, if F, G are two primitives,
then we must have G = F + c, for some c ∈ R, with this latter fact coming from a result
from chapter 3, saying that the derivative vanishes when the function is constant.
R
As for the convention at the end, F = f , this comes from the fundamental theorem
of calculus, which can be written as follows, by using this convention:
Z b Z  Z 
f (x)dx = f (b) − f (a)
a

RBy the way, observe that there is no contradiction here, coming from
R the indeterminacy
of f . Indeed, when adding a constant c ∈ R to the chosen primitive f , when conputing
the above difference the c quantities will cancel, and we will obtain the same result.
As an application, we can reformulate Theorem 4.21 in a more digest form, as follows:
Theorem 4.23. We have the following formulae for primitives,
xp+1
Z Z
p 1
x = , = log x
p+1 x
Z Z
sin x = − cos x , cos x = sin x
Z Z
x x
e =e , log x = x log x − x
allowing us to compute the corresponding definite integrals too.
Proof. Here the various formulae in the statement follow from Theorem 4.21, or
rather from the proof of Theorem 4.21, or even from chapter 3, for most of them, and the
last assertion comes from the integration formula given after Definition 4.22. □
96 4. INTEGRATION

Getting back now to theory, we have the following key result:

Theorem 4.24. We have the formula
Z Z
f g + f g′ = f g
′

called integration by parts.

Proof. This follows by integrating the Leibnitz formula, namely:
(f g)′ = f ′ g + f g ′
Indeed, with our convention for primitives, this gives the formula in the statement. □
It is then possible to pass to usual integrals, and we obtain a formula here as well, as
follows, also called integration by parts, with the convention [φ]ba = φ(b) − φ(a):
Z b Z b h ib
′ ′
f g+ fg = fg
a a a

In practice, the most interesting case is that when f g vanishes on the boundary {a, b}
of our interval, leading to the following formula:
Z b Z b
′
fg=− f g′
a a
Examples of this usually come with [a, b] = [−∞, ∞], and more on this later. Now
still at the theoretical level, we have as well the following result:
Theorem 4.25. We have the change of variable formula
Z b Z d
f (x)dx = f (φ(t))φ′ (t)dt
a c
−1 −1
where c = φ (a) and d = φ (b).
Proof. This follows with f = F ′ , from the following differentiation rule, that we
know from chapter 3, and whose proof is something elementary:
(F φ)′ (t) = F ′ (φ(t))φ′ (t)
Indeed, by integrating between c and d, we obtain the result. □
As a main application now of our theory, in relation with advanced calculus, and more
specifically with the Taylor formula from chapter 3, we have:
Theorem 4.26. Given a function f : R → R, we have the formula
n Z x+t (n+1)
X f (k) (x) k f (s)
f (x + t) = t + (x + t − s)n ds
k=0
k! x n!
called Taylor formula with integral formula for the remainder.
 4C. ADVANCED RESULTS 97

Proof. This is something which looks a bit complicated, so we will first do some
verifications, and then we will go for the proof in general:
(1) At n = 0 the formula in the statement is as follows, and certainly holds, due to
R x+t
the fundamental theorem of calculus, which gives x f ′ (s)ds = f (x + t) − f (x):
Z x+t
f (x + t) = f (x) + f ′ (s)ds
x

(2) At n = 1, the formula in the statement becomes more complicated, as follows:

Z x+t
′
f (x + t) = f (x) + f (x)t + f ′′ (s)(x + t − s)ds
x
As a first observation, this formula holds indeed for the linear functions, where we
have f (x + t) = f (x) + f ′ (x)t, and f ′′ = 0. So, let us try f (x) = x2 . Here we have:
f (x + t) − f (x) − f ′ (x)t = (x + t)2 − x2 − 2xt = t2
On the other hand, the integral remainder is given by the same formula, namely:
Z x+t Z x+t
′′
f (s)(x + t − s)ds = 2 (x + t − s)ds
x x
Z x+t
= 2t(x + t) − 2 sds
x
= 2t(x + t) − ((x + t)2 − x2 )
= 2tx + 2t2 − 2tx − t2
= t2
(3) Still at n = 1, let us try now to prove the formula in the statement, in general.
Since what we have to prove is an equality, this cannot be that hard, and the first thought
goes towards differentiating. But this method works indeed, and we obtain the result.
(4) In general, the proof is similar, by differentiating, the computations being similar
to those at n = 1, and we will leave this as an instructive exercise. □
So long for basic integration theory. As a first concrete application now, we can
compute all sorts of areas and volumes. Normally such computations are the business of
multivariable calculus, and we will be back to this later, but with the technology that we
have so far, we can do a number of things. As a first such computation, we have:
Proposition 4.27. The area of an ellipsis, given by the equation
 x 2  y 2
+ =1
a b
with a, b > 0 being half the size of a box containing the ellipsis, is A = πab.
98 4. INTEGRATION

Proof. The idea is that of cutting the ellipsis into vertical slices. First observe that,
according to our equation (x/a)2 + (y/b)2 = 1, the x coordinate can range as follows:
x ∈ [−a, a]
For any such x, the other coordinate y, satisfying (x/a)2 + (y/b)2 = 1, is given by:
r
x2
y = ±b 1 − 2
a
Thus the length of the vertical ellipsis slice at x is given by the following formula:
r
x2
l(x) = 2b 1 − 2
a
We conclude from this discussion that the area of the ellipsis is given by:
Z ar
x2
A = 2b 1 − 2 dx
a
Z−aa √
4b
= a2 − x2 dx
a 0
Z 1p
= 4ab 1 − y 2 dy
0
π
= 4ab ·
4
= πab
Finally, as a verification, for a = b = 1 we get A = π, as we should. □
Moving now to 3D, as an obvious challenge here, we can try to compute the volume
of the sphere. This can be done a bit as for the ellipsis, the answer being as follows:
Theorem 4.28. The volume of the unit sphere is given by:
4π
V =
3
More generally, the volume of the sphere of radius R is V = 4πR3 /3.
Proof. We proceed a bit as for the ellipsis. The equation of the sphere is:
x2 + y 2 + z 2 = 1
Thus, the range of the first coordinate x is as follows:
x ∈ [−1, 1]
Now when this first coordinate x is fixed, the other coordinates y, z vary on a circle,
given by the equation y 2 + z 2 = 1 − x2 , and so having radius as follows:
√
r(x) = 1 − x2
 4C. ADVANCED RESULTS 99

Thus, the vertical slice of our sphere at x has area as follows:

a(x) = πr(x)2 = π(1 − x2 )
We conclude from this discussion that the area of the sphere is given by:
Z 1
A = π 1 − x2 dx
−1
1 ′
x3
Z 
= π x− dx
−1 3
   
1 1
= π 1− − −1 +
3 3
 
2 2
= π +
3 3
4π
=
3
Finally, the last assertion is clear too, by multiplying everything by R, which amounts
in multiplying the final result of our volume computation by R3 . □
As another application of our integration methods, we can now solve the 1D wave
equation. In order to explain this, we will need a standard calculus result, as follows:
Proposition 4.29. The derivative of a function of type
Z h(x)
φ(x) = f (s)ds
g(x)

is given by the formula φ (x) = f (h(x))h (x) − f (g(x))g ′ (x).

′ ′

Proof. Consider a primitive of the function that we integrate, F ′ = f . We have:

Z h(x)
φ(x) = f (s)ds
g(x)
Z h(x)
= F ′ (s)ds
g(x)
= F (h(x)) − F (g(x))
By using now the chain rule for derivatives, we obtain from this:
φ′ (x) = F ′ (h(x))h′ (x) − F ′ (g(x))g ′ (x)
= f (h(x))h′ (x) − f (g(x))g ′ (x)
Thus, we are led to the formula in the statement. □
Now back to the 1D waves, the result here, due to d’Alembert, is as follows:
100 4. INTEGRATION

Theorem 4.30. The solution of the 1D wave equation φ̈ = v 2 φ′′ with initial value
conditions φ(x, 0) = f (x) and φ̇(x, 0) = g(x) is given by the d’Alembert formula:
Z x+vt
f (x − vt) + f (x + vt) 1
φ(x, t) = + g(s)ds
2 2v x−vt
Moreover, in the context of our previous lattice model discretizations, what happens is
more or less that the above d’Alembert integral gets computed via Riemann sums.
Proof. There are several things going on here, the idea being as follows:
(1) Let us first check that the d’Alembert solution is indeed a solution of the wave
equation φ̈ = v 2 φ′′ . The first time derivative is computed as follows:
−vf ′ (x − vt) + vf ′ (x + vt) 1
φ̇(x, t) = + (vg(x + vt) + vg(x − vt))
2 2v
The second time derivative is computed as follows:
v 2 f ′′ (x − vt) + v 2 f (x + vt) vg ′ (x + vt) − vg ′ (x − vt)
φ̈(x, t) = +
2 2
Regarding now space derivatives, the first one is computed as follows:
f ′ (x − vt) + f ′ (x + vt) 1
φ′ (x, t) = + (g ′ (x + vt) − g ′ (x − vt))
2 2v
As for the second space derivative, this is computed as follows:
f ′′ (x − vt) + f ′′ (x + vt) g ′′ (x + vt) − g ′′ (x − vt)
φ′′ (x, t) = +
2 2v
2 ′′
Thus we have indeed φ̈ = v φ . As for the initial conditions, φ(x, 0) = f (x) is clear
from our definition of φ, and φ̇(x, 0) = g(x) is clear from our above formula of φ̇.
(2) Conversely now, we can simply solve our equation, which among others will dou-
blecheck the computations in (1). Let us make the following change of variables:
ξ = x − vt , η = x + vt
With this change of variables, which is quite tricky, mixing space and time variables,
our wave equation φ̈ = v 2 φ′′ reformulates in a very simple way, as follows:
d2 φ
=0
dξdη
But this latter equation tells us that our new ξ, η variables get separated, and we
conclude from this that the solution must be of the following special form:
φ(x, t) = F (ξ) + G(η) = F (x − vt) + G(x + vt)
Now by taking into account the intial conditions φ(x, 0) = f (x) and φ̇(x, 0) = g(x),
and then integrating, we are led to the d’Alembert formula. Finally, in what regards the
last assertion, we will leave the study here as an instructive exercise. □
 4D. SOME PROBABILITY 101

4d. Some probability

As another application of the integration theory developed above, let us develop now
some theoretical probability theory. You probably know, from real life, what probability
is. But in practice, when trying to axiomatize this, in mathematical terms, things can be
quite tricky. So, here comes our point, the definition saving us is as follows:
Definition 4.31. A probability density is a function φ : R → R satisfying
Z
φ≥0 , φ(x)dx = 1
R

with the convention that we allow Dirac masses, δx with x ∈ R, as components of φ.

To be more precise, in what regards the convention at the end, which is something of
physics flavor, this states that our density function φ : R → R must be a combination as
follows, with ψ : R → R being a usual function, and with αi , xi ∈ R:
X
φ=ψ+ α i δx i
i

Assuming that xi are distinct, and with the usual convention that the Dirac masses
integrate up to 1, the conditions on our density function φ : R → R are as follows:
Z X
ψ ≥ 0 , αi ≥ 0 , ψ(x)dx + αi = 1
R i

Observe the obvious relation with intuitive probability theory, where the probability
for something to happen is always positive, P ≥ 0, and where the overall probability for
something to happen, with this meaning for one of the possible events to happen, is of
course ΣP = 1, and this because life goes on, and something must happen, right.

In short, what we are proposing with Definition 4.31 is some sort of continuous gen-
eralization of basic probability theory, coming from coins, dice and cards, that you surely
know. Moving now ahead, let us formulate, as a continuation of Definition 4.31:
Definition 4.32. We say that a random variable f follows the density φ if
Z b
P (f ∈ [a, b]) = φ(x)dx
a

holds, for any interval [a, b] ⊂ R.

With this, we are now one step closer to what we know from coins, dice, cards and so
on. For instance when rolling a die, the corresponding density is as follows:
1
φ= (δ1 + δ2 + δ3 + δ4 + δ5 + δ6 )
6
102 4. INTEGRATION

In what regards now the random variables f , described as above by densities φ, the
first questions regard their mean and variance, constructed as follows:
Definition 4.33. Given a random variable f , with probability density φ:
R
(1) Its mean is the quantity M = R xφ(x) dx.R
(2) More generally, its k-th moment is Mk = R xk φ(x) dx.
(3) Its variance is the quantity V = M2 − M12 .
Before going further, with more theory and examples, let us observe that, in both
Definition 4.32 and Definition 4.33, what really matters is not the density φ itself, but
rather the related quantity µ = φ(x)dx. So, let us upgrade our formalism, as follows:
Definition 4.34 (upgrade). A real probability
R measure isPa quantity of the following
type, with ψ ≥ 0, αi ≥ 0 and xi ∈ R, satisfying R ψ(x)dx + i αi = 1:
X
µ = ψ(x)dx + α i δx i
i
Rb
We say that a random variable f follows µ when P (f ∈ [a, b]) = a
dµ(x). In this case
Z
Mk = xk dµ(x)
R
are called moments of f , and M = M1 and V = M2 − M12 are called mean, and variance.
In practice now, let us look for some illustrations for this. The simplest random
variables are those following discrete laws, ψ = 0, and as a basic example here, when
flipping a coin and being rewarded $0 for heads, and $1 for tails, the corresponding law
is µ = 21 (δ0 + δ1 ). More generally, playing the same game with a biased coin, which lands
on heads with probability p ∈ (0, 1), leads to the following law, called Bernoulli law:
µ = pδ0 + (1 − p)δ1
Many more things can be said here, notably with a study of what happens when you
play the game n times in a row, leading to some sort of powers of the Bernoulli laws, called
binomial laws. Skipping some discussion here, and getting straight to the point, the most
important laws in discrete probability are the Poisson laws, constructed as follows:
Definition 4.35. The Poisson law of parameter 1 is the following measure,
1 X δk
p1 =
e k∈N k!
and more generally, the Poisson law of parameter t > 0 is the following measure,
X tk
pt = e−t δk
k∈N
k!
with the letter “p” standing for Poisson.
 4D. SOME PROBABILITY 103

Observe that our laws have indeed mass 1, as they should, and this due to:
X tk
et =
k∈N
k!
In general, the idea with the Poisson laws is that these appear a bit everywhere, in the
real life, the reasons for this coming from the Poisson Limit Theorem (PLT). However,
this theorem uses advanced calculus, and we will leave it for later. In the meantime,
however, we can have some fun with moments, the result here being as follows:
Theorem 4.36. The moments of p1 are the Bell numbers,
Mk (p1 ) = |P (k)|
where P (k) is the set of partitions of {1, . . . , k}. More generally, we have
X
Mk (pt ) = t|π|
π∈P (k)

for any t > 0, where |.| is the number of blocks.

Proof. The moments of p1 satisfy the following recurrence formula:
1 X (r + 1)k+1
Mk+1 =
e r (r + 1)!
k
1 X rk

1
= 1+
e r r! r
1 X rk X k −s
 
= r
e r r! s s
X k  1 X rk−s
= ·
s
s e r r!
X k  
= Mk−s
s
s
With this done, let us try now to find a recurrence for the Bell numbers, Bk = |P (k)|.
Since a partition of {1, . . . , k + 1} appears by choosing s neighbors for 1, among the k
numbers available, and then partitioning the k − s elements left, we have:
X k 
Bk+1 = Bk−s
s
s
Since the initial values coincide, M1 = B1 = 1 and M2 = B2 = 2, we obtain by
recurrence Mk = Bk , as claimed. Regarding now the law pt with t > 0, we have here a
104 4. INTEGRATION

similar recurrence formula for the moments, as follows:

X tr+1 (r + 1)k+1
−t
Mk+1 = e
r
(r + 1)!
X tr+1 rk  k
−t 1
= e 1+
r
r! r
X tr+1 rk X k  
−t
= e r−s
r
r! s
s
X k   X tr+1 rk−s
= · e−t
s
s r
r!
X k  
= t Mk−s
s
s
Regarding the initial values, the first moment of pt is given by:
X tr r X tr
M1 = e−t = e−t =t
r
r! r
(r − 1)!
Now by using the above recurrence we obtain from this:
X 1
M2 = t Mk−s = t(1 + t) = t + t2
s
s
On the other hand, some standard P combinatorics, a bit as before at t = 1, shows that
|π|
the numbers in the statement Sk = π∈P (k) t satisfy the same recurrence relation, and
with the same initial values. Thus we have Mk = Sk , as claimed. □
The above result looks quite exciting, making a link between subtle probability, and
subtle combinatorics. We will be back to it, after learning more calculus.
4e. Exercises
We had a lot of interesting theory in this chapter, and as exercises, we have:
Exercise 4.37. Clarify the integrability property of piecewise monotone functions.
Exercise 4.38. Find a geometric proof for 13 + . . . + N 3 = (1 + . . . + N )2 .
Exercise 4.39. Find estimates for the remainder, in the Taylor formula.
Exercise 4.40. Work out with full details the formula of Mk (pt ), at t > 0.
As a bonus exercise, compute areas and volumes. As many as you can.
 Part II

Complex functions
 How can I change the world
If I can’t even change myself
How can I change the way I am
I don’t know, I don’t know
 CHAPTER 5

Complex numbers

5a. Complex numbers

In this second part of the present book we discuss the functions of one complex variable
f : C → C. This is certainly something to be done before going to several variables, be
them real or complex, because we have some unfinished business with sin, cos, exp, log,
which remain quite mysterious objects. We will see here that these functions are much
better understood as complex functions f : C → C. In fact, even a dumb polynomial
P ∈ R[X] is better understood as complex polynomial, P ∈ C[X], because its roots, that
might not always exist as real numbers, always exist as complex numbers.

On top of this, we have physics. You might know this or not, but physics is all
about waves, which waves require complex numbers for their understanding. And there is
even more, because the R that we are so used to, from classical mechanics, dealing with
questions at our usual, human scale, naturally becomes, via some sort of complicated
procedure, C in quantum mechanics, that is, when zooming down, where the new science
and technologies are. But more on this later, towards the end of the present book.

In short, many interesting things to be discussed, and this not necessarily for the
sake of doing complex functions, but also with the goal of better understanding the real
functions themselves. Let us begin with the complex numbers. There is a lot of magic
here, and we will carefully explain this material. Their definition is as follows:
Definition 5.1. The complex numbers are variables of the form
x = a + ib
with a, b ∈ R, which add in the obvious way, and multiply according to the following rule:
i2 = −1
Each real number can be regarded as a complex number, a = a + i · 0.
In other words, we consider variables as above, without bothering for the moment
with their precise meaning. Now consider two such complex numbers:
x = a + ib , y = c + id
107
108 5. COMPLEX NUMBERS

The formula for the sum is then the obvious one, as follows:
x + y = (a + c) + i(b + d)
As for the formula of the product, by using the rule i2 = −1, we obtain:
xy = (a + ib)(c + id)
= ac + iad + ibc + i2 bd
= ac + iad + ibc − bd
= (ac − bd) + i(ad + bc)
Thus, the complex numbers as introduced above are well-defined. The multiplica-
tion formula is of course quite tricky, and hard to memorize, but we will see later some
alternative ways, which are more conceptual, for performing the multiplication.

The advantage of using the complex numbers comes from the fact that the equation
2
x = 1 has now a solution, x = i. In fact, this equation has two solutions, namely:
x = ±i
This is of course very good news. More generally, we have the following result, regard-
ing the arbitrary degree 2 equations, with real coefficients:
Theorem 5.2. The complex solutions of ax2 + bx + c = 0 with a, b, c ∈ R are
√
−b ± b2 − 4ac
x1,2 =
2a
with the square root of negative real numbers being defined as
√ √
−m = ±i m
and with the square root of positive real numbers being the usual one.
Proof. We can write our equation in the following way:
b c
ax2 + bx + c = 0 ⇐⇒ x2 + x + = 0
a a
2
b2

b c
⇐⇒ x+ − 2 + =0
2a 4a a
 2 2
b b − 4ac
⇐⇒ x+ =
2a 4a2
√
b b2 − 4ac
⇐⇒ x + =±
2a 2a
Thus, we are led to the conclusion in the statement. □
 5A. COMPLEX NUMBERS 109

We will see later that any degree 2 complex equation has solutions as well, and that
more generally, any polynomial equation, real or complex, has solutions. Moving ahead
now, we can represent the complex numbers in the plane, in the following way:
Proposition 5.3. The complex numbers, written as usual
x = a + ib
can be represented in the plane, according to the following identification:
 
a
x=
b
With this convention, the sum of complex numbers is the usual sum of vectors.
Proof. Consider indeed two arbitrary complex numbers:
x = a + ib , y = c + id
Their sum is then by definition the following complex number:
x + y = (a + c) + i(b + d)
Now let us represent x, y in the plane, as in the statement:
   
a c
x= , y=
b d
In this picture, their sum is given by the following formula:
 
a+c
x+y =
b+d
But this is indeed the vector corresponding to x + y, so we are done. □
Here we have assumed that you are a bit familiar with vector calculus. If not, no
problem, the idea is simply that vectors add by forming a parallelogram, as follows:
O

b+d •x+y

d •y

b •x

• /

c a a+c
110 5. COMPLEX NUMBERS

Observe that in our geometric picture from Proposition 5.3, the real numbers corre-
spond to the numbers on the Ox axis. As for the purely imaginary numbers, these lie on
the Oy axis, with the number i itself being given by the following formula:
 
0
i=
1
As an illustration for this, let us record now a basic picture, with some key complex
numbers, namely 1, i, −1, −i, represented according to our conventions:
O

•i

•−1 •1 /

•−i

You might perhaps wonder why I chose to draw that circle, connecting the numbers
1, i, −1, −i, which does not look very useful. More on this in a moment, the idea being
that that circle can be immensely useful, and coming in advance, some advice:
Advice 5.4. When drawing complex numbers, always begin with the coordinate axes
Ox, Oy, and with a copy of the unit circle.
We have so far a quite good understanding of their complex numbers, and their ad-
dition. In order to understand now the multiplication operation, we must do something
more complicated, namely using polar coordinates. Let us start with:
Definition 5.5. The complex numbers x = a + ib can be written in polar coordinates,
x = r(cos t + i sin t)
with the connecting formulae being as follows,
a = r cos t , b = r sin t
and in the other sense being as follows,
√ b
r = a2 + b 2 , tan t =
a
and with r, t being called modulus, and argument.
 5A. COMPLEX NUMBERS 111

There is a clear relation here with the vector notation from Proposition 5.3, because
r is the length of the vector, and t is the angle made by the vector with the Ox axis. To
be more precise, the picture for what is going on in Definition 5.5 is as follows:
O

b •x
r
t

• /

a
As a basic example here, the number i takes the following form:
π  π 
i = cos + i sin
2 2
The point now is that in polar coordinates, the multiplication formula for the complex
numbers, which was so far something quite opaque, takes a very simple form:
Theorem 5.6. Two complex numbers written in polar coordinates,
x = r(cos s + i sin s) , y = p(cos t + i sin t)
multiply according to the following formula:
xy = rp(cos(s + t) + i sin(s + t))
In other words, the moduli multiply, and the arguments sum up.
Proof. This follows from the following formulae, that we know well:
cos(s + t) = cos s cos t − sin s sin t
sin(s + t) = cos s sin t + sin s cos t
Indeed, we can assume that we have r = p = 1, by dividing everything by these
numbers. Now with this assumption made, we have the following computation:
xy = (cos s + i sin s)(cos t + i sin t)
= (cos s cos t − sin s sin t) + i(cos s sin t + sin s cos t)
= cos(s + t) + i sin(s + t)
Thus, we are led to the conclusion in the statement. □
The above result, which is based on some non-trivial trigonometry, is quite powerful.
As a basic application of it, we can now compute powers, as follows:
112 5. COMPLEX NUMBERS

Theorem 5.7. The powers of a complex number, written in polar form,

x = r(cos t + i sin t)
are given by the following formula, valid for any exponent k ∈ N:
xk = rk (cos kt + i sin kt)
Moreover, this formula holds in fact for any k ∈ Z, and even for any k ∈ Q.
Proof. Given a complex number x, written in polar form as above, and an exponent
k ∈ N, we have indeed the following computation, with k terms everywhere:
xk = x . . . x
= r(cos t + i sin t) . . . r(cos t + i sin t)
= rk ([cos(t + . . . + t) + i sin(t + . . . + t))
= rk (cos kt + i sin kt)
Thus, we are done with the case k ∈ N. Regarding now the generalization to the case
k ∈ Z, it is enough here to do the verification for k = −1, where the formula is:
x−1 = r−1 (cos(−t) + i sin(−t))
But this number x−1 is indeed the inverse of x, as shown by:
xx−1 = r(cos t + i sin t) · r−1 (cos(−t) + i sin(−t))
= cos(t − t) + i sin(t − t)
= cos 0 + i sin 0
= 1
Finally, regarding the generalization to the case k ∈ Q, it is enough to do the verifi-
cation for exponents of type k = 1/n, with n ∈ N. The claim here is that:
    
1/n 1/n t t
x =r cos + i sin
n n
In order to prove this, let us compute the n-th power of this number. We can use the
power formula for the exponent n ∈ N, that we already established, and we obtain:
    
1/n n 1/n n t t
(x ) = (r ) cos n · + i sin n ·
n n
= r(cos t + i sin t)
= x
Thus, we have indeed a n-th root of x, and our proof is now complete. □
 5A. COMPLEX NUMBERS 113

We should mention that there is a bit of ambiguity in the above, in the case of the
exponents k ∈ Q, due to the fact that the square roots, and the higher roots as well, can
take multiple values, in the complex number setting. We will be back to this.

As a basic application of Theorem 5.7, we have the following result:

Proposition 5.8. Each complex number, written in polar form,
x = r(cos t + i sin t)
has two square roots, given by the following formula:
√ √
    
t t
x = ± r cos + i sin
2 2
√ √
When x > 0, these roots are ± x. When x < 0, these roots are ±i −x.
Proof. The first assertion is clear indeed from the general formula in Theorem 5.7,
at k = 1/2. As for its particular cases with x ∈ R, these are clear from it. □
√
As a comment here, for x > 0 we are very used to call the usual x square root of x.
However, for x < 0,√or more generally for x ∈ C − R+ , there is less interest in choosing
one of the possible x and calling it “the” square root of x, because all this is based on
our convention that i comes up, instead of down, which is something rather arbitrary.
Actually, clocks turning clockwise, i should be rather coming down. All this is a matter
of taste, but in any case, for our math, the best is to keep some ambiguity, as above.

With the above results in hand, and notably with the square root formula from Propo-
sition 5.8, we can now go back to the degree 2 equations, and we have:
Theorem 5.9. The complex solutions of ax2 + bx + c = 0 with a, b, c ∈ C are
√
−b ± b2 − 4ac
x1,2 =
2a
with the square root of complex numbers being defined as above.
Proof. This is clear, the computations being the same as in the real case. To be
more precise, our degree 2 equation can be written as follows:
2
b2 − 4ac

b
x+ =
2a 4a2
Now since we know from Proposition 5.8 that any complex number has a square root,
we are led to the conclusion in the statement. □
As a last general topic regarding the complex numbers, let us discuss conjugation.
This is something quite tricky, complex number specific, as follows:
114 5. COMPLEX NUMBERS

Definition 5.10. The complex conjugate of x = a + ib is the following number,

x̄ = a − ib
obtained by making a reflection with respect to the Ox axis.
As before with other such operations on complex numbers, a quick picture says it all.
Here is the picture, with the numbers x, x̄, −x, −x̄ being all represented:

•−x̄ •x
r
t

• /

•−x •x̄

Observe that the conjugate of a real number x ∈ R is the number itself, x = x̄. In
fact, the equation x = x̄ characterizes the real numbers, among the complex numbers. At
the level of non-trivial examples now, we have the following formula:
ī = −i
There are many things that can be said about the conjugation of the complex numbers,
and here is a summary of basic such things that can be said:
Theorem 5.11. The conjugation operation x → x̄ has the following properties:
(1) x = x̄ precisely when x is real.
(2) x = −x̄ precisely when x is purely imaginary.
(3) xx̄ = |x|2 , with |x| = r being as usual the modulus.
(4) With x = r(cos t + i sin t), we have x̄ = r(cos t − i sin t).
(5) We have the formula xy = x̄ȳ, for any x, y ∈ C.
(6) The solutions of ax2 + bx + c = 0 with a, b, c ∈ R are conjugate.
Proof. These results are all elementary, the idea being as follows:
(1) This is something that we already know, coming from definitions.
(2) This is something clear too, because with x = a + ib our equation x = −x̄ reads
a + ib = −a + ib, and so a = 0, which amounts in saying that x is purely imaginary.
 5B. EXPONENTIAL WRITING 115

(3) This is a key formula, which can be proved as follows, with x = a + ib:
xx̄ = (a + ib)(a − ib)
= a2 + b 2
= |x|2
(4) This is clear indeed from the picture following Definition 5.10.
(5) This is something quite magic, which can be proved as follows:
(a + ib)(c + id) = (ac − bd) + i(ad + bc)
= (ac − bd) − i(ad + bc)
= (a − ib)(c − id)
However, what we have been doing here is not very clear, geometrically speaking,
and our formula is worth an alternative proof. Here is that proof, which after inspection
contains no computations at all, making it clear that the polar writing is the best:
r(cos s + i sin s) · p(cos t + i sin t)
= rp(cos(s + t) + i sin(s + t))
= rp(cos(−s − t) + i sin(−s − t))
= r(cos(−s) + i sin(−s)) · p(cos(−t) + i sin(−t))
= r(cos s + i sin s) · p(cos t + i sin t)
(6) This comes from the formula of the solutions, that we know from Theorem 5.2,
but we can deduce this as well directly, without computations. Indeed, by using our
assumption that the coefficients are real, a, b, c ∈ R, we have:
ax2 + bx + c = 0 =⇒ ax2 + bx + c = 0
=⇒ āx̄2 + b̄x̄ + c̄ = 0
=⇒ ax̄2 + bx̄ + c = 0
Thus, we are led to the conclusion in the statement. □

5b. Exponential writing

We discuss now the theory of complex functions f : C → C, in analogy with the theory
of the real functions f : R → R. We will see that many results that we know from the
real setting extend to the complex setting. Before starting, two remarks on this:

(1) Most of the real functions f : R → R that we know, such as sin, cos, exp, log,
extend into complex functions f : C → C, and the study of these latter extensions brings
some new light on the original real functions. Thus, what we will be doing here will be,
in a certain sense, a refinement of the theory developed in chapters 1-4.
116 5. COMPLEX NUMBERS

(2) On the other hand, since we have C ≃ R2 , the complex functions f : C → C that
we will study here can be regarded as functions f : R2 → R2 . This is something quite
subtle, but in any case, what we will be doing here will stand as well as an introduction
to the functions of type f : RN → RM , that we will study in chapters 9-16 below.
In short, one complex variable is something in between one real variable, and two
or more real variables, and we can only expect to end up with a mysterious mixture of
surprising and unsurprising results. Welcome to complex analysis. Let us start with:
Definition 5.12. A complex function f : C → C, or more generally f : X → C, with
X ⊂ C being a subset, is called continuous when, for any xn , x ∈ X:
xn → x =⇒ f (xn ) → f (x)
where the convergence of the sequences of complex numbers, xn → x, means by definition
that for n big enough, the quantity |xn − x| becomes arbitrarily small.
Observe that in real coordinates, x = (a, b), the distances appearing in the definition
of the convergence xn → x are given by the following formula:
p
|xn − x| = (an − a)2 + (bn − b)2
Thus xn → x in the complex sense means that (an , bn ) → (a, b) in the usual, intuitive
sense, with respect to the usual distance in the plane R2 , and as a consequence, a function
f : C → C is continuous precisely when it is continuous, in an intuitive sense, when
regarded as function f : R2 → R2 . But more on this later, in chapters 9-10 below.
At the level of examples, we have the following result:
Theorem 5.13. We can exponentiate the complex numbers, according to the formula
∞
x
X xk
e =
k=0
k!
and the function x → ex is continuous, and satisfies ex+y = ex ey .
Proof. We must first prove that the series converges. But this follows from:
∞
x
X xk
|e | =
k=0
k!
∞
X xk
≤
k=0
k!
∞
X |x|k
=
k=0
k!
|x|
= e <∞
 5B. EXPONENTIAL WRITING 117

Regarding the formula ex+y = ex ey , this follows too as in the real case, as follows:
∞
x+y
X (x + y)k
e =
k=0
k!
∞ X k  
X k xs y k−s
= ·
k=0 s=0
s k!
∞ X k
X xs y k−s
=
k=0 s=0
s!(k − s)!
= ex ey
Finally, the continuity of x → ex comes at x = 0 from the following computation:
∞
X tk
|et − 1| =
k=1
k!
∞
X tk
≤
k=1
k!
∞
X |t|k
=
k=1
k!
|t|
= e −1
x
As for the continuity of x → e in general, this can be deduced now as follows:
lim ex+t = lim ex et = ex lim et = ex · 1 = ex
t→0 t→0 t→0

Thus, we are led to the conclusions in the statement. □

We will be back to more functions later. As an important fact, however, let us point
out that, contrary to what the above might suggest, everything does not always extend
trivally from the real to the complex case. For instance, we have:
Proposition 5.14. We have the following formula, valid for any |x| < 1,
1
= 1 + x + x2 + . . .
1−x
but, unlike in the real case, the geometric meaning of this formula is quite unclear.
Proof. Here the formula in the statement holds indeed, by multiplying and cancelling
terms, and with the convergence being justified by the following estimate:
∞ ∞
X X 1
xn ≤ |x|n =
n=0 n=0
1 − |x|
118 5. COMPLEX NUMBERS

As for the last assertion, this is something quite informal. To be more precise, for
x = 1/2 our formula is clear, by cutting the interval [0, 2] into half, and so on:
1 1 1
1 + + + + ... = 2
2 4 8
More generally, for x ∈ (−1, 1) the meaning of the formula in the statement is some-
thing quite clear and intuitive, geometrically speaking, by using a similar argument. How-
ever, when x is complex, and not real, we are led into a kind of mysterious spiral there,
and the only case where the formula is “obvious”, geometrically speaking, is that when
x = rw, with r ∈ [0, 1), and with w being a root of unity. To be more precise here, by
anticipating a bit, assume that we have a number w ∈ C satisfying wN = 1, for some
N ∈ N. We have then the following formula, for our infinite sum:
1 + rw + r2 w2 + . . . = (1 + rw + . . . + rN −1 wN −1 )
+ (rN + rN +1 w . . . + r2N −1 wN −1 )
+ (r2N + r2N +1 w . . . + r3N −1 wN −1 )
+ ...
Thus, by grouping the terms with the same argument, our infinite sum is:
1 + rw + r2 w2 + . . . = (1 + rN + r2N + . . .)
+ (r + rN +1 + r2N +1 + . . .)w
+ ...
+ (rN −1 + r2N −1 + r3N −1 + . . .)wN −1
But the sums of each ray can be computed with the real formula for geometric series,
that we know and understand well, and with an extra bit of algebra, we get:
1 rw rN −1 wN −1
1 + rw + r2 w2 + . . . = + + . . . +
1 − rN 1 − rN 1 − rN
1
1 + rw + . . . + rN −1 wN −1


= N
1−r
1 1 − rN
= ·
1 − rN 1 − rw
1
=
1 − rw
Summarizing, as claimed above, the geometric series formula can be understood, in
a purely geometric way, for variables of type x = rw, with r ∈ [0, 1), and with w being
a root of unity. In general, however, this formula tells us that the numbers on a certain
infinite spiral sum up to a certain number, which remains something quite mysterious. □
Getting back now to less mysterious mathematics, which in fact will turn to be quite
mysterious as well, as is often the case with things involving complex numbers, as an
 5B. EXPONENTIAL WRITING 119

application of all this, let us discuss the final and most convenient writing of the complex
numbers, which is a variation on the polar writing, as follows:
x = reit
The point with this formula comes from the following deep result:
Theorem 5.15. We have the following formula,
eit = cos t + i sin t
valid for any t ∈ R.
Proof. Our claim is that this follows from the formula of the complex exponential,
and for the following formulae for the Taylor series of cos and sin, that we know well:
∞ 2l ∞ 2l+1
l t l t
X X
cos t = (−1) , sin t = (−1)
l=0
(2l)! l=0
(2l + 1)!

Indeed, let us first recall from Theorem 5.13 that we have the following formula, for
the exponential of an arbitrary complex number x ∈ C:
∞
x
X xk
e =
k=0
k!

Now let us plug x = it in this formula. We obtain the following formula:

∞
X (it)k
eit =
k=0
k!
X (it)k X (it)k
= +
k=2l
k! k=2l+1
k!
∞ ∞
X t2l l
X t2l+1
= (−1) +i (−1)l
l=0
(2l)! l=0
(2l + 1)!
= cos t + i sin t
Thus, we are led to the conclusion in the statement. □

As a main application of the above formula, we have:

Theorem 5.16. We have the following formula,
eπi = −1
and we have E = mc2 as well.
120 5. COMPLEX NUMBERS

Proof. We have two assertions here, the idea being as follows:

(1) The first formula, eπi = −1, which is actually the main formula in mathematics,
comes from Theorem 5.15, by setting t = π. Indeed, we obtain:
eπi = cos π + i sin π
= −1 + i · 0
= −1
(2) As for E = mc2 , which is the main formula in physics, this is something deep too.
Although we will not really need it here, we recommend learning it as well, for symmetry
reasons between math and physics, say from Feynman [31], [32], [33]. □
Now back to our x = reit objectives, with the above theory in hand we can indeed use
from now on this notation, the complete statement being as follows:
Theorem 5.17. The complex numbers x = a + ib can be written in polar coordinates,
x = reit
with the connecting formulae being
a = r cos t , b = r sin t
and in the other sense being
√ b
r= a2 + b 2 , tan t =
a
and with r, t being called modulus, and argument.
Proof. This is a reformulation of our previous Definition 5.5, by using the formula
eit = cos t + i sin t from Theorem 5.15, and multiplying everything by r. □
With this in hand, we can now go back to the basics, namely the addition and multi-
plication of the complex numbers. We have the following result:
Theorem 5.18. In polar coordinates, the complex numbers multiply as
reis · peit = rp ei(s+t)
with the arguments s, t being taken modulo 2π.
Proof. This is something that we already know, from Theorem 5.6, reformulated by
using the notations from Theorem 5.17. Observe that this follows as well directly, from
the fact that we have ea+b = ea eb , that we know from analysis. □
The above formula is obviously very powerful. However, in polar coordinates we do
not have a simple formula for the sum. Thus, this formalism has its limitations.

We can investigate as well more complicated operations, as follows:

 5C. EQUATIONS, ROOTS 121

Theorem 5.19. We have the following operations on the complex numbers, written
in polar form, as above:
it −1
(1) Inversion: (re√ ) = r−1 e−it .
√
(2) Square roots: reit = ± reit/2 .
(3) Powers: (reit )a = ra eita .
(4) Conjugation: reit = re−it .
Proof. This is something that we already know, from Theorem 5.7, but we can now
discuss all this, from a more conceptual viewpoint, the idea being as follows:
(1) We have indeed the following computation, using Theorem 5.18:
(reit )(r−1 e−it ) = rr−1 · ei(t−t)
= 1·1
= 1
(2) Once again by using Theorem 5.18, we have:
√ √
(± reit/2 )2 = ( r)2 ei(t/2+t/2) = reit
(3) Given an arbitrary number a ∈ R, we can define, as stated:
(reit )a = ra eita
Due to Theorem 5.18, this operation x → xa is indeed the correct one.
(4) This comes from the fact, that we know from Theorem 5.11, that the conjugation
operation x → x̄ keeps the modulus, and switches the sign of the argument. □

5c. Equations, roots

Getting back to algebra, recall from Theorem 5.9 that any degree 2 equation has 2
complex roots. We can in fact prove that any polynomial equation, of arbitrary degree
N ∈ N, has exactly N complex solutions, counted with multiplicities:
Theorem 5.20. Any polynomial P ∈ C[X] decomposes as
P = c(X − a1 ) . . . (X − aN )
with c ∈ C and with a1 , . . . , aN ∈ C.
Proof. The problem is that of proving that our polynomial has at least one root,
because afterwards we can proceed by recurrence. We prove this by contradiction. So,
assume that P has no roots, and pick a number z ∈ C where |P | attains its minimum:
|P (z)| = min |P (x)| > 0
x∈C
122 5. COMPLEX NUMBERS

Since Q(t) = P (z + t) − P (z) is a polynomial which vanishes at t = 0, this polynomial

must be of the form ctk + higher terms, with c ̸= 0, and with k ≥ 1 being an integer. We
obtain from this that, with t ∈ C small, we have the following estimate:
P (z + t) ≃ P (z) + ctk
Now let us write t = rw, with r > 0 small, and with |w| = 1. Our estimate becomes:
P (z + rw) ≃ P (z) + crk wk
Now recall that we assumed P (z) ̸= 0. We can therefore choose w ∈ T such that cwk
points in the opposite direction to that of P (z), and we obtain in this way:
|P (z + rw)| ≃ |P (z) + crk wk |
= |P (z)|(1 − |c|rk )
Now by choosing r > 0 small enough, as for the error in the first estimate to be small,
and overcame by the negative quantity −|c|rk , we obtain from this:
|P (z + rw)| < |P (z)|
But this contradicts our definition of z ∈ C, as a point where |P | attains its minimum.
Thus P has a root, and by recurrence it has N roots, as stated. □
All this is very nice, and we will see applications in a moment. As a word of warning,
however, we should mention that the above result remains something quite theoretical.
Indeed, the proof is by contradiction, and there is no way of recycling the material there
into something explicit, that can be used for effectively computing the roots.

Still talking polynomials and their roots, let us try however to understand what the
analogue of ∆ = b2 − 4ac is, for an arbitrary polynomial P ∈ C[X]. We will need:
Theorem 5.21. Given two polynomials P, Q ∈ C[X], written as follows,
P = c(X − a1 ) . . . (X − ak ) , Q = d(X − b1 ) . . . (X − bl )
the following quantity, which is called resultant of P, Q,
Y
R(P, Q) = cl dk (ai − bj )
ij

is a polynomial in the coefficients of P, Q, with integer coefficients, and we have

R(P, Q) = 0
precisely when P, Q have a common root.
Proof. Given P, Q ∈ C[X], we can certainly construct the quantity R(P, Q) in the
statement, and we have then R(P, Q) = 0 precisely when P, Q have a common root. The
whole point is that of proving that R(P, Q) is a polynomial in the coefficients of P, Q,
with integer coefficients. But this can be checked as follows:
 5C. EQUATIONS, ROOTS 123

(1) We can expand the formula of R(P, Q), and in what regards a1 , . . . , ak , which are
the roots of P , we obtain in this way certain symmetric functions in these variables, which
will be therefore polynomials in the coefficients of P , with integer coefficients.
(2) We can then look what happens with respect to the remaining variables b1 , . . . , bl ,
which are the roots of Q. Once again what we have here are certain symmetric functions,
and so polynomials in the coefficients of Q, with integer coefficients.
(3) Thus, we are led to the conclusion in the statement, that R(P, Q) is a polynomial
in the coefficients of P, Q, with integer coefficients, and with the remark that the cl dk
factor is there for these latter coefficients to be indeed integers, instead of rationals. □
All this might seem a bit complicated, and as an illustration, let us work out an
example. Consider the case of a polynomial of degree 2, and a polynomial of degree 1:
P = ax2 + bx + c , Q = dx + e
In order to compute the resultant, let us factorize our polynomials:
P = a(x − p)(x − q) , Q = d(x − r)
The resultant can be then computed as follows, by using the method above:
R(P, Q) = ad2 (p − r)(q − r)
= ad2 (pq − (p + q)r + r2 )
= cd2 + bd2 r + ad2 r2
= cd2 − bde + ae2
Finally, observe that R(P, Q) = 0 corresponds indeed to the fact that P, Q have a
common root. Indeed, the root of Q is r = −e/d, and we have:
ae2 be R(P, Q)
P (r) = − + c =
d2 d d2
Thus P (r) = 0 precisely when R(P, Q) = 0, as predicted by Theorem 5.21.
With this, we can now talk about the discriminant of any polynomial, as follows:
Theorem 5.22. Given a polynomial P ∈ C[X], written as
P (X) = cX N + dX N −1 + . . .
its discriminant, defined as being the following quantity,
N
(−1)( 2 )
∆(P ) = R(P, P ′ )
c
is a polynomial in the coefficients of P , with integer coefficients, and
∆(P ) = 0
happens precisely when P has a double root.
124 5. COMPLEX NUMBERS

Proof. This follows from Theorem 5.21, applied with P = Q, with the division by c
being indeed possible, under Z, and with the sign being there for various reasons, including
the compatibility with some well-known formulae, at small values of N ∈ N. □
As an illustration, let us see what happens in degree 2. Here we have:
P = aX 2 + bX + c , P ′ = 2aX + b
Thus, the resultant is given by the following formula:
R(P, P ′ ) = ab2 − b(2a)b + c(2a)2
= 4a2 c − ab2
= −a(b2 − 4ac)
With the normalizations in Theorem 5.22 made, we obtain, as we should:
∆(P ) = b2 − 4ac
As another illustration, let us work out what happens in degree 3. Here the result,
which is useful and interesting, and is probably new to you, is as follows:
Theorem 5.23. The discriminant of a degree 3 polynomial,
P = aX 3 + bX 2 + cX + d
is the number ∆(P ) = b2 c2 − 4ac3 − 4b3 d − 27a2 d2 + 18abcd.
Proof. We need to do some tough computations here. Let us first compute resultants.
Consider two polynomials, of degree 3 and degree 2, written as follows:
P = aX 3 + bX 2 + cX + d = a(X − p)(X − q)(X − r)
Q = eX 2 + f X + g = e(X − s)(X − t)
The resultant of these two polynomials is then given by:
R(P, Q) = a2 e3 (p − s)(p − t)(q − s)(q − t)(r − s)(r − t)
= a2 · e(p − s)(p − t) · e(q − s)(q − t) · e(r − s)(r − t)
= a2 Q(p)Q(q)Q(r)
= a2 (ep2 + f p + g)(eq 2 + f q + g)(er2 + f r + g)
By expanding, we obtain the following formula for this resultant:
R(P, Q)
= e3 p2 q 2 r2 + e2 f (p2 q 2 r + p2 qr2 + pq 2 r2 )
a2
+ e2 g(p2 q 2 + p2 r2 + q 2 r2 ) + ef 2 (p2 qr + pq 2 r + pqr2 )
+ ef g(p2 q + pq 2 + p2 r + pr2 + q 2 r + qr2 ) + f 3 pqr
+ eg 2 (p2 + q 2 + r2 ) + f 2 g(pq + pr + qr)
+ f g 2 (p + q + r) + g 3
 5C. EQUATIONS, ROOTS 125

Note in passing that we have 27 terms on the right, as we should, and with this kind
of check being mandatory, when doing such computations. Next, we have:
b c d
p+q+r =− , pq + pr + qr = , pqr = −
a a a
By using these formulae, we can produce some more, as follows:
b2 2c
p2 + q 2 + r2 = (p + q + r)2 − 2(pq + pr + qr) = 2
−
a a
bc 3d
p2 q + pq 2 + p2 r + pr2 + q 2 r + qr2 = (p + q + r)(pq + pr + qr) − 3pqr = − +
a2 a
c2 2bd
p2 q 2 + p2 r2 + q 2 r2 = (pq + pr + qr)2 − 2pqr(p + q + r) = −
a2 a2
By plugging now this data into the formula of R(P, Q), we obtain:
2
 2 
2 3 d 2 2 cd 2 2 c 2bd bd
R(P, Q) = a e · 2 − a e f · 2 + a e g 2
− 2 + a2 ef 2 · 2
a a a a a
 
bc 3d d
+ a2 ef g − 2 + − a2 f 3 ·
a a a
 2 
b 2c c b
+ a2 eg 2 2
− + a2 f 2 g · − a2 f g 2 · + a2 g 3
a a a a
Thus, we have the following formula for the resultant:
R(P, Q) = d2 e3 − cde2 f + c2 e2 g − 2bde2 g + bdef 2 − bcef g + 3adef g
− adf 3 + b2 eg 2 − 2aceg 2 + acf 2 g − abf g 2 + a2 g 3
Getting back now to our discriminant problem, with Q = P ′ , which corresponds to
e = 3a, f = 2b, g = c, we obtain the following formula:
R(P, P ′ ) = 27a3 d2 − 18a2 bcd + 9a2 c3 − 18a2 bcd + 12ab3 d − 6ab2 c2 + 18a2 bcd
− 8ab3 d + 3ab2 c2 − 6a2 c3 + 4ab2 c2 − 2ab2 c2 + a2 c3
By simplifying terms, and dividing by a, we obtain the following formula:
−∆(P ) = 27a2 d2 − 18abcd + 4ac3 + 4b3 d − b2 c2
But this gives the formula in the statement, and we are done. □
Still talking degree 3 equations, let us try to solve P = 0, with P = aX 3 +bX 2 +cX +d
as above. By linear transformations we can assume a = 1, b = 0, and then it is convenient
to write c = 3p, d = 2q. Thus, our equation becomes x3 + 3px + 2q = 0, and regarding
such equations, we have the following famous result, due to Cardano:
126 5. COMPLEX NUMBERS

Theorem 5.24. For a normalized degree 3 equation, namely

x3 + 3px + 2q = 0
the discriminant is ∆ = −108(p3 + q 2 ). Assuming p, q ∈ R and ∆ < 0, the number
q p q p
3 3
x = −q + p + q + −q − p3 + q 2
3 2

is a real solution of our equation.

Proof. The formula of ∆ is clear from definitions, and with 108 = 4 × 27. Now with
x as in the statement, by using (a + b)3 = a3 + b3 + 3ab(a + b), we have:
q p q p 3
3 3 3
x = −q + p3 + q 2 + −q − p3 + q 2
q p q p
3 3
= −2q + 3 −q + p + q · −q − p3 + q 2 · x
3 2
p
= −2q + 3 3 q 2 − p3 − q 2 · x
= −2q − 3px
Thus, we are led to the conclusion in the statement. □
There are many more things that can be said about degree 3 equations, along these
lines, and we will certainly have an exercise about this, at the end of this chapter.
5d. Roots of unity
We kept the best for the end. As a last topic regarding the complex numbers, which
is something really beautiful, we have the roots of unity. Let us start with:
Theorem 5.25. The equation xN = 1 has N complex solutions, namely
n o
wk k = 0, 1, . . . , N − 1 , w = e2πi/N
which are called roots of unity of order N .
Proof. This follows from the general multiplication formula for complex numbers
from Theorem 5.16. Indeed, with x = reit our equation reads:
rN eitN = 1
Thus r = 1, and t ∈ [0, 2π) must be a multiple of 2π/N , as stated. □
As an illustration here, the roots of unity of small order, along with some of their
basic properties, which are very useful for computations, are as follows:
N = 1. Here the unique root of unity is 1.
N = 2. Here we have two roots of unity, namely 1 and −1.
N = 3. Here we have 1, then w = e2πi/3 , and then w2 = w̄ = e4πi/3 .
 5D. ROOTS OF UNITY 127

N = 4. Here the roots of unity, read as usual counterclockwise, are 1, i, −1, −i.
N = 5. Here, with w = e2πi/5 , the roots of unity are 1, w, w2 , w3 , w4 .
N = 6. Here a useful alternative writing is {±1, ±w, ±w2 }, with w = e2πi/3 .
N = 7. Here, with w = e2πi/7 , the roots of unity are 1, w, w2 , w3 , w4 , w5 , w6 .
N = 8. Here the roots of unity, read as usual counterclockwise, are the√numbers
1, w, i, iw, −1, −w, −i, −iw, with w = eπi/4 , which is also given by w = (1 + i)/ 2.
The roots of unity are very useful variables, and have many interesting properties. As
a first application, we can now solve the ambiguity questions related to the extraction of
N -th roots, from Theorem 5.7 and Theorem 5.19, the statement being as follows:
Theorem 5.26. Any nonzero complex number, written as
x = reit
has exactly N roots of order N , which appear as
y = r1/N eit/N
multiplied by the N roots of unity of order N .
Proof. We must solve the equation z N = x, over the complex numbers. Since the
number y in the statement clearly satisfies y N = x, our equation is equivalent to:
zN = yN
Now observe that we can write this equation as follows:
 N
z
=1
y
We conclude that the solutions z appear by multiplying y by the solutions of tN = 1,
which are the N -th roots of unity, as claimed. □
The roots of unity appear in connection with many other interesting questions, and
there are many useful formulae relating them, which are good to know. Here is a basic
such formula, very beautiful, to be used many times in what follows:
Theorem 5.27. The roots of unity, {wk } with w = e2πi/N , have the property
N
X −1
(wk )s = N δN |s
k=0

for any exponent s ∈ N, where on the right we have a Kronecker symbol.

128 5. COMPLEX NUMBERS

Proof. The numbers in the statement, when written more conveniently as (ws )k with
k = 0, . . . , N − 1, form a certain regular polygon in the plane Ps . Thus, if we denote by
Cs the barycenter of this polygon, we have the following formula:
N −1
1 X ks
w = Cs
N k=0
Now observe that in the case N/| s our polygon Ps is non-degenerate, circling around
the unit circle, and having center Cs = 0. As for the case N |s, here the polygon is
degenerate, lying at 1, and having center Cs = 1. Thus, we have the following formula:
Cs = δN |s
Thus, we obtain the formula in the statement. □
As an interesting philosophical fact, regarding the roots of unity, and the complex
numbers in general, we can now solve the following equation, in a “uniform” way:
x1 + . . . + x N = 0
With this being not a joke. Frankly, can you find some nice-looking family of real
numbers x1 , . . . , xN satisfying x1 + . . . + xN = 0? Certainly not. But with complex
numbers we have now our answer, the sum of the N -th roots of unity being zero.
This was for our basic presentation of the complex numbers. We will be back to more
theory regarding them, and the roots of unity, later on. Among others, we will see later
some non-trivial applications of our above solution to x1 + . . . + xN = 0.
5e. Exercises
This was all beautiful material, and our exercises will be beautiful as well:
Exercise 5.28. Solve some easy plane geometry problems, using complex numbers.
Exercise 5.29. Solve more plane geometry problems, using complex numbers, this
time of scary type, involving circles, square angles, and equilateral triangles.
Exercise 5.30. Further build on the Cardano formula presented above.
Exercise 5.31. Diagonalize, beautifully, the all-one matrix.
As bonus exercise, already mentioned in the above, learn some physics, as much as
you can. This is where the complex numbers really shine.
 CHAPTER 6

Complex functions

6a. Functions, continuity

We discuss in this chapter and in the next two ones the theory of complex functions
f : C → C, in analogy with the theory of the real functions f : R → R. We will see that
many results that we know from the real setting extend to the complex setting, but there
will be quite a number of new phenomena too. We will need, in order to get started:
Definition 6.1. The distance between two complex numbers is the usual distance in
the plane between them, namely:
d(x, y) = |x − y|
With this, we can talk about convergence, by saying that xn → x when d(xn , x) → 0.
Here the fact that d(x, y) = |x − y| is indeed the usual distance in the plane is clear for
y = 0, because we have d(x, 0) = |x|, by definition of the modulus |x|. As for the general
case, y ∈ C, this comes from the fact that the distance in the plane is given by:
d(x, y) = d(x − y, 0) = |x − y|
Observe that in real coordinates, the distance formula is quite complicated, namely:
d(a + ib, c + id) = |(a + ib) − (c + id)|
= |(a − c) + i(b − d)|
p
= (a − c)2 + (b − d)2
However, for most computations, we will not need this formula, and we can get away
with the various tricks regarding complex numbers that we know. As a first result now,
regarding C and its distance, that we will need in what follows, we have:
Proposition 6.2. The complex plane C is complete, in the sense that any Cauchy
sequence converges.
Proof. Consider indeed a Cauchy sequence {xn }n∈N ⊂ C. If we write xn = an + ibn
for any n ∈ N, then we have the following estimates:
p
|an − am | ≤ (an − am )2 + (bn − bm )2 = |xn − xm |
p
|bn − bm | ≤ (an − am )2 + (bn − bm )2 = |xn − xm |
129
130 6. COMPLEX FUNCTIONS

Thus both the sequences {an }n∈N ⊂ R and {bn }n∈N ⊂ R are Cauchy, and since we
know that R itself is complete, we can consider the limits of these sequences:
an → a , bn → b
With x = a + ib, our claim is that xn → x. Indeed, we have:
p
|xn − x| = (an − a)2 + (bn − b)2
≤ |an − a| + |bn − b|
It follows that we have xn → x, as claimed, and this gives the result. □
Talking complex functions now, we have the following definition:
Definition 6.3. A complex function f : C → C, or more generally f : X → C, with
X ⊂ C being a subset, is called continuous when, for any xn , x ∈ X:
xn → x =⇒ f (xn ) → f (x)
Also, we can talk about pointwise convergence of functions, fn → f , and about uniform
convergence too, fn →u f , exactly as for the real functions.
Observe that, since xn → x in the complex sense means that (an , bn ) → (a, b) in the
usual, real plane sense, a function f : C → C is continuous precisely when it is continuous
when regarded as real function, f : R2 → R2 . But more on this later in this book. At the
level of examples now, we first have the polynomials, P ∈ C[X]. We already met such
polynomials in chapter 5, so let us recall from there that we have:
Theorem 6.4. Each polynomial P ∈ C[X] can be regarded as a continuous function
P : C → C. Moreover, we have the formula
P (x) = a(x − r1 ) . . . (x − rn )
with a ∈ C, and with the numbers r1 , . . . , rn ∈ C being the roots of P .
Proof. This is something that we know from chapter 5, the idea being that one root
can be always constructed, by reasoning by contradiction, and doing some analysis around
the minimum of |P |, and then a recurrence on the degree n ∈ N does the rest. □
Next in line, we have the rational functions, which are defined as follows:
Theorem 6.5. The quotients of complex polynomials f = P/Q are called rational
funtions. When written in reduced form, with P, Q prime to each other,
P
f=
Q
is well-defined and continuous outside the zeroes Pf ⊂ C of Q, called poles of f :
f : C − Pf → C
In addition, the rational functions, regarded as algebraic expressions, are stable under
summing, making products and taking inverses.
 6A. FUNCTIONS, CONTINUITY 131

Proof. There are several things going on here, the idea being as follows:
(1) First of all, we can surely talk about quotients of polynomials, f = P/Q, regarded
as abstract algebraic expressions. Also, the last assertion is clear, because we can indeed
perform sums, products, and take inverses, by using the following formulae:
 −1
P R P S + QR P R PR P Q
+ = , · = , =
Q S QS Q S QS Q P
(2) The question is now, given a rational function f , can we regard it as a complex
function? In general, we cannot say that we have f : C → C, for instance because
f (x) = x−1 is not defined at x = 0. More generally, assuming f = P/Q with P, Q ∈ C,
we cannot talk about f (x) when x is a root of Q, unless of course we are in the special
situation where x is a root of P too, and we can simplify the fraction.
(3) In view of this discussion, in order to solve our question, we must avoid the situation
where the polynomials P, Q have common roots. But this can be done by writing our
rational function f in reduced form, as follows, with P, Q ∈ C[X] prime to each other:
P
f=
Q
(4) Now with this convention made, it is clear that f is well-defined, and continuous
too, outside of the zeroes of f . Now since these zeroes can be obviously recovered from
the knowledge of f itself, as being the points where “f explodes”, we can call them poles
of f , and so we have a function f : C − Pf → C, as in the statement. □
As a comment here, the term “pole” does not come from the Poles who invented this,
but rather from the fact that, when trying to draw the graph of f , or rather imagine that
graph, which takes place in 2 + 2 = 4 real dimensions, we are faced with some sort of tent,
which is suspended by infinite poles, which lie, guess where, at the poles of f .

Getting back now to Theorem 6.5, as stated, that is obviously a mixture of algebra
and analysis. So, let us first further clarify the algebra part. We know that the rational
functions are stable under summing, making products and taking inverses, and this makes
the link with the following notion, from number theory and abstract algebra:
Definition 6.6. A field is a set F with a sum operation + and a product operation
×, subject to the following conditions:
(1) a + b = b + a, a + (b + c) = (a + b) + c, there exists 0 ∈ F such that a + 0 = 0,
and any a ∈ F has an inverse −a ∈ F , satisfying a + (−a) = 0.
(2) ab = ba, a(bc) = (ab)c, there exists 1 ∈ F such that a1 = a, and any a ̸= 0 has a
multiplicative inverse a−1 ∈ F , satisfying aa−1 = 1.
(3) The sum and product are compatible via a(b + c) = ab + ac.
132 6. COMPLEX FUNCTIONS

As basic examples of fields, we have the rational numbers Q, the real numbers R, and
the complex numbers C. Some further examples of fields of numbers, which are more
specialized, and useful in number theory, can be constructed as well. In view of this, it is
useful to think of any field F as being a “field of numbers”, and this because the elements
a, b, c, . . . ∈ F behave under the operations + and × exactly as the usual numbers do.

In what regards the various spaces of functions, such as the polynomials C[X], or the
continuous functions C(R), these certainly have sum and product operations + and ×,
but are in general not fields, because they do not satisfy the following field axiom:
f ̸= 0 =⇒ ∃f −1
However, and here comes our point, Theorem 6.5 tells us that the rational functions
form a field. This is quite interesting, and opposite to the general spirit of analysis and
function spaces, which are in general not fields. Let us record this finding, as follows:
Definition 6.7. We denote by C(X) the field of rational functions
P
f= , P, Q ∈ C[X]
Q
with the usual sum and product operations + and × for the rational functions.
To be more precise, this is some sort of reformulation of Theorem 6.5, or rather of
the algebraic content of Theorem 6.5, telling us that the rational functions form indeed a
field. And to the question, how can a theorem suddenly become a definition, the answer
is that this is quite commonplace in mathematics, and especially in algebra.

Back now to analysis, let us point out that, contrary to what the above might suggest,
everything does not always extend trivally from the real to the complex case. For instance,
we have the following result, that we already talked about a bit in chapter 5:
Proposition 6.8. We have the following formula, valid for any |x| < 1,
1
= 1 + x + x2 + . . .
1−x
but, for x ∈ C − R, the geometric meaning of this formula is quite unclear.
Proof. Here the formula in the statement holds indeed, by multiplying and cancelling
terms, exactly as in the real case, with the convergence being justified by:
∞ ∞
X
n
X 1
x ≤ |x|n =
n=0 n=0
1 − |x|
As for the last assertion, this is something rather informal, which hides however many
interesting things, that we discussed in some detail in chapter 5. □
 6A. FUNCTIONS, CONTINUITY 133

Getting now to more complicated functions, such as sin, cos, exp, log, again many
things extend well from real to complex, the basic theory here being as follows:
Theorem 6.9. The functions sin, cos, exp, log have complex extensions, given by
∞ ∞
X x2l+1 X x2l
sin x = (−1)l , cos x = (−1)l
l=0
(2l + 1)! l=0
(2l)!
∞ ∞
x
X xk X
k+1 x
k
e = , log(1 + x) = (−1)
k=0
k! k=1
k
with |x| < 1 needed for log, which are continuous over their domain, and satisfy the
formulae ex+y = ex ey and eix = cos x + i sin x.
Proof. This is a mixture of trivial and non-trivial results, as follows:
(1) We already know about ex from chapter 5, the idea being that the convergence of
the series, and then the continuity of ex , come from the following estimate:
∞
X |x|k
x
|e | ≤ = e|x| < ∞
k=0
k!
(2) Regarding sin x, the same method works, with the following estimate:
∞ ∞
X |x|2l+1 X |x|k
| sin x| ≤ ≤ = e|x|
l=0
(2l + 1)! k=0
k!
(3) The same goes for cos x, the estimate here being as follows:
∞ ∞
X |x|2l X |x|k
| cos x| ≤ ≤ = e|x|
l=0
(2l)! k=0
k!
(4) Regarding now the formulae satisfied by sin, cos, exp, we already know from chapter
5 that the exponential has the following property, exactly as in the real case:
ex+y = ex ey
We also have the following formula, connecting sin, cos, exp, again as before:
∞
ix
X (ix)k
e =
k=0
k!
X (ix)k X (ix)k
= +
k=2l
k! k=2l+1
k!
∞ ∞
X x2l X x2l+1
= (−1)l +i (−1)l
l=0
(2l)! l=0
(2l + 1)!
= cos x + i sin x
134 6. COMPLEX FUNCTIONS

(5) In order to discuss now the complex logarithm function log, let us first study some
more the complex exponential function exp. By using ex+y = ex ey we obtain ex ̸= 0 for
any x ∈ C, so the complex exponential function is as follows:
exp : C → C − {0}
Now since we have ex+iy = ex eiy for x, y ∈ R, with ex being surjective onto (0, ∞),
and with eiy being surjective onto the unit circle T, we deduce that exp : C → C − {0} is
surjective. Also, again by using ex+iy = ex eiy , we deduce that we have:
ex = ey ⇐⇒ x − y ∈ 2πiZ
(6) With these ingredients in hand, we can now talk about log. Indeed, let us fix a
horizontal strip in the complex plane, having width 2π:
n o
S = x + iy x ∈ R, y ∈ [a, a + 2π)

We know from the above that the restriction map exp : S → C − {0} is bijective, so
we can define log as to be the inverse of this map:
log = exp−1 : C − {0} → S
(7) In practice now, the best is to choose for instance a = 0, or a = −π, as to have
the whole real line included in our strip, R ⊂ S. In this case on R+ we recover the usual
logarithm, while on R− we obtain complex values, as for instance log(−1) = πi in the
case a = 0, or log(−1) = −πi in the case a = −π, coming from eπi = −1.
(8) Finally, assuming |x| < 1, we can consider the following series, which converges:
∞
X xk
f (x) = (−1)k+1
k=1
k

We have then ef (x) = 1 + x, and so f (x) = log(1 + x), when 1 + x ∈ S. □

As an interesting consequence of the above result, which is of great practical interest,
we have the following useful method, for remembering the basic math formulae:
Method 6.10. Knowing ex = k xk /k! and eix = cos x + i sin x gives you
P

sin(x + y) = sin x cos y + cos x sin y

cos(x + y) = cos x cos y − sin x sin y
right away, in case you forgot these formulae, as well as
∞ ∞
X x2l+1 X x2l
sin x = (−1)l , cos x = (−1)l
l=0
(2l + 1)! l=0
(2l)!
again, right away, in case you forgot these formulae.
 6A. FUNCTIONS, CONTINUITY 135

To be more precise, assume that we forgot everything trigonometry, which is something

that can happen to everyone, in the real life, but still know the formulae ex = k xk /k!
P
and eix = cos x + i sin x. Then, we can recover the formulae for sums, as follows:
ei(x+y) = eix eiy =⇒ cos(x + y) + i sin(x + y) = (cos x + i sin x)(cos y + i sin y)
(
cos(x + y) = cos x cos y − sin x sin y
=⇒
sin(x + y) = sin x cos y + cos x sin y
And isn’t this smart. Also, and even more impressively, we can recover the Taylor
formulae for sin, cos, which are certainly difficult to memorize, as follows:
X (ix)k X (ix)k
eix = =⇒ cos x + i sin x =
k
k! k
k!
( P∞ x2l
cos x = l=0 (−1)l (2l)!
=⇒
sin x = ∞ l x2l+1
P
l=0 (−1) (2l+1)!

Finally, in what regards log, there is a trick here too, which is partial, namely:
x2
 
log(exp x) = x =⇒ log 1 + x + + ... = x
2
y2
=⇒ log(1 + y) = y − + ...
2
To be more precise, log(1 + y) ≃ y is clear, and with a bit more work, that we will
leave here as an instructive exercise, you can recover log(1 + y) = y − y 2 /2 too. Of course,
the higher terms can be recovered too, with enough work involved, at each step.

Moving ahead, Theorem 6.9 leads us into the question on whether the other formulae
that we know about sin, cos, such as the values of these functions on sums x + y, or on
doubles 2x, extend to the complex setting. Things are quite tricky here, and in relation
with this, we have the following result, which is something of general interest:
Proposition 6.11. The following functions, called hyperbolic sine and cosine,
ex − e−x ex + e−x
sinh x = , cosh x =
2 2
are subject to the following formulae:
(1) ex = cosh x + sinh x.
(2) sinh(ix) = i sin x, cosh(ix) = cos x, for x ∈ R.
(3) sinh(x + y) = sinh x cosh y + cosh x sinh y.
(4) cosh(x + y) = cosh x cosh y + sinh x sinh y.
P x2l+1 P x2l
(5) sinh x = l (2l+1)! , cosh x = l (2l)! .
136 6. COMPLEX FUNCTIONS

Proof. The formula (1) follows from definitions. As for (2), this follows from:
eix − e−ix cos x + i sin x cos x − i sin x
sinh(ix) = = − = i sin x
2 2 2
eix + e−ix cos x + i sin x cos x − i sin x
cosh(ix) = = + = cos x
2 2 2
Regarding now (3,4), observe first that the formula ex+y = ex + ey reads:
cosh(x + y) + sinh(x + y) = (cosh x + sinh x)(cosh y + sinh y)
Thus, we have some good explanation for (3,4), and in practice, these formulae can
be checked by direct computation, as follows:
ex+y − e−x−y ex − e−x ey + e−y ex + e−x ey − e−y
= · + ·
2 2 2 2 2
−x−y −x −y −x
ex+y
+e x
e +e y
e +e x
e −e e − e−y
y
= · + ·
2 2 2 2 2
P k
Finally, (5) is clear from the definition of sinh, cosh, and from e = k xk! .
x
□
Finally, we can talk as well about powers, in the following way:
Fact 6.12. Under suitable assumptions, we can talk about xy with x, y ∈ C, and in
particular about the complex functions ax and xa , with a ∈ C.
To be more precise, in what regards xy , we already know from chapters 1-2 that things
are quite tricky, even in the real case. In the complex case the same problems appear,
along with some more, but these questions can be solved by using the above theory of
exp, log. To be more precise, in order to solve the first question, we can set:
xy = ey log x
We will be back to these functions later, when we will have more tools for studying
them. In fact, all of a sudden, we are now into quite complicated mathematics, and we
cannot really deal with the problems left open above, with bare hands. More later.
At the level of the general theory now, the main tool for dealing with the continuous
functions f : R → R was the intermediate value theorem. In the complex setting, that of
the functions f : C → C, we do not have such a theorem, at least in its basic formulation,
because there is no order relation for the complex numbers, or things like complex in-
tervals. However, the intermediate value theorem in its advanced formulation, that with
connected sets, extends of course, and we have the following result:
Theorem 6.13. Assuming that f : X → C with X ⊂ C is continuous, if the domain
X is connected, then so is its image f (X).
Proof. This follows exactly as in the real case, with just a bit of discussion being
needed, in relation with open and closed sets, and then connected sets, inside C. □
 6B. HOLOMORPHIC FUNCTIONS 137

6b. Holomorphic functions

Let us study now the differentiability of the complex functions f : C → C. Things
here are quite tricky, but let us start with a straightforward definition, as follows:
Definition 6.14. We say that a function f : X → C is differentiable in the complex
sense when the following limit is defined for any x ∈ X:
f (x + t) − f (x)
f ′ (x) = lim
t→0 t
In this case, we also say that f is holomorphic, and we write f ∈ H(X).
As basic examples, we have the power functions f (x) = xn . Indeed, the derivative of
such a power function can be computed exactly as in the real case, and we get:
n ′ (x + t)n − xn
(x ) = lim
t→0 t
nxn−1 t + n2 xn−2 t2 + . . . + tn


= lim
t→0 t
nxn−1 t
= lim
t→0 t
n−1
= nx
We will see later more computations of this type, similar to those from the real case.
To summarize, our definition of differentiability seems to work nicely, so let us start
developing some theory. The general results from the real case extend well, as follows:
Proposition 6.15. We have the following results:
(1) (f + g)′ = f ′ + g ′ .
(2) (λf )′ = λf ′ .
(3) (f g)′ = f ′ g + f g ′ .
(4) (f ◦ g)′ = f ′ (g)g ′ .
Proof. These formulae are all clear from definitions, following exactly as in the real
case. Thus, we are led to the conclusions in the statement. □
As an obvious consequence of (1,2) above, any poynomial P ∈ C[X] is differentiable,
with its derivative being given by the same formula as in the real case, namely:
Xn n
X
P (x) = ck xk =⇒ P ′ (x) = kck xk−1
k=0 k=1

More generally, any rational function f ∈ C(X) is differentiable on its domain, that
is, outsides its poles, because if we write f = P/Q with P, Q ∈ C[X], we have:
 ′
′ P P ′ Q − P Q′
f = =
Q Q2
138 6. COMPLEX FUNCTIONS

Let us record these conclusions in a statement, as follows:

Proposition 6.16. The following happen:
(1) Any polynomial P ∈ C[X] is holomorphic, and in fact infinitely differentiable in
the complex sense, with all its derivatives being polynomials.
(2) Any rational function f ∈ C(X) is holomorphic, and in fact infinitely differen-
tiable, with all its derivatives being rational functions.
Proof. This follows indeed from the above discussion. □

Let us look now into more complicated complex functions that we know. And here,
surprise, things are quite tricky, the result being as follows:
Theorem 6.17. The following happen:
(1) sin, cos, exp, log are holomorphic, and in fact are infinitely differentiable, with
their derivatives being given by the same formulae as in the real case.
(2) However, functions like x̄ or |x| are not holomorphic, and this because the limit
defining f ′ (x) depends on the way we choose t → 0.
Proof. There are several things going on here, the idea being as follows:
(1) Here the first assertion is standard, because our functions sin, cos, exp, log have
Taylor series that we know, and the derivative can be therefore computed by using the
same rule as in the real case, similar to the one for polynomials, namely:
∞
X ∞
X
f (x) = ck xk =⇒ f ′ (x) = kck xk−1
k=0 k=1

(2) Regarding now the function f (x) = x̄, the point here is that we have:
f (x + t) − f (x) x̄ + t̄ − x̄ t̄
= =
t t t
But this limit does not converge with t → 0, for instance because with t ∈ R we obtain
1 as limit, while with t ∈ iR we obtain −1 as limit. In fact, with t = rw with |w| = 1
fixed and r ∈ R, r → 0, we can obtain as limit any number on the unit circle:
f (x + rw) − f (x) rw̄
lim = lim = w̄2
r→0 rw r→0 rw

(3) The situation for the function f (x) = |x| is similar. To be more precise, we have:
f (x + rw) − f (x) |x + rw| − |x|
= · w̄
rw r
 6B. HOLOMORPHIC FUNCTIONS 139

Thus with |w| = 1 fixed and r → 0 we obtain a certain multiple of w̄, with the
multiplication factor being computed as follows:
|x + rw| − |x| |x + rw|2 − |x|2
=
r (|x + rw| + |x|)r
xrw̄ + x̄rw
≃
2|x|r
 
xw̄
= Re
|x|
Now by making w vary on the unit circle, as in (2) above, we can obtain in this way
limits pointing in all possible directions, so our limit does not converge, as stated. □
The above result is quite surprising, because we are so used, from the real case, to the
notion of differentiability to correspond to some form of “smoothness” of the function,
and to be more precisely, “smoothness at first order”. Or, if you prefer, to correspond to
the “non-bumpiness” of the function. So, we are led to the following dilemma:
Dilemma 6.18. It’s either that x̄ and |x| are smooth, as the intuition suggests, and
we are wrong with our definition of differentiability. Or that x̄ and |x| are bumpy, while
this being not very intuitive, and we are right with our definition of differentiability.
And we won’t get discouraged by this. After all, this is just some empty talking, and if
there is something to rely upon, mathematics and computations, these are the computa-
tions from the proof of Theorem 6.17. So, moving ahead now, based on that computations,
let us formulate the following definition, coming as a complement to Definition 6.14:
Definition 6.19. A function f : X → C is called differentiable:
(1) In the real sense, if the following two limits converge, for any x ∈ X:
f (x + t) − f (x) f (x + t) − f (x)
f1′ (x) = lim , fi′ (x) = lim
t∈R→0 t t∈iR→0 t
(2) In a radial sense, if the following limit converges, for any x ∈ X, and w ∈ T:
f (x + t) − f (x)
fw′ (x) = lim
t∈wR→0 t
(3) In the complex sense, if the following limit converges, for any x ∈ X:
f (x + t) − f (x)
f ′ (x) = lim
t→0 t
If f is differentiable in the complex sense, we also say that f is holomorphic.
We can see now more clearly what is going on. We have (3) =⇒ (2) =⇒ (1) in
general, and most of the functions that we know, namely the polynomials, the rational
functions, and sin, cos, exp, log, satisfy (3). As for the functions x̄, |x|, these do not satisfy
140 6. COMPLEX FUNCTIONS

(3), and do not satisfy (2) either, but they satisfy however (1). It is possible to say more
about all this, and we will certainly come back to this topic, later in this book.
Back to business now, all the examples of holomorphic functions that we have are
infinitely differentiable, and this raises the question of finding a function such that f ′
exists, while f ′′ does not exist. Quite surprisingly, we will see that such functions do not
exist. In order to get into this latter phenomenon, let us start with:
Theorem 6.20. Each power series f (x) = n cn xn has a radius of convergence
P

R ∈ [0, ∞]
which is such that f converges for |x| < R, and diverges for |x| > R. We have:
1 p
R= , C = lim sup n |cn |
C n→∞

Also, in the case |x| = R the function f can either converge, or diverge.
Proof. This
P follows from the Cauchy criterion for series, from chapter 1, which says
that a series n xn converges if c < 1, and diverges if c > 1, where:
p
c = lim sup n |xn |
n→∞
n
Indeed, with xn = |cn x | we obtain that the convergence radius R ∈ [0, ∞] exists, and
is given by the formula in the statement. Finally, for the examples and counterexamples
at the end, when |x| = R, the simplest here is to use f (x) = n xn , where R = 1.
P
□
Back now to our questions regarding derivatives, we have:
Theorem 6.21. Assuming that a function f : X → C is analytic, in the sense that it
is a series, around each point x ∈ X,
∞
X
f (x + t) = cn t n
n=0
it follows that f is infinitely differentiable, in the complex sense. In particular, f ′ exists,
and so f is holomorphic in our sense.
Proof. Assuming that f is analytic, as in the statement, we have:
X∞
′
f (x + t) = ncn tn−1
n=1
Moreover, the radius of convergence is the same, as shown by√the following computa-
tion, using the Cauchy formula for the convergence radius, and n n → 1:
1 p
n
p
n 1
′
= lim sup |ncn | = lim sup |cn | =
R n→∞ n→∞ R
′
Thus f exists and is analytic, on the same domain, and this gives the result. □
 6C. CAUCHY FORMULA 141

6c. Cauchy formula

Our goal in what follows will be that of proving that any holomorphic function is
analytic. This is something quite subtle, which cannot be proved with bare hands, and
requires lots of preliminaries. Getting to these preliminaries now, our claim is that a lot
of useful knowledge, in order to deal with the holomorphic functions, can be gained by
further studying the analytic functions, and even the usual polynomials P ∈ C[X].

So, let us further study the polynomials P ∈ C[X], and other analytic functions. We
already know from chapter 5 that in the polynomial case, P ∈ C[X], some interesting
things happen, because any such polynomial has a root, and even deg(P ) roots, after a
recurrence. Keeping looking at polynomials, with the same methods, we are led to:
Theorem 6.22. Any polynomial P ∈ C[X] satisfies the maximum principle, in the
sense that given a domain D, with boundary γ, we have:
∃x ∈ γ , |P (x)| = max |P (y)|
y∈D

That is, the maximum of |P | over a domain is attained on its boundary.

Proof. In order to prove this, we can split D into connected components, and then
assume that D is connected. Moreover, we can assume that D has no holes, and so is
homeomorphic to a disk, and even assume that D itself is a disk. But with this assumption
made, the result follows from by contradiction, by using the same arguments as in the
proof of the existence of a root, from chapter 5. To be more precise, assume deg P ≥ 1,
and that the maximum of |P | is attained at the center of a disk D = D(z, r):
|P (z)| = max |P (x)|
x∈D

We can write then P (z + t) ≃ P (z) + ctk with c ̸= 0, for t small, and by suitably
choosing the argument of t on the unit circle we conclude, exactly as in chapter 5, that
the function |P | cannot have a local maximum at z, as stated. □
A good explanation for the fact that the maximum principle holds for polynomials
P ∈ C[X] could be that the values of such a polynomial inside a disk can be recovered
from its values on the boundary. And fortunately, this is indeed the case, and we have:
Theorem 6.23. Given a polynomial P ∈ C[X], and a disk D, with boundary γ, we
have the following formulae, with the integrations being the normalized, mass 1 ones:
R
(1) P satisfies the plain mean value formula P (x) = D PR(y)dy.
(2) P satisfies the boundary mean value formula P (x) = γ P (y)dy.
Proof. As a first observation, the two mean value formulae in the statement are
equivalent, by restricting the attention to disks D, having as boundaries circles γ, and
using annuli and polar coordinates for the proof of the equivalence. As for the formulae
142 6. COMPLEX FUNCTIONS

themselves, these can be checked by direct computation for a disk D, with the formulation
in (2) being the most convenient. Indeed, for a monomial P (x) = xn we have:
Z Z 2π
n 1
y dy = (x + reit )n dt
γ 2π 0
Z 2π X n  
1 n k it n−k
= x (re ) dt
2π 0 k=0 k
n   Z 2π
X n k n−k 1
= x r ei(n−k)t dt
k=0
k 2π 0
n
X n  
= xk rn−k δkn
k=0
k
= xn
Here we have used the following key identity, valid for any exponent m ∈ Z:
Z 2π Z 2π
1 imt 1
e dt = cos(mt) + i sin(mt)dt
2π 0 2π 0
= δm0 + i · 0
= δm0
Thus, we have the result for monomials, and the general case follows by linearity. □

All the above is very nice, but we can in fact do even better, with a more powerful
integration formula. Let us start with some preliminaries. We first have:
Proposition 6.24. We can integrate functions f over curves γ by setting
Z Z b
f (x)dx = f (γ(t))γ ′ (t)dt
γ a

with this quantity being independent on the parametrization γ : [a, b] → C.

Proof. We must prove that the quantity in the statement is independent on the
parametrization. In other words, we must prove that if we pick two different parametriza-
tions γ, η : [a, b] → C of our curve, then we have the following formula:
Z b Z b
′
f (γ(t))γ (t)dt = f (η(t))η ′ (t)dt
a a

But for this purpose, let us write γ = ηϕ, with ϕ : [a, b] → [a, b] being a certain
function, that we can assume to be bijective, via an elementary cut-and-paste argument.
 6C. CAUCHY FORMULA 143

By using the chain rule for derivatives, and the change of variable formula, we have:
Z b Z b
′
f (γ(t))γ (t)dt = f (ηϕ(t))(ηϕ)′ (t)dt
a a
Z b
= f (ηϕ(t))η ′ (ϕ(t))ϕ′ (t)dt
a
Z b
= f (η(t))η ′ (t)dt
a

Thus, we are led to the conclusions in the statement. □

The main properties of the above integration method are as follows:
Proposition 6.25. We have the following formula, for a union of paths:
Z Z Z
f (x)dx = f (x)dx + f (x)dx
γ∪η γ η

Also, when reversing the path, the integral changes its sign.
Proof. Here the first assertion is clear from definitions, and the second assertion
comes from the change of variable formula, by using Proposition 6.24. □
Now by getting back to polynomials, we have the following result:
Theorem 6.26. Any polynomial P ∈ C[X] satisfies the Cauchy formula
Z
1 P (y)
P (x) = dy
2πi γ y − x
with the integration over γ being constructed as above.
Proof. This follows by using abstract arguments and computations similar to those
in the proof of Theorem 6.23. Indeed, by linearity we can assume P (x) = xn . Also, by
using a cut-and-paste argument, we can assume that we are on a circle:
γ : [0, 2π] → C , γ(t) = x + reit
By using now the computation from the proof of Theorem 6.23, we obtain:
Z 2π
yn (x + reit )n
Z
dy = it
rieit dt
γ y − x 0 re
Z 2π
= i (x + reit )n dt
0
= i · 2πxn
Thus, we are led to the formula in the statement. □
144 6. COMPLEX FUNCTIONS

All this is quite interesting, and obviously, we are now into some serious mathematics.
Importantly, Theorem 6.22, Theorem 6.23 and Theorem 6.26 provide us with a path for
proving the converse of Theorem 6.21. Indeed, if we manage to prove the Cauchy formula
for any holomorphic function f : X → C, then it will follow that our function is in fact
analytic, and so infinitely differentiable. So, let us start with the following result:
Theorem 6.27. The Cauchy formula, namely
Z
1 f (y)
f (x) = dy
2πi γ y − x
holds for any holomorphic function f : X → C.
Proof. This is something standard, which can be proved as follows:
(1) Our first claim is that given f ∈ H(X), with f ′ ∈ C(X), the integral of f ′ vanishes
on any path. Indeed, by using the change of variable formula, we have:
Z Z b
′
f (x)dx = f ′ (γ(t))γ ′ (t)dt
γ a
= f (γ(b)) − f (γ(a))
= 0
(2) Our second claim is that given f ∈ H(X) and a triangle ∆ ⊂ X, we have:
Z
f (x)dx = 0
∆

Indeed, let us call ∆ = ABC our triangle. Now consider the midpoints A′ , B ′ , C ′ of
the edges BC, CA, AB, and then consider the following smaller triangles:
∆1 = AC ′ B ′ , ∆2 = BA′ C ′ , ∆3 = CB ′ A′ , ∆4 = A′ B ′ C ′
These smaller triangles partition then ∆, and due to our above conventions for the
vertex ordering, which produce cancellations when integrating over them, we have:
Z 4 Z
X
f (x)dx = f (x)dx
∆ i=1 ∆i

Thus we can pick, among the triangles ∆i , a triangle ∆(1) such that:
Z Z
f (x)dx ≤ 4 f (x)dx
∆ ∆(1)

In fact, by repeating the procedure, we obtain triangles ∆(n) such that:

Z Z
n
f (x)dx ≤ 4 f (x)dx
∆ ∆(n)
 6C. CAUCHY FORMULA 145

(3) Now let z be the limiting point of these triangles ∆(n) , and fix ε > 0. By using
the fact that the functions 1, x integrate over paths up to 0, coming from (1), we obtain
the following estimate, with n ∈ N being big enough, and L being the perimeter of ∆:
Z Z
f (x)dx = f (x) − f (z) − f ′ (z)(x − z)dx
∆(n) (n)
Z∆
≤ |f (x) − f (z) − f ′ (z)(x − z)| dx
(n)
Z∆
≤ ε|x − z|dx
∆(n)
−n
≤ ε(2 L)2
Now by combining this with the estimate in (2), this proves our claim.
(4) The rest is quite routine. First, we can pass from triangles to boundaries of convex
sets, in a straightforward way, with the same conclusion as in (2), namely:
Z
f (x)dx = 0
γ
Getting back to what we want to prove, namely the Cauchy formula for an arbitrary
holomorphic function f ∈ H(X), let x ∈ X, and consider the following function:
(
f (y)−f (x)
y−x
(y ̸= x)
g(y) =
f ′ (x) (y = x)
Now assuming that γ encloses a convex set, we can apply what we found, namely
vanishing of the integral, to this function g, and we obtain the Cauchy formula for f .
(5) Finally, the extension to general curves is standard, and standard as well is the
discussion of what exactly happens at x, in the above proof. See Rudin [74]. □
As a main application of the Cauchy formula, we have:
Theorem 6.28. The following conditions are equivalent, for a function f : X → C:
(1) f is holomorphic.
(2) f is infinitely differentiable.
(3) f is analytic.
(4) The Cauchy formula holds for f .
Proof. This is routine from what we have, the idea being as follows:
(1) =⇒ (4) is non-trivial, but we know this from Theorem 6.27.
(4) =⇒ (3) is something trivial, because we can expand the series in the Cauchy
formula, and we conclude that our function is indeed analytic.
(3) =⇒ (2) =⇒ (1) are both elementary, known from Theorem 6.21. □
146 6. COMPLEX FUNCTIONS

As another application of the Cauchy formula, we have:

Theorem 6.29. Any holomorphic function f : X → C satisfies the maximum princi-
ple, in the sense that given a domain D, with boundary γ, we have:
∃x ∈ γ , |f (x)| = max |f (y)|
y∈D

That is, the maximum of |f | over a domain is attained on its boundary.

Proof. This follows indeed from the Cauchy formula. Observe that the converse is
not true, for instance because functions like x̄ satisfy too the maximum principle. We will
be back to this later, when talking about harmonic functions. □
As before with polynomials, a good explanation for the fact that the maximum prin-
ciple holds could be that the values of our function inside a disk can be recovered from
its values on the boundary. And fortunately, this is indeed the case, and we have:
Theorem 6.30. Given an holomorphic function f : X → C, and a disk D, with
boundary γ, the following happen:
R
(1) f satisfies the plain mean value formula f (x) = D fR(y)dy.
(2) f satisfies the boundary mean value formula f (x) = γ f (y)dy.
Proof. As usual, this follows from the Cauchy formula, with of course some care in
passing from integrals constructed as in Proposition 6.24 to integrals viewed as averages,
which are those that we refer to, in the present statement. □
Finally, as yet another application of the Cauchy formula, which is something nice-
looking and conceptual, we have the following statement, called Liouville theorem:
Theorem 6.31. An entire, bounded holomorphic function
f :C→C , |f | ≤ M
must be constant. In particular, knowing f → 0 with z → ∞ gives f = 0.
Proof. This follows as usual from the Cauchy formula, namely:
Z
1 f (y)
f (x) = dy
2πi γ y − x
Alternatively, we can view this as a consequence of Theorem 6.30, because given two
points x ̸= y, we can view the values of f at these points as averages over big disks
centered at these points, say D = Dx (R) and E = Dy (R), with R >> 0:
Z Z
f (x) = f (z)dz , f (y) = f (z)dz
D E
Indeed, the point is that when the radius goes to ∞, these averages tend to be equal,
and so we have f (x) ≃ f (y), which gives f (x) = f (y) in the limit. □
 6D. STIELTJES INVERSION 147

6d. Stieltjes inversion

We would like to end this chapter with an interesting application of the complex
functions to probability theory. We have learned some basic probability in chapter 4,
and in view of the material there, an interesting question is how to recover a probability
measure out of its moments. And the answer here, which is non-trivial, is as follows:

Theorem 6.32. The density of a real probability measure µ can be recaptured from
the sequence of moments {Mk }k≥0 via the Stieltjes inversion formula

1
dµ(x) = lim − Im (G(x + it)) · dx
t↘0 π

where the function on the right, given in terms of moments by

G(ξ) = ξ −1 + M1 ξ −2 + M2 ξ −3 + . . .

is the Cauchy transform of the measure µ.

Proof. The Cauchy transform of our measure µ is given by:

∞
X
−1
G(ξ) = ξ Mk ξ −k
k=0
Z −1
ξ
= −1
dµ(y)
R 1−ξ y
Z
1
= dµ(y)
R ξ −y

Now with ξ = x + it, we obtain the following formula:

Z  
1
Im(G(x + it)) = Im dµ(y)
R x − y + it
Z  
1 1 1
= − dµ(y)
R 2i x − y + it x − y − it
Z
t
= − 2 2
dµ(y)
R (x − y) + t
148 6. COMPLEX FUNCTIONS

By integrating over [a, b] we obtain, with the change of variables x = y + tz:

Z b Z Z b
t
Im(G(x + it))dx = − 2 2
dx dµ(y)
a R a (x − y) + t
Z Z (b−y)/t
t
= − 2 2
t dz dµ(y)
R (a−y)/t (tz) + t
Z Z (b−y)/t
1
= − 2
dz dµ(y)
R (a−y)/t 1 + z
Z  
b−y a−y
= − arctan − arctan dµ(y)
R t t
Now observe that with t ↘ 0 we have:
π π

 2
−2 =0 (y < a)
π π
2 − 0 = 2

    (y = a)
b−y a−y π π
lim arctan − arctan = 2 − (− 2 ) = π (a < y < b)
t↘0 t t  π π



 0 − (− 2 ) = 2 (y = b)
 π π
− 2 − (− 2 ) = 0 (y > b)
We therefore obtain the following formula:
Z b  
µ(a) + µ(b)
lim Im(G(x + it))dx = −π µ(a, b) +
t↘0 a 2
Thus, we are led to the conclusion in the statement. □
Before getting further, let us mention that the above result does not fully solve the
moment problem, because we still have the question of understanding when a sequence
of numbers M1 , M2 , M3 , . . . can be the moments of a measure µ. We have here:
Theorem 6.33. A sequence of numbers M0 , M1 , M2 , M3 , . . . ∈ R, with M0 = 1, is the
series of moments of a real probability measure µ precisely when:
M0 M1 M2
M0 M1
M0 ≥ 0 , ≥0 , M1 M2 M3 ≥ 0 , ...
M1 M2
M2 M3 M4
That is, the associated Hankel determinants must be all positive.
Proof. This is something a bit more advanced, the idea being as follows:
(1) As a first observation, the positivity conditions in the statement tell us that the
following associated linear forms must be positive:
Xn
ci c̄j Mi+j ≥ 0
i,j=1
 6D. STIELTJES INVERSION 149

(2) But this is something very classical, in one sense the result being elementary,
coming from the following computation, which shows that we have positivity indeed:
Z n 2 n
Z X n
X X
i
ci x dµ(x) = ci c̄j xi+j dµ(x) = ci c̄j Mi+j
R i=1 R i,j=1 i,j=1

(3) As for the other sense, here the result comes once again from the above formula,
this time via some standard functional analysis. □

Getting back now to more concrete things, the point is that we have:
Fact 6.34. Given a graph X, with distinguished vertex ∗, we can talk about the prob-
ability measure µ having as k-th moment the number of length k loops based at ∗:
n o
Mk = ∗ −i1 − i2 − . . . − ik = ∗

As basic examples, for the graph N the moments must be the Catalan numbers Ck , and
for the graph Z, the moments must be the central binomial coefficients Dk .
To be more precise, the first assertion, regarding the existence and uniqueness of µ,
follows from a basic linear algebra computation, by diagonalizing the adjacency matrix
of X. As for the examples, for the graph N we more or less already know this, from
our various Catalan number considerations from chapter 2, and for the graph Z this is
something elementary, that we will leave here as an instructive exercise.

Needless to say, counting loops on graphs, as in Fact 6.34, is something important in

applied mathematics, and physics. So, back to our business now, motivated by all this,
as a basic application of the Stieltjes formula, let us solve the moment problem for the
Catalan numbers Ck , and for the central binomial coefficients Dk . We first have:
Theorem
1 2k

6.35. The real measure having as even moments the Catalan numbers,
Ck = k+1 k , and having all odd moments 0 is the measure
1√
γ1 = 4 − x2 dx
2π
called Wigner semicircle law on [−2, 2].
Proof. In order to apply the inversion formula, our starting point will be the formula
from chapter 3 for the generating series of the Catalan numbers, namely:
∞ √
X
k 1 − 1 − 4z
Ck z =
k=0
2z
150 6. COMPLEX FUNCTIONS

By using this formula with z = ξ −2 , we obtain the following formula:

∞
X
G(ξ) = ξ −1 Ck ξ −2k
k=0
p
−1 1 − 4ξ −2
1−
= ξ ·
2ξ −2
ξ p 
= 1 − 1 − 4ξ −2
2
ξ 1p 2
= − ξ −4
2 2

Now let us apply Theorem 6.32. The study here goes as follows:

(1) According to the general philosophy of the Stieltjes formula, the first term, namely
ξ/2, which is “trivial”, will not contribute to the density.

(2) As for the second term, which is somethingpnon-trivial, this will contribute to the
rule here being that the square root ξ 2 − 4 will be replaced by the “dual”
density, the √
square root 4 − x2 dx, and that we have to multiply everything by −1/π.

(3) As a conclusion, by Stieltjes inversion we obtain the following density:

1 1√ 1√
dµ(x) = − ·− 4 − x2 dx = 4 − x2 dx
π 2 2π

Thus, we have obtained the mesure in the statement, and we are done. □

We have the following version of the above result:

Theorem 6.36.
The real measure having as sequence of moments the Catalan num-
1 2k
bers, Ck = k+1 k , is the measure

1 √ −1
π1 = 4x − 1 dx
2π

called Marchenko-Pastur law on [0, 4].

 6D. STIELTJES INVERSION 151

Proof. As before, we use the standard formula for the generating series of the Catalan
numbers. With z = ξ −1 in that formula, we obtain the following formula:
∞
X
−1
G(ξ) = ξ Ck ξ −k
k=0
p
−1 1− 1 − 4ξ −1
= ξ ·
2ξ −1
1  p 
= 1 − 1 − 4ξ −1
2
1 1p
= − 1 − 4ξ −1
2 2
With this in hand, let us apply now the Stieltjes inversion formula, from Theorem
6.32. We obtain, a bit as before in Theorem 6.35, the following density:
1 1 √ −1 1 √ −1
dµ(x) = − · − 4x − 1 dx = 4x − 1 dx
π 2 2π
Thus, we are led to the conclusion in the statement. □
Regarding now the central binomial coefficients, we have here:
Theorem 6.37. The
real probability measure having as moments the central binomial
2k
coefficients, Dk = k , is the measure
1
α1 = p dx
π x(4 − x)
called arcsine law on [0, 4].
Proof. We have the following computation, using formulae from chapter 3:
∞
X
−1
G(ξ) = ξ Dk ξ −k
k=0
∞  k
1X t
= Dk −
ξ k=0
4
1 1
= ·p
ξ 1 − 4/ξ
1
= p
ξ(ξ − 4)
But this gives the density in the statement, via Theorem 6.32. □
Finally, we have the following version of the above result:
152 6. COMPLEX FUNCTIONS

Theorem 6.38. The
real probability measure having as moments the middle binomial
k
coefficients, Ek = [k/2] , is the following law on [−2, 2],
r
1 2+x
σ1 = dx
2π 2 − x
called modified the arcsine law on [−2, 2].
Proof. In terms of the central binomial coefficients Dk , we have:
 
2k (2k)!
E2k = = = Dk
k k!k!
 
2k − 1 (2k − 1)! Dk
E2k−1 = = =
k k!(k − 1)! 2
Standard calculus based on the Taylor formula for (1 + t)−1/2 gives:
r ! ∞
1 1 + 2x X
−1 = Ek xk
2x 1 − 2x k=0

With x = ξ −1 we obtain the following formula for the Cauchy transform:

X∞
−1
G(ξ) = ξ Ek ξ −k
k=0
s !
1 1 + 2/ξ
= −1
ξ 1 − 2/ξ
s !
1 ξ+2
= −1
ξ ξ−2
By Stieltjes inversion we obtain the density in the statement. □
All this is very nice, and we are obviously building here, as this book goes by, some
good knowledge in probability theory. We will be back to all this later.
6e. Exercises
Here are a few exercises, in relation with the material from this chapter:
Exercise 6.39. Find a geometric proof for 1 + x + x2 + . . . = 1/(1 − x).
Exercise 6.40. Clarify the definition and properties of xy , with x, y ∈ C.
Exercise 6.41. Complete our proof of the Cauchy formula, in the general case.
Exercise 6.42. Learn more complex analysis, including the residue formula.
As a bonus exercise, learn some more probability, and random matrices.
 CHAPTER 7

Fourier analysis

7a. Function spaces

In this chapter we go back to the functions of one real variable, f : R → R or
f : R → C, with some applications, by using our complex function technology. We will
be mainly interested in constructing the Fourier transform, which is an operation f → fb
on such functions, which can solve various questions. Some advertisement first:
Advertisement 7.1. There is only one serious tool in mathematics, and that is the
Fourier transform.
Or at least, that’s what some people say. And as fans of calculus, we should of
course side with them. We will see for instance how, via Fourier transform, light can
be decomposed into color components, distant galaxies can be observed, atoms can be
distinguished from each other, the future of the universe can be predicted, and so on.

Before doing that, however, let us study the spaces that the functions f : R → C
can form. These functions can be continuous, differentiable, infinitely differentiable, and
so on, but there are many more properties that these functions can have, that we will
investigate now. This will lead to various spaces of functions f : R → C, that can be
used, among others, in order to well-define the Fourier transform operation f → fb.

Let us start with some well-known and useful inequalities, as follows:

Theorem 7.2. Given two functions f, g : R → C and an exponent p ≥ 1, we have
Z 1/p Z 1/p Z 1/p
p p p
|f + g| ≤ |f | + |g|
R R R

called Minkowski inequality. Also, assuming that p, q ≥ 1 satisfy 1/p + 1/q = 1, we have
Z Z 1/p Z 1/q
p q
|f g| ≤ |f | |g|
R R R

called Hölder inequality. These inequalities hold as well for ∞ values of the exponents.
Proof. All this is very standard, the idea being as follows:
153
154 7. FOURIER ANALYSIS

(1) As a first observation, at p = 2, which is a special exponent, we have q = 2 as well,

and the Minkowski and Hölder inequalities are as follows:
Z 1/2 Z 1/2 Z 1/2
2 2 2
|f + g| ≤ |f | + |g|
R R R
Z Z 1/2 Z 1/2
|f g| ≤ |f |2 |g|2
R R R
But the proof of the Hölder inequality, called Cauchy-Schwarz
R inequality in this case,
is something elementary, coming from the fact that I(t) = R |f + twg|2 with |w| = 1 is a
positive degree 2 polynomial in t ∈ R, and so its discriminant must be negative. As for
the Minkowski inequality, this follows from this, by taking squares and simplifying.
(2) In general, let us first prove Hölder, in the case of finite exponents, p, q ∈ (1, ∞).
By linearity we can assume that f, g are normalized, in the following way:
Z Z
p
|f | = |g|q = 1
R R
R
In this case, we want to prove that we have R |f g| ≤ 1. And for this purpose, we can
use the Young inequality from chapter 3, which gives, for any x ∈ R:
|f (x)|p |g(x)|q
|f (x)g(x)| ≤ +
p q
By integrating now over x ∈ R, we obtain from this, as desired:
|f |p |g|q
Z Z
1 1
|f g| ≤ + = + =1
R R p q p q
Let us prove now Minkowski, again in the finite exponent case, p ∈ (1, ∞). We have
the following estimate, using the Hölder inequality, and the conjugate exponent:
Z Z
p
|f + g| = |f + g| · |f + g|p−1
R
ZR Z
p−1
≤ |f | · |f + g| + |g| · |f + g|p−1
R R
Z 1/p Z 1/q
p (p−1)q
≤ |f | |f + g|
R R
Z 1/p Z 1/q
p (p−1)q
+ |g| |f + g|
R R
"Z 1/p Z 1/p # Z 1−1/p
p p p
= |f | + |g| |f + g|
R R R

Thus, we are led to the Minkowski inequality in the statement.

 7A. FUNCTION SPACES 155
R
(3) Finally, with the convention that ( R |f |p )1/p takes as value at p = ∞ the essential
supremum of f , the Minkowski inequality holds as well at p = ∞, trivially:
sup f + g ≤ sup f + sup g
The same goes for the Hölder inequality at p = ∞, q = 1, which is simply:
Z Z
|f g| ≤ sup f × |g|
R R

And finally, the same goes for the Hölder inequality at p = 1, q = ∞. □

As a consequence of the above results, we can formulate:
Theorem 7.3. Given an interval I ⊂ R and an exponent p ≥ 1, the following space,
with the convention that functions are identified up to equality almost everywhere,
 Z 
p p
L (I) = f : I → C |f (x)| dx < ∞
I

is a vector space, and the following quantity,

Z 1/p
p
||f ||p = |f (x)|
I

is a norm on it, in the sense that it satisfies the usual conditions for a vector space norm.
Moreover, Lp (I) is complete with respect to the distance d(f, g) = ||f − g||p .
Proof. This basically follows from Theorem 7.2, the idea being as follows:
(1) Again, let us first see what happens at p = 2. Here everything is standard from
what we have in Theorem 7.2, and with the remark that the space L2 (I) that we obtain is
more than just a normed vector space, √ because we have as well a scalar product, related
to the norm by the formula ||f ||2 = < f, f >, constructed as follows:
Z
< f, g >= f (x) g(x) dx
I

(2) In the general case now, where p ≥ 1 is still finite, but arbitrary, the proof is
similar, basically coming from the Minkowski inequality from Theorem 7.2.
(3) Finally, the extension at p = ∞ is clear too, coming from definitions, and with the
various conventions made at the end of the proof of Theorem 7.2. □
There are many more things that can be said about the above spaces L2 (I), which
are all good to know, and we refer here to any functional analysis book. Going ahead
now with our study of functions f : R → C, let us define an interesting operation on such
functions, called convolution, which is useful for many purposes. Let us start with:
156 7. FOURIER ANALYSIS

Definition 7.4. The convolution of two functions f, g : R → C is the function

Z
(f ∗ g)(x) = f (x − y)g(y)dy
R

provided that the function y → f (x − y)g(y) is indeed integrable, for any x.

There are many reasons for introducing this operation, that we will gradually discover,
in what follows. As a basic example, let us take g = χ[0,1] . We have then:
Z 1
(f ∗ g)(x) = f (x − y)dy
0

Thus, with this choice of g, the operation f → f ∗ g has some sort of “regularizing
effect”, that can be useful for many purposes. We will be back to this, later.

Goinh ahead with more theory, let us try to understand when the convolution opera-
tion is well-defined. We have here the following basic result:

Theorem 7.5. The convolution operation is well-defined on the space

n o
Cc (R) = f ∈ C(R) supp(f ) = compact

of continuous functions f : R → C having compact support.

Proof. We have several things to be proved, the idea being as follows:

(1) First we must show that given two functions f, g ∈ Cc (R), their convolution f ∗ g
is well-defined, as a function f ∗ g : R → C. But this follows from the following estimate,
where l denotes the length of the compact subsets of R:
Z Z
|f (x − y)g(y)|dy = |f (x − y)g(y)|dy
R supp(g)
Z
≤ max(g) |f (x − y)|dy
supp(g)
≤ max(g) · l(supp(g)) · max(f )
< ∞

(2) Next, we must show that the function f ∗ g : R → C that we constructed is indeed
continuous. But this follows from the following estimate, where Kf is the constant of
 7A. FUNCTION SPACES 157

uniform continuity for the function f ∈ Cc (R):

Z Z
|(f ∗ g)(x + ε) − (f ∗ g)(x)| = f (x + ε − y)g(y)dy − f (x − y)g(y)dy
R R
Z
= (f (x + ε − y) − f (x − y)) g(y)dy
R
Z
≤ |f (x + ε − y) − f (x − y)| · |g(y)|dy
R
Z
≤ Kf · ε · |g|
R
(3) Finally, we must show that the function f ∗ g ∈ C(R) that we constructed has
indeed compact support. For this purpose, our claim is that we have:
supp(f + g) ⊂ supp(f ) + supp(g)
In order to prove this claim, observe that we have, by definition of f ∗ g:
Z
(f ∗ g)(x) = f (x − y)g(y)dy
R
Z
= f (x − y)g(y)dy
supp(g)

But this latter quantity being 0 for x ∈

/ supp(f ) + supp(g), this gives the result. □
Here are now a few remarkable properties of the convolution operation:
Proposition 7.6. The following hold, for the functions in Cc (R):
(1) f ∗ g = g ∗ f .
(2) f ∗ (g ∗ h) = (f ∗ g) ∗ h.
(3) f ∗ (λg + µh) = λf ∗ g + µf ∗ h.
Proof. These formulae are all elementary, the idea being as follows:
(1) This follows from the following computation, with y = x − t:
Z
(f ∗ g)(x) = f (x − y)g(y)dy
R
Z
= f (t)g(x − t)dt
R
Z
= g(x − t)f (t)dt
R
= (g ∗ f )(x)
(2) This is clear from definitions.
(3) Once again, this is clear from definitions. □
158 7. FOURIER ANALYSIS

In relation with derivatives, and with the “regularizing effect” of the convolution
operation mentioned after Definition 7.4, we have the following result:
Theorem 7.7. Given two functions f, g ∈ Cc (R), assuming that g is differentiable,
then so is f ∗ g, with derivative given by the following formula:
(f ∗ g)′ = f ∗ g ′
More generally, given f, g ∈ Cc (R), and assuming that g is k times differentiable, then so
is f ∗ g, with k-th derivative given by (f ∗ g)(k) = f ∗ g (k) .
Proof. In what regards the first assertion, with y = x − t, then t = x − y, we get:
Z
′ d
(f ∗ g) (x) = f (x − y)g(y)dy
dx R
Z
d
= f (t)g(x − t)dt
dx R
Z
= f (t)g ′ (x − t)dt
ZR
= f (x − y)g ′ (y)dy
R
= (f ∗ g ′ )(x)
As for the second assertion, this follows form the first one, by recurrence. □
Finally, getting beyond the compactly supported continuous functions, we have the
following result, which is of particular theoretical importance:
Theorem 7.8. The convolution operation is well-defined on L1 (R), and we have:
||f ∗ g||1 ≤ ||f ||1 ||g||1
Thus, if f ∈ L1 (R) and g ∈ Cck (R), then f ∗ g is well-defined, and f ∗ g ∈ Cck (R).
Proof. In what regards the first assertion, this follows from the following computa-
tion, involving an intuitive manipulation on the double integrals, called Fubini theorem,
that we will use as such here, and that we will fully clarify later on, in this book:
Z Z Z
|(f ∗ g)(x)|dx ≤ |f (x − y)g(y)|dydx
R R R
Z Z
= |f (x − y)g(y)|dxdy
ZR R Z
= |f (x)|dx |g(y)|dy
R R
As for the second assertion, this follows from the first one, and from Theorem 7.7. □
 7B. FOURIER TRANSFORM 159

Summarizing, we have now some good knowledge of the various spaces that the func-
tions f : R → C can form, and we have as well an interesting regularization operation
f → f ∗ g on such functions, that can be used for various purposes.

7b. Fourier transform

We discuss here the construction and main properties of the Fourier transform, which
is the main tool in analysis, and even in mathematics in general. We first have:
Definition 7.9. Given f ∈ L1 (R), we define a function fb : R → C by
Z
fb(ξ) = eixξ f (x)dx
R

and call it Fourier transform of f .

As a first observation, even if f is a real function, fb is a complex function, which is
not necessarily real. Also, fb is obviously well-defined, because f ∈ L1 (R) and |eixξ | = 1.
Also, the condition f ∈ L1 (R) is basically needed for construcing fb, because:
Z
f (0) =
b f (x)dx
R

Generally speaking, the Fourier transform

R is there for helping with various computa-
tions, with the above formula f (0) = f being something quite illustrating. Here are
b
some basic properties of the Fourier transform, all providing some good motivations:
Proposition 7.10. The Fourier transform has the following properties:
(1) Linearity: f[ + g = fb + gb, λf
c = λfb.
(2) Regularity: fb is continuous and bounded.
(3) If f is even then fb is even.
(4) If f is odd then fb is odd.
Proof. These results are all elementary, as follows:
(1) The additivity formula is clear from definitions, as follows:
Z
[
f + g(ξ) = eixξ (f + g)(x)dx
ZR Z
= e f (x)dx + eixξ f (x)dx
ixξ
R R

= fb(ξ) + gb(ξ)
As for the formula λf
c = λfb, this is clear as well.
160 7. FOURIER ANALYSIS

(2) The continuity of fb follows indeed from:

Z
|f (ξ + ε) − f (ξ)| ≤
b b (eix(ξ+ε) − eixξ )f (x) dx
ZR
= eixξ (eixε − 1)f (x) dx
R
Z
ixε
≤ |e − 1| |f |
R

As for the boundedness of fb, this is clear as well.

(3) This follows from the following computation, assuming that f is even:
Z
f (−ξ) =
b e−ixξ f (x)dx
ZR
= eixξ f (−x)dx
ZR
= eixξ f (x)dx
R

= fb(ξ)
(4) The proof here is similar to the proof of (3), by changing some signs. □

We will be back to more theory in a moment, but let us explore now the examples.
Here are some basic computations of Fourier transforms:
Proposition 7.11. We have the following Fourier transform formulae,
2 sin(aξ)
f = χ[−a,a] =⇒ fb(ξ) =
ξ
1
f = e−ax χ[0,∞] (x) =⇒ fb(ξ) =
a − iξ
1
f = eax χ[−∞,0] (x) =⇒ fb(ξ) =
a + iξ
2a
f = e−a|x| =⇒ fb(ξ) =
a2 + ξ 2
2iξ
f = sgn(x)e−a|x| =⇒ fb(ξ) =
a2 + ξ2
valid for any number a > 0.
 7B. FOURIER TRANSFORM 161

Proof. All this follows from some calculus, as follows:

(1) In what regards first formula, assuming f = χ[−a,a] , we have, by using the fact
that sin(xξ) is an odd function, whose integral vanishes on centered intervals:
Z a
f (ξ) =
b eixξ dx
Z−aa Z a
= cos(xξ)dx + i sin(xξ)dx
−a −a
Z a
= cos(xξ)dx
−a
 a
sin(xξ)
=
ξ −a
2 sin(aξ)
=
ξ
−ax
(2) With f (x) = e χ[0,∞] (x), the computation goes as follows:
Z ∞
f (ξ) =
b eixξ−ax dx
Z0 ∞
= e(iξ−a)x dx
0
 (iξ−a)x ∞
e
=
iξ − a 0
1
=
a − iξ
(3) Regarding the third formula, this follows from the second one, by using the fol-
lowing fact, generalizing the parity computations from Proposition 7.10:
F (x) = f (−x) =⇒ Fb(ξ) = fb(−ξ)
(4) The last 2 formulae follow from what we have, by making sums and differences,
and the linearity properties of the Fourier transform, from Proposition 7.10. □
We will see many other examples, in what follows. Getting back now to theory, we
have the following result, adding to the various general properties in Proposition 7.10,
and providing more motivations for the Fourier transform:
Proposition 7.12. Given f, g ∈ L1 (R) we have fbg, f gb ∈ L1 (R) and
Z Z
f (ξ)b
g (ξ)dξ = fb(x)g(x)dx
R R
called “exchange of hat” formula.
162 7. FOURIER ANALYSIS

Proof. Regarding the fact that we have indeed fbg, f gb ∈ L1 (R), this is actually a bit
non-trivial, but we will be back to this later. Assuming this, we have:
Z Z Z
f (ξ)b
g (ξ)dξ = f (ξ)eixξ g(x)dxdξ
R R R
On the other hand, we have as well the following formula:
Z Z Z
fb(x)g(x)dx = eixξ f (x)g(ξ)dxdξ
R R R
Thus, with x ↔ ξ, we are led to the formula in the statement. □
As an important result, showing the power of the Fourier transform, this transforms
the derivative into something very simple, namely a multiplication by the variable:
Theorem 7.13. Given f : R → C such that f, f ′ ∈ L1 (R), we have:
fb′ (ξ) = −iξ fb(ξ)
More generally, assuming f, f ′ , f ′′ , . . . , f (n) ∈ L1 (R), we have
fd (k) (ξ) = (−iξ)k fb(ξ)

for any k = 1, 2, . . . , n.
Proof. These results follow by doing a partial integration, as follows:
(1) Assuming that f : R → C has compact support, we have indeed:
Z
fb′ (ξ) = eixξ f ′ (x)dx
R
Z
= − iξeixξ f (x)dx
R
Z
= −iξ eixξ f (x)dx
R

= −iξ fb(ξ)
(2) Regarding the higher derivatives, the formula here follows by recurrence. □
Importantly, we have a converse statement as well, as follows:
Theorem 7.14. Assuming that f ∈ L1 (R) is such that F (x) = xf (x) belongs to L1 (R)
too, the function fb is differentiable, with derivative given by:
(fb)′ (ξ) = iFb(ξ)
More generally, if Fk (x) = xk f (x) belongs to L1 (R), for k = 0, 1, . . . , n, we have
(fb)(k) (ξ) = ik F
ck (ξ)
for any k = 1, 2, . . . , n.
 7B. FOURIER TRANSFORM 163

Proof. These results are both elementary, as follows:

(1) Regarding the first assertion, the computation here is as follows:
Z
′ d
(fb) (ξ) = eixξ f (x)dx
dξ R
Z
= ixeixξ f (x)dx
R
Z
= i eixξ xf (x)dx
R

= iFb(ξ)
(2) As for the second assertion, this follows from the first one, by recurrence. □
As a conclusion to all this, we are on a good way with our theory, and we have:
Conclusion 7.15. Modulo normalization factors, the Fourier transform converts the
derivatives into multiplications by the variable, and vice versa.
And isn’t this interesting, because isn’t computing derivatives a difficult task. Here is
now another useful result, of the same type, this time regarding convolutions:
Theorem 7.16. Assuming f, g ∈ L1 (R), the following happens:

∗ g = fb · gb
f[
Moreover, under suitable assumptions, the formula fcg = fb ∗ gb holds too.
Proof. This is something quite subtle, the idea being as follows:
(1) Regarding the first assertion, this is something elementary, as follows:
Z
f ∗ g(ξ) =
[ eixξ (f ∗ g)(x)dx
ZR Z
= eixξ f (x − y)g(y)dxdy
ZR R Z 
iyξ i(x−y)ξ
= e e f (x − y)dx g(y)dy
R R
Z Z 
iyξ itξ
= e e f (t)dt g(y)dy
R R
Z
= eiyξ fb(ξ)g(y)dy
R

= fb(ξ)b
g (ξ)
164 7. FOURIER ANALYSIS

(2) As for the second assertion, this is something more tricky, and we will be back to
it later. In the meantime, here is however some sort of proof, not very honest:
Z
(f ∗ gb)(ξ) =
b fb(ξ − η)b g (η)dη
R
Z Z Z
= eix(ξ−η) f (x)eiyη g(y)dxdydη
ZR ZR ZR
= eixη ei(y−x)η f (x)g(y)dxdydη
ZR R R
= eixη f (x)g(x)dx
R

= fcg(η)
To be more precise, the point here is that we can pass from the triple to the single
integral by arguing that “we must have x = y”. We will be back to this later. □
As an updated conclusion to all this, we have, modulo a few bugs, to be fixed:
Conclusion 7.17. The Fourier transform converts the derivatives into multiplications
by the variable, and convolutions into products, and vice versa.
We will see applications of this later, after developing some more general theory. So,
let us develop now more theory for the Fourier transform. We first have:
Theorem 7.18. Given f ∈ L1 (R), its Fourier transform satisfies
lim fb(ξ) = 0
ξ→±∞

called Riemann-Lebesgue property of fb.

Proof. This is something quite technical, as follows:
(1) Given a function f : R → C and a number y ∈ R, let us set:
fy (x) = f (x − y)
Our claim is then that if f ∈ Lp (R), then the following function is uniformly continu-
ous, with respect to the usual p-norm on the right:
R → Lp (R) , y → fy
(2) In order to prove this, fix ε > 0. Since f ∈ Lp (R), we can find a function of type
g : [−K, K] → C which is continuous, such that:
||f − g||p < ε
Now since g is uniformly continuous, we can find δ ∈ (0, K) such that:
|s − t| < δ =⇒ |g(s) − g(t)| < (3K)−1/p ε
 7B. FOURIER TRANSFORM 165

But this shows that we have the following estimate:

Z 1/p
p
||gs − gt ||p = g(x − s) − g(x − t) dx
R
1/p
< (3K)−1 εp (2k + δ)


< ε
By using now the formula ||f ||p = ||fs ||p , which is clear, we obtain:
||fs − ft ||p ≤ ||fs − gs ||p + ||gs − gt ||p + ||gt − ft ||p
< ε+ε+ε
= 3ε
But this being true for any |s − t| < δ, we have proved our claim.
(3) Let us prove now the Riemann-Lebesgue property of fb, as formulated in the
statement. By using eπi = −1, and the change of variables x → x − π/ξ, we have:
Z
f (ξ) =
b eixξ f (x)dx
R
Z
= − eixξ eπi f (x)dx
ZR
= − eiξ(x+π/ξ) f (x)dx
ZR  
ixξ π
= − e f x− dx
R ξ
On the other hand, we have as well the following formula:
Z
f (ξ) =
b eixξ f (x)dx
R

Thus by summing, we obtain the following formula:

Z   
ixξ π
2fb(ξ) = e f (x) − f x − dx
R ξ
But this gives the following estimate, with notations from (1):
2|fb(ξ)| ≤ ||f − fπ/ξ ||1
Since by (1) this goes to 0 with ξ → ±∞, this gives the result. □
Quite remarkably, and as a main result now regarding Fourier transforms, a function
f : R → C can be recovered from its Fourier transform fb : R → C, as follows:
166 7. FOURIER ANALYSIS

Theorem 7.19. Assuming f, fb ∈ L1 (R), we have

Z
f (x) = e−ixξ fb(ξ)dξ
R

almost everywhere, called Fourier inversion formula.

Proof. This is something quite tricky, due to the fact that a direct attempt by double
integration fails. Consider the following function, depending on a parameter λ > 0:
Z
φλ (x) = e−ixξ−λ|ξ| dξ
R

We have then the following computation:

Z
(f ∗ φλ )(x) = f (x − y)φλ (y)dy
ZR Z
= f (x − y)e−iyξ−λ|ξ| dξdy
ZR R Z 
−λ|ξ| −iyξ
= e f (x − y)e dy dξ
R R
Z
= e−λ|ξ| e−ixξ fb(ξ)dξ
R

By letting now λ → 0, we obtain from this the following formula:

Z
lim (f ∗ φλ )(x) = e−ixξ fb(ξ)dξ
λ→0 R

On the other hand, by using Theorem 7.18 we obtain that, almost everywhere:
lim (f ∗ φλ )(x) = f (x)
λ→0

Thus, we are led to the conclusion in the statement. □

There are many more things that can be said about Fourier transforms, a key result
being the Plancherel formula, allowing us to talk about the Fourier transform over the
space L2 (R). Also, we can talk about the Fourier transform over the space S of functions
all whose derivatives are rapidly decreasing, called Schwartz space. It is beyond our scope
here to discuss all this, and we recommend instead any good book on the subject.

7c. Groups, extensions

All the above theory concerns the Fourier transform over R, but there are some other
types of Fourier transforms too, the idea being that we have such a transform over any
locally compact abelian group G. In order to discuss this, let us start with:
 7C. GROUPS, EXTENSIONS 167

Definition 7.20. An abelian group is a set G with a multiplication operation

(g, h) → gh
which must satisfy the following conditions:
(1) Commutativity: we have gh = hg, for any g, h ∈ G.
(2) Associativity: we have (gh)k = g(hk), for any g, h, k ∈ G.
(3) Unit: there is an element 1 ∈ G such that g1 = g, for any g ∈ G.
(4) Inverses: for any g ∈ G there is g −1 ∈ G such that gg −1 = 1.
There are many examples of abelian groups, and in order to understand what is going
on, and to develop our Fourier transform theory, let us first restrict the attention to the
finite group case. Here we have as basic examples the cyclic group ZN , or more generally
the products of such cyclic groups, which are clearly both finite and abelian:
G = ZN1 × . . . × ZNk
Our goal in what follows will be that of proving that these are in fact all the finite
abelian groups, and also that a Fourier transform theory, called “discrete Fourier trans-
form”, can be developed over such groups. Let us start our study with:
Proposition 7.21. Given a finite abelian group G, the group morphisms
χ:G→T
with T being the unit circle, called characters of G, form a finite abelian group G.
b

Proof. There are several things to be proved here, the idea being as follows:
(1) Our first claim is that G
b is a group, with the pointwise multiplication, namely:
(χρ)(g) = χ(g)ρ(g)
Indeed, if χ, ρ are characters, so is χρ, and so the multiplication is well-defined on
b Regarding the unit, this is the trivial character 1 : G → T, mapping g → 1, for any
G.
g ∈ G. Finally, we have inverses, with the inverse of χ : G → T being its conjugate:
χ̄ : G → T , g → χ(g)
(2) Our next claim is that Gb is finite. Indeed, given a group element g ∈ G, we can
talk about its order, which is smallest integer k ∈ N such that g k = 1. Now assuming
that we have a character χ : G → T, we have the following formula:
χ(g)k = 1
Thus χ(g) must be one of the k-th roots of unity, and in particular there are finitely
many choices for χ(g). Thus, there are finitely many choices for χ, as desired.
(3) Finally, the fact that G
b is abelian follows from definitions, because the pointwise
multiplication of functions, and in particular of characters, is commutative. □
168 7. FOURIER ANALYSIS

The above construction is quite interesting, and we have:

Theorem 7.22. The character group operation G → G b for the finite abelian groups,
called Pontrjagin duality, has the following properties:
(1) The dual of a cyclic group is the group itself, Z
b N = ZN .
(2) The dual of a product is the product of duals, G\ ×H =G b × H.
b
(3) Any product of cyclic groups G = ZN1 × . . . × ZN is self-dual, G = G.
k
b

Proof. We have several things to be proved, the idea being as follows:

(1) A character χ : ZN → T is uniquely determined by its value z = χ(g) on the
standard generator g ∈ ZN . But this value must satisfy:
zN = 1
Thus we must have z ∈ ZN , with the cyclic group ZN being regarded this time as
being the group of N -th roots of unity. Now conversely, any N -th root of unity z ∈ ZN
defines a character χ : ZN → T, by setting, for any r ∈ N:
χ(g r ) = z r
Thus we have an identification Z
b N = ZN , as claimed.
(2) A character of a product of groups χ : G × H → T must satisfy:
χ(g, h) = χ [(g, 1)(1, h)] = χ(g, 1)χ(1, h)
Thus χ must appear as the product of its restrictions χ|G , χ|H , which must be both
characters, and this gives the identification in the statement.
(3) This follows from (1) and (2). Alternatively, any character χ : G → T is uniquely
determined by its values χ(g1 ), . . . , χ(gk ) on the standard generators of ZN1 , . . . , ZNk ,
which must belong to ZN1 , . . . , ZNk ⊂ T, and this gives Gb = G, as claimed. □
At a more advanced level now, we have the following result:
Theorem 7.23. The finite abelian groups are the following groups,
G = ZN1 × . . . × ZNk
and these groups are all self-dual, G = G.
b

Proof. This is something quite tricky, the idea being as follows:

(1) In order to prove our result, assume that G is finite and abelian. For any prime
number p ∈ N, let us define Gp ⊂ G to be the subset of elements having as order a power
of p. Equivalently, this subset Gp ⊂ G can be defined as follows:
n k
o
Gp = g ∈ G ∃k ∈ N, g p = 1
 7C. GROUPS, EXTENSIONS 169

(2) It is then routine to check, based on definitions, that each Gp is a subgroup. Our
claim now is that we have a direct product decomposition as follows:
Y
G= Gp
p

(3) Indeed, by using the fact that our group G isQabelian, we have a morphism as
follows, with the order of the factors when computing p gp being irrelevant:
Y Y
Gp → G , (gp ) → gp
p p

Moreover, it is routine to check that this morphism is both injective and surjective,
via some simple manipulations, so we have our group decomposition, as in (2).
(4) Thus, we are left with proving that each component Gp decomposes as a product
of cyclic groups, having as orders powers of p, as follows:
Gp = Zpr1 × . . . × Zprs
But this is something that can be checked by recurrence on |Gp |, via some routine
computations, and we are led to the conclusion in the statement.
(5) Finally, the fact that the finite abelian groups are self-dual, G = G,
b follows from
the structure result that we just proved, and from Theorem 7.22 (3). □
In relation now with Fourier analysis, the result is as follows:
Theorem 7.24. Given a finite abelian group G, we have an isomorphism as follows,
obtained by linearizing/delinearizing the characters,
C ∗ (G) ≃ C(G)
b

where C ∗ (G) is the algebra of functions φ : G → C, with convolution product, and C(G)
b
b → C, with usual product.
is the algebra of functions φ : G
Proof. There are several things going on here, the idea being as follows:
(1) Given a finite abelian group G, we can talk about the complex algebra C(G)
formed by the complex functions φ : G → C, with usual product, namely:
(φψ)(g) = φ(g)ψ(g)
Observe that we have C(G) ≃ CN as an algebra, where N = |G|, with this being best
seen via the basis of C(G) formed by the Dirac masses at the points of G:
( )
X
C(G) = λg δg λg ∈ C
g∈G
170 7. FOURIER ANALYSIS

(2) On the other hand, we can talk as well about the algebra C ∗ (G) formed by the
same functions φ : G → C, but this time with the convolution product, namely:
X
(φ ∗ ψ)(g) = φ(gh−1 )ψ(h)
h∈G

Since we have δk ∗ δl = δkl for any k, l ∈ G, as you can easily check by using the above
formula, the Dirac masses δg ∈ C ∗ (G) behave like the group elements g ∈ G. Thus, we
can view our algebra as follows, with multiplication given by g · h = gh, and linearity:
( )
X
C ∗ (G) = λg g λg ∈ C
g∈G

(3) Now that we know what the statement is about, let us go for the proof. The first
observation is that we have a morphism of algebras as follows:

C ∗ (G) → C(G)
b , g → [χ → χ(g)]
Now since on both sides we have vector spaces of dimension N = |G|, it is enough
to check that this morphism is injective. But this is best done via Theorem 7.23, which
shows that the characters χ ∈ G
b separate the points g ∈ G, as desired. □

We can feel that Theorem 7.24 is related to Fourier analysis, and we have:
Fact 7.25. The following happen, regarding the locally compact abelian groups:
(1) What we did in the finite case, namely group characters, and construction and
basic properties of the dual, can be extended to them.
(2) As basic examples of this, besides what we have in the finite case, and notably
b N = ZN , we have Z
Z b = T, Tb = Z, and also Rb = R.
∗
(3) With some care for analytic aspects, C (G) ≃ C(G) b remains true in this setting,
and in the case G = R, this isomorphism is the Fourier transform.
Obviously, all this is a bit heavy, but you get the point, we have 3 types of Fourier
analysis in life, namely the “standard” one that we previously learned in this chapter,
corresponding to G = R, then another one that we skipped, and that we encourage
you to learn, called the “Fourier series” one, corresponding to G = Z, T, and finally the
“discrete” one that we started to learn, over G = ZN and other finite abelian groups.

In practice, all this is a bit complicated, and back now to the finite abelian groups, let
us work out a softer version of all the above, which is what is really needed, in practice,
when doing discrete Fourier analysis. For G = ZN , what we need is:
 7C. GROUPS, EXTENSIONS 171

Definition 7.26. The Fourier matrix FN is the following matrix, with w = e2πi/N :
 
1 1 1 ... 1
1
 w w2 . . . wN −1  
1 w 2 4 2(N −1) 
FN =  w . . . w
 .. .. .. .. 

. . . . 
N −1 2(N −1) (N −1)2
1 w w ... w
That is, FN = (wij )ij , with indices i, j ∈ {0, 1, . . . , N − 1}, taken modulo N .
Observe that this matrix is Hadamard, in the sense that its entries are on the unit
circle, and the rows are pairwise orthogonal. In fact, in general, we have:
b = {χ : G → T},
Theorem 7.27. Given a finite abelian group G, with dual group G
consider the corresponding Fourier coupling, namely:
FG : G × Gb → T , (i, χ) → χ(i)
(1) Via the standard isomorphism G ≃ G, b this Fourier coupling can be regarded as a
square matrix, FG ∈ MG (T), which is a complex Hadamard matrix.
(2) In the case of the cyclic group G = ZN we obtain in this way, via the standard
identification ZN = {1, . . . , N }, the Fourier matrix FN .
(3) In general, when using a decomposition G = ZN1 × . . . × ZNk , the corresponding
Fourier matrix is given by FG = FN1 ⊗ . . . ⊗ FNk .
Proof. This follows indeed by using the above finite abelian group theory:
(1) With the identification G ≃ G
b made our matrix is given by (FG )iχ = χ(i), and the
scalar products between the rows are computed as follows:
X X
< Ri , Rj >= χ(i)χ(j) = χ(i − j) = |G| · δij
χ χ

Thus, we obtain indeed a complex Hadamard matrix.

(2) This follows from the well-known and elementary fact that, via the identifications
2πi/N
ZN = ZcN = {1, . . . , N }, the Fourier coupling here is as follows, with w = e :
(i, j) → wij
(3) We use here the following formula that we know, for the duals of products:
×K =H
H\ b ×K
b
At the level of the corresponding Fourier couplings, we obtain from this:
FH×K = FH ⊗ FK
Now by decomposing G into cyclic groups, as in the statement, and by using (2) for
the cyclic components, we obtain the formula in the statement. □
172 7. FOURIER ANALYSIS

As a nice application of discrete Fourier analysis, we have:

Theorem 7.28. For a matrix M ∈ MN (C), the following are equivalent:
(1) M is circulant, Mij = ξj−i , for a certain vector ξ ∈ CN .
(2) M is Fourier-diagonal, M = FN QFN∗ , for a certain diagonal matrix Q.
Moreover, if these conditions hold, then ξ = FN∗ q, where q = (Q11 , . . . , QN N ).
Proof. This follows from some computations with roots of unity, as follows:
(1) =⇒ (2) Assuming Mij = ξj−i , the matrix Q = FN∗ M FN is indeed diagonal, as
shown by the following computation:
X
Qij = wjl−ik ξl−k
kl
X
= wj(k+r)−ik ξr
kr
X
= N δij wjr ξr
r

(2) =⇒ (1) Assuming Q = diag(q1 , . . . , qN ), the matrix M = FN QFN∗ is indeed

circulant, as shown by the following computation:
X X
Mij = wik Qkk w−jk = w(i−j)k qk
k k

P Indeed, since the last term depends only on j − i, we have Mij = ξj−i , with ξi =
−ik
kw qk = (FN∗ q)i . Thus, we are led to the conclusions in the statement. □

As an illustration for the above result, the all-one matrix diagonalizes as follows:
 
1 ... ... 1 
N

 .. .. 
1 0
. .   ∗
. . = FN  ..  FN
 .. ..  N  . 
1 ... ... 1 0

But you might know this already, since this was a recurrent exercise, in this book.

7d. Independence, limits

With the above theory in hand, let us go back to probability. Our claim is that things
are very interesting here, with the real-life notion of independence corresponding to our
mathematical notion of convolution, and with the Fourier transform being something very
useful in probability, in order to understand the independence. Let us start with:
 7D. INDEPENDENCE, LIMITS 173

Definition 7.29. Two variables f, g ∈ L∞ (X) are called independent when

E(f k g l ) = E(f k ) E(g l )
happens, for any k, l ∈ N.
This definition hides some non-trivial things. Indeed, by linearity, we would like to
have a formula as follows, valid for any polynomials P, Q ∈ R[X]:
E[P (f )Q(g)] = E[P (f )] E[Q(g)]
By a continuity argument, it is enough to have this formula for the characteristic
functions χI , χJ of the arbitrary measurable sets of real numbers I, J ⊂ R:
E[χI (f )χJ (g)] = E[χI (f )] E[χJ (g)]
Thus, we are led to the usual definition of independence, namely:
P (f ∈ I, g ∈ J) = P (f ∈ I) P (g ∈ J)
All this might seem a bit abstract, but in practice, the idea is of course that f, g must
be independent, in an intuitive, real-life sense. As a first result now, we have:
Theorem 7.30. Assuming that f, g ∈ L∞ (X) are independent, we have
µf +g = µf ∗ µg
where ∗ is the convolution of real probability measures.
Proof. We have the following computation, using the independence of f, g:
Mk (f + g) = E((f + g)k )
X k 
= E(f r g k−r )
r
r
X k 
= Mr (f )Mk−r (g)
r
r
On the other hand, we have as well the following computation:
Z Z
k
x d(µf ∗ µg )(x) = (x + y)k dµf (x)dµg (y)
R R×R
X k  Z Z
= x dµf (x) y k−r dµg (y)
r

r
r R R
X k  
= Mr (f )Mk−r (g)
r
r
Thus µf +g and µf ∗ µg have the same moments, and so they coincide, as claimed. □
Here is now a second result on independence, which is something more advanced:
174 7. FOURIER ANALYSIS

Theorem 7.31. Assuming that f, g ∈ L∞ (X) are independent, we have

Ff +g = Ff Fg
where Ff (x) = E(eixf ) is the Fourier transform.
Proof. We have the following computation, by using Theorem 7.30:
Z
Ff +g (x) = eixz dµf +g (z)
ZR
= eixz d(µf ∗ µg )(z)
ZR
= eix(z+t) dµf (z)dµg (t)
ZR×R Z
= e dµf (z) eixt dµg (t)
ixz
R R
= Ff (x)Fg (x)
Thus, we are led to the conclusion in the statement. □
In order to work out some basic illustrations, let us go back to the Poisson laws pt ,
from chapter 4. We can now say more about them, first with the following result:
Theorem 7.32. We have the following formula, for any s, t > 0,
ps ∗ pt = ps+t
so the Poisson laws form a convolution semigroup.
Proof. By using δk ∗ δl = δk+l and the binomial formula, we obtain:
X sk X tl
ps ∗ pt = e−s δk ∗ e−t δl
k
k! l
l!
X X sk t l
= e−s−t δn
n k+l=n
k!l!
X δn X n!
= e−s−t sk tl
n
n! k+l=n
k!l!
X (s + t)n
= e−s−t δn
n
n!
= ps+t
Thus, we are led to the conclusion in the statement. □
Next in line, we have the following result, which is fundamental as well:
 7D. INDEPENDENCE, LIMITS 175

Theorem 7.33. The Poisson laws appear as formal exponentials

X tk (δ1 − δ0 )∗k
pt =
k
k!

with respect to the convolution of measures ∗.

Proof. By using the binomial formula, the measure on the right is:
X tk X k!
µ = (−1)s δr
k
k! r+s=k
r!s!
X X δr
= tk (−1)s
k r+s=k
r!s!
X tr δr X (−1)s
=
r
r! s s!
1 X tr δr
=
e r
r!
= pt
Thus, we are led to the conclusion in the statement. □
Regarding now the Fourier transform computation, this is as follows:
Theorem 7.34. The Fourier transform of pt is given by
Fpt (y) = exp (eiy − 1)t

for any t > 0.

Proof. We have indeed the following computation:
X tk
Fpt (y) = e−t Fδ (y)
k
k! k
X tk
= e−t eiky
k
k!
X (eiy t)k
= e−t
k
k!
= exp(−t) exp(eiy t)
= exp (eiy − 1)t

Thus, we obtain the formula in the statement. □

176 7. FOURIER ANALYSIS

Observe that the above formula gives an alternative proof for Theorem 7.32, by the
using the fact that the logarithm of the Fourier transform linearizes the convolution. As
another application, we can now establish the Poisson Limit Theorem, as follows:
Theorem 7.35 (PLT). We have the following convergence, in moments,
  ∗n
t t
1− δ0 + δ1 → pt
n n
for any t > 0.
Proof. Let us denote by νn the measure under the convolution sign, namely:
 
t t
νn = 1 − δ0 + δ1
n n
We have the following computation, for the Fourier transform of the limit:
 
iry t t
Fδr (y) = e =⇒ Fνn (y) = 1 − + eiy
n n
  n
t t iy
=⇒ Fνn∗n (y) = 1− + e
n n
 iy
n
(e − 1)t
=⇒ Fνn∗n (y) = 1 +
n
iy


=⇒ F (y) = exp (e − 1)t
Thus, we obtain indeed the Fourier transform of pt , as desired. □
We have accumulated so far a lot of probability knowledge, going along with our
learning of calculus, in chapter 4, chapter 6, and in the present chapter. All this certainly
needs some systematic discussion, and we will be back to it once we will have the full
calculus tools that are needed, in chapter 14, which will be dedicated to probability.
7e. Exercises
This was a dense, introductory chapter to Fourier analysis, with several technical
things missing, due to some lack of space, and here are some exercises on all this:
Exercise 7.36. Clarify the proof of fcg = fb ∗ gb, say via Fourier inversion.
Exercise 7.37. Learn about the Plancherel formula, and Fourier over L2 (R).
Exercise 7.38. Learn about the Schwartz space S, and Fourier over it.
Exercise 7.39. Further clarify the Fourier transforms over ZN , Z, T, R.
As bonus exercise, learn some probability. Indeed, as mentioned above, we already
know some, but we will take a long break from this, until chapter 14 below.
 CHAPTER 8

Harmonic functions

8a. Laplace operator

Welcome to theoretical physics. We will discuss in this chapter some applications of
the theory of complex functions developed so far, to some questions from physics, along
with some more theory, needed in order to reach to these applications. We will be mainly
interested in two fundamental equations of physics, namely the wave equation, and the
heat equation. We will have a look as well into basic quantum mechanics.
All this sounds very exciting, doesn’t it. However, in practice, technically speaking,
we are not exactly ready yet for all these things. But the temptation remains high to talk
about them, a bit like physicists do. So, what to do. Ask the cat of course, who says:
Cat 8.1. The more you know, the better that is.
Thanks cat. So, we will talk indeed about all this, a bit like physicists do, by taking
some freedom with surfing on difficult mathematics. In short, this will be a physics
chapter, and for a more rigorous treatement of all this, which will be more general too,
do not worry, that will come, later in this book, once we will know more calculus.
So, let us start, with some mathematics. We have:
Definition 8.2. A function f : X → C with X ⊂ C is called differentiable in the real
sense if the following limits, called its partial derivatives,
df f (z + t) − f (z)
(z) = lim
dx t→0 t
df f (z + it) − f (z)
(z) = lim
dy t→0 t
computed by using t ∈ R, t → 0, exist at any z ∈ X, and are continuous.
We have already met this notion, in chapter 6, in a purely mathematical context,
when taking about holomorphic functions. Our observation there was that a holomorphic
function is differentiable in the above real sense, and also that there are further functions,
such as the conjugate f (z) = z̄ or the modulus f (z) = |z|, which are still differentiable in
the above real sense, but are not holomorphic. But more on this later.
We will need as well the following definition, complementing Definition 8.2:
177
178 8. HARMONIC FUNCTIONS

Definition 8.3. A function f : X → C is called doubly differentiable if its partial

derivatives are both differentiable. These double derivatives are denoted as follows:
d2 f d2 f
   
d df d df
= , =
dx2 dx dx dxdy dx dy
d2 f d2 f
   
d df d df
= , =
dydx dy dx dy 2 dy dy
We can in fact talk, in a similar way, about order k partial derivatives, for any k ∈ N.
As before with Definition 8.2, there are many things that can be said here, mathemat-
ically speaking, but we will leave this discussion for later, starting with chapter 9 below,
when systematically investigating functions of several variables. In fact, the material in
this chapter will be, among others, an introduction to that. However, we will need:
Theorem 8.4. The double derivatives satisfy the formula
d2 f d2 f
=
dxdy dydx
called Clairaut formula.
Proof. This is something very standard, the idea being as follows:
(1) Before pulling out a formal proof, as an intuitive justification for our formula, let
us consider a product of power functions, f (z) = xp y q . We have then:
d2 f d dxp y q
 
d
qxp y q−1 = pqxp−1 y q−1


= =
dxdy dx dy dx
d2 f d dxp y q
 
d
pxp−1 y q = pqxp−1 y q−1


= =
dydx dy dx dy
Next, let us consider a linear combination of power functions, f (z) = pq cpq xp y q ,
P
which can be finite or not. We have then, by using the above computation:
d2 f d2 f X
= = cpq pqxp−1 y q−1
dxdy dydx pq

Thus, we can see that our commutation formula for derivatives holds indeed, and this
due to the fact that the functions in x and y commute. Of course, all this does not prove
our formula, in general. But exercise for you, to have this idea fully working.
(2) Getting now to more standard techniques, given a point in the complex plane,
z = a + ib, consider the following functions, depending on h, k ∈ R small:
u(h, k) = f (a + h, b + k) − f (a + h, b)
v(h, k) = f (a + h, b + k) − f (a, b + k)
w(h, k) = f (a + h, b + k) − f (a + h, b) − f (a, b + k) + f (a, b)
 8A. LAPLACE OPERATOR 179

By the mean value theorem, for h, k ̸= 0 we can find α, β ∈ R such that:

w(h, k) = u(h, k) − u(0, k)
d
= h· u(αh, k)
dx 
d d
= h f (a + αh, b + k) − f (a + αh, b)
dx dx
d d
= hk · · f (a + αh, b + βk)
dy dx
Similarly, again for h, k ̸= 0, we can find γ, δ ∈ R such that:
w(h, k) = v(h, k) − v(h, 0)
d
= k· v(h, δk)
dy
 
d d
= k f (a + h, b + δk) − f (a, b + δk)
dy dy
d d
= hk · · f (a + γh, b + δk)
dx dy
Now by dividing everything by hk ̸= 0, we conclude from this that the following
equality holds, with the numbers α, β, γ, δ ∈ R being found as above:
d d d d
· f (a + αh, b + βk) = · f (a + γh, b + δk)
dy dx dx dy
But with h, k → 0 we get from this the Clairaut formula, at z = a + ib, as desired. □
Regarding the double derivatives in one of the variables, here our intuition from the
one-variable case can help, and in fact there is no need for more, at least of this stage of
the things. But, as already mentioned, we will be back to this, later in this book.

Getting now into physics, which will be our starting point for the considerations in
this chapter, we will be interested in the propagation of waves and heat. Let us start with
the waves. In analogy with what we saw in chapter 3, in one dimension, we have:
Theorem 8.5. The wave equation in the plane R2 is
φ̈ = v 2 ∆φ
where dots denote time derivatives, ∆ is the Laplace operator, given by
d2 φ d2 φ
∆φ = + 2
dx2 dy
and v > 0 is the propagation speed.
180 8. HARMONIC FUNCTIONS

Proof. We have already met this equation in chapter 3, in one dimension, and in 2
dimensions the study is similar, by using a lattice model, as follows:
(1) In order to understand the propagation of waves in 2 dimensions, we can model
the whole space R2 as a network of balls, with springs between them, as follows:

• • • •

• • • •

• • • •

As before in chapter 3 one dimension, let us send now an impulse, and zoom on one
ball. The situation here is as follows, with l being the spring length:
•φ(x,y+l)

•φ(x−l,y) •φ(x,y) •φ(x+l,y)

•φ(x,y−l)
We have two forces acting at (x, y). First is the Newton motion force, mass times
acceleration, which is as follows, with m being the mass of each ball:
Fn = m · φ̈(x, y)
And second is the Hooke force, displacement of the spring, times spring constant.
Since we have four springs at (x, y), this is as follows, k being the spring constant:
Fh = Fhr − Fhl + Fhu − Fhd
= k(φ(x + l, y) − φ(x, y)) − k(φ(x, y) − φ(x − l, y))
+ k(φ(x, y + l) − φ(x, y)) − k(φ(x, y) − φ(x, y − l))
= k(φ(x + l, y) − 2φ(x, y) + φ(x − l, y))
+ k(φ(x, y + l) − 2φ(x, y) + φ(x, y − l))
We conclude that the equation of motion, in our model, is as follows:
m · φ̈(x, y) = k(φ(x + l, y) − 2φ(x, y) + φ(x − l, y))
+ k(φ(x, y + l) − 2φ(x, y) + φ(x, y − l))
(2) Now let us take the limit of our model, as to reach to continuum. For this purpose
we will assume that our system consists of B 2 >> 0 balls, having a total mass M , and
 8A. LAPLACE OPERATOR 181

spanning a total area L2 . Thus, our previous infinitesimal parameters are as follows, with
K being the spring constant of the total system, taken to be equal to k:
M L
m= 2 , k=K , l=
B B
With these changes, our equation of motion found in (1) reads:
KB 2
φ̈(x, y) = (φ(x + l, y) − 2φ(x, y) + φ(x − l, y))
M
KB 2
+ (φ(x, y + l) − 2φ(x, y) + φ(x, y − l))
M
Now observe that this equation can be written, more conveniently, as follows:
KL2 φ(x + l, y) − 2φ(x, y) + φ(x − l, y)
φ̈(x, y) = ×
M l2
2
KL φ(x, y + l) − 2φ(x, y) + φ(x, y − l)
+ ×
M l2
With N → ∞, and therefore l → 0, we obtain in this way:
KL2 d2 φ d2 φ
 
φ̈(x, y) = + 2 (x, y)
M dx2 dy
Thus,
p we are led in this way to the following wave equation in two dimensions, with
v = K/M · L being the propagation speed of our wave:
 2
d φ d2 φ

2
φ̈(x, y) = v + 2 (x, y)
dx2 dy
But we recognize at right the Laplace operator, and we are done. As before in 1D,
there is of course some discussion to be made here, arguing that our spring model in (1)
is indeed the correct one. But do not worry, experiments confirm our findings.
(3) Finally, for completness, let us mention that the same argument, namely a lattice
model, carries on in arbitrary N dimensions, and we obtain here the same wave equation as
before, namely φ̈ = v 2 ∆φ, with the following straightforward definition for the Laplacian,
based on the obvious N -dimensional analogues of Definition 8.2 and Definition 8.3:
N
X d2 φ
∆φ =
i=1
dx2i
This is quite nice, because at N = 1 this Laplace operator is just the second derivative,
so what we have here is a unification of what we did in chapter 3 and in (1,2) above. In
addition, at N = 3 we obtain in this way the wave equation in the case of main interest,
namely the one of our real-life world. But more on this later in this book. □
Regarding now heat diffusion, we have here a similar equation, as follows:
182 8. HARMONIC FUNCTIONS

Theorem 8.6. Heat diffusion in R2 is described by the heat equation

φ̇ = α∆φ
where α > 0 is the thermal diffusivity of the medium, and ∆ is the Laplace operator.
Proof. We have already met this equation in chapter 3, in one dimension, and in 2
dimensions the study is similar, by using a lattice model, as follows:
(1) In order to understand the propagation of heat in 2 dimensions, we can model the
whole space R2 as a network as follows, with all lengths being l > 0:

◦ ◦ ◦ ◦

◦ ◦ ◦ ◦

◦ ◦ ◦ ◦

We have to implement now the physical heat diffusion mechanism, namely “the rate of
change of the temperature of the material at any given point must be proportional, with
proportionality factor α > 0, to the average difference of temperature between that given
point and the surrounding material”. In practice, this leads to a condition as follows,
expressing the change of the temperature φ, over a small period of time δ > 0:
αδ X
φ(x, y, t + δ) = φ(x, y, t) + 2 [φ(u, v, t) − φ(x, y, t)]
l
(x,y)∼(u,v)

In fact, we can rewrite our equation as follows, making it clear that we have here an
equation regarding the rate of change of temperature at x:
φ(x, y, t + δ) − φ(x, y, t) α X
= 2 [φ(u, v, t) − φ(x, y, t)]
δ l
(x,y)∼(u,v)

(2) Now, let us do the math. In the context of our 2D model the neighbors of x are
the points (x ± l, y ± l), so the equation above takes the following form:
φ(x, y, t + δ) − φ(x, y, t)
δ
αh i
= 2 (φ(x + l, y, t) − φ(x, y, t)) + (φ(x − l, y, t) − φ(x, y, t))
l h
α i
+ 2 (φ(x, y + l, t) − φ(x, y, t)) + (φ(x, y − l, t) − φ(x, y, t))
l
 8B. HARMONIC FUNCTIONS 183

Now observe that we can write this equation as follows:

φ(x, y, t + δ) − φ(x, y, t) φ(x + l, y, t) − 2φ(x, y, t) + φ(x − l, y, t)
= α·
δ l2
φ(x, y + l, t) − 2φ(x, y, t) + φ(x, y − l, t)
+ α·
l2
As it was the case before when modeling the wave equation, we recognize on the right
the usual approximation of the second derivative, coming from calculus. Thus, when
taking the continuous limit of our model, l → 0, we obtain the following equation:
 2
d φ d2 φ

φ(x, y, t + δ) − φ(x, y, t)
=α + 2 (x, y, t)
δ dx2 dy
Now with t → 0, we are led in this way to the heat equation, namely:
φ̇(x, y, t) = α · ∆φ(x, y, t)
(3) Finally, for completness, let us mention that in arbitrary N dimensions the same
argument, namely a basic lattice model, carries over, and we obtain a similar equation,
namely φ̇ = α∆φ, with the following straightforward definition for the Laplacian, based
on the obvious N -dimensional analogues of Definition 8.2 and Definition 8.3:
N
X d2 φ
∆φ =
i=1
dx2i
Which is quite nice, unfiying what we did in chapter 3 in one dimension, and in the
above in 2 dimensions. We will be back to this, later in this book. □

8b. Harmonic functions

To summarize, we have been doing some physics, with the conclusion that both the
wave and heat equation involve the Laplace operator. So, let us formulate, as a conclusion,
the definition of this operator, that we would like to further understand:
Definition 8.7. The Laplace operator in 2 dimensions is:
d2 f d2 f
∆f = +
dx2 dy 2
A function f : R2 → C satisfying ∆f = 0 will be called harmonic.
Here the formula of ∆ is the one coming from Theorem 8.5 and Theorem 8.6. As for
the notion of harmonic function, this is something quite natural. Indeed, we can think
of ∆ as being a linear operator on the space of functions f : R2 → C, and previous
experience with linear operators, and linear algebra in general, suggests looking first into
184 8. HARMONIC FUNCTIONS

the eigenvectors of ∆. But the simplest such eigenvectors are those corresponding to the
eigenvalue λ = 0, and these are exactly our harmonic functions, satisfying:
∆f = 0
Getting now to more concrete things, and to some mathematics that we can do, using
our knowledge, let us try to find the functions f : R2 → C which are harmonic. And here,
as a good surprise, we have an interesting link with the holomorphic functions:
Theorem 8.8. Any holomorphic function f : C → C, when regarded as real function
f : R2 → C
is harmonic. Moreover, the conjugates f¯ of holomorphic functions are harmonic too.
Proof. The first assertion follows from the following computation, for the power
functions f (z) = z n , with the usual notation z = x + iy:
d2 z n d2 z n
∆z n = +
dx2 dy 2
d(nz n−1 ) d(inz n−1 )
= +
dx dy
n−2
= n(n − 1)z − n(n − 1)z n−2
= 0
As for the second assertion, this follows from ∆f¯ = ∆f , which is clear from definitions,
and which shows that if f is harmonic, then so is its conjugate f¯. □

All this is quite interesting, and the idea in what follows will be that of developing
a theory of harmonic functions, as a generalization of the theory that we know for the
holomorphic functions, covering as well functions of type z̄. Them, we will go back to
physics, with some applications of this to the wave equation, and the heat equation.

As a first goal, in order to understand the harmonic functions, we can try to find the
homogeneous polynomials P ∈ R[x, y] which are harmonic. In order to do so, the most
convenient is to use the variable z = x + iy, and think of these polynomials as being
homogeneous polynomials P ∈ R[z, z̄]. With this convention, the result is as follows:
Theorem 8.9. The degree n homogeneous polynomials P ∈ R[x, y] which are har-
monic are precisely the linear combinations of
P = zn , P = z̄ n
with the usual identification z = x + iy.
 8B. HARMONIC FUNCTIONS 185

Proof. As explained above, any homogeneous polynomial P ∈ R[x, y] can be re-

garded as an homogeneous polynomial P ∈ R[z, z̄], with the change of variables z = x+iy,
and in this picture, the degree n homogeneous polynomials are as follows:
X
P (z) = ckl z k z̄ l
k+l=n

In oder to solve now the Laplace equation ∆P = 0, we must compute the quantities
∆(z k z̄ l ), for any k, l. But the computation here is routine. We first have the following
formula, with the derivatives being computed with respect to the variable x:

d(z k z̄ l )
= (z k )′ z̄ l + z k (z̄ l )′
dx
= kz k−1 z̄ l + lz k z̄ l−1

By taking one more time the derivative with respect to x, we obtain:

d2 (z k z̄ l )
= k(z k−1 z̄ l )′ + l(z k z̄ l−1 )′
dx2
= k (z k−1 )′ z̄ l + z k−1 (z̄ l )′ + l (z k )′ z̄ l−1 + z k (z̄ l−1 )′
   

= k (k − 1)z k−2 z̄ l + lz k−1 z̄ l−1 + l kz k−1 z̄ l−1 + (l − 1)z k z̄ l−2

   

= k(k − 1)z k−2 z̄ l + 2klz k−1 z̄ l−1 + l(l − 1)z k z̄ l−2

With respect to the variable y, the computations are similar, but some ±i factors
appear, due to z ′ = i and z̄ ′ = −i, coming from z = x + iy. We first have:

d(z k z̄ l )
= (z k )′ z̄ l + z k (z̄ l )′
dy
= ikz k−1 z̄ l − ilz k z̄ l−1

By taking one more time the derivative with respect to y, we obtain:

d2 (z k z̄ l )
= ik(z k−1 z̄ l )′ − il(z k z̄ l−1 )′
dy 2
= ik (z k−1 )′ z̄ l + z k−1 (z̄ l )′ − il (z k )′ z̄ l−1 + z k (z̄ l−1 )′
   

= ik i(k − 1)z k−2 z̄ l − ilz k−1 z̄ l−1 − il ikz k−1 z̄ l−1 − i(l − 1)z k z̄ l−2
   

= −k(k − 1)z k−2 z̄ l + 2klz k−1 z̄ l−1 − l(l − 1)z k z̄ l−2

186 8. HARMONIC FUNCTIONS

We can now sum the formulae that we found, and we obtain:

d2 (z k z̄ l ) d2 (z k z̄ l )
∆(z k z̄ l ) = +
dx2 dy 2
= k(k − 1)z k−2 z̄ l + 2klz k−1 z̄ l−1 + l(l − 1)z k z̄ l−2
−k(k − 1)z k−2 z̄ l + 2klz k−1 z̄ l−1 − l(l − 1)z k z̄ l−2
= 4klz k−1 z̄ l−1
In other words, we have reached to the following formula:
4klf
f = z k z̄ l =⇒ ∆f =
|z|2
Now let us get back to our homogeneous polynomial P , written as follows:
X
P (z) = ckl z k z̄ l
k+l=n

By using the above formula, the Laplacian of P is given by:

4 X
∆P (z) = 2 klckl z k z̄ l
|z| k+l=n

We conclude that the Laplace equation for P takes the following form:
∆P = 0 ⇐⇒ klckl = 0, ∀k, l
⇐⇒ [k, l ̸= 0 =⇒ ckl = 0]
⇐⇒ P = cn0 z n + c0n z̄ n
Thus, we are led to the conclusion in the statement. And with the observation that
the real formulation of the final result is something quite complicated, and so, for one
more time, the use of the complex variable z = x + iy is something very useful. □
We know that the holomorphic functions are harmonic, and it follows from this that the
real and imaginary parts of the holomorphic functions, as well as any linear combinations
of these real and imaginary parts, are harmonic too. That is, if f is holomorphic, then
the following function is harmonic, for any values of the parameters α, β ∈ C:
fαβ = αRe(f ) + βIm(f )
Observe that this result covers all the examples that we have so far, for instance with
the function z̄, that we know to be harmonic, appearing as follows:
z̄ = Re(z) − iIm(z)
Our main goal in what follows will be that of proving a converse to this, at least
locally. For this purpose, let us start with the following definition:
 8B. HARMONIC FUNCTIONS 187

Definition 8.10. The Cauchy-Riemann operators are

   
1 d d ¯ 1 d d
∂= −i , ∂= +i
2 dx dy 2 dx dy
d d
where dx
and dy
are the usual partial derivatives for complex functions.

There are many things that can be said about the Cauchy-Riemann operators ∂, ∂, ¯
the idea being that in many contexts, these are better to use than the usual partial
d d
derivatives dx , dy , and with this being a bit like the usage of the variables z, z̄, instead of
the decomposition z = a + ib, for many questions regarding the complex numbers.

We have already seen in fact some instances of this, in our computations above. At
the general level, the main properties of ∂, ∂¯ can be summarized as follows:
Proposition 8.11. Assume that f : X → C is differentiable in the real sense.
(1) f is holomorphic precisely when ∂f ¯ = 0.
′
(2) In this case, its derivative is f = ∂f .
(3) The Laplace operator is given by ∆ = 4∂ ∂. ¯
¯
(4) f is harmonic precisely when ∂ ∂f = 0.
Proof. We can assume by linearity that we are dealing with differentiability questions
at 0. Since our function f : X → C is differentiable in the real sense, we have a formula
as follows, with z = x + iy, and with a, b ∈ C being the partial derivatives at 0:
f (z) = ax + by + o(z)
Now observe that we can write this formula in the following way:
z + z̄ z − z̄
f (z) = a · +b· + o(z)
2 2i
z + z̄ iz̄ − iz
= a· +b· + o(z)
2 2
a − ib a + ib
= ·z+ · z̄ + o(z)
2 2
Now by dividing by z, we obtain from this the following formula:
f (z) a − ib a + ib z̄
= + · + o(1)
z 2 2 z
z̄
¯ (0) · + o(1)
= ∂f (0) + ∂f
z
But this gives the first two assertions, because in order for the derivative f ′ (0) to exist,
appearing as the z → 0 limit of the above quantity, the coefficient of z̄/z, which does not
188 8. HARMONIC FUNCTIONS

converge, must vanish. Regarding now the third assertion, this follows from:
d2 d2
∆ = +
dx2 dy 2
  
d d d d
= −i +i
dx dy dx dy
¯
= 4∂ ∂
As for the last assertion, this is clear from this latter formula of ∆. □
In analogy now with the theory of the holomorphic functions, we have:
Theorem 8.12. The harmonic functions obey to the same general principles as the
holomorphic functions, namely:
(1) The maximum modulus principle.
(2) The plain mean value formula.
(3) The boundary mean value formula.
(4) The Liouville theorem.
Also, locally, the real harmonic functions are the real parts of holomorphic functions.
Proof. This is something quite tricky, the idea being as follows:
(1) Regarding the maximum modulus principle, the statement here is that any har-
monic function f : X → C has the property that the maximum of |f | over a domain is
attained on its boundary. That is, given a domain D, with boundary γ, we have:
∃x ∈ γ , |f (x)| = max |f (y)|
y∈D

(2) Regarding the plain mean value formula, here the statement is that given an
harmonic function f : X → C, and a disk D, the following happens:
Z
f (x) = f (y)dy
D
(3) Regarding the boundary mean value formula, here the statement is that given an
harmonic function f : X → C, and a disk D, with boundary γ, the following happens:
Z
f (x) = f (y)dy
γ

(4) Regarding the Liouville theorem, the statement here is that an entire, bounded
harmonic function must be constant:
f :C→C , |f | ≤ M =⇒ f = constant
(5) Finally, regarding the proofs, these are not exactly trivial. According to our
previous experience with holomorphic functions, the above conditions (2,3) are equivalent,
and imply (1,4) via some simple arguments, and the same happens for the harmonic
 8B. HARMONIC FUNCTIONS 189

functions. However, establishing (2) and the last assertion is something which is not
exactly trivial, and we refer to Rudin [74] for proofs of this, and with the promise that
we will be back to this, later in this book, once we will know more calculus. □

Observe that we are in the process of a tactical retreat from mathematics. As a last
objective, however, before completely giving up with all this, and getting back to physics,
let us try to find the harmonic functions which are radial, in the following sense:

f (z) = φ(|z|)

However, things are quite tricky here, involving a blowup phenomenon at the dimen-
sion value N = 2, which is precisely the one that we are interested in. So, moving now to
N dimensions, with the straightforward definition for ∆ there, here is the result:

Theorem 8.13. The fundamental radial solutions of ∆f = 0 are

(
||x||2−N (N ̸= 2)
f (x) =
log ||x|| (N = 2)

with the log at N = 2 basically coming from log′ = 1/x.

Proof. Consider indeed a radial function, defined outside the origin x = 0. This
function can be written as follows, with φ : (0, ∞) → C being a certain function:

f : RN − {0} → C , f (x) = φ(||x||)

Our first goal will be that of reformulating the Laplace equation ∆f = 0 in terms of
the one-variable function φ : (0, ∞) → C. For this purpose, observe that we have:
qP
N 2
d||x|| d i=1 xi
=
dxi dxi
P 
N
1 1 d i=1 x2i
= · qP ·
2 N 2 dxi
i=1 xi
1 1
= · · 2xi
2 ||x||
xi
=
||x||
190 8. HARMONIC FUNCTIONS

By using this formula, we have the following computation:

df dφ(||x||)
=
dxi dxi
d||x||
= φ′ (||x||) ·
dxi
xi
= φ′ (||x||) ·
||x||
By differentiating one more time, we obtain the following formula:
d2 f
 
d ′ xi
= φ (||x||) ·
dx2i dxi ||x||
dφ′ (||x||) xi
 
′ d xi
= · + φ (||x||) ·
dxi ||x|| dxi ||x||
 
xi xi ||x|| − xi · xi /||x||
= φ′′ (||x||) · · + φ′ (||x||) ·
||x|| ||x|| ||x||2
x2 ||x||2 − x2i
= φ′′ (||x||) · i 2 + φ′ (||x||) ·
||x|| ||x||3
Now by summing over i ∈ {1, . . . , N }, this gives the following formula:
N N
X
′′ x2 X ||x||2 − x2i
∆f = φ (||x||) · i 2 + φ′ (||x||) ·
i=1
||x|| i=1
||x||3
||x||2 (N − 1)||x||2
= φ′′ (||x||) · + φ ′
(||x||) ·
||x||2 ||x||3
N −1
= φ′′ (||x||) + φ′ (||x||) ·
||x||
Thus, with r = ||x||, the Laplace equation ∆f = 0 can be reformulated as follows:
(N − 1)φ′ (r)
φ′′ (r) + =0
r
Equivalently, the equation that we want to solve is as follows:
rφ′′ + (N − 1)φ′ = 0
Now observe that we have the following formula:
(rN −1 φ′ )′ = (N − 1)rN −2 φ′ + rN −1 φ′′
= rN −2 ((N − 1)φ′ + rφ′′ )
Thus, the equation to be solved can be simply written as follows:
(rN −1 φ′ )′ = 0
 8C. LIGHT, SPECTROSCOPY 191

We conclude that rN −1 φ′ must be a constant K, and so, that we must have:

φ′ = Kr1−N
But the fundamental solutions of this latter equation are as follows:
(
r2−N (N ̸= 2)
φ(r) =
log r (N = 2)
Thus, we are led to the conclusion in the statement. □
And good news, that is the end. We have learned many things about the Laplace
operator and the harmonic functions, but all this clearly becomes too complicated, spilling
sometimes in arbitrary N dimensions, so time to stop. We will be back to this.

8c. Light, spectroscopy

Getting back now to physics, a natural question would be that of going back to the
equations that we started with, namely the wave and heat equations in 2D, and see if our
accumulated knowledge about harmonic functions, which is after all not that bad, can
help there. Unfortunately, this is not exactly the case, and we are still a long way to go,
from solving that equations. So, obviously, time to ask the cat. And cat says:
Cat 8.14. If you don’t understand physics, do more physics.
Thanks cat, and this seems wise indeed, let’s have some fun with more physics, still
related of course to the Laplace operator ∆, and for difficult mathematics, which is still
to be solved, we will see later, towards the end of the present book. Getting started now,
we will need for what we want to talk about something quite heavy, namely:
Theorem 8.15 (Maxwell theory). In regions of space where there is no charge or
current present the Maxwell equations for electrodynamics read
< ∇, E >=< ∇, B >= 0
∇ × E = −Ḃ , ∇ × B = Ė/c2
and both the electric field E and magnetic field B are subject to the wave equation
φ̈ = c2 ∆φ
d2 /dx2i is the Laplace operator, and c = 299, 792, 458.
P
where ∆ = i

Proof. This is something fundamental, appearing as a tricky mixture of physics facts

and mathematical results, the idea being as follows:
(1) To start with, electrodynamics is the science of moving electrical charges. And
this is something quite complicated, because unlike in classical mechanics, where the
Newton law is good for both the static and the dynamic setting, the Coulomb law, which
192 8. HARMONIC FUNCTIONS

is actually very similar to the Newton law, does the job when the charges are static, but
no longer describes well the situation when the charges are moving.
(2) The problem comes from the fact that moving charges produce magnetism, and
with this being visible when putting together two electric wires, which will attract or
repel, depending on orientation. Thus, in contrast with classical mechanics, where static
and dynamic problems are described by a unique field, the gravitational one, in electro-
dynamics we have two fields, namely the electric field E, and the magnetic field B.
(3) Fortunately, there is a full set of equations relating the electric field E and the
magnetic field B. These are the Maxwell equations, which look as follows:
ρ
< ∇, E >= , < ∇, B >= 0
ε0

∇ × E = −Ḃ , ∇ × B = µ0 J + µ0 ε0 Ė
(4) To be more precise, regarding first the math, the dots denote derivatives with
respect to time, and ∇ is the gradient operator, or space derivative, given by:
d
dx
d 
∇ =  dy
d
dz

(5) Regarding the physics, the first formula is the Gauss law, ρ being the charge, and
ε0 being a constant, and with this Gauss law more or less replacing the Coulomb law
from electrostatics. The second formula is something basic, and anonymous. The third
formula is the Faraday law. As for the fourth formula, this is the Ampère law, as modified
by Maxwell, with J being the volume current density, and µ0 being a constant.
(6) Without bothering too much about the precise meaning of all this, we can see
right away that under the circumstances in the statement, namely in the regions of space
where there is no charge or current present, the Maxwell equations take a simple and
comprehensible form, readable even without much physics background, namely:
< ∇, E >=< ∇, B >= 0

∇ × E = −Ḃ , ∇ × B = µ0 ε0 Ė
(7) Thus, we have reached to the equations in the statement, modulo a discussion
about the constant µ0 ε0 . And the point here is that, according to a remarkable discovery
of Biot and Savart, the main electrodynamics constants µ0 , ε0 are magically related to the
observed speed of light in vacuum c = 299, 792, 458 by the following formula:
1
µ 0 ε0 =
c2
 8C. LIGHT, SPECTROSCOPY 193

(8) Summarizing, we have our equations. Leaving aside the first two equations, by
applying the curl operator to the last two equations, we obtain:
∇ × (∇ × E) = −∇ × Ḃ = −(∇ × B)′ = −Ë/c2
∇ × (∇ × B) = ∇ × Ė/c2 = (∇ × E)′ /c2 = −B̈/c2
But the double curl operator is subject to the following formula:
∇ × (∇ × φ) = ∇ < ∇, φ > −∆φ
Now by using the first two equations, we are led to the conclusion in the statement. □
So, question now, what is light? Light is the wave predicted by Theorem 8.15, travel-
ling at speed c, and with the important extra property that it depends on a real positive
parameter, that can be called, upon taste, frequency, wavelength, or color:
Fact 8.16. An accelerating or decelerating charge produces electromagnetic waves,
travelling in vacuum at speed c = 299, 792, 458,
φ̈ = c2 ∆φ
and in non-vacuum at a lower speed v < c. These waves are called light, whose frequency
and wavelength can be explicitly computed.
This phenomenon can be observed is a variety of situations, such as the usual light
bulbs, where electrons get decelerated by the filament, acting as a resistor, or in usual fire,
which is a chemical reaction, with the electrons moving around, as they do in any chemical
reaction, or in more complicated machinery like nuclear plants, particle accelerators, and
so on, leading there to all sorts of eerie glows, of various colors.

Let us try now to understand the simplest solutions of the wave equation φ̈ = c2 ∆φ,
or more generally φ̈ = v 2 ∆φ, from Fact 8.16. In 1D, the situation is as follows:
Theorem 8.17. The 1D wave equation, with speed v, namely
d2 φ
φ̈ = v 2
dx2
has as basic solutions the following functions,
φ(x) = A cos(kx − wt + δ)
with A being called amplitude, kx−wt+δ being called the phase, k being the wave number,
w being the angular frequency, and δ being the phase constant. We have
2π 2π 1
λ= , T = , ν= , w = 2πν
k kv T
relating the wavelength λ, period T , frequency ν, and angular frequency w. Moreover, any
solution of the wave equation appears as a linear combination of such basic solutions.
194 8. HARMONIC FUNCTIONS

Proof. There are several things going on here, the idea being as follows:
(1) Our first claim is that the function φ in the statement satisfies indeed the wave
equation, with speed v = w/k. For this purpose, observe that we have:
d2 φ
φ̈ = −w2 φ , = −k 2 φ
dx2
Thus, the wave equation is indeed satisfied, with speed v = w/k:
 w 2 d2 φ 2
2 d φ
φ̈ = = v
k dx2 dx2
(2) Regarding now the other things in the statement, all this is basically terminology,
which is very natural, when thinking how φ(x) = A cos(kx − wt + δ) propagates.
(3) Finally, the last assertion is standard, coming from Fourier analysis, and we will
leave this as an instructive exercise, based on the material from chapter 7. □

As a first observation, the above result invites the use of complex numbers. Indeed,
we can write the solutions that we found in a more convenient way, as follows:
φ(x) = Re A ei(kx−wt+δ)
 

And we can in fact do even better, by absorbing the quantity eiδ into the amplitude
A, which becomes now a complex number, and writing our formula as:

φ = Re(φ)
e , φ e i(kx−wt)
e = Ae
Moving ahead now towards electromagnetism and 3D, let us formulate:
Definition 8.18. A monochromatic plane wave is a solution of the 3D wave equation
which moves in only 1 direction, making it in practice a solution of the 1D wave equation,
and which is of the special from found in Theorem 8.17, with no frequencies mixed.
In other words, we are making here two assumptions on our wave. First is the 1-
dimensionality assumption, which gets us into the framework of Theorem 8.17. And
second is the assumption, in connection with the Fourier decomposition result from the
end of Theorem 8.17, that our solution is of “pure” type, meaning a wave having a well-
defined wavelenght and frequency, instead of being a “packet” of such pure waves.

Summarizing, we have now a decent intuition about what light is, and more on this
later. Let us discuss now the examples. The idea is that we have various types of light,
depending on frequency and wavelength. These are normally referred to as “electromag-
netic waves”, but for keeping things simple and luminous, we will keep using the familiar
 8C. LIGHT, SPECTROSCOPY 195

term “light”. The classification, in a rough form, is as follows:

Frequency Type Wavelength

−
1018 − 1020 γ rays 10−12 − 10−10
1016 − 1018 X − rays 10−10 − 10−8
1015 − 1016 UV 10−8 − 10−7
−
1014 − 1015 blue 10−7 − 10−6
1014 − 1015 yellow 10−7 − 10−6
1014 − 1015 red 10−7 − 10−6
−
1011 − 1014 IR 10−6 − 10−3
109 − 1011 microwave 10−3 − 10−1
1 − 109 radio 10−1 − 108

Observe the tiny space occupied by the visible light, all colors there, and the many
more missing, being squeezed under the 1014 − 1015 frequency banner. Here is a zoom on
that part, with of course the comment that all this, colors, is something subjective:

Frequency THz = 1012 Hz Color Wavelength nm = 10−9 m

−
670 − 790 violet 380 − 450
620 − 670 blue 450 − 485
600 − 620 cyan 485 − 500
530 − 600 green 500 − 565
510 − 530 yellow 565 − 590
480 − 510 orange 590 − 625
400 − 480 red 625 − 750

With this in hand, we can now do some basic optics. Light usually comes in “bundles”,
with waves of several wavelenghts coming at the same time, from the same source, and
the first challenge is that of separating these wavelenghts. In order to discuss this, from
a practical perspective, let us start with the following fact:

Fact 8.19. When travelling through a material, and hitting a new material, some of
the light gets reflected, at the same angle, and some of it gets refracted, at a different
angle, depending both on the old and the new material, and on the wavelength.

Again, this is something deep, and there are many things that can be said here, ranging
from various computations based on the Maxwell equations, to all sorts of considerations
belonging to advanced materials theory. As a basic formula, we have the famous Snell
196 8. HARMONIC FUNCTIONS

law, which relates the incidence angle θ1 to the refraction angle θ2 , as follows:
sin θ2 n1 (λ)
=
sin θ1 n2 (λ)
Here ni (λ) are the refraction indices of the two materials, adjusted for the wavelength,
and with this adjustment for wavelength being the whole point, which is something quite
complicated. Now as a simple consequence of the above, we have:
Theorem 8.20. Light can be decomposed, by using a prism.
Proof. This follows from Fact 8.19. Indeed, when hitting a piece of glass, provided
that the hitting angle is not 90◦ , the light will decompose over the wavelenghts present,
with the corresponding refraction angles depending on these wavelengths. And we can
capture these split components at the exit from the piece of glass, again deviated a bit,
provided that the exit surface is not parallel to the entry surface. And the simplest device
doing the job, that is, having two non-parallel faces, is a prism. □
As an application of this, we can study events via spectroscopy, by capturing the light
the event has produced, decomposing it with a prism, carefully recording its “spectral
signature”, consisting of the wavelenghts present, and their density, and then doing some
reverse engineering, consisting in reconstructing the event out of its spectral signature.

8d. Atomic spectrum

Getting now to some truly exciting applications of light and spectroscopy, let us dis-
cuss the beginnings of the atomic theory. There is a long story here, involving many
discoveries, around 1890-1900, focusing on hydrogen H. We will present here things a bit
retrospectively. First on our list is the following discovery, by Lyman in 1906:
Fact 8.21 (Lyman). The hydrogen atom has spectral lines given by the formula
 
1 1
=R 1− 2
λ n
where R ≃ 1.097 × 107 and n ≥ 2, which are as follows,
n Name Wavelength Color
− −
2 α 121.567 UV
3 β 102.572 UV
4 γ 97.254 UV
.. .. .. ..
. . . .
∞ limit 91.175 UV
called Lyman series of the hydrogen atom.
 8D. ATOMIC SPECTRUM 197

Observe that all the Lyman series lies in UV, which is invisible to the naked eye. Due
to this fact, this series, while theoretically being the most important, was discovered only
second. The first discovery, which was the big one, and the breakthrough, was by Balmer,
the founding father of all this, back in 1885, in the visible range, as follows:
Fact 8.22 (Balmer). The hydrogen atom has spectral lines given by the formula
 
1 1 1
=R −
λ 4 n2
where R ≃ 1.097 × 107 and n ≥ 3, which are as follows,
n Name Wavelength Color
− −
3 α 656.279 red
4 β 486.135 aqua
5 γ 434.047 blue
6 δ 410.173 violet
7 ε 397.007 UV
.. .. .. ..
. . . .
∞ limit 346.600 UV
called Balmer series of the hydrogen atom.
So, this was Balmer’s original result, which started everything. As a third main result
now, this time in IR, due to Paschen in 1908, we have:
Fact 8.23 (Paschen). The hydrogen atom has spectral lines given by the formula
 
1 1 1
=R −
λ 9 n2
where R ≃ 1.097 × 107 and n ≥ 4, which are as follows,
n Name Wavelength Color
− −
4 α 1875 IR
5 β 1282 IR
6 γ 1094 IR
.. .. .. ..
. . . .
∞ limit 820.4 IR
called Paschen series of the hydrogen atom.
Observe the striking similarity between the above three results. In fact, we have here
the following fundamental, grand result, due to Rydberg in 1888, based on the Balmer
series, and with later contributions by Ritz in 1908, using the Lyman series as well:
198 8. HARMONIC FUNCTIONS

Conclusion 8.24 (Rydberg, Ritz). The spectral lines of the hydrogen atom are given
by the Rydberg formula, depending on integer parameters n1 < n2 ,
 
1 1 1
=R −
λn1 n2 n21 n22
with R being the Rydberg constant for hydrogen, which is as follows:
R ≃ 1.096 775 83 × 107
These spectral lines combine according to the Ritz-Rydberg principle, as follows:
1 1 1
+ =
λn1 n2 λn2 n3 λn1 n3
Similar formulae hold for other atoms, with suitable fine-tunings of R.
Here the first part, the Rydberg formula, generalizes the results of Lyman, Balmer,
Paschen, which appear at n1 = 1, 2, 3, at least retrospectively. The Rydberg formula
predicts further spectral lines, appearing at n1 = 4, 5, 6, . . . , and these were discovered
later, by Brackett in 1922, Pfund in 1924, Humphreys in 1953, and others afterwards,
with all these extra lines being in far IR. The simplified complete table is as follows:
n1 n2 Series name Wavelength n2 = ∞ Color n2 = ∞
− −
1 2−∞ Lyman 91.13 nm UV
2 3−∞ Balmer 364.51 nm UV
3 4−∞ Paschen 820.14 nm IR
− −
4 5−∞ Brackett 1458.03 nm far IR
5 6−∞ Pfund 2278.17 nm far IR
6 7−∞ Humphreys 3280.56 nm far IR
.. .. .. .. ..
. . . . .
Regarding the last assertion, concerning other elements, this was something conjec-
tured and partly verified by Ritz, and fully verified and clarified later, via many experi-
ments, the fine-tuning of R being basically R → RZ 2 , where Z is the atomic number.

From a theoretical physics viewpoint, the main result remains the middle assertion,
called Ritz-Rydberg combination principle. This is something at the same time extremely
simple, and completely puzzling, the informal conclusion being as follows:
Thought 8.25. The simplest observables of the hydrogen atom, combining via
1 1 1
+ =
λn1 n2 λn2 n3 λn1 n3
look like quite weird quantities. Why wouldn’t they just sum normally.
 8D. ATOMIC SPECTRUM 199

Fortunately, mathematics comes to the rescue. Indeed, the Ritz-Rydberg combination

principle reminds the formula en1 n2 en2 n3 = en1 n3 for the usual matrix units eij : ej → ei .
In short, we are in familiar territory here, and we can start dreaming of:
Thought 8.26. Observables in quantum mechanics should be some sort of infinite
matrices, generalizing the Lyman, Balmer, Paschen lines of the hydrogen atom, and mul-
tiplying between them as the matrices do, as to produce further observables.
Good news, time to put everything together. As a main problem that we would like
to solve, we have the understanding the intimate structure of matter, at the atomic level.
There is of course a long story here, regarding the intimate structure of matter, going
back centuries and even millennia ago, and our presentation here will be quite simplified.
As a starting point, since we need a starting point, let us agree on:
Claim 8.27. Ordinary matter is made of small particles called atoms, with each atom
appearing as a mix of even smaller particles, namely protons +, neutrons 0 and electrons
−, with the same number of protons + and electrons −.
As a first observation, this is something which does not look obvious at all, with
probably lots of work, by many people, being involved, as to lead to this claim. And so it
is. The story goes back to the discovery of charges and electricity, which were attributed
to a small particle, the electron −. Now since matter is by default neutral, this naturally
leads to the consideration to the proton +, having the same charge as the electron.

But, as a natural question, why should be these electrons − and protons + that small?
And also, what about the neutron 0? These are not easy questions, and the fact that it
is so came from several clever experiments. Let us first recall that careful experiments
with tiny particles are practically impossible. However, all sorts of brutal experiments,
such as bombarding matter with other pieces of matter, accelerated to the extremes,
or submitting it to huge electric and magnetic fields, do work. And it is such kind of
experiments, due to Thomson, Rutherford and others, “peeling off” protons +, neutrons
0 and electrons − from matter, and observing them, that led to the conclusion that these
small beasts +, 0, − exist indeed, in agreement with Claim 8.27.

Of particular importance here was as well the radioactivity theory of Becquerel and
Pierre and Marie Curie, involving this time such small beasts, or perhaps some related
radiation, peeling off by themselves, in heavy elements such as uranium 92 U, polonium
84 Po and radium 88 Ra. And there was also Einstein’s work on the photoelectric effect,
light interacting with matter, suggesting that even light itself might have associated to
it some kind of particle, called photon. All this goes of course beyond Claim 8.27, with
further particles involved, and more on this later, but as a general idea, all this deluge of
small particle findings, all coming around 1900-1910, further solidified Claim 8.27.
200 8. HARMONIC FUNCTIONS

So, taking now Claim 8.27 for granted, how are then the atoms organized, as mixtures
of protons +, neutrons 0 and electrons −? The answer here lies again in the above-
mentioned “brutal” experiments of Thomson, Rutherford and others, which not only
proved Claim 8.27, but led to an improved version of it, as follows:
Claim 8.28. The atoms are formed by a core of protons + and neutrons 0, surrounded
by a cloud of electrons −, gravitating around the core.
This is a considerable advance, because we are now into familiar territory, namely
some kind of mechanics. And with this in mind, all the pieces of our puzzle start fitting
together, and we are led to the following grand conclusion:
Claim 8.29 (Bohr and others). The atoms are formed by a core of protons and neu-
trons, surrounded by a cloud of electrons, basically obeying to a modified version of elec-
tromagnetism. And with a fine mechanism involved, as follows:
(1) The electrons are free to move only on certain specified elliptic orbits, labeled
1, 2, 3, . . . , situated at certain specific heights.
(2) The electrons can jump or fall between orbits n1 < n2 , absorbing or emitting light
and heat, that is, electromagnetic waves, as accelerating charges.
(3) The energy of such a wave, coming from n1 → n2 or n2 → n1 , is given, via the
Planck viewpoint, by the Rydberg formula, applied with n1 < n2 .
(4) The simplest such jumps are those observed by Lyman, Balmer, Paschen. And
multiple jumps explain the Ritz-Rydberg formula.
And isn’t this beautiful. Moreover, some further claims, also by Bohr and others,
are that the theory can be further extended and fine-tuned as to explain many other
phenomena, such as the above-mentioned findings of Einstein, and of Becquerel and Pierre
and Marie Curie, and generally speaking, all the physics and chemistry known.
And the story is not over here. Following now Heisenberg, the next claim is that
the underlying mathematics in all the above can lead to a beautiful axiomatization of
quantum mechanics, as a “matrix mechanics”, along the lines of Thought 8.26. We will
be back to all this in chapter 16 below, at the end of the present book.
8e. Exercises
We had a tough physics chapter here, and here are some physics exercises:
Exercise 8.30. Find proofs for the main properties of the harmonic functions.
Exercise 8.31. Learn about soap films, and their relation with harmonic functions.
Exercise 8.32. Explore a bit the wave and heat equations, in 1 and 2 dimensions.
Exercise 8.33. Work out the details for the decomposition of wave packets.
As bonus exercise, try developing quantum mechanics, based on the above. With this
being not a joke, Heisenberg did not know much more than that, when he did it.
 Part III

Several variables
 This is my church
This is where I heal my hurts
For tonight
God is a DJ
 CHAPTER 9

Linear maps

9a. Linear maps

In the remainder of this book we study the functions of several variables. Things here
are quite tricky, and we will do this slowly, over a whole 200 pages. Let us start with the
linear functions, which are the simplest. We have here the following result:
Theorem 9.1. The linear maps f : RN → RM are in correspondence with the matrices
A ∈ MM ×N (R), with the linear map associated to such a matrix being
f (x) = Ax
and with the matrix associated to a linear map being Aij =< f (ej ), ei >. Similarly, the
linear maps f : CN → CN are in correspondence with the matrices A ∈ MM ×N (C).
Proof. The first assertion is clear, because a linear map f : RN → RM must send a
vector x ∈ RN to a certain vector f (x) ∈ RM , all whose components are linear combina-
tions of the components of x. Thus, we can write, for certain numbers Aij ∈ R:
x1 A11 x1 + . . . + A1N xN
   

f  ...  =  ..
. 
xN AM 1 x1 + . . . + AM N xN
Now the parameters Aij ∈ R can be regarded as being the entries of a rectangular
matrix A ∈ MM ×N (R), and with the usual convention for the rectangular matrix multi-
plication, the above formula is precisely the one in the statement, namely:
f (x) = Ax
Regarding the second assertion, with f (x) = Ax as above, if we denote by e1 , . . . , eN
the standard basis of RN , then we have the following formula:
A1j
 

f (ej ) =  ... 
AM j
But this gives < f (ej ), ei >= Aij , as desired. As for the last assertion, regarding
the complex maps and matrices, the proof here is similar, and with the complex
P scalar
products being by definition given, here and in what follows, by < x, y >= i xi ȳi . □
203
204 9. LINEAR MAPS

At the level of examples, let us first discuss the linear maps f : R2 → R2 . We have:
Proposition 9.2. The rotation of angle t ∈ R, and the symmetry with respect to the
Ox axis rotated by an angle t/2 ∈ R, are given by the matrices
   
cos t − sin t cos t sin t
Rt = , St =
sin t cos t sin t − cos t
both depending on t ∈ R taken modulo 2π.
Proof. The rotation being linear, it must correspond to a certain matrix:
 
a b
Rt =
c d
1 0



We can guess this matrix, via its action on the basic coordinate vectors 0
and 1
.
Indeed, a quick picture in the plane shows that we must have:
         
a b 1 cos t a b 0 − sin t
= , =
c d 0 sin t c d 1 cos t
Guessing now the matrix is not complicated, because the first equality gives us the
first column, and the second equality gives us the second column:
       
a cos t b − sin t
= , =
c sin t d cos t
Thus, we can just put together these two vectors, and we obtain
our matrix

Rt . As
1 0
for the symmetry, the proof here is similar, again by computing St 0 and St 1 . □
Let us record as well a result regarding the projections, as follows:
Proposition 9.3. The projection on the Ox axis rotated by an angle t/2 ∈ R is
 
1 1 + cos t sin t
Pt =
2 sin t 1 − cos t
depending on t ∈ R taken modulo 2π.
Proof. A quick picture in the plane, using similarity of triangles, and the basic
trigonometry formulae for the duplication of angles, show that we must have:
t cos 2t
     
1 1 1 + cos t
Pt = cos =
0 2 sin 2t 2 sin t
Similarly, another quick picture plus trigonometry show that we must have:
t cos 2t
     
0 1 sin t
Pt = sin =
1 2 sin 2t 2 1 − cos t
Now by putting together these two vectors, and we obtain our matrix. □
 9A. LINEAR MAPS 205

Back to Theorem 9.1, our claim is that, no matter what we want to do with f or A,
we will run at some point into their adjoints f ∗ and A∗ , constructed as follows:
Theorem 9.4. The adjoint linear map f ∗ : CN → CN , which is given by
< f (x), y >=< x, f ∗ (y) >
corresponds to the adjoint matrix A∗ ∈ MN (C), given by
(A∗ )ij = Āji
via the correspondence between linear maps and matrices constructed above.
Proof. Given a linear map f : CN → CN , fix y ∈ CN , and consider the linear form
φ(x) =< f (x), y >. This form must be as follows, for a certain vector f ∗ (y) ∈ CN :
φ(x) =< x, f ∗ (y) >
Thus, we have constructed a map y → f ∗ (y) as in the statement, which is obviously
linear, and that we can call f ∗ . Now by taking the vectors x, y ∈ CN to be elements of
the standard basis of CN , our defining formula for f ∗ reads:
< f (ei ), ej >=< ei , f ∗ (ej ) >
By reversing the scalar product on the right, this formula can be written as:
< f ∗ (ej ), ei >= < f (ei ), ej >
But this means that the matrix of f ∗ is given by (A∗ )ij = Āji , as desired. □
Getting back to our claim, the adjoints ∗ are indeed ubiquitous, as shown by:
Theorem 9.5. The following happen:
(1) f (x) = U x with U ∈ MN (C) is an isometry precisely when U ∗ = U −1 .
(2) f (x) = P x with P ∈ MN (C) is a projection precisely when P 2 = P ∗ = P .
Proof. Let us first recall that the lengths, or norms, of the vectors x ∈ CN can be
recovered from the knowledge of the scalar products, as follows:
√
||x|| = < x, x >
Conversely, we can recover the scalar products out of norms, by using the following
difficult to remember formula, called complex polarization identity:
||x + y||2 − ||x − y||2 + i||x + iy||2 − i||x − iy||2
= ||x||2 + ||y||2 − ||x||2 − ||y||2 + i||x||2 + i||y||2 − i||x||2 − i||y||2
+2Re(< x, y >) + 2Re(< x, y >) + 2iIm(< x, y >) + 2iIm(< x, y >)
= 4 < x, y >
206 9. LINEAR MAPS

Finally, we will use Theorem 9.4, and more specifically the following formula coming
from there, valid for any matrix A ∈ MN (C) and any two vectors x, y ∈ CN :
< Ax, y >=< x, A∗ y >
(1) Given a matrix U ∈ MN (C), we have indeed the following equivalences, with the
first one coming from the polarization identity, and with the other ones being clear:
||U x|| = ||x|| ⇐⇒ < U x, U y >=< x, y >
⇐⇒ < x, U ∗ U y >=< x, y >
⇐⇒ U ∗U y = y
⇐⇒ U ∗U = 1
⇐⇒ U ∗ = U −1
(2) Given a matrix P ∈ MN (C), in order for x → P x to be an oblique projection, we
must have P 2 = P . Now observe that this projection is orthogonal when:
< P x − x, P y >= 0 ⇐⇒ < P ∗ P x − P ∗ x, y >= 0
⇐⇒ P ∗P x − P ∗x = 0
⇐⇒ P ∗P − P ∗ = 0
⇐⇒ P ∗P = P ∗
The point now is that by conjugating the last formula, we obtain P ∗ P = P . Thus we
must have P = P ∗ , and this gives the result. □
Summarizing, the linear operators come in pairs T, T ∗ , and the associated matrices
come as well in pairs A, A∗ . We will keep this in mind, and come back to it later.
9b. Matrix inversion
We have seen so far that most of the interesting maps f : RN → RN that we know,
such as the rotations, symmetries and projections, are linear, and can be written in the
following form, with A ∈ MN (R) being a square matrix:
f (v) = Av
We develop now more general theory for such linear maps. We will be interested in
the question of inverting the linear maps f : RN → RN . And the point is that this is the
same question as inverting the corresponding matrices A ∈ MN (R), due to:
Theorem 9.6. A linear map f : RN → RN , written as
f (v) = Av
is invertible precisely when A is invertible, and in this case we have f −1 (v) = A−1 v.
Proof. This comes indeed from the fact that, with the notation fA (v) = Av, we have
the formula fA fB = fAB . Thus, we are led to the conclusion in the statement. □
 9B. MATRIX INVERSION 207

In order to study invertibility questions, for matrices or linear maps, let us begin with
some examples. In the simplest case, in 2 dimensions, the result is as follows:
Theorem 9.7. We have the following inversion formula, for the 2 × 2 matrices:
 −1  
a b 1 d −b
=
c d ad − bc −c a
When ad − bc = 0, the matrix is not invertible.
Proof. We have two assertions to be proved, the idea being as follows:
(1) As a first observation, when ad − bc = 0 we must have, for some λ ∈ R:
b = λa , d = λc
Thus our matrix must be of the following special type:
   
a b a λa
=
c d a λc
But in this case the columns are proportional, so the linear map associated to the
matrix is not invertible, and so the matrix itself is not invertible either.
(2) When ad − bc ̸= 0, let us look for an inversion formula of the following type:
 −1  
a b 1 ∗ ∗
=
c d ad − bc ∗ ∗
We must therefore solve the following system of equations:
    
a b ∗ ∗ ad − bc 0
=
c d ∗ ∗ 0 ad − bc
But the solution to these equations is obvious, is as follows:
    
a b d −b ad − bc 0
=
c d −c a 0 ad − bc
Thus, we are led to the formula in the statement. □
In order to deal now with the inversion problem in general, for the arbitrary matrices
A ∈ MN (R), we will use the same method as the one above, at N = 2. Let us write
indeed our matrix as follows, with v1 , . . . , vN ∈ RN being its column vectors:
A = [v1 , . . . , vN ]
We know from the above that, in order for the matrix A to be invertible, the vectors
v1 , . . . , vN must be linearly independent. Thus, we are led into the question of under-
standing when a family of vectors v1 , . . . , vN ∈ RN are linearly independent. In order to
deal with this latter question, let us introduce the following notion:
208 9. LINEAR MAPS

Definition 9.8. Associated to any vectors v1 , . . . , vN ∈ RN is the volume

det+ (v1 . . . vN ) = vol < v1 , . . . , vN >
of the parallelepiped made by these vectors.
Here the volume is taken in the standard N -dimensional sense. At N = 1 this volume
is a length, at N = 2 this volume is an area, at N = 3 this is the usual 3D volume, and so
on. In general, the volume of a body X ⊂ RN is by definition the number vol(X) ∈ [0, ∞]
of copies of the unit cube C ⊂ RN which are needed for filling X. Now with this notion
in hand, in relation with our inversion problem, we have the following statement:
Proposition 9.9. The quantity det+ that we constructed, regarded as a function of
the corresponding square matrices, formed by column vectors,
det+ : MN (R) → R+
has the property that a matrix A ∈ MN (R) is invertible precisely when det+ (A) > 0.
Proof. This follows from the fact that a matrix A ∈ MN (R) is invertible precisely
when its column vectors v1 , . . . , vN ∈ RN are linearly independent. But this latter condi-
tion is equivalent to the fact that we must have the following strict inequality:
vol < v1 , . . . , vN >> 0
Thus, we are led to the conclusion in the statement. □
Summarizing, all this leads us into the explicit computation of det+ . As a first obser-
vation, in 1 dimension we obtain the absolute value of the real numbers:
det+ (a) = |a|
In 2 dimensions now, the computation is non-trivial, and we have the following result,
making the link with our main inversion result so far, namely Theorem 9.7:
Theorem 9.10. In 2 dimensions we have the following formula,
 
+ a b
det = |ad − bc|
c d
with det+ : M2 (R) → R+ being the function constructed above.
Proof. We must show that the area of the parallelogram formed by ac , db equals



|ad − bc|. We can assume a, b, c, d > 0 for simplifying,

b
the proof in general being similar.
a
Moreover, by switching if needed the vectors c , d , we can assume that we have:
a b
>
c d
 9B. MATRIX INVERSION 209

According to these conventions, the picture of our parallelogram is as follows:

O
c+d •

d •
c •
• /

b a a+b
Now let us slide the upper side downwards left, until we reach the Oy axis. Our
parallelogram, which has not changed its area in this process, becomes:

O
c+d ◦
c+x •
d ◦
x •
c •
• /
b a a+b
We can further modify this parallelogram, once again by not altering its area, by
sliding the right side downwards, until we reach the Ox axis:

O
c+x ◦
x • •
c ◦
• • /

b a a+b
Let us compute now the area. Since our two sliding operations have not changed the
area of the original parallelogram, this area is given by:

A = ax
210 9. LINEAR MAPS

In order to compute the quantity x, observe that in the context of the first move, we
have two similar triangles, according to the following picture:
O
c+d •

d ◦ • ◦
x •
/
b a a+b
Thus, we are led to the following equation for the number x:
d−x c
=
b a
By solving this equation, we obtain the following value for x:
bc
x=d−
a
Thus the area of our parallelogram, or rather of the final rectangle obtained from it,
which has the same area as the original parallelogram, is given by:
ax = ad − bc
Thus, we are led to the conclusion in the statement. □
All this is very nice, and obviously we have a beginning of theory here. However, when
looking carefully, we can see that our theory has a weakness, because:
(1) In 1 dimension the number a, which is the simplest function of a itself, is certainly
a better quantity than the number |a|.
(2) In 2 dimensions the number ad − bc, which is linear in a, b, c, d, is certainly a
better quantity than the number |ad − bc|.
So, let us upgrade now our theory, by constructing a better function, which takes
signed values. In order to do this, we must come up with a way of splitting the systems
of vectors v1 , . . . , vN ∈ RN into two classes, call them positive and negative. And here,
the answer is quite clear, because a bit of thinking leads to the following definition:
Definition 9.11. A system of vectors v1 , . . . , vN ∈ RN is called:
(1) Oriented, if one can continuously pass from the standard basis to it.
(2) Unoriented, otherwise.
The associated sign is + in the oriented case, and − in the unoriented case.
 9B. MATRIX INVERSION 211

As a first example, in 1 dimension the basis consists of the single vector e = 1, which
can be continuously deformed into any vector a > 0. Thus, the sign is the usual one:
(
+ if a > 0
sgn(a) =
− if a < 0
Thus, in connection with our original question, we are definitely on the good track,
because when multiplying |a| by this sign we obtain a itself, as desired:
a = sgn(a)|a|
In 2 dimensions now, the explicit formula of the sign is as follows:
Proposition 9.12. We have the following formula, valid for any 2 vectors in R2 ,
   
a b
sgn , = sgn(ad − bc)
c d
with the sign function on the right being the usual one, in 1 dimension.
Proof. According to our conventions, the sign of ac , db is as follows:

(1) The sign is + when these vectors come in this precise order with respect to the
counterclockwise rotation in the plane, around 0.
(2) The sign is − otherwise, meaning when these vectors come in this order with
respect to the clockwise rotation in the plane, around 0.
If we assume now a, b, c, d > 0 for simplifying, we are left with comparing the angles
having the numbers c/a and d/b as tangents, and we obtain in this way:
    (
a b + if ac < db
sgn , =
c d − if ac > db
But this gives the formula in the statement. The proof in general is similar. □
Once again, in connection with our original question, we are on the good track, because
when multiplying |ad − bc| by this sign we obtain ad − bc itself, as desired:
ad − bc = sgn(ad − bc)|ad − bc|
At the level of the general results now, we have:
Proposition 9.13. The orientation of a system of vectors changes as follows:
(1) If we switch the sign of a vector, the associated sign switches.
(2) If we permute two vectors, the associated sign switches as well.
Proof. Both these assertions are clear from the definition of the sign, because the
two operations in question change the orientation of the system of vectors. □
With the above notion in hand, we can now formulate:
212 9. LINEAR MAPS

Definition 9.14. The determinant of v1 , . . . , vN ∈ RN is the signed volume

det(v1 . . . vN ) = ±vol < v1 , . . . , vN >
of the parallelepiped made by these vectors.
In other words, we are upgrading here Definition 9.8, by adding a sign to the quantity
det+ constructed there, as to potentially reach to good additivity properties:
det(v1 . . . vN ) = ±det+ (v1 . . . vN )
In relation with our original inversion problem for the square matrices, this upgrade
does not change what we have so far, and we have the following statement:
Theorem 9.15. The quantity det that we constructed, regarded as a function of the
corresponding square matrices, formed by column vectors,
det : MN (R) → R
has the property that a matrix A ∈ MN (R) is invertible precisely when det(A) ̸= 0.
Proof. We know from the above that a matrix A ∈ MN (R) is invertible precisely
when det+ (A) = | det A| is strictly positive, and this gives the result. □
Let us try now to compute the determinant. In 1 dimension we have of course the
formula det(a) = a, because the absolute value fits, and so does the sign:
det(a) = sgn(a) × |a| = a
In 2 dimensions now, we have the following result:
Theorem 9.16. In 2 dimensions we have the following formula,
a b
= ad − bc
c d
with | . | = det being the determinant function constructed above.
Proof. According to our definition, to the computation in Theorem 9.10, and to the
sign formula from Proposition 9.12, the determinant of a 2 × 2 matrix is given by:
       
a b a b + a b
det = sgn , × det
c d c d c d
   
a b
= sgn , × |ad − bc|
c d
= sgn(ad − bc) × |ad − bc|
= ad − bc
Thus, we have obtained the formula in the statement. □
 9C. THE DETERMINANT 213

9c. The determinant

In order to discuss now arbitrary dimensions, we will need a number of theoretical
results. Here is a first series of formulae, coming straight from definitions:
Theorem 9.17. The determinant has the following properties:
(1) When multiplying by scalars, the determinant gets multiplied as well:
det(λ1 v1 , . . . , λN vN ) = λ1 . . . λN det(v1 , . . . , vN )
(2) When permuting two columns, the determinant changes the sign:
det(. . . , u, . . . , v, . . .) = − det(. . . , v, . . . , u, . . .)
(3) The determinant det(e1 , . . . , eN ) of the standard basis of RN is 1.
Proof. All this is clear from definitions, as follows:
(1) This follows from definitions, and from Proposition 9.13 (1).
(2) This follows as well from definitions, and from Proposition 9.13 (2).
(3) This is clear from our definition of the determinant. □
As an application of the above result, we have:
Theorem 9.18. The determinant of a diagonal matrix is given by:
λ1
... = λ1 . . . λN
λN
That is, we obtain the product of diagonal entries, or of eigenvalues.
Proof. The formula in the statement is clear by using Theorem 9.17, which gives:
λ1 1
.. = λ1 . . . λN .. = λ1 . . . λN
. .
λN 1
As for the last assertion, this is rather a remark. □
In order to reach to a more advanced theory, let us adopt now the linear map point
of view. In this setting, the definition of the determinant reformulates as follows:
Theorem 9.19. Given a linear map, written as f (v) = Av, its “inflation coefficient”,
obtained as the signed volume of the image of the unit cube, is given by:
If = det A
More generally, If is the inflation ratio of any parallelepiped in RN , via the transformation
f . In particular f is invertible precisely when det A ̸= 0.
214 9. LINEAR MAPS

Proof. The only non-trivial thing in all this is the fact that the inflation coefficient
If , as defined above, is independent of the choice of the parallelepiped. But this is a
generalization of the Thales theorem, which follows from the Thales theorem itself. □
As a first application of the above linear map viewpoint, we have:
Theorem 9.20. We have the following formula, valid for any matrices A, B:
det(AB) = det A · det B
In particular, we have det(AB) = det(BA).
Proof. The first formula follows from the formula fAB = fA fB for the associated
linear maps. As for det(AB) = det(BA), this is clear from the first formula. □
Getting back now to explicit computations, we have the following key result:
Theorem 9.21. The determinant of a diagonalizable matrix,
λ1
 

A∼ ..
. 
λN
is the product of its eigenvalues, det A = λ1 . . . λN .
Proof. We have not talked yet about diagonalization, and more on this in a moment,
but the idea is that a matrix is diagonalizable when it can be written in the form A =
P DP −1 , with D = diag(λ1 , . . . , λN ). Now by using Theorem 9.20, we obtain:
det A = det(P DP −1 )
= det(DP −1 P )
= det D
= λ1 . . . λN
Thus, we are led to the formula in the statement. □
In general now, at the theoretical level, we have the following key result:
Theorem 9.22. The determinant has the additivity property
det(. . . , u + v, . . .) = det(. . . , u, . . .) + det(. . . , v, . . .)
valid for any choice of the vectors involved.
Proof. This follows by doing some elementary geometry, in the spirit of the compu-
tations in the proof of Theorem 9.10, as follows:
(1) We can either use the Thales theorem, and then compute the volumes of all the
parallelepipeds involved, by using basic algebraic formulae.
 9C. THE DETERMINANT 215

(2) Or we can solve the problem in “puzzle” style, the idea being to cut the big
parallelepiped, and then recover the small ones, after some manipulations.
(3) We can do as well something hybrid, consisting in deforming the parallelepipeds
involved, without changing their volumes, and then cutting and gluing. □
As a basic application of the above result, we have:
Theorem 9.23. We have the following results:
(1) The determinant of a diagonal matrix is the product of diagonal entries.
(2) The same is true for the upper triangular matrices.
(3) The same is true for the lower triangular matrices.
Proof. All this can be deduced by using our various general formulae, as follows:
(1) This is something that we already know, from Theorem 9.21.
(2) This follows by using our various formulae, then (1), as follows:
λ1 ∗ λ1 0 ∗
λ2 λ2
.. = ..
. .
0 λN 0 λN
..
.
..
.
λ1 0
λ2
= ..
.
0 λN
= λ1 . . . λN
(3) This follows as well from our various formulae, then (1), by proceeding this time
from right to left, from the last column towards the first column. □
As an important theoretical result now, we have:
Theorem 9.24. The determinant of square matrices is the unique map
det : MN (R) → R
satisfying the conditions found above.
Proof. Any map det′ : MN (R) → R satisfying our conditions must indeed coincide
with det on the upper triangular matrices, and then all the matrices. □
Here is now another important theoretical result:
216 9. LINEAR MAPS

Theorem 9.25. The determinant is subject to the row expansion formula

a11 . . . a1N a22 . . . a2N a21 a23 . . . a2N

.. .. = a11 .. .. − a .. .. ..
. . . . 12 . . .
aN 1 . . . aN N aN 2 . . . aN N aN 1 aN 3 . . . aN N
a21 . . . a2,N −1
+ . . . . . . + (−1) N +1
a1N .. ..
. .
aN 1 . . . aN,N −1

and this method fully computes it, by recurrence.

Proof. This follows from the fact that the formula in the statement produces a
certain function det : MN (R) → R, which has the 4 properties in Theorem 9.24. □

We can expand as well over the columns, as follows:

Theorem 9.26. The determinant is subject to the column expansion formula

a12 . . . a1N
a11 . . . a1N a22 . . . a2N
.. .. = a11 .. .. − a a32 . . . a3N
. . . . 21 .. ..
. .
aN 1 . . . aN N aN 2 . . . aN N
aN 2 . . . aN N
a12 ... a1N
+ . . . . . . + (−1) N +1
aN 1 .. ..
. .
aN −1,2 . . . aN −1,N

and this method fully computes it, by recurrence.

Proof. This follows by using the same argument as for the rows. □

As a first application of the above methods, we can now prove:

Theorem 9.27. The determinant of the 3 × 3 matrices is given by

a b c
d e f = aei + bf g + cdh − ceg − bdi − af h
g h i

which can be memorized by using Sarrus’ triangle method, “triangles parallel to the diag-
onal, minus triangles parallel to the antidiagonal”.
 9C. THE DETERMINANT 217

Proof. Here is the computation, using the above results:

a b c
e f d f d e
d e f = a −b +c
h i g i g h
g h i
= a(ei − f h) − b(di − f g) + c(dh − eg)
= aei − af h − bdi + bf g + cdh − ceg
= aei + bf g + cdh − ceg − bdi − af h
Thus, we obtain the formula in the statement. □
Let us discuss now the general formula of the determinant, at arbitrary values N ∈ N
of the matrix size, generalizing the formulae that we have at N = 2, 3. We will need:
Definition 9.28. A permutation of {1, . . . , N } is a bijection, as follows:
σ : {1, . . . , N } → {1, . . . , N }
The set of such permutations is denoted SN .
There are many possible notations for the permutations, the simplest one consisting
in writing the numbers 1, . . . , N , and below them, their permuted versions:
 
1 2 3 4 5
σ=
2 1 4 5 3
Another method, which is better for most purposes, and faster too, remember that
time is money, is by denoting permutations as diagrams, going from top to bottom:

σ=

There are many interesting things that can be said about permutations. In what
concerns us, we will need the following key result:
Theorem 9.29. The permutations have a signature function
ε : SN → {±1}
which can be defined in the following equivalent ways:
(1) As (−1)c , where c is the number of inversions.
(2) As (−1)t , where t is the number of transpositions.
(3) As (−1)o , where o is the number of odd cycles.
(4) As (−1)x , where x is the number of crossings.
(5) As the sign of the corresponding permuted basis of RN .
218 9. LINEAR MAPS

Proof. Let us begin with the precise definition of c, t, o, x, as numbers modulo 2:

(1) The idea here is that given any two numbers i < j among 1, . . . , N , the permutation
can either keep them in the same order, σ(i) < σ(j), or invert them:
σ(j) > σ(i)
Now by making i < j vary over all pairs of numbers in 1, . . . , N , we can count the
number of inversions, and call it c. This is an integer, c ∈ N, which is well-defined.
(2) Here the idea, which is something quite intuitive, is that any permutation appears
as a product of switches, also called transpositions:
i↔j
The decomposition as a product of transpositions is not unique, but the number t of
the needed transpositions is unique, when considered modulo 2. This follows for instance
from the equivalence of (2) with (1,3,4,5), explained below.
(3) Here the point is that any permutation decomposes, in a unique way, as a product
of cycles, which are by definition permutations of the following type:
i1 → i2 → i3 → . . . . . . → ik → i1
Some of these cycles have even length, and some others have odd length. By counting
those having odd length, we obtain a well-defined number o ∈ N.
(4) Here the method is that of drawing the permutation, as we usually do, and by
avoiding triple crossings, and then counting the number of crossings. This number x
depends on the way we draw the permutations, but modulo 2, we always get the same
number. Indeed, this follows from the fact that we can continuously pass from a drawing
to each other, and that when doing so, the number of crossings can only jump by ±2.
Summarizing, we have 4 different definitions for the signature of the permutations,
which all make sense, constructed according to (1-4) above. Regarding now the fact that
we always obtain the same number, this can be established as follows:
(1)=(2) This is clear, because any transposition inverts once, modulo 2.
(1)=(3) This is clear as well, because the odd cycles invert once, modulo 2.
(1)=(4) This comes from the fact that the crossings correspond to inversions.
(2)=(3) This follows by decomposing the cycles into transpositions.
(2)=(4) This comes from the fact that the crossings correspond to transpositions.
(3)=(4) This follows by drawing a product of cycles, and counting the crossings.
Finally, in what regards the equivalence of all these constructions with (5), here sim-
plest is to use (2). Indeed, we already know that the sign of a system of vectors switches
when interchanging two vectors, and so the equivalence between (2,5) is clear. □
 9C. THE DETERMINANT 219

Now back to linear algebra, we can formulate a key result, as follows:

Theorem 9.30. We have the following formula for the determinant,
X
det A = ε(σ)A1σ(1) . . . AN σ(N )
σ∈SN

with the signature function being the one introduced above.

Proof. This follows by recurrence over N ∈ N, as follows:
(1) When developing the determinant over the first column, we obtain a signed sum of
N determinants of size (N −1)×(N −1). But each of these determinants can be computed
by developing over the first column too, and so on, and we are led to the conclusion that
we have a formula as in the statement, with ε(σ) ∈ {−1, 1} being certain coefficients.
(2) But these latter coefficients ε(σ) ∈ {−1, 1} can only be the signatures of the
corresponding permutations σ ∈ SN , with this being something that can be viewed again
by recurrence, with either of the definitions (1-5) in Theorem 9.29 for the signature. □
The above result is something quite tricky, and in order to get familiar with it, there is
nothing better than doing some computations. As a first, basic example, in 2 dimensions
we recover the usual formula of the determinant, the details being as follows:
a b
= ε(| |) · ad + ε(/\) · cb
c d
= 1 · ad + (−1) · cb
= ad − bc
In 3 dimensions now, we recover the Sarrus formula:
a b c
d e f = aei + bf g + cdh − ceg − bdi − af h
g h i
Let us record as well the following general formula, which was not obvious with our
previous theory of the determinant, but which follows from Theorem 9.30:
det(A) = det(At )
There are countless other applications of the formula in Theorem 9.30, and we will be
back to this, on several occassions. But, most importantly, that formula allows us to deal
now with the complex matrices too, by formulating the following statement:
Theorem 9.31. If we define the determinant of a complex matrix A ∈ MN (C) to be
X
det A = ε(σ)A1σ(1) . . . AN σ(N )
σ∈SN

then this determinant has the same properties as the determinant of the real matrices.
220 9. LINEAR MAPS

Proof. This follows by doing some sort of reverse engineering, with respect to what
has been done in this section, and we reach to the conclusion that det has indeed all the
good properties that we are familiar with. Except of course for the properties at the very
beginning of this section, in relation with volumes, which don’t extend well to CN . □

Good news, this is the end of the general theory that we wanted to develop. We have
now in our bag all the needed techniques for computing the determinant.

9d. Diagonalization
Let us discuss now the diagonalization question for linear maps and matrices. The
basic diagonalization theory, formulated in terms of matrices, is as follows:
Proposition 9.32. A vector v ∈ CN is called eigenvector of A ∈ MN (C), with
corresponding eigenvalue λ, when A multiplies by λ in the direction of v:
Av = λv
In the case where CN has a basis v1 , . . . , vN formed by eigenvectors of A, with correspond-
ing eigenvalues λ1 , . . . , λN , in this new basis A becomes diagonal, as follows:
λ1
 

A∼ ..
. 
λN
Equivalently, if we denote by D = diag(λ1 , . . . , λN ) the above diagonal matrix, and by
P = [v1 . . . vN ] the square matrix formed by the eigenvectors of A, we have:
A = P DP −1
In this case we say that the matrix A is diagonalizable.
Proof. This is something which is clear, the idea being as follows:
(1) The first assertion is clear, because the matrix which multiplies each basis element
vi by a number λi is precisely the diagonal matrix D = diag(λ1 , . . . , λN ).
(2) The second assertion follows from the first one, by changing the basis. We can
prove this by a direct computation as well, because we have P ei = vi , and so:
P DP −1 vi = P Dei = P λi ei = λi P ei = λi vi
Thus, the matrices A and P DP −1 coincide, as stated. □

In order to study the diagonalization problem, the idea is that the eigenvectors can
be grouped into linear spaces, called eigenspaces, as follows:
 9D. DIAGONALIZATION 221

Theorem 9.33. Let A ∈ MN (C), and for any eigenvalue λ ∈ C define the correspond-
ing eigenspace as being the vector space formed by the corresponding eigenvectors:
n o
Eλ = v ∈ CN Av = λv
These eigenspaces Eλ are then in a direct sum position, in the sense that given vectors
v1 ∈ Eλ1 , . . . , vk ∈ Eλk corresponding to different eigenvalues λ1 , . . . , λk , we have:
X
ci vi = 0 =⇒ ci = 0
i
P
In particular, we have λ dim(Eλ ) ≤ N , with the sum being over all the eigenvalues, and
our matrix is diagonalizable precisely when we have equality.
Proof. We prove the first assertion by recurrence on k ∈ N. Assume by contradiction
that we have a formula as follows, with the scalars c1 , . . . , ck being not all zero:
c1 v1 + . . . + ck vk = 0
By dividing by one of these scalars, we can assume that our formula is:
vk = c1 v1 + . . . + ck−1 vk−1
Now let us apply A to this vector. On the left we obtain:
Avk = λk vk = λk c1 v1 + . . . + λk ck−1 vk−1
On the right we obtain something different, as follows:
A(c1 v1 + . . . + ck−1 vk−1 ) = c1 Av1 + . . . + ck−1 Avk−1
= c1 λ1 v1 + . . . + ck−1 λk−1 vk−1
We conclude from this that the following equality must hold:
λk c1 v1 + . . . + λk ck−1 vk−1 = c1 λ1 v1 + . . . + ck−1 λk−1 vk−1
On the other hand, we know by recurrence that the vectors v1 , . . . , vk−1 must be
linearly independent. Thus, the coefficients must be equal, at right and at left:
λ k c1 = c1 λ 1
..
.
λk ck−1 = ck−1 λk−1
Now since at least one of the numbers ci must be nonzero, from λk ci = ci λi we obtain
λk = λi , which is a contradiction. Thus our proof by recurrence of the first assertion is
complete. As for the second assertion, this follows from the first one. □
In order to reach now to more advanced results, we can use the characteristic polyno-
mial, which appears via the following fundamental result:
222 9. LINEAR MAPS

Theorem 9.34. Given a matrix A ∈ MN (C), consider its characteristic polynomial:

P (x) = det(A − x1N )
The eigenvalues of A are then the roots of P . Also, we have the inequality
dim(Eλ ) ≤ mλ
where mλ is the multiplicity of λ, as root of P .
Proof. The first assertion follows from the following computation, using the fact that
a linear map is bijective when the determinant of the associated matrix is nonzero:
∃v, Av = λv ⇐⇒ ∃v, (A − λ1N )v = 0
⇐⇒ det(A − λ1N ) = 0
Regarding now the second assertion, given an eigenvalue λ of our matrix A, consider
the dimension dλ = dim(Eλ ) of the corresponding eigenspace. By changing the basis of
CN , as for the eigenspace Eλ to be spanned by the first dλ basis elements, our matrix
becomes as follows, with B being a certain smaller matrix:
 
λ1dλ 0
A∼
0 B
We conclude that the characteristic polynomial of A is of the following form:
PA = Pλ1dλ PB = (λ − x)dλ PB
Thus the multiplicity mλ of our eigenvalue λ, as a root of P , satisfies mλ ≥ dλ , and
this leads to the conclusion in the statement. □
Now recall that we are over C. We obtain the following result:
Theorem 9.35. Given a matrix A ∈ MN (C), consider its characteristic polynomial
P (X) = det(A − X1N )
then factorize this polynomial, by computing the complex roots, with multiplicities,
P (X) = (−1)N (X − λ1 )n1 . . . (X − λk )nk
and finally compute the corresponding eigenspaces, for each eigenvalue found:
n o
N
Ei = v ∈ C Av = λi v

The dimensions of these eigenspaces satisfy then the following inequalities,

dim(Ei ) ≤ ni
and A is diagonalizable precisely when we have equality for any i.
 9D. DIAGONALIZATION 223

Proof. This follows by combining the above results. Indeed, by summing the in-
equalities dim(Eλ ) ≤ mλ from Theorem 9.34, we obtain an inequality as follows:
X X
dim(Eλ ) ≤ mλ ≤ N
λ λ

On the other hand, we know from Theorem 9.33 that our matrix is diagonalizable
when we have global equality. Thus, we are led to the conclusion in the statement. □
As an illustration for all this, which is a must-know computation, we have:
Proposition 9.36. The rotation of angle t ∈ R in the plane diagonalizes as:
     −it  
cos t − sin t 1 1 1 e 0 1 −i
=
sin t cos t 2 i −i 0 eit 1 i
Over the reals this is impossible, unless t = 0, π, where the rotation is diagonal.
Proof. Observe first that, as indicated, unlike we are in the case t = 0, π, where our
rotation is ±12 , our rotation is a “true” rotation, having no eigenvectors in the plane.
Fortunately the complex numbers come to the rescue, via the following computation:
      
cos t − sin t 1 cos t − i sin t −it 1
= =e
sin t cos t i i cos t + sin t i
We have as well a second complex eigenvector, coming from:
      
cos t − sin t 1 cos t + i sin t it 1
= =e
sin t cos t −i −i cos t + sin t −i
Thus, we are led to the conclusion in the statement. □
At the level of basic examples of diagonalizable matrices, we first have the following
result, which provides us with the “generic” examples:
Theorem 9.37. For a matrix A ∈ MN (C) the following conditions are equivalent,
(1) The eigenvalues are different, λi ̸= λj ,
(2) The characteristic polynomial P has simple roots,
(3) The characteristic polynomial satisfies (P, P ′ ) = 1,
(4) The resultant of P, P ′ is nonzero, R(P, P ′ ) ̸= 0,
(5) The discriminant of P is nonzero, ∆(P ) ̸= 0,
and in this case, the matrix is diagonalizable.
Proof. The last assertion holds indeed, due to Theorem 9.35. As for the equivalences
in the statement, these are all standard, by using the theory of R, ∆ from chapter 5. □
As already mentioned, one can prove that the matrices having distinct eigenvalues are
“generic”, and so the above result basically captures the whole situation. We have in fact
the following collection of density results, which are quite advanced:
224 9. LINEAR MAPS

Theorem 9.38. The following happen, inside MN (C):

(1) The invertible matrices are dense.
(2) The matrices having distinct eigenvalues are dense.
(3) The diagonalizable matrices are dense.
Proof. These are quite advanced results, which can be proved as follows:
(1) This is clear, intuitively speaking, because the invertible matrices are given by the
condition det A ̸= 0. Thus, the set formed by these matrices appears as the complement
of the hypersurface det A = 0, and so must be dense inside MN (C), as claimed.
(2) Here we can use a similar argument, this time by saying that the set formed by
the matrices having distinct eigenvalues appears as the complement of the hypersurface
given by ∆(PA ) = 0, and so must be dense inside MN (C), as claimed.
(3) This follows from (2), via the fact that the matrices having distinct eigenvalues are
diagonalizable, that we know from Theorem 9.37. There are of course some other proofs
as well, for instance by putting the matrix in Jordan form. □
This was for the basic theory of the linear maps, that we will need in what follows.
There are of course far more things that can be said, and we will be back to this gradually,
when needed, and notably in chapter 12 below, when doing advanced calculus.
9e. Exercises
This was an easy chapter, and the start of something new, at least when compared
with the horrors from chapters 7-8, and here are some easy exercises for you:
Exercise 9.39. Work out all the details for the main properties of the determinant
function det : MN (R) → R, by using our geometric approach, and Thales.
Exercise 9.40. Prove that for H ∈ MN (±1) we have | det H| ≤ N N/2 , with equality
precisely when H is Hadamard, in the sense that its rows are pairwise orthogonal.
Exercise 9.41. Work out all the details for the properties of the determinant function
det : MN (C) → C, by using our algebraic approach, and permutations.
Exercise 9.42. Learn the Jordan form, and come up with an alternative proof for
our result stating that the diagonalizable matrices are dense.
As an important bonus exercise, diagonalize 3 × 3 matrices, that you can easily find
on the internet, as many as needed, as for the computation to take 20 minutes.
 CHAPTER 10

Partial derivatives

10a. Functions, continuity

We have seen so far the theory of linear maps f : RN → RM , and f : CN → CM . In
what follows we develop the theory of maps f : RN → RM and f : CN → CM in general,
in analogy with what we know about the maps f : R → R, from Part I of this book, and
with what we know about the maps f : C → C, from Part II of this book.

We will rely of course as well on the substantial help from the linear algebra that
we learned in chapter 9, which more or less solves all our potential questions regarding
analysis, when the map f in question happens to be linear, of the following form:
f (x) = Ax
However, things in general will be quite tricky, among others because a map f : C → C
can be regarded as a map f : R2 → R2 , leading to an obvious dimension mess. So, in
order not to mess up things with dimensions, we will first study the real case, that of
the maps f : RN → RM , with minimal reference to the complex numbers, and study the
maps f : CN → CM only afterwards. And with the remark that this plan, while certainly
reasonable, will stumble at some point into some difficulties, coming from the fact that
a linear map f : RN → RN can have complex eigenvalues, and so staying over R is not
enough, for fully understanding it. Welcome to the several variables mess.

But probably enough talking, let us get started. As a first objective, we would like to
talk about the continuity, and other basic analytic properties, of the maps f : RN → RM .
And for this, we will run all the time into the formula of the distance in RN , namely:
v
u N
uX
d(x, y) = t (x − y )2
i i
i=1

In order to avoid using all the time this formula, which quite often can lead into
complicated computations, and even into following wrong paths, it is convenient to relax
a bit, and take an abstract point of view on all this.

So, let us begin by axiomatizing the properties of the distance d(x, y) given above, by
generalizing what we know in RN . This leads us into the following notion:
225
226 10. PARTIAL DERIVATIVES

Definition 10.1. A metric space is a set X with a distance function d : X ×X → R+ ,

having the following properties:
(1) d(x, y) > 0 if x ̸= y, and d(x, x) = 0.
(2) d(x, y) = d(y, x).
(3) d(x, y) ≤ d(x, z) + d(y, z).
As a basic example, we have RN , as well as any of its subsets X ⊂ RN . Indeed, the
first two axioms are clear, and for the third axiom, we must prove that:
s s s
X X X
2
(xi − yi ) ≤ 2
(xi − zi ) + (yi − zi )2
i i i

But with a = x − z and b = y − z, this is the same as proving that:

s s s
X X X
2
(ai + bi )2 ≤ ai + b2i
i i i

Now by raising to the square, this is the same as proving that:

!2 ! !
X X X
ai b i ≤ a2i b2i
i i i

But this latter inequality is one of the many equivalent formulations of the Cauchy-
Schwarz inequality, that we know from chapter 7, and which follows as well by using the
fact that the following function being positive, its discriminant must be negative:
X
f (t) = (ai + tbi )2
i

As another example, we have CN , as well as any of its subsets X ⊂ CN . Indeed,

this follows either from CN ≃ R2N , which is an isomorphism of metric spaces, or directly,
along the lines of the above proof for RN , by making changes where needed. To be more
precise, after doing the algebra, we are left with proving the following inequality:
2 ! !
X X X
ai b̄i ≤ |ai |2 |bi |2
i i i

But this is the complex version of the Cauchy-Schwarz inequality, that we know also
from chapter 7, and which follows as well directly, by using the fact that the following
function, with t ∈ R and |w| = 1, being positive, its discriminant must be negative:
X
f (t) = |ai + twbi |2
i

Here is now another example, which at first looks new and interesting, but is in fact
not new, because it appears as a subspace of some suitable RN :
 10A. FUNCTIONS, CONTINUITY 227

Proposition 10.2. Given a finite set X, the following function is a metric on it,
called discrete metric: (
1 if x ≠ y
d(x, y) =
0 if x = y
This metric space is in fact the N -simplex, with N = |X|, and can be realized as a subspace
of RN −1 , or, more conveniently, as a subspace of RN .
Proof. There are several things going on here, the idea being as follows:
(1) First of all, the axioms from Definition 10.1 are trivially satisfied, and with the
main axiom, namely the triangle inequality, basically coming from:
1+1≤1
(2) At the level of examples, at |X| = 1 we obtain a point, at |X| = 2 we obtain a
segment, at |X| = 3 we obtain an equilateral triangle, at |X| = 4 we obtain a regular
tetrahedron, and so on. Thus, what we have in general, at |X| = N , is the arbitrary
dimensional generalization of this series of geometric objects, called N -simplex.
(3) In what regards now the geometric generalization of the N -simplex, our above
examples, namely segment, triangle, tetrahedron and so on, suggest to look for an em-
bedding X ⊂ RN −1 . This is something which is certainly possible, but the computations
here are quite complicated, involving a lot of trigonometry, as you can check yourself by
studying the problem at N = 4, that is, parametrizing the regular tetrahedron in R3 .
(4) However, mathematics, or perhaps physics, come to the rescue, via the idea “add a
dimension, for getting smarter”. Indeed, when looking for an embedding X ⊂ RN things
drastically simplify, because we can simply take X to be the standard basis of RN :
X = {e1 , . . . , eN }
Indeed, we have by definition d(ei , ej ) = 1 for any i ̸= j. So, we solved our embedding
problem, just like that, without doing any computations or trigonometry. □
Getting back now to Definition 10.1 as it is, the axioms there are satisfied as well for the
various spaces of infinite dimensions that we met in chapter 7. Thus, the whole discussion
following Definition 10.1 generalizes to arbitrary dimensions. Also, so does Proposition
10.2, to the case of the sets X of arbitrary cardinality, and with the realization of the
corresponding metric space being again by using a standard basis, as follows:
X = {ex }x∈X ⊂ l2 (X)
However, in what follows we will not insist much on these latter examples, because we
will be mainly interested, as planned before, in RN , CN and their subspaces.

Moving ahead now with some theory, and allowing us a bit of slopiness, we have:
228 10. PARTIAL DERIVATIVES

Proposition 10.3. We can talk about limits inside metric spaces X, by saying that
xn → x ⇐⇒ d(xn , x) → 0
and we can talk as well about continuous functions f : X → Y , by requiring that
xn → x =⇒ f (xn ) → f (x)
and with these notions in hand, all the basic results from the cases X = R, C extend.
Proof. All this is very standard, and we will leave this as an exercise, namely carefully
checking what we did so far in this book, in relation with limits and continuity, in the
cases X = R, C, and working out the metric space extensions of this. Of course I can hear
you screaming that this is too much work, but believe me, after some thinking, there is
in fact not much work to be done. Indeed, all that we have been doing of advanced type
requires sums x + y or multiplication by scalars λx, and such operations being not allowed
in the general context of Definition 10.1, we are just left with a handful of trivialities,
that you can surely work out, as a quick and instructive exercise. □
More interestingly now, we can talk about open and closed sets inside metric spaces
X, again in analogy with what we did for X = R, C, but with a whole lot of interesting
new phenomena appearing. So, we will do this in detail. Let us start with:
Definition 10.4. Let X be a metric space.
(1) The open balls are the sets Bx (r) = {y ∈ X|d(x, y) < r}.
(2) The closed balls are the sets B̄x (r) = {y ∈ X|d(x, y) ≤ r}.
(3) E ⊂ X is called open if for any x ∈ E we have a ball Bx (r) ⊂ E.
(4) E ⊂ X is called closed if its complement E c ⊂ X is open.
At the level of examples, you can quickly convince yourself, by working out a few of
them, that our notions above coincide with the usual ones, that we know well, in the
cases X = R, C. We will be back to this later, with some general results in this sense,
confirming all this. But for the moment, let us work out the basics. We first have the
following result, clarifying some terminology issues from Definition 10.4:
Proposition 10.5. The open balls are open, and the closed balls are closed.
Proof. This might sound a bit as a joke, but it is not one, because this is the kind
of thing that we have to duly check. Fortunately, all this is elementary, as follows:
(1) Given an open ball Bx (r) and a point y ∈ Bx (r), by using the triangle inequality
we have By (r′ ) ⊂ Bx (r), with r′ = r − d(x, y). Thus, Bx (r) is indeed open.
(2) Given a closed ball B̄x (r) and a point y ∈ Bx (r)c , by using the triangle inequality
we have By (r′ ) ⊂ Bx (r)c , with r′ = d(x, y) − r. Thus, B̄x (r) is indeed closed. □
Here is now something more interesting, making the link with our intuitive under-
standing of the notion of closedness, coming from our experience so far with analysis:
 10A. FUNCTIONS, CONTINUITY 229

Theorem 10.6. For a set E ∈ X, the following are equivalent:

(1) E is closed in our sense, meaning that E c is open.
(2) We have xn → x, xn ∈ E =⇒ x ∈ E.
Proof. We can prove this by double implication, as follows:
(1) =⇒ (2) Assume by contradiction xn → x, xn ∈ E with x ∈ / E. Since we
c c
have x ∈ E , which is open, we can pick a ball Bx (r) ⊂ E . But this contradicts our
convergence assumption xn → x, so we are done with this implication.
(2) =⇒ (1) Assume by contradiction that E is not closed in our sense, meaning that
E c is not open. Thus, we can find x ∈ E c such that there is no ball Bx (r) ⊂ E c . But
with r = 1/n this provides us with a point xn ∈ Bx (1/n) ∩ E, and since we have xn → x,
this contradicts our assumption (2). Thus, we are done with this implication too. □

Here is another basic theorem about open and closed sets:

Theorem 10.7. Let X be a metric space.
(1) If Ei are open, then ∪i Ei is open.
(2) If Fi are closed, then ∩i Fi is closed.
(3) If E1 , . . . , En are open, then ∩i Ei is open.
(4) If F1 , . . . , Fn are closed, then ∪i Fi is closed.
Moreover, both (3) and (4) can fail for infinite intersections and unions.
Proof. We have several things to be proved, the idea being as follows:
(1) This is clear from definitions, because any point x ∈ ∪i Ei must satisfy x ∈ Ei for
some i, and so has a ball around it belonging to Ei , and so to ∪i Ei .
(2) This follows from (1), by using the following well-known set theory formula:
!c
[ \
Ei = Eic
i i

(3) Given an arbitrary point x ∈ ∩i Ei , we have x ∈ Ei for any i, and so we have a

ball Bx (ri ) ⊂ Ei for any i. Now with this in hand, let us set:
B = Bx (r1 ) ∩ . . . ∩ Bx (rn )
As a first observation, this is a ball around x, B = Bx (r), of radius given by:
r = min(r1 , . . . , rn )
But this ball belongs to all the Ei , and so belongs to their intersection ∩i Ei . We
conclude that the intersection ∩i Ei is open, as desired.
230 10. PARTIAL DERIVATIVES

(4) This follows from (3), by using the following well-known set theory formula:
!c
\ [
Ei = Eic
i i

(5) Finally, in what regards the counterexamples at the end, these can be both found
on R. Indeed, for the infinite intersections of open sets, we can use:
\ 1 1
− , = {0}
n
n n
As for the infinite unions of closed sets, here we can use:
[ 1

0, 1 − = [0, 1)
n
n
Thus, we are led to the conclusions in the statement. □
Finally, still in relation with open and closed sets, we have as well:
Definition 10.8. Let X be a metric space, and E ⊂ X be a subset.
(1) The interior E ◦ ⊂ E is the set of points x ∈ E which admit around them open
balls Bx (r) ⊂ E.
(2) The closure E ⊂ Ē is the set of points x ∈ X which appear as limits of sequences
xn → x, with x ∈ E.
These notions are quite interesting, because they make sense for any set E. That is,
when E is open, it is open and end of the story, and when E is closed, it is closed and
end of the story too. In general, however, a set E ⊂ X is not open or closed, and what
we can best do to it, in order to study it with our tools, is to “squeeze” it, as follows:
E ◦ ⊂ E ⊂ Ē
In practice now, in order to use the above notions, we need to know a number of
things, including that fact that E open implies E ◦ = E, the fact that E closed implies
Ē = E, and many more such results, not to forget the fact that the closures of the open
balls Br (x) are the closed balls B̄x (r), clarifying an obvious notational issue which appears
with respect to Definition 10.4. But all this can be done, and the useful statement here,
summarizing all we need to know about interiors and closures, is as follows:
Theorem 10.9. Let X be a metric space, and E ⊂ X be a subset.
(1) The interior E ◦ ⊂ E is the biggest open set contained in E.
(2) The closure E ⊂ Ē is the smallest closed set containing E.
Proof. We have several things to be proved, the idea being as follows:
 10B. COMPACT SETS 231

(1) Let us first prove that the interior E ◦ is open. For this purpose, pick x ∈ E ◦ . We
know that we have a ball Bx (r) ⊂ E, and since this ball is open, it follows that we have
Bx (r) ⊂ E ◦ . Thus, the interior E ◦ is open, as claimed.
(2) Let us prove now that the closure Ē is closed. For this purpose, we will prove
that the complement Ē c is open. So, pick x ∈ Ē c . Then x cannot appear as a limit of a
sequence xn → x with xn ∈ E, so we have a ball Bx (r) ⊂ Ē c , as desired.
(3) Finally, the maximality and minimality assertions regarding E ◦ and Ē are both
routine, coming from definitions, and we will leave them as exercises. □
As an application of the theory developed above, and more specifically of the notion
of closure from Definition 10.8, we can talk as well about density, as follows:
Definition 10.10. We say that a subset E ⊂ X is dense when:
Ē = X
That is, any point of X must appear as a limit of points of E.
Obviously, this is something which is in tune with what we know so far from this book,
and with the intuitive notion of density. As a basic example, we have Q̄ = R, that we
know well from the beginning of this book. As another example, we have the fact that
the diagonalizable matrices are dense inside MN (C), that we know from chapter 9. There
are of course many other examples, and we will be back to this, in what follows.

10b. Compact sets

Moving ahead now, at a more subtle level, again in analogy with what we know about
X = R, C, we can talk about compact sets, and about connected sets. Again things here
are quite tricky, in the general metric space framework, actually substantially deviating
from what we know, and we will do this in detail. Let us start with:
Definition 10.11. A set K ⊂ X is called compact if any cover with open sets
[
K⊂ Ei
i

has a finite subcover, K ⊂ (Ei1 ∪ . . . ∪ Ein ).

This definition might seem overly abstract, and perhaps even sound like a joke, but
our claim is that this is the correct definition, and that there is no way of doing otherwise.
Let us start with some examples, with X = R. The situation here is as follows:

(1) A point is obviously compact, and we can choose that finite subcover with n = 1.
Similarly, 2 points are compact, and we can choose the subcover with n = 2. More
generally, N points are compact, and we can choose the subcover with n = N .
232 10. PARTIAL DERIVATIVES

(2) In contrast, the set N ⊂ R is not compact, because the following open cover of it
obviously has no finite subcover:
[ 1 1

N⊂ n− , n+
n
3 3
Similarly, the set {1/n|n ∈ N} is not compact either, with a similar cover, consisting
of a suitable union of open intervals around each point, doing the job.

(3) However, and here comes an interesting point, the following set is compact:
 
1
K= n ∈ N ∪ {0}
n
Indeed, any open cover of it ∪i Ei has to cover 0, and by selecting an open set Ei
covering 0, this set Ei will cover the whole K, except for finitely many points, due to
1/n → 0. But these finitely points left, say N of them, can be covered by suitable sets
Ei1 , . . . , EiN , and by adding to this family the set Ei , we have our finite subcover.

As a conclusion to this, Definition 10.11 seems to be in tune with what we know about
the compact subsets K ⊂ R, namely that these are the sets which are closed and bounded.
However, and here comes our point, such things are wrong in general, due to:
Theorem 10.12. Given an infinite set X with the discrete distance on it, namely
d(p, q) = 1 − δpq , this can be modeled as the basis of a suitable Hilbert space,
X = {ex }x∈X ⊂ l2 (X)
and this set is closed and bounded, but not compact.
Proof. Here the first part, regarding the modeling of X, that we will actually not
really need, is something that we already know. Regarding now the second part:
(1) X being the total space, it is by definition closed. As a remark here, that we will
need later, since the points of X are obviously open, any subset E ⊂ X is open, and by
taking complements, any set E ⊂ X is closed as well.
(2) X is also bounded, because all distances are smaller than 1.
(3) However, our set X is not compact, because its points being open, as noted above,
X = ∪x∈X {x} is an open cover, having no finite subcover. □
The above result is quite interesting, and is the source of many troubles with compact-
ness, in general, and can be subject to some further meditation. First, you might argue
that by declaring X to be the total space, we have “cheated” with its closedness, which
came like this for free, without computations. But, when using the modeling X ⊂ l2 (X)
mentioned above, X still remains closed, this time with no cheating involved.
 10B. COMPACT SETS 233

As a second thought that you might have, you might perhaps say okay, but why not
changing our definition of compactness as to include the above set X, which looks quite
nice, being both closed and bounded, into our class of compact sets. But this does not
work, because after some thinking, the above set X is in fact not nice at all, and all sorts
of things that you can try with it, for “confirming” its compactness, will simply fail.
So, this is the situation, and as a conclusion, Definition 10.11 is both in tune with
what we know about the compact sets K ⊂ R, which must be closed and bounded, and
with what we learned from Theorem 10.12, telling us that being closed and bounded is
worth nothing, or almost, in general. Thus, Definition 10.11 is the correct definition.
Moving away now from these philosophical thoughts, or rather with the promise to
come back to them later, when we will know more, let us develop the theory of compact
sets, as axiomatized in Definition 10.11, and see what we get. We first have the following
elementary result, confirming that we are on the good way, with our Definition 10.11:
Proposition 10.13. The following hold:
(1) Compact implies closed.
(2) Closed inside compact is compact.
(3) Compact intersected with closed is compact.
Proof. These assertions are all clear from definitions, as follows:
(1) Assume that K ⊂ X is compact, and let us prove that K is closed. For this
purpose, we will prove that K c is open. So, pick p ∈ K c . For any q ∈ K we set
r = d(p, q)/3, and we consider the following balls, separating p and q:
Uq = Bp (r) , Vq = Bq (r)
We have then K ⊂ ∪q∈K Vq , so we can pick a finite subcover, as follows:
K ⊂ (Vq1 ∪ . . . ∪ Vqn )
With this done, consider the following intersection:
U = Uq1 ∩ . . . ∩ Uqn
This intersection is then a ball around p, and since this ball avoids Vq1 , . . . , Vqn , it
avoids the whole K. Thus, we have proved that K c is open at p, as desired.
(2) Assume that F ⊂ K is closed, with K ⊂ X being compact. For proving our result,
we can assume, by replacing X with K, that we have X = K. In order to prove now that
F is compact, consider an open cover of it, as follows:
[
F ⊂ Ei
i
c
By adding the set F , which is open, to this cover, we obtain a cover of K. Now since
K is compact, we can extract from this a finite subcover Ω, and there are two cases:
234 10. PARTIAL DERIVATIVES

– If F c ∈ Ω, by removing F c from Ω we obtain a finite cover of F , as desired.

– If F c ∈
/ Ω, we are done too, because in this case Ω is a finite cover of F .
(3) This follows from (1) and (2), because if K ⊂ X is compact, and F ⊂ X is closed,
then K ∩ F ⊂ K is closed inside a compact, so it is compact. □
As a second batch of results, which are useful as well, we have:
Proposition 10.14. The following hold:
(1) If Ki ⊂ X are compact, satisfying Ki1 ∩ . . . ∩ Kin ̸= ∅, then ∩i Ki ̸= ∅.
(2) If K1 ⊃ K2 ⊃ K3 ⊃ . . . are non-empty compacts, then ∩i Ki ̸= ∅.
(3) If K is compact, and E ⊂ K is infinite, then E has a limit point in K.
(4) If K is compact, any sequence {xn } ⊂ K has a limit point in K.
(5) If K is compact, any {xn } ⊂ K has a subsequence which converges in K.
Proof. Again, these are elementary results, which can be proved as follows:
(1) Assume by contradiction ∩i Ki = ∅, and let us pick K1 ∈ {Ki }. Since any x ∈ K1
is not in ∩i Ki , there is an index i such that x ∈ Kic , and we conclude that we have:
[
K1 ⊂ Kic
i̸=1

But this can be regarded as being an open cover of K1 , that we know to be compact,
so we can extract from it a finite subcover, as follows:
K1 ⊂ Kic1 ∪ . . . ∪ Kicn

Now observe that this latter subcover tells us that we have:

K 1 ∩ K i1 ∩ . . . ∩ K in = ∅
But this contradicts our intersection assumption in the statement, and we are done.
(2) This is a particular case of (1), proved above.
(3) We prove this by contradiction. So, assume that E has no limit point in K. This
means that any p ∈ K can be isolated from the rest of E by a certain open ball Vp = Bp (r),
and in both the cases that can appear, p ∈ E or p ∈/ E, we have:
|Vp ∩ E| = 0, 1
Now observe that these sets Vp form an open cover of K, and so of E. But due to
|Vp ∩ E| = 0, 1 and to |E| = ∞, this open cover of E has no finite subcover. Thus the
same cover, regarded now as cover of K, has no finite subcover either, contradiction.
(4) This follows from (3) that we just proved, with E = {xn }.
(5) This is a reformulation of (4), that we just proved. □
 10B. COMPACT SETS 235

Getting now to some more exciting theory, here is a key result about compactness,
which is less trivial, and that we will need on a regular basis, in what follows:
Theorem 10.15. For a subset K ⊂ RN , the following are equivalent:
(1) K is closed and bounded.
(2) K is compact.
(3) Any infinite subset E ⊂ K has a limiting point in K.
Proof. This is something quite tricky, the idea being as follows:
(1) =⇒ (2) As a first task, in order to establish this implication, let us prove that
any product of closed intervals, as follows, is indeed compact:
N
Y
J= [ai , bi ] ⊂ RN
i=1

We can assume by linearity that we are dealing with the unit cube:
N
Y
C1 = [0, 1] ⊂ RN
i=1

In order to prove that C1 is compact, we proceed by contradiction. So, assume that

we have an open cover as follows, having no finite subcover:
[
C1 ⊂ Ei
i

Now let us cut C1 into 2N small cubes, in the obvious way, over the N coordinate axes.
Then at least one of these small cubes, which are all covered by ∪i Ei too, has no finite
subcover. So, let us call C2 ⊂ C1 one of these small cubes, having no finite subcover:
[
C2 ⊂ Ei
i
N
We can then cut C2 into 2 small cubes, and by the same reasoning, we obtain a
smaller cube C3 ⊂ C2 having no finite subcover. And so on by recurrence, and we end up
with a decreasing sequence of cubes, as follows, having no finite subcover:
C1 ⊃ C2 ⊃ C3 ⊃ . . .
Now since these decreasing cubes have edge size 1, 1/2, 1/4, . . . , their intersection must
be a point. So, let us call p this point, defined by the following formula:
\
{p} = Ck
k

But this point p must be covered by ∪i Ei , so we can find an index i such that:
p ∈ Ei
236 10. PARTIAL DERIVATIVES

Now observe that Ei must contain a whole ball around p, and so starting from a
certain K ∈ N, all the cubes Ck will be contained in this ball, and so in Ei :
Ck ⊂ Ei , ∀k ≥ K
But this is a contradiction, because CK , and in fact the smaller cubes Ck with k > K
as well, were assumed to have no finite subcover. Thus, we have proved our claim.
(1) =⇒ (2), continuation. But with this claim in hand, the result is now clear.
Indeed, assume that K ⊂ RN is closed and bounded. Then, since K is bounded, we can
view it as a subset as a suitable big cube, of the following form:
N
Y
K⊂ [−M, M ] ⊂ RN
i=1

But, what we have here is a closed subset inside a compact set, that follows to be
compact, as desired.
(2) =⇒ (3) This is something that we already know, not needing K ⊂ RN .
(3) =⇒ (1) We have to prove that K as in the statement is both closed and bounded,
and we will do both these things by contradiction, as follows:
– Assume first that K is not closed. But this means that we can find a point x ∈
/K
which is a limiting point of K. Now let us pick xn ∈ K, with xn → x, and consider the
set E = {xn }. According to our assumption, E must have a limiting point in K. But this
limiting point can only be x, which is not in K, contradiction.
– Assume now that K is not bounded. But this means that we can find points xn ∈ K
satisfying ||xn || → ∞, and if we consider the set E = {xn }, then again this set must have
a limiting point in K, which is impossible, so we have our contradiction, as desired. □
So long for compactness. As a last piece of general topology, in our metric space
framework, we can talk as well about connectedness, as follows:
Definition 10.16. We can talk about connected sets E ⊂ X, as follows:
(1) We say that E is connected if it cannot be separated as E = E1 ∪ E2 , with the
components E1 , E2 satisfying E1 ∩ Ē2 = Ē1 ∩ E2 = ∅.
(2) We say that E is path connected if any two points p, q ∈ E can be joined by a
path, meaning a continuous f : [0, 1] → X, with f (0) = p, f (1) = q.
All this looks a bit technical, and indeed it is. To start with, (1) is something quite
natural, but the separation condition there E1 ∩ Ē2 = Ē1 ∩ E2 = ∅ can be weakened into
E1 ∩ E2 = ∅, or strengthened into Ē1 ∩ Ē2 = ∅, depeding on purposes, and with our (1)
as formulated above being the good compromise, for most purposes.
 10B. COMPACT SETS 237

As for (2), this condition is obviously something stronger, and we have in fact the
following implications, which are both clear:
convex =⇒ path connected =⇒ connected
To be more precise, the first implication is clear, by taking as path as in Definition
10.16 (2) the segment joining p, q, and for the second implication, the idea is that a
separation of E as in Definition 10.16 (1) will produce a separation of the path joining
p, q, and so ultimately, a separation of the interval [0, 1], which is impossible.

The problem, however, is that connected does not imply path connected, and there
are as well various counterexamples in relation with the various versions of (1) that can
be formulated, as explained above. Anyway, leaving aside the discussion here, which is
something quite technical, once all these questions clarified, the idea is that any set E
can be written as a disjoint union of connected components, as follows:
G
E= Ei
i

However, the story is not over here, because when looking at the connected components
Ei , or simply at the connected sets E, if you prefer, there are many things that can happen,
in relation with the “holes” that E can have or not. Thus, the classification of connected
sets runs into the question of deciding how many holes can have such a set E, and this is
something quite subtle, that we will discuss in chapter 11 below.

Getting back now to more concrete things, remember that we are here in this book
for studying functions, and doing calculus. And, regarding functions, we have:
Theorem 10.17. Assuming that f : X → Y is continuous, the following happen:
(1) If O is open, then f −1 (O) is open.
(2) If C is closed, then f −1 (C) is closed.
(3) If K is compact, then f (K) is compact.
(4) If E is connected, then f (E) is connected.
Proof. This is something fundamental, which can be proved as follows:
(1) This is clear from the definition of continuity, written with ε, δ. In fact, the converse
holds too, in the sense that if f −1 (open) = open, then f must be continuous.
(2) This follows from (1), by taking complements. And again, the converse holds too,
in the sense that if f −1 (closed) = closed, then f must be continuous.
(3) This is something that took us some time to prove for the functions f : R → R,
earlier in this book, with our compactness technology there, but which is now clear,
by using our definition of compactness with open covers. Indeed, given an open cover
f (K) ⊂ ∪i Ei , we have by using (1) an open cover K ⊂ ∪i f −1 (Ei ), and so by compactness
238 10. PARTIAL DERIVATIVES

of K, a finite subcover K ⊂ f −1 (Ei1 ) ∪ . . . ∪ f −1 (Ein ), and so finally a finite subcover

f (K) ⊂ Ei1 ∪ . . . ∪ Ein , as desired. And isn’t this smart, hope you agree with me.
(4) This can be proved via the same trick as for (3). Indeed, any separation of f (E) into
two parts can be returned via f −1 into a separation of E into two parts, contradiction. □
As a comment here, Theorem 10.17 generalizes, and in a clever way, many things that
we know from one-variable calculus. Of particular interest is (3), which shows in particular
that any continuous function on a compact space f : X → R attains its minimum and its
maximum, and then (4), which can be regarded as being a general mean value theorem.
As for (1) and (2), these are useful in everyday life, and we will see examples of this.

In addition to this, all useful things, there are a few other things that can be said
about the continuity of the functions f : RN → RM . We will be back to this.

10c. Partial derivatives

Let us discuss now the differentiability in several variables. At order 1, the situation
is quite similar to the one in 1 variable, but this time involving matrices. In order to
explain this material, let us start with a straightforward definition, as follows:
Definition 10.18. We say that a map f : RN → RM is differentiable at x ∈ RN if
f (x + t) ≃ f (x) + f ′ (x)t
for some linear map f ′ (x) : RN → RM , called derivative of f at the point x ∈ RN .
But is this the correct definition. I can hear you screaming that we are probably going
the wrong way, because for functions f : R → R the derivative is something much simpler,
as follows, and that we should try to imitate, in our higher dimensional setting:
f (x + t) − f (x)
f ′ (x) = lim
t→0 t
However, this is not possible, for a number of reasons, that are worth discussing in
detail. So, here is the discussion, answering all kinds of questions that you might have:

(1) First of all, the above formula does not make any sense for a function f : RN → RM
with N ̸= M , because we cannot divide oranges by apples. And it doesn’t make sense
either at N = M ∈ N, because, well, here we have RN oranges, I agree with you, but
there is no way of dividing these oranges, unless we are in the special cases N = 1, 2.

(2) More philosophically know, we have seen that having f ′ (x) defined as a number is
difficult, but the question is, do we really want to have f ′ (x) defined as a number? And
my claim here is that, this would be a pity. Think at the case where f : RN → RM is
linear. Such a map is just “perfect”, and so should equal its own derivative, f = f ′ .
 10C. PARTIAL DERIVATIVES 239

(3) Summarizing, our Definition 10.18 is just perfection, and is waiting for some further
study, and this is what we will do. And in case you’re still secretly dreaming about having
f ′ (x) defined as some sort of number, wait for it. When N = M at least, there is indeed
a lucky number, namely det(f ′ (x)), called Jacobian, but more on this later.
Getting back now to Definition 10.18 as formulated, and agreed upon, we have there
a linear map f ′ (x) : RN → RM , waiting to be further understood. So, time now to use
our linear algebra knowledge from chapter 9. We know from there that such linear maps
correspond to rectangular matrices A ∈ MM ×N (R), and we are led in this way to:
Question 10.19. Given a differentiable map f : RN → RM , in the abstract sense of
Definition 10.18, what exactly is its derivative
f ′ (x) : RN → RM
regarded as a rectangular matrix, f ′ (x) ∈ MM ×N (R)?
Again, I might hear scream you here, arguing that you come after a long battle, just
agreeing that the derivative is a linear map, and not a number, and now what, we are
trying to replace this linear map by a matrix, and so by a bunch of numbers.
Good point, and I must admit that I have no good answer to this. In fact, what we
are doing here, namely Definition 10.18, then Question 10.19, and finally Theorem 10.20
to follow in a moment, are quite deep things, that took mankind several centuries to
develop, and that we are now presenting in a compressed form. So yes, all this is difficult
mathematics, when you first see it, I perfectly agree with you.
In any case, hope that you’re still with me, and in order to further clarify all this, here
is the answer to Question 10.19:
Theorem 10.20. The derivative of a differentiable function f : RN → RM , making
the approximation formula
f (x + t) ≃ f (x) + f ′ (x)t
work, is the matrix of partial derivatives at x, namely
 
′ dfi
f (x) = (x) ∈ MM ×N (R)
dxj ij

acting on the vectors t ∈ RN by usual multiplication.

Proof. As a first observation, the formula in the statement makes sense indeed, as
an equality, or rather approximation, of vectors in RM , as follows:
   df1 (x) . . . df1 (x)  
x1 + t1 x1 t1
 
dx1 dxN
f .
.. .
 ≃ f  ..  +  ..
 . .
..   ... 

dfM dfM
xN + tN xN dx1
(x) ... dxN
(x) tN
240 10. PARTIAL DERIVATIVES

In order to prove now this formula, which does make sense, the idea is as follows:

(1) First of all, at N = M = 1 what we have is a usual 1-variable function f : R → R,

and the formula in the statement is something that we know well, namely:

f (x + t) ≃ f (x) + f ′ (x)t

(2) Let us discuss now the case N = 2, M = 1. Here what we have is a function
f : R2 → R, and by using twice the basic approximation result from (1), we obtain:
   
x1 + t1 x1 + t 1 df
f ≃ f + (x)t2
x2 + t2 x2 dx2
 
x1 df df
≃ f + (x)t1 + (x)t2
x2 dx1 dx2
    t 
x1 df df 1
= f + dx1 (x) dx2 (x)
x2 t2

(3) More generally, we can deal in this way with the general case M = 1, with the
formula here, obtained via a straightforward recurrence, being as follows:

x1 + t1 x1
   

f .
.. ≃ f ..  + df (x)t + . . . + df (x)t
   . dx1
1
dxN
N
xN + tN xN
x1  t1
   

= f  ...  + dx (x) . . . dxdfN (x)  ... 


df
1
xN tN

(4) But this gives the result in the case where both N, M ∈ N are arbitrary too.
Indeed, consider a function f : RN → RM , and let us write it as follows:

f1
 

f =  ... 
fM

We can apply (3) to each of the components fi : RN → R, and we get:

x1 + t1 x1  t1
     
. ..  + dfi (x) . . . (x)  ... 

dfi
fi  ..  ≃ fi  . dx1 dxN
xN + tN xN tN
 10C. PARTIAL DERIVATIVES 241

But this collection of M formulae tells us precisely that the following happens, as an
equality, or rather approximation, of vectors in RM :
   df1 (x) . . . df1 (x)  
x1 + t1 x1 t1
 
dx1 dxN
f .
.. .
 ≃ f  ..  +  ..
 . .
..   ... 

dfM dfM
xN + tN xN dx1
(x) . . . dx (x) tN
N

Thus, we are led to the conclusion in the statement. □

The above result, while being very nice, clear and useful, does not close the discus-
sion regarding the derivative. Indeed, conversely, we would like to know if the existence
of partial derivatives guarantees the fact that f is differentiable, with derivative f ′ (x)
appearing as the rectangular matrix fomed by the partial derivatives at x.

So, let us discuss now this remaining theoretical question. The result here, which is
something quite technical, but which can be useful in practice, is as follows:
Theorem 10.21. For a function f : X → RM , with X ⊂ RN , the following conditions
are equivalent, and in this case we say that f is continuously differentiable:
(1) f is differentiable, and the map x → f ′ (x) is continuous.
(2) f has partial derivatives, which are continuous with respect to x ∈ X.
If these conditions are satisfied, f ′ (x) is the matrix fomed by the partial derivatives at x.
Proof. We already know, from Theorem 10.20, that the last assertion holds. Re-
garding now the proof of the equivalence, this goes as follows:
(1) =⇒ (2) Assuming that f is differentiable, we know from Theorem 10.20 that
′
f (x) is the matrix fomed by the partial derivatives at x. Thus, for any x, y ∈ X:
dfi dfi
(x) − (y) = f ′ (x)ij − f ′ (y)ij
dxj dxj
By applying now the absolute value, we obtain from this the following estimate:
dfi dfi
(x) − (y) = |f ′ (x)ij − f ′ (y)ij |
dxj dxj
= |(f ′ (x) − f ′ (y))ij |
≤ ||f ′ (x) − f ′ (y)||
But this gives the result, because if the map x → f ′ (x) is assumed to be continuous,
then the partial derivatives follow to be continuous with respect to x ∈ X.
(2) =⇒ (1) This is something more technical. For simplicity, let us assume M = 1,
the proof in general being similar. Given x ∈ X and ε > 0, let us pick r > 0 such that
242 10. PARTIAL DERIVATIVES

the ball B = Bx (r) belongs to X, and such that the following happens, over B:

df df ε
(x) − (y) <
dxj dxj N

Our claim is that, with this choice made, we have the following estimate, for any
t ∈ RN satisfying ||t|| < r, with A being the vector of partial derivatives at x:

|f (x + t) − f (x) − At| ≤ ε||t||

In order to prove this claim, the idea will be that of suitably applying the mean value
theorem, over the N directions of RN . Indeed, consider the following vectors:
 
t1
 .. 
.
t 
 
t(k) =  k
0
.
 .. 
0

In terms of these vectors, we have the following formula:

N
X
f (x + t) − f (x) = f (x + t(j) ) − f (x + t(j−1) )
j=1

Also, the mean value theorem gives a formula as follows, with sj ∈ [0, 1]:

df
f (x + t(j) ) − f (x + t(j−1) ) = (x + sj t(j) + (1 − sj )t(j−1) ) · tj
dxj

But, according to our assumption on r > 0 from the beginning, the derivative on the
df
right differs from dx j
(x) by something which is smaller than ε/N :

df df ε
(x + sj t(j) + (1 − sj )t(j−1) ) − (x) <
dxj dxj N
 10C. PARTIAL DERIVATIVES 243

Now by putting everything together, we obtain the following estimate:

N
X df
|f (x + t) − f (x) − At| = f (x + t(j) ) − f (x + t(j−1) ) − (x) · tj
j=1
dxj
N
X df
≤ f (x + t(j) ) − f (x + t(j−1) ) − (x) · tj
j=1
dxj
N
X df df
= (x + sj t(j) + (1 − sj )t(j−1) ) · tj − (x) · tj
j=1
dx j dxj

N
X df df
= (x + sj t(j) + (1 − sj )t(j−1) ) − (x) · |tj |
j=1
dxj dxj
N
X ε
≤ · |tj |
j=1
N
≤ ε||t||

Thus we have proved our claim, and this gives the result. □

This was for the basic theory of partial derivatives. In practice, there are far more
things that can be said, both at the abstract and the concrete level, including of course
many examples. We will be back to this, after developing some more general theory.

Before getting into this, however, let us formulate a definition that you will certainly
appreciate, bringing a bit of humanity, and more specifically a good old real number, in
this world of vectors, matrices and other beasts which is multivariable calculus:

Definition 10.22. Given a differentiable function f : RN → RN , its Jacobian at

x ∈ RN is the number
det(f ′ (x)) ∈ R

measuring the infinitesimal rate of the volume inflation by f , at the point x.

Here the first part is standard, because when N = M , as above, the derivative is a
linear map f ′ (x) : RN → RN , which is the same as a square matrix f ′ (x) ∈ MN (R), and
so we can consider the determinant of this matrix, det(f ′ (x)) ∈ R. As for the second
part, this comes from our fine knowledge of the determinant, from chapter 9.
244 10. PARTIAL DERIVATIVES

All this is very nice, and as a first observation, according to our formula of f ′ (x) as
being the matrix formed by the partial derivatives, we have:
df1 df1
dx1
(x) ... dxN
(x)
det(f ′ (x)) = .. ..
. .
dfN dfN
dx1
(x) . . . dxN
(x)
Thus, the Jacobian can be explicitly computed. However, in what regards the practical
uses of the Jacobian, these are quite complicated, and this will have to wait a bit, until
chapter 13 below. So, sorry for this, not yet time to enjoy Definition 10.22, and stay with
me, plenty of further linear algebra, and matrices instead of numbers, to follow.

10d. The chain rule

Generally speaking, Theorem 10.20 is what you need to know for upgrading from
calculus to multivariable calculus, of course with some help from linear algebra as well.
This is the general principle of multivariable calculus, that we will explore now.

So, let us first recall that we know from Theorem 10.20 that the derivative of a
differentiable function f : RN → RM , making the approximation formula f (x + t) ≃
f (x) + f ′ (x)t work, is the matrix of partial derivatives at x, namely:
 
′ dfi
f (x) = (x) ∈ MM ×N (R)
dxj ij

In practice, unless we are in the special case N = M , where our derivative is a

square matrix, f ′ (x) ∈ MN (R), and more can be said, as briefly explained at the end
of the previous section, we can always write our function as follows, with each of the
components f1 , . . . , fM being a scalar-valued function, fi : RN → R:
f1
 

f= .. 
 .
fM
With this convention, the derivative, as a rectangular matrix, is given by:
f1 (x)
 

f ′ (x) =  ... 
fM (x)
Moreover, assuming that the formula fi (x + t) ≃ fi (x) + fi′ (x)t works for any i, it
follows that the formula f (x + t) ≃ f (x) + f ′ (x)t works too. Thus, as a conclusion to this,
Theorem 10.20 is more of less equivalent to its M = 1 particular case.

So, let us record this M = 1 particular case of Theorem 10.20, as follows:

 10D. THE CHAIN RULE 245

Theorem 10.23. The derivative of a differentiable function f : RN → R, making the

approximation formula f (x + t) ≃ f (x) + f ′ (x)t work, is given by
f ′ (x) = (∇f (x))t
where ∇f , called gradient of f , is the column vector formed by the partial derivatives:
 
df
∇f (x) = (x) ∈ RN
dxj j

Equivalently, without reference to f ′ , we have the approximation formula

f (x + t) ≃ f (x)+ < ∇f (x), t >
where < , > is the usual scalar product on RN .
Proof. This is a reformulation of Theorem 10.20, in the particular case M = 1.
Indeed, we know from there that the derivative of our function f : RN → R, making the
approximation formula f (x + t) ≃ f (x) + f ′ (x)t work, is given by:
 
′ df
f (x) = (x) ∈ M1×N (R)
dxj j

But, what we have here is the horizontal vector formed by the partial derivatives at
x, which is transpose to the vertical vector formed by the same partial derivatives, which
is by definition ∇f (x). Thus, we are led to the formula in the statement, namely:
f ′ (x) = (∇f (x))t
Finally, regarding the last assertion, this follows from this, as follows:
f (x + t) ≃ f (x) + f ′ (x)t
= f (x) + (∇f (x))t t
X
= f (x) + (∇f (x))j tj
j

= f (x)+ < ∇f (x), t >

Thus, we are led to the conclusions in the statement. □
In practice, more things can be said about the gradient, which is a notion which is
quite intuitive. Also, getting back now to the general functions f : RN → RM , in the case
where we are in small dimension, N, M ≤ 3, we can do many other geometric things with
the partial derivatives. We will be back to all this, gradually, in what follows.

In general now, our first task will be that of extending to several variables our basic
results from one-variable calculus. As a standard result here, we have:
246 10. PARTIAL DERIVATIVES

Theorem 10.24. We have the chain derivative formula

(f ◦ g)′ (x) = f ′ (g(x)) · g ′ (x)
as an equality of matrices.
Proof. This is something standard in one variable, and in several variables the proof
is similar, by using the notion of derivative coming from Theorem 10.20. To be more
precise, consider a composition of functions, as follows:
f : RN → RM , g : RK → RN , f ◦ g : RK → RM
According to Theorem 10.20, the derivatives of these functions are certain linear maps,
corresponding to certain rectangular matrices, as follows:
f ′ (g(x)) ∈ MM ×N (R) , g ′ (x) ∈ MN ×K (R) (f ◦ g)′ (x) ∈ MM ×K (R)
Thus, our formula makes sense indeed. As for proof, this comes from:
(f ◦ g)(x + t) = f (g(x + t))
≃ f (g(x) + g ′ (x)t)
≃ f (g(x)) + f ′ (g(x))g ′ (x)t
Thus, we are led to the conclusion in the statement. □
The above formula is quite interesting, and working out its main particular cases,
when some of parameters M, N, K equal 1, is an instructive exercice. We typically obtain
some mysterious formulae involving the gradient from Theorem 10.23. But, the best is to
rather ignore all these particular cases, and memorize Theorem 10.24 as stated.

Along the same lines, observe also that we can as well reformulate Theorem 10.24 in
terms of partial derivatives, by performing the matrix multiplication on the right. But
again, the best is to rather ignore all this, and memorize Theorem 10.24 as stated.

This being said, nothing better than working out a few exercises, at this point of
learning. As a first result coming from Theorem 10.24, as a particular case, we have:
Proposition 10.25. We have the chain derivative formula
d(f ◦ g)i X dfi dgk
(x) = (g(x)) · (x)
dxj k
dx k dx j

as an equality of numbers.
Proof. This follows indeed from the formula in Theorem 10.24, namely:
(f ◦ g)′ (x) = f ′ (g(x)) · g ′ (x)
 10D. THE CHAIN RULE 247

Indeed, what we have here is an equality of matrices, and at the level of the individual
entries of these matrices, by performing the multiplication on the right, we obtain:
X
(f ◦ g)′ (x)ij = f ′ (g(x))ik · g ′ (x)kj
k

Thus, we are led to the formula in the statement. □

As already mentioned, there are many possible variations of the above formula, ob-
tained when some of dimension parameters M, N, K equal 1, and with the final formulae
involving tricky sums, or scalar products, whose components are partial derivatives, or
entries of the gradient. We will leave them as an instructive exercise.

Moving ahead now, we have several useful consequences of the above result, a standard
application, generalizing things that we know in one variable, being as follows:
Theorem 10.26. Assuming that f : X → RM is differentiable, with X ⊂ RN being
convex, we have the estimate
||f (x) − f (y)|| ≤ M ||x − y||
for any x, y ∈ X, where the quantity on the right is given by:
M = sup ||f ′ (x)||
x∈X

Moreover, this estimate can be sharp, for instance for the linear functions.
Proof. This is something quite tricky, which in several variables cannot be proved
with bare hands. However, we can get it by using our chain derivative formula. Consider
indeed the path γ : [0, 1] → RM given by the following formula:
γ(t) = tx + (1 − t)y
Now let us set g(t) = f (γ(t)). We have then, according to the chain rule formula:
g ′ (t) = f ′ (γ(t))γ ′ (t)
= f ′ (γ(t))(x − y)
But this gives the following estimate, with M > 0 being as in the statement:
|g ′ (t)| ≤ ||f ′ (γ(t))|| · ||x − y||
≤ M ||x − y||
Now by using one-variable results that we know, we obtain from this:
||g(1) − g(0)|| ≤ ||M || · ||x − y||
But since we have g(1) = f (x), g(0) = f (y), this gives the formula in the statement.
Finally, the last assertion is clear. □
248 10. PARTIAL DERIVATIVES

As a conclusion to all this, we have extended to the case of vector functions most of
what we know about the one-variable functions, at the general level, of basic calculus.
Still waiting to be discussed are the higher derivatives, and their various applications, and
we will come back to this in chapter 12, after a short geometric break, and then of course
the theory of integration, that we will discuss in chapters 13-16.
10e. Exercises
This was a quite straightforward chapter, assuming that you have understood well
Part I and chapter 9, and as exercises here, mostly straightforward, we have:
Exercise 10.27. Write a short essay about the basic theory of sequences and contin-
uous functions over metric spaces, following the material from Part I.
Exercise 10.28. Clarify the relation between the various notions of connectedness,
with the goal of decomposing each metric space into connected components.
Exercise 10.29. Enjoy defining the derivative of a function f : RN → RN as being a
number, namely its Jacobian. What can you do, and what not, with this approach?
Exercise 10.30. Look up the internet for various other formulations of the chain rule
for derivatives, and prove them all, quickly, by using what we know.
As bonus exercise, think a bit about higher derivatives, and what can you do with
them, following the material from Part I. We will be back to this, after a break.
 CHAPTER 11

Some geometry

11a. Equations, conics

As an application of the theory of partial derivatives that we developed, and of calculus
in general, as we know it at this point, we can now talk about some beautiful things,
namely geometry and physics. As a starting point, you have surely noticed that the Sun
moves around the Earth on a circle. However, when carefully measuring it, this circle is
not exactly a circle, but rather an ellipsis. Also, some further possible trajectories, of one
object with respect to another, due to gravity, include parabolas and hyperbolas, that
you can observe with a telescope, by looking at asteroids and comets.

So, before even starting to look at the equations of gravity, and having some fun in
solving them, we need a mathematical theory of curves like ellipses, parabolas and hyper-
bolas, which are what we can expect to find, as trajectories, from gravity computations.
And, good news, this theory exists, since the ancient Greeks. Let us start with:
Definition 11.1. A conic is a plane algebraic curve of the form
n o
C = (x, y) ∈ R2 P (x, y) = 0

with P ∈ R[x, y] being of degree ≤ 2.

As basic examples of conics, we have the ellipses, parabolas and hyperbolas. The
simplest examples of these are as follows, with the ellipsis actually being a circle:
x2 + y 2 = 1 , x2 = y , xy = 1
Observe that, due to our assumption deg P ≤ 2, we have as conics some degenerate
curves as well, such as lines, ∅, and R2 itself, coming from deg P ≤ 1, as follows:
x=0 , 1=0 , 0=0
This might suggest to replace our assumption deg P ≤ 2 by deg P = 2, but we will
not do so, because deg P = 2 does not rule out degenerate situations, such as:
x2 + y 2 = −1 , x2 + y 2 = 0 , x2 = 0 , xy = 0
In fact, what we get here are ∅, points, lines, and pairs of lines, so in the end, assuming
deg P = 2 instead of deg P ≤ 2 would only rule out R2 itself, which is not worth it.
249
250 11. SOME GEOMETRY

Summarizing, our notion of conic from Definition 11.1 looks quite reasonable, so let
us agree on this notion. Getting now to classification matters, we first have:
Theorem 11.2. Up to non-degenerate linear transformations of the plane,
   
x x
→A
y y
with det A ̸= 0, the conics fall into two classes, as follows:
(1) Non-degenerate: circles, parabolas, hyperbolas.
(2) Degenerate: ∅, points, lines, pairs of lines, R2 .
Proof. As a first observation, looks like we forgot the ellipses, but via linear trans-
formations these become circles, so things fine. As for the proof, this goes as follows:
(1) Consider an arbitrary conic, written as follows, with a, b, c, d, e, f ∈ R:
ax2 + by 2 + cxy + dx + ey + f = 0
(2) Assume first a ̸= 0. By making a square out of ax2 , up to a linear transformation
in (x, y), we can get rid of the term cxy, and we are left with:
ax2 + by 2 + dx + ey + f = 0
In the case b ̸= 0 we can make two obvious squares, and again up to a linear transfor-
mation in (x, y), we are left with an equation as follows:
x2 ± y 2 = k
In the case of positive sign, x2 + y 2 = k, the solutions are the circle, when k ≥ 0, the
point, when k = 0, and ∅, when k < 0. As for the case of negative sign, x2 − y 2 = k,
which reads (x − y)(x + y) = k, here once again by linearity our equation becomes xy = l,
which is a hyperbola when l ̸= 0, and two lines when l = 0.
(3) In the case b ̸= 0 the study is similar, with the same solutions, so we are left with
the case a = b = 0. Here our conic is as follows, with c, d, e, f ∈ R:
cxy + dx + ey + f = 0
If c ̸= 0, by linearity our equation becomes xy = l, which produces a hyperbola or two
lines, as explained before. As for the remaining case, c = 0, here our equation is:
dx + ey + f = 0
But this is generically the equation of a line, unless we are in the case d = e = 0,
where our equation is f = 0, having as solutions ∅ when f ̸= 0, and R2 when f = 0.
(4) So, this was the study of an arbitrary conic, and by putting now everything
together, we are led to the conclusions in the statement. □
 11A. EQUATIONS, CONICS 251

In order now to plainly classify the conics, without reference to a linear transformation
of the plane, we just need to apply linear transformations to the curves that we found in
Theorem 11.2. This leads to the following classification result:
Theorem 11.3. The conics fall into two classes, as follows:
(1) Non-degenerate: ellipses, parabolas, hyperbolas.
(2) Degenerate: ∅, points, lines, pairs of lines, R2 .
Also, the compact conics are ∅, the points, and the ellipses.
Proof. We have several assertions here, the idea being as follows:
(1) As said above, in order to get to such a classification result, we just need to apply
linear transformations to the curves that we found in Theorem 11.2. But this leaves the
list there unchanged, up to the circles becoming ellipses, as stated above.
(2) In what regards the last assertion, this is clear from the first one, but since this
assertion is quite interesting, let us give it a quick, independent proof as well. Consider
an arbitary conic, written as follows, with a, b, c, d, e, f ∈ R:
ax2 + by 2 + cxy + dx + ey + f = 0
Compacity rules then out the case c ̸= 0, and our conic must be in fact:
ax2 + by 2 + dx + ey + f = 0
But then with a, b ̸= 0 we must have by compacity a, b > 0 or a, b < 0, and we get an
ellipsis, then with a = 0, b ̸= 0 or a ̸= 0, b = 0 we get by compacity either ∅ or a point,
and finally with a = b = 0 the compacity rules out again everything, except for ∅. □

As a third main result now on the conics, also known since the ancient Greeks, and
which justifies the name “conics”, coming from “cone”, we have:
Theorem 11.4. Up to some degenerate cases, the conics are exactly the curves which
appear by cutting a 2-sided cone with a plane.
Proof. This is something quite tricky, the idea being as follows:
(1) By suitably choosing our coordinate axes (x, y, z), we can assume that our 2-sided
cone is given by an equation as follows, with k > 0:
x2 + y 2 = kz 2
In order to prove the result, we must intersect this cone with an arbitrary plane, which
has an equation as follows, with (a, b, c) ̸= (0, 0, 0):
ax + by + cz = d
252 11. SOME GEOMETRY

(2) However, before getting into computations, observe that what we want to find is a
certain degree 2 equation in the above plane, for the intersection. Thus, it is convenient
to change the coordinates, as for our plane to be given by the following equation:
z=0
(3) But with this done, what we have to do is to see how the cone equation x2 +y 2 = kz 2
changes, under this change of coordinates, and then set z = 0, as to get the (x, y) equation
of the intersection. But this leads, via some thinking or computations, to the conclusion
that the cone equation x2 + y 2 = kz 2 becomes in this way a degree 2 equation in (x, y),
which can be arbitrary, and so to the final conclusion in the statement.
(4) Alternatively, and perhaps more concretely, we can use the original coordinates,
with the cone being x2 + y 2 = kz 2 , and compute the intersection, with the conclusion
that what we get, depending on the slope of the cone, and modulo degenerate cases, is
an ellipsis, hyperbola or parabola. So, by invoking Theorem 11.3, we obtain the result.
(5) Summarizing, we have proved the result, modulo some details and interesting
computations which are left to you, reader. Left to you as well is the full discussion
concerning degree 2 curve degeneracy vs cone cutting degeneracy, with the remark that
in what regards the cone cuts, the degenerate cases are very easy to identity and list, with
the list consisting of ∅, the points, the lines, the pairs of lines, and R2 itself. □
All this is very nice, and as a conclusion to what we have so far about conics, we have
the following statement, containing all the needed essentials:
Theorem 11.5. The conics, which are the algebraic curves of degree 2 in the plane,
n o
2
C = (x, y) ∈ R P (x, y) = 0
with deg P ≤ 2, appear modulo degeneration by cutting a 2-sided cone with a plane, and
can be classified into ellipses, parabolas and hyperbolas.
Proof. This follows indeed by putting together the above results, and with the dis-
cussion concerning degeneration being left, as usual, as an instructive exercise. □
Moving ahead now, the most interesting conics, which are both compact and non-
degenerate, are the ellipses. So, let us study them more in detail. As a starting point, we
have the following statement, summarizing our knowledge about ellipses:
Theorem 11.6. The compact non-degenerate conics are the ellipses, which can be
written, modulo rotations and translations in the plane, as
 x 2  y 2
+ =1
a b
with a, b > 0 being half the size of a box containing the ellipsis. These ellipses also appear
by compactly cutting a cone with a plane. The area of such an ellipsis is A = πab.
 11A. EQUATIONS, CONICS 253

Proof. In this statement most of the mathematics is from above, and with our ex-
planations regarding the parameters a, b > 0 coming from the following picture:

•b

•−a •a

•−b

As for the formula A = πab, this comes from a computation from chapter 4, namely:
Z a r
x2
A = 2 b 1 − 2 dx
−a a
Z a√
4b
= a2 − x2 dx
a 0
Z 1p
= 4ab 1 − y 2 dy
0
π
= 4ab ·
4
= πab
Finally, as a verification, for a = b = 1 we get A = π, as we should. □

The above result is not the end of the story with ellipses, because we have as well, as
a complement to it, or even as a rival result, which is just fine on its own:
Theorem 11.7. The ellipses appear via equations of the following type, with p, q being
two points in the plane, and with l ≥ d(p, q) being a certain length:
d(z, p) + d(z, q) = l
2 2
√ + (y/b) = 1 with a ≥ b ≥ 0, the focal points
For an ellipsis parametrized as before, (x/a)
are p = (0, −r) and q = (0, r), with r = a2 − b2 , and the length is l = 2a.
Proof. As already mentioned, it is possible to take d(z, p)+d(z, q) = l as a definition
for the ellipses, which is nice because all you need for drawing such an ellipsis is a string
and a pencil, and then work out all the theory starting from this. In what concerns us,
we will rather further build on what we know from Theorem 11.6, as follows:
254 11. SOME GEOMETRY

(1) After some routine thinking, in order to fully prove the result, what we have to do
is to take an ellipsis as parametrized in Theorem 11.6, and look for the focal points:

•b

•−a •−r •r •a

•−b

To be more precise, we are looking for a number r > 0, and a number l > 0, such that
our ellipsis appears as d(z, p) + d(z, q) = l, with p = (0, −r) and q = (0, r).

(2) Let us first compute these numbers r, l > 0. Assuming that our result holds indeed
as stated, by taking z = (0, a), we see that the length l is:

l = (a − r) + (a + r) = 2a

As for the parameter r, by taking z = (b, 0), we conclude that we must have:
√ √
2 b2 + r2 = 2a =⇒ r = a2 − b2

(3) With these observations made, let us prove the result. Given l, r > 0, and setting
p = (0, −r) and q = (0, r), we have the following computation, with z = (x, y):

d(z, p) + d(z, q) = l
p p
⇐⇒ (x + r)2 + y 2 + (x − r)2 + y 2 = l
p p
⇐⇒ (x + r)2 + y 2 = l − (x − r)2 + y 2
p
⇐⇒ (x + r)2 + y 2 = (x − r)2 + y 2 + l2 − 2l (x − r)2 + y 2
p
⇐⇒ 2l (x − r)2 + y 2 = l2 − 4xr
⇐⇒ 4l2 (x2 + r2 − 2xr + y 2 ) = l4 + 16x2 r2 − 8l2 xr
⇐⇒ 4l2 x2 + 4l2 r2 + 4l2 y 2 = l4 + 16x2 r2
⇐⇒ (4x2 − l2 )(4r2 − l2 ) = 4l2 y 2
 11B. KEPLER AND NEWTON 255

(4) Now observe that we can further process the equation that we found as follows:
4x2 − l2 4y 2
(4x2 − l2 )(4r2 − l2 ) = 4l2 y 2 ⇐⇒ =
l2 4r2 − l2
2 2
4x − l y2
⇐⇒ =
l2 r2 − l2 /4
!2
 x 2 y
⇐⇒ −1= p
2l r2 − l2 /4
!2
 x 2 y
⇐⇒ + p =1
2l r − l2 /4
2

(5) Thus, our result holds indeed, and with√the numbers l, r > 0 appearing, and no
surprise here, via the formulae l = 2a and r = a2 − b2 , found in (2) above. □

The above results, which are old as modern mathematics itself, are foundational for
both algebraic and differential geometry. We will be back to them later, after doing some
physics, following Kepler and Newton, making the link with calculus.

11b. Kepler and Newton

As a continuation of the above, or rather as a complement, let us do some physics.
Theorem 11.5 was the foundational result of modern mathematics, and by a remarkable
twist of fate, the foundational result of physics is something related to it, as follows:
Theorem 11.8. Planets and other celestial bodies move around the Sun on conics,
n o
2
C = (x, y) ∈ R P (x, y) = 0

with P ∈ R[x, y] being of degree 2, which can be ellipses, parabolas or hyperbolas.

Proof. This is something quite long, due to Kepler and Newton, but no fear, we
know calculus, and therefore what can resist us. The proof goes as follows:
(1) According to observations and calculations performed over the centuries, since the
ancient times, and first formalized by Newton, following some groundbreaking work of
Kepler, the force of attraction between two bodies of masses M, m is given by:
Mm
||F || = G ·
d2
Here d is the distance between the two bodies, and G ≃ 6.674 × 10−11 is a constant.
Now assuming that M is fixed at 0 ∈ R3 , the force exterted on m positioned at x ∈ R3 ,
256 11. SOME GEOMETRY

regarded as a vector F ∈ R3 , is given by the following formula:

x
F = −||F || ·
||x||
GM m x
= − ·
||x||2 ||x||
GM mx
= −
||x||3
But F = ma = mẍ, with a = ẍ being the acceleration, second derivative of the
position, so the equation of motion of m, assuming that M is fixed at 0, is:
GM x
ẍ = −
||x||3
Obviously, the problem happens in 2 dimensions, and you can even find, as an exercise,
a formal proof of that, based on the above equation, if you really want to. But here the
most convenient is to use standard x, y coordinates, and denote our point as z = (x, y).
With this change made, and by setting K = GM , the equation of motion becomes:
Kz
z̈ = −
||z||3
(2) The idea now is that the problem can be solved via some calculus. Let us write
indeed our vector z = (x, y) in polar coordinates, as follows:
x = r cos θ , y = r sin θ
We have then ||z|| = r, and our equation of motion becomes:
Kz
z̈ = − 3
r
Let us differentiate now x, y. By using the standard calculus rules, we have:
ẋ = ṙ cos θ − r sin θ · θ̇
ẏ = ṙ sin θ + r cos θ · θ̇
Differentiating one more time gives the following formulae:
ẍ = r̈ cos θ − 2ṙ sin θ · θ̇ − r cos θ · θ̇2 − r sin θ · θ̈
ÿ = r̈ sin θ + 2ṙ cos θ · θ̇ − r sin θ · θ̇2 + r cos θ · θ̈
Consider now the following two quantities, appearing as coefficients in the above:
a = r̈ − rθ̇2 , b = 2ṙθ̇ + rθ̈
In terms of these quantities, our second derivative formulae read:
ẍ = a cos θ − b sin θ
ÿ = a sin θ + b cos θ
 11B. KEPLER AND NEWTON 257

(3) We can now solve the equation of motion from (1). Indeed, with the formulae that
we found for ẍ, ÿ, our equation of motion takes the following form:
K
a cos θ − b sin θ = − 2 cos θ
r
K
a sin θ + b cos θ = − 2 sin θ
r
But these two formulae can be written in the following way:
   
K K
a + 2 cos θ = b sin θ , a + 2 sin θ = −b cos θ
r r
By making now the product, and assuming that we are in a non-degenerate case,
where the angle θ varies indeed, we obtain by positivity that we must have:
K
a+ 2 =b=0
r
(4) Let us first examine the second equation, b = 0. This can be solved as follows:
b=0 ⇐⇒ 2ṙθ̇ + rθ̈ = 0
θ̈ ṙ
⇐⇒ = −2
θ̇ r
⇐⇒ (log θ̇) = (−2 log r)′
′

⇐⇒ log θ̇ = −2 log r + c
λ
⇐⇒ θ̇ = 2
r
As for the first equation the we found, namely a + K/r2 = 0, this becomes:
λ2 K
r̈ − 3 + 2 = 0
r r
As a conclusion to all this, in polar coordinates, x = r cos θ, y = r sin θ, our equations
of motion are as follows, with λ being a constant, not depending on t:
λ2 K λ
r̈ =3
− 2 , θ̇ = 2
r r r
2
Even better now, by writing K = λ /c, these equations read:
λ2 1 1
 
λ
r̈ = 2 − , θ̇ = 2
r r c r
(5) In order to study the first equation, we use a trick. Let us write:
1
r(t) =
f (θ(t))
258 11. SOME GEOMETRY

Abbreviated, and by reminding that f takes θ = θ(t) as variable, this reads:

1
r=
f
With the convention that dots mean as usual derivatives with respect to t, and that
the primes will denote derivatives with respect to θ = θ(t), we have:
f ′ θ̇ f′ λ
ṙ = − 2
= − 2
· 2 = −λf ′
f f r
By differentiating one more time with respect to t, we obtain:
′′ λ ′′λ2 ′′
r̈ = −λf θ̇ = −λf · 2 = − 2 f
r r
On the other hand, our equation for r̈ found in (4) above reads:
λ2 1 1 λ2
   
1
r̈ = 2 − = 2 f−
r r c r c
Thus, in terms of f = 1/r as above, our equation for r̈ simply reads:
1
f ′′ + f =
c
But this latter equation is elementary to solve. Indeed, both functions cos t, sin t satisfy
g” + g = 0, so any linear combination of them satisfies as well this equation. But the
solutions of f ′′ + f = 1/c being those of g ′′ + g = 0 shifted by 1/c, we obtain:
1 + ε cos θ + δ sin θ
f=
c
Now by inverting, we obtain the following formula:
c
r=
1 + ε cos θ + δ sin θ
(6) But this leads to the conclusion that the trajectory is a conic. Indeed, in terms of
the parameter θ, the formulae of the coordinates are:
c cos θ
x=
1 + ε cos θ + δ sin θ
c sin θ
y=
1 + ε cos θ + δ sin θ
Now observe that these two functions x, y satisfy the following formula:
c2 (cos2 θ + sin2 θ) c2
x2 + y 2 = =
(1 + ε cos θ + δ sin θ)2 (1 + ε cos θ + δ sin θ)2
 11B. KEPLER AND NEWTON 259

On the other hand, these two functions satisfy as well the following formula:

2
2 c2 ε cos θ + δ sin θ − (1 + ε cos θ + δ sin θ)
(εx + δy − c) =
(1 + ε cos θ + δ sin θ)2
c2
=
(1 + ε cos θ + δ sin θ)2
We conclude that our coordinates x, y satisfy the following equation:
x2 + y 2 = (εx + δy − c)2
But what we have here is an equation of a conic, and we are done. □
The above result is not the end of the story, because there is still some discussion
to be made, in relation with degeneration. There is as well a discussion concerning
normalization, because in the Kepler problem we assumed M to be fixed at 0. However,
when changing coordinates via a translation, we can obtain in this way all conics.

Finally, from a physical perspective, that of concretely solving the gravity equation,
there is a long discussion, and lots of additional formulae, regarding the trajectory and
its parameters, as functions of the initial data. Without getting into full details here, let
us record however the following result, coming as a useful version of Theorem 11.8:
Theorem 11.9. In the context of a 2-body problem, with M fixed at 0, and m starting
its movement from Ox, the equation of motion of m, namely
Kz
z̈ = −
||z||3
with K = GM , and z = (x, y), becomes in polar coordinates, x = r cos θ, y = r sin θ,
λ2 1 1
 
λ
r̈ = 2 − , θ̇ = 2
r r c r
for some λ, c ∈ R, related by λ2 = Kc. The value of r in terms of θ is given by
c
r=
1 + ε cos θ + δ sin θ
for some ε, δ ∈ R. At the level of the affine coordinates x, y, this means
c cos θ c sin θ
x= , y=
1 + ε cos θ + δ sin θ 1 + ε cos θ + δ sin θ
with θ = θ(t) being subject to θ̇ = λ2 /r, as above. Finally, we have
x2 + y 2 = (εx + δy − c)2
which is a degree 2 equation, and so the resulting trajectory is a conic.
260 11. SOME GEOMETRY

Proof. This is a sort of “best of” the formulae found in the proof of Theorem 11.8.
And in the hope of course that we have not forgotten anything. Finally, let us mention
that the simplest illustration for this is the circular motion, and for details on this, not
included in the above, we refer to the proof of Theorem 11.8. □
As a first concrete question, we would like to understand how the various parameters
appearing above, namely λ, c, ε, δ, which via some basic math can only tell us more about
the shape of the orbit, appear from the initial data. The formulae here are as follows:
Proposition 11.10. In the context of Theorem 11.9, and in polar coordinates, x =
r cos θ, y = r sin θ, the initial data is as follows, with R = r0 :
c
r0 = , θ0 = 0
1+ε
√ √
δ K Kc
ṙ0 = − √ , θ̇0 =
c R2
εK 4δK
r̈0 =2
, θ̈0 = 2
R R
The corresponding formulae for the affine coordinates
√ x, y can be deduced from this. Also,
the various motion parameters c, ε, δ and λ = Kc can be recovered from this data.
Proof. We have several assertions here, the idea being as follows:
(1) As mentioned in Theorem 11.9, the object m begins its movement on Ox. Thus
we have θ0 = 0, and from this we get the formula of r0 in the statement.
(2) Regarding the initial speed now, the formula of θ̇0 follows from:
√
λ Kc
θ̇ = 2 = 2
r r
Also, in what concerns the radial speed, the formula of ṙ0 follows from:
c(ε sin θ − δ cos θ)θ̇
ṙ =
(1 + ε cos θ + δ sin θ)2
√
c(ε sin θ − δ cos θ) Kc
= ·
c2 /r2 r2
√
K(ε sin θ − δ cos θ)
= √
c
√
(3) Regarding now the initial acceleration, by using θ̇ = Kc/r2 we find:
√
√ 2rṙ 4 Kc · ṙ
θ̈ = −2 Kc · 3 = −
r r2
 11C. ALGEBRAIC MANIFOLDS 261

In particular at t = 0 we obtain the formula in the statement, namely:

√ √ √
4 Kc · ṙ0 4 Kc δ K 4δK
θ̈0 = − 2
= 2
· √ = 2
R R c R
√
(4) Also regarding acceleration, with λ = Kc our main motion formula reads:
 
Kc 1 1
r̈ = 2 −
r r c
In particular at t = 0 we obtain the formula in the statement, namely:
 
Kc 1 1 Kc ε εK
r̈0 = 2 − = 2 · = 2
R R c R c R
(5) Finally, the last assertion is clear, and since the formulae look better anyway in
polar coordinates than in affine coordinates, we will not get into details here. □
With the above formulae in hand, which are a precious complement to Theorem 11.9,
we can do some reverse engineering at the level of parameters, and work out how various
inital speeds and accelerations lead to various types of conics. The computations here are
quite interesting, and we will leave them as an instructive exercise.

11c. Algebraic manifolds

We have seen so far the mathematical and physical meaning of the conics, which are
the simplest curves around, with in both cases, very simple answers. The continuation
of the story, however, is more complicated, because beyond conics, things ramify. A first
idea, in order to generalize the conics, is to look at zeroes of arbitrary polynomials:
P (x, y) = 0 , P ∈ R[x, y]
More generally, we can look at zeroes of polynomials in arbitrary N dimensions:
P (x1 , . . . , xN ) = 0 , P ∈ R[x1 , . . . , xN ]
Observe that, at N ≥ 3, what we have is not exactly a curve, but rather some sort
of (N − 1)-dimensional surface, called algebraic hypersurface. Due to this, in order to
have a full collection of beasts, of all possible dimensions, we must intersect such algebraic
hypersurfaces. We are led in this way to zeroes of families of polynomials, as follows:
Definition 11.11. An algebraic manifold is a space of the form
n o
N
X = (x1 , . . . , xN ) ∈ R Pi (x1 , . . . , xN ) = 0, ∀i

with Pi ∈ R[x1 , . . . , xN ] being a family of polynomials.

262 11. SOME GEOMETRY

And, good news, this is the good definition, and with the branch of mathematics
studying such manifolds being called algebraic geometry. However, it is in fact possible to
do even more generally, by looking at algebraic manifolds defined over an arbitrary field
F , by using a family of polynomials Pi ∈ F [x1 , . . . , xN ], as follows:
n o
X = (x1 , . . . , xN ) ∈ F N Pi (x1 , . . . , xN ) = 0, ∀i
These ideas are very old, going back to the old Greeks, and there are many things
that can be said about algebraic geometry, especially in its “arithmetic” version, over
arbitrary fields F , where the theory really shines, with many known advanced results in
number theory having been obtained in this way, via algebraic geometry.

Instead of pursuing with usual, affine geometry, which can quickly escalate into fairly
complicated things, let us take a look at projective geometry too, which is something fun,
and interesting, and quite often more fun and interesting than affine geometry itself.

You might have heard or not of projective geometry. In case you didn’t yet, the general
principle is that “this is the wonderland where parallel lines cross”. Which might sound
a bit crazy, and not very realistic, but take a picture of some railroad tracks, and look at
that picture. Do these parallel railroad tracks cross, on the picture? Sure they do. So,
we are certainly not into abstractions here, but rather into serious science. QED.

Mathematically now, here are some axioms, to start with:

Definition 11.12. A projective space is a space consisting of points and lines, subject
to the following conditions:
(1) Each 2 points determine a line.
(2) Each 2 lines cross, on a point.
As a basic example we have the usual projective plane PR2 , which is best seen as being
the space of lines in R3 passing through the origin. To be more precise, let us call each
of these lines in R3 passing through the origin a “point” of PR2 , and let us also call each
plane in R3 passing through the origin a “line” of PR2 . Now observe the following:

(1) Each 2 points determine a line. Indeed, 2 points in our sense means 2 lines in R3
passing through the origin, and these 2 lines obviously determine a plane in R3 passing
through the origin, namely the plane they belong to, which is a line in our sense.

(2) Each 2 lines cross, on a point. Indeed, 2 lines in our sense means 2 planes in R3
passing through the origin, and these 2 planes obviously determine a line in R3 passing
through the origin, namely their intersection, which is a point in our sense.

Thus, what we have is a projective space in the sense of Definition 11.12. More
generally now, we have the following construction, in arbitrary dimensions:
 11C. ALGEBRAIC MANIFOLDS 263

Theorem 11.13. We can define the projective space PRN −1 as being the space of lines
in RN passing through the origin, and in small dimensions:
(1) PR1 is the usual circle.
(2) PR2 is some sort of twisted sphere.
Proof. We have several assertions here, with all this being of course a bit informal,
and self-explanatory, the idea and some further details being as follows:
(1) To start with, the fact that the space PRN −1 constructed in the statement is indeed
a projective space in the sense of Definition 11.12 follows from definitions, exactly as in
the discussion preceding the statement, regarding the case N = 3.
(2) At N = 2 now, a line in R2 passing through the origin corresponds to 2 opposite
points on the unit circle T ⊂ R2 , according to the following scheme:

•
<

• •

Thus, PR1 corresponds to the upper semicircle of T, with the endpoints identified, and
so we obtain a circle, PR1 = T, according to the following scheme:

•o / •

(3) At N = 3, the space PR2 corresponds to the upper hemisphere of the sphere
SR2⊂ R3 , with the points on the equator identified via x = −x. Topologically speaking,
we can deform if we want the hemisphere into a square, with the equator becoming the
boundary of this square, and in this picture, the x = −x identification corresponds to a
264 11. SOME GEOMETRY

“identify opposite edges, with opposite orientations” folding method for the square:
◦O / ◦


◦o ◦
(4) Thus, we have our space. In order to understand now what this beast is, let us
look first at the other 3 possible methods of folding the square, which are as follows:
◦O / ◦O ◦O / ◦ ◦O / ◦O


◦ /◦ ◦ /◦ ◦o ◦
Regarding the first space, the one on the left, things here are quite simple. Indeed,
when identifying the solid edges we get a cylinder, and then when further identifying the
dotted edges, what we get is some sort of closed cylinder, which is a torus.
(5) Regarding the second space, the one in the middle, things here are more tricky.
Indeed, when identifying the solid edges we get again a cylinder, but then when further
identifying the dotted edges, we obtain some sort of “impossible” closed cylinder, called
Klein bottle. This Klein bottle obviously cannot be drawn in 3 dimensions, but with a
bit of imagination, you can see it, in its full splendor, in 4 dimensions.
(6) Finally, regarding the third space, the one on the right, we know by symmetry that
this must be the Klein bottle too. But we can see this as well via our standard folding
method, namely identifying solid edges first, and dotted edges afterwards. Indeed, we
first obtain in this way a Möbius strip, and then, well, the Klein bottle.
(7) With these preliminaries made, and getting back now to the projective space PR2 ,
we can see that this is something more complicated, of the same type, reminding the torus
and the Klein bottle. So, we will call it “sort of twisted sphere”, as in the statement, and
exercise for you to figure out how this beast looks like, in 4 dimensions. □
All this is quite exciting, and reminds childhood and primary school, but is however a
bit tiring for our neurons, guess that is pure mathematics. It is possible to come up with
some explicit formulae for the embedding PR2 ⊂ R4 , which are useful in practice, allowing
us to do some analysis over PR2 , and we will leave this as an instructive exercise.

There is some linear algebra to be done here too, by identifying the lines in RN with
the corresponding rank 1 projections, along with many other things, and we have:
 11C. ALGEBRAIC MANIFOLDS 265

Theorem 11.14. The projective space PRN −1 can be thought of as being the space of
rank 1 projections in the matrix algebra MN (R), given by
1
Px = (xi xj )ij
||x||2
by identifying the lines in RN passing through the origin with the corresponding rank 1
projections in MN (R), in the obvious way.
Proof. There are several things going on here, the idea being as follows:
(1) The main assertion is more or less clear from definitions, the point being that the
lines in RN passing through the origin are obviously in bijection with the corresponding
rank 1 projections. Thus, we obtain the interpretation of PRN −1 in the statement.
(2) Regarding now the formula of the rank 1 projections, which is a must-know, for
this, and in everyday life, consider a vector y ∈ RN . Its projection on Rx must be a
certain multiple of x, and we are led in this way to the following formula:
< y, x > 1
Px y = x= < y, x > x
< x, x > ||x||2
(3) But with this in hand, we can now compute the entries of Px , as follows:
(Px )ij = < Px ej , ei >
1
= < ej , x >< x, ei >
||x||2
xj xi
=
||x||2
Thus, we are led to the formula in the statement. □
All this is very interesting, but we will pause our study here, because we still have
many other things to say. Getting now to finite fields, we have:
Theorem 11.15. Given a field F , we can talk about the projective space PFN −1 , as
being the space of lines in F N passing through the origin. At N = 3 we have
|PF2 | = q 2 + q + 1
where q = |F |, in the case where our field F is finite.
Proof. This is indeed clear from definitions, with the cardinality coming from:
|F 3 − {0}| q3 − 1
|PF2 | = = = q2 + q + 1
|F − {0}| q−1
Thus, we are led to the conclusions in the statement. □
266 11. SOME GEOMETRY

As an example, let us see what happens for the simplest finite field that we know,
namely F = Z2 . Here our projective plane, having 4 + 2 + 1 = 7 points, and 7 lines, is a
famous combinatorial object, called Fano plane, which is depicted as follows:
•

• •
•

• • •
Here the circle in the middle is by definition a line, and with this convention, the basic
axioms in Definition 11.12 are satisfied, in the sense that any two points determine a line,
and any two lines determine a point. And isn’t this beautiful. Let us record:
Conclusion 11.16. For getting started with geometry, all you need is 7 points.
So long for algebraic geometry, real, complex or over arbitrary fields, and affine or
projective. For more, a good reference here is the book by Harris [45].

Finally, no discussion about algebraic geometry would be complete without a look into
algebraic topology. We have already seen, in the proof of Theorem 11.13, that “shape”,
taken in a basic topological sense, matters. So, let us further explore this.

Forgetting about manifolds, let us start with something that we know, namely:
Definition 11.17. A topological space X is called connected when any two points
x, y ∈ X can be connected by a path. That is, given any two points x, y ∈ X, we can find
a continuous function f : [0, 1] → X such that f (0) = x and f (1) = y.
The problem is now, given a connected space X, how to count its “holes”. And this
is quite subtle problem, because as examples of such spaces we have:

(1) The sphere, the donut, the double-holed donut, the triple-holed donut, and so on.
These spaces are quite simple, and intuition suggests to declare that the number of holes
of the N -holed donut is, and you guessed right, N .

(2) However, we have as well as example the empty sphere, I mean just the crust of
the sphere, and while this obviously falls into the class of “one-holed spaces”, this is not
the same thing as a donut, its hole being of different nature.
 11C. ALGEBRAIC MANIFOLDS 267

(3) As another example, consider again the sphere, but this time with two tunnels
drilled into it, in the shape of a cross. Whether that missing cross should account for 1
hole, or for 2 holes, or for something in between, I will leave it up to you.

Summarizing, things are quite tricky, suggesting that the “number of holes” of a
topological space X is not an actual number, but rather something more complicated.
Now with this in mind, let us formulate the following definition:
Definition 11.18. The homotopy group π1 (X) of a connected space X is the group
of loops based at a given point ∗ ∈ X, with the following conventions,
(1) Two such loops are identified when one can pass continuously from one loop to
the other, via a family of loops indexed by t ∈ [0, 1],
(2) The composition of two such loops is the obvious one, namely is the loop obtaining
by following the first loop, then the second loop,
(3) The unit loop is the null loop at ∗, which stays there, and the inverse of a given
loop is the loop itself, followed backwards,
with the remark that the group π1 (X) defined in this way does not depend on the choice
of the given point ∗ ∈ X, where the loops are based.
Here the fact that π1 (X) defined in this way is indeed a group is obvious, and obvious
as well is the fact that, since X is assumed to be connected, this group does not depend
on the choice of the given point ∗ ∈ X, where the loops are based.

As basic examples, for spaces having “no holes”, such as R itself, or RN , and so on,
we have π1 = {1}. In fact, having no holes can only mean, by definition, that π1 = {1}.
As further illustrations, here are now a few basic computations:
Theorem 11.19. We have the following computations of homotopy groups:
(1) For the circle, we obtain π1 = Z.
(2) For the torus, we obtain π1 = Z × Z.
(3) For the disk minus 2 points, we have π1 = Z ∗ Z.
Proof. These results are all standard, as follows:
(1) The first assertion is clear, because a loop on the circle must wind n ∈ Z times
around the center, and this parameter n ∈ Z uniquely determines the loop, up to the
identification in Definition 11.18. Thus, the homotopy group of the circle is the group of
such parameters n ∈ Z, which is of course the group Z itself.
(2) In what regards now the second assertion, the torus being a product of two circles,
we are led to the conclusion that its homotopy group must be some kind of product of
Z with itself. But pictures show that the two standard generators of Z, and so the two
copies of Z themselves, commute, gh = hg, and so we obtain the product of Z with itself,
subject to commutation, which is the usual product Z × Z.
268 11. SOME GEOMETRY

(3) This is quite clear, because the homotopy group is generated by the 2 loops around
the 2 missing points, which are obviously free, algebrically speaking. Thus, we obtain a
free product of the group Z with itself, also known as free group on 2 generators. □
There are many other interesting things that can be said about homotopy groups.
Also, another thing that can be done with the arbitrary spaces X, again in relation with
studying their “shape”, is that of looking at the fiber bundles over them, again up to
continuous deformation. We are led in this way into a group, called K0 (X). Moreover,
both π1 (X) and K0 (X) have in fact higher analogues πn (X) and Kn (X) as well, and the
general goal of algebraic topology is that of understanding all these groups.
But all this, obviously, starts to become too complicated. So, leaving now the gen-
eral manifolds and topological spaces aside, let us focus now on the simplest objects of
topology, namely the knots, with this meaning the smooth closed curves in R3 :
Definition 11.20. A knot is a smooth closed curve in R3 , regarded modulo smooth
transformations.
And isn’t this a beautiful definition. We are here at the core of everything that can be
called “geometry”, and in fact, thinking a bit on how knots can be knotted, in so many
fascinating ways, we are led to the following philosophical conclusion:
Conclusion 11.21. Knots are to geometry what prime numbers are to number theory.
At the level of questions now, once we have a closed curve, say given via its algebraic
equations, can we decide if is tied or not, and if tied, how complicated is it tied, how to
untie it, and so on? But these are, obviously, quite difficult questions.
Perhaps simpler now, experience with cables and ropes shows that a random closed
curve is usually tied. But can we really prove this? Once again, difficult question. So, we
will stop here, and exercise for you, to say something non-trivial about knots.
11d. Differential manifolds
Back now to applied mathematics, involving analysis and physics, the situation here
is a bit different. Although many interesting algebraic manifolds appear at the advanced
level, making algebraic geometry a key tool in advanced physics, in what concerns the
basics, here we are mostly in need of a different definition, as follows:
Definition 11.22. A smooth manifold is a space X which is locally isomorphic to
N
R . To be more precise, this space X must be covered by charts, bijectively mapping open
pieces of it to open pieces of RN , with the changes of charts being C ∞ functions.
As a basic example, we have RN itself, or any open subset X ⊂ RN . Another example
is the circle, or curves like ellipses and so on, for obvious reasons. To be more precise, the
unit circle can be covered by 2 charts as above, by using polar coordinates, in the obvious
 11D. DIFFERENTIAL MANIFOLDS 269

way, and then by applying dilations, translations and other such transformations, namely
bijections which are smooth, we obtain a whole menagery of circle-looking manifolds.

In particular, we can see from this that Definition 11.22 is a serious rival to Definition
11.11, because both generalize, in a natural way, the conics that we know well.

So, this is the situation, we have two geometric disciplines inspired by the findings of
the old Greeks, regarding the conics, called algebraic geometry and differential geometry.
With algebraic geometry being substantially older, but in what concerns us, we will rather
go for differential geometry, which is more relevant to basic modern physics.

Going back now to Definition 11.22 as stated, let us first explore the basic examples.
We have already talked about them in the above, but all that discussion needs to be
accompanied by some proofs, and here is a more precise statement in this sense:
Theorem 11.23. The following are smooth manifolds, in the plane:
(1) The circles.
(2) The ellipses.
(3) The non-degenerate conics.
(4) Smooth deformations of these.
Proof. All this is quite intuitive, the idea being as follows:
(1) Consider the unit circle, x2 + y 2 = 1. We can write then x = cos t, y = sin t, with
t ∈ [0, 2π), and we seem to have here the solution to our problem, just using 1 chart.
But this is of course wrong, because [0, 2π) is not open, and we have a problem at 0. In
practice we need to use 2 such charts, say with the first one being with t ∈ (0, 3π/2),
and the second one being with t ∈ (π, 5π/2). As for the fact that the change of charts
is indeed smooth, this comes by writing down the formulae, or just thinking a bit, and
arguing that this change of chart being actually a translation, it is automatically linear.
(2) This follows from (1), by pulling the circle in both the Ox and Oy directions, and
the formulae here, based on Theorem 11.6 or Theorem 11.7, are left to you, reader.
(3) We already have the ellipses, and the case of the parabolas and hyperbolas is
elementary as well, and in fact simpler than the case of the ellipses. Indeed, a parablola
is clearly homeomorphic to R, and a hyperbola, to two copies of R.
(4) This is something which is clear too, depending of course on what exactly we mean
by “smooth deformation”, and by using a bit of multivariable calculus if needed. □
In higher dimensions now, as basic examples here, we have the unit sphere in RN ,
and smooth deformations of it, once again, somehow by obvious reasons. In case you are
wondering on how to construct explicit charts for the sphere, the answer comes from:
270 11. SOME GEOMETRY

Theorem 11.24. We have spherical coordinates in N dimensions,



 x1 = r cos t1

x 2 = r sin t1 cos t2



..
 .
xN −1 = r sin t1 sin t2 . . . sin tN −2 cos tN −1





xN = r sin t1 sin t2 . . . sin tN −2 sin tN −1


with this guaranteeing that the sphere is indeed a smooth manifold.

Proof. There are several things going on here, the idea being as follows:
(1) The fact that we have indeed spherical coordinates is clear, with the only point to
be clarified being the identification of the precise ranges of the angles, which follows from
some geometric thinking, first at N = 2, 3, and then in general.
(2) With the remark that we use here mathematicians’ convention for the angles, which
works nicely in N arbitrary dimensions, as opposed to physicists’ convention, which works
best at N = 3. And with this being something quite subjective, because mathematicians’
convention is based on physicists’ finding that the more dimensions, the better.
(3) Finally, in what regards the last assertion, this can proved a bit like for the circle,
as we did in the proof of Theorem 11.23 (1), basically by cutting the sphere into 2N parts,
and we will leave the details here as an instructive exercise. □

Summarizing, we have the spheres as main examples of differential manifolds, as ex-

pected, but we can also see from the above that the technical verification of the manifold
axioms is something quite complicated. Welcome to differential geometry, where nothing
is really trivial, I mean even things which are supposed to be trivial are not.

In relation with these questions, namely parametrizing the spheres, we have the stere-
ographic projection as well, which works more directly, as follows:
Theorem 11.25. The stereographic projection is given by inverse maps
Φ : RN → SRN − {∞} , Ψ : SRN − {∞} → RN
given by the following formulae,
2 x
Φ(v) = (1, 0) + (−1, v) , Ψ(c, x) =
1 + ||v||2 1−c
with the convention RN +1 = R × RN , and with the coordinate of R denoted x0 , and with
the coordinates of RN denoted x1 , . . . , xN .
 11D. DIFFERENTIAL MANIFOLDS 271

Proof. We are looking for the formulae of the isomorphism RN ≃ SRN −{∞}, obtained
by identifying RN = RN × {0} ⊂ RN +1 with the unit sphere SRN ⊂ RN +1 , with the
convention that the point which is added is ∞ = (1, 0, . . . , 0), via the stereographic
projection. That is, we need the precise formulae of two inverse maps, as follows:
Φ : RN → SRN − {∞} , Ψ : SRN − {∞} → RN
In one sense, according to our conventions above, we must have a formula as follows
for our map Φ, with the parameter t ∈ (0, 1) being such that ||Φ(v)|| = 1:
Φ(v) = t(0, v) + (1 − t)(1, 0)
The equation for the parameter t ∈ (0, 1) can be solved as follows:
(1 − t)2 + t2 ||v||2 = 1 ⇐⇒ t2 (1 + ||v||2 ) = 2t
2
⇐⇒ t =
1 + ||v||2
We conclude that the formula of the map Φ is as follows:
2
Φ(v) = (1, 0) + (−1, v)
1 + ||v||2
In the other sense now we must have, for a certain α ∈ R:
(0, Ψ(c, x)) = α(c, x) + (1 − α)(1, 0)
But from αc + 1 − α = 0 we get the following formula for the parameter α:
1
α=
1−c
We conclude that the formula of the map Ψ is as follows:
x
Ψ(c, x) =
1−c
Here, as before, we use the convention in the statement, namely RN +1 = R × RN , with
the coordinate of R denoted x0 , and with the coordinates of RN denoted x1 , . . . , xN . □
There are of course many other possible parametrizations of the sphere, such as the
one using cylindrical coordinates, or the one peeling the sphere as an orange, and so on.
All this is quite interesting, and as question here, of practical interest, we have:
Question 11.26. What is the best parametrization of the unit sphere in R3 , for purely
mathematical reasons? What about for cartography reasons?
To be more precise, the first question, which is quite challenging, is that of finding the
simplest proof ever for the fact that the sphere in R3 is indeed a smooth manifold. As for
the second question, which is even more challenging, and not really solved by mankind,
despite centuries of work, and many bright ideas, the problem here is to have your charts
reflecting as nicely as possible useful things such as lengths, angles and areas.
272 11. SOME GEOMETRY

By the way, speaking lengths, angles and areas, observe that the general differential
manifold formalism from Definition 11.22 is obviously too broad for talking about these.
In order to do so, several more axioms must be added, and we end up with something
called Riemannian manifold, which is the main object of study of advanced differential
geometry. And, regarding such manifolds, there are many deep theorems, including a key
result of Nash, stating that we can always find an embedding, as follows:
X ⊂ RN
Thus, all in all, all this leads us into the good old RN , and multivariable calculus. But
the story here is quite long and technical, and it’s getting late, and we will stop here.
So long for geometry, in a large sense. As a conclusion to all this, geometry comes in
many flavors, algebraic or differential, Riemannian or not, over R, or C, or some other
field F , and in addition to this we can talk about affine or projective geometry, or about
discrete or continuous geometry, and so on. Many interesting things, and if excited by all
this, orient yourself towards physics, where geometry in all its flavors is needed.
11e. Exercises
This was a pure mathematics chapter, save for some old school computations coming
from Kepler and Newton, and as exercises, all of pure mathematics type, we have:
Exercise 11.27. Learn some algebraic geometry.
Exercise 11.28. Learn some differential geometry.
Exercise 11.29. Learn some Riemannian geometry.
Exercise 11.30. Learn some symplectic geometry too.
With apologies for this, but that is how things go in pure mathematics, “learn” being
the keyword. As bonus exercise, more concrete, have some fun with cartography.
 CHAPTER 12

Higher derivatives

12a. Higher derivatives

Welcome back to analysis, and in the hope that you survived the previous chapter.
Goow news, in this chapter we go for the real thing, namely the computation of maxima
and minima of functions. Want to optimize the characteristics of your building, bridge,
plane, engine, medicine, computer, or casino and other racketeering operations? It all
comes down, you guessed right, to maximizing or minimizing a certain function.

In one variable things are quite easy, at least from a theoretical viewpoint, due to the
following result, that we know well from Part I of the present book:
Theorem 12.1. The one-variable smooth functions are subject to the Taylor formula
∞
X f (k) (x)
f (x + t) = tk
k=0
k!

which allows, via suitable truncations, to determine the local maxima and minima.
Proof. This is a compact summary of what we know from Part I, with everything
being in fact quite technical, and with the idea being as follows:
(1) In order to compute the local maxima and minima, a first method is by using the
following formula, which comes straight from the definition of the derivative:
f (x + t) ≃ f (x) + f ′ (x)t
Indeed, this formula shows that when f ′ (x) ̸= 0, the point x cannot be a local minimum
or maximum, due to the fact that t → −t will invert the growth. Thus, in order to find
the local minima and maxima, we must compute first the points x satisfying f ′ (x) = 0,
and then perform a more detailed study of each solution x that we found.
(2) In relation with the problems left, the second derivative comes to the rescue.
Indeed, we can use the following more advanced formula, coming via l’Hôpital’s rule:
f ′′ (x) 2
f (x + t) ≃ f (x) + f ′ (x)t + t
2
273
274 12. HIGHER DERIVATIVES

To be more precise, assume that we have f ′ (x) = 0, as required by the study in (1).
Then this second order formula simply reads:
f ′′ (x) 2
f (x + t) ≃ f (x) + t
2
But this is something very useful, telling us that when f ′′ (x) < 0, what we have is
a local maximum, and when f ′′ (x) > 0, what we have is a local minimum. As for the
remaining case, that when f ′′ (x) = 0, things here remain open.
(3) All this is very useful in practice, and with what we have in (1), complemented
if needed with what we have in (2), we can in principle compute the local minima and
maxima, without much troubles. However, if really needed, more tools are available.
Indeed, we can use if we want the order 3 Taylor formula, which is as follows:
f ′′ (x) 2 f ′′′ (x) 3
f (x + t) ≃ f (x) + f ′ (x)t + t + t
2 6
To be more precise, assume that we are in the case f ′ (x) = f ′′ (x) = 0, which is where
our joint algorithm coming from (1) and (2) fails. In this case, our formula becomes:
f ′′′ (x) 3
f (x + t) ≃ f (x) + t
6
But this solves the problem in the case f ′′′ (x) ̸= 0, because here we cannot have a
local minimum or maximum, due to t → −t which switches growth. As for the remaining
case, f ′′′ (x) = 0, things here remain open, and we have to go at higher order.
(4) Summarizing, we have a recurrence method for solving our problem. In order to
formulate now an abstract result about this, we can use the Taylor formula at order n:
n
X f (k) (x) k
f (x + t) ≃ t
k=0
k!
Indeed, assume that we started to compute the derivatives f ′ (x), f ′′ (x), f ′′ (x), . . . of
our function at the point x, with the goal of finding the first such derivative which does
not vanish, and we found this derivative, as being the order n one:
f ′ (x) = f ′′ (x) = . . . = f (n−1) (x) = 0 , f (n) (x) ̸= 0
Then, the Taylor formula at x at order n takes the following form:
f (n) (x) n
f (x + t) ≃ f (x) + t
n!
But this is exactly what we need, in order to fully solve our local extremum problem.
Indeed, when n is even, if f (n) (x) < 0 what we have is a local maximum, and if f (n) (x) > 0,
what we have is a local minimum. As for the case where n is odd, here we cannot have a
local minimum or maximum, due to t → −t which switches growth. □
 12A. HIGHER DERIVATIVES 275

All the above, Theorem 12.1 and its proof, must be of course perfectly known, when
looking for applications of such things. However, for theoretical purposes, let us record
as well, in a very compact form, what is basically to be remembered:
Theorem 12.2. Given a differentiable function f : R → R, we can always write
f (n) (x) n
f (x + t) ≃ f (x) + t
n!
with f (n) (x) ̸= 0, and this tells us if x is a local minimum, or maximum of f .
Proof. This was the conclusion of the proof of Theorem 12.1, and with the extra
remark that local extremum means that n is even, with in this case f (n) (x) < 0 corre-
sponding to local maximum, and f (n) (x) > 0 corresponding to local minimum. □
In several variables now, things will be quite tricky, making full use of the material
that we learned in the previous 3 chapters, and even more, requiring some continuations
of that. Indeed, we need a lot of knowledge, in order to solve our problems:

(1) The linear algebra from chapter 9 is the backbone of multivariable calculus, so we
will surely need all that, and more. For instance, since the first derivatives f ′ (x) are now
matrices, expect some sort of positivity theory for the matrices, to be needed.

(2) Regarding the partial derivatives from chapter 10, no question about it, all that
material is certainly useful, and the more theory we have here, the better it will be, for
our questions. The problem is that of iterating that partial derivative operations.

(3) Finally, regarding the geometry from chapter 11, the functions that we want to
study will be naturally defined either on RN , or on spheres, ellipses, cylinders and other
differential manifolds X ⊂ RN , so we will certainly need that material too.

Getting started now, we can talk about higher derivatives, in the obvious way, simply
by performing the operation of taking derivatives recursively. As result here, we have:
Theorem 12.3. Given a continuous function f : RN → R, we can talk about its higher
derivatives, defined recursively as
dk f d d
= ··· (f )
dxi1 . . . dxik dxi1 dxik
provided that all these derivatives exist indeed. Moreover, due to the Clairaut formula,
d2 f d2 f
=
dxi dxj dxj dxi
the order in which these higher derivatives are computed is irrelevant.
276 12. HIGHER DERIVATIVES

Proof. There are several things going on here, the idea being as follows:
(1) First of all, we can talk about the quantities in the statement, with the remark
however that at each step of our recursion, the corresponding partial derivative can exist
of not. We will say in what follows that our function is k times differentiable if the
quantities in the statement exist at any l ≤ k, and smooth, if this works with k = ∞.
(2) Regarding now the second assertion, this is something more tricky. Let us first
recall from chapter 8 that the second derivatives of a twice differentiable function of two
variables f : R2 → R are subject to the Clairaut formula, namely:
d2 f d2 f
=
dxdy dydx
(3) But this result clearly extends to our function f : RN → R, simply by ignoring
the unneeded variables, so we have the Clairaut formula in general, also called Schwarz
formula, which is the one in the statement, namely:
d2 f d2 f
=
dxi dxj dxj dxi
(4) Now observe that this tells us that the order in which the higher derivatives
are computed is irrelevant. That is, we can permute the order of our partial derivative
computations, and a standard way of doing this is by differentiating first with respect to
x1 , as many times as needed, then with respect to x2 , and so on. Thus, the collection of
partial derivatives can be written, in a more convenient form, as follows:
dk f dk1 dkN
= k1 · · · kN (f )
dxk11 . . . dxkNN dx1 dxN
(5) To be more precise, here k ∈ N is as usual the global order of our derivatives,
the exponents k1 , . . . , kN ∈ N are subject to the condition k1 + . . . + kN = k, and the
operations on the right are the familiar one-variable higher derivative operations.
(6) This being said, for certain tricky questions it is more convenient not to order the
indices, or rather to order them according to what order best fits our computation, so
what we have in the statement is the good formula, and (4-5) are mere remarks.
(7) And with the remark too that for trivial questions, what we have in the statement
is the good formula, simply because there are less indices to be written, when compared
to what we have to write when using the ordering procedure in (4-5) above. □

All this is very nice, and as an illustration for the above, let us work out the case
k = 2. Here things are quite special, and we can formulate the following definition:
 12A. HIGHER DERIVATIVES 277

Definition 12.4. Given a twice differentiable function f : RN → R, we set

 2 
′′ df
f (x) =
dxi dxj ij
which is a symmetric matrix, called Hessian matrix of f at the point x ∈ RN .
To be more precise, we know that when f : RN → R is twice differentiable, its order
k = 2 partial derivatives are the numbers in the statement. Now since these numbers
naturally form a N × N matrix, the temptation is high to call this matrix f ′′ (x), and so
we will do. And finally, we know from Clairaut that this matrix is symmetric:
f ′′ (x)ij = f ′′ (x)ji
Observe that at N = 1 this is compatible with our previous definition of the second
derivative f ′′ , and this because in this case, the 1 × 1 matrix from Definition 12.4 is:
f ′′ (x) = (f ′′ (x)) ∈ M1×1 (R)
As a word of warning, however, never use Definition 12.4 for functions f : RN → RM ,
where the second derivative can only be something more complicated. Also, never attempt
either to do something similar at k = 3 or higher, for functions f : RN → R with N > 1,
because again, that beast has too many indices, for being a true, honest matrix.

Back now to business, with these notions, we have the following question to be solved:
Question 12.5. What is the Taylor formula for a function
f : RN → R
and how can this be used for computing the local minima and maxima of f ?
We will solve this slowly, a bit as we did in the proof of Theorem 12.1, in the N = 1
case. Let us start with something that we know well from chapter 10, namely:
f (x + t) ≃ f (x) + f ′ (x)t
To be more precise, we know that this formula holds indeed, with the derivative
f ′ (x) being by definition the horizontal vector formed by the partial derivatives, and with
t ∈ RN being regarded as usual as column vector, the formula of f ′ (x)t being:
 t1
 

df df  ... 
f ′ (x)t = ...
dx1 dxN
tN
N
X df
= ti
i=1
dx i

∈ R
278 12. HIGHER DERIVATIVES

Here we have of course identified the 1 × 1 matrices with their numeric content. As a
consequence, in analogy with what we know in 1 variable, we can formulate:
Theorem 12.6. The Taylor formula at order 1 for a function f : RN → R is
f (x + t) ≃ f (x) + f ′ (x)t
and in particular, in order for x to be a local extremum, we must have f ′ (x) = 0.
Proof. Here the first assertion is something that we know, as explained above, and
the second assertion follows from it. Indeed, let us look at the order 1 term, given by:
N
X df
f ′ (x)t = ti
i=1
dx i

Now since this linear combination of the entries of t ∈ RN can range among positives
and negatives, unless all the coefficients are zero, which means f ′ (x) = 0, we are led to
the conclusion that local extremum needs f ′ (x) = 0 to hold, as stated. □
Let us discuss now the Taylor formula at order 2. We have here:
Theorem 12.7. Given a twice differentiable function f : RN → R, we have
< f ′′ (x)t, t >
f (x + t) ≃ f (x) + f ′ (x)t +
2
′′
where f (x) ∈ MN (R) stands as usual for the Hessian matrix.
Proof. This is something more tricky, the idea being as follows:
(1) As a first observation, at N = 1 the Hessian matrix as constructed in Definition
12.4 is the 1 × 1 matrix having as entry the second derivative f ′′ (x), and the formula in
the statement is something that we know well from Part I, namely:
f ′′ (x)t2
f (x + t) ≃ f (x) + f ′ (x)t +
2
(2) In general now, this is in fact something which does not need a new proof, because
it follows from the one-variable formula above, applied to the restriction of f to the
following segment in RN , which can be regarded as being a one-variable interval:
I = [x, x + t]
N
To be more precise, let y ∈ R , and consider the following function, with r ∈ R:
g(r) = f (x + ry)
We know from (1) that the Taylor formula for g, at the point r = 0, reads:
′ g ′′ (0)r2
g(r) ≃ g(0) + g (0)r +
2
And our claim is that, with t = ry, this is precisely the formula in the statement.
 12A. HIGHER DERIVATIVES 279

(3) So, let us see if our claim is correct. By using the chain rule, we have the following
formula, with on the right, as usual, a row vector multiplied by a column vector:
g ′ (r) = f ′ (x + ry) · y
By using again the chain rule, we can compute the second derivative as well:
g ′′ (r) = (f ′ (x + ry) · y)′
!′
X df
= (x + ry) · yi
i
dxi
X X d2 f d(x + ry)j
= (x + ry) · · yi
i j
dx i dx j dr
X X d2 f
= (x + ry) · yi yj
i j
dxi dxj
= < f ′′ (x + ry)y, y >
(4) Time now to conclude. We know that we have g(r) = f (x + ry), and according to
our various computations above, we have the following formulae:
g(0) = f (x) , g ′ (0) = f ′ (x) , g ′′ (0) =< f ′′ (x)y, y >
Buit with this data in hand, the usual Taylor formula for our one variable function g,
at order 2, at the point r = 0, takes the following form, with t = ry:
< f ′′ (x)y, y > r2
f (x + ry) ≃ f (x) + f ′ (x)ry +
2
′′
< f (x)t, t>
= f (x) + f ′ (x)t +
2
Thus, we have obtained the formula in the statement.
(5) Finally, for completness, let us record as well a more numeric formulation of what
we found. According to our usual rules for matrix calculus, what we found is:
N N N
X df 1 X X d2 f
f (x + t) ≃ f (x) + ti + ti tj
i=1
dx i 2 i=1 j=1
dx i dx j

Observe that, since the Hessian matrix f ′′ (x) is symmetric, most of the terms on the
right will appear in pairs, making it clear what the 1/2 is there for, namely avoiding
redundancies. However, this is only true for the off-diagonal terms, so instead of further
messing up our numeric formula above, we will just leave it like this. □
We can now go back to local extrema, and we have, improving Theorem 12.6:
280 12. HIGHER DERIVATIVES

Theorem 12.8. In order for a twice differentiable function f : RN → R to have a

local minimum or maximum at x ∈ RN , the first derivative must vanish there,
f ′ (x) = 0
and the Hessian must be positive or negative, in the sense that the quantities
< f ′′ (x)t, t >∈ R
must keep a constant sign, positive or negative, when t ∈ RN varies.
Proof. This is clear from Theorem 12.7. Consider indeed the formula there, namely:
< f ′′ (x)t, t >
f (x + t) ≃ f (x) + f ′ (x)t +
2
It is clear then that, in order for our function to have a local minimum or maximum at
x ∈ RN , the first derivative must vanish there, f ′ (x) = 0. Moreover, with this assumption
made, the approximation that we have around x becomes:
< f ′′ (x)t, t >
f (x + t) ≃ f (x) +
2
Thus, we are led to the conclusion in the statement. □
As a conclusion to our study so far, our analytic questions lead us into a linear algebra
question, regarding the square matrices of type f ′′ (x) ∈ MN (R), and more specifically the
positivity properties of the following quantities, when t ∈ RN varies:
< f ′′ (x)t, t >∈ R
This is actually a quite subtle question, and many things can be said here, after some
linear algebra work. We will be back to this in a moment.

Finally, at higher order things become more complicated, as follows:

Theorem 12.9. Given an order k differentiable function f : RN → R, we have
′ < f ′′ (x)t, t >
f (x + t) ≃ f (x) + f (x)t + + ...
2
and this helps in identifying the local extrema, when f ′ (x) = 0 and f ′′ (x) = 0.
Proof. The study here is very similar to that at k = 2, from the proof of Theorem
12.7, with everything coming from the usual Taylor formula, applied on:
I = [x, x + t]
We will leave this as an instructive exercise, for your long Summer nights. □
 12B. MATRICES, POSITIVITY 281

12b. Matrices, positivity

According to what we know so far, the local extrema of a function f : RN → R come
under the assumption f ′ (x) = 0, and depend on the positivity properties of the Hessian
matrix f ′′ (x) ∈ MN (R). In view of this, and of the Taylor formula at order 2 in general,
the best method in order to deal with a function f : RN → R satisfying f ′ (x) = 0 is
that of changing the basis around x ∈ RN as to have the Hessian matrix f ′′ (x) ∈ MN (R)
diagonalized, if this diagonalization is of course possible. Thus, forgetting now about
functions and analysis, we are led into the following linear algebra question:
Question 12.10. Which symmetric matrices A ∈ MN (R) have the property
< At, t >≥ 0
for any t ∈ RN ? Also, when is a matrix A ∈ MN (R) diagonalizable?
Observe that in the second question we have lifted the assumption that A is sym-
metric, and this for good reason. Indeed, besides the local extremum problematics for
the functions f : RN → R, discussed above, we have as well the quite interesting prob-
lem of studying the functions f : RN → RN , whose derivatives are square matrices,
f ′ (x) ∈ MN (R), waiting to be diagonalized too. So, we are trying here two shoot 2
rabbits at the same time, with our Question 12.10, as formulated above.

Obviously, hunting matters, so time to ask the expert. And the expert says:
Cat 12.11. In the lack of speed and claws, yes, develop as much linear algebra as you
can. Without even caring for applications, these will come naturally.
Thanks cat, so this will be our plan for this section, develop as much linear algebra as
we can. And for applications, we will most likely leave them to you, reader, for later in
life, depending on the precise physics and engineering questions that you will be interested
in. And with the advice of course to follow the feline way, relax, and no mercy.

With this plan made, let us go back to the diagonalization question, from chapter 9.
We will need here diagonalization results which are far more powerful. We first have:
Theorem 12.12. Any matrix A ∈ MN (C) which is self-adjoint, A = A∗ , is diagonal-
izable, with the diagonalization being of the following type,
A = U DU ∗
with U ∈ UN , and with D ∈ MN (R) diagonal. The converse holds too.
Proof. As a first remark, the converse trivially holds, because if we take a matrix of
the form A = U DU ∗ , with U unitary and D diagonal and real, then we have:
A∗ = (U DU ∗ )∗ = U D∗ U ∗ = U DU ∗ = A
282 12. HIGHER DERIVATIVES

In the other sense now, assume that A is self-adjoint, A = A∗ . Our first claim is that
the eigenvalues are real. Indeed, assuming Av = λv, we have:
λ < v, v > = < λv, v >
= < Av, v >
= < v, Av >
= < v, λv >
= λ̄ < v, v >
Thus we obtain λ ∈ R, as claimed. Our next claim now is that the eigenspaces
corresponding to different eigenvalues are pairwise orthogonal. Assume indeed that:
Av = λv , Aw = µw
We have then the following computation, using λ, µ ∈ R:
λ < v, w > = < λv, w >
= < Av, w >
= < v, Aw >
= < v, µw >
= µ < v, w >
Thus λ ̸= µ implies v ⊥ w, as claimed. In order now to finish the proof, it remains to
prove that the eigenspaces of A span the whole space CN . For this purpose, we will use
a recurrence method. Let us pick an eigenvector of our matrix:
Av = λv
Assuming now that we have a vector w orthogonal to it, v ⊥ w, we have:
< Aw, v > = < w, Av >
= < w, λv >
= λ < w, v >
= 0
Thus, if v is an eigenvector, then the vector space v ⊥ is invariant under A. Moreover,
since a matrix A is self-adjoint precisely when < Av, v >∈ R for any vector v ∈ CN , as
one can see by expanding the scalar product, the restriction of A to the subspace v ⊥ is
self-adjoint. Thus, we can proceed by recurrence, and we obtain the result. □

Observe that, as a consequence of the above result, that you certainly might have
heard of, any symmetric matrix A ∈ MN (R) is diagonalizable. In fact, we have:
 12B. MATRICES, POSITIVITY 283

Proposition 12.13. Any matrix A ∈ MN (R) which is symmetric, A = At , is diago-

nalizable, with the diagonalization being of the following type,
A = U DU t
with U ∈ ON , and with D ∈ MN (R) diagonal. The converse holds too.
Proof. As before, the converse trivially holds, because if we take a matrix of the
form A = U DU t , with U orthogonal and D diagonal and real, then we have At = A. In
the other sense now, this follows from Theorem 12.12, and its proof. □
As basic examples of self-adjoint matrices, we have the orthogonal projections:
Proposition 12.14. The matrices P ∈ MN (C) which are projections, P 2 = P ∗ = P ,
are precisely those which diagonalize as follows,
P = U DU ∗
with U ∈ UN , and with D ∈ MN (0, 1) being diagonal.
Proof. Since we have P ∗ = P , by using Theorem 12.12, the eigenvalues must be
real. Then, by using P 2 = P , assuming that we have P v = λv, we obtain:
λ < v, v > = < λv, v >
= < P v, v >
= < P 2 v, v >
= < P v, P v >
= < λv, λv >
= λ2 < v, v >
We therefore have λ ∈ {0, 1}, as claimed, and as a final conclusion here, the diagonal-
ization of the projection matrices is as follows, with ei ∈ {0, 1}:
e1
 

P ∼ ... 
eN
To be more precise, the number of 1 values is the dimension of the image of P . □
In the real case, the result regarding the projections is as follows:
Proposition 12.15. The matrices P ∈ MN (R) which are projections, P 2 = P t = P ,
are precisely those which diagonalize as follows,
P = U DU t
with U ∈ ON , and with D ∈ MN (0, 1) being diagonal.
Proof. This follows indeed from Proposition 12.14, and its proof. □
284 12. HIGHER DERIVATIVES

An important class of self-adjoint matrices, which includes for instance all the projec-
tions, are the positive matrices. The theory here is as follows:
Theorem 12.16. For a matrix A ∈ MN (C) the following conditions are equivalent,
and if they are satisfied, we say that A is positive:
(1) A = B 2 , with B = B ∗ .
(2) A = CC ∗ , for some C ∈ MN (C).
(3) < Ax, x >≥ 0, for any vector x ∈ CN .
(4) A = A∗ , and the eigenvalues are positive, λi ≥ 0.
(5) A = U DU ∗ , with U ∈ UN and with D ∈ MN (R+ ) diagonal.
Proof. The idea is that the equivalences in the statement basically follow from some
elementary computations, with only Theorem 12.12 needed, at some point:
(1) =⇒ (2) This is clear, because we can take C = B.
(2) =⇒ (3) This follows from the following computation:
< Ax, x > = < CC ∗ x, x >
= < C ∗ x, C ∗ x >
≥ 0
(3) =⇒ (4) By using the fact that < Ax, x > is real, we have:
< Ax, x > = < x, A∗ x >
= < A∗ x, x >
Thus we have A = A∗ , and the remaining assertion, regarding the eigenvalues, follows
from the following computation, assuming Ax = λx:
< Ax, x > = < λx, x >
= λ < x, x >
≥ 0
(4) =⇒ (5) This follows indeed by using Theorem 12.12.
(5) =⇒ (1) Assuming
√ A = U DU ∗ with U ∈ UN , and with D ∈ MN (R+ ) diagonal,
∗
we can set B = U DU . Then B is self-adjoint, and its square is given by:
√ √
B 2 = U DU ∗ · U DU ∗
= U DU ∗
= A
Thus, we are led to the conclusion in the statement. □
Let us record as well the following technical version of the above result:
 12B. MATRICES, POSITIVITY 285

Theorem 12.17. For a matrix A ∈ MN (C) the following conditions are equivalent,
and if they are satisfied, we say that A is strictly positive:
(1) A = B 2 , with B = B ∗ , invertible.
(2) A = CC ∗ , for some C ∈ MN (C) invertible.
(3) < Ax, x >> 0, for any nonzero vector x ∈ CN .
(4) A = A∗ , and the eigenvalues are strictly positive, λi > 0.
(5) A = U DU ∗ , with U ∈ UN and with D ∈ MN (R∗+ ) diagonal.
Proof. This follows either from Theorem 12.16, by adding the above various extra
assumptions, or from the proof of Theorem 12.16, by modifying where needed. □

Let us discuss now the case of the unitary matrices. We have here:
Theorem 12.18. Any matrix U ∈ MN (C) which is unitary, U ∗ = U −1 , is diagonaliz-
able, with the eigenvalues on T. More precisely we have
U = V DV ∗
with V ∈ UN , and with D ∈ MN (T) diagonal. The converse holds too.
Proof. As a first remark, the converse trivially holds, because given a matrix of type
U = V DV ∗ , with V ∈ UN , and with D ∈ MN (T) being diagonal, we have:
U∗ = (V DV ∗ )∗
= V D∗ V ∗
= V D−1 V −1
= (V ∗ )−1 D−1 V −1
= (V DV ∗ )−1
= U −1
Let us prove now the first assertion, stating that the eigenvalues of a unitary matrix
U ∈ UN belong to T. Indeed, assuming U v = λv, we have:
< v, v > = < U ∗ U v, v >
= < U v, U v >
= < λv, λv >
= |λ|2 < v, v >
Thus we obtain λ ∈ T, as claimed. Our next claim now is that the eigenspaces
corresponding to different eigenvalues are pairwise orthogonal. Assume indeed that:
U v = λv , U w = µw
286 12. HIGHER DERIVATIVES

We have then the following computation, using U ∗ = U −1 and λ, µ ∈ T:

λ < v, w > = < λv, w >
= < U v, w >
= < v, U ∗ w >
= < v, U −1 w >
= < v, µ−1 w >
= µ < v, w >
Thus λ ̸= µ implies v ⊥ w, as claimed. In order now to finish the proof, it remains to
prove that the eigenspaces of U span the whole space CN . For this purpose, we will use
a recurrence method. Let us pick an eigenvector of our matrix:
U v = λv
Assuming that we have a vector w orthogonal to it, v ⊥ w, we have:
< U w, v > = < w, U ∗ v >
= < w, U −1 v >
= < w, λ−1 v >
= λ < w, v >
= 0
Thus, if v is an eigenvector, then the vector space v ⊥ is invariant under U . Now since
U is an isometry, so is its restriction to this space v ⊥ . Thus this restriction is a unitary,
and so we can proceed by recurrence, and we obtain the result. □

Let us record as well the real version of the above result, in a weak form:
Proposition 12.19. Any matrix U ∈ MN (R) which is orthogonal, U t = U −1 , is
diagonalizable, with the eigenvalues on T. More precisely we have
U = V DV ∗
with V ∈ UN , and with D ∈ MN (T) being diagonal.
Proof. This follows indeed from Theorem 12.18. □

Observe that the above result does not provide us with a complete characterization
of the matrices U ∈ MN (R) which are orthogonal. To be more precise, the question left
is that of understanding when the matrices of type U = V DV ∗ , with V ∈ UN , and with
D ∈ MN (T) being diagonal, are real, and this is something non-trivial.
 12B. MATRICES, POSITIVITY 287

As an illustration, for the simplest unitaries that we know, namely the rotations in
the real plane, we have the following formula, that we know well from chapter 9:
     −it  
cos t − sin t 1 1 1 e 0 1 −i
=
sin t cos t 2 i −i 0 eit 1 i
Back to generalities, the self-adjoint matrices and the unitary matrices are particular
cases of the general notion of a “normal matrix”, and we have here:
Theorem 12.20. Any matrix A ∈ MN (C) which is normal, AA∗ = A∗ A, is diagonal-
izable, with the diagonalization being of the following type,
A = U DU ∗
with U ∈ UN , and with D ∈ MN (C) diagonal. The converse holds too.
Proof. As a first remark, the converse trivially holds, because if we take a matrix of
the form A = U DU ∗ , with U unitary and D diagonal, then we have:
AA∗ = U DU ∗ · U D∗ U ∗
= U DD∗ U ∗
= U D∗ DU ∗
= U D∗ U ∗ · U DU ∗
= A∗ A
In the other sense now, this is something more technical. Our first claim is that a
matrix A is normal precisely when the following happens, for any vector v:
||Av|| = ||A∗ v||
Indeed, the above equality can be written as follows:
< AA∗ v, v >=< A∗ Av, v >
But this is equivalent to AA∗ = A∗ A, by expanding the scalar products. Our claim
now is that A, A∗ have the same eigenvectors, with conjugate eigenvalues:
Av = λv =⇒ A∗ v = λ̄v
Indeed, this follows from the following computation, and from the trivial fact that if
A is normal, then so is any matrix of type A − λ1N :
||(A∗ − λ̄1N )v|| = ||(A − λ1N )∗ v||
= ||(A − λ1N )v||
= 0
Let us prove now, by using this, that the eigenspaces of A are pairwise orthogonal.
Assume that we have two eigenvectors, corresponding to different eigenvalues, λ ̸= µ:
Av = λv , Aw = µw
288 12. HIGHER DERIVATIVES

We have the following computation, which shows that λ ̸= µ implies v ⊥ w:

λ < v, w > = < λv, w >
= < Av, w >
= < v, A∗ w >
= < v, µ̄w >
= µ < v, w >
In order to finish, it remains to prove that the eigenspaces of A span the whole CN .
This is something that we have already seen for the self-adjoint matrices, and for unitaries,
and we will use here these results, in order to deal with the general normal case. As a
first observation, given an arbitrary matrix A, the matrix AA∗ is self-adjoint:
(AA∗ )∗ = AA∗
Thus, we can diagonalize this matrix AA∗ , as follows, with the passage matrix being
a unitary, V ∈ UN , and with the diagonal form being real, E ∈ MN (R):
AA∗ = V EV ∗
Now observe that, for matrices of type A = U DU ∗ , which are those that we supposed
to deal with, we have the following formulae:
V =U , E = DD̄
In particular, the matrices A and AA∗ have the same eigenspaces. So, this will be
our idea, proving that the eigenspaces of AA∗ are eigenspaces of A. In order to do so, let
us pick two eigenvectors v, w of the matrix AA∗ , corresponding to different eigenvalues,
λ ̸= µ. The eigenvalue equations are then as follows:
AA∗ v = λv , AA∗ w = µw
We have the following computation, using the normality condition AA∗ = A∗ A, and
the fact that the eigenvalues of AA∗ , and in particular µ, are real:
λ < Av, w > = < λAv, w >
= < Aλv, w >
= < AAA∗ v, w >
= < AA∗ Av, w >
= < Av, AA∗ w >
= < Av, µw >
= µ < Av, w >
We conclude that we have < Av, w >= 0. But this reformulates as follows:
λ ̸= µ =⇒ A(Eλ ) ⊥ Eµ
 12B. MATRICES, POSITIVITY 289

Now since the eigenspaces of AA∗ are pairwise orthogonal, and span the whole CN ,
we deduce from this that these eigenspaces are invariant under A:
A(Eλ ) ⊂ Eλ
But with this result in hand, we can finish. Indeed, we can decompose the problem,
and the matrix A itself, following these eigenspaces of AA∗ , which in practice amounts
in saying that we can assume that we only have 1 eigenspace. By rescaling, this is the
same as assuming that we have AA∗ = 1, and so we are now into the unitary case, that
we know how to solve, as explained in Theorem 12.18. □
As a first application of all this, we have the following result:
Theorem 12.21. Given a matrix A ∈ MN (C), we can construct a matrix |A| as
follows, by using the fact that A∗ A is diagonalizable, with positive eigenvalues:
√
|A| = A∗ A
This matrix |A| is then positive, and its square is |A|2 = A∗ A. In the case N = 1, we
obtain in this way the usual absolute value of the complex numbers.
Proof. Consider indeed the matrix A∗ A, which is normal. According to Theorem
12.20, we can diagonalize this matrix as follows, with U ∈ UN , and with D diagonal:
A = U DU ∗
From A∗ A ≥ 0 we obtain D ≥ 0. But this √ means that the entries of D are real, and
positive. Thus we can extract the square root D, and then set:
√ √
A∗ A = U DU ∗
Thus, we are basically done. Indeed, if we call this latter matrix |A|, we are led to the
conclusions in the statement. Finally, the last assertion is clear from definitions. □
We can now formulate a first polar decomposition result, as follows:
Theorem 12.22. Any invertible matrix A ∈ MN (C) decomposes as
A = U |A|
√
with U ∈ UN , and with |A| = A∗ A as above.
Proof. This is routine, and follows by comparing the actions of A, |A| on the vectors
v ∈ CN , and deducing from this the existence of a unitary U ∈ UN as above. □
Observe that at N = 1 we obtain in this way the usual polar decomposition of the
nonzero complex numbers. More generally now, we have the following result:
Theorem 12.23. Any square matrix A ∈ MN (C) decomposes as
A = U |A|
√
with U being a partial isometry, and with |A| = A∗ A as above.
290 12. HIGHER DERIVATIVES

Proof. Again, this follows by comparing the actions of A, |A| on the vectors v ∈ CN ,
and deducing from this the existence of a partial isometry U as above. Alternatively, we
can get this from Theorem 12.22, applied on the complement of the 0-eigenvectors. □
So long for advanced linear algebra. There are actually many other decomposition
results for the real matrices, quite often in relation with positivity, and the Jordan form
too, which are all useful for questions in analysis, via derivatives and Hessians.

Good luck of course in learning all this, when needed later in life, for your various
math problems at that time. And always have in mind expert’s advice, Cat 12.11.

12c. Harmonic functions

As another application of the second derivatives, we can now talk about harmonic
functions in general, extending what we know from chapter 8. We first have:
Definition 12.24. The Laplace operator in N dimensions is:
N
X d2 f
∆f =
i=1
dx2i
A function f : RN → C satisfying ∆f = 0 will be called harmonic.
Observe that this generalizes what we know in two dimensions, from chapter 8. In
order to discuss this, let us first recall the handful of results that we know from there,
in N dimensions, with the comment of course that chapter 8 was a physics chapter, and
now that we learned mathematics in N dimensions, you can check all that stuff, which
was coming a bit in advance, and correct all that stuff is. First is the validity of the wave
equation in N dimensions, which is as follows, v > 0 being the propagation speed:
φ̈ = v 2 ∆φ
Then second is the validity of the heat equation in N dimensions, which is as follows,
with α > 0 being the thermal diffusivity of the medium:
φ̇ = α∆φ
And finally, we had as well some mathematics in chapter 8, but mostly 2-dimensional,
and not very useful for our purposes here. Let us recall however a key computation, that
we did in N dimensions, stating that the fundamental radial solutions of ∆f = 0 are as
follows, with the log at N = 2 basically coming from log′ = 1/x:
(
||x||2−N (N ̸= 2)
f (x) =
log ||x|| (N = 2)
Not bad for a start, all this. In analogy now with the one complex variable theory
that we know from chapter 6 and chapter 8, we have the following result:
 12C. HARMONIC FUNCTIONS 291

Theorem 12.25. The harmonic functions in N dimensions obey to the same general
principles as the holomorphic functions, namely:
(1) The plain mean value formula.
(2) The boundary mean value formula.
(3) The maximum modulus principle.
(4) The Liouville theorem.
Proof. This is something quite tricky, the idea being as follows:
(1) Regarding the plain mean value formula, here the statement is that given an
harmonic function f : X → C, and a ball B, the following happens:
Z
f (x) = f (y)dy
B

In order to prove this, we can assume that B is centered at 0, of radius r > 0. If we

denote by χr the characteristic function of this ball, nomalized as to integrate up to 1, in
terms of the convolution operation from chapter 7, we want to prove that we have:
f = f ∗ χr
For doing so, let us pick a number 0 < s < r, and a solution w of the following
equation, supported on B, which can be constructed explicitly:
∆w = χr − χs
By using the properties of the convolution operation ∗ from chapter 7, we have:
f ∗ χr − f ∗ χs = f ∗ (χr − χs )
= f ∗ ∆w
= ∆f ∗ w
= 0
Thus f ∗ χr = f ∗ χs , and by letting now s → 0, we get f ∗ χr = f , as desired.
(2) Regarding the boundary mean value formula, here the statement is that given an
harmonic function f : X → C, and a ball B, with boundary γ, the following happens:
Z
f (x) = f (y)dy
γ

But this follows as a consequence of the plain mean value formula in (1), with our two
mean value formulae, the one there and the one here, being in fact equivalent, by using
annuli and radial integration for the proof of the equivalence, in the obvious way.
(3) Regarding the maximum modulus principle, the statement here is that any holo-
morphic function f : X → C has the property that the maximum of |f | over a domain is
292 12. HIGHER DERIVATIVES

attained on its boundary. That is, given a domain D, with boundary γ, we have:
∃x ∈ γ , |f (x)| = max |f (y)|
y∈D

But this is something which follows again from the mean value formula in (1), first
for the balls, and then in general, by using a standard division argument.
(4) Finally, regarding the Liouville theorem, the statement here is that an entire,
bounded harmonic function must be constant:
f : RN → C , ∆f = 0 , |f | ≤ M =⇒ f = constant
As a slightly weaker statement, again called Liouville theorem, we have the fact that
an entire harmonic function which vanishes at ∞ must vanish globally:
f : RN → C , ∆f = 0 , lim f (x) = 0 =⇒ f =0
x→∞

But can view these as a consequence of the mean value formula in (1), because given
two points x ̸= y, we can view the values of f at these points as averages over big balls
centered at these points, say B = Bx (R) and C = By (R), with R >> 0:
Z Z
f (x) = f (z)dz , f (y) = f (z)dz
B C

Indeed, the point is that when the radius goes to ∞, these averages tend to be equal,
and so we have f (x) ≃ f (y), which gives f (x) = f (y) in the limit, as desired. □

12d. Lagrange multipliers

We discuss now some more specialized topics, in relation with optimization. First of
all, thinking well, the functions that we have to minimize or maximize, in the real life,
are often defined on a manifold, instead of being defined on the whole RN . Fortunately,
the good old principle f ′ (x) = 0 can be adapted to the manifold case, as follows:
Principle 12.26. In order for a function f : X → R defined on a manifold X to have
a local extremum at x ∈ X, we must have, as usual
f ′ (x) = 0
but with this taking into account the fact that the equations defining the manifold count
as well as “zero”, and so must be incorporated into the formula f ′ (x) = 0.
Obviously, we are punching here about our weight, because our discussion about man-
ifolds from chapter 11 was quite amateurish, and we have no tools in our bag for proving
such things, or even for properly formulating them. However, we can certainly talk about
all this, a bit like physicists do. So, our principle will be the one above, and in practice,
 12D. LAGRANGE MULTIPLIERS 293

the idea is that we must have a formula as follows, with gi being the constraint functions
for our manifold X, and with λi ∈ R being certain scalars, called Lagrange multipliers:
X
f ′ (x) = λi gi′ (x)
i

As a basic illustration for this, our claim is that, by using a suitable manifold, and a
suitable function, and Lagrange multipliers, we can prove in this way the Hölder inequality,
that we know well of course, but without any computation. Let us start with:
Proposition 12.27. For any exponent p > 1, the following set
( )
X
N p
Sp = x ∈ R |xi | = 1
i

is a submanifold of RN .
Proof. We know from chapter 11 that the unit sphere in RN is a manifold. In our
terms, this solves our problem at p = 2, because this unit sphere is:
( )
X
N 2
S2 = x ∈ R xi = 1
i

Now observe that we have a bijection Sp ≃ S2 , at least on the part where all the
coordinates are positive, xi > 0, given by the following function:
2/p
xi → xi
Thus we obtain that Sp is indeed a manifold, as claimed. □

We already know that the manifold Sp constructed above is the unit sphere, in the
case p = 2. In order to have a better geometric picture of what is going on, in general,
observe that Sp can be constructed as well at p = 1, as follows:
( )
X
S1 = x ∈ RN |xi | = 1
i

However, this is no longer a manifold, as we can see for instance at N = 2, where we

obtain a square. Now observe that we can talk as well about p = ∞, as follows:
 
N
S∞ = x ∈ R sup |xi | = 1
i

This letter set is no longer a manifold either, as we can see for instance at N = 2,
where we obtain again a square, containing the previous square, the one at p = 1.
294 12. HIGHER DERIVATIVES

With these limiting constructions in hand, we can have now a better geometric picture
of what is going on, in the general context of Proposition 12.27. Indeed, let us draw, at
N = 2 for simplifying, our sets Sp at the values p = 1, 2, ∞ of the exponent:
◦ ◦ ◦

◦ ◦ ◦

◦ ◦ ◦
We can see that what we have is a small square, at p = 1, becoming smooth and
inflating towards the circle, in the parameter range p ∈ (1, 2], and then further inflating,
in the parameter range p ∈ [2, ∞), towards the big square appearing at p = ∞.

With these preliminaries in hand, we can formulate our result, as follows:

Theorem 12.28. The local extrema over Sp of the function
X
f (x) = xi yi
i

can be computed by using Lagrange multipliers, and this gives

!1/p !1/q
X X X
xi y i ≤ |xi |p |yi |q
i i i

with 1/p + 1/q = 1, that is, the Hölder inequality, with a purely geometric proof.
Proof. We can restrict the attention to the case where all the coordinates are posi-
tive, xi > 0 and yi > 0. The derivative of the function in the statement is:
f ′ (x) = (y1 , . . . , yN )
On the other hand, we knowP that the manifold Sp appears by definition as the set of
p
zeroes of the function φ(x) = i xi − 1, having derivative as follows:

φ′ (x) = p(xp−1 p−1

1 , . . . , xN )

Thus, by using Lagrange multipliers, the critical points of f must satisfy:

(y1 , . . . , yN ) ∼ (xp−1 p−1
1 , . . . , xN )
 12D. LAGRANGE MULTIPLIERS 295

1/(p−1)
In otherPwords, the critical points must satisfy xi = λyi , for some λ > 0, and by
using now i xpi = 1 we can compute the precise value of λ, and we get:
!−1/p
X p/(p−1)
λ= yi
i

Now let us see what this means. Since the critical point is unique, this must be a
maximum of our function, and we conclude that for any x ∈ Sp , we have:
!1−1/p !1/q
X X 1/(p−1) X p/(p−1) X q
xi y i ≤ λyi · yi = yi = yi
i i i i

Thus we have Hölder, and the general case follows from this, by rescaling. □

As a second illustration for the method of Lagrange multipliers, this time in relation
with certain questions from linear algebra, let us go back to the Hadamard matrices, that
we met in chapter 7. In the real case, the basic theory of these matrices is as follows:
Theorem 12.29. The real Hadamard matrices, H ∈ MN (−1, 1) having pairwise or-
thogonal rows, have the following properties:
√
(1) The set of Hadamard matrices is XN = MN (−1, 1) ∩ N ON .
(2) In order to have XN ̸= ∅, the matrix size must be N ∈ {2} ∪ 4N.
N/2
(3) For H ∈ MN (−1, 1) we have | det
√H| ≤ N , with equality when√ H is Hadamard.
(4) For U ∈ ON we have ||U ||1 ≤ N N , with equality when H = N U is Hadamard.
Proof. Many things going on here, the idea being as follows:
(1) This is just a reformulation of the Hadamard matrix condition.
(2) This follows by playing with the first 3 rows, exercise for you.
(3) This follows from our definition of the determinant, as a signed volume.
√
(4) This follows from ||U ||2 = N and Cauchy-Schwarz, easy exercise for you. □

All the above is quite interesting, and (1,2) raise the question of finding the correct
generalizations of the Hadamard matrices, at N ∈ / {2} ∪ 4N. But the answer here comes
√ which suggest looking either at the maximizers of | det | on MN (−1, 1), or of
from (3,4),
||.||1 on N ON . By following this latter way, we are led to the following question:
Question 12.30. What are the critical points of the 1-norm on ON ?
And, good news, we can solve this latter question by using the theory of Lagrange
multipliers developed in the above, the result here being as follows:
296 12. HIGHER DERIVATIVES

Theorem 12.31. An orthogonal matrix with nonzero entries is a critical point of

X
||U ||1 = |Uij |
ij
t
precisely when SU is symmetric, where Sij = sgn(Uij ).
Proof. We regard ON as a real algebraic manifold, with coordinates Uij . This man-
ifold consists by definition of the zeroes of the following polynomials:
X
Aij = Uik Ujk − δij
k

Thus U ∈ ON is a critical point of F (U ) = ||U ||1 when the following is satisfied:

dF ∈ span(dAij )
Regarding the space span(dAij ), this consists of the following quantities:
X X
Mij dAij = Mij (Uik dUjk + Ujk dUik )
ij ijk
X X
= (M t U )ij dUij + (M U )ij dUij
ij ij

In order to compute dF , observe first that, with Sij = sgn(Uij ), we have:

q Uij dUij
d|Uij | = d Uij2 = = Sij dUij
|Uij |
P
Thus dF = ij Sij dUij , and so U ∈ ON is a critical point of F precisely when there
exists a matrix M ∈ MN (R) such that the following two conditions are satisfied:
S = M tU , S = MU
Now observe that these two equations can be written as follows:
M t = SU t , M = SU t
Thus, the matrix SU t must be symmetric, as claimed. □
12e. Exercises
This chapter was tough analysis as it comes, and as exercises, we have:
Exercise 12.32. Clarify the Taylor formula at order 3, and its applications.
Exercise 12.33. Learn the Jordan form, as a complement to the linear algebra here.
Exercise 12.34. Harmonic functions in 3 dimensions, and their applications.
Exercise 12.35. Clarify all the details, in relation with Lagrange multipliers.
As bonus exercise, learn the gradient method. You won’t escape from that.
 Part IV

Integration theory
 I’m the left eye
You’re the right
Would it not be madness to fight
We come one
 CHAPTER 13

Multiple integrals

13a. Multiple integrals

Welcome to advanced calculus. We have kept the best for the end, and in this whole
last part of the present book we will learn how to integrate functions of several variables.
The general, fascinating question that we will be interested in is as follows:
Question 13.1. Given a function f : RN → R, how to compute
Z
f (x)dx1 . . . dxN
RN

or at least, what are the rules satisfied by such integrals?

Here we adopt, somehow by definition, the convention that the above integral is con-
structed a bit like the one-variable integrals, via Riemann sums as in chapter 4, by using
this time divisions of the space RN into small N -dimensional cubes. There are of course
some theory and details to be worked out here, not exactly trivial, but since we have not
really done this in chapter 4, nor we will do this here. Let’s focus on computations.

At the first glance solving Question 13.1 looks like an easy task, because we can iterate
one-variable integrations, which are something that we know well, from chapter 4. For
instance the integral of a function f : R2 → R can be computed by using:
Z Z Z
f (z)dz = f (x, y)dxdy
R2 R R

This being said, we are faced right away with a difficulty, when using this method.
Indeed, we can use as well the following rival method, yielding the same answer:
Z Z Z
f (z)dz = f (x, y)dydx
R2 R R

So, which method is the best? Depends on f , of course. However, things do not stop
here, because in certain situations it is better to use polar coordinates, as follows:
Z Z 2π Z ∞
f (z)dz = f (r cos t, r sin t)Jdrdt
R2 0 0
299
300 13. MULTIPLE INTEGRALS

Here the factor on the left is J = dxdy/drdt, which remains to be computed. And for
the picture to be complete, we have as well the following fourth formula:
Z Z ∞ Z 2π
f (z)dz = f (r cos t, r sin t)Jdtdr
R2 0 0

In short, you got my point, I hope, things are quite complicated in several variables,
with the complications starting already in 2 variables. And actually, if you think a
bit about 3 variables, it is quite clear that the above complications can become true
nightmares, leading to long nights spent in computing integrals, there in 3 variables.

So, what to do? Work and patience, of course, and here is our plan:

(1) In this chapter we will get used to the multiple integrals, with some general rules
for their computation, including a rule for computing the above factor J = dxdy/drdt.
Then, we will enjoy all this by computing some integrals over the spheres in RN .

(2) In chapter 14 we will review probability theory, and further develop it, notably
with the theory of normal variables. And finally, in chapters 15-16 we will get back to
physics, with some sharp results in 3 dimensions, and then in infinite dimensions.

This sounds good, but as a matter of doublechecking what we are doing, make sure
that it is wise indeed, let us ask the cat about what he thinks. And cat answers:
Cat 13.2. Not quite sure about your formula
Z Z Z Z
f (x, y)dxdy = f (x, y)dydx
R R R R

and I doubt too that you can properly compute J = dxdy/drdt. Read Rudin.
Oh dear. What can I say. Sure I read Rudin, as a Gen X mathematician, but we
are now well into the 21st century, and shall I go ahead with heavy measure theory, for
having all this properly developed? Not quite sure, I’d rather stick to my plan.

So, ignoring what cat says, but get however a copy of Rudin’s red book, and don’t
forget about Mao’s too, and getting back to our plan, as a first goal, we would like to
compute the factor J = dxdy/drdt. Let us start with something that we know, in 1D:
Proposition 13.3. We have the change of variable formula
Z b Z d
f (x)dx = f (φ(t))φ′ (t)dt
a c
−1 −1
where c = φ (a) and d = φ (b).
 13A. MULTIPLE INTEGRALS 301

Proof. This follows with f = F ′ , via the following differentiation rule:

(F φ)′ (t) = F ′ (φ(t))φ′ (t)
Indeed, by integrating between c and d, we obtain the result. □

In several variables now, we can only expect the above φ′ (t) factor to be replaced by
something similar, a sort of “derivative of φ, arising as a real number”. But this can only
be the Jacobian det(φ′ (t)), and with this in mind, we are led to:
Theorem 13.4. Given a transformation φ = (φ1 , . . . , φN ), we have
Z Z
f (x)dx = f (φ(t))|Jφ (t)|dt
E φ−1 (E)

with the Jφ quantity, called Jacobian, being given by

"  #
dφi
Jφ (t) = det (x)
dxj ij

and with this generalizing the formula from Proposition 13.3.

Proof. This is something quite tricky, the idea being as follows:
(1) Observe first that this generalizes indeed the change of variable formula in 1
dimension, from Proposition 13.3, the point here being that the absolute value on the
derivative appears as to compensate for the lack of explicit bounds for the integral.
(2) In general now, we can first argue that, the formula in the
R statement being linear
in f , we can assume f = 1. Thus we want to prove vol(E) = φ−1 (E) |Jφ (t)|dt, and with
D = φ−1 (E), this amounts in proving vol(φ(D)) = D |Jφ (t)|dt.
R

(3) Now since this

R latter formula is additive with respect to D, it is enough to prove
that vol(φ(D)) = D Jφ (t)dt, for small cubes D, and assuming Jφ > 0. But this follows
by using the definition of the determinant as a volume, as in chapter 9.
(4) The details and computations however are quite non-trivial, and can be found for
instance in Rudin [73]. So, please read that. With this, reading the complete proof of
the present theorem from Rudin, being part of the standard math experience. □

Finally, speaking Rudin, and getting back to Cat 13.2, there are some deep truths
there. But, remember our Advice 1.3, from the beginning of this book. Full rigor does
not guarantee that your computation is correct, you always have to doublecheck. So an
alternative method is that of using less rigor, and more doublechecks at the end. And
this will be our philosophy in what follows, with all our formulae below being correct.
302 13. MULTIPLE INTEGRALS

13b. Spherical coordinates

Time now do some exciting computations, with the technology that we have. In what
regards the applications of Theorem 13.4, these often come via:
Proposition 13.5. We have polar coordinates in 2 dimensions,
(
x = r cos t
y = r sin t
the corresponding Jacobian being J = r.
Proof. This is elementary, the Jacobian being:
d(r cos t) d(r cos t)
dr dt
J =
d(r sin t) d(r sin t)
dr dt
cos t −r sin t
=
sin t r cos t
= r cos2 t + r sin2 t
= r
Thus, we have indeed the formula in the statement. □
We can now compute the Gauss integral, which is the best calculus formula ever:
Theorem 13.6. We have the following formula,
√
Z
2
e−x dx = π
R
called Gauss integral formula.
Proof. Let I be the above integral. By using polar coordinates, we obtain:
Z Z
2 2
2
I = e−x −y dxdy
R R
Z 2π Z ∞
2
= e−r rdrdt
0 0
Z ∞ 2
!′
e−r
= 2π − dr
0 2
  
1
= 2π 0 − −
2
= π
Thus, we are led to the formula in the statement. □
 13B. SPHERICAL COORDINATES 303

Moving now to 3 dimensions, we have here the following result:

Proposition 13.7. We have spherical coordinates in 3 dimensions,

x = r cos s

y = r sin s cos t

z = r sin s sin t

the corresponding Jacobian being J(r, s, t) = r2 sin s.

Proof. The fact that we have indeed spherical coordinates is clear. Regarding now
the Jacobian, this is given by the following formula:
J(r, s, t)
cos s −r sin s 0
= sin s cos t r cos s cos t −r sin s sin t
sin s sin t r cos s sin t r sin s cos t
cos s −r sin s cos s −r sin s
= r2 sin s sin t + r sin s cos t
sin s sin t r cos s sin t sin s cos t r cos s cos t
cos s −r sin s cos s −r sin s
= r sin s sin2 t + r sin s cos2 t
sin s r cos s sin s r cos s
cos s −r sin s
= r sin s(sin2 t + cos2 t)
sin s r cos s
= r sin s × 1 × r
= r2 sin s
Thus, we have indeed the formula in the statement. □
Let us work out now the general spherical coordinate formula, in arbitrary N dimen-
sions. The formula here, which generalizes those at N = 2, 3, is as follows:
Theorem 13.8. We have spherical coordinates in N dimensions,


 x1 = r cos t1

x 2 = r sin t1 cos t2



..
 .
xN −1 = r sin t1 sin t2 . . . sin tN −2 cos tN −1





xN = r sin t1 sin t2 . . . sin tN −2 sin tN −1


the corresponding Jacobian being given by the following formula,

J(r, t) = rN −1 sinN −2 t1 sinN −3 t2 . . . sin2 tN −3 sin tN −2
and with this generalizing the known formulae at N = 2, 3.
304 13. MULTIPLE INTEGRALS

Proof. As before, the fact that we have spherical coordinates is clear. Regarding
now the Jacobian, also as before, by developing over the last column, we have:
JN = r sin t1 . . . sin tN −2 sin tN −1 × sin tN −1 JN −1
+ r sin t1 . . . sin tN −2 cos tN −1 × cos tN −1 JN −1
= r sin t1 . . . sin tN −2 (sin2 tN −1 + cos2 tN −1 )JN −1
= r sin t1 . . . sin tN −2 JN −1
Thus, we obtain the formula in the statement, by recurrence. □

As a comment here, the above convention for spherical coordinates is one among many,
designed to best work in arbitrary N dimensions. Also, in what regards the precise range
of the angles t1 , . . . , tN −1 , we will leave this to you, as an instructive exercise.

As an application, let us compute the volumes of spheres. For this purpose, we must
understand how the products of coordinates integrate over spheres. Let us start with the
case N = 2. Here the sphere is the unit circle T, and with z = eit the coordinates are
cos t, sin t. We can first integrate arbitrary powers of these coordinates, as follows:
Proposition 13.9. We have the following formulae,
Z π/2 Z π/2  π ε(p)
p p!!
cos t dt = sinp t dt =
0 0 2 (p + 1)!!
where ε(p) = 1 if p is even, and ε(p) = 0 if p is odd, and where
m!! = (m − 1)(m − 3)(m − 5) . . .
with the product ending at 2 if m is odd, and ending at 1 if m is even.
Proof. Let us first compute the integral on the left in the statement:
Z π/2
Ip = cosp t dt
0

We do this by partial integration. We have the following formula:

(cosp t sin t)′ = p cosp−1 t(− sin t) sin t + cosp t cos t
= p cosp+1 t − p cosp−1 t + cosp+1 t
= (p + 1) cosp+1 t − p cosp−1 t
By integrating between 0 and π/2, we obtain the following formula:
(p + 1)Ip+1 = pIp−1
 13B. SPHERICAL COORDINATES 305

Thus we can compute Ip by recurrence, and we obtain:

p−1
Ip = Ip−2
p
p−1 p−3
= · Ip−4
p p−2
p−1 p−3 p−5
= · · Ip−6
p p−2 p−4
..
.
p!!
= I1−ε(p)
(p + 1)!!
But I0 = π2 and I1 = 1, so we get the result. As for the second formula, this follows from
the first one, with t = π2 − s. Thus, we have proved both formulae in the statement. □
We can now compute the volume of the sphere, as follows:
Theorem 13.10. The volume of the unit sphere in RN is given by
 π [N/2] 2N
V =
2 (N + 1)!!
with our usual convention N !! = (N − 1)(N − 3)(N − 5) . . .
Proof. Let us denote by B + the positive part of the unit sphere, or rather unit ball
B, obtained by cutting this unit ball in 2N parts. At the level of volumes, we have:
V = 2N V +
We have the following computation, using spherical coordinates:
Z
+
V = 1
B+
Z 1 Z π/2 Z π/2
= ... rN −1 sinN −2 t1 . . . sin tN −2 drdt1 . . . dtN −1
0 0 0
Z 1 Z π/2 Z π/2 Z π/2
N −1 N −2
= r dr sin t1 dt1 . . . sin tN −2 dtN −2 1dtN −1
0 0 0 0
1  π [N/2]
(N − 2)!! (N − 3)!! 2!! 1!!
= × × · ... · ·1
N 2 (N − 1)!! (N − 2)!! 3!! 2!!
1  π [N/2]
 1
= × ×
N 2 (N − 1)!!
 π [N/2] 1
=
2 (N + 1)!!
306 13. MULTIPLE INTEGRALS

Here we have used the following formula, for computing the exponent of π/2:
ε(0) + ε(1) + ε(2) + . . . + ε(N − 2) = 1 + 0 + 1 + . . . + ε(N − 2)
 
N −2
= +1
2
 
N
=
2
Thus, we obtain the formula in the statement. □

As main particular cases of the above formula, we have:

Theorem 13.11. The volumes of the low-dimensional spheres are as follows:
(1) At N = 1, the length of the unit interval is V = 2.
(2) At N = 2, the area of the unit disk is V = π.
(3) At N = 3, the volume of the unit sphere is V = 4π3
π2
(4) At N = 4, the volume of the corresponding unit sphere is V = 2
.
Proof. Some of these results are well-known, but we can obtain all of them as par-
ticular cases of the general formula in Theorem 13.10, as follows:
2
(1) At N = 1 we obtain V = 1 · 1
= 2.
π 4
(2) At N = 2 we obtain V = 2
· 2
= π.
π 8 4π
(3) At N = 3 we obtain V = 2
· 3
= 3
.
π2 16 π2
(4) At N = 4 we obtain V = 4
· 8
= 2
. □

There are many other applications of the above, as we will see in what follows.

13c. Stirling estimates

The formula in Theorem 13.10 is certainly nice, but in practice, we would like to have
estimates for that sphere volumes too. For this purpose, we will need:
Theorem 13.12. We have the Stirling formula
 N
N √
N! ≃ 2πN
e
valid in the N → ∞ limit.
Proof. This is something quite tricky, the idea being as follows:
 13C. STIRLING ESTIMATES 307

(1) Let us first see what we can get with Riemann sums. We have:
N
X
log(N !) = log k
k=1
Z N
≈ log x dx
1
= N log N − N + 1

By exponentiating, this gives the following estimate, which is not bad:

 N
N
N! ≈ ·e
e
(2) We can improve our estimate by replacing the rectangles from the Riemann sum
approach to the integrals by trapezoids. In practice, this gives the following estimate:
N
X
log(N !) = log k
k=1
Z N
log 1 + log N
≈ log x dx +
1 2
log N
= N log N − N + 1 +
2
By exponentiating, this gives the following estimate, which gets us closer:
 N
N √
N! ≈ ·e· N
e
(3) In order to conclude, we must take some kind of mathematical magnifier, and
carefully estimate the error made in (2). Fortunately, this mathematical magnifier exists,
called Euler-Maclaurin formula, and after some computations, this leads to:
 N
N √
N! ≃ 2πN
e
(4) However, all this remains a bit complicated, so we would like to present now
an alternative approach
√ to (3), which also misses some details, but better does the job,
explaining where the 2π factor comes from. First, by partial integration we have:
Z ∞
N! = xN e−x dx
0
308 13. MULTIPLE INTEGRALS

Since the integrand is sharply peaked at x = N , as you can see by computing the
derivative of log(xN e−x ), this suggests writing x = N + y, and we obtain:
log(xN e−x ) = N log x − x
= N log(N + y) − (N + y)
 y
= N log N + N log 1 + − (N + y)
N 
y2

y
≃ N log N + N − − (N + y)
N 2N 2
y2
= N log N − N −
2N
By exponentiating, we obtain from this the following estimate:
 N
N −x N 2
x e ≃ e−y /2N
e
Now by integrating, and using the Gauss formula, we obtain from this:
Z ∞
N! = xN e−x dx
0
Z N N

N 2
≃ e−y /2N dy
−N e
 N Z
N 2
≃ e−y /2N dy
e R
 N
N √ Z
2
= 2N e−z dz
e R
 N
N √
= 2πN
e
Thus, we have proved the Stirling formula, as formulated in the statement. □
With the above formula in hand, we have many useful applications, such as:
Proposition 13.13. We have the following estimate for binomial coefficients,
   N
N 1 1
≃ t 1−t
p
K t (1 − t) 2πt(1 − t)N
in the K ≃ tN → ∞ limit, with t ∈ (0, 1]. In particular we have
4N
 
2N
≃√
N πN
in the N → ∞ limit, for the central binomial coefficients.
 13C. STIRLING ESTIMATES 309

Proof. All this is very standard, by using the Stirling formula etablished above, for
the various factorials which appear, the idea being as follows:
(1) This follows from the definition of the binomial coefficients, namely:
 
N N!
=
K K!(N − K)!
 N N −K
√

N  e K 1 e 1
≃ 2πN √ p
e K 2πK N − K 2π(N − K)
s
NN N
= K N −K
K (N − K) 2πK(N − K)
s
NN N
≃ tN (1−t)N
(tN ) ((1 − t)N ) 2πtN (1 − t)N
 N
1 1
= t 1−t
p
t (1 − t) 2πt(1 − t)N
Thus, we are led to the conclusion in the statement.
(2) This estimate follows from a similar computation, as follows:
 
2N (2N )!
=
N N !N !
2N
√

2N  e 2N 1
≃ 4πN
e N 2πN
N
4
= √
πN
Alternatively, we can take t = 1/2 in (1), then rescale. Indeed, we have:
   N
N 1 1
≃ 1 1/2 1 1/2
q
[N/2] (2) (2) 2π · 21 · 12 · N
r
2
= 2N
πN
Thus with the change N → 2N we obtain the formula in the statement. □

Summarizing, we have so far complete estimates for the factorials. Regarding now the
double factorials, that we will need as well, the result here is as follows:
310 13. MULTIPLE INTEGRALS

Proposition 13.14. We have the following estimate for the double factorials,
 N/2
N
N !! ≃ C
e
√ √
with C = 2 for N even, and C = π for N odd. Alternatively, we have
 N/2
N
(N + 1)!! ≃ D
e
√ √
with D = πN for N even, and D = 2N for N odd.

Proof. Once again this is standard, the idea being as follows:

(1) When N = 2K is even, we have the following computation:

N !! = (2K − 1)(2K − 3) . . . 1
(2K)!
= K
2 K!
2K
√

1 2K  e K 1
≃ K 4πK √
2 e K 2πK
K
√

2K
= 2
e
 N/2
N √
= 2
e

(2) When N = 2K + 1 is odd, we have the following computation:

N !! = (2K)(2K − 2) . . . 2
= 2K K!
K
√

2K
≃ 2πK
e
K+1/2 r K
√
 
2K + 1 e 2K
= 2πK
e 2K + 1 2K + 1
 N/2 r
N e 1 √
≃ · √ · 2πK
e 2K e
 N/2
N √
= π
e
 13C. STIRLING ESTIMATES 311

(3) Back to the case where N = 2K is even, by using (2) we obtain:

(N +1)/2
√

N +1
(N + 1)!! ≃ π
e
N/2 r
N +1 √

N +1
= · π
e e
 N/2  N/2 r
N N +1 N +1 √
= · π
e N e
 N/2 r
N √ N √
≃ e· · π
e e
 N/2
N √
= πN
e
(4) Finally, back to the case where N = 2K + 1 is odd, by using (1) we obtain:
(N +1)/2
√

N +1
(N + 1)!! ≃ 2
e
N/2 r
N +1 √

N +1
= · 2
e e
 N/2  N/2 r
N N +1 N +1 √
= · 2
e N e
 N/2 r
N √ N √
≃ e· · 2
e e
 N/2
N √
= 2N
e
Thus, we have proved the estimates in the statement. □
We can now estimate the volumes of the spheres, as follows:
Theorem 13.15. The volume of the unit sphere in RN is given by
 N/2
2πe 1
V ≃ √
N πN
in the N → ∞ limit.
Proof. We use the formula for V found in Theorem 13.10, namely:
 π [N/2] 2N
V =
2 (N + 1)!!
312 13. MULTIPLE INTEGRALS

In the case where N is even, the estimate goes as follows:

 π N/2 2N
V =
2 (N + 1)!!
 π N/2  e N/2 1
≃ 2N √
2 N πN
 N/2
2πe 1
= √
N πN
In the case where N is odd, the estimate goes as follows:
 π (N −1)/2 2N
V =
2 (N + 1)!!
 π (N −1)/2  e N/2 1
≃ 2N √
2 N 2N
r  N/2
2 2πe 1
= √
π N 2N
 N/2
2πe 1
= √
N πN
Thus, we are led to the uniform formula in the statement. □
Getting back now to our main result so far, Theorem 13.10, we can compute in the
same way the area of the sphere, the result being as follows:
Theorem 13.16. The area of the unit sphere in RN is given by
 π [N/2] 2N
A=
2 (N − 1)!!
with the our usual convention for double factorials, namely:
N !! = (N − 1)(N − 3)(N − 5) . . .
In particular, at N = 2, 3, 4 we obtain respectively A = 2π, 4π, 2π 2 .
Proof. Regarding the first assertion, there is no need to compute again, because the
formula in the statement can be deduced from Theorem 13.10, as follows:
(1) We can either use the “pizza” argument from chapter 1, which shows that the area
and volume of the sphere in RN are related by the following formula:
A=N ·V
Together with the formula in Theorem 13.10 for V , this gives the result.
 13D. SPHERICAL INTEGRALS 313

(2) Or, we can start the computation in the same way as we started the proof of
Theorem 13.10, the beginning of this computation being as follows:
Z π/2 Z π/2
+
vol(S ) = ... sinN −2 t1 . . . sin tN −2 dt1 . . . dtN −1
0 0

Now by comparing with the beginning of the proof of Theorem 13.10, the only thing
that changes is the following quantity, which now dissapears:
Z 1
1
rN −1 dr =
0 N
Thus, we have vol(S + ) = N · vol(B + ), and so we obtain the following formula:
vol(S) = N · vol(B)
But this means A = N · V , and together with the formula in Theorem 13.10 for V ,
this gives the result. As for the last assertion, this can be either worked out directly, or
deduced from the results for volumes that we have so far, by multiplying by N . □

13d. Spherical integrals

Let us discuss now the computation of the arbitrary integrals over the sphere. We will
need a technical result extending Proposition 13.9, as follows:
Theorem 13.17. We have the following formula,
Z π/2  π ε(p)ε(q)
p q p!!q!!
cos t sin t dt =
0 2 (p + q + 1)!!
where ε(p) = 1 if p is even, and ε(p) = 0 if p is odd, and where
m!! = (m − 1)(m − 3)(m − 5) . . .
with the product ending at 2 if m is odd, and ending at 1 if m is even.
Proof. We use the same idea as in Proposition 13.9. Let Ipq be the integral in the
statement. In order to do the partial integration, observe that we have:
(cosp t sinq t)′ = p cosp−1 t(− sin t) sinq t
+ cosp t · q sinq−1 t cos t
= −p cosp−1 t sinq+1 t + q cosp+1 t sinq−1 t
By integrating between 0 and π/2, we obtain, for p, q > 0:
pIp−1,q+1 = qIp+1,q−1
314 13. MULTIPLE INTEGRALS

Thus, we can compute Ipq by recurrence. When q is even we have:

q−1
Ipq = Ip+2,q−2
p+1
q−1 q−3
= · Ip+4,q−4
p+1 p+3
q−1 q−3 q−5
= · · Ip+6,q−6
p+1 p+3 p+5
..
= .
p!!q!!
= Ip+q
(p + q)!!

But the last term comes from Proposition 13.9, and we obtain the result:

p!!q!!
Ipq = Ip+q
(p + q)!!
p!!q!!  π ε(p+q) (p + q)!!
=
(p + q)!! 2 (p + q + 1)!!
 π ε(p)ε(q) p!!q!!
=
2 (p + q + 1)!!

Observe that this gives the result for p even as well, by symmetry. Indeed, we have
Ipq = Iqp , by using the following change of variables:
π
t= −s
2
In the remaining case now, where both p, q are odd, we can use once again the formula
pIp−1,q+1 = qIp+1,q−1 established above, and the recurrence goes as follows:

q−1
Ipq = Ip+2,q−2
p+1
q−1 q−3
= · Ip+4,q−4
p+1 p+3
q−1 q−3 q−5
= · · Ip+6,q−6
p+1 p+3 p+5
..
= .
p!!q!!
= Ip+q−1,1
(p + q − 1)!!
 13D. SPHERICAL INTEGRALS 315

In order to compute the last term, observe that we have:

Z π/2
Ip1 = cosp t sin t dt
0
Z π/2
1
= − (cosp+1 t)′ dt
p+1 0
1
=
p+1
Thus, we can finish our computation in the case p, q odd, as follows:
p!!q!!
Ipq = Ip+q−1,1
(p + q − 1)!!
p!!q!! 1
= ·
(p + q − 1)!! p + q
p!!q!!
=
(p + q + 1)!!
Thus, we obtain the formula in the statement, the exponent of π/2 appearing there
being ε(p)ε(q) = 0 · 0 = 0 in the present case, and this finishes the proof. □
We can now integrate over the spheres, as follows:
Theorem 13.18. The polynomial integrals over the unit sphere SRN −1 ⊂ RN , with
respect to the normalized, mass 1 measure, are given by the following formula,
(N − 1)!!k1 !! . . . kN !!
Z
xk11 . . . xkNN dx =
N −1
SR (N + Σki − 1)!!
valid when all exponents ki are even. If an exponent ki is odd, the integral vanishes.
Proof. Assume first that one of the exponents ki is odd. We can make then the
following change of variables, which shows that the integral in the statement vanishes:
xi → −xi
Assume now that all exponents ki are even. As a first observation, the result holds
indeed at N = 2, due to the formula from Theorem 13.17, which reads:
Z π/2  π ε(p)ε(q) p!!q!! p!!q!!
cosp t sinq t dt = =
0 2 (p + q + 1)!! (p + q + 1)!!
In the general case now, where the dimension N ∈ N is arbitrary, the integral in the
statement can be written in spherical coordinates, as follows:
2N π/2
Z Z π/2
I= ... xk11 . . . xkNN J dt1 . . . dtN −1
A 0 0
316 13. MULTIPLE INTEGRALS

Here A is the area of the sphere, J is the Jacobian, and the 2N factor comes from the
restriction to the 1/2N part of the sphere where all the coordinates are positive. According
to Theorem 13.16, the normalization constant in front of the integral is:
 [N/2]
2N 2
= (N − 1)!!
A π
As for the unnormalized integral, this is given by:
Z π/2 Z π/2
′
I = ... (cos t1 )k1 (sin t1 cos t2 )k2
0 0
..
.
(sin t1 sin t2 . . . sin tN −2 cos tN −1 )kN −1
(sin t1 sin t2 . . . sin tN −2 sin tN −1 )kN
sinN −2 t1 sinN −3 t2 . . . sin2 tN −3 sin tN −2
dt1 . . . dtN −1
By rearranging the terms, we obtain:
Z π/2
′
I = cosk1 t1 sink2 +...+kN +N −2 t1 dt1
0
Z π/2
cosk2 t2 sink3 +...+kN +N −3 t2 dt2
0
..
.
Z π/2
coskN −2 tN −2 sinkN −1 +kN +1 tN −2 dtN −2
0
Z π/2
coskN −1 tN −1 sinkN tN −1 dtN −1
0
Now by using the above-mentioned formula at N = 2, this gives:
k1 !!(k2 + . . . + kN + N − 2)!!  π ε(N −2)
I′ =
(k1 + . . . + kN + N − 1)!! 2
k2 !!(k3 + . . . + kN + N − 3)!! π ε(N −3)

(k2 + . . . + kN + N − 2)!! 2
..
.
k !!(kN −2 + k + 1)!!  π ε(1)
N −1 N
(kN −2 + kN −1 + lN + 2)!! 2
kN −1 !!kN !!  π ε(0)
(kN −1 + kN + 1)!! 2
 13D. SPHERICAL INTEGRALS 317

Now let F be the part involving the double factorials, and P be the part involving the
powers of π/2, so that I ′ = F · P . Regarding F , by cancelling terms we have:

k1 !! . . . kN !!
F =
(Σki + N − 1)!!

π [N/2]


As in what regards P , by summing the exponents, we obtain P = 2
. We can
now put everything together, and we obtain:

2N
I = ×F ×P
A
 [N/2]
2 k1 !! . . . kN !!  π [N/2]
= (N − 1)!! × ×
π (Σki + N − 1)!! 2
(N − 1)!!k1 !! . . . kN !!
=
(Σki + N − 1)!!

Thus, we are led to the conclusion in the statement. □

We have the following useful generalization of the above formula:

Theorem 13.19. We have the following integration formula over the sphere SRN −1 ⊂
RN , with respect to the normalized, mass 1 measure, valid for any exponents ki ∈ N,
 Σ(k1 ,...,kN )
(N − 1)!!k1 !! . . . kN !!
Z
2
|xk11 . . . xkNN | dx =
SRN −1 π (N + Σki − 1)!!

with Σ = [odds/2] if N is odd and Σ = [(odds + 1)/2] if N is even, where “odds” denotes
the number of odd numbers in the sequence k1 , . . . , kN .

Proof. As before, the formula holds at N = 2, due to Theorem 13.17. In general,

the integral in the statement can be written in spherical coordinates, as follows:
π/2 π/2
2N
Z Z
I= ... xk11 . . . xkNN J dt1 . . . dtN −1
A 0 0

Here A is the area of the sphere, J is the Jacobian, and the 2N factor comes from
the restriction to the 1/2N part of the sphere where all the coordinates are positive. The
normalization constant in front of the integral is, as before:
 [N/2]
2N 2
= (N − 1)!!
A π
318 13. MULTIPLE INTEGRALS

As for the unnormalized integral, this can be written as before, as follows:

Z π/2
′
I = cosk1 t1 sink2 +...+kN +N −2 t1 dt1
0
Z π/2
cosk2 t2 sink3 +...+kN +N −3 t2 dt2
0
..
.
Z π/2
coskN −2 tN −2 sinkN −1 +kN +1 tN −2 dtN −2
0
Z π/2
coskN −1 tN −1 sinkN tN −1 dtN −1
0

Now by using the formula at N = 2, we get:

 δ(k1 ,k2 +...+kN +N −2)
′ π k1 !!(k2 + . . . + kN + N − 2)!! 2
I = ·
2 (k1 + . . . + kN + N − 1)!! π
 δ(k2 ,k3 +...+kN +N −3)
π k2 !!(k3 + . . . + kN + N − 3)!! 2
·
2 (k2 + . . . + kN + N − 2)!! π
..
.
 δ(kN −2 ,kN −1 +kN +1)
π kN −2 !!(kN −1 + kN + 1)!! 2
·
2 (kN −2 + kN −1 + kN + 2)!! π
 δ(kN −1 ,kN )
π kN −1 !!kN !! 2
·
2 (kN −1 + kN + 1)!! π
In order to compute this quantity, let us denote by F the part involving the double
factorials, and by P the part involving the powers of π/2, so that we have:

I′ = F · P

Regarding F , there are many cancellations there, and we end up with:

k1 !! . . . kN !!
F =
(Σki + N − 1)!!
As in what regards P , the δ exponents on the right sum up to the following number:
N
X −1
∆(k1 , . . . , kN ) = δ(ki , ki+1 + . . . + kN + N − i − 1)
i=1
 13D. SPHERICAL INTEGRALS 319

In other words, with this notation, the above formula reads:

 π N −1  ∆(k1 ,...,kN )
′ k1 !!k2 !! . . . kN !! 2
I =
2 (k1 + . . . + kN + N − 1)!! π
 ∆(k1 ,...,kN )−N +1
2 k1 !!k2 !! . . . kN !!
=
π (k1 + . . . + kN + N − 1)!!
 Σ(k1 ,...,kN )−[N/2]
2 k1 !!k2 !! . . . kN !!
=
π (k1 + . . . + kN + N − 1)!!
Here the formula relating ∆ to Σ follows from a number of simple observations, the
first of which being the fact that, due to obvious parity reasons, the sequence of δ numbers
appearing in the definition of ∆ cannot contain two consecutive zeroes. Together with
I = (2N /V )I ′ , this gives the formula in the statement. □
Let us discuss as well the complex versions of the above results. We have the following
variation of the formula in Theorem 13.19, dealing with the complex sphere:
Theorem 13.20. We have the following integration formula over the complex sphere
SCN −1 ⊂ CN , with respect to the normalized uniform measure,
(N − 1)!k1 ! . . . kn !
Z
|z1 |2k1 . . . |zN |2kN dz = P
SCN −1 (N + ki − 1)!
valid for any exponents ki ∈ N. As for the other polynomial integrals in z1 , . . . , zN and
their conjugates z̄1 , . . . , z̄N , these all vanish.
Proof. Consider an arbitrary polynomial integral over SCN −1 , containing the same
number of plain and conjugated variables, as to not vanish trivially, written as follows:
Z
I= zi1 z̄i2 . . . zi2k−1 z̄i2k dz
SCN −1

By using transformations of type p → λp with |λ| = 1, we see that this integral I

vanishes, unless each za appears as many times as z̄a does, and this gives the last assertion.
So, assume now that we are in the non-vanishing case. Then the ka copies of za and the
ka copies of z̄a produce by multiplication a factor |za |2ka , so we have:
Z
I= |z1 |2k1 . . . |zN |2kN dz
SCN −1

Now by using the standard identification SCN −1 ≃ SR2N −1 , we obtain:

Z
I = (x21 + y12 )k1 . . . (x2N + yN
2 kN
) d(x, y)
SR2N −1
X k1  kN  Z
1 −2r1 2r1 N −2rN 2rN
= ... x2k
1 y1 . . . x2k
N yN d(x, y)
r ...r
r 1 rN S 2N
R
−1
1 N
320 13. MULTIPLE INTEGRALS

By using the formula in Theorem 13.19, we obtain:

I
X k1  kN  (2N − 1)!!(2r1 )!! . . . (2rN )!!(2k1 − 2r1 )!! . . . (2kN − 2rN )!!
= ... P
r1 ...rN
r1 rN (2N + 2 ki − 1)!!
X k1  kN  2N −1 (N − 1)! Q(2ri )!/(2ri ri !) Q(2ki − 2ri )!/(2ki −ri (ki − ri )!)
= ... P P
r1 ...rN
r1 rN 2N + ki −1 (N + ki − 1)!
X k1  kN  (N − 1)!(2r1 )! . . . (2rN )!(2k1 − 2r1 )! . . . (2kN − 2rN )!
= ... P
ki (N +
P
r ...r
r1 rN 4 ki − 1)!r1 ! . . . rN !(k1 − r1 )! . . . (kN − rN )!
1 N

Now observe that can rewrite this quantity in the following way:
I
X k1 ! . . . kN !(N − 1)!(2r1 )! . . . (2rN )!(2k1 − 2r1 )! . . . (2kN − 2rN )!
= P P
r1 ...rN
4 ki (N + ki − 1)!(r1 ! . . . rN !(k1 − r1 )! . . . (kN − rN )!)2
X 2r1 2k1 − 2r1  X 2rN 2kN − 2rN  (N − 1)!k1 ! . . . kN !
= ... P
ki (N +
P
r
r1 k1 − r1 r
r N kN − rN 4 ki − 1)!
1 N

(N − 1)!k1 ! . . . kN !
= 4k1 × . . . × 4kN × P P
4 ki (N + ki − 1)!
(N − 1)!k1 ! . . . kN !
= P
(N + ki − 1)!
Thus, we are led to the formula in the statement. □
We will see applications of all this in the next chapter, with some quite conceptual
results regarding the spherical coordinates, in the N → ∞ limit, obtained by processing
the above results, by using standard tools from probability.
13e. Exercises
Here are some exercises on the above, often insisting on missing details:
Exercise 13.21. Clarify the details of the change of variable formula.
Exercise 13.22. Find the ranges of angles in the spherical coordinate formula.
Exercise 13.23. Further refine the Stirling formula, with more terms.
Exercise 13.24. Prove r 2rr 2k−2r



= 4k , used in the above, at the end.
P
k−r

As a bonus exercise, compute areas and volumes, as many as you can.

 CHAPTER 14

Normal variables

14a. Probability basics

Good news, with the calculus that we know, we can have a crash course in advanced
probability. We have already met some probability in this book, scattered at the end
of chapters 4, 6, R7, and we√will first review that material. Then, helped by the Gauss
2
integral formula R e−x = π that we just learned, and by multivariable integration in
general, we will develop the theory of normal variables, real and complex, which is central
to advanced probability. And then, there will certainly be more, all interesting things.

Sounds exciting, doesn’t it. Cat however seems unfazed, and declares:
Cat 14.1. Probability is the same thing as measure theory. Read Rudin.
Damn cat, looks like we disagree more and more as time goes by, especially on these
multivariable integration topics. So, let me ask you cat, how many times did you come
upon a function which is not integrable? How many times did you fail catching a mouse,
due to a failure of Fubini? What about catching birds, does the Zorn lemma really help
there? Things are nice and smooth in life, or at least that’s my belief.
Cat 14.2. Yes for mice and birds, but electrons can be quite tricky.
Humm, good point, but let’s leave electrons for later, for chapter 16. So, forgetting
now about philosophy, and pedagogy matters, but dear reader feel free to have your own
opinion here, and why not agreeing with cat, and going ahead with our plan, let us first
review the few things that we know about probability, from chapters 4, 6, 7.

There will be of course a bit of redundancy, but always good to talk again about that
things, with a bit less details of course, this time. As a starting point, we have:
Definition 14.3. Let X be a probability space, that is, a space with a probability
measure, and with the corresponding integration denoted E, and called expectation.
(1) The random variables are the real functions f ∈ L∞ (X).
k
(2) The moments of such a variable are the numbers Mk (f ) = E(fR ).k
(3) The law of such a variable is the measure given by Mk (f ) = R x dµf (x).
321
322 14. NORMAL VARIABLES

Here, as explained in chapter 7, the fact that µf as above exists indeed is not exactly
trivial. But we can do this by looking at formulae of the following type:
Z
E(φ(f )) = φ(x)dµf (x)
R
Indeed, having this for monomials φ(x) = xn , as above, is the same as having it
for polynomials φ ∈ R[X], which in turn is the same as having it for the characteristic
functions φ = χI of measurable sets I ⊂ R. Thus, in the end, what we need is:
P (f ∈ I) = µf (I)
But this formula can serve as a definition for µf , and we are done.
Regarding now independence, as explained in chapter 7, we can formulate here:
Definition 14.4. Two variables f, g ∈ L∞ (X) are called independent when
E(f k g l ) = E(f k ) E(g l )
happens, for any k, l ∈ N.
Again, this definition hides some non-trivial things, the idea being a bit as before,
namely that of looking at formulae of the following type:
E[φ(f )ψ(g)] = E[φ(f )] E[ψ(g)]
To be more precise, passing as before from monomials to polynomials, then to char-
acteristic functions, we are led to the usual definition of independence, namely:
P (f ∈ I, g ∈ J) = P (f ∈ I) P (g ∈ J)
As a first result now, that we know well from chapter 7, we have:
Theorem 14.5. Assuming that f, g ∈ L∞ (X) are independent, we have
µf +g = µf ∗ µg
where ∗ is the convolution of real probability measures.
Proof. We have the following computation, using the independence of f, g:
Z X k 
k k
x dµf +g (x) = E((f + g) ) = Mr (f )Mk−r (g)
R r
r
On the other hand, we have as well the following computation:
Z Z
k
x d(µf ∗ µg )(x) = (x + y)k dµf (x)dµg (y)
R R×R
X k 
= Mr (f )Mk−r (g)
r
r
Thus µf +g and µf ∗ µg have the same moments, so they coincide, as claimed. □
 14A. PROBABILITY BASICS 323

As a second result on independence, which is more advanced, we have:

Theorem 14.6. Assuming that f, g ∈ L∞ (X) are independent, we have
Ff +g = Ff Fg
where Ff (x) = E(eixf ) is the Fourier transform.
Proof. This is something that we know too from chapter 7, coming from:
Z
Ff +g (x) = eixz d(µf ∗ µg )(z)
ZR
= eix(z+t) dµf (z)dµg (t)
ZR×R Z
= e dµf (z) eixt dµg (t)
ixz
R R
= Ff (x)Fg (x)
Thus, we are led to the conclusion in the statement. □
So long for laws, moments and independence, at the theoretical level. In what follows
Definitions 14.3 and 14.4 and Theorems 14.5 and 14.6 are what we need, I mean we will
certainly get away with that, as general theory, and this no matter what cat says.

However, still staying general, we will need as well, from time to time, some key
complex analysis results that we established in chapter 6. First, we have:
Theorem 14.7. The density of a real probability measure µ can be recaptured from
the sequence of moments {Mk }k≥0 via the Stieltjes inversion formula
1
dµ(x) = lim − Im (G(x + it)) · dx
t↘0 π
where the function on the right, given in terms of moments by
G(ξ) = ξ −1 + M1 ξ −2 + M2 ξ −3 + . . .
is the Cauchy transform of the measure µ.
Proof. This is something quite subtle and heavy, and for the full proof, along with
some basic applications, we refer to chapter 6, the idea being as follows:
(1) Regarding the proof, the Cauchy transform of our measure µ is given by:
∞ Z
−1
X
−k 1
G(ξ) = ξ Mk ξ = dµ(y)
k=0 R ξ − y
324 14. NORMAL VARIABLES

Now with ξ = x + it, we obtain from this the following formula:

Z  
1
Im(G(x + it)) = Im dµ(y)
R x − y + it
Z
t
= − 2 2
dµ(y)
R (x − y) + t

By integrating over [a, b] we obtain, with the change of variables x = y + tz:

Z b Z Z (b−y)/t
t
Im(G(x + it))dx = − 2 2
t dz dµ(y)
a R (a−y)/t (tz) + t
Z Z (b−y)/t
1
= − 2
dz dµ(y)
R (a−y)/t 1 + z
Z  
b−y a−y
= − arctan − arctan dµ(y)
R t t

(2) The point now is that with t ↘ 0 we have the following estimates:
π π

 2
−2 =0 (y < a)
π π
2 − 0 = 2

    (y = a)
b−y a−y π π
lim arctan − arctan = 2 − (− 2 ) = π (a < y < b)
t↘0 t t  π π



 0 − (− 2 ) = 2 (y = b)
 π π
− 2 − (− 2 ) = 0 (y > b)

We therefore obtain the following formula, which proves our result:

Z b  
µ(a) + µ(b)
lim Im(G(x + it))dx = −π µ(a, b) +
t↘0 a 2

(3) As applications, the first thing that we did in chapter

6 was to find the measure
1 2k
having as even moments the Catalan numbers, Ck = k+1 k
, and having all odd moments
0. We found the following measure, called Wigner semicircle law on [−2, 2]:
1√
γ1 = 4 − x2 dx
2π
(4) We also computed the measure having as moments the Catalan numbers, Mk = Ck .
We found the following measure, called Marchenko-Pastur law on [0, 4]:
1 √ −1
π1 = 4x − 1 dx
2π
 14A. PROBABILITY BASICS 325

(5) Next, we
found that the measure having as moments the central binomial coeffi-
cients, Dk = 2kk
, is the following measure, called arcsine law on [0, 4]:
1
α1 = p dx
π x(4 − x)
k


(6) Finally, for the middle binomial coefficients, Ek = [k/2]
, we found the following
law on [−2, 2], called modified arcsine law on [−2, 2]:
r
1 2+x
σ1 = dx
2π 2 − x
(7) All this is very nice, and as already mentioned in chapter 6, these measures are
those appearing via random walks on basic graphs. In addition, these measures are as
well the main laws in Random Matrix Theory (RMT). More on them later. □

We have as well the following result, also from chapter 6, which can be useful too:
Theorem 14.8. A sequence of numbers M0 , M1 , M2 , M3 , . . . ∈ R, with M0 = 1, is the
series of moments of a real probability measure µ precisely when:
M0 M1 M2
M0 M1
M0 ≥ 0 , ≥0 , M1 M2 M3 ≥ 0 , ...
M1 M2
M2 M3 M4
That is, the associated Hankel determinants must be all positive.
Proof. This is something a bit heavier, and as a first observation, the positivity
conditions in the statement tell us that the following linear forms must be positive:
n
X
ci c̄j Mi+j ≥ 0
i,j=1

But this is something very classical, in one sense the result being elementary, coming
from the following computation, which shows that we have positivity indeed:
Z n 2 n
Z X n
X X
i
ci x dµ(x) = ci c̄j xi+j dµ(x) = ci c̄j Mi+j
R i=1 R i,j=1 i,j=1

As for the other sense, here the result comes once again from the above formula, this
time via some standard study, inspired from the positivity results from chapter 12. □

All this is very nice, and we have some interesting theory going on. Let us discuss
now some illustrations. We will first talk about discrete probability. We have:
326 14. NORMAL VARIABLES

Definition 14.9. The Poisson law of parameter t > 0 is the measure

X tk
pt = e−t δk
k∈N
k!

with the letter “p” standing for Poisson.

We have already talked about these laws, first in chapter 4, with an elementary discus-
sion, using binomials and factorials, and then in chapter 7 too. So, let us quickly review
what we know. Going directly for the kill, Fourier transform computation, we have:

Theorem 14.10. The Fourier transform of pt is given by:

Fpt (y) = exp (eiy − 1)t

In particular we have ps ∗ pt = ps+t , called convolution semigroup property.

Proof. We have indeed the following computation:

X tk
Fpt (y) = e−t Fδk (y)
k
k!
X tk
= e−t eiky
k
k!
X (eiy t)k
= e−t
k
k!
= exp (eiy − 1)t

As for the second assertion, this follows from the fact that log Fpt is linear in t, via
the linearization property for the convolution from Theorem 14.6. □

The above result suggests that the laws pt should appear as some sort of exponentials
with respect to convolution, and yes indeed, this is the case, as shown by:

Proposition 14.11. The Poisson laws appear as formal exponentials

X tk (δ1 − δ0 )∗k
pt =
k
k!

with respect to the convolution of measures ∗.

 14A. PROBABILITY BASICS 327

Proof. By using the binomial formula, the measure on the right is:
X tk X k!
µ = (−1)s δr
k
k! r+s=k r!s!
X tr δr X (−1)s
=
r
r! s s!
1 X tr δr
=
e r
r!
= pt
Thus, we are led to the conclusion in the statement. □
As a main result now, we have the Poisson Limit Theorem, as follows:
Theorem 14.12 (PLT). We have the following convergence, in moments,
  ∗n
t t
1− δ0 + δ1 → pt
n n
for any t > 0.
Proof. Indeed, if we denote by νn the measure under the convolution sign, we have
the following computation, for the Fourier transform of the limit:
 
iry t t
Fδr (y) = e =⇒ Fνn (y) = 1 − + eiy
n n
  n
t t iy
=⇒ Fνn∗n (y) = 1− + e
n n
 iy
n
(e − 1)t
=⇒ Fνn∗n (y) = 1 +
n
iy


=⇒ F (y) = exp (e − 1)t
Thus, we obtain indeed the Fourier transform of pt , as desired. □
Finally, one more thing that we know about the Poisson laws are some interesting
formulae for their moments, from chapter 4. The result there was as follows:
Theorem 14.13. The moments of p1 are the Bell numbers,
Mk (p1 ) = |P (k)|
where P (k) is the set of partitions of {1, . . . , k}. More generally, we have
X
Mk (pt ) = t|π|
π∈P (k)

for any t > 0, where |.| is the number of blocks.

328 14. NORMAL VARIABLES

Proof. We know that the moments of p1 are given by the following formula:
1 X rk
Mk =
e r r!
We therefore have the following recurrence formula for these moments:
k
1 X rk

1
Mk+1 = 1+
e r r! r
1 X rk X k −s
 
= r
e r r! s s
X k 
= Mk−s
s
s
But the Bell numbers Bk = |P (k)| satisfy the same recurrence, so we have Mk = Bk ,
as claimed. Next, we know that the moments of pt with t > 0 are given by:
X tr rk
Nk = e−t
r
r!
We therefore have the following recurrence formula for these moments:
X tr+1 rk  k
−t 1
Nk+1 = e 1+
r
r! r
X tr+1 rk X k 

−t
= e r−s
r
r! s
s
X k 
= t Nk−s
s
s
|π|
P
But the numbers Sk = π∈P (k) t are easily seen to satisfy the same recurrence, with
the same initial values, namely t and t + t2 , so we have Nk = Sk , as claimed. □
Summarizing, we know so far what probability theory is, modulo perhaps some an-
noying foundational details, that do not seem needed for computations. Then we have
some powerful tools, namely convolution, Fourier, Stieltjes, and Hankel determinants if
needed. Then, as a main achievement, we have a full theory in the main discrete case,
that of the Poisson laws. And finally, in the continuous case, we came across a number
of interesting laws, namely Wigner, Marchenko-Pastur, arcsine, and modified arcsine.

This is not bad, as a start. In what follows we will first do the necessary, namely
mirror what we know about the Poisson laws, with a complete study of the normal laws,
which are the central objects in continuous probability. And then, we will go for more.
 14B. NORMAL VARIABLES 329

14b. Normal variables

In order to introduce the normal variables, we will need the Gauss integral formula,
established in chapter 13. So, let us recall from there that we have:
Theorem 14.14. We have the following formula,
√
Z
2
e−x dx = π
R
called Gauss integral formula.
Proof. Let I be the above integral. By using polar coordinates, we obtain:
Z Z
2 2
2
I = e−x −y dxdy
R R
Z 2π Z ∞
2
= e−r rdrdt
0 0
1
= 2π ×
2
Thus, we are led to the formula in the statement. □
As a main application of the Gauss formula, we can now formulate:
Definition 14.15. The normal law of parameter 1 is the following measure:
1 2
g1 = √ e−x /2 dx
2π
More generally, the normal law of parameter t > 0 is the following measure:
1 −x2 /2t
gt = √ e dx
2πt
These are also called Gaussian distributions, with “g” standing for Gauss.
Observe that the above laws have indeed mass√ 1, as they should. This follows indeed
from the Gauss formula, which gives, with x = 2t y:
2√
Z Z
−x2 /2t
e dx = e−y 2t dy
R R
√ Z −y2
= 2t e dy
R
√ √
= 2t × π
√
= 2πt
Generally speaking, the normal laws appear as bit everywhere, in real life. The reasons
behind this phenomenon come from the Central Limit Theorem (CLT), that we will
explain in a moment, after developing some general theory. As a first result, we have:
330 14. NORMAL VARIABLES

Proposition 14.16. We have the variance formula

V (gt ) = t
valid for any t > 0.
Proof. The first moment is 0, because our normal law gt is centered. As for the
second moment, this can be computed as follows:
Z
1 2
M2 = √ x2 e−x /2t dx
2πt R
Z ′
1  2
= √ (tx) −e−x /2t dx
2πt R
Z
1 2
= √ te−x /2t dx
2πt R
= t
We conclude from this that the variance is V = M2 = t. □
Here is another result, which is the key one for the study of the normal laws:
Theorem 14.17. We have the following formula, valid for any t > 0:
2 /2
Fgt (x) = e−tx
In particular, the normal laws satisfy gs ∗ gt = gs+t , for any s, t > 0.
Proof. The Fourier transform formula can be established as follows:
Z
1 2
Fgt (x) = √ e−y /2t+ixy dy
2πt R
1
Z √ √ 2 2
= √ e−(y/ 2t− t/2ix) −tx /2 dy
2πt R
1
Z
2 2 √
= √ e−z −tx /2 2tdz
2πt R
Z
1 −tx2 /2 2
= √ e e−z dz
π R
1 −tx2 /2 √
= √ e · π
π
2 /2
= e−tx
As for the last assertion, this follows from the fact that log Fgt is linear in t. □
We are now ready to state and prove the CLT, as follows:
 14B. NORMAL VARIABLES 331

Theorem 14.18 (CLT). Given random variables f1 , f2 , f3 , . . . ∈ L∞ (X) which are

i.i.d., centered, and with variance t > 0, we have, with n → ∞, in moments,
n
1 X
√ fi ∼ gt
n i=1
2
where gt is the Gaussian law of parameter t, having as density √ 1 e−y /2t dy.
2πt

Proof. We use the Fourier transform, which is by definition given by:

Ff (x) = E(eixf )

In terms of moments, we have the following formula:

∞
!
X (ixf )k
Ff (x) = E
k=0
k!
∞
X (ix)k E(f k )
=
k=0
k!
∞ k
X i Mk (f ) k
= x
k=0
k!

Thus, the Fourier transform of the variable in the statement is:

  n
x
F (x) = Ff √
n
 2
n
tx −2
= 1− + O(n )
2n
n
tx2

≃ 1−
2n
2 /2
≃ e−tx

But this latter function being the Fourier transform of gt , we obtain the result. □

Let us discuss now some further properties of the normal law. We first have:

Proposition 14.19. The even moments of the normal law are the numbers

Mk (gt ) = tk/2 × k!!

where k!! = (k − 1)(k − 3)(k − 5) . . . , and the odd moments vanish.

332 14. NORMAL VARIABLES

Proof. We have the following computation, valid for any integer k ∈ N:

Z
1 2
Mk = √ y k e−y /2t dy
2πt R
Z ′
1  2
= √ (ty k−1 ) −e−y /2t dy
2πt R
Z
1 2
= √ t(k − 1)y k−2 e−y /2t dy
2πt R
Z
1 2
= t(k − 1) × √ y k−2 e−y /2t dy
2πt R
= t(k − 1)Mk−2
Now recall from the proof of Proposition 14.16 that we have M0 = 1, M1 = 0. Thus
by recurrence, we are led to the formula in the statement. □

We have the following alternative formulation of the above result:

Proposition 14.20. The moments of the normal law are the numbers
Mk (gt ) = tk/2 |P2 (k)|
where P2 (k) is the set of pairings of {1, . . . , k}.
Proof. Let us count the pairings of {1, . . . , k}. In order to have such a pairing, we
must pair 1 with one of the numbers 2, . . . , k, and then use a pairing of the remaining
k − 2 numbers. Thus, we have the following recurrence formula:
|P2 (k)| = (k − 1)|P2 (k − 2)|
As for the initial data, this is P1 = 0, P2 = 1. Thus, we are led to the result. □

We are not done yet, and here is one more improvement of the above:
Theorem 14.21. The moments of the normal law are the numbers
X
Mk (gt ) = t|π|
π∈P2 (k)

where P2 (k) is the set of pairings of {1, . . . , k}, and |.| is the number of blocks.
Proof. This follows indeed from Proposition 14.20, because the number of blocks of
a pairing of {1, . . . , k} is trivially k/2, independently of the pairing. □

Let us discuss now the complex analogues of all this, with a notion of complex normal,
or Gaussian law. To start with, we have the following definition:
 14B. NORMAL VARIABLES 333

Definition 14.22. The complex normal, or Gaussian law of parameter t > 0 is

 
1
Gt = law √ (a + ib)
2
where a, b are independent, each following the law gt .
In short, the complex normal laws appear as natural complexifications of the real
normal laws. As in the real case, these measures form convolution semigroups:
Proposition 14.23. The complex Gaussian laws have the property
Gs ∗ Gt = Gs+t
for any s, t > 0, and so they form a convolution semigroup.
Proof. This follows indeed from the real result, namely gs ∗ gt = gs+t , established in
Theorem 14.17, simply by taking the real and imaginary parts. □
We have as well the following complex analogue of the CLT:
Theorem 14.24 (CCLT). Given complex variables f1 , f2 , f3 , . . . ∈ L∞ (X) which are
i.i.d., centered, and with common variance t > 0, we have
n
1 X
√ fi ∼ Gt
n i=1
with n → ∞, in moments.
Proof. This follows indeed from the real CLT, established in Theorem 14.18, simply
by taking the real and imaginary parts of all the variables involved. □
Regarding now the moments, the situation here is more complicated than in the real
case, because in order to have good results, we have to deal with both the complex
variables, and their conjugates. Let us formulate the following definition:
Definition 14.25. The moments a complex variable f ∈ L∞ (X) are the numbers
Mk = E(f k )
depending on colored integers k = ◦ • • ◦ . . . , with the conventions
f ∅ = 1 , f ◦ = f , f • = f¯
and multiplicativity, in order to define the colored powers f k .
Observe that, since f, f¯ commute, we can permute terms, and restrict the attention
to exponents of type k = . . . ◦ ◦ ◦ • • • . . . , if we want to. However, our results about the
complex Gaussian laws, and other complex laws too, later on, will actually look better
without doing is, so we will use Definition 14.25 as stated. We first have:
334 14. NORMAL VARIABLES

Theorem 14.26. The moments of the complex normal law are given by
(
tp p! (k uniform, of length 2p)
Mk (Gt ) =
0 (k not uniform)

where k = ◦ • • ◦ . . . is called uniform when it contains the same number of ◦ and •.

Proof. We must compute the moments, with respect to colored integer exponents
k = ◦ • • ◦ . . . , of the variable from Definition 14.22, namely:
1
f = √ (a + ib)
2
We can assume that we are in the case t = 1, and the proof here goes as follows:
(1) As a first observation, in the case where our exponent k = ◦ • • ◦ . . . is not uniform,
a standard rotation argument shows that the corresponding moment of f vanishes. To
be more precise, the variable f ′ = wf is complex Gaussian too, for any complex number
w ∈ T, and from Mk (f ) = Mk (f ′ ) we obtain Mk (f ) = 0, in this case.
(2) In the uniform case now, where the exponent k = ◦ • • ◦ . . . consists of p copies of
◦ and p copies of • , the corresponding moment can be computed as follows:
Z
Mk = (f f¯)p
Z
1
= p (a2 + b2 )p
2
 Z Z
1 X p 2r
= p a b2p−2r
2 r r
 
1 X p
= p (2r)!!(2p − 2r)!!
2 r r
1 X p! (2r)! (2p − 2r)!
= p · r · p−r
2 r r!(p − r)! 2 r! 2 (p − r)!
  
p! X 2r 2p − 2r
= p
4 r r p−r

(3) In order to finish now the computation, let us recall that we have the following
formula, coming from the generalized binomial formula, or from the Taylor formula:
∞   q
1 X 2q −t
√ =
1+t q=0
q 4
 14B. NORMAL VARIABLES 335

By taking the square of this series, we obtain the following formula:

1 X  −t p X 2r2p − 2r
=
1+t p
4 r
r p−r

Now by looking at the coefficient of tp on both sides, we conclude that the sum on the
right equals 4p . Thus, we can finish the moment computation in (2), as follows:
p!
Mk = × 4p = p!
4p
We are therefore led to the conclusion in the statement. □
As before with the real Gaussian laws, a better-looking statement is in terms of parti-
tions. Given a colored integer k = ◦ • • ◦ . . . , we say that a pairing π ∈ P2 (k) is matching
when it pairs ◦ − • symbols. With this convention, we have the following result:
Theorem 14.27. The moments of the complex normal law are the numbers
X
Mk (Gt ) = t|π|
π∈P2 (k)

where P2 (k) are the matching pairings of {1, . . . , k}, and |.| is the number of blocks.
Proof. This is a reformulation of Theorem 14.26. Indeed, we can assume that we
are in the case t = 1, and here we know from Theorem 14.26 that the moments are:
(
(|k|/2)! (k uniform)
Mk =
0 (k not uniform)

On the other hand, the numbers |P2 (k)| are given by exactly the same formula. Indeed,
in order to have a matching pairing of k, our exponent k = ◦ • • ◦ . . . must be uniform,
consisting of p copies of ◦ and p copies of •, with p = |k|/2. But then the matching
pairings of k correspond to the permutations of the • symbols, as to be matched with
◦ symbols, and so we have p! such pairings. Thus, we have the same formula as for the
moments of f , and we are led to the conclusion in the statement. □
In practice, we also need to know how to compute joint moments. We have here:
Theorem 14.28 (Wick formula). Given independent variables fi , each following the
complex normal law Gt , with t > 0 being a fixed parameter, we have the formula
n o
k1 ks s/2


E fi1 . . . fis = t # π ∈ P2 (k) π ≤ ker i

where k = k1 . . . ks and i = i1 . . . is , for the joint moments of these variables, where

π ≤ ker i means that the indices of i must fit into the blocks of π, in the obvious way.
336 14. NORMAL VARIABLES

Proof. This is something well-known, which can be proved as follows:

(1) Let us first discuss the case where we have a single variable f , which amounts in
taking fi = f for any i in the formula in the statement. What we have to compute here
are the moments of f , with respect to colored integer exponents k = ◦ • • ◦ . . . , and the
formula in the statement tells us that these moments must be:
E(f k ) = t|k|/2 |P2 (k)|
But this is the formula in Theorem 14.27, so we are done with this case.
(2) In general now, when expanding the product fik11 . . . fikss and rearranging the terms,
we are left with doing a number of computations as in (1), and then making the product
of the expectations that we found. But this amounts in counting the partitions in the
statement, with the condition π ≤ ker i there standing for the fact that we are doing the
various type (1) computations independently, and then making the product. □
The above statement is one of the possible formulations of the Wick formula, and
there are many more formulations, which are all useful. For instance, we have:
Theorem 14.29 (Wick formula 2). Given independent variables fi , each following the
complex normal law Gt , with t > 0 being a fixed parameter, we have the formula
n o
∗ ∗ k


E fi1 . . . fik fj1 . . . fjk = t # π ∈ Sk iπ(r) = jr , ∀r
for the non-vanishing joint moments of these variables.
Proof. This follows from the usual Wick formula, from Theorem 14.28. With some
changes in the indices and notations, the formula there reads:
n o
E fIK1 1 . . . fIKs s = ts/2 # σ ∈ P2 (K) σ ≤ ker I

Now observe that we have P2 (K) = ∅, unless the colored integer K = K1 . . . Ks

is uniform, in the sense that it contains the same number of ◦ and • symbols. Up to
permutations, the non-trivial case, where the moment is non-vanishing, is the case where
the colored integer K = K1 . . . Ks is of the following special form:
K = ◦| ◦{z
. . . ◦} •| •{z
. . . •}
k k

So, let us focus on this case, which is the non-trivial one. Here we have s = 2k, and
we can write the multi-index I = I1 . . . Is in the following way:
I = i1 . . . ik j1 . . . jk
With these changes made, the above usual Wick formula reads:
n o
E fi1 . . . fik fj∗1 . . . fj∗k = tk # σ ∈ P2 (K) σ ≤ ker(ij)


 14C. HYPERSPHERICAL LAWS 337

The point now is that the matching pairings σ ∈ P2 (K), with K = ◦ . . . ◦ • . . . • , of

length 2k, as above, correspond to the permutations π ∈ Sk , in the obvious way. With
this identification made, the above modified usual Wick formula becomes:
n o
E fi1 . . . fik fj∗1 . . . fj∗k = tk # π ∈ Sk iπ(r) = jr , ∀r

Thus, we have reached to the formula in the statement, and we are done. □

Finally, here is one more formulation of the Wick formula, useful as well:
Theorem 14.30 (Wick formula 3). Given independent variables fi , each following the
complex normal law Gt , with t > 0 being a fixed parameter, we have the formula
n o
E fi1 fj∗1 . . . fik fj∗k = tk # π ∈ Sk iπ(r) = jr , ∀r

for the non-vanishing joint moments of these variables.

Proof. This follows from our second Wick formula, from Theorem 14.29, simply by
permuting the terms, as to have an alternating sequence of plain and conjugate variables.
Alternatively, we can start with Theorem 14.28, and then perform the same manipulations
as in the proof of Theorem 14.29, but with the exponent being this time as follows:
K = ◦| • ◦ • .{z
. . . . . ◦ •}
2k

Thus, we are led to the conclusion in the statement. □

So long for the normal laws, real and complex. These appear pretty much everywhere
in mathematics, and with the above, you are fully armed for dealing with them. In
particular the Wick formula is what you need for Random Matrix Theory (RMT), shall
you ever get interested in that, and with a good reference here being Mehta [68].

14c. Hyperspherical laws

We can do many things with the probability theory that we learned so far, and espe-
cially with the real and complex normal laws. As a first application, let us go back to the
spherical integrals, studied in chapter 13. In the real case, we have:
Theorem 14.31. The moments of the hyperspherical variables are
(N − 1)!!p!!
Z
xpi dx =
SRN −1 (N + p − 1)!!
√
and the rescaled variables yi = N xi become normal and independent with N → ∞.
338 14. NORMAL VARIABLES

Proof. The moment formula in the statement follows from the general formulae in
chapter 13. As a consequence, with N → ∞ we have the following estimate:
Z
xpi dx ≃ N −p/2 × p!!
SRN −1

= N −p/2 Mp (g1 )
√
Thus, the rescaled variables N xi become normal with N → ∞, as claimed. As for
the proof of the asymptotic independence, this is standard too, once again by using the
formulae in chapter 13. Indeed, the joint moments of x1 , . . . , xN are given by:
(N − 1)!!k1 !! . . . kN !!
Z
xk11 . . . xkNN dx =
N −1
SR (N + Σki − 1)!!
≃ N −Σki × k1 !! . . . kN !!
√
By rescaling, the joint moments of the variables yi = N xi are given by:
Z
kN
y1k1 . . . yN dx ≃ k1 !! . . . kN !!
SRN −1

Thus, we have multiplicativity, and so independence with N → ∞, as claimed. □

Importantly, we can recover the normal laws as well in connection with the rotation
groups. Indeed, we have the following reformulation of Theorem 14.31:
Theorem 14.32. We have the integration formula
(N − 1)!!p!!
Z
Uijp dU =
ON (N + p − 1)!!
√
and the rescaled variables Vij = N Uij become normal and independent with N → ∞.
Proof. We use the basic fact that the rotations U ∈ ON act on the points of the real
sphere z ∈ SRN −1 , with the stabilizer of z = (1, 0, . . . , 0) being the subgroup ON −1 ⊂ ON .
In algebraic terms, this gives an identification as follows:
SRN −1 = ON /ON −1
In functional analytic terms, this result provides us with an embedding as follows, for
any i, which makes correspond the respective integration functionals:
C(SRN −1 ) ⊂ C(ON ) , xi → U1i
With this identification made, the result follows from Theorem 14.31. □
In the complex case, the analogues of the above results are as follows:
 14C. HYPERSPHERICAL LAWS 339
√
Theorem 14.33. The rescalings N zi of the unit complex sphere coordinates
zi : SCN −1 → C
√
as well as the rescalings N Uij of the unitary group coordinates
Uij : UN → C
become complex Gaussian and independent with N → ∞.
Proof. We have several assertions to be proved, the idea being as follows:
(1) According to the formulae in chapter 13, the polynomials integrals in zi , z̄i vanish,
unless the number of zi , z̄i is the same. In this latter case these terms can be grouped
together, by using zi z̄i = |zi |2 , and the relevant integration formula is:
(N − 1)!k!
Z
|zi |2k dz =
N −1
SC (N + k − 1)!
Now with N → ∞, we obtain from this the following estimate:
Z
|zi |2k dx ≃ N −k × k!
SCN −1
√
Thus, the rescaled variables N zi become normal with N → ∞, as claimed.
(2) As for the proof of the asymptotic independence, this is standard too, again by
using the formulae in chapter 13. Indeed, the joint moments of z1 , . . . , zN are given by:
(N − 1)!k1 ! . . . kn !
Z
|z1 |2k1 . . . |zN |2kN dx = P
SRN −1 (N + ki − 1)!
≃ N −Σki × k1 ! . . . kN !
√
By rescaling, the joint moments of the variables yi = N zi are given by:
Z
|y1 |2k1 . . . |yN |2kN dx ≃ k1 ! . . . kN !
SRN −1

Thus, we have multiplicativity, and so independence with N → ∞, as claimed.

(3) Regarding the last assertion, we can use here the basic fact that the rotations
U ∈ UN act on the points of the sphere z ∈ SCN −1 , with the stabilizer of z = (1, 0, . . . , 0)
being the subgroup UN −1 ⊂ UN . In algebraic terms, this gives an equality as follows:
SCN −1 = UN /UN −1
In functional analytic terms, this result provides us with an embedding as follows, for
any i, which makes correspond the respective integration functionals:
C(SCN −1 ) ⊂ C(UN ) , xi → U1i
With this identification made, the result follows from (1,2). □
340 14. NORMAL VARIABLES

14d. Rotations, reflections

All the above is quite nice, and suggests that there are deep relations between group
theory and probability. So, let us discuss this now, by starting with the discrete case,
which is the simplest one. We have here the following beautiful, well-known result:
Theorem 14.34. The probability for a random σ ∈ SN to have no fixed points is
1
P ≃
e
in the N → ∞ limit, where e = 2.71828 . . . is the usual constant from analysis.
Proof. This is best viewed by using the inclusion-exclusion principle. Let us set:
n o
k
SN = σ ∈ SN σ(k) = k

The set of permutations having no fixed points, called derangements, is then:

!c
[
k
XN = SN
k

Now the inclusion-exclusion principle tells us that we have:

!c
[
k
|XN | = SN
k
X X X kN
k k l k1
= |SN | − |SN |+ |SN ∩ SN | − . . . + (−1)N |SN ∪ . . . ∪ SN |
k k<l k1 <...<kN
   
N N N
= N ! − N (N − 1)! + (N − 2)! − . . . + (−1) (N − N )!
2 N
N  
r N
X
= (−1) (N − r)!
r=0
r

Thus, the probability that we are interested in, for a random permutation σ ∈ SN to
have no fixed points, is given by the following formula:
N
|XN | X (−1)r
P = =
N! r=0
r!

Since on the right we have the expansion of 1/e, this gives the result. □

The above result looks quite exciting. In order to further explore and refine it, we will
need some notions from group theory, and more specifically, the following definition:
 14D. ROTATIONS, REFLECTIONS 341

Definition 14.35. Given a closed subgroup G ⊂ UN , the function

N
X
χ:G→C , χ(g) = gii
i=1

is called main character of G.

We should mention that there is a long story with these characters, the idea being that
a compact group G can have several representations π : G ⊂ UN , which can be studied
via their characters χπ : G → C, and with all this coming from Weyl [97], [98].

All this is quite heavy, but technically, we will not need all this here. Indeed, in
relation with our questions, we can formulate the following nice, elementary result:
Theorem 14.36. Consider the symmetric group SN , regarded as the permutation
group, SN ⊂ ON , of the N coordinate axes of RN .
(1) The main character χ ∈ C(SN ) counts the number of fixed points.
(2) The law of χ ∈ C(SN ) becomes Poisson (1), in the N → ∞ limit.
Proof. We have two things to be proved here, the idea being as follows:
(1) The permutation matrices σ ∈ ON , which give the embedding SN ⊂ ON in the
statement, being given by σij = δiσ(j) , we have the following computation:
X n o
χ(σ) = δσ(i)i = # i ∈ {1, . . . , N } σ(i) = i
i

(2) In order to establish now the asymptotic result in the statement, we must prove
the following formula, for any r ∈ N, in the N → ∞ limit:
1
P (χ = r) ≃
r!e
We already know, from Theorem 14.34, that this formula holds at r = 0. In the
general case now, we have to count the permutations σ ∈ SN having exactly r points.
Now since having such a permutation amounts in choosing r points among 1, . . . , N , and
then permuting the N − r points left, without fixed points allowed, we have:
  n
n o N o
# σ ∈ SN χ(σ) = r = # σ ∈ SN −r χ(σ) = 0
r
N! n o
= # σ ∈ SN −r χ(σ) = 0
r!(N − r)!
n o
1 # σ ∈ SN −r χ(σ) = 0
= N! × ×
r! (N − r)!
342 14. NORMAL VARIABLES

By dividing everything by N !, we obtain from this the following formula:

n o n o
# σ ∈ SN χ(σ) = r 1 # σ ∈ S N −r χ(σ) = 0
= ×
N! r! (N − r)!
Now by using the computation at r = 0, that we already have, from Theorem 14.34,
it follows that with N → ∞ we have the following estimate:
1 1 1
P (χ = r) ≃ · P (χ = 0) ≃ ·
r! r! e
Thus, we obtain as limiting measure the Poisson law of parameter 1, as stated. □
As a next step, let us try now to generalize what we have, namely Theorem 14.36,
as to reach to the Poisson laws of arbitrary parameter t > 0. Again we will need some
notions from group theory, and more specifically, the following definition:
Definition 14.37. Given a closed subgroup G ⊂ UN , the function
[tN ]
X
χt : G → C , χt (g) = gii
i=1

is called main truncated character of G, of parameter t ∈ (0, 1].

As before with plain characters, there is some theory behind this definition, involving
this time Random Matrix Theory (RMT), but technically, we will not need this here. In-
deed, we can now formulate the following result, which is nice and elementary, generalizing
Theorem 14.36, and which will be our final saying on the subject:
Theorem 14.38. Consider the symmetric group SN , regarded as the permutation
group, SN ⊂ ON , of the N coordinate axes of RN .
(1) The variable χt counts the number of fixed points among 1, . . . , [tN ].
(2) The law of this variable χt becomes Poisson (t), in the N → ∞ limit.
Proof. We already know from Theorem 14.36 that the results hold at t = 1. In
general, the proof is similar, the idea being as follows:
(1) We have indeed the following computation, coming from definitions:
[tN ] n o
X
χt (σ) = δσ(i)i = # i ∈ {1, . . . , [tN ]} σ(i) = i
i=1

(2) Consider indeed the following sets, as in the proof of Theorem 14.34:
n o
k
SN = σ ∈ SN σ(k) = k
 14D. ROTATIONS, REFLECTIONS 343

The set of permutations having no fixed points among 1, . . . , [tN ] is then:

 c
[
k
XN =  SN
k≤[tN ]

As before in the proof of Theorem 14.36, we obtain by inclusion-exclusion that:

[tN ]
1 X X
k1 kr
P (χt = 0) = (−1)r |SN ∩ . . . ∩ SN |
N ! r=0
k1 <...<kr <[tN ]
[tN ]  
1 X r [tN ]
= (−1) (N − r)!
N ! r=0 r
[tN ]
X (−1)r [tN ]!(N − r)!
= ·
r=0
r! N !([tN ] − r)!
Now with N → ∞, we obtain from this the following estimate:
[tN ]
X (−1)r
P (χt = 0) ≃ · tr
r=0
r!
[tN ]
X (−t)r
=
r=0
r!
−t
≃ e
More generally, by counting the permutations σ ∈ SN having exactly r fixed points
among 1, . . . , [tN ], as in the proof of Theorem 14.36, we obtain:
tr
P (χt = r) ≃
r!et
Thus, we obtain in the limit a Poisson law of parameter t, as stated. □
It is possible to establish as well continuous analogues of the above results, involving
the groups ON , UN , and the real and complex normal laws. However, the proofs here are
more technical. For an elementary introduction to this, have a look at my book [9].

Finally, let us end all this with some low-dimensional magics, involving physicists’
favorite group, SU2 , and one of their favorite laws, the Wigner law γ1 . We have:
Theorem 14.39. The main character of SU2 follows the following law,
1√
γ1 = 4 − x2 dx
2π
which is the Wigner law of parameter 1.
344 14. NORMAL VARIABLES

Proof. The group SU2 is by definition the group of unitary rotations U ∈ U2 of

determinant one, and by solving the equation U ∗ = U −1 , we are led to:
  
a + ib c + id 2 2 2 2
SU2 = a +b +c +d =1
−c + id a − ib
In this picture, the main character is given by the following formula:
 
a + ib c + id
χ = 2a
−c + id a − ib
We are therefore left with computing the law of the following variable:
a ∈ C(SR3 )
But this is something very familiar, namely a hyperspherical variable at N = 4, so we
can use here Theorem 14.31. We obtain the following moment formula:
Z
3!!(2k)!!
a2k =
SR3 (2k + 3)!!
(2k)!
= 2·
2k k!2k+1 (k
+ 1)!
 
1 1 2k
= k·
4 k+1 k
Ck
=
4k
Thus the variable 2a ∈ C(SR3 ) follows the Wigner semicircle law γ1 , as claimed. □
There are many more things that can be said, in relation with the above, all beautiful
mathematics, useful in relation with physics. If interested, have a look at my book [9].
14e. Exercises
Probability theory is a big business, and many other things can be said, in relation
with the above, which are all useful, and good to know. As exercises, we have:
Exercise 14.40. Learn about compound Poisson laws.
Exercise 14.41. Study more in detail the CLT, and the convergence there.
Exercise 14.42. Learn about the heat kernel, and solving the heat equation.
Exercise 14.43. Recover the Marchenko-Pastur law π1 , in relation with SO3 .
As bonus exercise, learn some compact group theory, and more specifically the Peter-
Weyl theory for such groups. This is quite tough, but is an excellent investment.
 CHAPTER 15

Partial integration

15a. Vector products

Good news, done with mathematics, and in the remainder of this book, last 50 pages,
we will get into physics. We already met some physics in this book, namely gravity, waves
and heat in one dimension, in chapters 3-4, then waves and heat again, in two dimensions,
along with some basic electrodynamics, light and early atomic theory, in chapter 8, and
finally harmonic functions, this time in arbitrary N dimensions, in chapter 12.

Obviously, time to have some cleanup here, review what we know, and do more, in a
more systematic way. We have divided what we have two say in two parts, as follows:

(1) In the present chapter we will discuss physics in low dimensions, N = 2, 3, 4,

with on the menu more classical mechanics, notably with the Coriolis force, then some
relativity following Einstein, and finally the basics of electrostatics, following Gauss and
others, which are quite interesting, mathematically speaking. We will be mostly interested
in N = 3, and the common feature of everything that we want to talk about will be the
vector product ×, which exists only in 3 dimensions, and is something quite magic.

(2) In the next and final chapter we will discuss physics in infinite dimensions, N = ∞.
Our goal here will be that of having some quantum mechanics theory started, along the
lines suggested at the end of chapter 8, and more specifically, solving the hydrogen atom.
There are actually several ways of proceeding here, following Heisenberg, Schrödinger and
others, and matter of annoying my cat, who seems to be a big fan of measure theory and
Hilbert spaces, we will opt here for the Schrödinger approach, which is elementary.

Getting now to what we want to do in this chapter, namely 3 dimensions, vector

products ×, and all sorts of related physics, there is some important mathematical interest
in doing this, because we will reach in this way to answers to the following question:
Question 15.1. What are the higher dimensional analogues of the formula
Z b
F ′ (x)dx = F (b) − F (a)
a
that is, of the fundamental theorem of calculus?
345
346 15. PARTIAL INTEGRATION

And isn’t this a fundamental question for us, mathematicians, because we have so
far in our bag multivariable extensions of all the main theorems of one-variable calculus,
except for this. So, definitely something to be solved, before the end of this book.
So, let us discuss this first. The fundamental theorem of calculus tells us that the
integral of a function on an interval [a, b] can be suitably recaptured from what happens
on the boundary {a, b} of this interval. Obviously, this is something quite magic, and
thinking now at what we can expect in N = 2, 3 or more dimensions, that can only be
quite complicated, involving curves, surfaces, solid bodies and so on, and with all this
vaguely reminding all sorts of physics things, such as field lines for gravity, magnets and
so on. In short, despite having no formal proof yet for all this, let us formulate:
Answer 15.2. Partial integration in several dimensions most likely means physics,
and we will probably only get some partial results here, at N = 2, 3.
Of course, all this remains to be confirmed, but assuming that you trust me a bit,
here we are now at the plan that we made before, for this chapter. That is, do physics in
low dimensions, guided by the beauty of the world surrounding us, and once this physics
done, record some mathematical corollaries too, in relation with Question 15.1.
Before getting started, however, as usual when struggling with pedagogical matters,
and other delicate dillemas, let us ask the cat. But here, unfortunately, no surprise:
Cat 15.3. Read Rudin.
Thanks cat, and guess this sort of discussion, that we started in chapter 13, looks
more and more cyclic. So, I’ll just go my way, on your side have a good hunt, and by the
way make sure to catch enough mice and birds, using your measure theory and differential
forms techniques, because there is a bit of shortage of cat food today, sorry for that.
Getting started now, here is what we need:
Definition 15.4. The vector product of two vectors in R3 is given by
x × y = ||x|| · ||y|| · sin θ · n
where n ∈ R3 with n ⊥ x, y and ||n|| = 1 is constructed using the right-hand rule:
↑x×y
←x
↙y
Alternatively, in usual vertical linear algebra notation for all vectors,
     
x1 y1 x2 y 3 − x3 y 2
 x2  ×  y 2  =  x3 y 1 − x1 y 3 
x3 y3 x1 y 2 − x2 y 1
the rule being that of computing 2 × 2 determinants, and adding a middle sign.
 15A. VECTOR PRODUCTS 347

Obviously, this definition is something quite subtle, and also something very annoying,
because you always need this, and always forget the formula. Here are my personal
methods. With the first definition, what I always remember is that:
||x × y|| ∼ ||x||, ||y|| , x×x=0 , e1 × e2 = e3
So, here’s how it works. We are looking for a vector x × y whose length is proportional
to those of x, y. But the second formula tells us that the angle θ between x, y must be
involved via 0 → 0, and so the factor can only be sin θ. And with this we are almost
there, it’s just a matter of choosing the orientation, and this comes from e1 × e2 = e3 .

As with the second definition, that I like the most, what I remember here is simply:
1 x1 y 1
1 x2 y2 =?
1 x3 y 3
Indeed, when trying to compute this determinant, by developing over the first column,
what you get as coefficients are the entries of x × y. And with the good middle sign.

In practice now, in order to get familiar with the vector products, nothing better than
doing some classical mechanics. We have here the following key result:
Theorem 15.5. In the gravitational 2-body problem, the angular momentum
J =x×p
with p = mv being the usual momentum, is conserved.
Proof. There are several things to be said here, the idea being as follows:
(1) First of all the usual momentum, p = mv, is not conserved, because the simplest
solution is the circular motion, where the moment gets turned around. But this suggests
precisely that, in order to fix the lack of conservation of the momentum p, what we have
to do is to make a vector product with the position x. Leading to J, as above.
(2) Regarding now the proof, consider indeed a particle m moving under the gravita-
tional force of a particle M , assumed, as usual, to be fixed at 0. By using the fact that
for two proportional vectors, p ∼ q, we have p × q = 0, we obtain:
J˙ = ẋ × p + x × ṗ
= v × mv + x × ma
= m(v × v + x × a)
= m(0 + 0)
= 0
Now since the derivative of J vanishes, this quantity is constant, as stated. □
348 15. PARTIAL INTEGRATION

While the above principle looks like something quite trivial, the mathematics behind
it is quite interesting, and has several notable consequences, as follows:
Theorem 15.6. In the context of a 2-body problem, the following happen:
(1) The fact that the direction of J is fixed tells us that the trajectory of one body
with respect to the other lies in a plane.
(2) The fact that the magnitude of J is fixed tells us that the Kepler 2 law holds,
namely that we have same areas sweeped by Ox over the same times.
Proof. This follows indeed from Theorem 15.5, as follows:

(1) We have by definition J = m(x × v), and since a vector product is orthogonal on
both the vectors it comes from, we deduce from this that we have:

J ⊥ x, v

But this can be written as follows, with J ⊥ standing for the plane orthogonal to J:

x, v ∈ J ⊥

Now since J is fixed by Theorem 15.5, we conclude that both x, v, and in particular
the position x, and so the whole trajectory, lie in this fixed plane J ⊥ , as claimed.

(2) Conversely now, forget about Theorem 15.5, and assume that the trajectory lies
in a certain plane E. Thus x ∈ E, and by differentiating we have v ∈ E too, and so
x, v ∈ E. Thus E = J ⊥ , and so J = E ⊥ , so the direction of J is fixed, as claimed.

(3) Regarding now the last assertion, we already know from the various formulae from
chapter 11 that the Kepler 2 law is more or less equivalent to the formula θ̇ = λ/r2 .
However, the derivation of θ̇ = λ/r2 was something tricky, and what we want to prove
now is that this appears as a simple consequence of ||J|| = constant.

(4) In order to to so, let us compute J, according to its definition J = x × p, but in

polar coordinates, which will change everything. Since p = mẋ, we have:
   
cos θ ṙ cos θ − r sin θ · θ̇
J = r  sin θ  × m ṙ sin θ + r cos θ · θ̇ 
0 0
Now recall from the definition of the vector product that we have:
     
a c 0
 b  × d =  0 
0 0 ad − bc
 15A. VECTOR PRODUCTS 349

Thus J is a vector of the above form, with its last component being:
cos θ ṙ cos θ − r sin θ · θ̇
Jz = rm
sin θ ṙ sin θ + r cos θ · θ̇
= rm · r(cos2 θ + sin2 θ)θ̇
= r2 m · θ̇
(5) Now with the above formula in hand, our claim is that the magnitude ||J|| is
constant precisely when θ̇ = λ/r2 , for some λ ∈ R. Indeed, up to the obvious fact that
the orientation of J is a binary parameter, who cannot just switch like that, let us just
agree on this, knowing J is the same as knowing Jz , and is also the same as knowing ||J||.
Thus, our claim is proved, and this leads to the conclusion in the statement. □
As another basic application of the vector products, still staying with classical me-
chanics, we have all sorts of useful formulae regarding rotating frames. We first have:
Theorem 15.7. Assume that a 3D body rotates along an axis, with angular speed w.
For a fixed point of the body, with position vector x, the usual 3D speed is
v =ω×x
where ω = wn, with n unit vector pointing North. When the point moves on the body
V = ẋ + ω × x
is its speed computed by an inertial observer O on the rotation axis.
Proof. We have two assertions here, both requiring some 3D thinking, as follows:
(1) Assuming that the point is fixed, the magnitude of ω × x is the good one, due to
the following computation, with r being the distance from the point to the axis:
||ω × x|| = w||x|| sin t = wr = ||v||
As for the orientation of ω × x, this is the good one as well, because the North pole
rule used above amounts in applying the right-hand rule for finding n, and so ω, and this
right-hand rule was precisely the one used in defining the vector products ×.
(2) Next, when the point moves on the body, the inertial observer O can compute its
speed by using a frame (u1 , u2 , u3 ) which rotates with the body, as follows:
V = ẋ1 u1 + ẋ2 u2 + ẋ3 u3 + x1 u̇1 + x2 u̇2 + x3 u̇3
= ẋ + (x1 · ω × u1 + x2 · ω × u2 + x3 · ω × u3 )
= ẋ + w × (x1 u1 + x2 u2 + x3 u3 )
= ẋ + ω × x
Thus, we are led to the conclusions in the statement. □
In what regards now the acceleration, the result, which is famous, is as follows:
350 15. PARTIAL INTEGRATION

Theorem 15.8. Assuming as before that a 3D body rotates along an axis, the accel-
eration of a moving point on the body, computed by O as before, is given by

A = a + 2ω × v + ω × (ω × x)

with ω = wn being as before. In this formula the second term is called Coriolis accelera-
tion, and the third term is called centripetal acceleration.
Proof. This comes by using twice the formulae in Theorem 15.7, as follows:

A = V̇ + ω × V
= (ẍ + ω̇ × x + ω × ẋ) + (ω × ẋ + ω × (ω × x))
= ẍ + ω × ẋ + ω × ẋ + ω × (ω × x)
= a + 2ω × v + ω × (ω × x)

Thus, we are led to the conclusion in the statement. □

The truly famous result is actually the one regarding forces, obtained by multiplying
everything by a mass m, and writing things the other way around, as follows:

ma = mA − 2mω × v − mω × (ω × x)

Here the second term is called Coriolis force, and the third term is called centrifugal
force. These forces are both called apparent, or fictious, because they do not exist in the
inertial frame, but they exist however in the non-inertial frame of reference, as explained
above. And with of course the terms centrifugal and centripetal not to be messed up.

In fact, even more famous is the terrestrial application of all this, as follows:
Theorem 15.9. The acceleration of an object m subject to a force F is given by

ma = F − mg − 2mω × v − mω × (ω × x)

with g pointing upwards, and with the last terms being the Coriolis and centrifugal forces.
Proof. This follows indeed from the above discussion, by assuming that the acceler-
ation A there comes from the combined effect of a force F , and of the usual g. □

We refer to any standard undergraduate mechanics book, such as Feynman [31], Kibble
[53] or Taylor [89] for more on the above, including various numerics on what happens
here on Earth, the Foucault pendulum, history of all this, and many other things. Let
us just mention here, as a basic illustration for all this, that a rock dropped from 100m
deviates about 1cm from its intended target, due to the formula in Theorem 15.9.
 15B. EINSTEIN ADDITION 351

15b. Einstein addition

As another application of the vector products, let us discuss now the speed addition
in relativity. Based on experiments by Fizeau, then Michelson-Morley and others, and
some physics by Maxwell and Lorentz, Einstein came upon the following principles:
Fact 15.10 (Einstein principles). The following happen:
(1) Light travels in vacuum at a finite speed, c < ∞.
(2) This speed c is the same for all inertial observers.
(3) In non-vacuum, the light speed is lower, v < c.
(4) Nothing can travel faster than light, v ̸> c.
The point now is that, obviously, something is wrong here. Indeed, assuming for
instance that we have a train, running in vacuum at speed v > 0, and someone on board
lights a flashlight ∗ towards the locomotive, an observer ◦ on the ground will see the light
travelling at speed c + v > c, which is a contradiction:

∗ / /
c v

⃝ ⃝ ⃝ ⃝ ⃝ ⃝
◦ /
c+v

Equivalently, with the same train running, in vacuum at speed v > 0, if the observer
on the ground lights a flashlight ∗ towards the back of the train, then viewed from the
train, that light will travel at speed c + v > c, which is a contradiction again:

o ◦ /
c+v v

⃝ ⃝ ⃝ ⃝ ⃝ ⃝
o ∗
c

Summarizing, Fact 15.10, while physically true, implies c + v = c, so contradicts

classical mechanics, which needs a fix. In the classical case, to start with, we have:
Proposition 15.11. The classical speeds add according to the Galileo formula
vAC = vAB + vBC
where vAB denotes the relative speed of A with respect to B.
Proof. This is clear indeed from the definition of speed, and very intuitive. □
352 15. PARTIAL INTEGRATION

In order to find the fix, we will first discuss the 1D case, and leave the 3D case, which
is a bit more complicated, for later. We will use two tricks. First, let us forget about
absolute speeds, with respect to a given frame, and talk about relative speeds only. In this
case we are allowed to sum only quantities of type vAB , vBC , and we denote by vAB +g vBC
the corresponding sum vAC . With this convention, the Galileo formula becomes:
u +g v = u + v
As a second trick now, observe that this Galileo formula holds in any system of units.
In order now to deal with our problems, basically involving high speeds, it is convenient
to change the system of units, as to have c = 1. With this convention our c + v = c
problem becomes 1 + v = 1, and the solution to it is quite obvious, as follows:
Theorem 15.12. If we define the Einstein sum +e of relative speeds by
u+v
u +e v =
1 + uv
in c = 1 units, then we have the formula 1 +e v = 1, valid for any v.
Proof. This is obvious indeed from our definition of +e , because if we plug in u = 1
in the above formula, we obtain as result:
1+v
1 +e v = =1
1+v
Thus, we are led to the conclusion in the statement. □
Summarizing, we have solved our problem. In order now to formulate a final result,
we must do some reverse engineering, by waiving the above two tricks. First, by getting
back to usual units, v → v/c, our new addition formula becomes:
u
u v +v
+e = c u c v
c c 1+ c · c
By multiplying by c, we can write this formula in a better way, as follows:
u+v
u +e v =
1 + uv/c2
In order now to finish, it remains to get back to absolute speeds, as in Proposition
15.11. And by doing so, we are led to the following result:
Theorem 15.13. If we sum the speeds according to the Einstein formula
vAB + vBC
vAC =
1 + vAB vBC /c2
then the Galileo formula still holds, approximately, for low speeds
vAC ≃ vAB + vBC
and if we have vAB = c or vBC = c, the resulting sum is vAC = c.
 15B. EINSTEIN ADDITION 353

Proof. We have two assertions here, which are both clear, as follows:
(1) Regarding the first assertion, if we are at low speeds, vAB , vBC << c, the correction
term vAB vBC /c2 dissapears, and we are left with the Galileo formula, as claimed.
(2) As for the second assertion, this follows from the above discussion. □
The Einstein summation formula, while looking very simple, is in fact quite subtle,
and must be handled with care. Indeed, getting back to c = 1 conventions, we have:
Proposition 15.14. The Einstein speed summation, written in c = 1 units as
u+v
u +e v =
1 + uv
has the following properties:
(1) u, v < 1 implies u +e v < 1.
(2) u +e v = v +e u.
(3) (u +e v) +e w = u +e (v +e w).
(4) However, λu +e λv = λ(u +e v) fails.
Proof. All these assertions are elementary, as follows:
(1) This follows from the following formula, valid for any speeds u, v:
u+v (1 − u)(1 − v)
1 − u +e v = 1 − =
1 + uv 1 + uv
(2) This is clear too, coming from the following computation:
u+v v+u
u +e v = = = v +e u
1 + uv 1 + vu
(3) We have indeed the following computation:
u + v + w + uvw
(u +e v) +e w = u +e (v +e w) =
1 + uv + uw + vw
(4) This is clear too, with the remark however that the formula λu +e λv = λ(u +e v)
works at λ = −1, 0, 1, or when u = 0, v = 0, or u + v = 0. □
All the above is very nice, but remember, takes place in 1D. So, time now to get
seriously to work, and see what all this becomes in 3D. Expect of course a lot a vector
calculus, as usual in relation with 3D problems, and in the hope that you love that. As a
main goal, we must review the Einstein speed summation formula:
Question 15.15. What is the correct analogue of the Einstein summation formula
u+v
u +e v =
1 + uv
in 2 and 3 dimensions?
354 15. PARTIAL INTEGRATION

In order to discuss this question, let us attempt to construct u +e v in arbitrary

dimensions, just by using our common sense and intuition. When the vectors u, v ∈ RN
are proportional, we are basically in 1D, and so our addition formula must satisfy:
u+v
u ∼ v =⇒ u +e v =
1+ < u, v >
However, the formula on the right will not work as such in general, for arbitrary speeds
u, v ∈ RN , and this because we have, as main requirement for our operation, in analogy
with the 1 +e v = 1 formula from 1D, the following condition:
||u|| = 1 =⇒ u +e v = u
Equivalently, in analogy with u +e 1 = 1 from 1D, we would like to have:
||v|| = 1 =⇒ u +e v = v
Summarizing, our u ∼ v formula above is not bad, as a start, but we must add a
correction term to it, for the above requirements to be satisfied, and of course with the
correction term vanishing when u ∼ v. So, we are led to a math puzzle:
Puzzle 15.16. What vanishes when u ∼ v, and then how to correctly define
u + v + γuv
u +e v =
1+ < u, v >
as for the correction term γuv to vanish when u ∼ v?
But the solution to the first question is well-known in 3D. Indeed, here we can use the
vector product u × v, that we met before, which notoriously satisfies:
u ∼ v =⇒ u × v = 0
Thus, our correction term γuv must be something containing w = u×v, which vanishes
when this vector w vanishes, and in addition arranged such that ||u|| = 1 produces a
simplification, with u +e v = u as end result, and with ||v|| = 1 producing a simplification
too, with u +e v = v as end result. Thus, our vector calculus puzzle becomes:
Puzzle 15.17. How to correctly define the Einstein summation in 3 dimensions,
u + v + γuvw
u +e v =
1+ < u, v >
with w = u × v, in such a way as for the correction term γuvw to satisfy
w = 0 =⇒ γuvw = 0
and also such that ||u|| = 1 =⇒ u +e v = u, and ||v|| =⇒ u +e v = v?
 15B. EINSTEIN ADDITION 355

In order to solve this latter puzzle, the first observation is that γuvw = w will not do,
and this for several reasons. First, this vector points in the wrong direction, orthogonal
to the plane spanned by u, v, and we certainly don’t want to leave this plane, with our
correction. Also, as a technical remark to be put on top of this, the choice γuvw = w will
not bring any simplifications, as required above, in the cases ||u|| = 1 or ||v|| = 1. Thus,
certainly wrong choice, and we must invent something more complicated.

Moving ahead now, as obvious task, we must “transport” the vector w to the plane
spanned by u, v. But this is simplest done by taking the vector product with any vector
in this plane, and so as a reasonable candidate for our correction term, we have:
γuvw = (αu + βv) × w
Here α, β ∈ R are some scalars to be determined, but let us take a break, and leave
the computations for later. We did some good work, time to update our puzzle:
Puzzle 15.18. How to define the Einstein summation in 3 dimensions,
u + v + γuvw
u +e v =
1+ < u, v >
with the correction term being of the following form, with w = u × v, and α, β ∈ R,
γuvw = (αu + βv) × w
in such a way as to have ||u|| = 1 =⇒ u +e v = u, and ||v|| =⇒ u +e v = v?
In order to investigate what happens when ||u|| = 1 or ||v|| = 1, we must compute the
vector products u × w and v × w. So, pausing now our study for consulting the vector
calculus database, and then coming back, here is the formula that we need:
u × (u × v) =< u, v > u− < u, u > v
As for the formula of v × w, that I forgot to record, we can recover it from the one
above of u × w, by using the basic properties of the vector products, as follows:
v × (u × v) = −v × (v × u)
= −(< v, u > v− < v, v > u)
= < v, v > u− < u, v > v
With these formulae in hand, we can now compute the correction term, with the result
here, that we will need several times in what comes next, being as follows:
Proposition 15.19. The correction term γuvw = (αu + βv) × w is given by
γuvw = (α < u, v > +β < v, v >)u − (α < u, u > +β < u, v >)v
for any values of the scalars α, β ∈ R.
356 15. PARTIAL INTEGRATION

Proof. According to our vector product formulae above, we have:

γuvw = (αu + βv) × w
= α(< u, v > u− < u, u > v) + β(< v, v > u− < u, v > v)
= (α < u, v > +β < v, v >)u − (α < u, u > +β < u, v >)v
Thus, we are led to the conclusion in the statement. □
Time now to get into the real thing, see what happens when ||u|| = 1 and ||v|| = 1,
if we can get indeed u +e v = u and u +e v = v. It is convenient here to do some reverse
engineering. Regarding the first desired formula, namely u +e v = u, we have:
u +e v = u ⇐⇒ u + v + γuvw = (1+ < u, v >)u
⇐⇒ γuvw =< u, v > u − v
⇐⇒ α = 1, β = 0, ||u|| = 1
Thus, with the parameter choice α = 1, β = 0, we will have, as desired:
||u|| = 1 =⇒ u +e v = u
In what regards now the second desired formula, namely u +e v = v, here the compu-
tation is almost identical, save for a sign switch, which after some thinking comes from
our choice w = u × v instead of w = v × u, clearly favoring u, as follows:
u +e v = v ⇐⇒ u + v + γuvw = (1+ < u, v >)v
⇐⇒ γuvw = −u+ < u, v > v
⇐⇒ α = 0, β = −1, ||v|| = 1
Thus, with the parameter choice α = 0, β = −1, we will have, as desired:
||v|| = 1 =⇒ u +e v = v
All this is mixed news, because we managed to solve both our problems, at ||u|| = 1
and at ||v|| = 1, but our solutions are different. So, time to breathe, decide that we did
enough interesting work for the day, and formulate our conclusion as follows:
Proposition 15.20. When defining the Einstein speed summation in 3D as
u + v + u × (u × v)
u +e v =
1+ < u, v >
in c = 1 units, the following happen:
(1) When u ∼ v, we recover the previous 1D formula.
(2) When ||u|| = 1, speed of light, we have u +e v = u.
(3) However, ||v|| = 1 does not imply u +e v = v.
(4) Also, the formula u +e v = v +e u fails.
 15B. EINSTEIN ADDITION 357

Proof. Here (1) and (2) follow from the above discussion, with the following choice
for the correction term, by favoring the ||u|| = 1 problem over the ||v|| = 1 one:
γuvw = u × w
In fact, with this choice made, the computation is very simple, as follows:
||u|| = 1 =⇒ γuvw =< u, v > u − v
=⇒ u + v + γuvw = u+ < u, v > u
u + v + γuvw
=⇒ =u
1+ < u, v >
As for (3) and (4), these are also clear from the above discussion, coming from the
obvious lack of symmetry of our summation formula. □

Looking now at Proposition 15.20 from an abstract, mathematical perspective, there

are still many things missing from there, which can be summarized as follows:
Question 15.21. Can we fine-tune the Einstein speed summation in 3D into
u + v + λ · u × (u × v)
u +e v =
1+ < u, v >
with λ ∈ R, chosen such that ||u|| = 1 =⇒ λ = 1, as to have:
(1) ||u||, ||v|| < 1 =⇒ ||u +e v|| < 1.
(2) ||v|| = 1 =⇒ ||u +e v|| = 1.
All this is quite tricky, and deserves some explanations. First, if we add a scalar λ ∈ R
into our formula, as above, we will still have, exactly as before:
1 + uv
u ∼ v =⇒ u +e v =
1+ < u, v >
On the other hand, we already know from our previous computations, those preceding
Proposition 15.20, that if we ask for λ ∈ R to be a plain constant, not depending on u, v,
then λ = 1 is the only good choice, making the following formula happen:
||u|| = 1 =⇒ u +e v = u
But, and here comes our point, λ = 1 is not an ideal choice either, because it would be
nice to have the properties (1,2) in the statement, and these properties have no reason to
be valid for λ = 1, as you can check for instance by yourself by doing some computations.
Thus, the solution to our problem most likely involves a scalar λ ∈ R depending on u, v,
and satisfying the following condition, as to still have ||u|| = 1 =⇒ u +e v = u:
||u|| = 1 =⇒ λ = 1
358 15. PARTIAL INTEGRATION

Obviously, as simplest answer, λ must be some well-chosen function of ||u||, or rather

of ||u||2 , because it is always better to use square norms, when possible. But then, with
this idea in mind, after a few computations we are led to the following solution:
1
λ= p
1 + 1 − ||u||2
Summarizing, final correction done, and with this being the end of mathematics, we
did a nice job, and we can now formulate our findings as a theorem, as follows:
Theorem 15.22. When defining the Einstein speed summation in 3D as
!
1 u × (u × v)
u +e v = u+v+ p
1+ < u, v > 1 + 1 − ||u||2
in c = 1 units, the following happen:
(1) When u ∼ v, we recover the previous 1D formula.
(2) We have ||u||, ||v|| < 1 =⇒ ||u +e v|| < 1.
(3) When ||u|| = 1, we have u +e v = u.
(4) When ||v|| = 1, we have ||u +e v|| = 1.
(5) However, ||v|| = 1 does not imply u +e v = v.
(6) Also, the formula u +e v = v +e u fails.
Proof. This follows from the above discussion, as follows:
(1) This is something that we know from Proposition 15.20.
p
(2) In order to
p simplify notation, let us set δ = 1 − ||u||2 , which is the inverse of the
quantity γ = 1/ 1 − ||u||2 . With this convention, we have:
< u, v > u − ||u||2 v
 
1
u +e v = u+v+
1+ < u, v > 1+δ
(1 + δ+ < u, v >)u + (1 + δ − ||u||2 )v
=
(1+ < u, v >)(1 + δ)
Taking now the squared norm and computing gives the following formula:
(1 + δ)2 ||u + v||2 + (||u||2 − 2(1 + δ))(||u||2 ||v||2 − < u, v >2 )
||u +e v||2 =
(1+ < u, v >)2 (1 + δ)2
p
But this formula can be further processed by using δ = 1 − ||u||2 , and by navigating
through the various quantities which appear, we obtain, as a final product:
||u + v||2 − ||u||2 ||v||2 + < u, v >2
||u +e v||2 =
(1+ < u, v >)2
 15C. CHARGES AND FLUX 359

But this type of formula is exactly what we need, for what we want to do. Indeed, by
assuming ||u||, ||v|| < 1, we have the following estimate:
||u +e v||2 < 1 ⇐⇒ ||u + v||2 − ||u||2 ||v||2 + < u, v >2 < (1+ < u, v >)2
⇐⇒ ||u + v||2 − ||u||2 ||v||2 < 1 + 2 < u, v >
⇐⇒ ||u||2 + ||v||2 − ||u||2 ||v||2 < 1
⇐⇒ (1 − ||u||2 )(1 − ||v||2 ) > 0
Thus, we are led to the conclusion in the statement.
(3) This is something that we know from Proposition 15.20.
(4) This comes from the squared norm formula established in the proof of (2) above,
because when assuming ||v|| = 1, we obtain:
||u + v||2 − ||u||2 + < u, v >2
||u +e v||2 =
(1+ < u, v >)2
||u||2 + 1 + 2 < u, v > −||u||2 + < u, v >2
=
(1+ < u, v >)2
1 + 2 < u, v > + < u, v >2
=
(1+ < u, v >)2
= 1
(5) This is clear, from the obvious lack of symmetry of our formula.
(6) This is again clear, from the obvious lack of symmetry of our formula. □
That was nice, all this mathematics, and hope you’re still with me. And good news,
the formula in Theorem 15.22 is the good one, confirmed by experimental physics.

15c. Charges and flux

Time for electricity. In analogy with the usual study of gravity, let us start with:
Definition 15.23. Given charges q1 , . . . , qk ∈ R located at positions x1 , . . . , xk ∈ R3 ,
we define their electric field to be the vector function
X qi (x − xi )
E(x) = K
i
||x − xi ||3
so that their force applied to a charge Q ∈ R positioned at x ∈ R3 is given by F = QE.
More generally, we will be interested in electric fields of various non-discrete config-
urations of charges, such as charged curves, surfaces and solid bodies. Indeed, things
like wires or metal sheets or solid bodies coming in all sorts of shapes, tailored for their
purpose, play a key role, so this extension is essential. So, let us go ahead with:
360 15. PARTIAL INTEGRATION

Definition 15.24. The electric field of a charge configuration L ⊂ R3 , with charge

density function ρ : L → R, is the vector function
ρ(z)(x − z)
Z
E(x) = K 3
dz
L ||x − z||

so that the force of L applied to a charge Q positioned at x is given by F = QE.

With the above definitions in hand, it is most convenient now to forget about the
charges, and focus on the study of the corresponding electric fields E.

These fields are by definition vector functions E : R3 → R3 , with the convention that
they take ±∞ values at the places where the charges are located, and intuitively, are best
represented by their field lines, which are constructed as follows:
Definition 15.25. The field lines of an electric field E : R3 → R3 are the oriented
curves γ ⊂ R3 pointing at every point x ∈ R3 at the direction of the field, E(x) ∈ R3 .
As a basic example here, for one charge the field lines are the half-lines emanating
from its position, oriented according to the sign of the charge:
↖ ↑ ↗ ↘ ↓ ↙
← ⊕ → → ⊖ ←
↙ ↓ ↘ ↗ ↑ ↖
For two charges now, if these are of opposite signs, + and −, you get a picture that
you are very familiar with, namely that of the field lines of a bar magnet:
↗ ↗ → → → → ↘ ↘
↖ ↑ ↗ → → ↘ ↓ ↙
← ⊕ → → → → ⊖ ←
↙ ↓ ↘ → → ↗ ↑ ↖
↘ ↘ → → → → ↗ ↗
If the charges are +, + or −, −, you get something of similar type, but repulsive this
time, with the field lines emanating from the charges being no longer shared:
← ↖ ↖ ↗ ↗ →
↑ ↗ ↖ ↑
← ⊕ ⊕ →
↓ ↘ ↙ ↓
← ↙ ↙ ↘ ↘ →
These pictures, and notably the last one, with +, + charges, are quite interesting,
because the repulsion situation does not appear in the context of gravity. Thus, we can
only expect our geometry here to be far more complicated than that of gravity.
 15C. CHARGES AND FLUX 361

The field lines, as constructed in Definition 15.25, obviously do not encapsulate the
whole information about the field, with the direction of each vector E(x) ∈ R3 being
there, but with the magnitude ||E(x)|| ≥ 0 of this vector missing. However, say when
drawing, when picking up uniformly radially spaced field lines around each charge, and
with the number of these lines proportional to the magnitude of the charge, and then
completing the picture, the density of the field lines around each point x ∈ R will give
you then the magnitude ||E(x)|| ≥ 0 of the field there, up to a scalar.

Let us summarize these observations as follows:

Proposition 15.26. Given an electric field E : R3 → R3 , the knowledge of its field
lines is the same as the knowledge of the composition
nE : R3 → R3 → S
where S ⊂ R3 is the unit sphere, and n : R3 → S is the rescaling map, namely:
x
n(x) =
||x||
However, in practice, when the field lines are accurately drawn, the density of the field
lines gives you the magnitude of the field, up to a scalar.
Proof. We have two assertions here, the idea being as follows:
(1) The first assertion is clear from definitions, with of course our usual convention
that the electric field and its problematics take place outside the locations of the charges,
which makes everything in the statement to be indeed well-defined.
(2) Regarding now the last assertion, which is of course a bit informal, this follows
from the above discussion. It is possible to be a bit more mathematical here, with a
definition, formula and everything, but we will not need this, in what follows. □
Let us introduce now a key definition, as follows:
Definition 15.27. The flux of an electric field E : R3 → R3 through a surface S ⊂ R3 ,
assumed to be oriented, is the quantity
Z
ΦE (S) = < E(x), n(x) > dx
S

with n(x) being unit vectors orthogonal to S, following the orientation of S. Intuitively,
the flux measures the signed number of field lines crossing S.
Here by orientation of S we mean precisely the choice of unit vectors n(x) as above,
orthogonal to S, which must vary continuously with x. For instance a sphere has two
possible orientations, one with all these vectors n(x) pointing inside, and one with all
these vectors n(x) pointing outside. More generally, any surface has locally two possible
362 15. PARTIAL INTEGRATION

orientations, so if it is connected, it has two possible orientations. In what follows the

convention is that the closed surfaces are oriented with each n(x) pointing outside.

As a first illustration, let us do a basic computation, as follows:

Proposition 15.28. For a point charge q ∈ R at the center of a sphere S,
q
ΦE (S) =
ε0
where the constant is ε0 = 1/(4πK), independently of the radius of S.
Proof. Assuming that S has radius r, we have the following computation:
Z
ΦE (S) = < E(x), n(x) > dx
S
Z  
Kqx x
= , dx
S r3 r
Z
Kq
= 2
dx
S r
Kq
= × 4πr2
r2
= 4πKq
Thus with ε0 = 1/(4πK) as above, we obtain the result. □
More generally now, we have the following result:
Theorem 15.29. The flux of a field E through a sphere S is given by
Qenc
ΦE (S) =
ε0
where Qenc is the total charge enclosed by S, and ε0 = 1/(4πK).
Proof. This can be done in several steps, as follows:
(1) Before jumping into computations, let us do some manipulations. First, by dis-
cretizing the problem, we can assume that we are dealing with a system of point charges.
Moreover, by additivity, we can assume that we are dealing with a single charge. And if
we denote by q ∈ R this charge, located at v ∈ R3 , we want to prove that we have the
following formula, where B ⊂ R3 denotes the ball enclosed by S:
q
ΦE (S) = δv∈B
ε0
(2) By linearity we can assume that we are dealing with the unit sphere S. Moreover,
by rotating we can assume that our charge q lies on the Ox axis, that is, that we have
 15C. CHARGES AND FLUX 363

v = (r, 0, 0) with r ≥ 0, r ̸= 1. The formula that we want to prove becomes:

q
ΦE (S) = δr<1
ε0

(3) Let us start now the computation. With u = (x, y, z), we have:
Z
ΦE (S) = < E(u), u > du
S
Z  
Kq(u − v)
= , u du
S ||u − v||3
< u − v, u >
Z
= Kq du
S ||u − v||3
Z
1− < v, u >
= Kq du
S ||u − v||3
1 − rx
Z
= Kq 2 3/2
du
S (1 − 2xr + r )

(4) In order to compute the above integral, we will use spherical coordinates for the
unit sphere S, which are as follows, with s ∈ [0, π] and t ∈ [0, 2π]:

x = cos s

y = sin s cos t

z = sin s sin t

We recall that the corresponding Jacobian, computed before, is given by:

J = sin s

(5) With the above change of coordinates, our integral from (3) becomes:

1 − rx
Z
ΦE (S) = Kq 2 3/2
du
S (1 − 2xr + r )
Z 2π Z π
1 − r cos s
= Kq 2 3/2
· sin s ds dt
0 0 (1 − 2r cos s + r )
Z π
(1 − r cos s) sin s
= 2πKq 2 3/2
ds
0 (1 − 2r cos s + r )
Z π
q (1 − r cos s) sin s
= ds
2ε0 0 (1 − 2r cos s + r2 )3/2
364 15. PARTIAL INTEGRATION

(6) The point now is that the integral on the right can be computed with the change
of variables x = cos s. Indeed, we have dx = − sin s ds, and we obtain:
Z π Z 1
(1 − r cos s) sin s 1 − rx
2 3/2
ds = 2 3/2
dx
0 (1 − 2r cos s + r ) −1 (1 − 2rx + r )
 1
x−r
= √
1 − 2rx + r2 −1
1−r −1 − r
= √ −√
1 − 2r + r2 1 + 2r + r2
1−r
= +1
|1 − r|
= 2δr<1
Thus, we are led to the formula in the statement. □

More generally now, we have the following key result, due to Gauss:
Theorem 15.30 (Gauss law). The flux of a field E through a surface S is given by
Qenc
ΦE (S) =
ε0
where Qenc is the total charge enclosed by S, and ε0 = 1/(4πK).
Proof. This basically follows from Theorem 15.29, or even from Proposition 15.28,
by adding to the results there a number of new ingredients, as follows:
(1) Our first claim is that given a closed surface S, with no charges inside, the flux
through it of any choice of external charges vanishes:
ΦE (S) = 0
This claim is indeed supported by the intuitive interpretation of the flux, as corre-
sponding to the signed number of field lines crossing S. Indeed, any field line entering as
+ must exit somewhere as −, and vice versa, so when summing we get 0.
(2) In practice now, in order to prove this rigorously, there are several ways. A standard
argument, which is quite elementary, is the one used by Feynman in [32], based on the
fact that, due to F ∼ 1/d2 , local deformations of S will leave invariant the flux, and so
in the end we are left with a rotationally invariant surface, where the result is clear.
(3) The point now is that, with this and Proposition 15.28 in hand, we can finish by
using a standard math trick. Let us assume indeed, by discretizing, that our system of
charges is discrete, consisting of enclosed charges q1 , . . . , qk ∈ R, and an exterior total
 15D. GAUSS, GREEN, STOKES 365

charge Qext . We can surround each of q1 , . . . , qk by small disjoint spheres U1 , . . . , Uk ,

chosen such that their interiors do not touch S, and we have:
ΦE (S) = ΦE (S − ∪Ui ) + ΦE (∪Ui )
= 0 + ΦE (∪Ui )
X
= ΦE (Ui )
i
X qi
=
i
ε0
Qenc
=
ε0
(4) To be more precise, in the above the union ∪Ui is a usual disjoint union, and
the flux is of course additive over components. As for the difference S − ∪Ui , this is by
definition the disjoint union of S with the disjoint union ∪(−Ui ), with each −Ui standing
for Ui with orientation reversed, and since this difference has no enclosed charges, the flux
through it vanishes by (2). Finally, the end makes use of Proposition 15.28. □
So long for the Gauss law. We will be back to it in a moment, with a new, better
proof, by using some advanced 3D calculus, that we will have to learn first.

15d. Gauss, Green, Stokes

Let us start with a standard definition, immersing us into 3D problematics, as follows:
Definition 15.31. Given a function f : R3 → R, its usual derivative f ′ (u) ∈ R3 can
be written as f ′ (u) = ∇f (u), where the gradient operator ∇ is given by:
d
dx
d 
∇ =  dy
d
dz

By using ∇, we can talk about the divergence of a function φ : R3 → R3 , as being

* d   +
dx φx
d   dφx dφy dφz
< ∇, φ >=  dy , φy  = + +
d φ dx dy dz
dz z

as well as about the curl of the same function φ : R3 → R3 , as being

 dφy

d dφz
ux dx φx dy
− dz
d dφ dφ 
∇ × φ = uy dy φy =  dzx − dxz 
d dφy
uz dz φz dx
− dφ
dy
x

where ux , uy , uz are the unit vectors along the coordinate directions x, y, z.

366 15. PARTIAL INTEGRATION

All this might seem a bit abstract, but is in fact very intuitive. The gradient ∇f points
in the direction of the maximal increase of f , with |∇f | giving you the rate of increase
of f , in that direction. As for the divergence and curl, these measure the divergence and
curl of the vectors φ(u + v) around a given point u ∈ R3 , in a usual, real-life sense.
Getting back now to calculus tools, what was missing from our picture was the higher
dimensional analogue of the fundamental theorem of calculus, and more generally of the
partial integration formula. In 3 dimensions, we have the following result:
Theorem 15.32. The following results hold, in 3 dimensions:
(1) Fundamental theorem for gradients, namely
Z b
< ∇f, dx >= f (b) − f (a)
a
(2) Fundamental theorem for divergences, or Gauss or Green formula,
Z Z
< ∇, φ >= < φ(x), n(x) > dx
B S
(3) Fundamental theorem for curls, or Stokes formula,
Z Z
< (∇ × φ)(x), n(x) > dx = < φ(x), dx >
A P
where S is the boundary of the body B, and P is the boundary of the area A.
Proof. This is a mixture of trivial and non-trivial results, as follows:
(1) This is something that we know well in 1D, namely the fundamental theorem of
calculus, and the general, N -dimensional formula follows from that.
(2) This is something more subtle, and we had a taste of it when dealing with the
Gauss law, and its various proofs. In general, the proof is similar, by using the various
ideas from the proof of the Gauss law, and this can be worked out.
(3) This is again something subtle, and again with a flavor of things that we know,
from the proof of the Gauss law, and which can be again worked out. □
Getting back now to electrostatics, as a main application of the above, we have the
following new point of view on the Gauss formula, which is more conceptual:
Theorem 15.33 (Gauss). Given an electric potential E, its divergence is given by
ρ
< ∇, E >=
ε0
where ρ denotes as usual the charge distribution. Also, we have
∇×E =0
meaning that the curl of E vanishes.
 15D. GAUSS, GREEN, STOKES 367

Proof. We have several assertions here, the idea being as follows:

(1) The first formula, called Gauss law in differential form, follows from:

Z Z
< ∇, E > = < E(x), n(x) > dx
B S
= ΦE (S)
Qenc
=
ε
Z 0
ρ
=
B ε0

Now since this must hold for any B, this gives the formula in the statement.

(2) As a side remark, the Gauss law in differential form can be established as well
directly, with the computation, involving a Dirac mass, being as follows:

 
ρ(z)(x − z)
Z
< ∇, E > (x) = ∇, K 3
dz
R3 ||x − z||
Z  
x−z
= K ∇, ρ(z) dz
R3 ||x − z||3
Z
= K 4πδx · ρ(z)dz
R3
Z
= 4πK δx ρ(z)dz
R3
ρ(x)
=
ε0

And with this in hand, we have via (1) a new proof of the usual Gauss law.
368 15. PARTIAL INTEGRATION

(3) Regarding the curl, by discretizing and linearity we can assume that we are dealing
with a single charge q, positioned at 0. We have, by using spherical coordinates r, s, t:
Z b Z b 
Kqx
< E(x), dx > = , dx
a a ||x||3
Z b 
Kq x
= · , dx
a r2 ||x||
Z b
Kq
= 2
dr
a r
 b
Kq
= −
r
 a 
1 1
= Kq −
ra rb
In particular the integral of E over any closed loop vanishes, and by using now Stokes’
theorem, we conclude that the curl of E vanishes, as stated.
(4) Finally, as a side remark, both the formula of the divergence and the vanishing of
the curl are somewhat clear by looking at the field lines of E. However, as all the above
mathematics shows, there is certainly something to be understood, in all this. □
So long for electrostatics, which provide a good motivation and illustration for our
mathematics. When upgrading to electrodynamics, things become even more interesting,
because our technology can be used in order to understand the Maxwell equations.
15e. Exercises
We have a tough chapter here, and as exercises, we have:
Exercise 15.34. Work out some numerics for the Coriolis force, on Earth.
Exercise 15.35. Learn about Einstein summation via Lorentz transformation.
Exercise 15.36. Prove the Gauss law by counting the flux lines. Can you?
Exercise 15.37. Find some further applications of Gauss, Green, Stokes.
As bonus exercise, learn full electrodynamics, say from Griffiths [41].
 CHAPTER 16

Infinite dimensions

16a. Operators, matrices

Welcome to quantum mechanics, and in the hope that we will survive. We already
talked in chapter 8 about the main idea of Heisenberg, namely using infinite matrices in
order to axiomatize quantum mechanics, based on the following key fact:
Fact 16.1 (Rydberg, Ritz). The spectral lines of the hydrogen atom are given by the
Rydberg formula, as follows, depending on integer parameters n1 < n2 :
 
1 1 1
=R −
λn1 n2 n21 n22
These spectral lines combine according to the Ritz-Rydberg principle, as follows:
1 1 1
+ =
λn1 n2 λn2 n3 λn1 n3
Similar formulae hold for other atoms, with suitable fine-tunings of the constant R.
We refer to chapter 8 for the full story with all this, which is theory based on some
key observations of Lyman, Balmer, Paschen, around 1890-1900. The point now is that
the above combination principle reminds the multiplication formula en1 n2 en2 n3 = en1 n3 for
the elementary matrices eij : ej → ei , which leads to the following principle:
Principle 16.2 (Heisenberg). Observables in quantum mechanics should be some sort
of infinite matrices, generalizing the Lyman, Balmer, Paschen lines of the hydrogen atom,
and multiplying between them as the matrices do, as to produce further observables.
All this is quite deep, and needs a number of comments, as follows:

(1) First of all, our matrices must be indeed infinite, because so are the series observed
by Lyman, Balmer, Paschen, corresponding to n1 = 1, 2, 3 in the Rydberg formula, and
making it clear that the range of the second parameter n2 > n1 is up to ∞.

(2) Although this was not known to Ritz-Rydberg and Heisenberg, let us mention too
that some later results of Brackett, Pfund, Humphreys and others, at n1 = 4, 5, 6, . . . ,
confirmed the fact that the range of the first parameter n1 is up to ∞ too.
369
370 16. INFINITE DIMENSIONS

(3) As a more tricky comment now, going beyond what Principle 16.2 says, our infinite
matrices must be in fact complex. This was something known to Heisenberg, and later
Schrödinger came with proof that quantum mechanics naturally lives over C.

(4) But all this leads us into some tricky mathematics, because the infinite matrices
A ∈ M∞ (C) do not act on the vectors v ∈ C∞ just like that. For instance the all-one
matrix Aij = 1 does not act on the all-one vector vi = 1, for obvious reasons.

Summarizing, in order to get to some mathematical theory going, out of Principle

16.2, we must assume that our matrices A ∈ M∞ (C) must be “bounded” in some sense.
Or perhaps the vectors v ∈ C∞ must be bounded. Or perhaps, both.

In order to fix all this, let us start with C∞ . We would like to replace it with its sub-
space H = l2 (N) consisting of vectors having finite norm, as for our various computations
to converge. But this being said, taking a look at what Schrödinger was saying too, a bit
later, why not including right away in our theory spaces like H = L2 (R3 ) too, which are
perhaps a bit more relevant than Heisenberg’s l2 (N). We are led in this way into:
Definition 16.3. A Hilbert space is a complex vector space H with a scalar product
< x, y >, which will be linear at left and antilinear at right,
< λx, y >= λ < x, y > , < x, λy >= λ̄ < x, y >
and which is complete with respect to corresponding norm
√
||x|| = < x, x >
in the sense that any sequence {xn } which is a Cauchy sequence, having the property
||xn − xm || → 0 with n, m → ∞, has a limit, xn → x.
Here our convention for the scalar products, written < x, y > and being linear at left,
is one among others, often used by mathematicians, and we will just use this, in the lack
of a physicist with an axe around. As further comments now on Definition 16.3, there is
some mathematics encapsulated there, needing some discussion. First, we have:
Theorem 16.4. Given an index set I, which can be finite or not, the space of square-
summable vectors having indices in I, namely
( )
X
l2 (I) = (xi )i∈I |xi |2 < ∞
i

is a Hilbert space, with scalar product as follows:

X
< x, y >= xi ȳi
i

When I is finite, I = {1, . . . , N }, we obtain in this way the usual space H = CN .

 16A. OPERATORS, MATRICES 371

Proof. We have already met such things in chapter 7, but let us recall all this:
(1) We know that l2 (I) ⊂ CI is the space of vectors satisfying ||x|| < ∞. We want
to prove that l2 (I) is a vector space, that < x, y > is a scalar product on it, that l2 (I) is
complete with respect to ||.||, and finally that for |I| < ∞ we have l2 (I) = C|I| .
(2) The last assertion, l2 (I) = C|I| for |I| < ∞, is clear, because in this case the sums
are finite, so the condition ||x|| < ∞ is automatic. So, we know at least one thing.
(3) Regarding the rest, our claim here, which will more or less prove everything, is
that for any two vectors x, y ∈ l2 (I) we have the Cauchy-Schwarz inequality:
| < x, y > | ≤ ||x|| · ||y||
But this follows from the positivity of the following degree 2 quantity, depending on
a real variable t ∈ R, and on a variable on the unit circle, w ∈ T:
f (t) = ||twx + y||2
(4) Now with Cauchy-Schwarz proved, everything is straightforward. We first obtain,
by raising to the square and expanding, that for any x, y ∈ l2 (I) we have:
||x + y|| ≤ ||x|| + ||y||
Thus l2 (I) is indeed a vector space, the other vector space conditions being trivial.
(5) Also, < x, y > is surely a scalar product on this vector space, because all the
conditions for a scalar product are trivially satisfied.
(6) Finally, the fact that our space l2 (I) is indeed complete with respect to its norm
||.|| follows in the obvious way, the limit of a Cauchy sequence {xn } being the vector
y = (yi ) given by yi = limn→∞ xni , with all the verifications here being trivial. □
Going now a bit abstract, we have, more generally, the following result, which shows
that our formalism covers as well the Schrödinger spaces of type L2 (R3 ):
Theorem 16.5. Given an arbitrary space X with a positive measure µ on it, the space
of square-summable complex functions on it, namely
 Z 
2 2
L (X) = f : X → C |f (x)| dµ(x) < ∞
X

is a Hilbert space, with scalar product as follows:

Z
< f, g >= f (x)g(x) dµ(x)
X

When X = I is discrete, meaning that the measure µ on it is the counting measure,

µ({x}) = 1 for any x ∈ X, we obtain in this way the previous spaces l2 (I).
372 16. INFINITE DIMENSIONS

Proof. This is something routine, remake of Theorem 16.4, as follows:

(1) The proof of the first, and main assertion is something perfectly similar to the
proof of Theorem 16.4, by replacing everywhere the sums by integrals.
(2) With the remark that we forgot to say in the statement that the L2 functions are
by definition taken up to equality almost everywhere, f = g when ||f − g|| = 0.
(3) As for the last assertion, when µ is the counting measure all our integrals here
become usual sums, and so we recover in this way Theorem 16.4. □
As a third and last theorem about Hilbert spaces, that we will need, we have:
Theorem 16.6. Any Hilbert space H has an orthonormal basis {ei }i∈I , which is by
definition a set of vectors whose span is dense in H, and which satisfy
< ei , ej >= δij
with δ being a Kronecker symbol. The cardinality |I| of the index set, which can be finite,
countable, or worse, depends only on H, and is called dimension of H. We have
H ≃ l2 (I)
P
in the obvious way, mapping λi ei → (λi ). The Hilbert spaces with dim H = |I| being
2 2
countable, including l (N) and L (R), are all isomorphic, and are called separable.
Proof. We have many assertions here, the idea being as follows:
(1) In finite dimensions an orthonormal basis {ei }i∈I can be constructed by starting
with any vector space basis {xi }i∈I , and using the Gram-Schmidt procedure. As for the
other assertions, these are all clear, from basic linear algebra.
(2) In general, the same method works, namely Gram-Schmidt, with a subtlety coming
from the fact that the basis {ei }i∈I will not span in general the whole H, but just a dense
subspace of it, as it is in fact obvious by looking at the standard basis of l2 (N).
(3) And there is a second subtlety as well, coming from the fact that the recurrence
procedure needed for Gram-Schmidt must be replaced by some sort of “transfinite recur-
rence”, using scary tools from logic, and more specifically the Zorn lemma.
(4) Finally, everything at the end is clear from definitions, except perhaps for the fact
that L2 (R) is separable. But here we can argue that, since functions can be approxi-
mated by polynomials, we have a countable algebraic basis, namely {xn }n∈N , called the
Weierstrass basis, that we can orthogonalize afterwards by using Gram-Schmidt. □
Moving ahead, now that we know what our vector spaces are, we can talk about
infinite matrices with respect to them. And the situation here is as follows:
 16A. OPERATORS, MATRICES 373

Theorem 16.7. Given a Hilbert space H, consider the linear operators T : H → H,

and for each such operator define its norm by the following formula:
||T || = sup ||T x||
||x||=1

The operators which are bounded, ||T || < ∞, form then a complex algebra B(H), which
is complete with respect to ||.||. When H comes with a basis {ei }i∈I , we have
B(H) ⊂ L(H) ⊂ MI (C)
where L(H) is the algebra of all linear operators T : H → H, and L(H) ⊂ MI (C) is the
correspondence T → M obtained via the usual linear algebra formulae, namely:
T (x) = M x , Mij =< T ej , ei >
In infinite dimensions, none of the above two inclusions is an equality.
Proof. This is something straightforward, the idea being as follows:
(1) The fact that we have indeed an algebra, satisfying the product condition in the
statement, follows from the following estimates, which are all elementary:
||S + T || ≤ ||S|| + ||T || , ||λT || = |λ| · ||T || , ||ST || ≤ ||S|| · ||T ||
(2) Regarding now the completness assertion, if {Tn } ⊂ B(H) is Cauchy then {Tn x}
is Cauchy for any x ∈ H, so we can define the limit T = limn→∞ Tn by setting:
T x = lim Tn x
n→∞

Let us first check that the application x → T x is linear. We have:

T (x + y) = lim Tn (x + y)
n→∞
= lim Tn (x) + Tn (y)
n→∞
= lim Tn (x) + lim Tn (y)
n→∞ n→∞
= T (x) + T (y)
Similarly, we have T (λx) = λT (x), and we conclude that T ∈ L(H).
(3) With this done, it remains to prove now that we have T ∈ B(H), and that Tn → T
in norm. For this purpose, observe that we have:
||Tn − Tm || ≤ ε , ∀n, m ≥ N =⇒ ||Tn x − Tm x|| ≤ ε , ∀||x|| = 1 , ∀n, m ≥ N
=⇒ ||Tn x − T x|| ≤ ε , ∀||x|| = 1 , ∀n ≥ N
=⇒ ||TN x − T x|| ≤ ε , ∀||x|| = 1
=⇒ ||TN − T || ≤ ε
But this gives both T ∈ B(H), and TN → T in norm, and we are done.
374 16. INFINITE DIMENSIONS

(4) Regarding the embeddings, the correspondence T → M in the statement is indeed

linear, and its kernel is {0}, so we have indeed an embedding as follows, as claimed:
L(H) ⊂ MI (C)
In finite dimensions we have an isomorphism, because any M ∈ MN (C) determines
an operator T : CN → CN , given by < T ej , ei >= Mij . However, in infinite dimensions,
we have matrices not producing operators, as for instance the all-one matrix.
(5) As for the examples of linear operators which are not bounded, these are more
complicated, coming from logic, and we will not need them in what follows. □
Finally, as a second and last result regarding the operators, we will need:
Theorem 16.8. Each operator T ∈ B(H) has an adjoint T ∗ ∈ B(H), given by:
< T x, y >=< x, T ∗ y >
The operation T → T ∗ is antilinear, antimultiplicative, involutive, and satisfies:
||T || = ||T ∗ || , ||T T ∗ || = ||T ||2
When H comes with a basis {ei }i∈I , the operation T → T ∗ corresponds to
(M ∗ )ij = M ji
at the level of the associated matrices M ∈ MI (C).
Proof. This is standard too, and can be proved in 3 steps, as follows:
(1) The existence of the adjoint operator T ∗ , given by the formula in the statement,
comes from the fact that the function φ(x) =< T x, y > being a linear map H → C, we
must have a formula as follows, for a certain vector T ∗ y ∈ H:
φ(x) =< x, T ∗ y >
Moreover, since this vector is unique, T ∗ is unique too, and we have as well:
(S + T )∗ = S ∗ + T ∗ , (λT )∗ = λ̄T ∗ , (ST )∗ = T ∗ S ∗ , (T ∗ )∗ = T
Observe also that we have indeed T ∗ ∈ B(H), because:
||T || = sup sup < T x, y >
||x||=1 ||y||=1

= sup sup < x, T ∗ y >

||y||=1 ||x||=1
∗
= ||T ||
(2) Regarding now ||T T ∗ || = ||T ||2 , which is a key formula, observe that we have:
||T T ∗ || ≤ ||T || · ||T ∗ || = ||T ||2
 16B. SCHRÖDINGER EQUATION 375

On the other hand, we have as well the following estimate:

||T ||2 = sup | < T x, T x > |
||x||=1

= sup | < x, T ∗ T x > |

||x||=1
∗
≤ ||T T ||
By replacing T → T ∗ we obtain from this ||T ||2 ≤ ||T T ∗ ||, as desired.
(3) Finally, when H comes with a basis, the formula < T x, y >=< x, T ∗ y > applied
with x = ei , y = ej translates into the formula (M ∗ )ij = M ji , as desired. □
So long for Hilbert spaces and operators. For more, you can check my book [10].
16b. Schrödinger equation
Time now to get into quantum mechanics. This is a bit of a nightmare, and I am
afraid that I will have to ask again the cat. With him, things going better since last time,
we sort of agree now on multivariable calculus, following a long philosophical debate, and
my willingness to start sharing his diet, namely raw mice and birds. But let’s see now in
relation with this highly divisive topic which is quantum mechanics:
Cat 16.9. Schrödinger.
Thanks cat, this is actually quite surprising, and I am particularly pleased to hear
this. So, before getting back to what Heisenberg was saying, based on Lyman, Balmer,
Paschen, namely developing some sort of “matrix mechanics”, let us hear as well the point
of view of Schrödinger, which came a few years later. His idea was to forget about exact
things, and try to investigate the hydrogen atom statistically. Let us start with:
Question 16.10. In the context of the hydrogen atom, assuming that the proton is
fixed, what is the probability density φt (x) of the position of the electron e, at time t,
Z
Pt (e ∈ V ) = φt (x)dx
V
as function of an intial probability density φ0 (x)? Moreover, can the corresponding equa-
tion be solved, and will this prove the Bohr claims for hydrogen, statistically?
In order to get familiar with this question, let us first look at examples coming from
classical mechanics. In the context of a particle whose position at time t is given by
x0 + γ(t), the evolution of the probability density will be given by:
φt (x) = φ0 (x) + γ(t)
However, such examples are somewhat trivial, of course not in relation with the com-
putation of γ, usually a difficult question, but in relation with our questions, and do not
apply to the electron. The point indeed is that, in what regards the electron, we have:
376 16. INFINITE DIMENSIONS

Fact 16.11. In respect with various simple interference experiments:

(1) The electron is definitely not a particle in the usual sense.
(2) But in most situations it behaves exactly like a wave.
(3) But in other situations it behaves like a particle.
Getting back now to the Schrödinger question, all this suggests to use, as for the
waves, an amplitude function ψt (x) ∈ C, related to the density φt (x) > 0 by the formula
φt (x) = |ψt (x)|2 . Not that a big deal, you would say, because the two are related by
simple formulae as follows, with θt (x) being an arbitrary phase function:
p
φt (x) = |ψt (x)|2 , ψt (x) = eiθt (x) φt (x)
However, such manipulations can be crucial, raising for instance the possibility that
the amplitude function satisfies some simple equation, while the density itself, maybe not.
And this is what happens indeed. Schrödinger was led in this way to:
Claim 16.12 (Schrödinger). In the context of the hydrogen atom, the amplitude func-
tion of the electron ψ = ψt (x) is subject to the Schrödinger equation
h2
ihψ̇ = − ∆ψ + V ψ
2m
m being the mass, h = h0 /2π the reduced Planck constant, and V the Coulomb potential
of the proton. The same holds for movements of the electron under any potential V .
Observe the similarity with the wave equation φ̈ = v 2 ∆φ, and with the heat equation
φ̇ = α∆φ too. Many things can be said here. Following now Heisenberg and Schrödinger,
and then especially Dirac, who did the axiomatization work, we have:
Definition 16.13. In quantum mechanics the states of the system are vectors of a
Hilbert space H, and the observables of the system are linear operators
T :H→H
which can be densely defined, and are taken self-adjoint, T = T ∗ . The average value of
such an observable T , evaluated on a state ξ ∈ H, is given by:
< T >=< T ξ, ξ >
In the context of the Schrödinger mechanics of the hydrogen atom, the Hilbert space is the
space H = L2 (R3 ) where the wave function ψ lives, and we have
Z
< T >= T (ψ) · ψ̄ dx
R3
which is called “sandwiching” formula, with the operators
ih h2 ∆ h2 ∆
x , − ∇ , −ih∇ , − , − +V
m 2m 2m
representing the position, speed, momentum, kinetic energy, and total energy.
 16C. SPHERICAL COORDINATES 377

In other words, we are doing here two things. First, we are declaring by axiom that
various “sandwiching” formulae found before by Heisenberg, involving the operators at
the end, that we will not get into in detail here, hold true. And second, we are raising
the possibility for other quantum mechanical systems, more complicated, to be described
as well by the mathematics of the operators on a certain Hilbert space H, as above.

So, this was the story of early quantum mechanics, over-simplified as to fit here in
a few pages. For more, you can check Feynman [33] for foundations, and everything,
including for some nice pictures and explanations regarding Fact 16.11. You have as well
Griffiths [42] or Weinberg [94], for further explanations on Definition 16.13, not to forget
Dirac’s original text [22], and all this is discussed as well in my book [11].

16c. Spherical coordinates

In order to solve now the hydrogen atom, the idea will be that of reformulating the
Schrödinger equation in spherical coordinates. And for this purpose, we will need:
Theorem 16.14. The Laplace operator in spherical coordinates is
d2
   
1 d 2 d 1 d d 1
∆= 2 · r · + 2 · sin s · + 2 2 · 2
r dr dr r sin s ds ds r sin s dt
with our standard conventions for these coordinates, in 3D.
Proof. There are several proofs here, a short, elementary one being as follows:
(1) Let us first see how ∆ behaves under a change of coordinates {xi } → {yi }, in
arbitrary N dimensions. Our starting point is the chain rule for derivatives:
d X d dyj
= ·
dxi j
dy j dxi

By using this rule, then Leibnitz for products, then again this rule, we obtain:
d2 f X d  df dyj 
= ·
dx2i j
dx i dy j dxi
X d  df  dyj df d

dyj

= · + ·
j
dxi dyj dxi dyj dxi dxi
!
X X d dyk  df  dyj df d2 yj
= · · + ·
j k
dyk dx i dy j dx i dy j dx2i
X d2 f dyk dyj X df d2 yj
= · · + ·
jk
dyk dyj dx i dxi j
dyj dx2i
378 16. INFINITE DIMENSIONS

(2) Now by summing over i, we obtain the following formula, with A being the deriv-
ative of x → y, that is to say, the matrix of partial derivatives dyi /dxj :
X d2 f dyk dyj X df d2 yj
∆f = · · + ·
ijk
dy k dyj dx i dxi ij
dyj dx2i
X d2 f X d2 yj df
= Aki Aji + ·
ijk
dyk dyj ij
dx2i dyj
X d2 f X df
= (AAt )jk + ∆(yj )
jk
dyk dyj j
dyj

(3) So, this will be the formula that we will need. Observe that this formula can be
further compacted as follows, with all the notations being self-explanatory:

∆f = T r(AAt Hy (f ))+ < ∆(y), ∇y (f ) >

(4) Getting now to spherical coordinates, (x, y, z) → (r, s, t), the derivative of the
inverse, obtained by differentiating x, y, z with respect to r, s, t, is given by:
 
cos s −r sin s 0
A−1 = sin s cos t r cos s cos t −r sin s sin t
sin s sin t r cos s sin t r sin s cos t

The product (A−1 )t A−1 of the transpose of this matrix with itself is then:
  
cos s sin s cos t sin s sin t cos s −r sin s 0
−r sin s r cos s cos t r cos s sin t sin s cos t r cos s cos t −r sin s sin t
0 −r sin s sin t r sin s cos t sin s sin t r cos s sin t r sin s cos t

But everything simplifies here, and we have the following remarkable formula, which
by the way is something very useful, worth to be memorized:
 
1 0 0
(A−1 )t A−1 = 0 r2 0 
0 0 r sin2 s
2

Now by inverting, we obtain the following formula, in relation with the above:
 
1 0 0
AAt = 0 1/r2 0 
2 2
0 0 1/(r sin s)
 16C. SPHERICAL COORDINATES 379

(5) Let us compute now the Laplacian of r, s, t. We first have the following formula,
that we will use many times in what follows, and is worth to be memorized:

dr dp 2
= x + y2 + z2
dx dx
1 2x
= ·p
2 x2 + y 2 + z 2
x
=
r
Of course the same computation works for y, z too, and we therefore have:

dr x dr y dr z
= , = , =
dx r dy r dz r

(6) By using the above formulae, twice, we can compute the Laplacian of r:
p 
∆(r) = ∆ x2 + y 2 + z 2
d x d y  d z 
= + +
dx r dy r dz r
2 2 2 2
r −x r −y r2 − z 2
= + +
r3 r3 r3
2
=
r

(7) In what regards now s, the computation here goes as follows:

  x 
∆(s) = ∆ arccos
r
 √ 2     
d r − x2 d xy d xz
= − + √ + √
dx r2 dy r2 r2 − x2 dz r2 r2 − x2
√
2x r2 − x2 r2 (z 2 − 2y 2 ) + 2x2 y 2 r2 (y 2 − 2z 2 ) + 2x2 z 2
= + √ + √
r4 r 4 r 2 − x2 r 4 r 2 − x2
√
2x r2 − x2 x(2x2 − r2 )
= + √
r4 r 4 r 2 − x2
x
= √
r r 2 − x2
2
cos s
= 2
r sin s
380 16. INFINITE DIMENSIONS

(8) Finally, in what regards t, the computation here goes as follows:

  
z
∆(t) = ∆ arctan
y
   
d d z d y
= (0) + − 2 +
dx dy y + z2 dz y 2 + z 2
2yz 2yz
= 0− 2 2 2
+ 2
(y + z ) (y + z 2 )2
= 0
(9) We can now plug the data from (4) and (6,7,8) in the general formula that we
found in (2) above, and we obtain in this way:
d2 f 1 d2 f 1 d2 f 2 df cos s df
∆f = 2
+ 2
· 2 + 2 2 · 2
+ · + 2 ·
dr r ds r sin s dt r dr r sin s ds
2
2 df df cos s df 1 d2 f 1 d2 f
= · + 2 + 2 · + 2· 2 + 2 2 · 2
r dr  dr r sin s ds r  ds r sin s dt
d2 f

1 d df 1 d df 1
= 2· r2 · + 2 · sin s · + 2 2 · 2
r dr dr r sin s ds ds r sin s dt
Thus, we are led to the formula in the statement. □
Still with me, I hope, and do not worry, one day you will have such computations
for breakfast. We can now reformulate the Schrödinger equation in spherical coordinates,
and separate the variables, which leads to a radial and angular equation, as follows:
Theorem 16.15. The time-independent Schrödinger equation in spherical coordinates
separates, for solutions of type ϕ = ρ(r)α(s, t), into two equations, as follows,
2mr2
 
d 2 dρ
r · − (V − E)ρ = Kρ
dr dr h2
d2 α
 
d dα
sin s · sin s · + 2 = −K sin2 s · α
ds ds dt
with K being a constant, called radial equation, and angular equation.
Proof. By using the formula in Theorem 16.14, the time-independent Schrödinger
equation reformulates in spherical coordinates as follows:
h2 1 d d2 ϕ
     
2 dϕ 1 d dϕ 1
(V − E)ϕ = · r · + 2 · sin s · + 2 2 · 2
2m r2 dr dr r sin s ds ds r sin s dt
Let us look now for separable solutions for this latter equation, consisting of a radial
part and an angular part, as in the statement, namely:
ϕ(r, s, t) = ρ(r)α(s, t)
 16C. SPHERICAL COORDINATES 381

By plugging this function into our equation, we obtain:

h2 α d d2 α
     
2 dρ ρ d dα ρ
(V − E)ρα = · r · + 2 · sin s · + 2 2 · 2
2m r2 dr dr r sin s ds ds r sin s dt
In order to solve this equation, we will do two manipulations. First, by multiplying
everything by 2mr2 /(h2 ρα), this equation takes the following more convenient form:
2mr2 d2 α
   
1 d 2 dρ 1 d dα 1
(V − E) = · r · + · sin s · + ·
h2 ρ dr dr α sin s ds ds α sin2 s dt2
Now observe that by moving the radial terms to the left, and the angular terms to the
right, this latter equation can be written as follows:
2mr2 d2 α
     
1 d 2 dρ 1 d dα
(V − E) − · r · = sin s · sin s · + 2
h2 ρ dr dr α sin2 s ds ds dt
Since this latter equation is now separated between radial and angular variables, both
sides must be equal to a certain constant −K, as follows:
2mr2
 
1 d 2 dρ
(V − E) − · r · = −K
h2 ρ dr dr
d2 α
   
1 d dα
sin s · sin s · + 2 = −K
α sin2 s ds ds dt
But this leads to the conclusion in the statement. □
Let us first study the angular equation. We first have the following result:
Proposition 16.16. The angular equation that we found before, namely
d2 α
 
d dα
sin s · sin s · + 2 = −K sin2 s · α
ds ds dt
separates, for solutions of type α = σ(s)θ(t), into two equations, as follows,
1 d2 θ
· = −m2
θ dt2
 
sin s d dσ
· sin s · + K sin2 s = m2
σ ds ds
with m being a constant, called azimuthal equation, and polar equation.
Proof. This is something elementary, the idea being as follows:
(1) Let us first recall that r ∈ [0, ∞) is the radius, s ∈ [0, π] is the polar angle, and
t ∈ [0, 2π] is the azimuthal angle. Be said in passing, there are several conventions and
notations here, and the above ones, that we use here, come from the general ones in N
dimensions, because further coordinates can be easily added, in the obvious way.
382 16. INFINITE DIMENSIONS

(2) Getting back now to our question, by plugging α = σ(s)θ(t) into the angular
equation, we obtain:
d2 θ
 
d dσ
sin s · θ · sin s · + σ · 2 = −K sin2 s · σθ
ds ds dt
By dividing everything by σθ, this equation can be written as follows:
1 d2 θ
 
sin s d dσ
− · 2 = · sin s · + K sin2 s
θ dt σ ds ds
Since the variables are separated, we must have, for a certain constant m:
1 d2 θ
· 2 = −m2
θ dt
 
sin s d dσ
· sin s · + K sin2 s = m2
σ ds ds
Thus, we are led to the conclusion in the statement. □
Regarding the azimuthal equation, things here are quickly settled, as follows:
Proposition 16.17. The solutions of the azimuthal equation, namely
1 d2 θ
· 2 = −m2
θ dt
are the functions as follows, with a, b ∈ C being parameters,
θ(t) = aeimt + be−imt
and with only the case m ∈ Z being acceptable, on physical grounds.
Proof. The first assertion is clear, because we have a second order equation, and
two obvious solutions for it, e±imt , and then their linear combinations, and that’s all.
Regarding the last assertion, the point here is that by using θ(t) = θ(t + 2π), which is a
natural physical assumption on the wave function, we are led to m ∈ Z, as stated. □
We are now about to solve the angular equation, with only the polar equation remain-
ing to be studied. However, in practice, this polar equation is 10 times more difficult that
everything what we did so far, so be patient. We first have:
Proposition 16.18. The polar equation that we found before, namely
 
sin s d dσ
· sin s · + K sin2 s = m2
σ ds ds
with m ∈ Z, translates via σ(s) = f (cos s) into the following equation,
m2
 
2 ′′ ′
(1 − x )f (x) − 2xf (x) = − K f (x)
1 − x2
where x = cos s, called Legendre equation.
 16C. SPHERICAL COORDINATES 383

Proof. Let us first do a number of manipulations on our equation, before making

the change of variables. By multiplying by σ, our equation becomes:
 
d dσ
= m2 − K sin2 s σ


sin s · sin s ·
ds ds
By differentiating at left, this equation becomes:
sin s (cos s · σ ′ + sin s · σ ′′ ) = m2 − K sin2 s σ

Finally, by dividing everything by sin2 s, our equation becomes:

 2 
′′ cos s ′ m
σ + ·σ = −K σ
sin s sin2 s
Now let us set σ(s) = f (cos s). With this change of variables, we have:
σ = f (cos s)
σ ′ = − sin s · f ′ (cos s)
σ ′′ = − cos s · f ′ (cos s) + sin2 s · f ′′ (cos s)
By plugging this data, our radial equation becomes:
 2 
2 ′′ ′ m
sin s · f (cos s) − 2 cos s · f (cos s) = − K f (cos s)
sin2 s
Now with x = cos s, which is our new variable, this equation reads:
m2
 
2 ′′ ′
(1 − x )f (x) − 2xf (x) = − K f (x)
1 − x2
But this is the Legendre equation, as stated. □
Here comes now the difficult point. We have the following non-trivial result:
Theorem 16.19. The solutions of the Legendre equation, namely
m2
 
2 ′′ ′
(1 − x )f (x) − 2xf (x) = − K f (x)
1 − x2
can be explicitely computed, via complicated math, and only the case
K = l(l + 1) : l∈N
is acceptable, on physical grounds.
Proof. The first part is something quite complicated, involving the hypergeometric
functions 2 F1 , that you don’t want to hear about, believe me. As for the second part,
analysis and physics, this is something not trivial either. See Griffiths [42]. □
In order to construct the solutions, we will need:
384 16. INFINITE DIMENSIONS

Theorem 16.20. The orthonormal basis of L2 [−1, 1] obtained by starting with the
Weierstrass basis {xl }, and doing Gram-Schmidt, is the family of polynomials {Pl }, with
each Pl being of degree l, and with positive leading coefficient, subject to:
Z 1
Pk (x)Pl (x) dx = δkl
−1

These polynomials, called Legendre polynomials, satisfy the equation

(1 − x2 )Pl′′ (x) − 2xPl′ (x) + l(l + 1)Pl (x) = 0
which is the Legendre equation at m = 0, and with K = l(l + 1). Moreover,
 l
1 d
Pl (x) = l (x2 − 1)l
2 l! dx
which is called the Rodrigues formula for Legendre polynomials.
Proof. As a first observation, we are not lost somewhere in abstract math, because
of the occurrence of the Legendre equation. As for the proof, this goes as follows:
(1) The first assertion is clear, because the Gram-Schmidt procedure applied to the
Weierstrass basis {xl } can only lead to a certain family of polynomials {Pl }, with each Pl
being of degree l, and also unique, if we assume that it has positive leading coefficient,
with this ± choice being needed, as usual, at each step of Gram-Schmidt.
(2) In order to have now an idea about these beasts, here are the first few of them,
which can be obtained say via a straightforward application of Gram-Schmidt:
P0 = 1
P1 = x
P2 = (3x2 − 1)/2
P3 = (5x3 − 3x)/2
P4 = (35x4 − 30x2 + 3)/8
P5 = (63x5 − 70x3 + 15x)/8
(3) Now thinking about what Gram-Schmidt does, this is certainly something by
recurrence. And examining the recurrence leads to the Legendre equation, as stated.
(4) As for the Rodrigues formula, by uniqueness no need to try to understand where
this formula comes from, and we have two choices here, either by verifying that {Pl } is
orthonormal, or by verifying the Legendre equation. And both methods work. □

Going ahead now, we can solve in fact the Legendre equation at any m, as follows:
 16C. SPHERICAL COORDINATES 385

Proposition 16.21. The general Legendre equation, with parameters m ∈ N and

K = l(l + 1) with l ∈ N, namely
m2
 
2 ′′ ′
(1 − x )f (x) − 2xf (x) = − l(l + 1) f (x)
1 − x2
is solved by the following functions, called Legendre functions,
 m
m m 2 m/2 d
Pl (x) = (−1) (1 − x ) Pl (x)
dx
where Pl are as before the Legendre polynomials. Also, we have
2 m/2
 l+m
m m (1 − x ) d
Pl (x) = (−1) l
(x2 − 1)l
2 l! dx
called Rodrigues formula for Legendre functions.
Proof. The first assertion is something elementary, coming by differentiating m times
the Legendre equation, which leads to the general Legendre equation. As for the second
assertion, this follows from the Rodrigues formula for Legendre polynomials. □
And this is the end of our study. Eventually. By putting together all the above results,
we are led to the following conclusion:
Theorem 16.22. The separated solutions α = σ(s)θ(t) of the angular equation,
d2 α
 
d dα
sin s · sin s · + 2 = −K sin2 s · α
ds ds dt
are given by the following formulae, where l ∈ N is such that K = l(l + 1),
σ(s) = Plm (cos s) , θ(t) = eimt
and where m ∈ Z is a constant, and with Plm being the Legendre function,
 m
m m 2 m/2 d
Pl (x) = (−1) (1 − x ) Pl (x)
dx
where Pl are the Legendre polynomials, given by the following formula:
 l
1 d
Pl (x) = l (x2 − 1)l
2 l! dx
These solutions α = σ(s)θ(t) are called spherical harmonics.
Proof. This follows indeed from all the above, and with the comment that everything
is taken up to linear combinations. We will normalize the wave function later. □
386 16. INFINITE DIMENSIONS

16d. The hydrogen atom

Hydrogen, eventually. In order now to finish our study, and eventually get to conclu-
sions about hydrogen, it remains to solve the radial equation, for the Coulomb potential
V of the proton. Let us begin with some generalities, valid for any time-independent
potential V . As a first manipulation on the radial equation, we have:
Proposition 16.23. The radial equation, written with K = l(l + 1),
2mr2
(r2 ρ′ )′ − (V − E)ρ = l(l + 1)ρ
h2
takes with ρ = u/r the following form, called modified radial equation,
h2 h2 l(l + 1)
 
′′
Eu = − ·u + V + u
2m 2mr2
which is a time-independent 1D Schrödinger equation.
Proof. With ρ = u/r as in the statement, we have:
u u′ r − u
ρ= , ρ′ = , (r2 ρ′ )′ = u′′ r
r r2
By plugging this data into the radial equation, this becomes:
2mr l(l + 1)
u′′ r − (V − E)u = ·u
h2 r
By multiplying everything by h2 /(2mr), this latter equation becomes:
h2 h2 l(l + 1)
· u′′ − (V − E)u = ·u
2m 2mr2
But this gives the formula in the statement. As for the interpretation, as time-
independent 1D Schrödinger equation, this is clear as well, and with the comment here
that the term added to the potential V is some sort of centrifugal term. □
Getting back now to the Coulomb potential of the proton, we have here:
Fact 16.24. The Coulomb potential of the hydrogen atom proton, acting on the elec-
tron by attraction, is given according to the Coulomb law by
Kep
V =−
r
where p is the charge of the proton, and K is the Coulomb constant. In practice however
we have p ≃ e up to order 10−7 , and so our formula can be written as
Ke2
V ≃−
r
and we will use this latter formula, and with = sign, for simplifying.
 16D. THE HYDROGEN ATOM 387

Getting back now to math, it remains to solve the modified radial equation, for the
above potential V . And we have here the following result, which does not exactly solve
this radial equation, but provides us instead with something far better, namely the proof
of the original claim by Bohr, which was at the origin of everything:
Theorem 16.25 (Schrödinger). In the case of the hydrogen atom, where V is the
Coulomb potential of the proton, the modified radial equation, which reads
h2 Ke2 h2 l(l + 1)
 
′′
Eu = − ·u + − + u
2m r 2mr2
leads to the Bohr formula for allowed energies,
2
m Ke2

1
En = − · 2
2 h n
with n ∈ N, the binding energy being
E1 ≃ −2.177 × 10−18
with means E1 ≃ −13.591 eV.
Proof. This is again something non-trivial, and we will be following Griffiths [42],
with some details missing. The idea is as follows:
(1) By dividing our modified radial equation by E, this becomes:
h2 Ke2 h2 l(l + 1)
 
′′
− ·u = 1+ − u
2mE Er 2mEr2
√
In terms of α = −2mE/h, this equation takes the following form:
u′′ Ke2 l(l + 1)
 
= 1+ + u
α2 Er (αr)2
In terms of the new variable p = αr, this latter equation reads:
αKe2 l(l + 1)
 
′′
u = 1+ + u
Ep p2
Now let us introduce a new constant S for our problem, as follows:
αKe2
S=−
E
In terms of this new constant, our equation reads:
 
′′ S l(l + 1)
u = 1− + u
p p2
(2) The idea will be that of looking for a solution written as a power series, but before
that, we must “peel off” the asymptotic behavior. Which is something that can be done,
388 16. INFINITE DIMENSIONS

of course, heuristically. With p → ∞ we are led to u′′ = u, and ignoring the solution
u = ep which blows up, our approximate asymptotic solution is:
u ∼ e−p
Similarly, with p → 0 we are led to u′′ = l(l +1)u/p2 , and ignoring the solution u = p−l
which blows up, our approximate asymptotic solution is:
u ∼ pl+1
(3) The above heuristic considerations suggest writing our function u as follows:
u = pl+1 e−p v
So, let us do this. In terms of v, we have the following formula:
u′ = pl e−p [(l + 1 − p)v + pv ′ ]
Differentiating a second time gives the following formula:
  
′′ l −p l(l + 1) ′ ′′
u =pe − 2l − 2 + p v + 2(l + 1 − p)v + pv
p
Thus the radial equation, as modified in (1) above, reads:
pv ′′ + 2(l + 1 − p)v ′ + (S − 2(l + 1))v = 0
(4) We will be looking for a solution v appearing as a power series:
X∞
v= cj p j
j=0

But our equation leads to the following recurrence formula for the coefficients:
2(j + l + 1) − S
cj+1 = · cj
(j + 1)(j + 2l + 2)
(5) We are in principle done, but we still must check that, with this choice for the
coefficients cj , our solution v, or rather our solution u, does not blow up. And the whole
point is here. Indeed, at j >> 0 our recurrence formula reads, approximately:
2cj
cj+1 ≃
j
But, surprisingly, this leads to v ≃ c0 e2p , and so to u ≃ c0 pl+1 ep , which blows up.
(6) As a conclusion, the only possibility for u not to blow up is that where the series
defining v terminates at some point. Thus, we must have for a certain index j:
2(j + l + 1) = S
In other words, we must have, for a certain integer n > l:
S = 2n
 16D. THE HYDROGEN ATOM 389

(7) We are almost there. Recall from (1) above that S was defined as follows:
√
αKe2 −2mE
S=− : α=
E h
Thus, we have the following formula for the square of S:
α2 K 2 e4 2mE K 2 e4 2mK 2 e4
S2 = = − · = −
E2 h2 E2 h2 E
Now by using the formula S = 2n from (6), the energy E must be of the form:
2mK 2 e4 mK 2 e4
E=− = −
h2 S 2 2h2 n2
Calling this energy En , depending on n ∈ N, we have, as claimed:
2
m Ke2

1
En = − · 2
2 h n
(8) Thus, we proved the Bohr formula. Regarding numerics, the data is as follows:
K = 8.988 × 109 , e = 1.602 × 10−19
h = 1.055 × 10−34 , m = 9.109 × 10−31
But this gives the formula of E1 in the statement. □
As a first remark, all this agrees with the Rydberg formula, due to:
Theorem 16.26. The Rydberg constant for hydrogen is given by
E1
R=−
h0 c
where E1 is the Bohr binding energy, and the Rydberg formula itself, namely
 
1 1 1
=R −
λn1 n2 n21 n22
simply reads, via the energy formula in Theorem 16.25,
1 En2 − En1
=
λn1 n2 h0 c
which is in agreement with the Planck formula E = h0 c/λ.
Proof. Here the first assertion is something numeric, coming from the fact that the
formula in the statement gives, when evaluated, the Rydberg constant:
−E1 2.177 × 10−18
R= = −34 8
= 1.096 × 107
h0 c 6.626 × 10 × 2.998 × 10
390 16. INFINITE DIMENSIONS

As a consequence, and passed now what the experiments exactly say, we can define
the Rydberg constant of hydrogen abstractly, by the following formula:
 2 2
m Ke
R=
2h0 c h
Regarding now the second assertion, by dividing R = −E1 /(h0 c) by any number of
type n2 we obtain, according to the energy convention in Theorem 16.25:
R En
2
=−
n h0 c
But these are exactly the numbers which are subject to substraction in the Rydberg
formula, and so we are led to the conclusion in the statement. □
Let us go back now to our study of the Schrödinger equation. Our conclusions are:
Theorem 16.27. The wave functions of the hydrogen atom are the following functions,
labelled by three quantum numbers, n, l, m,
ϕnlm (r, s, t) = ρnl (r)αlm (s, t)
where ρnl (r) = pl+1 e−p v(p)/r with p = αr as before, with the coefficients of v subject to
2(j + l + 1 − n)
cj+1 = · cj
(j + 1)(j + 2l + 2)
and αlm (s, t) being the spherical harmonics found before.
Proof. This follows indeed by putting together all the results obtained so far, and
with the remark that everything is up to the normalization of the wave function. □
In what regards the main wave function, that of the ground state, we have:
Theorem 16.28. With the hydrogen atom in its ground state, the wave function is
1
ϕ100 (r, s, t) = √ e−r/a
πa3

where a = 1/α is the inverse of the parameter appearing in our computations above,
√
−2mE
α=
h
called Bohr radius of the hydrogen atom. This Bohr radius is the mean distance between
the electron and the proton, in the ground state, and is given by the formula
h2
a=
mKe2
which numerically means a ≃ 5.291 × 10−11 .
 16D. THE HYDROGEN ATOM 391

Proof. There are several things going on here, as follows:

(1) According to the various formulae in the proof of Theorem 16.25, taken at n = 1,
the parameter α appearing in the computations there is given by:
√
−2mE 1 Ke2 mKe2
α= = ·m· =
h h h h2
Thus, the inverse α = 1/a is indeed given by the formula in the statement.
(2) Regarding the wave function, according to Theorem 16.27 this consists of:
2e−r/a 1
ρ10 (r) = √ , α00 (s, t) = √
a 3 2 π
By making the product, we obtain the formula of ϕ100 in the statement.
(3) But this formula of ϕ100 shows in particular that the Bohr radius a is indeed the
mean distance between the electron and the proton, in the ground state.
(4) Finally, in what regards the numerics, these are as follows:
1.0552 × 10−68
a= = 5.297 × 10−11
9.109 × 10−31 × 8.988 × 109 × 1.6022 × 10−38
Thus, we are led to the conclusions in the statement. □
Getting back now to the general setting of Theorem 16.25, the point is that the
polynomials v(p) appearing there are well-known objects in mathematics, as follows:
Proposition 16.29. The polynomials v(p) are given by the formula
v(p) = L2l+1
n−l−1 (p)

where the polynomials on the right, called associated Laguerre polynomials, are given by
 p
p p d
Lq (x) = (−1) Lp+q (x)
dx
with Lp+q being the Laguerre polynomials, given by the following formula:
 q
ex d
Lq (x) = (e−x xq )
q! dx
Proof. The story here is very similar to that of the Legendre polynomials. Consider
the Hilbert space H = L2 [0, ∞), with the following scalar product on it:
Z ∞
< f, g >= f (x)g(x)e−x dx
0

(1) The orthogonal basis obtained by applying Gram-Schmidt to the Weierstrass basis
{xq } is then the basis formed by the Laguerre polynomials {Lq }.
392 16. INFINITE DIMENSIONS

(2) We have the explicit formula for Lq in the statement, which is analogous to the
Rodrigues formula for the Legendre polynomials.
(3) The first assertion follows from the fact that the coefficients of the associated
Laguerre polynomials satisfy the equation for the coefficients of v(p).
(4) Alternatively, the first assertion follows as well by using an equation for the La-
guerre polynomials, which is very similar to the Legendre equation. □
With the above result in hand, we can now improve Theorem 16.25, as follows:
Theorem 16.30. The wave functions of the hydrogen atom are given by
s 
3  l  
2 (n − l − 1)! −r/na 2r 2l+1 2r
ϕnlm (r, s, t) = e Ln−l−1 αlm (s, t)
na 2n(n + l)! na na
with αlm (s, t) being the spherical harmonics found before.
Proof. This follows indeed by putting together what we have, namely Theorem 16.25
and Proposition 16.29, and then doing some remaining work, concerning the normalization
of the wave function, which leads to the normalization factor appearing above. □
And good news, that is all. The above formula is all you need, in everyday life.
16e. Exercises
Congratulations for having read this book, and no exercises for this final chapter.
However, if you enjoyed this book, and looking for more to read, have a look at the
various books referenced below. Normally these are all good books, which all sort of go to
the point, without bothering much with annoying details, and normally you should like
them too. So, have a look at them, and start with the one that you like the most.
Before everything, however, learn more calculus, and especially do many exercises, as
many as needed, first in order to be at ease with calculus, and then in order to really
love calculus. In fact, and in the hope that you got it, the present book was just an
introduction to calculus, not calculus itself. Just a beginning.
 Bibliography

[1] V.I. Arnold, Ordinary differential equations, Springer (1973).

[2] V.I. Arnold, Mathematical methods of classical mechanics, Springer (1974).
[3] V.I. Arnold, Catastrophe theory, Springer (1984).
[4] V.I. Arnold, Lectures on partial differential equations, Springer (1997).
[5] V.I. Arnold and B.A. Khesin, Topological methods in hydrodynamics, Springer (1998).
[6] M.F. Atiyah, K-theory, CRC Press (1964).
[7] M.F. Atiyah, The geometry and physics of knots, Cambridge Univ. Press (1990).
[8] M.F. Atiyah and I.G. MacDonald, Introduction to commutative algebra, Addison-Wesley (1969).
[9] T. Banica, Linear algebra and group theory (2024).
[10] T. Banica, Principles of operator algebras (2024).
[11] T. Banica, Introduction to modern physics (2024).
[12] R.J. Baxter, Exactly solved models in statistical mechanics, Academic Press (1982).
[13] S.J. Blundell and K.M. Blundell, Concepts in thermal physics, Oxford Univ. Press (2006).
[14] B. Bollobás, Modern graph theory, Springer (1998).
[15] S.M. Carroll, Spacetime and geometry, Cambridge Univ. Press (2004).
[16] A.R. Choudhuri, Astrophysics for physicists, Cambridge Univ. Press (2012).
[17] D.D. Clayton, Principles of stellar evolution and nucleosynthesis, Univ. of Chicago Press (1968).
[18] A. Connes, Noncommutative geometry, Academic Press (1994).
[19] J.B. Conway, A course in functional analysis, Springer (1985).
[20] W.N. Cottingham and D.A. Greenwood, An introduction to the standard model of particle physics,
Cambridge Univ. Press (2012).
[21] P.A. Davidson, Introduction to magnetohydrodynamics, Cambridge Univ. Press (2001).
[22] P.A.M. Dirac, Principles of quantum mechanics, Oxford Univ. Press (1930).
[23] M.P. do Carmo, Differential geometry of curves and surfaces, Dover (1976).
[24] M.P. do Carmo, Riemannian geometry, Birkhäuser (1992).

393
394 BIBLIOGRAPHY

[25] S. Dodelson, Modern cosmology, Academic Press (2003).

[26] R. Durrett, Probability: theory and examples, Cambridge Univ. Press (1990).
[27] A. Einstein, Relativity: the special and the general theory, Dover (1916).
[28] L.C. Evans, Partial differential equations, AMS (1998).
[29] W. Feller, An introduction to probability theory and its applications, Wiley (1950).
[30] E. Fermi, Thermodynamics, Dover (1937).
[31] R.P. Feynman, R.B. Leighton and M. Sands, The Feynman lectures on physics I: mainly mechanics,
radiation and heat, Caltech (1963).
[32] R.P. Feynman, R.B. Leighton and M. Sands, The Feynman lectures on physics II: mainly electro-
magnetism and matter, Caltech (1964).
[33] R.P. Feynman, R.B. Leighton and M. Sands, The Feynman lectures on physics III: quantum me-
chanics, Caltech (1966).
[34] R.P. Feynman and A.R. Hibbs, Quantum mechanics and path integrals, Dover (1965).
[35] P. Flajolet and R. Sedgewick, Analytic combinatorics, Cambridge Univ. Press (2009).
[36] A.P. French, Special relativity, Taylor and Francis (1968).
[37] W. Fulton, Algebraic topology, Springer (1995).
[38] W. Fulton and J. Harris, Representation theory, Springer (1991).
[39] C. Godsil and G. Royle, Algebraic graph theory, Springer (2001).
[40] H. Goldstein, C. Safko and J. Poole, Classical mechanics, Addison-Wesley (1980).
[41] D.J. Griffiths, Introduction to electrodynamics, Cambridge Univ. Press (2017).
[42] D.J. Griffiths and D.F. Schroeter, Introduction to quantum mechanics, Cambridge Univ. Press
(2018).
[43] D.J. Griffiths, Introduction to elementary particles, Wiley (2020).
[44] D.J. Griffiths, Revolutions in twentieth-century physics, Cambridge Univ. Press (2012).
[45] J. Harris, Algebraic geometry, Springer (1992).
[46] R.A. Horn and C.R. Johnson, Matrix analysis, Cambridge Univ. Press (1985).
[47] K. Huang, Introduction to statistical physics, CRC Press (2001).
[48] K. Huang, Quantum field theory, Wiley (1998).
[49] K. Huang, Quarks, leptons and gauge fields, World Scientific (1982).
[50] K. Huang, Fundamental forces of nature, World Scientific (2007).
[51] J.E. Humphreys, Introduction to Lie algebras and representation theory, Springer (1972).
 BIBLIOGRAPHY 395

[52] V.F.R. Jones, Subfactors and knots, AMS (1991).

[53] T. Kibble and F.H. Berkshire, Classical mechanics, Imperial College Press (1966).
[54] C. Kittel, Introduction to solid state physics, Wiley (1953).
[55] M. Kumar, Quantum: Einstein, Bohr, and the great debate about the nature of reality, Norton
(2009).
[56] T. Lancaster and K.M. Blundell, Quantum field theory for the gifted amateur, Oxford Univ. Press
(2014).
[57] L.D. Landau and E.M. Lifshitz, Mechanics, Pergamon Press (1960).
[58] L.D. Landau and E.M. Lifshitz, The classical theory of fields, Addison-Wesley (1951).
[59] L.D. Landau and E.M. Lifshitz, Quantum mechanics: non-relativistic theory, Pergamon Press (1959).
[60] S. Lang, Algebra, Addison-Wesley (1993).
[61] P. Lax, Linear algebra and its applications, Wiley (2007).
[62] P. Lax, Functional analysis, Wiley (2002).
[63] P. Lax and M.S. Terrell, Calculus with applications, Springer (2013).
[64] P. Lax and M.S. Terrell, Multivariable calculus with applications, Springer (2018).
[65] J.M. Lee, Introduction to topological manifolds, Springer (2011).
[66] J.M. Lee, Introduction to smooth manifolds, Springer (2012).
[67] J.M. Lee, Introduction to Riemannian manifolds, Springer (2018).
[68] M.L. Mehta, Random matrices, Elsevier (2004).
[69] M.A. Nielsen and I.L. Chuang, Quantum computation and quantum information, Cambridge Univ.
Press (2000).
[70] R.K. Pathria and and P.D. Beale, Statistical mechanics, Elsevier (1972).
[71] P. Petersen, Linear algebra, Springer (2012).
[72] P. Petersen, Riemannian geometry, Springer (1998).
[73] W. Rudin, Principles of mathematical analysis, McGraw-Hill (1964).
[74] W. Rudin, Real and complex analysis, McGraw-Hill (1966).
[75] W. Rudin, Functional analysis, McGraw-Hill (1973).
[76] W. Rudin, Fourier analysis on groups, Dover (1974).
[77] B. Ryden, Introduction to cosmology, Cambridge Univ. Press (2002).
[78] B. Ryden and B.M. Peterson, Foundations of astrophysics, Cambridge Univ. Press (2010).
[79] B. Ryden and R.W. Pogge, Interstellar and intergalactic medium, Cambridge Univ. Press (2021).
396 BIBLIOGRAPHY

[80] D.V. Schroeder, An introduction to thermal physics, Oxford Univ. Press (1999).
[81] J.P. Serre, A course in arithmetic, Springer (1973).
[82] J.P. Serre, Linear representations of finite groups, Springer (1977).
[83] I.R. Shafarevich, Basic algebraic geometry, Springer (1974).
[84] R. Shankar, Fundamentals of physics I: mechanics, relativity, and thermodynamics, Yale Univ. Press
(2014).
[85] R. Shankar, Fundamentals of physics II: electromagnetism, optics, and quantum mechanics, Yale
Univ. Press (2016).
[86] R. Shankar, Principles of quantum mechanics, Springer (1980).
[87] R. Shankar, Quantum field theory and condensed matter: an introduction, Cambridge Univ. Press
(2017).
[88] A.M. Steane, Thermodynamics, Oxford Univ. Press (2016).
[89] J.R. Taylor, Classical mechanics, Univ. Science Books (2003).
[90] J. von Neumann, Mathematical foundations of quantum mechanics, Princeton Univ. Press (1955).
[91] J. von Neumann and O. Morgenstern, Theory of games and economic behavior, Princeton Univ.
Press (1944).
[92] J. Watrous, The theory of quantum information, Cambridge Univ. Press (2018).
[93] S. Weinberg, Foundations of modern physics, Cambridge Univ. Press (2011).
[94] S. Weinberg, Lectures on quantum mechanics, Cambridge Univ. Press (2012).
[95] S. Weinberg, Lectures on astrophysics, Cambridge Univ. Press (2019).
[96] S. Weinberg, Cosmology, Oxford Univ. Press (2008).
[97] H. Weyl, The theory of groups and quantum mechanics, Princeton Univ. Press (1931).
[98] H. Weyl, The classical groups: their invariants and representations, Princeton Univ. Press (1939).
[99] H. Weyl, Space, time, matter, Princeton Univ. Press (1918).
[100] B. Zwiebach, A first course in string theory, Cambridge Univ. Press (2004).
 Index

abelian group, 166 central binomial coefficients, 49, 75, 308

absolute value, 289 chain rule, 61, 96, 245
adjoint map, 205 change of variable, 96, 300, 301
adjoint matrix, 205 characteristic polynomial, 221, 222
adjoint operator, 374 charge, 359
algebraic curve, 249 chart, 268
algebraic manifold, 261 closed and bounded, 41, 235
alternating series, 27, 29 closed set, 39, 40
analytic function, 140, 145 colored integers, 333
angular equation, 380 colored moments, 333
arctan, 63 column expansion, 216
area, 81 common roots, 122
area of sphere, 312 compact set, 41, 235
argument of complex number, 110 compact support, 156
atom, 200 complete space, 25, 129
average of function, 84 Complex CLT, 333
azimuthal equation, 381 complex conjugate, 113, 138, 184
complex eigenvalues, 223
barycenter, 127
complex function, 116, 137
Bell numbers, 327
complex matrix, 219
Bernoulli laws, 176, 327
complex number, 107, 109
binomial coefficient, 11
complex plane, 129
binomial coefficients, 308
binomial formula, 13, 48, 74 complex roots, 121
Bohr formula, 387 concave function, 70
Bohr radius, 390 conic, 249, 255
boundary of domain, 141 connected set, 41, 136, 236
bounded operator, 372 continuous function, 33, 39, 116, 237
bounded sequence, 23 continuously differentiable, 241
convergent sequence, 22, 129
cartography, 271 convergent series, 25
Catalan numbers, 49, 75 convex function, 70
Cauchy formula, 143–145 convolution, 155, 173, 322
Cauchy sequence, 25, 129 convolution exponential, 174, 326
Cauchy-Schwarz, 71 convolution semigroup, 326
CCLT, 333 coordinates, 268
397
398 INDEX

cos, 21, 35, 59, 76, 133 Fourier inversion, 165

cosh, 135 Fourier transform, 159, 173, 323, 326
cover, 41 fraction, 62
curl, 365 fractions, 11
cyclic group, 168 free group, 267
function, 33
decimal writing, 18 function space, 155
decreasing sequence, 23 fundamental theorem of calculus, 92, 93
Dedekind cut, 16
degree 2 equation, 17, 108, 113, 249 Gauss formula, 366
density, 223 Gauss law, 364
derivative, 57, 58, 238 generalized binomial formula, 48, 74
determinant, 211 generalized binomial numbers, 48
determinant formula, 219 geometric series, 26, 117, 132
determinant of products, 214 gradient, 365
deterrminant, 219 Gram-Schmidt, 372
diagonalizable matrix, 214, 223, 287 gravity, 77, 255
diagonalization, 220, 222 Green formula, 366
differentiable function, 57, 137, 238 ground state, 390
differential equation, 77
differential manifold, 268 Hölder inequality, 153
discriminant, 123, 223 harmonic function, 183, 290
distance, 47, 129, 225 heat equation, 79, 181
distribution, 321 Hessian, 278
divergence, 365 higher derivative, 96, 275
double derivatives, 177 higher derivatives, 72
double factorial, 305 Hilbert space, 370
double factorials, 304, 309, 315 hole, 267
double root, 123 holomorphic function, 137, 145
homogeneous polynomial, 184
e, 29, 53 homotopy group, 267
eigenvalue, 220, 222 Hooke law, 78, 179
eigenvector, 220, 222 hyperbola, 250
electric field, 359
electron, 376 i, 107
ellipsis, 250 increasing sequence, 23
equal almost everywhere, 155 independence, 172, 173, 322, 323
exchange of hat, 161 infinite matrix, 372
exp, 59, 76, 133 infinitely differentiable, 138, 145
expectation, 86 infinitesimal, 92
exponential, 53, 60, 116 inflation coefficient, 213
integrable function, 82
factorials, 11 integral, 81
Fano plane, 266 integral over curve, 142
field, 130 integration by parts, 96
field lines, 360 intermediate value, 42, 136
finite field, 265 inverse function, 43
flux, 361, 362, 364 inversion formula, 207
 INDEX 399

invertible matrix, 206, 208, 212 modulus of matrix, 289

moment, 102
Jacobian, 243, 301, 362 moments, 321
Jensen inequality, 70 monotone function, 42
Monte Carlo integration, 84
Kepler laws, 255 multiplication of complex numbers, 120
Klein bottle, 262
knot, 268 Nash theorem, 272
Netwon law, 255
L’Hôpital’s rule, 68, 72 Newton law, 78, 179
Lagrange multipliers, 295 normal matrix, 287
Laplace equation, 189 normed space, 155
Laplace operator, 179, 181, 183, 290, 377 number of inversions, 217
lattice model, 78, 79, 179
law, 321 observable, 376
Legendre function, 385 open set, 39, 40
lim inf, 25 oriented system of vectors, 210
lim sup, 25 orthonormal basis, 372
limit of continuous functions, 46
limit of sequence, 22 p-norm, 155
limit of series, 25 parabola, 250
linear function, 203 parallelogram rule, 109
linear map, 203 partial derivatives, 239
linear operator, 372 partial isometry, 289
Liouville theorem, 146, 188, 290 partitions, 327
local maximum, 64, 69, 278 Pascal triangle, 14
local minimum, 64, 69, 278 passage matrix, 220
locally affine, 58 permutation, 217
log, 59, 76, 133 pi, 20
loop, 267 piecewise continuous, 87
lower triangular matrix, 215 piecewise linear, 82
piecewise monotone, 87
Möbius strip, 262 Planck constant, 376
main value formula, 141, 146 Planck formula, 389
matching pairings, 335 PLT, 176, 327
matrix, 203 point charge, 362
maximum, 42, 64, 278 pointwise convergence, 45, 46, 130
maximum modulus, 188, 290 Poisson law, 325
maximum principle, 141, 146 Poisson Limit Theorem, 176, 327
Maxwell equations, 191 polar coordinates, 110, 120, 269, 302
mean, 102 polar decomposition, 289
mean value formula, 188, 290 polar equation, 381
mean value property, 89 polar writing, 119
metric space, 225 polarization formula, 205
minimum, 42, 64, 278 pole, 130
Minkowski inequality, 153 polynomial, 73, 130
modulus, 57, 138 positive eigenvalues, 284
modulus of complex number, 110 positive matrix, 278, 284
400 INDEX

power function, 43, 58, 136 signed volume, 211

powers of complex number, 111 sin, 21, 35, 59, 76, 133
probability 0, 19 single roots, 123
probability density, 101 sinh, 135
probability space, 321 smooth manifold, 268
product of cyclic groups, 168 spectral theorem, 287
product of eigenvalues, 214 spherical coordinates, 269, 303, 377
projection, 204, 205, 283 spherical harmonics, 385, 390, 392
projective space, 262 spherical integral, 315, 317, 319
Pythagoras’ theorem, 21 spiral, 117, 132
square root, 16, 17, 43, 49, 75, 108, 113, 284
radial equation, 380, 386 step function, 45
radial function, 189 stereographic projection, 270
radial limit, 139 Stirling formula, 306
random number, 84 Stokes formula, 366
random variable, 86, 321 strictly positive matrix, 284
rank 1 projection, 264 subcover, 41
rational function, 130, 138 subsequence, 23, 25
real number, 16 sum of vectors, 109
rectangular matrix, 203 support, 156
reflection, 113 symmetric group, 217
remainder, 96 symmetry, 204
resultant, 122, 223
Riemann integration, 83 tan, 63
Riemann series, 26 Taylor formula, 68, 72, 73, 76, 96
Riemann sum, 89, 306 time-independent equation, 380
Riemann-Lebesgue property, 164 totally discontinuous, 45
right-hand rule, 346 transposition, 217
roots of polynomial, 43, 121, 130 trigonometric integral, 304, 313
roots of unity, 126, 127
rotation, 204, 223 uniform convergence, 46, 130
row expansion, 215 union of intervals, 40
Rydberg formula, 389 unitary, 205, 285
unoriented system of vectors, 210
Sarrus formula, 216, 219 upper triangular matrix, 215
scalar product, 205, 370
variance, 102
Schrödinger equation, 376
vector, 109
second derivative, 68
vector product, 346
second order derivative, 278
volume, 81
self-adjoint matrix, 281
volume inflation, 243
self-dual group, 168
volume of parallelepiped, 207
separable space, 372
volume of sphere, 305, 306, 311
separation of variables, 380, 385
sequence, 22 wave equation, 78, 179
sequence of functions, 45 wave function, 376
series, 25
sign of system of vectors, 210
signature, 217