Fermats Last Theorem

Henri Darmon
(darmon@math.mcgill.ca)
Department of Mathematics
McGill University
Montreal, QC
Canada H3A 2K6
Fred Diamond
(fdiamond@pmms.cam.ac.uk)
D.P.M.M.S.
Cambridge University
Cambridge, CB2 1SB
United Kingdom
Richard Taylor
(taylorr@maths.ox.ac.uk)
Mathematics Institute
Oxford University
24-29 St. Giles
Oxford, OX1 3LB
United Kingdom
September 9, 2007
The authors would like to give special thanks to N. Boston, K. Buzzard, and
B. Conrad for providing so much valuable feedback on earlier versions of this
paper. They are also grateful to A. Agboola, M. Bertolini, B. Edixhoven, J.
Fearnley, R. Gross, L. Guo, F. Jarvis, H. Kisilevsky, E. Liverance, J. Manohar-
mayum, K. Ribet, D. Rohrlich, M. Rosen, R. Schoof, J.-P. Serre, C. Skinner,
D. Thakur, J. Tilouine, J. Tunnell, A. Van der Poorten, and L. Washington
for their helpful comments.
Darmon thanks the members of CICMA and of the Quebec-Vermont Num-
ber Theory Seminar for many stimulating conversations on the topics of this
paper, particularly in the Spring of 1995. For the same reason Diamond is
grateful to the participants in an informal seminar at Columbia University
in 1993-94, and Taylor thanks those attending the Oxford Number Theory
Seminar in the Fall of 1995.
1
Parts of this paper were written while the authors held positions at other
institutions: Darmon at Princeton University, Diamond at the Institute for
Advanced Study, and Taylor at Cambridge University. During some of the pe-
riod, Diamond enjoyed the hospitality of Princeton University, and Taylor that
of Harvard University and MIT. The writing of this paper was also supported
by research grants from NSERC (Darmon), NSF # DMS 9304580 (Diamond)
and by an advanced fellowship from EPSRC (Taylor).
This article owes everything to the ideas of Wiles, and the arguments pre-
sented here are fundamentally his [W3], though they include both the work
[TW] and several simplications to the original arguments, most notably that
of Faltings. In the hope of increasing clarity, we have not always stated
theorems in the greatest known generality, concentrating instead on what is
needed for the proof of the Shimura-Taniyama conjecture for semi-stable ellip-
tic curves. This article can serve as an introduction to the fundamental papers
[W3] and [TW], which the reader is encouraged to consult for a dierent, and
often more in-depth, perspective on the topics considered. Another useful
more advanced reference is the article [Di2] which strengthens the methods of
[W3] and [TW] to prove that every elliptic curve that is semistable at 3 and 5
is modular.
2
Introduction
Fermats Last Theorem
Fermats Last Theorem states that the equation
r
n
+
n
= .
n
. r. ,= 0
has no integer solutions when : is greater than or equal to 3. Around 1630,
Pierre de Fermat claimed that he had found a truly wonderful proof of this
theorem, but that the margin of his copy of Diophantus Arithmetica was too
small to contain it:
Cubum autem in duos cubos, aut quadrato quadratum in duos
quadrato quadratos, et generaliter nullam in innitum ultra qua-
dratum potestatem in duos ejusdem nominis fas est dividere; cujus
rei demonstrationem mirabile sane detexi. Hanc marginis exiguitas
non caperet.
Among the many challenges that Fermat left for posterity, this was to prove
the most vexing. A tantalizingly simple problem about whole numbers, it
stood unsolved for more than 350 years, until in 1994 Andrew Wiles nally
laid it to rest.
Prehistory: The only case of Fermats Last Theorem for which Fermat actu-
ally wrote down a proof is for the case : = 4. To do this, Fermat introduced
the idea of innite descent which is still one the main tools in the study of
Diophantine equations, and was to play a central role in the proof of Fermats
Last Theorem 350 years later. To prove his Last Theorem for exponent 4, Fer-
mat showed something slightly stronger, namely that the equation r
4
+
4
= .
2
has no solutions in relatively prime integers with r. ,= 0. Solutions to such
an equation correspond to rational points on the elliptic curve
2
= n
3
4n.
Since every integer : 3 is divisible either by an odd prime or by 4, the result
of Fermat allowed one to reduce the study of Fermats equation to the case
where : = / is an odd prime.
In 1753, Leonhard Euler wrote down a proof of Fermats Last Theorem for
the exponent / = 3, by performing what in modern language we would call
a 3-descent on the curve r
3
+
3
= 1 which is also an elliptic curve. Eulers
argument (which seems to have contained a gap) is explained in [Edw], ch. 2,
and [Dic1], p. 545.
It took mathematicians almost 100 years after Eulers achievement to han-
dle the case / = 5; this was settled, more or less simultaneously, by Gustav
3
Peter Lejeune Dirichlet [Dir] and Adrien Marie Legendre [Leg] in 1825. Their
elementary arguments are quite involved. (Cf. [Edw], sec. 3.3.)
In 1839, Fermats equation for exponent 7 also yielded to elementary meth-
ods, through the heroic eorts of Gabriel Lame. Lames proof was even more
intricate than the proof for exponent 5, and suggested that to go further, new
theoretical insights would be needed.
The work of Sophie Germain: Around 1820, in a letter to Gauss, Sophie
Germain proved that if / is a prime and = 2/+1 is also prime, then Fermats
equation r

= .

with exponent / has no solutions (r. . .) with r. ,= 0

(mod /). Germains theorem was the rst really general proposition on Fer-
mats Last Theorem, unlike the previous results which considered the Fermat
equation one exponent at a time.
The case where the solution (r. . .) to r

= .

satises r. ,= 0 (mod
/) was called the rst case of Fermats Last Theorem, and the case where /
divides r., the second case. It was realized at that time that the rst case
was generally easier to handle: Germains theorem was extended, using similar
ideas, to cases where //+1 is prime and / is small, and this led to a proof that
there were no rst case solutions to Fermats equation with prime exponents
/ 100, which in 1830 represented a signicant advance. The division between
rst and second case remained fundamental in much of the later work on the
subject. In 1977, Terjanian [Te] proved that if the equation r
2
+
2
= .
2
has
a solution (r. . .), then 2/ divides either r or , i.e., the rst case of Fermats
Last Theorem is true for even exponents. His simple and elegant proof used
only techniques that were available to Germain and her contemporaries.
The work of Kummer: The work of Ernst Eduard Kummer marked the
beginning of a new era in the study of Fermats Last Theorem. For the rst
time, sophisticated concepts of algebraic number theory and the theory of
1-functions were brought to bear on a question that had until then been
addressed only with elementary methods. While he fell short of providing
a complete solution, Kummer made substantial progress. He showed how
Fermats Last Theorem is intimately tied to deep questions on class numbers
of cyclotomic elds which are still an active subject of research. Kummers
approach relied on the factorization
(r +)(r +

) (r +
1

) = .

of Fermats equation over the ring Z[

] generated by the /th roots of unity.

One observes that the greatest common divisor of any two factors in the prod-
uct on the left divides the element (1

), which is an element of norm /.

4
Since the product of these numbers is a perfect /-th power, one is tempted to
conclude that (r+). . . . . (r+
1

) are each /-th powers in the ring Z[

] up
to units in this ring, and up to powers of (1

). Such an inference would be

valid if one were to replace Z[

] by Z, and is a direct consequence of unique

factorization of integers into products of primes. We say that a ring 1 has
property l1 if every non-zero element of 1 is uniquely a product of primes,
up to units. Mathematicians such as Lame made attempts at proving Fer-
mats Last Theorem based on the mistaken assumption that the rings Z[

]
had property l1. Legend even has it that Kummer fell into this trap, al-
though this story now has been discredited; see for example [Edw], sec. 4.1. In
fact, property l1 is far from being satised in general: one now knows that
the rings Z[

] have property l1 only for / < 23 (cf. [Wa], ch. 1).

It turns out that the full force of property l1 is not really needed in the
applications to Fermats Last Theorem. Say that a ring 1 has property l1

if the following inference is valid:

c/ = .

. and gcd(c. /) = 1 c and / are /th powers up to units of 1.

If a ring 1 has property l1, then it also has property l1

, but the converse

need not be true. Kummer showed that Fermats last theorem was true for
exponent / if Z[

] satised the property l1

(cf. [Wa]). The proof is far from

trivial, because of diculties arising from the units in Z[

] as well as from
the possible failure of property l1. (A number of Kummers contemporaries,
such as Cauchy and Lame, seem to have overlooked both of these diculties
in their attempts to prove Fermats Last Theorem.)
Kummer then launched a systematic study of the property l1

for the
rings Z[

]. He showed that even if Z[

] failed to have unique factorization,

it still possessed unique factorization into prime ideals. He dened the ideal
class group as the quotient of the group of fractional ideals by its subgroup
consisting of principal ideals, and was able to establish the niteness of this
class group. The order of the class group of Z[

], denoted /

, could be taken
as a measure of the failure of the ring Z[

] to satisfy l1. It was rather

straightforward to show that if / did not divide /

, then Z[

] satised the
property l1

. In this case, one called / a regular prime. Kummer thus showed

that Fermats last theorem is true for exponent / if / is a regular prime.
He did not stop here. For it remained to give an ecient means of com-
puting /

, or at least an ecient way of checking when / divides /

. The class
number /

can be factorized as a product

/

= /
+

.
5
where /
+

is the class number of the real subeld Q(

)
+
, and /

is dened as
/

,/
+

. Essentially because of the units in Q(

)
+
, the factor /
+

is somewhat
dicult to compute, while, because the units in Q(

)
+
generate the group of
units in Q(

) up to nite index, the term /

can be expressed in a simple

closed form. Kummer showed that if / divides /
+

, then / divides /

. Hence, /
divides /

if and only if / divides /

. This allowed one to avoid the diculties

inherent in the calculation of /
+

. Kummer then gave an elegant formula for /

by considering the Bernoulli numbers 1

n
, which are rational numbers dened
by the formula
r
c
x
1
=

1
n
:!
r
n
.
He produced an explicit formula for the class number /

, and concluded that

if / does not divide the numerator of 1
2i
, for 1 i (/ 3),2, then / is
regular, and conversely.
The conceptual explanation for Kummers formula for /

lies in the work

of Dirichlet on the analytic class number formula, where it is shown that /

can be expressed as a product of special values of certain (abelian) 1-series

1(:. ) =

n=1
(:):
s
associated to odd Dirichlet characters. Such special values in turn can be
expressed in terms of certain generalized Bernoulli numbers 1
1,
, which are
related to the Bernoulli numbers 1
i
via congruences mod /. (For more details,
see [Wa].)
These considerations led Kummer to initiate a deep study relating congru-
ence properties of special values of 1-functions and of class numbers, which
was to emerge as a central concern of modern algebraic number theory, and
was to reappear in a surprisingly dierent guise at the heart of Wiles
strategy for proving the Shimura-Taniyama conjecture.
Later developments: Kummers work had multiple ramications, and led
to a very active line of enquiry pursued by many people. His formulae re-
lating Bernoulli numbers to class numbers of cyclotomic elds were rened
by Kenneth Ribet [R1], Barry Mazur and Andrew Wiles [MW], using new
methods from the theory of modular curves which also play a central role in
Wiles more recent work. (Later Francisco Thaine [Th] reproved some of the
results of Mazur and Wiles using techniques inspired directly from a reading
of Kummer.) In a development more directly related to Fermats Last Theo-
rem, Wieferich proved that if /
2
does not divide 2
1
1, then the rst case
of Fermats Last Theorem is true for exponent /. (Cf. [Ri], lecture VIII.)
6
There were many other renements of similar criteria for Fermats Last
theorem to be true. Computer calculations based on these criteria led to a
verication that Fermats Last theorem is true for all odd prime exponents less
than four million [BCEM], and that the rst case is true for all / 8.858 10
20
[Su].
The condition that / is a regular prime seems to hold heuristically for about
61% of the primes. (See the discussion on p. 63, and also p. 108, of [Wa], for
example.) In spite of the convincing numerical evidence, it is still not known
if there are innitely many regular primes. Ironically, it is not too dicult to
show that there are innitely many irregular primes. (Cf. [Wa].)
Thus the methods introduced by Kummer, after leading to very strong
results in the direction of Fermats Last theorem, seemed to become mired in
diculties, and ultimately fell short of solving Fermats conundrum
1
.
Faltings proof of the Mordell conjecture: In 1985, Gerd Faltings [Fa]
proved the very general statement (which had previously been conjectured
by Mordell) that any equation in two variables corresponding to a curve of
genus strictly greater than one had (at most) nitely many rational solutions.
In the context of Fermats Last Theorem, this led to the proof that for each
exponent : 3, the Fermat equation r
n
+
n
= .
n
has at most nitely many
integer solutions (up to the obvious rescaling). Andrew Granville [Gra] and
Roger Heath-Brown [HB] remarked that Faltings result implies Fermats Last
Theorem for a set of exponents of density one.
However, Fermats Last Theorem was still not known to be true for an
innite set of prime exponents. In fact, the theorem of Faltings seemed ill-
equipped for dealing with the ner questions raised by Fermat in his margin,
namely of nding a complete list of rational points on all of the Fermat curves
r
n
+
n
= 1 simultaneously, and showing that there are no solutions on these
curves when : 3 except the obvious ones.
Mazurs work on Diophantine properties of modular curves: Although
it was not realized at the time, the chain of ideas that was to lead to a proof
of Fermats Last theorem had already been set in motion by Barry Mazur
in the mid seventies. The modular curves A
0
(/) and A
1
(/) introduced in
section 1.2 and 1.5 give rise to another naturally occurring innite family
of Diophantine equations. These equations have certain systematic rational
solutions corresponding to the cusps that are dened over Q, and are analogous
1
However, W. McCallum has recently introduced a technique, based on the method
of Chabauty and Coleman, which suggests new directions for approaching Fermats Last
Theorem via the cyclotomic theory. An application of McCallums method to showing the
second case of Fermats Last Theorem for regular primes is explained in [Mc].
7
to the so-called trivial solutions of Fermats equation. Replacing Fermat
curves by modular curves, one could ask for a complete list of all the rational
points on the curves A
0
(/) and A
1
(/). This problem is perhaps even more
compelling than Fermats Last Theorem: rational points on modular curves
correspond to objects with natural geometric and arithmetic interest, namely,
elliptic curves with cyclic subgroups or points of order /. In [Maz1] and [Maz2],
B. Mazur gave essentially a complete answer to the analogue of Fermats Last
Theorem for modular curves. More precisely, he showed that if / ,= 2. 3. 5
and 7, (i.e., A
1
(/) has genus 0) then the curve A
1
(/) has no rational points
other than the trivial ones, namely cusps. He proved analogous results for
the curves A
0
(/) in [Maz2], which implied, in particular, that an elliptic curve
over Q with square-free conductor has no rational cyclic subgroup of order /
over Q if / is a prime which is strictly greater than 7. This result appeared a
full ten years before Faltings proof of the Mordell conjecture.
Freys strategy: In 1986, Gerhard Frey had the insight that these construc-
tions might provide a precise link between Fermats Last Theorem and deep
questions in the theory of elliptic curves, most notably the Shimura Taniyama
conjecture. Given a solution c

+ /

= c

to the Fermat equation of prime

degree /, we may assume without loss of generality that c

1 (mod 4) and
that /

0 (mod 32). Frey considered (following Hellegouarch, [He], p. 262;

cf. also Kubert-Lang [KL], ch. 8, 2) the elliptic curve
1 :
2
= r(r c

)(r +/

).
This curve is semistable, i.e., it has square-free conductor. Let 1[/] denote the
group of points of order / on 1 dened over some (xed) algebraic closure

Q
of Q, and let 1 denote the smallest number eld over which these points are
dened. This extension appears as a natural generalization of the cyclotomic
elds Q(

) studied by Kummer. What singles out the eld 1 for special

attention is that it has very little ramication: using Tates analytic description
of 1 at the primes dividing c/c, it could be shown that 1 was ramied only at 2
and /, and that the ramication of 1 at these two primes was rather restricted.
(See theorem 2.15 of section 2.2 for a precise statement.) Moreover, the results
of Mazur on the curve A
0
(/) could be used to show that 1 is large, in the
following precise sense. The space 1[/] is a vector space of dimension 2 over the
nite eld F

with / elements, and the absolute Galois group G

Q
= Gal (

Q,Q)
acts F

-linearly on 1[/]. Choosing an F

-basis for 1[/], the action is described

by a representation

E,
: Gal (1,Q) G1
2
(F

).
8
Mazurs results in [Maz1] and [Maz2] imply that
E,
is irreducible if / 7
(using the fact that 1 is semi-stable). In fact, combined with earlier results
of Serre [Se6], Mazurs results imply that for / 7, the representation
E,
is
surjective, so that Gal (1,Q) is actually isomorphic to G1
2
(F

) in this case.
Serres conjectures: In [Se7], Jean-Pierre Serre made a careful study of mod
/ Galois representations : G
Q
G1
2
(F

) (and, more generally, of repre-

sentations into G1
2
(/), where / is any nite eld). He was able to make very
precise conjectures (see section 3.2) relating these representations to modular
forms mod /. In the context of the representations
E,
that occur in Freys
construction, Serres conjecture predicted that they arose from modular forms
(mod /) of weight two and level two. Such modular forms, which correspond to
dierentials on the modular curve A
0
(2), do not exist because A
0
(2) has genus
0. Thus Serres conjecture implied Fermats Last Theorem. The link between
elds with Galois groups contained in G1
2
(F

) and modular forms mod / still

appears to be very deep, and Serres conjecture remains a tantalizing open
problem.
Ribets work: lowering the level: The conjecture of Shimura and Taniya-
ma (cf. section 1.8) provides a direct link between elliptic curves and modular
forms. It predicts that the representation
E,
obtained from the /-division
points of the Frey curve arises from a modular form of weight 2, albeit a form
whose level is quite large. (It is the product of all the primes dividing c/c,
where c

+ /

= c

is the putative solution to Fermats equation.) Ribet [R5]

proved that, if this were the case, then
E,
would also be associated with a
modular form mod / of weight 2 and level 2, in the way predicted by Serres
conjecture. This deep result allowed him to reduce Fermats Last Theorem to
the Shimura-Taniyama conjecture.
Wiles work: proof of the Shimura-Taniyama conjecture: In [W3]
Wiles proves the Shimura-Taniyama conjecture for semi-stable elliptic curves,
providing the nal missing step and proving Fermats Last Theorem. After
more than 350 years, the saga of Fermats Last theorem has come to a spec-
tacular end.
The relation between Wiles work and Fermats Last Theorem has been
very well documented (see, for example, [R8], and the references contained
therein). Hence this article will focus primarily on the breakthrough of Wiles
[W3] and Taylor-Wiles [TW] which leads to the proof of the Shimura-Taniyama
conjecture for semi-stable elliptic curves.
From elliptic curves to /-adic representations: Wiles opening gambit
for proving the Shimura-Taniyama conjecture is to view it as part of the more
9
general problem of relating two-dimensional Galois representations and mod-
ular forms. The Shimura-Taniyama conjecture states that if 1 is an elliptic
curve over Q, then 1 is modular. One of several equivalent denitions of mod-
ularity is that for some integer ` there is an eigenform ) =

c
n

n
of weight
two on
0
(`) such that
#1(F
p
) = j + 1 c
p
for all but nitely primes j. (By an eigenform, here we mean a cusp form
which is a normalized eigenform for the Hecke operators; see section 1 for
denitions.)
This conjecture acquires a more Galois theoretic avour when one considers
the two dimensional /-adic representation

E,
: G
Q
G1
2
(Z

)
obtained from the action of G
Q
on the /-adic Tate module of 1: T

1 =
lim

1[|
n
](

Q). An /-adic representation of G

Q
is said to arise from an eigen-
form ) =

c
n

n
with integer coecients c
n
if
tr ((Frob
p
)) = c
p
.
for all but nitely many primes j at which is unramied. Here Frob
p
is a
Frobenius element at j (see section 2), and its image under is a well-dened
conjugacy class.
A direct computation shows that #1(F
p
) = j + 1 tr (
E,
(Frob
p
)) for
all primes j at which
E,
is unramied, so that 1 is modular (in the sense
dened above) if and only if for some /,
E,
arises from an eigenform. In
fact the Shimura-Taniyama conjecture can be generalized to a conjecture that
every /-adic representation, satisfying suitable local conditions, arises from a
modular form. Such a conjecture was proposed by Fontaine and Mazur [FM].
Galois groups and modular forms
Viewed in this way, the Shimura-Taniyama conjecture becomes part of a much
larger picture: the emerging, partly conjectural and partly proven correspon-
dence between certain modular forms and two dimensional representations
of G
Q
. This correspondence, which encompasses the Serre conjectures, the
Fontaine-Mazur conjecture, and the Langlands program for G1
2
, represents a
rst step toward a higher dimensional, non-abelian generalization of class eld
theory.
10
Two-dimensional representations of G
Q
: In the rst part of this century,
class eld theory gave a complete description of G
ab
Q
, the maximal (continu-
ous) abelian quotient of G
Q
. In fact the Kronecker-Weber theorem asserts that
G
ab
Q

=

p
Z

p
, and one obtains a complete description of all one-dimensional
representations of G
Q
. In the second half of this century much attention has
focused on attempts to understand the whole group G
Q
, or more precisely to
describe all its representations. Although there has been a fair degree of suc-
cess in using modular forms to construct representations of G
Q
, less is known
about how exhaustive these constructions are. The major results in the lat-
ter direction along these lines are the work of Langlands [Ll2] and the recent
work of Wiles ([W3] completed by [TW]). Both concern two-dimensional rep-
resentations of G
Q
and give signicant evidence that these representations are
parametrised (in a very precise sense) by certain modular forms. The purpose
of this article is to describe both the proven and conjectural parts of this the-
ory, give a fairly detailed exposition of Wiles recent contribution and explain
the application to Fermats Last theorem. To make this description somewhat
more precise let us distinguish three types of representation.
Artin representations and the Langlands-Tunnell theorem: Contin-
uous representations : G
Q
G1
2
(C) are called (two-dimensional) Artin
representations. Such representations necessarily have nite image, and are
therefore semi-simple. We restrict our attention to those which are irreducible.
They are conjectured to be in bijection (in a precise way) with certain new-
forms (a special class of eigenforms). Those which are odd (i.e. the deter-
minant of complex conjugation is 1), should correspond to weight 1 holo-
morphic newforms. Those which are even should correspond to certain non-
holomorphic (Maass) newforms. Two partial but deep results are known.
(a) (Deligne-Serre) If ) is a holomorphic weight one newform then the cor-
responding Artin representation can be constructed ([DS]).
(b) (Langlands-Tunnell) If is a two dimensional Artin representation with
soluble image then the corresponding modular form exists ([Ll2] and
[Tu]).
The proof of the latter result is analytic in nature, invoking the trace formula
and the theory of 1-functions.
/-adic representations and the Fontaine-Mazur conjecture: By an /-
adic representation we shall mean any continuous representation : G
Q

G1
2
(1) which is unramied outside a nite set of primes and where 1 is a
nite extension of Q

(generalizing slightly the notion of /-adic representation

11
that was introduced before). Given a holomorphic newform ) one can attach
to ) a system of /-adic representations, following Eichler, Shimura, Deligne and
Serre. These /-adic representations are called modular. The Fontaine-Mazur
conjecture (see [FM]) predicts if is an odd, irreducible, /-adic representation
whose restriction to the decomposition group at / is well enough behaved,
then is modular. (The restriction on the behaviour of the representation
on the decomposition group at / is essential in this conjecture; it is not true
that all odd, irreducible two dimensional /-adic representation are modular.)
Before Wiles work almost nothing was known about this conjecture, except
that certain very special cases could be deduced from the work of Hecke,
Langlands and Tunnell.
Mod / representations and Serres conjecture: A mod / representation
is a continuous representation : G
Q
G1
2
(

). For example if 1,Q is an

elliptic curve then the action of G
Q
on the /-division points of 1 gives rise to a
mod / representation
E,
which is just the reduction modulo / of
E,
. One can
use the work of Eichler, Shimura, Deligne and Serre to associate to each mod
/ eigenform a mod / representation of G
Q
. The mod / representations which
arise in this way are called modular. Serre has conjectured [Se7] that every
odd (absolutely) irreducible mod / representation is modular and should arise
from a mod / eigenform with certain very specic properties. This conjecture
can be thought of as having two parts.
The rst asserts that every odd irreducible mod / representation is modular.
About this very little is known. It is known for : G
Q
G1
2
(F
2
) by work
of Hecke. It is also known for : G
Q
G1
2
(F
3
). This latter result is an
application of the Langlands-Tunnell theorem using the two accidents that
there is a section to the homomorphism G1
2
(Z[

2]) G1
2
(F
3
) and that
G1
2
(F
3
) is soluble. Partial results for : G
Q
G1
2
(F
5
) follow from Wiles
work.
Given a mod / representation arising from a mod / eigenform, the second
part of Serres conjecture predicts the minimal weight and level for that mod
/ eigenform. Here the situation is much better. There has been a lot of work
over the last decade (including ideas from Mazur, Ribet, Carayol and Gross)
and the problem is nearly completely resolved (see [Di1]). As was pointed
out earlier, Ribets contribution [R5] implies that, if one can show that the
Galois representation
E,
arising from the (semi-stable) Frey curve attached
to a solution of Fermats equation with exponent / is modular, then one can
show that this representation does not existbecause it would be modular of
weight two and level two and hence one can deduce Fermats Last Theorem.
However we have seen that to show
E,
is modular it suces to show that
12
for some /
0
, the /
0
-adic representation
E,
0
is modular. In particular it suces
to verify that either
E,3
or
E,5
is modular. Hence the Shimura-Taniyama
conjecture can be reduced to (part of) the Fontaine-Mazur conjecture for / = 3
and 5. We have seen that for these primes part of Serres conjecture is known,
so it turns out it suces to prove results of the form Serres conjecture for /
implies the Fontaine-Mazur conjecture for /. This is the direction of Wiles
work, although nothing quite this general has been proven yet.
Deformation theory: Thus the problem Wiles faces is to show that if is
an odd /-adic representation which has irreducible modular reduction and
which is suciently well behaved when restricted to the decomposition group
at /, then is modular. In fact he only proves a weakened version of such a
result, but one which is sucient to conclude that all semistable elliptic curves
are modular.
Wiles approaches the problem by putting it in a more general setting. On
the one hand he considers lifts of to representations over complete noetherian
local Z

-algebras 1. For each nite set of primes , one can consider lifts of
type ; these are lifts which are well-behaved on a decomposition group at /,
and whose ramication at primes not in is rather restricted. In particular,
such a lift is unramied outside o where o is the set of ramied primes
of . A method of Mazur (see [Maz3]) can then be used to show that if is
absolutely irreducible, then there is a representation

univ

: G
Q
G1
2
(1

)
which is universal in the following sense. If : G
Q
G1
2
(1) is a lift of of
type , then there is a unique local homomorphism 1

1 such that is
equivalent to the pushforward of
univ

. Thus the equivalence classes of type

lifts to G1
2
(1) can be identied with Hom(1

. 1). The local ring 1

is called
the universal deformation ring for representations of type .
On the other hand Wiles constructs a candidate for a universal modular
lifting of type

mod

: G
Q
G1
2
(T

).
The ring T

is constructed from the algebra of Hecke operators acting on

a certain space of modular forms. The universal property of 1

gives a
map 1

. The problem thus becomes: to show that this map is an

isomorphism
2
. In fact, it can be shown to be a surjection without great dif-
2
Maps of this kind were already considered in [Maz3] and [BM], and it is conjectured in
[MT] that these maps are isomorphisms in certain cases, though not in exactly the situations
considered by Wiles.
13
culty, and the real challenge is to prove injectivity, i.e., to show, in essence,
that 1

is not larger than T

.
By an ingenious piece of commutative algebra, Wiles found a numerical
criterion for this map to be an isomorphism, and for the ring T

to be a
local complete intersection. This numerical criterion seems to be very close
to a special case of the Bloch-Kato conjecture [BK]. Wiles further showed
(by combining arguments from Galois cohomology and from the theory of
congruences between modular forms) that this numerical criterion was satised
if the minimal version T

of this Hecke algebra (obtained by taking = , i.e.,

allowing the least possible amount of ramication in the deformations) was a
complete intersection. Finally in [TW] it was proved that T

is a complete
intersection.
Outline of the paper
Chapter 1 recalls some basic notions from the classical theory of elliptic curves
and modular forms, such as modular forms and modular curves over C and Q,
Hecke operators and -expansions, and Eichler-Shimura theory. The Shimura-
Taniyama conjecture is stated precisely in section 1.8.
Chapter 2 introduces the basic theory of representations of G
Q
. We describe
Mazurs deformation theory and begin our study of the universal deformation
rings using techniques from Galois cohomology and from the theory of nite
at group schemes. We also recall some basic properties of elliptic curves,
both to explain Freys argument precisely and illustrate the uses of /-adic
representations.
Chapter 3 explains how to associate Galois representations to modular
forms. We then describe what was known and conjectured about associating
modular forms to Galois representations before Wiles work. After introducing
the universal modular lifts of certain mod / representations, we give the proof
of Wiles main theorems, taking for granted certain results of a more technical
nature that are proved in the last two chapters.
Chapter 4 explains how to prove the necessary results concerning the struc-
ture of Hecke algebras: the generalization by Taylor and Wiles of a result of
de Shalit, and the generalization by Wiles of a result of Ribet.
Chapter 5 establishes the fundamental results from commutative algebra
discovered by Wiles, following modications of the approach of Wiles and
Taylor-Wiles proposed by Faltings and Lenstra.
14
Contents
1 Elliptic curves and modular forms 16
1.1 Elliptic curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2 Modular curves and modular forms over C . . . . . . . . . . . 22
1.3 Hecke operators and Hecke theory . . . . . . . . . . . . . . . . 28
1.4 The 1-function associated to a cusp form . . . . . . . . . . . . 33
1.5 Modular curves and modular forms over Q . . . . . . . . . . . 34
1.6 The Hecke algebra . . . . . . . . . . . . . . . . . . . . . . . . 39
1.7 The Shimura construction . . . . . . . . . . . . . . . . . . . . 44
1.8 The Shimura-Taniyama conjecture . . . . . . . . . . . . . . . 47
2 Galois theory 50
2.1 Galois representations . . . . . . . . . . . . . . . . . . . . . . 50
2.2 Representations associated to elliptic curves . . . . . . . . . . 55
2.3 Galois cohomology . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4 Representations of G
Q

. . . . . . . . . . . . . . . . . . . . . . 62
2.5 The theory of Fontaine and Laaille . . . . . . . . . . . . . . . 69
2.6 Deformations of representations . . . . . . . . . . . . . . . . . 73
2.7 Deformations of Galois representations . . . . . . . . . . . . . 76
2.8 Special cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3 Modular forms and Galois representations 84
3.1 From modular forms to Galois representations . . . . . . . . . 84
3.2 From Galois representations to modular forms . . . . . . . . . 88
3.3 Hecke algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.4 Isomorphism criteria . . . . . . . . . . . . . . . . . . . . . . . 98
3.5 The main theorem . . . . . . . . . . . . . . . . . . . . . . . . 100
3.6 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4 Hecke algebras 106
4.1 Full Hecke algebras . . . . . . . . . . . . . . . . . . . . . . . . 106
4.2 Reduced Hecke algebras . . . . . . . . . . . . . . . . . . . . . 111
4.3 Proof of theorem 3.31 . . . . . . . . . . . . . . . . . . . . . . . 119
4.4 Proof of theorem 3.36 . . . . . . . . . . . . . . . . . . . . . . . 124
4.5 Homological results . . . . . . . . . . . . . . . . . . . . . . . . 133
5 Commutative algebra 137
5.1 Wiles numerical criterion . . . . . . . . . . . . . . . . . . . . 138
5.2 Basic properties of
A
and
A
. . . . . . . . . . . . . . . . . . 140
15
5.3 Complete intersections and the Gorenstein condition . . . . . 143
5.4 The Congruence ideal for complete intersections . . . . . . . . 148
5.5 Isomorphism theorems . . . . . . . . . . . . . . . . . . . . . . 149
5.6 A resolution lemma . . . . . . . . . . . . . . . . . . . . . . . . 152
5.7 A criterion for complete intersections . . . . . . . . . . . . . . 153
5.8 Proof of Wiles numerical criterion . . . . . . . . . . . . . . . 153
5.9 A reduction to characteristic / . . . . . . . . . . . . . . . . . . 154
5.10 J-structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
1 Elliptic curves and modular forms
1.1 Elliptic curves
We begin with a brief review of elliptic curves. A general reference for the
results discussed in this section is [Si1] and [Si2].
An elliptic curve 1 over a eld 1 is a proper smooth curve over 1 of genus
one with a distinguished 1-rational point. If 1,1 is an elliptic curve and if
is a non-zero holomorphic dierential on 1,1 then 1 can be realised in the
projective plane by an equation (called a Weierstrass equation) of the form
(\) 1
2
2 +c
1
A1 2 +c
3
1 2
2
= A
3
+c
2
A
2
2 +c
4
A2
2
+c
6
2
3
such that the distinguished point is (0 : 1 : 0) (sometimes denoted because
it corresponds to the point at innity in the ane model obtained by setting
2 = 1) and =
dx
2y+a
1
x+a
3
. We also dene the following quantities associated
to (\):
/
2
= c
2
1
+ 4c
2
/
4
= 2c
4
+c
1
c
3
/
6
= c
2
3
+ 4c
6
/
8
= c
2
1
c
6
+ 4c
2
c
6
c
1
c
3
c
4
+c
2
c
2
3
c
2
4
= 9/
2
/
4
/
6
/
2
2
/
8
8/
3
4
27/
2
6
, = (/
2
2
24/
4
)
3
,.
One can check that the equation (\) denes an elliptic curve if and only if
is nonzero. One can also check that such equations dene elliptic curves
which are isomorphic over

1 if and only if they give the same quantity ,. Thus
, only depends on 1 so we will denote it ,
E
. The quantity depends only
on the pair (1. ) so we shall denote it (1. ). If n belongs to 1

then
n
12
(1. n) = (1. ).
An elliptic curve 1,1 has a natural structure of a commutative algebraic
group with the distinguished 1-rational point as the identity element.
An algebraic map between two elliptic curves which sends the distinguished
point of one to the distinguished point of the other is automatically a morphism
16
of algebraic groups. A map between elliptic curves which has nite kernel (and
hence, is generically surjective) is called an isogeny.
Elliptic curves over C: If 1 = C, then the curve 1 is isomorphic as a
complex analytic manifold to the complex torus C,, where is a lattice
in C, i.e., a discrete Z-submodule of C of rank 2. The group law on 1(C)
corresponds to the usual addition in C,. In terms of , an ane equation
for 1 in
2
(C) is given by

2
= 4r
3
+p
2
r +p
3
.
where
p
2
= 60

0
1
.
4
. p
3
= 140

0
1
.
6
.
In terms of this equation, the map from C, to 1(C) sends . to (r. ) =
((.).
t
(.)), where (.) is the Weierstrass -function associated to the lattice
. (Cf. [Si1], ch. VI.) The inverse map is given by integrating the holomorphic
dierential , i.e., sending 1 1(C) to the image of
_

in C,, where is
any path on 1(C) from to 1, and is the lattice of periods
_

, where
ranges over the integral homology H
1
(1(C). Z). Replacing by n changes
to n, so that is determined by 1 only up to homotheties. We scale so
that one of its Z-generators is 1, and another, , has strictly positive imaginary
part. This gives the analytic isomorphism:
1(C) C,1. ).
The complex number in the complex upper half plane H is well dened,
modulo the natural action of o1
2
(Z) on H by Mobius transformations. (Thus
the set of isomorphism classes of elliptic curves over C can be identied with
the quotient H,o1
2
(Z).)
The map . c
2iz
identies C,1. ) with C

,
Z
, where = c
2i
is the
multiplicative Tate period. The analytic isomorphism
1(C) C

,
Z
has the virtue of generalizing to the j-adic setting in certain cases, as we will
see shortly.
Note that [[ < 1. The invariant , can be expressed in terms of by a
convergent power series with integer coecients:
, =
1
+ 744 + 196884 + . (1.1.1)
The following basic facts are a direct consequence of the analytic theory:
17
Proposition 1.1 The subgroup 1[:](C) of points of order : on 1(C) is iso-
morphic (non-canonically) to Z,:ZZ,:Z. More generally, if 1 is any eld
of characteristic zero, the subgroup 1[:](1) is contained in Z,:Z Z,:Z.
Proof: The analytic theory shows that 1(C) is isomorphic as an abstract group
to a product of two circle groups, and the rst statement follows. The second
statement follows from the Lefschetz principle (cf. [Si1], ch. VI, 6). 2
Proposition 1.2 The endomorphism ring End
C
(1) of an elliptic curve over
C is isomorphic either to Z or to an order in a quadratic imaginary eld. The
same is true if one replaces C by any eld of characteristic 0.
Proof: An endomorphism of 1(C) C, induces multiplication by complex
number on the tangent space. Hence End
C
(1) is isomorphic to the ring
of C satisfying . Such a ring is isomorphic either to Z or to a
quadratic imaginary order. The corresponding statement for elds of charac-
teristic 0 follows as in the proof of proposition 1.1. 2
If End
C
(1) Q is a quadratic imaginary eld, we say that 1 has complex
multiplication.
Remark 1.3 It follows from the arithmetic theory of complex multiplication
(cf. [Si2], ch. 1) that any elliptic curve 1 with complex multiplication is dened
over an abelian extension of the quadratic imaginary eld 1 = End
C
(1) Q.
If 1 is dened over Q, then 1 has class number one. There are only nitely
many elliptic curves over Q with complex multiplication, up to twists (i.e.,
C-isomorphism).
Elliptic curves over Q
p
: Now suppose that 1 is an elliptic curve dened
over the j-adic eld Q
p
. There is an equation
(\
min
) 1
2
2 +c
1
A1 2 +c
3
1 2
2
= A
3
+c
2
A
2
2 +c
4
A2
2
+c
6
2
3
for 1 with the property c
i
Z
p
for all i and [[ is minimal amongst all such
equations for 1. Although (\
min
) is not unique, the associated discriminant
depends only on 1 and is denoted
min
E
. Moreover the reduction of (\
min
)
modulo the uniformizer j denes a projective curve

1, which is independent
of the particular minimal equation chosen. If (\) is any equation for 1 with
coecients in Z
p
and with discriminant , then
min
E
divides .
If

1 is a smooth curve we say that 1 has good reduction at j. If

1 has a
unique singular point which is a node we say that 1 has multiplicative reduction
at j. Otherwise

1 has a unique singular point which is a cusp and we say that
18
1 has additive reduction at j. If 1 has good or multiplicative reduction we
say that it has semi-stable reduction at j, or simply that 1 is semi-stable.
If (\) denes a smooth curve mod j then 1 has good reduction at j
and (\) is a minimal equation. If 0 mod j but /
2
2
,= 24/
4
mod j, then
modulo j the equation (\) denes a curve with a node. In this case 1 has
multiplicative reduction at j and (\) is a minimal equation.
Curves with good reduction: In that case j does not divide
min
E
, and the
reduction

1 is an elliptic curve over F
p
.
If is any power of j, and F
q
is the eld with elements, we dene the
integer `
q
to be the number of solutions to the equation (\
min
) in the projec-
tive plane P
2
(F
q
). Thus `
q
is the order of the nite group

1(F
q
). We dene
the integer c
q
by the formula
c
q
= + 1 `
q
.
The integers c
q
are completely determined by c
p
: more precisely, we have
(1 c
p
j
s
+j
12s
)
1
= 1 +c
p
j
s
+c
p
2j
2s
+c
p
3j
3s
+ .
(1.1.2)
We call the expression on the left the (local) 1-function associated to 1
over Q
p
, and denote it by 1(1,Q
p
. :). Concerning the size of c
p
we have the
following fundamental result of Hasse, whose proof can be found in [Si1], ch.
V, 1:
Theorem 1.4 [c
p
[ 2

j.
A further division among curves of good reduction plays a signicant role
in our later discussion. We say that 1 has (good) ordinary reduction if j does
not divide c
p
, and that it has supersingular reduction if j divides c
p
.
When 1 has good reduction at j, we dene its local conductor at j to be
:
p
(1) = 0.
Curves of multiplicative reduction: Elliptic curves over Q
p
which have multi-
plicative reduction at j can be understood by using the j-adic analytic de-
scription discovered by Tate. More precisely, we can formally invert the power
series (1.1.1) expressing , in terms of , to obtain to a power series for in
,
1
, having integer coecients:
= ,
1
+ 744,
2
+ 750420,
3
+ 872769632,
4
+ .
(1.1.3)
19
If 1 has multiplicative reduction, then , Q
p
is non-integral, and hence the
power series (1.1.3) converges, yielding a well-dened value of in jZ
p
. This
is called Tates j-adic period associated to 1 over Q
p
. Note that we have

p
() =
p
(,) =
p
(
min
E
).
We say that 1 has split (resp. non-split) multiplicative reduction at j if
the two tangent lines to the node on

1(F
p
) have slopes dened over F
p
(resp.
F
p
2).
Proposition 1.5 (Tate) There is a j-adic analytic isomorphism
:

Q

p
,
Z
1(

Q
p
).
which has the property that
((r)) = (r
()
). G
Q
p
.
where : G
Q
p
1 is
the trivial character, if 1 has split multiplicative reduction;
the unique unramied quadratic character of G
Q
p
, if 1 has non-split
multiplicative reduction.
The proof of this proposition is explained in [Si2], ch. V, for example.
We dene the 1-function 1(1,Q
p
. :) to be
1(1,Q
p
. :) =
_
(1 j
s
)
1
if 1 has split reduction,
(1 +j
s
)
1
if 1 has non-split reduction.
(1.1.4)
In both cases the conductor :
p
(1) is dened to be 1.
Curves of additive reduction: If 1 has additive reduction at j, we simply dene
1(1,Q
p
. :) = 1. (1.1.5)
The conductor :
p
(1) is dened to be 2, if j 3. When j = 2 or 3, it is
determined by a somewhat more complicated recipe, given in [Ta].
Elliptic curves over Q: Let 1 be an elliptic curve dened over Q. In
particular 1 may be viewed as a curve over Q
p
for every j, and we dene its
(global) conductor by
`
E
=

p
j
m
p
(E)
.
20
The curve 1 is said to be semi-stable if it is semi-stable over all j-adic elds
Q
p
. Note that 1 is semi-stable if and only if its conductor `
E
is square-free.
Using the fact that Q has class number 1, one can show that 1 has a global
minimal Weierstrass model (\
min
) which gives the equation of a minimal
Weierstrass model over each Q
p
. The associated discriminant, denoted
min
E
,
depends only on 1. The associated dierential, denoted
Neron
E
, is called the
Neron dierential. It is well-dened up to sign.
The following, known as the Mordell-Weil theorem, is the fundamental
result about the structure of the group of rational points 1(Q). (Cf. For
example [Si1].)
Theorem 1.6 The group 1(Q) is a nitely generated abelian group. Hence
1(Q) 1 Z
r
.
where 1 is the (nite) torsion subgroup of 1(Q), and : 0 is the rank of 1
over Q.
Concerning the possible structure of 1, there is the following deep result of
Mazur, a variant of which also plays a crucial role in the proof of Fermats
Last Theorem:
Theorem 1.7 If 1,Q is an elliptic curve, then its torsion subgroup is iso-
morphic to one of the following possibilities:
Z,:Z. 1 : 10. : = 12. Z,2:Z Z,2Z. 1 : 4.
The proof is given in [Maz1] (see also [Maz2]). Thanks to this result, the
structure of the torsion subgroup 1 is well understood. (Recently, the tech-
niques of Mazur have been extended by Kamienny [Kam] and Merel [Mer] to
prove uniform boundedness results on the torsion of elliptic curves over general
number elds.)
Much more mysterious is the behaviour of the rank :. It is not known if :
can be arbitrarily large, although results of Mestre [Mes] and Nagao [Na] show
that it is greater or equal to 13 for innitely many elliptic curves over Q. It
turns out that many of the deep results on 1(Q) and on : are based on the
relation with 1-functions.
We dene the global 1-function of the complex variable : by:
1(1,Q. :) =

p
1(1,Q
p
. :). (1.1.6)
21
Exercise 1.8 Using theorem 1.4, show that the innite product dening the
1-function 1(1,Q. :) converges absolutely on the right half plane Real(:)
3,2.
Conjecture 1.9 (Birch-Swinnerton-Dyer) The 1-function 1(1,Q. :) has
an analytic continuation to the entire complex plane,and in particular is ana-
lytic at : = 1. Furthermore:
ord
s=1
1(1,Q. :) = :.
There is also a more precise form of this conjecture, which expresses the leading
coecient of 1(1,Q. :) at : = 1 in terms of certain arithmetic invariants of
1,Q. For more details, see [Si1], conj. 16.5.
As we will explain in more detail in section 1.8, the analytic continuation
of 1(1,Q. :) now follows from the work of Wiles and Taylor-Wiles and a
strengthening by Diamond [Di2], for a very large class of elliptic curves over
Q, which includes all the semi-stable ones.
Abelian varieties: Elliptic curves admit higher-dimensional analogues, called
abelian varieties, which also play a role in our discussion. Analytically, the set
of complex points on an abelian variety is isomorphic to a quotient C
g
,,
where is a lattice in C
g
of rank 2p, satisfying the so-called Riemann period
relations. A good introduction to the basic theory of abelian varieties can be
found in [CS] and [We1].
1.2 Modular curves and modular forms over C
Modular curves: The group o1
2
(Z) of two by two integer matrices of deter-
minant one acts by fractional linear (Mobius) transformations on the complex
upper half plane
H = . C [ Im(.) 0.
equipped with its standard complex analytic structure. The principal con-
gruence group (`) of level ` is the subgroup of matrices in o1
2
(Z) which
reduce to the identity matrix modulo the positive integer `. A subgroup of
o1
2
(Z) is called a congruence group if it contains (`) for some `. The level
of is the smallest ` for which this is true. The most important examples of
congruence groups are:
The group
0
(`) consisting of all matrices that reduce modulo ` to an
upper triangular matrix.
22
The group
1
(`) consisting of all matrices that reduce modulo ` to a
matrix of the form
_
1
0 1
_
.
The principal congruence group (`) of level ` consisting of all matrices
that reduce modulo ` to the identity.
Notice the natural inclusions of normal subgroups (`)
1
(`)
0
(`).
The quotient
0
(`),
1
(`) is canonically isomorphic to (Z,`Z)

via
_
c /
c d
_
d mod `.
For any subgroup H of (Z,`Z)

, we denote by
H
(`) the group of matrices
in
0
(`) whose image in
0
(`),
1
(`) belongs to H.
If is a congruence subgroup of o1
2
(Z), dene 1

to be the quotient of
the upper half plane H by the action of . One equips 1

with the analytic

structure coming from the projection map : H 1

. (More precisely, if
= (), and G

is the stabilizer of in , then the local ring O

Y

,y
is
identied with the local ring of germs of holomorphic functions at which are
invariant under the action of G

.) This makes 1

into a connected complex

analytic manifold of dimension one, i.e., a Riemann surface. If is
0
(`) (resp.

1
(`), or (`)), we will also denote 1

by 1
0
(`) (resp. 1
1
(`), or 1 (`)). One
compacties 1

by adjoining a nite set of cusps which correspond to orbits

of P
1
(Q) = Q under . Call A

the corresponding compact Riemann

surface. (For more details, notably on the denition of the analytic structure
on A

at the cusps, see for example [Kn], p. 311, or [Shi2], ch. 1.) It follows
from the denition of this analytic structure that the eld 1

of meromorphic
functions on A

is equal to the set of meromorphic functions on H satisfying

(Transformation property): )() = )(), for all ;
(Behaviour at the cusps): For all o1
2
(Z), the function )() has a
Puiseux series expansion

m
c
n

n/h
in fractional powers of = c
2i
.
Riemanns existence theorem (cf. for example [For], ch. 2) asserts that the
analytic structure on A

comes from an algebraic one, i.e., the eld 1

is a
nitely generated extension of C of transcendence degree 1. Thus we can, and
will, view A

as a complex algebraic curve over C. If is

0
(`) (resp.
1
(`),
or (`)), we will also denote A

by A
0
(`) (resp. A
1
(`), or A(`)).
Examples and exercises:
23
1. For ` = 1, the curve A
0
(`) = A
1
(`) = A(`) is a curve of genus 0, and
its eld of functions is the ring C(,), where , is the classical modular function,
,() =
1
+ 744 + 196884 + . = c
2i
.
(Cf., for example, [Se4], ch. 7.)
2. For H P
1
(Q) =

H, dene G

to be the stabilizer of in 1o1

2
(Z),
and let c

= #(G

,(G

)). Show that c

depends only on the -orbit

of in

H, and that c

= 1 for all but nitely many in

H,. Using the
Riemann-Hurwitz formula (cf. [Ki], sec. 4.3) show that the genus of A

is
given by
p() = 1 [1o1
2
(Z) : ] +
1
2

1/
(c

1).
Use this to compute the genus of A
0
(j), A
1
(j), and A(j) for j prime. For
details, see [Shi2], sec. 1.6 or [Ogg].
3. For = (2), show that A

is isomorphic to P
1
, and that 1

is isomorphic
to P
1
0. 1. . Show that ,1) is the free group on the two generators
p
1
=
_
1 2
0 1
_
and p
2
=
_
1 0
2 1
_
.
4. Dene a homomorphism (2) Z,:Z Z,:Z, by sending p
1
to (1. 0)
and p
2
to (0. 1), and let denote its kernel. Show that is not in general a
congruence subgroup and that the curve 1

:= H, is birationally isomorphic
to the Fermat curve of degree : with ane equation r
n
+
n
= 1.
Moduli interpretations: The points in 1

= H, can be interpreted as
elliptic curves over C with some extra level ` structure. More precisely,
If =
0
(`), then the -orbit of H corresponds to the complex
torus 1 = C,1. ) with the distinguished cyclic subgroup of order `
generated by
1
N
. Thus, points on 1
0
(`) parametrize isomorphism classes
of pairs (1. C) where 1 is an elliptic curve over C and C is a cyclic
subgroup of 1 of order `.
If =
1
(`), then the -orbit of corresponds to the complex torus
1 = C,1. ) with the distinguished point of order ` given by
1
N
. Hence,
points on 1
1
(`) parametrize isomorphism classes of pairs (1. 1) where
now 1 is a point on 1 of exact order `.
Remark 1.10 One checks that the above rules set up a bijection between
points on 1

and elliptic curves with the appropriate structures, and that the
24
projection 1
1
(`) 1
0
(`) sending
1
(`) to
0
(`) becomes the forget-
ful map sending (1. 1) to (1. 1)).
Remark 1.11 (This remark will be used in section 1.3 when discussing Hecke
operators.) Dene an :-isogeny of -structures to be an :-isogeny of the
underlying elliptic curves which sends one -structure to the other. If j is a
prime not dividing `, then there are exactly j + 1 distinct j-isogenies from
(C,. 1).
1
N
), whose images are the pairs:
_
C,
_
+i
j
. 1
_
.
1
`
_
(i = 0. . . . . j 1).
_
C,j. 1).
j
`
_
.
If j divides `, then there are only j distinct j-isogenies from (C,. 1).
1
N
),
since (C,j. 1).
p
N
) is not a
1
(`)-structure (the point j,` not being of exact
order ` on the complex torus C,j. 1)).
Modular forms: Let / be an even positive integer. A modular form of weight
/ on is a holomorphic function ) on H satisfying:
(Transformation property): )() = (c +d)
k
)(), for all =
_
c /
c d
_

.
(Behaviour at the cusps): For all 1o1
2
(Z), the function (c +
d)
k
)() has a Puiseux series expansion

0
c
n

n/h
in fractional powers
of = c
2i
. We call

c
n

n/h
the Fourier expansion of ) at the cusp

1
(i).
A modular form which satises the stronger property that the constant
coecient of its Fourier expansion at each cusp vanishes is called a cusp form.
We denote by `
k
() the complex vector space of modular forms of weight /
on , and by o
k
() the space of cusp forms on . (For examples, see [DI],
sec. 2.2 and the references therein, especially, [Shi2], ch. 2.)
This article is mainly concerned with modular forms of weight 2, and hence
we will focus our attention on these from now on. A pleasant feature of the case
/ = 2 is that the cusp forms in o
2
() admit a direct geometric interpretation
as holomorphic dierentials on the curve A

.
Lemma 1.12 The map )()
f
:= 2i)()d is an isomorphism between
the space o
2
() and the space
1
(A

) of holomorphic dierentials on the curve

A

.
25
Sketch of proof: One checks that the transformation property satised by )()
under causes the expression )()d to be -invariant, and that the condition
of vanishing at the cusps translates into holomorphicity of )()d. (Note, for
example, that 2id = d,, so that
f
is holomorphic at i precisely when
)() vanishes at = 0.) 2
As a corollary, we nd:
Corollary 1.13 The space o
2
() is nite-dimensional, and its dimension is
equal to the genus p of A

.
Proof: This follows directly from the Riemann-Roch theorem, cf. [Ki], sec. 6.3.
2
To narrow still further the focus of our interest, we will be mostly concerned
with the cases =
0
(`) and
1
(`). A slightly more general framework is
sometimes convenient, so we suppose from now on that satises

1
(`)
0
(`).
Such a group is necessarily of the form
H
(`) for some subgroup H of
(Z,`Z)

. Because the transformation + 1 belongs to the forms

in o
2
() are periodic functions on H of period 1, and hence their Fourier
expansions at i are of the form
)() =

n>0
c
n

n
. = c
2i
. c
n
C.
The Petersson inner product: The spaces o
2
() are also equipped with a
natural Hermitian inner product given by
). p) =
i
8
2
_
X

f

g
=
_
1/
)() p()drd.
where = r +i. This is called the Petersson inner product.
The diamond operators: Suppose now that =
1
(`) and let d be an ele-
ment of (Z,`Z)

. The map d) which sends an elliptic curve with -structure

(1. 1) to the pair (1. d1) gives an automorphism of 1

which extends to A

.
It is called the diamond operator. For in H and =
_
c /
c d
_
in
0
(`), we
have
d)() = ().
26
Hence d) acts on o
2
(), identied with the holomorphic dierentials on A

,
by the rule
d))() = (c +d)
2
)
_
c +/
c +d
_
.
In geometric terms, the diamond operators are the Galois automorphisms of
the natural (branched) covering A
1
(`) A
0
(`) whose Galois group is
isomorphic to
0
(`),
1
(`)) = (Z,`Z)

,1). Given an even Dirichlet

character : (Z,`Z)

, say that ) is a modular form of level ` and

character if it belongs to the -eigenspace in o
2
(
1
(`)) under this action.
Let o
2
(`. ) denote the space of all such forms. Thus a function ) in o
2
(`. )
is a cusp form on
1
(`) which satises the stronger transformation property:
)
_
c +/
c +d
_
= (d)(c +d)
2
)(). for all
_
c /
c d
_

0
(`).
Note that if 1 is the trivial character, then o
2
(`. 1) is canonically identied
with o
2
(
0
(`)), which we will also denote by o
2
(`). Finally note the direct
sum decomposition:
o
2
(
1
(`)) =

o
2
(`. ).
where the sum ranges over all the even Dirichlet characters modulo `.
Exercise 1.14 Show that if )() belongs to o
2
(`), then )(c) belongs to
o
2
(:`), for each integer c dividing :.
Jacobians of modular curves: Let \ be the dual space
\ = o
2
()

:= Hom(o
2
(). C).
It is a complex vector space of dimension p = genus(A

). The integral ho-

mology = H
1
(A

. Z) maps naturally to \ by sending a homology cycle c

to the functional
c
dened by
c
()) =
_
c

f
. The image of is a lattice in
\ , i.e., a Z-module of rank 2p which is discrete (cf. [Mu1], cor. 3.8). Fix a
base point
0
H, and dene the Abel-Jacobi map
AJ
: A

(C) \, by

AJ
(1)()) =
_
P

f
. Note that this is well-dened, i.e., it does not depend on
the choice of path on A

from
0
to 1, up to elements in .
We extend the map
AJ
by linearity to the group Div (A

) of divisors on
A

, and observe that the restriction of

AJ
to the group Div
0
(A

) of degree
0 divisors does not depend on the choice of base-point
0
. Moreover we have
the Abel-Jacobi theorem:
27
Theorem 1.15 The map

AJ
: Div
0
(A

) \,
has a kernel consisting precisely of the group 1(A

) of principal divisors on
A

. Hence
AJ
induces an isomorphism from Pic
0
(A

) := Div
0
(A

),1(A

)
to \,.
For the proof, see [Mu1], ch. 3. The quotient \, is a complex torus, and is
equal to the group of complex points of an abelian variety. We denote this
abelian variety by J

, the Jacobian variety of A

over C. If =
0
(`) or

1
(`), we will also write J
0
(`) or J
1
(`) respectively for the Jacobian J

.
1.3 Hecke operators and Hecke theory
We maintain our running assumption that satises

1
(`)
0
(`).
If j is a prime not dividing the level `, we dene the Hecke operator 1
p
on
o
2
() by the formula
1
p
()) =
1
j
p1

i=0
)
_
+i
j
_
+jj))(j).
We give a more conceptual description of 1
p
in terms of remark 1.11 in the
case =
1
(`). We have

T
p
(f)
=

i
(
f
).
where
i
() =
+i
p
, and

() = j)j represent the j + 1 curves with -

structure that are images of (C,. 1).
1
N
) by a j-isogeny, and the

i
are the
pull-back maps on dierential forms on H. (An isogeny of elliptic curves with
-structure is simply an isogeny between the underlying curves which sends
one -structure to the other.) Such a description makes it evident that 1
p
())
belongs to o
2
(), if ) does. In terms of the Fourier expansion of ) =

c
n

n
,
the formula for the operator 1
p
on o
2
(`. ) is given by:
1
p
()) =

p[n
c
n

n/p
+j(j)

c
n

pn
.
28
If j divides `, then we dene the Hecke operator l
p
analogously, by summing
again over all the cyclic j-isogenies of -structures. Since there are only j of
them, the formula becomes simpler:
l
p
()) =
1
j
p1

i=0
)
_
+i
j
_
=

p[n
c
n

n/p
.
The reader is invited to check that the Hecke operators of the form 1
p
or
l
q
commute with each other, and also that they commute with the diamond
operators introduced in the previous section.
We extend the denition of the Hecke operators to operators 1
p
n, with
: 1, by the inductive formulae
1
p
n+1 = 1
p
1
p
n j)j1
p
n1. if (j. `) = 1.
and 1
p
n = l
n
p
otherwise. We then dene the operator 1
n
, where : =

j
e
i
i
is
written as a product of powers of distinct primes j
i
, by
1
n
=

i
1
p
e
i
i
.
This denition makes the Hecke operators multiplicative, i.e., 1
m
1
n
= 1
mn
if
(:. :) = 1. (A more conceptual denition of the Hecke operator 1
n
is that
1
n
()) is obtained by summing the pullback of
f
over the maps describing
all the cyclic :-isogenies of -structures.) The relations among the dierent
Hecke operators can be stated succinctly by saying that they obey the following
(formal) identity:

p,[N
(1 1
p
j
s
+j)j
12s
)
1

p[N
(1 l
p
j
s
)
1
=

n
1
n
:
s
.
(1.3.1)
The reader can consult [DI], sec. 3 and the references therein (especially [Shi2],
ch. 3 or [Kn]) for more details and dierent points of view on Hecke opera-
tors. Let T be the subring of End
C
(o
2
()) generated over C by all the Hecke
operators 1
p
for j,[`, l
q
for [`, and d) acting on o
2
().
Denition 1.16 A modular form ) is an eigenform if it is a simultaneous
eigenvector for all the Hecke operators in T, i.e., if there exists a C-algebra
homomorphism : T C such that 1) = (1)), for all 1 T.
A direct calculation shows that the coecients c
n
of an eigenform ) can be
recovered from the homomorphism by the formula:
c
n
()) = c
1
())(1
n
).
29
It follows that the rst Fourier coecient c
1
of a non-zero eigenform is always
non-zero, and that the non-trivial eigenspaces for T are all one-dimensional:
Proposition 1.17 Given a non-zero algebra homomorphism : T C,
there is exactly one eigenform ) up to scaling, which satises 1) = (1)),
for all 1 T.
Sketch of proof: The proof of the existence of ) is an exercise in commutative
algebra (localize o
2
() at the kernel of ), and the uniqueness is clear from
the formula above. 2
We call an eigenform satisfying c
1
= 1 a normalized eigenform.
Atkin-Lehner theory: It is natural to ask whether o
2
() can be decomposed
into a basis consisting of distinct normalized eigenforms. Unfortunately, this
is not always possible, as the following exercise illustrates.
Exercise 1.18 Suppose that j
3
divides ` exactly. Let T
t
be the algebra of
Hecke operators (generated by the operators 1
q
with ,[`,j
3
, and l
q
with
[`,j
3
) acting on o
2
(`,j
3
). Let ) =

n=1
c
n

n
be a T
t
-eigenform of level
`,j
3
in o
2
(`,j
3
). Show that the space o
f
spanned by the forms )(), )(j),
)(j
2
), and )(j
3
) is contained in o
2
(`), and is stable for the action of the
Hecke operators 1
q
, ,[`, and l
q
, [`. Show that o
f
has no basis of simulta-
neous eigenforms for the Hecke algebra T of level `, so that the action of T
on o
f
is not semi-simple.
Let T
0
denote the subalgebra of T generated only by the good Hecke oper-
ators 1
q
with ,[`, and d).
Proposition 1.19 If does not divide `, the adjoint of the Hecke operator
1
q
with respect to the Petersson scalar product is the operator )
1
1
q
, and the
adjoint of ) is )
1
. In particular, the Hecke operators commute with their
adjoints.
Proof: See [Kn], th. 9.18 and 8.22, or [Ogg]. 2
Proposition 1.19 implies, by the spectral theorem for commuting operators
that commute with their adjoints:
Proposition 1.20 The algebra T
0
is semi-simple (i.e, it is isomorphic to a
product C C of a certain number of copies of C), and there is a basis
of o
2
() consisting of simultaneous eigenvectors for the operators 1
q
.
30
Thus, T
0
has the merit of being semi-simple, while T is not in general. The
cost of replacing T by T
0
, however, is that one loses multiplicity one, i.e.,
the eigenspaces for T
0
need not be one-dimensional. For example, the space
o
f
dened in the previous exercise is contained in a single eigenspace for T
0
.
The theory of Atkin-Lehner [AL] gives essentially a complete understanding
of the structure of the algebra T, and the structure of the space of eigenforms.
To motivate the main result, observe that the problem in the exercise above
seems to be caused by forms of level ` that are coming from forms of lower
level `,j
3
by a straightforward operation, and are therefore not genuinely
of level `. They are the analogues, in the context of modular forms, of non-
primitive Dirichlet characters.
Denition 1.21 We dene the old subspace of o
2
() to be the space spanned
by those functions which are of the form p(c.), where p is in o
2
(
1
(`)) for
some ` < ` and c` divides `. We dene the new subspace of o
2
()
to be the orthogonal complement of the old subspace with respect to the
Petersson scalar product. A normalized eigenform in the new subspace is
called a newform of level `.
The following result is the main consequence of the theory of Atkin-Lehner. It
gives a complete answer to the question of what is the structure of the algebra
T acting on o
2
().
Theorem 1.22 If ) is in the new subspace of o
2
() and is an eigenvector for
all the operators in T
0
, then it is also an eigenform for T, and hence is unique
up to scaling. More generally, if ) is a newform of level `
f
[`, then the space
o
f
dened by
o
f
= p o
2
() such that 1p =
f
(1)p. for all 1 T
0

is stable under the action of all the Hecke operators in T. It is spanned by the
linearly independent forms )(c.) where c ranges over the divisors of `,`
f
.
Furthermore, we have
o
2
() =
f
o
f
.
where the sum is taken over all newforms ) of some level `
f
dividing `.
See [AL] for the proof in the case =
0
(`), and [La2], ch. VIII for the
general case. (See also [DI], sec. 6 for an overview.)
Exercise 1.23 Consider the case where =
0
(22). Show that A
0
(22) is of
genus 2, and hence that o
2
(22) has dimension 2. Show that o
2
(22) is equal
to o
f
, where ) = (()(11))
2
is a newform of level 11, so that in particular
there are no newforms of level 22 on . Show that T
0
is isomorphic to C in
this case, and that T is isomorphic to the semisimple algebra C C.
31
Action on homology and Jacobians: Note that the Hecke operators act
on \ = o
2
()

by duality. One checks (cf. [Kn], props. 11.23, 11.24) that the
sublattice of \ is stable under the action of all the Hecke operators 1
n
, and
of the diamond operators d). Therefore the operators 1
n
and d) give rise
to endomorphisms of the torus \,, and hence the Jacobian variety J

, in
a natural way. The involution gives rise to an involution on A

(C)
(which is complex conjugation on the model of A

over R deduced from the

Q- model dened in section 1.5). Since complex conjugation is continuous it
also acts on the integral homology = H
1
(A

(C). Z). Let

+
and

be the
sublattices of on which complex conjugation acts by +1 and 1. These are
sublattices of of rank p which are stable under the Hecke operators, since
complex conjugation commutes with the Hecke action.
A more algebraic description of the action of 1
p
on J

is given via the

notion of an algebraic correspondence. A correspondence on a curve A is a
divisor C on A A which is taken modulo divisors of the form 1 A and
A Q. Let
1
and
2
denote the projections of A A onto each factor.
Then the correspondence C induces a map on divisors of A, by setting
C(1) =
2
(
1
1
(1) C).
(For the denition of the intersection 1
1
1
2
of two divisors, see [We1].) The
map C preserves divisors of degree 0, and sends principal divisors on A to
principal divisors. It gives a well dened algebraic endomorphism of the Jaco-
bian variety Jac (A). Given a correspondence C, its transpose C

is dened to
be the divisor of A A obtained by interchanging the two factors of A A.
One can dene a natural notion of composition for correspondences, and the
set of correspondences forms a ring. The general theory of correspondences
and the proofs of the above facts are given in [We1], particularly the second
chapter.
The Hecke correspondence 1
n
is dened as the closure in A

of the
locus of points (. 1) in 1

such that there is a degree : cyclic isogeny of

elliptic curves with -structure from to 1. For example, if j is a prime not
dividing `, then 1
p
is an algebraic curve in A
1
(`)A
1
(`) which is birational
to A

1
(N)
0
(p)
. The induced map on divisors in this case satises
1
p
((1. 1)) =

(1,C. 1 mod C)
where the sum runs over the subgroups C of 1 having order j. Note also that
if (. 1) belongs to 1
p
, then the isogeny dual to 1 gives a j-isogeny
from 1 to j), so that
1

p
= j)
1
1
p
.
32
1.4 The 1-function associated to a cusp form
For this section, let ) in o
2
(
1
(`)) be a cusp form with Fourier expansion
at i given by

n
c
n

n
. One has the following estimate for the size of the
Fourier coecients c
n
:
Theorem 1.24 The coecients c
n
C satisfy the inequality
[c
n
[ c())
0
(:)

:.
where c()) is a constant depending only on ), and
0
(:) denotes the number
of positive divisors of :.
Sketch of proof: This follows from proposition 1.51 of section 1.7 which relates
the j-th Fourier coecients of eigenforms, for j a prime not dividing the level
of , to the number of points on certain abelian varieties over the nite eld
F
p
. The estimates of Hasse and Weil for the number of points on abelian
varieties over nite elds (stated in theorem 1.4 of section 1.1 for the special
case of elliptic curves; see [We1], IV for the general case) thus translate into
asymptotic bounds for the Fourier coecients of these eigenforms. We note
that the cruder estimate [c
n
[ = C(:), which is enough for the purposes of this
section, can be derived by a more elementary, purely analytic argument; cf.
[Ogg], ch. IV, prop. 16. 2
The 1-function associated to ) is dened by the formula:
1(). :) =

c
n
:
s
.
As in exercise 1.8, one can show that the innite sum dening 1(). :) converges
absolutely in the right half-plane Re (:)
3
2
. A much better insight is gained
into the function 1(). :) by noting that it is essentially the Mellin transform of
the modular form ). More precisely, if we set (). :) = `
s/2
(2)
s
(:)1(). :),
then we have
(). :) = `
s/2
_

0
)(i)
s
d, (1.4.1)
Exercise 1.25 Check the formula above.
This integral representation for 1(). :) gives the analytic continuation of
1(). :) to the entire complex plane. The modular invariance of ) translates
into a functional equation for 1(). :): more precisely, let u
N
be the Atkin-
Lehner involution dened by u
N
() = 1,`. The reader may check that
33
u
N
induces an involution of A

, and hence of
1
(A

) = o
2
(). One nds that
1(). :) satises the functional equation:
(). :) = (u
N
()). 2 :).
For a proof of this, see [Ogg], ch. V, lemma 1. Eigenforms for T in o
2
(`. )
have a great importance in the theory because their associated 1-functions
have an Euler product expansion, in addition to an analytic continuation and
functional equation:
Theorem 1.26 If ) =

c
n

n
is a normalized eigenform in o
2
(`. ) for all
the Hecke operators, then the 1-function 1(). :) =

c
n
:
s
has the Euler
product expansion

p,[N
(1 c
p
j
s
+(j)j
12s
)
1

p[N
(1 c
p
j
s
)
1
.
Proof: This follows directly from equation (1.3.1) of section 1.3. 2
If ) is a newform of level `, then it is also an eigenform for u
N
, so that the
functional equation may be viewed as relating 1(). :) and 1(). 2 :). We can
also state the following more precise version of theorem 1.24 (see lemma 3.2
of [Hi2] for example for parts (b), (c) and (d)).
Theorem 1.27 Suppose that ) is a newform of level `
f
and let `

denote
the conductor of its character .
(a) If j does not divide `
f
then [c
p
[ 2

j.
(b) If j[[`
f
and j does not divide `

then c
2
p
=
0
(j) where
0
is the
primitive character associated to .
(c) If j divides `
f
and j does not divide `
f
,`

then [c
p
[ =

j.
(d) If j
2
divides `
f
and j divides `
f
,`

then c
p
= 0.
1.5 Modular curves and modular forms over Q
Modular curves: For between
0
(`) and
1
(`), the modular curve A

has a model over Q. We describe such a model in the case of =

0
(`); the
construction for general follows from similar considerations.
The key remark here is that, as was noted in section 1.2, the complex
points on the curve 1
0
(`) have a natural interpretation as moduli of elliptic
34
curves together with a cyclic subgroup of order `, given by sending the point
H,
0
(`) to the pair (C,1. ). 1,`)).
Consider the universal elliptic curve
1
j
:
2
+r = r
3

36
, 1728
r
1
, 1728
.
It is an elliptic curve over the function eld Q(,), with ,-invariant ,. Let d be
the order of P
1
(Z,`Z), and let C
1
. . . . . C
d
denote the set of all cyclic subgroups
of 1
j
of order `, dened over Q(,), an algebraic closure of Q(,). Fix one of
these subgroups, C. The Galois group Gal (Q(,),Q(,)) permutes the C
i
in a
natural way. Let 1
N
be the smallest extension of Q(,) (viewed as embedded in
Q(,)) with the property that (C) = C, for all Gal (Q(,),1
N
). It can be
seen (cf. [Shi2], thm. 6.6) that the Galois action on the C
i
is transitive so that
1
N
,Q(,) is of degree d, and that it is a regular extension, i.e., 1
N


Q = Q.
Geometrically, 1
N
can be viewed as the function eld of a curve A
/Q
over Q,
with the inclusion of Q(,) corresponding to a map from A
/Q
to the projective
,-line over Q. The pair (1
j
. C) is an elliptic curve over Q(,) with a subgroup
of order ` dened over 1
N
. Using (1
j
. C), each complex point of A gives an
elliptic curve over C with a subgroup of order `, provided the point does
not lie over , = 0, 1728 or . The resulting map to A
0
(`) extends to an
isomorphism from A to A
0
(`). The curve A thus constructed, together with
this identication, is the desired model of A
0
(`) over Q.
More concretely, the functions , = ,() and ,
N
= ,(`) are related by a
polynomial equation
N
(,. ,
N
) = 0 with coecients in Q, of bidegree d. The
eld 1
N
is the function eld of the ane curve
N
(A. 1 ) = 0, and the mapping
(,(). ,(`)) gives a birational equivalence between H,
0
(`) and the
complex curve dened by the equation
N
. In practice it is not feasible to
write down the polynomial
N
, except for certain very small values of `. To
study the models over Q of A
0
(`), more indirect methods are needed, which
rely crucially on the moduli interpretation of A
0
(`). Similar remarks hold for
A
1
(`).
Models over Z: The work of Igusa [Ig], Deligne-Rapoport [DR], Drinfeld [Dr],
and Katz-Mazur [KM] uses the moduli-theoretic interpretation to describe a
canonical proper model for A

over Spec Z. These models allow us to talk

about the reduction of A

over nite elds F

p
, for j prime. The curve has
good reduction at primes j not dividing `, with the non-cuspidal points of
A
/F
p
corresponding to elliptic curves over

F
p
with -structure. The singular
bers at primes j dividing ` can also be described precisely; an important
special case (see [DR]) is that of
0
(`) with j exactly dividing `. For further
discussion and references, see [DI], sec. 8, 9.
35
From now on, when we write A

, A
0
(`), or A
1
(`), we will mean the curve
over Q which are the models described above for the complex curves dened
in section 1.2.
Remark 1.28 When considering -expansions, it is more convenient to use
a dierent set of models over Z for these complex curves. We dene A

1
(`)
in the case of
1
(`) as a model over Z which parametrizes pairs (1. i) where
i is an embedding of j
N
in the (generalized) elliptic curve 1. (So A

1
(`) is
the model denoted A

(`) in [DI], sec. 9.3, assuming ` 4.) For between

0
(`) and
1
(`), we dene A

as the corresponding quotient of A

1
(`). This
model has good reduction at primes j not dividing `, but unlike the models
mentioned above, its bers at primes j dividing ` are smooth and irreducible,
but not proper. In the case of =
0
(`), the curve A

,Q
can be identied
with A
0
(`). However, this is not the case in general: the cusp is a rational
point of A

,Q
but not necessarily of A

.
Jacobians: Weils theory [We1] of the Jacobian shows that the Jacobians J

dened in section 1.2 as complex tori also admit models over Q. When we
speak of J

, J
0
(`) and J
1
(`) from now on, we will refer to these as abelian
varieties dened over Q. Thus, the points in J

(1), for any Q-algebra 1, are

identied with the divisor classes on A

of degree 0, dened over 1.

We let J
/Z
, denote the Neron model of the Jacobian J

over Spec (Z).

Using this model we dene J
/A
for arbitrary rings . In particular we can
consider J
/F
p
, the reduction of the Jacobian in characteristic j, which is closely
related to the reduction of the integral model of the curve A

mentioned above.
In particular, if j does not divide the level of , then J
/F
p
can be identied
with the Jacobian of A
/F
p
. For a treatment of the case =
0
(`) with j
exactly dividing `, see the appendix of [Maz1]; for more general discussion
and references, see [DI], sec. 10, especially sec. 10.3.
Hecke operators: The Hecke operators have a natural moduli interpretation,
which was already touched upon in section 1.3. In particular, one nds that
the operator 1
n
arises from a correspondence which is dened over Q, and
gives rise to an endomorphism of the Jacobian J

which is dened over Q.

This in turn gives rise to an endomorphism of the Neron model J
/Z
, and we
can then consider the endomorphism 1
n
on the reduction of the Jacobian in
characteristic j. Recall that if j is a prime not dividing `, we may identify
this reduction with the Jacobian of A
/F
p
. (Cf. [MW], ch. 2, sec. 1, prop. 2.)
Furthermore, one can show that the moduli-theoretic interpretation of the
Hecke operator remains valid in characteristic j; i.e., the endomorphism 1
n
of
36
J
/F
p
is induced by a map on divisors satisfying, for all ordinary elliptic curves
with -structure:
1
n
() =

i
i().
where the sum is taken over all cyclic isogenies of degree :. (See [DI], sec. 10.2
for further discussion and references.)
This description allows one to analyse, for example, the Hecke operator
1
p
over F
p
, when (j. `) = 1. Let us work with =
1
(`), to illustrate
the idea. For a variety A over

F
p
, let
X
be the Frobenius morphism on A
dened by raising coordinates to the jth power. Thus if (1. 1) corresponds
to a point of A
1
(`)
/F
p
, then
E
is an isogeny of degree j from (1. 1) to the
pair (1

. 1

) =
X
1
(N)
(1. 1). The graph of
X
1
(N)
in (A
1
(`) A
1
(`))
/F
p
is
a correspondence of degree j, which we call 1. Let 1
t
be the transpose of this
correspondence. The endomorphism 1 of J

induced by 1 is the Frobenius

endomorphism
J

, and the endomorphism 1

t
is the dual endomorphism (in
the sense of duality of abelian varieties). Now consider the divisor
1
t
((1. 1)) = (1
1
. 1
1
) + + (1
p
. 1
p
).
where the (1
i
. 1
i
) are elliptic curves with -structure in characteristic j. Since

E
i
is an isogeny of degree j from (1
i
. 1
i
) to (1. 1), we also have the dual
isogeny from (1. 1) to (1
i
. j1
i
). If 1 is ordinary at j, then the points
(1

. 1

), (1
1
. j1
1
), . . . , (1
p
. j1
p
) are a complete list of the distinct curves
with -structure which are j-isogenous to (1. 1). Hence one has the equality
of divisors on A
1
(`)
/F
p
:
1
p
((1. 1)) = (1

. 1

) + (1
1
. j1
1
) + + (1
p
. j1
p
) = (1 +j)1
t
)((1. 1)).
Since the ordinary points are dense on A
1
(`)
/F
p
, we deduce that 1
p
= (1 +
j)1
t
) as endomorphisms of J
1
(`)
/F
p
. This equation, known as Eichler-
Shimura congruence relation, plays a central role in the theory. (For more
details, see [DI], sec. 10.2, 10.3.)
Theorem 1.29 If j,[` then the endomorphism 1
p
of J
/F
p
satises
1
p
= 1 +j)1
t
.
Remark 1.30 This was proved by Eichler [Ei] to hold for all but nitely many
j in the case of
0
(`), and by Shimura ([Shi1], see also [Shi2], ch. 7) in the
case of
1
(`). The fact that it holds for all j not dividing ` follows from
work of Igusa [Ig].
37
Modular forms: In the same way that modular curves have models over
Q and over Z, the Fourier coecients of modular forms also have natural
rationality and integrality properties. We start by sketching the proof of:
Theorem 1.31 The space o
2
() has a basis consisting of modular forms with
integer Fourier coecients.
Proof: The Hecke operators act on the integral homology
+
in a way that
is compatible with the action on o
2
() and respects the natural (Poincare)
duality between these two spaces. Hence, if
n

nN
is a system of eigenvalues
for the 1
n
, then the
n
are algebraic integers in some nite extension 1 of Q,
and the system

nN
is a system of eigenvalues for the 1
n
for any Galois
automorphism of

Q,Q. Hence, we have shown:
Proposition 1.32 If ) o
2
(`. ) is a newform of some level ` dividing
`, then its Fourier coecients lie in a nite extension 1 of Q. Moreover,
if Gal (

Q,Q) is any Galois automorphism, then the Fourier series )

obtained by applying to the Fourier coecients is a newform in o

2
(`.

).
The explicit description of o
2
() given in section 1.3 implies that o
2
() is
spanned by forms having Fourier coecients which are algebraic integers in
some nite (Galois) extension 1 of Q, and that the space of forms with Fourier
coecients in 1 is stable under the natural action of Gal (1,Q) on Fourier
expansions. An application of Hilberts theorem 90 shows that o
2
() has a
basis consisting of forms with rational Fourier expansions, and the integrality
of the Fourier coecients of eigenforms yields the integrality statement of
theorem 1.31. 2
We dene o
2
(. Z) to be the space of modular forms with integral Fourier
coecients in o
2
(). Theorem 1.31 states that o
2
(. Z) C = o
2
(). Given
any ring , we dene
o
2
(. ) = o
2
(. Z) .
and dene o
2
(`. ) and o
2
(`. . ) (where now is a character with values
in

) in the obvious way. If is contained in C, the -expansion principle

below allows us to identify o
2
(. ) with the set of modular forms in o
2
()
with Fourier coecients in .
The -expansion principle: Because the modular curve A
0
(`) has a model
over Q, the space of modular forms o
2
(`) =
1
(A
0
(`)) has a natural rational
structure, given by considering the dierential forms on A
0
(`) dened over
Q. The fundamental -expansion principle (see [DR], ch. 7, or [Kat], sec. 1.6)
says that these algebraic structures are the same as those obtained analytically
38
by considering -expansions at . More generally, using the model A

, we
obtain the -expansion principle over Z for cusp forms on (cf. [DI], sec. 12.3).
Theorem 1.33 The map o
2
()
1
(A

) dened by )
f
induces an
isomorphism from o
2
(. Z) to
1
(A

). Furthermore, if is at over Z or ` is
invertible in , then the induced map o
2
(. )
1
(A

,A
) is an isomorphism.
Furthermore, if is a subring of C, then this isomorphism identies o
2
(. )
with set of modular forms in o
2
() having coecients in .
1.6 The Hecke algebra
It follows directly from the formulas for the Hecke operators acting on -
expansions that the 1
n
leave o
2
(
0
(`). Z) stable, as well as the subspace
of o
2
(`. ) with coecients in Z[]. Using the -expansion principle (the-
orem 1.33), one can also show ([DI], sec. 12.4) that the diamond operators
preserve the spaces of cusp forms on with integral Fourier expansions, and
hence that the space o
2
(. Z) is preserved by all the Hecke operators.
We dene T
Z
to be the ring generated over Z by the Hecke operators 1
n
and d) acting on the space o
2
(. Z). More generally, if is any ring, we dene
T
A
to be the -algebra T
Z
. This Hecke ring acts on the space o
2
(. )
in a natural way. Before studying the structure of the Hecke rings T
A
as we
vary the rings , we note the following general result (Cf. [Shi2], ch. 3.):
Lemma 1.34 The space o
2
(. )

= Hom
A
(o
2
(. ). ) is a free T
A
-module
of rank one.
Sketch of proof: One checks that the pairing T
Z
o
2
(. Z) Z dened by
(1. )) c
1
(1)) sets up a perfect, T
Z
-equivariant duality between T
Z
and
o
2
(. Z). The result for arbitrary follows. 2
Hecke rings over C: If = C, then the structure of the ring T = T
C
is completely described by theorem 1.22. More precisely, if T
C,f
denotes the
image of the Hecke algebra acting on the space o
f
dened in section 1.3, then
T
C
=
f
T
C,f
.
where the direct sum ranges over all distinct newforms ) of some level `
f
dividing `. Furthermore, the algebra T
C,f
can be described explicitly. In
particular, if ) is a newform of level ` then T
C,f
is isomorphic to C, but if `
f
is not equal to ` then the ring T
C,f
need not be a semi-simple algebra over C.
Lemma 1.34 in the case = C says that \ = o
2
()

is a free T
C
-module
of rank one, but we also have:
39
Lemma 1.35 The module o
2
() is a free T
C
-module of rank one.
Proof: Let p
1
. . . . . p
t
be a complete system of newforms of levels `
1
, . . . , `
t
dividing `. One can check that the form
p = p
1
(`,`
1
) + +p
t
(`,`
t
)
generates o
2
() as a T
C
-module. The map T
C
o
2
() dened by 1 1p
gives an isomorphism from T
C
to o
2
(), as T
C
-modules. 2
Remark 1.36 Lemmas 1.34 and 1.35 imply that T
C
is a Gorenstein C-algebra
(of nite rank), i.e., Hom
C
(T
C
. C) is isomorphic to T
C
as a T
C
-module.
Hecke rings over Q: Let [)] be the Galois orbit (under the action of G
Q
=
Gal (

Q,Q)) of a normalized newform ) of some level `

f
dividing `, and let
1
f
be the eld extension of Q generated by the Fourier coecients of ). The
space
g[f]
o
g
is a vector space of dimension [1
f
: Q]
0
(`,`
f
), which is
spanned by modular forms with rational Fourier coecients. Let o
[f]
be the
Q-subspace of forms in
g[f]
o
g
with rational Fourier coecients. The space
o
[f]
is stable under the action of T
Q
, and letting T
Q,[f]
be the image of T
Q
acting on o
[f]
, we obtain the direct sum decomposition
T
Q
=
[f]
T
Q,[f]
.
where the sum is taken over the distinct G
Q
-orbits of normalized newforms
) of some level `
f
dividing `. If `
f
is equal to `, then the algebra T
Q,[f]
is isomorphic to the eld 1
f
. If `
f
is a proper divisor of `, then, as in the
complex case, the algebra T
Q,[f]
is a (not necessarily semi-simple) algebra over
Q of rank
0
(`,`
f
)[1
f
: Q]. The nature of the elds 1
f
, and in general the
structure of T
Q
, is very poorly understood at this stage; for example, one does
not know how to characterize the number elds that occur as a 1
f
for some
) (but they are all known to be totally real or CM elds).
The ring T
Q
acts naturally on the rational homology H
1
(A

. Q) = Q,
and we have
Lemma 1.37 The module Q is free of rank two over T
Q
.
Sketch of proof: The modules
+
C \ and

C \ are free of rank

one over T
C
, by lemma 1.34. This implies that
+
Q and

Q are both
free of rank one over T
Q
. 2
40
Hecke rings over Z: The ring T
Z
is a certain (not necessarily maximal) order
in T
Q
. One still has an injection
T
Z

[f]
T
Z,[f]
.
where now T
Z,[f]
denotes the ring generated over Z by the Hecke operators
acting on o
[f]
. Of course the structure of T
Z
is even more mysterious than
that of T
Q
! The ring T
Z
acts naturally on , but it is not the case in general
that is free of rank two over T
Z
, i.e., that the integral analogue of lemma 1.37
is true. (See remark 1.42 below.)
Hecke rings over Q

: The study of the algebras T

Z

and T
Q

arises naturally
because of the Hecke action on the Tate module T

(J

),
T

(J

) := lim

(J

)[/
n
].
where the inverse limit is taken with respect to the multiplication by / maps.
The action of T
Z

on T

(J

) is compatible with that of G

Q
, and it is this pair of
actions on the Tate module which is used to associate two-dimensional Galois
representations to modular forms.
It will sometimes be more convenient to consider the ring T
Q

and its action

on
1 = T

(J

)
Z

.
We rst record a useful duality property enjoyed by the Tate modules. The
Weil pairings on the groups J

[/
n
] for : 1 induce a perfect pairing
. ) : T

(J

) T

(J

) Z

.
Since each Hecke operator 1 is adjoint to u1u where u = u
N
is the Atkin-
Lehner involution, we have the following lemma.
Lemma 1.38 The map r
x
where
x
() = r. u) denes an isomor-
phism of T
Z

-modules,
T

(J

)

= T

(J

= Hom
Z

(T

(J

). Z

).
and hence an isomorphism of T
Q

-modules
1

= 1

= Hom
Q

(1. Q

).
The following lemma allows us to regard T
Q

as a coecient ring for a

two-dimensional Galois representation
Gal (

Q,Q) Aut
T
Q

(1)

= G1
2
(T
Q

).
41
Lemma 1.39 The module 1 is free of rank 2 over T
Q

.
Proof: Lemma 1.37 implies directly that the module 1 = Q

is free of rank
two over T
Q

. 2
Corollary 1.40 The ring T
Q

is a Gorenstein Q

-algebra; i.e.,
T

= Hom
Q

(T
Q

. Q

)
is free of rank one over T
Q

.
Proof: Choosing a basis for 1 over T
Q

, we obtain an isomorphism
T
Q

T
Q

= T

.
Decomposing T
Q

as

i
1
i
where each factor 1
i
is a nite-dimensional local
Q

-algebra, we obtain an isomorphism

1
i
1
i

= 1

i
1

i
for each i. At least one of the four maps 1

i
1
i
deduced from this
isomorphism must be surjective, and by counting dimensions, we see that it
must be injective as well. It follows that T
Q

is isomorphic to T

. 2
Recall that for primes j not dividing `, the Jacobian J

has good reduction

mod j, and the Eichler-Shimura relation, theorem 1.29, states that on J
/F
p
,
we have
1
p
= 1 +j)1
t
.
For primes j not dividing `/, we may identify T

(J

) with the /-adic Tate

module of the reduction (see [ST]) and consider the Frobenius endomorphism
1 on the free rank two T
Q

-module 1. As a consequence of the Eichler-Shimura

relation, we nd:
Theorem 1.41 For j not dividing `/, the characteristic polynomial of 1 on
the T
Q

-module 1 is
A
2
1
p
A +j)j.
Proof: (We are grateful to Brian Conrad for showing us this argument.) Since
11
t
= j, it follows from the Eichler-Shimura relation that
1
2
1
p
1 +j)j = 0.
To conclude that this is in fact the characteristic polynomial, it suces to
compute the trace of 1. To do so, we use the T
Q

isomorphism
1 1

42
dened by the modied pairing . u). Under this modied pairing, 1 is
adjoint to u1
t
u = j)1
t
, so the trace of 1 on 1 is the same as that of
(j)1
t
) on 1

. Choosing bases for 1 and T

, one sees that this is

the same as the trace of j)1
t
. Hence
2 tr 1 = tr 1 + tr (j)1
t
) = tr (1
p
) = 2 1
p
.
2
Hecke rings over Z

: The ring T
Z

is free of nite rank over Z

. It therefore
decomposes as
T
Z

T
m
.
where the product runs over the maximal ideals m of T
Z

, and T
m
is the
localization of T
Z

at m. For each m, T
m
is a complete local Z

-algebra, free
of nite rank as a Z

-module.
Remark 1.42 While the analogue of lemma 1.39 does not always hold for
T
Z

(see [MRi], sec. 13) we shall see that it holds for certain localizations T
m
.
Results of this type are much deeper than lemma 1.39 and were rst obtained
by Mazur [Maz1], sec. 14, 15. We shall return in chapter 4 to explain Mazurs
result and its generalizations, which play a role in the arguments of [W3] and
[TW].
Example 1.43 The curve A
0
(19) has genus 1, and A
0
(57) has genus 5. By
consulting the tables in the Antwerp volume [Ant4] or Cremonas book [Cr],
one nds that there is exactly one newform of level 19, and that there are
three newforms of level 57, which all have rational Fourier coecients. Their
Fourier coecients c
p
, for the rst few primes j, are listed in the following
table:
2 3 5 7 11 13 17 19 23 29 31
19 0 2 3 1 3 4 3 1 0 6 4
57 2 1 3 5 1 2 1 1 4 2 6
571 1 1 2 0 0 6 6 1 4 2 8
57C 2 1 1 3 3 6 3 1 4 10 2
Setting ) = 19, we nd that a basis of simultaneous eigenforms in o
2
(57. C)
for the Hecke operators 1
p
, | ,= 3. 19, and l
3
, l
19
is:
)() + (1 +

2))(3). )() + (1

2))(3). 57. 571. 57C.

It appears from the table that the Fourier coecients corresponding to the
forms 571 and 57C are congruent modulo 3. This is in fact the case. One
43
nds that the Hecke ring T
Z
3
generated over Z
3
by the Hecke operators acting
on o
2
(57. Z
3
) is isomorphic to the subalgebra of Z
5
3
:
(r. . .. t. u) such that t u (mod 3).
The isomorphism sends 1
p
(for j ,= 3. 19) to the element
(c
p
(19). c
p
(19). c
p
(57). c
p
(571). c
p
(57C)).
It sends l
3
to (1 +

2. 1

2. 1. 1. 1) and l
19
to (1. 1. 1. 1. 1).
Thus there are exactly 4 distinct maximal ideals of T
Z
3
, and the localizations
at these maximal ideals are the rings Z
3
, Z
3
, Z
3
, and
T
m
= (t. u) such that t u (mod 3).
where m is the ideal generated by 3 and 1
n
c
n
(571) for all :.
1.7 The Shimura construction
Let ) =

c
n

n
be an eigenform on with (not necessarily rational) Fourier
coecients, corresponding to a surjective algebra homomorphism
f
: T
Q

1
f
, where 1
f
is the eld generated over Q by the Fourier coecients of ). We
briey review in this section a construction of Shimura ([Shi2], ch. 7) which
associates to ) (or rather, to the orbit [)] of ) under G
Q
) an abelian variety

f
dened over Q and of dimension [1
f
: Q].
Let 1
f
T
Z
be the ideal ker (
f
) T
Z
. The image 1
f
(J

) is a (connected)
subabelian variety of J

which is stable under T

Z
and is dened over Q.
Denition 1.44 The abelian variety
f
associated to ) is the quotient

f
= J

,1
f
J

From this denition one sees that

f
is dened over Q and depends only on
[)], and that its endomorphism ring contains T
Z
,1
f
which is isomorphic to an
order in 1
f
.
Remark 1.45 Using theorem 1.22, one can show that J
0
(`) is isogenous to

[f]

0
(N/N
f
)
[f]
.
We now describe the abelian variety
f
as a complex torus. Let \
f
be the
subspace of \ = o
2
()

on which T acts by
f
. Theorem 1.22 and lemma 1.34
show that \
f
is a one-dimensional complex vector space. Let
f
be the or-
thogonal projection of \ to \
f
, relative to the Petersson scalar product. The
projector
f
belongs naturally to T
K
f
.
44
Let [)] be the set of all eigenforms whose Fourier coecients are Galois
conjugate to those of ). The number of forms in [)] is equal to the degree d
of 1
f
over Q. Now, we set
\
[f]
=
g[f]
\
g
.
[f]
=

g[f]

g
.
Note that
[f]
is simply the orthogonal projection of \ to \
[f]
. Note also that

[f]
belongs to the Hecke algebra T
Q
.
Lemma 1.46 The abelian variety
f
is isomorphic over C to the complex
torus \
[f]
,
[f]
(), with the map
[f]
: \, \
[f]
,
[f]
() corresponding to
the natural projection from J

to
f
.
In particular, one sees that
f
is an abelian variety of dimension d = [1
f
: Q].
Hence if ) has rational Fourier coecients, then the abelian variety
f
is an
elliptic curve. This elliptic curve is called the strong modular elliptic curve
associated to ) if also ) is a newform of level ` and =
0
(`).
Example 1.47 If =
0
(26), one checks that the genus of A

is two, and
that there are two distinct normalized eigenforms )
1
and )
2
in o
2
(26). From
the tables in [Ant4] or [Cr], one sees that )
1
and )
2
have integral Fourier
coecients, whose values for the primes 31 are:
2 3 5 7 11 13 17 19 23 29 31
)
1
1 1 3 1 6 1 3 2 0 6 4
)
2
1 3 1 1 2 1 3 6 4 2 4
Hence the abelian varieties
f
1
and
f
2
are elliptic curves. The above table
suggests (and this can be checked directly by looking at the equations for these
curves given in [Ant4] or [Cr], or by using the discussion in [DO], lemma 2.1)
that the Fourier coecients of )
1
and )
2
are congruent modulo 2. The natural
projection J
0
(26)
f
1

f
2
is an isogeny whose kernel is isomorphic to
Z,2ZZ,2Z, and J
0
(26) is not isomorphic to
f
1

f
2
. More generally, one
knows that a Jacobian can never decompose as a non-trivial direct product of
two principally polarized abelian varieties, cf. [Maz1], prop. 10.6. (There are
no non-zero homomorphisms from
f
1
to
f
2
, and so a non-trivial product
decomposition of J

would have to induce a decomposition as a product of

principally polarized abelian varieties.)
Let T

(
f
) be the Tate module of the abelian variety
f
,
T

(
f
) := lim

(
f
)[/
n
].
45
where the inverse limit is taken with respect to the multiplication by / maps.
This module is naturally a module for T
f
Z

, and T

(
f
) Q

is a module
for 1
f
Q

as well.
Lemma 1.48 The module T

(
f
)Q

is a free module of rank 2 over 1

f
Q

.
Proof: This follows directly from lemma 1.39. 2
Proposition 1.49 The algebra End
Q
(
f
) Q is isomorphic to 1
f
. In par-
ticular
f
is simple over Q.
The proof is given in [R2], cor. 4.2. The main ingredient is the irreducibility of
the Galois representation attached to
f
as in section 3.1. (Cf. theorem 3.1.)
Properties of
f
: good reduction:
Theorem 1.50 If ) is an eigenform of level `, and j is a prime that does
not divide `, then the abelian variety
f
has good reduction at j.
Proof: This follows from the fact that J

has good reduction at such primes,

which in turn is a consequence of the good reduction of A

. 2
If j is a prime not dividing `, it then becomes natural to study the number
`
f,p
of points on the abelian variety
f
over the nite eld F
p
. It is given by
the following formula:
Proposition 1.51 The number of points `
f,p
is given by the formula
`
f,p
= Norm
K
f
/Q
(
f
(1 c
p
()) +j)j)).
where c
f
(j) 1
f
is the j-th Fourier coecient of ).
Proof: By Weils theory [We1], this number is given by the determinant
`
f,p
= det(1 1).
where 1 is the Frobenius endomorphism acting on the /-adic Tate module
T

(
f
). So theorem 1.41 implies
det(1 1) = Norm
K
f
/Q
(
f
(1 1
p
+j)j)).
and proposition 1.51 follows. 2
Dening the local Hasse-Weil 1-function of
f
over F
p
by the formula
1(
f
,F
p
. :) = det(1 1j
s
)
1
.
46
the proof above gives the formula:
1(
f
,F
p
. :) =

1
p
()

. :).
where the product is taken over all complex embeddings : 1
f
C, and
1
p
()

. :) is the Euler factor at j in the 1-function that was associated to )

in section 1.4.
In particular, if
f
= 1 is an elliptic curve, i.e., ) has rational Fourier
coecients and ) is on
0
(`), then the number of points on 1 over F
p
is
given by the formula
#1(F
p
) = j + 1 c
p
()). (1.7.1)
Properties of
f
: bad reduction: A fundamental quantity associated to
any abelian variety over Q is its arithmetic conductor, which measures the
amount of bad reduction of . If we factor this conductor as a product of
prime powers,

p
j
m
p
, then the exponents :
p
are equal to 0 precisely when
has good reduction at j. The denition of :
p
() was already given in
section 1.1 when is an abelian variety of dimension 1, i.e., an elliptic curve.
In general, the exponent :
p
coincides with :
p
(
A,
) (see section 2.1 below),
where
A,
is the Galois representation on the /-adic Tate module of .
Regarding the bad reduction of the abelian variety
f
, one has the following
consequence of the results of Langlands, Deligne and Carayol [Ca3] discussed
below in section 3.1.
Theorem 1.52 If ) is a newform of level `, then the conductor of the abelian
variety
f
is equal to `
g
.
1.8 The Shimura-Taniyama conjecture
Let 1 be any elliptic curve dened over Q, and let ` denote its arithmetic
conductor, dened as in section 1.1. Then we have:
Proposition 1.53 The following are equivalent:
(a) The curve 1 is isogenous over Q to
f
, for some newform ) on some
congruence group .
(b) The curve 1 is isogenous over Q to
f
, for a newform ) on
0
(`).
(c) There is a non-constant morphism dened over Q, from A
0
(`) to 1.
47
Sketch of proof: The implication (/) (c) is immediate, and (c) (/) follows
from the work of Carayol [Ca3] (cf. proposition 3.20 below). If (/) holds, then
there is a surjective map J
0
(`) 1. Composing with the Abel-Jacobi map
A
0
(`) J
0
(`) gives a map to 1 which is not constant, by the denition of
the Jacobian, and hence (c) holds. Conversely, a non-constant map of curves
: A
0
(`) 1 induces, by Albanese (covariant) functoriality, a surjective
map of Jacobians

: J
0
(`) 1, where we have identied 1 with its own
Jacobian in the natural way. Since J
0
(`) is isogenous to a product (possibly
with multiplicities) of abelian varieties
f
where ) run over newforms of level
`
f
dividing ` (cf. remark 1.45), it follows that there is a surjective map

f
1 for some
f
. Since
f
is simple (proposition 1.49), this map is an
isogeny, and part (/) follows. 2
We call an elliptic curve over Q satisfying the equivalent properties above
a modular elliptic curve. A startling conjecture that was rst proposed by
Taniyama in the 1950s and made more precise by Shimura predicts that the
Shimura construction explained in the previous section is surjective:
Conjecture 1.54 All elliptic curves dened over Q are modular.
This conjecture is now known to be true for a very wide class of elliptic curves.
The results of Wiles [W3] and Taylor-Wiles [TW] imply it is true for all semi-
stable elliptic curves, and a strengthening of the method, [Di2], implies the
conjecture for all elliptic curves which have semi-stable reduction at 3 and 5.
Remark 1.55 See [DI], sec. 13 for a more thorough list of equivalent forms
of the conjecture.
Relation with 1-functions: Dene numbers c
p
= c
p
(1) as in section 1.1.
Recall the (global) Hasse-Weil 1-function dened in section 1.1 by equations
(1.1.2), (1.1.4), (1.1.5), and (1.1.6). The following proposition reveals some of
the importance of the Shimura-Taniyama conjecture:
Proposition 1.56 If 1 is modular, i.e., is associated to a newform ) by the
Shimura construction, then
1(1,Q. :) = 1(). :).
In particular, 1(1,Q. :) has an analytic continuation to the entire complex
plane, and a functional equation.
48
Sketch of proof: If 1 is isogenous to an elliptic curve
f
associated to a
newform ) on
0
(`) by the Shimura construction, then the two 1-functions
1(
f
,Q. :) and 1(1,Q. :) are equal. On the other hand, formula (1.7.1)
directly implies that 1(
f
,Q. :) is equal to 1(). :), at least up to nitely
many Euler factors (corresponding precisely, by the work of Igusa, to the
primes dividing `). The equality of all the Euler factors follows from the
work of Deligne, Langlands, and Carayol (cf. [Ca1]). 2
Knowing the analytic continuation and functional equation of the Hasse-
Weil 1-function of an elliptic curve is of great importance in the theory. For
example, the conjecture of Birch and Swinnerton-Dyer (which was stated in
a weak form in section 1.1, conjecture 1.9) relates arithmetic invariants of 1
such as the rank of its Mordell-Weil group and the order of its Shafarevich-
Tate group to the behaviour at : = 1 of 1(1,Q. :). (Note that : = 1 is
outside the domain of absolute convergence of the innite product used to
dene 1(1,Q. :).)
One consequence of this conjecture is that 1(Q) is nite if 1(1,Q. 1) is
non-zero. This was proved by Coates and Wiles in [CW] for elliptic curves with
complex multiplication, and by Kolyvagin [Kol] for modular elliptic curves.
(Recently, a dierent proof for modular elliptic curves has been announced by
K. Kato.) Thanks to the breakthroughs of [W3], [TW], and [Di2], the results
of Kolyvagin and Kato are now unconditional for a very large class of elliptic
curves.
The Shimura-Taniyama conjecture for abelian varieties: We may view
conjecture 1.54 as asserting that the map
_
Newforms of weight 2 on A
0
(`)
with rational Fourier coecients
_

_
Isogeny classes of elliptic curves
over Q of conductor N
_
provided by the Shimura construction is a bijection. It is natural to extend
this result to all eigenforms, not just those having rational Fourier expansions.
We say that an abelian variety dened over Q is modular if it is isogenous to
an abelian variety of the form
f
for some newform ) on
1
(`). It would
not be reasonable to expect that all abelian varieties over Q are modular: the
abelian varieties arising from the Shimura construction are very special, in
many respects. Most importantly, they have a very large ring of endomor-
phisms. Following Ribet [R7], we make the denition:
Denition 1.57 An abelian variety over Q of dimension p is said to be of
G1
2
-type if its endomorphism ring over Q contains an order in a eld of
degree p over Q.
Then we have the following generalized Shimura-Taniyama conjecture:
49
Conjecture 1.58 Every simple abelian variety over Q which is of G1
2
-type
is modular.
This generalization of the Shimura-Taniyama conjecture is far from being
proved, and tackling it still seems to require some major new ideas, even
after the ideas introduced by Wiles.
2 Galois theory
2.1 Galois representations
If 1 is a perfect eld, we will let

1 denote an algebraic closure of 1 and
G
F
= Gal (

1,1) its absolute Galois group. Recall that G
F
is a pronite group:
more precisely, G
F
= lim

Gal (1,1) as 1 runs over nite Galois extensions of

1 contained in

1. Hence G
F
carries a natural topology. If / is a prime
dierent from the characteristic of 1 we will let c

: G
F
Z

denote the
/-adic cyclotomic character, i.e. if is an /-power root of unity in

1 then
() =

()
for all G
F
. We will simply write c if the choice of / is clear
from the context.
The nite elds F
p
: The group G
F
p
is isomorphic to

Z, the pronite com-
pletion of Z. A natural topological generator is provided by the Frobenius
element Frob
p
, the automorphism of

F
p
which raises elements of

F
p
to their
j
th
power. If / is dierent from j, then c

(Frob
p
) = j.
The j-adic elds Q
p
: The j-adic valuation
p
: Q

p
Z extends uniquely
to a valuation
p
:

Q

p
Q. Let O
Q
p
denote the ring of integers of

Q
p
and
m
Q
p
its maximal ideal. The residue eld O
Q
p
,m
Q
p
is an algebraic closure of
F
p
, which we will identify with

F
p
. The valuation
p
is compatible with the
action of G
Q
p
, and hence O
Q
p
and m
Q
p
are stable under G
Q
p
. In particular
G
Q
p
acts on O
Q
p
,m
Q
p
and we obtain a map : G
Q
p
G
F
p
, which is in fact
a surjection. We call the kernel the inertia group and denote it 1
p
. It has a
unique maximal pro-j subgroup: the wild inertia group, denoted 1
p
. There is
a canonical isomorphism
1
p
,1
p

=

,=p
Z

(1).
where the (1) indicates that if ) G
Q
p
,1
p
is a lifting of Frobenius and if
1
p
,1
p
then ))
1
=
p
.
The inertia group 1
p
is ltered by a series of subgroups called the higher
ramication groups. More precisely, there are closed normal subgroups 1
u
p
in
50
G
Q
p
for any n [1. ], with the following properties:
If n then 1
u
p
1
v
p
, i.e., the 1
u
p
form a decreasing ltration.
For n 0, 1
u
p
= 1
p
while 1

p
= 1.
1
p
=

u>0
1
u
p
.
1
u
p
=

v<u
1
v
p
.
These groups are dened in section 3 of chapter IV of [Se2], where they are
denoted G
u
Q
p
(except that there, G
1
Q
p
= G
Q
p
whereas we have 1
1
p
= 1
p
).
The local Kronecker-Weber theorem ([Se2], ch. 14, sec. 7, for example)
asserts that the map
c
p
: G
ab
Q
p
G
F
p
Z

p
.
is an isomorphism. (We will use G
ab
to denote the abelianisation of a pronite
group G, i.e. the unique maximal abelian continuous image.) Under it 1
p
goes
to Z

p
and for n 0, 1
u
p
goes to (1+j
u|
Z
p
) Z

p
, where n| denotes the least
integer greater than or equal to n. If / ,= j then c

is trivial on 1
p
and takes
Frob
p
to j Z

.
The eld Q: If j is a rational prime, the usual j-adic valuation
p
: Q

Z
extends to a valuation :

Q

Q. This extension is not unique, but all

extensions are permuted transitively by G
Q
. Fix one such extension
p
and
let G
p
denote its stabiliser in G
Q
. Then G
p
acts on the completion (

Q)
v
p
and preserves the subeld of elements algebraic over Q
p
. This eld is an
algebraic closure of Q
p
, which we shall identify with

Q
p
. One can check that
the resulting map G
p
G
Q
p
is an isomorphism. Thus we obtain subgroups
G
p
1
p
1
u
p
for n [1. ] and a distinguished element Frob
p
G
p
,1
p
.
These objects depend on the choice of
p
and vary by conjugation in G
Q
as
this choice is varied. We call an algebraic extension 1,Q unramied at j if
all conjugates of 1
p
lie in G
F
G
Q
; otherwise we say that 1,Q is ramied.
If 1,Q is Galois and unramied at j, then there is a well-dened conjugacy
class [Frob
p
] Gal (1,Q).
By replacing the j-adic completion by an Archimedean completion, one
has a well-dened conjugacy class [c] in G
Q
consisting of those elements that
arise as complex conjugation for some embedding

Q C. We will denote by
G

the subgroup 1. c for one such element c.

We have the following fundamental results concerning the structure of G
Q
(see for instance [La1]).
51
Theorem 2.1 If 1,Q is a nite extension then 1 is only ramied at nitely
many primes (those dividing the discriminant of 1,Q).
Theorem 2.2 (Hermite-Minkowski) If o is a nite set of primes and if
d Z
>0
then there are only nitely many extensions 1,Q of degree d which
are unramied outside o (contained in a xed algebraic closure

Q).
Theorem 2.3 (Chebotarev) If 1,Q is a Galois extension unramied out-
side a nite set o of primes then

p,S
[Frob
p
] is dense in Gal (1,Q).
Theorem 2.4 (Kronecker-Weber) The product of the j-adic cyclotomic
characters gives an isomorphism

p
c
p
: G
ab
Q

p
Z

p
.
We remark that the Chebotarev density theorem (theorem 2.3) applied
to the extension of Q generated by the :
th
roots of unity implies Dirichlets
theorem that if : and : are coprime integers then there are innitely many
primes j with j : mod :. We also remark that the Kronecker-Weber
theorem is equivalent to the fact that the maximal abelian extension of Q is
obtained by adjoining all roots of unity to Q.
Representations: A d-dimensional representation of G
Q
is a homomorphism
G
Q
G1
d
(1).
where 1 is any eld. (In our later discussion, we will also consider repre-
sentations with coecients in a ring.) Often 1 (and hence G1
d
(1)) comes
equipped with a natural topology, and it is then customary to restrict ones
attention to continuous homomorphisms G
Q
G1
d
(1).
Since any one-dimensional representation has abelian image, theorem 2.4
allows one to give a complete description of the one-dimensional representa-
tions of G
Q
together with the behaviour of these representations on the de-
composition groups at all primes. The aim of this article is to discuss attempts
to give a similar theory for the two-dimensional representations of G
Q
.
We will call a representation of G
Q
unramied at j if it is trivial on the
inertia group 1
p
. Otherwise we say that it is ramied at j. If is unramied
at j, then (Frob
p
) is well-dened (and its conjugacy class is independent of
the choice of
p
).
If is a representation of a group G and i is a non-negative integer, then
we let
i
denote the representation of G on the i
th
exterior power of the
52
underlying module of . If H is a subgroup of G, we will let
H
(resp.
H
)
denote the H-invariants (resp. H-coinvariants) of the underlying module of
. If H is normal in G, then we shall also use
H
and
H
to denote the
corresponding representation of G,H. In particular if j is a prime and is a
representation of G
Q
or G
Q
p
, then
I
p
(Frob
p
) and
I
p
(Frob
p
) are well-dened.
We will be primarily interested in three types of representation of G
Q
.
Artin representations, i.e. continuous representations G
Q
G1
d
(C).
Since all compact totally disconnected subgroups of G1
d
(C) are nite,
Artin representations have nite image. Hence they are semi-simple and
are unramied at all but nitely many primes, by theorem 2.1.
Mod / representations, i.e. continuous representations G
Q
G1
d
(/),
where / is a nite eld of characteristic /. These always have nite image
and hence, like Artin representations, are unramied at all but nitely
many primes.
/-adic representations, i.e. continuous representations G
Q
G1
d
(1),
where 1 is a nite extension of Q

. We require that an /-adic represen-

tation be unramied at all but nitely many primes.
Remark 2.5 (a) Continuous representations G
Q
G1
d
(1), unlike those
to G1
d
(C) or G1
d
(/), may be ramied at innitely many primes. We
shall not need to consider such representations. (One rarely if ever does.)
However, the term /-adic representation is often used in the more
general sense elsewhere in the literature.
(b) Note that the image of an /-adic representation can very well be in-
nite. For instance this is the case if d = 1 and the representation is the
cyclotomic character.
(c) Note that in the case of mod / and /-adic representations, the represen-
tations need not be semi-simple.
Proposition 2.6 Let o be any nite set of primes.
(a) An Artin representation : G
Q
G1
d
(C) is determined by the values
of tr (Frob
p
) on the primes j , o at which is unramied.
(b) A semi-simple mod / representation : G
Q
G1
d
(/) is determined by
the values of tr
i
(Frob
p
) (i = 1. .... d) on the primes j , o at which
is unramied. If / d then it is determined by tr (Frob
p
) on the
primes j , o at which is unramied.
53
(c) A semi-simple /-adic representation : G
Q
G1
d
(1) is determined by
the values of tr (Frob
p
) on the primes j , o at which is unramied.
Proof: Combining the Chebotarev density theorem (theorem 2.3) with the
continuity of we see that we must show that tr (resp. tr
i
) determines
up to conjugacy in the various settings. For the characteristic zero represen-
tations, see [Bour], ch. 8, sec. 12.1, prop. 3; for mod / representations, this is
the Brauer-Nesbitt theorem ([CR], (30.16)). 2
Let 1 denote a nite extension of Q

, let O be its ring of integers, the

maximal ideal of O and / the residue eld. If : G
Q
G1
d
(1) is an /-adic
representation, then the image of is compact, and hence can be conju-
gated to a homomorphism G
Q
G1
d
(O). Reducing modulo the maximal
ideal gives a residual representation : G
Q
G1
d
(/). This representation
may depend on the particular G1
d
(1)-conjugate of chosen, but its semi-
simplication
ss
(i.e. the unique semi-simple representation with the same
Jordan-Holder factors) is uniquely determined by , by proposition 2.6 (b).
Note that the kernel of the reduction map G1
d
(O) G1
d
(/) is a pro-
/-group. In particular we see that if j ,= / then (1
p
) is nite and (1
p
) is
isomorphic to (1
p
).
Suppose is a representation of G
Q
of one of the three sorts above. If j ,= /
then we dene the conductor, :
p
(), of at j by
:
p
() =
_

1
codim
I
u
p
dn = codim
I
p
+
_

0
codim
I
u
p
dn.
This is well-dened as (1
p
) is a nite group. If is an Artin representation
this makes sense also for j = /. (It would even make sense for a mod /
representation if j = /, but in this case it does not seem to be a useful notion.)
It is known that :
p
() is an integer (see chapter VI of [Se2]). Moreover it is
easily seen that :
p
() = 0 if and only if is unramied at j. We dene the
conductor `() of to be

p
j
m
p
()
.
where the product is over all primes j ,= / in the case of an /-adic or mod /
representation and over all primes in the case of an Artin representation. This
makes sense because is unramied almost everywhere.
The following lemma is an exercise.
Lemma 2.7 Suppose that : G
Q
G1
d
(1) is an /-adic representation and
that : G
Q
G1
d
(/) is a reduction of . Then for each prime j ,= / we have
54
_

0
codim
I
u
p
dn =
_

0
codim
I
u
p
dn. Thus
`() = `( )

p,=
j
(dim
I
p
dim
I
p
)
.
and in particular `( ) divides `().
We take this opportunity to introduce some important notation. Let G
denote a group and 1 a ring. If : G G1
d
(1) is a representation we
shall let `

denote the underlying 1[G]-module (so that `

= 1
d
as an 1-
module). If ` is an 1[G]-module we shall let End (`) denote the module of
1-linear endomorphisms of `. It is also an 1[G]-module with the G-action
being dened by
(p())(:) = p((p
1
:)).
If ` is a nitely generated free 1-module we will let End
0
(`) denote the
sub-1[G]-module of End (`) consisting of endomorphisms of trace zero. We
will use ad to denote End (`

) and ad
0
to denote End
0
(`

). Note that if
d is invertible in 1 then ad

= ad
0
1 as 1[G]-modules, where 1 has the
trivial action of G. Note also then that the kernel of ad
0
is the same as the
kernel of the composite map G

G1
d
(1) 1G1
d
(1).
We also remark that if 1 is an algebraically closed eld of characteristic
other than 2 and : G G1
2
(1) is irreducible, then either ad
0
is irreducible
or there is a subgroup H G of index 2 and a character : H 1

such that

= Ind
G
H
. In the latter case ad
0

= Ind
G
H
(,
t
), where is the nontrivial
character of G,H and
t
is the composite of and conjugation by an element
of G H. Moreover, either Ind
G
H
(,
t
) is irreducible or ad
0

=
1

2

3
where the
i
are distinct characters of G of order 2. In the latter case, is
induced from a character from each of the subgroups H
i
= ker
i
of index 2 in
G. We leave the verication of these facts as an exercise to the reader.
2.2 Representations associated to elliptic curves
Perhaps the simplest examples of non-abelian /-adic representations arise from
elliptic curves dened over Q.
We will let 1[:](

Q) denote the group of :-torsion points on 1(

Q). By
proposition 1.1 there is a non-canonical isomorphism 1[:](

Q)

= (Z,:Z)
2
.
Furthermore, 1[:](

Q) carries a natural action of G

Q
and so we get a represen-
tation (dened up to conjugation)

E,n
: G
Q
G1
2
(Z,:Z).
55
If / is a prime dierent from the characteristic of Q then we set T

1 =
lim

1[/
n
](

Q), which is non-canonically isomorphic to Z

2

. Again it has a nat-

ural continuous action of G
Q
, and so we get a representation (dened up to
conjugation)

E,
: G
Q
G1
2
(Z

).
Note that
E,
n

= (
E,
mod /
n
).
Global properties: We have the following basic properties of the represen-
tations
E,
n and
E,
:
Proposition 2.8 (a) The determinant of
E,
is c

.
(b) The representation
E,
is absolutely irreducible for all /. For xed 1,

E,
is absolutely irreducible for all but nitely many /.
(c) If 1 does not have complex multiplication then
E,
(and hence
E,
) is
surjective for all but nitely many /.
Proof: Part (a) follows from the existence of the non-degenerate alternating
Galois-equivariant Weil pairing
T

1 T

1 Z

(1) := lim

n.
Part (b) is proved in [Se5], ch. IV. Part (c) is the main result of [Se6]. 2
The following deep result, which is stronger than the second assertion of
part (b) of proposition 2.8, is due to Mazur [Maz2], thms. 1 and 2. (Parts (b)
and (c) can actually be deduced directly from theorem 1.7, using the fact that
a semi-stable curve with a rational subgroup of order / is necessarily isogenous
to a curve with a rational point of order /.)
Theorem 2.9 Suppose that 1,Q is an elliptic curve.
(a) If / 163 is a prime then
E,
is irreducible.
(b) If 1 is semistable then
E,
is irreducible for / 7.
(c) If 1 is semistable and
E,2
is trivial then
E,
is irreducible for / 3.
Remark 2.10 Combined with the results of Serre [Se6], Mazurs results imply
that if 1 is semistable everywhere then
E,
is surjective for / 7 ([Maz2],
thm. 4).
Local behaviour of
E,
and
E,
:
56
Proposition 2.11 Suppose 1 has good reduction at j.
(a) If / ,= j, then
E,
is unramied at j, and we have the formula
tr
E
,

(Frob
p
) = j + 1 #

1
p
(F
p
).
In particular tr
E,
(Frob
p
) belongs to Z and is independent of / ,= j.
(b) For all : 1 there is a nite at group scheme T
n
,Z
p
such that
1[j
n
](

Q
p
)

= T
n
(

Q
p
)
as G
p
-modules.
(c) If 1 has good ordinary reduction at j (i.e., c
p
is not divisible by j)
then

1
p
[j](

F
p
) has order j and

E,p
[
I
p

_
c
p

0 1
_
;
If 1 has supersingular reduction at j, then

1
p
[j](

F
p
) is trivial and

E,p
[
G
p
is irreducible.
Proofs: The proof of part (a) can be found in chapters V and VII of [Si1].
For part (b) one considers the nite at group scheme T
n
= c[j
n
], where c
is the model for 1 over Z
p
dened by (\
min
). (See [Sha] for an introduction
to theory of nite at group schemes.) Part (c) follows from the results in
chapters V and VII of [Si1] together with basic results on nite at group
schemes.
Proposition 2.12 Suppose that 1 has multiplicative reduction at j. Let =

E,p
jZ
p
be the multiplicative Tate period attached to 1, dened as in sec-
tion 1.1. Let : G
p
,1
p
1 be the unique non-trivial unramied quadratic
character of G
p
if 1 has non-split multiplicative reduction, and let be the
trivial character if 1 has split multiplicative reduction. Then
(a)

E,
[
G
p

_
c

0 1
_
.
and if / ,= j then :
p
(
E,
) = 1.
(b)

E,
[
G
p

_
c

0 1
_
.
and H
1
(G
Q
p
. Z,/Z(1))

= Q

p
,(Q

p
)

corresponds to the Tate period

E,p
Q

p
,(Q

p
)

.
57
(c) If / ,= j then :
p
(
E,
) = 0 (i.e.,
E,
is unramied at j) if and only if
/[
p
(
min
E
) =
p
(,
E
).
(d) There is a nite at group scheme T,Z
p
such that 1[j](

Q
p
)

= T(

Q
p
) if
and only if j[
p
(
min
E
) =
p
(,
E
).
Sketch of proof: By Tates proposition 1.5 of section 1.1, we have
1(

Q
p
)

= (

p
,
Z
)()
as G
p
-modules. (The () indicates that G
p
acts on (

p
,
Z
)() by : r
(r)
()
.) The proposition follows directly from this, together with [Edi], prop.
8.2. for the last part. (For further details, see ch. VII, prop. 5.1 of [Si1], ch.
IV, thm. 10.2. and ch. V of [Si2], and sec. 2.9 of [Se7].) 2
Proposition 2.13 Suppose that 1 has additive reduction at j. Then for any
/ ,= j the conductor of
E,
[
G
p
is at least 2, and if j 3 then it is equal to 2.
Remark 2.14 For an elliptic curve 1 with any type of reduction at j, the
conductor of
E,
[
G
p
(for / ,= j) coincides with the local conductor of 1 at j
given by a formula of Ogg (proved by Saito [Sa] in the case j = 2). In particular
the conductor of
E,
[
G
p
is independent of / ,= j. For further discussion and
references, see ch. IV, secs. 10 and 11 of [Si2].
The Frey curve: We now consider the elliptic curve
1
A,B
: 1
2
2 = A(A 2)(A +12).
where and 1 are non-zero coprime integers with + 1 ,= 0. Set C =
+ 1. This equation has discriminant = 16(1C)
2
. If j is an odd prime
then this curve has good reduction at j if j,[(1C) and has multiplicative
reduction at j if j[1C. Thus
min
E
A,B
= 2
412n
(1C)
2
for some : Z
0
.
If for instance 1 mod 4 and 1 0 mod 16 then 1
A,B
has a minimal
Weierstrass equation
1
2
2 +A1 2 = A
3
+
1 1
4
A
2
2 (1,16)A2
2
.
In this special case 1
A,B
has semi-stable reduction everywhere and
min
E
A,B
=
2
8
(1C)
2
.
The connection with Fermats Last Theorem is provided by the elliptic
curve
1 : 1
2
2 = A(A c

2)(r +/

2)
58
considered by Hellegouarch [He] and Frey [Fr], where c

+ /

= c

is a hy-
pothetical non-trivial solution to Fermats Last theorem, and where we have
supposed without loss of generality that c, / and c are pairwise coprime, that
/ is even and that c 1 mod 4. Frey suggested that properties of this hy-
pothetical elliptic curve 1 could lead to a contradiction if the curve were also
known to be modular. This suggestion was made precise by Serre [Se7], who
observed that the following list of properties would yield the desired contradic-
tion, when combined with his conjectures on Galois representations associated
to modular forms (to be discussed in section 3.2).
Theorem 2.15 Suppose that / 3 is prime. The representation
E,
has the
following properties.
(a)
E,
is irreducible.
(b)
E,
is unramied outside / and 2.
(c) There is a nite at group scheme T,Z

such that T(

)

= 1[/](

) as
G

-modules.
(d) #
E,
(1
2
) = /.
Proof: Part (a) follows from Mazurs theorem (theorem 2.9, (c)). To prove (b),
(c), and (d), we note the key fact that
min
E
= 2
8
(c/c)
2
is a perfect /-th power,
up to powers of 2. The fact that
E,
is unramied at j ,= /. 2 (part (b)) follows
from this, from proposition 2.11, (a) (for primes j ,= / of good reduction for 1)
and from proposition 2.12 (c) (for primes j ,= /. 2 of multiplicative reduction
for 1). Part (c) is a consequence of proposition 2.12 (d), and part (d) follows
from proposition 2.12 (c). 2
Remark 2.16 Part (a) implies that the image of
E,
is large, while parts (b)
and (c) state that the image has very limited ramication. We remark that
the conclusion of the theorem ensures that :
p
(
E,
) = 0 for j ,= 2. / and that
:
2
(
E,
) = 1.
2.3 Galois cohomology
In this section ` will denote a continuous discrete G
Q
-module of nite cardi-
nality. If ` and ` are two such modules we will endow the space Hom(`. `)
of homomorphisms of abelian groups with an action of G
Q
via the formula
(p())(:) = p((p
1
:)).
59
We will use `

to denote Hom(`. j
n
(

Q)), where : is an integer such that

:` = (0) and where j
n
(

Q) denotes the group of :

th
roots of unity in

Q.
We will use to denote Pontryagin duals. All cohomology groups of pro-
nite groups will mean continuous cohomology. The general references for this
section are Milne [Mi] and Serre [Se1].
We start by recalling Tates local duality theorem (see chapter 1, section 2
of [Mi]).
Theorem 2.17 Let be a place of Q (i.e. either a prime number or ).
(a) H
i
(G
v
. `) is nite for all i.
(b) For all integers : there are compatible embeddings H
2
(G
v
. j
n
(

Q))
Q,Z (i.e. compatible with the maps coming from j
n
(

Q) j
m
(

Q) if
:[:).
(c) For ,= and i = 0. 1. 2 the cup product and the above embeddings give
rise to a perfect pairing
H
i
(G
v
. `) H
2i
(G
v
. `

) H
2
(G
v
. j
n
(

Q)) Q,Z.
(This also holds when = , if one replaces H
0
(G
v
. `) = H
0
(R. `) by
Tates modied cohomology group:

H
0
(R. `) := `
=1
,(1 +)`, where
is complex conjugation.)
(d) If ,= is a prime, then H
i
(G
v
. `) = (0) for i 2 and
#H
1
(G
v
. `) = #H
0
(G
v
. `)#H
2
(G
v
. `)#(` Z
v
).
(e) If ,= and ` Z
v
= 0 then H
1
(G
v
,1
v
. `
I
v
) and H
1
(G
v
,1
v
. `
I
v
)
are the annihilators of each other under the pairing
H
1
(G
v
. `) H
1
(G
v
. `

) Q,Z.
By a collection of local conditions for ` we shall mean a collection L =
1
v
of subgroups 1
v
H
1
(G
v
. `) as runs over all places of Q, such that
1
v
= H
1
(G
v
,1
v
. `
I
v
) for all but nitely many . Note that if L is a collection
of local conditions for ` then L

= 1

v
is a collection of local conditions for
`

. If L is a collection of local conditions for ` we dene the corresponding

Selmer group, H
1
/
(Q. `), to be the subgroup of r H
1
(G
Q
. `) such that for
all places of Q we have
res
v
r 1
v
H
1
(G
v
. `).
60
Suppose that is a nite place of Q. We have an exact sequence
(0) H
0
(G
v
. `) `
I
v
Frob
v
1
`
I
v
H
1
(G
v
,1
v
. `
I
v
) (0).
and hence we see that #H
1
(G
v
,1
v
. `
I
v
) = #H
0
(G
v
. `).
We have the following important observation of Wiles [W3], inspired by a
formula of Greenberg [Gre]. An important theme in number theory has been
the calculation of Selmer groups. This result, although it does not allow the
absolute calculation of Selmer groups, allows the comparison of two (dual)
Selmer groups. In the applications in this paper we shall apply the theorem
where the various data have been chosen to make one of the Selmer groups
trivial. In such a situation it allows the exact calculation of the order of the
non-trivial Selmer group.
Theorem 2.18 If L is a collection of local conditions for ` then the Selmer
group H
1
/
(Q. `) is nite. Moreover we have the formula
#H
1
/
(Q. `)
#H
1
/
(Q. `

)
=
#H
0
(G
Q
. `)
#H
0
(G
Q
. `

v
#1
v
#H
0
(G
v
. `)
.
We note that all but nitely many terms in the product are 1 so that it makes
sense.
Proof of theorem 2.18: Choose a nite set, o, of places of Q, which contains ,
all the places whose residue characteristic divides the order of `, all places
at which ` is ramied and all places at which 1
v
,= H
1
(G
v
,1
v
. `
I
v
). Let
Q
S
denote the maximal extension of Q unramied outside o and let G
S
=
Gal (Q
S
,Q). Then we have an exact sequence
(0) H
1
/
(Q. `) H
1
(G
S
. `)

vS
H
1
(G
v
. `),1
v
.
Because H
1
(G
S
. `) is nite (see cor. 4.15 of ch. 1 of [Mi]), we see that
H
1
/
(Q. `) is nite.
Dualising this sequence for `

and using theorem 2.17, we see that we

have an exact sequence

vS
1
v
H
1
(G
S
. `

H
1
/
(Q. `

(0).
On the other hand we have the following part of the Poitou-Tate nine term
sequence (see thm. 4.10 of ch. 1 of [Mi]):
(0) H
0
(G
S
. `)
_
vS
H
0
(G
v
. `)
_
,(1 +c)` H
2
(G
S
. `

H
1
(G
S
. `

vS
H
1
(G
v
. `) H
1
(G
S
. `).
61
where we regard (1 +c)` as contained in H
0
(G

. `). Replacing H
1
(G
v
. `)
by 1
v
for o and combining this with the previous exact sequence we get
(0) H
0
(G
S
. `)
_
vS
H
0
(G
v
. `)
_
,(1 +c)`

vS
1
v
H
1
/
(Q. `) H
2
(G
S
. `

H
1
(G
S
. `

H
1
/
(Q. `

(0).
Theorem 2.18 follows from this and the global Euler characteristic formula
(see thm. 5.1 of ch. 1 of [Mi]):
#H
1
(G
S
. `

)
#H
0
(G
S
. `

)#H
2
(G
S
. `

)
=
(#`

)
#H
0
(G

. `

)
= #(1 +c)`.
2
We remark, but will not use elsewhere in this article, that the above argu-
ment works over any number eld:
Theorem 2.19 Let 1 be a number eld and ` a discrete G
F
module of nite
cardinality. For each place of 1 let G
v
denote a decomposition group at
and if is nite let 1
v
G
v
denote the inertia group. Fix subgroups 1
v

H
1
(G
v
. `), and such that for all but nitely many , 1
v
= H
1
(G
v
,1
v
. `
I
v
).
Let H
1
/
(1. `) (respectively, H
1
/
(1. `

)) denote the inverse image of

v
1
v
(respectively,

v
1

v
) under the map H
1
(G
F
. `)

v
H
1
(G
v
. `) (respec-
tively, H
1
(G
F
. `

)

v
H
1
(G
v
. `

)). Then H
1
/
(G. `) and H
1
/
(G. `

)
are nite and we have the formula
#H
1
/
(1. `)
#H
1
/
(1. `

)
=
#H
0
(G
F
. `)
#H
0
(G
F
. `

v
#1
v
#H
0
(G
v
. `)
.
2.4 Representations of G
Q

In this section we assume that / is an odd prime. If G is any topological

group then by a nite O[G]-module we shall mean a discrete O-module of
nite cardinality with a continuous action of G. By a pronite O[G]-module
we shall mean an inverse limit of nite O[G]-modules.
If ` is a pronite O[G

]-module then we will call `

good if for every discrete quotient `
t
of ` there is a nite at group
scheme T,Z

such that `
t

= 1(

) as Z

[G

]-modules;
62
ordinary if there is an exact sequence
(0) `
(1)
` `
(0)
(0)
of pronite O[G

]-modules such that 1

acts trivially on `
(0)
and by c
on `
(1)
(equivalently, if for all . 1

we have ( c())( 1) = 0
on `);
semi-stable if ` is either good or ordinary.
Suppose that 1 is a complete Noetherian local O-algebra with residue eld /.
We will call a continuous representation : G

G1
2
(1) good, ordinary or
semistable, if
det [
I

= c (2.4.1)
and if the underlying pronite O[G

]-module, `

is good, ordinary or semi-

stable. We write for mod m
R
. We record the following consequence of
Nakayamas lemma.
Lemma 2.20 If `

and are ordinary, then `

(1)

and `
(0)

are each free

of rank one over 1 and is ordinary.
Remark 2.21 (a) These denitions are somewhat ad hoc, but at the mo-
ment that is all that seems to be available (though the work of Fontaine
and Laaille [FL] and its generalisations may well provide a more sys-
tematic setting).
(b) For part of the motivation for our denitions, see proposition 2.23 and
remark 2.24. For further motivation, we shall see in the next chapter
that representations arising from certain modular forms are semistable.
Moreover the Fontaine-Mazur conjecture predicts that, conversely, any
representation of G
Q
with semistable restriction to G

arises from such

a modular form. We shall state a weak form of the Fontaine-Mazur
conjecture below (conjecture 3.17).
(c) The terminology in [W3] to describe representations of G

is slightly
dierent. In particular, we impose the condition (2.4.1) in our denitions
of good and ordinary, as this is all we shall need here. Assuming
satises this condition, the notion of ordinary in [W3] coincides with the
one here, and is at in the sense of [W3] if and only if it is good but not
ordinary. (Note that a representation may be both good and ordinary,
for instance
E,
for an elliptic curve with good, ordinary reduction; see
proposition 2.23.)
63
The following assertions, in the good case, are consequences of results of
Raynaud [Ray] (in particular, see sec. 2.1, prop. 2.3.1 and cor. 3.3.6 of [Ray]).
In the ordinary case they are elementary.
Lemma 2.22 (a) Stable pronite O[G

]-modules are closed under taking

sub-objects, quotients and direct products. The same is true for ordi-
nary pronite O[G

]-modules.
(b) Suppose that ` is a pronite O[G

]-module and that `

i
is a family of
sub-objects with trivial intersection, such that each `,`
i
is good (resp.
ordinary). Then ` is good (resp. ordinary).
(c) If ` is a nite O[G

]-module then ` is good if and only if there is a

nite at group scheme T,Z

such that `

= T(

) as Z

[G

]-modules.
(d) If ` and `
t
are pronite O[G

]-modules with `
t

= ` as Z

[1

]-modules
then ` is good (resp. ordinary) if and only if `
t
is good (resp. ordinary).
(e) Suppose that ` is a pronite O[G

]-module which is nite and free over

O. Then ` is good if and only if there exists an /-divisible group T,Z

such that ` is isomorphic to the Tate module of T as a Z

[G

]-module.
Together with results stated in section 2.2, we have the following.
Proposition 2.23 Suppose that 1 is an elliptic curve over Q and O = Z

.
if 1 has good (resp. semistable) reduction at /, then
E,
and
E,
are
good (resp. semistable);
if 1 has semistable reduction at /, then
E,
is ordinary if and only if

E,
is ordinary if and only if either 1 has multiplicative reduction or 1
has good ordinary reduction.
Remark 2.24 It is also true that if
E,
is good (resp. semistable), then 1
has good (resp. semistable) reduction at /, but the result is more dicult and
we shall not need it.
We will need a few more denitions. We will let : 1

denote
the character (), mod where =

2
1

/. If 1 is a eld of
characteristic other than / and if ` is a pronite Z

[G

]-module then we
set `(1) = lim

`
Z

n(

1). If : G

G1
2
(1) is ordinary the extension
(0) 1(1) `

1 (0)
64
of 1[1

]-modules gives rise to a class c

H
1
(1

. 1(1)). Kummer theory and

the valuation on (

Q
I

give rise to a map : H

1
(1

. 1(1)) H
Z

1 1,
where H denotes the /-adic completion of (

Q
I

. Then we also have the

following lemma.
Lemma 2.25 (a) If : G

G1
2
(/) is good then either is ordinary or
[
I

/ =

.
(b) If : G

G1
2
(1) is such that `

is good and is ordinary, then is

good and ordinary.
(c) If : G

G1
2
(1) is ordinary then is good if and only if (c

) = 0.
Note that we need only consider the case that 1 has nite cardinality to
prove the lemma. Parts (a) and (b) again follow from Raynauds results [Ray]
(for part (b) consider the connected-etale sequences for `

and `

).
We sketch the proof of part (c) using an argument suggested to us by
Edixhoven. As in [Edi], prop. 8.2 it suces to determine which extensions
(0) 1(1) ` 1 (0) (2.4.2)
of 1[1

]-modules arise from nite at group schemes over Z

nr

, the ring of
integers of the maximal unramied extension of Z

. By prop. 17.4 of [Oo],

the extension (2.4.2) arises from a nite at group scheme if and only if it
corresponds to an extension of sheaves of 1-modules in the fpqc topology over
Z
nr

. Therefore we must compute the image of

Ext
1
Rmod/fpqc
(1. 1(1)) Ext
1
R[I

]
(1. 1(1)).
Since the sheaf Ext
1
of 1 by 1(1) vanishes, this is equivalent to computing
the image of
H
1
fpqc
(Z
nr

. 1(1)) H
1
(1

. 1(1)).
Part (c) follows from the fact this is precisely the kernel of .
Remark 2.26 The authors expect that the theory of Fontaine and Laaille
[FL] discussed in the next section could be used to prove that if : G

G1
2
(1) is such that `

is good and is good, then det [

I

is cyclotomic, i.e.,
is good.
For semistable representations : G

G1
2
(O,
n
) we shall dene O-
submodules
H
1
f
(G

. ad) H
1
ss
(G

. ad) H
1
(G

. ad).
65
Before doing so, let us consider more generally a continuous representation
of a pronite group G with values in G1
d
(O,
n
). For each continuous cocycle
: G ad, we dene a representation

: G G1
2
(1
n
)
p (1 +(p))(p)
where 1
n
= (O,
n
)[],(
2
). We nd that the map

induces a bijec-
tion between H
1
(G. ad) and equivalence classes of representations
t
: G
G1
2
(1
n
) such that =
t
mod 1
n
, where
t
1
and
t
2
are deemed equivalent
if they are conjugate by an element of 1 + `
2
(1
n
). Now let `

denote the
(O,
n
)[G]-module corresponding to , and for each continuous cocycle , let
1

denote the 1
n
[G]-module corresponding to

. Note that multiplication by

denes an isomorphism from `

= 1

,1

to 1

of (O,
n
)[G]-modules.
We thus obtain an extension
0 `

0
and hence a class in Ext
1
(`

. `

) in the category of pronite (O,

n
)[G]-
modules. Moreover if

1
and

2
are equivalent, then 1

1
and 1

2
dene the
same extension class. We thus obtain a map
H
1
(G. ad) Ext
1
(`

. `

).
which the reader can check is an (O,
n
)-linear isomorphism. Classes in
H
1
(G. ad
0
) correspond to the equivalence classes of representations
t
sat-
isfying det
t
= det , and in Hom(G. O,
n
) corresponds to the twist of by
the character p 1+(p). In the case /,[d where we have the decomposition
H
1
(G. ad) = H
1
(G. ad
0
) Hom(G. O,
n
).
we see that if
t
corresponds to a class in H
1
(G. ad), then the projection to
Hom(G. O,
n
) is the homomorphism such that det
t
= (1 +d) det .
We now return to the case of a semistable representation
: G

G1
2
(O,
n
)
and dene the cohomology groups H
1
f
(G

. ad) and H
1
ss
(G

. ad) as follows.
Let H
1
ss
(G

. ad) denote the natural image in H

1
(G

. ad) of Ext
1
(`

. `

)
taken in the category of semi-stable pronite O,
n
[G

]-modules. If is not
good then H
1
f
(G

. ad) is dened to be H
1
ss
(G

. ad). If, however, is good

then H
1
f
(G

. ad) will denote the natural image in H

1
(G

. ad) of the group

66
Ext
1
(`

. `

) taken in the category of good pronite O,

n
[G

]-modules. We
dene H
1
f
(G

. ad
0
) (resp. H
1
ss
(G

. ad
0
)) as the intersection of H
1
f
(G

. ad)
(resp. H
1
ss
(G

. ad)) and H
1
(G

. ad
0
). Note that if is good (resp. semistable)
and : G

ad is a cocycle, then 1

is good (resp. semistable) if and only

if

is.
Suppose that : G

G1
2
(O,
n
) is ordinary. Consider the exact se-
quence
0 `
(1)

`
(0)

0.
where `
(1)

denote the maximal submodule of `

where 1

acts by c. Let
ad
(1)
denote the submodule Hom(`
(0)

. `
(1)

) of ad
0
. Then
H
1
ss
(G

. ad) = ker (H
1
(G

. ad) H
1
(1

. ad,ad
(1)
)).
The same is true with ad
0
replacing ad.
If : G

G1
2
(O) is semi-stable then we set
H
1
f
(G

. ad 1,O) = lim

H
1
f
(G

. ad (
n
,O)) H
1
(G

. ad 1,O).
We dene H
1
f
(G

. ad
0
1,O) as
H
1
f
(G

. ad 1,O) H
1
(G

. ad
0
1,O).
We make similar denitions for H
1
ss
(G

. ad1,O) and H
1
ss
(G

. ad
0
1,O).
We will need the following calculations.
Proposition 2.27 (a) Suppose that : G

G1
2
(/) is semi-stable. Then
#H
1
f
(G

. ad
0
) #H
0
(G

. ad
0
)#/
and equality holds if is ordinary
3
.
(b) Suppose that : G

G1
2
(O) is both good and ordinary. Let
1
and
2
be the unramied characters such that
_

1
c
0
2
_
. Let
c

= (
1
,
2
)(Frob

) 1.
If c

,= 0, then
H
1
ss
(G

. ad
0
1,O),H
1
f
(G

. ad
0
1,O)
is nite of order #(O,c

O).
3
The authors expect equality to hold without this hypothesis; cf. remark 2.26.
67
(c) If mod is either not good or not ordinary, then
H
1
ss
(G

. ad
0
1,O) = H
1
f
(G

. ad
0
1,O).
Proof: Part (c) is clear, and for parts (a) and (b) it suces to prove the
following two results.
Proposition 2.28 Suppose that : G

G1
2
(O,
n
) is good. Then
#H
1
f
(G

. ad) = #H
0
(G

. ad
0
)(#O,
n
)
2
.
Lemma 2.29 Suppose that : G

G1
2
(O,
n
) is ordinary, so that we have

_

1
c
0
2
_
, for some unramied characters
1
and
2
.
(a) We have
#H
1
ss
(G

. ad
0
) #H
0
(G

. ad
0
)#(O,
n
)#(O,(
n
. c

))
where c

= (
1
,
2
)(Frob

)1. Moreover if is good, then equality holds.

(b) If : = 1 and is not good then #H
1
ss
(G

. ad
0
) = #/.
End of proof of proposition 2.27: To deduce proposition 2.27 (in the good case)
from proposition 2.28 and lemma 2.29, note that if : G

G1
2
(O,
n
) is
good, then
H
1
(G

,1

. O,
n
) H
1
f
(G

. ad) H
1
(G

. O,
n
)
by lemma 2.22 (d), and this gives the inequality in proposition 2.27(a). Fur-
thermore if is also ordinary, then the above group is contained in
Im(H
1
f
(G

. ad) H
1
(G

. O,
n
)) H
1
(G

,1

. O,
n
)
where the last inclusion comes from lemma 2.25 (b). Therefore
#H
1
f
(G

. ad
0
) = #H
0
(G

. ad
0
)#(O,
n
).
2
We will prove proposition 2.28 in the next section using the theory of
Fontaine and Laaille. The proof of lemma 2.29 is a somewhat technical
exercise in the Galois cohomology of local elds, for which we refer the reader
to ch. 1 of [W3], prop. 1.9, parts (iii) and (iv). We remark that for our
purposes, inequalities would suce in part (b) of proposition 2.27 and part (a)
68
of lemma 2.29 (and this is all that is proved in [W3]). We have included the
precise formulas for the sake of completeness, since they are not much more
dicult to obtain. The only additional observation required is that if is
good, then the composite of the natural maps
H
1
(G

,1

. (ad
0
,ad
(1)
)
I

) H
1
(G

. ad
0
,ad
(1)
) H
2
(G

. ad
(1)
)
is trivial. To prove this, rewrite the composite as
H
1
(G

,1

. O,
n
) H
1
(G

. O,
n
) H
2
(G

. ad
(1)
)
with the second map given by c

where c

is the class in H
1
(G

. ad
(1)
)
dening the extension `

and apply lemma 2.25 (c). (In fact, one only needs
the easier half of lemma 2.25 (c): if is good then (c

) = 0.)
2.5 The theory of Fontaine and Laaille
In this section we again assume that / is an odd prime. We mentioned in
the last section the importance of understanding good representations of G

.
However the denition of good is somewhat indirect, and this makes computa-
tions dicult. The key result we use to address the problem is an equivalence
between the category TT
C
of good nite O[G

]-modules and a category /T

C
which we dene below following Fontaine and Laaille [FL]. The beauty and
utility of the result stems from the elementary algebraic nature of the def-
inition of /T
C
; we can convert questions about good representations into
questions in linear algebra.
Remark 2.30 Let (T
C
denote the category of nite at commutative group
schemes over Z

with an action of O. By results of Raynaud [Ray], taking

points denes an equivalence between the categories (T

C
and TT
C
. The
equivalence between (T
C
and a category closely related to /T
C
was rst
established by Fontaine [Fo2] and [Fo1]. While Fontaines results would suce
for our purposes here, our formulation will be closer to that in [FL], where
an equivalence between /T
C
and TT
C
is dened as part of a more general
construction of representations of G

from linear-algebraic data. We caution

however that our formulation is not exactly the same as that of [FL] since we
wish to work with covariant functors.
We now turn to the denition of /T
C
. The objects are O-modules 1
of nite cardinality together with a distinguished submodule 1
0
and O-linear
maps
1
: 1 1 and
0
: 1
0
1 which satisfy:
69

1
[
D
0 = /
0
,
Im
1
+ Im
0
= 1.
The morphisms are O-linear maps compatible with the additional data of the
distinguished submodules and maps .
It is useful to note that if 1 is an object of /T
C
, then there is a surjection

0
: 1,1
0
1
0
,/1
0
1,/
0
(1
0
).
and on counting orders we see that this is in fact an isomorphism. Thus there
is an isomorphism
1,(1
0
+1) 1
0
,1
0

1,1.
It follows that 1
0
,1
0
1,1 is injective, and hence that 1
0
is (non-
canonically) a direct summand of 1 as an O-module. Note also that
0
is
injective, and if 1 = 1
0
1
t
as O-modules, then also 1 =
0
(1
0
)
1
(1
t
)
as O-modules.
It is then straightforward to check that there is a contravariant functor

from /T
C
to itself dened by:
1

= Hom(1. Q

,Z

);
(1

)
0
= Hom(1,1
0
. Q

,Z

);

1
())(.) = )(/r +), where . =
1
(r) +
0
();

0
())(.) = )(r mod 1
0
), where .
1
(r) mod (
0
1
0
).
Moreover the canonical isomorphism 1

= (1

of O-modules denes a natu-

ral isomorphism in /T
C
.
We leave it as an exercise for the reader to use the above observations to
dene cokernels and then kernels of morphisms in /T
C
and verify that it is
an abelian category (or see [FL], sec. 1).
Theorem 2.31 There is a covariant functor D : TT
C
/T
C
which denes
an O-additive equivalence of categories. Moreover if ` is an object of TT
C
,
then we have
(a) ` and D(`) have the same cardinality;
(b) D(`) = D(`)
0
if and only if ` is unramied.
70
Remark 2.32 (a) It follows on applying part (a) to `,
i
` for each i that
` and D(`) are in fact (non-canonically) isomorphic as O-modules.
(b) As mentioned above, our formulation diers from that of Fontaine and
Laaille in that we are using covariant functors. To deduce theorem 2.31
from the results in [FL], sec. 9, we dene D as a quasi-inverse of the
functor l
S
(

[1]), where l
S
is dened in [FL] and [1] indicates a shift
by 1 in ltration degrees. In particular, if T is a nite at group scheme
over Z

with an action of O, then the underlying O-module of D(T(

))
can be identied with the covariant Dieudonne module of T
/F

, and
1
with 1.
Suppose now that : G

G1
2
(O,
n
) is a good continuous repre-
sentation. Let 1

= D(`

) be the corresponding object of /T

C
. Then
1

= (O,
n
)
2
as an O-module while 1
0

= O,
n
.
Lemma 2.33 The following are isomorphic.
(a) The group of extensions of `

by itself in the category of good nite

O,
n
[G

]-modules.
(b) The group of extensions of 1

by itself in the full subcategory of /T

C
consisting of objects which are O,
n
-modules.
(c) Pairs (
1
.
0
) where
1
Hom
C
(1

. 1

),
0
Hom
C
(1
0

. 1

) and
/
0
=
1
[
D
0

, modulo pairs of the form (c

,1

,1
c. c
,0

,0
c[
D
0)
where c Hom
C
(1

. 1

) and c1
0

1
0

.
Proof: The rst two groups of extensions are isomorphic by the Fontaine-
Laaille theorem. Following Ramakrishna [Ram] we explain how to calculate
the second group of extensions. We will write 1 for 1

. If
(0) 1 1 1 (0)
is an extension then we have as O-modules that 1

= 1
2
by an isomorphism
such that 1
0

(1
0
)
2
. Then 1 is determined by giving elements
E,1

Hom(1. 1) and
E,0
Hom(1
0
. 1) with /
E,0
=
E,1
[
D
0. Explicitly

E,1
=
_

1

E,1
0
1
_
and
E,0
=
_

0

E,0
0
0
_
.
Two such extensions 1 and 1
t
are isomorphic if there is an element c
End (1) such that c1
0
1
0
,
_
1 c
0 1
__

1

E,1
0
1
_
=
_

1

E

,1
0
1
__
1 c
0 1
_
.
71
and
_
1 c
0 1
__

0

E,0
0
0
_
=
_

0

E

,0
0
0
__
1 c[
D
0
0 1
_
;
or equivalently if c1
0
1
0
,
E,1

,1
=
1
c c
1
and
E,0

,0
=

0
c[
D
0 c
0
. The lemma follows. 2
Corollary 2.34 The group, H
1
f
(G

. ad), of extensions of `

by itself in the
category of good, nite O,
n
[G

]-modules is (non-canonically) isomorphic to

(O,
n
)
2
H
0
(1

. `

).
Proof: Choose generators c
0
. c
1
of 1, such that c
0
is a generator of 1
0
. With
respect to this basis let
0
have matrix
_
r
.
_
and
1
have matrix
_
/r
/. u
_
. If
c has matrix
_
c
1
c
2
0 c
3
_
then a direct calculation shows that
(c
1

1
c. c
0

0
c[
D
0) =
__
/c
2
. (c
1
c
3
) +c
2
(u /r)
(c
3
c
1
)/. /.c
2
_
.
_
c
2
.
(c
3
c
1
).
_
_
.
Thus the Ext group we want is the quotient of the set of pairs of matrices
(
1
.
0
) as above by the submodule generated by
__
/. u /r
0 /.
_
.
_
.
0
_
) and (
_
0
/. 0
_
.
_
0
.
__
.
Note that either . or u is a unit in O, from which it follows that the Ext
group we want is isomorphic to (O,
n
)
2
O,(..
n
). On the other hand
H
0
(1

. `

) corresponds to the largest submodule C 1

0
such that
0
C = C,
i.e. C = d 1
0
: .d = 0

= O,(..
n
). 2
Corollary 2.35 Suppose that : G

G1
2
(O,
n
) is a continuous good
representation. Then H
1
f
(G

. ad) is isomorphic to (O,

n
)
2
H
0
(G

. ad
0
).
Proof: If is not ordinary then H
0
(G

. ad
0
) and H
0
(1

. `

) are both triv-

ial, so suppose that is ordinary. In this case, let
t
denote the ordinary
representation dened by
`

= Hom(`
(0)

. `

) ad
0
()
and let `
(1)

= ad
0
()
(1)
. Then
H
0
(G

. ad
0
) = H
0
(G

. `

) H
0
(1

. `

)

= H
0
(1

. `

).
72
Since H
0
(1

. `
(1)

) is trivial, the restriction homomorphism

H
1
(G

. `
(1)

) H
1
(1

. `
(1)

)
is injective. We deduce from the long exact sequences associated to
(0) `
(1)

O,
n
0
that H
0
(G

. `

) = H
0
(1

. `

). 2
2.6 Deformations of representations
In this section we shall review Mazurs theory of deformations of representa-
tions of pronite groups (see [Maz3]).
Let (
C
denote the category whose objects are complete noetherian local
O-algebras with residue eld / and whose morphisms are O-algebra homomor-
phisms which are local (i.e. take maximal ideals into maximal ideals). (The
structure maps from O to every object of (
C
are also assumed to be local.) Let
G denote a topologically nitely generated pronite group and let denote an
absolutely irreducible representation of G into G1
d
(/). (In fact all we shall
use in the sequel is that / is the centraliser in `
d
(/) of the image of .) Let
T
0
denote the category of pronite O[G]-modules with continuous morphisms.
We will let T denote a full subcategory of T
0
which is closed under taking sub-
objects, quotients and direct products and which contains `

. Note that if
` is an object of T
0
and `
i
is a collection of subobjects which have trivial
intersection and such that each `,`
i
is an object of T, then ` is an object
of T, since `

i
`,`
i
.
Let : G O

be a continuous character such that det = mod . By

a lifting of of type 1 = (O. . T) we shall mean an object 1 of (
C
and a
continuous representation : G G1
d
(1) such that:
(a) mod m
R
= ,
(b) det = ,
(c) `

is an object of T.
Note that if : 1 1
t
and : G G1
d
(1) is of type 1 so is .
Theorem 2.36 There is a lifting
univ
D
: G G1
d
(1
D
) of of type 1 such
that if : G G1
d
(1) is any lifting of of type 1 then there is a unique
homomorphism of O-algebras : 1
D
1 such that is conjugate to
univ
D
.
73
The representation
univ
D
is referred to as the universal deformation of type
1. Mazur [Maz3] proved this theorem for T
0
and certain other categories T.
Ramakrishna [Ram] observed that the arguments work with any category T
satisfying the above hypotheses. We will sketch a proof which was suggested
by Faltings. (We remark that another explicit construction of the deformation
ring has been given by de Smit and Lenstra in [dSL].)
Proof of theorem 2.36: Choose a sequence p
1
. .... p
r
of topological generators
of G and liftings
1
. ....
r
of (p
1
). .... (p
r
) to `
d
(O). Dene a mapping
: `
d
(O) `
d
(O)
r
which sends r to (r
1

1
r. .... r
r

r
r). Since
has torsion-free cokernel, so we can decompose
`
d
(O)
r
= (`
d
(O)) \.
for some submodule \ `
d
(O)
r
. If : G G1
d
(1) is a lifting of of type
1 set

= ((p
1
)
1
. .... (p
r
)
r
) `
d
(1)
r
. Note that

0 mod m
R
and that

completely determines . We call the lifting well-placed if

belongs to \
C
1 `
d
(1)
r
. The crucial observation is the following result.
Lemma 2.37 If : G G1
d
(1) is a lifting of of type 1 then there is a
unique conjugate
t
of which is well-placed.
The lemma is rst proved for algebras 1 such that m
n
R
= (0) by induction on
:, and then one deduces the general case.
In virtue of the lemma it suces to nd a universal well-placed lifting of
type 1. Let c
1
. .... c
s
be a basis of \ as an O-module. If is a well-placed lifting
of of type 1 then we can write

s
i=1

,i
c
i
for unique elements
,i
m
R
and we can dene a homomorphism

: O[[1
1
. .... 1
s
]] 1 sending 1
i
to
,i
.
Note that is completely determined by

((p
i
) =
i
+

s
j=1

(1
j
)c
ji
, where
c
j
= (c
j1
. .... c
jr
)). Let 1 denote the intersection of all ideals J of O[[1
1
. .... 1
s
]]
such that there is a representation
J
: G G1
d
(O[[1
1
. .... 1
s
]],J) of type
1 with
J
(p
i
) =
i
+

s
j=1
1
j
c
ji
for all i. Let 1
D
denote the quotient of
O[[1
1
. .... 1
s
]] by 1. Then one can check that there is a representation
univ
:
G G1
d
(1
D
) with
univ
(p
i
) =
i
+

s
j=1
1
j
c
ji
for all i, and that this is the
desired universal representation. 2
We will need a few elementary properties of these universal deformations.
More precisely we will need to know how these universal rings change when
we change the base eld, we will need to know how to calculate the equi-
characteristic tangent space of these rings and more generally how to calculate
,
2
for certain prime ideals . The rst of these lemmas is a remark of
Faltings, the second is due to Mazur [Maz3] and the third to Wiles [W3].
74
Lemma 2.38 Let 1
t
,1 be a nite extension with ring of integers O
t
and
residue eld /
t
. Let T
t
denote the full sub-category of the category of pronite
O
t
[G]-modules such that the underlying object of T
0
is actually an object of T.
Let 1
t
= (O
t
. . T
t
). Then 1
D
= 1
D

C
O
t
and
univ
D
=
univ
D
1.
Proof: Let

1
D
denote the subring of 1
D
consisting of elements which reduce
modulo the maximal ideal to an element of / /
t
. Similarly let

1
D
denote the
subring of 1
D

C
O
t
consisting of elements which reduce modulo the maximal
ideal to an element of / /
t
. Then
univ
D
is in fact valued in G1
d
(

1
D
) and

univ
D
1 is in fact valued in G1
d
(

1
D
). The universal properties give natural
maps : 1
D


1
D
and : 1
D
1
D

C
O
t
. Moreover they show that the
composite ( 1) : 1
D
1
D
is the identity and that : 1
D


1
D
is the natural embedding. Thus is an isomorphism. 2
We will let T
(n)
denote the full subcategory of T whose objects are killed
by
n
. Suppose that ` is an object of T
0
which is nite and free over
O,
n
. Recall from section 2.4 that H
1
(G. End (`)) may be identied with
Ext
1
T
(n)
0
(`. `). If ` is an object of T
(n)
which is nite and free over O,
n
,
then we have a natural inclusion
Ext
1
T
(n)
(`. `) Ext
1
T
(n)
0
(`. `)

= H
1
(G. End (`)).
We dene H
1
T
(G. End (`)) to be the image of Ext
1
T
(n)
(`. `) in the group
H
1
(G. End (`)), and H
1
T
(G. End
0
(`)) to be the intersection
H
1
(G. End
0
(`)) H
1
T
(G. End (`)).
Lemma 2.39 There is a canonical isomorphism of /-vector spaces
Hom
k
(m
R
D
,(. m
2
R
D
). /)

= H
1
T
(G. ad
0
).
Proof: There is a natural bijection between Hom
k
(m
R
D
,(. m
2
R
D
). /) and the
set of O-algebra homomorphisms from 1
D
to the algebra /[] where
2
= 0 (the
correspondence associates : 1
D
/[] to [
m
R
D
). Hence there is a bijection
to the set of liftings : G G1
d
(/[]) of of type 1, modulo conjugation by
elements of 1 +`
d
(/). On the other hand, recall from section 2.4 that there
is a natural bijection between Ext
1
T
(1)
0
(`

. `

) and the set of of all continuous
liftings : G G1
d
(/[]) of modulo conjugation by elements of 1+`
d
(/).
Moreover a lifting is type 1 if and only if det = det and the corresponding
extension `

is an object of T. The lemma follows on checking linearity. 2

Now suppose that : 1
D
O is an O-algebra homomorphism. Let
denote the kernel of and let =
univ
D
. We set
H
1
T
(G. ad
0
1,O) = lim

H
1
T
(G. ad
0

n
,O) H
1
(G. ad
0
1,O).
75
Lemma 2.40 There is a canonical O-linear isomorphism
Hom
C
(,
2
. 1,O)

= H
1
T
(G. ad
0
1,O).
Proof: One shows in a very similar manner to the proof of lemma 2.39 that
for all : there is a natural isomorphism.
Hom
C
(,
2
. O,
n
)

= H
1
T
(G. ad
0
O,
n
).
2
2.7 Deformations of Galois representations
Again in this section we assume that / is an odd prime. Let : G
Q
G1
2
(/)
denote a continuous absolutely irreducible representation. Suppose moreover
that det = c and that is semi-stable in the sense that
[
G

is semi-stable,
and if j ,= / then # (1
p
)[/.
Note that if 1,Q is a semistable elliptic curve then
E,
: G
Q
G1
2
(F

), sat-
ises these conditions if it is irreducible. By a theorem of Mazur (theorem 2.9)
this will be the case if / 7.
Let denote a nite set of prime numbers. If 1 is an object of (
C
then
we say that a continuous lifting : G
Q
G1
2
(1) of is of type if the
following hold.
det = c.
[
G

is semi-stable.
If / , and [
G

is good then [
G

is good.
If j , / and is unramied at j then is unramied at j.
If j , / then [
I
p

_
1
0 1
_
.
Roughly speaking we require that at primes j , , is as unramied as could
be hoped for and we require that [
G

is semi-stable. Note that if

t
and
is a lifting of type then it is also a lifting of type
t
. Note also that if 1,Q
is an elliptic curve which is semi-stable at / and for which
E,
is irreducible
and semi-stable, then
E,
: G
Q
G1
2
(Z

) is a lifting of type if contains

all the primes for which 1 has bad reduction.
Now suppose that : G
Q
G1
2
(O,
n
) is a lifting of of type . We will
write H
1

(Q. ad
0
) for H
1
/

(Q. ad
0
), where
76
1
,p
= H
1
(G
p
,1
p
. (ad
0
)
I
p
) if j , /;
1
,p
= H
1
(G
p
. (ad
0
)) if j and j ,= /;
1
,
= H
1
f
(G

. (ad
0
)) if / , ;
1
,
= H
1
ss
(G

. (ad
0
)) if / .
Note that the pairing ad
0
ad
0
O,
n
given by (c. /) tr (c/) is perfect
and respects the action of G
Q
(i.e. tr ((ad
0
)(p)c. (ad
0
)(p)/) = tr (c/) for all
p G
Q
). Choosing a generator for the O-module Hom(O,
n
. Q

,Z

), we
obtain an isomorphism of O[G
Q
]-modules
ad
0
(1)

= Hom
C
(ad
0
. O,
n
)(1)

Hom(ad
0
. Q

,Z

)(1) = (ad
0
)

.
The O-submodule of H
1
(Q. ad
0
(1)) corresponding to H
1
/

(Q. (ad
0
)

) is in-
dependent of the choice of generator and we denote it H
1

(Q. ad
0
(1)). Thus
H
1

(Q. ad
0
(1)) is dened by the local conditions 1

,v
where the orthogo-
nality is with respect to the pairing
H
1
(G
v
. ad
0
) H
1
(G
v
. ad
0
(1)) Q

,Z

arising from the above isomorphism. We may equivalently regard the orthog-
onality as being with respect to the natural perfect O-bilinear pairing dened
by the composition
H
1
(G
v
. ad
0
) H
1
(G
v
. ad
0
(1)) H
2
(G
v
. ad
0

C
ad
0
(1))
H
2
(G
v
. O,
n
(1))

= O,
n
.
Note that if j ,= / we have that
1

,p
= H
1
(G
p
,1
p
. (ad
0
)(1)
I
p
) if j , ;
1

,p
= (0) if j .
If : G
Q
G1
2
(O) is a lifting of of type , we will write H
1

(Q. ad
0

1,O) for the direct limit lim

H
1

(Q. ad
0

n
,O), and H
1

(Q. ad
0
(1)1,O)
for the direct limit lim

H
1

(Q. ad
0
(1)
n
,O).
The main result we need about deformations of is the following.
Theorem 2.41 There is a universal lifting
univ

: G
Q
G1
2
(1

) of of
type , i.e.
univ

is a lifting of type and if : G

Q
G1
2
(1) is any lifting
of type then there is a unique O-algebra homomorphism : 1

1 such
that
univ

. Moreover we have the following.

77
(a) If 1
t
,1 is a nite extension and 1
t

is the corresponding deformation

ring then 1
t

= 1

C
O
t
and (
univ

)
t
=
univ

1.
(b) 1

can be topologically generated as an O-algebra by dim

k
H
1

(Q. ad
0
)
elements.
(c) If : 1

O is a O-algebra homomorphism, if =

and if
= ker then Hom(,
2
. 1,O)

= H
1

(Q. ad
0
1,O).
Proof: Let 1
0
denote the xed eld of . Let 1
n
(for : Z
>0
) denote the
maximal elementary abelian /-extension of 1
n1
unramied outside , /
and the primes where ramies. Let 1

n
1
n
and let G = Gal (1

,Q).
Note that any lifting of of type factors through G. Gal (1

,1
0
) is a pro-
/-group and its maximal elementary abelian quotient, Gal (1
1
,1
0
), is nite
by theorem 2.2. We deduce from the following lemma that Gal (1

,1
0
) and
hence G are topologically nitely generated. (See for instance [Koc] Satz 4.10
for a proof of this lemma.)
Lemma 2.42 Let H be a pro-/-group and

H its maximal elementary abelian
quotient. Suppose /
1
. .... /
r
H map to a set of generators of

H, then /
1
. .... /
r
topologically generate H.
Let T denote the category of pronite O[G]-modules ` for which
` is semi-stable as an O[G

]-module,
if / , and if is good then ` is good as an O[G

]-module,
if j , / and if is ramied at j then there is an exact sequence of
O[1
p
]-modules
(0) `
(1)
` `
(0)
(0).
such that 1
p
acts trivially on `
(1)
and `
(0)
.
Then we see that a lifting : G
Q
G1
2
(1) of is of type if and only if
factors through G,
det = c,
`

is an object of T.
78
The existence part of theorem 2.41 now follows from theorem 2.36 and part (a)
follows from lemma 2.38. Parts (b) and (c) follow from lemmas 2.39 and 2.40,
and the following observation:
If : G
Q
G1
2
(O,
n
) is a lifting of type and if j , | is a prime
where ramies then
ker (H
1
(G
p
. ad
0
) H
1
(1
p
. ad
0
,(ad
0
)
I
p
)) = H
1
(G
p
,1
p
. (ad
0
)
I
p
).
(2.7.1)
Equation (2.7.1) follows from the fact that the natural map H
1
(1
p
. ad
0
)
H
1
(1
p
. ad
0
,(ad
0
)
I
p
) is an injection (in fact an isomorphism). It is nothing
other than the map (ad
0
)
I
p
(ad
0
,(ad
0
)
I
p
)
I
p
. This completes the proof
of theorem 2.41. 2
Corollary 2.43 Suppose that if / = 3 then [
G
Q(

3)
is absolutely irreducible.
Then 1

can be topologically generated as an O-algebra by

dim
k
H
1

(Q. ad
0
(1)) +d

p
dim
k
H
0
(Q
p
. ad
0
(1))
elements, where d

= dim
k
H
1
ss
(Q

. ad
0
) dim
k
H
1
f
(Q

. ad
0
) if / , while
d

= 0 if / , .
Proof: Note that H
0
(Q. ad
0
(1)) = 0 unless / = 3 and [
G
Q(

3)
is not abso-
lutely irreducible. Thus according to theorem 2.18, dim
k
H
1

(Q. ad
0
) is the
sum of the following terms.
dim
k
H
1

(Q. ad
0
(1)).
dim
k
H
1
f
(Q

. ad
0
)dim
k
H
0
(Q

. ad
0
)dim
k
H
0
(R. ad
0
) 0 by propo-
sition 2.27.
d

.
For each j |, dim
k
H
1
(Q
p
. ad
0
) dim
k
H
0
(Q
p
. ad
0
) =
dim
k
H
0
(Q
p
. ad
0
(1)) by the local Euler characteristic formula (see the-
orem 2.17).
2
79
2.8 Special cases
In this section we will restrict attention to the deformation problems associated
with sets Q of certain special primes . These sets Q are chosen so that the
associated deformation ring 1
Q
enjoys a number of special properties which
will be crucial in the sequel. In particular we will only consider sets Q of
primes which satisfy the following two properties
1 mod /,
is unramied at and (Frob
q
) has distinct /-rational eigenvalues.
For such primes we will let
q
denote the maximal quotient of (Z,Z)

of /-power order. Then

q
is naturally a quotient of both G
q
and G
Q
via the
map

q
: G
Q
Gal (Q(
q
),Q)

= (Z,Z)

q
.
We will let
Q
=

qQ

q
,
Q
=

qQ

q
: G
Q

Q
and a
Q
denote the
augmentation ideal of O[
Q
]. Note that there is an isomorphism O[
Q
]

=
O[[o
q
: Q]],((1 + o
q
)
#
q
1 : Q) under which a
Q
corresponds to
(o
q
: Q).
For each Q, we choose an eigenvalue
q
of (Frob
q
) and denote the
other
q
.
Lemma 2.44 If Q then
univ
Q
[
G
q
is conjugate to
_
0
0 c
1
_
for some char-
acter such that

(Frob
q
) =
q
.
Proof: Choose a lifting ) G
q
of Frob
q
. As ()) has distinct /-rational
eigenvalues it is a simple application of Hensels lemma to see that we can
choose a basis such that
univ
Q
()) =
_

q
0
0

q
_
where
q
and

q
reduce to
q
and
q
in /. It suces to show that for any 1
q
,
univ
Q
() is diagonal in this
basis. Suppose
univ
Q
() = 1
2
+
_
c /
c d
_
with c. /. c. d m
R
Q
. Because
univ
Q
is
tamely ramied at we see that
univ
Q
())
univ
Q
()
univ
Q
())
1
=
univ
Q
()
q
. Thus
(
q
,

q
)/ and (

q
,
q
)c lie in m
R
Q
(/. c) so that (/. c) = m
R
Q
(/. c) and
(by Nakayamas lemma) / = c = 0. 2
We let
q,Q
denote the character : G
q
1

Q
in the conclusion of the
lemma. The restriction of
q,Q
to 1
q
factors through
q
. We let
Q
denote the
character G
Q
1

Q
which is unramied outside the primes of Q and whose
restriction to 1
q
is
q,Q
for each in Q. Thus
Q
factors through
Q
. We wish
to regard 1
Q
as an O[
Q
]-algebra, and it will be most convenient to do so
80
via the map which gives rise to
2
Q
. Note the following consequence of the
lemma 2.44.
Corollary 2.45 The natural map 1
Q
1

gives rise to an isomorphism

1
Q
,a
Q
1
Q

.
Lemma 2.46 (a) If Q then H
0
(F
q
. ad
0
) = H
0
(F
q
. ad
0
(1)) = / and
H
1
(F
q
. ad
0
) = H
1
(F
q
. ad
0
(1)) = /.
(b) 1
Q
can be topologically generated as an O-algebra by
#Q+ dim
k
H
1
Q
(Q. ad
0
(1))
elements.
(c) If
H
1

(Q. ad
0
(1))

qQ
H
1
(F
q
. ad
0
(1))
then #Q = dim
k
H
1

(Q. ad
0
(1)) and 1
Q
can be topologically generated
as an O-algebra by #Q elements.
Proof: The rst part is a direct calculation using the fact that Frob
q
acts
semi-simply on ad
0
with eigenvalues r. 1. r
1
for some r

/0. 1. The
same is true for ad
0
(1) as 1 mod /. The second and third parts follow
from this and corollary 2.43. 2
For the proof of the last theorem of the chapter, we shall need two results
on nite groups.
Theorem 2.47 (a) If H is a nite subgroup of 1G1
2
(C) then H is iso-
morphic to one of the following groups: the cyclic group C
n
of order :
(: Z
>0
), the dihedral group 1
2n
of order 2: (: Z
>1
),
4
, o
4
or
5
.
(b) If H is a nite subgroup of 1G1
2
(

) then one of the following holds:

H is conjugate to a subgroup of the upper triangular matrices;
H is conjugate to 1o1
2
(F

r ) or 1G1
2
(F

r ) for some : Z
>0
;
H is isomorphic to
4
, o
4
,
5
or the dihedral group 1
2r
of order
2: for some : Z
>1
not divisible by /.
In fact we shall only need part (b) which is due to Dickson [Dic2], secs. 255,
260 (see also [Hu] II.8.27), but we have included part (a) for later reference.
81
Lemma 2.48 Let F be a nite eld of odd characteristic /. If #F ,= 5, then
H
1
(o1
2
(F). End
0
(F
2
)) = 0.
Proof: This is a special case of results of [CPS], table 4.5 (assuming #F ,= 3).
In fact we shall only need it in the case / = 3, but the proof in the general
case is no more dicult, and we sketch it here for the readers convenience.
We let 1 (resp. l) denote the group of upper-triangular (resp. unipotent)
matrices in G = o1
2
(F). Since / does not divide the index of 1 in G, the
restriction homomorphism
H
1
(G. End
0
(F
2
)) H
1
(1. End
0
(F
2
))
is injective, so it suces to prove the latter group vanishes. Since / does not
divide the index of l in 1, we have
H
i
(1. `)

= H
i
(l. `)
B/U
all integers i 0 and F[1]-modules `. If #F = 3, then one checks directly
that for ` = End
0
(F
2
3
),
H
1
(l. `)

= ker `,( 1)` = 0
where generates l and ` = 1 + +
2
on `. If #F 5, then one proceeds
by writing
(0) = `
0
`
1
`
2
`
3
= End
0
(F
2
)
as F[1]-modules where the successive quotients `
i
,`
i1
, (for i = 1, 2, 3) are
one-dimensional over F. The calculation is then straightforward using long
exact sequences of cohomology, except in the case #F = 9 where one must
also check that the one-dimensional space of classes in H
1
(l. `
3
,`
2
) xed by
1,l maps injectively to H
2
(l. `
2
,`
1
) via the connecting homomorphism.
2
Theorem 2.49 Keep the assumptions of the last section and suppose more-
over that if 1 = Q(
_
(1)
(1)/2
/) then [
G
L
is absolutely irreducible. Then
there exists a non-negative integer : such that for any : Z
>0
we can nd a
nite set of primes Q
n
with the following properties.
(a) If Q
n
then 1 mod /
n
.
(b) If Q
n
then is unramied at and (Frob
q
) has distinct eigenvalues.
(c) #Q
n
= :.
82
(d) 1
Q
n
can be topologically generated by : elements as a O-algebra.
Proof: Take : = dim
k
H
1

(Q. ad
0
(1)). It suces to nd a set Q
n
with the
following properties.
(a) If Q
n
then 1 mod /
n
.
(b) If Q
n
then is unramied at and (Frob
q
) has distinct eigenvalues.
(c) H
1

(Q. ad
0
(1))

qQ
n
H
1
(F
q
. ad
0
(1)).
As each H
1
(F
q
. ad
0
(1)) is one dimensional we may replace the last condition
with
H
1

(Q. ad
0
(1))

qQ
n
H
1
(F
q
. ad
0
(1)).
Thus what we need show is that for each non-zero class [] H
1

(Q. ad
0
(1))
there is a prime (depending on []) such that
(a) 1 mod /
n
,
(b) is unramied at and (Frob
q
) has distinct eigenvalues,
(c) res
q
[] H
1
(F
q
. ad
0
(1)) is nontrivial.
Using the Chebotarev density theorem we see that it will do to nd G
Q
such that
(a) [
Q(

n)
= 1,
(b) ad
0
() has an eigenvalue other than 1,
(c) () , ( 1)ad
0
(1).
For : 0, let 1
m
denote the extension of Q(

m) cut out by ad
0
; i.e.,
the eld xed by the kernel of the representation ad
0
restricted to G
Q(

m)
.
We will show that (G
F
n
) is non-trivial. For this it suces to prove that
H
1
(Gal (1
n
,Q). ad
0
(1)) = (0). Consider the ination-restriction exact se-
quence
(0) H
1
(Gal (1
0
,Q). (ad
0
(1))
G
F
0
) H
1
(Gal (1
n
,Q). ad
0
(1))
H
1
(Gal (1
n
,1
0
). ad
0
(1))
G
Q
.
83
Since 1
1
,1
0
is an extension of degree prime to /, and since G
Q
acts trivially
on Gal (1
n
,1
1
), we have:
H
1
(Gal (1
n
,1
0
). ad
0
(1))
G
Q
Hom(Gal (1
n
,1
1
). ad
0
(1)
G
Q
).
(2.8.1)
Since [
G
L
is absolutely irreducible, the cohomology group in equation (2.8.1)
vanishes. On the other hand, G
F
0
acts trivially on ad
0
so the rst term van-
ishes as well unless Gal (1
0
,Q) has order divisible by / and has Gal (Q(

),Q)
as a quotient. Recall that Gal (1
0
,Q) is isomorphic to the projective image of
, so by theorem 2.47 we are reduced to the case / = 3 and the map
Gal (1
0
,Q(
3
)) 1o1
2
(

/)
has image conjugate to 1o1
2
(F
3
r ) for some : 1. It suces to prove that in
this case
H
1
(Gal (1
0
,Q(
3
)). ad
0

k

/)) = (0).
but this follows directly from lemma 2.48.
Now note that since [
G
L
is absolutely irreducible, so is [
G
Q(

n)
. Regarding
(G
F
n
) as a Gal (1
n
,Q(

n))-submodule of ad
0
( ), we nd that some p
Gal (1
n
,Q(

n)) of order not dividing / xes a non-zero element of (G

F
n
).
Let
0
be a lifting of p to G
Q(

n)
. We will look for =
0
with G
F
n
. We
only need choose so that (
0
) = () + (
0
) , (
0
1)ad
0
(1). This is
possible because (G
F
) , (p 1)ad
0
(1). 2
3 Modular forms and Galois representations
3.1 From modular forms to Galois representations
We suppose in this section that ) =

c
n
())
n
is a newform of weight two and
level `
f
(see denition 1.21). Let 1
f
denote the number eld in C generated
by the Fourier coecients c
n
()). Let
f
denote the character of ), i.e., the
homomorphism (Z,`
f
Z)

f
dened by mapping d to the eigenvalue of
d) on ).
Recall that a construction of Shimura (section 1.7) associates to ) an
abelian variety
f
of dimension [1
f
: Q]. This abelian variety is a certain
quotient of J
1
(`
f
), and the action of the Hecke algebra on J
1
(`
f
) provides an
embedding
1
f
End
Q
(
f
) Q.
84
We saw also that for each prime / the Tate module T

(
f
)
Z

becomes a
free module of rank two over 1
f
Q

(lemma 1.48). The action of the Galois

group G
Q
on the Tate module commutes with that of 1
f
, so that a choice of
basis for the Tate module provides a representation
G
Q
G1
2
(1
f
Q

). (3.1.1)
As 1
f
Q

can be identied with the product of the completions of 1

f
at its
primes over /, we obtain from ) certain 2-dimensional /-adic representations
of G
Q
.
/-adic representations: In this discussion, we x a prime / and a nite
extension 1 of Q

. We let O denote the ring of integers of 1, the maximal

ideal and / the residue eld. We shall consider /-adic representations with
coecients in nite extensions of our xed eld 1. We regard 1 as a subeld
of

Q

and x embeddings

Q

Q

and

Q

C. If 1
t
is a nite extension of
1 with ring of integers O
t
, then we say that an /-adic representation G

G1
2
(1
t
) is good (respectively, ordinary, semistable) if it is conjugate over
1
t
to a representation G

G1
2
(O
t
) which is good (respectively, ordinary,
semistable) in the sense of section 2.4.
Let 1
t
f
denote the 1-algebra in

Q

generated by the Fourier coecients of

). Thus 1
t
f
is a nite extension of 1, and it contains the completion of 1
f
at
the prime over / determined by our choice of embeddings. We let O
t
f
denote
the ring of integers of 1
t
f
and write /
f
for its residue eld. We dene

f
: G
Q
G1
2
(1
t
f
)
as the pushforward of (3.1.1) by the natural map 1
f
Q

1
t
f
. We assume
the basis is chosen so that
f
factors through G1
2
(O
t
f
). We also let
t
f
denote
the nite order /-adic character
G
Q
Gal (Q(
N
f
),Q) (1
t
f
)

obtained from
f
.
The following theorem lists several fundamental properties of the /-adic
representations
f
obtained from Shimuras construction. The result is a com-
bination of the work of many mathematicians. We discuss some of the proofs
and provide references below. In the statement we x ) as above and write
simply `, c
n
, , ,
t
and 1
t
for `
f
, c
n
()),
f
,
f
,
t
f
and 1
t
f
respectively.
Theorem 3.1 The /-adic representation
: G
Q
G1
2
(1
t
)
has the following properties.
85
(a) If j,[`/ then is unramied at j and (Frob
p
) has characteristic poly-
nomial
A
2
c
p
A +j(j).
(b) det() is the product of
t
with the /-adic cyclotomic character c, and
(c) is conjugate to
_
1 0
0 1
_
.
(c) is absolutely irreducible.
(d) The conductor `() is the prime-to-/-part of `.
(e) Suppose that j ,= / and j[[`. Let denote the unramied character
G
p
(1
t
)

satisfying (Frob
p
) = c
p
. If j does not divide the conductor
of , then [
G
p
is of the form
_
c
0
_
.
If j divides the conductor of , then [
G
p
is of the form

1
c
t
[
G
p
.
(f ) If /,[2`, then [
G

is good. Moreover [
G

is ordinary if and only if c

is
a unit in the ring of integers of 1
t
, in which case
I

(Frob

) is the unit
root of the polynomial A
2
c

A +/(/).
(g) If / is odd and /[[`, but the conductor of is not divisible by /, then
[
G

is ordinary and
I

(Frob

) = c

.
Proof: Part (a) was established by Shimura ([Shi2], [Shi3]). The key ingredi-
ent is the Eichler-Shimura congruence relation, theorem 1.29. Recall that
J
1
(`) has good reduction at primes j not dividing `. So the action of
G
p
on T

(
f
)
Z

is unramied and is in fact described by the action of

Frob
p
G
F
p
on the Tate module of the reduction. But this is given by the
Frobenius endomorphism 1 whose characteristic polynomial is computed in
corollary 1.41.
The rst assertion of (b) follows from (a) on applying the Chebotarev
density theorem. The second assertion then follows on noting that (1) = 1.
Part (c) was proved by Ribet (see section 2 of [R3]). Assuming reducibility
of the representation, he applies algebraicity results of Lang and Serre to
obtain a contradiction to the estimate on the Fourier coecients stated in
theorem 1.24.
86
Parts (d) and (e) follow from a deep result of Carayol [Ca1], Thm. (A),
building on the work of Langlands [Ll1], Deligne and others. In fact, this
result and the local Langlands correspondence characterize [
G
p
in terms of
[
G
p
and the 1- and c-factors at j of twists of ). The descriptions in the case
of j[[` are based on the analysis of Deligne-Rapoport of the reduction mod j
of J
1
(`) (see [DR], [Ll1]).
The rst assertion of (f) follows from the fact that
f
has good reduction
at / if / does not divide `. The second assertion of (f) (respectively, all of (g))
follows from the Eichler-Shimura congruence relation (respectively, the work of
Deligne-Rapoport), and general results on /-divisible groups and the reduction
of abelian varieties; see thm. 2.2 of [W2], lemma 2.1.5 of [W1], 12 of [Gro]
and thms. 2.5 and 2.6 of [Edi]. The restriction to odd / in (f) and (g) is made
primarily out of lack of suitable denitions. 2
Mod / representations: We maintain the notation used in the discussion of
/-adic representations. Dene

f
: G
Q
G1
2
(/
f
)
to be the semi-simplication of the reduction of
f
. (See the discussion fol-
lowing proposition 2.6.) Assertions analogous to those in theorem 3.1 hold for
=
f
, except that
The representation need not be absolutely irreducible (as in (c)). How-
ever if / is odd, one checks using (b) that is irreducible if and only if
it is absolutely irreducible.
In (d) one only has divisibility of the prime-to-/ part of `
f
by `( ) in
general.
The various possibilities for :
p
( ) to be strictly less than the exponent of
j in ` (where j ,= /) were classied independently by Carayol [Ca2] and Livne
[Liv]. We record the following consequence of their results (cf. the introduction
of [DT1]):
Proposition 3.2 Suppose that j is a prime such that j[`
f
, j , 1 mod / and

f
is unramied at j. Then tr (
f
(Frob
p
))
2
= (j + 1)
2
in /
f
.
Artin representations: The theory of Hecke operators and newforms (see
section 1.3) extends to modular forms on
1
(`) of arbitrary weight. The
construction of /-adic representations associated to newforms was generalized
to weight greater than 1 by Deligne [De] using etale cohomology. There are
also Galois representations associated to newforms of weight 1 by Deligne and
Serre [DS], but an essential dierence is that these are Artin representations.
87
Theorem 3.3 If p =

c
n
(p)
n
is a newform of weight one, level `
g
and
character
g
, then there is an irreducible Artin representation

g
: G
Q
G1
2
(C)
of conductor `
g
with the following property: If j,[`
g
, then the characteristic
polynomial of
g
(Frob
p
) is
A
2
c
p
(p)A +
g
(j).
Remark 3.4 Note that det(
g
) is the character of G
Q
corresponding to and

g
(c) is conjugate to
_
1 0
0 1
_
(see theorem 3.1).
Remark 3.5 A basis can be chosen so that the representation
g
takes values
in G1
2
(1
g
) (where 1
g
is the number eld generated by the c
n
(p)). Moreover
suppose that 1 is a nite extension of Q

in

Q

and we have xed embeddings

of

Q in C and

Q

). If 1
g
is contained in 1, then we can view
g
as giving rise
to an /-adic representation G
Q
G1
2
(1) and hence a mod / representation
G
Q
G1
2
(/).
Remark 3.6 A key idea in the construction of
g
is to rst construct the mod
/ representations using those already associated to newforms of higher weight.
More precisely, suppose that 1
g
1 as in remark 3.5. One can show that
for some newform ) of weight 2 and level `
f
dividing `/ we have
c
p
(p) c
p
()).
g
(j)) j
f
(j)
for all j,[`/, the congruence being modulo the maximal ideal of the ring
of integers of 1
t
f
. Thus
f
is the semisimplication of the desired mod /
representation (with scalars extended to /
f
).
3.2 From Galois representations to modular forms
It is conjectured that certain types of two-dimensional representations of G
Q
always arise from the constructions described in section 3.1. We now state
some of the conjectures and the results known prior to [W3] and [TW].
Artin representations:
Conjecture 3.7 Let : G
Q
G1
2
(C) be a continuous irreducible represen-
tation with det((c)) = 1. Then is equivalent to
g
for some newform p of
weight one.
88
Recall that
g
is the Artin representation associated to p by the Deligne-Serre
construction (theorem 3.3).
Remark 3.8 Conjecture 3.7 is equivalent to the statement that the Artin
1-functions attached to and to all its twists by one-dimensional characters
are entire. (The Artin conjecture predicts that the Artin 1-function 1(:. ) is
entire, for an arbitrary irreducible, non-trivial Artin representation : G
Q

G1
d
(C).)
A large part of conjecture 3.7 was proved by Langlands in [Ll2], and the
results were extended by Tunnell [Tu].
Theorem 3.9 Let : G
Q
G1
2
(C) be a continuous irreducible representa-
tion such that (G
Q
) is solvable and det((c)) = 1. Then is equivalent to

g
for some newform p of weight one.
Remark 3.10 Note that by theorem 2.47, part (a), the solvability hypothesis
excludes only the case where the projective image of is isomorphic to
5
.
Remark 3.11 If the projective image of is dihedral, then is induced from
a character of a quadratic extension of Q. In this case the result can already
be deduced from work of Hecke.
Mod / representations: We x notation as in the discussion of /-adic and
mod / representations in section 3.1.
Denition 3.12 We say that a representation : G
Q
G1
2
(/) is modular
(of level `) if for some newform ) of weight 2 (and level `), is equivalent
over /
f
to
f
.
By proposition 1.32 the notion is independent of the choices in section 3.1
of embeddings 1

Q

,

Q

Q

and

Q C. Moreover if 1
t
is a nite
extension of 1 with residue eld /
t
, then is modular if and only if
k
/
t
is
modular.
The following was conjectured by Serre [Se7], (3.2.3). (See also [Da1] for
further discussion and references.)
Conjecture 3.13 Let : G
Q
G1
2
(/) be a continuous absolutely irreducible
representation with det( (c)) = 1. Then is modular.
Some cases of Serres conjecture can be deduced from theorem 3.9.
89
Theorem 3.14 Let : G
Q
G1
2
(/) be a continuous absolutely irreducible
representation with det( (c)) = 1. Suppose that one of the following holds:
(a) / = F
3
;
(b) the projective image of is dihedral.
Then is modular.
Sketch of proof: For case (a), we consider the surjection
G1
2
(Z[

2]) G1
2
(F
3
)
dened by reduction mod (1 +

2). One checks that there is a section

: : G1
2
(F
3
) G1
2
(Z[

2]) and applies theorem 3.9 to : . The resulting

representation arises from a weight one newform, and hence its reduction is
equivalent to
f
for some ) (see remark 3.6).
In case (b), is equivalent to a representation of the form Ind
G
Q
G
F

where
1 is a quadratic extension of Q and

is a character G
F
/

. (We have
here enlarged 1 if necessary.) Let : be the order of

, choose an embedding
Q(c
2i/n
) 1 and lift

to a character : G
F
Z[c
2i/n
]

. We may
always choose so that the Artin representation = Ind
G
Q
G
F
is odd, i.e.,
det((c)) = 1. (In the case / = 2 and 1 real quadratic, we may have to
multiply by a suitable quadratic character of G
F
.) We then apply theorem 3.9
to and deduce as in case (a) that is modular. 2
Serre also proposed a renement ([Se7], (3.2.4)) of the conjecture which
predicts that is associated to a newform of specied weight, level and char-
acter. Through work of Mazur, Ribet [R5], Carayol [Ca2], Gross [Gro] and
others, this renement is now known to be equivalent to conjecture 3.13 if / is
odd. (One also needs to impose a mild restriction in the case / = 3.) See [R6]
and [Di1] for statements of the results and further references; here we give a
variant which applies to newforms of weight two. Before doing so, we assume
/ is odd and dene an integer ( ) as follows:
( ) = 0 if [
G

is good;
( ) = 1 if [
G

is not good and [

I

k

/ is of the form
_
c
a

0 1
_
.
_
c
0 c
a
_
or
_

a
0
0
a
_
for some positive integer c < /. (Recall that c is the cyclotomic character
and is the character of 1

dened after lemma 2.22.)

90
( ) = 2 otherwise.
Theorem 3.15 Suppose that / is odd and is absolutely irreducible and mod-
ular. If / = 3, then suppose also that [
G
Q(

3)
is absolutely irreducible. Then
there exists a newform ) of weight two such that
is equivalent over /
f
to
f
;
`
f
= `( )/
( )
;
the order of
f
is not divisible by /.
Proof: The existence of such an ) follows from [Di1] thm. 1.1, thm. 5.1 and
lemma 2.1, but with `
f
dividing `( )/
( )
. By lemma 2.7 above, we see that
`
f
is divisible by `( ). The divisibility of `
f
by ( ) follows from results in
sec. 8 of [Gro] and sec. 2.4 of [Edi]. 2
/-adic representations: We again use the notation of section 3.1. Let :
G
Q
G1
2
(1) be an /-adic representation.
Denition 3.16 We say that is modular if for some weight 2 newform ),
is equivalent over 1
t
f
to
f
.
The notion is independent of the choices of embeddings and well-behaved under
extension of scalars by proposition 1.32 (cf. denition 3.12).
The following is a special case of a conjecture of Fontaine and Mazur [FM].
Conjecture 3.17 If : G
Q
G1
2
(1) is an absolutely irreducible /-adic
representation and [
G

is semistable (in the sense of section 2.4), then is

modular.
(Recall that for us /-adic representations are dened to be unramied at all but
nitely many primes. Recall also that if [
G

is semistable, then by denition

det [
I

is the cyclotomic character c.)

Remark 3.18 Relatively little was known about this conjecture before Wiles
work [W3]. Wiles proves that under suitable hypotheses, the modularity of
implies that of .
Remark 3.19 The conjecture stated in [FM] is stronger than the one here;
in particular, the semistability hypothesis could be replaced with a suitable
notion of potential semistability. On the other hand, one expects that if [
G

is
semistable, then it is equivalent to
f
(over 1
t
f
) for some ) on
1
(`())
0
(/)
(and on
1
(`()) if [
G

is good).
91
The Shimura-Taniyama conjecture: Conjecture 1.54 can be viewed in the
framework of the problem of associating modular forms to Galois representa-
tions. Let 1 be an elliptic curve dened over Q. For each prime /, we let
E,
denote the /-adic representation G
Q
G1
2
(Q

) dened by the action of G

Q
on the Tate module of 1 (see section 2.2).
Proposition 3.20 The following are equivalent:
(a) 1 is modular.
(b)
E,
is modular for all primes /.
(c)
E,
is modular for some prime /.
Proof: If 1 is modular, then 1 is isogenous to
f
for some weight two newform
) with 1
f
= Q (see section 1.8). It follows that for each prime /,
E,
is
equivalent to the /-adic representation
f
. Hence (a) (b) (c).
To show (c) (b), suppose that for some / and some ), the representations

E,
and
f
are equivalent. First observe that for all but nitely primes j, we
have
tr (
f
(Frob
p
)) = tr (
E,
(Frob
p
)).
We deduce from proposition 2.11 and theorem 3.1, part (a) that for all but
nitely many primes j
c
p
()) = j + 1 #

1
p
(F
p
) Z. (3.2.1)
Applying proposition 2.6, we nd that for each prime /,
E,
is equivalent to

f
and is therefore modular.
We nally show that (b) (a). The equality (3.2.1) holds for all primes
j not dividing `
f
, which by theorem 3.1, part (d), is the conductor of 1.
Since det(
f
) = det(
E,
) = c, we see by 3.1, Part (b) that
f
is trivial. By
theorem 1.27 parts (b) and (d) (or [AL] thm. 3), c
p
is in 0. 1 for primes
j dividing `
f
. Thus 1
f
= Q and
f
is an elliptic curve. Faltings isogeny
theorem (see [CS], sec. II.5) now tells us that 1 and
f
are isogenous and we
conclude that 1 is modular. 2
Remark 3.21 Note that the equivalence (b) (c) does not require Faltings
isogeny theorem.
Proposition 3.22 If the Fontaine-Mazur conjecture (conjecture 3.17) holds
for some prime /, then the Shimura-Taniyama conjecture (conjecture 1.54)
holds. If Serres conjecture (conjecture 3.13) holds for innitely many /, then
conjecture 1.54 holds.
92
Proof: The rst assertion is immediate from proposition 3.20 and the irre-
ducibility of
E,
. See [Se7], sec. 4.6 for a proof of the second. (We have
implicitly chosen the eld 1 to be Q

in the statements of conjectures 3.17

and 3.13, but it may be replaced by a nite extension.) 2
Remark 3.23 Note that to prove a given elliptic curve 1 is modular, it suf-
ces to prove that conjecture 3.17 holds for a single / at which 1 has semistable
reduction. Wiles approach is to show that certain cases of conjecture 3.13 im-
ply cases of conjecture 3.17 and hence cases of conjecture 1.54.
3.3 Hecke algebras
In this section x the following notation. Let / be an odd prime, let 1 be a
nite extension of Q

, let O denote the ring of integers of 1, let denote its

maximal ideal and / its residue eld. Fix embeddings 1

Q

,

Q

Q

and

Q C. Let : G
Q
G1
2
(/) denote a continuous representation with the
following properties
(a) is irreducible,
(b) is modular,
(c) det = c,
(d) [
G

is semi-stable,
(e) and if j ,= / then # (1
p
)[/.
Let us rst record the following lemma.
Lemma 3.24 The representation [
G
L
is absolutely irreducible where 1 =
Q(
_
(1)
(1)/2
/).
Proof: If it were not then we see that /,[# (G
Q
) and so is unramied at
all j ,= /. Moreover we can check that [
I

_
c 0
0 1
_
. If / 3 we can use
theorem 3.15 to deduce that is modular of weight 2 and level 1 and hence
obtain a contradiction. If / = 3 we see that the splitting eld of is everywhere
unramied over Q(

3) and hence must equal Q(

3), a contradiction. 2
Let denote a nite set of primes. For the application to modularity of
semistable elliptic curves, it suces to consider sets contained in

where

is dened as follows
93
Denition 3.25 For a representation as above, we let

denote the set of
primes j satisfying
j = / and [
G

is good and ordinary; or

j ,= / and is unramied at j.
We shall sometimes assume that

in order to simplify statements and
proofs. Let ^

denote the set of newforms ) such that

f
: G
Q
G1
2
(O
t
f
)
is equivalent to a lifting of
k
/
f
of type and /
2
,[`
f
(the last condition is
presumably not necessary, cf. remark 3.19). From theorem 3.1 and lemma 2.7
one deduces the following description of ^

.
Lemma 3.26 The set ^

consists of newforms ) such that

f

=
k
/
f
,

f
is trivial,
`
f
divides /

`( )

p
j
dim
I
p
, where = 0 if is good and / , and
= 1 otherwise.
As is modular it follows from theorem 3.15 that for all , ^

,= . Set

fA

O
t
f
. If j is a prime not in and not dividing /`( ), we let 1
p
denote the element (c
p
()))
f
in

T

. Then dene T

to be the O-subalgebra of

generated by the elements 1

p
for such primes j. Then T

is a complete
noetherian local O-algebra with residue eld /. Moreover it is reduced and it
is a nitely generated free O-module.
Lemma 3.27 There is a continuous representation

mod

: G
Q
G1
2
(T

)
such that if j,[/`( ) and j , then
mod

is unramied at j and we have

tr
mod

(Frob
p
) = 1
p
. Moreover we have the following.
(a)
mod

is a lift of of type and there is a unique surjection

: 1

such that
mod

univ

.
(b) If
t
then there is a unique surjection T

such that
mod

pushes forward to
mod

and 1
p
maps to 1
p
for j,[/`( ) and j ,
t
.
94
(c) If 1
t
is a nite extension of 1 and T
t

is constructed in the same way

as T

but with 1
t
replacing 1 then T
t

= T

C
O
K
.
Proof: Consider
mod

f
: G
Q
G1
2
(

). Choose some complex con-

jugation, c, and conjugate so that (c) =
_
1 0
0 1
_
. By conjugating
mod

we may assume that modulo every maximal ideal it reduces to and that

mod

(c) =
_
1 0
0 1
_
. Call the corresponding basis c
+
. c

. Note that by con-

tinuity and the Chebotarev density theorem tr
mod

is valued in T

. Thus for
any p G
Q
the diagonal entries of
mod

(p) lie in T

, because tr
mod

(p) and
tr
mod

(cp) both do.

By irreducibility of we can nd some G
Q
such that () =
_
/

_
where / ,= 0. Rescaling c
+
we may assume that
mod

() =
_
c 1
c d
_
. Then for all
p G
Q
the lower left entry of
mod

(p) lies in T

(look at the upper left entry

of
mod

(p)). Again using the irreducibility of we can nd G

Q
such that

mod

() =
_

c
_
where c T

. Looking at the lower right entry of

mod

(p)
we see that for any p G
Q
the upper right entry of
mod

(p) lies in T

. Thus

mod

is now in fact valued in G1

2
(T

) and will be our candidate for

mod

. We
leave the verication of the other properties of
mod

as an exercise. 2
Example 3.28 Let =
f
57B
,3
where )
57B
is the newform of level 57 discussed
in the example of section 1.6. As )
57B
is not congruent modulo 3 to any form of
level 19 or 3 we see by theorem 3.15 that is ramied at 19 and [
G
3
is not good.
On the other hand is semi-stable. The facts that (Frob
2
) has order 8 (see the
table below) and 3[# (1
19
) (from the discussion above) imply that : G
Q

G1
2
(F
3
). The table below also shows that T

= (r. ) Z
2
3
: r mod 3.
There is a unique continuous representation
mod

: G
Q
G1
2
(T

) such that
if j,[57 then
mod

is unramied at j and tr
mod

(Frob
p
) = 1
p
. We have an
isomorphism T

,3T

F
3
[] (where
2
= 0) given by (r. ) r+
yx
3
. Thus
we get a representation : G
Q
G1
2
(F
3
[]). The following table lists the
traces of Frobenius elements for the rst few unramied primes under these
representations:
j 2 5 7 11 13 17 23 29
1 1 0 0 0 0 1 1

mod

(1. 2) (2. 1) (0. 3) (0. 3) (6. 6) (6. 3) (4. 4) (2. 10)

1 1 + 0 1 1
Exercise 3.29 Show that there are three algebra homomorphisms T

Z,9Z and hence show that there are at least three liftings of of type
to Z,9Z.
95
Exercise 3.30 What is the image of ?
We will need two deeper properties of the Hecke algebras T

(theorem 3.31
and theorem 3.36 below). These results will be proved in the next chapter
(sections 4.3 and 4.4).
Theorem 3.31 Let Q be a nite set of primes as in section 2.8. Then T
Q
is a free O[
Q
]-module, where T
Q
inherits the structure of an O[
Q
]-module
from 1
Q
via the homomorphism
Q
: 1
Q
T
Q
.
Corollary 3.32 T

= T
Q
,a
Q
.
Proof: From the denitions and corollary 2.45 we have that
T

C
1 = (T
Q

C
1),a
Q
.
and from the theorem we have that T
Q
,a
Q
is torsion free. The corollary
follows. 2
For the second of these results we will need some additional notation and
we restrict our attention to sets of primes contained in

. Suppose that


and that ) is an element of ^

whose Fourier coecients are in O.

Then O
t
f
= O and projection to the component corresponding to ) gives rise
to an O-algebra homomorphism
=
f
: T

O.
The pushforward of
mod

by is equivalent to

f
: G
Q
G1
2
(O).
Remark 3.33 Most of the objects considered in the rest of the section will
depend on the choice of newform ). We also remark that to give an O-algebra
homomorphism T

O is equivalent to giving a newform ) in ^

with
coecients in O. Indeed given such a homomorphism, there exists a newform
) in ^

such that the homomorphism is dened by 1

p
c
p
()) for all j ,
with j not dividing /`( ). The fact that c
p
()) O for such j implies (using
for example parts (b) and (d) of theorem 1.27 and lemma 4.1 below) that all
the Fourier coecients of ) are in O. The uniqueness of ) is a consequence of
the theory of newforms, theorem 1.22.
96
For
t
satisfying

t


.
let

denote the composite T

O. Let

(Ann
T

(ker

)). (3.3.1)
Note that because T

is reduced,

,= (0). Also let

denote the kernel of

. Recall that
#

,
2

= #H
1

(Q. ad
0

f

C
1,O).
Remark 3.34 We have not yet shown that these groups are nite, but if
either is nite, then so is the other and their cardinality is the same.
Note that if / is in

then 1

:= (c

(p))
g


T

is actually in T

.
This follows from theorem 3.1(f), which shows that 1

+ /
1

where

is the eigenvalue of Frob

on the free rank one T

-module `
(0)
where ` is
the module underlying
mod

(see lemma 2.20). Now that we have shown that

1

, we may characterize

as the unit root in T

of
A
2
1

A +/ = 0. (3.3.2)
For primes j in

we dene an element c
p
T

by
c
p
= (j 1)(1
2
p
(j + 1)
2
).
Note that (c
p
) = (j 1)(c
p
())
2
(j + 1)
2
), which is non-zero by theo-
rem 1.27(a). A calculation using theorem 2.17 and theorem 3.1(a) shows that
if j is in

( /) then
#H
1
(G
p
. ad
0

f

C
1,O),H
1
(G
p
,1
p
. ad
0

f

C
1,O)
= #H
0
(G
p
. ad
0

f

C
1,O(1)) = #O,(c
p
).
If / is in

then we see from proposition 2.27 and theorem 3.1(f) that
#H
1
ss
(G

. ad
0

f

C
1,O),H
1
f
(G

. ad
0

f

C
1,O) = #O,(c

).
For the next proposition dene groups H
p
for j

by
H
p
= H
1
(G
p
. ad
0

f

C
1,O),H
1
(G
p
,1
p
. ad
0

f

C
1,O) if j ,= /
and H

= H
1
ss
(G

. ad
0

f

C
1,O),H
1
f
(G

. ad
0

f

C
1,O).
Then we have the following result.
97
Proposition 3.35 If
t


, then
#H
1

(Q. ad
0

f

C
1,O),H
1

(Q. ad
0

f

C
1,O) #(O,(

c
p
)).
Moreover if we have equality then the sequence
(0) H
1

(Q. ad
0

f

C
1,O) H
1

(Q. ad
0

f

C
1,O)

H
p
(0)
is exact.
Finally we state the second theorem on Hecke algebras which shall be
proved in section 4.4.
Theorem 3.36 If
t


and ) is a newform in ^

with coecients
in O, then we have

c
p
)

.
Corollary 3.37 With the above notation
#H
1

(Q. ad
0

f

C
1,O),H
1

(Q. ad
0

f

C
1,O) #(

,
t

).
3.4 Isomorphism criteria
The main thrust of Wiles approach is to prove that in many circumstances the
map

: 1

is an isomorphism. For this we will need two criteria from

commutative algebra. The rst was found by Wiles [W3] (but is presented here
in a slightly stronger form due to Lenstra [Len]); the second was developed by
Faltings, from the original arguments of [TW]. In both criteria the notion of
complete intersection plays a vital part.
Proofs of all the results in this section, together with some background,
references and examples, is given in chapter 5.
Denition 3.38 Suppose that is an object of (
C
which is nitely generated
and free as an O-module. Then we call a complete intersection if and only
if for some : Z
0
and some )
1
. .... )
r
O[[A
1
. .... A
r
]] we have

= O[[A
1
. .... A
r
]],()
1
. .... )
r
)
(i.e. there are the same number of generators as relations).
98
We rst record a lemma about complete intersections.
Lemma 3.39 Suppose that 1
t
,1 is a nite extension with ring of integers
O
t
and that is an object of (
C
which is nitely generated and free as an
O-module. Then is a complete intersection if and only if
C
O
t
is.
For the proof, see chapter 5, lemma 5.30.
Now x objects 1 and 1 of (
C
and a surjection of O-algebras : 1 1.
Also assume that 1 is a nitely generated, free O-module. The rst criterion
is as follows.
Theorem 3.40 Suppose that : 1 O. Let = ker ( ) 1 and let
= (Ann
T
(ker )) O. Suppose also that ,= (0). Then the following are
equivalent.
(a) The inequality #,
2
#O, is satised.
(b) The equality #,
2
= #O, is satised.
(c) The rings 1 and 1 are complete intersections, and the map : 1 1
is an isomorphism.
The proof is explained in chapter 5, sections 5.1 to 5.8. (See theorem 5.3.)
For the second criterion let us also x a non-negative integer :. If J
O[[o
1
. .... o
r
]] is an ideal contained in (o
1
. .... o
r
), then by a J-structure we
mean a commutative diagram in (
C
O[[o
1
. .... o
r
]]

O[[A
1
. .... A
r
]] 1
t
1
t

1 1.
such that
(a) 1
t
,(o
1
. .... o
r
)

1 and 1
t
,(o
1
. .... o
r
) 1,
(b) for each ideal 1 J, 1 = ker (O[[o
1
. .... o
r
]] 1
t
,1).
Theorem 3.41 Suppose there exist a sequence of ideals J
n
O[[o
1
. .... o
r
]]
such that J
0
= (o
1
. .... o
r
), J
n
J
n+1
,

n
J
n
= (0) and for each : there exists
a J
n
-structure. Then the map 1 1 is an isomorphism and these rings are
complete intersections.
The proof of this statement is explained in sections 5.9 and 5.10.
99
3.5 The main theorem
We are now in a position to deduce the main theorems. We will keep the
notation and assumptions from the start of section 3.3.
Theorem 3.42 Keep the notation and assumptions of section 3.3. Then, for
all nite sets

,

: 1

is an isomorphism and these rings are

complete intersections.
Remark 3.43 There seems to be a deep link between the fact that

is an
isomorphism and the fact that T

is a complete intersection. The proof of the

theorem divides into two parts. One rst proves it in the minimal case where
= . One then deduces the full theorem from this special case by a dierent
argument. In both these steps the facts that

is an isomorphism and that

T

is a complete intersection are proved simultaneously.

Proof of theorem 3.42: Note that to prove the theorem we may extend scalars
if necessary (by lemma 3.39) and hence assume that both of the following hold:
The eigenvalues of all elements of the image of are rational over /.
There is a newform ) in ^

with coecients in O, hence an O-algebra

homomorphism T

O.
We rst prove that

: 1

is an isomorphism and that the rings

1

and T

are complete intersections.

Note that according to theorem 2.49 and lemma 3.24 we can nd an integer
: 0 and for each : Z
>0
we can nd a set Q
n
of : primes such that
if Q
n
then 1 mod /
n
;
if Q
n
then is unramied at and (Frob
q
) has distinct eigenvalues;
1
Q
n
can be topologically generated by : elements as a O-algebra.
Let J
t
n
= ((o
i
+ 1)

n
i
1 : i = 1. . . . . :) O[[o
1
. . . . . o
r
]], where the :
i
are
chosen such that O[[o
1
. . . . . o
r
]],J
n

= O[
Q
n
]. Then for each : we have a
diagram
O[[o
1
. . . . . o
r
]]

O[[A
1
. . . . . A
r
]] 1
Q
n
T
Q
n

1

.
where
100
(a) 1
Q
n
,(o
1
. . . . . o
r
)

1

(see corollary 2.45),

(b) T
Q
n
is nite and free over O[[o
1
. . . . . o
r
]],J
t
n
(see theorem 3.31),
(c) T
Q
n
,(o
1
. . . . . o
r
)

T

(see corollary 3.32).

Let J
n
= ((o
i
+ 1)

n
1 : i = 1. . . . . :). Replacing T
Q
n
and 1
Q
n
by T
Q
n
,J
n
and 1
Q
n
,J
n
we see that we have a J
n
-structure for every :. Theorem 3.42 for
= now follows from the criterion of theorem 3.41.
We now turn to the proof of theorem 3.42 in the general case. By theo-
rem 2.41 and theorem 3.40 we see that
#H
1

(Q. ad
0

f

C
1,O) = #O,

. (3.5.1)
and so applying corollary 3.37 we see that for any

#H
1

(Q. ad
0

f

C
1,O) #O,

. (3.5.2)
A second application of theorems 2.41 and 3.40 allows us to deduce theo-
rem 3.42. 2
Remark 3.44 In certain cases where

= (1), the bound (3.5.1) on the order

of the Selmer group H
1

(Q. ad
0

C
1,O) also follows from the previous work
of Flach, by a dierent method. See [Fl1] for details.
Corollary 3.45 Keep the notation of theorem 3.42 and suppose that ) is a
newform in ^

with coecients in O.
(a) We have
#H
1

(Q. ad
0

f

C
1,O) = #O,

< .
where

was dened in equation (3.3.1) after remark 3.33.

(b) If
t


then
(0) H
1

(Q. ad
0

f

C
1,O) H
1

(Q. ad
0

f

C
1,O)

H
p
(0)
is exact, where the groups H
p
and H

were dened before proposition 3.35.

Proof: The rst part now follows as a direct consequence of theorem 3.42 and
another application of theorems 2.41 and 3.40. The second part follows from
the rst, together with proposition 3.35 and theorem 3.36. 2
101
Corollary 3.46 Suppose : G
Q
G1
2
(1) is a continuous representation
and let denote its reduction. Suppose also that
(a) is irreducible and modular,
(b) if j ,= / then [
I
p

_
1
0 1
_
,
(c) [
G

is semi-stable,
(d) det = c.
Then is modular.
Proof: We let denote the set of primes in

at which is ramied. Then
: G
Q
G1
2
(O) is a deformation of of type , so there is an O-algebra
homomorphism 1

O such that =
univ

O. Since

: 1

is an
isomorphism by theorem 3.42, it follows that there is a homomorphism T

O sending 1
p
to tr ((Frob
p
)). Since such a homomorphism is necessarily of
the form 1
p
c
p
()) for some newform ), it follows that is equivalent to
f
and hence is modular. 2
Corollary 3.47 Suppose that 1,Q is a semi-stable elliptic curve such that

E,3
is irreducible. Then 1 is modular.
Proof: One need only apply the last corollary with / = 3 and theorem 3.14. 2
3.6 Applications
The Shimura-Taniyama conjecture for semi-stable elliptic curves:
Theorem 3.48 If 1,Q is a semistable elliptic curve, then 1 is modular.
Proof: By corollary 3.47, it is enough to show that 1 is modular when its
associated mod 3 representation
E,3
is reducible, i.e., when 1 has a subgroup
of order 3 dened over Q. Consider the group 1[5] of 5-division points of 1.
The mod 5 Galois representation
E,5
associated to 1[5] is irreducible: for
otherwise, 1 would have a subgroup of order 15 dened over Q, and would
give rise to a (non-cuspidal) rational point on the modular curve A
0
(15). This
curve is of genus one, and is known to have only 4 non-cuspidal rational points,
which do not correspond to semi-stable elliptic curves (and, at any rate, are
known to correspond to modular elliptic curves). Hence we know that
E,5
satises all the assumptions of corollary 3.46, except the (crucial!) modularity
property. To show that
E,5
is modular, one starts with
102
Lemma 3.49 There is an auxiliary (semi-stable) elliptic curve ,Q which
satises
(a) [5] 1[5] as G
Q
-modules.
(b) [3] is an irreducible G
Q
-module.
Proof: Let 1
t
(5) be the curve over Q which classies elliptic curves together
with an isomorphism 1[5] [5] compatible with Weil pairings. Elliptic
curves over Q satisfying (a) correspond to rational points in 1
t
(5)(Q). Ad-
joining a nite set of points to 1
t
(5) yields its compactication A
t
(5) which
is a twist of the modular curve A(5) with full level 5 structure. (I.e., it be-
comes isomorphic to this curve, over

Q.) As was shown by Klein, the modular
curve A(5) over C has genus 0. Since A
t
(5) has a point r
0
dened over Q
corresponding to 1, it is isomorphic over Q to P
1
. The rational points of 1
t
(5)
therefore give a plentiful supply of elliptic curves satisfying condition (a). Now
consider the curve 1
t
(5. 3) classifying elliptic curves with an isomorphism
1[5] [5] (respecting Weil pairings) and a subgroup of of order 3. One
checks that the compactication of 1
t
(5. 3) has genus greater than 1, hence
has only nitely many rational points by Faltings theorem (the Mordell con-
jecture). It follows that only nitely many points 1
t
(5)(Q) are in the image of
1
t
(5. 3)(Q) under the natural map 1
t
(5. 3) 1
t
(5). Hence for all but nitely
many points r in 1
t
(5)(Q), the corresponding elliptic curve satises (b)
since it has no rational subgroup of order 3. Choosing r arbitrarily close in
the 5-adic topology to r
0
, we nd that the elliptic curve associated to r is
semistable and satises the two conditions in the lemma. 2
We can now nish the proof of theorem 3.48. Applying corollary 3.47 to
the curve , we nd that is modular. Hence so is the mod 5 representation

A,5

E,5
. Now applying corollary 3.46 with / = 5 and the representation
of G
Q
acting on the 5-adic Tate module of 1, we nd that
E,5
is modular,
and hence, so is 1, as was to be shown.
Remark 3.50 Wiles original argument uses Hilberts irreducibility theorem
where we have used Faltings theorem. The alternative presented here is based
on a remark of Karl Rubin.
Remark 3.51 The results of [W3] and [TW] actually apply to a larger class
of elliptic curves than those which are semistable. In [Di2], their methods are
further strengthened to prove that all elliptic curves which have semi-stable
reduction at 3 and 5 are modular.
103
Remark 3.52 Rubin and Silverberg observed that an elliptic curve of the
form
2
= r(r c)(r + /) has a twist with semi-stable reduction at all odd
primes, hence is modular by [Di2]. In fact it is shown in [DK] that their
observation together with the general results of [W3] and [TW] already imply
modularity.
Fermats Last Theorem: As was already mentioned in the introduction,
the Shimura-Taniyama conjecture for semi-stable elliptic curves (and, more
precisely, for the elliptic curves that arise in Freys construction explained in
section 2.2) implies Fermats Last Theorem.
More precisely, suppose that there is a non-trivial solution to the Fermat
equation r

= .

, with / 3. By theorem 2.15 the Frey curve constructed

from this solution (cf. section 2.2) is a semistable elliptic curve 1,Q whose
associated mod / representation
E,
is irreducible, unramied outside 2/ and is
good at /. Serres conjecture predicts that
E,
arises from a newform of weight
2 and level 2; the lowering the level result of Ribet [R5] (cf. theorem 3.15)
actually proves this, once we know that 1 is modular, i.e.,
E,
arises from a
modular form of weight 2 and some level. But this is a contradiction, since
there are no modular forms of weight two and level two: such forms would
correspond to holomorphic dierentials on the modular curve A
0
(2) which is
of genus 0. This contradiction completes the proof.
Values of 1-functions: Also mentioned in the introduction was the relation-
ship between the calculation of the Selmer group (3.5.1) and certain cases of
a conjecture of Bloch-Kato [BK], called the Tamagawa number conjecture. It
was in this context that partial results were obtained by Flach in [Fl1] (cf.
remark 3.44).
If ) is a newform of weight 2, then one can associate to ) a certain sym-
metric square 1-function 1(Symm
2
). :). We shall recall the denition in sec-
tion 4.4 and explain how a method of Hida establishes a relationship between
1(Symm
2
). 2) and O,

in the setting of corollary 3.45. We may therefore

regard part (a) of that corollary as a relationship between 1(Symm
2
). 2) and
the size of a Selmer group. While the result is in the spirit of the Tamagawa
number conjecture of [BK], we have not veried that the relevant cases of the
conjecture can be deduced from it. We shall however state a partial result
in the context of semistable elliptic curves. The reader can consult [Fl1] and
[Fl2] for a discussion of the relation to the Tamagawa number conjecture.
104
Suppose that 1 is a semistable elliptic curve over Q of conductor `
E
and
(minimal) discriminant

E
=

p[N
E
j
d
p
.
The symmetric square 1-function associated to 1 is dened by
1(Symm
2
1. :) =

p
1
p
(Symm
2
1. :)
where the Euler factors 1
p
(Symm
2
1. :) are dened as follows:
If j `
E
, then
1
p
(Symm
2
1. :) = ((1
2
p
j
s
)(1 j
1s
)(1
2
p
j
s
))
1
where
p
and
p
are the roots roots of A
2
c
p
A+j with c
p
= j+1`
p
as in section 1.1).
If j[`
E
, then 1
p
(Symm
2
1. :) = (1 j
s
)
1
.
Let

E
=
_
E(C)

Neron
E

Neron
E
where
Neron
E
is the Neron dierential dened in section 1.1. Since 1 is modular
by theorem 3.48, a method of Shimura (see [Shi4] and the introduction of [St])
establishes the analytic continuation of 1(Symm
2
1. :) to an entire function
and shows that 1(Symm
2
1. 2) is a non-zero rational multiple of i
E
. We
now explain how to deduce the following theorem from Wiles results and a
formula of Hida, corollary 4.21.
Theorem 3.53 Suppose that 1 is a semistable elliptic curve and / is a prime
such that

E,
is irreducible, and
/ does not divide 2

p[N
E
d
p
.
Then the /-part of
`
E
1(Symm
2
1. 2)
i
E
is the order of
H
1

(Q. ad
0

E,

,Z

).
105
Sketch of proof: Since 1 is semistable, it is modular by theorem 3.48. Letting
) denote the associated newform, we have `
f
= `
E
and 1(Symm
2
1. :) =
1(Symm
2
). :). Furthermore the representations
E,
and

f
: G
Q
G1
2
(Z

)
are equivalent, since they are equivalent over Q

and =
E,
is irreducible.
The conditions on 1 and / ensure that and ) satisfy the hypotheses of
corollary 3.45 with = , so that
#H
1

(Q. ad
0

E,

,Z

) = #(Z

).
To apply Hidas formula, corollary 4.21, it remains to relate det and
E
where the matrix G1
2
(C) is dened in section 4.4. Using that 1 is
semistable, / is odd and
E,
is irreducible, one obtains a modular parametriza-
tion : A
0
(`
E
) 1 such that

: H
1
(1. Z

) H
1
(A
0
(`
E
). Z

) has torsion-free cokernel;

the Manin constant for is not divisible by / ([Maz2], sec. 4(a)).
One can then verify that
1
E
det is an /-adic unit and theorem 3.53 follows
from corollary 4.21. 2
4 Hecke algebras
4.1 Full Hecke algebras
Suppose that 1 is a nite extension of Q

for some prime /. Let O denote

its ring of integers and let / = O, where is the maximal ideal of O. Fix
embeddings 1

Q

,

Q

Q

and

Q C. Recall that we dened Hecke
algebras over O in two dierent contexts:
In section 1.6 as an algebra T
C
generated by the full set of Hecke oper-
ators acting on a space of modular forms;
In section 3.3 as a certain subring T

of a product of elds of Fourier

coecients of newforms giving rise to the same mod / representation.
The rst of these provides a concrete geometric description useful for establish-
ing properties of the ne structure of the algebra; the second yields a reduced
ring which is more easily interpreted as the coecient ring of a Galois repre-
sentation. In the next section we shall relate the two notions by identifying
106
the reduced Hecke algebras of the form T

as localizations of the full Hecke

algebras of the form T
C
. Before doing so we need to recall some fundamental
properties of the algebras T
K
, T
C
and T
k
.
Let =
H
(`) for some positive integer ` and subgroup H of (Z,`Z)

(see section 1.2). We let T

Z
denote the subring of End (o
2
()) generated by
the operators 1
n
for all positive integers : and d) for all d (Z,`Z)

. If 1
is a ring, then T
R
denotes the 1-algebra T
Z
1. Recall that T
R
acts faithfully
on o
2
(. 1) and is nitely generated and free as an 1-module. (This holds for
1 = Z, hence for arbitrary 1.)
We rst record the following lemma:
Lemma 4.1 (a) T
R
is generated as an 1-algebra by either of the following
sets of elements:
1
n
for all positive integers :.
1
p
for all primes j and d) for all d in (Z,`Z)

.
(b) Suppose that 1 is a positive integer relatively prime to `. If either 1
is odd or 2 is invertible in 1, then T
R
is generated as an 1-algebra by
either of the following sets of elements:
1
n
for all positive integers : relatively prime to 1.
1
p
for all primes j not dividing 1, and d) for all d in (Z,`Z)

.
For a proof of (a), see [DI], prop. 3.5.1; for (b), see p. 491 of [W3].
The spectrum of T
C
: First note that T
K
and T
k
are Artinian, hence have
only a nite number of prime ideals, all of which are maximal. Since T
C
is
nitely generated and free as an O-module, its maximal (resp. minimal) prime
ideals are those lying over the prime (resp. (0)) of O. (This follows from the
going-up and going-down theorems, [Mat] thms. 9.4 and 9.5, for example.) It
follows that the natural maps
T
C
T
C

C
1

= T
K
; and T
C
T
C

C
/

= T
k
induce bijections
maximal ideals of T
K
minimal primes of T
C
and
maximal ideals of T
k
maximal primes of T
C
.
Moreover, since O is complete we have (by [Mat] thms. 8.7 and 8.15, for
example) that the natural map
T
C

m
T
m
107
is an isomorphism, where the product is over the nite set of maximal ideals m
of T
C
and T
m
denotes the localization of T
C
at m. Furthermore each T
m
is a
complete local O-algebra which is nitely generated and free as an O-module,
and each minimal prime T of T
C
is contained in a unique m.
Now suppose that ) =

c
n

n
is a normalized eigenform in o
2
(.

1) for
the operators 1
n
for all : 1. Then 1
n
c
n
denes a map T
Z


1 and
induces a 1-algebra homomorphism
f
: T
K


1. The image is the nite
extension of 1 generated by the c
n
, and the kernel is a maximal ideal of T
K
which depends only on the G
K
-conjugacy class of ). Similarly a G
k
-conjugacy
class of normalized eigenforms in o
2
(.

/) gives rise to a maximal ideal of T

k
.
Recall also that a normalized eigenform ) in o
2
(.

1) has coecients in
O
K
, hence gives rise by reduction to a normalized eigenform

) in o
2
(.

/).
Furthermore if ) and p are G
K
-conjugate, then

) and p are G
k
-conjugate.
We have thus constructed a diagram of maps of nite sets whose commu-
tativity is easily veried.
_
normalized eigenforms in
o
2
(.

1) modulo G
K
-conjugacy
_

_
normalized eigenforms in
o
2
(.

/) modulo G
k
-conjugacy
_

maximal ideals of T
K
maximal ideals of T
k

minimal primes of T
C
maximal primes of T
C
.
(4.1.1)
Proposition 4.2 The vertical maps are bijective, and the horizontal maps are
surjective.
Proof: For the injectivity of the upper-left vertical arrow, note that if p is
a maximal ideal of T
K
, then all 1-algebra homomorphisms T
K
,p

1 are
obtained from a single one by composing with an element of G
K
. For the
surjectivity, let 1
t
= T
K
,p and p
t
denote the kernel of the natural 1
t
-algebra
homomorphism T
K
1
t
. Since T
K
acts faithfully on o
2
(. 1
t
), the local-
ization o
2
(. 1
t
)
p
is non-zero, hence so is o
2
(. 1
t
)[p
t
]. (For an 1-module `
and an ideal 1 of 1, we write `[1] for the intersection over the elements : in 1
of the kernels of : : ` `.) It follows that there is a normalized eigenform
) in o
2
(. 1
t
) so that p
t
is the kernel of the
t
f
: T
K
1
t
, and therefore p is
the kernel of
f
. To prove that the upper-right vertical arrow is bijective, note
that the above arguments carry over with 1 replaced by /. 2
Remark 4.3 The surjectivity of the top arrow is called the Deligne-Serre
lifting lemma ([DS], lemma 6.11).
108
Suppose that m is a maximal ideal of T
C
. Note that the maximal ideals
of T
K
mapping to m are precisely those p for which p T
C
is contained in m.
Note also that the natural map

m
: T
m

C
1

p
T
p
is an isomorphism, where the product is over such p.
It is straightforward to check that the above constructions are well-behaved
with respect to replacing the eld 1 by an extension 1
t
. More precisely,
for each set o in the above diagram, there is a natural surjective map from
: o
t
o where o
t
is dened by replacing 1 with 1
t
, and these maps are
compatible with the maps in the diagram. Furthermore the maps
T
p

K
1
t

1
(p)
T
p
and T
m

C
O
t

1
(m)
T
m

are isomorphisms by which

m

K
1
t
can be identied with

1
(m)

m
.
Associated Galois representations: Suppose that p is a maximal ideal of
T
K
and m is the associated maximal ideal of T
C
. By lemma 1.39,
T

(J
H
(`))
Z

1
is free of rank two over T
K
, so reduction mod p yields a two-dimensional vector
space over the eld T
K
,p endowed with an action of G
Q
. The resulting Galois
representation

p
: G
Q
G1
2
(T
K
,p)
is unramied at all primes j not dividing `/, and for such j the characteristic
polynomial of
p
(Frob
p
) is
A
2
1
p
A +jj) mod p.
If / is odd, then
p
is dened over T
C
,m and we write
m
for its semisimpli-
cation

m
: G
Q
G1
2
(T
C
,m).
Thus
m
is unramied at primes j not dividing `/ and the characteristic
polynomial of Frob
p
is
A
2
1
p
A +jj) mod m.
109
Suppose that p is in the G
K
-conjugacy class of eigenforms in o
2
() corre-
sponding to p. If p is a newform then T
K
,p is isomorphic to the eld denoted
1
t
g
in section 3.1 and
p
can be identied with the representation

g
: G
Q
G1
2
(1
t
g
)
considered in theorem 3.1. If also / is odd then
g
is obtained from
m
by
extending scalars. More generally suppose that p is not necessarily a newform
and consider the associated newform). Let 1 denote the product of the primes
which divide ` but not `
f
. Let T
(D)
K
denote the 1-subalgebra of T
K
generated
by the operators 1
n
for : relatively prime to 1 and d) for d in (Z,`Z)

. Let

t
=
H
(`
f
) where H
t
is the image of H in (Z,`
f
Z)

, and let T
t
K
= T
t
Z
1
where T
t
Z
is the Hecke algebra acting on o
2
(
t
). Restriction of operators denes
a natural homomorphism T
(D)
K
T
t
K
which is surjective by lemma 4.1. The
composite with T
t
K
1
t
f
factors through the eld T
(D)
K
,(pT
(D)
K
), so we may
identify 1
t
f
with a subeld of T
K
,p and
p
is then equivalent to the extension
of scalars of
f
. If / is odd, then
m
and
f
are dened and equivalent over a
common subeld of T
C
,m and /
f
.
The structure of T
K
: We now give an explicit description of T
K
in the case
that 1 contains the coecients of all eigenforms of level dividing `. Let ^

denote the set of newforms in o

2
(), i.e., the set of newforms ) of level `
f
dividing ` such that H is contained in the kernel of the character
(Z,`Z)

(Z,`
f
Z)

f
1

.
By theorem 1.22
o
2
(. 1) =

fA

o
K,f
.
where o
K,f
is spanned by the linear independent elements
)(c) [ c divides `,`
f
.
For each ) =

c
n
())
n
in ^

, let
T
K,f
denote the image of T
K
in End
K
o
K,f
;

K,f
denote the polynomial ring over 1 in the variables n
f,p
indexed by
the prime divisors of `,`
f
;
1
K,f
denote the ideal in
K,f
generated by the polynomials
1
f,p
(n
f,p
) = n
v
p
(N/N
f
)1
f,p
(n
2
f,p
c
p
())n
f,p
+
f
(j)j).
for primes j dividing `,`
f
(setting
f
(j) = 0 if j divides `
f
).
110
Consider the 1-algebra homomorphism
K,f
T
K,f
dened by mapping n
f,p
to the operator 1
p
. Since 1
f,p
is the characteristic polynomial of 1
p
on the span
of )(cj
i
) [ i = 1. . . . .
p
(`,`
f
) for each c dividing `,(`
f
j
v
p
(N/N
f
)
), we see
that 1
K,f
is contained in the kernel of
K,f
T
K,f
. Taking the product over
) in ^

, we have a surjective 1-algebra homomorphism

K,f
,1
K,f

f
T
K,f
.
Since the natural map T
K

f
T
K,f
is injective and
dim
K
T
K
= dim
C
o
2
(. C) =

0
(`,`
f
) =

f
dim
K

K,f
,1
K,f
.
we conclude
Lemma 4.4 There is an isomorphism of 1-algebras:
: T
K

fA

K,f
,1
K,f
dened by
(1
p
)
f
= c
p
()) if j is a prime not dividing `,`
f
;
(1
p
)
f
= n
f,p
mod 1
K,f
if j is a prime dividing `,`
f
;
(d))
f
=
f
(d) if d is relatively prime to `.
Remark 4.5 It follows that the algebra T
K
is a complete intersection over
1 in the sense that it is a nite-dimensional 1-algebra of the form
1[A
1
. . . . . A
r
],(1
1
. . . . . 1
r
)
for some :.
4.2 Reduced Hecke algebras
As in section 4.1, we suppose 1 is a nite extension of Q

with ring of integers

O and residue eld /, and x embeddings 1

Q

,

Q

Q

and

Q C. We
assume in this section that / is odd and we x a representation
: G
Q
G1
2
(/)
111
which is modular (denition 3.12). Thus is equivalent to
m
(over T
C
,m) for
some =
H
(`) and maximal ideal m of T
C
.
We suppose also that has the properties listed at the beginning of sec-
tion 3.3 and that is a nite set of primes contained in

(denition 3.25).
We shall show that and m can be chosen so that the reduced Hecke alge-
bra T

can be identied with a localization T

m
of the full Hecke algebra T
C
.
The main result is due to Wiles ([W3], prop. 2.15), but we also explain an
important variant ([TW], lemma 1) which arises when considering special sets
of primes Q as in section 2.8 above.
T

and the full Hecke algebra: Let =

0
(`

) where
`

= /

`( )

p
j
dim
I
p
= /

p[N( )
j

p
j
2
(4.2.1)
with = 0 if is good and / , and = 1 otherwise.
Suppose that ) is a newform in ^

. Recall that ^

, dened in section 3.3,

is the set of newforms ) in o
2
() such that is equivalent to
f
over /
f
(lemma 3.26). Note that these representations are equivalent if and only if
c
p
()) mod
t
= tr ( (Frob
p
)) for all j `

/,
where
t
is maximal ideal of the ring of integers of 1
t
f
. There is then a
normalized eigenform p in o
2
(. 1
t
f
) characterized by
c
p
(p) = c
p
()) if j does not divide `

,`
f
;
c
p
(p) = 0 if j ,= / and j divides `

,`
f
;
c

(p) is the unit root of A

2
c

())A +/ if / divides `

,`
f
.
Note that if / divides `

,`
f
, then = 1 and is at, hence is ordinary and
c

()) is a unit by theorem 3.1 (f).

Lemma 4.6 For ) and p as above, the reduction p is the normalized eigenform
in o
2
(. /
t
) characterized by:
c
p
( p) = tr
I
p
(Frob
p
) if j = / or j is not in (recall that
I
p
denotes
the representation on 1
p
-coinvariants);
c
p
( p) = 0 otherwise.
Furthermore, if 1 is as in lemma 4.4, then p is the unique normalized eigen-
form in o
K,f
with this reduction.
112
Proof: If j is not in / then the formula for c
p
( p) = c
p
(

)) follows from
theorem 3.1, parts (a) and (e). If j = /, then we use parts (f) and (g) of
theorem 3.1. Otherwise, we have c
p
( p) = 0 by theorem 1.27(d) (if j
2
[`
f
and
) has trivial character, then c
p
()) = 0). The uniqueness follows from the fact
that if j[[`
f
, then c
p
()) = 1 (by theorem 1.27(b)). 2
Note in particular that p has coecients in / and is independent of the
choice of ) in ^

. We let m denote the corresponding maximal ideal of T

C
.
Proposition 4.7 There is an isomorphism T

T
m
of O-algebras such
that 1
p
1
p
for all primes j not dividing `

/.
Proof: Replacing 1 by a larger eld 1
t
, we have T

C
O
t

T
t

(lemma 3.27)
and T
m

C
O
t

T
t
m
(since there is a unique maximal ideal m
t
of T
C
over
m). We are therefore reduced to the case where 1 is as in lemma 4.4. Note
that ^

is the set of newforms ) in ^

such that
c
p
(

)) = tr ( (Frob
p
)) for all j `

/.
(We have 1 = 1
t
f
and write

) for ) mod .)
We now dene an isomorphism of 1-algebras
: T
m
1

fA

1.
such that for each ) ^

:
(1
p
)
f
= c
p
()) if j is not in ;
(1
p
)
f
= 0 if j is in /;
(1

)
f
is the unit root of A
2
c

())A +
f
(/)/ if / is in .
Recall that T
m
1

=

p
T
p
where p runs over the primes of T
K
whose
preimage in T
C
is contained in m. Thus according to lemma 4.4 we have
T
m
1

fA

p/
f
(
K,f
,1
K,f
)
p
.
where /
f
is the set of prime ideals in
K,f
,1
K,f
whose preimage in T
C
is
contained in m. If ) is not in ^

, then /
f
is empty. If ) is in ^

, then
/
f
consists only of the kernel p
f
of the map dened by n
p,f
c
p
(p) where
p is the eigenform of lemma 4.6. Furthermore the maps n
p,f
c
p
(p) induce
113
isomorphisms (
K,f
,1
K,f
)
p
f

1, and taking the product over ) ^

, we
obtain the desired isomorphism .
Identifying T

with the O-subalgebra of

fA

1 generated by the ele-

ments 1
p
= (c
p
()))
f
for all j not dividing `

/, it suces to prove that T

contains (1
p
) for all j dividing `( )/. Now observe that for each ) in ^

the representation
f
is isomorphic to
G
Q
G1
2
(T

) G1
2
(1)
obtained by composing
mod

with the projection T

1. It follows that for j

dividing `( )/

, the image of (
mod

)
I
p
(Frob
p
) in 1 is (
f
)
I
p
(Frob
p
) = c
p
())
f
and therefore
(
mod

)
I
p
(Frob
p
) = (1
p
).
Finally if = 0, then we conclude from lemma 4.1 that T

contains (1

). 2
A twisted variant: Suppose now that is a nite set of primes contained
in

/ such that if j is in then (Frob
p
) has eigenvalues
p
and
p
in
/ satisfying

p
,
p
,= j
1
. (4.2.2)
Note that (4.2.2) is equivalent to
tr ( (Frob
p
))
2
,= (j + 1)
2
. (4.2.3)
Let Q denote the set of primes in such that 1 mod /. The set of primes Q
is therefore as in section 2.8. Choosing an eigenvalue
q
of (Frob
q
) for each
Q as in section 2.8, we regard 1
Q
and hence T
Q
as an O[
Q
]-algebra.
(Recall that
Q
is the the maximal quotient of (Z,(

qQ
)Z)

of /-power
order.)
Instead of working with
0
(`

) as in proposition 4.7, we shall now work

with the group
=
0
(`

)
1
(`). (4.2.4)
where
` =

pQ
j
2

qQ
.
Remark 4.8 We are about to relate T
Q
to a localization of T
C
, where T
C
is now dened using the Hecke operators on o
2
(), with dened by (4.2.4).
Recall that T
Q
is dened using modular forms with trivial character, but note
that the modular forms involved in the denition of T
C
may have non-trivial
character. The purpose of establishing this relationship is to give a concrete
realization of the image of
Q
in T
Q
for the purpose of proving theorem 3.31.
114
Suppose that ) is a newform pair in ^

. For each in Q, let

q
()) be the
root of A
2
c
q
())A + = 0 (in 1
t
f
) whose image in /
f
is
q
. We let p denote
the unique normalized eigenform in o
2
(. 1
t
f
) of trivial character such that
c
p
(p) = c
p
()) if j is not in ;
c
p
(p) =
p
()) if j is in Q;
c
p
(p) = 0 otherwise.
The reduction p is the unique normalized eigenform in o
2
(. /
t
) of trivial char-
acter such that
c
p
( p) = tr
I
p
(Frob
p
) if j is not in ;
c
p
( p) =
p
if j is in Q;
c
p
( p) = 0 otherwise.
Thus p has coecients in / and is independent of the choice of ) in ^

. We
let m denote the corresponding maximal ideal of T
C
.
Suppose for the moment that 1 contains the coecients of all eigenforms
of level dividing `

. Let ^
t
denote the set of newforms p in ^

such that

g
;
c
p
( p) =
p
for all j dividing `
g
,`

.
Suppose we are given a newform p ^
t
. Let
g
denote its character and
write Q
g
for the conductor of
g
. Note that
g
has trivial reduction and
hence /-power order, and that Q
g
divides Q. By proposition 3.2 we see that
if j Q, then
g
is unramied at j and hence `
g
is not divisible by
j. Furthermore, by theorem 3.1 (e) and (4.2.2), `
g
= `

Q
g
. Let
g
denote
the character of (Z,Q
g
Z)

of /-power order such that

2
g
is the primitive
character associated to
g
. Then
p
g
=

g
(:)c
n
(p)
n
is in ^
Q
g
^
Q
.
Lemma 4.9 The map p p
g
denes a bijection between ^
t
and ^
Q
.
115
Proof: Suppose we are given a newform ) in ^
Q
; i.e., ) is in ^

0
(Q)
and
f
.
For Q, we have (by lemma 2.44 for example) that

f
[
G
q

_

f,q
0
0 c
1
f,q
_
.
where
f,q
: G
q
1

is a character whose reduction is the unramied charac-

ter sending Frob
q
to
q
. Note that the characters
f,q
[
I
q
have /-power order. In
particular `
f
,`

is Q
2
f
where Q
f
is the product of the primes Q such that

f,q
is ramied (theorem 3.1 (d)). Note also that there is a unique character

f
: (Z,Q
f
Z)

such that
1
q
G
Q
Gal (Q(
Q
f
),Q)

f
1

coincides with the restriction of

f,q
to 1
q
for each [Q
f
. (We have written
f
for the character of G
Q
as well as the corresponding Dirichlet character. We
shall also write
f
for the corresponding character of
Q
.) Let p denote the
newform associated to the eigenform

1
f
(:)c
n
())
n
o
2
(
0
(`

)
1
(Q
2
f
). 1).
By proposition 2.6 and theorem 3.1 we have

g

f

1
f
,

g
=
2
f
,
`
g
= `

Q
f
and
c
q
(p) = (
f
[
1
G
q

f,q
)(Frob
q
) has reduction

f,q
(Frob
q
) =
q
for all divid-
ing Q
f
.
Therefore p is in ^
t
and ) = p
g
. In particular p p
g
is surjective and

g
=
g
g
. Injectivity follows as well on noting that if p
g
= p
t

g
then

g
=
g
, hence p = p
t
. 2
Recall that T
C
contains the operators d) for d in (Z,`Z)

, and note that

d) 1 m for all d. Let H
t
denote the kernel of
(Z,`Z)

(Z,(

qQ
)Z)

Q
.
116
Since the order of H
t
is not divisible by /, we nd that

dH

d) , m. It
follows that if d H
t
then d) = 1 in T
m
. We may therefore regard T
m
as an O[
Q
]-algebra via the map d d). Recall that T
Q
is considered an
O[
Q
]-algebra via the map
O[
Q
] 1
Q
T
Q
.
If j is a prime not in Q, let r
p
denote the unique element of
Q
such that
r
2
p
= j.
Proposition 4.10 There is an isomorphism T
Q

T
m
of O[
Q
]-algebras
such that 1
p
r
p
1
p
for all primes j , with j /`( ).
Proof: We may enlarge 1 so that we are in the setting of lemma 4.9. We then
dene an isomorphism of 1-algebras
: T
m
1

gA

1
such that for each p ^
t
:
(1
p
)
g
= c
p
(p) if j is not in or if j divides Q
g
,
(1
p
)
g
= 0 if j is in Q,
(1
p
)
g
is the root of A
2
c
p
(p) +
g
(j)j with reduction
p
if j is in Q
but does not divide Q
g
, and
(d)
g
=
g
(d) for all d
Q
.
The existence of such an isomorphism follows from lemma 4.4 which gives
T
m
1

gA

p/
g
(
K,g
,1
K,g
)
p
.
where /
g
is the set of prime ideals in
K,g
,1
K,g
whose preimage in T
C
is
contained in m. If p is not in ^
t
, then /
g
is empty. If p is in ^
t
, then /
g
consists of the prime ideal corresponding to the eigenform whose eigenvalues
are prescribed as above, and one checks that (
K,g
,1
K,g
)
p
= 1.
Viewing T
Q
as a subalgebra of

fA
Q
1 and matching indices via the bijec-
tion ) = p
g
p, we obtain an injective homomorphism of O[
Q
]-algebras

t
: T
Q

gA

1
117
such that

t
(1
p
) = (
g
(j)c
p
(p))
g
= (r
p
1
p
)
for primes j , such that j /`

. Since T
Q
is generated over O by the set
of such 1
p
, we see that
t
(T
Q
) is contained in (T
m
). On the other hand, the
image of
t
contains the image of
Q
, hence contains (d)) for d in (Z,`Z)

as well as (1
p
) for j , with j `

/.
It remains to prove that (1
p
) is in the image of
t
(T
Q
) for j dividing
`

qQ
. Consider the composite
1
Q

Q
T
Q

gA

1.
The pushforward of
univ
Q
is equivalent to

g

g
and therefore

g
is
equivalent to the pushforward of

univ
Q

1
Q
(where
Q
was dened in section 2.8). For primes j dividing `

, we recover
(1
p
) as the image of Frob
p
on the 1
p
-coinvariants, hence (1
p
) is in
t
(T
Q
).
For Q and p ^
t
, the pushforward of
Q
[
1
G
q

q,Q
to the p-component is an
unramied summand of
g
[
G
q
whose reduction sends Frob
q
to
q
. It follows
that this character maps Frob
q
to (1
q
) and we conclude that (1
q
)
t
(T
Q
).
Finally in the case that / does not divide `

, we appeal to lemma 4.1 to

conclude that (1

)
t
(T
Q
). 2
Auxiliary primes: For as in proposition 4.10, i.e. a set of primes satisfying
(4.2.2), a somewhat simpler argument provides a similar a description of T

with replaced by

t
=
0
(`

)
H
(`) (4.2.5)
where H is the /-Sylow subgroup of (Z,`Z)

. While we shall make no direct

use of this, the group
t
will play a role in the proof of theorem 3.31 and
we shall need to choose so that
t
has no elliptic elements; i.e., non-trivial
elements of nite order. This is the case for example if contains a prime
j , 1 mod / with j 3. The group
t
is then contained in
1
(`) for some
integer ` 3, and therefore has no elliptic elements. Indeed if
_
c /
c d
_

has nite order then the roots of A
2
(c +d)A +1 are roots of unity and we
deduce that this matrix is the identity.
In order to show that can be so chosen, we appeal to the following lemma
(cf. [DT2], lemma 3).
118
Lemma 4.11 Suppose that G is a nite group, : G

/

is a character of
order d and : G G1
2
(

/) is a representation. Suppose that for all p , ker

we have
(tr (p))
2
, det (p) = (1 +(p))
2
,(p). (4.2.6)
If d 3, then is reducible.
If d = 3, then is reducible or has projective image isomorphic to
4
.
If d = 2, then [
ker
is reducible.
Proof: Note that if (p) is a scalar then (G) = 1. Hence induces induces
a surjective homomorphism
t
: G
t
C
d
where G
t
is the projective image of
and C
d
is cyclic of order d. Furthermore if d = 2, then every element of
G
t
ker
t
has order 2. The lemma then follows from theorem 2.47(b). 2
4.3 Proof of theorem 3.31
We shall give a proof of theorem 3.31 which is based on the -expansion princi-
ple rather than the method of de Shalit [dS] employed in [TW]. It will be more
convenient to consider the action of the Hecke operators on the full space of
modular forms `
2
(). The Riemann-Roch theorem shows that the dimension
of `
2
() is p +: 1 where p is the genus of the modular curve A

associated
to and : is the number of cusps on A

(see for example [Shi2] thm. 2.23 or

[DI] (12.1.5)).
Eisenstein maximal ideals: One can give an explicit description of a space
of Eisenstein series G
2
() so that
`
2
() = o
2
() G
2
().
(See for example [Hi3], lemma 5.2.)
There is a natural action on `
2
() by the Hecke operators d) for all d
(Z,`Z)

and 1
p
for all primes j. We shall only need to use the operators d)
and 1
p
for j not dividing `, and we let

T
Z
denote the subring of End (`
2
())
these generate. The ring

T
Z
is commutative and is nitely generated and free
as a Z-module. That

T
Z
is nitely generated is proved for example by showing
that `
2
(. Z), the set of forms with integer Fourier coecients at , is stable
under the Hecke operators and contains a basis for `
2
() (see [DI], cor. 12.3.12
and prop. 12.4.1). Alternatively, one can show that

T
Z
acts faithfully on the
cohomology of the non-compact modular curve 1

(cf. section 1.3).

119
For any ring , we write `
2
(. ) for `
2
(. Z) and regard this as a
module for

T
A
:=

T
Z
. If n is a maximal ideal of

T
C
we write

T
n
for the
localization. If m is a maximal ideal of T
C
then we let m denote its preimage
in

T
C
.
We say that a maximal ideal n of

T
C
is Eisenstein if
1
p
j + 1 mod n for all j 1 mod `.
One sees from the explicit description of the Eisenstein series that 1
p
= j+1 on
G
2
() if j 1 mod `. We shall be interested in the non-Eisenstein maximal
ideals because of the following lemma (see [R5], thm. 5.2(c)).
Lemma 4.12 Suppose that / is odd. The representation

m
: G
Q
G1
2
(T
C
,m)
is absolutely irreducible if and only if m is not Eisenstein.
Proof: If
m
is not absolutely irreducible then
m

1

2
with ` divisible by
the product of the conductors of
1
and
2
. Furthermore if / does not divide
`, then one of the characters is unramied at / while the other coincides with
the cyclotomic character c on 1

. Therefore m is Eisenstein. Conversely if m is

Eisenstein then proposition 2.6 implies that
m
restricted to G
Q(
N
)
has trivial
semisimplication from which it follows that
m
is also reducible. 2
Dierentials: Suppose now that is a nite set of primes satisfying (4.2.2)
Moreover we assume that contains a prime j 3 with j , 1 mod /. Let
and
t
be dened as in (4.2.4) and (4.2.5); i.e.
=
0
(`

)
1
(`)

t
=
0
(`

)
H
(`)
where H is the /-Sylow subgroup of (Z,`Z)

. Let A and A
t
denote the
modular curves associated to and
t
.
We rst consider the case where / `

. The curve A has a smooth proper

model A over Z[1,`

`] such that the complement of the cusps parametrizes

cyclic isogenies (1
1
. i
1
) (1
2
. i
2
) of degree `

where 1
j
is an elliptic curve
and i
j
is an embedding j
M
1
j
. (See [DR] or [Kat] for example.) The action
of H on A extends to A and the quotient A
t
= A,H is a smooth model over
Z[1,`

`] for A
t
. The natural projection A A
t
is etale. (The fact that it is
etale on the complement of the cusps follows from the natural moduli-theoretic
description of A
t
using the fact that
H
(`)
1
(j) for some j 3. One
then need only verify that A A
t
is unramied at the cusps.)
For a Z[1,`

`]-algebra , we dene
120

A
= H
0
(A
A
.
1
.
A
/A
),

t
A
= H
0
(A
t
A
.
1
.

A
/A
),

A
= H
0
(A
A
.
1
.
A
/A
(T)),

t
A
= H
0
(A
t
A
.
1
.

A
/A
(T
t
)),
where T (resp. T
t
) denotes the reduced divisor dened by the cusps of A (resp.
A
t
). The -expansion principle and standard base-change arguments allow us
to identify these with o
2
(. ), o
2
(
t
. ), `
2
(. ) and `
2
(
t
. ) (see [Kat],
sec. 1.6, 1.7). We may therefore regard the rst two of these as modules for
T
A
and all of them as modules for

T
A
.
Lemma 4.13 Suppose that = / or 1.
(a)

C

C


=

A
and

t
C

C


=

t
A
.
(b) The natural map A A
t
induces an isomorphism

t
A

H
A
.
(c) If n is a non-Eisenstein maximal ideal of

T
C
, then the localization at n
of
C

C
is an isomorphism.
(d) If m is a maximal ideal of T
A
then
A
[m] is one-dimensional over T
A
,m.
Sketch of proof:
(a) In the case = 1 this follows from the fact that 1 is at over O, so
suppose = /. Identify

C
,

C
with the direct image of

k
under
A
k
A
C
and use that H
1
(A
k
.
1
.
k
/k
(T)) vanishes by Serre duality. The
case of

t
is similar. (See [Maz1] sec. II.3.)
(b) Using the fact that A A
t
is etale one identies the pull-back of
1
.

A
/A
with
1
.
A
/A
and that of T
t
with T.
(c) The map is injective and one proves that its cokernel is free and annihi-
lated by the operators 1
p
(j +1) for all j 1 mod `

`. To prove the
latter assertion, rst observe that it holds with O replaced by C, then
by Z[1,`

`].
121
(d) This follows from

A
[m]

= o
2
(. )[m]

= Hom
A
(T
A
. )[m]

= Hom
A
(T
A
,m. ).
where the middle isomorphism is that of proposition 1.34, but let us
reformulate the argument in a way more easily generalized to the case
of / dividing `

discussed below. Note that

A
[m] is non-zero since
T
m
acts faithfully on
m
. To prove that the dimension is at most one
we can enlarge the eld and assume T
A
,m

= . One then shows
that an eigenform ) in
A
(for all the Hecke operators 1
p
and d)) is
determined by its eigenvalues and the rst coecient of its -expansion.
This follows from the fact that ) is determined by its -expansion and
for all :, c
n
()) = c
1
(1
n
)) and 1
n
can be expressed in terms of the 1
p
and d). (See [Maz1], sec. II.9.) 2
We now explain how the situation changes if / divides `

. In that case
A has a regular model A over Z[/,`

`] with the same moduli-theoretic de-

scription as above. This model is smooth over Z[1,`

`], but A
F

has two
smooth irreducible components crossing at ordinary double points as in [DR]
sec. V.1. The quotient A
t
= A,H is a regular model for A
t
over Z[/,`

`]
and A A
t
is etale. For a Z[/,`

`]-algebra we dene
A
,
t
A
,

A
and

t
A
as before, except that
1
is replaced by the sheaf of regular dierentials
(see [DR] sec. I.2, [Maz1] sec. II.6 or [MRi] sec. 7). Formation of these mod-
ules again commutes with change of the base , but we can no longer identify
them with o
2
(. ), o
2
(
t
. ), `
2
(. ) and `
2
(
t
. ) if / is not invertible
in . For every , there is a natural action of T
A
(resp.

T
A
) on
A
and
t
A
(resp.

A
and

t
A
). In the case of T
A
this is proved by identifying
A
(resp.

t
A
) with the cotangent space at the origin for the Neron model over for
the Jacobian of A (resp. A
t
). In the case of

T
A
the action is dened using
Grothendieck-Serre duality (as in [Maz1] sec. II.6 or [MRi] sec. 7).
Lemma 4.14 If / divides `

then lemma 4.13 carries over with the above

notation and the additional hypothesis 1

, m for part (d).

The proof is essentially the same except that part (d) is more delicate in the
case = /. For the proof in that case we refer the reader to [W3], lemma 2.2
(from which the result stated here is immediate). We remark however that we
shall only use lemma 4.14 if
m
is not good. In that case one can also deduce
the statement here from the argument used in the proof of [MRi] prop. 20,
which shows instead that dim
k

k
[m
t
] = 1 where m
t
is a certain maximal ideal
corresponding to m, but contained in an algebra dened using the operators
d), 1
p
for j ,= / and u

.
122
Remark 4.15 The approach in [TW] to proving theorem 3.31 is based on the
result that under certain hypotheses (T

(J)
Z

O)
m
is free of rank two over
T
m
(see [W3], thm. 2.1 and its corollaries). This result generalizes work of
Mazur Ribet, and Edixhoven (see secs. 14 and 15 of ch. II of [Maz1], thm. 5.2
of [R5], [MRi] and sec. 9 of [Edi]), and the key to its proof is lemmas 4.13(d)
and 4.14. We shall instead give a proof of theorem 3.31 which is based directly
on lemma 4.13, (d).
The fact that (T

(J)
Z

O)
m
is free of rank two over T
m
actually underlies
Wiles approach to many intermediate results along the way to proving the
Shimura-Taniyama conjecture for semistable elliptic curves. We shall appeal
to a special case in the course of proving theorem 3.36 below.
Proof of the theorem: Suppose we are given a representation as in sec-
tion 3.3 and a set of primes Q as in the statement of theorem 3.31. We apply
lemma 4.11 with G = G
Q
and = c to choose an auxiliary prime j not dividing
6`

qQ
such that
j , 1 mod /
tr ( (Frob
p
))
2
,= (j + 1)
2
.
To prove theorem 3.31 we may replace 1 by a larger eld and assume that /
contains the eigenvalues of (Frob
p
).
Let = Q j. Thus is a set of primes as in proposition 4.10 and
contains a prime j 3 such that j , 1 mod /. We let be as in (4.2.4); thus
=
0
(`

)
1
(`)
where ` = j
2

qQ
, and we choose the maximal ideal m of T
C
as in propo-
sition 4.10. Let H denote the /-Sylow subgroup of (Z,`Z)

and regard T
C
and hence T
m
as an O[H]-algebra via d d). In view of proposition 4.10,
theorem 3.31 is equivalent to the following:
Theorem 4.16 T
m
is free over O[H].
Proof: Consider the

T
C
-module

1 = Hom
C
(

C
. O)
and the T
C
-module
1 = Hom
C
(
C
. O).
123
Since A A
t
is unramied, the Riemann-Hurwitz formula implies that
dim`
2
() = p +: 1 = #H(p
t
+:
t
1) = #H dim`
2
(
t
).
where p (resp. p
t
) is the genus and : (resp. :
t
) is the number of cusps of A
(resp. A
t
). It follows from lemma 4.13 (a) (and lemma 4.14) that
dim
K
(

1
C
1) = #H dim
k

t
k
(4.3.1)
On the other hand, by (b) we have that

1,(a. )

1

= Hom
k
(

H
k
. /)

= Hom
k
(

t
k
. /) (4.3.2)
where a = ker (O[H] O) is the augmentation ideal. By Nakayamas lemma
and (4.3.2),

1 is generated as an O[H]-module by d elements where d =
dim
k

t
k
. By (4.3.1) any surjective homomorphism
O[H]
d

1
is in fact an isomorphism. Hence

1 is free over O[H], as is
1
n
= Hom
C
(
n
. O)

= Hom
C
(

n
. O)
for each non-Eisenstein maximal ideal n of

T
C
(by part (c) of lemmas 4.13
and 4.14).
Since
m
is absolutely irreducible, m is not Eisenstein (by lemma 4.12),
and it follows that 1
m
is free over O[H].
Since 1
C
1 is free of rank one over T
K
(lemma 1.34), we have that
1
m

C
1 is free of rank one over T
m

C
1. If / `

then lemma 4.13 implies

that
1,m1

= Hom
k
(
k
[m]. /)
is one-dimensional over /. The same assertion holds if /[`

by lemma 4.14
(from the denition of m in this case, we see that it does not contain 1

). It
follows from Nakayamas lemma that 1
m
is free of rank one over T
m
. Therefore
T
m
is free over O[H]. 2
4.4 Proof of theorem 3.36
Our proof of theorem 3.36 is based on Wiles arguments in ch. 2 of [W3],
which in turn are based on a method of Ribet [R4]. We shall reformulate the
proof somewhat to underscore the relationship observed by Doi and Hida [Hi1]
between the size of O, and the value of an 1-function, but after doing so we
shall also sketch the more direct argument used by Wiles. Two important
ingredients appear in both versions of the argument, and we shall discuss their
proofs in section 4.5. These ingredients are
124
the generalization of a result of Ihara [Ih] used in Ribets argument;
the generalization of a result of Mazur [Maz1] on the structure of the
Tate module T

(J
0
(`)) as a module for T
Z

.
Assume now that we are in the setting of theorem 3.36. In particular, is
a representation as in section 3.3, is a nite set of primes contained in

and is an O-algebra homomorphism T

O arising from a newform ) in

^

with coecients in O.
The symmetric square 1-function: For each prime j let
p
()) and
p
())
denote the roots of the polynomial A
2
c
p
())A +
p
j = 0, where
p
= 0 or 1
according to whether j divides `
f
. Thus we have
1
p
(). :) = (1
p
())j
s
)
1
(1
p
())j
s
)
1
.
If j = / divides `,`
f
, then we require that

()) be the root which is a unit

in O, i.e. c

(p). We dene the symmetric square 1-function associated to ) as

1(Symm
2
). :) =

p
1
p
(Symm
2
). :).
where
1
p
(Symm
2
). :) = (1
2
p
())j
s
)
1
(1
p
())
p
())j
s
)
1
(1
2
p
())j
s
)
1
.
(We caution the reader we have dened the Euler factors rather naively at
primes j such that j
2
divides `
f
.) The product converges absolutely for real
part of : 2 and can be analytically continued to an entire function.
The calculation of

: We write for the kernel of and 1 for the annihilator

of in T

. Since T

is reduced, we have 1 = 0 and 1 has nite index

in T

. Note that
O,

= T

,( 1).
Suppose that we are given a T

-module 1 which is nitely generated and free

over O, and such that 1 1 is free of rank d over T

C
1. Suppose also
that 1 is endowed with a perfect O-bilinear pairing
1 1 O
(r. ) r. )
such that 1r. ) = r. 1) for all r. 1 and 1 T

. Note that to give

such a pairing is equivalent to giving an isomorphism
1

Hom
C
(1. O)
r r. )
125
of T

-modules. We shall refer to the module 1 together with the pairing . )

as a self-dual T

-module of rank d. We then can give a lower bound for the

size of O,

in terms of a basis r
1
. r
2
. . . . . r
d
for the free O-module 1[]:
Lemma 4.17 We have

Odet(r
i
. r
j
))
i,j
.
and equality holds if 1 is free over T

.
Proof: The modules 1[] and 1,1[1] are free of rank d over O and the pairing
. ) induces an isomorphism
1,1[1]

Hom
C
(1[]. O). (4.4.1)
The O-module ` = 1,(1[] 1[1]) is annihilated by

and is generated by
d elements. Furthermore ` is isomorphic to (O,

)
d
if 1 is free over T

. It
follows that
#(O,

)
d
#`.
with equality if 1 is free. The cardinality of ` is that of the cokernel of the
map
1[] Hom
C
(1[]. O)
arising from the pairing . ), and this is precisely
O, det(r
i
. r
j
))
i,j
.
2
Recall that proposition 4.7 establishes an isomorphism between T

and
T
m
, the localization at a certain maximal ideal m of T
C
= T
Z
O, where
T
Z
End (o
2
(
0
(`))).
and where ` = `

is dened in (4.2.1). Write A for A

0
(`) and J for J
0
(`).
Recall that H
1
(A. Z) is endowed with the structure of a T
Z
-module and
T

(J)
Z

O

= H
1
(A. Z) O
with that of a T
C
-module. We also regard
H
1
(A. O)

= Hom(H
1
(A. Z). O)
as a T
C
-module; as such it is naturally isomorphic to
Hom
C
(T

(J)
Z

O. O).
126
The Weil pairing denes a perfect pairing
T

(J) T

(J) Z

(with (c
2i/
n
)
n
as the chosen generator for lim

n(C)). The composite

T

(J)
w

(J) Hom
Z

(T

(J). Z

) (4.4.2)
is an isomorphism of T
Z

-modules where u

is induced by the involution u =

u
N
of A (see section 1.4). Tensoring with O and localizing at m, we regard
1
T
= (T

(J)
Z

O)
m
as a self-dual T

-module of rank 2 via the isomorphism T

= T
m
of proposi-
tion 4.7. We shall write . )
T
for the pairing obtained from (4.4.2). Similarly,
using u

, the cup product and Poincare duality, we regard

1
H
= H
1
(A. O)
m
as a self-dual T

-module of rank 2, with r. )

H
dened by the image of r
u

under the canonical isomorphism of H

2
(A. O) with O. Note that 1
H
is
canonically isomorphic as a T

-module to Hom
C
(1
T
. O). In the next section
we shall discuss the following generalization of a result of Mazur [Maz1].
Theorem 4.18 The T

-modules 1
T
and 1
H
are free.
Corollary 4.19 If r. is a basis for 1
T
[] (resp. 1
H
[]), then

is gen-
erated by r. )
T
(resp. r. )
H
).
This follows from lemma 4.17, theorem 4.18 and skew-symmetry of the pair-
ings.
Hidas formula: We now explain how the value of r. )
H
in corollary 4.19
is related to 1(Symm
2
). 2) by a formula of Hida (see [Hi1] sec. 5, [W3] sec.
4.1).
Recall that the homomorphism
T
C
T
m

= T

O
arises as 1
n
c
n
(p) for a normalized eigenform p in o
2
(
0
(`). 1) whose as-
sociated newform is ). We shall write T for the kernel of this homomorphism;
thus T is the preimage in T
C
of the ideal of T

. Note that T
m
corresponds
to under the isomorphism T
m

= T

.
127
Choose a number eld 1
0
containing the Fourier coecients of p and let
O
0
denote the valuation ring O1
0
. We choose the basis r. for 1
H
[] in
the image of
H
1
(A. O
0
)[T
0
] H
1
(A. O)[T]

H
1
(A. O)
m
[T] = 1
H
[].
where T
0
= T T
C
0
. Let T
C
denote the kernel of the homomorphism

g
: T
C
C
associated to the eigenform p. The two-dimensional complex vector space
(H
1
(A. O
0
)[T
0
])
C
0
C = H
1
(A. C)[T
C
]
has a canonical basis

g
.
g
c.
where we view these dierential forms as cohomology classes and the super-
script c indicates complex conjugation applied to the Fourier coecients. (Re-
call that
g
is the holomorphic dierential on A dened by

c
n
(p)
n1
d,
so
g
c is the antiholomorphic dierential dened by

c
n
(p)
n1
d where
= c
2i
.) Let denote the matrix in G1
2
(C) such that
(
g
.
g
c) = (r. ).
Theorem 4.20 With the above notation we have (up to sign)
r. ) = (i det )
1
`
f
1(Symm
2
). 2)

p[N/N
f
c
p
())).
where
c
p
()) =
_
j

p
1
p
(Symm
2
). 2)
1
if j ,= /
(

())

()))(1
2

())) if j = /.
Proof: We have
r. )
H
det =
_
X

g

wg
c = 8
2
ip. (up)
c
)
where the last . ) denotes the Petersson inner product and we have used that
up
c
= (up)
c
for forms on
0
(`). We then appeal to a formula of Shimura
(see [Shi5] (2.5) and [Hi1] (5.13)) to obtain
p. (up)
c
) = (48)
1
[o1
2
(Z) :
0
(`)]res
s=2
1(p. up. :).
128
By [Shi5] lemma 1, the Dirichlet series 1(p. /. :) is dened by

c
n
(p)c
n
(/):
s
and if p and / are normalized eigenforms then this has an Euler product
expression in which the factors are
(1
p
(p)
p
(p)
p
(/)
p
(/)j
2s
) (1
p
(p)
p
(/)j
s
)
1

(1
p
(p)
p
(/)j
s
)
1
(1
p
(p)
p
(/)j
s
)
1
(1
p
(p)
p
(/)j
s
)
1
.
(where the
p
s and
p
s are dened as they were for )). Using the recipe for
obtaining p from ) (see lemma 4.6) we nd that if / is not in then
1(p. up. :) = 1

(). ). :)

p[N/N
f
j
1
.
(where the subscript indicates that the primes in are removed from the
dening Euler product). If / is in , then we must also multiply by the factor
(1 /
s
)
1
if / divides `
f
;
(1 /
1s
)
1
(1
2

())/
s
)
1
(

())
2s

())) otherwise.
To obtain the formula in this last case, consider the eigenform
)

= )

()))(/)
on
0
(`
f
/). Since
u
N
f

= (/)(/)
1

())))
and
1()

. /)(/). :) =

())/
1s
1()

. ). :).
we get
1()

. u
N
f

. :) = /
1
(

())

())/
2s
)1()

. ). :)
and use the Euler product. The Euler product expression also gives
1

(). ). :) =

N
(: 1)1

(Symm
2
). :)

N
(2: 2)
.
where
N
(:) is the Riemann zeta function with the Euler factors removed at
primes dividing `. Thus
res
s=2
1(p. up. :) =
6

2
`1

(Symm
2
). 2)

p[N/N
f
j
[o1
2
(Z) :
0
(`)]
.
with the extra factor of (1 /
2
)
1
or
(

())

()))(1
2

()))
1
(1 /
1
)
1
if / is in . The theorem follows. 2
129
Corollary 4.21 With the above notation (in particular,

), we have
/
m
`
f
1

(Symm
2
). 2)
i det
O
0
is a generator of

, where
: =
_
0. if / , ,
2. if / and /[`
f
,
3. if / and / `
f
.
Remark 4.22 Without theorem 4.18 we obtain instead that

is contained
in the ideal generated by the expression in the corollary.
Comparing

s: Suppose that
t


is a nite set of primes containing .
In view of corollary 4.21, theorem 3.36 reduces to a comparison of det and
det
t
where
t
is dened using
t
instead of . We shall make the desired
comparison using Wiles generalization of a result of Ihara.
Applying proposition 4.7 to
t
, we have an isomorphism between T

and
T
t
m
where m
t
is a certain maximal ideal of T
t
C
= T
t
Z
O with T
t
Z
acting on
o
2
(
0
(`

)). We write T
t
for the preimage in T
t
C
of
t
. We let `
t
= `

,
A
t
= A
0
(`
t
), J
t
= J
0
(`
t
) and dene self-dual T
t
m
-modules 1
t
T
and 1
t
H
as
above.
We now dene a homomorphism

T
: 1
t
T
1
T
of T
t
m
-modules in the case that
t
= j for some prime j

. We rst
suppose that j ,= /, in which case `
t
= `j
2
and the surjective homomorphism
of O-algebras T
t
m
T
m
(corresponding to T

) is dened by 1
r
1
r
if : ,= j and 1
p
0. The maps , j and j
2
on the upper
half-plane induce maps A
t
A which we denote , and . These give rise
by functoriality to a map
J
t
J J J
r (

r.

r.

r)
which induces an O-linear homomorphism
T

(J
t
)
Z

O (T

(J)
Z

O)
3
.
We compose with the projection to
1
3
T
= (T

(J)
Z

O)
3
m
130
followed by the map 1
3
T
1
T
dened by the matrix (1. j
1
1
p
. j
1
). The
reason for this choice of matrix is that
p
t
= (

j
1

1
p
+j
1

)p (4.4.3)
is the eigenform corresponding to T
t
.
Lemma 4.23 The composite
T

(J
t
)
Z

O 1
T
(4.4.4)
is T
t
C
-linear where 1
T
is regarded as a T
t
C
-module via the surjection
T
t
C
T
t
m
T
m
.
We leave the proof to the reader after pointing out that the map on Hecke
algebras can be rewritten as
T
t
C
T
C
[n],()(n)) T
m
[n],()(n)) T
m
1
p
n n 0.
where )(n) = n
3
1
p
n
2
+jn.
It follows that (4.4.4) factors through 1
t
T
and we thus obtain the desired
homomorphism

T
: 1
t
T
1
T
of T
t
m
-modules. Note that
T
is also the composite of the canonical splitting
1
t
T
T

(J
t
)
Z

O
with (4.4.4).
The construction of
T
is similar for j = / except that we have only two
copies of J and the map 1
2
T
1
T
is given by the matrix (1.
1

) where

is
the unit root of (3.3.2). For arbitrary
t
satisfying
t


we dene
T
as a composite of the maps dened above. It is independent of the choice of
ordering of
t
. Recall that
t
T
is used to denote the dual map 1
T
1
t
T
.
We dene
H
: 1
t
H
1
H
as the adjoint of the map
t
H
which renders the
diagram

t
H
: H
1
(A. O)
m
H
1
(A
t
. O)
m

Hom
C
(1
T
. O) Hom
C
(1
t
T
. O)
commutative (where the vertical maps are the natural isomorphisms and the
bottom one is dual to
T
).
The second crucial result whose discussion we postpone until the next
section is the following generalization of a lemma of Ihara.
131
Lemma 4.24
T
and
H
are surjective.
Suppose again we are in the case
t
= j with j ,= /. One need only
unravel the denition of
t
H
to see that the diagram
H
1
(A. O
0
)[T
0
] H
1
(A
t
. O
0
)[T
t
0
]

1
H

H
1
t
H
.
commutes, where the top arrow is dened as

j
1
(1
p
)

+j
1

.
Extending scalars to C, this map sends
g
to
g
and
g
c to
g
c where p
t
is
dened by (4.4.3). In the case
t
= /, the same assertion holds if O
0
is
chosen so that it contains (

) = c

(p
t
) and the top arrow is dened as

)
1

.
We conclude that if r. is a basis for H
1
(A. O
0
)[T
0
] and the matrices
and
t
are dened using the bases r. for H
1
(A. O
0
)[T
0
] and
t
H
r.
t
H

for H
1
(A
t
. O
0
)[T
t
0
], then
t
= . Theorem 3.36 is now immediate from corol-
lary 4.21. In fact, we have proved that

= (

c
p
)

.
2
Remark 4.25 Note that to obtain the inclusion stated in theorem 3.36 it suf-
ces to apply theorem 4.18 for , rather than both and
t
(cf. remark 4.22).
Wiles argument: The method of [W3] sec. 2.2, like that of Ribet in [R4],
is more direct than the one given above but does not explicitly illustrate the
relation with values of 1-functions. The approach is simply to compute the
composite
T

t
T
. It suces to consider the case
t
= j and we suppose
rst that j ,= /. Let denote the endomorphism of J
3
dened by the matrix
=
_

_
u
t

)u

.
Using the relations u
t
= u and u
t
= u, we nd that
=
_

_
.
132
which can be computed for example by considering its eect on the cotangent
space o
2
(
0
(`))
3
. The result is that
=
_
_
1
2
p
(j + 1) j1
p
j(j + 1)
j1
p
j(j + 1) j1
p
j(j + 1) j1
p
1
2
p
(j + 1)
_
_
.
which we note commutes with the action of T
Z
on J
3
. One then checks that

t
T
= (1. j
1
1
p
. j
1
) (T

()
Z

O)
m
_
1
j
1
1
p
j
1
_
= j
2
(j 1)(1
2
p
(j + 1)
2
).
The case of j = / is similar but simpler. One uses
=
_

_
u
t

)u

=
_
1

/ + 1
/ + 1 1

_
and gets

t
T
= (

/
1

)(1
2

).
Thus in either case we nd that
T

t
T
acts on 1
t
T
[T] by an element of O =
T
m
,T which is a unit times (c
p
). Theorem 3.36 then follows from corol-
lary 4.19 and lemma 4.24.
4.5 Homological results
In this section we sketch the proofs of theorem 4.18 and lemma 4.24, but shall
often refer the reader to ch. 2 of [W3] for more details.
Multiplicities: We rst consider theorem 4.18, generalizing a result proved
by Mazur in sections II.14 and II.15 of [Maz1]. Recall that 1
T
(resp. 1
H
)
is dened as (T

(J
0
(`))
C
)
m
(resp. H
1
(A
0
(`). O)
m
) where m is a maximal
of the ideal of the Hecke algebra T
C
generated by the Hecke operators on
o
2
(
0
(`)). We are assuming moreover that
m
is irreducible and that one of
the following holds:
/ does not divide `;
/
2
does not divide ` and 1

, m.
We wish to prove that 1
T
and 1
H
are free over T
m
.
Since 1
H
and 1
T
are isomorphic as T
m
-modules, it suces to prove that
1
T
is free. Furthermore, to prove that 1
T
is free, it suces to prove that the
133
localization of T

(J
0
(`)) at m T
Z

is free; i.e., we may replace O by Z

and
m by its intersection with T
Z

and consider
T

(J
0
(`))
m
.
Since T

(J
0
(`))
Z

is free is free of rank two over T

Q

, it suces to prove
that
dim
T/m
(1
T
,m1
T
) = 2.
Since
T

(J
0
(`)),/T

(J
0
(`))

= J
0
(`)[/]

= Hom
F

(J
0
(`)[/]. F

)
as T
F

-modules, it suces to prove the following cases of [W3] thm. 2.1.

Theorem 4.26 Let m be a maximal ideal of T
Z

. Suppose that
m
is irre-
ducible and that either
(a) / does not divide `, or
(b) /
2
` and 1

, m.
Then
dim
T
Z

/m
J
0
(`)[m] = dim
T
Z

/m
(J
0
(`)[/],m) = 2.
The proof of theorem 4.26 in case (b) requires more of the theory of group
schemes and Neron models than we wish to delve into here. We shall therefore
only explain the proof in the case (a) (which is [R5] thm. 5.2(b), see Ribets
paper for more details) and refer to [W3] sec. 2.1 for the general case. We
remark however that the proof of theorem 3.42 appeals to theorem 3.36 only
in the case = (cf. remarks 4.22 and remark 4.25). Recall that / divides `

only if is not good. So for the purposes of proving the Shimura-Taniyama

conjecture for semistable elliptic curves, (b) can be replaced by the stronger
hypothesis
(/
t
) /
2
` and
m
[
G

is not good.
The result in this case is due to Mazur and Ribet (the main result of [MRi])
and the proof is slightly easier than in the case of (b).
Returning to case (a), we appeal to a general property of the functor D of
theorem 2.31 which follows from cor. 5.11 of [Oda].
Theorem 4.27 Suppose that is an abelian variety over Z

(i.e., the Neron

model of an abelian variety over Q

with good reduction). There is a canonical

and functorial isomorphism of vector spaces over F

D((

)[/]

)
0

= Cot
0
(,F

).
where Cot
0
denotes the cotangent space at the origin.
134
In the case that is the Jacobian of a smooth proper curve A over Z

, we
have also a canonical isomorphism
Cot
0
(,F

)

= H
0
(A.
1
X/F

)
and in the case A = A
0
(`) with / `, this provides an isomorphism of
T
F

-modules
D(J
0
(`)(

)[/]

)
0

= o
2
(
0
(`). F

). (4.5.1)
Consider now the action of G
Q
on the points of
\ = J
0
(`)[/]

[m]

= (J
0
(`)[/],m)

.
By the argument in [Maz1] prop. II.14.2 or by the main result of [BLR], we
know that every Jordan-Holder constituent of the representation of G
Q
on this
T,m-vector space is isomorphic to

=
m
. It follows that
dim
T/m
D(\ ) = 2 dim
T/m
D(\ )
0
where we now regard \ as a good G

-module. On the other hand (4.5.1)

implies that
D(\ )
0

= o
2
(
0
(`). F

)[m].
which is one-dimensional over T,m by the -expansion principle (lemma 1.34
for example). We have now shown that D(\ ) is two-dimensional over T,m,
and it follows that so is \ .
Iharas lemma: We now sketch the proof of lemma 4.24, which proceeds
by analyzing the behavior of the homology of modular curves under certain
degeneracy maps. Note that it suces to prove the lemma for
T
.
If ` and ` are positive integers, then we let
1
(`. `) =
1
(`)
0
(`)
and write 1
1
(`. `) (resp. A
1
(`. `)) for the associated non-compactied
(resp. compactied) modular curve. The key intermediate result is the fol-
lowing:
Lemma 4.28 Suppose that ` is a positive integer and j is a prime not di-
viding `.
(a) The map
H
1
(A
1
(`. j). Z) H
1
(A
1
(`). Z)
2
r (

r.

r)
is surjective, where is dened by and by j.
135
(b) If ` 3 and / is an odd prime dierent from j, then the sequence
H
1
(1
1
(`j. j
2
). Z

) H
1
(1
1
(`j). Z

)
2
H
1
(1
1
(`). Z

)
r (
1
r.
1
r); (. .)
2

2
.
is exact, where the maps
1
and
2
are dened by and
1
and
2
by j.
To prove the lemma, one rst translates the statement into one about the
homology (or cohomology) of the congruence subgroups involved. Part (a)
of the lemma is due to Ihara [Ih], but see also the proof of thm. 4.1 of [R4].
Part (b) due to Wiles is more elementary and is established in the course of
proving lemma 2.5 of [W3] (see the sequence (2.13)). The result stated there
is in terms of cohomology rather than homology, but the one given here is
immediate from it.
Remark 4.29 A method of Khare [Kh], sec. 2 using modular symbols yields
an alternate proof of part (b) (after localization at a non-Eisenstein maximal
ideal).
To deduce lemma 4.24 from lemma 4.28 we also need the following result:
Lemma 4.30 Suppose that O is the ring of integers of a nite extension of Q

.
Let ` be a positive integer and H a subgroup of (Z,`Z)

. Let

T
C
=

T
Z
O
where

T
Z
is the subring of endomorphisms of `
2
(
1
(`)) generated by the
operators d) and 1
r
for primes : `. If n is a non-Eisenstein maximal ideal
then the localization at n of the natural map
(a) H
1
(1
1
(`). O) H
1
(A
1
(`). O) is an isomorphism;
(b) H
1
(A
1
(`). O) H
1
(A
H
(`). O) is surjective.
To prove part (a), one checks that
H
1
(1
1
(`). Z) H
1
(A
1
(`). Z)
is surjective and that 1
r
= : + 1 on the kernel for primes : 1 mod ` (cf.
proposition 4.13, part (c)).
Part (b) is most easily proved by showing that G
Q
acts trivially on the
cokernel of the natural map
T

(J
1
(`)) T

(J
H
(`))
(using [LO], prop. 6 for example) and then applying lemma 4.12.
136
Finally we explain how to use lemmas 4.28 and 4.30 to prove lemma 4.24.
It suces to consider the case
t
= j. We discuss only the more dicult
case of j ,= /. We dene ` = `

, `
t
= `j
2
, m and m
t
as in the preceding
section. Thus m
t
is a maximal ideal of T
t
C
where T
t
Z
acts on o
2
(
0
(`
t
)). All the
maps we consider in the argument below will respect the action of the Hecke
operators in

T
tt
Z
where

T
tt
Z
is the subring of endomorphisms of `
2
(
1
(`j. j
2
))
generated by the operators d) and 1
r
for primes : not dividing `j. We let n
denote the preimage of m
t
in

T
tt
C
=

T
tt
Z
O. We rst apply lemma 4.28(b) to
show that
H
1
(1
1
(`j. j
2
). O)
n
H
1
(1
1
(`j). O)
2
n
H
1
(1
1
(`). O)
n
is exact. Next we apply lemma 4.30(a) to obtain the exactness of
H
1
(A
1
(`j. j
2
). O)
n
H
1
(A
1
(`j). O)
2
n
H
1
(A
1
(`). O)
n
.
Applying lemma 4.30(b) with ` = `j, we nd that
H
1
(A
1
(`. j
2
). O)
n
H
1
(A
1
(`. j). O)
2
n
H
1
(A
1
(`). O)
n
is exact. We then apply lemma 4.28(a) to conclude that the map
H
1
(A
1
(`. j
2
). O)
n
H
1
(A
1
(`). O)
3
n
r (

r.

r.

r)
is surjective where , and are dened respectively by , j and
j
2
. Finally, we use lemma 4.30(b) again (now with ` = `) to get the
surjectivity of
H
1
(A
0
(`j
2
). O)
n
H
1
(A
0
(`). O)
3
n
.
hence that of (4.4.4) upon localizing at m
t
.
5 Commutative algebra
In this section we collect some basic facts of commutative algebra that are
used in the proof. We recall that O is the ring of integers in a nite extension
1 of Q

, and that O has residue eld /. Let (

C
denote as in section 2.6 the
category of complete noetherian local O-algebras with residue eld /.
137
5.1 Wiles numerical criterion
In this section we state a numerical criterion discovered by Wiles for a map
between two rings in (
C
to be an isomorphism.
Complete intersections: We say that a ring in (
C
is nite at if it is
nitely generated and torsion free as an O-module. A key ingredient in Wiles
isomorphism criterion is played by the concept of complete intersection, for
which we give the following naive denition.
Denition 5.1 An object in (
C
which is nite at is called a complete
intersection if it can be expressed as a quotient
O[[A
1
. . . . . A
n
]],()
1
. . . . . )
n
).
where there are as many relations as there are variables.
Remark 5.2 It is also true that if an object O[[1
1
. . . . . 1
r
]],J in (
C
which is
nite at is a complete intersection, then necessarily J can be generated by
: elements. See for example [Mat], thm. 21.2 (and lemma 5.11 below). We
will not use this fact in this chapter, although a special case is proved in the
course of establishing lemma 5.30.
The category (

C
: The numerical criterion of Wiles is stated more naturally
in terms of rings in the category C
C
which are endowed with some extra
structure: namely, a surjective O-algebra homomorphism
A
: O. Let
(

C
be the category whose objects are pairs (.
A
), where is an object of
(
C
and
A
: O is a surjective O-algebra homomorphism, also called the
augmentation map attached to . Morphisms in (

C
are local ring homomor-
phisms which are compatible in the obvious way with the augmentation maps.
By abuse of notation one often omits mentioning the augmentation map
A
when talking of objects in (

C
, and simply uses to denote (.
A
), when this
causes no confusion. Objects of (

C
will also be called augmented rings.
The invariants
A
and
A
: One associates to an augmented ring (.
A
)
two basic invariants:

A
= (ker
A
),(ker
A
)
2
;

A
=
A
(Ann
A
ker
A
).
Here Ann
A
(1) denotes the annihilator ideal of the ideal 1 in .
The invariant
A
can be thought of as a tangent space for the object .
(More precisely, it is the cotangent space of the scheme spec() at the point
ker
A
.) It is a nitely generated O-module.
138
The invariant
A
seems less familiar at rst sight. It is called the congruence
ideal. (The reason for this terminology should become clearer shortly.) We
are now ready to state Wiles numerical criterion:
Theorem 5.3 Let : 1 1 be a surjective morphism of augmented rings.
Assume that 1 is nitely generated and torsion-free as an O-module, and that

T
,= (0). (And hence, in particular, #(O,
T
) < .) Then the following are
equivalent:
(a) The inequality #
R
#(O,
T
) is satised.
(b) The equality #
R
= #(O,
T
) is satised.
(c) The rings 1 and 1 are complete intersections, and the map : 1 1
is an isomorphism.
Remark 5.4 The above theorem is slightly dierent from the one that ap-
pears in [W3], where it is assumed from the outset that the ring 1 is Goren-
stein. In the form in which we state it above, the theorem is due to H. Lenstra
[Len]. Our presentation follows Lenstras very closely. For the original (and
slightly dierent) point of view, the reader should consult the appendix of
[W3].
Some examples: Before going further, it may be good to pause and consider
some examples of objects of (

C
and the invariants associated to them. While
logically independent of the proof, the examples should help the reader develop
some intuition. (For a systematic way to compute the tangent spaces
A
, see
the paragraph at the end of section 5.2.)
Example 1:
= (c. /) O O. c / (mod
n
)
O[[1]],(1(1
n
)). with
A
(c. /) := c.

A
O,
n
O.
A
= (
n
).
Example 2:
= (c. /. c) O O O. c / c (mod )
O[[A. 1 ]],(A(A ). 1 (1 ). A1 ). with
A
(c. /. c) := c.

A
O,O O,O.
A
= ().
139
Example 3:
= O[[A]],(A
2
). with
A
()) := )(0).

A
O.
A
= 0.
Example 4:
=
_
(c. /. c. d) O O.
c / c d (mod ).
c +d / +c (mod
2
)
_
O[[A. 1 ]],(A(A ). 1 (1 )). with
A
(c. /. c. d) := c.

A
O,O O,O.
A
= (
2
).
Example 5:
= O[[A
1
. . . . . A
n
]]. with
A
()) := )(0).

A
O
n
.
A
= (0).
Example 6: ( = / 1 (mod 4). O = Z

).
= (c. / +ci) Z

[i]. c / (mod /
2
). c 0 (mod /)
Z

[[A]],(A(A
2
+/
2
)). with
A
(c. / +ci) := c.

A
Z,/
2
Z.
A
= (/
2
).
Example 7:
= O[[1]],(1) O /1 /1
2
. with
A
()) = )(0).

A
/.
A
= ().
5.2 Basic properties of
A
and
A
In this section we collect some of the basic properties of the invariants
A
and

A
, and prove the equivalence of (c) and (/) in theorem 5.3.
Behaviour of
A
under morphisms: The assignment
A
is a functor
from the category (

C
to the category of O-modules; a morphism 1 in
(

C
induces a homomorphism
A

B
of O-modules. Moreover, if 1
is surjective, then so is the induced map on the tangent spaces. Therefore,
when maps surjectively onto 1 we have
#
A
#
B
. (5.2.1)
There is also a converse to this, which will be useful later:
140
Lemma 5.5 If the homomorphism
A

B
is surjective, then 1 is
also surjective.
Proof: This follows from Nakayamas lemma. (Cf. [Ha], ch. II, sec. 7.4 and
[Mat], th. 8.4.) 2
Behaviour of
A
under (surjective) morphisms: Unlike the assignment

A
, the assignment
A
is not functorial, but it does have a nice
behaviour under surjective morphisms: namely, if : 1 is surjective,
then

A

B
. i.e., #(O,
A
) #(O,
B
). (5.2.2)
This is simply because in that case induces a map
Ann
A
ker
A
Ann
B
ker
B
.
Relation between the invariants
A
and
A
: In general, we have the
following inequality:
#
A
#(O,
A
). (5.2.3)
The key behind proving this identity is to interpret #
A
in terms of Fitting
ideals.
Digression on Fitting ideals: If 1 is a ring (in (
C
, say) and ` is a nitely
generated 1-module, we express ` as a quotient of 1
n
for some ::
0 `
t
1
n
` 0. (5.2.4)
The Fitting ideal of `, denoted Fitt
R
(`), is the ideal of 1 generated by the
determinants det(
1
. . . . .
n
), where the vectors
i
1
n
range over all possible
choices of elements of `
t
1
n
. One checks that this ideal does not depend on
the choice of exact sequence (5.2.4), and hence is an invariant of the 1-module
`. For example, if ` is a nitely generated O-module, we may write
` = O
r
O,(
n
1
) O,(
n
2
) O,(
n
k
).
with :
1
:
2
:
k
, and the sequence :
i
is completely determined by
`. The Fitting ideal Fitt
C
(`) is then the ideal of O generated by
n
1
++n
k
,
if : = 0, and is the 0-ideal if : 0. Note in particular that, if ` is a nite
O-module, then
#` = #(O,Fitt
C
(`)). (5.2.5)
141
Furthermore, if ` is any 1-module, it follows directly from the denition that
Fitt
R
(`) Ann
R
(`). (5.2.6)
Fitting ideals behave well under tensor products: in particular, if ` is a
nitely generated -module, where is an object in (

C
, then:

A
(Fitt
A
(`)) = Fitt
C
(`
A
O). (5.2.7)
where the tensor product is taken with respect to the augmentation map
A
.
For more details and references on the Fitting ideal, see [Len] for example.
Now we are ready to prove equation (5.2.3). For, noting that
A
=
ker
A

A
O, where the tensor product is taken with respect to
A
, and applying
equation (5.2.7) with ` = ker
A
, we have:
Fitt
C
(
A
) =
A
(Fitt
A
(ker
A
))
A
(Ann
A
ker
A
) =
A
.
(5.2.8)
where the containment follows from equation (5.2.6). Now the inequality
(5.2.3) follows by combining (5.2.5) and (5.2.8).
As a consequence, we have
Corollary 5.6 The statements (c) and (/) in theorem 5.3 are equivalent.
Proof: If 1 1 is a surjective map of augmented rings, then #
R

#
T
, by equation (5.2.1). But equation (5.2.3) gives the inequality #
T

#(O,
T
). Hence the inequality #
R
#(O,
T
) always holds, so that (c)
implies (/) in theorem 5.3. The reverse implication is clear. 2
Remark 5.7 (Computing the tangent spaces
A
): Any object (.
A
) in (

C
can be expressed as a quotient of the object l = O[[A
1
. . . . . A
n
]] of ex-
ample 5 with augmentation map given by
U
()) = )(0). Indeed, one can
take c
1
. . . . . c
n
to be -module generators of the nitely generated -module
ker
A
, and obtain the desired quotient map by sending A
i
to c
i
.
The tangent space
U
of l is a free O-module of rank : which can be
written down canonically as
OA
1
OA
2
OA
n
.
the natural map from (ker
U
) being simply the map which sends a power
series ) l with no constant term to its degree 1 term, which we will denote
by

).
If is expressed as a quotient l,()
1
. . . . . )
r
), then one has

A
=
U
,(

)
1
. . . . .

)
r
). (5.2.9)
142
5.3 Complete intersections and the Gorenstein condi-
tion
In this section we show that complete intersections satisfy a Gorenstein con-
dition, and that (c) implies (c) and (/) in the statement of Wiles isomorphism
criterion (theorem 5.3).
Denition 5.8 An object in (
C
which is nite at is said to be Gorenstein
if
Hom
C
(. O) as -modules.
Proposition 5.9 Suppose in (
C
is nite at. If is a complete intersec-
tion, then is Gorenstein.
The remainder of this section will be devoted to proving proposition 5.9. Since
the proof is a bit long and involved, and the concepts it uses are not used
elsewhere, the reader is advised on a rst reading to take it on faith and skip
to the next section. A more direct proof of proposition 5.9 which bypasses the
arguments of this section is explained in [Len].
We let be a ring which is nite at, and is a complete intersection.
(Hence, can be written as O[[A
1
. . . . . A
n
]],()
1
. . . . . )
n
).) We assume that
the augmentation map for is induced from the map on O[[A
1
. . . . . A
n
]]
sending ) to )(0). This implies that )
i
(0) = 0, i.e., the )
i
have no constant
term.
We recall some denitions from commutative algebra that we will need.
An ideal 1 of a local ring 1 is said to be primary if 1 ,= 1 and every zero
divisor in 1,1 is nilpotent. If (r
1
. . . . . r
n
) generates a primary ideal of 1, and
: = dim1, then (r
1
. . . . . r
n
) is called a system of parameters for 1.
Lemma 5.10 The sequence ()
1
. . . . . )
n
. ) is a system of parameters for l =
O[[A
1
. . . . . A
n
]].
Proof: The quotient ring l,()
1
. . . . . )
n
. ) is local and is nitely generated as
a /-vector space; therefore every element in its maximal ideal is nilpotent. 2
A sequence (r
1
. . . . . r
n
) in a ring 1 is said to be a regular sequence if r
i
is
not a zero-divisor in 1,(r
1
. . . . . r
i1
) for i = 1. . . . . :.
Lemma 5.11 The sequence ()
1
. . . . . )
n
) is a regular sequence for l.
143
Proof: The ring l is Cohen Macaulay, since . A
1
. . . . . A
n
is a system of
parameters of l which is also a regular sequence. Hence, by theorem 17.4 (iii)
of [Mat], ()
1
. . . . . )
n
. ) is a regular sequence in l. A fortiori, the sequence
()
1
. . . . . )
n
) is also a regular sequence. 2
The proofs of lemma 5.10 and 5.11 use only the fact that is nitely
generated as an O-module, and not that is at. As a corollary of this proof,
we therefore have:
Corollary 5.12 If 1 is an object of (

C
which is nitely generated as an O-
module, and 1 O[[A
1
. . . . . A
n
]],()
1
. . . . . )
n
), then 1 is also at, and hence
is a complete intersection.
To go further, we will introduce the Koszul complex
1(r. 1) :=
n
p=0
1
p
(r. 1)
associated to a local ring 1 and a sequence r = (r
1
. . . . . r
n
) of elements in
its maximal ideal. For more details on the Koszul complex and its relation
to regular sequences, the reader can consult [Mat], especially 16, or [Bour]
X, or [Se3]. This complex is dened to be the free graded dierential algebra
generated by symbols n
1
. . . . . n
n
:
1
p
(r. 1) :=
i
1
<i
2
<<i
p
1 n
i
1
n
i
p
.
with dierential d : 1
p
1
p1
dened by
d(n
i
1
n
i
p
) =
p

t=1
(1)
t
r
t
n
i
1
n
i
t1
n
i
t+1
n
i
p
.
We denote by H
p
(r; 1) the homology groups of this complex. We record here
the main properties of this complex that we will use.
Proposition 5.13 (a) H
0
(r; 1) = 1,(r).
(b) There is a long exact homology sequence
H
p
(r; 1) H
p
(r. r
n+1
; 1) H
p1
(r; 1)
x
n+1

H
p1
(r; 1) H
p1
(r. r
n+1
; 1) H
p2
(r; 1)
(c) H
p
(r; 1) is annihilated by the ideal (r), i.e., it has a natural 1,(r)-
module structure.
144
(d) If r is a regular sequence, then H
p
(r; 1) = 0 for all j 0 (i.e., the
complex 1
p
(r; 1) is a free resolution of 1,(r).)
Proof: The rst assertion follows directly from the denition. For (/) and (c),
see [Mat], th. 16.4. The assertion (d) can be proved by a direct induction
argument on :, using the long exact homology sequence: For j 1, this
sequence becomes
0 H
p
(r. r
n+1
; 1) 0.
and for j = 1, it is
0 H
1
(r. r
n+1
; 1) H
0
(r; 1)
x
n+1
H
0
(r; 1).
But the assumption that r. r
n+1
is a regular sequence means that multiplica-
tion by r
n+1
is injective on H
0
(r; 1) = 1,(r). Hence, the assertion (d) follows.
2
Now, we turn to the proof of proposition 5.9, following a method of Tate
which is explained in the appendix of [MRo]. For any ring 1, write 1[[A]] :=
1[[A
1
. . . . . A
n
]]. Let c
1
. . . . . c
n
be the images in of A
1
. . . . . A
n
by the
natural map
: O[[A]] = O[[A]],()
1
. . . . . )
n
).
and let
: [[A]]
be the natural map which sends A
i
to c
i
. The sequence (p
i
) = (A
i
c
i
)
generates the kernel of . Since the )
i
, viewed as polynomials in [[A]], also
belong to ker , we have:
()
1
. . . . . )
n
) = (p
1
. . . . . p
n
)`. (5.3.1)
where ` is an : : matrix with coecients in [[A]]. Let
1 = det(`) [[A]].
Our goal is to construct an -module isomorphism
Hom
C
(. O) .
We begin by constructing a surjective (O-linear) map
Hom
C[[X]]
([[A]]. O[[A]]) .
145
Lemma 5.14 (Tate): The function ()) = ()(1)) induces an isomorphism
of O[[A]]-modules
Hom
C[[X]]
([[A]]. O[[A]]),(p
1
. . . . . p
n
) .
Proof: By lemma 5.11, the sequence ()) = ()
1
. . . . . )
n
) is a regular O[[A]]-
sequence. One can see directly from the denition that the sequence (p) =
(p
i
) = (A
i
c
i
) is a regular [[A]]-sequence. Let 1()) and 1(p) be the Koszul
complexes associated to these two sequences. It follows from proposition 5.13
that the Koszul complex 1()) is a resolution of by free O[[A]]-modules,
and the Koszul complex 1(p) is a resolution of by free [[A]]-modules, and
hence a fortiori, by free O[[A]]-modules. We dene a map : 1()) 1(p)
of complexes by letting

0
: 1
0
()) 1
0
(p)
be the natural inclusion of O[[A]] into [[A]], and letting

1
: 1
1
()) 1
1
(p)
be the map dened by
(
1
(n
1
). . . . .
1
(n
n
)) = (
1
. . . . .
n
)`.
and extending it by skew-linearity to a map of exterior algebras. One can
check that the resulting map is a morphism of complexes which induces the
identity map , and satises

n
(n
1
n
n
) = 1
1

n
.
Applying the functor Hom
C[[X]]
(. O[[A]]) to these two free resolutions, and
taking the homology of the resulting complexes, we nd that since is a homo-
topy equivalence of complexes it induces an isomorphism on the cohomology,
and in particular, on the :th cohomology:

n
: Hom
C[[X]]
([[A]]. O[[A]]),(p
1
. . . . . p
n
)

Hom
C[[X]]
(O[[A]]. O[[A]]),()
1
. . . . . )
n
) .
which is given explicitly by the formula:

n
()) = ()(1)).
2
We nally come to the proof of proposition 5.9, which we can state in a
more precise form.
146
Lemma 5.15 The map : Hom
C
(. O) dened by ()) = (

)(1)),
where

) : [[A]] O[[A]] is the base change of ), is an -module isomor-
phism, and hence, is Gorenstein.
Proof: The key point is to show that is -linear. By denition, if
c = (c
t
) . with c
t
O[[A]].
then
(c)) = (

)(c1)) = (

)((c c
t
)1)) +(

)(c
t
1)).
Since c c
t
ker , it can be written as an [[A]]-linear combination of the
p
i
. By multiplying the relation
()
1
. . . . . )
n
) = (p
1
. . . . . p
n
)`
by the matrix 1 `
1
`
n
([[A]]), one sees that the 1p
i
can be written as
[[A]]-linear combinations of the )
i
s. Hence, so can the expression (cc
t
)1.
By the O[[A]]-linearity of

), and the fact that each )
i
belongs to ker , it
follows that
(

)((c c
t
)1)) = 0.
Therefore,
(c)) = (c
t

)(1)) = c()).
This shows that is -linear. To show that is surjective, observe that if
)
1
. . . . . )
r
is a O-basis of Hom
C
(. O), then

)
1
. . . . .

)
r
is a O[[A]]-basis for
Hom
C[[X]]
([[A]]. O[[A]]). Hence, for all c , there exist j
1
. . . . . j
r
such
that

n
(j
1

)
1
+ +j
r

)
r
) = c.
But this means that
((j
1
))
1
+ +(j
r
))
r
) = c.
so that is surjective. Finally, since Hom
C
(. O) and are free O-modules
of the same nite rank, and is surjective, it must also be injective. Hence
is an isomorphism, as was to be shown. 2
Example 5.16 Let OO be the ring of example 1 in section 5.1. Then,
) = 1
2

n
1, and p = 1 (0.
n
). Hence, ) = (1 (
n
. 0))p, so that
1 = 1 (
n
. 0). It follows that
(/) = (

/(1 (
n
. 0))) = (1/(1. 1) /(
n
. 0)) = (0.
n
)/(1. 1) /(
n
. 0).
The reader can check directly that is indeed an -linear isomorphism from
the -module Hom
C
(. O) to .
147
Example 5.17 Let = O[c],(c
2
) be the ring of example 3 in section 5.1.
Then, ) = 1
2
, and p = 1 c. Hence, ) = (1 + c)p, so that 1 = 1 + c. It
follows that
()) = (

)(1 +c)) = (1)(1) +)(c)) = c)(1) +)(c).
Exercise 5.18 Write down explicitly an isomorphism

Hom
C
(. O) for
the complete intersection = O[[A. 1 ]],(A(A ). 1 (1 )) of example 4
of section 5.1.
5.4 The Congruence ideal for complete intersections
Let be an object of (

C
, which is nite at and is a complete intersection as
in the previous section, so that O[[A
1
. . . . . A
n
]],()
1
. . . . . )
n
).
Using the result of the previous section, we will give an explicit formula
for computing
A
in this case, and prove that (c) implies (/) in theorem 5.3.
Let

:= Hom
C
(. O), and let

A
: O

be the dual map. From

the Gorenstein property of , we may identify

with (as -modules). Fix

any identication :

. (Any two such dier by a unit in .) One
checks that

A
(O

) = Ann
A
ker
A
.
Hence,
A
is the image of the map
A

A
. By using the explicit construction
of given in lemma 5.15, we nd:

A
(O

) = (1).
where 1 is the determinant dened in section 5.3. By a direct calculation using
equation (5.3.1), one sees that the right hand side is equal to det()
i
,A
j
(0)).
Hence we have shown:
Proposition 5.19 Suppose that is an object of (

C
which is nite at. If
= O[[A
1
. . . . . A
n
]],()
1
. . . . . )
n
) is a complete intersection, then

A
= (det()
i
,A
j
(0))).
This proposition implies:
Corollary 5.20 Let in (

C
be nite at. If is a complete intersection,
then #
A
= #(O,
A
). Hence, the statement (c) implies statement (/) in
theorem 5.3.
Proof: The equation (5.2.9) of remark 5.7 implies that
#
A
= #O,(det()
i
,A
j
(0))).
The result follows. 2
148
5.5 Isomorphism theorems
The usefulness of the notion of complete intersections comes from the following
two (vaguely stated) principles:
1. Isomorphisms to complete intersections can often be recognized by looking
at their eects on the tangent spaces.
2. Isomorphisms from complete intersections can often be recognized by look-
ing at their eects on the invariants .
These vague principles are made precise in theorems 5.21 and 5.24 respec-
tively.
Theorem 5.21 Let : 1 be a surjective morphism of augmented rings,
with 1 a (nite, at) complete intersection. If induces an isomorphism from

A
to
B
, and these modules are nite, then is an isomorphism.
Remark 5.22 Let = O[[A. 1 ]],(A(A ). 1 (1 )) be the ring of ex-
ample 4, let 1 = O[[A. 1 ]],(A(A). 1 (1 ). A1 ) be the ring of example
2, and let : 1 be the natural projection. The map induces an
isomorphism
A

B
, even though is not an isomorphism. The assump-
tion that 1 is a complete intersection is crucial for concluding that is an
isomorphism.
Remark 5.23 The natural map
O[[A]],(A
3
) O[[A]],(A
2
)
is a surjective morphism inducing an isomorphism on tangent spaces, and the
target ring is a complete intersection. Yet this map is not an isomorphism.
This shows that the assumption on the niteness of the tangent spaces cannot
be dispensed with.
Proof of theorem 5.21: Recall that l = O[[A
1
. . . . . A
n
]] is the augmented ring
of example 5 of section 5.1. Let

B
: l 1
be a surjective morphism of augmented rings with ker
B
= ()
1
. . . . . )
n
). Let
/
1
. . . . . /
n
ker
B
denote the images of A
1
. . . . . A
n
by
B
, and let c
1
. . . . . c
n

ker
A
denote inverse images of /
1
. . . . . /
n
by . Since is an isomorphism
on tangent spaces, the elements c
i
generate (ker
A
),(ker
A
)
2
. Hence the
morphism

A
: O[[A
1
. . . . . A
n
]]
149
dened by
A
(A
i
) = c
i
induces a surjection
U

A
, and so it is surjective,
by lemma 5.5.
We claim that ker
B
is contained in ker
A
(and hence, ker
B
= ker
A
).
For, let p
1
. . . . . p
n
be elements of ker
A
whose linear terms p
1
. . . . . p
n
generate
the kernel of

A
:
U

A
.
Since ker
A
ker
B
, it follows that there exists an :: matrix ` `
n
(l)
with entries in l such that
(p
1
. . . . . p
n
) = ()
1
. . . . . )
n
)`.
Let

` be the matrix of constant terms of the matrix `. Then we have
( p
1
. . . . . p
n
) = (

)
1
. . . . .

)
n
)

`.
Since ( p
1
. . . . . p
n
) and (

)
1
. . . . .

)
n
) generate the same submodules of rank :
and nite index in
U
, it follows that det

` is a unit in O. Hence, ` is
invertible, and therefore the )
i
can be expressed as a l-linear combination of
the p
j
. This implies that ker
B
ker
A
. Now we see that
A

1
B
gives a
well-dened inverse to , so that is an isomorphism. 2
Theorem 5.24 Let : 1 be a surjective morphism of augmented rings.
Suppose that and 1 are nite at, and that a complete intersection. If

A
=
B
,= 0, then is an isomorphism.
Remark 5.25 The torsion-freeness assumption on 1 is essential: if : is large
enough, then 1 = ,(ker
A
)
n
satises
A
=
B
, although the natural map
1 is not injective when ,= O.
Proof: By proposition 5.9, is Gorenstein, i.e., it satises

:= Hom
C
(. O) as -modules.
Now, we observe that
ker
A
Ann
A
ker
A
= 0. (5.5.1)
and likewise for 1. For, let r be a non-zero element of
A
, and let r
t

Ann
A
ker
A
satisfy
A
(r
t
) = r. For all c ker
A
Ann
A
ker
A
, we have
0 = c(r r
t
) = cr.
150
the rst equality because c belongs to Ann
A
ker
A
and (r r
t
) belongs to
ker
A
, the second equality because c belongs to ker
A
and r
t
belongs to
Ann
A
ker
A
. Hence c belongs to the O-torsion submodule of , and therefore
is 0.
It follows from (5.5.1) that the homomorphism
A
(resp.
B
) induces an
isomorphism from Ann
A
ker
A
(resp. Ann
B
ker
B
) to
A
(resp.
B
). Since
A
is isomorphic to
B
, it follows that induces an isomorphism from Ann
A
ker
A
to Ann
B
ker
B
, i.e.,
Ann
A
ker
A
= Ann
B
ker
B
.
From (5.5.1) it also follows a fortiori that
ker Ann
A
ker
A
= 0.
hence there is an exact sequence of -modules:
0 ker Ann
A
ker
A
. (5.5.2)
The cokernel of the last map is
,(ker Ann
A
ker
A
) 1,(Ann
A
ker
A
) 1,(Ann
B
ker
B
).
which is torsion-free, since there is a natural injection
1,(Ann
B
ker
B
) End
C
(ker
B
).
Hence, the exact sequence (5.5.2) splits over O. Taking O duals in (5.5.2)
and using the Gorenstein condition for , we thus get an exact sequence of
-modules:
(ker )

(Ann
A
ker
A
)

0.
Applying the functor
A
/ (relative to the map /), we nd
1 = dim
k
(
A
/) dim
k
((ker )

A
/) + dim
k
((Ann
A
ker
A
)

A
/).
Since
A
,= 0, it follows that (Ann
A
ker
A
)

A
/ ,= 0, and hence we must have
(ker )

A
/ = 0.
Therefore by Nakayamas lemma and duality, ker = 0, which proves the
theorem. 2
151
5.6 A resolution lemma
It turns out that objects in (

C
can be resolved (in a weak sense) by a
complete intersection, namely,
Theorem 5.26 Let be an augmented ring which is nite at over O. Then
there is a morphism

in the category (

C
such that:
(a) the ring

is nite at over O and is a complete intersection;
(b) the map

induces an isomorphism

A

A
.
Proof: Write as a quotient of l = O[[A
1
. . . . . A
n
]] (with
U
: l O the
map which sends each A
i
to 0.) Let )
1
. . . . . )
n
be elements in the kernel of the
natural map l , such that

)
1
. . . . .

)
n
generate the kernel of
U

A
.
Now letting

= l,()
1
. . . . . )
n
)
would give the desired ring

, provided

is nitely generated: indeed, the
atness of

would follow from corollary 5.12.
Thus we need to show that the )
i
can be chosen so that

is nitely
generated. Let c
1
. . . . . c
n
be O-module generators of the nitely generated
module ker
A
, and dene a homomorphism from the polynomial ring
\ = O[A
1
. . . . . A
n
]
to by sending A
j
to c
j
. Clearly is surjective. Let )
1
, . . . , )
n
be elements
of ker chosen as above, and let : denote their maximal degree. Since the
elements c
2
i
belong to ker
A
, we may write
c
2
i
= /
i
(c
1
. . . . . c
n
).
where /
i
(A
1
. . . . . A
n
) is a linear polynomial. Now, replacing the relations )
i
by the relations
)
i
+A
m
i
/
i
A
m+2
i
.
we nd that the ring \,()
1
. . . . . )
n
) is a nitely generated O-module: it can
be generated by the images of the monomials of degree :(:+ 1), since the
relations allow us to rewrite any monomial of higher degree in terms of ones of
lower degree. Completing at the ideal (. A
1
. . . . . A
n
), we nd that the ring

= l,()
1
. . . . . )
n
)
has the desired properties: the natural homomorphism from

to induces
an isomorphism on the tangent spaces, since the linear terms of the )
i
generate
the kernel of the induced map
U

A
on the tangent spaces, and

is a
nitely generated O-module, since \,()
1
. . . . . )
n
) is. 2
152
5.7 A criterion for complete intersections
The results we have accumulated so far allow us to give an important criterion
for an object to be a complete intersection:
Theorem 5.27 Let be an augmented ring which is a nitely generated
torsion-free O-module. If #
A
#(O,
A
) < , then is a complete inter-
section.
Proof: Let :

be the surjective morphism given by the resolution
theorem (theorem 5.26). Then we have
#(O,
A
) (#
A
) = (#

A
) #(O,

A
).
where the rst inequality is by assumption, the second by the choice of

, and
the third is by the equation (5.2.3). On the other hand, by equation (5.2.2),
we have
#(O,

A
) #(O,
A
).
It follows that

A
=

A
.
so that is an isomorphism by theorem 5.24. It follows that is a complete
intersection. 2
5.8 Proof of Wiles numerical criterion
Theorem 5.28 Let 1 and 1 be augmented rings such that 1 is a nitely
generated torsion-free O-module, and let : 1 1 be a surjective morphism.
If
#
R
#(O,
T
) < .
then 1 and 1 are complete intersections, and is an isomorphism.
Proof: We have:
#(O,
T
) #
T
#
R
#(O,
T
).
where the rst inequality is by equation (5.2.3), the second follows from the
surjectivity of , and the third is by hypothesis. Therefore,
#
T
= #(O,
T
).
and hence 1 is a complete intersection by theorem 5.27. Since the orders of

R
and
T
are the same, induces an isomorphism between them. Hence
is an isomorphism 1 1, by theorem 5.21. This completes the proof. 2
153
Theorem 5.28 shows that the statements (c) and (/) in theorem 5.3 imply
the statement (c). Combining corollaries 5.6, 5.20, and theorem 5.28 completes
the proof of theorem 5.3.
5.9 A reduction to characteristic /
Let (
k
be the category of complete local noetherian /-algebras with residue
eld /. Again all morphisms are assumed to be local. There is a natural
functor

from (
C
to (
k
which send to

:= ,.
We say that an object of (
k
which is nite dimensional as a /-vector
space is a complete intersection if it is isomorphic to a quotient
= /[[A
1
. . . . . A
r
]],()
1
. . . . . )
r
).
Note that if an object of (
C
is a complete intersection, then

is a
complete intersection in (
k
. As a partial converse, we have:
Lemma 5.29 Suppose that 1 1 is a map in the category (
C
, and that 1 is
nitely generated and free as an O-module. Then 1 1 is an isomorphism
of complete intersections, if and only if

1

1 is.
Proof: This is an exercise and is left to the reader. 2
We now come to the proof of lemma 3.39 of section 3.4:
Lemma 5.30 Suppose that 1 1
t
are local elds with rings of integers
O O
t
and that is an object of (
C
which is nitely generated and free as
an O-module. Then is a complete intersection if and only if
C
O
t
is.
Proof: One implication is clear. Let / and /
t
be the residue elds of O and O
t
respectively. By lemma 5.29 it is enough to prove that, if 1 is an object of (
k
which is nite dimensional as a /-vector space, then
1
t
= 1
k
/
t
is a complete intersection 1 is a complete intersection.
Let m and m
t
denote the maximal ideals of 1 and 1
t
respectively.
By assumption, we have 1
t
= /
t
[[1
1
. . . . . 1
r
]],J, where the ideal J can be
generated by : elements. We can assume without loss of generality (by adding
extra variables and relations if necessary) that the images of 1
1
. . . . 1
r
generate
m
t
as a /
t
-vector space. Now, let
: /[[A
1
. . . . . A
r
]] 1
154
be a ring homomorphism, such that the images of A
1
. . . . . A
r
generate m as
a /-vector space, and let

t
: /
t
[[A
1
. . . . . A
r
]] 1
t
be the extension of scalars. Let 1 and 1
t
= 1
k
/
t
be the kernels of these
two maps. We claim that 1
t
can be generated by : elements. (In fact, this is
also true without the assumption that the images of A
1
. . . . . A
r
generate m
t
as a /
t
vector space, although we use this assumption in the proof below. Cf.
remark 5.2.) To see this, choose an isomorphism of /
t
-vector spaces /
t
A
1

/
t
A
r
/
t
1
1
/
t
1
r
such that
/
t
A
1
/
t
A
r
/
t
1
1
/
t
1
r

m
t
= m
t
commutes. Extending this map to an isomorphism of /
t
-algebras,
: /
t
[[A
1
. . . . . A
r
]] /
t
[[1
1
. . . . . 1
r
]].
one sees that 1
t
=
1
(J), and hence can be generated by : elements, as
claimed.
In particular we have dim
k
(1
t
,m
t
1
t
) :, and
dim
k
(1,m1) = dim
k
((1,m1)
k
/
t
) dim
k
(1
t
,m
t
1
t
) :.
Nakayamas lemma now implies that 1 can be generated by : elements, and
hence 1 is a complete intersection. This proves the lemma. 2
5.10 J-structures
We now turn to the proof of theorem 3.41. In view of the last section we will
work in characteristic /. Thus let : 1 1 be a surjective morphism in the
category (
k
, where 1 and 1 are nite dimensional as /-vector spaces. Let :
be a non-negative integer. If J /[[o
1
. .... o
r
]] and J (o
1
. .... o
r
) then by a
strong J-structure we shall mean a commutative diagram in (
k
/[[o
1
. .... o
r
]]

/[[A
1
. .... A
r
]] 1
t
1
t

1 1.
such that
155
(a) 1
t
,(o
1
. .... o
r
)1
t

1 and 1
t
,(o
1
. .... o
r
)1
t
1,
(b) for each ideal 1 J, 1 = ker (/[[o
1
. .... o
r
]] 1
t
,1),
(c) J = ker (/[[o
1
. .... o
r
]] 1
t
) and 1
t
1
t
1.
We refer to these as strong J-structures because we have slightly altered the
conditions from section 3.4 for technical convenience in the rest of this section.
We will prove the following result.
Theorem 5.31 Suppose there exist a sequence of ideals J
n
/[[o
1
. .... o
r
]] such
that J
0
= (o
1
. .... o
r
), J
n
J
n+1
,

n
J
n
= (0) and for each : there exists a
strong J
n
-structure. Then 1

1 and these rings are complete intersections.
Before proving theorem 5.31 we shall explain how to deduce theorem 3.41
from it. If J is an ideal of O[[o
1
. .... o
r
]] we will use

J to denote its image in
/[[o
1
. .... o
r
]]. If o given by
O[[o
1
. .... o
r
]]

O[[A
1
. .... A
r
]] 1
t
1
t

1 1.
is a J-structure for 1 1, let a = ker (1 1), and let 1
tt
denote the ring
Im(1
t
(1,m
R
a (1
t
,J))
C
/). Then

o given by
/[[o
1
. .... o
r
]]

/[[A
1
. .... A
r
]] 1
tt
(1
t
,J)
C
/

(1,m
R
a)
C
/ 1
C
/.
is a strong

J-structure for (1,m
R
a)
C
/ 1
C
/. If J
n
is a sequence of ideals
as in theorem 3.41 then

J
n
is a sequence of nested ideals with

J
0
= (o
1
. .... o
r
)
and

J
n
= (0). Moreover if for each : there is a J
n
-structure o
n
for 1 1
then for each : there is a strong

J
n
-structure

o
n
for
(1,m
R
a)
C
/ 1
C
/.
Then by theorem 5.31 we see that this map is an isomorphism of complete
intersections. Lemma 5.29 shows that theorem 3.41 follows. 2
Before returning to the proof of theorem 5.31, let us rst make some re-
marks about strong J-structures.
156
The set of strong J structures for all ideals J (o
1
. .... o
r
) /[[o
1
. .... o
r
]]
forms a category, with the obvious notion of morphism.
If o is a strong J-structure and if (o
1
. .... o
r
) J
t
J then there is a natural
J
t
-structure o mod J
t
obtained by replacing 1
t
by 1
t
,J
t
and 1
t
by the image
of 1
t
1 (1
t
,J
t
).
If 1 is a nite dimensional /-vector space and if J has nite index in
/[[o
1
. .... o
r
]] then there are only nitely many isomorphism classes of strong
J-structure. (This follows because we can bound the order of 1
t
in any J-
structure. Explicitly we must have #1
t
(#1)(#/[[o
1
. .... o
r
]],J)
dim
k
T
.)
Lemma 5.32 Suppose that 1 is a nite dimensional / vector space. Suppose
also that J
n
is a nested (decreasing) sequence of ideals and that J =

n
J
n
.
If for each : a strong J
n
structure exists then a strong J structure exists.
Proof: We may suppose that each J
n
has nite index in /[[o
1
. .... o
r
]]. Let o
n
denote a strong J
n
-structure. Let o
n,m
= o
n
mod J
m
if : :. Because there
are only nitely many isomorphism classes of strong J
m
structure, we may
recursively choose integers :(:) such that
o
n(m),m

= o
n,m
for innitely many :,
if : 1 then o
n(m),m1

= o
n(m1),m1
.
Let o
t
m
= o
n(m),m
. Then o
t
m
is a strong J
m
structure and if : :
1
then
o
t
m
mod J
m
1

= o
t
m
1
. One checks that o = lim

o
t
m
is the desired strong J-
structure. 2
Lemma 5.33 Suppose that a strong (0) structure exists. Then the map 1
1 is an isomorphism, and these rings are complete intersections.
Proof: Because /[[o
1
. .... o
r
]] 1
t
(and 1
t
is a nitely generated /[[o
1
. .... o
r
]]-
module by Nakayamas lemma, cf. [Mat], thm. 8.4) we see that the Krull
dimension of 1
t
is at least :. On the other hand /[[A
1
. .... A
r
]] 1
t
and so
by Krulls principal ideal theorem this map must be an isomorphism. Thus
/[[A
1
. .... A
r
]]

1
t

1
t
.
Hence we have that
/[[A
1
. .... A
r
]],(o
1
. .... o
r
)

1
t
,(o
1
. .... o
r
)

1.
and the lemma follows. 2
Theorem 5.31 follows at once from these two lemmas. 2
157
References
[Ant2] W. Kuyk, P. Deligne, eds., Modular Functions of One Variable II,
Lecture Notes in Math. 349, Springer-Verlag, New York, Berlin, Hei-
delberg, 1973.
[Ant3] W. Kuyk, J.-P. Serre, eds., Modular Functions of One Variable III,
Lecture Notes in Math. 350, Springer-Verlag, New York, Berlin, Hei-
delberg, 1973.
[Ant4] B. Birch, W. Kuyk, eds., Modular Functions of One Variable IV,
Lecture Notes in Math. 476, Springer-Verlag, New York, Berlin, Hei-
delberg, 1975.
[Ant5] J.-P. Serre, D. Zagier, eds., Modular Functions of One Variable V,
Lecture Notes in Math. 601, Springer-Verlag, New York, Berlin, Hei-
delberg, 1977.
[AL] A.O.L. Atkin and J. Lehner, Hecke operators on
0
(:), Math. Ann.
185 (1970), 134160.
[BCEM] J. Buhler, R. Crandall, R. Ernvall, T. Metsankyla, Irregular primes
and cyclotomic invariants to four million, Math. Comp. 61 (1993),
151153.
[BK] S. Bloch, K. Kato, L-functions and Tamagawa numbers of motives,
in: The Grothendieck Festschrift I, Progress in Math. 86, Birkhauser,
Boston, Basel, Berlin, 1990, pp. 333400.
[BLR] N. Boston, H. Lenstra, K. Ribet, Quotients of group rings arising
from two-dimensional representations, C. R. Acad. Sc. t312 (1991),
323328.
[BM] N. Boston, B. Mazur, Explicit universal deformations of Galois repre-
sentations, Algebraic Number Theory, 121, Adv. Stud. Pure Math.
17, Acad. Press, Boston MA, 1989.
[Bour] N. Bourbaki, Alg`ebre (

Elements de Mathematiques, Fasc. XXIII),

Hermann, Paris, 1958.
[Ca1] H. Carayol, Sur les representations j-adiques associees aux formes
modulaires de Hilbert, Ann. Sci. Ec. Norm. Super. 19 (1986), 409
468.
158
[Ca2] H. Carayol, Sur les representations galoisiennes modulo / attachees
aux formes modulaires, Duke Math. J. 59 (1989), 785801.
[Ca3] H. Carayol, Formes modulaires et representations galoisiennes `a
valeurs dans un anneau local complet, in: j-adic Monodromy and
the Birch-Swinnerton-Dyer Conjecture (B. Mazur and G. Stevens,
eds.), Contemporary Math. 165, 1994, pp. 213237.
[Co] J. Coates, S.T. Yau, eds., Elliptic Curves, Modular Forms and Fer-
mats Last Theorem, International Press, Cambridge, 1995.
[CPS] E. Cline, B. Parshall, L. Scott, Cohomology of nite groups of Lie
type I, Publ. Math. IHES 45 (1975), 169191.
[Cr] J. Cremona, Algorithms for Modular Elliptic Curves, Cambridge
Univ. Press, Cambridge, 1992.
[CR] C.W. Curtis, I. Reiner, Representation Theory of Finite Groups and
Associative Algebras, Wiley Interscience, New York, 1962.
[CS] G. Cornell and J.H. Silverman, eds., Arithmetic Geometry, Springer-
Verlag, New York, Berlin, Heidelberg, 1986.
[CW] J. Coates, A. Wiles, On the conjecture of Birch and Swinnerton-Dyer,
Inv. Math. 39 (1977), 223251.
[Da1] H. Darmon, Serres conjectures, in [Mu2], pp. 135153.
[Da2] H. Darmon, The Shimura-Taniyama Conjecture, (dapr`es Wiles), (in
Russian) Uspekhi Mat. Nauk 50 (1995), no. 3(303), 3382. (English
version to appear in Russian Math Surveys)
[De] P. Deligne, Formes modulaires et representations /-adiques, in: Lec-
ture Notes in Math. 179, Springer-Verlag, New York, Berlin, Heidel-
berg, 1971, pp. 139172.
[DR] P. Deligne, M. Rapoport, Les schemas de modules de courbes ellip-
tiques, in [Ant2], pp. 143316.
[dS] E. de Shalit, On certain Galois representations related to the modular
curve A
1
(j), to appear in Compositio Math.
[dSL] B. de Smit, H.W. Lenstra, Explicit construction of universal defor-
mation rings, to appear in the Proceedings of the Conference on
Fermats Last Theorem, Boston University, August 918, 1995.
159
[DS] P. Deligne, J.-P. Serre, Formes modulaires de poids 1, Ann. Sci. Ec.
Norm. Sup. 7 (1974), 507530.
[Di1] F. Diamond, The rened conjecture of Serre, in [Co], pp. 2237.
[Di2] F. Diamond, On deformation rings and Hecke rings, to appear in
Annals of Math.
[Dic1] L.E. Dickson, History of the Theory of Numbers, Vol. II, Chelsea
Publ. Co., New York, 1971.
[Dic2] L.E. Dickson, Linear Groups with an Exposition of the Galois Field
Theory, Teubner, Leipzig, 1901.
[Dir] P.G.L. Dirichlet, Memoire sur limpossibilite de quelques equations
indeterminees du cinqui`eme degre, Jour. f ur Math. (Crelle) 3 (1828),
354375.
[DI] F. Diamond and J. Im, Modular forms and modular curves, in [Mu2],
pp. 39133.
[DK] F. Diamond and K. Kramer, Modularity of a family of elliptic curves
Math. Research Letters 2 (1995), 299305.
[DO] K. Doi, M. Ohta, On some congruences between cusp forms on
0
(`),
in [Ant5], pp. 91106.
[Dr] V.G. Drinfeld, Elliptic modules, (Russian) Math Sbornik 94 (1974),
594627. (English translation: Math. USSR, Sbornik 23 (1973).)
[DT1] F. Diamond and R. Taylor, Non-optimal levels of mod / modular
representations, Invent. Math. 115 (1994) 435462.
[DT2] F. Diamond and R. Taylor, Lifting modular mod / representations,
Duke Math. J. 74 (1994) 253269.
[Edi] B. Edixhoven, The weight in Serres conjectures on modular forms,
Invent. Math. 109 (1992), 563594.
[Edw] H.M. Edwards, Fermats Last Theorem: A genetic introduction to al-
gebraic number theory, Graduate Texts in Math. 50, Springer-Verlag,
New York, Berlin, Heidelberg, 1977.
160
[Ei] M. Eichler, Quaternare quadratische Formen und die Riemannsche
Vermutung f ur die Kongruenzzetafunktion, Arch. Math. 5 (1954)
355366.
[Fa] G. Faltings, Endlichkeitssatze f ur abelsche Varietaten uber Zahlkorp-
ern, Inv. Math. 73 (1983), 349366. (English translation in [CS].)
[Fl1] M. Flach, A niteness theorem for the symmetric square of an elliptic
curve, Inv. Math. 109 (1992), 307327.
[Fl2] M. Flach, On the degree of modular parametrizations, Seminaire de
theorie des nombres, Paris, 1991-92, 23-36; Prog. in Math. 116,
Birkhauser, Boston, MA 1993.
[FL] J.-M. Fontaine and G. Laaille, Construction de representations j-
adiques, Ann. Sci. Ec. Norm. Super. 15 (1982), 547608.
[FM] J.-M. Fontaine and B. Mazur, Geometric Galois representations, in
[Co], 4178.
[Fo1] J.-M. Fontaine, Groupes nis commutatifs sur les vecteurs de Witt,
C. R. Acad. Sc. 280 (1975), 14231425.
[Fo2] J.-M. Fontaine, Groupes j-divisibles sur les corps locaux, Asterisque
4748 (1977).
[For] O. Forster, Lecures on Riemann Surfaces, Springer-Verlag, New York,
Berlin, Heidelberg, 1981.
[Fr] G. Frey, Links between solutions of 1 = C and elliptic curves,
in: Number Theory, Ulm 1987, Proceedings, Lecture Notes in Math.
1380, Springer-Verlag, New York, Berlin, Heidelberg, 1989, pp. 31
62.
[Gra] A. Granville, The set of exponents for which Fermats Last Theorem
is true, has density one, Comptes Rendus de lAcademie des Sciences
du Canada 7 (1985), 55-60.
[Gre] R. Greenberg, Iwasawa theory for j-adic representations, in: Alge-
braic Number Theory, in honor of K. Iwasawa, Adv. Studies in Pure
Math. 17, Academic Press, San Diego, 1989, pp. 97137.
[Gro] B.H. Gross, A tameness criterion for Galois representations associ-
ated to modular forms modj, Duke Math. J. 61 (1990), 445517.
161
[Ha] R. Hartshorne, Algebraic Geometry, Graduate Texts in Math. 52,
Springer-Verlag, New York, Berlin, Heidelberg, 1977.
[HB] D.R. Heath-Brown, Fermats Last Theorem for almost all expo-
nents, Bull. London. Math. Soc. 17 (1985), 15-16
[He] Y. Hellegouarch, Points dordre 2j
h
sur les courbes elliptiques, Acta
Arith. 26 (1974/75) 253263.
[Hi1] H. Hida, Congruences of cusp forms and special values of their zeta
functions, Inv. Math. 63 (1981), 225261.
[Hi2] H. Hida, A j-adic measure attached to the zeta functions associated
with two elliptic modular forms. I, Inv. Math. 79 (1985), 159195.
[Hi3] H. Hida, Galois representations into G1
2
(Z
p
[[A]]) attached to ordi-
nary cusp forms, Inv. Math. 85 (1986), 545613.
[Hu] B. Huppert, Endliche Gruppen I, Grundlehren Math. Wiss. 134,
Springer-Verlag, New York, Berlin, Heidelberg, 1983.
[Ig] J. Igusa, Kroneckerian model of elds of elliptic modular functions,
Amer. J. Math. 81, 561577 (1959).
[Ih] Y. Ihara, On modular curves over nite elds, in: Proc. Intl. Coll.
on Discrete Subgroups of Lie Groups and Applications to Moduli,
Bombay, 1973, pp. 161202.
[Kam] S. Kamienny, Torsion points on elliptic curves over elds of higher
degree, Inter. Math. Res. Not. 6, 1992, 129133.
[Kat] N.M. Katz, j-adic properties of modular schemes and modular forms,
in [Ant3], pp. 70189.
[Kh] C. Khare, Congruences between cusp forms: the (j. j) case, to appear
in Duke Math. J.
[Ki] F. Kirwan, Complex Algebraic Curves, LMS Student Texts 23, Cam-
bridge Univ. Press, Cambridge, 1993.
[KL] D.S. Kubert, S. Lang, Modular Units. Springer-Verlag, 1981.
[KM] N.M. Katz, and B. Mazur, Arithmetic Moduli of Elliptic Curves,
Annals of Math. Studies 108, Princeton Univ. Press, Princeton, 1985.
162
[Kn] A.W. Knapp, Elliptic Curves, Princeton Univ. Press, Princeton,
1992.
[Koc] H. Koch, Galoissche Theorie der j-Erweiterungen, Springer-Verlag,
New York, Berlin, Heidelberg, 1970.
[Kol] V.A. Kolyvagin, Finiteness of 1(Q) and :/c(1,Q) for a subclass of
Weil curves, Izv. Akad. Nauk. SSSR Ser. Mat. 52 (3) (1988), 522
540. (English translation: Math. USSR, Izvestiya 32 (1989).)
[Ku] E. Kunz, Introduction to Commutative Algebra and Algebraic Ge-
ometry, Birkhauser, Boston, Basel, Berlin, 1985.
[La1] S. Lang, Algebraic Number Theory, Addison-Wesley, Reading, MA
1970.
[La2] S. Lang, Introduction to Modular Forms, Grundlehren Math. Wiss.
222, Springer-Verlag, New York, Berlin, Heidelberg, 1976.
[Leg] A.M. Legendre, Sur quelques objets danalyse indeterminee et par-
ticuli`erement sur le theor`eme de Fermat, Mem. Acad. R. Sc. de
lInstitut de France 6, Paris, 1827.
[Len] H.W. Lenstra, Complete intersections and Gorenstein rings, in [Co],
pp. 99109.
[Liv] R. Livne, On the conductors of mod / representations coming from
modular forms, J. Number Theory 31 (1989), 133141.
[Ll1] R.P. Langlands, Modular forms and /-adic representations, in [Ant2],
pp. 361500.
[Ll2] R.P. Langlands, Base Change for G1(2), Annals of Math. Studies
96, Princeton Univ. Press, Princeton, 1980.
[LO] S. Ling, J. Oesterle, The Shimura subgroup of J
0
(`), Asterisque
196197 (1991), 171203.
[Mat] H. Matsumura, Commutative Ring Theory, Cambridge Studies in
Adv. Math. 8, Cambridge Univ. Press, Cambridge, 1986.
[Maz1] B. Mazur, Modular curves and the Eisenstein ideal, Publ. Math.
IHES 47 (1977), 33186.
163
[Maz2] B. Mazur, Rational isogenies of prime degree, Inv. Math. 44 (1978),
129162.
[Maz3] B. Mazur, Deforming Galois representations, in: Galois groups over
Q, Y. Ihara, K. Ribet, J-P. Serre, eds., MSRI Publ. 16, Springer-
Verlag, New York, Berlin, Heidelberg, 1989, pp. 385437.
[Mc] W. McCallum, On the method of Coleman and Chabauty, Math. Ann.
299 (1994), 565596.
[Mer] L. Merel, Bornes pour la torsion des courbes elliptiques sur les corps
de nombres, Inventiones Math., to appear.
[Mes] J-F. Mestre, Constructions polynomiales et theorie de Galois, Pro-
ceedings of the ICM, Zurich, August 1994.
[Mi] J.S. Milne, Arithmetic Duality Theorems, Perspectives in Math.,
Academic Press, San Diego, 1986.
[MRi] B. Mazur, K. Ribet, Two-dimensional representations in the arith-
metic of modular curves, Asterisque 196197 (1991), 215255.
[MRo] B. Mazur, L. Roberts, Local Euler characteristics, Inv. Math. 9
(1970), 201234.
[MT] B. Mazur, J. Tilouine, Representations galoisiennes, dierentielles
de Kahler et conjectures principales, Publ. Math. IHES 71 (1990),
65103.
[Mu1] V.K. Murty, Introduction to Abelian Varieties, CRM Monograph
Series 3, AMS Publ., Providence, 1993.
[Mu2] V.K. Murty, ed., Seminar on Fermats Last Theorem, CMS Confer-
ence Proceedings 17, AMS Publ., Providence, 1995.
[MW] B. Mazur, A. Wiles, Class elds of abelian extensions of Q, Inv.
Math. 76 (1984), 179330.
[Na] K. Nagao, An example of an elliptic curve over Q(1) with rank 13,
Proc. of The Japan Acad., 70, 1994, 152-153.
[Oda] T. Oda, The rst De Rham cohomology group and Dieudonne mod-
ules, Ann. Sci. Ec. Norm. Super. 2 (1969), 63135.
164
[Oe] J. Oesterle, Nouvelles aproches du theor`eme de Fermat, Seminaire
Bourbaki no. 694 (1987-88), Asterisque 161162 (1988), 165186.
[Ogg] A. Ogg, Modular Forms and Dirichlet Series, Benjamin, New York,
1969.
[Oo] F. Oort, Commutative Group Schemes, Lecture Notes in Math. 15,
Springer-Verlag, New York, Berlin, Heidelberg, 1966.
[Ram] R. Ramakrishna, On a variation of Mazurs deformation functor,
Compositio Math. 87 (1993), 269286.
[Ray] M. Raynaud, Schemas en groupes de type (j. j. .... j), Bull. Soc. Math.
France 102 (1974), 241280.
[R1] K. Ribet, A modular construction of unramied j-extensions of
Q(j
p
), Inv. Math. 34, 151-162.
[R2] K. Ribet, Twists of modular forms and endomorphisms of abelian
varieties, Math. Ann. 253 (1980), 43-62.
[R3] K.A. Ribet, The /-adic representations attached to an eigenform with
Nebentypus: a survey, in [Ant5], pp. 1752.
[R4] K.A. Ribet, Congruence relations between modular forms, Proc. ICM,
1983, 503514.
[R5] K.A. Ribet, On modular representations of Gal (

Q,Q) arising from

modular forms, Inv. Math. 100 (1990), 431476.
[R6] K.A. Ribet, Report on mod / representations of Gal (

Q,Q), in: Mo-

tives, Proc. Symp. in Pure Math. 55 (2), 1994, pp. 639676.
[R7] K.A. Ribet, Abelian varieties over Q and modular forms, 1992 Pro-
ceedings of KAIST Mathematics Workshop, Korea Advanced Insti-
tute of Science and Technology, Taejon, 1992, pp. 5379.
[R8] K.A. Ribet, Galois representations and modular forms, Bull. AMS.
32 (1995), 375402.
[Ri] P. Ribenboim, 13 Lectures on Fermats Last Theorem, Springer-
Verlag, New York, Berlin, Heidelberg, 1979.
[Sa] T. Saito, Conductor, discriminant and the Noether formula for arith-
metic surfaces, Duke Math. J. 57 (1988), 151173.
165
[Se1] J.-P. Serre, Cohomologie Galoisienne, Lecture Notes in Math. 5,
Springer-Verlag, New York, Berlin, Heidelberg, 1964.
[Se2] J.-P. Serre, Corps Locaux, Hermann, Paris, 1962.
[Se3] J.-P. Serre, Alg`ebre Locale, Multiplicites, Cours au Coll`ege de France,
1957-58, Lecture Notes in Math. 11, Springer-Verlag, New York,
Berlin, Heidelberg, 1965.
[Se4] J.-P. Serre, A Course in Arithmetic, Graduate Texts in Math. 7,
Springer-Verlag, New York, Berlin, Heidelberg, 1973.
[Se5] J.-P. Serre, Abelian /-adic Representations and Elliptic Curves, Ben-
jamin, New York, 1968.
[Se6] J.-P. Serre, Proprietes galoisiennes des points dordre ni des courbes
elliptiques, Inv. Math. 15 (1972), 259331.
[Se7] J.-P. Serre, Sur les representations modulaires de degre 2 de
Gal (

Q,Q), Duke Math. J. 54 (1987), 179230.

[Sha] S. Shatz, Group schemes, formal groups, and j-divisible groups, in
[CS], pp. 2978.
[Shi1] G. Shimura, Correspondances modulaires et les fonctions de courbes
algebriques, J. Math. Soc. Japan 10 (1958), 128.
[Shi2] G. Shimura, Introduction to the Arithmetic Theory of Automorphic
Functions, Princeton Univ. Press, Princeton, 1971.
[Shi3] G. Shimura, On the factors of the jacobian variety of a modular func-
tion eld, J. Math. Soc. Japan 25 (1973), 523544.
[Shi4] G. Shimura, On the holomorphy of certain Dirichlet series, Proc.
LMS (3) 31 (1975), 7998.
[Shi5] G. Shimura, The special values of the zeta functions associated with
cusp forms, Comm. Pure Appl. Math. 29 (1976), 783804.
[Si1] J. Silverman, The Arithmetic of Elliptic Curves, Graduate Texts in
Math. 106, Springer-Verlag, New York, Berlin, Heidelberg, 1986.
[Si2] J. Silverman, Advanced Topics in the Arithmetic of Elliptic Curves,
Graduate Texts in Math. 151, Springer-Verlag, New York, Berlin,
Heidelberg, 1994.
166
[St] J. Sturm, Special values of zeta functions, and Eisenstein series of
half integral weight, Amer. J. Math. 102 (1980), 219240.
[ST] J.-P. Serre, J.T. Tate, Good reduction of abelian varieties, Annals of
Math. 88 (1968), 492517.
[Su] J. Suzuki, On the generalized Wieferich criterion, Proc. Japan Acad.
70 (1994), 230-234.
[Ta] J.T. Tate, Algorithm for determining the type of a singular ber in
an elliptic pencil, in [Ant4], pp. 3352.
[Te] G. Terjanian, Sur lequation r
2p
+
2p
= .
2p
, C.R. Acad. Sci. Paris,
285 (1977) 973-975.
[Th] F. Thaine, On the ideal class groups of real abelian number elds,
Annals of Math. 128 (1988), 118.
[Tu] J. Tunnell, Artins conjecture for representations of octahedral type,
Bull. AMS 5 (1981), 173175.
[TW] R. Taylor, A. Wiles, Ring theoretic properties of certain Hecke alge-
bras, Annals of Math. 141 (1995), 553572.
[Wa] L. Washington, Introduction to Cyclotomic Fields, Graduate Texts
in Math. 83, Springer-Verlag, New York, Berlin, Heidelberg, 1982.
[We1] A. Weil, Varietes Abeliennes et Courbes Algebriques, Hermann,
Paris, 1948.
[We2] A. Weil,

Uber die Bestimmung Dirichletscher Reihen durch Funktion-
algleichungen, Math. Ann. 168 (1967), 165172.
[W1] A. Wiles, On j-adic representations for totally real elds, Annals of
Math. 123 (1986), 407456.
[W2] A. Wiles, On ordinary -adic representations associated to modular
forms, Inv. Math. 94 (1988), 529573.
[W3] A. Wiles, Modular elliptic curves and Fermats Last Theorem, Annals
of Math. 141 (1995), 443551.
167