Number Theory

and
Combinatorics
B SURY

Stat-Math Unit, Indian Statistical Institute, Bengaluru, India.

All rights reserved. No part of this publication may be reproduced, stored
in a retrieval system or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise, without
prior permission of the publisher.

c Indian Academy of Sciences 2017

Reproduced from Resonance–journal of science education
Reformatted by TNQ Books and Journals Pvt Ltd, www.tnq.co.in
Published by Indian Academy of Sciences
Foreword

The Masterclass series of eBooks brings together pedagogical articles on

single broad topics taken from Resonance, the Journal of Science Educa-
tion, that has been published monthly by the Indian Academy of Sciences
since January 1996. Primarily directed at students and teachers at the un-
dergraduate level, the journal has brought out a wide spectrum of articles
in a range of scientific disciplines. Articles in the journal are written in a
style that makes them accessible to readers from diverse backgrounds, and
in addition, they provide a useful source of instruction that is not always
available in textbooks.

The third book in the series, ‘Number Theory and Combinatorics’, is by

Prof. B Sury. A celebrated mathematician, Prof. Sury’s career has largely
been at the Tata Institute of Fundamental Research, Mumbai’, and the
Indian Statistical Institute, Bengaluru, where he is presently professor. He
has contributed pedagogical articles regularly to Resonance, and his arti-
cles on Number Theory and Combinatorics comprise the present book. He
has also served for many years on the editorial board of Resonance.

Prof. Sury has contributed significantly to research in the areas of linear al-
gebraic groups over global and local fields, Diophantine equations, division
algebras, central extensions of p-adic groups, applications of density theo-
rems in number theory, K-theory of Chevalley groups, combinatorial num-
ber theory, and generation of matrix groups over rings. The book, which
will be available in digital format, and will be housed as always on the
Academy website, will be valuable to both students and experts as a useful
handbook on Number Theory and Combinatorics.

Amitabh Joshi
Editor of Publications
Indian Academy of Sciences
August 2017

iii
About the Author

B Sury is a Professor of Mathematics at the Indian Statistical Institute in

Bangalore since 1999. He was earlier at the Tata Institute of Fundamental
Research, Mumbai, where he also got his PhD. Sury’s professional interests
have been very diverse: from the theory of algebraic groups and arithmetic
groups, to algebraic K-theory, and number theory. He has contributed to
these areas both through research papers and also through books.
Sury enjoys thinking about mathematical problems at all levels, and has
taken keen interest in promoting problem solving skills. As a result of his
abilities and willingness to shoulder responsibilities both in training stu-
dents, and in writing exposition for students, he is universally sought after
for any student activity in the country: he has been associated with sev-
eral student magazines in the country, perhaps the important of them all,
‘Resonance’ for the last two decades, as well the Ramanujan Math Soci-
eties Newsletter. He has been an important member of the Mathematical
Olympiad Program of the country. Sury is known to friends and colleagues
for his wit and humor, which seems to come almost instantaneously; one
can enjoy some of his limericks on his webpage at ISI. The present volume
brings together some of the writings of B Sury on Number Theory and
Combinatorics which have appeared in ‘Resonance’ during the last two
decades. Each of the articles is a masterpiece! I am sure it will be a feast
for the intellect of any reader with some mathematical inclination.

Dipendra Prasad
TIFR, Mumbai

v
Contents

1 Cyclotomy and Cyclotomic Polynomials: The Story of how Gauss 1

Narrowly Missed Becoming a Philologist
2 Polynomials with Integer Values 25
3 How Far Apart are Primes? Bertrand’s Postulate 37
4 Sums of Powers, Bernoulli and the Riemann Zeta function 47
5 Frobenius and His Density Theorem for Primes 55
6 When is a Decimal Expansion Irrational? 63
7 Revisiting Kummer’s and Legendre’s Formulae 65
8 Bessels Contain Continued Fractions of Progressions 69
9 The Prime Ordeal 75
10 Extending Given Digits to Make Primes or Perfect Powers 87
11 An Irrational Walk and Why 1 is Not Congruent 93
12 Covering the Integers 99
13 S Chowla and S S Pillai: The Story of Two Peerless Indian 103
Mathematicians
14 Multi-variable Chinese Remainder Theorem 127
15 Which Positive Integers are Interesting? 135
16 Counting, Recounting and Matching 149
17 Odd if it isn’t an Even Fit! Lighting up Tiling 161
18 Polya’s One Theorem with 100 pages of Applications 169

vii
Preface

Over the last two decades, my expositions in Resonance are on roughly

two topics – (i) number theory and combinatorics, and (ii) group theory. In
this volume, some of the expositions related to the former topic have been
put together. The chapter on the work of Chowla and Pillai is part of an
article written in collaboration with R Thangadurai that appeared in Reso-
nance. I would like to thank Thangadurai for allowing me to include it here.
I have attempted to retain the original write-ups of the articles as much as
possible other than carrying out some corrections and (minor) additions.
As this is a compendium of articles that appeared in Resonance, at times
some material is repeated in different chapters. Further, this makes the
compilation somewhat uneven in terms of levels of mathematical maturity
required of the reader despite my efforts to ensure uniformity.
Concerning my philosophy of mathematics dissemination, I have
staunchly held the opinion that many topics in mathematics which are
supposedly advanced in nature, can be exposed in a manner understand-
able to a motivated undergraduate or postgraduate student. It may be
somewhat of a challenge to make it comprehensible while still retaining the
essential depth and technicality of the content, but it should nevertheless
be possible. Over the years, I have tried to put this in practice. One of the
best compliments I have received from Professor K R Parthasarathy, who
told a visitor that he had learnt some number theory through these articles.
Despite this obvious exaggeration (or rather because of this perhaps!), I was
encouraged to keep trying to share some of the beautiful number theoretic
and combinatorial ideas I myself enjoyed learning. To my surprise, I have
occasionally found references to some of these articles in university courses
in various places, and that has been fulfilling. The years 1981-1999 at the
Tata Institute of Fundamental Research – where I first saw what excellence
means – have had a positive effect on me in terms of being able to appreci-
ate good mathematics (irrespective of whether I could originally contribute
to it or not). During those years, it was a lot of fun to learn over conversa-
tions near the sea-side, at the coffee-table, and during ping-pong sessions
from Madhav Nori, Venkataramana, Dipendra Prasad, C S Rajan, Ravi
Rao, Raja Sridharan, Amit Roy, R R Simha, Bhatwadekar, Subramaniam,
Parameswaran, Nitin Nitsure, Kapil Paranjape, Stephen Lobo, Kumare-
san, and many others. In most instances, I would excitedly come up to
them with something I had “seen”, and they would be willing listeners and
teachers. In later years, I have always benefited by talking to students and
realizing time and again what they found simultaneously enjoyable as well

ix
as not easy to understand. In almost all cases, after a meeting or discussion
with some students, I felt compelled to write on a certain topic. Inciden-
tally, some of the summer projects that students worked on have appeared
in Resonance as well and, many of them were rewritten by me – showing
my desire to communicate in a particular way. I personally know several
colleagues who are much better at writing such expositions but, not many
of them seem to find the time to write at this level. I wish they would.
Finally, I am indebted to Professor Ram Ramaswamy for suggesting the
publication of this volume, for convincing me that it could be useful, and
for pushing me to finish this process.

B Sury
July 2017

x
Cyclotomy and Cyclotomic Polynomials:
The Story of how Gauss Narrowly
Missed Becoming a Philologist

A person who does arouse

much admiration and a million wows!
Such a mathematicians’ prince
has not been born since.
I talk of Carl Friedrich Gauss!
Cyclotomy – literally ‘circle-cutting’ – was a puzzle begun more than
2000 years ago by the Greek geometers. In this pastime, they used two
implements – a ruler to draw straight lines and a pair of compasses to draw
circles. The problem of cyclotomy was to divide the circumference of a circle
into n equal parts using only these two implements.
As these n points on the circle are also the corners of a regular n-gon,
the problem of cyclotomy is equivalent to the problem of constructing the
regular n-gon using only a ruler and a pair of compasses. Euclid’s school
constructed the equilateral triangle, the square, the regular pentagon and
the regular hexagon. For more than 2000 years mathematicians had been
unanimous in their view that for no prime p bigger than 5 can the p-gon
be constructed by ruler and compasses. The teenager Carl Friedrich Gauss
proved a month before he was 19 that the regular 17-gon is constructible.
He did not stop there but went ahead to completely characterise all those
n for which the regular n-gon is constructible! This achievement of Gauss
is one of the most surprising discoveries in mathematics. This feat was
responsible for Gauss dedicating his life to the study of mathematics instead
of philology1 in which too he was equally proficient.
In his mathematical diary2 maintained from 1796 to 1814, he made his
first entry on the 30th of March and announced the construction of the
regular 17-gon. It is said that he was so proud of this discovery that he
requested that the regular 17-gon be engraved on his tombstone! This wish
was, however, not carried out.
An amusing story alludes to Kästner, one of his teachers at the university
of Göttingen, and an amateur poet. When Gauss told him of his discovery,
Kästner was skeptical and did not take him seriously. Gauss insisted that
The chapter is a modified version of an article that first appeared in Resonance, Vol. 4, No. 12,
pp. 41–53, December 1999.
1 By 18, Gauss was already an expert in Greek, Latin, French and German. At the age of 62, he

took up the study of Russian and read Pushkin in the original.

2 The diary was found only in 1898!

1
 Cyclotomy and Cyclotomic Polynomials

he could prove his result by reducing to smaller degree equations and,

being fond of calculations, also showed the co-ordinates of the 17 points
computed to several decimal places. Kästner is said to have claimed that
he already knew such approximations long before. In retaliation, some time
later, Gauss described Kästner as the best poet among mathematicians and
the best mathematician among poets!
Gauss’s proof was not only instrumental in making up his mind to take up
mathematics as a career but it is also the first instance when a mathematical
problem from one domain was rephrased in another domain and solved
successfully. In this instance, the geometrical problem of cyclotomy was
reset in algebraic terms and solved. So, let us see more in detail what
cyclotomy is all about.
The unit circle is given to us3 and we would like to divide it into n equal
parts using only the ruler and compasses. It should be noted that the ruler
can be used only to draw a line joining two given points and not for measuring
lengths. For this reason, one sometimes uses the word straightedge instead of
a ruler.
If we view the plane as the complex plane, the unit circle has the equation
z = eiθ . Since arc length is proportional to the angle subtended, the n com-
plex numbers e2πik/n , 1 ≤ k ≤ n cut the circumference into n equal parts.
As we might fix a diameter to be on the x-axis, the problem may also be
variously posed as the problem of using only the ruler and the compasses to:
(i) find the roots of z n = 1, or
(ii) construct the angle 2π/n.
Since, by coordinate geometry, a line and a circle have equations, respec-
tively, of the form ax + by = c and (x − s)2 + (y − t)2 = r2 , their points of
intersection (if any) are the common roots. Eliminating one of x, y leads to a
quadratic equation for the other. Therefore, the use of ruler and compasses
amounts in algebraic terms to solving a chain of quadratic equations.
Before we go further, we need to clarify one point. On the one hand, we
seem to be talking of constructing lengths and, on the other hand, we seem
to want to mark off certain specific points on the plane corresponding to
the vertices of a regular n-gon. To remove any confusion due to this, let us
explain how these are equivalent.

1. Some Easy Constructions Possible with a Ruler and a Pair

of Compasses
1. Drop a perpendicular on a given line l from a point P outside it
(Figure 1).
3 It is understood that the centre is also given.

2
 Cyclotomy and Cyclotomic Polynomials

l
B

Draw a circle centred at P cutting l at A and B. Draw circles centred at

A and B having radii AP and BP , respectively. The latter circles intersect
at P and Q and P Q is perpendicular to l.
2. Draw a perpendicular to a given line l through a point P on it
(Figure 2).
C

P l
B
A

Draw any circle centred at P intersecting l at A and B. Then, the circles

centred at A and B with the common radius AB intersect at two points C
and D. Then, CD passes through P and is perpendicular to l.
3. Draw a line parallel to a given line through a point outside.
This follows by doing the above constructions in succession.
4. Bisect a given segment AB.
This is obvious from Figure 2. The circles centred at A and B having
the common radius AB intersect at two points. The line joining these two
points is the perpendicular bisector.
Marking the point (a, b) on the plane is, by these observations, equivalent
to the construction of the lengths |a| and |b|. Further, one can view the same
as the construction of the complex number a + ib.
5. If a and b are constructed real numbers, then the roots of the polynomial
x2 − ax + b = 0 are constructible as well.
Actually, in the discussion of cyclotomy, we will need to deal only with
the case when the roots of such a quadratic equation are real. In this case
(Figure 3), draw the circle with the segment joining the points (0, 1) and
(a, b) as its diameter.
√
The points of intersection of this circle with tile x-axis
a± a2 −4b
are the roots 2 of the given quadratic equation x2 − ax + b = 0.

3
 Cyclotomy and Cyclotomic Polynomials

Even when the roots of x2 − ax + b = 0 are not real, they can √be con-
2
structed easily. In this case, we need to construct the points ( a2 , ± 4b−a
2 ).
But, as we

(0,1)

(0,0) x-axis
( a–√a2 –4b ,0) ( a+√a2 –4b ,0)
2 2

(a,b)
y-axis

observed earlier, we can drop perpendiculars

√ and it suffices to construct the
absolute value of these roots which is b. This is accomplished by drawing
√ (0, −b) as diameter and noting
the circle with the segment joining (0, 1) and
that it meets the x-axis at the points (± b, 0).
With this renewed knowledge, let us return to cyclotomy.
For n = 2, one needs to draw a diameter and this is evidently achieved
by the ruler.
For n = 3, the equation z 3 − 1 = 0 reduces to the
√ equations z − l = 0 or
2 −1±i 3
z + z + l = O. The roots of the latter are 2 . So, we have to bisect
the segment [−1, 0] and the points of intersection of this bisector with the
unit circle are the points we want to mark off on the circle.
For n = 4, again only bisection (of the x-axis) is involved. This already
demonstrates clearly that if the regular n-gon can be constructed, then so
can the 2r n-gon be for any r. In particular, the 2r -gons are constructible.
To construct the regular pentagon, one has to construct the roots of z 5 −1
= 0. These are the 5-th roots of unity ζ k ; i ≤ k ≤ 5 where ζ = e2iπ/5 . Now
the sum of the roots of z 5 − 1 = 0 is

0 = ζ + ζ 2 + ζ 3 + ζ 4 + ζ 5.

On using this and the fact that ζ 5 = 1, we get

(ζ 2 + ζ 3 )(ζ + ζ 4 ) = ζ + ζ 2 + ζ 3 + ζ 4 = −1.

On the other hand, we also have their sum

(ζ 2 + ζ 3 ) + (ζ + ζ 4 ) = −1.

4
 Cyclotomy and Cyclotomic Polynomials

This means that ζ 2 +ζ 3 and ζ +ζ 4 are the two roots of the quadratic poly-
nomial T 2 +√
T − 1 = 0. Thus, ζ + ζ (being positive, equal to 2Cos(2π/5))
−1+ 5
equals 2 . Multiplying this equality by ζ and using ζ 5 = 1, one gets a
quadratic equation for ζ!
This is the algebraic reasoning behind the construction. Following it, we
can geometrically make the construction also with the aid of the dictionary
between algebra and geometry that we have established above.

(– 12 ,√5
2 )

(–1,0) (1,0)

(– 12 , – √5
2 )

Let us now turn to the construction of the 17-gon. There are many ways
of doing it – [4], [5] contain explicit geometric algorithms; Gauss’s own
construction appears in [6], Art.365 – and all of them succeed essentially
because 17−1 is a power of 2! This is the reason why the degree 16 equation
x17 −1 16 15
x−1 = x +x +· · ·+x+1 = 0 reduces to a chain of quadratic equations.
It would be ideal to use the language of Galois theory (see Resonance,
Vol. 4, No. 10, 1999) to discuss the constructibility or non-constructibility of
a regular polygon. However, we will keep the discussion elementary and will
only make a few remarks for the reader familiar with basic Galois theory
so that she/he can grasp the conceptual reason behind various explicit
expressions, the appearance of which will seem magical without the added
understanding4 provided by Galois theory.
In the light of our dictionary, we describe a construction as follows:
Denote by ζ, the 17-th root of unity e2iπ/17 . Then, ζ 17 = 1 gives
ζ 16 + ζ 15 + · · · + ζ + 1 = 0.
That is,
(ζ + ζ −1 ) + (ζ + ζ −2 ) + · · · + (ζ 8 + ζ −8 ) = −1.
4 Perhaps
an outstanding feature of mathematics is that knowing the conceptual reason behind
a phenomenon is often much more important than a proof of the phenomenon itself.

5
 Cyclotomy and Cyclotomic Polynomials

Let us write α1 , α2 , α3 for the three real numbers

ζ 3 + ζ −3 + ζ 5 + ζ −5 + ζ 6 + ζ −6 + ζ 7 + ζ −7 ,
ζ 3 + ζ −3 + ζ 5 + ζ −5 ,
and
ζ + ζ −1 ,
respectively. Look at the sequence of four quadratic equations:
T 2 − α3 T + 1 = 0;
α3 − 6α2 − 2α1 + 1
T2 − 2 T + α2 = 0;
2
T 2 − α1 T − 1 = 0;
T 2 − T − 4 = 0.
A routine calculation shows that ζ, α3 , α2 , α1 are roots of the four equa-
tions in that order. The roots of the last equation are evidently constructible
as it has integer coefficients. This means that one can construct ζ by recur-
sively constructing the roots of these equations using a ruler and a pair of
compasses. The reader may geometrically make the construction based on
the dictionary above or consult one of the references quoted.
The reader familiar with basic Galois theory would recall that the four
successive quadratic extension fields generated by the polynomials give a
tower of corresponding Galois groups. The Galois group of the cyclotomic
field5 Q(ζ) over Q is a cyclic group of order 16 generated by the automor-
phism
σ : ζ 7→ ζ 3
(in other words, 3 is what is known as a primitive root modulo 17; in other
words, 16 is the smallest positive integer such that 316 − 1 is divisible by
17). The automorphisms σ 2 (that is, σ applied twice in succession), σ 4 , σ 8
fix α1 , α2 , α3 respectively.
Now, the question remains as to what made this work and which other n-
gons are constructible. Look at the regular n-gon for some n. To construct
it, one needs to mark off the complex number ζ = e2iπ/n . The question is
whether ζ can be expressed in terms of a nested chain of square roots.
For example, for the 17-gon, one gets
√ p √
−1 17 34 − 2 17
Cos(2π/17) = + +
16 r 16 16
√ √ √
q q
1
+ 17 + 3 17 − 34 − 2 17 − 2 34 + 2 17.
8
5 This consists of all polynomial expressions in ζ with rational coefficients.

6
 Cyclotomy and Cyclotomic Polynomials

Thus, sin(2π/17) can also be expressed similarly.

Of course, ζ is a root of the polynomial z n − 1, but it is a root of an
equation of smaller degree. What is the smallest degree equation of which
ζ is a root? By the division algorithm, there is such a unique monic polyno-
mial (that is, a polynomial with top coefficient 1) which divides any other
polynomial of which ζ is a root. This polynomial is called a cyclotomic poly-
nomial. The degree of this polynomial is of paramount importance because
if it is a power of two, we know from our earlier discussion that ζ can be
constructed. The cyclotomic polynomials are useful in many ways and have
several interesting properties some of which will be discussed in the last
three sections.
For a prime number p such that p − 1 is a power of 2, our discussion
shows that the regular p-gon is constructible. Such a prime p is necessarily
n
of the form 22 + 1 since 2odd + 1 is always a multiple of 3. Fermat thought
n
that the numbers 22 + 1 are primes for all n. However, the only primes of
5
this form found until now are 3, 5, 17, 257 and 65537(!). The number 22 + 1
was shown by Euler to have 641 as a proper factor. The primes of the form
n
22 + 1 are called Fermat primes. For coprime numbers m and n, if the
m-gon and the n-gon are constructible, then so is the mn-gon. The reason
is, if we write ma + nb = 1 for integers a, b, then
   
2π 2πb 2πa
Cos = Cos +
mn m n

which is constructible when Cos 2πb and Cos 2πa

m n are. Thus, Gauss’s
analysis shows that if n is a product of Fermat primes and a power of
2, the regular n-gon can be constructed by a ruler and compasses. The
converse is also true; that is, if the regular n-gon is constructible, then n is
of this special form. Gauss did not give a proof of this although he asserted
it to be true – see [5]. The construction of the regular 257-gon was published
in four parts in Crelle’s Journal. Details of the 65537-gon fill a whole trunk
kept at the University of Göttingen!
We must see Gauss’s feat in the light of the fact that complex analysis was
in its infancy at that time. In fact, Gauss was the first one to give a rigorous
proof (in his doctoral thesis) of the so-called fundamental theorem of alge-
bra which asserts that every nonconstant complex polynomial has a root.

2. Abel’s Theorem for the Lemniscate

Abel earned fame by proving that the general equation of degree at least
five is not solvable by a ‘formula’ involving only arithmetical operations and
extraction of square roots, cube roots and higher roots of the coefficients
– in other words, there is no such formula which can be prescribed so that

7
 Cyclotomy and Cyclotomic Polynomials

when we apply the formula to any collection of coefficients, we obtain the

roots of the corresponding polynomial equation. One of his lesser-known
achievements involves a problem analogous to cyclotomy viz., the division
of the lemniscate. The name lemniscate literally means a ribbon and comes
from its shape (Figure 5); this curve – also called the elastic curve – was
discovered by Bernoulli.

(–1,0) (1,0)

It has an equation of the form (x2 + y 2 )2 = 2

R 1x − y 2 . The total arc length
dt
of the lemniscate is given by the integral 4 0 √1−t 4
. Thus, the number ω
which is half of this integral is the analogue for the lemniscate of what π
is for the unit circle. It is approximately 2.6205 · · · · · · Gauss had already
asserted as entry 62 in his diary (see [7]) that the lemniscate is divisible into
five equal parts by a ruler and a pair of compasses. The previous two entries
show clearly that Gauss knew that the lemniscatic trigonometric functions
are doubly-periodic functions; they are called elliptic functions nowadays.
He hinted at a vast theory of his behind these functions but this work never
appeared. It was Abel who published a comprehensive treatise on elliptic
functions and in it he also looked at the problem of dividing the lemniscate
into 77, equal parts for any 77, (see [8] for a more modern discussion). He
discovered the remarkable fact that the answer is the same as for the circle!
In other words, the lemniscate can be divided into n equal parts with the
aid of a ruler and compasses if, and only if, n is a product of a power of
2 and distinct Fermat primes. The reason can again be understood using
Galois theory. In the case of the circle, the Galois group of the cyclotomic
extension is the multiplicative group of integers modulo n which are coprime
to n. This latter group is a group of order a power of 2 precisely when n is
a product as above. For the lemniscate, it turns out that one needs to know
when the unit group of Z[i]/nZ[i] is a group of order a power of 2 where
the set Z[i] of Gaussian integers consists of the complex numbers a + bi
with integers a, b. This is again a ‘ring’ – like the integers, we can add and
multiply elements.

3. Cyclotomic Polynomials
We introduced for any positive integer n, the cyclotomic polynomial
Φn (X) as the unique monic integer polynomial of least degree having ζ =
e2iπ/n as a root. What does Φn (X) look like? Obviously Φ1 (X) = X − 1

8
 Cyclotomy and Cyclotomic Polynomials

and Φ2 (X) = X + 1. Moreover, for a prime number p, Φp (X) = X p−1 +

X p−2 + · · · + X + 1. For any n, the n-th roots of unity Qn are the rcomplex
numbers e 2irπ/n n
; l ≤ r ≤ n. In other words, X − l = r=1 (X − ζ ) where
ζ = e2iπ/n . The crucial fact is that along with ζ, all the powers ζ r with r
coprime 6 to n are the roots of Φn (X)! So,
Y Y Y
Xn − 1 = (X − ζ r ) = Φd (X).
d|n (r,n)=d d|n

Here, we have denoted by (r, n) the greatest common divisor of r and

n. Note that the degree of Φn (X) is the number of positive integers r ≤ n
that are coprime to n; this is usually denoted by φ(n), and called Euler’s
totient. If we only look at the above expression, it is not clear that Φn (X)
has integer coefficients.QOne may use elementary number theory to invert
the identity X n − 1 = d|n Φd (X). This is accomplished by what is known
as the Möbius inversion formula and, yields the identity
Y
Φn (X) = (X d − 1)µ(n/d)
d|n

where the Möbius function µ(m) is defined to take the value 0, 1 or −1

according as whether m is divisible by a square, is a square-free product of
an even number of primes, or is a square-free product of an odd number
of primes. The inversion formula is a very easy and pleasant exercise in
elementary number theory. Note that from the above expression, it is not
clear that the fractional expression on the right side is indeed a polynomial
n
Q that this follows from induction on n using the expression X − 1 =
but
d|n Φd (X).
By observing this among the cyclotomic polynomials Φp (X) for prime
p and for Φn (X) for small n, an interesting feature is that the coefficients
seem to be among 0, 1 and −1. One might wonder whether this is true
about Φn (X) for any n. It turns out that Φ105 has one coefficient equal
to 2. This is not an aberration. Indeed, using some nontrivial results on
how prime numbers are distributed, one can show that every integer occurs
among the coefficients of the cyclotomic polynomials!

4. Infinitude of Primes Ending in 1

11, 31, 41, 61, 71, 101, · · · are primes – where does it stop? Are there in-
finitely many primes ending in 1? Equivalently, does the arithmetic pro-
gression {10n + 1; n ≥ 1} contain infinitely many prime numbers? Any
prime number other than 2 must obviously end in 1, 3, 7 or 9. The natural
6 These are the primitive n-th roots of unity i.e., they are not m-th roots of unity for any smaller m.

9
 Cyclotomy and Cyclotomic Polynomials

question is whether there are infinitely many of each type? The answer is
‘yes’ by a deep theorem due to Dirichlet – infinitely many primes occur in
any arithmetic progression {a + nd; n ≥ 1} with a, d coprime.
If d is a positive integer, then for the arithmetic progression {nd + 1;
n ≥ 1}, one can use cyclotomic polynomials to prove this! This is not
surprising because we have already noted in the last section that cyclotomic
polynomials are related to the way prime numbers are distributed. Let us
prove this now.
Suppose p1 , p2 , · · · , pr are prime numbers in this arithmetic progression.
We will use cyclotomic polynomials to produce another prime p in this
progression different from the above pi ’s. This would imply that there are
infinitely many primes in such a progression. We will use the simple obser-
vation that a polynomial p(X) with integer coefficients has the property
that p(m) − p(n) is an integer multiple of m − n.
Consider the number N = dp1 p2 · · · pr . Then, for any integer n, the two
values Φd (nN ) and Φd (0) differ by a multiple of N . But, Φd (0) is an integer
which is also a root of unity and must, therefore, be ±1. Moreover, as
n → ∞, the values Φd (nN ) → ∞ as well since Φd is a nonconstant monic
polynomial. In other words, n > 0, the integer Φd (nN ) has a prime factor
p. As Φd (nN ) is ±1 modulo any of the p1 , p2 , · · · , pr and modulo d, the
prime p is different from any of the pi ’s and does not divide d. One might
wonder which primes divide some value Φd (a) of a cyclotomic polynomial.
The answer is that these are precisely the primes occurring in the arithmetic
progression {nd + 1; n > 0}. To show this, we use the idea that the nonzero
integers modulo p form a group of order p − 1 under the operation of
multiplication modulo p. So, it is enough to prove that if p divides Φd (a)
for some integer a, then a has order d in this group (for, then Lagrange’s
theorem of finite group theory tells us that d divides the order p − 1 of
the group, which is just re-stating that p is in the arithmetic progression
{nd+1; n > 0}. Let us prove this now. Since X d −1 = l|d Φl (X), it follows
Q

that p which divides Φd (a) has to divide ad − 1 also. If d were not the order
divide d with k < d and p divides ak − 1. Once again, the relation
of a, let k Q
k
a − 1 = l|k Φl (a) shows that p divides Φl (a) for some positive integer l
dividing k. Therefore, p divides both Φd (a + p) and Φl (a + p). Now,
Y
(a + p)d − 1 = Φm (a + p) = Φd (a + p)Φl (a + p) (other terms).
m|d

The expression on the right hand side is divisible by p2 . On the other

hand, the left side is equal, modulo p2 , to ad + dpad−1 − 1. Since p2 divides
ad −1, it must divide dpad−1 as well. This is clearly impossible since neither
a nor d is divisible by p. This proves that any prime factor p of Φd (nN )

10
 Cyclotomy and Cyclotomic Polynomials

occurs in the arithmetic progression {1+nd; n > 0} and thereby, proves the
infinitude of the primes in this progression. Interestingly, Euclid’s classical
proof of the infinitude of prime numbers is the special case of the above
proof where we can use d = 2.

5. Sum of Primitive Roots

For a prime number p, Gauss defined a primitive root modulo p to be an
integer a whose multiplicative order modulo p is p − 1. In other words, a is
a generator of the multiplicative group of non-zero integers modulo p. More
generally, for a positive integer n, every integer a coprime to n is such that
aφ(n) is 1 modulo n. A primitive root modulo n is an integer a such that
φ(n) is the smallest r > 0 for which ar is 1 modulo n. Gauss also showed
that primitive roots modulo n exist if, and only if, n is 2, 4, pa or 2pa for
some odd prime p.
For instance, the primitive roots modulo 5 among the integers 1 to 4 are
2 and 3. Their sum is 0 modulo 5. Now, look at the primitive roots modulo
7 among 1 to 6. These are 3 and 5. Modulo 7, these sum to 1. What about
11? The primitive roots here are 2, 6, 7 and 8 and these give the sum 1
modulo 11. What is the pattern here? Without letting out the secret, let
us go on to investigate the problem for a general prime p.
When is an integer modulo p a primitive root? As we already observed,
an integer a is a primitive root modulo p precisely when p divides the
integer Φp−1 (a). This means when the polynomial Φp−1 (X) is regarded as a
polynomial with coefficients integers modulo p, a is a root. Hence the sum
of all the primitive roots modulo p is simply the sum modulo p of the roots
of Φp−1 modulo p. As we will prove below, the above sum is µ(p − 1), where
µ(n) is the Möbius function.

6. Cyclotomic Polynomials and Ramanujan Sums [9]

In his famous paper ([10]), Ramanujan discussed the properties of certain
finite sums – the so-called Ramanujan sums. Even though Dirichlet and
Dedekind had already considered these sums in the 1860’s, according to
G H Hardy, “Ramanujan was the first to appreciate the importance of
the sum and to use it systematically.” Ramanujan sums play a key role in
the proof of a famous result due to Vinogradov asserting that every large
odd number is the sum of three primes. These sums have numerous other
applications in diverse branches of mathematics as well as in some parts of
physics. So, what are these sums?
For integers n ≥ 1, k ≥ 0, the sum
X
cn (k) = e2ikrπ/n
(r,n)=1;r≤n

11
 Cyclotomy and Cyclotomic Polynomials

is called a Ramanujan sum. In other words, it is simply the sum of the k-th
powers of the primitive n-th roots of unity – ‘primitive’ here means that the
number is not an m-th root of unity for any m < n. Note that the primitive
n-th roots of unity are the numbers e2ikrπ/n for all those r ≤ n which are
relatively prime to n.
The first remarkable property cn (k) have is that they are integers. Ra-
manujan showed that several arithmetic functions (that is, functions de-
fined from the set of positive integers to the set of complex numbers) have
‘Fourier-like’ of expansions in terms of the sums; hence, nowadays these ex-
pansions are known as Ramanujan expansions. They often yield very pretty
elementary number-theoretic identities. Recently, the theory of group rep-
resentations of the permutation groups (specifically, the so-called super-
character theory has been used to re-prove old identities in a quick way
and also, to discover new identities.
It is convenient to write

∆n = {e2irπ/n : (r, n) = 1, 1 ≤ r ≤ n}.

Then, the set of all n-th roots of unity {e2ikπ/n : 0 ≤ k < n} is a union
of the disjoint sets ∆d as d varies over the divisors of n. This is because an
n-th root of unity is a primitive d-th root of unity for a unique divisor d of
n. It is also convenient to introduce the ‘characteristic’ function δk|n which
has the value 1 when k divides n and the value 0 otherwise. Before stating
some properties of the ck (n)’s, let us recall two arithmetic functions which
are ubiquitous in situations where elementary number-theoretic counting is
involved. The first one is Euler’s totient function

φ(n) = |{r : 1 ≤ r ≤ n, (r, n) = 1}|.

The other arithmetic function is the Möbius function defined by

µ(1) = 1, µ(n) = (−1)k or 0 for n > 1 according as to if n is a square-free
integer that is a product of k distinct primes or otherwise.
The Möbius function keeps tab when we use the principle of inclusion-
exclusion to do counting. The basic result which can be easily proved by
induction on the number of prime factors, is the Möbius inversion formula:
If g is an arithmetic function and
X
f (n) = g(d),
d|n

then
X
g(n) = f (d)µ(n/d).
d|n

12
 Cyclotomy and Cyclotomic Polynomials

With these notations, here are some elementary properties of the Ra-
manujan sums.
(i) cn (k) = cn (−k) = cn (n − k).
(ii) cn (0) = φ(n) and cn (1) = µ(n).
(iii) cn (ks) = cn (k) if (s, n) = 1; in particular, cn (s) = µ(n) if (s, n) = 1.
(iv) cn (k) = cn (k 0 ) if (k, n) = (k 0 , n); in particular, cn (k) ≡ cn (k 0 ) mod n
if k ≡ k 0 mod n.
P n−1
(v) Pk=0 cn (k) = 0. P P
(vi) d|n cd (k) = δn|k n and cn (k) = d|n dµ(n/d)δd|k = d|(n,k) dµ(n/d);
in particular, for prime powers pr , we have cpr (k) = pr − pr−1 if pr |k;
= −pr−1 if pr−1 ||k; and = 0 otherwise.
(vii) cmn (k) = cm (k)cn (k) if (m, n) = 1.
P n
(viii) k=1 cm (k)cn (k) = δmn nφ(n).

The property (vi) shows that these sums actually have integer values.
The proof of (i) follows already from the definition and, so do the first
parts of (ii) and (iii). The second parts of (ii), (iii) as well as the assertions
(iv) and (vii) will follow from (vi). We shall prove (v) and (vi).
For (v), we have
n−1
X n−1
X X X n−1
X
k
cn (k) = ζ = ζk = 0
k=0 k=0 ζ∈∆n ζ∈∆n k=0

where the last equality is because

n−1
X 1 − ζn
ζk = =0
1−ζ
k=0

for each ζ ∈ ∆n .
For proving (vi), we note that the second statement follows from the first
by the Möbius inversion formula. Let us prove the first one now. We have

X X X n−1
X
k
cd (k) = ζ = e2imkπ/n
d|n d|n ζ∈∆d m=0

because, as we observed, the disjoint union of ∆d as d varies over the

divisors of n is the set of all n-th roots of unity. Now, if the above sum
Pn−1 2imkπ/n
m=0 e is multiplied by e2ikπ/n , we get the same sum which means
that it is equal to 0 unless n|k. When k|n, the sum is clearly equal to n.
This proves (vi).
The other parts easily follow from (vi).

13
 Cyclotomy and Cyclotomic Polynomials

The equality cn (k) = d|n dµ(n/d)δd|k is very useful.7 For instance, if n

P
is a prime power pr , as we noted above in (vi), we have

cpr (k) = pr δpr |k − pr−1 δpr−1 |k .

Using this expression in (vii) above, we get

k
µ( (k,n) )φ(k)
ck (n) = k
.
φ( (k,n) )

The right hand side was studied by R D Von Sterneck in 1902 and is
known by his name. The equality above itself was known before Ramanujan
and is due to J C Kluyver in 1906.

7. Connection of Ramanujan Sums with Cyclotomic

Polynomials
Q
The cyclotomic polynomials Φn (x) = ζ∈∆n (x−ζ) have some fascinating
properties and have surprising consequences (see [9], where applications
such as the infinitude of primes in arithmetic progressions of the form {1 +
an} are proved). We have:
Y Y Y
xn − 1 = (x − ζ) = Φd (x)
d|n ζ∈∆d d|n

and – by Möbius inversion, we deduce

Y
Φn (x) = (xd − 1)µ(n/d) .
d|n

Taking the logarithmic derivative, we obtain

Φ0n (x) X dxd−1 µ(n/d)

= .
Φn (x) xd − 1
d|n

Multiplying by x(xn − 1), we get a polynomial in x, viz.,

Φ0n (x) X
x(xn − 1) = dµ(n/d)(xd + x2d + · · · + xn ).
Φn (x)
d|n

7 Note
that even computationally the defining sum for cn (k) requires approximately n operations
whereas the other sum requires roughly log(n) operations.

14
 Cyclotomy and Cyclotomic Polynomials

Thus, the coefficient of xk in the polynomial on the right is d|(n,k) dµ(n/d),

P
which is simply the Ramanujan sum cn (k). Hence, we have:
Proposition. For each k < n, the Ramanujan sum cn (k) is the coeffi-
Φ0n (x)
cient of xk−1 in the polynomial (xn − 1) Φn (x)
.
Note a special case of the above discussion.
The sum of the roots of Φn (X) is cn (1) = µ(n) as seen above. In partic-
ular, if n = p − 1 for a prime p, the sum of primitive roots modulo p is the
sum of the roots of the polynomial Φp−1 (X) modulo p, which is therefore
µ(p − 1). More generally, the sum of the r-th powers of the primitive roots
mod p equals the Ramanujan sum cp−1 (r).

8. Equal Sums of Powers via Cyclotomic Polynomials

One has the elementary identity

13 + 33 + · · · + n3 = (1 + 2 + · · · + n)2 .

By raising both sides to the k-th power, we have the identity

(13 + 23 + 33 + · · · + n3 )k = (1 + 2 + 3 + · · · + n)2k .
Are there other such identities? It turns out that there are no others. We
shall prove this now using cyclotomic polynomials. To be more precise, let
us set up the notation

pr (n) = 1r + 2r + · · · + nr .

for any natural numbers n, r. Let us look for natural numbers r1 < r2 <
· · · < rk and s1 < s2 < · · · < sl different from the ri ’s and, also some natural
numbers a1 , a2 , · · · , ak , b1 , b2 , · · · , bl such that, for any natural number n,
one has identities

pr1 (n)a1 pr2 (n)a2 · · · prk (n)ak = ps1 (n)b1 ps2 (n)b2 · · · psl (n)bl .

If we have such an identity, then (for n = 2),

(1 + 2r1 )a1 · · · (1 + 2rk )ak = (1 + 2s1 )b1 · · · (1 + 2sl )bl · · · (A)

Now, we look at the larger number among rk or sl , say sl . If sl ≤ 3, then

the identity can be shown to be just

p3 (n)a = p1 (n)2a ,

that is,
(13 + 23 + 33 + · · · + n3 )a = (1 + 2 + 3 + · · · + n)2a

15
 Cyclotomy and Cyclotomic Polynomials

which we have already seen. This is an easy check. Now, let us suppose
sl > 3. Below, we will prove the nice fact that any number of the form
1 + 2b with b > 3 always has a prime factor which is not a factor of any
1 + 2c for any c < b. This beautiful observation was first made by A S
Bang 120 years ago. This observation shows immediately that an equality
of the form (A) cannot hold good because the prime factor p of the largest
1 + 2sl cannot divide any term on the left hand side. Seeing why Bang’s
observation is valid requires some discussion about cyclotomic polynomials
which we proceed to do now.
Generally, if one has an infinite sequence of natural numbers u1 < u2 <
u3 < · · · such that for every n, there exists a prime factor of un which does
not divide um for every m < n usually called a primitive prime divisor of
un . From any such sequence admitting primitive prime divisors, we have a
proof of infinitude of primes because we find at least one new prime divisor
at each step as we move along the sequence of un ’s.
We show now that the sequence {2n + 1}n>3 has primitive divisors.
An advantage of knowing that the cyclotomic polynomials Φn (x) have in-
teger coefficients is the following. For any integer a and any natural number
n, one has
Y
an − 1 = Φd (a)
d|n

which is a product of integers. Thus, if p is any prime dividing an − 1 for

some a, then p divides Φd (a) for some d|n.
Now, we make the following interesting assertion :
Proposition. Let n > 2. If p is a prime dividing Φn (a) for some integer
a > 1, then p divides an −1 and in this case n is the smallest natural number
such that p divides an − 1 unless p|n in which case the smallest number is
of the form n/pi for some i ≥ 1. In the latter case, p is the largest prime
dividing n. Finally, given a > 1 and n > 2, if there are no primes p for
which a has order n mod p, then Φn (a) is actually a prime.
Let us prove this. Indeed, p divides an − 1 since Φn (a) divides an − 1.
Also, if p divides am − 1 for some m < n as well, then p divides a(m,n) − 1
where (m, n) is the GCD of m and n.
Therefore, if n were not the smallest for which p divides an − 1, we would
have a factor d of n such that d < n and p|(ad − 1). As d < n and d|n,
there is some prime q such that qd|n. Thus, d divides n/q and so p divides
an/q − 1. Writing b = an/q , we see that b leaves a remainder 1 on division
by p. So, we have
an − 1 bq − 1
= = 1 + b + b2 + · · · + bq−1
an/q − 1 b−1

16
 Cyclotomy and Cyclotomic Polynomials

leaves a remainder of q on division by p. On the other hand, the left hand

side is a multiple of Φn (a) which is a multiple of p. Thus, we must have
that p = q and that it divides n.
This also shows that an/q − 1 is not a multiple of p for any prime divisor
q 6= p of n. Thus, the order of a mod p (which means the smallest natural
number d such that p divides ad − 1) is either n or of the form n/pi for some
i ≥ 1.
To say that when the order of a mod p is < n, then the prime p (which
we have shown to be a divisor of n) is the largest prime divisor, we need to
use Fermat’s little theorem.
In the case when the order of a mod p is < n, we have seen that it is
of the form n/pi . Thus, n/pi divides p − 1 which means every other prime
divisor of n is < p. This proves the proposition except for the last assertion.
To see that the last assertion of the proposition also holds, consider n > 2,
a > 1 and a prime divisor p of Φn (a). Under the hypothesis that there are
no primes modulo which a has order n, we have seen that p is the largest
prime dividing n and that Φn (a) = pk for some k ≥ 1. We assert that k = 1.
Now, Φn (a) divides
an − 1
= 1 + an/p + a2n/p + · · · + a(p−1)n/p .
an/p − 1
As an/p = 1 + pb for some b, the right hand side above is
1 + (1 + pb) + · · · + (1 + pb)(p−1)
= p + p(b + 2b + · · · + (p − 1)b) + p2 c = p + p2 d
for some c, d if p > 2.
Therefore, p2 does not divide Φn (a); hence Φn (a) = p.
When p = 2, the argument is again easy remembering that n > 2 is a
power of 2 as p is the largest prime divisor of n.
This proves the proposition.
How does one relate what we need about primitive prime divisors for
the sequence {2n + 1}n>3 with the above discussion? For each n > 3, if we
find a prime p such that the order of 2 mod p is 2n, then from (22n − 1)
= (2n − 1)(2n + 1), we would have p|(2n + 1) because p does not divide
2n − 1. Also, if p divided 2m + 1 for some m < n, then it would divide
22m − 1 which would contradict the fact that the order of 2 mod p is 2n.
In order to get a prime p such that 2 has order 2n mod p, we need to get
a prime p dividing Φ2n (2) and not dividing 2n. If there is no such prime,
then as we saw in the discussion above, we must have that Φ2n (2) = p is
the largest prime dividing n, and p is odd. Writing 2n = pi d with d dividing
p − 1. On the other hand,
i Qφ(d) p
Φd (2p ) r=1 (b − ζr ) bp − 1 φ(d)
Φ2n (2) = = > ( ) ,
Φd (2pi−1 )
Qφ(d) b+1
r=1(b − ζr )

17
 Cyclotomy and Cyclotomic Polynomials

i−1
where b = 2p and ζr are the φ(d) primitive d-th roots of unity (the roots
of Φd (x)).
As bp − 1 ≥ bp−2 (b2 − 1), the right side above is > b(p−2)φ(d) (b − 1)φ(d) .
As b ≥ 2, this last expression is at least 2p−2 . Therefore, we have
p = Φ2n (2) > 2p−2 ,
which is possible only if p = 3. In that case we must also have 2n = 6
which we rule out. In other words, when n > 3, then there does exist
a prime divisor p of Φ2n (2) which does not divide n; the above discussion
then shows that n is the smallest natural number for which p divides 2n +1.
This finishes the whole argument.

9. Cyclotomic Polynomials and a Problem in Geometry

Let us first start with the following simple question. On the unit circle,
take n points dividing the circumference into n equal parts. From one of
these n points, draw the n−1 chords joining it to the other points. It is easy
to see that the product of the lengths of these chords is n. A more difficult
problem is to start from one of the points and – going in one direction (say,
the anticlockwise direction) – drawing the chords joining it to the k-th point
from it for each k relatively prime to n, what is the product of the lengths
of these chords in this case?
We prove:
Let n > 1 and let P1 , · · · , Pn be points on a circle of radius 1 dividing
the circumference into n equal
Q parts. Then, we have :
The product of lengths (l,n)=1,l<n |P1 Pl+1 | = p or 1 accordingly as to
whether n = pk for a prime p or n is not a power of a prime.
We may assume that the origin is the centre and that points are Pd+1
= e2idπ/n for d = 0, 1, · · · Q
, n − 1. Note that the product of lengths of all
the chords P1 Pi is simply n−1 d=1 |1 − e
2idπ/n |. Since the polynomial 1 + X +

··· + X n−1 has as roots all the n-th roots of 1 excepting 1 itself, we have
n−1
Y
(1 − e2idπ/n ) = n,
d=1

by evaluating at X = 1. Notice that we have the equality n−1 2idπ/n )

Q
d=1 (1 − e
= n as complex numbers; that is, even without considering absolute values.
Now, let us consider our problem. Here, the product under consideration
is Y
|1 − e2idπ/n |.
(d,n)=1

18
 Cyclotomy and Cyclotomic Polynomials

First, let us look at the case when n = pk for some prime p. Then,
Qpk −1 k
Y
2idπ/pk |1 − e2idπ/p | pk
|1 − e | = Q d=1 k = = p.
dp<pk |1 − e
2idpπ/p | pk−1
(d,pk )=1,d<pk

Now, suppose that n has at least Q two prime factors.

Let us start with the identity n−1 d=1 (1 − e 2idπ/n ) = n.

If p is a prime dividing
Qn−1 n, suppose pk is the highest power of p dividing n.
Then, the product d=1 (1 − e2idπ/n ) contains the products of terms corre-
Qpk −1 k
sponding to d running through multiples of n/pk ; that is, d=1 (1−e2idπ/p )
(which is pk ). We observe that factors occurring for a different prime q
dividing n are disjoint from those occurring corresponding to p. There-
fore,
Q thek factors corresponding to the various primes dividing n contribute
pk ||n p = n.
On removing these factors corresponding to each prime divisor of n, we
will get d∈D (1 − e2idπ/n ) = 1, where D consists of those d for which
Q

e2idπ/n does not have prime power order. Thus, if d ∈ D, then 1 − e2idπ/n
is a unit since n is not a prime power. Therefore, 1 − e2iπ/n is a unit in the
cyclotomic field Q(e2iπ/n ). From Galois theory, we have that the product
2idπ/n ) is the norm of 1 − e2iπ/n from Q(e2iπ/n ) to Q. As this
Q
(d,n)=1 (1 − e
element is a unit, this product is ±1. Hence we get (d,n)=1 |1−e2idπ/n | = 1
Q
which proves our assertion in the case when n is not a prime power. The
proof is complete.
In the above proof, the second part can also be deduced from the first
part of the proof inQa different fashion as follows.
Writing P (n) = n−1 l d
Q
l=1 (1 − ζ ) and Q(n) = (d,n)=1 (1 − ζ ), where ζ =
e2iπ/n , we can see that
Y
P (n) = Q(r).
r|n

By Möbius inversion, Q(n) = d|n P (d)µ(n/d) = d|n dµ(n/d) by the sim-

Q Q
pler first assertion observed at the beginning of the pr of the proposition.
The function
X
log Q(n) = µ(n/d) log(d),
d|n

can be identified with the so-called von Mangoldt function Λ(n) which
is defined to have the value log(p) if n is a power of p and 0 otherwise.
Using this identification, exponentiation gives also the value asserted in the
proposition; viz., Q(n) = p or 1 according as to whether n is a power of a
prime p or not.

19
 Cyclotomy and Cyclotomic Polynomials

vp (n)
P Q
To see why Λ(n) = d|n µ(n/d) log(d), we write n = p|n p and note
that
X
log(n) = vp (n) log(p)
p|n
P
But, the right hand side is clearly d|n Λ(d). Hence, Möbius inversion
yields
X
Λ(n) = log(d)µ(n/d).
d|n

We shall use the solution of the above elementary geometric problem

(obtained using cyclotomic polynomials) to unearth interesting information
about the so-called cyclotomic field Q(ζn ) which consists of polynomial
expressions in ζn = e2iπ/n . This section requires a bit of background in basic
field theory. It also implies by the so-called Dedekind–Kummer criterion,
the well-known fact that the primes ramifying in Q(ζn ) are exactly those
which divide n.
Discriminant of Q(ζ n) . Let n > 2 be a positive integer and ζn be a
primitive n-th root of unity. Then, the discriminant of the cyclotomic field
φ(n)
is (−1)φ(n)/2 Q pnφ(n)/(p−1) .
p|n

Recall that the ring OK of algebraic integers of K = Q(ζn ) is Z[ζn ].

The minimal polynomial of ζn is the cyclotomic polynomial
Y
Φn (X) = (X − ζnr ).
(r,n)=1

Thus, the discriminant of OK is that of the polynomial

Q Φn up to sign.
The polynomial Φn has another expression Φn (X) = d|n (X d − 1)µ(n/d)
which is obtained by Möbius inversion formula to the decomposition
Y
Xn − 1 = Φd (X).
d|n

Let us prove the first result.

Since
Y Y
Φn (X) = (X d − 1)µ(n/d) = (X n − 1) (X d − 1)µ(n/d) ,
d|n d|n,d<n

we may write
Xn − 1 Y
Ψ(X) := = (X d − 1)−µ(n/d).
Φn (X)
d|n,d<n

20
 Cyclotomy and Cyclotomic Polynomials

Now, differentiating X n − 1 = Φn (X)Ψ(X) and putting X = ζn , we get

nζn−1 = Φ0n (ζn )Ψ(ζn ).
We have the discriminant d(K) = ±NK/Q Φ0n (ζn ) = ±nφ(n) NK/Q (Ψ(ζn ))−1 .
Now Ψ(ζn )−1 = d|n,d<n (ζnd −1)µ(n/d) which is convenient to write (using
Q
n/d instead of d) as:
Y
Ψ(ζn )−1 = (ζnn/d − 1)µ(d) ,
d|n,d>1

Separating the terms corresponding to µ(d) = 1 and to µ(d) = −1, we

have Q n/d
−1 d|n,d>1,µ(d)=1 (ζn − 1)
Ψ(ζn ) = Q n/d
.
d|n,d>1,µ(d)=−1 (ζn − 1)
n/d
Now, for each divisor d of n, ζn is a primitive d-th root of unity. By
n/d
proposition 1 above, 1−ζn is a unit unless d is a prime power. In the above
expression for Ψ(ζn )−1 , a nontrivial term in the denominator corresponds
to µ(d) = −1 which can happen for a prime power d only if d is prime. In
the numerator, the condition µ(d) = 1 cannot happen for any prime power
d. In other words,
Y
Ψ(ζn )−1 = (unit). (ζnn/p − 1)−1 .
p|n

n/p
− 1)−1 as units have norm ±1.
Q
So, its norm is ± p|n NK/Q (ζn
n/p
As ζn is a primitive p-th root of unity, it is in the subfield Q(ζp ) gen-
erated by a primitive p-th root of unity, and we have

NK/Q (ζnn/p − 1) = (NQ(ζp )/Q (ζp − 1))[K:Q(ζp )] = (±p)φ(n)/(p−1) .

Thus, we get
nφ(n)
d(K) = ± Q φ(n)/(p−1)
.
p|n p

Finally, it is well-known (and easy to deduce from the definition) that

for any number field L, the discriminant d(L) has sign (−1)s where s is
the number of complex places of L. Our field K = Q(ζn ) has s = φ(n)/2
because primitive n-th roots of unity are all complex.

10. Reducibility of Cyclotomic Polynomials Modulo Primes

The cyclotomic polynomial Φn is the monic, irreducible polynomial of
a primitive n-th root of unity but it may happen to be reducible modulo
certain primes. In this section, we investigate when this happens.

21
 Cyclotomy and Cyclotomic Polynomials

We recall:
For a positive integer n > 2, if disc(Φn ) is a perfect square, then Φn is
reducible modulo every prime.
This proof is a standard application of Galois theory. Indeed, it is well-
known that if the discriminant of a Galois extension is a square, its Galois
group would be contained in the subgroup of even permutations ([1], Lemma
12.3). So, if Φn were irreducible modulo some prime p, then the reduction
of Φn mod p generates over Fp a Galois extension of degree φ(n); the Galois
group would contain a φ(n)-cycle which is an odd permutation since φ(n)
is even for n
We prove:
For n > 2, the polynomial Φn is reducible modulo every prime if, and
only if, disc(Φn ) is a perfect square. If disc(Φn ) is not a perfect square –
which happens if, and only if, n = 4, pk or 2pk – then there are infinitely
many primes p such that Φn is irreducible modulo p.
Let us prove this now.
We have already seen that if disc(Φn ) is a perfect square in Z, then
Φn is reducible modulo every prime. Conversely, suppose disc(Φn ) is not a
φ(n)
perfect square. Then, looking at the expression (−1)φ(n)/2 Q pnφ(n)/(p−1) for
p|n

the discriminant, we shall deduce that n = 4, pk or 2pk for some odd prime.
Indeed, write
n = pα1 1 pα2 2 · · · pαr r .
Firstly, if n is odd and r > 1, clearly,
pαi −1 (pi − 1)
Qr
φ(n)
= i=1 i ,
2 2
is even and the power of pi dividing the discriminant is
  Y r  Y 
αi (pi − 1) − 1 pαk k −1 (pj − 1) ,
k=1 j6=i

which is even.
Thus, if n > 2 is odd, then the discriminant is a perfect square unless
n = pk .
If n = 2pα1 1 · · · pαr r for some odd primes, Φn = Φn/2 and the discriminant
is a perfect square excepting the case r = 1; i.e., n = 2pk .
Now, if n = 2α pα1 1 · · · pαr r with either α > 2 or α = 2 and r ≥ 1, then
again the powers of 2 and each pi dividing the discriminant are all even.
Thus, the exceptional case is n = 4.
Therefore, we have deduced that the expression for discriminant is a
perfect square excepting the cases n = 4, pk and 2pk for an odd prime.

22
 Cyclotomy and Cyclotomic Polynomials

These exceptional cases are when the Galois group of the cyclotomic field
is cyclic.
The Galois group of Φn over Q is a cyclic group of order φ(n) and con-
tains a φ(n)-cycle. By the Frobenius density theorem discussed in a later
chapter, there are infinitely many prime numbers l such that the decompo-
sition group at l is cyclic of order φ(n) which means that Φn modulo l is
irreducible and generates the extension of degree φ(n) over Fl . This proves
the proposition.

References
[1] P Morandi, Field and Galois theory, Graduate texts in Mathematics
167, Springer-Verlag, 1996.
[2] M Ram Murty and J Esmonde, Problems in Algebraic Number Theory,
Graduate Texts in Mathematics 190, Springer-Verlag, New York, 2005.
[3] M Artin, Algebra, Prentice Hall, 1991.
[4] D Suryaramana, Resonance, Vol. 2, No. 6, 1997.
[5] Ian Stewart, Gauss, Scientific American, 1977.
[6] C F Gauss, Disquisitiones Arithmeticae, English Edition, Springer,
1985.
[7] J Gray, English translation and commentary on Gauss’s mathematical
diary, Expo. Math., Vol. 2, 1984.
[8] M Rosen, Amer. Math. Monthly, 1981.
[9] B Sury, Ramanujan’s Awesome Sums,Mathematics Newsletter,
Vol. 24, no. 2, pp. 31–36, September 2013.
[10] S Ramanujan, On certain trigonometrical sums and their applications
in the theory of numbers, Trans. Cambridge Philos. Soc., Vol. 22,
No. 13, 259-276, 1918.

23
Polynomials with Integer Values

A quote attributed to the famous mathematician L Kronecker is ‘Die

Ganzen Zahlen hat Gott gemacht, alles andere ist Menschenwerk.’ A trans-
lation might be ‘God gave us integers and all else is man’s work.’ All of us
are familiar already from middle school with the similarities between the
set of integers and the set of all polynomials in one variable. A paradigm
of this is the Euclidean (division) algorithm. However, it requires an as-
tute observer to notice that one has to deal with polynomials with real or
rational coefficients rather than just integer coefficients for a strict anal-
ogy. There are also some apparent dissimilarities – for instance, there is no
notion among integers corresponding to the derivative of a polynomial. In
this discussion, we shall consider polynomials with integer coefficients. Of
course a complete study of this encompasses the whole subject of algebraic
number theory, one might say. Most of this article sticks to fairly elemen-
tary methods (in fact, other than the discussion on Schur’s theorem) to
address a number of rather natural questions. To give a prelude, since the
square of a polynomial with integer coefficients takes perfect square values
at all integer points, one such natural question might be “if an integral
polynomial takes only values which are perfect squares, then must it be the
square of a polynomial? ” Note that for a natural number n, the polynomial

X(X−1)···(X−n+1)
X
n
= n(n−1)···1 takes integer values at all integers although it does
not have integer coefficients.

1. Prime Values and Irreducibility

The first observation about polynomials taking integral values is:
Lemma 1.1. A polynomial P
takes integer values at all integer points if,
X X


and only if, P (X) = a0 + a1 1 + · · · + an n for some ai in Z.
Proof. The sufficiency is evident. For the converse, we first note that
any polynomial whatsoever can be written in this form for some n and
some (possibly non-integral) a0i s. Writing P in this form and assuming that
P (Z) ⊂ Z, we have
P (0) = a0 ∈ Z
P (1) = a0 + a1 ∈ Z
P (2) = a0 + a1 21 + a2 ∈ Z

and so on. Inductively, since P (m) ∈ Z∀m, we get an ∈ Z ∀ n.

The chapter is a modified version of an article that first appeared in Resonance, Vol. 6, No. 9,
pp. 46–60, September 2001.

25
 Polynomials with Integer Values

COROLLARY 1.2
If a polynomial P takes integers to integers and has degree n, then n!P (X)
∈ Z[X].
Lemma 1.3. A nonconstant integral polynomial P (X) cannot take only
prime values.
Proof. If all values are composite, then there is nothing to prove. So,
assume that P (a) = p for some integer a and prime p. Now, as P is non-
constant, lim |P (a + np)| = ∞. So, for big enough n, |P (a + np)| > p. But
n→∞
P (a + np) ≡ P (a) ≡ 0 mod p, which shows P (a + np) is composite.
Remark 1.4. Infinitely many primes can occur as integral values of a
polynomial. For example, if (a, b) = 1, then the well-known (but deep)
Dirichlet’s theorem on primes in progression shows that the polynomial
aX +b takes infinitely many prime values. In general, it may be very difficult
to decide whether a given polynomial takes infinitely many prime values.
For instance, it is not known if X 2 + 1 represents infinitely many primes.
In fact, there is no known polynomial of degree ≥ 2 which takes infinitely
many prime values.
Lemma 1.5. If P is a nonconstant polynomial that takes integers to inte-
gers, the number of prime divisors of its value set {P (m)}m∈Z , is infinite
i.e. not all terms of the sequence P (0), P (1), · · · can be built from finitely
many primes.
Proof. It is clear from the note above that it is enough to prove this for
n
ai X i
P
P (X) ∈ Z[X], which we will henceforth assume. Now, P (X) =
i=0
where n ≥ 1. If a0 = 0, then clearly P (p) ≡ 0 mod p for any prime p. If
a0 6= 0, let us consider for any integer t, the polynomial
n n
( )
X X
P (a0 tX) = ai (a0 tX)i = a0 1 + ai a0i−1 ti X i = a0 Q(X).
i=0 i=1

There exists some prime number p such that Q(m) ≡ 0 mod p for some
m and some prime p, because Q can take the values 0,1,−1 only at finitely
many points. Since Q(m) ≡ 1 mod t, we have (p, t) = 1. Then P (a0 tm) ≡
0 mod p. Since t was arbitrary the set of p arising in this manner is infinite.
Remark 1.6. (a) Note that it may be possible to construct infinitely many
terms of the sequence {P (m)}m∈Z using only a finite number of primes.
For example take (a, d) = 1, a ≥ d ≥ 1. Since, by Euler’s theorem, aϕ(d) ≡
ϕ(d)n
1 mod d, the numbers a(a d −1) ∈ Z ∀ n. For the polynomial P (X) =
dX + a, the infinitely many values P ( ad (aϕ(d)n − 1)) = aϕ(d)n+1 have only
prime factors coming from primes dividing a.

26
 Polynomials with Integer Values

(b) In order that the values of an integral polynomial P (X) be prime for
infinitely many integers, P (X) must be irreducible over Z and of content 1.
By content, we mean the greatest common divisor of the coefficients. In
general, it is difficult to decide whether a given integral polynomial is irre-
ducible or not. We note that the irreducibility of P (X) and the condition
that it have content 1, are not sufficient to ensure that P (X) takes infinitely
many prime values. For instance, the polynomial X n + 105X + 12 is irre-
ducible, by Eisenstein’s criterion (see Box 1). But, it cannot take any prime
value because it takes only even values and it does not take either of the
values ±2 since both X n + 105X + 10 and X n + 105X + 14 are irreducible,
again by Eisenstein’s criterion.
Lemma 1.7. Let a1 , · · · , an be distinct integers.
Then P (X) = (X − a1 ) · · · (X − an ) − 1 is irreducible.
Proof. Suppose, if possible, P (X) = f (X)g(X) with deg .f, deg .g < n.
Evidently, f (ai ) = −g(ai ) = ±1 ∀1 ≤ i ≤ n. Now, f (X) + g(X) be-
ing a polynomial of degree < n which vanishes at the n distinct integers
a1 , · · · , an must be identically zero. This gives P (X) = −f (X)2 but this is
impossible as can be seen by comparing the coefficients of X n .
Exercise 1.8. Let n be odd and a1 , · · · , an be distinct integers. Prove that
(X − a1 ) · · · (X − an ) + 1 is irreducible.
Let us consider the following situation. Suppose p = an · · · a0 is a prime
number expressed in the usual decimal system i.e. p = a0 + 10a1 + 100a2
+ · · · + 10n an , 0 ≤ ai ≤ 9. Then, is the polynomial a0 + a1 X + · · · + an X n
irreducible? For example 1289 is a prime the following result due to A Cohn
and x3 + 2x2 + 8x + 9 is irreducible. This is, in fact, true more generally
and, we have:
Lemma 1.9. Let P (X) ∈ Z[X] and assume that there exists an integer n
such that
(i) the zeros of P lie in the half plane Re(z) < n − 21 .
(ii) P (n − 1) 6= 0
(iii) P (n) is a prime number.
Then P (X) is irreducible.
Proof. Suppose, if possible P (X) = f (X)g(X) over Z with f, g hav-
ing positive degrees. All the zeros of f (X) also lie in Re(z) < n − 12 .
Writing f as a product of its irreducible factors over R, we can observe
that f (x + n − 1/2) has ALL coefficients non-zero and of the same sign.
Thus, the coefficients of f (−x + n − 1/2) have alternate signs. Therefore,
|f (n − 21 − t)| < |f (n − 12 + t)|∀t > 0. Since f (n − 1) 6= 0 and f (n − 1) is
integral, we have |f (n − 1)| ≥ 1. Thus |f (n)| > |f (n − 1)| ≥ 1. A similar

27
 Polynomials with Integer Values

thing holding for g(X), we get that P (n) has proper divisors f (n), g(n)
which contradicts our hypothesis.
Remark: Michael Filaseta and collaborators have generalized
this vastly. They show that there exists an integer polynomial f of
degree 129 explicitly written down whose largest coefficient is
49598666989151226098104244512919 such that f (10) is prime but f has
the factor x2 − 20x + 101. Further, every integer polynomial g of any degree
whose coefficients are non-negative and strictly less than the above number
must be irreducible if g(10) is prime!

2. Irreducibility and Congruence Modulo p

For an integral polynomial to take the value zero at an integer or even
to be reducible, it is clearly necessary that these properties hold modulo
any integer m. Conversely, if P (X) has a root modulo any integer, it must
itself have a root in Z . In fact, if P (X) ∈ Z [X] has a linear factor modulo
all but finitely many prime numbers, the P (X) itself has a linear factor.
This fact can be proved only by deep methods viz. using the so-called
Ćebotarev density theorem. On the other hand, (see lemma 2.3) it was first
observed by Hilbert that the reducibility of a polynomial modulo every in-
teger is not sufficient to guarantee its reducibility over Z . Regarding roots
of a polynomial modulo a prime, there is following general result due to
Lagrange:
Lemma 2.1. Let p be a prime number and let P (X) ∈ Z [X] be of degree
n. Assume that not all coefficients of P are multiples of p. Then the number
of solutions mod p to P (X) ≡ 0 mod p is, at the most, n.
The proof is obvious using the division algorithm over Z/p. In fact, the
general result of this kind (provable by the division algorithm again) is that
a nonzero polynomial over any field has at the most its degree number of
roots.
Remark 2.2. Since 1, 2, · · · , p − 1 are solutions to X p−1 ≡ 1 mod p, we
have
X p−1 − 1 ≡ (X − 1)(X − 2) · · · (X − (p − 1)) mod p
For odd p, putting X = 0 gives Wilson’s theorem that (p − 1)! ≡
−1 mod p.
Note that we have observed earlier that any integral polynomial has a
root modulo infinitely many primes. However, as first observed by Hilbert,
the reducibility of a polynomial modulo every integer does not imply its
reducibility over Z. For example, we have the following result:
Lemma 2.3. Let p, q be odd prime numbers such that ( pq ) = ( pq ) = 1 and
p ≡ 1 mod 8. Here ( pq ) denotes the Legendre symbol defined to be 1 or −1

28
 Polynomials with Integer Values

according as p is a square or not modulo q. Then, the polynomial P (X) =

(X 2 − p − q)2 − 4pq is irreducible whereas it is reducible modulo any integer.
Proof.
P (X) = X 4 − 2(p + q)X 2 + (p − q)2
√ √ √ √ √ √ √ √
= (X − p − q)(X + p + q)(X − p + q)(X + p − q).
√ √ √ √ √
Since p, q, p ± q, pq are all irrational, none of the linear or
quadratic factors of P (X) are in Z[X] i.e. P (X) is irreducible. Note that it
is enough to show that a factorisation of P exists modulo any prime power
as we can use Chinese reminder theorem to get a factorisation modulo a
general integer.
Now, P (X) can be written in the following ways:
P (X) = X 4 − 2(p + q)X 2 + (p − q)2
= (X 2 + p − q)2 − 4pX 2
= (X 2 − p + q)2 − 4qX 2
= (X 2 − p − q)2 − 4pq.

The second and third equalities above show that P (X) is reducible mod-
ulo any pn and any q n . Also since p ≡ 1 mod 8, p is a quadratic residue
modulo any 2n and the second equality above again shows that P (X) is
the difference of two squares modulo 2n , and hence reducible mod 2n .
If ` is a prime 6= 2, p, q, let us show now that P (X) is reducible modulo
ln for any n.
At least one of ( p` ), ( q` ) and ( pq
` ) is 1 because, by the product formula for
p
Legendre symbols, ( ` ) · ( ` ) · ( ` ) = 1. According as ( p` ), ( q` ) or ( pq
q pq
` ) = 1, the
second, third or fourth equality shows that P (X) is reducible mod `n for
any n.
We end this section with a result of Schur whose proof is surprising and
elegant as well. This is:
Schur’s Theorem 2.4. For any n, the truncated exponential polynomial
2 n
En (X) = n!(1 + X + X2! + · · · + Xn! ) is irreducible over Z.
Just for this proof, we need some nontrivial number theoretic facts. A
reader unfamiliar with these notions but one who is prepared to accept at
face value a couple of results can still appreciate the beauty of Schur’s proof.
Here is where we have to take recourse to some very basic facts about prime
decomposition in algebraic number fields. Start with any (complex) root α
of f and look at the field K = Q(α) of all those complex numbers which
can be written as polynomials in α with coefficients from Q. The basic
fact that we will be using (without proof) is that any nonzero ideal in ‘the
ring of integers of K’ (i.e., the subring OK of K made up of those elements

29
 Polynomials with Integer Values

which satisfy a monic integral polynomial) is uniquely a product of nonzero

prime ideals and a prime ideal can occur only at the most the ‘degree’
times. This is a good replacement for K of the usual unique factorisation
of natural numbers into prime numbers. The proof also uses a fact about
prime numbers observed by Sylvester but is not trivial to prove.
Sylvester’s Theorem. If m ≥ r, then (m + 1)(m + 2) · · · (m + r) has a
prime factor p > r.
The special case m = r is known as Bertrand’s postulate. (see the next
chapter for two proofs).
Proof of Schur’s Theorem. Suppose, if possible, that En (X) = f (X)g(X)
for some nonconsant, irreducible integral polynomial f . Let us write f (X)
= a0 + a1 X + · · · + X r (evidently, we may take the top coefficients of f to
be 1).
Now, the proof uses the following observation which is interesting in its
own right:
Observation: Any prime dividing the constant term a0 of f is less than
the degree r of f .
To see this, note first that N (α), the ‘norm of α’ (a name for the product
of all the roots of the minimal polynomial f of α) is a0 upto sign. So, there
is a prime ideal P of OK so that (α) = P k I, (p) = P l J where I, J are
indivisible by P and k, l ≥ 1. Here, (α) and (p) denote, respectively, the
ideal of OK generated by α and p. Since En (α) = 0, we have
0 = n! + n!α + n!α2 /2! + · · · + αn .
We know that the exact power of p dividing n! is
hn = [n/p] + [n/p2 ] + · · · · · ·
Thus, in OK , the ideal (n!) is divisible by P lhn and no higher power of
P . Similarly, for 1 ≤ i ≤ n, the ideal generated by n!αi /i! is divisible by
P lhn −lhi +ki . Because of the equality
−n! = n!α + n!α2 /2! + · · · + αn ,
it follows that we cannot have lhn − lhi + ki cannot be strictly bigger than
lhn which is the exact power of P dividing the left hand side. Therefore,
there is some i so that −lhi + ki ≤ 0. Thus,
li
i ≤ ki ≤ lhi = l([i/p] + [i/p2 ] + · · · ) < .
p−1
Thus, p − 1 < l ≤ r i.e., p ≤ r. This confirms the observation.
To continue with the proof, we may clearly assume that the degree r of
f at most n/2. Now, we use Sylvester’s theorem to choose a prime q > r
dividing the product n(n − 1) · · · (n − r + 1). Note that we can use this

30
 Polynomials with Integer Values

theorem because the smallest term n − r + 1 of this r-fold consecutive

product is bigger than r as r ≤ n/2. Note also that the observation tells us
that q cannot divide a0 . Now, we shall write En (X) modulo the prime q.
By choice, q divides the coefficients of X i for 0 ≤ i ≤ n − r.
X n−1 X n−r+1
So, f (X)g(X) ≡ X n + n! (n−1)! + · · · + n! (n−r+1)! mod q.
Write f (X) = a0 + a1 X + · · · + X r and g(X) = b0 + b1 X + · · · + X n−r .
The above congruence gives a0 b0 ≡ 0, a0 b1 + a1 b0 ≡ 0 etc. mod q until
the coefficient of X n−r of f (X)g(X). As a0 6≡ 0 mod q, we get recursively
(this is just like the proof of Eisenstein’s criterion – see box) that
b0 ≡ b1 ≡ · · · bn−r ≡ 0 mod q.
This is impossible as bn−r = 1. Thus, Schur’s assertion follows.

3. Polynomials Taking Square Values

If an integral polynomial takes only values which are squares, is it true
that the polynomial itself is a square of a polynomial? In this section, we
will show that this, and more, is indeed true (see also [1]).
Lemma 3.1. Let P (X) be a Z-valued polynomial which is irreducible. If
P is not a constant, then there exist arbitarily large integers n such that
P (n) ≡ 0 mod p and P (n) 6≡ 0 mod p2 for some prime p.
Proof. First, suppose that P (X) ∈ Z[X]. Since P is irreducible, P and
P 0 have no common factors. Write f (X)P (X) + g(X)P 0 (X) = c for some
f, g ∈ Z[X] and some non-zero integer c. By lemma 1.5, there is a prime
p such that P (n) ≡ 0 mod p where n can be as large as we want. So,
P 0 (n) 6≡ 0 mod p as f (n)P (n) + g(n)P 0 (n) = c. Since P (n + p) − P (n) ≡
P 0 (n) mod p2 , either P (n + p) or P (n) is 6≡ 0 mod p2 . To prove the result
for general P , one can replace P by m! · P where m = deg P .
Lemma 3.2. Let P (X) be a Z-valued polynomial such that the zeros of
smallest multiplicity have multiplicity m. Then, there exist arbitrarily large
integers n such that P (n) ≡ 0 mod pm , P (n) 6≡ 0 mod pm+1 for some
prime p.
Proof. Let P1 (X), · · · , Pr (X) be the distinct irreducible factors of P (X).
Write P (X) = P1 (X)m1 · · · Pr (X)mr with m = m1 ≤ · · · mr . By the above
lemma, one can find arbitrarily large n such that for some prime p, P1 (n) ≡
0 mod p, P1 (n) 6≡ 0 mod p2 and, Pi (n) 6≡ 0 mod p for i > 1. Then, P (n) ≡ 0
mod pm and 6≡ 0 mod pm+1 .
COROLLARY 3.3.
If P (X) takes at every integer, a value which is the k-th power of an
integer, then P (X) itself is the k-th power of a polynomial.

31
 Polynomials with Integer Values

Proof. If P (X) is not an exact k-th power, then one can write P (X)
= f (X)k g(X) for polynomials f, g so that g(X) has a zero whose multiplic-
ity is < k. Once again, we can choose n and a prime p such that g(n) ≡ 0
mod p, 6≡ 0 mod pk . This contradicts the fact that P (n) is a k-th power.
Remark: The above results and much more general properties of polyno-
mials are consequences of the so-called Hilbert irreducibility criterion which
implies: if f (X, Y ) is an irreducible polynomial with rational coefficients,
then there exist infinitely many rational values a of x such that the polyno-
mials f (a, Y ) are irreducible in Q[Y ]. One application of the above theorem
is:
Given two non-constant polynomials f, g with rational coefficients such
that f (Q) is contained in g(Q), there exist a polynomial h with rational
coefficients so that f (X) = g(h(X)).

4. Cyclotomic Polynomials
These were already referred to in the earlier chapter. It was also shown
there that one could use these polynomials to prove the existence of in-
finitely many primes congruent to 1 modulo n for any n. For a natural
number d, recall that the cyclotomic polynomial Φd (X) is the irreducible,
monic polynomial whose roots are the primitive d-th roots of unity i.e.
Φd (X) = a≤d:(a,d)=1 (X − e2πa/d ). Note that Φ1 (X) = X − 1 and that
Q

for a primeQp, Φp (X) = X p−1 + · · · + X + 1. Observe that for any n ≥ 1,

X n − 1 = d/n Φd (X).
Exercise 4.1.
(i) Prove that for any d, Φd (X) has integral coefficients.
(ii) Prove that for any d, Φd (X) is irreducible over Q.
Factorising an integral polynomial into irreducible factors is far from
easy. Even if we know the irreducible factors, it might be difficult to decide
whether a given polynomial divides another given one.
Exercises 4.2.

Q positive integers a1 < · · · <Q

(a) Given an , consider the polynomials P (X)
= i>j (X ai −aj − 1) and Q(X) = i>j (X i−j − 1). By factorizing into
cyclotomic polynomials, prove that Q(X) divides P (X). Conclude that
Q ai −aj
i>j i−j is always an integer.
(b) Consider  the n ×n matrix A whose (i, j)-th entry is the Gaussian poly-
ai
nomial .
j−1
Compute detA to obtain the same conclusion as in part (a).

32
 Polynomials with Integer Values

Here, for m ≥ r > 0, the Gaussian polynomial is defined as

(X m − 1)(X m−1 − 1) · · · (X m−r+1 − 1)
 
m
= .
r (X r − 1)(X r−1 − 1) · · · (X − 1)
     
m m−1 r m−1
Note that = +X .
r r−1 r
Recall from the earlier chapter that from looking at Φp (X) for prime p,
it seems as though the coefficients of the cyclotomic polynomials Φd (X) for
any d are among 0, 1 or −1. However, the following rather amazing thing
was discovered by Schur. His proof uses a consequence of a deep result about
prime numbers known as the prime number theorem. The prime-number
theorem tells us that π(x) ∼ x/log(x) as x → ∞. Here π(x) denotes the
number of primes until x. The reader does not need to be familiar with the
prime number theorem but is urged to take on faith the consequence of it
that for any constant c, there is n such that π(2n ) ≥ cn.

PROPOSITION 4.3.
Every integer occurs as a coefficient of some cyclotomic polynomial.

Proof. First, we claim that for any integer t > 2, there are primes p1 < p2
< · · · < pt such that p1 + p2 > pt . Suppose this is not true. Then, for some
t > 2, every set of t primes p1 < · · · < pt satisfies p1 + p2 ≤ pt . So, 2p1 < pt .
Therefore, the number of primes between 2k and 2k+1 for any k is less than
t. So, π(2k ) < kt. This contradicts the prime-number theorem as noted
above. Hence, it is indeed true that for any integer t > 2, there are primes
p1 < p2 < · · · < pt such that p1 + p2 > pt .
Now, let us fix any odd t > 2. We shall demonstrate that both −t + 1 and
−t + 2 occur as coefficients. This will prove that all negative integers occur
as coefficients. Then, using the fact that for an odd m > 1, Φ2m (X) =
Φm (−X), we can conclude that all integers are coefficients.
Consider now primes p1 < p2 < · · · < pt such that p1 + p2 > pt . Write
pt = p for simplicity. Let n = p1 · · · pt and let us write Φn (X) modulo X p+1 .
Since X n − 1 = d/n Φd (X), and since p1 + p2 > pt , we have
Q

t
Y 1 − X pi
Φn (X) ≡ ≡ (1 + · · · + X p )(1 − X p1 ) · · · (1 − X pt )
1−X
i=1
≡ (1 + · · · + X p )(1 − X p1 − · · · − X pt ) mod X p+1 .

Therefore, the coefficients of X p and X p−2 are 1−t and 2−t respectively.
This completes the proof. Note that in the proof, we have used the fact that
if P (X) = (1−X r )Q(X) for a polynomial Q(X), then Q(X) = P (X)(1+X r
+ X 2r + · · · + · · · ) modulo any X k .

33
 Polynomials with Integer Values

Exercise 4.4.
(a) Let A = (aij ) be a matrix in GL(n, Z) i.e., both −1 have integer
Pn A and Aj
entries. Consider the polynomials pi (X) = j=0 aij X for 0 ≤ i ≤ n.
Prove that any integral polynomial of degree at most n is an integral
linear combination of the pi (X). In particular, if a0 , . . . , an ∈ Q are
distinct, show that any rational polynomial of degree at most n is of
n
λi (X + ai )n for some λi ∈ Q.
P
the form
i=0
[n
2
]
+ . . . + Xn (−1)i n−i
P
i n−2i . Conclude
(b) Prove that 1 + X = i X (1 + X)
i=0
P n−i
1+ρ+···+ρn
that i = (1+ρ)n where ρ is either root of X 2 + 3X + 1 = 0.
i≥0
(−1)i n−i
P

Further, compute i .
i≥0
n−i
P

Remark 4.5. It is easily seen by induction that i is just the
i≥0
(n+1)-th Fibonacci number Fn+1 . Thus, exercise (b) provides an expression
for Fn+1 . This expression makes it easy to prove the following identities:
(a) Fn + Fn+1 = Fn+2 .
(b) F
Pn+1 Fn−1 = Fn2 + (−1)n .
n z
(c) Fn z = 1−z−z 2.

(d) 2 2
Fn + Fn+1 = F2n+1 .
n−i
P

Notice that only (a) seems obvious from the expression Fn+1 = i .
i≥0
As we remarked earlier, even for a polynomial of degree 2 (like X 2 + 1)
it is unknown whether it takes infinitely many prime values. A general
conjecture in this context is:
Conjecture 4.6. (Bouniakowsky, Schinzel and Sierpinski.) A nonconstant
irreducible integral polynomial whose set of values has no nontrivial
common factor, always takes on a prime value.
It is appropriate to recall here that the polynomial X 2 + X + 41 takes
prime values at X = −40, −39, . . . , 0, 1, . . . , 39. We end with an open ques-
tion which is typical of number-theoretic questions – a statement which
can be understood by the proverbial layman but an answer which proves
elusive to this day to professional mathematicians.
For any irreducible, monic,
Q integral polynomial P (X), define its Mahler
measure to be M (P ) = i Max(|αi |, 1) where the product is over the roots
of P . The following is an easy exercise.
Exercise 4.7. M (P ) = 1 if, and only if, P is cyclotomic.
D H Lehmer posed the following question:

34
 Polynomials with Integer Values

Does there exist a constant C > 0 such that M (P ) > 1 + C for all
noncyclotomic (irreducible) polynomials P ?
This is expected to have an affirmative answer and, indeed, Lehmer’s
calculations indicate that the smallest possible value of M (P ) 6= 1 is
1.176280821...., which occurs for the polynomial

P (X) = X 10 + X 9 − X 7 − X 6 − X 5 − X 4 − X 3 + X + 1.

Lehmer’s question can be formulated in terms of discrete subgroups of

Lie groups. One may not be able to predict when it can be answered but it
is more or less certain that one will need tools involving deep mathematics.

References
[1] Polya and Szego, Problems in Analysis, I & II, Springer-Verlag, 1945.

35
 Polynomials with Integer Values

Box 1. Eisenstein’s and Dumas’s Criteria

A general criterion known to check whether an integral polynomial of a special

kind is irreducible is due to G Eisenstein, a student of Gauss and an outstanding
mathematician. Eisenstein died when he was 27.
Let f (X) = a0 +a1 X +· · ·+an X n be an integral polynomial satisfying the following
property with respect to some prime p. The prime p divides a0 , a1 , · · · , an−1 but does
not divide an . Also, assume that p2 does not divide a0 . Then, f is irreducible.
The proof is indeed very simple high school algebra. Suppose, if possible, that f (X)
= g(X)h(X) = (b0 +b1 X+· · ·+br X r )(c0 +c1 X+· · ·+cs X s ) with r, s ≥ 1. Comparing
coefficients, one has

a0 = b0 c0 , a1 = a0 b1 + b0 a1 , · · · , an = br cs , r + s = n.

Since a0 = b0 c0 ≡ 0 mod p, either b0 ≡ 0 mod p or c0 ≡ 0 mod p.

To fix notations, we may assume that b0 ≡ 0 mod p. Since a0 6≡ 0 mod p2 , we
must have c0 6≡ 0 mod p. Now a1 = b0 c1 + b1 c0 ≡ b1 c0 mod p; so b1 ≡ 0 mod p.
Proceeding inductively in this maner, it is clear that all the bi ’s are multiples of p.
This is a manifest contradiction of the fact that an = br cs is not a multiple of p.
This finishes the proof.
It may be noted that one may reverse the roles of a0 and an and obtain another
version of the criterion:
Let f (X) = a0 +a1 X +· · ·+an X n be an integral polynomial satisfying the following
property with respect to some prime p. The prime p divides a1 , a2 , · · · , an but does
not divide a0 . Also, assume that p2 does not divide an . Then, f is irreducible.
More generally, we have the following irreducibility criterion due to Gustave Du-
mas from 1906:
Let f = n i
P
i=0 ci X be a polynomial with integer coefficients and let p be a prime
such that vp (cn ) = 0, vp (c0 ) < nvp (cn−i )/i for 0 < i < n, and gcd(n, vp (c0 )) = 1.
Then, f is irreducible modulo p (hence irreducible over Z itself).
Here vp (a) denotes the power of p dividing a. Note that the polynomial X 6 + 2X 5
+ 3X 4 + 4x3 + 5X 2 + 6X + 7 is irreducible modulo 5 by Dumas’s criterion. Usually,
this criterion is stated in terms of so-called Newton polygons.

36
How Far Apart are Primes?
Bertrand’s Postulate

It is well-known that there are arbitrarily large gaps in between primes.

Indeed, given any natural number n, the numbers (n + 1)! + 2, (n + 1)!
+3, · · · , (n+1)!+(n+1) being large multiples of 2, 3, · · · , n+1 respectively,
are all composite numbers.
Let us now ask ourselves the following question. If we start with a natural
number n and start going through the numbers n + 1, n + 2, etc., how far
do we have to go before hitting a prime? Trying out the first few numbers,
we see that

1 → 2, 2 → 3, 3 → 5, 4 → 5, 5 → 7, 6 → 7, 7 → 11 etc.

Thus, it seems that we need to go ‘at most twice the distance’ i.e., we
seem to be able to find a prime between n and 2n for the first few values of
n. But there is absolutely no pattern here. In fact, although we have seen
above that there are arbitrarily large gaps between primes, it is nevertheless
true that ‘there is regularity in the distribution of primes’. It is this fascinat-
ing clash of tendency which seems to make primes at once interesting and
intriguing. It turns out indeed to be true that: there is always a prime be-
tween n and 2n. This statement, known as Bertrand’s postulate, was stated
by Bertrand (1822–1900) in 1843 and proved later by Chebychev in 1852.
Actually, Chebychev proves a much stronger statement which was further
generalised to yield a fundamental fact about the prime numbers known
as the prime number theorem. Interestingly, Bertrand’s motivation was to
group theory and not really number theory; he made many contributions
to differential geometry and probability theory as well.
Proving the prime number theorem is beyond the scope of this article
but stating it certainly lies within it. For a positive real number x, let us
denote by π(x), the number of primes which do not exceed x. The prime
number theorem states that the ratio π(x)logx x approaches the limit 1 as x
π(x)logx
grows indefinitely large i.e., limx→∞ x = 1.
x
One usually writes π(x) ∼ logx to describe such an asymptotic result.
What Chebychev proved was that there are some explicit positive constants
a, b so that
logx logx
a < π(x) < b .
x x
The chapter is a modified version of an article that first appeared in Resonance, Vol. 7, No. 6,
pp. 77–87, June 2002.

37
 How Far Apart are Primes? Bertrand’s Postulate

If pn denotes the n-th prime, Bertrand’s postulate is equivalent to the as-

sertion that pn+1 < 2pn while the prime number theorem itself is equivalent
pn
to the statement limn→∞ nlogn = 1.
It is rather startling to note that the great mathematician Gauss (1777-
1855) had, at the age of 15, already conjectured the truth of the prime
number theorem. Four years later, in 1796, Legendre also came indepen-
dently to conjecture something similar.
x
Legendre conjectured, based on empirical evidence, that π(x) ∼ Alogx+B
and also conjectured values of A, B which turnedR xout to be incorrect. Gauss,
dt
on the other hand, conjectured that π(x) ∼ 2 logt ; the right side is de-
noted by li(x) to stand for the ‘logarithmic integral’. This seems to have
x
exactly the content of the prime number theorem since clearly li(x) ∼ logx .
However, later research (following Riemann) has confirmed that Gauss’s
assertion is even more astute than what it appears to be on the face of it.
In fact, the function li(x) has the asymptotic expansion (for any fixed n)
x x x x x
li(x) = + 1! 2
+ 2! 3
+ · · · + (n − 1)! n
+ O( ).
logx (logx) (logx) (logx) (logx)n+1

A refined version of the prime number theorem indeed implies that π(x)
has the same asymptotic expansion.
In particular, this implies that the best possible values for A and B in
Legendre’s conjecture are A = 1, B = −1.
The prime number theorem was proved independently by Hadamard and
de la Vallee Poussin. A well-known mathematician quipped once that the
proof almost immortalised these two mathematicians – they lived to be 96
and 98 respectively!
Returning to Bertrand’s postulate, after Chebychev’s first proof, other
simpler proofs appeared. A generalization of Bertrand’s postulate is
Sylvester’s theorem which was stated and used in the previous chapter.
In this chapter, we shall discuss two of the simplest proofs due to two great
minds – Ramanujan and Erdos. Most of us are told stories about Ramanu-
jan and his discoveries and it is rarely that one can find a proof of his which
is elementary enough to be actually discussed at this level. Erdos’s proof is
even more elementary and we start with it.

Erdos’s Proof
Q
We start with any natural number n and look at the product p≤n p
over all primes p ≤ n.
We shall have occasion to use the well-known and easily proved fact
asserting that the highest power to which a prime divides n! is given by the

38
 How Far Apart are Primes? Bertrand’s Postulate

expression
[n/p] + [n/p2 ] + [n/p3 ] + · · · · · ·
Erdos’s proof starts with the following very beautiful observation:
Lemma. p≤n p ≤ 4n .
Q
Proof. We prove this by induction on n. It evidently holds good for small
n. Look at some n > 1 such that the result is assumed for all m ≤ n. Then,
Y Y Y
p= p p
p≤n 2p≤n+1 n+1<2p≤2n
n+1 Y
≤4 2 p
n+1<2p≤2n

by the induction hypothesis.

Now, the surprisingly simple observation that each prime in the last
product (i.e., each
prime between
Q (n + 1)/2 and n) divides the binomial
n (n+1)/2
coefficient [ n+1 ] , shows that p≤n p ≤ 4 2n−1 = 4n .
2
n n−1 used above is trivially seen to be true by


The bound [ n+1
2
] ≤ 2
induction. Thus, we have proved this lemma.
Since we are interested in the possible primes
between n and 2n, it is
2n
natural to consider the binomial coefficient n because it ‘captures’ all
these primes as its divisors. Now, obviously, the binomial coefficient 2n


n
is the largest term in the expansion (1 + 1)2n which has 2n + 1 terms.
Therefore, we have  
2n
(2n + 1) ≥ 22n . (1)
n
This gives a lower bound for this middle binomial coefficient and, the idea
of the rest of the proof is that lot of the contribution comes from primes
between n and 2n. More precisely, if we write
 
2n Y Y Y
= pep = pep p,
n
p≤2n p≤n n+1<p≤2n

then
Q weepshall use the lemma to give an upper bound for the first product
p≤n p . Here,
p
e denotes the power of p dividing the middle binomial
2n
coefficient n and, the second product stands for 1 if there are no terms.
We want to see which primes p ≤ n actually contribute to 2n


.
√ n
If p2 > 2n i.e., if n ≥ p > 2n, then clearly, ep = [ 2n
p ] − 2[ n
p ] = 0 or 1.
2n


Thus, such primes divide n either to a single power or not at all.
If n ≥ 3, then a prime p ≤ n with 2n/3 < p must be at least 3 and so
p2 ≥ 3p > 2n. As 1 ≤ n/p < 3/2 for 2n/3 < p ≤ n, we have [ np ] = 1 and
[ 2n
p ] = 2 i.e.,

39
 How Far Apart are Primes? Bertrand’s Postulate

2n
ep = [ 2n n


p ] − 2[ p ] = 2 − 2 = 0. Thus, these primes do not divide n when
n ≥ 3.
In other words,
√ we have:
ep ≤ 1 if 2n < p ≤ 2n/3, and ep√= 0 if 2n/3 < p ≤ n.
Finally, for the primes with p ≤ 2n, we simply take the trivial bound
pep ≤ 2n. Then, we have
  Y
2n Y Y Y Y
= pep p≤ (2n) p p
n √ √
p≤n n+1<p≤2n p≤ 2n 2n<p≤2n/3 n+1<p≤2n
Y Y
≤ (2n) 42n/3 p,
√
p≤ 2n n+1<p≤2n

using the lemma. We take√ n ≥ 8 as we have verified Bertrand’s postulate

explicitly for n ≤ 7;√so 2n ≥ 4 and thus, the number of terms in the first
product is at most 2n − 2 (as 1 and 4 are not primes). Therefore, we have
on using (1) that

22n √
 
2n Y
≤ ≤ (2n) 2n−2 42n/3 p.
2n + 1 n
n+1<p≤2n

22n
Replacing the first term by (2n)2
, we get

Y 4n/3
p≥ √ .
(2n) 2n
n+1<p≤2n

Thus, to show that the left side has terms (i.e., that it is not 1 according to
our convention), it suffices to see whether the right hand side is bigger than
1 for all n. As usual, this will turn out to be true for large enough values of
n and will fail for small values (this only means that the inequality is good
enough for large values of n and we need to verify the original assertion
directly for the smaller values left out).
After a few trials, √
we arrive at the number n = 450 and find that
4n/3 = 4150 > (2n) 2n = (900)30 since 45 > 900.
There is nothing special about 450 excepting the fact that 2n is a per-
fect square and 450 is large enough for the inequality to hold good. This √
inequality continues to hold for n > 450 as the difference 4n/3 − (2n) 2n
is an increasing function. This last statement is simple to see by looking
at the derivative of the difference of the corresponding logarithms. Now it
is an easy exercise to verify Bertrand’s postulate for n < 450. The above
proof was essentially due to Erdos; it is a slightly simplified version of his
original argument which appears in [2].

40
 How Far Apart are Primes? Bertrand’s Postulate

Ramanujan’s Proof
Let us turn to Ramanujan’s proof. It is also extremely clever and com-
pletely elementary apart from the use of what is known as Stirling’s formula
– a proof has been discussed in Resonance earlier Q[3].
In the previous proof we used an estimate for p≤n p. Here, we consider
an Padditive version of it viz., look at the so-called Chebychev function θ(x)
= p≤x log p defined for any real number x ≥ 2. One remark about why
one considers functions like the Chebychev function instead of just the
prime-counting function – weighted prime-counting is easier as the function
becomes smoother. We note two things to begin with:
Q
(i) θ(n) is simply the logarithm of p≤n p,
(ii) Bertrand’s postulate is true for a real number x if it is true for n = [x];
indeed, a prime between n and 2n is between x and 2x as well.
Let us also understand that since we are interested in primes between x
and 2x, we need a lower bound for θ(2x) − θ(x). In oher words, we need
reasonable lower as well P
as upper bounds for θ values.
Now, the expression i
i≥0 [n/p ] for the power of a prime dividing n!
gives us
X
log[x]! = Ψ(x/i),
i≥1

where Ψ is the function defined by

X
Ψ(x) = θ(x1/i ).
i≥1

This is the reason to introduce real x.

Using an elementary trick of old vintage, we have the following:
X
log[x]! − 2log[x/2]! = (−1)i−1 Ψ(x/i),
i≥1
√ X
Ψ(x) − 2Ψ( x) = (−1)i−1 θ(x1/i ).
i≥1

As θ, Ψ are increasing functions, we get inequalities by chopping off at

an odd stage and at an even stage as follows:
x x x x
Ψ(x) − Ψ( ) ≤ log[x]! − 2log[ ]! ≤ Ψ(x) − Ψ( ) + Ψ( ), (2)
2 2 2 3
√
Ψ(x) − 2Ψ( x) ≤ θ(x) ≤ Ψ(x). (3)
Here, Ramanujan takes recourse to Stirling’s formula which states that
√ n+ 1
n! ∼ 2π n en 2 .

41
 How Far Apart are Primes? Bertrand’s Postulate

We shall not use it but proceed as follows.

For any x > 1, we have the binomial coeficient
  X 
[x] [x]
x < = 2[x] .
[2] r
r≥0
Taking logarithms, we obtain
log[x]! − 2log[x/2]! < xlog 2.
3
But, clearly log 2 < 4 since 16 < e3 . In other words, we have
log[x]! − 2log[x/2]! < 3x/4 ∀ x > 0 · · · . (4)
Now, we find a lower bound. As we observed before Erdos’s proof, for x >
[x]

0, the binomial coefficient [x/2] (being the largest term in the expansion
2[x]
of (1 + 1)[x] ), must be bigger than [x]+1 since there are [x] + 1 terms in the
binomial expansion of (1 + 1)[x] .
2[x]
2[x]
If x is large enough (for instance, if x > 240), then [x]+1 >e 3 . Taking
logarithms, we get
log[x]! − 2log[x/2]! > 2[x]/3 ∀ x > 240. (5)
By (2),(4) and (5), we have
Ψ(x) − Ψ(x/2) < 3x/4 ∀ x > 0, (6)

Ψ(x) − Ψ(x/2) + Ψ(x/3) > 2[x]/3 ∀ x > 240. (7)

By replacing x by x/2, x/4, x/8 etc. in (6) and adding all the expressions
we get
Ψ(x) < 3x/2 ∀ x > 0. (8)
Note that since θ(x) < Ψ(x), we have a reasonable upper bound for θ(x)
by (8). For the lower bound, let us use the first inequality of (3) viz., Ψ(x)
√
≤ 2Ψ( x) + θ(x) and the inequality θ(x/2) ≤ Ψ(x/2) to write
√
Ψ(x) − Ψ(x/2) + Ψ(x/3) ≤ 2Ψ( x) + θ(x) − θ(x/2) + Ψ(x/3).
If we use the upper bound for Ψ given in (8), we obtain
x x x x √
Ψ(x) − Ψ( ) + Ψ( ) < θ(x) − θ( ) + + 3 x. (9)
2 3 2 2
For x > 240, the left side has a lower bound given in (7) so that we finally
obtain √
θ(x) − θ(x/2) > x/6 − 3 x − 2/3 ∀ x > 240.
Evidently, the right side is positive if x > 361. Therefore, there is a
prime between x and 2x if x > 181. For smaller values of x, we find primes
explicitly as before. This finishes the beautiful proof due to Ramanujan.

42
 How Far Apart are Primes? Bertrand’s Postulate

More Comments on Primes

The above elementary methods and their modifications are sufficient to
prove Chebychev’s theorem viz., the assertion
1 x x
< π(x) < 6 .
6 logx logx
The reader is urged to try and use this to prove the following bound for
the n-th prime:
1 12
nlogn < pn < 12(nlogn + nlog ).
6 e
PThis
1
last inequality (in fact, the upper bound) shows easily that the series
p over all primes diverges. How fast does it diverge?
Here is a proof showing that the divergence is at least as fast as loglogx
i.e., we prove
∃c > 0 such that p≤x p1 ≥ cloglogx ∀x.
P

To see this, given x > 1, let us look at the area under the curve y = x1
Rx P[x] 1
between 1 and x. Evidently, this area 1 dx x = logx is less than n=1 n .
(see the figure).
So, if we consider the product p≤x (1 − p1 )−1 , then clearly,
Q

Y 1 X[x] 1
(1 − )−1 ≥ ≥ logx.
p≤x p n=1 n

Now, (1 − p1 )−1 = (1 + p1 )(1 + p21−1 ). Therefore, p≤x (1 + p1 ) ≥ alogx

Q

where a = p (1 + p21−1 )−1 . Using ex > 1 + x for all x > 0, we get p≤x e1/p
Q Q

≥ p≤x (1 + p1 ) ≥ alogx so that p≤x p1 ≥ cloglogx for some c > 0.

Q P
This finishes the proof of the lower bound.
The more adventurous reader may like to use the Abel summation for-
mula (see [4]) and prove that this lower bound is of the correct order i.e.,
one has the rather interesting statement:
X1
∼ loglogx.
p
p≤x

Some Applications of Bertrand’s Postulate

The original application of Bertrand’s postulate was to Galois theory!
Let us talk about some elementary number-theoretic applications.

A Recursion for Primes

We recall a curious application of Bertrand’s postulate to finding a re-
cursive expression for primes which appeared in [5], p. 289.

43
 How Far Apart are Primes? Bertrand’s Postulate

We consider the function

  
2((n − 1)!) 2((n − 1)!)
fn = Sign −
n n

defined for all n ≥ 3. Here, the sign function is the function which takes the
x
value 0 at 0 and the value |x| for any x 6= 0. Clearly, fn = 1 or 0 according
as n is prime or composite. Now, by Bertrand’s postulate, if pn ≥ 3, then
pn+1 occurs as the first prime among pn + 2, pn + 4, · · · , 2pn − 1. Therefore,
(writing pn as p for simplicity of notation),

pn+1 = (p + 2)fp+2 + (p + 4)fp+4 (1 − fp+2 ) + · · ·

+(p+6)fp+6 (1−fp+2 )(1−fp+4 )+· · ·+(2p−1)f2p−1 (1−fp+2 ) · · · (1−f2p−3 ).

The Harmonic Sum

The sum 12 + 13 + · · · + n1 is not an integer for any n ≥ 2.
1 1 1
More generally, m+1 + m+2 + · · · + m+n is not an integer for positive
integers m, n. An application of Bertrand’s postulate gives a very quick
proof of this – we leave this as an exercise.

Prime Sums
For any positive integer n, consider the set {1, 2, · · · , 2n} of the first 2n
positive integers. We claim that this set can be written as the union of n
pairs of integers {ai , bi } (1 ≤ i ≤ n) such that ai + bi is prime! Indeed,
this is clear for n = 1 as 1 + 2 = 3 is prime, and we will apply induction
on n to prove it in general. Assume that n > 1 and that our assertion
is valid for every m < n. Now, Bertrand’s postulate ensures we have a
prime p among the numbers in the set {2n + 1, 2n + 2, · · · , 4n − 1}. Writing
p = 2n + r, we have r ∈ {1, 2, · · · , 2n − 1}. Thus, note that r is odd as
p must be an odd prime. If r > 1, then by induction hypothesis, the set
{1, 2, · · · , r − 1} can be split into pairs {ai , bi } (1 ≤ i ≤ r−12 ) such that
ai + bi is prime for each i. Now, {r, r + 1, · · · , 2n} is evidently split into the
pairs {r, 2n}, {r + 1, 2n − 1}, · · · whose sums are all equal to the prime p.
Another very interesting application is the following one. By refining the
above methods, one may prove that for any positive integer k, there is
a sufficiently large N such that there is a prime between n and 2n − k
for all n > N . Applying this to k = 11, Robert Dressler showed in 1972
that every positive integer other than 1,2,4,6,9 is a sum of distinct odd
primes.

44
 How Far Apart are Primes? Bertrand’s Postulate

References
[1] W H Mills, A prime-representing function, Bull. Amer. Math. Soc. 53,
pp. 604, 1947.
[2] I Niven and D Zuckermann, Introduction to number theory, John Wiley
and Sons, New York, 1960.
[3] S Ramasubramanian, Mathematical Analysis, Echoes from Resonance,
Universities Press, Hyderabad, 2001.
[4] T Apostol, Introduction to analytic number theory, Springer Interna-
tional Students Edition, Narosa Publishers, New Delhi, 1986.
[5] American Math. Monthly, 1975.

45
Sums of Powers, Bernoulli
and the Riemann Zeta function

Bernoulli truly stunned us with his number;

woke us up from a deep and ignorant slumber.
Its relation with Riemann zeta
makes us think nothing could be neater.
The connection is much deeper – ask any plumber!

1. Introduction
It is a beautiful discovery
Pn k due to Jakob Bernoulli that for any positive
integer k, the sum i=1 i can be evaluated in terms of, what are now
known as, Bernoulli numbers. It is said that the Bernoulli numbers were
discovered simultaneously and independently by Japanese mathematician
Seki Kowa. Seki Kowa’s discovery was posthumously published in 1712, in
his work Katsuyo Sampo. In this article, we shall discuss several methods
of evaluating the above sum. Apart from Bernoulli’s method which we
shall recall, we give a method akin to using integration, and one using
differentiation. These methods are often useful in evaluating more general
sums too as we shall indicate. We also discuss connections with the Riemenn
zeta function – some old and some new.

2. Bernoulli Polynomials and Numbers

To motivate the introduction of the Bernoulli polynomials, letPus start
n
ik
with the sum that we want to evaluate viz., ni=1 ik . Evidently, i=1
P
k! is
the coefficient of xk in the power series expansion of ex + e2x + · · · + enx .
In other words,
e(n+1)x − 1 X 1k + 2k + · · · + nk
= 1 + xk .
ex − 1 k!
k≥0
Now, the function exx−1 can be represented by a power series exx−1 =
P xr
r≥0 Br r! . The numbers Br are known as Bernoulli numbers and it is
easy to evaluate them as follows.
r
Since the power series x and (ex − 1) r≥0 Br xr! agree in an interval
P
around 0, the numbers are determined recursively as
X r
B0 = 1, Bs = 0 ∀ r ≥ 2.
s<r
s

The chapter is a modified version of an article that first appeared in Resonance, Vol. 8, No. 7,
pp. 54–62, July 2003.

47
 Sums of Powers, Bernoulli and the Riemann Zeta function

The first few values are B0 = 1, B1 = −1/2, B2 = 1/6 and B3 = B5 =

B7 = · · · = 0.
tx
Now, consider the function Ft (x) = exe
x −1 for x 6= 0 and Ft (0) = 1. Once
k
again, Ft has a power series expansion Ft (x) = k≥0 Bk (t) xk! .
P
The functions Bk (t) are actually polynomials in t since
X xk x X xk
Bk (t) = Ft (x) = etx x = etx Bk
k! e −1 k!
k≥0 k≥0

and thus
k  
X k
Bk (t) = Bl tk−l .
l
l=0

Bk (t) are called Bernoulli polynomials; note that Bk (0) = Bk .

k k +···+nk
Returning to our sum, we have that 1 +2 k! is the coefficient of xk in
e(n+1)x −1 x(e(n+1)x −1)
ex −1 i.e., it is the coefficient of xk+1 in ex −1 = Fn+1 (x) − F0 (x).
k k k Bk+1 (n+1)−Bk+1 Pk k+1

Thus, 1 +2 k!
+···+n
= (k+1)! = 1
(k+1)! l=0 l Bl (n + 1)
k+1−l .

In other words,
k  
k k 1 X k+1k
1 + 2 + ··· + n = Bl (n + 1)k+1−l .
k+1 l
l=0

Note that it is evident from this formula that the sum of the k-th powers
of the first n natural numbers is a polynomial function of n of degree k + 1.

3. Method of ‘Integration’
For convenience, let us denote Sk (n) = 1k + 2k + · · · + nk . This is a
polynomial function of n i.e., there is a polynomial Sk (x) of degree k + 1
such that the above equality holds for all n.
The basic idea of the method we will discuss now is that (since nk = Sk (n)
− Sk (n − 1)), xk can be thought of as a ‘derivative’ of the function Sk (x). In
other words, Sk (x) itself may be thought of as an ‘integral’ of the function
xk . Of course, this is only heuristic at the moment because xk will be the
derivative of Sk at some point between x−1 and x. The correct tool to make
this precise is the ‘method of differences’ which is really a discrete analogue
of differentiation. More precisely, let us recall that the ‘backward difference’
operator is defined on any function f by (∇f )(x) = f (x) − f (x − 1) for all
x. It is trivial to see that if Pr (x) = x(x + 1) · · · (x + r − 1) for r ≥ 1 and
for all x, then (∇Pr )(x) = rPr−1 (x) for all x.

48
 Sums of Powers, Bernoulli and the Riemann Zeta function

Let us call g an anti-difference of f if ∆g = f . Note that if f is a poly-

nomial such that (∇f )(n) = 0 for infinitely many n, then f is a constant.
So, if f1 , f2 are polynomials with ∇f1 = ∇f2 , then f1 − f2 is a constant.
Let us look at our sums Sk (n) now. Let us keep in mind that the polyno-
mial Sk (x) has no constant term. Writing fk (x) = xk and gk (x) for any anti-
difference of fk which is a polynomial function, then we have (∇gk )(n) =
fk (n) = nk = Sk (n) − Sk (n − 1) = (∇Sk )(n) for all n ≥ 2.
Hence, Sk (x) = gk (x)+c for some constant c. Since Sk (x) has no constant
term, we have c = −gk (0).
In other words, Sk (n) = gk (n)−gk (0) for any anti-difference (polynomial)
function gk of fk .
Note the similarity with the fundamental theorem of calculus.
So, our problem reduces to finding an anti-difference of the function xk .
We observed earlier that the function Pr (x) = x(x + 1) · · · (x + r − 1) has an
anti-difference Pr+1 (x) k
r+1 . Therefore, it is just a matter of writing x in terms
of the Pr ’s.
For instance, k = 1 gives f1 (x) = x = P1 (x) so that g1 (x) can be taken
to be P22(x) = x(x+1) 2 so that S1 (n) = g1 (n) − g1 (0) = n(n+1)
2 .
2
For k = 2, one has f2 (x) = x = x(x + 1) − x = P2 (x) − P1 (x) so that g2
can be taken as g2 (x) = P33(x) − P22(x) = x(x+1)(x+2) 3 − x(x+1)
2 = x(x+1)(2x+1)
6 .
n(n+1)(2n+1)
This gives S2 (n) = 6 for all n.
The fact that one can indeed write xk as an integer linear combination
of Pk , Pk−1 , · · · , P1 can be seen as follows.
Now Pr (x) = x(x + 1) · · · (x + r − 1) = xr + ar−1,r xr−1 + · · · + a0,r for
some integers ai,r . Indeed, these integers are the symmetric polynomials in
1, 2, · · · , r − 1.
Then, we have the matrix equation AF = P where A is the upper trian-
gular integer matrix
 
1 ak−1,k ak−2,k · · · a0,k
0
 1 ak−2,k−1 · · · a0,k−1 

 · · · 

 ··· 
0 0 ··· 0 1,

F is the column vector (xk , xk−1 , · · · , x) and, P is the column vector

(Pk (x), Pk−1 (x), · · · , P1 (x)).
The matrix A has an inverse which is also an upper triangular integer
matrix B with 1’s on the diagonal.
Thus, F = BP gives the required expression.
Let us remark here that the above method is general enough to work
atleast for any complex polynomial function f instead of fk . Thus, to

49
 Sums of Powers, Bernoulli and the Riemann Zeta function

evaluate f (1) + · · · + f (n), one writes f as a linear combination of the

polynomials Pr , say,

f (x) = a0 + a1 P1 (x) + · · · + ad Pd (x)

where d = deg f and ai are complex numbers. Then, one has

n(n + 1) n(n + 1) · · · (n + d)
f (1) + · · · + f (n) = a0 n + a1 + · · · + ad .
2 d+1

4. A Method Involving Differentiation

This is an elementary and pretty useful method involving the differential
d
operator x dx .
d
Note that (x dx )xn = nxn . Therefore, applying it repetitively, one obtains
d k n
(x dx ) x = nk xn .
d k
Hence 1k + 2k + · · · + nk = (x dx ) (1 + x + x2 + · · · + xn ) at x = 1.
This can be rewritten in a more convenient form as
n
X d k xn+1 − 1
ik = limx→1 (x ) .
dx x−1
i=1

Riemann Zeta Function

We end with some remarks on the sums of the infinite series n≥1 n1k
P
for integers k ≥ 2. This is a special value ofP the so-called Riemann Zeta
function ζ(s) defined as the sum of the series n≥1 n1s for any real number
s > 1 (actually, it can be defined as a complex valued function for any
complex number s with Re s > 1 by the same series).
2 4 π6
Some of the values are ζ(2) = π6 , ζ(4) = π90 , ζ(6) = 945 .
The reader will notice that we have not written ζ(k) for any odd value of
k and that, for even k, the value seems to be a rational multiple of π k . In
fact, the value ζ(3) is known to be irrational but it is still unkown if it can
be expressed in terms of ‘known’ constants! We shall show now that ζ(2k)
is indeed a rational multiple of π 2k for any natural number k. In fact, the
Bernoulli numbers will surface here again! The reasons for not being able
to evaluate ζ at odd values (or even say whether it is irrational in general)
are deep and we do not go into them here.
2
Now, for any complex number z, we have Sin z = z n≥1 (1 − nz2 π2 ).
Q
Its logarithmic derivative gives us
X z2 X X 1 z 2k
z Cot z = 1 + 2 = 1 − 2 .
z 2 − n2 π 2 n2k π 2k
n≥1 n≥1 k≥1

50
 Sums of Powers, Bernoulli and the Riemann Zeta function

On the other hand, in the definition of the Bernoulli numbers as exx−1 =

P xr
r≥0 Br r! , if we put x = 2iz, we obtain (recalling that B2r+1 = 0 for
r ≥ 1),
X 22k z 2k
z Cot z = 1 − (−1)k−1 B2k .
(2k)!
k≥1

Comparing the two expressions, we obtain

22k−1 2k
ζ(2k) = (−1)k−1 B2k π .
(2k)!

We remark in passing that the same cotangent series is the starting point
for obtaining an expression via theta series for the number of ways of writing
a positive integer as a sum of squares.
Here is a rather surprising observation. The Riemann zeta function ζ(s) is
−s
P
defined by the series n≥1 n for any complex number with Re s > 1. The
theory of the zeta function implies that its definition can be extended (not
by the same series, of course) to all values of s other than s = 1. Moreover,
the values at s and 1 − s are related by what is known as a functional
equation (thus there is the mysterious half line Re s = 1/2 in the middle
on which the Riemann hypothesis predicts all the nontrivial zeroes of ζ(s)
ought to lie). Let us now think of the naiveP idea that since ζ(k) for any
natural number k > 1 is given by the series n≥1 n−k , it is possible that
the value ζ(−k) is related to the partial sums n≤N nk . That this is indeed
P
so is a simple, beautiful observation due to J Minac [1]). Recall from the
previous discussion that there is a unique polynomial Sk (x) which coincides
with the sum 1k + · · · + nk at x = n for any natural number n and that Sk
has degree k + 1. In fact, we saw that

Bk+1 (x + 1) − Bk+1 (1)

Sk (x) = .
k+1
0 (x) = mB
As Bm m−1 (x) for all m, we see

1 1
Bk+1 (x + 1) − Bk+1 (1)
Z Z
Bk+1
Sk (x − 1)dx = = (−1)k .
0 0 k+1 k+1

We claim: Z 1
Bk+1
ζ(−k) = Sk (x − 1)dx = (−1)k .
0 k+1
Actually, one can use the functional equation for the zeta function to
conclude this but we follow a more elementary method of obtaining analytic
continuation of the zeta function which will also prove this claim.

51
 Sums of Powers, Bernoulli and the Riemann Zeta function

The analytic continuation of the zeta function to all s 6= 1 and the

fact that lims→1 (s − 1)ζ(s) = 1 are obtainable as follows. Now,Pthe zeta
function ζ(s) is defined for a complex variable s by the series ∞ n=1 n
−s

which converges for Re s) > 1. We shall use Abel’s partial summation

formula which is an elementary yet very powerful formula – the readers are
well aware of its continuous analogue – integration by parts.
If {an }, {bn } are two sequences of complex numbers, and if An = a1 +
· · · + an , then we have the identity
n
X
a1 b1 + · · · + an bn = An bn+1 − Ak (bk+1 − bk ).
k=1
P
Thus,
P∞ n an bn converges if both the sequence {An bn+1 } and the series
k=1 Ak (bk+1 − bk ) converge.
The proof follows simply by observing that
n
X n
X n
X n
X
ak bk = (Ak − Ak−1 )bk − Ak bk − Ak bk+1 + An bn+1 .
k=1 k=1 k=1 k=1

In our case, by using Abel’s partial summation formula, one has

Z ∞ Z ∞
[x] s {x}
ζ(s) = s s+1
dx = −s dx.
1 x s−1 1 xs+1
Here [x] and {x} respectively denote the integral part and the fractional
part of x. Note that the integral converges for Re (s) > 0 and thus the last
expression gives the analytic continuation of the zeta function to the region
Re (s) > 0. We shall proceed inductively now. On writing
∞ ∞ n+1 ∞ 1
{x} x−n X
Z Z Z
X udu
dx = =
1 xs+1 xs+1 (u + n)s+1
n=1 n n=1 0

and integrating the last integral by parts, we obtain

s(s + 1) ∞ {x}2
Z
s s
ζ(s) = − (ζ(s + 1) − 1) − dx.
s−1 2 2 1 xs+2
From this, we have analytic continuation of ζ for Re (s) > −1 and also
that ζ(0) = − 21 . Proceeding inductively, we get
m
X s(s + 1) · · · (s + q − 1)
1
ζ(s) = 1 + − (ζ(s + q) − 1)
s−1 (q + 1)!
q=1
∞ 1
s(s + 1) · · · (s + m) X um+1
Z
− .
(m + 1)! (u + n)s+m+1
n=1 0

52
 Sums of Powers, Bernoulli and the Riemann Zeta function

The infinite sum on the right hand side converges for Re (s) > −m and
thus we have an expression for ζ(s) for such s. At this point, we evaluate
it at s = 1 − m. Rather surprisingly, this pretty but simple idea does not
seem to have been thought of until very recently when it was done so by
Ram Murty and M Reece. We get
m−1
(−1)m
 
1 q m−1 1
X
ζ(1 − m) = 1 − + − (−1) (ζ(1 − m + q) − 1).
m m(m + 1) q q+1
q=1

The first few values at nonpositive integers are

1 1 1
ζ(0) = − , ζ(−1) = − , ζ(−2) = 0, ζ(−3) = .
2 12 120
On the other hand, for M = 1, 2, 3, · · · , we have
k  
k+1
X k+1
M = (−1)r Sk−r (M − 1).
r+1
r=0

Therefore, we get
k  Z 1
(−1)k+1

X k+1
(−1)r Sk−r (x − 1)dx = .
r+1 0 k+2
r=0

As ζ(0) = − 12 , we arrive at the formula

Z 1
Bk+1
ζ(−k) = Sk (x − 1)dx = (−1)k
0 k+1

which was claimed.

Let us finally remark that the Riemann zeta function vanishes at the
negative even integers −2, −4, −6, · · · and these are its so-called trivial
zeroes. The Riemann hypothesis asserts that all other zeroes lie on the line
Re(s) = 1/2.

References
[1] J Minac, Expo. Math., Vol. 12, pp. 459–462, 1994.

53
Frobenius and His Density Theorem
for Primes

A rare sight is seen; yes,

when we spot a genius.
We saw one who made sense
of prime numbers being dense.
This was the great George Frobenius!

1. Introduction
Our starting point is the following problem which appeared in the recent
IMO (International Mathematical Olympiad):
If p is a prime number, show that there is another prime number q such
that np − p is not a multiple of q for any natural number n.
Now, this problem itself can be solved using elementary mathematics
(otherwise, it would not be posed in the IMO). However, how does one
guess that such a thing ought to be true? Can we produce an abundance of
such problems in some systematic manner? We take this problem as a point
of reference to discuss some deep number theory (which is already a century
old) which not only solves this problem, but also gives us an understanding
of why such facts are true and what more one can expect. The main theorem
under discussion is known as Frobenius’s density theorem.

2. Rephrasing and Generalisation

Let us start by rephrasing the above problem. For a prime p, consider the
integral polynomial f (X) = X p −p. For any prime q, one may consider f as
a polynomial over Z/qZ, the integers modulo q by reducing the coefficients
of f modulo q. Then, the problem asks us to prove that there is some prime
q for which f does not have a root in Z/qZ. So, when is it true that an
integral polynomial has roots in Z/qZ for every prime p? Obviously, if the
integral polynomial already has an integral root, this happens. Can it also
happen when f has no integral root?
Before answering this, let us note that every nonconstant integral poly-
nomial has a root in Z/qZ for infinitely many primes q. Here is the simple
argument proving it. See also ‘Polynomials with integer values’, second
chapter in this book.
The chapter is a modified version of an article that first appeared in Resonance, Vol. 8, No. 12,
pp. 33–41, December 2003.

55
 Frobenius and His Density Theorem for Primes

Let P (X) = a0 + a1 X + · · · + an X n be an integral polynomial with n > 0

and an 6= 0. For any integer d, look at the polynomial

P (a0 dX) = a0 (1 + a1 dX + a0 a2 d2 X 2 + · · · + a0n−1 an dn X n ).

Since Q(X) = 1 + a1 dX + a0 a2 d2 X 2 + · · · + an−1

0 an dn X n takes the values
0, 1, −1 at the most for finitely many values of X, it takes a value Q(m)
6= 0, 1, −1 which must then be a multiple of some prime p. As Q(m) ≡ 1
mod d, p is coprime to d. Therefore, for any d, we have shown that there
is some m such that P (a0 dm) is zero modulo p for some prime p coprime
to d. Varying d, we have infinitely many such primes p.
The set of odd primes modulo which the polynomial X 2 + 1 has roots,
consists precisely of all primes in the arithmetic progression 4n + 1. In
general, every quadratic polynomial has a corresponding arithmetic pro-
gression such that the polynomial has roots modulo each prime in this
progression, and modulo no other primes. This follows from the famous
quadratic reciprocity law.
Returning to our case f (X) = X p −p, let us see whether we can explicitly
get an infinite set of primes modulo which f does have roots. Consider
any prime q and the group (Z/qZ)∗ of nonzero integers modulo q under
multiplication modulo q. If the p-th power map θ : a 7→ ap on (Z/qZ)∗ is
not 1 − 1, then there exists some a 6= 1 with ap = 1. Since aq−1 = 1, we
must have (q −p 1) . In other words, whenever q 6≡ 1 mod p, our polynomial f
has a root modulo q.
Let us now return to the possibility of producing a polynomial which
has no integral roots but has roots modulo every integer. Consider the
polynomial
g(X) = (X 2 − 13)(X 2 − 17)(X 2 − 221).
√ √ √
Evidently, its roots ± 13, ± 17, ± 221 are not integral (or even ratio-
nal). We show that it has roots modulo any nonzero integer. Recall that the
Chinese remainder theorem tells us that whenever m1 , · · · , mr are pairwise
coprime integers and a1 , · · · , ar are any integers, there is an integer a which
is simultaneously ≡ ai mod mi for i = 1, · · · , r. Therefore, by the Chinese
remainder theorem, it suffices to prove that g has roots modulo every prime
power. In what follows, for any prime p and a coprime to p, the notation
( ap ) stands for 1 or −1 according as whether a is a square or not mod p.
One also says in the respective cases that a is a quadratic residue modulo
p and a is a quadratic nonresidue modulo p.
Let us look at g now. If p is an odd prime such that ( 13 p ) = 1, then
t ≡ 13 mod p for some integer t. We show by induction on n that x2 ≡ 13
2

mod pn has a solution. Suppose t2 ≡ 13 mod pn−1 , say t2 = 13 + upn−1 .

Consider t0 = t + pn−1 t1 where we shall choose t1 so that t20 ≡ 13 mod pn .

56
 Frobenius and His Density Theorem for Primes

This requires u + 2tt1 ≡ 0 mod p; such a choice of t1 can be made since

2t is coprime to p. Thus, we have shown that if 13 is a quadratic residue
modulo an odd prime p, the polynomial g has a root modulo any power
pn . The same argument works if 17 or 221 is a quadratic residue modulo a
prime p. For powers of 2 we note that 17 ≡ 32 mod 23 and work as above
but with a minor change; we try t + 2n−2 t1 instead of the n − 1st power.
Note that 13 ≡ 82 mod 17 and 17 ≡ 22 mod 13. Further, for any p, one
of 13, 17 or 221 is a square modulo p. This is because the homomorphism
x 7→ x2 on (Z/pZ)∗ for an odd prime p, has kernel of order 2. Its image,
which is the subgroup of squares, is the unique subgroup of index 2. Hence
the cosets of 13 and 17 multiply to give the coset of 221. Thus, the above
argument goes through for all p and it follows that the polynomial g, indeed,
has roots modulo any nonzero integer.
Now, let us ask ourselves what is different about X p − p in comparison
with the above example. It is immediately evident that g is a reducible
polynomial over Z while the famous Eisenstein criterion shows that the
polynomial f (X) = X p − p is irreducible over Z. In fact, irreducibility of f
can be proved quite easily even without the Eisenstein criterion.
Ok, but let us look at another obvious irreducible polynomial over Z -
the linear polynomial h(X) = aX + b where (a, b) = 1. But, if p is any
prime not dividing a, then aX + b has a root modulo p. In other words,
h does have a root modulo all but finitely many primes even though it is
irreducible over Q. Thus, a reasonable guess for us could be:
(∗) An integral polynomial which is irreducible over Z and has degree
> 1 cannot have roots modulo all but finitely many primes. In other words,
for such a polynomial, there are infinitely many primes modulo which the
polynomial has no roots.
Our intention is to show that this is true. In fact, one may wonder maybe
whether we have the much stronger result that an irreducible polynomial
f over Z remains irreducible over all but finitely many primes. But, the
following example dashes this hope. It was already observed by Hilbert
(this example was already discussed in an earlier chapter).
Let p, q be odd prime numbers such that ( pq ) = ( pq ) = 1 and p ≡ 1 mod 8.
Then, the polynomial P (X) = (X 2 − p − q)2 − 4pq is irreducible whereas it
is reducible modulo any integer.
Now
P (X) = X 4 − 2(p + q)X 2 + (p − q)2
√ √ √ √ √ √ √ √
= (X − p − q)(X + p + q)(X − p + q)(X + p − q).
√ √ √ √ √
Since p, q, p ± q, pq are all irrational, none of the linear or

57
 Frobenius and His Density Theorem for Primes

quadratic factors of P (X) are in Z[X] i.e. P (X) is irreducible over Z.

Note, as before, that it is enough to show that a factorisation of P exists
modulo any prime power as we can use Chinese reminder theorem to get
a factorisation modulo a general integer. Now, P (X) can be written in the
following ways:

P (X) = X 4 − 2(p + q)X 2 + (p − q)2

= (X 2 + p − q)2 − 4pX 2
= (X 2 − p + q)2 − 4qX 2
= (X 2 − p − q)2 − 4pq.

The second and third equalities above show that P (X) is reducible mod-
ulo any q n and any pn respectively. Also since p ≡ 1 mod 8, p is a quadratic
residue modulo 2 and, therefore, modulo any 2n ; the second equality above
again shows that P (X) is the difference of two squares modulo 2n , and
hence reducible mod 2n .
If ` is a prime 6= 2, p, q, at least one of ( p` ), ( q` ) and ( pq
` ) is 1 by the product
formula ( ` ) · ( ` ) · ( ` ) = 1 that we noted earlier. According as ( p` ), ( q` ) or
p q pq

( pq
` ) = 1, the second, third or fourth equality shows that P (X) is reducible
mod `n for any n.
We mention, in passing, a very simple but important general method of
proving the irreducibility of an integral polynomial. This will also set up
the notation for our main statement when we address (∗). To illustrate it,
consider the polynomial p(X) = X 4 + 3X 2 + 7X + 4. Modulo 2, we have
p(X) = X(X 3 + X + 1) and both factors are irreducible over the field
Z/2Z. We say that the decomposition type of p(X) mod 2 is 1, 3. Therefore,
either p is irreducible over Z or if not, it is a product of a linear factor
and an irreducible factor of degree 3 over Z. But, modulo 11, we have
p(X) = (X 2 + 5X − 1)(X 2 − 5X − 4) where both factors are irreducible
over the field Z/11Z. That is, the decomposition type of p mod 11 is 2, 2.
Thus, it cannot be that p has a linear factor over Z. In other words, p must
be irreducible over Z.

3. The Notion of Density of Primes

Let us get back to the guess quoted above as (∗). We need some notations.
Let f be a monic integral polynomial of degree n. Suppose that f has dis-
tinct roots α1 , · · · , αn ∈ C; equivalently, the discriminant disc(f ) 6= 0. Let
K = Q(α1 , · · · , αn ), the subfield of C generated by the roots; that is, all
rational expressions in the αi ’s with coefficients from Q. This is the smallest
subfield of C which contains all the αi ’s; it is also known
Qn as the splitting field
of f for the reason that f splits into the product i=1 (X − αi ) over K. We

58
 Frobenius and His Density Theorem for Primes

look at the group G of those permutations of αi ’s which give rise to a field

automorphism of K. This is known as the Galois group of f and denoted
by Gal(f ). For instance, if f (X) = X 2 − a for some nonsquare integer a,
√ √
then K = Q( a) where a denotes a square root of a in C and G has two
√ √
elements I, σ where σ interchanges a and − a. In general, although G is
a subgroup of Sn , the permutations which belong to G are rather restricted;
for example if f is irreducible over Q, then a permutation in G is neces-
sarily transitive on the αi ’s. If p 6 |disc(f ), then the decomposition type of
f modulo p gives a partition of n as we saw above. On the other hand,
each element of G has a cycle decomposition as an element of Sn and, thus
defines a partition of n as well. Frobenius’s wonderful idea is to relate the
numbers of such partitions for a particular type. This will be expressed in
terms of a notion of density of a set of prime numbers. P
1/p
A set S of primes is said to have density δ if P p∈S 1/p → δ as s → 1+ .
all p
Here 1+ means the limit when s tends to 1 from the right. For instance,
any finite set of primes has density 0. Using this notion of density, we state
more precisely:

4. Frobenius Density Theorem

The set of primes p modulo which a monic integral, irreducible polyno-
mial f has a given decomposition type n1 , n2 , · · · , nr , has density equal to
N/O(Gal(f )) where N = |{σ ∈ Gal(f ) : σ has a cycle pattern n1 , n2 , · · · , nr }|.
As we point out now, our guess (∗) is vindicated by this theorem and a
little bit of group theory; in particular, this also solves the IMO problem.
If f is irreducible, and has roots modulo all but finitely many primes,
then the theorem shows that each σ has a cycle pattern of the form 1, n2 , · · ·
This means that each element of Gal(f ) fixes a root. Since the roots of f
are transitively moved around by Gal(f ), this group would be the union of
the conjugates of its subgroup H consisting of elements which fix a root of
f , say α1 . However, it is an elementary exercise that a finite group cannot
be the union of conjugates of a proper subgroup. Thus, in our case H =
Gal(f ). This means that Gal(f ) fixes each αi and is therefore trivial. That
is, f is linear.
A famous theorem of Dirichlet on primes in arithmetic progressions as-
serts that the density of the set of primes p ≡ a mod n is 1/φ(n) for any
(a, n) = 1. Dirichlet’s theorem implies Frobenius’s theorem for the polyno-
mial f (X) = X n − 1. The converse conclusion cannot quite be made. Thus,
Frobenius formulated a conjecture which generalises both his theorem and
Dirichlet’s theorem. This was proved 42 years later by Chebotarev and is
known now as the Chebotarev density theorem. This is an extremely useful
result and even effective versions are known (see the end of the article for

59
 Frobenius and His Density Theorem for Primes

what the word ‘effective’ means here). Chebotarev’s idea of proving this
has been described by two prominent mathematicians as “a spark from
heaven”. In fact, this theorem was proved in 1922 (“while carrying water
from the lower part of town to the higher part, or buckets of cabbages to
the market, which my mother sold to feed the entire family”) and Emil
Artin wrote to Hasse in 1925: “Did you read Chebotarev’s paper? ... If
it is correct, then one surely has the general abelian reciprocity laws in
one’s pocket...” Artin found the proof of the general reciprocity law in 1927
using Chebotarev’s technique (he had already boldly published the reci-
procity law in 1923 but admitted that he had no proof). Nowadays, Artin’s
reciprocity law is proved in some other way and Chebotarev’s theorem is
deduced from it!
To state Chebotarev’s theorem, we recall one notion – the Frobenius
map. The idea is that given a monic integral polynomial f and its split-
ting field K, we can associate to any prime p 6 |disc(f ), an element Φp
of Gal(f ) in a natural manner. If we can do this, one may expect that
the decomposition type of f modulo p coincides with the cycle pattern
of Φp . It can almost be done except that a prime p gives rise to a con-
jugacy class of elements in Gal(f ). We do not define the Frobenius con-
jugacy class here as it is somewhat technical and merely explain some
properties it has. For any prime number p, the p-th power map F robp
is an automorphism of the field F¯p which is identity on Fp . Therefore,
F robp permutes the roots of any polynomial over Fp . Indeed, the Ga-
lois theory of finite fields amounts to the statement that if g is a poly-
nomial over Fp with simple roots, then the cycle pattern of F robp viewed
as a permutation of the roots of g coincides with the decomposition type
of g over Fp . In our case, we start with an integral polynomial f and
look at it modulo p for various primes p. The basic theory of algebraic
numbers shows that whenever p 6 |disc(f ), the automorphism F robp gives
rise to a conjugacy class in Gal(f ), called the Frobenius conjugacy class
modulo p.
In Frobenius’s density theorem, one cannot distinguish between two primes
p, q defining different conjugacy classes C(x) and C(y) but some powers of
x and y are conjugate. For instance, for the polynomial X 10 −1, the decom-
position type modulo primes congruent to 1, 3, 7, 9 mod 10 are, respectively,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1; 1, 1, 4, 4; 1, 1, 4, 4; 1, 1, 2, 2, 2, 2.
Frobenius’s theorem cannot distinguish between primes which are 3 mod
10 and those which are 7 mod 10; they define different conjugacy classes
in Gal(X 10 − 1). Thus, it would imply that the number of primes ≡ 3 or 7
mod 10 is infinite but doesn’t say whether each congruence class contains
infinitely many primes. This is what Chebotarev’s theorem asserts.

60
 Frobenius and His Density Theorem for Primes

5. Chebotarev’s Density Theorem

Let f be monic integral and assume that disc (f ) does not vanish. Let C
be a conjugacy class of Gal(f ). Then, the set of primes p not dividing disc
|C|
(f ) for which σp ∈ C, has a well-defined density which equals |G| .
We state here without proof two results which can be proved with the
aid of Chebotarev’s density theorem. These concrete applications are:
(I) The set of primes which are expressible in the form 3x2 + xy + 4y 2
for integers x, y, has density 1/5.
(II) The set of primes p for which the decimal expansion of 1/p has odd
period, has density 1/3.
Finally, we end with the remark that a recent result due to Berend and
Bilu [1]) gives an ‘effective version’ of Chebotarev’s theorem. This means
in simple terms that given a nonconstant integral polynomial, one has a
certain number N , explicitly determined in terms of the irreducible factors
of f and their coefficients, so that f will have an integral root if, and only
if, it has a root modulo N . See also [2] for a nice historical introduction to
Frobenius’s and Chebotarev’s density theorems.
We conclude by stating that an integral polynomial of degree n is irre-
ducible but reducible modulo all positive integers if and only if, the cor-
responding Galois group has no element of order n. The proof uses the
Chebotarev density theorem.

References
[1] D Berend and Y Bilu, Polynomials with roots modulo every integer,
Proc. Amer. Math. Soc., Vol. 124, pp. 1663–1671, 1996.
[2] P Stevenhagen and H W Lenstra, Jr., Chebotarev and his density
theorem, Mathematical Intelligencer, Vol. 18, pp. 26–37, 1996.

61
When is a Decimal Expansion
Irrational?
√
Everyone learns in school that 2 is irrational. This, along with Euclid’s
proof of infinitude of primes, is probably the first time she encounters√a
proof by contradiction. Most students know in school that the value of 2
is approximately 1.414 but, more often than not, this aspect is not pursued
further in detail. The decimal 0.999 · · · where 9 recurs indefinitely is under-
stood (after some persuasion perhaps) to be none else than the number 1.
The problem here is that the concept of limit takes some time to sink in.
Given that start, they can see easily that numbers like 0.142857 where
the digits overlined recur indefinitely, are rational. Even if the recurring
string occurs after an initial string (for example, a decimal expansion like
27.142857), it still gives us a rational value only because it is the sum of a
geometric series with a ratio of the form 10k .
It is not hard to prove that this is a necessary condition as well, that
is, a decimal expansion of a real number represents a rational number if,
and only if, after the decimal place, there is a finite (possibly empty) string
after which the digits consist of a finite string (possibly consisting entirely
of zeroes) recurring indefinitely.
Thus, for instance, the number 0.101001000 · · · where the number of
zeroes keeps increasing by 1 has to represent an irrational number. However,
from the decimal expansion, sometimes it is not clear whether there is such
an eventual recurrence or not. This could be due to our present state of
knowledge. For instance, one could define a number 0.1010 · · · where the
number of zeroes occurring at the n-th step is either increased by one or
n
kept the same according as to whether the number 22 + 1 is prime or not.
Since one does not know whether there are infinitely many such primes, we
cannot say at present as to whether the above decimal represents a rational
or an irrational number.
The decimal 0.1234 · · · , where the natural numbers are written in se-
quence, is clearly irrational since, for instance, the number of zeroes occur-
ring in the powers of 10 keeps increasing. The decimal 0.235711 · · · , where
the set of primes is written down in sequence, is also irrational. This is
because there is a prime of the form 10n a + 1 for an arbitrary n – this
was proved in an earlier chapter. Here is another elementary general result
(see [1]).

The chapter is a modified version of an article that first appeared in Resonance, Vol. 9, No. 3,
pp. 78–80, March 2004.

63
 When is a Decimal Expansion Irrational?

Consider a decimal x = 0.a1 a2 · · · where {an } is aPstrictly increasing

r
sequence of natural numbers having the property that n ans diverges for
n
some r, s > 0. Then x is irrational.
Note that since the reciprocals of primes do not sum to a finite quantity,
this result also implies that 0.23571113 · · · is irrational.
To prove this, assume, if possible, that x is rational. Then, by throwing
out some of the first a’s and scaling, we may assume that the decimal is
actually periodic and not just eventually periodic. Let t denote a period.
Let N1 < N2 < · · · denote the natural numbers representing the different
numbers of digits occurring for the ai ’s. Let di denote the number of ai ’s
which have exactly Ni digits. In other words, a1 , · · · , ad1 are the a’s with
N1 digits and, for each i ≥ 1, the numbers
ad1 +···+di +1 , · · · , ad1 +···+di +di+1
are the a’s which have exactly Ni+1 digits. Let us write for simplicity d0 = 0.
Now, if some di+1 were bigger than t, then the numbers
ad1 +···+di +1 , · · · , ad1 +···+di +t+1 ,
would all have Ni+1 digits and since the length of the string
(ad1 +···+di +1 ) · · · (ad1 +···+di +t ),
is tNi+1 which is a multiple of t, it follows that
ad1 +···+di +1 = ad1 +···+di +t+1 ,
which is a manifest contradiction of the assumption that {an } is an increas-
ing sequence. Hence, we have shown that each di is ≤ t.
Now, we also have the evident inequalities
ad1 +···+di +di+1 ≥ · · · ≥ ad1 +···+di +1 ≥ 10Ni+1 −1 ,
P∞ nr
since these numbers have Ni+1 digits. We shall show that n=1 asn con-
verges. Now, for each i ≥ 0,
d1 +···+di +di+1 d1 +···+di +di+1
X jr X (d1 + · · · + di+1 )r di+1 (d1 + · · · + di+1 )r
≤ ≤ .
asj 10s(Ni+1 −1 ) 10s(Ni+1 −1 )
j=d1 +···+di +1 j=d1 +···+di +1
P Pd1 +···+di +di+1 jr P t(i+1)r tr tr+1 P (i+1)r
Thus, i≥0 j=d1 +···+di +1 asj ≤ i≥0 10s(Ni+1 −1) ≤ 10s(N1 −1) i≥0 10is .
nr
that ∞
P
But, the last series converges so n=1 asn also converges. This is a
contradiction of our assumption and the irrationality of x follows.

References
[1] A McD Mercer, American Mathematical Monthly, Vol. 101,
pp. 567–568, 1994.

64
Revisiting Kummer’s
and Legendre’s Formulae

In a beginning course in number theory, an elementary exercise is to

compute the largest power of a prime p dividing n!. This number, called
the p-adic valuation of n!, is easily proved to be
     
n n n
vp (n!) = + 2 + 3 + ··· . (1)
p p p
Note that this is a finite series. The number vp (n!) comes up naturally
in a few situations like the following. In the group of permutations of n
objects, this would give the power of p which is the order of a p-Sylow
subgroup. While discussing p-adic numbers as analogues of the usual real
numbers, one looks at the analogue of the exponential series. The expression
for vp (n!) leads one to determine that the exponential series has the radius
of convergence p−1/(p−1) .
Now, vp (n!) can also be computed in another manner by a beautiful ob-
servation due to the legendary mathematician Legendre. Legendre observed
that the p-adic valuation of n! can be read off from the base-p expansion
of n. It is simply n−s(n)
p−1 where s(n) is the sum of the “digits” of n in this
expansion. A related result that Kummer proved
is that, if r ≤ n, then
the p-adic valuation of the binomial coefficient nr is simply the number
of ‘carry-overs’ when one adds r and n − r in base-p. In ([1], pp.229–233)
Honsberger deduces Kummer’s theorem from Legendre’s result and refers
to Ribenboim’s lovely book [2], pp.30–32) for a proof of the latter. Riben-
boim’s proof is by verifying that Legendre’s base-p formula agrees with the
standard formula (1).
Is it possible to prove Legendre’s formula without recourse to the above
formula? We shall see that this is indeed possible and that the standard
formula above follows from such a proof. What is more, Kummer’s formula
also follows without having to use Legendre’s result. Let us start by recalling
Legendre’s formula.

Legendre’s Formula
Let p be a prime number and let ak · · · a1 a0 be the base-p expansion of
a natural number n. We shall show that if Legendre’s formula
n − ki=0 ai
P
n − s(n)
vp (n!) = = (2)
p−1 p−1
The chapter is a modified version of an article that first appeared in Resonance, Vol. 10, No. 2,
pp. 62–71, February 2005.

65
 Revisiting Kummer’s and Legendre’s Formulae

holds good for n, then it also holds good for pn + r for any 0 ≤ r < p. Note
that the base-p expansion of pn + r is

ak · · · a1 a0 r.
m−s(m)
Let us denote, for convenience, the number p−1 by f (m) for any
natural number m. Evidently,

pn − ki=0 ai
P
f (pn + r) = = n + f (n).
p−1
On the other hand, it follows by induction on n that

vp ((pn + r)!) = n + vp (n!). (3)

For, if it holds good for all n < m, then

vp ((pm + r)!) = vp (pm) + vp ((pm − p)!)

= 1 + vp (m) + m − 1 + vp ((m − 1)!) = m + vp (m!).

Since it is evident that f (m) = 0 = vp (m!) for all m < p, it follows that
f (n) = vp (n!) for all n. This proves Legendre’s formula.
Note also that the formula
     
n n n
vp (n!) = + 2 + 3 + ···
p p p

follows inductively on using (3).

Kummer’s Algorithm
As before p is any prime number. For any natural numbers r and s, let
us denote by g(r, s) the number of ‘carry-overs’ when the base-p expansions
of r and s are added. Kummer’s result is that for k ≤ n,
 
n
vp = g(k, n − k). (4)
k

Once again, this is clear if n < p, as both sides are then zero. We shall
show that if the formula holds good for n (and every k ≤ n), it does so for
pn + r for 0 ≤ r < p (and any k ≤ pn + r). This would prove the result for
all natural numbers.
pn+r


Consider any binomial coefficient pm+a for 0 ≤ a < p.
First, suppose a ≤ r.

66
 Revisiting Kummer’s and Legendre’s Formulae

Write m = bk · · · b0 and n − m = ck · · · c0 in base-p. Then the base-p

expansions of pm + a and p(n − m) + (r − a) are, respectively,

pm + a = bk · · · b0 a
p(n − m) + (r − a) = ck · · · c0 r − a.

Evidently, the corresponding number of carry-overs is

f (pm + a, p(n − m) + (r − a)) = f (m, n − m).

 
n

pn + r
By the induction hypothesis, f (m, n−m) = vp ( m ). Now vp
pm + a
is equal to

vp ((pn + r)!) − vp ((pm + a)!) − vp ((p(n − m) + r − a)!)

 
n
= n + vp (n!) − m − vp (m!) − (n − m) − vp ((n − m)!) = vp .
m

 a ≤ r.
Thus, we are through in the case when 
pn + r
Now suppose that r < a. Then vp is equal to
pm + a

vp ((pn + r)!) − vp ((pm + a)!) − vp ((p(n − m − 1) + (p + r − a))!)

= n + vp (n!) − m − vp (m!) − (n − m − 1) − vp ((n − m − 1)!)
= 1 + vp (n) + vp ((n − 1)!) − vp (m!) − vp ((n − m − 1)!)
 
n−1
= 1 + vp (n) + vp .
m

We need to show that

f (pm + a, p(n − m − 1) + (p + r − a)) = 1 + vp (n) + f (m, n − m − 1). (5)

Note that m < n. Write n = ak · · · a0 , m = bk · · · b0 and n − m − 1 =

ck · · · c0 in base-p. If vp (n) = d, then ai = 0 for i < d and ad 6= 0. In base-p,
we have
n = ak · · · ad 0 · · · 0
and, therefore,

n − 1 = ak · · · ad+1 ad − 1 p − 1 · · · p − 1.

Now, the addition m + (n − m − 1) = n − 1 gives bi + ci = p − 1 for i < d

(since they must be < 2p − 1). Moreover, bd + cd = ad − 1 or p + ad − 1.

67
 Revisiting Kummer’s and Legendre’s Formulae

Note the base-p expansions

pm + a = bk · · · b0 a,
p(n − m − 1) + (p + r − a) = ck · · · c0 p + r − a.

We add these using that fact that there is a carry-over in the beginning
and that 1 + bi + ci = p for i < d. Since there is a carry-over at the first
step as well as at the next d steps, we have

pn + r = ∗ ∗ · · · ad 0 · · · 0 r

where there are d zeroes before r, and

f (pm + a, p(n − m − 1) + (p + r − a)) = 1 + d + f (m, n − m − 1).

This proves Kummer’s assertion also.

We end with the remark that Kummer’s result gives an immediate proof
of the fact that the n-th Catalan number is odd if and only if n is 1 less
than a power of 2.

References
[1] R Honsberger, In Polya’s Footsteps, published and distributed by the
Mathematical Association of America, 1997.
[2] P Ribenboim, The Book of Prime Number Records, Springer-Verlag,
1996.

68
Bessels Contain Continued Fractions
of Progressions

1. Introduction
The January 2000 issue of Resonance carried a nice article on continued
fractions by Shailesh Shirali. After discussing various continued fractions
for numbers related to e, he left us with the intriguing question as to how
one could possibly evaluate the continued fraction
1 1 1
········· .
1+ 2+ 3+
The question is interesting because this continued fraction is simpler-
looking than the ones which were studied in that article. We answer this
question here and show that the discussion naturally involves the Bessel
functions, thus explaining the title. However, we shall begin with some
details about continued fractions which complement his discussion. One
place where continued fractions are known to appear naturally is in the
study of the so-erroneously-called Pell’s equation.
In a series of very interesting articles in Resonance, Amartya Kumar
Dutta had discussed various aspects of Mathematics in ancient India. In
particular, he discussed Brahmagupta’s and Bhaskara’s work on Samasab-
havana and the Chakravala method for finding solutions to ‘Pell’s equation’.
In fact, it is amusing to recall what Andre Weil, one of the great mathe-
maticians of the last century wrote once, while discussing Fermat’s writings
on the problem of finding integer solutions to x2 − Dy 2 = 1:

What would have been Fermat’s astonishment if some mission-

ary, just back from India, had told him that his problem had
been successfully tackled there by native mathematicians almost
six centuries earlier!

The Chakravala method can be described in terms of continued fractions.

Let us begin with some rather elementary things which were known so long
back and have gone out of fashion to such an extent that they are not as
widely known as they ought perhaps to be.

The chapter is a modified version of an article that first appeared in Resonance, Vol. 10, No. 3,
pp. 80–87, March 2005.

69
 Bessels Contain Continued Fractions of Progressions

2. Linear Diophantine Equations with SCF’s

Let us denote by
[a0 ; a1 , a2 , a3 , · · · ] (1)
the SCF (simple continued fraction)
1 1
a0 + ······ . (2)
a1 + a2 +
Here the ai are natural numbers. Evidently, any rational number has a
finite SCF. For instance,
763
= [1 ; 1, 12, 1, 1, 1, 9].
396
Its successive convergents are 11 , 21 , 25 27 52 79 763
13 , 14 , 27 , 41 , 396 . Note that if the n-
th convergent is pqnn , then pn qn−1 − pn−1 qn = (−1)n . This holds for any
continued fraction, as can be seen by induction. This gives a method of
finding all positive integral solutions (in particular, the smallest one) x, y
to a Diophantine equation of the form ax − by = c. For instance, consider
the equation
396x − 763y = 12.
763 79
Look at the SCF for 396 and compute its penultimate convergent 41 .
Now, if x, y are positive integers satisfying

396x − 763y = 12,

then combining with the fact that 396 × 79 − 763 × 41 = 1, we get

x − (79 × 12) = 763t, y − (41 × 12) = 396t

for some integer t. This gives all solutions, and the smallest solution in
natural numbers x, y is obtained by taking t = −1 and turns out to be
(185, 96).
The reader is left with deriving similarly the corresponding expression
for any linear equation.

3. Quadratic Equations from SCF’s

Evidently, finite CF’s give only rational numbers. Given the fact that
a periodic decimal expansion gives rational numbers too, a reader might
be tempted to guess that a periodic CF gives rationals. After just a lit-
tle thought, it becomes apparent that an eventually periodic SCF gives a
quadratic irrational number. For example, [1 ; 1, 1, · · · ] is the ‘golden ratio’

70
 Bessels Contain Continued Fractions of Progressions

√
(1 + 5)/2. This is because the value s satisfies p s = 1 + 1/s and is posi-
tive. Similarly, the SCF [1 ; 3, 2, 3, 2, · · · ] = 5/3, as it gives the√quadratic
equation s − 1 = (s + 1)/(3s + 4), and [0 ; 3, 2, 1, 3, 2, 1, · · · ] = ( 37 − 4)/7
as it gives the equation s = (3 + 2s)/(10 + 7s), etc.
Consider a quadratic Diophantine equation in two variables

ax2 + 2hxy + by 2 + 2gx + 2f y + c = 0 (3)

where a, b, c, f, g, h are integers. Thinking of this as a polynomial in x and

solving it, one obtains
p
ax + hy + g = ± (h2 − ab)y 2 + 2(hg − af )y + g 2 − ac).

For any integral solution, the expression inside the square root (which we
write as ry 2 + 2sy + t now) must be a perfect square, say v 2 . Once again,
solving this as a polynomial in y, we get
p
ry + s = ± (s2 − rt + rv 2 ).

Hence, s2 −rt+rv 2 must be a perfect square u2 . In other words, the origi-

nal equation does not have integral solutions unless the equation u2 −rv 2 =
w has a solution, where w is a constant defined in terms of a, b, c, f, g, h.
An equation of the form u2 + rv 2 = w for r positive has only finitely
many solutions. Therefore, let us discuss the equation u2 − rv 2 = ±w
where r, w are positive integers and r is not a perfect square. The SCF
√
for r provides a way of obtaining infinitely many solutions of the special
equation u2 − rv 2 = 1. Consequently, for given r, w if we find one solution
(u0 , v0 ) of u2 −rv 2 = w, one can find infinitely many by the samasabhavana
(composition) x = uu0 + rvv0 , y = uv0 + vu0 for any u, v with u2 − rv 2 =
1. However, the method of CF’s will provide even one solution only for
certain w’s; namely, those which appear as one of the denominators while
√
expressing r as a continued fraction.
Let us now show how u2 − rv 2 = 1 can always be solved in positive
√
integers using the SCF for r. It is a simple exercise to show that the SCF
√
for r has the form

[a1 ; b1 , b2 , · · · , bn , 2a1 , b1 , b2 , · · · , bn , 2a1 , · · · ]. (4)

If p/q is a penultimate convergent of a recurring period, then it is easy

to check that p2 − rq 2 = ±1. In fact, if the period is even, this is always 1.
If the period is odd, then the penultimate convergents of the first, second,
third period, . . . alternately satisfy the equations

x2 − ry 2 = −1, x2 − ry 2 = 1.

71
 Bessels Contain Continued Fractions of Progressions

For example, √
13 = [3 ; 1, 1, 1, 1, 6, · · · ].
The period is 5 which is odd. The penultimate convergent to the first
period is
1 1 1 1 18
3+ = .
1+ 1+ 1+ 1 5
Therefore, (18, 5) is a solution of u2 − 13v 2 = −1.
The penultimate convergent to the second period is computed to be
649/180. Therefore, (649, 180) is a solution of u2 − 13v 2 = 1.

4. SCF’s in Arithmetic Progressions

In his discussion, Shirali showed that the following SCF’s can be evalu-
ated in terms of the exponential function; he showed

e+1 e2 + 1
[2 ; 6, 10, 14, · · · ] = , [1 ; 3, 5, 7, · · · ] = .
e−1 e2 − 1
The SCF’s here involve terms in arithmetic progression. What about
a general SCF of the form [a ; a + d, a + 2d, · · · ]? For example, can the
SCF [0 ; 1, 2, 3, · · · ] be evaluated in terms of some ‘known’ numbers and
functions? Shirali started with the differential equation (1 − x)y 00 = 2y 0 + y
which he remarked “does not seem to solvable in closed form”.
We start with any arithmetic progression a, a + d, a + 2d, · · · where a is
any real number and d is any non-zero real number, and show how it can
be evaluated.
Let us consider the differential equation

dxy 00 + ay 0 = y. (5)

Actually, heuristic reasons can be given as to why one looks at this differ-
ential equation but we directly start with it here and show its relation to our
problem. Let y = y(x) be a solution of the above differential equation satis-
fying y(0) = ay 0 (0). Let us denote the r-th derivative of y by yr for simplicity
of notation. By repeated differentiation, we get dxyr+2 + (a + rd)yr+1 = yr
for all r ≥ 0 (with y0 denoting y). Therefore, we have

y0 dxy2 dx dx
=a+ =a+ ······ . (6)
y1 y1 a + d+ a + 2d+
Observe that
y(1/d)
[a ; a + d, a + 2d, a + 3d, · · · ] = .
y 0 (1/d)

72
 Bessels Contain Continued Fractions of Progressions

A solution function such as above can be very easily obtained as a series;

we get
X xn+1
y = c0 + c0 , (7)
(n + 1)!a(a + d) · · · (a + nd)
n≥1

for any c0 . This evaluates the SCF [a ; a + d, a + 2d, a + 3d, · · · ] in terms of

these series. As we shall see now, these series are special values of modified
Bessel functions and, for certain choices of a and d, the series are even
expressible in terms of e, etc.
Before proceeding further, let us note that the SCF whose evaluation was
asked for by Shirali is:
P
1/ ((n + 1)!n!)
[0 ; 1, 2, 3, · · · ] = . (8)
1/(n!)2
P

Its approximate value is 0.7.

For general a, d as above, the solution function
X xn+1
y = y(x) = c0 + c0 (9)
(n + 1)!a(a + d) · · · (a + nd)
n≥1

is related to Bessel functions in the following manner. First, the Bessel

differential equation x2 y 00 + xy 0 + (x2 − α2 )y = 0 has certain solutions
X (−1)n (x/2)2n+α
Jα (x) = ; (10)
n!Γ(n + 1 + α)
n≥0

these are usually referred to as Bessel functions of the first kind. Here Γ(s)
is the Gamma function. If α is not an integer, then J−α (defined in the
obvious manner) is another independent solution to the Bessel differential
equation above. Closely related to the Jα is the so-called modified Bessel
function of the first kind
X (x/2)2n+α
Iα (x) = . (11)
n!Γ(n + 1 + α)
n≥0

Thus, we have
1 1 1 I1 (2)
········· = .
1+ 2+ 3+ I0 (2)
The function Iα (x) is a solution of the differential equation x2 y 00 + xy 0 −
(x2 + α2 )y = 0. Indeed, Iα (x) = i−α Jα (ix) for each x. Using the relation
√
Γ(s + 1) = sΓ(s) and the value Γ(1/2) = π, it is easy to see that the
solution function

73
 Bessels Contain Continued Fractions of Progressions

X xn+1
y = c0 + c0 (12)
(n + 1)!a(a + d) · · · (a + nd)
n≥1

above, is related to the modified Bessel function of the first kind as:

y(x2 /d) = c0 Γ(a/d)(x/d)1−a/d Ia/d−1 (2x/d). (13)

In particular,

y(1/d) Ia/d−1 (2/d)

[a ; a + d, a + 2d, a + 3d, · · · ] = 0
= . (14)
y (1/d) Ia/d (2/d)

Conclusion
Before finishing, we recall some SCF’s evaluated out by Shirali:

e+1 e2 + 1
[2 ; 6, 10, 14, · · · ] = , [1 ; 3, 5, 7, · · · ] = .
e−1 e2 − 1
Our formula above yields for the same SCF’s the expressions:

I−1/2 (1/2)
[2 ; 6, 10, 14, · · · ] = , (15)
I1/2 (1/2)

I−1/2 (1)
[1 ; 3, 5, 7, · · · ] = . (16)
I1/2 (1)

It is clear from the definition that

r r
2 X 1 2 e + e−1
I−1/2 (1) = = , (17)
π (2n)! π 2
n≥0
r r
2 X 1 2 e − e−1
I1/2 (1) = = . (18)
π (2n + 1)! π 2
n≥0

Therefore, for these special parameters, the value of the modified Bessel
function is expressible in terms of e and one can recover Shirali’s expres-
sions.

74
The Prime Ordeal

Numbers in their prime

for no reason or rhyme
show up at a rhythm
with probability 1/logarithm.
If this is a law they knew,
they also break quite a few
but that is not a crime!
“There are two facts about the distribution of prime numbers. The first
is that, [they are] the most arbitrary and ornery objects studied by math-
ematicians: they grow like weeds among the natural numbers, seeming to
obey no other law than that of chance, and nobody can predict where the
next one will sprout. The second fact is even more astonishing, for it states
just the opposite: that the prime numbers exhibit stunning regularity, that
there are laws governing their behavior, and that they obey these laws with
almost military precision.”
Don Zagier

Prime numbers have fascinated mankind through the ages. In fact, one
may think that we know all about them. However, this is not so! One
does not know the answers to many basic questions on primes. We shall
concentrate here mainly on questions and discoveries whose statements are
elementary and accessible. Right at the end, we mention a result whose
statement is simple but whose proof uses rather sophisticated mathematics.
Even here, we do not try to be exhaustive. The subject is too vast for that
to be possible.

1. Introduction
Let us start with the first major discovery about primes, which is the
proof by Euclid’s school that there are infinitely many prime numbers.
Euclid’s proof of the infinitude of primes will eternally remain beautiful no
matter what advances modern mathematics makes. In spite of its simplicity,
it still retains quite a bit of mystery. For instance, it is unknown as yet
whether the product of the first few primes added to 1 takes a prime value
infinitely often. It is even unknown whether it takes a composite value
infinitely often! Do you see the mystery? What is the first time we get
some composite number? Does anyone know the answer already? Anyway,
The chapter is a modified version of an article that first appeared in Resonance, Vol. 13, No. 9,
pp. 866–881, September 2008.

75
 The Prime Ordeal

let me tell you that 2.3.5.7.11.13 + 1 is not a prime.

Actually, it is often the case that for any sequence of natural numbers
which does not obviously take only composite values, the question as to
whether it does take infinitely many prime values remains unanswered.
Here are some examples (the p1 , p2 , . . . are prime numbers):
(i) p1 p2 · · · pn + 1,
(ii) p1 p2 · · · pn − 1,
(iii) n! + 1,
(iv) n! − 1,
(v) 2n − 1,
(vi) 2n + 1,
(vii) n2 + 1,
(viii) f (n) for any polynomial of degree ≥ 2 such that there is no k dividing
all the values f (r), r ∈ Z.
Of course, in (v), it is obviously necessary that n itself be prime, and in
(vi), a necessary condition is that n is a power of 2. As for (vii), it was proved
by a contemporary mathematician Henryk Iwaniec in 1978 using some ad-
vanced mathematics that infinitely many numbers of the form n2 +1 can be
expressed as a product of at most 2 primes. Note that the condition in (viii)
cannot be weakened; for instance, if we merely say that all the coefficients
of f be not divisible by any k, it is not sufficient. Indeed, f (x) = x(x + 1) is
a counter example. That the sequence in the last example takes infinitely
many prime values was conjectured by Viktor Bouniakowsky in the 19th
century. In contrast to the last example, the degree one case is known to
take infinitely many prime values – this is the famous theorem of Lejeune
Dirichlet on primes in arithmetic progressions. Incidentally, here is a little
exercise: If we make the (apparently weaker) conjecture that under the hy-
pothesis of example (viii), every such f takes ONE prime value, it is actually
equivalent to asserting that each such f takes infinitely many prime values!
Here is another issue of importance – in cryptography, for example. Given
a natural number n, how does one recognize whether it is prime or not?
This is of crucial importance in many modern cryptosystems where the
belief is that it is comparatively much easier (computationally) to answer
this question than to factorize a given number. Basically, the idea would
be to unearth properties of prime numbers which characterize them (that
is, would not hold for even a single composite number). One such funda-
mental property (which is an easy exercise) is that a natural number n > 1
is prime if, and only if, n divides (n − 1)! + 1. This is known as Wilson’s
congruence. Another such property is that any prime p divides the binomial
coefficients pr for each r in the range 0 < r < p. That this is untrue for

every composite number is again a nice little exercise.

76
 The Prime Ordeal

Using the above property of primes, one can prove by induction on n

that np − n is a multiple of p for every n. Equivalently, if p does not divide
n, then p divides np−1 − 1. This is known as the little theorem of Fermat.
Interestingly, I found that one website in German translated ‘the little
theorem of Fermat’ as ‘the small sentence of Fermat’ ! It is even funnier
when we recall that Fermat was a judge who did pass sentences at times!
At this point, it is better to stop and point out the answer to a question
which would have crossed the minds of many people. Is there a ‘formula’
for the n-th prime? Indeed, there are many formulae for primes! However,
they are all worthless in a practical sense; that is, one cannot hope to
computationally produce primes by such formulae. However, later we do
talk about a recent algorithm by 3 Indians which tells us in polynomial
time whether a given number is prime or not. Here is a ‘formula’ for primes
x−y
based on Wilson’s congruence. Put f (x, y) = 12 {1 + |x−y| } if x 6= y, and
f (x, x) = 0. Note that f (x, y) P is simply 1 or 0 according as to whether
x > y or x ≤ y. Put π(n) = 1 + ni=3 {(i − 2)! − i[(i − 2)!/i]} for n ≥ 3 and
π(1) = 0, π(2) = 1. This counts the number of primes up to n. Then, the
n-th prime pn is given by the formula:
X2n
pn = 1 + f (n, π(i)).
i=1
After some thought, we can see that the formula, although perfectly valid,
is of no practical use in finding the n-th prime. A somewhat better formula
was given several years ago by an Indian named J M Gandhi.

2. Carmichael Numbers: ‘Carm’posites in Prime Clothing

Some avatar of Fermat’s little theorem is used in most primality tests
even today. But, unfortunately Fermat’s little theorem does not character-
ize primes! It does happen for some composite n that n divides an−1 − 1 for
some a co-prime to n. In the terminology of cryptography, one says that n
is a pseudo-prime to the base a and that a is a strong liar for n. Worse hap-
pens – there are, indeed, infinitely many numbers (known as Carmichael
numbers after Robert Carmichael) n such that n divides an−1 −1 for every a
co-prime to n. The smallest such number is 561. The proof of the infinitude
of the Carmichael numbers (as recently as 1994) also showed that there are
at least n2/7 such numbers ≤ n provided n is sufficiently large. The proof
used deep, modern-day mathematics. In this article, I will concentrate on
two conjectures (one made in 1950 and the other made in 1990) which aim
to characterize primes. Ironically, they have turned out to be equivalent! As
the conjectures involve Carmichael numbers also, we first prove a nice ele-
mentary criterion due to Theodor Korselt which characterizes Carmichael
numbers.

77
 The Prime Ordeal

In what follows, we will be using the following notations. We will say

a ≡ b mod m when a − b is a multiple of m. These congruences have a
calculus quite similar to equality. Namely, if a ≡ b mod m and c ≡ d mod
m (same m, of course), then a + b ≡ c + d and ab ≡ cd mod m.
Theorem. A composite number n is a Carmichael number if, and only
if, n is square-free and, for each prime divisor p of n, the number p − 1
divides n − 1.
Proof. We shall assume and use the following fact which was first proved
by Gauss. For any prime number p, there exist positive integers a < p and
b < p2 which have ‘orders’ p − 1 and p(p − 1) in the following sense:
when ar ≡ 1 mod p, then p − 1 divides r, and
when bs ≡ 1 mod p2 , then p(p − 1) divides s.
(It should be noted that neither of these statements is trivial to prove
although they are some 200 hundred years old.)
Now, first let n = p1 p2 · · · pr be a square-free number such that for each
i ≤ r, the number pi − 1 divides n − 1. Evidently, for every a co-prime to
n, a is co-prime to each pi . Thus, one has by Fermat’s little theorem that
api −1 ≡ 1 mod pi . So, an−1 = (api −1 )∗ ≡ 1 mod pi . In other words, pi
divides an−1 − 1 for each i ≤ r. Thus, n = p1 p2 · · · pr itself divides an−1 − 1.
This shows that n is a Carmichael number.
Conversely, let n be a Carmichael number. If p is a prime dividing n,
consider a natural number a of ‘order’ p − 1 mod p. We claim that we can
always choose such an a which is co-prime to n.
First, if a is co-prime to n, then by hypothesis, an−1 ≡ 1 mod n, which
implies an−1 ≡ 1 mod p, and thus p − 1 divides n − 1. If (a, n) > 1, then
look at the set of primes p = p1 , · · · , pk which divide n but not a. Consider
a + p1 · · · pk in place of a. Evidently, a + p1 · · · pk is co-prime to n. Moreover,
it is of the form a + pd, and so, its ‘order’ mod p is the same as that of a.
Now, let p2 divide n for some prime p, if possible. Let b be of order p(p−1)
mod p2 . If b is co-prime to n, then bn−1 ≡ 1 mod n which gives bn−1 ≡ 1
mod p2 which again implies that p(p − 1) divides n − 1. Thus p divides
(n − 1), an impossibility because p divides n. So, n must be square-free if
the b can be chosen co-prime to n. But, if (b, n) > 1, then once again we
look at the set of primes p = p1 , p2 , · · · , pk which divide n but not b. Then
b + p21 p2 · · · pk is co-prime to n and has the same order mod p2 as b has,
namely, p(p − 1).
The proof is complete.
We end with an easy exercise:
Suppose n = p1 · · · pr is a Carmichael number and m ≡ 1 mod L where
L = LCM of p1 − 1, · · · , pr − 1. If qi = 1 + m(pi − 1) are all primes, then
N = q1 · · · qr is also a Carmichael number.

78
 The Prime Ordeal

3. ‘Nava’ Giuga and long ‘Agoh’

Let us start with the first of the 2 conjectures we wish to discuss. If p is
a prime, then clearly

1p−1 + 2p−1 + · · · + (p − 1)p−1 ≡ −1 mod p.

Giuseppe Giuga conjectured in 1950 that this characterises primes; that is,
Conjecture (Giuga 1950): n−1 n−1 ≡ −1 mod n ⇒ n is prime.
P
k=1 k
As he showed, the conjecture can be reformulated as follows:
Theorem. n−1 n−1 ≡ −1 mod n if, and only if, for each prime divisor
P
k=1 k
p of n, both p and p − 1 divide np − 1.
Pn−1 n−1
Equivalently, a composite number n satisfies k=1 Pk ≡Q −1 mod n
if, and only if, it is a Carmichael number such that p|n p1 − p|n p1 is a
natural number.
In the above statement, the sum and the product run over primes and
p|n denotes ‘p divides n’.
Pp−1 r
Proof. Note that for any prime p, we have k=1 k ≡ −1 or 0 mod p
according as whether p − 1 divides r or not.
Therefore, for a prime p dividing n, we have
n−1
X p−1
X 2p−1
X n−1
X
n−1 n−1 n−1
k ≡ k + k + ··· + k n−1
k=1 k=1 k=p+1 k=n−p+1

≡ −n/p or 0 mod p according as to whether − 1 divides n − 1 or not.

Pn−1 p n−1
To prove the theorem, first suppose k=1 k ≡ −1 mod n. Then, for
every prime p|n, we have (p − 1)|(n − 1) and np ≡ 1 mod p. Note that
(p − 1)|(n − 1) implies p − 1 divides p( np − 1) = n − p = (n − 1) − (p − 1)
and so (p − 1) also divides np − 1.
Conversely, suppose p(p − 1) divides np − 1 for each prime divisor p of n.
First P
of all, this forces n to be square-free. Now, for any prime p|n, we also
have n−1 k=1 k
n−1 ≡ − n ≡ −1 mod p. This proves the first statement. The
p
second assertion is easy. If p(p − 1)|( np − 1) for each prime p|n, we have that
n is a Carmichael number (in particular, it is square-free). Then,
X1 Y1 X1 1
− = − .
p p p n
p|n p|n p|n
P n
So, multiplying by n, we must show that n divides P p|n p − 1. Thus, we
n
need to show that each prime divisor of n divides p|n p − 1. This follows
because each prime divisor p of n satisfies p|( np − 1) and p| nq for p 6= q.

79
 The Prime Ordeal

Remarks. A composite number n such that p|( np − 1) for each prime p|n,
is called a Giuga number. Equivalently, p|n p1 − p|n p1 ∈ N. Then, Giuga’s
P Q
conjecture amounts to the assertion that there is no Giuga number which is
also a Carmichael number. As of today, only 12 Giuga numbers are known
and all of them have sum minus product (of reciprocals of prime divisors)
equal to 1. The numbers 30, 858, 1722 are Giuga numbers. Until now, no
odd Giuga numbers have been found. Any possible odd Giuga number must
have at least 10 prime factors because the sum 13 + 15 + 71 + 11
1 1
+ 13 1
+ 17 +
1 1 1
23 + 29 + 31 < 1.
In an article in Volume 103 of the American Mathematical Monthly of
1996, David Borwein, Jonathan Borwein, Peter Borwein and Roland Gir-
gensohn propose that a good way to approach Giuga’s conjecture is to study
Giuga numbers in general. More generally, they define a Giuga sequence to
be a finite Qsequence n1 < n2 < · · · < nr of natural numbers such that
P r 1 r 1
i=1 ni − i=1 ni is a natural number. Thus, a Giuga sequence consist-
ing of primes gives rise to a Giuga number, viz., to the product of those
primes. The smallest Giuga sequence where the sum minus product is > 1,
has 59 factors! Here is an easy method to produce arbitrarily long Giuga
sequences.
Theorem. Suppose n1 < n2 < · · · < nr is a Giuga sequence satisfying
nr = r−1
Q
i=1 n i − 1. Then, the sequence n1 < n2 < · · · < ñr , ñr+1 is a Giuga
Qr−1
sequence whoseQr−1 sum minus product is the same, where ñ r = i=1 ni +
1, ñr+1 = ñr i=1 ni − 1.
Starting with a sequence like 2, 3, 5 say, this gives Giuga sequences of
arbitrary lengths whose sum minus product is 1. The proof is a simple
exercise of manipulation. In fact, one has the following nice result:

r 1 Qr 1Look at a sequence n1 < n2 < · · · < nr which

PProposition. Q satisfies
i=1 ni + i=1 ni = 1 - for example, the sequence n1 = 2, nk = i<k ni + 1
Qk
is such a sequence. Then, n1 < n2 < · · · < nk < nk+1 := i=1 ni − 1 is a
Giuga sequence.
The proof is straightforward verification.
Incidentally, note that the sequence given as an example above proves the
infinitude of primes because the pairwise GCD (ni , nj ) = 1 for all i 6= j.
The Giuga conjecture involved the sums n−1 n−1
P
k=1 kP . As we have seen in
an earlier chapter, in general, a sum of the form r−1 n
k=1 k can be ‘easily’
evaluated in terms of certain rational numbers called the Bernoulli numbers.
Let us recall these briefly. These ubiquitous numbers have so many connec-
tions that it is impossible to mention most of them here. Suffice it to say
that Fermat’s last theorem can be proved for a prime p (in an easy, natural

80
 The Prime Ordeal

manner) provided p does not divide the numerators of B2 , B4 , · · · , Bp−3 .

How are theP Bn ’s defined? Often, they are defined by means of the gener-
zn
ating series ∞ z
n=0 Bn
n! = ez −1 . The equality can be un-winded to give the
Pn n+1
recursion r=0 r Br = 0 and using B0 = 1, one can determine them.
It turns out that B1 = − 12 and Br = 0 for all odd Pr > 1.
More generally,
the Bernoulli polynomials are defined as Bn (x) = nk=0 nk Bk xn−k ; it is of
degree n. Note that Bn (0) = Bn .
We showed that
r−1
X 1
kn = (Bn+1 (r) − Bn+1 ).
n+1
k=1

In this manner, the sums of powers can be expressed in terms of Bernoulli

numbers.
The von Q Staudt–Clausen theorem says that the denominator of B2k is
precisely (p−1)|2k p; note this is square-free. In particular, it makes sense
to talk about (2k + 1)B2k mod 2k + 1; note that for (a, b) = 1, one talks of
1
a mod b – it is the unique c mod b for which ac ≡ 1 mod b.
For example, 15B14 = 15 × 67 = 35 2 ≡ 35 × 8 ≡ −5 mod 15.
−691
13B12 = 13 × 2.3.5.7.13 ≡ −1 mod 13.
Looking at such data, Takashi Agoh conjectured in 1990 (conjectured by
‘Agoh’ not long ‘ago’ !):
nBn−1 ≡ −1 mod n if, and only if, n is prime.
A few years later (in 1994) he used the von Staudt–Clausen theorem
and showed that his conjecture is actually equivalent to Giuga’s conjecture.
Then, in September 2004, Bernd Kellner gave a new proof of the equivalence
of the two conjectures (which gives another proof of von Staudt-Clausen
theorem) based on the following result:
Theorem (Kellner). If m > 1, and n is even, then
m−1
X X m
kn ≡ − ≡ mBn mod m.
p
k=1 p|m,(p−1)|n

The proof is elementary but rather involved and we do not discuss it

here. This theorem allows for a further reformulation of the Giuga and
Agoh conjectures, and may now be called:
Conjecture (Agoh-Giuga-Kellner): An integer n ≥ 2 is prime if, and only if,
X 1 1
− ∈ Z.
p n
p|n,(p−1)|(n−1)

81
 The Prime Ordeal

4. All’s Bell
In this section, we discuss a conjecture due to Djuro Kurepa which can
be stated in elementary language but the proof which appeared last year
involves some sophisticated mathematics. Those who have learnt Galois
theory would be able to appreciate it but others can also get a flow of the
argument. Of course, the fact that an elementary statement may require
very sophisticated methods should not come as a surprise. A case in point
is Fermat’s last theorem (FLT) which says that for an odd prime p, there
do not exist nonzero integers x, y, z such that xp +y p +z p = 0. The question
of Kurepa doesn’t quite require the kind of sophisticated mathematics re-
quired in FLT though.
Pp−1Kurepa conjectured in 1971 that for any odd prime
p, the sum Kp := n=0 n! is not a multiple of p. Of course K2 = 2. This
is, of course, not a characterisation of primes; for example, K4 = 10. The
proof (only in 2004) of Kurepa’s conjecture due to D Barsky and B Ben-
zaghou involves the so-called Bell numbers named after Eric Temple Bell.
One way of defining the Bell numbers is as follows. The n-th Bell number
Pn is the number of ways of writing an n-element set as s union of non-
empty subsets. We see that P1 = 1, P2 = 2, P3 = 5, P4 = 15, P5 = 52 etc.
There is a lot of combinatorics involving the Bell numbers.
Pn
From combi-
n
natorial considerations, one can prove that Pn+1 = k=0 k Pk , where we
have written P0 to stand for 1. From this, it is easy to prove (analogously
to the proof for Bernoulli numbers) that the generating function for Pn ’s is
given by
∞ ∞
X X xn
F (x) = Pn xn = · · · · · · (♠).
(1 − x)(1 − 2x) · · · (1 − nx)
n=0 n=0

The Kurepa question can be formulated in terms of the Bell numbers P eas-
ily. It turns out using some elementary combinatorics that Pp−1 ≡ p−2 n=0 n!
modulo p. Thus, since Kp is the sum of (p − 1)! with the right hand side
above, Kurepa’s conjecture amounts to the statement that Pp−1 6≡ 1 mod-
ulo p because (p − 1)! ≡ −1 modulo p. The idea of the proof Kurepa’s
conjecture is to consider what is known as the Artin–Schreier extension
Fp [θ] of the field Fp of p elements, where θ is a root (in the algebraic clo-
sure of Fp ) of the polynomial xp − x − 1. This is a cyclic Galois extension
of degree p over Fp . Note that the other roots of xp − x − 1 are θ + i for
i = 1, 2, · · · , p − 1. The theory of such extensions is named after Emil Artin
and Otto Schreier. The reason this field extension comes up naturally is
as follows. The generating series F (x) of the Bell numbers can be evalu-
ated modulo p; this means one computes a ‘simpler’ series Fp (x) such that
F (x) − Fp (x) has all coefficients multiples of p. Since Kurepa’s conjecture
is about the Bell numbers Pp−1 considered modulo p, it makes sense to

82
 The Prime Ordeal

consider Fp (x) rather than F (x). Reading the equality (♠) modulo p, one
gets
p−1 X
X xip+n
Fp (x) =
(1 − x) · · · (1 − (ip + n)x)
n=0 i≥0
p−1 X
X xn xip
=
1 − (ip + 1)x) · · · (1 − (ip + n)x) (1 − x) · · · (1 − ipx)
n=0 i≥0
p−1 X i
xn xp
X 
≡
1 − (ip + 1)x) · · · (1 − (ip + n)x) (1 − x) · · · (1 − px)
n=0 i≥0

modulo p. Therefore,
Pp−1 n (1
n=0 x − (n + 1)x) · · · (1 − (p − 1)x)
Fp (x) =
1 − xp−1 − xp

on simplification. Notice that θ−1 is a root of the polynomial 1 − xp−1 − xp

above. Thereafter, doing some algebra in the field extension Fp [θ] of Fp
expresses the various Bell numbers Pn modulo p as

Pn ≡ −T r(θcp )T r(θn−cp −1 ),

where T r denotes the trace to Fp from the Artin–Schreier extension Fp [θ]

pp −t p −1
and cp = p−1p and tp = pp−1 . Thereafter, the analysis of the properties of
the trace functions implies that if Pp−1 − 1 were to be zero modulo p, then
θcp would be zero, which is absurd since θ is not zero, as it generates a degree
p extension. This was one instance of proving an elementary statement on
primes which needs some sophisticated mathematics.

5. AKS --- A Case of Indian Expertise

Having said that there are no (practically) ‘nice’ formulae for primes,
and having also said that producing large primes is a basic requirement
in fields like cryptography, how does one reconcile one with the other?
The fact is that there are many probabilistic algorithms to certify primes
with very high probability. We shall not discuss them but we raise the
mathematical question as to whether there are deterministic algorithms to
decide in reasonable computational time whether a given number is prime
or not. Until very recently, no deterministic algorithm was known which
was polynomial-time and which could detect every prime. Recently, three
Indians (Manindra Agrawal, a professor of computer science from IIT Kan-
pur and his students Neraj Kayal and Nitin Saxena) stunned the world

83
 The Prime Ordeal

with the discovery of a polynomial-time deterministic primality testing al-

gorithm. We mention very briefly the Agrawal–Kayal–Saxena algorithm.
Most algorithms start with Fermat’s little theorem which, apart from other
shortcomings, are also infeasible on the first glance because of having to
compute p coefficients in order to check the validity of the congruence
(x − a)p ≡ xp − a mod p. The basic idea of the A-K-S algorithm is to make
it feasible by evaluating both sides modulo a polynomial of the form xr − 1.
Their algorithm would take O(r2 log3 p) time to verify (x−a)p ≡ xp −a mod
xr − 1 in Fp [x]. As there are composites also which satisfy this congruence,
one has to choose r and a suitably. One general comment to note is that
it is far easier to test a polynomial over Fp for irreducibility than to test
primality of a natural number. In a nutshell, here is the A-K-S algorithm:
A-K-S algorithm to check primality of n
Step I
Check if n is a perfect power; if not go to the next step.
Step II
Find a prime number r = O(log6 n) such that r − 1 has a prime divisor
√
q > 4 r logn where q divides the order of n mod r.
Step III
√
With r as above, check for each a ≤ 2 r logn, if

(x − a)n ≡ xn − a mod xr − 1 in (Z/nZ)[x].

If the congruence is not satisfied for some a, declare n is composite. If it

is satisfied for all a, declare n prime.

6. Sundries
We will finish with a few more remarks about primes. We mentioned
Bouniakowsky’s conjecture which asserts the infinitude of prime values.
Can a polynomial take only prime values? It is again an easy, elementary
exercise to prove that there is no nonconstant polynomial in some vari-
ables x1 , · · · , xr which takes only prime values at all integers. However, it
is a deep consequence of the solution of Hilbert’s 10th problem by Hilary
Putnam, Martin Davis, Julia Robinson and Yuri Matiyashevich that there
exist polynomials f (x1 , · · · , xr ) over integers such that the set of positive
values taken by f equals the set of prime numbers! Of course, the polyno-
mials does take negative values as well as take certain prime values more
than once. Indeed, one can take f to be of degree 25 and r to be 26. This
expresses the fact that the set of prime numbers is a Diophantine set.

84
 The Prime Ordeal

Joseph Bertrand stated that there is a prime among n + 1, n + 2, · · · , 2n.

This is Bertrand’s postulate which was discussed in an earlier chapter; it
was proved first by Pafnuty Chebychev and there are many simpler proofs.
Incidentally, a generalization of Bertrand’s postulate is a theorem of James
Joseph Sylvester which asserts that in any sequence n + 1, n + 2, · · · , n + r
with n ≥ r, there is a number which is divisible by a prime > r.
Of course, the twin prime problem (whether there are infinitely many
primes p with p + 2 also prime) is still open. Viggo Brun proved that
the series of reciprocals of twin primes converges. Note that the series
P of
reciprocals of all primes is divergent, as proved by Euler. Indeed, p≤x p1
behaves asymptotically like the function log log x for x tending to infinity.
Then, the Goldbach conjecture (named after Christian Goldbach and as-
serting that every even number >2 is a sum of two primes) is also open; Ivan
Matveevich Vinogradov proved using the Hardy–Ramanujan circle method
(named after G H Hardy and S Ramanujan) that every sufficiently large
odd number is a sum of three primes. The prime number theorem proved in
the beginning of the 20th century shows that the ‘prime counting function’
π(x) which counts the number of primes up to x, behaves asymptotically
like the function logx x as x tends to infinity. An equivalent formulation is
to say that the product of all the primes up to some x is asymptotically
like ex . Here, and elsewhere, one means by the statement f (x) is asymptot-
ically like g(x) that the ratio f (x)/g(x) approaches 1 as x tends to infinity.
One can deduce from the prime number theorem that the n-th prime is
approximately of size n log n for large n. That is, very roughly speaking,
the probability that a given n is prime is log1 n .
In connection with the fact we mentioned about Gauss showing that for
each prime p, there is an integer a whose order mod p is p − 1, here is
a famous conjecture due to Emil Artin. He conjectured that each natural
number a which is not a square is the order mod p for infinitely many
primes p. It is also open.
Shortly before his death, Paul Erdös, in collaboration with Takashi Agoh
and Andrew Granville, showed that any large composite n (n ≥ 400 would
do) satisfies   
X 1 Y
n≤  p .
√ p √
p≤ n p≤ n

Using this, and nothing more than the Chinese remainder theorem, they
showed that any prime n can be proved to be prime by expressing it as
n = N1 + N2 + · · · + Nk where p1 , · · · , pk are the first k primes and n is not
divisible by any of them while each Ni is divisible by all the pj with j 6= i
and not by pi .

85
Extending Given Digits to Make Primes
or Perfect Powers

Any sequence of digits can be amended

by adding a tail and extended
to get a power
more or less of whatever
and maybe even prime opposite ended!

1. Introduction
Start with any string of digits. Can we always put down some more digits
on the right of it to get a prime? Can we similarly get a power of 2? How
about a power of 3? It turns out that the answers to all these questions are
in the affirmative. Our discussion will be elementary excepting a concrete
consequence of a weak version of the prime number theorem. For a really
detailed analysis of the proportion of primes with given starting digits, the
interested reader is referred to look at ergodic theory. There is no prime-
producing polynomial in a single variable – this is trivial to see. However,
it turns out (as other concrete consequences of the properties of primes
like Bertrand’s postulate and the prime number theorem) that there are
exponential type of functions which produce infinitely many primes. One
n
such is the sequence of integer parts of t3 for a certain positive real number
s
..
t. Another is the sequence of integer parts of 22 for a certain real s > 0.
We shall prove these also.

2. Perfect Powers with a Given Beginning

Let us first deal with the problem of extending a given string of digits to
make a power of any natural number a which is at least 2 but not a power
of 10. Notice that these exceptions are clearly unavoidable; there is no way
to start with, say 11, and get a power of 10 by adding any number of digits.
Let a > 1 be any natural number other than a power of 10. Let A be any
given natural number in base 10. The only property we need is the following
observation which can be proved simply by using the pigeon-hole principle.
For any real number α, let us write {α} for the fractional part of α.
Observation. For any irrational number θ > 0, the sequence of fractional
parts {nθ} as n varies over natural numbers, is dense in the interval (0, 1).
The chapter is a modified version of an article that first appeared in Resonance, Vol. 15, No. 10,
pp. 941-947, October 2010.

87
 Extending Given Digits to Make Primes or Perfect Powers

Proof. For any n, consider the intervals [0, 1/n), [1/n, 2/n), · · · , [(n −
1)/n, 1). By the pigeon-hole principle, among the fractional parts of
θ, 2θ, · · · , nθ, (n + 1)θ, there must be at least two (say rθ, sθ with r <
s ≤ n + 1) which are in the same interval [(k − 1)/n, k/n). But then the
fractional part of (s − r)θ lies in [0, 1/n). Therefore, each [(m − 1)/n, m/n]
contains the fractional part of some dθ where d is a multiple of s−r. As every
subinterval (x, y) of (0, 1) contains an interval of the form [(m − 1)/n, m/n]
for some m, n, the claim asserted follows.
Using this, one may extend any given digits to produce powers as follows.
Lemma 1. Let a > 1 be not a power of 10 and let A be any given natural
number. Then, one may add digits to the right end of the digits of A to
obtain some power of a.
Proof. For any a as above, log10 (a) is irrational because, if it is u/v, then
10u = av which implies by uniqueness of prime decomposition that a must
be a power of 10, a contradiction of the hypothesis. So, it follows by the
above observation that each interval (x, y) ⊆ (0, 1) contains some fractional
part {n log10 (a)}. Now, suppose A has d+1 digits; that is, 10d ≤ A < 10d+1 .
Then, we consider x ∈ (0, 1) such that 10x = A/10d . Choosing some large
n so that 101/n < 1 + A1 (as limn→∞ 101/n = 1), we consider the point y ∈
1/n
(x, 1) with 10y = 1010d A . Note that y < 1 because 101/n A < A + 1 ≤ 10d+1 .
If the fractional part {r log10 (a)} ∈ (x, y), we have
x < r log10 (a) − k < y
for some positive integer k. Taking 10-th powers, we have
ar
10x < < 10y
10k
which gives
10k−d A < ar < 10k−d 101/n A < 10k−d A + 10k−d .
Thus, ar has been obtained by adding k − d digits to the right of the base
10 expansion of A.
Illustration. Let us see how to demonstrate the above lemma for a small
number. Let us begin with A = 4 and a = 2. Of course A itself is a
power of 2 but let us see what we get from the above lemma. In the above
notation, d = 0 and x = log10 (4). The choice n = 16 is large enough so that
101/n < 1 + 1/4 = 1.25. Then y = x + 1/16 and the choice of r, k such that
2r 1
10x = 4 < k
< 10y = 4 +
10 16
can be taken to be r = 12, k = 3. Hence 4 can be extended to 212 = 4096.

88
 Extending Given Digits to Make Primes or Perfect Powers

3. Primes with a Given Beginning

Now, we consider the problem of extending given digits on the right to
get a prime. Here, we will need the following property of prime numbers
which is a weak consequence of the so-called prime number theorem:
There exists n0 so that for n ≥ n0 , there is always a prime strictly between
n and n + logn n .
Lemma 2. Let A be any given natural number. Then, one may add digits
to the right end of the digits of A to obtain a prime number.
Proof. As our purpose is to add digits to the right end, we may assume
that A ≥ n0 where n0 is as above. Now, let r > log A(10) and consider a
e
10r A
prime p between 10r A and 10r A + r
loge (10r A) . Thus, p = 10 A + d where
r
d < log10(10Ar A) < 10r since A < loge (10r ) = r loge (10) by the choice of r.
e
Finally, note that the last lemma implies the following:
COROLLARY 1.
The fractional parts of log10 (p), as p runs over primes, is dense in (0, 1).
Proof. Let n be arbitrary and divide (0, 1) into the intervals (0, 1/n),
[1/n, 2/n), · · · , [(n − 1)/n, 1). Consider the numbers 10(m−1)/n , 10m/n ∈
[1, 10). By lemma 2, for each m < n, there is a prime p and some inte-
ger d so that
p
10(m−1)/n < d < 10m/n .
10
Thus, (m − 1)/n < log10 (p/10k ) < m/n which means the fractional part
of log10 (p) lies in ((m−1)/n, m/n). This completes the proof of the corollary
as m, n are arbitrary.
We remark that using a weak version of the prime number theorem for
arithmetic progressions, one may similarly prove that given any string of
beginning digits and any string of end digits which end in 1, 3, 7 or 9, one
may introduce digits in between to get a prime.

4. Some Exponential Functions Producing Primes

The so-called Bertrand postulate (which was discussed in an earlier chap-
ter) tells us that there is a prime between N and 2N for each N > 1. Using
this, one can write down a function which produces infinitely many primes.
Let us first discuss this 1951 result due to E M Wright [1]. This asserts:
Lemma 3. There exists a real number s > 0 such that the sequence
s
a0 = s, a1 = 2s , a2 = 22 , · · · , an+1 = 2an produces primes
s
2 ..
[an ] = [22 ] for all n > 0.

89
 Extending Given Digits to Make Primes or Perfect Powers

Proof. Let p1 = 2, p2 = 3 and choose primes pn for n > 2 such that

2pn < pn+1 < pn+1 + 1 < 2pn +1 .

Look at the sequences bn and cn defined as follows. Define bn = log2 log2 · · ·

log2 (pn ) and cn = log2 log2 · · · log2 (pn + 1) where there are n logarithms to
the base 2. Then, we have

pn < log2 pn+1 < log2 (pn+1 + 1) < pn + 1.

This means bn < bn+1 < cn+1 < cn which ensures that the sequence {bn }
converges to some real number s as n → ∞. Notice that for this number s,
s
2 ..
the sequence an = 22 satisfies pn < an < pn + 1. Hence pn = [an ]. This
completes the proof.
Remarks.
(i) The above formula is not a practical one. Since there is a choice of pn ’s
allowed, the real number s is not unique. One possible value of s is
1.9287800 · · · and the primes pn defined by the lemma grow much too
fast. For example, p4 has 5000 digits.
(ii) A result earlier to Wright’s result above (in fact, the result which mo-
tivated Wright’s theorem) is due to W H Mills [2] in 1947. This uses a
result on primes which is considerably deeper than Bertrand’s postu-
late. This deeper result alluded to is due to the British mathematician
A E Ingham; he derived in 1937 (see [3]) the following concrete conse-
quence of the prime number theorem:
5/8
There is a positive number c such that pn + cpn > pn+1 for all n, where
p1 < p2 < p3 < · · · is the sequence of all primes.
n
Lemma 4. There exists a real number t > 0 such that [t3 ] is a prime for
every n.
Proof. Start with Ingham’s result and choose a large N > c8 . Look at
the prime pn such that pn < N 3 < pn+1 . Then, we have

pn < N 3 < pn+1 < pn + cp5/8 3

n < N + cN
15/8
< N 3 + N 2 < (N + 1)3 − 1.

Take for N , a prime p > c8 . Thus, we have a sequence of primes pr0 =

p < pr1 < pr2 < · · · such that

p3rn < prn+1 < (prn + 1)3 − 1 · · · (♥)

−n −n
Then the sequences un = p3rn and vn = (prn + 1)3 satisfy
−n −n−1 −n−1 −n
vn = (prn + 1)3 > (prn+1 + 1)3 = vn+1 > p3rn+1 = un+1 > p3rn = un .

90
 Extending Given Digits to Make Primes or Perfect Powers

Indeed, the inequality vn > vn+1 is simply the second inequality in ♥; the
inequality un+1 > un is the first inequality of ♥ and the inequality vn+1 >
un+1 is obvious. Hence, the sequence {un } is a bounded, monotonically
increasing sequence and must have a limit t. Clearly, t = limn→∞ un satisfies
un ≤ t < vn . Thus,
n
prn ≤ t3 < prn + 1.
n
This proves that [t3 ] = prn for all n.

Remarks on Mills’s Constant

Mills proved only the existence of a constant t as above. Later, others
showed that there are uncountably many choices for t but it is still not
possible to produce a value of t which can be proven. Under the Riemann
hypothesis, one can prove that there is a value of t which is between 1.3
n
and 1.31 for which the sequence [t3 ] gives primes.

References
[1] E M Wright, A prime-representing function, Amer. Math. Monthly,
Vol. 58, No. 9, pp. 616–618, 1951.
[2] W H Mills, A prime-representing function, Bull. Amer. Math. Soc., 53,
pp. 604, 1947.
[3] A E Ingham, On the difference between consecutive primes, Quart.
J. Math. Oxford Ser., Vol. 8, pp. 255–266, 1937.

91
An Irrational Walk and Why 1
is Not Congruent

1. Introduction
The subject of Diophantine equations is an area of mathematics where
solutions to very similar-looking problems can vary from the elementary to
the deep. Problems are often easy to state, but it is usually far from clear
whether a given one is trivial to solve or whether it must involve deep ideas.
Fermat showed that the equations X 4 + Y 4 = Z 4 and X 4 − Y 4 = Z 4 do
not have nontrivial solutions in integers. He did this through a method now
known as the method of descent. In fact, he discovered this while working
on a Diophantine problem called the congruent number which we discuss
below. There are other situations where these equations arise naturally.
One such problem we discuss is the following.

#
#
#
#
#







Suppose we start walking from a corner of a unit square to reach the

diagonally opposite corner. The rule is to walk on a straight line to some
point of the middle vertical line as in the figure and, on reaching that point,
walk towards the opposite corner along a straight line. Thus, we have a path
as in the figure consisting of one segment until the middle line is reached and
the other from that point to the opposite corner of the square. The question
is whether we can follow such a path where both the segments we walked
can be rational in lengths. This was asked (and answered!) by Roy Barbara
in Article 93.21, Vol. 93 (2009), The Mathematical Gazette. We discuss this
problem also.

The chapter is a modified version of an article that first appeared in Resonance, Vol. 17, No. 1,
pp. 76–82, January 2012.

93
 An Irrational Walk and Why 1 is Not Congruent

2. The Congruent Number Problem

A natural number d is said to be a congruent number if there is a right-
angled triangle with rational sides and area d.
(Equivalently:) Can we have an arithmetic progression of three terms
which are all squares of rational numbers and the common difference d?
That is, x2 − d, x2 , x2 + d comprised of squares of rational numbers where
x is rational?
Indeed, Let u ≤ v < w be the sides of a right-triangle with rational sides.
Then x = w/2 is such that (v − u)2 /4, w2 /4, (u + v)2 /4 form an arithmetic
progression.
Conversely, if x2 − d = y 2 , x2 , x2 + d = z 2 are three rational squares
in arithmetic progression, then z − y, z + y are the legs of a right angled
triangle with rational legs, area (z 2 − y 2 )/2 = d and rational hypotenuse
2x because 2(y 2 + z 2 ) = 4x2 .
• For example, 5, 6, 7 are congruent numbers.
To see these, consider the following three right-angled triangles:
with sides 3/2, 20/3, 41/6 with area 5,
with sides 3, 4, 5 with area 6,
with sides 35/12, 24/5, 337/60.
• 1, 2, 3 are not congruent numbers.
The fact that 1, 2 are not congruent numbers is essentially equivalent to
Fermat’s last theorem for the exponent 4.
Indeed, if a2 + b2 = c2 , 21 ab = 1 for some rational numbers a, b, c then
x = c/2, y = |a2 − b2 |/4 are rational numbers satisfying y 2 = x4 − 1.
Similarly, if a2 + b2 = c2 , 21 ab = 2 for rational numbers a, b, c, then x =
a/2, y = ac/4 are rational numbers satisfying y 2 = x4 + 1.
These equations reduce to the equation x4 ± z 4 = y 2 over integers which
was proved by Fermat using the method of descent not to have nontrivial
solutions.
The unsolvability of y 2 = x4 ±1 in rational numbers are exactly equivalent
to showing 1, 2 are not congruent.
In fact y 2 = x4 − 1 for rational x, y gives a right-angled triangle with
sides y/x, 2x/y, (x4 + 1)/xy and area 1.
Similarly, y 2 = x4 + 1 for rational x, y gives a right-angled triangle with
sides 2x, 2/x, 2y/x and area 2.
Here is an amusing √ way of using the above fact that 1 is not a congruent
number to show that 2 is irrational! √ √
Indeed,
√ consider the right-angled triangle with legs 2, 2 and hypotenuse
2. If 2 were rational, this triangle would exhibit 1 as a congruent number!
Though it is an ancient problem to determine which natural numbers
are congruent, it is only in late 20th century that substantial results were

94
 An Irrational Walk and Why 1 is Not Congruent

obtained and progress has been made which is likely to lead to its complete
solution.
The rephrasing in terms of arithmetic progressions of squares emphasizes
a connection of the problem with rational solutions of the equation y 2 =
x3 − d2 x.
Such equations define elliptic curves.
It turns out that:
d is a congruent number if, and only if, the elliptic curve Ed : y 2 =
x3 − d2 x has a solution with y 6= 0.
In fact, a2 + b2 = c2 , 12 ab = d implies bd/(c − a), 2d2 /(c − a) is a rational
solution of y 2 = x3 − d2 x.
Conversely, a rational solution of y 2 = x3 − d2 x with y 6= 0 gives the
rational, right-angled triangle with sides (x2 − d2 )/y, 2xd/y, (x2 + d2 )/y
and area d.
In a nutshell, here is the reason we got this elliptic curve. The real solu-
tions of the equation a2 + b2 = c2 defines a surface in 3-space and so do the
real solutions of 21 ab = d. The intersection of these two surfaces is a curve
whose equation in suitable co-ordinates is the above curve.

A Rational Walk Which is Impossible

Now, we discuss a problem which, on the face of it, is very different,
but leads to the same impossibility problem as above. Recall the figure we
started with in the introduction.

#
#
#
#
#







Recall that the rule is to walk on a straight line to some point of the
middle vertical line as in the figure and, on reaching that point, walk to-
wards the opposite corner along a straight line. Thus, we have a path as
in the figure consisting of one segment of length r until the middle line is
reached and the other of length s from that point to the opposite corner.
The question is whether we can follow such a path with both the distances
r, s rational numbers. Suppose such a ‘rational’ path is possible. Let us call
the vertical distance x on the middle line from the bottom to the point we

95
 An Irrational Walk and Why 1 is Not Congruent

reach on it. Of course, the rest of the vertical distance is 1 − x. Now,

1
r2 − = x2
4
1
s2 − = (1 − x)2
4
This gives 2x = 1 + r2 − s2 , which is then rational. Then, writing r = p/q
and s = u/v, we have two equations
p
4p2 − q 2
=x
2q
√
4u2 − v 2
=1−x
2v
Therefore, since the above square-roots are rational, they must be in-
tegers and so, q = 2Q (if q were odd, the number 4p2 − q 2 would be −1
modulo 4 which cannot be a square).
Thus, p
p2 − Q2 l
x= =
2Q 2Q
for some l with (l, Q) = 1.
Similarly, v = 2V and
√
u2 − V 2 m
1−x= =
2V 2V
with (m, V ) = 1.
Thus,
l m
1= +
2Q 2V
gives 2QV = lV + mQ.
As (l, Q) = 1 = (m, V ), we get Q|V, V |Q; that is, Q = V as both are
positive integers.
Hence, we have obtained l2 + Q2 = p2 , m2 + Q2 = u2 , l + m = 2Q with
(l, Q) = 1 = (m, Q).
We show that this set of equations does not have any integral solutions.
We could try to use a characterization of primitive Pythagorean triples and
proceed but we take another approach. Now, the trivial identity (l + m)2 +
(l − m)2 = 2(l2 + m2 ) gives, on multiplication by (l + m)2 that

(l + m)4 + (l2 − m2 )2 = 2(l + m)2 (l2 + m2 ).

96
 An Irrational Walk and Why 1 is Not Congruent

The reason to do this is that we have an expression in terms of p, u

and Q as follows. Indeed, putting l + m = 2Q, l2 − m2 = p2 − u2 and
l2 + m2 = p2 + u2 − 2Q2 , we have

16Q4 + (p2 − u2 )2 = 8Q2 (p2 + u2 − 2Q2 ).

In other words,

(p2 − u2 )2
4Q4 − (p2 + u2 )Q2 + = 0.
8
This means that the discriminant (p2 +u2 )2 −2(p2 −u2 )2 = 6p2 u2 −p4 −u4
must be a perfect square, say d2 . But then the general algebraic identity

(p2 + u2 )4 − (6p2 u2 − p4 − u4 )2 = (4pu(p2 − u2 ))2

tells us that (p2 + u2 , d, 4pu(p2 − u2 )) is an integer solution of the equation

X 4 − Y 4 = Z 2 . As we already saw, this has only the trivial √ solution when
Y is 0. Note that d = (p2 + u2 )2 − 2(p2 − u2 ) 6= 0 since 2 is irrational.
Therefore, we have shown that a rational walk as above is impossible for
the same reason that 1 is not a congruent number; viz., that the Fermat
equation X 4 − Y 4 = Z 2 does not have integral solutions with Y 6= 0. It
would be interesting to directly relate the above ‘rational walk’ problem to
the congruent number problem for 1!

97
Covering the Integers

Let f : N → C be an arithmetic function. Then,P f can be recovered back

from the function (its Möbius transform) f (n) = d|n f (d) by the Möbius
b
inversion formula X
f (n) = µ(d)fb(n/d),
d|n

where the Möbius function is defined µ(n) = 1 if n = 1, µ(n) = (−1)r if

n = p1 p2 · · · pr for distinct primes p1 , · · · , pr and µ(n) = 0 if not. Here is an
elementary observation (in the spirit of the uncertainty principle as made
by Paul Pollack):
Lemma. Let f be any non-zero arithmetic function such that the support

{n : fb(n) 6= 0},

is a finite set. Then, the support

{n : f (n) 6= 0},

of f is infinite.
Suppose f is non-zero and that {n : f (n) 6= 0} is finite. Then,
Proof. P
F (z) := n≥1 fb(n)z n is a non-zero polynomial. But, if M = M ax(|f (n)| :
n ≥ 1), then for |z| < 1, we have
XX XX
|F (z)| = | ( f (d))z n | ≤ |f (d)||z|n
n d|n n d|n

XX X m|z|
≤ ( M )|z|n ≤ M n|z|n = < ∞.
n n
1 − |z|2
d|n

Therefore, we can interchange the summations to obtain

XX XX zd
F (z) = f (d)z rd = f (d) .
r r
1 − zd
d|n d|n

Now, if N = Max(n : f (n) 6= 0), then F (z) clearly has a pole at z =

e2iπ/N which contradicts the fact that F is an entire function.
This gives a proof of the infinitude of primes as follows.
The chapter is a modified version of an article that first appeared in Resonance, Vol. 17, No. 3,
pp. 284–290, March 2012.

99
 Covering the Integers

P
Note that µ b(n) = d|n µ(d) = 0 for n > 1. So, µ b has finite support.
So, if there were only finitely many primes, µ would have finite support
contradicting the ‘uncertainty’ result proved above.
The above type of argument has interesting applications to what are
known as covering congruences. These can be described as follows.
Just as every natural number is either odd or even, one can see that for
any k, the congruences
x ≡ 1 (mod 2), x ≡ 1 (mod 4), · · · , x ≡ 1 (mod 2k ), x ≡ 0 (mod 2k )
cover the set of all integers. In other words, every integer satisfies at least
one of these congruences. Similarly, for any positive integer N , the set of
congruences x ≡ i (mod N ) for i = 0, 1, · · · , N − 1 covers the integers.
Thus, these are called sets of ‘covering congruences’. The general defini-
tion is the following.
Let a1 , · · · , ak be integers and let n1 , · · · , nk be positive integers. The
set of congruences x ≡ ai (mod Sk ni ) for i = 1, · · · , k is called a set of
covering congruences if Z = i=1 (ai + ni Z). We write the set in short as
a1 (n1 ), a2 (n2 ), · · · ,
ak (nk ).
Note that in both the cases P above, the ‘moduli’ n1 , · · · , nk of the congru-
ences have the property that ki=1 n1i = 1.
A set of covering congruences a1 (n1 ), · · · , ak (nk ) is called a disjoint cov-
ering system if every integer satisfies exactly one of the congruences x ≡ ai
(mod ni ).
Note that the two sets of covering congruences above, viz.,

1(2), 1(22 ), · · · , 1(2k ), 0(2k )

and
0(N ), 1(N ), · · · , N − 1(N )
are disjoint covering systems. Note also that at least two of the moduli are
the same. That this must always be true is the assertion of the following
proposition whose proof follows the method we started the article with.
PROPOSITION
Let a1 (n1 ), a2 (n2 ), · · · , ak (nk ) be a disjoint system of covering
P congru-
ences where n1 ≤ n2 ≤ · · · ≤ nk . Then, nk−1 = nk . Moreover kj=1 n1j = 1.
Proof. We may assume without loss of generality that 0 ≤ aj < nj for
each j ≤ k. By the hypothesis, we have for |z| < 1,
k k ∞
X z aj XX
aj +rnj
X 1
n = z = zn = .
1−z j 1−z
j=1 j=1 r=0 n≥0

100
 Covering the Integers

Thus, this function has a pole at the point z = 0. But, if ni < nk for all
i < k, then this function has a pole at z = e2iπ/nk , a contradiction. This
proves the first assertion.
Let n = LCM (n1 , · · · , nk ). Then, we have
k
Y
1 − zn = (1 − e2iπaj /nj z n/nj ),
j=1

as each n-th root of unity is a simple root of the polynomial on the left
hand side. This can be rewritten as
X P P
1 − zn = (−1)|I| z j∈S n/nj e2iπ j∈I aj /nj .
I⊂{1,··· ,k}

Comparing degrees gives us kj=1 nnj = n which is the second assertion.

P

This leaves the question of the existence of a covering system (necessarily

not disjoint by the above proposition) with distinct moduli. The first set
of covering congruences for which the moduli are distinct was given by A
Schinzel. This is the set

0(2), 0(3), 1(4), 5(6), 7(12).

Here are two famous conjectures due to Erdös and others which are still
open.
Conjecture 1. (Erdös-Selfridge). There is no finite system of covering
congruences where all the moduli are distinct and odd.
Conjecture 2. (Erdös) For any M > 0, there exists a system of covering
congruences where all the moduli are distinct and > M .
Erdös mentioned in 1995 that the last conjecture is perhaps his favourite
problem!
We finish with a number-theoretic application which originally motivated
the study of covering congruences. This is the following beautiful result due
to (who else but!) Erdös.

Theorem. (Erdös, 1950). There are infinitely many odd integers which
are not expressible as the sum of a prime and a power of 2.
Proof. Consider a covering system of congruences ai (ni ) with 1 ≤ i ≤ k
where p1 , · · · , pk are distinct odd primes and ni the least positive number
satisfying 2ni ≡ 1 mod pi .
Erdös constructed such covering systems explicitly; one such is

0(2), 0(3), 1(4), 3(8), 7(12), 23(24)

101
 Covering the Integers

Note that the ni ’s here – 2, 3, 4, 8, 12, 24 – are the orders of 2 modulo the
primes 3, 7, 5, 17, 13, 241.
Given any such system, the Chinese remainder theorem provides a com-
mon solution to the congruences
x ≡ 2ai (mod pi ); 1 ≤ i ≤ k and x ≡ 1Q(mod 2).
Such a solution x is unique modulo 2 ki=1 pi . Consider the smallest pos-
itive integer solution, say x0 . Now, for each integer r, there exists some i
such that r ≡ ai (mod mi ).
So, if n ≡ n0 (mod 2p1 · · · pk ), then
n − 2r ≡ n0 − 2ai ≡ 0 (mod pi ).
Thus, if n − 2r > pi , then it must be composite.
This proves that all n not of the form 2r + pi for some r and some i ≤ k
are inexpressible in the form asserted. To dispose of the exceptional cases,
we may impose extra congruence conditions.
An example (taking the above covering system 0(2), 0(3), 1(4), 3(8), 7(12),
23(24) of Erdös) gives us:
No integer ≡ 7629217 (mod 1184810) is a sum of a power of 2 and an
odd prime.
Z-W Sun has done substantial work in the subject of covering congru-
ences and revealed connections with zero-sum problems. We finish with
a following amazing application completed by Sun of the above work by
Erdös and later work of Cohen, Selfridge on covering congruences.
Consider the 29-digit number

M = 66483084961588510124010691590
= 2.3.5.7.11.13.17.19.31.37.41.61.73.97.109.151.241.257.331.

Then, the solutions of the congruence

x ≡ 47867742232066880047611079 mod M

cannot be expressed as ±pa ± q b where p, q are primes and a, b are non-

negative integers.
As can be readily imagined, simple questions for integers arising out of
covering systems of congruences have many interesting analogues for groups
where the congruences are replaced by cosets of subgroups. That will be a
topic to discuss on another occasion.

102
S Chowla and S S Pillai:
The Story of Two Peerless
Indian Mathematicians1

Ramanujan may be a household name

in our country, but it is a shame
that not much is known about who later came.
Here, we talk about Chowla and Pillai
whose names in the mathematical landscape will lie
right at the top Any doubts? Illai Illai!
Sarvadaman Chowla (1907–1995) and S S Pillai (1901–1950) were two
of the foremost mathematicians to emerge from India in the generation
immediately after Ramanujan. The Mathematics Geneology Project lists
both Ramanujan and Chowla among the students of Littlewood! This arti-
cle specially features Chowla and Pillai. As a matter of fact, the June 2004
issue of Resonance journal had already featured Pillai [1] but we shall see
that a discussion of Chowla is necessarily intertwined with one of Pillai.
It has been mentioned by G H Hardy that after Ramanujan, the great-
est Indian mathematician was Pillai. We journey through some of the very
interesting and illuminating correspondences between Chowla and S S Pillai
which reveals also other personal and historical aspects. Apart from that,
we talk about Chowla’s and Pillai’s mathematical works. In a book of this
type, it is essential to select only those topics which are more elementary
or easy to describe while conveying some of the beauty and depth of the
ideas. Fortunately, in the works of Chowla and Pillai, we can find a verita-
ble treasury which is accessible at a level which can be enjoyed by even the
non-expert. Each of their works has an element of surprise and an element
of elegance and simplicity. They worked on a wide spectrum of areas of
number theory. We select some of their more elementary gems and discuss
their proofs. In fact, we attempt to retain as much of the original ideas in
the arguments as possible.
It is an enigma that even a layman may ask a question in elementary num-
ber theory which turns out to be nontrivial. The fact that several old prob-
lems in elementary number theory remain unsolved to this day, has been
referred to in different ways by people. To quote Professor K Ramachandra,

The chapter is a modified version of an article that first appeared in Resonance, Vol. 17, No. 9,
pp. 855–883, September 2012.
1 This chapter was part of an article written in collaboration with R Thangadurai.

103
 The Story of Two Peerless Indian Mathematicians

“in figurative terms, what has been solved can be likened to an egg-shell, and
what remains to be solved to the infinite space surrounding it.”

2. Chowla to Pillai Correspondence

Starting in the late 1920s, and up to one month before Pillai’s death,
Chowla and Pillai maintained regular correspondence. They published joint
papers starting in 1930 with a famous piece of work on the Euler’s totient
function. Some other themes that they collaborated on were concerned with
solutions to the Brahmagupta–Pell equation and the Waring problem. The
correspondence between these two stalwarts is mathematically illuminating
to read. It also reveals the intellectual honesty they possessed and the joy
each drew from the other’s successes.

104
The Story of Two Peerless Indian Mathematicians

105
The Story of Two Peerless Indian Mathematicians

106
 The Story of Two Peerless Indian Mathematicians

The letter shows the reluctance on Chowla’s part to be a co-author of

some result where he felt he had not contributed enough. The letter is
written by Chowla after he joined St.Stephen’s College in Delhi. Through
the years, Chowla consistently expresses almost in every letter his gladness
for the correspondence between them!
The number theorist K Ramachandra spoke of his first meeting with
Chowla at the Institute for Advanced Study in Princeton during the

107
 The Story of Two Peerless Indian Mathematicians

former’s first visit there. After discussing mathematics, Chowla got them
both bottles of ‘pepsi’ from a vending machine. After the meeting,
Ramachandra says that he ran around the premises muttering that he drank
pepsi with Chowla!
Chowla’s fertile imagination earned him the sobriquet of ‘poet of mathe-
matics’ from his associates. Chowla passed away in the U.S. in 1995 at the
age of 88. On the other hand, Pillai died tragically at the age of 49 in 1950.
Pillai was invited to visit the Institute for Advanced Study in Princeton
for a year. The flight which he boarded to participate in the 1950 Inter-
national Congress of Mathematicians tragically crashed near Cairo on the
31st of August. The readers are directed to Pillai’s collected works which
have appeared recently.

3. Mathematical Gems from Chowla and Pillai

Waring’s Problem
A discussion of Pillai’s mathematical work must start with Waring’s
problem and vice versa! However, since this has been written about in
detail in the June 2004 issue, we mention this problem in passing and refer
the readers to the above-mentioned Resonance article by C S Yogananda
[2].
Waring’s problem asks for the smallest number g(k) corresponding to
any k ≥ 2 such that every positive integer is a sum of g(k) numbers each
of which is the k-th power of a whole number. Hilbert had shown that such
a finite number g(k) does exist. The ideal Waring’s conjecture predicts
a particular value of g(k). Indeed, if 3k is divided by 2k , the quotient is
[(3/2)k ], and some remainder r, where [t] denotes the greatest integer less
than or equal to t. Now, the number

2k [(3./2)k ] − 1 = ([(3/2)k ] − 1)2k + (2k − 1)1k

is a sum of 2k + [(3/2)k ] − 2 numbers which are k-th powers and is not the
sum of a smaller number of k-th powers. Hence,

g(k) ≥ 2k + [(3/2)k ] − 2.

This ideal Waring conjecture asserts that this lower bound is the cor-
rect bound also. Pillai proved, among other things, that this ideal Waring
conjecture holds good under the condition on k that the remainder r on
dividing 3k by 2k satisfies r ≤ 2k − [(3/2)k ] − 2 (chapter 21 of [3]). This
is known to hold for all k ≤ 471600000. At this time, the ideal Waring
conjecture is known to hold for all large enough k.

108
 The Story of Two Peerless Indian Mathematicians

4. The Least Prime Quadratic Residue

Chowla’s lifelong pre-occupation with class number of binary quadratic
forms led him to discover some rare gems on the way, so to speak! An inter-
esting problem, useful in cryptography, for instance, is to find for a given
prime p, the smallest prime q which is a quadratic residue (that is, a square)
modulo p. For example, the quadratic reciprocity law tells us that if p ≡ ±1
modulo 8, then 2 is the least quadratic residue mod p. Chowla [4] proved
the following beautiful result:
Let p > 3 be a prime such that p ≡ 3 mod 8. Let l(p) denote the least
prime which is a quadratic residue mod p. If the number h(−p) of classes
q
p
of binary quadratic forms of discriminant −p is at least 2, then l(p) < 3.
p+1 p+1
If h(−p) = 1, then l(p) = 4 (and, therefore, 4 is prime!).
Remarks. (i) The theorem implies, in particular, that for primes
q p > 3,
p ≡ 3 mod 8, we have l(p) = 4 if and only if h(−p) = 1 because p3 < p+1
p+1
4
for p > 3.
(ii) The proof of the theorem is easy and uses Minkowski’s reduction
theory of quadratic forms which produces in each equivalence class of
positive-definite forms, a unique one ax2 + bxy + cy 2 which is ‘reduced’
in the sense that |b| ≤ a ≤ c.

5. Chowla’s Counter-examples to a Claim of Ramanujan and a

Disproof of Chowla’s Conjecture
Among Ramanujan’s numerous astonishing results, there are also occa-
sional lapses. One such was his ‘proof’ (in his very first paper of 1911) that
the numerators of Bernoulli numbers are primes. This is false; for instance,
denoting by Bn the Bernoulli number defined by
z X zn
= B n ,
ez − 1 n!
n≥0

and by Nn , the numerator of Bn /n, the numbers N20 , N37 are composite.
In 1930, Chowla showed [5] that Ramanujan’s claim has infinitely many
counter-examples. Surprisingly, Chowla returns to this problem 56 years
later(!) in a joint paper with his daughter [6] and poses as an unsolved
problem that Nn is always square-free. In a recent article, Dinesh Thakur
[7] points out that Chowla’s question has infinitely many negative
answers. Indeed, Thakur shows:
For any fixed irregular prime p less than 163 million, and any arbi-
trarily large k, there exists a positive integer n such that Nn is divisible
by pk .

109
 The Story of Two Peerless Indian Mathematicians

If we observe (from the existing tables) that 372 divides N284 , Chowla’s
question has a negative answer. The proof of the more general assertion uses
the so-called Kummer congruences which essentially assert that the value of
(pn−1 −1)Bn
n modulo pk depends (for even n) only on n modulo pk−1 (p − 1),
if p − 1 does not divide n. Using this as well as certain functions called
p-adic L-functions, the general assertion of arbitrarily large powers can
also be obtained.

6. Problem on Consecutive Numbers

Pillai proved in 1940 that any set of n consecutive positive integers where
n ≤ 16, contains an integer which is relatively prime to all the others.
However, there are infinitely many sets of 17 consecutive integers where
the above fact fails. For instance, N + 2184, N + 2185, · · · , N + 2200 is such
a set whenever N is a multiple of 2.3.5.7.11.13 = 30, 030. Moreover, Pillai
also proved [8] that for any m ≥ 17, there are infinitely many blocks of m
consecutive integers for which the above property fails. Now, generalizations
to arithmetic progressions instead of consecutive numbers are known.

7. How Spread-out are Perfect Powers?

Look at the sequence of all perfect powers of positive integers:
1, 4, 8, 9, 16, 25, 27, 32, 36, 49, 64, 81, 100, 121, 125, · · · .
We observe that differences between consecutive terms can be: 1 (9 − 8),
2 = (27 − 25), 3 (4 − 1), 4 (36 − 32), 5 (32 − 27) etc.
Pillai conjectured that consecutive terms can be arbitrarily far apart [9].
In other words, given any number, one can find consecutive terms whose
difference is larger than that given number. Equivalently, the conjecture
asserts:
Given a positive integer k, the equation xp − y q = k has only finitely
many solutions in positive integers x, y, p, q ≥ 2.
This is unproved as yet even for one value of k > 1 although it is known
now that if one of these 4 parameters is fixed, the finiteness holds.

8. Independent Values of the Cotangent Function

A typical aspect of Chowla’s works has been to come back to an old result
after several years and applying it in an unexpected manner. On 9.2.1949,
Chowla had written to Carl Ludwig Siegel about a certain non-vanishing of
a particular type of series. Three days later, he received a reply from Siegel,
improving the result. In 1970, Chowla, while wondering about relations
between the roots of a certain polynomial, realized that not only could he

110
 The Story of Two Peerless Indian Mathematicians

re-prove Siegel’s improved version in a simpler fashion [10], he could use

this old result to prove what he wanted about the roots! Let us discuss it
briefly.
If p is a prime number, consider the values xr = cot(rπ/p) of the cotan-
Pp−1
gent function, for 0 < r < p. Evidently, xr + xp−r = 0. Also, r=1 xr = 0
but this is easily deduced from the earlier relations. So, a natural ques-
tion is:
Pp−1
Are all the linear relations of the form i=1 ar xr = 0 with ai ∈ Q,
consequences of the relations xr + xp−r = 0 for 1 ≤ r < p?
Indeed, xr ’s are the roots of an irreducible polynomial over Q, of degree
p − 1 and, one may ask for possible linear relations among the roots of any
irreducible polynomial over Q. Chowla’s theorem asserts:
Let p be a prime number and xr = cot(πr/p) for r = 1, · · · , p−1
2 . If ai ∈ Q
P p−1
such that i=1 2
ai xi = 0, then ai = 0 for all i ≤ (p − 1)/2.
Chowla uses some very basic Galois theory to deduce that, under the as-
P p−1
2
sumption i=1 ai xi = 0 above, there are (p−1)/2 such linear relations, viz.,
a1 x1 + a2 x2 + · · · + a p−3 x p−3 + a p−1 x p−1 = 0
2 2 2 2

a1 x2 + a2 x3 + · · · + a p−3 x p−1 + a p−1 x1 = 0

2 2 2

..
.
a1 x p−1 + a2 x1 + · · · + a p−1 x p−3 = 0.
2 2 2

If the ai ’s are not all 0, this leads to the vanishing of the ‘cirulant’
determinant  
x1 x2 · · · x(p−1)/2
 x2 x3 · · · x1 
.
 
 .. ..
 . . 
x(p−1)/2 x1 · · · x(p−3)/2
At this point, Chowla quotes the well-known value of this determinant
and proceeds.
What is the value of this determinant?
In the 3 × 3 case  
x1 x2 x3
x2 x3 x1  ,
x3 x1 x2
has determinant 3x1 x2 x3 − x31 − x32 − x33 .
Let 1, ω, ω 2 be the cube roots of unity.
If x1 , x2 , x3 are replaced by ωx1 , ω 2 x2 , x3 or by ω 2 x1 , ωx2 , x3 , the expres-
sion remains the same. As x3 = −x1 −x2 leads to 3x1 x2 x3 −x31 −x32 −x33 = 0,

111
 The Story of Two Peerless Indian Mathematicians

the three expressions x1 + x2 + x3 , ωx1 + ω 2 x2 + x3 , ω 2 x1 + ωx2 + x3 are

factors.  the determinant in the 3 × 3 case is given as:
In other words,
x1 x2 x3
det x2 x3 x1  = −(x1 + x2 + x3 )(ωx1 + ω 2 x2 + x3 )(ω 2 x1 + ωx2 + x3 ).
x3 x1 x2
 
x1 x2 · · · x(p−1)/2
 x2 x3 · · · x1 
Similarly, in general, the determinant of 
 
. . . . 
 . . 
x(p−1)/2 x1 · · · x(p−3)/2
equals the product (up to sign) of
r(p−1)
ω r x1 + ω 2r x2 + · · · + ω 2 x p−1
2
4iπ
as r varies from 1 to p−1
2 where ω = e , (p − 1)/2-th root of unity.
p−1

Here, Chowla realizes with surprise that the above factors are (up to
certain non-zero factors) none else other than the special value at s = 1
of certain functions L(s, χ) = ∞ χ(n)
P
n=1 n called Dirichlet L-functions cor-
responding to Dirichlet characters χ modulo p which satisfy χ(−1) = −1.
The non-vanishing of these are the older result mentioned above and show,
thus, that the determinant is non-zero. Hence, the linear independence of
the cot(rπ/p) for 1 ≤ r ≤ (p − 1)/2 is established.
Remarks. (i) Kai Wang closely followed Chowla’s proof to generalize his
theorem to non-primes and to derivatives of the cotangent function [11] by
showing:
φ(k)
 s ≥0 and an arbitrary natural number k, the 2 real numbers
For any
ds rπ
dxs cot x+ k (where r ≤ φ(k)
2 and (r, k) = 1) are linearly independent
x=0
over Q.
(ii) The non-vanishing at s = 1 of L(s, χ) for nontrivial Dirichlet char-
acters is the key fact used in the proof of Dirichlet’s famous theorem on
existence of infinitely many primes in any arithmetic progression an + b
with (a, b) = 1.
(iii) Kenkichi Iwasawa (see [12]) showed later in 1975 that the above
result has connections with the so-called “regular” primes. The definition
of regular primes is not needed here and, we merely recall that Kummer had
proved that the Fermat equation xp + y p = z p has no non-zero solutions
in integers if p is a ‘regular prime’. It is unknown yet whether there are
infinitely many regular primes (although Fermat’s last theorem has been
proved completely) – surprisingly, it has been known for a long time that
there are infinitely many irregular primes! The connection of the above
result of Chowla with regular primes is the following.

112
 The Story of Two Peerless Indian Mathematicians

The linear independence of the (p − 1)/2 cotangent values ensures that

there exist rational numbers t1 , · · · , t p−1 so that
2

p−1
2
2π X
2 sin = tr cot(2rπ/p).
p
r=1

Iwasawa shows the prime p is regular if and only if, none of the tr ’s have
denominators which is a multiple of p and, at least one tr has numerator
also not divisible by p.

9. On the Number of Permutations of a Given Order

Chowla wrote a series of papers on generating functions for the number of
permutations of a given order etc. In the permutation group Sn , let An (d)
denote the number of permutations σ satisfying σ d = Id. In collaboration
with Herstein and Scott, he showed ([13]):
∞
X An (d)xn X xk
= exp( ).
n! k
n=0 k|d

Here, convenience of notation, one takes A0 (n) = 1. Let us prove this

beautiful, useful fact.
We look for a recursive relation for An (d) in terms of Ak (d) for k < n.
Look at what happens to the symbol n under any permutation contribut-
ing to An (d). If the symbol is fixed, then the rest of the n − 1 symbols can
be permuted in An−1 (d) ways. Now, suppose the symbol n is a part of a
k-cycle for some 1 < k ≤ n. Note that any permutation contributing to
An (d) has some order dividing d; thus, if it has a k-cycle in its decompo-
sition, then k|d. Now, each k-cycle contributes An−k (d) elements. As there
are (n − 1)(n − 2) · · · (n − k + 1) ways to choose such k-cycles, we get
X
An (d) = An−1 (d) + (n − 1) · · · (n − k + 1)An−k (d).
k|d;1<k≤n

This can be rewritten as

An (d) X An−k (d)
= .
n! (n − k)!
k|d;1≤k≤n

Therefore, the generating function f (x) = n≥0 Ann!(d) satisfies

P

X Ai (d)xi X  X Ai−k (d) 

0
xf (x) = = xi ,
i! (i − k)!
i≥1 i≥1 k|d,1≤k≤i

on using the recursion above.

113
 The Story of Two Peerless Indian Mathematicians

Combining the terms corresponding to a particular Aj (d), we have

therefore
X Aj (d) X X
xf 0 (x) = xj+k = f (x) xk .
j!
j≥0 k|d k|d

This is a differential equation

f 0 (x) X k−1
= x
f (x)
k|d

whose general solution is obtained by integration as

X k
x
f (x) = c.exp
k
k|d

for some constant c. Since f (0) = A0 (d) = 1 = c, we get the assertion

 
X An (d) X xk
= exp  .
n! k
n≥0 k|d

This formula is useful in a number of ways. For instance, one can get
an asymptotic estimate of how fast An (d) grows with n (for any fixed d).
Moreover, for d = p, a prime, this gives a simple-looking closed formula for
An (p) for any n.
9.1 Closed Form for the Prime Case
In the above identity
 
X An (d) X xk
= exp  .
n! k
n≥0 k|d

take d = p, a prime and note that

X An (p) p /p
X xi X xpj
= ex ex = .
n! i! pj j!
n≥0 i≥0 j≥0

Comparing the coefficients of xn , we obtain

X n!
An (p) = .
pj j!i!
i+pj=n

In particular, this number is a positive integer for each n ≥ p(!)

This is an exclamation mark, not a factorial!

114
 The Story of Two Peerless Indian Mathematicians

Also, a classical theorem in the theory of groups, due to Frobenius, asserts

that in any finite group G, the number of elements satisfying xd =identity
(for any divisor d of the order of G) is a multiple of d. Thus, we have An (d)
is a multiple of d for each n ≥ d.
This statement for An (p) gives, therefore, that i+pj=n pjn!j!i! is a multiple
P
of p for every n ≥ p and, the special case n = p is known as Wilson’s
theorem.
9.2 Applications to Finite Groups
Apart from being useful in its own right, the study of the numbers An (d)
has connections to some counting problems in groups. Notice that the num-
ber An (d) of permutations in Sn which satisfy xd = identity, is nothing
else than the number of group homomorphisms from a cyclic group of
order d to Sn – each homomorphism associates a permutation σ satisfying
σ d = Identity, to a fixed ‘generator’ of the cyclic group. For any group G
(not necessarily cyclic), knowledge of the numbers hn of group homomor-
phisms from G to Sn for various n, allows us to find a recursive expression
for the number of subgroups of G which have a given index in it. In fact,
if sn is the number of subgroups of index n in G, then
h1 h2 hn−1 hn
sn + sn−1 + sn−2 + · · · + s1 = .
1! 2! (n − 1)! (n − 1)!

10. Convenient Numbers and Class Number

Euler observed that 18518809 = 1972 + 1848(100)2 is a prime. In fact,
Euler was interested in producing large primes of the form x2 + ny 2 for
various values of n. It happens (and is easy to prove) that a number which
has a unique expression of the form x2 + y 2 is prime. Thus, one may hope
this is true for expressions of the form x2 + ny 2 also for any n. However,
as Euler noted [14], this holds only for a certain set of values of n. He con-
structed explicitly a set of 65 positive integers for which this is true (the
largest of which is 1848) – he called such numbers ‘idonean’ or ‘convenient’.
To this day, it is not proven that Euler’s list is complete [15]. However, a
beautiful result of Chowla shows at least that the list of idonean num-
bers is finite! To explain how it is done, we very briefly define and dis-
cuss binary quadratic forms – another name for expressions of the form
ax2 + bxy + cy 2 .
A binary, integral quadratic form is a polynomial f (x, y) = ax2 +bxy+cy 2
where a, b, c are integers. It is primitive if (a, b, c) = 1. The integer b2 − 4ac
is known as the discriminant.
As
4af (x, y) = (2ax + by)2 + (4ac − b2 )y 2 = (2ax + by)2 − dy 2 ,

115
 The Story of Two Peerless Indian Mathematicians

when d < 0 and a > 0, f (x, y) takes only positive values (excepting the
value 0 at x = y = 0). Thus, when a > 0, and the discriminant is negative,
the form is positive-definite.
For example, x2 + ny 2 is a positive-definite form with discriminant −4n.
Two forms f (x, y) and g(x, y) are said to be equivalent
or, said
 to be in the
α β
same class if f (αx+βy, γx+δy)) = g(x, y) where A = ∈ SL(2, Z),
γ δ
an integer matrix of determinant 1.
The motivation behind this definition is:
Equivalent forms take the same sets of values as x, y vary over integers.
This is clear because one may also write

g(x, y) = f (αx + βy, γx + δy))

in the form
f (x, y) = g(δx − βy, −γx + αy)).
 
δ −β
Notice that the matrix is simply the inverse of the matrix
−γ α
 
α β
.
γ δ
Moreover, equivalent forms have the same discriminant. Gauss showed:
For d < 0, the number h(d) of classes of primitive, positive-definite binary
quadratic forms of discriminant d is finite.
Gauss conjectured that h(d) → ∞ as −d → ∞. This was proved by
Heilbronn. By a modification of Heilbronn’s argument, Chowla proved the
following fact which was another conjecture of Gauss [16]:
h(d)
2t → ∞ as −d → ∞ where t is the number of primes dividing d.
This interesting fact is useful in a totally different context which we
indicate briefly now.
Euler obtained the following list of 65 numbers called ‘Numerus idoneus’
(or ‘convenient’ numbers):
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 15, 16, 18, 21,
22, 24, 25, 28, 30, 33, 37, 40, 42, 45, 48, 57, 58,
60, 70, 72, 78, 85, 88, 93, 102, 105, 112, 120, 130,
133, 165, 168, 177, 190, 210, 232, 240, 253, 273, 280,
312, 330, 345, 357, 385, 408, 462, 520, 760, 840, 1320, 1365, 1848.
Consider any odd number m co-prime to n which is expressible as
m = x2 + ny 2 with (x, ny) = 1. If each such m which has a unique ex-
pression of the form x2 + ny 2 in positive integers x, y is necessarily prime,
the number n is said to be idonean.

116
 The Story of Two Peerless Indian Mathematicians

As mentioned in the beginning of this section, Euler observed that

18518809 = 1972 + 1848(100)2 is a prime by showing 1848 is idonean.
Firstly, it is not even clear whether the list of idonean numbers is finite
or not. This can be analyzed (and was analyzed by Gauss) using the theory
of quadratic forms.
Firstly, we recall one more notion – the genus.
Two primitive, positive-definite forms of discriminant d are said to be
in the same genus if they take the same set of values modulo d. As forms
in the same class take the same set of values, they are in the same genus.
Each genus, therefore, consists of finitely many classes.
Gauss proved (modulo some gaps which were filled later by Grübe):
A positive integer n is idonean if and only if, for forms of discriminant
−4n, every genus consists of a single class.
Chowla’s theorem h(d)
2t → ∞ as −d → ∞ where t is the number of primes
dividing d, which was quoted above, implies that for large enough −d, each
genus has more than one class of forms. Therefore, by Chowla’s theorem,
the set of idonean numbers is finite!
As a matter of fact, Euler’s list is expected to be a complete one (this
has been proved to be so under the assumption of a deep conjecture known
as the generalized Riemann hypothesis).

11. Matrices and Quadratic Polynomials

As we saw earlier, the equivalence classes of integral, binary quadratic
forms are related to the group SL2 (Z) of integral matrices of determinant
1. Recall also that equivalent forms take the same sets of integer values as
x, y vary over integers. The following result of Chowla with J Cowles and
M Cowles [17] shows that the relation is an intimate one. We recall that
two matrices A, B are said to be conjugate if there is an invertible matrix
P such that B = P AP −1 . The trace of a matrix is the sum of its diagonal
entries. Then:
For all integers t 6= ±2, the number of conjugacy classes of matrices in
SL2 (Z) with trace t, equals the number of equivalence classes of integral,
binary quadratic forms with discriminant t2 − 4.
Recall the discriminant of a quadratic form ax2 + bxy + cy 2 was defined
in the last section to be the number b2 − 4ac. Here is the easy proof of the
above theorem.  
a b
Associate to each matrix M := ∈ SL2 (Z), the quadratic form
c d

m(x, y) := bx2 + (d − a)xy − cy 2 .

117
 The Story of Two Peerless Indian Mathematicians

Note that if the trace of M is t, then a + b = t and, therefore, the

discriminant of m(x, y) is
(d − a)2 + 4bc = (d + a)2 − 4(ad − bc) = t2 − 4.
 
−1 α β
For a conjugate matrix N := AM A where A = ∈ SL2 (Z),
γ δ
 0 0
a b
write N as . Then, the form n(x, y) = b0 x2 + (d0 − a0 )xy − c0 y 2 is
c0 d0
easily seen to be m(x0 , y 0 ) where
 0    
x t x αx + γy
=A = .
y0 y βx + δy
In other words, under the above association, conjugate matrices of trace
t correspond to equivalent forms of discriminant t2 − 4.
Conversely, associate to a quadratic form f (x, y) = px2 + qxy + ry 2 with
discriminant t2 − 4 (so, q 2 − 4pr = t2 − 4), the matrix
 t−q 
2 p
F := ∈ SL2 (Z).
−r t+q 2
2 2
Note that indeed, det F = t −q 4 + pr = 1 and trace F = t.
Further,
 look at any equivalent form f 0 (x, y) = f (ax + by, cx + dy) with
a b
M := .
c d ∈ SL2 (Z)
Write f 0 (x, y) = p0 x2 + q 0 xy + r0 y 2 .
Then, we compute and see that
!
  t−q t0 −q 0
p0
  
t t −1 a c 2 p d −c 2
M F (M ) = = 0 0 ,
b d −r t+q 2
−b a −r0 t +q 2

which shows that the corresponding matrices are conjugate in SL2 (Z).
The above associations are inverse to each other and proves the proposi-
tion.
Remarks. The above association is also useful in deciding if two
 matri-

1 3
ces are conjugate in SL2 (Z) or not. For instance, the matrices
3 10,
 
1 1
which have trace 11 are associated to the quadratic forms
9 10
3x2 + 9xy − 3y 2 , x2 + 9xy − 9y 2 ,
respectively. However, they are evidently in-equivalent because the first one
takes only multiples of 3 as values whereas the second one takes values like
1 at (x, y) = (1, 0).

118
 The Story of Two Peerless Indian Mathematicians

12. Average of Euler’s φ-function

Euler’s φ-function, denoted by φ(n) is an arithmetic function defined on
natural numbers that counts the number of natural numbers 1 ≤ m ≤ n
with (m, n) = 1. Euler gave a formula which can be proved using inclusion
– exclusion principle as follows.
Y 1

φ(n) = n 1− ,
p
p|n

where the product varies over all the distinct prime divisors. This formula
shows that the functional value fluctuates a lot.
In analytic number theory, to study such fluctuating arithmetical func-
tions, one often looks at their average behaviour. One can prove that the
average value from 1 to x is π32 x but the interesting part is to have an idea
of the error which would be introduced if we take this value. In analytic
number theory, this methodology of ‘determining the main term and esti-
mating the error term’ is fundamental because we cannot deduce anything
concrete if the error term is of the same order as the main term! One has
X 3
φ(n) = 2 x2 + E(x),
π
1≤n≤x

where E(x) is the remainder or error term in the average.

Dirichlet showed that for any given  > 0, there is a constant C > 0 so
that |E(x)| ≤ Cx1+ for all x > 0. Later, this was improved to |E(x)| ≤
C 0 x log x for all x > 0 by Mertens.
X 3
Sylvester prepared a table of values for φ(n) and 2 x2 for all x =
π
n≤x
1, 2, · · · , 1000. However, he failed to notice that E(820) < 0 and made
a conjecture that E(x) ≥ 0 for all x. In 1929, Chowla wrote a letter to
Pillai where he predicted that E(x) > 0 for infinitely many values of x and
E(x) < 0 for infinitely many values of x.
In order to prove an error term, say |E(x)| ≤ cg(x), is tight for some
non-negative function g(x), one needs to produce a positive constant c0
and infinitely many x’s such that |E(x)| > c0 g(x).
Such a result is called an ‘omega’-result in analytic number theory; we
write E(x) = Ω(g(x)).
Chowla and Pillai showed that E(x) = Ω(x log log log x).
Such a result took many years to generalize. There is a conjecture by
Montgomery that
E(x) = O(x log log x) and E(x) = Ω± (x log log x),
which is still open.

119
 The Story of Two Peerless Indian Mathematicians

13. A Variant of Tic-Tac-Toe!

In 1933, Pillai studied a variant of Tic-Tac-Toe game as follows. Let n ≥ 3
be an integer and t ≤ n be another integer. Suppose n × n grid with n2
squares is given in the plane. Let P and Q be two players playing. By turns
each mark a square. The rule of playing the game is as follows. Suppose
P is the starter and, let Pr and Qr be the squares marked respectively by
P and Q during their r-th turns. After P ’s (or Q’s) s-th turns, whoever
marked t squares in a straight line wins the game.
Pillai proves that when t = n and the game is carefully played, then it
will end always in a draw. However, if t < n, then for a given t, there is a
function f (n) depending on n such that if t ≥ f (n), then the game ends
in a draw. When t < f (n), he proved pthat the player who starts will win.
Also, he proved that f (n) ≤ n + 1 − n/6 and f (n) = n for all n = 3, 4, 5,
and 6. For large values of n, the correct order of f (n) is still unknown!

14. Smooth Numbers

Smooth numbers are numbers which have only ‘small’ prime factors. For
example, 1,620 has prime factorization 22 ×34 ×5; therefore 1620 is 5-smooth
because none of its prime factors is greater than 5. Smooth numbers have
a number of applications to cryptography. For example, the very smooth
hash functions are used constructively to get a provably secure design. They
also play a role in music theory apparently (Longuet–Higgins, H C, Letter
to a musical friend, Music Review (August: 244-248), 1962!)
For other applications, the interested reader may also consult [18] and [19].
For any real numbers x, y > 1 with y ≤ x, we define ψ(x, y) to be the
number of positive integers t ≤ x such that if a prime p|t, then p ≤ y. In
other words, ψ(x, y) counts all the y-smooth numbers up to x. Ramanujan
(in a letter to Hardy) was the first to study these smooth numbers when
y = 3(!)
He obtained a nice asymptotic formula for ψ(x, 3).
In the 7-th conference of Indian Mathematical Society during 3–5, April,
1931 at Trivandrum, Pillai extended the above result of Ramanujan which
implies an asymptotic formula for ψ(x, y) if y > 1 is a fixed real number.
This is technical to state but we mention it here in passing, for the interested
reader:
If p1 , p2 , · · · , pr ≤ y are all the prime numbers less than y, then
(log x)r x)r−1Qlog(p1 ···pr )
ψ(x, y) − r!
Q r + (log
2(r−1)! ri=1 log pi
)
i=1 log pi
r−1 →0
log x
when x → ∞.

120
 The Story of Two Peerless Indian Mathematicians

Around that time, Dickman obtained an asymptotic result for ψ(y u , y)

for any fixed u > 0. The word ‘asymptotic’ here refers to an assertion of
ψ(y u ,y)
the form ‘what is the limit of yu as u → ∞’ ?
A more rigorous proof of Dickman’s result, in modern standards, was sup-
plied by Chowla and T Vijayaraghavan in 1947 where they used an unpub-
lished result of Pillai which is more general than the above result of Pillai!
It should be mentioned that Ramanujan – see page 337 of S Ramanujan,
The lost notebook and other unpublished papers. Narosa Publishing House,
(1988) – had the following entry. We write in the standard notations as
above:  Z 1 
du
ψ(x, xc ) ∼ x 1 − if 1/2 ≤ c ≤ 1;
c u
!
Z 1 Z 1/2 Z 1−v
c du dv du
ψ(x, x ) ∼ x 1 − + if 1/3 ≤ c ≤ 1/2;
c u c v v u
!
Z 1 Z 1/2 Z 1−v Z 1/3 Z (1−z)/2 Z 1−v
c du dv du dz dv du
ψ(x, x ) ∼ x 1 − + −
c u c v v u c z z v v u

if 1/4 ≤ c ≤ 1/3; and so on.

This is nothing else than Dickman’s asymptotic formula for ψ(x, y)!

15. Chowla’s Argument and the Langlands Conjecture

In this last section, we mention how an argument due to Chowla plays a
role in the famous Langlands conjectures. The cognoscenti would know that
the latter conjectures drive much of the contemporary research in number
theory [20].
This section is mostly taken from a lecture of Professor Ram Murty at
Kerala School of Mathematics, Calicut – thanks go to Professor Ram Murty
for allowing inclusion of the contents in this write up.
For each integer n ≥ 1, define d(n) is the number of positive divisors of
n. In his famous paper on ‘highly composite numbers’, Ramanujan gave an
upper bound for the function d(n) as follows:
log N
d(N ) ≤ 2 log log N for all N ≥ 2

and he produced infinitely many integers N for which the above bound is
attained.
For any given  > 0, we can deduce from the above upper bound that
there is a constant N0 depending on  so that

d(N ) ≤ N  for all N ≥ N0 .

Chowla proved this deduction using another argument involving Dirichlet

series.

121
 The Story of Two Peerless Indian Mathematicians

Let r ≥ 1 be any integer and let

∞
X dr (n)
Lr (s) = where s ∈ C with <(s) > 1.
ns
n=1

Chowla observed that the series

Y 2r 3r

Lr (s) = 1 + s + 2s + · · ·
p
p p

(where the product runs over all the prime numbers) converges absolutely
for Re(s) > 1 for all r ≥ 1. In particular, when s = 2, this series converges.
So,
∞
X d(n)r
< ∞ for all integers r ≥ 1.
n2
n=1

d(n)r
Therefore, the n-th term which is tends to zero. In particular, it
n2
is bounded for all large enough n’s. Thus, we get

d(n)r ≤ cn2 for all n ≥ M

for some M > 0 and c > 0 constants and this is true for all r ≥ 1.
Thus, for all n ≥ M , we get d(n) ≤ c1/r n2/r . Also, note that c1/r → 1 as
r → ∞.
Given  > 0, we can find r0 such that 2/r <  for all r ≥ r0 and we get
d(n) ≤ n .
To show how this sort of argument plays a role in the famous
Langlands conjectures, we describe such a conjecture. This can be done
through the famous Ramanujan delta function. Ramanujan studied the
following q-series
∞
Y
∆(z) = q (1 − q n )24 where z ∈ C with =(z) > 0; q = e2πiz
n=1

which is often called Ramanujan’s delta function because of his funda-

mental contribution to it although it was already studied by Jacobi and
others.
The delta
 function
 satisfies the following property:
a b
for all
c d
∈ SL2 (Z) (that is, for integers a, b, c, d with ad − bc = 1),
 
az + b
∆ = (cz + d)12 ∆(z).
cz + d

122
 The Story of Two Peerless Indian Mathematicians

 
1 1
Since ∈ SL2 (Z) and by the above relation, we see that ∆(z+1) =
0 1
∆(z) and hence ∆ function is a periodic function.
Therefore, it has a Fourier expansion. It can be proved that the Fourier
expansion of ∆(z) is
X∞
∆(z) = τ (n)e2iπnz ,
n=1
where τ (n) is the Fourier coefficients which are integers.
Traditionally, one writes q = e2iπz so that ∆(z) = ∞ n
P
n=1 τ (n)q .
Ramanujan computed the initial τ values and conjectured the following
relations.
1. τ (mn) = τ (m)τ (n) whenever (m, n) = 1.
2. τ (pa+1 ) = τ (p)τ (pa ) − p11 τ (pa−1 ) for all primes p and a ≥ 1.
3. |τ (p)| < 2p11/2 for every prime p.
The first two conjectures were proved by Mordell in 1917 and the third one
was proved by P Deligne in 1975 using deep algebraic geometry.
Note that the third conjecture of Ramanujan is equivalent to the
assertion:
11
τ (n) = O(n 2 + )
for any given  > 0.
Our interest is in this version of Ramanujan’s conjecture.
Let us define for each integer n ≥ 1,
τn = τ (n)/n11/2 .
Then Ramanujan’s conjecture is equivalent to τn = O(n ) for any given
 > 0. Define the L-series attached to ∆ function as
∞
X τn
L(s, ∆) = ,
ns
n=1
where s ∈ C with =(s) > 0.
Since τ (n) is a multiplicative function (the first conjecture of Ramanujan
mentioned above and proved by Mordell), we see that τnr is also a multi-
plicative function and hence we get
Y τp τp2

L(s, ∆) = 1 + s + 2s + · · · .
p
p p

Using the second conjecture of Ramanujan, one notes that for all primes
p, we have
∞
X 1 1
τpa X a = 2
= ,
1 − τp X + X (1 − αp X)(1 − βp X)
a=0

123
 The Story of Two Peerless Indian Mathematicians

where αp and βp are the complex roots of X 2 − τp X + 1. Note that

αp + βp = τp and αp βp = 1.
1
!−1
Y αp −1 βp −1 Y Y αp1−m βpm
  
L(s, ∆) = 1− s 1− s = 1− .
p
p p p
ps
m=0

For any r ≥ 1, Langlands defined the function:

r
!−1
YY αpr−m βpm
Lr (s, ∆) = 1− .
p m=0
ps

He conjectured that for every r ≥ 1, Lr (s, ∆) defines a series which is

absolutely convergent for <(s) > 1.
Note that the Dirichlet series
∞
X τ 2rn
Sr :=
ns
n=1

can be written as a product of the Lk (s, ∆) for k ≤ r. Therefore, if the

conjecture of Langlands is true, then Sr converges absolutely for <(s) > 1
for every r ≥ 1. This implies, by Chowla’s argument, that
τn2r
≤ C ⇐⇒ τn ≤ C 1/2r n1/r
n2
for all n ≥ n0 . Thus, we arrive at τn = O(n ) for any given  > 0 (!).
However, at present Langlands’s conjecture is known only for all r ≤ 9.

Collected works of Chowla and of Pillai:

The collected works of Chowla and of Pillai contain unpublished papers
also. The interested readers can look at:

• Collected works of S Chowla, Vol.1,2,3, Edited by James G Huard

and Kenneth S Williams, CRM Univ. de Montreal, 1999.

• Collected works of S S Pillai, Edited by R Balasubramanian and

R Thangadurai, Ramanujan Mathematical Society Collected Works
Series, 2010.

References
[1] B Sury, Box 1, Resonance, Vol. 9, No. 6, 2004.
[2] Waring’s problem and the circle method, Resonance, Vol. 9, No. 6,
pp.51–55, 2004.

124
 The Story of Two Peerless Indian Mathematicians

[3] G H Hardy and E M Wright, An introduction to the theory of numbers,

Oxford Univ. Press, third ed., 1954.
[4] S Chowla, The least prime quadratic residue and the class number,
J. Number Theory, 22, 1–3, 1986.
[5] S Chowla, On a conjecture of Ramanujan, Tohoku Math. J. 33, 1–2,
1930.
[6] S Chowla and P Chowla, Some unsolved problems, Norske Vid. Selsk.
Forth. (Trondheim), p.7, 1986.
[7] D Thakur, A note on numerators of Bernoulli numbers, Proc. Amer.
Math. Soc., Electronically published on 24.02.2012.
[8] S S Pillai, On a linear diophantine equation, Proc. Indian Acad. Sci.
A, 12, pp.199–201, 1940.
[9] S S Pillai, On the inequality 0 < ax − by ≤ n, Journal Indian M. S.,
19, pp. 1–11, 1931.
[10] S Chowla, The nonexistence of nontrivial linear relations between the
roots of a certain irreducible equation, J. Number Theory, 2, 120–123,
1970.
[11] Kai Wang, On a theorem of Chowla, J. Number Theory, 15, 1–4, 1982.
[12] Raymond Ayoub, On a theorem of Chowla, J. Number Theory, 7,
108–120, 1975.
[13] S Chowla, I N Herstein and W R Scott, The solutions of xd = 1 in
symmetric groups, Norske Vid. Selsk. Forth. (Trondheim) 25, 29–31,
1952.
[14] L Euler, De Formulis specei mxx + nyy ad numeros primos exploran-
dos idoneis earumque mirabilis proprietatibus, Opera Omnia I, v. 4,
Teubner, pp. 269–289, 1916.
[15] G Frei, Leonhard Euler’s convenient numbers, Math. Intell., 7, No. 3,
55–58, 64, 1985.
[16] S Chowla, An extension of Heilbronn’s class-number theorem, Quart.
J. Math., 5, 304–307, 1934.
[17] S Chowla, J Cowles and M Cowles, J.Number Theory, 12, 372–377,
1980.
[18] A Granville, Smooth numbers: Computational number theory and
beyond, Proceedings of an MSRI workshop, 2004.
[19] A Hildebrand and G Tenenbaum, Integers without large prime factors,
J. Theor. Nombres Bordeaux, 5, No. 2, 411–484, 1993.
[20] Ram Murty, Topics in Number Theory, Mehta Research Institute
Lecture Note No. 1, 1993.

125
Multi-variable Chinese
Remainder Theorem

1. Introduction
The Chinese remainder theorem (CRT) seems to have originated in the
3rd century AD in the work of Sun-Tsu. There are also versions in Indian
5th century mathematics of Aryabhata. The classical versions dealt with
coprime moduli. Oystein Ore proved a version [1] for non-coprime moduli
in 1952 in the American Mathematical Monthly but, this does not seem
to be well-known because a paper published 50 years later by Howard [2]
proves the same result! However, a multi-variable version does not seem
to be known. We present such a version and point out that there are still
many questions open for investigation.

2. Classical CRT
A variant of a folklore tale goes as follows. Three thieves steal a number
of gold coins and go to sleep after burying the loot. During the night, one
thief wakes up and digs up the coins and, after distributing into 6 equal
piles, finds 1 coin left over which he pockets quietly after burying the rest
of the coins. He goes back to sleep and after a while, a second thief wakes
up and digs up the coins. After making 5 equal piles, he again finds 1 coin
left over which he pockets and buries the rest and goes to sleep. The 3rd
thief wakes up and finds the rest of the coins make 7 equal piles excepting
a coin which he pockets. If the total number of coins they stole is not more
than 200, what is the exact number?
With a bit of hit and trial, one can find that 157 is a possible number.
The Chinese remainder theorem gives a systematic way of solving this in
general.
In the above problem, the sought-for natural number N is so that N − 1
is a multiple of 6, N − 2 is a multiple of 5 and N − 3 is a multiple of 7. This
means that N leaves remainders 1, 2, 3 on division by 6, 5, 7 respectively.
Let us consider two coprime natural numbers m1 , m2 and suppose, we are
looking for a natural number N which leaves remainders a1 , a2 on division
by m1 , m2 respectively. Then, the Euclidean division algorithm tells us that

The chapter is a modified version of an article that first appeared in Resonance, Vol. 20, No. 3,
pp. 206–216, March 2015.

127
 Multi-variable Chinese Remainder Theorem

the smallest positive integer of the form m1 k1 + m2 k2 (for integers k1 , k2 )

is the greatest common divisor (GCD) of m1 and m2 (which is 1 in this
case). Therefore, we have integers (of opposite signs) k1 , k2 such that

m1 k1 + m2 k2 = 1.

The number
N = a1 m2 k2 + a2 m1 k1
has the property that (N −a1 ) is a multiple of m1 and (N −a2 ) is a multiple
of m2 . We have not yet got a number as sought since N could be negative.
However, we may add a suitable multiple of m1 m2 to N and that will satisfy
the requirements.
More generally, suppose there are r natural numbers m1 , · · · , mr which
are pairwise coprime. We seek a natural number N leaving given remainders
a1 , · · · , ar on divisions by m1 , · · · , mr respectively. An appropriate gener-
alization of the above argument for two numbers is the following. If Mi
denotes the product of all the mj ’s excepting mi , then the GCD of mi
and Mi is 1 for each i = 1, · · · , r. As above, the Euclidean algorithm gives
integers ni , ki such that
ni Mi + ki mi = 1,
for each i = 1, · · · , r.
We have therefore, ai ni Mi − ai is a multiple of mi for each i ≤ r.
As mi divides Mj for each j 6= i, the integer

N = a1 n1 M1 + a2 n2 M2 + · · · + ar nr Mr ,

is such that (N − ai ) is a multiple of mi for each i ≤ r. Adding a suitable

multiple of m1 m2 · · · mr , we can get a natural number N such that N leaves
the remainder ai on division by mi for each i ≤ r.
Note that any other natural number N0 satisfying the same property
must differ from N by a multiple of each mi and hence, of the product
m1 m2 · · · mr .
In other words, there is a unique solution for N in the range
[1, m1 m2 · · · mr ].
Here is a nice exercise.
If m1 , · · · , mr are natural numbers which are not necessarily pairwise
coprime, then there is a natural number N yielding given remainders ai on
division by mi if, and only if, the GCD of mi and mj divides ai − aj for
each i, j.
2.1 Gauss and Congruences
The great mathematician C F Gauss defined the algebraic notion of
‘congruence’ which generalizes the notion of equality of numbers and

128
 Multi-variable Chinese Remainder Theorem

dramatically simplifies the proofs of several number-theoretic results.

Given a natural number m, one says two integers x and y are ‘congruent
modulo m’ – written x ≡ y mod m – if x − y is an integral multiple
of m.
Not surprisingly, the notation is also due to Gauss! Congruences to a
fixed modulus behave much like equality. For instance, it is very easy to
verify:
x1 ≡ y1 mod m , x2 ≡ y2 mod m
implies
x1 + x2 ≡ y1 + y2 mod m , x1 x2 ≡ y1 y2 mod m.
Further, we note that negative numbers are dealt with on equal footing;
a statement such as x leaves a remainder 3 on division by 4 can be written
as x ≡ −1 mod 4 as well. Note that if x leaves a remainder 3 on division
by 4, then x2013 leaves the same remainder and, this is easier to see via
congruences because

x ≡ −1 mod 4 ⇒ x2013 ≡ (−1)2013 = −1 ≡ 3 mod 4.

The argument that for two coprime integers m, n, 1 = mu + nv for

integers u, v from the Euclidean division algorithm, can also be rephrased as
asserting that ‘Each of the two integers has a multiplicative inverse modulo
the other’. That is, there exists an integer u (unique mod n – meaning
unique up to adding multiples of n) such that mu ≡ 1 mod n and, similarly,
there is an integer v (unique mod m) such that nv ≡ 1 mod m.
We call u ‘the multiplicative inverse of m mod n’ keeping in mind that
it is defined only up to addition of multiples of n.
The calculus of congruences is highly efficient in formulating and solving
problems in elementary number theory. For instance, given a number with
the digits d1 d2 · ·P
· dk in base 10, its remainder on division by 9 is simply
that of the sum ki=1 di . Note that the given number could be as large as
10r−1 while the sum of its digits is at the most 9r which is much smaller
for large r; this reduces computation drastically.
In the language of congruences, the classical Chinese remainder theorem
we proved above can be recast as follows:

(Explicit) Chinese Remainder Theorem: Let m1 , · · · , mr are pair-

wise coprime Qnatural numbers, and ai (1 ≤ i ≤ r) be arbitrary integers.
Write Mi = j6=i mj . Let ni be the multiplicative inverse of Mi modulo
mi . Then, the unique solution N mod m1 m2 · · · mr to the system of con-
gruences N ≡ ai mod mi for all i ≤ r is given by

N = a1 n1 M1 + a2 n2 M2 + · · · + ar nr Mr .

129
 Multi-variable Chinese Remainder Theorem

3. Many Variable CRT

Let us observe first that if c is an integer coprime to m and d is its
multiplicative inverse modulo m, then x = da is a solution to a congruence
cx ≡ a mod m. Therefore, the classical CRT can be formulated also as a
system of congruences of the form ci x ≡ ai mod mi (1 ≤P i ≤ r) where
(ci , mi ) = 1 for each i; the solution above changes to N = ri=1 di ai ni Mi
where ci di ≡ 1 mod mi . Thus, it is easy to formulate a multivariable version
where the left hand sides are linear polynomials in several variables xi ’s
instead of a single one. However, we soon realize that a necessary and
sufficient condition of the existence of a solution in general is far from
obvious. We formulate and prove the following multivariable version and
analyze later what else needs to be done.

Theorem. Let k, n be arbitrary positive integers and suppose aij are

integers (for 1 ≤ i ≤ k, 1 ≤ j ≤ n). Suppose m1 , · · · , mk are pairwise
coprime integers and b1 , · · · , br be arbitrary integers. Then, the k simulta-
neous congruences

a11 x1 + a12 x2 + · · · + a1n xn ≡ b1 (mod m1 ),

a21 x1 + a22 x2 + · · · + a2n xn ≡ b2 (mod m2 ),

···············
ak1 x1 + ak2 x2 + · · · + akn xn ≡ bk (mod mk )
have a solution in integers x1 , · · · , xn if and only if, for each i ≤ k, the
GCD of ai1 , ai2 , · · · , ain , mi divides bi .

Proof. We apply induction on k to prove the theorem. The proof is

constructive modulo the Euclidean division algorithm (which is also
constructive).
Consider first the case k = 1.
If the integers x1 , · · · , xn satisfy the congruence

a11 x1 + a12 x2 + · · · + a1n xn ≡ b1 (mod m1 ),

Pn
we have j=1 a1j xj − b1 = m1 t for some integer t. Thus, the greatest
common divisor of a11 , a12 , · · · , a1n and m1 divides b1 . This condition is
also sufficient by the Euclidean division algorithm. For, if b1 = sd where
d = GCD(a11 , · · · , a1n , m1 ), then writing
n
X
d= a1j yj + m1 t,
j=1

130
 Multi-variable Chinese Remainder Theorem

we have a solution x1 = sy1 , · · · , xn = syn of the congruence

a11 x1 + a12 x2 + · · · + a1n xn ≡ b1 (mod m1 ).
Therefore, for a general k, a necessary condition for a common solution
is that, for each i ≤ k, the GCD of ai1 , ai2 , · · · , ain , mi divides bi .
This condition also ensures that each individual congruence has a solu-
tion.
Now, we suppose that the GCD condition holds and that we have al-
ready arrived at a common solution x1 , · · · , xn in integers for the first r
congruences (1 ≤ r < k):
ai1 x1 + ai2 x2 + · · · + ain xn ≡ bi (mod mi ) ∀ 1 ≤ i ≤ r.
Now, we first choose a solution y1 , · · · , yn of the (r + 1)-th congruence
ar+1,1 x1 + ar+1,2 x2 + · · · + ar+1,n xn ≡ br+1 (mod mr+1 ).
For each j ≤ n, choose Xj such that
m1 m2 · · · mr Xj ≡ yj − xj (mod mr+1 ).
These choices are possible because m1 m2 · · · mr and mr+1 relatively prime.
We observe that for the new choices
x0j = xj + m1 m2 · · · mr Xj (1 ≤ j ≤ n),
the first r congruences continue to hold. Moreover,
n
X n
X
ar+1,j x0j ≡ ar+1,j (xj + m1 m2 · · · mr Xj )
j=1 j=1
n
X
≡ ar+1,j yj ≡ br+1 (mod mr+1 ).
j=1

Therefore, the theorem is proved by induction.

Remarks.
(i) The classical Chinese remainder theorem can be thought of as the
special case when the matrix {aij } has only a single column which is
non-zero.
(ii) If the matrix {aij } has a left inverse (that is an n × k integer matrix
{bij } such that BA = In , then clearly the necessary condition of the
theorem holds for any choice of b1 , · · · , bk .
In particular, if k = n and {aij } is in GL(n, Z), each system of n
linear congruences in n variables with pairwise co-prime moduli has a
solution.
(iii) A special case of the above theorem which is of interest as it produces
a solution for arbitrary bi ’s, is the following one. In the theorem above,

131
 Multi-variable Chinese Remainder Theorem

if, for each i ≤ k, there is some j for which aij is coprime to mi , then
the necessary condition obviously holds.
(iv) In the classical case of one variable, there is a unique solution modulo
m1 m2 · · · mk . In the multivariable case, there is no natural unique-
ness assertion possible. The point is that homogeneous congruences in
more than one variable have many solutions. So, uniqueness can be
asked for only after specifying a box (more precisely, an n-dimensional
parallelotope) in which we seek solutions.
For example, both (1, 4) and (0, −1) are simultaneous solutions of
the congruences
x − y ≡ 1 (mod 2),
x + y ≡ 2 (mod 3).
(v) The Euclidean division algorithm is the principal reason behind these
classical versions of the Chinese remainder theorem. In particular, it
holds good over the polynomial ring in one variable over a field. If n
elements are coprime, then there is a linear combination which gives 1.
This is no longer if we consider, for instance, polynomials in two vari-
ables.
For example, in the polynomial ring C[X, Y ], consider the congru-
ences
t ≡ 0 (mod X),
t ≡ 1 (mod Y ).
Here, of course, by a congruence f (X, Y ) ≡ g(X, Y ) mod h(X, Y ),
we mean that there is a polynomial k(X, Y ) so that

h(X, Y )k(X, Y ) = f (X, Y ) − g(X, Y ).

There is no common solution of the two congruences mentioned

above as there do not exist polynomials f, g for which Xf + Y g = 1!.
(vi) The Chinese remainder theorem has been generalized to rings and
modules. But, none of the versions is an analogue of the many variable
case proved above.

4. General Moduli for Multivariable CRT

Here, we point out a criterion which is sufficient to ensure the existence
of a solution when the moduli mi ’s are general (that is, not necessarily
pairwise coprime).
The case of general moduli is equivalent to a system of congruences for
prime power moduli. By the above theorem, we need to look at only the case
when all the moduli are powers of a single prime p. If we can get a necessary

132
 Multi-variable Chinese Remainder Theorem

and sufficient criterion for these cases, we will get such a criterion for the
general case. In this section, we fix a prime p and moduli mi = pti . We
further consider only the special case of n congruences in n variables. That
is, let us look at an n×n integer matrix A = {aij } and at the corresponding
system of congruences
n
X
aij xj ≡ bi (mod pti ) ∀ i ≤ n.
j=1

We write bi = pβi u0i

and det(A) = pδ d where u0i , d are not divisible by p.
A sufficient criterion is the following one:
Lemma. If δ ≤ βi ≤ ti for all i ≤ n, then the simultaneous system of
congruences above has a common solution.
Proof. The proof is straight forward.
Write the system of congruences as an equality Ax = b + mu for some
u1 , · · · , un , where we have written x, b, m as columns and where x, u need
to be shown to exist. Here m is the column (pt1 , pt2 , · · · , ptn )t .
If B = adj(A), the adjoint matrix of A, then multiplying the matrix
equation on the left by B = {bij }, we have det(A)x = B(b + mu) i.e.
n
X
pδ dxi = bij (pβj u0j + ptj uj ) ∀ i ≤ n.
j=1

By the hypothesis, the equality below has all entries to be integers:

n
X n
X
dxi = bij (pβj −δ u0j + ptj −δ uj ) = bij pβj −δ (u0j + ptj −βj uj ) ∀ i ≤ n.
j=1 j=1

As (p, d) = 1, we may choose uj ’s satisfying

ptj −βj uj ≡ −u0j (mod d) ∀ j.
Write u0j + ptj −βj uj = dyj for j ≤ n; then, we have
n
X
xi = bij pβj −δ yj ∀ i ≤ n.
j=1

Hence, we have a simultaneous solution to the congruences.

5. A Question for Investigation

Although sufficient conditions such as the one above can be formulated,
it is not clear (unlike the single variable case) how to formulate a general
necessary and sufficient condition for the multivariable Chinese remainder
theorem when the moduli are not necessarily pairwise coprime.

133
 Multi-variable Chinese Remainder Theorem

References
[1] Oystein Ore, The general Chinese remainder theorem, The American
Mathematical Monthly, Vol. 59, No. 6, 1952.
[2] Fredric T Howard, A Generalized Chinese Remainder Theorem,
The College Mathematics Journal, Vol. 33, No. 4, pp. 279–282,
September 2002.

134
Which Positive Integers are Interesting?

To Ramanujan, each number was a personal friend

in whose company, a lifetime he did spend.
Let us too begin this quest
to befriend numbers of interest.
A friend of a friend is a friend, may we not pretend?!
Much has been written about the numbers π and e (which are them-
selves related by the beautiful identity eiπ = −1). However, if we are to
think of only those interesting numbers which are positive integers, each
of us comes with her or his own list. Indeed, the question “which posi-
tive integers are interesting?” is not well-defined for, if n were the smallest
uninteresting positive integer, it is interesting for that reason! Be that as
it may, we identify some positive integers which have certain unique char-
acteristics. In some of these examples, this number is unique with those
characteristics and, in other cases, it is the smallest or the largest positive
integer encountered where a pattern changes, although there may be other
numbers with those characteristics. We also use our discussion as an excuse
to unveil interesting mathematics behind some of these phenomena. The
numbers are not necessarily arranged according to size.

30031
We are exposed to a beautiful thought process in school when we learn
Euclid’s proof of the infinitude of primes. Recall that the argument pro-
ceeds by observing that once we have gotten hold of the first few prime
numbers, the number obtained by adding 1 to their product must have
a prime factor which must necessarily be larger than the previous ones.
As this large number leaves remainder 1 on division by any of these primes,
any of its prime factors is larger than the previous primes (and hence gives
a new prime). The ‘hope’ (if one may call it that) that this new num-
ber is itself a prime, leads quickly to disillusionment. The first example is
2 × 3 × 5 × 7 × 11 × 13 + 1 = 30031. I leave it to the reader to find the
prime factors of this number. The intriguing question as to whether we do
get primes infinitely often in this process is still open! The largest known
prime P for which the product of all the primes until P is 1 less than a
prime number is 42209.

The chapter is a modified version of an article that first appeared in Resonance, Vol. 20, No. 8,
pp. 680–698, August 2015.

135
 Which Positive Integers are Interesting?

In the above proof of infinitude of primes, we used the sequence of num-

bers of the form 2 × 3 × · · · pn + 1. One could as well have used the sequence
2×3×· · · pn −1 instead. In that case, 2×3×5×7−1 = 11×19 is composite.
Once again, it is unknown whether there are infinitely many primes of this
form.
Let us pause for a moment to mull over an irony – among all numbers
of the form 2 × 3 × · · · pn + 1, it is certain that either we have infinitely
many primes or infinitely many composite numbers but, we do not know
the answer to either of these at present!

561
In cryptography, one of the recurring themes is the employment of the
so-called Fermat little theorem – If p is a prime number and a is an integer
which is not a multiple of p, the number ap−1 − 1 is a multiple of p.
The thought that this property might characterize all primes perishes
soon. There are composite positive integers n such that every positive inte-
ger a co-prime to n possesses the property that an−1 − 1 is a multiple of n.
Such numbers – now known as Carmichael numbers – have a characterizing
property. This is the property:
n is a Carmichael number if and only if it is square-free and each prime
divisor p of n satisfies p − 1 divides n − 1.
Look at ‘Prime ordeal’, Resonance, pp.866-881, September 2008, for a
proof.
The smallest Carmichael number is 561. Indeed, 561 = 3 × 11 × 17 and
560 = 16 × 5 × 7.
If a is relatively prime to 561, then a560 −1 has factors a2 −1, a10 −1, a16 −1
which are multiples of 3, 11, 17 respectively, by Fermat’s little theorem.

15
For an integer n > 1, look at all its divisors including 1 and n. Let s(n)
denote the sum of all digits of all the divisors.
For example, s(10) = 1 + 2 + 5 + (1 + 0) = 9.
Let us iterate this process, that is, look at s2 (n) = s(s(n)), s3 (n) =
s(s2 (n)) etc. In general, sk+1 (n) = s(sk (n)).
For instance
s2 (10) = s(9) = 1 + 3 + 9 = 13,
s3 (10) = s(13) = 1 + 1 + 3 = 5,
s4 (10) = s(5) = 1 + 5 = 6,
s5 (10) = s(6) = 1 + 2 + 3 + 6 = 12,

136
 Which Positive Integers are Interesting?

s6 (10) = s(12) = 1 + 2 + 3 + 4 + 6 + (1 + 2) = 19,

s7 (10) = s(19) = 1 + (1 + 9) = 11,
s8 (10) = s(11) = 1 + (1 + 1) = 3,
s9 (10) = 1 + 3 = 4,
s10 (10) = s(4) = 1 + 2 + 4 = 7,
s11 (10) = 1 + 7 = 8,
s12 (10) = 1 + 2 + 4 + 8 = 15,
s13 (10) = 1 + 3 + 5 + (1 + 5) = 15.

Therefore, after 12 iterations, 10 leads to 15; note that 15 is a fixed point

for the function s. The beautiful thing that happens is that every positive
integer n > 1 leads to 15 (so, 15 is like a black hole!).
√
The proof is very simple. The integer n has less than 2 n divisors (for
√ √
each divisor d < n, the divisor n/d is a divisor > n). Any positive
integer m has [log10 (m)] + 1 digits (if it has d digits, then 10d−1 ≤ m < 10d
which gives, on taking logs to the base 10 what is asserted). As each digit is
at most 9, the sum of the digits of m is at most 9(log10 (m)] + 1). Therefore,
n is a positive integer, for any divisor m of n, the sum of digits of m is at
most 9([log10 (n)] + 1) which gives
√
s(n) < 18 n([log10 (n)] + 1).

Using this, we get s(n) < n if n ≥ 104 .

For n < 104 , we can use a better upper bound for the number of divisors
of n to again deduce s(n) < n if n > 15. We leave it to the ingenuity
of the reader to complete by herself the argument for proving s(n) < n
when n > 15. In fact, it turns out that s(n) < n excepting the six values
16, 18, 24, 28, 36, 48. For these six values, we have s2 (n) < n excepting 18
for which s4 (18) < 18. Thus, by descending, it follows that one needs to
check only the numbers 2 to 15. This can be done by hand. In fact, for large
n, the argument shows that sk (n) = 15 where k is of the order of log log n.

15 again!
The number 15 has a claim to fame for another reason too. To motivate
it, we recall a few things. Lagrange proved that every positive integer is
expressible as a sum of four squares of integers. On the other hand, Gauss
proved that a natural number n is expressible as a sum of three squares of
integers if and only if it is NOT of the form 4k (8r +7). Indeed, Gauss was so
excited about this discovery which he noted in his mathematical diary as:
“EYPHKA! ∆ + ∆ + ∆ = n.”

137
 Which Positive Integers are Interesting?

It is said that this was the single discovery that turned Gauss’s mind
into taking up mathematics as a career although he was a great philologist
as well. Fermat stated the result that a positive integer > 1 is a sum of
two squares of integers if and only if, in its prime decomposition, every
prime of the form 4k + 3 appears with an even power. Ramanujan wrote
down a list of 55 such ‘positive forms’ ax2 + by 2 + cz 2 + dw2 for positive
integers a, b, c, d which he claimed were the only ones of this form which
take ANY positive integer value as the variables take integer values. His
list was almost perfect – the one exception x2 + 2y 2 + 5z 2 + 5w2 takes all
values excepting the value 15(!)
The mathematician and puzzlist John Conway came up with the follow-
ing general observation which he proved along with his student William
Alan Schneeberger. Consider
n
X
q(x1 , x2 , · · · , xn ) = aij xi xj
i,j=1

in n variables which takes only strictly positive values for all real values of
the variables other than xi = 0 for all i (one calls it positive-definite),
where all aij ’s are integers and aij = aji for i 6= j. They proved the
remarkable theorem that if this function takes all the integer values from
1 to 15 when we consider integer values for the variables xi , then it takes
ALL integer values! Conway–Schneeberger’s proof was very involved and
the mathematician Manjul Bhargava who received a Fields medal in 2014
(not for this work though) came up with a much simpler proof of this result
and, what is more, vastly generalized the result. Thus, 15 is special for the
reason that:
Pn
If i,j=1 aij xi xj is positive-definite, and aij are integers such that
aij = aji , and if all integer values from 1 to 15 occur as values of the
form when evaluated at suitable integers x1 , x2 , · · · , xn then ALL positive
integers occur as values. Moreover, 15 is the smallest such number.

1729
Any list of interesting positive integers is likely to include the taxicab
number 1729. The story of Ramanujan coming up with the observation
that 1729 is the smallest positive integer which is the sum of two perfect
cubes in two different ways
103 + 93 = 1729 = 123 + 13 ,
is too well-documented to repeat here. However, what may not be so well-
known is that 1729 is also a Carmichael number! Indeed, 1729 = 7 × 13 × 19
and 1728 = 26 × 33 . If a is coprime to 1729, then a1728 − 1 has factors

138
 Which Positive Integers are Interesting?

a6 − 1, a12 − 1, a18 − 1 which are multiples of 7, 13, 19 respectively. So, by

the criterion for Carmichael numbers mentioned during the discussion on
561 tells us that 1093 is a Carmichael number as well.

1806
This number has a very curious origin. It turns out to be the unique
solution to the following problem.
Q
Find all the even numbers n which satisfy n = p prime,(p−1)|n p.
Note that this means n includes ALL possible primes p for which p − 1
divides n. Thus, numbers like n = 2, 6 are ruled out.
Note that n must be square-free. One easily sees that 2, 3, 6, 7, 43 divide n.
Moreover, any such n must be divisible by these numbers (and perhaps
others). Because of the hypothesis that a prime p|n if, and only if, (p−1)|n,
if a new prime factor of n arises, it must be one more than a product of
smaller prime factors of n. However, the above numbers cannot give a new
prime because the numbers
2 × 43 + 1, 2 × 3 × 43 + 1, 2 × 7 × 43 + 1, 2 × 3 × 7 × 43 + 1
are all composite. Therefore, the unique answer to the problem is the num-
ber 2. × 3 × 7 × 43 = 1806.
The discovery/appearance of this number is due to Kellner and is in
the context of Bernoulli numbers – the numerator of the n-th Bernoulli
number is the above product. Thus, 1806 is the unique number n for which
the numerator of Bn equals n.

6174
This was a discovery by D Kaprekar in the 1940’s. Starting with any
4-digit number (other than those with identical digits), apply the following
transformation. Arrange the digits in the descending order, say a > b >
c > d. Subtract the number with the digits dcba from abcd to obtain a
4-digit number (even if it is a 3-digit number, it should be regarded as a
4-digit number with 0 in the beginning). This transformation produces after
finitely many iterations (at the most 7), the number 6174 which has come
to be known as the Kaprekar constant. Note that 6174 is invariant under
this transformation.
Before proving that every 4-digit number leads to 6174, we should first
look for such constants among 2-digit and 3-digit numbers.
It is immediately seen that any 2-digit number (other than those with
identical digits) leads to the cycle
09 → 81 → 63 → 27 → 45 → 09.

139
 Which Positive Integers are Interesting?

The unique 3-digit Kaprekar constant is 495. So, 495 has at least as much
claim to fame as 6174 (!)
Indeed, if abc is a 3-digit number with a ≥ b ≥ c and a > c, and if we
write
abc − cba = pqr,
then
10 + c − a = r, 10 − 1 + b − b = q, a − 1 − c = p.
Thus, q = 9 and p + r = 9. Hence, we need to check only the numbers

990, 891, 792, 695, 594

each of which is seen to lead to 495 which is fixed by the iteration.

For a 4-digit number abcd with a ≥ b ≥ c ≥ d and a > d, write the first
iteration as abcd − dcba = pqrs.
Then, 10 + d − a = s.
Now, if b = c, we have r = 9+c−b = 9, q = 9+b−c = 9 and a−1−d = p.
Thus, if b = c, we get q = r = 9 and p + s = 9 which leaves us to check
only the five numbers 9990, 8991, 7992, 6993, 5994. Each of these is easily
seen to lead to 6174 which is fixed by the Kaprekar iteration.
Finally, in case b > c, we have

10 + d − a = s, 10 − 1 + c − b = r, b − 1 − c = q, a − d = p.

These imply q + r = 8 and p + s = 10. This means that one needs to

check only the 25 numbers

p80s, p71s, p62s, p53s, p44s

for
(p, q) = (9, 1), (8, 2), (7, 3), (6, 4), (5, 5).
Each of these leads to 6174.
After this, we could go in two different directions – look at a general
number d of digits or/and a general base b in place of 10. We mention a
few results and leave it to the interested reader to investigate further.
For instance, if the base b = 2r, then the only 3-digit Kaprekar constant
in base b has the digits r − 1, 2r − 1 and r – the proof of this generalization
is the same as that of 495 in base 10.
It can be proved that there are no odd bases b admitting a 3-digit
Kaprekar constant.
As for 4-digit Kaprekar constants, there is one in base 5 which is
3032 – so, once again this has as much claim to be of interest as
6174 has!

140
 Which Positive Integers are Interesting?

The other bases where 4-digit Kaprekar constants exist are of the form
b = 4k × 10. In this base, the 4-digit Kaprekar constant has digits
6 × 4k , 2(4k − 1) + 1, 8(4k − 1) + 7 and 4 × 4k .
This can be proved similarly to the case of base 10.
There is no 5-digit Kaprekar constant in base 10.
On the other hand, base 15 has the Kaprekar constant with the 5 digits
10, 4, 14, 9, 5.
There exist 5-digit Kaprekar constants in each base of the form
b = 6k + 3 ≥ 15; this is left to the interested reader to determine.

3435
This is sometimes known as the Ramachandra number. An eminent
number theorist K Ramachandra observed when he was in college that
his Professor’s car number 3435 has the property
3435 = 33 + 44 + 33 + 55 .
This is the only number > 1 with this property. However, this remains
just a curiosity and does not seem to unveil any serious mathematics.

1848
We will see that 1848 is the largest of 65 numbers written down by
Euler with a certain property. It is known that there could be at the most
two larger numbers with that property. It is easy to show that if an odd
number has a unique expression as a sum of two squares of positive integers
n = x2 + y 2 and, if x, y are coprime, then n must be a prime number. Euler
generalized this property in order to obtain a primality criterion. He defined
a positive integer m to be ‘Idoneal’ or ‘convenient’ (‘Idoneus Numerus’ in
Latin) if it satisfies the property:
If an odd positive integer n admits a unique expression n = x2 + my 2
with x, y > 0 and if, in addition, the GCD (x, my) = 1, then n must be
prime.
Euler wrote down a list of 65 convenient numbers (the smallest ‘inconve-
nient’ number is 11) based on a criterion he obtained. The largest in his list
is 1848. Until date, no bigger convenient number has been found. S Chowla
was the first to prove in 1934 that there are only finitely many convenient
numbers. This is based on deep methods (coming under the umbrella of
class field theory) outside the scope of our discussion. Later, in 1973, it
has been shown by Weinberger that Euler could have missed at most two
other convenient numbers. Indeed, assuming the truth of a deep unsolved

141
 Which Positive Integers are Interesting?

problem known as the generalized Riemann hypothesis, it follows that there

could be at the most one number missing in Euler’s list.
Using the fact that 1848 is idonean, Euler observed that 18518809 =
1972 + 1848(100)2 is a prime.

8191
We know that every positive integer can be represented in binary form
(that is, in base 2) in terms of 0’s and 1’s. There is nothing sacrosanct
(mathematically) about base 2 and, one may represent numbers in any
base one wants to use. Notice that the number 31 has the base 2 expansion
(11111)2
and the base 5 expansion
(111)5 .
So, it is natural to ask which natural numbers have all their digits to be
equal to 1 with respect to two different bases > 1.
It was observed by Goormaghtigh nearly a century ago that 8191 has
this property;
(111)90 = (1111111111111)2 .
In usual decimal (base 10) notation, this number is 8191.
The question can be posed in another form as follows. If b1 6= b2 are
two positive integers > 1, then the number with m ones in base b1 is
2 m−1 bm
1 −1
1 + b1 + b1 + · · · + b1 = b1 −1 . Therefore, we are asking if this number
can consist of n one’s in another base b2 .
This is equivalent to solving
xm − 1 yn − 1
=
x−1 y−1
in natural numbers x, y > 1 for some m, n > 2.
The largest known solution is 8191 mentioned above. It is still unknown
whether there are only finitely many solutions in all variables x, y, m, n. In
fact,

1093
Fermat’s last ‘theorem’ – asserting that the equation xn +y n = z n has no
solutions in positive integers x, y, z when n > 2 – took 350 years to be justi-
fiably called a theorem. However, there were several subjective results from
the old times. One of them due to Wieferich showed that the first case of
Fermat’s last theorem holds good for a prime p for which 2p−1 − 1 is not a
multiple of p2 . That is, for such a prime p, the equation xp + y p = z p

142
 Which Positive Integers are Interesting?

has no solutions in positive integers x, y, z coprime to p. If there were

no such ‘Wieferich primes’, we would have a relatively elementary proof
of (the first case of) Fermat’s last theorem. However, there are Wieferich
primes and 1093 is the smallest. The next is 3511. To this day, no others
are known although on probabilistic grounds one expects asymptotically
log log(x) Wieferich primes until x as x → ∞.
The relation of congruence modulo a positive integer is a very convenient
way to express many divisibility statements. If m is a fixed positive integer,
one calls two integers a and b to be congruent modulo m, if a − b is a
multiple of m (meaning a − b = mc for some integer c). The notation a ≡ b
mod m is due to the great mathematician C-F Gauss who also discussed
the notion in the first place. Congruence relation generalizes equality and,
it is an easy exercise to check that it satisfies natural properties like:
a ≡ b mod m ; c ≡ d mod m implies
a + c ≡ b + d mod m , ac ≡ bd mod m.
Fermat’s little theorem can be re-stated as the assertion:
ap−1 ≡ 1 mod p if p is a prime a 6≡ 0 mod p.
Then, Wieferich’s congruences are 2p−1 ≡ 1 mod p2 for p = 1093, 3511.
If 2 is replaced by some other positive integers, there are other examples
when analogous congruences hold; viz.,
310 ≡ 1 mod 112
74 ≡ 1 mod 52
316 ≡ 1 mod 72
To see that 21092 − 1 is a multiple of 10932 , we proceed as follows.
Now 37 = 2187 = (2 × 1093) + 1 = 2p + 1, say.
Then 314 ≡ 4p + 1 mod p2 .
Also, 214 = 16384 = 15p − 11 which gives 228 ≡ −330p + 121 mod p2 .
So, 32 × 228 ≡ −1876p − 4 mod p2 .
On dividing by 4, we have
32 × 226 ≡ −469p − 1 mod p2 .
Raising to the 7-th power, we have:
314 × 226×7 ≡ −(1 + 469p)7 ≡ −(1 + 7 × 469p) mod p2
≡ −(1 + 3283p) ≡ −(1 + 4p) ≡ −314 mod p2 as observed above.
Hence 226×7 ≡ −1 mod p2 which gives 21092 = 226×7×6 ≡ (−1)6 ≡
1 mod p2 .
On the other hand, we show that a prime p which is either of the form
b + 1 or of the form 1 + b + b2 + · · · + bn for some b, cannot satisfy
N

bp−1 6≡ 1 mod p2 .

143
 Which Positive Integers are Interesting?

In particular, we have the observation:

Neither Mersenne primes (that is, primes of the form 1 + 2 + 22 + · · · +
2n−1 ), nor Fermat primes (that is, primes of the form 2n + 1) can be
Wieferich primes.
More generally, we prove:
Let p be a prime whose expression in a base b > 1 is of the form

1 + bk + b2k + · · · + bnk

for some n, k ≥ 1. Then,

p−1
bp−1 ≡ 1 + (bk − 1)p 6≡ 1 mod p2 .
(n + 1)k

Here is the proof.

(n+1)k
Now p = 1 + bk + · · · + bnk = b bk −1−1 .
Now, p and bk − 1 are relatively prime because p is a prime and

p ≥ bk + 1 > bk − 1.

Since p divides b(n+1)k − 1, the order of b mod p is a divisor of (n + 1)k.

If it were smaller, say mr, with m|(n + 1) and r|k, then either m < n + 1
or r < k.
If r < k, then the assertion b(n+1)r ≡ 1 mod p means p divides

(1 + br + · · · + bnr )(br − 1).

Now, p and br − 1 are relatively prime because p is a prime and p ≥

bk + 1 > br − 1.
Hence p = 1 + bk + · · · + bnk divides 1 + br + · · · + bnr , which is impossible
as p is the bigger number.
Now, if m < n + 1, then the condition bmk ≡ 1 means p divides 1 + bk +
(mk −1
· · · + b(m−1)k = b bk −1 as p and bk − 1 are relatively prime because p is a
prime and p ≥ bk + 1 > bk − 1.
This is impossible, as p = 1 + bk + · · · + bnk is larger than 1 + bk + · · · +
b (m−1)k .
We have shown that the order of b mod p is (n + 1)k; hence, this order
(n + 1)k divides p − 1.
p−1
Now, raise b(n+1)k = 1 + p(bk − 1) to the (n+1)k -th power. We have

p−1
bp−1 ≡ 1 + p(bk − 1) mod p2 .
(n + 1)k

144
 Which Positive Integers are Interesting?

Now, again the observation that p is relatively prime to bk − 1 implies

p−1
that p does not divide (bk − 1) (n+1)k .
This completes the proof.
In view of the observation above, an elementary proof of the first case of
Fermat’s last theorem exists (thanks to Wieferich’s criterion) for Mersenne
primes and Fermat primes.
Wieferich’s criterion can be proved with a bit of knowledge of the
Eisenstein reciprocity law which generalizes the so-called quadratic reci-
procity law of Gauss (again!). This is somewhat outside the scope of our
discussion. However, we can fortunately give an elementary result which is
in the spirit of (but weaker than) Wieferich’s criterion and gives a sufficient
criterion for Fermat’s last theorem to hold good.
Let p be an odd prime and let x, y, z be integers such that (p, xyz) = 1
and xp + y p ≡ z p mod p2 . Then, there exists a positive integer a ≤ (p − 1)/2
such that (a+1)p −ap −1 ≡ 0 mod p2 . In particular, if none of the (p−1)/2
congruences hold, the first case of Fermat’s last theorem holds.
Proof. By Fermat’s little theorem, z ≡ z p = xp + y p ≡ x + y mod p.
As (p, x) = 1, there is an integer x0 such that xx0 ≡ 1 mod p (viz., write
1 = pu + xx0 for some x0 ).
Note that since z ≡ x + y mod p, we have zx0 ≡ 1 + yx0 mod p.
Consider the integer a ≡ yx0 mod p with 1 ≤ a ≤ (p − 1)/2. Writing
a = yx0 + pt and applying the binomial expansion, we have
ap ≡ y p (x0 )p mod p2 .
Also, a + 1 ≡ yx0 + 1 mod p which gives, on raising to the p-th power
and applying binomial theorem as before, that
(a + 1)p ≡ (yx0 )p + 1 ≡ ap + 1 mod p2 .
This proves the assertion.
Note that if the (p − 1)/2 congruences in the statement above are
replaced by the single congruence corresponding to a = 1, we have Wieferich
criterion.

71
John Conway discovered an amazing fact. Start with any positive integer
other than 22. Let us start with 1 say. Define the sequence which just reads
out the number of times each chain of digits is repeated in turn. That is,
after 1, we have 11 (meaning one 1) and after that we have 21 (to mean two
1’s) and 1211 (to mean one 2, one 1) and 111221 (meaning one 1, one 2,
two 1’s) etc. In general, if ak11 ak22 · · · akr r with ai 6= ai+1 , then the next term

145
 Which Positive Integers are Interesting?

of the sequence is defined to be

k1 a1 k2 a2 · · · kr ar .
For example, the sequence starting from 1 is:
1, 11, 21, 1211, 111221, 312211, 13112221, 1113213211, 31131211131221, · · ·
If dn is the number of digits in the n-th term, then Conway discovered
the remarkable fact that the ratio dn+1 /dn approaches a constant λ (called
Conway’s constant) which is the unique real root of the polynomial
x71 −x69 −2x68 −x67 +2x66 +2x65 +x64 −x63 −x62 −x61 −x60 −x59 +2x58 +
5x57 + 3x56 − 2x55 − 10x54 − 3x53 − 2x52 + 6x51 + 6x50 + x49 + 9x48 − 3x47 +
7x46 −8x45 −8x44 + 10x43 + 6x42 + 8x41 −5x40 −12x39 + 7x38 −7x37 + 7x36 +
x35 − 3x34 + 10x33 + x32 − 6x31 − 2x30 − 10x29 − 3x28 + 2x27 + 9x26 − 3x25 +
14x24 −8x23 −7x22 −7x21 +9x20 −3x19 −4x18 −10x17 −7x16 +12x15 +7x14 +
2x13 − 12x12 − 4x11 − 2x10 − 5x9 + x7 − 7x6 + 7x5 − 4x4 + 12x3 − 6x2 + 3x − 6
of degree 71. If this is remarkable, it is even more remarkable that every
starting number (other than 22) leads to this same constant λ of degree 71.
The proof is very involved and comes under the umbrella of what is now
known as the cosmological theorem.
I give a very rough explanation of this phenomenon for the sake of the
more mathematically precocious reader among high school students. Some
experimentation will tell us that all the numbers in this sequence from the
8th one onwards arise from certain basic strings of elements – 92 of them.
In effect, these 92 ‘atoms’ can be written down explicitly and all elements of
the Conway sequence can be described in a sense through these 92 elements.
Thus, each element of the sequence is a word in the 92 basic elements and
the number of digits can be described recursively. This amounts to having
a 92 × 92 matrix which describes the recursion. A well-known technique on
recursion shows that the n-th term is expressible in terms of ‘eigenvalues’
of the matrix. These eigenvalues are solutions of a polynomial equation
which is obtained from the matrix. In the case above, the polynomial has
degree 71 and the ratios dn+1 /dn approach the only positive real root of
this polynomial – this is the λ mentioned above!

Skewes’s constants
The great mathematician C F Gauss conjectured at the age of 15, what
is now called the prime number theorem. He conjectured that the number
π(x) of primes not exceeding a number R xx is asymptotically given by the
logarithmic integral function li(x) = 2 logdt t . Here, by ‘asymptotically’,
we mean that the ratio π(x)/li(x) approaches 1 as x grows unboundedly
large. However, the inequality π(x) < li(x) was seen to hold for values of

146
 Which Positive Integers are Interesting?

x when these functions could be calculated. J E Littlewood proved in 1914

that the difference actually changes signs infinitely often. Hence, there is
indeed a smallest natural number n for which π(n) ≥ li(n). But, this was
an existential proof. Littlewood had a doctoral student named Skewes who,
one day in 1933, presumably said “(Ex)Skewes me! Assuming the Riemann
1034
Hypothesis, I can show that there is a number N no larger than 1010 such
that π(N ) ≥ li(N )”. Twenty years later, Skewes himself showed without
assuming the Riemann Hypothesis, that there is a number M no larger
10964
than 1010 such that π(M ) ≥ li(M ). These two numbers have come to
be known as Skewes’s constants. The latest developments have brought the
constants down to e728 although no explicit value of n is known for which
π(n) ≥ li(n). In the above, we have used the phrase Riemann Hypothesis
for an (perhaps the most important) open problem in mathematics.

Graham’s constant G
The number is so gigantic that additional notation is needed to write it
down. This number arose as follows.
Consider a hypercube in n dimensions. This is the generalization of a
square in 2 dimensions and a cube in 3 dimensions; it has 2n vertices. If we
join every vertex to every other one, we get what is known as a complete
graph. R L Graham and B L Rothschild considered the following problem.
If we colour each edge with one of two available colours, is it always true
that there must exist a complete subgraph containing four coplanar vertices
such that all its six edges are of the same colour?
This is not necessarily true for 3-dimensional cubes – we leave it to the
reader to construct an example. On the other hand, Graham and Rothschild
proved the existence of a complete, monochromatic subgraph containing
four coplanar vertices in any colouring, if the dimension n is large enough.
Until now, one does not know the minimal possible value of n with this
property but the proof of Graham and Rothschild showed the existence
of an n which is at the most a constant G known as Graham’s constant.
To define what G is, we introduce Knuth’s up-arrow notation.
For positive integers a, b we already know the usual exponentiation ab as
a shorthand notation for multiplying a to itself b times. Knuth introduces
the up-arrow notation as: a ↑ b for ab . Next, define

a ↑↑ b = a ↑ (a ↑ (a ↑ (· · · ))) .
| {z }
b times

4 (33 ) 27
For example, 4 ↑↑ 3 = 4(4 ) while 3 ↑↑ 4 = 3(3 ) = 3(3 ) a much larger
number. In fact, the former has about 154 digits whereas the latter has

147
 Which Positive Integers are Interesting?

more than 1012 digits.

Now, the next stage is easy to define.
For positive integers a, b define

a ↑↑↑ b = a ↑↑ (a ↑↑ (a ↑↑ (· · · ))) .
| {z }
b times

More generally,

a ↑n b = a ↑n−1 (a ↑n−1 (a ↑n−1 (· · · )))

| {z }
b times

where ↑k stands for ↑↑ · · · ↑.

| {z }
k times
In terms of these notations, Graham’s constant G = g64 where

g1 = 3 ↑4 3, g2 = 3 ↑g1 3, · · · , gn = 3 ↑gn−1 3.

We cannot even have a reasonable comprehension of how big this number

is but it has appeared in a mathematical proof; such is the power of the
human mind!

148
Counting, Recounting and Matching

Counting problems are teeny-weeny

and don’t require a magical genie
until we encounter a count
that seems impossible to surmount.
But then there is always Fubini!

1. Introduction
Although we start our mathematical education by learning to count, it
is this very thing which is one of the most difficult things to carry out in
practice. In this article, we discuss a very simple method which is available
in a very general form under the name of Fubini’s principle. In its simplest
form, it is the obvious observation that counting in two different ways
produces the same sum. However, the various examples we discuss will
hopefully convince the reader that the principle is surprisingly effective.

2. Hare and Tortoise

Let us start with a simple example.
Suppose, we wish to find the length of the longest possible sequence of
real numbers such that the sum of every 7 consecutive terms is positive and
the sum of every 11 consecutive terms is negative.
How do we analyze this problem?
Consider such a possible sequence of length at least 17 (if one exists),
say a1 , · · · , a17 , · · · .
The trick is to write the following array made up from these numbers,
viz.:
a1 a2 a3 · · · · · · a7
a2 a3 a4 · · · · · · a8
.. ..
. .
a11 a12 a13 · · · · · · a17
Each row sum is positive while each column sum is negative, which is
impossible as seen by considering the sum of ALL these numbers. Hence,
the maximal possible length is less than 17.

The chapter is a modified version of an article that first appeared in Resonance, Vol. 21, No. 4,
pp. 353–368, April 2016.

149
 Counting, Recounting and Matching

Let me leave as an exercise the task of finding a sequence of 16 numbers

with the required property (turn to the last page if you cannot work it out).
The principle we have been discussing is sometimes known as Fubini’s
principle. In its simplest form, it is the obvious observation that counting
in two different ways produces the same sum. To further convince ourselves
that such a simple principle can indeed be powerful, here are some more
examples:

3. Valuable Tiles
Tile the plane by unit squares with vertices at integer lattice points.
Inside each unit square, fill in some real number and call it the value of
that unit square. Let A be a finite collection of unit squares having the
property that the total value of the ‘translate’ A + (i, j) is positive for
each lattice point (i, j). Here, by translate A + (i, j), we mean the set
{(x + i, y + j) : (x, y) ∈ A}. Then, we claim that for each finite collection
B of unit squares, some translate of B must have positive value.
The solution depending on the Fubini principle goes as follows:
Denote by [i, j], the unit square whose lower left corner has co-ordinates
(i, j) and let v(i, j) denote its value. Write

A = {[i1 , j1 ], · · · , [ir , jr ]}

B = {[k1 , l1 ], · · · , [ks , ls ]}.

PNow, for each 1 ≤ m ≤ s, the total value of the translate A + (km , lm ) is

r
n=1 v(in + km , jn + lm ) which is known to be positive. Hence,
s X
X r
v(in + km , jn + lm ) > 0
m=1 n=1

which gives
r X
X s
( v(in + km , jn + lm )) > 0.
n=1 m=1

Therefore, we must have some n ≤ r for which sm=1 v(in +km , jn +lm ) >
P
0; hence, the translate (in , jn ) + B has total positive value.

4. Partitions
A well-known application of the Fubini principle is the following stunning
fact about partitions of a number. Recall that the number p(n) of partitions
of n is the number of ways of partitioning n objects into smaller collections.
For instrance p(4) = 5 because there are 5 ways to partition 4 objects:

150
 Counting, Recounting and Matching

all 4, 3 + 1, 2 + 2, 2 + 1 + 1, 1 + 1 + 1 + 1. This number of partitions of

n grows extremely fast as n grows. The fact alluded to asserts:
The number of partitions of n into m parts equals the number of parti-
tions of n in which the largest part is m.
The proof is by plotting an array of points corresponding to a partition
in the following manner:
For a partition n1 + n2 + · · · + nr = n where n1 ≤ n2 ≤ · · · nr , draw an
array consisting of dots with n1 dots in the first row, n2 dots in the second
row (centered to the left) etc.
Now, count column-wise; for instance, 8 = 2 + 2 + 4 gives the conjugate
array corresponding to the partition 8 = 3 + 3 + 1 + 1.

5. Regular Solids
Here is yet another striking example of the Fubini principle:
There are exactly five platonic solids.
In a platonic solid, the faces meeting at a vertex are regular polygons
with the same number of sides, the number of faces meeting at each vertex
is the same, and the solid angles at each vertex is the same.
If there are v vertices, e edges and f faces in the solid, and the faces
are regular polygons with p sides, then let us count the edges by counting
faces.
We get pf edges but we have counted each edge twice as it is the edge
of exactly two faces.
Hence, the Fubini principle implies 2e = pf .
Similarly, counting the edges by means of the two end point vertices, this
number also equals qv where q is the number of faces meeting at any vertex.
Hence,
qv = 2e = pf.
Rewrite it as
v e f
= = .
1/q 1/2 1/p
Now, we use the famous formula of Euler: v − e + f = 2; see, for instance
Example 9 on p.601 of the article ‘Invariants’ by B V Rajarama Bhat in
the July 2010 issue of Resonance. So,

v e f v−e+f
= = = .
1/q 1/2 1/p 1/q − 1/2 + 1/p
This is equal to
2 4pq 4pq
= = .
1/q − 1/2 + 1/p 2p − pq + 2q (4 − (p − 2)(q − 2)

151
 Counting, Recounting and Matching

4pq
We had shown qv = 2e = pf = (4−(p−2)(q−2) .
Therefore, (p − 2)(q − 2) < 4 which gives exactly five solutions

(3, 3), (3, 4), (3, 5), (4, 3), (5, 3).

These are the five Platonic solids – tetrahedron, octahedron, icosahedron,

cube and dodecahedron respectively.

6. Fubini for Bounds

Even when it is not possible to count precisely, the Fubini principle can
be useful to get lower bounds for the actual count. Here is one problem
which looks simple but proves to be deceptively difficult unless we think of
Fubini.
Consider n points on a line and look at the sequence of n2 possible

pairwise distances. Suppose each pairwise distance appears at the most

twice. Then, there are at least [n/2] numbers which appear exactly once as
pairwise distances.
Let us view the points from left to right on a horizontal line and denote
them by P1 , P2 , · · · , Pn respectively. Let a distances appear exactly once
and b distances appear exactly twice; then a + 2b = n2 .

The distances P1 Pi for 1 < i ≤ n are n − 1 distinct numbers.

The distances P2 Pj for 2 < j ≤ n are n − 2 distinct numbers,
The distance P1 Pi can equal P2 Pj at the most for one choice (i, j); for,
if P1 Pk = P2 Pl , then
P1 P2 = Pi Pj = Pk Pl ,
which leads to a contradiction.
Thus, we have at least n − 3 distances P2 Pj not occurring as P1 Pi .
In this manner, P3 Pr for 3 < r ≤ n gives at least (n − 3) − 2 = n − 5
distances not occurring as some P1 Pi or some P2 Pj .
2
2 n

bound a + b ≥ (n − 1) + (n − 3) + · · · = [n /4].
Thus, we have a lower
So, a ≥ [n /2] − 2 = [n/2].

7. Teachers with Designs

Here is a problem which seems like a scenario which could happen. We
discuss it and solve it using the Fubini principle. However, we point out
after the discussion that this scenario cannot occur! We will also give an
example of a similar situation which can occur in practice.
Here is the situation.
In a school, suppose there are a total of 50 teachers and s students.
Suppose, it happens that each teacher teaches a total of exactly 57 students

152
 Counting, Recounting and Matching

and each pair of students has exactly one common teacher. Then, can we
determine the total number of students?
Let us suppose t is the total number of teachers (in our case t = 50) and
that the total number of students is s. Suppose each teacher teaches a total
of exactly s0 students (we have s0 = 57) and that each pair of students has
exactly t0 common teachers (we have t0 = 1).
We will find the relations between t, t0 , s, s0 .
Let us look at any teacher T and a pair of students Si , Sj taught by
her. Count the number of triples (T, Si , Sj ) as T varies over teachers and
Si , Sj vary over pairs of students such that T teaches both
Si and Sj . As a
s0
teacher teaches exactly s0 students, there are exactly 2 triples (T, Si , Sj )
containing any particular teacher T . Therefore, the total number of triples
s0


is t 2 .
On the other hand, for each pair of students Si , Sj there are exactly t0
common teachers which means
there are t0 triples (T, Si , Sj ) containing the
s
pair Si , Sj . As there are 2 ways to select the pair of students Si , Sj , the
total number of triples is t0 2s .

By Fubini’s principle, we get

   
s0 s
t = t0 .
2 2

In our example, t = 50, t0 = 1, s0 = 57. So, we have 50×57×56

2 = s(s−1)
2 .
Solving this quadratic equation, we obtain s = 400.
Now, what is wrong with it? Nothing excepting the fact that this situation
cannot occur! Indeed, a necessary condition for the scenario to arise is given
by Fisher’s inequality which does not hold good here. Indeed, if there are
v students and b teachers such that each teacher k students and any pair
of students have λ common teachers, then turns out that there exists a
constant r such that each student is taught by exactly r teachers and r
satisfies
bk = rv , r(k − 1) = λ(v − 1).
In this standard notation, (b, v, r, k, λ) is called a block design. There is
an inequality due to population geneticist and statistician Ronald Fisher
(the doctoral supervisor of the well known Indian statistician C R Rao)
which asserts that b ≥ v; this is not satisfied in our situation.
Let us give an example of a similar scenario which can actually occur.
We recall in passing the famous quote by John von Neumann in 1947: ‘If
people do not believe that mathematics is simple, it is only because they do
not realize how complicated life is.’
Suppose there are a certain number b of teachers in a school who offer
to supervise a project each on different subjects. The projects are offered

153
 Counting, Recounting and Matching

to an exclusive group of the 9 most talented students each of whom can

work on more than one project. Each teacher supervises exactly 3 stu-
dents. Every pair of students works on exactly one common project. Is this
possible? If so, what is the total number of teachers, and what is the (com-
mon) number of teachers supervising each student (the number r predicted
above)?
The answer turns out to be yes, the situation can arise and the numbers
b and r can be found. Just as in the earlier problem, let us count the
triples:
(T, S1 , S2 ) where T varies over the b teachers and S1 , S2 are two students
supervised by T .
For each teacher, there are 32 choices for (S1 , S2 ) so that the total num-

ber of triples is t 32 . On the other hand, for any fixed pair (S1 , S2 ) among

the 9 students,
there is exactly one common supervisor so that the number
of triples is 92 . By Fubini’s principle,
   
3 9
t = .
2 2

Hence t = 12. Moreover, the constraints bk = rv , r(k − 1) = λ(v − 1)

along with the values b = 12, k = 3, v = 9 give r = 4. A situation when this
happens is represented by the following 9 × 12 (v × b) matrix consisting of
0’s and 1’s. In this matrix, each row represents a student, each column a
teacher, and the number of 1’s in each row is 4 (r = 4); there are exactly
three 1’s in each column (k = 3), and each pair of rows has 1’s in exactly
one column (λ = 1). How is such a matrix (equivalently (b, v, r, k, λ) block
design) constructed? This is tricky, and the essentially unique such matrix
turns out to be
1 0 0 1 0 0 1 0 0 1 0 0
 
 1 0 0 0 1 0 0 1 0 0 1 0 
 
 1 0 0 0 0 1 0 0 1 0 0 1 
 
 0 1 0 1 0 0 0 0 1 0 1 0 
 
 0 1 0 0 1 0 1 0 0 0 0 1 .
 
 0 1 0 0 0 1 0 1 0 1 0 0 
 
 0 0 1 1 0 0 0 1 0 0 0 1 
 
 0 0 1 0 1 0 0 0 1 1 0 0 
0 0 1 0 0 1 1 0 0 0 1 0

In general, if M is a v × b matrix consisting of 1’s and 0’s, then it corre-

sponds to a (b, v, r, k, λ) block design if, and only if, M M t = λJv + (r − λ)Iv
and uv M = kub , where uv , ub are vectors comprised of 1’s and Jv is the
v × v matrix all of whose entries are 1’s.

154
 Counting, Recounting and Matching

8. Sperner’s Theorem from Fubini

A striking application of the Fubini principle is a proof of the Sperner
theorem on antichains which asserts:
n


If |S| = n, then no antichain can contain more than [n/2] subsets in its
collection.
Here, a collection of its subsets is said to be antichain no subset in this
collections contains another in the collection.
Let us first look at an application.
Let a1 , · · · , an be n real
Pnumbers satisfying |ai | ≥ 1 for all i. Look at the
2n numbers of the form ni=1 i ai where each 
i is ±1. Then, any interval
n
of length less than 2 contains at the most [n/2] of these 2n numbers. Here
is a proof. We may assume without loss of generality that each ai ≥ 1.
For 1 , · · · , n ∈ {1, −1}, P
put E() =P{i ≤ n : i = 1}. If 01 , · · · , 0n ∈
{1, −1}, then let us look at i=1 i ai − ni=1 0i ai .
n

If E(0 ) ⊂ E(), then the above difference is simply

X
2 ai ≥ 2|E() \ E(0 )| ≥ 2.
i∈E()\E(0 )

Hence, only one of the sums can be in an interval length < 2.

As the sums inside an interval of length < 2 correspond to antichains,
Sperner’s theorem immediately implies our assertion that there are at the
n


most [n/2] sums as above in such an interval.
Here is a proof of Sperner’s theorem using the Fubini principle.
Let |S| = n; any maximal chain

∅=
6 S1 ⊂ S2 ⊂ · · · ⊂ Sn = S,

evidently has |Si | = i. Also, clearly there are n! maximal chains. If T is

any subset of S (with i elements say), then there are i!(n − i)! maximal
chains which contain T as a term. Let T1 , T2 , · · · , Tm be an antichain; let
C1 , C2 , · · · , Cn! be all the maximal chains in S.
Consider the set Σ of pairs (Ci , Tj ) where Tj occurs in Ci .
Write a matrix M whose (i, j)-th entry is 1 if Ci contains Tj as a term;
write the entry 0 if not.
Then, the sum of j-thP column is kj !(n−kj )! where Tj has kj elements; so,
the sum of all entries is m j=1 kj !(n − kj )!. On the other hand, two different
Tj ’s cannot occur in the same Ci (as T1 , T2 , · · · , Tm is an antichain) which
means each row sum is at the most 1. Thus, sum of all entries is ≤ n!; hence
m
X
kj !(n − kj )! ≤ n!
j=1

155
 Counting, Recounting and Matching

In other words, Pm 1 n ≤ 1.
j=1 (kj )

Since any nr ≤ [n/2]

n n




, we have m ≤ [n/2] .

9. Counting Leads to Identities

Counting is likely to involve expressions in terms of binomial coefficients;
so, there are situations when the Fubini principle will provide us with beau-
tiful identities involving binomial coefficients.
We start with some simple, well-known identities which arise in this
manner.
A class contains g girls and b boys, where we write b ≤ g to fix notation.
Suppose one wants to choose s students from the class for the school quiz
team.
Let us see how many ways they
can be chosen.
The first naive way says g+b s ways.
Now, let us look at a choice involving r boys and s − r girls; there are
b

g

r s−r choices.
As r can vary from 0 to b, we have the total br=0 rb s−r
P
g

.
By Fubini’s principle, we get
b     
X b g b+g
= .
r s−r s
r=0

We usually rewrite it in the symmetric form

    
X b g b+g
=
u v s
u,v≥0,u+v=s

n


where it is understood that a binomial coefficient d = 0 if n < d.
Note the particular case
X g 2 2g 
= .
u g
u≥0

Here is a more complicated example. Look at all paths from (0, 0) of

length 2n + 1 where each step is of unit length either to the east or to the
west or to the south or to the north, and which ends on the Y -axis.
We wish to count the number of such paths.
As we begin and end on the same vertical line, the total number

2r
of east
moves and west moves are equal; if this is r each, we have 2n+1
2r r choices.
As the rest of the 2n + 1 − 2r steps are vertical (up or down), there are
22n+1−2r choices.

156
 Counting, Recounting and Matching

PTherefore, the
total

2n+1−2r 2n+1 2r .

number of paths of length 2n + 1 is
r 2 2r r
On the other hand, if we only look at the set of paths on the X-axis,
which start

and end at (0, 0) and go by unit distance east or west, there
4n+2
are 2n+1 paths.
Calling the latter steps as ex , wx , and the earlier steps E, W, S, N , the
correspondence

E 7→ ex ex , W 7→ wx wx , N 7→ ex wx , S 7→ wx ex

is a bijection.
Hence, we obtain
  X   
4n + 2 2n+1−2r 2n + 1 2r
= 2 .
2n + 1 r
2r r

10. Counting Tiles of Different Shapes

Now, we discuss a very interesting example coming from tiling. Suppose
we have a board of dimension 1 foot ×n feet.
We have white and black ‘tiles’ which are unit squares in dimension; we
also have grey tiles (dominos) of dimension 1 × 2.
Let us tile the board (that is, fill the board with these with these three
types of tiles without overlapping).
For instance, if n = 1, there are two possible ways to tile – one using the
white square and the other using the black square.
If n = 2, then there are five ways to tile (black-black, white-white, white-
black, black-white, grey).
So, if tn denotes the number of ways to tile the 1 × n board, we have
t1 = 2, t2 = 5.
Let us alsoPput t0 = 1 for convenience. Here is an observation:
t2n+1 = 2 nr=0 t2r .
Indeed, note that every tiling of a 1 × (2n + 1) board must contain an
odd number of squares (hence, at least one square).
Look at the right-most square in a tiling.
Since all the tiles to its right are the 1 × 2 dominos, the right-most square
occurs at an odd-numbered place, say (2r + 1)-th place.
Then, the number of tilings with the right-most square at the (2r + 1)-
th place is 2t2r because the left-most 2r tiles can be tiled in t2r with the
(2r + 1)-th tile being white or black.
Hence, the assertion t2n+1 = 2 nr=0 t2r is proved.
P
We also have:
tn+1 = 2tn + tn−1 ; in particular, tn+1 and tn−1 are of the same parity.

157
 Counting, Recounting and Matching

Indeed, the first term 2tn counts the number of tilings of the 1 × (n + 1)-
board with the last tile a square, and tn−1 counts those for which the last
tile is a domino.
We make a divisibility observation
Pn now:
tn divides t2n+1 as well as r=0 t2r .
In fact, look at the two possibilities for a tiling – one in which the n-th
and (n + 1)-th places are filled by a domino or not.
If they are not occupied by a domino, the board is “breakable” at the
n-th place, and there are tn tn+1 such tilings.
If a domino occupies the n-th and (n + 1)-th squares, the number of
tilings is tn−1 tn .
Hence, we have
t2n+1 = tn tn+1 + tn−1 tn = tn (tn+1 + tn−1 )
which is evidently a multiple of tn .
Note also that since tn+1 , tn−1 have the same
P parity, tn+1 /2 is a multiple
of tn which means that tn divides t2n+1 /2 = r≥0 t2r .
On the other hand, in any tiling of the 1 × n board, consider the number
d of dominos.
They occupy 2d squares, and in the rest of the n − 2d squares, one can
have a black or white square which gives 2n−2d possibilities.
Now, the number of tiles here is n − d (because there are d dominos and
n − 2d unit squares).

number of ways to choose d dominos from these n − d tiles
So, the
n−d
is d .
Hence, we get:  
n−2d n − d
X
tn = 2 .
d
d≥0

In particular, P the divisibility properties above give divisibility properties

n−2d n−d


for the integers d≥0 2 d .
I leave it as interesting exercises to discover other such relations by count-
ing tiles rather than using algebra; for instance, prove the following:
X n + 1
tn−1 + tn = 2r .
2r
r≥0

4n
X
(t2n−1 + t2n )2 = tr .
r=0
I also leave it to the interested person to discover such identities when
we count tilings by squares of a different colours and dominos of b different
colours.

158
 Counting, Recounting and Matching

11. A Higher Congruence by Counting

Counting in two different ways, we can prove the following beautiful
congruence:
For a prime p > 3, and n ≥ r, pn n 3



pr ≡ r mod p .
Proof. Consider a n × p grid of squares from which we select pr squares.
We may either choose r entire rows; otherwise, there are at least two rows
from which between 1 and p − 1 squares are chosen. Cyclically shifting the
squares in each row divides the choices into equivalence classes out of which
n


r classes are singletons; the other classes are all of cardinalities multiples
of p2 . Thus, we have, first of all,
   
pn n
≡ mod p2 .
pr r

We refine this argument now.

If a choice of pr squares has less than r −2 entire rows, the corresponding
equivalence class has cardinality a multiple of p3 . Therefore,
the asserted3
congruence mod p3 reduces to showing the special case 2p p ≡ 2 mod p
when p ≥ 5. To see this, note
  X  2 p−1
2p p X
= ≡2+p2
k −2 mod p3 .
p k
k=1
Pp−1 2
The latter sum is clearly ≡ k=1 k ≡ 0 mod p when p > 3.

12. Macaulay Expansion by Counting

We finish with a beautiful number-theoretic statement for with a Fubini-
type proof [1]. For each natural number r, denote by Sr , the set of all r-digit
numbers in some base b whose digits are in strictly decreasing order of size.

Evidently, Sr is non-empty if and only if b ≥ r; in this case, Sr has rb
elements.
Let us now write the elements of Sr in increasing order. For instance, in
base 10, the first few of the 120 members of S3 are:

(2, 1, 0), (3, 1, 0), (3, 2, 0), (3, 2, 1), (4, 1, 0), (4, 2, 0), (4, 2, 1), (4, 3, 0), · · · .

Then, we have:
Given any positive integer n, and any base b such
that rb
> n, the (n+1)-

th member of Sr is (ar , · · · , a2 , a1 ) where n = arr + ar−1 + · · · + a11 . In

r−1

particular, for each n, the Diophantine equation arr + ar−1 +· · ·+ a11 = n

r−1

has a unique solution in positive integers ar > ar−1 > · · · > a1 ≥ 0.

159
 Counting, Recounting and Matching

We leave the proof as an
exercise.

The expression n = arr + ar−1 + · · · + a11 is known as Macaulay’s
r−1

expansion and can simply be proved by the greedy algorithm but the above
statement gives a combinatorial interpretation.
Here are a couple of examples to illustrate the statement.
(i) Let r = 3 and n = 12.
We may take any base b so that 3b > 12. For example, b = 6 is allowed

because 63 = 20.

Among the 20 members in S3 , the 13-th member is (5, 2, 1).

Note that      
5 2 1
+ + = 12.
3 2 1
(ii) Let r = 3, n = 74.
We may take b = 10 as 10


3 = 120. The 75-th member of S3 is (8, 6, 3).
Note that      
8 6 3
+ + = 74.
3 2 2

Remark. Informally, Fubini’s theorem gives conditions under which a

function f of two variables satisfies the property that the integral of f
under the product measure equals
Z Z Z Z
( f (x, y)dy)dx = ( f (x, y)dx)dy.

An example to show this subtlety is the function

x2 − y 2
f (x, y) = ∀ 0 < x, y < 1.
(x2 + y 2 )2
Then,
Z Z Z Z
( f (x, y)dy)dx = π/4 6= −π/4 = ( f (x, y)dx)dy.

Answer to the Puzzle

A sequence of 16 numbers with each subsequence of 7 consecutive terms
adding to a positive number and every subsequence of 11 consecutive terms
adding to a negative number is:
−5, −5, 13, −5, −5, −5, 13, −5, −5, 13, −5, −5, −5, 13, −5, −5.

References
[1] B Sury, Macaulay expansion, Amer. Math. Monthly, Vol. 121, 2014.

160
Odd if it isn’t an Even Fit!
Lighting up Tiling

There was a chap who could tile a square

whom I was perfectly willing to hire.
“Used triangles – all of areas same,
and needed but eleven for this game”,
he said, and I knew he was a liar!

Tiling Squares by Triangles of Given Area

Try to cut a square into finitely many triangles (possibly of different
shapes) of equal area. You would find that – no matter what the shapes
are – the number of triangles is always even. Here is an example.

There is some interesting history behind the discovery of the above fact.
In 1965, Fred Richman from the university of New Mexico had decided to
pose this in an examination in the master’s programme. He had observed
this in some cases but when he tried to prove it in general prior to posing
it in the exam, he was unsuccessful. So, the problem was not posed in the
exam. His colleague and bridge partner John Thomas tried for a long time
and finally came up with a proof that it is impossible to break the unit
square with corners at (0, 0), (1, 0), (0, 1)(1, 1) cannot be broken into an
odd number of triangles when the vertices of all the triangles have rational
co-ordinates with odd denominators. He sent the paper to the Mathematics
Magazine where the referee thought the result may be fairly easy (but could

The chapter is a modified version of an article that first appeared in Resonance, Vol. 20, No. 1,
pp. 23–33, January 2015.

161
 Lighting up Tiling

not find a proof himself) and perhaps known (but could not find a refer-
ence to it). On the referee’s suggestion, Richman and Thomas posed this
as a problem [1] in the American Mathematical Monthly which nobody
could solve. Subsequently, Thomas’s paper appeared in the Mathematics
Magazine [2] in 1968. Finally, in 1970, Paul Monsky proved the complete
version in a paper [3] in the American Mathematical Monthly removing the
restriction imposed in Thomas’s paper.
We shall discuss Monsky’s proof of this beautiful fact presently.
Amazingly, it uses some nontrivial mathematical objects called 2-adic
valuations.
In fact, one may consider generalizations of squares like cubes and hyper-
cubes of higher dimensions. If an n-dimensional cube is cut into simplices
(generalizations of triangles like tetrahedra etc. in higher dimensions) of
equal volumes, it turns out that the number of simplices must be a multi-
ple of n!
One also says that a region like the interior of a square is ‘tiled’ by
triangles if the square can be broken into triangular pieces.
There is also a generalization of the result on tiling by triangles of
squares to the (so-called) polyominos. Polyominos are just unions of unit
squares.
Connected subsets of the square lattice tiling of the plane are called
special polyominos. That is, they have standard edges – edges of the unit
squares are parallel to the co-ordinate axes. The generalization alluded to
is due to S K Stein [4] and asserts:
Consider a special polyomino which is the union of an odd number of
unit squares. If this polyomino is a union of triangles of equal areas, then
the number of triangles is even.
We discuss the proof of this statement after discussing the solution by
Paul Monsky of the first problem we started with. It can be noticed that
the proof uses crucially that the number of unit squares in the polyomino
is odd. Interestingly, this question is still unanswered when the number of
unit squares in the polyomino is even.
The proof of Monsky as well as Stein’s result above use the so-called
‘2-adic valuation function’. The 2-adic valuation is a function from the set
of non-zero rational numbers to the set of integers; it simply counts the
power of 2 dividing any integer or, more generally, any rational number.
What is meant by the power of 2 dividing a rational number? Writing any
non-zero rational number as p/q with p, q having no common factors, one
may look at the power of 2 dividing p or q. If p, q are both odd, this is
simply 0. If p is even, then the 2-adic valuation of p/q is defined to be the
power of 2 dividing p. If q is even, then the 2-adic valuation of p/q is defined
to be the negative of the power of 2 dividing q, Formally, we write:

162
 Lighting up Tiling

φ : Q∗ → Z
a
defined by φ( 2c b ) = a where b, c are odd; define also φ(0) = ∞ so that φ is
defined for all rational numbers. We keep in mind that 0 has larger 2-adic
value than any other rational number.
Colour a point (x, y) ∈ Q × Q by the colour:

red, if φ(x), φ(y) > 0,

blue, if φ(x) ≤ 0; φ(x) ≤ φ(y),
green, if φ(x) > φ(y) and φ(y) ≤ 0.

In this manner, all points of Q × Q are coloured by these three

colours.
For example:
(2, 0) is red, (1, 3) is blue and (1, 1/2) is green. Also, (0, 0) is red while
(1, 0) is blue and (0, 1) is green.
Now, we proceed to assert something which is easy to believe but not
that easy to prove. The assertion is that it is possible to extend the above
function to a function on the whole of real numbers (but the values can
be non-integers). In our further discussion, we assume without further ado,
the existence of an extension

φ:R→R

which satisfies:
φ restricts to the 2-adic valuation on Q;
φ(xy) = φ(x) + φ(y);
φ(x + y) ≥ min(φ(x), φ(y)).
This is important to have such an extension because we would really like
to colour all points in a square.
For instance, as φ(3/4) = φ(2−2 3) = −2, the second property above
implies that √
φ(3/4) = 2φ( 3/2) = −2.
√
Hence φ( 3/2) = −1.
Let us start Monsky’s proof by first considering the unit square with a
left lower corner at (0, 0) and making a few easy observations:
(i) If a point a is red, then any point x and x + a have the same colour.
(ii) On any line, there are at the most two colours.
(iii) The boundary of the square has an odd number of segments which
have a red end and a blue end.
(iv) If a triangle is not ‘complete’ (that is, has vertices only of one or two
colours), then it has 0 or 2 red-blue edges.

163
 Lighting up Tiling

Proof. Recall

φ(xy) = φ(x) + φ(y)

φ(x + y) ≥ min(φ(x), φ(y))

In particular, if φ(x) > φ(y), then φ(x + y) = φ(y).

In particular, if (x, y) is blue, then φ(x) ≤ φ(y) and so, φ(y/x) ≥ 0 and,
if (x, y) is green, then φ(x) > φ(y) and so, φ(y/x) < 0.
Let us prove (i) now.
As a is red, its co-ordinates have positive φ and it is easy to check in
each of the three cases of colouring for a point x that x and x + a have the
same colour.
For (ii), without loss of generality, we may assume that the line passes
through the origin. But two other points (xi , yi ); i = 1, 2 on the line y = tx
have colours blue and green respectively, say.
But then φ(y1 /x1 ) = φ(y2 /x2 ) = φ(t) is impossible as the former is ≥ 0
while the latter is < 0.
To prove (iii), note that (ii) implies that such segments on the bound-
ary must be on the segment from (0, 0) to (1, 0) which are red and blue
respectively. But, this is clear.
The proof of (iv) is completely clear by considering the various possibil-
ities RRB,RBB,RGG,RRG,BBG,BGG.
Now, we are ready to prove:
Let a square be tiled by n triangles of equal areas. Then, n is even.
Proof (Monsky). Counting the red-blue edges on the square, we are count-
ing the interior edges twice and the boundary edges once.
Thus, (iii) above would be contradicted unless there is a complete
triangle. But then a complete triangle has area A with φ(A) < 0 – let
us check this now.
Firstly, note that the triangle can be moved so that the vertices are at
(0, 0), (a, b) and (c, d) where (a, b) is blue and (c, d) is green. Thus, the area
is (ad − bc)/2.
As (a, b) is blue, φ(a) ≤ φ(b) and as (c, d) is green, φ(c) > φ(d).
Therefore, φ(ad) = φ(a) + φ(d) < φ(b) + φ(c) = φ(bc) which gives
φ(ad − bc) = φ(ad) = φ(a) + φ(d) ≤ 0.
Hence φ(A) = φ((ad − bc)/2) ≤ −1.
So, if there are n triangles, then φ(A) = φ(1/n) < 0; that is, n is
even.
This completes Monsky’s wonderful proof.
Let us now prove the more general version on polyominos mentioned
above:

164
 Lighting up Tiling

Consider a polyomino which is the union of an odd number of unit

squares. If it is tiled by triangles of equal areas, then the number of tri-
angles is even.
Proof. It can be seen that if a line segment made up of segments parallel
to the axes has a blue end and a green end, then each of the individual
segments has ends only coloured blue or green and, an odd number of them
have both colours as ends.
The key observation is:
If a polyomino made up of standard squares as above is made up of n
triangles of equal areas and, if an odd number of standard edges on its
boundary have ends coloured blue and green, then φ(2A) ≤ φ(n), where A
is the area of the polyomino.
The proof of this in turn depends on the following fact:
Let (xi , yi ); i = 0, 1, 2 be the vertices of a triangle T where (xi , yi ) ∈ Si
with
S0 = {(x, y) : φ(x), φ(y) > 0},
S1 = {(x, y) : φ(x) ≤ 0, φ(y)},
S2 = {(x, y) : φ(y) < φ(x), φ(y) ≤ 0}.
Then, φ(area(T ) ≤ −φ(2).
Proof. As translation by (−x1 , −y1 ) does not change areas, and Pi − P0 ∈
Si for any Pi ∈ Si , we may assume that (x0 , y0 ) = (0, 0).
Then, area(T ) = 21 |x1 y2 = x2 y1 |.
Now φ(x1 ) ≤ 0, φ(y1 ). Also φ(y2 ) ≤ 0 and φ(y2 ) < φ(x2 ).
Thus, φ(x1 y2 ) < φ(x2 y1 ) and φ(x1 y2 ) ≤ 0.
Hence φ( area of T ) = φ(1/2) + φ(x1 y2 ) ≤ φ(1/2) = −φ(2).
Next, we observe:
If a polyomino made up of standard squares as above is made up of n
triangles of equal areas and, if an odd number of standard edges on its
boundary have ends coloured blue and green, then φ(2A) ≤ φ(n), where A
is the area of the polyomino:
Look at a triangle of the dissection which has all three colours and let B
denotes its area.
Note that points in S0 , S1 , S2 have different colours.
Now, nB = A and φ(B) ≤ −φ(2); that is, φ(A) − φ(n) ≤ −φ(2).
Therefore,
φ(n) ≥ φ(2A).
We now proceed to show that a special polyomino which is the union of
an odd number of unit squares and is a union of triangles of equal areas,
then the number of triangles is even.

165
 Lighting up Tiling

We note that a standard (unit) edge with a blue end and a green
end must be parallel to the X-axis and lies on a line whose height
is odd.
Therefore, on the border of each standard square, there is an edge with
a blue end and a green end.
Edges in the interior of the polyomino are adjacent to two standard
squares whereas those on the boundary are adjacent to one standard square
of the polyomino.
As there is an odd number of standard squares, the above observation
applies and, implies that φ(2A) ≤ φ(n). But, φ(2A) ≥ 1; so n must be a
multiple of 2.

2. Tiling Rectangles by Rectangles

Let us discuss tiling integer rectangles with integer rectangles now. Can
we tile a rectangle of size 28 × 17 by rectangles of size 4 × 7?
At least, the area of the smaller rectangle divides that of the larger one
(a necessary requirement for tiling). But, in fact, we don’t have a tiling.
Why?
Look at each row of the big rectangle. If we have managed to tile as
required, then 17 would be a positive linear combination of 4 and 7. This
is impossible.
Thus, two necessary conditions for tiling an m × n rectangle with a × b
rectangles are:
(i) ab divides mn and,
(ii) each of m, n should be expressible as positive linear combinations of
a, b.
Are these conditions sufficient for tiling?
Look at a 10 × 15 rectangle which we wish to tile with copies of a 1 × 6
rectangle.
The two necessary conditions mentioned clearly hold true in this case.
However, a tiling is obviously impossible. Let us see why.
In fact, more generally, we claim that for copies of an a × b rectangle to
tile an m × n rectangle, a third condition that is also necessary is that a
must divide either m or n and b also must divide m or n.
To demonstrate this, look at a possible tiling.
We may suppose a > 1 (if a = b = 1, there is nothing to prove).
We colour the unit squares of the m × n rectangle with the different a-th
roots of unity 1, ζ, ζ 2 , · · · , ζ a−1 as follows.
Think of the rectangle as an m × n matrix of unit squares and colour the
(i, j)-th unit square by ζ i+j−2 . Here is a suggestive figure:

166
 Lighting up Tiling

a a+1

1 ζ ζ 2 · · · · · · 1/ζ 1 ζ ···
ζ ζ2 ζ3 · · ·
ζ2 ζ3 · · ·
..
.

Since each tile (copy of the smaller rectangle used) contains all the a-th
roots of unity exactly once, and as the sum 1 + ζ + · · · + · · · ζ a−1 = 0, the
sum of all theP entries
Pnof the m × n rectangle must be 0.
Therefore, m i=1 j=1 ζ i+j−2 = 0.

But this sum is the same as ( m i−1 )( n ζ j−1 ) = 0 which means

P P
i=1 ζ j=1
one of P
these two sums must be 0.
But m i=1 ζ
i−1 = 0 if, and only if, ζ m − 1 = 0; that is, a|m. Similarly, the

other sum is 0 if, and only if, a|n.

Thus, this condition that a divides m or n is necessary and, by the same
reasoning it is necessary for tiling that b divides m or n.
Looking at the above proof, it is also easy to see how to tile when these
conditions hold good. That is, we have the necessary and sufficient criterion:
Proposition. An m × n rectangle can be tiled with copies of a × b rect-
angles if, and only if,
(i) ab divides mn,
(ii) m and n are expressible as non-negative linear combinations of a
and b,
(iii) a divides m or n and b divides m or n.
This generalizes in an obvious way to any dimension and we leave it to
the reader to investigate this.
We discuss now the following result for which several proofs are available.
If a rectangle is tiled by rectangles each of which has at least one of its
sides integral, then the big rectangle must also have a side of integral length.
We place the co-ordinate system such that all the sides of the rectangles
have sides parallel to the co-ordinate axes.
Consider the function f (x, y) = e2iπ(x+y) for (x, y) ∈ R2 .

167
 Lighting up Tiling

For a rectangle defined by [a, b] × [c, d], we have

Z Z Z b Z d
2iπx
f (x, y) = e dx e2iπy
a c
e2iπb − 2iπa
 2iπd
− e2iπc
 
e e
= .
2iπ 2iπ

Thus, the integral of f over a rectangle is zero if and only if it has at

least one integer side is zero; hence, in case of a tiling by such rectangles,
the integral is zero which means that the big rectangle has an integer side.
One of the beautiful results proved by Max Dehn using methods from
topology (outside our scope here) is:
A rectangle of size l×b is tileable by squares if and only if l/b is a rational
number.
He proved more generally:
Let R be a rectangle which has at least one side of rational length. If R
is tiled by rectangles each of which has rational ratio of length to breadth.
Then, all the sides of all the rectangles (including R) are of rational lengths.
Interestingly, this result of Dehn was re-proved by Brooks by associat-
ing an electrical network consisting of currents, voltages and resistances
with the tiling and using well known properties of such networks. The dis-
cussions in this article would convince the reader that the subject draws
from several areas of mathematics. However, we have mostly included only
proofs which involve some simple algebra and basic number theory. Topo-
logical arguments require a more detailed discussion. In the next part of
the article, we hope to discuss such aspects.

References
[1] F Richman and J Thomas, Problem 5471, The American Mathematical
Monthly, 74, p. 329, 1967.
[2] J Thomas, A dissection problem, The Mathematics Magazine, 41,
pp. 187–190, 1968.
[3] P Monsky, On dividing a square into triangles, The American Mathe-
matical Monthly, 77, pp. 161–164, 1970.
[4] S K Stein and S Szabo, Algebra & Tiling, The Carus Mathematical
Monographs 25, Published by the Mathematical Association of America,
1994.

168
Polya’s One Theorem with 100 pages
of Applications

Believe me, Mathematics is an asset

in counting (it can aid and abet)
of different isomers even stereo.
We’d have bid it cheerio
but for this magic PET!
In 1937, George Polya wrote a paper which is considered one of the most
significant papers in 20th-century mathematics [1]. The article contained
one theorem and 100 pages of applications. It introduced a combinatorial
method and led to hitherto unexpected applications to diverse problems
in science. Very interestingly, we mention in passing that it was noticed
and pointed out to the mathematical community only as late as 1960 by
Frank Harary that Polya’s work had already been anticipated in 1927 by
J H Redfield [2].
Polya’s theory of enumeration was discussed in detail in an earlier issue of
Resonance by Shriya Anand [3], a summer student of the author. In what
follows, we recall briefly some of this theory and complement the earlier
article by adding some other applications not discussed there.
Let us start with an example. Consider the problem of painting the faces
of a cube either black or white. How many such distinct coloured cubes are
there? Since the cube has 6 faces, and we have 2 colours to choose from, the
total number of possible coloured cubes is 26 . But, painting the top face
white and all the other faces black produces the same pattern as painting
the bottom face white and all the other faces black as we can simply invert
the cube and it looks the same! The answer to the above question is not so
obvious. To find the various possible colour patterns which are inequivalent,
we shall exploit the fact that the rotational symmetries of the cube have
the structure of a group.
Before explaining how the above problem is dealt with by Polya’s theory
in more precise terms, we mention that the scope of Polya’s theory is extra-
ordinarily wide because of its very simple and very general expression. This
theory deals with enumeration of mathematical configurations which can
be thought of as placement of shapes in receptacles. More abstractly, we
have mappings from a set D of receptacles to a set R of shapes. Thus, in a
configuration, two elements of D may have the same image in R; that is, the

The chapter is a modified version of an article that first appeared in Resonance, Vol. 19, No. 4,
pp. 338–346, April 2014.

169
 Polya’s One Theorem with 100 pages of Applications

same shape can be placed in more than one receptacle. Each shape is given
a value and the value of a configuration is the total value of all the shapes.
A typical problem would be to determine the number of configurations with
a given value. Those permutations of the receptacles which yield another
configuration which is equivalent to the original one, give rise to a group
and the theory is stated in terms of this group. For instance, in the cube-
colouring problem, let D be the set of 6 faces of the cube and R is the set
of two colours black and white. A configuration here is a colouring of the
faces by the two colours; that is, a mapping φ : D → R.
The group of rotations of the 6 faces has 24 elements (see the figure):
(i) the identity element;
(ii) rotating clockwise by 90 degree about the line through the centers of
opposite faces like A and D (there are three such rotations);
(iii) rotating by 180 degrees about the line through the centers of opposite
faces like A and D (there are three such);
(iv) rotating anti-clockwise by 90 degrees about the line through the cen-
ters of opposite faces like A and D (there are three such);
(v) rotating by 180 degrees about lines through the centers of diagonally
opposite edges like the edge of the faces A and B and the edge of the
faces D and E (there are six such);
(vi) rotating clockwise by 120 degrees about the line connecting diagonally
opposite vertices like the vertex where A,B,C meet and the vertex
where D,E,F meet (there are four such);
(vii) rotating anti-clockwise by 120 degrees about the line connecting
diagonally opposite vertices like the vertex where A,B,C meet and
the vertex where D,E,F meet (there are four of these).

A B

Returning to the cube-colouring problem, let X be the set of all

colourings. With respect to the group G of permutations of D, we can
define an equivalence of elements in X as follows:

170
 Polya’s One Theorem with 100 pages of Applications

φ1 ∼ φ2 if, and only if, there exists some g ∈ G such that φ1 g = φ2 .

As G is a group, ∼ is an equivalence relation on X. So, it partitions X
into disjoint equivalence classes. It is clear that the orbits of the action
i.e., the equivalence classes under ∼ are precisely the different colour pat-
terns. Therefore, we need to find the number of orbits of the action of G
on X. Polya’s theorem has as its starting point a lemma known popularly
as Burnside’s lemma although it was already known due to Cauchy and
Frobenius. That lemma says thatP for a group G of transformations on a
1 g g
set X, the number of orbits is |G| g∈G |X | where X is the set of points
of X fixed by g. In the problem of colouring cubes with two colours, this
lemma suffices to find the number of configurations. We will describe this
now. Finer information like how many configurations have 2 white faces
and 4 black faces need the force of Polya’s theorem.
In the cube-colouring problem with two colours, X has 26 elements. Let
us see how the above mentioned 24 transformations affect X. The identity,
of course, does not change anything; that is, fixes all the 26 elements of X.
The rotation mentioned in (ii) above fixes all those colourings where the
faces B,C,E,F have the same colour and A,D could be arbitrarily coloured.
The description for transformations in (iv) is similar. For the transformation
in (iii), the colours of B and E should match and so should the colours of
C and F. The transformation described in (v) fixes those colourings where
A and B have the same colour, D and E have the same colour and C and F
have the same colour. Under the transformations of type (vi) and (vii), a
colouring which is fixed must have three ‘top’ faces of P the same colour and
the ‘bottom’ three of the same colour. Thus, the sum g∈G |X g | equals

26 + 6 · 23 + 8 · 22 + 3 · 22 · 22 + 6 · 22 · 2 = 240.
Therefore, Cauchy–Frobenius–Burnside lemma gives the number of orbits
to be 240/24 = 10.
To describe Polya’s theorems, we shall consider only finite sets D, R like
in the example of the cube. For a group G of permutations on a set of n
elements and variables s1 , s2 , . . . , sn , one defines a polynomial expression
(called the cycle index) for each g ∈ G. If g ∈ G, let λi (g) denote the number
of i-cycles in the disjoint cycle decomposition of g. Then, the cycle index of
G, denoted by z(G; s1 , s2 , . . . , sn ) is defined as the polynomial expression
1 X λ1 (g) λ2 (g)
z(G; s1 , s2 , . . . , sn ) = s1 s2 . . . snλn (g) .
|G|
g∈G

For instance,
X sλ1 1 sλ2 2 . . . sλk k
z(Sn ; s1 , s2 , . . . , sn ) =
1λ1 λ1 !2λ2 λ2 ! . . . k λk λk !
λ1 +2λ2 +...+kλk =n

171
 Polya’s One Theorem with 100 pages of Applications

We may view the above-mentioned configurations obtained by placing a

set R of shapes in D receptacles also as a colouring problem – viz., the
shapes can be thought of as colours and the receptacles as various objects
to be coloured. Polya’s theorem asserts:
Suppose D is a set of m objects to be coloured using a range R of k
colours. Let G be the group of symmetries of D. Then, the number of colour
1
patterns = |G| z(G; k, k, . . . , k).
The cycle index of the group G of symmetries of the 6 faces of the cube
turns out to be
1
z(G; s1 , · · · , s6 ) = (6s2 s4 + 3s21 s22 + 8s23 + 6s32 + s61 ).
24 1
So, in our example of the cube, the number of distinct coloured cubes
 
1
= 26 + 6 · 23 + 8 · 22 + 3 · 22 · 22 + 6 · 22 · 2 = 10.
24

There are 10 distinct coloured cubes in all, using two colours, as we saw
above.
Incidentally, if we look at a regular octahedron, then the group of sym-
metries of the 6 vertices is the same group above of symmetries of the 6
faces of the cube.

A ‘Valuable’ Version of Polya’s Theorem

The above version of Polya’s theorem gives us the total number of config-
urations but we can retrieve finer information from other versions. We look
at one example before proceeding to some other applications of Polya’s
theorem. As mentioned earlier, one could assign a value for each shape/
colour in R and enumerate the number of configurations with a given value.
It is convenient to give values of shapes to be non-negative integers. One
forms the generating function

c(x) = c0 + c1 x + c2 x2 + c3 x3 + · · ·

which is a polynomial where ck is the number of shapes which have the

value k. The finer version of Polya’s theorem we are alluding to asserts
that if ak is the number of configurations whose total value is k, then the
generating function

a0 + a1 x + a2 x2 + a3 x3 + · · ·

is obtained by substituting c(xr ) for sr in the cycle index.

172
 Polya’s One Theorem with 100 pages of Applications

For simplicity, suppose R has two elements Black and white which have
values 0 and 1. Then, the generating polynomial above is simply 1 + x.
Let us discuss the example of chlorination of Benzene where some hydro-
gen atoms get substituted by Chlorine atoms. Give the values 0 and 1 to Cl
and H, and note that the group of symmetries of the Benzene molecule is
the group of rotations of the regular hexagon, which is the so-called dihedral
group D6 of order 12. The cycle index of D6 is
 
1 6 3 2 2 2
z(D6 ) = s + 4s2 + 2s3 + 3s1 s2 + 2s6 .
12 1
Substituting 1 + xr for sr ’s, we obtain the polynomial
 
1 6 2 3 3 2 2 2 2 6
(1 + x) + 4(1 + x ) + 2(1 + x ) + 3(1 + x) (1 + x ) + 2(1 + x )
12
= 1 + x + 3x2 + 3x3 + 3x4 + x5 + x6 .
Therefore, the number of configurations which have 2 chlorine atoms
is the coefficient of 4 which is 3. These are the ortho dichlorobenzenes,
meta dichlorobenzenes and para dichlorobenzenes where the gap between
the vertices corresponding to the carbon atoms to which the two chlorine
atoms are attached are 1, 2 and 3 edges respectively.
More general weighted versions of Polya’s theorem as well as the immedi-
ate applications to enumerating isomers of chemical compounds have been
discussed in detail in the earlier article in 2002.

Graph Enumeration
A key application of Polya’s theorem is to the enumeration of graphs.
Indeed, the introduction of Polya’s paper begins with the words (as trans-
lated by Read):
“This paper presents a continuation of work done by Cayley. Cayley has
repeatedly investigated combinatorial problems regarding the determination
of the number of certain trees. Some of his problems lend themselves to
chemical interpretation: the number of trees in question is equal to the num-
ber of certain theoretically possible chemical compounds.”
Indeed, a chemical compound with no multiple bonds corresponds to a
tree where different types of vertices correspond to different atoms. In case
of multiple bonds, one may regard different kinds of edges also.
A tree consists of vertices and edges and is a connected graph where each
edge connects two vertices. There can be several edges meeting at a vertex.
There is no closed path. Therefore, the number of edges is one less than the
number of vertices. A vertex is called r-edged if there are exactly r edges
originating there.

173
 Polya’s One Theorem with 100 pages of Applications

Consider an alkane – this has a formula Cn H2n+2 . The carbon atoms are
usually assumed to have valency 4 which means that the structure of the
alkane is determined (that is the positions of the hydrogen atoms are uniquely
determined) by the structure formed by the carbon atoms. Topologically dif-
ferent trees with n four-edged vertices and 2n + 2 one-edged vertices corre-
spond to the different isomers with the molecular formula Cn H2n+2 .
Thus, the enumeration of isomers as above is equivalent to the enumer-
ation of trees as above. Interestingly, in Polya’s paper, he describes the
groups of symmetries for certain chemical compounds as so-called wreath
products. Polya calls wreath products as coronas. As far as one can ascer-
tain, this is the first introduction and study of finite wreath products as
permutation groups.
Polya’s theorem was generalized by de Bruijn in a way which allows one
to permute the shapes in R also.

Musings on Music
Polya’s theorem has been applied to the theory of music. One may deter-
mine the number of chords. To define this, one takes the n-scale to be the
integers from 0 to n − 1 under addition modulo n. There are translations
a 7→ a + i, where 0 ≤ i < n. An equivalence class (that is an orbit) is
called a chord and, one wishes to determine for each r < n, the number
of r-chords; that is, the number of orbits consisting of r elements. This is
equivalent to colouring the n-notes by two colours – we choose the notes
in the chord by colouring them by one colour and those which are not in it
by the other colour. The group is simply the cyclic group of order n whose
cycle index is:
1X n/d
z(Cn ; s1 , · · · , sn ) = φ(d)sd .
n
d|n

In this, we substitute 1 + xdfor sd and obtain the generating function

r
whose coefficient of x is the number of r-chords.
We obtain the number of r-chords to be
 
1 X n/d
φ(d) .
n r/d
d|(n,r)

Sometimes, one allows for a bigger group of transformations of the scale

by allowing inversion a 7→ −a also. Then, the group becomes the dihedral
group Dn of order 2n formed by the translations and the transposition
above. In this case, as we have the cycle index of Dn to be
1 X n d 1 n
−1 n
z(Dn ; s1 , · · · , sn ) = φ( )sn/d + (s21 s22 + s22 )
2n d 4
d|n

174
 Polya’s One Theorem with 100 pages of Applications

if n is even and
1 X n d 1 n−1
z(Dn ; s1 , · · · , sn ) = φ( )sn/d + s1 s2 2
2n d 2
d|n

if n is odd, we may determine the number of r-chords in this dihedral case

to be:    
1 X n/d 1 [n/2]
φ(d) +
2n r/d 2 [r/2]
d|(n,r)

if n is odd;    
1 X n/d 1 n/2
φ(d) +
2n r/d 2 r/2
d|(n,r)

if n, r are even and;

1 n2 − 1
   
1 X n/d
φ(d) +
2n r/d 2 [r/2]
d|(n,r)

if n is even and r is odd.

For instance, in 12-tone music with dihedral symmetry, the number of
pentachords is computed to be 38.
For more details, the interested reader may see the paper [4] by Reiner.
The paper of Polya was in German and was translated by RC Read and
this appears in a book now (see[5]). Also, several interesting generalizations
and applications have been discussed by Krishnamurthy in the book [6]
available in an Indian edition.

References
[1] G Polya, Kombinatorische Anzahlbestimmungen fur Gruppen, Graphen
und chemische Verbindungen, Acta Mathematica, Vol. 68, pp. 145–254,
1937.
[2] J H Redfield, The theory of group reduced distributions, Amer. J.Math.,
49, pp. 433–455, 1927.
[3] Shriya Anand, How to count – an exposition of Polya’s theory of enu-
meration, Resonance, pp. 19–35, September 2002.
[4] David L Reiner, Enumeration in music theory, The American Mathe-
matical Monthly, pp. 51–54, January 1985.
[5] G Polya and R C Read, Combinatorial enumeration of groups, graphs,
and chemical compounds, Springer-Verlag, 1987.
[6] V Krishnamurthy, Combinatorics – Theory and Applications, Affiliated
East-West Press Private Limited, 1985.

175