SOCIEDADE BRASILEIRA DE MATEMÁTICA ENSAIOS MATEMÁTICOS

2009, Volume 17, 1–104

The remarkable effectiveness

of ergodic theory in number
theory

Part I. Green-Tao theorem

by Alexander Arbieto, Carlos Matheus and
Carlos G. Moreira

Part II. Elkies-McMullen theorem

by Carlos Matheus

Abstract. The main goal of this survey is the description of the fruitful
interaction between Ergodic Theory and Number Theory via the study of
two beautiful results: the first one by Ben Green and Terence Tao (about
long arithmetic progressions of primes) and the second one by √ Noam Elkies
and Curtis McMullen (about the distribution of the sequence { n} mod 1).
More precisely, during the first part, we will see how the ergodic-theoretical
ideas of Furstenberg about the famous Szemerédi theorem were greatly
generalized by Green and Tao in order to solve the classical problem of
finding arbitrarily long arithmetical progression of prime numbers, while
the second part will focus on how Elkies and McMullen used the ideas
of Ratner’s theory (about the classification of ergodic measures related to
unipotent
√ dynamics) to compute explicitly the distribution of the sequence
{ n} on the unit circle.

2000 Mathematics Subject Classification: 37-02, 11-02.

Preface
The Nobel laureate physicist Eugene Wigner wrote (in 1960) an article
entitled The Unreasonable Effectiveness of Mathematics in the Natural Sci-
ences. In this paper, he argues that the fruitful interaction between Math-
ematics and Physics often points towards profound advances in Physics
and he claims that this is not just a coincidence: instead, this beautiful
interaction reflects a deeper truth about both Mathematics and Physics.
In the present book, we will discuss a much more modest interaction
between two areas of Mathematics, namely Ergodic Theory and Number
Theory, leading to the solution of several interesting problems of Num-
ber Theory (see the next paragraph below). Of course, this interaction
has a much more recent history when compared with the interaction of
Mathematics and Physics, so that it would be exaggerated to call it “un-
reasonable”. On the other hand, since the Ergodic Theory certainly sheds
light into some deep questions in Number Theory, we believe it is fair to
say that the interaction between these two areas of Mathematics is “re-
markable” (so that this justifies the choice for the title of this book).
More concretely, one can quote the following results where some ideas
from Ergodic Theory helped the understanding of a problem in Number
Theory:

• Weyl estimates of exponential sums – the study of exponential sums

P
n
e2πixj is related to the equidistribution properties of the real num-
j=1
bers xn (mod 1);
2
• Lévy’s constant eπ /12 ln 2 ≃ 3, 27582291872... giving the asymptotic
√
exponential growth n qn of the denominators qn of the continued
fraction expansion of (Lebesgue) almost every real number x can
be very clearly explained in terms of the Ergodic
 Theory of the so-
called Gauss map G : [0, 1] → [0, 1], G(x) = x1 , where {.} denotes
the fractionary part of . (namely, via Pesin’s formula relating the
entropy and the Jacobian of G with respect to its natural absolutely
continuous invariant measure).

• Margulis’ solution (1989) of Oppenheim conjecture – the study of

quadratic forms Q of n variables restricted to the integer lattice Zn
is an important subject of Number Theory; for instance, Lagrange
theorem says that the image of the quadratic form Q(a, b, c, d) =
a2 + b2 + c2 + d2 is precisely the set of all natural numbers. In gen-
eral, when Q is positive definite, it is easy to see that Q(Zn ) is a
discrete subset of the positive real numbers. On the other hand, the
4

study of indefinite forms can be a little bit more sophisticate, e.g.,

the study of the values of Q(a, b)√= a2 − db2 for some integer d in-
volves the class field theory of Q( −d) as pointed out by Gauss. In
general, when the coefficients of Q are commensurable, we see that
the image Q(Zn ) is discrete, so that a natural question arises: what
happens when the coefficients of Q are incommensurable? In two
variables, we still have some discreteness phenomena (as√the reader
can check with the quadratic form Q(a, b) = a2 − ((1 + 5)/2)2 b2 ).
Nevertheless, Oppenheim (1929) conjectured that this discreteness
phenomena doesn’t occur in higher dimensions: for n ≥ 3, the im-
age Q(Zn ) of Zn by any indefinite quadratic form Q on n variables
is dense in R. This number-theoretical conjecture was solved by
Margulis via the Ergodic Theory of certain homogenous flows on the
space of lattices (and, in fact, to the best of our knowledge, all proofs
of Oppenheim conjecture are ergodic-theoretical in nature).
√
• Elkies-McMullen theorem (2004) on the gaps of n (mod 1) – given a
sequence of real numbers on the circle S 1 = R/Z, a natural question
concerns the study of its distribution: namely, we can remove these
points from the circle and look at the lengths of the intervals one gets
in this process. For certain sequences of numbers (e.g., a sequence
of randomly chosen real numbers), we know that the distribution
law is given by an exponential function. However, for the sequence
{nα } with 0 < α < 1, we know that it is equidistributed on S 1
and it is conjectured that its distribution law is exponential for any
α 6= 1/2. Furthermore, some numerical experiments performed by
Boshernitzan (around 1993) indicated an special distribution for the
particular case α = 1/2. In this direction, Elkies and McMullen
used the Ergodic Theory of homogenous flows (and, in particular, √
Ratner’s theorem) to explicitly compute the distribution
√ of { n}.
A consequence of their result is the fact that { n} doesn’t have an
exponential distribution.

• Green-Tao theorem (2004) on existence of arbitrarily large arithmetic

progressions of primes – a classical subject of Number Theory is the
study of prime numbers. A particularly interesting problem is to
understand the additive properties of the prime numbers. In this di-
rection, B. Green and T. Tao were able to use the ergodic-theoretical
ideas of Furstenberg’s proof of Szemerédi theorem to show that there
are arbitrarily large arithmetic progressions of primes.

• Einsiedler-Katok-Lindenstrauss theorem (2006) on Littlewood con-

jecture – a fascinating problem of Number Theory is the study of
rational approximations of irrational numbers. Although we have
a lot of information about the rational approximations of a single
 5

irrational number (due to the marvelous properties of the contin-

ued fraction algorithm), the situation becomes more delicate when
one asks about simultaneous approximations of irrational numbers.
In this sense, Littlewood (1930) proposed the following conjecture:
lim inf n→∞ n{nα}{nβ} = 0 for all α, β ∈ R. Again using the Ergodic
Theory of homogenous flows, Einsiedler, Katok and Lindenstrauss
were able to give the following almost complete answer to Little-
wood conjecture: the set of exceptional pairs (α, β) of Littlewood
conjecture has Hausdorff dimension zero!

The initial plan of this book was to cover the Green-Tao, Elkies-McMul-
len and Einsiedler-Katok-Lindenstrauss theorems. However, due to the
usual problem of limitation of space and time, we were forced to make a
choice between these three beautiful results. Because Einsiedler-Katok-
Lindenstrauss theorem is a little bit more difficult to explain from the
technical point of view (in the author’s opinion), we have chosen to discuss
Green-Tao and Elkies-McMullen theorems.
More precisely, this book has two parts: the first part, by A. Arbieto, C.
G. Moreira and C. Matheus, consists of the first two chapters and concerns
Green and Tao theorem, while the last one, by C. Matheus, consists of the
third chapter and concerns Elkies and McMullen theorem. The resume of
the contents of these chapters is:
• in the first chapter (part I), we will make a historical review of the
basic questions and theorems about the additive and multiplicative
properties of integer numbers. In particular, we are going to see that
there are several problems about the additive properties of prime
numbers which are very easy to state but very difficult to solve: e.g.,
it is not known whether there are infinitely many pairs of twin prime
numbers, i.e., pairs of prime numbers whose difference is 2 (this is
known as the Twin Prime Conjecture) and it is still open the so-
called Goldbach conjecture saying that every even natural number
≥ 4 is the sum of two prime numbers. Also, we will see that another
classical conjecture (solved by Ben Green and Terence Tao) says
that there are arbitrarily large arithmetical progressions of prime
numbers. The biggest known progression of primes (to the best of
the authors’ knowledge) is

6171054912832631 + k · 81737658082080, 0 ≤ k ≤ 24,

formed by 25 prime numbers (this arithmetical progression of primes

was discovered in May 17, 2008 by Raanan Chermoni and Jaroslaw
Wroblewski). Observe that this conjecture was solved by Ben Green
and Terence Tao in 2004 and, in fact, Terence Tao was awarded
a Fields Medal in 2006 partly due to this outstanding work with
6

Ben Green. The basic plan of the first chapter is to discuss some
“preparatory” results in the direction of Green and Tao theorem such
as Szemerédi theorem and its ergodic-theoretical proof by Fursten-
berg.

• in the second chapter (part I), we will present the ergodic-theoretical

component of Green and Tao argument, namely, the proof of Green-
Tao-Szemerédi theorem (via the introduction of the Gowers norms
and the adaptation of Furstenberg proof of Szemerédi theorem).
Once this theorem is proved, it remains to perform a number-theo-
retical argument to conclude the Green-Tao theorem from the Green-
Tao-Szemerédi. However, in order to be coherent with the title of this
book, we will omit the very interesting part related to the construc-
tion of a pseudorandom majorant of the (modified) von Mangoldt
function (based on the works of Goldston and Yildirim).

• finally, in the last chapter (part II), we will completely change the
subject from the additive properties of prime numbers
√ to the Elkies-
McMullen calculation of the distribution law of n (mod 1). In
particular, we subdivide this chapter into three sections: the first
two concerns√the translation of the problem of computing the dis-
tribution of n (mod 1) into an ergodic-theoretical problem and
the last section concerns the solution of the corresponding ergodic
problem via Ratner theory of homogenous flows.

Evidently, as the reader can infer from this summary, the parts I and
II are completely independent, so that the reader can chose where he/she
wants to start reading the book.
Finally, we would like to apologize for the omission of Einsiedler-Katok-
Lindestrauss theorem: as a form of compensation, C. Matheus would
like to say that he’s planning to include some notes about Einsiedler-
Katok-Lindenstrauss theorem in his mathematical blog “Disquisitiones
Mathematicae” (http://matheuscmss.wordpress.com/) in a near future.
Acknowledgments
The first two chapters are strongly based on our book [1] “Aspectos
Ergódicos da Teoria dos Números” (in Portuguese) for a mini-course of
the 26th Brazilian Mathematical Colloquium, which was, in turn, based
on our discussions with Jimmy Santamaria in an Ergodic-Theory seminar
at IMPA around Green-Tao theorem. In particular, we would like to thank
Jimmy Santamaria for his decisive participation in this seminar. Also, we
would like to thank Yuri Lima and Jimmy Santamaria for a careful revision
of previous versions of this work.
Furthermore, this book was written during the pos-doctoral stay of
Carlos Matheus at Collège de France (under the supervision of Jean-
Christophe Yoccoz). In particular, he is thankful to Jean-Christophe for
the several interesting mathematical (and non-mathematical) discussions,
advices and friendship, and the two secretaries of the Mathematics section
of College de France, Dominique Bidois and Veronique Sainz, for their
invaluable support and friendship. Moreover, Carlos Matheus is grateful
to his beloved wife Aline Gomes Cerqueira for her patience and support
(both mathematical and non-mathematical).
Last, but not least, it is a great pleasure to acknowledge Étienne Ghys,
Maria Eulália Vares and Vladas Sidoravicius, current editors of “Ensaios
Matemáticos”, for their invitation to start this project (after our mini-
course during the 26th Brazilian Math. Colloquium). In particular, we
are grateful to Maria Eulália Vares whose strong encouragement and kind
words kept us working on this project.

A. Arbieto, C. Matheus, C. Moreira,

Rio de Janeiro and Paris, Oct. 30, 2009
Contents

1 Additive properties of prime numbers 13

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Classical problems . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.1 The twin prime conjecture . . . . . . . . . . . . . . . 14
1.2.2 Goldbach conjecture . . . . . . . . . . . . . . . . . . 14
1.2.3 Some results about arithmetical progressions vs. prime
numbers . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Arithmetic progressions in certain subsets of Z . . . . . . . 16
1.3.1 Van der Waerden theorem . . . . . . . . . . . . . . . 16
1.3.2 Szemerédi theorem . . . . . . . . . . . . . . . . . . . 16
1.3.3 Prime number theorem . . . . . . . . . . . . . . . . 17
1.3.4 Erdös-Turán conjecture . . . . . . . . . . . . . . . . 18
1.4 Proof of Van der Waerden theorem . . . . . . . . . . . . . . 18
1.4.1 Combinatorial proof . . . . . . . . . . . . . . . . . . 19
1.4.2 Dynamical proof . . . . . . . . . . . . . . . . . . . . 20
1.5 Furstenberg theorem and its applications . . . . . . . . . . 22
1.5.1 Crash course in Ergodic Theory . . . . . . . . . . . . 22
1.5.2 Furstenberg theorem . . . . . . . . . . . . . . . . . . 24
1.5.3 Szemerédi theorem via multiple recurrence . . . . . 26
1.6 Quantitative Szemerédi theorem . . . . . . . . . . . . . . . 27
1.7 Further results . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.7.1 von Mangoldt function . . . . . . . . . . . . . . . . . 31
1.7.2 Constellations of primes and polynomial progressions 33
1.7.3 Gaps in the set of prime numbers . . . . . . . . . . . 33
1.7.4 Magnitude of N0 (k, δ) . . . . . . . . . . . . . . . . . 34
1.8 Appendix to Chapter 1 . . . . . . . . . . . . . . . . . . . . . 34
1.8.1 Proof of Theorem 1.4.3 . . . . . . . . . . . . . . . . 34
1.8.2 F. Behrend’s example . . . . . . . . . . . . . . . . . 36

2 Green-Tao-Szemerédi theorem 38
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

9
10 CONTENTS

2.2 Strategy of the proof of the Green-Tao

theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.1 Proof of Green and Tao theorem . . . . . . . . . . . 42
2.2.2 Some comments . . . . . . . . . . . . . . . . . . . . 43
2.3 Proof of Roth theorem . . . . . . . . . . . . . . . . . . . . . 43
2.4 Proof of Green-Tao-Szemerédi theorem . . . . . . . . . . . . 51
2.4.1 Gowers norms . . . . . . . . . . . . . . . . . . . . . . 51
2.4.2 Anti-Uniformity . . . . . . . . . . . . . . . . . . . . 58
2.4.3 Sigma-algebras generated by basic anti-uniform func-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.4.4 The energy increment argument . . . . . . . . . . . 65
2.4.5 End of the proof of Green-Tao-Szemerédi theorem . 70

3 Elkies-McMullen theorem 75
3.1 Distribution of sequences on the circle . . . . . . . . . . . . 75
3.2 Ergodic version of Elkies-McMullen theorem . . . . . . . . . 78
3.2.1 Scheme of the proof of Elkies and McMullen theorem 78
3.2.2 Some preliminary reductions . . . . . . . . . . . . . 80
3.2.3 Geometrical interpretation of Le N . . . . . . . . . . . 83
3.3 Study of L via Ratner theorems . . . . . . . . . . . . . . . . 85
3.3.1 Computation of F assuming Theorem 3.3.1 . . . . . 85
3.3.2 Theorem 3.3.1 and homogenous flows . . . . . . . . 91
3.4 Equidistribution of non-linear horocycles . . . . . . . . . . . 92
3.4.1 The distribution law of a loop . . . . . . . . . . . . . 95
3.4.2 Ratner theorem and classification of µ . . . . . . . . 97
3.4.3 Non-linearity and torsion points . . . . . . . . . . . 100
3.4.4 End of the proof of Theorem 3.4.3 . . . . . . . . . . 102

Bibliography 103
 Part I

Green-Tao theorem
Chapter 1

Additive properties of
prime numbers

1.1 Introduction
One of the oldest concepts in Mathematics is the the notion of prime
number. By definition, an integer number p is prime if it is divisible only
by 1 and by itself.
Prime numbers are important objects in Number Theory due the unique
factorization theorem (saying that any integer number can be written as
a product of prime numbers in an essentially unique way). Another ele-
mentary property of prime numbers is the fact that they are precisely the
integer numbers p such that Z/pZ is a field.
Obviously, due to the multiplicative character of the definition of prime
numbers, it is fairly easy to extract its multiplicative properties. For in-
stance, the product of two primes is certainly not a prime number and
there are no geometrical progressions of primes of length ≥ 3 formed only
by prime numbers.
On the other hand, the situation changes considerably when one poses
some questions about the additive character of the prime numbers. For
example, one can ask whether the sum of two prime numbers is still a
prime number. Of course, the answer is: it depends. In fact, 2+3=5 is
prime and 2+5=7 is prime, but neither 3+5=8 isn’t prime nor 7+2=9.
However, the Bertrand’s postulate says that, for every natural number N ,
there exists a prime number between N and 2N = N + N . In particular,
this shows that the following question deserves a little bit of attention:
Are there arithmetical progressions of length ≥ 3 formed only by prime
numbers? In the case of an affirmative answer to this question, how many
of them exist once the length of the arithmetical progression is fixed?

13
14 Chapter 1. Additive properties of prime numbers

We are going to see (in the first two chapters of this book) that this prob-
lem was solved by Ben Green and Terence Tao. However, before entering
this issue, let us take a little trip around the world of the prime numbers in
order to see some related questions about the additive properties of prime
numbers and its partial solutions.

1.2 Classical problems about the additive

properties of prime numbers
1.2.1 The twin prime conjecture
As the examples of the introduction showed us, the sum of a prime number
with 2 is not always a prime number. However, we can ask whether there
are infinitely many primes of this form. We say that p and p + 2 are twin
primes whenever both of them are prime numbers. Some examples of twin
primes are: (3 and 5), (5 and 7), (11 and 13), (17 and 19), (29 and 31), (41
and 43), etc. One of the most famous open conjectures in Number Theory
is the so-called twin prime conjecture:

Are there infinitely many twin prime numbers?

An important result due to Brun [2] gives a flavor of the difficulty of this
conjecture: namely, Brun proved that, even in the case of the existence of
infinitely many pairs of twin primes, it is a very hard task to locate them
because they are very rare. More precisely, Brun’s theorem says that the
series of the inverse of the twin primes converges (to a certain number
called Brun’s constant):

1 1 1 1 1 1 1 1
( + ) + ( + ) + ( + ) + ( + ) + ... < +∞.
3 5 5 7 11 13 17 19
Later on, we will reformulate this conjecture in a more analytical language.

1.2.2 Goldbach conjecture

In a letter addressed to Euler (in 1742), Goldbach asked whether every
integer number ≥ 2 is the sum of 3 prime numbers. In his formulation,
Goldbach assumed that 1 is a prime number, although nowadays this con-
vention is not used anymore. In modern terms, Goldbach’s conjecture can
be reformulated as:

Can we write every even integer number n ≥ 4 as a sum of 2 prime

numbers?
A. Arbieto, C. Matheus and C. G. Moreira 15

Although this conjecture is fairly easy to state, Goldbach’s conjecture

is still one of the big challenges in Number Theory. Nowadays, we have
several partial results, but none of them seem to extend to a full solution
of this conjecture.
For example, Schnirelman [13] showed that every integer number can be
written as a sum of prime numbers, where the quantity of terms can be
bounded by ∼ 300000 (a little bit far from 2, don’t you think?).
A related conjecture (also called Goldbach conjecture) is:
Can we write every odd integer number n ≥ 9 as a sum of 3 prime
numbers?
In this direction, we have a famous result of Vinogradov [20] solving
this conjecture for any sufficiently large odd integers (for instance, the
15
conjecture holds for any odd integer ≥ 33 ).
Another interesting result is Chen’s theorem [3] showing that any suffi-
ciently large even integer is the sum of a prime number and a quasi-prime
number (i.e., an integer number with 2 prime factors at most).
A stronger version of Goldbach conjecture (called Levy’s conjecture) is:
Can we write every odd integer n ≥ 7 as a sum of a prime number p and
the double 2q of another prime number q?
Later on, we will also reformulate these conjectures in an analytical
language.

1.2.3 Some results about arithmetical progressions vs.

prime numbers
A classical result in this subject is Dirichlet’s theorem:
If a and b are relatively prime, then the arithmetical progression a + nb
contains infinitely many primes.
During the proof of this result, Dirichlet introduce the important concept
of L-series. In particular, we will omit the proof of this beautiful theorem
in order to keep the coherence with the purpose of this book.
Observe that Dirichlet theorem doesn’t say that this arithmetical pro-
gression is entirely formed by prime numbers. A natural question arises:
are there arithmetical progressions of infinite length formed only by prime
numbers? The negative answer to this problem is provided by the following
theorem of Lagrange and Waring:
Consider an arithmetical progression formed only by prime numbers of
length k and ratio d. Then, d is necessarily divisible by every prime
number smaller than k. In particular, there is no infinite arithmetical
progression formed only by prime numbers.
16 Chapter 1. Additive properties of prime numbers

1.3 Arithmetical progressions in certain sub-

sets of Z
The question of existence of arithmetical progressions (AP) of finite length
of prime numbers can be extended as follows:

Let A ⊂ Z be a given infinite subset of integers. Are there APs of

arbitrarily large length formed only by elements of A?

In a certain sense, we will see that the subset P of prime numbers is

very “thin”. Thus, we can start to attack the previous question by the
“easy” case of “fat” subsets A, where there is a nice chance of finding
AP, and then one could try to adapt the method to work with “thin” sets
(such as the set of primes P ). Of course, the central problem lies in the
formalization of the definition of “thin” and “fat” subsets. In this section,
we will revise some results in this direction, although we should stress that
we are not going to follow the chronological order in the exposition of these
theorems.

1.3.1 Van der Waerden theorem

Suppose that you are given a finite number, say k, of colors and you use
them to color the integer numbers. In this process, you get a partition
of the integer numbers into k disjoint subsets. Van der Waerden theorem
says that:

At least one of these subsets is so “fat” that it should contain arbitrarily

large arithmetical progressions.

In particular, if we take two colors and we give one color to the set of
prime numbers P and the other color to the composite (i.e., non-prime)
numbers, we obtain:

Either the set of prime numbers or the set of composite numbers

possesses arbitrarily large arithmetical progressions.

Later in this chapter, we will see two proofs of Van der Waerden theorem
(one of them is combinatorial and the other is ergodic-theoretical).

1.3.2 Szemerédi theorem

Logically, the subset of even integers possesses arithmetical progressions
of arbitrarily large length (with ratio 2, for instance). Observe that the
even integers occupy essentially 1/2 of any interval [1, N ] := {n ∈ Z; 1 ≤
n ≤ N }. Similarly, the odd integers share the same property and they
A. Arbieto, C. Matheus and C. G. Moreira 17

also possess arbitrarily large arithmetical progressions. More generally, we

can pick an integer number k and we can look at the subset of multiples
of k: this subset essentially occupies 1/k of any interval [1, N ] and it has
arbitrarily large arithmetical progressions.
Based on these simple remarks, we are tempted to say that a subset A
is “fat” if it occupies a definite positive fraction of the intervals [1, N ] for
a infinite sequence of N s:

Definition 1.3.1. Let A ⊂ N. The (upper) density of A is:

|[1, N ] ∩ A|
d(A) = lim sup .
N →∞ N

Here, |B| denotes the cardinality of a given B ⊂ N.

Of course, this definition extends to subsets of integer numbers. The

first theorem dealing with the existence of arithmetical progressions in
“fat” subsets (i.e., subsets with positive density) is Roth’s theorem [12]
(1956):

If A ⊂ Z has positive density, then A contains infinitely many arithmetic

progressions of length 3.

In general, the problem of existence of arbitrarily large arithmetical pro-

gressions in a positive density subset was solved by Szemerédi [14] (1975):

Theorem 1.3.1 (Szemerédi). If A ⊂ Z has positive density, then A has

arbitrarily large arithmetical progressions.

The extension of Szemerédi theorem to a more general context is the

subject of the chapter 2 of this book (since, as stated, this theorem can’t
be applied to the subset of prime numbers). In the subsequent sections,
we will give a proof of Szemerédi theorem assuming an important result
of Ergodic Theory (namely, Furstenberg multiple recurrence theorem).

1.3.3 Prime number theorem

The basic obstruction to the application of Szemerédi theorem to the sub-
set of prime numbers is provided by the famous prime number theorem:

Theorem 1.3.2 (Prime number theorem). It holds:

|P ∩ [1, N ]| 1
= (1 + o(1)).
N log N

Here P is the subset of prime numbers and o(1) is a quantity converging

to zero when N → ∞. In particular, d(P ) = 0.
18 Chapter 1. Additive properties of prime numbers

Although the prime numbers form a subset of zero density, the existence
of infinitely many arithmetical progressions of length 3 formed only by
prime numbers was showed in 1939 by Van der Corput (before Roth’s
theorem):

There are infinitely many arithmetical progressions of length 3 composed

only by prime numbers.

Finally, in 2004, Ben Green and Terence Tao [9] proved the general result
(about arbitrarily large arithmetic progressions formed only by primes).
This theorem is the main object of the first two chapters of this book:
Theorem 1.3.3 (Green-Tao). The primes contains arbitrarily large arith-
metical progressions.

1.3.4 Erdös-Turán conjecture

P 1 2
We know that n2 converges (to π /6), but, in 1737, Euler showed that
the serie of the inverse of the primes is divergent:
X 1
= +∞.
p
p prime

This means that the prime numbers are less sparse than the perfect squares.
The Erdös-Turán conjecture claims that any set with this property con-
tains arbitrarily large arithmetical progressions (so that Green-Tao theo-
rem solves a particular case of this conjecture):
Conjecture 1 (Erdös-Turán). Let A ⊂ N be a subset such that
X 1
= +∞.
n
n∈A

Then, A contains arbitrarily large arithmetical progressions.

This conjecture is completely open (in general): we don’t know even
when such subsets contain arithmetic progressions of length 3.

1.4 Proof of Van der Waerden theorem

During this section, we will present two proofs of Van der Waerden theo-
rem:
Theorem 1.4.1 (Van der Waerden). For any coloring of the integers with
a finite number of colors m, we can find arbitrarily large monochromatic
(i.e., one color) arithmetic progressions.
A. Arbieto, C. Matheus and C. G. Moreira 19

1.4.1 Combinatorial proof

In this subsection we prove Van der Waerden theorem via the coloring
method in Combinatorics. In order to alleviate the notation, we denote
the arithmetical progression a, a + r, . . . , a + (k − 1)r by a + [0, k)r, and
we assume that one disposes of m colors to assign to the natural numbers
from 1 to N .
Definition 1.4.1. Let c : {1, . . . , N } → {1, . . . , m} be a coloring. Given
k ≥ 1, d ≥ 0 and a ∈ {1 . . . , N }, a fan of radius k, degree d with base point
a is a d-tuple of arithmetic progressions (a+[0, k)r1 , . . . , a+[0, k)rd ) where
r1 , . . . , rd > 0. For each 1 ≤ i ≤ d, the progressions a + [1, k)ri are called
spokes of the fan. We say that a fan is polychromatic if its base point and
its spokes are monochromatic, i.e., there are distinct colors c0 , c1 , . . . , cd
such that c(a) = c0 and c(a + jri ) = ci for j = 1, . . . , k and i = 1, . . . , d.
Remark 1.4.1. Observe that, by the distinction between the colors, if we
have m colors, it is not possible to construct a polychromatic fan whose
degree is ≥ m.
Of course, we see that the van der Waerden theorem is a direct conse-
quence of the following result:
Theorem 1.4.2. Let k, m ≥ 1. Then, there exists N such that any col-
oring of {1, . . . , N } with m colors contains a monochromatic arithmetic
progression of length k.
Proof. The argument consists into a double induction scheme. Firstly, we
make an inductive argument on k: observe that the case k = 1 is trivial, so
that we can take k ≥ 2 and we can assume that the theorem holds for k−1.
Secondly, we perform an induction on d, i.e., we will show the following
claim by induction: given d, there exists N such that for any coloring
of {1, . . . , N } with m colors, we have either a monochromatic arithmetic
progression of length k or a polychromatic fan of radius k and degree d.
Note that the case d = 0 is trivial and once we prove this claim for d = m,
one can use the remark 1.4.1 in order to obtain the desired monochromatic
arithmetic progression of length k (so that the double inductive argument
is complete).
Let us take d ≥ 1 and suppose that this claim is true for d − 1. Let
N = 4kN1 N2 , where N1 and N2 are large integers to be chosen later,
and consider A = {1, . . . , N }. Fix c : {1, . . . , N } → {1, . . . , m} a coloring
of A. Obviously, {bkN1 + 1, . . . , bkN1 + N1 } is a subset of A with N1
elements for each b = 1, . . . , N2 . By our inductive hypothesis on k and d,
if N1 is sufficiently large, we can find either a monochromatic arithmetic
progression of length k or a polychromatic fan of radius k and degree d−1.
Of course, if we find a monochromatic arithmetic progression of length
k inside {bkN1 + 1, . . . , bkN1 + N1 } for some b = 1, . . . , N2 , we are done.
20 Chapter 1. Additive properties of prime numbers

Thus, one can suppose that we find a polychromatic fan inside {bkN1 +
1, . . . , bkN1 + N1 } for every b = 1, . . . , N2 . In other words, for each b =
1, . . . , N2 , we have a(b), r1 (b), . . . , rd−1 (b) ∈ {1, . . . , N1 } and distinct colors
c0 (b), c1 (b), . . . , cd−1 (b) ∈ {1, . . . m} such that c(bkN1 + a(b)) = c0 (b) and
c(bkN1 +a(b)+jri (b)) = ci (b) for every j = 1, . . . , k−1 and i = 1, . . . , d−1.
We say that these are the first and second properties of the fan associated
to b. In particular, the map

b → (a(b), r1 (b), . . . , rd−1 (b), c0 (b), . . . , cd−1 (b))

is a coloring with md N1d colors of the set {1, . . . , N2 }. Using again our in-
ductive hypothesis on k, if N2 is sufficiently large, there exists some arith-
metic progression b+[0, k−1)s which is monochromatic with respect to this
new coloring, say that its color has the form (a, r1 , . . . , rd−1 , c1 , . . . , cd−1 ).
Up to reversing the position of the progression, we can suppose that s is
negative.
At this point, the idea is to convert this huge progression of identical
polychromatic fans of degree d − 1 (in the sense that their combinatorial
type is fixed by the coloring (a, r1 , . . . , rd−1 , c1 , . . . , cd−1 )) in a new poly-
chromatic fan with degree d in order to close the inductive argument. Let
b0 = (b − s)kN1 + a ∈ {1, . . . , N } and consider:

(b0 + [0, k)skN1 , b0 + [0, k)(skN1 + r1 ), . . . , b0 + [0, k)(skN1 + rd−1 )).

We affirm that this is a fan of radius k, degree d and base point b0 .

Indeed, let us verify that the spokes are monochromatic. In the first
spoke we have c(b0 + jskN1 ) = c((b + (j − 1)s)kN1 + a) by direct sub-
stitution. By the first property of the fan associated to b + (j − 1)s, it
follows that c((b + (j − 1)s)kN1 + a) = c0 (b + (j − 1)s) = c0 (b) (since the
arithmetic progression b + [0, k − 1)s is monochromatic if 1 ≤ j ≤ k − 1).
Similarly, in an arbitrary spoke, using the second property of the fans, we
have that, if 1 ≤ j ≤ k − 1 and 1 ≤ t ≤ d, then

c(b0 + j(skN1 + rt )) = c((b + (j − 1)s)kN1 + a + jrt ) = ct (b + (j − 1)s) = ct .

If the base point b0 has the same color of a spoke, we found a monochro-
matic arithmetic progression of length k. Otherwise, the base point has
a distinct color from the spokes, so that we found a polychromatic fan of
radius k and degree d. This ends the inductive step and, a fortiori, the
proof of the theorem.

1.4.2 Dynamical proof

An useful tool in Dynamical Systems is the so-called symbolic dynamics
consisting on the study of a specific map called shift. In the sequel, we
A. Arbieto, C. Matheus and C. G. Moreira 21

will introduce the precise definition of the shift map and we will see how
this important tool was applied by Furstenberg and Weiss to give a proof
of van der Waerden theorem.
Let A = {a1 , . . . , ak } be a finite alphabet. Consider the set Ω of all
infinite words obtained from the letters of this alphabet:

Ω = {(x1 , x2 , . . . , xn , . . . ) ; xi ∈ A, ∀ i}.

This set has a natural structure of metric space with respect to the follow-
ing distance: given x = (x1 , x2 , . . . ) and y = (y1 , y2 , . . . ), define
1
d(x, y) := if l is the smallest integer such that xl 6= yl .
l
The shift map T : Ω → Ω is:

T (x1 , x2 , x3 , . . . ) = (x2 , x3 , x4 , . . . ).

It is a simple exercise to show that the shift map is continuous with respect
to the distance d.
From these concepts, Furstenberg proved the van de Waerden theorem
via the following topological dynamical theorem (whose complete proof is
presented in the appendix to this chapter):
Theorem 1.4.3 (Topological Multiple Recurrence - Furstenberg and Weiss).
Let T : X → X be a continuous dynamical system on a compact metric
space X. For all k ∈ N and ε > 0, there exist x ∈ X and n ∈ N such that
d(T in (x), x) < ε for every i = 1, . . . , k. Moreover, given any dense subset
Z ⊂ X, we can take x ∈ Z.
Assuming this result, let us see how one can prove van der Waerden
theorem. Let A = {c1 , . . . , cs } be the set of colors and z = (z1 , z2 , z3 , . . . ) ∈
AN a given coloring of N, where zi ∈ A is the color of the integer i. Consider
T : AN → AN the shift map. From the definition of the distance d, we have
that, for x, y ∈ AN and m, l ∈ N, it holds d(T m (x), T l (y)) < 1 if and only
if xm+1 = yl+1 .
In particular, for a given coloring z ∈ AN , an arithmetic progression
m, m + n, . . . , m + kn is monochromatic if and only if zm = zm+n = · · · =
zm+kn , that is, if and only if:

d(T m−1 (z), T m−1+in (z)) = d(T m−1 (z), T in (T m−1 (z)))
< 1, for i = 1, . . . , k.

Taking X = {T m (z)}∞ m=0 , we see that X is a compact metric space, T is

a continuous dynamical system on X and the subset Z = {T m (z)}∞ m=0 is
dense in X. Thus, the van der Waerden theorem follows directly from the
topological multiple recurrence theorem (Theorem 1.4.3).
22 Chapter 1. Additive properties of prime numbers

1.5 Furstenberg theorem and its application

to Szemerédi theorem
In this section we present a proof of Szemerédi theorem inspired by the
dynamical proof of van der Waerden theorem: firstly, we will make a quick
revision of basic elements in Ergodic Theory, then we will state a deep
multiple recurrence result of Furstenberg, and finally we will get Szemerédi
theorem as a consequence of Furstenberg theorem.

1.5.1 Crash course in Ergodic Theory

Ergodic Theory studies the statistics of the dynamics of a (measurable)
map T : X → X, where X is a probability space, from the point of view
of a T -invariant probability measure µ (i.e., for any measurable subset A
we have µ(A) = µ(T −1 (A))).
Usually, the mere existence of a T -invariant probability measure gives
us a lot of information about the statistics of generic orbits of T (i.e., the
subsets {T n (x)}∞n=0 , for almost every x ∈ X with respect to µ). For in-
stance, Poincaré recurrence theorem says that if T : X → X is µ-invariant
and µ(A) > 0, then for µ-almost every x ∈ A, there exists n(x) ≥ 1 such
that T n(x) (x) ∈ A. Consequently, there exists N such that

µ(A ∩ T −N (A)) > 0.

In particular, we see that, independently of the size of a given subset, if it

has positive measure, then there are plenty of orbits starting at this subset
and coming back to it infinitely many times. In the case of a topological
probability space, one can reformulate Poincaré recurrent theorem as:

Let T : X → X be a dynamical system of a probability space (X, µ).

Assume that X is also a compact metric space and µ is T -invariant.
Then, µ-almost every point is recurrent, i.e., for a µ-generic point x,
there exists a sequence nk → ∞ of natural numbers such that
d(T nk (x), x)) → 0 when k → ∞.

After knowing that the dynamics T enjoys nice statistical properties

once it has an invariant probability measure, a natural question arises:
which dynamical systems possess invariant probability measures? When
the space X is compact and T is continuous, the answer is yes. The idea of
the proof of this fact is quite simple: take any probability measure (e.g., a
Dirac measure at some arbitrary point) and let us analyze the evolution of
this measure under the dynamics (that is, by the action of the iterates of
T ). By picking the average of the measures obtained up to a large iterate
N , we hope to get an almost insensitive probability measure with respect
to the action of T (as N grows). Thus, the natural argument is to consider
A. Arbieto, C. Matheus and C. G. Moreira 23

an accumulation point of this sequence of almost insensitive measures and

(by crossing fingers) one can expect that the limit probability measure is
the desired T -invariant probability.
Now we put some details into the previous rough scheme. Firstly, since
our scheme involves the process of taking limits of probability measures,
let us introduce a notion of convergence of probability measures. Since
the space of (Radon) measures is the dual space of continuous functions,
it is natural to use the weak-* topology (because the Functional Analysis
results can help us with the compactness issues we are going to face in a
few moments).

Definition 1.5.1. We say that a sequence of measures µk on X converges

(weakly-*) to µ whenever, for any continuous function f : X → R, it holds

Z Z
f dµk → f dµ.
X X

Because this is the so-called weak-* topology on the concrete space of

Radon measures, we can apply Banach-Alaoglu theorem to obtain:

The space of probability measures on X is compact with respect to the

weak-* topology.

Now let us come back to the question of existence of invariant measures.

Let η be an arbitrary probability measure. The action of T on η occurs by
push-forward, i.e., ((T n )∗ η)(A) := η(T −n (A)) for every measurable subset
A. A simple observation is: a given probability η is T -invariant if and only
if T ∗ η = η.
Next, let us consider the sequence of probabilities

k−1
1X i ∗
µk = (T ) η.
k i=0

In other words, we are taking temporal averages of the measures ob-

tained by the push-forward of η under the first k − 1 iterates. By compact-
ness, there is a convergent subsequence µnk accumulating some probability
measure µ. We claim that µ is invariant. In fact, we have the following
equalities (explained below):
24 Chapter 1. Additive properties of prime numbers

T ∗ µ = T ∗ (lim µnk )
= lim(T ∗ (µnk ))
nk −1
1 X
= lim( (T i+1 )∗ (η))
nk i=0
nk
1 X
= lim( ( (T i )∗ (η) − η + (T nk )∗ η))
nk i=0
nk
1 X
= lim (T i )∗ (η)
nk i=0
= µ.
In the second equality, we used the fact that the push-forward opera-
tor T ∗ is continuous in the weak-* topology. This is true because T is
continuous: indeed, suppose that µk → µ weakly-* and fix a continuous
f : X → R, so that we have that f ◦ T is also continuous and, a fortiori,
Z Z Z Z
∗
f d(T µk ) = f ◦ T dµk → f ◦ T dµ = f d(T ∗ µ).
X X X X

In the fifth equality, we observe that, for every continuous f : X → R,

by compactness of X, we have:
Z Z Z
1 1 nk ∗ 1
f dµ → 0 and f d((T ) µ) = f ◦ T nk dµ → 0.
nk X nk X nk X
Hence, the two last parts of this sum go to zero (weakly-*).
A concrete interesting example for our purposes is the shift map T on
X = {0, 1}N . Considering a Dirac measure associated to a point x ∈ X,
that is, δx (A) = 0 if x ∈
/ A and δx (A) = 1 if x ∈ A, then we know that
Pk−1
the sequence µk = k1 j=0 δT j (x) accumulates (weakly-*) some probability
measure and any such accumulation point is an invariant measure of the
shift map.

1.5.2 Furstenberg theorem

Going back to Poincaré recurrence theorem, given a positive measure sub-
set A, one can ask whether there is some structure on the set of return
times to A. More precisely, we know that this set is infinite, but, does
it have any arithmetic structure? For instance, can it coincide with the
set of prime numbers?1 In this direction, a beautiful result of Furstenberg
(called Furstenberg multiple recurrence theorem) gives us a precise answer:
1 Actually, Birkhoff theorem says that the density of the set of return times is positive,

so that this set can’t coincide with the prime numbers.

A. Arbieto, C. Matheus and C. G. Moreira 25

Theorem 1.5.1 (Furstenberg’s multiple recurrence theorem). Let T :

X → X be a µ-invariant map, k ≥ 3 an integer and µ(A) > 0. Then,
there exists N such that

µ(A ∩ T −N (A) ∩ · · · ∩ T −(k−1)N (A)) > 0.

This deep theorem is the heart of Furstenberg ergodic-theoretical proof

of Szemerédi theorem. Unfortunately, the complete proof of this result
would lead us too far away from the scope of this book, so that we will
content ourselves with a proof of this result into two important represen-
tative cases.
The first case is the Bernoulli system. Again, T is the (full) shift map
on the set X = AN , where A is a finite P alphabet. Next, we pick p1 , . . . , pr
non-negative real numbers such that pi = 1. This provides a probability
measure on A and, by taking the associated product measure, we get a
probability measure on X. The dynamical system T equipped with this
probability measure is called Bernoulli system.
By definition, the product σ-algebra is generated by the cylinders subsets
one obtains by fixing a finite number n of coordinates, i.e., a cylinder is a
subset of the form C = {w ∈ Z; wi1 = j1 , . . . , win = jn } and its Bernoulli
measure is µ(C) = pj1 . . . pjn . After extending this definition to the whole
σ-algebra, it is a simple task to show that the Bernoulli measure is invariant
by the shift map (in fact, it suffices to prove that µ(B) = µ(T −1 (B)) when
B is a cylinder and this fact is easy to check).
In the same manner, since the cylinders generate the σ-algebra, it suffices
to show Furstenberg multiple recurrence theorem to every cylinder subset.
Consider C0 , C1 , . . . , Ck some cylinders and observe that, for a sufficiently
large integer n, the fixed coordinates in the definitions of the cylinders
T −nl (Cl ) are distinct. Hence, we obtain:

µ(C0 ∩ T −n (C1 ) ∩ · · · ∩ T −kn (Ck )) = µ(C0 )µ(C1 ) . . . µ(Ck ) > 0.

This proves Furstenberg theorem in this first example.

Another example is a periodic system, i.e., a dynamical system T such
that T p = T for some p. In this case, Furstenberg theorem is totally
trivial. A slightly less trivial dynamics (still along this line of reasoning)
is the following quasi-periodic example: the space X = S 1 = R/Z is the
circle, µ is the Lebesgue measure and T (x) = x + α(mod 1) is a rotation
of α (of course, the name quasi-periodic is motivated by the fact that T
is periodic if and only if α is a rational number and the dynamics of T
is close to periodic when α is irrational by taking approximations of α by
rational numbers).
R Given a measurable subset A with µ(A) > 0, note that the function
1A (x + y)dµ(x) is continuous on y. Hence, for all ε > 0, there exists δ
26 Chapter 1. Additive properties of prime numbers

such that, if |y| < δ, then µ(A ∩ (A − y)) > µ(A) − ε. Hence

µ(A ∩ (A − y) ∩ (A − 2y) ∩ · · · ∩ (A − ky)) > µ(A) − (k + 1)ε.

Choosing ε < µ(A)

k+1 and fixing the corresponding δ, we can define Dδ =
{n ≥ 1; nα ∈ (−δ, δ)(mod 1)}. Observe that, if n ∈ Dδ , it holds:

µ(A ∩ T −n (A) ∩ · · · ∩ T −nk (A)) > µ(A) − (k + 1)ε > 0.

This proves Furstenberg theorem in this second example.

Notice that the first example belongs to the class of weak-mixing sys-
tems, i.e., the class of dynamical systems verifying the following equality
N
1 X
lim (µ(A ∩ T −n B) − µ(A)µ(B))2 = 0
N →∞ N
n=1

for any two measurable subsets A and B. By an adaptation of the ideas

used to treat the case of Bernoulli shifts, it is possible to show that Fursten-
berg theorem holds for any weak-mixing (i.e., pseudorandom) system.
On the other hand, the second example belongs to the class of compact
systems, i.e., the class of dynamical systems such that, for any function
f ∈ L2 (µ), the closure of {f, T f, T 2 f, . . . , T n f, . . . } is compact in L2 .
Again, by an adaptation of the ideas used to deal with the case of quasi-
periodic rotations of the circle, it is possible to prove that Furstenberg
theorem also holds for any compact (i.e., structured) system.
In general, Furstenberg theorem is a consequence of the so-called Fursten-
berg structure theorem saying that we can decompose an arbitrary dynami-
cal system into several levels (i.e., factors) along a tower of extensions such
that each level (factor) is weakly mixing or compact and any two distinct
factors doesn’t correlate much. In fact, since these factors don’t interact
and we know Furstenberg theorem for weakly-mixing and compact factors,
we are morally done. Evidently, the existence of this tower of extensions
is a highly non-trivial fact beyond the scope of this book (so that we will
end our comments on Furstenberg theorem here). However, it is worth to
point out that this structure theorem will reappear later (on chapter 2) in
a finitary context during the proof of Green-Tao theorem.

1.5.3 Szemerédi theorem via multiple recurrence

Once we have Furstenberg multiple recurrence theorem in our toolbox, one
can quickly give a proof of Szemerédi theorem using the shift dynamics (in
the spirit of the ergodic proof of van der Waerden theorem).
Put X = {0, 1}N and let T : X → X be the shift. Take (xn ) =
(1A (n)), where 1A (x) is the characteristic function of A, and consider
A. Arbieto, C. Matheus and C. G. Moreira 27

Pk−1
µk = k1 j=0 δT j (x) . Then, as we already know, up to passing to a sub-
sequence, we can assume that µ = lim µk is a T -invariant probability
measure.
Define Y = {(yn ); y1 = 1}. Since Y is a compact subset, we have
that µ(Y ) = lim µk (Y ) = lim k1 |A ∩ [1, k]| > 0 (by hypothesis). Thus, by
Furstenberg multiple recurrence theorem, it follows that there exists N
such that µ(Y ∩ T −N (Y ) ∩ · · · ∩ T −(k−1)N (Y )) > 0. In particular, there
is z ∈ Y ∩ T −N (Y ) ∩ · · · ∩ T −(k−1)N (Y ). I.e., there exists some integer x
such that x, x + N, . . . , x + (k − 1)N ∈ A. This ends the proof of Szemerédi
theorem.

1.6 Quantitative Szemerédi theorem

In this section, we will see some reformulations of Szemerédi theorem 1.3.1
(which are very useful for our future purposes).
Let’s start with the remark that Szemerédi theorem is equivalent to the
following statement:

For any k ≥ 1 and 0 < δ ≤ 1, there exists a large integer NSZ (k, δ) ≥ 1
such that, for every N ≥ NSZ , any subset A ⊂ {1, . . . , N } of cardinality
|A| ≥ δN contains some arithmetic progression of length k.

Logically, this statement is a priori certainly stronger than Szemerédi

theorem.
On the other direction, we will use an abstract non-sense argument:
admitting that the statement is false for a certain pair (k, δ), we affirm
that there is a subset Y ⊂ N∗ satisfying |Y ∩ {1, 2, . . . , n}| ≥ δr, ∀ r ∈ N∗
so that Y doesn’t contain any arithmetical progression of length k.
In order to prove this affirmation, let us first prove that the non-existence
of NSZ (k, δ) implies that, for each n ∈ N∗ , there exists a subset Xn ⊂
{1, 2, . . . , n} verifying |Xn ∩ {1, 2, . . . , k}| ≥ δk for 1 ≤ k ≤ n such that
Xn doesn’t contain any arithmetic progression of length k. In fact, let
εn = max ((⌈δk⌉ − 1)/k) < δ. We claim that, if N is sufficiently large and
1≤k≤n
A ⊂ {1, 2, . . . , N } has cardinality |A| ≥ δN , then there exists m ≤ N − n
such that |A ∩ {m + 1, . . . , m + k}| ≥ δk, for each 1 ≤ k ≤ n. Indeed,
if this is not the case, there are s ∈ N∗ , k1 , k2 , . . . , ks ∈ {1, 2, . . . , n}
such that N ≥ k1 + k2 + · · · + ks > N − n and, for 1 ≤ r ≤ s, it holds
P P  1 P P 
|A∩ kj , kj | < δkr , so that |A∩ kj , kj | ≤ εn < δ, and, a
j<r j≤r kr j<r j≤r
n
fortiori, δN ≤ |A| ≤ n + εn · N , an absurd when N > · Observe that
δ − εn
this proves the desired affirmation about the existence of Xn (assuming
that NSZ (k, δ) doesn’t exist) because, by the previous discussion, it suffices
to consider an appropriate translation of a certain subset of any A ⊂
28 Chapter 1. Additive properties of prime numbers

{1, 2, . . . , N } with cardinality |A| ≥ δN such that A doesn’t contain any

arithmetic progression of length k.
Next, once we know about the existence of these special subsets Xn ,
let us cook up the desired Y . To do so, for each r ∈ N∗ , let πr : 2N →
2{1,2,...,r} be given by πr (A) = A ∩ {1, 2, . . . , r}. We construct inductively
some sets Y1 , Y2 , Y3 , . . . with Yr ⊂ {1,2,. . ., r} for each r ∈ N∗ such that
Yr+1 ∩ {1, 2, . . . , r} = Yr , ∀ r ∈ N∗ in the following way. Firstly, we
put Y1 := {1} ⊂ Xn , for all n ∈ N∗ . Next, given Yr , r ∈ N∗ such that
Yr = πr (Xn ) for infinitely many n ∈ N, there exists Yr+1 ⊂ {1, 2, . . . , r+1}
with Yr+1 ∩ {1, 2, . . . , r} = Yr such that Yr+1 = πr+1 (Xn ) for infinitely
many n ∈ N (indeed, if πr (Xn ) = Yr , there are S only two possibilities for
πr+1 (Xn )). Now, it is easy to see that Y = Yn verifies πr (Y ) = Yr ,
n∈N∗
∀ r ∈ N∗ , so that πr (Y ) = πr (Xn ) for infinitely many n ∈ N. In particular,
|Y ∩ {1, 2, . . . , r}| ≥ δr, ∀ r ∈ N∗ and Y doesn’t contain any arithmetic
progression of length k.
After this, we introduce a more analytical and finitary language in or-
der to get another reformulation of Szemerédi theorem (Theorem 1.3.1).
Keeping this goal in mind, we recall the following definition:

Definition 1.6.1. Let f : A → C be an arbitrary function where A is a

finite set. Then, the expectation of f is
1 X
E(f ) = E(f (n); n ∈ A) = f (n).
|A|
n∈A

Given f : (Z/N Z) → R any function, we can define the shift T n f :

(Z/N Z) → R of this function f by an integer n ∈ Z/N Z (or n ∈ Z) via
the formula T n f (x) := f (x + n).
Using this notation, we can reformulate Szemerédi theorem as follows:

Theorem 1.6.1 (Szemerédi theorem – quantitative version). For every

integer number k ≥ 1 and real number 0 < δ ≤ 1, there are N0 (k, δ) a
large integer and c(k, δ) > 0 a small real number such that, for any N ≥
N0 (k, δ) a large prime number and any f : Z/N Z → R+ with 0 ≤ f ≤ 1,
E(f |Z/N Z) ≥ δ, it holds
k−1
Y
E( T jr f (x)|x, r ∈ Z/N Z) ≥ c(k, δ).
j=0

Remark 1.6.1. Concerning this quantitive version of Szemerédi theo-

rem, we will construct in the appendix to this chapter some examples
due to F.√ Behrend of some subsets S of the interval [1, N ] such that
1−2 2 log 2+ε
√
|S| ≥ N log N and S doesn’t contains arithmetic progressions of length
A. Arbieto, C. Matheus and C. G. Moreira 29

3. Moreover, by a slight modification of the scheme of Behrend’s argu-

ment, we will see that, concerning the behavior of c(k, δ) above with re-
spect to δ, one can’t expect that c(k, δ) has a polynomial behavior in the
variable δ (i.e., c(k, δ) ≥ δ Ck for some Ck > 0): indeed, we will show that
c(3, δ) ≤ δ c log(1/δ) .
Observe that the statement of theorem 1.6.1 provides (a priori ) a much
stronger conclusion than the usual Szemerédi theorem. In fact, while the
usual Szemerédi allows only to conclude the existence of one k-AP (i.e.,
arithmetic progression of length k), the quatitative version permits to infer
the existence of c(k, δ)N 2 k-APs (at least). However, although the quan-
titative Szemerédi theorem is apparently better than the usual one, we
claim that the theorems 1.3.1 and 1.6.1 are equivalents.
We start by showing that the usual Szemerédi theorem follows from its
quantitative version. Fix k, δ and take N a large prime number. Let
us suppose that A ⊂ {1, . . . , N } has cardinality |A| ≥ δN (this is plau-
sible since A has positive density). Pick N ′ a prime number between
kN and 2kN (its existence is assured by Bertrand’s postulate). Consider
{1, . . . , N } as a subset of Z/N ′ Z and denote by A′ the subset of Z/N ′ Z
corresponding to A.
By our choices, we have E(1A′ |Z/N Z) ≥ δ/2k. By the quantitative
version of Szemerédi theorem, it follows that:
k−1
Y
E( T jr 1A′ (x)|x, r ∈ Z/N ′ Z) ≥ c(k, δ/2k).
j=0

Rewriting this expression, we get:

|{(x, r) ∈ (Z/N ′ Z)2 ; x, x + r, . . . , x + (k − 1)r ∈ A′ }| ≥ c(k, δ/2k)(N ′ )2 .

Since N ′ ≥ kN and A′ ⊂ {1, . . . , N }, we have that 1 ≤ x ≤ N and −N ≤

r ≤ N . Observe that the contribution of the (degenerate) progressions
with r = 0 is N (at most). Removing these degenerate progressions and
taking N large, the right-hand side is still positive, so that A must contain
a progression x, x + r, . . . , x + (k − 1)r.
Now let us prove that the usual Szemerédi theorem implies its quanti-
tative version. Recall that we know that the usual version is equivalent to
its finitary version, i.e., the existence of the integer NSZ (k, δ), ∀ k ∈ N∗ ,
δ > 0 with the previously discussed properties. Thus, our task is reduced
to the proof of the following proposition:
δ

Proposition 1.6.1. Suppose that the integer NSZ k, with the previous
2
properties does exist. Then, there are N0 ∈ N and α(k, δ) > 0 such that,
if N ≥ N0 for all A ⊂ {1, 2, . . . , N } with |A| ≥ δN , there are α(k, δ) N 2
k-APs, i.e., arithmetic progressions of length k (at least).
30 Chapter 1. Additive properties of prime numbers

Proof. Let m0 = NSZ (k, δ/2). Then, for every m ≥ m0 , any subset of
{1, 2, . . . , m} with cardinality ≥ δm/2 contains some arithmetic progres-
sion of length k. Let N be a large integer. For each 1 ≤ r ≤ ⌊N/m0 ⌋,
we divide {1, 2, . . . , N } into r arithmetic progressions of ratio r, e.g.,
{1 ≤ n ≤ N | n ≡ a(mod r)}, for each a with 0 ≤ a ≤ r − 1. Each
of these APs has ⌊N/r⌋ elements (at least), and, therefore, they can be
decomposed into a union of ⌊⌊N/r⌋/m0 ⌋ disjoint arithmetic progressions of
ratios r, lengths ≥ m0 (and almost equal), so that their diameters belong
to the interval [r(m0 − 1), r(2m0 − 1)].
Now, if A ⊂ {1, 2, . . . , N } satisfies |A| ≥ δN , we have that, for each r,
3δ
#{0 ≤ a ≤ r − 1 | #A ∩ {1 ≤ n ≤ N | n ≡ a(mod r)} ≥ ⌊N/r⌋} ≥
4
δr δ 3δ
(since t < ⇒ t + (1 − t) < δ).
4 − 3δ 4 − 3δ 4
δ δ 3δ
On the other hand, because t < ⇒ t+ (1−t) < , if #A∩{1 ≤
4 − 2δ 2 4
3δ
n ≤ N | A ≡ a(mod r)} ≥ ⌊N/r⌋ for a certain r, then, it follows
4
δ
that · ⌊⌊N/r⌋/m0 ⌋ of the created arithmetic progressions of length
4 − 2δ
≥ m0 (at least) intersects A with a relative proportion of δ/2 (at least),
⌊N/m
P 0 ⌋ δr
so that a fortiori, it should contain a k-AP. This gives us ·
r=1 4 − 3δ
δ
⌊⌊N/r⌋/m0 ⌋ > β(δ, m0 )N 2 k-APs (at least) contained in A (for a
4 − 2δ
large N ), where β(δ, m0 ) = δ 2 /64m20 . Of course, some of these APs can
be double-counted sometimes (for different choices of r), but once we fix
d
the diameter d of the AP, r must be a divisor of d between and
2m0 − 1
d d
, i.e., r = ′ , where k − 1 ≤ r′ ≤ 2m0 − 1. Consequently, there are
k−1 r
2m0 − k + 1 possibilities for r′ (at most) and a fortiori for r, so that each
AP is counted 2m0 − k + 1 times (at most). Hence, A contains at least
δ2
α(k, δ)N 2 k-APs, where α(k, δ) = ·
64m20 (2m0 − k + 1)

Remark 1.6.2. The basic difference between the proof of the quantitative
Szemerédi theorem and its previous versions is the finitary nature of the
arguments (allowing to explicit bounds on NSZ ). The proofs of the other
versions are infinitary arguments (they use to some extend the Axiom of
Choice) only permits us to show the existence of NSZ without any bound
on its magnitude. The strategy of the proof of the quantitative Szemerédi
is used during the proof of Green-Tao theorem as we are going to see in
the next chapter (we also recommend [16]).
A. Arbieto, C. Matheus and C. G. Moreira 31

1.7 Further results

During this section, we indicate (without proofs) equivalent analytic for-
mulations of some of the conjectures cited above. After that, we present
some further results (historically more recent than Green-Tao theorem).
Finally, we will make some comments about the nature of the integer
N0 (k, δ).

1.7.1 von Mangoldt function

In order to reformulate some number-theoretical theorems and conjectures,
we need to introduce the so-called von Mangoldt function.

Definition 1.7.1. The von Mangoldt function Λ : Z → R+ is given by

Λ(n) = log p if n = pr (for some r ≥ 1), and Λ(n) = 0 otherwise.

In these terms, observe that the unique factorization theorem can be

expressed as:
X
log n = Λ(d). (1.1)
d|n

Now we recall the definition of expectation of a function on a finite set:

Definition 1.7.2. Given f : X → R and A ⊂ X a finite set, we define

the (conditional) expectation of f with respect to A via the formula:

1 X
E(f (n)|n ∈ A) = E(f |A) = f (n).
|A|
n∈A

In this setting, the prime number theorem can be seen as an estimate

for the conditional expectations of the von Mangoldt function:

Theorem 1.7.1. The prime number theorem is equivalent to

E(Λ|[1, N ]) = 1 + o(1).

Proof. By the definition of the von Mangoldt function, we have:

X log N X
N E(Λ|[1, N ]) = [ ] log p ≤ log N 1
log p
p≤N p≤N

= log N · (|primes in [1, N ]|).

Dividing by N , we get that the prime number theorem implies the state-
ment about the expectations of Λ.
32 Chapter 1. Additive properties of prime numbers

On the other hand, if 1 < M < N , then

X
|primes in [1, N ]| = |primes in [1, M ]| + 1
M <p≤N
X log p
≤ |primes in [1, M ]| +
log M
M <p≤N
1
< M+ N E(Λ|[1, N ]).
log M
Now, if N is large, we have 1 < M = logN2 N < N . Combining this
inequality with the previous estimate, we obtain:
N N E(Λ|[1, N ])
|primes in [1, N ]| < 2 + .
log N log N − 2 log log N
Therefore,
|primes in [1, N ]| 1 1
< E(Λ|[1, N ])( )+ .
N log N − 2 log log N log2 N
log x
This ends the proof because log x−2 log log x → 1 when x → ∞.
In fact, the precise knowledge of the expectations of von Mangoldt func-
tion2 actually implies several famous conjectures. Let us give a list of these
conjectures (without further details):
• The Riemann hypothesis 3 is equivalent to the following claim:
E(Λ|[1, N ]) = 1 + O(N −1/2 log2 N ).

• The twin prime conjecture follows from the following affirmation:

lim inf E(Λ(n)Λ(n + 2)|1 ≤ n ≤ N ) > 0.
N →∞

• The Goldbach conjecture is equivalent to:

E(Λ(n1 )Λ(n2 )|n1 , n2 ∈ [1, N ] and n1 + n2 = N ) > 0 ∀ N even.

• The odd Goldbach conjecture is equivalent to:

E(Λ(n1 )Λ(n2 )Λ(n3 )|n1 , n2 , n3 ∈ [1, N ] and n1 +n2 +n3 = N ) > 0

∀ N odd.
2 I.e., an explicit estimate for the speed of convergence to zero of the term o(1) in

Thereom 1.7.1.
3 This famous conjecture, one of the seven Millennium prize problems of Clay Mathe-

matical Institute (who offers 1 million dollars for its solution), says that the (non-trivial)
zeros of Riemann zeta function ζ(s) (a complex-analytic function related to the prime
∞
P
numbers obtained by analytic continuation of 1/ns , ℜ(s) > 1) are located in the
n=1
line ℜ(s) = 1/2.
A. Arbieto, C. Matheus and C. G. Moreira 33

1.7.2 Constellations of primes and polynomial pro-

gressions
A well-known set with a nice notion of primality is the set of Gaussian
integers Z[i] := {a + bi; a, b ∈ Z}. Here, by definition, p is a Gaussian
prime if it is only divisible by ±1, ±i, ±p and ±ip.
A shape on Z[i] is a finite set (vj )j∈J ∈ (Z[i])J of distinct Gaussian inte-
gers. A constellation in Z[i] with this shape is any J-tuple (a + rvj )j∈J ∈
(Z[i])J of distinct Gaussian integers (where a ∈ Z[i] and r ∈ Z[i]).
The notion of constellation extends the concept of arithmetic progres-
sions to the context of Gaussian integers. The abundance of arbitrar-
ily shaped constellations formed by Gaussian primes was proved by T.
Tao [17]:

Let (vj )j∈J be an arbitrary shape of Gaussian integers. Then, the set of
Gaussian primes contains infinitely many constellations with this
prescribed shape.

On the other hand, an alternative generalization of the concept of arith-

metical progressions is: since any arithmetical progression has the form
x + P1 (m), . . . , x + Pk (m) where Pi (m) = (i − 1)m, one can extend this
notion by simply allowing Pi ∈ Z[m] to be some integer-valued polyno-
mials with Pi (0) = 0 (for each i = 1, . . . , k). These type of generalized
progression are called polynomial progressions.
The existence of infinitely many polynimial progressions formed by prime
numbers was showed by T. Tao and T. Ziegler [18]:

Let P1 , . . . Pk be integer-valued polynomials with Pi (0) = 0. Given ε > 0,

there are infinitely many integers x and m such that x + Pi (m) are prime
numbers for any i = 1, . . . , k and 1 ≤ m ≤ xε .

1.7.3 Gaps in the set of prime numbers

In a certain sense, all the theorems presented in this chapter have a com-
mon feature of searching patterns in the set P of prime numbers. In this
direction, a natural problem is to know how sparse can P be.
In order to investigate this issue, one can denote by pn the n-th prime
number, so that the size of the n-th gap of P is pn+1 − pn . The prime
number theorem says that the average size of this gap is morally log pn .
We define ∆ as the smallest number such that there are infinitely many
gaps of size less than (∆ + ε) times the average gap, i.e.,
 
pn+1 − pn
∆ = lim inf .
n→∞ log pn
34 Chapter 1. Additive properties of prime numbers

It was conjectured that ∆ = 0 and this fact was recently proved by D.

Goldston, J. Pintz and C. Yıldırım [7]. Also, in a more recent work,
D. Goldston, J. Pintz and C. Yıldırım [8] were able to show the slightly
stronger theorem:
 
pn+1 − pn
lim inf √ < ∞.
n→∞ log pn (log log pn )2

In these works, the authors proposed a method to show the existence of

large primes which are very close4 .

Remark 1.7.1. Just to stress the difficulty of the twin prime conjecture,
let us observe that this conjecture is much stronger than the result ∆ = 0
of Goldston, Pintz and Yıldırım (whose proof is highly non-trivial!).

1.7.4 Magnitude of N0 (k, δ)

Concerning the size of N0 (k, δ) in the quantitative version of Szemerédi
theorem, we have the following results:
δ −ck k+9
• T. Gowers showed that N0 (k, δ) ≤ 22 , where ck = 22 ;

• R. Rankin proved that N0 (k, δ) ≥ exp(C(log 1δ )1+⌊log2 (k−1)⌋ );

−2
• J. Bourgain proved that N0 (3, δ) ≤ 2Cδ log(1/δ)
;
−1
• it is expected that N0 (k, δ) ≤ 2ck δ , but this is an open problem
related to the Erdös-Turán conjecture.

1.8 Appendix to Chapter 1

1.8.1 Proof of Theorem 1.4.3
A preliminary observation is: if, for some k and ε, the first part of theorem
holds for a certain x, then the same statement is true for an entire small
neighborhood of x, and, a fortiori, the theorem works with some element
of any fixed dense subset Z. Hence, it suffices to show the first part of the
theorem to get a full proof of it.
Next, we notice that Zorn’s lemma implies that one can assume that X
is minimal, i.e., X doesn’t possess any proper closed subset Y such that
T (Y ) ⊂ Y . Observe that, in this situation, the subsets {T m (x)}∞ m=0 are
dense in X, so that the theorem is true for k = 1 (since, by denseness,
there exists some n ∈ N with d(T n (x), x) < ε).
4 Actually, they can show that the existence of infinitely many primes at a bounded

distance assuming a conjecture of Elliot-Halberstam.

A. Arbieto, C. Matheus and C. G. Moreira 35

At this stage, the proof proceeds by induction (on k). Suppose that the
theorem holds for some k ≥ 1, i.e., for all ε > 0 there exists x ∈ X and
n ∈ N such that d(T in (x), x) < ε for each i = 1, . . . , k. We claim that the
set of such points x is actually dense in X.
Indeed, let U ⊂ X be an arbitrary open subset and pick B ⊂ U a small
ball of radius strictly less than ε. Define Bm = (T m )−1 (B), so that these
subsets form an open cover of X (by the minimality assumption). Using
the compactness of X, we can extract a finite subcover {Bm1 , . . . , Bmr }.
Let δ > 0 be the Lebesgue number of this open subcover, that is, a number
such that any ball of radius δ is contained inside some element of the
subcover. Take x and n such that d(T in (x), x) < δ for i = 1, . . . , k (whose
existence is assured by the inductive hypothesis) and denote by D the ball
of center x and radius δ. Then, by our choice of δ, there exists j such that
D ⊂ Bmj . In particular, T mj (D) ⊂ B, that is, the elements T mj (T in (x))
belong to the ball of radius ε centered on T mj (x) ∈ U . This proves our
denseness claim.
Now, let’s go back to the proof of the theorem. Fix ε > 0. By the induc-
tive hypothesis, there are x0 and n0 such that d(T in0 x0 , x0 ) < ε/2 for i =
1, . . . , k. Taking x1 such that T n0 (x1 ) = x0 , we have d(T (i+1)n0 x1 , x0 ) <
ε/2 for i = 1, . . . , k. Hence, d(T in0 (x1 ), x0 ) < ε/2 for i = 1, . . . , k + 1.
By continuity, there exists ε1 < ε such that d(y, x1 ) < ε1 implies
d(T in0 (y), x0 ) < ε/2 for i = 1, . . . , k + 1. By our denseness claim, there
are y1 and n1 such that d(y1 , x1 ) < ε1 /2 and d(T in1 (y1 ), y1 ) < ε1 /2 for
i = 1, . . . , k. By the triangular inequality, we have:

d(T in0 (T (i−1)n1 (y1 )), x0 ) < ε2 for i = 1, . . . , k + 1.

Proceeding in this way (taking x2 such that T n1 (x2 ) = y1 , etc.), we find

a sequence of points x2 , x3 , . . . ∈ X and a sequence of natural numbers
n2 , n3 , . . . such that, for each l, we have:

d(T inl−1 (xl ), xl−1 ) < ε/2

d(T i(nl−1 +nl−2 ) (xl ), xl−2 ) < ε/2
...
d(T i(nl−1 +···+n0 ) (xl ), x0 ) < ε/2 for i = 1, . . . , k + 1.

By compactness, there are l > m such that d(xl , xm ) < ε/2. By the
triangular inequality, we have:

d(T i(nl−1 +···+nm ) (xl ), xl ) < ε , for i = 1, . . . , k + 1.

Therefore, it suffices to take x = xl and n = nl−1 + · · · + nm to conclude

the proof of the theorem.
36 Chapter 1. Additive properties of prime numbers

1.8.2 F. Behrend’s example

As we announced in the remark 1.6.1, firstly we will construct some exam-
ples of subsets S of natural numbers ≤ N such that S doesn’t contain
√
any
1− 2 2 log 2+ε
√
arithmetic progression of length 3 and its cardinality is |S| ≥ N log N ;
after that, we will adapt this technique to study the behavior of the func-
tion c(3, δ).
Given d ≥ 2, n ≥ 2 and k ≤ n(d − 1)2 , consider Sk (n, d) the subset of
all integer of the form

x = a1 + a2 (2d − 1) + · · · + an (2d − 1)n−1

whose digits ai in the basis (2d − 1) are subject to the constraints

0 ≤ ai < d and kxk2 := a21 + · · · + a2n = k.

Notice that Sk (n, d) doesn’t contain 3-APs (arithmetic progressions of

length 3): in fact, if this is not the case, there are x, x′ , x′′ ∈ Sk (n, d) such
that x + x′ = 2x′′ , so that
√
kx + x′ k = k2x′′ k = 2 k

and √
kxk + kx′ k = 2 k.
Thus, since the equality in the triangular inequality kx + x′ k ≤ kxk +
kx′ k can only occur when the vectors (a1 , . . . , an ) and (a′1 , . . . , a′n ) are
proportional, we see that x = x′ = x′′ (because the norms of these vectors
are the same by hypothesis).
On the other hand, there are dn vectors (a1 , . . . , an ) satisfying the con-
straint 0 ≤ ai < d and there are n(d − 1)2 + 1 possible values of k. Con-
sequently, for some k = K0 , the subset Sk (n, d) must have cardinality (at
least)
dn dn−2
> ·
n(d − 1)2 + 1 n
Since every element of Sk (n, d) has modulus < (2d − 1)n , if we define

ν(N ) := max{|S| ; S ⊂ [1, N ], S without any 3-AP},

it holds
ν((2d − 1)n ) > dn−2 /n.

qNow, for a fixed ε > 0 and for a given (large) N , we choose n =

⌊ 2log
log N
2 ⌋ and d satisfying

(2d − 1)n ≤ N < (2d + 1)n ,

A. Arbieto, C. Matheus and C. G. Moreira 37

so that
dn−2 (N 1/n − 1)n−2
ν(N ) ≥ ν((2d − 1)n ) > >
n 2n−2 n
N 1−(2/n)
= (1 − N −1/n )n−2
2n−2 n
N 1−(2/n) log n (n−1) log 2
> n−1
= N 1−(2/n)− log N − log N
2 √n
1− 2 2 log 2+ε
√
>N log N .

Next, we will slightly modify this reasoning to study the behavior of

c(3, δ): fix d, n ≥ 1 integers (to be chosen later) and define φ : {1, . . . , N } →
{0, . . . , 2d − 1}n by
n−1
φ(x) := (⌊x/(2d − 1)i ⌋mod(2d − 1))i=0 .

For each 1 ≤ k ≤ n(d − 1)2 , consider again the subsets

Sk (n, d) := {(x1 , . . . , xn ) ∈ {0, . . . , d − 1}n : x21 + · · · + x2n = k}

and define Ak (n, d) := φ−1 (Sk (n, d)). As we already know, Sk (n, d) is
free from 3-APs (except for the trivial 3-APs {x, x, x}). This implies that
Ak (n, d) can only contain arithmetic progressions (n, n + r, n + 2r) when r
is a multiple of (2d − 1)n . In particular, the maximal number of 3-APs in
Ak (n, d) is N 2 /(2d − 1)n . On the other hand, when φ(x) ∈ {0, . . . , d − 1}n ,
1
the probability of x to belong to Ak (n, d) is n(d−1) 2 +1 . Therefore, we get

the following lower bound for the cardinality of Ak (n, d):

c −n
|Ak (n, d)| ≥ 2 N.
nd2
Taking n = c log(1/δ) and d = δ −c , we obtain that, for some k, the set
Ak (n, d) satisfies |Ak (n, d)| ≥ δ c N and the maximum number of 3-APs
inside Ak (n, d) is δ c log(1/δ) N 2 . In other words, c(3, δ) ≤ δ c log(1/δ) .

Remark 1.8.1. Actually, a careful analysis

√
of Behrend’s construction
1 1− 2√ 2 log 2
above shows that ν(N ) ≫ log1/4 N
·N log N . Recently, this lower bound

was improved by M. Elkin [4] who proved that

√
2 log 2
1− 2√log
ν(N ) ≫ log1/4 N · N N .

For a brief exposition of M. Elkin estimate see the paper [10] of B. Green
and J. Wolf.
Chapter 2

Green-Tao-Szemerédi
theorem

2.1 Introduction
The main goal of this chapter is the discussion of some of the ideas behind
the proof of Green-Tao theorem.
Roughly speaking, the proof has two main steps:

• firstly one generalizes Szemerédi theorem to the more general context

of pseudorandom measures: this is the content of the Green-Tao-
Szemerédi theorem (see the section 2.2 for further details);

• secondly, one shows the existence of pseudorandom measures on the

set of primes (along the lines of the works of Goldston-Yıldırım).

Once these two facts are established, we will see that the Green-Tao
theorem follows immediately (see the section 2.2).
However, before entering (in section 2.4) into the details of the previous
outline, we will try to motivate the concepts and tactics of the proof of
the Green-Tao-Szemerédi theorem via a complete proof of Roth theorem
(which corresponds to the particular case k = 3 of Szemerédi theorem) in
the section 2.3, while (unfortunately) we will skip completely the discussion
of the results related to the work of Goldston-Yıldırım.
The organization of this chapter is:

• In section 2.2 we will present in more details the scheme of the

proof of Green-Tao theorem; in particular, we will state precisely
the Green-Tao-Szemerédi theorem and Goldston-Yıldırım theorems;
finally, we will prove Green-Tao theorem assuming these two results.

38
A. Arbieto, C. Matheus and C. G. Moreira 39

• In section 2.3 we give a proof of Roth theorem about the existence

of infinitely many arithmetic progressions of length 3 (i.e., 3-AP)
in subsets of positive density to serve as an inspiring model for the
proof of the Green-Tao-Szemerédi theorem.

• Ending this chapter, in section 2.4, we will prove the Green-Tao-

Szemerédi theorem.

Closing this brief introduction, let us remark that when we reach the
end of this chapter, the Green-Tao theorem will be proved except for the
results of Goldston and Yıldırım whose beautiful proof will be omitted
because it is purely number-theoretical (and, thus, outside the scope of
this “ergodic-theoretical” book).

2.2 Strategy of the proof of the Green-Tao

theorem
During this chapter, we will fix k ≥ 3 the length of the arithmetic progres-
sion (AP) of primes (we are searching for) and we take N := |ZN | a (large)
prime number, so that the elements 1, . . . , N − 1 can be inverted in ZN .
We write o(1) to denote any quantity converging to zero when N → ∞
and O(1) to denote any quantity staying bounded when N → ∞. At cer-
tain moments of the text, the o(1) (resp., O(1)) quantities will converge
to zero (resp., stay bounded) depending on certain extra parameters (e.g.,
j, ε). In this case, we will put these extra parameters as subscript indices
(e.g., oj,ε (1)). Moreover, we will abreviate any quantity of the form O(1)X
(resp., o(1)X) as O(X) (resp., o(X)).
Keeping these notations in mind, let us put Green-Tao theorem in an
appropriate context. Recall that this theorem says that for any k ≥ 3,
there are infinitely many k-APs (i.e., arithmetic progressions of length
k) formed (only) by prime numbers. Moreover, we know that Szemerédi
theorem ensures the existence of several k-APs in subsets of integers with
positive density. However, we already know that Szemerédi theorem doesn’t
imply Green-Tao theorem because the set of prime numbers has density
zero. Nevertheless, the idea of Green and Tao is:

• although we can’t apply directly Szemerédi theorem, we can modify

it to work with certain subsets with a weakly random (additive) be-
havior1 (or, more precisely, pseudorandom); this result will be called
Green-Tao-Szemerédi in this book;
1 Of course, the point here consists into the choice of a nice definition of pseudoran-

domness in order to include the set of prime numbers.

40 Chapter 2. Green-Tao-Szemerédi theorem

• this reduces our task to show that the set of prime numbers has a
pseudorandom behaviour; this fact follows more or less directly from
the works of Goldston and Yıldırım.

In the sequel, we give the details for these items. To do so, we start
with the definition of pseudorandomness:

Definition 2.2.1. • We say that ν : ZN → R+ is a measure if E(ν) =

1 + o(1).

• A measure ν : ZN → R+ satisfies the (m0 , t0 , L0 )-linear forms con-

t
P of m ≤ m0 linear forms ψi : ZN → ZN ,
dition if, for any family
t ≤ t0 , say ψi (x) = Lij xj + bi , where Lij are rational numbers
with numerators and denominators ≤ L0 (in absolute value), no t-
tuple (Lij )1≤j≤t is a rational multiple of any other and for arbitrary
bi ∈ Z, then:

E(ν(ψi (x)) · · · ν(ψm (x))|x ∈ ZtN ) = 1 + om0 ,t0 ,L0 (1).

• A measure ν : ZN → R+ satisfies the m0 -correlation condition if,

for every m ≤ m0 , there are weights τm : ZN → R+ such that
q
E(τm ) = Om,q (1) (moment conditions) for all 1 ≤ q < ∞ and
X
E(ν(x + h1 ) · · · ν(x + hm )|x ∈ Z) ≤ τm (hi − hj ),
i<j≤m

for any hi ∈ ZN (not necessarily distinct).

• A measure ν : ZN → R+ is k-pseudorandom if ν : ZN → R+ satisfy

the (k·2k−1 , 3k−4, k) linear forms condition and the 2k−1 correlation
condition.

This definition may be strange at first, but it is essentially based on the

works of Goldston-Yıldırım, where one studies majorants for modified ver-
sions of the von Mangoldt function (which, as we saw, is intimately related
to the prime numbers). Intuitively, the two conditions above mean that
the set of integers supporting ν has weakly random (additive) arithmetic
properties. The main advantage of these conditions is the fact that they
are only slightly weaker than the usual randomness condition, so that Sze-
merédi theorem can be generalized (see Theorem 2.2.1 below), although
they are sufficiently weak to include the case of the prime numbers (even
though these two facts are very far from trivial).
Once we have a precise definition of pseudorandomness, we can state
one of the main results of the work of Green-Tao [9, Theorem 3.5]:
A. Arbieto, C. Matheus and C. G. Moreira 41

Theorem 2.2.1 (Green-Tao-Szemerédi). Let k ≥ 3 and 0 < δ ≤ 1. Then,

for any k-pseudorandom measure ν, it holds that, for every f : ZN → R+
such that 0 ≤ f (n) ≤ ν(n) and E(f ) ≥ δ, we have:
E(f (n)f (n + r) · · · f (n + (k − 1)r)|n, r ∈ ZN ) ≥ c(k, δ) − ok,δ (1).
Remark 2.2.1. Taking ν ≡ νconst ≡ 1 in this statement, we recover
Szemerédi theorem as a corollary. However, it turns out we are going to
use Szemerédi theorem in the proof of this more general result, so that
the proof of Green-Tao-Szemerédi theorem doesn’t give a new proof of
Szemerédi theorem.
The proof of Theorem 2.2.1 is based on Furstenberg ideas (i.e., multiple
recurrence theorem) and the so-called Gowers norms. For the moment,
we postpone the proof of Theorem 2.2.1 to section 2.4.
Since we aim to apply Theorem 2.2.1 to conclude the Green-Tao theo-
rem, let us see how one can construct pseudorandom measures related to
the prime numbers. We recall the definition:
Definition 2.2.2. The von Mangoldt function is
(
log p if n = pm
Λ(n) :=
0 otherwise .

We remember that this function is (essentially) supported on the prime

numbers (since the contribution of the powers of prime numbers is very
small), so that we can say that this function works as a “characteristic
function” of the primes. In terms of this function, we know that the prime
number theorem can be reformulated as E(Λ(n)) = 1 + o(1) (see chapter
1). In order to be able to use Green-Tao-Szemerédi theorem in the context
of prime numbers, we need to find a k-pseudorandom measure ν such that
ν(n) ≥ c(k)Λ(n). However, it is known that such measures don’t exist!2
In order to overcome this technical difficulty, Green and Tao use the“W-
trick ”. Let w = w(N ) → ∞ be a parameter growing very slowly with N
and let W = Πp≤w(N ); p prime p. The modified von Mangoldt function is:
(
φ(W )
e
Λ(n) = W log(W n + 1) if W n + 1 is prime
0 otherwise .

Now we have a function still seeing (some) prime numbers such that it
doesn’t see neither the powers of primes nor the annoying non-uniformity
2 Basically this occurs because the prime numbers (and the von Mangoldt function)

are concentrated, for any q > 1 integer, on the φ(q) residual classes a(mod q) with
(a, q) = 1 (where φ(q) is the Euler totient function), while any pseudorandom measure
must be equidistributed along all congruence classes modulo q; since the quocient φ(q)/q
can be made arbitrarily small, we see that the von Mangoldt function doesn’t have
pseudorandom majorants.
42 Chapter 2. Green-Tao-Szemerédi theorem

due to the presence of products of small primes. If w(n) has a sufficiently

slow growth 3 , say w(n) << log log n, we see that Dirichlet theorem im-
plies:
1 X e
Λ(n) = 1 + o(1).
N
n≤N
e is a measure. In this situation, the second key result
In other words, Λ
of the work of Green and Tao [9, Proposition 9.1] (based on the results of
Goldston and Yıldırım) is:
Theorem 2.2.2. If ǫk = 1/(k + 4)!2k and N is a very large prime
number, then there exists ν a k-pseudorandom measure such that ν(n) ≥
e
2−k−5 k −1 Λ(n) for ǫk N ≤ n ≤ 2ǫk N .
Following a long-standing tradition in (analytical) Number Theory, the
proof of Theorem 2.2.2 (as any important fact about prime numbers) uses
zero-free regions of the so-called Riemann zeta function (besides other
machineries). In particular, the nature of the argument is purely number-
theoretical. Because we want to keep the coherence between the title and
the content of this book, we will completely skip the proof of this beautiful
result (referring the curious reader to the original paper of Green and Tao)
and we will admit its validity during the rest of our discussion.
Finally, assuming momentarily the two key theorems 2.2.1 and 2.2.2,
we can show the Green-Tao theorem.

2.2.1 Proof of Green and Tao theorem

Suppose that the theorems 2.2.1 and 2.2.2 are valid.
If
1 e
f (n) = k+5 Λ(n)1 [ǫk N,2ǫk N ]
k2
then X
1 e 1
E(f ) = k+5
Λ(n) = k+5 ǫk (1 + o(1)).
N k2 k2
ǫk N ≤n≤2ǫk N
Since the theorem 2.2.2 guarantees the existence of a k-pseudorandom
e in [ǫk N, 2ǫk N ], we can apply the Green-Tao-
majorant of 2−k−5 k −1 Λ
Szemerédi theorem 2.2.1 to conclude that:
E(f (n) · · · f (n + (k − 1)r)|n, r ∈ ZN ) ≥ c(k, k −1 2−k−3 ǫk ) − o(1).
Because the case r = 0 contributes with a factor of O( N1 logk N ) = o(1), we
obtain the existence of a k-AP of primes in ZN (for a large N ). Moreover,
since ǫk < 1/k and k ≥ 3, we see that this k-AP is a legitimate k-AP in Z
(in the sense that it doesn’t wrap around in ZN ).
3 Although we ask for slow growth of w(n), it is possible to check (at the end of

the proof of Green-Tao theorem) that it suffices to take w(n) a very large constant
depending only on k (but not on N ).
A. Arbieto, C. Matheus and C. G. Moreira 43

2.2.2 Some comments

Once we reduced the Green-Tao theorem to the two key theorems 2.2.1
and 2.2.2, we will dedicate the rest of this chapter to the proof of the
ergodic part of the argument, namely, the Green-Tao-Szemerédi theo-
rem 2.2.1 (while we skip the number-theoretical part corresponding to
the theorem 2.2.2).
However, in order to illustrate the ideas behind the Green-Tao-Szemerédi
theorem (which can be technical and hard in a first reading), we will
present a proof of Roth theorem (i.e., Szemerédi theorem in the particular
case k = 3) and, in the sequel, we prove the Green-Tao-Szemerédi theorem
(closing the chapter).

2.3 Proof of Roth theorem

Define Λ3 (f, g, h) = E(f (n)g(n + r)h(n + 2r)|n, r ∈ ZN ). In this language,
Roth theorem can be reformulated as:
Theorem 2.3.1. For all non-negative f : ZN → R with

0 < δ ≤ kf kL1 (ZN ) ≤ kf kL∞ (ZN ) ≤ 1

it holds
Λ3 (f, f, f ) ≥ c(3, δ) − oδ (1).
In other words, we want some lower bounds on Λ3 (f, f, f ). We begin
with the simple remark that it is fairly easy to get upper bounds: e.g., by
Young inequality,

|Λ3 (f, g, h)| ≤ kf kLp kgkLq khkLr ,

if 1 ≤ p, q, r ≤ ∞ and p1 + 1q + 1r ≤ 2.
On the other hand, we are only interested in lower bounds for Λ3 and,
a priori, upper bounds are not useful for our interests. However, we can
decompose f into a “good” part g = E(f ) and a“bad” part b = f − E(f ).
Using the multilinearity of Λ3 , we can use this decomposition to split up
Λ3 (f, f, f ) into eight pieces:

Λ3 (f, f, f ) = Λ3 (g, g, g) + · · · + Λ3 (b, b, b).

By hypothesis, E(f ) ≥ δ, so that the first term is Λ3 (g, g, g) ≥ δ 3 . Thus,

some good upper estimates of the remaining terms (e.g., if the sum of the
seven remaining terms is smaller than δ 3 ) would lead us to a proof of Roth
theorem.
Unfortunately, the previous upper bound (via Young inequality) isn’t
sufficient to conclude the argument, unless δ is close to 1, say δ > 2/3
44 Chapter 2. Green-Tao-Szemerédi theorem

(a non-realistic case). But, we can refine this upper bound argument via
Harmonic Analysis, or, more precisely the Fourier transform:

fb(ξ) := E(f (x)eN (−xξ) : x ∈ ZN ),

where eN (x) := exp(2πix/N ). From the inversion formula

X
f (x) = fb(ξ)eN (xξ)
ξ∈ZN

we obtain that
X
Λ3 (f, g, h) = fb(ξ1 )b
g (ξ2 )b
h(ξ3 )×
ξ1 ,ξ2 ,ξ3

E(eN (nξ1 + (n + r)ξ2 + (n + 2r)ξ3 ) : n, r ∈ ZN ).

Note that the expectations on the right-hand side are 1 if ξ1 = ξ3 and

ξ2 = −2ξ1 , and 0 otherwise. In particular,
X
Λ3 (f, g, h) = fb(ξ)b
g (−2ξ)b
h(ξ).
ξ∈ZN

From Plancherel formula kf kL2 (ZN ) = kfbkl2 (ZN ) and Hölder inequality,
it follows that

g kl4 (ZN ) kb
|Λ3 (f, g, h)| ≤ kf kL2 (ZN ) kb hkl4 (ZN ) . (2.1)

Using this estimate, we can prove:

Proposition 2.3.1. Fix f and consider a decomposition f = g + b where

kgkL∞ (ZN ) , kbkL∞ (ZN ) = O(1)

and
kgkL1 (ZN ) , kbkL1 (ZN ) = O(δ).
Then
Λ3 (f, f, f ) = Λ3 (g, g, g) + O(δ 5/4 kbbkl4 (ZN ) )
and
Λ3 (f, f, f ) = Λ3 (g, g, g) + O(δkbbkl∞ (ZN ) ).

Proof. By hypothesis, kgkL2 (ZN ) , kbkL2 (ZN ) = O(δ 1/2 ), so that Plancherel
theorem says that

g kl2 (ZN ) , kbbkl2 (ZN ) = O(δ 1/2 ).

kb
A. Arbieto, C. Matheus and C. G. Moreira 45

Furthermore, the L1 bounds on g and b imply

g kl∞ (ZN ) , kbbkl∞ (ZN ) = O(δ).

kb

Thus, by Hölder inequality,

g kl4 (ZN ) , kbbkl4 (ZN ) = O(δ 3/4 ).

kb

The proposition follows by decomposing Λ3 (f, f, f ) into eight pieces and

using (2.1).

This proposition suggests that a possible strategy to get non-trivial lower

bounds on Λ3 (f, f, f ) passes by a decomposition f = g + b into a good
function in the sense that Λ3 (g, g, g) is “large” and a bad function b in the
sense that the l4 norm of its Fourier transform is small.
The attentive reader noticed that we already indicated one simple possi-
bility for the decomposition: g = E(f ) and b = f − E(f ). Observe that we
have the relatively nice estimate Λ3 (g, g, g) ≥ δ 3 , but we don’t dispose of
any good estimate for kbbkl4 (ZN ) ; our best estimates at the present moment
are O(δ 3/4 ), which is essentially useless since it allows the “error” term to
dominate the “main” term.4
However, we can eliminate the case of a linearly uniform b = f − E(f ),
i.e., kbbkl∞ ≤ δ 2 /100. Of course, the problem is: what to do when b isn’t
linearly uniform? In this situation, we adopt the so-called energy increment
argument, i.e.,

• Energy increment argument: once b isn’t uniform, we replace g

by another function whose L2 norm is strictly greater than kgkL2 .
After repeating this process a finite number of times, we expect to
reach an uniform function b (since the energy is finite).

Logically, one should work in details this idea in order to see that it leads to
a proof of Roth theorem. In this direction, let us introduce the definition:

Definition 2.3.1. Given K a positive integer, we call any function f :

ZN → C of the form
K
X
f (x) = cj exp(2πiξj x/N ),
j=1

where |cj | ≤ 1 and ξj ∈ ZN , a K-quasi-periodic function. Furthermore,

for a fixed σ > 0, we say that a function f is (σ, K)-quasi-periodic if
kf − fqp kL2 (ZN ) ≤ σ for some K-quasi-periodic function fqp .
4 Indeed, f = χ[1,δN ] verifies Λ3 (f, f, f ) ∼ δ 2 and Λ3 (g, g, g) = δ 3 , for instance.
46 Chapter 2. Green-Tao-Szemerédi theorem

Remark 2.3.1. Notice that the product f g is (2σ, K 2 )-quasi-periodic when-

ever f and g are (σ, K)-quasi-periodic functions.
An important point around the concept of quasi-periodic functions f is
the fact that it is possible to prove good lower bounds on Λ3 (f, f, f ) in
this context:
Lemma 2.3.1 (“Multiple Recurrence” of quasi-periodic functions). Given
0 < δ < 1, M ≥ 1, 0 < σ ≤ δ 3 /100M and 0 ≤ f ≤ M a non-negative
bounded (σ, K)-quasi-periodic function with E(f ) ≥ δ, it holds

Λ3 (f, f, f ) ≥ c(K, M, δ) − oK,M,δ (1),

for some c(K, M, δ) > 0.

PK
Proof. Let fqp (x) = j=1 cj exp(2πixξj /N ) be a K-quasi-periodic func-
tion close to f and take ε = ε(K, δ) > 0 a small constant. By Dirichlet
simultaneous approximation theorem (which is a consequence of Dirichlet’s
pigeonhole principle), we have

E(krξj k ≤ ε for every j ; r ∈ ZN ) ≥ c(ε, K). (2.2)

Here, it is fundamental to stress out that the constant c(ε, K) > 0 doesn’t
depend on N . On the other hand, by considering the shift dynamics
T (x) := x + 1 and by fixing r such that krξj k ≤ ε (where 1 ≤ j ≤ K), we
get:
kfqp ◦ T r − fqp kL2 (ZN ) ≤ C(K)ε.
Combining this information with the triangular inequality, it follows:

kf ◦ T r − f kL2 (ZN ) ≤ δ 3 /50M + C(K)ε.

Applying T r again to this estimate, we also obtain

kf ◦ T 2r − f ◦ T r kL2 (ZN ) ≤ δ 3 /50M + C(K)ε.

Since f is bounded, from these estimates, we can infer:

kf · (f ◦ T r ) · (f ◦ T 2r ) − f 3 kL1 (ZN ) ≤ δ 3 /2 + C(K)M ε.

However, because f ≥ 0, using our hypothesis E(f ) ≥ δ and Hölder in-

equality, we obtain

kf 3 kL1 (ZN ) ≥ kf k3L1 (ZN ) ≥ δ 3 .

Thus, E(f · (f ◦ T r ) · (f ◦ T 2r )) ≥ δ 3 /2 − C(K)M ε. Choosing ε > 0 small

depending on δ, K and M , it follows that

E(f · (f ◦ T r ) · (f ◦ T 2r )) ≥ δ 3 /4.
A. Arbieto, C. Matheus and C. G. Moreira 47

Because f ≥ 0, taking the average on r and using (2.2), we conclude

E(f (n) · (f ◦ T r )(n) · (f ◦ T 2r )(n) | n, r ∈ ZN ) ≥ δ 3 c(ε, K)/4.

Since Λ3 (f, f, f ) = E(f (n) · (f ◦ T r )(n) · (f ◦ T 2r )(n) | n, r ∈ ZN ), the proof

of the lemma is complete.
Looking for possible applications of this lemma, we will try to write arbi-
trary functions as a sum of a quasi-periodic function and a linearly uniform
function. Keeping this goal in mind, we construct a class of sigma-algebras
such that its measurable functions are always quasi-periodic functions:
Lemma 2.3.2. Take 0 < ε ≪ 1 and let χ be a function of the form
χ(x) := exp(2πixξ/N ). Then, there exists a sigma-algebra Bε,χ such that
kχ − E(χ|Bε,χ )kL∞ (ZN ) ≤ Cε and, for every σ > 0, there is K = K(σ, ε) >
0 with the following property: every Bε,χ -measurable function f satisfying
the estimate kf kL∞ (ZN ) ≤ 1 is (σ, K)-quasi-periodic.
Proof. We will use a random process to get the desired sigma-algebra:
take α a complex number in the unit square and let Bε,χ be the sigma-
algebra whose atoms have the form χ−1 (Q), where Q is a square such that
the vertices of Q − εα lies over the lattice εZ2 . Notice that this sigma-
algebra has O(1/ε) atoms and kχ − E(χ|Bε,χ )kL∞ (ZN ) ≤ Cε. In particular,
it remains only to verify the second part of the lemma. Observe that
it suffices to show the desired fact for σ = 2−n (where n ≫ 1) with a
probability of 1 − O(σ) on α. Since Bε,χ possess O(1/ε) atoms, we can
restrict ourselves to the case f equals to the characteristic function of an
atom A of Bε,χ so that our task is reduced to prove the desired property
with a probability of 1 − O(c(ε)σ). Observe that, in this situation, we can
rewrite f as f (x) = 1Q (χ(x) − εα). Applying Weierstrass approximation
theorem on the disc |z| ≤ O(1/ε), we can find a polynomial P (z, z) with
C(σ, ε) terms whose coefficients are bounded by C(σ, ε) such that |P | ≤ 1
on the disc |z| ≤ O(1/ε) and |1Q (z) − P (z, z)| = O(c(ε)σ) for all z on this
disc with the exception of a subset of measure O(c(ε)2 σ 2 ). This implies
that

k1Q (χ(x) − εα) − P (χ(x) − εα, χ(x) − εα)kL2 (ZN ) ≤ c(ε)σ

with a probability of 1 − O(c(ε)σ) on α. However P (χ(x) − εα, χ(x) − εα)

is a linear combination of C(ε, σ) functions of the form exp(2πixξ/N )
where the coefficients are bounded by C(ε). In other words, it follows that
P (χ(x)−εα, χ(x)−εα) is a K-quasi-periodic function (where K = C(ε, σ)).
In particular, f is a (σ, K)-quasi-periodic function. This concludes the
proof of the lemma.
A useful corollary of this lemma is:
48 Chapter 2. Green-Tao-Szemerédi theorem

Corollary 2.3.1. Let 0 < εj ≪ 1 and χj (x) = exp(2πixξj /N ), where

j = 1, . . . , n. Denote by Bεj ,χj the sigma-algebras provided by the previous
lemma. Then, for all σ > 0, there is some K = K(n, σ, ε1 , . . . , εn ) such
that any Bε1 ,χ1 ∨ · · · ∨ Bεn ,χn -measurable function f satisfying the estimate
kf kL∞ (ZN ) ≤ 1 is (σ, K)-quasi-periodic.

Proof. Because the number of atoms of the sigma-algebra Bε1 ,χ1 ∨ · · · ∨

Bεn ,χn is C(n, ε1 , . . . , εn ), it suffices to show the corollary for the particular
case of characteristic functions f of some atom of Bε1 ,χ1 ∨· · ·∨Bεn ,χn . But,
in this setting, f is a product of n characteristic functions of certain atoms
of the sigma-algebras Bεj ,χj . Thus, the corollary follows from the previous
lemma and Remark 2.3.1.

Another interesting property of these sigma-algebras (besides the fact

that they contain quasi-periodic functions) is their usefulness in the iden-
tification of obstructions to linear uniformity:

Lemma 2.3.3. Let b be a bounded function with kbbkl∞ (ZN ) ≥ σ > 0 and
0 < ε ≪ σ. Then, there exists a function of the form χ(x) = exp(2πixξ/N )
such that the associated sigma-algebra Bε,χ satisfies

kE(b|Bε,χ )kL2 (ZN ) ≥ σ/2.

Proof. By definition, there exists a frequency ξ such that |bb(ξ)| ≥ σ, i.e.,

|E(b(n) exp(−2πinξ/N )|n ∈ ZN )| ≥ σ.

Define χ(x) := exp(2πixξ/N ) and rewrite the previous inequality as

|hb, χiL2 (ZN ) | ≥ σ.

By the previous lemma, we know that there is a sigma-algebra Bε,χ such

that
kχ − E(χ|Bε,χ )kL∞ (ZN ) ≤ Cε.
On the other hand, since b is bounded and the conditional expectation is
self-adjoint, we can combine these two estimates to conclude

hE(b|Bε,χ ), χi = hb, E(χ|Bε,χ )i ≥ σ − Cε.

This shows that kE(b|Bε,χ )kL2 (ZN ) ≥ σ − Cε ≥ σ/2, so that the proof of
the lemma is complete.

The last ingredient in the proof of Roth theorem is the following struc-
ture proposition:
A. Arbieto, C. Matheus and C. G. Moreira 49

Proposition 2.3.2 (“quantitative Koopman-von Neumann theorem”).

Let F : R+ × R+ → R+ be an arbitrary function, 0 < δ ≤ 1, f be a
non-negative bounded function verifying E(f ) ≥ δ and σ := δ 3 /100. Then,
there are a constant 0 < K ≤ C(δ, F ) and a decomposition f = g + b such
that g is non-negative, bounded, E(g) = E(f ), g is (σ, K)-quasi-periodic
and b verifies
kbbkl∞ (ZN ) ≤ F (δ, K). (2.3)
Proof. We will use the energy increment argument during the construction
of g and b. For this argument, we need two sigma-algebras B and Be which
are always of the form Bε1 ,χ1 ∨ · · · ∨ Bεn ,χn during the entire argument.
Moreover, we will need some estimatives of the form
e 22
kE(f |B)k 2 2
L (ZN ) ≤ kE(f |B)kL2 (ZN ) + σ /4. (2.4)

Observe that, by Pythagoras theorem, this estimate is equivalent to

e − E(f |B)k2 2
kE(f |B) L (ZN ) ≤ σ/2.

For the proof of this proposition, we will employ the following algorithm:
• Stage 0 : We start with B and Be equal to the trivial sigma-algebra
{0, ZN }. Note that the inequality (2.4) is automatically satisfied at
this stage.
• Stage 1 : Consider B a sigma-algebra of the form Bε1 ,χ1 ∨ · · · ∨
Bεn ,χn , where χj (x) = exp(2πixξj /N ). Since the function E(f |B)
is bounded and B-measurable, the corollary 2.3.1 says that one can
find K depending on δ, n, ε1 , . . . , εn such that E(f |B) is (σ/2, K)-
quasi-periodic.
e If kbbkl∞ (Z ) ≤ F (δ, K),
e and b = f −E(f |B).
• Stage 2 : Put g = E(f |B) N
we end the algorithm. Otherwise, we go to the third stage.
• Stage 3 : Since we didn’t end the algorithm at the stage 2, we have
kbbkl∞ > F (δ, K). By Lemma 2.3.3, we can find ε ≪ F (δ, K) and
a function χ of the form χ(x) = exp(2πixξ/N ) whose associated
sigma-algebra Bε,χ verifies

kE(b|Bε,χ )kL2 (ZN ) ≥ F (δ, K)/2.

From the identity

E(b|Bε,χ ) = E(E(f |Be ∨ Bε,χ ) − E(f |B)|B

e ε,χ )

and Pythagoras theorem, we obtain also that

kE(f |Be ∨ Bε,χ ) − E(f |B)k

e L2 (Z ) ≥ F (δ, K)/2.
N
50 Chapter 2. Green-Tao-Szemerédi theorem

Applying Pythagoras theorem again, we get an energy increment

estimate:

kE(f |Be ∨ Bε,χ )k2L2 (ZN ) ≥ kE(f |B)k

e 22 2
L (ZN ) − F (δ, K) /4.

• Stage 4 : We replace Be by B∨B

e ε,χ . If one still don’t have the estimate
(2.4), we return to the stage 2; otherwise, we replace B by Be and we
go to the stage 1.
We claim that this algorithm stops. In fact, for a fixed B (and consequently
K), each time we pass by the stage 4, the energy kE(f |B)ke 22
L (ZN ) increases
2
by a factor of F (δ, K) /4. Thus, either the algorithm ends or the estimate
(2.4) is violated in C(δ, K, F ) = Cσ 2 /F (δ, K)2 steps. In the second case,
we replace B by the sigma-algebra associated to the C(δ, K, F ) functions χ
and parameters ε ≥ C(δ, F, K)−1 appearing in this process. This implies
that the quantity K associated to this new sigma-algebra B will be changed
by a quantity of the form C(δ, K, F ) and the energy kE(f |B)k2L2 (ZN ) will
increase by (at least) σ 2 /4 due to the violation of (2.4). On the other
hand, the fact that f is bounded ensures that the energy can’t be bigger
than O(1). Therefore, these replacements of B just described can only be
perfomed at most O(σ −2 ) times. Putting these informations together, we
see that the whole algorithm ends in C(δ, F ) steps (and the quantity K is
never bigger than C(δ, F ) during the entire process). This concludes the
proof of the proposition.
Finally, let us complete the proof of Roth theorem: let F : R+ × R+ →
+
R be a function to be chosen in a few moments and let us apply the
previous propostion to get the corresponding decomposition f = g + b. By
the lemma 2.3.1, we know that

Λ3 (g, g, g) ≥ c(δ, K) − oδ,K (1).

Combining this inequality with (2.3) and the proposition 2.3.1, we have

Λ3 (f, f, f ) ≥ c(δ, K) + O(δ · F (δ, K)) − oδ,K (1).

Taking F “sufficiently small”, we can absorb the second term of the right-
hand side via the first term, so that

Λ3 (g, g, g) ≥ c(δ, K)/2 − oδ,K (1).

Since K ≤ C(δ, F ) = C(δ), Roth theorem is proved.

Closing this section, let us review below the two main steps of the proof
of Roth theorem (which are going to inspire our discussion of the Green-
Tao-Szemerédi theorem):
A. Arbieto, C. Matheus and C. G. Moreira 51

• 1st step: to define a class of norms (Gowers norms k.kU k−1 ) in order
to control the expectation of a k-AP to reside in the support of f ;
observe that, by proposition 2.3.1, in the particular case k = 3, the
l4 norm of the Fourier transform is a good candidate;5
• 2nd step: to make an energy increment argument.

2.4 Proof of Green-Tao-Szemerédi theorem

In this last subsection of the present chapter, we will prove (along several
subsections) the key results allowing us to formalize the previous ideas.
However, since they are somewhat technical, the reader may get lost during
the discussion. In view of this possibility, we included at the end of each
subsection a “resume” of the main results demonstrated and how they are
connected to the big picture of the energy increment strategy.

2.4.1 Gowers norms

Let {0, 1}d be the d-dimensional discrete cube, and w = (w1 , . . . , wd ) ∈
{0, 1}d . If h ∈ ZdN , then w.h := w1 .h1 + . . . wd .hd . If {fw }w∈{0,1}d , then
the Gowers inner product is:
h(fw )iU d := E(Πw fw (n + w.h)|n ∈ ZN , h ∈ ZdN ).
Firstly we remark that if fw = f for every w then h(fw )iU d ≥ 0. Thus, we
can define the Gowers norms (using fw = f ):
1
kf kU d := h(f )iU2dd .
A basilar tool for the analysis of the Gowers norms is the Gowers-Cauchy-
Schwarz inequality:
|h(fw )iU d | ≤ Πw kfw kU d .
The proof of this inequality follows from the fact that, when fw doesn’t
depend on wd , it holds
Y
h(fw )iU d = E(E( fw′ ,0 (y + w′ .h′ ) : y ∈ ZN ) ×
w′ ∈{0,1}d−1
Y
E( fw′ ,1 (y + w′ .h′ ) : y ∈ ZN |h′ ∈ (ZN )d−1 )).
w′ ∈{0,1}d−1

Therefore, by the Cauchy-Schwarz inequality, we have

1/2 1/2
|h(fw )iU d | ≤ h(fw′ ,0 )iU d h(fw′ ,1 )iU d .
5 Indeed, in the case k = 3, the Gowers norm k.k 4
U 2 is the l norm of the Fourier
transform; see the remark 2.4.1 of the next subsection.
52 Chapter 2. Green-Tao-Szemerédi theorem

Since we can exchange wd by any other digit, applying this estimate d

times, we obtain the Gowers-Cauchy-Schwarz inequality.
Furthermore, the binomial formula and the multilinearity of the inner
product lead us to the Gowers triangular inequality:

kf + gkU d ≤ kf kU d + kgkU d .

Finally, we have the following monotonicity relation:

kf kU d−1 ≤ kf kU d ,

which is a direct consequence of the Gowers-Cauchy-Schwarz inequality

applied to the case fw := 1 when wd = 1 and fw := f when wd = 0.

Remark 2.4.1. Since the norms k.kU d are homogenous, this shows that
k.kU d are semi-norms. However, k.kU 1 isn’t a norm because kf kU 1 = E(f ).
However, one can prove (by direct calculation):
X 1
kf kU 2 = ( fb(ξ)4 ) 4 ,

where fb(ξ) = E(f (x)e−2πiξ/N ; x ∈ ZN ) and the inversion formula f (x) =

Pb
f (ξ)e2πixξ/N holds. Consequently, the Gowers norms are genuine norms
for d ≥ 2.

Using this notation, the natural generalization of proposition 2.3.1 is:

Theorem 2.4.1 (generalized von Neumann theorem). If ν is a k-pseudoran-

dom measure and f0 , . . . , fk−1 ∈ L1 (ZN ) are some functions such that
|fj (x)| ≤ 1 + ν(x), then, if c0 , . . . , ck−1 ∈ ZN are distinct, we have:

E(Πj fj (n + cj r)|n, r ∈ ZN ) = O(inf kfj kU k−1 ) + o(1).

Proof. We begin with some preliminary reductions: up to replacing ν by

(ν + 1)/2, rearranging fj , cj and translating x by c0 r, we can assume that

|fj (x)| ≤ ν(x), ∀ x ∈ ZN , j = 0, . . . , k − 1,

inf kfj kU k−1 = kf0 kU k−1

0≤j≤k−1

and
c0 = 0.
This reduces our problem to prove that
 
k−1
Y
E fj (x + cj r)|x, r ∈ ZN  = O(kf0 kU k−1 ) + o(1).
j=0
A. Arbieto, C. Matheus and C. G. Moreira 53

We divide the proof of this inequality into two parts: in the first part
we prove a Cauchy-Schwarz type inequality and we apply this inequality
k − 1 times to
Qthe left-hand side of the previous
 identity in order to get a
k−1
control of E j=0 fj (x + cj r)|x, r ∈ ZN via a weighted sum of f0 over
(k −1)-dimensional cubes; in the second part we show that the linear forms
conditions implies that these weights are equal to 1 in average (so that the
theorem follows).
In order to state the Cauchy-Schwarz type inequality in a reasonable
way, we introduce a little bit more of notation. Given 0 ≤ d ≤ k − 1,
two vectors y = (y1 , . . . , yk−1 ) ∈ (ZN )k−1 and y = (yk−d
′ ′
, . . . , yk−1 ) ∈
d (S)
(ZN ) , and a subset S ⊂ {k − d, . . . , k − 1}, we define the vector y =
(S) (S)
(y1 , . . . , yk−1 ) ∈ (ZN )k−1 by

(S) yi if i ∈/S
yi :=
yi′ if i ∈ S.
In other words, S indicates the components of y (S) coming from y ′ instead
of y.
Lemma 2.4.1. Let ν : ZN → R+ be a measure and φ0 , . . . , φk−1 :
(ZN )k−1 → ZN some functions of the (k − 1) variables yi such that φi
doesn’t depend on yi for each 1 ≤ i ≤ k − 1. Suppose that f0 , . . . , fk−1 ∈
L1 (ZN ) are some functions satisfying |fi (x)| ≤ ν(x) for every x ∈ ZN and
0 ≤ i ≤ k − 1. For each 0 ≤ d ≤ k − 1 and 1 ≤ i ≤ k − 1, define

Y k−d−1
Y
Jd :=E  ( fi (φi (y (S) ))×
S⊂{k−d,...,k−1} i=0
!
k−1
Y 

× ν 1/2 (φi (y (S) ))y ∈ (ZN )k−1 , y ′ ∈ (ZN )d ,
i=k−d

and
 
Y
Pd := E  ν(φk−d−1 (y (S) ))|y ∈ (ZN )k−1 , y ′ ∈ (ZN )d  .
S⊂{k−d,...,k−1}

Then, for all 0 ≤ d ≤ k − 2, we have the inequality

|Jd |2 ≤ Pd Jd+1 .
Proof of Lemma 2.4.1. Consider Jd . Since φk−d−1 doesn’t depend on
yk−d−1 , we can extract the quantities depending on φk−d−1 from the av-
erage on yk−d−1 , so that we can write

Jd = E(G(y, y ′ )H(y, y ′ )|y1 , . . . , yk−d−2 , yk−d , . . . , yk−1 ,

′ ′
yk−d , . . . , yk−1 ∈ ZN ),
54 Chapter 2. Green-Tao-Szemerédi theorem

where
Y
G(y, y ′ ) := fk−d−1 (φk−d−1 (y (S) ))ν −1/2 (φk−d−1 (y (S) ))
S⊂{k−d,...,k−1}

and

Y k−d−2
Y
′
H(y, y ) := E( fi (φi (y (S) ))
S⊂{k−d,...,k−1} i=0
k−1
Y
× ν 1/2 (φi (y (S) ))|yk−d−1 ∈ ZN ).
i=k−d−1

Using the Cauchy-Schwarz inequality,

|Jd |2 ≤ E(|G(y, y ′ )|2 |y1 , . . . , yk−d−2 , yk−d , . . . , yk−1 ,

′ ′
yk−d , . . . , yk−1 ∈ ZN )×
× E(|H(y, y ′ )|2 |y1 , . . . , yk−d−2 , yk−d , . . . , yk−1 ,
′ ′
yk−d , . . . , yk−1 ∈ ZN ).

On the other hand, since |fk−d−1 (x)| ≤ ν(x) for every x,

E(|G(y, y ′ )|2 |y1 , . . . , yk−d−2 , yk−d , . . . , yk−1 , yk−d

′ ′
, . . . , yk−1 ∈ ZN ) ≤ Pd .

Moreover, by writing the definition of H(y, y ′ ) and expanding the squares

′
by changing the variable yk−d−1 by the new variables yk−d−1 and yk−d−1 ,
we see that

E(|H(y, y ′ )|2 |y1 , . . . , yk−d−2 , yk−d , . . . , yk−1 , yk−d

′ ′
, . . . , yk−1 ∈ ZN )
= Jd+1 .

This completes the proof.

Applying this lemma (k − 1) times, we obtain

k−2
Y
k−1 k−2−d
|J0 |2 ≤ Jk−1 Pd2 .
d=0

Observe that, by definition,

k−1
!
Y
k−1
J0 = E fi (φi (y))|y ∈ (ZN ) .
i=0
A. Arbieto, C. Matheus and C. G. Moreira 55

In order to prove the desired inequality, we choose6

X
k−1
ci

φi (y) := 1− yj ,
j=1
cj

so that φ0 (y) = y1 + · · · + yk−1 , φi (y) doesn’t depend on yi and, for all y,

φ(y) = x + ci r where
k−1
X yi
r=− .
c
i=1 i

Now, the surjective map Φ : (ZN )k−1 → (ZN )2 defined by

y1 yk−1
Φ(y) := (y1 + · · · + yk−1 , + ··· + )
c1 ck−1
has a constant number of pre-images, so that a simple calculation shows
that
  !
k−1
Y k−1
Y
E fj (x + cj r)|x, r ∈ ZN  = E fi (φi (y))|y ∈ (ZN )k−1
= J0 .
j=0 i=0

However, Pd = 1 + o(1) for each 0 ≤ d ≤ k − 2, since ν satisfies the

(2d , k − 1 + d, k)-linear forms condition. In particular, from the previous
estimates, we get
k−1
J02 ≤ (1 + o(1))Jk−1 .
Fix y. When S varies over the subsets of {1, . . . , k − 1}, φ0 (y (S) ) varies
on the (k − 1)-dimensional cube {x + w · h : w ∈ {0, 1}k−1 }, where x =
y1 + · · · + yk−1 and hi = yi′ − yi , i = 1, . . . , k − 1. Hence,
 
Y
Jk−1 = E W (x, h) f0 (x + w · h)|x ∈ ZN , h ∈ (ZN )k−1  ,
w∈{0,1}k−1

with the weight W (x, h) given by

k−1
Y k−2
Y
W (x, h) = E( ν 1/2 (φi (y + wh)) ×
w∈{0,1} i=1
1/2
ν (φk−1 (y + wh))|y1 , . . . , yk−2 ∈ ZN )
k−2
Y Y
= E( ν(φi (y + wh)) ×
i=1 w∈{0,1}k−1 ,wi =0
Y
ν(φk−1 (y + wh))|y1 , . . . , yk−2 ∈ ZN ),
w∈{0,1}k−1 ,wk−1 =0

6 Here we are using our hypothesis that cj are distinct.

56 Chapter 2. Green-Tao-Szemerédi theorem

where wh ∈ (ZN )k−1 is the vector of coordinates (wh)j := wj hj and

y ∈ (ZN )k−1 is the vector with components yj for 1 ≤ k − 2 and yk−1 :=
x − y1 − · · · − yk−2 . On the other hand, the definition of the Gowers norms
says that
 
Y
2k−1
E f0 (x + w · h)|x ∈ ZN , h ∈ (ZN )k−1  = kf0 kU k−1 .

w∈{0,1}k−1

Therefore, it suffices to show that

 
Y
E (W (x, h) − 1) f0 (x + w · h)|x ∈ ZN , h ∈ (ZN )k−1  = o(1).
w∈{0,1}k−1

Because |fj (x)| ≤ ν(x), we see that our task is reduced to prove
 
Y
E |W (x, h) − 1| ν(x + w · h)|x ∈ ZN , h ∈ (ZN )k−1  = o(1).
w∈{0,1}k−1

By the Cauchy-Schwarz inequality, this follows directly from the following

lemma:
Lemma 2.4.2 (ν uniformly covers its own cubes). For n = 0, 2, it holds
 
Y
E |W (x, h) − 1|n ν(x + w · h)|x ∈ ZN , h ∈ (ZN )k−1  = o(1).
w∈{0,1}k−1

Proof. By expanding the square, we see that it suffices to prove that, for
q = 0, 1, 2, it holds
 
Y
E W (x, h)q ν(x + w · h)|x ∈ ZN , h ∈ (ZN )k−1  = o(1).
w∈{0,1}k−1

However, this is a immediate consequence of the linear forms condition:

• for q = 0, we apply the (2k−1 , k, 1)-linear forms condition with vari-
ables x, h1 , . . . , hk−1 and linear forms x + w · h, w ∈ {0, 1}k−1 ;
• for q = 1, we apply the (2k−2 (k + 1), 2k − 2, k)-linear forms condition
with variables x, h1 , . . . , hk−1 , y1 , . . . , yk−2 and linear forms
(
φi (y + w · h), w ∈ {0, 1}k−1 , wi = 0 for 1 ≤ i ≤ k − 1
x + w · h, w ∈ {0, 1}k−1 ;
A. Arbieto, C. Matheus and C. G. Moreira 57

• for q = 2, we apply the (k2k−1 , 3k − 4, k)-linear forms condition with

variables
x, h1 , . . . , hk−1 , y1 , . . . , yk−2 , y1′ , . . . , yk−2
′

and linear forms


 k−1
φi (y + w · h), w ∈ {0, 1} , wi = 0 for 1 ≤ i ≤ k − 1
φi (y + w · h), w ∈ {0, 1}k−1 , wi = 0 for 1 ≤ i ≤ k − 1
′


x + w · h, w ∈ {0, 1}k−1 ;

Here we are adopting the conventions yk−1 = x − y1 − · · · − yk−2 and

′
yk−1 = x − y1′ − · · · − yk−2
′
. Clearly, this completes the proof of the
lemma.
As we told before, this ends the proof of the generalized von Neumann
theorem (Theorem 2.4.1).
Remark 2.4.2. Note that we used only the linear forms condition during
the proof of Theorem 2.4.1.
Closing the study of the Gowers norms, we state the following simple
and useful lemma about the Gowers distance k.kU k−1 between an arbitrary
k-pseudorandom measure ν and the k-pseudorandom measure νconst ≡ 1:
Lemma 2.4.3. Suppose that ν is a k-pseudorandom measure. Then,
kν − 1kU d = o(1),
for every d ≤ k − 1.
Proof. Observe that the linear forms condition for ν easily implies that
kνkU k−1 = 1 + o(1). But, we can refine this reasoning: indeed, let us note
that the monotonicity property of the Gowers norms means that it suffices
to prove kν − 1kU k−1 = o(1). Multiplying by the factor 2k−1 , this reduces
our task to prove
 
Y 
E ν(x + w · h)x ∈ ZN , h ∈ ZN
k−1 
= o(1).
w∈{0,1}k−1

The left-hand side of this identity can be expanded as

!
X Y 
(−1)|A|
E ν(x + w · h)x ∈ ZN , h ∈ Zk−1
N .
A⊂{0,1}k−1 w∈A

Looking at the expression

!
Y 
E ν(x + w · h)x ∈ ZN , h ∈ ZN
k−1

w∈A
58 Chapter 2. Green-Tao-Szemerédi theorem

for some fixed A ⊂ {0, 1}k−1 , we see that it has the form


E ν(φ1 (x)) . . . ν(φ|A| (x))x ∈ ZkN ,
where x := (x, h1 , . . . , hk−1 ) and φ1 , . . . , φ|A| are an ordering of the |A|
linear forms x 7→ x + w · h with w ∈ A. Obviously, each of these linear
forms isn’t a rational multiple of any other, so that the (2k−1 , k, 1)-linear
forms condition can be used to conclude the proof of the lemma.
Let us resume the discussion of this subsection.
Resume of the subsection “Gowers norms”:
In this subsection we identified a class of norms (namely, Gowers norms)
naturally associated to the problem of counting arithmetic progressions
whose elements belong to the support of a given family of functions and
we proved Theorem 2.4.1 saying that the Gowers norms can effectively
give upper bounds on the number of such progressions up to a negligible
error. As we saw during the proof of Roth theorem, this good upper bound
is useful during the task of getting lower bounds of certain expectations
(our primary goal). The next step will be to introduce the concept of
anti-uniformity, which plays a fundamental role during the decomposition
of arbitrary functions into good and bad parts.

2.4.2 Anti-Uniformity
Since the Gowers norms (for d ≥ 2) are genuine norms, we can consider
the dual norms:
kgk(U k−1 )∗ := sup |hf, gi|,
kf kU k−1 ≤1

where hf, gi denotes the usual L2 product. We say that g is anti-uniform

if kgk(U k−1 )∗ = O(1) and kgkL∞ = O(1).
Remark 2.4.3. Although we aren’t going to use this fact, observe that,
for k = 3, the remark 2.4.1 gives the formula:
 3/4
X
kgk(U 2 )∗ =  g (ξ)|4/3  .
|b
ξ∈ZN

Observe that, if g is anti-uniform and |hf, gi| is large then f can’t

be uniform (i.e., kf kL∞ = O(1) and kf kU k−1 = O(1)) since |hf, gi| ≤
kf kU k−1 kgk(U k−1 )∗ . Thus, this gives us an obstruction to the uniformity.
Moreover, we have a canonical way to construct anti-uniform functions:
given F ∈ L1 (ZN ), we define the dual of F by:
k−1
DF (x) := E(Πw6=0 F (x + w.h)|h ∈ ZN ).
Among the several elementary properties of these functions, we can quote:
A. Arbieto, C. Matheus and C. G. Moreira 59

Lemma 2.4.4. Let ν be a k-pseudorandom measure and F ∈ L1 (ZN ) be

an arbitrary function. It holds:
k−1
2
• hF, DF i = kF kU k−1 ,

k−1
• kDF k(U k−1 )∗ = kF k2U k−1−1 and
k−1
• if |F | ≤ 1 + ν, then kDF kL∞ ≤ 22 −1
+ o(1).
k−1
2
Proof. The identity hF, DF i = kF kU k−1 follows directly from the defini-

tions of Gowers norms and DF , and we left its verification as an exercise.

Concerning the second identity, consider F 6= 0 (since the case F = 0 is
trivial) and note that the definition of the dual norms combined with the
2k−1
identity hF, DF i = kF kU k−1 say that it is sufficient to show that, for any

function f , we have
k−1
2 −1
|hf, DF i| ≤ kf kU k−1 kF kU k−1 .

However, the definition of DF shows that hf, DF i is the Gowers inner

product of h(fw )w∈{0,1}k−1 iU k−1 where fw := f when w = 0 and fw := F
otherwise, so that the desired inequality above follows from the Gowers-
Cauchy-Schwarz inequality.
Finally, the last item follows from the linear forms condition. In fact,
since F is bounded by 2(1 + ν)/2 := 2ν1/2 , we see our task is reduced to
prove
Dν1/2 (x) ≤ 1 + o(1)

uniformly for every x ∈ ZN . On the other hand, the definition of the dual
function says that Dν1/2 can be expanded as
  
Y 

E k−1 
ν1/2 (x + w · h)  h ∈ ZN .

w∈{0,1}k−1 −{0}

Since ν1/2 is a k-pseudorandom measure, the linear forms condition implies

that this term is 1 + o(1).

Remark 2.4.4. The proof of this lemma is the unique place where we
use the linear forms condition with inhomogeneuos terms bi 6= 0; indeed,
during the previous argument, all of the bi ’s are equal to x.

We call basic anti-uniform functions any dual function DF where F is

pointwise bounded by 1 + ν; a relevant property of these functions is its
good distribution with respect to ν:
60 Chapter 2. Green-Tao-Szemerédi theorem

Proposition 2.4.1. If ν is k-pseudorandom, Φ : I K → R is continuous

and DF1 , . . . , DFK are basic anti-uniform functions, we define
Ψ(x) = Φ(DF1 (x), . . . , DFK (x)).
Then, hν − 1, Ψi = ok,Φ (1). Furthermore, the right-hand side quantity can
be taken uniform on any compact set of Φ’s.
Proof. The basic idea is to use Weierstrass approximation theorem and the
fact that ν is a measure to reduce our task to the proof of this proposition
to the “simpler” case of a polynomial Φ.
Observe that, by replacing ν by (ν + 1)/2, we can assume |Fj (x)| ≤ ν(x)
for all x ∈ ZN , 1 ≤ j ≤ K.
Lemma 2.4.5. Let d ≥ 1. For any polynomial function P of degree d with
real-valued coefficients (independents of N ) it holds
kP (DF1 , . . . , DFK )k(U k−1 )∗ = OK,d,P (1).
Proof. By linearity (and increasing K to dK if necessary), it suffices to
show the lemma when P (x1 , . . . , xK ) = x1 . . . xK . In other words, we want
to verify
YK
hf, DFj i = OK (1)
j=1

for any f : ZN → R with kf kU k−1 ≤ 1. Expanding the left-hand side as

  
YK Y 

E f (x) E( Fj (x + w · h(j) )|h(j) ∈ (ZN )k−1 )x ∈ ZN 

j=1 k−1
w∈{0,1} :w6=0

Making the change of variables h(j) = h + H (j) for any h ∈ (ZN )k−1 ,
taking the averages on h, we expand the products on j and interchanging
the expectations, we can rewrite this expression in terms of Gowers inner
product:
E(h(fw,H )w∈{0,1}k−1 iU k−1 |H ∈ ((ZN )k−1 )K ),
where H = (H (1) , . . . , H (K) ), f0,H := f , fw,H := gw·H for w 6= 0 with
w · H = (w · H (1) , . . . , w · H (K) ) and
K
Y
gu(1) ,...,u(K) (x) := Fj (x + u(j) ) for all u(1) , . . . , u(K) ∈ ZN .
j=1

In particular, the Gowers-Cauchy-Schwarz inequality and the fact kf kU k−1

≤ 1 reduce the proof of the lemma to prove the estimate
 
Y
E kgw·H kU k−1 |H ∈ ((ZN )k−1 )K  = OK (1).
w∈{0,1}k−1 :w6=0
A. Arbieto, C. Matheus and C. G. Moreira 61

By Hölder inequality, it suffices to show

k−1
E(kgw·H k2U k−1−1 |H ∈ ((ZN )k−1 )K )) = OK (1),

for each w ∈ {0, 1}k−1 − 0.

Fix w. Since 2k−1 − 1 ≤ 2k−1 and we have dealing with probability
spaces, it suffices to prove
k−1
2 k−1 K
E(kgw·H kU k−1 |H ∈ ((ZN ) ) )) = OK (1).

This last estimate is true by the following argument: w 6= 0 implies that

w → w · H is a covering map; this allows us to use it to perform a change
of variables so that the left-hand side of the previous identity is
k−1
2 (1)
E(kgu(1) ,...,u(K) kU k−1 |u , . . . , u(K) ∈ ZN ).

Using the definitions of the Gowers norms and gu(1) ,...,u(K) , we can expand
this term as
  
Y K
Y 
(j)  (1) (K)
E Fj (x + u + h · w)
e  x, u , . . . , u k−1 
∈ ZN , h ∈ ZN .
k−1 j=1

e
w∈{0,1}

By factorizing, we obtain
  
K
Y 

E E(Fj (x + u(j) + h · w) k−1 
e | u(j) ∈ ZN )  x ∈ ZN , h ∈ ZN .

j=1

Using the assumption |Fj (x)| ≤ ν(x), our task is reduced to the verification
of the estimate
 !

 k−1
e | u ∈ ZN )K  x ∈ ZN , h ∈ ZN
E E(ν(x + u + h · w) .


Performing the change of variables y = x + u and taking the average over

x, our goal is to prove
 !

K k−1
e | y ∈ ZN )  h ∈ ZN
E E(ν(y + h · w) = OK (1).


At this point, we are ready to use the correlation condition, which says
that
 !
 X

E ν(y + h · w)
e  y ∈ ZN ≤ τ (h · (w e′ )),
e−w
 k−1
′ w,
e we ∈{0,1}
′ ,w6
e =w
e
62 Chapter 2. Green-Tao-Szemerédi theorem

where τ is a weight function satisfying E(τ q ) = Oq (1). Using the triangular

k−1
inequality in LK (ZN ), we see that it suffices to prove
 !

′ K k−1
E τ (h · (w e−w e ))  h ∈ ZN = OK (1),


for all w,
e we′ ∈ {0, 1}k−1 distinct. But, since h 7→ h · (w e′ ) is a covering
e−w
K
map, the left-hand side is E(τ ) = OK (1).

Now let’s go back to the proof of the proposition. Recall that the
lemma 2.4.4 says that a basic anti-uniform function takes its values in
k−1 k−1
the interval I = [−22 , 22 ]. By Weierstrass approximation theorem,
given ε > 0, there exists a polynomial P close to the continuous function
Φ in the sense that

kΦ(DF1 , . . . , DFK ) − P (DF1 , . . . , DFK )kL∞ ≤ ε.

Since ν is a measure (i.e., E(ν) = 1 + o(1)), we have

|hν − 1, Φ(DF1 , . . . , DFK ) − P (DF1 , . . . , DFK )i| ≤ (2 + o(1))ε.

On the other hand, combining the lemmas 2.4.3, 2.4.5, we obtain

|hν − 1, P (DF1 , . . . , DFK )i| = oK,ε (1)

because P depends only on K and ε. Making N large (depending on K, ε),

we see that
|hν − 1, Φ(DF1 , . . . , DFK )i| ≤ 4ε.
This ends the proof of the proposition 2.4.1.

Remark 2.4.5. The unique place in this book where we used the correla-
tion condition was during the final part of the proof of Lemma 2.4.5.

Resume of the subsection “Anti-Uniformity”:

In this subsection we introduced the notion of anti-uniformity, which is

useful for the identification of non-uniform functions (that is, they give
obstructions to uniformity). In fact, we saw that any function F naturally
generates a (basic) anti-uniform function DF such that the correlation
hF, DF i is large whenever F isn’t uniform; moreover, we showed a result
saying that a pseudorandom measure is well-distibuted with respect to the
algebra generated by the basic anti-uniform functions.
In the sequel, we will study the sigma-algebras generated by the level sets
of anti-uniform functions, which are the basic pieces of the sigma-algebra
whose conditional expectations provides good (=anti-uniform) functions.
A. Arbieto, C. Matheus and C. G. Moreira 63

2.4.3 Sigma-algebras generated by basic anti-uniform

functions
The following proposition says that basic anti-unform functions naturally
generate sigma-algebras where they are well-behaved (i.e., one can use
Szemerédi theorem in its original form).

Proposition 2.4.2. If ν is k-pseudorandom measure and DF1 , . . . , DFK

are basic anti-uniform functions, then, for every ǫ < 1 and σ < 1/2, there
exists a sigma-algebra B such that, if N is a large prime, it holds:

• kDFj − E(DFj |B)kL∞ ≤ ǫ ∀j.

• There is an exceptional subset Ω ⊂ B such that E((ν + 1)1Ω ) =

OK,ǫ (σ 1/2 ).

• k(1 − 1Ω )E(ν − 1|B)kL∞ = OK,ǫ (σ 1/2 ).

Proof. The starting point of the argument is the following lemma ensuring
that each function generates a sigma-algebra:

Lemma 2.4.6. Let ν be a k-pseudorandom measure, 0 < ǫ < 1 and

0 < σ < 1/2 be parameter, and G ∈ L∞ (ZN ) be a function taking its
k−1 k−1
values on the interval I := [−22 , 22 ]. Then, there exists Bǫ,σ (G) a
sigma-algebra such that

• (G belongs to its own σ-algebra) For every σ-algebra B, we have

kG − E(G|B ∨ Bǫ,σ (G))kL∞ (ZN ) ≤ ǫ.

• (Bounded complexity) Bǫ,σ (G) has O(1/ǫ) atoms.

• (Nice approximation by continuous functions of G) If A is an atom

of Bǫ,σ (G), then there is a ΨA : I → [0, 1] such that

k(1A − ΨA (G))(ν + 1)kL1 (ZN ) = O(σ).

Moreover, ΨA belongs to a compact subset E ⊂ C 0 (I) independent

of G, ν, N and A.

Proof of lemma 2.4.6. Putting together Fubini’s theorem with the fact
that ν is a measure, we have
Z 1 X
E(1G(x)∈[ǫ(n−σ+α),ǫ(n+σ+α)] (ν(x) + 1)|x ∈ ZN )dα
0 n∈Z

= 2σE(1 + ν(x)|x ∈ ZN ) = O(σ).

64 Chapter 2. Green-Tao-Szemerédi theorem

Hence, we can fix α such that

X
E(1G(x)∈[ǫ(n−σ+α),ǫ(n+σ+α)] (ν(x) + 1)|x ∈ ZN ) = O(σ). (2.5)
n∈Z

Define Bǫ,σ (G) as the σ-algebra whose atoms are G−1 ([ǫ(n + α), ǫ(n + α +
1)]) for n ∈ Z. Note that Bǫ,σ (G) is well-defined because the intervals
[ǫ(n + α), ǫ(n + α + 1)] are a partition of the real line.
Clearly, if B is a σ-algebra, then the function G restricted to a atom of
B ∨ Bǫ,σ (G) takes its values on an interval of diameter ǫ, which gives the
first item of the lemma (G belongs to its own σ-algebra). Now, let A :=
G−1 ([ǫ(n + α), ǫ(n + α + 1)]) be a atom of Bǫ,σ (G). Since G takes its values
on I, we have n = O(1/ǫ) (otherwise, A = ∅). This proves the second item
of the lemma (bounded complexity). Finally, let ψσ : R → [0, 1] be a fixed
bump function such that ψσ = 1 on [σ, 1 − σ] and ψσ = 0 on [−σ, 1 + σ],
and define ΨA (x) := ψσ ( xǫ − n − α). Obviously, ΨA varies on a compact
subset Eǫ,σ of C 0 (I) (since n and α are bounded) and the identity (2.5)
implies the third item of the lemma (nice approximation by continuous
functions of G).
At this point, we come back to the proof of Proposition 2.4.2. We take
B := Bǫ,σ (DF1 ) ∨ · · · ∨ Bǫ,σ (DFK ), where Bǫ,σ (DFj ) is the sigma-algebra
provided by Lemma 2.4.6. Clearly the first item of Proposition 2.4.2 fol-
lows from the first item of Lemma 2.4.6. On the other hand, since each
Bǫ,σ (DFj ) has O(1/ǫ) atoms, B is generated by OK,ǫ (1) atoms. We say
that an atom A of B is small if E((ν + 1)1A ) ≤ σ 1/2 . Denote by Ω the
union of all small atoms. Then, Ω ∈ B and the second item of Propo-
sition 2.4.2 holds. In order to prove the last item of this proposition, it
suffices to show
E((ν − 1)1A )
= E(ν − 1|A) = oK,ǫ,σ (1) + OK,ǫ (σ 1/2 )
E(1A )
for all non-small atom A. From the smallness definition, we have

E((ν − 1)1A ) + 2E(1A ) = E((ν + 1)1A ) ≥ σ 1/2

for a non-small A. Thus, since σ is small and N is large, our task is reduced
to the verification of

E((ν − 1)1A ) = oK,ǫ,σ (1) + OK,ǫ (σ 1/2 ).

However, since A is the intersection of K atoms Aj ∈ Bǫ,σ (DFj ), j =

1, . . . , K, Lemma 2.4.6 and Hölder inequality show that there exists ΨA :
I K → [0, 1] such that

k(ν + 1)(1A − ΨA (DF1 , . . . , DFK ))kL1 (ZN ) = OK (σ),

A. Arbieto, C. Matheus and C. G. Moreira 65

so that

k(ν − 1)(1A − ΨA (DF1 , . . . , DFK ))kL1 (ZN ) = OK (σ).

Furthermore, ΨA belongs to a compact set Eǫ,K,σ of C 0 (I K ). This fact

and Proposition 2.4.1 (of uniform distribution with respect to basic anti-
uniform functions) imply

E((ν − 1)ΨA (DF1 , . . . , DFK )) = oK,ǫ,σ (1) = OK,ǫ (σ 1/2 ),

because N is large (depending on K, ǫ, σ). Of course, this estimate and

the triangular inequality conclude the argument.

Resume of the subsection “Sigma-Algebras generated by basic

anti-uniform functions”:

In this subsection we associated to each basic anti-uniform function DF

a sigma-algebra B so that the conditional expectation E(DF |B) approxi-
mates DF (i.e., DF is almost constant on each atom of B) and the pseu-
dorandom measure ν takes values close to 1 on average (with respect to
B).
The next stage consists into the utilization of the machinery of basic
anti-uniform functions and its associated sigma-algebras to formalize the
decomposition process of an arbitrary function into good (anti-uniform)
and bad (uniform) parts via an inductive scheme. The key point will be to
guarantee that this inductive procedure stops in a finite number of steps.
This fact will be a consequence of an energy increment argument.

2.4.4 The energy increment argument

Using the sigma-algebras generated by basic anti-uniform functions, we can
obtain the desired decomposition into uniform and anti-uniform parts:

Theorem 2.4.2 (generalized Koopman-von Neumann theorem). Con-

sider ν a k-pseudorandom measure, f ∈ L1 such that 0 ≤ f ≤ ν, ǫ << 1
and N a large prime number. Then, there exists a sigma-algebra B and an
exceptional subset Ω ∈ B such that:

• E(ν · 1Ω ) = oǫ (1) (the exceptional set is small).

• k(1 − 1Ω )E(ν − 1|B)kL∞ = oǫ (1) (good distribution of functions out-

side the exceptional subset).
k
• k(1 − 1Ω )(f − E(f |B)kU k−1 ≤ ǫ1/2 (uniformity on B)
66 Chapter 2. Green-Tao-Szemerédi theorem

Proof. The basic strategy is the same of the structure theorem of Fursten-
berg7 : we start with the trivial sigma-algebra B = {∅, ZN } and we look
at the function f − E(f |B). If it is uniform (i.e., the third item above
holds), we are done. Otherwise, we use the results about anti-uniformity
to find an anti-uniform function G1 with non-trivial correlation with f
and we add the level sets of G1 to the sigma-algebra B. The non-trivial
correlation property will tell us that the L2 norm of E(f |B) increases by a
non-trivial amount8 , while the pseudorandomness of ν shows that E(f |B)
stays uniformly bounded. At this point, we repeat this procedure until
f − E(f |B) becomes sufficiently uniform; note that the algorithm stops
k
in a finite number of steps (of order 22 /ǫ) due to the definite energy
increment at each step.
Now let us write this strategy in a more organized manner. Fix ǫ and let
k
K0 be the smallest integer greater than 1+22 /ǫ. We will need a parameter
0 < σ ≪ ǫ and we will take N large depending on ǫ and σ. We construct
B and Ω via a sequence of basic anti-uniform functions DF1 , . . . , DFK
on ZN , exceptional subsets Ω0 ⊂ · · · ⊂ ΩK ⊂ ZN and sigma-algebras
B0 ⊂ · · · ⊂ BK for some 0 ≤ K ≤ K0 as follows:

• Stage 0: We begin with K=0, B0 := {∅, ZN } and Ω0 := ∅.

• Stage 1: We put FK+1 := (1 − 1ΩK )(f − E(f |BK )). If

k
kFK+1 kU k−1 ≤ ǫ1/2

we define Ω := ΩK , B = BK and we end the algorithm successfully.

Otherwise, we go to the stage 2.

• Stage 2: If
k
kFK+1 kU k−1 > ǫ1/2
we define BK+1 := BK ∨ Bǫ,σ (DFK+1 ) and we go to the stage 3.

• Stage 3: We look for an exceptional subset ΩK+1 ⊃ ΩK in BK+1

with
E((ν + 1)1ΩK+1 ) = OK,ǫ (σ 1/2 ) (2.6)
and
k(1 − 1ΩK+1 )E(ν − 1|BK+1 )kL∞ = OK,ǫ (σ 1/2 ).
If such an exceptional subset can’t be located, we end the algorithm
with an error. Otherwise, we proceed to the stage 4.
7 This result says that we can decompose any dynamical system in a weak-mixing

extension of a tower of compact extensions.

8 The idea that non-trivial correlation implies an augmentation of the L2 norm is

precisely the energy increment argument.

A. Arbieto, C. Matheus and C. G. Moreira 67

• Stage 4: We replace K by K + 1. If K > K0 , we end the algorithm

with an error. Otherwise, we go back to stage 1.
Assume momentarily that the algorithm ends without any errors in the
stages 3 or 4. Then, it is clear that after K0 + 1 interactions (at most), we
constructed a σ-algebra B and an exceptional subset Ω with the desired
properties, except by the fact that the error terms are OK,ǫ (σ 1/2 ) instead
of oǫ (1), for a large N depending on σ, K, ǫ. However, this little difficulty
is easily overcome by making σ converging to zero very slowly.
In other words, we reduced the proof of this theorem to show that the
algorithm ends without any errors. The argument is inductive: as an in-
duction hypothesis on 0 ≤ K1 ≤ K0 , suppose that the algorithm ends
without any errors or it attains the stage 2 at the K1 -th interaction with-
out any errors. Note that this is obvious for K1 = 0. Assuming this
is valid for some K1 < K0 , we would like to show that this is true for
K1 + 1. Observe that we can assume that the algorithm doesn’t end be-
fore the stage 2 of the K1 -th interaction. At this stage, we have σ-algebras
B0 , . . . , BK1 +1 , basic anti-uniform functions DF1 , . . . , DFK1 +1 and excep-
tional subsets Ω0 , . . . , ΩK1 already constructed. We claim that
k−1
kDFj kL∞ ≤ 22 + Oj,ǫ (σ 1/2 ), (2.7)

for any 1 ≤ j ≤ K1 + 1. This follows from the stage 3 of the previous

interactions (or the stage 0 when j = 1) because they say that

k(1 − 1Ωj−1 )E(ν − 1|Bj−1 )kL∞ = Oj,ǫ (σ 1/2 ),

so
E(ν|Bj−1 )(x) = 1 + Oj−1,ǫ (σ 1/2 ),
for all x ∈
/ Ωj−1 . Since 0 ≤ f (x) ≤ ν(x), we conclude the pointwise
estimates

0 ≤ (1 − 1Ωj−1 (x))E(f |Bj−1 )(x) ≤ 1 + Oj,ǫ (σ 1/2 ), (2.8)

so that, we can see that, by the definition of Fj ,

|Fj (x)| ≤ (1 + Oj,ǫ (σ 1/2 ))(ν(x) + 1). (2.9)

In particular, a simple application of Lemma 2.4.4 proves our claimed

estimate (2.7).
On the other hand, since BK1 +1 is the σ-algebra generated by

Bǫ,α1 (DF1 ), . . . , Bǫ,αK1 +1 (DFK1 +1 ),

Proposition 2.4.2 permits to find some subset Ω such that

E((ν + 1)1Ω ) = OK1 ,ǫ (σ 1/2 )

68 Chapter 2. Green-Tao-Szemerédi theorem

and
k(1 − 1Ω )E(ν − 1|BK1 +1 )kL∞ = OK1 ,ǫ (σ 1/2 ).
Define ΩK1 +1 := Ω ∪ ΩK1 . Obviously, ΩK1 +1 has the required properties
to execute the stage 3 without errors, so that we can go to the stage 2 of
the K1 + 1-th interaction (or we end the algorithm without any errors), so
that the inductive argument is complete.
In other words, at this moment, we proved that there are only two
possibilities for the algorithm: either it ends without errors or it goes all
the way to the K0 -th interaction. In order to conclude the proof of the
theorem, we affirm that, if the algorithm reaches the stage 3 of the K0 -
th iterate, then the following key-property (namely, the energy increment
estimate) holds:

k(1 − 1Ωj )E(f |Bj )k2L2

≥ k(1 − 1Ωj−1 )E(f |Bj−1 )k2L2 (2.10)
2k −2 1/2 2
+2 ǫ − Oj,ǫ (σ ) − O(ǫ ),

for every 1 ≤ j ≤ K0 (if N is large depending on K0 and ǫ). Of course,

this key property is sufficient to conclude the proof of the theorem because
the estimate (2.8) gives us

0 ≤ k(1 − 1Ωj )E(f |Bj )k2L2 ≤ 1 + Oj,ǫ (σ 1/2 ), (2.11)

k
for any 0 ≤ j ≤ K0 . Since K0 is the smallest integer greater than 22 /ǫ+1,
the pigeonhole principle leads us to a contradiction for sufficiently small ǫ
and σ, and for a sufficiently large N depending on ǫ, σ.
Finally, it remains only to prove the energy increment estimate. The
idea is to exploit the fact that the algorithm doesn’t stop at the stage 2 of
the (j − 1)-th iteraction: indeed, observe that this fact implies
k
kFj kU k−1 ≥ ǫ1/2 .

Combining this estimate with the definition of Fj and Lemma 2.4.4, we

obtain
k−1
2 1/2
|h(1 − 1Ωj−1 )(f − E(f |Bj−1 )), DFj i| = kFj kU k−1 ≥ ǫ .

On the other hand, the pointwise estimates (2.7), (2.9), (2.6) above show
that
h(1Ωj − 1Ωj−1 )(f − E(f |Bj−1 ), DFj )i = Oj,ǫ (σ 1/2 ),
while Lemma 2.4.6 and the estimate (2.9) tell us

h(1 − 1Ωj )(f − E(f |Bj−1 ), DFj − E(DFj |Bj ))i = O(ǫ).
A. Arbieto, C. Matheus and C. G. Moreira 69

In particular, by the triangular inequality, we obtain the lower bound:

|h(1 − 1Ωj )(f − E(f |Bj−1 )), E(DFj |Bj )i| ≥ ǫ1/2 − Oj,ǫ (σ 1/2 ) − O(ǫ).

Since the functions (1−1Ωj ), E(DFj |Bj ) and E(f |Bj−1 ) are Bj -measurable,
we can replace f by E(f |Bj ), so that we get

|h(1 − 1Ωj )(E(f |Bj ) − E(f |Bj−1 )), E(DFj |Bj )i| ≥ ǫ1/2 − Oj,ǫ (σ 1/2 ) − O(ǫ).

Using the Cauchy-Schwarz inequality and the estimate (2.7), we conclude:

k−1
k(1 − 1Ωj )(E(f |Bj ) − E(f |Bj−1 ))kL2 ≥ 2−2 +1 1/2
ǫ− Oj,ǫ (σ 1/2 ) − O(ǫ).
(2.12)
This inequality morally implies, by the Pythagorean theorem, the energy
increment estimate, although there is a little problem due to the presence
of the exceptional subsets Ωj−1 , Ωj , which are not a priori trivial to control
because we don’t dispose of good L2 bounds for ν. In order to treat this
little technicality, we notice that (2.6) and (2.8) imply

k(1Ωj − 1Ωj−1 )E(f |Bj−1 )kL2 = Oj,ǫ (σ 1/2 ).

Thus, the triangular inequality and (2.11) show that the energy increment
estimate (2.10) follows from the following estimate

k(1 − 1Ωj )E(f |Bj )k2L2

≥ k(1 − 1Ωj−1 )E(f |Bj−1 )k2L2 + ǫ1/2 − Oj,ǫ (σ 1/2 ) − O(ǫ).

However, the left-hand side above can be expanded (by the cosine law) as

k(1 − 1Ωj )E(f |Bj−1 )k2L2 + k(1 − 1Ωj )(E(f |Bj ) − E(f |Bj−1 ))k2L2
+ 2h(1 − 1Ωj )E(f |Bj−1 ), (1 − 1Ωj )(E(f |Bj ) − E(f |Bj−1 ))i.

Therefore, by (2.12), we see that it suffices to prove the following quasi-

orthogonality relation:

h(1 − 1Ωj )E(f |Bj−1 ), (1 − 1Ωj )(E(f |Bj ) − E(f |Bj−1 ))i = Oj,ǫ (σ 1/2 ).

Since (1 − 1Ωj )2 = (1 − 1Ωj ), we can rewrite this identity as

h(1 − 1Ωj )E(f |Bj−1 ), E(f |Bj ) − E(f |Bj−1 )i = Oj,ǫ (σ 1/2 ).

Now we observe that, since the function (1−1Ωj−1 )E(f |Bj−1 ) is measurable
with respect to Bj−1 , it is orthogonal to the function E(f |Bj ) − E(f |Bj−1 )
because Bj−1 is a sub-sigma-algebra of Bj (by construction). In particular,
we can again rewrite the previous expression as

h(1Ωj − 1Ωj−1 )E(f |Bj−1 ), E(f |Bj ) − E(f |Bj−1 )i = Oj,ǫ (σ 1/2 ).
70 Chapter 2. Green-Tao-Szemerédi theorem

Using again the fact that (1Ωj − 1Ωj−1 )E(f |Bj−1 ) is a Bj -measurable (so
that it must be orthogonal to f −E(f |Bj )), we see that the previous identity
is equivalent to

h(1Ωj − 1Ωj−1 )E(f |Bj−1 ), f − E(f |Bj−1 )i = Oj,ǫ (σ 1/2 ).

However, this equality is certainly true in view of 0 ≤ f ≤ ν and the

estimates (2.6), (2.8). This proves the energy increment estimate (2.10)
and, consequently, the proof of Theorem 2.4.2.

Resume of the subsection “The energy increment argument”:

In this subsection, we used the machinery of sigma-algebras associated

to the anti-uniform functions to exhibit, for a density f bounded by a
pseudo-random measure, an algorithmic construction of small exceptional
subsets and a sigma-algebra such that the function f has an uniform be-
havior outside the exceptional subset. In particular, this shows us how
to decompose f into an uniform (bad) part and an anti-uniform (good)
part. This was the content of the generalized Koopman-von Neumann
theorem (Theorem 2.4.2). Moreover, the algorithm related to the proof of
the generalized Koopman-von Neumann theorem was finite (i.e., it stops
after a finite steps of interaction) due to the energy increment argument
at each step (indeed, since the energy was always increasing by a definite
amount and it is a priori bounded during the whole process, this gives
immediately the desired conclusion).
The last step of the proof of Green-Tao-Szemerédi theorem is the ap-
plication of this decomposition (provided by Theorem 2.4.2) in order to
conclude the argument.

2.4.5 End of the proof of Green-Tao-Szemerédi theo-

rem
Once we formalized (and quantified) the machinery related to the unifor-
mity and anti-uniformity issues, and the decomposition of arbitrary func-
tions into uniform and non-uniform pieces, it is fairly easy to imitate the
scheme proposed in the section 2.3 during the proof of Roth theorem in
order to get a proof of Green-Tao-Szemerédi theorem (Theorem 2.2.1):
Let f , ν and δ be the objects appearing in the statement of Theo-
rem 2.2.1. Take ǫ << δ and consider B the sigma-algebra provided by the
generalized Koopman-von Neumann theorem (Theorem 2.4.2). Define the
functions:

• fU = (1 − 1Ω )(f − E(f |B))

• fAU = (1 − 1Ω )E(f |B).

A. Arbieto, C. Matheus and C. G. Moreira 71

Recall that, by hypothesis, 0 ≤ f ≤ ν (pointwise) and E(f ) ≥ δ. Hence,

Theorem 2.4.2 ensures that

E(fAU ) = E((1 − 1Ω )f ) ≥ E(f ) − E(ν1Ω ) ≥ δ − oǫ (1).

Moreover, we have fAU ≤ 1+oǫ (1), so that we can use Szemerédi theorem9 .
In particular:

E(fAU (n) . . . fAU (n + (k − 1)r)|n, r ∈ ZN ) ≥ c(k, δ) − oǫ (1).

k
On the other hand, we know that kfU kU k−1 ≤ ǫ1/2 . Thus, by the gener-
alized von Neumann theorem 2.4.1, we have:
k
E(f∗1 (n) . . . f∗k (n + (k − 1)r)|n, r ∈ ZN ) = O(ǫ1/2 )

where ∗j = U or AU , and ∗j = U for at least one index j.

Therefore, we obtain:
k
E(f (n) . . . f (n + (k − 1)r)|n, r ∈ ZN ) ≥ c(k, δ) − O(ǫ1/2 ) − oǫ (1).

Since ǫ is arbitrary, Green-Tao-Szemerédi theorem is proved!

Remark 2.4.6. The careful reader noticed the strong analogy between the
previous estimates and the estimates of the proposition 2.3.1. In fact, as
it is natural to expect, we separated, in both arguments, a “good” term (it
was Λ3 (g, g, g) in Roth case and E(fAU (n) . . . fAU (n + (k − 1)r)|n, r ∈ ZN )
in Green-Tao case) which is relatively big (its order was δ 3 in Roth case
and c(k, δ) − oǫ (1) in Green-Tao case) and our task was to control the
“bad” terms (Λ3 (., ., .) where the bad function b must appear in some entry
in Roth case and E(f∗1 (n) . . . f∗k (n + (k − 1)r)|n, r ∈ ZN ) where the bad
function fU must appear in some entry in Green-Tao case). In order
to accomplish this goal, we used Hölder inequality in Roth case and the
generalized von Neumann theorem in Green-Tao case to reduce the problem
to the (non-trivial) “fact” that b in Roth case and fU in Green-Tao can
be taken uniform. Logically, this fact was obtained from the generalized
Koopman-von Neumann theorem in Green-Tao case, which uses (in its
proof ) the energy increment argument, as we annouced in the beginning to
the section.

9 In fact we are omitting a little detail here: since f

AU is not exactly bounded from
above by 1 and E(fAU ) is not exactly bounded from below by δ, Szemerédi theorem
doesn’t apply directly. However, this is easily overcome by using Szemerédi theorem for
a function equal to f modulo a term of the form oǫ (1) which can be trivially controlled
in this case.
 Part II

Elkies-McMullen theorem
Chapter 3

Elkies-McMullen theorem

3.1 Distribution of sequences on the circle

A basic problem in Number Theory consists into the study of the distri-
bution of a given sequence of real numbers on the circle R/Z.
More precisely, given a real number x ∈ R, let us denote by
{x} = x(mod 1) ∈ S 1 := R/Z
its fractionary part. In this setting, the problem quoted above can be
described as follows: given a sequence of real numbers xn , n = 1, 2, . . . ,
what can we say about the behavior of the sequence {xn } on the circle S 1 ?
The distribution problem of given sequences (mod 1) is very old, so
that we will make just a few comments. Firstly, a classical result (due to
Kronecker) says that the sequence {nθ} is dense in S 1 for all irrational
θ ∈ R. Another classical result (due to Weyl) says that the same sequence
{nθ} (for irrational θ) is equidistributed, i.e., for all interval I ⊂ S 1 ,
#{1 ≤ n ≤ N : {nθ} ∈ I} |I|
→ 1 when N → ∞,
N |S |
where |J| denotes the length of J. Furthermore, it is known that the
sequence {κn } is equidistributed for almost every κ > 1 (although certain
specific cases such as {(3/2)n } being still open). More generally, we have
Weyl’s criterion saying that a given sequence is equidistributed (mod 1) if
and only if certain associated exponential sums converge to zero.
A particularly interesting example for our future purposes is the se-
quence {nα } where 0 < α < 1. A preliminary simple result about this
sequence is the fact that it is equidistributed:
Exercise 3.1.1. Show that {nα } is equidistributed on S 1 for each 0 < α <
1. Hint: By making a suitable re-interpretation of the problem on the real
line R, use the fact that (n + 1)α − nα → 0 and nα → ∞ when n → ∞.

75
76 Chapter 3. Elkies-McMullen theorem

A quite popular fashion of studying the distribution of a sequence xn

consists into the analysis of the gaps J1 , . . . , JN determined by these points
on S 1 , i.e., J (N ) := {J1 , . . . , JN } are the connected components of S 1 −
{x1 , . . . , xN }. Observe that the sum of the lengths |Ji | of the gaps Ji
equals to 1, so that the average length of these gaps is:
N
1 X
|Ji | = 1/N.
N i=1

In other words, the “natural scale” for the study of the lengths |Ji | of
the gaps is 1/N . Taking this “natural scale” into account, we introduce
the following definition:

Definition 3.1.1. We say that a sequence xn is exponentially distributed

whenever, for any 0 ≤ a ≤ b, we have
Z b
1
#{J ∈ J (N ) : |J| ∈ [a/N, b/N ]} → e−t dt
N a

when N → ∞.

Example 3.1.1. A random choice of points on the circle gives an exponen-

tially distributed sequence (for further details see the page 158 of Feller’s
book [5]).

Coming back to the sequence {nα } with 0 < α < 1, some numerical
experiments suggest that it is exponentially distributed for every α 6= 1/2.
Furthermore, M. Boshernitzan observed (numerically) in 1993 a special
distribution for the specific case α = 1/2. However, a rigorous confirmation
of this numerical observation of Boshernitzan was only recently obtained
by N. Elkies and C. McMullen.
More precisely, N. Elkies and C. McMullen [6] showed √ the following
theorem about the distribution of the gaps of the sequence { n} (mod 1):

Theorem √ 3.1.1 (Elkies and McMullen (2004)). The distribution of the

gaps of { n} is given by the continuous function

 2
6/π , t ∈ [0, 1/2],
F (t) := F2 (t), t ∈ [1/2, 2],


F3 (t), t ∈ [2, ∞),

where F2 (t) and F3 (t) are explicit real-analytic functions. Moreover, F (t)
isn’t analytic (and it isn’t even C 3 ) at the points t = 1/2 and t = 2. In
other words, F (t) exhibits a genuine phase transition at the points t = 1/2
and t = 2.
C. Matheus 77

For the reader’s convenience, we recall that G(t) is the distribution of

the gaps of a given sequence xn whenever, for any 0 ≤ a ≤ b, we have
Z b
1
#{J ∈ J (N ) : |J| ∈ [a/N, b/N ]} → G(t)dt when N → ∞.
N a

In the statement of Elkies and McMullen theorem, we avoided the intro-

duction of the explicit expressions of the functions F2 (t) and F3 (t) in order
to get a short statement. However, it is not hard to write the formulas of
these functions. To do so, we denote by r := 1/2x and define

6 2 3 1
F2 (x) = ( (4r − 1) 2 ψ(r) + (1 − 6r) log r + 2r − 1) if ≤ x ≤ 2,
π2 3 2

6
F3 (x) = (f (α) − g(α) − h(α)) if x ≥ 2.
π2
√ √ √
Here ψ(r) = tan−1 [(2r−1)/ 4r − 1]−tan−1 [1/ 4r − 1], 2α = 1− 1 − 4r,
f (α) = 4(1 − 4α)(1 − α)2 log(1 − α), g(α) = 2(1 − 2α)3 log(1 − 2α) and
h(α) = 2α2 .
In particular, we see that F (t) is continuous at t = 1/2 and t = 2 (with
values F (1/2) = 6/π 2 and F (2) = 6(log 2 − 1/2)/π 2 ). Also, F (t) is C 1
but F (t) is not C 2 at t = 2 and F (t) is not C 3 at t = 1/2 (as a simple
argument with Taylor series around these points shows). Another direct
consequence of this√explicit formula of F (t) is the fact that the “tail” of
the distribution of n (mod 1) is not exponential: F (t) ∼ 3t−3 /π 2 when
t → ∞. Comparing this fact with the example 3.1.1, we √ see that the
appearance of large gaps is more frequent for the sequence n (mod 1)
than it is for a random sequence of points.
This being said, we will concentrate our efforts on the discussion of the
beautiful proof of Elkies and McMullen theorem. Since the proof of this
result involves a number of technical steps, we will describe in the next
paragraph a very rough sketch of the main arguments, leaving the details
to the subsequent sections.
Roughly speaking, Elkies and McMullen idea consists into a translation
of √the problem of the computation of the distribution F (t) of the gaps
of n (mod 1) to the calculation of the probability of a random affine
lattice of R2 to intersect a certain fixed triangle. The advantage of this
apparently artificial approach resides in the fact that the Ergodic Theory
of random (affine) lattices is well-understood due to the so-called Ratner
theory, which allows us to compute precisely the desired probability (of a
random lattice to meet a fixed triangle).
At this point, we are ready to discuss this scheme with a little bit more
of details.
78 Chapter 3. Elkies-McMullen theorem

3.2 Ergodic version of Elkies-McMullen the-

orem
√
Before starting the translation of the calculation of the distribution of n
(mod 1) to an ergodic-theoretical problem, let us introduce some notation.
We remember that Λ0 ⊂ R2 is a lattice if Λ0 is a discrete subgroup iso-
morphic to Z2 . We say that a lattice Λ0 is unimodular if the torus R2 /Λ0
has unit area. Furthermore, an affine lattice Λ ⊂ R2 is a subset of the
following form: Λ = Λ0 + v where v ∈ R2 and Λ0 is a lattice.
We denote by E the space of unimodular affine lattices. As the reader
can easily check, E is naturally identified with the space

E = ASL(2, R)/ASL(2, Z),

where ASL(2, R) is the group of affine transformations g : R2 → R2 of
the form g(v) = Av + b with det A = 1 and ASL(2, Z) is the discrete
subgroup of ASL(2, R) whose elements g(v) = Av + b verify A ∈ SL(2, Z)
and b ∈ Z2 .
An immediate consequence of this identification is the fact that E comes
equipped with an unique probability measure µE which is invariant by the
left action of ASL(2, R) (namely, µE is the so-called Haar measure of E).
In particular, µE gives a sense to the notion of a random lattice of E: a
random lattice is a lattice with generic properties with respect to µE , i.e.,
a lattice belonging to a µE -full measure subset of µE .
Keeping these notations in mind, we are ready to give an informal
scheme of the proof of Elkies and McMullen theorem.

3.2.1 Scheme of the proof of Elkies and McMullen the-

orem
Let us fix N a large integer, t > 0 and I = [x, x + t/N ] ⊂ R/Z, where
x ∈ [0, 1] is randomly choosen (with respect to Lebesgue measure). In
these terms, the Elkies and McMullen program is:
√
1. computing the gap distribution of n (mod√1) is equivalent to cal-
culate the probability PN (t) of some point { n} with 0 ≤ n ≤ N to
belong to I;
√ √
2. on the other hand, { n} ∈ I ⇐⇒ n ∈ I + a for some a ∈ Z ⇐⇒
n ∈ (I + a)2 for some a ∈ Z;
3. now,
√ for the purpose of the computation of the gap distribution of
{ n}, it is possible to show that one can harmlessly replace (I + a)2
by its linear approximation
(I + a)2 ∼ (a + x)2 + 2(a + x)(I − x) = a2 − x2 + 2(a + x)I;
C. Matheus 79

4. also, one can show that we can assume that N is a square (without
loss of generality);
5. under these assumptions, we look at the linear approximation a2 −
x2 + 2(a + x)I of (I + a)2 and note that
n ∈ a2 − x2 + 2(a + x)I (for 0 ≤ n ≤ N )
m
2
√
(Z + x ) ∩ 2(a + x)I 6= ∅ (for 0 ≤ a + x ≤ N );
6. this last condition can be rewritten as T ∩ Z2 6= ∅ where T ⊂ R2 is
the triangle of area t given by
√
T := {(a, b) : b + x2 ∈ 2(a + x)I and a + x ∈ [0, N ]};

7. denoting by St the “standard” triangle of area t with vertices (0, 0),

(1, 0), (0, 2t) and considering g ∈ ASL(2, R) the unique afinne trans-
formation with g(T ) = St and g(−x, −x2 ) = (0, 0), we can rephrase
the above condition as
g(Z2 ) ∩ St 6= ∅;

8. in resume, this (sketchy) discussion allows us to translate

√ the com-
putation of the probability PN (t) of some element { n} with 0 ≤
n ≤ N to belong to I = [x, x + t/N ] into the problem of calculating
the probability of the affine lattice ΛN (x) := g(Z2 ) to intersect the
standard triangle St ;
9. on the other hand, for N ≫ 1 large, we expect that ΛN (x) ∈ E
behaves like a random affine lattice: in fact, we will see that, as
any good random affine lattice, the sequence ΛN (x) is uniformly
distributed - for every compactly supported continuous function f of
E it holds
Z 1 Z
f (ΛN (x))dx → f (Λ)dµE (Λ) when N → ∞;
0 E

10. using this uniform distribution result, we get that PN (t) → p(t) when
N → ∞, where p(t) is the probability of a random affine lattice to
intersect the standard triangle St ;
11. finally, by reversing this translation, one can√show that p′′ (t) =
−F (t), where F (t) is the gap distribution of{ n}, which ends the
proof of Elkies and McMullen theorem because p′′ (t) can be explicitly
calculated from natural expressions for the Haar probability µE .
After this informal description of Elkies and McMullen arguments, we
will start a more or less detailed discussion of all of these topics.
80 Chapter 3. Elkies-McMullen theorem

3.2.2 Some preliminary reductions

For each N ≥ 1 integer, we define
√ a function λN : [0, ∞) → [0, 1] as follows.
We consider the N points { n}, n = 1, . . . , N , of the circle R/Z. This
gives
√ us a partition of the circle into N intervals J1 , . . . , JN (among them,
⌊ N ⌋ − 1 have length zero). In this setting,
1
λN (x) := #{1 ≤ i ≤ N : |Ji | < x/N }.
N
We left the verification of the following elementary properties of λN (x)
as an exercise to the reader:
Exercise 3.2.1. Show that λN (x) has the following properties:
• λN is a non-decreasing left-continuous function (indeed, λN is con-
stant except for a finite number of jump discontinuities); moreover,
λN (0) = 0 and λN (∞) = 1;
R∞ R∞ R∞
• 0 (1 − λN (x))dx = 0 xdλN (x) = 1 (because 0 xdλN (x) is the
sum of the lengths of the gaps).
In these terms, Elkies and McMullen theorem is equivalent to:
Theorem 3.2.1. There exists a continuous function λ∞ : [0, ∞) → [0, 1]
such that λN → λ∞ uniformly in compact sets when N → ∞. Moreover,
Z x
λ∞ (x) = F (ξ)dξ,
0

where F is an explicit function to be computed later.

√
Observe that F is the desired gap distribution
√ of { n}: in fact, for all
0 ≤ x1 < x2 < ∞, the quantity of gaps ofR { n} (1 ≤ n ≤ N ) with length
x
between x1 /N and x2 /N is asymtotic to x12 F (ξ)dξ.
In order to study λN , we introduce LN : R/Z → [0, ∞) defined by
( √
0, if t = { n} for some 0 ≤ n ≤ N
LN (t) =
N × length of the gap containing t, otherwise.

Using LN we can write the union of the gaps of length < x/N as

IN (x) := {t ∈ R/Z : LN (t) < x}.

Rx
In particular, |IN (x)| = 0 ξdλN (ξ), so that the theorem 3.2.1 is equiv-
alent to:
Rx
Theorem 3.2.2. |IN (x)| → 0 ξF (ξ)dξ uniformly on x varying on com-
pact sets when N → ∞.
C. Matheus 81

Now let us talk about the preliminary reductions of these theorems. The
first simplification of the statement of Theorem 3.2.1 was annouced at the
fourth item of Elkies and McMullen scheme discussed above: namely, in
the statement of this theorem, we can assume that N is a square, i.e.,
N = s2 with s ∈ Z. More precisely, we have the following result:

Lemma 3.2.1 (Lemma 3.1 of Elkies and McMullen). Suppose that λs2
converges (uniformly on compact sets) to a continuous function λ∞ when
s → ∞. Then, λN also converges (uniformly on compact sets) to λ∞ when
N → ∞.

Proof.√ Observe that any integer N is far from some square s2 by a distance
2 2
√ O( N ). On the other hand, by replacing N by s , we change 3|N −s | .
of
N of the lengths of the gaps (at
p most) and we multiply the normalizing
factor 1/N by N/s2 = 1 + O( 1/N ). In particular, the desired lemma
follows.

The second simplification was described at the second and third items of
Elkies and McMullen program: during the study of λN (x) we can replace
the quadratic expressions by its linear approximations without changing
the asymptotic results. In order to formalize this fact, we will need more
2
notation. Recall
√ that
√ each√integer n can be uniquely written as a + b
where a = ⌊ n⌋ = n − { n}. Using Lemma 3.2.1, we can assume that
N is a perfect square, say N = s2 . In this situation, we see that

LN (t) = N (t2 − t1 )

where t2 is the smallest real number ≥ t with (a2 + t2 )2 ∈ Z for some

integer a2 < s and t1 is the biggest real number ≤ t with (a1 + t1 )2 ∈ Z for
some integer a1 < s. In order to understand the properties of the function
LN we make the following arithmetical remark:

Remark 3.2.1. For aj ∈ Z (j = 1, 2), we have (aj +tj )2 = a2j +2aj tj +t2j ∈
Z if and only if bj := 2aj tj + t2j ∈ Z. Furthermore, we have 0 ≤ bj ≤
(aj + 1)2 − a2j = 2aj + 1 when 0 ≤ tj ≤ 1.

Using this remark, we can rewrite the identity LN (t) = N (t2 − t1 ) as

 
LN (t) = N ((t2 − t) − (t1 − t)) = N min rt (a, b) − max rt (a, b)
rt (a,b)≥0 rt (a,b)≤0

where a, b vary over the set of integers satisfying

0 < a < s and 0 ≤ b ≤ 2a + 1

√
and rt (a, b) := a2 + b − a − t.
82 Chapter 3. Elkies-McMullen theorem

Remark 3.2.2. In fact, this last condition on b is superfluous: on one

hand, 0 ≤ b ≤ 2a + 1 is equivalent to 0 ≤ rt (a, b) + t ≤ 1 and, on the other
hand, min rt (a, b) + t and max rt (a, b) belong to the interval [0, 1]
rt (a,b)≥0 rt (a,b)≤0
since rt (1, 0) + t = 0 and rt (1, 3) + t = 1.

Going back to the analysis of LN , we will apply the idea discussed in the
third item of the Elkies and McMullen program: in the definition of bj (see
remark 3.2.1), we replaced t2j by its linear approximation t2 + 2t(tj − t) =
2tj t − t2 nearby t. Consequently, bj is substituted by 2aj tj + (2tj t − t2 ) =
2(a + t)tj − t2 . By this reason, we will consider τj = (bj + t2 )/2(a + t) the
solution of the equation

2(a + t)τj − t2 = bj

and we will replace tj by τj in the definition of LN , so that we get the

function
 
e
LN (t) := N min rt (a, b) − max rt (a, b)
ρt (a,b)≥0 ρt (a,b)≤0

b+t 2 a2 +b−(a+t)2
where ρt (a, b) := 2(a+t) −t = 2(a+t) and a, b vary over the set of
integers satisfying

0 < a < s and 0 ≤ b ≤ 2a + 1.

Remark 3.2.3. Similarly to Remark 3.2.2, this last condition on b is

superfluous.

Analogously to the definition of the set IN (x), we introduce

IeN (x) := {t ∈ R/Z : L

e N (t) < x}.

As we told in the third item of the Elkies and McMullen program, re-
placing t2j by its linear approximation (or equivalently the replacement of
tj by τj ) in the definition of LN doesn’t affect the asymptotics. More
precisely, we have the following theorem:

Theorem 3.2.3 (Proposition 3.2 of Elkies and McMullen). Suppose that

Rx
|IeN (x)| converges (uniformly on compact sets) to 0 ξF (ξ)dξ with F con-
tinuous. Then, the same happens to IN :
Z x
|IN (x)| → ξF (ξ)dξ
0

when N → ∞ (uniformly on compact sets).

C. Matheus 83

A “conceptual” explanation for the validity of this theorem goes as

follows: typically, we expect that tj = t + O(1/N ), so that |tj − τj | =
O(1/aj N 2 ) (since τj is the solution of the equation 2(aj + t)τj − t2 = bj =
2aj tj + t2j ). Moreover, given ε > 0, we get 1/aj < ε for the “majority” of
the pairs (aj , bj ). Hence, the natural expectations is the replacement of tj
by τj changes the lengths of most of the gaps by O(ε/N 2 ), which certainly
doesn’t affect the asymptotic behavior of |IN |.
In order to formalize the previous heuristic, we consider the quocient
√  
ρt (a, b) a2 + b + a + t 2a + 1 2a + 1
= ∈ , .
rt (a, b) 2(a + t) 2a + 2 2a
Manipulating this information (see Lemma 3.3 of Elkies and McMullen
paper), it follows that
e N (t) ≤
Lemma 3.2.2. For all t ∈ [0, 1], it holds 34 LN (t) ≤ L 3
2 LN (t).
Furthermore, for all A ∈ N, we have the estimate
2A + 1 e 2A + 1
LN (t) ≤ LN (t)
2A + 2 2A
for all t ∈ [0, 1] except for a subset of size ≤ (A + 2)(A − 1)/(s − 1).
As the reader can check, this lemma easily implies the following esti-
mates
|IeN (3x/4)| ≤ |IN (x)| ≤ |IeN (3x/2)|
and
2A + 1 2A + 1
|IeN ( x)| − O(A2 /s) ≤ |IN (x)| ≤ |IeN ( x)| + O(A2 /s)
2A + 2 2A
for all x ∈ [0, ∞) and A = 1, 2, . . . . Using these estimates with A =
1 + ⌊s1/3 ⌋ (or any other increasing function of s with A2 /s → 0), we get
the proof of Theorem 3.2.3.
Once we have the theorem 3.2.3 in our toolbox, our goal is reduced to
the asymptotic study of Lf N . In this direction, we are going to interpret (in
the next subsection) this quantity in terms of the Ergodic Theory of affine
lattices (as proposed in the items 5, 6, 7 and 8 of Elkies and McMullen
scheme.

3.2.3 eN
Geometrical interpretation of L
As we already told in the item 6 of the program, we are going to use a
convenient triangle T (whose specific form was described above) in our
arguments. In our current notation, T is the triangle of the (a, b)-plane
whose interior is determined by the inequalities
0 < a + t < s, 2c− (a + t)/s2 < b − 2ta − t2 < 2c+ (a + t)/s2 ,
84 Chapter 3. Elkies-McMullen theorem

for c− < 0 < c+ . Making a direct translation between our notations, the
e N is interpreted in terms of T as described
reader can easily check that L
in the following lemma:
Lemma 3.2.3. For each N = s2 and t ∈ [0, 1], we have the following
possibilities:
• if Le N (t) 6= 0, then L
e N (t) is the area c+ − c− of the biggest triangle
T as above whose interior doesn’t intersect Z2 − {(0, 0), (0, 1)};
e N (t) = 0, then all triangles T as above contains the point (a, b)
• if L
whose coordinates satisfy b − 2ta − t2 = 0 with 0 < a < s.
Now, we will “clean” the statement of this lemma by observing that the
e N (t) except for t ∈ [0, 1] in
possibilities (a, b) = (0, 0), (0, 1) don’t affect L
a subset of size O(1/s):
Lemma 3.2.4. The characterization of the quantity L e N (t) in lemma 3.2.3
isn’t affected by the inclusion of the extra cases (a, b) = (0, 0), (0, 1) except
when 0 ≤ t < 1/(s − 1) or t−1 − t < 1/(s − 1).
The proof of this lemma is a simple case-by-case verification (for further
details, see Lemma 3.7 of Elkies and McMullen). Using this lemma, it
follows that the inclusion of the extra cases (a, b) = (0, 0), (0, 1) doesn’t
affect the asymptotic of |IeN (t)| (because this minor modification alters the
values of Le N (t) only on a subset of size O(1/s)).
Now we will apply the idea discussed in the item 7 of the Elkies and
McMullen scheme: consider the affine transformation g of R2 defined by
g(a, b) = (w1 , w2 ) = (s(b − 2ta − t2 ), (a + t)/s).
Observe that g sends the vertex (−t, −t2 ) to the origin (0, 0), the triangle
T into the “standard” triangle
∆c− ,c+ := {(w1 , w2 ) ∈ R2 : 0 < w2 < 1, 2c− w2 < w1 < 2c+ w2 }
and the lattice Z2 to the lattice Λs2 (t) = g(Z2 ) given by
Λs2 (t) := {(s(b − 2ta − t2 ), (a + t)/s) : (a, b) ∈ Z2 )}.
Note that the “standard” triangle ∆c− ,c+ depends on c− , c+ but it doesn’t
depend on s, t (for the effect of comparison, this standard triangle corre-
sponds to the triangle St of item 7 of Elkies and McMullen program).
Closing our series of translations, we introduce the following definition:
Definition 3.2.1. Given an affine lattice Λ in the (w1 , w2 )-plane, we de-
note by L(Λ) the area c+ − c− of the biggest triangle of the form ∆c− ,c+
which is disjoint from Λ with the conventions L(Λ) = 0 when such a tri-
angle doesn’t exist and L(Λ) = ∞ when all of these triangles are disjoint
from Λ.
C. Matheus 85

Using this notation, we can apply Lemma 3.2.4 to resume the discussion
of this subsection into the following proposition:

Proposition 3.2.1 (Proposition 3.8 of Elkies and McMullen). For all

s ∈ N, x ∈ R, the set of t ∈ [0, 1] with L(Λs2 (t)) ≤ x has size

|Ies2 (x)| + O(1/s).

In other words, Proposition 3.2.1 says that the issue of studying the
asymptotic of Ies2 is reduced to the study of the behavior of the function
L on the family of affine lattices Λs2 (t).
At this point, it remains “only” to complete the details of the items 9,
10, 11 of the program. This will be performed in the next two subsections.
Morally, these items essentially say that the study of L on the affine lattices
Λs2 (t) can be done through the Ergodic Theory of random affine lattices
(in particular, Ratner’s theorems will be useful during this task).

3.3 Study of L via Ratner theorems

In the sequel, we will use Ratner’s theorems (about the Ergodic Theory of
homogeneous flows) in order to understand the values of L along the family
of (affine) lattices Λs2 (t). Namely, we will invoke the following result saying
that the family {Λs2 (t) : t ∈ [0, 1]} of “circles” of affine lattices become
equidistributed on the space of affine lattices E when s → ∞:

Theorem 3.3.1. For any f ∈ C0 (E) we have

Z 1 Z
f (Λs2 (t))dt → f dµE when s → ∞.
0 E

For the discussion of this subsection, we will assume this equidistribution

theorem and we will see√ how it helps in the determination of the asymptotic
gap distribution F of { n}.

3.3.1 Computation of F assuming Theorem 3.3.1

We recall that Proposition 3.2.1 proved in the previous section says that the
size of the subset of t ∈ [0, 1] such that L(Λs2 (t)) ≤ x is |Ies2 (x)| + O(1/s).
Putting this information together with the theorem 3.3.1 above, we have
the following consequence:

Proposition 3.3.1. For x ∈ [0, ∞) we have

|Ies2 (x)| → µE ({Λ ∈ E : L(Λ) ≤ x} when s → ∞.

86 Chapter 3. Elkies-McMullen theorem

Proof. Consider Ex := {Λ ∈ E : L(Λ) ≤ x}. Using this notation, the fact

that the size of the subset of t ∈ [0, 1] with L(Λs2 (t)) ≤ x is |Ies2 (x)| +
O(1/s) can be rewritten as:
Z 1
χEx (Λs2 (t))dt = |Ies2 (x)| + O(1/s)
0
R1
This reduces our task to show that 0 χEx (Λs2 (t))dt converges to µE (Ex ).
To do so, the basic idea is to use the theorem 3.3.1. However, it is not
possible to make a direct application of this result because the character-
istic function χEx is not continuous. A simple remedy to this difficulty
is the classical L1 -approximation argument of χEx and 1 − χEx by some
continuous function in the space C0 (E) combined with the theorem 3.3.1.
At this stage, it remains only to know whether such compactly supported
continuous L1 -approximations exist. From the basic result of Real Analy-
sis, we know that the functions χEx and χE−Ex can be approximated by
some functions in C0 (E) whenever µE (∂Ex ) = 0.
In resume, the proof of this proposition is complete once we can show
that µE (∂Ex ) = 0. In this direction, we invite the reader to prove that
L : E → [0, ∞] is a submersion for almost all points of E: more precisely,
L isn’t a submersion only at the affine lattices Λ containing the origin
(0, 0) or some point of the horizontal side w2 = 1 of its maximal triangle
∆c− ,c+ . In particular, for each x, the corresponding points Ex where L
isn’t a submersion is a closed subset of µE -measure zero. Thus, by the
inverse theorem, we see that the level sets of L have zero µE -measure and,
a fortiori, it follows that µE (∂Ex ) = 0. This ends the proof.

A direct corollary of this proposition (and the results of the previous

section) is:

Proposition
√ 3.3.2. Suppose that the asymptotic gap distribution F (ξ) of
{ n} is continuous. Then,

lim |IN (x)| = lim |IeN (x)| = µE ({Λ ∈ E : L(Λ) ≤ x})

N →∞ N →∞

for x ∈ [0, ∞). Moreover, this convergence is uniform on compact sets

(with respect to x).

Proof. Assuming that F is continuous, we can combine Lemma 3.2.1, The-

orem 3.2.3 and Proposition 3.3.1 to get the desired result.

Although the statement of Proposition 3.3.2 seems quite promising (since

we wrote IN asymptotically in terms of the measure µE of the subset
L−1 ([0, x])), we are still not ready to compute the gap distribution F : in
fact, the discussion of the previous section tells us only that IN (x) →
C. Matheus 87

Rx
0
ξF (ξ)dξ, so that in order to get a concrete formula for F in terms of
µE (L−1 ([0, x])) we should extract the two derivatives (with respect to the
x variable) of this function. However, at the present moment, it is not
clear even that µE (L−1 ([0, x])) is differentiable!
Hence, we should analyze more carefully the subsets L−1 ([0, x]). Keep-
ing this goal in mind, we introduce the subset Sc− ,c+ of E formed by
the affine lattices Λ with some point inside the triangle ∆c− ,c+ , where
c− < 0 < c+ . Observe that µE (Sc− ,c+ ) depends only on the area c+ −c− of
the triangle ∆c− ,c+ because any two triangles with the same area are equiv-
alent under some element of ASL2 (R) and the measure µE is ASL2 (R)-
invariant. In particular, we can define a function p : [0, ∞] → [0, 1] by

p(c+ − c− ) := µE (Sc− ,c+ )

with the conventions that p(0) = 0 and p(∞) = ∞.

As we already mentioned, we will eventually take two derivatives of p in
order to find an explicit formula for F :
Lemma 3.3.1. Suppose that p ∈ C 2 (i.e., p is twice differentiable and p′′
is continuous). Then,
F (x) = −p′′ (x).
Proof. We write µE ({Λ ∈ E : L(Λ) < x}) as a telescopic sum:
M
X −1
µE (S0,x ) − lim [µE (S (j+1)x −x, jx ) − µE (S jx −x, jx )].
M →∞ M M M M
j=0

Putting this equation in terms of p, we get

 x 
µE ({Λ ∈ E : L(Λ) < x}) = p(x) − lim M p(x) − p(x − ) .
M →∞ M
Since p is differentiable, it follows that

µE ({Λ ∈ E : L(Λ) < x}) = p(x) − xp′ (x).

On the other hand, assuming that p is twice differentiable, we know that

d ′ ′′
dx (p(x) − xp (x)) = −xp (x). Moreover, since p(0) = 0, we see that
′
p(x) − xpR(x) = 0 on x = 0. Combining these two facts, we get p(x) −
x
xp′ (x) = 0 −ξp′′ (ξ)dξ.
Using these two identities, we obtain
Z x
µE ({Λ ∈ E : L(Λ) < x}) = −ξp′′ (ξ)dξ.
0

This completes the proofR xof the lemma in view of Proposition 3.3.2 and
the fact that |IN (x)| → 0 ξF (ξ)dξ.
88 Chapter 3. Elkies-McMullen theorem

Remark 3.3.1. Still assuming that p ∈ C 2 , from the definition of p and

Lemma 3.3.1, we see that

∂2
F (x) = −p′′ (x) = − µE (Sc− ,c+ )
∂c− ∂c+

for any c− < 0 < c+ with c+ − c− = x. This gives the following geomet-
rical interpretation of F (x) in terms of µE : the value F (c+ − c− )dc− dc+
is the measure of the subset of affine lattices Λ ∈ E intersecting ∆c− ,c+
into exactly two points - one of them with coordinates (w1 , w2 ) verifying
w1 /2w2 ∈ (c− , c− + dc− ) and the other one with coordinates (w1 , w2 ) ver-
ifying w1 /2w2 ∈ (c+ − dc+ , c+ ).
√
From Lemma 3.3.1, the calculation of the gap distribution of F of { n}
is reduced to the explicit computation of the function p and the verification
of the fact p ∈ C 2 . In this direction, we will recall some known facts about
the theory of unimodular lattices.
We denote by B the space of unimodular lattices of R2 (i.e., discrete
subgroups Λ0 isomorphic to Z2 with covolume 1) and µB the Haar prob-
ability measure of B. A vector w ∈ Λ0 of a lattice Λ0 ∈ B is called
primitive whenever one can find w′ ∈ Λ0 such that {w, w′ } is a Z2 -basis
of Λ0 . Equivalently, w ∈ Λ0 is primitive when w/k ∈ / Λ0 for any k > 1. In
the sequel, we are going to use the following facts:

• the subset Zw ⊂ B of lattices with w as a primitive vector is a circle

(in fact it is a closed horocycle);

• Rgiven K ⊂ R2 a convex compact subset, the area of K is ζ(2) ×

B K
f (Λ0 )dµB where fK (Λ0 ) is the quantity of primitive vectors of
0
Λ inside K;

• in particular, taking K sufficiently small so that fK (Λ0 ) ≤ 1 for all

Λ0 ∈ B, we see that the subset of lattices with a primitive vector
inside K has µB -measure equal to 1/ζ(2) times the area of K;

• moreover, we can desintegrate the µBR-measure of a measurable sub-

e ⊂ B as follows: µB (B)
e = 1 e
set B ζ(2) w∈K µw (B ∩ Zw ), where µw is
the (normalized) Lebesgue measure on the circle Zw .

At this point, our objective is to combine the remark 3.3.1 with the desin-
tegration of µB in order to express F as a double integral. In this direction,
in view of the geometrical interpretation of F (see remark 3.3.1), we look
at the lattices Λ ∈ E intersecting the triangle ∆c− ,c+ into two points whose
(i) (i) (1) (1)
coordinates (w1 , w2 ) (i = 1, 2) verify w1 /2w2 ∈ (c− , c− + dc− ) and
(2) (2)
w1 /2w2 ∈ (c+ − dc+ , c+ ). Note that the difference between these two
points is a primitive vector of Λ: otherwise, Λ would contain a third point
C. Matheus 89

inside the line segment determined by these two points; since ∆c− ,c+ is
convex (since it is a triangle) it would follow that Λ intersects ∆c− ,c+ into
three points, a contradiction with our hypothesis. Using this primitive
vector, we apply the desintegration of µB to write F as an integral on
the w2 -coordinates v− , v+ of the vectors of Λ at the boundary of ∆c− ,c+ :
for v− , v+ ∈ (0, 1), we write w = (2c+ v+ , v+ ) − (2c− v− , v− ) and we recall
that Zw parametrize the subset of lattices containing w; next, we denote
by qx (v− , v+ ) ∈ [0, 1] the (µw )-measure of the subset of Zw formed by the
lattices do not touching the interior of ∆c− ,c+ . Observe that we write qx
instead of qc− ,c+ because this quantity depends only on x = c+ − c− . In
this notation, we can express F as a double integral:

Proposition 3.3.3. The function (x, v− , v+ ) 7→ qx (v− , v+ ) is continuous

except at a certain subset of {v− = v+ }. Moreover, for x ∈ [0, ∞), we
have
Z 1Z 1
1
−p′′ (x) = F (x) = 4v− v+ qx (v− , v+ )dv− dv+ .
ζ(2) 0 0

In particular, it follows that F is continuous.

Proof. The fact that qx (v− , v+ ) is continuous is immediate except when

the vector w is horizontal (in particular, it is parallel to the third side of
∆c− ,c+ ). This shows the first part of this proposition since w is horizontal
implies v− = v+ . Next, since 0 ≤ qx (v− , v+ ) ≤ 1, the double integral
above exists and it varies continuously on the x-variable. Finally, in order
to see that this integral coincides with −p′′ (x) and F (x), we use the ge-
ometrical interpretation of F (see the paragraph before the statement of
this proposition) combined with the fact that 4v− v+ is the product of the
lengths of the line segments

{(w1 , v− ) : 2c− v− < w1 < 2(c− + dc− )v− }

and
{(w1 , v+ ) : 2(c+ − dc+ )v+ < w1 < 2c+ v+ }

where the vectors of the lattices reside, and the desintegration formula of
µB .

In order to make Proposition 3.3.3 more useful, we need to compute

qx (v− , v+ ). The idea is to do a series of geometrical considerations after
a affine change of variables from (w1 , w2 ) to (z, z ′ ) sending the triangle
∆c− ,c+ on the isosceles triangle

∆0 := {(z, z ′ ) ∈ R2 : z, z ′ > 0, z + z ′ < 1}

90 Chapter 3. Elkies-McMullen theorem

of area 1/2. Since the triangle ∆c− ,c+ has area c+ − c− , this affine trans-
formation multiplies the area by the factor
r := 1/2x.
Although the argument is not complicated, we will refer the curious reader
to Lemma 3.12 of Elkies and McMullen for a detailed proof of the following
fact:
Lemma 3.3.2. For any 0 < v, v ′ ≤ 1 and x > 0 it holds qx (v, v ′ ) =
qx (v ′ , v). Moreover, for v ≥ v ′ , we have
  r   
′ v(1 − v ′ ) − r
qx (v, v ) = max 0, min 1, ′ − max 0,
vv v(v − v ′ )
with r = 1/2x. Here we are using the following convention
  (
v(1 − v ′ ) − r ∞ if v = v ′ and r < v(1 − v ′ )
max 0, =
v(v − v ′ ) 0 if v = v ′ and r ≥ v(1 − v ′ )

Once this fact is available, the task√of finding an explicit formula for F
(the asymptotic gap distribution of { n}) becomes a Calculus I exercise.
Indeed, combining Proposition 3.3.3 with Lemma 3.3.2 and computing
some integrals (as in the proof of Theorem 3.14 of Elkies and McMullen
paper), the reader will eventually prove the following result:
Theorem 3.3.2. It holds

 2
6/π , t ∈ [0, 1/2],
F (t) = F2 (t), t ∈ [1/2, 2],


F3 (t), t ∈ [2, ∞),

where F2 (t) and F3 (t) are

6 2 3
F2 (x) = ( (4r − 1) 2 ψ(r) + (1 − 6r) log r + 2r − 1)
π2 3
and
6
F3 (x) = (f (α) − g(α) − h(α)).
π2
√ √
Here r :=√1/2x and ψ(r) = tan−1 [(2r − 1)/ 4r − 1] − tan−1 [1/ 4r − 1],
α = (1 − 1 − 4r)/2, f (α) = 4(1 − 4α)(1 − α)2 log(1 − α), g(α) = 2(1 −
2α)3 log(1 − 2α) and h(α) = 2α2 .
In other words, we completed the proof of Elkies and McMullen The-
orem 3.1.1 modulo Theorem 3.3.1 (which we were assuming during the
entire subsection)!
Now, we end this subsection and we pass to the study of the relation
between Theorem 3.3.1 with the Ergodic Theory of homogenous flows.
C. Matheus 91

3.3.2 Theorem 3.3.1 and homogenous flows

We recall that Theorem 3.3.1 concerns the equidistribution of the family
of circles of affine lattices {Λs2 (t) : t ∈ [0, 1]} when s → ∞. In order
to reformulate this theorem into a more appropriate language, we observe
that the entire action takes places at the special affine group ASL2 (R)
which we will denote by
  
 a b x 
G(R) :=  c d y  : ad − bc = 1 ⊂ SL3 (R).
 
0 0 1

Note that this group acts on R2 via the conservative affine transformations
      
X a b X x
7→ + .
Y c d Y y

Denote by G(Z) ⊂ G(R) the subgroup of matrices with integer entries

and observe that the space of unimodular affine lattices E is naturally
identified with G(R)/G(Z): we take the integral lattice Z2 as a base point
and for each g ∈ G(R) we associate the affine lattice
    
 w1 Z 
Λ(g) := (w1 , w2 ) ∈ R2 :  w2  ∈ g  Z  .
 
1 1

This map is surjective and Λ(g) = Λ(h) if and only if h ∈ g · G(Z) (as the
reader can easily check), so that this map is an isomorphism between E
and G(R)/G(Z).
In the particular case of the affine lattices Λs2 (t), the corresponding
elements of G(R)/G(Z) under this isomorphism can be explicitly calculated
as follows: recall that

Λs2 (t) := {(s(b − 2ta − t2 ), (a + t)/s)},

so that the points (w1 , w2 ) ∈ Λs2 (t) in matricial notation are:

      
w1 s −2st −st2 b b
 w2  =  0 1/s t/s   a  = As U (t)  a  ,
1 0 0 1 1 1

where As = diag(s, 1/s, 1) is the diagonal matrix

 
s 0 0
As :=  0 1/s 0 
0 0 1
92 Chapter 3. Elkies-McMullen theorem

and  
1 −2t −t2
U (t) :=  0 1 t .
0 0 1

Thus, Λs2 (t) is the affine lattice

    
 w1 Z 
Λs2 (t) = (w1 , w2 ) ∈ R2 :  w2  ∈ As U (t)  Z  .
 
1 1

In other words, Λs2 (t) is identified with As U (t) via the isomorphism Λ. In
resume, we see that Theorem 3.3.1 is equivalent to:

Theorem 3.3.3. The circles {As U (t) : t ∈ [0, 1]} become equidistributed
in G(R)/G(Z) when s → ∞, i.e., for all f ∈ C0 (E), it holds
Z 1 Z
lim f (As U (t))dt = f dµE .
s→∞ 0 E

Once we translated Theorem 3.3.1 to Theorem 3.3.3, our plan is to

use the Ergodic Theory of the homogenous flow As on the space E =
G(R)/G(Z): more precisely, we will exploit the fact that the circle {U (t) :
t ∈ [0, 1]} is a non-linear horocycle (a concept to be introduced later) to
derive Theorem 3.3.3 from a more general result about the equidistribution
of the non-linear horocycles of the geodesic flow As .
A detailed explanation of this plan is the content of the next section.

3.4 Equidistribution of non-linear horocycles

During this section, we will discuss the Ergodic Theory of the homogenous
flow As in the space of affine lattices G(R)/G(Z). To do so, we recall some
definitions of the last subsection. As we saw in the last section, the special
affine group ASL2 (R) is naturally identified with the following subgroup
of SL3 (R):
  
 a b x 
G(R) :=  c d y  : ad − bc = 1
 
0 0 1

which is the semi-direct product G(R) = SL2 (R) ⋉ V2 (R) where

   
 a b 0   1 0 x 
SL2 (R) ≃  c d 0  and V2 (R) =  0 1 y  ≃ R2 .
   
0 0 1 0 0 1
C. Matheus 93

Moreover, we identify the space of affine lattices E with G(R)/G(Z) and

we define
   
s 0 0 1 −2t −t2
As :=  0 1/s 0  and U (t) :=  0 1 t . (3.1)
0 0 1 0 0 1

Finally, we concluded that these identifications reduce our task to the proof
of the following result:
Theorem 3.4.1. For all f ∈ C0 (E) it holds
Z 1 Z
f (As · U (t))dt → f dµE .
0 E

As we already told, this theorem will be a consequence of a more general

result about the equidistribution of non-linear horocycles. In order to state
precisely this general result, let us introduce the following definition:
Definition 3.4.1. A horocyclic section (or horocycle) is a map σ : R →
G(R) of the form  
1 t x(t)
σ(t) =  0 1 y(t)  (3.2)
0 0 1
such that
σ(t + p0 ) = σ(t)γ0
for some integer p0 ≥ 1 and some element γ0 ∈ G(Z).
Remark 3.4.1. Given a horocycle σ, there exists a minimal integer p ≥ 1
such that σ(t + p) = σ(t)γ for some γ ∈ G(Z). This integer p is the period
of σ in E = G(R)/G(Z).
Remark 3.4.2. The name horocycle is motivated by the fact that the
natural projection of the space of affine lattices E to the space of lattices
B sends a horocyclic section of E to a (usual) horocycle around a cusp of
B.
Definition 3.4.2. A horocycle σ is called linear (over the rationals Q)
whenever for all α, β ∈ Q it holds

m ({t ∈ [0, p] : x(t) = αt + β}) > 0,

where m stands for the Lebesgue measure. Otherwise, the horocycle σ is

called non-linear.
Remark 3.4.3. The behavior of y(t) doesn’t have any influence in the
definition of linear horocycles.
94 Chapter 3. Elkies-McMullen theorem

Remark 3.4.4. A real-analytic horocycle σ is linear if and only if x(t) ≡

αt + β for some α, β ∈ Q (since any non-constant real-analytic function
has a discrete set of zeroes).
Comparing the equation (3.1) and (3.2) and using the remark 3.4.4, we
see that  
1 t −t2 /4
σ(t) := U (−t/2) :=  0 1 −t/2 
0 0 1
is a non-linear horocycle with period p = 2 and x(t) = −t2 /4. There-
fore, the theorem 3.4.1 is an immediate consequence of the following more
general fact:
Theorem 3.4.2 (Equidistribution of non-linear horocycles). Let σ : R →
G(R) be a non-linear horocycle of period p. Then, the circle As · σ become
equidistributed in E, i.e.,
Z Z
1 p
lim f (As · σ(t))dt = f (x)dµE (x).
s→∞ p 0 E

Remark 3.4.5. The main ingredients in this result are: the ”linear part”
of the horocycle is an unipotent matrix and the horocycle is non-linear.
Indeed, during the proof of Theorem 3.4.2, we will use the fact that the
linear part of the horocycle is unipotent to apply Ratner’s theorem in order
to reduce the list of candidates to the distribution law µ of the horocycle to
a countable quantity of possibilities (among them µE ). Then, we will use
the non-linearity to exclude all exotic possibilities.
Remark 3.4.6. The assumption of non-linearity of the horocycle is es-
sential: when it is linear, the conclusion of Theorem 3.4.2 is simply false!
We will come back to this point after the proof of this theorem.
After this considerations, we will dedicate the rest of this last section of
the last chapter of this book to the proof of Theorem 3.4.2. To do so, we
will use the following scheme:
• during the next subsection, we will revise some basic facts about
invariant measure and we will see some properties of the probability
measure µ associated to the distribution law of As · σ(t);
• in the sequel, we will use Ratner theorem to show that there are only
a countable quantity of possibilities to the distribution law µ;
• finally, in the last subsection, we will use the non-linearity of the
horocycle σ to show that µ = µE .
Now we start to formalize this program.
C. Matheus 95

3.4.1 The distribution law of a loop

Given a loop σ : R/pZ → E, we denote by m(σ) the natural probability
measure supported on the image of σ:
Z Z
1 p
f dm(σ) := f (σ(t))dt
E p 0
for any f ∈ C0 (E).
Furthermore, given σ : R/pZ → E a non-linear horocycle of period p,
we denote by σs := As · σ, so that Theorem 3.4.2 is equivalent to:
Theorem 3.4.3 (Equidistribution of non-linear horocycles, 2nd version).
For any non-linear horocycle σ it holds

m(σs ) = (As )∗ m(σ) → µE

when s → ∞.
Here the convergence means weak-* convergence. By the Banach-Alaoglu
theorem, we know that m(σs ) possesses a subsequence converging to a mea-
sure µ. In particular, our task consists into showing that µ = µE is the
unique possible limit of all convergent subsequences.
In this direction, consider the “derivative” map D from the space of
affine lattices E to the space of lattices B assigning to each element g ∈ E
its linear part D(g) ∈ B, i.e.,
   
a b x a b 0
D  c d y  :=  c d 0  .
0 0 1 0 0 1
Observe that, a priori, the projection of the Haar probability measure µE
of E by D isn’t necessarily equal to the Haar probability of µB of B. Thus,
as a preliminary work in the direction of the proof of Theorem 3.4.3, let
us verify that the projection of µ by D is correct:
Proposition 3.4.1. We have D∗ µ = µB .
Proof. The image H of D ◦ σ is a horocycle (in the usual sense) of the
space B. On the other hand, D sends the orbits of the ”Teichmuller
geodesic flow” As of E to the geodesics of B and D sends the measure
m(σ) to the Haar measure µH of H. Finally, a simple argument shows
that the geodesic flow of B pushes H far from the cusps of B so that H
becomes equidistributed (for further details see Theorem 2.4 of Elkies and
McMullen paper). Putting these facts together, it follows that

D∗ µ = lim(As )∗ µH = µB .

This concludes the proof.

96 Chapter 3. Elkies-McMullen theorem

Remark 3.4.7. A direct consequence of Proposition 3.4.1 is the fact that

µ is a probability measure on E, i.e., µ(E) = 1. In particular, the mass
of the probabilities m(σs ) is conserved at the limit. This is a non-trivial
remark because E is non-compact (so that the mass could escape to infinity
a priori)!

As we are going to see later, in order to fit the context of Ratner theorem,
we need to know that µ is invariant by an unipotent subgroup of SL2 (R).
In this direction, we introduce the subgroup
 
1 t 0
N (t) :=  0 1 0  .
0 0 1

Note that this unipotent subgroup appears naturally in view of the formula
D ◦ σ(t) = N (t) whenever σ(t) is a horocycle. The following preparatory
result will put us in the setting of Ratner theorem:

Proposition 3.4.2. The probability measure µ is N (R)-invariant.

Proof. Fix τ ∈ R. Consider σs (t) = As · σ(t) and ηs (t) = Nτ · σs (t) where

σ(t) is a horocycle. We have that
   τ τ y(t) 
s st sx(t) s st + s sx(t) + s
σs (t) =  0 1s y(t)
s
 , ηs (t) =  0 1 y(t) .
s s
0 0 1 0 0 1

In order to compare σs (t) and ηs (t), we perform a change of variables to

get the same linear parts. More precisely, we define u = τ /s2 and we
consider
 
s st sx(t − u) + s−1 τ y(t − u)
ρs (t) := ηs (t − u) :=  0 1/s y(t − u) .
0 0 1

Recall that m(σs ) → µ, so that

m(ρs ) = m(ηs ) = (Nτ )∗ m(σs ) → (Nτ )∗ µ. (3.3)

On the other hand, we have that D ◦ ρs = D ◦ σs , so that the distance

between ρs and σs is given by the distance between the vectors obtained
from the third column of the matrices:
   
 sx(t − u) + τ y(t − u)/s sx(t) + τ y(t)/s 

d(ρs , σs ) =  y(t − u)/s − y(t)/s 

 1 1 
C. Matheus 97

Next, we use the fact that x(t) is Lipschitz, y(t) is bounded and u = τ /s2
to get

|sx(t) − sx(t − u)| ≤ s|x(t) − x(t − u)| ≤ O(su) = O(1/s)

and
|y(t)/s − y(t − u)/s| ≤ (|y(t)| + |y(t − u)|)/s = O(1/s).
Thus, we see that d(ρs , σs ) → 0 when s → ∞. In particular, it follows that
lim m(ρs ) = lim m(σs ) = µ. Putting this information together with (3.3),
we get
(Nτ )∗ µ = µ
so that the proof is complete.

Once we know that µ is invariant by the unipotent subgroup N (R), we

move to the discussion of Ratner theorem.

3.4.2 Ratner theorem and classification of µ

Ratner theorem can be stated as follows:

Theorem 3.4.4. Let Γ be a discrete subgroup of a connected Lie group G

and N be an unipotent subgroup. Let ν be a N -invariant ergodic probability
on G/Γ and denote by J the biggest subgroup of G leaving ν invariant.
Then, there exists x ∈ G/Γ such that ν(J · x) = 1. Furthermore, ν is the
Haar measure of J · x and its support is J · x (so that J · x is closed in
G/Γ).

The relevance of Ratner’s theorem in the context of Elkies and McMullen

is evident: since µ is invariant by the unipotent subgroup N , we can classify
µ by listing all closed subgroups of E.
Obviously Ratner’s theorem has a beautiful history including several
applications to several areas of Mathematics. In particular, it is nearly
impossible to describe its importance in a brief discussion, so that we
recommend the interested reader the nice exposition of D. Morris [11] (and
also Terence Tao’s blog “What’s new” http://terrytao.wordpress.com/ for
a series of blog posts about Ratner’s theorems).
In any case, we are going to use Ratner theorem as follows. Denoting
by F a fiber of E → B, we observe that F
is a complex torus C/Λ. For
each integer n ≥ 1 we define F [n] = n1 · Λ /Λ ⊂ F the torsion points of
order n with respect to the group structure of F and we denote by E[n]
the subbundle of E whose fibers are F [n].
S
Definition 3.4.3. E[n] is the subset of torsion points of E.
98 Chapter 3. Elkies-McMullen theorem

Next we introduce H(R) ⊂ G the subgroup of horizontal translations,

i.e., translations by vectors of the form (x, 0) ∈ R2 and H(r, ε) ⊂ G the
subset of translations by vectors (x, y) of the form |x| < r and |y| < ε.
The goal of this subsection is the application of Ratner theorem to show
the following classification result:
Theorem 3.4.5 (Classification of µ). We have that µ = µE or µ(H(R) ·
E[n]) > 0 for some n ≥ 1.
Unfortunately, this result isn’t an immediate consequence of Ratner the-
orem because we don’t know that µ is ergodic. However, this is not a big
deal since we can use the ergodic decomposition theorem to write µ as a
(“unique”) convex combination of N (R)-invariant ergodic measures:
Z
µ = νdP (ν).

Remark 3.4.8. Usually the ergodic decomposition theorem concerns com-

pact spaces. In the specific case of E (a non-compact space), we apply the
ergodic decomposition theorem to the one-point compactification of E and
we restrict it to E.
Now, for each N (R)-invariant ergodic probability ν on E, we define

J(ν) := {g ∈ G(R) : g∗ ν = ν},

that is, J(ν) is the biggest subgroup of G(R) leaving ν invariant. Observe
that J(ν) is closed and N (R) ⊂ J(ν).
Proposition 3.4.3. For almost every ν of the ergodic decomposition of
µ, we have
D∗ ν = µB and D(J(ν)) = SL2 (R).
R
Proof. From Proposition 3.4.1, we know that µB = D∗ µ = D∗ νdP (ν).
Since the action of N (R) on (B, µB ) is ergodic (because it is the action of
the horocyclic flow on B), it follows that D∗ ν = µB for almost every ν.
On the other hand, by Ratner theorem, we know that ν is supported on
a closed orbit J(ν) · x ⊂ E. Thus,

D(J(ν)) · D(x) = D(J(ν) · x) = D(supp(ν)) = supp(D∗ ν).

Since D∗ ν = µB , we obtain

D(J(ν)) · D(x) = supp(µB ) = B = SL2 (R)/SL2 (Z).

Therefore, D(J(ν)) = SL2 (R). This completes the proof.

Now we recall the following proposition about SL2 (R)-actions:
C. Matheus 99

Proposition 3.4.4. Every affine action of SL2 (R) on Rk has fixed points.

Proof. By Weyl’s unitary trick, this action can be extended to a SL2 (C)-
action on Ck . On the other hand, a fixed point p ∈ Ck of the compact
subgroup SU2 (C) can be easily constructed (e.g., by averaging). Since
C · su2 (C) = sl2 (C), the point p is also fixed by the SL2 (C) and, a fortiori,
by the SL2 (R). Hence, the real part of p is a fixed point of SL2 (R) on
Rk .

Proposition 3.4.5. If H ⊂ G(R) is a subgroup such that D(H) = SL2 (R),

then H = G(R) or H is conjugated to SL2 (R).

Proof. Since D(H) = SL2 (R), the kernel K of the derivative map D :
H → SL2 (R) is a SL2 (R)-invariant subgroup of V2 (R) ≃ R2 , so that one
of the following two possibilities occurs:

• K = V2 (R): in this case, H = G(R);

• K = {e}: in this case, we have an affine action D−1 : SL2 (R) →

H ⊂ G(R) = ASL2 (R) of SL2 (R) on R2 which has fixed points by
Proposition 3.4.4; up to conjugation with an appropriate element of
V2 (R), we can assume that this fixed point is the origin and H =
SL2 (R).

This ends the proof.

Corollary 3.4.1. J(ν) = G(R) or J(ν) = g · SL2 (R) · g −1 for some

horizontal translation g ∈ H(R).

Proof. Since ν is N (R)-invariant, we know that N (R) ⊂ J(ν). Moreover,

by Proposition 3.4.3, we have that D(J(ν)) = SL2 (R). Hence, using
Proposition 3.4.5, it follows that J(ν) = G(R) or J(ν) = g · SL2 (R) · g −1 .
This concludes the argument.

Proposition 3.4.6. ν = µE or supp(ν) ⊂ g · E[n] for some integer n ≥ 1

and g ∈ H(R).

Proof. From the previous corollary, we have that J(ν) = G(R) or g ·

SL2 (R) · g −1 . In the first case, we see that ν = µE by J(ν)-invariance
of ν. In the second case, g −1 supp(ν) = SL2 (R) · x is a closed SL2 (R)-orbit
in E. Since such orbits are always contained in E[n] for some n ≥ 1, this
ends the proof.

At this stage, we can conclude this subsection with the proof of Theo-
rem 3.4.5:
100 Chapter 3. Elkies-McMullen theorem

R
Proof. We write the ergodic decomposition of µ as µ = νdP (ν). By the
proposition 3.4.6, almost every ergodic component ν of µ satisfy ν = µE
or supp(ν) ⊂ H(R) · E[n] for some n. Hence, we can write µ as:
∞
X
µ = a0 µE + an µn ,
n=1

P
∞
where an = 1 and supp(µn ) ⊂ H(R) · E[n]. In particular, if µ 6= µE
n=0
then an 6= 0 for some n ≥ 1, so that µ(H(R) · E[n]) > 0. This completes
the proof of the theorem.

In view of the classification of µ provided by Theorem 3.4.5, we see

that Theorem 3.4.3 (about the equidistribution of non-linear horocycles)
follows once we show that µ doesn’t see the torsion points of E. This is
the topic of the next subsection.

3.4.3 Non-linearity and torsion points

The main theorem of this subsection is

Theorem 3.4.6. Given σ a non-linear horocycle and µ an accumulation

point of the sequence of probability measures m(As · σ) (when s → ∞), we
have
µ(H(R) · E[n]) = 0
for every n ≥ 1.

Proof. Given ε > 0 and r > 0, define

U = H(r, ε) · E[n]

and
Ts = {t ∈ [0, p] : σs (t) ∈ U }.
We claim that
lim sup m(Ts ) = O(ε), (3.4)
s→∞

where m is the Lebesgue measure. In order to compute m(Ts ), it is con-

venient to pass to the universal cover G = G(R) of E = G/G(Z). We
start with
S the observation that E[n] is covered by the SL2 (R)-orbit of
G[n] = G[n]i,j where
 j
 
 a b ni a + n b 
G[n]i,j =  c d ni c + nj d  : ad − bc = 1 .
 
0 0 1
C. Matheus 101

In particular, the points of G[n] belonging to the same fiber of σs (t) are
 
s st ni s + nj st
ρi,j
s (t) =
 0 s−1 j −1 .
ns
0 0 1
Taking the Euclidean
S metric on the third column of the matrices above,
we see that Ts = Tsi,j where
    i  
i,j sx(t) n s + nj st
Ts = t : − ∈ H(r, ε) .
s−1 y(t) s−1 nj

In particular, Tsi,j ⊂ Xsi,j ∩ Ysi,j where

i j
Xsi,j = {t : |x(t) − − t| < r/s}
n n
and
j
Ysi,j = {t : |y(t) − | < εs}.
n
At this point, we will use the non-linearity of σ to get that the subset
of t with x(t) = ni + nj t has zero Lebesgue measure, so that, for each i, j
fixed, we have
lim m(Xsi,j ) = 0 (3.5)
s→∞

On the other hand, we can use that x(t) is Lipschitz to estimate m(Xsi,j )
when j is large: more precisely, whenever |j| > M := 2n sup |x′ (t)|, the
0≤t≤p
subset Xsi,j is the pre-image of an interval of size 1/s by a map whose
derivative has order j/n. Hence,
m(Xsi,j ) = O(1/s|j|) for all |j| > M. (3.6)
Moreover, we note that
Ysi,j = ∅ when |j| ≥ Js := n(sε + sup |y(t)|). (3.7)
0≤t≤p

and
r j
Xsi,j = ∅ when |i| ≥ Is (j) := n( + | | + sup |x(t)|). (3.8)
s n 0≤t≤p

Finally, we observe that

Js = O(sε) and Is (j) = O(|j| + 1) for s large. (3.9)
Keeping these facts in mind, we can estimate m(Ts ): by (3.7) and (3.8) it
follows that X X
m(Ts ) ≤ m(Xsi,j ) (3.10)
|j|<Js |i|<Is (j)
102 Chapter 3. Elkies-McMullen theorem

Now we split the right-hand sum into two parts:

X X X X X X
m(Xsi,j ) = m(Xsi,j ) + m(Xsi,j )
|j|<Js |i|<Is (j) M <|j|<Js |i|<Is (j) |j|≤M |i|<Is (j)

Next, we note that the equation (3.6) implies that the first part of this
sum is O(|Js |ε/s) = O(ε2 ) (because (3.9) says that |Is | = O(|j| + 1) and
|Js | = O(sε)) and the second part of this sum occurs over a finite set of
indices i, j so that (3.5) says that it tends to zero (when s increases). Thus,
putting these two estimates together with (3.10), we see that, for large s,
it holds
m(Ts ) = O(ε),
so that the desired estimate (3.4) follows.
Finally, we recall that ms (U ) = m(Ts )/p, so that the estimate (3.4)
implies µ(H(r, ε) · E[n]) = O(ε) for all r, ε > 0. Making ε → 0 first and
r → ∞ after, it follows that µ(H(R)·E[n]) = 0, so that the desired theorem
is proved.

Once we proved Theorem 3.4.6, it is an easy task to conclude the proof

of Theorem 3.4.3 (or equivalently, Theorem 3.4.2). In fact, this is the
content of the next (short) final subsection below.

3.4.4 End of the proof of Theorem 3.4.3

Given σ a non-linear horocycle, consider any accumulation point µ of
m(As · σ) when s → ∞. By Theorem 3.4.6,Sµ gives zero mass to the
horizontal translations of the torsion points E[n] of E. Hence, the
n≥1
classification theorem (Theorem 3.4.5) implies that µ = µE . In other
words, we have that µE is the unique accumulation point of the sequence
m(As · σ). This shows that

m(As · σ) → µE

so that the proof of Theorem 3.4.3 is complete.

At this point, our exposition of the proof of Elkies and McMullen is
finished! Closing the last section of the final chapter of this book, we
make the following remark:

Remark 3.4.9. The equidistribution theorem (Theorem 3.4.3) is optimal,

i.e., it never holds for linear horocycles σ: if x(t) = ni + nj t for a positive
measure subset of t then µ(E[n]) > 0 so that m(As · σ) can’t converge to
µE .
Bibliography

[1] A. Arbieto, C. Matheus and C.G. Moreira, Aspectos Ergódicos

da Teoria dos Números. Pub. Mat. IMPA., 26.o Colóquio Brasileiro
de Matemática, 2007.

[2] V. Brun, La serie 1/5 + 1/7 + 1/11 + 1/13 + 1/17 + 1/19 + 1/29 +
1/31 + 1/41 + 1/43 + 1/59 + 1/61 + ..., les denominateurs sont nombres
premiers jumeaux est convergente ou finie. Bull. Sci. Math. 43, 124-
128, 1919.

[3] J. Chen, On the representation of a larger even integer as the sum

of a prime and the product of at most two primes. Sci. Sinica 16,
157-176, 1973.

[4] M. Elkin, An improved construction of progression-free sets,

Preprint 2008, http://arxiv.org/abs/0801.4310.

[5] W. Feller, Introduction to Probability Theory and its applications.

Vol. I, 3rd Edition, Wiley, New York. 1968.
√
[6] N. Elkies and C. McMullen, Gaps in n mod 1 and Ergodic
Theory. Duke Math. Journal 123, 95-139, 2004.

[7] A. Goldston, J. Pintz and Y. Yıldırım, Primes in Tuples I. An-

nals of Math. 170, 819-862, 2009.

[8] A. Goldston, J. Pintz and Y. Yıldırım, Primes in Tuples II.

Preprint 2007. http://arxiv.org/abs/0710.2728.

[9] B. Green and T. Tao, The primes contain arbitrarily long arith-
metic progressions. Annals of Math. 167, 481–547, 2008.

[10] B. Green and J. Wolf, A note on Elkin’s improvement of Behrend’s

construction, Preprint 2008, http://arxiv.org/abs/0810.0732.

[11] D. Morris, Ratner’s Theorems on Unipotent Flows. Chicago Lec-

tures in Mathematics Series, University of Chicago Press. 2005.

103
104 Bibliography

[12] K. Roth, On certain sets of integers. J. London Math. Soc. 28, 245-
252, 1953.
[13] L. G. Schnirelman, Uspekhi Math. Nauk 6, 3-8, 1939.
[14] E. Szemerédi, On sets of integers containing no k elements in arith-
metic progression. Acta Arith. 27, 299-345, 1975.
[15] T. Tao, Arithmetic Progressions and the Primes (El Escorial Lec-
tures), Collect. Math., 37-88, 2006.
[16] T. Tao, The ergodic and combinatorial approaches to Szemerédi’s
theorem. CRM Proc. Lecture Notes 43, 145-193, 2007.
[17] T. Tao, The Gaussian primes contain arbitrarily shaped constella-
tions. J. d’Analyse Mathematique 99, 109-176, 2006.
[18] T. Tao and T. Ziegler, The primes contain arbitrarily long poly-
nomial progressions. Acta Math. 201, 213-305, 2008.
[19] J. Van der Corput, Uber Summen von Primzahlen und
Primzahlquadraten. Math. Ann. 116, 1-50, 1939.
[20] I. Vinogradov, Representation of an odd prime as a sum of three
primes. Comptes Rendus (Doklady) de l’Académie des Sciences de
PURSS 15, 291-294, 1937.

Alexander ARBIETO
UFRJ, Universidade Federal do Rio de Janeiro
Av. Athos da Silveira Ramos, 149, Ilha do Fundão, CEP 68530
Rio de Janeiro, RJ, Brazil
alexande@impa.br
http://www.im.ufrj.br/∼arbieto
Carlos MATHEUS
Collège de France
3, Rue d’Ulm, CEDEX 05
Paris, France
matheus@impa.br
http://www.impa.br/∼cmateus
Carlos Gustavo (Gugu) MOREIRA
IMPA, Instituto de Matemática Pura e Aplicada
Estrada D. Castorina, 110, CEP 22.460-320
Rio de Janeiro, RJ, Brazil
gugu@impa.br
http://www.impa.br/∼gugu