Prime Numbers

and the
Riemann Hypothesis

Barry Mazur
William Stein

1
Contents

Preface 5

I The Riemann Hypothesis 11

1 Thoughts about numbers 12

2 What are prime numbers? 15

3 “Named” prime numbers 20

4 Sieves 22

5 Questions about primes 25

6 Further questions about primes 28

7 How many primes are there? 32

8 Prime numbers viewed from a distance 37

9 Pure and applied mathematics 39

10 A probabilistic first guess 41

11 What is a “good approximation”? 45

12 Square root error and random walks 47

13 What is Riemann’s Hypothesis? 49

14 The mystery moves to the error term 51

15 Cesàro smoothing 52

16 A view of | Li(X) − π(X)| 54

2
CONTENTS 3

17 The Prime Number Theorem 56

18 The staircase of primes 60

19 Tinkering with the staircase of primes 62

20 Computer music files and prime numbers 65

21 The word “Spectrum” 71

22 Spectra and trigonometric sums 73

23 The spectrum and the staircase of primes 75

24 To our readers of Part I 77

II Distributions 78
25 Slopes of graphs that have no slopes 79

26 Distributions 86

27 Fourier transforms: second visit 92

28 Fourier transform of delta 95

29 Trigonometric series 97

30 A sneak preview of Part III 99

III The Riemann Spectrum of the Prime Numbers 105

31 On losing no information 106

32 From primes to the Riemann spectrum 109

33 How many θi ’s are there? 114

34 Further questions about the Riemann spectrum 117

35 From the Riemann spectrum to primes 119

IV Back to Riemann 121

36 Building π(X) from the spectrum 122

37 As Riemann envisioned it 128

4 CONTENTS

38 Companions to the zeta function 135

Endnotes 141
Preface

The Riemann Hypothesis is one of the great unsolved problems of mathematics,

and the reward of $1,000,000 of Clay Mathematics Institute prize money awaits
the person who solves it. But—with or without money—its resolution is crucial
for our understanding of the nature of numbers.

There are several full-length books recently published, written for a general
audience, that have the Riemann Hypothesis as their main topic. A reader
of these books will get a fairly rich picture of the personalities engaged in the
pursuit, and of related mathematical and historical issues.1

This is not the mission of the book that you now hold in your hands. We
aim—instead—to explain, in as direct a manner as possible and with the least
mathematical background required, what this problem is all about and why it
is so important. For even before anyone proves this hypothesis to be true (or
false!), just getting familiar with it, and with some of the ideas behind it, is
exciting. Moreover, this hypothesis is of crucial importance in a wide range of
mathematical fields; for example, it is a confidence-booster for computational
mathematics: even if the Riemann Hypothesis is never proved, assuming its
truth (and that of closely related hypotheses) gives us an excellent sense of how
long certain computer programs will take to run, which, in some cases, gives us
the assurance we need to initiate a computation that might take weeks or even
months to complete.

1 See, e.g., The Music of the Primes by Marcus du Sautoy (2003) and Prime Obsession:

Bernhard Riemann and the Greatest Unsolved Problem in Mathematics by John Derbyshire
(2003).

5
6 CONTENTS

Figure 1: Peter Sarnak

Here is how the Princeton mathematician Peter Sarnak describes the broad
impact the Riemann Hypothesis has had2 :

“The Riemann hypothesis is the central problem and it implies many,

many things. One thing that makes it rather unusual in mathemat-
ics today is that there must be over five hundred papers—somebody
should go and count—which start ‘Assume the Riemann hypothe-
sis3 ,’ and the conclusion is fantastic. And those [conclusions] would
then become theorems ... With this one solution you would have
proven five hundred theorems or more at once.”

So, what is the Riemann Hypothesis? Below is a first description of what it is

about. The task of our book is to develop the following boxed paragraph into
a fuller explanation and to convince you of the importance and beauty of the
mathematics it represents. We will be offering, throughout our book, a number
of different—but equivalent—ways of precisely formulating this hypothesis (we
display these in boxes). When we say that two mathematical statements are
“equivalent” we mean that, given the present state of mathematical knowledge,
we can prove that if either one of those statements is true, then the other is true.
The endnotes will guide the reader to the relevant mathematical literature.

2 See page 222 of The Riemann hypothesis: the greatest unsolved problem in mathematics

by Karl Sabbagh (2002).

3 Technically, a generalized version of the Riemann hypothesis (see Chapter 38 below).
CONTENTS 7

What sort of Hypothesis is the Riemann Hypothesis?

Consider the seemingly innocuous series of questions:

• How many prime numbers (2, 3, 5, 7, 11, 13, 17, . . .) are

there less than 100?
• How many less than 10,000?
• How many less than 1,000,000?
More generally, how many primes are there less than any given
number X?

Riemann proposed, a century and half ago, a strikingly simple-to-describe

“very good approximation” to the number of primes less than a given num-
ber X. We now see that if we could prove this Hypothesis of Riemann we
would have the key to a wealth of powerful mathematics. Mathematicians
are eager to find that key.

Figure 2: Raoul Bott (1923–2005). Photograph by George M. Bergman, Cour-

tesy of the Department of Mathematics, Harvard University.

The mathematician Raoul Bott—in giving advice to a student—once said that

whenever one reads a mathematics book or article, or goes to a math lecture,
one should aim to come home with something very specific (it can be small,
but should be specific) that has application to a wider class of mathematical
problems than was the focus of the text or lecture. If we were to suggest some
possible specific items to come home with, after reading our book, three key
phrases—prime numbers, square-root accurate, and spectrum—would
head the list. As for words of encouragement to think hard about the first of
these, i.e., prime numbers, we can do no better than to quote a paragraph of
Don Zagier’s classic 12-page exposition, The First 50 Million Prime Numbers:
8 CONTENTS

Figure 3: Don Zagier

“There are two facts about the distribution of prime numbers of

which I hope to convince you so overwhelmingly that they will be
permanently engraved in your hearts. The first is that, [they are]
the most arbitrary and ornery objects studied by mathematicians:
they grow like weeds among the natural numbers, seeming to obey
no other law than that of chance, and nobody can predict where the
next one will sprout. The second fact is even more astonishing, for
it states just the opposite: that the prime numbers exhibit stunning
regularity, that there are laws governing their behavior, and that
they obey these laws with almost military precision.”

Mathematics is flourishing. Each year sees new exciting initiatives that extend
and sharpen the applications of our subject, new directions for deep exploration—
and finer understanding—of classical as well as very contemporary mathemat-
ical domains. We are aided in such explorations by the development of more
and more powerful tools. We see resolutions of centrally important questions.
And through all of this, we are treated to surprises and dramatic changes of
viewpoint; in short: marvels.
And what an array of wonderful techniques allow mathematicians to do their
work: framing definitions; producing constructions; formulating analogies re-
lating disparate concepts, and disparate mathematical fields; posing conjectures,
that cleanly shape a possible way forward; and, the keystone: providing unas-
sailable proofs of what is asserted, the idea of doing such a thing being itself one
of the great glories of mathematics.
Number theory has its share of this bounty. Along with all these modes of
theoretical work, number theory also offers the pure joy of numerical experi-
mentation, which—when it is going well—allows you to witness the intricacy of
CONTENTS 9

numbers and profound inter-relations that cry out for explanation. It is strik-
ing how little you actually have to know in order to appreciate the revelations
offered by numerical exploration.
Our book is meant to be an introduction to these pleasures. We take an exper-
imental view of the fundamental ideas of the subject buttressed by numerical
computations, often displayed as graphs. As a result, our book is profusely
illustrated, containing 131 figures, diagrams, and pictures that accompany the
text.4
There are few mathematical equations in Part I. This first portion of our book
is intended for readers who are generally interested in, or curious about, math-
ematical ideas, but who may not have studied any advanced topics. Part I is
devoted to conveying the essence of the Riemann Hypothesis and explaining
why it is so intensely pursued. It requires a minimum of mathematical knowl-
edge, and does not, for example, use calculus, although it would be helpful to
know—or to learn on the run—the meaning of the concept of function. Given
its mission, Part I is meant to be complete, in that it has a beginning, middle,
and end. We hope that our readers who only read Part I will have enjoyed the
excitement of this important piece of mathematics.
Part II is for readers who have taken at least one class in calculus, possibly
a long time ago. It is meant as a general preparation for the type of Fourier
analysis that will occur in the later parts. The notion of spectrum is key.
Part III is for readers who wish to see, more vividly, the link between the
placement of prime numbers and (what we call there) the Riemann spectrum.
Part IV requires some familiarity with complex analytic functions, and returns
to Riemann’s original viewpoint. In particular it relates the “Riemann spec-
trum” that we discuss in Part III to the nontrivial zeroes of the Riemann zeta
function. We also provide a brief sketch of the more standard route taken by
published expositions of the Riemann Hypothesis.
The end-notes are meant to link the text to references, but also to provide
more technical commentary with an increasing dependence on mathematical
background in the later chapters. References to the end notes will be in brackets.
We wrote our book over the past decade, but devoted only one week to it each
year (a week in August). At the end of our work-week for the book, each year,
we put our draft (mistakes and all) on line to get response from readers.5 We

4 We created the figures using the free SageMath software (see http://www.sagemath.org).

Complete source code is available, which can be used to recreate every diagram in this book
(see http://wstein.org/rh). More adventurous readers can try to experiment with the pa-
rameters for the ranges of data illustrated, so as to get an even more vivid sense of how the
numbers “behave.” We hope that readers become inspired to carry out numerical experimen-
tation, which is becoming easier as mathematical software advances.
5 See http://library.fora.tv/2014/04/25/Riemann_Hypothesis_The_Million_Dollar_
Challenge which is a lecture—and Q & A—about the composition of this book.
10 CONTENTS

therefore accumulated much important feedback, corrections, and requests from

readers.6 We thank them infinitely.

6 Including Dan Asimov, Bret Benesh, Keren Binyaminov, Harald Bögeholz, Louis-Philippe

Chiasson, Keith Conrad, Karl-Dieter Crisman, Nicola Dunn, Thomas Egense, Bill Gosper,
Andrew Granville, Shaun Griffith, Michael J. Gruber, Robert Harron, William R. Hearst
III, David Jao, Fredrik Johansson, Jim Markovitch, David Mumford, James Propp, Andrew
Solomon, Dennis Stein, and Chris Swenson.
 Part I

The Riemann Hypothesis

11
Chapter 1

Thoughts about numbers:

ancient, medieval, and
modern

If we are to believe the ancient Greek philosopher Aristotle, the early Pythagore-
ans thought that the principles governing Number are “the principles of all
things,” the concept of Number being more basic than earth, air, fire, or water,
which were according to ancient tradition the four building blocks of matter.
To think about Number is to get close to the architecture of “what is.”
So, how far along are we in our thoughts about numbers?

Figure 1.1: René Descartes (1596–1650)

The French philosopher and mathematician René Descartes, almost four cen-
turies ago, expressed the hope that there soon would be “almost nothing more

12
 13

to discover in geometry.” Contemporary physicists dream of a final theory.1 But

despite its venerability and its great power and beauty, the pure mathematics
of numbers may still be in the infancy of its development, with depths to be
explored as endless as the human soul, and never a final theory.

Figure 1.2: Jean de Bosschere, “Don Quixote and his Dulcinea del Toboso,”
from The History of Don Quixote De La Mancha, by Miguel De Cervantes.
Trans. Thomas Shelton. George H. Doran Company, New York (1923).

Numbers are obstreperous things. Don Quixote encountered this when he re-
quested that the “bachelor” compose a poem to his lady Dulcinea del Toboso,
the first letters of each line spelling out her name. The “bachelor” found2

“a great difficulty in their composition because the number of letters

in her name was 17, and if he made four Castilian stanzas of four
octosyllabic lines each, there would be one letter too many, and if
he made the stanzas of five octosyllabic lines each, the ones called
décimas or redondillas, there would be three letters too few...”

“It must fit in, however you do it,” pleaded Quixote, not willing to grant the
imperviousness of the number 17 to division.

1 See Weinberg’s book Dreams of a Final Theory: The Search for the Fundamental Laws

of Nature, by Steven Weinberg (New York: Pantheon Books, 1992)

2 See Chapter IV of the Second Part of the Ingenious Gentleman Don Quixote of La Man-
cha.
14 CHAPTER 1. THOUGHTS ABOUT NUMBERS

Seventeen is indeed a prime number: there is no way of factoring it as the prod-

uct of smaller numbers, and this accounts—people tell us—for its occurrence in
some phenomena of nature, as when the seventeen-year cicadas all emerged to
celebrate a “reunion” of some sort in our fields and valleys.

Figure 1.3: Cicadas emerge every 17 years

Prime numbers, despite their primary position in our modern understanding of

numbers, were not specifically doted over in the ancient literature before Euclid,
at least not in the literature that has been preserved. Primes are mentioned as
a class of numbers in the writings of Philolaus (a predecessor of Plato); they are
not mentioned specifically in the Platonic dialogues, which is surprising given
the intense interest Plato had in mathematical developments; and they make
an occasional appearance in the writings of Aristotle, which is not surprising,
given Aristotle’s emphasis on the distinction between the composite and the
incomposite. “The incomposite is prior to the composite,” writes Aristotle in
Book 13 of the Metaphysics.
Prime numbers do occur, in earnest, in Euclid’s Elements!

Figure 1.4: Euclid

There is an extraordinary wealth of established truths about whole numbers;

these truths provoke sheer awe for the beautiful complexity of prime numbers.
But each of the important new discoveries we make gives rise to a further richness
of questions, educated guesses, heuristics, expectations, and unsolved problems.
Chapter 2

What are prime numbers?

Primes as atoms. To begin from the beginning, think of the operation of mul-
tiplication as a bond that ties numbers together: the equation 2 × 3 = 6 invites
us to imagine the number 6 as (a molecule, if you wish) built out of its smaller
constituents 2 and 3. Reversing the procedure, if we start with a whole number,
say 6 again, we may try to factor it (that is, express it as a product of smaller
whole numbers) and, of course, we would eventually, if not immediately, come
up with 6 = 2 × 3 and discover that 2 and 3 factor no further; the numbers 2
and 3, then, are the indecomposable entities (atoms, if you wish) that comprise
our number.
6

2 3

Figure 2.1: The number 6 = 2 × 3

By definition, a prime number (colloquially, a prime) is a whole number,

bigger than 1, that cannot be factored into a product of two smaller whole
numbers. So, 2 and 3 are the first two prime numbers. The next number along
the line, 4, is not prime, for 4 = 2 × 2; the number after that, 5, is. Primes are,
multiplicatively speaking, the building blocks from which all numbers can be
made. A fundamental theorem of arithmetic tells us that any number (bigger
than 1) can be factored as a product of primes, and the factorization is unique
except for rearranging the order of the primes.
For example, if you try to factor 12 as a product of two smaller numbers—
ignoring the order of the factors—there are two ways to begin to do this:
12 = 2 × 6 and 12 = 3 × 4
But neither of these ways is a full factorization of 12, for both 6 and 4 are

15
16 CHAPTER 2. WHAT ARE PRIME NUMBERS?

not prime, so can be themselves factored, and in each case after changing the
ordering of the factors we arrive at:

12 = 2 × 2 × 3.

If you try to factor the number 300, there are many ways to begin:

300 = 30 × 10 or 300 = 6 × 50

and there are various other starting possibilities. But if you continue the fac-
torization (“climbing down” any one of the possible “factoring trees”) to the
bottom, where every factor is a prime number as in Figure 2.2, you always end
up with the same collection of prime numbers1 :

300 = 22 × 3 × 52 .

300 300
3 100 20 15
10 10 2 10 5 3
2 5 2 5 2 5

Figure 2.2: Factor trees that illustrate the factorization of 300 as a product of
primes.

6469693230

690 9376367

30 23 13 721259

3 10 1309 551

2 5 119 11 19 29

17 7

Figure 2.3: Factorization tree for the product of the primes up to 29.
1 See
Section 1.1 of Stein’s Elementary Number Theory: Primes, Congruences, and Secrets
(2008) at http://wstein.org/ent/ for a proof of the “fundamental theorem of arithmetic,”
which asserts that every positive whole number factors uniquely as a product of primes.
 17

The Riemann Hypothesis probes the question: how intimately can we know
prime numbers, those atoms of multiplication? Prime numbers are an important
part of our daily lives. For example, often when we visit a website and purchase
something online, prime numbers having hundreds of decimal digits are used to
keep our bank transactions private. This ubiquitous use to which giant primes
are put depends upon a very simple principle: it is much easier to multiply
numbers together than to factor them. If you had to factor, say, the number
391 you might scratch your head for a few minutes before discovering that 391 is
17 × 23. But if you had to multiply 17 by 23 you would do it straightaway. Offer
two primes, say, P and Q each with a few hundred digits, to your computing
machine and ask it to multiply them together: you will get their product N =
P × Q with its hundreds of digits in about a microsecond. But present that
number N to any current desktop computer, and ask it to factor N , and the
computer will (almost certainly) fail to do the task. See [1] and [2].
The safety of much encryption depends upon this “guaranteed” failure!2
If we were latter-day number-phenomenologists we might revel in the discovery
and proof that

p = 243,112,609 − 1 = 3164702693 . . . . . . (millions of digits) . . . . . . 6697152511

is a prime number, this number having 12,978,189 digits! This prime, which
was discovered on August 23, 2008 by the GIMPS project,3 is the first prime
ever found with more than ten million digits, though it is not the largest prime
currently known.
Now 243,112,609 − 1 is quite a hefty number! Suppose someone came up to you
saying “surely p = 243,112,609 − 1 is the largest prime number!” (which it is not).
How might you convince that person that he or she is wrong without explicitly
exhibiting a larger prime? [3]
Here is a neat—and, we hope, convincing—strategy to show there are prime
numbers larger than p = 243,112,609 − 1. Imagine forming the following humon-
gous number: let M be the product of all prime numbers up to and including
p = 243,11,2609 − 1. Now go one further than M by taking the next number
N = M + 1.
OK, even though this number N is wildly large, it is either a prime number
itself—which would mean that there would indeed be a prime number larger
than p = 243,112,609 − 1, namely N ; or in any event it is surely divisible by some
prime number, call it P .
Here, now, is a way of seeing that this P is bigger than p: Since every prime
number smaller than or equal to p divides M , these prime numbers cannot

2 Nobody has ever published a proof that there is no fast way to factor integers. This is an

article of “faith” among some cryptographers.

3 The GIMPS project website is http://www.mersenne.org/.
18 CHAPTER 2. WHAT ARE PRIME NUMBERS?

divide N = M + 1 (since they divide M evenly, if you tried to divide N = M + 1

by any of them you would get a remainder of 1). So, since P does divide N it
must not be any of the smaller prime numbers: P is therefore a prime number
bigger than p = 243,112,609 − 1.
This strategy, by the way, is not very new: it is, in fact, well over two thousand
years old, since it already occurred in Euclid’s Elements. The Greeks did know
that there are infinitely many prime numbers and they showed it via the same
method as we showed that our p = 243,112,609 −1 is not the largest prime number.
Here is the argument again, given very succinctly: Given primes p1 , . . . , pm , let
n = p1 p2 · · · pm + 1. Then n is divisible by some prime not equal to any pi , so
there are more than m primes.
You can think of this strategy as a simple game that you can play. Start with
the bag of prime numbers that contains just the two primes 2 and 3. Now each
“move” of the game consists of multiplying together all the primes you have in
your bag to get a number M , then adding 1 to M to get the larger number
N = M + 1, then factoring N into prime number factors, and then including all
those new prime numbers in your bag. Euclid’s proof gives us that we will—
with each move of this game—be finding more prime numbers: the contents of
the bag will increase. After, say, a million moves our bag will be guaranteed to
contain more than a million prime numbers.
For example, starting the game with your bag containing only one prime number
2, here is how your bag grows with successive moves of the game:
{2}
{2, 3}
{2, 3, 7}
{2, 3, 7, 43}
{2, 3, 7, 43, 13, 139}
{2, 3, 7, 43, 13, 139, 3263443}
{2, 3, 7, 43, 13, 139, 3263443, 547, 607, 1033, 31051}
{2, 3, 7, 43, 13, 139, 3263443, 547, 607, 1033, 31051, 29881, 67003,
9119521, 6212157481}
etc.4
Though there are infinitely many primes, explicitly finding large primes is a
major challenge. In the 1990s, the Electronic Frontier Foundation http://www.
eff.org/awards/coop offered a $100,000 cash reward to the first group to find
a prime with at least 10,000,000 decimal digits (the group that found the record
prime p above won this prize5 ), and offers another $150,000 cash prize to the

4 The sequence of prime numbers we find by this procedure is discussed in more detail with

references in the Online Encyclopedia of Integer Sequences http://oeis.org/A126263.

5 See http://www.eff.org/press/archives/2009/10/14-0. Also the 46th Mersenne prime

was declared by Time Magazine to be one of the top 50 best “inventions” of 2008: http://www.
time.com/time/specials/packages/article/0,28804,1852747_1854195_1854157,00.html.
 19

first group to find a prime with at least 100,000,000 decimal digits.

The number p = 243,112,609 − 1 was for a time the largest prime known, where
by “know” we mean that we know it so explicitly that we can compute things
about it. For example, the last two digits of p are both 1 and the sum of the
digits of p is 58,416,637. Of course p is not the largest prime number since there
are infinitely many primes, e.g., the next prime q after p is a prime. But there is
no known way to efficiently compute anything interesting about q. For example,
what is the last digit of q in its decimal expansion?
Chapter 3

“Named” prime numbers

Prime numbers come in all sorts of shapes, some more convenient to deal with
than others. For example, the number we have been talking about,

p = 243,112,609 − 1,

is given to us, by its very notation, in a striking form; i.e., one less than a power
of 2. It is no accident that the largest “currently known” prime number has
such a form. This is because there are special techniques we can draw on to
show primality of a number, if it is one less than a power of 2 and—of course—if
it also happens to be prime. The primes of that form have a name, Mersenne
Primes, as do the primes that are one more than a power of 2, those being called
Fermat Primes. [4]
Here are two exercises that you might try to do, if this is your first encounter
with primes that differ from a power of 2 by 1:

1. Show that if a number of the form M = 2n −1 is prime, then the exponent

n is also prime. [Hint: This is equivalent to proving that if n is composite,
then 2n − 1 is also composite.] For example: 22 − 1 = 3, 23 − 1 = 7 are
primes, but 24 − 1 = 15 is not. So Mersenne primes are numbers that are

• of the form 2prime number − 1, and

• are themselves prime numbers.

2. Show that if a number of the form F = 2n + 1 is prime, then the exponent

n is a power of two. For example: 22 + 1 = 5 is prime, but 23 + 1 = 9 is
not. So Fermat primes are numbers that are

• of the form 2power of two + 1, and

• are themselves prime numbers.

20
 21

Not all numbers of the form 2prime number − 1 or of the form 2power of two + 1
are prime. We currently know only finitely many primes of either of these
forms. How we have come to know what we know is an interesting tale. See,
for example, http://www.mersenne.org/.
Chapter 4

Sieves

Eratosthenes, the mathematician from Cyrene (and later, librarian at Alexan-

dria) explained how to sift the prime numbers from the series of all numbers:
in the sequence of numbers,

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26,

for example, start by circling the 2 and crossing out all the other multiples of 2.
Next, go back to the beginning of our sequence of numbers and circle the first
number that is neither circled nor crossed out (that would be, of course, the
3), then cross out all the other multiples of 3. This gives the pattern: go back
again to the beginning of our sequence of numbers and circle the first number
that is neither circled nor crossed out; then cross out all of its other multiples.
Repeat this pattern until all the numbers in our sequence are either circled, or
crossed out, the circled ones being the primes.

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

In Figures 4.1–4.4 we use the primes 2, 3, 5, and finally 7 to sieve out the primes
up to 100, where instead of crossing out multiples we grey them out, and instead
of circling primes we color their box red.

22
 23

2 3 4 5 6 7 8 9 10

11 12 13 14 15 16 17 18 19 20

21 22 23 24 25 26 27 28 29 30

31 32 33 34 35 36 37 38 39 40

41 42 43 44 45 46 47 48 49 50

51 52 53 54 55 56 57 58 59 60

61 62 63 64 65 66 67 68 69 70

71 72 73 74 75 76 77 78 79 80

81 82 83 84 85 86 87 88 89 90

91 92 93 94 95 96 97 98 99 100

Figure 4.1: Using the prime 2 to sieve for primes up to 100

Since all the even numbers greater than two are eliminated as being composite
numbers and not primes they appear as gray in Figure 4.1, but none of the odd
numbers are eliminated so they still appear in white boxes.

2 3 5 7 9

11 13 15 17 19

21 23 25 27 29

31 33 35 37 39

41 43 45 47 49

51 53 55 57 59

61 63 65 67 69

71 73 75 77 79

81 83 85 87 89

91 93 95 97 99

Figure 4.2: Using the primes 2 and 3 to sieve for primes up to 100

2 3 5 7

11 13 17 19

23 25 29

31 35 37

41 43 47 49

53 55 59

61 65 67

71 73 77 79

83 85 89

91 95 97

Figure 4.3: Using the primes 2, 3, and 5 to sieve for primes up to 100

Looking at Figure 4.3, we see that for all but three numbers (49, 77, and 91) up
24 CHAPTER 4. SIEVES

to 100 we have (after sieving by 2,3, and 5) determined which are primes and
which composite.
2 3 5 7

11 13 17 19

23 29

31 37

41 43 47 49

53 59

61 67

71 73 77 79

83 89

91 97

Figure 4.4: Using the primes 2, 3, 5, and 7 to sieve for primes up to 100

Finally, we see in Figure 4.4 that sieving by 2, 3, 5, and 7 determines all primes
up to 100. See [5] for more about explicitly enumerating primes using a com-
puter.
Chapter 5

Questions about primes

that any person might ask

We become quickly stymied when we ask quite elementary questions about the
spacing of the infinite series of prime numbers.
For example, are there infinitely many pairs of primes whose difference is 2?
The sequence of primes seems to be rich in such pairs
5 − 3 = 2, 7 − 5 = 2, 13 − 11 = 2, 19 − 17 = 2,
and we know that there are loads more such pairs1 but the answer to our
question, are there infinitely many?, is not known. The conjecture that there
are infinitely many such pairs of primes (“twin primes” as they are called)
is known as the Twin Primes Conjecture. Are there infinitely many pairs of
primes whose difference is 4, 6? Answer: equally unknown. Nevertheless there
is very exciting recent work in this direction, specifically, Yitang Zhang proved
that there are infinitely many pairs of primes that differ by no more than 7 ×
107 . For a brief account of Zhang’s work, see the Wikipedia entry http://
en.wikipedia.org/wiki/Yitang_Zhang. Many exciting results have followed
Zhang’s breakthrough; we know now, thanks to results2 of James Maynard and
others, that there are infinitely many pairs of primes that differ by no more than
246.
Is every even number greater than 2 a sum of two primes? Answer: unknown.
Are there infinitely many primes which are 1 more than a perfect square? An-
swer: unknown.
1 For example, according to http://oeis.org/A007508 there are 10,304,185,697,298 such

pairs less than 10,000,000,000,000,000.

2 See https://www.simonsfoundation.org/quanta/20131119-together-and-alone-closing-the-prime-gap/

and for further work http://michaelnielsen.org/polymath1/index.php?title=Bounded_

gaps_between_primes.

25
26 CHAPTER 5. QUESTIONS ABOUT PRIMES

Figure 5.1: Yitang Zhang

Remember the Mersenne prime p = 243,112,609 − 1 from Chapter 3 and how

we proved—by pure thought—that there must be a prime P larger than p?
Suppose, though, someone asked us whether there was a Mersenne Prime larger
than this p: that is, is there a prime number of the form

2some prime number

−1

bigger than p = 243,112,609 − 1? Answer: For many years we did not know;
however, in 2013 Curtis Cooper discovered the even bigger Mersenne prime
257,885,161 − 1, with a whopping 17,425,170 digits! Again we can ask if there is a
Mersenne prime larger than Cooper’s. Answer: we do not know. It is possible
that there are infinitely many Mersenne primes but we’re far from being able
to answer such questions.

Figure 5.2: Marin Mersenne (1588–1648)

Is there some neat formula giving the next prime? More specifically, if I give you
a number N , say N = one million, and ask you for the first number after N that
 27

is prime, is there a method that answers that question without, in some form or
other, running through each of the successive odd numbers after N rejecting the
nonprimes until the first prime is encountered? Answer: unknown.
One can think of many ways of “getting at” some understanding of the place-
ment of prime numbers among all numbers. Up to this point we have been
mainly just counting them, trying to answer the question “how many primes
are there up to X?” and we have begun to get some feel for the numbers behind
this question, and especially for the current “best guesses” about estimates.
What is wonderful about this subject is that people attracted to it cannot resist
asking questions that lead to interesting, and sometimes surprising numerical
experiments. Moreover, given our current state of knowledge, many of the
questions that come to mind are still unapproachable: we don’t yet know enough
about numbers to answer them. But asking interesting questions about the
mathematics that you are studying is a high art, and is probably a necessary
skill to acquire, in order to get the most enjoyment—and understanding—from
mathematics. So, we offer this challenge to you:
Come up with your own question about primes that

• is interesting to you,
• is not a question whose answer is known to you,
• is not a question that you’ve seen before; or at least not exactly,

• is a question about which you can begin to make numerical investigations.

If you are having trouble coming up with a question, read on for more examples
that provide further motivation.
Chapter 6

Further questions about

primes

In celebration of Yitang Zhang’s recent result, let us consider more of the nu-
merics regarding gaps between one prime and the next, rather than the tally of
all primes. Of course, it is no fun at all to try to guess how many pairs of primes
p, q there are with gap q − p equal to a fixed odd number, since the difference of
two odd numbers is even, as in Chapter 5. The fun, though, begins in earnest
if you ask for pairs of primes with difference equal to 2 (these being called twin
primes) for it has long been guessed that there are infinitely many such pairs of
primes, but no one has been able to prove this yet.
As of 2014, the largest known twin primes are

3756801695685 · 2666669 ± 1.

These enormous primes, which were found in 2011, have 200,700 digits each.1
Similarly, it is interesting to consider primes p and q with difference 4, or 8,
or—in fact—any even number 2k. That is, people have guessed that there are
infinitely many pairs of primes with difference 4, with difference 6, etc. but none
of these guesses have yet been proved.
So, define
Gapk (X)
to be the number of pairs of consecutive primes (p, q) with q < X that have “gap
k” (i.e., such that their difference q − p is k). Here p is a prime, q > p is a prime,
and there are no primes between p and q. For example, Gap2 (10) = 2, since the
pairs (3, 5) and (5, 7) are the pairs less than 10 with gap 2, and Gap4 (10) = 0

1 See http://primes.utm.edu/largest.html#twin for the top ten largest known twin

primes.

28
 29

because despite 3 and 7 being separated by 4, they are not consecutive primes.
See Table 6.1 for various values of Gapk (X) and Figure 6.1 for the distribution
of prime gaps for X = 107 .

Table 6.1: Values of Gapk (X)

X Gap2 (X) Gap4 (X) Gap6 (X) Gap8 (X) Gap100 (X) Gap246 (X)
10 2 0 0 0 0 0
102 8 7 7 1 0 0
103 35 40 44 15 0 0
104 205 202 299 101 0 0
105 1224 1215 1940 773 0 0
106 8169 8143 13549 5569 2 0
107 58980 58621 99987 42352 36 0
108 440312 440257 768752 334180 878 0

The recent results of Zhang as sharpened by Maynard (and others) we mentioned

above tell us that for at least one even number k among the even numbers
k ≤ 246, Gapk (X) goes to infinity as X goes to infinity. One expects that this
happens for all even numbers k. We expect this as well, of course, for Gap246 (X)
despite what might be misconstrued as discouragement by the above data.

90000

60000

20000

2 4 6 8 10 20 30 40 48

Figure 6.1: Frequency histogram showing the distribution of prime gaps of size
≤ 50 for all primes up to 107 . Six is the most popular gap in this data.
Gap 6

200

Gap 2
150 Gap 4

100
Gap 8

50

1000 2000 3000 4000 5000 6000 7000

Figure 6.2: Plots of Gapk (X) for k = 2, 4, 6, 8. Which wins?

30 CHAPTER 6. FURTHER QUESTIONS ABOUT PRIMES

Here is yet another question that deals with the spacing of prime numbers that
we do not know the answer to:
Racing Gap 2, Gap 4, Gap 6, and Gap 8 against each other:

Challenge: As X tends to infinity which of Gap2 (X), Gap4 (X),

Gap6 (X), or Gap8 (X) do you think will grow faster? How much
would you bet on the truth of your guess? [6]

Here is a curious question that you can easily begin to check out for small
numbers. We know, of course, that the even numbers and the odd numbers are
nicely and simply distributed: after every odd number comes an even number,
after every even, an odd. There are an equal number of positive odd numbers
and positive even numbers less than any given odd number, and there may be
nothing else of interest to say about the matter. Things change considerably,
though, if we focus our concentration on multiplicatively even numbers and
multiplicatively odd numbers.
A multiplicatively even number is one that can be expressed as a product
of an even number of primes; and a multiplicatively odd number is one that
can be expressed as a product of an odd number of primes. So, any prime is
multiplicatively odd, the number 4 = 2 · 2 is multiplicatively even, and so is
6 = 2 · 3, 9 = 3 · 3, and 10 = 2 · 5; but 12 = 2 · 2 · 3 is multiplicatively odd. Below
we list the numbers up to 25, and underline and bold the multiplicatively odd
numbers.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Table 6.2 gives some data:

Table 6.2: Count of multiplicatively odd and even positive numbers ≤ X

X 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
m. odd 0 1 2 2 3 3 4 5 5 5 6 7 8 8 8 8
m. even 1 1 1 2 2 3 3 3 4 5 5 5 5 6 7 8

Now looking at this data, a natural, and simple, question to ask about the
concept of multiplicative oddness and evenness is:
Is there some X ≥ 2 for which there are more multiplicatively even numbers less
than or equal to X than multiplicatively odd ones?
Each plot in Figure 6.3 gives the number of multiplicatively even numbers be-
tween 2 and X minus the number of multiplicatively odd numbers between 2
and X, for X equal to 10, 100, 1000, 10000, 100000, and 1000000. The above
question asks whether these graphs would, for sufficiently large X, ever cross
the X-axis.
 31

0 0

-2
-1
-4
-2
-6

-3 -8

-10
-4
2 4 6 8 10 12 14 16 0 20 40 60 80 100
0 0
-5 -20

-10 -40
-60
-15
-80
-20
-100
-25
-120
0 200 400 600 800 1000 0 2000 4000 6000 8000 10000
0 0
-200
-100
-400
-200 -600
-800
-300
-1000

-400 -1200
0 2e4 4e4 6e4 8e4 1e5 0 2e5 4e5 6e5 8e5 1e6

Figure 6.3: Racing Multiplicatively Even and Odd Numbers.

A negative response to this question—i.e., a proof that any plot as drawn in

Figure 6.3 never crosses the X-axis—would imply the Riemann Hypothesis! In
contrast to the list of previous questions, the answer to this question is known2 :
alas, there is such an X. In 1960, Lehman showed that for X = 906,400,000
there are 708 more multiplicatively even numbers up to X than multiplicatively
odd numbers (Tanaka found in 1980 that the smallest X such that there are
more multiplicative even than odd numbers is X = 906,150,257).
These are questions that have been asked about primes (and we could give
bushels more3 ), questions expressible in simple vocabulary, that we can’t answer
today. We have been studying numbers for over two millennia and yet we are
indeed in the infancy of our understanding.
We’ll continue our discussion by returning to the simplest counting question
about prime numbers.

2 For more details, see P. Borwein, “Sign changes in sums of the Liouville Function” and

the nice short paper of Norbert Wiener “Notes on Polya’s and Turan’s hypothesis concerning
Liouville’s factor” (page 765 of volume II of Wiener’s Collected Works); see also: G. Pólya
“Verschiedene Bemerkungen zur Zahlentheorie,” Jahresbericht der Deutschen Mathematiker-
Vereinigung, 28 (1919) 31–40.
3 See, e.g., Richard Guy’s book Unsolved Problems in Number Theory (2004).
Chapter 7

How many primes are

there?

2 3 5 7 11 13
17 19 23
29 31 37 41
43 47 53
59 61 67
71 73 79 83
89 97
101 103 107 109
113
127 131 137 139
149 151
157 163 167
169 173 179 181
191 193

Figure 7.1: Sieving primes up to 200

Slow as we are to understand primes, at the very least we can try to count them.
You can see that there are 10 primes less than 30, so you might encapsulate this
by saying that the chances that a number less than 30 is prime is 1 in 3. This
frequency does not persist, though; here is some more data: There are 25 primes
less than 100 (so 1 in 4 numbers up to 100 are prime), there are 168 primes less
than a thousand (so we might say that among the numbers less than a thousand
the chances that one of them is prime is roughly 1 in 6).

32
 33

0.6

0.5

0.4

0.3

0.2

0.1

20 40 60 80 100

Figure 7.2: Graph of the proportion of primes up to X for each integer X ≤ 100

There are 78,498 primes less than a million (so we might say that the chances
that a random choice among the first million numbers is prime have dropped to
roughly 1 in 13).

0.6

0.5

0.4

0.3

0.2

0.1

200 400 600 800 1000

Figure 7.3: Proportion of primes for X up to 1,000

34 CHAPTER 7. HOW MANY PRIMES ARE THERE?

0.6

0.5

0.4

0.3

0.2

0.1

2000 4000 6000 8000 10000

Figure 7.4: Proportion of primes for X up to 10,000

There are 455,052,512 primes less than ten billion; i.e., 10,000,000,000 (so we
might say that the chances are down to roughly 1 in 22).

Primes, then, seem to be thinning out. We return to the sifting process we

carried out earlier, and take a look at a few graphs, to get a sense of why that
might be so. There are a 100 numbers less than or equal to 100, a thousand
numbers less than or equal to 1000, etc.: the graph in Figure 7.5 that looks like
a regular staircase, each step the same length as each riser, climbing up at, so
to speak, a 45 degree angle, counts all numbers up to and including X.

Following Eratosthenes, we have sifted those numbers, to pan for primes. Our
first move was to throw out roughly half the numbers (the even ones!) after the
number 2. The graph that is labeled “Sieve by 2” in Figure 7.5 that is, with
one hiccup, a regular staircase climbing at a smaller angle, each step twice the
length of each riser, illustrates the numbers that are left after one pass through
Eratosthenes’ sieve, which includes, of course, all the primes. So, the chances
that a number bigger than 2 is prime is at most 1 in 2. Our second move was
to throw out a good bunch of numbers bigger than 3. So, the chances that a
number bigger than 3 is prime is going to be even less. And so it goes: with
each move in our sieving process, we are winnowing the field more extensively,
reducing the chances that the later numbers are prime.
 35

100 All Numbers

80
60
Sieve by 2
40
20 Primes

20 40 60 80 100

Figure 7.5: Sieving by removing multiples of 2 up to 100

1000 All Numbers

800
600
400
200
Primes
200 400 600 800 1000

Figure 7.6: Sieving for primes up to 1000

The red curve in these figures actually counts the primes: it is the beguilingly
irregular staircase of primes. Its height above any number X on the horizontal
line records the number of primes less than or equal to X, the accumulation of
primes up to X. Refer to this number as π(X). So π(2) = 1, π(3) = 2, π(30) =
10; of course, we could plot a few more values of π(X), like π(ten billion) =
455, 052, 512.
Let us accompany Eratosthenes for a few further steps in his sieving process.
Figure 7.7 contains a graph of all whole numbers up to 100 after we have removed
the even numbers greater than 2, and the multiples of 3 greater than 3 itself.

35 Sieve by 2, 3
30
25
20
15
10
5
20 40 60 80 100

Figure 7.7: Sieving out multiples of 2 and 3.

From this graph you can see that if you go “out a way” the likelihood that
a number is a prime is less than 1 in 3. Figure 7.8 contains a graph of what
Eratosthenes sieve looks like up to 100 after sifting 2, 3, 5, and 7.
36 CHAPTER 7. HOW MANY PRIMES ARE THERE?
Sieve by 2, 3, 5, 7
25
20
15
10
5

20 40 60 80 100

Figure 7.8: Sieving out multiples of 2, 3, 5, and 7.

This data may begin to suggest to you that as you go further and further out
on the number line the percentage of prime numbers among all whole numbers
tends towards 0% (it does).
To get a sense of how the primes accumulate, we will take a look at the staircase
of primes for X = 25 and X = 100 in Figures 7.9 and 7.10.

0
5 10 15 20 25

Figure 7.9: Staircase of primes up to 25

25
20
15
10
5

20 40 60 80 100

Figure 7.10: Staircase of primes up to 100

Chapter 8

Prime numbers viewed

from a distance

The striking thing about these figures is that as the numbers get large enough,
the jagged accumulation of primes, those quintessentially discrete entities, be-
comes smoother and smoother to the eye. How strange and wonderful to watch,
as our viewpoint zooms out to larger ranges of numbers, the accumulation of
primes taking on such a smooth and elegant shape.

150

100

50

200 400 600 800 1000

Figure 8.1: Staircases of primes up to 1,000

37
38 CHAPTER 8. PRIME NUMBERS VIEWED FROM A DISTANCE

1200

1000

800

600

400

200

2000 4000 6000 8000 10000

Figure 8.2: Staircases of primes up to 10,000

8000

6000

4000

2000

50000 100000

Figure 8.3: Primes up to 100,000.

But don’t be fooled by the seemingly smooth shape of the curve in the last
figure above: it is just as faithful a reproduction of the staircase of primes as
the typographer’s art can render, for there are thousands of tiny steps and risers
in this curve, all hidden by the thickness of the print of the drawn curve in the
figure. It is already something of a miracle that we can approximately describe
the build-up of primes, somehow, using a smooth curve. But what smooth curve?
That last question is not rhetorical. If you draw a curve with chalk on the
blackboard, this can signify a myriad of smooth (mathematical) curves all en-
compassed within the thickness of the chalk-line, all—if you wish—reasonable
approximations of one another. So, there are many smooth curves that fit the
chalk-curve. With this warning, but very much fortified by the data of Fig-
ure 8.3, let us ask: what is a smooth curve that is a reasonable approximation
to the staircase of primes?
Chapter 9

Pure and applied

mathematics

Mathematicians seems to agree that, loosely speaking, there are two types of
mathematics: pure and applied. Usually—when we judge whether a piece of
mathematics is pure or applied—this distinction turns on whether or not the
math has application to the “outside world,” i.e., that world where bridges are
built, where economic models are fashioned, where computers churn away on
the Internet (for only then do we unabashedly call it applied math), or whether
the piece of mathematics will find an important place within the context of
mathematical theory (and then we label it pure). Of course, there is a great
overlap (as we will see later, Fourier analysis plays a major role both in data
compression and in pure mathematics).
Moreover, many questions in mathematics are “hustlers” in the sense that, at
first view, what is being requested is that some simple task be done (e.g., the
question raised in this book, to find a smooth curve that is a reasonable approx-
imation to the staircase of primes). And only as things develop is it discovered
that there are payoffs in many unexpected directions, some of these payoffs be-
ing genuinely applied (i.e., to the practical world), some of these payoffs being
pure (allowing us to strike behind the mask of the mere appearance of the math-
ematical situation, and get at the hidden fundamentals that actually govern the
phenomena), and some of these payoffs defying such simple classification, inso-
far as they provide powerful techniques in other branches of mathematics. The
Riemann Hypothesis—even in its current unsolved state—has already shown
itself to have all three types of payoff.
The particular issue before us is, in our opinion, twofold, both applied, and pure:
can we curve-fit the “staircase of primes” by a well approximating smooth curve
given by a simple analytic formula? The story behind this alone is marvelous,
has a cornucopia of applications, and we will be telling it below. But our

39
40 CHAPTER 9. PURE AND APPLIED MATHEMATICS

curiosity here is driven by a question that is pure, and less amenable to precise
formulation: are there mathematical concepts at the root of, and more basic
than (and “prior to,” to borrow Aristotle’s use of the phrase) prime numbers—
concepts that account for the apparent complexity of the nature of primes?
Chapter 10

A probabilistic first guess

Figure 10.1: Gauss

The search for such approximating curves began, in fact, two centuries ago when
Carl Friedrich Gauss defined a certain beautiful curve that, experimentally,
seemed to be an exceptionally good fit for the staircase of primes.

50

40

30

20

10

50 100 150 200 250

Figure 10.2: Plot of π(X) and Gauss’s smooth curve G(X)

41
42 CHAPTER 10. A PROBABILISTIC FIRST GUESS

Let us denote Gauss’s curve G(X); it has an elegant simple formula comprehen-
sible to anyone who has had a tiny bit of calculus. If you make believe that the
chances that a number X is a prime is inversely proportional to the number of
digits of X you might well hit upon Gauss’s curve. That is,

X
G(X) is roughly proportional to .
the number of digits of X

But to describe Gauss’s guess precisely we need to discuss the natural logarithm1
“log(X)” which is an elegant smooth function of a real number X that is roughly
proportional to the number of digits of the whole number part of X.

4
3
2
1

20 40 60 80 100
-1

Figure 10.3: Plot of the natural logarithm log(X)

Euler’s famous constant e = 2.71828182 . . ., which is the limit of the sequence

 2  3  4
1 1 1
1+ , 1+ , 1+ ,...,
2 3 4

is used in the definition of log:

A = log(X) is the number A for which eA = X.

Before electronic calculators, logarithms were frequently used to speed up cal-

culations, since logarithms translate difficult multiplication problems into easier
addition problems which can be done mechanically. Such calculations use that
the logarithm of a product is the sum of the logarithms of the factors; that is,

log(XY ) = log(X) + log(Y ).

1 In advanced mathematics, “common” logarithms are sufficiently uncommon that “log”

almost always denotes natural log and the notation ln(X) is not used.
 43

Figure 10.4: A slide rule computes 2X by using that log(2X) = log(2) + log(X)

In Figure 10.4 the numbers printed (on each of the slidable pieces of the rule) are
spaced according to their logarithms, so that when one slides the rule arranging
it so that the printed number X on one piece lines up with the printed number
1 on the other, we get that for every number Y printed on the first piece, the
printed number on the other piece that is aligned with it is the product XY ; in
effect the “slide” adds log(X) to log(Y ) giving log(XY ).

Figure 10.5: A Letter of Gauss

In 1791, when Gauss was 14 years old, he received a book that contained log-
arithms of numbers up to 7 digits and a table of primes up to 10,009. Years
later, in a letter written in 1849 (see Figure 10.5), Gauss claimed that as early
as 1792 or 1793 he had already observed that the density of prime numbers over
intervals of numbers of a given rough magnitude X seemed to average 1/ log(X).

Very very roughly speaking, this means that the number of primes up to X is
approximately X divided by twice the number of digits of X. For example, the
44 CHAPTER 10. A PROBABILISTIC FIRST GUESS

number of primes less than 99 should be roughly

99
= 24.75 ≈ 25,
2×2
which is pretty amazing, since the correct number of primes up to 99 is 25. The
number of primes up to 999 should be roughly
999
= 166.5 ≈ 167,
2×3
which is again close, since there are 168 primes up to 1000. The number of
primes up to 999,999 should be roughly
999999
= 83333.25 ≈ 83,333,
2×6
which is close to the correct count of 78,498.
Gauss guessed that the expected number of primes up to X is approximated by
the area under the graph of 1/ log(X) from 2 to X (see Figure 10.6). The area
under 1/ log(X) up to X = 999,999 is 78,626.43 . . ., which is remarkably close
to the correct count 78,498 of the primes up to 999,999.

1/log(x) 1/log(x) 1/log(x)

1 1 1

Area 11.977468 Area 29.080978 Area 176.564494

0.5 0.5 0.5

0 0
5 10 15 20 25 30 0 20 40 60 80 100 200 400 600 800 1000

10 Primes 25 Primes 168 Primes

Figure 10.6: The expected tally of the number of primes ≤ X is approximated

by the area underneath the graph of 1/ log(X) from 1 to X.

Gauss was an inveterate computer: he wrote in his 1849 letter that there are
216,745 prime numbers less than three million. This is wrong: the actual number
of these primes is 216,816. Gauss’s curve G(X) predicted that there would be
216,970 primes—a miss, Gauss thought, by

225 = 216970 − 216745.

But actually he was closer than he thought: the prediction of the curve G(X)
missed by a mere 154 = 216970 − 216816. Gauss’s computation brings up two
queries: will this spectacular “good fit” continue for arbitrarily large numbers?
and, the (evidently prior) question: what counts as a good fit?
Chapter 11

What is a “good
approximation”?

If you are trying to estimate a number, say, around ten thousand, and you get it
right to within a hundred, let us celebrate this kind of accuracy
√ by saying that
you have made an approximation with square-root error ( 10,000 = 100). Of
course, we should really use the more clumsy phrase “an approximation with at
worst square-root error.” Sometimes we’ll simply refer to such approximations
as good approximations. If you are trying to estimate a number in the millions,
and you get it right to within a thousand, let’s √ agree that—again—you have
made an approximation with square-root error ( 1,000,000 = 1,000). Again,
for short, call this a good approximation. So, when Gauss thought his curve
missed by 226 in estimating the number of primes less than three million, it was
well within the margin we have given for a “good approximation.”
More generally, if you are trying to estimate a number that has D digits and
you get it almost right, but with an error that has no more than, roughly, half
that many digits, let us say, again, that you have made an approximation with
square-root error or synonymously, a good approximation.
This rough account almost suffices for what we will be discussing below, but to
be more precise, the specific gauge of accuracy that will be important to us is
not for a mere single estimate of a single error term,

Error term = Exact value − Our “good approximation”

but rather for infinite sequences of estimates of error terms. Generally, if you
are interested in a numerical quantity q(X) that depends on the real number
parameter X (e.g., q(X) could be π(X), “the number of primes ≤ X”) and if
you have an explicit candidate “approximation,” qapprox (X), to this quantity, let
us say that qapprox (X) is essentially a square-root accurate approxima-
tion to q(X) if for any given exponent greater than 0.5 (you choose it: 0.501,

45
46 CHAPTER 11. WHAT IS A “GOOD APPROXIMATION”?

0.5001, 0.50001, . . . for example) and for large enough X—where the phrase
“large enough” depends on your choice of exponent—the error term—i.e., the
difference between qapprox (X) and the true quantity, q(X), is, in absolute value,
less than X raised to that exponent (e.g. < X 0.501 , < X 0.5001 , etc.). Readers
who know calculus and wish to have a technical formulation of this definition
of good approximation might turn to the endnote [7] for a precise statement.
If you found the above confusing, don’t worry: again, a square-root accurate
approximation is one in which at least roughly half the digits are correct.
Remark 11.1. To get a feel for how basic the notion of approximation to data
being square root close to the true values of the data is—and how it represents
the “gold standard” of accuracy for approximations, consider this fable.
Imagine that the devil had the idea of saddling a large committee of people
with the task of finding values of π(X) for various large numbers X. This
he did in the following manner, having already worked out which numbers are
prime himself. Since the devil is, as everyone knows, in the details, he has made
no mistakes: his work is entirely correct. He gives each committee member a
copy of the list of all prime numbers between 1 and one of the large numbers
X in which he was interested. Now each committee member would count the
number of primes by doing nothing more than considering each number, in turn,
on their list and tallying them up, much like a canvasser counting votes; the
committee members needn’t even know that these numbers are prime, they just
think of these numbers as items on their list. But since they are human, they
will indeed be making mistakes, say 1% of the time. Assume further that it is
just as likely for them to make the mistake of undercounting or overcounting.
If many people are engaged in such a pursuit, some of them might overcount
π(X); some of them might undercount it. √ The average error (overcounted or
undercounted) would be proportional to X.
In the next chapter we’ll view these undercounts and overcounts as analogous
to a random walk.
Chapter 12

Square root error and

random walks

To take a random walk along a (straight) east–west path you would start at
your home base, but every minute, say, take a step along the path, each step
being of the same length, but randomly either east or west. After X minutes,
how far are you from your home base?
The answer to this cannot be a specific number, precisely because you’re mak-
ing a random decision that affects that number for each of the X minutes of
your journey. It is more reasonable to ask a statistical version of that ques-
tion. Namely, if you took many random walks X minutes long, then—on the
average—how far would you be from your home base? The answer, as is illus-
trated by the figures below, is that the average distance you will find yourself
from home base after (sufficiently many
q of)√these excursions is proportional to
√
X. (In fact, the average is equal to π2 · X.)

To connect this with the committee members’ histories of errors, described in

the fable in Chapter 11, imagine every error (undercount or overcount by 1)
the committee member makes, as a “step” East for undercount and West for
overcount. Then if such errors were made, at a constant frequency over the
duration of time spent counting, and if the over and undercounts were equally
likely and random, then one can model the committee members’ computational
accuracy by a random walk. It would be—in the terminology we have already
discussed—no better than square-root accurate; it would be subject to square-
root error.
To get a real sense of how constrained random walks are by this “square-root
law,” here are a few numerical experiments of random walks. The left-hand
squiggly (blue) graphs in Figures 12.1–12.4 below are computer-obtained ran-
dom walk trials (three, ten, a hundred, and a thousand random walks). The

47
48 CHAPTER 12. SQUARE ROOT ERROR AND RANDOM WALKS

blue curve in the right-hand graphs of those four figures is the average distance
from home-base of the corresponding (three, ten, a hundred, and a thousand)
random
q √ walks. The red curve in each figure below is the graph of the quantity
2
π · X over the X-axis. As the number of random walks increases, the red
curve better and better approximates the average distance.

30
30
20 25

10 20

200 400 600 800 1000 15

-10
10
-20
5
-30
-40 200 400 600 800 1000

Figure 12.1: Three Random Walks

60 30
40 25
20
20
200 400 600 800 1000 15
-20
10
-40
-60 5

-80
200 400 600 800 1000

Figure 12.2: Ten Random Walks

25
60
40 20
20
15
200 400 600 800 1000
-20 10
-40
5
-60
-80 200 400 600 800 1000

Figure 12.3: One Hundred Random Walks

25
100

20
50
15

200 400 600 800 1000 10

-50 5

200 400 600 800 1000

Figure 12.4: One Thousand Random Walks

Chapter 13

What is Riemann’s
Hypothesis? (first
formulation)

Recall from Chapter 10 that a rough guess for an approximation to π(X), the
number of primes ≤ X, is given by the function X/ log(X). Recall, as well, that
a refinement of that guess, offered by Gauss, stems from this curious thought:
the “probability” that a number N is a prime is proportional to the reciprocal
of its number of digits; more precisely, the probability is 1/ log(N ). This would
lead us to guess that the approximate value of π(X) would be the area of the
region from 2 to X under the graph of 1/ log(X), a quantity sometimes referred
to as Li(X). “Li” (pronounced Li, so the same as “lie” in “lie down”) is short for
logarithmic integral, because the area of the region from 2 to X under 1/ log(X)
RX
is (by definition) the integral 2 1/ log(t)dt.

Figure 13.1 contains a graph of the three functions Li(X), π(X), and X/ log X
for X ≤ 200. But data, no matter how impressive, may be deceiving (as we
learned in Chapter 6). If you think that the three graphs never cross for all large
values of X, and that we have the simple relationship X/ log(X) < π(X) <
Li(X) for large X, read http://en.wikipedia.org/wiki/Skewes’_number.

49
50 CHAPTER 13. WHAT IS RIEMANN’S HYPOTHESIS?
50

40

30

20

10

50 100 150 200

Figure 13.1: Plots of Li(X) (top), π(X) (in the middle), and X/ log(X) (bot-
tom).

It is a major challenge to evaluate π(X) for large values of X. For example, let
X = 1024 . Then (see [8]) we have:

π(X) = 18,435,599,767,349,200,867,866
Li(X) = 18,435,599,767,366,347,775,143.10580 . . .
X/(log(X) − 1) = 18,429,088,896,563,917,716,962.93869 . . .
Li(X) − π(X) = 17,146,907,277.105803 . . .
√
X · log(X) = 55,262,042,231,857.096416 . . .

Note that several of the left-most digits of π(X) and Li(X) are the same (as
indicated in red), a point we will return to on page 57.
More fancifully, we can think of the error in this approximation to π(X), i.e.,
| Li(X) − π(X)|, (the absolute value) of the difference between Li(X) and π(X),
as (approximately) the result of a walk having roughly X steps where you move
by the following rule: go east by a distance of 1/ log N feet if N is not a prime
and west by a distance of 1 − log1 N feet if N is a prime. Your distance, then,
from home base after X steps is approximately | Li(X) − π(X)| feet.
We have no idea if this picture of things resembles a truly random walk but at
least it makes it reasonable to ask the question: is Li(X) essentially a square
root approximation to π(X)? Our first formulation of Riemann’s Hypothesis
says yes:

The Riemann Hypothesis (first formulation)

For any real number X the number of prime numbers less than X is
approximately Li(X) and this approximation is essentially square root
accurate (see [9]).
Chapter 14

The mystery moves to the

error term

Let’s think of what happens when you have a mysterious quantity (say, a func-
tion of a real number X) you wish to understand. Suppose you manage to ap-
proximate that quantity by an easy to understand expression—which we’ll call
the “dominant term”—that is also simple to compute, but only approximates
your mysterious quantity. The approximation is not exact; it has a possible
error, which happily is significantly smaller than the size of the dominant term.
“Dominant” here just means exactly that: it is of size significantly larger than
the error of approximation.

Mysterious quantity(X) = Simple, but dominant quantity(X) + Error (X).

If all you are after is a general estimate of size your job is done. You might
declare victory and ignore Error (X) as being—in size—negligible. But if you
are interested in the deep structure of your mysterious quantity perhaps all you
have done is to manage to transport most questions about it to Error (X). In
conclusion, Error (X) is now your new mysterious quantity.
Returning to the issue of π(X) (our mysterious quantity) and Li(X) (our dom-
inant term) the first formulation of the Riemann Hypothesis (as in Chapter 13
above) puts the spotlight on the Error term | Li(X) − π(X)|, which therefore
deserves our scrutiny, since—after all—we’re not interested in merely counting
the primes: we want to understand as much as we can of their structure.
To get a feel for this error term, we shall smooth it out a bit, and look at a few
of its graphs.

51
Chapter 15

Cesàro smoothing

Often when you hop in a car and reset the trip counter, the car will display
information about your trip. For instance, it might show you the average speed
up until now, a number that is “sticky”, changing much less erratically than your
actual speed, and you might use it to make a rough estimate of how long until
you will reach your destination. Your car is computing the Cesàro smoothing
of your speed. We can use this same idea to better understand the behavior of
other things, such as the sums appearing in the previous chapter.

Figure 15.1: The “Average Vehicle Speed,” as displayed on the dashboard of

the 2013 Camaro SS car that one of us drove during the writing of this chapter.

Suppose you are trying to say something intelligent about the behavior of a
certain quantity that varies with time in what seems to be a somewhat erratic,
volatile pattern. Call the quantity f (t) and think of it as a function defined
for positive real values of “time” t. The natural impulse might be to take
some sort of “average value”1 of f (t) over a time interval, say from 0 to T .
This would indeed be an intelligent thing to do if this average stabilized and
depended relatively little on the interval of time over which one is computing
this average, e.g., if that interval were large enough. Sometimes, of course, these
averages themselves are relatively sensitive to the times T that are chosen, and

1 For 1
RT
readers who know calculus: that average would be T 0 f (t)dt.

52
 53

don’t stabilize. In such a case, it usually pays to consider all these averages as
your chosen T varies, as a function (of T ) in its own right2 . This new function
F (T ), called the Cesàro Smoothing of the function f (t), is a good indicator
of certain eventual trends of the original function f (t) that is less visible directly
from the function f (t). The effect of passing from f (t) to F (T ) is to “smooth
out” some of the volatile local behavior of the original function, as can be seen
in Figure 15.2.

15

10

10 20 30 40

Figure 15.2: The red plot is the Cesàro smoothing of the blue plot

2 For readers who know calculus: this is

1
Z T
F (T ) := f (t)dt.
T 0
Chapter 16

A view of | Li(X) − π(X)|

Returning to our mysterious error term, consider Figure 16.1 where the volatile
blue curve in the middle is Li(X) − π(X), its Cesàro smoothing is the red
relatively
q psmooth curve on the bottom, and the curve on the top is the graph
of π2 · X/ log(X) over the range of X ≤ 250,000.

100

80

60

40

20

10000 100000 250000

Figure
q 16.1: Li(X) − π(X) (blue middle), its Cesàro smoothing (red bottom),
2
p
and π · X/ log(X) (top), all for X ≤ 250,000

Data such as this graph can be revealing and misleading at the same time. For
example, the volatile blue graph (of Li(X) − π(X)) seems to be roughly sand-
wiched between two rising functions, but this will not continue for all values of
X. The Cambridge University mathematician, John Edensor Littlewood (see
Figure 16.2), proved in 1914 that there exists a real number X for which the
quantity Li(X) − π(X) vanishes, and then for slightly larger X crosses over
into negative values. This theorem attracted lots of attention at the time (and

54
 55

continues to do so) because of the inaccessibility of achieving good estimates

(upper or lower bounds) for the first such number. That “first” X for which
Li(X) = π(X) is called Skewes Number in honor of the South African math-
ematician (a student of Littlewood) Stanley Skewes, who (in 1933) gave the
first—fearfully large—upper bound for that number (conditional on the Rie-
mann Hypothesis!). Despite a steady stream of subsequent improvements, we
currently can only locate Skewes Number as being in the range:

1014 ≤ Skewes Number < 10317 ,

and it is proven that at some values of X fairly close to the upper bound 10317
π(X) is greater than Li(X). So the trend suggested in Figure 16.1 will not
continue indefinitely.

Figure 16.2: John Edensor Littlewood (1885–1977)

Chapter 17

The Prime Number

Theorem

Take a look at Figure 13.1 again. All three functions, Li(X), π(X) and X/ log(X)
are “going to infinity with X” (this means that for any real number R, for all
sufficiently large X, the values of these functions at X exceed R).
Are these functions “going to infinity” at the same rate?
To answer such a question, we have to know what we mean by going to infinity
at the same rate. So, here’s a definition. Two functions, A(X) and B(X), that
each go to infinity will be said to go to infinity at the same rate if their
ratio
A(X)/B(X)
tends to 1 as X goes to infinity.
If for example two functions, A(X) and B(X) that take positive whole number
values, have the same number of digits for large X and if, for any number you
give us, say: a million (or a billion, or a trillion) and if X is large enough, then
the “leftmost” million (or billion, or trillion) digits of A(X) and B(X) are the
same, then A(X) and B(X) go to infinity at the same rate. For example,

A(X) 281067597183743525105611755423
= = 1.00000000007963213762060 . . .
B(X) 281067597161361511527766294585

While we’re defining things, let us say that two functions, A(X) and B(X), that
each go to infinity go to infinity at similar rates if there are two positive
constants c and C such that for X sufficiently large the ratio

A(X)/B(X)

is greater than c and smaller than C.

56
 57
30000 1
25000
0.8
20000
B(X) 0.6 A(X)/B(X)
15000
0.4
10000
A(X)
5000 0.2

20 40 60 80 100 20 40 60 80 100

Figure 17.1: The polynomials A(X) = 2X 2 + 3X − 5 (bottom) and B(X) =

3X 2 − 2X + 1 (top) go to infinity at similar rates.

For example, two polynomials in X with positive leading coefficients go to in-

finity at the same rate if and only if they have the same degrees and the same
leading coefficient; they go to infinity at similar rates if they have the same
degree. See Figures 17.1 and 17.2.
10000 6
8000
A(X) 5

6000 4
B(X) 3
4000
2
2000
1
20 40 60 80 100
A(X)/B(X)
0 20 40 60 80 100

Figure 17.2: The polynomials A(X) = X 2 + 3X − 5 (top) and B(X) = X 2 −

2X + 1 (bottom) go to infinity at the same rate.

Now a theorem from elementary calculus tells us that the ratio of Li(X) to
X/ log(X) tends to 1 as X gets larger and larger. That is—using the definition
we’ve just introduced—Li(X) and X/ log(X) go to infinity at the same rate (see
[10]).
Recall (on page 50 above in Chapter 13) that if X = 1024 , the left-most twelve
digits of π(X) and Li(X) are the same: both numbers start 18,435,599,767,3 . . ..
Well, that’s a good start. Can we guarantee that for X large enough, the “left-
most” million (or billion, or trillion) digits of π(X) and Li(X) are the same, i.e.,
that these two functions go to infinity at the same rate?
The Riemann Hypothesis, as we have just formulated it, would tell us that the
difference between Li(X) and π(X) is pretty small in comparison with the size of
X. This information would imply (but would be much more precise information
58 CHAPTER 17. THE PRIME NUMBER THEOREM

than) the statement that the ratio Li(X)/π(X) tends to 1, i.e., that Li(X) and
π(X) go to infinity at the same rate.

This last statement gives, of course, a far less precise relationship between Li(X)
and π(X) than the Riemann Hypothesis (once it is proved!) would give us. The
advantage, though, of the less precise statement is that it is currently known
to be true, and—in fact—has been known for over a century. It goes under the
name of

The Prime Number Theorem: Li(X) and π(X) go to infinity at the same
rate.

Since Li(X) and X/ log(X) go to infinity at the same rate, we could equally
well have expressed the “same” theorem by saying:

The Prime Number Theorem: X/ log(X) and π(X) go to infinity at the

same rate.

This fact is a very hard-won piece of mathematics! It was proved in 1896

independently by Jacques Hadamard and Charles de la Vallée Poussin.

A milestone in the history leading up to the proof of the Prime Number Theorem
is the earlier work of Pafnuty Lvovich Chebyshev (see http://en.wikipedia.
org/wiki/Chebyshev_function) showing that (to use the terminology we in-
troduced) X/ log(X) and π(X) go to infinity at similar rates.

The elusive Riemann Hypothesis, however, is much deeper than the Prime Num-
ber Theorem, and takes its origin from some awe-inspiring, difficult to inter-
pret, lines in Bernhard Riemann’s magnificent 8-page paper, “On the number
of primes less than a given magnitude,” published in 1859 (see [11]).

Figure 17.3: From a manuscript of Riemann’s 1859 paper written by a contem-

porary of Riemann (possibly the mathematician Alfred Clebsch).

Riemann’s hypothesis, as it is currently interpreted, turns up as relevant, as a

key, again and again in different parts of the subject: if you accept it as hy-
pothesis you have an immensely powerful tool at your disposal: a mathematical
magnifying glass that sharpens our focus on number theory. But it also has a
wonderful protean quality—there are many ways of formulating it, any of these
formulations being provably equivalent to any of the others.
 59

Figure 17.4: Bernhard Riemann (1826–1866)

The Riemann Hypothesis remains unproved to this day, and therefore is “only
a hypothesis,” as Osiander said of Copernicus’s theory, but one for which we
have overwhelming theoretical and numerical evidence in its support. It is the
kind of conjecture that contemporary Dutch mathematician Frans Oort might
label a suffusing conjecture in that it has unusually broad implications: many,
many results are now known to follow, if the conjecture, familiarly known as
RH, is true. A proof of RH would, therefore, fall into the applied category, given
our discussion above in Chapter 9. But however you classify RH, it is a central
concern in mathematics to find its proof (or, a counter-example!). RH is one of
the weightiest statements in all of mathematics.
Chapter 18

The information contained

in the staircase of primes

We have borrowed the phrase “staircase of primes” from the popular book The
Music of the Primes by Marcus du Sautoy, for we feel that it captures the
sense that there is a deeply hidden architecture to the graphs that compile the
number of primes (up to N ) and also because—in a bit—we will be tinkering
with this carpentry. Before we do so, though, let us review in Figure 18.1 what
this staircase looks like for different ranges.

60
 61

25
8
20
6
15
4
10
2 5

5 10 15 20 25 20 40 60 80 100
1200
150
1000
800
100
600

50 400
200

200 400 600 800 1000 2000 4000 6000 8000 10000

8000

5000

2000

20000 60000 100000

Figure 18.1: The Staircase of Primes

The mystery of this staircase is that the information contained within it is—in
effect—the full story of where the primes are placed. This story seems to elude
any simple description. Can we “tinker with” this staircase without destroying
this valuable information?
Chapter 19

Tinkering with the

carpentry of the staircase of
primes

For starters, notice that all the (vertical) risers of this staircase (Figure 18.1
above) have unit height. That is, they contain no numerical information except
for their placement on the x-axis. So, we could distort our staircase by changing
(in any way we please) the height of each riser; and as long as we haven’t brought
new risers into—or old risers out of—existence, and have not modified their
position over the x-axis, we have retained all the information of our original
staircase.
A more drastic-sounding thing we could do is to judiciously add new steps to
our staircase. At present, we have a step at each prime number p, and no step
anywhere else. Suppose we built a staircase with a new step not only at x = p
for p each prime number but also at x = 1 and x = pn where pn runs through
all powers of prime numbers as well. Such a staircase would have, indeed, many
more steps than our original staircase had, but, nevertheless, would retain much
of the quality of the old staircase: namely it contains within it the full story of
the placement of primes and their powers.
A final thing we can do is to perform a distortion of the x-axis (elongating or
shortening it, as we wish) in any specific way, as long as we can perform the
inverse process, and “undistort” it if we wish. Clearly such an operation may
have mangled the staircase, but hasn’t destroyed information irretrievably.
We shall perform all three of these kinds of operations eventually, and will see
some great surprises as a result. But for now, we will perform distortions only
of the first two types. We are about to build a new staircase that retains the
precious information we need, but is constructed according to the following

62
 63

architectural plan.

• We first build a staircase that has a new step precisely at x = 1, and

x = pn for every prime power pn with n ≥ 1. That is, there will be a new
step at x = 1, 2, 3, 4, 5, 7, 8, 9, 11, . . .

• Our staircase starts on the ground at x = 0 and the height of the riser of
the step at x = 1 will be log(2π). The height of the riser of the step at
x = pn will not be 1 (as was the height of all risers in the old staircase of
primes) but rather: the step at x = pn will have the height of its riser equal
to log p. So for the first few steps listed in the previous item, the risers
will be of height log(2π), log 2, log 3, log 2, log 5, log 7, log 2, log 3, log 11, . . .
Since log(p) > 1, these vertical dimensions lead to a steeper ascent but no
great loss of information.
Although we are not quite done with our architectural work, Figure 19.1
shows what our new staircase looks like, so far.

8 80

6 60

4 40

2 20

2 4 6 8 20 40 60 80 100

Figure 19.1: The newly constructed staircase that counts prime powers

Notice that this new staircase looks, from afar, as if it were nicely approximated
by the 45 degree straight line, i.e., by the simple function X. In fact, we have—
by this new architecture—a second equivalent way of formulating Riemann’s
hypothesis. For this, let ψ(X) denote the function of X whose graph is depicted
in Figure 19.1 (see [12]).

The Riemann Hypothesis (second formulation)

This new staircase is essentially square root close to the 45 degree
straight line; i.e., the function ψ(X) is essentially square root close to
the function f (X) = X. See Figure 19.2.
64 CHAPTER 19. TINKERING WITH THE STAIRCASE OF PRIMES

1000

800

600

400

200

200 400 600 800 1000

Figure 19.2: The newly constructed staircase is close to the 45 degree line.

Do not worry if you do not understand why our first and second formula-
tions of Riemann’s Hypothesis are equivalent. Our aim, in offering the second
formulation—a way of phrasing Riemann’s guess that mathematicians know to
be equivalent to the first one—is to celebrate the variety of equivalent ways we
have to express Riemann’s proposed answers to the question “How many primes
are there?”, and to point out that some formulations would reveal a startling
simplicity—not immediately apparent—to the behavior of prime numbers, no
matter how erratic primes initially appear to us to be. After all, what could be
simpler than a 45 degree straight line?
Chapter 20

What do computer music

files, data compression, and
prime numbers have to do
with each other?

Sounds of all sorts—and in particular the sounds of music—travel as vibrations

of air molecules at roughly 768 miles per hour. These vibrations—fluctuations of
pressure—are often represented, or “pictured,” by a graph where the horizontal
axis corresponds to time, and the vertical axis corresponds to pressure at that
time. The very purest of sounds—a single sustained note—would look something
like this (called a “sine wave”) when pictured (see Figure 20.1), so that if you
fixed your position somewhere and measured air pressure due to this sound at
that position, the peaks correspond to the times when the varying air pressure
is maximal or minimal and the zeroes to the times when it is normal pressure.

0.5

-6 -4 -2 2 4 6

-0.5

-1

Figure 20.1: Graph of a sine wave

65
66 CHAPTER 20. COMPUTER MUSIC FILES AND PRIME NUMBERS

You’ll notice that there are two features to the graph in Figure 20.1.

1. The height of the peaks of this sine wave: This height is referred to as the
amplitude and corresponds to the loudness of the sound.

2. The number of peaks per second: This number is referred to as the fre-
quency and corresponds to the pitch of the sound.

Of course, music is rarely—perhaps never—just given by a single pure sustained

note and nothing else. A next most simple example of a sound would be a simple
chord (say a C and an E played together on some electronic instrument that
could approximate pure notes). Its graph would be just the sum of the graphs
of each of the pure notes (see Figures 20.2 and 20.3).
1

0.5

5 10 15

-0.5

-1

Figure 20.2: Graph of two sine waves with different frequencies

2
1.5
1
0.5

5 10 15
-0.5
-1
-1.5

Figure 20.3: Graph of sum of the two sine waves with different frequencies

So the picture of the changing frequencies of this chord would be already a

pretty complicated configuration. What we have described in these graphs are
two sine waves (our C and our E) when they are played in phase (meaning they
start at the same time) but we could also “delay” the onset of the E note and
play them with some different phase relationship, for example, as illustrated in
Figures 20.4 and 20.5.
 67
1

Phase 0.5

5 10 15

-0.5

-1

Figure 20.4: Graph of two “sine” waves with different phases.

1.5
1
0.5

5 10 15
-0.5
-1
-1.5

Figure 20.5: Graph of the sum of the two “sine” waves with different frequencies
and phases.

So, all you need to reconstruct the chord graphed above is to know five numbers:

• the two frequencies—the collection of frequencies that make up the sound

is called the spectrum of the sound,
• the two amplitudes of each of these two frequencies,
• the phase between them.

Now suppose you came across such a sound as pictured in Figure 20.5 and
wanted to “record it.” Well, one way would be to sample the amplitude of the
sound at many different times, as for example in Figure 20.6.

1.5
1
0.5

5 10 15
-0.5
-1
-1.5

Figure 20.6: Graph of sampling of a sound wave

Then, fill in the rest of the points to obtain Figure 20.7.

68 CHAPTER 20. COMPUTER MUSIC FILES AND PRIME NUMBERS

1.5
1
0.5

5 10 15
-0.5
-1
-1.5

Figure 20.7: Graph obtained from Figure 20.6 by filling in the rest of the points

But this sampling would take an enormous amount of storage space, at least
compared to storing five numbers, as explained above! Current audio compact
discs do their sampling 44,100 times a second to get a reasonable quality of
sound.
Another way is to simply record the five numbers: the spectrum, amplitudes,
and phase. Surprisingly, this seems to be roughly the way our ear processes
such a sound when we hear it.1
Even in this simplest of examples (our pure chord: the pure note C played
simultaneously with pure note E) the efficiency of data compression that is the
immediate bonus of analyzing the picture of the chords as built just with the
five numbers giving spectrum, amplitudes, and phase is staggering.

Figure 20.8: Jean Baptiste Joseph Fourier (1768–1830)

This type of analysis, in general, is called Fourier Analysis and is one of the
glorious chapters of mathematics. One way of picturing spectrum and amplitudes
of a sound is by a bar graph which might be called the spectral picture of the
sound, the horizontal axis depicting frequency and the vertical one depicting

1 We recommend downloading Dave Benson’s marvelous book Music: A Mathematical Of-

fering from https://homepages.abdn.ac.uk/mth192/pages/html/maths-music.html. This is

free, and gives a beautiful account of the superb mechanism of hearing, and of the mathematics
of music.
 69

amplitude: the height of a bar at any frequency is proportional to the amplitude

of that frequency “in” the sound.
So our CE chord would have the spectral picture in Figure 20.9.

1.5

0.5

5 10 15
-0.5

-1
C D E

Figure 20.9: Spectral picture of a CE chord

This spectral picture ignores the phase but is nevertheless a very good portrait
of the sound. The spectral picture of a graph gets us to think of that graph as
“built up by the superposition of a bunch of pure waves,” and if the graph is
complicated enough we may very well need infinitely many pure waves to build
it up! Fourier analysis is a mathematical theory that allows us to start with any
graph—we are thinking here of graphs that picture sounds, but any graph will
do—and actually compute its spectral picture (and even keep track of phases).
The operation that starts with a graph and goes to its spectral picture that
records the frequencies, amplitudes, and phases of the pure sine waves that, to-
gether, compose the graph is called the Fourier transform and nowadays there
are very fast procedures for getting accurate Fourier transforms (meaning ac-
curate spectral pictures including information about phases) by computer [13].
The theory behind this operation (Fourier transform giving us a spectral analy-
sis of a graph) is quite beautiful, but equally impressive is how—given the power
of modern computation—you can immediately perform this operation for your-
self to get a sense of how different wave-sounds can be constructed from the
superposition of pure tones.
The sawtooth wave in Figure 20.10 has a spectral picture, its Fourier transform,
given in Figure 20.11:
1
0.8
0.6
0.4
0.2

1 2 3 4

Figure 20.10: Graph of sawtooth wave

70 CHAPTER 20. COMPUTER MUSIC FILES AND PRIME NUMBERS
1
0.8
0.6
0.4
0.2
0
5 10 15

Figure 20.11: The Spectrum of the sawtooth wave has a spike of height 1/k at
each integer k

5 10 15
-1

-2

Figure 20.12: A Complicated sound wave

Suppose you have a complicated sound wave, say as in Figure 20.12, and you
want to record it. Standard audio CDs record their data by intensive sampling
as we mentioned. In contrast, current MP3 audio compression technology uses
Fourier transforms plus sophisticated algorithms based on knowledge of which
frequencies the human ear can hear. With this, MP3 technology manages to
get a compression factor of 8–12 with little perceived loss in quality, so that you
can fit your entire music collection on your phone, instead of just a few of your
favorite CDs.
Chapter 21

The word “Spectrum”

It is worth noting how often this word appears in scientific

literature, with an array of different uses and meanings. It
comes from Latin where its meaning is “image,” or “appear-
ance,” related to the verb meaning to look (the older form
being specere and the later form spectare). In most of its
meanings, nowadays, it has to do with some procedure, an
analysis, allowing one to see clearly the constituent parts of
something to be analyzed, these constituent parts often orga-
nized in some continuous scale, such as in the discussion of
the previous chapter.
The Oxford Dictionary lists, as one of its many uses:

Used to classify something, or suggest that it can

be classified, in terms of its position on a scale
between two extreme or opposite points.

This works well for the color spectrum, as initiated by Newton (as in the figure
above, sunlight is separated by a prism into a rainbow continuum of colors):
an analysis of white light into its components. Or in mass spectrometry, where
beams of ions are separated (analyzed) according to their mass/charge ratio
and the mass spectrum is recorded on a photographic plate or film. Or in
the recording of the various component frequencies, with their corresponding
intensities of some audible phenomenon.
In mathematics the word has found its use in many different fields, the most
basic use occurring in Fourier analysis, which has as its goal the aim of either
analyzing a function f (t) as being comprised of simpler (specifically: sine and
cosine) functions, or synthesizing such a function by combining simpler functions
to produce it. The understanding here and in the previous chapter, is that an

71
72 CHAPTER 21. THE WORD “SPECTRUM”

analysis of f (t) as built up of simpler functions is meant to provide a significantly

clearer image of the constitution of f (t). If, to take a very particular example,
the simpler functions that are needed for the synthesis of f (t) are of the form
a cos(θt) (for a some real number which is the amplitude or size of the peaks of
this periodic function)—and if f (t) is given as the limit of

a1 cos(θ1 t) + a2 cos(θ2 t) + a3 cos(θ3 t) . . .

for a sequence of real numbers θ1 , θ2 , θ3 , . . . (i.e., these simpler functions are

functions of periods 2π 2π 2π
θ1 , θ2 , θ3 , . . . ) it is natural to call these θi the spectrum
of f (t). This will eventually show up in our discussion below of trigonometric
sums and the Riemann spectrum.
Chapter 22

Spectra and trigonometric

sums

As we saw in Chapter 20, a pure tone can be represented by a periodic sine

wave—a function of time f (t)—the equation of which might be:

f (t) = a · cos(b + θt).

-15 -10 -5 5 10 15
-1

-2

Figure 22.1: Plot of the periodic sine wave f (t) = 2 · cos(1 + t/2)

The θ determines the frequency of the periodic wave, the larger θ is the higher
the “pitch.” The coefficient a determines the envelope of size of the periodic
wave, and we call it the amplitude of the periodic wave.
Sometimes we encounter functions F (t) that are not pure tones, but that can
be expressed as (or we might say “decomposed into”) a finite sum of pure tones,
for example three of them:

F (t) = a1 · cos(b1 + θ1 t) + a2 · cos(b2 + θ2 t) + a3 · cos(b3 + θ3 t)

73
74 CHAPTER 22. SPECTRA AND TRIGONOMETRIC SUMS
10
8
6
4
2
-15 -10 -5 5 10 15
-2
-4

Figure 22.2: Plot of the sum 5 cos (−t − 2) + 2 cos (t/2 + 1) + 3 cos (2t + 4)

We refer to such functions F (t) as in Figure 22.2 as finite trigonometric sums,

because—well—they are. In this example, there are three frequencies involved—
i.e., θ1 , θ2 , θ3 —and we say that the spectrum of F (t) is the set of these frequen-
cies, i.e.,

The spectrum of F (t) = {θ1 , θ2 , θ3 }.

More generally we might consider a sum of any finite number of pure cosine
waves—or in a moment we’ll also see some infinite ones as well. Again, for
these more general trigonometric sums, their spectrum will denote the set of
frequencies that compose them.
Chapter 23

The spectrum and the

staircase of primes

25
20
15
10
5

20 40 60 80 100

Figure 23.1: The Staircase of primes

In view of the amazing data-compression virtues of Fourier analysis, it isn’t

unnatural to ask these questions:

• Is there a way of using Fourier analysis to better understand the compli-

cated picture of the staircase of primes?

• Does this staircase of primes (or, perhaps, some tinkered version of the
staircase that contains the same basic information) have a spectrum?

• If such a spectrum exists, can we compute it conveniently, just as we have

done for the saw-tooth wave above, or for the major third CE chord?

• Assuming the spectrum exists, and is computable, will our understand-

ing of this spectrum allow us to reproduce all the pertinent information
about the placement of primes among all whole numbers, elegantly and
faithfully?

75
76 CHAPTER 23. THE SPECTRUM AND THE STAIRCASE OF PRIMES

• And here is a most important question: will that spectrum show us order
and organization lurking within the staircase that we would otherwise be
blind to?

Strangely enough, it is towards questions like these that Riemann’s Hypothesis

takes us. We began with the simple question about primes: how to count them,
and are led to ask for profound, and hidden, regularities in structure.
Chapter 24

To our readers of Part I

The statement of the Riemann Hypothesis—admittedly as elusive as before—

has, at least, been expressed elegantly and more simply, given our new staircase
that approximates (conjecturally with essential square root accuracy) a 45 degree
straight line.
We have offered two equivalent formulations of the Riemann Hypothesis, both
having to do with the manner in which the prime numbers are situated among
all whole numbers.
In doing this, we hope that we have convinced you that—in the words of Don
Zagier—primes seem to obey no other law than that of chance and yet exhibit
stunning regularity. This is the end of Part I of our book, and is largely the
end of our main mission, to explain—in elementary terms—what is Riemann’s
Hypothesis?
For readers who have at some point studied differential calculus, in Part II we
shall discuss Fourier analysis, a fundamental tool that will be used in Part III
where we show how Riemann’s hypothesis provides a key to some deeper struc-
ture of the prime numbers, and to the nature of the laws that they obey. We
will—if not explain—at least hint at how the above series of questions have been
answered so far, and how the Riemann Hypothesis offers a surprise for the last
question in this series.

77
 Part II

Distributions

78
Chapter 25

How Calculus manages to

find the slopes of graphs
that have no slopes

Differential calculus, initially the creation of Newton and Leibniz in the 1680s,
acquaints us with slopes of graphs of functions of a real variable. So, to discuss
this we should say a word about what a function is, and what its graph is.

Figure 25.1: Isaac Newton and Gottfried Leibniz created calculus

A function (let us refer to it in this discussion as f ) is often described as a kind

of machine that for any specific input numerical value a will give, as output, a
well-defined numerical value.
This “output number” is denoted f (a) and is called the “value” of the function
f at a. For example, the machine that adds 1 to any number can be thought of
as the function f whose value at any a is given by the equation f (a) = a + 1.
Often we choose a letter—say, X—to stand for a “general number” and we

79
80 CHAPTER 25. SLOPES OF GRAPHS THAT HAVE NO SLOPES

denote the function f by the symbol f (X) so that this symbolization allows to
“substitute for X any specific number a” to get its value f (a).
The graph of a function provides a vivid visual representation of the function in
the Euclidean plane where over every point a on the x-axis you plot a point above
it of “height” equal to the value of the function at a, i.e., f (a). In Cartesian
coordinates, then, you are plotting points (a, f (a)) in the plane where a runs
through all real numbers.
8
6
4
2
0
-2
-4
-4 -2 0 2 4 6 8

Figure 25.2: Graph of the function f (a) = a + 1

In this book we will very often be talking about “graphs” when we are also
specifically interested in the functions—of which they are the graphs. We will
use these words almost synonymously since we like to adopt a very visual atti-
tude towards the behavior of the functions that interest us.

What is the slope of the tangent line?

2

1.5

1
Here it is!
0.5

0
How to compute the slope? This is Calculus.
-0.5
0 1 2 3 4 5 6 7 8

Figure 25.3: Calculus

Figure 25.3 illustrates a function (blue), the slope at a point (green straight
line), and the derivative (red) of the function; the red derivative is the function
whose value at a point is the slope of the blue function at that point. Differential
calculus explains to us how to calculate slopes of graphs, and finally, shows us
the power that we then have to answer problems we could not answer if we
couldn’t compute those slopes.
 81

Usually, in elementary calculus classes we are called upon to compute slopes

only of smooth graphs, i.e., graphs that actually have slopes at each of their
points, such as in the illustration just above. What could calculus possibly do
if confronted with a graph that has jumps, such as in Figure 25.4:

(
1 x≤3
f (x) =
2 x > 3?

(Note that for purely aesthetic reasons, we draw a vertical line at the point
where the jump occurs, though technically that vertical line is not part of the
graph of the function.)

1.5

0.5

1 2 3 4 5 6

Figure 25.4: The graph of the function f (x) above that jumps—it is 1 up to 3
and then 2 after that point.

The most comfortable way to deal with the graph of such a function is to just
approximate it by a differentiable function as in Figure 25.5.

1.8

1.6

1.4

1.2

1
0 1 2 3 4 5 6

Figure 25.5: A picture of a smooth graph approximating the graph that is 1 up

to some point x and then 2 after that point, the smooth graph being flat mostly.

Then take the derivative of that smooth function. Of course, this is just an
approximation, so we might try to make a better approximation, which we do
in each successive graph starting with Figure 25.6 below.
82 CHAPTER 25. SLOPES OF GRAPHS THAT HAVE NO SLOPES
2

1.5

0.5

1 2 3 4 5 6

Figure 25.6: A picture of the derivative of a smooth approximation to a function

that jumps.

Note that—as you would expect—in the range where the initial function is
constant, its derivative is zero. In the subsequent figures, our initial function
will be nonconstant for smaller and smaller intervals about the origin. Note also
that, in our series of pictures below, we will be successively rescaling the y-axis;
all our initial functions have the value 1 for “large” negative numbers and the
value 2 for large positive numbers.

3.5
3
2.5
2
1.5
1
0.5

1 2 3 4 5 6

Figure 25.7: Second picture of the derivative of a smooth approximation to a

function that jumps.

14
12
10
8
6
4
2

1 2 3 4 5 6

Figure 25.8: Third picture of the derivative of a smooth approximation to a

function that jumps.
 83

70
60
50
40
30
20
10

1 2 3 4 5 6

Figure 25.9: Fourth picture of the derivative of a smooth approximation to a

function that jumps.

Notice what is happening: as the approximation gets better and better, the
derivative will be zero mostly, with a blip at the point of discontinuity, and the
blip will get higher and higher. In each of these pictures, for any interval of real
numbers [a, b] the total area under the red graph over that interval is equal to

the height of the blue graph at x = b

minus
the height of the blue graph at x = a.

This is a manifestation of one of the fundamental facts of life of calculus relating

a function to its derivative:

Given any real-valued function F (x)—that has a derivative—for any

interval of real numbers [a, b] the total signed1 area between the
graph and the horizontal axis of the derivative of F (x) over that
interval is equal to F (b) − F (a).

What happens if we take the series of figures 25.6–25.9, etc., to the limit? This
is quite curious:

• the series of red graphs: these are getting thinner and thinner and
higher and higher: can we make any sense of what the red graph might
mean in the limit (even though the only picture of it that we have at
present makes it infinitely thin and infinitely high)?

• the series of blue graphs: these are happily looking more and more
like the tame Figure 25.4.

Each of our red graphs is the derivative of the corresponding blue graph. It is
tempting to think of the limit of the red graphs—whatever we might construe
this to be—as standing for the derivative of the limit of the blue graphs, i.e., of
the graph in Figure 25.4.

1 When F (x) < 0 we count that area as negative.

84 CHAPTER 25. SLOPES OF GRAPHS THAT HAVE NO SLOPES

Well, the temptation is so great that, in fact, mathematicians and physicists of

the early twentieth century struggled to give a meaning to things like the limit of
the red graphs—such things were initially called generalized functions, which
might be considered the derivative of the limit of the blue graphs, i.e., of the
graph of Figure 25.4.
Of course, to achieve progress in mathematics, all the concepts that play a role
in the theory have to be unambiguously defined, and it took a while before
generalized functions such as the limit of our series of red graphs had been
rigorously introduced.
But many of the great moments in the development of mathematics occur when
mathematicians—requiring some concept not yet formalized—work with the
concept tentatively, dismissing—if need be—mental tortures, in hopes that the
experience they acquire by working with the concept will eventually help to put
that concept on sure footing. For example, early mathematicians (Newton, Leib-
niz)—in replacing approximate speeds by instantaneous velocities by passing to
limits—had to wait a while before later mathematicians (e.g., Weierstrass) gave
a rigorous foundation for what they were doing.

Figure 25.10: Karl Weierstrass (1815–1897) and Laurent Schwartz (1915–2002)

Karl Weierstrass, who worked during the latter part of the nineteenth century,
was known as the “father of modern analysis.” He oversaw one of the glorious
moments of rigorization of concepts that were long in use, but never before
systematically organized. He, and other analysts of the time, were interested
in providing a rigorous language to talk about functions, and more specifically
continuous functions and smooth (i.e., differentiable) functions. They wished
to have a firm understanding of limits (i.e., of sequences of numbers, or of
functions).
For Weierstrass and his companions, even though the functions they worked with
needn’t be smooth, or continuous, at the very least, the functions they studied
 85

had well-defined—and usually finite—values. But our “limit of red graphs” is

not so easily formalized as the concepts that occupied the efforts of Weierstrass.
Happily however, this general process of approximating discontinuous functions
more and more exactly by smooth functions, and taking their derivatives to get
the blip-functions as we have just seen in the red graphs above was eventually
given a mathematically rigorous foundation; notably, by the French mathemati-
cian, Laurent Schwartz who provided a beautiful theory that we will not go
into here, that made perfect sense of “generalized functions” such as our limit
of the series of red graphs, and that allows mathematicians to work with these
concepts with ease. These “generalized functions” are called distributions in
Schwartz’s theory [14].
Chapter 26

Distributions: sharpening
our approximating
functions even if we have to
let them shoot out to
infinity

The curious limit of the red graphs of the previous section, which you might be
tempted to think of as a “blip-function” f (t) that vanishes for t nonzero and is
somehow “infinite” (whatever that means) at 0 is an example of a generalized
function (in the sense of the earlier mathematicians) or a distribution in the
sense of Laurent Schwartz.

This particular limit of the red graphs also goes by another name (it is officially
called a Dirac δ-function (see [15]), the adjective “Dirac” being in honor of
the physicist who first worked with this concept, the “δ” being the symbol
he assigned to these objects). The noun “function” should be in quotation
marks for, properly speaking, the Dirac δ-function is not—as we have explained
above—a bona fide function but rather a distribution.

86
 87

Figure 26.1: Paul Adrien Maurice Dirac (1902–1984)

Now may be a good time to summarize what the major difference is between
honest functions and generalized functions or distributions.
An honest (by which we mean integrable) function of a real variable f (t) pos-
sesses two “features.”

• It has values. That is, at any real number t, e.g., t = 2 or t = 0 or t = π

etc., our function has a definite real number value (f (2) or f (0) or f (π)
etc.) and if we know all those values we know the function.
• It has areas under its graph. If we are given any interval of real
numbers, say the interval between a and b, we can talk unambiguously
about the area “under” the graph of the function f (t) over the interval
between a and b. That is, in the terminology of integral calculus, we can
talk of the integral of f (t) from a to b. And in the notation of calculus,
this—thanks to Leibniz—is elegantly denoted
Z b
f (t)dt.
a

0.8

R ∞ 0.6

−∞ f (x)dx 0.4

0.2

-10 -5 5 10

R∞
Figure 26.2: This figure illustrates −∞ f (x)dx, which is the signed area between
the graph of f (x) and the x-axis, where area below the x-axis (yellow) counts
negative, and area above (grey) is positive.

In contrast, a generalized function or distribution

88 CHAPTER 26. DISTRIBUTIONS

• may not have “definite values” at all real numbers if it is not an

honest function. Nevertheless,
• It has well-defined areas under portions of its “graph.” If we are
given any interval of real numbers, say the (open) interval between a and
b, we can still talk unambiguously about the area “under” the graph of the
generalized function D(t) over the interval between a and b and we will
denote this–extending what we do in ordinary calculus—by the symbol
Z b
D(t)dt.
a

This description is important to bear in mind and it gives us a handy way of

thinking about “generalized functions” (i.e., distributions) as opposed to func-
tions: when we consider an (integrable) function of a real variable, f (t), we may
invoke its value at every real number and for every interval [a, b] we may consider
Rb
the quantity a f (t)dt. BUT when we are given a generalized function D(t) we
only have at our disposal the latter quantities. In fact, a generalized function of
a real variable D(t) is (formally) nothing more than a rule that assigns to any
Rb
finite interval (a, b] (a ≤ b) a quantity that we might denote a D(t)dt and that
behaves as if it were the integral of a function and in particular—for three real
numbers a ≤ b ≤ c we have the additivity relation
Z c Z b Z c
D(t)dt = D(t)dt + D(t)dt.
a a b

SO, any honest function integrable over finite intervals clearly is a distribution
(forget about its values!) but . . . there are many more generalized functions,
and including them in our sights gives us a very important tool.
It is natural to talk, as well, of Cauchy sequences, and limits, of distributions.
We’ll say that such a sequence D1 (t), D2 (t), D3 (t), . . . is a Cauchy sequence
if for every interval [a, b] the quantities
Z b Z b Z b
D1 (t)dt, D2 (t)dt, D3 (t)dt, . . .
a a a

form a Cauchy sequence of real numbers (so for any ε > 0 eventually all terms in
the sequence of real numbers are within ε of each other). Now, any Cauchy se-
quence of distributions converges to a limiting distribution D(t) which is defined
by the rule that for every interval [a, b],

Z b Z b
D(t)dt = lim Di (t)dt.
a i→∞ a

If, by the way, you have an infinite sequence—say—of honest, continuous, func-
tions that converges uniformly to a limit (which will again be a continuous
 89

function) then that sequence certainly converges—in the above sense—to the
same limit when these functions are viewed as generalized functions. BUT,
there are many important occasions where your sequence of honest continuous
functions doesn’t have that convergence property and yet when these contin-
uous functions are viewed as generalized functions they do converge to some
generalized function as a limit. We will see this soon when we get back to the
“sequence of the red graphs.” This sequence does converge (in the above sense)
to the Dirac δ-function when these red graphs are thought of as a sequence of
generalized functions.
The integral notation for distribution is very useful, and allows us the flexibility
to define, for nice enough—and honest—functions c(t) useful expressions such
as Z b
c(t)D(t).
a

Figure 26.3: The Dirac δ-“function” (actually distribution), where we draw a

vertical arrow to illustrate the delta function with support at a given point.

For example, the Dirac δ-function we have been discussing (i.e., the limit of
the red graphs of Chapter 25) is an honest function away from t = 3 and—in
fact—is the “trivial function” zero away from 3. And at 3, we may say that it
has the “value” infinity, in honor of it being the limit of blip functions getting
taller and taller at 3. The feature that pins it down as a distribution is given
by its behavior relative to the second feature above, the area of its graph over
the open interval (a, b) between a and b:

• If 3 is not in the open interval spanned by a and b, then the “area under
the graph of our Dirac δ-function” over the interval (a, b) is 0.
• If 3 is in the open interval (a, b), then the “area under the graph of our
Dirac δ-function” is 1—in notation
Z b
δ = 1.
a

We sometimes summarize the fact that these areas vanish so long as 3 is not
included in the interval we are considering by saying that the support of this
δ-function is “at 3.”
90 CHAPTER 26. DISTRIBUTIONS

Once you’re happy with this Dirac δ-function, you’ll also be happy with a Dirac
δ-function—call it δx —with support concentrated at any specific real number x.
This δx vanishes for t 6= x and intuitively speaking, has an infinite blip at t = x.
So, the original delta-function we were discussing, i.e., δ(t) would be denoted
δ3 (t).
A question: If you’ve never seen distributions before, but know the Riemann
Rb
integral, can you guess at what the definition of a c(t)D(t) is, and can you
formulate hypotheses on c(t) that would allow you to endow this expression
with a definite meaning?
A second question: If you have not seen distributions before, and have an-
swered the first question above, let c(t) be an honest function for which your
definition of
Z b
c(t)D(t)
a

applies. Now let x be a real number. Can you use your definition to compute

Z +∞
c(t)δx (t)?
−∞

R +∞
The answer to this second question, by the way, is: −∞
c(t)δx (t) = c(x). This
will be useful in the later sections!
The theory of distributions gives a partial answer to the following funny ques-
tion:

How in the world can you “take the derivative” of a function F (t)
that doesn’t have a derivative?

The short answer to this question is that this derivative F 0 (t) which doesn’t
exist as a function may exist as a distribution. What then is the integral of that
distribution? Well, it is given by the original function!

Z b
F 0 (t)dt = F (b) − F (a).
a

Let us practice this with simple staircase functions. For example, what is the
derivative—in the sense of the theory of distributions—of the function in Fig-
ure 26.4? Answer: δ0 + 2δ1 .
 91
3
2.5
2
1.5
1
0.5

-1 -0.5 0.5 1 1.5 2

Figure 26.4: The staircase function that is 0 for t ≤ 0, 1 for 0 < t ≤ 1 and 3 for
1 < t ≤ 2 has derivative δ0 + 2δ1 .

We’ll be dealing with much more complicated staircase functions in the next
chapter, but the general principles discussed here will nicely apply there [16].
Chapter 27

Fourier transforms: second

visit

In Chapter 20 above we wrote:

The operation that starts with a graph and goes to its spectral pic-
ture that records the frequencies, amplitudes, and phases of the pure
sine waves that, together, compose the graph is called the Fourier
transform.

Now let’s take a closer look at this operation Fourier transform.

We will focus our discussion on an even function f (t) of a real variable t.
“Even” means that its graph is symmetric about the y-axis; that is, f (−t) =
f (t). See Figure 27.1.
2.5
2
1.5
1
0.5

-4 -3 -2 -1 1 2 3 4

Figure 27.1: The graph of an even function is symmetrical about the y-axis.

When we get to apply this discussion to the staircase of primes π(t) or the
tinkered staircase of primes ψ(t), both of which being defined only for positive
values of t, then we would “lose little information” in our quest to understand
them if we simply “symmetrized their graphs” by defining their values on neg-
ative numbers −t via the formulas π(−t) = π(t) and ψ(−t) = ψ(t) thereby
turning each of them into even functions.

92
 93
14
12
10
8
6
4
2
-40 -20 20 40

Figure 27.2: Even extension of the staircase of primes.

The idea behind the Fourier transform is to express f (t) as made up out of sine
and cosine wave functions. Since we have agreed to consider only even func-
tions, we can dispense with the sine waves—they won’t appear in our Fourier
analysis—and ask how to reconstruct f (t) as a sum (with coefficients) of cosine
functions (if only finitely many frequencies occur in the spectrum of our func-
tion) or more generally, as an integral if the spectrum is more elaborate. For
this work, we need a little machine that tells us, for each real number θ, whether
or not θ is in the spectrum of f (t), and if so, what the amplitude is of the cosine
function cos(θt) that occurs in the Fourier expansion of f (t)—this amplitude
answers the awkwardly phrased question:
How much cos(θt) “occurs in” f (t)?
We will denote this amplitude by fˆ(θ), and refer to it as the Fourier transform
of f (t). The spectrum, then, of f (t) is the set of all frequencies θ where the
amplitude is nonzero.

f (t) fˆ(θ)

Figure 27.3: The Fourier transform machine, which transforms f (t) into fˆ(θ)
R +∞
Now in certain easy circumstances—specifically, if −∞ |f (t)|dt (exists, and) is
finite—integral calculus provides us with an easy construction of that machine
(see Figure 27.3); namely:

Z +∞
fˆ(θ) = f (t) cos(−θt)dt.
−∞

This concise machine manages to “pick out” just the part of f (t) that has
frequency θ! It provides for us the analysis part of the Fourier analysis of our
function f (t).
94 CHAPTER 27. FOURIER TRANSFORMS: SECOND VISIT

But there is a synthesis part to our work as well, for we can reconstruct f (t)
from its Fourier
R +∞ transform, by a process intriguingly similar to the analysis part;
namely: if −∞ |fˆ(θ)|dθ (exists, and) is finite, we retrieve f (t) by the integral

Z +∞
1
f (t) = fˆ(θ) cos(θt)dθ.
2π −∞

R +∞
We are not so lucky to have −∞ |f (t)|dt finite when we try our hand at a
Fourier analysis of the staircase of primes, but we’ll work around this!
Chapter 28

What is the Fourier

transform of a delta
function?

Consider the δ-function that we denoted δ(t) (or δ0 (t)). This is also the “general-
ized function” that we thought of as the “limit of the red graphs” in Chapter 26
above. Even though δ(t) is a distribution and not a bona fide function, it is
symmetric about the origin, and also
Z +∞
|δ(t)|dt
−∞

exists, and is finite (its value is, in fact, 1). All this means that, appropriately
understood, the discussion of the previous section applies, and we can feed
this delta-function into our Fourier Transform Machine (Figure 27.3) to see
what frequencies and amplitudes arise in our attempt to express—whatever
this means!—the delta-function as a sum, or an integral, of cosine functions.
So what is the Fourier transform, δ̂0 (θ), of the delta-function?
Well, the general formula would give us:
Z +∞
δ̂0 (θ) = cos(−θt)δ0 (t)dt.
−∞

For any nice function c(t) we have that the integral of the product of c(t) by
the distribution δx (t) is given by the value of the function c(t) at t = x. So:

Z +∞
δ̂0 (θ) = cos(−θt)δ0 (t)dt = cos(0) = 1.
−∞

95
96 CHAPTER 28. FOURIER TRANSFORM OF DELTA

In other words, the Fourier transform of δ0 (t) is the constant function

δ̂0 (θ) = 1.

One can think of this colloquially as saying that the delta-function is a perfect
example of white noise in that every frequency occurs in its Fourier analysis and
they all occur in equal amounts.
To generalize this computation, let us consider for any real number x the sym-
metrized delta-function with support at x and −x, given by

dx (t) = (δx (t) + δ−x (t))/2

in Figure 28.1.

−x x

Figure 28.1: The sum (δx (t) + δ−x (t))/2, where we draw vertical arrows to
illustrate the Dirac delta functions.

What is the Fourier transform of this dx (t)? The answer is given by making the
same computation as we’ve just made:

Z +∞ Z +∞ 
1
dˆx (θ) = cos(−θt)δx (t)dt + cos(−θt)δ−x (t)dt
2 −∞ −∞
1

= cos(−θx) + cos(+θx)
2
= cos(xθ)

To summarize this in ridiculous (!) colloquial terms: for any frequency θ the
amount of cos(θt) you need to build up the generalized function (δx (t)+δ−x (t))/2
is cos(xθ).
So far, so good, but remember that the theory of the Fourier transform has—
like much of mathematics—two parts: an analysis part and a synthesis part.
We’ve just performed the analysis part of the theory for these symmetrized
delta functions (δx (t) + δ−x (t))/2.
Can we synthesize them—i.e., build them up again—from their Fourier trans-
forms?
We’ll leave this, at least for now, as a question for you.
Chapter 29

Trigonometric series

Given our interest in the ideas of Fourier, it is not surprising that we’ll want to
deal with things like
∞
X
F (θ) = ak cos(sk · θ)
k=1

where the sk are real numbers tending (strictly monotonically) to infinity. These
we’ll just call trigonometric series without asking whether they converge in
any sense for all values of θ, or even for any value of θ. The sk ’s that occur
in such a trigonometric series we will call the spectral values or for short,
the spectrum of the series, and the ak ’s the (corresponding) amplitudes. We
repeat that we impose no convergence requirements at all. But we also think of
these things as providing “cutoff” finite trigonometric sums, which we think of
as functions of two variables, θ and C (the “cutoff”) where
X
F (θ, C) := ak cos(sk · θ).
sk ≤C

These functions F (θ, C) are finite trigonometric series and therefore “honest
functions” having finite values everywhere.
Recall, as in Chapter 28, that for any real number x, we considered the sym-
metrized delta-function with support at x and −x, given by

dx (t) = (δx (t) + δ−x (t))/2,

and noted that the Fourier transform of this dx (t) is

dˆx (θ) = cos(xθ).

97
98 CHAPTER 29. TRIGONOMETRIC SERIES

−x x

Figure 29.1: The sum (δx (t) + δ−x (t))/2, where we draw vertical arrows to
illustrate the Dirac delta functions.

It follows, of course, that a cutoff finite trigonometric series, F (θ, C) associated

to an infinite trigonometric series
∞
X
F (θ) = ak cos(sk · θ)
k=1

is the Fourier transform of the distribution

X
D(t, C) := ak dsk (t).
sk ≤C

Given the discreteness of the set of spectral values sk (k = 1, 2, . . . ) we can

consider the infinite sum
∞
X
D(t) := ak dsk (t),
k=1

viewed as distribution playing the role of the “inverse Fourier transform” of our
trigonometric series F (t).
Definition 29.1. Say that a trigonometric series F (θ) has a spike at the real
number θ = τ ∈ R if the set of absolute values |F (τ, C)| as C ranges through
positive number cutoffs is unbounded. A real number τ ∈ R is, in contrast, a
non-spike if those values admit a finite upper bound.

In the chapters that follow we will be exhibiting graphs of trigonometric func-

tions, cutoff at various values of C, that (in our opinion) strongly hint that
as C goes to infinity, one gets convergence to certain (discrete) sequences of
very interesting spike values. To be sure, no finite computation, exhibited by
a graph, can prove that this is in fact the case. But on the one hand, the
vividness of those spikes is in itself worth experiencing, and on the other hand,
given RH, there is justification that the strong hints are not misleading; for some
theoretical background see the endnotes.
Chapter 30

A sneak preview of Part III

In this chapter, as a striking illustration of the type of phenomena that will

be studied in Part III, we will consider two infinite trigonometric sums—that
seem to be related one to the other in that the frequencies of the terms in the
one trigonometric sum give the spike values of the other, and vice versa: the
frequencies of the other give the spike values of the one: a kind of duality as in
the theory of Fourier transforms. We show this duality by exhibiting the graphs
of more and more accurate finite approximations (cutoffs) of these infinite sums.
More specifically,

1. The first infinite trigonometric sum F (t) is a sum1 of pure cosine waves
with frequencies given by logarithms of powers of primes and with ampli-
tudes given by the formula
X log(p)
F (t) := − cos(t log(pn ))
pn
pn/2

the summation being over all powers pn of all prime numbers p.

The graphs of longer and longer finite truncations of these trigonometric
sums, as you will see, have “higher and higher peaks” concentrated more
and more accurately at a certain infinite discrete set of real numbers that
we will be referring to as the Riemann spectrum, indicated in our
pictures below (Figures 30.2–30.5) by the series of vertical red lines.

2. In contrast, the second infinite trigonometric sum H(t) is a sum of pure

cosine waves with frequencies given by what we have dubbed above the
Riemann spectrum and with amplitudes all equal to 1.

1 Here we make use of the Greek symbol

P
as a shorthand way of expressing a sum of
many terms. We are not requesting this sum to converge.

99
100 CHAPTER 30. A SNEAK PREVIEW OF PART III

X
H(s) := 1 + cos(θ log(s)).
θ

These graphs will have “higher and higher peaks” concentrated more and
more accurately at the logarithms of powers of primes indicated in
our pictures below (see Figure 30.6) by the series of vertical blue spikes.

That the series of blue lines (i.e., the logarithms of powers of primes) in our pic-
tures below determines—via the trigonometric sums we describe—the series of
red lines (i.e., what we are calling the spectrum) and conversely is a consequence
of the Riemann Hypothesis.

1. Viewing the Riemann spectrum as the spike values of a trigono-

metric series with frequencies equal to (logs of ) powers of the
primes:
To get warmed up, let’s plot the positive values of the following sum of
(co-)sine waves:

log(2) log(3)
f (t) = − cos(t log(2)) − 1/2 cos(t log(3))
21/2 3
log(2) log(5)
− 1/2 cos(t log(4)) − 1/2 cos(t log(5))
4 5

1.5

0.5

20 40 60 80 100

Figure 30.1: Plot of f (t)

Look at the peaks of Figure 30.1. There is nothing very impressive about
them, you might think; but wait, for f (t) is just a very “early” piece of
the infinite trigonometric sum F (t) described above.
Let us truncate the infinite sum F (t) taking only finitely many terms, by
choosing various “cutoff values” C and forming the finite sums
X log(p)
F≤C (t) := − n/2
cos(t log(pn ))
pn ≤C
p
 101

and plotting their positive values. Figures 30.2–30.5 show what we get for
a few values of C.

In each of the graphs, we have indicated by red vertical arrows the real
numbers that give the values of the Riemann spectrum that we will be
discussing. These numbers at the red vertical arrows in Figures 30.2–30.5,

θ1 , θ2 , θ3 , . . .

are spike values—as described in Chapter 29—of the infinite trigonometric

series
X log(p)
− n/2
cos(t log(pn )).
n
p <C
p

They constitute what we are calling the Riemann spectrum and are key
to the staircase of primes [17].

• The sum with pn ≤ C = 5

In Figure 30.2 we plot the function f (t) displayed above; it consists
of the sum of the first four terms of our infinite sum, and doesn’t yet
show very much “structure”:

1.5

0.5

0
0 20 40 60 80 100

P log(p)
Figure 30.2: Plot of − pn ≤5 pn/2 cos(t log(pn )) with arrows pointing to the
spectrum of the primes

• The sum with pn ≤ C = 20

Something, (don’t you agree?) is already beginning to happen in the
graph in Figure 30.3:
102 CHAPTER 30. A SNEAK PREVIEW OF PART III

3
2.5
2
1.5
1
0.5
0
0 20 40 60 80 100

P log(p)
Figure 30.3: Plot of − pn ≤20 pn/2
cos(t log(pn )) with arrows pointing to the
spectrum of the primes

• The sum with pn ≤ C = 50

Note that the high peaks in Figure 30.4 seem to be lining up more
accurately with the vertical red lines. Note also that the y-axis has
been rescaled.

0
0 20 40 60 80 100

P log(p)
Figure 30.4: Plot of − pn ≤50 pn/2
cos(t log(pn )) with arrows pointing to the
spectrum of the primes

• The sum with pn ≤ C = 500

Here, the peaks are even sharper, and note that again they are higher;
that is, we have rescaled the y-axis.
 103

8
7
6
5
4
3
2
1
0 20 40 60 80 100

P log(p)
Figure 30.5: Plot of − pn ≤500 pn/2
cos(t log(pn )) with arrows pointing to the
spectrum of the primes
We will pay attention to:

• how the spikes “play out” as we take the sums of longer and longer
pieces of the infinite sum of cosine waves above, given by larger and
larger cutoffs C,
• how this spectrum of red lines more closely matches the high peaks
of the graphs of the positive values of these finite sums,
• how these peaks are climbing higher and higher,
• what relationship these peaks have to the Fourier analysis of the
staircase of primes,
• and, equally importantly, what these mysterious red lines signify.

2. Toward (logs of ) powers of the primes, starting from the Rie-

mann spectrum:
Here we will be making use of the series of numbers

θ1 , θ2 , θ3 , . . .

comprising what we called the spectrum. We consider the infinite trigono-

metric series

G(t) := 1 + cos(θ1 t) + cos(θ2 t) + cos(θ3 t) + . . .

P
or, using the notation,
X
G(t) := 1 + cos(θt)
θ

where the summation is over the spectrum, θ = θ1 , θ2 , θ3 , . . . . Again

we will consider finite cutoffs C of this infinite trigonometric sum (on a
logarithmic scale),
X
H≤C (s) := 1 + cos(log(s)θi )
i≤C
104 CHAPTER 30. A SNEAK PREVIEW OF PART III

and to see the spikes in H≤1000 (s) consider Figure 30.6.

2 3 4 5 7 8 9 11 13 1617 19 23 25 27 29

P1000
Figure 30.6: Illustration of − i=1 cos(log(s)θi ), where θ1 ∼ 14.13, . . . are the
first 1000 contributions to the spectrum. The red dots are at the prime powers
pn , whose size is proportional to log(p).

This passage—thanks to the Riemann Hypothesis—from spectrum to prime

powers and back again via consideration of the “high peaks” in the graphs of
the appropriate trigonometric sums provides a kind of visual duality emphasiz-
ing, for us, that the information inherent in the wild spacing of prime powers, is
somehow “packaged” in the Riemann spectrum, and reciprocally, the informa-
tion given in that series of mysterious numbers is obtainable from the sequence
of prime powers.
 Part III

The Riemann Spectrum of

the Prime Numbers

105
Chapter 31

On losing no information

To manage to repackage the “same” data in various ways—where each way

brings out some features that would be kept in the shadows if the data were
packaged in some different way—is a high art, in mathematics. In a sense
every mathematical equation does this, for the “equal sign” in the middle of
the equation tells us that even though the two sides of the equation may seem
different, or have different shapes, they are nonetheless “the same data.” For
example, the equation

log(XY ) = log(X) + log(Y )

which we encountered earlier in Chapter 10, is just two ways of looking at

the same thing, yet it was the basis for much manual calculation for several
centuries.

Now, the problem we have been concentrating on, in this book, has been—in
effect—to understand the pattern, if we can call it that, given by the placement
of prime numbers among the natural line-up of all whole numbers.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

Figure 31.1: Prime numbers up to 37

There are, of course, many ways for us to present this basic pattern. Our initial
strategy was to focus attention on the staircase of primes which gives us a vivid
portrait, if you wish, of the order of appearance of primes among all numbers.

106
 107

12
10
8
6
4
2

5 10 15 20 25 30 35

Figure 31.2: Prime numbers up to 37

As we have already hinted in the previous sections, however, there are various
ways open to us to tinker with—and significantly modify—our staircase without
losing the essential information it contains. Of course, there is always the danger
of modifying things in such a way that “retrieval” of the original data becomes
difficult. Moreover, we had better remember every change we have made if we
are to have any hope of retrieving the original data!
With this in mind, let us go back to Chapter 18 (discussing the staircase
of primes) and Chapter 19, where we tinkered with the original staircase of
primes—alias: the graph of π(X)—to get ψ(X) whose risers look—from afar—
as if they approximated the 45 degree staircase.
At this point we’ll do some further carpentry on ψ(X) without destroying the
valuable information it contains. We will be replacing ψ(X) by a generalized
function, i.e., a distribution, which we denote Φ(t) that has support at all posi-
tive integral multiples of logs of prime numbers, and is zero on the complement
of that discrete set. Recall that by definition, a discrete subset S of real num-
bers is the support of a function, or of a distribution, if the function vanishes
on the complement of S and doesn’t vanish on the complement of any proper
subset of S.
Given the mission of our book, it may be less important for us to elaborate on
the construction of Φ(t) than it is (a) to note that Φ(t) contains all the valuable
information that ψ(X) has and (b) to pay close attention to the spike values of
the trigonometric series that is the Fourier transform of Φ(t).
For the definition of the distribution Φ(t) see the end-note [18].
A distribution that has a discrete set of real numbers as its support—as Φ(t)
does—we sometimes like to call spike distributions since the pictures of func-
tions approximating it tend to look like a series of spikes.
We have then before us a spike distribution with support at integral multiples of
logarithms of prime numbers, and this generalized function retains the essential
108 CHAPTER 31. ON LOSING NO INFORMATION

information about the placement of prime numbers among all whole numbers,
and will be playing a major role in our story: knowledge of the placement of the
“blips” constituting this distribution (its support), being at integral multiples of
logs of prime numbers, would allow us to reconstruct the position of the prime
numbers among all numbers. Of course there are many other ways to package
this vital information, so we must explain our motivation for subjecting our
poor initial staircase to the particular series of brutal acts of distortion that we
described, which ends up with the distribution Φ(t).
Chapter 32

Going from the primes to

the Riemann spectrum

We discussed the nature of the Fourier transform of (symmetrized) δ-functions

in Chapter 28. In particular, recall the “spike function”

dx (t) = (δx (t) + δ−x (t))/2

that has support at the points x and −x. We mentioned that its Fourier trans-
form, dˆx (θ), is equal to cos(xθ) (and gave some hints about why this may be
true).
Our next goal is to work with the much more interesting “spike function”

Φ(t) = e−t/2 Ψ0 (t),

which was one of the generalized functions that we engineered in Chapter 31,
and that has support at all nonnegative integral multiples of logarithms of prime
numbers.
As with any function—or generalized function—defined for non-negative values
of t, we can “symmetrize it” (about the t-axis) which means that we can define
it on negative real numbers by the equation

Φ(−t) = Φ(t).

Let us make that convention, thereby turning Φ(t) into an even generalized
function, as illustrated in Figure 32.1. (An even function on the real line is a
function that takes the same value on any real number and its negative as in
the formula above.)

109
110 CHAPTER 32. FROM PRIMES TO THE RIEMANN SPECTRUM

log(16) log(5) log(2) log(2) log(5) log(16)

Figure 32.1: Φ(t) is a sum of Dirac delta functions at the logarithms of prime
powers pn weighted by p−n/2 log(p) (and log(2π) at 0).

We may want to think of Φ(t) as a limit of this sequence of distributions,

Φ(t) = lim Φ≤C (t)
C→∞

where Φ≤C (t) is the following finite linear combination of (symmetrized) δ-

functions dx (t):
X
Φ≤C (t) := 2 p−n/2 log(p)dn log(p) (t).
prime powers pn ≤C

Since the Fourier transform of dx (t) is cos(xθ), the Fourier transform of each
dn log(p) (t) is cos(n log(p)θ), so the Fourier transform of Φ≤C (t) is
X
Φ̂≤C (θ) := 2 p−n/2 log(p) cos(n log(p)θ).
prime powers pn ≤C

So, following the discussion in Chapter 29 above, we are dealing with the cutoffs
at finite points C of the trigonometric series1
X
Φ̂(θ) := 2 p−n/2 log(p) cos(n log(p)θ).
prime powers pn

For example, when C = 3, we have the rather severe cutoff of these trigonometric
series: Φ̂≤3 (θ) takes account only of the primes p = 2 and p = 3:
2 2
Φ̂≤3 (θ) = √ log(2) cos(log(2)θ) + √ log(3) cos(log(3)θ),
2 3

1 The trigonometric series in the text—whose spectral values are the logarithms of prime

powers—may also be written as

∞
X ∞
X
Λ(m)m−s + Λ(m)m−s̄
m=2 m=2
1
for s = 2
+ iθ, where Λ(m) is the von-Mangoldt function.
 111

which we plot in Figure 32.2.

2

0
5 10 15 20 25 30 35 40

-1

-2

Figure 32.2: Plot of Φ̂≤3 (θ)

We will be interested in the values of θ that correspond to higher and higher

peaks of our trigonometric series Φ̂≤C (θ) as C → ∞. For example, the value of
θ that provides the first peak of Φ̂0≤3 (θ) such that |Φ̂0≤3 (θ)| > 2 is

θ = 14.135375354 . . . .

So in Figure 32.2 we begin this exploration by plotting Φ̂≤3 (θ), together with
its derivative, highlighting the zeroes of the derivative.
2
1
0 5 10 15 20 25 30 35 40
-1
-2

Figure 32.3: Plot of Φ̂≤3 (θ) in blue and its derivative in grey

As we shall see in subsequent figures of this chapter, there seems to be an

eventual convergence of the values of θ that correspond to higher and higher
peaks, the red dots in the figure above, as C tends to ∞. These “limit” θ-values
we now insert as the endpoints of red vertical lines into Figure 32.4 comparing
them with the red dots for our humble cutoff C = 3.
112 CHAPTER 32. FROM PRIMES TO THE RIEMANN SPECTRUM

10 20 30 40

-1

-2

Figure 32.4: Φ̂≤3 (θ)

We give a further sample of graphs for a few higher cutoff values C (introducing
a few more primes into the game!).

Figures 32.5–32.7 contain graphs of various cutoffs of Φ̂≤C (θ). As C increases

a sequence of spikes down emerge which we indicate with red vertical arrows.

4
5
2

10 20 30 40
10 20 30 40

-2 -5

-4
-10

Figure 32.5: Plot of Φ̂≤C (θ) for C = 5 and 10

Given the numerical-experimental approach we have been adopting in this book,

it is a particularly fortunate (and to us: surprising) thing that the convergence
to those vertical red lines can already be illustrated using such small cutoff
values C. One might almost imagine making hand computations that exhibit
this phenomenon! Following in this spirit, see David Mumford’s blog post http:
//www.dam.brown.edu/people/mumford/blog/2014/RiemannZeta.html.

To continue:
 113

10 10

5 5

10 20 30 40 10 20 30 40
-5 -5

-10 -10

-15 -15

Figure 32.6: Plot of Φ̂≤C (θ) for C = 10 and 100

10 10

5 5

10 20 30 40 10 20 30 40
-5 -5

-10 -10

-15 -15

Figure 32.7: Plot of Φ̂≤C (θ) for C = 200 and 500

For a theoretical discussion of these spikes, see endnote [19].

The θ-coordinates of these spikes seem to be vaguely clustered about a discrete
set of positive real numbers. These “spikes” are our first glimpse of a certain
infinite set of positive real numbers

θ1 , θ2 , θ3 , . . .

which constitute the Riemann spectrum of the primes. If the Riemann

Hypothesis holds, these numbers would be the key to the placement of primes
on the number line.
By tabulating these peaks we compute—at least very approximately—. . .

θ1 = 14.134725 . . .
θ2 = 21.022039 . . .
θ3 = 25.010857 . . .
θ4 = 30.424876 . . .
θ5 = 32.935061 . . .
θ6 = 37.586178 . . .

Riemann defined this sequence of numbers in his 1859 article in a manner some-
what different from the treatment we have given. In that article these θi appear
as “imaginary parts of the nontrivial zeroes of his zeta function;” we will discuss
this briefly in Part IV, Chapter 37 below.
Chapter 33

How many θi’s are there?

The Riemann spectrum, θ1 , θ2 , θ3 , . . . clearly deserves to be well understood!

What do we know about this sequence of positive real numbers?
Just as we did with the prime numbers before, we can count these numbers
θ1 = 14.1347 . . ., θ2 = 21.0220 . . ., etc., and form the staircase of Figure 33.1,
with a step up at θ1 , a step up at θ2 , etc.
3

2.5

1.5

0.5

0
5 10 15 20 25 30

Figure 33.1: The staircase of the Riemann spectrum

Again, just as with the staircase of primes, we might hope that as we plot this
staircase from a distance as in Figures 33.2 and 33.3 that it will look like a
beautiful smooth curve.
In fact, we know, conditional on RH, the staircase of real numbers θ1 , θ2 , θ3 , . . .
is very closely approximated by the curve
T T
log ,
2π 2πe
(the error term being bounded by a constant times log T ).

114
 115

10
25
8
20
6
15
4
10
2 5

0
0 10 20 30 40 50 20 40 60 80 100

T T
Figure 33.2: The staircase of the Riemann spectrum and the curve 2π log 2πe

600

500

400

300

200

100

200 400 600 800 1000

Figure 33.3: The staircase of the Riemann spectrum looks like a smooth curve

Nowadays, these mysterious numbers θi , these spectral lines for the staircase
of primes, are known in great abundance and to great accuracy. Here is the
smallest one, θ1 , given with over 1,000 digits of its decimal expansion:

14.134725141734693790457251983562470270784257115699243175685567460149
9634298092567649490103931715610127792029715487974367661426914698822545
8250536323944713778041338123720597054962195586586020055556672583601077
3700205410982661507542780517442591306254481978651072304938725629738321
5774203952157256748093321400349904680343462673144209203773854871413783
1735639699536542811307968053149168852906782082298049264338666734623320
0787587617920056048680543568014444246510655975686659032286865105448594
4432062407272703209427452221304874872092412385141835146054279015244783
3835425453344004487936806761697300819000731393854983736215013045167269
6838920039176285123212854220523969133425832275335164060169763527563758
9695376749203361272092599917304270756830879511844534891800863008264831
2516911271068291052375961797743181517071354531677549515382893784903647
4709727019948485532209253574357909226125247736595518016975233461213977
3160053541259267474557258778014726098308089786007125320875093959979666
60675378381214891908864977277554420656532052405
116 CHAPTER 33. HOW MANY θI ’S ARE THERE?

and if, by any chance, you wish to peruse the first 2,001,052 of these θi ’s calcu-
lated to an accuracy within 3 · 10−9 , consult Andrew Odlyzko’s tables:

http://www.dtc.umn.edu/~odlyzko/zeta_tables
Chapter 34

Further questions about the

Riemann spectrum

Since people have already computed1 the first 10 trillion θ’s and have never
found one with multiplicity > 1, it is generally expected that the multiplicity of
all the θ’s in the Riemann spectrum is 1.
But, independent of that expectation, our convention in what follows will be
that we count each of the elements in the Riemann spectrum repeated as many
times as their multiplicity. So, if it so happens that θn occurs with multiplicity
two, we view the Riemann spectrum as being the series of numbers

θ1 , θ2 , . . . , θn−1 , θn , θn , θn+1 , . . .

It has been conjectured that there are no infinite arithmetic progressions among
these numbers. More broadly, one might expect that there is no visible correla-
tion between the θi ’s and translation, i.e., that the distribution of θi ’s modulo
any positive number T is random, as in Figure 34.1.

0.15 1
0.8
0.1 0.6
0.4
0.05
0.2

1 2 3 4 5 6 0.2 0.4 0.6 0.8 1

Figure 34.1: Frequency histogram of Odlyzko’s computation of the Riemann

spectrum modulo 2π (left) and modulo 1 (right)

1 See http://numbers.computation.free.fr/Constants/constants.html for details.

117
118CHAPTER 34. FURTHER QUESTIONS ABOUT THE RIEMANN SPECTRUM

In analogy with the discussion of prime gaps in Chapter 6, we might com-

pute pair correlation functions and the statistics of gaps between successive θ’s
in the Riemann spectrum (see http://www.baruch.cuny.edu/math/Riemann_
Hypothesis/zeta.zero.spacing.pdf). This study was begun by H. L. Mont-
gomery and F. J. Dyson. As Dyson noticed, the distributions one gets from the
Riemann spectrum bears a similarity to the distributions of eigenvalues of a
random unitary matrix.2 This has given rise to what is know as the random
matrix heuristics, a powerful source of conjectures for number theory and other
branches of mathematics.
Here is a histogram of the distribution of differences θi+1 − θi :

1500

1000

500

0
0 0.5 1 1.5 2

Figure 34.2: Frequency histogram of the first 99,999 gaps in the Riemann spec-
trum

2 Any connection of the Riemann spectrum to eigenvalues of matrices would indeed be

exciting for our understanding of the Riemann Hypothesis, in view of what is known as
the Hilbert-Pólya Conjecture (see http://en.wikipedia.org/wiki/Hilbert%E2%80%93P%C3%
B3lya_conjecture).
Chapter 35

Going from the Riemann

spectrum to the primes

To justify the name Riemann spectrum of primes we will investigate graphically

whether in an analogous manner we can use this spectrum to get information
about the placement of prime numbers. We might ask, for example, if there is a
trigonometric function with frequencies given by this collection of real numbers,

θ1 , θ2 , θ3 , . . .

that somehow pinpoints the prime powers, just as our functions

X
Φ̂(θ)≤C = 2 p−m/2 log(p) cos(n log(p)θ)
prime powers pn ≤C

for large C pinpoint the spectrum (as discussed in the previous two chapters).
To start the return game, consider this sequence of trigonometric functions that
have (zero and) the θi as spectrum
X
GC (x) := 1 + cos(θi · x).
i<C

As we’ll see presently it is best to view these functions on a logarithmic scale so

we will make the substitution of variables x = log(s) and write
X
HC (s) := GC (log(s)) = 1 + cos(θi · log(s)).
i<C

The theoretical story behind the phenomena that we will see graphically in
this chapter is a manifestation of Riemann’s explicit formula. For modern text
references that discuss this general subject, see endnote [20]

119
120 CHAPTER 35. FROM THE RIEMANN SPECTRUM TO PRIMES

2 3 4 5 7 8 9 11 13 1617 19 23 25 27 29

P1000
Figure 35.1: Illustration of − i=1 cos(log(s)θi ), where θ1 ∼ 14.13, . . . are the
first 1000 contributions to the Riemann spectrum. The red dots are at the prime
powers pn , whose size is proportional to log(p).

27 29 31 32

P1000
Figure 35.2: Illustration of − i=1 cos(log(s)θi ) in the neighborhood of a twin
prime. Notice how the two primes 29 and 31 are separated out by the Fourier
series, and how the prime powers 33 and 25 also appear.

1013 1019 1021 1024

Figure 35.3: Fourier series from 1, 000 to 1, 030 using 15,000 of the numbers θi .
Note the twin primes 1,019 and 1,021 and that 1,024 = 210 .
 Part IV

Back to Riemann

121
Chapter 36

How to build π(X) from the

spectrum (Riemann’s way)

We have been dealing in Part III of our book with Φ(t), a distribution that—
we said—contains all the essential information about the placement of primes
among numbers. We have given a clean restatement of Riemann’s hypothesis,
the third restatement so far, in term of this Φ(t). But Φ(t) was the effect of a
series of recalibrations and reconfigurings of the original untampered-with stair-
case of primes. A test of whether we have strayed from our original problem—to
understand this staircase—would be whether we can return to the original stair-
case, and “reconstruct it” so to speak, solely from the information of Φ(t)—or
equivalently, assuming the Riemann Hypothesis as formulated in Chapter 19—
can we construct the staircase of primes π(X) solely from knowledge of the
sequence of real numbers θ1 , θ2 , θ3 , . . . ?
The answer to this is yes (given the Riemann Hypothesis), and is discussed very
beautifully by Bernhard Riemann himself in his famous 1859 article.
Bernhard Riemann used the spectrum of the prime numbers to provide an exact
analytic formula that analyzes and/or synthesizes the staircase of primes. This
formula is motivated by Fourier’s analysis of functions as constituted out of
RX
cosines. Recall from Chapter 13 that Gauss’s guess is Li(X) = 2 dt/log(t). To
continue this discussion, we do need some familiarity with complex numbers,
for the definition of Riemann’s exact formula requires extending the definition
of the function Li(X) to make sense for complex numbers X = a + bi. In fact,
RX
more naturally, one might work with the path integral li(X) := 0 dt/log(t).
Riemann begins his discussion (see Figure 36.1) by defining

X∞ XN
µ(n) 1
(N ) µ(n) 1
R(X) = li(X ) = lim R (X) := lim
n li(X n ),
n=1
n N →∞ N →∞
n=1
n

122
 123

where R(N ) (X) denotes the truncated sum, which one can compute as an ap-
proximation.

Figure 36.1: Riemann defining R(X) in his manuscript

In all the discussion of this section the order of summation is important. For
such considerations and issues regarding actual computation we refer to Riesel-
Gohl (see http://wstein.org/rh/rg.pdf).
Here µ(n) is the Möbius function which is defined by
 if n is a square-free positive integer with an even

1

 number of distinct prime factors,



µ(n) = −1 if n is a square-free positive integer with an odd

 number of distinct prime factors,





0 if n is not square-free.

See Figure 36.2 for a plot of the Möbius function.

1
0.5

10 20 30 40 50
-0.5
-1

Figure 36.2: The blue dots plot the values of the Möbius function µ(n), which
is only defined at integers.

In Chapter 17 we encountered the Prime Number Theorem, which asserts that

X/ log(X) and Li(X) are both approximations for π(X), in the sense that both
go to infinity at the same rate. That is, the ratio of any two of these three
functions tends to 1 as X goes to ∞. Our first formulation of the Riemann
Hypothesis (see page 50) was that Li(X) is an essentially square root accurate
approximation of π(X). Figures 36.3–36.4 illustrate that Riemann’s function
R(X) appears to be an even better approximation to π(X) than anything we
have seen before.
124 CHAPTER 36. BUILDING π(X) FROM THE SPECTRUM

25 150
20
100
15

10
50
5

0
0 20 40 60 80 100 200 400 600 800 1000

Figure 36.3: Comparisons of Li(X) (top), π(X) (middle), and R(X) (bottom,
computed using 100 terms)

1340
1320
1300
1280
1260
1240

10000 10200 10400 10600 10800 11000

Figure 36.4: Closeup comparison of Li(X) (top), π(X) (middle), and R(X)
(bottom, computed using 100 terms)

Think of Riemann’s smooth curve R(X) as the fundamental approximation to

π(X). Riemann offered much more than just a (conjecturally) better approxi-
mation to π(X) in his wonderful 1859 article (see Figure 36.5). He found a way
to construct what looks vaguely like a Fourier series, but with sin(X) replaced
by R(X) and its spectrum the θi , which conjecturally equals π(X) (with a slight
correction if the number X is itself a prime).

Figure 36.5: Riemann’s analytic formula for π(X)

In this manner, Riemann gave an infinite sequence of improved guesses, begin-

ning with R0 (X) (see equation (18) of Riesel-Gohl at http://wstein.org/rh/
rg.pdf) a modification of R(X) that takes account of the pole and the trivial
zeroes of the Riemann zeta-function, and then considered a sequence:
 125

R0 (X), R1 (X), R2 (X), R3 (X), ...

and he hypothesized that one and all of them were all essentially square root
approximations to π(X), and that the sequence of these better and better ap-
proximations converge to give an exact formula for π(X).
Thus not only did Riemann provide a “fundamental” (that is, a smooth curve
that is astoundingly close to π(X)) but he viewed this as just a starting point,
for he gave the recipe for providing an infinite sequence of corrective terms—
call them Riemann’s harmonics; we will denote the first of these “harmonics”
C1 (X), the second C2 (X), etc. Riemann gets his first corrected curve, R1 (X),
from R0 (X) by adding this first harmonic to the fundamental,

R1 (X) = R0 (X) + C1 (X),

he gets the second by correcting R1 (X) by adding the second harmonic

R2 (X) = R1 (X) + C2 (X),

and so on
R3 (X) = R2 (X) + C3 (X),
and in the limit provides us with an exact fit.
The Riemann Hypothesis, if true, would tell us that these correction terms
C1 (X), C2 (X), C3 (X), . . . are all square-root small
The elegance of Riemann’s treatment of this problem is that the corrective terms
Ck (X) are all modeled on the fundamental R(X) and are completely described
if you know the sequence of real numbers θ1 , θ2 , θ3 , . . . of the last section.
Assuming the Riemann Hypothesis, the Riemann correction terms Ck (X) are
defined to be
1 1
Ck (X) = −R(X 2 +iθk ) − R(X 2 −iθk ),
where θ1 = 14.134725 . . . , θ2 = 21.022039 . . . , etc., is the spectrum of the prime
numbers [21].
In sum, Riemann provided an extraordinary recipe that allows us to work out
the harmonics,
C1 (X), C2 (X), C3 (X), . . .
without our having to consult, or compute with, the actual staircase of primes.
As with Fourier’s modus operandi where both fundamental and all harmonics
are modeled on the sine wave, but appropriately calibrated, Riemann fashioned
his higher harmonics, modeling them all on a single function, namely R(X).
The convergence of Rk (X) to π(X) is strikingly illustrated in the plots in Fig-
ures 36.6–36.11 of Rk for various values of k.
126 CHAPTER 36. BUILDING π(X) FROM THE SPECTRUM

25

20

15

10

0 20 40 60 80 100

Figure 36.6: The function R1 approximating the staircase of primes up to 100

25

20

15

10

0
0 20 40 60 80 100

Figure 36.7: The function R10 approximating the staircase of primes up to 100

25

20

15

10

0
0 20 40 60 80 100

Figure 36.8: The function R25 approximating the staircase of primes up to 100
 127
25

20

15

10

0
0 20 40 60 80 100

Figure 36.9: The function R50 approximating the staircase of primes up to 100

80

60

40

20

100 200 300 400 500

Figure 36.10: The function R50 approximating the staircase of primes up to 500
84

82

80

78

76

74

72

70

350 360 370 380 390 400

Figure 36.11: The function Li(X) (top, green), the function R50 (X) (in blue),
and the staircase of primes on the interval from 350 to 400.
Chapter 37

As Riemann envisioned it,

the zeta function relates the
staircase of primes to its
Riemann spectrum

In the previous chapters we have described—using Riemann’s Hypothesis—how

to obtain the spectrum
θ1 , θ2 , θ3 , . . .
from the staircase of primes, and hinted at how to go back. Roughly speaking, we
were performing “Fourier transformations” to make this transit. But Riemann,
on the very first page of his 1859 memoir, construes the relationship we have
been discussing, between spectrum and staircase, in an even more profound way.
To talk about this extraordinary insight of Riemann, one would need to do two
things that might seem remote from our topic, given our discussion so far.

• We will discuss a key idea that Leonhard Euler had (circa 1740).

• To follow the evolution of this idea in the hands of Riemann, we would

have to assume familiarity with basic complex analysis.

We will say only a few words here about this, in hopes of giving at least a shred
of a hint of how marvelous Riemann’s idea is. We will be drawing, at this point,
on some further mathematical background. For readers who wish to pursue the
themes we discuss, here is a list of sources, that are our favorites among those
meant to be read by a somewhat broader audience than people very advanced
in the subject. We list them in order of “required background.”

128
 129

1. John Derbyshire’s Prime Obsession: Bernhard Riemann and the Greatest

Unsolved Problem in Mathematics (2003). We have already mentioned
this book in our foreword, but feel that it is so good, that it is worth a
second mention here.
2. The Wikipedia entry for Riemann’s Zeta Function (http://en.wikipedia.
org/wiki/Riemann_zeta_function). It is difficult to summarize who
wrote this, but we feel that it is a gift to the community in its clarity.
Thanks authors!
3. Enrico Bombieri’s article [22]. To comprehend all ten pages of this excel-
lent and fairly thorough account may require significant background, but
try your hand at it; no matter where you stop, you will have seen good
things in what you have read.

Leonhard Euler’s idea ('1740): As readers of Jacob Bernoulli’s Ars Con-

jectandi (or of the works of John Wallis) know, there was in the early 18th
century already a rich mathematical theory of finite sums of (non-negative k-th
powers) of consecutive integers. This sum,

Sk (N ) = 1k + 2k + 3k + · · · + (N − 1)k ,

is a polynomial in N of degree k + 1 with no constant term, a leading term equal

1
to k+1 N k+1 , and a famous linear term. The coefficient of the linear term of the
polynomial Sk (N ) is the Bernoulli number Bk :

N (N − 1) N2 1
S1 (N ) = 1 + 2 + 3 + · · · + (N − 1) = = − · N,
2 2 2
N3 1
S2 (N ) = 12 + 22 + 32 + · · · + (N − 1)2 = + ··· − · N,
3 6
N4
S3 (N ) = 13 + 23 + 33 + · · · + (N − 1)3 = + · · · − 0 · N,
4
N5 1
S4 (N ) = 14 + 24 + 34 + · · · + (N − 1)4 = + ··· − · N,
5 30
etc. For odd integers k > 1 this linear term vanishes. For even integers 2k the
x2k
Bernoulli number B2k is the rational number given by the coefficient of (2k)! in
the power series expansion

∞
x x X k+1 x2k
= 1 − + (−1) B 2k .
ex − 1 2 (2k)!
k=1
130 CHAPTER 37. AS RIEMANN ENVISIONED IT

So
1 1 1 1
B2 = , B4 = , B6 = , B8 = ,
6 30 42 30
and to convince you that the numerator of these numbers is not always 1, here
are a few more:
5 691 7
B10 = , B12 = , B14 = .
66 2730 6

If you turn attention to sums of negative k-th powers of consecutive integers,

then when k = −1,
1 1 1 1
S−1 (N ) = + + + ··· +
1 2 3 N
tends to infinity like log(N ), but for k < −1 we are facing the sum of reciprocals
of powers (of exponent > 1) of consecutive whole numbers, and Sk (N ) converges.
This is the first appearance of the zeta function ζ(s) for arguments s = 2, 3, 4, . . .
So let us denote these limits by notation that has been standard, after Riemann:
1 1 1
ζ(s) := + s + s + ···
1s 2 3
The striking reformulation that Euler discovered was the expression of this in-
finite sum as an infinite product of factors associated to the prime numbers:
X 1 Y 1
ζ(s) = = ,
n
ns 1 − p−s
p prime

where the infinite sum on the left and the infinite product on the right both
converge (and are equal) if s > 1. He also evaluated these sums at even positive
integers, where—surprise—the Bernoulli numbers come in again; and they and
π combine to yield the values of the zeta function at even positive integers:

1 1
ζ(2) = + 2 + · · · = π 2 /6 ' 1.65 . . .
12 2
1 1
ζ(4) = 4 + 4 + · · · = π 4 /90 ' 1.0823 . . .
1 2
and, in general,

1 1 22n−1
ζ(2n) = + 2n + · · · = (−1)n+1 B2n π 2n · .
12n 2 (2n)!

A side note to Euler’s formulas comes from the fact (only known much later)
that no power of π is rational: do you see how to use this to give a proof that
 131

there are infinitely many primes, combining Euler’s infinite product expansion
displayed above with the formula for ζ(2), or with the formula for ζ(4), or, in
fact, for the formulas for ζ(2n) for any n you choose?
Pafnuty Lvovich Chebyshev’s idea ('1845): The second moment in the
history of evolution of this function ζ(s) is when Chebyshev used the same
formula as above in the extended range where s is allowed now to be a real
variable—not just an integer—greater than 1. Making use of this extension of
the range of definition of Euler’s sum of reciprocals of powers of consecutive
whole numbers, Chebyshev could prove that for large x the ratio of π(x) and
x/ log(x) is bounded above and below by two explicitly given constants. He
also proved that there exists a prime number in the interval bounded by n
and 2n for any positive integer n (this was called Bertrand’s postulate; see
http://en.wikipedia.org/wiki/Proof_of_Bertrand%27s_postulate).

Riemann’s idea (1859): It is in the third step of the evolution of ζ(s) that
something quite surprising happens. Riemann extended the range of Cheby-
shev’s sum of reciprocals of positive real powers of consecutive whole numbers
allowing the argument s to range over the entire complex plane s (avoiding
s = 1). Now this is a more mysterious extension of Euler’s function, and it is
deeper in two ways:

• The formula
1 1 1
ζ(s) := s
+ s + s + ···
1 2 3
does converge when the real part of the exponent s is greater than 1
(i.e., this allows us to use the same formula, as Chebyshev had done, for
the right half plane in the complex plane determined by the condition
s = x + iy with x > 1 but not beyond this). You can’t simply use the
same formula for the extension.

• So you must face the fact that if you wish to “extend” a function beyond
the natural range in which its defining formula makes sense, there may be
many ways of doing it.

To appreciate the second point, the theory of a complex variable is essential.

The uniqueness (but not yet the existence) of Riemann’s extension of ζ(s) to the
entire complex plane is guaranteed by the phenomenon referred to as analytic
continuation. If you are given a function on any infinite subset X of the complex
plane that contains a limit point, and if you are looking for a function on the
entire complex plane1 that is differentiable in the sense of complex analysis,
there may be no functions at all that have that property, but if there is one,
that function is unique. But Riemann succeeded: he was indeed able to extend

1 or to any connected open subset that contains X

132 CHAPTER 37. AS RIEMANN ENVISIONED IT

Euler’s function to the entire complex plane except for the point s = 1, thereby
defining what we now call Riemann’s zeta function.

Those ubiquitous Bernoulli numbers reappear yet again as values of this extended
zeta function at negative integers:

ζ(−n) = −Bn+1 /(n + 1)

so since the Bernoulli numbers indexed by odd integers > 1 all vanish, the
extended zeta function ζ(s) actually vanishes at all even negative integers.
The even integers −2, −4, −6, . . . are often called the trivial zeroes of the
Riemann zeta function. There are indeed other zeroes of the zeta function, and
those other zeroes could—in no way—be dubbed “trivial,” as we shall shortly
see.

It is time to consider these facts:

1. Riemann’s zeta function codes the placement of prime powers

among all numbers. The key here is to take the logarithm and then
the derivative of ζ(s) (this boils down to forming dζ
ds (s)/ζ(s)). Assuming
that the real part of s is > 1, taking the logarithm of ζ(s)—using Euler’s
infinite product formulation—gives us
X

log ζ(s) = − log 1 − p−s ,
p prime

and we can do this term-by-term, since the real part of s is > 1. Then
taking the derivative gives us:

X∞
dζ
(s)/ζ(s) = − Λ(n)n−s
ds n=1

where
(
log(p) when n = pk for p a prime number and k > 0, and
Λ(n) :=
0 if n is not a power of a prime number.

In particular, Λ(n) “records” the placement of prime powers.

2. You know lots about an analytic function if you know its zeroes
and poles. For example for polynomials, or even rational functions: if
 133

someone told you that a certain rational function f (s) vanishes to order
1 at 0 and at ∞; and that it has a double pole at s = 2 and at all other
points has finite nonzero values, then you can immediate say that this
mystery function is a nonzero constant times s/(s − 2)2 .
Knowing the zeroes and poles (in the complex plane) alone of the Riemann
zeta function doesn’t entirely pin it down—you have to know more about
its behavior at infinity since—for example, multiplying a function by ez
doesn’t change the structure of its zeroes and poles in the finite plane.
But a complete understanding of the zeroes and poles of ζ(s) will give all
the information you need to pin down the placement of primes among all
numbers.
So here is the score:

• As for poles, ζ(s) has only one pole. It is at s = 1 and is of order 1

(a “simple pole”).
• As for zeroes, we have already mentioned the trivial zeroes (at nega-
tive even integers), but ζ(s) also has infinitely many nontrivial zeroes.
These nontrivial zeroes are known to lie in the vertical strip

0 < real part of s < 1.

And here is yet another equivalent statement of Riemann’s Hypothesis—this

being the formulation closest to the one given in his 1859 memoir:

The Riemann Hypothesis (fourth formulation)

All the nontrivial zeroes of ζ(s) lie on the vertical line in the complex
plane consisting of the complex numbers with real part equal to 1/2.
These zeroes are none other than 21 ± iθ1 , 21 ± iθ2 , 12 ± iθ3 , . . . , where
θ1 , θ2 , θ3 , . . . comprise the spectrum of primes we talked about in the
earlier chapters.

The “ 12 ” that appears in this formula is directly related to the fact—correspondingly

conditional on RH—that π(X) is “square-root accurately” approximated by
1
Li(X). That is, the error term is bounded by X 2 + . It has been conjectured
that all the zeroes of ζ(s) are simple zeroes.
Here is how Riemann phrased RH:

“One now finds indeed approximately this number of real roots

within these limits, and it is very probable that all roots are real.
Certainly one would wish for a stricter proof here; I have meanwhile
temporarily put aside the search for this after some futile attempts,
as it appears unnecessary for the next objective of my investigation.”
134 CHAPTER 37. AS RIEMANN ENVISIONED IT

In the above quotation, Riemann’s roots are the θi ’s and the statement that
they are “real” is equivalent to RH.
The zeta function, then, is the vise, that so elegantly clamps together informa-
tion about the placement of primes and their spectrum!
That a simple geometric property of these zeroes (lying on a line!) is directly
equivalent to such profound (and more difficult to express) regularities among
prime numbers suggests that these zeroes and the parade of Riemann’s correc-
tions governed by them—when we truly comprehend their message—may have
lots more to teach us, may eventually allow us a more powerful understanding of
arithmetic. This infinite collection of complex numbers, i.e., the nontrivial ze-
roes of the Riemann zeta function, plays a role with respect to π(X) rather like
the role the spectrum of the Hydrogen atom plays in Fourier’s theory. Are the
primes themselves no more than an epiphenomenon, behind which there lies,
still veiled from us, a yet-to-be-discovered, yet-to-be-hypothesized, profound
conceptual key to their perplexing orneriness? Are the many innocently posed,
yet unanswered, phenomenological questions about numbers—such as in the
ones listed earlier—waiting for our discovery of this deeper level of arithmetic?
Or for layers deeper still? Are we, in fact, just at the beginning?
These are not completely idle thoughts, for a tantalizing analogy relates the
number theory we have been discussing to an already established branch of
mathematics—due, largely, to the work of Alexander Grothendieck, and Pierre
Deligne—where the corresponding analogue of Riemann’s hypothesis has indeed
been proved. . . .
Chapter 38

Companions to the zeta

function

Our book, so far, has been exclusively about Riemann’s ζ(s) and its zeroes. We
have been discussing how (the placement of) the zeroes of ζ(s) in the complex
plane contains the information needed to understand (the placement of) the
primes in the set of all whole numbers; and conversely.
It would be wrong—we think—if we don’t even mention that ζ(s) fits into a
broad family of similar functions that connect to other problems in number
theory.
For example—instead of the ordinary integers—consider the Gaussian integers.
This is the collection of numbers
{a + bi}
√
where i = −1 and a, b are ordinary integers. We can add and multiply two
such numbers and get another of the same form. The only “units” among the
Gaussian integers (i.e., numbers whose inverse is again a Gaussian integer) are
the four numbers ±1, ±i and if we multiply any Gaussian integer a + bi by any
of these four units, we get the collection {a + bi, −a − bi, −b + ai, b − ai}. We
measure the size of a Gaussian integer by the square of its distance to the origin,
i.e.,
|a + bi|2 = a2 + b2 .
This size function is called the norm of the Gaussian integer a + bi and can also
be thought of as the product of a + bi and its “conjugate” a − bi. Note that the
norm is a nice multiplicative function on the set of Gaussian integers, in that
the norm of a product of two Gaussian integers is the product of the norms of
each of them.
We have a natural notion of prime Gaussian integer, i.e., one with a > 0
and b ≥ 0 that cannot be factored as the product of two Gaussian integers of

135
136 CHAPTER 38. COMPANIONS TO THE ZETA FUNCTION

smaller size. Given that every nonzero Gaussian integer is uniquely expressible
as a unit times a product of prime Gaussian integers, can you prove that if a
Gaussian integer is a prime Gaussian integer, then its size must either be an
ordinary prime number, or the square of an ordinary prime number?

Figure 38.1 contains a plot of the first few Gaussian primes as they display
themselves amongst complex numbers:

10

1 2 3 4 5 6 7 8 9 10

Figure 38.1: Gaussian primes with coordinates up to 10

Figure 38.2 plots a much larger number of Gaussian primes:

 137

100

80

60

40

20

0
0 20 40 60 80 100

Figure 38.2: Gaussian primes with coordinates up to 100

Figures 38.3–38.6 plot the number of Gaussian primes up to each norm:

138 CHAPTER 38. COMPANIONS TO THE ZETA FUNCTION

2 4 6 8 10 12

Figure 38.3: Staircase of Gaussian primes of norm up to 14

25

20

15

10

20 40 60 80

Figure 38.4: Staircase of Gaussian primes of norm up to 100

 139

150

100

50

200 400 600 800 1000

Figure 38.5: Staircase of Gaussian primes of norm up to 1000

1200

1000

800

600

400

200

2000 4000 6000 8000 10000

Figure 38.6: Staircase of Gaussian primes of norm up to 10000

The natural question to ask, then, is: how are the Gaussian prime numbers
distributed? Can one provide as close an estimate to their distribution and
structure, as one has for ordinary primes? The answer, here is yes: there is a
companion theory, with an analogue to the Riemann zeta function playing a role
similar to the prototype ζ(s). And, it seems as if its “nontrivial zeroes” behave
similarly: as far as things have been computed, they all have the property that
their real part is equal to 12 . That is, we have a companion to the Riemann
Hypothesis.
This is just the beginning of a much larger story related to what has been
come to be called the “Grand Riemann Hypotheses” and connects to analogous
problems, some of them actually solved, that give some measure of evidence for
the truth of these hypotheses. For example, for any system of polynomials in
140 CHAPTER 38. COMPANIONS TO THE ZETA FUNCTION

a fixed number of variables (with integer coefficients, say) and for each prime
number p there are “zeta-type” functions that contain all the information needed
to count the number of simultaneous solutions in finite fields of characteristic
p. That such counts can be well-approximated with a neatly small error term
is related to the placement of the zeroes of these “zeta-type” functions. There
is then an analogous “Riemann Hypothesis” that prescribes precise conditions
on the real parts of their zeroes—this prescription being called the “Riemann
Hypothesis for function fields.” Now the beauty of this analogous hypothesis is
that it has, in fact, been proved!
Is this yet another reason to believe the Grand Riemann Hypothesis?
Endnotes

[1] How not to factor the numerator of a Bernoulli number:

As mentioned in Chapter 37, the coefficient Bk of the linear term of the
polynomial
Sk (n) = 1k + 2k + 3k + · · · + (n − 1)k
is (up to sign) the k-th Bernoulli number. These numbers are rational
numbers and, putting them in lowest terms, their numerators play a role
in certain problems, and their denominators in others. (This is an amazing
story, which we can’t go into here!)
One of us (Barry Mazur) in the recent article How can we construct abelian
Galois extensions of basic number fields? (see http://www.ams.org/
journals/bull/2011-48-02/S0273-0979-2011-01326-X/) found himself
dealing (for various reasons) with the fraction −B200 /400, where B200 is the
two-hundredth Bernoulli number. The numerator of this fraction is quite
large: it is—hold your breath—

389 · 691 · 5370056528687 times this 204-digit number:

N := 3452690329392158031464109281736967404068448156842396721012
9920642145194459192569415445652760676623601087497272415557
0842527652727868776362959519620872735612200601036506871681
124610986596878180738901486527

and he incorrectly asserted that it was prime. Happily, Bartosz Naskrȩcki

spotted this error: our 204-digit N is not prime.
How did he know this? By using the most basic test in the repertoire of
tests that we have available to check to see whether a number is prime:
we’ll call it the “Fermat 2-test.” We’ll first give a general explanation of
this type of test before we show how N fails the Fermat 2-test.
The starting idea behind this test is the famous result known as Fermat’s
Little Theorem where the “little” is meant to alliteratively distinguish it
from you-know-what.

141
142 ENDNOTES

Theorem 39.1 (Fermat’s Little Theorem). If p is a prime number, and a

is any number relatively prime to p then ap−1 − 1 is divisible by p.

A good exercise is to try to prove this, and a one-word hint that might lead
you to one of the many proofs of it is induction.1
Now we are going to use this as a criterion, by—in effect—restating it in
what logicians would call its contrapositive:
Theorem 39.2 (The Fermat a-test). If M is a positive integer, and a is
any number relatively prime to M such that aM −1 − 1 is not divisible by
M , then M is not a prime.

Well, Naskrȩcki computed 2N −1 − 1 (for the 204-digit N above) and saw

that it is not divisible2 by N . Ergo, our N fails the Fermat 2-test so is not
prime.
But then, given that it is so “easy” to see that N is not prime, a natural
question to ask is: what, in fact, is its prime factorization? This—it turns
out—isn’t so easy to determine; Naskrȩcki devoted 24 hours of computer
time setting standard factorization algorithms on the task, and that was not
sufficient time to resolve the issue. The factorization of the numerators of
general Bernoulli numbers is the subject of a very interesting website run by
Samuel Wagstaff (http://homes.cerias.purdue.edu/~ssw/bernoulli).
Linked to this web page one finds (http://homes.cerias.purdue.edu/
~ssw/bernoulli/composite) which gives a list of composite numbers
whose factorizations have resisted all attempts to date. The two-hundredth
Bernoulli number is 12th on the list.
The page http://en.wikipedia.org/wiki/Integer_factorization_
records lists record challenge factorization, and one challenge that was
completed in 2009 involves a difficult-to-factor number with 232 digits; its
factorization was completed by a large team of researchers and took around
2000 years of CPU time. This convinced us that with sufficient motivation
it would be possible to factor N , and so we asked some leaders in the field
to try. They succeeded!
Factorisation of B200
by Bill Hart on 4 Aug 05, 2012 at 07:24pm
We are happy to announce the factorization of the numerator of the
200th Bernoulli number:

1 Here’s the proof:

(N + 1)p ≡ N p + 1 ≡ (N + 1) mod p,
where the first equality is the binomial theorem and the second equality is induction.
2 The number 2N −1 − 1 has a residue after division by N of

3334581100595953025153969739282790317394606677381970645616725285996925
6610000568292727335792620957159782739813115005451450864072425835484898
565112763692970799269335402819507605691622173717318335512037457.
ENDNOTES 143

N = 389 · 691 · 5370056528687 · c204

c204 = p90 · p115
p90 = 149474329044343594528784250333645983079497454292
= 838248852612270757617561057674257880592603
p115 = 230988849487852221315416645031371036732923661613
= 619208811597595398791184043153272314198502348476
= 2629703896050377709

The factorization of the 204-digit composite was made possible with

the help of many people:
• William Stein and Barry Mazur challenged us to factor this
number.
• Sam Wagstaff maintains a table of factorizations of numerators
of Bernoulli numbers at http://homes.cerias.purdue.edu/
~ssw/bernoulli/bnum. According to this table, the 200th
Bernoulli number is the 2nd smallest index with unfactored nu-
merator (the first being the 188th Bernoulli number).
• Cyril Bouvier tried to factor the c204 by ECM up to 60-digit
level, using the TALC cluster at Inria Nancy - Grand Est.
• yoyo@home tried to factor the c204 by ECM up to 65-digit level,
using the help of many volunteers of the distributed computing
platform http://www.rechenkraft.net/yoyo/. After ECM
was unsuccessful, we decided to factor the c204 by GNFS.
• Many people at the Mersenne forum helped for the polyno-
mial selection. The best polynomial was found by Shi Bai,
using his implementation of Kleinjung’s algorithm in CADO-
NFS: http://www.mersenneforum.org/showthread.php?
p=298264#post298264. Sieving was performed by many vol-
unteers using NFS@home, thanks to Greg Childers. See http:
//escatter11.fullerton.edu/nfs for more details about
NFS@home. This factorization showed that such a distributed
effort might be feasible for a new record GNFS factorization,
in particular for the polynomial selection. This was the largest
GNFS factorization performed by NFS@home to date, the sec-
ond largest being 21040 + 1 at 183.7 digits.
• Two independent runs of the filtering and linear algebra were
done: one by Greg Childers with msieve (http://www.boo.
net/~jasonp/qs.html) using a 48-core cluster made available
by Bill Hart, and one by Emmanuel Thomé and Paul Zimmer-
mann with CADO-NFS (http://cado-nfs.gforge.inria.
fr/), using the Grid 5000 platform.
144 ENDNOTES

• The first linear algebra run to complete was the one with CADO-
NFS, thus we decided to stop the other run.
Bill Hart

We verify the factorization above in SageMath as follows:

sage: p90 = 1494743290443435945287842503336459830794974542928382

48852612270757617561057674257880592603
sage: p115 = 230988849487852221315416645031371036732923661613619
2088115975953987911840431532723141985023484762629703896050377709
sage: c204 = p90 * p115
sage: 389 * 691 * 5370056528687 * c204 == -numerator(bernoulli(200))
True
sage: is prime(p90), is prime(p115), is prime(c204)
(True, True, False) .

[2] Given an integer n, there are many algorithms available for trying to write
n as a product of prime numbers. First we can apply trial division, where
we simply divide n by each prime 2, 3, 5, 7, 11, 13, . . . in turn, and see what
small prime factors we find (up to a few digits). After using this method
to eliminate as many primes as we have patience to eliminate, we typi-
cally next turn to a technique called Lenstra’s elliptic curve method, which
allows us to check n for divisibility by bigger primes (e.g., around 10–15
digits). Once we’ve exhausted our patience using the elliptic curve method,
we would next hit our number with something called the quadratic sieve,
which works well for factoring numbers of the form n = pq, with p and
q primes of roughly equal size, and n having less than 100 digits (say,
though the 100 depends greatly on the implementation). All of the above
algorithms—and then some—are implemented in SageMath, and used by
default when you type factor(n) into SageMath. Try typing factor(some
number, verbose=8) to see for yourself.
If the quadratic sieve fails, a final recourse is to run the number field
sieve algorithm, possibly on a supercomputer. To give a sense of how
powerful (or powerless, depending on perspective!) the number field
sieve is, a record-setting factorization of a general number using this al-
gorithm is the factorization of a 232 digit number called RSA-768 (see
https://eprint.iacr.org/2010/006.pdf):
n = 12301866845301177551304949583849627207728535695953347921973224521
517264005072636575187452021997864693899564749427740638459251925573263
034537315482685079170261221429134616704292143116022212404792747377940
80665351419597459856902143413
which factors as pq, where
p = 334780716989568987860441698482126908177047949837137685689124313889
82883793878002287614711652531743087737814467999489
ENDNOTES 145

and
q = 367460436667995904282446337996279526322791581643430876426760322838
15739666511279233373417143396810270092798736308917.
We encourage you to try to factor n in SageMath, and see that it fails.
SageMath does not yet include an implementation of the number field sieve
algorithm, though there are some free implementations currently available
(see http://www.boo.net/~jasonp/qs.html).
[3] We can use SageMath (at http://sagemath.org) to quickly compute the
“hefty number” p = 243,112,609 − 1. Simply type p = 2^43112609 - 1 to
instantly compute p. In what sense have we computed p? Internally, p is
now stored in base 2 in the computer’s memory; given the special form
of p it is not surprising that it took little time to compute. Much more
challenging is to compute all the base 10 digits of p, which takes a few
seconds: d = str(p). Now type d[-50:] to see the last 50 digits of p. To
compute the sum 58416637 of the digits of p type sum(p.digits()).
[4] In contrast to the situation with factorization, testing integers of this size
(e.g., the primes p and q) for primality is relatively easy. There are fast al-
gorithms that can tell whether or not any random thousand digit number is
prime in a fraction of second. Try for yourself using the SageMath command
is prime. For example, if p and q are as in endnote 2, then is prime(p)
and is prime(q) quickly output True and is prime(p*q) outputs False.
However, if you type factor(p*q, verbose=8) you can watch as Sage-
Math tries forever and fails to factor pq.
[5] In Sage, the function prime range enumerates primes in a range.
For example, prime range(50) outputs the primes up to 50 and
prime range(50,100) outputs the primes between 50 and 100. Typing
prime range(10^8) in SageMath enumerates the primes up to a hundred
million in around a second. You can also enumerate primes up to a bil-
lion by typing v=prime range(10^9), but this will use a large amount of
memory, so be careful not to crash your computer if you try this. You can
see that there are π(109 ) = 50,847,534 primes up to a billion by then typ-
ing len(v). You can also compute π(109 ) directly, without enumerating
all primes, using the command prime pi(10^9). This is much faster since
it uses some clever counting tricks to find the number of primes without
actually listing them all.
In Chapter 19 we tinkered with the staircase of primes by first counting
both primes and prime powers. There are comparatively few prime powers
that are not prime. Up to 108 , only 1,405 of the 5,762,860 prime powers
are not themselves primes. To see this, first enter a = prime pi(10^8);
pp = len(prime powers(10^8)). Typing (a, pp, pp-a) then outputs
the triple (5761455, 5762860, 1405).
[6] Hardy and Littlewood give a nice conjectural answer to such questions
about gaps between primes. See Problem A8 of Guy’s book Unsolved Prob-
146 ENDNOTES

lems in Number Theory (2004). Note that Guy’s book discusses counting
the number Pk (X) of pairs of primes up to X that differ by a fixed even
number k; we have Pk (X) ≥ Gapk (X), since for Pk (X) there is no require-
ment that the pairs of primes be consecutive.

[7] If f (x) and g(x) are real-valued functions of a real variable x such that
for any  > 0 both of them take their values between x1− and x1+ for x
sufficiently large, then say that f (x) and g(x) are good approximations
of one another if, for any positive  the absolute value of their difference
1
is less than x 2 + for x sufficiently large. The functions Li(X) and R(X)
are good approximations of one another.

[8] This computation of π(X) was done by David J. Platt in 2012, and is the
largest value of π(X) ever computed. See http://arxiv.org/abs/1203.
5712 for more details.

√ fact, the Riemann Hypothesis is equivalent to | Li(X) − π(X)| ≤

[9] In
X log(X) for all X ≥ 2.01. See Section 1.4.1 of Crandall and Pomer-
ance’s book Prime Numbers: A Computational Perspective.

[10] For a proof of this here’s a hint. Compute the difference between the deriva-
tives of Li(x) and of x/ log x. The answer is 1/ log2 (x). So you must show
RX RX
that the ratio of 2 dx/ log2 (x) to Li(x) = 2 dx/ log(x) tends to zero as x
goes to infinity, and this is a good calculus exercise.

[11] See http://www.maths.tcd.ie/pub/HistMath/People/Riemann/Zeta/

for the original German version and an English translation.

[12] We have X
ψ(X) = log p
pn ≤X

where the summation is over prime powers pn that are ≤ X.

[13] See http://en.wikipedia.org/wiki/Fast_Fourier_transform.

[14] http://en.wikipedia.org/wiki/Distribution_%28mathematics%29
contains a more formal definition and treatment of distributions. Here is
Schwartz’s explanation for his choice of the word distribution:

“Why did we choose the name distribution? Because, if µ is a

measure, i.e., a particular kind of distribution, it can be con-
sidered as a distribution of electric charges in the universe. Dis-
tributions give more general types of electric charges, for exam-
ple dipoles and magnetic distributions. If we consider the dipole
placed at the point a having magnetic moment M , we easily see
that it is defined by the distribution −DM δ(a) . These objects oc-
cur in physics. Deny’s thesis, which he defended shortly after,
introduced electric distributions of finite energy, the only ones
ENDNOTES 147

which really occur in practice; these objects really are distribu-

tions, and do not correspond to measures. Thus, distributions
have two very different aspects: they are a generalization of the
notion of function, and a generalization of the notion of distribu-
tion of electric charges in space. [...] Both these interpretations
of distributions are currently used.”

[15] David Mumford suggested that we offer the following paragraph from
http://en.wikipedia.org/wiki/Dirac_delta_function
on the Dirac delta function:

An infinitesimal formula for an infinitely tall, unit impulse delta

function (infinitesimal version of Cauchy distribution) explicitly
appears in an 1827 text of Augustin Louis Cauchy. Siméon Denis
Poisson considered the issue in connection with the study of wave
propagation as did Gustav Kirchhoff somewhat later. Kirchhoff
and Hermann von Helmholtz also introduced the unit impulse as
a limit of Gaussians, which also corresponded to Lord Kelvin’s
notion of a point heat source. At the end of the 19th century,
Oliver Heaviside used formal Fourier series to manipulate the
unit impulse. The Dirac delta function as such was introduced
as a “convenient notation” by Paul Dirac in his influential 1930
book Principles of Quantum Mechanics. He called it the “delta
function” since he used it as a continuous analogue of the discrete
Kronecker delta.

[16] As discussed in
http://en.wikipedia.org/wiki/Distribution_%28mathematics%29,
“generalized functions” were introduced by Sergei Sobolev in the 1930s,
then later independently introduced in the late 1940s by Laurent Schwartz,
who developed a comprehensive theory of distributions.

[17] If the Riemann Hypothesis holds they are precisely the imaginary parts of
the “nontrivial” zeroes of the Riemann zeta function.

[18] The construction of Φ(t) from ψ(X):

Succinctly, for positive real numbers t,

Φ(t) : = e−t/2 Ψ0 (t),

where Ψ(t) = ψ(et ) (see Figure 39.2), and Ψ0 is the derivative of Ψ(t),
viewed as a distribution. We extend this function to all real arguments t
by requiring Φ(t) to be an even function of t, i.e., Φ(−t) = Φ(t). But, to
review this at a more leisurely pace,
148 ENDNOTES

1. Distort the X-axis of our staircase by replacing the variable X by et

to get the function

Ψ(t) := ψ(et ).

No harm is done by this for we can retrieve our original ψ(X) as

ψ(X) = Ψ(log(X)).

Our distorted staircase has risers at (0 and) all positive integral mul-
tiples of logs of prime numbers.

2. Now we’ll do something that might seem a bit more brutal: take the
derivative of this distorted staircase Ψ(t). This derivative Ψ0 (t) is a
generalized function with support at all nonnegative integral multiples
of logs of prime numbers.

log(1) log(2) log(3) log(5) log(8) log(13) log(19)

Figure 39.1: Ψ0 (t) is a (weighted) sum of Dirac delta functions at the logarithms
of prime powers pn weighted by log(p) (and by log(2π) at 0). The taller the
arrow, the larger the weight.

3. Now—for normalization purposes—multiply Ψ0 (t) by the function

e−t/2 which has no effect whatsoever on the support.
ENDNOTES 149

200

150

100

50

50 100 150 200

Figure 39.2: Illustration of the staircase ψ(X) constructed in Chapter 19 that

counts weighted prime powers.

In summary: The generalized function that resulted from the above car-
pentry:
Φ(t) = e−t/2 Ψ0 (t),

retains the information we want (the placement of primes) even if in a

slightly different format.

[19] A version of the Riemann–von Mangoldt explicit formula gives some theo-
retical affirmation of the phenomena we are seeing here. We thank Andrew
Granville for a discussion about this.

Figure 39.3: Andrew Granville

Even though the present endnote is not the place to give anything like a
full account, we can’t resist setting down a few of Granville’s comments
that might be helpful to people who wish to go further. (This discussion
can be modified to analyze what happens unconditionally, but we will be
assuming the Riemann Hypothesis below.) The function Φ̂≤C (θ) that we
are graphing in this chapter can be written as:
X
Φ̂≤C (θ) = Λ(n)n−w
n≤C
150 ENDNOTES

1
where w = 2 + iθ. This function, in turn, may be written (by Perron’s
formula) as
Z s=σo +iT X  s
1 C ds
lim Λ(n)n−w
2πi T →∞ s=σo −iT n
n s
Z s=σo +iT X
1 ds
= lim Λ(n)n−w−s C s
2πi T →∞ s=σo −iT n
s
Z s=σo +iT  
1 ζ0 Cs
=− lim (w + s) ds.
2πi T →∞ s=σo −iT ζ s

Here, we assume that σo is sufficiently large and C is not a prime power.

One proceeds, as is standard in the derivation of Explicit Formulae, by
moving the line of integration to the left, picking up residues along the
way. Fix the value of w = 21 + iθ and consider
 
1 ζ0 Cs
Kw (s, C) := (w + s) ,
2πi ζ s

which has poles at

s = 0, 1 − w, and ρ − w,

for every zero ρ of ζ(s). We distinguish five cases, giving descriptive names
to each:

1. Singular pole: s = 1 − w.
2. Trivial poles: s = ρ − w with ρ a trivial zero of ζ(s).
3. Oscillatory poles: s = ρ − w = i(γ − θ) 6= 0 with ρ = 1/2 + iγ(6= w) a
nontrivial zero of ζ(s). (Recall that we are assuming the Riemann Hy-
pothesis, and our variable w = 21 + iθ runs through complex numbers
of real part equal to 12 . So, in this case, s is purely imaginary.)
4. Elementary pole: s = 0 when w is not a nontrivial zero of ζ(s)—i.e.,
when 0 = s 6= ρ − w for any nontrivial zero ρ.
5. Double pole: s = 0 when w is a nontrivial zero of ζ(s)—i.e., when
0 = s = ρ − w for some nontrivial zero ρ. This, when it occurs, is
indeed a double pole, and the residue is given by m · log C + . Here m
is the multiplicity of the zero ρ (which we expect always—or at least
usually—to be equal to 1) and  is a constant (depending on ρ, but
not on C).

The standard technique for the “Explicit formula” will provide us with a
formula for our function of interest Φ̂≤C (θ). The formula has terms result-
ing from the residues of each of the first three types of poles, and of the
ENDNOTES 151

Elementary or the Double pole—whichever exists. Here is the shape of the

formula, given with terminology that is meant to be evocative:

(1) Φ̂≤C (θ) = Sing≤C (θ) + Triv≤C (θ) + Osc≤C (θ) + Elem≤C (θ)

Or:

(2) Φ̂≤C (θ) = Sing≤C (θ) + Triv≤C (θ) + Osc≤C (θ) + Double≤C (θ),

the first if w is not a nontrivial zero of ζ(s) and the second if it is.
The good news is that the functions Sing≤C (θ), Triv≤C (θ) (and also
Elem≤C (θ) when it exists) are smooth (easily describable) functions of the
two variables C and θ; for us, this means that they are not that related to
the essential information-laden discontinuous structure of Φ̂≤C (θ). Let us
bunch these three contributions together and call the sum Smooth(C, θ),
and rewrite the above two formulae as:

(1) Φ̂≤C (θ) = Smooth(C, θ) + Osc≤C (θ)

Or:

(2) Φ̂≤C (θ) = Smooth(C, θ) + Osc≤C (θ) + m · log C + ,

depending upon whether or not w is a nontrivial zero of ζ(s).

We now focus our attention on the Oscillatory term, Osc≤C (θ), approxi-
mating it by a truncation:
X ei log C·(γ−θ)
Oscw (C, X) := 2 .
i(γ − θ)
|γ|<X

Here if a zero has multiplicity m, then we count it m times in the sum. Also,
in this formula we have relegated the “θ” to the status of a subscript (i.e.,
w = 12 + iθ) since we are keeping it constant, and we view the two variables
X and C as linked in that we want the cutoff “X” to be sufficiently large,
say X  C 2 , so that the error term can be controlled.
At this point, we can perform a “multiplicative version” of a Cesàro
RC
summation—i.e., the operator F (c) 7→ (CesF )(C) := 1 F (c)dc/c. This
has the effect of forcing the oscillatory term to be bounded as C tends to
infinity.
This implies that for any fixed θ,
• CesΦ̂≤C (θ) is bounded independent of C if θ is not the imaginary part
of a nontrivial zero of ζ(s) and
152 ENDNOTES

• CesΦ̂≤C (θ) grows as m 2

2 · (log C) + O(log C) if θ is the imaginary part
of a nontrivial zero of ζ(s) of multiplicity m,

giving a theoretical confirmation of the striking feature of the graphs of our

Chapter 32.

[20] A reference for this is:

[I-K]: H. Iwaniec; E. Kowalski, Analytic Number Theory, American

Mathematical Society Colloquium Publications 53 (2004).
(See also the bibliography there.)

Many ways of seeing the explicit relationship are given in Chapter 5 of

[I-K]. For example, consider Exercise 5 on page 109.

X X
φ̂(ρ) = − Λ(n)φ(n) + I(φ),
ρ n≥1

where

• φ is any smooth complex-valued function on [1, +∞) with compact

support,
• φ̂ is its Mellin transform
Z ∞
φ̂(s) := φ(x)xs−1 dx,
0

• the last term on the right, I(φ), is just

Z ∞
1
I(φ) := (1 − 3 )φ(x)dx
1 x −x

(coming from the pole at s = 1 and the “trivial zeroes”).

• The more serious summation on the left hand side of the equation is
over the nontrivial zeroes ρ, noting that if ρ is a nontrivial zero so is ρ̄.

Of course, this “explicit formulation” is not immediately applicable to the

graphs we are constructing since we cannot naively take φ̂ to be a function
forcing the left hand side to be GC (x).
See also Exercise 7 on page 112, which discusses the sums

X x 21 +iθ − 1
x− 1 .
|θ|≤C 2 + iθ

.
ENDNOTES 153

[21] You may well ask how we propose to order these correction terms if RH is
false. Order them in terms of (the absolute value of) their imaginary part,
and in the unlikely situation that there is more than one zero with the
same imaginary part, order zeroes of the same imaginary part by their real
parts, going from right to left.

[22] Bombieri, The Riemann Hypothesis, available at http://www.claymath.

org/sites/default/files/official_problem_description.pdf.