Lectures on

Analytic Number Theory

By
H. Rademacher

Tata Institute of Fundamental Research,

Bombay
1954-55
 Lectures on
Analytic Number Theory

By
H. Rademacher

Notes by

K. Balagangadharan
and
V. Venugopal Rao

Tata Institute of Fundamental Research, Bombay

1954-1955
Contents

I Formal Power Series 1

1 Lecture 2

2 Lecture 11

3 Lecture 17

4 Lecture 23

5 Lecture 30

6 Lecture 39

7 Lecture 46

8 Lecture 55

II Analysis 59
9 Lecture 60

10 Lecture 67

11 Lecture 74

12 Lecture 82

13 Lecture 89

iii
CONTENTS iv

14 Lecture 95

15 Lecture 100

III Analytic theory of partitions 108

16 Lecture 109

17 Lecture 118

18 Lecture 124

19 Lecture 129

20 Lecture 136

21 Lecture 143

22 Lecture 150

23 Lecture 155

24 Lecture 160

25 Lecture 165

26 Lecture 169

27 Lecture 174

28 Lecture 179

29 Lecture 183

30 Lecture 188

31 Lecture 194

32 Lecture 200
CONTENTS v

IV Representation by squares 207

33 Lecture 208

34 Lecture 214

35 Lecture 219

36 Lecture 225

37 Lecture 232

38 Lecture 237

39 Lecture 242

40 Lecture 246

41 Lecture 251

42 Lecture 256

43 Lecture 261

44 Lecture 264

45 Lecture 266

46 Lecture 272
 Part I

Formal Power Series

1
Lecture 1

Introduction
In additive number theory we make reference to facts about addition in 1
contradistinction to multiplicative number theory, the foundations of which
were laid by Euclid at about 300 B.C. Whereas one of the principal concerns
of the latter theory is the deconposition of numbers into prime factors, addi-
tive number theory deals with the decomposition of numbers into summands.
It asks such questions as: in how many ways can a given natural number be
ecpressed as the sum of other natural numbers? Of course the decompostion
into primary summands is trivial; it is therefore of interest to restrict in some
way the nature of the summands (such as odd numbers or even numbers or per-
fect squares) or the number of summands allowed. These are questions typical
of those which will arise in this course. We shall have occasion to study the
properties of V-functions and their numerous applications to number theory,
in particular the theory of quadratic residues.
Formal Power Series
Additive number theory starts with Euler (1742). His tool was power series.
His starting point was the simple relation xm . xn = xm+n by which multiplica-
tion of powers of x is pictured in the addition of exponents. He therefore found
it expedient to use power series. Compare the situation in multiplicative num-
ber theory; to deal with the product n.m, one uses the equation n s m s = (nm) s ,
thus paving the way for utilising Dirichlet series.
While dealing with power series in modern mathematics one asks ques- 2
tions about the domain of convergence. Euler was intelligent enough not to ask
this question. In the context of additive number theory power series are purely
formal; thus the series 0! + 1! x + 2! x2 + · · · is a perfectly good series in our

2
1. Lecture 3

theory. We have to introduce the algebra of formal power series in order to

vindicate what Euler did with great tact and insight.
A formal power series is an expression a0 + a1 x + a2 x2 + · · · . Where the
symbol x is an indeterminate symbol i.e., it is never assigned a numerical value.
Consequently, all questions of convergence are irrelevant.
Formal power series are manipulated in the same way as ordinary power
series. We build an algebra with these by defining addition and multiplication
in the following way. If
∞
X ∞
X
A= a n xn , B= b n xn ,
n=0 n=0
∞
P ∞
P
we define A + B = C where C = cn xn and AB = D where D = d n xn ,
n=0 n=0
with the stipulation that we perform these operations in such a way that these
equations are true modulo xN , whatever be N. (This reauirement stems from
the fact that we can assign a valuation in the set of power series by defining
∞
P
the order of A = an xn to be k where ak is the first non-zero coefficient).
n=0
Therefore cn and dn may be computed as for finite polynomials; then
cn = a n + b n ,
dn = a0 bn + a1 bn−1 + · · · + an−1 b1 + an b0 .
A = B means that the two series are equal term by term, A = 0 means that
all the coefficiants of A are zero. It is easy to verify that the following relations 3
hold:
A+B= B+a AB = BA
A + (B + C) = (A + B) + C A(BC) = (AB)C
A(B + C) = AB + AC
We summarise these facts by saying that the formal power series form a
commutative ring. This will be the case when the coefficients are taken from
such a ring, eg. the integres, real numbers, complex numbers.
The ring of power series has the additional property that there are no divi-
sors of zero (in case the ring of coefficients is itself an integrity domain), ie. if
A, B = 0, either A = 0 or B = 0. We see this as follows: Suppose A = 0, B = 0.
Let ak be the first non-zero coefficient in A, and b j the first non-zero coefficient
∞
P
in B. Let AB = dn xn ; then
n=0
   
dk+ j = a◦ bk+ j + · · · + ak−1 b j+1 + ak b j + ak+1 b j−1 + · · · + ak+ j b0 .
1. Lecture 4

In this expression the middle term is not zero while all the other terms are
zero. Therefore dk+ j , 0 and so A.B , 0, which is a contradiction.
From this property follows the cancellation law:
If A , ◦ and A.B = A.C, then B = C. For, AB − AC = A(B − C). Since
A , 0, B − C = ◦ or B = C.
If the ring of coefficients has a unit element so has the ring of power series.
As an example of multiplication of formal power series, let, 4

A=1−x and B = 1 + x + x2 + · · ·
∞
X
A= a n xn , where a0 = 1, a1 = −1, and an = 0 for n ≥ 2,
n=0
X∞
B= b n xn , where bn = 1, n = 0, 1, 2, 3, . . .
n=0
X∞
C= cn xn , where cn = a0 bn + a1 bn−1 + · · · + an b0 ;
n=0

then

c0 = a0 b0 = 1, cn = bn − bn−1 = 1 − 1 = 0, n = 1, 2, 3, . . . ;
so (1 − x)(1 + x + x2 + · · · ) = 1.

We can very well give a meaning to infinite sums and products in certain
cases. Thus

A1 + A2 + · · · = B,
C1C2 · · · = D,

both equations understood in the sense module xN , so that only a finite number
of A′ s or (C − 1)′ s can contribute as far as xN .
Let us apply our methods to prove the identity:

1 + x + x2 + x3 + · · · = (1 + x)(1 + x2 )(1 + x4 )(1 + x8 ) · · ·

Let

C = (1 + x)(1 + x2 )(1 + x4 ) . . .
(1 − x)C = (1 − x)(1 + x)(1 + x2 )(1 + x4 ) . . .
= (1 − x2 )(1 + x2 )(1 + x4 ) . . .
= (1 − x4 )(1 + x4 ) . . .
1. Lecture 5

Continuing in this way, all powers of x on the right eventually disappear,

and we have (1−x)C = 1. However we have shown that (1−x)(1+x+x2 +· · · ) =
1, therefore (1 − x)C = (1 − x)(1 + x + x2 + · · · ), and by the law of cancellation,
C = 1 + x + x2 + · · · which we were to prove.
This identity easily lends itself to an interpretation which gives an example 5
of the application of Euler’s idea. Once again we stress the simple fact that
xn · xm = xn+m . We have

1 + x + x2 + x3 + · · · = (1 + x)(1 + x2 )(1 + x4 )(1 + x8 ) · · ·

This is an equality between two formal power series (one represented as a

product). The coefficients must then be identical. The coefficient of xn on the
right hand side is the number of ways in which n can be written as the sum
of powers of 2. But the coefficient of xn on the left side is 1. We therefore
conclude: every natural number can be expressed in one and only one way as
the sum of powers of 2.
We have proved that

1 + x + x2 + x3 + · · · = (1 + x)(1 + x2 )(1 + x4 ) · · ·

If we replace x by x3 and repeat the whole story, modulo x3N , the coeffi-
cients of these formal power series will still be equal:

1 + x3 + x6 + x9 + · · · = (1 + x3 )(1 + x2.3 )(1 + x4.3 ) · · ·

Similarly

1 + x5 + x2.5 + x3.5 + · · · = (1 + x5 )(1 + x2.5 )(1 + x4.5 ) · · ·

We continue indefinitely, replacing x by odd powers of x. It is permissible

to multiply these infinitely many equations together, because any given power
of x comes from only a finite number of factors. On the left appears
Y
(1 + xk + x2k + x3k + · · · ).
k odd

On the right side will occur factors of the form (1 + xN ). But N can be
written uniquely as xλ .m where m is odd. That means for each N, 1 + xN will
occur once and only once on the right side. We would like to rearrange the
factors to obtain (1 + x)(1 + x2 )(1 + x3 ) · · ·
This may be done for the following reason. For any N, that part of the 6
formal power series up to xN is a polymial derived from a finite number of
1. Lecture 6

factors. Rearranging the factors will not change the polynomial. But since this
is true for any N, the entire series will be unchanged by the rearrangement of
factors. We have thus proved the identity
Y ∞
Y
(1 + xk + x2k + x3k + · · · ) = (1 + xn ) (1)
k odd n=1

∞
P
This is an equality of two formal power series and could be written a n xn
n=0
∞
P
= bn xn . Let us find what an and bn are. On the left we have
n=0

(1 + x1.1 + x2.1 + x3.1 + · · · )(1 + x1.3 + x2.3 + x3.3 + · · · )

× (1 + x1.5 + x2.5 + x3.5 + · · · ) · · ·

xn will be obtainded as many times as n can be expressed as the sum of odd

numbers, allowing repetitions. On the right side of (1), we have (1 + x)(1 +
x2 )(1 + x3 ) · · · xn will be obtained as many times as n can be expressed as the
sum of integers, no two of which are equal.
an and bn are the number of ways in which n can be expressed respectively
in the two manners just stated. But an = bn . Therefore we have proved the
following theorem of Euler:
Theorem 1. The number of representations of an integer n as the sum of dif-
ferent parts is the same as the number of representations of n as the sum of odd
parts, repetitions permitted.
We give now a different proof of the identity (1).
∞
Y ∞
Y ∞
Y ∞
Y
(1 + xn ) (1 − xn ) = (1 − xn )(1 + xn ) = (1 − x2n ).
n=1 n=1 n=1 n=1

Again this interchange of the order of the factors is permissible. For, up to 7

any given power of x, the formal series is a polynomial which does not depend
on the order of the factors.
∞
Y ∞
Y ∞
Y
(1 + xn ) (1 − xn ) = (1 − x2n ),
n=1 n=1 n=1
∞
Y ∞
Y ∞
Y ∞
Y
(1 + xn ) (1 − x2n−1 ) (1 − x2n ) = (1 − x2n ).
n=1 n=1 n=1 n=1
1. Lecture 7

∞
Q
Now (1 − x2n ) , 0, and by the law of cancellation, we may cancel it from
n=1
both sides of the equation obtaining,
∞
Y ∞
Y
(1 + xn ) (1 − x2n−1 ) = 1.
n=1 n=1

Multiplying both sides by

∞ 
Y 
1 + x2n−1 + x2(2n−1) + x3(2n−1) + · · ·
n=1
∞
Y ∞ 
Y ∞ 
Y 
(1 + xn ) 1 + x2n−1 1 + x2n−1 + x2(2n−1) + · · ·
n=1 n=1 n=1
∞ 
Y 
= 1 + x2n−1 + x2(2n−1) + · · · .
n=1

For the same reason as before, we may rearrange the order of the factors on
the left.
∞
Y ∞ 
Y  
(1 + xn ) 1 + x2n−1 1 + x2n−1 + x2(2n−1) + · · ·
n=1 n=1
∞ 
Y 
= 1 + x2n−1 + x2(2n−1) + · · · .
n=1

However,
∞ 
Y  
1 + x2n−1 1 + x2n−1 + x2(2n−1) + · · · = 1,
n=1

because we have shown that (1 − x)(1 + x + x2 + · · · ) = 1, and this remains true

when x is replaced by x2n−1 . Therefore the above equation reduces to
∞
Y ∞
Y   Y 
(1 + xn ) 5 1 + x2n−1 + x2(2n−1) + · · · = 1 + xn + x2n + x3n + · · ·
n=1 n=1 n odd

which is the identity (1).

Theorem 1 is easily verified for 10 as follows:10, 1+9, 2+8, 3+7, 4+6, 8
1+2+7, 1+3+6, 1+4+5, 2+3+5, 1+2+3+4 are the unrestricted partitions. Par-
titions into odd summands with repetitions are
1. Lecture 8

1+9, 3+7, 5+5, 1+1+1+7, 1+1+3+5, 1+3+3+3, 1+1+1+1+1+5,

1+1+1+1+3+3, 1+1+1+1+1+1+1+3, 1+1+1+1+1+1+1+1+1+1.
We have ten partitions in each category.
It will be useful to extend the theory of formal power series to allow us to
find the reciprocal of the series a0 + a1 x + a2 x2 + · · · where we assume that
a0 , 0. (The coefficients are now assumed to form a field). If the series
1
b 0 + b 1 x + b 2 x2 + · · · = ,
a 0 + a 1 x + a 2 x2 + · · ·
we would have (a0 + a1 x + a2 x2 + · · · )(b0 + b1 x + b2 x2 + · · · ) = 1. This means
that a0 b0 = 1 and since a0 , 0, b0 = 1/a0 . All other coefficients on the left
vanish:
a0 b1 + a1 b0 = 0,
a0 b2 + a1 b1 + a2 b0 = 0
..............................
We may now find b1 from the first of these equations since all the a′ s and
b0 are known. Then b2 can be found from the next equation, since b1 will
then be known. Continuing in this, manner all the b′ s can be computed by
successively solving linear equations since the new unknown of any equation
is always accompanied by aν , 0. The uniquely determined formal series
b0 + b1 x + b2 x2 + · · · is now called the reciprocal of a0 + a1 x + a2 x2 + · · · (We 9
can not invert if a0 = 0 since in that case we shall have to introduce negative
exponents and so shall be going out of our ring of power series). In view of this
1
definition it is meaningful to write = 1+ x+ x2 +· · · since we have shown
1−x
1
that (1 − x)(1 + x + x2 + · · · ) = 1. Replacing x by xk , = 1 + xk + x2k + · · ·
1 − xk
Using this expression, identity (1) may be written
Y∞ Y  Y 1
(1 + xn ) 1 + xk + x2k + · · · = .
n=1 k odd k odd
1 − xk
For any N,
Y 1 1
= Q
k odd
1 − xk (1 − xk )
k≤N k odd,k≤N

Since this is true for any N, we may interchange the order of factors in the
entire product and get
∞
Y 1 1
k)
= Q
n odd
(1 − x (1 − xk )
k odd
1. Lecture 9

Therefore, in its revised form identity (l) becomes:

∞
Y 1
(1 − xn ) = Q
n=1
(1 − xn )
n odd

In order to determine in how many ways a number n can be split into k

parts, Wuler introduced a parameter z into his formal power series. (The prob-
lem was proposed to Euler in St.Petersburgh: in how many ways can 50 be
decomposed into the sum of 7 summands?). He considered such expression as
(1 + xz)(1 + x2z) · · · This is a formal power series in x. The coefficients of x are
now polynomials in z, and since these polynomials form a ring they porvide an
sdmissible set of coefficients. The product is not a formal power series in z 10
however. The coefficient of z for example, is an infinite sum which we do not
allow.

(1 + xz)(1 + x2 z)(1 + x3 z) · · ·
= 1 + zx + zx2 + (z + z2 )x3 + (z + z2 )x4 + (z + 2z2 )x5 + · · ·
= 1 + z(x + x2 + x3 + · · · ) + z2 (x3 + x4 + 2x5 + · · · ) + · · ·
= 1 + zA1 (x) + z2 A2 (x) + z3 A3 (x) + · · · (2)

The expressions A1 (x), A2 (x), · · · are themselves formal power series in x.

They begin with higher and higher powers of x, for the lowest power of x
occurring in Am (x) is x1+2+3+···+m = xm(m+1)/2 . This term arises by multiplying
(xz)(x2 z)(x3 z) · · · (xm z). The advantage in the use of the parameter z is that any
power of x multiplying zm is obtained by multiplying m different powers of x.
Thus each term in Am (x) is the product of m powers of x. The z′ s therefore
record the number of parts we have used in building up a number.
QN
Now we consider the finite product PN (z, x) ≡ (1 + zxn ).
n=1
PN (z, x) is a polynomial in z: PN (z, x) = 1 + zA(N) 2 (N)
1 (x) + z A2 (x) + · · · +
N (N) (N) N(N+1)/2
z AN (x), where AN (x) = x . Replacing z by Zx, we have
N
Y
(1 + zxn+1 ) = PN (zx, x)
n=1

= 1 + zxA(N) 2 2 (N)
1 (x) + z x A2 (x) + · · ·

So
 
(1 + zx)PN (zx, x) = 1 + zxN+1 PN (z, x),
1. Lecture 10
 
(1 + zx) 1 + zxA(N) N (N)
1 (x) + · · · + (zx) AN (x)
  
= 1 + zxN+1 1 + A(N) 2 (N)
1 (x) + z A2 (x) + · · ·

We may now compare powers of z on both sides since these are polynomi- 11
als. Taking zk , k ≤ N, we have

xk A(N) k (N) (N)

k (x) + x Ak−1 (x) = Ak (x) + x
N+1 (N)
Ak−1 (x);
(N) (N)
 
Ak (x)(1 − x ) = ak−1 (x)x 1 − xN+1−k ,
k k

xk  
A(N)
k (x) = k
1 − xN+1−k A(N)
k−1 (x),
1−x
xk
A(N)
k (x) ≡ A(N) (x) (mod xN ).
1 − xk k−1
From this recurrence relation we immediately have
x
A(N)
1 (x) ≡ (mod xN ),
1−x
x · x2
A(N)
2 (x) ≡ (1 − x)(1 − x2 ) (mod xN )

x3
≡ (mod xN )
(1 − x)(1 − x2 )
.....................
.....................
xk(k+1)/2
A(N)
k ≡ (mod xN )
(1 − x)(1 − x2 ) · · · (1 − xk )
Hence
∞
Y zx z 2 x3 z 3 x6
(1 + zxn ) ≡ 1 + + +
n=1
1 − x (1 − x)(1 − x ) (1 − x)(1 − x2 )(1 − x3 )
2

+··· (mod xN )
Lecture 2

In the last lecture we proved the identity: 12

∞
Y ∞
X
(1 + zxk ) = zk Ak (x), (1)
n=1 k=0

where
xk(k+1)/2
Ak (x) = (2)
(1 − x)(1 − x2 ) · · · (1 − xk )
We shall look upon the right side of (1) as a power series in x and not as a
power-series in z, as otherwise the infinite product on the left side would have
no sense in our formalism. Let us inerpret (1) arithmetically. If we want to
decompose m into k summands, we have evidently to look for zk and then for
xm , and the coefficient of zk xm on the right side of (1) gives us exactly what we
want. We have
∞ ∞ ∞
1 X
n1
X
2n2
X
= x x · · · xknk
(1 − x)(1 − x2 ) · · · (1 − xk ) n =0 n =0 n =0
1 2 k
∞
X
= p(k) m
m x ,
m=0

say, with p(k)

0 = 1.

Therefore m occurs only in the form

m = n1 + 2n2 + · · · + knk , n j ≥ 0,

and p(k)
m tells us how often m can be represented by k dfferent summands (with
possible repetitions). On the other hand the coefficient of xm on the left-side

11
2. Lecture 12

of (1) gives us the number of partitions of m into summands not exceeding k.

Hence,
Theorem 2. m can be represented as the sum of k different parts as often as 13
k(k + 1)
m− can be expressed as the sum of parts not exceeding k (repetition
2
being allowed).
(In the first the number of parts is fixed, in the second, the size of parts).
In a similar way, we can extablish the identity
∞
1 X
Q∞ n
= zk Bk (x), (3)
n=1 (1 − zx ) k=0

with B0 = 1, which again can be interpreted arithmetically as follows.

The left side is
∞
X ∞
X ∞
X
(zx)n1 (zx2 )n2 (zx3 )n3 · · ·
n1 =0 n2 =0 n3 =0

and
xk
Bk (x) = (4)
(1 − x)(1 − x2 ) · · · (1 − xk )
The left-side of (3) gives m with the representation

m = n1 + 2n2 + 3n3 + · · ·

i.e., as a sum of parts with repetitions allowed. Exactly as above we have:

Theorem 3. m can be expressed as the sum of k parts (repetitions allowed) as
often as m − k as the sum of parts not exceeding k.
We shall now consider odd summands which will be of interest in connex-
ion with V-function later. As earlier we can establish the identity
Y ∞
X
(1 + zxV ) = zk Ck (x) (5)
V odd k=0

with the provide that C◦ (x) = 1. The trick is the same. One studies temporatily 14
QV
a truncated affair (1 + zxV ), replaces z by zx2 and evaluates Ck (x) as in
V=1
Lecture 1. This would be perfectly legitimate. However one could proceed as
2. Lecture 13

Euler did - this is not quite our story. Multiplying both sides by 1 + zx2 , we
have
X∞ ∞
X
zk Ck (x) = (1 + zx2 ) zk x2k Ck (x).
k=0 k=0
Now compare powers of z on both sides - and this was what required some
2
extra thought. Ck (x) begins with x1+3+···+(2k−1) = xk ; in fact they begin with
later and later powers of x and so can be added up. We have
C0 = 1,
Ck (x) = x2k Ck (x) + x2k−1Ck−1 (x), k > 0,
x2k−1
or Ck (x) = Ck−1 (x)
1 − x2k
from this recurrence relation we obtain
x
C1 (x) = ,
1 − x2
3
x x4
C2 (x) = C 1 (x) = ,
1 − x4 (1 − x2 )(1 − x4 )
x5 x9
C3 (x) = C 2 (x) = ,
1 − x6 (1 − x2 )(1 − x4 )(1 − x6 )
...............
...............
2
xk
Ck (x) = ,
(1 − x2 )(1 − x4 ) · · · (1 − x2k )
carrying on the same rigmarole.
Now note that all this can be retranslated into something.
Let us give the number theoretic interpretation. The coefficient of zk xm 15
gives the number of times m can be expressed as the sum of k different odd
1
summands. On the other hand, the coefficients in the expansion of (1−x2 )···(1−x 2k )

give the decomposition into even summands, with repetitions. Hence,

Theorem 4. m is the sum of k different odd parts as often as m − k2 is the sum
2
of even parts not exceeding 2k, or what is the same thing, as m−k
2 is the sum
of parts not exceeding k. (since m and k are obviously of the same parity, it
2
follows that m−k
2 is an integer).
Finally we can prove that
∞
1 X
Q V
= zk Dk (x) (6)
V odd (1 − zx ) k=0
2. Lecture 14

Replacing z by zx2 , we obtain

xk
Dk (x) = ,
(1 − x2 ) · · · (1 − x2k )
leading to the
Theorem 5. m is the sum of k odd parts as often as m − k is the sum of even
parts not exceeding 2k, or m−k m−k
2 is the sum of even parts not exceeding k. ( 2
again is integral).
Some other methods
Temporarily we give up power series and make use of graphs to study par-
titions. A partition of N may be represented as an array of dots, the number of
dots in a row being equal to the magnitude of a summand. Let us arrange the
summands according to size.
For instance, let us consider a partition of 18 into 4 different parts 16

If we read the diagram by rows we get the partiton 18=7+5+4+2. On the

other hand reading by columns we have the pertition 18= 4+4+3+3+2+1+1.
In general it is clear that if we represent a partition of n into k parts graphically,
then reading the graph vertically yields a partition of n with the largest part
k, and conversely. This method demonstrates a one-to-one correspondence
between partitions of n with k parts and partitions sees that the number of
partitions of n with largest part k is equal to the number of partitions of n − k
into parts not exceeding k.
2. Lecture 15

Draw a diagonal upward starting from the last but one dot in the column on
the extreme left. All the dots to the right of this diagonal constitute a partition of
12 into 4 parts. For each partition of 18 into 4 different parts there corresponds
thus a partition of 18 − 4.32 = 12 into parts. This process works in general for a 17
partition of n with k different parts. If we throw away the dots on and to the left
of the diagonal (which is drawn from the last but one point from the bottom in
order to exsure that the number of different parts constinues to be exactly k),
we are left with a partition of n − (1 + 2 + 3 + · · · + (k − 1)) = n − k(k−1)
2 . This
partition has exactly k parts because each row is longer by at least one dot than
the row below it, so an entire row is never discarded. Conversely, starting with
a partition of n − k(k−1)
2 into k parts, we can build up a unique partition of n into
k different parts. Add 1 to the next to the smallest part, 2 to the next longer, 3 to
the next and so on. This one-to-one correspondence proves that the number of
partitions of n into k different parts equals the number of partitions of n − k(k−1)
2
into k parts.
We can prove graphically that the number of partitons of n into k odd sum-
mands is the same as the number of partitions of n − k2 into even summands
not exceeding k. The last row of the

diagram contains at least one dot, the next higher at least three, the one
above at least five, and so on. Above and on the diagonal there are 1 + 3 + 5 +
· · · + (2k − 1) = k2 dots. When these are removed, an even number of dots is
left in each row, althogether adding up to n − k2 . This proves the result.
Theorem 1 can also be proved graphically, although the proof is not quite 18
as simple. The idea of the proof is examplified by considering the partitons of
35. We have

35 = 10 + 8 + 7 + 5 + 4 + 1
=5×2+1×8+7+5+1×4+1
= 5(2 + 1) + 7 × 1 + 1(8 + 4 + 1)
2. Lecture 16
 
 
 · · · + }1
7 + 5 + 5 + 5 + 1| + {z
13 times

Thus to each unrestricted partition of 35 we can make correspond a parti-

tion into add summands with possible repetitions. Conversely
7 × 1 + 5 × 3 + 1 × 13 = 7 × 1 + 5(1 + 2) + 1(23 + 22 + 20 )
= 7 + 5 + 10 + 8 + 4 + 1.
Now consider the following diagram

2 4 6 3

13 times

20

Each part is represented by a row of dots with the longest row at ehe top,
second longest next to the top, etc. The oddness of the parts allows uo to
place the rows symmetrically about a central vertical axis. Now connect the 19
dots in the following way. Connect the dots on this vertical axis with those on
the left half of the top row. Then connect the column to the right of this axis
to the other half of the top row. We continue in this way as indicated by the
diagram drawing right angles first on one side of the centre and then on the
other. We now interpret this diagram as a new partition of 35 each part being
represented by one of the lines indicated. In this way we obtain the partition
20+6+4+3+2 of 35 into different parts. It can be proved that this method works
in general. That is, to prove that given a partition of n into odd parts, this
method transforms it into a unique partition of n into distinct parts; conversely,
given a partation into distinct parts, the process can be reversed to find a unique
partition into odd parts. This establishes a one-to-one correspondence between
the two sorts of partitions. This proves our theorem.
Lecture 3
∞
P
The series zk Ak (x) that we had last time is itself rather interesting; the Ak (x) 20
k=0
have a queer shape:

xk(k−1)/2
Ak (x) =
(1 − x)(1 − x2 ) · · · (1 − xk )
Such series are called Euler series. Such expressions in which the factors in
the denominator are increasing in this way have been used for wide generalisa-
tions of hypergeometric series. Euler indeed solved the problem of computing
the coefficients numerically. The coefficient of zk xm is obtained by expanding
1
(1−x)···(1−xk )
as a power series. This is rather trivial if we are in the field of com-
plex numbers, since we can then have a decomposition into partial fractiions.
Euler did find a nice sort of recursion formula. There is therefore a good deal
to be said for a rather elementary treatment.
We shall, however, proceed to more important discussions the problem of
unrestricted partitions. Consider the infinite product (this is justifiable modulo
xN )
∞ ∞ ∞
Y 1 YX
= xmn
m=1
1 − xm m=1 n=0
∞
X ∞
X ∞
X
= xn1 x2n2 · x3n3 · · ·
n1 =0 n2 =0 n j =0
2
= 1 + x + 2x + · · ·
X∞
=1+ p n xn (1)
n=1

17
3. Lecture 18

What does pn signify? pn appeared in collecting the term xn . Following 21

Euler’s idea of addition of exponents, we have
n = n1 + 2n2 + 3n3 + 4n4 + · · · n j ≥ 0, (2)
so that pn is the number of solutions of a finite Diophantine equation (since the
right side of (2) becomes void after a finite stage) or the number of ways in
which n can be expressed in this way, or the number of unrestricted partitions.
Euler wrote this as
∞
1 X
Q∞ m)
= p(n)xn , (3)
m=1 (1 − x n=0
with the provide that p(0) = 1.
We want to find as much as possible about p(n). Let us calculate p(n).
Expanding the product,
∞
Y
(1 − xn ) = (1 − x)(1 − x2 )(1 − x3 ) · · ·
n=1
= 1 − x − x2 + x5 + x7 − x12 − x15 + + − − · · ·
(Note Euler’s skill and patience; he calculated up to xn and found to this
surprise that the coefficients were always 0, ±1, two positive terms followed by
two negative terms). We want to find the law of exponents, as every sensible
man would. Writing down the first few coeffieicnts and taking differences, we
have
0 1 2 5 7 12 15 22 26
1 1 3 2 5 3 7 4
the sequence of odd numbers interspersed with the sequence of natural num-
bers. Euler forecast by induction what the general power would be as follows. 22

7 2 0 1 5 12 22
−5 −2 1 4 7 10
3 3 3 3 3
Write down the coefficients by picking up 0, 1 and every other alternate
term, and continue the row towards the left by putting in the remaining coeffi-
cients. Now we find that the second differences have the constant value 3. But
an arithmetical progression of the second order can be expressed as a polyno-
mial of the second degree. The typical coefficient will therefore be given by an
expression of the form
3. Lecture 19

aλ2 + bλ + c a(λ + 1)2 + b(λ + 1) + c a(λ + 2)2 + b(λ + 2) + c

a(2λ + 1) + b a(2λ + 3) + b

2 a (the constant second difference)

Hence 2a = 3 or a = 3/2. Taking λ = 0 we find that c = 0 and b = − 12 ,
so that the general coefficient has the form λ(3λ−1)2 . Observing that when λ is
changed to −λ, λ(3λ−1)
2 becomes λ(3λ+1)
2 , the coefficient of xλ(3λ−1)/2 is (−)λ , and
hence
Y∞ Y ∞
(1 − xn ) = (−)λ xλ(3λ−1)/2 , (4)
n=1 λ=−∞

which is Euler’s theorem.

This sequence of numbers λ(3λ−1)
2 played a particular role in the middle
ages. They are called pentagonal numbers and Euler’s theorem is called the
pentagonal numbers theorem. We have the so-called triangular numbers:

1 3 6 10 15
2 3 4 5
1 1 1

where the second differences are all 1; the square-numbers 23

1 4 9 16 25
3 5 7 9
2 2 2

for which the second difference are always 2; and so on.

The triangular numbers can be represented by dots piled up in the form of

equilateral triangles; the square numbers by successively expanding squares.
3. Lecture 20

The pentagons however do not fit together like this. We start with one pen-
tagon; notice that the vertices lie perspectively along rays through the origin.
So take two sides as basic and magnify them and add successive shelves. The
second differences now are always 3:

1 5 12 22

4 7 10
3 3

In general we can have r-gonal numbers where the last difference are all
r − 2.
We go back to equation (4): 24

∞
Y ∞
X
(1 − xm ) = (−)λ xλ(3λ−1)/2
m=1 λ=−∞

It is quite interesting to go into the history of this. It appeared in Euler’s

Introductio in Analysin Infinitorum, Caput XVI, de Partitio numerorum, 1748
(the first book on the differential and integral calculus). It was actually discov-
ered earlier and was mentioned in a paper communicated to the St. Petersburgh
Academy in 1741, and in letters to Nicholas Bernoulli (1742) and Goldbach
(1743). The proof that bothered him for nine years was first given in a letter
dated 9th june 1750 to Goldvach, and was printed in 1750.
The identity (4) is remarkable; it was the first time in history that an identity
belonging to the V-functions appeared (later invented and studied systemati-
cally by Jocobi). The interesting fact is that we have a power-series in which
the exponents are of the second degree in the subscripts. The V-functions have
a representation as a series and slso as an infinite porduct.
The proof of identity (4) is quite exciting and elementary. By using dis-
tributivity we break up the product

(1 − x)(1 − x2 )(1 − x3 )(1 − x4 ) · · ·

in the following way:

(1 − x)(1 − x2 )(1 − x3 )(1 − x4 ) · · · = 1 − x − (1 − x)x2 − (1 − x)(1 − x2 )x3 −

3. Lecture 21

(1 − x)(1 − x2 )(1 − x3 )(1 − x4 ) − (1 − x) · · · (1 − x4 )x5 − · · ·

which may be re-arranged, opening first parenthesis, as

1 − x −Ex2 − (1 − x2 )x 3 2 3 4
MMM − (1 − x )(1 − xRR)x − (1 − x2 )(1 − x3 )(1 4 5
JJJ − x )x
EE RRRRR
EE MMM RRR JJJ
EE MMM RRR JJJ
EE M R JJJ
+x3 +(1 − x2 )x4 +(1 − x2 )(1 − x3 )x5

So 25

1 − x − x2 + x5 + x7 (1 − x2 ) + x9 (1 − x2 )(1 − x3 ) + · · ·
= 1 − x − x2 + x5 + 27 :+(1 − x3 )x 9 3
MMM +(1 − x )(1 −CCx )x
4 11
:: MMM CC
:: MMM CC
:: M CC
CC
9 3 11
−x −(1 − x )x
1 − x − x2 + x5 + x7 − x12 (1 − x3 )x15 − (1 − x3 )(1 − x4 )x18 − · · ·

When this is continued, we get some free terms at the beginning followed
by a typical remainder

(1 − xk )xm + (1 − xk )(1 − xk+1 )xm+k + (1 − xk )(1 − xk+1 )(1 − xk+2 )xm+2k ,

which may be rearranged into

xm + (1 − xk+1 )xm+k + (1 − xk+1 )(1 − xk+2 )xm+2k − xm+k − (1 − xk+1 )xm+2k (*)
= xm − xm+2k+1 − (1 − xk+1 )xm+3k+2 − (1 − xk+1 )(1 − xk+2 )
xm+4k+3 − · · · (**)

We have two free terms with opposite signs at the beginning. In (*) the
difference between exponents in successive terms is k, while in (**) this in-
creases to k + 1; this difference is in both cases the exponent of x in the first
factor. The remainder after the free terms begine with −, so that the sequence
of signs is + − − + + − − · · · This process perpetuates itself and the question 26
remains which powers actually appear. It is sufficient to mark down a scheme
for the exponents which completely characterises the expansion. The scheme
is illustrated by what follows.
3. Lecture 22

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
2 3 4 5 6 7 8 9 10 11 12 13 14 15
5 7 9 11 13 15 17 19 21 23 25 27 29 31
3 4 5 6 7 8 9 10 11 12 13 14
12 15 18 21 24 27 30 33 36 39 42 45
4 5 6 7 8 9 10 11 12 13
22 26 30 34 38 42 46 50 54 58
5 6 7 8 9 10 11 12
35 40 45 50 55 60 65 70
6 7 8 9 10 11
51 57 63 69 75 81
We write down the sequence of natural numbers in a row; the sequence less
the first two membere is repeated in a parallel row below leaving out the first
three placess at the beginning. Adding up we get

5 7 9 11 ... ... ... ,

below which is placed the original sequence less the first three members, again
translating the whole to the right by two places. We again add up and repeat
the procedure. A typical stage in the procedure is exhibited below.

m m+k m + 2k m + 3k m + 4k m + 5k
k+1 k+2 k+3 k+4 k+5
m + 2k + 1 m + 3k + 2 m + 4k + 3 m + 5k + 4 m + 6k + 5

The free indices then appear successively as 27

2+3=5 3 + 4 + 5 = 12
3+4=7 4 + 5 + 6= 15,

and in general:
λ(3λ − 1)
λ + (λ + 1) + · · · + (2λ − 1) = ,
2
λ(3λ + 1)
(λ + 1) + (λ + 2) + · · · + 2λ = ,
2
which are the only exponents appearing. We thus have
∞
Y ∞
X
(1 − xn ) = (−)λ xλ(3λ−1)/2
n=1 λ=−∞
Lecture 4

In the last lecture we proved the surprising theorem on pentagonal numbers: 28

∞
Y ∞
X
(1 − xm ) = (−)λ xλ(3λ−1)/2 (1)
m=1 λ=−∞

We do not need these identities for their own sake, but for their applications
to number theory. We have the same sort of power-series on both sides; let us
compare the coefficients of xn . On the left side n appears as the sum of different
exponents. But in contradiction to previous situations, the coefficients appear
with both positive and negative signs, so that when we collect the terms there
may be cancellations. There are gaps in the powers that appear, but among
those which appear with non-zero coefficients, we have a pair of positive terms
followed by a pair of negative terms and vice versa. In most cases the coef-
ficients are zero; this is because of cancellations, so that roughly terms with
positive and negative signs are in equal number. A positive sign appears if we
multiply an even number of times. otherwise a negative sign. So an even num-
ber of different summands is as frequent generally as an odd number. Hence
the following theorem:
The number of decompositions of n into an even number of different parts
is the same as the number of decompositions into an odd number, with the
exception that there is a surplus of one sort or the other if n is a pentagonal
number of the form λ(3λ − 1)/2.
Before proceeding further let us examine a number of concrete instances.
Take 6 which is not a pentagonal number. The partitions are 6, 1 + 5, 2 + 4, 29
1 + 2 + 3, so that there are two decompositions into an even number of different
parts, and two into an odd number. Next take 7, which is a pentagonal number,
7 = λ(3λ+1)
2 with λ = 2. We can actually foresee that the excess will be in the
even partitions. The partitions are 7, 1 + 6, 2 + 5, 3 + 4, 1 + 2 + 4. Take 8 which

23
4. Lecture 24

again is not pentagonal. We have three in each category: 8, 1 + 7, 2 + 6, 3 + 5,

1 + 2 + 5, 1 + 3 + 4.
This is a very extraordinary property of pentagonal numbers. One would
like to have a direct proof of this. A proof is due to Fabian Franklin (Comptes
Rendus, Paris. 1880), a pupil of the famous Sylvester. The proof is combi-
natorial. We want to establish a one-one correspondence between partitions
containing an even number of summands and those containing an odd number
- except for pentagonal numbers.
Consider a partition with the summands arranged in increasing order, each
summand being denoted by a horizontal row of dots. Mark specifically the first
row,

with r dots, and the last slope, with s dots i.e., points on or below a segn-
ment starting from the dot on the extreme right of the last row and inclined at
45◦ (as in the diagram). We make out two cases.
1. s < r. Transfer the last slope to a position immediately above the first
row. The diagram is now as shown below:

The uppermost row is still shorter than the others. (because in our case 30
s < r). By this procedure the number of rows is changed by 1. This
establishes the one-one correspondence between partition of the ‘odd’
type and ‘even’ type.
4. Lecture 25

2. s ≥ r. As before consider the first row and the last slope.

Take the uppermost row away and put it parallel to the last slope. This
diminishes the number of rows by 1, so that a partition is switched over
from the ‘even’ class to the ‘odd’ class or conversely.
Therefore there exists a one-one correspondence between the two classes.
So we have proved a theorem, which is a wrong one! because we have not
taken account of the exceptional case of pentagonal numbers. The fallacy lies
in having overlooked the fact that the last slope may extend right up to the
first row; the slope and the row may very well interfere. Let us take one such
instance. Let again s < r.

If we place the last slope above the first row this works because the number 31
of points in the first row is also diminished by one, in fact by the disputed point
(notice again that no two rows are equal for s < r − 1). So the interference is of
no account. With s ≥ r we may again have an interfering case. We again place
the top row
4. Lecture 26

behind the last slope, this time with a punishment. We have now shortened the
slope by 1. For s − 1 ≥ r the method is still good. So the only cases of earnest
interference are:

(i) s < r but x ≥ r − 1. Then r − 1 ≤ s ≤ r and hence s = r − 1

(ii) s ≥ r but s − 1 < r. Then s ≥ r > s − 1 and hence s = r.

Here we have something which can no longer be overcome. These are the
cases of pentagonal numbers. In (ii) the total number of dots is equal to

s + (s + 1) + (s + 2) + · · · + (2s − 1)
s(3s − 1)
=
2
In (i) this number = (s + 1) + (s − 2) + · · · + 2s
s(3s + 1)
=
2
These decompositions do not have companions. In general every partition
into one parity of different summands has a companion of the other parity of
different summands; and in the case of pentagonal numbers there is just one in 32
excess in one of the classes.
We now come to the most important application of identity (1). Since
∞
1 X
Q∞ m
= p(n)xn ,
m=1 (1 − x ) n=0

we have on combining this with (1),

∞
X ∞
X
1= p(n)xn (−)λ xλ(3λ−1)/2 (2)
n=0 λ=−∞
4. Lecture 27

This tells us the follwing story. All the coefficients on the right side of (2)
excepting the first must be zero. The typical exponent in the second factor on
the right side is λ(3λ − 1)/2 = ωλ , say. (The first few ω′λ s are 0, 1, 2, 5, 7, 12,
15, . . .). Now look for xn . Since the coefficient in the first factor is p(n) and
that in the second always ±1, we have, since xn (n , 0) does not appear on the
left side

p(n) − p(n − 1) − p(n − 2) + p(n − 5) + p(n − 7) − − + + · · · = 0

or
X
p(n − ωλ )(−)λ = 0 (3)
0≤ωλ ≤n

This is a formula of recursion. Omitting the first index of summation (3)

gives X
p(n) = (−)λ−1 p(n − ωλ ) (4)
0<ωλ ≤n

Let us calculate the first few p(n).

p(0) = 1
p(1) = p(1 − 1) = p(0) = 1
p(2) = p(2 − 1) + p(2 − 2) = 2
p(3) = p(3 − 1) + p(3 − 2) = 3
p(4) = p(4 − 1) + p(4 − 2) = 5
p(5) = p(5 − 1) + p(5 − 2) − p(5 − 5) = 7

(Watch! a pentagonal number - and a negative sign comes into action!). These 33
formulae get longer and longer, but not excessively so. Let us estimate how
λ(3λ − 1)
long these will be. Since ωλ ≤ n we have to look for λ satisfying ≤
2
n, which gives

12λ(3λ − 1) ≤ 24n,
36λ2 − 12λ ≤ 24n,
(5λ − 1)2 = 24n + 1,
√
|6λ − 1| = 24n + 1,
4. Lecture 28

1 1√
|λ − | ≤ 24n + 1.
6 6
1√ 2√
Hence roughly there will be 24n = 6n summands on the left side
3 3
of (3). So their number increases with the square root of n- the expressions do
not get too long after all (for n = 100, we have 17 terms).
These formulae have been used for preparing tables of p(n) which have
been quite useful. For instance Ramanujan discovered some of the divisibility
properties of p(n) by using them. In the famous paper of Hardy and Ramanu- 34
jan (1917) there is a table of p(n) for n ≤ 200. These were computed by
Macmahon, by using the above formulae and the values were checked with
those given by the Hardy-Ramanujan formula. The asymptotic values were
found to be very close to what Macmahon computed. Gupta has extended the
table for p(n) up to 600.
Before making another application of Euler’s pentagonal theorem, we pro-
ceed a bit further into the theory of formal power series. We add now one more
formal procedure, that of formal differentiation. Let

A = a ◦ + a 1 x + a 2 x2 + · · ·

The derivative A′ of A is by definition

A′ = a1 + 2a2 x + 3a3 x2 + · · ·

This is again a power series in our sense. This operation of differentiation

which produces one power series from another is a linear operation:

(A + B)′ = A′ + B′ ,

where B is a second power series. This is easy to verify; actually we need do

this only for polynomials as everything is true modulo xN . Again,

(c A)′ = c A′

as can be seen directly. Also

(A · B)′ = A′ B + A B′.

Let us look into this situation. Start with the simplest case, A = xm , B = xn .
Then

A′ = mxm−1 , B′ = nxn−1
4. Lecture 29

and (AB)′ = (xm+n )′ = (m + n)xm+n−1 ,

also A′ B + AB′ = mxm−1+n + nxm+n−1
= (m + n)xm+n−1

So this is true also for polynomials by linearity, we can do it piecemeal. 35

And as it is enough if we stop short at xN , it is true in general,
Let us add one more remark. Let us write down a special case where A
and B have reciprocals. Then AB has a reciprocal too (since the units form a
group). In this case we have

(AB)′ A′ B′
= + ,
AB A B
which is the rule for logarithmic differentiation. (It is identical with the proce-
dure in the calculus, as soon as we speak of functions). For A, B and C,

(ABC)′ = A′ (BC) + A(BC)′ = A′ BC + AB′C + ABC ′

(ABC)′ A′ B′ C ′
or = + + ,
ABC A B C
and so on; in general, Q ′
K K
n=1 Ak A′ K
X
QK =
k=1 Ak k=1
Ak
We can do this for infinite products also if the products are permissible.
K
Q
Indeed Ak is legitimate if Aℓ = 1 + aℓ(k) xℓ + · · · Consider modulo xN ; break
k=1
at a finite spot and the factors 1 will come into action.
Lecture 5

Let us consider some applications of formal differentiation of power series. 36

Once again we start from the pentagonal numbers theorem:
∞
Y ∞
X
(1 − xm ) = (−)λ xλ(3λ−1)/2
m=1 λ=−∞
X∞
= (−)λ xωλ , (1)
λ=−∞

λ(3λ − 1)
with ωλ = . Taking the logarithmic derivative - and this can be done
2
piecemeal-
∞
P
∞ (−)λ ωλ xωλ−1
X −mxm−1 λ=−∞
= ∞
1 − xm P
m=1 (−)λ xωλ
λ=−∞
Multiplying both sides by x,
∞
P
∞ (−)λ ωλ xωλ
X −mxm λ=−∞
= ∞ (2)
1 − xm P
m=1 (−)λ xωλ
λ=−∞

The left side here is an interesting object called a Lambert series, with a
structure not quite well defined; but it plays some role in number theory. Let
us transform the Lambert series into a power series; it becomes
∞
X ∞
X X∞X
− m xkm = − mxkm ,
m=1 k=1 k1 m=1

30
5. Lecture 31

and these are all permissible power series, because though there are infinitely
many of them, the inner ones begin with later and later terms.
Rearranging, this gives 37

∞
XX ∞
X X
− mxn = − xn m
n=km m=1 n=1 m/n
X∞
=− σ(n)xn ,
n=1
P
where σ(n) denotes the sum of the divisors of n, σ(n) = d.
d|n
(Let us study σ(n) for a moment.

σ(1) = 1, σ(2) = 3, σ3 = 4, σ(5) = 6; indeed σ(p) = p + 1

for a prime p. And σ(n) = n + 1 implies that n is prime. σ(n) is not too big;
there can be at most n divisers of n and so roughly σ(n) = O(n2 ). In fact it is
known that σ(n) = O(n1+epsilon ), ∈> 0, that is, a little larger than the first power.
We shall however not be studying σ(n) in detail).
Equation (2) can now be rewritten as
∞
X ∞
X ∞
X
σ(n)xn (−)λ xωλ = (−)λ−1 ωλ xωλ
n=1 λ=−∞ λ=−∞

Let us look for the coefficient of xm on both sides. Remembering that the
first few ω′λ s are 0, 1, 2, 5, 7, 12, 15 · · · , the coefficient of xm on the left side is

σ(m) − σ(m − 1) − σ(m − 2) + σ(m − 6) + σ(m − 7) − − + + · · ·

On the right side the coefficient is 0 most frequently, because the pentago-
nal numbers are rather rare, and equal to (−)λ−1 ωλ exceptionally, when m = ωλ .

 0

 usually,
σ(m) − σ(m − 1) − σ(m − 2) + + − − · · · = 
(−)λ−1 ωλ for m = ωλ .


We now single out σ(m). 38

We may write

X
λ−1 0

 usually,
σ(m) = (−) σ(m − ωλ ) + 
(−)λ−1 ωλ
 for m = ωλ
0<ωλ <m
5. Lecture 32

This is an additive recursion formula for σ(n). We can make it even more
striking. The inhomogeneous piece on the right side is a little annoying. σ(m −
m) can occur on the right side only for m = ωλ ; σ(0) does not make sense;
however, for our purpose let us define

σ(m − m) = m.

Then σ(ωµ − ωµ ) = ωµ , and the previous formula can now be written

uninterruptedly as X
σ(m) = (−)λ−1 σ(m − ωλ ) (3)
0<ωλ ≤m

We have proved earlier that

X
p(m) = (−)λ−1 p(m − ωλ ) (4)
0<ωλ ≤m

which is a formula completely identical with (3). Here p(m − m) = p(0) = 1.

It is extraordinary that σ(m) and p(m) should have the same recursion formula,
differing only in the definition of the term with n = 0. This fact was noted by
Euler. In fact p(m) is increasing monotonically, while the growth of σ(m) is
more erratic.
There are more relations between p(m) and σ(m). Let us start again with
the identity
Y∞ ∞
X
(1 − xm ) p(m)xm = 1 (5)
m=1 m=0

We know that for a pair of power series A, B such that AB = 1, on taking 39

A′ B′ A′ B′
logarithmic derivatives, we have + = 0 or = − . So from (5),
A B A B
∞
P
∞
X np(n)xn
n=0
σ(m)xm = ∞ ,
P
m=1 p(n)xn
n=0
∞
X ∞
X ∞
X
or σ(m)xm p(k)xk = np(n)xn .
m=1 k=0 n=0

Comparing coefficients of xn ,
X
np(n) = σ(m)p(k),
m+k=n
5. Lecture 33

or more explicitly,
∞
X
np(n) = σ(m)p(n − m) (6)
m=1

This is a bilinear relation between σ(n) and p(n). This can be proved di-
rectly also in the following way. Let us consider all the partitions of n; there
are p(n) such:

n = h1 + h2 + · · ·
n = k1 + k2 + · · ·
n = ℓ1 + ℓ2 + · · ·
.........

Adding up, the left side gives np(n). Let us now evaluate the sum of the
right sides. Consider a particular summand h and let us look for those partitions
in which h figures. These are p(n −h) partitions in which h occurs at least once,
p(n − 2h) in which h occurs at least twices; in general, p(n − rh) in which h
occurs at least r times. Hence the number of those partitions which contain h
exactly r times is p(n − nh) − p(n − n + 1h). Thus the number of times h occurs 40
in all partitions put together is
X n o
n p(n − nh) − (n − n + 1h)
nh≤n

Hence the contribution from these to the right side will be

X n o X
h n p(n − nh) − (n − n + 1h) = h p(n − nh)
nh≤n nh≤n

on applying partial summation. Now summing over all summands h, the right
side becomces
X X XmX
h p(n − nh) = p(n − m),
h nh≤n n/m
n m≤n

on putting rh = m; and this is

X Xm n
X
p(n − m) = p(n − m)σ(m).
m≤n n.m
n m=1

Let us make one final remark.

5. Lecture 34

Again from the Euler formula,

∞
P
∞
X (−)λ−1 ωλ xωλ
m λ=−∞
σ(m)x = ∞
P
m=1 (−)λ xωλ
λ=−∞
∞
P
(−)λ−1 ωλ xωλ
λ=−∞
= ∞
Q
(1 − xm )
m=1
∞
X ∞
X
= (−)λ−1 ωλ xωλ p(m)xm
λ=−∞ m=0

Comparing the coefficients of xm on both sides, 41

σ(m) = p(m) − 1 · p(m − 1) − 2 · p(m − 2) + 5 · p(m − 5)

+ 7 · p(m − 7) − + · · ·
X
= (−)λ−1 ωλ p(m − ωλ )
0≤ωλ ≤m

This last formula enables us to find out the sum of the divisors provided
that we know the partitions. This is not just a curiosity; it provides a useful
check on tables of pertitions computed by other means.
We go back to power series leading up to some of Ramanujan’s theorems.
Jacobi introduced the products
∞
Y
(1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ).
n=1

This is a power series in x; though these are infinitely many factors they
start with progressively higher powers. The coefficients this time are not poly-
nomials in z but from the field R(z), the field of rational functions of z, which
is a perfectly good field. Let us multiply out and we shall have a very nice
surprise. The successive coeffieicnts are:
1
x : z + z−1 (note that this is unchanged when z → z−1 )
x2 : (1 + 1) = 0
x3 : (z + z−1 − z − z−1 ) = 0
x4 : (−1 − 1 + z2 + 1 + 1 + z−2 ) = z2 + z−2 (again unchanged when z → z−1 )
............
5. Lecture 35

We observe that non-zero coeffieicnts are associated only with square ex- 42
ponents. We may threfore provisionally write
∞
Y ∞
X 2
−1 2n−1
2n
(1 − x )(1 + zx 2n−1
)(1 + z x )=1+ (zk + z−k )xk
n=1 k=1
∞
X 2
= z k xk (7)
k=−∞

(with the terms corresponding to ±k folder together). This is a V- series; only

quadratic exponents occur.
We shall now prove the identity (7). But we have got to be careful. Consider
the polynomial
N
Y
ΦN (x, z) = (1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 )
n=1

This consists of terms z j xk with −N ≤ j ≤ N, 0 ≤ k ≤ N(N + 1) + 2N 2 =

2
3N + N. We can rearrange with respect to powers of z. The coefficients are
now polynomials in x. z and z−1 occur symmetrically.

ΦN (x, z) = C◦ (x) + (z + z−1 )C1 (x) + (z2 + z−2 )C2 (x) + · · · + (zN + z−N )C N (x).

Let us calculate the C ′ s. It is cumbersome to look for C◦ , for so many

cancellations may occur. It is easier to calculate C N . Since the highest power
of z can occur only from the terms with the highest power of x, we have
N
Y
C N (x) = (1 − x2n ) × x1+3+···+(2N−1)
n=1
N
Y
N2
=x (1 − x2n )
n=1

Now try to get a recursion among the C ′ s. Replacing z by zx2 , we get 43

N
Y
ΦN (x, zx2 ) = (1 − x2n )(1 + zx2n+1 )(1 + z−1 x2n−3 ).
n=1

Compare ΦN (x, zx2 ) and ΦN (x, z); these are related by the equation

ΦN (x, zx2 )(1 + zx)(1 + z−1 x2N−1 ) = ΦN (x, z)(1 + zx2n+1 )(1 + z−1 x−1 )
5. Lecture 36

The negative power in the last factor on the right is particularly disgusting;
to get rid of it we multiply both sides by xz, leading to

ΦN (x, zx2 )(xz + x2N ) = ΦN (x, z)(1 + zx2N+1 ),

or (1 + zx2N+1 )(C◦ (x) + (z + z−1 )C1 (x) + · · · + (zN + z−N )C N (x))
= (xz + x2n )(C◦ (x) + (zx2 + z−1 x−2 )C1 (x)+
+ (z2 x4 + z−2 x−4 )C2 (x) + · · · + (zN x2N + z−N x−2N )C N (x))
These are perfectly harmless polynomials in x; we may compare coeffi-
cients of zk . Then
Ck (x) + Ck−1 (x)x2N+1 = Ck (x)x2k+2N + x2k−1Ck−1 (x),
or Ck (x)(1 − x2N+2k ) = Ck−1 (x)x2k−1 (1 − x2N−2k+2 )
(We proceed from Ck to Ck−1 since C N is already known).
x−2k+1 (1 − x2N+2k )
Ck−1 (x) = Ck (x)
1 − x2N−2k+2
2
N
Q
Since C N (x) = xN (1 − x2n ), we have in succession 44
n=1

N
2
−2N+1 1 − x4N Y
C N−1 (x) = xN (1 − x2n )
1 − x2 n=1
N
Y
2
= x(N−1) (1 − x2n ) · (1 − x4N );
n=2
N
Y
2
C N−2 (x) = x(N−2) (1 − x2n ) · (1 − x4N )(1 − x4N−2 )
n=3
............

In general,
N
Y j−1
Y
2
C N− j (x) = x(N− j) (1 − x2n ) (1 − x4N−2m )
n= j+1 m=0

or, with j = N − n,
N
Y N−n−1
Y
2
Cn (x) = xn (1 − x2n ) (1 − x4N−2m ) (8)
n=N−n+1 m=0
5. Lecture 37

Equation (8) leads to some congruence relations. The lowest terms of Cn (x)
have exponent

n2 + 2(N − n + 1) = 2N + (n2 − 2n + 1) + 1 ≥ 2N + 1

Hence
2
Cn (x) ≡ xk (mod x2N+1 ) (9)
From the original formula,
N
Y
ΦN (x, z) = (1 − x2n )(1 + zx2n+1 )(1 + z−1 x2n−1 )
n=1
≡ 1 + (z + z−1 )x + (z2 + z−2 )x4 + · · · (mod x2N+1 )
X∞
2
≡ zk xk (mod x2N+1 ),
k=−∞

since the infinite series does not matter, the higher powers being absorbed in 45
the congruence. Hence
∞
Y
ΦN (x, z) ≡ (1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) (mod x2N+1 )
n=1

The new terms x2N+2 , . . ., are absorbed by mod x2N+1 . We have

∞
Y ∞
X 2
(1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) ≡ z k xk (mod x2N+1 )
n=1 k=−∞

Thus both expansions agree as far as we wish, and this is what we mean
by equality of formal power series. Hence we can replace the congruence by
equality, and Jacobi’s identity (7) is proved.
As an application of this identity, we shall now give a new proof of the
pentagonal numbers theorem. We replace x by y3 , as we could consistently in
the whole story; only read modulo y6N+3 . Then we have
∞
Y ∞
X 2
(1 − y6n )(1 + zy6n−3 )(1 + z−1 y6n−3 ) = zk y3k
n=1 k=−∞

We now do something which needs some justification. Replace z by −y.

This is something completely strange, and would interfere seriously with our
reasoning. For ΦN (y3 , z) we had congruences modulo y6N+3 . If we replaced z
5. Lecture 38

by y3 nobody could forbid that. Since z occurs in negative powers, the powers
of y might be lowered too by as much as N. We obtain polynomials in y alone
on both sides, but true modulo y5N+3 , because we may have lowered powers of 46
y. With this proviso it is justified to replace z by −y; so that ultimately we have
∞
Y ∞
X 2
(1 − y6 )(1 − y6n−2 )(1 − y6n−4 ) = (−)k y3k +k
(mod y5N+3 )
n=1 k=−∞

We can carry over the old proof step by step. Since we now have only even
powers of y, this leads to
∞
Y ∞
X
(1 − y2m ) = (−)k yk(3k+1)
m=1 k=−∞

These are actually power series in y2 . Set y2 = x, then

∞
Y ∞
X
(1 − xm ) = (−)K xk(3k+1)/2
m=1 k=−∞

which is the pentagonal numbers theorem.

Lecture 6

In the last lecture we used the Jacobi formula: 47

∞
Y ∞
X 2
(1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) = z k xk (1)
n=1 k=−∞

to give a new proof of Euler’s pentagonal numbers theorem. We proceed to

give another application. We observe again that the right side of (1) is a power
series in x; we cannot do anything about the z′ s and no formal differentiation
can be carried out with respect to z. Let us make the substitution z → −zx. This
again interferes greatly with our variable x. Are we entitled to do this? Let us
look back into our proof of (1). We started with a curtailed affair
∞
Y
ΦN (x, z) = (1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 )
n=1

and this was a polynomial of the proper size and everything went through.
When we replace z by −zx and multiply out, the negative powers might accu-
mulate and we might be destroying xN possibly; nevertheless the congruence
relations would be true this time modulo xN+1 instead of x2N+1 as it was previ-
ously; but this is all we went. So the old proof can be reproduced step by step
and every thing matches modulo xN+1 . (Let us add a side remark. In the proof
of (1) we had to replace z by zx2 - and this was the essential step in the proof.
We cannot do the same here as this would lead to congruences mod x only.
Before we had the congruences we had identities and there we could carry out
any substitution. Then we adopted a new point of view and introduced congru- 48
ences; and that step bars later the substitution z → zx2 .
So let us make the substitution z → −zx without further compuncton. This

39
6. Lecture 40

gives us
∞
Y ∞
X 2
(1 − x2n )(1 − zx2n )(1 − z−1 x2n−2 ) = (−)k zk xk +k

n=1 k=−∞

This is not nicely arranged. There appears an extraordinary term without x-

corresponding to n = 1 in the last factor on the left side; let us keep this apart.
Also on the right side the exponent of x is k(k + 1), so that every number occurs
twice; let us keep these two pieces together. We then have
∞
Y
(1 − z)−1 (1 − x2n )(1 − zx2n )(1 − z−1 x2n )
n=1
∞
X ∞
X
= (−)k zk xk(k+1) + (−)−k−1 z−k−1 xk(k+1)
k=0 k=0
(where in the second half we have replaced k by −k − 1),
∞
X
= (−)k xk(k+1) (zk − z−k−1 )
k=0
∞
X
= (−)k xk(k+1) zk (1 − z−2k−1 )
k=0
∞
X
= (−)k xk(k+1) zk (1 − z−1 )(1 + z−1 + z−2 + · · · + z−2k )
k=0

We now have an infinite series in x equal to another. Now recollect that

our coefficients are from the field R(z) which has no zero divisors. So we may
cancel 1 − z−1 on both sides; this is a non-zero factor in R(z) and has nothing to 49
do with differentiation. This leads to
Y∞ ∞
X
(1 − x2n )(1 − zx2n )(1 − z−1 x2n ) = (−)k xk(k+1) (zk + zk−1 + · · · + z−k ).
n=1 k=0

In the field R(z) we can replace z by 1. We can do what we like in the field
and that is the essence of the power series method. So putting z = 1,
∞
Y ∞
X
(1 − x2n )3 = (−)k xk(k+1) (2k + 1).
n=1 k=0

This is a power series in x ; give it a new name, x2 = y. Then

2

∞
Y ∞
X
(1 − yn )3 = (−)k (2k + 1)yk(k+1)/2 (2)
n=1 k=0
6. Lecture 41

This is a very famous identity of Jacobi, originally proved by him by an

altogether different method using the theory of functions. Let us juxtapose it
with the Euler pentagonal formula:
∞
Y ∞
X
(1 − yn ) = (−)λ xλ(3λ−1)/2 (2a)
n=1 λ=−∞

Let us proceed to yet another application of the triple product formula; we

shall obtain some of Ramanujan’s formulas. Taking away the first part of the
triple product formula we have
∞ ∞
Y X 2 1
(1 + zx2n−1 )(1 + z−1 x2n−1 ) = z k xk ∞
(3)
Q
n=1 k=−∞ (1 − x2n )
n=1

The second part on the right side here is of interest, because it is the gener- 50
ating function of the partition. We had earlier the formula
∞
Y ∞
X
(1 + zx2n−1 ) = zmCm (n),
n=1 m=0
2
(4)
xm
Cm (x) =
(1 − x2 ) · · · (1 − x2m )
and these are permissible power series, beginning with later and later powers
of x, and so the right side of (4) makes sense, as a formal power series in x.
Substituting (4) in (3), we have
∞ ∞ ∞
X X X 2 1
znCr (x) z−sC s (x) = z k xk ∞ (5)
Q
r=0 s=0 k=−∞ (1 − x2n )
n=1

We can compare zO on both sides for, for very high xN the left side will
contain only finitely many terms and all otheres will disappear below the hori-
zon; we can also add as many terms as we wish. So equating coefficients of zO ,
we have
∞
X 1
Cr (x)Cr (x) = ∞ ,
Q
r=0 (1 − x2n )
n=1
∞ 2
X x2r 1
or = ∞
(1 − x ) · · · (1 − x2n )2
2 2 Q
r=0 (1 − x2n )
n=1
6. Lecture 42

We have even powers of x consistently on both sides; so replace x2 by y,

and write down the first few terms explicitly:

y y4 y9
1+ 2
+ 2 2 2
+ +···
(1 − y) (1 − y) (1 − y ) (1 − y) (1 − y2 )2 (1 − y3 )2
2

1
= ∞ (6)
Q
(1 − yn )
n=1

This formula is found in the famous paper of Hardy and Ramanujan (1917) 51
and ascribed by them to Euler. It is very useful for rough appraisal of asymp-
totic formulas. Hardy and Ramanujan make the cryptic remark that it is “a
formula which lends itself to wide generalisations”. This remark was at first
not very obvious to me; but it can now be interpreted in the following way. Let
us look for zk in (5). Then
2
X xk
Cr (x)C s (x) = ∞
Q
r,s
r−s=k
(1 − x2n )
n=1

or, replacing r by s + k, and writting C s for C s (x), the left side becomes

∞ 2 2
X xk x1+(k+1)
C sC s+k =1· + +
s=0
(1 − x2 ) · · · (1 − x2k ) (1 − x2 )2 (1 − x4 ) · · · (1 − x2k+2 )
2
x4+(k+2)
+ +···
(1 − x2 )2 (1 − x4 )2 (1 − x6 ) · · · (1 − x2k+4 )
2
Let us divide by xk . The general exponent on the right side is ℓ2 + (k + ℓ)2 , 52
so on division it becomes 2ℓ2 + 2kℓ. Every exponent is even, which is a very
nice situation. Replace x2 by y, and we get the ‘wide generalisation’ of which
Hardy and Ramanujan spoke:

1 yk+1
+
(1 − y)(1 − y2 ) · · · (1 − yk ) (1 − y)2 (1 − y2 ) · · · (1 − yk+1 )
y2(k+2)
+ +···
(1 − y)2 (1 − y2 )2 (1 − y3 ) · · · (1 − yk+2 )
yl(k+l) 1
2 l 2 l+1 k+l
+··· = ∞ (7)
(1 − u) · · · (1 − y ) (1 − y ) · · · (1 − y ) Q
(1 − yn )
n=1
6. Lecture 43

k is an assigned number and it can be taken arbitrarily.

So such expansions are not unique.
Thus (6) and (7) give two different expansions for
1
∞
.
Q
(1 − yn )
n=1

We are now slowly coming to the close of our preoccupation with power
series; we shall give one more application due to Ramanujan (1917). In their
paper Hardy and Ramanujan gave a surprising asymptotic formula for p(n).
It contained an error term which was something unheard of before, O(n−1/4 ),
error term decreasing as n increases. Since p(n) is an integer it is enough to take
a few terms to get a suitable value. The values calculated on the basis of the 53
asymptotic formula were checked up with those given by Macmahon’s tables
and were found to be astonishingly close. Ramanujan looked at the tables and
with his peculiar insight discovered something which nobody else could have
noticed. He found that the numbers p(4), p(9), p(14), in general p(5k + 4) are
all divisible by 5; p(5), p(12), · · · p(7k + 5) are all divisible by 7; p(11k + 6) by
11. So he thought thiswas a general property. A divisibility property of p(n) is
itself surprising, because p(n) is a function defined with reference to addition.
The first and second of these results are simpler than the third. Ramanujan in
fact suggested more. If we chose a special progression modulo 5λ , then all the
terms are divisible by 5λ . There are also special progressions modulo 72λ−1 ; so
for 11. Ramanujan made the general conjecture that if δ = 5a 7b 11c and 24n ≡ 1
(mod δ), then p(n) ≡ 0 (mod δ). In this form the conjecture is wrong. These
things are deeply connected with the theory of modular forms; the cases 5 and
7 relate to modular forms with subgroups of genus 1, the case 11 with genus 2.
Let us take the case of 5. Take p(5k + 4). Consider Σp(n)xn ; it is nicer
to multiply by x and look for x5k . We have to show that the coefficients of
x5k in xΣp(n)xn are congruent to zerp modulo 5. We wish to juggle around
with series a bit. Take Σan xn ; we want to study x5k . Multiply by the series
1 + b1 x5 + b2 x10 + · · · where the b′ s are integers. We get a new power series
X X
an xn · (1 + b1 x5 + b2 x10 + · · · ) = cn xn ,

which is just as good. It is enough if we prove that for this series every fifth 54
coefficient ≡ 0 (mod 5).
For,
P
X cn xn
a n xn =
1 + b1 x + b2 x10 + · · ·
5
6. Lecture 44
X
= cn xn , (1 + d1 x5 + d2 x10 + · · · ), say.
Then if every fifth coefficient of Σcn xn is divisible by 5, multiplication by
Σdn x5n will not disturb this. For a prime p look at
! ! ! !
p p 2 p 3 p p
(1 + x) p = 1 + x+ x + x +···+ x .
1 2 3 p
All except the first and
 lastp!coefficients on the right side are divisible by p,
for in a typical term qp = (p−q)!q! , the p inm the numerator can be cancelled
only by a p in the denominator. So
(1 + x) p ≡ 1 + x p (mod p).
This means that the difference of the two sides contains only coefficients
divisible by p. This
(1 − x)5 ≡ 1 + x5 (mod 5)
We now go to Ramanujan’s proof that p(5k + 4) ≡ 0 (mod 5) We have 55
X x
x p(n)xn = Q
(1 − xn )
It is irrelevant here if we multiply both sides by a series containing only
x5 , x10 , x15 , · · · . This will not ruin our plans as we have declared in advance.
So
∞ ∞
X Y x Y
x p(n)xn (1 − x5m ) = Q n)
(1 − x5m )
m=1
(1 − x m=1
∞
x Y
≡Q n
(1 − xm )5 modulo 5
(1 − x ) m=1
Y Y 
(1 − x5m ) − (1 − xm )5 has only coefficients divisible by 5
∞
Y
≡x (1 − xm )4 modulo 5
m=1
∞
Y ∞
Y
=x (1 − xm ) (1 − xm )3 .
m=1 m=1

For both products on the right side we have available wonderful expres-
sions. By (2) and (2a),
Y Y ∞
X ∞
X
x (1 − xm ) (1 − xm )3 = x (−)λ(3λ−1)/2 (−)k (2k + 1)xk(k+1)/2
λ=−∞ k=0
6. Lecture 45

The typical term on the right side is

∞
X
(−)λ+k x1+λ(3λ−1)/2 + k(k + 1)/2
k=0

The exponent = 1 + λ(3λ − 1)/2 + k(k + 1)/2, and we want this to be of the
form 5m. Each such combination contributes to x5m . We want 56

λ(3λ − 1) k(k + 1)
1+ + ≡0 (mod 5)
2 2
Multiply by 8; that will not disturb it. So we want

8 + 12λ2 − 4λ + 4k2 + 4k ≡ 0(5),

3 + 2λ2 − 4λ + 4k2 + 4k ≡ 0(5),
2(λ − 1)2 + (2k + 1)2 ≡ 0(5).

This is of the form:

2. a square + another square ≡ 0(5)
Now

A2 ≡ 0, 1, 4(5),
2B2 ≡ 0, 2, 3(5);

and so A2 + 2B2 ≡ 0(5) means only the combination A2 ≡ 0(5) and 2B2 ≡ 0(5);
each square must therefore separately be divisible by 5, or

2k + 1 ≡ 0(5)

So to x5m has contributed only those combinations in which 2k+1 appeared;

and every one of these pieces carried with it a factor of 5. This porves the result.
The case 7k + 5 is even simpler. We multiply by a series in x7 leading to
(1 − xm )6 which is to be broken up into two Jacobi factors (1 − xm )3 . These are
examples of very beautiful theorems proved in a purely formal way.
We shall deal in the next lecture with one more starting instance, the
Rogers-Ramanujan identities which one cannot refrain from talling about.
Lecture 7

We wish to say something about the celebrated Rogers-Ramanujan identities: 57

x x4 x9
1+ + +
1 − x (1 − x)(1 − x ) (1 − x)(1 − x2 )(1 − x3 )
2

1
+··· = Q ; (1)
n>0 (1 − xn )
n≡±1 (mod 5)
x2 x2·3 x3·4
1+ + +
1 − x (1 − x)(1 − x2 ) (1 − x)(1 − x2 )(1 − x3 )
1
+··· = Q (2)
n>0 (1 − xn )
n≡±2 (mod 5)

The right hand sides of (1) and (2), written down explicitly, are respectively
1
(1 − x)(1 − x4 )(1 − x6 )(1 − x9 ) . . .
1
(1 − x )(1 − x )(1 − x7 )(1 − x8 ) . . .
2 3

One immediately observes that ±1 are quadratic residues modulo 5, and

±2 quadratic non-residues modulo 5. These identities were first communicated
by Ramanujan in a letter written to Hardy from India in February 1913 be-
fore he embarked for England. No proofs were given at that time. It was a
remarkable fact, nevertleless, to have even written down such identities. It is 58
true that Euler himself did some experimental work with the pentagonal num-
bers formula. But one does not see the slightest reason why anybody should
have tried ±1, ±2 modulo 5. Then in 1917 something happened. In an old

46
7. Lecture 47

volume of the Proceedings of the London Mathematical Society Ramanujan

found that Rogers (1894) had these identities along with extensions of hyper-
geometric functions and a wealth of other formulae. In 1916 the identities
were published in Macmahon’s Combinatory Analysis without proof, but with
a number-theoretic explanation. This was some progress. In 1917 I.Schur gave
proofs, one of them combinatorial, on the lines of F.Franklin’s proofof Euler’s
theorem. Schue also emphasized the mathematical meaninf of the identities.
Let us look at the meaning of these identities. Let us write the right side of
(1) as a power-series, say,
∞
1 X
= q′ (n)xn ,
(1 − x)(1 − x4 )(1 − x6 )(1 − x9 ) . . . n=0

q′ (n) is the number of terms collected from summands 1, 4, 6, . . . with rep-

etitions, or, what is the same thing, the number of times in which n can be
expressed as the sum of parts ≡ ±1 (mod 5), with repetitions. Likewise, if we
write
∞
1 X
Q = q′′ (n)xn ,
(1 − xn ) n=0
n≡±2(5)
′′
then q (n) is the number of representations of n as the sum of parts ≡ ±2
(mod 5), with repetitions.
The expressions on the other side appear directly.
Take 2
59
xk
(1 − x)(1 − x2 ) · · · (1 − x4 )
If we write
1
= a 0 + a 1 x + a 2 x2 + · · ·
(1 − x)(1 − x2 ) · · · (1 − xk )
then the coefficient an gives us the number of partitions of n into parts not
exceeding k. Let us represent the partitions by dots in a diagram, each vertical
column denoting a summand. Then there are at most k rows in the diagram.
Since k2 is the sum of the k first odd numbers,

k2 = 1 + 3 + 5 + · · · + (2k − 1),

each partition of n into summands not exveeding k can be enlarged into a par-
tition of n + k2 into summands which differ by at least two, for we can adjoin
k2 dots on the left side, putting one in the lowest row, three in the next, five
7. Lecture 48

in the one above and so on finally 2k − 1 in the top most row. Conversely any
partition of n into

parts with minimal difference 2 can be mutilated into a partition of n − k2 into 60

summands not exceeding k. Hence there is a one one correspondence between
these two types. So the coefficients in the expansion of
k2
x
represent the number of times that a number N can
(1 − x)(1 − x2 ) · · · (1 − xk )
be decomposed into k parts (the partitions are now read horizontally in the di-
agram) differing by two at least. When this is done for each k and the results
added up, we get the following arithmetical interpretation of (1): The num-
ber of partitions of n with minimal difference two is equal to the number of
partitions into summands congruent to ±1 (mod 5) allowing repetitions.
A similar explanation is possible in the case of (2). On the left side we
can account for the exponents 2.3, 3.4, . . ., k(k + 1), . . . in the numerator by
means of triangular numbers. In the earlier diagram we adjoin on the left 2,
4, 6, . . ., 2k dots beginning with the lowerst row. The number thus added is
2+4+· · · · · · · · ·+2k = k(k+1); this disposes of xk(k+1) in the numerator. So read
horizontally, the diagram gives us a decomposition into parts which differ by
xk(k+1)
2 at least, but the summand 1 is no longer tolerated. gives
(1 − x) · · · (1 − xk )
N
us therefore the enumeration of x by parts differeing by 2 at least, the part 1
being forbidden. We have in this way the following arithmetical interpretation
of (2): The number of partitions of n into parts not less than 2 and with minimal
difference 2, is equal to the number of partitions of n into parts congruent ±2
(mod 5), repetitions allowed.
By a similar procedure we can construct partitions where 1 and 2 are for- 61
bidden, partitions differing by at least three, etc. In the case where the differ-
ence is 3, we use 1, 4, 7, . . ., so that the number of dots adjoined on the left is
1 + 4 + 7 + · · · to k terms = k(3k − 1)/2, so a pentagonal number, and this is
7. Lecture 49

P xk(3k−1)/2
no surprise. In fact would give us the number of
(1 − x)(1 − x2 ) · · · (1 − xk )
partitions into parts differing by at least 3. And for 4 the story is similar.
The unexpected element in all these cases is the association of partitions
of a definite type with divisibility properties. The left-side in the identities
is trivial. The deeper part is the right side. It can be shown that there can
be no corresponding identities for moduli higher than 5. All these appear as
wide generalisations of the old Euler theorem in which the minimal difference
between the summands is, of course, 1. Euler’s theorem is therefore the nucleus
of all such results.
We give here a proof of the Roger-Ramanujan identities which is in line
with the treatment we have been following, the motiod of formal power series.
It is a transcription of Roger’s proof in Hardy’s ‘Ramanujan’, pp.95-98. We
use the so-called Gaussian polynomials.
Let us introduce the Gaussian polynomials in a much neater notation than
usual. Consider for first the binomial coefficients:
!
n n(n − 1)(n − 2) · · · (n − k + 1)
=
m 1 · 2 · 3 · ···k

(Observe that both in the numerator and in the denominator there are k fac- 62
tors, which are consecutive integers, and that the factors of equal rank in both
numerator and denominator always add up to n + 1). The nk are all integers, as
is obvious from the recursion formula
! ! !
n+1 n n
= +
k k k−1
n n 
n = 1, of course, and by definition, 0 = 1 We also define nk = 0 for k > n
 n
or for k < 0. Observe also the eymmetry: nk = n−k
The Gaussian polynomials are something of a similar nature. We define the
Gaussian polynomial
" # " #
n n
=
k k x
" #
n (1 − x )(1 − xn−1 ) · (1 − xn−k+1 )
n
by =
k (1 − x)(1 − x2 ) · · · (1 − xk )

The sum of the indices of x in corresponding

" # factors in the numerator r and
n n
denominator is n + 1, as in k . That the are polynomials in x is obvious
k
7. Lecture 50

from the recursion formula

" # " # " #
n+1 n n+1 k
= + x
k k k−1
" # " #
n n
where = 1 and = 1 by definition. The recursion formula is just the same
n 0
n
as that for k except for the factor in the second term on the right. Also define
" # " #
0 n
= 1; also let = 1 for k > n or k < 0. 63
0 0
" # " # " #
1 0 0 ◦
= + x = 1,
0 0 −1
" # " #
1 1
= ;
1 0
" #
2 1 − x2
= = 1 + x;
1 1−x
and so on. We also have the symmetry:
" # " #
n n
=
k n−k
The binomial coefficients appear in the expansion
n !
2
X n k
(1 + y) = y.
k=0
k
" #
n
Likewise, the Gaussian polynomial appear in expansion:
k

(1 + y)(1 + xy)(1 + x2 y) · · · (1 + xn−1 y) = 1 + yG1 (x) + y2G2 (x) + · · · + ynGn (x)

" #
k(k−1)/2 n
where Gk (x) = x
k
" #
n 
Notice that for x = 1, = nk . Changing y to yx we get the recursion
k
formula stated earlier.
We now go back to an identity we has porved sometime back:
∞
Y
(1 + zx2n−1 ) = 1 + zC1 (x) + z2C2 (x) + · · · (1)
n=1
7. Lecture 51

where 2
xk
Ck (x) =
(1 − x2 ) · · · (1 − x2k )
Now write 64

x2 = X, 1 − X = X1 − X 2 = X2 , . . . , 1 − X k = Xk ;
(1 − X)(1 − X 2 ) · · · (1 − X k ) = X, X2 . . . Xk = Xk !

With this notation,

2
xk
Ck (x) =
Xk !
From Jacobi’s triple porduct formula, we have
∞
P 2
∞
Y z l xl
ℓ=−∞
(1 + zx2n−1 )(1 + z−1 x2n−1 ) = ∞ (2)
Q
n=1 (1 − x2n )
n=1

By (1), the left side of (2) becomes

∞ ∞ ∞
X X X Bn (z, x)
zrCr (x) z−sC s (x) = ,
r=0 s=0 n=0
Xn !

where X◦ ! is put equal to 1. Bn (z, x) is the term corresponding to r + s = n when

the left side is multiplied out in Cauchy fashion. Thus
X
Bn (z, x) = Xn ! zr−sCr (x)C s (x)
r+s=n
n 2 2
X xr +s
= Xn ! zn−2r (r + s = n)
r=0
Xr !Xn−r !
n " #
X n 2
+r2 n−2r
= x(n−r) z
r X
r=0

Notice that the powers of z occur with the same parity as n. Now (2) can 65
be re-written as
∞
P 2
∞ z l xl
X Bn (z, x) l=−∞
= ∞
Xn ! Q
n=0 (1 − x2n )
n=1
7. Lecture 52

Both sides are formal power series in x of the appropriate sort. The Bn (z, x)
are linear combinations of power series in x with powers of z for coeffieicnts.
We can now compare powers of z. We first take only even exponents z2m ; we
then have infinitely many equations of formal power series. We multiply the
equation arising from z2m by (−)m xm(m−1) and add all these equations together;
(amd that is the trick, due to Rogers) we can do this because of linearity. Then
∞
P 2
∞ (−)m xm(m−1) x(2m)
X β2l (x) m=0
= ∞ , (3)
X2l ! Q
l=0 (1 − x2n )
n=1
2l "
X #
2l 2 2
where β2l (x) = x(2l−r) +r (−)l−r x(l−r)(l−r−1)
r X
r=0

Writting l − r = s,
X l " #
2l 2 2
β2l (x) = x2l +2s (−) s x s(s−1)
l−s
s=−l
Xl " #
2l2 2l 2
=x (−) s x3s − s
l+s
s=−l

(because of the symmetry between l − s and l + s). Separating out the term
corresponding to s = 0 and folding together the terms corresponding to s and 66
−s,
" # l " # 
 
2l2  2l
 X 2l 
s s(3s−1) 2s 
β2l (x) = x  + (−) x (1 + x )
 l
 l+s 

s=1
 l " # l " # 

X s 2l X 2l 

2l2  s(3s−1) s s(3s+1) 
=x  (−) x + (−) x 

 l+s l+s 

s=1 s=0
 l " # l " # 
X
 2l X 2l 

2l2  s+1 (s+1)(3s+2) s s(3s+1) 
=x  (−) x + (−) x  (4)

 l+s+1 l+s 

s=0 s=0

Then
l # " !
2l2
X 2ls s(3s+1) 1 − X l−s 4s+2
β2l (x) = x (−) x 1− x
l+s 1 − X l+s+1
s=0
l " #
2
X 2l 1 − X 2s+1
= x2l (−) s x s(3s+1)
l+s 1 − X l+s+1
s=0
7. Lecture 53

2 l " #
x2l X
s 2l + 1
= (−) x s(3s+1) (1 − x4s+2 ) (5)
1 − X 2l+1 s=0 l+s+1

Let us now compute β2l+1 (x). For this we compare the coefficients of z2m+1 ,
multiply the resulting equations by (−)m xm(m−1) and add up. Then
∞
P 2
∞ (−)m xm(m−1) x(2m+1)
X β2l+1 (x) m=0
= ∞ , (6)
X2l+1 ! Q
l=0 (1 − x2n )
n=1

where 67
2l+1
X" #
2l + 1 (2l+1−r)2
β2l+1 (x) = x + r2 (−)l−r x(l−r)(l−r−1)
r
r=0
Writting l − r = s, this gives
l " #
X 2l + 1 (l+1−s)2 +(l−s)2 s s(s−1)
β2l+1 (x) = x (−) x
l−s
s=−l−1
l " #
X 2l + 1 2 2 2
= (−) s x3s +s+l +(l+1)
l−s
s=−l−1
 l " #
X
 2l + 1
2l2 +2l+1  s
=x  (−) x s(3s+1)

 l+s+1
s=0
l " # 
X 2l + 1 

s+1 (−s−1)(−3s−2) 
+ (−) x 
l+s+1 

s=0
l " #
2l2 +2l+1
X
s 2l + 1
=x (−) x s(3s+1) (1 − x4s+2 ) (7)
l+ s+1
s=0

This expression for β2l+1 (x) is very neat; it is almost the same as β2l (x) but
for trivial factors. Let us go back to β2l+1 (x) in its best shape.
(" #
2l2 +2l+1 2l + 1
β2l+1 (x) = x
l
X l " # " # !

2l + 1 s s(3s+1) 2l + 1 
s s(3s−1) 
+ (−) x + (−) x 
l+s+1 l+s 

s=1
" # X l " # !

2l2 +2l+1  2l + 1
 2l + 1 s s(3s−1) 1 − X l−s+1 2s 


=x  + (−) x 1+ x 
 l
 l+s 1−X l+s+1 

s=1
7. Lecture 54

Since 68

1 − X l−s+1 2s 1 − X l+s+1 + X s − X l+1 (1 − X l+1 )(1 + X s )

1+ x = = ,
1 − X l+s+1 1 − X l+s+1 1 − X l+s+1
2 1 − X l+1
β2l+1 (x) = x2l +2l+1
1 − X 2l+2
" # l+1 " # 

 2l + 2 X 2l + 2
 

s s(3s−1) 2s 
 + (−) x (1 + x )
 l+1
 l+s+1 

s=1

This fits with β2l+2 . Now we can read off the recursion formulae. The
consequences are too very nice facts. The whole thing hinges upon the courage
to tackle these sums. We did not do these things ad hoc.
Let us compare β2l+1 with β2l

β2l+1 = x2l+1 (1 − X 2l+1 )β2l ;

1 − X l+1
β2l+1 = x−2l−1 β2l+2 ;
1 − X 2l+2
1 − X 2l+2
so β2l+2 = x2l+1 β2l+1 ,
1 − X l+1
and β◦ = 1. These things collapse beautifully into something which we could
not foresee before. Of course the older proof was shorter. This proof fits very
well into our scheme.
Lecture 8

Last time we obtained the two fundamental formulae for β2l , β2l+1 , from which 69
we deduced the recurrence relations:
β2m+1 = x2m+1 (1 − x2(2m+1) )β2m ,
1 − x2(2m+2) (1)
β2m+2 = x2m+1 β2m+1
1 − x2(m+1)
β2m came from B2m by a substitution which was not yet plausible. Let us cal-
culate the first few β′ s explicitly. By definition
B0 = 1 = β0
β1 = x(1 − x2 ) β0 = x(1 − x2 )
1 − x4
β2 = x β1 = x2 (1 − x4 )
1 − x2
β3 = x (1 − x6 )
3
β2 = x5 (1 − x4 )(1 − x6 )
1 − x8
β4 = x3 β3 = x8 (1 − x6 )(1 − x8 );
1 − x4
and in general,
2
β2m = x2m (1 − x2m+2 )(1 − x2m+4 ) · · · (1 − x4m )
2 X2m !
= Xm (with X = x2 ); (2)
Xm !
and similarly,
2
β2m+1 = x2m +2m+1
(1 − x2m+2 )(1 − x2m+4 ) · · · (1 − x4m+2 )
2 X2m+1 !
= X m +m x · (3)
Xm !

55
8. Lecture 56

This is a very appealing result. We got the β′ s in the attempt of ours to 70

utilise the Jacobi formula. We actually had
∞
P 2
(−)l x5l −l ∞
l=0
X β2m
∞ = ,
Q X
m=0 2m
!
(1 − x2m )
m=1

so that by (2)
∞
P
(−)l X l(5l−l)/2 ∞ 2
l=0
X Xm
∞ = (4)
Q X !
m=0 m
(1 − xm )
m=1
Similarly we had
∞
P 2
(−)l x5l +3l+1
∞
l=0
X β2m+1
∞
= ,
Q X
m=0 2m+1
!
(1 − x2m )
m=1

so that by (3)
∞
P
(−)l X l(5l+3)/2 ∞
l=0
X X m(m+1)
∞ = (5)
Q Xm !
(1 − xm ) m=0
m=1
Now the right side in the Rogers-Ramanujan formula is
∞
Q
(1 − x5m )(1 − x5m−2 )(1 − x5m−3 )
1 m=1
∞
= ∞
Q Q
(1 − x5m−1 )(1 − x5m−4 ) (1 − xm )
m=1 m=1

which becomes, on replacing x by x2 , 71

∞
Q
(1 − x10m )(1 − x10m−4 )(1 − x10m−6 )
m=1
∞
Q
(1 − x2m )
m=1

The numerator is the same as the left side of Jacobi’s triple product formula:
∞
Y ∞
X 2
(1 − x2m )(1 − zx2m−1 )(1 − z−1 x2m−1 ) = (−)l zl xl ,
m=1 l=−∞
8. Lecture 57

with x replaced by x5 and z by x. Hence

∞
Q ∞
P 2
∞
P 2
(1 − x10mm )(1 − x10m−4 )(1 − x10m−6 ) (−)l X 5l +l (−)l X (5l +l)/2
l=−∞ l=−∞ l=−∞
∞ = ∞ = ∞
Q Q Q
(1 − x2m ) (1 − x2m ) (1 − X m )
m=1 m=1 m=1

now
∞
P 2
∞
P 2
∞
P 2
+l (2l)2
(−)l x5l +l (−)k zk xk ∞ (−)l xl x
l=−∞ k=−∞
X Bn (z, x) l=−∞
∞ = ∞ = = ∞ ,
Q Q Xn ! Q
(1 − x2m ) (1 − x2m ) n=0 (1 − x2m )
m=1 m=1 m=1

on replacing z2l by (−)l xl(l+1) , and this we can do because of linearity. Hence 72

∞
P
(−)l X l(5l−1)/2
l=−∞ 1
∞
= ∞
Q Q
(1 − X m ) (1 − x5m−1 )(1 − x5m−4 )
m=1 m=1

Similarly,
∞
Q
(1 − x10m )(1 − x10m−2 )(1 − x10m−8 )
1 m=1
∞ = ∞
Q Q
(1 − X 5m−2 )(1 − X 5m−3 ) (1 − X m )
m=1 m=1
∞
P 2
(−)l x5l +3l
l=−∞
= ∞ .
Q
(1 − X m )
m=1

This time we have to replace z2k+1 by (−)k xk(k−1) . Then

∞
P
(−)l X l(5l+3)/2
1 l=−∞
∞ = ∞
Q Q
(1 − X 5m−2 )(1 − X 5m−3 ) (1 − X m )
m=1 m=1

These formulae are of extreme beauty. The present proof has at least to do
with things that we had already handled. The pleasant surprise is that these
things do come out. The other proofs by Watson, Ramanujan and other use 73
8. Lecture 58

completely unplausible combinations from the very start. Our proof is sub-
stantilly that by Rogers given in Hardy’s Ramanujan, pp.96-98, though one
may not recognize it as such. The proof there contains completely foreign
elements, trigonometric functions which are altogether irrelevant here.
We now give up formal power series and enter into an entirely different
chapetr - Analysis.
 Part II

Analysis

59
Lecture 9

Theta-functions
A power series hereafter shall for us mean something entirely different from 74
∞
P
what it did hitherto. x is a complex variable and an xn will have a value, its
n=0
sum, which is ascertained only only after we introduce convergence. Then
∞
X
f (x) = a n xn ;
n=0

x and the series are coordinated and we have a function on the complex domain.
We take for granted the theory of analytic functions of a complex variable; we
shall be using Caushy’s theorem frequently, and in a moment we shall have
occasion to use Weierstrass’s double series theorem.
Let us go back to the Jacobi identity:
∞
Y ∞
X 2
(1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) = z k xk
n=1 k=−∞
∞
X 2
=1+ (zk + z−k )xk , (z , 0),
k=1

which is a power series in x. Two questions arise. First, what are the domains
of convergence of both sides? Second, what does equality between the two
sides mean? Formerly, equality meant agreement of the coefficients up to any 75
stage; what it means now we have got to explore. The left side is absolutely
convergent - and absolute convergence is enough for us - for |x| < 1; (for the
Q P
infinite product (1+an ) is absolutely convergent if |an | < ∞; z is a complex

60
9. Lecture 61

variable which we treat as a parameter). For the right side we use the Cauchy-
Hadamard criterion for the radius of convergence:
1
ρ= √
lim n |an |
1
= p
k2
lim |z + z−k |
Suppose |z| > 1; then *****************, and

|zk + z−k | < 2|z|k ,

k√
p
k2 2 pk
and |zk + z−k | < 2 |z| → 1 as k → ∞
p
k2
∴ lim( |zk + z−k |) ≤ 1.

It is indeed = 1, not < 1, because ultimately, if k is large enough, |z|k > 1,

and so
1 k
|z| < |zk + z−k |,
2
and we have the reverse inequality. By symmetry in z and 1/z, this holds also
for |z| < 1. The case |z| = 1 does not present any serious difficulty either. So
in all cases ρ = 1. Thus both sides are convergent for |x| < 1, and indeed
uniformly in any closed circle |x| ≤ 1 − δ < 1.
The next question is, why are the two sides equal in the sense of function 76
theory? This is not trivial. Here equality of values of coefficients up to any
definite stage is not sufficient as it was before; the unfinished coefficients before
multiplication may go up and cannot be controlled. Here, however, we ae in a
strong position. We have to prove that
N
Y ∞
X 2
(1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) → 1 + (zk + z−k )xk
n=1 k=1

with increasing N, when |x| < 1, and indeed uniformly so in |x| ≤ 1 − δ < 1.
On the left side we have a sequence of polynomials:
N
Y ∞
X
fN (x) = (1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) = a(N) m
m x , say.
n=1 m=0

(of course the coefficients are all zero beyond a certain finite stage). Now we
know that the left side is a partial product of a convergent infinite product; in
fact fN (x) tends uniformly to a series, f (x), say. Now what do we know about
9. Lecture 62

a sequence of analytic functions on the same domain converging uniformly

to a limit function? The question is answered by Weierstrass’s double series
theorem. We can assert that f (x) is analytic in the same domain at least, and
∞
P
further if f (x) = am xm , then
m=0

am = lim a(N)
m .
N→∞

The coefficients of the limit function have got something to do with the 77
original coefficients. Now
Z
(N) 1 fN (x)
am = dx
2πi xm+1
|x|=1−δ

Let N → ∞; this is permissible by uniform convergence and the a(N)

m , s in
fact converge to Z
1 f (x)
am = dx.
2πi xm+1
|x|=1−δ

(Weirestrass’ own proof of this theorem was what we have given here, in
some disguise; he takes the values at the roots of unity and takes a sort of mean
value).
P 2
Now what are the coefficients in 1 + (zk + z−k )xk ? Observe that the con-
(N) (N)
vergence of am to am is a peculiar and simple one. am indeed converges to a
(N)
known am ; as a matter of fact am = am for N sufficiently large. They reach a
limit and stay put. And this is exactly the meaning of our formal identity. So
the identity has been proved in the function-theoretic sense:
∞
Y ∞
X ∞
X
2 2
(1 − x2n )(1 + zx2n−1 )(1 + z−1 x2n−1 ) = 1 + (zk + z−k )xk = z k xk .
n=1 k=1 k=−∞

These things were done in full extension by Jacobi. Let us employ the ususl 78
symbols; in place of x write q, |q| < 1, and put z = e2πiv . Notice that the right
side is a Laurent expansion in z in 0 < |z| < ∞ (v is unrestricted because we
hace used the exponential). We write in the traditional notation
∞
Y
(1 − q2n )(1 + q2n−1 e2πiv )(1 + q2n−1 e−2πiv )
n=1
∞
X 2
= qn e2πinv
n=−∞
9. Lecture 63

= v3 (v, q)

v3 (and in fact all the theta functions) are entire functions of v. We have taken
|q| < 1; it is customary to write q = eπiτ , so that |q| < 1 implies

|eπiτ | = eRπiτ , Rπiτ < 0

i.e., Riτ < 0 or I mτ > 0

τ is a point in the upper half-plane. τ and q are equivalent parameters. We also

write
V3 (V , q) = V3 (V /τ)
(An excellent accout of the V -functions can be found in Tannery and Molk:
Fonctiones Elliptiques, in 4 volumes; the second volume contains a very well
orginized collection of formulas).
One remark is immediate from the definition of V3 , viz.

V3 (V + 1, q) = V3 (V , q)

On the other hand, 79

∞
Y
V3 (V + τ, q) = (1 − q2n )(1 − q2n−1 e2πiV e2πiτ ) × (1 + q2n−1 e−2πiV e−2πiτ )
n=1
X∞
2
= qn e2πinV e2πinV ,
n=−∞

and since q = eπiτ ,

∞
Y ∞
X 2
(1 − q2n )(1 + q2n+1 e2πiV )(1 + q2n−3 e−2πiV ) = qn +2n e2πinV
n=1 n=−∞

or
∞
1 + q−1 e−2πiV Y
(1 − q2n )(1 + q2n−1 e2πiV )(1 + q2n−1 e−2πiV )
1 + qe2πiV n=1
∞
X 2
= q−1 e−2πiV q(n+1) e2πi(n+1)V
n=−∞
−1 −2πiV
=q e V3 (V , q)
2πiV −1
= (qe ) V3 (V , q)
9. Lecture 64

So we have the neat result:

V3 (V + τ, q) = q−1 e−2πiV V3 (V , q)

1 is a period of V3 and τ resembles a period. It is quite clear that we cannot 80

expect 2 peroids in the full sense, because it is imposible for an entire function
to have two periods. Indeed if ω1 and ω2 are two periods of f , then f (V +
ω1 ) = f (V ), f (V + ω2 ) = f (V ), and f (V + ω1 + ω2 ) = f (V ) and the whole
module generated by ω1 and ω2 form periods. Consider the fundamental region
which is the parallelogram with vertices at 0, ω1 , ω2 , ω1 + ω2 . If the function
is entire it has no poles in the parallelogram and is bounded there (because
the parallelogram is bounded and closed), and therefopre in the whole plane.
Hence by Liouville’s theorem the function reduces to a constant.
While dealing with trigonometric functions one is not always satisfied with
the cosine function alone. It is noce to have another function: cos(x − π/2) =
sin x. A shift by a half-period makes it concenient for us. Let us consider
analogously V3 (V + 21 , q), V3 (V + τ/2, q), and V3 (V + 12 + 2τ , q). Though τ
is not strictly a period we can still speak of the funcdmental region, because
on shifting by τ we change only by a trivial factor. Replace V by V + 12 and
everything is fine as 1 is a period.
∞
1 Y
V3 (V + , q) = (1 − q2n )(1 − q2n−1 e2πiV )(1 − q2n−1 e−2πiV )
2 n=1
∞
X 2
= (−)n qn e2πinV
n=−∞

which is denoted V4 (V , q)
Again 81

∞
τ Y
V3 (V + , q) = (1 − q2n )(1 + q2n−1 e2πiV eπiτ )(1 + q2n−1 e−2πiV e−πiτ )
2 n=1
∞
X 2
= qn e2πinV eπinτ
n=−∞
∞
Y
i.e., (1 + e−2πiV ) (1 − q2n )(1 + q2n e2πiV )(1 + q2n e−2πiV )
n=1
∞
X 2
= qn +n e2πinV
n=−∞
9. Lecture 65

∞
X 2
= q−1/4 e−πiV q(n+1/2) e(2n+1)πiV
n=−∞
= q−1/4 e−πiV V2 (V , q)
∞
P 2
where V2 (V , q) = q(n+1/2) e(2n+1)πiV , by definition. (Here q−1/4 does not
n=−∞
contain an unknown 4th root of unity as factor, but is an abbreviation for e−πiτ/4 ,
so that it is well defined). So
∞
Y
V2 (V , q) = 2q1/4 cos πV (1 − q2n )(1 + q2n e2πτV )(1 + q2n e−2πiV )
n=1

Finally 82
!
1+τ 1 1
V3 (V + , q) = q−1/4 e−πi(V + 2 ) V2 V + , q
2 2
!
1 1
= q−1/4 e−πiV V2 V + , q
i 2
X∞
2n+1 2
= q1/4 e−πiV (−)n q( 2 ) e(2n+1)πiV
n=−∞
!
2 1 −πiV
= cos π V + e
i 2
Y∞    
(1 − q2n ) 1 − q2n e2πiV × 1 − q2n e−2πiV
n=1

Now define !
1
V1 (V , q) = V2 V + , q ,
2
or
∞
Y
V1 (V , q) = 2q1/4 sin πV (1 − q2n )(1 + q2n e2πiV )(1 − q2n e−2πiV )
n=1
∞
X 2n+1 2
= iq−1/4 (−)n q( 2 ) e(2n+1)πiV
m=−∞

Collecting together we have the four V -functions: 83

∞
X 2n+1 2
V1 (V , q) = iq−1/4 (−)n q( 2 ) e(2n+1)πiV
m=−∞
9. Lecture 66

∞
X 2n+1 2
= (−)n q( 2 ) sin(2n + 1)πV
n=0
∞
X 2n+1 2
V2 (V , q) = 2 q( 2 ) cos(2n + 1)πV
n=0
∞
X 2
V3 (V , q) = 1 + 2 qn cos 2nπV
n=1
∞
X 2
V4 (V , q) = 1 + 2 (−)n qn cos 2πnV
n=1

Observe that the sine function occurs only in V1 . Also if q, V are rel these
reduce to trigonometric expansions.
Lecture 10

Let us recapitulate the formulae we has last time. 84

∞
1 X 2n+1 2
V1 (V , q) = (−)n q( 2 ) e(2n+1)πiV
i n=−∞
∞
X 2n+1 2
=2 (−)n q( 2 ) sin(2n + 1)πV
n=∞
∞ 
Y   
= 2q1/4 sin πV 1 − q2m 1 − q2m e2πiV 1 − q2m e−2πiV (1)
m=1
∞
X 2n+1 2
V2 (V , q) = q( 2 ) e(2n+1)πiV
n=−∞
X∞
2n+1 2
=2 q( 2 ) cos(2n + 1)πV
n=0
∞ 
Y   
= 2q1/4 cos πV 1 − q2m 1 + q2m e2πiV 1 + q2m e−2πiV (2)
m=1
∞
X 2
V3 (V , q) = qn e2πiV
n=−∞
∞
X 2
=1+2 qn cos 2nπV
n=1
∞ 
Y   
= 1 − q2m 1 + q2m−1 e2πiV 1 + q2m−1 e−2πiV (3)
m=1

67
10. Lecture 68

∞
X 2
V4 (V , q) = (−)n qn e2nπiV
n=−∞
∞
X 2
=1+2 (−)n qn cos 2nπV
n=1
∞
Y
= (1 − q2m )(1 − q2m−1 e2πiV )(1 − q2m−1 e−2πiV ) (4)
m=1

We started with V3 and shifted the argument V by ‘periods’, and we had, 85

writing q = eπiτ ,
V3 (V + 1, q) = V3 (V , q)
(5)
V3 (V + τ, q) = q−1 e−2πiV V3 (V , q).
Then we took ‘half-periods’ and then something new happened, and we
gave names to the new functions:
!
1
V3 V + , q = V4 (V , q)
2
 τ 
V3 V + , q = q−1/4 e−2πiV V2 (V , q) (6)
2 !
1+τ
V3 V + , q = iq−1/4 e−πiV V1 (V , q)
2

Let us study how these functions alter when the argument V is changed by
1, τ, 1/2, τ/2, (1 + τ)/2. V → V + 1 is trivial; V → V + 1/2 is also easy to see
by inspection. Let us take V + τ. (We suppress the argument q for convenience
of writing).
!
1 1/4 2πiV 1+τ
V1 (V ) = q e V3 V +
i 2
!
1 1/4 πi(V +τ) 1+τ
∴ V1 (V + τ) = q e V3 V + τ +
i 2
!
1 1/4 πiV −1 −2πi(V +1+τ/2) 1+τ
= q e qq e V3 V +
i 2
= e−2πiV e−πi(1+τ) V1 (V , q)
= −AV1 (V , q),

where A = q−1 e−2πiV ; the other conspicuous factor which occurs in similar 86
contexts is denoted B = q−1/4 e−2πiV .
10. Lecture 69

The other transformations can be worked out in a similar way by first going
over to V3 . We collect the results below in tabular form.

1 3 1+τ
V +1 V +τ V + 2 V + 2 V + 2

V1 −V1 −AV1 V2 iBV4 BV3

V2 −V2 AV2 −V1 BV3 −iBV4

V3 V3 AV3 V4 BV2 BV1

V4 V4 −AV4 V3 iBV1 BV2

It may be noticed that each column in the table contains all the four func-
tions; so does each now.
The systematique of the notation for the V -functions is rather questionable.
Whittaker and Watson write V instead of πV , which has the unpleasant con-
sequence that the ‘periods’ are then π and πτ. Our notation is the same as in 87
Tannery and Molk. An attempt was made by Kronecker to systematise a little
the unsystematic notation. Charles Hermite introduced the following notation:
∞
X  2r+µ 2
Vµν (V , q) = (−)νn q 2 e(2n+µ)πiV
n=−∞
X∞  2n+µ 2
πiτ (2n+µ)πiV
= (−)νn e 2 e
n=−∞

where µ, ν = 0, 1 ∗ ∗ ∗ ∗∗. In this notation,

V00 (V , q) = V3 (V , q)
V01 (V , q) = V4 (V , q)
V10 (V , q) = V2 (V , q)
V11 (V , q) = iV1 (V , q).

This, however, has not found any followers.

While writing down derivatives, we always retain the convention that a
prime refers to differentiation with respect to V :

∂
Vα′ (V , q) = Vα (V , q) (α = 1, 2, 3, 4)
∂ν
10. Lecture 70

Taking partial derivatives, we have

∞ !
∂ X 2n + µ  2n+µ 2
Vµν (V /τ) = (−)νn πi e 2 πiτ e(2n+µ)πiV ,
∂τ n=−∞
2
∞
∂2 X
νn
 2n+µ 2
πiτ 2 2
and V µν (V /τ) = (−) e 2 π i (2n + µ)2 e(2n+µ)πiV ,
∂V 2 n=−∞

Comparing these we see that they agree to some extent; in fact, 88

∂ ∂2
4πi Vµν (V /τ) = 2 Vµν (V /τ) (7)
∂τ ∂ν
This is a partial differential equation of the second order, a parabolic equa-
tion with constant coefficients. It is fundamental to write iτ = −t; (7) then be-
comes the differential equation for heat conduction. V -functions are thus very
useful tools in applied mathematics; they were used by Poisson and Fourier in
this connection.
Again,
∞ !2  2
∂ X 2n + µ 2n+µ
Vµν (V , q) = (−)ν q 2 −1 e(2n+µ)πiV ,
∂q n=−∞
2
∂ ∂2
− 4π2 q Vµν (V , q) = 2 Vµν (V , q), (8)
∂q ∂ν
which is another form of (7). Here the uniformity of notation was helpful; it
was not necessary to discuss the different functions separately.
We now pass on to another important topic. The zeros of the theta - func-
tions.
The V -functions are more or less periodic. The exponential factor that is
picked up on passing from one parallelogram to another is non-zero and can
accumulate. It is evident from the definition that

V , (0, q) = 0.

On the other hand V2 , V3 , V4 , 0. (when the argument V is 0 we write 89

hereafter simply V ). This is so because the infinite products are absolutely
convergent. (Let us recall that a product like 1 · 21 · 13 · · · is not properly conver-
gent in the product sense). Again from the definitions,
!
1
V2 =0
2
10. Lecture 71
!
1+τ
V3 =0
2
τ
V4 =0
2
So far we have one zero per parallelogram for each of the functions; and
there can be no other in a parallelogram, as can be seen from the infinite prod-
uct expansions. The zeros of V1 (V , q) are m1 + m2 τ(m1 , m2 integers), for
1 − e2πimτ e2πiV = 0 implies mτ + V1 = m1 or V = m1 − mτ. The zeros
of V1 (V , q), V2(V , q), V3 (V , q), V4(V , q) in the fundamental parallelogram are
τ
nicely arranged in order at the points 0, 21 , 1+τ2 , 2 respectively.

0 1

All the zeros are therefore given by the formulae: 90

V1 (m1 + m2 τ) = 0
!
1
V 2 m1 + m2 τ + =0
2
!
1+τ
V 3 m1 + m2 τ + =0
2
 τ
V 4 m1 + m2 τ + =0
2
It is of interest to study Vα (0, q) (usually written Vα ).

V1 (0) = 0
X∞
1 2
V2 (0) = q(2n+ 2 )
n=−∞
∞
Y
= 2q1/4 (1 − q2m )(1 + q2m )2
m=1
10. Lecture 72

= V2
X∞
2
V3 (0) = qn
n=−∞
∞
Y
= (1 − q2m )(1 + q2m−1 )
m=1
X∞
2
V4 (0) = (−)n qn
n=−∞
Y∞
= (1 − q2m )(1 − q2m−1 )2
m=1

We cannot anything of interest in V1 . Let us look at the others. 91

∞
X 2n+1 2
V1′ (0, q) = V1′ = 2π (−)n (2n + 1)q( 2 )
n=0
 ∞
∞ ′ 
 Y Y 
= 2q1/4 π cos πV (· · · ) + sin πV  (· · · ) 
m=1 m=1 V =0
∞
Y
= 2πq1/4 (1 − q2m )3
m=1

Immediately we see that this yields the interesting identity of Jacobi.

∞
Y ∞
X 2
(1 − q2m )3 = (−)n (2n + 1)qn +n ,
m=1 n=0

or, replacing q2 by x,
∞
Y ∞
X
(1 − xn )3 = (−)n (2n + 1)xn(n+1)/2
n=1 n=0

We had proved this earlier by the method of formal power series. Here we
can differentiate with good conscience.
Now
∞ 2
′
Y 2m 2m−1 2m−1 

πV2 V3 V4 = V1  (1 + q )(1 + q
 )(1 − q )
m=1
∞ 2
′
Y 2m 4m−2

= V1  (1 + q )(1 − q ) ,
m=1
10. Lecture 73

which becomes, on replacing q2 by x, 92

∞ 
Y  2
V1′ n
 (1 + x ) 1 − x 2n−1 

m=1

∞
Q
However, (1 + xn )(1 − x2n−1 ) = 1. We therefore have the very useful and
m=1
pleasant formula
V1′ = πV2 V3 V4
Lecture 11

We found that Vα (V , q) changes at most its sign when V is replaced by V + 1, 93

while it picks up a trivial factor A when V is replaced by V + τ. If we form
quotients, A will cancel out and we may therefore expect to get doubly-periodic
functions. Let us form some useful quotients:
V2 (V , q)
f2 (V ) =
V1 (V , q)
V3 (V , q)
f3 (V ) =
V1 (V , q)
V4 (V , q)
f4 (V ) =
V1 (V , q)
For simplicity of location of poles it is convenient to take V1 in the denom-
inator since it has a zero at the origin. From the table of the V -functions we
find that these functions are not quite doubly periodic:
f2 (V + 1) = f2 (V ) f3 (V + 1) = − f3 (V )
f2 (V + τ) = − f2 (V ) f3 (V + τ) = − f3 (V )
f4 (V + 1) = − f4 (V )
f4 (V + τ) = f4 (V )
So the functions are not doubly periodic; they do not return to themselves. 94
And we cannot expect that either. For suppose any of the functions f were
actually doubly periodic. We know that each has a pole of the first order per
parallelogram. Integrating round the parallelogram with vertices at ± 1+τ 1−τ
2 ,± 2
(so that the origin which is the pole is enclosed), we have
Z
f (V )dV = 0

74
11. Lecture 75

0 1

i.e., the sum of the residues at the poles =0. This means that either the
pole is a double is a double pole with zero residue, or there are two simple
poles with residues equal in magnitude but opposite in sign. However neither
of these is the case. So there is no necessity for any further experimentation.
Let us therefore consider the squares
f22 (V ), f32 (V ), f42 (V )
these are indeed doubly periodic functions. And they are even functions. So
the expansion in the neighbourhood of the pole will not contain the term of
power −1. Hence the pole must be a double pole with residue zero. So they
are closely related to the Weierstrassian function P(V ), and must indeed be
of the form CP(V ) + C1 .
So we have constructed doubly periodic functions. They are essentially 95
P(V ). ω1 and ω2 of P(V ) are our 1 and τ. In order to get a better insight
we need the exact values of the functions. Let us consider their pole terms.
Expanding in the neighbourhood of the origin,
V ′′
Vα (V , q) Vα + 2!α V 2 + · · ·
= V′ V1′′′
V1 (V , q) 3
1! V + 3! V + · · ·
1

 ′′ 
1 Vα 2
Vα  1 + 2 Vα ν + · · · 
 
=  ′′′ 
νV1′  1 + V1 ν2 + · · · 

6V1′
! !2 
Vα 1 Vα′′ 2  V ′′′ 
= 1 + ν +··· 1 − 1 ν2 + · · · + (· · · )3 − · · · 
νV1′ 2 Vα 6V1 ′

′′ ′′′ ! !
Vα V V
= 1 + ν2 − 1 ′ +···
νV1′ 2Vα 6V1
!2
Vα (V , q)
∴ fα2 =
V1 (V , q)
11. Lecture 76
! !
Vα2 1 2 Vα
′′ V1′′′
= ′ 1+V − +···
V 21 V 2 V1 3V1′

Let us now specialise α. We have a special interest in V3 because it is such a

′
P∞ 2 V 2 1
nice function: V3 = qn . We have 1 2 fα2 (V ) = + non-negative powers
n=−∞ Vα V2
of V .
If we take two such and take the difference, the difference will no longer 96
have a pole. Taking α = 2, 4, for instance,
′ ! ′ !
V1 2 V2 (V , q) 2 V1 2 V4 (V , q) 2 V2′′ V4
− = − +positive powers of V (*)
V22 V1 (V , q) V42 V1 (V , q) V2 V4′′

The left side is a doubly periodic function without a pole and so a constant
V ′′ V ′′
C; the right side is therefore just 2 − 4 . The vanishing of the other terms
V2 V4
on the other terms on the right side, of course, implies lots of identities.
So we have already computed C in one way:

V2′′ V4′′
C= −
V2 V4
1 τ
To evaluate C in other ways we may take in (*) V = , V = or V =
2 2
(1 + τ)/2. From the table,
! !
1 1+τ
V1 , q = V2 V1 , q = q−1/4 V3
2 2
! !
1 1+τ
V2 , q = −V1 = 0 V2 , q = −iq−1/4 V4
2 2
! !
1 1+τ
V4 , q = V3 V4 , q = q−1/4 V2
2 2

So again from the left side of (*), 97

′
V1 2 V12 V32
C= ×0−
V22 V42 V22
π2 V12 V34
=− = −π2 V34
π2 V22 V32 V42
11. Lecture 77

Also
′   ′
V1 2  V42  V1 2 V22
C = 2 − 2  − 2 2
V2 V3 V4 V3
′ ′
π2 V1 2 V44 π2 V1 2 V24
=− −
π2 V22 V32 V42 π2 V22 V32 V42
= −π2 V44 − π2 V24

From these we get an identity which is particularly striking:

V34 = V24 + V44 (1)

We have also
V4′′ V2′′
π2 V34 = − (2)
V4 V2
Now let us look at (1) and do a little computing. Explicitly (1) states:
 ∞   4  ∞ 4
 X n2 4  1/4 X
∞
n(n+1)
  X n n2 
 q  = q q  +  (−) q  (3)
n=−∞ n=−∞ n=−∞

This is an identity of some interest.

Let us look for qN on both sides. The left side gives N in the form N = 98
n + n22 + n23 + n24 , that is, as the sum of four squares. So dies the second term
2

on the right. If N is even, it is trivial that both sides are in agreement because
the first term on the right gives only odd powers of q, and the coefficient of qn
in the second term on the right is
X
(−)n1 +n2 +n3 +n4
n21 n22 +n23 +n24 =N

Since N is even either all ni ’s are odd, or two of them odd, or none. It is not
transperent. What happens when N is odd.
Take the more interesting formula (2):
V4′′ V2′′
πV34 = −
V4 V2
By the differential equation,
" #
∂2
Vα′′ = Vα (V , q)
∂V 2 V =0
11. Lecture 78
" #
∂
= −4π2 q Vα (V , q)
∂q V =0
!
4 1 ∂V 2 1 ∂V 4
∴ V3 = 4q −
V2 ∂q V4 ∂q
∂ V2
= 4q log
∂q V4
Now 99
∞
Q
(1 − q2n )(1 + q2n )2
V2 n=1
= 2q1/4 ∞
V4 Q
(1 − q2n )(1 − q2n−1 )
n=1
Q∞
(1 − q2n )2 (1 + q2n )2
n=1
= 2q1/4 ∞
Q
(1 − q2n )2 (1 − q2n−1 )2
n=1
Q∞
(1 − q4n )2
n=1
= 2q1/4 ∞
Q
(1 − qn )2
n=1
1
2q /4
= Q
(1 − qn )2
4∤n

Taking the logarithmic derivative,

 
 n−1 

 1 X −nq 

V34 = 4q 
 
− 2 n 
 4q


4∤n
1 − q 


X nqn
=1+8
4∤n
1 − qn
X X∞
=1+8 n qnk
4∤n k=1
X X
m
=1+8 q n
4∤m n|m
m=1
X∞
=1+8 σ∗ (m)qm
m=1
11. Lecture 79
X
with the previous that σ∗ (m) = , d , that is the divisor sum with those 100
4∤α
d/m
divisors omitted which are divisible by 4. This is an interesting identity:
 ∞
 X n2 4
 ∞
X
 q  = 1 + 8 σ∗ (m)qm (4)
n=−∞ m=1
2 2 2 2
On the left qm can be obtained only as qn1 +n2 +n3 +n4 , so that the coefficient
of qm on the right is the number of ways in which this representation for m is
possible; m is as often the sum of four squares as 8σ∗ (m). Clearly σ∗ (m) , 0,
since among the admissible divisors, 1 is always present. So σ∗ (m) ≥ 1, or
every m does admit at least one such representation. We have thus proved
Lagrange’s theorem: Every integer is the sum of at most four squares.
If m is odd, σ∗ (m) = σ(m); if m is even,
X X
σ∗ (m) = d+2 d
d|m, d odd d|m, d odd
X
=3 d
d|m, d odd

If we denote by r4 (m) the number of representations of m as the sum of four

squares, then

r4 (m) = 8 times the sum of odd divisors of m, m odd;

24 times the sum of odd divisors of m, m even.

We have not partitions this time, but representation as the sum of squares.
We agree to consider as distinct these representations in which the order of the
components has been changed. In partitions we abstracted from the order of 101
the summands; here we pay attention to order, and also to the sign (i.e., one
representation n21 + n22 + n23 + n24 is actually counted, order apart, as 16 different
representations (±n1 )2 +(±n2 )2 +(±n3 )2 +(±n4 )2 , if n1 , n2 , n3 , n4 are all different
from 0).
As an example, take m = 10. The different representations as the sum of
four squares are

(±1)2 + (±1)2 + (±2)2 + (±2)2 ,

(±1)2 + (±3)2 + (0)2 + (0)2 ,

along with their rearrangements, six in each. Thus altogether

r4 (10) = 6 × 16 + 6 × 8 = 144
11. Lecture 80

8σ∗ (10) = 3(1 + 2 + 5 + 10) = 8 × 18 = 144

Lagrange’s theorem was first enunciated by Fermat in the seventeenth cen-

tury. Many mathematicians tried to solve it without success; eventually Jacobi
found out the identity
r4 (m) = 8σ∗ (m)
Before that, the fact that every integer is the sum of four squares was con-
jectured by Fermat, Euler did not succeed in proving it. It was proved by La-
grange, and later Euler gave a mere elementary proof. Euler proved that if two
numbers are each the sum of four squares, then so is their product, by means
of the identity:

(x21 + x22 + x23 + x24 )(y21 + y22 + y23 + y24 )

= (x1 y1 + x2 y2 + x3 y3 + x4 y4 )2 + (x1 y2 − x2 y1 + x3 y4 − x4 y3 )2 +
+ (x1 y3 − x3 y1 + x4 y2 − x2 y4 )2 + (x1 y4 − x4 y1 + x2 y3 − x3 y2 )2 .

We do not proceed to discuss in detail the representability of a number as 102

the sum of two sequences.
If we return not to fα2 but to fα we are not helpless to deal with them. f4 is
not doubly periodic in the fundamental parallelogram, but is doubly periodic in
a parallelogram of twice this size with vertices at 0, 2, 2 + τ, τ. It has got a pole
at the vertex 0 and another at the vertex 1, with residues adding up to zero.

0 1 2

We may write down another identity:

  .   . 
′ V τ
V1′ V4 (V /τ) 1   V1 2 2

 V1′ V 2+1 2τ  


· =   .  −  .  
V4 V1 (V /τ) 2   V1
 V τ V +1
V1 2 2  τ 

2 2

This may be deduced by checking that the poles on both sides are the same,
Further they are odd functions and so the constant term in the difference must
1
vanish. Put V = on both sides.
2
11. Lecture 81

Then we get   .   .  103

1


 V1′ 41 τ2 V1′ 43 τ2  

πV32 = 
 
 .  −  .  
2  V1

 1 τ 3
V1 4 2  τ 

4 2

By straightforward calculation, taking logarithmic derivatives, we obtain,

∞
X
V32 = 4 qm (σ(1) (3)
◦ (m) − σ◦ (m)),
m=1

where the notation employed is:

X
σk (m) = dk ,
d|m
X
σ◦ (m) = d◦ = number of divisors of m;
d|m
X
σ(◦j) (m) = d◦
d|m, d≡ j (mod 4)

comparing coefficients of qm , and observing that on the left m occurs only in

the form n21 + n22 , we get the beautiful theorem:
m can be represented as the sum of two squares as often
as 4(σ(1) (3)
◦ (m) − σ◦ (m)).

Notice that σ(1) (3) (1)

◦ (m) − σ◦ (m) is always non negatives; hence σ◦ (m) ≥
σ(3)
◦ (m) (i.e., the number of divisors of the form 4r + 1 is never less than the
number of divisors of the form 4r + 1), which is by no means a trivial fact.
In some cases we can actually find out what the difference σ(1) (3)
◦ (m)−σ◦ (m) 104
will be. Suppose that m is a prime p. Then the only divisors are 1 and p.
The divisor 1 goes into σ(1) (3)
◦ ; and p goes into σ◦ if p ≡ 3 (mod 4). So the
difference is zero. However, if p ≡ 1 (mod 4). p goes into σ(1) ◦ . Hence the
number of representations of a prime p ≡ 1 (mod 4) as the sum of two squares
is 4 × 2 = 8. That the number of representations of a prime p ≡ 1 (mod 4) as
the sum of two squares is 8 is a famous theorem of Fermat, proved for the first
time by Euler. It is usually proved by using the Gaussian complex numbers.
So far we have been looking upon V as the variable in the V - functions;
now we proceed to consider q as the variable and go to deeper things like the
Jacobi transformation.
Lecture 12

We now come to a rather important topic, the transformation of V -functions. 105

So far we have been looking upon Vα (V /τ) as a function of V only; hereafter
we shall be interfering with the ‘period’ τ also. We want to study how Vα (V /τ)
changes when V is replaced by V + 1/τ. For this it is enough if we replace V
by V τ = ω and see how the function behaves when ω is changed to ω + 1. This
would amount to turning the whole plane around in the positive sense about
the origin through arg τ. We take V1 , because it is easier to handle, since the
zeros become the periods too. Consider

f (V ) = V1 (V τ/τ)

Then

f (V + 1) = V1 ((V + 1)τ/τ)
= V1 (V τ + τ/τ)
= −e−πiτ e−2πiV τ V1 (V τ/τ)
= e−πiτ e−2πiV τ f (V )

τ Similarly consider f (V − 1/τ) (We choose to take − 1τ rather than 1

τ since
we want the imaginary part of the parameter to be positive:
1 τ̄ 1
Im = Im < 0 and so Im − > 0)
τ ττ̄ τ
1 1
f (ν − ) = V1 ((ν − )τ/τ)
τ τ
= V1 (V τ − 1/τ)
= −V1 (V τ/τ)

82
12. Lecture 83

= − f (V )

So f is a sort of V -function which picks up simple factors for the ‘periods’ 1 106
and − 1τ . f (V ) has clearly zeros at 0 and τ′ = − τ1 , or generally at V = m1 +m2 τ′ ;
m1 , m2 integers, which is a point-lattice similar to the old one turned around.
Similarly let us define
!
′ 1
g(V ) = V1 (V /τ ) = V1 V / −
τ
g(V + 1) = V1 (V + 1/τ′ )
= −V1 (V /τ′ )
= −g(V )
!
1
g V − = g(V + τ′ )
τ
= V1 (V + τ′ /τ′ )
′
= −e−πiτ e−2πiV V1 (V /τ′ )
= −eπi/τ e−2πiV g(V )

Let us form the quotient: 107

f (V )
Φ(V ) =
g(V )
f (V + 1)
φ(V + 1) =
g(V + 1)
= −e−πiτ e−2πiV τ φ(V )
!
1 f (V + τ′ )
Φ V − =
τ g(V + τ′ )
f (V )
= πi/τ −2πiV
e e g(V )
= e−πi/τ e2πiV Φ(V )

Φ takes on simple factors in both cases of this peculiar sort that we can
eliminate them both at one stroke. We write

e−πiτ(2V +1) Φ(V ) = Φ(V + 1),

eπi(2V −1/τ) Φ(V ) = Φ(V − 1/τ)
12. Lecture 84

Let us try the following trick. Let us supplement Φ(V ) by an outside func-
tion h(V ) so that the combined function Φ(V ) is totally doubly periodic. Write

Ψ(V ) = Φ(V )eh(V )

We want to choose h(V ) in such a manner that 108

Ψ(V + 1) = Ψ(V + τ′ ) = Ψ(V )

This implies two equations:

e−πiτ(2V +1) eh(V +1)−h(V ) = 1

eπi(2V −1τ) eh(V −1/τ)−h(V ) = 1;

or

h(V + 1) − h(V ) = πiτ(2V + 1) + 2πim

h(V + τ′ − h(V ) = −πi(2V + τ′ ) + 2πim′ .

We can solve both at one stroke. Since on the right side we have a linear
function of V in both cases, a quadratic polynomial will do what we want.

(V + δ)2 − V 2 = 2V δ + δ2 = δ(2V + δ),

and taking h(V ) = πiτV 2 ,

h(V + 1) − h(V ) = πiτ(2V + 1)

h(V + τ ) − h(V ) = πiττ′ (V + τ′ ) = −πi(2V + τ′ ),
′

so that both the equations are satisfied. Putting it in, we have

2 V1 (V τ/τ)
Ψ(V ) = eπiτV  
V1 V / − 1τ

This has the property that 109

ψ(V + 1) = ψ(V + τ) = ψ(V )

So we have double periodicity. This function is also an entire function

because the numerator and denominator have the same simple zeros. So this is
a pole-free function and hence a constant C, C a constant with respect to the
variable V , but may be a function of the parameter τ, C = C(τ). We thus have
12. Lecture 85

I. !
2 1
eπiτV V1 (V τ/τ) = C(τ)V1 V / −
τ
What we need now are the corresponding formulas for the other functions.
Replacing V by V + 12 ,
! ! !
πiτ(V + 21 )2 1 1. 1
e V1 V + τ/τ = C(τ)V1 V + − ,
2 2 τ
!
πiτ(V 2 +V +1/4) −πiτ/4 −πiV τ 1
or e ie e V4 (V τ/τ) = C(τ)V2 V / −
τ
We notices here that two different V -functions are related. This gives
II. !
πiτV 2 1
ie V4 (V τ/τ) = C(τ)V2 V / − .
τ
Replacing in I V by V + τ′ /2 = V − 1/(2τ), we get
III. !
πiτV 2 1
ie V2 (V τ/τ) = C(τ)V4 V / −
τ
1
Finally putting V + 2 for V In, III,
IV. !
2 1
ieπiτV V3 (V τ/τ) = C(τ)V3 V / −
τ
The way the functions change over in I-IV is quite plausible. For consider 110
the location of the zeros.
When we take the
parallelogram and
turn it around what
was originally a
zero for V4 becomes
one for V2 and vice 3
4
versa; and what
used to be in the
middle, the zero of 1 2
V3 , Remains in the
middle. So the for-
mulae are plausible
in structure.
The most important thing now is, what is C(τ)? To evaluate C(τ) let us
differentiate I and put V = 0. We have
12. Lecture 86

V. !
1
τV11 (0/τ) = C(τ)V11 0/ −
τ
From II, III, and IV, putting V = 0,
!
1
iV4 (0/τ) = C(τ)V2 0/ −
τ
!
1
iV2 (0/τ) = C(τ)V4 0/ −
τ
!
1
iV3 (0/τ) = C(τ)V3 0/ −
τ

Multiplying these together and recalling that πV1′ = V2 V3 V4 , we obtain

VI. !
1 3 ′ 1
−iV1 (0/τ) = (C(τ)) V1 0/ − .
τ
Dividing by VI, by V, 111

1
= C 2 (τ),
iτ
r
1
or C(τ) = ±
iτ
C(τ)
In II, III, IV, it is i that appears; so let us write this is
r r
C(τ) 1 1 i
=± =±
i i iτ τ
C(τ)
Now k(i/τ) > 0 · i is completely determined, analytically, in particular
by IV:
P πiτn2
C(τ) V3 (0/τ) e
= 1
= P πiτ′ n2
i V3 (0/ − τ ) e
Both the numerator and denominator are analytic functions if Im τ > 0. So
C(τ)
i is analytic and therefore continuous. i/τ must lie in the right half-plane,
q
and thus τi in either of the sectors with central angle π/2, but because of
continuity it cannot lie on the border lines. So it is in the interior of entirely
one sector. To decide which one it is enough if we make one choice.
12. Lecture 87

Take τ = it, t > 0; then 112

P −πtn2
C(it) e
= P −(π/t)n2
i e

Both numerator and denominator are positive. So C(τ)

i lies in the right half.
q q
i π i
So |arg τ | < 4 and τ denotes the principal branch. The last equality gives:

∞ r ∞
X
−πin2 /τ τ X πin2 τ
e = e
n=−∞
i n=−∞

This is a very remarkable formula. It gives a functional relation: the trans-

formation τ → −1/τ almost leaves the function unchanged; it changes only by
a simple algebraic function. This is one of the achievements
√ of Jacobi.
In the earlier equations we can now put C(τ) = (i/τ). In particular envis-
age V1′ :
 r 3 !
 i  ′ 1
V1′ (0/τ) =   V 0/ −
τ 1 τ
! r
1 τ τ ′
or V1′ 0/ − = · V (o/τ)
τ i i 1
But
∞
Y
V1′ (0/τ) = 2πeπiτ/4 (1 − e2πimτ )3
m=1
∞ r ∞
Y τ τ πiτ/4 Y
∴ e−πiτ/4 (1 − e−2πim/τ )3 = · e (1 − e2πimτ )3
m=1
i i m=1
12. Lecture 88

Extracting cube roots on both sides, 113

∞ r ∞
−πiτ/12
Y
−2πim/τ τ πiτ/12 Y
e (1 − e )=ǫ e (1 − e2πimτ )
m=1
i m=1

where ǫ 3 = 1. Dedekind first introduced the function

∞
Y
η(τ) = eπiτ/12 (1 − e2πimτ )
m=1

Then ! r
1 τ
η − =ǫ η(τ)
τ i
3
This . pis challenging; we have to decide which ǫ to take: ǫ = 1. The quotient
η(− 1τ ) τi η(τ) is an analytic (hence contains) function in the upper half-plane
and so must be situated in on of the three open sectors. Now make a special
choice; put τ = i. Then η(i) = ǫ(+1)η(i), or ǫ = 1.
! r
1 τ
∴ η − = η(τ)
τ i
What we have done by considering the lattice of periods can be done in 114
more sophisticated ways. One can have a whole general theory of the transfor-
aτ + b
mations from 1, τ, to 1, . The quotients appear first and can be carried
cτ + d
over. We start with V1 and come back to it; there may be difficulty, however in
deciding the sign.
Lecture 13

We arrived at the following result last time: 115

! r
1 τ
η − = η(τ).
τ i
We began by investigating a transformation of V1 (V , τ). Instead of looking
upon 1 and τ as generators of the period lattice, we looked upon τ and −1 as
generators (turning the plane around through arg τ): 1, τ → τ, −1. We have
of course still the same parallelogram of periods. Since we should like to keep
the first period 1, we reduced everything by τ : τ, −1 → 1, − τ1 ; so we had to
investigate V1 (V τ/τ). V1 (V τ/τ) and V1 (V / − 1τ ) have the same parallelogram
of periods.
We could do this a little more generally. Let us introduce linear combina-
tions:
ω1 = cτ + d, ω2 = aτ + b,
and go from ω1 to ω2 in the
positive sense. In order that we
must have these also as generat-
ing vectors for the same lattice,
we should have a, b, c, d inte-
gers with
 
a b
  = 1.
 
c d
Moreover we want the first period to be always 1. (This is the difference
between our case and the Weierstrassian introduction of periods, where we
have complete homogeneity). So replacing by linearity, the periods are 1 and
τ′ = aτ+b
cτ+d .

89
13. Lecture 90

Be sure that we want to go from 1 to τ′ through an angle less than π in the 116
positive sense. For this we want τ′ to have a positive imaginary parts Imτ̄′ > 0,
or
τ′ − τ̄′
>0
i !
1 aτ + b aτ̄ + b
i.e., − >0
i cτ̄ + d cτ̄ + d
1 adτ + bcτ̄ − ad τ̄ − bcτ
i.e., >0
i |cτ + d|2
1 (ad − bc)(τ − τ̄)
i.e., >0
i |cτ + d|2
or since τ − τ̄ is purely imaginary,
ad − bc = ±1.
We could do the same thing in all our different steps. The most important
step, however, cannot be carried through, because we get lost at an important
point; and rightly so, it becomes cumber some because a number-theoretic
problem is involved there. Let us see what we have done. Compare
. aτ + b !
V1 ((cτ + d)V /τ) and V1 V
cτ + d
We want periods 1, τ′ ; indeed all things obtainable from ω1 = cτ + d and
ω2 = aτ + b; or m1 ω1 + m2 ω2 must in their totality comprise all periods. For
the first cτ + d is indeed a period, and for the second aτ + b.
Now define 117
f (V ) = V1 ((cτ + d)V /τ)
f (V + 1) is essentially f (V ):
f (V + 1) = V1 ((cτ + d)V + cτ + d/τ)
2
= (−)c+d e−c πiτ −∗∗∗∗∗(cτ+d)V
e V1 ((cτ + d)V /τ), from the table,
= (· · · · · · ) f (V )
′
f (V + τ ) = V1 ((cτ + d)ν + aτ + b/τ)
2
= (−)a+b e−a πiτ −2πia(cτ+d)ν
e V1 ((cτ + d)V /τ)
= (· · · · · · ) f (V )
′
f (V ) has, leaving
 trivial
.  factors aside, periods 1, τ *****. So too for the
aτ+b
second function V1 V cτ+d .
13. Lecture 91

We can form quotients and proceed as we did earlier.

Let us consider for a moment the V ′ s with double subscripts. This is a
digression, but teaches us a good deal about how to work with V -functions.
Recall that
X 1 2
(−)νn eπiτ(n+ 2 ) e2πiν(n+ 2 )
1
Vµν (ν/τ) =
n
V1 (ν/τ) = Vn (ν/τ)
V2 (ν/τ) = V10 (ν/τ)
V3 (ν/τ) = V00 (ντ)
V4 (ν/τ) = V01 (ν/τ)

We take one liberty from now on. Take µ, ν to be arbitrary integers, no 118
longer 0, 1. That will not do very much harm either. In fact,

Vµ,ν+2 (ν/τ) = Vµ,ν (v/τ)

It is unfortunately not quite so easy for the other one:

Vµ+2,ν (v/τ) = (−)ν Vµ,ν (v/τ)

For
X 2
Vµ+2,ν (v/τ) = (−)ν eπiτ(n+µ/2) e2πiν(n+1+µ/2)
n
X 2
= (−)ν (−)v(n+1) eπiτ(n+1+µ/2) e2πiv(n+1+µ/2)
n
= (−)v Vµν (v/τ),

on shifting the summation index from n to n + 1. The original table will be

considerably reduced now; only in place of ν + 1, ν + 12 , ν + τ2 , ν + 1+τ 2 it
will be now necessary! to have the combination ν + 2k + 2l τ. The expression
k l
for Vµν ν + + τ/τ will include everything that we have done so far in one
2 2
single formula.
! X
k l µ 2 µ µ
Vµν ν + + τ/τ = (−)νn eπiτ(n+ 2 ) e2πiν(n+ 2 ) eπi(k+lτ)(n+ 2 )
2 2 n
X n 2 2
=i kµ
(−)(ν+k) eπiτ(n+µ/2+l/2) e−πiτl /4 e2πiν(n+µ/2+l/2) e−πilν
n
kµ −πiτl2 /4 −πilν
=i e e Vµ+l,ν+k (ν/τ) (*)
13. Lecture 92

119
This one formula has the whole table in it.
We now turn to our purpose, viz. To consider the quotient
V1 ((cτ + d)ν/τ)
 . 
V1 ν aτ+b
cτ+d

We wish to discuss the behaviour a little more explicitly of f (ν).

f (v) = V11 ((cτ + d)ν/τ)

f (v + 1) = V11 ((cτ + d)ν + cτ + d/τ)
f (v + τ′ ) = V11 ((cτ + d)ν + aτ + b/τ)

putting k = 2c, l = 2d, µ = ν = 1 in (*), 120

2
f (ν + 1) = (−)d e−πiτc e−2πic(cτ+d)ν V1+2c,1+2d ((cτ + d)ν/τ)
2
= (−)c+d e−πiτc e−πic(cτ+d)ν f (v)

Similarly, putting k = 2a, l = 2b, µ = ν = 1,

2
f (ν + τ′ ) = (−)a+b e−πiτa e−2πia(cτ+d)ν f (v).

Also defining g(ν):

. aτ + b !
V1 v = g(v) = V11 (ν/τ′ ),
cτ + d
we have
g(ν + 1) = V11 (ν + 1/τ′ ) = −V11 (ν/τ′ ).
And putting k = 0, l = 2, µ = 3, ν = 1 in (*),
′
g(ν + τ′ ) = e−πiτ e−2πiν V31 (ν/τ′ )
′
= −e−πiτ e−2πiν g(ν).

We form now in complete analogy with the old procedure

f (ν)
Φ(ν) =
g(ν)
2
Φ(ν + 1) = (−)c+d+1e−πiτc e−2πic(cτ+d)νΦ(ν) ,
2 ′
Φ(ν + τ′ ) = (−)a+b+1 e−πiτc e−2πia(cτ+d)ν eπiτ +2πiv Φ(ν)
13. Lecture 93

Φ takes up exponential factors which contain ν linearly. As before ewe can 121
submerge this under a general form. Define

Ψ(ν) = Φ(ν)eh(ν) ,

where h(ν) is to be so determined that

Ψ(ν + 1) = Ψ(ν + τ′ ) = Ψ(ν)

we therefore want
2
eh(ν+1)−h(ν) (−)c+d+1 e−c πiτ−2πic(cτ+d)ν = 1,
′ 2
πiτ+πiτ′ +2πiν −2πia(cτ+d)ν
eh(ν+τ )−h(ν) (−)a+b+1 e−a e = 1.

It will be convenient to observe that c + d + cd + 1 = (c + 1)(d + 1) is even,

for at least one of c, d should be odd as otherwise c, d would not be co-prime
and we would not have  
a b
  = 1
 
c d
So (−)c+d+1 = (−)cd = eπicd . h is given by the equations:

h(ν + 1) − h(ν) = 2πic(cτ + d)ν + πic(cτ + d),

h(ν + τ′ ) − h(ν) = 2πia(cτ + d) + πia(aτ + b) − πiτ′
= 2πc(aτ + b)ν + πicτ′ (aτ + b).

We have to introduce a suitable function h(ν). Since the difference equation

can be solved by means of a second degree polynomial, put

h(ν) = Aν2 + B

for each separately and see whether it works for both. 122

h(ν + δ) − h(ν) = 2Aνδ + Aδ2 + Bδ

= δ(2Aν + Aδ + B)

Putting δ = 1, τ′ , we find that A = πic(cτ + d) works in both cases. Also

for δ = 1,

A + B = πic(cτ + d),
!2 !
aτ + b aτ + b πic(aτ + b)2
A +B =
cτ + d cτ + d cτ + d
13. Lecture 94

So B = 0 fits both. Hence

h(ν) = Aν2 , A = πi(cτ + d)c

2 f (ν)
∴ Ψ(ν) = eπic(cτ+d)ν
g(ν)
And this is a doubly periodic entire function (because the numerator and
denominator have the same simple zeros) and therefore a constant. We thus
have the transformation formula
. aτ + b ! 2
V11 ν = Ceπic(cτ+d)ν V11 ((cτ + d)ν/τ)
cτ + d

where C may depend on the parameters τ, a, b, c, d:

C = C(τ; a, b, c, d)

More generally we can have a parallel formula for any µ, ν. As before we 123
get an equation for C 2 . And there the thing stops. Formerly we were in a very
good position with the special matrix
   
a b 0 −1
  =   .
   
c d 1 0

For general a, b, c, d we get into trouble.

Lecture 14

We were considering the behaviour of V11 (ν/τ) under the general modular 124
transformation:
. aτ + b ! 2
V11 ν = C(τ)eπiC(cτ+d)ν V11 ((cτ + d)ν/τ), (1)
cτ + d
 
 
a b
a, b, c, d integers with   = +1.
 
c d
We want to determine C(τ) as far as possible. We shall do this up to a ±
sign. ν is unimportant at the moment; even if we put ν = 0, C(τ) survives. Put
′ ′
ν = 21 , τ2 , 1+τ
2 in succession, ans use out auxiliary formula which contracted
the whole table into one thing:
!
k lτ . 2
Vµν v + + τ = ikµ e−πiτl /4 e−πiν Vµ+l,v+k (v/τ) (*)
2 2
1 aτ + b
Putting ν = 2 in (1), and writing τ′ = ,
cτ + d
! !
1. ′ πic(cτ+d)/4 cτ + d .
V11 τ = C(τ)e V11 τ (2)
2 2
This is the right moment to call for formula (*). From (*) with ν = 0,
µ = ν = 1, k = 1, l = 0, we get
!
1. ′
V11 τ = iV12 (0/τ′ )
2
Also from (*) with ν = 0, µ = ν = 1, k = d, l = C, we get
!
cτ + d . 2
V11 τ = id C −πic /4 V1+c,1+d (o/τ).
2

95
14. Lecture 96

Substituting these two formulas in the left and right sides of (2) respec- 125
tively, we get
2
iV12 (0/τ′ ) = C(τ)eπic(cτ+d)/4 id e−πiτc /4
V1+c,1+d (0/τ)

Now, recalling that

Vµ,ν+2 (ν/τ) = Vµν (ν/τ)

(**)
Vµ+2,ν (ν/τ) = (−)ν Vµν (ν/τ),

the last formula becomes

iV10 (0/τ′ ) = C(τ)eπicd/4 id V1+c,1+d (0, τ) (3)

Putting ν = τ′ /2 in (1), we have

! !
τ′ . ′ πic/4τ′ (aτ+b) aτ + b .
V11 τ = C(τ)e V11 τ .
2 2

Making use of (*) in succession on the left and right sides (with proper
choice of indices) as we did before, this gives
′ ′ 2
e−πiτ /4 V12 (0/τ′ ) = C(τ)eπicτ (aτ+b)/4i be −πia τ/4
V1+a,1+b (0/τ),

and this, after slight simplification of the exponents on the right sides, gives in
view of (**),
−V01 (0/τ′ ) = C(τ)ib eπiab/4 V1+a,1+b (0/τ) (4)
Putting ν = (1 + τ′ )/2 in (1), 126
! !
1 + τ′ . ′ πic/4(1+τ′ )(la+c)τ+b+d) (a + c)τ + l + d .
V11 τ = C(τ)e V11 τ
2 2

Again using (*) and (**) as we did earlier, this gives

′ ′ 2 (0/τ)
ie−πiτ /4 V22 (0/τ′ ) = C(τ)eπic/4(1+τ )((a+c)τ+l+d) il+d e−πi(0+c) /4
V1+a+c,1+l+d

This of course can be embellished a little:

′ (0/τ)2
iV00 (0/τ′ ) = C(τ)eπi/4(1+τ )(c(a+c)τ+cb+id+1) ib+d e−πi/4 e−πiτ(a+c) /4 V1+a+c,1+b+d
2
= C(τ)eπi/4(a+c)((a+c)τ+b+d) il+d e−πi/4 e−πiτ(a+c) /4 V1+a+c,1+b+d
(0/τ)
∴ iV00 (0/τ′ ) = C(τ)eπi/4(a+c)(b+d) ib+d e−πi/4 V1+a+c,1+b+d (5)
14. Lecture 97

Now utilise the formula:

V1′ (0/τ) = πV2 (0/τ)V3(0/τ)V4 (0/τ)

Multiplying (3), (4) and (5),

V11′ (0/τ′ ) = (C(τ))3 (−)b+d eπi/4(ab+cd+(a+b)(b+d)−1)

× iπV1+c,1+d (0/τ)V1+a,1++b(0/τ)V1+a+c,1+b+d (0/τ)

Observe that the sum of the first subscripts on the right side = 3+2a+2c ≡ 1 127
(mod 2). So either all three numbers 1 + a, 1 + c, 1 + a + c are odd, or one of
them is odd and two even. Then first case is impossible since we should then
have both a and c even and so  ac db  , 1. So two of them are even and one
odd. The even suffixes can be reduced to zero and the odd one to 1 by repeated
application of (**). Similarly for the second suffixes. So the V -factors on the
right will be V00 , V01 , V10 . What we hate is the combination 1, 1 and this does
not occur. (If it did occur we should have V11 which vanishes at the origin).
Although we can not identify the V -factors on the right, we are sure that we
get exactly the combinations that are desirable: 01, 10, 00. The dangerous
combination is just out.
Let us reduce the subscripts by stages to 0 or 1 as the case may be. When
we reduce the second subscript nothing happens, whereas when we reduce
index by steps of 2, each time a factor ±1 is introduced, by virtue of (**). By
1+c
the time the subscript 1 + c is reduced to 0 or 1, a factor (−)[ 2 ] (1 + d) will
have accumulated in the case of V1+c,1+d . Similarly in the case of V1+a,1+b and
V1+a+c,1+b+d . Altogether therefore we have a factor 128

1+c
(−)[ 2 ](1+d)+[ 1+a 1+a+c
2 ](1+b)+[ 2 ](1+b+d) ,

and when this compensating factor is introduced we can write V00 , V11′ and V10 .
Hence our formula becomes

V11′ (0/τ) = (C(τ))3 (−)α eπi/4(ab+cd+(a+c)(b+d)−1) iπV00 (0/τ)V01 (0/τ)V10(0/τ) (6)

where
# " " # " #
1+c 1+a 1+a+c
α=b+d+ (1 + d) + (1 + b) + (1 + b + d)
2 2 2

From (1), differentiating and putting ν = 0, we have

V11′ (0/τ′ ) = C(τ)(Cτ + d)V11′ (0/τ) (7)

14. Lecture 98

Dividing (6) by (7)

(C(τ))2 = (cτ + d)(−)α e−πi/4(ab+cd+(a+c)(b+d)−1)

cτ + d α −πi/4(ab+cd+(a+c)(b+d)−3)
= (−) e
i
(we may assume c > 0, since c = 0 implies ad = 1 or a, d = ±1, which give
just translations).
r
cτ + d α −πi/8(ab+cd+(a+c)(b+d)1−3)
∴ C(τ) = ± i e
i
For the square root we take the principal branch. Since Im(cτ + d) > 0,
cτ + d cτ + d
R > 0, so that is a point in the right half-plane. The sign is still
i i
uncertain.
The factor e−πi/8(··· ) looks like a 16th root of unity, but is really not so. Since 129
ad + bc has the same parity as ad − bc = 1, the exponent is even, and therefore
what we really have is only an 8th root of unity.
What could we do now? We really do not know of any fruitful way. We
cannot copy what we did formerly. There we had a very special case: τ′ =



−1/τ, or the modular substitution involved was ac bd = 10 −1 0 . The ± sign
depends only on a, b, c, d, not on τ, so that it is enough if we make a special
choice of τ in the equation. Formerly we could take τ = τ′ = i and it worked
so beautifully because τ is a study the fixed points of the transformation τ′ =
aτ + b
. The fixed points ξ are given by
cτ + d
aξ + b
ξ= ,
cξ + d
or cξ2 + (d − a)ξ − b = 0
p
(a − d) ± (a − d2 ) + 4bc
i.e., ξ=
2c
p
(a − d) ± (a + d)2 − 4
=
2
since ad − bc = 1.
Hence we have several possibilities. If the square root is imaginary we have
twp points one in each of the upper and√ lower√half-planes, and for this |a + d| < 130
2, so that the square root becomes −4 or −3 according as |a + d| = 0or 1.
This is the elliptic case. If |a + d| = 2, we have one rational fixed point; this is
the parabolic case. And in the huge infinity of cases, |a + d| > 2, we have two
14. Lecture 99

real fixed points -the hyperbolic case. Here the fixed are not accessible to us
because they are quadratic algebraic numbers on the real axis.
In the elliptic case with |a + d| < 2 we could finish the thing without much
trouble. In the parabolic case we are already in a fix. Much more difficult is
the hyperbolic case.
If ξ1 and ξ2 are the fixed
points, τ and τ′ will lie on the
same circle through ξ1 and ξ2 ,
and repetitions of the transfor-
mation would give a sequence
of points on the same circle
which may converge to either ξ1
or ξ2 . So the ambiguity in the ±
sign will remain.
It will be much more difficult when we pass from V to η, because then we
shall have to determine a cube root.
Lecture 15

We were discussing the possibility !of getting a root of unity determined for 131
. aτ + b
the transformation of V11′ ν . There do exist methods for determining
cτ + d
this explicitly. Only we tried to carry out the analogue with the special case
as far as possible, not with complete success. other methods exist. The first of
these is due to Hermite, done nearly 100 years ago. He used what are called
Gaussian sums. There are difficulties there too and we want to avoid them.
Another method is that of Dedekind using Dedekind sums.
In the special case of the transformation from τ to τ′ = − 1τ we were faced
 
with an elliptic substitution ac bd . These are of two sorts:
1. a + d = 0
2. a + d = ±1
In both cases we can completely forget about the root of unity if we remember
the following fact. Our formula had the following shape:
r
. aτ + b ! cτ + d cτ + d
′
V11 0 = · ρ(a, b, c, d)V11′ (0/τ) (*)
cτ + d i i
where ρ is a root of unity which is completely free of τ. we can then get things
straightened out. We have only to consider the fixed points of the transforma-
tion given by p
a − d ± (a + d)2 − 4
ξ=
2c
aξ + b
Put ξ on both sides of the formula; since ξ = , both sides look alike 132
cξ + d
′
and V11 does not vanish for appropriate τ in the upper half-plane (we may take
c > 0), so that ρ is given directly by the formula.

100
15. Lecture 101

Case 1. a + d = 0 √
−2d ± −d −d + i
ξ= =
2c c
(reject the negative sign since we want a point in the upper half-plane).
cξ + d i
= =i
i 1
∴ V11′ (0/ξ) = ρ(a, b, c, d)V11′ (0/ξ)

So ρ(a, b, c, d) = 1 and remains 1 in the general formula when we go away

from ξ.
Case 2. a + d = ±1
√
±1 − 2d + −3
ξ=
2c
√
±1 + i 3
∴ cξ + d = = eπi/3 or e2πi/3
2
cξ + d
= eπi/3−πi/2 or e2πi/3−πi/2
i
= e±πi/6
r
cξ + d
= e∓πi/12
i
Putting ξ on both sides of (*), 133

1 = e∓πi/4 ρ(a, b, c, d)

∴ ρ(a, b, c, d) = e±πi/4
when a + d = ±1, (we may take c > 0; the case c = 0 is uninteresting and
if c < 0 we can make it c > 0).
There are unfortunately no more cases like these.
Parabolic case. The analysis here is a little longer but it is worth while working
it out. Now a + d = ±2, and there is only one fixed point
a − d −2d ± 2 −d ± 1 δ
ξ= − = =−
2c 2c c γ
where (γ, δ) = 1 and we may choose γ > 0. The fixed point is now a rational
point on the real axis. We try to approach it. This is a little difficult because
15. Lecture 102

we do not know what the function will do thee. But by an auxiliary transfor-
mation we can throw this point into the point at infinity. Consider the auxiliary
transformation  
Aτ + B A B
T= , =1
γτ + δ  γ δ 

The denominator becomes zero for τ = ξ. Let

Aτ′ + B
T′ =
γτ′ + δ
(notice that γ and δ have got something to do with the properties of two other 134
numbers c, d). Now (*) gives
r 3
′
 γτ + δ 
V11 (0/T ) =    ρ(A, B, γ, δ)V11′ (0/τ),
i 
r 3
′ ′
 γτ′ + δ 
V11 (0/T ) =   ρ(A, B, γ, δ)V11′ (0/τ′ ).
i 

Dividing, we get
 q ′ 3
 γτ +δ 
V11′ (0/T ′ )  i 
′ ′
 V11 (0/τ )
=  q (1)
V11′ (0/T )  γτ+δ  V11′ (0/τ)
 
i

The left side gives the behaviour at infinity. We cannot of course put τ = ξ.
Put τ = ξ + it, t > 0, and later make t → 0. τ is a point in the upper half-plane.
τ − ξ = it,
aτ + b aξ + b
τ′ − ξ = −
cτ + d cξ + d
τ−ξ
=
(cτ + d)(cξ + d)
it
=
1 ± ict
This is also in the upper half-plane. τ′ → ξ as t → 0
Let us calculate T and T ′ . For this consider

Aτ′ + B Aτ + B
T′ − T = −
γτ′ + δ γτ + δ
15. Lecture 103

τ′ − τ
=
(γτ′ + δ)(γτ + δ)
c
=∓
γ2
135
This is quite nice; the difference is a real number.
Aτ + B A(ξ + it) + B
T= =
γτ + δ γ(ξ + it) + δ
Ait + Aξ + B
= , since γξ + δ = 0,
γit
δ
A −A γ + B
= +
γ γit
A Bγ − Aδ
= +
γ γτt
A 1
= + , since Bγ − Aδ = −1
γ γ2 t

→ i∞ (along the ordinate x = Aγ ) as t → 0

A c i
T′ = ± +
γ γ2 r 2 t
→ i∞ as t → 0

Now recall the infinite product formula 136

∞
Y
V11′ (0/T ) = 2πieπiT/4 (1 − e2πinT )3
n=1

A i
Let T= γ + γ2 t
. Then
2
e2πinT = e2πinA/γ e−2πnt/γ → 0
∴ V11′ (0/T ) = 2πieπiT/4 (a factor tending to 1)

We do not know what happens to eπiT/4 . But we need only the quotient. So
V11′ (0/τ) ′
∼ eπi(T −T )/4
V11′ (0/τ)
2
= e∓πic/4γ (2)
15. Lecture 104

Consider similarly the quotient V11′ (0/τ′ )/V11′ (0/τ). We have, since γξ+δ =
0,
γτ + δ γ(ξ + it) + δ
= = γt
i i
r
γτ + δ √
or = νt, where we take the positive square root
i
 
it
γτ′ + δ γ ξ + 1±ict + δ γt
= =
i i 1 ± ict
r r
γτ′ + δ √ 1
∴ = rt (both branches principal)
i 1 ± ict
√
∼ γt as t → 0.
r
γτ′ + δ . γτ+δ
q
Hence the quotient i behaves like 1. 137
i
And so we have what we were after:
V11′ (0/τ′ ) 2
→ e∓πic/4γ as t → 0 (3)
V11′ (0/τ)
a−d
cτ + d c(ξ + it) + d + cit + d
= = 2
i i i
a + d2 + cit 1 + cit
= =±
i i
= ∓i + ct
→ ∓i as t → 0.
What will the square root of this do?
r
cτ + d √
= ct ∓ i, and this
i
does lie in the proper half plane
because ct > 0. For small t
it will be very near the imagi-
nary axis near ∓i. So the square
root lies in the sector, in the
lower half plane if we choose
√
−i = e−πi/4 , and in the√upper
half-plane if r
we choose +i =
cτ + d
eπi/4 Hence → e∓πi/4
i
as t → 0.
15. Lecture 105

Using this fact as well as (2) and (3) in (*) we get 138

2
e∓πic/4γ = e∓3πi/4 ρ(a, b, c, d)
πi
∴ ρ(a, b, c, d) = e∓ (c/γ2 − 3)
4
We observe that the common denominator (γ, δ) = 1 plays a role, however
a and b do not enter.
Hyperbolic case. The thing could also be partly considered in the hyperbolic
case. It will take us into deeper things like real quadratic fields and we do not
propose to do it.
Let us return to what we had achieved in the specific case. We had a formula
for η(τ): r
′ cτ + d
η(τ ) = ǫ(a, b, c, d)η(τ),
i
where ǫ(a, b, c, d) is a 24th root of unity. This shape we have in all circum-
stances. The difficulty is only to compute ǫ. We shall not determine it in
general, and we can do away with it even for the purpose of partitions by using
a method developed recently by Selberg.
However in each specific case we can compute ǫ.
! r
1 τ
η − = η(τ)
τ i

Now
∞
Y
η(τ) = eπiτ/12 (1 − e2πinτ ),
n=1
η(τ + b) = eπib/12 η(τ)

Out of these two facts we can get every other one, because the two substi- 139
tutions    
1 1 0 −1
S =   ,
 T =  

0 1 1 0
form generators of the full modular group. This can be shown as follows. Take
c > 0.

aτ + b = q0 (cτ + d) − a, τ − b1 , c > |a1 |,

15. Lecture 106
 
aτ + d a1 τ + b1  c d 
or = q0 − ,   = 1
cτ + d cτ + d 
a1 b1 

(if a < 0 this step is unnecessary). Similarly

cτ + d a2 τ + b2
= q1 −
a1 τ + b1 a1 τ + b1
We thus get a continued fraction expansion. The partial quotients get simi-
τ+b
lar and simpler every time and end with = τ + qk . so we can go back and
0+1
take linear combinations; all that we have to do is either to add an integer to τ
or take −1/τ.
As an example, let us consider
!
3τ + 4
η
2τ + 3

3τ + 4
Let us break into simpler substitutions,
2τ + 3
3τ + 4 1
τ3 = =1− ,
2τ + 3 τ2
τ2 = −2 + τ1 ;
1
τ1 = −
τ+1
! r
1 τ+1
η(τ1 ) = η − = η(τ + 1)
τ+1 i
r
τ + 1 πi/12
= e η(τ).
i
η(τ2 ) = η(τ1 − 2) = e−πi/6 η(τ1 )
r
τ + 1 −πi/12
= ·e η(τ)
i
! r
1 τ2
η − = η(τ2 )
τ2 i
r r
τ + 1 τ2 −πi/12
= e η(τ)
i i
The two square roots taken separately are each a principal branch, but taken 140
15. Lecture 107

together they may exceed one. We can write this as

s
! r 1
1 τ + 1 −2 − τ+1 −πi/12
η − = e η(τ)
τ2 i i
r s
τ + 1 −2τ − 3 −πi/12
= e η(τ)
i i(τ + 1)
r
−2τ − 3 −πi/12
= e η(τ)
−1
√
= ± 3 + 2τe−πi/12 η(τ)

Here we are faced with a question: which square root are we to take? 141
√ πi/4
q
2τ+3
We write 3 + 2τ = e i
Let us look into each root singly. For τ = it where do they go?
r r
τ+1 it + 1
=
i i
→ ∞ with argument 0 as t → ∞.
s s
−2τ − 3 −2it − 3 √
= → 2i as t → ∞,
i(τ + 1) i(τ + 1)
or its argument = π/4
r r
τ + 1 −2τ − 3
The product has here argument π/4, so that it continues
i i(τ + 1)
to be the principal branch. Of course in a less favourable case, if we had two
other arguments, together they would have run into something which was no
longer a principal branch. Finally,
! r
3τ + 4 πi/4 2τ + 3
η =e η(τ)
2τ + 3 i

and here there is no ambiguity. Actually in every specific case that occurs one
can compute step and make sure what happens.
There does exist a complete formula which determines ǫ(a, b, c, d) explic-
itly by means of Dedekind sums S (h, k).
 Part III

Analytic theory of partitions

108
Lecture 16

Our aim will be now to get an explicit formula for p(n) and things connected 142
with it. Later we shall return to the function η(τ) and the discussion of the sign
of the square root. That will again lead us into some aspects of the theory of
V -functions connected with quadratic residues.
Let us come to our topic. Euler had, as we know, the identity:
∞
X 1
p(n)xn = ∞
.
Q
n=0 (1 − xm )
m=1

This is the starting point of the function-theoretic treatment of p(n).

Z
1 f (x)
p(n) = dx,
2πi c xn+1
∞
Q
where f (x) = (1 − xm )−1 and C is a suitable closed path contained in the
m=1
unit circle, in which the function is analytic, and enclosing the origin. Since
P
p(n)xn is a power series beginning with 1, this means a little more. n may be
negative also; and when n is negative f (x)x−n−1 is regular at x = 0. Therefore
we include negative exponents also in our discussion; we put p(−n) = 0, n > 0,
when is convenient. Hereafter we shall take n to be an integer R 0; we shall
choose n and keep it fixed throughout our discussion.
It is a little more comfortable to change the variable and put x = e2πiτ , 143
Im τ > 0, which is familiar to us.
dx = e2πiτ · 2πidτ and the whole thing boils down to
Z α+1
p(n) = f (e2πiτ )e−2πinτ dτ
α

109
16. Lecture 110

It is enough to take the integral along a path from an arbitrary point α to

the point α + 1, because the integrand is periodic, with period 1. (This path
replaces the original path C that we had in the x-plane before we changed the
variable). The method of Hardy and Ramanujan was to take a curve rather
close to the unit circle which is a natural boundary for the function (this will
come out in the course of the argument). They cut up the path of integration
into pieces called Farey arcs, and the trick was to replace the function by a
simpler approximating function on each specific Farey arc. We shall use not
exactly this method, but consider a special path from α to α + 1, which we shall
discuss.
We shall keep our formula in abeyance for a moment and give a short dis-
cussion of Farey series (‘series’ here is not to be understood in the sense of
infinite series, but as just an aggregate of numbers). Cauchy did make all the 144
observation attributed by Hardy and Wright to Farey; Farey made his remarks
in the Philosophical Magazine, 1816. He put only questions; Cauchy had all
the answers earlier.
We deal with the interval (0, 1). Choose all reduced fractions whose de-
nominators do not exceed 1, 2, 3, · · · in succession. Let us write down the first
few, with the fractions arranged in increasing order of magnitude.
0 1
1 1 order 1
0 1 1
1 2 1 order 2
0 1 1 2 1
1 3 2 3 1 order 3
0 1 1 1 2 3 1
1 4 3 2 3 4 1 order 4
0 1 1 1 2 1 3 2 3 4 1
1 5 4 3 5 2 5 3 4 5 1 order 5
The interesting fact is that we can write down a new in the following way.
h h′
We repeat the old row and introduce some new fractions. If < ′ are adjacent
k k
fractions in a row, the new one introduced between these in the next row is
16. Lecture 111

h + h′
, provided that the denominator is of the proper size. Following Hardy
k + k′
h + h′ h h′ h
and Littlewood we call ′
the ‘mediant’ between and ′
. We have <
k + k k k k
h + h′ h′
< ′ , so that the order is automatically the natural order. We call
k + k′ k
that row which has denominator k ≤ N, the Farey series of order N. We get
this by forming mediants from the preceding row. Farey made the following
observation. Take two adjacent fractions in a row; then the determinant formed
by their numerators and denominators   is equal to −1. For instance, in the fifth
1 2 
1 2
row and are adjacent and   = −1***********. If we now prove that
3 5 
3 5
new fractions are always obtained by using mediants, then we can be sure, by 145
induction, that this determinant is always −1. For, let
 
h h′ 
  = −1; then
k k′ 
 
   
h h + h′  h + h′ h′ 
  = −1 =  
k k + k′  k + k′ k′ 
   

If indeed only mediants occur, Farey’s observation is justified. And this is

so. Observe that these fractions must all appear in their lowest terms; other-
wise, the common factor will show up and the determinant would not be −1.
Suppose that we want to find out where a particular fraction appears. Say, we
H
have in mind a specific fraction . It should occur for the first time in the
K
Farey series of order N = K and it should not be present on any series of order
H
< K. Now look at N = K − 1 where is not present. If we put it in, it will
K
h1 h2
belong somewhere according to its size, i.e., we can find fractions , , with
k1 k2
h1 H h2
k, k2 < N such that < < . Assume that the determinant property and
k1 K k2
the mediant property are true for order N < K. (They are clearly true up to
order 5, as we verify by inspection, so that we can start induction). Now prove
h1
them for N = K. Try to determine H and K by interpolation between and
k1
h2
. Put
k2
Hk1 − Kh1 = λ,
16. Lecture 112

−Hk2 + Kh2 = µ,

so that λ and µ are integers > 0. Solving for H and K by Cramer’s rule, 146
   
λ −h1  λ −k 
  1
   
µ h2  µ k2 
H =   , K =  
h k2  
h2 k2 
 2   
 h1 k1 
h1 k1 

By induction hypothesis, the denominator = 1, and so

H = λh2 + µh1
K = λk2 + µk1
H λh2 + µh1
or = .
K λk2 + µk1
What do we know about K? K did not appear in a series of order K − 1;
k1 and k2 are clearly less than K. What we have found out so far is that any
h1 h2 λh2 + µh1
fraction lying between and can be put in the form . Of these
k1 k2 λk2 + µk1
only one interests us - that one with lowest denominator. This comes after the
ones used so far. Look for the one with lowest denominator; this corresponds
to the smallest possible λ, µ, i.e., λ = µ = 1. Hence first among the many later
H h1 + h2
appearing ones is = , i.e., if in the next Farey series a new fraction
K k1 + k2
is called for, that is produced by a mediant. So what was true for K − 1 is true
for K; and the thing runs on.
One remark is interesting, which was used in the Hardy - Littlewood- Ra-
h1 h2
manujan discussion. In the Farey series of order N, let and be adjacent
k1 k2
h1 h2 h1 + h2
fractions. < · does not being these. It is of higher order. This
k1 k2 k1 + k2
says that k1 + k2 > N. For two adjacent fractions in the Farey series of order N, 147
the sum of the denominators exceeds N. Both k1 and k2 ≤ N, so

2N ≥ k1 + k2 > N.

k1 and k2 are equal only in the first row, otherwise it would ruin the deter-
minant rule. So
2N > k1 + k2 > N, N > 1.
16. Lecture 113

This was very often used in the Hardy - Ramanujan discussion. (The Farey
series is an interesting way to start number theory with. We can derive from it
Euclid’s lemma of decomposition of an integer into primes. This is a concrete
way of doing elementary number theory).
We now come to the special path of integration. For this we use Ford
Circles (L.R. Ford, American Mathematical Monthly, 45 (1938), 568-601).
We describe a series of circles in the upper half-plane. To each proper fraction
h h i 1
we associate a circle Chk with centre τhk = + 2 and radius 2 , so the
k k 2k 2k
circles all touch the real axis.

Take another Ford circle Ch′ k′ , with centre at τh′ k′ . Calculate the distance
between the centres.
!2 !2
2 h h′ 1 1
|τhk − τh′ k′ | = − + − .
k k′ 2k2 2k′2
1 1
The sum of the radii = 2k2
+ 2k′2
148

! !2
2 1 1 h h′ 1
|τhk − τ | −
h′ k′ + = − − 2 ′2
2k2 2k′2 k k′ k h
(hk′ − h′ k)2 − 1
= ≥ 0,
k2 k′2
 
since h, k are coprime and so  h k  is an integer , 0. So two Ford circles never
h′ k′
intersect. And they touch if and only if
 
 h k 
  = ±1,
 ′ ′ 
h k

h h′
i.e., if in a Farey series , ′ have appeared as adjacent fractions.
k k
16. Lecture 114

Now we come to the description of the path of integration from α to α + 1.

For this consider the Ford circle Chk .

h1 h
In a certain Farey series of order N we have adjacent fractions < <
k1 k
h2
. (We know exactly which are adjacent ones in a specific series). Draw also 149
k2
the Ford circles Ch1 k1 and Ch2 k2 . These touch Chk . Take the arc γhk of Chk from
one point of contact to the other in the clockwise sense (the arc chosen is the
one not touching the real axis). This we do for all Farey fractions of a given
order. We call the path belonging to Farey series of order N PN . Let us describe
this in a few cases.
We fix α = i and pass to α + 1 = i + 1. Take N = 1; we have two circles of
i i
radii 2 each with centres at and 1 +
2 2

ρ1 will be the path consisting of arcs from i to 21 + 2i and 21 + 2i to i + 1. Later

because of the periodicity of f (e2πiτ ) we shall replace the second piece by the
16. Lecture 115

0 1
arc from − 12 + 2i to i. Next consider Farey series of order 2; and are no
1 1
longer adjacent. The path now comprises the arc of C01 from i to the point of
contact with C12 , the arc of C12 from this point to the point of contact with C11 150
and the arc of C11 from this point to i + 1. Similarly at the next stage we pass
from i on C01 to i+1 on C11 through the appropriate arcs on the circles C13 , C12 ,
C23 in order. So the old arcs are always retained but get extended and new arcs
spring into being and the path gets longer and longer. At no stage does the path
intersect itself, but these are points of contact. The path is complicated and was
not invented in one sitting. The Farey dissection of Hardy and Ramanujan can
be pictured as composed of segments parallel to the imaginary axis. Here it is
more complicated.
We need a few things for our consideration. We want the point of contact
of Chk and Ch′ k′ . This is easily seen to be the point
1 1 ! !
2k′2 2k2 h i k2 h′ i k′2
τhk + τh′ k′ = + 2 2 ′2
+ ′
+
1
2k2
+ 1
2k′2
1
2k2
+ 1
2k′ 2
k 2k k + k k 2k k + k′2
′2 2
!
h h′ h k′2 i
= + ′ − +
k k k k2 + k′2 k2 + k′2

and this, since 151

 
′
h h h′ h h k′ i
< ′ and   = 1, is = + + 2
k k  ′
k k k k(k 2 + k ′2 ) k + k′ 2
h ′
= + ξhk
k
16. Lecture 116

′ k′ i
where ζhk = + 2 . We notice that the imaginary part 1/(k2 +k′2 )
k(k2 + k ) k + k′2
′2
gets smaller and smaller as k + h′ lies between N and 2N. Each arc runs from
h ′ h ′′
k + ζhk to k + ζhk . Such an arc is the arc νhk . No arc touches the real axis.
We continue our study of the integral. Choose a number N, the order of the
Farey series, and cut the path of integration PN into pieces γhk .
Z
p(n) = f (e2πiτ )e−2πinτ dτ
PN
X Z
= f (e2πiτ )e−2πinτ dτ
(h,k)=1 γ
0≤h<k≤N hk

Now utilise the points of contact: put

h
τ= + ζ;
k
′′
X Zζhk
h h
p(n) = f (e2πi( k +ζ) )e−2πin( k +S ) dS
(h,k)=1 ζ ′
0≤h<k≤N hk

h ′ h
′′
(γhk goes from k + ζhk + ζhk
to k ; these are all arcs of radii 1/2k2 ). We make 152
iz
a further substitution: put ζ = 2 , so that we turn round and have everything
k
in the right half-plane, instead of the upper half-plane. (All these are only
preparatory changes; there is no actual mathematical progress as yet). Then
′′
Zzhk
X e−2πinh/k h iz 2
p(n) = i f (e2πi( k + k2 ) )e2πnz/k dz
(h,k)=1
k2
z′hk
0≤h<k≤N

Now find out z′hk and z′′hk .

k2 + ikk′
z′hk =
k2 + k′2
k2 − ikk′′
z′′hk = 2
k + k′′2
So what we have achieved so for is to cut the integral into pieces. 153
16. Lecture 117

The whole thing lies on the right

half-plane. The original point
of contact is 0 and everything
1 1
lies on the circle |z − | = .
2 2
This is a normalisation. We now
study the complicated function
on each arc separately. We shall ǫ be-
0
find that it is practically the
function η(τ) about which we
know a good deal:
! r
aτ + b cτ + d
η =ǫ η(τ),
cτ + d i
ing a complicated 24th root of unity.
Lecture 17

We continue our discussion of p(n). Last time we obtained 154

′′
Zzhk  
X i −2πinh/k h z 2
p(n) = e f e2πi( k + k2 ) e2πinz/k dz
(h,k)=1
k2
z′hk
0≤h<k≤N

n is a fixed integer here, n R 0 and p(n) = 0 for n < 0; and this will be of some
use later, trivial as it sounds. N is the order of the Farey series. We have to deal
with a complicated integrand and we can foresee that there will be difficulties
as we approach the real axis. However, f is closely related to η:

f (e2πiτ ) = eπiτ/12 (η(τ))−1 ,

1
since f (x) = ∞ ,
Q
(1 − xm )
m=1
∞
Y
η(τ) = eπiτ/12 (1 − e2πimτ ).
m=1

h iz
For us τ = + 2 .
k k
We can now use the modular transformation. We want to make Im τ large so
aτ + b
that we obtain a big negative exponent. So we put τ′ = , a, b, c, d being
′
cτ + d
chosen in such a way for small τ, τ becomes large. This is accomplished by 155
taking kτ − h in the denominator; kτ − h = 0 when z = 0 and close to zero
aτ + b
when z is close to the real axis. We can therefore put τ′ = where a, b
  kτ − h
a b
should be integers such that  k −h  = 1. This gives −ah − bk = 1 or ah ≡ −1

118
17. Lecture 119

(mod k). Take a solution of this congruence, say h′ i.e., choose h′ in such a
way that h′ h ≡ −1 (mod k), which is feasible since (h, k) = 1. As soon as h′ is
found, we can find b. Thus the matrix of the transformation would be
   
a b h′ − hh′ +1 
  =  k 

   
c d k −h

So we have found a suitable τ′ for our purpose.

!
′ h iz hh′ + 1
h + 2 −
k k k
τ′ = !
h iz
k + −h
k k2
iz
h′ − 1
= k
iz
h′ i
= + .
k z
If z is small this is big.
Now recall the transformation formula for η: if c > 0,
! r
aτ + b cτ + i
η =ǫ η(τ)
cτ + d i
In our case 156
′ ′
f (e2πiτ ) = eπiτ /12 (η(τ′ ))−1
!−1/2
′ cτ + d
= eπiτ /12 ǫ −1 (η(τ))−1
i
!−1/2
′ cτ + d
eπiτ /12 ǫ −1 eπiτ/12 f (e2πiτ )
i
And this is what we were after. Since
!
h iz iz
cτ + d = kτ − h = k + 2 −h= ,
k k k
this can be rewritten in the form:
 iz

h
f (e2πiτ ) = f (e2πi k + k2 )
17. Lecture 120
  ′  
  h i  
 
πi  h h′  πi  iz − i  r
k− k
 2πi +  
z z f
 k2 
= ǫe 12 e 12 e k z 
k  

and there is no doubt about the square root - it is the principal branch. We write

πi  h h′ 
 
 − 
= ǫe 12 k k
 
ωhk

So something mathematical has happened after all this long preparation;

and we can make some use of our previous knowledge. We have
z′′
Zhk 
X′ iωhk
−2πinh/k π 1 z
12 z − k2

√  2πi h′ + i   2πnz/k2
p(n) = e e zf e k z e dz
o≤h<k≤N
k5/2
z′hk

P′
where denotes summation over h and k, (h, k) = 1. Now what is the 157
advantage we have got? Realise that
∞
X
f (x) = p(n)xn = 1 + x + · · ·
n=0

So for small x, f (x) is close to 1. It will be a good approximation for small

arguments at least to replace f (x) by 1. Let us write
√ 
π 1 z
z − k2

Ψk (z) = ze 2

Then

Zhk z′′
X′ iωhk 2
−2πinh/k
p(n) = 5/2
e e2πnz/k Ψk (z)dz+
0≤h<k≤N
k
z′hk
′′
Zzhk  
X iωhk −2πinh/k  ′ 
h i 2
e Ψk (z) f (e2πi k + z ) − 1 e2πnz/k dz
o≤h<k≤N
k5/2
z′hk

where the second term compensates for the mistake committed on taking 158
f (x) = 1. The trick will be now to use the first term as the main term and
17. Lecture 121

to use an estimate for the small contribution from the second term. We have
∗
now to appraise this. Writing Ihk and Ihk for the two integrals, we have
X′ iωhk X′ iωhk
p(n) = 5/2
e−2πinh/k Ihk + 5/2
e−2πinh/k Ihk
∗

o≤h<k≤N
k o≤h<k≤N
k
∗
Here we stop for a moment and consider only Ihk and see what great ad-
vantage we got from our special path.

This is the arc of the circle |z − 12 | = 21 from z′hk to z′′hk described clockwise.
∞
P ∗
Since f (x)−1 = p(ν)xν , the integrand in Ihk is regular, and so for integration
m=1
we can just as well run across, along the chord from z′hk to z′′hk . Let us see what
happens on the chord. We have

 f (e2πih′ /k−2πz ) − 1 Ψk (z)e2πnz/k2  =  f (e2πih′ /k−2π/z ) − 1 √ze 12zπ − 12k2 + k2 

     πz 2πnz


 
∞
√ R π Rz π2 (− 12 +2n) X (2πi h′ 2π 
− )ν
= ze 12z e k ×  p(ν)e k z 
 ν=1 
∞
√ X 1 π π 1
≤ | z| p(ν)e−R z (2πν− 12 ) e k2 (− 12 +2n)Rz
ν=1

Let us determine Rz and R 1z on the path of integration o < Rz ≤ 1 on the 159

chord. And R 1z > 1; for

1 1 x
R =R = 2 ,
z x + iy x + y2
17. Lecture 122

while the equation to the circle is (x − 21 )2 + y2 = 14 or x2 + y2 = x; the interior

of the circle is x2 + y2 < x, and so R 1z ≤ 1, equality on the circle.
√
| z| ≤ the longer of the distances of z′hk , z′′hk from 0.
We have already computed z′hk and z′′hk : 160

k2 kk1
z′hk = 2 2
+i 2 ,
k1 + k k1 + k 2
k2 kk2
z′′hk = +i
k22 + k2 k22 + k2
k4 + k2 k12 k2
|z′hk |2 = =
(k12 + k2 )2 k2 + k12
Now we wish to appraise this in a suitable way.

2(k12 + k2 ) = (k1 + k2 )2 + (k1 − k)2

≥ (k1 + k)2
≥ N2,

from our discussion of adjacent fractions. So

2k2
|z′hk |2 ≤
N2
√
′ 2·k
or |zhk | ≤ ;
N
√
2·k
also |z′′hk | ≤
N
So the inequality becomes 161

  √ 1/2 X∞
 f (e2πih′ /k−2π/z ) − 1 Ψk (z)e2πnz/k2  ≤ 4 2 · k
 1 2
p(ν)e(2πν−π/12) eπ(− 2 +2/n1)/k
N 1/2 ν=1
k1/2
≤ Ce2π|n|
N 1/2
∞
P
where C is independent of ν, since the series p(ν)e−(2πν−π/12) is convergent.
ν=1√
Since the length of the chord of integration < 2 2 · k/(N + 1)m, we have
 ∗  k3/2
Ihk  < C1 e2π|n| 3/2
N
17. Lecture 123

Then
 
 X′ iωhk −2πnh/k ∗  X′ 1
 5/2
e Ihk  ≤ C1 e2π|n| 3/2
0≤h<k≤N k  0≤h<k≤N
kN
X 1
≤ C1 e2π|n| ,
0<k≤N
N 3/2

Since the summation is over h < k with (h, k) = 1, so that there are only 162
ϕ(k) terms and this is ≤ k. So the last expression is

C1 e2π|n| N −1/2

Hence X iωhk −2πinh/k

p(n) = e Ihk + RN
o≤h<k≤N
k5/2
where
−1/2
|RN | < C1 e2π|n|N
Lecture 18

The formula that we had last time looked like this: 163
X′
p(n) = iωhk e−2πinh/k k−5/2 Ihk + RN ,
o≤h<k≤N

and it turned out that

|RN | ≤ Ce2π|n| N −1/2
We had ′′
Zzhk
2
Ihk = Ψ(z)e2πnz/k dz
z′hk

and the path of the integration

was the arc from z′hk to z′′hk in the
sense indicated. And now what
we do is this. We shall add the
missing piece and take the in-
tegral over the full circle, how 0
over excluding the origin. Now
the path is taken in the negative
sense, and we indicate this by
writing
Z
2
Ψk (z)e2πnz/k dz.
k(−)

This is an improper integral with both ends going to zero. That it exists is
′
Rzhk
clear, for what do we have to compensate for that? we have to subtract · · ·
0

124
18. Lecture 125

R0
and · · · , and we prove that these indeed contribute very little. What is after 164
z′′hk
all Ψk (z)?
√ 12π ( 13 − z2 )
Ψk (z) = ze k

0 < Rz ≤ 1 and R1/z = 1 on the circle, so that

√
|Ψk (z)| ≤ | zeπ/12 |
2
and |e2πnz/k | ≤ e2π|n| ;

so that the integrand is bounded. Hence the limit exists. This is indeed very
astonishing, for Ψ has an essential singularity at the origin; but on the circle it
does not do any harm. Near the origin there are value which are as big as we
want, but we can approach the origin in a suitable way. This is the advantage
of this contour. The earlier treatment was very complicated.

We can now estimate the inte-

√
grals. Since |z′hk | ≤ 2 · k/N, the
chord can be a little longer, in
π
fact times the chord at most.
2
So 0
 z′ 
Zhk √ √ !1
  ≤ 2 · π k e2π|n| z · k 2 eπ/12
  2N N
0 
≤ Ce2π|n| k3/2 N −3/2 .

R0
The same estimate holds good for · · · . Hence introducing.
z′′hk

Page missing page No 165 165

Now everything is under our control. N appeared previously tacitly in z′hk , 166
because z′hk depends on the Farey arc. Now N appears in only two places. So
p(n) is the limit of the sum which we write symbolically as
∞
X Z
2
p(n) = i Ak (n)k−5/2 Ψk (z)e2πnz/k dz
k=1
K (−)
18. Lecture 126

(n R 0, integral, and p(n) = 0 for n < 0). So we have an exact infinite series
for p(n).

A thing of lesser significance is

to determine the sum of this se-
ries. So we have to speak about
the integral. Let us take one
more step. Let us get away from
the circle. Replace z by w1 . We 0
do know what this will mean. w
will now run on a line parallel to
the imaginary axis, from 1 − i∞
to 1 + i∞. So

∞ 1+i∞
Z
X
−5/2 π 2 2πn dω
p(n) = − Ak (n)k ω−1/2 e 12 (ω−1/k ω)
e k2 ω ·
k=1
ω2
1−i∞

∞ 1+i∞
Z
1X π π
= Ak (n)k−5/12 ω−5/2 e 2 + 12k2 ω (24n−1) dω
i k=1
1−i∞

One more step is advisable to get a little closer to the customary notation. 167
πω
We then get traditional integrals known in literature. Put = s,
12
π
12 +i∞
∞ Z
1  π 3/2 X π2
p(n) = Ak (n)k−5/2 s−5/2 e s+ 12k2s (24n − 1)ds
i 12 k=1 π
12 −i∞

One could look up Watson’s ‘Bessel Functions’ and write down this inte-
gral as a Bessel function. But since we need the series anyway we prefer to
compute it directly. So we have to investigate an integral of the type
c+i∞
Z
1 ν
L(ν) = s−ν−1 e s+ s ds
2πi
c−i∞

It does not matter what c > 0 is because it means only a shift to a parallel
line, and the integrand goes to zero for large imaginary part of s. For absolute
convergence it is enough to have a little more than s−1 . So take Rν > 0; in our
case ν = 3/2.
18. Lecture 127

So let us study the integral

c+i∞
Z
1 v
Lν (v) = s−ν−1 e s+ s ds
2πi
c−i∞

leaving it to the future what to do with v. The integration is along a 168

line parallel to the imaginary

axis. We now bend the path
into a loop as in the figure
and push the contour out, so
0
that along the arcs we get
negligible contributions.

The contribution from the arc |s| = R is

!
1
O
Rν+1 · R
ν
since |e s+ s | ≤ ec eRν/R, for a fixed v; this is O(R−ν ) → 0 as R → ∞, since
ν > 0. So the integral along the ordinate becomes a ‘loop integral’, starting
from −∞ along the lower bank of the real axis, looping around the origin and
proceeding along the upper bank towards −∞; the loop integral is written in a
fashion made popular by Watson as
(0+)
Z
1 ν
s−ν−1 e s+ s ds
2πi
−∞

For better understanding we take a specific loop. On the lower bank of the 169
negative real axis we proceed only up to −ǫ,

then go round a circle of radius ǫ in the positive sense and proceed thence
along the upper bank, the integrand now having acquired a new value-unless ν
is an integer. This we take as a standardised loop. We now prove that Lν (V ) is
actually differentiable and that the derivative can be obtained by differentiating
18. Lecture 128

under the integral sign. For this we take {Lν (V + h) − Lν (V )} /h and compare
it with what we could foresee to be L′ν (V ) and show that the difference goes to
zero as h → 0.
(0+)
Z
Lν (V + h) − Lν (V ) 1 ν
− s−ν−2 e s+ s ds
h 2πi
−∞
Z(0+)  v+h ν ν

1  s −e s
−ν−1 s  e
 es 


= s e  −  ds
2πi  h
 s
−∞
(0+)
Z  h 
1 ν es − 1 1

 
s−ν−1 e s+ s

=  −  ds
2πi  h
 s

−∞

Now 170
2
h h h
es − 1 1 s + s2 ,s! + · · · 1
− = −
h s ( h s )
1 h
=h 2 + + ···
s · 2! s3 · 3!

On the path of integration, |s| ≥ ǫ > 0; so

 h 
 e s −1 1 
 −  ≤ C|h|,
 h s

since we are having a quickly converging power series.

 
 Z(0+) 
 νL (ν + h) − Lν (ν) 1 −v−2 s+ sv 
∴ − s e ds

 h 2πi 

−∞
 ∞ 
 Z 


 1 −x ǫRν 1 Rν



≤ C|h| 2 v+1
e e dx + 2πǫ · ν+1
e ǫ
 = 0(h)


 x e 


ǫ

Lv (ν+h)−Lv (ν)
So the limit lim h exists and Lν (ν) is differentiated uniformly in a
h→0
circle of any size. Since the differential integral is of the same shape we can
differentiate under the integral as often as we please.
Lecture 19

The formula for p(n) looked like this: 171

∞ Z(0+)
1  π 3/2 X 1 π 2
p(n) = Ak (n)k−5/2 s−5/2 e s+ s ( 12k ) (24n−1) ds
i 12 k=1 −∞

We discussed the loop integral

(0+)
Z
1 v
Lv (ν) = s−v−1 e s+ s ds, Rν > 0.
2πi
−∞

We can differentiate under the integral sign and obtain

Z(0+)
1 v
L′v (ν) = s−v−2 e s+ s ds = Lv+1 (ν)
2πi
−∞

This integral is again of the same sort as before; so we can repeat differenti-
ation under the integral sign. Clearly then Lv (ν) is an entire function of ν · Lv (ν)
has the expansion in a Taylor series:
∞
X L(r)
v (0) r
Lv (ν) = ν
r=0
r!

L(r)
v (ν) can be foreseen and is clearly

(0+)
Z
1 v
s−v−1−r e s+ s ds
2πi
−∞

129
19. Lecture 130

So 172
(0+)
∞ Z
X νr 1
Lv (ν) = s−v−1−r e s ds
r=0
r! 2πi
−∞

We now utilise a famous formula for the Γ-function - Hankel’s formula,

viz,
Z(0+)
1 1
= s−µ e s ds.
Γ(µ) 2πi
−∞
π
This is proved by means of the formula Γ(s)Γ(1 − s) = and the Euler
sin πs
integral. Using the Hankel formula we get L explicitly:
∞
X νr
Lv (ν) =
r=0
r!Γ(v + r + 1)

What we have obtained is something which we could have guessed earlier.

Expanding ev/s as a power series, we have
(0+)
Z ∞
1 X (v/s)r
Lv (ν) = s−v−1 e s ds
2πi r=0
r!
−∞
(0+)
Z ∞
1 X vr s −v−1−r
= es ds,
2πi r!
−∞ r=0

and what we have proved therefore is that we can interchange the integration
and summation. We have
L′v (ν) = Lv+1 (ν).
Having this under control we can put it back into our formula and get a
final statement about p(n).
 π 3/2 X∞  π 2 !
p(n) = 2π Ak (n)k−5/2 L3/2 (24n − 1)
12 k=1
12k
This is not yet the classical formula of Hardy and Ramanujan. One trick 173
one adopts is to replace the index. Remembering that L′v (ν) = Lv+1 (ν), we have
 π 2 !  π 2 !
′
L3/2 (24n − 1) = L1/2 (24n − 1)
12k 12k
19. Lecture 131
!
6k2 d  π 2
= 2 L1/2 (24n − 1)
π dn 12k

Let us write the formula for further preparation closer to the Hardy Ra-
manujan notation:
 π 1/2 X∞  π 2 !
−1/2 d
p(n) = Ak (n)k L1/2 (24n − 1)
12 k=1
dn 12k

1
Now it turns out that the L-functions for the subscript 2 are elementary
functions. We introduce the classical Bessel function
∞
X (−)r (z/2)2r+v
Jv (z) =
r=0
r!Γ(v + r + 1)

and the hyperbolic Bessel function (or the ‘Bessel function with imaginary
argument’)
∞
X (z/2)2r+v
Iv (z) =
r=0
r!Γ(v + r + 1)
How do they belong together? We have 174
!
z2  z −v
Lv = Iv (z) ,
4 2
!
z2  z −v
Lv − = Jv (z) ,
4 2

connecting our function with the classical functions. In our case therefore we
could write in particular
 π 2 ! π √  π √ −1/2
L1/2 (24n − 1) = I1/2 24n − 1 24n − 1
12k 6k 12k

This is always good, but we would come into trouble if we have 24n−1 ≤ 0.
It is better to make a case distinction; the above holds for n ≥ 1, and for n ≤ 0,
n = −m, we have
 π 2 !  π 2 !
L1/2 (24n − 1) = L1/2 − (24m + 1)
12k 12k
π √  π √ −1/2
= J1/2 24m + 1 24m + 1
6k 12k
19. Lecture 132

So we have: n ≥ 1.
  √ 
 π 1/2 X∞  I1/2 π 24n − 1 
d 6k
Ak (n)k−1/2   √
 
p(n) = 1/2 
12 k=1
dn  π
24n − 1
12k

n = −m ≤ 0 175
  √ 
 π 1/2 X∞  J1/2 π 24m + 1 
d 
p(n) = p(−m) = − Ak (−m)k−1/2 ·   √6k 
1/2 
12 dm  π
k=1 12k 24m + 1

We are not yet quite satisfied. It is interesting to note that the last expression
is 1 for n = 0 and 0 for n < 0. We shall pursue this later.
We have now more or less standardised functions. We can even look up
tables and compute the Bessel function. However I1/2 and J1/2 are more ele-
mentary functions.
∞
X (−)r (z/2)2r+1/2
J1/2 (z) =
r=0 r!Γ(r + 23 )
∞ 1
X (−)r (z/2)2r+ 2
=     
1 1 1 1
r=0 r! r + 2 r − 2 · · · 2 Γ 2
!1/2 X∞
2 (−)r z2r+1
=
πz r=0
(2r + 1)!
!1/2
2
= sin z.
πz

Similarly if we has abolished (−)r we should have 176

!1/2
2
I 21 (z) = sinh z
πz
!1/2
I1/2 (z) 2 2 sinh z
1/2
= I1/2 (z) = √ =
(z/2) z 2 z
J1/2 (z) 2 sin z
= √
(z/2)1/2 π z

We are now at the final step in the deduction of our formula:

n≥1
19. Lecture 133

∞  π
√ 
1 X −1/2 d 
 sinh 6k 24n − 1 
p(n) = √ Ak (n)k  √ 
3 k=1 dn  π 24n − 1 
6k
r r r
π √ π 2 1 c 1 2
or with 24n − 1 = (n − ) = n − ,C = π ,
6k k 3 24 k 24 3
 q 
c 1 
d  sinh k n − 24 
∞  
1 X
p(n) = √ Ak (n)k1/2  q 
π 2 k=1 dn  
n− 124

n = −m ≤ 0
 q 
∞  sin c m + 1 
1 X 1 d   k 24 
p(n) = p(−m) = − √ Ak (−m)k 2  q 
π 2 k=1 dm 
 m+ 1 24

This is the final shape of our formula -a convergent series for p(n). 177
The formula can be used for independent computation of p(n). The terms
become small. It is of interest to find what one gets if one breaks the series off,
say at k = N
N
π5/2 X
p(n) = √ · · · + RN
12 3 k=1
Let us appraise RN · |Ak (n)| ≤ k, because there are only ϕ(k) roots of unity.
We want an estimate for L3/2 . For n ≥ 1,
 2 r
! X ∞ π
 π 2
6k2
n
L3/2 (24n − 1) ≤  
12k r=0 r!Γ r + 2
5
 2 r
∞ π
X 6(N+1) 2 n
≤    
1 1 3
r=0 r!Γ 2 · 2 · r+ 2
 2 r
∞ π 2r+1
1 X 6(N+1)2 n · 2
(sincek > N in RN ) = √
π r=0 (2r + 1)!(r + 32 )
 r
∞ 2π2
2 X 3(N+1)2 n
≤ √
π r=0 (2r + 1)!
19. Lecture 134
 r
∞ 2π2
2 2 X 3(N+1)2 n
≤ · √
3 π r=0 (2r)!
4 π
√ 3n
< √ e N+1 3
3 π

π2 π
√ 2n X ∞
1
∴ |RN | ≤ √ e N+1 3 3/2
9 3 k=N+1
k

π2 π
√ 2n Z∞ dk
≤ √ e N+1 3
9 3 k3/2
N
2π2 π √ 1
∴ |RN | < √ e N+1 2n3 1/2
9 3 N

This tells us what we have in mind.

√ Make N suitably large. Then one gets 178
something of interest. Put N = [α n], α constant. Then
1
RN = O(n− 4 )

And this is what Hardy and Ramanujan did. Their work still looks different.
They did not have infinite series. They had replaced the hyperbolic sine by the
most important part of it, the exponential.
 q The series
 converges in our case
c 1
since sinh x ∼ x as x → 0, so that sinh k n− 24 behaves roughly like kc . On
differentiation we have k12 so that along with k in the numerator we get k−3/2
1/2

and we have convergence. In the Hardy-Ramanujan paper they had

√
 √ 
[X n]  c n− 1 
1 d  e k 24 
−1 ∗
Ak (n)k1/2  q

p(n) = √  + O(n 4 ) + RN
2π 2 k=1 dn   1 
n − 24

sinh was replaced by exp.; so they neglected 179

 − c √n− 1 

√
 
[X n]
1 d  e k 24 
R∗ = √ Ak (n)k1/2  q 
2π 2 k=1 dn  n − 1 

24
√  √ √ 
[X n]  e− kc n− 241 c e− ck n− 241 
|R∗ | = O  k3/2  1
· + 1 


k=1 n − 24
k (n − 24 )
19. Lecture 135

The exponential is strongly negative if k is small; so it is best for k = 1.

Hence
 √ √ 
 1 [Xn]
1
[X n] 
|R∗ | = O   k1/2 + √ k3/2 
n k=1 n k=1
N
X
k1/2 = O(N 3/2 )
k=1
N
X
k3/2 = O(N 5/2 )
k=1

So
!!
1 3/4 1
|R∗ | = O n + √ n5/4
n n
 1
−4
=O n

The constants in the O-term were not known at that time so that numerical 180
computation was difficult. If the series was broken off at some other place the
terms might have increased. Hardy and Ramanujan with good instinct broke
off at the right place.
We shall next resume our function-theoretic discussion and cast a glance at
the generating function for p(n) about which we know a good deal more now.
Lecture 20

We found a closed expression for p(n); we shall now look back at the generat- 181
ing function and get some interesting results.
∞
1 X
f (x) = ∞ = p(n)xn ,
Q
(1 − xn ) n=0
n=1

and we know p(n). p(n) in its simplest form before reduction to the traditional
Bessel functions is given by
 π 3/2 X∞  π 2 !
p(n) = 2π Ak (n)k−5/2 L3/2 (24n − 1) ,
12 k=1
12k
!! ∞
X π2 (n − 1 )r
π2 1
where L3/2 n − = 6k
 24 
6k2 24 5
r=0 r!Γ 2 + r

We wish first to give an appraisal of L and show that the series for p(n)
converges absolutely. The series is
∞ ∞ !
 π 3/2 X X π2
f (x) = 2π xn Ak (n)k−5/2 L3/2 (n − α) ,
12 n=0 k=1
6k2

1
where we write 24 = α for abbreviation - it will be useful for some other
purposes also to have a symbol there instead of a number.
We make only a crude estimate.
 ! X  2 r
∞ π
 π2  6n
 L3/2 2
(n − α)  ≤  
6k 
r=0 r!Γ 5 + r
2

136
20. Lecture 137
 
∞ π2 r
6 n
X
=    
1 1
r=0 r!Γ 2 · · 23 · · ·
2
3
2 +r
 r
∞ 2π
22 X 3 πn
√
π r=0 (2r + 1)!(3 + 2r)
∞ √ r
X (C n)2r 2
≤4 ,C = π ,
r=0
(2r)! 3
∞ √
X (C n)ρ
≤4
ρ=0
ρ!
√
= 4eC n

So f (x) is majorised by 182

∞ √ ∞
X X 1
constant x |x|n eC n
3/2
n=1 k=1
k

√ convergent for |x| < 1, indeed uniformly so for |x| ≤ 1 −δ,

and this is absolutely
δ > 0, because eC n = 0(eδn ), δ > 0, so that we need take |xeδ | < 1. We can
therefore interchange the order of summation:
∞ ∞ !
 π 3/2 X X π2
f (x) = 2π k−5/2 Ak (n)xn L3/2 (n − α)
12 k=1 n=0
6k2
∞ ∞  !
 π 3/2 X X′ X h n π2
= 2π k−5/2 ωhk xe−2πi k L3/2 (n − α)
12 k=1 h mod k n=0
6k2

where the middle sum is a finite sum. This is all good for |x| < 1. Now call 183

∞ !
X π2
Φk (z) = L3/2 (n − α) zn
n=0
6k2

So in a condensed form f (x) appears as

 π 3/2 X∞ X′  h
f (x) = 2π k−5/2 ωhk Φk xe−2πi k
12 k=1 h mod k

We have now a completely new form for our function. It is of great in-
terest to consider Φk (z) for its own sake; it is a power series (|z| < 1) and the
20. Lecture 138

coefficients of zn are functions of n − α.

∞
X νr
L3/2 (ν) =  
5
r=0 r!Γ 2 +r
This is an entire function of ν, for the convergence is rapid enough in the
whole plane. Looking into the Hadamard theory of entire functions, we could
1
see that the order of this function is . This is indeed plausible, for the de-
P νr 2 P ( √ν)2r √
ν
nominator is roughly (2r)! and (2r)!
√
= (2r)! ∼ e ; or the function grows
like e ν , and this is characteristic of the growth of an entire function or order
1 n
2 . The coefficients of z are themselves entire functions in the subscript n.
We now quote a theorem of Wigert to the following effect. Suppose that we
∞
P
have a power series Φ(z) = g(n)zn where g(ν) is in entire function of order 184
n=0
less than 1; then we can say something about Φ(z) which has been defined so
far for |z| < 1. This function can be continued analytically beyond the circle of
convergence, and Φ(z) has only z = 1 as a singularity; it will be an essential
singularity in general, but a pole if g(ν) is a rational function. We can extract
the proof of Wigert’s theorem from our subsequent arguments; so we do not
give a separate proof here.
Φk (z) is a double series :
 2 r
∞ ∞ π
X X 6k 2 (n − d)
Φk (z) = zn   , |z| < 1
5
n=0 r=0 r!Γ 4 +r

This is absolutely convergent; so we can interchange summations and write

 2r
∞ π
√ ∞
X k 6 X
Φk (z) =   (n − α)r zn
5
r=0 r!Γ 2 + r n=0
 2r
∞ π
√
X k 6
=   ϕr (z)
5
r=0 r!Γ 2 + r

∞
P
where ϕr (z) is the power series (n−α)r zn . Actually it turns out to be a rational
n=0
function. Φk (z) can be extended over the whole plane.
∞
X 1
ϕr (z) = zn = .
n=0
1−z
20. Lecture 139

Differentiating ϕr (z), 185

∞
X
ϕ′r (z) = n(n − α)r zn−1 ,
n=0
X∞
zϕ′r (z) = n(n − α)r zn ,
n=0
∞
X
αϕr (z) = α(n − α)r zn ;
n=0

so,
∞
X
zϕ′r (z) − αϕr (z) = (n − α)r+1 zn = ϕr+1 (z)
n=0

This says that we con derive ϕr+1 (z) from ϕr (z) by rational processes and
differentiation. This will introduce no new pole; the old pole z = 1 (pole
for ϕ◦ (z)) will be enhanced. So ϕr (z) is rational. Let us express the function
1
a little more explicitly in terms of the new variable u = z−1 or 1u + 1 = z.
Introduce (−)r+1 ϕr (z) = (−)r+1 ϕr (1 + u) = ψr (u), say he last equation which
was a recursion formula now becomes
!
1
(−)r+2 ψr+1 (u) = + 1 (−)r u2 ψ′r (u) − α(−)r+1 ψr (u)
u
! ! !
1 1 1 1
because ψ′r (u) = (−)r+1 ϕ′r 1 + − 2 = (−)r ϕ′r 1 +
u u u u2
′
∴ ψr+1 (u) = u(u + 1)ψr (u) + αψr (u)

This is a simplified version of our recursion formula. We have a mind to

expand about the singularity z = 1. Let us calculate the ψ′ s.

ψ0 (u) = u
ψ1 (u) = u(u + 1) + αu = (1 + α)u + u2
ψ2 (u) = u(u + 1)(2u + 1 + α) + α(1 + α)u + αu2
= (1 + α)2 u + (2α + 3)u3 + 2u3

ψr (u) is a polynomial of degree r+1 without the constant term. The coefficients 186
are a little complicated. If we make a few more trials we get by induction the
following:
20. Lecture 140

Theorem.
r
X
ψr (u) = ∆ j (α + 1)r u j+1 ,
j=0

where ∆ j is the jth difference.

By definition,

∆ f (x) = f (x + 1) − f (x),
∆2 f (x) = ∆∆ f (x) = ∆ f (x + 1) − ∆ f (x)
= f (x + 2) − 2 f (x + 1) + f (x)

The binomial coefficients appear, and

k !
X k
∆k f (x) = (−)k−ℓ f (x + ℓ)
ℓ=0
ℓ

How does the formula for ψr fit? For induction one has to make sure that 187
the start is good.

ψ0 (u) = (α + 1)◦ u = u
ψ1 (u) = (α + 1)′ u′ + ∆(α + 1)′ u2 = (α + 1)u + u2
ψ2 (u) = (α + 1)2 u′ + ∆(α + 1)2 u2 + ∆2 (α + 1)2 u3
 
= (α + 1)2 u + (α + 2)2 − (α + 1)2 u2 + 2u3
= (α + 1)2 u + (2α + 3)u2 + 2u3

So the start is good. We assume the formula up to r.

r n
X o
ψr+1 (u) = (u2 + u)( j + 1)∆ j (α + 1)r u j + α∆ j (α + 1)r u j+1
j=0
r+1 n
X o
= j∆ j−1 (α + 1)r u j+1 + ( j + 1 + α)∆ j (α + 1)r u j+1
j=0

(A Seemingly negative difference need not bother us because it is accom-

panied by the term j = 0).
r+1
X  
= u j+1 j∆ j−1 (α + 1)r + ( j + 1 + α)∆ j (α + 1)r
j=0
20. Lecture 141

To show that the last factor is ∆ j (α + 1)r+1 , we need a side remark. Intro- 188
duce a theorem corresponding to Leibnitz’s theorem on the differentiation of a
product. We have

∆ f (x)g(x) = f (x + 1)g(x + 1) − f (x)g(x)

= f (x + 1)∆g(x) + f (x + 1)g(x) − f (x)g(x)
= f (x + 1)∆g(x) + ∆ f (x) · g(x)

The general rule is

k !
X k k−ℓ
∆k f (x)g(x) = ∆ f (x + ℓ)∆ℓ g(x)
ℓ=0
ℓ

This is true for k = 1. We prove it by induction,

∆k+1 f (x)g(x) = ∆(∆k f (x)g(x),

and since ∆ is a linear process, this is equal to

k !
X k n k−ℓ o
∆ f (x + ℓ + 1)∆ℓ+1 g(x) + ∆k+1−ℓ f (x + ℓ)∆ℓ g(x) ,
ℓ=0
ℓ

which becomes, on rearranging summands,

k+1 ( ! !)
X k k
∆k+1−ℓ f (x + ℓ)∆ℓ g(x) + ,
ℓ=0
ℓ ℓ−1
   k   k  
and the last factor is k+1
ℓ , −1 = k+1 = 0 This proves the rule.
Applying this to (α + 1)r , 189

(α + 1)r+1 = (α + 1)(α + 1)r ; write f = α + 1, g = (α + 1)r ,

and observe that f being linear permits only 0th and 1 st differences;
! !
k k
∆k (α + 1)r = ∆k−1 (α + 1)r + (α + k + 1)∆k (α + 1)r
k−1 k
= k∆k−1 (α + 1)r + (α + k + 1)∆k (α + 1)r
Xr
∴ ψr (u) = ∆ j (α + 1)r u j+1
j=0
20. Lecture 142

We can now rewrite the ϕ′ s:

ϕr (z) = (−)r+1 ϕr (n)

r
X 1
= (−)r+1 ∆ j (α + 1)r
j=0
(z − 1) j+1

ϕr has now been defined in the whole plane.

Lecture 21

We have rewritten the generating function f (x) as a sum consisting of certain 190
functions which we called Φk (x):
∞
 π  23 X X′  h
f (x) = 2π k−5/2 ωhk Φk xe−2πi k
12 k=1 h mod k
∞ !
X π2
where Φk (z) = L3/2 2
(n − α) zn ,
n=0
6k

1
with α = 24 . Φk (z) could also be written as
 2r
∞ π
√
X k 6
Φk (z) =   ϕr (z′′ )
5
r=0 r!Γ 2 +r

where ϕr (z) is a rational function as we found out. We got ϕ explicitly by means

of a certain ψ:
r
X
j 1
ϕr (z) = (−)r+1 ∆α (α + 1)r
j=0
(z − 1) j+1
What we need for questions of convergence is an estimate of ϕr ; this is not
difficult.
∆ f (x) = f (x + 1) − f (x) = f ′ (ξ1 ), x < ξ1 < x + 1,
by the mean-value theorem; and because ∆ is a finite linear process we can in-
terchange it with the operation of applying the mean value theorem and obtain 191

∆2 f (x) = ∆(∆ f (x)) = ∆ f ′ (ξ1 ) = f ′ (ξ1 + 1) − f ′ (ξ1 )

= f ′′ (ξ2 ), x < ξ1 < ξ2 < ξ1 + 1 < x + 2;

143
21. Lecture 144

∆3 f (x) = ∆(∆2 f (x)) = ∆2 f ′ (ξ), x < ξ < x + 1,

= f ′′′ (ξ3 ), x < ξ < ξ3 < ξ + 2 < x + 3;

and in general
∆k f (x) = f k (ξ), x < ξ < x + k.
This was to be expected. Take |1 − z| ≥ δ, 0 < δ < 1 so that z is not too
1
close to 1. > 1 and 0 < α < 1
δ
r
X 1
|ϕr (z| ≤ r(r − 1) · · · (r − j + 1)(1 + α + j)r− j ·
j=0
δ j+1
r
X (α + 1 + r)r
<
j=0
δ j+1
(α + 1 + r)r
< (r + 1)
δr+1
(α + 1 + r)r+1
<
δr+1
Originally we know that the formula for f (x) was good for |x| < 1. From
this point on we give a new meaning to ϕr (z) for all z , 1.
This is a new step. We prove that the series for Φk (z) is convergent not 192
merely for |z| < 1 but also elsewhere. The sum in Φk (z) is majorised by
 2 r
∞ π
1X 6 (α + 1 + r)r+1
 ·
δ r=0 r!Γ 5 + r δr
2

This is convergent, for thought the numerator increases with r, we have by

Stirling’s formula
rr rr er
∼ √ 1
= √
r! 2πrr+ 2 e−r 2πr
So as far as convergence is concerned it is no worse than
 2 r
∞ eπ !r
1X 6δ α+1
  (α + 1 + r) 1 +
δ r=1 Γ 5 + r r
2

which is ≤ Cδ , the power series still being rapidly

 converging because of the
α+1 r
2

factorial in the denominator and eπ
6δ is fixed and 1 + r is bounded. So we
have absolute convergence and indeed uniformly so for |1 − z| ≥ δ.
21. Lecture 145

We have now a uniformly convergent series outside the point z = 1, and

Φk (z) is explained at every point except z = 1 which is an essential singularity.
1
Φk (z) is entire in . From this moment if we put it back into our argument
1−z
we have f (x) in the whole plane if xe−2πih/k keep away from 1. And we are sure 193
of that; either |x| ≤ 1 − δ or |x| ≥ 1 + δ. Originally x was only inside the unit
circle; now it can be outside also. In both cases f (x) is majorised by
∞
X ∞
X
5
k− 2 , k · C δ = C δ k−3/2 ,
k=1 k=1

which is absolutely convergent.

Therefore we have now a very peculiar situation. In this notation of Φk we
have obtained a function which represents two analytic functions separated by a
natural boundary which is full of singularities and cannot be crossed. They are
not analytic continuations. The outer function is something new; it is analytic
because the series is uniformly convergent in each compact subset.
Consider the circle. We state something more explicit which explains the
behaviour at each point near the boundary. Since every convergence is absolute
there are no difficulties and convergence prevails even if we take each piece
separately.

 π 3/2 X∞ X′
5
f (x) = −2π k− 2 ωhk
12 k=1 h mod k
 2 r
X∞ − π6k ∞
X
j 1
  ∆α (α + 1)r · −2πih/k − 1) j+1
5
r=0 r!Γ 2 + r j=0
(xe

We can now rearrange at leisure. 194

 π 3/2 X∞ X′
5
f (x) = −2π k− 2 ωhk
12 k=1 h mod k
 2 r
∞ h ∞ π
X e−2πi k ( j+1) X j − 6k 2
r
2πih/k j+1
∆α (α + 1)  
j=0
(x − e ) r= j r!Γ 5 + r
2
P P∞
However, if we replaced ∞ r= j by r=0 it would not to any harm because the
summation is applied to a polynomial of degree r and the order of the difference
is one more than the power. We can therefore write, taking ∆ outside,
21. Lecture 146

 π 3/2 X∞ X′
5
f (x) = −2π k− 2 ωhk
12 k=1 h mod k
 
∞ ∞ π2 r
X e2πi hk ( j + 1) X − 6k 2
∆j
2πih/k ) j+1 α
(α + 1)r  
j=0
(x − e r=0 r!Γ 52 + r
∞ ∞ h !
 π −5/2 X 5
X′ X e2πi k ℓ ℓ−1 π2
= −2π k2 ωhk ∆ L 3/2 − (α + 1)
12 k=1 h mod k ℓ=1
(x − e2πih/k )ℓ α 6k2

It is quite clear what has happened. x appears only in the denominator,

a root of unity is subtracted and the difference raised to a power 1. Choose 195
B
specific h, x, 1; then we have a term . We have a conglomerate of
(x − e2πih/k )ℓ
terms which look like this, a conglomerate of singularities at each root of unity.
So we have a partial fraction decomposition not exactly of the Mittag-Leffler
type. Here of course the singularities are not poles, and they are everywhere
∞
P
dense on the unit circle. Each series represents one specific point e2πih/k .
ℓ=1
Let us return to our previous statement. f (x) is regular and analytic outside
∞
Q
the unit circle. What form has it there? Inside it is (1 − xm ). We shall expand
m=1
f (x) about the point at infinity. We want the ϕ′ s explicitly.
1
ϕ0 (z) =
1−z
ϕr+1 (z) = zϕ′r (z) − αϕr (z)
∞
z−1 z−1 X
ϕ0 (z) = = − = − z−m
z−1 − 1 1 − z−1 m=1
∞
X ∞
X ∞
X
ϕ1 (z) = mz−m + α z−m = (m + α)z−m
m=1 m=1 m=1

The following thing will clearly prevail

∞
X
ϕr (z) = (−)r+1 (m + α)r z−m
m=1

This speaks for itself.

∞
X ∞
X
ϕr+1 (z) = (−)r m(m + α)z−m + (−)r α (m + α)r z−m
m=1 m=1
21. Lecture 147

So the general formula is justified by induction. 196

 2 r
∞ π ∞
X − 6k 2
X
Φk (z) = −   (m + α)r z−m
5
r=0 r!Γ 2 + r m=1

for all |z| > 1. Exchanging summations,

 r
∞ ∞ π2
X X 6k2
(−m− α)
Φk (z) = − z−m  
5
m=1 r=0 r!Γ 2 + r
∞ !
X
−m π2
=− z L3/2 (−m − α)
m=1
6k2

Put this back into f (x); we get for |x| > 1,

∞ ∞  !
 π 3/2 X 5
X′ X h m π2
f (x) = −2π k− 2 ωhk x−1 e2πi k L3/2 (−m − α) ,
12 k=1 h mod k m=1
6k2
X′
and since Ak (n) = ωhk e−2πih/k ,
h mod k
∞ ∞ !
 π 3/2 X X π2
f (x) = −2π k−5/2 Ak (−m)x−m L3/2 (−m − α)
12 k=1 m=1
6k2

Again interchanging summations, 197

∞ ∞ !
 π 3/2 X X π2
f (x) = −2π x−m Ak (−m)k−5/2 L3/2 (−m − α)
12 m=1 k=1
6k2

The inner sum we recognize immediately; it is exactly what we had for

p(n); so
 π 3/2 X∞
f (x) = −2π p(−m)x−m
12 m=1

And here is a surprise which could not be foreseen! By its very meaning
p(−m) = 0. So
f (x) ≡ 0
outside the unit circle. This was first conjectured by myself and proved by
H.Petersson by a completely different method. Such expressions occur in the
theory of modular forms. Petersson got the outside function first and then the
inner one, contrary to what we did.
21. Lecture 148

The function is represented by a series inside the circle, and it is zero out-
side, with the circle being a natural boundary. There exist simpler examples of
this type of behaviour. Consider the partial sums:

x 1
1+ =
1−x 1− x
x x2 1 x2 1
1+ + = + =
1 − x (1 − x)(1 − x2 ) 1 − x (1 − x)(1 − x2 ) (1 − x)(1 − x2 )
x x2 x3
1+ + + + · · · to n + 1 terms
1 − x (1 − x)(1 − x2 ) (1 − x)(1 − x2 )(1 − x3 )
1
=
(1 − x)(1 − x2 ) · · · (1 − xn )
1
For |x| < 1, the partial sum converges to ∞ . For |x| > 1 also 198
Q
(1 − xm )
m=1
it has a limit; the powers of x far outpace 1 and so the denominator tends to
infinity and the limit is zero. The Euler series here is something just like our
complicated function. Actually the two are the same. For suppose we take the
1
partial sum and break it into partial fractions. We
(1 − x)(1 − x2 ) · · · (1 − xn )
get the roots of unity in the denominator, so that we have a decomposition
X Bh,k,l,n
h ℓ
 
x − e2πi k

k ≤ n and ℓ not too high. For a higher n we get a finer expression into partial
fractions. Let us face one of these, keeping h, k, ℓ fixed:
Bh,k,l,n
 h
ℓ
x − e2πi k

Let n → ∞. Then I have the opinion that

!
 π 3/2
− 25 2πi hk ℓ ℓ−1 π2
Bh,k,l,n → −2π ωhk k e ∆α L3k − 2 (α + 1)
12 6k
The B′ s all appear from algebraic relations and so are algebraic numbers
- in sufficiently high cyclotomic fields. And this is equal to something which
looked highly transcendental! though we cannot vouch for this. The verifi-
cation is difficult even in simple cases - and no finite number of experiments
would prove the result.
21. Lecture 149

B0,1,1,n
is itself very complicated. Let us evaluate the principal formula for 199
x−1
f (x) and pick out the terms corresponding to h = 0, k = l, ℓ = √1.
6 12 3 1
L3/2 is just the sine function and terns out to be − − . Since 1−x =
25 75π
1
− x−1 , −1 is the first approximation to B0,1,1,n. If we take the partial fraction
decomposition for
1 1 ·· ·· ··
, = + + ,
(1 − x)(1 − x2 ) (1 − x)(1 − x2 ) (x − 1)2 (x − 1) (1 + x)
the numerator of the second term would give the second approximation. If in-
deed these successive approximations converge to B0,1,1,n we could get a whole
new approach to the theory of partitions. We could start with the Euler series
and go to the partition function.
We are now more prepared to go into the structure of ωhk . We shall study
next time the arithmetical sum Ak (n) and the discovery of A.Selberg. We shall
then go back again to the η-function.
Lecture 22

We shall speak about the important sum Ak (n) which appeared in the formula 200
for p(n), defined as X′
Ak (n) = ωhk e−2πinh/k .
h mod k

we need the explanation of the ωhk ; they appeared as factors in a transfor-

mation formula in the following way:
 h+iz  √ π  1   h′ +i/z 
f e2πi k = ωhk ze 12k z −z f e2πi k ,
hh′ + 1 ≡ 0 (mod k)

Here, as we know,
1
f (x) = Q∞
m=1 (1 − xm )
Y∞
and as η(τ) = eπiτ/12 (1 − e2πimτ ),
m=1
f (e2πiτ ) = eπiτ/12 (η(τ))−1

We know how η(τ) because. ωhk is something belonging to the behaviour

of the modular form η(τ). What is ωhk explicitly? We had a formula
! r
aτ + b cτ + d
η =ǫ η(τ), c > 0,
cτ + d i
and 201
epsilonn is just the question. Our procedure will be to study ǫ and η and then
go back to f where ωhk appeared. The trick in the discussion will be that we

150
22. Lecture 151

shall not use the product formula for η(τ), but the infinite series from the pen-
tagonal numbers theorem. This was carried out at my suggestion by W.Fischer
(Pacific Journal of Mathematics; vol. 1). However we shall not copy him. We
shall make it shorter and dismiss for our purpose all the long and complicated
discussions of Gaussian sums
k
X 2
G(h, k) = e2πiν h/k

v=1

which are of great interest arithmetically, having to do with law of reciprocity

to which we shall return later.
We are able to infer that a formula of the sort quoted for η should exist from
the discussion of V1′ (0/τ). We had the formula (see hechire 14)
. aτ + b !
V1 0 = ···
cτ + d
where the right side contains a doubtful root of unity, which we could discuss in
some special cases, and by iteration in all cases. We shall use as further basis of
our argument that such a formula has been established with the proviso |ǫ| = 1.
We then make a statement about ǫ and use it directly.
After all this long talk let us go to work. We had τ′ = (h′ + i/z/k), τ =
(h + iz)/k. The question is how is τ′ produced from τ? It was obtained by 202
means of the substitution
   
   ′ hh′ +1 
a b = h − k 
   
c d k −h

We can therefore get what we are after if we specify the formula by these
particular values. ! !
h′ + iz √ h + iz
η = ǫ zη
k k
√
with the principal value for z. We wish to determine ǫ defined by this. We
shall expand both sides and compare the results. For expansion we do not use
the infinite product but the pentagonal numbers formula.
∞
X
η(τ) = eπiτ/12 (−)λ e2πiτλ(3λ−1)/2
λ=−∞
∞
X πiτ 2
−12λ)
= (−)λ e 12 (1+36λ
λ=−∞
22. Lecture 152

∞
X 2
= (−)λ e3πiτ(λ−1/6)
λ=−∞

Most determinations of η(τ) make use of the infinite product formula; the
infinite series is simpler here
! X ∞
h + iz h+iz 2
η = (−)λ e3πi k (λ−1/6)
k λ=−∞

In order to get the root of unity a little more clearly exhibited, we replace λ 203
mod 2k.
λ = 2kq + j, j = 0, 1, . . . , 2k − 1 and q runs from −∞ to ∞. So
! X ∞ 2k−1
h + iz X h 1 2 z 1 2
η = (−) j e3πi k (2kq+ j− 6 ) e−3π k (2kq+ j− 6 )
k q=−∞ j=0

The product term in the exponent = 4kq( j − 16 ).3πi kh

= 2πihq(6 j − 1)
= an integral multiple of 2πi

(This is the reason why we used mod 2k).

! 2k−1 ∞
h + iz X h 1 2
X j−1/6 2
η = (−) j e3πi k ( j− 6 ) e−12πzk(q+ 2k ) .
k j=0 q=−∞

We did this purposely in order to make it comparable to what we did in the

theory of V -functions. For Rt > 0, we have
∞ ∞
X 2 1 X − π m2 2πimα
e−πt(q+α) = √ e t e
q=−∞ t m=−∞

This is a consequence of a V -formula we had:

r !
πiτν2 1 1
e V3 (ντ/τ) = V3 ν/ −
τ τ

If we write this explicitly, 204

∞
X 2
V3 (ν/τ) = eπiτn e2πinν ,
n=−∞
22. Lecture 153

and put iτ = −t,

∞ ∞
2
X 2 1 X −πn2 /t 2πinν
e−πtν e−πtn e−2πntν = √ e e ,
n=−∞ t n=−∞
∞ ∞
X 2 1 X − π n2 2πinν
or e−πt(ν+n) = √ e t e ,
n=−∞ t n=−∞

which is the formula quoted. We now apply this deep theorem and get some-
j − 1/6
thing completely new. Putting t = 12zk and α = ,
k
! 2k−1 ∞
h + iz X 1 2 1 X πm2
(−) j e2πi k ( j− 6 ) √ e− 12zk e k ( j− 6 )
h πim 1
η =
k j=0 12kz m=−∞

We rewrite this, emphasizing the variable and exchanging the orders of

summation. Then
! ∞ 2k−1
X πi j+ 2h ( j− 1 )2 + m (6 j−1)
h + iz 1 X πm2
η = √ e− 12kz e k 6 12k

k 2 3kz m=−∞ j=0

205
Let us use an abbreviation.
! ∞
h + iz 1 X − πm 2
η = √ e 12kz T (m) ,
k 2kz m=−∞
2K−1
1 X πi( j+ 2h ( j− 1 )2 + m (6 j−1))
where T (m) = e k 6 12k .
2 j=0
!  ∞

h + iz 1  
 X
− πm
2 


η = √  T (0) + e 12kz (T (m) + T (−m))
k 3kz 

m=1



This is a function in 1z . Also

! ∞
h + iz ǫ −1 X π 2 πih′ 2
η = √ (−)λ e− 12kz (6λ−1) e 12k (6λ−1)
k z λ=−∞
 
Now η h+iz
k has been obtained in two different ways. We have in both cases
a power series in e−π/(12zk) = x, both for |x| < 1. But an analytic function has
only one power series; so they are identical. This teaches us something. The
22. Lecture 154

second teaches us that by no means do all sequences appear in the exponent. 206
Only m2 = (6λ − 1)2 can occur. There is no constant term in the second expres-
sion. So m has the form |6λ − 1| = 6λ ± 1, λ > 0. Make the comparison; the
coefficients are identical. They are almost always zero. In particular T (0) = 0.
T (m) for m other than ±1 (mod 6) also vanish. So we have the following iden-
tification.
1 πih′
√ (T (6λ − 1) + T (−6λ + 1)) = ǫ −1 (−)λ e 12k (6λ − 1)2
3k
Realise that we have acknowledged here that a transformation formula ex-
ists. The root of unity ǫ is independent of λ. This we can assume but W.
Ruscher does not. Take in particular λ = 0. Then we have for m = ±1,
1 πih′
√ (T (−) + T (1)) = ǫ −1 e 12k
3k
This is proved by Fischer by using Gaussian sums. Therefore

′ 

− πih 2k−1 2k−1
X  3h 1 2 6 j−1  
e 12k 
 X 3h i 2 6 j−1


πi j+ k ( j− 6 ) + 6k 
ǫ −1 = √  πi( j+ k ( j− 6 ) )− 6k

 e + e 

3k  j=0

j=0



Now j matters only mod 2k. We can beautify things slightly:

′
 
− πih πih 
12k + 12k 

e  πi
X πi πi
X πi


k (3h j + j(k−h−1))
2 2
ǫ −1 = k (3h j + j(k−h−1)) + e− 6k
 6k 
√ 
e e e 

2 3k  j mod 2k

j mod 2k



The sum appears complicated but will collapse nicely; however compli-
cated it should be a root of unity. In Ak (n) the sums are summed over h and for
that purpose we shall not need to compute the sums explicitly.
Lecture 23

Last time we obtained the formula 207

1 πi(h−h′ ) πi
X πi

2

ǫ −1 = √ e 12k e− 6k e k 3h j j (k−h+1)
2 3k j mod 2k
1 πi(h−h′ ) πi
X πi 2
+ √ e 12k e 6k e k (3h j + j(k−h−1))
2 3k j mod 2k

ωh,k was defined by means of the equation

 h+iz  √ π  1   h′ +i/z 
f e2πi k = ωhk ze 12k z −z f e2πi k

ωhk came from the ǫ in the transformation formula

! r
aτ + b cτ + d
η =ǫ η(τ)
cτ + d i
In particular,
! !
h′ + 1/z √ h + iz
η = ǫ zη ,
k k
 
f e2πiτ = eπiτ/12 (η(τ))−1
Substituting in the previous formula,
( !)−1 ( !)−1
πi h + iz h + iz √ π 1 h′ +iz h′ + i/z
= ωhk ze 12k ( 3 −z) e 12
πi
e 12 η k η
k k k
! !
h′ + i/z √ πi ′ h + iz
i.e., η = ωhk ze 12k (h −h) η
k k

155
23. Lecture 156

πi ′
∴ ǫ = ωhk e 12k (h −h)
π ′
or ωhk = ǫe− 12k (h −h)

208
In the first formula we have obtained an expression for ǫ −1 . However, we
could make a detour and act ǫ directly instead of ǫ −1 . Even otherwise this could
be fixed up, for after all it is a root of unity. We have ǫ ǭ = 1 or ǫ = ǭ −1 . So
consistently changing the sign in the exponents, we have

πi ′ 1 πi
X πi 2
ωhk = ǭ −1 e 12k (h−h ) = √ e 6k e− k (3h j + j(k−h+1))
2 3k j mod 2k
1 πi
X πi 2
+ √ e− 6k e− k (3h j + j(k−h−1))
2 3k j mod 2k

We now have the ωhk that we need. But the ωhk are only of passing interest;
we put them back into Ak (n);
X′
Ak (n) = ωhk e−2πinh/k
h mod k

This formula has one unpleasant feature, viz. (h, k) = 1. But this would not
do any harm. We can use a lemma from an unpublished paper by Whiteman
which status that if (h, k) = d > 1, then
X πi 2
e− k (3h j + j(k−h±1)) = 0
j mod 2k

For proving Whiteman status put h = dh∗ , k = dk∗ and j = 2k∗ l + r, 209
0 ≤ 1 ≤ d − 1, 0 ≤ r ≤ 2k∗ − 1. Then
∗
X d−1 2k
X X −1
− πik (3h j2 + j(k−h±1)) πi ∗ ∗
1+r)2 +(2k∗ ℓ+r)(dk∗ −dh∗ ±1)))
e = e− dk∗ (3dh (2k
j mod 2k ℓ=0 r=0
∗
2k
X −1 d−1
X
πi 2
= e− k (3hr +r(k−h±1))
e∓2πiℓ/d ,
r=0 ℓ=0

and the inner sum = 0 because it is a full sum of roots of unity and d , 1.
This simplifies the matter considerably. We can now write

1 πi
X X πi 2 h
Ak (n) = √ e 6k e− k (3h j + j(k−h+1)) e−2πin k
2 3k h mod k j mod 2k
23. Lecture 157

1 πi
X X πi 2 h
+ √ e− 6k e− k (3h j + j(k−h−1)) e−2πin k
2 3k h mod k j mod 2k

Rearranging, this gives 210

1 πi
X πi
X 2πi j(3 j−1)
Ak (n) = √ e 6k e− k (k+1) j e− k (n+ k )h
2 3k j mod 2k h mod k
1 πi
X πi
X 2πi j(3 j−1)
+ √ e− 6k e− k (k−1) j e− k (n+ 2 )h
2 3k j mod 2k h mod k

The inner sum is equal to the sum of the kth roots of unity, which is 0 or k,
k if all the summands are separately one, i.e., if
j(3 j − 1)
n+ ≡0 (mod k)
2
Hence
r r
1 k πi X
j − πik j 1 k − πi X πi j
Ak (n) = e 6k (−) e + e 6k (−) j e k
2 3 j mod 2k
2 3 j mod 2k
j(3 j−1) j(3 j−1)
2 ≡−n
(mod k) 2 ≡−n
(mod k)

In the summation here we first take all j′ s modulo 2k (this is the first sieving
out), and then retain only those j which satisfy the second condition modulo k
(this is the second sieving out). Combining the terms,
r
1 k X n πi πi o
Ak (n) = (−) j e− 6k (6 j−1) + e 6k (6 j−1)
2 3 j mod 2k
j(3 j−1)
2 ≡−n (mod k)
r
k X π(6 j − 1)
= (−) j cos
3 j mod 2k
6k
j(3 j−1)
2 ≡−n(mod k)
211
This formula is due to A.Selberg. It is remarkable how simple it is. We shall
change it a little, so that it could be easily computed. We shall show that the
Ak (n) have a certain multiplicative property, so that they can be broken up into
prime parts which can be computed separately. Let us rewrite the summation
condition in the following way.

12 j(3 j − 1) ≡ −24n (mod 24k)

23. Lecture 158

i.e., 36 j2 − 12 j + 1 ≡ 1 − 24n (mod 24k)

i.e., (6 j − 1)2 ≡ ν (mod 24k)
where we have written ν = 1 − 24n. In the formula
r
1 k X n πi πi o
Ak (n) = (−) j e− 6k (6 j−1) + e 6k (6 j−1)
2 3 j mod 2k
j(3 j−1)
2 ≡−n (mod k)

replace j by 2k − j in the popint term (where j runs through a full system of

residues, so does 2k − j). Further, observe that we have now 212

(12k − 6 j − 1)2 ≡ (mod 24k)

2
i.e., (6 j + 1) ≡ (mod 24k)
Then
 
r 







1 k
 X πi
X πi


(−) j e− 6k (6 j+1) + (−) j (6 j−1) 

Ak (n) =  e 6k

2 3


 j mod 2k j mod 2k





 

(6 j−1)2 ≡ν (mod 24k) (6 j−1)2 ≡ν (mod 24k)

In both terms the range of summation is j mod 2k and there is the further
condition which restricts j. So
r
1 k X πi
Ak (n) = (−) j e− 6k (6 j±1)
2 3 j mod 2k
(6 j±1)2 ≡ν (mod 24k)

Write 6 j ± 1 = ℓ. 6 j ± 1 thus modulo 24k. j = ℓ+1 6 , so it is the integer

( )
ℓ
nearest to 6ℓ since (ℓ, 6) = 1. So write j = where {x} denotes the integer
6
nearest to x. Then
r
1 k X
(−){ 6 } e 6k
ℓ πiℓ
Ak (n) =
2 3 ℓ mod 2k
(ℓ,6)=1,ℓ2 ≡ν (mod 24k)

And one final touch. The ranges for ℓ in the two conditions are modulo 12k 213
and modulo 24k. Make these ranges the same. Then
r
1 k X
(−){ 6 } e 6k
ℓ πiℓ
Ak (n) =
4 3 ℓ mod 24k
ℓ2 ≡ν (mod 24k)
23. Lecture 159

We prefer the formula in this form which is much handler. We shall utilise
this to get the multiplicative property of Ak (n).
Lecture 24

We derived Selberg’s formula, and it looked in our transformation like this: 214
r
1 k X
(−){ 6 } e 6k ,
ℓ πiℓ
Ak (n) =
4 3 2
l ≡γ (mod 24k)

where ν = 1 − 24n, or ν ≡ 1 (mod 24). We write this Bk (ν); this is defined for
ν ≡ 1 (mod 24), and we had tacitly (ℓ, 6) = 1. We make an important remark
about the symbol (−){ 6 } . This repeats itself for ℓmodulo12. The values are
ℓ

ℓ= 1 3 7 11
(−){ } = 1
ℓ
6 −1 −1 1
{ ℓ
}
But (−) 6 can be expressed in terms of the Legendre symbol:
! !
{ ℓ
} ℓ −1
(−) = 6
3 ℓ
  ℓ−1
when (ℓ, 6) = 1. We can test this, noticing that −1 ℓ = (−1) 2 . Since 1, 7
are quadratic residues and 5, 11 quadratic non-residues modulo 3, we have for
ℓ
ℓ = 1, 5, 7, 11, (−){ 6 } = 1, −1, −1, 1 respectively; this agrees with the previous
ℓ
list. It is sometimes simpler to write (−){ 6 } in this way, though it is an after-
thought. It shows the periodicity.
Let us repeat the formula: 215
r ! !
1 k X ℓ −1 πiℓ
Bk (ν) = e 6k
4 3 2 3 ℓ
ℓ ≡ν (mod 24k)

This depends upon how k behaves with respect to 24. It has to be done
separately for 2, 3, 4, 6. For this introduce d = (24, k3). We have

160
24. Lecture 161

d = 1 if (k, 24) = 1,
3 if 3 | 4, k odd,
8 if k is even and 3 ∤ k
24 if 6 | k.
Let us introduce the complementary divisor e, de = 24. So e = 24, 8, 3 or
1. (d, e) = 1. Also (c, k) = 1.
All this is a preparation for our purpose. The congruence ℓ2 ≡ ν (mod 24k)
can be re-written separately as two congruences: ℓ2 ≡ ν (mod dk), ℓ2 ≡ ν
(mod e).
The latter is always fulfilled if (ℓ, 6) = 1. Now break the condition into two
subcases. Let r be a solution of the congruence
(er)2 ≡ ν (mod dk);
then we can write ℓ = er + dk j, where j runs modulo e and moreover ( j, e) = 1.
To different pairs modulo dk and e respectively belong different ℓ modulo 24k.
Bk (ν) can then be written as
r
1 k X X er + dk j ! −1
!
πi
Bk (ν) = e 6k (er+dk j)
4 3 2 j mod e
3 er + dk j
(er) ≡ν (mod dk)
( j,e)=1

Separating the summations, this gives 216

r
1 k X πiℓk
Bk (ν) = e 6k S k (r),
4 3 2
(er) ≡ν (mod dk)

where ! !
X′ er + dh j −1 πid j
S k (r) = e 6k
j mod e
3 er + dk j
We compute this now in the four different cases implied in the possibilities
d = 1, 3, 8, 24.
Case 1. d = 1, e = 24
! !
X′ k j −1 πi j
S k (r) = e6
j mod 24
3 kj
! !
k −1 X′  j  j−1 πi j
= (−) 2 e 6
3 k j mod 24 3
24. Lecture 162

There are eight summands, but effectively only four, because they can be
folded together.
! !
k −1 X′  j  π−1
S k (r) = 2 (−) 2 eπi j
3 k j mod 12 3
! !
h −1 n πi 5πi 7πi 11πi o
=2 e6 −e 6 −e 6 +e 6
3 k
ℓ
(We replaced the nice symbol (−){ 6 } by the Legendre symbol because we
did not know a factorisation law for the former. So we make use of one special
character that we know).
! ! !
k −1 π 5π
S k (r) = 4 cos − cos
3 k 6 6
! !
k −1 √
=4 3
3 k
   k−1
 
and since 3k 3k = (−) 2 ·1 = −1 k , this gives gives 217
!
√ 3
S k (r) = 4 3
k
Case 2. d = 3, e = 8.
! !
X′ 8r −1 πi2 j
S k (r) = e
j mod 8
3 3k j
 −r  −1 ! X′ −1 ! πi j
= e2
3 3k j mod 8 j
 r  −1 ! X′ −1 ! πi j
=2 e2
3 k j mod 4 j
 r  −1 !
=2 (i + i)
3 k
 r  −1 !
= 4i .
3 k
Case 3. d = 8, e = 3.
! !
X′ 8k j −1 4πi j
S k (r) = e 3
j mod 3
3 3r
24. Lecture 163
! ! X  
k −1 ′ j 4πi
= e3
3 r
j mod 3
3
! !
k −1  4πi 8πi 
= e 3 −e 3
3 r
! !
k −1 2π
= −2i sin
3 r 3
! !
1 √ k −1
= 3
i 3 r
218
Case 4. d = 24, e = 1.
! ! !
k −1 3
S k (r) = =
3 r r

Now utilise these; we get a handier definition for Ak (n).

Case 1.
! √k
3 X 4πir
Bk (ν) = e k
k
(24r)2 ≡ν (mod k)

Case 2. r !
k −1 X r 4πir
Bk (ν) = i e 3k
3 k 3
(8k)2 ≡ν (mod 3k)

The i should not bother us because r and −r are solutions together, so they
combine to give a real number.
r ! r
k −1 X 4πr
Bk (ν) = − sin
3 k 2
3 3k
(8r) ≡v (mod 3k)

Case 3.
! !
1 √ k X −1 πir
Bk (ν) = k e 3k
4i 3 r
(3k)2 ≡ν (mod 8k)
! !
1√ k X −1 πr
= k sin
4 3 2
r 2k
(3r) ≡ν (mod 8k)
219
24. Lecture 164

Case 4. r !
1 k X 3 πir
Bk (ν) = e 6k
4 3 r
r2 ≡ν (mod 24k)

This is the same as the old definition.

This makes it possible to compute Ak (n). We break k into prime factors and
because of the multiplicative property which we shall prove, have to face only
the task of computing for prime powers.
Lecture 25

We wish to utilise the formula for Bk (ν) that we had: 220

r ! !
1 k X ℓ −1 πiℓ
Ak (n) = Bk (ν) = e 6k ,
4 3 2 3 ℓ
ℓ ≡ν (mod 24k)

with ν = 1 − 24n (and so ≡ 1 modulo 24). Some cases were considerably

simpler. Writing d = (24, k3), de = 24, we have four cases: d = 1, 3, 8, 24.
d=1 !
3 √ X
Bk (ν) = k e4πir/k
k
(24r)2 ≡ν (mod k)

d=3 !r
−1 k X r
Bk (ν) = 2i e4πir/3k
k 3 3
(8r)2 ≡ν (mod 3k)

d=8 ! !
1 k √ X −1 πir/2k
Bk (ν) = k e
4i 3 r
(8r)2 ≡ν (mod 8k)

d = 24
There is nothing new; we get the old formula back.
We wish first to anticipate what we shall use later and get An (n) for prime
powers which will be the ultimate elements. Again we have to discuss several
cases.
First take k = pλ , p a prime exceeding 3. Then, by case 1 above (since 221
(24, k3 ) = 1),
!λ
3 X λ
Bk (ν) = pλ/2 e4πir/p
p 2 λ
(24r) ≡ν (mod p )

165
25. Lecture 166

Look into the condition of summation. It is quite clear that this implies
(24r)2 ≡ ν (mod p) i.e., ν is a quadratic residue modulo p. Hence
!
v
B p λ(ν) = 0 if = −1. (1)
p

On the other hand, if x2 ≡ ν (mod p) is solvable, then x2 ≡ ν (mod pλ )

is also solvable (we take for granted the structure of the cyclic residue group).
x2 ≡ ν (mod pλ ) has two solutions, and now we want only x = 24r (mod pλ ).
Let r be a solution, −r is the other solution: (24r)2 ≡ (mod pλ ). Then
!λ
3 n λ λ
o
Bk (ν) = pλ/2 e4πir/p + e−4πir/p
p
!λ
3 4πr
=2 pλ/2 cos λ (2)
p p
p
This is roughly of the order of pλ
Next, suppose that p/ν. This is a special case of pλ /ν. Then (24r)2 ≡ 0
(mod pλ ), and the solutions are
λ+1
r = p[ 2 ] · j,
λ+1
j = 0, 1, 2, . . . , pλ−[ 2 ] − 1.
h i
λ+1
when λ = 1, 2 = λ and we have only one summand. Hence 222
!
3 1/2
Bk (ν) = p (3)
p
Now let λ > 1. Then
!λ p [ ]
λ− λ+1
2
3 λ
X λ+1
Bk (ν) = p 2 e4πi j/p[ 2 ]
p j=1

This again involves two cases, λ even and λ odd. If λ is even, λ = 2µ and
the sum becomes
X pµ
µ
e4πi j/p
j=1

and this is 0, being a full sum of roots of unity. Hence in this case

Bk (ν) = 0 (4)
25. Lecture 167

Now let λ be odd: λ = 2µ + 1.

r = pµ+1 · j, j = 0, 1, . . . , pµ − 1.

Then the sum becomes µ

p
X µ
e4πi j/p
j=1

which is again zero; hence

Bk (ν) = 0 (5)
Now suppose that pµ | ν, µ < λ and pλ ∤ ν. r2 ≡ ν (mod pλ )ν = pµ ν, p + ν, 223
or ν1 ≡ pµ ν (mod pλ ) ν = pµ ν1 , p ∤ ν1 ; or ν2 ≡ pµ ν1 (mod pλ ). If is odd,
µ < λ, then pµ | ν; and again
Bk (ν) = 0 (6)
There remain the case in which µ is even, µ = 2ρ. Then r2 ≡ p2ρ ν,
(mod pλ ). Writing r = pρ j, p2ℓ j2 ≡ p2ρ ν1 (mod pλ ), or j2 ≡ ν1 (mod pλ−2ρ )
If νp1 = −1, then again
Bk (ν) = 0 (7)
ν 
However p1 = 1 implies j2 ≡ ν1 (mod pλ−2ρ ) has two solutions, j and
− j. Then
 
r ≡ pρ j + ℓpλ−2ρ (mod pλ )
or τ r ≡ pρ j + ℓpλ−ρ (mod pλ )
where ℓ = 0, 1, . . . , pρ − 1.

Then the sum becomes

ρ ρ
pX −1 pX −1
4πi ± p4πi
λ −ρ j
4πi
ρ
e (±p j + ℓp
pλ λ−ρ
)=e e pρ ℓ
ℓ=0 ℓ=0
=0

Again
Bk (ν) = 0 (8)
We now take up the case p = 3. This corresponds to p = 3. If k = p = 3λ , λ
224

λ−1
X r
λ+1
B3 λ(ν) = i(−)λ 3 2 e4πir/3
2 λ+1
3
(8r) ≡ (mod 3 )
25. Lecture 168
 
ν
ν ≡ 1 (mod 24) or ν ≡ 1 (mod 3). So 3 = 1. There are two solutions, r and
   
2
−r for the congruence (8r) ≡ ν (mod 3 λ+1
). Since −r
3 =− 3 ,
r

rλ−1
 4πir 4πir

B3λ (ν) = i(−)λ 3 2 e 3λ+1 − e− 3λ+1
3
  λ−1
λ+1 r 4πr
= 2(−) 3 2 sin λ+1 (9)
3 3
Finally, we take p = 2; then d is 8. Let k = 2λ . Then
!
1 λ λ/2
X −1 4πir/2λ+1
B2 λ(ν) = (−) 2 e
4i 2 λ+3
r
(3r) ≡ν (mod 2 )

ν ≡ 1 (mod 8) implies that (3r2 ) ≡ ν (mod 8) has four solutions, and these
solutions are inherited by the higher powers of the modulus. The solutions are
r ≡ 1, 3, 5, 7 (mod 8). In general the congruence x2 ≡ ν (mod 2µ ), µ ≥ 3 has
four solutions
±r + h2µ−1 , h = 0, 1
Then 225

o −1 !
1 n λ+1 λ+1 λ+1 λ+1
B2λ (ν) = (−)λ 2λ/2 e4πir/2 − e−4πir/2 + e4πir/2 − e−4πir/2
4i r
  r−1
and since −1
r = (−)
2 ,

!
−1 4πr
B2λ (ν) = (−)λ eλ/2 sin λ+! (10)
r 2

We have thus computed the fundamental cases explicitly.

Lecture 26

We had the formula for Bk (ν): 226

r ! !
1 k X ℓ −1 πiℓ/6k
Bk (ν) = e ,
4 3 3 3
ℓ2 ≡ν (mod 24k)

with ν ≡ 1 (mod 24). Writing d = (24, k3 ), we had the following cases:

1) d = 1 !
3 √ X
Bk (ν) = k e4πir/k
k
(24π)2 ≡ν (mod k)

2) d = 3
!r !
−1 k X −1 πir/2k
Bk (ν) = i e
k 3 r
(24r)2 ≡ν (mod 3k)

3) d = 8 ! !
1 k √ X −1 πir/2k
Bk (ν) = k e
4i 3 r
(3r)2 ≡ν (mod 8k)

4) d = 24. We do not get anything new.

Assume k = k1 k2 , (k1 , k2 ) = 1. We desire to write Bk (ν1 ). Bk2 (ν2 ) = Bk (ν),

with a suitable ν to be found out from ν1 and ν2 . It cannot be foreseen. It is a
multiplication of a peculiar sort. Two cases arise.
(i) At least one of k1 , k2 is prime to 24 and therefore to 6, say (k1 , 6) = 1. 227

(ii) None is prime to 6. But since (k1 , k2 ) = 1, 2/k1, 3/k1 . Under the circum-
stances prevailing these are the two mutually exclusive cases.

169
26. Lecture 170

Case 1. Utilise d = 1.
! r
3 p 1 k2 X
Bk1 (ν1 ) · Bk2 (ν2 ) = k1 · e4πir/k1
k1 4 3
(24r)2 ≡ν 1 (mod k1 )
! !
X ℓ−1 πiℓ/6k2
· e
3 ℓ
ℓ2 ≡ν2 (mod 24k2 )
!r ! !
1 3 k1 k2 XX πi
6k1 k2 (24k2 r+k1 l)
l −1
= e
4 k1 3 2
3 l
(24r) ≡ν1 (mod k1 )
ℓ2 ≡ν2 (mod 24k2 )

k1 and 24k2 are coprime moduli. If r runs modulo k1 and ℓ runs modulo 24k2 ,
24k2 r + k1 ℓ would then run modulo 24k1 k2 .
Write
24k2 r + k, ℓ ≡ t (mod 24k1 k2 )
Then
t2 = (24k2 + k1 ℓ)2 ≡ (24k2 r)2 (mod k1 )
≡ k22 ν1 (mod k1 ), since (24r)2 ≡ ν1 (mod k1 )
Similarly
t2 ≡ (k1 ℓ)2 (mod 24k2 )
≡ k12 ν2 (mod 24k2 ), since ℓ2 ≡ ν2 (mod 24k2 ).
So in order to get both conditions of summation, we need only choose 228
t2 ≡ ν (mod 24k1 k2 ); and this can be done by the Chinese remainder theorem.
So !r ! !
1 3 k X ℓ −1 πit/6k
Bk1 (ν1 )Bk2 (ν2 ) = , e
4 k 3 2 3 ℓ
t ≡ (mod 24k,k2 )

This already looks very much like the first formula though not quite. What
we have in mind is to compare it with
r  t  −1 !
1 k X
Bk (ν) = eπit/6k
4 3 2 3 t
t ≡ν (mod 24k)

So find out
 t  −1 ! 24k2 r + k1 ℓ
!
−1
!
=
3 t 3 24k2 r + k1 ℓ
26. Lecture 171
! !
k1 ℓ −1
=
3 k1 ℓ
! ! ! !
k1 −1 ℓ −1
=
3 k1 3 ℓ
! ! !
3 ℓ −1
= ,
k1 3 ℓ
by the reciprocity law. So the formulas agree: Bk1 (ν1 )Bk2 (ν2 ) = Bk (ν); and we
have settled the affair in this case by
Theorem 1. If k22 ν1 ≡ ν (mod k1 ) and k12 ν2 ≡ ν (mod 24k2 ), (k, 6) = 1, then 229

Bk1 (ν1 )Bk2 (ν2 ) = Bk1 k2 (ν)

Case 2. This corresponds to d = d1 = 8 and d = d2 = 3.
! !r
1 k1 p −1 k2
Bk1 (ν1 ) · Bk2 (ν2 ) = k1
4 3 k2 3
!
X −1 πir/2k1 X
e eπis/3k2 .
r
(3r)2 ≡ν1 (mod 8k1 ) (8r)2 ≡ν2 (mod 3k2 )
! !r
1 k1 −1 k1 k2
=
4 3 k2 3
!
XX −1  s  6kπik (3k2 r+8k1 s)
e 12
2
3 3
(3r) ≡ν1 (mod 8k1 )
(8r)2 ≡ν2 (mod 3k2 )

Since (k1 , k2 ) = 1, (8k1 , 3k2 ) = 1 and so 3k2 r + 8k1 s = t runs through a full
system of residues modulo 24k1 k2 . So
! !r !
1 k1 −1 k X −1  s  πit/(6k1 k2 )
Bk1 (ν1 )Bk2 (ν2 ) = e
4 3 k2 3 2 r 3
t ≡ν (mod 24k1 k2 )

As before 230

t2 = (3k2 r + 8k1 s)2 ≡ (3k2 r)2 ≡ (3k2 r)2 ≡ k22 ν1 (mod 8k1 )
2 2
t = (8k1 s) ≡ k12 ν2 (mod 3k2 )
Now determine ν such that ν ≡ k22 ν1 (mod 8k1 ) and ν ≡ k12 ν2 (mod 3k2 ),
again by the Chinese remainder theorem. So t2 ≡ (mod 24k1 k2 ). Now
 t  −1 ! !
8k1 s −1
!
=
3 t 3 3k1 r
26. Lecture 172
! ! !
k1 −1  s  −1
=
3 k2 3 r
(since 8 and −1 are quadratic non-residues modulo 3). So
r  t  −1 !
1 k X
Bk1 (ν1 )Bk2 (ν2 ) = eπit/6k
4 3 2 3 t
t ≡ν (mod 24k)

= Bk (ν)

where ν is given. Hence

Theorem 2. If k22 ν1 ≡ ν (mod 8k1 ) and k12 ν2 ≡ ν (mod 3k2 ), then

Bk1 (ν1 )Bk2 (ν2 ) = Bk1 k2 (ν)

Let us give an example of what this is good for. Calculate A10 (26). Since
we can reduce modulo 10, A10 (26) = A10 (6).

ν = 1 − 24n = −143.
A10 (26) = A10 (6) = B10 (−143) = B10 (−23)
= B5 (ν1 )B2(ν2 )

where ν1 , ν2 are determined by the conditions 231

4ν1 ≡ −23 (mod 5) or − ν1 ≡ −3 (mod 5)

and 25ν2 ≡ −23 (mod 48) or ν2 ≡ 1 (mod 48)
 
3
So A10 (26) = B5 (3)B2(1), and these are explicitly known. Since 5 = −1,
B5 (3) = 0. It is actually not necessary now to calculate B2 (1).
!
λ −1 πr
B2 (1) = (−) 2λ/2 sin λ+1
r 2
where (3r)2 ≡ ν (mod 2λ+3 ), (3r)2 ≡ 1 (mod 16),

or 3r ≡ 1 (mod 16), r ≡ 11 (mod 16). (there being four solutions). Then

√ 11π √ 1
B2 (1) = (−)(−) 2 sin =1× 2· √ =1
4 2
A10 (26) = 0.

One more thing can be established now. We have the inequalities:

|B2λ (ν)| ≤ 2λ/2 ,

26. Lecture 173

λ √
|B3λ (ν)| ≤ 3 2 2 3,
λ
|B pλ (ν)| ≤ 2p 2 , p > 3.

By the multiplicative property, 232

√ √ λ(k)
|Bk (ν)| = |Ak (ν)| ≤ k(2 3)
X
where λ(k) = 1.
p|k

This is a rough appraisal, but λ(k) is in any case a small number. So

√
|Bk (ν) < C k · kǫ , ǫ > 0, C = Cǫ .

We see that although An (n) has ϕ(k) summands and in general all that one
knows is that ϕ(k) ≤ k − 1, because of strong mutual cancellations among the
1
roots of unity, the order is brought down to that of k 2 +ǫ . This reminds us of
other arithmetical sums like the Gaussian sums and the Kloosterman sums.
Lecture 27

We now give a proof of the transformation formula for η(τ). η(τ) we first in- 233
troduced by Dedekind in his commentary on a fragment on modular functions
by Riemann; it is natural in the theory of elliptic functions.
∞
Y
πiτ
η(τ) = e 12 (1 − e2πimτ )
m=1

aτ + b
We want to replace τ by τ′ = . Actually in the whole literature there
cτ + d
is no full account except in a paper by W.Fischer (Pacific Journal of Mathe-
matics, Vol. 1). We know what happens in the special cases − τ1 and τ + 1. We
get the explicit form in which the root of unity appears in the transformation
formula if we put together some things from the theory of modular functions.
There some discussion in Tannery-Molk; they write h(τ) instead of η(τ). (η(τ))3
is up to a factor V11 (o/τ). It turns out for quite other reasons that (η(τ))8 can
be discussed too; it has to do with the modular invariant J(τ). Dedekind did
something more than what is needed here. He studied log η(τ). For Im τ > 0,
η(τ) is a function in the interior of the unit circle (if we set x = e2πiτ ) free from
zeros and poles. So the logarithm has no branch points and is fully defined
without ambiguity.
∞
πiτ X
log η(τ) = + log(1 − e2πimτ )
12 m−1

(For purely imaginary τ, the logarithms on the right side are real).
The multiplicative root of unity now appear as something additive. This 234
is what Dedékind investigated. Recently (Mathematika, vol.1, 1954) Siegel
1
published a proof for the particular case − , using logarithms. Actually Siegel
τ

174
27. Lecture 175

proves much more than the functional equation for η(τ). He proves that
1 τ
log η(−τ−1 ) = log η(τ) + log
2 i
We shall extend his proof to the more general case. The interesting case
where a root of unity appears explicitly has not been dealt with by Siegel.
We write the general modular transformation in the form
h + iz ′ h′ + i/z
τ= ,τ = , hh′ ≡ −1 (mod k)
k k
We wish to prove that
! !
h′ + i/z h + iz 1
log η = log η + log z + πiC(h, k) (*)
k k 2

where C(h, k) is a real constant.

From the definition of η(τ),
! ∞ ∞
h + iz π(h + iz) X X 1 2πimr(h+iz)/k
log η = − e
k 12k m=1 r=1
r
∞ ∞
πih πz X X 1 2πimrh/k −2πmrz/k
= − − e e
12k 12k m=1 r=1 r

e2πimrh/k is periodic with period k; we emphasize this and write 235

m = qk + µ; µ = 1, . . . , k; q = 0, 1, 2, . . . .

Then
! ∞ k ∞
h + iz πih πz X X X 1 2πiµ rh −2π(1k+µ) rz
log η = − − e k e k ,
k 12k 12k q=0 ν=1 r=1 r

and taking the summation over q inside, this becomes

∞ ∞ ∞
πih πz X X 1 2πiµ rh −2πµ rz X −2πqrz
− − e k e k e
12k 12k µ=1 r=1 r q=0
k ∞
πih πz X X 1 2πiµ rh e−2πµrz/k
= − − e k
12k 12k µ=1 r=1 r 1 − e−2πrz
27. Lecture 176
 h′ +i/z 
Substituting in (*), with similar expansion for η k , we have

k ∞ 2πνr
πih π XX 1 h′ e− kz
− − e2πiνr k
12k 12kz ν=1 r=1 r 1 − e−2πr/z
k ∞ rz
1 πih πz X X 1 2πiµ rhk e−2πµ k
= log z + πiC(h, k) + − − e
2 12k 12k µ=1 r=1 r 1 − e−2πrz

Rearranging this, we get 236

∞
k X k ∞
X 1 h′ e−2πνr/kz XX 1
2πiµrh/k e
−2πµrz/k
e2πiνr k . − e
ν=1 r=1
r 1 − e−2πr/z µ=1 r=1 r 1 − e−2πr/z
!
π 1 πi 1
+ −z + (h − h′ ) + πiC(h, k) = − log z.
12k z 12k 2
We now follow Siegel’s idea to get the whole thing as a sum of residues of
a certain function. Clearly there is r in it. Being integers r can be produced
1 1
by something like which has poles with residue − 2πi at every integral
1 − e2πix
valued x. So let us study a function like
−2πµxz/k
1 1 2πiµxh/k e
e
x 1 − e2πix 1 − e−2πxz
We may have to sum this from µ = 1 to µ = k. This should somehow be
the form of the function that we wish to integrate. We do not want it in the
whole plane. In fact, we can either take a wider and wider path of integration,
or multiply the function by a factor and magnify it; we prefer to do the latter. 237
We shall put xN for x, keep the path fixed and take N = n + 12 , n integer,
to avoid integral points, and then make n → ∞. The term corresponding to
µ = k should be treated separately, as otherwise the factor e−2πxz would stop
convergence. Also µh and µ should appear symmetrically for reasons which
we shall see. So introduce µ∗ ≡ µh (mod k), µ = 1, 2, . . . , k − 1, and choose
1 ≤ µ∗ ≤ k − 1. It turns out, taking all this together, that the following thing
will do. Write
k−1
1 πN x X 1 e2πµN x/k e−2πiµN x/kz
Fn (x) = − cot hπN x cot + · ·
4ix z µ=1
x 1−e 2πN x 1 − e−2πiN x/z

The first term is a consequence of the term for µ = k:

1 e2πN xi e−2πxNz
× ×
x 1 − e2πixN 1 − e−2πxNz
27. Lecture 177

The poles will not change if we write this as

! !
1 e2πN xi 1 e−2πxNz 1 1 1 + e2πiN x 1 + e−2πxNz
+ −2πxNz
+ = ·
x 1−e 2πiN x 2 1−e 2 x 2(1 − e 2πiN x ) 2(1 − e−2πxNz )
1
= cot πxN · cot hπxNz.
4xi
We integrate Fn (x) along a certain parallelogram P, a little different from 238
Siegel’s. P has vertices at ±z, ±i (since Jm τ > 0, Re z > 0). Then
Z
1 X
Fn (x)dx = (Residues).
2πi p

We then let n → ∞.
The poles of Fn (x) are indicated by the denominators and the cotangent
factors. These are
rz ir
x = 0, x=− , x= , r integer.
N N
x = 0 is a triple pole for the first summand.

1 πN x 1 1 z
− cot hπN x cot =− ·
4ix z 4ix πN x πN x
( ) ( )
(πN x)2 (πN x/z)2
1+ +· × 1− +···
3 3
27. Lecture 178

Residue for this term at x = 0 239

!
iz 1 2 2 π2 N 2
= · π N −
4π2 N 2 3 z
!
i 1
= z− .
12 z

which had been foreshadowed already.

Lecture 28

We had 240

k−1 ∗
1 πN x X 1 e2πµN x/k e−2πiµ N x/kz
Fn (x) = − cot hπN x cot + · × ,
4ix z µ=1
x 1−e 2πN x 1 − e−2πiN x /z

N = n + 12 , n integer > 0, µ∗ ≡ hµ (mod k) and 1 ≤ µ∗ ≤ k − 1. At the triple

 
1
pole x = 0 the residue from the first summand = − 12i z − 1z - Let us find the
residues from the more interesting pieces of the sum. The general term on the
right has in the neighbourhood of x = 0 the expansion
( )
1 2πµN x (2πµN2 /k)2
1+ + +···
x k 2!
( )−1
−1 2πN x (2πN x)2
× 1+ + +···
2πN x 2 6
( )
2πiµ∗ N x (2πµ∗ N x/kz)2
× 1− − +···
kz 2
( )−1
1 2πiN x (2πN x/z)2
× 1− − + ···
2πiN x/z 2z 6
 !2 
−z 

 2πµN x 1 2πµN x 


= 2 2 3 1+ + − ···
4π iN x   k 2 k 

( 2
! )
2πN x (2πN x) 2
× 1− + + · · · + (· · · ) + · · ·
2 6
 !2 


 2πiµ∗ N x 1 2πµ∗ N x 


× 1− − + ···

 kz 2 kz 


179
28. Lecture 180
( )
2πiNz (2πN 2 /z)2
× 1+ + +···
z 2
241
1
Fishing out the term in , the residue at x = 0 from this summand becomes
x
 !2 !2 !2
iz   1 2πµN
 1 2 1 2πµ∗ N 1 2πN N
 + (2πN) − − − 2πµ πN
4π2 N 2 
2 k 12 2 kz 12 z k
2 2 2 2 ∗ 2 2 2 2 ∗ 2
)
N i 2π iµN 2π iµ N π iN 2π µ N
−4π2 µµ∗ 2 + + − +
k z kz kz z kz2
( )  2

iz 2 2 1 2µ  −2µ∗
i 
 1 2µ∗  

= k + − +  − + 
4 µ2 3 k  k2
4z  3 k 
( )
i 4iµµ∗ 2iµ 2iµ∗
+ − 2 + + −i
4 k k k
( 2 )  2  ! !
∗ ∗
µ µ 1 1 µ
 µ 1 
 µ 1 µ∗ 1
= iz − + +  − + + − − (*)
2k2 2k 12  2k2 2k 12 
iz   k 2 k 2

We have to sum this up from µ = 1 to µ = k − 1. Let us prepare a few 242

things.
Let us remark that
k−1 k−1
X (k − 1)k X (k − 1)k(2k − 1)
µ= ; µ2 =
µ=1
2 µ=1
6

Also if µ runs through∗ a full system of residues, so would µ∗ because

µ µ∗
(h, k) = 1. Further 0 < k < 1, and k and hµ k differ only by an integer, so
∗ h hµ i
that µk = hµk − k . Hence summing up the last expression (∗) from µ = 1 to
µ = k − 1, we have
( )
(k − 1)(2k − 1) k − 1 k − 1
ız − +
12k 4 12
( ) X k−1 ! " # !
1 (k − 1)(2k − 1) k − 1 k − 1 µ 1 hµ hµ 1
+ − + + − − −
iz 12k 4 12 µ=1
k 2 k k 2
! !
2k − 1 1 1
= (k − 1) − iz + + s(h, k)
12k 6 iz
28. Lecture 181

where s(h, k) stands for the arithmetical sum

k−1 ! " # !
X µ 1 hµ hµ 1
− − −
µ=1
k 2 k k 2

which appears here very simply as a sum of residues. The last expression
becomes !
k−1 1
− iz + + s(h, k)
12k iz
So the total residue at x = 0 is 243
! ! !
1 1 k−1 1 1 1
iz + − iz + + s(h, k) = iz + + s(h, k)
12 iz 12k iz 12k iz

Next, we consider the simple poles of Fn (x) at the points x = Nir (r , 0).
The coth factor is periodic and so the residue at any of these poles is the same
as that at the origin, which is π1 . Hence the residue of Fn (x) at x = Nir (r , 0)
becomes
k−1 ∗
N 1 πir X N −1 2πiµ r e2πµ r/kz
· cot + e k
4r πN z µ=1
ir 2πN 1 − e2πr/z

(There is a very interesting juxtaposition of an arithmetical term and a func-

tion theoretic term in the last part; this gets reversed for the next set of poles)
k−1 ∗
1 πr 1 X 1 2πi µr e2πµ r/kz
= coth − e k
4πir z 2πi µ=1 r 1 − e2πr/z
 
x remains between ±i on the imaginary axis. So  Nr  < 1; so we need consider
only r = ±1, ±2, . . . , ±n. Again,

ey + e−y 2e−y
coth y = −y
=1+ y
y
e −e e − e−y
−2y
2e
=1+
1 − e−2y
coth y is an odd function so that 1y cot h y is even. Hence summing up over all 244
the poles corresponding to r = ±1, . . . , ±n, we get the sum of the residues

n ( ) k−1 n ∗
1 X1 2e−2πr/z 1 X X 1 2πih′ µ∗ r/k e−2πµ r/kz
= 1+ + e
2πi r=1 r 1 − e−2πr/z 2πi µ∗ =1 r=1 r 1 − e−2πr/z
28. Lecture 182

k−1 n ∗
1 X X 1 2πih′ (k−µ∗ )r/k e2πµ r/kz
− e ,
2πi µ∗ =1 r=1 r 1 − e2πr/z

where we have made use of the fact that hh′ ≡ −1 (mod k), so h′ µ∗ ≡ hh′ µ ≡
−µ (mod k), or µ ≡ −h′ µ∗ (mod k). In the last sum replace µ∗ by k − µ∗ ; then
the previous sum is duplicated and we get
 
n  k−1 n ∗ 
1 X1 
 2e−2πr/z 1 X X 1 2πih′ µ∗ r/k e−2πµ r/kz 


 1 + −2πr/z
+ e −2πr/z 
2πi r=1 r 

 1−e πi µ∗ =1 r=1 r 1−e 


n k n
1 X 1 1 X X 1 2πih′ νr/k e−2πνn/kz
= + e
2πi r=1 r πi ν=1 r=1 r 1 − e−2πr/z

This accounts for all the poles on the imaginary axis (except the origin 245
which has been considered separately before).
Finally we have poles x = rz
N (e , 0) on the other diagonal of the parallelo-
gram. The same calculation goes through verbatim and we get the sum of the
residues at these poles to be
n k n
i X 1 i X X 1 2πihνr/k e−2πνrz/k
+ e
2π r=1 r π ν=1 r=1 r 1 − e−2πrz
Lecture 29

We had 246
k−1 ∗
1 πN x X 1 e2πµN x/k e−2πiµ N x/kz
Fn (x) = − cot hπN x cot +
4ix z µ=1
x 1 − e2πN x 1 − e−2πN x/z

The residue at x = 0 is
!
1 1
iz + + s(h, k),
12k iz
s(h, k), which will interest us for some time, being
k−1 ! " # !
X µ 1 hµ hµ 1
− − − .
µ=1
k 2 k k 2

ir
The residues at the points x = N (r , 0) amount to
n k n
1 X 1 1 X X 1 2πih′ ν r e−2πνr/kz
+ e k ;
2πi r=1 r πi ν=1 r=1 r 1 − e−2πr/z
zr
and the residues at the points x = N (r , 0)
n k n
i X 1 i X X 1 2πihµ r e−2πµrz/k
+ e k
2π r=1 r π µ=1 r=1 r 1 − e−2πrz
P
When we add up, the sums nr=1 1r , the disagreeable ones which would have 247
gone to infinity, fortunately destroy each other; so the sum of the residues of
Fn (x) at all its poles becomes

183
29. Lecture 184

! k n
1 1 1 X X 1 2πih′ νr/k e−2πνr/kz
− z + s(h, k) + e
12ki z πi ν=1 r=1 r 1 − e−2πr/z
k n
1 X X 1 2πihµr/k e−2πµrz/h
− e
πi µ=1 r=1 r 1 − e−2πrz

We had prepared in advance what we were going to obtain. s(h, k) is what

we had called C(h, k) + (h − h′ )/12k. We have to prove that the sum of the
′
residues above, with C(h, k) = s(h, k) − h−h 1
12k , is equal to − 2πi log z, as n → ∞.
But there is one difference. The sums we have earlier were sums from r = 1
to r = ∞; whereas here they are sums from r = 1 to r = n. But this does not
matter as convergence is guaranteed since we have an exponential factor e−z
with Re z > 0. We have to see what becomes R of our sum when we evaluate it
in another way. We have to consider lim p Fn (x)dx. So in effect we have to
n→∞
prove that Z
1 1
lim Fn (x)dx = − log z.
n→∞ 2πi p 2πi
248
Now this is a question of direct
computation. Let us look at the
path of integration. Fn (x) will
be seen to have simple limits on
the sides of the parallelogram.
We consider xFn (x) broken into
pieces. Take the first piece
1 πN x
cot hπN x cot
4i z
On the side from x = i to x = z,

x = ρi + σz; ρ, σ ≤ 0, ρ + σ = 1.
Actually we take only ρ, σ > 0; we shall exclude the points i and z them-
selves. Then this becomes
1 eπN(ρi+σz) + e−πN(ρi+σz) eπiN(ρi+σz)/z + e−πiN(ρi+σz)/z
− × i ×
4i eπN(ρi+σz) − e−πN(ρi+σz) eπiN(ρi+σz)/z − e−πiN(ρi+σz)/z
The size of the first factor is determined by the terms eπNσz and e−Nπσz
in the numerator; the first term becomes big and the other goes to zero as
N → ∞(σ > 0 and Re z > 0). So we divide by the first term. Similarly for the
29. Lecture 185

second factor. We therefore get

ρ 
σ
1 1 + e−2πN(ρi+σz) e−2πN z + i + 1
− ·  
4 1 − e−2πN(ρi+σz) e−2πN ρz + σi − 1
249
As N → ∞ the exponential factors go to zero; so the whole expression
tends to 14 . It will further remain on its way bounded, because the numerators
in either factor are at most equal to 2, while the denominators remain away
from zero by a fixed amount, as we shall be showing in a moment - and for this
it is essential to have N = n + 12 .
Since the functions concerned are even functions, what was good here
would also be good on the apposite side, from x = −i to x = −z. So on
this side also the expression will tend to 14 . We cannot say uniformly; indeed
if σ = 0, here is no convergence in the first factor, and if ρ = 0 none in the
second factor, though there is boundedness: the thing would oscillate finitely.
Now take the other pieces of xFn (x) on the same sides of ρ. We have to
consider ∗N
N
e2πµ k (ρi+σz) e−2πiµ kz (ρi+σz)
×
1 − e2πN(ρi+σz) 1 − e−2πi Nz (ρi+σz)
Remember, what is now important, that 0 < µ < k, but neither 0 nor k.
The denominator in the first factors goes more strongly to infinity as N → ∞
than the numerator because µk is a proper fraction; so too in the second factor
because µ > 1. So the whole function tends to zero. Hence on these two sides
xFn (x) → 41 .
Now consider the other two sides; it looks different here and has got to be 250
inspected. On the side from x = −i to x = z, x = −ρi + σz; σ, ρ > 0, σ + ρ = 1,
and the first part of xFn (x) is

1 πN x 1 eπN(−ρi+σz) + −e−πN(−ρi+σz)
− cot hπN x cot = − πN(−ρi+−σz)
4i z 4e − e−πN(−ρi+σ−z)
 −ρi   −ρi 
eπiN z +σ
+ e−πiN z +−σ
×  −ρ   ρi

eπiN z
i+σ
− e −πiN − z +σ
ρi
1 1 + e−2πN(−ρi+σz) 1 + e−2πiN − z +σ
=− −2πN(−ρi+σz)
−×  ρ−i 
41−e 1 − e−2πiN − z +σ
Let N → ∞. Assuming that the denominator is going to behave decently,
this goes to − 41 . The other pieces go to zero for the same reason as before. And
all this is good for the opposite side too.
29. Lecture 186

We now have got to show that the convergence it nice and the denominators
do not make any fuse. This we can clarify in the following way. Consider the
denominator 1 − e−2πN(ρi+σz) .
Difficulties will arise if the exponent comes close to an even multiple of πi.
So we should see that it stays safely away from these points.

251

And actually it stays away from the danger spots by the same distance, for 251
the exponent is −2N(πiρ + πzσ) i.e., a point on the segment joining (2r + 1)πi
and (2r + 1)πz. Since ez is periodic there is a minimal amount by which it
stays away from 1. The second denominator looks a little different. We have
π
z instead of πz. But we have only to turn the whole thing around. We see how
essential it was to take N = n + 21 = (2n + 1) 21 = on odd multiple of 21 .
So the convergence is nice, but not uniform. We can nevertheless say that
xFn (x) → ± 14 boundedly on the sides of ρ except for the vertices where it does
not converge but oscillates finitely. But bounded convergence is enough for
1
interchanging integration and summation. Fn (x) → ± 4x and the x does not
ruin anything because it stays away from zero everywhere on ρ. Hence
Z
1
lim Fn (x)dx
n→∞ 2πi p

exists and we have

Z Z
1 1 1
lim Fn (x)dx = ± dx
n→∞ 2πi p 2πi p 4x
 

Zi Z −z Z −i Z z  
1 
 dx dx dx dx 


=  − + − 
2πi 


z 4x i 4x −z 4x −i 4x 



(Z i Z z Z i Z z )
1 dx dx dx dx
= − + −
8πi z x −i x z x −i x
29. Lecture 187
(Z i Z z )
1 dx dx
= −
4πi z x i x

z is in the positive half-plane; we can take the principal branch of the logarithm, 252
so that we get on integration, since log i is completely determined,
1 π  πi  1
− log z − log z + =− log z
4πi 2 2 2πi

So we have proved the foreseen formula with the particular substitution

′
C(h, k) = s(h, k) − h−h
12k :
! !
h′ + i/z h + iz 1 h′ − h
log η = log η + log z + πis(h, k) + πi ,
k k 2 12k

which is the complete formula in all its details. The mysterious s(h, k) enjoys 253
certain properties. It has the group properties of the modular group behind it
and so must participate in them.
Lecture 30

Last time we had the formula of transformation of log η in the following shape: 254
! !
h′ + i/z h + iz 1 πi ′
log η = log η + log z + (h − h) + πis(h, k),
k k 2 12k
where s(h, k) is the Dedekind sum, which, by direct computation of residues,
was seen to be
k−1 ! " # !
X µ 1 hµ hµ 1
− − − .
µ=0
k 2 k k 2

We use the abbreviation: for real x,



 1
 x− [x] − 2 , if x is not an integer,


((x)) = 



 0 , if x is an integer.

Then
k   !!
X µ hµ
s(h, k) = .
µ=1
k k

Now ((x)) is an odd function; for x integer, trivially ((−x)) = −((x)), and
for x not an integer,
1
((−x)) = −x − [−x] −
2
1
= −x + [x] + 1 − , since [−x] = −[x] − 1,
2
= −((x)).
∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗

188
30. Lecture 189

((x)) is the familiar function whose graph is as indicated. 255

We now prove that

k  
X µ
=0
µ=1
k

Because of periodicity we can write

k   X  µ 
X µ
=
µ=1
k µ mod k
k
X  µ 
=
µ mod k
k
X  µ 
=−
µ mod k
k
k  
X µ
∴ =0
µ=1
k

We can also write s(h, k) in the form

k ! !!
X µ 1 hµ
s(h, k) = −
µ=1
k 2 k
k !! k !!
X µ hµ 1X hµ
= − ,
mu=1
k k 2 µ=1 k

and since hµ also runs through a full system of residues mod k when µ does 256
so, as (h, k) = 1, the second sum is zero, and we can therefore write
k !!
X µ hµ
s(h, k) =
µ=1
k k
30. Lecture 190

Let us now rewrite this in a form in which the modular substitution comes
into play
h′ + i/z h + iz
τ′ = , τ= ;
k k
so kτ − h = iz, and
h′ − 1/(kτ − h) h′ kτ − hh′ − 1
τ′ = =
k k(kτ − h)
h′ τ − (hh′ + 1)/k
=
kτ − h
′
( hhk+1 is necessarily integral for hh′ ≡ −1 mod k). So the modular substitution
is    
 ′ −hh′ +1   
 h k  = a b  , c > 0.
   
k −h c d
The transformation formula for log η now reads
!
aτ + b 1 cτ + d πi
log η = log η(τ) + log + (a + d) − πis(d, c),
cτ + d 2 i 12c
since s(−d, c) = − − s(d, c).
Let us take in particular 257
   
a b 0 −1
  =   ;
  
c d 1 0
then we obtain !
1 1 τ
log η = log η(τ) + log ,
τ 2 i
the special case discussed by Siegel.
Let us now make two substitutions in succession:
aτ′ + b 1
τ′′ = , τ′ = − .
cτ′ + d τ
Then
−a/τ + b bτ − a
τ′′ = =
−c/τ + d dτ − c
We suppose c > 0, d > 0; (c, d) = 1. Then
1 cτ′ + d πi
log η(τ′′ ) = log η(τ′ ) + log + (a + d) − πis(d, c);
2 i 12c
30. Lecture 191

1 dτ − c πi
log η(τ′′ ) = log η(τ) + log + (b − c) − πis(−c, d).
2 i 12d
Sub tracting, and observing that
1 τ
log η(τ′ ) − log ηn(τ) = log ,
2 i
we have 258

1 τ 1 cτ′ + d 1 dτ − c
0= log + log − log
2 i 2 i 2 i !
πi a + d b − c
+ − − πi(s(d, c) − s(c, d))
12 c d
The sum of the logarithms on the right side is determinate only up to a
multiple of 2πi:
τ cτ′ + d dτ − c τ (−c/τ + d)/i
log + log − log = log + 2πik
1 i i i (dτ − c)/i
!
1
= log + 2πik
i
πi
= − + 2πik
2
Now each logarithm above has an imaginary part which is strictly less than
π
2 in absolute value; so
 ( )
 τ cτ′ + d dτ − c  3π
Im log + log − log  <
i i i 2
So the only admissible value of k is zero.
Hence we have 259
!
πi πi a + d b − c
0=− + − − πi(s(d, c) + s(c, d)) ,
4 12 c d
or since ad − bc = 1,
!
1 1 d c 1
s(d, c) + s(c, d) = − + + + .
4 12 c d cd
This is the reciprocity law for Dedekind sums. It is a purely arithmetical
formula for which I have given several proofs; here I reproduce the proof that
I gave originally, by lattice-point enumeration.
30. Lecture 192

We have to prove that

k−1 ( " # ) Xh−1 ( " # )
X µ hµ hµ 1 ν kν hν 1
− − + − −
µ=1
k k k 2 µ=1
h h k 2
!
1 1 h k 1
=− + + + ,
4 12 k h hk
or
k−1 k−1 h−1 h−1 k−1 " # h−1 " #
h X 2 1 X k X 2 1 X 1 X hµ 1 X kν
µ − µ + ν − ν − µ − ν
k2 µ−1 2k µ−1 h2 ν−1 2h ν=1 k µ=1 k h ν=1 h
!
1 1 h h 1
=− + + + ;
4 12 k h hk
or 260

h2 (k − 1)(2k − 1) h k(k − 1) k2 (h − 1)(2h − 1) k h(h − 1)

− + −
6 2 2 6 2 2
k−1 " # h−1 " #
X hµ X kν
−h µ −k ν
µ=1
k ν=1
h
−3hk + h2 + k2 + 1
=
12
k−1 " # h−1 " #
X hµ X kν
i.e., 12h µ + 12k ν
µ=1
k ν=1
h
= h(k − 1)(2h(2k − 1) − 3k) + k(h − 1)(2k(2h − 1) − 3h) + 3hk − h2 − k2 − 1
= 8h2 k2 − 9h2 k − 9hk2 + h2 + k2 + 9hk − 1
= (h − 1)(k − 1)(8hk − h − k − 1)

So the whole thing is equivalent to proving that

k−1 " # h−1 " #
X hµ X kν
12h µ + 12k ν = (h − 1)(k − 1)(8hk − h − k − 1).
µ=1
k ν=1
h

This reduces to something that looks familiar; indeed the square brackets
appear in lattice-point enumeration. Here (h, k) = 1, but in a paper with White-
man I have also discussed the case where h, k are not coprime.
Enumerating by rows and columns parallel to the µ− and ν− axes, the num- 261
ber of lattice-points in the integer a the rectangle
30. Lecture 193

with sides of length k, h along the axes of µ and ν respectively is seen to be

(h−1)(k−1). This can be enumerated in another way also. The number of lattice
points in the interior, with abscissa µ and lying below the diagonal through the
k−1
P h hµ i
origin is the full integer in hµ
k . So we have k lattice points below the
µ=1
h−1
Ph i
kν
diagonal. Similarly there are h points above the diagonal. Since (h, k) = 1
ν=1
there are no points on the diagonal. Hence
k−1 " # h−1 " #
X hµ X kν
(h − 1)(k − 1) = +
µ=1
k ν=1
h

In out case we have quadratic summands; but something which goes so

well here in the plane should go well in space also.
Lecture 31

We want to prove directly the reciprocity formula 262

!
1 1 h k 1
s(h, k) + s(k, h) = − + + +
4 12 k h hk
k !!
X µ hµ
with s(h, k) =
µ=1
k k

The reciprocity formula is equivalent to proving that

k−1 " # h−1 " #
X hµ X kν
12h µ + 12k ν = (h − 1)(k − 1)(8hk − h − k − 1)
µ=1
k ν=1
h

We made a little digression and spoke of similar sums which occur in

lattice-point summations:
k−1 " # h−1 " #
X hµ X hν
+ = (h − 1)(k − 1)
µ=1
k ν=1
h

If we use a rectangle of sides 2h , 2k , (h, k odd) we obtain

k−1
" # h−1
" #
2 2
X hµ X hν 1
+ = (h − 1)(k − 1).
µ=1
k ν=1
h 4

This is made use of the theory of quadratic residues.

The summands in our case are ‘quadratic’ in µ and ν.

194
31. Lecture 195

Consider the rectangular parallelopiped with three concurrent edges along 263
the axes of µ, ν and ρ, the lengths of these edges being h, k, hk respectively. Dis-
sect the parallelopiped into three pyramids having a common apex at the origin
and having for bases the three rectangular faces which do not pass through the
origin, viz. ABCD, BCFE and CDGF. We now compute the number of lattice
points in each pyramid. Take for example the pyramid O(BEFC). Consider
a section parallel to the (ρ, ν)-plane at a distance µ along the µ-axis. The lat-
tice points lie in such sheets. The edges of this section are hµ and µ hk . The
number of lattice points on this sheet (including possibly those on the edges) is
h i k−1 h i
hµ µh hµ µh
P
k . So for the whole pyramid the number = k . For the pyramid
µ=1
h−1
P h i
νk
O(ABCD), the one facing us, the number is kν h
ν=1
Of course are some points on the common edge. Finally there is a pyramid 264
of exceptional sort which lies upside down. Consider a section at a height
h parallel to the (µ, ν) plane the numberr of lattice points on and inside this
pyramid is seen to be
hk−1
X ρ ρ
.
ρ=1
h k
So altogether we have
k−1 " # Xh−1 " # hk−1
X ρ ρ
X µh νk
hµ + kν +
µ=1
k ν=1
k ρ=1
h k
31. Lecture 196

points, including some points which have been counted twice over. But the
number of lattice points inside: the parallelopiped is equal to (h − 1)(k − 1)(hk −
1). Hence making a correction for the lattice points on the cleaving surfaces
through the edges CF and CD which have been counted twice (the surface
along BC has no points on it because (h, k) = 1), we have
k−1 " # Xh−1 " # hk−1
X ρ ρ
X µh νk
hµ + kν +
µ=1
k ν=1
h ρ=1
k h
= (h − 1)(k − 1)(hk − 1) + (h − 1)(k − 1)
= hk(h − 1)(k − 1)

Now write
hk−1
Xρ ρ
S =
ρ=1
h k
 ρ  ρ 1  ρ  ρ  ρ 
= − − , if h ∤ ρ; − , if h ∤ ρ.
h h 2 h h h
So 265
hk−1
X( )( )
ρ 1  ρ  ρ 1  ρ 
S = − − − −
ρ=1
h 2 h 2 2 k

With some correction. Indeed h | ρ, k | ρ do not happen together: Let

ρ = hσ, ρ = kτ. In the first case. i.e., h | ρ, we have to correct the above by an
amount
k−1 ( !!)
X 1 hσ 1 hσ
− − ,
σ=1
2 k 2 k
and in the second case, k | ρ, by
h−1 ( !!)
X 1 kτ 1 kτ
− −
τ=1
2 h 2 h

So
hk−1
X( )( ) X hk   ! X hk   !
ρ 1 ρ 1 ρ ρ 1 ρ ρ 1
S = − − − − − −
ρ=1
h 2 k 2 ρ=1
h k 2 ρ=1
k h 2
hk     k−1 ( ) h−1 ( )
X ρ ρ 1 X hσ 1 1 X 1 kτ 1
+ + − + −
ρ=1
h k 2 σ=1 k 2 2 τ=1 2 h 2
31. Lecture 197

P  µ 
Since k = 0, this becomes
µ mod k

hk−1
X( 2 ) hk hk
ρ 1 ρ ρ 1 1 X  ρ  1 X  ρ 
S = − + + − ρ − ρ
ρ=1
hk 2 h k 4 4 ρ=1 k h ρ=1 k
hk     ! !
X ρ ρ 1 h(k − 1) k − 1 1 k(h − 1) h − 1
+ + − + −
ρ=1
h k 2 2 2 2 2 2

we use the periodicity in the non-elementary pieces; so write 266

ρ = hr + s; r = 0, 1, . . . , k − 1; s = 1, . . . , h.
hk  ρ  X k−1 Xh !!
X hr + s
ρ = (hr + s)
ρ=1
h r=0 s=1
h
k−1 X
X h  s  k−1 X
X h  s 
= hr + s
r=0 s=1
h r=0 s=1
h
Xh  s 
=k s
s=1
h

(since the first sum is zero, as we see by summing over s first)

h−1 !
X s 1
=k s −
s=1
h 2
( )
(h − 1)(2h − 1) 1
=k − h(h − 1)
6 4
k(h − 1)(h − 2)
=
12
Similarly 267
Xhk  ρ  h(k − 1)(k − 2)
ρ =
ρ=1
k 12

next, consider
ik    
X ρ ρ
ρ=1
h k

Write ρ = hα + kβ; when α, β run through complete systems of residues

modulo h, k respectively, hα + kβ runs through a complete system modulo hk,
31. Lecture 198

by the Chinese remainder theorem. Then

hk     !! !!
X ρ ρ X X hα + kβ hα + kβ
=
ρ=1
h k α mod k β mod h
h k
!! !!
X X kβ hα
=
α mod k β mod h
h k
!! X !!
X hα kβ
=
α mod k
k β mod h h
=0

since each sum is separately zero. Hence

1 1 1
S = (hk − 1)(2hk − 1) − (hk − 1)(k + h) + (hk − 1)
6 4 4
1 1 1
− (h − 1)(h − 2) − (k − 1)(k − 2) + (k − 1)(h − 1)
12 12 2
1 1
= (hk − 1)(4hk − 3h − 3k + 1) − (h − 1)(h − 2)
12 12
1 1
− (k − 1)(k − 2) + (k − 1)(h − 1)
12 2
1
= (h − 1)(k − 1)(4hk + h + k + 1)
12
Thus 268

k−1 " # h−1 " #

X hµ X kν 1
h µ +k ν + (h − 1)(k − 1)(4hk + h + k + 1)
µ=1
k ν=1
h 12
= (h − 1)(k − 1)hk
k−1 " # h−1 " #
X hµ X kµ
∴ 12h µ + 12k ν
µ=1
k ν=1
h
= (h − 1)(k − 1)(8hk − h − k − 1)

We make some elementary remarks about quadratic residues. The reci-

procity formula gives, on multiplication by 12h2 k.

12h2 ks(h, k) + 12h2 ks(k, h) == −3h2 k + h3 + k2 h

Look at the denominator of s(h, k). At worst it can have for factors 2 and 269
31. Lecture 199

k2 . So 2k2 s(h, k) is integral. 2h2 s(k, h) is also integral.

12h2 ks(h, k) ≡ h3 + k2 + h (mod 3k)
2
≡ h(h + 1) (mod k),
and since h2 cannot help to make an integer of the left side,
12hks(h, k) ≡ h2 + 1 (mod k).
Sp 12ks(h, k) is an integer. The highest possible denominator for s(h, k) is
(2k2 , 12k) = 2k(k, 6). So the denominator which at first glance could conceiv-
ably be as big as 2k2 is actually at most only 2k(k, 6). This is achieved, for
instance, in s(1, 3) = 1/18, where 6(6, 3) = 18. In fact s(1, 3) can be computed
from the reciprocity formula:
!
1 1 1 3 1
s(1, 3) + s(3, 1) = − + + + s(3, 1) = 0
4 12 3 1 3
1
since an integer is involved and so s(1, 3) = 18 .
In general,
!
1 1 1 k 1
s(1, k) = − + + +
4 12 k 1 k
(k − 1)(k − 2)
=
12k
s(2, k) is also easily obtained. k is odd; so we have
!
1 1 2 k 1
s(2, k) + s(1, 2) = − + + +
4 12 k 2 2k
and as s(1, 2) = 0 (by direct computation), we get 270

(k − 1)(k − 5)
s(2, k) =
24k
Let us calculate s(5, 27).
1 12 + 52 + 272
s(5, 27) + s(27, 5) = − +
4 12 × 5 × 27
1 12 + 22 + 52
s(2, 5) + s(5, 2) = − +
4 12 × 2 × 5
s(5, 2) = 0 = s(1, 2), and on sub traction,
s(5, 27) = 35/(6 × 27); and we know that
the denominator could be at most 2.27(27, 6) = 6 × 27.
Lecture 32

We shall study a few more properties of Dedekind sums. We had the reciprocity 271
law !
1 1 h k 1
s(h, k) + s(k, h) = − + + + .
4 12 k h hk
From this we deduced as a consequence

12hk s(h, k) ≡ h2 + 1 (mod k) (*)

Now when do the Dedekind sums vanish? Let us write s(h, k) in the more
flexible form:
X  µ  µh !!
s(h, k) =
µ mod k
k k

Let hh∗ ≡ 1 (mod k). Since (h∗ , k) = 1, h∗ µ runs through a full residue
system modulo k, and so
!! !!
X µh∗ µhh∗
s(h, k) =
µ mod k
k k
X  µ  µh∗ !!
=
µ mod k
k k
= s(h∗ , k)
 
This is of some significance. We came to s from the substitution a b and
c d
since ad ≡ 1 (mod c), s(d, c) = s(a, c). hh′ ≡ −1 (mod k), and
X  µ  µh !!
s(h, k) =
µ mod k
k k

200
32. Lecture 201
!! !!
X h′ µ µhh′
=
µ mod k
k k
X  µ  h′ µ !!
= −
µ mod k
k k
= −s(h′ , k)

When h = h′ i.e., h2 ≡ − (mod k) (cg. 22 ≡ − (mod 5)) 272

s(h, k) = −s(h, k)
or s(h, k) = 0 if h2 ≡ −1 (mod k)

(in particular if h2 + 1 = k. I have a conjecture that s(h, k) ≥ 0 if h2 < k). In

fact we can say more. We have the
Theorem. 12 s(h, k) is en integer only for h2 ≡ −1( (mod k)) and is then equal
to zero.
For assume that 12s(h, k) = integer; this implies, because of (*), that 0 ≡
h2 + 1 (mod k)
In such cases, therefore, we can make a direct statement about the value of
s(h, k) without going through the rigmarole of the Euclidean algorithm. Thus
s(2, 5) = 0, s(5, 26) = 0.
In a recent issue of the Duke Mathematical Journal (1954), I gave a gen-
eralisation of the reciprocity formula for Dedekind sums. It takes into account
three summands. The formula is very elegant and throws some light on the
reciprocity relation itself. We quote it without proof.
Theorem . If a, b, c are pairwise coprime and aa∗ ≡ 1 (mod bc), bb∗ ≡ 1
(mod ca), cc∗ ≡ 1 (mod ab), then

T ≡ s(bc∗ , a) + s(ca∗ , b) + s(ab∗ , c)

!
1 1 a b c
=− + + +
4 2 bc ca ab
273
The proof is by an algebraic method due to Rédel. The formula is very
gratifying as a generalisation of the reciprocity formula is which latter there is
some non-homogeneity. Put c = 1; then c∗ = 1, and s(ab∗ , c) = 0; so we get the
reciprocity formula. The right side above is 12 1abc − 3abc + a2 + b2 + c2 . Hence
T = 0 if and only if a2 + b2 + c2 = 3abc. This combination of three integers
plays some role the theory of quadratic forms; it is called a Markoff triple. It
has reappeared in literature in connection with the geometry of numbers. It has
32. Lecture 202

to do with the existence if certain quadratic forms with minimum values close
to zero for integers. 1, 1, 2 is a Markoff triple. If we keep two of them fixed, for
the third we get a quadratic equation of which one root we know to be rational.
So the other root is rational too. For instance if a, b = 1 are fixed, we have
c2 − 3c + 2 = 0 or (c − 1)(c − 2) = 0; – the triples pre 1, 1, 1 and 1,1,2. If we
take the triple a, 1, 2, then a2 + 5 = 6a or a = 1, 5; we have the triples b, 1, 2;
1, 1, 2. T = 0 only if a, b, c being to a Markoff triple. For such a triple,

b2 + c2 ≡ 0 (mod a), c2 + a2 ≡ 0 (mod b), a2 + b2 ≡ 0 (mod c)

So b2 ≡ −c2 (mod a), or (c∗ b)2 ≡ −1 (mod a), etc.

Then s(bc∗ , a) = 0, and each summand in T is zero.

Dedekind
  sums have something to do with Farey fractions. Let us suppose
that  ac db  = 1, c, d > 0.
!
1 1 c d 1
s(c, d) + s(d, c) = − + + +
4 12 d c cd
cb ≡ −1 (mod d) and ad ≡ 1 (mod c)
so s(c, d) = −s(−b, d) and s(d, c) = s(a, c).
!
1 c d 1
So s(a, c) − s(b, d) = −4 + + +
12 d c cd
 
Now if hk11 , hk22 the adjacent Farey Fractions, then  hk11 hk22  = −1 274
so !
1 1 k1 k2 1
s(h1 , k1 ) − s(h2 , k2 ) = − + +
4 12 k2 k1 k1 k2
h  h 
Write the left side as s k11 − s k22 .
Suppose hk22 is fixed. Let us look at all possible adjacent fractions hk11 . They
are obtainable by forming mediants; replace hk11 successively by hk11 +λk +λh2
. Make
 2  
k2
k1 larger and larger. Then k1 and k1 k2 → 0. So k2 → ∞. Thus s k1 − s hk22
1 k1 h1
 
goes unboundedly by −∞, and so s hk11 → −∞. Therefore only on the left side
of hk22 can we get a sequence of rational fractions for which the Dedekind sums
tend to −∞.
We now give  µanother
 proof of the reciprocity law, by the method of finite
Fourier series, k is a number- theoretic periodic function. It has a finite
Fourier expansion:
 µ  X k
µ
= c j e2πi j k
k j=1
32. Lecture 203

In fact this is always solvable for c1 , c2 , . . . , ck . For writing down µ = 275

1, 2, 3, . . . , k in succession, we have a system of k linear equations whose de-
terminant is a Vandermonde determinant which is non-zero since the roots of
unity are different. We have
k   k k
X µ ℓ
X X ( j−ℓ)
e−2πiµ k = cj e2πiµ k

µ=1
k j=1 µ=1

= kcl ,
k
1 X  µ  −2πiµ ℓ
i.e., cl = e k
k µ=1 k

This was done by Eisenstein, We can also write

k−1 !
1 X µ 1 −2πiµ l
cl = − e k
k µ=1 k 2

k−1 
 k−1

1 X
−2πiµ kl

 2k , if k | l;
= 2 µe −
k µ=1 

1,

2k if k ∤ l.

So if k | l, then
1 k(k − 1) k − 1
cℓ = − = 0.
k2 2 2k
In particular
ck = 0.
If k ∤ l, then writing
k−1
X ℓ
S = µe−2πiµ k ,
µ=1
k−1
X
ℓ µ+1
S e−2πi k = µe−2πi k ℓ

µ=1
k
X ℓ
= (ν − 1)e−2πiν k
ν=2
k
X
l l l
= S − e−2πi k + k − e−2πiν k + e−2πi k
ν=1
32. Lecture 204

So 276
k
S =
e−2πiℓ/k − 1
Hence, if k ∤ l, then
−1 1
cℓ = +
k(1 − e2πiℓ/k ) 2k
−2 + 1 − e−2πiℓ/k
=
2k(1 − e−2πiℓ/k )
ℓ
1 1 + e−2πi k
=− ·
2k 1 − e−2πi kℓ
i πℓ
= cot
2k k
So we have what is essentially Eisenstein’s formula:
 µ  k−1
i X πℓ µ
= cot e2πi j k
k 2k j=1 k
 
This is an explicit formula for µk as a finite Fourier series. We utilise it
for Dedekind sums.
X  µ  hµ !!
s(h, k) =
µ mod k
k k
k−1 k−1
1 X X π j 2πi j µ X πℓ µ
=− cot e k × cot e2πiℓh k
4k2 µ mod k j=1
k ℓ=1
k
k−1 k−1
1 XX πj πl X 2πi µ ( j+hℓ)
=− 2
cot cot e k
4k j=1 ℓ=1 k k µ mod k
k−1
1 X πℓ −πhℓ
=− cot cot ,
4k ℓ=1 k k

since in the summation with respect to µ only those terms remain for which 277
j + hℓ ≡ (mod k). Then
k−1
1 X πℓ πhℓ
s(h, k) = cot cot .
4k ℓ=1 k k
32. Lecture 205

The reciprocity formula can be tackled immediately by the powerful

method of residues. We have to construct the proper function for which these
become the residues. Take
πz πhz
f (z) = cot πz cot cot
k k
and integrate over a rectangle with vertices ±iΩ, ±i(k +iΩ), indented at o and k.
The poles of the first factor all in the contour one 0, 1, . . . , k − 1, for the second
0; and for the third 0, k/h, 2k/h, . . . , (h − i)12/h. We have
!
1 ω2
cot ω = 1− −···
ω 3
278
About the triple pole z = 0,
! ! !
1 k 1 π2 z2 πz2 π2 h2 z2
f (z) = · · 1− +··· 1− +··· 1− +···
πz πz πhz 3 3 3k2

So the residue at z = 0 is
! !
k2 π2 π2 π2 h2 k k 1 h
− − 2− =− + +
π2 h 3 3k 3k2 3π h hk k
So
! k−1
X k k h 1 1X πℓ πhℓ
(Res) = − + + + cot cot
3π h k hk π ℓ=1 k k
32. Lecture 206

h−1
k X πrh πh
+ cot cot
πh k=1 h h
! !
k k h 1
= − + + + 12s(h, k) + 12s(k, h)
3π h k hk
And this is equal to Z 279
1
f (z)dz
2πi R
where R is the rectangle. On the vertical lines the function the same value (by
periodicity) end so the integrals cancel out. Hence
 
Z 
 −iΩ+k
Z iΩ+k
Z 
1 1  
 


f (z)dz =  − 
2πi R 2πi 






−iΩ iΩ

Now
eiω + e−iω
cot ω = i , ω = x + iy,
eiω − e−iω
eix−y + e−ix+y
= i ix−y ;
e − e−ix+y
x varies from o to k and y = ±Ω, for this



−i, as y = Ω → ∞


→  uniformly


 i, as y = −Ω → −∞
Therefore
Z
1 1 n3 o
lim f (z)dz = i k − (−i)3 k
Ω→∞ 2πi R 2πi
2ki3 k
= =−
2πi π
280
! !
k h k 1
∴ − + + + 12s(h, k) + 12s(k, h)
3π k h hk
k
=−
π !
k h 1
or 12s(h, k) + 12s(k, h) = −3 + + + ,
h k hk
which is the reciprocity formula.
 Part IV

Representation by squares

207
Lecture 33

We wish to begin the study of the representation of a number as the sum of 281
squares:
n = n21 + n22 + · · · + n2r
We shall develop in this connection the Hardy-Littlewood circle method.
Historically it is an off shoot of the Hardly-Ramanujan method in partition-
theory, though we did not develop the latter in its original form in our treat-
ment. The circle method has been applied to very many cases, and the problem
of squares is a very instructive one for finding out the general thread. We
shall later replace the problem by that of the representation of n by a posi-
tive quadratic form. This would involve only the general Poisson summation
formula. In the case of representation as the sum of squares there is some
simplification, because the generating ing function is the rth power of a simple
V function. We shall deal with the asymptotic theory. Later we may go into
Siegel’s theory of quadratic forms.
Let us write
X∞ X∞
2 2
Θ(x) = xn = 1 + 2 xn ,
n=−∞ n=1

|x| < 1. For r at least equal to 4, we consider

 ∞ 
r
 X n2 r X∞
2 2 2
Θ (x) =   x  = xn1 +n2 +·+nr
n=−∞ n j =−∞
∞
X
= Ar (n)xn ,
n=0

on collecting the terms with exponent n, where Ar (n) is the number of times n 282

208
33. Lecture 209

appears as the sum of r square:

X
Ar (n) = 1
n21 +···+n2r =n

It is clear that ni can be positive or negative. The more serious thing is

that we have to count the representations differently when the summands are
interchanged, in contradiction to the situation in the case of partitions. The
problem of partition into squares would be a more complicated problem; the
generating function would be more complicated, and what is worse, all the
help one gets in partition theory from the theory of modular forms would break
down here.
An (n) is the nth coefficient of a power-series;
Z
1 Θr (x)
Ar (n) = dx
2πi C xn+1
where C is a suitable circle inside and close to the unit circle. The trick of
Hardy and Littlewood was to break the circle |x| = e−2πδN where N is the order
of a certain Farey dissection, into Farey arcs and write
Z
1 X′ Θn (x)
Ar (n) = dx,
2πio≤h<k≤N ξhk xn+1

where ξhk are the arcs over which one integration piecemeal the prime denoting
that (h, k) = 1. Consider on each piece ξhk the neighbourhood of a root of unity: 283
h
x = e2πi k −2πξ

Re z < 0, and set z = δN − iϕ, so chat we have a little freedom along both real
and imaginary axes.
h
x = e2πi k −2πδN +2πiϕ .
The choice of the little arc ϕ is also classical. hk is a certain Farey fraction,
with adjacents hk11 and hk22 , say. hk11 < hk < hk22 . We limit ϕ on the seperate arcs.
Introduce the mediants:
h1 h1 + h h h2 + h h2
< < < < ,
k1 k1 + k k k2 + k k2
 h +h h +h 
h
So that the interval k11 +k , k22 +2 gives the movement of k + ϕ. So ϕ runs
between
h1 + h h h2 + h h
−Vhk′ = − ≤ϕ≤ − = Vhk′′
k1 + k k h2 + k k
33. Lecture 210

1 1
−Vhk′ = − ; Vhk′′ = ;
(k1 + k)k (k2 + k)k
k1 +k
and since 2N > k2 +k > N, we have necessarily
1 1
≤ |Vhk | ≤
2Nk Nk
Now changing the variable of integration to ϕ, we can write
′′
X′ ZVhk  h 
−2πi hk n
Ar (n) = e Θr e2πi k −2πz e2πnz dϕ
0≤h<k≤N
−Vhk′
284
The trick is to overcome the difficulty in the integral by replacing on each
arc the highly transcendental function by a simpler function. Here we stop for
a moment to see what we can do with the integrand.
 h  X∞
e(2πi k −2πz)n
h 2
Θ · e2πi k −2πz =
n=−∞
k−1
X X
h 2 2
= e2πi k j e−2πzn
j=0 n≡ j (mod k)
k−1
X ∞
X
h 2 2
(q+ kj )2
= e2πi k j e2πzk ,
j=0 q=−∞

where we have written n = kq + j. We can now handle this from our V -series
formula. We proved (Lecture 12) that
C(τ) 1 2
V3 (V / − ) = eπiτV V3 (V τ/τ)
i τ
r
C(τ) i
and = , Imτ > 0.
i τ
Since 285
∞
X
πin2 τ 2πinV
V3 (V /τ) = e e ,
n=−∞

writing τ = it, Re t > 0, we have from the above,

∞ ∞
1 X −π n2 2πinV −πtV 2
X 2
√ e t e = e e−πn t e−2πV nt
t n=−∞ n=−∞
33. Lecture 211

∞
X 2
= e−πt(n+V )
n=−∞

j
Replacing n by q, V by k and t by 2zk2 , we have
k−1 ∞ 2
 h  X h 2 i X − πq j
Θ e2πi k −2πz = e2πi k j p e 2zk2 e2πiq k
2
2zk q=−∞
j=0
∞ πq2
1 X − 2zk
= √ e 2 T q (h, k)
k 2z q=−∞
k−1
X h j2 +q j
where T q (h, k) = e2πi k

j=0

This is already a good reduction. T q (h, k) depends on q modulo k, so it is

periodic. We shall approximate to it in general.
One special case, however, is of interest: for q = 0, 286

k−1
X h 2
T 0 (h, k) = e2πi k j = G(h, k),
j=0

where G(h, k) are the so-called Gaussian sums which we shall study in detail.
They are sums of roots of unity raised to a square power, Θ is actually a V3 ,
and when we evaluate T q we get some other V .
We now write
 h  1
Θ e2πi k −2πz = √ {G(h, k) + H(h, k; z)}
k 2z
∞
X 2
− πq
where H(h, k; z) = T q (h, k)e 2k2 z
q=−∞
q,0

We shall throw H into the error term. Let us appraise T q (h, k), not explic-
itly: that will take us into Gaussian sums.
k−1
X h j2 +q j
T q (h, k) = e2πi k

j=0
k−1 X
X k−1
2πi 2 i 2
|T q (h, k)|2 = e k (h j +q j) e−2π k (hℓ +qℓ)

j=0 ℓ=0
33. Lecture 212
X X i 2
−ℓ2 )+q( j−ℓ))
= e2π k (h(ℓ j
j mod k ℓ mod k
X X i
= e2π k ( j−ℓ)(h( j+ℓ)+q) ,
j mod k ℓ mod k

which, an rearranging according to the difference j − ℓ, becomes 287

X X a
= e2πi k (h( j+ℓ)+q)
a mod k j−ℓ≡a (mod k)
X X a
= e2πi k (h(a+2ℓ)+q)
a mod k ℓ mod k
X X
2π ki (ha2 +aq) h
= e e4πia k ℓ
a mod k ℓ mod k

The inner sum is a sum of the roots of unity. Two cases arises, according
as k | 2a or k ∤ 2a. k odd implies that a = 0 and k even implies that a = 0 or
k | 2. In case k | 2a, the sum is zero. We then have
 !
2
2 2π ki h k4 + 2k q
|T q (h, k)| = k, if k is odd; k 1 + e , if k is even
 
= k 1 + eπi( 2 +q) , if k is even
hk

= o or 2k if k is even
hk
It is of interest to notice that T q = 0 only if k is even and 2 + q is an odd
integer. In any case, √
|T q (h, k)| ≤ 2k,
and this cannot be improved. We then have
∞ √
X πq2 1
|H(h, k; k; z)| ≤ 2 2ke− 2k2 R z (q = 0 is not involved here).
q=1
√ − π R1 X ∞
(q2 −1) 1
= 2 2ke 2k 2 z e−π 2k2 R z
q=1
√ ∞
X
π 1 3πm 1
= 2 2ke− 2k2 R z e− 2k2 R z
m=0
√ π 1 1
= 2 2ke− 2k2 R z 3π
− R 1z
1−e 2k2

Since z = δN − iϕ, 288

33. Lecture 213

1 1 1 1 δN
2
R =R 2 =R 2 = 2 2
k z k z k (δN − iϕ) k (δN + ϕ2 )
1 1 δN 1
∴ 2
R ≥ 2 1
, since |Vhk | ≤ ,
k z 2
k δN + N 2 kN
δN 1
≥ =
N 2 δ2N + N12 N 2 δN + N 21δN

We want to make this keep away from 0 as far as possible. This gives a
desirable choice of δN . Make the denominator as small as possible. Since x + 1x
is minimised when x = 1, we have
1 1 1
R ≥ ,
k2 z 2

this minimum corresponding to N 2 δN = 1. So if we choose the radius of the

circle in terms of the Farey order, we shall have secured the best that we can:
√ π 1
|H(h, k; z)| ≤ 2 2ke− 2k2 R z × C

It would be unwise to appraise the remaining exponential now.

Lecture 34
 h 
We had discussed the sum Θ e2πi k −2πz and written it equal to 289

1
√ {G(h, k) + H(h, k; z)}
k 2z
√ π 1
where |H(h, k; z)| < C ke− 2k2 R z

If we apply this to the integral in which Θr appears,

r !
 h r 1 X r
Θ e2πi k −2πz = k
G(h, k)r−λ H(h, k; z),
r
k (2z) λ=0
2 λ

or, keeping the piece corresponding to λ = 0 apart,

r !
 h r 1 1 X r
Θ e2πi k −2πz − r G(h, k) r
= r G(h, k)r−λ H(h, k; z)λ
kr (2z) 2 kr (2z) 2 λ=1 λ

Let us appraise this. Since

  r
  2πi h −2πz r 1 r  1 X √ r−λ λ − 2π2 1
Θ e k − r G(h, k)  < C r ( k) k 2 e 2k R
kr (2z) 2 kr |z| 2 λ=1 z
1 − 2π − 2kπ2 R 1z
<C· r e e
(k|z|) 2

Now ′′
290
ZVhk  h 
P′ h
Ar (n) = e−2πi k n 2πznΘ e2πi k n−2πz
0≤h<k≤N
−Vhk′

214
34. Lecture 215

where, of course, z = δN − iϕ. Hence

 ′′

 ZVhk 2πnz 
 P′ −2πi hk n e r 
Ar (n) − e
r r G(h, k) dϕ
 0≤h<k≤N k (2z) 2 
−Vhk′ 
′′
ZVhk 2π 1
P′ e− k 2 R z
≤C e2πnRz dϕ
0≤h<k≤N kr/2 |z|r/2
−Vhk′
′′
ZVhk − π δν
2k2 δ2 +ϕ2
P′ 2πnδN e N
≤C e h i 4r dϕ
0≤h<k≤N
−Vhk′ h2 (δ2N + ϕ2 )
′′
ZVhk   4r
P′ 2πnδN − 4r  δN  − 2kπ2 δ2δ+ϕ
N
2
=C e δN 
2 2 2
 e N dϕ
0≤h<k≤N k (δN + ϕ )
−Vhk′

1 1
Now 2kN ≤ Vhk ≤ kN and Vhk′ ≤ ϕ ≤ Vhk′′ , while δN = N12 . Putting 291
δN r π
X = k2 (δ2 +ϕ2 ) , the integrand becomes X 4 e− 2 X which remains bounded. (It was
N
for this purpose that in our estimate of H(h, k; z) earlier we retained the factor
2 1
e−π/(2k )·R z ). Hence the last expression is less than or equal to
′′
X ZVhk
2π Nn2 r n r
C e N 2 dϕ = Ce2π N2 N 2 ,
o≤h<k≤N −Vhk′

since the whole Farey dissection exactly fills the interval (0, 1).
In the next stage of our argument we take the integral
′′
ZVhk
e2πnz
k
z4
Vhk′

and write it as  
Z∞ Z∞ Z −Vhk′
 2πnz

 − −  e
 r/2 dϕ
  z
−∞ Vhk′′ −∞

The infinite integrals are conditionally convergent if r > 0 (because the

numerator is essentially trigonometric), and absolutely convergent for r > 2, 292
34. Lecture 216

so that we take r at least equal to 3. Then

 
Z∞ 2πnz  Z∞
 e  2π n2 dϕ
 r/2
dϕ ≤ e N 2 + ϕ2 ) 2r
 ′′ z 
 (δ N
Vhk Vhk′′
Z∞
2π n2 dϕ
≤e N 2 r
(δN + ϕ2 ) 4
1
2kN

′
−V
R hk
(Here and in the estimate of the other integral , we make use of the
−∞
fact that the interval from Vhk′ to Vhk′′ is neither too long nor too short. This
argument arises also in Goldbach’s problem and Waring’s problem). The right
side is equal to
Z∞ Z∞
r−2 2π Nn2 N 2 dϕ 2π Nn2 r−2 dψ
N e r = e N
4 2
(1 + N ϕ ) 4 (1 + ψ2 )r/4
1 N
− 2kN 2k

Z∞
2π Nn2 r−2 dψ
<e N
ψr/2
N
2k

This appears crude but is nevertheless good since ϕ never comes near 0;
N/2k > 12 , and the ratio of ψ2 to 1 + ψ2 is at least 13 and so we lose no essential
order of magnitude. The last integral is equal to

n
 N − 2r +1
Ce2π N2 N r−2 , r ≥ 3,
2k
n r r
= Ce2π N2 N 2 −1 k 2 −1

−Vhk′
R
A similar estimate holds for also. So, 293
−∞
 
 !r Z∞ 2πnz 
P ′ h G(h, k) e 
Ar (n) − e−2πi k n √ r/2
dϕ
 0≤h<k≤N k 2 z 
−∞
2π n
r/2 P′ 1 2π n
< Ce N2 N +C e N2 N r/2−1 kr/2−1
0≤h<k≤N kr/2
34. Lecture 217

n h
< Ce2π N2 N r/2 + Ce2π N2 N r/2−1
P′
0≤k≤N
n
= Ce2π N2 N r/2 .
This, however, does not go to zero as N → ∞; we have no good luck here as
we had in partitions. So we make the best of it, and obtain an asymptotic result.
Let n also tend to infinity. We shall keep n/N 2 bounded, without lotting; it go
to zero, as in the latter case the exponential √ factor would become 1. We have
to see to it that n ≤ CN 2 i.e., N is at least n. Otherwise the error term would
increase fast. Making N bigger would not help in the first factor and √ would 294
make the second worse. So the optical choice for N would be N = [ N]. The
error would now be  r
O n4
We next evaluate the integral
Z∞
e2πnz
dϕ
zr/2
−∞

This is the some as

Z∞ Z−∞
e2πn(δN −iϕ) e2πn(δN +iα)
dϕ = − dϕ
(δN − iϕ)r/2 (δN + iα)r/2
−∞ ∞
δZ
N +i∞
1 e2πns
= ds
i sr/2
δN −i∞

After a little embellishment this becomes a well-known integral. It is equal

to
2πnδ
ZN +i∞
(2πn) r/2 eω
dω
i ωr/2
2πnδN −i∞

which exists for r > 2, and is actually the Hankel loop integral, and hence equal
to
2π(2πn)r/2 − 1
Γ(r/2)
Hence, for f ≥ 3. We hence the number of representations of n as the sum 295
of r squares:
r
(2π)r/2 n 2 −1 P′ G(h, k)r −2ri h
Ar (n) = · r/2 · e k + O(nr/4 ).
Γ(r/2) 2 0≤h<k≤N kr
34. Lecture 218

One final step. Let us improve this a little further. Write

X G(h, k)r −2πi h n
r
e k = Vk(r) (n) = Vk (n)
h mod k
k

We have to sum Vk (n) from k = 1 to k = N. However, we sum from k = 1

to k = ∞, thereby incurring an error
 
 X ∞  X∞
r
 Vk (n) ≤ k− 2 +1 ,
k=N+1  k=N+1

and this converging absolutely for r ≥ 5 is

 r   r 
O N − 2 +2 = O n− 4 +1
r
This along with the factor n 2 −1 would give exactly O(nr/4 ). (We could have
saved this for r = 4 also if we had been a little more careful). Thus, for r ≥ 5,
we have 296

πr/2 r −1
Ar (n) = n 2 S r (n) + O(nr/4 ),
Γ(r/2)
X∞
where S r (n) = Vk (n)
k=1

S r (n) is the singular series. We shall show that S r (n) remains bounded at
least for r ≥ 5.
Lecture 35

After we reduced our problem to the singular series in which the Gaussian 297
sums appear conspicuously, we have to do something about them before we
proceed further. The Gaussian sums are defined as
X h 2
G(h, k) = e2πi k ℓ , (h, k) = 1
ℓ mod k

They obey a simple multiplication rule: if k = k1 k2 , (k1 , k2 ) = 1, then

G(h, k1 k2 ) = G(hk1 , k2 ) · G(hk2 , k1 ).

For, put ℓ = rk1 + sk2 ; when r runs modulo k2 and s modulo k1 , ℓ runs
through a full residue system modulo k1 k2 . Hence
X X
2πi h (k r+k s)2
G(h, k1 k2 ) = e k1 k2 1 2
k mod k2 s mod k1
X X
2πi k hk (k12 r2 +k22 s2 )
= e 1 2

r mod k2 s mod k1
X hk1 2 X hk2 2
2πi k2 r 2πi k1 s
= e e
r mod k2 s mod k1

= G(hk1 , k2 )G(hk2 , k1 ).

Ultimately, therefore, only prime powers have to be considered to denomi-

nators. We have to distinguish the cases p = 2 and p > 2, p prime.
1) Let p ≥ 3, k = pα with α > 1 298
X h 2
G(h, pα ) = e2πi pα ℓ
l mod pr

219
35. Lecture 220

write ℓ = mpα−1 + r;

m = 0, 1, . . . , p − 1; r = 0, 1, . . . pα−1 − 1. Then this becomes

α−1 α−1−1
p−1 pX
X −1 p−1 pX
X
2πi phα (mpα−1 +r)2 h 2 2α−1
+2mrpα−1 +r2 )
e = e2πi pα (m p

m=0 r=0 m=0 r=0

Since α ≥ 2, 2α − 2 ≥ α and so the first term in the exponent may be

omitted. This gives
α−1
pX −1 p−1
X
2πi phα r2 h
e e2πi p 2mr
r=0 m=0

The inner sum is a sum of pth roots of unity; so it depends on whether p

divides 2rh or not. But (h, p) = 1 and p ∤ 2. So we need consider only the
cases: p | r and p ∤ r. However in the latter case this sum is 0 while in the
former it is p. We therefore get, when p | r, r = ps,
α−1 α−2
pX −1 pX −1
h 2 h 2
p e2πi pα r = p e2πi pα p s

r=0,p|r s=0
α−2
pX−1
h
2πi s
=p e pα−2

s=0
= pG(h, pα−2 )

We have therefore reduced the never of the denominator by 2. We can 299

repeat the process and proceed as long as we end with either the 0th or the 1 st
power. So we have two chances. In the former case, evidently G(h, 1) = 1. So
for α even,
G(h, pα ) = pα/2
On the other hand, if α is odd, we have
α−1
G(h, pα ) = p 2 G(h, p).

These may be combined into the single formula

α  α 
G(h, pα ) = p[ 2 ] G h, pα−2[ 2 ] (1)

2) p = 2λ , λ ≥ 2. h is now odd. Write

ℓ = m2λ−1 + r; m = 0, 1; r = 0, 1, . . . , 2λ−1 − 1
35. Lecture 221

2λ−1
X−1 2λ−1
X−1
h 2 h λ−1
+r)2
G(h, 2λ ) = e2πi 2λ r + e2πi 2λ (2
r=0 r=0

since λ ≥ 2, 2λ − 2 ≥ λ, in the second sum it is only the exponent r2 that

contributes a non-zero term; and this is then the same the first. Altogether we 300
have then
λ−1
2X −1
h 2
2 eπi 2λ −1 r (*)
r=0

This, however is not a Gaussian sum. The substitution for ℓ does not work;
to be effective, then we take

ℓ = m2λ−2 + r; m = 0, 1, 2, 3; r = 0, 1, . . . , 2λ−2 − 1.

Now take λ ≥ 4 and start again all over.

λ−2
3 2X
X −1
h λ−2
+r)2
λ
G(h, 2 ) = e2πi 2λ (m2
m=0 r=0
λ−2
3 2X
X −1
h λ−1
mr+r2 )
= e2πi 2λ (2 , (for λ ≥ 4 i.e., 2λ − 4 ≥ λ).
m=0 r=0
λ−2
2X −1 3
X
h 2
= e2πi 2λ r eπihmr
r=0 m=0
λ−2
2X −1 3
X
h 2
= e2πi 2λ r (−)mn
r=0 m=0
λ−2
2X −1 2λ−2
X−1
h 2 h 2
=2 (−)r e2πi 2λ r + 2 e2πi 2λ r
r=0 r=0
λ−3
2X −1
h 2
=4 eπi 2λ−3 s
s=0

This is not Gaussian sum either. But is is of the form (*). We therefore 301
have, for λ ≥ 4, G(h, 2λ) = 2G(h, 2λ−2). If λ = 4, we need go down to only
22 = 4 and if λ = 5 to 23 = 8. So we need separately G(h, 8) and G(h, 4); and
of course G(h, 2). These cases escape us, while formerly only G(h, p) did. For
λ ≥ 4, we may write
λ  λ 
G(h, 2λ) = 2[ 2 ]−1G h, 2λ−2[ 2 +2] (2)
35. Lecture 222

This supplements formula (1).

We now consider the special cases, k = 2, 4, 8. Here h is odd.
h
G(h, 2) = 1 + e2πi 2 = 0
h h h
G(h, 4) = 1 + e2πi 4 ·1 + e2πi 4 ·4 + e2πi 4 ·9
h
= 2 + 2eπi 2
 
= 2 1 + ih
h
G(h, 8) = 1 + 1 + 2eπih + 4e2πi 8
(since 12 , 32 , 52 , 72 are all ≡ 1 modulo 8)
!2
h 1+i
= 4eπi 4 = 4 √
2

Before we return to G(h, p), p > 2, we shall a digression an connect to the 302
whole thing with the Legendre-Jacobi symbols
p−1
X h 2
G(h, p) = e2πi p ℓ
ℓ=0
X h
=1+2 e2πi p a ,
a

the summation over all quadratic residues a modulo p, since along with ℓ, p − ℓ
is also a quadratic residue. We can write this in a compact form, so arranging
it that the non-residues get cancelled and the residues appear twice:
X ( r
!)
h
G(h, p) = 1+ e2πi p r
r mod p
p
!
X r 2πi hp r
= e
r mod p
p

This would appear in a completely new aspect if we utilised the fact that hr
runs through a full system of residues modulo p. Then
X h ! hr ! h
G(h, p) = e2πi p r
k mod p
p p
! X !
h r 2πi rp
= e
p r mod p p
35. Lecture 223
!
h
= G(h, p).
p
This is very useful if we new go to the Jacobi symbol. For prime p, the 303
Legendre symbol has the multiplicative property:
! ! !
r1 r2 r1 r2
=
p p p
Jacobi has the
 following generalisation.
r
Define pq by
! ! !
r r r
= .
pq p q
Si it is ±1; if it is +1 it does not necessarily mean that r is a quadratic residue
modulo pq. The Jacobi symbol no longer discriminates between residues and
non residues. From the definition then
! !α !β
a a a
= ··· .
pα qβ · · · p q
The Jacobi symbol has the properties of a character, as can be verified by
using the Chinese remainder theorem.
We can now write
!α
h
G(h, pα ) = G(1, pα )
p
under all circumstances. How does this come about? Separate the cases: α
even, α odd.

G(g, pα ) = G(1, pα ), α even;

α−1
=p 2 G(h, p), α odd,
! !
h α−1 h
= p 2 G(1, p) = G(1, pα )
p p
We can write both in one sweep as 304
!α
h
G(h, pα ) = G(1, pα )
p
!
h
= α G(1, pα )
p
35. Lecture 224

Now use the multiplicative law. If p, q are odd primes, then

G(h, pα qβ ) = G(hpα , qβ )G(hqβ, pα )

! !
hpα β hq
β
= G(1, q ) G(1, pα )
qβ pα

Since the Jacobi symbol is separately multiplicative in numerator and de-

nominator, but not both, this is equal to
! ! ! β! ! !
h pα β h q h h
G(1, q ) α G(1, pα ) = G(pα , qβ )G(qβ, pα ),
qβ q β p pα qβ pα

taking the second and third factors together, and also the last two. And this is
!
h
G(1, pα qβ )
pα qβ

according to the multiplication law.

Suppose that we have
! !
h h
G(h1 k1 ) = G(1, k1 ); G(h, k2) = G(1, k2 ).
k1 k2

We go through the above worker; literally and get 305

!
h
G(h, k1k2 ) = G(1, k1 , k2 ).
1, h2

So we have proved in general that for odd k,

!
h
G(h, k) = G(1, k)
k

We can now return to G(h, p).

Lecture 36

We were discussing Gaussian sums and it remained to evaluate 306

!
h
G(h, p) = G(1, p)
p

We shall do a little more than that; we shall study them in a more flexible
form. Define
k−1
X hg 2
S (h, k) = eπi k ℓ ,
ℓ=0

h, k > 0 but not necessarily coprime. We cannot now take the summation over
ℓ modulo k. For if k is odd, (ℓ + k)2 = ℓ2 + 2ℓk + k2 and k2 may give rise to
an odd multiple of πi in the exponent and hence introduce a change of sign,
We should therefore insist on this particular range of summation. S (h, k) are
connected with the Gaussian sums; indeed

G(h, k) = S (2h, k)

We shall now produce S (h, k) as a sum of residues. To get the integers

as poles we should clearly take e2πiz − 1 in the denominator; so we integrate
h 2 2πiz
eπi k z /(e −1) over such a contour as has in its interior the desired poles z =
0, 1, 2, . . . , k − 1. Indeed
Z h 2
eπi k z
S (h, k) = 2πiz
dz
C e −1

225
36. Lecture 226

0 1 2

Where C is the parallelogram with vertices at ±(1 + i)Ω, ±(1 + i)Ω + k, 307
with the slant sides inclined at 45◦ (infact this may be anything less than 90◦ )
to the real axis, and making a detour round 0 and k. When we push Ω to ∞, the
integrals along the horizontal sides will tend to zero. For instance on the upper
side, z = (1 + i)Ω + x, 0 ≤ x ≤ k, and the integrand is therefore
h 2 h 2 2
eπi k ((1+i)Ω+x) eπi k (2iΩ +2(1+i)Ωx+x )
2πi((1+i)Ω+x)
= 2πi(Ω+x)−2πΩ
e −1 e −1
h 2 h 2
e−π k (2Ω +2Ωx)+πi k (2Ωx+x )
=
e−2πΩ+2πi(Ω+x) − 1
h
→ 0 uniformly as Ω → ∞ since k > 0. Hence the integral can be written as 308

(1+i)∞+k
Z (1+i)∞
Z h 2
eπi k (z)
− dz
e2πz − 1
−(1+i)∞+k −(1+i)∞

where, of course, we have to make a small detour round 0 and k. Replacing z

by z + k in the first integral, this becomes
(1+i)∞ (1+i)∞ h 2
 h 2

Z h 2
eπi k (z+k) − eπi k z
h 2 Z eπi k z eπi k (2zk+k ) − 1
dz = dz
e2πiz − 1 e2πiz − 1
−(1+i)∞ −(1+i)∞
(1+i)∞
eπ k z ( e −1)
Z h 2 2πihz+πihk

= dz
e2πiz − 1
−(1+i)∞
36. Lecture 227

Let us assume from now on that hk is even. Then we can actually divide 309
out and the integral becomes
(1+i)∞
Z  
 h 2 X h−1 
eπi k z e 2πiλz  dz
 
−(1+i)∞ λ=0

The denominator has now disappeared. There is a further advantage that

the integral can now be stretched along the whole line and the detour can be
avoided. We then have
(1+i)∞
h−1
X Z
λk 2
−πiλ2 hk
eπi k (z+ h ) dz
h
e
λ=0 −(1+i)∞

Write z + λk/h = ω; and shift the integral back to the line from −(1 + i)∞
to (1 + i)∞ - this we can do since the integrand tends to zero along a horizontal
segment. This gives
(1+i)∞
h−1
X Z
−πi hk λ2 h 2
e eπi k ω dω,
λ=0 −(1+i)∞
q q
h h
or writing t = ω k, k > 0, 310

r (1+i)∞ r
h−1 Z h−1
k X −πi h λ2 πit2 k X −πi k λ2
e k e dt = A e h
h λ=0 h λ=0
−(1+i)∞

where A is the specific constant:

(1+i)∞
Z
2
A= eπit dt
−(1+i)∞

Hence r
k
S (h, k) = AS (−k, h).
h
In order to evaluate A, take a simple case: h = 1, k = 2
√
S (1, 2) = A 2S (−2, 1)
πi √
i.e., 1 + e 2 = A 2,
36. Lecture 228
√
So A = (1 + i)/ 2, an eighth root or unity.
So our reciprocity formula becomes complete:
r
1+i k
S (h, k) = √ S (−k, h).
2 h
Let us develop some corollaries.
1) h = 2, k arbitrary: 311

S (2, k) = G(1, k), so

r
1+i k
G(1, k) = S (2, k) = √ S (−k, 2)
2 2
1+i √ k
= k(1 + e−πi 2 )
2
1+i √
= k(1 + (−i)k )
2
We then have explicitly the value of G(1, k)

(1 + i)(1 + (−i)k ) √
G(1, k) = k.
2
We mention the four cases separately:


 √



 k if k ≡ 1 (mod 4)





0

 if k ≡ 2 (mod 4)
G(1, k) = 


 √



 i k if k ≡ 3 (mod 4)




 √
(1 + i) k if k ≡ 0 (mod 4)
√
Hence the absolute value of G(1, k) can be 0, k or 2k.
So far k was only positive. The case k odd deserves some special mention.
k − 1 is even and 
 √
 k if k−1

2 is even


G(1, k) = 
 √
i k if k−1 is odd.

2
 2
k−1
2 ≡ 0, 1 (mod 4) according as k−12 is even or odd; so we can write this 312
as
k−1 2 √
G(1, k) = i( 2 ) k.
36. Lecture 229

This we have obtained by a purely function-theoretical argument. From our

arithmetical augment, we had, for odd k,
!
h ( k−1 )2 √
G(h, k) = i 2 k
k
 
where hk is the Jacobi symbol. We can get a little more out of it.
!
−1 ( k−1 )2 √
G(−1, k) = i 2 k.
k
Multiplying this and the equation for G(1, k) together,
!
−1 k−1 2
G(1, k)G(−1, k) = (−)( 2 ) k
k
!
−1 k−1
= (−) 2 k
k

But the left side is only G(l, k)G(1, k), and this is always > 0. So
!
−1 k−1
(−) 2 k > 0,
k
and since k > 0 by nature, !
−1 k−1
= (−) 2
k
which is Euler’s criterion for the Jacobi symbol.
2) h = 2, k odd.

r
1+1 k
G(2, k) = S (4, k) = √ b S (−k, 4)
2 4
1+i√ n πik πik o
= √ k 1 + e− 4 + e−πik + e− 4
2 2
1 + i 1 √ √ −πi k
= √ 2 ke 4
2 i
πi √
= e− 4 (k−1) k
πi k−1
√
= e− 2 2 2

k−1 √
= i− 2 k
36. Lecture 230

On the other hand ! 313

2 ( k−1 )2 √
G(2, k) = i 2 k
k
Hence
!
2 k−1 2
= i− 2 −( 2 )
k−1

k
= i− 2 (1+ 2 )
k−1 k−1

k2 −1
= i− 4

k2 −1
= i−2 8

k2 −1
= (−) 8

3) (h, k) = 1; h, k both odd:

r
1+i k
G(h, k) = S (2h, k) = √ S (−k, 2h)
2 2h
r 2h−1
1+i k X πi k λ2
= √ e 2h
2 2h λ=0
r
1+i k X k 2
= √ e−πi 2h λ
2 2h λ mod 2h

Here it is no longer necessary to insist on the special range of summation, 314

for changing λ by λ + 2h would introduce only an even multiple of πi in the
exponent. Separating the odd and even λ′ s, this becomes
r  
1+i k 
 X −πi 2hk (2ℓ)2
 X k
−πi 2h


(2ℓ+h)2 
√  e + e 
2 2h ℓ mod h


ℓ mod h


r
1+i k  hk 
X k 2
= √ 1 + e−πi 2 e−2πi h ℓ
2 2h ℓ mod h
r
1+i   k
= √ 1 + (−i)hk G(−k, h)
2 2h
r
hk−1 2 k
= i( 2 ) G(−k, h)
h
r
( hk−1 2
) k
=i 2 G(k, h)
h
36. Lecture 231

Then we have 315

! r !
h ( k−1 2 √ hk−1 2 k h −( h−1 2 √
i 2 ) (
k=i 2 ) i 2 ) h
k h k
! !
h k 2 h−1 2 k−1 2
= i( 2 ) −( 2 ) −( 2 )
hk−1
i.e.,
k h
ib
1 2 2 
where b= h k − h2 − k2 + 1 − 2(hk − h − k − 1)
4
1
= (h − 1)(k − 1) {(h + 1)(k + 1) − 2}
4 " #
1 (h + 1)(k + 1)
= [(h − 1)(k − 1)] −1
2 2

So
(h−1)(k−1)
ib = i2 4 an odd number
(h−1)(k−1) (h−1)(k−1)
= (−) 4 (odd number) = (−) 4
! !
h k (h−1)(k−1)
∴ = (−) 4 .
k h

which is Jacobi’s law of reciprocity.

We shall use all this in the singular series. It may be worth while to do what 316
Gauss himself did and evaluate G(1, k) by an arithmetical method. To distin-
guish between the different primitive roots of unity is, however, algebraically
impossible; in the analytical method we can use the exponential function to
uniformise the roots of unity.
Lecture 37

We have finished to some extent Gaussian sums; we treated then only in view 317
of their occurrence in the singular series defined as
∞
X
(r)
S (n) = Vk(r) (n)
k=1
!r
X G(h, k) −2πi h n
with Vk(r) (n) = Vk (n) = e k ,
h mod k
k
(h,k)=1

which appeared as the principal term in the expression for the number of rep-
resentation of n as the sum of r squares:

πr/2 r −1 (r)  
Ar (n) =   n 2 S (n) + O nr/4 ,
r
Γ 2

r ≥ 5. We did not bother to do this for lower r, although we could for r = 4,

in which case we know an exact formula; but this is another question. We
consider first a fundamental property of the singular series, viz. its expression
as an infinite product.
Fundamental Lemma.
Yn o
(r)
S (n) = 1 + V p (n) + V p2 (n) + V p3 (n) + · · · ,
p

p prime.
We first prove the multiplicative property of Vk (n): for (k1 , k2 ) = 1,

Vk1 (n)Vk2 (n) = Vk1 k2 (n)

232
37. Lecture 233

We had a similar situation in connection with Ak (n) for the partition func- 318
tion; but there the multiplication was more complicated. Here we have
1 X
−2πi hn
Vk1 k2 (n) = r
G(h, k1 k2 )r e k1 k2 .
(k1 , k2 ) h mod k k
1 2
(h,k1 k2 )=1

Writing h = k2 h1 + k1 h2 with the conditions (h1 , k1 ) = 1 = (h2 , k2 ), h,

running modulo h1 and h2 modulo k2 , this becomes

1 XX −2πi h n
G(k2 h1 + k1 h2 , k1 k2 )e k1 k2
(k1 k2 )r h h
1 2

1 XX
= G ((k2 h1 + k1 h2 )k1 , k2 )r
(k1 k2 )r h h
1 2
−2πi(k2 h1 +k1 h2 ) k nk
G((k2 h1 + k1 h2 )k2 k1 )r e 1 2

on using the multiplicativity of the Gaussian sums; and suppressing multiples

of k1 , k2 , as we may, this gives
1 X X h h
−2πi 1 n−2πi k2 n
r r G(h2 k12 , k2 )r G(k22 h1 , k1 )r e k1 2
k1 k2 h mod k h mod k
1 1 2 2

Now X 319
h 2 2
ℓ
G(ha2 , h) = e2πi k a
ℓ mod k
If (a, k) = 1, al also runs modulo k when ℓ does, so that the right side is
X h 2
e2πi k m = G(h, k)
n mod k

In our case (k1 , k2 ) = 1. So we have

1 X
r −2πi k11 n 1
h X h
−2πi k2 n
G(h 1 , k 1 ) e G(h2 , k2 )r e 2
k1r h mod k k2r h2 mod k2
1 1

= Vk1 (n)Vk2 (n)

We can then break each summand in S n(r) into factors corresponding to

prime powers and multiply them again together, and the rearrangement does
not count because of absolute convergence; so
Yn o
(r)
S (n) = 1 + V p (n) + V p2 (n) + V p3 (n) + · · ·
p
37. Lecture 234
Y
= γ p (n),
p

say; this is an absolutely convergent product. This simplifies matters consid- 320
erably. We have to investigate V only for those G′ s is which prime powers
appear.
We first take p = 2. then
γ2 (n) = 1 + V2 (n) + V22 (n) + · · ·
1 X n
V2λ (n) = λr G(h, 2λ)r e−2πih 2λ
2 λ
h mod 2
2∤h

(i) λ = 1 Since G(h, 2) = 0 for odd h,

V2 (n) = 0
(ii) λ even. For λ ≥ 4,

λ λ
G(h, 2λ) = 2 2 −1 2(1 + ih ) = 2 2 (1 + ih )
1 X h
V2λ (n) = λr 2λr/2 (1 + ih )r e−2πi 2λ n
2 h mod 2λ
2∤h
 


 


1  
 X
r −2πi 2hλ n
X 

r −2πi 2hλ n 
= λr/2  (1 + i) e + (1 − i) e 
2 

 


h≡1 (mod λ4)
 h≡− (mod 4) 

h mod 2 h mod 2λ
 
 
2r/2 
 X
 πi 4r −2πi 2hλ n
X 

−πi 4r −2πi 2hλ n 
= λr/2  e e + e e 
2 

h≡1 (mod 4) h≡−1 (mod 4)




1  
 πi 4r −2πi 2rλ X s
= λ−1  e e−2πi 2λ−2 n +
2 2 r

s mod 2λ−2

X 
−πi 4r +2πi 2rλ s
−2πi λ−2 n

+e e 2 

λ−2

s mod 2
= 0, if 2λ−2 + n;
2λ−2  r ν
λ−1
cos π − 2π , if 2λ−2 /n, n = 2λ−2 .ν
2 2 r 4 4
1 π
i.e., (λ−1)( r −1) cos (2ν − r)
2 4 4
37. Lecture 235

321
Hence, for λ even, λ ≥ 4,


if 2λ−2 + n;

 0,


V2λ (n) = 
 π
(*)
 cos 4 (2ν−r)


r , if 2λ−2 .ν = n.
2(λ−1)( 2 −1)

(iii) λ odd, λ ≥ 3.
λ−3
G(h, 2λ) = 2G(h, 2λ−2) = 2 2 G(h, 23)
λ−3 λ+1
= 2 2 4eπih/4 = 2 2 eπih/4
1 λ+1 X r h
V2λ (n) = λr 2 2 eπih 4 e−2πi 2λ n ,
2 λ
h mod 2
2∤h

or, writing h = 8s + t, t = 1, 3, 5, 7, 322

2λ−3
1 XX n
= λ−1
eπitr/4 e−2πi(8s+t) 2λ
2 2 r
t s=1
2λ−3
1 X X λ−3
= λ−1
e πitr/4−2πitn/2λ
e−2πisn/2
2 2 r
t s=1
= 0, if 2λ−3 ∤ n.

If, however, 2λ−3 |n, n = 2λ−3 .ν, this is

2λ−3 X
λ−1
eπit∤4(r−ν) = o, if 4/(r − ν);
2 2 r
t
λ−1
2
λ−1
eπi(r−ν)|4 , if 4/(r − ν)
2 2 r

1 ν−r
i.e., r · (−) 4 .
2(λ−1)( 2 −1)
Hence for λ odd, λ ≥ 3, 323






 0, if 2λ−3 ∤ n;




V2λ (n) = 
 0, if 2λ−3 . | n, n = 2λ−3 ν, 4 ∤ (ν − r); (**)




 ν−r
 (−) 4

 (λ−1)( r −1) , if 2λ−3 | n, 4 | (ν − r)
2 2
37. Lecture 236

Now, given n, only a finite number of powers of 2 can divide it. So the
situation 2λ−3 /n will occur sometime or the other, so that γ2 (n) is always a
finite sum.
∞
X 1
|γ2 (n) − 1| ≤ r
λ=2
2(λ−1)( 2 −1)
1 1
= ·
2 2 −1 1 − 1/2 2r − 1
r

1
= r/2−1 ;
2 −1
and this is valid for r ≥ 3. so the singular series behaves much better than we
expected.
Lecture 38

It would be of interest to study γ2 (n) also for r = 3, 4. 324

γ2 (n) = 1 + V2 (n) + V22 (n) + · · ·

First consider the case r = 3, 2/n. Then

V2 (n) = 0.

For V2λ (n), λ > 1, we have to make a distinction between λ even and λ odd.
λ even.


if 2λ−2 ∤ n;

 0,


V2λ (n) = 
 π
 cos 4 (2ν−r)


r , if 2λ−2 ∤ n, n = 2λ−2 .ν.
2(λ−1)( 2 −1)
λ odd.







 0, if 2λ−3 ∤ n;




V2λ (n) = 
 0, if 2λ−3 | n, n = 2λ−3 ν, ν − r . 0 (mod 4)




 ν−r
 (−) 4

 (λ−1)( r −1) , if 2λ−3 | n, n = 2λ−3 ν, ν − r ≡ 0 (mod 4)
2 2

So for r = 3,

γ2 (n) = 1 + V4 (n) + V8 (n)

cos π4 (2n − 3) (−) 4
n−3

=1+ √ + ,
2 2

237
38. Lecture 238

where the last summand has to be replaced by 0 if (n − 3)/4 is not an integer. 325
Since 2n − 3 is odd, we have
π 1
| cos (2n − 3)| = √ ,
4 2
and thus clearly,
1 1
|γ2 (n)| ≤ 1 ++ =2
2 2
Moreover, γ2 (n) can vanish. This would require
n−3
(−) 4 = 1
π 1
and cos (2n − 3) = − √
4 2
simultaneously. But this is the case for

n ≡ 7 (mod 8),

as is easily seen. This corresponds to the fact that a number n, n ≡ 7 (mod 8)

cannot be represented as the sum of three squares.
Next take r = 4. We distinguish between the cases 2 ∤ n and 2 | n.

1. 2 ∤ n. Then from relations (*) and (**) proved in lecture 37, we have

γ2 (n) = 1 + V4 (n) + V8 (n)

cos π4 (2n − 4) 1 πn
=1+ = 1 − cos
2 2 2
=1

2. 2 | n Let n = 2α n′ , 2 | n′ . Then (*) and (**) show that V2λ (n) = 0 for 326
λ > α + 3. But actually V2λ (n) = 0 also for λ = α + 3. Indeed, for α odd,
λ = α + 1 is the last even, λ = α + 2 the last odd index for non-vanishing
V2λ (n). For α even, λ = α + 2 is the last even index: λ = α + 3 is odd and
since 4 ∤ (n′ − 4), we have also V2λ (n) = 0 for λ = α + 3.
α+2
X
∴ γ2 (2α n′ ) = 1 + V2λ (n)
λ=2

Now, in V2λ (n), for λ even,

π π
cos (2ν − r) = − cos n′ 2α−λ+2
4 2
38. Lecture 239

= − cos πn′ 2α−λ+1







 −1, for λ ≤ α,



= 1, for λ = α + 1,






0, for λ = α + 2.

Similarly in V2λ (n), for λ odd,

ν−4 1 α−λ+1
(−) 4 = −(−)n .2



−1, for λ ≤ α;


=


1, for λ = α + 1,

and V2λ (n) = 0 for λ = α + 2 since then 4 ∤ 2α−λ+1 . The numerators of the 327
non-vanishing V2λ (n) are −1 upto the last one, which is 1. And thus
1 1 1 1
γ2 (2α n′ ) = 1 − − 2 − · · · − α−1 + α
2 2 2 2
1 1 3
= α−1 + α = α
2 2 2
Although here γ2 (2α n′ ) > 0, we see that for α sufficiently large γ2 (n) can
come arbitrarily close to 0.
We now consider γ p (n) for p ≥ 3.

γ p (n) = 1 + V p (n) + V p2 (n) + · · · ,

1 X −2πi phλ n
where V pλ (n) = λr G(h, pλ )r e
p λ
h mod p
p∤h

Now
!
h
G(h, pλ ) = G(1, pλ)
pλ
!λ  pλ −1 2
h λ
= i 2 p2
p
 2
pλ −1
r 2
!
i X h −2πi phλ n
∴ V pλ (n) = e
pλr/2 p
h mod pλ
p∤h
38. Lecture 240

We have to distinguish between λr odd; and λr even 328

1) λr even. If pλ ≡ 1 (mod 4), then
 2
r pλ −1 λ
r p −1
(−) 2 2
= (−) 2 2

So  2
pλ −1
r 2
i X −2πi phλ n
V pλ (n) = e
pλr/2
h mod pλ
p∤h

2) λr odd. In this case

 2
pλ −1
r 2
!
i X h −2πi phλ n
V pλ (n) = e
pλr/2 p
h mod pλ
p∤h

The inner sum here is a special case of the so-called Ramanujan sums:
X h
Ck (n) = e2πi k n
h mod k
(h,k)=1

There sums can be evaluated. Look at the simpler sums

X λ
S k (n) = e2πi k n
λ mod k



k,

 if k | n;
=



0, if k ∤ n.

Classify the λ′ s in S k (n) according to their common divisor with k. Then 329
X X λ
S k (n) = e2πi k n
d|k λ mod k
(λ,k)=d
X X λ n
= e2πi d · k/d
d|k λ mod k
( λk , dk )=1
X X µn
= e2πi k/d
d|k µ mod k
d
( µ, dk )=1
38. Lecture 241
X
= C dk (n)
d|k
X
= Cd (n).
d|k

Now by Möbious inversion formula,

!
X k
Ck (n) = µ S d (n),
d|k
d

and S d (n) is completely known- it is either 0 or d; hence

!
X k
Ck (n) = dµ
d|k,d|n
d
!
X k
= dµ .
d|(n,k)
d

So these are integers. 330

The Möbious function which appears here arises as a coefficient in a certain
Dirichlet series; in fact
∞
1 X µ(n)
=
ζ(s) n=1 n s
It is possible to build up a complete formal theory of Dirichlet series as we
had in the case of power series. Formal Dirichlet series form a ring without
null-divisors. The multiplication law is given by
X a n X b n X cn
=
n2 n2X ns
where cn = ak b j
k j=n

The relation
X µ(n) X 1
s
=1
nX ns X
then implies that 0= µ( j) · 1 = µ(d), n > 1.
jk=n d|n
Lecture 39

For p ≥ 3 we had 331

γ p (n) = 1 + V p (n) + V p2 (n) + · · ·
where
1 X −2π phλ n
V pλ (n) = G(h, pλ)r e
pλr
h mod pλ
(h,p)=1


 P −2πi phλ n
 2r 

 e , λr even;
pλ −1 
2

 h mod pλ
i 

 p∤h
= 
pλr/2 

 P  h  −2πi phλ n



 p e , λr odd.
h mod pλ


p∤h

For λr odd we have to evaluate this directly. If λr is even it is simpler; it is

a special case of the Ramanujan sums:
X h
Ck (n) = e2πi k n
h mod k
(h,k)=1

which could be evaluated by means of the Möbious inversion formula:

!
X k
Ck (n) = dµ
d|(k,n)
d

So if k is a prime power, k = pλ , 332

!
X −2πi phλ n
X pλ
e = dµ
λ λ
d
h mod p d|(p ,n)
p∤h

242
39. Lecture 243


if α < λ − 1, n = pα n′ , p ∤ n′ ;




0,




−1 × pλ−1 = − − pα .


 if α = λ − 1;
=


−1 × pλ−1 + pλ









= pλ (1 − 1 ),

if α ≥ λ.
p

For obtaining these values we observe that in the summation on the right
λ
side we have to take into account only such divisors d that pα is at most p. This
leads in the first case α < λ − 1 to a vacuous sum. In the second case the only
admissible divisor is pλ−1 ; in the last we have two divisors pλ−1 and pλ . Thus
V pλ (n) = 0
for λ > α + 1; we get again a finite sum for γ p (n)
We now take λr odd. We want
!
X h −2πi phλ n
e
λ
p
h mod p
p∤h

h modulo p is periodic, and we emphasize this by writing

h = rp + s; s = 1, 2, . . . , p − 1; r = 1, . . . , pλ−1
333
So the above sum becomes
λ
pX −1 X
p−1 ! p−1 ! pλ −1
s −2πi (rp+s)
λ
X s −2πi psλ X −2πi pλr−1 n
e p = e e
r=1 s=1
p s=1
p r=1

This is zero when pλ−1 ∤ n (because the inner sum vanishes). Otherwise,
let n = pλ−1 ν and p ∤ ν; then it is again zero because we have only a sum of
quadratic residue symbols (since the character is not the principal character).
If p | ν, the sum becomes
!
ν  p−1 2
√
pλ−1G(ν, p) = pλ−1 i 2 p
p
So if n = pα · n′ where p ∤ n′ , then







 0, if λ − 1 > α;


  n′   p−1 2 √
V pλ (n) = 
 pα p i 2 p, if λ − 1 = α;







 0, if 0 ≤ λ − 1 < α.
39. Lecture 244

So the only non vanishing term in the case α + 1 odd is V pα+1 (n).
Let us put things together now. Let r be even. If p ∤ n, then
 p−1 2
r
i 2
γp = 1 + Vp = 1 −
pr/2
r p−1
(−) 2 2
=1−
pr/2
334
If p | n, n = pα · n′ , then

γ p = 1 + V p + V p2 + · · · + V pα + V pα +1
ǫp ǫ p2
=1+ (p − 1) + p(p − 1) + · · ·
pr/2 p2 r/2
ǫ pα ǫ pα+1
+ α pα−1 (p − 1) − (α+1)r/2 pα ,
p r/2 p
where ǫ p = (−)r(p−1)/4 for r , 4
! !
ǫp ǫp ǫp
= 1 − r/2 + r/2 1 − r/2
p p −1 p
2
ǫ pα
! !
ǫp ǫp ǫp
+   1 − r/2 + · · · + 1 −
pα( 2 −1)
r
p2 2r − 1 p pr/2
! 
ǫ pα+1  ǫ p −1
!
ǫ p 
= 1 − r/2 1 −  1 − r/2−1

p(α+1)( 2 −1)
r
p p

For r = 4, the thing becomes critical: Let us look at it more specifically. 335
r(p − 1)
is even now and so ǫ p = 1. Hence
4
! 1
1 1 − pα+1
γp = 1 − 2
p 1 − 1p

We go to the full singular series.

Y Y
S 4 (n) = γ p = γ2 γp
p p≥3
! 1
Y 1 Y 1 − pα+1
= γ2 1− 2
p≥3
p p≥3 1 − 1p
39. Lecture 245

P 1
The product is convergent since p2
< ∞. So
! 1
Y 1 Y 1 − p2
|S 4 (n)| ≥ γ2 1− 2
p
p p|n 1 − 1p
!2
Y 1 Y 1
≤ γ2 1− 2 1
p
p p|n 1 − p
Q 
1 − 1p diverges to zero in the infinite product senses. So S 4 (n) is not
bounded. S 4 (n) could become very small if we keep the odd factors fixed
and introduce more even factors.
S 4 (n) is unbounded in both senses; it can be as large as we please or as
small as zero.
For r ≥ 5 we are again on the safe side. In this case the first term comes
from V pλ . We have
!
Vp
S 5 (n) ∼ 1 ± 5/2
p
! !
Y 1 Y 1
or C2 1 + 2 < S 5 (n) < C1 1− 2
p p
336
For r = 7 the situation is similar. For r = 6 the series again converges. So
for r ≥ 5.
0 < C1 < S r (n) < C2
This is of importance in the application to our problem.
We had
πr/2 r −1
Ar (n) = n 2 S r (n) + O(nr/4 )
Γ(r/2)
If r ≥ 5, 2r − 1 > 4r , and since S r(n) being bounded does not raise the order
in the term,
πr/2 r −1
Ar (n) ∼ n 2 S r (n)
Γ(r/2)
If, however, if r = 4, the sharpness of the analysis is lost. Both the first
factor and the error term are O(r) and S r (n) may contribute to a decrease in the
first term. If there are many odd factors for n, the main term is still good. But
if there are many powers of 2, it would be completely submerged.
For r = 4 the exact formula was given by Jacobi.
We shall consider also representation of n in the form an21 + bn22 + cn23 + dn24
in which connection the Kloosherman sums appear. We shall also cast a glance
at the meaning of the singular series in the sense of Siegel’e p-edic density.
Lecture 40

Let us look at S r (n) a little more explicitly. 337

S r (n) = γ2 (n)γ3 (n)

r ≡ 0 (mod 4).
In this case we need not bother about the sign of the Gaussian sums; the
fourth power of the coefficient becomes 1.

γ2 (n) = 1 + V2 (n) + V22 (n) + · · ·

which is a finite sum. If 2 ∤ n, then γ2 (n) = 1. If 2 | n, n = 2α n′ , 2 ∤ n′ , then

V2 (n) = 0


 (−)r/4



 (λ−1)( 2r −1)
. if λ < α + 1;


 2


V2λ (n) =  (−)r/4
 − (λ−1) , if λ = α + 1;



 2 ( 2r −1)




0, if λ > α + 1

So
 

 1 

r/4  1 1 1 


γ2 (n) = 1 + (−)  r + + · · · + −   
22( 2 −1) 2(α−1)( 2 −1) 2α 2r − 1 
 −1 r r 
22
 
α n
r
X (−) 2µ
= 1 + (−) 4  ,
µ=1 2 2r − 1
µ

246
40. Lecture 247

if 2α ||n (2α is the highest power of 2 dividing n). If 2 ∤ n, γ2 (n) = 1. 338

! 
1  1 1  α
γ p (n) = 1 − r/2 1 − r/2−1 + · · ·  , p ||n
pα( 2 −1)
r
p p
! Y  
Y 1  1 1 
S r (n) = γ2 (n) 1 − r/2 1 − r/2−1 + · · · + α( r −1) 
p≥3
p p|n,p odd
p p 2
= γ2 (n)P1 · P2 (n),

where P1 is a fixed factor and

Y  1 1 

P2 (n) = 1 − r +···+ 
pα( 2 −1)
r
p|n,p odd
p 2 −1
X 1
= r .
d 2 −1
d|n,d odd
!−1 Y !
1 1
P1 = 1 − 1−
2r/2 p≥2
pr/2
r/2
2 1
= ×  .
2r/2−1 ζ r
2

It is known (vide: Whittaker & Watson) that

(2π)2k B2k
ζ(2k) = (−)k−1 , k ≥ 1,
2(2k)!
where B2k are the Bernoulli numbers. 339
!
1
B1 = − , B3 = B5 = B7 = · · · = 0; B2k , 0; sgnB2k = (−)k−1
2
 
2r/2 2 2r !
P1 = r/2 ×
2 − 1 (2π)r/2|Br/2 |

So for r > 4, the principal term

πr/2 r
Ar (n) ∼   n 2−1 S n (n)
Γ 2r
= Cr (n),
40. Lecture 248

say, where
 
πr/2 2r/2 2 2r ! r
X 1
Cr (n) = r/2 r/2
n 2 −1 γ2 (n) r
−1
Γ(r/2) 2 − 1 (2π) |Br/2| d|n,dodd
d 2

(a divisor sum! which is interesting, but not surprising, because the Jacobi 340
formula contains it).
r r
X 1
Cr (n) = r n 2 −1 r
2 2 −1|Br/2| d −1
2
d|n,d odd

r ≡ 0 (mod 8)

r
X 1
n 2 −1 γ2 (n)·
d|n,d odd
dr/2−1
 
1 + r1 + · · · + 1 1 1 
r
 X
=n 2 −1 −  .
(α−1)( 2r −1) r
2 2 −1)
(
−1 r
 2 2 2 α
d|n,d odd
d 2 −1 
r
X (−) nδ
= n 2 −1 r , if n is even;
δ|n
δ 2 −1
r
X 1
n 2 −1 r , if n is odd.
δ|n
δ 2 −1

So for any n,

r
X 1 r
X (−) nδ
n 2 −1 γ2 (n) =n 2 −1 (−) n
r r
d|n,dodd
d 2 −1 δ|n
δ 2 −1
X  n  2r −1
= (−)n (−)n/δ
δ|n
δ
X r
= (−)n (−)t t 2 −1
t|n

So 341
X r
Cr (n) = Qr (−)n (−)t t 2 −1 ,
t|n
40. Lecture 249

r
here Qr = r
2 2 −1 |B 2r |
This is exactly what appears for r = 4 in the Jacobi formula.
r ≡ 4 (mod 8)

r
X 1
n 2 −1 γ2 (n) r
d 2 −1
d|n,dodd
!X
r
−1 1 1 1 1
=n 2 1− r −···− r + , d odd r −1
2 2 −1 2(α−1)( 2 −1) α(
2 2
r
−1) d2
d|n

r
X (−)δ+ nδ +1
= n 2 −1 r
δ|n
δ 2 −1
X n
 n  2r −1
= (−)δ+ δ +1
δ|n
δ
X n r
= (−) t +t+1 2 −1
t , if n is even;
t|n
X r
t 2 −1 , if n is odd;
t|n

or in either case 342

X n r
(−)n (−)t+ t +1 t 2 −1
t|n
X n r
So Cr (n) = (−)n Qr (−) t +t−1 t 2 −1
t|n
X
(−)t+ 4 ( t +1) t 2 −1 ;
r n r
n
Ar (n) ∼ Qn (−)
t|n
r
where Ar = r
2 2 −1 |Br/2 |
The Bernoulli numbers are all rational numbers and we can show that
2(2r/2 − 1)Br/2 is an odd integral i.e., 2(22k − 1)B2k (k integral) is an all inte- 343
ger. Suppose q is an odd prime; then, by Fermat’s theorem,
2q−1 ≡ 1( mod q)
Let (q − 1) | 2k. Then
22k ≡ 1 (mod q)
40. Lecture 250

22k − 1 ≡ 0 (mod q)

We now appeal to the non-Steadt-Clausen theorem, which is a beautiful

theorem describing fully the denominators of the Bernoulli numbers:
X 1
B2k = G2k −
(p−1)|2k
p

where G2k is an integer.

X 1
∴ (22k − 1)B2k = (22k − 1)G2k − (22k − 1)
(p−1)|2k
p
1
= integer + integer
2
So 2(22k − 1)B2k is an odd integer.
Let us obtain some specimens of
2r
Qr =
(2(2r/2 − 1)|Br/2|)
32
A4 = 8, Q8 = 16, Q12 = 8, Q16 = ,
17
8 16 8 64
Q20 = , Q24 = , Q28 = , q32 =
31 691 5461 929569
The conspicuous prime 691 appears in connection with the representation 344
as the sum 24 squares; it has to do with η24 .
Can Ar (n) be exactly equal to the asymptotic expression? (as for r = 4).
A4 (n) = C4 (n), A8(n) = C8 (n). From Q16 on wards, A16 (n) , C16 (n). This is
because Q16 has an odd prime factor in the denominator. Suppose p divides
the denominator. Then the fraction produced by Q16 cannot be destroyed by
the other factor and Cr (n) is not always an integer. If p | n. the numerator of
Cr (pα ) is congruent to ±1 mod p.
Lecture 41

It might be of interest to take Cr (n), the main term in the formula for Ar (n) and 345
make some remarks about it.
X r −1
(−)n+d+ 4 ( d +1)d 2
r n
Cr (n) = Qr
d|n

Let us form the generating function

∞
X
Hr (x) = 1 + Cr (n)xn ;
n=1

this will give a sort of partial fraction decomposition. In the case where r ≡ 0
(mod 8), it is simpler:
∞
X X r
Hr (x) = 1 + Qr xn (−)n+d d 2 −1
n=1 d|n
∞
X X r
= 1 + Qr (−x)n (−)d d 2 −1
n=1 d|n
∞
X ∞
X
r
= 1 + Qr (−)d d 2 −1 (−x)qd
d=1 q=1
∞
X r (−x)d
= 1 + Qn (−)d d 2 −1
d=1
1 − (−x)d
∞
X r xd
= 1 + Qr d 2 −1
d=1
1 − (−x)d
( )
1.x r x2 r x3
= 1 + Qr + 2 2 −1 + 3 2 −1 +···
1+x 1 − x2 1 + x3

251
41. Lecture 252

This is a Lambert Series. Replacing x by eπiτ , it becomes

( πiτ )
e r
−1 e2πiτ
1 + Qr + 22 + ···
1 + eπiτ 1 − e2πiτ
The series above can be transformed into an Eisen stein series. If r is taken 346
to be 8, it is actually the 8th power of the V −function
Next, take r ≡ 4 (mod 8)
∞
X X n r
Gr (x) = 1 + Qr xn (−)n+d+ d +1 d 2 −1
n=1 d|n
∞
X X
r
= 1 − Qr (−)d d 2 −1 (−x)n (−)n/d
d=1 d|n
∞
X ∞
X
r
= 1 − Qr (−)d d 2 −1 (−x)qd (−)q
d=1 q=1
∞
X r (−x)d
= 1 + Qr (−)d d 2 −1
d=1
1 + (−x)d
∞
X r xd
= 1 + Qr d 2 −1
d=1
1 − (−x)d
( )
1·x r x2 r x3
= 1 + Qr + 2 2 −1 + 3 2 −1 +···
1−x 1 + x2 1 − x3
This is again a Lambert Series. This shows that a V −power has to do with
Lambert series which appears as an evaluation of certain Eisenstein series not
that they are identical.
We now go to something quite different. We had for r ≥ 5,

πr/2 r −1
Ar (n) ∼   n 2 S r (n) (*)
Γ 2r

This comes out as a nice formula. Now could we not make some sense out 347
of
 this formula?
 r What is its inner meaning? We shall show that the first factor
πr/2 /Γ(r/2) n 2 −1 gives the average value of the number of representations of
n as the sum of r squares; the second factor also is an average, in the p-adic
measurement. We shall show that
X πr/2 X r −1
Ar (n) ∼ n2
n≤x
Γ(r/2) n≤x
41. Lecture 253

 So for each
 r individual n, S r (n) gives the deviation of Ar (n) from
πr/2 /Γ(r/2) n 2 −1 ; but on the average there is no deviation.
P
Let us first look at n≤x Ar (n).
X X X
ar (n) = 1
n≤x n≤x m21 +···m2r =n
X
= 1 ,
m21 +···+m2r ≤x

which is the√number of lattice-points in the r-sphere with centre at the origin

and radius x, and so is proportional asymptotically to a certain volume (be-
cause the point lattice has cells or volume 1 and to each
√ points belongs a cell).
So this is roughly the volume of the sphere of radius x which is
Z Z
· · · dx1 · · · dxr
x21 +···+x2n ≤x

πr/2 r/2
= x
Γ(r/2)
348
The difference will not bezero but of the order of magnitude of the surface
of the sphere, i.e., O xr/2 − 1
Now consider the other side.
Z x
πr/2 X r −1 πr/2 r
n2 ∼ V 2 −1 dV
Γ(r/2) n≤x Γ(r/2) 0
πr/2 xr/2
=  
Γ 2r + 1
So the first factor on the right of (*) gives the average. S r (n) has to be
adjusted. S r (n) is also, surprisingly, an average. It was defined as
S r (n) = γ2 (n)γ3 (n)γ5 (n) · · · γ p (n) · · · ,
and γ p (n) in turn was given by 349
∞
X
γ p (n) = 1 + V pλ (n)
λ=1
∞
X 1 X −2πi phλ n 1 X −2πi phλ n
=1+ G(h, pλ)r e G(h, pλ )r e
pλr pλr
λ=1 h mod pλ h mod pλ
p∤h p∤h
41. Lecture 254


1 X 
 X
 2πi h ℓ
X 2πi h ℓ
= λr  e pλ 1 e pλ 2
p 


h mod pλ ℓ mod pλ
1 ℓ mod pλ 2
p∤h

X h


2πi pλ
ℓr 2  −2πi phλ n

e 
 e


ℓr mod pλ
1 X X 2πi h
= λr e pλ (ℓ12 + · · · + ℓr2 − n)
p
ℓ1 ,...,ℓ4 mod pλ h mod pλ
1 X
= λr
p
ℓ1 ,...,ℓr mod pλ
 

 

 X
 2πi psλ (ℓ12 +···+ℓr2 −n)
X t
2πi λ−1 (ℓ12 +···+ℓr2 −n) 


 e − e p



 s mod pλ 

t mod pλ−1
1 X X 2πi s (ℓ2 +···+ℓr2 −n)
= λr e pλ 1
p
ℓ1 ,...ℓn mod pλ s mod pλ
X X 2πi t (ℓ2 +···+ℓ22 −n)
− e pλ−1 1
ℓ1 ,...,ℓr mod pλ−1 t mod pλ−1

= W pλ (n) − W pλ −1 (n), say,

∴ 1 + V p (n) + V p2 (n) + · · · + V pλ (n) = W pλ (n) → γ p (n)
So for λ large enough W pλ (n) = V pλ (n): the partial sums get identical. The 350
value of λ for which this occurs depends on the structure of n, on how many
primes that specific n contains. Now
X 2πi s (ℓ2 +···+ℓ2 −n)
e pλ 1 r
= 0 or pλ
pλ X
∴ W pλ (n) = λr
p λ
ℓ1 ,...,ℓr mod p
ℓ12 +···+ℓr2 ≡n (mod pλ )

The sum on the right gives the number of times the congruence
ℓ12 + · · · + ℓr2 ≡ n (mod pλ ) can be solved, N pλ (n), say. Then
1
W pλ (n) = N pλ (n)
pλ(r−1)
We have therefore divided the number of solutions of the congruence by
pλ(r−1) . Now how many ℓ1 , . . . , ℓr mod pλ do we have? There are pλr possi-
bilities discarding n. n is one of the numbers modulo pλ . So dividing by pr , the
41. Lecture 255

N pλ (n)
average number of possibilities is pλ(r−1) . Hence is the average density
pλ(r−1)
modulo pλ of the solution of the congruence. And since the W pλ (n) eventu-
ally becomes γ p (n), each factor γ p (n) acquires a density interpretation, viz. the
p-adic density of the lattice points modulo pλ .
Lecture 42

The error term in the formula for the number of representations of n as the 351
sum of r squares, r ≥ 5, was O(nr/4 ). For r = 4 this did not suffice. We
shall therefore study the problem by Kloosterman’s method and find out what
happens when we want to decompose n in the form n = n21 + n22 + n23 + n24 .
1
We shall see that we can diminish the order in the error term by nearly 18 .
When Kloosterman did this for the first time (Act a Mathematic as 1927) he
took a slightly more general problem, that of representing n in the form n =
an21 + b22 + cn23 + dn24 , a, b, c, d integers. This works nicely; we get the singular
series and an error term which is smaller than before. The difficult not will be
about the arithmetical interpretation. The singular series will now be a difficult
phenomenon; we shall have multiplicativity, but the interpretation of the factors
γ p becomes complicated. We shall content ourselves with the analytical power
of the discussion. The generating function which will have to be discussed is
quite clear:

F(x) = Θ(xa ) Θ(xb ) Θ(xc ) Θ(xd )

X∞
2
where Θ(x) = xn
n=−∞

And we will have

Z
1 Θ(xa )Θ(xb )Θ(xc )Θ(xd )
A4 (n) = dx
2πi C xn+1
and the analysis goes on as before with Farey series.
We are here representing n by a positive definite quadratic form which is a
diagonal form. Let us make the problem more general.
Let us represent n by a positive definite form with integral coefficients. 352

256
42. Lecture 257

(We could very well unsedes also the ‘semi-integral’ case). Let S be a pos-
itive definite integral symmetric matrix and x a column vector with elements
x1 , x2 , . . . xr in r-space. x′ is the transposed row-vector. x′ S x is a quadratic
form in r variables. The question is how often can we express an integer n by
integer vectors with respect to this quadratic form in r variables.
The generating function to be studied this time is
X ′
Fr (x) = xn S n , |x| < 1,
n

the summation over all integral vectors n. Convergence is easily assured by

positive definiteness. Indeed
x′ S x ≥ C(x21 + · · · + x2r ), C > 0
For x′ S x has a minimum C > 0 on |x| = 1 by positive definiteness; the in-
P 2 2
equality follows from the homogeneity of the quadratic form. And xc(n1 +···+nr )
is trivially a product of convergent series.
In a later paper (Hamburger Abhandlungen, 1927) Kloosterman, on the
advice of Hecke, took up a more general problem. This would require a lit-
tle more preparation on modular forms. The generating function will now be
a modular form of dimension − 2r of a certain ‘stafe’; so we have to discuss
modular forms not only with respect to the full modular group, but also the
substitutions    
a b 1 0
  ≡   (mod N),
   
c d 0 1
(N will the ‘stafe’) which from a subgroup finite index in the modular group. 353
Kloosterman’s work goes through for all modular forms of this sort, but we
should want generalisations of η(τ) and V (τ). To do this we need a good deal
of Heeke’s theory about Eisenstein series of higher stafe of the type:
X 1
m1 ≡a (mod N)
(m1 + m2 τ)r
m2 ≡b (mod N)

which is a modular form of dimension − 2r and stafe N. These were investigated

by Hecke in a famous paper (Hamburger Abhandlungen 1927). Kloosterman
could carry out his theory for these also. We shall, however, compromise on
the quadratic form.
We had the generating function
X ′
Fr (x) = xn S n , |x| < 1,
n
42. Lecture 258

∞
X
=1+ Ar (n)xn .
n=1

Fr (x) is a modular form. This can be seen directly by the transformation

formulae. Let us start with Kloosterman’s method and see what happens. The
problem is to get Z
1 Fn (x)
Ar (n) = dx
2πi C xn+1
At a certain moment later on we shall need a greater knowledge of Fr (x)
Let us carry out the Farey dissection:
h h
x = e2πi k −2πz = e2πi k −2π(δN −iϕ)
′′
X ZVhk
−2πi hk n h
Ar (n) = e Fr (e2πi k −2πz )e2πnzdϕ
0≤h<k≤N −Vhk′

with (h, k) = 1, Vhk′ = k(k11+k) , Vhk′′ = k(k+k

1
2)
where in the Farey situation, 354
h1 h h2
k1 < k < k2 . The refinement of Kloosterman consists in not merely making
the rough remark that
1 1
≤ Vhk′ , Vhk′′ ≤ ,
2kN k(N + 1)
but in a finer following up of the number theoretical determination of the adja-
cent fractions. We have

h1 k − hk1 = −1, hk2 − h2 k = −1;

i.e., hk1 ≡ 1 mod k, hk2 ≡ −1 mod k
h
k is given. What we are worried about is, how long is its environment. k1 and
k2 are given as solutions of certain congruences. We have the habit of calling
′
h a number such that

hh′ ≡ −1 (mod k); so let us write

k1 ≡ −h′ (mod k), k2 ≡ h′ (mod k)

So we know in which residue class modulo k k1 and k2 have to lie. k1 + k, 355

being the denominator of a mediant, had to exced N. N < k1 + k ≤ N + k, or
N − k < k1 ≤ N. So k1 has a span of size k. This along with k1 ≡ −h mod k
determines k1 completely. Similarly, for k2 , N − k < k2 ≤ N So there is no
42. Lecture 259

uncertainty at all about Vhk′ , Vhk′′ ; and we could single them out if we insisted
on that.
For example, let hk = 95 , N = 12; what are the neighbours? hk11 < 5/9 < hk22 .
First determine h′ . 5h′ ≡ −1 (mod 9) or h′ = 7. Then 12 − 9 < k1 ≤ 12 and
k1 ≡ −7 (mod 9), so k1 = 11. Similarly 3 < k2 ≤ 12, k2 ≡ 7 (mod 9) so
k2 = 7. We need only k1 and k2 ; but for our own enjoyment let us calculate h1
and h2 .    
 h, 5 5 h 
  = −1,  2  = −1,
   
11 9 9 7
6
or h1 = 6, h2 = 4, so that we have 11 < 59 < 47 as adjacent fractions in the Farey
series of order 12. We do not need to display the whole Farey series.
Now utilise this in the following way.
1
k(k+k2 )
X′ Z  h 
2πi hk n
Ar (n) = e Fr e2πi k −2πz e2πnz dϕ
o≤h<k≤N
− k(k 1+k)
1

Kloosterman does the following investigation. In any case we are sure that k1 ,
k2 can at most become N. If we take k1 and k2 big we have a small interval of
integration. Since
k1 + k < k1 + 1 + k < · · · < N + k,
k2 + k < k2 + 1 + k < · · · < N + k,
1 1 1 1
> , > ,
k1 + k N + k k2 + k N + k
so that the interval of integration should be at least as big as the interval 356
−1/k(k + N) to 1/k(k + N). This interval is always present whatever be k1
and k2 . So Ar (n) is equal to the always present kernel
1
k(k+N)
Z
X′ h
e2πi k n (· · · )dϕ,
0≤h<k≤N −1
k(k+N)

with the possible additional terms

1 1
k(k+ℓ)
X′ N−1 Z
X X N−1 k(k+ℓ+1)
X Z
h
−2πi hk n
e−2πi k n (· · · )dϕ e (· · · )dϕ
0≤h<k≤N ℓ=k2 1 0≤h<k≤N ℓ=k1 −1
k(k+ℓ+1) k(k+ℓ)
42. Lecture 260

There is no doubt about the integrals. The limits are all well-defined. This
will help us to appraise certain roots of unity closely-by the Kloosterman sums. 357
We shall now return to the integrand; that is a V -function and requires the
usual V treatment. Consider the r-fold V -series:
X ′
Θ(t) = e−πtn S n , Re t > 0.
n

Modify this slightly by introducing a vector α of real numbers;

α′ = (α1 , . . . , αr ). Let
X ′ ′
Θ(t; α1 , . . . , αr ) = e−πt(n +α )S (n+α)
n

This is periodic in α j , of period 1, and so permits a Fourier expansion. The

convergence is so good that the function is analytic in each α j and so we are
sure that it is equal to the sum
X ′
C(m)e2πim α
m

where C(m) is the Fourier coefficient:

Z 1 Z 1
′
C(m) = ··· Θ(t; β1 , . . . , βr )e−2πim β dβ1 , . . . dβr
0 0
Z 1 Z 1 X ′ ′ ′
= ··· e−πt(n +β )S (n+β) e−2πim β dβ1 , . . . , dβr
0 0 n
Z 1 Z 1 X ′ ′ ′
= ··· e−πt(n +β )S (n+β) e−2πim (n+β) dβ1 , . . . , dβr
0 0 n

which is an integral over the unit cube W, and so on translation with respect to 358
the vector n, becomes
XZ Z
′
· · · e−πt(V S V ) e−2πim V dV 1 · · · dVr
n
W+n

(the exchange of integration and summation orders being trivial)

Z ∞ Z ∞
′ ′
= ··· e−πtV DV e−2πim V dV1 . . . dVr .
−∞ −∞
Lecture 43

Let us return to the generalised theta-formula: 359

X ′
Θ(t; α1 , . . . αr ) = e−πt(n +α)S (n+α)
n
X ′
= c(m)e2πim α
m

where
Z ∞ Z
′
S V −2πim′ V
c(m) = ··· e−πtV e dV1 . . . dVr
−∞
To get this into shape, consider the quadratic complement
π π
− (tV ′ + im′ S −1 )S (tV + iS −1 m) = −πtV ′ S V − πiV ′ m − πim′ V + m′ S −1 m
t t
Since m′ V = V ′ m,
Z ∞ Z
π ′ −1
m − πt (tV ′ +im′ S −1 )S (tV +iS −1 m)dV1 ...dVr
c(m) = ··· e− t m S e
−∞
Z ∞ Z √ √
− πt m′ S −1 m −π( tV ′ + √i t m′ S −1 )S ( tV + √i t S −1 m)
=e ··· e dV1 · · · dVr
−∞
√ 1
Put tV = w and µ = √
t
m′ S −1 . Then 360

π ′ −1 Z ∞ Z
e− t m S m ′ ′
c(m) = √ ··· e−π(w +iµ )S (w+iµ) dw1 · · · dwr
( t)r
−∞

261
43. Lecture 262

Since every positive definite quadratic form may be turned into a sum of
squares, we can put S = A′ A, so that the exponent in the integrand become
−π(w′ A′ + iµA′ )(Aw + iAµ); and writing Aw = z, we have

π ′ −1 Z ∞ Z
e− t m S m
′ ′
)(z+iV ) dz1 . . . d zr
c(m) = √ ··· e−π(z +iV
( t)r |A|
−∞

where V = µA, and |A| = determinant of A. Let D = |A|2 = |S |, z′ = (z1 , . . . , zr ).

Then
π ′ −1 r Z
e− t m S m Y ∞ −π(z j +iV j )2
c(m) = 1/2 r/2 e dz j
D t j=1 −∞
π ′ −1 Z ∞ !r
e− t m S m 2
= 1/2 r/2 e−πz dz
D t −∞
π
e− t m′ S −1 m
= ,
D1/2 tr/2
the last factor being unity. So we have ultimately
1 X π ′ −1
m 2πim′ α
Θ(t; α1 , . . . , αr ) = e− t m S e
D1/2 tr/2 m

Let us now we back to our study of Ar (n). We had integrals with now limits 361
which were the special feature of the Kloosterman method.
1
k(k+N)
Z  h  N−1
X N−1
X
P′ −2πi hk n
Ar (n) = e Fr e2πi k −2πz e2πnz dz + ···+ ···
0≤h<k≤N ℓ=0 ℓ=0
1
− k(k+N)

Now
X ∞
X
′
Fr (x) = xn S n = 1 + Ar (n)xn
n n=1
  X
2πi hk −2πz (2πi hk −2πz) ′
Fr e = e n Sn
n
X h ′ ′
= e2πi k n S n e−2πzn S n
n
43. Lecture 263

n is of interest only modulo k, so put

n = kq + ℓ, ℓ = (ℓ1 , . . . , ℓn ), 0 ≤ ℓ j < k.

So dismissing multiples of k,
X X ℓ′
 ℓ
h ′
h ′ 2
(q′ + k )S q+ k
Fr (e2πi k −2πz ) = e2πi k ℓ S ℓ e−2πzk ,
ℓ mod k q

and applying the transformation formula we derived earlier, with t = 2zk2 and
α = 1k ℓ, this becomes

1 X h ′
X − π
m′ S −1 m 2πim′ ℓ
′

√ e2πi k ℓ S ℓ e 2zk2 e k

Dkr er/2 zr/2 ℓ m

1 X − 2zkπ 2 m S −1 m
′
= √ e T k (h, m),
Dkr (2z)r/2 m

on exchanging summations, where 362

X πi ′ ′
T k (h, m) = e2 k (hℓ S ℓ+m ℓ)
ℓ

T k (h, 0) will be the most important; the others we only estimate. We require
a little more number theory for this. We cannot tolerate the presence of a both
a quadratic form and a linear form in the exponent. There will be a common
denominator in m′ S −1 m and that will have to be discussed.
Lecture 44

We had 363
2πi hk −2πz 1 X − π m′ S −1 m
Fr (e )= r r/2 1/2
e 2zk2 T k (h, m),
k (2z) D m
X 2πi ′ ′
and T k (h, m) = e k (hℓ S ℓ+m ℓ)
ℓ mod k
′ −1
The common denominator in m S m will be at most D, the determinant;
define k∗ and Dk by
kD = k · (k, D) · Dk = k∗ Dk , (Dk , k) = 1,
so that Dk is D stripped of all its common divisors with k. Suppose first that k
is odd. Let ρ be a solution of the congruence
4hDk ρ ≡ 1 (mod k∗ )
X
2πi h (ℓ′ S ℓ+4Dk ρm′ ℓ)
T k (h, m) = e k
ℓ mod k
X h ′ ′ −1
)S (ℓ+2Dk ρS −1 m) −(4D2k ρ2 m′ S −1 m)2πi hk
= e2πi k (ℓ +2Dk ρm S e .
ℓ mod k
′ −1
X ′ ′ −1
h 2 2 h
)S (ℓ+2Dk ρS −1 m)
= e−2πi k ·4Dk ρ m S m
e2πi k (ℓ +2Dk ρm S
ℓ mod k
D ρ
−2π kk m′ S −1 m
=e Uk ,
say, (using the definition of ρ), where Uk = Uk (h, m) is periodic in m with 364
period (k, D); it is enough if we take this period to be D itself. So
 
2πi hk −2πz 1 X X D ρ
− π 2 +2πi kk m′ S −1 m
Fr (e )= r Uk (h, s) e 2zk
k (2z)r/2 D1/2 s mod D m≡s (mod D)

264
44. Lecture 265

This is a linear combination of finitely many V -series of the form

X ′ −1
xm S m
m≡s (mod D)

x 1
The power series goes in powers of D because D remains silent inside. This
is for k odd.
For k even, define. σ by

hDk σ ≡ 1 (mod 4k∗ )

X h ′ ′
T k (h, m) = e2πi 4k (4ℓ S ℓ+4Dk σm ℓ)
ℓ mod k
h 2 2 ′ −1
X h ′ ′ −1
−2πi 4k Dk σ m S m )S (2ℓ+Dk σS −1 m)
=e e2πi 4k (2ℓ +Dk σm S
ℓ mod k
D ′ −1
−2πi 4kk σm S m
=e Uk (h, m),

where Uk again has a certain periodicity; we can take the period to be 2D and 365
forget about the refinement. So
 
 h  1 X X − π
+ 2πi ′ −1
4k Dk σ m S m
Fr e2πi k −2πz = Uk (h, s) e 2zk2
kr (2z)r/2 D1/2 s mod 2D m≡s (mod 2D)

which is again a linear combination of theta-series with coefficients Uk . Ob-

serve that T k and Uk differ only by a purely imaginary quantity:

|T k (h, m)| = |Uk (h, m)|,

and for m = 0, T k (h, 0) = Uk (h, 0).

We shall use as essential only those theta-series which are congruent to
zero modulo D or 2D; and the rest will be thrown into the error term. Only
these corresponding to o have a constant term. The general shape in both cases
is X X ′ −1
Uk (h, s) = xm S m
s mod 2D m≡s (mod 2D)
Lecture 45

We have to get a clear picture of that we are aiming at. We are discussing the 366
function under the integral sign. We get it as
 h  1 X
Fr e2πi k −2πz = Uk (h, s)
kr (2z)r/2 D1/2 s mod 2D
 
X π Dk σ
− +2πi 4k
e 2k2 z m′ S −1 m
m≡s (mod 2D)

where k · D = k∗ Dk , (k, Dk ) = 1. k is even; if k is odd the formula looks finitely

many different values. This most important fact we formulate as a lemma.
Lemma 1. For k even, Uk (h, s) depends only on h modulo 2D.
This depends on a theorem on the behaviour of quadratic forms the equiv-
alence of quadratic forms modulo a given number. This is a lemma of Siegel’s
(Annals of Mathematics, 1935, 527-606).
Let us recall that for k even
X i ′ ′
T k (h, m) = e2π k (h ℓ S ℓ +m ℓ)
ℓ mod k
h
−2πi 4k Dk σm′ S −1 m
=e Uk (h, m)

Lemma 2.
|T k (h, m)| ≤ Ckr/2
We have
X h ′
X h ′
S λ +σm′ ℓ)
|T k (h, m)|2 = e2πi k (ℓS ℓ+σm ℓ) e−2πi k (λ
ℓ mod k λ mod k

266
45. Lecture 267
X ′
h
S ℓ − λ′ S λ +σm′ (ℓ − λ))
= e2πi k (ℓ ,
ℓ,λ

and since 367

ℓ′ S ℓ − λ′ S λ = (ℓ′ − λ′ )S (ℓ + λ) + λ′ S ℓ − ℓ′ S λ
= (ℓ′ − λ′ )S (ℓ + λ) + ℓ′ S ℓ − ℓ′ S λ
= (ℓ′ − λ′ )S (ℓ + λ),

this is equal to
X ′
h
− λ′ )(S (ℓ + λ)+σm)
e2πi k (ℓ
ℓ,λ
X X h ′
= e2πi k α (S (ℓ + λ)+σm)
α mod k ℓ − λ≡α mod k
X X h ′
= e2πi k α (S (2 λ +α)+σm)
α mod k ℓ −λ≡ mod k
X h ′
X h ′
= e2πi k α (S α+σm) e2π k 2α S λ
α mod k λ mod k

If we write 2α′ S = β′ , the inner sum is 368

X h
e2πi k (β1 λ1 +···+βr λr ) = k2 , if k | β1 , . . . , k | βr ;
λ1 ,...λr mod k

0 otherwise

So |T k (h, m)|2 = 0 if at least one β is not divisible by k; otherwise it is equal

to X h ′
kr e2πi k α (S α+σm)
α mod k

Writing S = (s jk ), the system of congruences

2α1 s11 + 2α2 s21 + · · · + 2αr sr1 ≡ 0 (mod k)

· · · · ·
· · · · ·
2α1 s1r + 2α2 s2r + · · · + 2αr srr ≡ 0 (mod k)

has at most 2r |S |r solutions, and thus

|T k (h, m)|2 ≤ 2r |S |r kr ,
45. Lecture 268

i.e., |T k (h, m)| ≤ 2r/2 |S |r/2 kr/2 .

We shall now outline the main argument a little more skillfully, putting the 369
thing back on its track. Ar (n) is the sum of integrals over the finer-prepared
Farey arcs of Kloosterman:

1
k(k+N)
Z
1 P′ −2πi hk n e2πnz
Ar (n) = r/2 1/2 e
2 D 0≤h<k≤N zr/2
1
− k(k+N)
X  π 2πi  1
− + D σ
Uk (h, s)Θ s e 2k2 z 4k k dϕ +
s mod 2D
2r/2 D1/2
1 1
k(k+ℓ) − k(k+ℓ+1)
N−1 Z N−1 Z
X h
X 1 X h
X
e−2πi k n + e−2πi k n ··· ,
h,k ℓ=k2
2r/2 D1/2 h,k ℓ=k1
1 1
k(k+ℓ+1) − k(k+ℓ)

where
X ′ −1
Θ s (x) = xm S m
;
m≡s (mod 2D)

= S 0 + S 2 + S 1 , say,
     
 X   X   X 
= S 00 + S 0m  + S 20 + S 2m  + S 10 + S 1m 
m,0 m,0 m,0

in an obvious notation. Now treat the things separately. By inspection of

Uk (h, m) we find how it depends on h, it is only modulo 4 k∗ . We have to
reconcile Lemma 1 with this. This actual period therefore is neither 2D nor
4k∗ but the greatest common divisor
!
∗ ∗ kD
(2D, 4k ) = 2(D, 2k ) = 2 d,
Dk
2 2D
= (DDk , 2kD) = (Dk , 2k)
Dk Dk
2D 4D
= or
Dk Dk
So we have 370
45. Lecture 269

2D
Corollary of Lemma 1. Uk (h, m) for k even depends on h only modulo Dk =
∧, say.
1
l(k+N)
Z
1 X h e2πnz
S 00 = r/2 1/2 e−2πi k n T k (h, o) dϕ
2 D 0≤h<k≤N zr/2
1
− k(k+N)

This goes into the principal term. We shall make it a little more explicit
later.
1
k(k+N)
Z − 2kπ2 z
1 P′ −2πi hk n
D
2πi 4kk σm′ S −1 m e m′ S −1 m
S om = r/2 1/2 e Uk (h, m)e dϕ
2 D 0≤h<k≤N zr/2
1
− k(k+N)
1
k(k+N)
N Z
1 X
= r/2 1/2 Kk (n, m) ··· ,
2 D k=1
1
− k(k+N)

where 371
P′ −2πi hk 2π 4ki Dk σm′ S −1 m
Kk (n, m) = e Uk (h, m)e
h mod k
1 X X 2πiahn.4+2πiV σ
= Uk (λ, m) e− 4k∗ ,
a λ mod ∧ h≡λ (mod ∧)
h mod 4k∗
∗
where 4k∗ = ak, a ≤ 4D, and V = kk Dk m−1 S −1 m
We defined σ by
hDk σ ≡ 1 (mod 4k∗ )
Let
DD̄k ≡ 1 (mod 4k∗ ), hH̄ ≡ 1 (mod 4k∗ )
Then
1 X X −2πiahn+2πiV σ
Kk (n, m) = Uk (λ, m) e 4k∗ Dk D̄k
a λ mod ∧ h≡λ (mod ∧)
h mod k∗ ,(h,k∗ )=1
1 X
e 4k (−4anh+V D̄k h̄)
P′ 2πi
= Uk (λ, m)
a λ mod ∧ h≡λ (mod ∧)
h mod k∗

The inner sum here is a Kloosterman sum. It has essentially 4k∗ terms. A 372
trivial estimate of this would be O(k), and this is what we had tacitly assumed
in the older method. The advantage here is, however, that they can be appraised
letter. We shall not estimate them here but only quote the result as
45. Lecture 270

Lemma 3.
X i
Kk (u, ν) = e2π k (uh+νh̄) , ∧ | k, hh̄ ≡ 1 (mod k)
h≡λ (mod ∧)
h mod k
 
= O k1−α+ǫ (u, k)α

There has been a lot of discussion about the size of the α in this formula.
Kloosterman and Esternann proved that α = 14 (Hamb, Ab. 1930), Salie’ that
α = 1/3 and A.Weil that α = 12 (P.N.A.S’, 48) Weil’s was a very complicated
and deep method going into the zeta-functions of Artin type and the Riemann
hypothesis for these functions.
We thus save a good deal in the order of magnitude. The further S’s will be
nearly similar; the complete Kloosterman sums will be replaced by sums with
certain conditions.
1  
k(k+N) π
N Z −R ·1
(n, k)α π ′ −1 1
X e 2k2 z D
|S om | ≤ C k1−α+ǫ r/2 e− 4 (m S m− D ) dϕ
k=1
k |z|r/2
1
− k(k+N)

Since R k12 z ≥ 1
k on the Farey arc, the integrand is majorised by

π δN
 r/4 π δN
− 2D 1  δN  − 2D
δ−r/4  e k2 (δN +ϕ2 )
k2 |δ2 +ϕ2 2 2 −r/4 2
e N |k (δ2N + ϕ )| = N

2 2 2
k (δN + ϕ )
= O(nr/4 )
√
n
X π ′ −1 1
|S om | ≤ Cn r/4
k1−α+ǫ (n, k)α e− 4 m S m
√ ,
k=1
k n
√
since the path of integration has a length of the order 1/k n. Now summing 373
over all m , 0,
  √
X  X n
r 1
 S om  ≤ Cn 4 − 2 k−α+ǫ (n, k)α
 
m,0 k=1
r 1
X X
−
< Cn 4 2 dα (dt)−α+ǫ
√
d|n dt≤ n
r 1
X X
= Cn 4 − 2 dǫ t−α+ǫ
√
d|n t≤ n
d
45. Lecture 271

√ !1−α+ǫ
r 1
X n
< Cn 4 − 2 dǫ
d|n
d
r α ǫ
X
= Cn 4−2+2 dα−1 ,
d|n

and since the number of divisors of n is O(nǫ/2 ). This is 374

r α ǫ ǫ
= Cn 4 − 2 + 2 + 2
r α
= Cn 4 − 2 +ǫ

Improving α has been the feature of many investigations.

Lecture 46

All the other sums that we have to estimate behave some what similarly. We 375
take as specimen S 20 .
1
k(k+ℓ)
N−1 Z
1 P′ h 1 X e2πnz
S 20 = e−2πi k n T k (h, o) × r × dϕ
D1/2 2r/2 o≤h<k≤N k ℓ=k
zr/2
2 1
k(k+ℓ+1)

1
k(k+ℓ)
N N−1 Z
1 X 1 X e2πnz X h
= 1/2 r/2 e−2πi k n T k (h, o)dϕ
D 2 k=1 kr ℓ=N−k+1
zr/2 h mod k
1
k(k+ℓ+1) N−k<k2 ≤ℓ

The original interval for k2 was bigger: N − k < k2 ≤ N. Now the full
interval is not permissible, i.e., we have admitted not all residues modulo k, but
only a part of these, and the N may lie in two adjacent classes of residues.
Here we have a new type of sum of interest. We know how to discuss T k ; h
plays a role there. The sums we have now get are
X P′ h
T k (λ, o) e−2πi k n
λ mod ∧ h≡λ (mod ∧)
N−k<k2 ≤ℓ

The inner sum is an incomplete Ramanujan sum, with restriction on k2 376

implying (see lecture 45) actually a restriction on h! The Kloosterman sums
are a little more general:
X i
e2π k (uh+V h̄)
hh̄≡1 (mod k)

Our present sums are incomplete Kloosterman sums (with V = o and u =

272
46. Lecture 273

1), and the interesting fact is that they also permit the same appraisal, viz.
 
O kr/2 k1−α+ǫ (k, n)α
From there on things go just as smoothly as before.
 1

 √n k(N+1)
Z 
X dϕ 
S 2o = O  k−r/2 k1−α+ǫ (k, n)α 

k=1
(δ2N + ϕ2 )r/4 
1
k(k+N)

and here for convergence of the integral we want r ≥ 3. This would give again
the old order. Similar estimates hold for the other pieces:
X  α 
S 2m = O nr/4 − + ǫ
m,o
2

(The incomplete Kloosterman sums here are actually incomplete Ramanu-

jan sums and so we may got a slightly better estimate; but this is of no conse-
quence as the other terms have a higher order). 377
We then have
  1
Ar (n) = S oo + O nr/4−α/2−ǫ , α = .
2
Let us look at S oo . It is classical, but not quite what we like it to be.
1
k(k+N)
Z
1 X′ h T k (h, o) e2πnz  
S oo = e−2πi k n dϕ + O nr/4−α/2+ǫ
D1/2 2r/2 o≤h<k≤N
kr z r/2
1
k(k+N)

Replace the integral by an infinite integral:

√
n Z∞
1 X Hk (n) 2πnz 
r/4−α/2+ǫ

dϕ + O n ,
D1/2 2r/2 k=1 kr zr/2
−∞
X h
with Hk (n) = e−2πi k nT k (h,o)
h
 
= O kr/2 k1−α+ǫ (k, n)α ,
thereby adding an error term of order
√ 
 n Z∞ 
X − r +1−α+ǫ α dϕ 
O  k  2 (k, n) 
 k=1 (δ2N + ϕ2 )r/4 
1
kN
46. Lecture 274

Now
Z∞ Z∞
dϕ dϕ 1−r/2
2
(δN + ϕ2 )r/4
=   ϕ 2 r/4 δN
1 1 1 + δN
kN kN
 
 Z∞ 
 r −1 dψ 
= O n  2
2 r/4

 (1 + ϕ ) 
N
k

with ψ = N 2 ϕ. ψ is never smaller than 1 as Nk > 1. So we can drop 1 in the 378

denominator without committing any error in the order of magnitude. So this
gives  Z ∞ 
 r dψ 
O n 2 −1 
N
k
ψr/2
and the integral converging for r ≥ 3, it is equal to
 √ !− 2r +1 
 r −1 n   r 1 r 
O n 2  = O n 4 − 2 k 2 −1
k

Hence our new error term is

√ 
Xn 
r 1  
O  k−α+ǫ (n, k)α n 4 − 2  = O nr/4−α/2+ǫ
k=1

which is what has already appeared.

We than have on writing 2πnz = ω,
√ c+i∞
n Z
1 X Hk (n) 1 r
−1 eω  
Ar (n) = 1/2 r/2 r
(2πn) 2 r/2
dω + O nr/4−α/2+ǫ ,
D 2 k=1 k i ω
c−i∞

and the integral being the Hankel integral for the gamma-function, 379
√
n
(2π)r/2nr/2 − 1 X Hk (n) 1  
Ar (n) = 1/2 r/2 r r/2
+ O nr/4−α/2+ǫ
D 2 k=1
k Γ( )
∞
πr/2 X Hk (n)  
=   nr/2−1 r
+ O nr/4−α/2+ǫ
r
Γ 2 D1/2 k=1
k
46. Lecture 275
 
∞ r
2 +1−α+ǫ (k, n)α 
 r X k 
 −1 
+ O n 2 r 
√ k
k= n+1

This new error term is

   
   
 r −1 X α X r  
 r X X 
(qd)1− 2 −α+ǫ  = O n 2 −1 d1+ǫ−r/2 q−r/2 

O n 2 d
 √   √ 
d|n q> n d|n q> n
d d

(This is because for the Ramanujan sum we have)

!
X′
−2πi hk n
X k
e = dµ
h mod k d|(k,n)
d
 
 X   
= O (k, m) 1 = O (k, n)1+ǫ ;
d|(k,n)

and then we use the old appraisal (k1−α+ǫ (k, n)α with α = 1 + ǫ). So we have 380
 √ !−r/2+1   
 r X n   r 1 X   
O n 2 −1 d1+ǫ−r/2  = O n 4 − 2 dǫ  = O nr/4−1/2+2ǫ

d|n
d d|n

This is of smaller order than the old error term. So we have our final result:
∞
πr/2 r/2−1
X Hk (n) 
r/4−α/2+ǫ

Ar (n) = n + O n ;
Γ(r/2)D1/2 k=1
kr

the singular series plus the error term.

What remains to be shown is that the singular series again enjoys the mul-
tiplicative property:
Hk1 k2 (n) = Hk1 (n)Hk2 (n)
We shall then have it as the product
Y
γp
p

Hk (n) H p2 (n)
where γp = 1 + + +···
pr p2r
The arithmetical interpretation now becomes difficult, because all the prop- 381
erties that the quadratic form may have will have to show up. One or other of
the factors γ p may be zero in which case we have no representation.
46. Lecture 276

We should like to throw some light on the Kloosterman sums. We take for
granted the estimate
P′ i  
e2π k (uh+V h̄) = O k1−α+ǫ · (k, u)α
h mod k
hh̄≡1 (mod k)

Kloosterman and Esterman (Hamburger Abhandlungen Vol.7) proved α = 14 ;

Salie’ (Math. Zeit., vol. 36) proved α = 31 . Using the multiplicativity, in a
certain sense, of the sums, Salie’ could prove that if k = pβ , p prime and β ≥ 2,
then α = 12 but he could not prove this in the other cases. The difficult case was
that of
P′
e2πi/p(uh+V h̄) .
h mod p
 
For this nothing better than O p2/3+ǫ (p, u)1/3 could be obtained; and it de-
fied all efforts until A.Weil proved α = 1/2 in all cases by using deep methods
(Proc. Nat. Acad. Sc.1948). Further application of the Kloosterman sums offer
no difficulty.
The (generalised) Kloosterman sums are symmetrical in u and V , for
X′ 2πi X′ 2πi
e k (uh+V h̄) = e k (uh̄+V h)
h≡λ(∧) h̄≡λ̄(∧)
h mod k h mod k

since (λ, ∧) = 1, h ≡ λ (mod ∧) and hh̄ ≡ 1 (mod ∧) imply h̄ ≡ λ̄ (mod ∧) 382

and λλ̄ ≡ 1 (mod ∧). The last we can write as
P′ 2πi
g(h̄)e k (uh+V h̄) ,
h mod k

where g(m) is the periodic function defined as




1 if m ≡ λ̄ (mod ∧)


g(m) = 



0 otherwise.

g(m) has therefore the finite Fourier expansion

∧
X m
g(m) = C j e2πi j ∧
j=1

The coefficients c j can be calculated in the usual way:

1 −2πiqλ̄
cq = e ∧ mq = 1, 2, . . . , ∧
∧
46. Lecture 277

Substituting for Cq , the sum becomes

 
X P′ h̄ i 1 X λ̄
X jk
2π ki uh+(V + ∧ )h̄
Cj e2πi j ∧ e2π k (uh+V h̄) = e−2πi j ∧ e
j mod ∧
h mod k ∧ j mod ∧ h mod k

so that the generalised sum becomes a finite  combination  of undisturbed

Kloosterman sums and so has the estimate O k1−α+ǫ (k, u)α
This works just as well in the other case when there is an inequality on h̄. 383
X 2πi
X 2πi
e k (uh+V h̄) = f (h̄)e k (uh+V h̄)
h≡λ (mod ∧),h mod k h≡λ (mod ∧)
a≤h̄≤b h mod k



1, 0 < m ≤ a,


where f (m) = 



0, a < m ≤ k,

and f (m) is periodic modulo k.

Then
k
X m
f (m) = c j e2πi j k
j=1

where
2πi j
1 e− k − e−2πi j(a+1)/k
cj = , j , k,
k 1 − e−2πi j/k
a
ck =
k
2
|c j | ≤
k sin π j/k
The sum becomes 384

k−1
X P′ 2πi P′ 2πi
cj e k (uh+(V + j)h̄)
+ ck e k (uh+V h̄)

j=1 h mod k h mod k

h≡λ (mod ∧) h≡λ (mod ∧)
 
 k−1 
 
 1 X 1 

= O k1−α+ǫ (h, k)α 
 
1 + π j 


 k j=1 | sin |  

k

Since sin α ≥ π2 ,
k−1
j
X 1 πX 1
2 ≤2
j=1 sin πkj 2 πj
k
46. Lecture 278

X 1
=k = O(k log k)
k−1
j
j≤ 2

so that again the sum becomes

 
O k1−α+ǫ (k, u)α

Kloosterman first discussed his method for a diagonal quadratic form. Later
on he applied it to modular forms and for this he could derive on the investiga-
tions by Hecke comparing modular forms with Eisenstein series. In this case
the theory becomes simpler: we can subtract suitable Eisenstein series and the
principle term then becomes zero. The r-fold theta-series that we had are in
fact modular forms.