Anda di halaman 1dari 222

“Superior Mathematics from an Elementary point of view” course notes

Undergraduate course, 2017-2018, University of Pisa


Jack D’Aurizio

Contents
0 Introduction 3

1 Creative Telescoping and DFT 4

2 Convolutions and ballot problems 17

3 Chebyshev and Legendre polynomials 33

4 The glory of Fourier, Laplace, Feynman and Frullani 44

5 The Basel problem 65

6 Special functions and special products 75

7 The Cauchy-Schwarz inequality and beyond 102

8 Bessel functions and the Gauss circle problem 126


8.1 The Gauss circle problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

9 Remarkable results in Linear Algebra 145

10 The Fundamental Theorem of Algebra 151

11 Quantitative forms of the Weierstrass approximation Theorem 160

12 Elliptic integrals and the AGM 165

13 Dilworth, Erdos-Szekeres, Brouwer and Borsuk-Ulam’s Theorems 174

14 Continued fractions and elements of Diophantine Approximation 185

15 Symmetric functions and elements of Analytic Combinatorics 200

16 Spherical Trigonometry 210


0 INTRODUCTION

0 Introduction
This course has been designed to serve University students of the first and second year of Mathematics. The purpose
of these notes is to give elements of both strategy and tactics in problem solving, by explaining ideas and techniques
willing to be elementary and powerful at the same time. We will not focus on a single subject among Calculus,
Algebra, Combinatorics or Geometry: we will just try to enlarge the “toolbox” of any professional mathematician
wannabe, by starting from humble requirements:

• knowledge of number sets (N, Z, Q, R, C) and their properties;

• knowledge of mathematical terminology and notation;

• knowledge of the main mathematical functions;

• knowledge of the following concepts: limits, convergence, derivatives, Riemann integral;

• knowledge of basic Combinatorics and Arithmetics.

A collection of problems in Analysis and Advanced integration techniques, kindly provided by Tolaso J. Kos and Zaid
Alyafeai, are excellent sources of exercises to match with the study of these notes.

These notes are distributed under a Creative Commons Share-Alike (CCSA) license. Personal use is allowed,
distribution is allowed with the only constraint of making a proper mention of the author (Jack D’Aurizio).
Commercial use, modification or inclusion in other works are not allowed.

Page 3 / 222
1 Creative Telescoping and DFT
It is soon evident, during the study of Mathematics, that the bijectivity of some function f does not grant that the
explicit computation of f −1 (y) is “just as easy” as the explicit computation of f (x). Some examples are related to the
ease of multiplication, against the hardness of factorization; the possibility of computing derivatives in a algorithmic
fashion, against the lack of a completely algorithmic way to find indefinite integrals; the determination of a Galois
group of an irreducible polynomial over Q, against the difficult task of finding a polynomial having a given Galois
group. In the present section we will outline two interesting techniques for solving (or getting arbitrarily close to an
actual “solution”) a peculiar inverse problem, that is the computation of series.

Definition 1. We give the adjective telescopic to objects of the form a1 + a2 + . . . + an , where each ai can be written
as bi − bi+1 for some sequence b1 , b2 , . . . , bn , bn+1 . With such assumption we have:

a1 + a2 + . . . + an = (b1 − b2 ) + (b2 − b3 ) + . . . + (bn − bn+1 ) = b1 − bn+1 .

Essentially, every telescopic sum is simple to compute, just as any convergent series with terms of the form bi − bi+1 .
A peculiar example is provided by Mengoli series: the identity
N N  
X 1 X 1 1 1
= − =1−
n=1
n(n + 1) n=1 n n + 1 N +1

grants we have
X 1
= 1.
n(n + 1)
n≥1

The following generalization is also straightforward:

Lemma 2. X 1 1
∀k ∈ N+ , = .
n(n + 1) · . . . · (n + k) k · k!
n≥1

In a forthcoming section we will also see a proof not relying on telescoping, but on properties of Euler’s Beta function.

The first technical issue we meet in this framework is related to the fact that recognizing a contribution of the form
bi − bi+1 in the general term of a series is not always easy, just like in the continuous analogue: if the task is to
Rb
find a f (x) dx, it is not always easy to devise a function g such that f (x) = g 0 (x). Here there are some non-trivial
examples:

Lemma 3.  
X 1 π
arctan = .
1 + n + n2 4
n≥1

Proof. If we use “backwards” the sum/subtraction formulas for the tangent function, we have that

tan(x) ± tan(y)
tan(x ± y) =
1 ∓ tan(x) tan(y)
implies: !
1 1

 
1 1 n n+1 1
arctan − arctan = arctan 1 = arctan
n n+1 1+ n(n+1)
1 + n + n2
π
so the given series is telescopic and it converges to arctan(1) = 4.

Page 4 / 222
1 CREATIVE TELESCOPING AND DFT

Exercise 4. The sequence of Fibonacci numbers {Fn }n≥0 is defined through F0 = 0, F1 = 1


and Fn+2 = Fn+1 + Fn for any n ≥ 0. Show that:

(−1)n+1 √
X  
arctan = arctan( 5 − 2).
Fn+1 (Fn + Fn+2 )
n≥1

Exercise 5. Show that:


   
X sinh 1 π X 1 π π
arctan = , arctan = − arctan tanh .
cosh(2n) 2 8n2 4 4
n∈Z n≥1

Exercise 6. Prove that:


N    
X 1 2n N + 1 2N + 2
= ,
n=0
4n n 22N +1 N + 1
for instance by considering that by De Moivre’s formula we have
  Z π
1 2n 1
An = n = cos2n (x) dx
4 n 2π −π
P2n Rπ
since cos(x)2n = 41n j=0 2n
 (2n−j)ix −jix
j e e and −π ekix dx = 2π · δ(k),
so the only non-vanishing contribution is related to the j = n term, and
N Z π
1 − cos2N +2 (x)
   
X 1 2n 1 N + 1 2N + 2
= dx = (2N + 2) AN +1 = 2N +1
n=0
4n n 2π −π sin2 (x) 2 N +1

dx
R
follows from the integration by parts formula ( sin2 x
= − cot x). Prove also that
X 2k

1
=1
k (k + 1)4k
k≥0

by recognizing in the main term a telescopic contribution. Give a probabilistic interpretation to the proved identities,
by considering random paths on a infinite grid (Z × Z) where only unit movements towards North or East are allowed.

We now outline the first (really) interesting idea, namely creative telescoping: even if we are not able to write the
main term of a series in the bi − bi+1 form, it is not unlikely there is an accurate approximation of the main term that
can be represented in such a telescopic form. By subtracting the accurate telescopic approximation from the main
term, the original problem boils down to computing/approximating a series that is likely to converge faster than the
original one, and the same approximation-by-telescopic-series trick can be performed again. For instance, we might
employ creative telescoping for producing very accurate approximations of the series
X 1
ζ(2) =
n2
n≥1

that will be the main character of a forthcoming section.


In particular, for any n > 1 the term n12 is quite close to the telescopic term 1
n2 − 14
:
 
1 4 1 1
1 = =2 −
n2 − 4
(2n − 1)(2n + 1) 2n − 1 2n + 1

Page 5 / 222
1 1
and we have n2 − n2 − 41
= − (2n−1)n12 (2n+1) , so:

X 1 1
X X 1
ζ(2) = 1 + = 1+ 1 −
n2 n 2−
4
(2n − 1)n 2 (2n + 1)
n≥2 n≥2 n≥2
2 X 1
= 1+ −
3 (2n − 1)n2 (2n + 1)
n≥2

gives us ζ(2) < 35 (that we will prove to be equivalent to π 2 < 10), and the magenta “residual series” can be manipu-
lated in the same fashion (by extracting the first term, approximating the main term with a telescopic contribution,
considering the residual series) or simply bounded above by:
X 1 X 1 3 22
< 1
 1
 = ζ(2) −
(2n − 1)n2 (2n + 1) (2n − 1) n − 2 n + 2 (2n + 1) 2 9
n≥2 n≥2

from which the lower bound ζ(2) > 74 74 5


45 follows. We may notice that the difference between 45 and 3 is already pretty
small. In this framework the iteration of creative telescoping leads to two interesting consequences: the identity

X 1 X 3
ζ(2) = = ,
n≥1
n 2
n≥1
n 2n
2
n

providing a remarkable acceleration of the series defining ζ(2), and Stirling’s inequality:

 n n √ 
1
  n n √ 
1

2πn exp ≤ n! ≤ 2πn exp
e 12n + 1 e 12n

further details will be disclosed soon.


It is also possible to employ creative telescoping for proving that:

X 1 5 X (−1)n+1
ζ(3) = =
3
n3 2n

n 2 n
n≥1 n≥1

a key identity in Apery’s proof of ζ(3) 6∈ Q.


We may notice that:
1 1 (−1)
3
= +
n (n − 1)n(n + 1) (n − 1)n3 (n + 1)
1 1 −22
3
= +
(n − 1)n (n + 1) (n − 2)(n − 1)n(n + 1)(n + 2) (n − 2)(n − 1)n3 (n + 1)(n + 2)
Continuing on telescoping we get that:
m
1 (−1)m m!2 X (−1)j−1 (j − 1)!2
= +
n3 (n − m) . . . n3 . . . (n + m) j=1 (n − j) . . . (n + j)

So by setting m = n − 1:
n−1
1 (−1)n−1 (n − 1)!2 X (−1)j−1 (j − 1)!2
= +
n3 n2 (2n − 1)! j=1
(n − j) . . . (n + j)

The terms of the last series can be managed through partial fraction decomposition:
1 1 1 1
= − + − ...
(n − j) . . . (n + j) (2j)!(n − j) (2j − 1)!1!(n − j + 1) (2j − 2)!2!(n − j + 2)

Page 6 / 222
1 CREATIVE TELESCOPING AND DFT

2j 2j 2j

(n − j − 1)! X (−1)k 1 X (−1)k k
= =
(n + j)! (2j − k) k! (n − j + k) (2j)! n−j+k
k=0 k=0

and since:
2j 2j 2j 2j−1

(−1)h−1 1

(−1)k
Z
XX X h−1 1
k
= = (1 − x)2j−1 dx =
n>j k=0
n−j+k h 0 2j
h=1

we get:
+∞ +∞
X (−1)n−1 n!2 X X (−1)j−1 (j − 1)!2
ζ(3) = +
n=1
n4 (2n − 1)! j=1 n>j (n − j) . . . (n + j)

+∞ +∞ +∞
X (−1)n−1 n!2 X (−1)j−1 j!2 5 X (−1)n−1
ζ(3) = + =
n4 (2n − 1)! j=1 2j 3 (2j)! 2 n=1 n3 2n

n=1 n

as wanted.

Exercise 7. Prove that the following identity (about the acceleration of an “almost-geometric” series) holds.
X 1 1 X 8m + 1
= + .
n≥2
2n − 1 4
m≥2
(2m − 1)2m2 +m

As proved by Tachiya, this kind of acceleration tricks provide a simple way for proving the irrationality
of n≥1 qn1+1 and n≥1 qn1−1 for any q ∈ Z such that |q| ≥ 2.
P P

Creative telescoping can also be used for a humble purpose, like proving the divergence of the harmonic series.
By recalling that the n-th harmonic number Hn is defined through
n
X 1
Hn =
k
k=1

and by recalling that over the interval (0, 1] we have:


 x 
x 1+ 2
x < 2 arctanh = log x
2 1− 2

it follows that:
n  
X 2k + 1
Hn < log = log(2n + 1).
2k − 1
k=1

On the other hand 2 arctanh x2 − x = O(x3 ) in a neighbourhood of the origin, and the series k≥1 k13 = ζ(3) is
 P

convergent, so there is an absolute constant C granting Hn ≥ log(2n + 1) − C for any n ≥ 1. In a similar way, by
(j)
defining the n-th generalized harmonic number Hn through
n
X 1
Hn(j) = ,
kj
k=1

we may easily check that the sequence {an }n≥1 defined by


n
√ √ X 1
an = 2 n − Hn(1/2) = 2 n − √
k=1
k

is increasing and never exceeds a constant close to 1 + √1 . About an+1 ≥ an we have:


5

√ √ 1 1 1 1
an+1 − an = 2 n + 1 − 2 n − √ = 1 √ √ −√ =√ √ √ 2 > 0
n+1 2 n+ n+1 n+1 n+1 n+ n+1

Page 7 / 222
and
n
X 1 X 1
an = √ √ √ −→
2 n→+∞ √ √ √ 2 .
m=0 m+1 m+ m+1 n≥0 n+1 n+ n+1

The claim then follows from considering that the main term of the last series is well-approximated by the telescopic
term
1 1
√ −√
4n + 1 4n + 5
for any n ≥ 1. In similar contexts, by exploiting creative telescoping and the Cauchy-Schwarz inequality we may get
surprising results, like the following one:
v
n n u n  
X 1 X 1 CS u X 1 1 1
< √ √ ≤ tn − =√
n+k n+k−1 n+k n+k−1 n+k 2
k=1 k=1 k=1

but the limit of the LHS for n → +∞ is log(2), hence log(2) ≤ √12 . In general, by mixing few ingredients among
creative telescoping, the Cauchy-Schwarz inequality, convexity arguments and Weierstrass products we may achieve
short and elegant proofs of highly non-trivial claims, like:

Lemma 8. The sequence {an }n≥1 defined through


q
1


2n
 π n+ 4
an =
n 4n

is increasing and convergent to 1, due to the identity

(2n + 1)2 (4n + 5) − 4(n + 1)2 (4n + 1) = 1.

That implies
Γ x + 12

x
∼p
Γ(x) x + 1/4
for any x > 0, that is a strengthening of Gautschi’s inequality.

Creative telescoping is also a key element in the Wilf and Zeilberger algorithm for the symbolic computation of binomial
sums (http://mathworld.wolfram.com/Wilf-ZeilbergerPair.html), further extended by Gosper to the hyperge-
ometric case and by Risch (https://en.wikipedia.org/wiki/Risch_algorithm) to the symbolic computation of
elementary antiderivatives.

Exercise 9. Prove by creative telescoping that for any k ∈ {2, 3, 4, . . .} we have:


X 1
= k − ζ(2) − . . . − ζ(k)
n(n + 1)k
n≥1

1
P
where ζ(m) = n≥1 nm .

The following exercise is particularly exemplary, since it stresses some interesting relations among creative telescoping,
the Cauchy-Schwarz inequality, the Maclaurin series of arcsin2 (z), ζ(2) and Catalan numbers: all these topics will be
deeply investigated in the following sections.

Page 8 / 222
1 CREATIVE TELESCOPING AND DFT

Exercise 10. Prove that the value of the series


X 1
S= √
(n + 1) n
n≥1

1 π

is extremely close to 2 + 4 3.

Proof. It is not difficult to realize that


2n

1 √ X n 1 3√
S≈ + π n
= + π
2 4 (n + 1) 2 4
n≥2
2n
since 41n n ≈ √1πn is a pretty good approximation for any n ≥ 1 and the generating function for Catalan numbers


is fairly well-known. This can be improved by exploting the more accurate


√   
1 π 2n 1 1
√ ≈ n 1+ + .
n 4 n 8n 128n(n + 2)
1 √
Creative telescoping provides us a more elementary approach: indeed, (n+1) n
< √2n − √n+1
2
immediately proves
q
1 √
S < 2, and the more accurate (n+1) n
≈ √ 2 1 − √ 2 7 gives S ≈ 12 + 2 13
6
. On the other hand we may also combine
n+ 6 n+ 6
the approximation through central binomial coefficients with the Cauchy-Schwarz inequality to get an exceptionally
simple and very accurate approximation:
v 
u
2n
  r
1 u X 4n 1 π2 3
u X
n
S ≤ +t 2n
  
n
 = + ·
2 n(n + 1) n (n + 1)4 2 4 4
n≥2 n≥2

gives S ≈ 1
2 + 3
4 π, whose absolute error is less than 4 · 10−4 .

Before introducing a second tool (the discrete Fourier transform, DFT ), it might be interesting to consider an appli-
cation of creative telescoping to the computation of an integral.

Exercise 11. Prove that the following identity holds:


1
log(x) log2 (1 − x)
Z
1X 1 ζ(4)
dx = − 4
=− .
0 x 2 n 2
n≥1

Proof. The dilogarithm function is defined, for any x ∈ [0, 1], through:
X xn
Li2 (x) = .
n2
n≥1

We may notice that Li02 (x) = − log(1−x)


x , so, by integration by parts:
Z 1 Z 1  
log(x)Li2 (x) Li2 (x) 0
dx = log(1 − x) + log(x)Li2 (x) dx
0 1−x 0 x
log(x) log2 (1 − x)
Z 1 Z 1
= − Li02 (x)Li2 (x) dx − dx
0 0 x
In particular the opposite of our integral equals:
log(x) log2 (1 − x)
Z 1 Z 1 X xn X
1 2
− dx = Li (1) + xk log(x) dx
0 x 2 2 0 n≥1 n2
k≥0
1 X 1 X 1
= ζ(2)2 −
2 n2 m>n m2
n≥1

Page 9 / 222
where, by symmetry:  
2
X 1 1 X 1 X 1
=  −
m2 n2 2 n2 n4

m>n≥1 n≥1 n≥1

and the claim readily follows. We may notice that:


log2 (1 − x) X 2Hn n
= x ,
x (n + 1)
n≥0

since: X xn − log(1 − x) X 1 X Hn
− log(1 − x) = = Hn x n log2 (1 − x) = xn+1
n 1−x 2 n+1
n≥1 n≥1 n≥1
R1 n 1
By termwise integration (through 0
(− log x)x dx = (n+1)2 ) the proved identities lead to:
X Hn 1X 1 ζ(4)
3
= = .
(n + 1) 4 n4 4
n≥1 n≥1

A keen reader might ask why this virtuosity1 has been included in the creative telescoping section. The reason is
the following: in order to make the magic work, we actually do not need the dilogarithm function (a mathematical
function with the sense of humour, according to D.Zagier) or integration by parts. As a matter of fact:
n X1  X
X 1 1 n
Hn = = − =
k m m+n m(m + n)
k=1 m≥1 m≥1

hence it follows that:


X Hn X  Hn+1 1
 X Hn
= − = −ζ(4) +
(n + 1)3 (n + 1) 3 (n + 1) 4 n3
n≥1 n≥1 n≥1
 
X 1 1 X 1 1
= −ζ(4) + = −ζ(4) + +
mn2 (m + n) 2 mn2 (m + n) m2 n(m + n)
n,m≥1 m,n≥1
1 X 1 1
= −ζ(4) + = −ζ(4) + ζ(2)2
2 m2 n2 2
m,n≥1

and by comparing the last identity to the identities we already know, we get that ζ(4) = 52 ζ(2)2 .
Some questions might naturally arise at this point: is it possible, in a similar fashion, to relate the value of ζ(2k+1 )
to the value of ζ(2k )? Or: is it possible to find the explicit value of ζ(2) by simply squaring the Taylor series at the
origin of the arctangent function? Answers to such questions are postponed.

We directly introduce the DFT through a problem.

Exercise 12. Let A be a finite set with cardinality ≥ 4. Let P0 be the set of subsets of A with 3j elements, let P1 be
the set of subsets of A with 3k + 1 elements, let P2 be the set of subsets of A with 3h + 2 elements. Prove that any
two numbers among |P0 |, |P1 |, |P2 | differ at most by 1, no matter what |A| is.

R log(1−x) log2 (x)


1 Itis worth mentioning that just like 01 dx is associated to an Euler sum with weight 4, namely n≥1 H
P n
1−x n3
, the similar
R 1 log(1+x) log2 (x) P Hn n+1
integral 0 1+x
dx is associated to the alternating series n≥1 n3 (−1) . On the other hand, while the first series is clearly
given by the values of the Riemann ζ function at s = 2 or s = 4, the alternating series has a much more involved closed form:
1 1 log(1 + x) log2 (x)
Z
X Hn n+1 1  4
−π − 4π 2 log2 (2) + 4 log4 (2) + 96 Li2 21 + 84 log(2)ζ(3)
 
3
(−1) = =
n≥1
(n + 1) 2 0 1 + x 48

has been proved by De Doelder in 1991. See also Flajolet and Salvy, Euler sums and contour integral representations.

Page 10 / 222
1 CREATIVE TELESCOPING AND DFT

The claim appears to be a (more or less) direct generalization of a well-known fact: the number of subsets of I =
{1, 2, . . . , n} with even/odd cardinality is the same. In that framework, we may consider the map sending B ⊆ I in
B \ {1} when 1 ∈ B, and in B ∪ {1} when 1 6∈ B (“if there is 1, we remove it, otherwise we insert it”): such map is
an involution and provides a bijection between the subsets with even cardinality and the subsets with odd cardinality.
As an alternative, by recalling that in I we have nk subsets with k elements, we may simply check that


n  
X n
(−1)k = 0
k
k=0

holds as a trivial consequence of the binomial Theorem applied to (1 − 1)n .


In the ternary case we have to compare the sums
     
X n X n X n
|P0 | = , |P1 | = , |P2 | =
k k k
k≡0 (mod 3) k≡1 (mod 3) k≡2 (mod 3)

and we would like to have a tool allowing us to isolate the contributions given by elements in particular positions
(positions given by an arithmetic progression) in a sum. The DFT is precisely such a tool.

2πi

Lemma 13 (DFT). If n ≥ 2 is a natural number and ω = exp n , the function f : Z → C given by

n−1
1 X km
f (m) = ω
n
k=0

is the indicator function of nZ. As a consequence,


n−1
1 X −hk km
χh (m) = ω ω
n
k=0

is the indicator function of the integers ≡ h (mod n).

The possibility of writing some indicator functions as weighted power sums has deep consequences.
In our case, if we take ω as a primitive third root of unity, we have:
n   n  
X n 1X n  (1 + 1)n + (1 + ω)n + (1 + ω 2 )n
|P0 | = χ0 (k) = 1k + ω k + ω 2k =
k 3 k 3
k=0 k=0

due to the binomial Theorem. Since both (1 + ω) and its conjugate (1 + ω 2 ) lie on the unit circle, we have that |P0 |
n
is an integer number whose distance from 23 is bounded by 13 . The reader can easily check the same holds for |P1 |
and |P2 | and the claim readily follows. The discrete Fourier transform proves so the reasonable proposition claiming
the almost-uniform distribution of the cardinality (mod 3) of subsets of {1, . . . , n}. Perfect uniformity is clearly not
possible, since |P0 | + |P1 | + |P2 | = 2n never belongs to 3Z.
Exercise left to the reader: prove the claim of Exercise 11 by induction on |A|.

Exercise 14. Find the explicit value of the series:


X 1
S= .
(3n)!
n≥0

We may consider that the complex exponential function


X zn
ez =
n!
n≥0

Page 11 / 222
is defined by an everywhere-convergent power series, then apply a ternary DFT to such series and get, like in the
previous exercise (here we are manipulating an infinite sum, but there is no issue since ez is an entire function):
X z 3n X zn X zn 1 X z n + (ωz)n + (ω 2 z)n 1 z 2

= = χ0 (n) = = e + eωz + eω z
(3n)! n! n! 3 n! 3
n≥0 n≡0 (mod 3) n≥0 n≥0

√ √
−1+i 3 −1−i 3
and since ω = 2 and ω 2 = 2 , for any z ∈ C we have:
√ !
X z 3n 1 z 3
z −z/2
= e + 2e cos ,
(3n)! 3 2
n≥0

that by an evaluation at z = 1 leads to: √


e 2 3
S = + √ cos .
3 3 e 2
Was the introduction of the complex z variable really necessary? It clearly was not: a viable alternative would have
been to just re-write 1 as 1n in the definition of S. Besides the identity 1 = 1n being really obvious, the idea of tackling
the original problem through such identity and the DFT is not obvious at all: similar situations explain just fine the
subtle difference between the adjectives elementary and easy in a mathematical context. Another famous application
of the DFT is related to the Frobenius coin problem:

Exercise 15. Given n ∈ N, let Un be the number of natural solutions of the (diophantine) equation a + 2b + 3c = n,
i.e.
Un = (a, b, c) ∈ N3 : a + 2b + 3c = n .


(n+3)2
Prove that for any n, Un equals the closest integer to 12 .

This claim will be proved in the section about Analytic Combinatorics, since few elements of Complex Analysis and
manipulation of formal power series are required. However we remark that the key idea is the same key idea of
Hardy-Littlewood’s circle method, a really important tool in Additive Number Theory: for instance, it has been used
for proving that any odd natural number large enough is the sum of three primes (Chen’s theorem, also known as
ternary Goldbach). Now we will focus on a typical application of the DFT in Arithmetics, i.e. a proof of a particular
case of Dirichlet’s Theorem.

Theorem 16 (Dirichlet). If a and b are coprime positive integers, there are infinite prime numbers ≡ a (mod b).

The particular case we are going to study is the proof of the existence of infinite primes of the form 6k + 1. We recall
that the infinitude of primes of the form 6k − 1 follows from a minor variation on Euclid’s proof of the infinitude of
primes:

Let us assume the set of primes of the form 6k − 1 is finite and given by {p1 = 5, 11, 17, . . . , pM } = E.
Let us consider the huge number
YM
N = −1 + 6 pm .
m=1

By construction, no element of E divides N . On the other hand, N is a number of the form 6K − 1, hence it
must have some prime divisor ≡ −1 (mod 6). Such contradiction leads to the fact that the set of primes of the
form 6k − 1 is not finite (aka infinite).

Page 12 / 222
1 CREATIVE TELESCOPING AND DFT

We may notice that a number of the form 6k + 1 is not compelled to have a prime divisor of the same form (for
instance, 55 = 5 · 11), so the previous argument is not well-suited for covering the 6k + 1 case, too. 2 We then take a
step back and a step forward: we provide an alternative proof of the infinitude of primes, then prove it can be adjusted
to prove the existence of infinite primes of the form 6k + 1, too. Let us recall the main statement in Analytic Number
Theory:

Theorem 17 (Euler’s product for the ζ function). If P is the set of prime numbers and s is a complex number with
real part greater than one,
Y −1 X
1 1
1− s = = ζ(s).
p ns
p∈P n≥1

 −1
Since 1 − p1s = 1 + p1s + p12s + . . ., such result is just the analytic counterpart of the Fundamental Theorem of
Arithmetics, stating that Z is a UFD.
In such framework the following argument is pretty efficient: if there were just a finite number of primes, given Euler’s
product the harmonic series would be convergent. But we know it is not, so there have to be an infinite number of
primes. The Theorem just outlined has an interesting generalization:

Theorem 18 (Euler’s product for Dirichlet’s L-functions). If P is the set of prime numbers, s is a complex number with
real part greater than one and χ(n) is a totally multiplicative function (i.e. a function such that χ(nm) = χ(n)χ(m)
holds for any couple (n, m) of positive integers), we have:

Y −1
χ(p) X χ(n)
1− = = L(s, χ).
ps ns
p∈P n≥1

We may consider a simple totally multiplicative function: the function that equals 1 over natural numbers ≡ 1 (mod 6),
−1 over natural numbers ≡ −1 (mod 6) and zero otherwise. Such function is the non-principal (Dirichlet) character
(mod 6). We may notice that:
X zn
= − log(1 − z)
n
n≥1

for any complex number z having modulus less than one. By applying the DFT with respect to a primitive sixth root
of unity:
X 1 1

π
L(1, χ) = − = √ .
6k + 1 6k + 5 2 3
k≥0

As an alternative:
X  XZ 1
1 1
L(1, χ) = − = (x6k − x6k+4 ) dx
6k + 1 6k + 5 0
k≥0 k≥0
1 1
1 − x4
Z X Z
= (1 − x4 ) x6k dx = dx
0 0 1 − x6
k≥0
1
1 + x2
Z
π
= dx = √
0 1 + x2 + x4 2 3
2 However,there is a light that never goes out: the infinitude of primes of the form 6k + 1 can be proved in a algebraic fashion by
considering cyclotomic polynomials. For instance, every prime divisor of Φ6 (3n) = 9n2 − 3n + 1 is a number of the form 6k + 1.

Page 13 / 222
Let us assume that prime numbers ≡ 1 (mod 6) are finite and consider Euler’s product for L(s, χ):
 −1  −1
Y 1 Y 1
L(s, χ) = 1− s 1+ s
p p
p≡1 (mod 6) p≡−1 (mod 6)
  −1  −1
Y 1 Y 1
≤ 1− 1+ s
p p
p≡1 (mod 6) p≡−1 (mod 6)
Y  −1
1
= C 1+ s
p
p6=2,3
Y −1
1
= D 1+ s
p
p
ζ(2s)
= D
ζ(s)

for some constant D > 0. From the divergence of the harmonic series we would have:

lim L(s, χ) = 0,
s→1+

but we already know that L(1, χ) > 0 (we computed its explicit value).
Such contradiction leads to the fact that the set of primes of the form 6k + 1 is infinite.
We underline some points in the proof just outlined:

• we used Euler’s product, analytic counterpart of the Fundamental Theorem of Arithmetics, for studying the
distribution of primes in the arithmetic progressions (mod 6). It looks highly unlikely that there is just a finite
number of primes ≡ 1 (mod 6) and infinite primes ≡ −1 (mod 6), so we just need to show that such awkward
“imbalance” does not really occur;

• through the DFT, we may compute the value of L(1, χ) (with χ being the non-principal character (mod 6)) in
a explicit way, and check it is a positive number;

• from Euler’s product we have that the previous “imbalance” would lead to L(1, χ) = 0. It is not so, hence there
is no “imbalance”.

At last, we mention that both the DFT and the existence of Dirichlet’s characters are instances of Pontryagin’s duality
(https://en.wikipedia.org/wiki/Pontryagin_duality). The DFT is also of great importance for algorithms, since
it gives methods for the fast multiplication of polynomials (or integers): in such a context it is also known as FFT
(Fast Fourier Transform). Summarizing:

• The key idea is to exploit interpolation/extrapolation. A polynomial with degree m is fixed by its values at
m + 1 distinct points. If we assume to have a(x) and b(x) and we need to compute c(x) = a(x) · b(x), we may. . .

• compute in a explicit way the values of a and b at the 2n -th roots of unity, then the values of c at such points. . .

• and compute the coefficients of c(x) through such values. Nicely, both the evaluation process and the extrapola-
tion process are associated with a matrix-vector-product problem, where the involved matrix is Vandermonde’s
matrix given by the 2n -th roots of unity;

• the structure of such matrix depends in a simple way from the structure of Vandermonde’s matrix given by the
2n−1 -th roots of unity, hence the needed matrix-vector-products can be computed through a recursive, divide et
impera approach, with a significant improvement in computational costs.

For further details, please see http://en.wikipedia.org/wiki/Cooley-Tukey_FFT_algorithm.

Page 14 / 222
1 CREATIVE TELESCOPING AND DFT

The formulas of Koecher, Leshchiner and Bailey-Borwein-Bradley.


We have studied how to use the creative telecoping machinery for producing fast-convergent series representing
ζ(2) or ζ(3). Three formulas provide a wide generalization of such statement. The first one is due to Koecher
(1979):
k−1 
1 X (−1)k+1 5k 2 − a2 Y a2

X
2n
X 1
ζ(2n + 3)a = =  · 2 1− 2 ,
n≥0 k≥1
k(k 2 − a2 ) 2
k≥1
k 3 2k
k
k − a2 m=1 m

the second one is due to Leschiner (1981):

X (−1)n+1 k−1 
X 3k 2 + a2 Y a2
 
1 1X 1
1− ζ(2n + 2)a2n = = · 1 − ,
22n+1 n2 − a2 k 2 2k k 2 − a2 m=1 m2

2 k
n≥0 n≥1 k≥1

the third one is due to Bailey-Borwein-Bradley (2006):


k−1
Y m2 − 4a2
X X 1 X 1
ζ(2n + 2)a2n = 2 2
=3 2k
.
k −a (k − a ) m=1 m2 − a2

2 2
n≥0 k≥1 k≥1 k

They hold for any a ∈ (−1, 1): by comparing the coefficients of ah in the LHS/RHS one gets that ζ(m), for
any m ≥ 2, can be represented as a fast-convergent series involving central binomial coefficients and generalized
harmonic numbers. It is straightforward to recover the well-known results
X 1 5 X (−1)n+1
ζ(2) = 3 2n , ζ(3) =
n3 2n
 
n2 n
2 n
n≥1 n≥1

together with the lesser known results


n
36 X 1 X (−1)k 1X 2n X 1
ζ(4) = , G= = 2n .
k 4 2k (2k + 1)2

17 k
2 (2n + 1) n
2k + 1
k≥1 k≥0 n≥0 k=0

The last identity can also be proved by computing integrals involving the arcsin2 (x) function or by computing the
1
binomial transform of (2k+1) 2.

Exercise 19. Prove the following identity:


 
X 1 log(3) − log(2)
arctanh = ,
n3 2
n≥2

1
trivially leading to ζ(3) < 1 + 2 log 32 .

Exercise 20. By exploiting Euler’s product prove that


X 1 ζ(s)3
∀s > 1, = .
lcm(m, n)s ζ(2s)
m,n≥1

αk
Proof. For any M ∈ N+ of the form M = pα 1 · · · pk , the number of solutions of lcm(n, m) = M
1

is given by (2α1 + 1) · · · (2αk + 1). It follows that the given series equals
X 1 Y
(2νp (M ) + 1)
Ms
M ≥1 p|M

Page 15 / 222
Q
and since M 7→ p|M (2νp (M ) + 1) clearly is a multiplicative function, by Euler’s product

X 1 Y 3 5 7
 Y s s
p (p + 1) Y 1 − p12s ζ(s)3
= 1 + + + + . . . = = 3 = .
lcm(m, n)s ps p2s p3s (ps − 1)2 ζ(2s)
 
m,n≥1 p∈P p∈P p∈P 1 − 1s
p

Exercise 21. Prove the following identity:


2n
1
log2 (1 − x)

H2n−1 π3
Z X
√ =π n
= + π log2 (2).
−1 1−x 2 n4n 3
n≥1

Exercise 22. The analytic continuation for the Riemann ζ function to the region Re(s) > 0 gives us the identity

1
 X 1
ζ 2 = −2 + √ √ √ .
k≥1
k( k + k + 1)2

Use creative telescoping to show that:

1
 3 1X 1
ζ 2 = − + √ √ √ √
2 2 k k + 1( k + k + 1)3
k≥1
35 7 X 1 1 X 1
= − − √ √ √ √ + √ √ √ √ .
24 96 k k(k + 1) k + 1( k + k + 1) 3 96 k k(k + 1) k + 1( k + k + 1)7
k≥1 k≥1

1 1
Q 
Exercise 23. Find a rational approximation of r≥1 1+ 2r within 100 from the exact value.

Proof. We may check that for any x ∈ (0, 1) we have


3
1 + 2x

3
1+x>
1 + x3

hence it follows that


3 3
Y 1 + 3·21r−1
  
1 3 5 Y 15 1 50
1+ r > · · 1 = 1+ > .
2 2 4 1 + 3·2 r 8 12 21
r≥1 r≥3

Similarly
1 + 2x + 34 x2 + 87 x3
1+x<
1 + x + 31 x2 + 17 x3
implies
Y 1

15

4 8

74
1+ < 1 + 2x + x2 + x3 < .
2r 8 3 7 x=1/8 31
r≥1

The difference between the upper bound and the lower bound is already less than 7 · 10−3 .

Page 16 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

Exercise 24. Let Ω(n) be the function returning the number of prime factors of n, counted according to their
multiplicity, such that, for instance, Ω(24) = Ω(23 · 3) = 3 + 1 = 4. Prove that
X 1 7π 2
2
= .
n 60
n≥1
Ω(n) is even

Proof. By Euler’s product we have

Y −1 Y −1 X (−1)Ω(n)


1 X 1 1
1− 2 = , 1+ 2 = ,
p n2 p n2
p∈P n≥1 p∈P n≥1

hence by averaging/DFT we have


7π 2
 
X 1 1 ζ(4)
2
= ζ(2) + = .
n 2 ζ(2) 60
n≥1
Ω(n) is even

2 Convolutions and ballot problems


We start this section by recalling a well-known identity:

Lemma 25.
n  2  
X n 2n
= .
k n
k=0

Proof. The first proof we provide is based on a double counting argument. Let us assume to have a parliament
with n politicians in the left wing and n politicians in the right wing, and to be asked to count how many committees
with n politicians we may have. It is pretty clear such number is given by 2n

n , i.e. the number of subsets with n
elements in a set with 2n elements. On the other hand, we may count such committees according to the number of
politicians from the left wing (k ∈ [0, n]) in them. There are nk ways for choosing k politicians of the left wing from


the n politicians we have. If in a committee there are k politicians from the left wing, there are n − k politicians from
n
= nk for selecting them. It follows that:
 
the right wing, and we have n−k
  n    Xn  2
2n X n n n
= =
n k n−k k
k=0 k=0

as wanted. The second proof is based on the fact that


n
def
X
(f ∗ g)(n) = f (k) · g(n − k)
k=0

is the convolution between f and g.


Since nk is the coefficient of xk in the Taylor series of (1 + x)n at the origin3 ,


n    
X n n
(1 + x)n = xk =⇒ [xk ](1 + x)n =
k k
k=0

3 The notation [xk ] f (x) stands for the coefficient of xk in the Taylor/Laurent series of f (x) at the origin.

Page 17 / 222
implies that:
n    Xn  
X n n k n n−k n n n n n 2n 2n
= [x ](1 + x) · [x ](1 + x) = [x ] [(1 + x) · (1 + x) ] = [x ](1 + x) = .
k n−k n
k=0 k=0

The second approach leads to a nice generalization of the first identity in the current section:

Theorem 26 (Vandermonde’s identity).


X a b  a + b
= .
j k n
j+k=n

In the introduced convolution context the last identity simply follows from the trivial (1 + x)a · (1 + x)b = (1 + x)a+b .
We may notice that    
X X X
 c(n) xn  ·  d(n) xn  = (c ∗ d)(n) xn
n≥0 n≥0 n≥0

is the Cauchy product between two power series. The interplay between analytic and combinatorial arguments

allows us to prove interesting things. For instance we may consider the function f (x) = 1 − x = (1 − x)1/2 , analytic
in a neighbourhood of the origin. It is not difficult to compute its Taylor series by the extended binomial theorem.
Moreover f (x)2 = (1 − x) has a trivial Taylor series, hence by defining a(n) as the coefficient of xn in the Taylor series
of f (x), (a ∗ a)(n) always takes values in {−1, 0, 1}.
1 1 1
 
n 2 · 2 − 1 · ... · 2 − n + 1
X 1/2 X
1/2 n n
(1 − x) = (−1) x = (−1) xn
n n!
n≥0 n≥0
X 1 · (1 − 2) · . . . (1 − 2n + 2) n
= (−1)n x
2n n!
n≥0
X (2n − 1)!! X (2n)!
= 1− xn = 1 − xn
(2n − 1) · (2n)!! (2n − 1) · (2n)!!2
n≥1 n≥1
X 2n xn
= 1−
n 4n (2n − 1)
n≥1

By differentiating with respect to x,


1 X 2n xn
√ =
1 − x n≥0 n 4n

1 1 2n

follows, and since 1−x = 1 + x + x2 + x3 + . . ., if we set a(n) = 4n n we have (a ∗ a)(n) = 1, i.e.:

Lemma 27.
n   
X 2k 2n − 2k
= 4n .
k n−k
k=0

We may also notice that !


n
1 X X X X
a(n)xn = (a ∗ 1)(n) xn = a(k) xn
1−x
n≥0 n≥0 n≥0 k=0

Page 18 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

where the LHS equals


1 d

1
 X 2n nxn−1 X 2n + 2 2n + 2
√ =2 √ =2 = xn .
(1 − x) 1 − x dx 1−x n 4n n+1 4n+1
n≥1 n≥0

By comparing the last two RHSs we have an alternative proof of an identity claimed by Exercise 6:
N    
X 2n 1 2N + 2 N + 1
= .
n=0
n 4n N + 1 22N +1

Exercise 28 (Stars and bars). Prove that for any k ∈ N we have:

1 X n + k 
= xn .
(1 − x)k+1 k
n≥0

Proof. We may tackle this question both in a combinatorial and in an analytic way. The coefficient of xn in the
product of (k + 1) terms of the form (1 + x + x2 + . . .) is given by the number of ways for writing n as the sum of k + 1
natural numbers. By stars and bars we know the number of ways for writing n as the sum of k + 1 positive natural
numbers is n−1

k and it is not difficult to finish from there. As an alternative, we may proceed by induction on k.
The claim is trivial in the k = 0 case, and since
1 1 1 X n + k  
= · = ∗ 1 xn
(1 − x)k+2 1 − x (1 − x)k+1 k
n≥0

the inductive step follows from the hockey stick identity


n    
X k+j n+k+1
= .
j=0
k k+1

Exercise 29. Given the sequences of Fibonacci and Lucas numbers {Fn }n≥0 and {Ln }n≥0 ,
prove the following convolution identity:
n
X nLn − Fn
Fk Fn−k = .
5
k=0

Proof. Since Fibonacci numbers fulfill the relation Fn+2 = Fn+1 + Fn , by defining their generating function as
X
f (x) = Fn x n
n≥0

we have that (1 − x − x2 ) · f (x) is a linear polynomial (a similar idea leads to the Berkekamp-Massey algorithm).
On the other hand, if
X ax + b
f (x) = Fn x n =
1 − x − x2
n≥0

b = 0 has to hold to grant f (0) = F0 = 0 and a = 1 has to hold to grant f 0 (0) = F1 = 1. It follows that:
  √ √
x 1 1 1 1+ 5 1− 5
f (x) = =√ − , ϕ= , ϕ̄ =
1 − x − x2 5 1 − ϕx 1 − ϕ̄x 2 2
1 1
and by computing the Taylor series (that is a geometric series) of 1−ϕx and 1−ϕ̄x we immediately recover
Binet’s formula
ϕn − ϕ̄n
Fn = √ .
5

Page 19 / 222
The identity Ln = ϕn + ϕ̄n has a similar proof. By starting the convolution machinery:
n  2  
X x 1 n 1 1 2
Fk Fn−k = [xn ] = [x ] + −
1 − x − x2 5 (1 − ϕx)2 (1 − ϕ̄x)2 (1 − ϕx)(1 − ϕ̄x)
k=0

and the claim follows from simple fraction decomposition. To find a combinatorial proof is an exercise left to the
reader: we recall that Fibonacci numbers are related to subsets of {1, 2, . . . , n} without consecutive elements.

A note in mathematical folklore: Alon’s Combinatorial Nullstellensatz has further tightened the interplay
between combinatorial arguments and generating functions arguments. We invite the reader to delve into the
bibliography to find a generalization of Cauchy-Davenport’s theorem, once known as Kneser’s conjecture, now
known as Da Silva-Hamidoune’s Theorem:
def
Theorem 30 (Da Silva, Hamidoune). If A ⊆ Fp and A ⊕ A = {a + a0 : a, a0 ∈ A, a 6= a0 }, we have:

|A ⊕ A| ≥ min(p, 2|A| − 3).

The convolution machinery applies very well to another kind of coefficients given by Catalan numbers. We introduce
them in a combinatorial fashion, assuming to have two people involved in a ballot and to check the votes one by one.

Theorem 31 (Bertrand’s ballot problem). If the winning candidate gets A votes and the loser gets B votes (so we are
clearly assuming A > B), the probability that the winning candidate had the lead during the whole scrutiny equals:

A−B
A+B

Proof. The final outcome is so simple due to a slick symmetry argument, applied in a double counting framework:
instead of trying to understand what happens or might happen once a single vote is checked, it is more effective to
consider which orderings of the votes favour A or not. Let us consider just the first vote: if it is a vote for B, at some
point of the scrutiny there must be a tie, since the winning candidate is A. If the first vote is for A and at some point
of the scutiny there is a tie, by switching the votes for A and for B till the tie we return in the previous situation. It is
B
pretty clear that the probability the first vote is a vote for B is A+B . It follows that the probability of a tie happening
2B
during the scrutiny is A+B , and the probability that A always leads is:

2B A−B
1− = .
A+B A+B

Theorem 32 (Catalan numbers). The number of strings made by n characters 0 and n characters 1,
with the further property that no initial substring has more 1s than 0s, is:
 
1 2n
Cn = .
n+1 n

Proof. Any string with the given property can be associated (in a bijective way) with a path on a n×n grid, starting in
the bottom left corner and ending in the upper right corner, made by unit steps towards East (for each 1 character) or
North (for each 0 character) and never crossing the SW-NE diagonal (this translates the substrings constraint). These
paths can be associated in a bijective way with ballots that end in a tie, in which at every moment of the scrutiny the

Page 20 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

votes for B are ≤ than the votes for A. If a deus ex machina adds an extra vote for A before the scrutiny begins, we
have a situation in which A gets n + 1 votes, B gets n votes and A is always ahead of B. There are 2n+1 = 2n+1
 
n n+1
possible scrutinies in which A gets n + 1 votes and B gets n votes: by the previous result (Bertrand’s ballot problem)
the number of the wanted strings is given by:
     
(n + 1) − n 2n + 1 1 2n + 1 1 2n
= = .
(n + 1) + n n 2n + 1 n n+1 n
For a slightly different perspective on the same subject, the reader is invited to have a look at Josef Rukavicka’s
“third proof” on the Wikipedia page about Catalan numbers.

Theorem 33. Given two natural numbers a and b with a ≥ b, the following identity holds:
b      
X 1 2k a−b a + b − 2k 1+a−b a+b+1
= .
k + 1 k a + b − 2k b−k 1+a+b b
k=0

Proof. It is enough to count scrutinies for a ballot between two candidates A and B, with A getting a votes, B getting
b votes and A being always ahead of B, without excluding the chance of a tie at some point. We get the RHS by
adding an extra vote for A before the scrutiny begins and mimicking the previous proof. On the other hand we may
count such scrutinies according to the last moment in which we have a tie. If the last tie happens when 2k votes have
been checked, we simply need to assign a − k votes for A and b − k votes for B: what happens before the tie can be
accounted through Catalan numbers and what happens next through Bertrand’s ballot problem. This leads to the
LHS.
 Pr
The last identity is a particular case of a remarkable generalization of Vandermonde’s identity m+n = k=0 m
 n 
r k r−k :

Theorem 34 (Rothe-Hagen).
n      
X x x + kz y y + (n − k)z x+y x + y + nz
= .
x + kz k y + (n − k)z (n − k) x + y + nz n
k=0

Proof. This identity is usually proved through generating functions and that approach is not terribly difficult. We may
point that a purely combinatorial proof is also possible, by following the lines of the previous proof. It is enough to
slightly modify the constraint at any point, the votes for B are ≤ than the votes for A by replacing it with something
involving the ratio of such votes. This is surprising both for experts and for newbies: Micheal Spivey has written an
interesting lecture about it on his blog.

The following problems are equivalent:

Exercise 35 (Balanced parenthesis). How many strings with 2n characters over the alphabet Σ = {(, )} have as
many open parenthesis as closed parenthesis, and in every initial substring the number of closed parenthesis is
always ≤ the number of open parenthesis?

Exercise 36 (Sub-diagonal paths). Let us consider the paths from (0; 0) to (n; n) where each step is a unit step
towards East or North. How many such paths belong to the region y ≤ x?

Exercise 37 (Triangulations of a polygon). Given a convex polygon, a triangulation of such polygon is a


partition in almost-disjoint triangles, with the property that every triangle has its vertices on the boundary of the
original polygon. How many triangulations are there for a convex polygon with n + 2 sides?

Page 21 / 222
Exercise 38 (Complete binary trees). A tree is a connected, undirected and acyclic graph. It is said binary
and complete if each vertex has two neighbours (in such a case it is an inner node) or no neighbours (in such a
case it is a leaf ). How many complete binary trees with n inner nodes are there?

It is not difficult to prove the above claiming by exhibiting three combinatorial bijections:

TP ←→ CBT ←→ BP ←→ SDP.

In each case the answer is given by the Catalan number


 
1 2n
Cn =
n+1 n
where the sub-diagonal paths interpretation proves the identity
n
X
Cn+1 = Ck Cn−k
k=0

Cn xn ,
P
in a very straightfoward way. Such identity is a convolution formula: if we set c(x) = n≥0
we have c(x) = 1 + x · c(x)2 . By solving such quadratic equation in c(x) we get:

1 − 1 − 4x
c(x) =
2x
(the square root sign is chosen in such a way that c(x) is continuous at the origin) hence the coefficients of the
power series c(x) can be computed from the extended binomial theorem, extended since it is applied to (1 − x)α with
α = 21 6∈ N.

Exercise 39. Prove that:


n  π/2
16n
 
2n − 2k
Z
X 2k 1 n
= 2n = 4
 cos(x)2n+1 dx.
n−k k 2k + 1 (2n + 1) n 0
k=0

The last identity is not trivial at all, and it has very deep consequences. A possible proof of such identity (that is
not the most elementary one: to find an elementary proof is an exercise we leave to the reader) exploits a particular
class of orthogonal polynomials. We have not introduced the L2 space yet, nor the usual techniques for dealing with
square-integrable objects, so such “advanced” proof is postponed to the end of this section, in a dedicated box. What
we can say through the convolution machinery is that the above identity is related to a Taylor coefficient in the product
1
between arcsin(x) and its derivative √1−x 2
. By termwise integration it implies:

1 X (4x2 )n
arcsin2 (x) =
n2 2n

2 n
n≥1

1
and by evaluating the last identity at x = 2 we get that:

π2
 
X 3 2 1
2n = 6 arcsin =

n2 n
2 6
n≥1

2 2
where the LHS equals ζ(2) by creative telescoping, as seen in the previous section, where we proved ζ(4) = 5 ζ(2)
too. It follows that:

Page 22 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

X 1 π2 X 1 π4
ζ(2) = = , ζ(4) = = .
n2 6 n4 90
n≥1 n≥1

We have just solved Basel problem through a very creative approach, i.e. by combining creative telescoping with
convolution identities for Catalan-ish numbers. Plenty of other approaches are presented in a forthcoming section.

Theorem 40 (Euler’s acceleration technique).


n
∆n a0
 
X
n
X X n
(−1) an = (−1)n n+1 , n
∆ a0 = (−1) k
an−k .
2 k
n≥0 n≥0 k=0

This identity is simple to prove and it is really important in series manipulation and numerical computation: for
instance, it is the core of Van Wijngaarden’s algorithm for the numerical evaluating of series with alternating signs.
Let us study a consequence of Euler’s acceleration technique, applied to:
X (−1)n XZ 1 Z 1
dx π
= (−1)n x2n dx = 2
= arctan(1) =
2n + 1 0 0 1+x 4
n≥0 n≥0

We may compute ∆n a0 in a explicit way:


n   n  
Z 1X
X n 1 n
∆n a0 = (−1)n−k = (−1)n (−x2 )k dx
k 2k + 1 0 k
k=0 k=0
Z 1 n
4
= (−1)n (1 − x2 )n dx = (−1)n 2n

0 (2n + 1) n

and derive that:


2n+1 X
π=
(2n + 1) 2n

n≥0 n

where the main term of the last series behaves like 21n nπ for n → +∞, with a significative boost for the convergence
p

speed of the original series 4 . About the series defining ζ(2), Euler’s acceleration technique leads to:
X Hn ζ(2) π2
= = .
n2n 2 12
n≥1

n
P
By recalling Hn = m≥1 m(m+n) , the last identity turns out to be equivalent to

1
π2 − log(1 − x2 )
Z
= dx,
12 0 x
2
that is trivial by applying termwise integration to the Taylor series at the origin of − log(1−x
x
)
.

4 Byapplying Euler’s acceleration technique to the Taylor series of the arctangent function we get something equivalent to the functional
identity  
x
arctan x = arcsin √ .
1 + x2

Page 23 / 222
A convolution involving the Riemann ζ function.
We now present a result about an ubiquitous Euler sum:

Theorem 41. For any q ∈ {3, 4, 5, . . .} the following identity holds:


∞ q−2
X Hm q+2 1X
= ζ(q + 1) − ζ(q − j)ζ(j + 1).
m=1
mq 2 2 j=1

Proof.
k
X
ζ(k + 2 − j)ζ(j + 2)
j=0
∞ X∞ X
k
X 1
(expand ζ) =
m=1 n=1 j=0
mk+2−j nj+2
(pull out the terms for m = n and use the ∞ 1 1
X 1 mk+1
− nk+1
formula for finite geometric sums on the rest) = (k + 1)ζ(k + 4) + 1 1
m2 n2 −
m,n=1 m n
m6=n

X 1 1
(simplify terms) = (k + 1)ζ(k + 4) + −
m,n=1
nmk+2 (n − m) mnk+2 (n − m)
m6=n
+∞ X +∞
X 1 1
(exploit symmetry) = (k + 1)ζ(k + 4) + k+2 (n − m)
− k+2 (n − m)
m=1 n=m+1
nm mn
∞ X ∞
X 1 1
(n 7→ n + m and switch sums) = (k + 1)ζ(k + 4) + 2 k+2 n

m=1 n=1
(n + m)m m(n + m)k+2 n
1 1 1
By exploiting mn = n(m+n) + m(m+n) we get:
∞ X
∞   ∞ X
∞  
X 1 1 X 1 1
(k + 1)ζ(k + 4) + 2 k+3
− −2 +
m=1 n=1
m n (m + n)mk+3 m=1 n=1
m(n + m) k+3 n(n + m)k+3
 
1 1 1 1
P
and since Hm = n≥1 n − n+m , by exploiting the symmetry of m(n+m)k+3
+ n(n+m)k+3
we get:
∞ ∞ X ∞
X Hm X 1
(k + 1)ζ(k + 4) + 2 k+3
− 4
m=1
m n=1 m=1
n(n + m)k+3
∞ ∞ ∞
X Hm X X 1
(m 7→ m − n) = (k + 1)ζ(k + 4) + 2 k+3
− 4 k+3
m=1
m n=1 m=n+1
nm

X Hm ∞ ∞
XX 1
(reintroducing terms ) = (k + 1)ζ(k + 4) + 2 k+3
−4 + 4ζ(k + 4)
m=1
m n=1 m=n
nmk+3
∞ ∞ X m
X Hm X 1
(switching sums) = (k + 5)ζ(k + 4) + 2 k+3
− 4 k+3
m=1
m m=1 n=1
nm
∞ ∞
X Hm X Hm
= (k + 5)ζ(k + 4) + 2 k+3
− 4 k+3
m=1
m m=1
m

X Hm
(combining sums) = (k + 5)ζ(k + 4) − 2
m=1
mk+3
Letting q = k + 3 and reindexing j 7→ j − 1 yields
q−2 ∞
X X Hm
ζ(q − j)ζ(j + 1) = (q + 2)ζ(q + 1) − 2
j=1 m=1
mq

and the claim is proved.

Page 24 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

Exercise 42. Do we get something interesting (like an approximated functional identity for the ζ function)
from the previous convolution identity, by recalling that
1 1 1 1
Hm = log(m) + γ + − 2
+ 4
− + ...
2m 12m 120m 252m6
and that X log m X 1 X
q
= −ζ 0 (q) = Λ(d)
m mq
m≥1 m≥1 d|m

 
1 1
P
Exercise 43. By exploiting Hm = n≥1 n − n+m and symmetry prove that

X Hm X 1
= 2 ζ(3) = 2 .
m2 m3
m≥1 m≥1

Proof.
log2 (1 − x)
X Hm Z 1 X m+n−1 Z 1
XX 1 x
= = dx = dx
m2 (m + n)mn 0 mn 0 x
m≥1 m≥1 n≥1 m,n≥1

by the substitution x 7→ 1 − x turns into

1
log2 (x) XZ 1
Z X 2
dx = xn log2 (x) dx = .
0 1−x 0 (n + 1)3
n≥0 n≥0

Exercise 44. By generalizing the previous approach prove that


X Hn2
= 3 ζ(3).
n(n + 1)
n≥1

Vandermonde’s identity and Bessel functions. Bessel functions are important mathematical functions: they
are associated with the coefficients of the Fourier series of some inverse trigonometric functions and they arise in
the study of the diffusion of waves, like in the vibrating drum problem. Bessel functions of the first kind with
integer order can be simply defined by giving their Taylor series at the origin:
X (−1)l
Jn (z) = z 2l+n
22l+n l!(m + l)!
l≥0

from which it is trivial that Jn (z) is an entire function, a solution of the differential equation z 2 f 00 + zf 0 + (z 2 −
n2 )f = 0 and much more. In this paragraph we will see how Vandermonde’s identity plays a major role in dealing
with the square of a Bessel function of the first kind.

Page 25 / 222
Exercise 45. Prove the identity:
Z π
2 2
Jn2 (z) = J2n (2z cos(θ))dθ.
π 0

Proof. Let us try the brute-force approach. We have:


X (−1)a  x 2a+n
Jn (x) =
a!(a + n)! 2
a≥0

hence:
X X (−1)m (x/2)2m+2n
Jn2 (x) =
a!b!(a + n)!(b + n)!
m≥0 a+b=m

where:
X 1 1 X mm + 2n
=
a!b!(a + n)!(b + n)! m!(m + 2n)! a b+n
a+b=m a+b=m
1
= [xm+n ](1 + x)m (1 + x)m+2n
m!(m + 2n)!
 
1 2m + 2n
= (♣)
m!(m + 2n)! m + n

leads to:
X (−1)m (x/2)2m+2n 2m + 2n
Jn2 (x) = . (♥)
m!(m + 2n)! m+n
m≥0
R π/2
Since π2 0 cos2h (θ) dθ = 41h 2h

h follows from De Moivre’s formula, in order to prove the claim it is enough to
expand J2n (2z cos θ) as a power series in 2z cos θ, perform terwmwise integration and exploit (♥). The claim is
ultimately a consequence of Vandermonde’s identity proved in (♣).

This technique also shows that the Laplace transform (an important tool we will introduce soon) of J02 (x) is
related to the complete elliptic integral of the first kind (another object we will study in a forthcoming section)
through the identity  
2
 2 4
L J0 (s) = K − 2 .
πs s

Exercise 46. Prove that the inequality


4n
 
2n

n n+1
is a trivial consequence of the Cauchy-Schwarz inequality.


1+ 5
Exercise 47. Prove that if ϕ is the golden ratio 2 , we have:

6! · log2 ϕ < 167.


P (−1)n+1
Hint: it might be useful to consider the rapidly convergent series n≥1 2n2 (2n) .
n

Page 26 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

Exercise 48. Prove that:


X π2

1 1
+ = .
(6n + 1)2 (6n + 5)2 9
n≥0

Exercise 49. Prove that:


π/2
5π 2
Z  
1 dθ
log 1 + sin θ = .
0 2 sin θ 72

346
Exercise 50. Prove that log(3) > 315 follows from computing/approximating the integral
1
x4 (1 − x2 )2
Z
dx.
0 4 − x2

There is another kind of convolution, that is known as multiplicative convolution or Dirichlet’s convolution. We say
that a function f : Z+ → C is multiplicative when gcd(a, b) = 1 grants f (ab) = f (a) · f (b). A multiplicative function
has to fulfill f (1) = 1, and since Z is a UFD the values of a multiplicative function are fixed by the values of such
function over prime powers. Common examples of multiplicative functions are the constant 1, the divisor function
d(n) = σ0 (n) and Euler’s totient function ϕ(n). Given two multiplicative functions f, g : Z+ → C, their Dirichlet
convolution is defined through X n
(f ∗ g)(n) = f (d) · g .
d
d|n

It is a simple but interesting exercise to prove that the convolution between two multiplicative functions still is a mul-
tiplicative function. Just like additive convolutions are related to products of power series, multiplicative convolutions
are related with products of Dirichlet series. If we state that
X f (n)
L(f, s) =
ns
n≥1

is the Dirichlet series associated with f , the following analogue of Cauchy’s product holds:
X (f ∗ g)(n) X X f (d) g n

d
L(f ∗ g, s) = = = L(f, s) · L(g, s).
ns ns
n≥1 n≥1 d|n

The main difference between additive and multiplicative convolutions is that in the multiplicative context, given f (n)
P
and H(n) = d|n h(d), we can always find a multiplicative function g such that f ∗g = H, and such function is unique.
There is no additive analogue of Möbius inversion formula, allowing us to solve such problem. The extraction process
of a coefficient from a given power/Dirichlet series is similar and relies on the residue theorem, applied to the original
function multiplied by x1h or to the Laplace transform of the original function. Let us see how to use this machinery
for solving actual problems.

Exercise 51. Prove that for any n ∈ Z+ the following identity holds:
X
n= ϕ(d).
d|n

Page 27 / 222
Proof. The usual combinatorial proof starts by considering the n-th roots of unity on the unit circle. Every n-th root
of unity is a primitive d-th root of unity for some d | n, and the number of primitive d-th roots of unity is exactly ϕ(d),
so the claim follows from checking that no overcounting or undercounting occurs. With the multiplicative convolution
machinery, we do not have to find a combinatorial interpretation for both sides, we just have to find the Dirichlet
series associated with both sides. In equivalent terms, in order to show that Id = ϕ ∗ 1 it is enough to compute:
X n X 1
L(Id, s) = s
= = ζ(s − 1),
n ns−1
n≥1 n≥1

X 1
L(1, s) = = ζ(s)
ns
n≥1

ζ(s−1)
then prove that L(ϕ, s) = ζ(s) . By Euler’s product:

Y 1− 1
Y ϕ(p) ϕ(p2 ) ϕ(p3 ) Y ps − 1

ps
L(ϕ, s) = 1+ + 2s + 3s + . . . = = 1 ,
p
ps p p p
ps − p p
1− ps−1

Y  Y −1
1 1 1 1
ζ(s) = L(1, s) = 1 + s + 2s + 3s + . . . = 1− s
p
p p p p
p

and the claim is proved.

1
The keen reader might observe the ζ(s) function played an import role in the previous proof, and ask about the
multiplicative function associated to such Dirichlet series. Well, by defining ω(n) as the number of distinct prime
factors of n and µ(n) as (
(−1)ω(n) if n is square-free
µ(n) =
0 otherwise

we have that µ (Möbius’ function) is a multiplicative function and


X µ(n) Y  µ(p)

1
L(µ, s) = = 1+ s =
ns p
p ζ(s)
n≥1

1
as wanted. Then the trivial 1 = ζ(s) · ζ(s) leads to the following convolution identity:
(
X 1 if n = 1
µ(d) =
d|n
0 otherwise.

Theorem 52 (Möbius inversion formula). If we have


X
f (n) = g(d)
d|n

then X n
g(n) = µ(d) · f
d
d|n

holds.

Proof. Let us denote with ε the multiplicative function that equals 1 at n = 1 and zero otherwise. Since f = g ∗ 1,

µ ∗ g = µ ∗ (f ∗ 1) = µ ∗ (1 ∗ f ) = (µ ∗ 1) ∗ f = ε ∗ f = f

as wanted, by just exploiting a ∗ b = b ∗ a and the associativity of ∗.

Page 28 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

Corollary 53. Y Y
F (n) = f (d) =⇒ f (n) = F (d)µ(n/d) .
d|n d|n

The last identity encodes an algebraic equivalent of the inclusion-exclusion principle. For instance, by denoting through
Φm (x) the m-th cyclotomic polynomial (i.e. the minimal polynomial over Q of a primitive m-th root of unity) we have
Y
xn − 1 = Φd (x)
d|n

and from Möbius inversion formula it follows that:


Y
Φn (x) = (xd − 1)µ(n/d) .
d|n

By comparing the degrees of the LHS and RHS we also get:


Xn X µ(d)
ϕ(n) = · µ(d) = n
d d
d|n d|n

corresponding to ϕ = Id ∗ µ. The explicit formula for cyclotomic polynomials has many interesting consequences,
for instance:
∀n > 1, Φn (0) = (−1)(µ∗1)(n) = 1
that also follows from the fact that Φn (x) is a palindromic polynomial (if ξ is a root of Φn , ξ −1 is a root of Φn too).
We also have
Φ0n (z) d X  n  dxd−1
= log Φn (z) = µ
Φn (z) dz d xd − 1
d|n

hence for any n > 1:

Φ0n (0) X  n  dxd−1


[z 1 ]Φn (z) = Φ0n (0) = Φn (0) = lim µ = −µ(n)
Φn (0) x→0 d xd − 1
d|n

since only the term d = 1 may provide a non-zero contribution to the limit. By putting together the following facts:

• Φn (z) is a palindromic polynomial with degree ϕ(n);

• by Vieta’s formulas, for a monic polynomial with degree q the sum of the roots
equals the opposite of the coefficient of xq−1 ;

• we know where the roots of Φn (z) lie

it follows that µ(n) can be represented as an exponential sum, too:


 
X 2πim
µ(n) = exp .
n
1≤m≤n
gcd(m,n)=1

This sum is a particular case of Ramanujan sum. Thanks to Srinivasa Ramanujan we also know that the σ3 function,
σ3 (n) = d|n d3 , fulfills at the same time an additive convolution identity (due to the fact that the Eisenstein series
P

E4 (τ ) depends on σ3 ) and a multiplicative convolution identity:

n−1
X σ7 (n) − σ3 (n) X n
σ3 (m) σ3 (n − m) = , σ3 (d) σ3 = (1 ∗ 1 ∗ . . . ∗ 1 ∗ 1) = σ7 (n).
m=1
120 d | {z }
d|n
8 times

Thanks to Giuseppe Melfi and his work on the modular group Γ(3) we also know that:
n
X 1
σ1 (3k + 1)σ1 (3n − 3k + 1) = σ3 (3n + 2).
9
k=0

Page 29 / 222
Exercise 54. Prove that for any M ∈ {3, 4, 5, . . .} the following identity holds:
X  πn  ϕ(M ) − µ(M )
sin2 = .
M 2
1≤n≤M
gcd(n,M )=1

Hint: convert the LHS into something depending on the roots of a cyclotomic polynomial,
then recall the representation of the Möbius function as an exponential sum.

A proof of the identity presented in Exercise 32.


If we define the sequence of Legendre polynomials through Rodrigues’ formula
n  2
1 dn 2 n 1 X n
Pn (x) = (x − 1) = (x − 1)n−k (x + 1)k
2n n! dxn 2n k
k=0

we have that these polynomials, as a consequence of integration by parts,


give a orthogonal and complete base of L2 (−1, 1) with respect to the usual inner product:
Z 1
2δ(n, m)
Pn (x) Pm (x) dx = .
−1 2n + 1

Additionally, their generating function is pretty simple:


1 X
√ = Pn (x)tn .
1 − 2xt + t2
n≥0

In a similar way, the shifted Legendre polynomials P̃n (x) = Pn (2x − 1)


can be defined through Rodrigues’ formula
n  
1 dn

2 n n
X n n+k
P̃n (x) = (x − x ) = (−1) (−x)k
n! dxn k k
k=0

and they give a complete and orthogonal base of L2 (0, 1) with respect to the usual inner product:
Z 1
δ(m, n)
P̃n (x)P̃m (x) dx = .
0 2n + 1

By integration by parts it follows that:


Z 1 Z 1
n 1 2n+1 n!2
x Pn (x) dx = n (1 − x2 )n dx = .
−1 2 −1 (2n + 1)!

In particular: Z 1 X Z 1 
dx
√ = xn Pn (x) dx tn ,
−1 1 − 2x2 t + x2 t2 n≥0 −1
q
4 arcsin 2t X 2n+1 n!2
p = tn ,
t(2 − t) (2n + 1)!
n≥0

X (4t2 )n
2 arcsin2 (t) = .
n≥1
n2 2n
n

Page 30 / 222
2 CONVOLUTIONS AND BALLOT PROBLEMS

Exercise 55. Prove that for any x ∈ (0, 1) we have:


1 √ X X 2n + 1
√ = 2 Pn (2x − 1), − log(1 − x) = 1 + Pn (2x − 1).
1−x n(n + 1)
n≥0 n≥1

The acceleration of the ζ(2) series from another perspective.


We already mentioned that n≥1 n12 = n≥1 n2 32n can be proved by creative telescoping. In this paragraph
P P
(n)
we prove it is a consequence of a change of variable in a integral, in particular the tangent half-angle substitution
(sometimes known as Weierstrass’ substitution, apparently for no reason) x = 2 arctan 2t , sending (0, π/2) into

dx dt
(0, 1) and sin x in t . Let us set
Z π/2  
1 2 dx
I=− log 1 − sin x .
0 4 sin x
By expanding − log 1 − 14 sin2 x as a Taylor series in sin x we have that


X 1 Z π/2 X 1 1X 1
2n−1
I= sin(x) dx = 2n−2
=
n4n 0 n 2n

4n(2n − 1) n−1
2 2
n
n≥1 n≥1 n≥1

and by applying the tangent half-angle substitution we get:


Z 1  2 !
t dt
I= − log 1 −
0 1 + t2 t
 2
t
where the rational function 1 − 1+t 2 can be written in terms of products and ratios
m
of polynomials of the form 1 − t . Eureka, since:
Z 1
log(1 − tm ) X 1 1 mn−1
Z X 1 ζ(2)
Im = − = t dt = 2
=
0 t n 0 mn m
n≥1 n≥1

implies:
1
I = (I2 − 2I4 + I6 ) = ζ(2).
6

Exercise 56. Is it possible to prove the identity


X 1 5 X (−1)n+1
ζ(3) = =
3
n3 2n

n 2 n
n≥1 n≥1

through a change of variable in a integral?

The Taylor series of arcsin2 (x) from the Complex Analysis point of view.
The function f (z) = sin(z) is an entire function of the form z + o(z), hence it gives a conformal map
between two neighbourhoods of the origin (this crucial observation is the same leading to
the Buhrmann-Lagrange inversion formula). In particular,
I I
arcsin(x) 1 arcsin(z) 1 z
a2n−1 = [x2n−1 ] √ = √ dz = dz
1−x 2 2πi z 2n 1−z 2 2πi sin(z)2n

Page 31 / 222
and   Z 
z dz
a2n−1 = Res , z = 0 = −Res ,z = 0 .
sin(z)2n sin(z)2n
d 1
Since dz cot(z) = sin2 z
, the last integral can be computed through repeated integration by parts:
Z n−1 k  
dz X cot(z) Y 1
= − 1 +
sin(z)2n (2n) sin(z)2n−2k j=0 2n − 2j − 1
k=0

and for every m ≥ 1 we have:


     
cot z cos z 1
Res , z = 0 = Res , z = 0 = Res , z = 0 = 0,
sin2m (z) sin2m+1 (z) z 2m+1

so there is a single term really contributing to the value of the residue at the origin,
and since Res(cot(z), z = 0) = 1,
n 
4n
Z  
dz 1 Y 1 (2n)!!
−Res , z = 0 = 1 + = =
sin(z)2n 2n 2n

2n j=0 2n − 2j − 1 (2n) · (2n − 1)!! n

from which we get:


arcsin(x) X 4n X (4x2 )n
√ =  2n−1 ,
2n x arcsin2 (x) = .
1 − x2 n≥1
2n n n≥1
2n2 2n
n

Exercise 57. Prove that the value of the rapidly convergent series

13 X (−1)n (2n + 1)!



8 n!(n + 2)!42n+3
n≥0

1+ 5
is the golden ratio 2 .

xk

k, prove that limx→+∞ efx(x)
√ = 1. Hint: compute the series of f (x)2
P
Exercise 58. Given f (x) = k≥0 k! x
p
and exploit the inequality 2z(1 − z) ≤ z(1 − z) ≤ 12 , holding for any z ∈ [0, 1].

Exercise 59. Prove that the series


X (−1)k bk/2c
X 2m (−1)m
2k + 1 m=0 m 4m
k≥0

is convergent to √1 log(1 + 2).
2

Proof.
X (−1)m 2m 2 π/2
Z
dθ 1
m
= 2
=√
4 m π 0 1 + cos θ 2
m≥0

and Dirichlet’s test ensure that the given series is convergent. If k is even (say k = 2n) we have
n 
2m (−1)m

X 1
m
= [xn ] √
m=0
m 4 (1 − x) 1 + x

Page 32 / 222
3 CHEBYSHEV AND LEGENDRE POLYNOMIALS

and if k is odd (say k = 2n + 1) we have the same identity, where [xn ]f (x) stands for the coefficient of xn in the
Maclaurin series of f (x). In particular the original series can be written as
Z 1
X 1 1 1−x
· [x2n ] √ = √ dx
(2n + 1)(2n + 2) (1 − x 2 ) 1 + x2
0 (1 − x 2 ) 1 + x2
n≥0
R1 1
since 0
x2n (1 − x) dx = (2n+1)(2n+2) . It turns out that the original series is just
Z 1

dx arcsinh(1) log(1 + 2)
√ = √ = √ .
0 (1 + x) 1 + x2 2 2

3 Chebyshev and Legendre polynomials

Lemma 60. For any n ∈ N there exists a polynomial Tn (x) ∈ Z[x] such that:

cos(nθ) = Tn (cos θ).

It is not difficult to prove the claim by induction on n. It is trivial for n = 0 ed n = 1, and by the cosine addition
formulas:
cos((n + 2)θ) + cos(nθ) = 2 cos(θ) cos((n + 1)θ)
such that:

Tn+2 (x) = 2x · Tn+1 (x) − Tn (x)

for any n ≥ 0. The Tn (x) polynomials are Chebyshev polynomials of the first kind
and they have many properties, simple to prove:

• (uniform boundedness)
∀x ∈ [−1, 1], |Tn (x)| ≤ 1

• (distibution of roots)
n  
Y (2k − 1)π
Tn (x) = 2n−1 x − cos
2n
k=1

• (orthogonality)
Z 1
Tn (x) Tm (x) π
√ dx = δ(m, n)(1 + δ(n, 0))
−1 1 − x2 2

• (a simple generating function)


1 − xt X
2
= Tn (x)tn
1 − 2xt + t
n≥0

• (an explicit representation)


1h p p i
Tm (x) = (x + i 1 − x2 )m + (x − i 1 − x2 )m .
2

Chebyshev polynomials of the second kind, Un (x), are similarly defined through sin((n+1)θ)
sin θ = Un (cos θ):
they share with Chebyshev polynomials of the first kind the recurrence relation Un+2 (x) = 2x Un+1 (x) − Un (x)
and similar properties:

Page 33 / 222
• (boundedness)
∀x ∈ [−1, 1], |Un (x)| ≤ (n + 1)

• (distribution of roots)
n  
Y kπ
Un (x) = 2n x − cos
n+1
k=1

• (orthogonality)
Z 1 p π
Un (x) Um (x) 1 − x2 dx = δ(m, n)
−1 2

• (a simple generating function)


1 X
2
= Un (x)tn
1 − 2xt + t
n≥0

• (an explicit representation)


bm/2c  
X m−r
Um (x) = (−1)r (2x)m−2r .
r=0
r

By combining Vieta’s formulas (about the interplay between roots and coefficients of a polynomial) with the explicit
form of the roots of Un (x) or Tn (x), we have that many trigonometric sums or products can be easily evaluated
through Chebyshev polynomials.

Lemma 61.
n−1 n−1 n−1
n2 − 1
 
X πk n X 1 Y πk 2n
sin2 = , = , sin = .
k=1
n 2
k=1
sin2 πk
n
3
k=1
n 2n

The last identity is related to the combinatorial broken stick problem and it provides an unexpected way for tackling
(through Riemann sums!) the following integral:

Lemma 62. Z π
log sin(x) dx = −π log 2.
0

The first proof we provide relies on a “hidden symmetry”:


Z π Z π/2 Z π/2 Z π/2
log sin(x) dx = 2 log sin(x) dx = log sin2 (x) dx = log cos2 (x) dx
0 0 0 0
Z π/2 Z π/2
sin(2x)
= log [sin(x) cos(x)] dx = log dx
0 0 2
1 π
Z
π
(2x = z) = − log(2) + log sin(z) dz.
2 2 0

Then we exploit the previous closed form for a trigonometric product:


Z π n−1   n−1  
πX πk π Y πk
log sin(x) dx = lim log sin = lim log sin
0 n→+∞ n n n→+∞ n n
k=1 k=1
 
π 2n
= lim log = −π log 2.
n→+∞ n 2n

Page 34 / 222
3 CHEBYSHEV AND LEGENDRE POLYNOMIALS

Exercise 63. Prove the following identities:


2N N
X 1 X 1 2N 2 + 1
  = 4N (N + 1), πk
= .
cos2 πk 1 − cos N
6
k=1 2n+1 k=1

Another famous application of Chebyshev polynomials is related to the determination of the spectrum of tridiagonal
Toeplitz matrices. Due to the Laplace expansion and the recurrence relation for Chebyshev polynomials
 
2x 1 0 ... 0
1 2x 1 ... 0
 
 .. .. 
det  0

1 . . 0  = Un (x)

 .. .. .. ..
 
. .

. . 1
0 0 0 1 2x
so the spectrum of the n × n matrix with C on the diagonal, 1 on the sup- and sub-diagonal and zero anywhere else
is given by:
πk
λk = C + 2 cos , k = 1, 2, . . . n.
n+1
These matrices are deeply involved in the numerical solution of differential equations depending on the Laplacian
operator and in extensions of the rearrangement inequality, like:

π
|a1 a2 + a2 a3 + . . . + an−1 an | ≤ a21 + . . . + a2n cos2

,
n+1
that combined with the shoelace formula can be used to prove the isoperimetric inequality in the polygonal case.
Another important (but lesser-known) application of Chebyshev polynomials is the proof of the uniform convergence
of the Weierstrass products for the sine and cosine functions, over compact subsets of R:

Theorem 64 (Weierstrass). For any x ∈ R the following identities holds


Y x2 4x2
 Y 
sinc(x) = 1− 2 2 , cos(x) = 1−
π n (2n + 1)2 π 2
n≥1 n≥0

and the convergence is uniform over any compact K ⊂ R.

Exercise 65 (Uniform convergence of the Weierstrass product for the cosine function).
Let I = [a, b] ⊆ R and {fn (x)}n∈N the sequence of real polynomials defined through:
n 
4 x2
Y 
fn (x) = 1− .
j=0
(2j + 1)2 π 2

Prove that on I the sequence of functions {fn (x)}n∈N is uniformly convergent to cos x.

Proof. The factorization of Chebyshev polynomials of the first and second kind leads to the following identities:
n
! n−1
!
sin x Y sin2 2n+1
x Y sin2 2n
x

x = 1− , cos x = 1− ,
(2n + 1) sin 2n+1
k=1
sin2 2n+1

j=0 sin2 (2j+1)π
4n

Page 35 / 222
holding for every x ∈ R and every n ∈ Z+ .
We may assume without loss of generality that 0 < x < m < n holds, with m and n being positive natural numbers.
Since for every θ in the interval 0, π2 we have 2θ

π < sin θ < θ, it follows that:

n
! n n
sin2 x
x2 x2
 
Y Y X 1
1> 1− 2n
> 1− > 1 − x2 >1− ,
(2k+1)π
sin2 4n (2k + 1)2 (2k + 1) 2 4m
k=m+1 k=m+1 k=m+1

and by defining Hm (x) as !


m
Y sin2 x
2n
Hm (x) = 1− ,
j=0 sin2 (2j+1)π
4n

cos x belongs to the interval:


x2
  
1− Hm (x), Hm (x) .
4m
By sending n towards +∞ we get that cos x belongs to the interval:
 
 2
Y m  2
 Y m  2

 1− x 1−
4x
, 1−
4x ,
4m j=0 (2j + 1)2 π 2 j=0
(2j + 1)2 π 2

so, by sending m towards +∞, the pointwise convergence of the Weierstrass product for the cosine function is proved.
Additionally, by the last line it follows that:

4x2 /m 4x2
|fm (x) − cos x| ≤ |cos x| 2
≤ ,
1 − 4x /m m − 4x2

and such inequality proves the uniform convergence. The proof of the uniform convergence (over compact subsets of
the real line) of the Weierstrass product for the sine function is analogous.

Chebyshev polynomials can also be employed to prove the following statement (a first density result in Functional
Analysis) through an approach due to Lebesgue.

Theorem 66 (Weierstrass approximation Theorem). If f (x) is a continuous function on the interval [a, b],
for any ε > 0 there exists a polynomial pε (x) such that:

∀x ∈ [a, b], |f (x) − pε (x)| ≤ ε.

We may clearly assume [a, b] = [−1, 1] without loss of generality. Moreover, any continuous function over a compact
interval of the real line is uniformly continuous, so f can be uniformly approximated by a piecewise-linear function of
the form
n
X k
gn (x) = ck x −

n
k=−n

and it is enough to prove the statement for the function f (x) = |x| on the interval [−1, 1]. For such a purpose, we
may consider the projection of f (x) on the subspace of L2 (−1, 1) (equipped with the “Chebyshev” inner product
R1
hu(x), v(x)i = −1 u(x)

v(x)
1−x2
dx) spanned by T0 (x), T1 (x), . . . , T2N (x). Since

1 1 π
|x| T2k+1 (x) |x| T2k (x)
Z Z Z
√ dx = 0, √ dx = |cos θ| cos(2kθ) dθ
−1 1 − x2 −1 1 − x2 0

such projection is given by:


N
2 4 X (−1)k+1
pN (x) = + T2k (x)
π π 4k 2 − 1
k=1

Page 36 / 222
3 CHEBYSHEV AND LEGENDRE POLYNOMIALS

and it can be proved that the maximum difference, in absolute value, between pN (x) and |x| occurs at x = 0 and
equals:
N  
2 2X 1 1 4 X 1 2
pN (0) = − − = = ,
π π 2k − 1 2k + 1 π 4k 2 − 1 π(2N + 1)
k=1 k>N

so the sequence of polynomials {pN (x)}N ≥0 provides a uniform approximation of |x| as wanted.
As an alternative we may consider:
N 
2n (1 − x2 )n
X 
qN (x) = 1 −
n=1
n 4n (2n − 1)

from the truncation of the Taylor series at the origin of 1 − z, evaluated at z = 1 − x2 . In such a case it is trivial
that ||x| − qN (x)| achieves its maximum value at the origin, but the approximation we get this way, according to the
degree of the approximating polynomial, is worse than the approximation we got through Chebyshev polynomials,
1
since |qN (0)| ≈ √πN .

R1
The projection technique on L2 (−1, 1) (equipped with the non-canonical inner product hu(x), v(x)i = −1 u(x)

v(x)
1−x2
dx)
is also known as Fourier-Chebyshev series expansion. About applications, it is important to mention that many
functions have a pretty simple Fourier-Chebyshev series expansion:

Lemma 67. For any x ∈ (−1, 1),


X Tn (x)
− log(1 − x) = log(2) + 2
n
n≥1

p 2 4 X T2n (x)
1 − x2 = −
π π 4n2 − 1
n≥1

4 X T2n−1 (x)
arcsin(x) = .
π (2n − 1)2
n≥1

We many notice that the last identity provides an interesting way for the explicit evaluation of ζ(2) and ζ(4).
Since T2n−1 (1) = 1,
3 X 1 π π2
ζ(2) = = arcsin(1) = .
4 (2n − 1)2 4 8
n≥1

Additionally, due to orthogonality relations (i.e. by Parseval’s theorem),


Z 1 Z π/2
8X 1 arcsin2 (x) π3
4
= √ dx = θ2 dθ =
π (2n − 1) −1 1 − x2 −π/2 12
n≥1

so:
16 X 1 16 π 4 π4
ζ(4) = 2
= · = .
15 (2n − 1) 15 96 90
n≥1

We will now study some applications of Chebyshev polynomials in Arithmetics.

Lemma 68. If q ∈ Q and cos(πq) ∈ Q, then


 
1 1
cos(πq) ∈ −1, − , 0, , 1 .
2 2

Page 37 / 222
Proof. We may assume without loss of generality q = ab with a, b ∈ Z+ and gcd(a, b) = 1.
Let us study the case in which b is odd first. With such assumption:

Tb (cos(πq)) = cos(bπq) = cos(aπ) = (−1)a

so cos(πq) is a root of Tb (x) − (−1)a . The value at the origin of such polynomial is ±1 and the leading term is 2b−1 :
due to the rational root Theorem,
1
cos(πq) ∈ Q =⇒ cos(πq) = ± k .
2
However, due to the cosine duplication formula, if α = cos(πq) is a rational number, 2α2 − 1 = cos(2πq) is a rational
number too. Such observation leads to a proof of the given claimwhen b is odd. If ν2 (b) ≥ 1, it is enough to consider
q
1+cos θ
that by the duplication/bisection formulas cos(θ) = ± 2 we have:
 
1 1
cos(πq) ∈ Q =⇒ cos(2ν2 (b) πq) ∈ Q =⇒ cos(2ν2 (b) πq) ∈ −1, − , 0, , 1
2 2
so cos(πq) ∈ {−1, 0, 1}.

ϕ(n)
Lemma 69. If n ≥ 3, the number α = cos 2π

n is an algebraic number over Q with degree 2 .
Additionally, the Galois group of its minimal polynomial over Q is cyclic.

Proof. It is enough to check that the algebraic conjugates of α are given by


 
2πk n
αk = cos , 1 ≤ k < , gcd(k, n) = 1.
n 2

We may notice that the previous result is actually just a corollary of this Lemma.

Exercise 70. Prove that


2π 3π 5π
α = cos − cos − cos
19 19 19
is an algebraic number over Q with degree 3, i.e. it is a root of a cubic polynomial with integer coefficients.

Legendre polynomials share many properties with Chebyshev polynomials. Our opinion is that the most efficient
way for introducing Legendre polynomials is to do that through Rodrigues formula:

Definition 71.
1 dn 2
Pn (x) = (x − 1)n
2n n! dxn

Such definition provides a simple way for proving the following statements about Legendre polynomials:

• (orthogonality)
Z 1
2 δ(m, n)
Pn (x) Pm (x) dx =
−1 2n + 1

• (Legendre differential equation)


 
d 2 d
(1 − x ) Pn (x) + n(n + 1)Pn (x) = 0
dx dx

Page 38 / 222
3 CHEBYSHEV AND LEGENDRE POLYNOMIALS

• (a simple generating function)


1 X
√ = Pn (x)tn
1 − 2xt + t2 n≥0

• (an interesting convolution formula)


n
X
Un (x) = Pl (x)Pn−l (x)
l=0

• (Bonnet’s recursion formula)


d
(n + 1)Pn+1 (x) = (2n + 1)xPn (x) − Pn−1 (x), (2n + 1)Pn (x) = (Pn+1 (x) − Pn−1 (x))
dx
• (an explicit representation)
n  2  n−k  k
X n 1+x 1−x
Pn (x) = (−1)k .
k 2 2
k=0

It is very practical to introduce shifted Legendre polynomials too, defined by P̃n (x) = Pn (2x − 1).
Due to such affine transform, shifted Legendre polynomials fulfill the following properties:
• (Rodrigues formula)
1 dn 2
P̃n (x) = (x − x)n
n! dxn
• (orthogonality)
Z 1
δ(m, n)
P̃n (x) P̃m (x) dx =
0 2n + 1
• (a simple generating function)
1 X
p = P̃n (x)tn
(t + 1)2 − 4tx n≥0

• (Bonnet’s recursion formula)


d  
(n + 1)P̃n+1 (x) = (2n + 1)(2x − 1)P̃n (x) − P̃n−1 (x), (4n + 2)P̃n (x) = P̃n+1 (x) − P̃n−1 (x)
dx
• (an explicit representation)
n   
X n n+k
Pn (x) = (−1)n (−x)k .
k k
k=0

The explicit representation and orthogonality are really important. Shifted Legendre polynomials give an orthogonal
R1
base (with respect to the canonical inner product hf, gi = 0 f (x) g(x) dx) of the space of square-integrable functions
over (0, 1), and every C 1 function over (0, 1) has a representation of the form:
X Z 1
f (x) = cn P̃n (x), cn = (2n + 1) f (x) P̃n (x) dx.
n≥0 0

We may define a Fourier-Legendre series expansion just like we did for the Fourier-Chebyshev series expansion:
we just have to change the integration range into (0, 1) and the inner product into the canonical one, since the existence
of a complete orthogonal base of polynomials is unchanged.
Let us study an application of the Fourier-Legendre series expansion.

Exercise 72. Prove that for any f ∈ C 1 [−1, 1] we have


Z 1 Z 1 2 Z 1 2
2 f (x)2 dx ≥ 3 x f (x) dx + f (x) dx
−1 −1 −1

and find all the functions for which equality occurs.

Page 39 / 222
Proof. We may assume to have X
f (x) = cn Pn (x)
n≥0

and by the orthogonality relations we have:


Z 1 X c2 Z 1 Z 1
n 2c1
f (x)2 dx = 2 , f (x) dx = 2c0 , P1 (x) f (x) dx = ,
−1 2n + 1 −1 −1 3
n≥0

so the inequality to prove is equivalent to:


X c2n 4c2
4 ≥ 1 + 4c20 ,
2n + 1 3
n≥0

that is trivial and holds as an equality iff c2 = c3 = c4 = . . . = 0, i.e. iff f (x) is a linear polynomial of the form
ax + b.

Like in the Chebyshev case, the Fourier-Legendre series expansion of many functions can be simply derived by ma-
nipulating the generating function for our sequence of polynomials. For any x ∈ (0, 1) we have, for instance:

1 √ X X (−1)n (2n + 1) X (2n + 1)


√ = 2 Pn (x), − log(x) = 1 + P̃n (x), − log(1 − x) = 1 + P̃n (x)
1−x n(n + 1) n(n + 1)
n≥0 n≥1 n≥1

and the lesser known: Z π/2 X Pn (2k 2 − 1)



K(k) = p =2
0 1 − k 2 sin2 θ 2n + 1
n≥0

about the complete elliptic integral of the first kind.


Legendre and Chebyshev polynomials share the uniform boundedness property

∀x ∈ [−1, 1], |Pn (x)| ≤ 1

but that is not entirely trivial for Legendre polynomials.


A possible proof relies on the application of Cauchy-Schwarz inequality to the integral representation
1 π
Z p n
Pn (x) = x + i 1 − x2 cos θ dθ
π 0
that is a consequence of the generating function. An improved inequality is due to Tricomi:
2 1
∀x ∈ (−1, 1), |Pn (x)| ≤ p · √
4
.
(2n + 1)π 1 − x2

Readers can find a sketch of its (highly non-trivial) proof on the Whittaker&Watson book or on MSE (thanks to
M.Spivey). From these remarks (and/or from Bonnet’s recursion formula) it follows that the Legendre polynomial
Pn (x), just like the Chebyshev polynomial Tn (x), has only real roots.

Theorem 73 (Turán). For any x in the interval (−1, 1) the following inequality holds:

Pn (x)2 > Pn−1 (x) Pn+1 (x).

We will later see a proof of this brilliant result (a key ingredient for the Askey-Gasper inequality, that led De
Branges in 1985 to the proof of Bieberbach conjecture) based on the following remarks:

• due to the integral representation, {Pn (x)}n≥0 is a sequence of moments;

Page 40 / 222
3 CHEBYSHEV AND LEGENDRE POLYNOMIALS

• due to the Cauchy-Schwarz inequality, every sequence of moments is log-convex.

In this paragraph we will outline an alternative and really elegant approach, due to Szegö.
We just need few preliminary lemmas.

Lemma 74 (Newton’s inequality). If a1 , . . . , an are distinct real numbers and σk


−1
is the k-th elementary symmetric function of a1 , . . . , an , by setting Sk = σk nk it follows that Sk2 > Sk−1 Sk+1 .

Lemma 75 (Pólya, Schur). If X an


f (z) = zn
n!
n≥0

is an entire function and the zeroes of f are real and simple, as soon as these zeroes ζ1 , ζ2 , ζ3 , . . . fulfill
X 1
< +∞
ζk2
k≥1

it happens that a2n+1 > an an+2 .

Lemma 76 (Gauss, Lucas). The Bessel function of the first kind


X (−1)n z 2n
J0 (z) =
4n n!2
n≥0

fulfills the hypothesis of the previous Lemma.

Szegö’s remark is that the following identity is easy to derive from the generating function for Legendre polyno-
mials: X Pn (x) p
z n = ezx J0 (z 1 − x2 )
n!
n≥0

hence Turán’s inequality is a straightforward consequence of Pólya-Schur’s Lemma. Szegö’s idea is quite deep
since it implies the existence of many “Turán-type” inequalities, not only for Legendre polynomials, but for many
solutions of second-order differential equations with polynomial coefficients: Chebyshev, Hermite, Laguerre, Jacobi
polynomials, Bessel functions . . .

Exercise 77. Prove the identity:


n  
n n + j (−1)n+j (−1)n
X 
= .
j=0
j j (j + 1)2 n(n + 1)

Exercise 78 (A Fejér-Jackson-like inequality). Prove that for any natural number n,


n
X
∀x ∈ (−1, 1), Pk (x) > 0.
k=0

Exercise 79. Prove the following combinatorial identity:


n  2 n   
X n 2l 2n−2l
X 2l 2n − 2l 2l 2n−2l
(x + y) (x − y) = x y .
l l n−l
l=0 l=0

Page 41 / 222
Exercise 80 (Ramanujan-like formulas for π1 and π12 ). Use Rodrigues’ formula or the generating function for Legendre
polynomials to show that
1 L2 (0,1) X 4n + 1 2n2
= π P2n (2x − 1).
16n
p
x(1 − x) n
n≥0

Use Bonnet’s formula to deduce how the Fourier-Legendre series of a function g(x) changes if g(x) is replaced by
g(1 − x) or x · g(x). Use such transformations for showing that
 2
p L2 (0,1) πX 4n + 1 2n
x(1 − x) = P2n (2x − 1).
8 (n + 1)(1 − 2n)16n n
n≥0

1
Compute P2n (0) in explicit terms, then use the evaluation of the previous line at x = 2 and Parseval’s identity for
showing that
4 X (4n + 1)(−1)n 2n3
= ,
π (n + 1)(1 − 2n)64n n
n≥0
 4
32 X (4n + 1) 2n
2
= 2 2 n
.
3π (n + 1) (2n − 1) 256 n
n≥0

A note on Delannoy numbers and their asymptotic behaviour. Let us assume to travel in Z × Z, having
the origin as a starting point and the allowed steps

(n, m) 7→ (n + 1, m) or (n, m) 7→ (n, m + 1) or (n, m) 7→ (n + 1, m + 1).

Let us denote with Dn the number of paths from the origin to (n, n).
By this way we define the sequence of Delannoy numbers

{Dn }n≥0 = {1, 3, 13, 63, 321, 1683, 8989, . . .}

which are straightforward to describe through a bivariate generating function:

1 1 X 1  k 2
n n n n
Dn = [x y ] = [x y ] = .
1 − x − y − xy 2 − (1 + x)(1 + y) 2k+1 n
k≥0

Through Cauchy’s integral theorem or the orthogonality relations in L2 (0, 2π) the RHS of the previous line can
be written as Z 2π
1 dθ
Dn = √
2π 0 (3 − 2 2 cos θ)n+1
giving that {Dn }n≥0 is a sequence of moments, hence a log-convex sequence.
This integral representation leads to the ordinary generating function
X 1
D n xn = √
n≥0
1 − 6x + x2

and by the Hayman/Laplace method we have the asymptotic approximation


1 √
Dn ∼ √ (1 + 2)2n
3n
as n → +∞. Given the generating function for Legendre polynomials, we have Dn = Pn (3).

Page 42 / 222
3 CHEBYSHEV AND LEGENDRE POLYNOMIALS

Exercise 81. The n-th Fibonacci number can be computed in at most (2 + ε) log2 (n) integer multiplications by
exploiting the relations between Fn and Ln and the duplication formulas

L2n = L2n + 2(−1)n+1 , L2n+1 = Ln Ln+1 + (−1)n+1 .

Is it possible to compute the n-th Delannoy number in O(log n) integer multiplications?

Exercise 82. Prove that if we restrict the paths defining Dn to lie in the region m ≤ n, we get the sequence of
Schroeder numbers {Sn }n≥0 , with ordinary generating function

X
n 1 − x − 1 − 6x + x2
Sn x =
2x
n≥0

and related to Legendre polynomials via


1
Sn = (−Pn−1 (3) + 6Pn (3) − Pn+1 (3)) .
2
Through the last identity, find the asymptotic behaviour of Sn as n → +∞.

Exercise 83. Prove that for any prime p we have

Dp−1 ≡ 1 (mod p).

Page 43 / 222
4 The glory of Fourier, Laplace, Feynman and Frullani
In the computation of many limits it is often very practical to exploit Taylor series and Landau notation, i.e., loosely
speaking, the fact that a wide class of functions is well-approximated by polynomials. A key observation in Analysis
(due to Joseph Fourier) is that the same holds for trigonometric polynomials: every real function that is 2π-periodic
and regular enough can be written as a combination of terms of the form sin(nx) or cos(mx) (in Signal Processing these
terms are often called harmonics). For every couple (m, n) of positive natural numbers the following orthogonality
R 2π
relations, with respect to the canonical inner product hf, gi = 0 f (x) g(x) dx, hold:
Z 2π Z 2π Z 2π
sin(nx) sin(mx) dx = cos(nx) cos(mx) dx = π δ(m, n), sin(nx) cos(mx) dx = 0.
0 0 0

In particular, the coefficients of the involved harmonics can be easily computed through integrals: the determination
of such coefficients is a problem equivalent to finding the coordinates of a vector in a infinite-dimensional vector space,
equipped with an inner product and an orthogonal base. We perform such computation in three examplary cases.

Lemma 84 (Fourier series of the sawtooth wave).


Let g(x) be the 2π-periodic function that equals π−x
2 on the interval (0, 2π).

X sin(nx)
∀x ∈ R \ 2πZ, f (x) = .
n
n≥1

Lemma 85 (Fourier series of the rectangle wave).


Let f (x) be the 2π-periodic function that equals 1 on the interval (0, π) and −1 on the interval (π, 2π).

4 X sin((2n + 1)x)
∀x ∈ R \ πZ, f (x) = .
π 2n + 1
n≥0

Lemma 86 (Fourier series of the triangle wave).


Let h(x) be the 2π-periodic function that equals π − |x| on the interval [−π, π].

π 4 X cos((2n + 1)x)
∀x ∈ R, f (x) = + .
2 π (2n + 1)2
n≥0

We may notice that 85 follows from 84 by mapping g(x) into g(x) − g(x + π), while 86 follows from 85 by integration.
We remark that the computation of Fourier coefficients for a 2π-periodic and piecewise-polynomial function is
always pretty simple, but there might be converge issues nevertheless:

• The Fourier series of f (x) might fail to be pointwise convergent to f (x) for some x ∈ [0, 2π]: for instance, such
lack-of-convergence phenomenon occurs for sure if limx→0+ f (x) and limx→0− f (x) exist but do not agree;

• Even if pointwise convergence holds, the Fourier series of f might fail to be uniformly convergent to f (x) on
[0, 2π]: that happens for sure when Gibbs phenomenon occurs:
Z
π − x X N 1
sin(nx) sin(πx)
lim sup − = dx 6= 0.

2 n πx

N →+∞ x∈(π/(N +1),π) 0
n=1

Page 44 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

Moreover, given a trigonometric series, it might be extremely difficult to find a 2π-periodic function with the given
series as a Fourier series. These subtleties will be investigated in full detail during Calculus courses. Now we are just
interested in proving some consequences of the above results, by starting from this observation: if f (x) is a 2π-periodic
function with mean zero, it is a continuous function on (0, 2π) and its Fourier series is pointwise convergent to f (x) on
R \ πZ, by performing a termwise integration on the Fourier series of f (x) we get the Fourier series of an antiderivative
for f (x), and the convergence becomes uniform. In particular, for any x ∈ (0, 2π) we have:
Z x
πx x2 π−y X 1 − cos(nx) X cos(nx)
− = dy = 2
= c0 −
2 4 0 2 n n2
n≥1 n≥1
2
where c0 has to be the mean value of the function πx x
2 − 4 on the interval (0, 2π), since every term of the form cos(nx)
has mean zero. It follows that:
Z 2π Z 1
π2
 
1 1 1
2πx − x2 dx = π 2 x − x2 dx = π 2
 
c0 = ζ(2) = − =
8π 0 0 2 3 6
and such identity further leads to:
X cos(2πnx) π2
∀x ∈ (0, 1), 2
= (1 − 6x + 6x2 ).
n 6
n≥1

Due to the orthogonality relations


Z 1
1
cos(2πmx) cos(2πnx) dx = δ(m, n)
0 2
we also have that:
Z 1 2 2
π4 1 π4
X 1 Z
π 2 2 2
ζ(4) = = 2 (1 − 6x + 6x ) dx = (1 − 6x + 6x ) dx = .
n4 0 6 18 0 90
n≥1

The approach just outlined has an interesting generalization:

Lemma 87. By defining the sequence of Bernoulli polynomials through B0 (x) = 1 and
Z x
∀n ≥ 0, Bn+1 (x) = κn+1 + (n + 1) Bn (y) dy,
0

where the κn+1 constant is chosen in such a way that Bn+1 (x) has mean zero over (0, 1),
R1
we have that for any n ≥ 1 the value of ζ(2n) is a rational multiple of π 2n 0 Bn (x)2 dx, hence:

ζ(2n)
∈ Q.
π 2n

n
If we apply the same technique to the Fourier series of the triangle wave n≥0 (−1) (2n+1)
sin((2n+1)x)
P
2 ,
we may prove in a similar way that:
X (−1)n
∀m ∈ N, ∈ π 2m+1 Q.
(2n + 1)2m+1
n≥0

In particular, from the identity


X (−1)n π3
=
(2n + 1)3 32
n≥0
3
it follows that π ≈ 31 is a pretty accurate approximation. Let us see how to prove the last identity through a very
powerful tool, the Laplace transform. Let us assume that f (x) is a continuous and “vaguely integrable” function
on R+ , meaning that the following limit
Z M
lim f (x)e−sx dx
M →+∞ 0

Page 45 / 222
is finite for any s ∈ R+ . Its Laplace transform, denoted by Lf , is defined through:
Z +∞
∀s ∈ R+ , (Lf )(s) = f (x)e−sx dx.
0

In the given hypothesis, the map sending a continuous and vaguely integrable function into its Laplace transform
is injective, so if Lg = f holds we may say that g is the inverse Laplace transform of f , by using the notation
g = L−1 f .
If f (x) = xk with k ∈ N, we may notice that
k!
Lf (s) = k+1 ,
s
hence by exploiting the linearity of the Laplace transform and the integration by parts formula we may state that:

s2 −s/2
 
1
L−1 3
= e .
(2x + 1) 16
P (−1)n
That allows us to convert the series n≥0 (2n+1)3 into an (indefinite) integral:
+∞ Z +∞
(−1)n X s2 s2
Z
X 1
= e−s/2 (−1)n e−ns ds = ds
(2n + 1)3 0 16 32 0 cosh(s/2)
n≥0 n≥0

and the equality between the last integral and π 3 is a straightforward consequence of the residue Theorem. In a similar
way,
1 +∞
Z
3 X 1 s ds
ζ(2) = = .
4 (2n + 1)2 8 0 sinh(s/2)
n≥0

The trick of converting a series into an integral through the (inverse) Laplace transform is widespread in Analytic
Number Theory: for instance, it is the key ingredient of Ankeny and Wolf’s proof of the fact that, by assuming the
generalized Riemann hypothesis (GRH), for any prime p large enough the minimum quadratic non-residue (mod p) is
≤ 2 log2 (p). It is interesting to underline that the sharpest, GRH-independent upper bound actually known, arising
1

from the combination of Burgess inequality with Vinogradov’s amplification trick, is extremely weaker: ηp  p 4 e .
Another crucial property of the Laplace transform is the following one:

Lemma 88. Under suitable (and not particularly restrictive) hypothesis on the regularity of f and g and their speed
of decay, we have: Z +∞ Z +∞
f (x) · g(x) dx = (Lf )(s) · (L−1 g)(s) ds.
0 0

There are many famous applications of this Lemma, that can be seen as a “regularized” version of the integration
by parts formula. For instance, Dirichlet and Fresnel’s integrals are very simple to compute by using the Laplace
transform.

Lemma 89 (Dirichlet).
Z M
sin x π
lim dx = .
M →+∞ 0 x 2

Proof. Since L−1 1 1



x = 1 and, due to the integration by parts formula, L(sin x) = s2 +1 , we have:
Z M Z +∞
sin x ds π
lim dx = = lim arctan(N ) = .
M →+∞ 0 x 0 s2 + 1 N →+∞ 2

Page 46 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

Lemma 90 (Fresnel).
M M
Z Z r
2 2 π
lim sin(x ) dx = lim cos(x ) dx = .
M →+∞ 0 M →+∞ 0 8

1
 √
Proof. Since the equality Γ 2 =
π holds (that can be proved by considering the integral of a gaussian function
 or,
−1 √1
equivalently, as a consequence of Legendre’s duplication formula for the Γ function), we have that L x
= √1πs
(i.e. √1 is an eigenfunction for the Laplace transform). In particular:
x
Z +∞ Z +∞ Z +∞ Z +∞
2 sin x 1 ds 1 dt
sin(x ) dx = √ dx = √ √ 2 =√ 4
,
0 0 2 x 2 π 0 s(s + 1) π 0 t +1
Z +∞ Z +∞ Z +∞ √ Z +∞ 2
cos x 1 s ds 1 t dt
cos(x2 ) dx = √ dx = √ = √ .
0 0 2 x 2 π 0 (s2 + 1) π 0 t4 + 1
We remark that the integrals of rapidly oscillating functions have been converted into integrals of positive and very
well-behaved functions. Additionally, from Sophie Germain’s identity

1 + 4u4 = (1 − 2u + 2u2 )(1 + 2u + 2u2 )


2
we can derive the partial fraction decomposition of t41+1 , t4t+1 and the computation of Fresnel integrals
boils down to the computation of some values of the arctangent function, like in the previous case.

We also have that the Laplace transform (or the Fourier transform, that we have not introduced yet) can be used
to define derivatives of fractional order, or fractional derivatives. We have indeed that due to the integration by
parts formula, the Laplace transform of f (n) (x) depends in a very simple way on sn (Lf ) (s), hence we may define a
half-differentiation operator (also known as semiderivative) in the following way:

D1/2 f (x) = L−1 s (Lf ) (s) (x)


√ 

d
As soon as all the involved transforms and inverse transforms are well-defined, we have D1/2 ◦ D1/2 = dx .
For instance, by setting Z +∞
2
C= e−x dx
−∞
√ C 2√
 
we have L( x) = and D 1/2
x= Additionally L √1 C
C x. = and

2s3/2 x s

d 1/2 1
D3/2 x = D x= √ .
dx C x
However this is not the only way for defining fractional derivatives.
For instance, for every C ∞ function on the whole real line we have that:
 
1 X n
f (n) (x) = lim+ n (−1)k f (x − kh)
h→0 h k
k≥0

n

where the binomial coefficient k is well-defined also if n 6∈ N.
So we might introduce the semiderivative of f (in the Grünwald-Letnikov sense) also as
 
  X 2k  f (x − kh)
1 X 1/2 1
lim √ (−1)k f (x − kh) = lim √ f (x) − 
h→0+ h k≥0 k h→0+ h k 4k (2k − 1)
k≥1

but in order that the involved series is (at least conditionally) convergent, the f function has to be smooth and with a
sufficiently rapid decay to zero. For functions in the Schwartz space S(R), the two given definitions of semiderivative
are equivalent, but they are not in full generality.

Page 47 / 222
Exercise 91. Investigate about the interplay between the semiderivative of the sine function
(defined through the Laplace transform) and Fresnel integrals.

Exercise 92. Prove that:


Z +∞ Z +∞  4
sin x π
lim 2α sin(xα ) dx = π, dx = .
α→+∞ 0 0 x 3

A series related to the Trigamma function.

Lemma 93. For any a > 0, the integration by parts formula leads to:

1 +∞
Z Z +∞
1 sin(nx) −ax
cos(nx) e−ax dx = 2 = e dx.
a 0 a + n2 0 n

Lemma 94. The series


+∞
X sin(nx)
n=1
n

is pointwise convergent on R \ 2πZ to the function:


 
1 nxo
f (x) = π −
2 2π

where {x} stands for the fractional part of x, i.e. x − bxc.

Due to the dominated convergence Theorem we have:


+∞ Z +∞  
X 1 1 n x o −ax
2 + n2
=π − e dx,
n=1
a 0 2 2π

and by partitioning [0, +∞) as [0, 2π) ∪ [2π, 4π) ∪ . . . it follows that:

+∞ 2π
e2aπ π − x −ax πa coth(πa) − 1
Z
X 1
2 2
= 2aπ e dx = .
n=1
a +n e −1 0 2 2a2

π2
By computing the limit of both sides as a → 0+ , we find ζ(2) = 6 again.

Exercise 95. Prove the following identity through the Laplace transfom:
Z +∞
dt 1
e−t sin2 (t) = log 5.
0 t 4

Page 48 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

Exercise 96. Compute the explicit values of these indefinite integrals:


Z +∞ Z +∞
sin(s) sin(x) − x cos(x)
s−1
ds, dx.
0 e 0 x2

Exercise 97. By considering the Laplace transform of the indicator function of (0, a),
prove that for any a > 0 we have: Z +∞
 sin t
arctan(a) = 1 − e−at dt.
0 t

Exercise 98. Prove that for any u ∈ (0, 1) we have:


Z +∞
1 dx 1 1
· = + .
0 x + 1 − u π 2 + log2 (x) u log(1 − u)

Exercise 99. By the Laplace transform and the Cauchy-Schwarz inequality, prove that:
Z +∞
sin x
dx ≤ 1.
0 x(x + 1)

Exercise 100 (The Laplace transform as an acceleration trick).


Prove that, due to the Laplace transform:
X (−1)n Z +∞
1 ds
√ =√
2n + 1 π 0 cosh(s2 )
n≥0

where a slowly convergent series has been converted into the integral of a function in S(R+ ).
Notice that it is possible to state:
Z +∞ X (−1)n Z +∞
1 ds 1 ds
√ s4 ≤ √ ≤ √ 4 .
1 + s2

π 0 exp 2 2n + 1 π 0
n≥0

The Laplace transform also encodes an important Theorem known as differentiation under the integral sign or
Feynman’s trick. In his biography Surely you’re joking, Mr. Feynman, the American physicist Richard Feynman
talks about instruments learned by reading Woods Advanced Calculus book, that in Feynman’s opinion did not get
proper credit in University courses: differentation under the integral sign granted some fame to Feynman, since he
proved many logarithmic integrals can be tackled in a very slick way by exploiting such instrument.

Lemma 101. As soon as the hypothesis of the dominated convergence Theorem are met,
Z Z
d ∂f
f (x, y) dx = (x, y) dx.
dy E E ∂y

Page 49 / 222

R
The very deep consequences of the fact that, under suitable assumptions, the operators ∂y and E dx commute might
not be evident at first sight. Practical applications are needed to fully understand the hidden “magic”.

Exercise 102. Prove that Z 1


arctan(x) π √
I= √ dx = log(1 + 2).
0 x 1−x 2 2

Proof. If we define I(k) as


Z π/2
arctan(k sin θ)
I(k) = dθ
0 sin θ
we have that I = I(1) and
Z π/2
1 π
I 0 (k) = = √
0 1 + k 2 sin2 θ 2 1 + k2
holds by the substitution θ = arctan(t). Since limk→0+ I(k) = 0, it follows that:

π
Z 1
dk π π √
I = I(1) = √ = arcsinh(1) = log(1 + 2)
2 0 1+k 2 2 2

as wanted.

Exercise 103. Prove that for any k ∈ R


π/2  
1 + |k|
Z
2 2 2
log(sin θ + k cos θ) dθ = π log .
0 2

R π/2
Proof. By denoting as I(k) the integral in the LHS and exploiting 0
log tan(θ) dθ = 0 we have that:

π/2 +∞
log(k 2 + t2 )
Z Z
2 2
I(k) = log(k + tan θ) dθ = dt
0 0 1 + t2

and by differentiation under the integral sign:


Z +∞
2k dt π
I 0 (k) = = .
0 (k 2 2 2
+ t )(1 + t ) k + Sign(k)

Since I(1) = 0, the claim follows in a straightforward way.

Exercise 104. Prove that for any couple (a, b) of distinct and positive natural numbers we have:
+∞
log2 (b) − log2 (a)
Z
log x
I(a, b) = dx = .
0 (x + a)(x + b) 2(b − a)

Proof. By exploiting partial fraction decomposition and integration by parts,


Z +∞ Z +∞  
log x 1 x x log(x)
dx = − dx
0 (x + a)(x + b) b−a 0 x + a x + b x
Z +∞  
1 b a
= − log2 (x) dx.
2(b − a) 0 (x + b)2 (x + a)2

Page 50 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

If we define J(c) as
+∞
c log2 (x) +∞
log2 (cz)
Z Z
J(c) = dx = dz
0 (x + c)2 0 (z + 1)2
and notice that
+∞
log2 (c) +∞
Z Z
2 log(c) log(z)
dz = log2 (c), dz = 0
0 (z + 1)2 0 (z + 1)2
we immediately have:
log2 (z) +∞
log2 (z)
Z Z 1
2 2
J(c) = log (c) + dz = log (c) + 2 dz
0 (z + 1)2 0 (z + 1)
2

and the claim, since the last integral does not depend on c.
Anyway we may observe that by the integration by parts formula
log2 (z)
Z 1 Z 1 Z 1 X Z 1 (−1)n+1 z n−1 X (−1)n+1
2 log z log(1 + z)
2
dz = − dz = 2 dz = 2 dz = 2 = ζ(2).
0 (1 + z) 0 (1 + z) 0 z 0 n n2
n≥1 n≥1

Exercise 105. Prove that: Z +∞ √  dt


 1
1 − cos t e − 1 = .
0 tet 2

Proof. We may introduce the parametric integral


+∞
1 − cos(ωt) −t
Z
I(ω) = e dt.
0 t
It is trivial that limω→0+ I(ω) = 0. We also have:
Z +∞
ω
I 0 (ω) = ω sin(ωt)e−t dt =
0 1 + ω2
hence for any ω > 0 Z ω
u 1
I(ω) = du = log(1 + ω 2 )
0 1 + u2 2

and by considering ω = e − 1 the claim follows.

Exercise 106. Prove that, if a > b > 0,


Z π
dθ πa
I(a, b) = 2
= 2 .
0 (a + b cos θ) (a − b2 )3/2

Proof. We may notice that:


π/2 π/2 π/2
a2 + b2 cos2 θ
Z Z Z
dθ dθ
I(a, b) = + =2 dθ.
0 (a + b cos θ)2 0 (a − b cos θ)2 0 (a2 − b2 cos2 θ)2
By enforcing the substitution θ = arctan t we get:
Z +∞ 2 Z +∞ 2
a (1 + t2 ) + b2 2 (a + b2 ) + (a2 − b2 )t2
I(a, b) = 2 2 dt = dt
0 (a2 (1 + t2 ) − b2 ) 2
a(a − b )2 3/2
0 (1 + t2 )2
R +∞ dt R +∞ t2 dt π
so I(a, b) only depends on the integrals H0 = 0 (1+t2 )2 and H2 = 0 (1+t2 )2 = 2 − H0 .
On the other hand, for any β > 0
Z +∞ Z +∞ Z +∞
dt ∂ 1 d dt d π π
2 )2
=− 2
dt = − 2
=− √ = 3/2
0 (β + t 0 ∂β β + t dβ 0 β + t dβ 2 β 4β

Page 51 / 222
π
hence H0 = H2 = 4 and we have:
2 πa2 πa
I(a, b) = · = 2
a(a2 2
−b )3/2 2 (a − b2 )3/2
as wanted. The given problem can be solved through an interesting geometric approach, too. If a simple closed curve
around the origin is regular and has a parametrization of the form (ρ(θ) cos θ, ρ(θ) sin θ) for θ ∈ [0, 2π], the enclosed
area is given by:
1 2π
Z
A= ρ(θ)2 dθ.
2 0
1
In our case the curve decribed by ρ(θ) = a+b cos θ is an ellipse and
Z π Z 2π
dθ 1
I(a, b) = = ρ(θ)2 dθ
0 (a + b cos θ)2 2 0

is precisely the area enclosed by such ellipse. Since affine maps preserve the ratios of areas, the area enclosed by an
1 1
ellipse is π times the product between the lengths of semi-axis. The length of the major axis is given by a−b − a+b =
2a
a2 −b2 and the ratio between the square of the minor axis and the major axis (also known as semi-latus rectum) equals
1
a . It follows that:
a 1
I(a, b) = A = π · 2 ·√
a − b2 a2 − b2
just like we found before.

Exercise 107. Prove that for any couple (A, B) of positive real numbers the following identity holds:
Z +∞
dx π
I(A, B) = x2
 x2
= HM(A, B)
0 1+ A2 1+ B2
4

2AB
where HM(A, B) is the harmonic mean of A and B, i.e. A+B .

Exercise 108. Find an accurate numerical approximation of the following integral:


Z +∞
sin(3x) sin(4x) sin(5x)
dx.
0 x sin2 (x) cosh(x)
Proof. It is useful to notice that:
Z +∞
def sin(2nx)  πn 
I(n) = dx = 2 arctan tanh .
0 x cosh(x) 2
1
By expanding cosh x as a geometric series we get

1 2 2e−x
= 2 e−x − e−3x + e−5x − e−7x + . . .

= x =
cosh x e + e−x 1 + e−2x
and it follows that Z +∞
X sin(2nx) −(2k+1)x
I(n) = 2 (−1)k e dx.
0 x
k≥0

By differentiating with respect to the n parameter:


X Z +∞ X (2k + 1)(−1)k
0
I (n) = 4 (−1)k
cos(2nx)e−(2k+1)x dx = 4
0 (2k + 1)2 + 4n2
k≥0 k≥0
 
X
k 1 1
= 2 (−1) + .
(2k + 1) + 2in (2k + 1) − 2in
k≥0

Page 52 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

The last series can be interpreted as a logarithmic derivative, and it leads to the identity
π
I 0 (n) = .
cosh(πn)

Since limn→0+ I(n) = 0, or by exploiting limn→+∞ I(n) = 0 that follows from the Riemann-Lebesgue Lemma, we
get: Z n
π  πn 
I(n) = du = 2 arctan tanh .
0 cosh(πu) 2
On the other hand, by exploiting Chebyshev polynomials of the second kind it is not difficult to check that:

sin(3x) sin(4x) sin(5x)


= 2 sin(2x) + 3 sin(4x) + 3 sin(6x) + 2 sin(8x) + sin(10x),
sin(x)2
Z +∞
sin(3x) sin(4x) sin(5x) 11π
2 dx = 2 I(1) + 3 I(2) + 3 I(3) + 2 I(4) + I(5) ≈
0 x sin (x) cosh(x) 2
π
where the last approximation follows from the fact that I(n), for n  1, converges quite fast to the value 2.

Exercise 109. Prove that for any t > 0 we have:


1
xt − 1
Z
I(t) = dx = log(t + 1).
0 log x

Proof. Since xt − 1 = exp(t log x) − 1 = t log(x) + O t2 log2 x for t → 0+ and − log(x) is a positive and integrable


function over the interval (0, 1), we have that limt→0+ I(t) = 0. In order to prove the claim it is enough to show that:
1
xt − 1
Z
d 1 d
dx = = log(t + 1)
dt 0 log x t+1 dt

∂ xt −1
R1 1
which is a trivial consequence of ∂t log x = xt and 0
xt dx = t+1 .

Exercise 110. Compute the following integral by using Riemann sums:


Z π
log(7 + cos θ) dθ.
0

Proof. We may notice that for any n ∈ N+ :


2n   n−1
Y 
2n
Y πik 2 2 πk
x −1= x − exp = (x − 1) x + 1 − 2x cos .
n n
k=1 k=1

x2 +1

If we choose x in such a way that 2x = −7 holds, for instance through x = 4 3 − 7, we get:

n−1
π √
Z  
π Y
πk
 7
log(7 + cos θ) dθ = lim log 7 + cos n = π log +2 3 .
0 n→+∞ n 2
k=1

Page 53 / 222
Exercise 111. Prove that for any a ∈ (−1, 1)
Z π
log(1 − 2a cos x + a2 ) dx = 0.
0

Exercise 112. Prove that Z 1


arctan x π
= log 2.
0 x+1 8

Exercise 113. Prove that


π/2
π3
Z
π
log(sin x) log(cos x) dx = log2 (2) − .
0 2 48

At this point it should be clear that Feynman’s trick is a really powerful tool.
In the following formal manipulation
Z Z Z 1Z Z 1
˜ ∂ ˜
f (x) dx = f (x, 1) dx = f (x, α) dx = g(α) dα
E E 0 E ∂α 0

we have a complete freedom in choosing where to introduce the dummy parameter α in the definition of f , in such a
way that f˜(x, α)|α=1 = f (x) holds. In particular, Feyman’s trick is really effective in the computation of logarithmic
d α

integrals, since log(x) = dα x α=0+ , or in proving identities like


X log n d X 1
=− = −ζ 0 (2).
n2 dα nα
n≥1 n≥1
α=2

Additionally differentiation under the integral sign, the Laplace transform and Frullani’s Theorem
provide interesting integral representations for logarithms.

Theorem 114 (Frullani). If f ∈ C 1 (R+ ) and limx→+∞ f (x) = 0, for any couple (a, b) of positive real numbers we
have: Z +∞  
f (ax) − f (bx) b
dx = log · lim f (x).
0 x a x→0+

The most typical case of Frullani’s Theorem is related to the function f (x) = e−x :

Lemma 115. If b ≥ a > 0,


+∞
e−ax − e−bx
Z  
b
dx = log .
0 x a

Let us prove this Corollary through Feynman’s trick and the Laplace transform.
It is enough to show that: Z +∞ −x
e − e−ax
∀a > 0, I(a) = dx = log(a).
0 x

Page 54 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

It is pretty clear that lima→1 I(a) = 0. Since


Z +∞
1
I 0 (a) = e−ax dx =
0 a

we have: Z a Z a
0 dξ
I(a) = I (ξ) dξ = = log(a).
1 1 ξ
−1 1 −ax 1

As an alternative, by exploiting L x = 1 and L(e )= a+s ,
Z +∞  
1 1
I(a) = − ds = lim [log(M + 1) − log(M + a) + log(a)] = log(a).
0 s+1 a+s M →+∞

 
1 1
P
We remark the strong analogy between the last identity and Hn = m≥1 m − m+n . Frullani’s Theorem can be
used to provide integral representations for the Euler-Mascheroni constant γ or the constant log π2 :

Lemma 116 (An integral representation for the Euler-Mascheroni constant).


X1   Z +∞  
def 1 1 1
γ = lim (Hn − log n) = − log 1 + = − dx.
n→+∞ n n 0 ex − 1 xex
n≥1

Proof. Due to Frullani’s Theorem:


 Z +∞ 
e−nx − e−(n+1)
 
1 1 −nx
− log 1 + = e − dx
n n 0 x

and the claim follows by summing both sides on n ≥ 1.


We may notice that by Weierstrass product for the Γ function we have γ = −Γ0 (1).
It follows that by Feynman’s trick we also have:
Z +∞ Z +∞
d s−1 −x
log x
γ=− x e dx =− dx.
ds 0 s=1 0 exp(x)

π
Lemma 117 (An integral representation for the log 2 constant).
+∞
ex − 1 dx
Z
π
log = · .
2 0 ex + 1 xex

Proof. By Weierstrass product for the sinc function it is simple to prove that:

2 Y 1

= 1− 2 .
π 4n
n≥1

By switching to logarithms and exploiting Frullani’s Theorem,


Z +∞ X −nx Z +∞ x
π X
n+1 n+1 e − e−(n+1)x e − 1 dx
log = (−1) [log(n + 1) − log(n)] = (−1) dx = ·
x + 1 xex
.
2 0 x 0 e
n≥1 n≥1

Page 55 / 222
Exercise 118. Prove that
X  +∞
sinh x − x 1 − log 2
Z
2n + 1
n log −1 = 2 dx = .
2n − 1 0 2x sinh (x) 2
n≥1

Hint: the first equality follows from the Laplace transform, and the equality between the first term and the last one
is a consequence of Stirling’s inequality, since:

N   N
X 2n + 1 X
SN = n log −1 = −N + n [log(2n + 1) − log(2n − 1)]
n=1
2n − 1 n=1
N
X −1
= −N + N log(2N + 1) − log(2n + 1)
n=1
= −N + N log(2N + 1) − log [(2N − 1)!!]
= −N + N log(2N + 1) − log [(2N )!] + N log 2 + log [N !] .

As an alternative, we may notice that:

X 1
 XX
1 X ζ(2m)
S∞ = 2n arctanh −1 = 2m
=
2n (2m + 1)(2n) 4m (2m + 1)
n≥1 n≥1 m≥1 m≥1

R 1/2
and recognize in the last series the integral 0 [1 − πx cot(πx)] dx that is simple to compute by integration by parts:
R π/2
by this way the problem boils down to the computation of the well-known integral 0 log sin(θ) dθ = − π2 log(2).

We now introduce another very powerful tool for dealing with series, i.e. the discrete equivalent of the integration by
parts formula.

Theorem 119 (Abel, Summation by parts formula).


(ΣΣ) version: if {an }n≥1 and {bn }n≥1 are two sequences of complex numbers, by setting An = a1 + . . . + an we have:
N
X N
X −1
an bn = AN bN − An (bn+1 − bn ).
n=1 n=1

(Σ ) version: if φ(x) is a C 1 function on R+ , for any sequence {an }n≥1 of complex numbers we have:
R

X Z x
an φ(n) = A(x)φ(x) − A(u)φ0 (u) du
1≤n≤x 1

where X
A(x) = an .
1≤n≤x

The existence of a discrete-discrete and a discrete-continuous analogue of the integration by parts formula is very
practical, almost comforting. Abel’s theorem allows to study the convergence of a power series on the boundary of the
region of convergence, or to regularize series whose convergence is uncertain at first sight. The following Lemma, for
instance, is a straightforward consequence of the summation by parts formula:

Page 56 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

Lemma 120 (Dirichlet’s test).


(Discrete version) If {an }n≥1 and {bn }n≥1 are two sequence of real numbers, such that An = a1 + . . . + an is bounded
and bn is decreasing towards zero, the series X
an bn
n≥1

is convergent.
Rx
(Continuous version) If f (x), g(x) ∈ C 0 (R+ ), where g(x) is weakly decreasing towards zero and 0 f (u) du is bounded,
the improper Riemann integral Z +∞
f (x) g(x) dx
0
is convergent.

The convergence of
X sin n X cos n Z +∞
sin x
, , dx
n n 0 x
n≥1 n≥1
immediately follows from Dirichlet’s test, once we show two simple results on the behaviour of particular trigonometric
sums.

Lemma 121. For any couple (k, N ) of positive real numbers we have:

N N 1
X
1 k
X 1
sin(kn) ≤ 2 cot 4 , cos(kn) ≤ + .

2 2 sin k2


n=1 n=1

Proof. Both the mentioned sums are telescopic sums in disguise.


As a matter of fact, by the sine/cosine addition formulas:
N N h i
X X
k 1 1 1 1
cos k2 − cos (2N 2+1)k ,
    
sin 2 sin(kn) = 2 cos n− 2 k − cos n+ 2 k = 2
n=1 n=1
N N h i
X X
k 1 1 1 1
sin k2 − sin (2N 2+1)k ,
    
sin 2 cos(kn) = 2 sin n− 2 k − sin n+ 2 k = 2
n=1 n=1
so, in particular:
N N
X 1 + cos k X 1 + sin k 1 1
1
sin(kn) ≤ 2
= cot k , cos(kn) ≤ 2
≤ + .


2 sin k2 2 4
2 sin k2 2 2 sin k2
n=1 n=1

Exercise 122 (about Van Der Corput’s trick and Weyl’s inequality). Prove that the series
X sin(n2 )
n
n≥1

is convergent. Hint: it is enough to show that for any N large enough the inequality
N
X √
exp(in2 ) ≤ 4 N log2 (N )



n=1

holds, then apply summation by parts.

Page 57 / 222
sin n cos n
P P
Let us see how to compute n≥1 n and n≥1 n in a explicit way, now that we know they are convergent series.
Since we have:
X zn
= − log(1 − z)
n
n≥1

for any complex number z with modulus less than 1, it is reasonable to expect that, by evaluating both sides at z = ei ,
X sin n π−1 X cos n
= −Im log(1 − ei ) = = −Re log(1 − ei ) = − log 2 sin 21

,
n 2 n
n≥1 n≥1

n
holds. That is correct, indeed, but we have to explain why we are allowed to evaluate the Taylor series n≥1 zn at a
P

point on the boundary of the convergence region |z| < 1. We may notice that the function − log(1 − z) is continuous
(and much more, actually: holomorphic) in a neighbourhood of z = ei : that is enough to justify the “wild” evaluation
we performed. Such observation is also known as Abel’s Lemma. We may also proceed through a deformation
1
of an integration path. Let us consider the function g(z) = 1−z : it is a meromorphic function with a simple pole at
z = 1, and at the interior of the unit disk we have:

1 X
= 1 + z + z2 + z3 + . . . = zn.
1−z
n≥0

By termwise integration it follows that:


Z z X zn Z ei
du X sin n
− log(1 − z) = = , = Im g(z) dz.
0 1−u n n 0
n≥1 n≥1

Since g(z) is holomorphic far from z = 1, the integral along the


straight segment joining the origin with ei equals the integral along
the path given by the concatenation of the straight segment joining
z = 0 with z = −1 and the depicted circle arc γ joining −1 with ei .
In particular:
X sin n Z −1 Z
dz dz
= Im + Im
n 0 1−z γ 1−z
n≥1

and the first integral appearing in the RHS is a real number.

As a consequence:
Z ei
sin θ2
Z 1 Z π Z π
X sin n dz ieiθ eiθ/2 dθ π−1
= Im = Im iθ
dθ = Re iθ/2 −iθ/2
= θ
dθ = .
n≥1
n −1 1 − z π 1−e 1 e −e 1 2 sin 2 2

The same argument (finally!) proves the pointwise conver-


gence of the Fourier series of π−x
2 on the interval (0, π). More-
P sin n
R +∞ sin x
over, both the series n≥1 n and the integral 0 x dx
can be computed through the properties of the Fejér kernel.

Page 58 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

The Fejér kernel. In Functional Analysis density and regularization (or “mollification”) tricks are often used.
They are based on the fact that the convolution (f ∗ g)(x), defined through
Z +∞
(f ∗ g)(x) = f (τ ) g(x − τ ) dτ
−∞

inherits the best regularity between the behaviours of f (x) and g(x). In particular, if g(x) is a non-negative
function with unit integral, belonging to C k and concentrated enough around the origin, (f ∗ g)(x) is an excellent
approximation of f (x) that may be way more regular than f (x). Functions g(x) fulfilling the previous constraints
are regular approximations of the Dirac δ distribution, and they are said approximated identities or convolution
kernels. The Fejér kernel is a classical example.

Definition 123.
N −1 k N −1 k
!
def
X  |j|

1 X X isx 1 X X
ijx
FN (x) = 1− e = e = 1+2 cos(sx)
N N N s=1
|j|≤N k=0 s=−k k=0
!2
sin N2x
 
1 1 − cos(N x) 1
= = .
N 1 − cos x N sin x2

This particular trigonometric function is non-negative due to the last identity (for short: it is a square).

Moreover limx→0 FN (x) = N and −π FN (x) dx = 2π for any N ≥ 1. By termwise integration,

Z 1 N −1 k N −1
1 X X sin s X sin s  s
FN (x) dx = 1 + =1+2 1− .
0 N s=1
s s=1
s N
k=1

On the other hand


N N N  
X sin s X sin s  s 1 X 1
− 1− = sin(s) = O ,
s=1
s s=1
s N N s=1 N
so:  Z 1 
X sin s 1 π−1
= lim −1 + FN (x) dx =
s 2 N →+∞ 0 2
s≥1

since limN →+∞ 1
FN (x) dx = 0. It is not difficult to locate the stationary points of the function sin(N x/2)
sin(x/2) and
 2
state that for any large enough N the ratio sin(N x/2)
sin(x/2) is bounded by an absolute constant on the interval (1, π),
Rπ 1

from which it follows that 1 FN (x) dx = O N . With the same approach based on termwise integration and by
exploiting a Riemann sum we have:
k
Z Z k
N sin x
lim FN (x) dx = 2 · Si(k) = 2 dx
N →+∞ 0 0 x

hence by considering the limit as k → +∞ we get:


Z +∞
sin x π
dx = .
0 x 2

The pointwise convergence of the Fourier series of the sawtooth wave can also be studied through the Laplace transform.
If f (x) is a bounded and vaguely integrable function over R+ , the dominated convergence theorem ensures
Z +∞
lim+ f (x) = lim se−sx f (x) dx = lim s · Lf (s).
x→0 s→+∞ 0 s→+∞

Page 59 / 222
P sin(nx)
If this manipulation (convolution with an approximate identity) is applied to n≥1 n , it produces:
X sin(nx) X s
lim+ = lim
x→0 n s→+∞ s + n2
2
n≥1 n≥1

−1+πs coth(πs)
where the last series equals 2s . But even ignoring this identity, Riemann sums grant
Z +∞
X s dx π
lim 2 2
= 2
= .
s→+∞ s +n 0 1+x 2
n≥1

Poisson summation formula.

C
Lemma 124 (Poisson). If f ∈ C 2 (R), |f (x)| ≤ 1+x2 and the first two derivatives of f are integrable on R,

+∞ +∞ Z +∞
def
X X
f (n) = fˆ(n), fˆ(ν) = f (x)e−2πiνx dx.
n=−∞ n=−∞ −∞

This result has a great importance in Harmonic Analysis and Number Theory. It follows from the fact that the
distribution known as Dirac comb is a fixed point of the Fourier transform, and the identity
+∞ +∞
X 1 X
2
 π 
∀a > 0, exp(−πan + 2bin) = √ exp − (n − b)2
n=−∞
a n=−∞ a

leads to the reflection formula for the Riemann ζ function. It is possible to employ the Poisson summation formula
to prove interesting identities related to modular forms, like:
X coth(πn) 7π 3
=
n3 180
n≥1

or identities already proved, like:


X 1 X 1 X
−1 + 2 = = πe−2π|n| = π coth(π),
n2 +1 2
n +1
n≥1 n∈Z n∈Z

X 1 π π
= −1 + tanh .
n2 + (n + 1)2 2 2
n≥1

R 2π
The Poisson kernel. Since log kzk = Re log z and 0
eniθ dθ = 2πδ(n), for any r ∈ R we have:

Z 2π
log 1 − reiθ dθ = 2π log max(1, |r|).

0

Such identity can be re-written in terms of the integral of a real function:

Z 2π
∀r ∈ R, log(1 + r2 − 2r cos θ) dθ = 4π log max(1, |r|).
0

Page 60 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

By differentiating with respect to the r parameter, it follows that:

(
2π 2π
r − cos θ if |r| > 1
Z
∀r ∈ R \ {−1, +1}, dθ = r
0 1 + r2 − 2r cos θ 0 if |r| < 1,
(

1 − r cos θ if |r| < 1
Z

∀r ∈ R \ {−1, +1}, dθ =
0 1 + r2 − 2r cos θ 0 if |r| > 1.

That is not entirely surprising: the last result also follow from the behavior of complex homographies
or Cayley transforms, since

1 + reiθ 1 − r2
X  
|n| inθ
∀r ∈ [0, 1), Pr (θ) = r e = Re = .
1 − reiθ 1 + r2 − 2r cos θ
n∈Z

Exercise 125. Prove that the function


Z 2π
1 (1 + 3r2 ) − (5r + r3 ) cos(θ) + (1 + r2 ) cos(2θ)
f (r) = dθ
2π 0 (1 + r2 − 2r cos θ)2
1
equals 1 on the interval (−1, 1) and r2 outside the previous interval.

The Fourier series of log Γ.


1
 1 1 X sin 2πnz · log n
log Γ(z) = 2 − z (γ + log 2) + (1 − z) ln π − log sin πz + , 0<z<1
2 π n=1 n

is an important identity that can be derived through Weierstrass product for the Γ function and the Laplace
transform. For a long time it was credited to Ernst Kummer, that proved it in 1847. Only recently Iaroslav
Blagouchine has pointed out the same result was known to the Swedish mathematician Carl Johan Malmsten
since 1842.

Exercise 126. {an }n≥0 is a sequence defined by a0 = 0 and


q
1
an + a2n + 4n
an+1 =
2
for any n ≥ 0. Prove that: r
5 2
lim an ≤ , lim an = .
n→+∞ 12 n→+∞ π

Exercise 127 (Kronecker’s Lemma). A sequence {an }n≥1 of real numbers is such that
N
X an
lim = C < +∞.
N →+∞
n=1
n

Page 61 / 222
Prove that {an }n≥1 is necessarily a sequence with mean zero, i.e.
a1 + a2 + . . . + aN
lim = 0.
N →+∞ N

a1 a2 aM
Proof. Let us set AM = 1 + 2 + ... + M . By summation by parts:

n n n−1
1X 1X ak 1X
ak = k· = An − Ak
n n k n
k=1 k=1 k=1

and the convergence of the original series ensures that for any ε > 0, there is some N such that |An − C| ≤ ε holds
for any n ≥ N . By picking an ε and considering the associated N , the previous RHS can be written as
N −1 n−1
1 X n−N 1 X
An − Ak − C− (Ak − C).
n n n
k=1 k=N

If now we consider the limit as n → +∞, the first term goes to C, which cancels with the third term; the second term
goes to zero (as the sum is a fixed value) and the last term is bounded in absolute value by n−N
n ε ≤ ε.

A remark about nuking mosquitoes: since the sequence 1, 1, 1, . . . has not mean zero, the harmonic series is divergent.

Exercise 128. Find a function f ∈ C 0 (R+ ) such that the following equality holds for any t > 0:
Z t
f (t) = e−3t + e−t e−τ f (τ )dτ
0

Proof. Let g(s) = (Lf )(s). The Laplace transform of e−x f (x) is given by g(s + 1) and the Laplace transform of
R x −u
0
e f (u) du is given by 1s g(s + 1), hence the given differential equation can be written in terms of g as

1 g(s + 2)
g(s) = +
s+3 s+1
leading to:    
1 1 1 1 1 1 1
g(s) = + + + + ...
s+3 s+1 s+5 s+3 s+7 s+5 s+9

1 1 1 1
g(s) = + + + + ...
s + 3 (s + 1)(s + 5) (s + 1)(s + 3)(s + 7) (s + 1)(s + 3)(s + 5)(s + 9)
It is clear that g(s) is a meromorphic function with poles at the negative odd integers. Additionally it is simple to
compute the closed form of Rξ = Ress=ξ g(s) for any ξ ∈ Ξ = {−1, −3, −5, . . .}, then to consider the inverse Laplace
transform of g:
X
f (x) = Rξ eξx .
ξ∈Ξ

The closed form for g(s) is related to the incomplete Γ function:


1+s√ 
e −Γ 3+s + Γ 3+s 1
 
1+2 2
2 2 , 2
g(s) =
1+s
" Z 1/2 #
1 1+s √ s+1
−u
= 1−2 2 e u 2 e du
1+s 0
√ 1√ 1√ 1 √ 1 √
2− e 2 e 8 e 48 e 384 e
= + − + − + ...
s+1 s+3 s+5 s+7 s+9

Page 62 / 222
4 THE GLORY OF FOURIER, LAPLACE, FEYNMAN AND FRULLANI

and from the last line it is straightforward to recover a solution f (x):

√ X (−1)n+1 −(2n+1)x

f (x) = (2 − e e)e−x + e
(2n)!!
n≥1
 −x

= e−x 2 − ee sinh(x) .

The last entry of this section is a surprising consequence of the Laplace transform and the residue Theorem:

Theorem 129 (Ramanujan’s Master Theorem). Under suitable regularity assumptions for φ, the identity
X φ(n)
f (x) = (−x)n
n!
n≥0

implies: Z +∞
xs−1 f (x) dx = Γ(s) φ(−s).
0

In particular, given the series


X ζ(k)
log Γ(1 + x) = −γx + (−x)k
k
k≥2

following from the Weierstrass product for the Γ function, we have:


Z +∞
dx π ζ(1 + s)
∀s ∈ (0, 1), (γx + log Γ(1 + x)) s+2
= · .
0 x sin(πs) 1+s

Exercise 130 (The Russian Integral). Prove that for any a ∈ R+ and any b ∈ (0, 2) the following identity holds:

sinh a arccos 2b
Z +∞ 
x−ia 2π
dx = √ · .
0 x2 + bx + 1 4 − b2 sinh(aπ)

Proof. We may immediately notice that

eBs − eBs
 
−ia ia−1 −1 1
L(x )(s) = s Γ(1 − ia), L (s) = √
x2 + bx + 1 b2 − 4

where B is the root of x2 + bx + 1 with a positive imaginary part. By the properties of the Laplace transform, the
original integral is converted into
Γ(1 − ia) +∞ ia−1  Bs
Z 
√ s e − eBs ds
b2 − 4 0
which can be evaluated in terms of the Γ function. Due to the reflection formula, the final outcome simplifies into
−ia
 
B −ia − B π

sinh(πa) 4 − b2

and we may notice that B = exp i arccos 2b allows a further simplification, proving the claim.


Page 63 / 222
P cos(2nx)
Exercise 131. Find a closed form for n≥1 2n and use it to prove the following identity:

π/2
π 2e2 + 1
Z

2 = · 2 .
0 1 + 8 sin (tan θ) 6 2e − 1

Remark. By the inverse Laplace transform, for any p > 0,


X (−1)n+1 Z +∞ p−1 −s  
s e ds 1
η(p) = = · = E
np 0 Γ(p) 1 + e−s 1 + exp(−X)
n≥1
1
where X is a random variable with a Γ(p, 1) distribution. Given the RHS, the inequality 2 < η(p) < 1 is trivial.

The inverse Laplace transform of the central binomial coefficients.


Given the identity
2 π/2 1 +∞ e−nx
  Z Z
1 2n 2
n
= sin θ dθ = √ x dx
4n n π 0 π 0 e −1
we have the equality L−1 41s 2s = π√e1x −1 , hence the asymptotic behaviour of 41n 2n
 
s n depends on the Maclaurin
q
series of exx−1 . On the other hand

x x x  x X ζ(2n)
= coth − 1 = 1 − − (−1)n x2n
ex −1 2 2 2 (2π)2n
n≥1

implies  
x x X ζ(2n)
log =− + (−1)n x2n
ex − 1 2 n(2π)2n
n≥1

and  
r
x x X ζ(2n)
= exp − + (−1)n x2n  .
ex − 1 4 2n(2π)2n
n≥1

From the local approximation


e−x/4 x2
    
1 2s
L−1 ≈ √ 1 −
4s s π x 48
it is straightforward to recover the very accurate asymptotic approximation (which actually is a lower bound)
 
1 2n (8n + 1)(8n + 3)
n
∼ √ .
4 n 2 π(4n + 1)5/2

The evaluation of both sides at n = 2 produces 1292 729 as a rational approximation of π, whose absolute error
is about 1.63 · 10−4 . The evaluation at n = 6 produces the approximation 60928 34375 , whose absolute error is less
than 3 · 10 . An interesting exercise is to check that the approximation produced via L, L−1 outperforms the
−6

approximation produced via creative telescoping:


  2 n  2 n  −1  
1 2n 1Y 1 1 Y 1 1 Y 1
= 1 − = 1 − = 1 −
4n n 4 2k 4n (2k − 1)2 πn (2k + 1)2
k=2 k=2 k≥n

1 XX −1 CT 1 Y 8k − 1 8k + 9 1 8n − 1
= exp 2m
≈ · = · .
πn m(2k + 1) πn 8k + 7 8k + 1 πn 8n + 1
k≥n m≥1 k≥n

On the other hand a refinement of this approach leads to a tighter upper bound:
  2
(8n + 1)2 (8n + 3)2 L 1 2n CT 1 64n2 − 8n + 3
5
≤ n
≤ · .
4π(4n + 1) 4 n πn 64n2 + 8n + 3

Page 64 / 222
5 THE BASEL PROBLEM

5 The Basel problem


The Basel problem has been posed by Pietro Mengoli in 1644 and solved by Leonhard Euler in 1735 with a not
entirely rigorous argument, fixed in 1741. This section is dedicated to many classical and alternative approaches for
showing that:

X 1 π2
ζ(2) = 2
= .
n 6
n≥1

Proof # 1 (Euler, 1735-1741).


Let us consider the function of complex variable usually denoted as sinc(z), i.e. the function that equals 1 at the origin
and sinz z anywhere else. Due to the formula 5 sin(z) = 2i 1
eiz − e−iz , all the zeroes of the function sinc(z) belong


to the real line, are simple and exactly located at the elements of πZ \ {0}. We may recall that by the Fundamental
Theorem of Algebra, any even polynomial p(z) with simple zeroes, such that p(0) = 1, can be written in the form
n 
z2
Y 
p(z) = 1− 2
ζk
k=1

where ±ζ1 , . . . , ±ζn stands for the zeroes of p(z). Given such identity,
n
X 1
= −[z 2 ]p(z)
ζk2
k=1

holds as a consequence of Vieta’s formulas.


By assuming to be allowed to deal with sinc(z) as a “polynomial with an infinite degree”, from the identity
Y z2

sinc(z) = 1− 2 2
π k
k≥1

it follows that:
z2 z4 π2
 
2 2 2 2
ζ(2) = −π [z ] sinc(z) = −π [z ] 1 − + − ... = .
6 120 6
Quoting Sullivan, you may now sit back and smile smugly at his brilliance.
However, besides the last proof being absolutely brilliant and incredibly efficient, it is based on a unproven assumption:
the statement that we are allowed to deal with sinc(z) by regarding it as an infinite-degree polynomial. Which is true,
indeed, since the Weierstrass product for the entire function sinc(z) has no exponential part, also as a consequence of
Mittag-Leffler’s Theorem. However this part of Complex Analysis has been developed only in the middle nineteenth
century, and it was most certainly unknown to Euler. Such “flaw” was probably the reason for Euler to fix his original
proof, based on such inspired guess. In 1741, by starting from the property of uniform convergence of the Weierstrass
product for the sinc(z) function, he proved that in a neighbourhood of the origin we have:

z2
X  
log sinc(πz) = log 1 − 2
n
n≥1

uniformly. Once the uniform convergence of derivatives is proved, it follows that


1 X 2z
− π cot(πz) =
z n2 − z 2
n≥1

holds for any z 6= 0 close enough to the origin. In particular:


π2
 
X 1 1 1 π cos(πz)
ζ(2) = lim = lim − =
z→0 n2 − z 2 2 z→0 z2 z sin(πz) 6
n≥1

5 usually attributed to De Moivre, even if we are pretty sure Newton knew it yet, back in 1676.

Page 65 / 222
d
and rigour is safe. We may state that considering the logarithmic derivative dz log(·) is a big detour on the original
idea, nevertheless Euler’s second proof exploits an instrument that few years later will become essential in Complex
Analysis (deeply related with the topological degree of curves), and is able to show that all the values of the ζ function
at positive even integers can be computed through the coefficients of the Taylor series of z cot z at the origin:

1 − πz cot(πz) X
= ζ(2n) z 2n .
2
n≥1

In other terms, Euler’s “fixed” proof exhibits a generating function for the sequence {ζ(2n)}n≥1 . To produce deeper
results with less than two pages of written math is barely human: this is one of the reasons for which, since the
nineteenth century, concise and elementary proofs, able to prove highly non-trivial statements, have been mentioned
as Eulerian.

We remark that, by following the Eulerian approach of differentiating the logarithm of a Weierstrass product,
then exploiting the substitution z 7→ iz, we have:
X 1 −1 + πz coth(πz)
=
n2 + z 2 2z 2
n≥1

for any z ∈ C. Additionally, by differentiating both sides with respect to z we have that:

π2 z2
 
X 1 1 πz cosh(πz)
= −2 + + .
(n2 + z 2 )2
n≥1
4z 4 sinh(πz) sinh2 (πz)

Before studying other “classical attacks” to Basel problem, we outline a further proof due to Euler, related to manip-
ulations of the squared arcsine function.

Proof # 2 (Euler, 174?). It is trivial that:


Z 1
arcsin(x) 1 π2
√ dx = arcsin2 (1) = .
0 1 − x2 2 8
Additionally the Taylor series at the origin of arcsin(x) and √ 1 are well-known. It follows that:
1−x2
P∞ (2n−1)!! x2n+1
π2 4
Z 1
arcsin x 4
Z 1 x+ n=1 (2n)!! 2n+1
= √ dx = √ dx
6 3 0 1−x 2 3 0 1 − x2
1 ∞ Z 1
4 X (2n − 1)!!
Z
4 x x
= √
dx + x2n √ dx
03 1−x 2 3 n=1 (2n)!!(2n + 1) 0 1 − x2
∞  
4 4 X (2n − 1)!! (2n)!!
= +
3 3 n=1 (2n)!!(2n + 1) (2n + 1)!!

4X 1
=
3 n=0 (2n + 1)2
∞ ∞
!
4 X 1 1X 1
= −
3 n=1
n2 4 n=1 n2

X 1
= 2
= ζ(2).
n=1
n
We remark we already exploited a similar idea: we showed it is possible to prove the identity
X 3
ζ(2) = 2 2n

n≥1
n n

Page 66 / 222
5 THE BASEL PROBLEM

by creative telescoping or by exploiting the tangent half-angle substitution in a logarithmic integral. Once the coeffi-
cients of the Taylor series of arcsin2 (x) have been computed, one may recognize in the RHS the quantity 43 arcsin2 (1).

Proof # 3 (Cauchy, 1821) The following proof comes from its author’s Cours d’Analyse, note VIII.
Let x ∈ 0, π2 and let n be an odd natural number. By De Moivre’s formula and the definition of cotangent we have:


cos(nx) + i sin(nx) (cos x + i sin x)n


=
(sin x)n (sin x)n
 n
cos x + i sin x
=
sin x
= (cot x + i)n .

Due to the binomial Theorem we also have:

(cot x + i)n
       
n n n n−1 n n−1 n n
= cot x + (cot x)i + . . . + (cot x)i + i
0 1 n−1 n
         
n n n n
= cotn x − cotn−2 x ± . . . + i cotn−1 x − cotn−3 x ± . . . .
0 2 1 3

By comparing the equations above and considering the imaginary parts of the involved terms:
    
sin(nx) n n−1 n n−3
= cot x− cot x ± ... .
(sin x)n 1 3

Given the last identity, we may fix a positive integer m, set n = 2m + 1 and define xr = 2m+1
per r = 1, 2, . . . , m. Since nxr is an integer multiple of π, we have sin(nxr ) = 0. As a consequence:
     
2m + 1 2m + 1 2m + 1
0= cot2m xr − cot2m−2 xr ± · · · + (−1)m
1 3 2m + 1

holds for r = 1, 2, . . . , m. The numbers x1 , . . . , xm are distinct elements of the interval 0, π2 and the function cot2 (x)


is injective on such interval, hence the numbers tr = cot2 xr (per r = 1, 2, . . . , m) are distinct. Given the previous
equation, these m numbers are the roots of the following polynomial with degree m:
     
2m + 1 m 2m + 1 m−1 m 2m + 1
p(t) = t − t ± · · · + (−1) .
1 3 2m + 1

Due to Vieta’s formulas we may compute the sum of the roots of p through its coefficients:
2m+1

2 2 2 3 2m(2m − 1)
cot x1 + cot x2 + · · · + cot xm = 2m+1 = .
1
6

Since csc2 x = cot2 x + 1 we further have:


2m(2m − 1) 2m(2m + 2)
csc2 x1 + csc2 x2 + · · · + csc2 xm = +m= .
6 6
1 rπ
By considering the inequality cot2 (x) < x2 < csc2 (x) and summing its terms on xr = 2m+1 we get:
 2  2  2
2m(2m − 1) 2m + 1 2m + 1 2m + 1 2m(2m + 2)
< + + ... + < .
6 π 2π mπ 6
 2
π
By multiplying both sides by 2m+1 :

π2 π2
     
2m 2m − 1 1 1 1 2m 2m + 2
< 2 + 2 + ··· + 2 < .
6 2m + 1 2m + 1 1 2 m 6 2m + 1 2m + 1

Page 67 / 222
π2
If we consider the limits of the RHS and LHS as m → +∞, both limits equal 6 , hence

π2
X 1  
1 1 1
ζ(2) = 2
= lim 2
+ 2
+ . . . + 2
=
k m→∞ 1 2 m 6
k≥1

1−z
follows by (squeezing). Cauchy’s proof can be shortened a bit by recalling that the Cayley transform z 7→ 1+z
 n
is an involution on C \ {−1}. In particular a solution of 1−z
1+z = eiθ is given by

1 − eiθ/n θ
z= = −i tan 2n
1 + eiθ/n
and the minimal polynomial of tan nπ or cot nπ over Q is simple to derive. A proof relying on squeezing, but not
π
requiring the explicit construction of the minimal polynomial of cot2 2m+1 , is presented in Proofs from The Book :

Proof # 3 (Aigner, Ziegler, 1998).


Assuming 0 < x < π/2 we have that
1 1 1
< 2 <
tan2 x x sin2 x
and we may notice that tan12 x = sin12 x − 1. If we partition the interval 0, π2 in 2n equal subintervals,


then sum both sides of the previous inequality evaluated at xk = π2 · 2kn , we get:
n n n n
2X −1 2X−1 2X−1 2X−1
1 1 1
2 − 1 < 2 < .
k=1
sin xk k=1 k=1
xk
k=1
sin2 xk

By denoting the RHS with Sn we may state that:


n
2X −1  2
n 2 · 2n 1
Sn − (2 − 1) < < Sn .
π k2
k=1

At first sight, an explicit computation of Sn might seem to be difficult, however:

1 1 cos2 x + sin2 x 4
2 + 2 π = = .
sin x sin ( 2 − x) cos2 x · sin2 x sin2 2x

As a consequence, by coupling terms appearing in Sn , except the contribution given by term associated with π4 , we get
4 times a sum with the same structure, but with a doubled “step” and a halved number of terms. The contribution
provided to Sn by the term associated with π4 equals sin2 (π/4)
1
= 2, hence we have the following recurrence formula:

Sn = 4Sn−1 + 2

that together with the initial condition S1 = 2 produces the following explicit formula:

2(4n − 1)
Sn = .
3
As a consequence, the following inequality holds:
n
2 −1
2(4n − 1) 4n+1 X 1 2(4n − 1)
− (2n − 1) ≤ ≤
3 π2 k2 3
k=1

π2
and by considering the limit as n → +∞ we reach, like in Cauchy’s proof, the wanted identity ζ(2) = 6 .

Many solutions to Basel problem are based on trigonometric identities and bits of Theory of Hilbert spaces,
in particular Parseval’s Theorem.

Page 68 / 222
5 THE BASEL PROBLEM

Proof # 4 (Fourier, 1817) On the interval (0, π) the following identity holds pointwise:
π − x X sin(nx)
= .
2 n
n≥1

Additionally the convergence is uniform on any compact subinterval. Due to the orthogonality relations
Z 2π
sin(nx) sin(mx) dx = π δ(m, n)
0

we have that:
2π 2
π3

π−x
Z
= dx
6 0 2
X Z 2π  sin(nx) 2
= dx
n
n≥1 0
X π
= = π ζ(2).
n2
n≥1

R1
There are many orthogonal bases of L2 (0, 1), equipped with the canonical inner product hf, gi = 0 f (x) g(x) dx, so
the above proof admits many variations. For instance, we may consider the Fourier-Legendre expansions of K(m) (the
1
complete elliptic integral of the first kind, having the elliptic modulus as a variable) and √1−m :
X 2 1 X
K(m) = Pn (2m − 1), √ = 2 · Pn (2m − 1).
2n + 1 1 − m n≥0
n≥0

R1 K(m)
These expansions lead to 3 ζ(2) = 0

1−m
dm. On the other hand, by the Taylor series of K,

1 2n 2 Z 1 2n Z 1
 
xn
Z
K(m) πX n X
n dx
√ dm = n
√ dx = π n (2n + 1)
=π √
0 1−m 2 16 0 1−x 4 0 1 − x2
n≥0 n≥0

π2
hence 3 ζ(2) = π arcsin(1). There is a deep connection between Fourier-analytique proofs of the identity ζ(2) = 6
and Euler’s proof #2: f (x) = arcsin (sin x) is a triangle wave.

Proof # 5 (Ritelli at al., 2001)


π2
By combining different ideas from Apostol, Pace and Ritelli, the identity ζ(2) = 6 turns out to be a consequence
of simple manipulations of a double integral:
+∞
4 1 log y
Z
4X 1
ζ(2) = = dy
3 n=0 (2n + 1)2 3 0 y2 − 1
+∞
2 1 1 1 + x2 y 2
Z  
= log dy
3 0 y2 − 1 1 + x2 x=0
4 1 +∞
Z Z
x
= dx dy
3 0 0 (1 + x2 )(1 + x2 y 2 )
4 1 +∞ π2
Z Z
dx dz 4 π π
= 2 2
= · · = .
3 0 0 (1 + x )(1 + z ) 3 4 2 6
Proof # 6 (D’Aurizio, 2015)
The identity
3 X 1 π X (−1)k π π
ζ(2) = 2
= = ·
4 (2n + 1) 2 2k + 1 2 4
n≥0 k≥0

holds as a consequence of the residue Theorem. The integral


Z ∞  y 
y e −1 1 1
I= e 2
− 2y + 1
dy
−∞ y y e

Page 69 / 222
is clearly real, hence the imaginary part of the sum of residues of the integrand function equals zero:
∞ ∞ ∞
! !
1 X (−1)n 2 X 1 2i X (−1)n
I = 2πi − − 2
π n=0 2n + 1 π 2 n=0 (2n + 1)2 π n=0 (2n + 1)2

This argument first appeared at this MSE thread, where the identity
Z +∞ 
4 X (−1)n

x−1 1 dx
− =
0 log2 x log x x2 + 1 π
n≥0
(2n + 1)2

is also shown. We may find a similar symmetry trick in a further proof due to Euler.

Proof # 7 (Euler, 174?)


If we consider the reflection formula for the Γ function
π
Γ(z)Γ(1 − z) =
sin(πz)

d2
and apply the operator dz 2 log(·) to both sides, we get:

π2
ψ 0 (z) + ψ 0 (1 − z) =
sin2 (πz)

From which:  
1 X 1
π2 = 2 ψ0 =2  = 6 ζ(2).
2 1 2
n≥1 n − 2

It is interesting to remark that the previous approach through residues can be encoded in a combinatorial Lemma
not making any explicit mention of residues:

a2n is convergent,
P
Lemma 132. If {an }n≥0 is a weakly decreasing sequence of positive numbers and n≥0
the following series are also convergent
∞ ∞ ∞
def def def
X X X
s = (−1)n an , δk = an an+k , ∆ = (−1)k−1 δk
n=0 n=0 k=1

and we have: X
a2n = s2 + 2∆.
n≥0

1
Proof # 8 (Knopp, 1950) If we consider the sequence defined through an = 2n+1 , we have that:
X 1 π
s= (−1)n =
2n + 1 4
n≥0

∞    
X 1 1 X 1 1 1 1 1
δk = = − = 1 + + ... + .
(2n + 1)(2n + 2k + 1) 2k 2n + 1 2n + 2k + 1 2k 3 2k − 1
n≥0 n≥0

In particular:
X 1  π 2 X (−1)k−1 
1 1

π2 π2 π2
= + 1 + + ... + = + =
(2n + 1)2 4 k 3 2k − 1 16 16 8
n≥0 k≥1

from which it follows that:


4X 1 π2
ζ(2) = 2
= .
3 (2n + 1) 6
n≥0

Page 70 / 222
5 THE BASEL PROBLEM

A crucial step has been hidden “under the carpet”: the equality of blue terms is not entirely trivial. A proof of such
equality (not really efficient, but hopefully interesting) has been included in a box at the end of this section. We finish
by presenting a proof that follows from the contents of the previous sections in a very straightforward way.

Proof # 9 (MSE, 2016) We proved that from the identity


n−1  
Y πk 2n
sin = ,
n 2n
k=1

by switching to logarithms and considering Riemann sums, it follows that:


Z π/2
π
log sin(θ) dθ = − log(2),
0 2
π
hence by the substitution θ 7→ 2 − θ we have:
Z π/2
0= log (2 cos θ) dθ.
0

Due to De Moivre’s formula, 2 cos θ = eiθ + e−iθ = eiθ 1 − e−2iθ .




By exploiting the Taylor series of log(1 + x) at the origin and termwise integration we have:

Z π/2 Z π/2 X (−1)n+1 Z π/2


Im log (2 cos θ) dθ = θ dθ + Im e−2niθ dθ
0 0 n 0
n≥1

(−1)n+1 i 1 − e−πin

2
π X
= − Im
8 2n2
n≥1
2X (−1)n+1 (1 − (−1)n )
π
= −
8 2n2
n≥1

but since this integral has to be zero,


π2 X 1
= .
8 (2m + 1)2
m≥0

2017 addendum. Proof # 10 (MSE, 2017) Here it is another crazy approach by symmetry. For any s > 1 we have
−1 X −1 Z +∞ s−1
(−1)n+1 L−1 1
X 1  
2 2 t
ζ(s) = = 1− s = 1− s dt
ns 2 ns Γ(s) 2 0 e t+1
n≥1 n≥1

where the RHS is converging for any s > 0, providing an analytic continuation of the LHS over such region. By
applying integration by parts twice, we get the following integral representation for the ζ function over the region
s > −2:
 −1 Z +∞ s+1 t t
1 2 t e (e − 1)
ζ(s) = 1− s dt
Γ(s + 2) 2 0 (et + 1)3
ζ(1−s) 2 Γ(s)
cos πs

and due to the reflection formula ζ(s) = (2π)s 2 we have

+∞ 2−s t
(2π)s e (et − 1)
Z
−1 t
ζ(s) = (1 − 2s ) dt
2 Γ(3 − s) Γ(s) cos πs
2 0 (et + 1)3
for any s < 3. By evaluating the previous line at s = 2 and by enforcing the substitution t = log u we get:
+∞ 1
2π 2 u−1 1
2π 2 2π 2 1 π2
Z Z
u7→ 1−v v
ζ(2) = du = dv = · = .
3 1 (u + 1)3 3 0 (2 − v)3 3 4 6

Page 71 / 222
In equivalent terms, the value of ζ(2) can be derived from the value of ζ(−1), which on its turn is related to a Bernoulli
number.

2018 addendum. Proof # 11. We start with the statement of the MacMahon master theorem.
Let A = (aij )m×m be a complex matrix, and let x1 , . . . , xm be formal variables. Consider a coefficient
m
Y ki
G(k1 , . . . , km ) = xk11 · · · xkmm

ai1 x1 + · · · + aim xm .
i=1

Let t1 , . . . , tm be another set of formal variables, and let T = (δij ti )m×m be a diagonal matrix. Then
X 1
G(k1 , . . . , km ) tk11 · · · tkmm = ,
det(Im − T A)
(k1 ,...,km )

where the sum runs over all nonnegative integer vectors (k1 , . . . , km ), and Im denotes the identity matrix of size m.
By considering the matrix  
0 1 −1
A = −1 0 1
 

1 −1 0
it is not difficult to derive an important combinatorial identity due to Dixon:

   
X
k a+b b+c c+a (a + b + c)!
(−1) =
a+k b+k c+k a!b!c!
k∈Z

holding for any a, b, c ∈ N+ . By logarithmic convexity and the Bohr-Mollerup theorem, the range of validity of Dixon’s
identity can be extended to a, b, c ∈ (−1, +∞). If we consider the instance a = b = c = 12 we get:
X 1 3π 2
2 3
= ,
(1 − 4k ) 32
k∈Z

where the LHS, by partial fraction decomposition, can be written as:


X 1 1 1 1 3 1 3 1 3 1 3 1

· − · + · + · + · − ·
8 (2k + 1)3 8 (2k − 1)3 16 (2k + 1)2 16 (2k − 1)2 16 2k + 1 16 2k − 1
k∈Z

3 1
P
which by cancellation and symmetry boils down to 4 k≥0 (2k+1)2 .
π2
The identity ζ(2) = 6 is a straightforward consequence.

R +∞ arctan x 1 +∞ 2
Proof # 12. We may consider that J = 0 2
1+x2 dx = 2 arctan x 0 = π8 .
On the other hand, by Feynman’s trick or Fubini’s theorem
Z +∞ Z 1 Z 1
x − log a
J = 2 2 2
da dx = 2
da
0 0 (1 + x )(1 + a x ) 0 1−a
R1 1 1
and since 0
− log(x)xn dx = (n+1)2 , by expanding 1−a2 as a geometric series we have

π2 X 1
=J = .
8 (2n + 1)2
n≥0

Page 72 / 222
5 THE BASEL PROBLEM

Exercise 133. In order to stress the importance of the dominated convergence Theorem, prove that:
XX m2 − n2 π X X m2 − n2 π
2 2 2
= , 2 2 2
=− .
(m + n ) 4 (m + n ) 4
n≥1 m≥1 m≥1 n≥1

It is possible to tackle the given problem through

m2 − n2 π2
 
X 1 1
= −
m≥1
(m2 + n2 )2 2 n2 sinh2 (πn)

and by exploiting the identity in the next exercise, or by exploiting Poisson summation formula.

Exercise 134 (First steps towards modular forms). Prove the identity:
X 1 1 1
2 = − .
sinh (πn) 6 2π
n≥1

d2 sinh z
Hint: apply the dz 2 log(·) operator to both sides of the equality representing the Weierstrass product for the z
function.

Exercise 135 (First steps towards modular forms). Prove the identity:
X n(−1)n+1 1
= .
sinh(πn) 4π
n≥1

Exercise 136 (Ramanujan’s first steps towards modular forms). Prove the identity:

113 213 313 1



+ 4π
+ 6π
+ ... = .
e −1 e −1 e −1 24

Page 73 / 222
The identity under the carpet. As promised, we are going to prove that:
X (−1)k+1  1 1

π2
1 + + ... + = .
k 3 2k − 1 16
k≥1

According to the terminology introduced by Gioffrè, Iandoli and Scandone, the following proof is a “level 4 proof ”,
since at some point four consecutive symbols for series or integrals appear:
X (−1)k+1  1 1
 X X 2
1 + + ... + = (−1)k+1
k 3 2k − 1 (2n + 1)(2n + 2k + 1)
k≥1 k≥1 n≥0
Z 1Z 1XX
= 2 (−1)k+1 x2n y 2n+2k dx dy
0 0 k≥1 n≥0
x y 2 dx dy
= 2
(1 + y 2 )(1 − x2 y 2 )
(0,1)2
x dx dy x dx dy
= 2 2 2
−2 2 )(1 − x2 y 2 )
1 − x y (1 + y
(0,1)2 (0,1)2
Z 1
X 1 arctanh(y)
= 2 2
− 2 dy
(2n + 1) 0 y(y 2 + 1)
n≥0
Z +∞
X 1 u coth(u) du
(y 7→ tanh u) = 2 −2
(2n + 1)2 0 cosh(2u)
n≥0
Z 1
X 1 z(1 + z 2 )2 log z
(u 7→ − log z) = 2 + 4 dz
(2n + 1)2 0 1 − z8
n≥0
X 1 X 1 1 2

(Feynman’s trick) = 2 − 4 + +
(2n + 1)2 (8k + 2)2 (8k + 6)2 (8k + 4)2
n≥0 k≥0
1X 1
= = 38 ζ(2).
2 (2n + 1)2
n≥0

Pn
In a simpler way, we may set Hn = 1
k=0 2k+1 and notice that:

X arctanh (x) X arctan x


Hn x2n+1 = , (−1)n Hn x2n+1 = ,
1 − x2 1 + x2
n≥0 n≥0

from which it immediately follows that:


X (−1)n Z 1
arctan x
Hn = 2 dx = arctan2 (1).
n+1 0 1 + x2
n≥0

Exercise 137. Prove that: Z +∞


dx 8
=π− .
−∞ (1 + x2 ) cosh2 (πx) π

Exercise 138. Prove the following identity:


XX 1 π6
= .
m2 n2 (m + n)2 2835
m≥1 n≥1

Page 74 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

6 Special functions and special products


This section is dedicated to an introduction the Γ function and its properties. We start by investigating about the
interplay among values of the Riemann ζ function, Bernoulli numbers and power series. We recall that, by considering
the logarithmic derivative of the Weierstrass product for the sine function (Euler docet):

1 π X 1 X z 2h X X
2h+1
− cot(πz) = z = z ζ(2h + 2) = ζ(2k)z 2k−1
2z 2 n2 n2h+2
n≥1 h≥0 h≥0 k≥1
 2k−1  
1 d 1 π
ζ(2k) = − cot(πz)
(2k − 1)! dz 2k−1 2z 2 z=0

By introducing the Bernoulli numbers Bn through their (exponential6 ) generating function

z X Bn
= zn
ez −1 n!
n≥0

we immediately have that every Bernoulli number with odd index equals zero, with the only exception of B1 = − 21 .
This follows from the fact that
z z z z
+ = coth
ez − 1 2 2 2
is clearly an odd function. Additionally, since coth is the logarithmic derivative of
Y z2

sinh(z) = z 1+ 2 2 ,
π n
n≥1

Bernoulli numbers with an even index, B0 = 1, B2 = 21 , B4 = − 30


1 1
, B6 = 42 1
, B8 = − 30 5
, B10 = 66 691
, B12 = − 2730 , . . .,
have alternating signs from B2 onward. Due to the ordinary generating function for {ζ(2n)}n≥1 introduced in the
Basel problem section,

(2π)2k |B2k |
ζ(2k) = ∈ π 2k Q.
2 · (2k)!

Since for any k ≥ 1 we have 1 ≤ ζ(2k) ≤ ζ(2), the previous formula allows a simple estimation of the magnitude of
|Bn |. We may further notice that:
   
x
e −1 x X 1 X Bn n 
1= · x =  xn  ·  x
x e −1 (n + 1)! n!
n≥1 n≥0
n
!
X X Bk
= xn
k!(n + 1 − k)!
n≥0 k=0
n   !
X X n+1 xn
= Bk
k (n + 1)!
n≥1 k=0

implies that for any n > 1 we have:

6 According to the usual terminology, the ordinary and exponential generating functions

for a sequence of complex numbers {an }n≥0 are respectively given by


X X an
(OGF) an z n , (EGF) zn .
n≥0 n≥0
n!

Page 75 / 222
n−1  
1 X n+1
Bn = − Bk ,
n+1 k
k=0

an identity allowing the evaluation of Bernoulli numbers in a recursive fashion. Due to such relation, Bernoulli numbers
play a major role in problems related to (finite) power sums:

Theorem 139 (Faulhaber’s formula). For any natural number p,


n p  
X
p 1 X j p+1
k = (−1) Bj np+1−j .
p + 1 j=0 j
k=1

Since for any k ≥ 1 we have Bk = −k ζ(1 − k), the previous statement can also be written as:
n p−1  
X np+1 X p
kp = − ζ(−j)np−j
p + 1 j=0 j
k=1

from which it follows that ζ(s) for s < 1 can be computed through the regularization of a divergent series. 7 For
√ (1/2)
instance we proved that the sequence given by an = 2 n − Hn is convergent as n → +∞. Due to Faulhaber’s
formula, we may further state that:
 √  √ X (−1)n+1
2 n − Hn(1/2) = (1 + 2) 1

lim √ = −ζ 2
n→+∞ n
n≥1

1

then use the (inverse) Laplace transform to get an integral representation for −ζ 2 :
√ Z
1
 √ Z +∞ X (−1)n+1 e−ns 2 + 2 2 +∞ ds
−ζ 2 = (1 + 2) √ ds = √ 2)
.
0 πs π 0 1 + exp(s
n≥1

A de facto equivalent technique is to apply Ramanujan’s Master Theorem to the exponential generating function
of Bernoulli numbers.

The recurrence relation fulfilled by Bernoulli numbers, together with Lucas Theorem on the behaviour of binomial
coefficients (mod p), leads to an interesting consequence: the denominator of the rational number B2k is a squarefree
integer, given by the product of primes p such that p − 1 is a divisor of 2k:

Theorem 140 (Von Staudt-Clausen). By denoting as P the set of prime numbers,


X 1
∀n ≥ 1, B2n + ∈ Z.
p
p∈P
(p−1)|(2n)

The core of this section is summarized in the following four points.

7 The
P 1 1
famous “claim” n≥1 n = − 12 is preposterous since the LHS is a Pdivergent series. ζ(−1) = − 12 holds, but ζ(s) for s < 1 is
1
defined through an analytic continuation and not directly through the series n≥1 ns , which is convergent only if Re(s) > 1.

Page 76 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

Theorem 141. If s is a complex number with a positive real part, the following definitions are equivalent:

• (A)(Integral representation)
Z +∞
Γ(s) = xs−1 e−x dx
0

• (B)(Euler product)
n!ns
Γ(s) = lim
n→+∞ s(s + 1) · . . . · (s + n)

• (C)(Weierstrass product)

e−γs Y  s −1 s/n


Γ(s) = 1+ e , γ = lim (Hn − log n)
s n n→+∞
n≥1

• (D)(Bohr-Mollerup characterization) On the positive real semiaxis, there is a unique function


with a convex logarithm that fulfills Γ(1) = 1 and Γ(s + 1) = s Γ(s) for any s > 0.

Sketch of proof. (A) ↔ (B). Due to the integration by parts formula, the function defined by (A) fulfills Γ(1) = 1
and Γ(s + 1) = s Γ(s). Due to the dominated convergence Theorem we have:
Z +∞ Z n
s−1 −x
 x n
x e dx = lim xs−1 1 − dx
0 n→+∞ 0 n

where the integral appearing in the RHS can be computed by integration by parts too, proving (B) and (B) → (A).
(B) ↔ (C) follows from simple algebraic manipulations, since any n ∈ Z+ can be written as a telescopic product:
n−1
Y 
1
n= 1+ .
k
k=1

(A) → (D). If s is a positive real number, from (A) it follows that Γ(s) is a moment for a positive random variable,
i.e. an integral of the form R+ xs ω(x) dx with ω(x) being a positive and locally integrable function. In particular,
R

Γ(s) is a continuous positive function and due to the Cauchy-Schwarz inequality:


Z +∞ Z +∞ Z +∞ 2  2
dx dx s1 +s2 dx s1 + s2
Γ(s1 )Γ(s2 ) = xs1 xs2 ≥ x 2 =Γ .
0 xex 0 xex 0 xex 2

It follows that log Γ is a midpoint-convex function, and since it is a continuous function on R+ , it is tout court convex.
(D) → (B). Due to logarithmic convexity, the function of real variable defined through (D) has a representation as
an infinite product and it is not only continuous, but analytic. Due to the analytic continuation principle, the
function defined by (D) can be extended to the whole half-plane Re(s) > 0 and the statement (B) holds in that region
too. Summarizing:

A B

D C
We may also notice that once the Γ(s) function is defined on the half-plane Re(s) > 0, the functional identity
Γ(s + 1) = s Γ(s) allows to extend Γ to the whole complex plane. In particular, from the Weierstrass product it follows
that:

• Γ(s) is a meromorphic function on C, not vanishing at any point;

Page 77 / 222
• the singularities of Γ(s) are simple poles and they are located at the elements of {0} ∪ Z− . Additionally:
(−1)n
∀n ∈ N, Res (Γ(s), s = −n) =
n!
• the Γ function fulfills the following reflection formula:
π
∀z 6∈ Z, Γ(z) Γ(1 − z) = .
sin(πz)

The logarithmic convexity allows to produce tight approximations of the Γ function, like:

Theorem 142 (Gautschi’s inequality). For any real number x > 0 and for any s ∈ (0, 1) we have:

Γ(x + 1)
x1−s < < (x + 1)1−s .
Γ(x + s)

To have extended the factorial function allows us to define binomial coefficients also for non-integer parameters.
In particular, from n! = Γ(n + 1) it follows that:
 
2n (2n)! Γ(2n + 1)
= 2
= .
n n! Γ(n + 1)2
Given these important notions, we are ready to study the behavior of central binomial coefficients.
We may start by noticing that for any n ∈ Z+ ,
  n  
1 2n (2n)! (2n − 1)!! Y 1
= = = 1 − .
4n n (2n)!!2 (2n)!! 2k
k=1

By squaring both sides and performing some algebraic manipulations:


  2 n   n   n   n  −1
1 2n 1Y 1 1 1Y 1 Y 1 1 Y 1
= 1 − + = 1 − 1 + = 1 − .
4n n 4 k 4k 2 4 k 4k(k − 1) 4n (2k − 1)2
k=2 k=2 k=2 k=2
 −1
1
Q
If now we denote as W the infinite product k≥2 1− (2k−1)2 we have:
  2    −1
1 2n W Y 1 W Y 1
= 1− = 1+ .
4n n 4n (2k − 1)2 4n 4k(k − 1)
k>n k>n

Since in a right neighbourhood of the origin we have e > 1 + x (truth to be told, such inequality holds for any x ∈ R∗
x

by convexity), we may state:


  2  
1 2n W X −1 W 1
> exp = exp −
4n n 4n 4k(k − 1) 4n 4n
k>n

as well as:   2  


1 2n W X −1 W 1
< exp = exp −
4n n 4n (2k − 1)2 4n 4n + 2
k>n

by creative telescoping. Essentially, it is enough to find an explicit value for W (also known as Wallis product) to
have accurate approximations of central binomial coefficients. Due to the Weierstrass product for the cosine function,
Y 4z 2

cos(πz) = 1−
(2k + 1)2
k≥0

we have:
1 Y 1

cos(πz) d.H. π
= 1− 2
= lim = ,
W (2k + 1) z→1/2 1 − 4z 2 4
k≥1

hence:

Page 78 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

    
1 2n 1 1 1
=√ 1− +O
4n n πn 8n n2

where the asymptotic expansion still holds if n ≥ 1 is not an integer. By the reflection formula we clearly have
 √
Γ 21 = π, leading to a remarkable consequence:

Lemma 143. √
+∞ +∞
e−t
Z Z
−x2 (A) 1 1
 π
e dx = √ dt = 2 Γ 2 = .
0 0 2 t 2

2
In this framework we will see soon that the area of the unit circle being equal to Γ 12 is not accidental at all. Before
introducing the Beta function and the multiplication formula for the Γ function, let us investigate about the Eulerian
thought “if something is defined by an infinite product, it might be the case to consider its logarithmic derivative ”.
For any complex number with a positive real part, let us set:
d def
ψ(s) =
log Γ(s).
ds
The ψ function is also known as Digamma function. Due to the Weierstrass product for the Γ function, we have:
 
Γ0 (s) X1 
d  Xs  s  1 1
ψ(s) = = − log s − γs + − log 1 + =− −γ+ −
Γ(s) ds k k s k k+s
k≥1 k≥1

where the function relation Γ(s + 1) = s Γ(s) translates into ψ(s + 1) = 1s + ψ(s) for the Digamma function. Such
identity allows an analytic continuation of the ψ function to the whole complex plane. The Digamma function turns
out to be a meromorphic function with simple poles with residue −1 at each non-positive integer. Far enough from
the singularities, the following identities hold:


X 1 1

ψ(b) − ψ(a) = −




 (n + a) (n + b)

 n≥0
ψ(b) − ψ(a)


 X 1

 =
b−a (n + a)(n + b)


n≥0


  
1 1

 X X
ψ(−a )Res =

 k z=a k
Q Q
k (z − ak ) k (n − ak )




 k n≥0
(pi −1)
 pi
(−1) ψ (−ζi ) 1

 X X
=

 Q Q
j6=i i − ζj ) − ζi )pi

Main properties of


i
pi ! (ζ i (n
n≥0
X 1
the Digamma function 
 ψ 0 (a) =

 (n + a)2
n≥0



ψ (n) (a) 1

 X
= (−1)n+1


(k + a)n+1


 n!


 k≥0
ψ(z) − ψ(1 − z)



 = π cot(πz)

 n−1  

 1X k
ψ(nz) = log(n) + ψ z+


n n




 k=0

 ψ(1/2), ψ(1), ψ(n) = −γ − log 4, −γ, Hn−1 − γ.

They are clearly interesting from a combinatorial point of view. On the half-plane Re(s) > 0 we also have the following
integral representations:
Z +∞  −t Z 1
e−st 1 − xs

e
ψ(s) = − dt, ψ(s + 1) = −γ + dx
0 t 1 − e−t 0 1−x

Page 79 / 222
and the Taylor series of ψ(x + 1) at the origin is really simple:
X
ψ(x + 1) = −γ − ζ(k + 1)(−z)k
k≥1

just like the asymptotic expansion for x  1:

1 X B2n
ψ(x) = log(x) − − .
2x 2nx2n
n≥1

The last identity is a typical consequence of the Euler-McLaurin summation formula or, as we will see, of
techniques based on creative telescoping. We remark that by combining the discrete Fourier transform with the
multiplication and reflection formulas for the Γ and ψ functions we get a deep result:

Theorem 144 (Gauss Digamma Theorem). If r, m are positive integers and r < m, we have:

r bX2 c   m−1

π πr 2πnr  πn 
ψ = −γ − log(2m) − cot +2 cos log sin
m 2 m n=1
m m

allowing an explicit evaluation of ψ(s) for any s ∈ Q. A proof can be found on PlanetMath.

The multiplication formulas for the Γ function follow from the multiplication formulas for the ψ function
in a very straightforward way. Since:

s n−1
XZ s 
log Γ(ns) − log Γ(n)
Z
k
= ψ(nz) dz = (s − 1) log n + ψ z+ dz
n 1 n
k=0 1
n−1   n−1  
X k X k
= (s − 1) log n + log Γ s + − log Γ 1 +
n n
k=0 k=0

by exponentiating both sides we get another result due to Gauss:

Theorem 145 (Multiplication formulas for the Γ function). For any s ∈ C with positive real part and for any n ∈ Z+ ,
n−1  
1−n
ns−1/2
Y k
Γ(ns) = (2π) 2 n Γ s+ .
n
k=0

The n = 2 case is also known as Legendre duplication formula:

22z−1
 
1
Γ(2z) = √ Γ(z)Γ z + .
π 2

These identities can also be proved through a slick technique known as Herglotz trick: if two meromorphic functions
f (z), g(z) only have simple poles at the same points with the same residues, they share a value at a regularity point
and they fulfill the same functional equation, they are the same function. In particular, for any z ∈ C with positive
real part we may define
def Γ(2z)
g(z) =
Γ(z)Γ(z + 1/2)

Page 80 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

and check that g(z) is an entire function such that g(1) = √2π and g(z + 1) = 4 g(z). Legendre duplication formula
and, in a similar fashion, Gauss multiplication formulas are so straightforward consequences.
Back to central binomial coefficients, we may notice that:

Γ n + 12
  
1 2n Γ(2n + 1) Γ(2n)
= n = 2n−1 =√
4n n 4 Γ(n + 1)2 2 Γ(n + 1)2 π Γ(n + 1)

1
  √
= π4 is equivalent to the identity Γ 12 = π, and loosely speaking:
Q
hence the identity k≥1 1 − (2k+1) 2

1
 √
Wallis product ←→ Γ 2 = π ←→ Integral of the Gaussian function.

Theorem 146 (Stirling’s approximation). For any n > 0 (without assuming n ∈ Z) the following inequality
holds:  n n √    n n √  
1 1
2πn exp ≤ n! ≤ 2πn exp .
e 12n + 1 e 12n

1 1 1
Proof. Since n2 − n(n+1) = n2 (n+1) , for any m ∈ Z+ we have:

X 1 X 1 1

1 X 1

1

= − + −
n2 n (n + 1) 2 n2 (n + 1)2
n≥m n≥m n≥m
 
1 X 1 1 1 X 1
+ 3
− 3

6 n (n + 1) 6 n3 (n + 1)3
n≥m n≥m

hence: X 1 1 1 1
ψ 0 (m) = 2
≤ + 2
+
n m 2m 6m3
n≥m

and the inequality still holds if m ≥ 1 does not belong to Z. Additionally, in a similar fashion:
1 1 1 1
ψ 0 (m) ≥ + + − .
m 2m2 6m3 30m5
By integrating both sides with respect to the m variable twice, we get that log Γ(m) has the following behaviour:
 
1 1
log Γ(m) ≈ m − log(m) − αm + β +
2 12m

where α = 1 follows from the functional relation log Γ(m + 1) − log Γ(m) = log m. That gives Stirling’s approx-

imation up to a multiplicative constant: β = log 2π then follows from Legendre duplication formula and the
 √
identity Γ 12 = π.

The coefficients found through creative telescoping, due to Faulhaber’s formula, directly depend on Bernoulli numbers.
In particular the Trigamma function ψ 0 (z) has the following asymptotic expansion:
X 1 1 1 X B2t
ψ 0 (z) = 2
= + 2+
(z + m) z 2z z 2t+1
m≥0 t≥1

and by termwise integration:


1 X B2t
ψ(z) = log(z) − − ,
2z 2tz 2t
t≥1

Page 81 / 222

 
1 X B2t
log Γ(z) = z− log(z) − z + log 2π +
2 2t(2t − 1)z 2t−1
t≥1
√ +∞
2 arctan zt
  Z
1
= z− log(z) − z + log 2π + dt.
2 0 e2πt − 1

where the last identity, following from the (inverse) Laplace transform, is also known as second Binet’s Theorem
for log Γ.

We now consider the problem of computing


Z 1
I(a, b) = xa−1 (1 − x)b−1 dx
0

for positive integer values of a and b. By exploiting the substitution x = e−t and the identity
b−1 b−1

−1 −t b−1
X
k k (b − 1)!
L (1 − e ) = (−1) =
s+k s(s + 1) · · · (s + b − 1)
k=0

we have:
+∞
(b − 1)! (a − 1)!(b − 1)!
Z
Γ(a) Γ(b)
I(a, b) = e−ta (1 − e−t )b dt = = = .
0 a(a + 1) · · · (a + b − 1) (a + b − 1)! Γ(a + b)
The last identity holds in a more general context:

Theorem 147 (Eulero). If a, b are complex numbers with positive real parts,
Z 1
def Γ(a)Γ(b)
B(a, b) = xa−1 (1 − x)b−1 dx = .
0 Γ(a + b)

Proof. The function B(a, b) is clearly continuous and non-vanishing on its domain of definition.
Through the substitution x 7→ (1 − x) we have B(a, b) = B(b, a) and due to the integration by parts formula:
a
B(a + 1, b) = B(a, b).
a+b
R1
B(1, 1) = 0 1 dx = 1 holds and on the positive real line both B(·, b) and B(a, ·) are log-convex, since they are moments.
The claim hence follows from the Bohr-Mollerup characterization: B(a, b) is known as Euler’s Beta function.

Euler’s Beta function immediately gives extremely useful integral representations. For instance:

Lemma 148. If α and β are complex numbers with real parts greater than −1,
  β+1 
Z π/2 Γ α+1
2 Γ 2
sinα (θ) cosβ (θ) dθ =   .
0 2 Γ α+β+2
2

Lemma 149. If m and n are natural numbers, we have:


Z π
π(−1)n
 
2m 2m
sin (θ) cos(2nθ) dθ = .
0 4n m+n
Lemma 150. If Re(a) > −1 and Re(b) > Re(a) + 1 we have:
Z +∞ a
t dt π
b
=  .
0 1+t b sin π(a+1) b

Page 82 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

By combining Euler’s Beta function with Feynman’s trick, we get that many non-trivial integrals can be computed in
an explicit way. We immediately study a highly non-trivial example.

Exercise 151. Prove that:


Z π/2
π 2
log3 (sin θ) dθ = − π log 2 + 4 log3 2 + 6 ζ(3) .

0 8

Proof. By differentiation under the integral sign we have that:


√ 
π d3 Γ 12 + α2
Z π/2
3
log (sin θ) dθ = .
2 dα3 Γ 1 + α2

0
α=0

In particular the value of our integral just depends on the values of Γ, Γ0 , Γ00 and Γ000 at the points 21 and 1.
 √
We know that Γ(1) = 1, Γ 21 = π and by differentiating both sides of Γ0 (x) = Γ(x) ψ(x) multiple times we get:

Γ0 (x) = Γ(x)ψ(x),
00
Γ (x) = Γ(x)ψ(x)2 + Γ(x)ψ 0 (x),
Γ000 (x) = Γ(x)ψ(x)3 + 3Γ(x)ψ(x)ψ 0 (x) + Γ(x)ψ 00 (z)

The problem boils down to computing the values of ψ, ψ 0 and ψ 00 at the points 1
2 and 1.
By Gauss Theorem we have:
ψ 12 = −γ − 2 log 2

ψ(1) = −γ,
and we also know that: X 1
ψ 0 (a) =
(n + a)2
n≥0
0 π2 0 1 π2

from which ψ (1) = ζ(2) = 6 and ψ = 3 ζ(2) =
2 follow.2
By differentiating again with respect to the a variable, we get:
X 1
ψ 00 (a) = −2
(n + a)3
n≥0

from which it is simple to derive ψ 00 (1) = −2 ζ(3) and ψ 00 21 = −14 ζ(3).




At this point, in order to prove the claim it is enough to trust in a Computer Algebra System to perform the needed
simplifications, or just perform them by hand with a bit of patience. In a similar way it is possible to prove:
Z π/2
π π3
log(sin θ) log(cos θ) dθ = log2 (2) − ,
0 2 48
Z π/2
π3 π
log2 (sin θ) dθ = + log2 (2),
0 24 2
Z π/2 5
19π π3 π
log4 (sin θ) dθ = + log2 (2) + log4 (2) + 3π log(2) ζ(3).
0 480 4 2

As mentioned before, we investigate now about the interplay between the Γ function and the area of the unit circle.

Since the graph of f (x) = 1 − x2 over [−1, 1] is a half-circle,

1 1
1 √
Z p Z
π = 4 √
1 − x2 dx = 2 1 − u du
0 0 u
Z 1 1 3
 
1 3 Γ Γ 1 2
u 2 −1 (1 − u) 2 −1 du = 2 B 21 , 32 = 2 2 2
 
= 2 =Γ 2 .
0 Γ(2)

Page 83 / 222
The last identity has an interesting generalization: let Vn (ρ) and An (ρ) be the volume and the surface area of the
Euclidean ball with radius ρ ≥ 0 in Rn . For starters, we may notice that:
d
Vn (ρ) = ρn Vn (1), An (ρ) = ρn−1 An (1), An (ρ) = Vn (ρ) = nρn−1 Vn (1),

then we consider the integral Z
exp −(x21 + x22 + . . . + x2n ) dµ.
 
In =
Rn
Due to Fubini’s Theorem, the following identity clearly holds:
Z n
2
In = e−x dx = π n/2 .
R

Due to Cavalieri’s principle we also have:


Z +∞ Z +∞ Z +∞
−ρ2 n−1 −ρ2 An (1) n
In = An (ρ)e dρ = An (1) ρ e dρ = u 2 −1 e−u du
0 0 2 0

hence we get the identity:


2π n/2 = An (1)Γ(n/2)

and finally:

π n/2 2π n/2 n−1


Vn (ρ) = ρn , An (ρ) = ρ .
Γ(1 + n/2) Γ(n/2)

As n ranges over N+ , Vn (1) attains a maximum at n = 5 and An (1) attains a maximum at n = 7.


Additionally, if n ≥ 13 the volume of the unit balls is less than the volume of the unit cube.

Theorem 152 (Bailey-Borwein-Plouffe, BBP formula).


X 1  4 2 1 1

π= − − − .
16k 8k + 1 8k + 4 8k + 5 8k + 6
k≥0

This quite recent result (1995) has opened new frontiers in the problem of finding in a efficient way the digits in the
binary expansion of π. The proof of such identity through the instruments so far acquired is pretty simple:
Z 1 X 8k
X 1  4 4x − 2x8k+3 − x8k+4 − x8k+5

2 1 1
− − − = dx
16k 8k + 1 8k + 4 8k + 5 8k + 6 0 16k
k≥0 k≥0

1
1−x
Z
= 16 dx
0 4 − 4x + 2x3 − x4
Z 1
x dx
(x 7→ 1 − x) = 16
0 (1 + x2 )(1 + 2x − x2 )
Z 1 Z 1
1+x x
= 4 dx − 4 dx
0 1 + x2 0 2 − x2

= (π + log 4) − log 4 = π.

The author conjectures the BBP formula (or series expansions with a similar structure) might be the key for proving
the base-2 normality of the π constant, i.e., loosely speaking, the fact that any binary string appears in the binary

Page 84 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

expansion of π an infinite number of times, possibly with a regular frequency. On its behalf, the base-2 normality of π
n
is deeply related to the (actually unknown) convergence properties of series like n≥1 sin(2 )
P
n and similar ones, linked
with universally bad averaging sequences.

Exercise 153. Prove that: Z 1/2


X n! 2 1 2 1/4
= 2e1/4 e−x dx ≤ + e .
(2n + 1)! 0 3 3
n≥0

Proof. We may notice that:


X n! X Γ(n + 1) X B(n + 1, n + 1)
= =
(2n + 1)! Γ(2n + 2) n!
n≥0 n≥0 n≥0
Z 1X n
x (1 − x)n
= dx
0 n≥0 n!
Z 1
= exp [x(1 − x)] dx
0
Z 1/2
= 2 exp [x(1 − x)] dx
0
1
1 − x2
Z  
= exp dx.
0 4
The last inequality follows from the fact that, by convexity, on the interval [0, 1] we have:
1 − x2
 
exp ≤ 1 + (e1/4 − 1)(1 − x2 ).
4

Exercise 154 (Reflection formula for the Dilogarithm function). Prove that for any x ∈ (0, 1) we have:

ζ(2) − log(x) log(1 − x) = Li2 (x) + Li2 (1 − x)

xn
P
where Li2 (x) = n≥1 n2 .

Proof. Let us define:


f (x) = log(x) log(1 − x) + Li2 (x) + Li2 (1 − x).
We are interested in showing that f is constant, hence we compute f 0 :
 
0 log(1 − x) log(x) log(1 − x) log x
f (x) = − − + = 0.
x 1−x x 1−x
In order to prove the claim it is enough to compute f (x) at a point of regularity, or to compute the limit:

lim f (x) = ζ(2) + lim− log(x) log(1 − x) = ζ(2).


x→1− x→1

We have an interesting corollary:


π2 log2 2
     
X 1 1 1 1
= Li2 = f − log2 2 = − .
2n n2 2 2 2 12 2
n≥1

Page 85 / 222
Exercise 155 (“Continuous” binomial theorem). Prove that for any n ∈ N we have:
Z +∞  
n
dx = 2n .
−∞ x

Proof. By exploiting the reflection formula for the Γ function we have:


 
n n! sin(πx)
= ·
x π (n − 1)(n − 1 − x) · . . . · (1 − x)x
and since
n Z +∞
1 X ck sin(πx)
= , = (−1)k π
(n − 1)(n − 1 − x) · . . . · (1 − x)x x−k −∞ x−k
k=0
(−1)k n
follows from computing a partial fraction decomposition through the residue Theorem, ck = n! k , we have:
Z +∞   n  
n X n
dx = = 2n .
−∞ x k
k=0

sin x
Exercise 156. By recalling that sinc(x) is the function that equals 1 at x = 0 and x anywhere else, prove that for
any couple (α, β) of real numbers in (0, 1) we have:
X π
sinc(nα) sinc(nβ) = .
max(α, β)
n∈Z

Proof. Since
X 2 X sin(nα) sin(nβ)
sinc(nα) sinc(nβ) = 1 + ,
αβ n2
n∈Z n≥1

due to the addition formulas for the sine and cosine functions it is enough to prove the equality

def
X cos(nθ) π2 θ(2π − θ)
∀θ ∈ [0, 2π], g(θ) = = −
n2 6 4
n≥1

following by termwise integration of the Fourier series of a sawtooth wave. In particular,


X g(|α − β|) + g(α + β)
sinc(nα) sinc(nβ) = 1+
αβ
n∈Z
1 
(α − β)2 − (α + β)2 + 2π(α + β − |α − β|)

= 1+
4αβ
1 π
= 1+ [−4αβ + 4π min(α, β)] = .
4αβ max(α, β)

Exercise 157. If for any n ≥ 1 we define


n
Y 3k
An = ,
6k − 4
k=1
P
what is the value of n≥1 An ?

Page 86 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

n
· B n, 13 , we have:

Proof. Since An = 2n

X X n Z 1
An = xn−1 (1 − x)−2/3 dx
2n 0
n≥1 n≥1
Z 1
2(1 − x)−2/3
= dx
0 (2 − x)2
Z 1
2x−2/3
(x 7→ 1 − x) = 2
dx
0 (1 + x)
Z 1
6 du
(x 7→ u3 ) = 3 2
0 (1 + u )

and the last integral can be (tediously) computed through partial fraction decomposition:
X 4π 4 log 2
An = 1 + √ + .
3 3 3
n≥1

Exercise 158. Prove that by setting


def
X sin(n2 x)
f (x) =
n
n≥1

we have that:
π
lim f (x) = .
x→0+ 2

Proof. f (x) is defined by a pointwise convergent series by Dirichlet’s test and by Weyl’s inequality. It follows
that the wanted limit can be computed through a convolution trick, i.e. by expoiting an approximated identity:
Z +∞ X mn X 2m2 n
lim+ f (x) = lim f (x)me−mx dx = lim = lim .
x→0 m→+∞ 0 m→+∞ m2 + n4 m→+∞ 4m4 + n4
n≥1 n≥1

Due to Sophie Germain’s identity we have:


2m2 n
 
m 1 1
= − ,
4m4 + n4 2 2m2 + 2mn + n2 2m2 − 2mn + n2
hence:
X 2m2 n i 
lim = lim H−m(1+i) − Hm(−1+i) − Hm(1−i) + Hm(1+i)
m→+∞ 4m4 + n4 m→+∞ 4
n≥1
i π
= [log(−1 − i) − log(−1 + i) − log(1 − i) + log(1 + i)] = .
4 4
In a similar way it is possible to show that:
X sin(nk x) π
∀k ∈ Z+ , lim+ = .
x→0 n 2k
n≥1

Exercise 159 (Raabe’s log Γ Theorem). Prove that for any positive real number α we have:
Z 1 √
log Γ(x + α) dx = α log α − α + log 2π.
0

Page 87 / 222
Exercise 160 (Glasser’s Master Theorem). Prove that if F and F ◦ ϕ are integrable functions on the real line, where
N
X |ak |
ϕ(x) = |a|x − ,
x − bk
k=1

then the following identity holds: Z +∞ Z +∞


F (x) dx = F (ϕ(x)) dx.
−∞ −∞

For the proof presented here, the author is really grateful to achillehui.

Proof. Let φ(z) be any meromorphic function over C which

1. preserve the extended real line R∗ = R ∪ {∞} in the sense:



φ(R) ⊂ R∗
def
P = φ−1 (∞) = p ∈ C : p poles of φ(z) ⊂ R

=⇒
−1
φ (R) ⊂ R

S
2. Split R \ P as a countable union of its connected components (an , bn ) . Each connected component is an open
n

interval (an , bn ) and on such an interval, φ(z) increases from −∞ at a+
n to ∞ at bn .

3. There exists an ascending chain of Jordan domains D1 , D2 , . . . that cover C,



[
{0} ⊂ D1 ⊂ D2 ⊂ · · · with Dk = C
k=1

whose boundaries ∂Dk are ”well behaved”, ”diverge” to infinity and |z − φ(z)| is bounded on the boundaries.
More precisely, let

 
def
lim R = ∞

 Rk

 = inf |z| : z ∈ ∂Dk  k→∞ k


def
lim Lk2 = 0
R
Lk = |dz| < ∞ and
 ∂Dk  k→∞ Rk
 def  
Mk = sup |z − φ(z)| : z ∈ ∂Dk M = supk Mk < ∞
 

Given such a meromorphic function φ(z) and any Lebesgue integrable function f (x) on R, we have the following
identity: Z ∞ Z ∞
f (φ(x))dx = f (x)dx (∗1)
−∞ −∞

In order to prove this, we split our integral into a sum over the connected components of R \ P .
Z Z XZ bn
f (φ(x))dx = f (φ(x))dx = f (φ(x))dx
R R\P n an

For any connected component (an , bn ) of R \ P and y ∈ R, consider the roots of the equation φ(x) = y. Using
properties (1) and (2) of φ(z), we find there is a unique root for the equation y = φ(x) over (an , bn ). Let us call this
root as rn (y). Enforcing the substitution y = φ(x) the integral becomes
!
XZ ∞ Z ∞
drn (y) X drn (y)
f (y) dy = f (y) dy.
n −∞ dy −∞ n
dy
drn (y)
We can use the obvious fact dy ≥ 0 and dominated convergence theorem to justify the switching of order of
summation and integral.

Page 88 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

This means to prove (∗1), one only need to show


X drn (y) ?
=1 (∗2)
n
dy

For any y ∈ R, let R(y) = φ−1 (y) ⊂ R be the collection of roots of the equation φ(z) = y.
Over any Jordan domain Dk , we have the following expansion

φ0 (z) X 1 X 1
= − + something analytic
φ(z) − y z−r z−p
r∈R(y)∩Dk p∈P ∩Dk

This leads to
φ0 (z)
Z  
X X 1
r− p= z dz
2πi ∂Dk φ(z) − y
r∈R(y)∩Dk p∈P ∩Dk

As long as R(y) ∩ ∂Dk = ∅, we can differentiate both sides and get

φ0 (z)
Z   Z  
X drn (y) 1 1 d 1
= z dz = − z dz
dy 2πi ∂Dk (φ(z) − y)2 2πi ∂Dk dz φ(z) − y
rn (y)∈Dk
Z
1 dz
=
2πi ∂Dk φ(z) − y

For those k large enough to satisfy Rk > 2(M + |y|), we can expand the integrand in last line as

1 1 1 X (y + z − φ(z))j
= = +
φ(z) − y z − (y + z − φ(z)) z j=1 z j+1

and obtain a bound

 
∞ Z ∞  j
(|y| + |z − φ(z)|)j

X drn (y)
 − 1 ≤ 1
X (M + |y|)Lk X M + |y| M + |y| Lk
 |dz| ≤ 2 ≤

rn (y)∈Dk dy 2π
j=1 ∂Dk
|z|j+1 2πR k j=0
R k π Rk2

Lk
Since lim 2 = 0, this leads to
k→∞ Rk
X drn (y) X drn (y)
= lim =1
n
dy k→∞ dy
rn (y)∈Dk

drn (y)
This justifies (∗2) and hence (∗1) is proved. Notice all the dy are positive, there is no issue in rearranging the order
of summation in last line.

Corollary 161.
+∞ +∞ +∞
x2 dx
Z Z Z
parity 1 dx G.M.T. 1 dx π
= = = √ .
x4 + x2 + 1 2 1 2 2 x2+3

0 −∞ x− x +3 −∞ 2 3

Corollary 162. Observe that, for x 6= nπ, n = 0, ±1, ±2, . . ., we have


 
1 1 1 1 1
cot x = lim + + + ··· + +
N →+∞ x x+π x−π x + Nπ x − Nπ
leading to: Z +∞ Z +∞
f (x − cot x) dx = f (x) dx.
−∞ −∞
π
By the substitution x 7→ 2 − x it follows that:
Z +∞ Z +∞
dx dx
= = π.
−∞ 1 + (x + tan x)2 −∞ 1 + x2

Page 89 / 222
Corollary 163. For any a, b ∈ R+ ,
+∞
π −2√ab
Z r
−ax2 + xb2
e dx = e .
0 4a

The Riemann ζ function and its analytic continuation. For any complex number with real part greater
than one the Riemann ζ function is defined through the absolutely convergent series
X 1
ζ(s) =
ns
n≥1

that through the (inverse) Laplace transform admits the following integral representation:
+∞
xs−1
Z
1
ζ(s) = dx.
Γ(s) 0 ex − 1

In the region Re(s) > 1 we further have:


X (−1)n+1
def

2

η(s) = = 1 − s ζ(s),
ns 2
n≥1

but the series defining η(s) has a larger domain of (conditional) convergence, Re(s) > 0.
Due to such fact we are allowed to extend the ζ function to the half-plane Re(s) > 0 through
−1 X
(−1)n+1

2
ζ(s) = 1− s
2 ns
n≥1

 −1 Z +∞ s−1
−1 1 2 x
(by L ) = 1− s dx
Γ(s) 2 0 ex + 1
+∞
4s xs dx
Z
(by parts) = s
2(2 − 2)Γ(s + 1) 0 cosh2 (x)
1
4s
Z
(by substitution) = arctanh(x)s dx
2(2s − 2)Γ(s + 1) 0

1
2s−1
Z  
1+x
= s
logs dx.
(2 − 2)Γ(s + 1) 0 1−x

These integral representations shed a good amount of light on the tight interplay between logarithmic integrals
and values of the ζ function. They also underline that s = 1 is a simple pole with residue 1, that ζ(s) attains
negative values for any s ∈ [0, 1) and that ζ(0) = − 12 . By exploiting Euler’s product

Y −1
1
∀s : Re(s) > 0, ζ(s) = 1− s
p
p∈P

and its logarithmic derivative, we get that there is a very deep correspondence between the distribution of zeroes
of the ζ function in the region 0 < Re(s) < 1 and the distribution of prime numbers:

X X X ρ X X −2n
log p = X − (b + 1) − lim + .
T →+∞ ρ 2n
pm <X |Im(ρ)|<T n≥1

In particular, the Prime Number Theorem (PNT)

Page 90 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

|P ∩ [1, n]| · log n


lim =1
n→+∞ n

is substantially equivalent to the statement the ζ function is non-vanishing on the line Re(s) = 1, statement
that was almost simultaneously (but independently) proved by Hadamard and de la Vallée-Poussin through
a trigonometric trick. By defining π(x) as the number of prime numbers in the interval [1, x], the following
strengthening of the PNT Z x


dt
 x log2 (x)

π(x) −
2 log t

is substantially equivalent to Riemann Hypothesis (RH): all the zeroes of the ζ(s) function in the region
0 < Re(s) < 1 lie on the critical line Re(s) = 21 . Conversely, inequalities of the form
Z x

π(x) − dt α
x
2 log t

for some α ∈ 21 , 1 imply the absence of zeroes of the ζ function in subsets of the critical line 0 < Re(s) < 1.


The best result actually known about the zero-free region for the ζ function is due to Korobov and Vinogradov:
!!
X log3/5 X
log p = X + O X exp −c .
pm <X
(log log X)1/5

This result (pitifully very far from RH) comes from a sophisticated combinatorial manipulation of exponential sums
(variants on Van Der Corput’s trick), combined with classical inequalities in Complex Analysis (Hadamard,
Borel-Caratheodory). The reader can find on Terence Tao’s blog a very detailed lecture. The ζ function can
be further extended to the whole complex plane. By setting
X 2
θ(z) = eπin z
n∈Z

for any z ∈ C with positive real part, we have


 
1 1
θ(z) = √ θ
−iz z

as a consequence of Poisson summation formula, hence the function


s
def
ξ(s) = π −s/2 Γ ζ(s)
2
has the following integral representation
+∞
θ(iy) − 1  s/2
Z
1  dy
ξ(s) = − + y + y (1−s)/2
s(1 − s) 1 2 y

and ξ turns out to be a meromorphic function (with simple poles at s = 0 and s = 1) such that ξ(s) = ξ(1 − s).
This is known as the reflection formula for the ζ function.

Exercise 164 (Exploiting Euler’s product in a quantitative way).


By defining P as the set of prime numbers, prove that for any real number s > 1 we have:
Y −1
1 X 1
1− s = ζ(s) =
p ns
p∈P n≥1

Page 91 / 222
then prove the following inequality: r
X 1 5
≤ log .
p2 2
p∈P

Hint: for any x ∈ (0, 1) we have x ≤ arctanh x.

Exercise 165. Prove that: X 3


(ζ(2n) − 1) = .
4
n≥1

Proof. By the generating function for the sequence {ζ(2n)}n≥1 ,

z2
 
X 1 − πz cot(πz) d.H. 3
(ζ(2n) − 1) = lim− − = .
z→1 2 1 − z2 4
n≥1

As an alternative,
+∞ +∞ +∞
x2n−1
XZ   Z Z
X 1 1 sinh(x) t+1 3
(ζ(2n) − 1) = x
− x dx = dx = dt =
0 (2n − 1)! e −1 e 0 e (ex − 1)
x
1 2t3 4
n≥1 n≥1

or simply:
 
X XX 1 X 1 1 X 1 1 H2
(ζ(2n) − 1) = = = − = .
m2n m2 − 1 2 m−1 m+1 2
n≥1 n≥1 m≥2 m≥2 m≥2

Exercise 166. Prove that for any s ≥ 1 the following inequality holds:
Z +∞ r
dx π
2 )s

0 (1 + x 4s −3

and notice that it implies π < 10 (Hint: evaluate both sides at s = 3).

Exercise 167. Let A be the set (Z × Z) \ {0, 0}. Prove that:


X 1 √
≤ 4 ζ(4) + π 2 ζ(3).
m4 +n 4
(m,n)∈A

Exercise 168. Prove that: X (−1)n ζ(4n + 2)


π
= · .
2 2n + 1 4n
n≥0

Page 92 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

Exercise 169. Prove the inequality:  


X
2 1
log 1+ < 1.
n
n≥1

2Hn−1 n
Proof. By recalling log2 (1 − z) =
P
n≥1 n z and the integral representation for harmonic numbers we have:

X 
1 X X 2Hn−1 (−1)n

2
S= log 1+ =
m nmn
m≥1 m≥1 n≥1
Z 1XX
2Hn−1 (z n−1 − 1)(−1)n
= dz
0 m≥1 n≥1 nmn (z − 1)
Z 1X z 1
 
log 1 + m − z log 1 + m
= 2 dz
0 m≥1 z(1 − z)
Z 1
log Γ(1 + z)
(by Euler’s product for the Γ function) = −2 dz
0 z(1 − z)
Z 1  
z
(by integration by parts) = 2 ψ(z + 1) log dz
0 1−z
Z 1 
1 1
(by the reflection formula for ψ) = 2 − − π cot(πz) log(z) dz
0 z 1−z
 
The magic now comes from studying the function g(z) = z1 − 1−z 1
− π cot(πz) over the interval (0, 1).
It is extremely close to 2z − 1 by “cancellation of singularities”, hence
Z 1
S≈2 (2z − 1) log(z) dz = 1.
0
1
g(z) is symmetric with respect to z = 2 and log(z) is negative and increasing over (0, 1), hence ≈ is indeed a <.

Exercise 170. Prove that the previous statement implies


 
X n+1
log2 ≤ 2 log 2 + 2 − log2 2.
n−1
n≥2

Exercise 171 (Putnam 2016, Problem B6). Prove that:


X (−1)k+1 X 1
= 1.
k k2n + 1
k≥1 n≥0

Proof. The dominated convergence Theorem allows us to perform the following manipulations:
X (−1)k+1 X 1 X X (−1)k  1 1 1
 X h X (−1)k+1
h+1 2
0
= − + − . . . = (−1)
k k2n + 1 k k2n k 2 4n k 3 8n 2h − 1 k h+1
k≥1 n≥0 n≥0 k≥1 h≥1 k≥1
Z +∞ Z +∞
X0 1 X (−1)h+1 xh
= (−1)h+1 ζ(h + 1) = dx = e−x dx = 1
0 ex − 1 h! 0
h≥1 h≥1
P0
where denotes a regularized sum in the Cesàro sense.

Page 93 / 222
An exercise on generalized Euler sums, logarithmic integrals and values of ζ.

Exercise 172. Prove that



X (−1)k 3
Hk Hk−1 = ζ(4).
k2 16
k=2

Proof. We are going to see a generalization of the approach used (11) to show that ζ(4) = 25 ζ(2)2 . This is also
an opportunity to make a tribute to Pieter J. de Doelder (1919-1994) from Eindhoven University of Technology,
who evaluated in closed form the given series in a somewhat famous paper published in 1991. One may start by
using the following identity coming from the Cauchy product,

X Hn n+1
ln2 (1 + x) = 2 (−1)n−1 x
n=1
n+1

giving

1
ln(1 − x) ln2 (1 + x)
Z Z 1
n−1 Hn
X
dx = 2 (−1) xn ln(1 − x) dx,
0 x n=1
n + 1 0

then using the standard evaluation


Z 1
Hn+1
xn ln(1 − x) dx = − , n ≥ 0,
0 n+1
one gets

1
ln(1 − x) ln2 (1 + x)
Z X Hn Hn−1
dx = 2 (−1)n−1 .
0 x n=2
n2
Here are the main steps which de Doelder took to evaluate the related integral. We clearly have

1 1
ln3 (1 + x) 1
ln(1 − x) ln2 (1 + x)
Z   Z Z
1+x dx
ln3 = dx − 3 dx
0 1−x x 0 x 0 x

1
ln2 (1 − x) ln(1 + x) 1
ln3 (1 − x)
Z Z
+3 dx − dx
0 x 0 x

and

ln3 1 − x2
Z 1 3
1
ln(1 − x) ln2 (1 + x)
Z  Z 1
ln (1 + x)
dx = dx + 3 dx
0 x 0 x 0 x

1
ln2 (1 − x) ln(1 + x) 1
ln3 (1 − x)
Z Z
+3 dx + dx,
0 x 0 x

subtracting the two equalities,

ln(1 − x) ln2 (1 + x) ln3 1 − x2


1 1
 Z 1   Z 1 3
ln (1 − x)
Z Z
1 + x dx
6 dx = dx − ln3 −2 dx
0 x 0 x 0 1−x x 0 x

= I1 − I2 − 2I3 .

Page 94 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

It is easy to obtain

1
ln3 1 − x2 1 1 ln3 (1 − u)
Z  Z
I1 = dx = du (u = x2 )
0 x 2 0 u
1 1 ln3 v
Z
= dv (v = 1 − u)
2 0 1−v
∞ Z
1X 1 n 3
= v ln v dv
2 n=0 0

X 1 π4
= −3 = − ,
n=1
n4 30

similarly

1
ln3 (1 − x) π4
Z
I3 = dx = − .
0 x 15

1−x dx −2 du
By the change of variable, u = , one has = getting
1+x x 1 − u2

1
ln3 u
1
Z   Z
1+x dx
I2 = ln3 = −2 2
du
0 1−x x 0 1−u
∞ Z 1
X
= −2 u2n ln3 u dv
n=0 0

X 1 π4
= 12 4
= .
n=0
(2n + 1) 8

Then,

1
ln(1 − x) ln2 (1 + x) π4
Z
dx = −
0 x 240
and

X Hn Hn−1 3 π4
(−1)n = ζ(4) =
n=2
n2 16 480
as wanted.

An interesting integral related to Lambert’s function.

Exercise 173. Prove that ∞


ex + 1
Z
I1 = dx = 1.
−∞ (ex − x + 1)2 + π 2

Proof. We are going to exploit the following Lemma:


Lemma 174. If a > 0 and b ∈ R,
Z +∞
a2 dx 1
=
1 + W a1 e−b/a
x 2 2

−∞ (e − ax − b) + (aπ)

Page 95 / 222
where W is Lambert’s function, i.e. the principal branch of the inverse function of x 7→ xex .

1
Let us consider the function f (z) = a log(−z)+b−z ·
1 ≥0
z on the region D = C \ R . For any z ∈ D it is
possible to pick some r > 0 and some ϕ ∈ − π2 , π2


such that −z = reiϕ . This leads to:

a log(−z) + b − z = a log reiϕ + b + reiϕ




= a log r + iaϕ + b + r cos ϕ + ir sin ϕ

= (a log r + r cos ϕ + b) + i(aϕ + r sin ϕ)

where the imaginary part of the RHS has the


same sign as ϕ and the real part is a mono-
tonic function of the r variable. It follows that
z = −a · W a1 e−b/a is the only pole of f , at-


tained at ϕ = 0 and r = a · W a1 e−b/a . A simple




computation gives

  
1 −b/a 1 1
Res f (z), z = −a · W e = · ,
a a 1 + W a1 e−b/a

therefore, according to the residue Theorem:


Z Z Z Z
1 2πi
f dz + f dz + f dz + f dz = ·
a 1 + W a1 e−b/a

γR,ε CR δR,ε cε

R
From L(CR ) ∼ 2πR and M (CR ) = maxz∈CR |f (z)| ∼ R12 it follows that CR f dz ≤ L(CR ) M (CR ) ∼ 2π R.

R
Therefore the contribution of CR f dz is negligible as R → +∞. From L(cε ) = πε and M (cε ) = maxz∈cε |f (z)| ∼
R
−1 1
≤ L(cε ) M (cε ) ∼ a −π +
R
· it follows that f dz log ε , hence the contribution of cε f dz is negligible as r → 0 .

a log ε ε cε
By letting R → +∞ and ε → 0+ we get the equality
Z ∞  
1 1 dx 1 2πi
− = ·
a 1 + W a1 e−b/a

0 a log(−x − i0+ ) + b − x a log(−x + i0+ ) + b − x x

and since:
1 1 2πi · a
− = ,
a(log x − iπ) + b − x a(log x + iπ) + b − x (a log x + b − x)2 − (iπa)2

the previous identity has the following consequence:



a2
Z
dx 1
· = .
0 (a log x + b − x)2 + (aπ)2 x 1 + W a1 e−b/a

The claim then follows by enforcing the substitution x 7→ ex , leading to:



a2
Z
1
dx = .
−∞
x 2
(ax + b − e ) + (aπ) 2 1 + W a1 e−b/a

In our case, by choosing a = 1 and b = −1 we get that I1 depends on W (e) = 1 and equals 1.

Page 96 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

Due to the identity


Z +∞
u+1 du
I1 = ·
0 u (u + 1 − log u)2 + π 2
Z +∞ Z +∞
1
= (u + 1)ux−1 e−(u+1)x sin(πx) dx du
π 0 0
2 +∞ −x −x
Z
= e x Γ(x) sin(πx) dx
π 0
Z +∞ −x −x
e x
= 2 dx
0 Γ(1 − x)

the previous Lemma also proves the highly non-trivial identity


Z +∞
(ex)−x 1
dx = (HNT)
0 Γ(1 − x) 2

equivalent to the claim I1 = 1. It would be interesting to find an independent proof of (HNT), maybe based
on Glasser’s master theorem, Ramanujan’s master theorem or Lagrange inversion. There also is an interesting
discrete analogue of (HNT),
X nn
=1
n≥1
n!(4e)n/2

that comes from the Lagrange inversion formula.

A remark on some series involving sinh or cosh.


From the Weierstrass product
Y z2

cosh(πx/2) = 1+
(2m + 1)2
m≥0
2
d
by applying dz 2 log(·) to both sides we get:

π2 X (2m + 1)2 − z 2
=
8 cosh2 (πx/2) m≥0 ((2m + 1)2 + z 2 )

If we replace z with (2n + 1) and sum over n ≥ 0,


X π2 X X (2m + 1)2 − (2n + 1)2
2 =
n≥0
8 cosh (π(2n + 1)/2) n≥0 m≥0 ((2m + 1)2 + (2n + 1)2 )2

where the RHS of (3) can also be written as


XXZ +∞ XZ +∞
−(2m+1)x x cos((2n + 1)x)
cos((2n + 1)x)xe dx = dx
0 0 2 sinh(x)
n≥0 m≥0 n≥0

or, by exploiting integration by parts,


X Z +∞ x cosh(x) − sinh(x) sin((2n + 1)x)
· dx
0 2 sinh(x) 2n + 1
n≥0

On the other hand, n≥0 sin((2n+1)x) π


P
2n+1 is the Fourier series of a 2π-periodic rectangle wave that equals 4 over (0, π)
π
and − 4 over (π, 2π). That implies, by massive cancellation:
X 1 1
2 = .
cosh (π(2n + 1)/2) 2π
n≥0

Page 97 / 222

1 s 8π
On the other hand, the Fourier transform of cosh2 (πx)
is given by by sinh(πs) .
By Poisson’s summation formula,
X (−1)n+1 n 1
= .
sinh(πn) 4π
n≥1

Exercise 175. Prove that √ !


Z π/4
2 cos(3φ)
arctan p dφ = 0.
0 (3 + 2 cos(2φ)) cos(2φ)

Proof. By replacing φ with arctan(t), then using integration by parts, we have:


Z 1 √ ! Z 1 √
1 2(1 − 3t2 ) π2 3 2 t arctan(t)
I= 2
arctan √ dt = − √ dt.
0 1+t 8
2
(5 + t ) 1 − t 2 2 2
0 (3 − t ) 1 − t

Now comes the magic. Since:


√ r
1 − t2
Z
3 2t
√ dt = −3 arctan
(3 − t2 ) 1 − t2 2
integrating by parts once again we get:
r
1
π2 1 − t2
Z
1
I= −3 arctan dt
8 0 1 + t2 2

hence we just need to prove that:


1
r √1 √
1 − t2 arctan 1 − 2t2 π2
Z Z
dt 2
arctan = dt =
0 1 + t2 2 0 1+t2 24

and this is not difficult since both


1 2m+1
1 √
(1 − 2t2 )
Z Z
dt 2m+1 2 2
(1 − t2 ) 2 , dt
0 1 + t2 0 1 + t2

can be computed through the residue theorem or other techniques. For instance:
2m+1
1 1
m + 32 Γ n + 12
 
(1 − t) nΓ
Z Z
2 2m+1
n− 21
X X
n
1 dt = (−1) (1 − t) 2 t dt = (−1)
0 t 2 (1 + t) 0 Γ(m + n + 2)
n≥0 n≥0

or just:
q
Z 1 1−t2  
2 π 1
2 dt = 1− √
1 + 1−t 2(1 + u2 )

0 (1 + t2 ) 2 u
2 2 + u2
from which: r
1 1
1 − t2 π2
Z Z  
dt π du 1
arctan = 1− √ =
0 1 + t2 2 2 0 1 + u2 2 + u2 24
as wanted, since: Z
du u
√ = arctan √ .
(1 + u2 ) 2 + u2 2 + u2

Page 98 / 222
6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

Exercise 176. Prove that for any a > −2 the following identity holds:
Z 1  √ 
1−x dx 1 a+2
·√ =√ log 1 + .
0 1+x x4 + ax2 + 1 a+2 2

Proof. Simple algebraic manipulations allow us to write the given integral in the following form:
Z +∞  
2 du
−1 + √ √
4+u 2 2
u u +a+2
0

and by setting u = a + 2 sinh θ we get:
 
Z +∞
1 2  dθ
√ −1 + q
a+2 0 2 sinh θ
4 + (a + 2) sinh θ

We may get rid of the last term through the ”hyperbolic Weierstrass substitution”
 v 
e +1
θ = 2 arctanh(e−v ) = log
ev − 1

that wizardly gives  


Z +∞
1 2
√ −1 + q  dv
a+2 0 4+ a+2
sinh2 v

i.e. a manageable integral through differentiation under the integral sign.


To fill in the missing details is an exercise we leave to the reader.

Exercise 177 (Ahmed’s integral). Prove that:


1

arctan x2 + 2 5π 2
Z
√ dx = .
0 (x2 + 1) x2 + 2 96

Proof. In 2001-2002, Zahar Ahmed proposed the above integral in the American Mathematical Monthly (AMM). Here
we present his maiden solution. Let us call the given integral as I and use arctan z = π2 − arctan z1 to split I as I1 − I2 .
Using the substitution x = tan θ, we can write
Z π/4
cos θ π
I1 = p dθ
0 2
2 − sin2 θ
π2

which can be evaluated as I1 = 12 by using the substitution sin θ = 2 sin ϕ. Next we use the representation
Z 1
1 1 dx
arctan =
a a 0 x2 + a2
to express
Z 1 Z 1
dx dy
I2 = .
0 0 (1 + x2 )(2 + x2 + y 2 )
By partial fraction decomposition I2 can be re-written as:
Z 1 Z 1 Z 1 Z 1
dx dy dx dy
I2 = − .
0 0 (1 + x2 )(1 + y 2 ) 0 0 (1 + x2 )(2 + x2 + y 2 )

Page 99 / 222
Using the symmetry of the integrands and the domains for x and y, the second integral in the RHS of the last identity
equals I2 itself. This leads to:
Z 1 2
π2 1 dx π2 π2 5π 2
I= − 2
= − = .
12 2 0 1+x 12 32 96

Exercise 178. Prove the following identities:


π/2 √ π2 π/2 √ π2
Z Z
arccot 1 + csc θ dθ = , arccsc 1 + cot θ dθ = .
0 8 0 12

Exercise 179. Prove the following identity:



1/ 2
arcsin(x2 ) π2
Z
√ dx = .
0 (2x2 + 1) x2 + 1 144

Exercise 180. Prove the following identity:


Z 1
√ !
88 21 dx
ζ(2) = arctan √ .
0 215 + 36x2 1 − x2

Exercise 181. Prove the following identity:


X H2 17π 4
n
= .
n2 360
n≥1

Proof. It is very practical to recall that


 
(2)
X 3 Hn2 − Hn
− log3 (1 − z) = z n+1 ,
n+1
n≥1

immediately leading to the following intermediate identity:

X H 2 − Hn(2) 1 1 − log3 (1 − z)
Z
1 1 − log3 z
Z
π4
n
= dz = dz = .
(n + 1)2 3 0 z 3 0 1−z 45
n≥1

We may notice that

X H2 2
X Hn−1 X Hn−1
n
S= = +2 + ζ(4)
n2 n 2 n3
n≥1 n≥1 n≥1
2 (2) (2)
X Hn−1 − Hn−1 X Hn−1 π4
= 2
+ 2
+
n n 60
n≥1 n≥1

Page 100 / 222


6 SPECIAL FUNCTIONS AND SPECIAL PRODUCTS

P Hn−1
since the value of ζ(4) is known and the Euler sum n≥1 n3 can be tackled through the Theorem (41).
On the other hand, by symmetry:
(2)
X Hn−1 X 1 1 π4
ζ(2)2 − ζ(4) =

2
= 2 2
=
n m n 2 120
n≥1 m,n≥1,
m<n

from which it follows that


π4 π4 π4 17π 4
S= + + =
45 120 60 360
as wanted.

Exercise 182. Prove that by applying Feynman’s trick to an integral representation for ζ(s) in the region Re(s) ∈ (0, 1)
we have: Z +∞
log(x) log(2) log(8π 2 )
dx = − .
0 e2πx + 1 4π

Exercise 183. By exploiting Euler’s Beta function, convolutions and the discrete Fourier transform, prove that for
any n ∈ N the following identity holds:
2n   −1
X 4n 2n 1
(−1)k = .
2k k 1 − 2n
k=0

Proof.
2n  Z 1
X 4n k
S(n) = (2n + 1) (−1) (1 − x)k x2n−k dx
2k 0
k=0
2n  Z 1
X 4n
= (2n + 1) (−1)k 2z(1 − z 2 )k z 4n−2k dx
2k 0
k=0
Z π/2
sin(θ) cos(θ) e4niθ + e−4niθ dθ
 
= −(2n + 1)
0
Z π/2
= −(2n + 1) sin(2θ) cos(4nθ) dθ
0
2n + 1 1
= − =− .
4n2 − 1 2n − 1

Exercise 184. By using suitable substitutions, the Laplace transform and special values of Γ0 , ζ, ζ 0 , prove that:
Z 1 
1 1 dt 1 log 2 log π 3 X log n
+ 2
= + − − 2 .
0 log t 1 − t (1 + t) 2 3 4 2π n2
n≥1

Exercise 185. By enforcing the substitution x 7→ 1−t


1+t , check that
Z 1
arctan x q
5 2
dx = π2 Γ

p 4 .
0 x(1 − x2 )

Page 101 / 222


7 The Cauchy-Schwarz inequality and beyond

Definition 186. A Hilbert space is a vector space H (real or complex) equipped with a positive definite inner
product h·, ·i such that (H, k · k), where kuk2 = hu, ui, is a complete metric space.

Hilbert spaces like `2 or L2 (respectively the space of square-summable sequences and the space of square-integrable
functions) are the most natural places to extend the theory of inner products and orthogonal projections on Rn :
Fourier series and Fourier transforms are so natural products of this viewpoint. The context of Hilbert spaces is
also the typical framework for important inequalities like Bessel’s inequality (becoming Parseval’s identity under the
completeness assumption) and the Cauchy-Schwarz inequality:

Theorem 187 (Cauchy-Schwarz inequality). If (a1 , . . . , an ) and (b1 , . . . , bn ) are n-uples of real numbers, we have:

n
!2 n
! n
!
X X X
ak bk ≤ a2k · b2k
k=1 k=1 k=1

and equality holds if and only if a n-uple is a real multiple of the other one.

The integral form of Cauchy-Schwarz inequality is sometimes mentioned as Bunyakovskii’s inequality:

Theorem 188. If f, g are real, square-integrable functions over (a, b),


!2 ! !
Z b Z b Z b
2 2
f (x) g(x) dx ≤ f (x) dx · g(x) dx
a a a

f (x) g(x)
and equality holds if and only if g(x) = λ, or f (x) = λ, almost everywhere in (a, b).

We start this section by studying multiple proofs of Cauchy-Schwarz inequalities, among “classical” approaches and
less typical ones. The first proof depends on a amplification trick.
Proof # 1. Le us assume that u, v are non-zero vectors. From the trivial inequality

ku − vk2 ≥ 0

it follows immediately:
hu, vi ≤ 21 kuk2 + 12 kvk2
where by considering some λ > 0, the inner product hu, vi stays unchanged if u is replaced by λu and v is replaced by
v
λ , implying:
2
∀λ > 0, hu, vi ≤ λ2 kuk2 + 2λ1 2 kvk2
and we have complete freedom in choosing λ in such a way that the RHS is as small as possible.
kvk
With the optimal choice λ = kuk we get:
hu, vi ≤ kuk · kvk
that is precisely Cauchy-Schwarz inequality: the inner product between two vectors is bounded by the product of their
lengths 8 . It is pretty clear that equality holds only if u and v are linearly dependent, and this is even more evident
as a consequence of the next proof.
8 This is without doubt a really efficient mnemonic trick for recalling if CS holds as ≤ or ≥ when needed.

Page 102 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

Theorem 189 (Lagrange’s identity).

n
! n
! n
!2
X X X X
a2i · b2i = ai bi + (aj bk − bj ak )2 .
i=1 i=1 i=1 j6=k

To check the last identity is straightforward, and what is the best way for proving an inequality, than deriving it from
an identity? Given Lagrange’s identity, it is clear that Cauchy-Schwarz inequality holds as an equality if and only if
aj bk = ak bj for any k 6= j. 9
The most well-known proof exploits the concept of discriminant for a quadratic polynomial. The function
n
X
p(t) = (ai − tbi )2
i=1

is clearly a quadratic polynomial, not assuming any negative value for any t ∈ R (as a sum of squares).
That implies the discriminant of p(t) is non-positive, and from
n
X n
X n
X
2
[t ]p(t) = b2i , 1
[t ]p(t) = −2 a i bi , 0
[t ]p(t) = a2i
i=1 i=1 i=1

Cauchy-Schwarz inequality immediately follows. Another approach, usually proposed to students as a tedious exercise,
is to prove Cauchy-Schwarz inequality by induction on n. Such approach is not tedious anymore (quite the opposite,
indeed) if we bring the Cauchy-Schwarz inequality into a (almost) equivalent form:

Lemma 190 (Titu). If (a1 , . . . , an ) and (b1 , . . . , bn ) are n-uples of positive real numbers, we have:
n
X a2 k (a1 + . . . + an )2
≥ .
bk b1 + . . . + bn
k=1

We may firstly notice that Titu’s Lemma is a consequence of the Cauchy-Schwarz inequality:
2 n
! n
!
a2k

2 a1 p an p CS X X
(a1 + . . . + an ) = √ · b1 + . . . + √ · bn ≤ · bk .
b1 bn bk
k=1 k=1

Conversely, Titu’s Lemma implies that Cauchy-Schwarz inequality holds for positive n-uples.
On the other hand the n = 1 case of Titu’s Lemma is trivial and the n = 2 case
a21 a2 (a1 + a2 )2
+ 2 ≥
b1 b2 b1 + b2
is equivalent to the inequality (a1 b2 − a2 b1 )2 ≥ 0. By combining the cases n = 2 and n = N ,
N +1 N
X a2k a2N +1 X a2k
= +
bk bN +1 bk
k=1 k=1
a2N +1 (a1 + . . . aN )2
(case n = N ) ≥ +
bN +1 b1 + . . . + bN
(a1 + . . . + aN +1 )2
(case n = 2) ≥
b1 + . . . + bN +1
the proof of Titu’s lemma turns out to be straightforward.
9 This
rises a hystorical/epistemological caveat: since the Cauchy-Schwarz inequality immediately follows from Lagrange’s identity,
why such inequality was not attributed to Lagrange, whose work came before Cauchy’s work by at least thirty years?

Page 103 / 222


We may notice that the triangle inequality is a consequence of the Cauchy-Schwarz inequality:

hu, vi ≤ kuk · kvk −→ ku + vk2 ≤ kuk2 + kvk2 + 2kuk · kvk −→ ku + vk ≤ kuk + kvk.

The Cauchy-Schwarz inequality can also be used to reconcile the typical definitions of inner product in Physics and
Linear Algebra. By assuming that u, v ∈ R2 , u = (ux , uy ), v = (vx , vy ), hu, vi can be equivalently defined as ux vx +uy vy
or as kukkvk cos θ, where θ is the angle between the u and v vectors. How they come to be equivalent? Simple: on
one hand, kukkvk cos θ is the product between the length of v and the length of the projection of u along v, that we
may name w. w is a vector of the form λv minimizing ku − λvk or, equivalently:

ku − λvk2 = kuk2 − 2λ(ux vx + uy vy ) + λ2 kvk2 .


u v +u v
Such quadratic polynomial in λ attains its minimum value at λ = x kvk x
2
y y
, reconciling the previous definitions and
proving that two vectors have a zero inner product if and only if they are orthogonal.

Exercise 191. Given A, B ∈ R+ , prove that


p
max (A sin θ + B cos θ) = A2 + B 2 .
θ

Proof. Due to Cauchy-Schwarz inequality,


2
(A sin θ + B cos θ) ≤ A2 + B 2
A
and equality holds as soon as tan θ = B.


Exercise 192. Prove that for any N ≥ 1 we have HN ≤ 2N .

Proof. Due to Cauchy-Schwarz inequality


v
N u N
X 1 u X 1 p
HN = ≤ tN · 2
≤ N ζ(2).
n=1
N n=1
n

Exercise 193. Prove that if X


an
n≥1

is a convergent series with positive terms, then


X 1
n2 an
n≥1

is a divergent series.

Proof. Due to Cauchy-Schwarz inequality,


N
! N
! N
!2
X X 1 X 1
an · 2a
≥ ≥ log2 (N ).
n=1 n=1
n n n=1
n

Page 104 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND


Exercise 194. Prove that 3 log(3) < 2.

Proof. s
Z 2 Z 2 Z 2
dx CS dx 2
log(3) = < dx = √ .
0 1+x 0 (1 + x)2 0 3

Exercise 195. Prove that


+∞
Z r
sin x π
I= dx ≤ .
0 x+1 8

Proof. We are dealing with the integral of an oscillating function:


 in order to improve the convergence speed,
1 −1 1 −s
it is very practical to consider that L(sin x) = s2 +1 and L x+1 = e , from which:
sZ
+∞ +∞ +∞
e−s
Z Z r
CS ds π
I= 2
ds ≤ e−2s ds = .
0 s +1 0 0 (1 + s2 )2 8

This approximation is quite accurate, since the functions e−2s ed 1


(1+s2 )2 have a very similar behaviour in a right
neighbourhood of the origin.

Exercise 196. In a acute-angled triangle ABC, LA , LB , LC are the feet of the angle bisectors from A, B, C.
By denoting as r the inradius, prove that:

ALA + BLB + CLC ≥ 9r.

Proof. By naming I the incenter of ABC, from Van Obel’s theorem it follows that:

AI b+c
=
ILA a

and ILA ≥ r is trivial. By combining these observations:


 
1 1 1 CS
ALA + BLB + CLC ≥ r · (a + b + c) · + + ≥ 9r.
a b c

Exercise 197. Prove that if p(x) ∈ R[x] is a polynomial with non-negative coefficients,
for any couple (x, y) of positive real numbers we have:
√ p
p( xy) ≤ p(x)p(y).

With greater generality, the concept of convexity is a cornerstone in the theory of inequalities.

Page 105 / 222


Theorem 198 (Arithmetic-geometric inequality, AM-GM). If a1 , . . . , an are non-negative real numbers, we have:

def √ a1 + . . . + an def
GM(a1 , . . . , an ) = n
a1 · . . . · an ≤ = AM(a1 , . . . , an )
n
and equality holds iff a1 = . . . = an .

Proof. We may clearly assume that all the variables are positive without loss of generality.
In such a case, by setting bk = log ak , the arithmetic-geometric inequality turns out to be equivalent to
n
! n
1X 1X
exp bk ≤ exp (bk )
n n
k=1 k=1
2
d x
that is Jensen’s inequality for the exponential function, holding since dx 2e = ex > 0. An alternative and really
interesting approach is due to Cauchy. In his Cours d’analyse he observes the arithmetic-geometric inequality can be
proved through an “atypical induction”:

• AM-GM is trivial in the n = 2 case;

• if the AM-GM inequality holds for n variables, it also holds for 2n variables;

• if the AM-GM inequality holds for (n + 1) variables, it also holds for n variables.

Lemma 199 (Superadditivity of the geometric mean). If a1 , . . . , an are positive and distinct numbers,

∀τ > 0, GM(a1 + τ, . . . , an + τ ) > τ + GM(a1 , . . . , an ).

Proof. The given inequality is equivalent to:


n
Y √ n
(τ + ak ) > (τ + n
a1 · . . . · an )
k=1
ak
and by setting bk = τ and ck = log bk it is also equivalent to:
n
1X  c1 +...+cn

log(1 + eck ) > log 1 + e n ,
n
k=1

i.e. Jensen’s inequality for f (x) = log(1 + ex ). In order to prove the claim it is enough to notice that:
d2 x ex
log(1 + e ) = > 0.
dx2 (ex + 1)2
In a similar way we have that, if (a1 , . . . , an ) and (b1 , . . . , bn ) are n-uples of non-negative numbers,

GM(a1 + b1 , . . . , an + bn ) ≥ GM(a1 , . . . , an ) + GM(b1 , . . . , bn ).

Exercise 200 (Huygens inequality). Prove that for any θ ∈ [0, 1] we have:

tan θ + 2 sin θ ≥ 3θ.

Page 106 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

Proof.
Z θ   Z θ
1 AM-GM
tan θ + 2 sin θ = + cos u + cos u du ≥ 3 1 dθ = 3θ.
0 cos2 u 0

Exercise 201 (Weitzenbock’s inequality). ABC is a triangle with area ∆ and side lengths a, b, c.

Prove that the equality a2 + b2 + c2 = 4∆ 3 implies that ABC is an equilateral triangle.

Proof. Due to Heron’s formula


p
4∆ = (a2 + b2 + c2 )2 − 2(a4 + b4 + c4 ),

hence a2 + b2 + c2 = 4∆ 3 implies

(a2 + b2 + c2 )2 = (1 + 1 + 1)(a4 + b4 + c4 ),
CS
that is the equality case in (a2 + b2 + c2 )2 ≤ (1 + 1 + 1)(a4 + b4 + c4 ).
For other interesting proofs have a look at this thread on MSE.

Exercise 202. Prove that for any N ≥ 1 we have


N  
2 X
1
2
3
5 + log N ≥ n1/n − 1 ≥ 1
2 log2 N.
n=1

Proof. Since  
2 n AM-GM Hn−1
n1/n = GM 1, , . . . , < 1+
1 n−1 n
and
N (2)
X Hn−1 X 1 H 2 − HN
= = N
n=1
n kn 2
1≤k<n≤N

the inequality on the left follows from HN ≤ γ + log N and γ < 35 .


In order to prove the inequality on the right it is enough to show that for any N ≥ 1 we have:
1  2
1
log (N + 1) − log2 (N ) .

(N + 1) N +1 − 1 ≥ 2

To complete the proof is a task left to the reader. It might be useful to exploit:
Z N +1
 2 log x
1
log (N + 1) − log2 (N ) =

2 dx.
N x

Exercise 203. Prove that the sequence {an }n≥1 defined by


 n
1
an = 1 +
n

is increasing and bounded.

Page 107 / 222


Proof. We have an+1 > an by AM-GM:

    
n+1 1 1 AM-GM 1 1 1
1 · an = GM 1, 1 + , . . . , 1 + < 1+n· 1+ =1+ .
n n n+1 n n+1

Additionally:  n
a2n 1 1 1
= 1+ ≤ 1 =1+
an 4n(n + 1) 1− 4(n+1)
4n + 3
hence for any N ≥ 1 we have:
Y    1
1 16 Y 1 16 Y 1 + 2k+1 20
aN ≤ a1 1+ = 1+ ≤ 1 = .
4 · 2k + 3 7 k
4·2 +3 7 1+ 2k+2
7
k≥0 k≥1 k≥1

Exercise 204 (from Baby Rudin, 1). Let {an }n∈N be an increasing and unbounded sequence of positive real numbers.
P
Prove that there is a sequence {bn }n∈N of positive real numbers such that the series n∈N bn is convergent but the
P
series n∈N an bn is divergent.
P
Exercise 205 (from Baby Rudin, 2). Let {an }n∈N be a sequence of positive real numbers, such that the series n∈N an
is convergent. Prove the existence of a sequence {bn }n∈N of positive real numbers, increasing and unbounded, such
P
that the series n∈N an bn is convergent.

Proof. In the first exercise we may consider the sequence {bn }n∈N defined through:
1 1 1
b0 = √ , bn = √ −√ .
a0 an−1 an

The sequence {bn }n∈N is clearly increasing and we have:


N  
X 1 X 1 1 2 1
bn = √ + √ −√ =√ −√ .
a0 n=1 an−1 an a0 aN
n≤N

√2 .
P
Since the sequence {an }n∈N is unbounded, the previous identity proves the series n∈N bn is convergent to a0

We further have: √ √
an − an−1 √ √
an bn = an · √ ≥ an − an−1 .
an−1 an
PN √ √ P
The last identity proves n=1 an bn ≥ aN − a0 , hence the series n∈N an bn is divergent.
P
About the second exercise, we may assume n∈N an = 1 without loss of generality, by multiplying every element of
{an }n∈N by a suitable constant. Under such assumption, we may consider the sequence {bn }n∈N defined through:
n−1
1 X
b0 = c0 = 1, bn = √ , cn = 1 − am .
cn m=0
P
Since n∈N an is convergent to 1 and has positive terms, {cn }n∈N is a sequence with positive terms decreasing to 0,
hence {bn }n∈N is positive and unbounded. We also have an = cn − cn+1 , from which:
√ √  √ √ 
cn − cn+1 cn − cn+1 · cn + cn+1 √ √ 
an bn = √ = √ ≤ 2 cn − cn+1 .
cn cn

It follows that: X √
X √  √ 
an bn ≤ a0 + 2 · cn − cn+1 = a0 + 2 1 − cN +1 ,
n≤N n≤N

Page 108 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

P
hence the series n∈N an bn is convergent to some value ≤ (a0 + 2).
Both the exercises studied here follow from a more general inequality:
 −β !1−β
N N
X X 1 X
∀β ∈ (0, 1), ∀N ≥ 1, an  am  < an ,
n=1
β n=1
m≥n

P+∞
holding for any sequence {an }+∞
n=1 with positive terms such that n=1 an is convergent.

Exercise 206. {an }n≥1 is a sequence of real numbers with the following property: for any sequence of real numbers
PN
{bn }n≥1 such that n≥1 b2n is convergent, limN →+∞ n=1 an bn exists and it is finite. Prove that {an }n≥1 is square-
P
P 2
P 2 an
summable, i.e. n≥1 an is convergent. Hint: assume that n≥1 an is divergent and consider bn = An , where
AN = a1 + a2 + . . . + aN .

Exercise 207 (Hermite-Hadamard inequality). Prove that if f : [a, b] → R is a convex function,


  Z b
a+b 1 f (a) + f (b)
f ≤ f (x) dx ≤ .
2 b−a a 2

In other terms: for convex (or concave) functions it is not difficult to estimate the magnitude of the error
Rb
in computing a f (x) dx through the rectangle or trapezoid numerical methods.

Corollary 208. Since f (x) = log(x) is a concave function on R+ ,


Z n
n log n − n + 1 = log(x) dx ≥ f (1) + f (2) + . . . + f (n) = log(n!).
1

Corollary 209. For any n ∈ N+ ,


π2 1 π2 1
− 1 ≤ Hn(2) ≤ − .
6 n+ 2
6 n+1

log b−log a
Exercise 210. Given b > a > 0, provide an accurate lower bound for b−a in terms of a rational function.

Proof. As a rule of thumb, when dealing with objects like f (b) − f (a) it always is the case to wonder if such difference
has a nice/practical integral representation. That is certainly the case here:
b b−a 1
log b − log a
Z Z Z
def 1 dx 1 dx du
δ(a, b) = = = =
b−a b−a a x b−a 0 x+a 0 (1 − u)a + ub

where by symmetry (i.e. by enforcing the substitution u → 1 − u):


Z 1
a+b du
δ(a, b) = .
2 0 ((1 − u)a + ub) (ua + (1 − u)b)

This “folding trick ” leads to an integrand function that is convex, almost-constant on [0, 1] and symmetric with respect
to u = 21 . By applying the same trick a second time we get:
Z 1
2(a + b)
δ(a, b) = dt,
0 (a + b)2 − (b − a)2 t2

Page 109 / 222


2(a+b)
hence by denoting ga,b (t) = (a+b)2 −(b−a)2 t2 and by exploiting the Hermite-Hadamard inequality we get:

1
 8(a + b)
δ(a, b) ≥ ga,b 2 = .
(a + 3b)(3a + b)
24
As a straightforward corollary, log(2) ≥ 35 holds, and such inequality is pretty tight.

Theorem 211 (Karamata, Hardy-Littlewood).


Let (a1 , . . . , ak ) and (b1 , . . . , bk ) be sequences of real numbers with the following properties:

• ∀i < j, ai ≥ aj and bi ≥ bj (weakly decreasing);


def Pi Pi def
• ∀i ∈ [1, k], Ai = j=1 aj ≥ j=1 bj = Bi (the first sequence majorizes the second one);
Pk
• j=1 (aj − bj ) = 0 (same sum).

Karamata’s inequality, also known as Hardy-Littlewood’s inequality in its integral form,


states that the previous assumptions grant that for every real convex function f :
k
X k
X
f (ai ) ≥ f (bi ).
i=1 i=1

Proof. If f is convex, the function


f (b) − f (a)
δf (a, b) =
b−a
is symmetric in its arguments and increasing with respect to each argument.
In the given hypothesis we may define
ci = δf (ai , bi ),
then notice that:
k
X k
X k
X k−1
X
(f (ai ) − f (bi )) = ci (ai − bi ) = ci (Ai − Ai−1 − Bi + Bi−1 ) = (ci − ci+1 )(Ai − Bi ),
i=1 i=1 i=1 i=1
so
ci = δf (ai , bi ) ≥ δf (bi , ai+1 ) ≥ δf (ai+1 , bi+1 ) = ci+1
and the claim is proved.

1 1
Theorem 212 (Young’s inequality). If a and b are positive real numbers, p > 1 and p + q = 1, we have:

ap bq
ab ≤ + ,
p q
where equality holds iff ap = bq .

Proof. Due to the concavity of the function f (x) = log x we have:


 
1 1 1 p 1 q
log(ab) = log(ap ) + log(bq ) ≤ log a + b ,
p q p q
then the claim follows by exponentiation. We may notice that, if ap 6= bq , the last inequality is tight.
As an alternative, one might consider that on R+ the function f (x) = xp−1 has the inverse function g(x) = xq−1 .
By a well-known theorem on the integration of inverse functions (or simply by integration by parts) it follows that:
Z a Z b
ap bq
+ = f (x) dx + g(x) dx ≥ ab.
p q 0 0

Page 110 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

Hölder’s inequality can be obtained from Young’s inequality through an amplification trick:

1
Theorem 213 (Hölder’s inequality). If x1 , . . . , xn and y1 , . . . , yn are non-negative real numbers, p > 1 and p + 1q = 1,
we have: !1/p !1/q
Xn Xn Xn
p q
xi yi ≤ xi · yi ,
i=1 i=1 i=1

where equality holds if and only if for any k ∈ [1, n] we have xk = λyk .

Proof. By setting
n
!1/p
X
kxkp = xpi ,
i=1

Hölder’s inequality immediately follows from Young’s inequality, since:


Pn n
i=1 xi yi X xi yi 1 kxkpp 1 kykqq 1 1
= ≤ p + q = + = 1.
kxkp · kykq i=1
kxkp kykq p kxkp q kykq p q

1 1 1
Morever, it is not difficult to prove the following generalization: if p, q, r are positive real numbers and p + q + r = 1,
then:
! p1 ! q1 ! r1
X X X X
p q r
|xi yi zi | ≤ |xi | · |xi | · |xi | .
i i i i

Theorem 214 (Minkowski’s inequality, triangle inequality for the Lp norm). If x1 , . . . , xn and y1 , . . . , yn are non-
negative real numbers and p > 1, by setting
n
!1/p
X p
kxkp = xi
i=1

we have:
kx + ykp ≤ kxkp + kykp .

Proof.
kx + ykpp = k(x + y)p k1 ≤ kx(x + y)p−1 k1 + ky(x + y)p−1 k1 ,

and due to Hölder’s inequality we have:

kx(x + y)p−1 k1 ≤ kxkp · k(x + y)p−1 kq = kxkp · kx + ykp−1


p ,

from which:
kx + ykpp ≤ (kxkp + kykp ) kx + ykpp−1 .

Page 111 / 222


Theorem 215 (Carleman’s inequality). If {an }+∞
n=1 is a sequence of positive real numbers and the serie

+∞
X
an
n=1

is convergent to some C ∈ R, we have:


+∞ Y
n
!1/n
X
ai ≤ e C.
n=1 i=1

Proof. By denoting through GM the geometric mean, through AM the arithmetic mean :

GM(a1 , . . . , an ) = GM(a1 , 2a2 , . . . , nan )(n!)−1/n ≤ AM(a1 , 2a2 , . . . , nan )(n!)1/n ,

e
where by Stirling’s inequality we have (n!)1/n ≤ n+1 for any n ≥ 1, hence:
n
e X
GM(a1 , . . . , an ) ≤ kak .
n(n + 1)
k=1

It follows that:  
+∞ +∞ +∞
X X X 1 X
GM(a1 , . . . , an ) ≤ e   kak = e ak .
n=1
n(n + 1)
k=1 n≥k k=1

We may notice that such inequality always holds as a strict inequality. As a matter of fact, by assuming

GM(a1 , 2a2 , . . . , nan ) = AM(a1 , 2a2 , . . . , nan ),

D
we have ak = k for some constant D, but the harmonic series is divergent.

P
Exercise 216 (Indam test 2014, Exercise B3). Given n≥1 an , convergent series with positive terms, prove that
X n−1
(an ) n

n≥1

is also a convergent series.

n−1
Proof. We may notice that (an ) n ≤ 1+(n−1)a by AM-GM, but n≥1 n1 is a divergent series, so such argument does
n
P
n
not prove the claim directly. By tweaking it a bit:
 
n−1 1 3 n
an = GM
n
, 2an , an , . . . , an
n 2 n−1

leads, by AM-GM, to: !


n−1  
n−1 1 1 X k+1 1 log n
an n
≤ + an ≤ 2 + 1+ an
n n k n n
k=1

hence
π2
 
X n−1 1 X
ann
≤ + 1+ an .
6 e
n≥1 n≥1

An alternative approach is the following one: let α > 1 some real number and let
   
1 1
S = n : an ≤ n , L = n : an > n .
α α

Page 112 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

P an P an
Our purpose is to prove that both n∈S √n a
n
and n∈L √ n a
n
are convergent. By comparison with a geometric series,
P a P 1 α −1/n
n∈S n an ≤ n≥1 αn−1 = α−1 . On the other hand, if n ∈ L then an ≤ α, hence:
√n

X n−1 α X
(an ) n
≤ +α an
α−1
n≥1 n≥1

and by minimizing the RHS on α,


 2
sX
X n−1
(an ) n
≤ 1 + an  .
n≥1 n≥1

Theorem 217 (Hilbert’s inequality). If {am }m∈N and {bn }n∈N are two sequences in `2 (R) the following inequality
holds: v v
+∞ X +∞ u +∞ u +∞
X am bn uX
2
uX
≤π t am · t b2n .
n=1 m=1
n + m m=1 n=1

Proof. The proof we are going to present is essentially due to Schur, whose key idea is to prove the claim by a weighted
version of the Cauchy-Schwarz inequality. For any positive real number λ we have:
+∞ X+∞
!2 +∞ X +∞ +∞ +∞
X am bn X a2m  m 2λ X X b2n  n 2λ
≤ · .
m=1 n=1
n+m m=1 n=1
n+m n n=1 m=1
n+m m

We may notice that the first term in the RHS can be written as
+∞ +∞
X X 1  m 2λ
a2m ,
m=1 n=1
m+n n

so, by symmetry, it is enough to prove the existence of some λ > 0 such that, for any m ≥ 1:
+∞
X 1  m 2λ
≤ π.
n=1
m+n n

Since the terms of the series on the LHS are positive and decreasing, the following inequality holds:
+∞ Z +∞ Z +∞
X 1  m 2λ dx m2λ dy π
≤ 2λ
= 2λ
= ,
n=1
m+n n 0 m+x x 0 (1 + y) y sin(2πλ)

so it is enough to choose λ = 14 to finish. The π constant is optimal:


we leave to the reader to check that, if an = bn = n−1/2−ε ,
v v
+∞ X +∞ u +∞ u +∞
X am bn uX
2
uX
= (π − O(ε)) t am · t b2n .
n=1 m=1
n + m m=1 n=1

Exercise 218 (Hilber’s inequality, `p version).


Prove that if p > 1, p1 + 1q = 1, {am }m∈N ∈ `p (R), {bn }n∈N ∈ `q (R), then

+∞ X+∞

+∞
!1/p +∞
!1/q
X am bn π X X
≤ apm · bqn .
 
n+m


n=1 m=1
sin πp m=1 n=1

Page 113 / 222


Exercise 219 (A variant of Hilbert’s inequality).
Prove that if {an }n≥1 and {bn }n≥1 are two sequences in `2 (R+ ), then:

Γ 41 2
X X 
a b
m n
√ ≤ 2√π kak2 kbk2 .

m≥1 n≥1 m2 + n2

Theorem 220 (Hardy’s inequality). Let p > 1 and let a1 , . . . , aN be positive real numbers.
Pk
By setting Ak = i=1 ai , we have:
N  p  p XN
X An p
≤ apn .
n=1
n p − 1 n=1

In integral form: for every non-negative function in Lp (R+ ),


Z +∞  Z x p  p Z +∞
1 p
f (u) du dx ≤ f (x)p dx.
0 x 0 p − 1 0

Proof. The proof we are going to present is essentially due to Eliott.


We may firstly prove:
N N
X Apn p X an Ap−1 n
(♠) p
≤ p−1
.
n=1
n p − 1 n=1
n
An
Let Bn = n and let ∆n be the difference between the n-th terms in the RHS and LHS of (♠). We have:

def p p
∆n = Bnp − an Bnp−1 = Bnp − (nBn − (n − 1)Bn−1 ) Bnp−1 ,
p−1 p−1
or:  
np p(n − 1)
∆n = Bnp 1 − + Bn−1 Bnp .
p−1 p−1
As a consequence of Young’s inequality (212) we have:
p
Bn−1 Bp
Bn−1 Bnp−1 ≤ + (p − 1) n ,
p p

from which it follows that:


n−1 p n
∆n ≤ Bn−1 − Bp ,
p−1 p−1 n
and by creative telescoping we may state:
N p
X N BN
∆n ≤ − < 0,
n=1
p−1

proving (♠). By applying Hölder’s inequality (213) to the RHS of (♠) we have:

N N N
!1/p N
!(p−1)/p
X Apn p X an Ap−1n p X X Apn
p
≤ p−1
≤ apn ,
n=1
n p − 1 n=1
n p − 1 n=1 n=1
np

immediately proving the claim.

Page 114 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

P+∞
Exercise 221. Let {pn }+∞
n=1 be a sequence of positive real numbers such that the series
1
n=1 pn is convergent.
Prove that such assumptions grant that the series
+∞
X n2
pn
n=1
(p1 + . . . + pn )2

is convergent as well.

Proof. For the sake of brevity, let us set:


+∞ N N N
X 1 X X n2 X n2 p n
C2 = , PN = pi , SN = pn 2
= .
p
n=1 n i=1 n=1
(p1 + . . . + pn ) n=1
Pn2

Since {PN }+∞


N =1 is an increasing sequence, we have:

N
X n2 (Pn − Pn−1 ) N
X n2 (Pn − Pn−1 ) N
n2 n2
 
1 1 1 X
SN = + < + = + − ,
p1 n=2 Pn2 p1 n=2 Pn Pn−1 p1 n=2 Pn−1 Pn

−1
N
! N
5 X 2n + 1 N2 X n
SN < + − <5 .
P1 n=2
Pn PN P
n=1 n
By exploiting the Cauchy-Schwarz inequality we also have:
v v
N uN uN
X n2 pn
X n u X 1 u p
≤ t ·t 2 ≤ C Sn ,
P
n=1 n
p
n=1 n n=1
PN

from which it follows that:


p
SN < 5C SN ,
or:
SN < 25 C 2 .
Since such inequality holds for any N , and since the sequence {SN }+∞
N =1 is increasing, the following series is convergent
by the monotone convergence Theorem:
+∞ 2
X n pn
.
n=1
Pn2
It is not difficult to prove that we actually have the sharper inequality
N N
X 2n + 1 X 1
<4 ,
n=1
Pn p
n=1 n

from which we may derive the more accurate bound:


2
SN < + 4C 2 .
a1

P+∞
Exercise 222. Prove that if n=1 a1n is a convergent series with positive real terms,
there exists a constant C ∈ R such that:
+∞ +∞
X n X 1
≤C .
a + . . . + an
n=1 1
a
n=1 n

Page 115 / 222


Proof. Due to the AM-GM inequality we have that:
+∞ +∞ +∞  
X n X 1 X 1 1
≤ = GM ,..., ,
a + . . . + an
n=1 1 n=1
GM(a1 , . . . , an ) n=1 a1 an

where the RHS, by Carleman’s inequality (215), is bounded by:


+∞
X 1
e ,
a
n=1 n

so the given inequality holds for sure by taking C = e. In the next exercise we will see that such result can be
improved: the given inequality holds for C = 2, that is the optimal constant.

Exercise 223. Prove that if a1 , . . . , an are positive real numbers,


n n
X 2k + 1 X 1
<4 .
a1 + a2 + . . . + ak ak
k=1 k=1

Proof. We recall that, by exploiting the Cauchy-Schwarz inequality in the form of Titu’s Lemma,

Lemma 224. If r1 , r2 , α, β, γ are positive real numbers and γ = α + β, we have:


γ2 α2 β2
≤ + .
r1 + r2 r1 r2
This Lemma implies:
(n + 1)2 2n − 1 4 2n − 1 + (n − 1)2 4 n2
+ ≤ + = + ,
a1 + . . . + an a1 + . . . + an−1 an a1 + . . . + an−1 an a1 + . . . + an−1
from which it follows that:
n n−1
n2 X 2k + 1 4 (n − 1)2 X 2k + 1
+ ≤ + + ,
a1 + . . . + an a1 + a2 + . . . + ak an a1 + . . . + an−1 a1 + a2 + . . . + ak
k=1 k=1

and by induction:
n n
! n
n2 X 2k + 1 X 4 1 3 X 4
+ ≤ + + = .
a1 + . . . + an a1 + a2 + . . . + ak ak a1 a1 ak
k=1 k=2 k=1

Exercise 225. Prove that for any p > 1 and for any a, b, α, β > 0 we have:
1/p 1/p 1/p
(α + β)p+1 αp+1 β p+1
  
≤ + .
ap + bp ap bp

Proof. If we set b/a = x, it is enough to show that the minimum of the function f : R+ → R+
defined by:
p+1 p+1
f (x) = α p (1 + xp )1/p + β p (1 + x−p )1/p
p+1
is exactly (α + β) p . In order to do that, it is enough to check that f 0 (x) vanishes only at
 1/p
β
x= .
α

Page 116 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

Theorem 226 (Knopp). For any real number p ≥ 1 there exists some constant Cp ∈ R+ such that
for any sequence a1 , . . . , aN of positive real numbers,
N  1/p N
X n X 1
< Cp .
n=1
ap1 + . . . + apn a
n=1 n

Proof. In a similar way to Exercise 223, we prove there is a positive and increasing function f : N0 → R+ such that:
!1/p N  1/p !1/p N −1  1/p
f (N ) X n Cp f (N − 1) X n
(♦) PN + ≤ + + ,
ap1 + . . . + apn ap1 + . . . + apn
p PN −1 p
n=1 an n=1
aN n=1 an n=1

granting that, by induction, we have:


!1/p N  1/p N
f (N ) X n 1 + f (1)1/p X Cp
+ ≤ + .
ap1 + . . . + apn
PN p
n=1 an n=1
a1 a
n=2 n

In order that (♦) implies the claim it is enough that f (1) ≤ (Cp − 1)p holds and:
!1/p !1/p !1/p
f (N ) N Cp f (N − 1)
∀N ≥ 2, PN p
+ PN ≤ + PN −1 p .
n=1 an n=1 apn an n=1 an

p/(p+1)
By exploiting the inequalities proved in 225, by assuming f (N )1/p + N 1/p ≥ Cp we have:
 p
 p+1 p  p+1
p

f (N )1/p
+N 1/p
Cp f (N )1/p + N 1/p − Cpp+1
P 1/p ≤ a + P 1/p ,
N p N N −1
n=1 an n=1 apn

hence it is enough to find some function f such that:


 p
 p+1 p
1
f (N )1/p + N 1/p ≤ f (N − 1) p+1 + Cpp+1 .
1
Now we may consider Cp = (1 + p) p , that is the best constant we may put in the RHS of the initial inequality, if
an = n. Then we consider f (N ) = k · N p+1 : the previous inequality becomes:
 p
 p+1
1 1 1 1
(♠) k p+1 N 1+ ≤ k p+1 (N − 1) + (1 + p) p+1 .
N k 1/p
Due to Bernoulli’s inequality we have:
 p
 p+1
1 p
1+ ≤1+ ,
N k 1/p N (p + 1)k 1/p
so if we have some k such that:
p 1 1
p
1
(♥) k − p(p+1) + k p+1 ≤ Cpp+1 = (p + 1) p+1
p+1
the inequality (♠) is fulfilled. By studying the stationary points of the function g(x) = Ax−α + xβ
it is simple to derive that the choice
k = (p + 1)−p
N p+1
leads to an equality in (♥). It just remains to prove that with the choice f (N ) = (p+1)p
1
we have f (1) ≤ (Cp − 1)p , or Cp ≥ 1 + p+1 , i.e.:

1 1
(p + 1) p ≥ 1 + .
p+1

Page 117 / 222


By multiplying both sides by (p + 1) we get that such inequality is equivalent to:

p+1
(p + 1) p ≥ p + 2,

again a consequence of Bernoulli’s inequality, since:

p+1 p+1
(p + 1) p ≥1+ · p = p + 2.
p

Summarizing, we proved that for any p ≥ 1 and for any sequence a1 , . . . , aN


of positive real numbers we have:

p+1 N  1/p N
N p X n 1 X 1
+ ≤ (1 + p) p ,
1/p
(p + 1) (ap1 + . . . + apN ) n=1
ap1 + . . . + apn a
n=1 n

that essentially is a generalization of Hardy’s inequality (220) to negative exponents. We leave to the reader to prove
that the shown Cp constant is optimal, since by assuming an = n and letting N → +∞, it is clear it cannot be
replaced by any smaller real number.

Theorem 227 (Brunn-Minkowski). If A and B are two compact subsets of Rn and µ is the n-dimensional Lebesgue
measure,
1 1 1
µ(A + B) n ≥ µ(A) n + µ(B) n .

If both A and B are given by cartesian products of closed intervals (let us say boxes) the given inequality is trivial
by AM-GM. By an ingenious trick known as Hadwiger-Ohmann’s cut it is possible to show it continues to hold for
disjoint unions of boxes, hence the claim follows from the regularity of the Lebesgue measure.
Equality is achieved only if A and B are homothetic shapes, i.e. can be brought one into the other by the composition
of a uniform dilation and a translation. This inequality is extremely powerful and it can be employed to prove the
isoperimetric inequality in n dimensions.

Theorem 228 (Isoperimetric inequality in the plane). If γ is a regular, closed and simple curve, we have:

4πA ≤ L2

where L is the length of γ and A is the area enclosed by γ. Equality holds if and only if γ is a circle.

For a start, we show the most classical and elementary approaches.


A key observation is that if a simple, regular and closed curve with a
given length encloses the maximum area, it has to be convex.
Otherwise we might apply a reflection to an arc of such curve, increas-
ing the enclosed area without affecting the length.

Page 118 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

On the other hand, given a closed and convex curve, we may consider
an arbitrary chord and apply a reflection with respect to the perpen-
dicular bisector of such chord to one of the arcs cut. Through such
transform, both the perimeter and the enclosed area stay unchanged.
In particular any curve with a given length that is regular, simple,
closed and maximizes the enclosed area is mapped by any of such
transforms into a convex curve: otherwise it would be possible to in-
crease the enclosed area without affecting the length, as depicted on
the right.

Given these considerations it is not difficult to show that the circle is the only solution of the isoperimetric prob-
lem in the plane. As an alternative, one may follow a discretization approach.

Lemma 229. Among all the simple and closed polygonal lines P1 P2 . . . Pn (Pn+1 = P1 ) having sides li = Pi Pi+1 with
fixed lengths, the polygonal line enclosing the greatest area is such that P1 , . . . , Pn are vertices of a cylic polygon.

We may prove the claim by induction on n, by considering first the case n = 4.


That case can be tackled through Ptolemy’s inequality.

Theorem 230 (Ptolemy). If A, B, C, D (in this ordering) are the vertices of a convex quadrilateral in the plane, we
have:
AB · CD + BC · DA ≥ AC · BD

and equality holds if and only if ABCD is a cyclic quadrilateral.

Proof. It is enough to apply a circle inversion with respect to a unit circle centered at A, then consider how distances
change under circle inversion.

Let us consider the configuration on the side and set p = p1 + p2 and q = q1 + q2 .


We have:  2

 a = p21 + q12 − 2p1 q1 cos(π − θ)
 b2 = p2 + q 2 − 2p q cos(θ)

1 2 1 2
2 2 2


 c = p2 + q2 − 2p2 q2 cos(π − θ)
 2 2 2
d = p2 + q1 − 2p2 q1 cos(θ)
from which it follows that:

(a2 − b2 + c2 − d2 ) = 2(p1 + q1 )(p2 + q2 ) cos θ = 2pq cos θ.

By denoting as A the area of the depicted quadrilateral we have:

16A2 = (2pq sin θ)2


= (2pq)2 − (2pq cos θ)2
≤ (2ac + 2bd)2 − (a2 − b2 + c2 − d2 )2
= (a + b + c − d)(a + b − c + d)(a − b + c + d)(−a + b + c + d)

and ≤ holds as an equality iff the previous quadrilateral is cyclic. Moreover, given some closed and simple polygonal
line P1 , P2 , . . . , Pn (Pn+1 = P1 ) having sides with fixed lengths, it is always possible to rearrange its vertices in such a

Page 119 / 222


way they lie on the same circle. Indeed we may place Q1 , Q2 , . . . , Qn , Qn+1 on a circle with a huge radius, in such a way
that Q1 Q2 = P1 P2 , . . . , Qn Qn+1 = Pn Pn+1 hold, then slowly decrease the radius of such circle, letting Q1 , . . . , Qn+1
“slide” on it while preserving the mutual distances. By continuity, there exists some radius such that Qn+1 ≡ Q1 ,
leading to a cyclic rearrangement of the polygonal line.
In particular, if P1 P2 . . . Pn is a closed and simple polygonal line with n ≥ 4 vertices, enclosing the largest possible
area, all the quadrilaterals Pk Pk+1 Pk+2 Pk+3 are cyclic. Otherwise by leaving Pk and Pk+3 where they are and by
rearranging Pk+1 and Pk+2 we would increase the enclosed area. This proves Lemma (229).

Since every regular curve is uniformly approximated by a polyg-


onal line, and since all the solutions to the isoperimetric problem
in the plane (i.e. the curves of fixed length enclosing the largest
possible area) are regular functions, the isoperimetric inequality
follows from Lemma (229) “by sending n towards +∞”. How-
ever the Brunn-Minkowski inequality provides a less involved
approach.

Lemma 231. If P1 , P2 , . . . , Pn are vertices of a convex polygon


K having perimeter L and area A, and Br is a circle with radius
r, the area of K + Br is given by A + Lr + πr2 .

Cyclic rearrangement of 4 vertices:


the enclosed area increases.

In order to prove the Lemma it is enough to decompose K + Br as


the union of K, n rectangles with height r and bases on the sides of
K, n circle sectors corresponding to a partition of Br .

By approximating regular curves through polygonal lines, if K is the


region enclosed10 by a regular, simple and closed curve with length L,
we have:
µ(K + Br ) = µ(K) + rL + πr2 .

where µ is the 2-dimensional Lebesgue measure.


Due to Minkowski’s inequality with n = 2:
p p √
µ(K) + rL + πr2 ≥ µ(K) + πr2 ,

inequality which is equivalent to L2 ≥ 4π · µ(K), i.e. to the isoperimetric inequality.

Equality holds if and only if, up to translations, K = λBr , meaning that K is a circle with radius λr. The last
approach can be easily extended to the n > 2 case: if the surface area of the boundary is fixed, the suitable closed balls
with respect to the Euclidean norm are convex sets enclosing the maximum volume. The last proof of the isoperimetric
inequality we are going to see has a more analytic flavour, and it is deeply related to the Poincaré-Wirtinger
inequality.

10 We used the letter K twice on purpose. The solutions to the isoperimetric problem are given by convex sets and, since every regular
and convex curve is a limit of boundaries of convex polygons, in the isoperimetric problem the continuous and discrete approaches are
equivalent.

Page 120 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

Theorem 232 (Wirtinger’s inequality for functions). (Version I) If f ∈ C 1 (R) is a 2π-periodic function such that
R 2π
0
f (θ) dθ = 0, we have:
Z 2π Z 2π
2
f (θ) dθ ≤ f 0 (θ)2 dθ
0 0

where equality holds iff f (θ) = a sin θ + b cos θ.


(Version II) If f is a function of class C 1 on the interval [0, 2π] and f (0) = f (2π) = 0, we have:
Z 2π Z 2π
f (θ)2 dθ ≤ 4 f 0 (θ)2 dθ
0 0

θ
where equality holds iff f (θ) = κ sin 2.

Proof. By assuming the following identities


X X
f (θ) = sn sin(nθ) + cn cos(nθ),
n≥1 n≥1
X X
f 0 (θ) = nsn cos(nθ) − ncn sin(nθ)
n≥1 n≥1

hold in the L2 -sense, due to Parseval’s identity we get:


Z 2π X X Z 2π
f (θ)2 dθ = π (s2n + c2n ) ≤ π n2 (s2n + cn )2 = f 0 (θ)2 dθ
0 n≥1 n≥1 0

and ≤ holds as an equality iff c2 = c3 = . . . = s2 = s3 = . . . = 0. This proves the first version of Wirtinger’s inequality.
When proving the second version we may assume without loss of generality
X
f (θ) = sn sin(nθ/2)
n≥1
X
0 1
f (θ) = 2 nsn cos(nθ/2)
n≥1

(in the L2 -sense) and the claim follows again from Parseval’s identity.

We may now consider that any regular, simple and closed curve with length L has an arc length parametrization,
i.e. a couple of piecewise-C 1 , L-periodic functions x(s), y(s) such that
 2  2
dx dy
+ = 1.
ds ds
By introducing    
def Lθ def Lθ
f (θ) = x , g(θ) = y
2π 2π
R 2π
we have that the area enclosed by γ, as a consequence of Green’s Theorem, is given by the integral 0 f (θ) g 0 (θ) dθ.
By defining f as the mean value of f on the interval [0, 2π], and by noticing that the mean value of g 0 (θ) is zero, we
have:

Z 2π Z 2π
A= f (θ) g 0 (θ) dθ = (f (θ) − f ) g 0 (θ) dθ
0 0
Z 2π
1
(f (θ) − f )2 + g 0 (θ)2 dθ
 

2 0
Z 2π
Wirt 1
f 0 (θ)2 + g 0 (θ)2 dθ
 

2 0

L2 dθ L2
Z
1
= = .
2 0 4π 2 4π

Page 121 / 222


Equality holds if and only if f (θ)−f = a sin θ +b cos θ and g 0 (θ) = f (θ)−f , i.e. only when γ is a circle. It is interesting
to point out that the given proof can also be reversed, proving that Wirtinger’s inequality for functions (or, at least,
its first version) is a consequence of the isoperimetric inequality in the plane.

Exercise 233 (Dido’s problem). A rich landowner has bought 1 Km of metal fence. He wants to use such fence and
a wall of his huge home to enclose the largest possible area for his flock. What is the optimal shape he may choose
for the fence, and how large is the largest area he can dedicate to his flock?

Exercise 234 (Lhuilier’s inequality). Prove that if P is a convex polygon with n sides, having external angles
α1 , . . . , αn , we have:
n α 
i
X
L(∂P )2 ≥ 4 A(P ) tan
2
k=1

and equality holds if and only if P is circumscribed to a circle.

Exercise 235. Prove that if P is a convex polygon with n sides containing a unit circle, we have:
π
A(P ) ≥ n tan
n
and equality holds if and only if P is a regular polygon.

Exercise 236. Prove that if P is a convex polygon with n sides, we have


π
L(∂P )2 ≥ 4n tan A(P )
n
and equality holds if and only if P is a regular polygon.

Exercise 237 (Bonnesen’s inequality). Let γ be a regular, simple and closed curve enclosing a convex region P .
Let R, r denote the radii of the circumscribed and inscribed circle. Prove that the following strengthening of the
isoperimetric inequality holds:
L(γ)2 − 4π A(P ) ≥ π 2 (R − r)2 .

The isoperimetric inequality through Lagrange multipliers.


Among the convex polygons enclosing the origin, having sides of fixed lengths, the cyclic polygon encloses the
maximum area.

Proof. Setting θi = Pi+1


\ Pi O, Li = d(Pi , Pi+1 ) and φi = OP
\i+1 Pi , we want to find the maximum of

1 X sin θi sin φi 2
A= L
2 sin(θi + φi ) i
subject to the constraint: X
(θi + φi ) = (n − 2)π.

Page 122 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

We are in position of applying Lagrange multipliers, from which we have, for any i:
 2      2
sin φi ∂ sin θi sin φi λ ∂ sin θi sin φi sin θi
= = 2 = = .
sin(θi + φ1 ) ∂θi sin(θi + φi ) Li ∂φi sin(θi + φi ) sin(θi + φ1 )

It follows that all the triangles OPi Pi+1 are isosceles triangles with vertex at O, hence the cyclic polygon P1 . . . Pn
certainly has the greatest area.

Theorem 238 (Jung, isodiametric inequality).


Every compact set K ⊂ Rn of diameter d is contained in some closed ball of radius
r
n
R≤d
2(n + 1)

with equality attained only by regular n-simplex of side d.

Proof. Given a set of points of diameter d in Rn it is trivial to see that it can be covered by a ball of radius d.
But the above Theorem by Jung improves the result by a factor of about √12 , and is the best possible.
We first prove this Theorem for sets of points S with |S| ≤ n + 1 and then extend it to an arbitrary point set. If
|S| ≤ n + 1 then the smallest ball enclosing S exists. We assume that its center is the origin and denote its radius
by R. Denote by S 0 ⊆ S the subset of points such that kpk = R for p ∈ S 0 . It is easy to see that S 0 is in fact non empty.

Observation: The origin must lie in the convex hull of S 0 . Assuming the contrary, there is a separating hyperplane H
such that S 0 lies on one side and the origin lies on the other side of H (strictly). By assumption, every point in S \ S 0
has a distance strictly less than R from the origin. Move the center of the ball slightly from the origin, in a direction
perpendicular to the hyperplane H towards H such that the distances from the origin to every point in S \ S 0 remains
less than R. However, now the distance to every point of S 0 is decreased and so we will have a ball of radius strictly
less than R enclosing S which is a contradiction to the minimality of R.

Let S 0 = {p1 , p2 , . . . , pm } where m ≤ n ≤ d+1 and because the origin is in the convex hull of S 0 so we have non-negative
λi such that,
X X
λi pi = 0, λi = 1.

Fix a k, 1 ≤ k ≤ m. Then we have:


P
1 − λk = i6=k λi
1
Pm 2
≥ d2 P i=1 λi kpi − pk k
1 m 2

= d2 i=1 λi (2R − 2hpi , pk i)
1 2
Pm 
= d2 2R − 2 h i=1 λi pi , pk i
2R2
= d2

Adding up the above inequalities for all values of k, we get

2mR2
m−1≥
d2
2
Thus we get R m−1 n x−1
d2 ≤ 2m ≤ 2n+2 since m ≤ n + 1 and the function 2x is monotonic. So we have immediately
q
n
R ≤ d 2n+2 . The remainder of the proof uses the beautiful theorem of Helly. So assume S is any set of points of
q
n
diameter d. With each point as center draw a ball of radius R = d 2n+2 . Clearly any n + 1 of these balls intersect.
This is true because the center of the smallest ball enclosing n + 1 of the points is at most R away from each of those

Page 123 / 222


points. So we have a collection of compact convex sets, any n + 1 of which have a nonempty intersection. By Helly’s
theorem all of them have an intersection. Any point of this intersection can be chosen to be the center of a ball of
radius R that will enclose all of S.

Theorem 239 (Helly). Let {X1 , . . . , Xd } be a finite collection of convex subsets of Rd , with n > d. If the intersection
of every d + 1 of these sets is non-empty, then the whole collection has a nonempty intersection; that is,
n
\
Xj 6= ∅.
j=1

For infinite collections one has to assume compactness: Let {Xα } be a collection of compact convex subsets of Rd ,
such that every subcollection of cardinality at most d + 1 has a non-empty intersection, then the whole collection has
a non-empty intersection.

Proof. The proof is by mathematical induction:

Base case: Let n = d + 2. By our assumptions, for every j = 1, . . . , n there is a point xj that is in the common
intersection of all Xi with the possible exception of Xj . Now we apply Radon’s Theorem to the set A = {x1 , . . . , xn },
which furnishes us with disjoint subsets A1 , A2 of A such that the convex hull of A1 intersects the convex hull of A2 .
Suppose that p is a point in the intersection of these two convex hulls. We claim that
n
\
p∈ Xj .
j=1

Indeed, consider any j ∈ {1, . . . , n}. We shall prove that p ∈ Xj . Note that the only element of A that may not be in
Xj is xj . If xj ∈ A1 , then xj 6∈ A2 , and therefore Xj ⊃ A2 . Since Xj is convex, it then also contains the convex hull
of A2 and therefore also p ∈ Xj . Likewise, if xj 6∈ A1 , then Xj ⊃ A1 , and by the same reasoning p ∈ Xj . Since p is in
every Xj , it must also be in the intersection.
Above, we have assumed that the points x1 , . . . , xn are all distinct. If this is not the case, say xi = xk for some i 6= k,
then xi is in every one of the sets Xj , and again we conclude that the intersection is nonempty. This completes the
proof in the case n = d + 2.

Inductive Step: Suppose n > d + 2 and that the statement is true for n − 1. The argument above shows that any
subcollection of d + 2 sets will have nonempty intersection. We may then consider the collection where we replace the
two sets Xn−1 and Xn with the single set Xn−1 ∩ Xn . In this new collection, every subcollection of d + 1 sets will have
nonempty intersection. The inductive hypothesis therefore applies, and shows that this new collection has nonempty
intersection. This implies the same for the original collection, and completes the proof.

Theorem 240 (Radon, 1921). Any set of d + 2 points in Rd can be partitioned into two disjoint sets whose convex
hulls intersect. A point in the intersection of these convex hulls is called a Radon point of the set.

Proof. Consider any set {x1 , x2 , . . . , xd+2 } ⊂ Rd of d + 2 points in a d-dimensional space. Then there exists a set of
multipliers a1 , a2 , . . . , ad+2 , not all of which are zero, solving the system of linear equations
d+2
X d+2
X
ai xi = 0, ai = 0,
i=1 i=1

because there are d + 2 unknowns (the multipliers) but only d + 1 equations that they must satisfy (one for each
coordinate of the points, together with a final equation requiring the sum of the multipliers to be zero). Fix some

Page 124 / 222


7 THE CAUCHY-SCHWARZ INEQUALITY AND BEYOND

particular nonzero solution a1 , a2 , . . . , ad+2 . Let I be the set of points with positive multipliers, and let J be the set
of points with multipliers that are negative or zero. Then I and J form the required partition of the points into two
subsets with intersecting convex hulls. The convex hulls of I and J must intersect, because they both contain the
point
X ai X −aj
p= xi = xj ,
A A
i∈I j∈J

where X X
A= ai = − aj .
i∈I j∈J

The left hand side of the formula for p expresses this point as a convex combination of the points in I, and the
right hand side expresses it as a convex combination of the points in J. Therefore, p belongs to both convex hulls,
completing the proof.

Van Der Corput’s trick for lower bounds. The purpose of this paragraph is to prove that the partial sums
of the sequence {sin(n2 )}n≥1 are not bounded. We have:

n
!2 n n
X X X
2 sin(k 2 ) = cos(j 2 − k 2 ) − cos(j 2 + k 2 )
k=1 j,k=1 j,k=1
2 2
nX −1 2n
X
= n+2 d1 (m) cos(m) − 2 d2 (m) cos m
m=1 m=2

where d1 (m) accounts for the number of ways to write m as j 2 − k 2 with 1 ≤ k < j ≤ n and d2 (m) accounts for
the number of ways to write m as j 2 + k 2 with 1 ≤ j, k ≤ n. Since both these arithmetic functions do not deviate
much from their average order (by Dirichlet’s hyperbola method d1 (m) behaves on average like log m and d2 (m)
behaves on average like π4 ), It is not terribly difficult to prove that for infinitely many ns

n
X √
sin(k 2 ) ≥ C n



k=1

holds for some absolute constant C ≈ √12 through summation by parts and the Cauchy-Schwarz inequality.
A detailed exposition on Dirichlet’s hyperbola method can be found on Terence Tao’s blog.
We recall that Van Der Corput’s trick is usually employed to produce upper bounds: for instance
n
X √
sin(k 2 ) ≤ D n log n,



k=1

for some absolute constant D > 0, holds for any n large enough.
2
In particular n≥1 sin(n )
is convergent for any α > 12 .
P

Page 125 / 222


8 Bessel functions and the Gauss circle problem
Bessel functions naturally arise in the solution of the problem ∆u = f with certain boundary conditions, and they are
extremely relevant in Harmonic Analysis. We may introduce them by studying the Fourier sine series of the arcsin
function.
Z 1 Z π/2
arcsin(x) sin(πnx) dx = Im eiπn sin z z cos(z) dz
−1 −π/2
X (−1)m (πn)2m+1 Z π/2 2m+1
= z cos(z) (sin z) dz
(2m + 1)! −π/2
m≥0
!
IBP
X (−1)m (πn)2m+1  z sin(z)2m+2 π/2 Z π/2
2m+2
= − (sin z) dz
(2m + 1)! 2m + 2 −π/2 −π/2
m≥0
" 2m+2
#
X (−1)m (πn)2m+1 π m+1
= · 1 − m+1
(2m + 1)! 2m + 2 4
m≥0

1 − cos(nπ) X (−1)m (πn)2m+1


= −π
n 4m+1 (m + 1)!2
m≥0
 
m 2m
1 X (−1) (πn) 
= (−1)n+1 +
n 4m m!2
m≥0

hence by defining the following even, entire function

X (−1)n z 2n
J0 (z) =
4n n!2
n≥0

we have
L2 (−1,1) X (−1)m+1 + J0 (πm)
arcsin(z) = sin(πmz),
m
m≥1

π L2 (−1,1) X J0 (πm) sin(πmz)


arcsin(z) − z = .
2 m
m≥1

It is straightfoward to check from the power series that J0 is the unique solution of the following differential equation
of order 2:
z f 00 (z) + f 0 (z) + z f (z) = 0, f (0) = 1, f 0 (0) = 0

which allows to state many interesting facts about the behaviour of J0 far from the origin.
From the power series definition it also follows that

1
(LJ0 )(s) = √
1 + s2

and we have already seen that Vandermonde’s identity leads to


 
2 2 4 1
(LJ0 )(s) = K − 2 = √
πs s AGM(s, s2 + 4)

where the RHS behaves like − log(s)


π in a right neighbourhood of the origin and like 1s in a left neighbourhood of +∞.
Anyway the properties of J0 are better understood by introducing the Bessel functions of the first kind Jn and
their generating function: for such a purpose, we study the Fourier cosine series of eiz cos θ , which will lead to the
Jacobi-Anger expansion.

Page 126 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

Z 2π X in z n Z 2π
eiz cos θ cos(mθ) dθ = (cos θ)n cos(mθ) dθ
0 n! 0
n≥0
X in z n Z 2π
= (eiθ + e−iθ )n cos(mθ) dθ
2n n! 0
n≥0
X πin z n m + 2k 
= π
2n n! k
n=m+2k
X (−1)k z 2k m + 2k 
m m
= πi (z/2)
4k k!(m + k)! k
k≥0

hence by defining
X (−1)n (z/2)2n+m
Jm (z) =
n!(n + m)!
n≥0

we have

Theorem 241 (Jacobi-Anger expansion).


X
eiz cos θ = J0 (z) + 2 im Jm (z) cos(mθ).
m≥1

By considering the real or imaginary part of both sides, we get that Bessel functions of the first kind provide the
coefficients of the Fourier series of sin(z cos θ) and cos(z cos θ). By applying Cauchy’s integral formula to the Jacobi-
Anger expansion we find the following integral representation:

1 π
Z
Jn (z) = cos(z sin θ − nθ) dθ
π 0

with the following equivalent form:


I   
1 z 1 dt
Jn (z) = exp t− n+1
.
2πi kzk=ε 2 t t

Bessel functions share with the families of orthogonal polynomials many interesting properties. The recurrence relations

1
J00 (z) = −J1 (z), Jn0 (z) = (Jn−1 (z) − Jn+1 (z)) ,
2

2n d
Jn−1 (z) + Jn+1 (z) =
Jn (z), (xm Jm (x)) = xm Jm−1 (x)
z dx
are straightforward to prove through the series definition or the integral representation. In the same way the entire
function Jn (z) can be checked to be a solution of Bessel’s differential equation

x2 f 00 (x) + xf 0 (x) + (x2 − n2 )f (x) = 0

whose structure recalls the structure of Legendre’s differential equation. The identities

X X n
X X
1 = J0 (x)2 + 2 Jk (x)2 , 1 = J0 (x) + 2 J2k (x), Jn (2z) = Jk (z)Jn−k (z) + 2 (−1)k Jk (z)Jn+k (z)
k≥1 k≥1 k=0 k≥1

can be proved through the orthogonality relations in L2 (−π, π) applied to the Jacobi-Anger expansion. The integral
representation or Bessel’s differential equation allow to define Jν (x) also for non-integer values of ν. For n ∈ N, the

Page 127 / 222


function J−n (x) is defined as (−1)n Jn (x). According to this convention, Bessel functions of the first kind have a very
simple addition formula of the convolution-type:
X
Jn (y + z) = Jm (y)Jn−m (z).
m∈Z

The substitution g(x) = xf (x) turns the differential equation defining J0 , namely xf 00 (x) + f 0 (x) + xf (x) = 0, into:
 
1
g 00 (x) + 1 + 2 g(x) = 0.
4x

Since the term 4x1 2 is non-negative and negligible for large values of x, it is reasonable to assume that g(x) =
A(x) cos(x) + B(x) sin(x) is such that A(x) → A and B(x) → B for x → +∞. In order to find the value of these
constants, we may notice that
√ Z √2x r
0 2 2 2 t2
cos(x)J0 (x) − sin(x)J0 (x) = √ cos(t ) 1 − dt,
π x 0 2x
√ Z √2x r
0 2 2 2 t2
sin(x)J0 (x) + cos(x)J0 (x) = √ sin(t ) 1 − dt,
π x 0 2x
hence by the dominated convergence theorem and Fresnel integrals we have A = B = √1π .
d2 1

By studying the action of the operator dx 2 + 1 + 4x2 on the conjectural expression

√ A0 A00 B0 B 00
   
πx J0 (x) = A + + 2 + . . . cos(x) + B + + 2 + . . . sin(x)
x x x x
Poisson derived the formal series
   
1 9 9 · 25 1 9 9 · 25
1− − + + . . . cos(x) + 1 + − − + . . . sin(x)
8x 2 · 82 x2 2 · 3 · 8 3 x3 8x 2 · 82 x2 2 · 3 · 83 x3

which is not convergent, but whose truncations allow to devise arbitrarily accurate approximations of J0 (x) as x → +∞.
1
As a side note, we may notice that over the real line the Laplace transform of J0 (x), namely √1+s 2
, and the Laplace
sin(x)+cos(x)
q
1 1

transform of √
πx
, namely 1+s2 + √1+s2 , differ by a term bounded between 0 and 2 − 1 and behaving like
√1 for |s| → +∞.
|s|

Theorem 242 (RH for dummies). All the zeroes of the entire functions Jn (z), Jn0 (z) in C \ {0} are real and simple.
The average distance between a zero and the next one approaches π.

Proof. Assuming that for some z ∈ C we have Jn (z) = 0 and Jn0 (z) = 0, Bessel’s differential equation implies Jn00 (z) = 0,
(m)
then Jn (z) = 0 by induction, hence Jn (z) = 0 by the principle of analytic continuation, which is a contradiction.
Since the coefficients of the Maclaurin series of Jn (z) are real, assuming Jn (z0 ) = 0 with z0 6∈ R implies that z0 is a
zero, too. The same holds for Jn0 . Now we may consider the identity
Z z
2 2
(a − b ) tJn (at)Jn (bt) dt = z [bJn (az)Jn0 (bz) − aJn0 (az)Jn (bz)]
0

which follows from Bessel’s differential equation. By considering a = z0 , b = z0 and z = 1 we get


Z 1
2
0 = (z02 − z0 2 ) t |Jn (z0 t)| dt
0

which can only be true for z0 = iy with y ∈ R \ {0}. In such a case, however,
X 1  y 2m
m!(m + n)! 2
m≥0

Page 128 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

is clearly positive. We may invoke the Gauss-Lucas theorem (the zeroes of f 0 (z) lies in the convex hull of the zeroes
of f (z)) to deduce that all the zeroes of Jn0 (z) in C \ {0} are real and simple. The asymptotic formula
r
2  πn π  π
Jn (z) ∼ cos z − − for |z| → +∞ with |arg z| ≤ −ε
πz 2 4 2

finishes the proof.

Exercise 243. Show that J1/2 (x) is an elementary function.

Proof. We just have to notice that the Bessel differential equation


 
1
x2 y 00 + xy 0 + x2 − y = 0,
4

for which y = J 21 is a solution, is mapped into a well-known homogeneous differential equation with constant coefficients
f (x)
by the substitution y(x) = √ .
x

R +∞  J1 (x) 2
Exercise 244 (Exploiting the Fourier transform). Compute the value of the integral 0 x dx.

Proof. Since J1 is a solution of the Bessel differential equation:

x2 f 00 + xf 0 + x2 f = f

by exploiting integration by parts we have that:


Z +∞ Z +∞  2
1 J1 (x)
2
J1 (x) J100 (x)

J1 (x) + dx = dx
0 2 0 x

J1 (x)
so we just need to recall that the Fourier transform of x is given by:
  r
J1 (x) 2p
F (t) = 1 − t2 · 1(−1,1) (t)
x π

to be able to state: Z +∞ Z 1
1 2
J1 (x)2 + J1 (x) J100 (x) dx = 1 − t2 dt =
 
0 π 0 3π
as a consequence of Parseval’s theorem.

Exploiting the Laplace transform.


Z +∞ Z +∞
J1 (x) 1 sα−1
dx = √ √ ds
0 xα Γ(α) 0 1 + s2 (s + 1 + s2 )
Z π/2 α−1 Z π/2 h
(tan θ) α−1 −1−α α −1−α
i
= dθ = (sin θ) (cos θ) − (sin θ) (cos θ) dθ
0 1 + sin θ 0
πα
= 2
1 + α2 sin πα

2α+1 Γ 2

holds for any α ∈ − 12 , 2 , as soon as the original integral is intended as an improper Riemann integral.


Similar manipulations driven by the Laplace transform of J1 (x)2 lead to



π(1 − α)Γ α2
Z +∞ 
J1 (x)2
dx = 3
xα 2(1 + α)Γ 1+α cos πα

0
2 2

Page 129 / 222


for any α ∈ (0, 1). We also have relations with Fourier-Legendre series and hypergeometric functions. For instance:

2 K − s42
Z +∞ Z +∞  Z +∞
J0 (x)2 ds
√ dx = 3/2
ds = √ √
0 x 0 (πs) 0 πs AGM(s, s2 + 4)
Z +∞ Z +∞
ds ds
= √ √ = q
2πs AGM(s, s + 1) 2 p
0 0 2πs(s2 + 1) AGM(1, 1 − s21+1 )
√    
Z +∞ 2 K 21 ds Z +∞ K 1 ds Z 1
s +1 s+1 1 K (s) ds
= p = p √ = √ 3/4
π 3 s(s2 + 1) 3
2π s(s + 1) s 2π 3
0 0 0 (s(1 − s))
2
Γ 41 X Γ n + 12 Γ n + 14 Γ 41
  
· 2 F1 14 , 12 ; 1; 1

= √ 2
= √
3
8π n≥0 Γ(n + 1) 8π 2

1 4
= Γ 14 .
4π 5/2
The last integrals are just special instances of the Sonine-Schafheitlin integral formula.
Once the modified Bessel functions of the first kind are defined through
X (x/2)2n+m 1
Im (x) = = m Jm (ix)
n!(n + m)! i
n≥0

the Laplace transform allows a simple evaluation of the integrals


Z +∞
def
I(α, β, m) = xα e−βx Im (x) dx
0

for β > 1 and α, m ∈ N. Indeed, from

1s>1 (s)  p m
L(Im (x))(s) = √ s − s2 − 1
s2 − 1
we have:
" √ m #
dα s − s2 − 1
I(α, β, m) = √
dsα s2 − 1 s=β
2
xm−1 dx
I
s7→ x 2x
+1
α!
= √ α+1
2πi x2 +1
2x − β
kx−(β− β 2 −1)k=ε
 p m+α
z + β − β 2−1
α!
= Res α+1
2α+1 z=0 α+1
 p
z z − 2 β2 − 1
 p m+α
α+1 x + β − β 2−1
α!(−1) α+2
= p α+1 · [x ]  α+1
4α+1 2
β −1 1 − √ z2
2 β −1

α+1 α+2
X  
α!(−1) m+α 2α + 2 − k p p k−1
= α+1
(β + β 2 − 1)m+α−k 2k−1 β 2 − 1 .
8 (β 2 − 1)α+1 k α
k=0

Besides this characterization of the inner products against the elements of Span(xα e−βx : α ∈ N, β > 0),
modified Bessel functions of the first kind have a simple integral representation:

Z π
1
In (α) = cos(nx) eα cos x dx.
π 0

Page 130 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

By the cosine addition formula and integration by parts we may easily get the following recurrence relation:
α
In = (In−1 − In+1 ) .
2n
From this identity, we have that the ratio between In and In−1 is the continued fraction:

In 1
(α) = 2 1 .
In−1 α n+ 2
(n+1)+...
α

In particular
I1 1 x
r(x) = (x) = 2 1 = x2
.
I0 x + 4 1 2+ x2
x+ 6+ 1 4+ 2
x ... 6+ x
8+...

We also have
d 1
In = (In−1 + In+1 )
dα 2
mimicking the recurrence relation for Jn0 .

Corollary 245.
1
P
I1 (2) m≥0 m!(m+1)!
[0; 1, 2, 3, 4, 5, 6, 7, 8, 9, . . .] = = P 1 ≈ 0.697774657964.
I0 (2) m≥0 m!2

In general, if A, D > 0 we have


2

IA/D
D 
[A + D; A + 2D, A + 3D, A + 4D, . . .] = 2 .
I1+A/D D

The asymptotic behaviour of I0 (x) for x → +∞ can be derived from Bessel’s differential equation:

ex
 
1 9 9 · 25
I0 (x) = √ 1+ + + + ...
2πx 1! · 8x 2! · 82 x2 3! · 83 x3

the identity holds in the Poisson sense. If combined with the continued fraction representation for Im+1
Im ,
it leads to the asymptotic behavior of Im (x) as x → +∞, for any m ∈ N. For a fixed x ∈ R+ , the continued fraction
representation immediately leads to the upper bound

(x/2)n
In (x) ≤ I0 (x).
n!
If one is just interested in the first term of the asymptotic expansion of I0 , it can be recovered in a very simple way:
for s → 1+ ,
 
1 1 1
(LI0 )(s) = √ ∼√ √ =L √ (s).
s2 − 1 2 s−1 ex 2πx

P+∞ (−1)m 2m
Relations with the sine and cosine integrals. Since J0 (x) = m=0 4m m!2 x we have:

π/2 +∞
(−1)m x2m π/2
Z X Z
J0 (x cos θ) cos θ dθ = m m!2
(cos θ)2m+1 dθ
0 m=0
4 0
+∞
X (−1)m x2m 4m m!2
= ·
m=0
4m m!2 (2m + 1)!
+∞
X (−1)m x2m sin x
= = .
m=0
(2m + 1)! x

Page 131 / 222


P+∞ (−1)m 2m+1
In a similar way J1 (x) = m=0 2·4m (m+1) m!2 x gives:

π/2 +∞ π/2
(−1)m x2m+1
Z X Z
J1 (x cos θ) dθ = (cos θ)2m+1 dθ
0 m=0
2 · 4m (m + 1) m!2 0
+∞
X (−1)m x2m+1 4m m!2
= ·
m=0
2 · 4m (m + 1) m!2 (2m + 1)!
+∞
X (−1)m x2m+1 1 − cos x
= = .
m=0
(2m + 2)! x

By exploiting Fubini’s theorem we get:


Z +∞ Z π/2 Z +∞ Z π/2
sin x cos θ π
dx = cos(θ) J0 (x cos θ) dx dθ = dθ =
0 x 0 0 0 cos θ 2

and similarly
+∞ π/2 +∞ π/2
1 − cos x
Z Z Z Z
J1 (x cos θ) π
dx = dx dθ = 1 dθ = .
0 x2 0 0 x 0 2

A curious series. We investigate about the convergence of the series


X X 1 Z 2π
J0 (n) = eni cos θ dθ.
2π 0
n≥0 n≥0

The partial sums of this series can be written in the following form:
N
sin N + 21 cos θ
Z 2π Z π   
X 1 1 − e(N +1)i cos θ 1 1
SN = J0 (n) = dθ = + dθ
1 − ei cos θ sin cos2 θ
 
n=0
2π 0 2 2π 0
1
by invoking the identities cos(θ + π) = − cos θ and + 1−e1−z = 1. Rearranging,
1−ez
Z 1
1 1 D (x)
SN = + √N dx
2 2π −1 1 − x2
1 (x)
where DN (x) is Dirichlet’s kernel. If the coefficients cn of the Fourier cosine series of (−1,1)

1−x2
were in `1 ,
we could immediately state
Z 1 Z 1
1 D (x) 1 dx 1 1
√N

(♥) dx → √ + · √ =1 as N → +∞.
2π −1 1 − x2 2π −1 1 − x2 2 1 − x2 x=0
Unluckily, our case is not the case: it can be shown that |cn | ≤ π28√n for any n ≥ 1, but the density of ns such

that 21 ≤ n|cn | ≤ 1 is positive, hence {cn }n≥1 6∈ `1 . On the other hand, in order to prove (♥) it is enough to
1 (x)
show that the Fourier series of (−1,1)

1−x2
is pointwise convergent at the origin. Since

X 2n Z 1
 2
1 − e−s X (2n)!
    
1(0,1) (x) n 2n −sx 1 1
L √ (s) = x e dx = + + w
1 − x2 4n 0 s 2n n! s2n+1 s
n≥0 n≥1

where w(z) is an analytic function fulfilling w(0) = w0 (0) = 0, we have


 
1(0,1) (x)
lim s · L √ (s) = 1
s→+∞ 1 − x2
and the pointwise convergent at the origin for the Fourier series of √ 1 is proved.
1−x2
As a straightforward corollary, we have
X 3
J0 (n) = .
2
n≥0

Page 132 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

In particular J0 is an even entire function fulfilling


Z +∞
X f (0)
f (n) = + + f (x) dx
2 0
n≥0

while sinc is an even entire function fulfilling


Z +∞
X f (0)
f (n) = − + f (x) dx.
2 0
n≥0

R +∞
The same argument applied to Jk (x) with k ∈ N+ leads to
P
n≥0 Jk (n) = 0
Jk (x) dx = 1.

Series for fixed argument. From the Jacobi-Anger expansion:


+∞
X
eiz cos θ = J0 (z) + 2 in Jn (z) cos(nθ)
n=1

we have, by considering the imaginary part:


+∞
X
sin(z cos θ) = 2 (−1)m J2m+1 (z) cos((2m + 1)θ)
m=0

and we can remove the cosine-dependent term by exploiting the identities:


Z π/2
π
cos((2n + 1)x) cos((2m + 1)x)dx = δm,n ,
0 4
+∞
X 1
(−1)m cos((2m + 1)θ) = ,
m=0
2 cos θ

that give (integrating against the proper kernel):


Z π/2 +∞
4 sin(z cos θ) X
dθ = 2 J2m+1 (z),
π 0 2 cos θ m=0

hence:
+∞ Z π/2 Z 1
X 1 sin(z cos θ) 1 sin(zt)
J2m+1 (z) = dθ = √ dt
m=0
π 0 cos θ π 0 t 1 − t2
+∞
(−1)r 1 z
Z
1X
= z 2r+1 = J0 (u)du.
2 r=0
(2r + 1)4r (r!)2 2 0

For the alternating sum, it is sufficient to take θ = 0 in (1) in order to have:


+∞
X 1
(−1)m J2m+1 (z) = sin z.
m=0
2

There are no issues in exploiting the pointwise convergence of a Fourier series since g(θ) = eiz cos θ is an analytic
function.

Page 133 / 222


The integral of sin2 sin2 . We have:
π π
Z2 Z2
1 − cos(2 sin2 (x))
sin2 (sin2 (x))dx = dx
2
0 0
π
Z2 X
π 1 (−1)k 4k
= − sin4k (x)dx
4 2 (2k)!
0 k≥0
π
Z2
π 1 X (−1)k 4k
= − sin4k (x)dx
4 2 (2k)!
k≥0 0
π 1 X (−1)k 4k (4k)! π
= −
4 2 (2k)! (4k (2k)!)2 2
k≥0

π π X (−1)k (4k)!
= − .
4 4 ((2k)!)3 4k
k≥0

On the other hand:


π π π
Z2 Z2 Z2
1 − cos(2 sin2 (x)) 1 − cos(1 − cos(2x))
sin2 (sin2 (x))dx = dx = dx
2 2
0 0 0
π π
Z2 Z2
π cos(1) cos(cos(2x)) sin(1) sin(cos(2x))
= − dx − dx
4 2 2
0 0
π
Z 2 Zπ
π cos(1) π cos(1)
= − cos(cos(2x))dx = − cos(cos(x))dx
4 2 4 4
0 0
π cos(1)π
= − J0 (1).
4 4

Exercise 246. Prove the following identity:


X (2n)!
= e2 I0 (2).
n!3
n≥0

Proof. By exploiting the integral representation for central binomial coefficients we have
π/2
(2 cos θ)2n 2 π/2 4 cos2 θ
X (2n)! Z Z
1
 2
= 1 F1 2 ; 1; 4 = dθ = e dθ
n!3 0 π
n! π 0
n≥0
Z π/2
2e2 e2 π 2 cos ϕ 2e2 π/2
Z Z
2 cos(2θ)
= e dθ = e dϕ = cosh(2 cos ϕ) dϕ
π 0 π 0 π 0
2e2 X 1 Z π/2 X 1
= (2 cos ϕ)2m dϕ = e2 = e2 1 F2 (1; 1, 1; 1) = e2 I0 (2).
π (2m)! 0 m!2
m≥0 m≥0

Exercise 247. Investigate about the series


X (2n)!
S=
n!4
n≥0

and hypergeometric identities provided by such a series.

Page 134 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

Proof. Through the series representation for I0 (2z) and the integral representation for central binomial coefficients we
have
2 π/2 2 1 I0 (4x)
Z Z
S = 1 F2 21 ; 1, 1; 4 =

I0 (4 cos θ) dθ = √ dx
π 0 π 0 1 − x2
and the RHS can be written as the following double series:
2m n X 2n
 
2 X m 4 1
X
= n
= = I0 (2)2 = 1 F2 (1; 1, 1; 1)2 .
π 4m n!2 (2m + 2n + 1) n!2 m!2 n!2
m,n≥0 n≥0 m,n≥0

Actually S is just one of the coefficients of the Fourier cosine series of I0 (4 cos θ):
1 2π X 4n (cos θ)2n
X Z
I0 (4 cos θ) = I0 (2)2 + cm cos(2mθ), cm = cos(2mθ) dθ.
π 0 n!2
m≥1 n≥0
iθ −iθ
e +e
Using both cos θ = Re(eiθ ) = 2 and the binomial theorem we have
2n

X n+m
cm = 2 = 2Im (2)2
n!2
n≥0

and the pointwise convergence of the involved Fourier series leads to


X
I0 (4) = I0 (2)2 + 2 Im (2)2 ,
m≥1

i.e. to an instance of the duplication formula for the I0 function, which can be seen as a direct consequence of Parseval’s
2
identity, too. Since the Fourier coefficients of I0 (4 cos θ) depend on Im , Parseval’s identity provides a curious identity
4
about the series of Im (2) .
X 4n (2n)!
I0 (4z)2 = z 2n
n!4
n≥0
gives
π/2
π X (2n)!2
Z
π
I0 (4 cos θ)2 dθ = 1 1

= · 2 F3 2 , 2 ; 1, 1, 1; 16
0 2 n!6 2
n≥0
hence by the orthogonality relations
X (2n)!2 X
1 1
= I0 (2)4 + 2 Im (2)4 .

= 2 F3 2 , 2 ; 1, 1, 1; 16
n!6
n≥0 m≥1

Some series from Ramanujan and some generalizations. We have


√ X  1/2   
4 1/2
x + 1 − 1 = −4 +2 xm
m−1 m
m≥3

r r !4
1 1 X  1/2  
1/2

1+ − = −4 +2 xm
x x m+1 m+2
m≥1

hence:
X √ √ 4 X  1/2  
1/2

n+1− n = −4 +2 ζ(m)
m+1 m+2
n≥1 m≥1
Z +∞
xm−1
   
dx X 1/2 1/2
= −4 x
+ 2
0 e −1 m+1 m+2 (m − 1)!
m≥1
Z +∞ −x/2
x I0 x2 − 4 I1 x2
 
2e
= dx
0 x2 (ex − 1)
Z +∞
I2 (x) dx
= 2
0 xe (ex − 1)(ex + 1)
x

Page 135 / 222


1 −2x
where the last integrand function is pretty close to 16 e , from which it follows that the original series is pretty
1
close to 16 . The following integral representation for the Bessel function I2
π
x2
Z
I2 (x) = exp (x cos θ) sin4 (θ) dθ
3π 0

leads to:
X √ √ 4
Z π
1
ψ0 3−cos θ
sin4 (θ) dθ

n+1− n = 2
6π 0
n≥1
Z π
1
ψ 0 1−cos θ
 4
= −1 + 2 sin (θ) dθ
6π 0
π π/2 sin4 (θ)
Z
= −1 +  dθ
6 0 sin2 π sin2 θ2
π π/4 sin4 (2θ) dθ
Z
= −1 +
sin2 π sin2 θ

3 0

by the reflection formula for the trigamma function. The blue integral can be written in the more symmetric form

X √ 1 1
√ 4 x3/2 (1 − x)3/2 (1 − x2 )3/2 dx
Z Z
4π π
n+1− n = dx = .
3 0 sin2 (πx) 6 0 cos2 πx2
n≥0

Let us tackle the case s = 3 with a similar approach. We have


√ 3
X  1/2  1/2 
( x + 1 − 1) = 4 + xm
m m−1
m≥3

r r !3
1 1 X  1/2  1/2 
1+ − = 4 + xm−3/2
x x m m−1
m≥3

X √ √ 3 X  1/2  1/2   3

n+1− n = 4 + ζ m−
m m−1 2
n≥1 m≥3
     Z +∞ m−5/2
X 1/2 1/2 1 x
= 4 + 5
 dx
m m−1 m− 2 ! 0 ex − 1
m≥3
Z +∞ −x
3e (2 − 2ex + x + ex x)
= √ dx
0 2 πx5/2 (ex − 1)
Z +∞  
3 1 1 1
= √ − 5/2 x + 5/2 x x dx
π 0 2x3/2 ex x e x e (e − 1)

hence
X √ √ 3 3 3

n+1− n = 2π ζ 2
n≥0

just follows from integration by parts, Frullani’s Theorem and the integral representation for the ζ function.
The same approach allows an explicit evaluation in terms of the ζ function for any odd value of s. For instance:
X √ √ 5 X √ √ 7
15 5 7
ζ 32 − 2π
  105 7

n + 1 − n = 2π 2 ζ 2 , n + 1 − n = 2π 3 ζ 2
n≥0 n≥0

X √ √ 9 90 5 945 9
 
n+1− n = 2π 2 ζ 2 − 2π 4 ζ 2 .
n≥0

Page 136 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

Relations between Hermite polynomials and Bessel functions Starting with the generating function
2 X tn
e2xt−t = Hn (x)
n!
n≥0

then replacing t with teiθ we have


 X tn
exp 2xteiθ − t2 e2iθ = Hn (x)eniθ

n!
n≥0

and by Parseval’s identity


π
t2n
Z X
exp 4xt cos θ − 2t2 cos(2θ) dθ = 2π Hn (x)2 2 .
 
−π n!
n≥0

2
Now we can multiply both sides of (3) by e−x ekix and apply
R
R
(. . .) dx to get
π
√ X Z t2n
Z 
2 2 2
πe−k /4
e2t −2ikt cos θ
dθ = 2π Hn (x)2 e−x ekix dx
−π R n!2
n≥0

which simplifies into  2n


√ −k2 /4 2t2
X Z
2 −x2 kix t
πe e J0 (2kt) = Hn (x) e e dx
R n!2
n≥0
2
and the wanted integral can be recovered from the Cauchy product between the Taylor series of e2t and the
Taylor series of J0 (2kt):
Z
2

√ 2 X 2a (−1)b k 2b t2a+2b
Hn (x)2 e−x ekix dx = πe−k /4 n!2 · [t2n ]
R a!b!2
a,b≥0

such that:
n
√ 2n−b (−1)b k 2b
Z
2 −x2 kix −k2 /4
X
2
Hn (x) e e dx = πe n! .
R (n − b)!b!2
b=0

One could place the right hand side of the last equation into the more known form of
 2
√ n −k2 /4 k
π 2 n! e Ln ,
2

where Ln (x) are the Laguerre polynomials.

8.1 The Gauss circle problem


Disclaimer : most of the contents of this subsection are freely adapted from the books, notes and articles of T. Jameson,
A. Ivic, J. Bell and G.N. Watson. Let N (R) be the number of lattice points in the region {(x, y) : x2 + y 2 ≤ R2 } and
let r2 (N ) be the following arithmetic function:

r2 (N ) = {(x, y) ∈ Z2 : x2 + y 2 = N } .

A double-counting argument easily leads to the following identity:


R 
X jp R2
k X
N (R) = 1+2 R 2 − x2 = r2 (N )
x=−R N =0

and a reasonable claim is that N (R), for sufficiently large values of R, is close to the area of a circle with radius R,
i.e. πR2 . Since Z[i] (the ring of Gaussian integers) is an Euclidean domain, r2 (N ) only depends on the prime factors

Page 137 / 222


8.1 The Gauss circle problem

of N . Such numbers of representations can be shown to be four times a multiplicative function, since Z[i] has four
invertible elements:

X 
 1 if d ≡ 1 (mod 4),
r2 (N ) = 4(χ4 ∗ 1)(N ) = 4 χ4 (d), χ4 (d) = −1 if d ≡ 3 (mod 4),

d|N 0 if d ≡ 0 (mod 2).

The algebra of Dirichlet series hence ensures

def
X r2 (N ) X (−1)n+1
L(r2 , s) = = 4ζ(s)L(χ4 , s), L(χ4 , s) =
Ns (2n + 1)s
N ≥1 n≥0

for any s ∈ C such that Re(s) > 1. We may notice that

X (−1)n+1 Z 1
dx π
L(χ4 , 1) = = 2
=
2n + 1 0 1 + x 4
n≥1

and that summation by parts grants


1
ζ(s) = + γ + O(s − 1)
s−1
as s → 1+ . Summation by parts also gives
√ 

  X  
X 1 1 N ( n) 1 log n
L(r2 , s) = N ( n) s − = + (s − 1) · O
n (n + 1)s n n+1 n2
n≥1 n≥1

hence by assuming that NR(R)


2 is convergent, its limit has to be 4L(χ4 , 1) = π. This can be shown through a simple
geometric argument. Let us consider U (R) as the union of the squares centered at the lattice points √
of {(x, y) :
2 2 2 2
x + y ≤ R }, all of them having unit side length. U (R) is contained in a circle having radius R + 2 . Conversely,

2
U R− 2 is contained in a circle with radius R. It follows that

√ !2 √ !2
2 2
π R− ≤ N (R) ≤ π R +
2 2

hence

Theorem 248 (Gauss). For any ε > 0,



N (R) − πR2 ≤ ( 2 + ε)R = O(R)

as R → +∞.

Soon after Gauss’ work, mathematicians wondered about the optimality of the bound O(R) for the difference between
N (R) and the area of the circle with radius R. Since the Dirichlet L-function L(χ4 , s) has an infinitude of zeroes in
the strip 0 ≤ Re(s) ≤ 1, it can be shown (as done by Hardy) that the bound for the error term cannot be improved
beyond O(R1/2 ). The purpose of the final part of this section is to prove that it can be improved as follows:

Theorem 249 (Sierpinski,Voronoi). As R → +∞,

N (R) − πR2 = O(R2/3 ).


Page 138 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

The classical proofs revolve around a few key ingredients. The first fact is that the sum
R 
X p 
1+2 R 2 − x2
x=−R

can be estimated through the trapezoid method or Simpson’s rule, hence the problem is reduced to an accurate
estimation of
XR np o
R 2 − x2
x=−R

where {z} stands for the fractional part of z. The Fourier sine series of ρ(z) = 21 − {z} is well-known, the discrete

Fourier transform of { R2 − x2 } is related to Bessel functions of the first kind, with a known asymptotic behavior. In
order to partially compensate the erratic behaviour of r2 (n) over the integers, actual bounds for N (R) are produced
by considering an averaged version of r2 (n) over suitably short intervals. Let N 0 (R) denote the number of pairs of

integers (m, n) satisfying m2 + n2 ≤ R2 , m > 0, n ≥ 0 and let M = b √R2 c, f (x) = R2 − x2 − x. Then
M
X M
X M
X
N 0 (R) = M + bRc + 2 bf (m)c = bRc + 2 f (m) + 2 ρ(f (m))
m=1 m=1 m=1

and integration by parts leads to the fundamental formula


M
πR2 X
N 0 (R) = +2 ρ(f (m)) + O(1).
4 m=1
√ √
If we define P (x) as N 0 ( x) − πx 0
4 , from the fact that N ( x) is increasing we get
πy πy
P (X) ≤ P (X + y) + , P (X) ≥ P (X − y) − ,
4 4
hence by integration and the triangle inequality we have
Z Z !
1 X X+Y πY
|P (X)| ≤ max P (x) dx , P (x) dx +

Y X−Y X 8

yielding a bound for P (X), given a bound for the average of P (x) over short intervals. In the following manipulations
√ √ √
we will assume Y  X, such that x = X + O(Y ) ensures x = X + O(1). We have:
X p
P (x) = 2 ρ( x − m2 ) + O(1),

m≤ X/2
Z X+Y X Z X+Y p
P (x) dx = 2 ρ( x − m2 ) dx + O(Y ).
X−Y √ X−Y
m≤ X/2

By using the Fourier series of the fractional part provided by Bernoulli polynomials we obtain
Z X+Y √ 2 X 1 X p h p  p i
P (x) dx = O( X) + 2 Re X − m 2 e X − Y − m 2 −e X + Y − m2 (1)
X−Y π h2 √
h≥1 m≤ X/2

where e(z) is the common shorthand notation for e2πiz . The original problem is now converted into the approximated
evaluation of exponential sums. Such task can be accomplished by recalling two classical results due to Kusmin,
Landau and Van der Corput.

Theorem 250 (Kusmin-Landau). Let fA , . . . , fB be real numbers (where A, B are integers with A < B) and let
δn = fn − fn−1 . Suppose that δn is a weakly monotonic function of n. Suppose also that there is some integer K and
some δ ∈ 0, 21 such that δn ∈ [K + δ, K + 1 − δ] for all n. Then


B
X
def
2
S = e(fn ) ≤ .

sin(πδ)
n=A

Page 139 / 222


8.1 The Gauss circle problem

If fn is decreasing we may negate all the fn s, which simply turns S into S, which has the same size as S. Then we
may assume without loss of generality that δn is non-decreasing. We may also assume that K = 0, by replacing fn by
fn − Kn. The proof of the Kusmin-Landau theorem relies on a clever trick to express e(fn ) as a difference, then on
the application of summation by parts. We begin by writing (for A < n ≤ B)

e(fn ) − e(fn−1 ) = e(fn )(1 − e(−δn ))

then setting
1 1 i
gn = = − cot(πδn )
1 − e(−δn ) 2 2
such that
e(fn ) = gn e(fn ) − gn e(fn−1 )

and
1
|1 − gn | = |g n | = |gn | = .
2 sin(πδn )
The expression on the RHS is positive because from our assumptions we have 0 < δn < 1.
By using summation by parts we get
B−1
X
S = (1 − gA+1 )e(fA ) + gB e(fB ) + (gn − gn−1 )e(fn )
n=A+1

and the claim now follows from the triangle inequality:


B−1
X 1 1 2
|S| ≤ |1 − gA+1 | + |gB | + |gn − gn−1 | ≤ + ≤ .
sin(πδA+1 ) sin(πδB ) sin(πδ)
n=A+1

Suppose we have a differentiable function f (x) with f 0 (x) being monotonic and fulfilling f 0 (x) ∈ [K + δ, K + 1 − δ]
Rn
for all x ∈ [A, B]. Then with fn = f (n) we have δn = f (n) − f (n − 1) = n−1 f 0 (x) dx which clearly satisfies the
conditions for the applicability of the Kusmin-Landau theorem. Since we also have sin(πδ) ≥ 2δ for δ ∈ 0, 12 , we


may write B
X 1
e(f (n)) ≤ .

δ


n=A

Also, if A = B then this sum contains only one term and it has modulus 1, so that this bound holds trivially (since
1 00
δ ≥ 2). Suppose now that we have a twice differentiable f (x) defined for x ∈ [a, b], such that 0 < λ ≤ f (x) ≤ hλ.
Here we have a < b and a, b need not be integers. This implies that h ≥ 1 and that f 0 (x) is strictly increasing. From
the Kusmin-Landau theorem we will derive a bound for
def
X
S = e(f (n)).
a≤n≤b

Let f 0 (a) = α and f 0 (b) = β. For a free parameter δ ∈ 0, 21 we partition the interval [α, β] into sub-intervals of the


form In = (n − δ, n + δ) (containing the reals close to the integer n), and of the form Jn = [n + δ, n + 1 − δ] (containing
reals at least δ-apart from the integers). Accordingly, we split the sum S into subsums over ranges for x corresponding
to these ranges for f 0 (x). The condition that In ⊆ [α, β] is equivalent to the condition n ∈ (α − δ, β + δ), and the
number of ns satisfying this is ≤ β − α + 2. Similarly, the condition that Jn ⊆ [α, β] is equivalent to the condition
n ∈ (α − 1 + δ, β − δ) and the number of ns satisfying this is ≤ β − α + 2 as well. By invoking the Kusmin-Landau
theorem we have  
1 2δ
|S| ≤ (β − α + 2) + +1
δ λ
p
and this expression is minimized by choosing δ = λ/2, leading to:

Page 140 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

Theorem 251 (Van der Corput). !


r
8
|S| ≤ (β − α + 2) +1 .
λ

The square root appearing in the RHS is crucial in encoding the cancellations in the involved exponential sum.
When dealing with weighted exponential sums, we may observe that

X Z b+ X


g(n)cn =
g(x) d cn
a +
a<n≤b a<n≤x

X Z b X
0

= g(b)
cn − g (x) cn dx
a<n≤b a a<n≤x
!
Z b X
0

≤ |g(b)| + |g (x)| dx max cn
a a<x≤b a<n≤x

which is very practical in inserting (or removing, depending on which way around we read things) a smooth slowly
varying weight g(x). In the current case Van der Corput’s theorem leads to
 −1/2
X  p  h
e h X − m2  h √ = X 1/4 h1/2
m≤t
X
p
for t ≤ X/2 + O(1), then summation by parts grants
Xp p
X − m2 e(h X − m2 )  X 3/4 h1/2
m≤t

and X p h p p i √ hY
X − m2 e(h X − Y − m2 ) − e(h X + Y − m2 )  X 3/4 h1/2 · X · = X 1/4 h3/2 Y.
√ X
m≤ X/2

Recalling (1) now we state


Z X+Y X X
P (x) dx  Y X 1/4 h−1/2 + X 3/4 h−3/2
X−Y
h≤X 1/2 Y h>X 1/2 Y

1/2 1/2
−1/2
X 1/2
  
X
 Y X 1/4 + X 3/4
Y Y
 (Y X)1/2

and the mean-to-max trick gives us


 1/2
X
P (X)  Y +
Y
which is clearly minimized by taking Y = X 1/3 , finally giving

P (X)  X 1/3 .

It should not take too much imagination to envisage that there is nothing spectacularly important about us restricting
to a circle in all of this, so that (using mainly just a little notational gameplay) we can get similar results for simple
closed curves satisfying certain conditions (like twice differentiability and radius of curvature bounded above and
below). We have had formulae with square root signs all over the place coming from Pythagoras, but in essence we
are just using linearising tricks. More usually a proof is given using the Poisson summation formula and bounds for

Page 141 / 222


8.1 The Gauss circle problem

the resulting “exponential integrals”: this is exactly the second approach we are going to outline.

The Sonine-Schafheitlin integrals ensure that


2s−p−1 Γ(s/2) 3

0 < Re(s) = σ < p + 2
Γ(p + 1 − s/2)

is the Mellin transform of x−p Jp (x). This brings to the table the Perelli S-class of Dirichlet series.
If an is a arithmetic function and the associated Dirichlet series
X an
L(a, s) = Re(s) > 1
ns
n≥1

has an analytic continuation to C whose only possible pole is at s = 1, and there are A, a1 , . . . , ag such that
g
Y
γ(s) = As Γ(aj s)
j=1

fulfills the reflection formula γ(s)L(a, s) = γ(1 − s)L(a, 1 − s), then by assuming f ∈ S(R) and defining
Z +∞ Z Z +∞
1 γ(s)
M (f )(s) = s−1
x f (x) dx, H(x) = −s
x ds, g(x) = f (y)H(xy) dy
0 2πi Re(s)= 32 γ(1 − s) 0

we have:

X X
an f (n) = f (0)L(a, 0) + Res M (f )(s)L(a, s) + an g(n).
s=1
n≥1 n≥1

Since f is a Schwartz function its Mellin transform M (f ) is holomorphic on Re(s) > 0. Over such region integration
by parts ensures
1
M (f )(s) = − M (f 0 )(s + 1),
s
hence M (f ) has an analytic continuation to C possibly with poles at 0, −1, −2, . . ..
Denoting M (f ) as F , the Mellin inversion formula grants
X X an Z
an f (n) = n−s F (s) ds
2πi Re(s)= 23
n≥1 n≥1
Z
1
= F (s)L(a, s) ds.
2πi Re(s)= 32

The only possible pole of L(a, s) is at s = 1. From M (f )(s) = − 1s M (f 0 )(s + 1) the only possible pole of F (s) in the
half-plane Re(s) > −1 is at s = 0, and the residue of F (s) at s = 0 is
Z +∞
−M (f 0 )(1) = − f 0 (x) dx = f (0),
0

so the residue of F (s)L(a, s) at s = 0 is f (0)L(a, 0). By the residue theorem, taking as given that F (s)L(a, s) → 0
uniformly in − 21 ≤ Re(s) ≤ 23 as |Im(s)| → ∞, we have
Z
X 1
an f (n) = f (0)L(a, 0) + Res F (s)L(a, s) + F (s)L(a, s) ds
s=1 2πi Re(s)=− 21
n≥1

γ(s)
by shifting the integration line. Now we may exploit the reflection formula for L(a, s), introducing G(s) = F (s) γ(1−s) .
R
The last integral in the previous line turns into Re(s)= 3 G(s)L(a, s) ds, and it is tedious but straightforward to check
2
that Z
1
x−s G(s) ds = g(x),
2πi Re(s)= 32

Page 142 / 222


8 BESSEL FUNCTIONS AND THE GAUSS CIRCLE PROBLEM

proving the Lemma through the Mellin inversion formula. Since L(r2 , s) belongs to the Perelli S-class and

Γ 2−s
(s+1)/2

1−2s π 2 
L(χ4 , s) = 2 L(χ4 , 1 − s)
π (2−s)/2 Γ 1+s
2

holds as a consequence of the Poisson summation formula, we have the following resummation formula for r2 :

Theorem 252 (Voronoi).

X0 Z b X Z b √
r2 (n)f (n) = π f (x) dx + r2 (n) f (x)J0 (2π xn) dx
a≤n≤b a n≥1 a

P0
where f (x) is a suitably smooth function and denotes that at n = a or n = b the summand is to be halved if a or
b is an integer.

11
In view of the non-negativity of r2 (n) we have
X X X
f− (n)r2 (n) ≤ r2 (n) ≤ f+ (n)r2 (n)
n≥1 X<n≤2X n≥1

where f− is a smooth, non-negative function supported in [X, 2X] such that f (x) = 1 for x ∈ [X + G, 2X − G]

(X ε ≤ G ≤ X), while similarly f+ is supported in [X − G, 2X + G] and satisfies f (x) = 1 for x ∈ [X, 2X]. If
henceforth we denote by f (x) either f− (x) or f+ (x), then f (r) (x) r G−r (r = 0, 1, 2, . . .) and by the Voronoi
summation formula Z 2X+G
X X √
f (n)r2 (n) = πX + O(G) + r2 (n) f (x)J0 (2π xn) dx.
n≥1 n≥1 X−G

d
From the theory of Bessel functions we recall the identity dz [z ν Jν (z)] = z ν Jν−1 (z) and the bound Jν (z)  √1z
for z → +∞. It follows that for small values of n (n ≤ Y ) the integral appearing in the RHS of the last line is
 X 1/4 n−3/4 , and by invoking n≤x r2 (n)  x and summation by parts we have
P

X
r2 (n)X 1/4 n−3/4  (XY )1/4 .
n≤Y

By using the recurrence relation for Jν0 , performing two integration by parts and noting that the support of f 00 has
measure  G, we obtain that
X Z 2X+G √ X r2 (n) Z 2X+G √
r2 (n) f (x)J0 (2π xn) dx = 2
f 00 (x)xJ2 (2π xn) dx
X−G π n X−G
n>Y n>Y
X
 r2 (n)n−5/4 G−1 X 3/4  X 3/4 G−1 Y −1/4 .
n>Y

By combining the bounds obtained for small/large values of n we get


X
f (n)r2 (n) = πX + O(G) + O((XY )1/4 ) + O(X 3/4 G−1 Y −1/4 )
n≥1

which simplifies into X


r2 (n) = πX + O(X 1/3 )
X<n≤2X

by choosing G = Y = X 1/3 . Replacing X with 2−j R2 and summing over j ≥ 1,

N (R) − πR2 = O(R2/3 )


11 A resummation formula exists every time a Dirichlet L-series has a suitably structured reflection formula, and vice-versa.

Page 143 / 222


8.1 The Gauss circle problem

is finally proved. Various authors have exploited the oscillations of J0 and J1 to refine the previous bounds. For
instance Huxley has proved (in 2003) that the exponent 23 can be replaced with the slightly smaller 131 208 . In the

opposite direction, Hardy has shown in 1925 that N (R) − πR2 is as large as R1/2 infinitely often, by exploiting the
reflection formula X r2 (n) √ X r2 (n) √
√ e−2π (n+a)b = √ e−2π (n+b)a
n≥0
n+a n≥0
n+b

due to Ramanujan. The techniques outlined in this section can be applied to the divisor problem too, concerning the
number of lattice points in the first quadrant under a rectangular hyperbola. We have
X0
D(x) = d(n) = x(log x + 2γ − 1) + 41 + ∆(x)
n≤x

where ∆(x) = O(x1/2 ) follows from simple geometric arguments. Voronoi’s summation formula for d(n) involves Bessel
functions of the second kind and it allows to prove that ∆(x) = O(x1/3 log x). In the opposite direction, ∆(x) is as
large as x1/4 log x infinitely often.

Page 144 / 222


9 REMARKABLE RESULTS IN LINEAR ALGEBRA

9 Remarkable results in Linear Algebra


This brief section is devoted to some important results in Linear Algebra: we will outline proofs by density for the
Hamilton-Cayley Theorem and the identity Tr(AB) = Tr(BA), then we will outline a recent elementary proof (due to
Suk-Geun Hwang) of Cauchy’s interlace theorem, admitting Sylvester’s criterion as a straightforward corollary.

Theorem 253 (Hamilton-Cayley). If p ∈ C[x] is the characteristic polynomial of a n × n matrix A with complex
entries,
p(A) = 0.

Proof. If A is a diagonalizable matrix the claim is trivial: if A = J −1 DJ is the Jordan normal form of A, the diagonal
entries of D are the eigenvalues λ1 , . . . , λn of A. We have p(λj ) = 0 by the very definition of characteristic polynomial,
and since p(M k ) = p(M )k ,
p(A) = p(J −1 DJ) = J −1 p(D)J = 0.

On the other hand diagonalizable matrices form a dense subspace in the space of n × n matrices with complex entries.
Assuming that A is not a diagonalizable matrix it follows that for any ε > 0 there exist some diagonalizable matrix
Aε such that kA − Aε k2 ≤ ε (actually the choice of the Euclidean norm is immaterial, any induced norm does the job
equally fine). The eigenvalues of Aε converge to the eigenvalues of A and p is a continuous function, hence

p(A) = lim p(Aε ) = 0.


ε→0

Exercise 254. We have that {xn }n≥1 , {yn }n≥1 , {zn }n≥1 are three sequences of real numbers such that any term
among xn , yn , zn can be written as a linear combination of xn−1 , yn−1 , zn−1 with constant coefficients, for instance:

xn = 4yn−1 + zn−1 , yn = 3yn−1 − zn−1 , zn = xn−1 + yn−1 − 2zn−1 .

Prove that there exists an order-3 linear recurrence relation simultaneously fulfilled by each one of the sequences
{xn }n≥1 , {yn }n≥1 , {zn }n≥1 .

Exercise 255. The sequence {an }n≥1 is defined through

2an
a1 = 1, an+1 = .
7 + an
log an
Prove that limn→+∞ an = 0, then find limn→+∞ log n .

Theorem 256 (“The trace is Abelian”). For any couple (A, B) of n × n matrices with complex entries, the following
identity holds:
Tr(AB) = Tr(BA).

Proof. Assuming A is an invertible matrix, AB and BA share the same characteristic polynomial, since they are
conjugated matrices due to BA = A−1 (AB)A. In particular they have the same trace. Equivalently, they share the

Page 145 / 222


same eigenvalues (counted according to their algebraic multiplicity) hence they share the sum of such eigenvalues. On
def
the other hand, if A is a singular matrix then Aε = A + εI is an invertible matrix for any ε 6= 0 small enough. It
follows that Tr(Aε B) = Tr(BAε ), and since Tr is a continuous operator, by considering the limits of both sides as
ε → 0 we get Tr(AB) = Tr(BA) just as well.

Corollary 257. If A is a n × n matrix with real entries and B = AT , A and B have the same rank. Prove this
statement by showing that Tr(Ak ) = Tr(B k ) for any k ∈ N. Hint: notice that Tr(M ) = Tr(M T ) and that the power
sums of the eigenvalues fix the coefficients of the characteristic polynomial of a matrix.

Exercise 258. Prove that any matrix T with real entries such that Tr(T ) = 0 can be written in the form T = AB −BA
for a suitable choice of the matrices A, B.

Unexpected applications of the Hamilton-Cayley theorem: if you are able to draw it,
you also know its asymptotic behaviour.

Exercise 259. Let Tn be the number of strings over the alphabet Σ = {0, 1} with length n
and exactly one occurrence of the “11” substring. Find an explicit formula for Ln .

Proof. It is pretty simple to construct a finite automaton accepting the strings of our (regular) language.
The automaton depicted above has the following transition matrix:
 
1 1 0 0
1 0 1 0
M = .
 
0 0 0 1
0 0 1 1

Since the starting state is A and the accepting states are C, D, we simply have:

Tn = (1 0 0 0) M n (0 0 1 1)T

and by the Hamilton-Cayley Theorem the sequence √


{Tn }n≥0 and the matrix M share the same characteristic
polynomial. Since the eigenvalues of M are 1±2 5 and they both have algebraic multiplicity 2 and geometric
multiplicity 1, by the Jordan decomposition of M we have:
 √ n  √ n
Tn = (a + bn) 1+2 5 + (c + dn) 1−2 5

and the constants a, b, c, d can be found by interpolation, through T0 = T1 = 0, T2 = 1, T3 = 2.


In particular we have Tn = 51 (nLn − Fn ). To prove the same in a purely combinatorial fashion is a bit more
involved, but certainly not impossible. By stars and bars, the number of strings with length n − 1 and exactly k
non-adjacent 1s is given by n−k

k . In particular:

X n − k  n−1
X
Tn = k= Ak An−k
k
k≥1 k=1

Page 146 / 222


9 REMARKABLE RESULTS IN LINEAR ALGEBRA

where Ak is the number of strings with length k, having no adjacent 1s and starting with a 1. By the convolu-
tion machinery the generating function of {Tn }n≥0 is simply given by the square of the generating function of
{An }n≥0 = {Fn+1 }n≥0 . The latter is a meromorphic function with two simple poles and we may recover the
previous closed form by partial fraction decomposition.

A
Exercise 260. ( ) Let s be a non-empty string over the alphabet Σ = {0, 1}. Let As be the finite automaton
accepting the strings over Σ that do not contain s as a substring. Let us define Spec(s) as the spectrum of the
transition matrix of As . Investigate about the relations between Spec(s.t) and Spec(s), Spec(t), where . denotes
the concatenation of strings.

A
Exercise 261. ( ) According to the notation introduced in the previous exercise, prove or disprove the existence
of a string s such that 2 cos 2π
7 ∈ Spec(s).

Definition 262. We say that a weakly increasing sequence of real numbers a1 ≤ a2 ≤ . . . ≤ an interlaces another
weakly increasing sequence of real numbers b0 ≤ b1 ≤ . . . ≤ bn if

b0 ≤ a1 ≤ b1 ≤ a2 ≤ . . . ≤ an ≤ bn

holds.

Theorem 263 (Cauchy’s interlace Theorem). The eigenvalues of a Hermitian matrix A of order n are interlaced with
those of any principal submatrix of order n − 1.

Proof. Hermitian matrices have real eigenvalues. Let A be a Hermitian matrix of order n and let B be a principal
submatrix of A of order n − 1. If λn ≤ λn−1 ≤ . . . ≤ λ1 lists the eigenvalues of A and µn ≤ µn−1 ≤ . . . ≤ µ2 the
eigenvalues of B, we shall prove that

λn ≤ µn ≤ λn−1 ≤ µn−1 ≤ . . . ≤ λ2 ≤ µ2 ≤ λ1 .

Proofs of this theorem have been based on Sylvester’s law of inertia and the Courant-Fischer min-max theorem.
Here we will give a simple, elementary proof of the theorem by using the intermediate value theorem.
Simultaneously permuting rows and columns, if necessary, we may assume that the submatrix B occupies rows 2, 3, . . . n
and columns 2, 3, . . . , n, so that A has the form
!
a y∗
A=
y B

where ∗ stands for the conjugate transpose of a matrix. Let D = diag(µ2 , µ3 , . . . , µn ). Then, since B is also Hermitian,
by the spectral Theorem there exists a unitary matrix U of order n − 1 such that U ∗ BU = D. Let U ∗ y = z =
(z2 , z3 , . . . , zn )T . We first prove the theorem for the special case where µn < µn−1 < . . . < µ3 < µ2 and zi = 0 for
i = 2, 3, . . . , n. Let !
1 0T
V =
0 U

Page 147 / 222


in which 0 denotes the zero vector. Then V is a unitary matrix and
!
∗ a z∗
V AV = .
z D

Let f (x) = det(xI − A) = det(xI − V ∗ AV ), where I denotes the identity matrix. Expanding det(xI − V ∗ AV ) along
the first row, we get
n
X
f (x) = (x − a)(x − µ2 )(x − µ3 ) · · · (x − µn ) − fi (x)
i=2

2
where fi (x) = |zi | (x − µ2 ) · · · (x\ − µi ) · · · (x − µn ) for i = 2, 3, . . . , n, with the hat-sign denoting a missing term.
We may notice that fi (µj ) = 0 when j 6= i and fi (µi ) is strictly positive or strictly negative according to i being,
respectively, even or odd. It follows that f (µi ) is positive if i is odd and negative if i is even. Since f (x) is a
polynomial of degree n with positive leading coefficient, the intermediate value Theorem ensures the existence of n
roots λ1 , λ2 , . . . , λn of the equation f (x) = 0 such that λn < µn < λn−1 < µn−1 < . . . < λ2 < µ2 < λ1 .
For the proof of the general case, let ε1 , ε2 , . . . be a sequence of positive real numbers such that εk is decreasing towards
zero, zi + εk 6= 0 for i = 2, 3, . . . , n and k = 1, 2, . . . and the diagonal entries of D + εk diag(2, 3, . . . , n) are distinct for
fixed k. For k = 1, 2, . . . let !
a z(εk )∗
Ck =
z(εk ) D(εk )
where z(εk ) = z + εk (1, 1, . . . , 1)T and D(εk ) = D + εk diag(2, 3, . . . , n), and let Ak = V Ck V ∗ . Then Ak is Hermitian
(k) (k) (k) (k)
and Ak converges towards A. Let λn ≤ λn−1 ≤ . . . ≤ λ2 ≤ λ1 list the eigenvalues of Ak . Then

(k) (k) (k)


λ(k)
n < µn + nεk < λn−1 < µn−1 + (n − 1)εk < . . . < λ2 < µ2 + 2εk < λ1 .
(k) (k) (k)
Since λn , λn−1 , . . . , λ1 are n distinct roots of det(xI − Ak ) = 0 for each k and since the graph of y = det(xI − Ak )
 close to that of y= det(xI − A), it follows that the proof is complete by invoking the implicit function
is sufficiently
(k) (k) (k)
Theorem: λn , λn−1 , . . . , λ1 → (λn , λn−1 , . . . , λ1 ).

Theorem 264 (Sylvester’s criterion). A Hermitian matrix M is positive-definite


if and only if the determinants of the leading principal minors are positive.

Proof. If some minor has a negative or zero determinant the original matrix M cannot be positive definite by Cauchy’s
interlace theorem. This proves that the positivity of the mentioned determinants is a necessary condition. The converse
implication can be easily shown by induction on the dimension of M , or by exploiting the Cholesky decomposition
A = B T B with B being a non-singular matrix.

Theorem 265 (Banach-Steinhaus uniform boundedness theorem). Let F be a family of bounded linear operators
from a Banach space X to a normed linear space Y . If F is pointwise bounded (i.e., supT ∈F kT xk < ∞ for all x ∈ X),
then F is norm-bounded (i.e., supT ∈F kT k < ∞).

Proof. The following proof is due to Alan D.Sokal. Let T be a bounded linear operator from a normed linear space X
to a normed linear space Y . Then for any x ∈ X and r > 0, we have

sup kT x0 k ≥ kT kr,
x0 ∈B(x,r)

where, as usual, B(x, r) = {x ∈ X : kx0 − xk < r}. Indeed, for any ξ ∈ X we have
kT (x + ξ)k + kT (x − ξ)k
max{kT (x + ξ)k, kT (x − ξ)k} ≥ ≥ kT ξk
2

Page 148 / 222


9 REMARKABLE RESULTS IN LINEAR ALGEBRA

where the second ≥ uses the triangle inequality in the form kα − βk ≤ kαk + kβk, and we may consider the supremum
over ξ ∈ B(0, r). If we assume that supT ∈F kT k = ∞, we may construct a sequence {Tn }n≥1 such that kTn k ≥ 4n .
Setting x0 = 0, we may use the previous lemma on supx0 ∈B(x,r) kT x0 k to choose inductively xn ∈ X such that
kxn − xn−1 k ≤ 3−n and
2
kTn xn k ≥ 3−n kTn k.
3
{xn }n≥1 is a Cauchy sequence, hence it is convergent to some x ∈ X: it is easy to check that kx − xn k ≤ 12 3−n , such
n
that kTn xk ≥ 16 3−n kTn k ≥ 16 43 → ∞, contradicting the pointwise-boundedness of F .

Corollary 266. There is a 2π-periodic and continuous function f whose Fourier series
Z 2π
X 1
b inx
f (n)e , f (n) =
b f (x)e−nix dx
2π 0
n∈Z

does not converge at x = 0.

Proof. In order to invoke the Banach-Steinhaus theorem, we consider the functionals given by the partial sums of the
Fourier series of f , evaluated at x = 0: X
λN (f ) = fb(n).
|n|≤N

There is a simple upper bound, namely


Z 1 X X
|λN (f )| ≤ e−2πinx · |f (x)| dx ≤ kf k∞ · e−2πinx .


0 1
|n|≤N |n|≤N

If we take g(x) as the sign of the Dirichlet kernel


1

X
−2πinx sin 2πx N + 2
e =
sin(πx)
|n|≤N

and {gj (x)}j≥1 as a sequence of periodic continuous functions, such that |gj (x)| ≤ 1 and gj (x) → g(x), by dominated
convergence Z 1 Z 1 X
X
lim λN (gj ) = g(x) e−2πinx dx = e−2πinx dx,


j→+∞ 0 0
|n|≤N |n|≤N

2
hence the previous bound for the norm of λN holds as an equality. Since the mean value of |sin x| is π, integration by
parts leads to X 2
e−2πinx ∼ log N,

π

1
|n|≤N

hence there is no uniform bound for the L1 -norm of the Dirichlet kernel. By Banach-Steinhaus, there is some f in the
unit ball of C 0 (T) such that
sup |λN (f )| = +∞.
N

In fact, the collection of such f s is dense in the unit ball, and it is an intersection of a countable collection of dense
open sets (a Gδ ).

The same phenomenon does not occur if the Dirichlet kernel is replaced by the Fejér kernel: if f ∈ C 0 (T),
the sequence of trigonometric polynomials defined by
X  
|n| b
pN (x) = 1− f (n)e−2πinx
N
|n|≤N

converges uniformly to f (x).

Page 149 / 222


P
Corollary 267. If A = {an }n≥1 is a sequence of real numbers such that |hΛ, Ai| = n≥1 λn an is finite

for any Λ = {λn }n≥1 ∈ `2 , then B ∈ `2 .

Proof. We work by contradiction: we assume that A = n≥1 a2n is unbounded and we show that for some sequence Λ ∈
P

`2 the series n≥1 λn an is positively divergent. By picking the multipliers λn with the same sign as the corresponding
P

an , we may as well assume that an ≥ 0. For any real number c > 0, the subsequence {aσ(n) }n≥1 made by the terms
≥ c has a finite number of terms: otherwise the inner product with the sequence {λσ(n) = n1 }n≥1 ∈ `2 would be
unbounded. In particular 0 is the only accumulation point of {an }n≥1 and limn→+∞ an = 0. Let us define a0 = 1,
n
X
Sn = a2n
k=0
q
1 1
In this case Λ ∈ `2 by construction, since λ2n is a telescopic series and
P
and let us consider λn = Sn−1 − Sn . n≥1
Sn → +∞ by the original assumption B 6∈ `2 . If we manage to prove that
s
X 1 1 X a2 X a2
an − = p n ≥ n
Sn−1 Sn Sn−1 Sn Sn
n≥1 n≥1 n≥1

is divergent we are done. Let us define, by induction, τ (0) as the smallest n such that Sn ≥ 2, τ (m) as the smallest n
such that Sτ (m) is ≥ 2 · 2 · Sτ (m−1) . We have
X a2n 1  1
≥ Sτ (m+1) − Sτ (m) ≥
Sn Sτ (m+1) 2
τ (m)<n≤τ (m+1)
q
1 1
P
hence by summing both sides on m ≥ 0 we have that n≥1 an Sn−1 − Sn is divergent.

The last proof can be easily adapted to the continuous case through very few adjustements:

Corollary 268. If f : R+ → R is a function such that f · g ∈ L1 (R+ ) for any g ∈ L2 (R+ ), then f ∈ L2 (R+ ).

Rx
Proof. We work by contradiction: we assume that 0 f (t)2 dt is unbounded and we show that for some function
x
g ∈ L2 (R+ ) the integral 0 f (t)g(t) dt is positively divergent as x → +∞. By picking the multipliers g(x) such that
R

f /g has almost everywhere the same sign, we may as well assume that f (x) ≥ 0. For any real number c > 0, the set
of x ∈ R+ such that f (x) ≥ c has finite measure, otherwise the integral of √fx(x)
2 +1
over such set would be unbounded.
By replacing f with the convolution between f and a non-negative, compact supported and smooth kernel we may
also assume that f (x) is continuous on R+ . Let us define
Z x
F (x) = f (t)2 dt
0
r  
d
and let us consider g(x) = dx − 1+F1 (x) . In this case g ∈ L2 (R+ ) by construction, since the integral of g(x)2 can
be computed through the fundamental Theorem of Calculus and F (x) is increasing to +∞. If we manage to prove
that s 
Z +∞ Z +∞
f (x)2

d 1
f (x) − dx = dx
0 dx 1 + F (x) 0 1 + F (x)
is divergent we are done. Let us define, by induction, τ (0) as the infimum of the set {x : F (x) ≥ 2},
τ (m) as the infimum of the set {x : F (x) ≥ 2F (τ (m − 1))}. We have
Z τ (m+1) Z τ (m+1)
f (x)2 1 F (τ (m + 1)) − F (τ (m)) 1
dx ≥ f (x)2 dx = ≥
τ (m) 1 + F (x) 1 + F (τ (m + 1)) τ (m) 1 + F (τ (m + 1)) 3

Page 150 / 222


10 THE FUNDAMENTAL THEOREM OF ALGEBRA

r  
R +∞ d
hence by summing both sides on m ≥ 0 we have that 0
f (x) dx − 1+F1 (x) dx is divergent.

The powerful lemma 267 is a Corollary of the Banach-Steinhaus theorem: given the sequence A = {an }n≥1 and any
Λ ∈ `2 , the operators
XN
TN (Λ) = an λn
n=1

are linear, continuous and pointwise bounded. The uniform boundedness criterion ensures they are norm-bounded,
i.e. A ∈ `2 . The lemma 267 immediately leads to the fact that `2 and L2 (R+ ) are complete spaces. Since Fourier
series give an isometry between L2 (0, 2π) and `2 , such lemma also provides a proof of the completeness of L2 (I) for
any bounded interval I.

10 The Fundamental Theorem of Algebra


The purpose of this section is to shortly introduce some elements of Complex Analysis and use them to produce a
proof of the Fundamental Theorem of Algebra, stating that C is an algebraically closed field, or, in layman’s terms:

Theorem 269. Any non-constant p ∈ C[z] vanishes at some z ∈ C.

The first serious attempt to such problem is due to Gauss: it was mainly geometric, but it had a topological gap,
filled by Alexander Ostrowski in 1920. A rigorous proof was first published by Argand in 1806 (and revisited in 1813).
We will actually show many approaches and focus on geometric and analytic insights and their consequences. All
proofs below involve some analysis, or at least the topological concept of continuity of real or complex functions. Some
also use differentiable or even analytic functions. This fact has led to the remark that the Fundamental Theorem of
Algebra is neither fundamental, nor a theorem of algebra.

Definition 270. Suppose we are given a closed, oriented curve in the xy plane, not going through the origin. We
can imagine the curve as the path of motion of some object, with the orientation indicating the direction in which
the object moves. Then the winding number of the curve is equal to the total number of counterclockwise turns
that the object makes around the origin. When counting the total number of turns, counterclockwise motion counts
as positive, while clockwise motion counts as negative. For example, if the object first circles the origin four times
counterclockwise, and then circles the origin once clockwise, then the total winding number of the curve is three.

But given a closed curve γ : [0, 2π] → R2 \ {(0, 0)} represented by γ(t) = (x(t), y(t)) (which we may temporarily
assume to be smooth, too), how can we find its winding number around the origin? We may notice that in the open
y(t)
first quadrant arctan x(t) gives an “angular displacement” with respect to the origin: in order to compute the winding
number of γ we just need to find a continuous determination of such angular displacement. For brevity we will not
delve into the theory of differential forms, we just mention that such continuous determination can be achieved through
a step of differentiation and a step of integration, namely:
d y(t)
d y(t) dt x(t) x(t)y 0 (t) − y(t)x0 (t)
arctan = y(t)2
=
dt x(t) 1 + x(t) 2
x(t)2 + y(t)2

leading to

x(t)y 0 (t) − y(t)x0 (t)
Z
1
dt
2π 0 x(t)2 + y(t)2
as an expression for the winding number of γ around the origin. What if our curve is given by some f : S 1 → C∗ ,
with f being a holomorphic function? In such a case the previous winding number takes the following form:
f 0 (z)
I
1
dz.
2πi |z|=1 f (z)

Page 151 / 222


Now it comes an interesting remark, clarifying the interplay between winding numbers and zeroes of holomorphic
d
functions: in the previous line, the integrand function is formally dz log f (z), hence if f = hg with h and g being
holomorphic functions, the winding number of f is just the sum between the winding number of h and the winding
number of g. It is very simple to check that for any m ∈ N the winding number of z m is exactly m and for every
w ∈ C\S 1 , by setting f (z) = z −w we get that the winding number of f is 1 or 0 according to |w| < 1 or |w| > 1: in the
first case f (S 1 ) encloses the origin, in the latter it does not. Conversely, if f (z) is a holomorphic and non-vanishing
0 0
function over D, then ff (z) (z)
is a holomorphic function over D and the integral ∂D ff (z) (z)
H
dz equals zero by Stokes’
theorem.

Lemma 271. If f is a holomorphic function on D = {z ∈ C : |z| ≤ 1}, non-vanishing over ∂D,

f 0 (z)
I
1
Nf = dz
2πi ∂D f (z)

equals the number of zeroes of f in D, counted according to their multiplicity.

This Lemma can be seen both as a consequence of the residue Theorem or as a remark in Differential Geometry that
can be used to prove the residue Theorem. Such Lemma has a crucial role in the usual proof of the Jordan curve
Theorem, since it gives that a smooth, simple and closed curve cannot partition R2 in more than two connected
components. A curve fulfilling such constraints splits R2 in at least two connected components by the existence of
a tubular neighbourhood, then the chance to drop the previous smoothness assumption is granted by invoking Sard’s
Theorem.

Theorem 272 (The double leash principle). Assume that you are connected to a thin tree by a leash of fixed length
L. Assume that you dog is connected to you by a leash of fixed length l < L. If you take a walk and after some time
you and your dog return at the starting points, your winding number around the tree and your dog’s are the same.

The above statement is usually known as Rouché Theorem. We opted for such fancy introduction since we believe
the previous formulation might help the reader to grasp the geometric idea faster and better. In a more rigorous way:

Theorem 273 (Rouché). If f (z) and f (z) + g(z) are holomorphic function on the closed unit disk D centered at the
origin, and for every z ∈ ∂D we have 0 < |g(z)| < |f (z)|, then f and f + g have the same number of zeros inside D,
where each zero is counted as many times as its multiplicity. The same holds if D is replaced by some compact region
K whose boundary ∂K is a simple, piecewise-smooth and closed curve.

Proof. We have already shown that the wanted number of zeroes is given by a winding number. Since for any z ∈ ∂D
we may write  
g(z)
f (z) + g(z) = f (z) 1 +
f (z)
the winding number of f + g is given by the sum between the winding number of f and the winding number of the
curve
def g(eiθ )
h : S 1 → C∗ , h(eiθ ) = 1 + .
f (eiθ )
However, the winding number of h is clearly zero, since h stays “on the right” of the origin. As a matter of fact,

g(z)

f (z)

Page 152 / 222


10 THE FUNDAMENTAL THEOREM OF ALGEBRA

is a continuous function on a compact set (S 1 ) attaining a maximum M < 1, hence Re(h) ≥ 1 − M > 0 by the triangle
inequality. It follows that f and f + g have the same winding number, hence the same number of zeroes in D.

We are ready to prove the following statement:

Lemma 274. The Fundamental Theorem of Algebra is a simple consequence of the double leash principle.

Proof. We may assume without loss of generality that

p(z) = z n + an−1 z n−1 + . . . + a0 ∈ C[z]

is a monic polynomial with degree n ≥ 1 and p(0) = a0 6= 0. Our purpose is to prove it has a zero somewhere.
We prove first that if it has a zero, it cannot be too far from the origin. Let us consider

M = 1 + |an−1 | + . . . + |a0 |

and show that |z| ≥ M implies p(z) 6= 0. Since M > 1, for any z ∈ C such that |z| ≥ M we have:

an−1 z n−1 + . . . + a0 ≤ |an−1 ||z|n−1 + . . . + |a0 | ≤ (|an−1 | + . . . + |a0 |) |z|n−1 = (M − 1)|z|n−1 < |z|n


hence p(z) = 0 cannot occur, since it would imply |z|n = an−1 z n−1 + . . . + a0 . This proves all the complex zeroes of p,
if existing, lie in the region |z| < M . But the inequality above also shows that f (z) = z n and g(z) = an−1 z n−1 +. . .+a0
meet the hypothesis of Rouché’s Theorem for the the region K = {z ∈ C : |z| ≤ M }. In particular the number of
zeroes of p(z) = f (z) + g(z) in K equals the number of zeroes of f (z) = z n in K and we are done:

p(z) has exactly n zeroes in the region |z| < M, counted according to their multiplicity.

The inclusion
{z ∈ C : p(z) = 0} ⊂ {z ∈ C : |z| ≤ M }

can also be proved by applying the Gershgorin circle Theorem to the companion matrix of p.

Exercise 275. Prove that for any n ≥ 5 the polynomial

pn (z) = z n + z + 1

has approximately n3 roots in in the region |z| ≤ 1, with an error


not greater than one.

Sketch of proof. If n ≡ 2 (mod 3) the primitive third roots of


unity ω, ω 2 are also roots of pn . Conversely, if we denote as D
the unit disk centered at the origin, we may notice that ∂D and
−(∂D + 1) intersect only at ω, ω 2 : it follows that the n ≡ 2
(mod 3) case is the only case in which pn (z) has roots at ∂D.
The number of roots we want to approximate is given by the
winding number of the curve γn : [0, 2π] → C, γn (θ) = pn (eiθ ).
The diagram to the right depicts the n = 13 case, for instance.
The graph of e13iθ + eiθ + 1 for θ ∈ [0, 2π].

Page 153 / 222


It is pretty clear that the graph of γn is given by the union of n approximated
circles, completing a revolution around the point z = 1 in n steps. Assuming
n 6≡ 2 (mod 3) the number of roots we are interested in is exactly given by
the number of the previous approximated circles enclosing the origin. If such
portions of γn were perfect circles, from the diagram on the left it would be
clear that about n3 of them would enclose the origin. To fill in the missing
details is a task we leave to the reader.

We now outline another classical proof of the Fundamental Theorem of Algebra, relying on the following results:

Theorem 276 (Maximum modulus principle). If D is a closed disk in the complex plane and f is a non-constant
holomorphic function over D,
max |f (z)|
z∈D

is attained at ∂D.

Proof. If we assume that maxz∈D |f (z)| is attained at some z0 belonging to the interior of D we get a contradiction,
since by Cauchy’s integral formula or termwise integration of a Taylor series we have
I
1 f (z)
f (z0 ) = dz
2πi |z−z0 |=ε z
for any ε > 0 small enough, implying I
1
|f (z0 )| ≤ |f (z)| dz.
2πε |z−z0 |=ε

If equality holds for any ε small enough then |f (z)| is constant in a neighbourhood of z0 and f (z) is constant as
well.

Theorem 277 (Liouville). If a holomorphic function over C is bounded, it is constant.

Proof. The theorem follows from the fact that holomorphic functions are analytic.
If f is an entire function, it can be represented by its Taylor series about 0:

X
f (z) = ak z k
k=0

where by Cauchy’s integral formula


f (k) (0)
I
1 f (ζ)
ak = = dζ
k! 2πi Cr ζ k+1
and Cr is the circle about 0 of radius r > 0. Suppose f is bounded: i.e. there exists a constant M such that |f (z)| ≤ M
for all z. We can estimate directly
|f (ζ)|
I I I
1 1 M M M M
|ak | ≤ k+1
|dζ| ≤ k+1
|dζ| = k+1
|dζ| = k+1
2πr = k ,
2π Cr |ζ| 2π Cr r 2πr Cr 2πr r
where in the second inequality we have used the fact that |z| = r on the circle Cr . But the choice of r above is arbitrary.
Therefore, letting r tend to infinity gives ak = 0 for all k ≥ 1. Thus f (z) = a0 and this proves the theorem.

Page 154 / 222


10 THE FUNDAMENTAL THEOREM OF ALGEBRA

Corollary 278 (Casorati-Weierstrass). If f is a non-constant entire function, then its image is dense in C.

Proof. If the image of f is not dense, then there is a complex number w and a real number r > 0 such that the open
1
disk centered at w with radius r has no element of the image of f . Define g(z) = f (z)−w . Then g is a bounded entire
function, since
1 1
∀z ∈ C, |g(z)| = <
|f (z) − w| r
So, g is constant, and therefore f is constant.

Corollary 279. If p(z) ∈ C[z] is a monic polynomial with degree n ≥ 1 and p(0) 6= 0, it has a complex root.

Proof. Since |p(z)| → +∞ as |z| → +∞, there is some closed disk D centered at the origin such that |p(z)| > |p(0)|
for any z outside D. Assuming that p(z) is non-vanishing, it follows that minz∈C |p(z)| is attained at some z0 ∈ D
1
and q(z) = p(z) is an entire function such that

1
∀z ∈ C, |q(z)| ≤ .
|p(z0 )|
By Liouville’s Theorem we get that both q and p are constant functions, leading to a contradiction.

A shortened proof. Assume that p(z) = z n + an−1 z n−1 + . . . + a0 ∈ C[z] with n ≥ 1 is non-vanishing over C.
By the residue Theorem, for any r > 0 we have that
I
dz 2πi
= 6= 0
|z|=r z p(z) p(0)

but the limit of the LHS as r → +∞ is clearly 0.

Theorem 280 (Open mapping Theorem). Any non-constant holomorphic function on C is an open map,
i.e. it sends open subsets of C to open subsets of C.

Proof. Assume f : U → C is a non-constant holomorphic function and U is a domain of the complex plane. We have
to show that every point in f (U ) is an interior point of f (U ), i.e. that every point in f (U ) has a neighbourhood (open
disk) which is also in f (U ). Consider an arbitrary w0 in f (U ). Then there exists a point z0 in U such that w0 = f (z0 ).
Since U is open, we can find d > 0 such that the closed disk B around z0 with radius d is fully contained in U .
Consider the function g(z) = f (z) − w0 . Note that z0 is a root of the function. We know that g(z) is not constant and
holomorphic. The roots of g are isolated by the identity theorem, and by further decreasing the radius of the image
disk d, we can assure that g(z) has only a single root in B (although this single root may have multiplicity greater
than 1). The boundary of B is a circle and hence a compact set, on which |g(z)| is a positive continuous function, so
the extreme value Theorem guarantees the existence of a positive minimum e, that is, e is the minimum of |g(z)| for
z on the boundary of B and e > 0. Denote by D the open disk around w0 with radius e. By Rouché’s theorem, the
def
function g(z) = f (z) − w0 will have the same number of roots (counted with multiplicity) in B as h(z) = f (z) − w1 for
any w1 in D. This is because h(z) = g(z) + (w0 − w1 ), and for z on the boundary of B, |g(z)| ≥ e ≥ |w0 − w1 |. Thus,
for every w1 in D, there exists at least one z1 in B such that f (z1 ) = w1 . This means that the disk D is contained in
f (B). The image of the ball B, f (B), is a subset of the image of U , f (U ). Thus w0 is an interior point of f (U ). Since
w0 was arbitrary in f (U ) we know that f (U ) is open. Since U was arbitrary, the function f is open.

Lemma 281. Any non-constant polynomial p(z) ∈ C[z] is a closed map.

Proof. We have that |p(z)| → +∞ as |z| → +∞. Suppose that p(zk ) → w ∈ C as k → +∞: then {zk } is bounded,
so taking a subsequence if necessary, there is z ∈ C such that zk → z. By continuity p(zk ) → p(z), concluding that
w = p(z).

Page 155 / 222


Corollary 282. If p(z) ∈ C[z] is a non-constant polynomial, p(C) is unbounded and simultaneously open and closed.
If follows that p(C) = C, i.e. any non-constant polynomial with complex coefficients is surjective. In particular there
is at least a complex solution of p(z) = 0.

Exercise 283. Given a non-constant polynomial p(z) ∈ C[z] such that p(0) 6= 0, prove that the following statements
are equivalent forms of the Fundamental Theorem of Algebra:

1. The companion matrix of p has at least an eigenvector in Cn ;


1
2. q(z) = p(z) is an analytic function in a neighbourhood of the origin with a finite radius of convergence;

3. If p is the characteristic polynomial of a recurrent sequence {an }n≥0 , for some M ∈ R+


the following limit does not exist:
Xn
lim ak M k .
n→+∞
k=0

An interesting part of Complex Analysis is related to the problem of extending Rolle’s Theorem (if f is a differentiable
function on [a, b] and f (a) = f (b) = 0 holds, there is some ξ ∈ (a, b) such that f 0 (ξ) = 0) to the complex case.

Theorem 284 (Gauss-Lucas). If ζ1 , . . . , ζn ∈ C (n ≥ 3) are the roots of a polynomial p(z) ∈ C[z],


all the roots of p0 (z) lie inside the convex hull of ζ1 , . . . , ζn .

Proof. We may assume without loss of generality that p(z) is a monic polynomial with simple roots. Then by
considering the logarithmic derivative of p(z) we have the following identity
n
X 1
p0 (z) = p(z)
z − ζk
k=1
Pn
and p0 (z) vanishes iff s(z) = k=1 z−ζ
1
k
vanishes, since p(z) and p0 (z) have no common root. Let us assume that s(z)
vanishes at a point w lying outside the convex hull of ζ1 , . . . , ζk , or on its boundary. By the Hann-Banach theorem
there is some line ` through the origin such that all the vectors w − ζ1 , . . . , w − ζk lie on the same side of `. By
1 1
conjugation, the same applies to the vectors w−ζ 1
, . . . , w−ζ k
, hence for some θ ∈ R the complex number eiθ s(w) has
a non-zero real/imaginary part. s(w) 6= 0 leads to a contradiction, completing the proof.

About cubic polynomials, a remarkable and way sharper result is well-known:

Theorem 285 (Marden). If A, B, C are three distinct points in the


complex plane and we denote as D, E the roots of
d
(z − A)(z − B)(z − C),
dz
then D, E are the foci of the Steiner inellipse of ABC, centered at
A+B+C
3 and tangent to the sides of ABC at their midpoints.

What is the best possible improvement of the Gauss-Lucas Theorem still is an open problem in the general case.
The following result has only been proved for polynomials having degree ≤ 8 and for some other special cases:

Page 156 / 222


10 THE FUNDAMENTAL THEOREM OF ALGEBRA

The Ilieff-Sendov conjecture. If all the zeros of a polynomial p(z) lie in kzk ≤ 1 and if r is a zero of p(z),
then there is a zero of p0 (z) in the disk kz − rk ≤ 1.

Exercise 286. By exploiting the Gauss-Lucas theorem, show that the following entire functions only have real zeroes:
√ √ X (−1)n z 2n+1 X (−z 2 )n
sin(z) + 2 sin(z 2), Si(z) = , J0 (z) = .
(2n + 1) · (2n + 1)! 4n n!2
n≥0 n≥0

Exercise 287. By exploiting the Gauss-Lucas theorem, show that for any λ ∈ [1, +∞) all the solutions of
cot(x) + λx = 0 are real numbers.

Yet another way for approaching Calculus. The exponential function is usually introduced by proving that
n def n
limn→+∞ 1 + n1 exists, then showing that ex = limn→+∞ 1 + nx is a differentiable function, then noticing
d x
that dx e = ex leads to very-known Taylor series and to De Moivre’s formula. Here we outline a different way for
approaching the early stages of Calculus.

1. One may directly introduce the complex exponential function through an everywhere-convergent power
series,
def
X zn
∀z ∈ C, ez = ,
n!
n≥0

2. then check through the convolution machinery that such function fulfills ea · eb = ea+b for any a, b ∈ C;

3. In particular, for any θ ∈ R we have that eiθ ∈ S 1 , since |eiθ |2 = eiθ · e−iθ = 1

4. and the map γ : R → S 1 given by γ(θ) = eiθ is a parametrization of S 1 with constant speed, since by the
d iθ
series definition dθ e = ieiθ and the last quantity has unit modulus by the previous point (Pythagorean
Theorem);

5. Since γ is an arc-length parametrization of S 1 , we may define sin(θ) and cos(θ) through


def def
sin(θ) = Im eiθ , cos(θ) = Re eiθ

and derive the addition formulas for sin and cos from the point (2.);

6. Since γ travels S 1 counter-clockwise, by defining π as


def
π = inf θ ∈ R+ : sin(θ) = 0


we get that π equals half the length of the unit circle, or, equivalently, the area of the unit circle.
Additionally, eiπ + 1 = 0;

7. By the previous points ez is an entire function and a solution of the differential equation f 0 (z) = f (z);

8. By the series definition it also follows that


X (−1)n X (−1)n
sin(z) = z 2n+1 , cos(z) = z 2n
(2n + 1)! (2n)!
n≥0 n≥0

are entire functions and solutions of the differential equation f 00 (z) + f (z) = 0;

Page 157 / 222


9. From the Pythagorean Theorem sin2 θ + cos2 θ = 1, hence

Z 1 √ Z 1 Z 1/ 2
dx x7→ u du symmetry dx
π=2 √ = p = 4 √
0 1 − x2 0 u(1 − u) 0 1 − x2

and by integrating termwise the Taylor series of √ 1 we get the following series representation for π:
1−x2

√ X 2n

n
π=2 2 = 3.1415926535897932384626433832795 . . .
8n (2n + 1)
n≥0


As an alternative, the integral of 1 − x2 over some sub-interval of [−1, 1] is clearly related to the area of
a circle sector. By computing a Taylor series and applying termwise integration again, we may easily derive
Newton’s identity
2n

X
n
π =4−4 .
(4n2 − 1)4n
n≥1

The Lagrange inversion theorem in a nutshell.


Let us assume to have a holomorphic function which is z + o(z) in a neighbourhood of the origin, like
X (−1)n z 2n+1
sin(z) = (1)
(2n + 1)!
n≥0

and to want to compute the coefficients of the Maclaurin series of its inverse function arcsin(z), say the coefficient
of z 7 . By Cauchy’s integral formula
I
7 1 arcsin(z)
[z ] arcsin(z) = dz (2)