Elliptic Modular Forms and Their Applications

Don Zagier

Max-Planck-Institut für Mathematik, Vivatsgasse 7, 53111 Bonn, Germany

E-mail: [email protected]

Foreword
These notes give a brief introduction to a number of topics in the classical
theory of modular forms. Some of theses topics are (planned) to be treated
in much more detail in a book, currently in preparation, based on various
courses held at the Collège de France in the years 2000–2004. Here each topic
is treated with the minimum of detail needed to convey the main idea, and
longer proofs are omitted.
Classical (or “elliptic”) modular forms are functions in the complex upper
half-plane which transform in a certain way under the action of a discrete
subgroup Γ of SL(2, R) such as SL(2, Z). From the point of view taken here,
there are two cardinal points about them which explain why we are interested.
First of all, the space of modular forms of a given weight on Γ is finite dimen-
sional and algorithmically computable, so that it is a mechanical procedure
to prove any given identity among modular forms. Secondly, modular forms
occur naturally in connection with problems arising in many other areas of
mathematics. Together, these two facts imply that modular forms have a huge
number of applications in other fields. The principal aim of these notes – as
also of the notes on Hilbert modular forms by Bruinier and on Siegel modular
forms by van der Geer – is to give a feel for some of these applications, rather
than emphasizing only the theory. For this reason, we have tried to give as
many and as varied examples of interesting applications as possible. These
applications are placed in separate mini-subsections following the relevant
sections of the main text, and identified both in the text and in the table of
contents by the symbol ♠ . (The end of such a mini-subsection is correspond-
ingly indicated by the symbol ♥ : these are major applications.) The subjects
they cover range from questions of pure number theory and combinatorics to
differential equations, geometry, and mathematical physics.
The notes are organized as follows. Section 1 gives a basic introduction
to the theory of modular forms, concentrating on the full modular group
2 D. Zagier

Γ1 = SL(2, Z). Much of what is presented there can be found in standard

textbooks and will be familiar to most readers, but we wanted to make the
exposition self-contained. The next two sections describe two of the most im-
portant constructions of modular forms, Eisenstein series and theta series.
Here too most of the material is quite standard, but we also include a number
of concrete examples and applications which may be less well known. Sec-
tion 4 gives a brief account of Hecke theory and of the modular forms arising
from algebraic number theory or algebraic geometry whose L-series have Eu-
ler products. In the last two sections we turn to topics which, although also
classical, are somewhat more specialized; here there is less emphasis on proofs
and more on applications. Section 5 treats the aspects of the theory con-
nected with differentiation of modular forms, and in particular the differential
equations which these functions satisfy. This is perhaps the most important
single source of applications of the theory of modular forms, ranging from
irrationality and transcendence proofs to the power series arising in mirror
symmetry. Section 6 treats the theory of complex multiplication. This too is
a classical theory, going back to the turn of the (previous) century, but we try
to emphasize aspects that are more recent and less familiar: formulas for the
norms and traces of the values of modular functions at CM points, Borcherds
products, and explicit Taylor expansions of modular forms. (The last topic
is particularly pretty and has applications to quite varied problems of num-
ber theory.) A planned seventh section would have treated the integrals, or
“periods,” of modular forms, which have a rich combinatorial structure and
many applications, but had to be abandoned for reasons of space and time.
Apart from the first two, the sections are largely independent of one another
and can be read in any order. The text contains 29 numbered “Propositions”
whose proofs are given or sketched and 20 unnumbered “Theorems” which
are results quoted from the literature whose proofs are too difficult (in many
cases, much too difficult) to be given here, though in many cases we have tried
to indicate what the main ingredients are. To avoid breaking the flow of the
exposition, references and suggestions for further reading have not been given
within the main text but collected into a single section at the end. Notations
are standard (e.g., Z, Q, R and C for the integers, rationals, reals and complex
numbers, respectively, and N for the strictly positive integers). Multiplication
precedes division hierarchically, so that, for instance, 1/4π means 1/(4π) and
not (1/4)π.
The presentation in Sections 1–5 is based partly on notes taken by Chris-
tian Grundh, Magnus Dehli Vigeland and my wife, Silke Wimmer-Zagier, of
the lectures which I gave at Nordfjordeid, while that of Section 6 is partly
based on the notes taken by John Voight of an earlier course on complex
multiplication which I gave in Berekeley in 1992. I would like to thank all of
them here, but especially Silke, who read each section of the notes as it was
written and made innumerable useful suggestions concerning the exposition.
And of course special thanks to Kristian Ranestad for the wonderful week in
Nordfjordeid which he organized.
 Elliptic Modular Forms and Their Applications 3

1 Basic Definitions
In this section we introduce the basic objects of study – the group SL(2, R)
and its action on the upper half plane, the modular group, and holomorphic
modular forms – and show that the space of modular forms of any weight and
level is finite-dimensional. This is the key to the later applications.

1.1 Modular Groups, Modular Functions and Modular Forms

The upper half plane, denoted H, is the set of all complex numbers with
positive imaginary part:
 
H = z ∈ C | I(z) > 0 .

The special linear group SL(2, R) acts on H in the standard way by Möbius
transformations (or fractional linear transformations):
 
ab az + b
γ = : H → H, z → γz = γ(z) = .
cd cz + d

To see that this action is well-defined, we note that the denominator is non-
zero and that H is mapped to H because, as a sinple calculation shows,

I(z)
I(γz) = . (1)
|cz + d|2

The transitivity of the action also follows 

by direct
 calculations, or alterna-
tively we can view H as the set of classes of ω 1
ω2 ∈ C 2
| ω2 = 0, I(ω1 /ω2 )>0
under the equivalence relation of multiplication by a non-zero scalar, in which
case the action is given by ordinary matrix multiplication from the left. Notice
that the matrices ±γ act in the same way on H, so we can, and often will,
work instead with the group PSL(2, R) = SL(2, R)/{±1}.
Elliptic modular functions and modular forms are functions in H which
are either invariant or transform in a specific way under the action of a dis-
crete subgroup Γ of SL(2, R). In these introductory notes we will consider
only the group Γ1 = SL(2, Z) (the “full modular group”) and its congruence
subgroups (subgroups of finite index of Γ1 which are described by congruence
conditions on the entries of the matrices). We should mention, however, that
there are other interesting discrete subgroups of SL(2, R), most notably the
non-congruence subgroups of SL(2, Z), whose corresponding modular forms
have rather different arithmetic properties from those on the congruence sub-
groups, and subgroups related to quaternion algebras over Q, which have
a compact fundamental domain. The latter are important in the study of
both Hilbert and Siegel modular forms, treated in the other contributions in
this volume.
4 D. Zagier

The modular group takes its name from the fact that the points of the
quotient space Γ1 \H are moduli (= parameters) for the isomorphism classes
of elliptic curves over C. To each point z ∈ H one can associate the lattice
Λz = Z.z + Z.1 ⊂ C and the quotient space Ez = C/Λz , which is an elliptic
curve, i.e., it is at the same time a complex curve and an abelian group.
Conversely, every elliptic curve over C can be obtained in this way, but not
uniquely: if E is such a curve, then E can be written as the quotient C/Λ
for some lattice (discrete rank 2 subgroup) Λ ⊂ C which is unique up to
“homotheties” Λ → λΛ with λ ∈ C∗ , and if we choose an oriented basis (ω1 , ω2 )
of Λ (one with I(ω1 /ω2 ) > 0) and use λ = ω2−1 for the homothety, then we see
that E ∼ = Ez for some z ∈ H, but choosing a different oriented basis replaces
z by γz for some γ ∈ Γ1 . The quotient space Γ1 \H is the simplest example of
what is called a moduli space, i.e., an algebraic variety whose points classify
isomorphism classes of other algebraic varieties of some fixed type. A complex-
valued function on this space is called a modular function and, by virtue of the
above discussion, can be seen as any one of four equivalent objects: a function
from Γ1 \H to C, a function f : H → C satisfying the transformation equation
f (γz) = f (z) for every z ∈ H and every γ ∈ Γ1 , a function assigning to every
elliptic curve E over C a complex number depending only on the isomorphism
type of E, or a function on lattices in C satisfying F (λΛ) = F (Λ) for all
lattices Λ and all λ ∈ C× , the equivalence between f and F being given in
one direction by f (z) = F (Λz ) and in the other by F (Λ) = f (ω1 /ω2 ) where
(ω1 , ω2 ) is any oriented basis of Λ. Generally the term “modular function”, on
Γ1 or some other discrete subgroup Γ ⊂ SL(2, R), is used only for meromorphic
modular functions, i.e., Γ -invariant meromorphic functions in H which are
of exponential growth at infinity (i.e., f (x + iy) = O(eCy ) as y → ∞ and
f (x + iy) = O(eC/y ) as y → 0 for some C > 0), this latter condition being
equivalent to the requirement that f extends to a meromorphic function on
the compactified space Γ \H obtained by adding finitely many “cusps” to Γ \H
(see below).
It turns out, however, that for the purposes of doing interesting arithmetic
the modular functions are not enough and that one needs a more general class
of functions called modular forms. The reason is that modular functions have
to be allowed to be meromorphic, because there are no global holomorphic
functions on a compact Riemann surface, whereas modular forms, which have
a more flexible transformation behavior, are holomorphic functions (on H and,
in a suitable sense, also at the cusps). Every modular function can be repre-
sented as a quotient of two modular forms, and one can think of the modular
functions and modular forms as in some sense the analogues of rational num-
bers and integers, respectively. From the point of view of functions on lattices,
modular forms are simply functions Λ → F (Λ) which transform under homo-
theties by F (λΛ) = λ−k F (Λ) rather than simply by F (λΛ) = F (Λ) as before,
where k is a fixed integer called the weight of the modular form. If we translate
this back into the language of functions on H via f (z) = F (Λz ) as before, then
we see that f is now required to satisfy the modular transformation property
 Elliptic Modular Forms and Their Applications 5
 
az + b
f = (cz + d)k f (z) (2)
cz + d
 
for all z ∈ H and all ac db ∈ Γ1 ; conversely, given a function f : H → C satis-
fying (2), we can define a funcion on lattices, homogeneous of degree −k with
respect to homotheties, by F (Z.ω1 + Z.ω2 ) = ω2−k f (ω1 /ω2 ). As with modular
functions, there is a standard convention: when the word “modular form” (on
some discrete subgroup Γ of SL(2, R)) is used with no further adjectives, one
generally means “holomorphic
 modular form”, i.e., a function f on H satisfy-
ing (2) for all ac db ∈ Γ which is holomorphic in H and of subexponential
growth at infinity (i.e., f satisfies the same estimate as above, but now for
all rather than some C > 0). This growth condition, which corresponds to
holomorphy at the cusps in a sense which we do not explain now, implies that
the growth at infinity is in fact polynomial; more precisely, f automatically
satisfies f (z) = O(1) as y → ∞ and f (x + iy) = O(y −k ) as y → 0. We denote
by Mk (Γ ) the space of holomorphic modular forms of weight k on Γ . As we
will see in detail for Γ = Γ1 , this space is finite-dimensional, effectively
 com-
putable for all k, and zero for k < 0, and the algebra M∗ (Γ ) := k Mk (Γ ) of
  over C.
all modular forms on Γ is finitely generated
If we specialize (2) to the matrix 10 11 , which belongs to Γ1 , then we see
that any modular form on Γ1 satisfies f (z + 1) = f (z) for all z ∈ H, i.e., it is
a periodic function of period 1. It is therefore a function of the quantity e2πiz ,
traditionally denoted q ; more precisely, we have the Fourier development
∞ ∞
 
f (z) = an e2πinz = an q n z ∈ H, q = e2πiz , (3)
n=0 n=0

where the fact that only terms q n with n ≥ 0 occur is a consequence of (and
in the case of Γ1 , in fact equivalent to) the growth conditions on f just given.
It is this Fourier development which is responsible for the great importance of
modular forms, because it turns out that there are many examples of modular
forms f for which the Fourier coefficients an in (3) are numbers that are of
interest in other domains of mathematics.

1.2 The Fundamental Domain of the Full Modular Group

In the rest of §1 we look in more detail
 at the modular group. Because Γ1
contains the element −1 = −1 0 −1
0
which fixes every point of H, we can
also consider the action of the quotient group Γ 1 = Γ1 /{±1} = PSL(2, Z) ⊂
PSL(2, R) on H. It is clear from (2) that a modular form of odd weight on Γ1
(or on any subgroup of SL(2, R) containing −1) must vanish, so we can restrict
our attention to even k. But then the “automorphy factor” (cz + d)k in (2) is
unchanged when  wereplace γ ∈ Γ1 by −γ, so that we can consider equation (2)
for k even and ac db ∈ Γ 1 . By a slight abuse of notation, we will use the same
notation for an element γ of Γ1 and its image ±γ in Γ 1 , and, for k even, will
not distinguish between the isomorphic spaces Mk (Γ1 ) and Mk (Γ 1 ).
6 D. Zagier
   
The group Γ 1 is generated by the two elements T = 10 11 and S = −0 1 10 ,
with the relations S 2 = (ST )3 = 1. The actions of S and T on H are given by

S : z → −1/z , T : z → z + 1 .

Therefore f is a modular form of weight k on Γ1 precisely when f is periodic

with period 1 and satisfies the single further functional equation
 
f −1/z = z k f (z) (z ∈ H) . (4)

If we know the value of a modular form f on some group Γ at one point

z ∈ H, then equation (2) tells us the value at all points in the same Γ1 -orbit
as z. So to be able to completely determine f it is enough to know the value at
one point from each orbit. This leads to the concept of a fundamental domain
for Γ , namely an open subset F ⊂ H such that no two distinct points of F
are equivalent under the action of Γ and every point z ∈ H is Γ -equivalent to
some point in the closure F of F .

Proposition 1. The set

 
F1 = z ∈ H | |z| > 1, | (z)| < 1
2

is a fundamental domain for the full modular group Γ1 . (See Fig. 1A.)

Proof. Take a point z ∈ H. Then {mz + n | m, n ∈ Z} is a lattice in C.

Every lattice has a point different from the origin of minimal modulus. Let
cz + d be such a point. The integers c, d must be relatively prime (otherwise
we could divide cz + d by an integer to get a new point in thelattice
 of even
smaller modulus). So there are integers a and b such that γ1 = ac db ∈ Γ1 . By
the transformation property (1) for the imaginary part y = I(z) we get that
I(γ1 z) is a maximal member of {I(γz) | γ ∈ Γ1 }. Set z ∗ = T n γ1 z = γ1 z + n,

Fig. 1. The standard fundamental domain for Γ 1 and its neighbors

 Elliptic Modular Forms and Their Applications 7

where n is such that | (z ∗ )| ≤ 12 . We cannot have |z ∗ | < 1, because then

we would have I(−1/z ∗ ) = I(z ∗ )/|z ∗ |2 > I(z ∗ ) by (1), contradicting the
maximality of I(z ∗ ). So z ∗ ∈ F1 , and z is equivalent under Γ1 to z ∗ .
Now suppose that we had two Γ1 -equivalent points z1 and z2 = γz1 in F1 ,
with γ = ±1. This γ cannot be of the form Tn since  this would contradict
the condition | (z1 )|, | (z2 )| < 12 , so γ = ac db with c = 0. Note that
√
I(z) > 3/2 for all z ∈ F1 . Hence from (1) we get
√
3 I(z1 ) I(z1 ) 2
< I(z2 ) = ≤ 2 < √ ,
2 |cz1 + d|2 c I(z1 )2 c2 3
which can only be satisfied if c = ±1. Without loss of generality we may
assume that Iz1 ≤ Iz2 . But |±z1 +d| ≥ |z1 | > 1, and this gives a contradiction
with the transformation property (1).
Remarks. 1. The points on the borders of the fundamental region are Γ1 -
equivalent as follows: First, the points on the two lines (z) = ± 12 are equiva-
lent by the action of T : z → z + 1. Secondly, the points on the left and right
halves of the arc |z| = 1 are equivalent under the action of S : z → −1/z. In
fact, these are the only equivalences for the points on the boundary. For this
reason we define F1 to be the semi-closure of F1 where we have added only the
boundary points with non-positive real part (see Fig. 1B). Then every point
of H is Γ1 -equivalent to a unique point of F1 , i.e., F1 is a strict fundamental
domain for the action of Γ1 . (But terminology varies, and many people use
the words “fundamental domain” for the strict fundamental domain or for its
closure, rather than for the interior.)
2. The description of the fundamental domain F1 also implies the above-
mentioned fact that Γ1 (or Γ 1 ) is generated by S and T . Indeed, by the very
definition of a fundamental domain we know that F1 and its translates γF1 by
elements γ of Γ1 cover H, disjointly except for their overlapping boundaries
(a so-called “tesselation” of the upper half-plane). The neighbors of F1 are
T −1 F1 , SF1 and T F1 (see Fig. 1C), so one passes from any translate γF1
of F1 to one of its three neighbors by applying γSγ −1 or γT ±1 γ −1 . In parti-
cular, if the element γ describing the passage from F1 to a given translated
fundamental domain F1 = γF1 can be written as a word in S and T , then
so can the element of Γ1 which describes the motion from F1 to any of the
neighbors of F1 . Therefore by moving from neighbor to neighbor across the
whole upper half-plane we see inductively that this property holds for every
γ ∈ Γ1 , as asserted. More generally, one sees that if one has given a fundamen-
tal domain F for any discrete group Γ , then the elements of Γ which identify
in pairs the sides of F always generate Γ .

♠ Finiteness of Class Numbers

Let D be a negative discriminant, i.e., a negative integer which is congruent

to 0 or 1 modulo 4. We consider binary quadratic forms of the form Q(x, y) =
8 D. Zagier

Ax2 + Bxy + Cy 2 with A, B, C ∈ Z and B 2 − 4AC = D. Such a form is

definite (i.e., Q(x, y) = 0 for non-zero (x, y) ∈ R2 ) and hence has a fixed sign,
which we take to be positive. (This is equivalent to A > 0.) We also assume
that Q is primitive, i.e., that gcd(A, B, C) = 1. Denote by QD the set of
these forms. The group Γ1 (or indeed Γ 1 ) acts on QD by Q → Q ◦ γ, where
(Q ◦ γ)(x, y) = Q(ax + by, cx + dy) for γ = ± ac db ∈ Γ1 . We claim that the
number of equivalence classes under this action is finite. This number, called
the class number of D and denoted h(D), also has an interpretation as the
number of ideal classes (either for the ring of integers or, if D is a non-trivial
square multiple of some other √ discriminant, for a non-maximal order) in the
imaginary quadratic field Q( D), so this claim is a special case – historically
the first one, treated in detail in Gauss’s Disquisitiones Arithmeticae – of
the general theorem that the number of ideal classes in any number field is
finite. To prove it, we observe
√ that we can associate to any Q ∈ QD the
unique
√ root zQ = (−B + D)/2A of Q(z, 1) = 0 in the upper half-plane (here
D = +i |D| by definition and A > 0 by assumption). One checks easily
that zQ◦γ = γ −1 (zQ ) for any γ ∈ Γ 1 , so each Γ 1 -equivalence class of forms
Q ∈ QD has a unique representative belonging to the set
 
Qred
D = [A, B, C] ∈ QD | −A < B ≤ A < C or 0 ≤ B ≤ A = C
(5)
of Q ∈ QD for which zQ ∈ F1 (the so-called reduced quadratic forms of
discriminant D), and this set is finite because C ≥ A ≥ |B| implies |D| =
4AC − B 2 ≥ 3A2 , so that both A and B are bounded in absolute value by
|D|/3, after which C is fixed by C = (B 2 − D)/4A. This even gives us a way
to compute h(D) effectively, e.g., Qred −47 = {[1, 1, 12], [2, ±1, 6], [3, ±1, 4]}
and hence h(−47) = 5. We remark that the class numbers h(D), or a small
modification of them, are themselves the coefficients of a modular form (of
weight 3/2), but this will not be discussed further in these notes. ♥

1.3 The Finite Dimensionality of Mk (Γ )

We end this section by applying the description of the fundamental domain

to show that Mk (Γ1 ) is finite-dimensional for every k and to get an upper
bound for its dimension. In §2 we will see that this upper bound is in fact the
correct value.
If f is a modular form of weight k on Γ1 or any other discrete group Γ ,
then f is not a well-defined function on the quotient space Γ \H, but the
transformation formula (2) implies that the order of vanishing ordz (f ) at
a point z ∈ H depends only on the orbit Γ z. We can therefore define a local
order of vanishing, ordP (f ), for each P ∈ Γ \H. The key assertion is that the
total number of zeros of f , i.e., the sum of all of these local orders, depends
only on Γ and k. But to make this true, we have to look more carefully at
the geometry of the quotient space Γ \H, taking into account the fact that
some points (the so-called elliptic fixed points, corresponding to the points
 Elliptic Modular Forms and Their Applications 9

z ∈ H which have a non-trivial stabilizer for the image of Γ in PSL(2, R)) are
singular and also that Γ \H is not compact, but has to be compactified by the
addition of one or more further points called cusps. We explain this for the
case Γ = Γ1 .
In §1.2 we identified the quotient space Γ1 \H as a set with the semi-
closure F1 of F1 and as a topological space with the quotient of F1 obtained
by identifying the opposite sides (lines (z) = ± 12 or halves of the arc |z| = 1)
of the boundary ∂F1 . For a generic point of√F1 the stabilizer subgroup of Γ 1
is trivial. But the two points ω = 12 (−1 + i 3) = e2πi/3 and i are stabilized
by the cyclic subgroups of order 3 and 2 generated by ST and S respectively.
This means that in the quotient manifold Γ1 \H, ω and i are singular. (From
a metric point of view, they have neighborhoods which are not discs, but
quotients of a disc by these cyclic subgroups, with total angle 120◦ or 180◦
instead of 360◦.) If we define an integer nP for every P ∈ Γ1 \H as the order
of the stabilizer in Γ 1 of any point in H representing P , then nP equals 2
or 3 if P is Γ1 -equivalent to i or ω and nP = 1 otherwise. We also have to
consider the compactified quotient Γ1 \H obtained by adding a point at infinity
(“cusp”) to Γ1 \H. More precisely, for Y > 1 the image in Γ1 \H of the part of H
above the line I(z) = Y can be identified via q = e2πiz with the punctured
disc 0 < q < e−2πY . Equation (3) tells us that a holomorphic modular form
of any weight k on Γ1 is not only a well-defined function on this punctured
disc, but extends holomorphically to the point q = 0. We therefore define
Γ1 \H = Γ1 \H ∪ {∞}, where the point “∞” corresponds to q = 0, with q as
a local parameter. One can also think of Γ1 \H as the quotient of H by Γ1 , where
H = H ∪ Q ∪ {∞} is the space obtained by adding the full Γ1 -orbit Q ∪ {∞} of
∞ to H. We define the order of vanishing at infinity of f , denoted ord∞ (f ),
as the smallest integer n such that an = 0 in the Fourier expansion (3).

Proposition 2. Let f be a non-zero modular form of weight k on Γ1 . Then

1 k
ordP (f ) + ord∞ (f ) = . (6)
nP 12
P ∈Γ1 \H

Proof. Let D be the closed set obtained from F1 by deleting ε-neighborhoods

of all zeros of f and also the “neighborhood of infinity” I(z) > Y = ε−1 , where
ε is chosen sufficiently small that all of these neighborhoods are disjoint (see
Fig. 2.) Since f has no zeros in D, Cauchy’s theorem implies that the integral
  f  (z)
of d log f (z) = dz over the boundary of D is 0. This boundary consists
f (z)
of several parts: the horizontal line from − 12 + iY to 12 + iY , the two vertical
lines from ω to − 12 + iY and from ω + 1 to 12 + iY (with some ε-neighborhoods
removed), the arc of the circle |z| = 1 from ω to ω + 1 (again with some
ε-neighborhoods deleted), and the boundaries of the ε-neighborhoods of the
zeros P of f . These latter have total angle 2π if P is not an elliptic fixed point
10 D. Zagier

Fig. 2. The zeros of a modular form

(they consist of a full circle if P is an interior point of F1 and of two half-circles

if P corresponds to a boundary point of F1 different from ω, ω + 1 or i), and
total angle π or 2π/3 if P ∼ i or ω. The corresponding contributions to the
integral are as follows. The two vertical lines together give 0, because f takes
on the same value on both and the orientations are opposite. The horizontal
line from − 12 +iY to 12 +iY gives a contribution 2πi ord∞ (f ), because d(log f )
is the sum of ord∞ (f ) dq/q and a function of q which is holomorphic at 0,
and this integral corresponds to an integral around a small circle |q| = e−2πY
around q = 0. The integral on the boundary of the deleted ε-neighborhood of
a zero P of f contributes 2πi ordP (f ) if nP = 1 by Cauchy’s theorem, because
ordP (f ) is the residue of d(log f (z)) at z = P , while for nP > 1 we must
divide by nP because we are only integrating over one-half or one-third of the
full circle around P . Finally, the integral along the bottom arc contributes
πik/6, as we see by breaking up this arc into its left and right halves and
applying the formula d log f (Sz) = d log f (z) + kdz/z, which is a consequence
of the transformation equation (4). Combining all of these terms with the
appropriate signs dictated by the orientation, we obtain (6). The details are
left to the reader.

Corollary 1. The dimension of Mk (Γ1 ) is 0 for k < 0 or k odd, while for

even k ≥ 0 we have

[k/12] + 1 if k ≡ 2 (mod 12)

dim Mk (Γ1 ) ≤ (7)
[k/12] if k ≡ 2 (mod 12) .

Proof. Let m = [k/12] + 1 and choose m distinct non-elliptic points Pi ∈

Γ1 \H. Given any modular forms f1 , . . . , fm+1 ∈ Mk (Γ1 ), we can find a linear
 Elliptic Modular Forms and Their Applications 11

combination f of them which vanishes in all Pi , by linear algebra. But then

f ≡ 0 by the proposition, since m > k/12, so the fi are linearly dependent.
Hence dim Mk (Γ1 ) ≤ m. If k ≡ 2 (mod 12) we can improve the estimate by 1
by noticing that the only way to satisfy (6) is to have (at least) a simple
zero
 at i and a double zero at ω (contributing a total of 1/2 + 2/3 = 7/6 to
ordP (f )/nP ) together with k/12 − 7/6 = m − 1 further zeros, so that the
same argument now gives dim Mk (Γ1 ) ≤ m − 1.

Corollary 2. The space M12 (Γ1 ) has dimension ≤ 2, and if f, g ∈ M12 (Γ1 )
are linearly independent, then the map z → f (z)/g(z) gives an isomorphism
from Γ1 \H ∪ {∞} to P1 (C).

Proof. The first statement is a special case of Corollary 1. Suppose that f

and g are linearly independent elements of M12 (Γ1 ). For any (0, 0) = (λ, μ) ∈
C2 the modular form λf − μg of weight 12 has exactly one zero in Γ1 \H ∪ {∞}
by Proposition 2, so the modular function ψ = f /g takes on every value
(μ : λ) ∈ P1 (C) exactly once, as claimed.

We will make an explicit choice of f , g and ψ in §2.4, after we have introduced

the “discriminant function” Δ(z) ∈ M12 (Γ1 ).
The true interpretation of the factor 1/12 multiplying k in equation (6)
is as 1/4π times the volume of Γ1 \H, taken with respect to the hyperbolic
metric. We say only a few words about this, since these ideas will not be
used again. To give a metric on a manifold is to specify the distance be-
tween any two sufficiently near points. The hyperbolic metric in H is defined
by saying that the hyperbolic distance between two points in a small neigh-
borhood of a point z = x + iy ∈ H is very nearly 1/y times the Euclidean
distance between them, so the volume element, which in Euclidean geometry is
given by the 2-form dx dy, is given in hyperbolic geometry by dμ = y −2 dx dy.
Thus
  1/2  ∞ 
 dy
Vol Γ1 \H) = dμ = √ 2
dx
F1 −1/2 1−x2 y
 1/2 1/2
dx  π
= √ = arcsin(x) = .
−1/2 1−x 2
−1/2 3

Now we can consider other discrete subgroups of SL(2, R) which have a fun-
damental domain of finite volume. (Such groups are usually called Fuchsian
groups of the first kind, and sometimes “lattices”, but we will reserve this lat-
ter term for discrete cocompact subgroups of Euclidean spaces.) Examples
are the subgroups Γ ⊂ Γ1 of finite index, for which the volume of Γ \H is π/3
times the index of Γ in Γ1 (or more precisely, of the image of Γ in PSL(2, R)
in Γ 1 ). If Γ is any such group, then essentially the same proof as for Pro-
position 2 shows that the number of Γ -inequivalent zeros of any non-zero
12 D. Zagier

modular form f ∈ Mk (Γ ) equals k Vol(Γ \H)/4π, where just as in the case

of Γ1 we must count the zeros at elliptic fixed points or cusps of Γ with
appropriate multiplicities. The same argument as for Corollary 1 of Propos-
ition 2 then tells us Mk (Γ ) is finite dimensional and gives an explicit upper
bound:

Proposition 3. Let Γ be a discrete subgroup of SL(2, R) for which Γ \H has

kV
finite volume V . Then dim Mk (Γ ) ≤ + 1 for all k ∈ Z.
4π
In particular, we have Mk (Γ ) = {0} for k < 0 and M0 (Γ ) = C, i.e., there
are no holomorphic modular forms of negative weight on any group Γ , and
the only modular forms of weight 0 are the constants. A further consequence
is that any three modular forms on Γ are algebraically dependent. (If f, g, h
were algebraically independent modular forms of positive weights, then for
large k the dimension of Mk (Γ ) would be at least the number of monomials
in f , g, h of total weight k, which is bigger than some positive multiple of k 2 ,
contradicting the dimension estimate given in the proposition.) Equivalent-
ly, any two modular functions on Γ are algebraically dependent, since every
modular function is a quotient of two modular forms. This is a special case
of the general fact that there cannot be more than n algebraically indepen-
dent algebraic functions on an algebraic variety of dimension n. But the most
important consequence of Proposition 3 from our point of view is that it is
the origin of the (unreasonable?) effectiveness of modular forms in number
theory: if we have two interesting arithmetic sequences {an }n≥0 and {bn }n≥0
and conjecture that they are identical (and clearly many results of  number
theory can be formulated in this way), then if we can show that both an q n
n
and bn q are modular forms of the same weight and group, we need only
verify the equality an = bn for a finite number of n in order to know that it
is true in general. There will be many applications of this principle in these
notes.

2 First Examples: Eisenstein Series

and the Discriminant Function
In this section we construct our first examples of modular forms: the Eisenstein
series Ek (z) of weight k > 2 and the discriminant function Δ(z) of weight 12,
whose definition is closely connected to the non-modular Eisenstein series
E2 (z).

2.1 Eisenstein Series and the Ring Structure of M∗ (Γ1 )

There are two natural ways to introduce the Eisenstein series. For the first,
we observe that the characteristic transformation equation (2) of a modular
 Elliptic Modular Forms and Their Applications 13

form can be written in the form f |k γ = f for γ ∈ Γ , where f |k γ : H → C is

defined by
   
   az + b  ab 
f k g (z) = (cz + d)−k f z ∈ C, g = ∈ SL(2, R) .
cz + d cd
(8)
One checks easily that for fixed k ∈ Z, the map f → f |k g defines an operation
of the group SL(2, R) (i.e., f |k (g1 g2 ) = (f |k g1 )|k g2 for all g1 , g2 ∈ SL(2, R))
on the vector space of holomorphic functions in H having subexponential
or polynomial growth. The space Mk (Γ ) of holomorphic modular forms of
weight k on a group Γ ⊂ SL(2, R) is then simply the subspace of this vector
space fixed by Γ .
If we have a linear action v → v|g of a finite group G on a vector space V ,
then an obvious way to construct a G-invariant vector  in V is to start with
an arbitrary vector v0 ∈ V and form the sum v = g∈G v0 |g (and to hope
that the result is non-zero). If the vector v0 is invariant under some sub-
group G0 ⊂ G, then the vector v0 |g depends only  on the coset G0 g ∈ G0 \G
and we can form instead the smaller sum v = g∈G0 \G v0 |g, which again is
G-invariant. If G is infinite, the same method sometimes applies, but we now
have to be careful about convergence. If the vector v0 is fixed by an infinite
subgroup G0 of G, then this improves our chances because thesum over G0 \G
is much smaller than a sum over all of G (and in any case g∈G v|g has no
chance of converging since every term occurs infinitely often). In the context
when G = Γ ⊂ SL(2, R) is a Fuchsian group (acting by |k ) and v0 a ra-
tional function, the modular forms obtained in this way are called Poincaré
series. An especially easy case is that when v0 is the constant function “1”
and Γ0 = Γ∞ , the stabilizer of the cusp at infinity. In this case the series

Γ∞ \Γ 1|k γ is called an Eisenstein series.  
Let us look at this series more carefully when Γ = Γ1 . A matrix ac db ∈
SL(2, R) sends ∞ to a/c, and hence belongs  to the stabilizer of ∞ if and only
if c = 0. In Γ1 these are the matrices ± 10 n1 with n ∈ Z, i.e., up to sign the
matrices T n . We can assume that k is even (since there are no modular forms
of odd weight on Γ1 ) and hence work with Γ1 = PSL(2, Z), in which case the
stabilizer Γ ∞ is the infinite
 a b  cyclic group 1generated
 by T . If we multiply an
arbitrary matrix γ = on the left by n , then the resulting matrix γ  =
 a+nc b+nd  c d 0 1   
c d has the same bottom row as γ. Conversely, if γ  = ac bd ∈ Γ1 has
the same bottom row as γ, then from (a −a)d−(b −b)c = det(γ)−det(γ  ) = 0
and (c, d) = 1 (the elements of any row or column of a matrix in SL(2, Z) are
coprime!) we see that a − a = nc, b − b = nd for some n ∈ Z, i.e., γ  = T n γ.
Since every coprime pair of integers occurs as the bottom row of a matrix in
SL(2, Z), these considerations give the formula
  1 1
Ek (z) = 1k γ = 1k γ = (9)
2 (cz + d)k
γ∈Γ∞ \Γ1 γ∈Γ ∞ \Γ 1 c, d∈Z
(c,d) = 1
14 D. Zagier

for the Eisenstein series (the factor 12 arises because (c d) and (−c − d) give
the same element of Γ1 \Γ 1 ). It is easy to see that this sum is absolutely
convergent for k > 2 (the number of pairs (c, d) with N ≤ |cz + d| < N + 1
is the number of lattice points in an annulusof area π(N + 1)2 − πN 2 and
∞
hence is O(N ), so the series is majorized by N =1 N 1−k ), and this absolute
convergence guarantees the modularity (and, since it is locally uniform in z,
also the holomorphy) of the sum. The function Ek (z) is therefore a modular
form of weight k for all even k ≥ 4. It is also clear that it is non-zero, since for
I(z) → ∞ all the terms in (9) except (c d) = (±1 0) tend to 0, the convergence
of the series being sufficiently uniform that their sum also goes to 0 (left to
the reader), so Ek (z) = 1 + o(1) = 0.
The second natural way of introducing the Eisenstein series comes from
the interpretation of modular forms given in the beginning of §1.1, where we
identified solutions of the transformation equation (2) with functions on lat-
tices Λ ⊂ C satisfying the homogeneity condition F (λΛ) = λ−k F (Λ) under
homotheties Λ → λΛ. An obvious way to produce such a homogeneous  func-
tion – if the series converges – is to form the sum Gk (Λ) = 12 λ∈Λ0 λ−k of
the (−k)th powers of the non-zero elements of Λ. (The factor “ 12 ” has again
been introduce to avoid counting the vectors λ and −λ doubly when k is
even; if k is odd then the series vanishes anyway.) In terms of z ∈ H and its
associated lattice Λz = Z.z + Z.1, this becomes

1 1
Gk (z) = (k > 2, z ∈ H) , (10)
2 (mz + n)k
m, n∈Z
(m,n)=(0,0)

where the sum is again absolutely and locally uniformly convergent for k > 2,
guaranteeing that Gk ∈ Mk (Γ1 ). The modularity can also be seen directly by
noting that (Gk |k γ)(z) = m,n (m z + n )−k where (m , n ) = (m, n)γ runs
over the non-zero vectors of Z2  {(0, 0)} as (m, n) does.
In fact, the two functions (9) and (10) are proportional, as is easily seen:
any non-zero vector (m, n) ∈ Z2 can be written uniquely as r(c, d) with r (the
greatest common divisor of m and n) a positive integer and c and d coprime
integers, so

Gk (z) = ζ(k) Ek (z) , (11)


where ζ(k) = r≥1 1/rk is the value at k of the Riemann zeta function. It
may therefore seem pointless to have introduced both definitions. But in fact,
this is not the case. First of all, each definition gives a distinct point of view
and has advantages in certain settings which are encountered at later points
in the theory: the Ek definition is better in contexts like the famous Rankin-
Selberg method where one integrates the product of the Eisenstein series with
another modular form over a fundamental domain, while the Gk definition is
better for analytic calculations and for the Fourier development given in §2.2.
 Elliptic Modular Forms and Their Applications 15

Moreover, if one passes to other groups, then there are σ Eisenstein series of
each type, where σ is the number of cusps, and, although they span the same
vector space, they are not individually proportional. In fact, we will actually
want to introduce a third normalization
(k − 1)!
Gk (z) = Gk (z) (12)
(2πi)k

because, as we will see below, it has Fourier coefficients which are rational
numbers (and even, with one exception, integers) and because it is a normal-
ized eigenfunction for the Hecke operators discussed in §4.
As a first application, we can now determine the ring structure of M∗ (Γ1 )
Proposition 4. The ring M∗ (Γ1 ) is freely generated by the modular forms E4
and E6 .
Corollary. The inequality (7) for the dimension of Mk (Γ1 ) is an equality for
all even k ≥ 0.

Proof. The essential point is to show that the modular forms E4 (z) and
E6 (z) are algebraically independent. To see this, we first note that the forms
E4 (z)3 and E6 (z)2 of weight 12 cannot be proportional. Indeed, if we had
E6 (z)2 = λE4 (z)3 for some (necessarily non-zero) constant λ, then the
meromorphic modular form f (z) = E6 (z)/E4 (z) of weight 2 would satisfy
f 2 = λE4 (and also f 3 = λ−1 E6 ) and would hence be holomorphic (a function
whose square is holomorphic cannot have poles), contradicting the inequal-
ity dim M2 (Γ1 ) ≤ 0 of Corollary 1 of Proposition 2. But any two modular
forms f1 and f2 of the same weight which are not proportional are necessarily
algebraically independent. Indeed, if P (X, Y ) is any polynomial in C[X, Y ]
such that P (f1 (z), f2 (z)) ≡ 0, then by considering the weights we see that
Pd (f1 , f2 ) has to vanish identically for each homogeneous component Pd of P .
But Pd (f1 , f2 )/f2d = p(f1 /f2 ) for some polynomial p(t) in one variable, and
since p has only finitely many roots we can only have Pd (f1 , f2 ) ≡ 0 if f1 /f2
is a constant. It follows that E43 and E62 , and hence also E4 and E6 , are alge-
braically independent. But then an easy calculation shows that the dimension
of the weight k part of the subring of M∗ (Γ1 ) which they generate equals the
right-hand side of the inequality (7), so that the proposition and corollary
follow from this inequality.

2.2 Fourier Expansions of Eisenstein Series

Recallfrom (3) that any modular form on Γ1 has a Fourier expansion of the
∞
form n=0 an q n , where q = e2πiz . The coefficients an often contain interesting
arithmetic information, and it is this that makes modular forms important
for classical number theory. For the Eisenstein series, normalized by (12), the
coefficients are given by:
16 D. Zagier

Proposition 5. The Fourier expansion of the Eisenstein series Gk (z) (k even,

k > 2) is
∞
Bk
Gk (z) = − + σk−1 (n) q n , (13)
2k n=1
where Bk is the kth Bernoulli number and where σk−1 (n) for n ∈ N denotes
the sum of the (k − 1)st powers of the positive divisors of n.
We
∞recall that the Bernoulli numbers are defined by the generating function
k=0 Bk x /k! = x/(e − 1) and that the first values of Bk (k > 0 even) are
k x

given by B2 = 6 , B4 = − 30
1 1 1
, B6 = 42 , B8 = − 30
1 5
, B10 = 66 , B12 = − 2730
691
,
7
and B14 = 6 .
Proof. A well known and easily proved identity of Euler states that
1 π  
= z ∈C\Z , (14)
z+n tan πz
n∈Z

where the sum on the left, which is not absolutely

 convergent, is to be inter-
preted as a Cauchy principal value ( = lim N −M where M, N tend to infinity
with M − N bounded). The function on the right is periodic of period 1 and
its Fourier expansion for z ∈ H is given by
 ∞ 
π cos πz eπiz + e−πiz 1+q 1
= π = πi πiz = −πi = −2πi + q r
,
tan πz sin πz e − e−πiz 1−q 2 r=1

where q = e2πiz . Substitute this into (14), differentiate k − 1 times and divide
by (−1)k−1 (k − 1)! to get
  ∞
1 (−1)k−1 dk−1 π (−2πi)k
= = rk−1 q r
(z + n)k (k − 1)! dz k−1 tan πz (k − 1)! r=1
n∈Z
(k ≥ 2, z ∈ H) ,
an identity known as Lipschitz’s formula. Now the Fourier expansion of Gk
(k > 2 even) is obtained immediately by splitting up the sum in (10) into the
terms with m = 0 and those with m = 0:
∞ ∞ ∞
1 1 1 1 1 1
Gk (z) = + = +
2 nk 2 (mz + n)k n=1
n k
m=1 n=−∞
(mz + n)k
n∈Z m, n∈Z
n=0 m=0
k ∞ ∞
(2πi)
= ζ(k) + rk−1 q mr
(k − 1)! m=1 r=1
 ∞ 
(2πi)k Bk
= − + σk−1 (n) q n ,
(k − 1)! 2k n=1

where in the last line we have used Euler’s evaluation of ζ(k) (k > 0 even) in
terms of Bernoulli numbers. The result follows.
 Elliptic Modular Forms and Their Applications 17

The first three examples of Proposition 5 are the expansions

1
G4 (z) = + q + 9q 2 + 28q 3 + 73q 4 + 126q 5 + 252q 6 + · · · ,
240
1
G6 (z) = − + q + 33q 2 + 244q 3 + 1057q 4 + · · · ,
504
1
G8 (z) = + q + 129q 2 + 2188q 3 + · · · .
480
The other two normalizations of these functions are given by
16 π 4 π4
G4 (z) = G4 (z) = E4 (z) , E4 (z) = 1 + 240q + 2160q 2 + · · · ,
3! 90
64 π 6 π6
G6 (z) = − G6 (z) = E6 (z) , E6 (z) = 1 − 504q − 16632q 2 − · · · ,
5! 945
256 π 8 π8
G8 (z) = G8 (z) = E8 (z) , E8 (z) = 1 + 480q + 61920q 2 + · · · .
7! 9450
Remark. We have discussed only Eisenstein series on the full modular group
in detail, but there are also various kinds of Eisenstein series for subgroups
Γ ⊂ Γ1 . We give one example. Recall that a Dirichlet character modulo
N ∈ N is a homomorphism χ : (Z/N Z)∗ → C∗ , extended to a map χ : Z → C
(traditionally denoted by the same letter) by setting χ(n) equal to χ(n mod N )
if (n, N ) = 1 and to 0 otherwise. If χ is a non-trivial Dirichlet character and
k a positive integer with χ(−1) = (−1)k , then there is an Eisenstein series
having the Fourier expansion
∞  
Gk,χ (z) = ck (χ) + χ(d) d k−1
qn
n=1 d|n

which is a “modular form of weight k and character χ on Γ0 (N ).” (This  ameans


that Gk,χ ( az+b
cz+d ) = χ(a)(cz + d)k
G k,χ (z) for any z ∈ H and any b
c d ∈
SL(2, Z) with c ≡ 0 (mod N ).) Here ck (χ) ∈ Q is a suitable constant, given
L(1 − k, χ), where L(s, χ) is the analytic continuation
explicitly by ck (χ) = 12
∞
of the Dirichlet series n=1 χ(n)n−s .
The simplest example, for N = 4 and χ = χ−4 the Dirichlet character
modulo 4 given by
⎧
⎪
⎨+1 if n ≡ 1 (mod 4) ,
χ−4 (n) = −1 if n ≡ 3 (mod 4) , (15)
⎪
⎩
0 if n is even
and k = 1, is the series
⎛ ⎞
∞
⎝ 1
G1,χ−4 (z) = c1 (χ−4 ) + χ−4 (d)⎠ q n = + q + q 2 + q 4 + 2q 5 + q 8 + · · · .
n=1
4
d|n
(16)
18 D. Zagier

1
(The fact that L(0, χ−4 ) = 2c1 (χ−4 ) = is equivalent via the functional
2
1 1
equation of L(s, χ−4 ) to Leibnitz’s famous formula L(1, χ−4 ) = 1 − + −
3 5
π
· · · = .) We will see this function again in §3.1.
4

♠ Identities Involving Sums of Powers of Divisors

We now have our first explicit examples of modular forms and their Fourier
expansions and can immediately deduce non-trivial number-theoretic identi-
ties. For instance, each of the spaces M4 (Γ1 ), M6 (Γ1 ), M8 (Γ1 ), M10 (Γ1 ) and
M14 (Γ1 ) has dimension exactly 1 by the corollary to Proposition 2, and is
therefore spanned by the Eisenstein series Ek (z) with leading coefficient 1, so
we immediately get the identities

E4 (z)2 = E8 (z) , E4 (z)E6 (z) = E10 (z) ,

E6 (z)E8 (z) = E4 (z)E10 (z) = E14 (z) .

Each of these can be combined with the Fourier expansion given in Proposit-
ion 5 to give an identity involving the sums-of-powers-of-divisors functions
σk−1 (n), the first and the last of these being
n−1
σ7 (n) − σ3 (n)
σ3 (m)σ3 (n − m) = ,
m=1
120
n−1
σ13 (n) − 11σ9 (n) + 10σ3 (n)
σ3 (m)σ9 (n − m) = .
m=1
2640

Of course similar identities can be obtained from modular forms in higher

weights, even though the dimension of Mk (Γ1 ) is no longer equal to 1. For
instance, the fact that M12 (Γ1 ) is 2-dimensional and contains the three modu-
lar forms E4 E8 , E62 and E12 implies that the three functions are linearly
dependent, and by looking at the first two terms of the Fourier expansions we
find that the relation between them is given by 441E4 E8 + 250E62 = 691E12 ,
a formula which the reader can write out explicitly as an identity among sums-
of-powers-of-divisors functions if he or she is so inclined. It is not easy to obtain
any of these identities by direct number-theoretical reasoning (although in fact
it can be done). ♥

2.3 The Eisenstein Series of Weight 2

In §2.1 and §2.2 we restricted ourselves to the case when k > 2, since then
the series (9) and (10) are absolutely convergent and therefore define modular
forms of weight k. But the final formula (13) for the Fourier expansion of Gk (z)
converges rapidly and defines a holomorphic function of z also for k = 2, so
 Elliptic Modular Forms and Their Applications 19

in this weight we can simply define the Eisenstein series G2 , G2 and E2 by

equations (13), (12), and (11), respectively, i.e.,
∞
1 1
G2 (z) = − + σ1 (n) q n = − + q + 3q 2 + 4q 3 + 7q 4 + 6q 5 + · · · ,
24 n=1
24
6
G2 (z) = −4π G2 (z) ,
2
E2 (z) = G2 (z) = 1 − 24q − 72q 2 − · · · .
π2
(17)
Moreover, the same proof as for Proposition 5 still shows that G2 (z) is given
by the expression (10), if we agree to carry out the summation over n first
and then over m :
1 1 1 1
G2 (z) = 2
+ . (18)
2 n 2 (mz + n)2
n=0 m=0 n∈Z

The only difference is that, because of the non-absolute convergence of the

double series, we can no longer interchange the order of summation to get
the modular transformation equation G2 (−1/z) = z 2 G2 (z). (The equation
G2 (z + 1) = G2 (z), of course, still holds just as for higher weights.) Never-
theless, the function G2 (z) and its multiples E2 (z) and G2 (z) do have some
modular properties and, as we will see later, these are important for many
applications.
 
Proposition 6. For z ∈ H and ac db ∈ SL(2, Z) we have
 
az + b
G2 = (cz + d)2 G2 (z) − πic(cz + d) . (19)
cz + d
Proof. There are many ways to prove this. We sketch one, due to Hecke, since
the method is useful in many other situations. The series (10) for k = 2
does
 not converge absolutely, but it is just at the edge of convergence, since
−λ
m,n |mz + n| converges for any real number λ > 2. We therefore modify
the sum slightly by introducing
1  1
G2,ε (z) = (z ∈ H, ε > 0) . (20)
2 m, n (mz + n) |mz + n|2ε
2


(Here means that the value (m, n) = (0, 0) is to be omitted from
the summation.)
 The new series converges absolutely and transforms by
G2,ε az+b
cz+d = (cz + d)2 |cz + d|2ε G2,ε (z). We claim that lim G2,ε (z) exists
ε→0
and equals G2 (z) − π/2y, where y = I(z). It follows that each of the three
non-holomorphic functions
π 3 1
G∗2 (z) = G2 (z) − , E2∗ (z) = E2 (z) − , G∗2 (z) = G2 (z) +
2y πy 8πy
(21)
transforms like a modular form of weight 2, and from this one easily deduces
the transformation equation (19) and its analogues for E2 and G2 . To prove
20 D. Zagier

the claim, we define a function Iε by

 ∞
dt  
Iε (z) = z ∈ H, ε > − 21 .
−∞ (z + t) |z + t|
2 2ε

Then for ε > 0 we can write

∞ ∞
1
G2,ε − Iε (mz) =
m=1 n=1
n2+2ε
∞ ∞   n+1 
1 dt
+ − .
m=1 n=−∞
(mz + n)2 |mz + n|2ε n (mz + t)2 |mz + t|2ε

Both sums on the right converge absolutely and locally uniformly  for ε > −21
(the second one because the expression in square brackets is O |mz+n|−3−2ε
by the mean-value theorem, which tells us that f (t) − f (n) for any differen-
tiable function f is bounded in n ≤ t ≤ n + 1 by maxn≤u≤n+1 |f  (u)|), so
the limit of the expression on the right as ε → 0 exists and can be obtained
simply by putting ε = 0 in each term, where it reduces to G2 (z) by (18). On
the other hand, for ε > − 21 we have
 ∞
dt
Iε (x + iy) = 2 2 2 ε
−∞ (x + t + iy) ((x + t) + y )
 ∞
dt I(ε)
= 2 2 2 ε
= 1+2ε ,
−∞ (t + iy) (t + y ) y
∞ 
where I(ε) = −∞ (t+i)−2 (t2 +1)−ε dt , so ∞ m=1 Iε (mz) = I(ε)ζ(1+2ε)/y
1+2ε

for ε > 0. Finally, we have I(0) = 0 (obvious),

 ∞  ∞
log(t2 + 1) 1 + log(t2 + 1) 

I (0) = − 2
dt = − tan t 
−1
= −π,
−∞ (t + i) t+i −∞

1
and ζ(1 + 2ε) = + O(1), so the product I(ε)ζ(1 + 2ε)/y 1+2ε tends to −π/2y
2ε
as ε → 0. The claim follows.
Remark. The transformation equation (18) says that G2 is an example of what
is called a quasimodular form, while the functions G∗2 , E2∗ and G∗2 defined in
(21) are so-called almost holomorphic modular forms of weight 2. We will
return to this topic in Section 5.

2.4 The Discriminant Function and Cusp Forms

For z ∈ H we define the discriminant function Δ(z) by the formula

∞
  24
Δ(z) = e2πiz 1 − e2πinz . (22)
n=1
 Elliptic Modular Forms and Their Applications 21

(The name comes from the connection with the discriminant of the elliptic
curve Ez = C/(Z.z + Z.1), but we will not discuss this here.) Since |e2πiz | < 1
for z ∈ H, the terms of the infinite product are all non-zero and tend exponen-
tially rapidly to 1, so the product converges everywhere and defines a holomor-
phic and everywhere non-zero function in the upper half-plane. This function
turns out to be a modular form and plays a special role in the entire theory.
Proposition 7. The function Δ(z) is a modular form of weight 12 on SL(2, Z).

Proof. Since Δ(z) = 0, we can consider its logarithmic derivative. We find

∞ ∞
1 d n e2πinz
log Δ(z) = 1−24 = 1−24 σ1 (m) e2πimz = E2 (z) ,
2πi dz n=1
1 − e 2πinz
m=1

e2πinz
where the second equality follows by expanding as a geometric
∞ 2πirnz 1 − e2πinz
series r=1 e and interchanging the order of summation, and the third
equality from the definition of E2 (z) in (17). Now from the transformation
equation for E2 (obtained by comparing (19) and(11)) we find
     
1 d Δ az+b
cz+d 1 az + b 12 c
log 12
= 2
E2 − − E2 (z)
2πi dz (cz + d) Δ(z) (cz + d) cz + d 2πi cz + d
= 0.

In other words, (Δ|12 γ)(z) = C(γ) Δ(z) for all z ∈ H and all γ ∈ Γ1 , where
C(γ) is a non-zero complex number depending only on γ, and where Δ|12 γ is
defined as in (8). It remains to show that C(γ) = 1 for all γ. But C : Γ1 → C∗
is a homomorphism because Δ → Δ|12γ is a group  action,
 so it suffices to
check this for the generators T = 10 11 and S = 01 −1 0 of Γ1 . The first is
obvious since Δ(z) is a power series in e2πiz and hence periodic of period 1,
while the second follows by substituting z = i into the equation Δ(−1/z) =
C(S) z 12 Δ(z) and noting that Δ(i) = 0.

Let us look at this function Δ(z) more carefully. We know from Corollary 1
to Proposition 2 that the space M12 (Γ1 ) has dimension at most 2, so Δ(z)
must be a linear combination of the two functions E4 (z)3 and E6 (z)2 . From
the Fourier expansions E43 = 1 + 720q + · · · , E6 (z)2 = 1 − 1008q + · · · and
Δ(z) = q + · · · we see that this relation is given by
1  
Δ(z) = E4 (z)3 − E6 (z)2 . (23)
1728
This identity permits us to give another, more explicit, version of the fact that
every modular form on Γ1 is a polynomial in E4 and E6 (Proposition 4). In-
deed, let f (z) be a modular form of arbitrary even weight k ≥ 4, with Fourier
expansion as in (3). Choose integers a, b ≥ 0 with 4a + 6b = k (this is always
22 D. Zagier
 
possible) and set h(z) = f (z)− a0 E4 (z)a E6 (z)b /Δ(z). This function is holo-
morphic in H (because Δ(z) = 0) and also at infinity (because f − a0 E4a E6b
has a Fourier expansion with no constant term and the Fourier expansion of
Δ begins with q), so it is a modular form of weight k − 12. By induction on
the weight, h is a polynomial in E4 and E6 , and then from f = a0 E4a E6b + Δh
and (23) we see that f also is.
In the above argument we used that Δ(z) has a Fourier expansion begin-
ning q+O(q 2 ) and that Δ(z) is never zero in the upper half-plane. We deduced
both facts from the product expansion (22), but it is perhaps worth noting
that this is not necessary: if we were simply to define Δ(z) by equation (23),
then the fact that its Fourier expansion begins with q would follow from the
knowledge of the first two Fourier coefficients of E4 and E6 , and the fact that
it never vanishes in H would then follow from Proposition 2 because the total
number k/12 = 1 of Γ1 -inequivalent zeros of Δ is completely accounted for
by the first-order zero at infinity.
We can now make the concrete normalization of the isomorphism between
Γ1 \H and P1 (C) mentioned after Corollary 2 of Proposition 2. In the notation
of that proposition, choose f (z) = E4 (z)3 and g(z) = Δ(z). Their quotient is
then the modular function

E4 (z)3
j(z) = = q −1 + 744 + 196884 q + 21493760 q 2 + · · · ,
Δ(z)
called the modular invariant. Since Δ(z) = 0 for z ∈ H, this function is finite
in H and defines an isomorphism from Γ1 \H to C as well as from Γ1 \H to
P1 (C).
The next (and most interesting) remarks about Δ(z) concern its Fourier
expansion. By multiplying out the product in (22) we obtain the expansion
∞
 ∞

Δ(z) = q 1 − q n )24 = τ (n) q n (24)
n=1 n=1

where q = e2πiz as usual (this is the last time we will repeat this!) and the
coefficients τ (n) are certain integers, the first values being given by the table
n 1 2 3 4 5 6 7 8 9 10
τ (n) 1 −24 252 −1472 4830 −6048 −16744 84480 −113643 −115920
Ramanujan calculated the first 30 values of τ (n) in 1915 and observed several
remarkable properties, notably the multiplicativity property that τ (pq) =
τ (p)τ (q) if p and q are distinct primes (e.g., −6048 = −24 · 252 for p = 2,
q = 3) and τ (p2 ) = τ (p)2 − p11 if p is prime (e.g., −1472 = (−24)2 − 2048 for
p = 2). This was proved by Mordell the next year and later generalized by
Hecke to the theory of Hecke operators, which we will discuss in §4.
√
Ramanujan also observed that |τ (p)| was bounded by 2p5 p for primes
p < 30, and conjectured that this holds for all p. This property turned out
 Elliptic Modular Forms and Their Applications 23

to be immeasurably deeper than the assertion about multiplicativity and was

only proved in 1974 by Deligne as a consequence of his proof of the famous Weil
conjectures (and of his previous, also very deep, proof that these conjectures
implied Ramanujan’s). However, the weaker inequality |τ (p)| ≤ Cp6 with some
effective constant C > 0 is much easier and was proved in the 1930’s by Hecke.
We reproduce Hecke’s proof, since it is simple. In fact, the proof applies to
a much more general class of modular forms. Let us call a modular form on Γ1
a cusp form if the constant term a0 in the Fourier expansion (3) is zero. Since
the constant term of the Eisenstein series Gk (z) is non-zero, any modular form
can be written uniquely as a linear combination of an Eisenstein series and
a cusp form of the same weight. For the former the Fourier coefficients are
given by (13) and grow like nk−1 (since nk−1 ≤ σk−1 (n) < ζ(k − 1)nk−1 ). For
the latter, we have:
Proposition
 8. Let f (z) be a cusp form of weight k on Γ1 with Fourier expan-
sion ∞ n=1 na q n
. Then |an | ≤ Cnk/2 for all n, for some constant C depending
only on f .
Proof. From equations (1) and (2) we see that the function z → y k/2 |f (z)|
on H is Γ1 -invariant. This function tends rapidly to 0 as y = I(z) → ∞
(because f (z) = O(q) by assumption and |q| = e−2πy ), so from the form of
the fundamental domain of Γ1 as given in Proposition 1 it is clearly bounded.
Thus we have the estimate
|f (z)| ≤ c y −k/2 (z = x + iy ∈ H) (25)
for some c > 0 depending only on f . Now the integral representation
 1
an = e2πny f (x + iy) e−2πinx dx
0

for an , valid for any y > 0, show that |an | ≤ cy −k/2 e2πny . Taking y = 1/n (or,
optimally, y = k/4πn) gives the estimate of the proposition with C = c e2π
(or, optimally, C = c (4πe/k)k/2 ).
Remark. The definition of cusp forms given above is actually valid only for
the full modular group Γ1 or for other groups having only one cusp. In general
one must require the vanishing of the constant term of the Fourier expansion
of f , suitably defined, at every cusp of the group Γ , in which case it again
follows that f can be estimated as in (25). Actually, it is easier to simply define
cusp forms of weight k as modular forms for which y k/2 f (x + iy) is bounded,
a definition which is equivalent but does not require the explicit knowledge of
the Fourier expansion of the form at every cusp.

♠ Congruences for τ (n)

As a mini-application of the calculations of this and the preceding sections

we prove two simple congruences for the Ramanujan tau-function defined by
24 D. Zagier

equation (24). First of all, let us check directly that the coefficient τ (n) of q n
of the function defined by (23) is integral for all n. (This fact is, of course,
obvious from equation (22).) We have

(1 + 240A)3 − (1 − 504B)2 A−B

Δ = = 5 + B + 100A2 − 147B 2 + 8000A3
1728 12
∞ ∞ (26)
with A = n=1 σ3 (n)q n and B = n=1 σ5 (n)q n . But σ5 (n)−σ3 (n) is divisib-
le by 12 for every n (because 12 divides d5 − d3 for every d), so (A − B)/12
has integral coefficients. This gives the integrality of τ (n), and even a congru-
ence modulo 2. Indeed, we actually have σ5 (n) ≡ σ3 (n) (mod 24), because
d3 (d2 − 1) is divisible by 24 for every d, so (A − B)/12 has  even coeffi-
cients
 and (26) gives Δ ≡ B + B 2
(mod 2)
 or, recalling that ( an q n )2 ≡
2n n
an q (mod 2) for every power series an q with integral coefficients,
τ (n) ≡ σ5 (n) + σ5 (n/2) (mod 2), where σ5 (n/2) is defined as 0 if 2  n. But
σ5 (n), for any integer n, is congruent modulo 2 to the sum of the odd divisors
of n, and this is odd if and only if n is a square or twice a square, as one
sees by writing n = 2s n0 with n0 odd and pairing the complementary divisors
of n0 . It follows that σ5 (n) + σ5 (n/2) is odd if and only if n is an odd square,
so we get the congruence:

1 (mod 2) if n is an odd square ,

τ (n) ≡ (27)
0 (mod 2) otherwise .

In a different direction, from dim M12 (Γ1 ) = 2 we immediately deduce the

linear relation
 
691 E4 (z)3 E6 (z)3
G12 (z) = Δ(z) + +
156 720 1008

and from this a famous congruence of Ramanujan,

 
τ (n) ≡ σ11 (n) (mod 691) ∀n ≥ 1 , (28)

where the “691” comes from the numerator of the constant term −B12 /24 of
G12 . ♥

3 Theta Series
If Q is a positive definite integer-valued quadratic form in m variables, then
there is an associated modular form of weight m/2, called the theta series
of Q, whose nth Fourier coefficient for every integer n ≥ 0 is the number of
representations of n by Q. This provides at the same time one of the main
constructions of modular forms and one of the most important sources of
applications of the theory. In 3.1 we consider unary theta series (m = 1), while
 Elliptic Modular Forms and Their Applications 25

the general case is discussed in 3.2. The unary case is the most classical, going
back to Jacobi, and already has many applications. It is also the basis of the
general theory, because any quadratic form can be diagonalized over Q (i.e.,
by passing to a suitable sublattice it becomes the direct sum of m quadratic
forms in one variable).

3.1 Jacobi’s Theta Series

The simplest theta series, corresponding to the unary (one-variable) quadratic

form x → x2 , is Jacobi’s theta function

2
θ(z) = q n = 1 + 2q + 2q 4 + 2q 9 + · · · , (29)
n∈Z

where z ∈ H and q = e2πiz as usual. Its modular transformation properties

are given as follows.
Proposition 9. The function θ(z) satisfies the two functional equations
  
−1 2z
θ(z + 1) = θ(z) , θ = θ(z) (z ∈ H) . (30)
4z i

Proof. The first equation in (30) is obvious since θ(z) depends only on q.
For the second, we use the Poisson transformation formula. Recall that this
formula says that  f : R → C which is smooth
for any function and small at
 ∞ 2πixy
infinity, we have n∈Z f (n) = n∈Z f (n), where  f (y) = −∞
e f (x) dx is
the Fourier transform of f . (Proof : the sum n∈Z f (n + x) is convergent and
defines a function g(x) which is periodic of period 1 and hence has a Fourier
 1
expansion g(x) = n∈Z cn e
2πinx
with cn = 0 g(x)e−2πinx dx = f (−n),
   
so n f (n) = g(0) = n cn = n f (−n) = n f (n).) Applying this to
2
the function f (x) = e−πtx , where t is a positive real number, and noting
that
 ∞ 2  ∞ 2
−πtx2 +2πixy e−πy /t −πu2 e−πy /t
f (y) = e dx = √ e du = √
−∞ t −∞ t
√
(substitution u = t (x − iy/t) followed by a shift of the path of integration),
we obtain ∞ ∞
−πn2 t 1 2
e = √ e−πn /t (t > 0) .
n=−∞
t n=−∞

This proves the second equation in (30) for z = it/2 lying on the posi-
tive imaginary axis, and the general case then follows by analytic continu-
ation.
26 D. Zagier

The point is now that the two transformations z → z + 1 and z → −1/4z

generate a subgroup of SL(2, R) which is commensurable with SL(2, Z), so
(30) implies that the function θ(z) is a modular form of weight 1/2. (We
have not defined modular forms of half-integral weight and will not discuss
their theory in these notes, but the reader can simply interpret this statement
as saying that θ(z)2 is a modular form of weight 1.) More specifically, for
every N ∈ N we have the“congruence subgroup” Γ0 (N ) ⊆ Γ1 = SL(2, Z),
consisting of matrices ac db ∈ Γ1 with c divisible by N , and the larger group
 
Γ0+ (N ) = Γ0 (N ), WN  = Γ0 (N ) ∪ Γ0 (N )WN , where WN = √1N N0 −1 0
(“Fricke involution”) is an element of SL(2, R) of order  2 which normalizes
Γ0 (N ). The group Γ0+ (N ) contains the elements T = 10 11 and WN for any
N . In general they generate a subgroup of infinite index, so that to check the
modularity of a given function it does not suffice to verify its behavior just
for z → z + 1 and z → −1/N z, but for N = 4 (like for N = 1 !) they generate
the full group and this is sufficient. The proof is simple. Since WN2 = −1, it is
 
sufficient to show that the two matrices T and T = W4 T W4−1 = 1401 generate
the image of Γ0 (4) in PSL(2, R), i.e., that any element γ = ac db ∈ Γ0 (4) is,
up to sign, a word in T and T . Now a is odd, so |a| = 2|b|. If |a| < 2|b|, then
either b+a or b−a is smaller than b in absolute value, so replacing γ by γ ·T ±1
decreases a2 + b2 . If |a| > 2|b| = 0, then either a + 4b or a − 4b is smaller than
a in absolute value, so replacing γ by γ · T ±1 decreases a2 + b2 . Thus we can
keep multiplying γ on the right by powers of T and T until b = 0, at which
point ±γ is a power of T .
Now, by the principle “a finite number of q-coefficients suffice” formu-
lated at the end of Section 1, the mere fact that θ(z) is a modular form is
already enough to let one prove non-trivial identities. (We enunciated the
principle only in the case of forms of integral weight, but even without know-
ing the details of the theory it is clear that it then also applies to half-
integral weight, since a space of modular forms of half-integral weight can
be mapped injectively into a space of modular forms of the next higher in-
tegral weight by multiplying by θ(z).) And indeed, with almost no effort we
obtain proofs of two of the most famous results of number theory of the
17th and 18th centuries, the theorems of Fermat and Lagrange about sums of
squares.

♠ Sums of Two and Four Squares

 
Let r2 (n) = # (a, b) ∈ Z2 | a2 + b2 = n be the number of repre-
sentations of an integer n ≥ 0 as a sum of two squares. Since θ(z)2 =
 a2
 b2

a∈Z q b∈Z q , we see that r2 (n) is simply the coefficient of q n in
2
θ(z) . From Proposition 9 and the just-proved fact that Γ0 (4) is generated by
−Id2 , T and T , we find that the function θ(z)2 is a “modular form of weight 1
and character χ−4 on Γ0 (4)” in the sense explained in the paragraph preceding
equation (15), where χ−4 is the Dirichlet character modulo 4 defined by (15).
 Elliptic Modular Forms and Their Applications 27

Since the Eisenstein series G1,χ−4 in (16) is also such a modular form, and
since the space of all such forms has dimension at most 1 by Proposition 3
(because Γ0 (4) has index 6 in SL(2, Z) and hence volume 2π), these two func-
tions must be proportional. The proportionality factor is obviously 4, and we
obtain:
Proposition 10. Let n be a positive integer. Then the number of represen-
tations of n as a sum of two squares is 4 times the sum of (−1)(d−1)/2 , where
d runs over the positive odd divisors of n .
Corollary (Theorem of Fermat). Every prime number p ≡ 1 (mod 4) is
a sum of two squares.
 
Proof of Corollary. We have r2 (p) = 4 1 + (−1)(p−1)/4 = 8 = 0.
The same reasoning applies to other powers of θ. In particular, the number
r4 (n) of representations of an integer n as a sum of four squares is the coef-
ficient of q n in the modular form θ(z)4 of weight 2 on Γ0 (4), and the space
of all such modular forms is at most two-dimensional by Proposition 3. To
find a basis for it, we use the functions G2 (z) and G∗2 (z) defined in equa-
tions (17) and (21). We showed in §2.3 that the latter function transforms
with respect to SL(2, Z) like a modular form of weight 2, and it follows easily
that the three functions G∗2 (z), G∗2 (2z) and G∗2 (4z) transform like modular
forms of weight 2 on Γ0 (4) (exercise!). Of course these three functions are not
holomorphic, but since G∗2 (z) differs from the holomorphic function G2 (z) by
1/8πy, we see that the linear combinations G∗2 (z)−2G∗2 (2z) = G2 (z)−2G2(2z)
and G∗2 (2z) − 2G∗2 (4z) = G2 (2z) − 2G2 (4z) are holomorphic, and since they
are also linearly independent, they provide the desired basis for M2 (Γ0 (4)).
Looking at the first two Fourier coefficients of θ(z)4 = 1 + 8q + · · · , we find
that θ(z)4 equals 8 (G2 (z)−2G2(2z))+16 (G2(2z)−2G2(4z)). Now comparing
coefficients of q n gives:
Proposition 11. Let n be a positive integer. Then the number of representa-
tions of n as a sum of four squares is 8 times the sum of the positive divisors
of n which are not multiples of 4.
Corollary (Theorem of Lagrange). Every positive integer is a sum of four
squares. ♥
For another simple application of the q-expansion principle, we introduce
two variants θM (z) and θF (z) (“M” and “F” for “male” and “female” or “minus
sign” and “fermionic”) of the function θ(z) by inserting signs or by shifting the
indices by 1/2 in its definition:
2
θM (z) = (−1)n q n = 1 − 2q + 2q 4 − 2q 9 + · · · ,
n∈Z
2
θF (z) = qn = 2q 1/4 + 2q 9/4 + 2q 25/4 + · · · .
n∈Z+1/2
28 D. Zagier

These are again modular forms of weight 1/2 on Γ0 (4). With a little experi-
mentation, we discover the identity

θ(z)4 = θM (z)4 + θF (z)4 (31)

due to Jacobi, and by the q-expansion principle all we have to do to prove it

is to verify the equality of a finite number of coefficients (here just one). In
this particular example, though, there is also an easy combinatorial proof, left
as an exercise to the reader.
The three theta series θ, θM and θF , in a slightly different guise and
slightly different notation, play a role in many contexts, so we say a little more
about them. As well as the subgroup Γ0 (N ) of Γ1 , one also has the principal
congruence subgroup Γ (N ) = {γ ∈ Γ1 | γ ≡ Id2 (mod N )} for every integer
N ∈ N, which is more basic than Γ0 (N ) because it is a normal subgroup
(the kernel of the map Γ1 → SL(2, Z/N Z) given by reduction modulo N ).
Exceptionally,
  the group Γ0 (4) is isomorphic to Γ (2), simply by conjugation
by 20 01 ∈ GL(2, R), so that there is a bijection between modular forms on
Γ0 (4) of any weight and modular forms on Γ (2) of the same weight given by
f (z) → f (z/2). In particular, our three theta functions correspond to three
new theta functions

θ3 (z) = θ(z/2) , θ4 (z) = θM (z/2) , θ2 (z) = θF (z/2) (32)

on Γ (2), and the relation (31) becomes θ24 + θ44 = θ34 . (Here the index “1” is
 2
missing because the fourth member of the quartet, θ1 (z) = (−1)n q (n+1/2) /2
is identically zero, as one sees by sending n to −n − 1. It may look odd that
one keeps a whole notation for the zero function. But in fact the functions
θi (z) for 1 ≤ i ≤ 4 are just the “Thetanullwerte” or “theta zero-values” of the
 2
two-variable series θi (z, u) = εn q n /2 e2πinu , where the sum is over Z or
Z + 12 and εn is either 1 or (−1)n , none of which vanishes identically. The
functions θi (z, u) play a basic role in the theory of elliptic functions and are
also the simplest example of Jacobi forms, a theory which is related to many
of the themes treated in these notes and is also important in connection with
Siegel modular forms of degree 2 as discussed in Part III of this book.) The
quotient group Γ1 /Γ (2) ∼ = SL(2, Z/2Z), which has order 6 and is isomorphic
to the symmetric group S3 on three symbols, acts as the latter on the modular
forms θi (z)8 , while the fourth powers transform by
⎛ ⎞ ⎛ ⎞
θ2 (z)4 1 0 0
Θ(z) := ⎝− θ3 (z)4 ⎠ ⇒ Θ(z + 1) = − ⎝ 0 0 1 ⎠ Θ(z),
θ4 (z)4 0 1 0
⎛ ⎞
  0 0 1
1
z −2 Θ − = − ⎝ 0 1 0 ⎠ Θ(z) .
z
1 0 0
 Elliptic Modular Forms and Their Applications 29

This illustrates the general principle that a modular form on a subgroup of

finite index of the modular group Γ1 can also be seen as a component of
a vector-valued modular form on Γ1 itself. The full ring M∗ (Γ (2)) of modu-
lar forms on Γ (2) is generated by the three components of Θ(z) (or by any
two of them, since their sum is zero), while the subring M∗ (Γ (2))S3 =
8 8 8
M∗ (Γ14) is generated
 by the modular
 forms θ2 (z)
 + θ3 (z) + θ4 (z) and
θ2 (z) + θ3 (z) θ3 (z) + θ4 (z) θ4 (z) − θ2 (z) of weights 4 and 6 (which
4 4 4 4 4

are then equal to 2E4 (z) and 2E6 (z), respectively). Finally, we see that
1 8 8 8
256 θ2 (z) θ3 (z) θ4 (z) is a cusp form of weight 12 on Γ1 and is, in fact, equal
to Δ(z).
This last identity has an interesting consequence. Since Δ(z) is non-zero
for all z ∈ H, it follows that each of the three theta-functions θi (z) has the
same property. (One can also see this by noting that the “visible” zero of θ2 (z)
at infinity accounts for all the zeros allowed by the formula discussed in §1.3,
so that this function has no zeros at finite points, and then the same holds for
θ3 (z) and θ4 (z) because they are related to θ2 by a modular transformation.)
This suggests that these three functions, or equivalently their Γ0 (4)-versions
θ, θM and θF , might have a product expansion similar to that of the function
Δ(z), and indeed this is the case: we have the three identities

η(2z)5 η(z)2 η(4z)2

θ(z) = , θM (z) = , θF (z) = 2 , (33)
η(z)2 η(4z)2 η(2z) η(2z)
where η(z) is the “Dedekind eta-function”
∞
  
η(z) = Δ(z)1/24 = q 1/24 1 − qn . (34)
n=1

The proof of (33) is immediate by the usual q-expansion principle: we multiply

the identities out (writing, e.g., the first as θ(z)η(z)2 η(4z)2 = η(2z)5 ) and then
verify the equality of enough coefficients to account for all possible zeros of
a modular form of the corresponding weight. More efficiently, we can use our
knowledge of the transformation behavior of Δ(z) and hence of η(z) under Γ1
to see that the quotients on the right in (33) are finite at every cusp and
hence, since they also have no poles in the upper half-plane, are holomorphic
modular forms of weight 1/2, after which the equality with the theta-functions
on the left follows directly.
More generally, one can ask when a quotient of products of eta-functions
is a holomorphic modular form. Since η(z) is non-zero in H, such a quotient
never has finite zeros, and the only issue is whether the numerator vanishes to
at least the same order as the denominator at each cusp. Based on extensive
numerical calculations, I formulated a general conjecture saying that there
are essentially only finitely many such products of any given weight, and
a second explicit conjecture giving the complete list for weight 1/2 (i.e., when
the number of η’s in the numerator is one bigger than in the denominator).
Both conjectures were proved by Gerd Mersmann in a brilliant Master’s thesis.
30 D. Zagier

For the weight 1/2 result, the meaning of “essentially”

 is that the product
should be primitive, i.e., it should have the form η(ni z)ai where the ni
are positive integers with no common factor. (Otherwise one would obtain
infinitely many examples by rescaling, e.g., one would have both θM (z) =
η(z)2 /η(2z) and θM (2z) = η(2z)2 /η(4z) on the list.) The classification is then
as follows:
Theorem (Mersmann). There are precisely 14 primitive eta-products which
are holomorphic modular forms of weight 1/2 :

η(z)2 η(2z)2 η(z) η(4z) η(2z)3 η(2z)5

η(z) , , , , , ,
η(2z) η(z) η(2z) η(z) η(4z) η(z)2 η(4z)2
η(z)2 η(6z) η(2z)2 η(3z) η(2z) η(3z)2 η(z) η(6z)2
, , , ,
η(2z) η(3z) η(z) η(6z) η(z) η(6z) η(2z) η(3z)
η(z) η(4z) η(6z)2 η(2z)2 η(3z) η(12z) η(2z)5 η(3z) η(12z)
, , ,
η(2z) η(3z) η(12z) η(z) η(4z) η(6z) η(z)2 η(4z)2 η(6z)2
η(z) η(4z) η(6z)5
.
η(2z)2 η(3z)2 η(12z)2

Finally, we mention that η(z) itself has the theta-series representation

∞
2
η(z) = χ12 (n) q n /24
= q 1/24 − q 25/24 − q 49/24 + q 121/24 + · · ·
n=1

where χ12 (12m ± 1) = 1, χ12 (12m ± 5) = −1, and χ12 (n) = 0 if n is divisible
by 2 or 3. This identity was discovered numerically by Euler (in the simpler-
∞ ∞ 2
looking but less enlightening version n=1 (1 − q n ) = n=1 (−1)n q (3n +n)/2 )
and proved by him only after several years of effort. From a modern point of
view, his theorem is no longer surprising because one now knows the following
beautiful general result, proved by J-P. Serre and H. Stark in 1976:
Theorem (Serre–Stark). Every modular form of weight 1/2 is a linear
combination of unary theta series.
Explicitly, this means that every modular form of weight 1/2 with respect
to any subgroup of finite index of SL(2, Z) is a linear combination of sums of
 2
the form n∈Z q a(n+c) with a ∈ Q>0 and c ∈ Q. Euler’s formula for η(z) is
a typical case of this, and of course each of the other products given in Mers-
mann’s theorem must also have a representation as a theta series. For instance,
the last function on the list, η(z)η(4z)η(6z)5 /η(2z)2 η(3z)2 η(12z)2 , has the ex-
 2
pansion n>0, (n,6)=1 χ8 (n)q n /24 , where χ8 (n) equals +1 for n ≡ ±1 (mod 8)
and −1 for n ≡ ±3 (mod 8).
We end this subsection by mentioning one more application of the Jacobi
theta series.
 Elliptic Modular Forms and Their Applications 31

♠ The Kac–Wakimoto Conjecture

For any two natural numbers m and n, denote by Δm (n) the number of
representations of n as a sum of m triangular numbers (numbers of the form
a(a − 1)/2 with a integral). Since 8a(a − 1)/2 + 1 = (2a − 1)2 , this can also be
odd
written as the number rm (8n+m) of representations of 8n+m as a sum of m
odd squares. As part of an investigation in the theory of affine superalgebras,
Kac and Wakimoto were led to conjecture the formula

Δ4s2 (n) = Ps (a1 , . . . , as ) (35)

r1 , a1 , ..., rs , as ∈ Nodd
r1 a1 +···+rs as = 2n+s2

for m of the form 4s2 (and a similar formula for m of the form 4s(s + 1) ),
where Nodd = {1, 3, 5, . . . } and Ps is the polynomial
   2
i ai · a2i − a2j
i<j
Ps (a1 , . . . , as ) = 2s−1 .
4s(s−1) s! j=1 j!

Two proofs of this were subsequently given, one by S. Milne using elliptic func-
tions and one by myself using modular forms. Milne’s proof is very ingenious,
with a number of other interesting identities appearing along the way, but is
quite involved. The modular proof is much simpler. One first notes that, Ps
being a homogeneous polynomial of degree 2s2 − s and odd in each argument,
2
the right-hand side of (35) is the coefficient of q 2n+s in a function F (z) which
is a linear combination
 of products gh1 (z) · · · ghs (z) with h1 + · · · + hs = s2 ,
where gh (z) = r, a∈Nodd a2h−1 q ra (h ≥ 1). Since gh is a modular form (Eisen-
stein series) of weight 2h on Γ0 (4), this function F is a modular form of weight
2
2s2 on the same group. Moreover, its Fourier expansion belongs to q s Q[[q 2 ]]
(because Ps (a1 , . . . , as ) vanishes if any two ai are equal, and the smallest
value of r1 a1 + · · · + rs as with all ri and ai in Nodd and all ai distinct is
1 + 3 + · · · + 2s − 1 = s2 ), and from the formula given in §1 for the number of
zeros of a modular form we find that this property characterizes F (z) uniquely
2
in M2s2 (Γ0 (4)) up to a scalar factor. But θF (z)4s has the same property, so
the two functions must be proportional. This proves (35) up to a scalar factor,
easily determined by setting n = 0. ♥

3.2 Theta Series in Many Variables

We now consider quadratic forms in an arbitrary number m of variables. Let

Q : Zm → Z be a positive definite quadratic form which takes integral values
on Zm . We associate to Q the theta series
∞
ΘQ (z) = q Q(x1 ,...,xm ) = RQ (n) q n , (36)
x1 ,...,xm ∈Z n=0
32 D. Zagier

where of course q = e2πiz as usual and RQ (n) ∈ Z≥0 denotes the number of
representations of n by Q, i.e., the number of vectors x ∈ Zm with Q(x) = n.
The basic statement is that ΘQ is always a modular form of weight m/2.
In the case of even m we can be more precise about the modular transfor-
mation behavior, since then we are in the realm of modular forms of in-
tegral weight where we have given complete definitions of what modularity
means. The quadratic form Q(x) is a linear combination of products xi xj with
1 ≤ i, j ≤ m. Since xi xj = xj xi , we can write Q(x) uniquely as
m
1 1
Q(x) = xt Ax = aij xi xj , (37)
2 2 i,j=1

where A = (aij )1≤i,j≤m is a symmetric m × m matrix and the factor 1/2 has
been inserted to avoid counting each term twice. The integrality of Q on Zm
is then equivalent to the statement that the symmetric matrix A has integral
elements and that its diagonal elements aii are even. Such an A is called an
even integral matrix. Since we want Q(x) > 0 for x = 0, the matrix A must
be positive definite. This implies that det A > 0. Hence A is non-singular
and A−1 exists and belongs to Mm (Q). The level of Q is then defined as the
smallest positive integer N = NQ such that N A−1 is again an even integral
matrix. We also have the discriminant Δ = ΔQ of A, defined as (−1)m det A.
It is always congruent to 0 or 1 modulo 4, so there is an associated character
(Kronecker symbol)χΔ , which is the unique Dirichlet character modulo N
Δ
satisyfing χΔ (p) = (Legendre symbol) for any odd prime p  N . (The
p
character χΔ in the special cases Δ = −4, 12 and 8 already occurred in §2.2
(eq. (15)) and §3.1.) The precise description of the modular behavior of ΘQ
for m ∈ 2Z is then:
Theorem (Hecke, Schoenberg). Let Q : Z2k → Z be a positive definite
integer-valued form in 2k variables of level N and discriminant Δ. Then ΘQ
is a modular
 form on Γ0 (N ) of weight k and character
 a b χΔ , i.e., we have
cz+d = χΔ (a) (cz + d) ΘQ (z) for all z ∈ H and c d ∈ Γ0 (N ).
ΘQ az+b k

The proof, as in the unary case, relies essentially on the Poisson sum-
mation formula, which gives the identity ΘQ (−1/N z) = N k/2 (z/i)k ΘQ∗ (z),
where Q∗ (x) is the quadratic form associated to N A−1 , but finding the pre-
cise modular behavior requires quite a lot of work. One can also in principle
reduce the higher rank case to the one-variable case by using the fact that
every quadratic form is diagonalizable over Q, so that the sum in (36) can
be broken up into finitely many sub-sums over sublattices or translated sub-
lattices of Zm on which Q(x1 , . . . , xm ) can be written as a linear combination
of m squares.
There is another language for quadratic forms which is often more conve-
nient, the language of lattices. From this point of view, a quadratic form is no
longer a homogeneous quadratic polynomial in m variables, but a function Q
 Elliptic Modular Forms and Their Applications 33

from a free Z-module Λ of rank m to Z such that the associated scalar prod-
uct (x, y) = Q(x + y) − Q(x) − Q(y) (x, y ∈ Λ) is bilinear. Of course we can
always choose a Z-basis of Λ, in which case Λ is identified with Zm and Q
is described in terms of a symmetric matrix A as in (37), the scalar product
being given by (x, y) = xt Ay, but often the basis-free language is more conve-
nient. In terms of the scalar product, we have a length function x2 = (x, x)
(actually this is the square of the length, but one often says simply “length”
for convenience) and Q(x) = 12 x2 , so that the integer-valued case we are
considering corresponds to lattices in which all vectors have even length. One
often chooses the lattice Λ inside the euclidean space Rm with its standard
length function (x, x) = x2 = x21 + · · · + x2m ; in this case the square root
of det A is equal to the volume of the quotient Rm /Λ, i.e., to the volume of
a fundamental domain for the action by translation of the lattice Λ on Rm . In
the case when this volume is 1, i.e., when Λ ∈ Rm has the same covolume as
Zm , the lattice is called unimodular. Let us look at this case in more detail.

♠ Invariants of Even Unimodular Lattices

If the matrix A in (37) is even and unimodular, then the above theorem tells
us that the theta series ΘQ associated to Q is a modular form on the full
modular group. This has many consequences.

Proposition 12. Let Q : Zm → Z be a positive definite even unimodular

quadratic form in m variables. Then
(i) the rank m is divisible by 8, and
(ii)the number of representations of n ∈ N by Q is given for large n by the
formula

2k  
RQ (n) = − σk−1 (n) + O nk/2 (n → ∞) , (38)
Bk

where m = 2k and Bk denotes the kth Bernoulli number.

Proof. For the first part it is enough to show that m cannot be an odd multiple
of 4, since if m is either odd or twice an odd number then 4m or 2m is an
odd multiple of 4 and we can apply this special case to the quadratic form
Q ⊕ Q ⊕ Q ⊕ Q or Q ⊕ Q, respectively. So we can assume that m = 2k with
k even and must show that k is divisible by 4 and that (38) holds. By the
theorem above, the theta series ΘQ is a modular form of weight k on the full
modular group Γ1 = SL(2, Z) (necessarily with trivial character, since there
are no non-trivial Dirichlet characters modulo 1). By the results of Section 2,
this modular form is a linear combination of Gk (z) and a cusp form of weight k,
and from the Fourier expansion (13) we see that the coefficient of Gk in this
decomposition equals −2k/Bk , since the constant term RQ (0) of ΘQ equals 1.
(The only vector of length 0 is the zero vector.) Now Proposition 8 implies the
34 D. Zagier

asymptotic formula (38), and the fact that k must be divisible by 4 also follows
because if k ≡ 2 (mod 4) then Bk is positive and therefore the right-hand side
of (38) tends to −∞ as k → ∞, contradicting RQ (n) ≥ 0.

The first statement of Proposition 12 is purely algebraic, and purely alge-

braic proofs are known, but they are not as simple or as elegant as the modu-
lar proof just given. No non-modular proof of the asymptotic formula (38) is
known.
Before continuing with the theory, we look at some examples, starting in
rank 8. Define the lattice Λ8 ⊂ R8 to be the set of vectors belonging to either
Z8 or (Z+ 12 )8 for which the sum of the coordinates is even. This is unimodular
because the lattice Z8 ∪ (Z + 12 )8 contains both it and Z8 with the same
index 2, and is even because x2i ≡ xi (mod 2) for xi ∈ Z and x2i ≡ 14 (mod 2)
for xi ∈ Z + 12 . The lattice Λ8 is sometimes denoted E8 because, if we choose
the Z-basis ui = ei − ei+1 (1 ≤ i ≤ 6), u7 = e6 + e7 , u8 = − 21 (e1 + · · · + e8 ) of
Λ8 , then every ui has length 2 and (ui , uj ) for i = j equals −1 or 0 according
whether the ith and jth vertices (in a standard numbering) of the “E8 ” Dynkin
diagram in the theory of Lie algebras are adjacent or not. The theta series of
Λ8 is a modular form of weight 4 on SL(2, Z) whose Fourier expansion begins
with 1, so it is necessarily equal to E4 (z), and we get “for free” the information
that for every integer n ≥ 1 there are exactly 240 σ3 (n) vectors x in the E8
lattice with (x, x) = 2n.
From the uniqueness of the modular form E4 ∈ M4 (Γ1 ) we in fact get that
rQ (n) = 240σ3 (n) for any even unimodular quadratic form or lattice of rank 8,
but here this is not so interesting because the known classification in this rank
says that Λ8 is, in fact, the only such lattice up to isomorphism. However, in
rank 16 one knows that there are two non-equivalent lattices: the direct sum
Λ8 ⊕ Λ8 and a second lattice Λ16 which is not decomposable. Since the theta
series of both lattices are modular forms of weight 8 on the full modular group
with Fourier expansions beginning with 1, they are both equal to the Eisen-
stein series E8 (z), so we have rΛ8 ⊕Λ8 (n) = rΛ16 (n) = 480 σ7 (n) for all n ≥ 1,
even though the two lattices in question are distinct. (Their distinctness, and
a great deal of further information about the relative positions of vectors of
various lengths in these or in any other lattices, can be obtained by using the
theory of Jacobi forms which was mentioned briefly in §3.1 rather than just
the theory of modular forms.)
In rank 24, things become more interesting, because now dim M12 (Γ1 ) = 2
and we no longer have uniqueness. The even unimodular lattices of this rank
were classified completely by Niemeyer in 1973. There are exactly 24 of them
up to isomorphism. Some of them have the same theta series and hence the
same number of vectors of any given length (an obvious such pair of lattices
being Λ8 ⊕Λ8 ⊕Λ8 and Λ8 ⊕Λ16 ), but not all of them do. In particular, exactly
one of the 24 lattices has the property that it has no vectors of length 2.
This is the famous Leech lattice (famous among other reasons because it has
a huge group of automorphisms, closely related to the monster group and
 Elliptic Modular Forms and Their Applications 35

other sporadic simple groups). Its theta series is the unique modular form of
weight 12 on Γ1 with Fourier expansion starting 1 + 0q + · · · , so it must equal
E12 (z) − 21736
691 Δ(z), i.e., the
 number rLeech  (n) of vectors of length 2n in the
Leech lattice equals 21736
691 σ 11 (n) − τ (n) for every positive integer n. This
gives another proof and an interpretation of Ramanujan’s congruence (28).
In rank 32, things become even more interesting: here the complete clas-
sification is not known, and we know that we cannot expect it very soon,
because there are more than 80 million isomorphism classes! This, too, is
a consequence of the theory of modular forms, but of a much more sophisti-
cated part than we are presenting here. Specifically, there is a fundamental
theorem of Siegel saying that the average value of the theta series associated
to the quadratic forms in a single genus (we omit the definition) is always an
Eisenstein series. Specialized to the particular case of even unimodular forms
of rank m = 2k ≡ 0 (mod 8), which form a single genus, this theorem says
that there are only finitely many such forms up to equivalence for each k and
that, if we number them Q1 , . . . , QI , then we have the relation
I
1
ΘQi (z) = mk Ek (z) , (39)
i=1
wi

where wi is the number of automorphisms of the form Qi (i.e., the number of

matrices γ ∈ SL(m, Z) such that Qi (γx) = Qi (x) for all x ∈ Zm ) and mk is
the positive rational number given by the formula
Bk B2 B4 B2k−2
mk = ··· ,
2k 4 8 4k − 4
where Bi denotes the ith Bernoulli number. In particular, by comparing
the constant terms on the left- and right-hand sides of (39), we see that
I
i=1 1/wi = mk , the Minkowski-Siegel mass formula. The numbers m4 ≈
1.44 × 10−9 , m8 ≈ 2.49 × 10−18 and m12 ≈ 7, 94 × 10−15 are small, but
m16 ≈ 4, 03 × 107 (the next two values are m20 ≈ 4.39 × 1051 and m24 ≈
1.53 × 10121 ), and since wi ≥ 2 for every i (one has at the very least the
automorphisms ± Idm ), this shows that I > 80000000 for m = 32 as as-
serted.
A further consequence of the fact that ΘQ ∈ Mk (Γ1 ) for Q even and
unimodular of rank m = 2k is that the minimal value of Q(x) for non-zero
x ∈ Λ is bounded by r = dim Mk (Γ1 ) = [k/12] + 1. The lattice L is called
extremal if this bound is attained. The three lattices of rank 8 and 16 are
extremal for trivial reasons. (Here r = 1.) For m = 24 we have r = 2 and the
only extremal lattice is the Leech lattice. Extremal unimodular lattices are
also known to exist for m = 32, 40, 48, 56, 64 and 80, while the case m = 72
is open. Surprisingly, however, there are no examples of large rank:
Theorem (Mallows–Odlyzko–Sloane). There are only finitely many non-
isomorphic extremal even unimodular lattices.
36 D. Zagier

We sketch the proof, which, not surprisingly, is completely modular. Since

there are only finitely many non-isomorphic even unimodular lattices of any
given rank, the theorem is equivalent to saying that there is an absolute bound
on the value of the rank m for extremal lattices. For simplicity, let us suppose
that m = 24n. (The cases m = 24n + 8 and m = 24n + 16 are similar.) The
theta series of any extremal unimodular lattice of this rank must be equal to
the unique modular form fn ∈ M12n (SL(2, Z)) whose q-development has the
form 1 + O(q n+1 ). By an elementary argument which we omit but which the
reader may want to look for, we find that this q-development has the form
 
nbn
fn (z) = 1 + n an q n+1 + − 24 n (n + 31) an q n+2 + · · ·
2

where an and bn are the coefficients of Δ(z)n in the modular functions j(z)
and j(z)2 , respectively, when these are expressed (locally, for small q) as Lau-
rent series in the modular form Δ(z) = q − 24q 2 + 252q 3 − · · · . It is not
hard to show that an has the asymptotic behavior an ∼ An−3/2 C n for some
constants A = 225153.793389 · · · and C = 1/Δ(z0 ) = 69.1164201716 · · ·,
where z0 = 0.52352170017992 · · ·i is the unique zero on the imaginary axis
of the function E2 (z) defined in (17) (this is because E2 (z) is the logarithmic
derivative of Δ(z)), while bn has a similar expansion but with A replaced by
2λA with λ = j(z0 ) − 720 = 163067.793145 · · ·. It follows that the coefficient
2 nbn − 24n(n + 31)an of q
1 n+2
in fn is negative for n larger than roughly 6800,
corresponding to m ≈ 163000, and that therefore extremal lattices of rank
larger than this cannot exist. ♥

♠ Drums Whose Shape One Cannot Hear

Marc Kac asked a famous question, “Can one hear the shape of a drum?” Ex-
pressed more mathematically, this means: can there be two riemannian mani-
folds (in the case of real “drums” these would presumably be two-dimensional
manifolds with boundary) which are not isometric but have the same spectra
of eigenvalues of their Laplace operators? The first example of such a pair
of manifolds to be found was given by Milnor, and involved 16-dimensional
closed “drums.” More drum-like examples consisting of domains in R2 with
polygonal boundary are now also known, but they are difficult to construct,
whereas Milnor’s example is very easy. It goes as follows. As we already men-
tioned, there are two non-isomorphic even unimodular lattices Λ1 = Γ8 ⊕ Γ8
and Λ2 = Γ16 in dimension 16. The fact that they are non-isomorphic means
that the two Riemannian manifolds M1 = R16 /Λ1 and M2 = R16 /Λ2 , which
are topologically both just tori (S 1 )16 , are not isometric to each other. But
the spectrum of the Laplace operator on any torus Rn /Λ is just the set of
norms λ2 (λ ∈ Λ), counted with multiplicities, and these spectra agree for
 2  2
M1 and M2 because the theta series λ∈Λ1 q λ and λ∈Λ2 q λ coincide.
♥
 Elliptic Modular Forms and Their Applications 37

We should not leave this section without mentioning at least briefly that
there is an important generalization of the theta series (36) in which each term
q Q(x1 ,...,xm ) is weighted by a polynomial P (x1 , . . . , xm ). If this polynomial
is homogeneous of degree d and is spherical with respect to Q (this means
that ΔP = 0, where Δ is the Laplace operator with respect to a system of
coordinates in which . . . , xm ) is simply x21 + · · · + x2m ), then the theta
 Q(x1 ,Q(x)
series ΘQ,P (z) = x P (x)q is a modular form of weight m/2 + d (on the
same group and with respect to the same character as in the case P = 1),
and is a cusp form if d is strictly positive. The possibility of putting non-
trivial weights into theta series in this way considerably enlarges their range
of applications, both in coding theory and elsewhere.

4 Hecke Eigenforms and L-series

In this section we give a brief sketch of Hecke’s fundamental discoveries that
the space of modular forms is spanned by modular forms with multiplicative
Fourier coefficients and that one can associate to these forms Dirichlet series
which have Euler products and functional equations. These facts are at the
basis of most of the higher developments of the theory: the relations of modular
forms to arithmetic algebraic geometry and to the theory of motives, and the
adelic theory of automorphic forms. The last two subsections describe some
basic examples of these higher connections.

4.1 Hecke Theory

For each integer m ≥ 1 there is a linear operator Tm , the mth Hecke operator,
acting on modular forms of any given weight k. In terms of the description
of modular forms as homogeneous functions on lattices which was given in
§1.1, the definition of Tm is very simple: it sends a homogeneous function F
of degree −k on lattices Λ ⊂ C to the function
 Tm F defined (up to a suitable
normalizing constant) by Tm F (Λ) = F (Λ ), where the sum runs over all
sublattices Λ ⊂ Λ of index m. The sum is finite and obviously still homoge-
neous in Λ of the same degree −k. Translating from the language of lattices to
that of functions in the upper half-plane by the usual formula f (z) = F (Λz ),
we find that the action of Tm is given by
 
az + b
Tm f (z) = mk−1 (cz + d)−k f (z ∈ H) , (40)
  cz + d
a b ∈Γ1 \Mm
c d

where Mm denotes the set of 2 × 2 integral matrices of determinant m and

where the normalizing constant mk−1 has been introduced for later conve-
nience (Tm normalized in this way will send forms with integral Fourier co-
efficients to forms with integral Fourier coefficients). The sum makes sense
38 D. Zagier

because the transformation

  law (2) of f implies that the summand associated
to a matrix M = ac db ∈ Mm is indeed unchanged if M is replaced by γM
with γ ∈ Γ1 , and from (40) one also easily sees that Tm f is holomorphic
in H and satisfies the same transformation law and growth properties as f ,
so Tm indeed maps Mk (Γ1 ) to Mk (Γ1 ). Finally, to calculate the effect of Tm
on Fourier developments, we note that a set of representatives of Γ1 \Mm is
given by the upper triangular matrices a0 db with ad = m and 0 ≤ b < d
(this is an easy exercise), so
 
k−1 1 az + b
Tm f (z) = m f . (41)
dk d
ad=m b (mod d)
a, d>0

If f (z) has the Fourier development (3), then a further calculation with (41),
again left to the reader, shows that the function Tm f (z) has the Fourier ex-
pansion
 
k−1 mn/d2 k−1
Tm f (z) = (m/d) an q = r amn/r2 q n .
d|m n≥0 n≥0 r|(m,n)
d>0 d|n r>0
(42)
An easy but important consequence of this formula is that the operators Tm
(m ∈ N) all commute.
Let us consider some examples. The expansion (42) begins σk−1 (m)a0 +
am q + · · · , so if f is a cusp form (i.e., a0 = 0), then so is Tm f . In particular,
since the space S12 (Γ1 ) of cusp forms of weight 12 is 1-dimensional, spanned
by Δ(z), it follows that Tm Δ is a multiple of Δ for every m ≥ 1. Since the
Fourier expansion of Δ begins q + · · · and that of Tm Δ begins τ (m)q + · · · ,
the eigenvalue is necessarily τ (m), so Tm Δ = τ (m)Δ and (42) gives
 mn
τ (m) τ (n) = r11 τ for all m, n ≥ 1 ,
r2
r|(m,n)

proving Ramanujan’s multiplicativity observations mentioned in §2.4. By the

same argument, if f ∈ Mk (Γ1 ) is any simultaneous eigenfunction of all of
the Tm , with eigenvalues λm , then am = λm a1 for all m. We therefore have
a1 = 0 if f is not identically 0, and if we normalize f by a1 = 1 (such an
f is called a normalized Hecke eigenform, or Hecke form for short) then we
have

T m f = am f , am a n = rk−1 amn/r2 (m, n ≥ 1) . (43)

r|(m,n)

Examples of this besides Δ(z) are the unique normalized cusp forms f (z) =
Δ(z)Ek−12 (z) in the five further weights where dim Sk (Γ1 ) = 1 (viz. k = 16,
18, 20, 22 and 26) and the function Gk (z)for all k ≥ 4, for which we have
Tm Gk = σk−1 (m)Gk , σk−1 (m)σk−1 (n) = r|(m,n) rk−1 σk−1 (mn/r2 ). (This
 Elliptic Modular Forms and Their Applications 39

was the reason for the normalization of Gk chosen in §2.2.) In fact, a theorem
of Hecke asserts that Mk (Γ1 ) has a basis of normalized simultaneous eigen-
forms for all k, and that this basis is unique. We omit the proof, though it
is not difficult (one introduces a scalar product on the space of cusp forms
of weight k, shows that the Tm are self-adjoint with respect to this scalar
product, and appeals to a general result of linear algebra saying that com-
muting self-adjoint operators can always be simultaneously diagonalized), and
content ourselves instead with one further example, also due to Hecke. Con-
sider k = 24, the first weight where dim Sk (Γ1 ) is greater than 1. Here Sk is
2-dimensional, spanned by ΔE43 = q + 696q 2 + · · · and Δ2 = q 2 − 48q 3 + . . . .
Computing the first two Fourier expansions of the images under T2 of these
3 3 2
two functions by (42), we find that T2 (ΔE 4 ) = 696 ΔE
 696  4 + 20736000 Δ
2 3 2 20736000
and T2 (Δ ) = ΔE4 + 384 Δ . The matrix 1 has distinct eigen-
√ 384 √
values λ1 = 540 + 12 144169 and λ2 = 540 − 12 144169, so there are
precisely two normalized eigenfunctions
√ of T2 in S24 (Γ1 ), namely the func-
tions f1 = ΔE4√ 3
− (156 − 12 144169)Δ2 = q + λ1 q 2 + · · · and f2 =
ΔE43 − (156 + 12 144169)Δ2 = q + λ2 q 2 + · · · , with T2 fi = λi fi for i = 1, 2.
The uniqueness of these eigenfunctions and the fact that Tm commutes with
T2 for all m ≥ 1 then implies that Tm fi is a multiple of fi for all m ≥ 1, so G24 ,
f1 and f2 give the desired unique basis of M24 (Γ1 ) consisting of normalized
Hecke eigenforms.
Finally, we mention without giving any details that Hecke’s theory gen-
 a b  to congruence groups of SL(2, Z) like the group Γ0 (N ) of matrices
eralizes
c d ∈ Γ1 with c ≡ 0 (mod N ), the main differences being that the defi-
nition of Tm must be modified somewhat if m and N are not coprime and
that the statement about the existence of a unique base of Hecke forms be-
comes more complicated: the space Mk (Γ0 (N )) is the direct sum of the space
spanned by all functions f (dz) where f ∈ Mk (Γ0 (N  )) for some proper di-
visor N  of N and d divides N/N  (the so-called “old forms”) and a space
of “new forms” which is again uniquely spanned by normalized eigenforms of
all Hecke operators Tm with (m, N ) = 1. The details can be found in any
standard textbook.

4.2 L-series of Eigenforms

Let us return to the full modular group. We have seen that M k (Γ1 )mcontains,
and is in fact spanned by, normalized Hecke eigenforms f = am q satisfy-
ing (43). Specializing this equation to the two cases when m and n are coprime
and when m = pν and n = p for some prime p gives the two equations (which
together are equivalent to (43))
amn = am an if (m, n) = 1 , apν+1 = ap apν − pk−1 apν−1 (p prime, ν ≥ 1) .
The first says that the coefficients an are multiplicative and hence that the
∞ a
n
Dirichlet series L(f, s) = s
, called the Hecke L-series of f , has an Eu-
n=1 n
40 D. Zagier
  ap ap2 
ler product L(f, s) = 1 + s + 2s + · · · , and the second tells us
p p
p prime
∞
that the power series ν=0 apν xν for p prime equals 1/(1 − ap x + pk−1 x2 ).
Combining these two statements gives Hecke’s fundamental Euler product
development
 1
L(f, s) = −s
(44)
1 − ap p + pk−1−2s
p prime

for the L-series of a normalized Hecke eigenform f ∈ Mk (Γ1 ), a simple ex-

ample being given by
 1
L(Gk , s) = = ζ(s) ζ(s − k + 1) .
p
1 − (pk−1 + 1)p−s + pk−1−2s

For eigenforms on Γ0 (N ) there is a similar result except that the Euler factors
for p|N have to be modified suitably.
The L-series have another fundamental property, also discovered by Hecke,
which is that they can be analytically continued in s and then satisfy func-
tional equations. We again restrict to Γ = Γ1 and also, for convenience, to
cusp forms, though not any more just to eigenforms. (The method of proof ex-
tends to non-cusp forms but is messier there since L(f, s) then has poles, and
since Mk is spanned by cusp forms and by Gk , whose L-series is completely
known, there is no loss in making the latter restriction.) From the estimate
an = O(nk/2 ) proved in §2.4 we know that L(f, s) converges absolutely in the
half-plane (s) > 1 + k/2. Take s in that half-plane and consider the Euler
gamma function  ∞
Γ (s) = ts−1 e−t dt .
0
∞
Replacing t by λt in this integral gives Γ (s) = λs 0 ts−1 e−λt dt or λ−s =
 ∞
Γ (s)−1 0 ts−1 e−λt dt for any λ > 0. Applying this to λ = 2πn, multiplying
by an , and summing over n, we obtain
∞  ∞  ∞
(2π)−s Γ (s) L(f, s) = an ts−1 e−2πnt dt = ts−1 f (it) dt
n=1 0 0
 
k
(s) > + 1 ,
2

where the interchange of integration and summation is justified by the abso-

lute convergence. Now the fact that f (it) is exponentially small for t → ∞
(because f is a cusp form) and for t → 0 (because f (−1/z) = z k f (z) ) im-
plies that the integral converges absolutely for all s ∈ C and hence that the
function
∞
an
L∗ (f, s) := (2π)−s Γ (s) L(f, s) = (2π)−s Γ (s) (45)
n=1
ns
 Elliptic Modular Forms and Their Applications 41

extends holomorphically from the half-plane (s) > 1 + k/2 to the entire
complex plane. The substitution t → 1/t together with the transformation
equation f (i/t) = (it)k f (it) of f then gives the functional equation

L∗ (f, k − s) = (−1)k/2 L∗ (f, s) (46)

of L∗ (f, s). We have proved:

∞ n
Proposition 13. Let f = n=1 an q be a cusp form of weight k on the
full modular group. Then the L-series L(f, s) extends to an entire function
of s and satisfies the functional equation (46), where L∗ (f, s) is defined by
equation (45).
It is perhaps worth mentioning that, as Hecke also proved, the converse of
Proposition 13 holds as well: if an (n ≥ 1) are complex numbers of polynomial
growth and the function L∗ (f, s) defined by (45) continues analytically to the
whole complex plane and satisfies the functional equation (46), then f (z) =
 ∞ 2πinz
n=1 an e is a cusp form of weight k on Γ1 .

4.3 Modular Forms and Algebraic Number Theory

In §3, we used the theta series θ(z)2 to determine the number of represen-
tations of any integer n as a sum of two squares. More generally, we can
study the number r(Q, n) of representations of n by a positive definite binary
quadratic Q(x, y) = ax2 + bxy + cy 2 with integer coefficients
∞ by considering
Q(x,y) n
the weight 1 theta series ΘQ (z) = x,y∈Z q = n=0 r(Q, n)q . This
theta series depends only on the class [Q] of Q up to equivalences Q ∼ Q ◦ γ
with γ ∈ Γ1 . We showed in §1.2 that for any D < 0 the number h(D) of Γ1 -
equivalence classes [Q] of binary quadratic forms of discriminant b2 − 4ac = D
is finite. If D is a fundamental discriminant (i.e., not representable as D r2
with D congruent to 0 or 1 mod 4 and r > 1), √then h(D) equals the class
number of the imaginary quadratic field K = Q( D) of discriminant D and
there is a well-known bijection between the Γ1 -equivalence classes of binary
quadratic forms of discriminant D and the ideal classes of K such that r(Q, n)
for any form Q equals w times the number r(A, n) of integral ideals a of K
of norm n belonging to the corresponding ideal class A, where w is the num-
ber of roots of unity in K (= 6 or 4 if D = −3 or D = −4 and 2 oth-
erwise). The L-series L(ΘQ , s) of ΘQ is therefore w times the “partial zeta-
function” ζK,A (s) = a∈A N (a)−s . The ideal classes of K form an abelian
group. If χ  is a homomorphism from this group to C∗ , then the L-series
s
= a χ(a)/N (a) (sum over all integral ideals of K) can be writ-
LK (s, χ)
ten as A χ(A) ζK,A (s) (sum over all ideal classes  of K) and hence is the
L-series of the weight 1 modular form fχ (z) = w−1 A χ(A)ΘA (z). On the
other hand, from the unique prime decomposition of ideals in K it follows
that LK (s, χ) has an Euler product. Hence fχ is a Hecke eigenform. If χ = χ0
is the trivial character, then LK (s, χ) = ζK (s), the Dedekind zeta function
42 D. Zagier

of K, which factors as ζ(s)L(s, εD ), the product of theRiemann

 zeta-function
D
and the Dirichlet L-series of the character εD (n) = (Kronecker sym-
 n
bol). Therefore in this case we get [Q] r(Q, n) = w d|n εD (d) (an identity
known to Gauss) and correspondingly
∞   
1 h(D) D
fχ0 (z) = ΘQ (z) = + qn ,
w 2 n=1
d
[Q] d|n

an Eisenstein series of weight 1. If the character χ has order 2, then it

is a so-called genus character and one knows that LK (s, χ) factors as
L(s, εD1 )L(s, εD2 ) where D1 and D2 are two other discriminants with prod-
uct D. In this case, too, fχ (z) is an Eisenstein series. But in all other cases, fχ
is a cusp form and the theory of modular forms gives us non-trivial information
about representations of numbers by quadratic forms.

♠ Binary Quadratic Forms of Discriminant −23

We discuss an explicit example, taken from a short and pretty article written
by van der Blij in 1952. The class number of the discriminant D = −23
is 3, with the SL(2, Z)-equivalence classes of binary quadratic forms of this
discriminant being represented by the three forms

Q0 (x, y) = x2 + xy + 6y 2 ,
Q1 (x, y) = 2x2 + xy + 3y 2 ,
Q2 (x, y) = 2x2 − xy + 3y 2 .

Since Q1 and Q2 represent the same integers, we get only two distinct theta
series

ΘQ0 (z) = 1 + 2q + 2q 4 + 4q 6 + 4q 8 + · · · ,
ΘQ1 (z) = 1 + 2q 2 + 2q 3 + 2q 4 + 2q 6 + · · · .

The linear combination corresponding to the trivial character is the Eisenstein

series
∞   
1  3 −23
fχ0 = ΘQ0 + 2ΘQ1 = + qn
2 2 n=1
d
d|n
3
= + q + 2q 2 + 2q 3 + 3q 4 + · · · ,
2
 
in accordance with the general identity w−1 [Q] r(Q, n) = d|n εD (d) men-
tioned above. If χ√is one of the two non-trivial characters, with values
e±2πi/3 = 12 (−1 ± i 3) on Q1 and Q2 , we have
 Elliptic Modular Forms and Their Applications 43

1 
fχ = ΘQ0 − ΘQ1 = q − q 2 − q 3 + q 6 + · · · .
2
This is a Hecke eigenform in the space S1 (Γ0 (23), ε−23 ). Its L-series has the
form  1
L(fχ , s) = −s + ε −2s
p
1 − a p p −23 (p) p

where ε−23 (p) equals the Legendre symbol (p/23) by quadratic reciprocity
and
⎧
⎪
⎪ 1 if p = 23 ,
⎪
⎨ 0 if (p/23) = −1 ,
ap = (47)
⎪
⎪ 2 if (p/23) = 1 and p is representable as x2 + xy + 6y 2 ,
⎪
⎩
−1 if (p/23) = 1 and p is representable as 2x2 + xy + 3y 2 .
On the other hand, the space S1 (Γ0 (23), ε−23 ) is one-dimensional, spanned by
the function ∞
   
η(z) η(23z) = q 1 − q n 1 − q 23n . (48)
n=1
We therefore obtain the explicit “reciprocity law”
Proposition 14 (van der Blij). Let p be a prime. Then the number ap
defined in (47) is equal to the coefficient of q p in the product (48).
As an application of this, we observe that the q-expansion
 on the right-
hand side of (48) is congruent modulo 23 to Δ(z) = q (1 − q n )24 and hence
that τ (p) is congruent modulo 23 to the number ap defined in (47) for every
prime number p, a congruence for the Ramanujan function τ (n) of a somewhat
different type than those already given in (27) and (28). ♥
Proposition 14 gives a concrete example showing how the coefficients of
a modular form – here η(z)η(23z) – can answer a question of number theory√ –
here, the question whether a given prime number which splits in Q( −23)
splits into principal or non-principal ideals. But actually the connection goes
much deeper. By elementary algebraic number theory we have that the L-
series L(s) = LK (s, χ) is the quotient ζF (s)/ζ(s) of the Dedekind function
of F by the Riemann zeta function, where F = Q(α) (α3 − α − 1 = 0) is the
cubic field of discriminant −23. (The composite K · F is the Hilbert class field
of K.) Hence the four cases in (47) also describe the splitting of p in F : 23 is
ramified, quadratic non-residues of 23 split as p = p1 p2 with N (pi ) = pi , and
quadratic residues of 23 are either split completely (as products of three prime
ideals of norm p) or are inert (remain prime) in F , according whether they
are represented by Q0 or Q1 . Thus the modular form η(z)η(23z) describes
not only the algebraic number theory of the quadratic field K, but also the
splitting of primes in the higher degree field F . This is the first non-trivial
example of the connection found by Weil–Langlands and Deligne–Serre which
relates modular forms of weight one to the arithmetic of number fields whose
Galois groups admit non-trivial two-dimensional representations.
44 D. Zagier

4.4 Modular Forms Associated to Elliptic Curves

and Other Varieties

If X is a smooth projective algebraic variety defined over Q, then for al-

most all primes p the equations defining X can be reduced modulo p to
define a smooth variety Xp over the field Z/pZ = Fp . We can then count
the number of points in Xp over the finite field Fpr for all r ≥ 1 and,
putting
∞all this information together, define a “local zeta function” Z(Xp , s) =
−rs
exp r=1 |Xp (Fpr )| p /r ( (s)  0) and a “global zeta function” (Hasse–
Weil zeta function) Z(X/Q, s) = p Z(Xp , s), where the product is over all
primes. (The factors Zp (X, s) for the “bad” primes p, where the equations
defining X yield a singular variety over Fp , are defined in a more complicated
but completely explicit way and are again power series in p−s .) Thanks to
the work of Weil, Grothendieck, Dwork, Deligne and others, a great deal is
known about the local zeta functions – in particular, that they are rational
functions of p−s and have all of their zeros and poles on the vertical lines
(s) = 0, 12 , . . . , n − 12 , n where n is the dimension of X – but the global
zeta function remains mysterious. In particular, the general conjecture that
Z(X/Q, s) can be meromorphically continued to all s is known only for very
special classes of varieties.
In the case where X = E is an elliptic curve, given, say, by a Weierstrass
equation

y 2 = x3 + Ax + B (A, B ∈ Z, Δ := −4A3 − 27B 2 = 0) , (49)

the local factors can be made completely explicit and we find that Z(E/Q, s) =
ζ(s)ζ(s − 1)
where the L-series L(E/Q, s) is given for (s)  0 by an Euler
L(E/Q, s)
product of the form
 1  1
L(E/Q, s) = ·
1−ap (E) p−s +p1−2s (polynomial of degree ≤ 2 in p−s )
pΔ p|Δ

 (50)

with ap (E) defined for p  Δ as p −  (x, y) ∈ (Z/pZ)2 | y 2 = x3 + Ax + B  .
In the mid-1950’s, Taniyama noticed the striking formal similarity between
this Euler product expansion and the one in (44) when k = 2 and asked
whether there might be cases of overlap between the two, i.e., cases where
the L-function of an elliptic curve agrees with that of a Hecke eigenform of
weight 2 having eigenvalues ap ∈ Z (a necessary condition if they are to agree
with the integers ap (E)).
Numerical examples show that this at least sometimes happens. The sim-
plest elliptic curve (if we order elliptic curves by their “conductor,” an invariant
in N which is divisible only by primes dividing the discriminant Δ in (49))
is the curve Y 2 − Y = X 3 − X 2 of conductor 11. (This can be put into the
form (49) by setting y = 216Y − 108, x = 36X − 12, giving A = −432,
 Elliptic Modular Forms and Their Applications 45

B = 8208, Δ = −28 · 312 · 11, but the equation in X and Y , the so-called “min-
imal model,” has much smaller coefficients.) We can compute the numbers ap
by counting solutions of Y 2 − Y = X 3 − X 2 in (Z/pZ)2 . (For p > 3 this is
equivalent to the recipe given above because the equations relating (x, y) and
(X, Y ) are invertible in characteristic p, and for p = 2 or 3 the minimal model
gives the correct answer.) For example, we have a5 = 5 − 4 = 1 because the
equation Y 2 − Y = X 3 − X 2 has the 4 solutions (0,0), (0,1), (1,0) and (1,1)
in (Z/5Z)2 . Then we have
 −1  −1  −1
2 2 1 3 1 5
L(E/Q, s) = 1 + s + 2s 1 + s + 2s 1 − s + 2s ···
2 2 3 3 5 5
1 2 1 2 1
= s − s − s + s + s + · · · = L(f, s) ,
1 2 3 4 5
where f ∈ S2 (Γ0 (11)) is the modular form
∞
  2  2
f (z) = η(z)2 η(11z)2 = q 1−q n 1−q 11n = x−2q 2 −q 3 +2q 4 +q 5 + · · · .
n=1

In the 1960’s, one direction of the connection suggested by Taniyama was

proved by M. Eichler and G. Shimura, whose work establishes the following
theorem.
Theorem (Eichler–Shimura). Let f (z) be a Hecke eigenform in S2 (Γ0 (N ))
for some N ∈ N with integral Fourier coefficients. Then there exists an elliptic
curve E/Q such that L(E/Q, s) = L(f, s).
Explicitly, this means that ap (E) = ap (f ) for all primes p, where an (f ) is
the coefficient of q n in the Fourier expansion of f . The proof of the theorem
is in a sense quite explicit. The quotient of the upper half-plane by Γ0 (N ),
compactified appropriately by adding a finite number of points (cusps), is
a complex curve (Riemann surface), traditionally denoted by X0 (N ), such
that the space of holomorphic 1-forms on X0 (N ) can be identified canoni-
cally (via f (z) → f (z)dz) with the space of cusp forms S2 (Γ0 (N )). One can
also associate to X0 (N ) an abelian variety, called its Jacobian, whose tangent
space at any point can be identified canonically with S2 (Γ0 (N )). The Hecke
operators Tp introduced in 4.1 act not only on S2 (Γ0 (N )), but on the Jaco-
bian variety itself, and if the Fourier coefficients ap = ap (f ) are in Z, then so
do the differences Tp − ap · Id . The subvariety of the Jacobian annihilated by
all of these differences (i.e., the set of points x in the Jacobian whose image
under Tp equals ap times x; this makes sense because an abelian variety has
the structure of a group, so that we can multiply points by integers) is then
precisely the sought-for elliptic curve E. Moreover, this construction shows
that we have an even more intimate relationship between the curve E and the
form f than the L-series equality L(E/Q, s) = L(f, s), namely, that there is
an actual map from the modular curve X0 (N ) to the elliptic curve E which
46 D. Zagier


∞ a (f )
n
is induced by f . Specifically, if we define φ(z) = e2πinz , so that
n=1 n
φ (z) = 2πif (z)dz, then the fact that f is modular of weight 2 implies that
the difference φ(γ(z)) − φ(z) has zero derivative and hence is constant for all
γ ∈ Γ0 (N ), say φ(γ(z)) − φ(z) = C(γ). It is then easy to see that the map
C : Γ0 (N ) → C is a homomorphism, and in our case (f an eigenform, eigenval-
ues in Z), it turns out that its image is a lattice Λ ⊂ C, and the quotient map
C/Λ is isomorphic to the elliptic curve E. The fact that φ(γ(z)) − φ(z) ∈ Λ
φ pr
then implies that the composite map H −→ C −→ C/Λ factors through the
projection H → Γ0 (N )\H, i.e., φ induces a map (over C) from X0 (N ) to E.
This map is in fact defined over Q, i.e., there are modular functions X(z) and
Y (z) with rational Fourier coefficients which are invariant under Γ0 (N ) and
which identically satisfy the equation Y (z)2 = X(z)3 +AX(z)+B (so that the
map from X0 (N ) to E in its Weierstrass form is simply z → (X(z), Y (z) ) )
as well as the equation X  (z)/2Y (z) = 2πif (z). (Here we are simplifying
a little.)
Gradually the idea arose that perhaps the answer to Taniyama’s orig-
inal question might be yes in all cases, not just sometimes. The results
of Eichler and Shimura showed this in one direction, and strong evidence
in the other direction was provided by a theorem proved by A. Weil in
1967 which said that if the L-series of an elliptic curve E/Q and certain
“twists” of it satisfied the conjectured analytic properties (holomorphic con-
tinuation and functional equation), then E really did correspond to a modu-
lar form in the above way. The conjecture that every E over Q is modu-
lar became famous (and was called according to taste by various subsets
of the names Taniyama, Weil and Shimura, although none of these three
people had ever stated the conjecture explicitly in print). It was finally
proved at the end of the 1990’s by Andrew Wiles and his collaborators and
followers:

Theorem (Wiles–Taylor, Breuil–Conrad–Diamond–Taylor). Every el-

liptic curve over Q can be parametrized by modular functions.
The proof, which is extremely difficult and builds on almost the entire
apparatus built up during the previous decades in algebraic geometry, repre-
sentation theory and the theory of automorphic forms, is one of the pinnacles
of mathematical achievement in the 20th century.

♠ Fermat’s Last Theorem

In the 1970’s, Y. Hellegouarch was led to consider the elliptic curve (49) in
the special case when the roots of the cubic polynomial on the right were
nth powers of rational integers for some prime number n > 2, i.e., if this
cubic factors as (x − an )(x − bn )(x − cn ) where a, b, c satisfy the Fermat
equation an + bn + cn = 0. A decade later, G. Frey studied the same el-
liptic curve and discovered that the associated Galois representation (we
 Elliptic Modular Forms and Their Applications 47

do not explain this here) had properties which contradicted the properties
which Galois representations of elliptic curves were expected to satisfy. Pre-
cise conjectures about the modularity of certain Galois representations were
then made by Serre which would fail for the representations attached to the
Hellegouarch-Frey curve, so that the correctness of these conjectures would
imply the insolubility of Fermat’s equation. (Very roughly, the conjectures
imply that, if the Galois representation associated to the above curve E is
modular at all, then the corresponding cusp form would have to be congru-
ent modulo n to a cusp form of weight 2 and level 1 or 2, and there aren’t
any.) In 1990, K. Ribet proved a special case of Serre’s conjectures (the gen-
eral case is now also known, thanks to recent work of Khare, Wintenberger,
Dieulefait and Kisin) which was sufficient to yield the same implication. The
proof by Wiles and Taylor of the Taniyama-Weil conjecture (still with some
minor restrictions on E which were later lifted by the other authors cited
above, but in sufficient generality to make Ribet’s result applicable) thus suf-
ficed to give the proof of the following theorem, first claimed by Fermat in
1637:

Theorem (Ribet, Wiles–Taylor). If n > 2, there are no positive integers

with an + bn = cn . ♥
Finally, we should mention that the connection between modularity and
algebraic geometry does not apply only to elliptic curves. Without going
into detail, we mention only that the Hasse–Weil zeta function Z(X/Q, s)
of an arbitrary smooth projective variety X over Q splits into factors corre-
sponding to the various cohomology groups of X, and that if any of these
cohomology groups (or any piece of them under some canonical decompo-
sition, say with respect to the action of a finite group of automorphisms
of X) is two-dimensional, then the corresponding piece of the zeta func-
tion is conjectured to be the L-series of a Hecke eigenform of weight i + 1,
where the cohomology group in question is in degree i. This of course in-
cludes the case when X = E and i = 1, since the first cohomology group
of a curve of genus 1 is 2-dimensional, but it also applies to many higher-
dimensional varieties. Many examples are now known, an early one, due to
R. Livné, being given by the cubic hypersurface x31 + · · · + x310 = 0 in the
projective space {x ∈ P9 | x1 + · · · + x10 = 0}, whose zeta-function equals
7
j=0 ζ(s − i)
mi
· L(s − 2, f )−1 where (m0 , . . . m7 ) = (1, 1, 1, −83, 43, 1, 1, 1)
and f = q + 2q − 8q 3 + 4q 4 + 5q 5 + · · · is the unique new form of weight 4
2

on Γ0 (10). Other examples arise from so-called “rigid Calabi-Yau 3-folds,”

which have been studied intensively in recent years in connection with the
phenomenon, first discovered by mathematical physicists, called “mirror sym-
metry.” We skip all further discussion, referring to the survey paper and book
cited in the references at the end of these notes.
48 D. Zagier

5 Modular Forms and Differential Operators

The starting point for this section is the observation that the derivative of
a modular form is not modular, but nearly is. Specifically, if f is a modular
form of weight k with the Fourier expansion (3), then by differentiating (2)
we see that the derivative
∞
1 df df
Df = f  := = q = n an q n (51)
2πi dz dq n=1

(where the factor 2πi has been included in order to preserve the rationality
properties of the Fourier coefficients) satisfies
 
az + b k
f = (cz + d)k+2 f  (z) + c (cz + d)k+1 f (z) . (52)
cz + d 2πi
If we had only the first term, then f  would be a modular form of weight k + 2.
The presence of the second term, far from being a problem, makes the theory
much richer. To deal with it, we will:
• modify the differentiation operator so that it preserves modularity;
• make combinations of derivatives of modular forms which are again modu-
lar;
• relax the notion of modularity to include functions satisfying equations
like (52);
• differentiate with respect to t(z) rather than z itself, where t(z) is a modu-
lar function.
These four approaches will be discussed in the four subsections 5.1–5.4, re-
spectively.

5.1 Derivatives of Modular Forms

As already stated, the first approach is to introduce modifications of the op-
erator D which do preserve modularity. There are two ways to do this, one
holomorphic and one not. We begin with the holomorphic one. Comparing
the transformation equation (52) with equations (19) and (17), we find that
for any modular form f ∈ Mk (Γ1 ) the function
k
ϑk f := f  − E2 f , (53)
12
sometimes called the Serre derivative, belongs to Mk+2 (Γ1 ). (We will often
drop the subscript k, since it must always be the weight of the form to which
the operator is applied.) A first consequence of this basic fact is the following.
We introduce the ring M∗ (Γ1 ) := M∗ (Γ1 )[E2 ] = C[E2 , E4 , E6 ], called the ring
of quasimodular forms on SL(2, Z). (An intrinsic definition of the elements of
this ring, and a definition for other groups Γ ⊂ G, will be given in the next
subsection.) Then we have:
 Elliptic Modular Forms and Their Applications 49

Proposition 15. The ring M∗ (Γ1 ) is closed under differentiation. Specifi-

cally, we have

E22 − E4 E2 E4 − E6 E2 E6 − E42
E2 = , E4 = , E6 = . (54)
12 3 2
Proof. Clearly ϑE4 and ϑE6 , being holomorphic modular forms of weight 6
and 8 on Γ1 , respectively, must be proportional to E6 and E42 , and by looking
at the first terms in their Fourier expansion we find that the factors are −1/3
and −1/2. Similarly, by differentiating (19) we find the analogue of (53) for
E2 , namely that the function E2 − 12
1
E22 belongs to M4 (Γ ). It must therefore
be a multiple of E4 , and by looking at the first term in the Fourier expansion
one sees that the factor is −1/12 .

Proposition 15, first discovered by Ramanujan, has many applications. We de-

scribe two of them here. Another, in transcendence theory, will be mentioned
in Section 6.

♠ Modular Forms Satisfy Non-Linear Differential Equations

An immediate consequence of Proposition 15 is the following:

Proposition 16. Any modular form or quasi-modular form on Γ1 satisfies
a non-linear third order differential equation with constant coefficients.

Proof. Since the ring M∗ (Γ1 ) has transcendence degree 3 and is closed under
differentiation, the four functions f , f  , f  and f  are algebraically dependent
for any f ∈ M∗ (Γ1 ).

As an example, by applying (54) several times we find that the function

E2 satisfies the non-linear differential equation f  − f f  + 32 f  = 0 . This is
2

called the Chazy equation and plays a role in the theory of Painlevé equations.
We can now use modular/quasimodular ideas to describe a full set of solu-
tions of this equation. First,
 define  a b operator” f → f 2 g by
 π a c“modified slash
(f 2 g)(z) = (cz + d)−2 f az+b
cz+d + 12 cz+d for g = c d . This is not linear in f
(it is only affine), but it is nevertheless a group operation,
  as one checks easily,
and, at least locally, it makes sense for any matrix ac db ∈!SL(2,"C). Now one
checks by a direct, though tedious, computation that Ch f 2 g = Ch[f ]|8 g
(where defined) for any g ∈ SL(2, C), where Ch[f ] = f  − f f  + 32 f  . (Again
2

this is surprising, because the operator Ch is not linear.) Since E2 is a so-

lution of Ch[f ] = 0, it follows that E2 2 g is a solution of the same equation
(in g −1 H ⊂ P1 (C)) for every g ∈ SL(2, C), and since E2 2 γ = E2 for γ ∈ Γ1
and 2 is a group operation, it follows that this function depends only on
the class of g in Γ1 \SL(2, C). But Γ1 \SL(2, C) is 3-dimensional and a third-
order differential equation generically has a 3-dimensional solution space (the
values of f , f  and f  at a point determine all higher derivatives recursively
50 D. Zagier

and hence fix the function uniquely in a neighborhood of any point where it

is holomorphic), so that, at least generically, this describes all solutions of the
non-linear differential equation Ch[f ] = 0 in modular terms. ♥
Our second “application” of the ring M∗ (Γ1 ) describes an unexpected ap-
pearance of this ring in an elementary and apparently unrelated context.

♠ Moments of Periodic Functions

This very pretty application of the modular forms E2 , E4 , E6 is due to

P. Gallagher. Denote by P the space of periodic real-valued functions on R,
i.e., functions f : R → R satisfying f (x + 2π) = f (x). For each ∞-tuple
n = (n0 , n1 , . . . ) of non-negative integers (all but finitely many equal to 0)
we define a “coordinate” In on the infinite-dimensional space P by In [f ] =
 2π n0 
0 f (x) f (x)
n1
· · · dx (higher moments). Apart
  from the relations
  among
these
 coming from integration by parts, like f (x)dx = 0 or f (x)2 dx =
− f (x)f  (x)dx, we also have various inequalities. The general problem, cer-
tainly too hard to be solved completely, would be to describe all equalities and
inequalities among the In [f ]. As a special case we can ask for the complete 
list of inequalities satisfied by the four moments A[f ], B[f ], C[f ], D[f ] :=
 2π  2π 2  2π 3  2π  2 
0
f, 0 f , 0 f , 0 f as f ranges over P. Surprisingly enough, the
answer involves quasimodular forms on SL(2, Z). First, by making a lin-
ear shift f → λf + μ with λ, μ ∈ R we can suppose that A[f ] = 0 and
D[f ] = 1. The problem is then to describe the subset X ⊂ R2 of pairs
   2π  2π 
B[f ], C[f ] = 0 f (x)2 dx, 0 f (x)3 dx where f ranges over functions in
 2π  2π
P satisfying 0 f (x)dx = 0 and 0 f  (x)2 dx = 1.

Theorem (Gallagher). We have X = (B, C) ∈ R2 | 0 < B ≤ 1, C 2 ≤
Φ(B) where the function Φ : (0, 1] → R≥0 is given parametrically by
  
G2 (it) (G4 (it) − G2 (it))2  
Φ  = 0<t≤∞ .
G4 (it) 2 G4 (it)3

The idea of the proof is as follows. First, from A = 0 and D = 1 we deduce

0 < B[f ] ≤ 1 by an inequality of Wirtinger (just look at the Fourier expansion
of f ). Now let f ∈ P be a function – but one must prove that it exists! – which
maximizes C = C[f ] for given values of A, B and D. By a standard calculus-
of-variations-type argument (replace f by f + εg where g ∈ P is orthogonal
to 1, f and f  , so that A, B and D do not change to first order, and then
use that C also cannot change to first order since otherwise its value could
not be extremal, so that g must also be orthogonal to f 2 ), we show that the
four functions 1, f , f 2 and f  are linearly dependent. From this it follows by
integrating once that the five functions 1, f , f 2 , f 3 and f  are also linearly
2

dependent. After a rescaling f → λf + μ, we can write this dependency as

f  (x)2 = 4f (x)3 − g2 f (x) − g3 for some constants g2 and g3 . But this is the
 Elliptic Modular Forms and Their Applications 51

famous differential equation of the Weierstrass ℘-function, so f (2πx) is the

restriction to R of the function ℘(x, Zτ + Z) for some τ ∈ H, necessarily of
the form τ = it with t > 0 because everything is real. Now the coefficients g2
and g3 , by the classical Weierstrass theory, are simple multiples of G4 (it) and
 2π
G6 (it), and for this function f the value of A(f ) = 0 f (x)dx is known to be
a multiple of G2 (it). Working out all the details, and then rescaling f again
to get A = 0 and D = 1, one finds the result stated in the theorem. ♥

We now turn to the second modification of the differentiation operator

which preserves modularity, this time, however, at the expense of sacrificing
holomorphy. For f ∈ Mk (Γ ) (we now no longer require that Γ be the full
modular group Γ1 ) we define

k
∂k f (z) = f  (z) − f (z) , (55)
4πy
where y denotes the imaginary part of z. Clearly this is no longer holomorphic,
but from the calculation
   
1 |cz + d|2 (cz + d)2 ab
= = − 2ic(cz+d) γ = ∈ SL(2, R)
I(γz) y y cd

and (52) one easily sees that it transforms like a modular form of weight k + 2,
i.e., that (∂k f )|k+2 γ = ∂k f for all γ ∈ Γ . Moreover, this remains true even if f
1 ∂f
is modular but not holomorphic, if we interpret f  as . This means that
2πi ∂z
we can apply ∂ = ∂k repeatedly to get non-holomorphic modular forms ∂ nf
of weight k + 2n for all n ≥ 0. (Here, as with ϑk , we can drop the subscript k
because ∂k will only be applied to forms of weight k; this is convenient because
we can then write ∂ nf instead of the more correct ∂k+2n−2 · · · ∂k+2 ∂k f .) For
example, for f ∈ Mk (Γ ) we find
  
1 ∂ k+2 k
∂2f = − f − f
2πi ∂z 4πy 4πy
k  k k+2  k(k + 2)
= f  − f − f − f + f
4πy 16π 2 y 2 4πy 16π 2 y 2
k+1  k(k + 1)
= f  − f + f
2πy 16π 2 y 2
and more generally, as one sees by an easy induction,
n  
n n−r n (k + r)n−r r
∂ f = (−1) Df, (56)
r=0
r (4πy)n−r

where (a)m = a(a+1) · · · (a+m−1) is the Pochhammer symbol. The inversion

of (56) is
52 D. Zagier
 
n
n n (k + r)n−r r
D f = ∂f, (57)
r=0
r (4πy)n−r
and describes the decomposition of the holomorphic but non-modular form
f (n) = Dn f into non-holomorphic but modular pieces: the function y r−n ∂ rf
is multiplied by (cz + d)k+n+r (cz̄ + d)n−r when z is replaced by az+b
cz+d with
a b
c d ∈ Γ .
Formula (56) has a consequence which will be important in §6. The usual
way to write down modular forms is via their Fourier expansions, i.e., as
power series in the quantity q = e2πiz which is a local coordinate at infin-
ity for the modular curve Γ \H. But since modular forms are holomorphic
functions in the upper half-plane, they also have Taylor series expansions in
the neighborhood of any point z = x + iy ∈ H. The “straight” Taylor se-
ries expansion, giving f (z + w) as a power series in w, converges only in
the disk |w| < y centered at z and tangent to the real line, which is un-
natural since the domain of holomorphy of f is the whole upper half-plane,
not just this disk. Instead, we should remember that we can map H iso-
morphically to the unit disk, with z mapping to 0, by sending z  ∈ H to

w = zz −z  z−z̄w
−z̄ . The inverse of this map is given by z = 1−w , and then if f is
a modular form of weight k we should also include the automorphy factor
(1 − w)−k corresponding to this fractional linear transformation (even though
it belongs to PSL(2, C) and not Γ ). The most
 natural way to study f near
z is therefore to expand (1 − w)−k f z−z̄w1−w in powers of w. The following
proposition describes the coefficients of this expansion in terms of the opera-
tor (55).

Proposition 17. Let f be a modular form of weight k and z = x + iy a point

of H. Then
  ∞
 −k z − z̄w (4πyw)n
1−w f = ∂ nf (z) (|w| < 1) . (58)
1−w n=0
n!

Proof. From the usual Taylor expansion, we find

   
 −k z − z̄w  −k 2iyw
1−w f = 1−w f z +
1−w 1−w
 
 −k ∞ Drf (z) −4πyw r
= 1−w ,
r=0
r! 1−w

and now expanding (1 − w)−k−r by the binomial theorem and using (56) we
obtain (58).

Proposition 17 is useful because, as we will see in §6, the expansion (58), after
some renormalizing, often has algebraic coefficients that contain interesting
arithmetic information.
 Elliptic Modular Forms and Their Applications 53

5.2 Rankin–Cohen Brackets and Cohen–Kuznetsov Series

Let us return to equation (52) describing the near-modularity of the deriva-
tive of a modular form f ∈ Mk (Γ ). If g ∈ M (Γ ) is a second modu-
lar form on the same group, of weight , then this formula shows that
the non-modularity of f  (z)g(z) is given by an additive correction term
(2πi)−1 kc (cz + d)k+ +1 f (z) g(z). This correction term, multiplied by , is
symmetric in f and g, so the difference [f, g] = kf g  − f  g is a modular form
of weight k +  + 2 on Γ . One checks easily that the bracket [ · , · ] defined
in this way is anti-symmetric and satisfies the Jacobi identity, making M∗ (Γ )
into a graded Lie algebra (with grading given by the weight + 2). Further-
more, the bracket g → [f, g] with a fixed modular form f is a derivation with
respect to the usual multiplication, so that M∗ (Γ ) even acquires the structure
of a so-called Poisson algebra.
We can continue this construction to find combinations of higher deriva-
tives of f and g which are modular, setting [f, g]0 = f g, [f, g]1 = [f, g] =
kf g  − f  g,
k(k + 1)  ( + 1) 
[f, g]2 = f g − (k + 1)( + 1)f  g  + f g,
2 2
and in general
  
k+n−1 +n−1
[f, g]n = (−1)r Drf Dsg (n ≥ 0) , (59)
s r
r, s≥0
r+s=n

the nth Rankin–Cohen bracket of f and g.

Proposition 18. For f ∈ Mk (Γ ) and g ∈ M (Γ ) and for every n ≥ 0, the
function [f, g]n defined by (59) belongs to Mk+ +2n (Γ ).
There are several ways to prove this. We will do it using Cohen–Kuznetsov
series. If f ∈ Mk (Γ ), then the Cohen–Kuznetsov series of f is defined by
∞
Dnf (z) n
fD (z, X) = X ∈ Hol0 (H)[[X]] , (60)
n=0
n! (k)n

where (k)n = (k + n − 1)!/(k − 1)! = k(k + 1) · · · (k + n − 1) is the Pochhammer

symbol already used above and Hol0 (H) denotes the space of holomorphic
functions in the upper half-plane of subexponential growth (see §1.1). This
series converges for all X ∈ C (although for our purposes only its properties
as a formal power series in X will be needed). Its key property is given by:
Proposition 19. If f ∈ Mk (Γ ), then the Cohen–Kuznetsov series defined by
(60) satisfies the modular transformation equation
   
az + b X k c X
fD , = (cz + d) exp fD (z, X) . (61)
cz + d (cz + d)2 cz + d 2πi
 
for all z ∈ H, X ∈ C, and γ = ac db ∈ Γ .
54 D. Zagier

Proof. This can be proved in several different ways. One way is direct: one
shows by induction on n that the derivative Dnf (z) transforms under Γ
by
  n  
n az + b n (k + r)n−r n−r
D f = n−r
c (cz + d)k+n+r Drf (z)
cz + d r=0
r (2πi)

for all n ≥ 0 (equation (52) is the case n = 1 of this), from which the claim
follows easily. Another, more elegant, method is to use formula (56) or (57)
to establish the relationship

fD (z, X) = eX/4πy f∂ (z, X) (z = x + iy ∈ H, X ∈ C) (62)

between fD (z, X) and the modified Cohen–Kuznetsov series

∞
∂ nf (z) n
f∂ (z, X) = X ∈ Hol0 (H)[[X]] . (63)
n=0
n! (k)n

The fact that each function ∂ nf (z) transforms like a modular form of weight
k + 2n on Γ implies that f∂ (z, X) is multiplied by (cz + d)k when z
and X are replaced by az+b X
cz+d and (cz+d)2 , and using (62) one easily de-
duces from this the transformation formula (61). Yet a third way is to
observe that fD (z, X) is the unique solution of the differential equation
 ∂2 
∂
X ∂X 2 + k ∂X − D fD = 0 with the initial condition fD (z, 0) = f (z) and
 
that (cz + d)−k e−cX/2πi(cz+d) fD az+b X
cz+d , (cz+d)2 satisfies the same differential
equation with the same initial condition.
Now to deduce Proposition 18 we simply look at the product of fD (z, −X)
with gD (z, X). Proposition 19 implies that this product is multiplied by
(cz + d)k+ when z and X are replaced by az+b X
cz+d and (cz+d)2 (the factors
involving an exponential in X cancel), and this means that the coefficient of
[f,g]n
X n in the product, which is equal to (k) n ( )n
, is modular of weight k +  + 2n
for every n ≥ 0.

Rankin–Cohen brackets have many applications in the theory of modu-

lar forms. We will describe two – one very straightforward and one more
subtle – at the end of this subsection, and another one in §5.4. First, how-
ever, we make a further comment about the Cohen–Kuznetsov series at-
tached to a modular form. We have already introduced two such series: the
series fD (z, X) defined by (60), with coefficients proportional to Dnf (z),
and the series f∂ (z, X) defined by (63), with coefficients proportional to
∂ nf (z). But, at least when Γ is the full modular group Γ1 , we had de-
fined a third differentiation operator besides D and ∂, namely the opera-
tor ϑ defined in (53), and it is natural to ask whether there is a correspond-
ing Cohen–Kuznetsov series fϑ here also. The answer is yes, but this series
 Elliptic Modular Forms and Their Applications 55

n≥0 ϑ f (z) X /n! (k)n . Instead, for f ∈ Mk (Γ1 ),
n n
is not simply given by
we define a sequence of modified derivatives ϑ[n]f ∈ Mk+2n (Γ1 ) for n ≥ 0
by
  E4 [r−1]
ϑ[0]f = f, ϑ[1]f = ϑf, ϑ[r+1]f = ϑ ϑ[r]f − r(k + r − 1) ϑ f for r ≥ 1
144
(64)
(the last formula also holds for r = 0 with f [−1] defined as 0 or in any other
way), and set
∞
ϑ[n]f (z) n
fϑ (z, X) = X .
n=0
n! (k)n
Using the first equation in (54) we find by induction on n that
n    n−r
n E2
Dnf = (k + r)n−r ϑ[r]f (n = 0, 1, . . . ) , (65)
r=0
r 12

(together with similar formulas for ∂ nf in terms of ϑ[n]f and for ϑ[n]f in
terms of Dnf or ∂ nf ), the explicit version of the expansion of Dnf as
a polynomial in E2 with modular coefficients whose existence is guaran-
teed by Proposition 15. This formula together with (62) gives us the rela-
tions
∗
fϑ (z, X) = e−XE2 (z)/12 fD (z, X) = e−XE2 (z)/12 f∂ (z, X) (66)
between the new series and the old ones, where E2∗ (z) is the non-holomorphic
Eisenstein series defined in (21), which transforms like a modular form of
weight 2 on Γ1 . More generally, if we are on any discrete subgroup Γ of
SL(2, R), we choose a holomorphic or meromorphic function φ in H such that
the function φ∗ (z) = φ(z) − 4πy
1
transforms like a modular form of weight 2
 
on Γ , or equivalently such that φ az+b 1
= (cz + d)2 φ(z) + 2πi c(cz + d)
a b cz+d
for all c d ∈ Γ . (Such a φ always exists, and if Γ is commensurable
1
with Γ1 is simply the sum of 12 E2 (z) and a holomorphic or meromorphic
modular form of weight 2 on Γ ∩ Γ1 .) Then, just as in the special case
φ = E2 /12, the operator ϑφ defined by ϑφ f := Df − kφf for f ∈ Mk (Γ )
sends Mk (Γ ) to Mk+2 (Γ ), the function ω := φ − φ2 belongs to M4 (Γ ) (gen-
eralizing the first equation in (54)), and if we generalize the above defini-
[n]
tion by introducing operators ϑφ : Mk (Γ ) → Mk+2n (Γ ) (n = 0, 1, . . . )
by
[0] [r+1] [r] [r−1]
ϑφ f = f , ϑφ f = ϑφ f + r(k + r − 1) ω ϑφ f for r ≥ 0 , (67)
1
then (65) holds with 12 E2 replaced by φ, and (66) is replaced by
∞ [n]
ϑφ f (z) X n ∗
fϑφ (z, X) := = e−φ(z)X fD (z, X) = e−φ (z)X
f∂ (z, X) .
n=0
n! (k)n
(68)
56 D. Zagier

These formulas will be used again in §§6.3–6.4.

We now give the two promised applications of Rankin–Cohen brackets.

♠ Further Identities for Sums of Powers of Divisors

In §2.2 we gave identities among the divisor power sums σν (n) as one of
our first applications of modular forms and of identities like E42 = E8 .
By including the quasimodular form E2 we get many more identities of

the same type, e.g., the relationship
 E22 = E4 + 12E
 2 gives the identity
n−1
m=1 σ1 (m)σ1 (n − m) = 12 5σ3 (n) − (6n − 1)σ1 (n) and similarly for the
1

other two formulas in (54). Using the Rankin–Cohen brackets we get yet more.
For instance, the first Rankin–Cohen bracket of E4 (z) and E6 (z) is a cusp form
of weight 12 on Γ1 , so it must be a multiple of Δ(z), and this leads to the for-
n−1
7
mula τ (n) = n( 12 σ5 (n) + 125
σ3 (n)) − 70 m=1 (5m − 2n)σ3 (m)σ5 (n − m) for
the coefficient τ (n) of q n in Δ. (This can also be expressed completely in terms
of elementary functions, without mentioning Δ(z) or τ (n), by writing Δ as
a linear combination of any two of the three functions E12 , E4 E8 and E62 .) As
in the case of the identities mentioned in §2.2, all of these identities also have
combinatorial proofs, but these involve much more work and more thought
than the (quasi)modular ones. ♥

♠ Exotic Multiplications of Modular Forms

A construction which is familiar both in symplectic geometrys and in quantum

theory (Moyal brackets) is that of the deformation of the multiplication in an
algebra. If A is an algebra over some field k, say with a commutative and asso-
ciative multiplication μ : A ⊗ A → A, then we can look at deformations με of
μ given by formal power series με (x, y) = μ0 (x, y)+μ1 (x, y)ε+μ2 (x, y)ε2 +· · ·
with μ0 = μ which are still associative but are no longer necessarily commu-
tative. (To be more precise, με is a multiplication on A if there is a topol-
ogy and the series above is convergent and otherwise, after being extended
k[[ε]]-linearly, a multiplication on A[[ε]].) The linear term μ1 (x, y) in the ex-
pansion of με is then anti-symmetric and satisfies the Jacobi identity, making
A (or A[[ε]]) into a Lie algebra, and the μ1 -product with a fixed element
of A is a derivation with respect to the original multiplication μ, giving the
structure of a Poisson algebra. All of this is very reminiscent of the zeroth
and first Rankin–Cohen brackets, so one can ask whether these two brack-
ets arise as the beginning of the expansion of some deformation of the or-
dinary multiplication of modular forms. Surprisingly, this is not only true,
but there is in fact a two-parameter family of such deformed multiplica-
tions:
 Elliptic Modular Forms and Their Applications 57

Theorem. Let u and v be two formal variables and Γ any subgroup of


∞
SL(2, R). Then the multiplication μu,v on Mk (Γ ) defined by
k=0
∞
μu,v (f, g) = tn (k, ; u, v) [f, g]n (f ∈ Mk (Γ ), g ∈ M (Γ )) , (69)
n=0

where the coefficients tn (k, ; u, v) ∈ Q(k, l)[u, v] are given by

 1  u  u 
−2 v −1 −v
 
n j j j
tn (k, ; u, v) = v n    ,
2j −k − 12 − − 12 n + k + l − 32
0≤j≤n/2
j j j
(70)
is associative.
Of course the multiplication given by (u, v) = (0, 0) is just the usual mul-
tiplication of modular forms, and the multiplications associated to (u, v) and
(λu, λv) are isomorphic by rescaling f → λk f for f ∈ Mk , so the set of
new multiplications obtained this way is parametrized by a projective line.
Two of the multiplications obtained are noteworthy: the one corresponding to
(u : v) = (1 : 0) because it is the only commutative one, and the one cor-
responding to (u, v) = (0, 1) because it is so simple, being given just by
f ∗ g = n [f, g]n .
These deformed multiplications were found, at approximately the same
time, in two independent investigations and as consequences of two quite
different theories. On the one hand, Y. Manin, P. Cohen and I studied
Γ -invariant and twisted Γ -invariant pseudodifferentialoperators in the up-
per half-plane. These are formal power series Ψ (z) = n≥h fn (z)D−n , with
   az+b 
D as in (52), transforming under Γ by Ψ az+b cz+d = Ψ (z) or by Ψ cz+d =
(cz + d)κ Ψ (z)(cz + d)−κ , respectively, where κ is some complex parameter.
(Notice that the second formula is different from the first because multiplica-
tion by a non-constant function of z does not commute with the differentiation
operator D, and is well-defined even for non-integral κ because the ambiguity
of argument involved in choosing a branch of (cz + d)κ cancels out when one
divides by (cz+ d)κ on the other side.) Using that D transforms under the
action of ac db to (cz + d)2 D, one finds that the leading coefficient fh of F is
then a modular form on Γ of weight 2h and that the higher coefficients fh+j
(j ≥ 0) are specific linear combinations (depending on h, j and κ) of Dj g0 ,
Dj−1 g1 , . . . , gj for some modular forms gi ∈ M2h+2i (G), so that (assuming
that Γ contains −1 and hence has only modular forms of even weight) we
can canonically identify  the space of invariant or twisted invariant pseudodif-
∞
ferential operators with k=0 Mk (Γ ). On the other hand, pseudo-differential
operators can be multiplied just by multiplying out the formal series defining
them using Leibnitz’s rule, this multiplication clearly being associative, and
this then leads to the family of non-trivial multiplications of modular forms
given in the theorem, with u/v = κ − 1/2. The other paper in which the same
58 D. Zagier

multiplications arise is one by A. and J. Untenberger which is based on a cer-

tain realization of modular forms as Hilbert-Schmidt operators on L2 (R). In
both constructions, the coefficients tn (k, ; u, v) come out in a different and
less symmetric form than (70); the equivalence with the form given above is
a very complicated combinatorial identity. ♥

5.3 Quasimodular Forms

We now turn to the definition of the ring M∗ (Γ ) of quasimodular forms on Γ .

In §5.1 we defined this ring for the case Γ = Γ1 as C[E2 , E4 , E6 ]. We now give
an intrinsic definition which applies also to other discrete groups Γ .
We start by defining the ring of almost holomorphic modular forms on Γ .
By definition, such a form is a function in H which transforms like a modular
form but, instead of being holomorphic, is a polynomial in 1/y (with y = I(z)
as usual) with holomorphic coefficients, the motivating examples being the
non-holomorphic Eisenstein series E2∗ (z) and the non-holomorphic derivative
∂f (z) of a holomorphic modular form as defined in equations (21) and (55),
respectively. More precisely, an almost holomorphic modular p form of weight k
and depth ≤ p on Γ is a function of the form F (z) = r=0 fr (z)(−4πy)−r
with each fr ∈ Hol0 (H) (holomorphic functions of moderate growth) satisfying
F |k γ = F for all γ ∈ Γ . We denote by M #(≤p) = M #(≤p) (Γ ) the space of such
 k k
forms and by M #∗ = #k , M
M #k = ∪p M #(≤p) the graded and filtered ring of all
k k
almost holomorphic modular forms, usually omitting Γ from the notations.
For the two basic examples E2∗ ∈ M #(≤1) (Γ1 ) and ∂k f ∈ M #(≤1) (Γ ) (where
2 k+2
f ∈ Mk (Γ )) we have f0 = E2 , f1 = 12 and f0 = Df , f1 = kf , respectively.
(≤p) (≤p)
We now define the space Mk = Mk (Γ ) of quasimodular forms of
weight k and depth ≤ p on Γ as the space of “constant terms” f0 (z) of F (z)
as F runs over M #(≤p) . It is not hard to see that the almost holomorphic
k
modular form F is uniquely determined by its constant term f0 , so the ring
 (≤p)
M∗ = k Mk (Mk = ∪p Mk ) of quasimodular forms on Γ is canonically
#∗ of almost holomorphic modular forms on Γ . One can
isomorphic to the ring M
also define quasimodular forms directly, as was pointed out to me by W. Nahm:
a quasimodular form of weight k and depth≤ pon Γ is a function f ∈ Hol  0 (H)
such that, for fixed z ∈ H and variable γ = ac db ∈ Γ , the function f |k γ (z) is
a polynomial of degree ≤ p in cz+dc
. Indeed, if f (z) = f0 (z) ∈ Mk corresponds
 #k , then the modularity of F implies the
to F (z) = rfr (z) (−4πy)−r ∈ M r

identity f |k γ (z) = r fr (z) cz+d c
with the same coefficients fr (z), and
conversely.
The basic facts about quasimodular forms are summarized in the following
proposition, in which Γ is a non-cocompact discrete subgroup of SL(2, R) and
φ ∈ M2 (Γ ) is a quasimodular form of weight 2 on Γ which is not modular,
e.g., φ = E2 if Γ is a subgroup of Γ1 . For Γ = Γ1 part (i) of the proposition
reduces to Proposition 15 above, while part (ii) shows that the general defi-
 Elliptic Modular Forms and Their Applications 59

nition of quasimodular forms given here agrees with the ad hoc one given in
§5.1 in this case.
Proposition 20. (i) The space of quasimodular forms on Γ is closed un-
 (≤p)  (≤p+1)
der differentiation. More precisely, we have D Mk ⊆ Mk+2 for all
k, p ≥ 0.
(ii) Every quasimodular form on Γ is a polynomial in φ with modular co-
(≤p) p
efficients. More precisely, we have Mk (Γ ) = r=0 Mk−2r (Γ ) · φr for all
k, p ≥ 0.
(iii) Every quasimodular form on Γ can be written uniquely as a linear com-
bination of derivatives of modular forms and of φ. More precisely, for all
k, p ≥ 0 we have
⎧  
⎨ p Dr Mk−2r (Γ ) if p < k/2 ,
(≤p) r=0
Mk (Γ ) = 
⎩ k/2−1 Dr M 
k−2r (Γ ) ⊕ C · D φ if p ≥ k/2 .
k/2−1
r=0

 #k correspond to f = f0 ∈ Mk . The
Proof. Let F = fr (−4πy)−r ∈ M
#
almost
! holomorphic form ∂ " k F ∈ Mk+2 then has the expansion ∂k F =
D(fr ) + (k − r + 1)fr−1 (−4πy)−r , with constant term Df . This proves
the first statement. (One can also prove it in terms of the direct definition
of quasimodular forms by differentiating the formula expressing the transfor-
mation behavior of f under Γ .) Next, one checks easily that if F belongs to
#(≤p) , then the last coefficient fp (z) in its expansion is a modular form of
M k
weight k − 2p. It follows that p ≤ k/2 (if fp = 0, i.e., if F has depth exactly p)
and also, since the almost holomorphic modular form φ∗ corresponding to φ
is the sum of φ and a non-zero multiple of 1/y, that F is a linear combi-
nation of fp φ∗ p and an almost holomorphic modular form of depth strictly
smaller than p, from which statement (ii) for almost holomorphic modular
forms (and therefore also for quasimodular forms) follows by induction on p.
Statement (iii) is proved exactly the same way, by subtracting from F a mul-
tiple of ∂ p fp if p < k/2 and a multiple of ∂ k/2−1 φ if p = k/2 to prove by
induction on p the corresponding statement with “quasimodular” and D re-
placed by “almost holomorphic modular” and ∂, and then again using the
isomorphism between M #∗ and M∗ .

There is one more important element of the structure of the ring of quasi-
modular (or almost holomorphic modular) forms. Let F = fr (−4πy)−r and
f = f0 be an almost holomorphic modular form of weight k and depth ≤ p
and the quasimodular form which corresponds to it. One sees easily using the
properties above that each coefficient fr is quasimodular (of weight k − 2r and
depth ≤ p − r) and that, if δ : M∗ → M∗ is the map which sends f to f1 , then
fr = δ rf /r! for all r ≥ 0 (and δ rf = 0 for r > p), so that the expansion of F (z)
in powers of −1/4πy is a kind of Taylor expansion formula. This gives us three
operators from M∗ to itself: the differentiation operator D, the operator E
60 D. Zagier

which multiplies a quasimodular form of weight k by k, and the operator δ.

Each of these operators is a derivation on the ring of quasimodular forms, and
they satisfy the three commutation relations

[E, D] = 2 D , [D, δ] = −2 δ , [D, δ] = E

(of which the first two just say that D and δ raise and lower the weight
of a quasimodular form by 2, respectively), giving to M∗ the structure of
an sl2 C)-module. Of course the ring M #∗ , being isomorphic to M∗ , also be-
comes an sl2 C)-module, the corresponding operators being ϑ (= ϑk on Mk ),
E (= multiplication by k on M #k ) and δ ∗ (= “derivation with respect to
−1/4πy”). From this point of view, the subspace M∗ of M∗ or M #∗ appears sim-
∗
ply as the kernel of the lowering operator δ (or δ ). Using this sl2 C)-module
structure makes many calculations with quasimodular or almost holomorphic
modular forms simpler and more transparent.

♠ Counting Ramified Coverings of the Torns

We end this subsection by describing very briefly a beautiful and unexpected

context in which quasimodular forms occur. Define
  2  2 
Θ(X, z, ζ) = (1 − q n ) 1 − en X/8 q n/2 ζ 1 − e−n X/8 q n/2 ζ −1 ,
n>0 n>0
n odd

expand Θ(X, z, ζ) asa Laurent series n∈Z Θn (X, z) ζ n , and expand Θ0 (X, z)
∞ 2r
as a Taylor series r=0 Ar (z) X . Then a somewhat intricate calculation
involving Eisenstein series, theta series and quasimodular forms on Γ (2) shows
that each Ar is a quasimodular form of weight 6r on SL(2, Z), i.e., a weighted
homogeneous polynomial in E2 , E4 , and E6 . In particular, A0 (z) = 1 (this is
a consequence of the Jacobi triple product identity, which says that Θn (0, z) =
2 ∞
(−1)n q n /2 ), so we can also expand log Θ0 (X, z) as r=1 Fr (z) X 2r and Fr (z)
is again quasimodular of weight 6r. These functions arose in the study of the
“one-dimensional analogue of mirror symmetry”: the coefficient of q m in Fr
counts the generically ramified coverings of degree m of a curve of genus 1
by a curve of genus r + 1. (“Generic” means that each point has ≥ m − 1
preimages.) We thus obtain:
Theorem. The generating function of generically ramified coverings of a torus
by a surface of genus g > 1 is a quasimodular form of weight 6g−6 on SL(2, Z).
As an example, we have F1 = A1 = 103680 1
(10E23 − 6E2 E4 − 4E6 ) = q 2 +
8q + 30q + 80q + 180q + · · · . In this case (but for no higher genus), the
3 4 5 6
1
function F1 = 1440 (E4 + 10E2 ) is also a linear combination of derivatives of
Eisenstein series and we get a simple explicit formula n(σ3 (n) − nσ1 (n))/6 for
the (correctly counted) number of degree n coverings of a torus by a surface
of genus 2. ♥
 Elliptic Modular Forms and Their Applications 61

5.4 Linear Differential Equations and Modular Forms

The statement of Proposition 16 is very simple, but not terribly useful, because
it is difficult to derive properties of a function from a non-linear differential
equation. We now prove a much more useful fact: if we express a modular form
as a function, not of z, but of a modular function of z (i.e., a meromorphic
modular form of weight zero), then it always satisfies a linear differentiable
equation of finite order with algebraic coefficients. Of course, a modular form
cannot be written as a single-valued function of a modular function, since the
latter is invariant under modular transformations and the former transforms
with a non-trivial automorphy factor, but in can be expressed locally as such
a function, and the global non-uniqueness then simply corresponds to the
monodromy of the differential equation.
The fact that we have mentioned is by no means new – it is at the heart
of the original discovery of modular forms by Gauss and of the later work of
Fricke and Klein and others, and appears in modern literature as the theory of
Picard–Fuchs differential equations or of the Gauss–Manin connection – but
it is not nearly as well known as it ought to be. Here is a precise statement.
Proposition 21. Let f (z) be a (holomorphic or meromorphic) modular form
of positive weight k on some group Γ and t(z) a modular function with re-
spect to Γ . Express f (z) (locally) as Φ(t(z)). Then the function Φ(t) satisfies
a linear differential equation of order k + 1 with algebraic coefficients, or with
polynomial coefficients if Γ \H has genus 0 and t(z) generates the field of
modular functions on Γ .
This proposition is perhaps the single most important source of applications
of modular forms in other branches of mathematics, so with no apology we
sketch three different proofs, each one giving us different information about
the differential equation in question.
Proof 1: We want to find a linear relation among the derivatives of f with
respect to t. Since f (z) is not defined in Γ \H, we must replace d/dt by the
operator Dt = t (z)−1 d/dz, which makes sense in H. We wish to show that
the functions Dtn f (n = 0, 1, . . . , k + 1) are linearly dependent over the field
of modular functions on Γ , since such functions are algebraic functions of t(z)
in general and rational functions in the special cases when Γ \H has genus 0
and t(z) is a “Hauptmodul” (i.e., t : Γ \H → P1 (C) is an isomorphism). The
difficulty is that, as seen in §5.1, differentiating the transformation equation
(2) produces undesired extra terms as in (52). This is because the automorphy
factor (cz+d)k in (2) is non-constant and hence contributes non-trivially when
we differentiate. To get around this, we replace f (z) by the vector-valued
function F : H → Ck+1 whose mth component (0 ≤ m ≤ k) is Fm (z) =
z k−m f (z). Then
  m
az + b
Fm = (az + b)k−m (cz + d)m f (z) = Mmn Fn (z) , (71)
cz + d n=0
62 D. Zagier
 
for all γ = ac db ∈ Γ , where M = S k (γ) ∈ SL(k + 1, Z) denotes the kth
symmetric power of γ. In other words, we have F (γz) = M F (z) where the new
“automorphy factor” M , although it is more complicated than the automorphy
factor in (2) because it is a matrix rather than a scalar, is now independent of
z, so that when we differentiate we get simply (cz + d)−2 F  (γz) = M F  (z). Of
course, this equation contains (cz + d)−2 , so differentiating again with respect
to z would
 again
 produce unwanted terms. But the Γ -invariance of t implies
that t az+b
cz+d = (cz +d)2 t (z), so the factor (cz +d)−2 is cancelled if we replace
d/dz by Dt = d/dt. Thus (71) implies Dt F (γz) = M Dt F (z), and by induction
Dtr F (γz) = M Dtr F (z) for all r ≥ 0. Now consider the (k + 2) × (k + 2) matrix
 
f Dt f · · · Dtk+1 f
.
F Dt F · · · Dtk+1 F
The top and bottom rows of this matrix are  identical, so its determinant
is 0. Expanding by the top row, we find 0 = k+1 (−1)n det(An (z)) Dtn f (z),
 n=0 
where An (z) is the (k + 1) × (k + 1) matrix F Dt F · · · D  t F · · · Dt
n k+1
F .
From Dtr F (γz) = M Dtr F (z) (∀r) we get An (γz) = M An (z) and hence, since
det(M ) = 1, that An (z) is a Γ -invariant function. Since An (z) is also mero-
morphic (including at the cusps), it is a modular function on Γ and hence an
algebraic or rational function of t(z), as desired. The advantage of this proof is
that it gives us all k + 1 linearly independent solutions of the differential equa-
tion satisfied by f (z): they are simply the functions f (z), zf (z), . . . , z k f (z).

Proof 2 (following a suggestion of Ouled Azaiez): We again use the differ-

entiation operator Dt , but this time work with quasimodular rather than
vector-valued modular forms. As in §5.2, choose a quasimodular form φ(z) of
1
weight 2 on Γ with δ(φ) = 1 (e.g., 12 E2 if Γ = Γ1 ), and write each Dtn f (z)
(n = 0, 1, 2, . . . ) as a polynomial in φ(z) with (meromorphic) modular coeffi-
f ϑφ f
cients. For instance, for n = 1 we find Dt f = k  φ +  where ϑφ denotes
t t
the Serre derivative with respect to φ. Using that φ − φ2 ∈ M4 , one finds
by induction that each Dtn f is the sum of k(k − 1) · · · (k − n + 1) (φ/t )n f
and apolynomial  of degree < n in φ. It follows that the k + 2 func-
tions Dtn (f )/f 0≤n≤k+1 are linear combinations of the k + 1 functions
{(φ/t )n }0≤n≤k with modular functions as coefficients, and hence that they
are linearly dependent over the field of modular functions on Γ .

Proof 3: The third proof will give an explicit differential equation satisfied
by f . Consider first the case when f has weight 1. The funcion t = D(t) is
a (meromorphic) modular form of weight 2, so we can form the Rankin–Cohen
brackets [f, t ]1 and [f, f ]2 of weights 5 and 6, respectively, and the quotients

A = [f,t
f t2
]1 [f,f ]2
and B = − 2f 2 t 2 of weight 0. Then

 
f f t − 2f  t f  f f  − 2f 
2
1
Dt2 f + A Dt f + Bf = + − f = 0.
t t ft 2 t 
t 2 f 2
 Elliptic Modular Forms and Their Applications 63

Since A and B are modular functions, they are rational (if t is a Hauptmodul)
or algebraic (in any case) functions of t, say A(z) = a(t(z)) and B(z) =
b(t(z)), and then the function Φ(t) defined locally by f (z) = Φ(t(z)) satisfies
Φ (t) + a(t)Φ (t) + b(t)Φ(t) = 0.
Now let f have arbitrary (integral) weight k > 0 and apply the above
construction to the function h = f 1/k , which is formally a modular form of
weight 1. Of course h is not really a modular form, since in general it is not
even a well-defined function in the upper half-plane (it changes by a kth root
of unity when we go around a zero of f ). But it is defined locally, and the

functions A = [h,t ht 2
]1
and B = − 2h [h,h]2
2 t 2 are well-defined, because they are

both homogeneous of degree 0 in h, so that the kth roots of unity occur-

ring when we go around a zero of f cancel out. (In fact, a short calculation
shows that they can be written directly in terms of Rankin–Cohen brack-

ets of f and t , namely A = [f,t ]1
kf t 2
[f,f ]2
and B = − k2 (k+1)f 2 t 2 .) Now just as

before A(z) = a(t(z)), B(z) = b(t(z)) for some rational or algebraic func-
tions a(t) and b(t). Then Φ(t)1/k is annihilated by the second order differ-
ential operator L = d2/dt2 + a(t)d/dt + b(t) and Φ(t) itself is annihilated
by the (k + 1)st order differential operator Symk (L) whose solutions are
the kth powers or k-fold products of solutions of the differential equation
LΨ = 0. The coefficients of this operator can be given by explicit expres-
sions in terms of a and b and their derivatives (for instance, for k = 2 we
find Sym2 (L) = d3/dt3 + 3a d2/dt2 + (a + 2a2 + 4b) d/dt + 2(b + 2ab) ), and
these in turn can be written as weight 0 quotients of appropriate Rankin–
Cohen brackets.
Here are two classical examples of Proposition
 n2 /2 21. For the first, we take
Γ = Γ (2), f (z) = θ3 (z)2 (with θ3 (z) = q as in (32)) and t(z) = λ(z),
where λ(z) is the Legendre modular function

 4
η(z/2)8 η(2z)16 η(z/2)16 η(2z)8 θ2 (z)
λ(z) = 16 24
= 1− = , (72)
η(z) η(z)24 θ3 (z)

which is known to be a Hauptmodul for the group Γ (2). Then

 
∞ n  
2 2n λ(z) 1 1
θ3 (z) = = F , ; 1; λ(z) , (73)
n=0
n 16 2 2

∞ (a)n (b)n n
where F (a, b; c; x) = n=0 n! (c)n x with (a)n as in eq. (56) denotes the
Gauss hypergeometric function, whichsatisfies the second order differential
equation x(x − 1) y  + (a + b + 1)x − c y  + aby = 0. For the second example
we take Γ = Γ1 , t(z) = 1728/j(z), and f (z) = E4 (z). Since f is a modular
form of weight 4, it should satisfy a fifth order linear differential equation with
respect to t(z), but by the third proof above one should even have that the
64 D. Zagier

fourth root of f satisfies a second order differential equation, and indeed one
finds
  1 · 5 12 1 · 5 · 13 · 17 122
4
E4 (z) = F 1 5
12 , 12 ;
1; t(z) = 1 + + + ··· ,
1 · 1 j(z) 1 · 1 · 2 · 2 j(z)2
(74)
a classical identity which can be found in the works of Fricke and Klein.

♠ The Irrationality of ζ(3)

In 1978, Roger Apéry created a sensation by proving:


Theorem. The number ζ(3) = n≥1 n−3 is irrational.
What he actually showed was that, if we define two sequences {an } =
{1, 5, 73, 1445, . . . } and {bn } = {0, 6, 351/4, 62531/36, . . . } as the solutions
of the recursion

(n + 1)3 un+1 = (34n3 + 51n2 + 27n + 5) un − n3 un−1 (n ≥ 1) (75)

with initial conditions a0 = 1, a1 = 5, b0 = 0, b1 = 6, then we have

bn
an ∈ Z (∀n ≥ 0) , Dn3 bn ∈ Z (∀n ≥ 0) , lim = ζ(3) , (76)
n→∞ an
where Dn denotes the least common multiple of 1, 2, . . . , n. These three as-
sertions together imply the theorem. Indeed, both an and bn grow like C n by
the recursion, where C = 33.97 . . . is the larger root of C 2 − 34C + 1 = 0. On
the other hand, an−1 bn − an bn−1 = 6/n3 by the recursion and induction, so
the difference between bn /an and its limiting value ζ(3) decreases like C −2n .
Hence the quantity xn = Dn3 (bn −an ζ(3)) grows like by Dn3 /(C +o(1))n , which
tends to 0 as n → ∞ since C > e3 and Dn3 = (e3 + o(1))n by the prime num-
ber theorem. (Chebyshev’s weaker elementary estimate of Dn would suffice
here.) But if ζ(3) were rational then the first two statements in (76) would
imply that the xn are rational numbers with bounded denominators, and this
is a contradiction.
Apéry’s own proof of the three properties (76), which involved complicated
explicit formulas for the numbers an and bn as sums of binomial coefficients,
was very ingenious but did not give any feeling for why any of these three
properties hold. Subsequently, two more enlightening proofs were found by
Frits Beukers, one using representations of an and bn as multiple integrals
involving Legendre polynomials and the other based on modularity. We give
a brief sketch of the latter one. Let Γ = Γ0+ (6) as in §3.1 be the group
Γ0 (6) ∪ Γ0 (6)W , where W = W6 = √16 06 −1 0 . This group has genus 0 and the
Hauptmodul
 12
η(z) η(6z)
t(z) = = q − 12 q 2 + 66 q 3 − 220 q 4 + . . . ,
η(2z) η(3z)
 Elliptic Modular Forms and Their Applications 65

where η(z) is the Dedekind eta-function defined in (34). For f (z) we take the
function
 7
η(2z) η(3z) 2 3 4
f (z) =  5 = 1 + 5 q + 13 q + 23 q + 29 q + . . . ,
η(z) η(6z)

which is a modular form of weight 2 on Γ (and in fact an Eisenstein series,

namely 5G2 (z) − 2G2(2z) + 3G2(3z) − 30G2(6z) ). Proposition 21 then implies
that if we expand f (z) as a power series 1 + 5 t(z) + 73 t(z)2 + 1445 t(z)3 + · · ·
in t(z), then the coefficients of the expansion satisfy a linear recursion with
polynomial coefficients (this is equivalent to the statement that the power
series itself satisfies a differential equation with polynomial coefficients), and
if we go through the proof of the proposition to calculate the differential
∞ we find that the recursion is exactly the one defining an .
equation explicitly,
Hence f (z) = n=0 an t(z)n , and the integrality of the coefficients an follows
immediately since f (z) and t(z) have integral coefficients and t has leading
coefficient 1.
To get the properties of the second sequence {bn } is a little harder. Define

g(z) = G4 (z) − 28 G4 (2z) + 63 G4(3z) − 36 G4 (6z)

= q − 14 q 2 + 91 q 3 − 179 q 4 + · · · ,

an Eisenstein series of weight 

 4 on Γ . Write the Fourier expansion as g(z) =
∞ ∞ −3
n=1 nc q n
and define g̃(z) = n=1 n cn q n , so that g̃  = g. This is the so-
called Eichler integral associated to g and inherits certain modular properties
from the modularity of g. (Specifically, the difference g̃|−2 γ − g̃ is a polynomial
of degree ≤ 2 for every γ ∈ Γ , where |−2 is the slash operator defined in (8).)
Using this (we skip the details), one finds by an argument analogous to the
proof of Proposition 21 that if we expand the product f (z)g̃(z) as a power
8 t(z) + 216 t(z) + · · · in t(z), then this power series again
series t(z) + 117 2 62531 3

satisfies a differential equation. This equation turns out to be the same one
as the one satisfied by f (but with right-hand side 1 instead of 0), so the
coefficients of the new expansion satisfy the same recursion and hence (since
they begin 0,1) must be one-sixth of Apéry’s coefficients bn , i.e., we have
6f (z)g̃(z) = ∞ n 3
n=0 bn t(z) . The integrality of Dn bn (indeed, even of Dn bn /6)
3
3
follows immediately: the coefficients cn are integral, so Dn is a common de-
nominator for the first n terms of f g̃ as a power series in q and hence also as
a power series in t(z). Finally, from the definition (13) of G4 (z) we find that
the Fourier coefficients of g are given by the Dirichlet series identity
∞  
cn 28 63 36
s
= 1 − s + s − s ζ(s) ζ(s − 3) ,
n=1
n 2 3 6

so the limiting value of bn /an (which must exist because {an } and {bn } satisfy
the same recursion) is given by
66 D. Zagier
 ∞  ∞ 
1 bn  cn n  cn  1
lim = g̃(z) = q = = ζ(3) ,
6 n→∞ an t(z)=1/C n=1
n3 q=1 n=1
s 
n s=3 6

proving also the last assertion in (76). ♥

♠ An Example Coming from Percolation Theory

Imagine a rectangle R of width r > 0 and height 1 on which has been drawn
a fine square grid (of size roughly rN × N for some integer N going to in-
finity). Each edge of the grid is colored black or white with probability 1/2
(the critical probability for this problem). The (horizontal) crossing probability
Π(r) is then defined as the limiting value, as N → ∞, of the probability that
there exists a path of black edges connecting the left and right sides of the
rectangle. A simple combinatorial argument shows that there is always either
a vertex-to-vertex path from left to right passing only through black edges or
else a square-to-square path from top to bottom passing only through white
edges, but never both. This implies that Π(r) + Π(1/r) = 1. The hypothesis
of conformality which, though not proved, is universally believed and is at the
basis of the modern theory of percolation, says that the corresponding prob-
lem, with R replaced by any (nice) open domain in the plane, is unchanged
under conformal (biholomorphic) mappings, and this can be used to compute
the crossing probability as the solution of√a differential equation.
 The result,
due to J. Cardy, is the formula Π(r) = 2π 3 Γ ( 13 )−3 t1/3 F 13 , 23 ; 43 ; t), where
t is the cross-ratio of the images of the four vertices of the rectangle R when it
is mapped biholomorphically onto the unit disk by the Riemann uniformiza-
tion theorem. This cross-ratio is known to be given by t = λ(ir) with λ(z) as
in (72). In modular terms, using Proposition 21,  ∞we find that this translates
to the formula Π(r) = −27/3 3−1/2 π 2 Γ ( 13 )−3 r η(iy)4 dy ; i.e., the deriva-
tive of Π(r) is essentially the restriction to the imaginary axis of the modular
form η(z)4 of weight 2. Conversely, an easy argument using modular forms
on SL(2, Z) shows that Cardy’s function is the unique function satisfying the
functional equation Π(r) + Π(1/r) = 1 and having an expansion of the form
e−2παr times a power series in e−2πr for some α ∈ R (which is then given by
α = 1/6). Unfortunately, there seems to be no physical argument implying
a priori that the crossing probability has the latter property, so one cannot
(yet?) use this very simple modular characterization to obtain a new proof of
Cardy’s famous formula. ♥

6 Singular Moduli and Complex Multiplication

The theory of complex multiplication, the last topic which we will treat in
detail in these notes, is by any standards one of the most beautiful chapters
in all of number theory. To describe it fully one needs to combine themes
relating to elliptic curves, modular forms, and algebraic number theory. Given
 Elliptic Modular Forms and Their Applications 67

the emphasis of these notes, we will discuss mostly the modular forms side, but
in this introduction we briefly explain the notion of complex multiplication in
the language of elliptic curves.
An elliptic curve over C, as discussed in §1, can be represented by a quo-
tient E = C/Λ, where Λ is a lattice in C. If E  = C/Λ is another curve
and λ a complex number with λΛ ⊆ Λ , then multiplication by λ induces an
algebraic map from E to E  . In particular, if λΛ ⊆ Λ, then we get a map
from E to itself. Of course, we can always achieve this with λ ∈ Z, since Λ
is a Z-module. These are the only possible real values of λ, and for generic
lattices also the only possible complex values. Elliptic curves E = C/Λ where
λΛ ⊆ Λ for some non-real value of λ are said to admit complex multiplication.
As we have seen in these notes, there are two completely different ways
in which elliptic curves are related to modular forms. On the one hand, the
moduli space of elliptic curves is precisely the domain of definition !Γ1 \H of "
modular functions on the full modular group, via the map Γ1 z ↔ C/Λz ,
where Λz = Zz + Z ⊂ C. On the other hand, elliptic curves over Q are sup-
posed (and now finally known) to have parametrizations by modular functions
and to have Hasse-Weil L-functions that coincide with the Hecke L-series of
certain cusp forms of weight 2. The elliptic curves with complex multiplication
are of special interest from both points of view. If we think of H as parametriz-
ing elliptic curves, then the points in H corresponding to elliptic curves with
complex multiplication (usually called CM points for short) are simply the
numbers z ∈ H which satisfy a quadratic equation over Z. The basic fact is
that the value of j(z) (or of any other modular function with algebraic coef-
ficients evaluated at z) is then an algebraic number; this says that an elliptic
curve with complex multiplication is always defined over Q (i.e., has a Weier-
strass equation with algebraic coefficients). Moreover, these special algebraic
numbers j(z), classically called singular moduli, have remarkable properties:
they give explicit generators of the class fields of imaginary quadratic fields,
the differences between them factor into small prime factors, their traces are
themselves the coefficients of modular forms, etc. This will be the theme of the
first two subsections. If on the other hand we consider the L-function of a CM
elliptic curve and the associated cusp form, then again both have very special
properties: the former belongs to two important classes of number-theoretical
Dirichlet series (Epstein zeta functions and L-series of grossencharacters) and
the latter is a theta series with spherical coefficients associated to a binary
quadratic form. This will lead to several applications which are treated in the
final two subsections.

6.1 Algebraicity of Singular Moduli

In this subsection we will discuss the proof, refinements, and applications of

the following basic statement:
Proposition 22. Let z ∈ H be a CM point. Then j(z) is an algebraic number.
68 D. Zagier

Proof. By definition, z satisfies a quadratic equation over Z, say Az2 + Bz +

C = 0. There is then always a matrix M ∈ M (2, Z), not proportional
 B C to
the identity, which fixes z. (For instance, we can take M = −A 0 .) This
matrix has a positive determinant, so it acts on the upper half-plane in the
usual way. The two functions j(z) and j(M z) are both modular functions on
the subgroup Γ1 ∩ M −1 Γ1 M of finite index in Γ1 , so they are algebraically
dependent, i.e., there is a non-zero polynomial P (X, Y ) in two variables such
that P (j(M z), j(z)) vanishes identically. By looking at the Fourier expansion
at ∞ we can see that the polynomial P can be chosen to have coefficients
in Q. (We omit the details, since we will prove a more precise statement
below.) We can also assume that the polynomial P (X, X) is not identically
zero. (If some power of X − Y divides P then we can remove it without
affecting the validity of the relation P (j(M z), j(z)) ≡ 0, since j(M z) is not
identically equal to j(z).) The number j(z) is then a root of the non-trivial
polynomial P (X, X) ∈ Q[X], so it is an algebraic number.
More generally, if f (z) is any modular function (say, with respect to a sub-
group of finite index of SL(2, Z)) with algebraic Fourier coefficients in its q-
expansion at infinity, then f (z) ∈ Q, as one can see either by showing that
f (M z) and f (z) are algebraically dependent over Q for any M ∈ M (2, Z)
with det M > 0 or by showing that f (z) and j(z) are algebraically dependent
over Q and using Proposition 22. The full theory of complex multiplication
describes precisely the number field in which these numbers f (z) lie and the
way that Gal(Q/Q) acts on them. (Roughly speaking, any Galois conjugate
of any f (z) has the form f ∗ (z∗ ) for some other modular form with algebraic
coefficients and CM point z∗ , and there is a recipe to compute both.) We
will not explain anything about this in these notes, except for a few words
in the third application below, but refer the reader to the texts listed in the
references.
We now return to the j-function and the proof above. The key point was
the algebraic relation between j(z) and j(M z), where M was a matrix in
M (2, Z) of positive determinant m fixing the point z. We claim that the poly-
nomial P relating j(M z) and j(z) can be chosen to depend only on m. More
precisely, we have:
Proposition 23. For each m ∈ N there is a polynomial Ψm (X, Y ) ∈ Z[X, Y ],
symmetric up to sign in its two arguments and of degree σ1 (m) with respect
to either one, such that Ψm (j(M z), j(z)) ≡ 0 for every matrix M ∈ M (2, Z)
of determinant m.
Proof. Denote by Mm the set of matrices in M (2, Z) of determinant m. The
group Γ1 acts on Mm by right and left multiplication, with finitely many
orbits. More precisely, an easy and standard argument shows that the finite
set
$  %
a b 
M∗m = a, b, d ∈ Z, ad = m, 0 ≤ b < d ⊂ Mm (77)
0d
 Elliptic Modular Forms and Their Applications 69

is a full set of representatives for Γ1 \Mm ; in particular, we have

   
Γ1 \Mm  = M∗  = d = σ1 (m) . (78)
m
ad=m

We claim that we have an identity

  
X − j(M z) = Ψm (X, j(z)) (z ∈ H, X ∈ C) (79)
M∈Γ1 \Mm

for some polynomial Ψm (X, Y ). Indeed, the left-hand side of (79) is well-
defined because j(M z) depends only on the class of M in Γ1 \Mm , and it is
Γ1 -invariant because Mm is invariant under right multiplication by elements
of Γ1 . Furthermore, it is a polynomial in X (of degree σ1 (m), by (78)) each of
whose coefficients is a holomorphic function of z and of at most exponential
growth at infinity, since each j(M z) has these properties and each coefficient
of the product is a polynomial in the j(M z). But a Γ1 -invariant holomorphic
function in the upper half-plane with at most exponential growth at infinity
is a polynomial in j(z), so that we indeed have (79) for some polynomial
Ψm (X, Y ) ∈ C[X, Y ]. To see that the coefficients of Ψm are in Z, we use the
set of representatives
∞ (77) and the Fourier expansion of j, which has the form
j(z) = n=−1 c n q n
with cn ∈ Z (c−1 = 1, c0 = 744, c1 = 196884 etc.).
Thus

 d−1 
az + b

Ψm (X, j(z)) = X −j
d
ad=m b=0
d>0
   ∞ 
= X − cn ζdbn q an/d ,
ad=m b (mod d) n=−1
d>0

where q α for α ∈ Q denotes e2πiαz and ζd = e2πi/d . The expression in

parentheses belongs to the ring Z[ζd ][X][q −1/d , q 1/d ]] of Laurent series in q 1/d
with coefficients in Z[ζd ], but applying a Galois conjugation ζd → ζdr with
r ∈ (Z/dZ)∗ just replaces b in the inner product by br, which runs over the
same set Z/dZ, so the inner product has coefficients in Z. The fractional pow-
ers of q go away at the same time (because the product over b is invariant
under z → z + 1), so each inner product, and hence Ψm (X, j(z)) itself, belongs
to Z[X][q −1 , q]]. Now the fact that it is a polynomial in j(z) and that j(z)
has a Fourier expansion with integral coefficients and leading coefficient q −1
imlies that Ψm (X, j(z)) ∈ Z[X, j(z)]. Finally, the
 symmetry
 of Ψm (X, Y ) up
to sign follows because z  = M z with M = ac db ∈ Mm is equivalent to
 d −b 
z = M  z  with M  = −c a ∈ Mm .
70 D. Zagier

An example will make all of this clearer. For m = 2 we have

   z   
 z+1   
X − j(z) = X − j X −j X − j 2z
2 2
M∈Γ1 \Mm

by (77). Write this as X 3 − A(z)X 2 + B(z)X − C(z). Then

z  
z+1  
A(z) = j +j + j 2z
2 2
     
= q −1/2 + 744 + o(q) + −q −1/2 + 744 + o(q) + q −2 + 744 + o(q)

= q −2 + 0 q −1 + 2232 + o(1)

= j(z)2 − 1488 j(z) + 16200 + o(1)

as z → i∞, and since A(z) is holomorphic and Γ1 -invariant this implies that
A = j 2 − 1488j + 16200. A similar calculation gives B = 1488j 2 + 40773375j +
8748000000 and C = −j 3 + 162000j 2 − 8748000000j + 157464000000000, so

Ψ2 (X, Y ) = − X 2 Y 2 + X 3 + 1488X 2Y + 1488XY 2 + Y 3 − 162000X 2

+ 40773375XY − 162000Y 2 + 8748000000X + 8748000000Y

− 157464000000000 . (80)

Remark. The polynomial Ψm (X, Y ) is not in general irreducible: if m is not

square-free, then it factors into the product of all Φm/r2 (X, Y ) with r ∈ N
such that r2 |m, where Φm (X, Y ) is defined exactly like Ψm (X, Y ) but with
Mm replaced by the set M0m of primitive matrices of determinant m. The
polynomial Φm (X, Y ) is always irreducible.
To obtain the algebraicity of j(z), we used that this value was a root
of P (X, X). We therefore should look at the restriction of the polynomial
Ψm (X, Y ) to the diagonal X = Y . In the example (80) just considered, two
properties of this restriction are noteworthy. First, it is (up to sign) monic, of
degree 4. Second, it has a striking factorization:

Ψ2 (X, X) = −(X − 8000) · (X + 3375)2 · (X − 1728) . (81)

We consider each of these properties in turn. We assume that m is not a square,

since otherwise Ψm (X, X) is identically zero because Ψm (X, Y ) contains the
factor Ψ1 (X, Y ) = X − Y .
Proposition 24. For m not a perfect square, the polynomial Ψm (X, X) is, up
to sign, monic of degree σ1+ (m) := d|m max(d, m/d).
 Elliptic Modular Forms and Their Applications 71

Proof. Using the identity b (mod d) (x − ζdb y) = xd − y d we find

    
az + b
Ψm (j(z), j(z)) = j(z) − j
d
ad=m b (mod d)
   
= q −1 − ζd−b q −a/d + o(1)
ad=m b (mod d)
  
+
−d −a
= q −q + (lower order terms) ∼ ± q −σ1 (m)
ad=m

as I(z) → ∞, and this proves the proposition since j(z) ∼ q −1 .

Corollary. Singular moduli are algebraic integers. 

Now we consider the factors of Ψm (X, X). First let us identify the three
roots √in the factorization
√ for m = 2 just given. The three CM points  0 −1 i,
(1 + i
 1 1 7)/2 and i
 0 −1  2 are fixed, respectively, by the three matrices 1 0 ,
−1 1 and 2 0 of determinant 2, so each of the corresponding j-values
must be a root of the polynomial (81). Computing these values numerically
to low accuracy to see which is which, we find
√
1 + i 7 √
j(i) = 1728 , j = −3375 , j(i 2) = 8000 .
2
(Another way to distinguish the roots, without doing any transcendental
√ cal-
culations, would be to observe, for√instance, that i and i 2 are fixed by
matrices of determinant 3 but (1 + i 7)/2 is not, and that X + 3375 does not
occur as a factor of Ψ3 (X, X) but both X − 1728 and X − 8000 do.)
The same method can be used for any other √ CM point. Here is a table
√ of
the first few values of j(zD ), where zD equals 12 D for D even and 12 (1 + D)
for D odd:
D −3 −4 −7 −8 −11 −12 −15 √ −16 −19
j(zD ) 0 1728 −3375 8000 −32768 54000 − 191025+85995
2
5
287496 −884736
We can make this more precise. For each discriminant (integer congruent
to 0 or 1 mod 4) D < 0 we consider, as at the end of §1.2, the set QD of prim-
itive positive definite binary quadratic forms of discriminant D, i.e., functions
Q(x, y) = Ax2 + Bxy + Cy 2 with A, B, C ∈ Z, A > 0, gcd(A, B, C) = 1 and
B 2 − 4AC = D. To each Q ∈ QD we associate the root zQ of Q(z, 1) = 0
in H. This gives a Γ1 -equivariant bijection between QD and the set ZD ⊂ H
of CM points of discriminant D. (The discriminant of a CM point is the
smallest discriminant of a quadratic polynomial over Z of which it is a root.)
In particular, the cardinality of Γ1 \ZD is h(D), the class number of D. We
choose a set of representatives {zD,i }1≤i≤h(D) for Γ1 \ZD (e.g., the points of
72 D. Zagier

ZD ∩ F1 , where F1 is the fundamental domain constructed in §1.2, corre-

sponding to the set Qred
D in (5)), with zD,1 = zD . We now form the class
polynomial
     
HD (X) = X − j(z) = X − j(zD,i ) . (82)
z∈Γ1 \ZD 1≤i≤h(D)

Proposition 25. The polynomial HD (X) belongs to Z[X] and is irreducible.

In particular, the number j(zD ) is algebraic of degree exactly h(D) over Q,
with conjugates j(zD,i ) (1 ≤ i ≤ h(D)).

Proof. We indicate only the main ideas of the proof. We have already proved
that j(z) for any CM point z is a root of the equation Ψm (X, X) = 0 whenever
m is the determinant of a matrix M ∈ M (2, Z) fixing z. The main point is that
the set of these m’s depends only on the discriminant D of z, not on z itself, so
that the different numbers j(zD,i ) are roots of the same equations and hence
are conjugates. Let Az2+ Bz  + C = 0 (A > 0) be the minimal equation of z
and suppose that M = ac db ∈ Mm fixes z. Then since cz 2 + (d − a)z − b = 0
we must have (c, d − a, −b) = u(A, B, C) for some u ∈ Z. This gives
1 
2 (t − Bu) −Cu t2 − Du2
M = 1 , det M = , (83)
Au 2 (t + Bu) 4

where t = tr M . Convesely, if t and u are any integers with t2 − Du2 = 4m,

then (83) gives a matrix M ∈ Mm fixing z. Thus the set of integers m = det M
with M z = z is the set of numbers 14 (t2 − Du2 ) with t ≡ Du (mod 2) or, more
invariantly, the set of norms of elements of the quadratic order
$ √  %
t + u D 
OD = Z[zD ] =  t, u ∈ Z, t ≡ Du (mod 2) (84)
2

of discriminant D, and this indeed depends only on D, not on z. We can

then obtain HD (X), or at least its square, as the g.c.d. of finitely many
polymials Ψm (X, X), just as we obtained H−7 (X)2 = (X + 3375)2 as the
g.c.d. of Ψ2 (X, X) and Ψ3 (X, X) in the example above. (Start, for example,
with a prime m1 which is the norm of an element of OD – there are known to
be infinitely many – and then choose finitely many further m’s which are also
norms of elements in OD but not any of the finitely many other quadratic
orders in which m1 is a norm.)
This argument only shows that HD (X) has rational coefficients, not that
it is irreducible. The latter fact is proved most naturally by studying the
arithmetic of the corresponding elliptic curves with complex multiplication
(roughly speaking, the condition of having complex multiplication by a given
order OD is purely algebraic and hence is preserved by Galois conjugation),
but since the emphasis in these notes is on modular methods and their appli-
cations, we omit the details.
 Elliptic Modular Forms and Their Applications 73

The proof just given actually yields the formula


Ψm (X, X) = ± HD (X)rD (m)/w(D) (m ∈ N, m = square) , (85)
D<0
 
due
 to Kronecker. Here rD (m) =  { ( t, u ) ∈ Z2 | t2 − Du2 = 4m }  =
{λ ∈ OD | N (λ) = m} and w(D) is the number of units in OD , which
is equal to 6 or 4 for D = −3 or −4, respectively, and to 2 otherwise. The
product in (85) is finite since rD (m) = 0 ⇒ 4m = t2 − u2 D ≥ |D| because u
can’t be equal to 0 if m is not a square.
There is another form of formula (85) which will be used in the second ap-
plication below. As well as the usual class number h(D), one has the Hurwitz
class number h∗ (D) (the traditional notation is H(|D|)), defined as the num-
ber of all Γ1 -equivalence classes of positive definite binary quadratic forms of
discriminant D, not just the primitive ones, counted with multiplicity equal to
one over the order of their stabilizer in Γ1 (which is 2 or 3 if the corresponding
point in the fundamental
 domain for Γ1 \H is at i or ρ and is 1 otherwise). In
formulas, h∗ (D) = r2 |D h (D/r2 ), where the sum is over all r ∈ N for which
r2 |D (and for which D/r2 is still congruent to 0 or 1 mod 4, since otherwise
h (D/r2 ) will be 0) and h (D) = h(D)/ 12 w(D) with w(D) as above. Simi-
∗
larly, we can define a modified class “polynomial” HD (X), of “degree” h∗ (D),
by 
∗
HD (X) = HD/r2 (X)2/w(D) ,
r 2 |D

∗
e.g., H−12 (X) = X 1/3 (X − 54000). (These are actual polynomials unless |D|
or 3|D| is a square.) Then (85) can be written in the following considerably
simpler form:

Ψm (X, X) = ± Ht∗2 −4m (X) (m ∈ N, m = square) . (86)
t2 <4m

This completes our long discussion of the algebraicity of singular moduli.

We now describe some of the many applications.

♠ Strange Approximations to π

We start with an application that is more fun than serious. The discriminant
D = −163 has class number one (and is in fact known to be the smallest
such discriminant), so Proposition 25 implies that j(zD ) is a rational integer.
Moreover, it is large (in absolute value) because j(z) ≈ q −1 and the q = e2πiz
corresponding to z = z163 is roughly −4 × 10−18 . But then from j(z) =
q −1 + 744 + O(q) we find that q −1 is extremely close to an integer, giving the
formula
√
eπ 163
= 262537412640768743.999999999999250072597 · · ·
74 D. Zagier

which would be very startling if one did not know about complex multiplica-
tion. By taking logarithms, one gets an extremely good approximation to π:
1
π = √ log(262537412640768744) − (2.237 · · · × 10−31 ) .
163
(A poorer but simpler approximation, based on the fact that j(z163 ) is a perfect
cube, is π ≈ √1633
log(640320), with an error of about 10−16 .) Many further
identities of this type were found, still using the theory of complex multipli-
cation, by Ramanujan and later by Dan Shanks, the most spectactular one
being
6 ! "
π − √ log 2 ε(x1 ) ε(x2 ) ε(x3 ) ε(x4 ) ≈ 7.4 × 10−82
3502
√ √ √
where x1 = 429 + 304 2, x2 = 627 + 221 2, x3 = 1071 + 92 34, x4 =
√ √ 2 2
1553
2 + 133 34, and ε(x) = x + x2 − 1. Of course these formulas are more
curiosities than useful ways to compute π, since the logarithms are no easier to
compute numerically than π is and in any case, if we allow complex numbers,
then Euler’s formula π = √1−1 log(−1) is an exact formula of the same kind!
♥

♠ Computing Class Numbers

By comparing the degrees on both sides of (85) or (86) and using Proposit-
ion 24, we obtain the famous Hurwitz-Kronecker class number relations

h(D)
σ1+ (m) = rD (m) = h∗ (t2 − 4m) (m ∈ N, m = square) .
w(D)
D<0 t2 <4m
(87)
(In fact the equality of the first and last terms is true also for m square if
we replace the summation condition by t2 ≤ 4m and define h∗ (0) = − 12 1
, as
one shows by a small modification of the proof given here.) We mention that
this formula has a geometric interpretation in terms of intersection numbers.
Let X denote the modular curve Γ1 \H. For each m ≥ 1 there is a curve
Tm ⊂ X × X, the Hecke correspondence, corresponding to the mth Hecke
operator Tm introduced in §4.1. (The preimage of this curve in H × H consists
of all pairs (z, M z) with z ∈ H and M ∈ Mm .) For m = 1, this curve is just
the diagonal. Now the middle or right-hand term of (87) counts the “physical”
intersection points of Tm and T1 in X × X (with appropriate multiplicities
if the intersections are not transversal), while the left-hand term computes
the same number homologically, by first compactifying X to X̄ = X ∪ {∞}
(which is isomorphic to P1 (C) via z → j(z)) and Tm and T1 to their closures
T̄m and T̄1 in X̄ × X̄ and then computing the intersection number of the
homology classes of T̄m and T̄1 in H2 (X̄ × X̄) ∼
= Z2 and correcting this by the
contribution coming from intersections at infinity of the compactified curves.
 Elliptic Modular Forms and Their Applications 75

Equation (87) gives a formula for h∗ (−4m) in terms of h∗ (D) with

|D| < 4m. This does not quite suffice to calculate class numbers recursively
since only half of all discriminants are multiples of 4. But by a quite similar
type of argument (cf. the discussion in §6.2 below) one can prove a second
class number relation, namely
(m − t2 ) h∗ (t2 − 4m) = min(d, m/d)3 (m ∈ N) (88)
t2 ≤4m d|m

(again with the convention that h∗ (0) = − 121

), and this together with (87)
∗
does suffice to determine h (D) for all D recursively, since together they ex-
press h∗ (−4m) + 2h∗ (1 − 4m) and mh∗ (1 − 4m) + 2(m − 1)h∗ (1 − 4m) as
linear combinations of h∗ (D) with |D| < 4m − 1, and every negative discrim-
inant has the form −4m or 1 − 4m for some m ∈ N. This method is quite
reasonable computationally if one wants to compute a table of class numbers
h∗ (D) for −X < D < 0, with about the same running time (viz., O(X 3/2 )
operations) as the more direct method of counting all reduced quadratic forms
with discriminant in this range. ♥

♠ Explicit Class Field Theory for Imaginary Quadratic Fields

Class field theory, which is the pinnacle of classical algebraic number theory,
gives a complete classification of all abelian extensions of a given number
field K. In particular, it says that the unramified abelian extensions (we omit
all definitions) are the subfields of a certain finite extension H/K, the Hilbert
class field, whose degree over K is equal to the class number of K and whose
Galois group over K is canonically isomorphic to the class group of K, while
the ramified abelian extensions have a similar description in terms of more
complicated partititons of the ideals of OK into finitely many classes. However,
this theory, beautiful though it is, gives no method to actually construct the
abelian extensions, and in fact it is only known how to do this explicitly in two
cases: Q and imaginary quadratic fields. If K = Q then the Hilbert class field
is trivial, since the class number is 1, and the ramified abelian extensions are
just the subfields of Q(e2πi/N ) (N ∈ N) by the Kronecker-Weber theorem. For
imaginary quadratic fields, the result (in the unramified case) is as follows.
Theorem. Let K be an imaginary quadratic field, with discriminant D and
Hilbert class field H. Then the h(D) singular moduli j(zD,i ) are conjugate to
one another over K (not just over Q), any one of them generates H over K,
and the Galois group of H over K permutes them transitively. More precisely,
if we label the CM points of discriminant D by the ideal classes of K and
identify Gal(H/K) with the class group of K by the fundamental isomorphism
of class field theory, then for any two ideal classes A, B of K the element σA
of Gal(H/K) sends j(zB ) to j(zAB ).
The ramified abelian extensions can also be described completely by com-
plex multiplication theory, but then one has to use the values of other modu-
lar functions (not just of j(z)) evaluated at all points of K ∩ H (not just
76 D. Zagier

those of discriminant D). Actually, even for the Hilbert class fields it can
be advantageous to use modular functions other than j(z). For instance, for
D = −23, the first discriminant with class number not a power of 2 (and
therefore the first non-trivial example, since Gauss’s theory of genera de-
scribes the Hilbert class field of D when the class group
√ has exponent
√ √2 as
a composite quadratic extension, e.g., H for K = Q( −15) is Q( −3, 5) ),
the Hilbert class field is generated over K by the real root α of the polynomial
X 3 − X − 1. The singular modulus j(zD ), which also generates this field, is
equal to −53 α12 (2α − 1)3 (3α + 2)3 and is a root of the much more compli-
cated irreducible polynomial H−23 (X) = X 3 + 3491750X 2 − 5151296875X +
12771880859375, but using more detailed results from the theory of com-
plex multiplication one can show that the number 2−1/2 eπi/24 η(zD )/η(2zD )
also generates H, and this number turns out to be α itself! The improve-
ment is even more dramatic for larger values of |D| and is important in
situations where one actually wants to compute a class field explicitly, as
in the applications to factorization and primality testing mentioned in §6.4
below. ♥

♠ Solutions of Diophantine Equations

In §4.4 we discussed that one can parametrize an elliptic curve E over Q

by modular functions X(z), Y (z), i.e., functions which are invariant under
Γ0 (N ) for some N (the conductor of E) and identically satisfy the Weier-
strass equation Y 2 = X 3 + AX + B defining E. If K is an imaginary quadratic
field with discriminant D prime to N and congruent to a square modulo 4N
(this is equivalent to requiring that OK = OD contains an ideal n with
OD /n ∼ = Z/N Z), then there is a canonical way to lift the h(D) points of
Γ1 \H of discriminant D to h(D) points in the covering Γ0 (N )\H, the so-
called Heegner points. (The way is quite simple: if D ≡ r2 (mod 4N ), then
one looks at points zQ with Q(x, y) = Ax2 + Bxy + Cy 2 ∈ QD satisfying
A ≡ 0 (mod N ) and B ≡ r (mod 2N ); there is exactly one Γ0 (N )-equivalence
class of such points in each Γ1 -equivalence class of points of ZD .) One can
then show that the values of X(z) and Y (z) at the Heegner points be-
have exactly like the values of j(z) at the points of ZD , viz., these values
all lie in the Hilbert class field H of K and are permuted simply transi-
tively by the Galois group of H over K. It follows that we get h(D) points
Pi = (X(zi ), Y (zi )) in E with coordinates in H which are permuted by
Gal(H/K), and therefore that the sum PK = P1 + · · · + Ph(D) has coor-
dinates in K (and in many cases even in Q). This method of constructing
potentially non-trivial rational solutions of Diophantine equations of genus 1
(the question of their actual non-triviality will be discussed briefly in §6.2)
was invented by Heegner as part of his proof of the fact, mentioned above,
that −163 is the smallest quadratic discriminant with class number 1 (his
proof, which was rather sketchy at many points, was not accepted at the
time by the mathematical community, but was later shown by Stark to
 Elliptic Modular Forms and Their Applications 77

be correct in all essentials) and has been used to construct non-trivial so-
lutions of several classical Diophantine equations, e.g., to show that every
prime number congruent to 5 or 7 (mod 8) is a “congruent number” (= the
area of a right triangle with rational sides) and that every prime number
congruent to 4 or 7 (mod 9) is a sum of two rational cubes (Sylvester’s
problem). ♥

6.2 Norms and Traces of Singular Moduli

A striking property of singular moduli is that they always are highly factoriz-
able numbers. For instance, the value of j(z−163 ) used in the first application
in §6.1 is −6403203 = −218 33 53 233 293 , and the value of j(z−11 ) = −32768
is the 15th power of −2. As part of our investigation about the heights of
Heegner points (see below), B. Gross and I were led to a formula which ex-
plains and generalizes this phenomenon. It turns out, in fact, that the right
numbers to look at are not the values of the singular moduli themselves, but
their differences. This is actually quite natural, since the definition of the
modular function j(z) involves an arbitrary choice of additive constant: any
function j(z) + C with C ∈ Z would have the same analytic and arithmetic
properties as j(z). The factorization of j(z), however, would obviously change
completely if we replaced j(z) by, say, j(z) + 1 or j(z) − 744 = q −1 + O(q)
(the so-called “normalized Hauptmodul” of Γ1 ), but this replacement would
have no effect on the difference j(z1 ) − j(z2 ) of two singular moduli. That the
original singular moduli j(z) do nevertheless have nice factorizations is then
due to the accidental fact that 0 itself is a singular modulus, namely j(z−3 ),
so that we can write them as differences j(z) − j(z−3 ). (And the fact that the
values of j(z) tend to be perfect cubes is then related to the fact that z−3 is
a fixed-point of order 3 of the action of Γ1 on H.) Secondly, since the singu-
lar moduli are in general algebraic rather than rational integers, we should
not speak only of their differences, but of the norms of their differences, and
these norms will then also be highly factored. (For instance, the norms of
the singular moduli j(z−15 ) and j(z−23 ), which are algebraic integers of de-
gree 2 and 3 whose values were given in §6.1, are −36 53 113 and −59 113 173 ,
respectively.)
If we restrict ourselves for simplicity to the case that the discriminants D1
and D2 of z1 and z2 are coprime, then the norm of j(z1 ) − j(z2 ) depends only
on D1 and D2 and is given by
   
J(D1 , D2 ) = j(z1 ) − j(z2 ) . (89)
z1 ∈Γ1 \ZD1 z2 ∈Γ1 \ZD2

(If h(D1 ) = 1 then this formula reduces simply to HD2 (zD1 ), while in general
J(D1 , D2 ) is equal, up to sign, to the resultant of the two irreducible polyno-
mials HD1 (X) and HD2 (X).) These are therefore the numbers which we want
to study. We then have:
78 D. Zagier

Theorem. Let D1 and D2 be coprime negative discriminants. Then all prime

factors of J(D1 , D2 ) are ≤ 14 D1 D2 . More precisely, any prime factors √ of
J(D1 , D2 ) must divide 14 (D1 D2 − x2 ) for some x ∈ Z with |x| < D1 D2
and x2 ≡ D1 D2 (mod 4).
There are in fact two proofs of this theorem, one analytic and one arith-
metic. We give some brief indications of what their ingredients are, without
defining all the terms occurring. In the analytic proof, one looks at the Hilbert
modular group SL(2, OF ) (see the notes by√J. Bruinier in this volume) asso-
ciated to the real quadratic field F = Q( D1 D2 ) and constructs a certain
Eisenstein series for this group, of weight 1 and with respect to the “genus
character” associated to the decomposition of the discriminant of F as D1 ·D2 .
Then one restricts this form to the diagonal z1 = z2 (the story here is actually
more complicated: the Eisenstein series in question is non-holomorphic and
one has to take the holomorphic projection of its restriction) and makes use
of the fact that there are no holomorphic modular forms of weight 2 on Γ1 .
In the arithmetic proof, one uses that p|J(D1 , D2 ) if and only the CM ellip-
tic curves with j-invariants j(zD1 ) and j(zD2 ) become isomorphic over Fp .
Now it is known that the ring of endomorphisms of any elliptic curve over
Fp is isomorphic either to an order in a quadratic field or to an order in the
(unique) quaternion algebra Bp,∞ over Q ramified at p and at infinity. For
the elliptic curve E which is the common reduction of the curves with com-
plex multiplication by OD1 and OD2 the first alternative cannot occur, since
a quadratic order cannot contain two quadratic orders coming from differ-
ent quadratic fields, so there must be an order in Bp,∞ which contains two
elements α1 and α2 with square D1 and D2 , respectively. Then the element
α = α1 α2 also belongs to this order, and if x is its trace then x ∈ Z (because
α is in an order and hence integral), x2 < N (α) = D1 D2 (because Bp,∞ is
ramified at infinity), and x2 ≡ D1 D2 (mod p) (because Bp,∞ is ramified at p).
This proves the theorem, except that we have lost a factor “4” because the
elements αi actually belong to the smaller orders O4Di and we should have
worked with the elements αi /2 or (1 + αi )/2 (depending on the parity of Di )
in ODi instead.
The theorem stated above is actually only the qualitative version of the
full result, which gives a complete formula for the prime factorization of
J(D1 , D2 ). Assume for simplicity that D1 and D2 are fundamental. For
each positive integer n of the form 14 (D1 D2 − x2 ), we define a function ε
from the set of divisors of n to {±1} by the  requirement
 rs that ε is com-
r1
pletely multiplicative (i.e., ε pr11 · · · prss = ε p1 · · · ε ps for any divisor
pr11 · · · prss of n) and is given on primes p|n by ε(p) = χD1 (p) if p  D1 and
by ε(p) = χD2 (p) if p  D2 , where χD is the Dirichlet character modulo D
introduced at the beginning of §3.2. Notice that this makes sense: at least one
of the two alternatives must hold, since (D1 , D2 ) = 1 and p is prime, and if
they both hold then the two definitions agree because D1 D2 is then congruent
to a non-zero square modulo p if p is odd and to an odd square modulo 8 if
p = 2, so χD1 (p)χD2 (p) = 1. We then define F (n) (still for n of the form
 Elliptic Modular Forms and Their Applications 79

1
4 (D1 D2 − x2 )) by

F (n) = dε(n/d) .
d|n

This number, which is a priori only rational since ε(n/d) can be positive
or negative, is actually integral and in fact is always a power of a sin-
gle prime number : one can show easily that ε(n) = −1, so n contains
an odd number of primes p with ε(p) = −1 and 2  ordp (n), and if we
write
n = p12α1 +1 · · · pr2αr +1 p2β 2βs γ1 γt
r+1 · · · pr+s q1 · · · qt
1

with r odd and ε(pi ) = −1, ε(qj ) = +1, αi , βi , γj ≥ 0, then

p(α1 +1)(γ1 +1)···(γt +1) if r = 1, p1 = p ,

F (n) = (90)
1 if r ≥ 3 .

The complete formula for J(D1 , D2 ) is then

  
D1 D2 − x2
J(D1 , D2 )8/w(D1 )w(D2 ) = F . (91)
2
4
x <D1 D2
x2 ≡D1 D2 (mod 4)

As an example, for D1 = −7, D2 = −43 (both with class number one) this
formula gives
  
301 − x2
J(−7, −43) = F
4
1≤x≤17
x odd
= F (75)F (73)F (69)F (63)F (55)F (45)F (33)F (19)F (3)
= 3 · 73 · 32 · 7 · 52 · 5 · 32 · 19 · 3 ,

and indeed j(z−7 ) − j(z−43 ) = −3375 + 884736000 = 36 · 53 · 7 · 19 · 73. This

is the only instance I know of in mathematics where the prime factorization
of a number (other than numbers like n! which are defined as products) can
be described in closed form.

♠ Heights of Heegner Points

This “application” is actually not an application of the result just described,

but of the methods used to prove it. As already mentioned, the above theorem
about differences of singular moduli was found in connection with the study
of the height of Heegner points on elliptic curves. In the last “application”
in §6.1 we explained what Heegner points are, first as points on the modular
curve X0 (N ) = Γ0 (N )\H ∪ {cusps} and then via the modular parametrization
as points on an elliptic curve E of conductor N . In fact we do not have to
80 D. Zagier

pass to an elliptic curve; the h(D) Heegner points on X0 (N ) corresponding

to complex multiplication
√ by the order OD are defined over the Hilbert class
field H of K = Q( D) and we can add these points on the Jacobian J0 (N )
of X0 (N ) (rather than adding their images in E as before). The fact that
they are permuted by Gal(H/K) means that this sum is a point PK ∈ J0 (N )
defined over K (and sometimes even over Q). The main question is whether
this is a torsion point or not; if it is not, then we have an interesting solution of
a Diophantine equation (e.g., a non-trivial point on an elliptic curve over Q).
In general, whether a point P on an elliptic curve or on an abelian variety
(like J0 (N )) is torsion or not is measured by an invariant called the (global)
height of P , which is always ≥ 0 and which vanishes if and only if P has
finite order. This height is defined as the sum of local heights, some of which
are “archimedean” (i.e., associated to the complex geometry of the variety
and the point) and can be calculated as transcendental expressions involving
Green’s functions, and some of which are “non-archimedean” (i.e., associated
to the geometry of the variety and the point over the p-adic numbers for some
prime number p) and can be calculated arithmetically as the product of log p
with an integer which measures certain geometric intersection numbers in
characteristic p. Actually, the height of a point is the value of a certain positive
definite quadratic form on the Mordell-Weil group of the variety, and we can
also consider the associated bilinear form (the “height pairing”), which must
then be evaluated for a pair of Heegner points. If N = 1, then the Jacobian
J0 (N ) is trivial and therefore all global heights are automatically zero. It
turns out that the archimedean heights are essentially the logarithms of the
absolute values of the individual terms in the right-hand side of (89), while
the p-adic heights are the logarithms of the various factors F (n) on the right-
hand side of (91) which are given by formula (90) as powers of the given prime
number p. The fact that the global height vanishes is therefore equivalent to
formula (91) in this case. If N > 1, then in general (namely, whenever X0 (N )
has positive genus) the Jacobian is non-trivial and the heights do not have to
vanish identically. The famous conjecture of Birch and Swinnerton-Dyer, one
of the seven million-dollar Clay Millennium Problems, says that the heights
of points on an abelian variety are related to the values or derivatives of
a certain L-function; more concretely, in the case of an elliptic curve E/Q,
the conjecture predicts that the order to which the L-series L(E, s) vanishes
at s = 1 is equal to the rank of the Mordell-Weil group E(Q) and that the
value of the first non-zero derivative of L(E, s) at s = 1 is equal to a certain
explicit expression involving the height pairings of a system of generators
of E(Q) with one another. The same kind of calculations as in the case N = 1
permitted a verification of this prediction in the case of Heegner points, the
relevant derivative of the L-series being the first one:
Theorem. Let E be an elliptic curve defined over Q whose L-series vanishes
at s = 1. Then the height of any Heegner point is an explicit (and in general
non-zero) multiple of the derivative L (E/Q, 1).
 Elliptic Modular Forms and Their Applications 81

The phrase “in general non-zero” in this theorem means that for any
given elliptic curve E with L(E, 1) = 0 but L (E, 1) = 0 there are Heegner
points whose height is non-zero and which therefore have infinite order in the
Mordell-Weil group E(Q). We thus get the following (very) partial statement
in the direction of the full BSD conjecture:
Corollary. If E/Q is an elliptic curve over Q whose L-series has a simple
zero at s = 1, then the rank of E(Q) is at least one.
(Thanks to subsequent work of Kolyvagin using his method of “Euler systems”
we in fact know that the rank is exactly equal to one in this case.) In the
opposite direction, there are elliptic curves E/Q and Heegner points P in E(Q)
for which we know that the multiple occurring in the theorem is non-zero but
where P can be checked directly to be a torsion point. In that case the theorem
says that L (E/Q, 1) must vanish and hence, if the L-series is known to have
a functional equation with a minus sign (the sign of the functional equation
can be checked algorithmically), that L(E/Q, s) has a zero of order at least 3
at s = 1. This is important because it is exactly the hypothesis needed to
apply an earlier theorem of Goldfeld which, assuming that such a curve is
known, proves that the class numbers of h(D) go to infinity in an effective
way as D → −∞. Thus modular methods and the theory of Heegner points
suffice to solve the nearly 200-year problem, due to Gauss, of showing that
the set of discriminants D < 0 with a given class number is finite and can be
determined explicitly, just as in Heegner’s hands they had already sufficed to
solve the special case when the given value of the class number was one. ♥
So far in this subsection we have discussed the norms of singular moduli
(or more generally, the norms of their differences), but algebraic numbers also
have traces, and we can consider these too. Now the normalization of j does
matter; it turns out to be best to choose the normalized Hauptmodul j0 (z) =
j(z) − 744. For every discriminant D < 0 we therefore define T (D) ∈ Z to be
the trace of j0 (zD ). This is the sum of the h(D) singular moduli j0 (zD,i ), but
just as in the case of the Hurwitz class numbers it is better to use the modified
trace T ∗ (d) which is defined as the sum of the values of j0 (z) at all points z ∈
Γ1 \H satisfying a quadratic equation, primitive or not, of discriminant D (and
with the points i and ρ, if they occur  at all, being counted with multiplicity
1/2 or 1/3 as usual), i.e., T ∗ (D) = 2 1 2
r 2 |D T (D/r )/( 2 w(D/r )) with the
∗
same conventions as in the definition of h (D). The result, quite different
from (and much easier to prove than) the formula for the norms, is that
these numbers T ∗ (D) are the coefficients of a modular form. Specifically, if we
denote by g(z) the meromorphic modular form θM (z)E4 (4z)/η(4z)6 of weight
 2
3/2, where θM (z) is the Jacobi theta function n∈Z (−1)n q n as in §3.1, then
we have:
Theorem. The Fourier expansion of g(z) is given by

g(z) = q −1 − 2 − T ∗ (−d) q d . (92)

d>0
82 D. Zagier

We can check the first few coefficients of this by hand: the Fourier ex-
pansion of g begins q −1 − 2 + 248 q 3 − 492 q 4 + 4119 q 7 − · · · and indeed we
have T ∗ (−3) = 13 (0 − 744) = −248, T ∗ (−4) = 12 (1728 − 744) = 492, and
T ∗ (−7) = −3375 − 744 = −4119.
The proof of this theorem, though somewhat too long to be given here,
is fairly elementary and is essentially a refinement of the method used to
prove the Hurwitz-Kronecker class number relations (87) and (88). More pre-
cisely, (87) was proved by comparing the degrees on both sides of (86), and
these in turn were computed by looking at the most negative exponent oc-
curring in the q-expansion of the two sides when X is replaced by j(z); in
particular, the number σ1+ (m) came from the calculation in Proposition 24
of the most negative power of q in Ψm (j(z), j(z)). If we look instead at the
+ 
next coefficient (i.e., that of q −σ1 (m)+1 ), then we find − t2 <4m T ∗ (t2 − 4m)
for the right-hand side of (86), because the modular functions HD (j(z)) and
∗
HD (j(z)) have Fourier expansions beginning q −h(D) (1 − T (D)q + O(q 2 )) and
−h∗ (D)
q (1 − T ∗ (D)q + O(q 2 )), respectively, while by looking carefully at the
“lower order terms” in the proof of Proposition 24 we find that the corre-
sponding coefficient for the left-hand side of (86) vanishes for all m unless m
or 4m + 1 is a square, when it equals +4 or −2 instead. The resulting identity
can then be stated uniformly for all m as

T ∗ (t2 − 4m) = 0 (m ≥ 0) , (93)

t∈Z

where we have artificially defined T ∗ (0) = 2, T ∗ (1) = −1, and T ∗ (D) = 0 for
D > 1 ( a definition made plausible by the result in the theorem we want to
prove). By a somewhat more complicated argument involving modular forms
of higher weight, we prove a second relation, analogous to (88):

1 if m = 0 ,
(m − t2 )T ∗ (t2 − 4m) = (94)
t∈Z
240σ3 (n) if m ≥ 1 .

(Of course, the factor m − t2 here could be replaced simply by −t2 , in

view of (93), but (94) is more natural, both from the proof and by anal-
ogy with (88).) Now, just as in the discussion of the Hurwitz-Kronecker class
number relations, the two formulas (93) and (94) suffice to determine all
T ∗ (D) by recursion. But in fact we can solve these equations directly, rather
than recursively (though this was not done in the original paper). Write the
∞ ∗ 4m
right-hand side ∞ of (92) as t 0 (z) + t 1 (z) where t 0 (z) = m=0 T (−4m) q
∗
and t1 (z) = m=0 T (1 − 4m) q 4m−1
, and define two unary theta series θ0 (z)
 t2
and θ1 (z) (= θ(4z) and θF (4z) in the notation of §3.1) as the sums q
with t ranging over the even or odd integers, respectively. Then (93) says
that t0 θ0 + t1 θ1 = 0 and (94) says that [t0 , θ0 ] + [t1 , θ1 ] = 4E4 (4z), where
[ti , θi ] = 32 ti θi − 12 ti θi is the first Rankin–Cohen bracket of ti (in weight 3/2)
and θi . By taking a linear combination of the second equation with the deriva-
 Elliptic Modular Forms and Their Applications 83
    
θ0 (z) θ1 (z) t0 (z) 0
tive of the first, we deduce that = 2 , and
θ0 (z) θ1 (z) t1 (z) E4 (4z)
from this we can immediately solve for t0 (z) and t1 (z) and hence for their
sum, proving the theorem.

♠ The Borcherds Product Formula

This is certainly not an “application” of the above theorem in any reasonable

sense, since Borcherds’s product formula is much deeper and more general
and was proved earlier, but it turns out that there is a very close link and
that one can even use this to give an elementary proof of Borcherds’s formula
in a special case. This special case is the beautiful product expansion
∞

∗ ∗  AD (n2 )
HD (j(z)) = q −h (D)
1 − qn , (95)
n=1

where AD (m) denotes the mth Fourier coefficient of a certain meromor-

phic modular form fD (z) (with Fourier expansion beginning q D + O(q)) of
weight 1/2. The link is that one can prove in an elementary way a “duality
formula” AD (m) = −Bm (|D|), where Bm (n) is the coefficient of q n in the
Fourier expansion of a certain other meromorphic modular form gm (z) (with
Fourier expansion beginning q −m + O(1)) of weight 3/2. For m = 1 the func-
tion gm coincides with the g of the theorem above, and since (95) immediately
implies that T ∗ (D) = AD (1) this and the duality imply (92). Conversely, by
applying Hecke operators (in half-integral weight) in a suitable way, one can
give a generalization of (92) to all functions gn2 , and this together with the
duality formula gives the complete formula (95), not just its subleading coef-
ficient. ♥

6.3 Periods and Taylor Expansions of Modular Forms

In §6.1 we showed that the value of any modular function (with rational or
algebraic Fourier coefficients; we will not always repeat this) at a CM point z
is algebraic. This is equivalent to saying that for any modular form f (z), of
weight k, the value of f (z) is an algebraic multiple of Ωzk , where Ωz depends
on z only, not on f or on k. Indeed, the second statement implies the first
by specializing to k = 0, and the first implies the second by observing that if
f ∈ Mk and g ∈ M then f /g k has weight 0 and is therefore algebraic at z,
so that f (z)1/k and g(z)1/ are algebraically proportional. Furthermore, the
number Ωz is unchanged (at least up to an algebraic number, but it is only
defined up to an algebraic number) if we replace z by M z for any M ∈ M (2, Z)
with positive determinant, because f (M z)/f (z) is a modular function, and
since any two CM points which generate the same imaginary quadratic field
are related in this way, this proves:
84 D. Zagier

Proposition 26. For each imaginary quadratic field K there is a number

ΩK ∈ C∗ such that f (z) ∈ Q · ΩK k
for all z ∈ K ∩ H, all k ∈ Z, and all
modular forms f of weight k with algebraic Fourier coefficients. 
To find ΩK , we should compute f (z) for some special modular form f (of
non-zero weight!) and some point or points z ∈ K ∩H. A natural choice for the
modular form is Δ(z), since it never vanishes. Even better, to achieve weight 1,
is its 12th root η(z)2 , and better yet is the function Φ(z) = I(z)| η(z)|4 (which
at z ∈ K ∩H is an algebraic multiple of ΩK 2
), since it is Γ1 -invariant. As for the
choice of z, we can look at the CM points of discriminant D (= discriminant
of K), but since there are h(D) of them and none should be preferred over
the others (their j-invariants are conjugate algebraic numbers), the only rea-
sonable choice is to multiply them all together and take the h(D)-th root – or
rather the h (D)-th root (where h (D) as previously denotes h(D)/ 12 w(D),
i.e., h (D) = 13 , 12 or h(K) for D = −3, D = −4, or D < −4), because the
elliptic fixed points ρ and i of Γ1 \H are always to be counted with multiplic-
ity 13 and 12 , respectively. Surprisingly enough, the product of the invariants
Φ(zD,i ) (i = 1, . . . , h(D)) can be evaluated in closed form:
Theorem. Let K be an imaginary quadratic field of discriminant D. Then
|D|−1
  2/w(D)   χD (j)
4π |D| Φ(z) = Γ j/|D| , (96)
z∈Γ1 \ZD j=1

where χD is the quadratic character associated to K and Γ (x) is the Euler

gamma function.

Corollary. The number ΩK in Proposition 26 can be chosen to be

⎛ ⎞ 
|D|−1
  χD (j) 1/2h (D)
1 ⎝ j ⎠
ΩK = Γ . (97)
2π|D| j=1
|D|

Formula (96), usually called the Chowla–Selberg formula, is contained in

a paper published by S. Chowla and A. Selberg in 1949, but it was later no-
ticed that it already appears in a paper of Lerch from 1897. We cannot give
the complete proof here, but we describe the  main−sidea, which is quite sim-
ple. The Dedekind zeta function ζK (s) = N (a) (sum over  all non-zero
integral ideals of K) has two decompositions: an additive one as A ζK,A (s),
where A runs over the  ideal classes of K and ζK,A (s) is the associated “par-
tial zeta function” (= N (a)−s with a running over the ideals in A), and
a multiplicative one as ζ(s)L(s,
 χD ), where ζ(s) denotes the Riemann zeta
function and L(s, χD ) = ∞ n=1 χD (n)n
−s
. Using these two decompositions,
one can compute in two different ways the two leading terms of the Laurent
expansion ζK (s) = s−1A
+ B + O(s − 1) as s → 1. The residue at s = 1 of
ζK,A (s) is independent of A and equals π/ 21 w(D) |D|, leading to Dirichlet’s
 Elliptic Modular Forms and Their Applications 85

class number formula L(1, χD ) = π h (D)/ |D|. (This is, of course, precisely
the method Dirichlet used.) The constant term in the Laurent expansion of
ζK,A (s) at s = 1 is given by the famous Kronecker limit formula and is (up to
some normalizing constants) simply the value of log(Φ(z)), where z ∈ K ∩ H is
the CM point corresponding to the ideal class A. The Riemann zeta function
has the expansion (s − 1)−1 − γ + O(s − 1) near s = 1 (γ = Euler’s constant),
and L (1, χD ) can be computed by a relatively
 elementary analytic argument
and turns out to be a simple multiple of j χD (j) log Γ (j/|D|). Combining
everything, one obtains (96).
As our first “application,” we mention two famous problems of transcend-
ence theory which were solved by modular methods, one using the Chowla–
Selberg formula and one using quasimodular forms.

♠ Two Transcendence Results

In 1976, G.V. Chudnovsky proved that for any z ∈ H, at least two of the three
numbers E2 (z), E4 (z) and E6 (z) are algebraically independent. (Equivalently,
the field generated by all f (z) with f ∈ M∗ (Γ1 )Q = Q[E2 , E4 , E6 ] has tran-
scendence degree at least 2.) Applying this to z = i, for which E2 (z) = 3/π,
E4 (z) = 3 Γ ( 41 )8 /(2π)6 and E6 (z) = 0, one deduces immediately that Γ ( 41 )
is transcendental (and in fact algebraically independent of π). Twenty years
later, Nesterenko, building on earlier work of Barré-Sirieix, Diaz, Gramain and
Philibert, improved this result dramatically by showing that for any z ∈ H at
least three of the four numbers e2πiz , E2 (z), E4 (z) and E6 (z) are algebraically
independent. His proof used crucially the basic properties of the ring M∗ (Γ1 )Q
discussed in §5, namely, that it is closed under differentiation and that each of
its elements is a power series in q = e2πiz with rational coefficients of bounded
denominator and polynomial growth. Specialized to z = i, Nesterenko’s result
implies that the three numbers π, eπ and Γ ( 14 ) are algebraically indepen-
dent. The algebraic independence of π and eπ (even without Γ ( 14 )) had been
a famous open problem. ♥

♠ Hurwitz Numbers

Euler’s famous result of 1734 that ζ(2r)/π 2r is rational for every r ≥ 1 can
be restated in the form
1
= (rational number) · π k for all k ≥ 2 , (98)
nk
n∈Z
n=0

where the “rational number” of course vanishes for k odd since then the con-
tributions of n and −n cancel. This result can be obtained, for instance, by
looking at the Laurent expansion of cot x near the origin. Hurwitz asked the
corresponding question if one replaces Z in (98) by the ring Z[i] of Gaussian
86 D. Zagier

integers and, by using an elliptic function instead of a trigonometric one, was

able to prove the corresponding assertion
1 Hk k
k
= ω for all k ≥ 3 , (99)
λ k!
λ∈Z[i]
λ=0

1 3
for certain rational numbers H4 = 10 , H8 = 10 , H12 = 567
130 , . . . (here Hk = 0
for 4  k since then the contributions of λ and −λ or of λ and iλ cancel),
where  1
dx Γ ( 1 )2
ω = 4 √ = √4 = 5.24411 · · · .
0 1 − x4 2π
Instead of using the theory of elliptic functions, we can see this result as a spe-
cial case of Proposition 26, since the sum on the left of (99) is just the special
value of the modular form 2Gk (z) defined in (10), which is related by (12)
to the modular
√ form Gk (z) with rational Fourier coefficients, at z = i and
ω is 2π 2 times the Chowla–Selberg period ΩQ(i) . Similar considerations, of
 −k
course, apply to the sum λ with λ running over the non-zero elements of
the ring of integers (or of any other ideal) in any imaginary quadratic field, and
more generally to the special values of L-series of Hecke “grossencharacters”
which we will consider shortly. ♥
We now turn to the second topic of this subsection: the Taylor expansions
(as opposed to simply the values) of modular forms at CM points. As we al-
ready explained in the paragraph preceding equation (58), the “right” Taylor
expansion for a modular form f at a point z ∈ H is the one occurring on
the left-hand side of that equation, rather than the straight Taylor expansion
of f . The beautiful fact is that, if z is a CM point, then after a renormal-
ization by dividing by suitable powers of the period Ωz , each coefficient of
this expansion is an algebraic number (and in many cases even rational). This
follows from the following proposition, which was apparently first observed by
Ramanujan.
Proposition 27. The value of E2∗ (z) at a CM point z ∈ K ∩ H is an algebraic
2
multiple of ΩK .

Corollary. The value of ∂ nf (z), for any modular form f with algebraic Fourier
coefficients, any integer n ≥ 0 and any CM point z ∈ K ∩ H, is an algebraic
k+2n
multiple of ΩK , where k is the weight of f .

Proof. We will give only a sketch,  since

 the proof is similar to that al-
ready given for j(z). For M = ac db ∈ Mm we define (E2∗ |2 M )(z) =
m(cz + d)−2 E2∗ (M z) ( = the usual slash operator for the matrix m−1/2 M ∈
SL(2, R)). From formula (1) and the fact that E2∗ (z) is a linear combination
of 1/y and a holomorphic function, we deduce immediately that the difference
E2∗ − E2∗ |2 M is a holomorphic modular form of weight 2 (on the subgroup of
 Elliptic Modular Forms and Their Applications 87

finite index M −1 Γ1 M ∩ Γ1 of Γ1 ). It then follows by an argument similar to

that in the proof of Proposition 23 that the function
  
Pm (z, X) := X + E2∗ (z) − (E2∗ |2 M )(z) (z ∈ H, X ∈ C)
M∈Γ1 \Mm

is a polynomial in X (of degree σ1 (m)) whose coefficients are modular

forms of appropriate weights on Γ1 with rational coefficients. (For example,
P2 (z, X) = X 3 − 34 E4 (z) X + 14 E6 (z).) The algebraicity of E2∗ (z)/ΩK 2
for
a CM point z now follows from Proposition 26, since if M ∈ Z · Id2 fixes z then
E2∗ (z) − (E2∗ |2 M )(z) is a non-zero algebraic multiple of E2∗ (z). The corollary
follows from (58) and the fact that each non-holomorphic derivative ∂ nf (z) is
a polynomial in E2∗ (z) with coefficients that are holomorphic modular forms
with algebraic Fourier coefficients, as one sees from equation (66).
Propositions 26 and 27 are illustrated for z = zD and f = E2∗ , E4 , E6 and Δ
in the following table, in which α in the penultimate row is the real root of
α3 − α − 1 = 0. Observe that the numbers in the final column of this table
are all units; this is part of the general theory.

|D|1/2 E2∗ (zD ) E4 (zD ) E6 (zD ) Δ(zD )

D
ΩD2 ΩD 4
|D|1/2 ΩD6 12
ΩD
−3 0 0 24 −1
−4 0 12 0 1
−7 3 15 27 −1
−8 4 20 28 1
−11 8 32 56 −1
√ √ √
−15 6+3 5 15 + 12 5 42 + √63
5
3− 5
2
−19 24 √ 96 √ 216 √ −1
−20 12 + 4 5 40 + 12 5 72 + 112
√
5
5−2
7+11α+12α2 469+1176α+504α2
−23 a1/3 √
5α1/3 (6 + 4α 2
√+ α ) 23 √ −α−8
√
−24 12 + 12 2 60 + 24 2 84 + 72 2 3−2 2

In the paragraph preceding Proposition 17 we explained that the non-

holomorphic derivatives ∂ nf of a holomorphic modular form f (z) are more
natural and more fundamental than the holomorphic derivatives Dnf , because
the Taylor series Dnf (z) tn /n! represents f only in the disk |z  − z| < I(z)
whereas the series (58) represents f everywhere. The above corollary gives
a second reason to prefer the ∂ nf : at a CM point z they are monomials in
the period Ωz with algebraic coefficients, whereas the derivatives Dnf (z) are
polynomials in Ωz by equation (57) (in which there can be no cancellation
since Ωz is known to be transcendental). If we set
∂ nf (z)
cn = cn (f, z, Ω) = k+2n (n = 0, 1, 2, . . . ) , (100)
Ω
where Ω is a suitably chosen algebraic multiple of ΩK , then the cn (up to
a possible common denominator which can be removed by multiplying f by
88 D. Zagier

a suitable integer) are even algebraic integers, and they still can benconsidered
to be “Taylor coefficients of f at z” since by (58) the series ∞ c t /n! equals
−k
 At+B  A B  −z̄ z  1/4πyΩ 0  n=0 n
(Ct + D) f Ct+D where C D = −1 1 0 Ω
∈ GL(2, C). These
normalized Taylor coefficients have several interesting number-theoretical ap-
plications, as we will see below, and from many points of view are actually
better number-theoretic invariants
 than the more familiar coefficients an of
the Fourier expansion f = an q n . Surprisingly enough, they are also much
easier to calculate: unlike the an , which are mysterious numbers (think of the
Ramanujan function τ (n)) and have to be calculated anew for each modular
form, the Taylor coefficients ∂ nf or cn are always given by a simple recursive
procedure. We illustrate the procedure by calculating the numbers ∂ nE4 (i),
but the method is completely general and works the same way for all modular
forms (even of half-integral weight, as we will see in §6.4).

Proposition 28. We have ∂ nE4 (i) = pn (0)E4 (i)1+n/2 , with pn (t) ∈ Z[ 16 ][t]
defined recursively by

t2 − 1  n+2 n(n + 3)
p0 (t) = 1, pn+1 (t) = pn (t)− tpn (t)− pn−1 (t) (n ≥ 0) .
2 6 144

Proof. Since E2∗ (i) = 0, equation (66) implies that f∂ (i, X) = fϑ (i, X) for any
f ∈ Mk (Γ1 ), i.e., we have ∂ nf (i) = ϑ[n]f (i) for all n ≥ 0, where {ϑ[n]f }n=0,1,2,...
is defined by (64). We use Ramanujan’s notations Q and R for the modular
forms E4 (z) and E6 (z). By (54), the derivation ϑ sends Q and R to − 13 R and
2
− 21 Q2 , respectively, so ϑ acts on M∗ (Γ1 ) = C[Q, R] as − R
3 ∂Q − 2 ∂R . Hence
∂ Q ∂

ϑ[n]Q = Pn (Q, R) for all n, where the polynomials Pn (Q, R) ∈ Q[Q, R] are
given recursively by

R ∂Pn Q2 ∂Pn QPn−1

P0 = Q , Pn+1 = − − − n(n + 3) .
3 ∂Q 2 ∂R 144

Since Pn (Q, R) is weighted homogeneous of weight 2n + 4, where Q and R

have weight 4 and 6, we can write Pn (Q, R) as Q1+n/2 pn (R/Q3/2 ) where pn
is a polynomial in one variable. The recursion for Pn then translates into the
recursion for pn given in the proposition.

The first few polynomials pn are p1 = − 31 t, p2 = 36 5

, p3 = − 72 5 5 2
t, p4 = 216 t +
5 2 5 2 4 5 3
288 , . . . , giving ∂ E4 (i) = E
36 4 (i) , ∂ E4 (i) = E
288 4 (i) , etc. (The values for
n odd vanish because i is a fixed point of the element S ∈ Γ1 of order 2.) With
the same method we find ∂ nf (i) = qn (0)E4 (i)(k+2n)/4 for any modular form
f ∈ Mk (Γ1 ), with polynomials qn satisfying the same recursion as pn but with
n(n + 3) replaced by n(n + k − 1) (and of course with a different initial value
q0 (t)). If we consider a CM point z other than i, then the method and result
are similar but we have to use (68) instead of (66), where φ is a quasimodular
form differing from E2 by a (meromorphic) modular form of weight 2 and
chosen so that φ∗ (z) vanishes. If we replace Γ1 by some other group Γ , then
 Elliptic Modular Forms and Their Applications 89

the same method works in principle but we need an explicit description of the
ring M∗ (Γ ) to replace the description of M∗ (Γ1 ) as C[Q, R] used above, and if
the genus of the group is larger than 0 then the polynomials pn (t) have to be
replaced by elements qn of some fixed finite algebraic extension of Q(t), again
satisfying a recursion of the form qn+1 = Aqn +(nB+C)qn +n(n+k−1)Dqn−1
with A, B, C and D independent of n. The general result says that the values
of the non-holomorphic derivatives ∂ nf (z0 ) of any modular form at any point
z0 ∈ H are given “quasi-recursively” as special values of a sequence of algebraic
functions in one variable which satisfy a differential recursion.

♠ Generalized Hurwitz Numbers

In our last “application” we studied the numbers Gk (i) whose values, up to

a factor of 2, are given by the Hurwitz formula (99). We now discuss the
meaning of the non-holomorphic derivatives ∂ nGk (i). From (55) we find
 
1 k mz̄ + n
∂k = ·
(mz + n) k 2πi(z − z̄) (mz + n)k+1

and more generally

 
1 (k)r (mz̄ + n)r
∂kr
= ·
(mz + n)k (2πi(z − z̄))r (mz + n)k+r

for all r ≥ 0, where ∂kr = ∂k+2r−2 ◦· · ·◦∂k+2 ◦∂k and (k)r = k(k+1) · · · (k+r−1)
as in §5. Thus

(k)n λ̄n
∂ nGk (i) = (n = 0, 1, 2, . . . )
2 (−4π)n λk+n
λ∈Z[i]
λ=0

(and similarly for ∂ nGk (z) for any CM point z, with Z[i] replaced by Zz+Z and
−4π by −4πI(z)). If we observe that the class number of the field K = Q(i)
is 1 and that any integral ideal of K can be written as a principal ideal (λ)
for exactly four numbers λ ∈ Z[i], then we can write

λ̄n λ̄k+2n ψk+2n (a)

= = 4
λk+n (λλ̄)k+n a
N (a)k+n
λ∈Z[i] λ∈Z[i]
λ=0 λ=0

where the sum runs over the integral ideals a of Z[i] and ψk+2n (a) is defined as
λ̄k+2n , where λ is any generator of a. (This is independent of λ if k+2n is divis-
ible by 4, and in the contrary case the sum vanishes.) The functions ψk+2n are
called Hecke “grossencharacters” (the German original of this semi-anglicized
word is “Größencharaktere”, with five differences of spelling, and means liter-
ally “characters of size,” referring to the fact that these characters, unlike the
90 D. Zagier

usual ideal class characters of finite order, depend on the 

size of the generator
of a principal ideal) and their L-series LK (s, ψk+2n ) = a ψk+2n (a)N (a)−s
are an important class of L-functions with known analytic continuation and
functional equations. The above calculation shows that ∂ nGk (i) (or more gen-
erally the non-holomorphic derivatives of any holomorphic Eisenstein series
at any CM point) are simple multiples of special values of these L-series at
integral arguments, and Proposition 28 and its generalizations give us an al-
gorithmic way to compute these values in closed form. ♥

6.4 CM Elliptic Curves and CM Modular Forms

In the introduction to §6, we defined elliptic curves with complex multiplica-

tion as quotients E = C/Λ where λΛ ⊆ Λ for some non-real complex num-
ber λ. In that case, as we have seen, the lattice Λ is homothetic to Zz + Z for
some CM point z ∈ H and the singular modulus j(E) = j(z) is algebraic, so
E has a model over Q. The map from E to E induced by multiplication by λ
is also algebraic and defined over Q. For simplicity, we concentrate on those z
whose discriminant is one of the 13 values −3, −4, −7, . . . , − 163 with class
number 1, so that j(z) ∈ Q and E (but not the complex multiplication) can
be defined over Q. For instance, the three elliptic curves

y 2 = x3 + x , y 2 = x3 + 1 , y 2 = x3 − 35x − 98 (101)

have j-invariants √ 1728, 0 and −3375, corresponding to multiplication by the

! " ! √ "
orders Z[i], Z 1+ 2 −3 , and Z 1+ 2 −7 , respectively. For the first curve (or
more generally any curve of the form y 2 = x3 + Ax with A ∈ Z) the mul-
tiplication by i corresponds to the obvious endomorphism (x, y) → (−x, iy)
of the curve, and similarly for the second curve (or any curve of the form
y 2 = x3 + B) we have the equally obvious endomorphism (x, y) → (ωx, y)
of order 3, where ω is a non-trivial cube root of 1. For the third curve (or
any of its “twists” E : Cy 2 = x3 − 35x − 98) the existence of a non-trivial
endomorphism is less obvious. One checks that the map
    
β2 β2
φ : (x, y) → γ 2 x + , γ3 y 1 − ,
x+α (x + α)2
√ √ √
where α = (7 + −7)/2, β = (7 + 3 −7)/2, and γ = (1 + −7)/4, maps E
to itself and satisfies φ(φ(P )) − φ(P ) + 2P = 0 for any point P on E, where
the addition is with respect to the group law on the curve, so√that we have
a map from O−7 to the endomorphisms of E sending λ = m 1+ 2 −7 + n to the
endomorphism P → mφ(P ) + nP .
The key point about elliptic curves with complex multiplication is that
the number of their points over finite fields is given by a simple formula.
For the three curves above this looks as follows. Recall that the number
of points over Fp of an elliptic curve E/Q given by a Weierstrass equation
 Elliptic Modular Forms and Their Applications 91

y 2 = F (x) = x3 + Ax + B equals p + 1 − ap where ap (for p odd and not di-

  
viding the discriminant of F ) is given by − x (mod p) F (x)
p . For the three
curves in (101) we have
 3 
x +x 0 if p ≡ 3 (mod 4) ,
= (102a)
x (mod p)
p −2a if p = a2 + 4b2 , a ≡ 1 (mod 4)
 
x3 + 1 0 if p ≡ 2 (mod 3) ,
= (102b)
x (mod p)
p −2a if p = a2 + 3b2 , a ≡ 1 (mod 3)
 
x3 − 35x − 98 0 if (p/7) = −1 ,
= (102c)
x (mod p)
p −2a if p = a2 + 7b2 , (a/7) = 1 .

(For other D with h(D) = 1 we would get a formula for ap (E) as ±A where
4p = A2 + |D|B 2 .) The proofs of these assertions, and of the more general
statements needed when h(D) > 1, will be omitted since they would take us
too far afield, but we give the proof of the first (due to Gauss), since it is
elementary and quite pretty. We prove a slightly more general but less precise
statement.
Proposition 29. Let p be an odd prime and A an integer not divisible by p.
Then
⎧
 3  ⎪
⎨0 if p ≡ 3 (mod 4) ,
x + Ax
= ±2a if p ≡ 1 (mod 4) and (A/p) = 1 , (103)
p ⎪
⎩
x (mod p) ±4b if p ≡ 1 (mod 4) and (A/p) = −1 ,
where |a| and |b| in the second and third lines are defined by p = a2 + 4b2 .
Proof. The first statement is trivial since if p ≡ 3 (mod 4) then (−1/p) = −1
and the terms for x and −x in the sum cancel, so we can suppose that p ≡ 1
(mod 4). Denote the sum on the left-hand side of (103)   by sp (A). Replacing
x by rx with r ≡ 0 (mod p) shows that sp (r2 A) = pr sp (A), so the number
sp (A) takes on only four values, say ±2α for A = g 4i or A = g 4i+2 and ±2β
for A = g 4i+1 or A = g 4i+3 , where g is a primitive root modulo p. (That
sp (A) is always even is obvious by replacing x by −x.) Now we take the sum
of the squares of sp (A) as A ranges over all integers modulo p, noting that
sp (0) = 0. This gives
 3  3 
x + Ax y + Ay
2(p − 1)(α + β ) =
2 2
p p
A, x, y ∈ Fp
   2 
xy (x + A)(y 2 + A)
=
p p
x, y ∈ Fp A ∈ Fp
 
xy  
= −1 + p δx2 ,y2 = 2p(p − 1) ,
p
x, y ∈ Fp
92 D. Zagier
  
where in the last line we have used the easy fact that z∈Fp z(z+r) p equals
p − 1 for r ≡ 0 (mod p) and −1 otherwise. (Proof: the first statement is
obvious; the substitution z → rz shows that the sum is independent of r if
r ≡ 0; and the sum of the values for all integers r mod p clearly vanishes.)
Hence α2 + β 2 = p. It is also obvious that α is odd (for (A/p) = +1 there
 x3 +Ax 
2 − 1 values of x between 0 and p/2 with
are p−1 non-zero and hence
p
 3 
odd) and that β is even (for (A/p) = −1 all 2 values of x +Ax
p−1
p with
0 < x < p/2 are odd), so α = ±a, β = ±2b as asserted.
An almost exactly similar proof works for equation (102b): here one defines
  3   
sp (B) = x x +B p and observes that sp (r3B) = pr sp (B), so that sp (B)
takes on six values ±α, ±β and ±γ depending on the class of i (mod 6), where
p ≡ 1 (mod 6) (otherwise all sp (B) vanish) and B = g i with g a primitive root
modulo p. Summing sp (B) over all quadratic residues B (mod p) shows that
α+β+γ = 0, and summing sp (B)2 over all B gives α2 +β 2 +γ 2 = 6p. These two
equations imply that α ≡ β ≡ γ ≡ 0 (mod 3) and that 4p = α2 +3((β−γ)/3)2 ,
which gives (102b) since it is easily seen that α is even and β and γ are odd.
I do not know of any elementary proof of this sort for equation (102c) or for
the similar identities corresponding to the other imaginary quadratic fields of
class number 1. The reader is urged to try to find such a proof.

♠ Factorization, Primality Testing, and Cryptography

We mention briefly one “practical” application of complex multiplication the-
ory. Many methods of modern cryptography depend on being able to identify
very large prime numbers quickly or on being able to factor (or being relatively
sure that no one else will be able to factor) very large composite numbers. Sev-
eral methods involve the arithmetic of elliptic curves over finite fields, which
yield finite groups in which certain operations are easily performed but not
easily inverted. The difficulty of the calculations, and hence the security of the
method, depends on the structure of the group of points of the curve over the
finite fields, so one would like to be able to construct, say, examples of elliptic
curves E over Q whose reduction modulo p for some very large prime p has an
order which itself contains a very large prime factor. Since counting the points
on E(Fp ) directly is impractical when p is very large, it is essential here to
know curves E for which the number ap , and hence the cardinality of E(Fp ),
is known a priori, and the existence of closed formulas like the ones in (102)
implies that the curves with complex multiplication are suitable. In practice
one wants the complex multiplication to be by an order in a quadratic field
which is not too big but also not too small. For this purpose one needs effective
ways to construct the Hilbert class fields (which is where the needed singular
moduli will lie) efficiently, and here again the methods mentioned in 6.1, and
in particular the simplifications arising by replacing the modular function j(z)
by better modular functions, become relevant. For more information, see the
bibliography. ♥
 Elliptic Modular Forms and Their Applications 93

We now turn from elliptic curves with complex multiplication to an impor-

tant and related topic, the so-called CM modular forms. Formula (102a) says
that the coefficient ap of the L-series L(E, s) = an n−s of the elliptic curve
E : y 2 = x3 + x for a prime p is given by ap = Tr(λ) = λ + λ̄, where λ is one
of the two numbers of norm p in Z[i] of the form a + bi with a ≡ 1 (mod 4),
b ≡ 0 (mod 2) (or is zero if there is no such λ). Now using the multiplicative
properties of the an we find that the full L-series is given by

L(E, s) = LK (s, ψ1 ) , (104)


where LK (s, ψ1 ) = 0=a⊆Z[i] ψ1 (a)N (a)−s is the L-series attached as in §6.3
to the field K = Q(i) and the grossencharacter ψ1 defined by

0 if 2 | N (a) ,
ψ1 (a) = (105)
λ if 2  N (a) , a = (λ), λ ∈ 4Z + 1 + 2Z i .

In fact, the L-series of the elliptic curve E belongs to three important classes
of L-series: it is an L-function coming from algebraic geometry (by definition),
the L-series of a generalized character of an algebraic number field (by (104)),
and also the L-series of a modular form, namely L(E, s) = L(fE , s) where
2
+b2
fE (z) = ψ1 (a) q N (a) = a qa (106)
0=a⊆Z[i] a≡1 (mod 4)
b≡0 (mod 2)

which is a theta series of the type mentioned in the final paragraph of §3,
associated to the binary quadratic form Q(a, b) = a2 + b2 and the spherical
polynomial P (a, b) = a (or a + ib) of degree 1. The same applies, with suitable
modifications, for any elliptic curve with complex multiplication, so that the
Taniyama-Weil conjecture, which says that the L-series of an elliptic curve
over Q should coincide with the L-series of a modular form of weight 2, can
be seen explicitly for this class of curves (and hence was known for them long
before the general case was proved).
The modular form fE defined in (106) can also be denoted θψ1 , where ψ1 is
the grossencharacter (105). More generally, the L-series of the powers ψd = ψ1d
of ψ1 are the L-series of modular forms, namely LK (ψd , s) = L(θψd , s) where
2
+b2
θψd (z) = ψd (a) q N (a) = (a + ib)d q a , (107)
0=a⊆Z[i] a≡1 (mod 4)
b≡0 (mod 2)

which is a modular form of weight d + 1. Modular forms constructed in this

way – i.e., theta series associated to a binary quadratic form Q(a, b) and
a spherical polynomial P (a, b) of arbitrary degree d – are called CM modular
forms, and have several remarkable properties. First
 of all, they are always
linear combinations of the theta series θψ (z) = a ψ(α) q N (a) associated to
some grossencharacter ψ of an imaginary quadratic field. (This is because any
94 D. Zagier

positive definite binary quadratic form over Q is equivalent to the norm form
λ → λλ̄ on an ideal in an imaginary quadratic field, and since the Laplace
2
operator associated to this form is ∂λ∂ ∂λ , the only spherical polynomials of
degree d are linear combinations of λd and λ̄d .) Secondly, since the L-series
L(θψ , s) = LK (ψ, s) has an Euler product, the modular forms θψ attached to
grossencharacters are always Hecke eigenforms, and this is the only infinite
family of Hecke eigenforms, apart from Eisenstein series, which are known
explicitly. (Other eigenforms do not appear to have any systematic rule of
construction, and even the number fields in which their √ Fourier coefficients
lie are totally mysterious, an example being the field Q( 144169) of coeffi-
cients of the cusp form of level 1 and weight 24 mentioned in §4.1.) Thirdly,
sometimes modular forms constructed by other methods turn out to be of
CM type, leading to new identities. For instance, in four cases the modular
form defined in (107) is a product of eta-functions: θψ0 (z) = η(8z)4 /η(4z)2 ,
θψ1 (z) = η(8z)8 /η(4z)2 η(16z)2 , θψ2 (z) = η(4z)6 , θψ4 (z) = η(4z)14 /η(8z)4 .
More generally, we can ask which eta-products are of CM type. Here I
do not know the answer, but for pure powersthere is a complete result,
2
due to Serre. The fact that both η(z) = (−1)(n−1)/6 q n /24 and
n≡1 (mod 6)
3
 n2 /8
η(z) = nq are unary theta series implies that each of the
n≡1 (mod 4)
functions η , η and η 6 is a binary theta series and hence a modular form
2 4

of CM type (the function η(z)6 equals θψ2 (z/4), as we just saw, and η(z)4
equals fE (z/6), where E is the second curve in (101)), but there are other,
less obvious, examples, and these can be completely classified:
Theorem (Serre). The function η(z)n (n even) is a CM modular form for
n = 2, 4, 6, 8, 10, 14 or 26 and for no other value of n.
Finally, the CM forms have another property called “lacunarity” which is
not shared by any other modular forms. If we look at (106) or (107), then
we see that the only exponents which occur are sums of two squares. By
the theorem of Fermat proved in §3.1, only half of all primes (namely, those
congruent to 1 modulo 4) have this property, and by a famous theorem of
Landau, only O(x/(log x)1/2 ) of the integers ≤ x do. The same applies to
any other CM form and shows that 50% of the coefficients ap (p prime) vanish
if the form is a Hecke eigenform and that 100% of the coefficients an (n ∈ N)
vanish for any CM form, eigenform or not. Another difficult theorem, again
due to Serre, gives the converse statements:

Theorem (Serre). Let f = an q n be a modular form of integral weight
≥ 2. Then:
1. If f is a Hecke eigenform, then the density of primes p for which ap = 0
is equal to 1/2 if f corresponds to a grossencharacter of an imaginary
quadratic field, and to 1 otherwise. √
2. The number of integers n ≤ x for which an = 0 is O(x/ log x) as x → ∞
if f is of CM type and is larger than a positive multiple of x otherwise.
 Elliptic Modular Forms and Their Applications 95

♠ Central Values of Hecke L-Series

We just saw above that the Hecke L-series associated to grossencharacters

of degree d are at the same time the Hecke L-series of CM modular forms
of weight d + 1, and also, if d = 1, sometimes the L-series of elliptic curves
over Q. On the other hand, the Birch–Swinnerton-Dyer conjecture predicts
that the value of L(E, 1) for any elliptic curve E over Q (not just one
with CM) is related as follows to the arithmetic of E : it vanishes if and
only if E has a rational point of infinite order, and otherwise is (essentially)
a certain period of E multiplied by the order of a mysterious group Ш, the
Tate–Shafarevich group of E. Thanks to the work of Kolyvagin, one knows
that this group is indeed finite when L(E, 1) = 0, and it is then a stan-
dard fact (because Ш admits a skew-symmetric non-degenerate pairing with
values in Q/Z) that the order of Ш is a perfect square. In summary, the
value of L(E, 1), normalized by dividing by an appropriate period, is al-
ways a perfect square. This suggests looking at the central point ( = point
of symmetry with respect to the functional equation) of other types of L-
series, and in particular of L-series attached to grossencharacters of higher
weights, since these can be normalized in a nice way using the Chowla–
Selberg period (97), to see whether these numbers are perhaps also always
squares.
An experiment to test this idea was carried out over 25 years ago by
B. √Gross and myself and confirmed this expectation. Let K be the field
Q( −7) and for each d ≥ 1 let ψd = ψ1d be the grossencharacter √ of K
which sends an ideal a ⊆ OK to λd if a is prime to p7 = ( −7) and to
0 otherwise, where λ in the former case is the generator of a which is con-
gruent to a square modulo p7 . For d = 1, the L-series of ψd coincides with
the L-series of the third elliptic curve in (101), while for general odd values
of d it is the L-series of a modular form of weight d + 1 and trivial charac-
ter on Γ0 (49). The L-series L(ψ2m−1 , s) has a functional equation sending s
to 2m − s, so the point of symmetry is s = m. The central value has the
form
2Am  2π m 2m−1
L(ψ2m−1 , m) = √ ΩK (108)
(m − 1)! 7
where   √ 
√  1 + −7 2 1 2 4
ΩK =
4
7 η  = Γ ( 7 )Γ ( 7 )Γ ( 7 )
2  4π 2
is the Chowla–Selberg period attached to K and (it turns out) Am ∈ Z for all
m > 1. (We have A1 = 14 .) The numbers Am vanish for m even because the
functional equation of L(ψ2m−1 , s) has a minus sign in that case, but the nu-
merical computation suggested that the others were indeed all perfect squares:
A3 = A5 = 12 , A7 = 32 , A9 = 72 , . . . , A33 = 447622863272552. Many years
later, in a paper with Fernando Rodriguez Villegas, we were able to confirm
this prediction:
96 D. Zagier

Theorem. The integer Am is a square for all m > 1. More precisely, we have
A2n+1 = bn (0)2 for all n ≥ 0, where the polynomials bn (x) ∈ Q[x] are defined
recursively by b0 (x) = 12 , b1 (x) = 1 and 21 bn+1(x) = −(x− 7)(64x− 7)bn(x)+
(32nx − 56n + 42)bn (x) − 2n(2n − 1)(11x + 7)bn−1 (x) for n ≥ 1.
The proof is too complicated to give here, but we can indicate the main
idea. The first point is that the numbers Am themselves can be computed
by the method explained in §6.3. There we saw (for the full modular group,
but the method works for higher level) that the value at s = k + n of the
L-series of a grossencharacter of degree d = k + 2n is essentially equal to the
nth non-holomorphic derivative of an Eisenstein series of weight k at a CM
point. Here we want d = 2m − 1 and s = √ m, so k = 1, n = m − 1. More
7)m m−1
precisely, one finds that L(ψ2m−1 , m) = (2π/ (m−1)! ∂ G1,ε (z7 ), where G1,ε is
the Eisenstein series
∞  
1 1
G1,ε = + ε(n) q n = + q + 2q 2 + 3q 4 + q 7 + · · ·
2 n=1
2
d|n
n
of weight 1 associated to the character ε(n) = and z7 is the CM point
  7
1 i
1 + √ . These coefficients can be obtained by a quasi-recursion like
2 7
the one in Proposition 28 (though it is more complicated here because the
analogues of the polynomials pn (t) are now elements in a quadratic extension
of Q[t]), and therefore are very easy to compute, but this does not explain why
they are squares. To see this, we first observe that the Eisenstein series G1,ε is
 2 2
one-half of the binary theta series Θ(z) = m, n∈Z q m +mn+2n . In his thesis,
Villegas proved a beautiful formula expressing certain linear combinations of
values of binary theta series at CM points as the squares of linear combinations
of values of unary theta series at other CM points. The same turns out to be
true for the higher non-holomorphic derivatives, and in the case at hand we
find the remarkable formula
 2
∂ 2n Θ(z7 ) = 22n 72n+1/4 ∂ n θ2 (z∗7 ) for all n ≥ 0 , (109)
 2
where θ2 (z) = n∈Z q (n+1/2) /2 is the Jacobi theta-series defined in (32) and
√
z∗7 = 12 (1 + i 7). Now the values of the non-holomorphic derivatives ∂ n θ2 (z∗7 )
can be computed quasi-recursively by the method explained in 6.3. The result
is the formula given in the theorem.
Remarks 1. By the same method, using that every Eisenstein series of
weight 1 can be written as a linear combination of binary theta series, one
can show that the correctly normalized central values of all L-series of grossen-
characters of odd degree are perfect squares.
2. Identities like (109) seem very surprising. I do not know of any other
case in mathematics where the Taylor coefficients of an analytic function at
some point are in a non-trivial way the squares of the Taylor coefficients of
another analytic function at another point.
 Elliptic Modular Forms and Their Applications 97

3. Equation (109) is a special case of a yet more general identity, proved in

the same paper, which expresses the non-holomorphic derivative ∂ 2n θψ (z0 ) of
the CM modular form associated to a grossencharacter ψ of degree d = 2r as
a simple multiple of ∂ n θ2 (z1 ) ∂ n+r θ2 (z2 ), where z0 , z1 and z2 are CM points
belonging to the quadratic field associated to ψ. ♥
We end with an application of these ideas to a classical Diophantine equa-
tion.

♠ Which Primes are Sums of Two Cubes?

In §3 we gave a modular proof of Fermat’s theorem that a prime number can be

written as a sum of two squares if and only if it is congruent to 1 modulo 4.
The corresponding question for cubes was studied by Sylvester in the 19th
century. For squares the answer is the same whether one considers integer or
rational squares (though in the former case the representation, when it exists,
is unique and in the latter case there are infinitely many), but for cubes one
only considers the problem over the rational numbers because there seems
to be no rule for deciding which numbers have a decomposition into integral
cubes. The question therefore is equivalent to asking whether the Mordell-
Weil group Ep (Q) of the elliptic curve Ep : x3 + y 3 = p is non-trivial. Except
for p = 2, which has the unique decomposition 13 + 13 , the group Ep (Q) is
torsion-free, so that if there is even one rational solution there are infinitely
many. An equivalent question is therefore: for which primes p is the rank rp
of Ep (Q) greater than 0 ?
Sylvester’s problem was already mentioned at the end of §6.1 in connection
with Heegner points, which can be used to construct non-trivial solutions if
p ≡ 4 or 7 mod 9. Here we consider instead an approach based on the Birch–
Swinnerton-Dyer conjecture, according to which rp > 0 if and only if the L-
series of Ep vanishes at s = 1. By a famous theorem of Coates and Wiles, one
direction of this conjecture is known: if L(Ep , 1) = 0 then Ep (Q) has rank 0.
The question we want to study is therefore: when does L(Ep , 1) vanish? For
five of the six possible congruence classes for p (mod 9) (we assume that p > 3,
since r2 = r3 = 0) the answer is known. If p ≡ 4, 7 or 8 (mod 9), then the
functional equation of L(Ep , s) has a minus sign, so L(Ep , 1) = 0 and rp is
expected (and, in the first two cases, known) to be ≥ 1; it is also known
by an “infinite descent” argument to be ≤ 1 in these cases. If p ≡ 2 or 5
(mod 9), then the functional equation has a plus sign and L(Ep , 1) divided
by a suitable period is ≡ 1 (mod 3) and hence = 0, so by the Coates-Wiles
theorem these primes can never be sums of two cubes, a result which can
also be proved in an elementary way by descent. In the remaining case p ≡ 1
(mod 9), however, the answer can vary: here the functional equation has a plus
sign, so ords=1 L(Ep , s) is even and rp is also expected to be even, and descent
gives rp ≤ 2, but both cases rp = 0 and rp = 2 occur. The following result,
again proved jointly with F. Villegas, gives a criterion for these primes.
98 D. Zagier

Theorem. Define a sequence of numbers c0 = 1, c1 = 2, c2 = −152,

c3 = 6848, . . . by cn = sn (0) where s0 (x) = 1, s1 (x) = 3x2 and sn+1 (x) =
(1−8x3 )sn (x)+(16n+3)x2 sn (x)−4n(2n−1)xsn−1 (x) for n ≥ 1. If p = 9k +1
is prime, then L(Ep , 1) = 0 if and only if p|ck .
For instance, the numbers c2 = −152 and c4 = −8103296 are divis-
ible by p = 19 and p = 37, respectively, so L(Ep , 1) vanishes for these
two primes (and indeed 19 = 33 + (−2)3 , 37 = 43 + (−3)3 ), whereas
c8 = 532650564250569441280 is not divisible by 73, which is therefore not
a sum of two rational cubes. Note that the numbers cn grow very quickly
(roughly like n3n ), but to apply the criterion for a given prime number
p = 9k + 1 one need only compute the polynomials sn (x) modulo p for
0 ≤ n ≤ k, so that no large numbers are required.
We again do not give the proof of this theorem, but only indicate the
main ingredients. The central value of L(Ep , s) for p ≡ 1 (mod 9) is given by
3Ω Γ ( 13 )3
L(Ep , 1) = √ 3 p
S p where Ω = √ and Sp is an integer which is supposed
2π 3
to be a perfect square (namely the order of the Tate-Shafarevich group of Ep
if rp = 0 and 0 if rp > 0). Using the methods from Villegas’s thesis, one can
show that this is true and that both Sp and its square root can be expressed
as the traces of certain algebraic numbers defined as special values√of modu-
3 p
Θ(pz0 )
lar functions at CM points: Sp = Tr(αp ) = Tr(βp )2 , where αp =
√ 54 Θ(z0 )
6 p
η(pz1 )  m2 +mn+n2 1 i
and βp = √ with Θ(z) = q , z0 = + √ and
±12 η(z1 /p) 2 6 3
√ m,n
r + −3
z1 = (r ∈ Z, r2 ≡ −3 (mod 4p)). This gives an explicit formula for
2
L(Ep , 1), but it is not very easy to compute since the numbers αp and βp have
large degree (18k and 6k, respectively if p = 9k + 1) and lie in a different
number field for each prime p. To obtain a formula in which everything takes
place over Q, one observes that the L-series √ of Ep is the L-series of a cubic
twist of a grossencharacter ψ1 of K = Q( −3) which is independent of p. More
precisely, L(Ep , s) = L(χ√ p ψ1 , s) where ψ1 is defined (just like the grossenchar-
acters for Q(i) and Q( −7) defined in (105) and in the previous “application”)
by ψ1 (a) = λ if a = (λ) with λ ≡ 1 (mod 3) and ψ1 (α) = 0 if 3|N (a), and χp
is the cubic character which sends a = (λ) to the unique cube root of unity
in K which is congruent to (λ̄/λ)(p−1)/3 modulo p. This means that formally
the L-value L(Ep , 1) for p = 9k + 1 is congruent modulo p to the central
value L(ψ 12k+1 , 6k + 1). Of course both of these numbers are transcendental,
but the theory of p-adic L-functions shows that their “algebraic parts” are in
39k−1 Ω 12k+1 Ck
fact congruent modulo p: we have L(ψ112k+1 , 6k + 1) =
(2π)6k (6k)!
for some integer Ck ∈ Z, and Sp ≡ Ck (mod p). Now the calculation of Ck
proceeds exactly like that of the number Am defined in (108): Ck is, up to nor-
malizing constants, equal to the value at z = z0 of the 6k-th non-holomorphic
 Elliptic Modular Forms and Their Applications 99

derivative of Θ(z), and can be computed quasi-recursively, and Ck is also

equal to c2k where ck is, again up to normalizing constants,
√ the value of the
3k-th non-holomorphic derivative of η(z) at z = 12 (−1 + −3) and is given by
the formula in the theorem. Finally, an estimate of the size of L(Ep , 1) shows
that |Sp | < p, so that Sp vanishes if and only if ck ≡ 0 (mod p), as claimed.
The numbers ck can also be described by a generating function rather than
a quasi-recursion, using the relations between modular forms and differential
equations discussed in §5, namely
 1/2 ∞  n
 1 1 2 cn x F ( 23 , 23 ; 43 ; x)3
1 − x)1/24 F , ; ;x = ,
3 3 3 n=0
(3n)! 8 F ( 13 , 13 ; 23 ; x)3

where F (a, b; c; x) denotes Gauss’s hypergeometric function. ♥

References and Further Reading

There are several other elementary texts on the theory of modular forms
which the reader can consult for a more detailed introduction to the field.
Four fairly short introductions are the classical book by Gunning (Lectures
on Modular Forms, Annals of Math. Studies 48, Princeton, 1962), the last
chapter of Serre’s Cours d’Arithmétique (Presses Universitaires de France,
1970; English translation A Course in Arithmetic, Graduate Texts in Math-
ematics 7, Springer 1973), Ogg’s book Modular Forms and Dirichlet Series
(Benjamin, 1969; especially for the material covered in §§3–4 of these notes),
and my own chapter in the book From Number Theory to Physics (Springer
1992). The chapter on elliptic curves by Henri Cohen in the last-named book
is also a highly recommended and compact introduction to a field which is
intimately related to modular forms and which is touched on many times in
these notes. Here one can also recommend N. Koblitz’s Introduction to Elliptic
Curves and Modular forms (Springer Graduate Texts 97, 1984). An excellent
book-length Introduction to Modular Forms is Serge Lang’s book of that
title (Springer Grundlehren 222, 1976). Three books of a more classical na-
ture are B. Schoeneberg’s Elliptic Modular Functions (Springer Grundlehren
203, 1974), R.A. Rankin’s Modular Forms and Modular Functions (Cam-
bridge, 1977) and (in German) Elliptische Funktionen und Modulformen by
M. Koecher and A. Krieg (Springer, 1998).
The point of view in these notes leans towards the analytic, with as many
results as possible (like the algebraicity of j(z) when z is a CM point) be-
ing derived purely in terms of the theory of modular forms over the complex
numbers, an approach which was sufficient – and usually simpler – for the
type of applications which I had in mind. The books listed above also belong
to this category. But for many other applications, including the deepest ones
in Diophantine equations and arithmetic algebraic geometry, a more arith-
metic and more advanced approach is required. Here the basic reference is
100 D. Zagier

Shimura’s classic Introduction to the Arithmetic Theory of Automorphic Func-

tions (Princeton 1971), while two later books that can also be recommended
are Miyake’s Modular Forms (Springer 1989) and the very recent book A First
Course in Modular Forms by Diamond and Shurman (Springer 2005).
We now give, section by section, some references (not intended to be in
any sense complete) for various of the specific topics and examples treated in
these notes.
1–2. The material here is all standard and can be found in the books listed
above, except for the statement in the final section of §1.2 that the class
numbers of negative discriminants are the Fourier coefficients of some kind of
modular form of weight 3/2, which was proved in my paper in CRAS Paris
(1975), 883–886.
3.1. Proposition 11 on the number of representations of integers as sums of
four squares was proved by Jacobi in the Fundamenta Nova Theoriae Ellip-
ticorum, 1829. We do not give references for the earlier theorems of Fermat
and Lagrange. For more information about sums of squares, one can consult
the book Representations of Integers as Sums of Squares (Springer, 1985) by
E. Grosswald. The theory of Jacobi forms mentioned in connection with the
two-variable theta functions θi (z, u) was developed in the book The Theory
of Jacobi Forms (Birkhäuser, 1985) by M. Eichler and myself. Mersmann’s
theorem is proved in his Bonn Diplomarbeit, “Holomorphe η-Produkte und
nichtverschwindende ganze Modulformen für Γ0 (N )” (Bonn, 1991), unfortu-
nately never published in a journal. The theorem of Serre and Stark is given
in their paper “Modular forms of weight 1/2 ” in Modular Forms of One Vari-
able VI (Springer Lecture Notes 627, 1977, editors J-P. Serre and myself; this
is the sixth volume of the proceedings of two big international conferences on
modular forms, held in Antwerp in 1972 and in Bonn in 1976, which con-
tain a wealth of further material on the theory). The conjecture of Kac and
Wakimoto appeared in their article in Lie Theory and Geometry in Honor of
Bertram Kostant (Birkhäuser, 1994) and the solutions by Milne and myself
in the Ramanujan Journal and Mathematical Research Letters, respectively,
in 2000.
3.2. The detailed proof and references for the theorem of Hecke and Schoen-
berg can be found in Ogg’s book cited above. Niemeier’s classification of
unimodular lattices of rank 24 is given in his paper “Definite quadratische
Formen der Dimension 24 und Diskriminante 1” (J. Number Theory 5, 1973).
Siegel’s mass formula was presented in his paper “Über die analytische Theo-
rie der quadratischen Formen” in the Annals of Mathematics, 1935 (No. 20 of
his Gesammelte Abhandlungen, Springer, 1966). A recent paper by M. King
(Math. Comp., 2003) improves by a factor of more than 14 the lower bound
on the number of inequivalent unimodular even lattices of dimension 32. The
paper of Mallows-Odlyzko-Sloane on extremal theta series appeared in J. Al-
gebra in 1975. The standard general reference for the theory of lattices, which
contains an immense amount of further material, is the book by Conway and
 Elliptic Modular Forms and Their Applications 101

Sloane (Sphere Packings, Lattices and Groups, Springer 1998). Milnor’s ex-
ample of 16-dimensional tori as non-isometric isospectral manifolds was given
in 1964, and 2-dimensional examples (using modular groups!) were found by
Vignéras in 1980. Examples of pairs of truly drum-like manifolds – i.e., do-
mains in the flat plane – with different spectra were finally constructed by
Gordon, Webb and Wolpert in 1992. For references and more on the history
of this problem, we refer to the survey paper in the book What’s Happening
in the Mathematical Sciences (AMS, 1993).
4.1–4.3. For a general introduction to Hecke theory we refer the reader to
the books of Ogg and Lang mentioned above or, of course, to the beautifully
written papers of Hecke himself (if you can read German). Van der Blij’s
example is given in his paper “Binary quadratic forms of discriminant −23” in
Indagationes Math., 1952. A good exposition of the connection between Galois
representations and modular forms of weight one can be found in Serre’s article
in Algebraic Number Fields: L-Functions and Galois Properties (Academic
Press 1977).
4.4. Book-length expositions of the Taniyama–Weil conjecture and its proof
using modular forms are given in Modular Forms and Fermat’s Last The-
orem (G. Cornell, G. Stevens and J. Silverman, eds., Springer 1997) and,
at a much more elementary level, Invitation to the Mathematics of Fermat-
Wiles (Y. Hellegouarch, Academic Press 2001), which the reader can con-
sult for more details concerning the history of the problem and its solution
and for further references. An excellent survey of the content and status
of Serre’s conjecture can be found in the book Lectures on Serre’s Conjec-
tures by Ribet and Stein (http://modular.fas.harvard.edu/papers/serre/ribet-
stein.ps). For the final proof of the conjecture and references to all earlier
work, see “Modularity of 2-adic Barsotti-Tate representations” by M. Kisin
(http://www.math.uchicago.edu/∼kisin/preprints.html). Livné’s example ap-
peared in his paper “Cubic exponential sums and Galois representations”
(Contemp. Math. 67, 1987). For an exposition of the conjectural and known
examples of higher-dimensional varieties with modular zeta functions, in par-
ticular of those coming from mirror symmetry, we refer to the recent paper
“Modularity of Calabi-Yau varieties” by Hulek, Kloosterman and Schütt in
Global Aspects of Complex Geometry (Springer, 2006) and to the monograph
Modular Calabi-Yau Threefolds by C. Meyer (Fields Institute, 2005). However,
the proof of Serre’s conjectures means that the modularity is now known in
many more cases than indicated in these surveys.
5.1. Proposition 15, as mentioned in the text, is due to Ramanujan (eq. (30)
in “On certain Arithmetical Functions,” Trans. Cambridge Phil. Soc., 1916).
For a good discussion of the Chazy equation and its relation to the “Painlevé
property” and to SL(2, C), see the article “Symmetry and the Chazy equation”
by P. Clarkson and P. Olver (J. Diff. Eq. 124, 1996). The result by Gallagher
on means of periodic functions which we describe as our second application is
102 D. Zagier

described in his very nice paper “Arithmetic of means of squares and cubes”
in Internat. Math. Res. Notices, 1994.
5.2. Rankin–Cohen brackets were defined in two stages: the general conditions
needed for a polynomial in the derivatives of a modular form to itself be
modular were described by R. Rankin in “The construction of automorphic
forms from the derivatives of a modular form” (J. Indian Math. Soc., 1956),
and then the specific bilinear operators [ · , · ]n satisfying these conditions were
given by H. Cohen as a lemma in “Sums involving the values at negative
integers of L functions of quadratic characters” (Math. Annalen, 1977). The
Cohen–Kuznetsov series were defined in the latter paper and in the paper
“A new class of identities for the Fourier coefficients of a modular form” (in
Russian) by N.V. Kuznetsov in Acta Arith., 1975. The algebraic theory of
“Rankin–Cohen algebras” was developed in my paper “Modular forms and
differential operators” in Proc. Ind. Acad. Sciences, 1994, while the papers
by Manin, P. Cohen and myself and by the Untenbergers discussed in the
second application in this subsection appeared in the book Algebraic Aspects
of Integrable Systems: In Memory of Irene Dorfman (Birkhäuser 1997) and
in the J. Anal. Math., 1996, respectively.
5.3. The name and general definition of quasimodular forms were given in
the paper “A generalized Jacobi theta function and quasimodular forms” by
M. Kaneko and myself in the book The Moduli Space of Curves (Birkhäuser
1995), immediately following R. Dijkgraaf’s article “Mirror symmetry and el-
liptic curves” in which the problem of counting ramified coverings of the torus
is presented and solved.
5.4. The relation between modular forms and linear differential equations was
at the center of research on automorphic forms at the turn of the (previous)
century and is treated in detail in the classical works of Fricke, Klein and
Poincaré and in Weber’s Lehrbuch der Algebra. A discussion in a modern lan-
guage can be found in §5 of P. Stiller’s paper in the Memoirs of the AMS 299,
1984. Beukers’s modular proof of the Apéry identities implying the irrational-
ity of ζ(2) and ζ(3) can be found in his article in Astérisque 147–148 (1987),
which also contains references to Apéry’s original paper and other related
work. My paper with Kleban on a connection between percolation theory and
modular forms appeared in J. Statist. Phys. 113 (2003).
6.1. There are several references for the theory of complex multiplication.
A nice book giving an introduction to the theory at an accessible level is
Primes of the form x2 + ny 2 by David Cox (Wiley, 1989), while a more ad-
vanced account is given in the Springer Lecture Notes Volume 21 by Borel,
Chowla, Herz, Iwasawa and Serre. Shanks’s approximation to π is given in
a paper in J. Number Theory in 1982. Heegner’s original paper attacking the
class number one problem by complex multiplication methods appeared in
Math. Zeitschrift in 1952. The result quoted about congruent numbers was
proved by Paul Monsky in “Mock Heegner points and congruent numbers”
 Elliptic Modular Forms and Their Applications 103

(Math. Zeitschrift 1990) and the result about Sylvester’s problem was an-
nounced by Noam Elkies, in “Heegner point computations” (Springer Lecture
Notes in Computing Science, 1994). See also the recent preprint “Some Dio-
phantine applications of Heegner points” by S. Dasgupta and J. Voight.
6.2. The formula for the norms of differences of singular moduli was proved
in my joint paper “Singular moduli” (J. reine Angew. Math. 355, 1985) with
B. Gross, while our more general result concerning heights of Heegner points
appeared in “Heegner points and derivatives of L-series” (Invent. Math. 85,
1986). The formula describing traces of singular moduli is proved in my pa-
per of the same name in the book Motives, Polylogarithms and Hodge Theory
(International Press, 2002). Borcherds’s result on product expansions of auto-
morphic forms was published in a celebrated paper in Invent. Math. in 1995.
6.3. The Chowla–Selberg formula is discussed, among many other places, in
the last chapter of Weil’s book Elliptic Functions According to Eisenstein and
Kronecker (Springer Ergebnisse 88, 1976), which contains much other beauti-
ful historical and mathematical material. Chudnovsky’s result about the tran-
scendence of Γ ( 14 ) is given in his paper for the 1978 (Helskinki) International
Congress of Mathematicians, while Nesterenko’s generalization giving the al-
gebraic independence of π and eπ is proved in his paper “Modular functions
and transcendence questions” (in Russian) in Mat. Sbornik, 1996. A good
summary of this work, with further references, can be found in the “featured
review” of the latter paper in the 1997 Mathematical Reviews. The algorith-
mic way of computing Taylor expansions of modular forms at CM points is
described in the first of the two joint papers with Villegas cited below, in
connection with the calculation of central values of L-series.
6.4. A discussion of formulas like (102) can be found in any of the general ref-
erences for the theory of complex multiplication listed above. The applications
of such formulas to questions of primality testing, factorization and cryptogra-
phy is treated in a number of papers. See for instance “Efficient construction
of cryptographically strong elliptic curves” by H. Baier and J. Buchmann
(Springer Lecture Notes in Computer Science 1977, 2001) and its bibliogra-
phy. Serre’s results on powers of the eta-function and on lacunarity of modu-
lar forms are contained in his papers “Sur la lacunarité des puissances de η”
(Glasgow Math. J., 1985) and “Quelques applications du théorème de densité
de Chebotarev” (Publ. IHES, 1981), respectively. The numerical experiments
concerning the numbers defined in (108) were given in a note by B. Gross
and myself in the memoires of the French mathematical society (1980), and
the two papers with F. Villegas on central values of Hecke L-series and their
applications to Sylvester’s problems appeared in the proceedings of the third
and fourth conferences of the Canadian Number Theory Association in 1993
and 1995, respectively.