Straight from the Book

Titu Andreescu
Gabriel Dospinescu

Straight from the Book

Press
Titu Andreescu Gabriel Dospinesca
University of Texas at Dallas Ecole Normale Supérieure, Lyon

The only way to learn mathematics is to do mathematics.

Library of Congress Control Number: 2012951362

~Paul Halmos
ISBN-13: 978-0-9799269-3-8 ISBN-10: 0-9799269-3-9

© 2012 XYZ Press, LLC

All rights reserved. This work may not be translated or copied in whole or in
part without the written permission of the publisher (XYZ Press, LLC, 3425
Neiman Rd., Plano, TX 75025, USA) and the authors except for brief excerpts
in connection with reviews or scholarly analysis. Use in connection with any
form of nformation storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter
developed is forbidden. The use in this publication of tradenames, trademarks,
service marks and similar terms, even if they are not identified as such, is not
to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights,

987654321

www.awesomemath.org

Cover design by Lury Ulzutuev

Foreword

This book is a follow-on from the authors’ earlier book, ‘Problems from
the Book’. However, it can certainly be read as a stand-alone book: it is not
vital to have read the earlier book.
The previous book was based around a collection of problems. In contrast,
this book is based around a collection of solutions. These are solutions to
some of the (often extremely challenging) problems from the earlier book.
The topics chosen reflect those from the first twelve chapters of the previous
book: s0 we have Cauchy-Schwarz, Algebraic Number Theory, Formal Series,
Lagrange Interpolation, to name but a few.
The book is one of the most remarkable mathematical texts I have ever
seen. First of all, there is the richness of the problems, and the huge variety
of solutions. The authors try to give several solutions to each problem, and
moreover give insight about why each proof is the way it is, in what way the
solutions differ from each other, and so on. The amount of work that has been
put in, to compile and interrelate these solutions, is simply staggering. There
is enough here to keep any devotee of problems going for years and years.
Secondly, the book is far more than a collection of solutions. The solu-
tions are used as motivation for the introduction of some very clear expositions
of mathematics. And this is modern, current, up-to-the-minute mathematics.
For example, a discussion of Extremal Graph Theory leads to the celebrated
Szemerédi-Trotter theorem on crossing numbers, and to the amazing applica-
tions of this by Székely and then on to the very recent sum-product estimates
of Elekes, Bourgain, Katz, Tao and others. This is absolutely state-of-the-art
material. It is presented very clearly: in fact, it is probably the best exposition
of this that I have seen in print.
viii

As another example, the Cauchy-Schwarz section leads on to the devel-

opements of sieve theory, like the Large Sieve of Linnik and the Turan-Kubilius
inequality. And again, everything is incredibly clearly presented. The same
applies to the very large sections on Algebraic Number Theory, on p-adic
Analysis, and many others.
It is quite remarkable that the authors even know so much current math- Preface
ematics — 1 do not think any of my colleagues would be so well-informed over
50 wide an area. It is also remarkable that, at least in the areas in which I
am competent to judge, their explanations of these topics are polished and
This bock is a compilation of many suggestions, much advice, an even
exceptionally well thought-out: they give just the right words to help someone
more hard work. Its main objective is to provide solutions to the problems
understand what is going on.
which were originally proposed in the first 12 chapters of Problems from the
Overall, this seems to me like an ‘instant classic’. There is so much
Book. We were not able to provide full solutions in our first volume due to the
material, of such a high quality, wherever one turns. Indeed, if one opens the
lack of space. In addition, the statements of the proposed problems contained -
book at random (as I have done several times), one is pulled in immediately
typos and some elementary mistakes which needed further editing. Finally,
by the lovely exposition. Everyone who loves mathematics and mathematical
the problems were also considered to be quite difficult to tackle. With these
thinking should acquire this book.
points in mind, we came up with a two-part plan: to correct the identified
errors and to publish comprehensive solutions to the problems.
Imre Leader The first task, editing the statements of the proposed problems, was sim-
ple and has already been completed in the second edition of Problems from the
Professor of Pure Mathematics
Book. Although we focused on changing several problems, we also introduced
University of Cambridge meny new ones. The second task, providing full solutions, however, proved to
be more challenging than expected, so we asked for help. We created a forum
on www.mathlinks.ro (a familiar site for problem-solving enthusiasts) where
solutions to the proposed problems were gathered. It was a great pleasure to
witness the passion with which some of the best problem-solvers on mathlinks
attacked these tough-nuts. This new book is the result of their common effort,
and we thank them.
Providing solutions to every problem within the limited space of one vol-
ume turned oui to be an optimistic plan. Only the solutions to problems
from the first 12 chapters of the second edition of Problems from the Book are
presented here. Furthermore, many of the problems are difficult and require
a rather extensive mathematical background. We decided, therefore, to com-
plement the problems and solutions with a series of addenda, using various
problems as starting points for excursions into *real mathematics”. Although
 xi

we never underestimated the role of problem solving, we strongly believe that a beneficial elementary result called lifting the exponent lemma, as well as
the reader will benefit more from embarking on a mathematical journey rather more advanced techniques from p-adic analysis. One of the most beautiful re-
than navigating a huge list of scattered problems. This book tries to reconcile sults presented in this chapter in the celebrated Skolem-Mahler-Lech theorem
problem-solving with ” professional mathematics”. concerning the zeros of a linearly recursive sequence. Readers will perhaps
Let us now delve into the structure of the book which consists of 12 chap- appreciate the applications of p-adic analysis, which is covered extensively in
ters and presents the problems and solutions proposed in the corresponding a long addendum. This addendum discusses, from a foundational level, the
chapters of Problems from the Book, second edition. Many of these problems arithmetic of p-adic numbers, a subject that plays a central role in modern
are fairly difficult and different approaches are presented for a majority of number theory. Once the basic groundwork has been laid, we discuss the
them. At the end of each chapter, we acknowledge those who provided so- p-adic analogues of classical functions (exponential, logarithm, Gamma func-
lutions. Some chapters are followed by one or two addenda, which present tion) and their applications to difficult congruences (for instance, Kazandzidis™
topics of more advanced mathematics stemming from the elementary topics famous supercongruence). This serves as a good opportunity to explore the
discussed in the problems of that chapter. arithmetic of Bernoulli numbers, Volkenborn’s theory of p-adic integration,
The first two chapters focus on elementary algebraic inequalities (a no- Mahler and Amice’s classical theorems characterizing continuous and locally
table caveat is that some of the problems are quite challenging) and there is analytic functions on the ring of p-adic integers or Morita’s construction of the
not much to comment regarding these chapters except for the fact that al- p-adic Gamma function. A second addendum to this chapter discusses various
gebraic inequalities have proven fairly popular at mathematics competitions. classical estimates on prime numbers, which are used throughout the book.
To provide relief from this rather dry landscape (the reader will notice that Chapter 4 discusses problems and elementary topics related to prime num-
most of the problems in these two chapters start with "Let a,b, ¢ be positive bers of the form 4k + 1 and 4k + 3. The most intriguing problem discussed
real numbers”), we included an addendum presenting deep applications of the is, without any doubt, Cohn’s renowned theorem characterizing the perfect
Cauchy-Schwarz inequality in analytic number theory. For instance, we dis- squares in the Lucas sequence. Chapter 5 is dedicated again to the yoga of
cuss Gallagher’s sieve, Linnik’s large sieve and its version due to Montgomery. algebraic inequalities and is followed by an addendum discussing applications
We apply these results to the distribution of prime numbers (for instance of Holder’s inequality.
Brun's famous theorem stating that the sum of the inverses of the twin primes
Chapter 6 focuses on extremal graph theory. Most of the problems revolve
converges). The note-worthy Turdn-Kubilius inequality and its classical appli-
around Turdn’s theorem, however the reader will also be exposed to topics
cations (the Hardy-Ramanujan theorem on the distribution of prime factors of
such as chromatic numbers, bipartite graphs, etc. This chapter is followed
n, Erdés’ multiplication table problem, Wirsing’s generalization of the prime
by a relatively advanced addendum, discussing various topics related to the
number theorem) are also discussed. The reader will be exposed 1o the power
Szemerédi-Trotter theorem, which gives bounds for the number of incidences
of the Cauchy-Schwarz inequality in real mathematics and, hopefully, will un-
between a set of points and a set of curves. We discuss the theorem’s classical
derstand that the gymmnastics of three-variables algebraic inequalities is not
probabilistic proof, its generalization to multi-graphs due to Székely and its
the Holy Grail.
application to the sum-product problem due to Elekes, as well as more recent
Chapter 3 discusses problems related to the unique factorization of inte- developments due to Bourgain, Katz, and Tao. These results are then applied
gers and the p-adic valnation maps. Among the topics discussed, we cover to natural and nontrivial geometric questions (for instance: what is the least
the local-global principle (which is extremely helpful in proving divisibilities number of distinct distances determined by n points in the plane? what is the
or arithmetic identities), Legendre’s formula giving the p-adic valuation of !, maximal number of triangles of the same area?). Finally, another addendum
xii xiii

completes this chapter and is dedicated to the powerful probabilistic method. instance, we discuss arithmetic properties of cyclotomic polynomials (includ-
After a short discussion on finite probability spaces, we provide many examples ing Mann’s beautiful theorem on linear equations in roots of unity), rationality
of combinatorial applications. problems, and various applications of the theorem of symmetric polynomials.
Chapter 7 involves combinatorial and number theoretic applications of In addition, we present techniques rooted in the theory of ideals in number
finite Fourier analysis. The central principles of this chapter include the roots fields, finite fields and p-adic methods. We also give an overview of the el-
of unity and the fact that congruences between integers can be expressed in ementary algebraic number theory in the addendum following this chapter.
terms of sums of powers of roots of unity. To provide the reader with a broader After a brief review on ideals, field extensions, and algebraic numbers, we
view, we beefed-up this chapter with an addendum discussing Fourier analysis proceed with a discussion of the primitive element theorem and embeddings
on finite abelian groups and applying it to Gauss sums of Dirichlet characters, of number fields into C. We also briefly survey Galois theory, and the fun-
additive problems, combinatorial or analytic number theory. For instance, damental theorem on the prime factorization of ideals in number fields, due
the reader will find a discussion of the Pélya-Vinogradov inequality and Vino- to Kummer and Dedekind. Once the foundation has been set, we discuss the
gradov’s beautiful use of this inequality to deduce a rather strong bound on prime factorization in quadratic and cyclotomic fields and apply these tech-
the least quadratic non-residue mod p. At the same time, we explore Dirich- niques to basic problems that explore the aforementioned theories. Finally, we
let’s L-functions, culminating in a proof of Dirichlet’s theorem on arithmetic discuss various applications of Bauer’s theorem and of Chebotarev’s theorem.
progressions. This section is structured to first present the usual analytic proof The next addendum is concerned with the fascinating topic of counting the
(since it is really a masterpiece from all points of view), up to some important number of solutions modulo p of systems of polynomial equations. We use
simplification due to Monsky. At the same time, we also discuss how to turn this as an opportunity to state and prove the basic structural results on finite
this into an elementary proof that avoids complex analysis and dramatically fields and introduce the Gauss and Jacobi sums. We go ahead and count the
uses Abel’s summation formula. number of points over a finite field of a diagonal hypersurface and to compute
Chapter 8 focuses on diverse applications of generating functions. This s zeta function. This is a beautiful theorem of Weil, the very tip of a massive
is an absolutely crucial tool in combinatorics, be it additive or enumerative. iceberg.
In this section, the reader will have the opportunity to explore enumerative Chapter 10 focuses on the arithmetic of polynomials with integer coeffi-
problems (Catalan’s problem, counting the number of solutions of linear dio- cients. An essential aspect of the discussion concentrates on Mahler expan-
phantine equations or the number of irreducible polynomials mod p, of fixed sions, the theory of finite differences, and their applications. The techniques
degree). We also discuss exotic combinatorial identities or recursive sequences, used in this chapter are rather diverse. Although the problems can be consid-
which can be solved elegantly using generating functions, but also rather chal- ered basic, they are challenging and require advanced problem-solving skills.
lenging congruences that appear so often in number theory. The chapter is Chapter 11 provides respite from the difficult tasks mentioned above. It
followed by an addendum presenting a very classical topic in enumerative dis Lagrange’s interpolation formula, allowing a unified presentation of
combinatorics, Lagrange’s inversion formula. Among the applications, let us various estimates on polynomials.
mention Abel’s identity, other derived combinatorial identities, Cayley’s the- The longest and certainly most challenging chapter is the last one. It
orem on labeled trees, and various related problems. explores several algebraic techniques in combinatorics. The methods are stan-
Chapter 9 is rather extensive, due to the vast nature of the topic covered, dard but powerful. The last part of the chapter deals with applications to
algebraic number theory. While we are able only to scratch the surface, nev- geometry, presenting some of Dehn’s wonderful ideas. The last problem pre-
ertheless the reader will find a variety of intriguing techniques and ideas. For sented in the book is the famous Freiling, Laczkovich, Rinne, Szekeres theorem,
xiv

a stunningly beautiful application of algebraic combinatorics.

We would like to thank, again, the members of the mathlinks site for
their invaluable contribution in providing solutions to many of the problems
in this book. Special thanks are due to Richard Stong, who did a remarkable
job by pointing out many inaccuracies and suggesting numerous alternative
solutions. We would also like to thank Joshua Nichols-Barrer, Kathy Cordeiro
and Radu Sorici, who gave the manuscript a readable form and corrected Contents
several infelicities. Many of the problems and results in this book were used by
the authors in courses at the AwesomeMath Summer Program, and students’
reactions guided us in the process of simplifying or adding more details to the
discussed problems. We wish to thank them all, for their courage in taking
and sticking with these courses, as well as for their valuable suggestions. 1 Some Useful Substitutions 1
1.1 Therelation a® +b2 4+ =abc+4 . .. ... .......... 2
1.2 The relations abc = a + b+ c+ 2 and ab + be + ca + 2abe -9
Titu Andreescu Gabriel Dospinescu 1.3 Therelationa? + 6% +c2+2abc=1 ... ............ 21
[email protected] [email protected] L4 Notes ... ... ... ... ... 27

2 Always Cauchy-Schwarz. .. 29
21 Notes . ... ..........
.. ... . ... 81
62
3 Look at the Exponent 91
3.1 Introduction 91
3.2 92
3.3 96
34 104
3.5 Lifting exponent lemma . . . .. ... ... .. .. ... ... . 110
3.6 p-adic techniques 116
3.7 Miscellaneous problems 125
3.8 Notes
Addendum 3.A Classical Estimates on Prime Numbers . . . . . . . 135
Addendum 3.B An Introduction to p-adic Numbers . . . . ..... 141

4 Primes and Squares 189

41 Notes . ... ... ... ... ... 203
xvi Contents Contents xvii

5 Ty’s Lemma 205 10 Arithmetic Properties of Polynomials 485

5.1 Notes 10.1 The a — b|f{a) — f(b) trick . . . .. .. ... ..., .. 485
Addendum 5.A Holder’s Inequality in Action 10.2 Derivatives and p-adic Taylor expansions . . . . . . .. 494
10.3 Hilbert polynomials and Mahler expansions . . . . . .497
6 Some Cla 10.4 p-adic estimates .. 506
6.1 Notes 10.5 Miscellaneous problems .. 513
Addendum 6.A Some Pearls of Extremal Graph Theory 5 106 Notes . ... .. ... .. .. ... ... ... . 520
Addendum 6.B Probabilities in Combinatorics 5
11 Lagrange Interpolation Formula 521
7 Complex Combinatorics 281 TLI Notes ... ... ... 537
7.1 Tiling and coloring problems . . . . ... ... ... ... ... 281
7.2 Counting problems . . . . . 12 Higher Algebra in Combinatorics 539
7.3 Miscellaneous problems 12.1 The determinant trick . . . ... ... ... ... ... ..., 541
T4 Notes . . ... ... 12.2 Matricesover Fg . . . . ..... 546
Addendum 7.A Finite Fourier Analysis . . . ... ... ....... 310 12.3 Applications of bilinear algebra .. 552
12.4 Matrix equations . . . ..... .. 561
8 Formal Series Revisited 337 12.5 The linear independence trick . .. 568
81 Countingproblems . . ... ... ... ... .. ..... . ... 339 12.6 Applications to geometry . 576
8.2 Proving identities using generating functions . . . .. ..... 349 127 Notes . . . oot e 583
8.3 Recurrence relations
8.4 Additive properties . . . . Bibliography 585
8.5 Miscellaneous problems
86 Notes . ..........
Addendum 8.A Lagrange’s Inversion Theorem

9 A Little Introduction to Algebraic Number Theory 395

9.1 Tools from linear algebra . . .. .....
9.2 Cycotomy . .. ......
... ....
93 Thegedtrick ... .. ...........
9.4 The theorem of symmetric polynornials
9.5 Ideal theory and local methods . . . ... .. ... .......
9.6 Miscellaneous problems . . . . ... ... ... ... ...
9.7 Notes . ... ... ... ... ... ..
Addendum 9.A Equations over Finite Fields
Addendum 9.B A Glimpse of Algebraic Number Theory . ... .. 456
Chapter 1

Some Useful Substitutions

Let us first recall the classical substitutions that will be used in the fol-
lowing problems. All of these are discussed in detail in [3], chapter 1 and the
reader is invited to take a closer look there.
Consider three positive real numbers a,b,c. If abe = 1, a classical substi-
tution is
z
a=-, b:g, c:z
y z z
A less classical one is
z Y z
a= , b= , e=
¥tz z+z z+y
(for some positive real numbers z,y, z) whenever ab + bc + ca + 2abe = 1, or
its equivalent version
a= y+z, b= z+:v:’ P Tty
z y z
when abc = a + b + ¢ + 2 (the equivalence between the two relations follows
from the substitution (a,b,¢) — (£,1,1)). Two other very useful substitutions
concern the relations a2+ b2 +¢? = gbc+4 and a® + 42 + ¢? + 2abc = 1. In the
first case, with the extra assumption max(a,b,c) > 2, one can find positive
numbers z,y, z such that zyz = 1 and
1 1 1
a=z+-, b=y+-, c=z+-.
z Y z
2 Chapter 1. Some Useful Substitutions
1.1. The relation a® + b? + ¢? = abe + 4 3

In the second case one can find an acute-angled triangle ABC' such that
Thus (% — 4)(c* - 4) = (be+ 2a)?, which can also be written as
a = cos(A), b=cos(B), c=-cos(C).
(be + 20)°
b=2Me—2) = ———— 5
Of course, in practice one often needs to use a mixture of these substitutions b-2-2) br2)c+r2) =
and to be rather familiar with classical identities and inequalities. But expe-
‘Writing similar expressions for the other two variables, we are done. g
rience comes with practice, so let us delve into some exercises and problems
to see how things really work. The following exercise is trickier and one needs some algebraic skills in
order to solve it. We present two solutions, neither of which is really easy.
1.1 The relation a® + b + ¢? = abc + 4
2. Find all triples z, ¥, z of positive real numbers such that
‘We start with an easy exercise, based on the resolution of a quadratie
equation.
22+
+ 2 =ayr +4
zy+yz+ze
=2z +y+z)
1. Prove that if a,b,¢ > 0 satisfy ja + b% + ¢ — 4] = abe, then
Preof. By the second equation we have max(z,y,z) > 2 and so the first equa-
(@a=2)(b=2)+(b—2(c—2)+ (c-2)(a~2) 2 0. tion yields the existence of positive real numbers @, b, ¢ such that

Titu Andreescu, Gazeta Matematicd

Proof. If max(a,b, ¢} < 2, then everything is clear, so assume that and abe = 1.

max(a,b,¢) > 2.
Since abc = 1, we have
Then a2 + b% + ¢ - abe = 4, so there exist positive numbers z,y, z such that
zyz =1 and
a=z+~-,
1
x
b=yt-
1
y
1
c=z+-—.
z
Si-Ya Y-
s0 the second equation can be written
But then a,b,¢ > 2 and we are done again. 0

Proof. The most natural idea is to consider the hypothesis as s quadratic }:(g+§)=§:a+zm
equation in a, for instance. It becomes 4 % abe+ b% + ¢2 — 4 = 0, and solving
the equation yields The left-hand side is also equal to
gbes JETTHECD
Zc(a2 +b?) = (Za) (Zab) -3,
4 Chapter 1. Some Useful Substitutions 1.1. The relation
a® + b+ = abc+4 5

because Proof. The hypothesis yields the existence of positive real numbers z,y, 2 such
that O v,z Tz
Iy =¥ T2
y oz z oy 2z
We deduce that (35 —1)(3 ab—1) =4, Since . a > 3 and ¥ ab> 3 (by
the AM-GM inequality and the fact that ebc = 1), this can only happen if The miracle is that both sides of the inequality have very nice factorizations.
For the right-hand side, this is easy to observe, since
a=b=ec=]land thus whenz =y =2 = 2. o
1 1 1
Proof. 1f 2 +y = 2, the second equation yields zy = 4, so that (z —y)* = ~12 at+b+c+3=(zy+yz+2z) (I—y+~+7)
which is & contradiction. Thus +y # 2 and similarly y +2 # 2, z + 2 # 2.
The second equation yields and
4—zy
r+y—2
a2+b2+c2—3:2(§+g—)+3=(212) (Eé)
and a rather brutal insertion of this expression in the first equation gives For the left-hand side, things are more subtle, but one finally reaches the
identity
— )2 Ld—azy \*_
G (4—ap)?
S 20
-9 +<x+y72) 2—z—y atbt+ctabtbetca= (212) (EI—;)‘F(Z%) (Zzy).
Unless ¢ = y = 2, this implies the inequality 2 > z +y. If two of the numbers
.4,z are equal to 2, then trivially so is the third one. If not, the previous The desired inequality becomes simply the AM-GM inequality for two num-
argument shows that 2 > z -+ 4,2 > y+ 2 and 2 > z + z. But then the second bers! a
equation yields
The following problems are rather tricky mixtures of algebraic manipula-
tions and elementary number theory.
m+y+z>2x<y-¥>21y+yz+zzz2(z+y+2),
4. Find all triplets of positive integers (k,[, m) with sum 2002 and for which
the system
a contradiction. Thus, the only solutionis z =y =2 =2. [m}
Ly,
Y T
The following problem hides under a clever algebraic manipulation a very
simple AM-GM argument. The inequality is quite strong, as the reader can
easily see by trying a brute-force approach.
¥LZo
z ¥y
3. Prove that if a, b, ¢ > 2 satisfy @? + 8 + ¢ = abe + 4, then
Z24%cm
z T

)a?
+ b2+
a+b+ctabtac+be> 2/ (a+b+c+3 %~ 3), T z

has real sclutions.

Marian Tetiva Titu Andreescu, proposed for IMO 2002
6 Chapter 1. Some Useful Substitutions 1.1. The relation a® + b* + ¢? = abc + 4 7

Proof. The system has solutions if and only if

K2 2 m? = lhm+ 4.
An easy computation shows that this relation is equivalent to
ol ()
Part of the following problem can be dealt in a classical way, but we do
not know how to solve it entirely without using the trick of substitutions.

(k+2)(1+2)(m
+2) = (k+ 1 +m+2)% 6. The sequence {an)n>p is defined by ag = a; = 97 and

As k414 m = 2002, we deduce that any solution of the problem satisfies 41 = @nln-1 + /(a2 —1)(a2_; — 1)
k+1+m = 2002 and
for all n > 1. Prove that 2 + /2
F 2a,, is a perfect square for all n.
(E+2)(1+2(m+2) = (k+1+m+2)* = 20042 = 2* . 3% . 167 Titu Andreescu
A simple case analysis shows that the only solutions are k = I = 1000,m = 2 Proof. Writing
and its permutations. The result follows. O
(@n41 — anap-1)* = (“12; - 1)(“3.4 -9
5. Solve in positive integers the equation and simplifying this expression yields

(@+ 2y +2)(2+2) = (z+y+2+2)>% G+ a}+aki - 20n0010ns; = 1,

thus
Titu Andreescu 2an1) + (200)° + (200417 = (220-1)(202)(20011) = 4.
Since we clearly have a,, > 2 for all n, this implies the existence of a sequence
Proof. A simple algebraic manipulation shows that the equation is equivalent
Ta > 1 such that 20, = z, + z;* and such that 2,,; = z,7,_;. Thus logz,
t0 22 +y?+ 2% = 2yz-+4 and, seeing this as a quadratic equation in z, we obtain satisfies a Fibonacci-type recursive relation and so we can immediately find
the equivalent form (2% —4)(y%~4) = (ay—22), 1f 2% < 4, then y? < 4 as well out the general term of the sequence (an)n. Namely, a small computation
and so ¢ = y = 1, yielding z = 2. If x = 2, then y = z. In all other cases, we
shows that if we define a = 2 + /3, then z, = o where F,, is the nth
can find a positive square-free integer D (which is easily seen to be different
Fibonacci number. Thus
from 1) and positive integers u, v such that «* — 4 = Du? and y? — 4 = Do?.
Thus, solving the problem comes down to solving the generalized Pell equation 1( ar 1
a? — Db = 4, which is a classical topic: this equation always has nontrivial =g ( aAFu)
integer solutions and if (ag,bo) is the smallest solution with ag,by > 0, then and so
all solutions are given by
. 2
2+\/2*+2a,.:2+(a”"+%):(a""+ 1 ) .
_(ll(H‘bo\/D')" (ao*‘ bm/fi)” OF abs
S -2 B )
2
I The result follows, since a” +a™" € Z for all n, by the binomial formula. [
8 Chapter 1. Some Useful Substitutions 1.2. The relations abc = a +b+c+2 and ab+ bc+
ca + 2abe = 1 9

Remark 1.1. Here is an alternative proof of the fact that all terms of the 1.2 The relations abc =a+b+c+ 2 and
sequence are integers, without the use of substitutions. The method that we
ab + be+ ca + 2abe = 1
will use for this problem appears in many other problems. As we saw in the
previous solution, the sequence satisfies the recursive relation
The first inequality in the following problem is very useful in practice and
2 2, .2
Qg+ O+ Gy = 20y 10n0n1 = 1
we will meet it very often in the following problems.

Writing the same relation for 741 instead of n and subtracting the two yields 7. Prove that if ,3.2 > 0 and zyz =z + y + 2 + 2, then
the identity
“fwrz - aZL—l = 20n0n+1(@nt2 — An1). wy+y2+zmz2(w+y+z)andfi+fi+fi§g
Note that (ap), is an increasing sequence (this follows trivially by induction
from the recursive relation), so that we can divide by @np49 — @p—1 # 0 in the Proof. With the usual substitutions
previous relation and get ani2 = 2apan41 — Gn—1. The last relation clearly
implies that all terms of the sequence are integers (since one can immediately b+c cta a+b
T=——,
a
y= b
.= c
check that this is the case with the first three terms of the sequence). Note
the first inequality comes down (after clearing denominators and canceling out
square using this method. similar terms) to Schur’s inequality
Remark 1.2. There are a lot of examples of very complicated recurrence
relations that rather unexpectedly yield integers. For instance, the reader a{a —b){a—¢) +b{b— a){b— ¢} + c(c — a){c—b) > 0,
can try to prove the following result concerning Somos-5 sequences: let
ay=ag=--=as=1andlet while the second one follows.by adding up the inequalities

+10n44 + Gy 2@ /_T_ ja b 1 a 4 b o
ol =
Gn me“vaJ»c b+c” 2\c+a btc)
for n > 0. Then a, is an integer for all n. Similarly one defines Somos-6,
The reader may find a bit strange the first method of proof of the following
Somos-7, ete sequences by the formulas
problem, but it is actually a quite powerful one. We will use again this kind
of argument, see problem 11 for instance. Also, the third solution uses a very
useful technique.

Gng1lntf + 0ng20nt5 + Cnadlngg

a=ay=-=ag=1, Guu7= - s 8. Let z,y,2z > 0 be such that oy + yz + 22 = 2(z + y + z). Prove that
n
etc. One can prove (though this is not easy) that all terms of Somos-6 and zyz<rt+y+z+2.
Somos-7 sequences are integers. Surprisingly, this fails for Somos-8 sequences
{in which case a;7 is no longer an integer!), Gabriel Dospinescu, Mircea Lascu
10 Chapter 1, Some Useful Substitutions 1.2, The relations abc = a + b+ ¢+ 2 and ab + be + ca + 2abe = 1 1

Proof. We will argue by contradiction, assuming that zyz > z+y+2+2. We

and defining z,¥, z to be the three roots of this polynomial (in some order).
Increasing o3, i.e. lowering the constant term, corresponds geometrically to
claim that we can find 0 < r < 1 such that X = ra,Y = ry, Z = rz satisfy
lowering the graph. As we lower the graph, the smallest root increases, thus
XYZ = X +Y + Z+2. Indeed, this comes down to the vanishing of
we maintain three positive real roots until the smallest root becomes a double
flr)= riayz - rlz+y+z)—2 root. If the double root is at ¢ = a and the larger root at t = b, then we have
o1 =2a+b, 03 = a® + 2ab and 03 = a2b.
between ( and 1, and this is clear, since f(0) < 0 and f(1) > 0. Next, the
If we fix 0y and o3 with o2 = 20y as hypothesized, then we find b = ;4’")1“)-
condition zy + yz + 2z = 2(x + y + 2) yields
and because 0 < @ < b, we see that 1 < a < 2. By the discussion above
XY +YZ 4 ZX =2(X+Y +Z)<2AX
+Y + Z). zyz = 03 < a?b, so it suffices to show that

This contradicts the first inequality of problem 7. £ 2 (4 -a)a® <2+b+2=2a+ (4—a)a

o ) 2a+ 2=2a
Proaf. The condition can also be rewritten in the form
Na—p T2
for 1 <« < 2. But this rearranges to (¢ —2)%(a®
~ 1) > 0 and we are done.
E-Dy-D+E-—De~D+-Yz-1)=3 [J

or in the form The technique used in the first solution of the next problem is rather ver-
satile and the reader is invited to read the addendum 5.A for more examples.
ayz—w—y—z-1l=(z—1)(y—-1){z~1).

We will discuss several cases. If 2, g,z > 1, then by the AM-GM inequality 9. Let z,y,2 > 0 be such that zy + yz + 2z + zyz = 4. Prove that
and the first identity, we get
1 1 1\?
12 o=y = D2 - 17, 3<V§+fi+$) 2@ +2)(y+2)(z+2).
which yields, thanks to the second identity, the desired estimate.
If z,y, 2 < 1 or if only one of the numbers «,y, z is smaller than or equal Gabriel Dospinescu
to 1, then {(z ~ 1){y — 1)(z — 1) < O and so xyz < z + ¥ + 2 + 1 in this case.
Finally, if two of the numbers are smaller than 1, say z,y < 1, the desired Proof. Using the usnal substitution
inequality can be written in the form 0 < & + y + 2(1 — zy) + 2, which is
obvicus. o
Thre YT ore TTarw
Proof. For three positive real numbers z,y, 2 consider fixing the first two el-
ementary symmetric polynomials oy = © + y + 2 and 03 = zy + yz + 27 and the problem reduces to proving the inequality
letting o3 = @yz vary. This amounts to varying only the constant term in the
2
polynomial b+c {a+b+c)®
Pty =~ a1t + ot — 03 = (£ — 3)(t — ) (t — 2) 3(ZV " ) 2 T acTa
12 Chapter 1. Some Useful Substitutions 1.2, The relations abc = a + b+ ¢+ 2 and-ab + bc + ca + 2abc = 1 13

This is a quite strong inequality and it is easy to convince oneself that the inequality becomes
most applications of classical techniques fail. However, the following smart T+ R >_2rs
application of Holder’s inequality does the job: Vi 2
or (R+7)RV3 > 2rs. This splits trivially into R +r > 3r (Euler’s inequality)
(Z L )2-(Za(b+c)?)>(z(b+5))3» and s < 332@1{, the last one being well-known and easy. O
Here is yet another easy application of problem 7.
s0 it is enough to prove that
10. Let u,v,w > ( be positive real numbers such that
3[Je+b) 223 alb+c)
u+v+w+ Vuow =4,
This reduces after expanding to Y- a(b— ¢)? > 0, which is clear. |} Prove that
Proof. First, we get rid of those nasty square roots, via the substitution

Ty =4, yz=4d®, 2z = 4b%

China TST 2007
Then Proof. To get 1id of those nasty square roots, let us perform the substitution

fan fow Juw

V“:c’ \/—:a, v_:b_
and replacing these values in the inequality yields the equivalent form w u v
3a+ b+ e > 16(a + be)(b + ca)(c + ab). Then u = be,v = ca and w = ab, so that the inequality becomes @ + b + ¢ >
ab+be+ ca and the hypothesis is ab+
bc + ca + abe = 4. Another substitution
The hypothesis becomes a2 + 5% 4 ¢? + 2abe = 1, so that there exists an acute-
angled triangle ABC such that ¢ = cos4, b = cosB, ¢ = cosC. Next,
observe that
yields zyz = 2 +y +2+2. We need to prove that zy +yz
+ 22 > 2z +y+ z),
c+ab=cosC+cos
A cos B=—cos{A+ B)+cosA-cosB=sinA-sinB. which is the first component of problem 7. [m]
Using this (and similar identities obtained by permuting the variables), the The following problem has some common points with problem 7 and one
desired inequality becomes can actually deduce it from that result. But the proof is not formal.

fiZcosA Z4HsinA. 11. Prove that ifa,b,(:>01mda::a+%,y:b+l,z:c+1,then

c a
Using the well-known identities
Ty +yztzr>2Az+y+z).
§ :4RHCOS§, r:4RHsin§, ZcosA: 1 +%, Vasile Cartoaje
14 Chapter 1. Some Useful Substitutions 1.2. The relations abc = a+ b+ c+2 and ab + bc+ ca + 2abc = 1 15

Proof. The method is the same as the one used in problem 8. The starting 12. Prove that for all a,b,¢ > 0,
point is the observation that zyz > 2+ = + ¥ + 2. Indeed,
(bte—a)® (c+a-b? (at+b—c)? 3
1 111 :
zyz = abc + tatbtct-+r4+->24z+y+a. (b+e)?+a® (c+a)+0 (a+b)?+c2 5
abe a b c
Let us assume for sake of contradiction that zy + yz + 22 < 2(z +y + z) and Japan 1997
let us choose r € (0,1] such that X = rz,Y = ry,Z = rz satisfy XY Z =
2+ X +Y + Z. This equality is equivalent to 73y, 2+r(z+y+2z) and such Proof. With the usual substitution
7 exists by continuity of the function f(r) = r¥zyz~2~r(z+y+2) and by the
fact that f(1) > 0 and f(0) < 0. The hypothesis zy + y2 + 22 < 2(z +y + 2) 2= b+c . a+ec a+b
can also be written as a Y b 4
the problem asks to prove that for any positive numbers z,y.z such that
XY +YZAZX <20(X +Y+ Z2)<2AX +Y + Z).
zyz =1 +y+ 2+ 2 we have
This contradicts the first inequality in problem 7. 0
-1 y-1? -17
Proof. As in the previcus proof, we obtain that 2yz > z +y+ 2+ 2. Since we z?+1 yr+1 2+1
obviously have min(zy,yz, 2z) > 1, this implies that z > %fl and similar
inequalities obtained by permuting the variables. Next, we have Applying the Cauchy-Schwarz inequality {or what is also called Titu’s lemma},
we obtain the bound
. 1 1 1
m+y+z:a+a-+b + b3 +c+ ~>6.
e
@-12
5
-1 (-1 >
(@+y+z-3? .
2241 ¥+l Z2+1 T2 +y7+2243
In particular, there are two numbers, say .y, such that z +y > 4. The
inequality to be proved is equivalent to z > %f—pj%‘f:zfl, so we are done if we It remains to prove that the last quantity is at least 35 . Let S=z+y+zand
can prove that P = xy + yz + zz, so that the inequality becomes
24z +y . 2z +y) -y
zy-1 7 z+y-2 (-3 3
_— >
But this is equivalent, after an easy computation {in which it is convenient S$2-2P+3 75
todenote S =z +yand P = ay), to (ay —y —z)? > (z ~ Dy - 2). If
Notice that as § + £ > 2 and cyclic permutations of this, we have § > 6. Also,
(z - 2)(y ~ 2) < 0, this is clear; otherwise z,y > 2 as 2 +y > 4. Let
by problem 7 we have P > 2S. Therefore,
u =g - 2,v =y~ 2, then the inequality becomes (uv +u +v)? > wv, with
u,v > 0 and it is obvious. i} .(S a23) N (S 332
:«‘x) _S _ 3:17 2 .
. O

o
The form of the following inequality strongly suggests the use of the §2-2P+37 §2-45+3 S-1 5§-175
Cauchy-Schwarz inequality. It turns out that this approach works, but in Remark 1.3. There are other solutions for this problem: some of them are
a rather indirect and mysterious way, which makes the problem rather hard. shorter, but not easy to find. Here is a particularly elegant one, based on
16 Chapter 1. Some Useful Substitutions 1.2, The relations abc = a+ b+ c+ 2 and ab + be+ ca + 2abc = 1 17

the linearization technique: the inequality is homogeneous, so we may assume Letting ¢ -+ 0, we deduce that any & as in the statement must satisfy
that a -+ b~ ¢ = 1. We need to prove that
K(k+1) 2 (k+%)3,
which is equivalent to 4k* + 2k > 1. We claim that any such & is a solution of
the problem. )
(ta)? . X
—‘l—l{f)—g by an affine function of a, suitably Pick such k and perform the usual substitution
The point is to bound from below (T~aytrat
chosen. The best choice is the following e a . b . c
J c+a a+b
(1~ 2a)® 69 bda 1 2 1
P
T+
> =
725
= -
ag)
=1
\*tg) =>0 to reduce the problem fo

Well, it FEape AN .
S+ By 4 2) + R(zy fyz +o2x) b ayz > (k+§)
Of course, the question is how one came up with something ke this,
is actually very easy: we need constants A, B such that
Now, we know that 2y + yz + zz + 2zy2 = 1 and it is by now a classical fact
(1~ 2a)? (use problem 7 for 1/x,1/y,1/z) that = + y + > 2(zy + y= + 2z). Thus, it
(T~ o T > Aa+B
is enough to ensure that ’

for all @ € [0,1], with equality for ¢ = Tmposing also the vanishing of the K ) QR 4 Ry _ B. )
by 4 2a) ¢ L (k 1)2/ s "
rields the -
derivative of the difference between the left and right-hand si
very easy
desired constants A, B. Once we have the previous This can be rewritten as
a+b+c = 1.
to conclude by adding them and taking into account that 1 3
( 2k2+k—~)5 (ac ytyztzz—
tyztzz—— 220
>
13. Find all real numbers % with the following property: for all positive
numbers a, b, ¢ the following inequality holds But the last inequality follows from the fact that 2k + & — £ > 0 and

a b ¢ 1\* Ty +yztzz > Z’

4 ¢ ; >lk+ . <]
(k i b—re) <k+c+a) (k+a+b> - <L+2>
which, by returning to the substitution
Vietnam TST 2009 _a b
x
e Y T etal a+b
Proof. Take first of all @ = - 1 and arbitrary ¢ to get that
is equivalent to Y- a(b—c)? 2 0.
In '(-mw,lusion‘ the solutions of the problem are the real numbers k such
(m %1)2 (i %) > (m%)x. that 4k% + 2k > 1. ]
 Chapter 1. Some Useful Substitutions 1.2, The relations abc = a + b+ c+ 2 and ab + be + ca + 2abe = 1 19
18

the ",l[‘he crucial claim is that under the previous hypothesis, the system
We end this section with a very challenging problem, which answers has
be the analogue of the classical sub- so%nt.mns whose unknowns x; are positive. Probably the easiest way to prove
following natural question: what would
the this is to exhibit such a solution. Well, simply take z; = 1, )
stitution x = Q;’i Y=ye o, =% © ,yb when we have five variables? We warn
st solution is really not natural. ..
reader that the
_lratatatan
1+ as + agas + aszas
14. Let ay,ag, ..., a5 be positive real numbers such that
and the define
arag.as = ay(l 4 ag) +az(l +as) +--- + as(l+a) +2. 2y = :n4+15; o = T3+ T
az ag
What is the least possible value of al} + aiz 4+ %"
The question is, of course, how on earth did we choose the value of 757 Well.
Gabriel Dospinescu, Mathematical Reflections simply by solving the system, as indicated in the beginning of the‘ ;‘C;lllli(,‘;n‘
Note that we clearly have z; > 0 and an easy computation shows that thm(‘\
are solutions of the system (we only have to check two equations, since tllr;
Proof. Consider the linear system
are satisfied by construction). ’ -
g b 4y = @yF1, £y L4 = a2 T4 + x5 = a3rs, The conclusion is that if the e;’s satisfy the condition
5 5
25+ I = GaT4, Ty T2 = A5T5.
=24 a; + ayas + asaz + azas + asas + aza;,
all variables in i=1 =1
We can try to solve this system by expressing, for example,
Namely, we can use the fifth, first and second equation to then we can find z; > 0 such that
terms of xy, 75
in terms of z1,2s. Replacing the obtained values in the
eXpress Ty, al:»’Cz*Ia M713+x4 a Ty + 2o
obtain that the s = cenan = .
other two equations and eliminating z1, a5 between them, we zy E2S x5
system has a nontrivial solution if and only if
So, the problem reduces to finding the minimal value of
5 xy za x5
[1e:
41
+ L a3 -+ @10g
=l
+ 0407 + 0265 + ase3 + a3y zotaz
+
w3
+o
+11+Iz
.
Wo claim that the previous expression is always at least 5/2. By Titu's lemma.
they
All the previous (painfull) computations are left to the reader, since this reduces to proving that ’
can be
are far from having anything conceptual. Note that the same result 5
the
obtained by computing the determinant of the associat ed matrix. Now,
(w13 + 222 + 25+ 24+ 25)% 2 > 5 ;q z:3;.
a permutation
previous relation is almost the one given in the staterment, up to
, we
of the variables. So, since the conclusion is symmetric in the five variables But since
instead of the
will assume from now on that the previous relation is satisfied
one given in the statement. D> =(Tw) -y,
20 Chapter 1. Some Useful Substitutions 1.8. The relation a® + b? + ¢* + 2abe = 1 21

53" #¢, which is Cauchy-Schwarz. Using this lemma for the inverses of the a;’s and denoting
this reduces to (3 @)% <
Putting everything together, we obtain that the minimal value of .11 1
S=—4
=4 +—,
1 1 a ap as
S —1
a ag a5 we deduce that
is 5/2, attained when all a;'s are equal to 2. ‘Who wants to play now the same
O %> 752 25M1_,
=
game with 7 variables? @i0i+10i4+2 ayaz:--as

Proaf. Here is another proof, also far from being evident. .. We will use the
On the other hand, by Mac-Laurin’s and AM-GM inequalities we also have
following nontrivial
555 ayt+az+---+as
Lemma 1.4. For all nonnegative real numbers x1, %, ... a5 we have S° > 4> 125
apay” 5 aiaz---as
(w1 + 22+ 25)°2 25(wyz0ms + maaTy + X5 T2).
Taking into account these inequalities and the hypothesis, we deduce that

Proof. This is not easy at all, but here is a very elegant (but unnatura
l) proof: st g8 955
consider the identity iyt 2!
ayEes + TaTaLy + L b E4E5TL F TRTIXD which immediately implies that § > g O
(a3 + z3){(@2 + 24) -+ wmpralzn + 71—
minz; (since the inequality is cyc! so that 1.3 The relation o® + ¥ + ¢ + 2abc = 1
We may assume that s =
> 0. DPnotmg 21 4 o + 23 + x4 = 4t and vsing the AM GM
T b kg — The following problem is a rather tricky application of the AM-GM in-
mcquahw we deduce that equality.

xs(@1 + z3) (2 + ©a) + Baxs(ay 4+ 24— ws) < 4 15. Prove that in all acute-angled triangles the following inequality holds

cosC
2
Thus, it Temains to prove that ) +8cos Acos BeosC > 4.
cos A
5 4t — xg
4ty + ( 2 Titu Andreescu, MOSP 2000

By homogeneity, we may assume that 25 =1 and then cqunding everything Proof. Since
the inequality becomes, with the substitution 4t —1 = 3z, (z—1) Y(8z+T) 20,
] cos? A+ cos® B+ cos? C + 2cos Acos Beos C = 1,
which is clear.
22 Chapter 1. Some Useful Substitutions 1.3. The relation a® + b + ¢ + 2abe = 1 23

the inequality can be written for some positive real z,y.2z, we obtain the relation zyz =z +y + 2 + 2 from
the al identity 3 cos? A+2]]cos A = 1. Thus we can find positive real
numbers a, b, ¢ such that
cos A? cos B\? cosC\? 3
(CosB) + (COSC) + (cosA) 22 4{cos?
4(cos” AA++ cos”
cos® BB + cos” 2 C) C). b+ec cta a+b
T = s Y= y 2= .
a b c
Let
The desired inequality can be written
recoss A, y=cos’B, z=cos:C,

s0 we need to prove that ZEE}*Z 1xygg<=)2x+z\/x—ysgzyz

r oy %
—+y+z+m_
=+ - >4z(x4 y+ - 2) Z).
=Y a+y V< g(1+y+z+2)4:> (2fi)252(21+3).
The point is that we actually have This follows from Canchy-Schwarz

(Zfi)zg(Z(b+c))~(2%):2zz+6. [u)
Proof. Since the triangle is acute, there are positive numbers z,y, z such that
for any positive real numbers z,y, z, as follows by adding up the AM-GM
inequalities a2=y+zA b2:z+:r, cz=x+y.

Then
x
cos
A =
ViE+y)z+2)
Thus, it is enough to prove that zyz < é‘ But this is well-known and
easy to prove. [} and similar identities for cos B,cosC. The desired inequality can be written
as
The following problem is a preparation for a hard problem to come, but
has also an independent interest. — z2 4z
DB e KD Dy
oy s £

[
16. Prove that in every acute-angled triangle ABC,
which (after clearing denominators) is equivalent to
(cos A + cos B)? + (cos B + cos C)? 4+ (cos C + cos A)? < 3.
Y (P+ 2+ VETAETY) € St DA +a)
Proof. Leiting and then to

Z yz{z +y)-yz{z+ 2) < 3zyz+ %Zyz(yp} z).

eye e
24 Chapter 1. Some Useful Substitutions 1.3, The relation a® +b* + ¢? + 2abc = 1 25

However, this follows immediately from the AM-GM inequality: Thus, it is enough to prove that ¢ + be < 2. But the given condition and the
fact that a is nonnegative imply that
Ve ) -yl
oo het VI -9(F -1
=
and two similar inequalities. [}

Even though the following problem seems classical, it is actually rather . it is enough to prove that /(62 — 4){c? — 4) < 4 — bc. Note that be < 4,
hard. Fortunately, we did the difficult job in the previous problem. We > b? < 4 and ¢? < 4. Squaring the previous inequality thus yields an
also present an independent and very elegant approach due to Oaz Nair and equivalent one
Richard Stong. B2 — 46 — 4¢* + 16 < b — 8be+ 16,
which is obvious. [m]
17. Prove that if a, b, ¢ > 0 satisfy a® + b% - ¢ + abe = 4, then

0 <ab+bc+ca—abe <2 some tricky algebraic manipulations to the infamous “Iran 1996 inequality.”

Titu Andreescu, USAMO 2001 18. Prove that in all acute-angled triangles the following inequality holds

Proof. The inequality on the left is easy: the hypothesis and the AM-GM
R252 42, 2
eo + e+~
o + 2— >2 IR’ QRZ.
inequality imply that abe < 1, so that min{a, b, ¢) < 1. We may me that
¢ = minfa, b, ¢}, so that Nguyen Son Ha
ab -+ bedca—abe=cla+b)+ab(l —c) = 0. Proaf. Tt ,C are the angles of the triangle, the sine law and the identity
sin x + cos x = 1 yield the equivalent form of the inequality
The hard point is proving that ab-be+ ca —abe < 2. Taking into account
the hypothesis, this can also be written as Z(l—coszA)(l—cuszB) 29
1—cos?C =4

(57 () + (57 =2
a+b\?
i,
bte)? a+e)\?
< 3. Write
cos’ A=yz, cos?B=z2, cos’C= Ty,
If we denote a = 2x.b = 2y,¢ = 2z then 22 + 4% + 2% + 2zyz = 1 and so
Then +yz+ zz + 2zyz = 1 and so there exist positive numbers X, Y. Z
there exists a triangle ABC such that « s(A),y = cos(B),z = cos{C).
such that
Thus, the problem reduces to the previous one (since a, b, ¢ are nonnegative,
the triangle ABC is acute angled). )

Proof. Here is a very elegant proof for the hard part of the inequality. Among
the three munbers a—1,b~1, ¢~ 1, two have the same sign, say b—1 and e~ 1.
Thus (1~ b)(1 — ¢) > 0, implying that b+ ¢ < be+ 1 and b+ ac — abe < a. 9b3
26 Chapter 1. Some Useful Substitutions 1.4. Notes 27

On the other hand, we have Using the identities

abtbetea 1 cla+b) 2ab+be+ca) 2¢2

4ab
=2+
4 4ab
. (a+c)b+c)
—2- .
(a+c)b+c)
$0 that the inequality is equivalent to this becomes
cla+b) N 202
XY(X+Y+2) > 9 dab T {a+c)b+c)
ZX+Y)?: T4 But this simply the standard inequality (a + b){a + c}(b+¢) > 8abe and we
are done.
This can also be written in the form
1 9 This completes the solution to the problem. 0
XYZ(X+Y~}~Z)-E‘<WEZ
1.4 Notes
and it is a consequence of the following famous inequality

Lemma 1.5. For ol positive numbers a. b, ¢ we have We would like to thank the following people for their solutions: Vo Quoc
Ba Can (problems 13, 18), Xiangyi Huang (problems 4, 5), Logeswaran La-
ot1 11 1
s 29 janugen {problem 1, 9}, Oaz Nair (problem 17), Dusan Sobot {problems 2, 9,
m"””“‘”((bfl T “(mb)?)w 16}, Richard Stong (problems 8, 17), Gjergji Zaimi (problems 2, 3, 6, 7, 8, 11,
12, 15, 16, 17).
Proof. We may assume that ¢ > b > c. First, we show that

1 1 1 1 2
2 — .
(a+0)? * (a+c)? + (b+c)? ™ 4ab * (a+c)(b+¢)
This can be rewritten

o1 2 (a—b)?
atc b+e) T dabla+b)?’

or equivalently dab{a+b)? > (a-+¢)?(b-+¢)%. This s clear, as (6 4+)* > (a+¢)?

and 4ab 2 (b+ c)%
Thus, it remains to prove that

1 2 9
(ab + be + ca) [m E W (3“(:7)} > e
Chapter 2

Always Cauchy-Schwarz. . .

As the title suggests, all problems in this chapter can be solved using the
Cauchy-Schwarz inequality, even though sometimes this will require quite a
lot of work. Let us recall the statement of the Cauchy-Schwarz inequality: if
@1.a2,...,a, and by, by, ... b, are real numbers, then

(@ +ad+ o+ ad) R+ 8+ +82) > (arby +azhy + -+ anba)?.

This follows easily from the fact that 3.1 @iz +5;) > 0 for all real numbers
z. Indeed, the left-hand side is a quadratic function of = with nonnegative
values, so its discriminant is negative or 0. But this is precisely the content of
the Cauchy-Schwarz inequality. Another proof is based on Lagrange’s identity

(zaz)-(zbz)—(zaim) S51<i<i<n (b
n n 3 2

=1 i=1 =1

This useful identity will be used several times in this chapter.

As the best way to get familiar with this inequality is via a lot of examples
of all levels of difficulty, we will not insist on any theoretical aspects and go
directly to battle. We start with two easy examples, destined to give the
reader some confidence. He will surely need it for the more difficult problems
to come. . .
30 Chapter 2. Always Cauchy-Schwarz. .. 31

1. Let a, b, ¢ be nonnegative real numbers. Prove that Proof. The inequality being symmetric in b, ¢, but not in a, it is natural to
deal first with /&~ 1+ +/c — 1. This is easy to bound using Cauchy-Schwarz:
(az® + bz + ¢)(cr® + br + a) 2 (a + b + )*z?
VE—i+4ve-1<
-1+ D1 +ec-1) = Vi
for all nonnegative real numbers 2.
So, it is enough to prove that vhe + va -1 < v/a{bc+ 1}. But this is once
Titu Andre :u, Gazeta Matematica
more the Cauchy-Schwarz inequality. =)
Proof. This is just a matter of re-arranging terms and applying Cauchy- Another easy, but a bit exotic application of the Cauchy-Schwarz inequal-
Schwarz: ity is the following Chinese olympiad problem.

(ax® + bz + ¢)(cx? + be + a) = (az® + bz + ¢)(a+ bz + cx?) 4. Let n be a positive integer. Find the number of ordered n-tuples of
> {ax + bz + cz)? integers (a1,as,...,a,) such that
=(a+b+ )t aytaz+t+an>niandal +ad+ - +al <+l
O
China 20602
2. Let p be a polynomial with positive real coefficients. Prove that if proof By the Cauchy-Schwarz inequality we have
1 1. L
Pl = 2 — is true for z = 1, then it is true for all z > 0.
z) "~ plx) aitaytota, < VR <nl+ 1
Titu Andreescu, Revista Matematicd Timisoara
Since a; are integers and a3 +az + -+ + a, > n%, we must have
Proof. Write p(X)} = ag + a1X + -+« + @, X™ and observe that a1+a2+~--+a,,=n2.

1 3 a, But then (again by Cauchy-Schwarz) we have af + a3 +--- +a2 > n®, forcing
plz)p <;) = {ag + a1z+ -+ F ane™) ((1(, + ;! +oet 7)
al+ad+--+al € {ndnf+1} fad +ad+---+a2 = n, then
we must have
> (ag+ay 4+ 4 ag)? equality in Cauchy-Schwarz. implying that all a;’s are equal to n. Assume

=p(1)%. now that af +af + -+ a2 = n® + 1 and letb; = a; — n. Then

B+t =n'+1-2n-n2+n®=1,
The result follows. o

The following exercise is already a bit trickier, due to the lack of symmetry. forcing all but ouc b; vanish. This is however impossible, as by +by+- - +b, = 0.
Therefore, this second case will not occur and the only solution is
3. Prove that for all real numbers a,b, ¢ > 1 the following inequality holds: ar=ay=-=a,=n. 0
a~1+vVh—1+ve—1</albe+1). We continue the series of easy exercises:
32 Chapter 2. Always Cauchy-Schwarz. .. 33

5. Let 21,29, .... @19 be real numbers between 0 and 3 such that The equation

sin?z; +sin’mg 4+ AN/

(mt-2->
BY?
+(lmt§) +(3cot;)
C\? o=(fi)
res\?
Prove that
can also be rewritten as
3(sinzy +sinag + - +sinzig) < ¢ §41 + Ccosmy + - -+ COS T,
22
24V 2 0%2t (Seryt 2 T
Saint Petersburg 2001 2 Tt r2 7r

Proof. The Cauchy-Schwarz inequality and the hypothesis yield This simplifies to

Cos iy = sin® & -+ sin? 2 R sin’ Zio 49(2% + 49% + 927%) = 36(z + y + 2)%.
1, 5 .
> §(Sm g 4 sinxg + -+ sinzig) Now, by Cauchy-Schwarz we have

and similarly for the other variables. The result follows by adding up these (l+§+%) (zz+4f+91'2)2(1:+y+z)2,
inequalities. . )

The following exercise combines an easy application of the Cauchy- vielding

Schwarz inequality with some classical formulae from geometry. Recall that v 49(2% + 49* + 927) > 36(z + y+2)%
is the inradius and that s is the semi-perimeter of a triangle.
Thus, we are in the equality case of the Cauchy-Schwarz inequality, which is
precisely the case when there is a & > 0 such that = = 4k,z = 36k,y = 9.
6. The triangle ABC satisfies
Then the sides of the triangle are 13k, 40k, 45k. Thus the triangle is similar
to the triangle with sidelengths 13,40,45, and the problem is solved. 0
(.02ot 2 2+ 2c0t
(,()28 L+ 3ot & ey
(LOQ = [ .

The statement of the following problem looks rather classical. There are
Show that ABC is similar to a triangle whose sides are integers, and however some technical problems which make the problem more difficult than
find the smallest set of such integers. expected.

Titu Andreescu, USAMO 2002

7. Let n > 2 be an even integer. We consider all polynomials of the form
2"+ ap 12" 4 -+ @z + 1, with real coefficients and having at least
Proof. Let the incircle be tangent to the sides of the triangle at points which
one real zero. Determine the least possible value of af + a3 +---+a2_,.
split the sides into segments of length & and y, « and z, and y and z. Thus,
the sides of the triangle have lengths « +y, z + 2, y + 2. Czech-Polish-Slovak Competition 2002
34 Chapter 2. Always Cauchy-Schwarz. . . 35

Proof. Suppose that the corresponding polynomial has a real zero . Using The denominators in the following inequality look awful, but a clever
Cauchy-Schwarz, we obtain application of the Cauchy-Schwarz inequality can make all of them equal.
The method of proof is worth remembering, since it appears quite often.
(@ + 12 = (e +aga? + o+ a1z
<(ad+adt e+ ado )@+ at 4o 422D 8. Prove that for any positive real numbers z,y, z such that zyz > 1 the
following inequality holds
Thus ( 2
adrad b hal > e
z2+at
+ o+
2" +1
g2e = @)
3
Bryt+z
+
P24z
Yy 4
Py
z

and so we need to find first the minimal value of f. Now, looking at the
zeros of the derivative of f suggests that the minimal value might be taken at Tuan Le, K6MaL magazine
x = 1, so that it equals ;%7. This is not easy to prove using derivatives, as
the computations are a blt ndsty Instead, we will prove in an elementary way Proof. Using Cauchy-Schwarz, we obtain
that
n—1 Z2n~1) 3 1
(@ + 1P za? 4ot b+ (#+y7+2) sHltz) 2@y ta)2
4
In order to prove this, note that we have the inequalities
Thus
Pl e Felzateat L, 1zt z PEAtEa ]
Btttz (z+y+2)?
Adding them shows that
Writing down similar inequalities for the second and third terms of the left-
Ei-z(x”+1)zxz+z4+»-» a2, hand side, we reduce the problem to proving that

On the other hand, multiplying the last inequality by =™ and adding the result Z(1+z + zz) S(z+y+z)24zbzz2+21y23+z:c.
1o the previous inequality gives the inequality
This is immediate from
?j(x" L R R At REEE =t
Zzy >3 (xy2)2 >3
Thus it remains to prove that (¢ + 1)? > 42®, which is clear.
The previous paragraph shows that and
2o Ery+2)?
Zfz(*%)—z\/%w(myfl)zan. |
a4 +ag >
Here is a rather nice-looking inequality taken from a Romanian Team
if 2 + @ 12" 1 4+ -+ a1z+ 1 has at least one real zero. On the other hand, Selection Test. There are plenty of ways to prove it, but what follows is
choosing a; = ag = = Gp-p = A% shows that Fé‘i is optimal. [m] particularly elegant and natural.
36 Chapter 2. Ahways Couchy-Schwarz. .. 37

9. Let n > 2 and let ay,as,...,a, and by, bo, ..., b, be real numbers such Proof. This is a simple application of the Cauchy-Schwarz inequality. Namely,
that we have '
A taf+ohdd =0 bR b
@ +52 > {c+d+e )? ,(a+b)?
and aiby + agby + - + apby 5= 0. Prove that 2
e >2
Erd =
3
so that
ag)? + (b
(a1 +ag+ -+ by oo+ by)? <
\/(?+b2+<:2+42+e22(a+b)\/§
Cezar and Tudorel Lupu, Romanian TST 2007
Ou the other hand,
Proof. The proof mimics the proof of Cauchy-Schwarz: take any real number
x and apply the Cauchy-Schwarz inequality to obtain
Va+ Vi< Valatd), VerVa+ve< et dte = 3ath),
SRR » SR » T . 2
therefore
PR
#=1
7 (Va+Vh+ e+ Vd+ve) < (V2 +V3)Ha+b).
Comibining these two inequalities yields the estimate
By hypothesis, the left-hand side equals 1 + 22, Therefore

Va2 2+ +d e 6(\[‘/50‘/
> -)2 Va+Vo+
e+ Vd+ Vel
(Z(h% er,) < nfx? +1)
To see that this is optimal. it suffices to keep track of the equalities in the
for any real number . The difference between the right-hand side and the previous inequalities. For instance, we can take ¢ =b=3andc=d—¢ = 2.
left-hand side being a quadratic polynomial which takes nonnegative values Thus the answer is T = Wi}l*;\/?)" [
on the whole real line, its discriminant has to be negative or zero, thus
The mnext problem requires a whole series of applications of Cauchy-
4A°B% < 4{(n — A%)(n ~ B?), Schwarz.

where 4 = % a; and B = 31, b It is immediate to check that this is 11. Let 2y, 23,....2, be positive real numbers such that
equivalent to the desired inequality. a
IT
T4z 14z
S'+1+1,1
The following problem is eas but the lack of symmetry might make it
appear more difficult than it reall .
Prove the inequality
10. Find the largest real number 7' with the following property: if a, b,
are nonnegative real numbers such that a +b = c+d + e, then VI VT2 +\/1_,.>(n—1)(
\/_\/_
V@ T2+ E g d e > T(Va+ Vbt Vet Va+ el Vojtech Jarnik Competition 2002
38 Chapter 2. Always Cauchy-Schwarz. .. 39

Proof. Let a; = 50 that 3 a; = 1 and z; = é-’—f—“-’ Then, using Cauchy- where the first inequality is simply the hypothesis, while the second one is
Schwarz we can write: Cauchy-Schwarz applied to n — 1 terms, grouping the terms indexed 1 and
R — 2 in each multiplicand. We deduce that % < %5, which immediately
faz ozt
- +an V@2 + /a3 + -+ /an
SvE=3yT = /
o 23
>
T . implies that
max(a;,
az) < 4minfa;, ag).
Rearranging terms in the previous sum gives
Since everything is symmetric in @i, @z, - . . , @, the result follows. o
1 1 1 1
. a + R A bit more difficult, but with a similar flavor is the following problem.
V' E Ve (\/ Gz +/as N )
and using again Cauchy-Schwarz we can bound this from below by 13. Let n > 2 and let 1, 3. ...,z be positive real numbers such that

1 Z (n~1*var 1 1 1
(@1 + a2+ +an) (— +—+---+—-) =n’+1
w1 e fag b fag b i Ty T2 In

Using once more Cauchy-Schwarz for the denominators of each fraction Prove that
52 1 1 1 2
Vaz + az + o+ V/a (zi2 2, + x5+
rh2 4+ I")<x§+1§+
Sttt +I%)
5> 2 44—
n+ +"(n7”
vields the desired inequality. ]
Gabriel Dospinescu
The idea of the following example is worth keeping in mind, since it turns
out to be useful in a wide range of problems. Proof. The crucial idea is to write the hypothesis in a different way: expanding
the product and rearranging terms shows that the hypothesis can also be
12. For n > 2 let ay, ag, ... ap, be positive real numbers such that written
[ 1 1\? . 2
(a+ar+-to )| —+—+-+—1<In+35} .
a; a2 an 2 > ( /E_T_ \/i! ) =1
Prove that max({ay,az,....an) < 4minfag, az,. .., an). Wi \Wa Vi
Titu Andreescu, USAMO 2009 Let us write

Proof. The idea is to fix two of the variables, say as, ag and apply the Cauchy-
Schwarz inequality to get tid of the remaining variables. Explicitly, this can Now, expanding again gives us
be written in the form

?)(5=8)2 (
40 Chapter 2. Always Cauchy-Schwarz. . . 41

Since Thus, we need to prove that for any ay,as....,a, > 0 we have

(‘/ &
the inequality to be proved becomes

Define S; = a1 4-az+- - -+a;, with the convention Sy = 0. This

3 Attty >4+ sequence and so
s an increasing
n{n—1)"
1<i<j<n \ )
which is equivalent {taking into account the hypothesis that the sum of all ; ;
- L3 Si-1
is 1) to >—‘(1+a;+ St a)? z:(lfi»S)Q‘
9
Sy ——. Si =S _ ( 1 1 1
1<i<j<n nln—1) Z(wsflu».fiz T+ 8 1+s,-)*1’1+sn<1- o
This follows from Cauchy-Schwarz inequality, but there is still a detail to be Proof. First note that for any nonnegative real numbers ¢ and A we have
explained: we have to show that we cannot have equal y. But if we had
equality, all £; ; would be equal to "(13] and so all numb(*rs + fi would be I+\/E'm
equal. Since 1 > 2, this would force an equality z; = 2, fior some ] # k. But
then #;x = 0, a contradiction. ]
VI AT A+
- Vite Vit AT542
We continue with an inequality which combines a rather direct application TV A2 )
of the Cauchy-Schwarz inequality with a nice telescopic identity. We also
1+¢
present. a beautiful alternate solution, due to Richard Stong.
1+ 427
14. Prove that for any real numbers 2,29, ..., &, the following inequality Here the first inequality is just 1 + A2 + 2% > 1 + A% with equality if and
holds: only if © = 0; the second is Cauchy-Schwarz with equality if and only if
E zy/e= I+ A% Thus we cannot have equality in both cases and the final
vy P
&y o
e A
L1 +
Rl v inequality is actually strict. Applying this inequality repeatedly, it is easy to
see by downwards induction on k that
Bogdan Enescu, IMO Shortlist 2001 E3) . T
1+t L+ai+--+a2
Proof. The solution is very short, but far from obvious. The idea is to use the
Cauchy-Schwarz inequality to reduce the problem to S T n—k
< 5+t +
1+4} L+al+- +af 1+a3+
3 LE Ty,
(14292
—_ " (1+af+ad)?
: ot (I4of+ 422
5 The case k = 0 is the desired result. jm}
42 Chapter 2. Always Cauchy-Schwarz. . .
43

The trick of making the denominators equal thanks to a smart application Proof. Let us write the condition in the form
of Cauchy-Schwarz or AM-GM inequality is also used in the following problem.
23 + (@1 +22)? + (22 +23)2 4+ + (Tno1 +a)? a2 =2
15. Prove that if a, b, ¢, d are positive real numbers, then
I we fix 1 <k < n and apply the Cauchy-Schwarz inequality twice, we obtain
a " b . c " d N 4
P +d? @+ d @ dr @R+ atbtetd 23+ (@1 4 w2)? oo (e )
S @@t ze) + (@2 v a) — - (D (s ay))? _ T
P.K. Hung = k Tk
and
Proof. Using the AM-GM inequality in the form (x + y)? > dzy, we obtain
(g + To1) 2+ + (Tt + @) + 22
a . 4a®
Bt 4 d2 T (a2 + 82 4 dR)2 S (et o)— Gen t o)
o -2t 22
- n—k+1 Tn—k+1
Adding these inequalities yields Taking into account the hypothesis and these two inequalities, we deduce
that 2 > EG:%Tf for all &, so that
Xp 2k(n+1-k)
By Cauchy-Schwarz, [z < ¢/ n+1 )
2
z a3 (X% By studying the equality case in the previous inequalities, one immediately
—_—
= YNa sees tha value is attainable for each fixed k. Specifically, if we define
Inserting this in the previous inequality yields the desired result (note that =) and take g = Zns1 = 0 by convention, then equality occurs
the inequality is strict, since otherwise the equality in the Cauchy-Schwarz if (~1)F 7z; interpolates linearly and evenly between these endpoints and
inequality would yield @ = b = ¢ = d, for which the inequality is strict}. Note Tz = ag. The explicit formula is
that even though the inequalities we used were very rough, the constant 4 is
optimal: simply take a = b and ¢, d close to 0. i T = ; -
(—1)F-ide i<k
.
One needs rather good gymnastics with Cauchy-Schwarz to deal with the
PTG ik 0

following problem. Another ingenious application of the Cauchy-Schwarz inequality can be

found in the following problem.
16. Let n > 2 be an integer and let z1,23,.. ., %, be real numbers satisfying
17. Let a,b, ¢ be positive real numbers. Prove that
mf—kzg+--- ‘(»xf,-kmwg +aprg b Tpe1da = L
1,1 11 7(1 1 1 1 2
Stmta t————s2 (i
For a fixed 1 < k < n, find the maximum value that {z;] can take. a? B2 (a+b+c)2_25(a+b+c ta+b+c)
China 1998 Iran 2010
44 Chapter 2. Always Cauchy-Schwarz. .. 45

Proof. We see that we have equality when a = b = ¢, so we have to apply Another application of Cauchy-Schwarz thus yields
the Cauchy-Schwarz inequality in a smart way for the left-hand side of the
inequality. Narmely, start with zsin A +ysin B + zsinC' < V/(] +y?)(a? + 22 + 2wz cos B + sin® B),
Thus, it remains to prove that
1 1
(1+1+1+§>(— 22+ 2% 4 222cos
B+ 1~ cos® B < 1+1'2+22+x222,
_+(a+b§u7>
1 2 which is equivalent to (zz — cos B)? > 0. [m}
7+3(a+b+(.‘)> : We now enter the zone of challenging problems, with an unusual one.
The first solution seems to come from nowhere (well, it actually came from
the author’s imagination. .. ), but we also present a beautiful geometric proof
of Richard Stong, which makes things rather clear.

19. Let a.b,¢,z.5, 2 be real numbers and let

A=ar+by+cz, B=ay+bzter, C=az+br+cy.

Assuming that min(|4 ~ B|,|B - C],iC — Al) = 1, find the smallest
and we are done. )
possible value of (a? + 82 + ) (a2 + 42 + 22).
Adrian Zahariue, Mathematical Reflections
The form of the following problem strongly suggests using Cauchy-
Schwarz. However, it is rather easy to check that many atterpts fail. .. Proof. Note that by Lagrange's identity and the Cauchy-Schwarz inequality
we can write
18. Let x,y,z be real numbers and let A, B,C be the angles of a triangle:
Prove that (@ + 0+ )2+ 2 + 2%
= {az + by + c2)? + (ay - bz)? + (bz — )’ + (ex —az)?
wsin A +ysin B+ zsinC < mm
S A4 {ay —bz + bz — oy + cx — az)?
Proof. Using that ¢ = r — A — B, we obtain the identity - 3

g B-C
IB-CF
zsin A+ ysinB 4+ zsinC = (2 + zcos B)sinA + zsin Beos A + ysin B. + 3

Using Cauchy-Schwarz and the fact that sin® A + cos® A = 1, we obtain Since everything is symmetric, we obtain

(a® + 8 + A a? + 7 + )
(z+ zcos B)sin A+ zsin Beos 4 < \/ms B)? + 7 n’ B
2n{m(AszlB;?!
—CcrR —
132+,(73Ai A2
e — B2
3B|)‘
s /2% 4 22 + 202 08 B.
46 Chapter 2. Always Cauchy-Schwarz. .. 47

Using the hypothesis, it is immediate to check that the last quantity is at A+ B+C = 0. Recall that cos a + cos(e
+ 27 /3) + cos(a +47/3) = 0 for any
least % a. I we now take w to be the angle between u and v, then we compute
To see that the answer of the problem is %, it remains to find a 6-tuple
(a,b,¢,z,y, 2) satisfying the conditions of the problem and for which A%+ B? 4 O = ||u)l? - vl }cos® w + cos?(w + 27/3) + cos*(w + 4rn/3))
1
P P 4 éfiluhz - HulP(8 + cos 2w + cos 2w + 27/3) + cos 2(w + 4n/3))
(@ 4+ 6%+ D)2+ + 25 = 3

i
Sl
3
- ol

il
Taking 4 = 1, B = 0,C = ~1 (this is a triple which minimizes
2 a4 _pe The conditions A+ B+ C = 0 and min(|4A — B|,|B-C|,|C —A}) > 1
max(AQ-}-‘B 3C[ A,BZ+$C 3 A CA+M 383 > imply that A%+ B2+ C? > 2 (without loss of generality, assume that 4 and B
are nonnegative. As |A— B| > 1, we have max(A4, B) > 1,s0 A2+ B2+ (% =
under the restrictions of the problem), we must ensure that we have equality 24%+2B% + 2AB > 2max(4, B)? > 2) and so
in the Cauchy-Schwarz inequality. Solving the corresponding system yields,
after some tedious but easy work, suitable values for a,b, ¢, 7,y, 2, namely fuall? - fol® = (@ + 8 + ) a? + 4% + 2%) > §
] It is easy to see from the proof above that equality is attained. Simply choose
any (z,y,2) with 2 + y + 2 = 0 and choose {a,b,¢) with a + b+ ¢ = 0,
Proof. Let u be the column vector with entries (a,b,¢), v the column vector orthogonal to (z,y. z) so that A =0, and scaled so that B = 1. For example,
with entries (z,y,z) and let R be the linear map taking (z,y,2) = (1, ~2,1) we see that {a.b,c) = (—1,0,1)/3 suffices. o

x g 10 T Y ‘We present three short solutions for the following problem. However, none
Riyl=10 0 1 yl=1z]. of them is really easy or natural.
z 1 6 g z x
20. Let a,b,c¢,d be real numbers such that
Then A = uTv, B =u" Ry, C = v R*» and we want to minimize |ulf® - [[v}%.
(@ + 1) + (S + 1)(@® +1) = 16.
Note that R acts on R? by fixing the line z = y = z and rotating 120° in the
orthogonal plane z + y + 2 = 0. Prove that
Write % = uy + ug, where u; lies on the line z = y = 2z and ug lies in the
plane z +y + z = 0 and similarly for v. Then 3<ab+be+cd+da+ac+bd—abed
< 5.

el = Hurf]? + [fuzli® and wT B*v = wo; + ul R*vy. Titu Andreescu, Gabriel Dospinescu

Thus the contributions of u; and vy to A, B,C are all the same (hence cancel Proof. Write the inequality in the form
in {A ~ B, etc.). Thus clearly the minimum occurs when u; = v; = 0. In
this case, R acts as a 120° rotation on v, hence v + Rv + R%v = 0 and thus (ab+ be + od + da + ac + bd — abed — 1)? < 16.
48 Chapter 2. Always Cauchy-Schwarz. .. 49

Next, apply Cauchy-Schwarz in the form from where the conclusion follows, as the desired inequality can be written as

(a(b+ ¢+ d — bed) + be + bd + ed ~ 1)? [1=3" ab+abedi < 4.

<@+ 1) [(b+c+d~bed)?
+ (be+bd +cd—1)?].
Straight from the Book, isn’t it? o
The miracle is that we actually have
‘We continue with a really nice, but rather technical problem. It re-
(b+c+d = bed)? + (be +bd+ ed — 1)? = (B + 1)(c* + 1)(d® +1). quires some delicate algebraic computations and a rather exotic application
of Cauchy-Schwarz. We also present a more conceptual proof, due to Richard
Of course, this can be checked by brute force, but perhaps nicer is determining Stong, which uses more advanced tools, but proves much more.
it through two applications of Lagrange’s identity. a
21. Let a,b, ¢, d, e be nonnegative real numbers such that a2 +52+4¢? = d2+¢?
Proof. We will use Lagrange’s identity and Cauchy-Schwarz in the form and @® + 4 + ¢ =d*
+ €', Prove that @® + 83 + 3 < d® + €%,
IMC 2006
T+ 1) = [ +52+ (1 = ab)*] [(c+ @) + (1 — cd)’]
zla+b)e+d) — (1 —ab)(l—ed)?. Proof. Since we only deal with even exponents in the hypothesis, let us square
the desired inequality and write it in the form
Note that
Za5+22a363§d6+e6+2d333.
{a+b}{e+d) — (1 - ab)(1 —cd) = —1+ ab+ bc + cd + da + ac + bd — abed.
Next, the identity
A moment of thought shows that what we have just proved is exactly the
desired inequality. =
Yo - sare = L (S a?) (32a4 -(Z4))
Proof. Here is a rather unnatural, but very elegant proof. Consider the poly-
nomial combined with the hypothesis easily yield
P(X) = (X —a}(X = b)(X —e}(X ~d)
a®+ 5%+ F = 3% + & + 8.
and observe that the hypothesis becomes P{i)P(~i) = 16. This can also be
written as |P(4)[? = 16. On the other hand, we have Thus, we need to prove that

PG =13 b abed+5 (Y a— 3 abe). 23" % + 3a%Pc? < 2(de)® = 2(a%? + B2 + Fa?)¥2.

We deduce that ‘With the obvicus substitutions, this is equivalent to the inequality

(1= Y ab+ abed? + (Y a— Yabe) =16, 2 Zx3 +3zyz <2 (212)3/2 B

50 Chapter 2. Always Cauchy-Schwarz. .. 51

Using Cauchy-Schwarz Apply this to z = a%, y = b*, z = ¢, and f(u) = «*/2. This amounts to
choosing z,y, z to be the roots of p(t) = 2 — (d° + €2)t? + d2e®t — 3. Since
(229:3 + Bayz)? = (Zz(Zmz + yz))2 < (217) (Z(?mz + yz)2> s (¢ < Ofor all { > 0, we see that g is a decreasing function of o for o3 > 0.
Thus g(0) > g(o3) which, unwinding definitions, is the desired inequality.
it remains to prove that This argument shows that the same inequality holds for any exponent
strictly between 2 and 4 and gives similar (sometimes reversed) inequalities
2(212 +yz)t <4 (212)2, for other exponents. [

which follows immediately from z%y2 -+ 4222 + z%2% > ayz{z + y + 2), itself The reader will probably appreciate the beauty of the following inequality.
equivalent to Y (xy — x2)* > 0. 1t is more difficult than it appears at first sight, as the obvious application of
Cauchy-Schwarz fails rather badly.
Proof. Consider the polynomial p(t) = % — 71t2 + oot — o3 with o7 and a2
fixed and o3 allowed to vary. Regard the roots @, %, z of p(t) as functions of 3. 22. Prove that for any real numbers z;,25,...,
2, the following inequality
Suppose f : R —+ R is any smooth function (at least three times differentiable holds '
sion below). Define a function

glow) = f(z)+ fly) + f(2). 2D znjfjlxi-z,-fi).

i=13=1
Then g is a differentiable function of o3 and

g'loy) =
@ rw) 1 1w IMO 2003
Pylrams v Rl s e R M Proof. The first step is to order the z;s, say z; < 23 < --- < Zn. Then
for some ¢ with min(z,y,2) < { < max(z,y,2z). To prove the first equality
one merely checks that the curve Z s - 25 QZ(IJ' —zi)=2~(n— Dz — (n— a2+ -+ (n— Vza).
i Ed 8
ig=1 i<y
o) = (“’ (g Ty S s ey AL e ~1/)> Thus, applying Cauchy-Schwarz yields
2
preserves gy, has ’%"fl 0 = 0, and has *l
& o= 1. For the second equality
note that
fo= (Z fei—ajl] <4 [Z(n —2i+ 1)7]
w7 =1
£~ W=yt-2) fEt-at— FEE-—aE-y)
It is easy to compute
—yle-2) (y—2)0y—7) (z-2)z—y) the last expression and the final estimate that we
obtain is R
vanishes at ¢ = x,y,z. Hence by two applications of Rolle’s Theorem, its
second derivative vanishes at some ¢ in (min(z,y, 2), max(z,y, 2)) and this is POIEERT] B an(n®
LU — 1) &
o)
the required ¢. & =1
52 Chapter 2. Always Cauchy-Schwarz. .. 53

On the other hand, Legendre's identity shows that the last relation being obvious by definition of the ny’s. Finally, and most
importantly we have
n n n 2

Z (z; ~ :Ej)z = Znsz -2 (Z:fi) i

hj=1 =1 i=1 ’V( > o <§n:ak=1.
‘Well, unfortunately, when we combine all this we see that we are not 2-i<a, <2t k=1

done, because of the bad term 2 (Y%, #;)°. Fortunately, it is easy to repair Putting these inequalities together, we deduce that
the argument: indeed, we can always add the same number to all z;’swithout
changing the hypothesis or the conclusion of the problem. Thus, we may % 7
Z‘/7 < ‘/Zw+m.
assume that @ + g + - - + 2, = 0. But then the previous inequalities allow
21
us to conclude the proof. )
Taking N = logy(n), we obtain an even stronger (and strict!) inequality, in
The following problem is a very tricky application of the Cauchy-Schwarz
which 4 is replaced by 1. o
inequality. The technique used in the proof is worth remembering, since it is
quite useful. It is also a standard tool in analysis and probability (it is actually We end the series of moderately difficult problems with a very nice-locking
likely that the following problem is inspired by probability theory). improvement of an IMO 2004 problem.

23. Let ay,az,...,a, be positive real numbers which add up to 1. Let n; be 24. Let n > 2 be an integer. Find the greatest real number k with the
the number of integers k such that 211 > gp > 27 Prove that following property: if the positive real numbers zy, 23, ..., 2, satisfy

Z ‘/g < 4+ y/logy(n).

1
k> (zl+a:3+<~+3:,‘)(;—+l+~--+-1—),
i1 1

then any three of them are sides of a triangle.

L. Leindler, Miklos Schweitzer Competition
Adapted after IMO 2004
Proof. Choose a positive integer N and split the sum in two parts: the one
for i < N and the one for ¢ > N. Apply the Cauchy-Schwarz inequality for Proof. The main point is to solve this problem for n = 3. If b, ¢ are positive
each of them to obtain real numbers, let us lock at the possible values of

f(fr)=(x+b+c)(%+%+%)
when # > b+ c. It is not difficult to check {either directly or by computing
On the other hand, we have the derivative) that f is increasing in this domain of z, so that

Z%;:ZLV and EnLSZm:n,

>N >N izl
54 Chapter 2. Always Cauchy-Schwarz. .. 55

‘We deduce that if 2, b, ¢ satisfy f(z) < 10, then «,b, ¢ are the sides of a triangle ‘We give two proofs for the next difficult problem. The first is based on a
(since everything is symmetric in 2, b,c). Thus, for n = 3 the answer is 10. very tricky application of Cauchy-Schwarz combined with a mixing-variables
To reduce the general problem to the case n = 3, we use Cauchy-Schwarz, argument, while the second one is a pure, but very technical, mixing-variables
which allows us to obtain information about 1, zg, z3 knowing that argument.

k> (e 4+ ap+ o +an) (i + i IS i) X 25. If a,b,c,d, e are real numbers

such that ¢ + b+ c+d+e
= 0, then
Ty Ty Tn

Indeed, using Cauchy-Schwarz and the inequality @+t +F+d*+ )< ?(a“ +b 4t dt ).

1 1
(m4+-<-+yn)-<—+v-~+-—) > (n-3)% Vasile Cartoaje
24 Ty,

we obtain Proof. First of all, we may assume that among q,b,c,d,e at least three of
the numbers are nonnegative, say a,b, c. This follows immediately from the
1 1 1 pigeonhole principle, possibly after changing the sign of all numbers. The key
VE>n—3+ \/(xl +ar 2 (; + > * 1{)’ step is the following very tricky application of Cauchy-Schwarz:

so that 1 . 1 9(a® + b+ chrdt + ¢f) = [9(a® + b1+ )+ 2(dt + )] + (7dY) + (7€)

(w1 + m2 + 23) ( = 2) < (VE-n+3)% o VIO TV T )+ 2@ T e9) + 20(d + 2)?
’ - = 84+63+63 :
Using this and the previous case (n = 3), we deduce that if
A small computation yields the equivalent form of this inequality
= 1 1 1
(n~ 3+ VIO? > (21 + 22+ -+ 2n) <—+-+~~+7)(
I T2 Za ?(a4+b4+c“+d4+e4)
then any three among the numbers 21,2, . .., %, are the sides of a triangle. I T T T E T 2
Tt remains to check that this is indeed optimal. To see this, choose > (2V M)_j?(d_—iwz") 4+ e’) 3
21
Vo V10 :
Ty=xp=co=msy =l mpmap e o, 23 g This reduces the problem to showing that
Then 23,73,T3 are not the sides of a triangle and 2/0(af £ B+ A) +2d +eF) > V2i(a? + B2+ &

(@1 + @2+ -+ 2n) <i + 1 R i) = (n— 3+ V10)% To exploit the relationship between a,b,c,d, e, we use the fact that
Ty g Tn

i ' (d+e)* {a+b+e)t

Thus, the answer is ky, = (n — 3 + v/10)%. a d* et > 8 ) = § ) ) .
56 Chapter 2. Always Cauchy-Schwarz. . . 57

Thus, it remains to prove that Proof. We will use a mixing variables argument. Let

36(a® + 6% + ') + (a+b+¢)f 2 21(a® + b2 + 2)? Flabe,dey=30(@* + b+ +d' +e!) —T(®

+ B + 2 + > + %)%

for all nonnegative numbers a, b, c. ‘We want to prove that for ¢ +b+c+d+e = 0 we have F(a,b,c,d,e) > 0.
This inequality does not seem to follow easily from well-known results, so The basic formula we need is that
we will employ the powerful technique of mixing variables to prove it. Let
F(a,bc,d,e)~ F (a,b,l:, d+e’*d+e)
Fla,b, = 36(a’ + b1+ )+ (0 + b+ o)t - 21(a® 4+ 0P+ 2)? 2 2
= 15(a* + b* + ¢*) - 42(a%? + B2 + 2a?) 4+ (a+ b+ )L = (d — e)*(21d + 34de + 21€% — T(a® + b* +¢%)),

We will first show that for a = min{e, b, ¢) we have

which can be checked by tedious computation.
By the pigeonhole principle three of a,b.¢,d, e must have the same sign

flaho > £ (w5000

btcbte {counting zero as having either sign}. By symmetry, we may assume a,b,¢ > 0.
2 2 ! Then

For this we compute T+ +%) <T(a+b+6)? = 7(d +e)®

4 <17(d + € + 4(d® + %) = 21d° + 34de + 2162
fla,be)—f (a,b—fl, b_;_c) =15 (b"—i et - Q’%)
2 Therefore by the formula above F(a,b,c,d,€) > F {a,b,¢, 4%, 442}, Thus we
Canat may assume three of ¢, b, ¢, d, e are nonnegative and the other two are equal. To
42 (b (g2 + e (B —42 (b2c2 - M)
6
keep the same basic formula, we invoke symmetry and switch our assumption
to ¢,d,e > 0 and a = b= —(¢+d +¢)/2. Now suppose
¢ < e. Then we have
= g@ — ¢)? [426% + 92bc + 422 ~ 56a%] > 0,
T(a®+ 5+ %) = %(c+d+e)2+7cz
where the final inequality follows since 2862 +28¢% —56a? > 0. Thus we reduce
to the case where @ < b = ¢. In this case we compute < %(d +2€)% + 7% = %d2 + 14de + 2167
fla,b,b) = 4(b* + 8b% — 15b%a® + 2ba® + da¥) < 21d° + 3dde + 21¢%.
= 4(b— a)*(6* + 10ab + 4a?)
Therefore we again have F(a,b,c,d,e) > F{a,b,c, d*‘ 4 —1‘—) We conclude
>0 that we can repeatedly average the largest of ¢,d, e w1th either of the other
two and we will only lower F. By continuity of F, we therefore reduce to
The result follows. the case where ¢ = b and ¢ = d = e. In this case it is easy to check that
Note that we have equality for a = b = = 2andd=¢=-3. ] F(a,b,c,d,e) = 0 and we are done. in]
58 Chapter 2. Always Cauchy-Schwarz. .. 59

‘We end this chapter with a challenging inequality, which combines some and a second application of Cauchy-Schwarz yields
clever uses of Cauchy-Schwarz with a tricky homogeneity argument. This
result also appears in [35] and it is a generalization of a problem discussed in
[3], chapter 2, example 11.
V/Z (aris) -\ Saen i=1

26. Prove that for any positive real numbers ay, ag,...,8y, 1, T2,...,In
such that
5w = (3):
A T (otem)
ii<j<n 2 So, it remains to prove that

the following inequality holds

a
+an(zg+~~+xn)+m — —-y](mr s Tpor) 2
(Corta)
ag+ - +an i
n—1 S ar (ata)
- +ay
as+ A very elegant approach for this inequality was proposed by Darij Grinberg
in [35], where a more general result is proved. We may assume that
Vasile Céartoaje, Gabriel Dospinescu
@ tay+-+a,=1
Proof. First of all, it is enough to prove that for any z1,%s2,...,Zn > 0 and ‘We need to prove that
any a1,az,...,0, > 0 we have

me*
] P .. +2n) znv Zl<1<_’7<n
e ) LiZj
.
Using T2’ lemma (a form of Cauchy-Schwarz), we reduce this to proving that

Since this is homogeneous, it is enough to prove it when z; +@a+ -+ Ty = 1. 2

In this case, it is equivalent to
n
Z"maj > 2nfi2zaiaj(1_fli)(lfaj).
i<i i<y
a1 a1y
> +n Note that
Za2+~--+an72a2+~~+un
L(L.a] 2 2Za.,(]—a. .
But Cauchy-Schwarz shows that i<j
Thus, if we let b; = a;(1 — a;), it remains to prove that

by +bz 4 +ba) > fiZb,bJ.

i<j
60 Chapter 2. Always Cauchy-Schwarz. .. 2.1. Notes 61

This follows directly from Cauchy-Schwarz and the identity But this is easy, since the left-hand side is at least 23saa = 2(1 — 22)
and 2(1 - 2z) > 1—2;’ is equivalent to (1 — 2z)?> > 0. Thus, we have
23 biby = By by 4ot ba)? - (B B B, Flay,az,...,as) > F(z,z,0a3,...,a,). Continuing to mix variables in this
i<y way and using the continuity of F implies that F(a;,aq,...,a,) is at least
F(m,m,...,m), where m is the arithmetic mean of the a;’s, namely 1/n. The
Remark 2.1. We can also prove the inequality
result follows.
Z a;a; ._n
(t—a){l—a;) " 2n—2
i<j 2.1 Notes
by mixing variables: consider the map The following people provided solutions to the problems discussed in
9ia; this chapter: Vo Quoc Ba Can (problem 20), Ta Minh Hoang (problem 17),
Flay,az,...,an) 22 (1 . alm Mitchell Lee {problem 6), Dung Tran Nam (problem 25), Dusan Sobot. {(prob-
lems 1, 2, 5, 7), Richard Stong (problems 14, 19, 21, 25), Gjergji Zaimi {prob-
and set z = @LJZ‘—"Z We claim that F(ai,az,...,ax) = F(z,7,83,...,00). lems3, 4, 10, 13).
A small computation shows that this is equivalent to

ay a2
(1»(1141«&2 1—1) Zl—az

e fme %
(1 —a(1— ag) (1 e

Another computation shows that

@, e 2 (as
— a)®
1-a l-a 11—z 2(1-a)(l-a){l—2)
and
22 2a1az _{m — ag)? » 1-2z )
- {I-a)l-aj 2 (1—2)*(1
~ a1)(1 — az)
Thus, the inequality F{ay,as,...,an) = F(2,7,a3,...,an) is equivalent to

1w2r
T l-z’
 2.A. Cauchy-Schwarz in Number Theory 63

Addendum 2.A Cauchy-Schwarz in Theorem 2.A.1. (Gallagher’s larger sieve) Let S be a finite nonempty set of
Number Theory integers and let P be a finite set of prime powers. Assume that for eachp € P
we can find o real number u(p) > {{s (mod p)|s € S}| such that
This addendum shows some very beautiful applications of the Cauchy-
Schwarz inequality to number theory problems. We present Gallagher’s sieve > ((”))
- > 2logX,
and a beautiful arithmetic application; we discuss the large sieve, an amaz- pEP
ing tool invented by Linnik and extensively developed by a series of brilliant where X = maxeesls|. Then
mathematicians; finally we discuss another famous result, the Turdn-Kubilius
EpEP Alp) ~ log 2X
inequality. The Cauchy-Schwarz inequality plays an important role in the ISl <
proofs of all these theorems, which are elementary but quite powerful: this Spep a8 ~log2X
ought to show the reader how Cauchy-Schwarz appears in “real mathematics” Proof. Let p € P and let s(r,p) be the number of elements of S that are
and not only in olympiad-type problems. Analytic number theory has the rep- congruent to r modulo p. Then by Cauchy-Schwarz and the fact that
utation of being rather technical and this addendum is no exception. We hope
that the results presented here (especially their applications) will compensate u(p) 2 |{s (mod p)|s € S}|
for this nonetheless. we have
The following two sections try o give satisfactory answers to the following p—1 2 -1
natural problem: suppose that A is a set of integers such that |st*= (Zs(np)) <ulp)y_ srnp),
r=0 r=0
A (modp) = {z (mod p)lzec A} thus
152 fm
is relatively small for all primes p in a finite set P. Are there nontrivial bounds P DEED =0 51,5268
VL IR
s1#52E8
on the size of A? si=sp=r (mod p) plsi—sz
Multiplying this by A(p)and summing over all p yields
2.A.1 Gallagher’s sieve and a nice application
ISPy =& lS[ZA(p)+Z 3 A,
Recall that the von Mangoldt function A is defined by A(p™) = logp if peP s1#s2plsi—s2
pis a prime and n > 1 and A(z) = 0 for any other integer z. The crucial As
property of A is that > AW) <log(|s: — s2f) <log2X,
Z A(d) =logn plsi-s2
din we deduce that

is1* Z <8} ZA@) + (ISP - D log 2X,

for all 2. The following theorem is a pretty tricky application of the Cauchy-
Schwarz inequality, but the application given below reveals its usefulness and i
power. from which the result follows m)medmt.ely m]
64 Chapter 2. Always Cauchy-Schwarz. .. 2.A. Cauchy-Schwarz in Number Theory 65

The promised application (taken from [19}) requires some preliminaries. This follows from Euler’s celebrated formula 3°,, ;1; = 1'6: and the following
The following result is very classical, being related to the following natural computation®
question: given two positive integers, what is the probability that they are
relatively prime? Zu(d) Z E #{d) zkz Zu(d)—l
2
Proposition 2.A.2. As z — 00, we have d>1 a1 nd21 (ndy k=1

Zg‘o(k) 2r +O(zIng). Finally, as

k<e
Proof. The key point is the equality

olk) _ 5~ uld)
d>z -o(3) )
z

edik we obtain the desired result by combining the previous observations. ]

which follows easily from the classical formula Theorem 2.A.3. Let a.b > 1 be integers such that for any prime power p
there exists k > 1 (depending on p) such that b = a¥ (mod p). Then b is an
integral power of a.

Proof. We may assume that @ > 3 (as we may work with a2 and b2 instead of
a and b). The most difficult step is to establish that Ina and Inb are linearly
dependent over Q. Let us assume that this is not true and consider a large
number z. Let S; be the set of numbers smaller than z and of the form o’ - 57,
with 4, j nonnegative integers. As Ina and Inb are linearly independent over
Q, the set S, has the same number of elements as the number of pairs (i, j) for
which ilne+ jinb < Inz. So, there is an absolute constant ¢ > 0 depending
only on a and b such that |S;| > c(Inz)?. We will bound |S;| from above
using Gallagher’s sieve.
For any positive integer y let P, be the set of prime powers dividing at
least one of the numbers a —1,a% ~1,...,a¥ — 1. For each p € Py, let u(p) be
the order of a mod p. Then u(p) <y for all p € P,, and, since & is a power of
a modulo p, we have |S; (mod p})f < u{p) for all p € P,. To continue, we need
Note that 3, § = Oz lnz). Next, we claim that a technical lemma.

Zu(d = 6 “Which uses the absolute convergence of the double series, as well as the fact that
d>1
Sy 1) equals O for all £ > 1 and 1 for k= 1.
66 Chapter 2. Always Cauchy-Schwarz. .. 2.4, Cauchy-Schwarz in Number Theory 67

Lemma 2.A.4. There exist constants ¢y,ca > 0, depending only on a > 3, for an absolute constant c3. This contradicts the first paragraph for large z.
such that for all y > 1 we have Hence Ina and Inb are linearly dependent over Q and so there exists an
integer ¢ > 1 and positive integers i, j, relatively prime, such that ¢ = ¢! and
ay’ <Y Ap) € oyt b= ¢ If pis a prime power divisor of ¢ — 1, there is ky such that p divides
peP, @~ — 1 and so p divides ¢/ — 1. We deduce that ¢ — 1 divides ¢/ — 1 and so
Proof. One estimate is very easy, since i divides 7. The result follows. a

EA(p)<ZZAd:§: w-ma Remark 2.A.5. The result does not hold if we consider only primes instead of
PEPy j=1djai -1 prime powers. For instance, using properties of quadratic residues (not more
than the multiplicativity of Legendre’s symbol) one can easily prove that 16
The other estimate is more delicate and crucially uses properties of cyclotomic is an 8th power modulo any prime. Of course, 16 is not an eighth power of
polynomials? and proposition 2.A.2. Let ¢, be the nth cyclotomic polynomial. an integer. In the beautiful papers [4] and [32},% the following general result
Remark 9.5 implies that is proved fully using techniques of algebraic number theory:
S Az Y AW ZA(p) =10 gg(a) — Ind.
Theorem 2.A.6. Let n > 1 be an integer and let a be an integer such that
u(p)=d plbala)
a is an nth power modulo any sufficiently large* prime. Then either g is the
By the very definition of ¢ we have In d)d(a) = ¢{d) - In{a —~ 1). Hence nih power of an integer or 8in and a = 235 for some integer b.

SAp=3 Y APz (Zw)) ‘n(e-1) - Ind 2.A.2 The large sieve

PEFy sy ulp)=d d<y d<y
This rather long section presents a very deep result in analytic mumber
Since Z Ind = O(yIny), it suffices to use proposition 2.A.2 to conclude. theory, known as the large sieve. Introduced by Y. Linnik and extensively
dsy developed by Renyi, Bombieri, Davenport, Montgomery, Vaughan, Gallagher
The previous lemma shows that by choosing ¢ correctly and y about clnz,
(see [8], [9], [14], [25], [37], [44], [46], [57], [55], [56] to cite only a few refer-
ences}, this became a basic tool in modern analytic number theory, with rather
we can ensure that
spectacular results. We start by presenting a vast generalization of a famous
Alp) > i”chP Alp) > 2In(2z)
3 e inequality of Hilbert, due to Montgomery and Vaughan, which, combined with
some very clever tricks, yields the analytic form of the large sieve inequality.
Combined with standard results in finite Fourier analysis (for which the reader
and so by Gallagher’s sieve is invited to read addendum 7.A) this yields arithmetic forms of the large sieve.
This has some amazing applications to the distribution of prime numbers, twin
18]
S, < 111(21} (;A n( ar) | <caloz
—In(2z) < ezl
#4150 sec theorem 9.B.60.
*Actually, the proofs show that it is enough to assume that this holds for a set of primes.
}I‘or more details the reader is referred to section 9.2. of Dirichlet density 1.
 2.A. Cauchy-Schwarz in Number Theory 69
Chapter 2. Always Cauchy-Schwarz. ..
68

’I.‘he reader can easily check that A = (a;;) € M,(C) is hermitian if and
in arithmetic progressions, etc.
primes, least quadratic residue, prime numbers only if a;; = @ for all 4, j. A fundamental result of linear algebra is that for
article [56}.
We follow rather closely the amazingly well-written any- such matrix 4 we can find real numbers A;, A2, ..., A, and an orthonormal
basis vy, v vp of C" (ie. {v;,v;) = 0 if i # j and 1 otherwise) such that
The analytic form of the large sieve inequality Av; = ;v for all i. This can also be stated as: all eigenvalues of a hermitian
deep theorem, known as xx}atrlx are rc:d‘l numbers and there exists an orthonormal basis consisting of
We will spend some time proving the following
= minnez |2 - 7| be
the analytic form of the large sieve inequality. Let [z}l
eigenvectors.” The following result will be very useful in the next sections.
e set Z.
the distance from the real number @ to the discret Pmpogition 2.A.9. Let A be a hermitian matriz and let C > 0. If the
2.A.7. Let x1,32,...,Tn be real numbers such that inequality [{Av,v)| < Clvf? holds for any eigenvector.v of A, then it holds for
Theorem
any v € C™.
flai -zl Ze >0
Proof. Let vi,vs....,v, be an orthonormal basis of eigenvectors, with cor-
for alli#j. Let respondi
‘poxjnmg ige
eigenvalues S Ap,Ag....,A,. it
Consider any v € C" ' and write i
T(z) = Z 2 - 2T v =3"" 2 for some z; € C. Then
M<kSMAN
- 2M4N
& N and zp41,- € C.
be a trigonometric polynomial, where M,N (Av,v) =3 @Fih - (wv) = > M-z
Then 1 ' i i=1
n
Sreps(vel) ¥ oabl
M<k<MAN
By hypothesis we have {Av;,
i
must have [\ < C. Hence ’ :
5
f all 7,. and since
v:)| << Clu;{?{* for Av;vy == XA; - v, Wee
=t
Theorem 2.A.7 has a long history and many mathe
maticians contributed -
a very simple proof of the {Av,0} <C Y Il = Clof. o
to it: Davenport-Halberstamn, Gallagher {who gave
Montg omery and Vanghan, Selberg. =1
inequality with 7V +§ instead of N+ %),
N + % can be improved
One can prove (this is due to Selberg) that the factor ] ‘We end this section with a very useful technique. Though elementary, it
2.A.7 will span over
to N-1+ i and that this is sharp. The proof of theorem will play a key role in the proof of theorem 2.A.7. '
the next sections. Proposition 2.A.10. (Duality principle) Let (@i;)i<icmi<icn be complez
numbers and let C > 0. The following are equivalent:
Tools from linear algebra
1} For all z; € C we have
ities concerning
Tn this section we recall a few standard facts about inequal
Recall that the standard her- W
n 2 om
matrix norms and we prove a duality principle.
mitian product on " is defined by (@,y) = Y iw; Tilli,
Where Z; (respectively > Z%‘Z"} <CY lal
=1 li=1
(respectively y). If v € C", we denote )2 = {v,0). =1
) are the coordinates of —_—
called hermitian if for all “Recall that if 4 = (as;) € Ma(C) is any matrix and if A € C and v € € — {0 satisfy
Definition 2.A.8. A matrix A = (a;) € Ma(C) is Av = - v, then we say that v is an eigenvector of A associated to the eigenvalue A.
2,y € C" we have (Az,y) = {z, Ay).
70 Chapter 2. Always Cauchy-Schwarz. .. 2.A. Cauchy-Schwarz in Number Theory . 71

2) For all y; € C we have Since

| 2
. P
|ty
1)% 1| 2z(~1)%k

S auy| <O Il
=1 | {j=1 =1 &)
(the last equality is immediate, since 72_—,{,-" ~ 0 as n — oc), we can also

(-
write
Proof. Assume for instance that 1) bolds, Then for all z;,3 € C we have, by
Cauchy-Schwarz

Thus, it is enough to prove that for all N > n we have

ZZ“HZIJJ < Z “JJ CZ

< aaf? Z fo; 2.
i=l je=l = Y (1F TSN
ZZ1ZJ Z (IAF kdzi—z; Sgglll‘l
i#i k=N
By choosing for z; the complex conjugate of 3 1 @iy, we obtain the inequality
in 2). The converse is proved in exactly the same way. [m} Now, since there are N — lk{ solutions of the equation j; — j» = k with 7. /2 €
{1,2,..., N}, it is not difficult to see that the inequality is equivalent to
Montgomery and Vaughan'’s theorem
(~l)j‘+j22clzw_' 1 N
The key technical ingredient in the proof of theorem 2.A.7 is the following
delicate result [57).
>
12
>
1<sz<1\r
—_—r
“T Ak
T2
T
{ 2

Theorem 2.A.11. (Montgomery and Vaughan) Let ©i,22,...,0, be real This follows from the following vast generalization of a famous inequality due
numbers such that |lz; ~ ;|| > € > 0 for all i # 7. to Hilbert, applied to the family of real numbers (x; + Fici<ni<i<n and to
Then for all z;,22,...,2n. € C the family of complex numbers ({(— I)"’z,)1<,<n,1<_,<n
Theorem 2.A.12. (Montgomery-Vaughan) Let e >0 and let z1,x2, ..., 2, €R

Ss LR
i
be such that |z; — x;] > € for all i # j. Then for any compler numbers
b %mfl(r,vc]E Ev 2152255%0
=]

The proof of theorem 2.A.11 is a very nice mixture of elementary, analytic Z;‘ = {s le.l
and algebraic arguments. A first crucial ingredient is Euler's famous identity
(that we will take for granted) Proof. By homogeneity and symmetry we may assume that £ = 1 and that
T3 < ¥ < --- < x,. Then the hypothesis irnplies that z;—xz; > j~ifori <j.
Consider the matrix A, where a;;= Ligi s x . Note that A is antisymmetric,
thus i- A is a hermitian matrix, to which proposition 2.A.9 can be applied with
72 Chapter 2. Always Cauchy-Schwarz. ... 2.A. Cauchy-Schwarz in Number Theory 73

€ = . Thus, we may assume that z = (21,...,2n) is an eigenvector of i4, so Now, using the AM-GM inequality |257zx| < |2;{% + {2[? and rearranging
also of A, say Az =i Az for some A € R. terms, we obtain
By Cauchy-Schwaiz, the square of the left-hand side of the desired in- LHS <33 152 > ak. .
equality is bounded by 7, {2l - 30, 205 @i %5 , 0 it is enough to prove the B #j
inequality As |@; — x;] > li — j|, we have for any fixed j that®

T |k i=1
Expanding brutally the left-hand side and using the crucial observation that Combining the last two inequalities yields the desired resuit.

[}
ajaq = ajp{aig ~ o) yields

LHS =Y %a- . agan Proof of theorem 2.A.7

L3 stk Let ag; = e¥™% for 1 < j < nand M < k < M + N. We need to prove
that for all (2x)arcr<nrpn we have
.:Ziz 2 S ek Y Gaman | 3y — o 24
K i#j iFk i ik |
i
Z Grj2k < (N+ é) zkzizdz
It is now time to use the fact that z is an eigenvector: =1 M<kSM4+N
and by the duality principle {proposition 2.A.10) it is enough to prove that
S Zman | Y 0u =% Saga | | Y ey for all y1, ..., ¥, we have
i i 7 kA #5
- 12
o1 E
= i*z ‘Z]‘Z an S ey < (Iv + ;) >l
i i#i M<kSM4N [j=1 Ej
Doing the same with the other sum, we finally obtain that the term Expanding the left-hand side, we obtain the equivalent inequality

e bt ¥
TutE Y
M<ksM+N
e clyp F]
itk i
'We use here Euler’s famous identity
cancels with the other similar term, so that

LHS = Z |Zj\2 . Z af] + ZZTjZka]zk

i i# 7k
74 Chapter 2. Always Cauchy-Schwarz. .. 2.4. Cauchy-Schwarz in Number Theory k6]

Using the formula for the sum of a geometric series, a small computation shows Note that the only difference between this theorem and theorem 2.A.7
is
that for all u the factor IV, which is NV in theorem 2.A.7.

eRimku _ (ezflu(ZM—{A'ZN +1) _ E17m(2M+1)) . Proof. Assume that f is a continuously differentiable complex-valued
map on
M<kSMAN R. Integrating by parts, it is easy to establish the equality

By the triangle inequality, it is thus enough to prove that for any

& flz;) =/;’t§f(t)dt+f:; (t==+5) roe
$€{2M +2N +1,2M + 1} 7%

and for any y1,...,¥yn we have [ o) o

2
Using the triangle inequality and the fact that
i 1

i
P s,
Y
7) £ ;
fuit?
g s
t—z;+ &|l <

But this follows from theorem 2.A.11 with 2; = y; - 7%, forte (.’c] - g,x]] (and a similar inequality for ¢ € [a:j,zj + %]), we obtain

-
A ¢uick proof of a weaker form of the sieve inequality
i<y [z-% e+ 3 [ o

win
In this section we present Gallagher’s short and beautiful proof of the
following weaker form of the large sieve inequality. It is much simpler than
Take f(z) = T{2)* and add the corresponding inequalities. The hypothesis
the proof presented in the previous sections and the result it establishes is
Ilz: — ;] 2 ensures that the intervals (z; — §,2;5+ §) do not overlap mod
good enough in most applications of the large sieve. 1, so using the 1-periodicity of f we obtain
Theorem 2.A.13. Let 21, Ty, be real numbers such that n
5 1 1
o —ailize >0 S @<t /0 [T(x)Pdz + /D [T@)]
- [T (@) da.
=1
forallistj. Let Using Parseval’s equality and the Cauchy-Schwarz inequality, we obtain
T(z) = Z 2 - etihe
M<k<M+N

be a trigonometric polynomial, where M,N € N and zy41,... 2man € C.

Then n
]Z;[T(x,)p < (fl\ué) S jal
M<kSMEN
76 Chapter 2. Always Cauchy-Schwarz. .. 2.4. Cauchy-Schwarz in Number Theory 7

So we are done if we can ensure that maxar<p<arin [kl < % Of course, for Now, if the a, were uniformly distributed, one would expect that
arbitrary M and N it is unreasonable to hope for such an inequality, but a =h (mod ) % behaves as E; = 2Zk ag. Actually, a small computation
moment of thought shows that M plays no role in the theorem we are trying reveals that
to prove, so we can simply choose M = — [##1
”J to finish the proof. [m} q ] 2 e 2

Zj > oa-EB =3 3 a| —dlEf

Arithmetic forms of the large sieve inequality =1 {k=h (mod q) h=1 |k=h (mod q)
We will .apply theorem 2.A.7 for a special family of well-spaced numbers which combined with the previous formula yields

OO Pl
z;: pick an integer @ > 1 and consider all numbers of the form &, with
~1
ged(a,q) = land 1 < @ < ¢ < Q. It is clear that the difference between
two such numbers is (in absolute value) at least ;. Applying the large sieve
a=1 =1 {k=h (mod g)
inequality to this collection, we deduce the following

Theorem 2.A.14. Let Unfortunately, 397} |T (g)

T(z) = Z ape?e
M<ESMAN >
15a<gged{ag)=1
be a trigonometric polynomial. Then for all @ > 1 we have but if ¢ is a prime number, they are actually equal! Using these remarks and
the previous theorem, we obtain the following strong inequality:

S¢=13 ISasg
2 ()
(W <ycwviaer ¥ oP
M<ESMAN
Theorem 2.A.15.
Then for all integers
Let {(ax)arcr<m+n
@ > 1 we have
be @ sequence of complex numbers.

ged{a,g)=1

Next, let us observe that S0 ¥

PSQ A=l
)
k=h (mod p)
ak‘—zflk <(N+Q2)Z‘flklz

( ) Zakfi”f": N the sum being taken over all primes p < Q.

A=l \E=h (mod q) By specializing aj = lxea for a subset A C [A + 1, M + N}, we obtain
the following equidistribution result:
Therefore, using the techniques of addendum 7.A, more precisely Plancherel’s
formula (theorem 7.A.5, but this can also be done by expanding everything),
Corollary 2.A.16. Let A C [M + 1,M + N] be a set of iniegers. Then for
we obtain all integers Q > 1 we have
2
NGz~ J1al] < v+ @i
1 2

Zizri
P<Qh=1
78 Chapter 2. Always Cauchy-Schwarz. .. 2.4. Cauchy-Schwarz in Number Theory 79

This finally yields a “sieve inequality.” and

Theorem 2.A.17. Let A C [M + 1, M + N] be a set of integers and suppose ANz{ngm(%) =1 forall pEP]sl};
that for each prime p < Q at least w(p) residue classes mod p contain no
element of A. Then ‘We will prove the following technical result:
Lemma 2.A.19. There exist positive constants c(e) and c;(e) such that for
all N > ¢1(e) we have |Ax| > c(e)}N.
Proof. Use the previous corollary and the observation that for each p we have Let us assume for a moment that this holds. Since Ay (mod p) misses all
P 2 2
quadratic non-residues mod p, we can choose w(p) = b = & in the previous
AN (h +pT) ~ %IAli 2 w(P)%’w theorem and obtain [Ayx}< ,| Hence for all N > cl(e) we have 1Pnl< g
he=1 which is enough to conclude the theorem.
Let us prove the lemma. Note that Ay contains all numbers n < N all
as at least w(p) terms in the sum are equal to }-’;?7;. o of whose prime factors are smaller than N¢, as Legendre’s symbol is mul-
tiplicative. We will consider all numbers n = kp; ---pg smaller than N,
The previous theorem helps understanding the name “large sieve.” Tn-
deed, in typical applications A (mod p) will miss a good part of the residue with pl € [N® g , N*¢] in nondecreasing order and some integer k. Note chat
classes modulo p (i.e. the integers w(p) will be quite large) and the sieve k< N: as n < N, so k is relatively prime to any prime greater than N*®
inequality will yield nontrivial bounds on [A|. The next result, an amazing This easlly implies that all such numhe!s are different and they clearly beloug
theorem of Linnik, is at the very origin of the large sieve and one of its best to Ay. Since there are more than — 1 possible values for £, it follows
FIPZ Pl
applications. 1t concerns the distribution of the least quadratic non residue that we have at least
modulo p. Vinogradov conjectured that it is smaller than c()p® for any ¢ > 0 d
and if one assumes the Generalized Riemann Hypothesis, then one can actu-
ally prove that it is smaller than c(log p)?. The following deep theorem due to —aveizan
~dl
| S LP e
Linnik [47] is a first step in the direction of Vinogradov’s conjecture (which is
still largely open):
peNTFNe]
Theorem 2.A.18. (Linnik) Let n{p) be the least positive integer which is not such numbers. Taking into account the fact” that 7(z)= o{z) and
@ quadratic residue mod p. Then for all ¢ > 0 there is a constant c(s) such
that for all N we have Z S= + o)
wwn?
Kp < Nia(p) > N} < c(e)
as a,b — o0, the last quantity is easily seen to be greater than c(¢)N for
Proof. Fix € > 0, which for simplicity (but without loss of generality) we take
some positive constant c(¢) and all sufficiently large (depending on ) N. The
of the form a} for some integer
g d &l greater than 2. For a ¥ positive integer
g N , let
desired result follows. Im]

Py = {p < VNin(p) > N°} Both these resulis are proved in addendum 3.A.
80 Chapter 2. Always Cauchy-Schwarz. .. 2.A. Cauchy-Schwarz in Number Theory 81

Though theorem 2.A.17 is already a very strong result, in most applica- fact that the theorem holds for ¢ (with 7 (% + m) instead of T'(z)) and then
tions one needs more refined estimates, in which one takes into account all
for ¢/, we have
g < Q, not only prime numbers. In order to do this, we need another nice ap-

Gl - 2, 2 G
plication of finite Fourier analysis and Cauchy-Schwarz, but before doing that 2 3
recall that 4 is Mobius’s function, defined by p(n) = (—1)F if n is a product
of k > 0 distinct prime numbers and u(n) = 0 otherwise, a€(Z/qq'T)" > flE(Z/cZ)
Theorem 2.A.20. Let A C [M + 1, M + N| be o set of integers and suppose
that for each prime p < Q we have
> 9T
bE(Z/q’Z)' [ (ql)

Ha (mod p)la € AH < p—w(p). 2 9@ NTOF = g(aq)|T(O)2.

Then for all trigonometric polynomials T(x) = Y ¢ 4 axe® ™= and all positive o
integers q

= (o) 2w T2t mor

Combining the special case T(z) = ¥, , €% of the previous theorem
with the large sieve inequality (theorem 2.A.14), we obtain the following strong
ged{a,g)=1"
g form of theorem 2.A.17:
Proof. Let Theorem 2.A.21. (Montgomery) Let A C {M+1, M +N) be a set of integers
o) = m@? ] o and suppose that for each prime p < Q we have
pig
a multiplicative function. We will prove the theorem in two steps: first we
o (modplla€ A} <p—wip).
will prove it when ¢ is prime and then we will show that if the theorem holds
Then [A] < ME 2, where
for two relatively prime numbers ¢, ¢, then it also holds for g¢. First, assume
that ¢ is prime. Let
Zp = Z ag. L=3 wa?]] p*f(gzp) .
keAk=h (mod q) <@ plg
Then at least w(g) of the numbers x;, are zero (by hypothesis), so by Cauchy-
Some applications
Schwarz and Plancherel’s identity we get
We present here two classical applications of the large sieve: a uniform

()
q a\ |2
sz (a=st) Sl = (1-%2)
/5
3 w
g
upper bound on 7{m + n) — 7(n) and an upper bound for the number of twin
primes smaller than n.

from where the result follows easily. Next, assume that the result holds for ¢ Theorem 2.A.22. If m > /n, then there are at most 1 L primes between
and ¢, with ged(g,¢’) = 1. Using the Chinese Remainder Theorem and the m+1 and m+n.